IFSR International Series on Systems Science and Engineering

George J. Friedman
Phan Phan

Constraint
Theory
Multidimensional Mathematical Model
Management
Second Edition
CONSTRAINT THEORY
MULTIDIMENSIONAL
MATHEMATICAL MODEL
MANAGEMENT
International Federation for Systems Research International
Series on Systems Science and Engineering
Editor-in-Chief: George J. Klir
Binghampton State University

Series Editor: George J. Klir

Binghampton State University

Editorial Board
Gerrit Broekstra Ivan M. Havel
Erasmus University, Rotterdam, Charles University, Prague,
The Netherlands Czech Republic
John L. Castii Klaus Kornwachs
Sante Fe Institute, New Mexico Technical University of Cottbus,
Brian Gaines Germany
University of Calgary, Canada Franz Pichler
University of Linz Austria
This is the last volume in the IFSR series with the above stated editorial board. A reconstituted editorial
board with George Mobus as Editor-in-Chief will continue the series.

IFSR was established “to stimulate all activities associated with the scientific study of systems and to coordinate
such activities at international level.” The aim of this series is to stimulate publication of high-quality
monographs and textbooks on various topics of systems science and engineering. This series complements
the Federations other publications.

A Continuation Order Plan is available for this series. A continuation order will bring delivery of each
new volume immediately upon publication. Volumes are billed only upon actual shipment. For further
information please contact the publisher.

Volumes 1–6 were published by Pergamon Press

More information about this series at http://www.springer.com/series/6104

 CONSTRAINT THEORY
MULTIDIMENSIONAL
MATHEMATICAL MODEL
MANAGEMENT

Second Edition

George J. Friedman • Phan Phan

George J. Friedman Phan Phan
Viterbi School of Engineering Viterbi School of Engineering
University of Southern California University of Southern California
Los Angeles, CA, USA Los Angeles, CA, USA

ISSN 1574-0463
IFSR International Series on Systems Science and Engineering
ISBN 978-3-319-54791-6 ISBN 978-3-319-54792-3 (eBook)
DOI 10.1007/978-3-319-54792-3

Library of Congress Control Number: 2017933682

© Springer International Publishing AG 2005, 2017

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.

Printed on acid-free paper

This Springer imprint is published by Springer Nature

The registered company is Springer International Publishing AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Fronticepiece

24” Doily, designed and constructed by Regina Oberlander (Dr. Friedman’s

mother) circa 1915 in Chinedeev, near Muncacevo, Austro-Hungarian
Empire.

A single circuit cluster with over 1,000 independent simple circuits (thus
there are 1000 more edges than vertices)
 This book is dedicated to our
wives, who have helped us find
the time and priority for our
mathematical dreams.
Preface

At first glance, this might appear to be a book on mathematics, but it is

really intended for the practical engineer who wishes to gain greater control
of the multidimensional mathematical models which are increasingly an
important part of his environment. Another feature of the book is that it
attempts to balance left- and right-brain perceptions; the authors have
noticed that many graph theory books are disturbingly light on actual
topological pictures of their material.
Constraint Theory was originally defined by George Friedman in his PhD
dissertation at UCLA in 1967 and subsequent papers written over the
following decade. There was a dearth of constraint theory publication after
the 1970’s as Dr. Friedman was working on several classified aerospace
programs wherein publication of any kind was most difficult. The first
edition of this book was published in 2005. Constraint Theory was further
extended by Phan Phan in his PhD dissertation at USC in 2011, leading to
this second edition.
Acknowledgments

Over the past several years, Constraint Theory has been a substantial part
of a special graduate course in the Systems Architecture and Engineering
program at the University of Southern California, where the authors are
adjunct faculty members. The feedback from this group of bright graduate
students was invaluable. Special thanks are given to these graduate students
who performed research studies – summarized in Appendix A – on
Constraint Theory: Leila Habibibabadi, Kathya Zamora-Diaz and Elliott
Morgan [1]. Extra special thanks are given to Mr. Gary L. Friedman who
provided the extremely valuable service of intellectual reflector, editor and
digital format manager for this book.
Introduction

Many thousands of papers have been written about the accelerating pace
of increased complexity and interactivity in virtually every walk of life in the
developed world. Domains which previously could have been studied and
managed separately – such as energy, the environment and economics – must
now be dealt with as intimately intertwined disciplines. With its multitude of
additional capabilities, complex systems also provide a treacherous array of
fragile failure modes, and both the development and operation of new
systems are an increasing challenge to the systems engineer. Advanced
technology is the primary driving force behind the increasing complexity and
the enthusiastic pushing of immature technologies is behind most of the early
failures in the development phases.
Perhaps the most significant advanced technology employed in new
complex systems is the computer science family with its ancillary disciplines
of communications and software. Fortunately, computer science also repre-
sents a major opportunity to control the design and operation of complex
systems because of its ability to perform multi-dimensional modeling to any
level of detail desired. Math models have been used in support of every phase
of systems engineering, including requirements management, preliminary
design, interface and integration, validation and test, risk management,
manufacturing, reliability and maintainability, training, configuration man-
agement and developing virtual universes to measure customer preferences
prior to the implementation of the design. Properly used, the enormous power
of modern computers can even furnish researchers with a synthetic world
where theories can be tried and tested on validated models, thus avoiding far
more expensive tests in the real world. A wide variety of questions – or
“tradeoffs’ – can be asked of the models and, at least in theory, the analyst
has a free choice as to which computations he wishes to observe and which
variables he desires to be independent. Philosophically, it can even be argued
that the math model employed in this fashion provides the technologist a
virtual extension of the scientific method itself.
Those who have actually wrestled with large-scale models will complain
that the above description is far too rosy. Submodels which are developed by
xiv Constraint theory

separate organizations are normally very difficult to integrate into an overall

working model; they often must be dealt with as “islands of automation.”
The greatest of care must be taken to make sure that the definition of each
variable is clear and agreeable to every member of the team. In general, it is
difficult to distinguish between a model and the computer program, and if a
computational request is made which reverses dependent and independent
variables, then the model must be reprogrammed. To say the least, much
diligent effort must be undertaken to obtain the many advantages promised
by mathematical modeling.
However, even after the diligence, there exists a much deeper problem
that often diminishes the utility of math modeling; it is associated with the
traditional “well posed” problem in mathematics. We need to know whether
the model is internally consistent and whether computational requests made
on it are allowable. The alarming facts are that models constructed by diverse
teams – and this is normally the case for very large models – have internal
inconsistencies and that most of the possible computational requests which
can be made on even consistent models are not allowable. This problem is
the domain addressed by Constraint Theory and is the subject of this book.
Chapter 1 provides an example of low dimension, showing how problems
of consistency and computational allowability can arise in even simple
situations. The reader is introduced to the two main characters of the book –
an experienced manager and an analyst – whose dialogue will hopefully
illuminate the book’s many concepts. The bipartite graph is introduced, as are
a few simple rules. However, the analyst argues that, in order to expand the
tools to models of very high dimension, and in order to trust the reliability of
these tools, the theory must be based on a more rigorous foundation. “Only
the simplest 5% of graph theory and set theory are required”, he claims.
Chapter 2 begins to establish the rigorous foundation by defining four
“views” of a mathematical model: 1) set theoretic, 2) submodel family, 3)
bipartite graph, and 4) constraint matrix. The first two views are full models;
the last two views are metamodels. Then, rigorous definitions of consistency
and computational allowability are made in the context of these views.
Chapter 3 discusses the similarities between language and mathematics
and provides some general consistency and computability results with respect
to any class of relation. In order to provide a basis for the next three chapters,
three classes of exhaustive and mutually exclusive relations are defined:
discrete, continuum, and interval.
Due to the amount of new materials resultant from Dr. Phan’s research as
part of his doctoral dissertation, previous Chapter 4 in the first edition has
been split into two new Chapters 4 and 5 in this second edition. Chapters 4
and 5 represent the core of constraint theory at its present stage.
Introduction xv

As before, the new Chapter 4 addresses the constraint theoretic properties

of regular relations, the most important type within the continuum class, and
the most often employed in the development of multidimensional math
models. The simple rules presented in Chapter 1 are rigorously proved
employing the foundations of Chapters 2 and 3. The topological properties of
the bipartite graph are analyzed to provide key conclusions of the model’s
consistency and computational properties.
A specific type of subgraph within the bipartite graph, called the Basic
Nodal Square (BNS) is identified as the “kernel of intrinsic constraint” and is
accused of being the culprit in model inconsistency and unallowable com-
putability. Trivially easy computations on the bipartite graph – such as circuit
rank and constraint potential – are shown to have enormous utility in locating
the BNSs which hide in tangled circuit clusters.
Additionally, the newly updated Section 4.6 extends, in more prescriptive
details with graphical illustrations, the step by step algorithm for locating
BNSs within a model graph.
The new Chapter 5 discusses the general issue of constraint propagation
through a connected model graph of regular relations. A detailed procedure
for determining model consistency and computational allowability in such a
model is introduced.
In particular, Section 5.4 introduces new mathematical definitions and
theorems to enable the detection of overlapping BNSs. Section 5.5 outlines
techniques to relieve over-constraint among them. Section 5.6 describes a
precise procedure to expand their resultant constraint domains. Section 5.7
prescribes, in details with graphical illustrations, a step-by-step algorithm to
process computation requests made on a model. And Section 5.8 provides a
constraint theory toolkit to employ the rules and theorems in an orderly
manner and which can find BNSs trillions of times faster than brute force
approaches.
Chapter 6 addresses the constraint properties of discrete and interval
functions such as those from Boolean algebra, logic and inequalities. These
classes of relations are less important in support of modern math modeling,
but strangely, it was the first that the author studied in his development of
Constraint Theory. It was easier for him to imagine multidimensional sets of
points than multidimensional sets of continuous functions. Interval relations
require the greatest interaction between models and metamodels, and the
concept of constraint potential is less useful than for regular relations.
Chapter 7 provides a compact structure of constraint theory. All postu-
lates, definitions and theorems are listed and their logical interrelationships
are displayed in the form of quasi-bipartite graphs.
xvi Constraint theory

Chapter 8 presents detailed examples of the application of constraint

theory to the areas of operations analysis, kinematics of free-fall weapon
delivery systems and the dynamics of deflecting asteroids with mass drivers.
Chapter 9 summarizes the book and provides the manager and analyst a
final opportunity to dialogue and discuss their common background.
Problems for the interested student are presented at the end of most
chapters, so this book could be employed as a text for a graduate course – or
senior level undergraduate course – in Systems Engineering or mathematical
modeling.
Of course, a complete list of references is provided, as well as an index.
Several appendices treat detailed material to a depth that would slow
down the natural rhythm of the exposition if they were included in the
chapters themselves. Appendix A is noteworthy in that it summarizes the
research projects on “computational request disappointments.” On models
approximately the size of Chapter 1’s “simple example” – eight variables –
the percentage of allowable computational requests based on the total number
of possible computational requests is only on the order of 10%. It is presently
“Friedman’s conjecture” that as the dimensionality, K, of the model
increases, the number of allowable computational requests also increases,
perhaps as fast as the square of the model’s dimension or K2. However, the
number of possible computational requests increases far faster: 2K. Thus, for
a 100-dimension model, only 10-26 of all possible computational requests will
be allowable! Models of thousands of dimensions have been built and are
planned; so the ratio of allowable to possible computational requests is
enormously worse that even this incredibly low number. The technologist
who wishes to gain maximum benefit from asking his model to perform any
computation his imagination conjures up will certainly be disappointed! A
tool such as constraint theory which will lead him to the 10,000 computatio-
nal requests (K=100) or 1,000,000 requests (K=1,000) which are allowable
should be valuable.
Appendix B provides a very brief overview of graph theory with the
objective of justifying why the bipartite graph was chosen as the primary
metamodel for constraint theory.
Appendix C describes the rigorous logic of the difference between “if and
only if” and “if” types of theorems. Most of constraint theory’s theorems are
of the latter category – a source of confusion to many students.
The newly updated Appendix D establishes fundamental algebraic struc-
tures which are essential to implement constraint theory. These include defini-
tions and properties of general vector spaces, and binary set operations.
A Warmup Problem in Complexity

This book makes substantial use of a mathematical structure from graph

theory called a bipartite graph. In the past, bipartite graphs have been
employed to solve “pairing” problems associated with various social
situations such as picnics or dinner parties.
Out of respect for this tradition, let us consider a set of five men – named
Jack, Jake, Jude, Juan, and Jobe – and a set of five women – named Jane,
Joan, June, Jean, and Jenn. Let us define a relationship pattern as a complete
description of all heterosexual relationships between the five men and five
women. For example:
• In the communal pattern, every man has a relationship with every
woman. There is one such pattern.
• In the celibacy pattern, none of the men have a relationship with
any of the women. Again, there is one such pattern.
• In the male harem patterns, one of the men has a relationship
with each of the women, but all the other men are devoid of
relationships, except perhaps to be eunuchs. There are five such
patterns. Similarly, there are five possible female harem patterns.
• In the monogamy patterns, each man has a relationship with
exactly one woman and vice versa. There are 5!=120 such
distinct patterns.
And so on. There are many more patterns. The question is: What is the
total number of possible heterosexual relationship patterns between five men
and five women?
The answer – discussed in Chapter 4 and Appendix A – may surprise
you: it’s over 30 million (!). It certainly surprised the author and changed an
xviii Constraint theory

important objective of his research agenda. Moreover it represents the

hidden depths possible in apparently simple problems of low dimension as
well as a challenge to one’s belief in intuition or rational mathematics.
About the Authors

George Friedman is a Professor of Practice in the

Astronautical Engineering Department of the
Viterbi School of Engineering of the University of
Southern California. He has developed and taught
graduate courses in systems engineering with
emphasis on the management of complexity and
decision science. This book is the product of one of
these courses.
He has had over 45 years of experience in
industry, retiring from the Northrop Corporation as
their Corporate Vice President of Engineering and
Technology. He worked on a wide variety of
aerospace programs and served as a consultant to all branches of the
Department of Defense, NASA, the National Science Foundation, and
Department of Energy as well as to the NATO industrial advisory group.
He was a founder of the International Council on Systems Engineering
(INCOSE), served as its third president, was elected a fellow and is on the
editorial board of INCOSE’s journal, Systems Engineering.
He has also been a member of the Institute of Electrical and Electronic
Engineers (IEEE) since its formation from the IRE and AIEE, was elected a
fellow and was the vice president for publications of the IEEE Transactions
on Aerospace and Electronics Systems. He received the Baker Prize for the
best paper published by all societies of the IEEE in 1970 – the subject of the
paper was Constraint Theory.
xx Constraint theory

He was a former director of research at the Space Studies Institute at

Princeton, and had supported several new technologies associated with the
long range development of space.
He received the Bachelor’s degree in engineering at the University of
California at Berkeley and the Masters and Doctorate at UCLA. The topic of
his PhD dissertation was constraint theory [2, 3].

Phan Phan is a Lecturer in the Astronautical

Engineering Department of the Viterbi School of
Engineering of the University of Southern
California (USC). He has assisted and taught
graduate courses in systems engineering, systems
management, lean operations and economic
analysis.
He has had over 36 years of technical and
managerial experience in government, military and
various industries, including oil & gas exploration,
commercial and military aircraft, unmanned
sensors, and major weapon systems.
His industry assignments included Mobil Research & Development,
General Dynamics, Lockheed Aeronautical Systems, McDonnell Douglas
and Boeing Integrated Defense Systems. As a registered Professional
Engineer in California, he currently works as a reliability analyst with the
Naval Surface Warfare Center – Corona Division.
He has also served in the U.S. Navy Reserve as an Engineering Duty
Officer, attaining the rank of Captain, and currently assigned to Naval Sea
Systems Command. His previous Navy assignments included Office of
Naval Research/Naval Research Laboratory, Program Executive Office
Integrated Warfare Systems, Naval Space & Warfare Systems Command,
Mobile Mine Assembly Group, Naval Weapons Station Seal Beach and
Naval Shipyard Long Beach.
He received his B.S. in Engineering from the University of Alabama in
Huntsville, Master of Engineering from the University of Texas at Arlington,
MBA from California State University – Fullerton, M.S. in Systems
Architecture & Engineering from USC, Master of Engineering Acoustics
from the Naval Postgraduate School, and Ph.D. in Industrial & Systems
Engineering from USC.
The topic of his doctoral dissertation was “Expanding Constraint Theory
to Determine Well-Posedness of Large Mathematical Models” [22], the main
contribution to the second edition of this book.
Contents

CHAPTER 1 – MOTIVATIONS 1
WHAT IS CONSTRAINT THEORY AND WHY IS IT IMPORTANT?
1.1 TRENDS AND PROBLEMS IN SYSTEM TECHNOLOGIES
1.2 AN EXAMPLE OF LOW DIMENSION
1.3 THE MANAGER AND ANALYST CONTINUE THEIR DIALOGUE
1.4 PRELIMINARY CONCLUSIONS
1.5 A LITTLE WINDOW INTO FUTURE CHAPTERS
1.6 PROBLEMS FOR THE CURIOUS READER

CHAPTER 2: THE FOUR-FOLD WAY 25

HOW TO PERCEIVE COMPLEX MATHEMATICAL MODELS AND
WELL-POSED PROBLEMS
2.1 PROLOGUE: THE MANAGER AND ANALYST DISCUSS THE
ORIGINS OF MULTIDIMENSIONAL MODELS AND WELL-
POSEDNESS
2.2 THE FIRST VIEW: SET THEORETIC
2.3 THE SECOND VIEW: FAMILY OF SUBMODELS
2.4 THE THIRD VIEW: THE BIPARTITE GRAPH
2.5 THE FOURTH VIEW: THE CONSTRAINT MATRIX
2.6 MODEL CONSISTENCY AND COMPUTATIONAL ALLOWABILITY
2.7 THE MANAGER AND ANALYST CONTINUE THEIR DIALOGUE
2.8 CHAPTER SUMMARY
2.9 PROBLEMS FOR THE INTERESTED STUDENT
xxii Constraint theory

CHAPTER 3: GENERAL RESULTS 49

FROM PROTOMATH TO MATH TO METAMATH
3.1 LANGUAGE AND MATHEMATICS
3.2 MOST GENERAL TRUSTWORTHY RESULTS
3.3 CLASSES OF RELATIONS
3.4 MANAGER AND ANALYST REVISITED
3.5 CHAPTER SUMMARY
3.6 PROBLEMS FOR THE GENERAL STUDENT

CHAPTER 4: REGULAR RELATIONS 61

SEARCHING FOR THE KERNELS OF CONSTRAINT
4.1 COGNITIVE BARRIERS TO CIRCUITS
4.2 NODE, KNOT AND BASIC NODAL SQUARE SANCTIFICATION
4.3 USEFUL PROPERTIES OF BIPARTITE GRAPHS
4.4 CORNERING THE CULPRIT KERNELS; TEN EASY PIECES
4.5 CONTINUING THE PURSUIT INSIDE THE CIRCUIT CLUSTERS (CC)
4.6 LOCATING BNSS WITHIN A MODEL GRAPH
4.7 QUERIES FOR THE REGULAR STUDENT

CHAPTER 5: MODEL CONSISTENCY AND COMPUTATIONAL 101

ALLOWABILITY
5.1 ZERO CONSTRAINT ALL ALONG THE COMPUTATIONAL PATH
5.2 RECAPITULATION OF COMPUTATIONAL FLOW
5.3 GENERAL PROCEDURE FOR DETERMINING CONSISTENCY
AND ALLOWABILITY IN A MODEL OF REGULAR RELATIONS
5.4 DETECTION OF OVERLAPPING BNS
5.5 RELIEF OF OVER-CONSTRAINT
5.6 EXPANSION OF RESULTANT CONSTRAINT DOMAINS
5.7 PROCESSING OF COMPUTATIONAL REQUESTS
5.8 SUMMARY OF CHAPTER AND CONSTRAINT THEORY TOOLKIT
5.9 QUERIES FOR THE REGULAR STUDENT

CHAPTER 6: DISCRETE AND INTERVAL RELATIONS 137

THE DIMINISHED UTILITY OF METAMODELS
6.1 METAMODEL ISSUES AND PERSPECTIVES
6.2 THE GENERAL TAXONOMY AND PRIMARY PROPERTY OF
DISCRETE RELATIONS
6.3 BOOLEAN RELATIONS
6.4 TOPOLOGICAL IMPLICATIONS
6.5 ALLOWABILITY OF DISCRETE COMPUTATIONS
6.6 INEQUALITY RELATIONS
Table of Contents xxiii

6.7 SUMMARY
6.8 PROBLEMS FOR THE DISCRETE STUDENT

CHAPTER 7: THE LOGICAL STRUCTURE OF CONSTRAINT THEORY 155

A COMPACT SUMMARY
7.1 OVERVIEW
7.2 POSTULATES AND PHILOSOPHICAL ASSUMPTIONS
7.3 DEFINITIONS
7.4 THEOREMS
7.5 GRAPHS OF THE LOGICAL STRUCTURE OF CONSTRAINT THEORY
7.6 COMPLETENESS

CHAPTER 8: EXAMPLES OF CONSTRAINT THEORY APPLIED TO REAL-

WORLD PROBLEMS 163
8.1 APOLOGIES NOT REQUIRED
8.2 COST AS AN INDEPENDENT VARIABLE (CAIV)
8.3 THE KINEMATICS OF FREE-FALL WEAPONS
8.4 THE DEFLECTION OF AN EARTH-THREATENING ASTEROID
EMPLOYING MASS DRIVERS

CHAPTER 9: MANAGER AND ANALYST MEET AGAIN 183

GISTS AND SCHIZOPHRENIA

APPENDICES

APPENDIX A: COMPUTATIONAL REQUEST DISAPPOINTMENTS; 189

RESULTS OF THE USC ALLOWABILITY PROJECT

APPENDIX B: GRAPH THEORY OVERVIEW 197

APPENDIX C: THE LOGIC OF “IF” AND “IF AND ONLY IF” 201

APPENDIX D: ALGEBRAIC STRUCTURES 203

REFERENCES 211

INDEX 215
Chapter 1 MOTIVATIONS

What is Constraint Theory and why is it important?

1.1 TRENDS AND PROBLEMS IN SYSTEM

TECHNOLOGIES

Gone forever are the simple days! Virtually every identifiable trend is
driving humanity’s enterprises into more intimate interaction and conflict.
Increased population, accelerated exploitation of resources, and expanded
transportation have brought the previously decoupled worlds of economics,
energy and the environment into direct conflict. With the greater efficiency
of travel and communication, the emergence of global marketplaces and the
revolution in military strategies, the international world is incredibly more
interactive, multidimensional and complex than even a decade ago. Locally,
we observe ever tighter coupling between emerging problems in crime,
poverty, education, health and drug misuse. All these issues have been
aggravated by an explosion of new technology and – especially in the United
States – a compulsion to force these new technologies into early and often
simultaneous application. The most vigorous of these advancing
technologies – digital computation – brings with it an unexpected
complexity challenge: software and the management of highly complex and
multidimensional mathematical models.
Fortunately, this most rapidly advancing technology of computer science
not only adds to the complexity of designed systems, it also contributes
enormously to designing these systems themselves. A host of new “computer
assisted” software packages are published each year, running the gamut from
Computer Assisted Design (CAD), Computer Assisted Engineering (CAE),
Computer Assisted Systems Engineering (CASE), Computer Assisted
Manufacturing (CAM), Computer Assisted Instruction (CAI), and eventually
© Springer International Publishing AG 2017
G.J. Friedman, P. Phan, Constraint Theory, IFSR International Series on Systems
Science and Engineering, DOI 10.1007/978-3-319-54792-3_1
2 Constraint theory

to Computer Assisted Enterprise Management (CAEM). This family of tools

permits the design engineers to control a virtually unlimited number of
variables, to predict behaviors and performance of systems still in their early
conceptual stages, to optimize with respect to detailed criteria, to effect
interdisciplinary integration and to perform design changes with
unprecedented speed and accuracy. It is not an exaggeration to claim that
without this array of computer-based tools, many systems that exist today
would have been impossible to design and implement.
However, as most systems engineers who attempt to gain benefit from
these tools are well aware, computer-based design is a mixed blessing. A
common complaint is that the various programs which support facets of the
total problem are “islands of automation” – they are difficult to integrate into
a total system problem solving capability. Another problem is that the tools
are virtually useless in sorting out the variety of languages and technical
shades of meaning, especially on highly interdisciplinary systems. Yet
another challenge for the engineering and program managers is the vigorous
and frequent upgrading of every hardware and software package causing
unprecedented costs of initial installation and training to the overall design
process, not to mention the inevitable bugs in the early versions. Many
companies have even established new organizations whose members are
expert in computer-assisted programs, and not expert in the technical design
itself.
Observers who watched the agonizing entry of computers over the last
several decades into many diverse worlds such as financial management,
stock market trading, airline ticketing, air traffic control, and education
should be optimistic that eventually computer-based design will also become
an efficient tool which will become so easy to use that investing in it will be
clearly justified. But in order for this dream to occur, there are more
problems to solve – deeper problems than getting the definitions sorted out
and software packages to work together.
Even when the willing but cognitively challenged computational giants
of computer based tools are completely manageable, several fundamental
problems will still exist, mostly on the cognition and mathematical levels.
Nothing is said in any of the present set of textbooks on Systems
Engineering about the regrettable “subdimensionality” of the human
intellect. Despite the fact that a thorough description of a modern complex
system requires the understanding and integration of hundreds to thousands
of variables, cognitive scientists have known for decades that the human
mind is limited in its perceptive powers to a mere half-dozen dimensions.
Regardless of all our other miraculous gifts such as language, art, music,
imagination, judgment, and conscience, our dimensional perceptive power is
1. Motivations 3

tiny compared to the challenges of designing complex systems, and sadly, it

appears that this is a “wired in” shortcoming of our nervous system and thus
we cannot hope to be trained to attain a higher perceptive capability. This
“dimensionality gap” is severely aggravated by the habit many self-styled
“decisive managers” have in further suppressing their limited perception by
searching for simplifications such as “the bottom line,” “the long pole in the
tent” or “getting the right angle” in attempting to make complex descriptions
more comprehensible to them. Typically, when the dimensionality of the
model overwhelms that of the decision maker, and he sees results which
appear to be anti-intuitive, he will tend to distrust these results as the product
of software bugs or other errors. Thus, major opportunities to learn from the
enhanced power of modeling are lost because the operation of the computer
becomes more and more opaque to the decision maker as the dimensionality
increases.
As an aside, we humans also have problems with numbers: we cannot
perceive 29 in the same way we perceive 5. The raw arithmetic perception of
the average person is “the magic number seven, plus or minus two,”
according to the cognitive scientist George Miller. However, after a lifetime
of dependable experience with arithmetic algorithms, we have the illusion
that we can understand and manage entities such as 29, or even
29,000,000,029.
The other fundamental problem was previously referred to as the “well-
posed” problem in mathematics. That is, when a mathematical model is
established, is it internally consistent? When computational requests are
desired based on this model, are they allowable? If the answer to either
question is “no,” then we have a situation which is not “well posed” and we
can expect nonsensical results or jammed up attempts to program. This
problem is made worse by the fact that in most digital computer programs,
models are built with a unidirectional computational flow that was
anticipated by the programmers, but is not necessarily responsive to the
needs of the decision makers. It was a source of great irritation to this author
to be told many times over his career that a computational request was
“impossible” because the model was programmed with another
computational flow in mind. However, when the reprogramming was done
in an attempt to be more responsive, more fundamental problems frequently
arose.
An example of these problems will be useful at this point. The example
given in the next section was chosen to be as simple as possible, but still
indicating aspects of the well posed problem that can arise even without our
entering a dimension so high that our perceptions are boggled.
4 Constraint theory

1.2 AN EXAMPLE OF LOW DIMENSION

A decision-making manager was authorized to initiate the preliminary

design of a new system development by his board of directors. In the true
spirit of systems engineering, he realized the importance of making the best
decisions as early in the system development process as possible.
Accordingly, he gathered a team of the best specialists available, along with
a systems analyst to help organize the math model that he hoped would
guide him to strategic systems tradeoffs and decisions.
The chief systems engineer stressed that, in order for an “optimum
design” to exist, it was necessary to define a total systems optimization
criterion, T:

T = PE/C (1)

where:
P was the political index of acceptability by the board of directors,
E was the system effectiveness, and
C was the life cycle cost of the system.

The operational chief, expressing a weariness with the overly aggressive

use of new and unproved technology on most of his previous systems,
wanted to stress that most of the total system cost should be applied to
operations and support, not new systems development. Thus, he contributed
this limitation:

D = k1C, where k1=0.3 (2)

where D, the development cost, was to be limited to 30% of the total cost.

The operations and support specialist, attempting to predict the level of

cost after production and delivery were complete, provided:

S = X + 0.5D (3)

where:
S is the total support cost
X is the cost of ops and support if there were no new technology
D is the development cost for the system, including new technology.

The systems costing and estimating specialist contributed the obvious:

1. Motivations 5

C=D+S (4)

taking care that all ambiguities between development, operations and

support costs were clearly defined and resolved.
The reliability, maintainability and availability specialist provided:

S = K2E/(1-A) (5)

where:
K2 is a constant,
A is the probability that the system is ready when called upon.

Finally, the operations analyst provided this definition of effectiveness:

E = MA(D/Dmax)1/2 (6)

where:
M is the mission success probability, given the equipment is available
D is the amount spent on development
Dmax is the budget requested by the developers

All these inputs from the specialists were reviewed for reasonableness by
the systems analyst and integrated into the “model” shown in Table 1-1.

Table 1-1. The Mathematical Model

1) T = PE/C
2) D = k1C where k1 = 0.3
3) S = x + 0.5D
4) C=D+S
5) S = k2E/(1-A)
6) E = MA(D/Dmax)1/2

where:
T = Top-level systems criterion
P = Political index of acceptability
C = Life cycle cost of a system
D = Development costs of system
S = Support and operations cost
E = Effectiveness of system
M = Mission success probability (working)
A = Availability of system
6 Constraint theory

“There appear to be no internal inconsistencies,” reported the analyst to

the manager. “Indeed, this model is enormously simpler than any I have ever
dealt with for years.”
The manager, who claimed many years of systems engineering
experience, observed, “I see the model is imbedded in an eight-dimensional
space and is constrained by six equations. Therefore, there should be two
“degrees of freedom.” Since I’m most concerned with the total system
optimization criterion, please compute plots of T = f7(S,P) for me.”
“Sorry, said the analyst, that is not an allowable computation on this
model. Although the total model seems to have two degrees of freedom, that
freedom does not exist uniformly throughout all parts of the model. In
particular, the submodel composed of relations 2, 3 and 4 is concerned only
with the variables C, D and S. Therefore, in the three-dimensional subspace
of CDS, we have three equations and three unknowns; thus there are no
degrees of freedom, and these variables are constrained to a point or a set of
points. Since it is such a constrained variable, S obviously cannot act as an
independent variable for the computational request, T = f7(S,P).”
The manager did not like the word, “obviously”. “There must be
something wrong with the model,” he asserted. The specialists got huffy.
The analyst assured the manager, “There are no inconsistencies or
internal contradictions in this model. Once we’ve agreed to accept some
inaccuracy due to simplification, all the equations are ‘correct’ and perfectly
valid mathematically. Each of the relations referring to CDS space was
contributed by a separate specialist. Because of this interaction between
three disciplines, C, D and S are determined and can no longer be considered
as variables. Any computational request which includes C, D or S as an
independent variable must be considered unallowable. Your request was not
mathematically well-posed.”
“All right,” conceded the manager, “then let me see T = f8(M,A).”
“Sorry again,” said the analyst, “that request is also not allowable.
Consider the relations 5 and 6 which are concerned with variables S, A, E, D
and M. As we’ve just discussed, S and D are held to constant values because
of the internal constraint applied from another part of the model. By
applying M and A as independent variables, we are applying external
constraint to the SAEDM space. Thus, we have only one variable, E, which
is neither internally or externally constrained, and which must conform to
the two equations, 5 and 6. Having two equations with only one unknown is
a clear case of local overconstraint. This computational request is also not
well posed.”
The manager sighed, “Then what computations are allowable of the
form, T = f(p,q)?”
1. Motivations 7

The analyst replied, “only these three: T = f9(E,P), T = f10(M,P) and T =

f11(A,P). All other computations of the form T = f(p,q) either overconstrain
or underconstrain some part of the model.”
These computations were plotted and given to the manager, along with
the constant values for C, D and S. After studying these results, the manager
said, “OK, next I’d like to see some tradeoff curves. Please show me the
tradeoff between M and P, everything else being equal.”
“By ‘everything else being equal’ do you mean: ‘hold all other variables
at constant values?’” asked the analyst.
“Yes, I suppose so,” responded the manager, fearing the worst.
“In that case the desired tradeoff is not allowable,” said the analyst.
“Once we agree that C can no longer be considered as a variable, E is the
only variable that connects the submodel containing M with the submodel
containing P. If E is held to a constant value as you want, then the two
submodels are essentially disconnected and the M vs. P tradeoff cannot be
computed.”
“What tradeoffs would be allowable?”
“If we hold T only at a constant value, then M vs. P, A vs. P and E vs. P
are allowable computations.”
After all the allowable computations were performed and examined, the
manager asked, “How were you able to come to your conclusions on the
various computational allowabilities so rapidly? Do you have a method that
provides you special insight?”
The analyst showed the manager Figure 1-1. “Fundamentally, I attempt
to get a right-brain view of the topology of the model. Look at relation 5 for
example. In Figure 1-1a, I represent the relation by a square and show its
three relevant variables S,E,A, – represented by circles – connected to it.
Note that there are no arrowheads on these connections, since we don’t know
in advance what the computational path will be. In Figure 1-1b, the
arrowheads indicate that S can be computed if A and E are known inputs.
Similarly, Figures 1-1c and 1-1d show the computational flow directions for
the computations of A and E respectively.”
“Now, let’s expand our perspective to include relation 6 and its relevant
variables, A,E,M and D. Looking at Figure 1-2, we see that variables A and
E are common to both relations 5 and 6; they do not have to be repeated.
Note that the topology has developed a little circuit.”
“Continuing in this manner, we can include the entire model, shown in
Figure 1-3. This structure is called a bipartite graph which provides a right-
brained view of the topological structure of the model of Table 1-1. It’s
really a metamodel since it does not contain the actual equations of the
original model – just the structural information necessary to determine
internal consistency and computational allowability. As in all bipartite
8 Constraint theory

graphs, there are two distinct types of junctions: squares represent the
relations, and circles represent the variables. The arcs, or “graph edges”
connect each relation to the variables that are relevant to it. The “degree,” d,
of each junction is defined as the number of edges which intersect it.”

Figure 1-1. A right-brain view of relation 5, introducing the concept of representing relations
by squares and variables by circles, as well as demonstrating that computation can flow
through the relations in many directions.

“This bipartite graph can be considered as a network for information

flow. The squares are essentially multidirectional function generators – or
algorithmic processors – such that any output can be generated if all the
other edges provide input. The circles are essentially scalar measurements of
the value of the variable that they represent.”
“The above use of the bipartite graph for the representation of a math
model can be easily extended to the representation of a computational
request. In the general format of a computational request, one specifies a
dependent variable (the output) and a set of independent variables and
variables held constant (the input). These input variables essentially have
constraint applied to them – in addition to the constraint applied to them by
1. Motivations 9

their relevant relations – and thus additional squares are appended to the
bipartite graph.

Figure 1-2. The perspective is expanded to include both relations 5 and 6, which share
variables A and E.

Figure 1-3. The Bipartite Graph: A Metamodel displaying consistency and computability.

For example, assume that it is desired to have variables M and P be

independent variables and variable A be held constant. As is shown in Figure
1-4, the squares identified with “I” are appended to M and P, while the
10 Constraint theory

square identified with “C” is appended to variable A. In summary, the

squares representing relations of the model imply intrinsic constraint, while
the squares representing inputs to a computational request apply extrinsic
constraint. To emphasize this difference, the intrinsic constraint squares have
a single border while the extrinsic constraint squares have a double border.”

Figure 1-4. Symbology and rules of propagation for bipartite model graph.

“Before a computational request is made, the edges of a bipartite graph

model have no directionality. Once the request is made, the input variables
1. Motivations 11

now apply constraint to their neighbors and the edges take on a directionality
which is determined by the request. Essentially, the computation or
constraint “flows” across the graph.”
“For treelike graphs, the rules to map the sequential propagation of
information – or computation, or constraint – are simple in the extreme: for
d edges intersecting a square, information will propagate if there are (d-1)
inputs and one output; for d edges intersecting a circle, information will
propagate if there is one input and (d-1) outputs.” (See Figures 1-1 and 1-4.)
Note that a square with but a single arc intersecting it will automatically
transmit constraint to the circle it is connected with, since in this case, d=1
and (d-1)=0, thereby requiring no inputs to generate its output.
“For graphs which contain circuits, mapping propagation is a little more
complex. Once we have gone as far as we can with the sequential rules
above, we may require the rule of simultaneous propagation in the vicinity of
a circuit: if a connected subgraph exists which contains an equal number of
unpropagated squares and circles, then all its variables may be computed as
if they were within a set of simultaneous equations.”
“Now I can show you how easy it was to determine the computability of
your computational requests. Looking at Figure 1-5, we can easily see that
relations 2, 3 and 4 form a submodel with an equal number of squares and
circles. This denotes three simultaneous equations covering three unknown
variables and we should expect to be able to solve for the three variables,
converting them from unknown variables to fixed parameters. Thus, your
request, T = f7(S,P) is unallowable since it assumes S is variable rather than
a fixed parameter. For the same reason, C and D cannot be independent
variables either. This type of constraint imposed on the model is called
intrinsic because it existed even before you made any computational
request.”
“Now look at Figure 1-6, which shows constraint propagating from left to
right along variables C, D and S for the above explained reason. When you
requested the computation T = f8(M,A) you established the two independent
variables as a source of extrinsic constraint which propagates into the model
and hopefully gives us a computation of T. Let’s see what happens when we
employ the sequential propagation rules for constraint flow. Since M and A
are extrinsic sources of constraint and D is an intrinsic source, we can satisfy
the (d-1) inputs and one output rule for equation 6, thus producing a
computation for variable E. With A as an extrinsic source and S as an
intrinsic source, we can satisfy the (d-1) inputs and one output rule for
equation 5, thus producing another computation for variable E. This is a case
of local overconstraint, making the requested computation unallowable.”
12 Constraint theory

Figure 1-5. T = f7(S,P) is not allowable.

S cannot be an independent variable.

“Don’t get discouraged, allowable computations exist also. Figure 1-7

displays the computational paths to compute T = f9(E,P). Note how the
extrinsic constraint flows from E and P combine with the intrinsic flow from
C to satisfy the (d-1) inputs and one output rule to equation 1, resulting in
the computation of T. Also note that the entire model was not necessary for
this computation, as equations 5 and 6 were irrelevant to it.”
1. Motivations 13

Figure 1-6. T = f8(M,A) is not allowable

because of overconstraint on E.

“Figure 1-8 shows a case where both sequential and simultaneous

computational flow is used to satisfy the request, T = f10(M,P). As originally
constructed, the submodel comprised of the equations 5 and 6, together with
their relevant variables S,A,E,D and M had two equations and four variables
– ‘two degrees of freedom’ as you put it. Now, with the application of
intrinsic constraint from S and D and extrinsic constraint from M, the extra
degrees of freedom collapse to zero and we are left with two equations and
two variables for this submodel. We can expect to solve these two
simultaneous equations in two unknowns to obtain both A and E. Now
applying the intrinsic constraint flow from C, the simultaneous constraint
flow from E and the extrinsic constraint flow from P to equation 1, we can
compute T. Thus, this request is allowable. In this case, the entire model was
involved with the computation.”
“Figure 1-9 shows the computational paths for the allowable request T =
f11(A,P). The intrinsic input from S with the extrinsic input from A permit
equation 5 to compute E which, using the 1 in, (d-1) out rule, propagates to
both equations 6 and 1. This input from E plus the intrinsic input from C and
the extrinsic input from P permits 1 to compute T as requested. By the way,
this same bipartite graph shows that M = f12(A) is also allowable.”
14 Constraint theory

Figure 1-7. T = f9(E,P) is allowable.

The constraint flow rules work OK.

“Now the term, ‘tradeoffs, with everything else held equal’ is fraught
with ambiguity and most often used with insufficient rigor. Figure 1-10
displays how holding E at some constant value effectively decouples M and
P into different, non-communicating subgraphs, rendering this tradeoff
request unallowable.
If any combination of C, D and S were to be held constant, two types of
problem emerge. First of all, the value to which they were held constant
might not agree with the values computed from the 2, 3 and 4 simultaneous
equations. Even if they did agree, then propagating constraint across the
graph would yield an underconstraint at equation 1 – there would be an
insufficient number of inputs to provide the desired computation of P. The
same underconstraint situation occurs if the variable held constant is A. In
fact, the only variable that can be held constant in order to provide the M vs
P tradeoff is T.”
1. Motivations 15

Figure 1-8. T = f10(M,P) is allowable.

A new BNS is formed; then the flow is OK.

Figure 1-9. T = f11(A,P) is allowable.

M = f12(A) is also allowable.
16 Constraint theory

Figure 1-10. M = f13(P) “Tradeoff” is not allowable

holding E constant decouples M and P is allowable.

1.3 THE MANAGER AND ANALYST CONTINUE

THEIR DIALOGUE

The manager absorbed all these inputs soberly and after reviewing the
results of his requested computations as well as others which were allowable
on the original model, he complained, “I certainly can’t argue with your
mathematical rigor. But I’m still disappointed that I didn’t get more insight
out of this model for my preliminary design phase – I expected more,
somehow. It seems that whether a computation is allowable or not is like the
flip of a coin.”
“It’s worse than a coin flip; much worse. You have just observed a very
common and generally unappreciated feature of most math models,” the
analyst responded. “Indeed, the vast majority of all possible computational
requests on almost all models are unallowable. Some of the author’s
graduate students performed an exhaustive analysis of the likelihood of
computational allowability on 6- and 8-dimensional connected models of a
variety of topologies. The results are presented in detail in Appendix A, but
the general allowability likelihoods were surprisingly low. Of the 150
1. Motivations 17

computational requests made on the 6-dimensional model, less than 15%

were allowable. Of the 1078 computational requests made on the 8-
dimensional model, less than 6% were allowable. As the dimensionality of
the model increases to dozens or hundreds of variables to address modern
complex systems, then we can expect the allowability likelihood to diminish
even further. Needless to say, this is far worse than a coin flip.”
“But don’t lose hope, you can still ‘negotiate’ a more useful model with
your team of specialists – after all, they have an even more limited systems
view than you about the structure of the model and know even less about
your intended use of it.”
“For example, let’s examine the three relations, 2, 3 and 4, that caused
the inadvertent source of intrinsic constraint. Equation 4 is merely a
definition between three types of cost – this certainly seems OK. Equation 3
is the result of experience of how support costs increase with new
technology developments – this is OK based on the experience of many past
systems and assumes that there will be no investment in the development
phase to reduce supportability costs. But now look at equation 2; it is not a
representation of a definition or an experienced relationship, it is a policy
statement by a person who is attempting to limit development costs so he can
spend more on operations and support. If, instead of demanding that K1 =
0.3, he permitted K1 to merely be another variable in the model, the model
dimensionality would increase to nine, and more importantly, the intrinsic
source of constraint due to equations 2, 3 and 4 would be relieved (see
Figure 1-11). By this reasonable negotiation, we increase the candidates for
independent variables to include C, D, S and also K1. In fact, the operational
chief who furnished equation 2 in the first place can now run studies to
determine what value of K1 will maximize the systems level criterion, T,
rather than arbitrarily fixing it at 0.3.”
“Very interesting,” admitted the manager, “I can see a constructive
integration between left- and right-brain views. It would be extremely
difficult to initiate negotiations of this type without the visibility provided by
your bipartite graphs.
I’m surprised I haven’t seen this methodology before. But most real
world problems are vastly more complex than this example, are they not?
Wouldn’t analysts be driven crazy if they tried to work with snake charts that
were many square meters in area?”
“You’re absolutely correct,” agreed the analyst. “Meaningful models
quickly get large and even rigorous graphs become like a bundle of snakes,
as you put it, and not really amenable to the analysis by inspection shown in
the example. Figure 1-12 displays the flow graph for a specific
computational request on a model that’s about S times larger than the
example. The author actually was able to do consistency and allowability
18 Constraint theory

analysis on this size model without using computer aids. This was possible
because the rules, and later, the theorems were developed using
comprehensible models of low dimension, and then extending them
rigorously to higher dimensionality. For example, the model depicted in
Figure 1-13 – again about another S times larger than Figure 1-12 – would
undoubtedly boggle the mind of even the most focused analyst unless
computer aids were available. In order to communicate with the computer, a
new construct – called the ‘constraint matrix’, which has all the information
inherent in the bipartite graph – will be employed.”

Figure 1-11. Changing K1 from a constant to a variable permits C, D, S, and K1 to act as

independent variables in computations.
1. Motivations 19

Figure 1-12. A sequential computational flow in a model about S times larger than the
example.

“Is there a name to this process? How difficult would it be to become

proficient in this technique? What would the mathematical prerequisites be?”
asked the manager.
“The name of the process is ‘Constraint Theory’. It is based on the
author’s PhD dissertation [2] and subsequent published papers [3]. The only
mathematical prerequisites are the simplest 5% of set theory and graph
theory. Just read this book; it is written for practical engineers, not
professors or journal editors.”
20 Constraint theory

Figure 1-13. A model about another S times larger certainly invites computer assistance.

1.4 PRELIMINARY CONCLUSIONS

Today’s dynamic world of systems development is virtually in an

explosion of complexity and multidimensionality. Computer science
represents our best hope of controlling the complex multidimensionality;
however a major barrier to its trustworthy use is the “well posed problem.”
Constraint Theory addresses the two fundamental issues of the well-
posed problem:

(a) Is a mathematical model internally consistent?

(b) Are computational requests made of the model allowable?

Even with simple 6- or 8-dimensional models, the vast majority of

computational requests are not allowable from a well-posed standpoint.
Constraint Theory also provides decision-making managers greater
visibility into the assumptions and structure of the contributing relations
underlying models and a basis for negotiating alterations to the model in
order to attain greater benefit through desired computations.
Constraint Theory’s only mathematical prerequisites are elementary set
theory and graph theory.
1. Motivations 21

Compared to the present practice of digital simulation which tightly

integrates the mathematical model with rigid directions of algorithmic flow,
Constraint Theory distinguishes the mathematical model from computational
flow and permits multidirectional flow at the request of the analysts.
The bipartite graph and its companion, the constraint matrix, provide
insightful topological metamodels to both the model and its computations.
They provide a viewpoint to establish a rigorous mathematical extension of
the methodology to any number of dimensions.
22 Constraint theory

1.5 A LITTLE WINDOW INTO FUTURE

CHAPTERS1

For those systems engineers and managers who wish to gain

MULTIDIMENSIONAL INSIGHT
almost comparable to their perceived understanding of
NUMBER AND ARITHMETIC
and learn the activities of
NORBERT WIENER BEFORE CYBERNETICS
and
CLAUDE SHANNON BEFORE INFORMATION THEORY
and ponder the deep nature of
RELATIONS
with their many diverse characterizations including
BIPARTITE GRAPHS AND OTHER RIGHT BRAIN METAMODELS
which assist in solving the venerable
WELL-POSED PROBLEM
by detecting and correcting the dual villains
OVERCONSTRAINT AND UNDERCONSTRAINT
and ferreting out the chief culprit of inadvertent constraint
THE BASIC NODAL SQUARE
from its hiding place deep within
HOPELESSLY TANGLED CIRCUIT CLUSTERS
by employing easy graph theoretic measurements of
CIRCUIT RANK
and
CONSTRAINT POTENTIAL
which locate these kernels of constraint within medium-sized models
TRILLIONS OF TIMES
more rapidly than would a random search or the use of Hall’s 1914
theorem,
THEN DO AS THE ANALYST SUGGESTS:
READ THE BOOK!

Inspired by Edgar Palmer’s [4] delightful title page for his book, Graphical
1

Evolution, John Wiley and Sons, New York, 1985

1. Motivations 23

1.6 PROBLEMS FOR THE CURIOUS READER

1. Provide three examples from your knowledge or experience of small

models being developed and then integrated into total system
models.

2. Provide four examples of the utility of computers in the design,

production or operation of complex systems, where humans would
find it difficult or impossible to handle.

3. Provide five examples of problems associated with computers

applied to complex systems.

4. Regarding the mathematical model depicted in Figure 1-3, which of

the following computational requests are allowable and which are
not allowable?

For the allowable requests, draw the directed bipartite graph which
depicts the computational flow. For the unallowable requests,
discuss the reason(s) for the unallowability.

Computational Requests:
M=f(A), M=F(E), E=f(A,M), P=f(T,E), P=f(T,A)

5. For those computational requests which you deemed to be allowable,

switch the dependent variable with one of the independent variables
and check again for allowability. Does this suggest a possible
generalization?
Chapter 2 THE FOUR-FOLD WAY

How to Perceive Complex Mathematical Models and Well-Posed

Problems

2.1 PROLOGUE: THE MANAGER AND ANALYST

DISCUSS THE ORIGINS OF
MULTIDIMENSIONAL MODELS AND WELL-
POSEDNESS

“Since complexity has grown so enormously in modern times,” the

manager commented, “I presume that the motivations to develop techniques
to manage it are relatively recent.”
“On the contrary,” replied the analyst, “many of the concepts and
examples of problem recognition are quite old – ancient even.”
Consider the old Indian story of the blind men trying to “understand” an
elephant. Depending on what is touched – the leg, ear, tail, trunk, or tusk –
the unknown object takes on the attributes of a tree, a leaf, a rope, a snake or
a spear. Thus, touching an aspect of a complex object is far removed from
understanding the total integrated concept of “elephant.”
A more recent story – but still almost 2000 years old – comes from the
Talmud [5]. According to a commentary on the book of Genesis, on the day
that the Lord created Man, He took truth and hurled it to the ground,
smashing it into thousands of jagged pieces. From then on, truth was
dispersed, splintered into fragments like a jigsaw puzzle. While a person
might find a piece, it held little meaning until he joined with others who had
painstakingly gained different pieces of the puzzle. Only then, slowly and

© Springer International Publishing AG 2017

G.J. Friedman, P. Phan, Constraint Theory, IFSR International Series on Systems
Science and Engineering, DOI 10.1007/978-3-319-54792-3_2
26 Constraint theory

deliberately, could they try to fit their pieces of Truth together – to make
some sense of things.
Mankind’s yearning to understand the world over the eons has been aided
by the development of mathematical models. Groups of researchers,
sometimes spanning centuries contribute their little fragments of data or
understanding and eventually a general theory emerges. In many cases, the
consequences of the new theory are unexpected by the original contributors,
but such is the trust given to mathematics, the unexpected, nonintuitive
results are accepted given they are mathematically sound. Examples:
• In the 16th century, Tycho Brahe organized and extended the
astronomical observations of Copernicus and others into the world’s
finest set of data on stellar and planetary objects. Johann Kepler
took this data and formulated his famous three laws of planetary
motion. Despite his disappointment that planetary orbits were
elliptical – rather than the circles the Greeks maintained were
necessary for “celestial perfection” – he convinced himself and the
scientific world that the ellipse was the correct mathematical form
for all the orbits in Tycho’s data base.
• Decades later, Isaac Newton, with his greater mathematical
understanding, was able to generalize Kepler’s laws into his law of
universal gravitation – a gigantic intellectual feat which unified the
laws of the heavens and earth.
• Centuries later, Albert Einstein provided a refinement of Newton’s
theory of universal gravitation with his general theory of relativity.
Alexander Friedmann solved Einstein’s equations and concluded
that the universe began in a monstrous big bang. This was so against
Einstein’s instincts that he added a cosmological constant to his
equations of relativity to remove the possibility of an expanding
universe or the big bang. However, the rationality of mathematics,
as well as new data by Hubble and others have established
Friedmann correct and Einstein has referred to the cosmological
constant as his greatest blunder.
So in Man’s quest to understand, mathematical modeling has taken an
increasingly central role in building theories, and indeed in the scientific
method itself. The jagged shards of data, incomplete observations and
subdimensional theories are pieced together rationally – often resulting in
unexpected conclusions and a deeper view of the world. With the advent of
modern computer technology, this central importance promises to increase
far more.
“You certainly won’t get arguments from most practitioners of science
and technology about the importance of computers,” remarked the manager,
2. The Four-Fold Way 27

attempting to be agreeable. “What you have said would be obvious to most

observers.”
“What is not obvious is that there are many barriers to the future efficient
use of computers in the modeling of complex system,” rebutted the analyst.
“I knew you’d say that,” said the manager, remembering the example of
Chapter 1. “What are these barriers?”
“First of all,” began the analyst, “with all the increased capability and
flexibility that the digital computer offers over the analog, there comes a
subtle but pervasive disadvantage: the model and the computational requests
placed on it are inextricably intertwined. In almost all cases, the model is
programmed to execute a specific computational flow, and when asked to
alter the computation or switch input and output variables on the same
model, the programmers tend to tell the managers, “can’t be done” or “too
much trouble, or “can’t you make do with all that I’ve given you?”
“Amen,” agreed the manager, “I’ve been told that many a time. The
programmers love to overwhelm you with data to show off their powerful
computation. Their love of being responsive to your deep needs to
understand what the model is teaching us is unfortunately much less.”
Second, until early this century, the general concept of a relation has been
quite fuzzy and philosophical. Then in 1913, Norbert Wiener [6], before he
became the father of cybernetics, suggested that the definition of a relation
be imbedded within set theory – one of the foundations of all mathematics.
This served to add needed clarity and rigor to the concept of “relation.”
Third, there was a general expectation that once a model was developed,
there were no limitations on what computations could be asked of it. Which
questions are “well posed” and which are not? In 1942, Claude Shannon [7],
before he became the father of information theory, studied these issues on
the recently developed mechanical differential analyzer – the most powerful
computer of its time, analog or digital. He discovered that some of the
variables desired to be dependant, or output variables – based on the rotation
of a shaft assigned to that variable – were “free running”, providing no
useful results. In other cases the entire network of rotating shafts, gear trains
and integrators would just “lock up” – again providing no useful results.
These instances of “free running” and “lockup” are directly related to the
concepts of under constraint and over constraint, which we will discuss later.
Fourth, as was mentioned in Chapter 1, there is a vast dimensionality gap
between the cognitive capability of man and machine. Our challenge is to
make the best partnerships between these cognitive entities. As George
Gamow [8] related in his charming book, “One, Two, Three, Infinity,” it was
possible to survive with very limited numerical perceptions during our
primitive beginnings, but the advent of mathematics, starting with
28 Constraint theory

arithmetic, enormously enriched our lives and ability to understand and

control the world.
“I can see the issues are not as new as I thought,” admitted the manager.
“Speaking of all the fathers, you mentioned Gamow – wasn’t he also a
student of Alexander Friedmann, the father of the big bang theory as well as
the father of George Friedman, this book’s author?”
“Almost correct! You continue to amaze me, I should develop more
respect for you,” beamed the analyst. “Yes, George Friedman’s father was
Alexander Friedmann, but he was Friedmann the tailor, not Friedmann the
cosmologist. But let’s proceed to some substance.”
“OK,” challenged the manager. “I’m ready to enter the mathematical
world you tell me that is necessary to bring order to this confusion and
ambiguity. Let’s see if the work will prove to be a worthwhile expenditure of
intellectual energy.”
“Fair enough,” agreed the analyst. “In the remainder of this chapter, I
want to introduce you to the very simplest foundations of set theory and
graph theory, which will define for us with rigor and clarity the formerly
vague concepts of relation, well-posed, consistent, allowable computation,
overconstraint and underconstraint. I believe it will be worth your effort.”
We will begin our exploration of the foundations of constraint theory by
presenting four interrelated views of the mathematical model: set theoretic,
families of submodels, bipartite graph, and constraint matrix. The first and
second are complete and contain all the model’s detail. The third and fourth
are metamodels and contain only those abstractions which illuminate the
model’s structure as it relates to consistency and computability.

2.2 THE FIRST VIEW: SET THEORETIC

Definition 1: A set is a collection of elements. A subset is a portion of

this collection. The number of elements may be finite, such as the planets of
the solar system, or infinite, such as the points on a line. A set with no
elements at all is the null set. (Figure 2-1)
Definition 2: A variable is an abstraction of one of the model’s
characteristics which the analyst considers essential. Associated with each
variable is an allowable set of values. (Figure 2-2)
The set of variables which define the model can have enormous
flexibility. The variables can be continuous and quantitative, such as force,
length, or temperature; they can be discrete, such as the variables in Boolean
Algebra, or the solutions of Diophantine equations; they can be qualitative,
such as hot, rich, salty or sick; or combinations of these.
2. The Four-Fold Way 29

Figure 2-1. It all begins with the simple concept of sets, subsets, and the null set.

Figure 2-2. The sets of variables and their allowable values have enormous flexibility.

Definition 3: The model hyperspace is that multidimensional coordinate

system formed by all the variables as axes, each of which is orthogonal to all
the others. (Figure 2-3) This is simply a generalization of Descartes’
30 Constraint theory

unification of geometry and algebra. We will frequently refer to these axes

as Cartesian coordinates.

Figure 2-3. The Model Hyperspace formed by the orthogonal axes of the variables is a useful
abstraction, although in general it is impossible to be perceived.

Definition 4: The product set of a set of variables is the set containing all
possible combinations of the allowable values of all the variables. In the case
where all the variables are continuous over an infinite range, the product set
is merely every point within the hyperspace defined by the set of variables.
(Figure 2-4)
2. The Four-Fold Way 31

Figure 2-4. The Product Set contains all possible combinations of the allowable sets of the
variables’ values.

Definition 5: As suggested by Wiener and amplified by Bourbaki [9] and

Ashby [10], a relation between a set of variables is defined as that subset
within the product set of the variables which satisfies that relation. (Figure 2-
5) This relation can be between any number of variables and is not restricted
to the binary relations of “relation theory.”
The relations can also have enormous flexibility. They can be linear or
nonlinear, differential equation, partial differential equations, integral-
differential equations, logical equations, binary, ternary, etc., deterministic
or probabilistic, inequality relations, or any combination of these. In many
cases the relations can be represented by data or “truth tables.”
Definition 6: Since a relation reduces the size of the original product set
to a smaller, relation set, the relation can be said to constrain or apply a
constraint to the original product set. (Figure 2-6)
Now that we have embedded the concept of mathematical models within
set theory, we will need these four set theoretic operations for further
developments: (see definitions 7 and 8; Figure 2-7)
32 Constraint theory

Figure 2-5. Wiener suggested that a “relation” between variables can be defined as the subset
within the product set of these variables which satisfies it. This not only provides rigor, but
permits a tremendous variety of relation types.

Figure 2-6. Relations, as well as variables held constant, constrain the product set into a much
smaller subset.
2. The Four-Fold Way 33

Figure 2-7. Only four operations from set theory are employed for Constraint Theory.

Definitions 7: The union of sets A and B is the set of all points which are
either in set A or in set B or both. Symbolically:

x ңAыB if: x ң A or x ң B
34 Constraint theory

The intersection of sets A and B is the set of all points which are in
both A and B. Symbolically:

x ңAъB if: x ң A and x ң B

Definitions 8: The projection of set A onto dimension x is the set of

points within set A with all coordinates except x suppressed. For example, if
set A is the point (2,4) in the xy plane, then Prx (2,4) = 2 on the x axis only.
Projection is a dimension reducing operation. Symbolically:

If A = (2,4) in xy-space (point)

PrxA = (x=2) in x-space (point)

PryA = (y=4) in y-space (point)

The extension of set A into dimension y is the set of all points within set
A plus all possible values of the dimension y. For example, if set A is the
point (2,4) in the xy plane, then Exy(2,4) is the line x=2 where y varies over
all its possible values. Extension is a dimension increasing operation.
Symbolically:
ExyA = (x=2) in xy-space (line)

ExzA = (x=2)ъ(y=4) in xyz-space (line)

Definition 9: y is a relevant variable with respect to relation I in xyz

space means that there exist lines in xyz space parallel to the y axis that are
neither entirely within nor entirely outside of the relation set. Thus y has an
effect on I, or equivalently, the relation I constrains y. Symbolically:

If: Exy(PrxzAI) z $I, then y is relevant to I

Similarly, y is irrelevant with respect to relation I if:

Exy(PrxzAI) = AI (Figure 2-8)

2. The Four-Fold Way 35

2.3 THE SECOND VIEW: FAMILY OF

SUBMODELS

The set theoretic definition of relation was chosen to provide the firmest
and broadest mathematical foundation for the work to follow. Unfortunately,
it cannot also be claimed that this viewpoint is a practical way to describe
the relation. There are some occasions, such as tabulated or plotted
functions, when it is necessary to list every point within the relation subset
exhaustively. In these cases, the relation subset is merely the union of all the
listed points within the hyperspace of the model. However, in the vast
majority of mathematical models, far more efficient means are used to define
the usually infinite number of points comprising the relation subset.
These efficient means almost invariably involve the concept of
describing the total model as the intersection or union (or both) of a set of
submodels or algorithms. (Figure 2-9) This is necessary for at least two
reasons: First, and more obvious, a practical way of specifying infinite sets is
required. Second, and deeper, model builders cannot conceive of the entire
model with their limited perceptual dimensionality and thus attempt to
construct higher dimensional models by aggregating in some fashion a series
of lower dimensional submodels. The rules of aggregation employ the union,
intersection, projection and extension operators defined previously.
Frequently, a function is specified in a piecewise fashion; for example:

x = 0 when t<0
x = t2 when t>0

In cases like this, the meaning is that the contribution of these two sets to
the total model is the union of the sets.
More frequently, a collection of “simultaneous equations” attempt to
define the model; for example:

x + y + z = 13
x - y= 8

In cases like this, the meaning is that the contribution of these two sets to
the total model is the intersection of the sets.
In general, the dimensionality of the total model is far greater than any of
the contributing submodels. Thus, the contributing submodels specify only a
subset of the total model and, in order for them to be able to intersect in the
total dimensional space, they must be extended into all the unspecified
directions. For example, let the total model space be xyz and let the relation
subset for f1(x,y)=0 be A1 and the relation subset for f2(x,z)=0 be A2. Thus,
36 Constraint theory

before these two relations intersect, A1 must be extended in the missing z

direction, and A2 must be extended in the missing y direction. Defining A6 as
the total model relation, then:

Figure 2-8. Relevancy of a variable with respect to a relation can be defined in terms of the
projection and extension operations.

A6 = Exz(A1) ъ Exy (A2)

Now, once the model is constructed in the above fashion, an analyst

wishes to have a subdimensional “view” – or computational request – of this
multidimensional relation. In order for him to view the relation – as AV –
with respect to the xy plane, he must ask for a projection of A6 onto the xy
plane. (Figure 2-10). Symbolically:

AV = Prxy(A6)
2. The Four-Fold Way 37

Figure 2-9. Total model relations are generated by families of submodels which are combined
by the union, intersection, extension and projection operations.

If the analyst wishes to impose additional restrictions on his view, or

computation, prior to the projection, he may hold any number of variables at
a constant value. In these cases, the relations corresponding to these
variables held constant intersect the total model relation prior to the
application of the projection operation.
38 Constraint theory

Figure 2-10. After each generating relation is developed, it is extended into the full
hyperdimensional space of the total model, forming the relation AȈ, which in turn is projected
onto the subspace of the computational request, where it can be “viewed” by managers,
analysts and others interested in learning from the model.

2.4 THE THIRD VIEW: THE BIPARTITE GRAPH

Although there exist strong implications of topological structure in

mathematical models and their computations, neither of the two views
described above provides topological insight. In order to provide this
additional insight – as well as allow a right-brain perspective to aid the
dominantly left-brain views already presented – graph theory will be applied.
2. The Four-Fold Way 39

Definitions 10: A graph is a topological network of points, called

junctions, or vertexes, and lines connecting some of them, called arcs, or
edges. A bipartite graph is a special graph having two disjoint sets of
vertexes, {K} and {N}, such that any edge is only allowed to connect a vertex
in {K} to a vertex in {N}.
Definitions 11: A model graph, is a bipartite graph with one set of
vertexes, called nodes, corresponding to the model’s relations and the other
set of vertexes, called knots, corresponding to the model’s variables. A knot
will be connected by an edge to a node only if the corresponding variable is
relevant to the corresponding relation. As an additional visual aid, nodes will
be shown as squares and knots will be shown as circles. (Figure 2-11)
A model graph can be thought of merely as the circuit diagram of a
computer hookup of the mathematical model: the nodes are function
generators and the knots are wired connections that permit the values of the
variables to pass from one function generator to another. Thus, when the
edges indicate no direction, the bipartite graph represents a model. When the
edges indicate specific directions, then the bipartite graph represents a
computation on that model, tracking the flow of computation or constraint
across the topological structure.

2.5 THE FOURTH VIEW: THE CONSTRAINT

MATRIX

The fourth and final viewpoint of the mathematical model is introduced

primarily to provide a format amenable to computer processing. As will be
seen later, however, it also furnishes yet another mathematical perspective
from which the proof of certain theorems can most easily be made.
Definitions 12: A constraint matrix is a rectangular array of elements
that presents exactly the information inherent in a bipartite model graph, but
is a form that can be easily stored and operated upon by a computer. The
columns correspond to variables and the rows correspond to relations. An
element in the ith column and the Vth row will be filled if the variable i is
relevant to the relation V, and empty if it is not. (Figure 2.12) Compactly
stated, the rows, columns and elements of the constraint matrix are
homomorphic to the nodes, knots and edges of the bipartite graph. In order
to indicate the direction of computational flow, the elements of the constraint
matrix can take on the values: +1 or -1.
To further emphasize the essential similarity between the bipartite graph
and the constraint matrix, Figure 2-13 shows a logical evolutionary transition
between the two representations. As was stated earlier, both the bipartite
graph and constraint matrix are “metamodels” and do not contain the full
model information inherent in the set theoretic and family of submodels
40 Constraint theory

versions. Rather, they emphasize the structure and topology important to

model consistency and computational allowability.

2.6 MODEL CONSISTENCY AND

COMPUTATIONAL ALLOWABILITY

We are now prepared to present rigorous definitions in the area of the

“well posed problem.” Saying a problem is well posed means that the
mathematical model is consistent and the computation is allowable.
Definition 13: A mathematical model is consistent means that its
multidimensional relation set contains at least one point. Symbolically,

A6 z the null set

Definition 14: A computational request made on a model is allowable

means that the projection of A6 onto the view space of the computation
contains at least one point and in addition, each variable involved in the
computation must be relevant to this projection in the sense of definition 9.
Thus, if the projection onto the desired subspace that the analyst wants to
view has been nulled out to no points at all, then the computation, the
application of variables held constant or even the total model relation has
been overconstrained. On the other hand, if the projection has variables that
are not relevant, these variables take on all their possible values, and are
therefore underconstrained. (Figure 2-14)

2.7 THE MANAGER AND ANALYST CONTINUE

THEIR DIALOGUE

“You started off simply enough,” commented the manager. “What can be
easier than the definition of sets and their operations? Also, your extension
of Cartesian coordinates into hyperspace can be grasped by extrapolating
from what we know of one, two and three dimensional spaces.
As a teenager, I was inspired by Abbot’s Flatland [11] and Burger’s
Sphereland [12]. These extensions are interesting philosophically, but do
they really represent the real world and should they be the basis for applied
mathematics? I’ve heard professors argue that Descartes himself only was
thinking of our familiar three dimensional space, not even the four
dimensions for relativity theory, the eleven dimensions for string theory, and
certainly not the hundreds of dimensions we need for a modern mathematical
model.”
2. The Four-Fold Way 41

Figure 2-11. The bipartite graph is a metamodel of the full model which illuminates the
model’s structural and computational properties.
42 Constraint theory

Figure 2-12. The Constraint Matrix is the companion to the bipartite graph and displays
exactly the same information. It is also a metamodel which contains only that information
relating to the model’s structural and computational properties.
2. The Four-Fold Way 43

Figure 2-13. Evolutionary transition from the bipartite graph to the constraint matrix.

“What do we really know about ‘reality,’” asked the analyst, launching

into a minor tirade. “Irrational numbers were once thought to be irrational,
and still bear the label. Negative numbers were once thought to be
imaginary, but now play an essential role in every walk of science and
finance. Zero was originally thought to be completely unworthy of serious
consideration, but we would be crippled if we stuck to the awkward Roman
44 Constraint theory

numerals. Imaginary numbers were originally considered as only interesting

abstractions, but today form the basis of several very practical integral
transforms and are essential to almost every walk of electrical engineering.
Then there are loops (which we will consider in Chapter Four). Early in
control theory, feedback loops were considered impossible, in logic, self-
referential statements were considered illogical, and in decision theory,
intransitive preferences are still considered irrational.”
“So, in light of the above, we face this philosophical question: ‘Does
hyperdimensional space have to correspond to the space-time continuum of
our universe in order to be useful for the understanding of mathematical
models?’ I claim the answer is ‘no’ and most mathematicians use as many
dimensions as they need. Descartes’ crucial intellectual leap was to enrich
the algebraic relations with geometric concepts; the extrapolation to any
number of dimensions should be trusted as a straightforward extension.”
“OK,” agreed the manager, feeling a little over-answered. “Your use of
the projection and extension operators was less familiar to me and I never
really thought how families of submodels contributed to the total model. The
concept that a computational request is really a projection of the total model
onto a subspace was really beyond my experience. But now I can see the
value of this construct. The projection operator provides the dimensionally-
limited human an understandable perspective of an inconceivable multi-
dimensional relation.”
“Or to further extend the lingo that managers like to use,” added the
analyst, “if you choose the right subdimensional viewspace, you get the ‘best
angle’ on a complex problem – something you guys are always trying to do.”
“Figure 2-14 is a good summary of most of the previous ideas. Referring
to the ancient stories at the beginning of this chapter, the blind men each
observing the elephant from different aspects form pitifully incomplete
shards of truth, generating relations which are combined into the ‘total truth.’
However, since we are hopelessly subdimensional, we cannot perceive or
understand the total truth, A. Instead, we turn it every which way and attempt
to observe it from many different angles, some of which may help us to
understand ‘more deeply.’ Disappointingly, for most of the directions we
attempt to look at the relation, we will get no more information. That is the
agony of asking questions which are not well-posed.”
“Bottom line, to use more management jargon,” summarized the analyst,
“these four views were deemed necessary by the author to understand the
underlying foundations of models, computations and well-posedness, rather
than rely on the rather opaque, algorithmic crankturning he had been taught
in all his courses in mathematics.”
2. The Four-Fold Way 45

Figure 2-14. In order for a problem to be well-posed, the mathematical model must be
consistent and the computation must be allowable. Consistency requires that the hyper-
dimensional relation not be the null set. Allowability requires that all the variables of the
requested computation be relevant to the projection of the total relation onto the
computational subspace.
46 Constraint theory

2.8 CHAPTER SUMMARY

• Four interrelated views of a mathematical model and its

computations have been presented: set theoretic, family of
submodels, bipartite graph and constraint matrix. (Figure 2-15).
The first two are full models, containing all the detail necessary
for final construction and computation; the latter two are
metamodels, and are abstractions of the first two which
concentrate on the topological and computational features. In a
strong sense, the metamodels can be considered to provide an
overarching management perspective on consistency and
computability issues. Without this perspective, those who
attempt to build models and make computations on them will
blunder into difficulties due to either inconsistencies in the model
or unallowabililties in the computations – in short, the traditional
well posed problems. In any case, once the “well-posedness” of
the models and computations have been analyzed and managed
by the metamodels, the full models must then be employed for
the actual computations.
• The concept of “set” has been used as frequently in this chapter
as the word “system” has been used in a book on systems
engineering. This was done deliberately because – despite the
apparent simplicity of the concept – it is far more precise a
concept than “system” and its applicability is wide ranging.
• The set was used to define a relation between variables. The
concept of set was also used to identify:
o the allowable values of a variable
o the possible values of a product set
o collections of variables – which can represent
computational requests
o collections of relations – which can represent submodels
o subsets of bipartite graph vertexes called knots
o subsets of bipartite graph vertexes called nodes
o collection of edges connecting subsets of the knots and
nodes
o constraint matrix columns; homomorphic to knots and
variables
o constraint matrix rows; homomorphic to nodes and
relations
o constraint matrix elements; homomorphic to edges and
relevancies
2. The Four-Fold Way 47

Figure 2-15. The four representations of a mathematical model. The first two are full models
and the latter two are metamodels.
48 Constraint theory

o power set of the knots: all definable computational

requests
o power set of the nodes: all possible submodels

• Perhaps Friedman’s greatest contribution was the recognition

that very useful metamodels of the mathematical model’s
variables, relations and relevancies are the bipartite graph’s
knots, nodes and edges and the companion matrix’s rows,
columns and elements.
• This chapter provided the foundation for a building; its
construction and use will continue in subsequent chapters.

2.9 PROBLEMS FOR THE INTERESTED STUDENT

1. Provide a real-world example of a three-dimensional product set

where one dimension is continuous, another is discrete and a third is
defined in intervals.

2. Employing detailed algebraic equations in three-dimensional space,

show an example of y being relevant to the model and another
example of y being irrelevant.

3. For a three-dimensional model, show how ъыPr and Ex can be

used to combine a family of three algebraic equations into the total
model.

4. Draw the constraint matrix for Figures 1, 5, 6 and 7 of Chapter One.

5. Draw the constraint matrix for Figure 3 of Chapter One. Can you
suggest how the term, “Basic Nodal Square” was developed?

6. Regarding the mathematical model depicted in Figure 1-3, which of

the following computational requests are allowable and which are
not allowable?

For the allowable requests, draw the directed bipartite graph which
depicts the computational flow. For the unallowable requests,
discuss the reason(s) for the unallowability.

Computational Requests:
E=f(T,M), A=f(T,E), A=f(P,M)
Chapter 3 GENERAL RESULTS

From Protomath to Math to Metamath

3.1 LANGUAGE AND MATHEMATICS

“You’ve opened my mind to many new concepts and definitions, but I

don’t see where this is all leading,” complained the manager. “Are there
results I can use? I feel as if I’m learning the vocabulary of a new and rich
language, but I can’t make sentences.”
“Your language/mathematics analogy is very apt,” complimented the
analyst. “Before the emergence of modern language humanity had probably
millions of years of protolanguage. Words were formed to represent
abstractions from the observed world and originally served to communicate
basic ideas such as danger warnings or cooperation in game hunts. Full
language appeared when the words were organized into complete-thought
sentences employing grammar and syntax – which has a remarkable
worldwide structural similarity over all known languages. Similarly, before
the emergence of mathematics there were probably thousands of years of
protomathematics. Concepts of number and geometry were employed in
prehistoric times in practical ways for commerce and property surveys.
Full mathematics appeared with the organization of “math facts” into a
logical structure of definitions, relations and proofs – which also have a
remarkable similarity across all cultures and languages.”
“Although the two appear different superficially, mathematics is
completely imbedded within language. All the rules of math and logic of
proofs are linguistic. The applicability of language to the world is orders of
magnitude greater than math; however when math is applicable, it offers
these important advantages: precision, consistency, calculatability,
generalizability, stability across cultures and languages, and perhaps most
© Springer International Publishing AG 2017
G.J. Friedman, P. Phan, Constraint Theory, IFSR International Series on Systems
Science and Engineering, DOI 10.1007/978-3-319-54792-3_3
50 Constraint theory

importantly, provides a trustworthy vehicle to deduce conclusions from a

great diversity of inputs. True, it can be argued that math treats only a tiny
percentage of the world compared to language, but it was crucial to all the
advances of science, technology and economics which contribute to our
modern civilization.”
“Now just as math employed higher order abstractions based in language,
metamathematics employs abstractions of objects based in mathematics.
Constraint Theory is a form of metamathematics which employs the bipartite
graph and constraint matrix whose elements are the mathematical objects of
variable and relation. Thus, constraint theory is yet one more step removed
from the understanding we attain from natural language. That may be a
disadvantage if it appears the mathematical complexity is formidable – but I
argue the complexity is well below that of other branches of math. The
advantage is that metamodels of this type actually can bring us closer to
methods of plausible reasoning and further enable mathematics to intensify
its beneficial augmentation of language in all the dimensions listed above.”
“There is a fundamental mystery pondered by writers such as Devlin
[13]: ‘Why does language come so easily to virtually every human and why
is math so hard – even terrifying?’ A typical child acquires a vocabulary of
several thousand words and speaks in a respectable grammar even before
formal education begins. Protomath is also acquired quite early. But
education in math comes much later and never finds a comfortable place in
the minds of most people. Devlin argues that the primary reason is that these
people never live in what he calls the “math house” where the objects and
logic of mathematics become as familiar as the everyday objects around
which we form language. Even highly trained engineers who successfully
apply integral transforms in the design of sophisticated control systems in
their early careers are loath to trust math applied to model building or
decision making as they take on management roles in their later careers.
You, sir, are an excellent example of this math phobia in highly educated
people.”
The manager’s glazed over eyes sharpened. “I would resent that remark,
but I don’t disagree enough. I’ve been exposed too many times to
mathematicians claiming to serve me by narrowly focusing on a small part
of the problem – the only part where math is applicable – and by attempting
to impress by presenting their incomprehensible results within a forest of
incomprehensible derivations. I’ve often wondered if the word ‘analyst’ had
its roots in ‘anal retentive.’ Invariably, I had to come in to add the necessary
additional dimensions and provide management judgment – whatever that
means. Admittedly, when I did do math myself, I trusted that the proofs I
was given were correct, and tended to skip over them as I was a student.
3. General Results 51

Even the formulas and rules which I applied were not always provided with
the necessary assumptions to clarify their domains of applicability.”
“A major advantage that math has over language is its generizability into
domains which were previously incomprehensible,” the analyst pontificated.
As was mentioned in Chapter 1, we cannot really perceive numbers over
seven or so – the way that we (and many animals) can perceive 1, 2 and 3.
But we so trust the algorithms of arithmetic that we at least have a feeling or
control of understanding that which we need to know about numbers into the
millions or billions: ‘which number is greater and by how much?’ ‘is number
a slightly greater than b or is it an order of magnitude greater?’ These useful
answers can be obtained without fully perceiving large numbers. Similarly,
the objective of constraint theory is to provide the manager of large models a
trustworthy method to obtain answers to certain important properties of
models: ‘Is the model consistent?’, ‘Is the computation a=f(b,z) allowable on
this model?’ These answers are useful whether or not the manager can
actually perceive the very high dimensions involved.”
“OK, thanks for the sermon,” said the manager, feeling that the
explanations were somewhat long. “Let’s see how you can bring me into
Devlin’s “math house” and what trustworthy results you have to provide.”

3.2 MOST GENERAL TRUSTWORTHY RESULTS

Theorem 1: If a model is inconsistent, then no computational requests on

it are allowable.
Proof: By Definition 13, the relation set of an inconsistent model is the
null set. By Definition 14, for a computational request to be allowable, the
projection of the model relation set onto the view space must have at least
one point. But the projection of a null set onto any view space must also be
the null set. Thus, if the model is inconsistent, any request is unallowable.
QED
“Is the reverse also true?” asked the manager. “Depends on what you
mean by ‘reverse’,” responded the analyst. “If you mean, ‘does
noncomputabililty imply inconsistency?,’ the answer is ‘no’ – there are other
reasons for noncomputability than inconsistency; refer to Appendix A. If you
mean, ‘does consistency imply computability?’ the answer is ‘no’ again for
the same reason. But if you mean, “does computability imply consistency,’
then the answer is ‘yes’. We have to be careful as to how we employ double
negatives; not all natural languages recognize a double negative as a
positive. A mathematical implication which is two-way is called ‘iff’ – if and
only if.” See Appendix C for a more thorough discussion of this logic. But I
fear we are digressing from the main thread here.
52 Constraint theory

“I couldn’t fail to disagree with you less,” commented the manager, self-
referentially.
Theorem 2: If any submodel of a total model is inconsistent, then the
entire model is inconsistent.
Proof: The total model relation set is the intersection of the relation sets
of all its submodels. Since the intersection of any set with the null set is the
null set, if any submodel relation is the null set, then the total model will also
become the null set and thus, by Definition 13, will be inconsistent. QED.
“These are a good start it seems,” commented the manager, but aren’t
they rather intuitive? And thus perhaps not so useful?”
“Well I hope all the remaining theorems will be as intuitive, or at least as
plausible as these,” the analyst responded. “As far as useful goes, it tells us
to check the consistency of a model before we attempt computations on it. In
my experience, most of the time computations fail, the model is checked
only after much wasted effort. The utility goes further: even if a part of the
model that is not used in the computation is inconsistent, then the
computation is not allowable. Look at Figure 3-1. At first glance, it would
appear that the computational request d=f(a,c) is allowable. However,
observe that on the left side of the model, we have over three relations
constraining the two variables e and f; this is a serious case of overconstraint
– that is, the relation set for the submodel containing e and f is the null set.
Therefore, no computations, including d=f(a,c), are allowable.”

Figure 3-1. Inconsistency in one part of the model poisoning the whole model.

“OK, that seems nontrivial,” admitted the manager. “I would have treated
the d=f(a,c) computation as a work in progress and suppressed the remainder
of the model as requiring repair. I note that even if the overconstrained
submodel were reduced to just one basic nodal square, a would be
intrinsically constrained, thereby not permitting a to be an independent
variable.”
“Excellent observation,” complimented the analyst, “but you rarely know
in advance all the computational requests you wish to make on a model and
3. General Results 53

what portions of the model will be required for the request’s computational
paths. Thus, any submodel with a null set relation set can poison the entire
model. An even more severe case is shown in Figure 3-2. Here the
overconstrained submodel is in a completely separated component – there
being no possible computational paths between it and the computational
request – and the total model is still inconsistent, rendering all computational
requests unallowable.”

Figure 3-2. Inconsistency even in a disconnected component poisoning the whole model.

“Checking total model consistency seems to be a formidable task; every

pair of relations should be examined to see if they produce a null set,”
worried the manager.
“It’s nowhere near that bad,” the analyst reassured, “look at Theorems 3
and 4.”
Theorem 3: If two relations have no relevant variables in common, they
are consistent with each other.
Proof: Assume two relations: relation 1 with relevant variables a,b,c,..
and relation 2 with relevant variables r,s,t,.. . Choose any point in relation 1
– say a1,b1,c1... – and extend it into r,s,t.. space, resulting in the set defined
by: {a1,b1,c1....r,s,t...}. Similarly, choose any point in relation 2 – say
r2,s2,t2... – and extend it into a,b,c... space, resulting in the set defined by
{a,b,c..r2,s2,t2..}.
Now since the r,s,t of the relation 1 extension can take on any value, set
them equal to the r2,s2,t2 of the relation 2 extension, and set the a,b,c of the
relation 2 extension equal to the a1,b1,c1 of the relation 1 extension. Thus the
two sets of coordinates are identical and we have guaranteed that there is at
least one point in the intersection of the extensions of the two relations.
QED. (See Figure 3-3 for a simple example of this process.)
“This substantially eases the task implied by Theorem 2; only relations
which are ‘adjacent’ – that is, have relevant variables in common – can
engender inconsistency.”
54 Constraint theory

Figure 3-3. A simple example of the consistency of relations without common relevant
variables.

Definitions 15: If a pair of relations lie in subgraphs such that there

exists at least one path of edges connecting them, they lie in a connected
component. If the pair of relations lie in subgraphs with no path connecting
them, they lie in disconnected components.
Theorem 4: If two disconnected components are internally consistent,
they are consistent with each other.
Proof: Any two relations which lie in separate components cannot share
variables in common. Therefore, by Theorem 3, they are consistent. QED.
Theorem 5: No computations involving variables from disconnected
components are allowable.
Proof: For computational allowability, Definition 14 requires that the
intersection of the projections of the relations not be the null set and that the
request’s variables be relevant to it. The first requirement is satisfied but –
because there are no relevant variables across disconnected sets – the second
requirement is not met. Thus all computations across disconnected sets are
not allowable. QED.
“Now that’s certainly plausible,” said the manager. “I wouldn’t expect
that one could compute across disconnected components which are really
islands in separate universes.”
“Agreed,” agreed the analyst.
Theorem 6: The allowability of a computational request is independent
of permutations of its dependent and independent variables.
Proof: An allowable computational request will have a satisfactory
relation set in the sense of Figure 2.14. Choosing any dependent variable out
of the request’s variables by merely rotating the view space will not alter the
validity of the relation set. QED.
Theorem 7: All possible computational requests on a model with K knots
(variables) is the power set of the knots and its number is equal to 2K.
Proof: Each computational request is a subset of the set of knots and can
be uniquely identified with a binary number whose length equals the number
3. General Results 55

of knots. Therefore the number of subsets in the set of K knots equals 2K.
QED. Refer to Figure 3-4 for demonstrations and simple examples.

Figure 3-4. Three demonstrations that the power set of the set of N elements contains 2N
subsets.

“If it weren’t for Theorem 6,” commented the manager, “the number of
possible computational requests predicted by Theorem 7 would be far larger,
I presume.”
“Certainly,” agreed the analyst, “for each of the 2K subsets of dimension
d, one could choose d dependent variables, enormously increasing the
possible computational requests. Of course, one could still choose different
ways to plot the results of the computational requests, but that is really
56 Constraint theory

outside the concern of constraint theory, which is mainly concerned about

the fundamental allowability of any of the plots.”
Theorem 8: All possible submodels of a model with N nodes (relations)
is the power set of N and its number equals 2N.
Proof: This proof is identical to that of Theorem 7; merely replace K
with N.
Definition 16: A tree is that structure within a connected component of a
graph such that there is exactly one path connecting every pair of vertices.
Definition 17: A circuit cluster is that structure within a connected
component of a graph such that there are two or more independent paths
connecting every pair of vertices. Independent paths share only their initial
and terminal vertices. Adjacent circuits are circuits which share at least one
edge.
Refer to Figure 3-5 for examples of trees, circuits, adjacent circuits and
circuit clusters.
Definition 18: A universal relation is a relation which does not limit any
of its relevant variables to a given range. For example, x+y=5 and m=n3 are
universal relations, but s2+t2=3 and z>4 are not.
Theorem 9: Any set of universal relations whose bipartite graph has a
tree structure is consistent.
Proof: First, prove that any two universal relations that have only one
relevant variable in common are consistent. (Inconsistency could occur if the
common variable had incompatible constraints placed on it by the two
relations, but all the variables, by Definition 18, have unlimited ranges.)
Then append additional relations to the model in each case with only one
variable in common, forming a tree. Consistency will be maintained at each
step. QED.

3.3 CLASSES OF RELATIONS

Thus far, all the results and discussions of consistency and computability
have been on the basis of any conceivable type of general relation. As was
discussed in Chapter 2, relations can take on an extremely wide variety of
properties. In order to progress towards effective tools for the management
of multidimensional math models, it will be necessary to define three
important classes of relations.
3. Generral Res
R sultss 57

F
Figurre 3-5. Exam
E mplees of
o treees, circuits, adjjaceent circuuits and
a circuuit clust
c ters..

Deefinnitioon 19:: Reelattion

n claassees: Coonsiderr thee geenerral relaatio
on A2. Lett l be
b
a llinee thrrouggh any
y poointt in A2. Leet A2ъlъ beb tthe inteerseectiion of A2 witth l,, an nd
lett PrlA2 be thee pro ojecctioon of
o A2 onto
o o l. The
T en:
• A2 is a disscreete rellatio on if both
b h A2ъl aand PrlA2 aree poointt seets of o
m asurre zeroo.
mea
58 Constraint theory

• A2 is a continuum relation if A2ъl is a point set and PrlA2 is an

interval set of non-zero measure.
• A2 is an interval relation if both A2ъl and PrlA2 are interval sets
of non-zero measure.
Examples of these three relation classes are given in Figures 3-6 and 3-
6A.

Figure 3-6. Examples of the three relation classes.

These three relation classes will be treated in the next two chapters. We
will discuss the continuum relation class first – in Chapter 4 – because it can
be argued that it is the most important for math model building and also
because constraint theory happens to be most useful for this class. Chapter 6
will treat discrete and interval relations.
3. General Results 59

Figure 3-6A. Three relation classes can be defined – at least locally – by the number of points
in their intersections with a line and the number of points in their projections onto a line.

3.4 MANAGER AND ANALYST REVISITED

“I’m beginning to see what you mean by ‘living in a math house,’” the
manager complained mildly. “Although all the definitions and Theorem
proofs are imbedded in natural language, they appear far ‘tighter and
rigorous’ than the ambiguities of normal conversation. It certainly doesn’t
make for fast reading. Especially if one wants to understand the proofs.”
“I’m trying to make it as plausible and painless as I can,” responded the
analyst. “I encourage you to read the proofs and get in the spirit with the
mathematical rhythm of this theory. Otherwise, you won’t gain as much
confidence in the trustworthiness of extending these results into the high
dimensions that we need in order to manage modern models. So far, you’ve
just entered the foyer of the ‘math house’; I want to show you at least three
more rooms.”
60 Constraint theory

3.5 CHAPTER SUMMARY

• Language and mathematics are two of humanity’s greatest gifts and

are remarkably similar. However mathematics is more precise and is
embedded within language which is more general. We need to
become familiar with ‘the math house’ in order to extend our results
to unimaginably high dimensions.
• Model consistency is a necessary requirement for computational
allowability. If any part of a model is inconsistent, the entire model
is therefore inconsistent, but relations with no common relevant
variables are consistent. Universal relations with a tree graph
structure will always be consistent. Computations across
disconnected components are not computable.
• All possible computational requests are the power set of the set of
the variables and they number 2K; all possible submodels are the
power set of the relations and they number 2N.
• In order to provide a basis for more specific results, three classes of
relations are defined:
o discrete, dealing mainly with point sets such as Boolean
logic;
o continuum, dealing mainly with continuous curves; and
o interval, dealing mainly with densely packed sets such as
x>5.

3.6 PROBLEMS FOR THE GENERAL STUDENT

1. Construct a simple example for Theorem 1 showing that

consistency is necessary for computability.
2. In the example given in chapter 1, was the model provided
consistent? If so, why weren’t all the computational requests
allowable? If not, why were some of the requests allowable?
3. Construct a simple example for Theorem 2 showing that any
submodel inconsistency “poisons” the entire model.
4. Given a model with K=4, show all the possible computational
requests and show that they number 2K. (For the sake of
completeness, the full set and the empty set are considered valid
“subsets”.)
5. Construct a simple example demonstrating the validity of Theorem
9.
Chapter 4 REGULAR RELATIONS

Searching for the Kernels of Constraint

4.1 COGNITIVE BARRIERS TO CIRCUITS

“I must admit that the foyer of the math house was fascinating,” said the
manager, “and the rigorous structure based on the previous definitions was
quite impressive. With the theorems and their proofs, I guess I’ve progressed
from protomath to full math. However, it was a little like the ten
commandments: after an admonition on how much I should respect and
revere this central philosophy, all I got was a series of negative statements:
can’t kill, can’t commit adultery, can’t compute if inconsistent... I’m still
looking for useful rules which will permit me to manage, as this book’s title
promises.”
“This Chapter will present many rules and procedures by which you can
more effectively manage large math models,” assured the analyst. “In fact, it
will end with a Constraint Theory Toolkit which summarizes the most useful
of the theorems, rules and procedures.”
“Central to many of the rules will be the treatment of circuits and loops
within extremely tangled bipartite graphs. Many researchers appear to have a
basic phobia about circuits. Logicians dislike self-referential loops because
of the potential for paradoxes. Early in the field of control systems, feedback
was thought to be illogical. Even von Neumann classified intransitive loops
of preferences as ‘irrational.’ (See Figure 4-1) Devlin [12] noted that the
common structure of language worldwide was treelike and hypothesized that
‘our cognitive wiring favors trees.’ So I’m asking you to keep an open mind
on the concept and value of circuits and loops.”

© Springer International Publishing AG 2017

G.J. Friedman, P. Phan, Constraint Theory, IFSR International Series on Systems
Science and Engineering, DOI 10.1007/978-3-319-54792-3_4
62 Constraint theory

Figure 4-1. The concept of circuits has initially appeared to be contrary to rational thought,
but their careful management has led to many advances.

“I’ll try, but this is the briefest philosophical introduction you’ve given to
any chapter so far,” chided the manager. “I’ve attempted to enjoy your
pontifications on cognitive science, language, origins of mathematics and the
universe. Are we properly warmed up for the meat of this chapter?”
4. Regular Relations 63

“Yes, I hope so,” responded the analyst. “We have more material in this
chapter than any other. It is the heart of Constraint Theory so far in its
development. I’m anxious to get going.”

4.2 NODE, KNOT AND BASIC NODAL SQUARE

SANCTIFICATION

Definition 20: A pair of relations are locally universal if the ranges and
domains of their relevant variables are mutually compatible. In other words,
any output from one relation is an acceptable input to the other. For example
the circle x2+y2=1 is locally universal with x=0, but not with x=3.
Postulate 1: Model builders inherently wish their relations to be locally
universal.
Although an important assumption for constraint theory analysis, the
actual testing of universality can only be accomplished on the full model; the
constraint theory metamodels do not have sufficient information.
Definition 21: A set of regular relations are continuum relations which
are locally universal with all their interacting relations.
Definitions 22: The constraint potential of a graph G is defined as p(G)
and equals the excess of nodes over knots: p(G)=N-K. It is the negative of
the “degrees of freedom” notion used in Chapter 1. It will be useful to define
two circumstances where constraint potential is insightful:

pi(G)=intrinsic constraint potential; prior to any computational

request,

pr(G)=resultant constraint potential;: after the application of

independent variables, constants and computational flow from
neighboring portions of the bipartite graph.

Definition 23: The degree of a vertex, d(v), equals the number of arcs
that intersect that vertex. For the node vertex, d(n), it is the number of
variables that are relevant to the relation; for the knot vertex, d(k), it is the
number of relations which employ that variable. Refer to Figure 4-2 for
examples.
Thus, for example, the constraint potential of a node with degree d(n) is
merely 1-d(v) since it represents one node attached to d(n) knots. In general,
the average degree of all of a graph’s vertices is a strong indication of the
graph’s connectivity.
Theorem 10: For a model graph of regular relations, with a tree-like
topological structure, the computational rules are:
64 Constraint theory

for nodes: d(n)-1 inputs permit 1 output;

for knots: 1 input permits d(k)-1 outputs.

Proof: Consider the d(n)-dimensional space of any given node: if d(n)-1

of these dimensions are chosen as inputs, then this will define a line which
will intersect the relation in a point (or set of points), by definition 19, since
the relation was assumed to be regular and thus a continuum relation. QED1.
The single output of each of these computations can then propagate to all
other d(k)-1 nodes for which this variable is relevant and, by Definition 21,
will be compatible with all of them because a regular relation is locally
universal. QED2.
Thus the intuitively appealing rules which were employed so extensively
in Chapter 1 are now rigorously “sanctified” by Theorem 10. This concept
will be generalized to all topological structures later in the chapter.

Figure 4-2. The degree of a vertex is merely the number of edges which intersect it.

Definitions 24: If any node on the path has a resultant constraint

potential greater than zero, the computational request is not allowable due to
overconstraint. If any node on the path has a resultant constraint potential
less than zero, the computational request is not allowable due to
underconstraint. If the entire path has a resultant constraint potential of zero,
the computational request is perfectly constrained and is allowable. Refer to
Figure 4-3.
However, most math models – even those which could be called very
loosely connected – will have bipartite graphs with circuit structures. When
the computational flow described by the above rules reaches the vicinity of a
circuit, over- and underconstraint cannot always be determined
4. Regular Relations 65

unambiguously, and a new rule will be required. For this purpose, it is

necessary to define a special type of structure within the bipartite graph and
constraint matrix:
Definitions 25: A Nodal Square, NS, is a submodel of a math model such
that its constraint potential, p(NS)=0. A Basic Nodal Square, BNS, is a nodal
square which does not contain a smaller nodal square within it.

Figure 4-3. Computational (constraint) flow in trees requires only the simple rules: “d(n)-1 in,
1 out for nodes; and 1 in, d(k)-1 out for knots.”

Fundamentally, nodal squares and basic nodal squares have the property
that, in their local submodel, the number of variables equal the number of
relations. Examples are shown in Figure 4-4. Recall that submodels are
formed from the total model by grouping subsets of the nodes (or rows of the
constraint matrix); thus the term, “nodal” squares if all the elements of the
constraint matrix are captured in a “square.” In this context, groupings of the
66 Constraint theory

knots (or columns of the constraint matrix) have no meaning. As defined

earlier, groupings of the variables form the power set of all possible
computational requests.

Figure 4-4. A Nodal Square, NS, is a grouping of the rows of a constraint matrix, [C], such
that all the relevant elements are captured within a square. That is, the submodel has an equal
number of relations and variables. A Basic Nodal Square, BNS, is an NS which does not have
a smaller NS within it.

This brings us to the second computational theorem of this chapter:

Theorem 11: Every Basic Nodal Square (BNS) of regular relations exerts
point constraint on each of its relevant variables. That is, all its variables are
constrained to either points or sets of points.
4. Regular Relations 67

Proof: First, note the “dimension reducing” property of regular relations.

A regular relation applied to k-dimensional space will form an allowability
set of dimension k-1. If a second regular relation is applied, the intersection
of the two allowability sets will have dimension k-2. In general, if n sets are
intersected, the resulting dimension will be k-n. (Refer to Figure 4-5) Since
the BNS has by definition a constraint potential of zero, K-N=0 and thus the
intersection of all the N relations has a dimensionality of zero. QED.

Figure 4-5. Regular relations have the property that each application of a new relation reduces
the dimensionality of the total relation by one.

“I have no argument with this result,” commented the manager.

“Theorem 11 merely tells us that if we have n simultaneous equations
involving n variables, then we should be able to solve for all these variables.
In fact, I observe that the BNS can even be a 1x1 square: one equation and
one unknown, which I also expect to always be able to solve.”
“Yes, I agree that it is intuitively appealing,” responded the analyst, “but
that intuition is based on the career-long experience of mindlessly
manipulating algebraic rules. Theorem 11 “sanctifies” this algebraic rule
from the broader perspective of multidimensional relation theory. Perhaps
what is less intuitive – and therefore more valuable – is that the BNS is the
68 Constraint theory

“kernel of constraint” in multidimensional math models. Figure 4-6 shows

why the nodal square (NS) is not the kernel of constraint. In some cases the
‘shell’ between the NS and BNS is merely the resultant constraint domain
emanating from the BNS’ sources of constraint; in other (more serious)
cases, part of the NS is overconstrained with overlapping BNSs and the
remaining part is tree-like and underconstrained.”

Figure 4-6. The Nodal Square (NS) is not the kernel of constraint; only the Basic Nodal
Square (BNS) is. In NS-type a, constraint flows from the BNS to fill out the NS. In NS-type
b, multiple BNSs overconstrain some of the knots and constraint doesn’t even reach the other
knots.

“Thinking back to the simple example of Chapter 1, let’s examine the

general propagation of constraint across a math model,” suggested the
analyst, referring to Figure 4-7:
4. Regular Relations 69

Even before a computational request is made, we must determine if the

model is consistent. As we observed earlier, there may be intrinsic sources
of constraint in the form of BNSs (including 1x1 BNSs) that either point
constrain the model or, worse yet, overconstrain it if the BNSs overlap
common variables. If the BNSs overlap and contribute to overconstraint,
then the model is not consistent and we can go no further until the
overconstraint is relieved. Worse yet, the constraint may propagate into
resultant constraint domains, possibly overconstraining variables which are
in two or more of these resultant domains. Again, if this occurs, the model is
not consistent; overconstraint must be relieved before we can ask for
computational requests.
When checking for computational allowability, the procedure is quite
similar, except that, in addition to the constraint sources, we now must
superimpose constraints in the form of independent variables and variables
held at some selected constant value. As is shown in Appendix A, the
application of the rules of Theorems 10 and 11 will generally not yield an
allowable computational request. In tree structures, the Theorem 10 rules can
be applied very rapidly, but they will likely bog down in the vicinity of
circuit structures where the BNSs are hiding.
Therefore, for both the determination of model consistency and
computational allowability, the location of the BNSs become critical; they
are the “kernels of constraint” around which inconsistency and
unallowability occur.
“Well,” observed the manager, “they weren’t hard to find in the example
of Chapter 1. And for higher dimensioned models, all I need to do is merely
examine the subsets of the constraint matrix rows to see where the BNSs are
hiding.”
“There’s that word ‘merely’ again,” chided the analyst. Appendix A deals
with sizable models of thousands of variables and relations. However let’s
look at just a medium size model of say, 100 dimensions. The number of
possible submodels – or subsets of the rows of the constraint matrix, as you
put it – is the power set of the set of nodes and is equal to 2100. Even if your
computer could examine one of these subsets for a possible BNS every
nanosecond, it would still take about 1014 years, or a thousand lifetimes of the
universe to go through this power set exhaustively.”
“Attempting to find BNSs by staring at the bipartite graph would be even
worse,” guessed the manager. “Hundreds and thousands of vertices would
appear as a monstrous ‘snake chart.’”
70 Constraint theory

Figure 4-7. Even before a computational request is made, intrinsic constraint sources –
formed on the BNS kernels – may exist in consistent models. These sources flow their
constraint outwards into resultant constraint domains. If these domains overlap, overconstraint
of the variables within the overlap is likely, rendering the overall model inconsistent.

“Absolutely,” agreed the analyst. “The low dimensional bipartite graphs

are certainly useful to develop general theories, but they would be hopeless
to use as a tool for realistically complex models. Figure 4-8 is an example of
4. Regular Relations 71

how messy it can get. There is an important theorem in set theory by Hall
[14] which we modify into bipartite graph theory as:
Theorem 12: In order that a distinct output variable be associated with
each of m relations it is sufficient that, for each K=1,2,3..m, any selection of
K of the relations shall contain between them at least K relevant variables.
The computer aided method is the only one feasible and we need the
constraint matrix for communication with the computer. However, if we are
to address realistic dimensions of modern models, we must avoid the trap of
exhaustively searching for the BNS culprits through power sets with their
attendant exponential explosions.

Figure 4-8. A bipartite graph so large that gravity forces it into a spherical shape.

Our strategy will be to understand how the properties of the BNS fits
within the easily computed properties of the bipartite graph. If we are alert
72 Constraint theory

and understand what we’re doing, we can improve the computation time to
find the BNS foxes in the dense forests of the bipartite graph by factors of
trillions. But first, in Section 4.3 we must summarize some properties of the
bipartite graph, before we develop practical rules to locate the BNSs in
section 4.4.

4.3 USEFUL PROPERTIES OF BIPARTITE

GRAPHS

The most obvious properties of any graph are its connectedness, its tree-
ness and its circuit-ness. Each of these has important consistency and
computability consequences and we will treat them in this order.
In order to focus on the structure of graphs, let us repeat these three
important definitions from Chapter 3:
Definitions 15: If a pair of relations lie in subgraphs such that there
exists at least one path of edges connecting them, they lie in a connected
component. If the pair of relations lie in subgraphs with no path connecting
them, they lie in disconnected components.
Definition 16: A tree is that structure within a connected component of a
graph such that there is exactly one path connecting every pair of vertices.
Definition 17: A circuit cluster is that structure within a connected
component such that there are two or more independent paths connecting
every pair of vertices. Independent paths have no vertices in common except
their end points.
Examples of definitions 15, 16 and 17 are provided in Figure 3-5. Let us
first address an automatic algorithm to determine the connected components
of a graph.
Definition 26: A node and knot are adjacent when there is an edge
connecting them. A vertex is a separating vertex if its removal disconnects
the graph.
From the viewpoint of the constraint matrix, the existence of a relevant
element in the xth row and yth column denotes the existence of the xy edge
and thus the adjacency of the x node and y knot.
Definition 27: The connectedness algorithm, by repeatedly applying the
adjacency definition, will partition any graph into disjoint connected
components.
Definition 28: The separating vertex algorithm, by employing trial
eliminations of vertices and determining whether the remaining graph is still
connected, will locate the graph’s separating vertices.
These algorithms are exhaustive and will always work to determine the
connected components and separating vertices of any graph, including
bipartite graphs. However, it may sometimes be useful to employ more
4. Regular Relations 73

global results which are a function of only the total number of vertices and
edges and provide faster conclusions on some occasions. These results,
showing when the graph must be connected, may be connected, and cannot
be connected are presented in Figure 4-9.

Figure 4-9. Domains of connectedness for bipartite graphs as a function of the number of
edges, E, and the number of vertices, V. (Recall that V=K+N.)

Theorem 13: A connected graph which has V-E=1 is a tree. A tree is a

minimally connected component and every vertex of a tree is a separating
vertex.
Proof: Consider the simplest possible case for a tree: two vertices
connected by one edge; thus V=2 and E=1, yielding V-E=1. Now add more
fragments which are connected but do not contribute more paths between
vertex pairs. In all cases, these fragments will add one more vertex and one
74 Constraint theory

more edge; thus V-E=1 is still true no matter how many more fragments are
added. QED1.
Since a tree provides only a single path between any vertex pair,
removing any vertex will interrupt that path and disconnect the tree into
separate components. QED2.
Theorem 13 provides us with a simple and powerful tool for the analysis
of trees within graphs, no matter how highly dimensional and complex they
become. Figure 4-10 presents examples of how it can be applied.

Figure 4-10. Theorem 13 is unerring in its ability to identify trees. Initial layouts of bipartite
graphs may frequently disguise the fundamental “tree-ness” of the structure, especially when
the dimension is large.

Definition 29: The tree algorithm for graph G with V vertices and E
edges: If V-E=1, then G is a tree; if V-E>1, then G is disconnected, if V-
E<1, then G has at least one circuit.
4. Regular Relations 75

A companion algorithm which is even simpler and in many cases can

locate the very ends – or “twigs” – of the trees easily, is:

Definition 30: The tree-twig pruning algorithm: search the entire

constraint matrix for vertices of degree one – solitary elements in a row or
column – and remove the row (which is the node or relation). Repeat until no
vertices of degree one remain. The residue will be a graph completely
comprised of circuits, which may be connected by “internal trees” (without
twigs).
Finally, let us consider the most complex case of circuits within graphs:

Definition 31: The circuit rank, c(G), of a graph, G, with V vertices, E

edges and P connected components is: c(G)=E-V+P.

Definition 32: A simple circuit, Cj is a directed sequence of connected

edges that does not use any vertex or edge twice. A circuit vector for circuit
Cj is a sequence of elements, {e1,e2,e3...} where:

+1, if Cj traverses edge i in positive sense

ei = -1, if Cj traverses edge i in negative sense
0, if Cj does not traverse edge i

For the purpose of determining the ei, the edges of G are arbitrarily
assigned directions.

Definition 33: Circuit Vector Addition and Independence have exactly

the same meaning as vector addition and independence in linear algebra.
Refer to Appendix D for a brief summary of the relevant portions of
linear algebra which we will use for graph theory analysis. This relationship
between vector spaces and graph theory can be viewed as another example
of the unity and even beauty of mathematics. Its utility is demonstrated by
the following simple and powerful theorem:

Theorem 14: The number of simple, independent circuits of a graph G,

equals c(G), its circuit rank.
Proof: See Reference [2].
This theorem is truly amazing, providing us with very useful information
about the complexity of circuit clusters, requiring only a knowledge of the
number of the graph’s vertices and edges – these are merely the
semiperimeter and the number of non-trivial elements of the constraint
matrix. The theorem essentially tells us the dimensionality of the circuit
vector space. In linear algebra by comparison, there is no comparable
76 Constraint theory

method to compute the dimensionality of the vector basis of a space of many

vectors.
A dramatic example of Theorem 14’s application is shown in Figure 4-
11.
A relatively simple circuit cluster has 13 simple circuits, but only 4 of
them are independent circuits in the sense of Definition 30. This is analogous
to a circumstance where 13 vectors are defined but they can be captured
within a 4-dimensional vector space. In graph theory we can compute the 4
dimensions by a trivial computation of the observable variables V and E; but
in linear algebra, the computation is far more arduous and becomes rapidly
worse with higher dimensions. Figure 4-10a provides examples of circuit
rank, c(G), as well as constraint potential, p(G).
4. Regular Relations 77

Figure 4-10a. Both the circuit rank and the constraint potential of a bipartite graph will be
useful to locate the kernels of constraint in a complex model.
78 Constraint theory

Figure 4-11. A bipartite graph with 13 simple circuits. However, a simple computation of the
circuit rank, c(G)=E-V+1 reveals that there are only four independent circuits.

Definition 34: The circuit rank of G algorithm: compute the number of

independent circuits of G by: c(G)=E-V+P.
4. Regular Relations 79

Definition 35: A taxonomy of graph structures. The taxonomy listed

below classifies all possible graphs by defining these five properties in
sequence: a) number of paths between vertices, b) number of independent
paths, c) existence of trees, d) whether the trees are internal or external, and
e) whether the trees connect to other trees or to circuit clusters.

1. Zero paths connecting vertex pairs: disconnected components

2. One path connecting vertex pairs: isolated trees
3. Many paths connecting vertex pairs: circuit/tree structures
3.1. All paths are independent: circuit cluster
3.2 Not all paths are independent: circuit/tree networks
3.2.1 Circuit clusters w/o trees: “kissing” circuit clusters
3.2.2 Circuit clusters with trees:
3.2.2.1 Trees are external: circuit clusters with “twigs”
3.2.2.2 Trees are internal: “doily” structures
3.2.2.2.1 Trees-trees
3.2.2.2.2 Trees-circuit clusters

These definitions are summarized in Figure 4-12, along with a Venn

Diagram demonstrating the nested nature of the sequential categories.

Theorem 15: In the taxonomy of Definition 35, the following categories

are mutually exclusive and exhaustive: 1, 2, 3.1, 3.2.1, 3.2.2.1, 3.2.2.2.1 and
3.2.2.2.2.
Proof: Examination of the nested category specifications – with the aid
perhaps of Figure 4-12 – demonstrates that the specifications are mutually
exclusive and exhaustive at every level of classification. Specifically,
3.2.2.2.1 and 3.2.2.2.2 make up 3.2.2.2, which together with 3.2.2.1 make up
3.2.2, which together with 3.2.1 make up 3.2, which together with 3.1 make
up 3, which together with 1 and 2 make up all possibilities. QED.
In the next section, we apply the above properties of graphs in order to
locate the BNS kernels of constraint in a far more efficient manner than the
brute force approach mentioned above.

4.4 CORNERING THE CULPRIT KERNELS; TEN

EASY PIECES

Now that we’re somewhat familiar with some aspects of “the math house
of graphs” we can extend the previous section’s results on general graph
theory to bipartite graphs and corner the BNSs which may be lurking deep
within the tangled web of models like Figure 4-8.
80 Constraint theory

Figure 4-12. Classification chart and Venn diagram for the structures of a bipartite graph.
Categories 1, 2, 3.1, 3.2.1, 3.2.2.1, 3.2.2.2.1 and 3.2.2.2.2 are mutually exclusive and
exhaustive.

First of all, let us treat the trivial case of trees which terminate with nodes
of degree = 1. These are really 1x1 BNSs and do not support the spirit of
building multidimensional models. In every case, they represent a relation
with a single variable (because the nodal degree = 1) and this variable can be
solved for prior to its incorporation into the total multidimensional model.
4. Regular Relations 81

In other words, these “nodal twigs” are single dimensional models and serve
only to clutter the total model. They can be easily identified since they are
the rows of the constraint matrix which contain only one element. Once that
single relevant variable has been determined, its value can be “absorbed”
into all its other relevant nodes as a constant, thus eliminated from the
constraint matrix and bipartite graph.
Next, we will prove a simple theorem about integers which will help
speed many of the proofs in this section:
n
Theorem 16: Let p1, p2, p3.... pi be a set of integers. If ¦p i > n then
1
there exists at least one pi such that pi>-1.
Proof: Consider the trivial case of n=1: obviously if p1>-1, then p1>-1.
Next, consider n=2, then as can be seen in the
diagram on the right, if p1+p2>-2, then either
p1>-1, or p2>-1. Assume that the theorem is
true for n=1, and we add a (n+1)th term equal
to -1 (the most stress-full case) to both sides:
thus we obtain: p1+p2+... pn -1> -n-1, which is
n
identical to ¦p i > n . Thus we have shown
1
that the theorem is true for n=1 and n=2, and
furthermore, if it is true for n, then it is also
true for n+1. QED.
“I believe that this theorem is the least interesting one I have ever seen,”
sulked the manager.
“I would tend to agree with you,” said the analyst. “That’s why I
frequently refer to it as the ‘trivial’ theorem and it’s not counted among the
‘Ten Easy Pieces.’ However, as you will soon see, it helps prove other
theorems more compactly which will enable us to find BNSs trillions of
times faster than the method you suggested earlier (the ‘merely’ method).
“Trillions?!” mumbled the manager under his breath.
We are now prepared to relate the properties of a general BNS to its
constraint potential and topological graph structures defined in the previous
section by way of these ten compact results:

Theorem 17: If p(G) = N – K • 0, G contains at least one BNS.

Proof: If G has p(G) = 0, its constraint matrix will be a nodal square
(NS). If G has N – K > 0, its constraint matrix will be a rectangle with more
rows than columns, in which case there will always be rows (sub-graphs)
which can be removed to bring p(G) = 0, thus forming a NS. If there is no
82 Constraint theory

smaller NS within it, it is a BNS. If there is a smaller NS within it, it is still a

BNS within G. QED.

Theorem 18: A BNS cannot have a subgraph, Gs, with p(Gs)>0.

Proof: If p(Gs)>0, it must have a BNS within it, and by Definition 25
cannot be a BNS itself. QED.

Theorem 19: Every BNS must be connected; i.e.,

it cannot straddle two disjoint connected components.
Proof: Assume G is a BNS which is not connected
and has two subgraphs G1 and G2. Since p(G)=0, then
p(G1)+p(G2)=0. But then, by Theorem 16, either
p(G1)>0, or p(G2)>0. By Theorem 17, G has a BNS within it, contradicting
the assumption. QED.

Theorem 20: No BNS can be a tree.

Proof: Recalling that terminal nodes are not allowed,
the simplest tree is a node between two knots. Adding
more tree fragments with N-K<0 will either keep p(G)=-
1 or further decrease p(G). Thus p(any tree)<0 and it
can’t be a BNS. QED.

Theorem 21: No BNS can have a subgraph which is a

tree.
Proof: Assume that G is a BNS comprised of a tree
subgraph, Gt, and another subgraph, G2, joined at a
knot. By this assumption, p(G)=0, and from theorem 20,
p(Gt)<0. Since Gt and G2 share a knot,
p(G)=p(Gt)+p(G2)+1=0, resulting in p(G2)>-1. By
Theorem 18, G2 must have a BNS within in it resulting
in the conclusion that G cannot be a BNS since there is
a smaller BNS within it. The proof is similar if the
subgraphs are connected by a node. QED.

Theorem 22: No BNS can lie across circuit clusters with a separating
vertex.
Proof: Assume G is a BNS which lies across circuit
clusters G1and G2 with either a node or a knot as a
separating vertex. If the vertex is a knot, then
p(G1)+p(G2)=-1; if it’s a node, p(G1)+p(G2)=+1. In either
case, by Theorem 16, one or the other of the subgraphs
has constraint potential equal to or greater than zero.
4. Regular Relations 83

Thus, at least one BNS lies within the subgraphs and no BNS can lie across
both. QED.

Theorem 23: No BNS can lie across a tree-like network of circuit

clusters linked by trees which are attached to trees.
Proof: Assume G is a BNS which lies across
a chain of circuit clusters linked by trees. Let G1
and G3 be the circuit clusters and G2 be the tree.
Since the total constraint potential equals zero,
and the constraint potential of the tree subgraph
is negative, by Theorem 16, the constraint
potential must be equal or greater than zero in at
least one of the circuit clusters. Thus, a BNS
must lie in a sub-graph, and the total graph
cannot be a BNS. QED.

Theorem 24: No BNS can lie across a tree-like network of circuit

clusters which are linked by trees to other circuit clusters.
Proof: Since, by definition, the trees linking the circuit clusters together
are internal trees, they can have no external twigs and the constraint potential
of each tree equals +1, at most. A “meta-tree” can be formed by letting the
circuit clusters be vertices and the connecting trees become edges. By
Theorem 13, V-E=1 for a tree, thus in the metatree, the number of circuit
clusters minus the number of connecting trees equals 1. As before, let us
n
assume that the network is a BNS, so: p(G ) = 0 = ¦p i + (n  1) . This
1
n
yields ¦ pi = n + 1 > n . Thus, by Theorem 16, at least one of the pi>-1

and thus there must be a BNS within one of the circuit clusters. QED.

Theorem 25: Every BNS is the union of adjacent circuits within a circuit
cluster. (The “inside out” BNS location theorem.)
Proof: Recall the taxonomy of Definition 35 and Theorem 15 which
listed an exhaustive and mutually exclusive set of graph structures. BNSs
cannot lie across connected components, can’t be a tree or have tree
appendages, or be linked by trees, and can’t be within circuit clusters linked
by separating vertices. The only remaining structure is the union of circuits
within a single circuit cluster. QED.
Refer to Figure 4-13 for examples of BNSs in graph structures. Note that,
although every BNS must lie across adjacent circuits, not every circuit or
union of circuits necessarily contains a BNS. This is another example of
84 Constraint theory

the non-symmetry of certain theorems which are not “if and only if”
theorems. This misunderstanding has been the largest cause of confusion
among students.

Figure 4-13. Example of Theorem 25, the “inside out” theorem. All BNSs are the union of
adjacent circuits and lie completely within a circuit cluster.

Definition 36: A set of BNSs is independent if no single BNS is a linear

function of any combination of the other BNSs. (See Figure 4-14 for
examples.)

Theorem 26: Every circuit cluster (cc) with a constraint potential of

p(cc)>0 contains at least p(cc)+1 independent BNSs. (Proof provided later.)
Refer to Figure 4-14 for an example of BNSs in cc’s.
“Are you done with machine-gunning theorems at me?” complained the
manager. “I told you before that even in my most scholarly days, I tended to
skip over the proofs of theorems. Here, you haven’t avoided any proofs and
you’re even lapsing into using acronyms more frequently. That may speed
the exposition, but it also impedes the understanding somewhat.”
4. Regular Relations 85

“Sorry,” apologized the analyst, “but I’m really trying to communicate

the compactness and rigor of this little corner of mathematics. This hopefully
provides a foundation for the rules of the next section. If this were a more
typical book on mathematics, the ten theorems would have been aggregated
into two or three at the most and the rest demoted to lemmas in longer proofs
that would have been far more difficult to follow. Regarding acronyms, it is
really worth your effort to learn and master them; it’s like using language at
a higher level of abstraction. So from now on, we’ll use the shorthand ‘T12’
for ‘Theorem 12’ and ‘D9’ for ‘Definition 9’, etc.”
Table 4-1 provides an overview of the ten theorems, using this notation.
86 Constraint theory

Figure 4-14. Examples of Definition 36. Sets of independent BNSs can have no members
which are linear combinations2 of other members.

2
From the standpoint of their linear algebra representations.
4. Regular Relations 87

Table 4-1. Ten Easy Pieces Summary; Cornering the BNS within the Bipartite Graph
88 Constraint theory

4.5 CONTINUING THE PURSUIT INSIDE THE

CIRCUIT CLUSTERS (cc)

The analyst continued: “Now that we have cornered the culprit BNSs to
be inside circuit clusters, we must resort to both topological and constraint
properties of the bipartite graph to locate them precisely.”
“I presume that the constraint potential will be much more valuable
here,” the manager ventured.
“True,” agreed the analyst, “but not nearly as valuable as we would like.
For example, one would hope that if p(cc)<0, we could be assured that there
are no BNSs within that cc, and if p(cc)=0, then there are exactly p(cc)+1
BNSs within that cc. Unfortunately, we cannot reach either of these two
conclusions. Referring to Figure 4-15a, we see that a cc with p(cc)<0 can
still have BNSs and a cc with p(cc)>0 can have any number of BNSs equal
or greater than p(cc)+1. Look at the outer ring of the cc; because it is a
bipartite graph, p(outer ring)=0. We can form Y inner loops, each of which
increases p(cc) by 1, while at the same time we can form Z other inner loops
each of which decreases p(cc) by 1. Thus, Y additional BNSs are formed but
the p(cc) = Y-Z. Rearranging, we see that Y=p(cc)+Z and that the number of
BNSs=Y+1=p(cc)+Z+1. Since Z=0, we conclude: the number of
BNSs=p(cc)+1. (This is the proof of T26 promised above.)”
“No more use than that?” complained the manager.
“Well, we can squeeze out a little more: From the viewpoint of the
constraint matrix, if a cc has p(cc)>0 then it has N-K more rows than
columns and thus N!/(N-K)!K! nodal squares can be found within this
rectangle (Figure 4.15b). That’s a help, but beware; just because every NS
must contain at least one BNS does not allow us to conclude that there are at
least N!/(N-K)!K! BNSs in the cc. In Figure 4.15c, we see that a single BNS
may be common to two or more NSs, and in Figure 4.15d, we see there can
be more than one BNS in an NS.”
“So we must also resort to the topology of the cc. From T25, we know
that every BNS lies across the union of circuits and from T14, we know that
the number of independent circuits in a cc is its circuit rank, c(cc). This
allows us to establish the maximum number of BNSs in a cc as the power set
of the cc’s circuit rank:
Theorem 27: The maximum number of BNSs in a cc equals the power set
of the circuit rank of the cc: Max # of BNSs = 2c(cc)
“Thus, T26 sets the lower limit on the number of BNSs as a function of
the cc’s constraint potential and T27 sets the upper limit as a function of the
cc’s circuit rank,” the analyst concluded.
4. Regular Relations 89

Figure 4-15a. Demonstration of T26: the number of BNSs in a circuit cluster is equal or
greater than p(cc)+1.

“So what do you suggest a practical guy to do?” queried the manager.
“Constraint theory provides at least two complementary robust
procedures for locating the BNSs within the cc’s:

a) The brute force procedure which merely examines every submodel

within the cc and checks for p(submodel)=0, the definition of a
BNS.
b) The T25 procedure employing the fact that every BNS must be the
union of adjacent circuits.
90 Constraint theory

Figure 4-15 b-d. In a circuit cluster with p(cc)>0, it is easy to determine the number of nodal
squares within it, but this does not lead directly to the number of basic nodal squares.

The brute force procedure develops the power set of the N relations –
each one being a submodel – and tests for p(submodel)=0 and that there are
no submodels with p(submodel)=0 within it. There will be 2N such
submodels to test and the testing of each one is simple in the extreme. True,
we seem to be trapped again in an exponential number of trials but in general
the number of nodes in even the largest of circuit clusters will be far smaller
4. Regular Relations 91

than the number of nodes in the total model. For example, if the total model
had 100 relations, it would be very unusual to have a circuit cluster as large
as 30 relations. The computational advantage of performing brute force
testing within a cc rather than within the total model equals 2100/230 ~ 1021 –
truly enormous! Assuming that each test can be accomplished in a
nanosecond, the brute force examination of each of the cc’s submodels
would require only a second.
The T25 procedure offers greater efficiency as the number of relations
within the cc grows much larger than 30. By definition, it makes use of the
fact that every BNS is the union of adjacent circuits. The first step in this
procedure is to develop the set of independent circuits within the cc, which
by T14 is equal to the cc’s circuit rank, c(cc). Then examine the submodels
which reside on these circuits – including the twig knots which may be
attached – one at time, then two at a time with adjacent circuits, then three at
a time with adjacent circuits, etc. The disadvantage of this procedure is that
we must find the independent set of circuits and form the potential sets of
combinations of these circuits, before we test for zero constraint potential of
any of the submodels. The advantage of this procedure is that, instead of
searching through the power set of the N relations of the cc, we are searching
through the power set of the cc’s independent circuits. Assuming again that
N=30 and that c(cc)=6 (an average of 5 relations per circuit), then the
computational leverage is about 230/26, a factor of several million. Of course
this must be weighed against the additional “overhead” of forming the
independent set of circuits and its power set.
Generally speaking, it is judged the brute force procedure would be more
practical for cc’s with N<30 if it could be accomplished on the order of a
second. As N grows to 50 and beyond for the cc, it would become
increasingly attractive to employ the T25 procedure. (Refer to Figure 4.16).

4.6 LOCATING BNSs WITHIN A MODEL GRAPH

Let us now apply the definitions and theorems of the previous sections to
the challenge of finding BNSs within a model graph.

Definition 37: General procedure for locating BNSs in a model graph,

which involves the sequential application of these “sieves”: connected
components, tree structures, circuits, circuit clusters and constraint potential.

This procedure is outlined in the following steps a) – h) and illustrated

graphically, with a notional model graph, in Figures 4-16 through 4-21.
92 Constraint theory

Step a) trims down the model by eliminating all nodes of degree one. These
twig nodes represent single-variable equations which can be solved
separately by employing the full model. Without this key topological
property, the cornerstone theorems (“ten easy pieces”) of constraint theory
could not have been proven. As a subset of the “ten easy pieces”, T-21
through T-25 are specifically employed in the D-37 process. As such, no
terminal nodes are allowed before the rules of constraint theory can be
applied. Not only are the terminal nodes trimmed from the model graph but
their incident edges, which would become “dangling”, must also be cleaned
up explicitly.
Once the relevant single variables (knots) of the above terminal nodes have
been determined, analytically or numerically, their values can be
incorporated into other relevant relations as coefficients. In effect, these are
1 × 1 BNS which only serve to clutter a multi-dimensional model and can be
temporarily eliminated. The solution time for this process is linear as each
and every node will be examined once, and only once, for checking of d(n) =
1 and removal as such. It should be clarified that the terminal nodes are only
temporarily eliminated, or hidden, for the purpose of topological
simplification while we are searching for other higher-order BNS in the
subsequent steps of D-37. Afterwards, these terminal nodes will still need be
re-integrated back into the original, complete model to check for possible
over-constraint among all BNS (Phan, pp. 101-104).
4. Regular Relations 93

Step b) decomposes the total graph into connected components by

employing graph theoretic concepts of spanning tree and full-spanning forest
to improve solution time over that of D-27. In particular, to identify each
component in a model graph, the widely-accepted depth-first search (DFS)
algorithm is used to grow a spanning tree. This algorithm was originally
devised by Tarjan (1974), [27] and has been extensively referenced by Cormen
et al. (2001), Dechter (2003), Gross & Yellen (2006) and many others in the
literature. The DFS also lends itself naturally as an effective and efficient
application in steps (d) and (e) of D-37 and thus can even save more solution
time (Phan, pp. 104-114).

Figure 4-16. DFS algorithm used to decompose the BPG into two connected components.
94 Constraint theory

Step c) discards those components without any circuit as identified in Step

(b) since, per T-25, purely tree-like structures cannot contain any BNS. D-31
stipulates computation of the circuit rank, c(G), for every component and
discard those with c(G) = 0. However, the inherent products coming out of
Step (b) include only those components with c(G) > 0. Thus, repeated
calculation of c(G) is not necessary (Phan, pp. 114-115).

Figure 4-17. Retain only the lower connected component with c(G) > 0.
4. Regular Relations 95

Step d) uses D-30 to trim repeatedly each spanning tree of every external
twigs, i.e. vertex with dC(v) = 1, since no BNS can exist in tree twigs by
T21. The symbol dC(v) is meant to emphasize the degree of a vertex within
the context of its parent component coming out of Step (c), and not that of
the spanning tree which is just a subgraph of the component. Even though
this step only operates on spanning trees, it is important to keep in mind that
the relative complement of each spanning tree must always be carried and
fully accounted for, as an imperative part of every component, to be
simplified and decomposed through every sequentially-related step within
the D-37 process (Phan, pp. 115-117).

Figure 4-18. D-30 used to trim external trees from a connected component with c(G) > 0.
96 Constraint theory

Step e) can employ one of several methods to identify and remove all
internal trees from every spanning tree of each component coming out of
Step (d) since no BNS can exist within internal trees by T23 and T24. A
spanning tree will be decomposed into smaller spanning trees, corresponding
to newly-formed respective sub-components. The relative complement of
each resultant smaller spanning tree will also be automatically identified as
by-products. In an edge-centric method, a spanning tree needs to be grown
only once for a component to save solution time (Phan, pp. 117-122). Other
methods to determine internal trees include those by Tarjan, Cormen et al.
(pp.558-559) and Gross & Yellen (pp. 182-184).

Figure 4-19. All internal trees to be removed from a connected component with c(G) > 0.
4. Regular Relations 97

Step f) identifies separating vertices of every BPG coming out of Step (e)
and partitions its kissing circuits, or circuit clusters, at such points. D-28
stipulates a trial-elimination procedure with polynomial runtime of 4(V·E),
where V and E are the numbers of vertices and edges, respectively (Shirey,
1969). However, a more efficient DFS-based algorithm with linear runtime
can be used to separate kissing circuits and circuit clusters. The notion of a
separating vertex in Constraint Theory (D-26) is identical to that of a cut-
vertex, cut-point or articulation point in Graph Theory. And the structure of
a circuit cluster in Constraint Theory is the same as that of a biconnected
component in Graph Theory (Phan, pp. 123-143). For each partitioned
circuit cluster, if p(cc)>0 then the cc has at least p(cc)+1 BNS within it.

Figure 4-20. Kissing circuits and circuit clusters to be separated at articulation points.
98 Constraint theory

Step g) can follow one of several methods to search for intrinsic BNS within
each cc coming out of Step (e). For smaller cc with N ” 14, the brute-force
approach is to examine its nodal power set, which would result in an
exponential solution-time of 2N (Phan, 2011, p. 381). For larger cc, a more
efficient method for locating potential BNS is to examine only the fruitful
unions of overlapping nodes, either directly or transitively, i.e. avoiding
combinations of non-overlapping nodes. The search for BNS should also be
systematically implemented in a bottom-up approach, smaller nodal unions
before larger ones. As such, smaller nodal unions identified as BNS, or BNS
containers, can be tagged to not be re-used as components in the construction
of any larger union. Such a larger union can never be a BNS and should not
be unnecessarily constructed and examined (Phan, pp. 144-231).

Figure 4-21. Search for BNS among simple circuits and within circuit clusters.
4. Regular Relations 99

Step h) validates the potential BNSs located by step (g) in simplified circuit
clusters, which are isolated sub-graphs of the overall original model graph.
These potential BNS must still be validated, as complete sub-models, within
the context of a connected component of the model graph identified in step
(b). The validation process is implemented by re-integration of tree twigs
pruned in step (d), internal trees removed in step (e) and circuit clusters
decomposed at separating vertices in step (f) (Phan, pp. 323-325).

“Surely, for those not familiar with D37, there must be other ways of
managing high dimension models on computers”, insisted the manager.
“Yes, there are ways,” admitted the analyst. “But they require far more
attention by the people developing the model and programming it than one
would like in this world of highly computer-automated analysis that we like
to believe we’re in. A pragmatic approach would be to start with models
whose constraint potential is enormously negative – which have perhaps a
100 or so fewer equations than variables. Then, once the computational
request is made known to the model manager, the potential computational
paths – which started out under-constrained – are examined to see where
variables can be set at constant values so that the constraint potential is
brought up to zero along the entire path. This would be very labor intensive
and fraught with the danger that the programmers would make these
‘constraint enhancing’ decisions without properly consulting those with the
responsibility of model fidelity. I have a very sad story in my past – among
many others I’m certain – wherein a programmer decided to add a key
constraint to the model without informing management, and as a result my
company lost a crucial contract we had worked on for years. Typically, the
math model results were presented to management with a minimum of
visibility.”
“Getting off the war stories for a moment,” said the manager changing
the subject, “doesn’t step f) of D37 leave us hanging a bit? It says that if
p(cc)>1 then the cc has multiple BNSs. What happens then?”
“Thanks for the question,” the analyst said. “We have spent most of this
chapter on the issue of the location of BNSs but, as crucial as that is, it is still
only part of the job of math model management. It’s time to broaden our
sights to the whole problem.”

4.7 QUERIES FOR THE REGULAR STUDENT

1. Consider the following mathematical model of regular relations:

a = f1(c, d), h = f2(g, m, w), d = f3(r, g), c = f4(a), d = f5(c),
t = f6(r, s)
100 Constraint theory

Is the model connected?

Is it tree-like, circuit-like or both?
What is the circuit rank of the model?
What is the constraint potential of the model?
Are there BNS’s in the model? If so, identify it (or them).

2. Consider the following mathematical model of regular relations:

x = f1(y), z = f2(w, y), y = f3(x), x = f4(y)
What is the circuit rank of the model?
What is the model’s constraint potential?
Are there circuit cluster(s) in the model? If so, identify it (or them).
Are there BNS’s in the model? If so, identify it (or them).

3. Consider the following math model of regular relations:

m = f1(r, p), n = f2(q, r), r = f3(m, n), q = f4(p, r)
Is there a circuit cluster in this model? If so, identify it.
What is the model’s circuit rank?
What is the total number of simple circuits in the model?
What is the number of independent circuits?
What is the model’s constraint potential?

4. Provide a shorter proof than that given in the text for Theorem 21,
using only the constraint matrix.

5. If “sc” refers to a simple circuit, prove that p(sc) = 0. If “nsc” refers

to a non-simple circuit, show examples how p(nsc) > 0 and p(nsc) <
0.

6. Prove that, in a circuit cluster, the number of independent circuits is

equal to or greater than the number of BNSs.

7. Draw the bipartite graph of a circuit cluster which is a nodal square

that contains two BNSs within it.

8. Draw the bipartite graph of two overlapping nodal squares which

share a single BNS within them.

9. There is an easy way to count independent circuits much of the time:

merely locate all the “white areas” inside the bipartite graph which
are completely surrounded by edges. Why will this method not work
in general?
Chapter 5 MODEL CONSISTENCY AND
COMPUTATIONAL ALLOWABILITY

5.1 ZERO CONSTRAINT ALL ALONG THE

COMPUTATIONAL PATH

Now that we have developed a general process to locate intrinsic BNSs,

we can return to the concepts expressed in Figure 4-7 and discuss the general
issue of constraint propagation through a connected graph of regular
relations. The general rule, as depicted in Figure 5-1, is compactly stated as:
In order for a computational request on a consistent model to be
allowable, the entire computational path, from independent variables and
constants to dependent variable, must have a resultant constraint potential
of zero.
If the resultant constraint potential exceeds zero at any point, the
computation is over-constrained; if the resultant constraint potential is less
than zero at any point it is under-constrained. It is possible for the same
computational request to be both over- and under-constrained – at different
places along the computational path. In short, the resultant constraint
potential must be just right along the entire path. Thus, the designation: The
Goldilocks rule.
Referring to Figure 5-1, examine the nodes first. Recall that the local
degree of any node, d(N) is simply the number of edges that intersect that
node. The intrinsic constraint potential of that node pi(N) is by definition N
 K = 1  d(N). Now if constraint flows into this node from elsewhere in the
model, pi(N) will be increased by I(N), the number of edges which propagate
constraint into the node. The resultant constraint potential then becomes:

© Springer International Publishing AG 2017

G.J. Friedman, P. Phan, Constraint Theory, IFSR International Series on Systems
Science and Engineering, DOI 10.1007/978-3-319-54792-3_5
102 Constraint theory

pr(N) = pi(N) + I(N) = [1  d(N)] + I(N)

Let I(N) = d(N)  1; we see that this drives pr(N) to zero, and we have
just derived the “(d  1) in / 1 out” rule for nodes.

Figure 5-1. The Goldilocks rules for constraint flow through a network. In order for a
computational request to be allowable, the resultant constraint must equal zero at every vertex
and circuit along the computational paths from independent variables and variables held
constant to the dependent variable.
5. Consistency and Allowability 103

Next, examine the knots of Figure 5-1. The intrinsic constraint potential
of a knot is by definition: Pi(K) = N  K = d(K)  1. If this knot produces
O(K) outputs, then the resultant constraint potential of the knot is decreased
by O(K): pr(K) = pi(K) – O(K). Recognizing that O(K) + I(K) = d(K), we see
that setting O(K) = d(K)  1 will drive pr(K) to zero. Thus we have just
derived the “1 in / (d  1) out” rule for knots.
Finally, examine the circuits on the Figure 5-1 computational paths. As
we have seen above, if p(circuits) > 0, we will have multiple BNS and
therefore over-constraint on their common variables. On the other hand, if
p(circuits) < 0, then we will experience under-constraint. If p(circuits) = 0,
then a single BNS will provide constraint at all its variables, permitting the
flow of constraint though these circuits.
In summary, the “Goldilocks” rule stating the necessity of zero constraint
all along the computational path is a unifying concept for computational
allowability.

5.2 RECAPITULATION OF COMPUTATIONAL

FLOW

Let us attempt a recapitulation of what is involved in a computational

request at this point.
The first order of business is to determine the model’s consistency, for if
it is inconsistent then no computational request will be allowable. All
disconnected components and tree structures of universal relations are
intrinsically consistent. However, in circuit-like bipartite graphs there may
exist one or more BNSs which will exert intrinsic constraint on all of their
relevant knots. A single BNS merely restricts the number of computational
requests, but multiple BNSs often drive the model into inconsistency and
thus prevents all computational requests. These over-constraints must be
relieved before any computational requests can be entertained. Furthermore
the constraint flowing out of each BNS may intersect in a larger constraint
flow domain, and will also require relief by the model builders.
Once the model’s consistency has been established, computational
requests on it may be examined. The general format of these requests will
be: “Please compute the dependent variable X as a function of the set of
independent variables {Y}, with the set of {Z} variables held at the constant
values {Zo}.”
As part of the consistency check, the bipartite graph already has its
domains of intrinsic constraint mapped out. For each computational request,
add the extrinsic constraint sources due to the independent variables {Y} and
the variables held constant {Z}. The effect of applying extrinsic constraint to
the variables held constant {Z} will be to cause them to disappear from
104 Constraint theory

the graph completely (the {Zo} constants will be “absorbed” into their
relevant nodes as parameters in their equations rather than variables). This
action will tend to simplify the model and sometimes even disconnect it,
making computations across disconnected components impossible. See
Figure 5-2.

Figure 5-2. Applying extrinsic constraint by holding some variables constant tends to simplify
the bipartite graph. Sometimes the graph is disconnected into separate components, rendering
computational requests involving variables in different components unallowable.
5. Consistency and Allowability 105

Next, start the flow of extrinsic constraint from each of the {Y}
independent variables into the remainder of the bipartite graph. Several
circumstances may occur, as is shown in Figure 5-3.

a) In the vicinity of tree structures, the (d  1) in / 1 out rule for

nodes and the 1 in / (d  1) out rule for knots will always be
sufficient to determine over or under constraint.
b) Even in the vicinity of some circuits, these rules will still be
sufficient.
c) On many occasions, the constraint will flow into a circuit
structure and appear to stop. But closer examination shows that
the constraint flow into a node effectively increases the resultant
constraint potential of the circuits, forming them into a resultant
BNS, which by applying constraint to all its relevant knots,
permits the constraint flow to continue.
d) Sometimes, the intrinsic and resultant BNSs reside in close
proximity.

See Figure 5-4 for examples of interactions in these cases.

5.3 GENERAL PROCEDURE FOR DETERMINING

CONSISTENCY AND ALLOWABILITY IN A
MODEL OF REGULAR RELATIONS

“Recall that in Chapter 1, the existence of a single BNS was an irritant;

you were disappointed that you couldn’t accomplish all the computational
requests you desired, but it didn’t drive the model into inconsistency,” the
analyst continued. However, the existence of overlapping BNSs will cause
the model to be intrinsically inconsistent; therefore no computations at all
are allowed. That is not just irritating, it’s devastating. Multiple, overlapping
BNSs are against nature.”
“My, my; in addition to being a dealer in hyperbole,” commented the
manager, “you’re waxing philosophical, too. What exactly do you mean by
‘against nature’?”
“I mean at least two related things,” responded the analyst, “which can be
captured in the following postulate:”
Postulate 2: First, the laws of physics and other descriptions of the world
are fundamentally consistent if they are fully understood, and second, it is
the intent of model builders to represent phenomena accurately and thus if
over-constraint occurs, it is unintentional.
106 Constraint theory

Figure 5-3. Flow of constraint in the vicinity of trees and circuits.

“So, I was referring to the nature of the world as well as the nature of
model builders. Since the natural world itself is fundamentally consistent,
over-constraint is invariably contributed by the fragmented understanding of
the human model builders, either by inadvertently applying excessive
relations to the description of a phenomenon or, more frequently, adding too
5. Consistency and Allowability 107

many relations of policy, design rules, optimization criteria or desired

outcome.”

Figure 5-4. Examples of constraint flow in the vicinity of BNSs.

In order to bring a circuit cluster with p(cc) to consistency, at least p(cc)

nodes must be removed. In other words, of the N nodes in the cc, N  K of
them must be removed. From simple combination and permutation theory,
we conclude that there are N!/(NK)!K! ways to remove N  -K nodes from
the original N nodes. Unfortunately, these removals lie outside the domain of
Constraint Theory and require the group of human model builders who
contributed to the circuit cluster. Negotiations of the type discussed in
Chapter 1 will be required and the discussions could get tense. Better that
than having no working model at all!
108 Constraint theory

We are now prepared to describe the procedure of determining

consistency and computational allowability in a bipartite graph of regular
relations:

Definition 38: General procedure for regular relations.

a) Locate the BNS by employing D-37. If no BNS found, including 1 ×

1 BNS (a twig node and its relevant knot) and c(G) = 0, the model is
inherently consistent. Skip steps (b) – (f), go to step (g).
b) If overlapping BNSs are found, eliminate nodes through negotiation
so that all remaining BNSs are non-overlapping.
c) Propagate constraint emanating from the BNSs to their resultant
constraint domains, employing T10: for nodes, d(n)-1 in and 1 out,
and for knots, 1 in and d(k)-1 out.
d) If any two of these resultant constraint domains overlap, determine
whether any knots are over-constrained. If so, by negotiation,
remove sufficient nodes to relieve the over-constraint.
e) After the resultant domains have expanded as far as T10 applies,
analyze the remaining graph for resultant BNSs, using D36. If new
BNSs arise, continue with T10 and then apply D36 again. Continue
until the constraint domains no longer increase.
f) If all the overlapping BNSs are reconciled and resultant constraint
domains expand without over-constraint, then the model is
CONSISTENT.
g) For each computational request, treat all independent variables and
variables held constant as extrinsic sources of constraint which are
added to the intrinsic sources of constraint developed above.
h) Propagate computational paths from all constraint sources
throughout the model employing the T-10 rules. If any knot or node
is over-constrained, the computational request is NOT
ALLOWABLE.
i) If the computational path does not reach the dependent variable, then
examine the residue for BNSs, using D36, and continue with T10,
followed by D36 as necessary until the computational path can go no
further.
j) If the path does not reach the dependent variable, or over-constrains
a knot or node, the computational request is NOT ALLOWABLE.
k) If the path reaches the dependent variable by employing all
independent variables, without either over- or under-constraint along
the way, then the computational request is ALLOWABLE. It is
acceptable to have local under-constraint elsewhere in the model. If
5. Consistency and Allowability 109

the dependent variable can be reached without having to propagate

the constraint externally imposed on one or more independent
variables, then the computation is UNALLOWABLE.

Steps (a) – (f) address the issue of model consistency, and steps (g) – (k)
that of computational allowability. The following sections of this chapter
will examine the above general procedure in more details, step by step, with
the goal of realizing the utility of Constraint Theory. D-37 and D-38 will be
further extended, refined and improved into a set of more effective and
efficient algorithms, ready for implementation. As important as these issues
are to discuss, they do not represent exponential explosions of computational
time.

5.4 DETECTION OF OVERLAPPING BNS

Step (a) of D-38 states that: “Locate the BNS by employing D-37. If no
BNS found, including 1 × 1 BNS (a twig node and its relevant knot) and
c(G) = 0, the model is inherently consistent. Skip steps (b) – (f), go to step
(g).”

Per T-25, all BNS exist within simple circuits or across clusters of
adjacent circuits. If a model graph G with circuit rank c(G) = 0, it has no
circuits within. As such, G contains only tree structures. By T-9, any set of
universal relations whose BPG has a tree structure is consistent. As a global
criterion, and a quick check, to distinguish between trees and circuit clusters,
T-13 asserts that “a connected model graph having V – E = 1 is a tree”. This
assertion comes directly from the definition of circuit rank (Phan, pp. 31-32).

Otherwise, after all the intrinsic BNS have been identified and validated
within the context of a connected component in step (h) of D-37, they need
be compared against one another for any direct over-lapping. To check for
over-lapping, a set of definitions and vectorial operations will be developed
and illustrated herein.

Definition 39: Two BNS are said to directly overlap if they share at least
one variable in common, i.e. the intersection of their knot sets is non-null. In
other words, the common variable(s) are said to be over-constrained (or
over-specified), which causes the parent model graph to be inherently
inconsistent.

Figure 5-5 illustrates a model graph G and its constraint matrix [ CG ].

Within G, three intrinsic BNS can be identified:
110 Constraint theory

9 BNS #1 consisting of nodes 6 and 7, and their relevant knots e and f.

9 BNS #2 consisting of nodes 1 through 6, and their relevant knots a
through f.
9 BNS #3 consisting of nodes 1 through 5, and 7, and their relevant
knots a through f.

Since these intrinsic BNS overlap one another, i.e. sharing and thus over-
constraining at least one common variable, G is inherently inconsistent and
no computational requests made on G are allowable.

Figure 5-5. Inherently inconsistent model graph with multiple intrinsic BNS
overlapping one another.

Definition 40: The characteristic vector (or charvec) of a BNS is the

union, based on the bitwise inclusive OR operation, of all the charvecs of its
nodes (see Definition D-9 for charvec, and Definition D-10 for bitwise
inclusive OR). Symbolically,

| N BNS |
charvec(BNS) = *i =1
charvec(ni ) (5 – 1)
5. Consistency and Allowability 111

In the constraint matrix [ CG ] of Figure 5-5,

charvec(n6) = ( 0, 0, 0, 0, 1, 1 )
charvec(BNS1) = charvec(n6) U charvec(n7) = (0, 0, 0, 0, 1, 1 )
charvec(n7) = (0, 0, 0, 0, 1, 1 )

charvec(n1) = ( 1, 1, 0, 0, 0, 0 )

charvec(n2) = ( 1, 0, 1, 0, 1, 0 )

charvec(n3) = ( 0, 0, 1, 1, 0, 0 )
charvec(BNS2) = charvec(n1) U . . . U charvec(n6)
charvec(n4) = ( 0, 1, 0, 1, 0, 0 )
= ( 1, 1, 1, 1, 1, 1 )
charvec(n5) = ( 0, 0, 0, 1, 0, 1 )

charvec(n6) = ( 0, 0, 0, 0, 1, 1 )

charvec(n1) = ( 1, 1, 0, 0, 0, 0 )

charvec(n2) = ( 1, 0, 1, 0, 1, 0 )

charvec(n3) = ( 0, 0, 1, 1, 0, 0 )
charvec(BNS3) = charvec(n1) U . . . U charvec(n5) U charvec(n7)
charvec(n4) = ( 0, 1, 0, 1, 0, 0 )
= ( 1, 1, 1, 1, 1, 1 )
charvec(n5) = ( 0, 0, 0, 1, 0, 1 )

charvec(n7) = ( 0, 0, 0, 0, 1, 1 )

Definition 41: The BNS matrix of a bipartite graph G, denoted :G, is the
rectangular array of binary numbers whose rows correspond to the BNS and
columns correspond to the knots in G. In :G, the entry ei, j = 1 if knot kj 
KG has been specified by BNSi, and 0 otherwise.

Figure 5-6 presents the BNS matrix :G for the example BPG in Figure 5-
5. Note that each row of :G represents the characteristic vector of the
corresponding BNS.
112 Constraint theory

Knots
a b c d e f

BNS1 0 0 0 0 1 1
BNS2 1 1 1 1 1 1
BNS3 1 1 1 1 1 1

Figure 5-6. BNS matrix :G of BPG in Figure 5-5.

Definition 42: The overlapping factor between BNSi and BNSj, denoted
as Zi, j, is defined as the dot product between their characteristic vectors.
Symbolically,

Zi, j = charvec(BNSi) Ɣ charvec(BNSj) (5 – 2)

Furthermore, by definition of dot product, the overlapping factor Zi, j also

indicates the number of common variables (knots) over-constrained by the
two BNS.

Theorem 28: BNSi and BNSj are overlapping if Zi, j • 1. They are not
overlapping if Zi, j = 0.

Proof: If two BNS share any common knot in ki then the corresponding
ith elements in both of their respective characteristic vectors are equal to one.
Therefore, the dot product between their characteristic vectors must be equal
to one, or larger, if there are one, or more, common knot between the BNS.
QED

In Figure 5-6, Z1, 2 = charvec(BNS1) Ɣ charvec(BNS2) = (0, 0, 0, 0, 1, 1 )

Ɣ (1, 1, 1, 1, 1, 1 ) = 2. Thus, BNS1 and BNS2 overlap by two variables,
which confirms a visual inspection of Figure 5-5. This technique of
employing dot product provides a quick check for potential overlapping
between two BNS. Once any overlapping has been confirmed, the over-
constrained variables can be identified via an element-by-element
comparison between the BNS charvecs.

Definition 43: Algorithm to detect overlapping BNS in a model graph

and to identify variables commonly shared among them.
Input: A matrix :G with R rows and K columns representing all
BNS in a graph G.
5. Consistency and Allowability 113

Output: A subset S Ž KG consisting of overlapped knots (over-

constrained variables).
Initialize S = Ø.
For i ranging from 1 to (R – 1)
For r ranging from (i + 1) to R
If Zi, r • 1
For j ranging from 1 to | KG |
If the jth element of BNSi = that of BNSr = 1
Add the corresponding knot kj to S.
Return S (if S  Ø, the model is inconsistent).

Theorem 29: To detect overlapping BNS and identify over-constrained

variables using a matrix :G with R rows and K columns representing all
BNS in a model graph G, the required solution time will be:

4 detection = K (R2 – R/2) ~ K · R2 (5 – 3)

Proof: The matrix :G has R rows representing all BNS and K columns
representing knots (variables) in G. There are (R – 1) + (R – 2) + . . . + 2 + 1
= R2 – R/2 pairwise comparisons of BNS. For each pairwise comparison of
BNS, all K columns must be examined to look for commonly shared, and
thus over-constrained, variable(s) between the two corresponding BNS.
QED.

The above algorithm needs be applied to all BNS, including those 1 × 1

BNS born of terminal nodes initially trimmed by step (a) of D37 process
defined in Section 4.6. Figure 5-7 illustrates an example model graph back
into which a previously trimmed terminal node (nodal twig) is re-integrated
as a 1 × 1 BNS. This re-integration reveals knot c as over-constrained
between BNS1 = { 3, c } and BNS2 = { 1, c, 2, b }.
114 Constraint theory

1
Re-integrated terminal node

b BNS2 c 3
BNS1

Figure 5-7. Re-integration of previously trimmed terminal node reveals

model inconsistency

At this point in the process, the model builder needs to examine the set of
over-constrained variables output by the algorithm in Definition 43. Any
inconsistency, i.e. directly overlapping intrinsic BNS, inherent within the
model graph must be resolved before proceeding with the D-38 process. This
requirement for reconciliation has been established in steps (b) and (d) of D-
38.

5.5 RELIEF OF OVER-CONSTRAINT

Steps (b) and (d) of D-38 aim to relieve over-constraint inherent within a
model graph by eliminating nodes in overlapping BNS. In their current form,
the procedures for these steps state:

b) If overlapping BNS are found, eliminate nodes through

negotiation so that all remaining BNS become non-overlapping.

d) If any two of these resultant constraint domains overlap,

determine whether any knots are over-constrained. If so, by
negotiation, remove sufficient nodes to relieve the over-
constraint.

However, the removal of relations in their entirety from a model could be

a bit offensive to the stakeholders who have contributed the relations to the
model. This contribution represents their voice, their “piece of the pie”, into
the system design process. The exclusion of such input might not be
5. Consistency and Allowability 115

practical or feasible in some instances, especially when this is the voice of

the resource sponsor or policy maker who funds the project.

A less dramatic approach, and perhaps a more productive negotiation, to

achieve buy-in from the affected stakeholder(s), would be to adjust some
relations, rather than a total elimination of them. One possible technique
would be to parameterize some constant coefficients to variables.
Mathematically, the effect on p(G) should be identical, whether we remove a
node from, or add a knot to a model graph. Such an adjustment technique,
previously discussed in Section 1.3, could be employed as an alternative to
steps (b) and/or (d) above.

Let’s examine the example BPG of Figure 5-8, an inherently inconsistent

model with two indirectly overlapping BNS. By adding a knot to the over-
constrained node 5, the model now becomes inherently consistent. However,
no computational request can be made on this extended model yet because it
is perfectly constrained, consisting of two intrinsic 2 × 2 BNS and one
resultant 1 × 1 BNS. In order to entertain any computational request, another
knot needs be added to the BPG.

The above technique of extending adjustment to a BPG does not only

allow more flexibility in system modeling and simulation for the model
builder, but also facilitates thoughtful inquiry and cooperation among the
systems engineering and program management team to expand the trade
space in systems design.

5.6 EXPANSION OF RESULTANT CONSTRAINT

DOMAINS

After all inherent over-constraints have been resolved by the model

builder (with the aid of above method to detect directly overlapping BNS),
those “variables” identified as parts of an intrinsic BNS have essentially
become internal point-constraints per T-11. Their pre-determined values
must now be propagated throughout the graph network to form constraint
domains. This flow of constraint will fix other knots along the propagation
path, and may even result in additional BNS. If two constraint domains share
at least one common variable, they are said to be overlapping. In other
words, the commonly-shared variable(s) are over-constrained. As illustrated
in Figure 4-7, this phenomenon has been briefly discussed. As such, a
constraint domain is now formally defined.
116 Constraint theory

Definition 44: Let G be a model graph, and H be an intrinsic BNS within

G, i.e. H is a subgraph of G. The constraint domain emanated from H is the
maximum knot-set that includes KH and all other knots in KG which are also
fixed as a result of constraint propagated from KH throughout G.

1 3

(!)
Intrinsic Intrinsic
a b 5 c d
BNS #1 BNS #2

2 4

(a) Inherently inconsistent BPG with two intrinsic BNS over-constraining node 5.

1 e
3

a b 5 c d

Resultant

2 1 × 1 BNS 4

(b) Addition of knot e relieves over-constraint and renders BPG inherently consistent

1 e
3

Intrinsic Intrinsic
a b 5 c d
BNS #1 BNS #2

2 4
f

(c) Addition of knot f allows some computational requests made on BPG.

Figure 5-8. Adding knots to BPG relieves over-constraint and allows for
computational requests
5. Consistency and Allowability 117

It must be emphasized that constraints may flow out of a BNS even

further than the Duality Rules (T-10) imply. Even though the Duality Rules
only work around tree-like structures, constraint can propagate through both
tree-like and circuit-like structures. In tree-like structures, the Duality Rules
can be applied very rapidly, but they will likely bog down in the vicinity of
circuit structures where the BNS are hiding. It is quite possible that once a
constraint domain is expanded as far as the Duality Rules allow, new
resultant BNS may be formed. With all the knots reached by the Duality
Rules now fixed (no longer “variables”), the remainder matrix can be used to
search for more resultant BNS. Subsequently, the Duality Rules can be
employed again to re-start constraint propagation from the newly discovered,
and validated, resultant BNS. This expanded process, beyond the Duality
Rules, has been termed the Goldilocks Rule to manage constraint flow
through both tree-like and circuit-like structures. A constraint domain can
certainly expand even beyond the circuit cluster in which the BNS reside, as
far as its parent connected component will allow.

Figure 5-9 illustrates a model graph with several constraint domains. On

the left-hand side, constraint flows from three intrinsic BNS into neighboring
sub-graphs, which results in two additional BNS. The path of constraint
propagation runs along the bold edges. On the right-hand side, each resultant
constraint domain is shown to consist of its intrinsic BNS and the maximum
sub-graph expanded by propagating constraint from the intrinsic BNS. Note
that a constraint domain may not only contain BNS, which are simple
circuits or unions of adjacent circuits, but also tree structures. This is the
case with constraint domain #2 below. All variables within a constraint
domain have been pre-determined (fixed) even before any computational
request is made on the model.
118 Constraint theory

Figure 5-9. Model graph with intrinsic BNS and resultant constraint
domains

Also note that any interface between a sub-graph which represents a

constraint domain and the remainder of a model graph is through the knot set
of the sub-graph. In other words, a constraint domain can be considered as a
maximum complete sub-model all of whose variables have been inherently
fixed. To expand the resultant constraint domain within a connected graph,
the following algorithm refines steps (c) – (f) of D-38 with more details.

Definition 45: Procedure to expand the constraint domain Kcd from a

BNS within a connected graph G.
1) Per step (c) of D-38, propagate constraint from each and every knot of
the source BNS by employing the Duality Rules of T-10 to the farthest
extent possible within G.
5. Consistency and Allowability 119

a. For each knot discovered along the constraint flow path, add it to
Kcd.
b. For each vertex along the path, including the source BNS, remove it
from G via a vertex-deletion operation (Phan, pp. 101-102). This
may fragment the remainder graph into several disjoint connected
sub-graphs as evidenced from Figure 5-9. As a quick check for this
fragmentation, or disconnectivity, of the remainder graph, Friedman
[1967, p. 91] has proven that “a model graph with V vertices and
more than (V 2 4 ) – V + 2 edges is connected”.
c. If any over-constraint encountered during propagation, then G is
inherently inconsistent. Employ technique(s) outlined in Section 5.5
to relieve local over-constraints.
d. Re-process G through steps (a) – (h) of D-37 to identify a new set of
intrinsic BNS since the model topology may be unpredictably
altered as a result of step (c). Also repeat steps (1a) – (1c) after the
new set of intrinsic BNS has been validated.
2) Per step (e) of D-38, locate potential resultant BNS, and possibly
continue expanding the constraint domain further, by re-processing the
remainder subgraph(s) through steps (a) – (h) of D-37.
a. To facilitate computation, a vertex-deletion subgraph can be
represented by a remainder matrix (see Definition 46 below).
b. Re-apply the algorithm outlined in Definition 43 to detect
overlapping among newly located BNS and to assist the model
builder to resolve any inherent over-constraint among them.
c. Similarly to steps (1c) and (1d) above, if G is altered to relieve any
local over-constraint among the newly located BNS, it must be re-
processed through D-37.
d. Resume propagation of constraint from resultant BNS by re-iterating
steps (1) – (2c) until no more BNS can be located. As such, Kcd has
reached its maximum and cannot expand any farther.
3) Repeat steps (1) – (2) for each and every intrinsic BNS within G to
define its respective constraint domain.
4) Per step (d) of D-38, re-apply the algorithm outlined in Definition 43 to
detect overlapping among the constraint domains defined in step (3).
a. If any overlapping detected, then G is inherently inconsistent.
Employ technique(s) outlined in Section 5.5 to relieve over-
constraints.
120 Constraint theory

b. If G is altered to relieve any overlapping among the constraint

domains, it must be re-processed through steps (a) – (h) of D-37.
Also re-iterate steps (1a) – (4a) above.
5) By step (f) of D-38, upon successful completion of step (4b) without any
more overlapping among the BNS or among their resultant constraint
domains, G can be concluded as an inherently consistent model graph.

Definition 46: The remainder matrix representing a vertex-deletion

subgraph H, denoted as &H, of a model graph G is defined as the constraint
matrix [ CG ] less the rows and columns corresponding to the nodes and
knots, respectively, which have been fixed within a constraint domain and
thus removed from G.

Figure 5-10 graphically demonstrates an application of the high-level

algorithm outlined in Definition 45 above.
• After the “kissing” circuits and circuit cluster of original model
graph G have been separated at knots b and d by step (f) of D-37,
intrinsic BNS1 = { 1, a, 2, b } and BNS2 = { 7, d, 8, g } can be
identified by step (f).
• Upon removal of BNS1 and BNS2 from G, the D-37 process is re-
applied to vertex-deletion subgraph H. Again, its “kissing” circuits
are separated at knot f via step (f) of D-37. And this separation
results in BNS3 = { 3, c, 4, f } and BNS4 = { 5, e, 6, f } by step (g) of
D-37.
• Note that BNS3 and BNS4 over-constrain knot f, which can be
subsequently detected by applying the algorithm in Definition 43.
5. Consistency and Allowability 121

3 4
1 7

Intrinsic Intrinsic
a b f d g
BNS1 BNS2

2 8
6 5

c
Original model graph G with intrinsic BNS1 and BNS2.
3 Resultant 4
BNS3

Removal of BNS1 and BNS2 from G results in vertex- Resultant

6 BNS4 5
deletion subgraph H and reveals BNS3 and BNS4.
e

Figure 5-10. Removal of intrinsic BNS and repeating D-37 locate resultant
BNS

Step by step, Figure 5-11 below demonstrates another application of the

algorithm in Definition 45 by employing the example BPG of Figure 5-8:

• The flow of constraint is started from knot b of the intrinsic BNS, as the
source of internal constraint, into node n5. The vertices of this source
BNS are removed from the model graph via a vertex-deletion operation
(Phan, pp. 101-102) to simplify the graph during propagation.

• By the T-10 rule of d(n) – 1 in and 1 out for nodes, constraint continues
to flow from n5 into knot c, after which n5 is removed to further reduce
the graph.

• By the T-10 rule of 1 in and d(k) – 1 out for knots, constraint can
simultaneously flow from c into n3 and n4. However, as a matter of
practical implementation, the branches of propagation needs be executed
one of a time, in a single thread of execution. Parallel threads of
execution for multiple branches may result in various issues with
concurrency, e.g. unpredictable racing condition. And it would be
122 Constraint theory

difficult, if not impossible, to maintain (configuration control over) the

state of the remainder matrix when it may be modified by several threads
of execution. As such, for the purpose of demonstration herein, we will
just trace through one branch of constraint flow from c to n3.

• As knot c is removed to further simplify the graph, the vertex-deletion

operation stipulates that all edges incident upon c must also be removed.
And this includes the edge between c and n4 (Phan, pp. 101-102).

• Again, by the T-10 rule of d(n) – 1 in and 1 out for nodes, constraint
continues to flow from n3 into knot d, after which n3 is removed to
reduce the graph even further.

• Again, by the T-10 rule of 1 in and d(k) – 1 out for knots, constraint can
propagate from d to n4.

• Without any outlet from n4, the flow of constraint stalls here. Remember
that the edge between n4 and c has already been deleted above, and no
longer exists in the remainder matrix. This violates the T-10 rule of d(n)
– 1 in and 1 out for nodes. Therefore, the original model can be
concluded as inherently inconsistent, and no computational requests can
be made on it.

The procedure outlined in Definition 45 and demonstrated above should

have better efficiency in terms of both run-time and space-bound than steps
(c) – (f) of D-38 in their current form. The removal of every vertex
discovered along the constraint propagation path can repeatedly simplify the
model graph, and thus reduce the size of the remainder matrix further and
further. As such, the computational load for subsequent iterations will
require less and less CPU time and memory resources.
5. Consistency and Allowability 123

1 3

Source
a b 5 c d
BNS

Constraint flows from b into n 5,

2 source BNS is deleted from G. 4

1 3

Source
a b 5 c d
BNS

Constraint flows from n5 into c,

2 n5 is deleted from G. 4

1 3

Source
a b 5 c d
BNS

Constraint flows from c into n3,

2 c is deleted from G. 4

1 3

Source
a b 5 c d
BNS

Constraint flows from n3 into d,

(!)
2 n3 is deleted from G. 4

Figure 5-11. Violation of the d(n) – 1 in and 1 out rule for nodes exposes an
inherently inconsistent model
124 Constraint theory

5.7 PROCESSING OF COMPUTATIONAL

REQUESTS

“As we observed from Chapter 1 and even more in Appendix A, even

though the model may be consistent, the probability that any given
computational request will be allowable is still very small,” remarked the
manager. “Can constraint theory assist the math model manager in steering
him to those computational requests which are allowable?”
“Certainly,” the analyst assured. “We can consider another ‘brute force’
approach here, where constraint theory can theoretically examine all the
possible 2K computational requests. But, as we have already seen, 2N
computations are really out of the question and 2K would be roughly the
same size.”
“We can enormously reduce the number of computational requests to be
analyzed by noting two things,” the analyst continued. “First, the cognitive
limitations of any human attempting to understand a result from a math
model are limited to ‘view spaces’ whose dimensionality ranges from 2 to 5
at the most. So the 2K computations are reduced to ~K5 – a tremendous
reduction. Second, not all the K variables are equally interesting or
important; for large models, probably only a few percent of the total number
of variables would be considered as providing a ‘system level’ overview. We
can take a hint as to which are the most interesting (or valuable) variables by
just looking at the computational requests which were disallowed. Assume
for example that a failed computational request involved the variables
a,c,f,m. We could then ask D38 to examine all combinations of three of these
and any other of the K  4 variables, and two of these and any two other of
the K  4 variables, and finally any one of these and any three other of the K
 4 variables. This amounts to less than 4K3 examinations, which is over a
million possible computational requests but can be computed by D38 in just
a few minutes. This should provide the manager with many computational
requests which are in the ‘vicinity’ of the one he wanted and couldn’t have.”
Given an inherently consistent model graph by steps (a) – (f) of D-38, we
are now ready to examine in further details the general procedure for
processing computational requests as outlined in steps (g) – (k).

5.7.1 INITIAL SIMPLIFICATION OF MODEL GRAPH

We can again leverage T-11 to simplify a model graph before

entertaining any computational request made upon it. Since all the knots
belonging to a resultant constraint domain, previously identified by the
method outlined in Definition 45, have been internally fixed in values, they
5. Consistency and Allowability 125

can no longer be manipulated as part of any computational request.

Therefore, the sub-graphs spanning the various constraint domains can be
removed from the overall model, via a vertex-deletion operation (Phan, pp.
101-102) after step (f) and before step (g) of D-38.

This reduction can be safely implemented without losing any essential

information in determining the allowability of any computational request, or
the processing of an allowable request. Within the framework of such a
reduced graph, one can avoid many repetitious and unnecessary calculations.
And computational complexity can be exponentially lessened in terms of
both run-time and memory-space requirements. This improvement in
efficiency follows the same concept as presented in step (1b) of Definition
45 where vertices discovered along a constraint flow path are removed
during propagation of internal constraint.

Figure 5-12 illustrates above simplification technique. Upon the removal

of the intrinsic BNS = { 2, 3, 4, c, d, s } which just happens to be the same as
its resultant constraint domain in this case, model graph G is reduced to sub-
graph H. All computational requests made on G can be evaluated with the
remainder matrix associated with H. Any request involving c, d or s, as
dependent or independent variables, can be readily concluded as unallowable
since their corresponding knots  KH. This immediate determination will
help avoid the unnecessary propagation of constraint as stipulated by steps
(h) – (j) of D-38. Allowability of other requests, as well as the processing of
all allowable requests, can also be computed more efficiently in H than in G
by the fact that | VH | ” | VG | and | EH | ” | EG |.

By T-7, the number of possible computational requests on a model with

K variables is equal to 2K. In the above simple example BPG, with | KG | = 8,
there would have been 28 = 256 computational requests that could be
possibly made on G. With | KH | = 5, however, the number of possible
computational requests on H is now only 25 = 32, an order of magnitude less
than that on G. For real-world models involving hundreds, if not thousands,
of variables, the reduction in computational complexity would be even more
significant.

Another advantage of the above simplification technique is that the

removal of sub-graphs spanning various constraint domains may also break
up a BPG into several disjoint connected components. In such cases, a
computational request involving variables not all of which belongs to the
same connected component can also be quickly determined as unallowable
126 Constraint theory

without having to propagate needlessly throughout the graph network. This

is in accordance with T-5.

5.7.2 SIMPLIFYING MODEL GRAPH DURING

CONSTRAINT PROPAGATION

To evaluate allowability of a computational request made on a model,

steps (h) and (i) of D-38 repeatedly apply the Duality Rules of T-10 and the
BNS search process of D-37 to propagate constraint throughout its graph
network. In their current form, the general procedures for these steps state:

h) Propagate computational paths from all constraint sources throughout

the model employing the T-10 rules. If any knot is over-constrained, the
computational request is not allowable.

i) If the computational path does not reach the dependent variable then
examine the residue for BNS, using D-37, and continue with T-10,
followed by D-37 as necessary until the computational path can go no
further.

For computational requests with multiple independent variables as input,

the processing should initiate the propagation of constraint externally applied
to one independent variable at a time in a single thread of execution, and
complete the D-38 procedure as far as T-10 and D-37 will allow before
starting propagation of constraint from another independent variable. To
maintain flow control, independent variables should be processed
sequentially one at a time. If several independent variables are processed in
parallel, their computational paths may unknowingly collide with
unpredictable results. Given that, as vertices are discovered one by one along
the computational path originated from one independent variable, they can
be dynamically removed from the BPG by a vertex-deletion operation (Phan,
pp. 101-102). In several ways, this reduction can simplify computational
complexity, and thus improve efficiency of the iterative process stipulated by
steps (h) and (i) of D-38. The operational benefits include:

• With the ever-reduced remainder matrix, the D-37 process to locate

resultant BNS can only execute faster and faster, and require less
and less computer memory resource with each subsequent iteration,
than processing the entire model graph each and every time. The
same can also be said of the T-10 process to propagate constraint.
5. Consistency and Allowability 127

2 3
Model graph G

d a m

c s 5 6

e
4 p

Intrinsic 3 × 3BNS

Removal of intrinsic BNS from G

results in simplified sub-graph H.

a m

5 6

e
p

Vertex-deletion sub-graph H

Figure 5-12. Removal of BNS simplifies a model graph before processing

any computational request
128 Constraint theory

• Deletion of vertices may break up the BPG into several disjoint

connected components. Before initiating another path of propagation
from the next independent variable, a quick and easy check can be
performed to verify this independent variable as belonging to the
same component as the dependent variable. If not, it can be readily
concluded that the computational request is not allowable per T-5.
This immediate determination will help avoid the repetitive and
unnecessary propagation of constraint from this independent
variable, as well as others subsequently.

• If the computational path initiated from one independent variable

discovers the knot associated with another independent variable,
then the request can be immediately rendered as unallowable.

Theorem 30: If the constraint flow path originated from an independent

variable x1 as input to a computational request discovers, by the Duality
Rules of T-10 for propagating constraint throughout a graph network, the
knot associated with another distinct independent variable x2 of the same
request, then the request is not allowable.

Proof: Independent variable x2, as input to the computational request just

as x1, will also be externally constrained by definition. If x2 lies on the
computational path originated from x1 then x2 is over-constrained once the
flow of constraint from x1 flows into it. This violates the propagation rule of
“1 in and d(k) – 1 out” for knots under T-10. Over-constraint on x2 causes
the model to become inconsistent. And no computational request on an
inconsistent model is allowable per T-1.

5.7.3 UNALLOWABLE COMPUTATIONAL

REQUESTS

Steps (h) and (j) of D-38 have pointed out two possible scenarios in
which a computational request becomes unallowable due to over-constraint
of a knot along the computational path. The following sub-sections will
discuss other classes of unallowable computational requests.

5.7.3.1 OVER-CONSTRAINT OF NODES

For a given computational request, there may exist multiple possible

computational paths, depending on the order (or sequence) of propagating
externally-applied constraints. Per T-10, a node along a computational path
will become over-constrained (or under-constrained) if the propagating rule
5. Consistency and Allowability 129

of “d(n) – 1 in and 1 out” for nodes is violated. Accordingly, such a

computational request may also be rendered as unallowable.

Figure 5-1 graphically illustrates this situation. If the two constants in the
bottom half of the graph are propagated first, then knots d and g will be
discovered as over-constrained as stipulated by steps (h) and (j) of D-38. At
this time, the procedure can be halted and the computational request declared
unallowable. However, if the two independent variables in the top half of the
graph are applied first as external constraints, then node a will be discovered
as over-constrained. Accordingly, the same computational request can also
be rendered unallowable and the procedure immediately halted.

5.7.3.2 RELEVANCY OF DEPENDENT VARIABLES

Step (k) of D-38 states that: “If the path reaches the dependent variable
by employing all independent variables, without either over- or under-
constraint along the way, then the computational request is allowable. It is
acceptable to have local under-constraint elsewhere in the model. If the
dependent variable can be reached without having to propagate the
constraint externally imposed on one or more independent variables, then the
computation is unallowable”.

Per the following concepts in Constraint Theory, another class of

unallowable computational requests involves extra independent variables
whose externally-applied constraints are not necessarily used to reach the
desired dependent variable.

Definition 9: (p. 34): “y is a relevant variable with respect to relations φ in

xyz space means that there exist lines in xyz space parallel to the y-axis that
are neither entirely within nor entirely outside of the relation set. Thus, y has
an effect on I, or equivalently, the relation p constrains y”.

Definition 14: (p. 40): “a computational request on a model is allowable

means that the projection of A6 onto the view space of the computation
contains at least one point and, in addition, each variable involved in the
computation must be relevant to this projection in the sense of Definition 9”.

It has been previously illuminated and asserted (pp. 40, 45) that: “If the
projection has variables that are not relevant, these variables take on all their
possible values, and are therefore under-constrained. Allowability requires
that all the variables of the requested computation be relevant to the
projection of the total relation onto the computational sub-space.”
130 Constraint theory

Let’s consider the computational request e = f1(m, t) made on sub-graph

H of Figure 5-12. Figure 5-13 illustrates the propagation of constraints
externally imposed on the independent variables m and t (highlighted with
double squares). As knot m is initially fixed, its constraint flows into node 6,
resulting in a 2 × 2 BNS. In turn, this resultant BNS fixes knots a and e, and
the computational request is thus completed. It was not necessary, or useful,
to propagate the external constraint imposed on t. Had the constraint on t
been propagated first (before that on m), the computational path would never
pass node 1 to reach e. Violating the “d(n) – 1 in, 1 out” rule for propagating
constraint through nodes, node 1 would have been under-constrained.
Therefore, this computational request is not allowable since t is irrelevant in
determining e. In other words, e can be sufficiently determined without t.

a m

Independent variable
Resultant
5 6
BNS

e
p
Dependent variable

Independent variable
1

Figure 5-13. All independent variables must be relevant for a computational

request to be allowable

Per T-2, over-constraint must not exist anywhere in the model, for a
computational request to be allowable, since any sub-model inconsistency
would “poison” the entire model. However, it is acceptable to have locally
under-constrained vertices outside of the computational path. For example,
in Figure 5-14, the computational request m = f2(a) is allowable even though
variables p and t are locally under-constrained elsewhere.
5. Consistency and Allowability 131

Independent variable

a m Dependent variable

5 6

e
p

Node 1 local under-constrained

1
outside of computational path
t

Figure 5-14. Locally under-constrained vertices outside

of the computational path are acceptable

5.8 SUMMARY OF CHAPTER AND CONSTRAINT

THEORY TOOLKIT

The computational rules used in Chapter 1, “d(n)  1 in and 1 out” for

nodes and “1 in and d(k)  1 out” for knots, as well as the “BNSs being the
kernel of constraint” were formally sanctified from the viewpoint of the
math model’s being a set of relations within the multidimensional space
defined by the model’s variables. All these results were seen to be aspects of
the Goldilocks rule which stated that computational allowability requires that
the resultant constraint along the entire computational path from independent
to dependent variables be exactly zero.
The exhaustive search for BNSs requires 2N examinations of math model
relation subsets; even for moderate model sizes of N = 100, thousands of
universe lifetimes are required, even for nanosecond examinations.
Instead, D-37 can locate the BNSs in only seconds instead of trillions of
years. This is done by analyzing the topology of the BNSs which are
imbedded within the topology of the bipartite graph meta-model, employing
easily computed features such as connectedness, tree-ness, circuit rank and
constraint potential. A key result is the proof that, if there is a BNS within a
bipartite graph, it can only exist within a circuit cluster (cc) and if the
132 Constraint theory

constraint potential of that circuit cluster p(cc) is equal or greater than zero,
then that circuit cluster contains p(cc) + 1 BNSs. See Figure 5-15.
Once the BNSs are located, the model consistency and computational
allowability are easily determined, as is summarized in Tables 5-1 and 5-2,
the Constraint Theory Toolkit.
Figure 5-16 demonstrates a typical scenario for the flow of constraint
across a small bipartite graph model.
The many definitions and theorems of Chapters 4 and 5 may appear
onerous but they are necessary to establish the rules of D-37 and D-38 as
applicable to any mathematical model of any size. This is a demonstration of
the power of generalizability of mathematics. We start in tiny domains of
low dimension which we can comprehend, and then extend our
understanding and tools to dimensions of any size. As large as the numbers
2N and 2K are, the number of possible topologies of a bipartite graph is
unimaginably larger: 2NK.
5. Consistency and Allowability 133

Figure 5-15. By employing easily computed topological properties of a model’s bipartite

graph, consistency and allowability checks can be reduced from universe lifetimes to seconds.
134 Constraint theory

Figure 5-16. A walk through the trees and tangled clusters.

5. Consistency and Allowability 135

Table 5-1. Constraint Theory Toolkit (Part I).

To Determine Mathematical Model Consistency:
Organize the set of regular relations and their relevant variables into a bipartite
graph (BPG) metamodel with two types of vertices:
– Nodes (N) which represent relations, and Knots (K) which represent
variables
– Edges connect relations (nodes) to their relevant variables (knots); a BPG
with undirected edges represent the math model; a BPG with directed edges
represents the computation
– The degree of a vertex, d(v) is the number of edges which intersect that
vertex.
Employ the companion Constraint Matrix (CM) to communicate with computer
analysis.
The constraint propagation rules across a BPG are:
– for nodes: d(n)-1 edges flowing into the node, 1 edge flowing out.
– for knots: 1 edge flowing into the knot, d(k)-1 edges flowing out.
Propagate connectivity along BPG edges (or CM dots) to determine connected
components.
– If K+N-E=1 in a connected component, it is tree-like; no intrinsic
constraint in component.
– If not tree-like, circuit rank=E-K-N+1=number of independent circuits in
component.
– If the constraint potential, p(SM), of a submodel = N-K>0, intrinsic
constraint exists there.
A submodel wherein N=K is a nodal square (NS); a nodal square with no NS within
it is a basic nodal square (BNS) which is the kernel of constraint in the math model.
– Overlapping BNSs (with p(BNSs)>0) indicate overconstraint causing
INCONSISTENCY.
– Constraint propagates from the BNSs to resultant constraint domains via
the above rules.
– If resultant constraint domains overlap, they will either be redundant or
INCONSISTENT.
– Inconsistency must be negotiated by the human model builders to resolve
overconstraint.
BNSs can only exist within circuit clusters (cc’s).
A systematic search for BNSs involves separating the connected components,
trimming external trees, eliminating internal trees, separating kissing cc’s by
removing separating vertices, and finally computing the constraint potential of the
remaining cc’s.
136 Constraint theory

Table 5-2. Constraint Theory Toolkit (Part II)

To determine computational allowability:

If a mathematical model is inconsistent, no computations on it are allowable.

Add the sources of extrinsic constraint – independent variables and variables held
constant – to the sources of intrinsic constraint and propagate constraint through the
BPG using the propagation rules above. If overconstraint occurs at any vertex, the
computation is NOT ALLOWABLE. If the constraint does not propagate to the
dependent variable, using the above procedure, search for BNSs which can continue
the propagation. If the constraint flow still does not reach the dependent variable, the
computation is NOT ALLOWABLE due to underconstraint. If neither over- or
underconstraint occurs, the computation is ALLOWABLE.

5.9 QUERIES FOR THE

REGULAR STUDENT

1. For the mathematical model of Problem 4.1, which of these

computational requests are allowable? If not, why not?
g = f(r, c), g = f(s, t), m = f(w, h, s, t), h = f(r, w), d = f(g, r)

2. Is the mathematical model of Problem 4.2 consistent? If so, develop

at least two computational requests on this model.

3. Is the mathematical model of Problem 4.3 consistent? Which of these

computational requests are allowable? If not, why not?
n = f(r), q = f(p), m = f(n, q), p = f(m, r), m = f(p)

4. Prove that the bipartite graph of a model with N relations and K

variables has 2KN different topologies.

5. Derive the number of possible computational requests can be made

on a model with N relations and K variables if it were not for T6 and
if the relations which formed the model were not counted as
computational requests.
Chapter 6 DISCRETE AND INTERVAL
RELATIONS

The diminished utility of metamodels

6.1 METAMODEL ISSUES AND PERSPECTIVES

“You recall that in Chapter Three, we defined three types of relations,”

said the analyst. “The most important of these types – from the standpoint of
math modeling – is the regular relation and was treated in Chapter Four. In
this chapter, we will look at the other types, called ‘discrete’ and ‘interval’.
“This must have represented quite an intellectual leap, starting I presume
from the findings of regular relations,” the manager suggested.
“On the contrary,” the analyst differed. “The original research in
constraint theory actually started within the domain of discrete relations. It
was stimulated by a paper on the solution of simultaneous equations in
Boolean algebra, written by Antonin Svoboda [10], who was a member of
Friedman’s PhD committee. The use of the extension and projection
operators in set theory was more easily visualized when only points – rather
than multidimensional curves and surfaces – were involved. To generalize
even further, we should remember that thinking about ordinals is more
fundamental and closer to metamathematical language than thinking and
operating on cardinals. Logic precedes analysis.”
The perspectives and tools which we found to be useful for regular
relations will now be examined for their utility regarding discrete and
interval relations.

© Springer International Publishing AG 2017

G.J. Friedman, P. Phan, Constraint Theory, IFSR International Series on Systems
Science and Engineering, DOI 10.1007/978-3-319-54792-3_6
138 Constraint theory

6.2 THE GENERAL TAXONOMY AND PRIMARY

PROPERTY OF DISCRETE RELATIONS

The discrete relation defined by D19 may appear in a variety of forms, as

is displayed in Figure 6-1. Relation 1, a polynomial in a single variable is
similar to a regular relation, yet it will point constrain its one-dimensional
space at each of its roots. Relation 2 is an example of Diophantine equations
which permit only integers as allowable solutions. Note that most of the
equations in Chapter 4 – such as the definitions of constraint potential and
circuit rank – are this type of equation. Relation 3 is described by a “truth
table” which lists every point in its allowable space. Relation 4 is a logical or
Boolean equation, wherein all the variables take on the values of true or
false, or more compactly, 1 or 0, respectively. Finally, Relation 5 –
represented by a matrix – can represent even more abstract mathematical
forms such as bipartite graphs, via their companion matrices. In this sense,
discrete relations can be considered as meta-metamodels for the graph theory
described in Chapters 2, 3 and 4.
All these diverse characterizations of discrete relations can be covered
by:
Theorem 31: Every discrete relation is an intrinsic source of point
constraint with respect to each of its relevant variables.
Proof: By D19, in a discrete relation the intersection of any line with that
relation is a point or set of points. By choosing that line to be, in turn, each
axis of the relation’s space, each of the variables is seen to be point
constrained by that relation. QED.
Thus, the problem of finding sources of intrinsic point constraint and the
search for BNSs which was given so much attention in Chapter 4 is almost
trivial for discrete relations.
In pursuit of constraint theory’s general goal of determining consistency
and computational allowability, considerations of overconstraint and
multidirectional constraint propagation, as well as computational flow rules
will be treated in the following sections.

6.3 BOOLEAN RELATIONS

Boolean relations represent a significant portion of the class of discrete

relations and can be described by logical equations, truth tables, Venn and
Veitch diagrams, or by the complete mapping of allowable states in the
hyperspace of relevant variables. Figure 6-2 presents a simple example of
the logical equation: A=BC. In words, this merely states that A is true (1) if
B is true (1) and if C is not true (0). The truth table of Figure 6-2b provides
the value of A for all possible combinations of B and C. The Veitch diagram
6. Discrete and Interval Relations 139

Figure 6-1. Examples of Discrete Relations

– essentially a rectangular version of the Venn diagram – in Figure 6-2c also

shows all possible values of the implied independent variables B and C and
provides the value of A for each case. Sometimes, “forbidden domains” on
the Veitch diagram are specified. These are combinations of the input
variables which can never occur; when forbidden domains intersect domains
140 Constraint theory

defined by the logical equation, these forbidden domains take priority. This
convention permits the simplification of logical equations which is important
in the logical design of digital computers. In accordance with the
hyperspatial viewpoints discussed in Chapter 2, when the total model is
defined by a set of simultaneous Boolean equations, the intersection of the
Boolean equations and the union of the forbidden areas are taken.

Figure 6-2. Boolean relations; and their representations.

6. Discrete and Interval Relations 141

Both the truth table and Veitch diagram representations suffer from
implications of independent/variables and forbidden zone ambiguity.
Figure 6-2d displays the most complete representation of any discrete
function: the identification of the set of points within the hyperspace of
relevant variables. For any binary Boolean equation, this set can reside only
at the corners of the hypercube formed by the points 0 and 1 along the axis
of every relevant variable. That is, these are the values of the variables that
are allowable by the defined relation. In the spirit of the four fold way
presented in Chapter 2, call the set of points that satisfy the equation “A.”
Let us now examine how the computational rule: (d-1)in/1 out that was
developed for regular relations applies to Boolean relations.
Figure 6-3 displays all possible computational paths through the relation
A=BC. It shows that for B and C taken as input variables, every one of the
four possible cases, the (d-1)in/1 out rule works. (This should not be
surprising , since the format of the Boolean equation, truth table and Veitch
diagram were predisposed towards this computational flow.) However, in the
cases of A,B or A,C as input variables, it is seen that only half of the cases
follow the (d-1)in/1 out rule, while the other half result in either multiple
answers or represent forbidden inputs. Noting that multiple results also can
occur in regular relations and can still be carried forward to other
downstream computations, and that regular relations were defined to exclude
forbidden inputs, it can be concluded that the (d-1)in/1 out rule weakly
applies to discrete functions.
Figure 6-3 explores the more extreme case – from our regular relation
perspective – of a single input variable. There exists one case out of six
where the single variable A, set equal to 1, can uniquely determine the
values of B and C. This is a case of 1 in/(d-1)out for nodes, which is
surprising if we are used to thinking only of regular relations.
In summary, it can be seen that the computational flow across discrete
relations depends far more on the precise nature of the relation than on
metamodel rules such as (d-1)in/1 out.
Next, let us examine the applicability of constraint potential, (N-K), to
determine the consistency of Boolean relations with K variables and N
relations.
As noted above, a Boolean relation of K variables can exist only at the
corners of a hypercube formed by the coordinates 0 and 1 along each of the
K dimensions of the relation. Therefore, each of the N relations can be
represented by a “binary number” containing 2K states of A’s for points
contained by the relation and empty spaces which do not contain the relation.
Figure 6-4 provides the general format of a Boolean model. Note that,
despite the fact that it is a rectangular array with the rows referring to
variables and columns referring to relations, this is not a constraint matrix as
142 Constraint theory

defined in chapter 2. The total number of rows equals the number of possible
states of any given Boolean function, and equals 2K, which is the power set
of the K states. The total number of columns equals all possible Boolean
functions which is the power set of the 2K possible states and equals
2^(2^K). In other words, the number of columns is the power set of the
number of states which is the power set of the number of variables.

Figure 6-3. Computational Flows of A=BC

6. Discrete and Interval Relations 143

We now pose the question: If all the N columns refer to non-redundant

Boolean equations, what is the largest N which will not cause the total model
to become inconsistent? The answer to this question results in:
Theorem 32: For a 2-valued Boolean model consisting of K variables
and N independent (non-redundant) relations, the largest N which will not
2K 1
overconstrain the model is: Nmax = 2 .

Proof: Recall that, from Chapter 2, the model will be inconsistent if the
total allowability set is the null set. Also recall that the intersection of a null
set with any other set results in a null set. By examining Figure 5-4 (b), we
note that exactly half of the columns can be paired with the other half such
that their intersections are null sets. (Specifically, look at columns 1 and 16,
2 and 15, …etc.) Also note that no other columns have null set intersections.
Therefore the maximum number of relations which will not result in a null
set is exactly half the total number of columns. In short, the maximum N
which assures consistency is half the number of columns, or: Nmax = ( ½ )(
K K K
22 ) = ( 2-1 )( 22 ) = 22 1
. QED.
144 Constraint theory

Figure 6-4. Proof of T32

6. Discrete and Interval Relations 145

Figure 6-5. The Constraint Potential for Boolean relations is weaker than for regular relations.

Thus, the constraint potential (N-K) for regular relations can be replaced
2K 1
by (N- 2 ) for Boolean relations, and as is shown in Figure 6-5, it is a
weaker indication of the model’s inconsistency. Once again, we see that the
specific properties of the relations must be examined and the metamodels, so
useful in Chapters 4 and 5, are less useful for discrete relations.
Additional properties and methods of discrete relations are therefore
discussed in the following.

6.4 TOPOLOGICAL IMPLICATIONS

These two types of discrete relations will have importance from a

constraint standpoint:
Definition 47: A full discrete relation is an explicit discrete equation in
which each point in the space of independent variables provides at least one
value of the dependent variable.
146 Constraint theory

Figure 6-6. Examples of Full and Perfect Expressions.

Definition 48: A perfect discrete relation is a full discrete relation in

which each point in the space of independent variables determines a unique
value of the dependent variable.
Examples of these definitions are shown in Figure 6-6. Relation 1 is not
full since only three of the possible eight points in the abc-space provide
values for x. Relation 2 is not full because the point def is specified as being
forbidden and thus has no output to y. The map, 3, relating to the 4-state
6. Discrete and Interval Relations 147

variables u and v, is full if v is the dependent variable but is not full if u is

made the dependent variable. Thus, “fullness” and “perfection” are
properties of the explicit expression of a relation rather than of the relation
itself. Relation 4 is full with respect to x but is not perfect because the
“input” point (y=0, z=1) yields two values for x: =+1 and –1. Finally,
Relation 5 is perfect with respect to the ternary variable a because every
point in bc-space determines a unique value for a.
Theorem 33: Every Boolean relation which is perfect with respect to at
least one of its variables, is a universal relation.
Proof: Express the relation in its perfect, explicit form. By D48, every
point in the product space of the input variables produces an output; thus the
relation is universal. QED
Theorem 34: Every tree subgraph of Boolean relations each of which is
perfect with respect to at least one variable, is consistent.
Proof: By T33, the tree has only universal relations, and by T9, this tree
of universal relations is consistent. QED
Examples of the application of these Theorems are shown in Figure 6-7.
In Figure 6-7a, the mere specification that Relations 1, 2 and 3 are Boolean,
and that they are written in explicit form without specific forbidden domains
allow us to conclude that they are perfect with respect – at least – to the
dependent variables shown. By T34, then, the tree in Figure 6-7a is
consistent. Note that the dependent variables of Relations 1, 2 and 3 do not
indicate an obvious direction of computation or constraint flow.
Figure 6-7b presents a more general case of a discrete tree, mixing the
binary variable y with the ternary variables v, w, x and z, and the quaternary
variable r. Inspection of the truth tables for Relations 4, 5 and 6 discloses
that each relation is universal. Thus, by T9, this discrete tree is consistent.
That these rules cannot be generally extended to submodels whose graphs
are circuits is demonstrated in Figure 6-7c. Despite the fact that both of the
Relations 7 and 8 are perfect with respect to all their variables, and both are
universal, neither T9 nor T34 are applicable since they are connected in a
circuit cluster. In fact, a plot on a klm-space Boolean map will show that the
intersection of their relation sets is the null set for the given Relations 7 and
8, rendering the model inconsistent.

6.5 ALLOWABILITY OF DISCRETE

COMPUTATIONS

A general approach to the determination of the allowability of any

computational request made on discrete relations is described below.
The basic concept employed is a modification of the principles outlined
by A. Svoboda [15] in solving simultaneous systems of Boolean equations.
148 Constraint theory

Figure 6-7. Trees of Discrete Relations.

Let us begin by applying a viewpoint established in Chapter 2. Refer to

the top of Figure 6-8, where the discrete Relation 1 is shown in perfect form
with respect to the 4-state variable z. Given this representation of the
Relation 1, let us request the computation: x=f(y,z), where x is binary and y
6. Discrete and Interval Relations 149

is ternary. By Figure 2-14, we must first take the xyz-space “view” of A1 by

taking the projection: PrxyzA1. This is accomplished in Figure 6-8 by
replotting the 12 points of A1 from the z by wxy map to the w by xyz map,
suppressing the value of variable w, and replotting in an x by yz map. Note
that PrxyzA1 contains only 11 points due to the fact that two points in A1
project onto a single point, (x,y,z)=(1,1,1) in the xyz view. The final step is to
transform the x by yz map of PrxyzA1 into a “function map” wherein the
outputs for x are plotted as a function of the 12 possible points in the yz input
space. As can be seen in the lower right hand corner of Figure 6-8, only 5
input points in yz-space yield unique results, while 3 input points yield
multiple outputs and the remaining 4 yield no outputs at all; they are
“forbidden.”

Figure 6-8. Determination of X=f(y,z) by projection and replot.

6.6 INEQUALITY RELATIONS

As defined by D19, an interval relation is any relation such that its

intersection with a line is an interval containing an infinity of points.
150 Constraint theory

For all practical cases, the interval relation may be represented by a

system of inequality equations, or more generally by a combination of
inequality and “equals or greater (less) than” equations.
Definition 49: Inequality relations are defined by “<” and “>”. The
boundary set of equations are formed by replacing each < and > with equal
signs.
Figure 6-9 provides examples of D49.

Figure 6-9. Comparisons between pure inequality equations, equals or greater (less) than
relations, and boundary sets.

It was previously noted that, for regular relations, the BNS was required
to provide intrinsic point constraint, and for discrete relations, intrinsic point
constraint was always provided. We are now prepared to summarize the
point constraint situation for all classes of relations:
Theorem 35: In discrete relations, intrinsic point constraint always
occurs; in regular relations, intrinsic point constraint occurs in the presence
of a BNS; in inequality relations, intrinsic point constraint never occurs.
Proof: The first portion is true by T31; the second is true by T11; the last
is true because inequality relations do not include their boundary sets and
thus all intersections either include an infinite number of points or the null
set. QED
It is noteworthy that in a broad range of applications of math models, it is
the boundary set which is of the greatest interest. An example of this is the
general optimization techniques of linear programming where the
optimization of a linear criterion occurs only at the intersection of the
6. Discrete and Interval Relations 151

boundary sets of the constraining functions. Thus for these important

applications, the inequality sets collapse into sets of regular relations.
The next consideration is to examine the utility of the constraint potential
in determining consistency. As can be seen in Figure 6-10, the number of
inequality constraints can increase indefinitely without driving the overall
model into inconsistency. This observation permits us to state another
generalization across all relation classes:
2K −1
Theorem 36: For model consistency, in discrete relations, Nmax= 2 ;
in regular relations, Nmax=K; in inequality relations, there is no Nmax.
Proof: By T32, by T26, and by inspection of Figure 6-10. QED

Figure 6-10. The number of inequality constraints can increase to infinity without driving the
model’s relation set to null (inconsistency).

Next, examining the propagation of constraint or computation through

the metamodel, we invoke the transitivity rule of inequalities to prove:

Theorem 37: If a>b, and b>c, then a>c.

152 Constraint theory

As a side observation, if a, b and c are preferences, then the intransitive

preference ordering of a>b>c>a is considered irrational by decision
theorists.
Let us consider the implications of T37 on the allowability subset A in K-
space. In the upper part of Figure 6-11, the “A in K-space” interpretation of
the Theorem is presented. For ease of visualization, only the unit cube in the
upper octant is shown. The two intersecting planes are the boundary sets for
Relations 1 and 2 and the resultant allowability set A1 ъ A2 is the
tetrahedron with the corners o,x,y and z.
The projection of A1 ъ A2 on the ac plane is seen to cover half the plane
and corresponds to the interval relation a>c, substantiating T37. Thus we see
that a and c are mutually constraining: constraint on a will propagate to c
and vice versa.
On the other hand, consider the lower part of Figure 6-11, where the
inequality b>c has been replaced by b<c rendering T37 inapplicable.

6.7 SUMMARY

• Although the general principles of constraint theory are still applicable

to discrete and interval relations, the metamodel approach is less
powerful than for regular relations. Therefore the metamodel
viewpoints must be augmented by direct examination and computation
of the full model itself.
• Certain generalizations can be made across all relation classes:

Regarding: Discrete Relations Regular Interval

Relations Relations
Intrinsic Point
At every node At a BNS only Never
Constraint
Maximum N
2K 1
without Nmax = 2 , Nmax=K Nmax=infinite
overconstraint

The new allowability space becomes the pyramid with corners at

o,x,v,w,and z. The projection of this pyramid onto the ac plane covers the
entire plane and the variables a and c now do not exert constraint on one
another, in the view defined in Chapter 2.
6. Discrete and Interval Relations 153

Figure 6-11. Transitive and intransitive inequality relations.

154 Constraint theory

6.8 PROBLEMS FOR THE DISCRETE STUDENT

6.1 Given: a model of random Boolean relations with K variables and N

relations. Derive: Probability of inconsistency as a function of K,N.
6.2 Construct a model comprised of three ternary discrete relations and
determine under what circumstances the constraint flow follows the
nodal rule: (d-1)in/1 out.
6.3 Construct two Boolean functions and find a third function which
satisfy both.
6.4 Derive the general n-ary version of Theorem 32
Chapter 7 THE LOGICAL STRUCTURE OF
CONSTRAINT THEORY

A Compact Summary

7.1 OVERVIEW

This chapter provides a summary of all the postulates, definitions and

theorems presented in Chapters two through Six. Finally, two graph
structures are presented, showing the interrelationships between the elements
of constraint theory’s logical structure.

7.2 POSTULATES AND PHILOSOPHICAL

ASSUMPTIONS

Mathematical model building is a rapidly expanding activity. Since

computers do not suffer the dimensionality limitations of the human mind,
modeling is the best hope for systems engineers to manage the complexity of
modern and future systems. All model builders and users share (or should
share) these philosophical views and postulates:
Mathematics has been able to capture an incredible portion of natural and
man-made phenomena with amazing depth and accuracy. Just as valuable is
the ability of mathematics to logically integrate many diverse views of the
world, generalizing new findings and helping to manage the vast
dimensionality of complex systems.

© Springer International Publishing AG 2017

G.J. Friedman, P. Phan, Constraint Theory, IFSR International Series on Systems
Science and Engineering, DOI 10.1007/978-3-319-54792-3_7
156 Constraint theory

However, mathematics usually presents a simplified picture of the real

world and the variables in its models may not be defined coherently across
all members of a diverse team. These key issues are not within the domain of
constraint theory and it is presumed that they have been attended to prior to
the methods presented in this book.
Even when the issues of model accuracy and definitions have been
handled perfectly, there still exist crucial issues of well-posedness. These are
in the domain of constraint theory and the structure presented here addresses
them. Specifically, are models consistent and are the computational requests
made of them allowable?
Table 7-1 presents the postulates which have been directly involved in
the development of constraint theory.

Table 7-1. Postulates

1. Model Builders inherently wish the relations in their model to

be locally universal. (That is, if any relation applies a
constraint to a variable, that variable will be able to propagate
that constraint to an adjacent relation without causing
inconsistency.)
2a. The laws of physics and other sciences, if they are fully
understood, are fundamentally consistent.
2b. The intent of mathematical model builders is to represent
phenomena accurately; thus if overconstraint – or
inconsistency – occurs, it is unintentional.

7.3 DEFINITIONS

The forty definitions employed in Chapter Two through Six are listed in
Table 7-2.
Most of these definitions either contribute to other, more complex
definitions, or to theorems. In seven cases – marked by an asterisk – the
definitions are a result of a set of theorems and represent a computational
procedure.

7.4 THEOREMS

The thirty-three theorems derived in Chapters two through five are listed
in Table 7-3.
The proofs of all these theorems – except T12, “Hall’s Theorem” – are
provided as they are first introduced in the chapters. The proof of Hall’s
theorem can be found in the reference.
7. The Logical Structure of Contstraint Theory 157

Table 7-2. Definitions

D1: Set, Subset D26: Adjacency

D2: Variable * D27: Connectedness algr’m
D3: Model Hyperspace * D28: Sep. Vertex Algorithm
D4: Product Set * D29: Tree Algorithm
D5: Relation * D30: Twig pruning algr’m
D6: Constraint D31: Circuit Rank
D7: Union, Intersection D32: Simple Circuit
D8: Projection, Extension D33: Circuit Vector
D9: Relevance * D34: Circuit Rank Algr’m
D10: Bipartite Graph D35: Graph Taxonomy
D11: Model Graph D36: Independent BNS
D12: Constraint Matrix * D37: BNS Location
D13: Consistency * D38: General Procedure
D14: Allowability D39: Overlap BNS
D15: Connected Component D40: CharVec of BNS
D16: Tree Structure D41: BNS Matrix
D17: Circuit Cluster D42: Overlapping Factor
D18: Universal Relation D43: Algorithm to detect
D19: Relation Classes Overlapping BNS
D20: Locally Universal D44: Constraint Domain
D21: Regular Relations D45: Expanding Constraint
D22: Constraint Potential Domain
D23: Degree of a Vertex D46: Remainder Matrix
D24: Over Constraint D47: Full Discrete Rel’n
Under Constraint D48: Perfect Discrete
D25: Nodal Square Relation
Basic Nodal Square D49: Inequality Relations

7.5 GRAPHS OF THE LOGICAL STRUCTURE OF

CONSTRAINT THEORY

The graphical portrayal of the relationships between definitions and

theorems is presented in Figure 7-1. Although this graph has two disjoint
sets of vertices, it is not a bipartite graph because in several cases, there are
edges connecting definitions to other definitions, and edges connecting
theorems to other theorems.
Another graphical portrayal is presented in Figure 7-2. This graph is a
generalization of Figure 7-1 and suppresses many of the details but
emphasizes the major logical thrusts of the theory. The more important
definitions – such as the BNS – and the more important theorems – such as
the BNS location theorem – are highlighted.
158 Constraint theory

7.6 COMPLETENESS

The author makes no claim that these summaries and Constraint Theory
are complete.
Indeed, it his hope that this work will stimulate further research and study
into the increasingly important objective of managing complexity.
7. The Logical Structure of Contstraint Theory 159

Table 7-3. Theorems

T1: If a model is inconsistent, no computational requests on it

are allowable
T2: If any submodel is inconsistent, the total model is
inconsistent
T3: If two relations have no variables in common, they are
consistent
T4: If two disconnected components are internally consistent,
they are consistent with each other
T5: No computations across disconnected components are
allowable
T6: Allowability of a computational request is independent of
the dependent variable
T7: There exist 2K possible computational request of a K-
variable model
T8: There exist 2N possible submodels of an N-relation model
T9: Any set of universal relations whose graph is a tree is
consistent
T10: Computational rules for models with a tree structure:
nodes: d(n)-1 inputs --> 1 output
knots: 1 input --> d(k)-1 output
T11: Every BNS exerts point constraint on each of its variables
T12: Hall’s theorem
T13: A subgraph is a tree if and only if V-E=1
T14: The number of independent circuits in a graph = circuit
rank of graph
T15: The graph taxonomy of D35 is mutually exclusive and
exhaustive
T16: If SUM Pi > -n, then, at least one of the Pi > -1
T17: If p(G)=N-K>0, G contains at least one BNS
T18: No BNS can have a subgraph with a constraint potential>0
T19: Every BNS must be connected
T20: No BNS can be a tree
T21: No BNS can have a tree which is not part of a circuit
(continued)
160 Constraint theory

Table 7.3 Continued – Theorems

T22: No BNS can lie across circuit clusters containing a

separating vertex
T23: No BNS can lie across a tree-like network of circuit
clusters linked by a tree
T24: No BNS can lie across a tree-like network of circuit cluster
linked to other circuit clusters by trees
T25: Every BNS is the union of adjacent circuits within a circuit
cluster
T26: If a circuit cluster (cc) has p(cc) > 0, then it has at least
p(cc)+1 BNSs
T27: The maximum number of BNSs in a cc=2c(cc)
T28: BNSi and BNSj are overlapping if Zi, j • 1
T29: Solution time to detect overlapping BNS and identify over-
constrained variables using a matrix :G with R rows and K
columns ~ K · R2
T30: If the constraint flow path originated from an independent
variable x1 as input to a computational request discovers, by
the Duality Rules of T-10 for propagating constraint
throughout a graph network, the knot associated with
another distinct independent variable x2 of the same
request, then the request is not allowable
T31: Every discrete relation is an intrinsic source of point
constraint
2 K 1
T32: For a consistent model of Boolean relations, Nmax = 2
T33: Every Boolean relation which is perfect with respect to at
least one of its variables is a universal relation
T34: Every tree subgraph of perfect Boolean relations is
consistent
T35: In discrete relations, intrinsic constraint always occurs
In regular relations, intrinsic constraint sometimes occurs
(in a BNS)
In interval relations, intrinsic constraint never occurs
2 K 1
T36: For consistency in discrete relations, Nmax= 2
For consistency in regular relations, Nmax=K
For consistency in interval relations, Nmax=infinity
T37: If a > b, b > c then a > c
7. The Logical Structure of Contstraint Theory 161

Figure 7-1. Graph of Constraint Theory Definitions and Theorems

162 Constraint theory

Figure 7-2. Logical Thrusts of Constraint Theory

Chapter 8 EXAMPLES OF CONSTRAINT
THEORY APPLIED TO REAL-WORLD
PROBLEMS

8.1 APOLOGIES NOT REQUIRED

The examples provided in this Chapter are drawn from applications in the
aerospace industry that the author has experienced. Rather than apologize for
this apparently undemocratic representation, it should be noted that it was
the aerospace industry which gave the first major fertile ground for the
discipline of systems engineering to manage the ever increasing complexity
of integrating new technologies, equipment and missions.
The first example applies constraint theory to the operations analysis of
new systems; it is an elaboration of the example presented in Chapter 1. The
second example deals with the kinematics of free-fall weapons, which,
despite the advent of “smart bombs” still has utility in many operational
scenarios. The third example is a dynamic analysis of the control of the
trajectory of an asteroid employing mass drivers.

8.2 COST AS AN INDEPENDENT VARIABLE

(CAIV)

For several decades, the development process of new complex systems

by the United States Department of Defense (DOD), has permitted the
dominance of technical performance over cost and schedule. As a
consequence, the vast majority of programs suffered major cost overruns and
© Springer International Publishing AG 2017
G.J. Friedman, P. Phan, Constraint Theory, IFSR International Series on Systems
Science and Engineering, DOI 10.1007/978-3-319-54792-3_8
164 Constraint theory

schedule slips. Essentially, performance was the primary driving

requirement – or “independent variable.” In order to achieve a greater
measure of cost control and containment, the DOD initiated a management
thrust called, “Affordability, or Cost as an independent variable (CAIV)”.
Figure 8-1 provides a briefing chart used by the US Air Force Research
Laboratories (AFRL) that is intended to represent this new emphasis.

Figure 8-1. US Air Force Research Laboratories Briefing chart on Cost as an Independent
Variable.

At this broad level of detail, the math model employed is similar to the
example of Chapter 1. Mathematically, this type of problem is ideally suited
to the viewpoint and tools of Constraint Theory. Simply stated, a model is
built, and rather than follow the original computational flow from
performance specifications to cost, the flow is reversed; the independent and
dependent variables are switched. The new problem can be stated as:
“Given budgetary constraints on newly developed systems, what is the
optimum system design which provides the most acceptable performance
without exceeding these budgetary constraints?”
8. Examples of Constraint Theory 165

Let us flesh out in more detail the AFRL chart to attain a model which
can perform the desired analysis. Table 8-1 presents a list of variables; note
that K=14 across three levels of detail. Table 8-2 lists the relations, with
N=16. Noting that N>K, we should be immediately concerned with model
consistency.

Table 8-1. Variables for Model

166 Constraint theory

The top-level bipartite graph shown in Figure 8-2 is completely tree-like

and has K=14 (representing all the variables) and N=5. Clearly, there is no
concern about overconstraint or consistency at this level.

Table 8-2. Relations for Model

8. Examples of Constraint Theory 167

However, the 9 variables at the bottom of the graph are all inter-related,
as is shown in Figure 8-3 where 6 additional nodes are added to indicate
these relations. For example, node 7 describes how the cost of development,
CD, influences the radar signature, and node 9 describes how CD influences
the accuracy (CEP) of weapon delivery. Although we now have accumulated
several circuits, the total constraint potential is still less than zero, and there
are no local domains where the constraint potential exceeds zero – thus,
there are no BNSs. This can be verified by examining Figure 8-5 which
displays the constraint matrices for the various levels of detail in the model.

Figure 8-2. Top-Level Bipartite Graph.

The constraint situation gets far more serious when the final 5 nodes at
the bottom of Table 8-2 are included in the model. Note that each of these
last five nodes are “policy demands” specifying additional requirements on
reliability, weight, radar signature, etc., without directly relating them to a
higher level system criterion. The addition of these is devastating to the
consistency of the model. As is shown in Figure 8-4, the application of nodes
168 Constraint theory

12, 14 and 16 drives CD into inconsistency. Similarly, the application of

nodes 10 and 15 drives CP into inconsistency.
Thus, the bipartite graph and constraint matrix provide the managers and
analysts visibility regarding the model’s consistency, paving the way
towards an eventual “cost as an independent variable” analysis. When
overconstraint does occur, it can be pinpointed and – just as was done in
Chapter 1 – allowing policy nodes such as 12-16 of Table 8-2 to be
reconsidered for model inclusion. In short, rather than specifying a
mandatory level of reliability, signature, etc., let these variables “run free” in
the network of all other relations and variables. In this fashion, optimum
values can be computed rather than dictated.

Figure 8-3. Top- and Middle-level bipartite graph.

8. Examples of Constraint Theory 169

Figure 8-4. Full Bipartite Graph.

170 Constraint theory

Figure 8-5. Constraint Matrix for CAIV

8. Examples of Constraint Theory 171

8.3 THE KINEMATICS OF FREE-FALL WEAPONS

Consider the traditional bombing problem: an aircraft attempts to attain a

kinematic state which allows a free-falling bomb to impact on a target. A
great diversity of avionic equipment exists to measure the aircraft
kinematics, covering a great range of accuracy and cost. “What are the best
variables to measure?” is a common question. Before this question can be
properly answered, a more fundamental one must be addressed: “What sets
of variables are necessary and sufficient for a bombing solution, i.e., exactly
‘constrain’ the trajectory?” That this latter question is nontrivial is shown in
the following:
Define the following ten kinematic parameters of an aircraft with respect
to its target (See Figure 8-6):

X. horizontal distance between aircraft and target

X horizontal velocity of aircraft
Z. vertical distance between aircraft and target
Z vertical velocity of aircraft
r range from aircraft to target
.
r range rate, aircraft to target
S. angle from horizontal of range vector
S angle rate of range vector
V aircraft velocity magnitude
d aircraft velocity angle from horizontal.

For the sake of the example, only the kinematics within the bomb trajectory
plane are considered. All the variables listed previously except X are
measurable by at least one type of airborne instrument. We may now state
the problem more specifically.
. . .
.
Which subsets of (X, Z, Z, r, r, S, S, V, d) form complete descriptions of
the aircraft in-plane kinematics? Call these describing sets.
. .
From physical reasoning, the immeasurable set (X, X, Z, Z) completely
defines the kinematics
. . where (X, Z) define a two-dimensional positional
vector, and (X, Z) define a two-dimensional velocity vector. Assuming for
the moment that every describing set has four variables, then

9!
C49 = = 126
4!(9  4)!

four-element subsets of the measurable set must be examined to determine

if it is a describing set. This can be an extremely tedious problem since
only a few of these sets can be merely examined by inspection for easy
172 Constraint theory
. . .
classification
. as a describing set (Z, r, V, d), or not a describing set (X, Z, V,
d)

Figure 8-6. Weapon Delivery Variables (initial conditions).

Constraint theory provides both a point of view and a procedure to solve

this type of problem whether it be ten-dimensional or 500-dimensional. First,
a mathematical model which relates all the relevant variables is established.
Second, meta models consisting of a bipartite graph and a constraint matrix
are established to analyze the constraint properties of the original model.
Third, these meta models are tested and adjusted, if necessary, for internal
consistency. Fourth, the classification of any subset of measurables as a
describing set or not a describing set can be made by systematically
propagating constraint through the meta model. In this procedure, the subset
to be analyzed
. . acts as the input variables and any known describing set such
as (X, X, Z, Z ) acts as the output variables of a computational network.
Whether constraint theory is applied or not, the stated problem cannot be
solved unless the couplings and interactions between the kinematic variables
are clearly delineated as a set of relations. One such set may be the
following:
8. Examples of Constraint Theory 173

1: r = (X. 2 + Z2.)1/2
.
2: r = (XX + ZZ)(X2 + Z2)-1/2
3: X = r cos S
4: Z = r sin S
-1
5: . (Z/X)
S. = tan . 2
6: S = (ZX.2 – XZ)(X
. + Z2)-1
2 1/2
7: V = (X. + Z )
8: Z = -X tan d.

Each relation is correct. More correct relations can be written, but the
preceding list seems to be sufficient, whatever that means. Normally, the
analyst must rely on his intuition or judgment regarding the point at which
he should cease adding relations to his model. The bipartite graph and
constraint matrix meta models are presented in Figure 8-7.
The fundamental criterion for math model consistency is that the total
model relation set is not the null set. This simply means that the set of points
in the total ten-dimensional space which satisfied every one of the eight
relations must include at least one point, otherwise some part of the model is
intrinsically incompatible with another part. From the constraint point of
view, model inconsistency is due to overlapping domains of intrinsic
constraint.
Each relation in the example above is an n-1 dimensional regular surface
where n is the dimension of the relation. Thus they are all regular relations,
and the source of intrinsic constraint in this type of model is always a basic
nodal square (BNS), a special subgraph of the total bipartite graph which is
recognized by an equal number of vertices of each type. A logical first step
in the consistency testing is to search the meta models for BNSs, thereby
locating constraint sources.
Since for even this small model, a systematic search through all possible
submodels involves examination of 256 cases, it is clear that more powerful
means must be used to search for the BNS.
Every BNS is the union of simple nodal circuits. We know that every
BNS must be a circuit cluster and can contain no treelike appendages. . .
Examination of either Figures 8-7a or 8-7b reveals that the variables r, s, v,
and d form trees attached to the main circuit cluster; therefore, the nodes 2,
7, 8, and 6 cannot be part of a BNS. Temporarily stripping these nodes
(“trimming the trees”) from the model, we get the circuit cluster depicted in
Figure 8-8.
174 Constraint theory

Figure 8-7. Constraint Matrix Meta-model.

Since Figure 8-8 has the same number of each type of vertex (constraint
potential equal zero), it is a nodal square. Since every node has a degree of
three, there cannot exist smaller nodal squares of dimension 2x2 within
8. Examples of Constraint Theory 175

Figure 8-8. (The dimension of a nodal square must always be equal to or

greater than the degree of any of its nodes.) Since the degree of every knot is
also three, nodal squares of dimension 3x3 cannot exist within Figure 8-8
either. (In order for a nodal square of dimension n to contain a nodal square
of dimension n-1, there must exist a knot of degree 1.) Therefore, Figure 8-8
is a nodal square which does not contain smaller nodal squares and is thus a
basic nodal square. Moreover, it is the only BNS in the entire model.

Figure 8-8. Circuit Cluster.

We can generally expect that a BNS subgraph of a model of regular

relations will point constrain each of its variables. This does not violate the
fundamental criterion for model consistency.
But let us examine this BNS in the context of the purpose of the model.
We certainly did not expect the model to determine specific values of X, Z, r,
and S. Yet this is just what four equations covering four unknowns threaten
to do. Trying to resolve this absurdity, we focus our attention on this portion
of the original model and discover that relations 3 and 4 are redundant to
relations 1 and 5. 3 and 4 are therefore removed.
The new constraint matrix is shown in Figure 8-9. The total constraint
potential (relations minus variables) is now equal to –4, which agrees with
our physical reasoning that the total system should have 4 degrees of
freedom. Moreover, the subgraph 1-5 dealing only with position, and the
subgraph 7-8 dealing only with velocity, each have constraint potentials of
–2, further agreeing with our physical notions of position and velocity
vectors in a plane. . .
Certainly the set (X, X, Z, Z) is a describing set. Any other set (a, b, c, d)
will also be a describing set if, by
. setting
. unique values for (a,b,c,d), we can
compute unique values for (X, X, Z, Z). From the constraint matrix point of
view, we apply extrinsic constraint at the four knots a, b, c, d, and allow the
constraint to propagate throughout the matrix. If constraint propagates to (X,
176 Constraint theory
. .
X, Z, Z) and no other variable is over-constrained, then the trial set (a, b, c,
d) is. a describing
. set since it constrains the trajectory just as completely as
(X, X, Z, Z).

Figure 8-9. Final Constraint Matrix.

Constraint propagation through a constraint matrix follows very simple

rules which can be easily mechanized on a computer. When constraint is
transmitted to a variable via one of its edges, it is transmitted through the
variable via all its other edges. When constraint is transmitted to a relation
via all but one of its edges, it is transmitted through the relation via its
remaining edge. If this procedure does not completely fill the graph, then the
residue is investigated for BNSs.
Using this procedure, the describing sets can be found rapidly. It will be
more compact to list the measurable sets which are not describing sets. (See
Table 8-3.)
. . .Each
. . . .of the remaining
.. . 99 sets of four measurables, even the unlikely
XZrS, ZrSV, and ZrSd, perfectly constrains the trajectory and is a describing
set.

8.4 THE DEFLECTION OF AN EARTH-

THREATENING ASTEROID EMPLOYING
MASS DRIVERS

Since the early 1990’s there has been an increasing awareness of the risk
to earth by a strike from a large comet or asteroid. It is estimated that if such
a near-earth object (NEO) of 1 km or larger struck the earth it would disrupt
our global ecology and cause a billion human casualties. Much has been
8. Examples of Constraint Theory 177

accomplished regarding the detection of these NEOs; the number of

catalogued NEO’s has grown from under 5% to over 50% of their estimated
population from the early 90’s to the early 00’s.

Table 8-3. The set that are non-describing sets

These sets overconstrain velocity:

. . .. ..
XZZV XZSV
.. XrVd
. . .
XZZd XZSd XSVd
. . . . ..
XZVd XZSV
.. . XSVd
.
..
XZrV XZSd ZrVd
.. .. ..
XZrd XZVd ZrVd
. .. . .
XZrV ZZVd ZSVd
... . ..
XZrd XrvD ZSVd

These sets overconstrain position:

. .
XZrS ZrrS ZrSV
. .
ZZrS ZrSS ZrSd

However, much less attention has been paid to the deflection or intercept
of these threats to human civilization. The methods which provide the
highest energy density delivered to the threatening object are nuclear. But
there is great resistance to testing and putting nuclear devices into space, the
bomb may split the NEO in unpredictable ways, it may pollute the
atmosphere, and the coupling of the energy to the NEO is critically
dependent on the NEO’s structure and composition.
The most robust and lowest technical risk approach is to employ
traditional chemical propulsion to the NEO by attaching a conventional
rocket to the object and transfer orbit-changing momentum. But the cost of
this approach is enormous; not only must the chemical energy be delivered
to the NEO, the reaction mass must be transported there also. Imagine a
Saturn V rocket engine firing continuously for a year or more!
A concept which has received far less attention is to employ a mass
driver – essentially a linear motor which converts electrical energy into
kinetic energy. Once a mass driver is transported and installed on the NEO,
the required energy can be obtained with a solar collector and the required
reaction mass can be obtained from the NEO itself. The following analysis
178 Constraint theory

employs constraint theory to perform trade studies of mass driver designs for
NEO deflection.

Figure 8-10. Schematic of Mass Driver installation on asteroid.

Refer to Figure 8-10 for a schematic of the mass driver installation on the
NEO. The NEO, or asteroid, has mass M and diameter da. A solar collector
of diameter ds collects energy for the mass driver. The mass driver is a long
coil which propels mass from the NEO in buckets with a velocity, v. The
8. Examples of Constraint Theory 179

action of the mass driver ejecta causes a reaction by the asteroid, moving it
in the opposite direction with velocity, V. In order to miss the earth within a
time-to-go, tgo, the NEO must move a distance S – normally considered to be
the earth’s radius. (Generally, the optimum application of force for orbit
changing is parallel to the asteroid’s velocity vector. However this analysis
stresses the case of final approaches with short time-to-go and small angles
between the orbits of earth and asteroid.)
The kinematic relations can be captured by the bipartite graph in Figure
8-11. The variables are:

S= distance moved perpendicular to the asteroid trajectory

a= acceleration of asteroid
t= time
Vf= asteroid final velocity

and the relations are:

s=f1(a,t)=(1/2)at2 Vf=f2(a,t)=at

Figure 8-11. Bipartite Graph for initial kinematic model.

We can now determine whether computational requests can be made of

this simple model:
The computational request, v=f3(a,s) is allowable and is: v = 2as
which is a quite familiar formula. The computational request, v=f4(s,t) is also
allowable and it is: Vf=2s/t, but in the memory of the author, it was not
familiar. He does not recall ever seeing this formula, and doesn’t remember
that, in a system undergoing constant acceleration with zero initial velocity,
the final velocity is twice the average velocity. This was an unexpected
180 Constraint theory

bonus of this simple application of Constraint Theory. The value of the f4

relation is that we can easily compute the total energy requirement to deflect
the NEO from impact with the Earth, for any given time-to-go.
Let us now expand the model from kinematics to dynamics and write:

Total energy imparted to the asteroid = Er=(1/2) MVf2=2M(Vave)2

The energy available from the solar collector, Es=AFEtgo, where:

A=the solar collector area

F=the solar flux density at that distance from the sun
E=the flux-to-electrical energy conversion efficiency
tgo = time-to-go

Expanding the bipartite graph to include these variables and relations

yields Figure 8-12.

Figure 8-12. NEO Kinematics, NEO energy required and solar energy available.

It would be tempting to assume that all the energy collected by the solar
array is available to provide the asteroid with the required energy, that is, the
relation designated as “*” in Figure 8-12 is simply: Er=Es. Unfortunately,
this is far from correct, since both energy and momentum relations must be
satisfied:
8. Examples of Constraint Theory 181

Energy: Es=(1/2)MVf2+(1/2)mv2
Momentum: MVf=mv

Define a new variable, k=v/V=M/v – that is, the ratio of asteroid mass to
ejecta mass, or equivalently, the ratio of ejecta velocity to asteroid velocity.
Substituting k into the energy equation yields:

(Asteroid kinetic energy/ejecta kinetic energy)=1/k

In other words, only 1/(1+k) of the available solar energy couples to the
asteroid, the remainder accelerates the ejecta.
Paradoxically, the situation gets worse as the ejecta velocity increases
and the time-to-go increases. These results run counter to the “general
wisdom” that high ejecta velocity is good – after all, much hard work has
been expended in driving up the exhaust velocity of rocket engines and mass
drivers – and that the more time we have for deflection, the better off we are.
The general wisdom is still correct in their original domains and problem
statements, but does not precisely apply to the issues of asteroid deflection.
All these results can be summarized on a single chart, Figure 8-13.
Following the standard conventions employed to compare NEO deflection
concepts, the two fundamental independent variables are time-to-go and
asteroid diameter. The dependent variable is the diameter of the solar
collector, ds. At the center of the chart it can be seen that a solar collector of
only 100 meter diameter is required to deflect a 1 kilometer asteroid if one
year of time-to-go is available. The chart also presents an additional relation:
the locus of ds=da. On this locus, the solar collector is the same size as the
asteroid and the time-to-go becomes the dependent variable.
In summary, although the number of variables of this model was not
exceptionally high, the constraint theory methodology provided useful
kinematic and dynamic insight, as well as managing the computational flow
through a confusing set of equations.
182 Constraint theory

Figure 8-13. Results of the mass-driver analysis.

Chapter 9 MANAGER AND ANALYST MEET
AGAIN

Gists and Schizophrenia

The analyst and manager met again to celebrate their newfound

friendship and camaraderie.
“Would it be fair to ask if there is an overarching perspective to this
book?” asked the manager, still a little overwhelmed.
“Not really fair,” replied the analyst, “but I’ll try:”

THE GIST OF IT ALL

Technology’s ratchet is forcing systems into ever higher complexity.
Our greatest hope in managing complexity is math modeling.
But these models’ dimensionality is incomprehensible to any human.
By performing computations on the model, we can optimize designs as
well as understand the systems’ performance and phenomena.
But most of these computations are not allowable,
and it gets far worse with increasing dimension.
Computations are not allowable because they are not well-posed.
Allowable computations can only be made on a consistent model and
they must possess a proper constraint flow.
The kernel of constraint in a model is the basic nodal square (BNS)
and overlapping BNS’s cause inconsistency and overconstraint.
The topological properties of the model’s bipartite graph metamodel
provide practical clues for locating (and reconciling) BNS’s.
Bipartite graphs are most useful for regular relations.

“Alternatively, another summary from a somewhat different viewpoint is

presented in Figure 9-1,” the analyst contributed.

© Springer International Publishing AG 2017

G.J. Friedman, P. Phan, Constraint Theory, IFSR International Series on Systems
Science and Engineering, DOI 10.1007/978-3-319-54792-3_9
184 Constraint theory

“These are not very brief ‘gists,’” complained the manager. “Haven’t you
heard the saying: ‘That which is good and short is doubly good’?”
“Yes, I have,” agreed the analyst, “but brevity depends on a shared
experience – which we didn’t have originally – and a deep understanding of
the language we’re using. Normally, the richer the language, the briefer the
meaningful messages can become. If nothing else, this book has introduced
you to new directions of mathematics, which is a special type of language.”
“Speaking of language, it can reasonably be argued that constraint theory
sits on a bridge between mathematics and qualitative language. All
languages have built into their grammatical structure the primitive concepts
of “none,” “one” and “many.” – I have no horse, I have one horse, I have
many horses. The rhythm and pivotal concepts of constraint theory follow in
analogous ways:

Regarding paths between parts of a bipartite graph:

• no paths: the parts are in disjoint connected components
• one path: the parts are connected by a tree-like structure
• many paths: the parts are within a simple circuit or a circuit cluster
Regarding circuit rank, c(G):
• c(G) = 0: no circuits exist in G
• c(G) = 1: one independent circuit must exist in G
• c(G) > 1: multiple independent circuits must exist in G
Regarding constraint potential, p(G):
• p(G) ≥ 0: BNS must exist in G.

“But it’s regrettable, isn’t it, that the bipartite graph couldn’t be as useful
for discrete and interval relations as it is for regular relations?”
“I do wish it would be more broadly applicable,” agreed the analyst, but
can you name me any branch of mathematics that is equally useful over all
applications? Math tends to be very specialized into specific application
domains. Consider the vast body of math that analyzes linear systems,
despite the fact that most real-world systems are nonlinear. On the bright
side, regular relations presently represent the most common type of relation
used in math modeling, so if it’s limited to a domain, at least it’s the most
significant domain.”
“As I promised in Chapter One,” continued the analyst, “Constraint Theory
really occupies just a simple portion of all mathematics. Consider John
Barrow’s ‘Structure of Modern Mathematics’ [16] shown in Figure 9-1,
where Constraint Theory is represented only in the simplest lower right-hand
corner. We made substantial use of bipartite graphs which are with graph
theory and closely related to hypergraphs since the constraint matrix of a
9. Manager and Analyst Meet Again 185

bipartite graph is essentially identical to the incidence matrix of a

hypergraph [17].”
“There is another bright side,” the analyst added. “This book is only the
first venture into a new world of applied mathematics, and I’m hopeful that
far more research will be accomplished in the future.”
“I hope so too. We’ve been through a lot together,” commented the
manager, “more than I had expected at the beginning.”
“Yes,” agreed the analyst, “our interactions seemed to take on a life of
their own, far exceeding the expectations from the original seeds of our
discussions. In a way, that’s a testimony to the richness and importance of
the area.”
“Speaking of richness, you seem to have had a great diversity of prior
experience to provide the basis for such an invention as constraint theory,”
complemented the manager.
“Well yes, I suppose I have,” admitted the analyst. “I had the good
fortune to have worked in the fields of systems, mechanical, electrical and
civil engineering, applied mathematics, computer science, control systems,
surveying, soil mechanics, structural design, rockets, guidance and control,
instrumentation, steam power plants, environmental systems, pressure
sensors, accelerometers, altimeters, air data computers, missile fueling
systems, stellar inertial navigation, electro-optical sensors, acoustic terminal
homing sensors, optical gyroscopes, electronic countermeasures, airborne
radar, artificial intelligence, and low observables technology, to name a
few.”
“Very impressive,” beamed the manager. “I too have had the good
fortune to be involved with a great variety of programs, including the V-2
ballistic missile, the Redstone, Jupiter and Thor missiles, the Baldwin Hills
reservoir, the San Fernando Valley Steam Electric Plant, the Skybolt air-
launched ballistic missile, the F-5 fighter aircraft, the Advanced medium
range air-to-air missile, the brilliant antitank submunition and the B-2 stealth
bomber. In fact, I believe I’m the only person to have worked on both the V-
2 and the B-2. Additionally, I have contributed to workshops for all branches
of the US Department of Defense, NASA, DOE, NSF and NATO.”
186 Constraint theory

Figure 9-1. The Structure of Modern Mathematics

(From: John D. Barrow, The Book of Nothing, Pantheon Books, 2000, p152)
9. Manager and Analyst Meet Again 187

“Do you think it a coincidence that the list of technologies are generally
relevant to the list of programs?” asked the analyst.
“Not if you realize that we’re really the same guy,” both observed. “And
the dialogues we used were merely an expository mechanism to describe a
new field and illuminate unfamiliar concepts. We were attempting – without
daring to be compared to them – to use dialogue in the manner of Plato,
Galileo and even Mark Twain in his philosophical work, ‘What is Man?’

Roses are red,

violets are blue.
I am schizophrenic,
and so am I.
Appendix A COMPUTATIONAL REQUEST
DISAPPOINTMENTS; RESULTS OF
THE USC ALLOWABILITY
PROJECT

In order to construct a useful and trustworthy mathematical model, one

must gather authoritative data and relations, ask the advice of experts in the
domains of each submodel, assure that all uses of each variable agrees
semantically, and laboriously imbed the structure into a computer. After all
this painstaking work, it would be hoped that one could derive new
understandings across the many domains of the submodels and exercise wide
liberty in making computational requests of the math model.
It has been said of Carl Friedrich Gauss [19] that he “achieves greatness
in his work not through deep, abstract mathematical thinking, but rather
through an incredible vision of how the various quantities in the problem are
related, a vision that guides him through extraordinary computations that
others would likely abandon as futile.”
As was thoroughly discussed in the main text, a crucial impediment to
the allowability of computational requests is that the model is not consistent
due to intrinsic overconstraint. Clearly, no computational requests are
allowable on an inconsistent model. If this inconsistency is removed by
eliminating an appropriate number of submodels there still may exist
intrinsic basic nodal squares (BNSs) which apply constraint to all their
relevant variables.
Although the model has been rendered consistent, none of the variables
relevant to the BNSs can participate as an independent variable, dependent

© Springer International Publishing AG 2017

G.J. Friedman, P. Phan, Constraint Theory, IFSR International Series on Systems
Science and Engineering, DOI 10.1007/978-3-319-54792-3
190 Constraint theory

variable, or variable held constant. This significantly reduces the number of

allowable computational requests.
Unfortunately, even if all the BNSs were to be removed – making the
model completely free of intrinsic overconstraint and point constraint, the
likelihood of any given computational request being allowable is still quite
low. Examples may clarify this disappointing fact.
Consider the trivial case for N=1; the model has
only one node and K knots. There are 2K possible
computational requests, but only one allowable
computation (identical to the original contributing
submodel.) Thus, the probability of allowability,
p(a) equals 1/2K = 2-K. If K=4, then p(a) = 1/16.
Next, consider a model with N=2 having a tree
structure and with K1 variables relevant to N1 and
K2 variables relevant to N2. Since one of the
variables is shared, the total number of variables is
K1+K2-1 and the total number of possible
computational requests = 2 K1 + K 2 1 . The total
allowable computational requests are the two
contributing submodels, plus K1-1, plus K2-1,
which merely sums to K1+K2. Thus, for this case,
( K1 + K 2 )
p(a ) =
2 K1 + K 2 1
If K1=K2=4, then p(a) = 1/16 again.
If the N=2 model has a circuit structure with a
single circuit, then two of the knots are shared and
the total knots in the model = K1+K2-2. The total
number of allowable computations here is only 4,
4
so in this case, p ( a ) = K1 + K 2  2
, and if the K’s =
2
4, p(a) = 1/16 (again!!)
This remarkable coincidence of p(a) = 1/16 does not extend to more
complex models. Continuing this mode of analysis becomes exceedingly
complex with higher N and K; the number of possible computational
requests for a simple model of only N=10 already exceeds 1000. Thus let us
skip ahead a bit to the region of K=6,8 and N=4,6 which was employed in
Chapter 1 to demonstrate how even these low dimensions can exhibit
interesting behavior. We will employ the results of the USC Computational
Allowability project [1] wherein three engineering graduate students
employed the techniques of Chapters 4 and 5 and exhaustively performed
allowability analyses on the models shown below.
Appendix A 191

In the analyses to follow, a key “cognitive limitation” was placed on the

total number of computational requests to be considered: only 2, 3 and 4
dimensional requests were analyzed. (For example, x=f(y,z) is defined as a
three dimensional request.) This limitation was employed for two reasons:

(a) most importantly, dimensions of 5 and higher require complex

“carpet plots” to understand and are very difficult for most people
to perceive, and
(b) with the constraint potential of the examples equaling -2, it is
highly unlikely that computational requests of dimension higher
than 4 would be allowable anyway.
192 Constraint theory

Figure A-1. Bipartite Graph Topological Structures Examined.

Thus, instead of examining all 2K conceivable computational requests we

K! K! K!
only examine + + of them. For K=6,
2! ( K  2)! 3! ( K  3)! 4! ( K  4)!
only 50 out of the total possible of 64 were examined, for K=8, only 154 out
of the total possible of 256 were examined.
Figure A-1 summarizes the ten bipartite graph topological structures
which the team examined. Table A-1 summarizes the results of the 1228
different computational requests which were made on these ten topologies.
Appendix A 193

Table A-1. Summary of Computational Request Allowability Analysis

Struct- K N E C B # # %
ure Examined Allowable Allowable
A 6 4 12 3 0 50 13 26%
B 6 4 11 2 0 50 12 24
C 6 4 10 1 0 50 9 18
D 8 6 18 5 1 154 6 4
E 8 6 18 5 0 154 23 15
F 8 6 17 4 1 154 6 4
G 8 6 16 3 1 154 6 4
H 8 6 15 2 1 154 6 4
J 8 6 14 1 1 154 3 2
K 8 6 13 0 0 154 6 4
TOTALS 1228 90 7.3%

LEGEND:

Structure: Bipartite graph topological structure shown in Figure A-1

K: Number of Knots, or variables
N: Number of Nodes, or relations
E: Number of Edges, or relevancies
C: Number of independent circuits, or circuit rank of the graph
B: Number of Basic Nodal Squares
# Examined: Number of computational requests which were analyzed
# Allowable: Number of requests which were computationally allowable
%: Percent of the requests which were allowable

The conclusion is rather dramatic: the overall probability of allowability

was only 7.3%. If the analysis had been extended to all 2K computational
requests, the percent allowability would have been much less.
Not surprisingly, those topologies which were devoid of BNSs had the
highest allowabilities, but even then it was only about 20%. Therefore it
should be abundantly clear that an allowability analysis must be performed
on every computational request before the programmers jam it into the
machines.
194 Constraint theory

Worse yet, it is the author’s conjecture that – as the model dimensionality

increases to 100, 1000 and beyond – the absolute number of allowable
computational requests will increase, but the percentage of allowability
based on the total number of possible requests will decrease as illustrated in
Figures A-2 and A-3.
Appendix A 195

Figure A-2. Typical values of A&T for K = 10, 100, 1000

196 Constraint theory

Figure A-3. A, T, & A/T vs. K

Appendix B GRAPH THEORY OVERVIEW

Why was the Bipartite Graph Chosen?

A general graph, or graph, consists of a set of vertices, some pairs of

which are connected by a set of arcs, or edges.
A bipartite graph (BPG) is a graph such that its vertices can be
decomposed into two disjoint sets, X and Y, and its edges only connect
vertices in set X to vertices in set Y.
Any property of a general graph is also a property of a bipartite graph,
but not vice versa (Table B-1). For example, both the general and bipartite
graphs have connected components, trees and circuit structures. However only
the bipartite graph has the property of constraint potential and the structure
of basic nodal squares.
Graphs have been used for many decades as mathematical metamodels.
Mason, et al [20], applied signal flow graphs (SFG) to the analysis of
complex systems where the functions were linearly separable. Friedman [21]
applied inverse signal flow graphs (SFG) to a broader class of systems.
Figure B-1 provides a comparison between how the signal flow graph, the
inverse signal flow graph and the bipartite graph assign portions of their
structure to the mathematical concepts of variable, relation and relevancy.

Figure B-1. The definitions of Three Types of Graph Theory Metamodels.

© Springer International Publishing AG 2017

G.J. Friedman, P. Phan, Constraint Theory, IFSR International Series on Systems
Science and Engineering, DOI 10.1007/978-3-319-54792-3
198 Constraint theory

For the purposes of Constraint Theory, the bipartite graph metamodel

was chosen over the other alternatives because it is more general
topologically and functionally, as well as the fact that it can represent both
the model and the computations performed on it.
Figure B-2 demonstrates the greater generality of the BPG by showing
that any SFG can be represented by a BPG, but most BPGs cannot be
represented by a SFG or an SFG. Figure B-3 demonstrates the greater
generality of the BPG by showing that the SFG is limited only to linearly
separable functions where the BPG can represent any function.

Table B-1. Graph Theory Overview.

General Graph: A set of vertices, some pairs of which are

connected by a set of arcs (or edges)

Bipartite Graph: The vertices have two species and the

arcs connect only two different species

Any property of a general graph is also a property of a

bipartite graph, but not vice versa

Graphs have been used for many decades to describe

computations in complex networks; i.e.; Signal Flow Graphs

The Bipartite Graph was chosen over the Signal Flow

Graph because it is more general topologically and
functionally and because it represents both model and
computation
Appendix B 199

Figure B-2. BPG Characteristics (1).

200 Constraint theory

Figure B-3. BPG Characteristics (2).

Appendix C THE LOGIC OF “IF” AND “IF AND
ONLY IF”

Theorem 1 states that inconsistency implies computational non-

allowability. More compactly, we can state: “nonallowability if
inconsistent.” If the reverse, “inconsistent if nonallowable” were true then
we could state, “nonallowability if and only if inconsistent.” (For even
increased compactness, “if and only if” is often contracted to “iff”.)
Actually, Theorem 1 is an “if” statement, not an “iff” statement. If a
model is inconsistent, then no computation on it is permitted. However, the
reverse is not always true because there are many other possible
computability problems even with consistent models. Actually, the vast
majority of the theorems presented in Chapters 3 and 4 are “if” theorems, not
“iff” ones. Another example of a one-way “if” theorem is Theorem 25 which
states that a BNS must always lie within a circuit cluster. Many students then
wrongfully assumed that – whenever they identified a circuit cluster—they
have found a BNS.
On the other hand, Theorems 13 and 14, dealing with the topological
properties of trees and circuits, are indeed two-way “iff” theorems.
In order to appreciate the logic of “if” and “iff” statements more deeply,
let us examine Figure C-1 which presents a Venn Diagram that displays the
events of consistency, allowability and their interactions.

© Springer International Publishing AG 2017

G.J. Friedman, P. Phan, Constraint Theory, IFSR International Series on Systems
Science and Engineering, DOI 10.1007/978-3-319-54792-3
202 Constraint theory

Figure C-1. Logical Analysis of Theorem 1.

Appendix D ALGEBRAIC STRUCTURES

DEFINITIONS AND PROPERTIES OF GENERAL VECTOR SPACES

(Gross & Yellen, 2006, pp. 687 & 690; Anton, 1977, p. 127)

Definition D-1: A binary operation * on a non-empty set A is a function

f: A × A Æ A, given by f [ (a, b) ] = a * b. The set A together with a binary
operation * is denoted ( A, * ).

Definition D-2: The binary operation * on set A is said to be associative

if for a, b, c  A, ( a * b ) * c = a * ( b * c ), and commutative if for a, b 
A, a * b = b * a.

Definition D-3: An element e  A is an identity element in ( A, * ) if for

all a  A, a * e = e * a = a. And for all a  A, an element a’  A is an
inverse of a in ( A, * ) if a * a’ = a’ * a = e.

Definition D-4: A group G = ( G, * ) is a non-empty set G and a binary

operation * that satisfy the following conditions:
¾ The operation * is associative.
¾ G has an identify element.
¾ Each g  G has a unique inverse in ( G, * ), denoted g-1.

Definition D-5: An abelian group is a group whose operation is

commutative. In an abelian group, the binary operation is commonly denoted
“+” and called sum.

Definition D-6: A field F = ( F, +, • ) is a set F together with two

operations, + and • (generically called addition and multiplication), that meet
the following conditions:
¾ ( F, + ) is an abelian group.
¾ ( F – {0}, • ) is an abelian group, where 0 is the additive identity.
© Springer International Publishing AG 2017
G.J. Friedman, P. Phan, Constraint Theory, IFSR International Series on Systems
Science and Engineering, DOI 10.1007/978-3-319-54792-3
204 Constraint theory

¾ a • ( b + c ) = ( a • b + a • c ) and ( a + b ) • c = ( a • c ) + ( b • c ).

Definition D-7: The finite field GF(2) consists of the set Z2 = { 0, 1 }

together with the mod-2 operations +2 and •. Thus:
¾ 0 +2 0 = 1 +2 1 = 0 (this operation is identical to the logical operation
of exclusive OR)
¾ 0 +2 1 = 1 +2 0 = 1
¾ 0•0=1•0=0•1=0
¾ 1•1=1

Definition D-8: A vector space over a field (of scalars) F is a set V (of
vectors) together with an operation + on V and a mapping, called scalar
multiplication from the Cartesian product F × V to V ( (a, v) Æ av ), such
that the following conditions are satisfied for all scalars a, b  F and all
vectors v, w  V:
¾ ( V, + ) is an abelian group, where the notation “+” is being used to
denote both addition of scalars in field F and addition of vectors in
set V.
¾ ( a • b ) v = a ( bv )
¾ ( a + b ) v = av + bv
¾ a ( v + w ) v = av + aw
¾ ev = v, where e is the multiplicative identity of field F.
BINARY SET OPERATIONS (Gross & Yellen, 2006, pp. 197-198)
Appendix D 205

Definition D-9: Let s1, s2, . . . , sn be any sequence of objects, and let A
be a subset of S = { s1, s2, . . . , sn }. The characteristic vector of subset A,
denoted charvec ( A ), is the n-tuple whose jth component is 1 if sj  A, and 0
otherwise.

In the model graph G of Figure D-1 below, EG = { e1, e2, e3, e4, e5, e6, e7 }.
Cycle A has as its edge set EA = { e1, e2, e3, e4 }, and cycle B has as its
edge set EB = { e3, e5, e6, e7 }, both of which are subsets of EG. By Definition
D-9 above, charvec ( EA ) = ( 1, 1, 1, 1, 0, 0, 0 ), and charvec ( EB ) = ( 0, 0,
1, 0, 1, 1, 1 ).

Characteristic Edges
Cycle Vector of
Edge Set e1 e2 e3 e4 e5 e6 e7

A charvec (EA) 1 1 1 1 0 0 0

B charvec (EB) 0 0 1 0 1 1 1

Figure D-1. Cycles and characteristic vectors of their edge sets

206 Constraint theory

Definition D-10: For graphs A and B, the union of their edge sets EA and
EB is the set of all edges which are either in EA or in EB, or both.
Symbolically, e  EA U EB if e  EA or e  EB.

In Figure D-2 below, edge set EA = { e1, e2, e3, e4 } and edge set EB = {
e3, e5, e6, e7 }. Thus, their union set EA U EB = { e1, e2, e3, e4, e5, e6, e7 }. Note
that charvec(EA U EB) is formed by combining the respective components of
charvec(EA) and charvec(EB) with the bitwise inclusive OR operator. In
C/C++, this operator has the syntactical symbol of “ | ”, e.g. ( 0 | 0 ) = 0; and
( 0 | 1 ) = ( 1 | 0 ) = ( 1 | 1 ) = 1. In this case, charvec ( EA U EB ) = ( 1, 1, 1,
1, 1, 1, 1 ).

Characteristic Edges
Cycle
Vectors e1 e2 e3 e4 e5 e6 e7
A charvec (EA) 1 1 1 1 0 0 0
B charvec (EB) 0 0 1 0 1 1 1
charvec ( EA U EB ) 1 1 1 1 1 1 1

Figure D-2. Union of edge sets and its characteristic vector

Appendix D 207

Definition D-11: For graphs A and B, the intersection of their edge sets
EA and EB is the set of all edges which are in both EA and EB. Symbolically:
e  EA  EB if: e  EA and e  EB .

In Figure D-3 below, edge set EA = { e1, e2, e3, e4 } and edge set EB = {
e3, e5, e6, e7 }. Thus, their intersection set EA  EB = { e3 }. Note that
charvec ( EA  EB ) is formed by combining the respective components of
charvec ( EA ) and charvec ( EB ) with the logical operator AND where 1
means true, and 0 means false. Symbolically, ( 0 AND 0 ) = ( 0 AND 1 ) = ( 1
AND 0 ) = 0; ( 1 AND 1 ) = 1. In this case, charvec ( EA  EB ) = ( 0, 0, 1, 0,
0, 0, 0).

Characteristic Edges
Cycle
Vectors e1 e2 e3 e4 e5 e6 e7

A charvec (EA) 1 1 1 1 0 0 0
B charvec (EB) 0 0 1 0 1 1 1
charvec ( EA ∩ EB ) 0 0 1 0 0 0 0

Figure D-3. Intersection of edge sets and its characteristic vector

208 Constraint theory

Definition D-12: For graphs A and B, the difference EA  EB is the set of

all edges which are in EA but not in EB. Note that the difference is not
necessarily commutative, i.e. ( EA  EB )  (EB – EA ).

As demonstrated in Figure D-4 below, edge set EA = { e1, e2, e3, e4 } and
edge set EB = { e3, e5, e6, e7 }, therefore the difference set EA  EB = { e1, e2,
e4 }. However, the difference set EB  EA = { e5, e6, e7 }.

Figure D-4. Difference set ( EA  EB )  difference set (EB – EA )

Appendix D 209

Definition D-13: For a graph G, let WE(G) denote the set of all subsets of
EG, i.e. the power set of all edges in G. The ring sum of two elements E1 , E2
 WE(G) is defined as:
E1 † E2 = ( E1  E2 ) U ( E2  E1 )
Figure D-5 below illustrates graphically an application of ring sum in
which two cycles A and B are combined to form a new cycle. Note that
charvec ( EA † EB ) is formed by combining the respective components of
charvec ( EA ) and charvec ( EB ) with the bitwise exclusive OR operator. In
C/C++, this operator has the syntactical symbol of “ ^ ”, e.g. ( 0 ^ 0 ) = ( 1 ^
1 ) = 0; and ( 0 ^ 1 ) = ( 1 ^ 0 ) = 1.

Figure D-5. Ring sum EA † EB = ( EA  EB ) U ( EB  EA ) = ( 1, 1,

0, 1, 1, 1, 1 )
References

1. Leila Habibibabadi, Kathya Zamora-Diaz, Elliott Morgan,

Computational Allowability on Bipartite Graphs, projects for the
graduate course, Systems Engineering, ISE 541, University of
Southern California, 1998, 1999.

2. George Friedman, Constraint Theory Applied to Mathematical

Model Consistency and Computational Allowability, PhD
Dissertation, University of California at Los Angeles, 1967.
University Microfilms, Inc, Ann Arbor, MI

3. George Friedman, Constraint Theory, an Overview, International

Journal of Systems Science, 1976, Vol 7, No 10, pp 1113-1151.

4. Edgar Palmer, Graphical Evolution, John Wiley and Sons, New

York, 1985

5. Roald Hoffmann and Shira Schmidt, Old Wine, New Flasks, W. H.

Freeman and Company, New York, 1997

6. Norbert Wiener, A Simplification of the Logic of Relations,

Proceedings of the Cambridge Philosophical Society, Vol 17, pp
387-390, 1914

7. Claude E. Shannon, The Theory and Design of Linear Differential

Equation Machines, National Defense Research Committee Report,
Princeton University, Jan 1942

8. George Gamow, One, Two, Three...Infinity, Viking Press, New

York, 1947

9. Bourbaki, Elements de Mathematique; Theorie des Ensembles,

ASEI 1141; Hermann & Cie, Paris, 3me. ed., 1958

© Springer International Publishing AG 2017

G.J. Friedman, P. Phan, Constraint Theory, IFSR International Series on Systems
Science and Engineering, DOI 10.1007/978-3-319-54792-3
212 Constraint theory

10. W. Ross Ashby, The Set Theory of Mechanism and Homeostasis,

General Systems, Yearbook of the Society for General Systems
Research Volume IX, Bedford, MA 1964

11. Edwin A. Abbott, Flatland, A Romance of Many Dimensions,

Harper Collins, 1994; (Originally published circa 1880)

12. Dionys Burger, Sphereland, A Fantasy about Curved Spaces and an

Expanding Universe, Harper Collins, New York, NY, 1994

13. K. Devlin, The Math Gene, Basic Books, New York, NY, 2000

14. P. J. Hall, Representation Theorem in Set Theory, J. Lond. Math.

Soc, 10, 26. 1934

15. Antonin Svoboda, “An Algorithm for Solving Boolean Equations,”

IEEE Trans. Electron. Comput, Oct 1963, pp 557-559.

16. John Barrow, The Book of Nothing, Pantheon Books, 2000

17. Claude Berge, Hypergraphs, North Holland, 1989

18. T. Gehrels, Hazards due to Comets and Asteroids, University of

Arizona Press, 1995

19. Ivars Peterson, Math Trek, Science News Online, April 17, 1999

20. S. J. Mason, Feedback Theory -- Some Properties of Signal Flow

Graphs Proceedings of the IRE, Sept 1953, vol 41, pp 1144-56

21. G. J. Friedman, Constraint Algebra -- a Supervisory Programming

Technique and a Cognitive Process, IEEE Transactions on Military
Electronics, April 1963, pp 163- 167

22. Phan, Phan (2011). Expanding constraint theory to determine well-

posedness of multi-dimensional math models (Doctoral dissertation).
University of Southern California.

23. Tarjan, Robert Endre (1972). Depth-first search and linear graph
algorithms. SIAM Journal on Computing, Vol. 2, pp. 146-160.
References 213

24. Cormen, T. H., C. E. Leiserson, R. L. Rivest and C. Stein (2001).

Introduction to Algorithms. Cambridge, MA: The MIT Press.

25. Dechter, Rina (2003). Constraint Processing. San Francisco, CA:

Morgan Kaufmann Publishers.

26. Gross, Jonathan L. and Jay Yellen (2006). Graph Theory and Its
Applications. Boca Raton, FL: Chapman & Hall/CRC.

27. Tarjan, Robert Endre (1974, March). A note on finding the bridges
of a graph. Information Processing Letters, Vol. 2, No. 6, pp. 160-
161.

28. Shirey, R. W. (1969). Implementation and analysis of efficient graph

planarity testing algorithms (Doctoral dissertation). University of
Wisconsin, Madison, WI.

29. Anton, Howard (1977). Elementary Linear Algebra. New York, NY:
John Wiley & Sons, Inc.
Index

Abbot, 40 Boolean
abelian group, 203 algebra, 137
acceleration of asteroid, 179 equations, 138
accuracy, 167 relations, 138
adjacent, 72 Boolean equations, 141, 147
adjacent circuits, 89 Boolean relations, 141
aerospace industry, 163 Bourbaki, 31
Air Force Research Laboratories (AFRL), Brahe, 26
164 brute force, 124
airborne radar, 185 brute force procedure, 89, 90
aircraft, 171 budgetary constraints, 164
aircraft kinematics, 171 Burger, 40
aircraft velocity, 171 cardinals, 137
allowability, 147 Cartesian coordinates, 30
space, 152 charvec (characteristic vector)
subset, 152 definition, 110
allowable, 7, 40, 124, 156 chemical energy, 177
space, 138 chemical propulsion, 177
ambiguity, 141 circuit cluster, 56, 72, 79, 107, 131, 147,
analysis, 137 184
angle rate, 171 circuit rank, 75, 76, 78, 138, 184
applied mathematics, 185 Circuit Vector Addition, 75
articulation point, 97 circuit vector space, 75
artificial intelligence, 185 circuits, 105
asteroid, xiii, 163, 176, 179, 180 cognitive limitations, 124, 191
avionic equipment, 171 comet, 176
Barrow, 184 complex systems, xi
Basic Nodal Square, 65 computability, 46
biconnected component, 97 computational
binary operation *, 203 allowability, xii, 108
bipartite graph, xii, 8, 39, 46, 88, 103, flow, 141, 181
105, 108, 131, 157, 166, 168, 173, paths, 141
183, 184, 185, 197 request, xii, 66, 103, 105, 124, 156,
BNS, 65, 68, 79, 83, 105, 157, 173 179
detection of overlapping, 109 request disappointments, xiv, 189
kernels, 79 computational allowability, 103
location theorem, 157 Computer Assisted Engineering, 1
matrix, 111 computer program, xii
bombing problem, 171 computer science family, xi
bombing solution, 171 connected components, 54, 72, 184

© Springer International Publishing AG 2017

G.J. Friedman, P. Phan, Constraint Theory, IFSR International Series on Systems
Science and Engineering, DOI 10.1007/978-3-319-54792-3
216 Constraint theory

connectedness, 131 exclusive OR, 204, 209

connectedness algorithm, 72 exhaust velocity, 181
consistency, xii, 40, 46, 103, 107, 108, explicit expression, 147
124, 156, 166, 167, 168, 184 extension, 34
constrain (definition), 31 extrinsic constraint, 10, 103, 105
constraint domain, 114, 115, 118 FAMILY OF SUBMODELS, 35
constraint matrix, 28, 39, 46, 69, 72, 167, field, 203, 204
168 finite field, 204
constraint potential, 63, 76, 131, 138, forbidden, 149
145, 167 areas, 140
Constraint Theory, xii, 1, 19, 156, 164, inputs, 141
172 zone, 141
continuum relation, 58 fragmented understanding, 106
control systems, 185 free-falling bomb, 171
Cost as an independent variable, 164 Friedmann, 26
coupling of the energy, 177 full discrete relation, 145
crucial contract, 99 function map, 149
cut-point, 97 fundamental criterion, 173
cut-vertex, 97 Galileo, 187
Definitions summary, 157 Gamow, 27
deflection, 177, 181 Gauss, 189
degree of a vertex, 63 generalizability, 49, 132
Department of Defense, 163 generizability, 51
dependent variable, 181 global ecology, 176
depth-first search (DFS), 93 Goldilocks rule, 101, 103, 131
Descartes, 44 grammatical structure, 184
design rules, 107 graph, 39, 197
Devlin, 50, 61 graph theory, xiv, 20
DFS, 93, 97 group, 203
difference, 208 Hall, 71
dimensionality, 75 Hall’s Theorem, 156
limitations, 155 homomorphic, 39, 46
Diophantine equations, 28, 138 Hubble, 26
directly overlap, 109, 114, 115 hypercube, 141
disconnect, 104 hypergraphs, 184
disconnected components, 103 hyperspace, 29, 141
discrete, 137, 184 of relevant variables, 138
discrete relations, 57, 137, 138, 145 hyperspatial viewpoints, 140
disjoint sets, 157 identity element, 203
doily, 79 iff, 51, 201
dynamic analysis, 163 Imaginary numbers, 44
dynamic insight, 181 incidence matrix, 185
earth’s radius, 179 inclusive OR, 110, 206
edges, 157 inconsistency, 52, 105, 147, 168, 184
Einstein, 26 Independent paths, 72
ejecta, 179 independent variables, 105, 164
ejecta velocity, 181 inequality equations, 150
energy density, 177 Inequality relations, 150
energy equation, 181 integers, 138
Index 217

intercept, 177 near-earth object (NEO), 176

internal trees, 83, 96, 99 NEO deflection, 178
intersection, 34, 140 nested category, 79
interval, 137 Newton, 26
relations, 58, 184 nodal squares, 65, 174
intrinsic nodal twig, 81, 113
BNS, 98, 101, 110, 116 nodes, 39, 46, 107, 167
constraint, 10, 103 NS, 68, 88
constraint potential, 63, 101 nuclear devices, 177
point constraint, 138, 150 null set, 147
sources of constraint, 69 operations analysis, xiii, 163
inverse signal flow graphs, 197 optical gyroscopes, 185
irrational numbers, 152 optimization criteria, 107
Irrational numbers, 43 optimum
Kepler, 26 system design, 164
kernel of constraint, 61, 183 values, 168
kinematic, xiii, 163 orbit-changing momentum, 177
relations, 179 overconstraint, 7, 40, 64, 101, 105, 106,
state, 171 166, 168
knots, 39, 46 overlapping BNSs, 105
language, xii overlapping factor, 112
linear Palmer, 22
algebra, 75 paradoxes, 61
criterion, 150 parameters, 104
programming, 150 perfect discrete relation, 146
systems, 184 perfectly constrained, 64
local degree, 101 permutation, 107
locally universal, 63 Plato, 187
logical, 137 point constrain, 69, 138
design, 140 policy, 107
operator AND, 207 demands, 167
thrusts, 157 nodes, 168
mass drivers, 163, 177 pollute the atmosphere, 177
math model, 164 polynomial, 138
math phobia, 50 Postulate 1, 63
matrix, 138 Postulate 2, 105
measurable, 171 power set, 48, 54, 56
meta models, 173 product set, 30, 46
Metamath, 49 product space, 147
metamathematics, 50, 137 projection, 34, 152
meta-metamodels, 138 protolanguage, 49
metamodel, xii, xiv, 7, 46, 141, 145, 152, protomath, 61
198 protomathematics, 49
Miller, 3 pyramid, 152
model consistency, 175 quaternary variable, 147
model graph, 39 radar signature, 167
momentum, 180 range, 171
multi-dimensional models, xi, 25 range rate, 171
nanosecond examinations, 131 reaction, 179
218 Constraint theory

mass, 177 ternary, 149

reaction mass, 177 tetrahedron, 152
regular, 63 Theorem 1, 51
regular relations, xiii, 108, 137, 138, 141, Theorem 10, 63
184 Theorem 11, 66
relation, 31, 46 Theorem 12, 71
classes, 57 Theorem 13, 73
relevant, 34 Theorem 14, 75
knots, 103 Theorem 16, 81
relations, 9 Theorem 18, 82
variables, 138, 141, 172 Theorem 19, 82
reliability, 167 Theorem 2, 52
remainder matrix, 117, 119, 120 Theorem 20, 82
required energy, 180 Theorem 21, 82
resultant Theorem 22, 82
constraint domains, 69 Theorem 23, 83
constraint potential, 63, 101, 105 Theorem 24, 83
resultant BNS, 105 Theorem 25, 83
ring sum, 209 Theorem 26, 84
run free, 168 Theorem 3, 53
Saturn V, 177 Theorem 31, 138
schizophrenic, 187 Theorem 32, 143
self-referential loops, 61 Theorem 33, 147
semiperimeter, 75 Theorem 34, 147
separating vertex, 72 Theorem 35, 150
set, 28, 46 Theorem 36, 151
set theory, 20 Theorem 37, 151
Shannon, 22, 27 Theorem 4, 54
signal flow graphs, 197 Theorem 5, 54
simple circuit, 75 Theorem 6, 54
solar collector, 178 Theorem 7, 54
solar energy, 181 Theorem 8, 56
spanning tree, 93, 95, 96 Theorem 9, 56
squares, 65 Theorems summary, 159
stellar inertial navigation, 185 time-to-go, 181
subdimensional, 26 Toolkit, 61
sub-graphs, 99 topological
submodels, xi, 46, 52 properties, xiii
subset, 28 structure, 7
Svoboda, 137, 147 topology, 131
system criterion, 167 trade studies, 178
systems engineering, xi, 2 tradeoffs, xi, 14
T25 procedure, 89, 91 trajectory plane, 171
Talmud, 25 transitivity rule, 151
taxonomy of graph structures, 79 tree, 56, 72, 173
technical risk, 177 algorithm, 74
Technology’s ratchet, 183 structures, 103, 105, 184
Ten Easy Pieces, 81 tree-ness, 131
terminal node, 82, 113 tree-twig pruning algorithm, 75
Index 219

trimming the trees, 173 Veitch diagram, 138, 141

truth table, 138, 141 Venn diagram, 79, 139, 201
Twain, 187 vertex pairs, 79
twig node, 92, 108, 109 vertexes, 46
underconstraint, 7, 14, 40, 64, 101 view spaces, 124
union, 33, 140 von Neumann, 61
universal, 147 weapon delivery, 167
universal relation, 56 weight, 167
USC Allowability Project, 189 well-posedness, 22, 156
variable, 28 Wiener, 22, 27
vector space, 76