Contents

1 Introduction 1
1.1 What is a Model? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Reasons for Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2.1 Scientific Models that Describe . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2.2 Engineering Models that Predict . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2.3 Regulatory Models that Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 The Modeling Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3.1 The Objective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3.2 The Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3.3 The Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3.4 The Evaluation Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.3.5 The Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.4 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2 Analog Models 8
2.1 Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.1.1 Units . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.1.2 Exponents and Logarithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.2 Process Analogs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2.1 Fluid Flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2.2 Electricity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2.3 Heat . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2.4 Solutes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3 Scale Analogs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3.1 Dimensional Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3.2 Example: Fractal Scaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.4 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

3 Network Models 14
3.1 Percolation Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.1.1 Frequency Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.1.2 Markov Chains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.1.3 Discrete Event Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.1.4 Example: Extreme Values . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.2 Linear Systems Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.2.1 System Inputs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.2.2 Convolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.2.3 Deconvolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.2.4 Example: Unit Hydrographs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.3 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

i
CONTENTS ii

4 Statistical Models 19
4.1 Probability and Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
4.1.1 Sample Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
4.1.2 Higher Moments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
4.1.3 Variation of Sample Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
4.1.4 Hypothesis Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
4.2 Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
4.2.1 Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
4.2.2 Ordinary Least Squares . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
4.2.3 Transformations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
4.2.4 Factor Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
4.2.5 Model Updating . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
4.3 Time Series Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
4.3.1 Autoregressive (AR) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.3.2 Moving Average (MA) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.3.3 External Inputs (X) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.3.4 Differenced (D) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.3.5 Integrated (I) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.3.6 Example: Salt River Project . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.4 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

5 Differential Equations 29
5.1 System Specification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
5.1.1 Types of Differential Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
5.1.2 Initial and Boundary Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
5.1.3 Material Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
5.1.4 Example: Nash Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
5.2 Analytic Solution Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
5.2.1 Integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
5.2.2 Laplace Transform Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
5.3 Numerical Solution Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
5.3.1 Finite Difference Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
5.3.2 Finite Element Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
5.3.3 Boundary Element Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
5.3.4 Analytic Element Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
5.4 Example: Linear Reservoir . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
5.5 Example: Channel Routing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
5.5.1 General Routing Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
5.5.2 Nonlinear Outflow-Storage Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
5.5.3 Muskingum Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
5.5.4 Muskingum-Cunge Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
5.5.5 Muskingham Variants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
5.6 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

6 Vectors, Complex Variables, and Quaternions 39

6.1 Vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
6.1.1 Algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
6.1.2 Geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
6.1.3 Calculus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
6.1.4 Tensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
6.2 Complex Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
6.2.1 Algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
6.2.2 Geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
6.2.3 Calculus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
6.3 Quaternions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
6.3.1 Algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
6.3.2 Geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
CONTENTS iii

6.3.3 Calculus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
6.3.4 Relationship to Vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
6.4 The Dual . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
6.4.1 Problem Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
6.4.2 Dual Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
6.4.3 Cauchy Integral Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
6.4.4 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
6.5 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

7 Additional Reading 50
7.1 Mathematical References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
7.2 Management Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
7.3 Systems/Statistical Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
7.4 Watershed/Surface Water Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
7.5 Groundwater Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
List of Tables

1.1 The Basis of Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.2 General Systems Terms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

2.1 Metric System Units . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

2.2 Metric System Prefixes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.3 Physical analogs between processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

3.1 State Transition Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

3.2 Method for Determining Return Periods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.3 Example convolution problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

4.1 Statistical Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

4.2 Variation in Student Heights, cm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

5.1 General differentiation rules. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

5.2 General integration rules. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
5.3 Laplace transforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

6.1 Vector Identities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

6.2 Quaternion Multiplication Table . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
6.3 Quaternion Hyperbolic Identities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

iv
List of Figures

3.1 Triangular Unit Hydrograph. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

4.1 Venn Diagram of Events A and B . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

4.2 Normal Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

5.1 Linear Reservoir Definition Sketch. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

5.2 Cascade of linear reservoirs. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
5.3 Linear reservoir with a constant input. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
5.4 Sketch diagram for Muskingum storage. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

6.1 Diagram illustrating boundary flux components. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

v
Chapter 1
Introduction
1.1 What is a Model?
From birth, each of us must learn to negotiate the envi-
ronment that surrounds us. By experimentation we learn
how to interact with our environment - we learn hot from
cold, up from down. And once we gain confidence in our
knowledge, we learn how to control and manipulate our
surroundings. With maturity we learn that certain ap-
proaches are more effective than others - we learn right
from wrong.
Thus, our world view begins with basic observations -
data that we learn and remember. As data becomes more
abundant, we organize these data in ways that help us
remember details. The real world is a complex web of en-
tities and relationships. Once we have sufficient data, we
form relationships between entities. When we are young,
we may not understand the reason for these relationships.
But with education we begin to understand why these re-
lationships exist, and learn how to use these relationships
to control our environment. Finally, through experience
and training, we gain sufficient wisdom to use this knowl-
edge for our own, and perhaps society’s, betterment.
A model, then, is a representation of our environment -
the observations we collect, the relationships we infer, the
knowledge we gain, and the wisdom we draw upon to act.
From the very beginning, we are building mental models
of our environment, trying to make sense of our place on
Planet Earth, and what we can do to better deal with life’s
complexities.
When someone says they wish to build a model, what
they are really trying to do is capture the rules by which
a system operates. By explicitly stating these rules, they
wish to articulate a specific mental construct of their ob-
servations, experiences and knowledge.
Many of these rules are absolute, such as the force of
gravity, which can have no ambiguity. Such absolute rules
can be quantified mathematically, and are called laws of
nature. Other rules are less absolute, such as the path
a leaf takes as it falls to the ground. While controlled
by gravity, many uncertain variables - such as leaf shape,
random gusts of wind, and intervening branches - affect
the path in complicated ways. In these cases a statistical
representation may be more useful in describing how the
leaf falls.
1
Table 1.1: The Basis of Modeling
Data - Observations of our environment, collected by design
(hard data: a monitoring program, experimentation) or
by circumstance (soft data: ad hoc, incidental)
Information - Establishing relationships between observa-
tions by interpretation of data (hard information: sta-
tistical tests), or circumstantial relationships (soft infor-
mation: anecdotal)
Knowledge - Understanding why relationships exist - leading
to the ability to predict outcomes - based on scientific
principles (hard knowledge: hypothesis testing), or by
experience (soft knowledge: trail and error)
Wisdom - Using knowledge to successfully manipulate our
environment for personal or public good
Regardless of the approach one uses for documenting
the behavior of a system, the process of model building
provides the basis for documenting and communicating
the specific approach used to represent a system.
In the most general sense, a model is a simplified rep-
resentation of a system, where a system consists of entities
and relationships between entities. Each entity is a physi-
cally or logically distinct object, such as a person or event.
Relationships are the associations between entities, such
as the predator-prey relationships in a food web, or an
event tree in failure analysis.
Each entity and relationship has attributes, such as
size, color, intensity, or severity. Parameters are used to
quantify attributes, by providing quantitative or qualita-
tive values for each entity or relationship. Parameters nor-
mally remain constant, while state variables describe at-
tributes that change over time.
Inputs are external stimuli that cause changes to the
system, where the resulting change may be instantaneous
or delayed. The stimulus affects state variables as well as
outputs from the system. Outputs are external responses
of the system. If the response is delayed, then the maxi-
mum time required for the response to dissipate is called
the memory of the system.
CHAPTER 1. INTRODUCTION
Table 1.2: General Systems Terms
System - A set of relationships between entities
Entity - An object, person, or event
Relationship - An association between objects
Attribute - A quality of an entity or relationship
Parameter - A quantitative measure of the attribute of an
entity or a relationship
State Variable - A measurable attribute of an entity or rela-
tionship that changes with time or space. To distinguish
it from an input and output, state variables do not affect
other entities or relationships.
Input - A measurable quantity that behaves as a stimulus an
entity or relationship.
Output - A measurable response of an entity or relationship
to a stimulus.
Memory - The length of time that an input affects the output
1.2 Reasons for Modeling
Hydrologic models are commonly used to explain and pre-
dict complex behavior associated with the management of
environmental systems. Often, these models are used to
evaluate the effectiveness of regulatory controls, such as
for human health [49], pesticide and herbicide registration
[7]; waste isolation at proposed, existing, and abandoned
waste disposal sites [12], and to examine the effects of al-
ternative regulatory policies [53].
One reason is to learn how a system operates. The
process of building a model assists in data interpretation.
Data collected in time and space are analyzed for the pur-
pose of establishing relationships. These relationships, in
turn, provide information about the system being mod-
eled:
• Water levels in certain wells respond faster in one area
than another;
• Floods move faster in one section of a river than another;
and
• Fertilizers applied during one part of the year contami-
nate ground water, but not in another part of the year.
A second reason is to understand why observed rela-
tionships exist. This understanding is useful for generating
knowledge. We can employ basic physics to explain that:
• The recharge rate is dependent upon the amount of rain-
fall and the soil moisture content;
• The flood wave velocity is a function of the channel cross-
section; and
• Nutrient transport through the unsaturated zone is mod-
ified by plant root uptake.
Another reason for constructing or using a model is to
identify a cause, or to assign blame. The detection of a
contaminant in ground water must be the result of some
release. Where was the release, who caused it, when did
it occur, and why did it occur? A model is often used to
2
reconstruct a sequence of events, constrained by available
data, information, and knowledge, to assess the root cause
of some observed condtion.
A common reason to model is to predict the conse-
quences of alternative actions. We are often faced with
the situation that we have several options, but are not
sure which course of action is best. In these cases we turn
to models to help explore the possible outcomes of each
alternative action. Which action is most likely to succeed?
What are the possible adverse impacts of the alternatives?
Given a set of options, which will cause the least amount
of damage, or maximize the benefits? How can we design
a system that achieves our goals?
Clearly, there are many reasons for modeling. Each
application may have a different reason, thus requiring a
different model - no one model can be used for all potential
applications.
Different models and modeling strategies have been de-
veloped in response to these multiple modeling purposes.
While scientists generally seek to remove noise and un-
certainty during experiments by careful use of controls, a
regulator seeks to account for uncertainty by focusing on
worst-case scenarios. Engineers, as well, try to account for
uncertainty by using conservative measures.
Thus, regulatory and engineering models can be read-
ily separated from scientific models. The use of a scien-
tific model that accounts for the myriad forms of physical,
chemical and biological interactions may be fascinating
for a scientist, but not useful for the design of a simple
structure. In the same respect, an engineering model that
grossly oversimplifies a system may provide an adequate
basis for designing a structure, but ignores the components
that most interest a scientist.
The regulatory model differs from the engineering model
in that the engineer is not able to alter standards; the engi-
neer designs a structure to meet a specified standard. The
regulator, on the other hand, has the special obligation of
standards specification.
The regulator seeks to define standards that simulta-
neously achieve several objectives. One objective is to
safeguard the public interest. This objective often links
a standard to a critical measures of exposure, which can
be addressed using criteria that adequately account for
uncertainties and technological capabilities.
On the other hand, standards could be so specific that
only limited decision-making is given to the engineer, the
result of this which is to restrain efficiency. In an ideal
environment, standards should enable engineers to eval-
uate alternatives using measures of cost-effectiveness or
reliability, without the standards placing an undo burden.
The role of regulatory modeling, therefore, should be to
identify those measures of system performance that clearly
represent the public interest with a minimum of interdic-
tion by the state.
CHAPTER 1. INTRODUCTION 3

1.2.1 Scientific Models that Describe the parameter is near a threshold that determines failure
or success. An engineer may be willing to accept a margin
One objective of scientific modeling is to identify a new
of error so that a minimal likelihood of failure is possible.
or unanticipated process, or to identify novel interactions
Also, the stability of a specific parameter may be consid-
between processes. Keen observation or inference is re-
ered when designing materials or components. A more
quired to sort through known knowledge about previously
variable result generates concern, even though the value
unknown mechanisms. An additional objective is to char-
of the parameter is otherwise far from the performance
acterize the magnitude of environmental parameters.
threshold.
For both objectives, models are used in the inverse
Like the scientist, the engineer relies on inverse tech-
sense to identify system structure or identify parameters.
niques to estimate system behavior, but unlike the sci-
A forward model is generally used only to form predictions
entist, the engineer is satisfied with statistical certainty,
that are compared with observed state variables or model
rather physical certainty. Knowing that a parameter is
outputs.
bounded may mean more to an engineer than knowing the
Several hazards arise when scientific models are used
nature of the expected value. Because the engineer is seek-
in the regulatory environment. In these cases, model un-
ing a system design for optimal operation, the specification
certainties may not be explicitly quantified. Parameter
of an absolute constant is less meaningful than knowing
distributions and uncertainties, which are minimized dur-
the worst-case possibility. By containing the behavior to
ing a scientific experiment, may not approximate the true
desirable states, one can avoid unwanted outcomes.
diversity of parameter estimates expected under natural
The ultimate endeavor of any design effort, therefore, is
conditions.
to provide a system that offers predictable outcomes. The
Because the scientific estimation problem is inverse,
outcomes must be desirable, and they must allow for con-
rather than forward, extension of modeling results from
trol. By evaluating alternative designs, an optimal design
the calibration phase to the application phase is not nec-
can be selected based upon a performance measure. The
essarily available. That is, a model that has been con-
performance measure can have various forms; expected
structed using focused laboratory techniques, based on
value (benefits), extreme value (worst case), expected risk
sound scientific method, may never have been used to form
(consequences), etc. Because each of these measures are
a prediction at field scales in a way that can be verified
incommensurate in that they generally yield different se-
and tested. Even when the model may have been tested
lections, some balance of these measures is used. A stan-
for a specific field-scale experiment, the degree of agree-
dard mechanism is to apply benefit-risk analyses, which
ment between model prediction and observed results is left
balances the benefits to some against the costs to many.
unresolved.
Like scientific models, the use of engineering models by
While inverse models may be used in the regulatory
the regulatory community results in real dangers. Specif-
process to assign culpability, such as when sources of pol-
ically, model uncertainties may not conform with regula-
lution need to be identified for the determination of re-
tory intent. An engineering design that minimizes costs
sponsible parties, these inverse techniques generally lack
given a set of environmental standards may, or may not,
the designed controls used in scientific studies.
be better than a model that is free to balance costs against
Inference, in this case, is made difficult by the inherent
standards.
uncertainties in defining the amount and location of con-
An engineering model may yield an optimal solution
tamination (i.e., source terms), subsurface material prop-
for a specific standard, yet ignores a solution that costs
erties, and historical water levels. The exactitude of sci-
insignificantly more but reduces risks substantially. Ide-
entific hypothesis testing is further weakened by the legal
ally, tradeoffs between alternative engineering designs are
requirements of liability, thus placing this type of inverse
identified so that the relative merits of the alternative de-
problem outside the scientific venue.
signs can be evaluated. Due to public perceptions of risk,
the regulated and regulatory community, along with the
1.2.2 Engineering Models that Predict general public, should be involved in finalizing the optimal
The design of human structures may not require an ad- design.
vanced theoretical understanding of the components em-
bedded within it. Engineers are most concerned about 1.2.3 Regulatory Models that Control
the attributes of the system that are most likely to con-
To better understand the regulatory role of modeling, it is
tribute to performance and failure. Thus, those aspects of
important to note that models often serve as the interface
a design that affect system performance or failure require
between data and decision-making. Few individuals are
stringent characterization.
capable of making a sensible conclusion by examining large
Estimating the likelihood of failure, as a function of
volumes of data.
cost and performance, are an important aspect of engi-
Models assist in data interpretation by extracting in-
neering analysis. An engineer may devote resources to-
formation, and thence understanding, from data. Once
ward estimating a parameter, especially when the value of
experiments generate data, rules are applied via models
CHAPTER 1. INTRODUCTION
to gain insight into what happened where, when, why and
how. The information and knowledge gained from the in-
sight is then used to form a decision. Models are used,
therefore, to improve our knowledge of the system.
Many regulators commonly employ models to help learn
about processes in the environment. In this sense, the
models serve as true simulators; to teach the user how the
system behaves. Like pilots who use aircraft simulators
to improve their performance when learning how to fly
new aircraft, most modelers understand that the model
they use is not an exact replication of the real system.
With time, many modelers either move on to newer, more
powerful codes that provide additional insight and under-
standing, or they bypass the simulation route by learning
how to interpret data directly.
Veteran regulators acknowledge that simulation mod-
els are abstractions of reality. The abstraction is based
on simplified representations of the real world. It is also
recognized that model predictions are imperfect due to
uncertainties. At a minimum the regulatory model should
be sensitive to uncertainties, in that the performance of
a system should be tested under a wide range of possible
inputs, states, and material properties.
The complexity of material properties can be readily
quantified if the project is limited in scale, or the magni-
tude of the variation is small. Unfortunately, the effects of
environmental change cover a wide area and induce sub-
stantial perturbations to the environment in many ways.
The complexity of the model structure can also be evalu-
ated if the system functions deterministically, but is virtu-
ally impossible if the physical, hydraulic, chemical, ther-
mal or biological responses are poorly known. Also, the
complexity due to coupling between processes can be eval-
uated if the coupling are weak and limited in number, but
are difficult when the system is strongly coupled, or is
nonlinear in the effects of coupling.
From a regulatory viewpoint, simulation is straightfor-
ward if there are no uncertainties, but uncertainties tend
to introduce ambiguity in reaching a decision. A scientific
model that artificially controls or minimizes uncertainties
to allow identification of relevant processes and parame-
ters may not be useful for a decision-maker who wishes
to incorporate uncertainty. The engineering design that
incorporates uncertainties using conservative designs does
not necessarily allow the optimum balance of risk vs. cost
to be achieved. Each of these models therefore restrict
the regulator in their ability to incorporate uncertainties
in the decision-making process.
As noted above, a good performance measure should
conform with the regulatory intent of preserving and pro-
tecting the public health and welfare. As such, regulations
should protect public health, while minimizing intrusion
into the design process. The ability of a performance mea-
sure to evaluate the suitability of a proposed design, or to
assist in the approval of a permit, should reflect the ability
of the design to meet the regulatory intent.
A regulatory policy that enhances the efficiency of the
4
performance measure will require the use of regulatory
models. These models should fully account for model and
parameter uncertainties, while providing design flexibility,
and also incorporate societal goals.
1.3 The Modeling Process
1.3.1 The Objective
The modeling process begins with trying to articulate the
modeling objective. Is it a design objective that minimizes
cost while conforming to regulatory requirements? Is it a
scientific objective to understand how a system works? Is
it a regulatory objective whose purpose is to establish a
limit that is protective of human health and the environ-
ment? Ideally, a model should be as robust as possible
(i.e., suitable for use at many locations). Also, a model
that minimizes uncertainty is more useful than one which
gives highly uncertain predictions.
The greatest challenge to modeling results from the
confusion that commonly arises when scientists, engineers,
and regulators discuss modeling. This can be expected
given their disparate objectives. Simulation models vary
depending upon their purpose. Scientists employ simu-
lation models to formulate an understanding of complex
systems and to identify previously unknown phenomena or
processes. Engineers use simulation models to test alter-
nate designs and to predict system behavior. Engineering
models are generally coupled with prototype designs to
test model assumptions. The regulatory community uses
simulation models to evaluate the ability of proposed ac-
tions or activities to meet regulatory objectives. In many
regulatory applications, the model is used to extract in-
formation about regulatory performance from data so that
informed decisions can be made.
Clearly, an important first step is to clearly define the
objectives of your modeling effort. Examples include:
Developing a Well Field for Water Supplies. Models are
needed to meet an engineering objective - to provide
the most efficient way to develop ground-water resources
while minimizing impacts on local wells and wetlands
and complying with existing environmental laws.
Establishing Liability at a Superfund Site. Models are needed
to meet an identification objective - to identify the source
of ground-water contamination so that responsibility for
the contamination can be assigned. The models must be
physically-based because of the likelihood of litigation
over any liability that is assigned.
Establishing Nutrient Loading Limits. Models are needed
to meet a regulatory objective - to establish an upper
limit that is protective of the aquatic resource and hu-
man health. The models may be physically or statis-
tically based. Oftentimes, sufficient information is not
available to form a physically based model, so that sta-
tistical relationships are often used.
Approving a Pesticide License Application. Models are
needed to meet a regulatory objective - to evaluate whether
CHAPTER 1. INTRODUCTION
a chemical can be safely applied without undo harm to
human health and the environment.
1.3.2 The Approach
The second step is to define the model approach. Early
efforts to develop predictive models relied on scale models
that reduced the size of the system to a scale that fit in
the laboratory. Entire watersheds have been constructed
in miniature - with dams, bridges, and miniature obstruc-
tions - for the purpose of evaluating alternative designs
for channel structures. Laboratory column experiments
are a type of scale model still commonly used today to
evaluate the effects of contaminant transport through the
subsurface.
Another early effort used resistors and capacitors to
mimic the permeability and storage properties of an aquifer.
These process analog models were based on the similarity
between fluid flow and the flow of electricity. Heat con-
ductivity and capacity are also analogs for water flow, and
many heat flow models are commonly employed as surro-
gates for fluid flow.
Statistical procedures are useful when the behavior of a
system is highly uncertain, such as in rainfall-runoff mod-
els. In these cases, the complexities associated with ex-
plicitly modeling the physical processes overwhelm the in-
formation available to constrain model parameters.
In other cases, the physical processes are known and
appropriate parameters are readily available, so that both
analytic and numeric models can be used to evaluate and
predict behavior. While analytic models are usually pre-
ferred, due to their greater simplicity, site conditions may
not be conducive for direct application. Instead, numerical
models are used to provide greater flexibility in modeling
the unique conditions at the site.
The focus of this book is to provide methods and pro-
cedures to assist in understanding the relative strengths
and weaknesses these modeling approaches.
Scale Analogs - Models that employ a smaller (or larger)
version of the original system as an experimental tool.
Examples included model airplanes, Hele-Shaw models,
miniature flumes, and laboratory tests.
Process Analogs - Models that replace the flow of water with
another physical or chemical process. Examples include
electric-analogs, temperature-analogs, and solute-transport
analogs.
Statistical Models - Models based upon information, which
use observed data to identify relationships and to esti-
mate model parameters. Examples include regression,
convolution, time series, fourier, wavelet, and percola-
tion
Analytical Models - Models that provide analytic solutions
for specific physically-based systems. The solutions are
usually based on governing equations that take the form
of algebraic, differential, or integral expressions.
Numerical Models - Models that provide solutions when an- Instantaneous vs. Lagged Response - An instantaneous sys-
alytic solutions for physically-based systems are not pos- tem is one in which an input causes an immediate output,
5
sible. Examples include iterative, Finite Difference, Fi-
nite Element, Analytic Element, and other methods.
1.3.3 The Structure
Once a general class of model has been selected, a large
number of issues still need to be resolved. What processes
will be considered? What are the important relationships?
What state variables, inputs, outputs, and parameters can
be measured? How stable are these measurements? What
range of application is required?
Physical vs. Abstract - A physical model is a scaled replica
of the modeled system, such as a model airplane, meant
to perform identically to the original. An abstract model
is a mathematical representation of the modeled system,
meant to yield predictions that can be applied to the
original.
Natural vs. Devised - A natural model describes the phys-
ical environment, while a devised model is one, such as
law, that creates artificial rules for human conduct.
Open vs. Closed - A closed system has no input or outputs,
and is entirely self-contained. An open system interacts
with its external environment. Planet Earth is a closed
hydrologic system, but is open with respect to energy.
Steady vs. Dynamic - A steady system is one in which there
is no temporal variation in inputs, outputs, or state vari-
ables, while a dynamic system is one which changes over
time.
Stable vs. Unstable - A stable system is one which can be
controlled, while an unstable system is one which is dif-
ficult or impossible to control. One property of a stable
system may be that, in the absence of inputs, it returns
to a constant condition, while an unstable system grows
or shrinks without stabilizing.
Discrete vs. Continuous - A discrete system has inputs, out-
puts, or state variables that take on integer values, such
as the sex of a person, while a continuous system is one
in which components can vary over an infinite number of
conditions, such as a dimmer switch vs. an on-off switch.
Deterministic vs. Probabilistic - A deterministic system
is one in which all inputs, outputs, and attributes can
be described with great certainty. A probabilistic model
explicitly incorporates uncertainties in the model using
assumed distributions and statistical parameters.
Lumped vs. Distributed - Lumped parameters are used when
an attribute of a system can be described using a single
parameter or state variable. In some situations, however,
a single attribute can not fully describe the variability
within the system, and the entity or relationship must
be divided into multiple, or distributed, components -
each with their own attribute.
Causal vs. Noncausal - A causal system implies that an ac-
tion at one point or time causes a response at another
point or time. A noncausal system is one in which any
response between two points or times is the result of a
random coincidence.
CHAPTER 1. INTRODUCTION
and there is no memory. A lagged response occurs when
there is a delay between an input and a response.
Time Variant vs. Time Invariant - A time-variant model
has parameters that change with time, while a time-
invariant model has stable parameters. When faced with
a problem where a parameter is time-variant, the model
can be altered by specifying changing the parameter to
a state variable.
Linear vs. Nonlinear - A linear system is one in which the
magnitude of the response is a constant multiple of the
magnitude of the input. That is, a doubling of the input
causes a doubling of the output. A nonlinear system
results when the ratio of the response to the stimulus is
not constant, but varies as a function of the magnitude
of the input.
The definition of the model structure is called the con-
ceptual model. For some applications, this conceptual model
may be uncertain, and alternative conceptual models may
be required. For these cases, additional information may
need to be collected to evaluate which of the alternative
conceptual models is more appropriate.
1.3.4 The Evaluation Strategy
No phrase seems to induce greater discussion than model
validation. Accurate predictions hinge on a correct model
[10] [54]. While some argue that models can never be
validated, due to the inability to accept a hypothesis, oth-
ers argue that model validation depends upon the regula-
tory context. Efforts to evaluate suitability of model per-
formance focus on the key aspects of model verification,
which are used to evaluate the accuracy of coded state-
ments (software control), and model evaluation, which fo-
cuses on the ability of the theory to replicate observed
behavior. The issue of model validation comprises the
core of performance assessment; risk predictions depend
on accurate and precise measurements of anticipated out-
comes. The performance assessments are then used for
decision making, establishing regulatory policies, and for
evaluating regulatory compliance.
The greatest challenge in modeling is the proper math-
ematical specification of the conceptual model. The iden-
tification and quantification of relevant processes can be
performed using laboratory and limited-scale field exper-
iments. In general, the mathematical equations that re-
sult are greatly simplified from the real world. The quan-
tity, quality and types of data must be specified, both in
time and space. Once data are obtained, model results
are compared with experimental results to evaluate model
assumptions. The conceptual model is iteratively revised
until a specified performance measure is achieved. These
steps do not guarantee success. The calibration and evalu-
ation phases are both limited in areal and temporal extent.
Extrapolation beyond the range of experimental testing is
subject to great uncertainties, that are generally neglected
when running performance assessment simulations.
6
The performance assessment simulations are used to
develop alternative designs for a wide array of conditions,
beginning with limited-scale testing under relatively con-
trolled conditions, to extended-scale testing under open
conditions. These design comparisons can be used to pro-
vide model results that limit the prototype testing to a
smaller set of alternative system designs. The alternative
prototype designs are then evaluated using a variety of
performance measures, recognizing the uncertainties in-
herent in open testing. Finally, the sensitivity of model
predictions to the magnitude of uncertainties is evaluated.
Confidence in the validity or appropriateness of an en-
vironmental model is enhanced if the model accurately
and concisely represents the relevant physical, biological,
chemical, hydrologic, geologic, and thermal processes in-
herent in the system being modeled. Failure to account
in a meaningful way in any of these components may
compromise the robustness of model predictions. Estima-
tion of model single-phase and coupling parameters must
be performed under a wide range of ambient conditions,
especially those appropriate to the target environment.
And finally, extended prototype testing at multiple spa-
tial scales, time horizons, and thermal loadings is required
to identify unanticipated events. Reliance on natural ana-
logues that mimic the long-term effects of environmental
disturbance, such as volcanic and meteorite-impact events
are a possible mechanism for extending prototype testing.
Validation of model performance lies in the ability to
quantify model uncertainties. While some model uncer-
tainties can be quantified, many uncertainties related to
model structure remain unquantifiable. Regulatory poli-
cies that rely on model predictions may not provide greater
confidence than other approaches. Alternative regulatory
strategies include an increased reliance on prototype test-
ing, improved linkages between hydrologic modeling ca-
pabilities and standards, and decoupling models devel-
oped for scientific purposes from the regulatory process.
In some cases, regulatory policies and simulation model
capabilities may be incompatible, leading to the improper
use of simulation models in the regulatory environment.
Engineers routinely establish tradeoffs between model
performance and standards by incorporating margins of
safety to account for uncertainties [53]. The degree of
confidence reduces the need for margins of safety, and al-
lows for economies in the design and operation of a struc-
ture. For a new design, large margins of safety are rec-
ommended, which are slowly relaxed as the understand-
ing of the system increases. A regulatory policy that ac-
knowledges uncertainties, or develops methods for explic-
itly linking uncertainties to standards, is one means for
addresses incomplete understanding. Uncertainty magni-
tudes are especially large early in the design state, and are
solidified as the prototype testing reduces unknowns.
A regulatory policy that focuses exclusively on risk
minimization may fail to account for identification of un-
known uncertainties. A process of uncertainty characteri-
zation and minimization can be monitored using the var-
CHAPTER 1. INTRODUCTION
ious performance measures addressed above. The process
of testing need not be static, however. As new processes or
behaviors are observed, increased surveillance and testing
may be required. To define, a priori, the magnitude of un-
certainty may not be possible. In fact, the magnitude will
certainly be underestimated. Minimization and control of
uncertainties leads to reliability of operation, the primary
concern of the regulator.
1.3.5 The Limitations
A primary inadequacy of environmental models stems from
their inability to provide long-term predictions with suf-
ficient accuracy to meet regulatory requirements [49] [82]
[84]. Assuring the long-term continuation of sustainable
human and ecologic conditions for millions of years is diffi- to identify unknown processes or scenarios, or to quan-
cult, given unforeseeable climate changes, human and an-
imal interactions, geologic and astronomic instability, and
other, as yet unknown, modifications of the global environ-
ment. For a regulatory requirement of predictability for
only ten thousand years, many scenarios of climate change
or volcanic activity could adversely affect the integrity
of global systems. Incorporating the potential stresses of
changes in the external environment is rendered more diffi-
cult when observed behaviors are limited by an incomplete
understanding of potential future states [61].
An additional problem arises when more than a single
interpretive model can be used to explain observed exper-
imental data [37] [23] [80]. In the case of any experiment
in which data are incomplete, alternative scenarios can
be devised that may account for the observed distribution
of state variables or measured fluxes. Thus, alternative
hypotheses may not be distinguishable due to incomplete 1.4
control of the experiment.
Calibration data sets and model structure substan- Select a journal article of your choice that uses a model
tively affect model prediction accuracy. For data sets with in the analysis of field data. For the model you select,
high multicollinearity, the parameter covariances become identify the:
large, resulting in poor forecasts. Also, extrapolative mod-
els in which parameter values that lie far from the mean
results in larger forecasting errors. And finally, the model
shows that the objective of reducing model calibration er-
rors can result in a poorly structured parameter covariance
matrix, offsetting any reduction in forecast error.
A further confounding factor lies in the situation where 4. Parameters and state variables within the model
a regulatory policy may not be amenable to quantitative
analysis [26]. Without being specific, it may be safe to say
that a quantitative definition of safe, is not scientifically
possible. A policy which requires safety requires the speci-
fication of a quantifiable measure. A quantifiable measure
may be the specification of a risk envelope, which is the
product of the probability of an undesirable state with the
likely consequence of the state, for all possible states [17].
The total risk is the sum of risks of all possible adverse
events.
The total risk can be limited to specific subsets, e.g.,
the risk from carbon accumulation in the atmosphere. Risk
calculations incorporate all outcomes, with the obvious
7
bias due to the inability to specify the universe of out-
comes. Also, the risk computation should account for risk
transfers (e.g., between present and future generations),
perceptions of risk (e.g., brief, high exposures may be more
acceptable than chronic, low exposures given the same to-
tal consequence), and public voice (e.g., poor, minority
populations may have less access to the political process
than established political groups).
While mathematically simple, the use of risk as a mea-
sure of safety suffers from the inability to calculate either
the likely probability of a single hazard, or even the re-
sulting consequence for that hazard [21] [22]. The advan-
tage of risk assessments lies in their ability to incorporate
known uncertainties by coupling them with the magnitude
of the consequence, yet they are biased because they fail
tify unknown uncertainties [60]. The risk measure is a
relative measure, requiring the balancing of risks between
alternative decisions [11]. Also, community perceptions
of acceptable risks commonly differ from calculated risks
[81]. And finally, risk is transferable, allowing individuals
to benefit from the increased exposure of others.
The dilemma arises when a regulatory policy may not
adequately incorporate model capabilities and uncertain-
ties [60]. A model which superficially treats complex sys-
tems by using expected inputs, states, or material prop-
erties, may yield precise estimates of expected outcomes.
Yet this model will undoubtedly fail to accurately repre-
sent the wide range of possible outcomes by ignoring the
less likely, but still possible, alternatives.
Problems
1. Modeling objectives (e.g., scientific, engineering, reg-
ulatory)
2. Processes described by the model
3. Input information used to construct the model
5. Outputs predicted by the model
6. Measures of model performance
7. Construct a visual diagram of the model showing
the entities, relationships, parameters, state vari-
ables, inputs, outputs, feedback loops, and calibra-
tion checks.
8. Describe alternative modeling methods, and discuss
how they might be useful for meeting the modeling
objective.
Chapter 2

Analog Models

An analog model is an artificial physical representa-

Table 2.2: Metric System Prefixes
tion of the system that is to be studied. For example,
a process analog substitutes a different process, such as d deci 10−1 D deca 101
electricity or heat, to represent fluid flow. Another class c centi 10−2 h hecto 102
of analog models are scale analogs which use a miniature m milli 10−3 k kilo 103
(or perhaps an expanded) version of the system. Exam- µ micro 10−6 M mega 106
ples of scale analogs include miniature dams and rivers, n nano 10−9 G giga 109
or miniature aquifers. These models can be constructed p pico 10−12 T tera 1012
in a laboratory so that experiments can be conducted to a femto 10−15 P peta 1015
evaluate alternative designs.

2.1.2 Exponents and Logarithms

2.1 Review The exponent is used to indicate the repeated multiplica-
Before we launch this baby, let’s do a bit of review. I’m tion of a number:
hoping that you’ve had (at least some of) this before. If x3 = x · x · x (2.1)
not, please give it a try. . .
A negative exponent indicates the reciprocal of the num-
2.1.1 Units ber:

The metric system is the internationally recognized system 1

x−3 = 3 (2.2)
of units. All countries in the world (except for Liberia, x
Burma/Myanmar, and the United States) use the metric Non-integer exponents are also possible. For example, the
system as their official way to measure things. Mixing exponent one-half indicates the square root of a number:
English with other units can result in engineering failure,
1 √
such as the failure of the $125 million Mars Climate Or- x /2 = x (2.3)
biter spacecraft in 1999. In this case, two teams of scien-
tists used two different systems of units, which were never A general rule for exponents is:
reconciled. a √
a
x /b = xx (2.4)

Table 2.1: Metric System Units There are two standard bases for exponents. One is the
base-10 system, such that any number can be represented
Mass kilogram kg using:
Length meter m
Volume liter L x = 10b (2.5)
Time second s
Energy joule J (kg m2 )/s2 For example, b = 2 represents the number 102 = 100. A
Power watt W (kg m2 )/s3 J/s second base is the natural system, represented using e:
Force newton N (kg m)/s2 J/m
Pressure pascal Pa kg/(m s2 ) J/m3 x = ea (2.6)

where e ≈ 2.71828.
The logarithm of a number is equal to the exponent:

log10 x = log10 (10b ) = b (2.7)

8
CHAPTER 2. ANALOG MODELS 9

ln(x) = loge x = loge (ea ) = a (2.8) 2.2.2 Electricity

There are several basic rules of exponentials and loga- The corresponding equations for electricity are:
rithms:
x · y = 10b · 10c = 10b+c (2.9) ~i = − ∇V (2.19)
R
x 10b ∂V
= c = 10b−c (2.10) ∇ · ~i = −C (2.20)
y 10 ∂t
These relationships also hold for the natural system: ∂V
D ∇2 V = (2.21)
b
x·y = e ·e = ec b+c
(2.11) ∂t

x eb where ~i is the electrical flux, or amperage, vector, R is the

b−c
= c =e (2.12) electrical resistivity parameter, ∇V is the gradient in the
y e
electrical potential, or voltage, D = 1/RC is the electrical
Logarithms also follow similar rules: diffusivity, and C is the capacitance.
log(b · c) = log(b) + log(c) (2.13) In this case, the flux of electrons replaces the flux of wa-
ter, the reciprocal of the resistivity replaces the hydraulic
log(b)
log(b − c) = (2.14) conductivity, the voltage replaces the hydraulic head, and
log(c) the capacitance replaces the specific storage coefficient.
for both the base-10 and natural systems. Another loga- As an example, one can construct a ground-water flow
rithmic relationship is: model using a network of resistors. Resistors are inversely
related to the hydraulic conductivity, so a low permeabil-
log(axb ) = log(a) + b log(x) (2.15) ity aquifer would be constructed using resistors with high
resistances. A voltage can be applied at any point within
the network to simulate the addition or removal of water.
2.2 Process Analogs A ground is applied at either a point, or along a line, or
The rules that govern fluid flow are remarkably similar to at multiple points to simulate the return of the water to
those that govern heat, electrical, and solute flow. All are the system.
governed by conservation and flux equations, with similar Another means for studying fluid flow is to use electri-
laws and parameters. cally conductive paper. The flow domain is drawn on the
The general conservation equation requires that any paper and then the non-conducting part of the domain is
accumulation or loss in mass, heat, or current be offset cut off. An electrode is placed at one or more locations
by a corresponding increase or decrease in storage. The on the paper, while a second electrode is placed elsewhere.
general flux law states that the rate of movement of mass, The voltage on the paper is used to simulate the resulting
heat, or current is related to the product of a parameter flow field. This analog is limited to uniform conditions,
with the gradient of a state variable (head, temperature, unless papers with different electrical conductivities are
voltage, or concentration). used.
The ratio of the flux to the conservation parameters
is termed the diffusivity, and always has units of length 2.2.3 Heat
squared per unit time (m2 /s). Thus, the flux parameter
units must be paired with the conservation variable pa- The flux equation of heat is:
rameter so that concomitant units are used. ~j = −K ∇T (2.22)

2.2.1 Fluid Flow ∂T

∇ · ~j = −Cp (2.23)
∂t
The steady and unsteady flow equations for water in the
subsurface are: ∂T
D ∇2 T = (2.24)
∂t
~q = −K ∇h (2.16)
where ~j is the heat flux vector, K is the thermal conduc-
∂h tivity parameter, ∇T is the gradient in the temperature,
∇ · ~q = −Ss (2.17)
∂t D = K/Cp is the thermal diffusivity, and Cp is the heat
∂h capacity.
D ∇2 h = (2.18) A similar analog using heat can also be constructed. A
∂t
where ~q is the flux vector, K is the hydraulic conductivity heat-conducting surface is constructed in the shape of the
parameter, ∇h is the gradient in the hydraulic head, D = system to be studied. A constant heat source is placed
K/Ss is the hydraulic diffusivity, θ is the volumetric water at one or more locations, and the resulting temperature
content, and Ss = dθ/dH is the specific storage coefficient. is measured. Materials with variations in conductivity
CHAPTER 2. ANALOG MODELS 10

Table 2.3: Physical analogs between processes

Process Conservation Variable (parameter) Flux Law (parameter)

Water Mass, Volume (storage coefficient) Darcy’s (hydraulic conductivity)
Heat Calories, Joules (heat capacity) Fourier’s (thermal conductivity)
Electricity Amps, Electrons (capacitance) Ohm’s (resistance)
Solutes Mass, Moles Fick’s

can be used to reproduce variations in water transmitting
properties.
If steady flow is desired, then only the conductance,
or flux, component is needed. However, the storage must
be included when modeling dynamic conditions, i.e., when
there are temporal changes in the system. For the electri-
cal analogs, capacitors must be placed alongside the resis-
tors. For inertial components, inductors (or coils) must be
used. For thermal analogs, the heat capacity of the sur-
face should match the storage coefficient of the hydraulic
system.
2.2.4 Solutes
The flux equation of solutes due to a concentration gradi- fluid density, µ is the dynamic viscosity, and ν = µ/ρ is
ent is:
J~ = −D ∇C (2.25) µ = 1 g/m/s, and ν = 10
∂C
∇ · J~ = − (2.26)
∂t
∂C
D ∇2 C = (2.27) with lower viscosity. A Newtonian fluid is one which obeys
∂t
where J~ is the solute flux vector, D is the solute diffusivity,
∇C is the gradient in the solute concentration.
2.3 Scale Analogs
A scale model is a common tool for studying large sys-
tems. Like a map, the original system is reduced in size so
that it fits within a laboratory or other reasonably sized
structure. For a system with an original length of Lo , the
modeled scale, Lm , is:
Lm = λ Lo (2.28)
where λ is the model scale. Note that λ < 1, meaning
that a smaller scale involves a larger reduction - a 1:1000
model is a smaller scale than a 1:100 model.
For example, a river with a width of 1 km, a depth
of 10 m, and a length of 1000 km can be scaled down
to a manageable size using a λ = 1:10,000 scale, so that
the model river is 10 cm wide, 1 mm deep, and 100 m
long. While this is still the length of a football field, model
simulations might be performed within an indoor arena of
sufficient size.
The features of the flowing river would also have to be
reproduced. If the original bed material within the river is
a sand with a diameter of 0.3 mm, then the bed material
of the scale model would have to have a diameter of 30 nm,
or 300 Å, which is a fine clay particle. Also, a velocity of
3 m/s in a natural setting would correspond to a velocity
of 0.3 mm/s, or about 1 m/hr.
The viscosity of the fluid would also have to change be-
cause the shear stress of fluids is a function of the length
scale used. As the depth of water, or any fluid, gets thin-
ner, then the effects of viscosity increase. The Reynold’s
number, R, describes the ratio of inertial to viscous forces:
vLρ vL
R= = (2.29)
µ ν
where v is the fluid velocity, L is the length scale, ρ is the
the kinematic viscosity. For water at 4◦ C, ρ = 1 kg/L,
−6
kg/m2 /s. We know that
water flows is in a laminar manner if R < 10, and flows in
a turbulent manner when R > 100.
The viscosity of water is a measure of its ability to flow
- fluids with high viscosity flow more slowly than fluids
the relationship that:
∂vx
Υ = −µ (2.30)
∂y
where µ (Pa s), is the dynamic viscosity of water, and Υ
is the shear stress within the fluid, and ∂vx /∂y is the rate
of change in the y-direction of the fluid velocity in the x-
direction. The shear stress is a measure of the forces on
the fluid, and turbulent flow occurs if the forces exceed a
threshold value.
Thus, in order to maintain the proper flow conditions,
laminar or turbulent, then the viscosity of the fluid must
change at a rate equal to the product of the velocity and
length scales.
2.3.1 Dimensional Analysis
Whenever you use an equation, pay particular attention
to the units - they must balance each other on each side
of the equation. For example:
dx (m)
v (m/s) = (2.31)
dt (s)
where v is the velocity, in units of meters per second, dx
is the change in distance, in units of meters, and dt is
CHAPTER 2. ANALOG MODELS 11

the change in time, in units of seconds. Note how the Tortuosity, τ : The ratio of the path length ∆s, to the
units on the left-hand side perfectly balance those on the ruler (straight-line) length, ∆x:
right. One can check the accuracy of the equation, or your
calculations, by examining whether the units are correct. ∆s
τ= (2.38)
When modeling a new system, one can develop a func- ∆x
tional relationship just based on the units. For example, if The path length along a streamline between two points is
we know that the velocity is based on distance and time, generally unknown, however. In general, the parameters
there is only one combination that yields a balanced set of which may be measured using experimental tests are:
units. So, rather than having to remember the equation,
one only needs to remember the units.
Total Head Difference, ∆h: The change in head be-
Taking this one step further, the Buckingham-π theo-
tween two points:
rem states that one can establish a set of possible phys-
ical relationships if the number of independent variables ∆h = hb − ha (2.39)
are known. For example, if we know that the velocity of
falling water is affected by distance and time, as well as Hydraulic Gradient, i: The change in head, ∆h, per
gravitational acceleration, then, by dimensional analysis, unit distance, ∆x:
we arrive at the following possible combinations of vari-
ables: ∆h
i= (2.40)
2
∆x
v (m/s) = f {g (m/s ), dt (s)} (2.32)
Hydraulic Conductivity, K: The ability of a geologic
v (m/s) = f {g (m/s2 ), dx (m)} (2.33) medium to transmit water, calculated using:
g (m/s2 ) = f {v (m/s), dt (s)} (2.34) q
K= (2.41)
2
where g is the gravitational acceleration in units of m/s . i
Using a bit of math, we can show that these are equivalent
to: Travel Time, tt : The time required for a particle of
water to move from one point to another, equal to the
v = g dt (2.35) distance, ∆x, divided by the fluid velocity, v:

v 2 = 2g dx (2.36) ∆x
tt = (2.42)
v
dv
g= (2.37)
dx Fluid Velocity, v: The rate at which water moves through
the aquifer, equal to the flux, q, divided by the effective
2.3.2 Example: Fractal Scaling porosity, n:
Some systems display self-similar behavior, meaning that q
it looks the same regardless of the scale of measurement. v= (2.43)
n
Examples include:
Effective Porosity, n: The volume of voids, VV , per
Shorelines: Length of land-ocean boundary increases as unit volume of geologic medium, V :
T
ruler length decreases
VV
River densities: Number and length of waterways in- n= (2.44)
VT
creases as map scale becomes finer
The hydraulic gradient can be calculated at two scales,
Soil Physics: Scaling of particles shifts soil-moisture char- as a straightline gradient along ∆x, or along the curve
acteristic curves to common shape associated with the true path described by ∆s
Geophysical Measurements: Bulk resistivity is not just ∆h ∆h
product of resistivity and porosity ix = and is = (2.45)
∆x ∆s
Fractured Media: Fracture density changes as the scale It is easy to see that
of measurement changes
∆h ∆h ∆s
We will examine tortuosity which is a commonly ob- ix = = = is τ (2.46)
∆x ∆s ∆x
served property of environmental systems. The definition
of tortuosity used here (recognizing that there are several or
different definitions in the literature) is: ix
is = (2.47)
τ
CHAPTER 2. ANALOG MODELS 12

The hydraulic conductivity at the experimental scale Summary. This example demonstrates the effect of tor-
can also be related to the value at the streamline scale: tuosity and the scale of measurement on the hydraulic
conductivity, hydraulic gradient, and travel time. A crit-
q q Ks
Kx = = = (2.48) ical parameter in this analysis is the tortuosity, a result
ix τ is τ
of the geometry of the flow regime. Additional research
q q is required related to the effects of spatial variability on
Ks = =τ = τ Kx (2.49)
is ix tortuosity, and the relationship between scale and the es-
For extrapolating tests from one scale, say at a field or timated tortuosity.
laboratory scale of size, ∆x1 , to a different scale, ∆x2 , the The spatial variability of tortuosity along the stream-
following relationship can be used: line may or may not have significant effects on the esti-
mated travel time. It may be possible that local fluctu-
K1 i2 τ2 ations in this parameter may not significantly affect re-
= = (2.50)
K2 i1 τ1 gional travel times. Also, if the tortuosity is scale invari-
ant, then laboratory and field parameter estimates can be
Similar to the gradient and the hydraulic conductivity, directly applied to regional-scale models without the need
the calculated travel time may also be affected by the scale for incorporating scale effects.
of measurement. The travel time is defined here as the
integral of the inverse velocity along a one-dimensional
streamline: 2.4 Problems
Z sb
tt = v −1 ds (2.51) 1. Exponents and Logarithms
sa
(a) What is the range of ex if the range of x is
where tt is the fluid travel time, v is the fluid velocity along −∞ < x < ∞?
streamline, s is the distance along streamline, and sa and (b) What is the range of x if the range of log(x) is
sb are the particle starting and ending positions, respec- −∞ < log(x) < ∞?
tively. The fluid velocity is the volumetric flow rate per
unit area (i.e., the darcian flux) divided by the porosity, 2. Process Analogs Using the provided electrical con-
or: ducting paper:
q Q K s is (a) Construct a fluid flow model using a simple ge-
v= = =− (2.52)
n nA n ometry, like a stream with pools and narrows,
where q = −Ks is is the fluid flux, n is the porosity, Q is or an aquifer with varying shape.
the total flow, A is the cross-sectional area, Ks is the local (b) Attach a low-voltage electrical source to the pa-
hydraulic conductivity, and is is the local hydraulic gradi- per - such as a 9-V battery - so that the neg-
ent. By assuming constant velocity along the streamline, ative terminal is connected to one end of the
we obtain: paper and the positive terminal is attached to
the opposite end.
∆s n∆s n∆s n∆s2
tt = = =− =− (2.53) (c) Map lines of constant voltage on the paper.
v q Ki Ks ∆h
(d) Discuss how one could construct a heat analog
where i = ∆h/∆s, and where ∆s = sb − sa is the distance model using various metals.
along the streamline, and ∆h = hb − ha is the head drop
along the streamline. Switching to ruler lengths, we have: (e) Discuss the advantages and disadvantages of as-
tronauts using swimming pools to simulate the
nτ 2 ∆x2 nτ ∆x2 effects of zero gravity.
tt = − =− (2.54)
τ Kx ∆h Kx ∆h
3. Scale Analogs Consider a miniature aquifer mea-
If the concept of fractal scaling is employed, then a suring 1-m long, 10-cm wide, and 1-cm thick, in
relationship between the tortuosity at one scale can be which food-coloring is added to show contaminant
related to the tortuosity at a different scale: transport:

τ1 = τoβ∆x1 (2.55) (a) Discuss the effects on model results if water and
a sand medium is used.
and (b) Discuss how these effects might be reduced or
τ2 = τoβ∆x2 (2.56) eliminated by using alternate materials.
4. Tortuosity Consider an aquifer with a tortuosity of
where β is a fractal scaling parameter and τo is a dimen-
2.
sionless fractal tortuosity parameter.
CHAPTER 2. ANALOG MODELS 13

(a) Discuss the impacts of a travel time prediction

if the hydraulic conductivity was estimated us-
ing flow between two wells.
(b) Explain how the tortuosity (sometimes called
the formation factor) might be estimated using
electrical methods.
Chapter 3

Network Models

We are often faced with the challenge of organizing 3.1.1 Frequency Distributions
data. We start by grouping like quantities together; put
There are two classes of distributions - discrete and contin-
all lakes in one group, all rivers in another, and all ground-
uous. A discrete distribution is used when only countable
water in a third. These groupings could be considered to
number of outcomes are possible, such as the tosses of a
be entities. We then try to find relationships between
coin, or the number of students in a class. Continuous dis-
these entities; surface and ground water interact, rivers
tributions are used for describing outcomes can have any
flow into and out of reservoirs.
fractional value, such as the monthly or annual rainfall
This organization is called a network, in that each en-
depth.
tity is qualitatively different from the others, and there are
unique relationships between each entity. As the model
develops, one may begin to distinguish between specific Discrete Distributions
elements within each entity; small lakes behave differently Examples of discrete distributions that can be used to
than large lakes, surficial aquifers behave differently from model these problems include the uniform, binomial, and
confined aquifers. We then create new entities and rela- geometric distributions.
tionships between these entities. The discrete uniform distribution, as the name implies,
This chapter focuses on network models, and how they has a constant probability for all outcomes between x1 and
can be used to represent environmental systems. x2 . The corresponding probability for each outcome, x, is:
1
3.1 Percolation Models P (x) = (3.1)
n
Percolation models use nodes to describe locations that where n are the number of outcomes.
have discrete values, such as occupied or vacant, on or off. Another discrete distribution is the binomial, which
They also uses bonds that connect the nodes, which can we used earlier to predict the probability of the number of
be open or closed, and can have distinct attributes. In this heads when a coin is repeatedly tossed n times:
section, we describe several types of percolation models,  
such as: n
P (x) = px (1 − p)n−x (3.2)
x
• Water percolating through a soil column
where
• Contaminants percolating a fractured rock network  
n n!
• Wells connected through a layered aquifer system = (3.3)
x x! (n − x)!
These models are a function of scale and dimension: is the combinatorial operator that accounts for the number
of opportunities for getting the same outcome.
• As distance decreases between boundaries, connected
Another common discrete distribution is the geometric,
network probability changes
which is used to calculate the probability of a failure after
• For 1-d flow, only one bond needs to be broken to x attempts:
shut down the network, fewer bonds are needed to
complete a network in higher dimension P (x) = p (1 − p)x (3.4)

• Fracture length strongly affects network connectiv- where p is the likelihood of failure on each attempt.
ity, long fractures form a backbone

14
CHAPTER 3. NETWORK MODELS 15

Continuous Distributions 3.1.2 Markov Chains

Examples of continuous distributions include the uniform, A Markov Chain model has discrete (or finite) states, such
normal, exponential, gamma, and Gumbel extreme value. as rainy and dry days. We can use the concept of prob-
One can increase the number of distributions by taking ability to say that the sum of all states has to equal one.
the logarithm of the random variable, resulting in distri- That is, if p is the probability that a day is rainy, and q is
butions such as the log-normal, log-gamma, etc. the probability that it is dry, then p + q = 1.
The continuous uniform distribution takes the form: Once the states are defined, we need to define a state
1 transition matrix, which is simply the function that indi-
P (x) = (3.5) cates the probability of moving from one state to another.
b−a
when a ≤ x ≤ b. This is a two-parameter distribution,
meaning that two numbers, a and b, are sufficient to de- Table 3.1: State Transition Matrix
scribe the distribution.
The normal distribution is another two-parameter dis- Today
tribution, requiring the mean and standard deviation, x̄ Rainy Dry
and sx , respectively. The distribution is unbounded be- Tomorrow Rainy Pww Pdw
low and above, defined using: Tomorrow Dry Pwd Pdd
1 2
P (x) = √ ez /2 (3.6) where Pww is the probability of a rainy day followed by
2π x
e rainy day, etc.
and where It is easy now to estimate the probability of different
x − x̄ types of events:
z= (3.7)
x
e
• Two dry days followed by a rainy day:
is the standard normal variable with mean zero, z̄ = 0,
and unit standard deviation, ze = 1.
P (d, d, w) = Pdd × Pdw (3.13)
The log-normal distribution is a two-parameter distri-
y
bution obtained by setting y = ln(x) so that x = e is • n dry days followed by a rainy day:
log-normally distributed. The log-normal distribution is
unbounded above, but bounded by zero on the left. P (d, d, ..., d, w) = (P )n−1 × P (3.14)
dd dw
1 z 2 /2
P (x) = √ e (3.8) • n wet days followed by a dry day:
g
2π ln x
ln x − ln x P (w, w, ..., w, d) = (Pww )n−1 × Pwd (3.15)
z= (3.9)
g
ln x
The exponential distribution has just one parameter, Note also, the following properties:
x̄, and takes the form: • P (w) = P (w|w)P (w) + P (w|d)P (d)
1 x/x̄
P (x) = e (3.10) • P (w) = P (w|d)/[1 − P (w|w) + P (w|d)]
x̄
Note that this distribution is bounded by 0+ on the left, which are forms of Bayes Theorem. Also note that:
and ∞ on the right.
The gamma distribution, is a two-parameter distribu- • The transition probability decreases for increasing
tion that represents the sum of exponentials: precipitation depths (0.01, 0.10, 1.00) because the
likelihood of two days in a row with heavy rain is
β −α xα−1 e−x/β much less likely than days with light rain.
P (x) = (3.11)
Γ(α)
• P (w|w) 6= f (z, t), is relatively stable geographically
where α = (x̄/e x)2 , β = x e2 /x̄, and Γ(α) = (α − 1)! is the and seasonally. Note that P (w|w) is the storm per-
gamma function. The gamma function is the same as the sistence probability, and P (d|w) = 1 − P (w|w) is
χ2 distribution when α = ν/2 and β = 2. the storm departure probability. If the probability
The log-gamma function is also known as the Log- is constant, this means that storm persistence is rel-
Pierson Type III distribution which is widely used for atively uniform in space and time.
extreme value problems. Another common extreme-value
distribution is the Gumbel : • P (d|d) = f (z, t), varies geographically and season-
1 z
ally. Again, this is the drought persistence proba-
P (x) = ez e−e (3.12) bility and P (w|d) = 1 − P (d|d) is the storm arrival
x
e
probability. It is clear that desert climates have a
where z = (x − x̄)/e x is the standard normal variable.
CHAPTER 3. NETWORK MODELS 16

much lower chance of storms arriving than in a wet-

ter climate. Also, areas with strong seasonal varia-
tion in weather, such as Monsoonal and the Mediter-
ranean climates, will have different rates of storm
arrivals over the course of a year.
• Problem of persistence: extremely dry and wet weather
may reinforce themselves. The solution is to add ad-
ditional states that represent drought and flood con-
ditions. This results in additional transition proba-
bilities.
3.1.3 Discrete Event Modeling
Some things happen as distinct events - heads vs. tails,
on vs. off, boy vs. girl, day vs. night. While other
things happen along a gradient - shades of gray, moisture
content, income. Models that describe the discrete nature
of things are often easier to understand and build than
ones that handle continuously changing variables.
• The roadway over a river will be flooded or dry.
• The water level in a well will drop below the bottom
of the well.
3.1.4 Example: Extreme Values
The objective is to fit observed frequencies to a probability
model. Extreme value distributions are commonly applied
to maximum or minimum annual discharge. In this case
the largest or smallest event in each year is identified. Al-
ternatively, a data set could be constructed using all daily
observations above, or below, a threshold discharge. This
alternative technique is called a partial duration series.
Table 3.2: Method for Determining Return Periods
1. For discharge, and other heteroscedastic variables,
transform the observations, x, using the logarithm
(base-10), y = log x to make them homoscedastic.
2. Rank each observation from largest to smallest, m =
1, 2, · · · , n, where n is the number of events
3. Calculate the exceedence probability of each obser-
vation, Pi , using one of the following ranking statis-
tics:
Method Exceedence Probability
Weibull m / (n + 1)
Hazen (m − 0.5) / n
Cunnane (m − 0.4) / (n + 0.2)
4. Plot the exceedence probability, Pi , against the ob-
servations, xi
5. Calculate the recurrence interval, Ti = 1/Pi .
6. Plot the recurrence interval against the observations.
7. Specify a probability model (log-normal, log-gamma,
etc.)
8. Estimate the sample moments and use these to esti-
mate the model parameters
9. Show the calculated probabilities on the above plots.

3.2.1 System Inputs

3.2 Linear Systems Models
While many systems have random inputs, it is often help-
A system is composed of inputs, outputs, state variables, ful to examine the response to specific types of inputs,
and parameters. One can draw a box with an input ar- such as:
row, an output arrow, parameter boxes, and state-variable
dials. Dirac Delta
A linear system is one in which the value of the output 
varies linearly with the value of the input. For example: ∞ t = to
δ(to ) = continuous time (3.20)
0 t 6= to
y = f (x) = 2x (3.16)
is a linear system, because doubling the input, x∗ = 2x, Kronecker Delta
results in a doubling of the output y ∗ = 2x∗ ). Yet, 
1 t = to
δ(to ) = discrete time (3.21)
y = f (x, b) = 2x + b (3.17) 0 t 6= to
is not a linear system because 2y ∗ 6= f (2x∗ , b):
Heaviside (step)
y ∗ = f (x∗ , b) = 2x∗ + b 6= 2(x∗ + b) = 2x∗ + 2b (3.18) Z ∞
One can convert the non-linear system into a linear one H(t o ) = δ(to ) dt
◦ −∞
by removing the constant, b, so that y = y − b:

y ◦ = y − b = f (x) = 2x (3.19) 0 t < to
= (3.22)
1 t ≥ to
which is now a linear system in y ◦ .
CHAPTER 3. NETWORK MODELS 17

Pulse

P (ta , tb ) = H(ta ) − H(tb )


 0 t < ta
= 1 t a ≥ t < tb (3.23)

0 tb ≥ t

Ramp
Z tb
1
R(ta , tb ) = P (ta , tb ) dt
tb − ta ta

 0 t < ta Figure 3.1: Triangular Unit Hydrograph.
t−ta
= t a ≥ t < tb (3.24)
 tb −ta
1 tb ≥ t
3.2.3 Deconvolution
3.2.2 Convolution • An inverse method to estimate impulse response func-
tion
Convolution is a type of mathematical operator, defined
using y = h∗x where ∗ is the convolution operator and h is • Estimate h(τ ) values using linear regression, where
the unit response function. For the case where the input
is a Delta function (either continuous or discrete), then y(i) = h(0) x(i) + h(1) x(i − 1) + h(2) x(i − 2)
y = h. This means that a spike input causes an output
+ h(3) x(i−3) + · · · + h(n) x(i−n) (3.27)
equal to the unit response function.
Applications include:
3.2.4 Example: Unit Hydrographs
• Rainfall-Runoff
Unit hydrographs are used when the shape of a stormwater
• Soil Moisture Movement hydrograph is desired. It is used to predict the stormflow
hydrograph for conditions where one unit of effective pre-
• Contaminant Transport cipitation (net runoff) falls on a watershed during one time
period. The duration, time to peak, and peak discharge
• Groundwater Modeling
are all represented using a unit hydrograph.
The convolution operator is defined using differently A common shape to use is the triangular unit hydro-
for the discrete and continuous cases. For the discrete graph - where the duration of the hydrograph, tc , is the
case, it is: time of concentration within the watershed, and is the base
of the triangle. The peak discharge, Qp , is the height of
Xm
the triangle, and the maximum height occurs at the time
y(t) = h(i) x(t − i) (3.25) to peak, tp .
i=0
The area of the triangle, Q = tc Qp /2, represents the
while for the continuous case, it is: total volume of stormwater runoff. We normally set the
total volume equal to one unit of runoff, hence the name
Z t
unit hydrograph. Note that reducing the time of concen-
y(t) = h(τ ) x(t − τ ) dτ (3.26) tration requires a higher peak in order to maintain a unit
−∞
runoff.
The steps involved are: Other shapes besides a triangle can certainly be used -
the only constraint is that the area under the curve must
1. Multiply unit response function, h(i), by input, x(t), equal zero. The basic concept that a unit of runoff has a
for time step t specific shape that is determined by the watershed.
2. Shift to next hour and repeat Step 1, starting re-
sponse at beginning of input interval
3.3 Problems
3. When all inputs have been multiplied by the Unit
Response Function, add all responses 1. Unit Hydrographs
• Find the outflow hydrograph for:
• An input x = [0, 10, 12, 3, 0 5]
CHAPTER 3. NETWORK MODELS 18

Table 3.3: Example convolution problem

Input, x = [1, 2, 3, 0, 1]

Response function, h = [0.1, 0.4, 0.3, 0.2]

Time
Response to input 1 2 3 4 5 6 7 8
t=1 0.1 0.4 0.3 0.2 0.0 0.0 0.0 0.0
t=2 0.0 0.8 0.6 0.6 0.4 0.0 0.0 0.0
t=3 0.0 0.0 0.3 1.2 0.9 0.6 0.0 0.0
t=4 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
t=5 0.0 0.0 0.0 0.0 0.1 0.4 0.3 0.2
SUM 0.1 1.2 1.2 2.0 1.4 1.0 0.3 0.2

• A response function h = [0.0, 0.5, 1.0, 1.3, 1.2, 3. Extreme Events

0.8, 0.5, 0.2, 0.1, 0.0]
• Find the median (50th percentile), 10-year, and
2. Percolation Model 100-year precipitation depths for the following
data.
• Assume that water flow through a (two-dimensional)
porous medium can be represented as a finite Maximum annual daily precipitation
number of nodes that are either wet or dry. at Athens Ben-Epps Airport.
• Use an node configuration with the following
transition probabilities to model water move- Year Max (inches) Year Max (inches)
ment: 1948 4.05 1976 3.27
1949 2.34 1977 3.83
p3 p3 1950 3.77 1978 2.86
p1 p2 p1 p2 p1 1951 2.82 1979 3.36
p3 p3 1952 2.03 1980 3.67
p1 p2 p1 p2 p1 1953 2.02 1981 2.71
p3 p3 1954 2.02 1982 2.21
1955 1.72 1983 2.74
• where 1956 5.34 1984 2.39
– p1 = 0.4 1957 2.42 1985 1.66
– p2 = 0.2 1958 2.21 1986 4.31
1959 5.47 1987 2.99
– p3 = 0.1 1960 2.63 1988 1.94
• The transition probabilities may need to be ad- 1961 2.84 1989 5.43
justed near boundaries. 1962 2.62 1990 3.16
1963 5.16 1991 2.54
• Assume that the drainage and wetting transi- 1964 4.12 1992 2.38
tions are identical. 1965 2.16 1993 2.42
1966 3.04 1994 7.34
1967 9.93 1995 4.32
1968 2.16 1996 2.11
1969 2.97 1997 3.80
1970 4.00 1998 3.51
1971 2.48 1999 2.20
1972 2.72 2000 2.11
1973 4.48 2001 6.22
1974 2.96 2002 3.54
1975 3.26
Chapter 4

Statistical Models

Statistics is the science of understanding uncertainty.

Table 4.1: Statistical Definitions
Will it rain today? Given that it has not rained for three
months, what is the probability that it might rain in the Frequency vs. Probability - The frequency of an outcome
next week? How does a dam (or ground-water pumping, is the observed number of occurrences based on a finite
wetland construction, timber harvesting) affect stream- sample size, fi = ni /N , while a probability is the like-
flows? What are the health risks from drinking contami- lihood of that outcome based on an infinite sample size,
nated water? These are all questions that are commonly pi = limN→∞ ni /N . We might say that the probability is
the predicted frequency based on complete information.
asked.
While the goal of science is to separate fact from fic- Discrete vs. Continuous Distributions - Discrete distri-
tion, we are often limited to providing statistical measures butions have a countable number of outcomes, while con-
of truth and error. Thus science often rests on the edge of tinuous distributions have a smooth (infinite) number of
outcomes
certainty, not entirely sure, yet nor entirely unsure. Many
of the early studies in statistics were performed by com- Bounded vs. Unbounded Distributions - Bounded dis-
pulsive gamblers who wished to improve their odds of win- tributions have a maximum and/or minimum limit (such
ning. Rather than accept the “roll of the dice”, these in- as the uniform distribution), while unbounded distribu-
tions (such as the normal distribution) are unlimited in
dividuals wanted to better understand the risks they were
their range. Some distributions (such as the lognormal
taking, and place their bets in ways that maximized their distribution) may be bounded on one side, and not on
likelihood of winning. the other.
Heteroscedasticity vs. Homoscedasticity -
4.1 Probability and Statistics Homoscedastic errors are uncorrelated to either
magnitude of x or y, while heteroscedastic errors occur
Probability analysis is used to describe random behavior, when they increase (or decrease) with the magnitude
such as the chance that an event will occur, or the likeli- of x or y, and must be transformed (perhaps by taking
logarithms) to make them homoscedastic.
hood that an event will exceed a certain magnitude. While
much of nature is not entirely random, we can often apply
probability models to natural systems. We can make these
applications more readily in cases where: We can also say that the probability of two events, A and
B, occurring together is just the product of the probability
• Events are independent of each other, and do not of each event:
affect each other. That is, the result of one coin toss
does not affect the following coin toss. P (A ∩ B) = P (A and B) = P (A) · P (B) (4.2)

• Events are stationary - they are not a function of For example, let us assume that the probability of landing
time. That is, heads are not more likely in the morn- either a heads (H) or a tails (T) when a coin is flipped are
ing than in the evening. equal, so that there are two outcomes, with the probability
of obtaining one or the other being p = 0.5.
• Events are identically distributed. That is, the vari-
When the coin is tossed twice, n = 2, there are four
ability of heads is the same under all conditions.
outcomes; H ∩ H, T ∩ T , H ∩ T , and T ∩ H. Each outcome
If these assumptions are satisfied, then we can say that has an equal probability because these are independent
the likelihood of either one event or another occurring is events. The probability of two heads in a row is:
just the sum of the individual events:
P (H ∩ H) = p2 = 0.25 (4.3)
P (A ∪ B) = P (A or B) = P (A) + P (B) (4.1)

19
CHAPTER 4. STATISTICAL MODELS 20

and snow in a day is five percent, P (A ∩ B) = 0.05, then

the probability of getting either rain or snow in a day is:

P (A ∪ B) = 0.3 + 0.1 − 0.05 = 0.35 (4.9)

Conditional probabilities arise when an event, A, may

be more (or less) likely given that another event, B, has
happened. In this case:

P (A)
P (A|B) = P (B|A) (4.10)
P (B)

where P (A|B) means the probability of Event A given

that Event B has already occurred.
For example, if the probability of rain and snow are
Figure 4.1: Venn Diagram of Events A and B
again 30 and 10 percent, respectively, and the probability
of snow given that rainfall has occurred is 20 percent, then
which is the same for landing two tails. The probability the probability of rain given that snow has occurred is:
of landing one of each has two outcomes, so that:  
0.3
P (A|B) = 0.2 = 0.6 (4.11)
P ((T ∩ H) ∪ (H ∩ T )) = P (T ∩ H) + P (H ∩ T ) 0.1

= 2 p2 = 0.5 (4.4) or sixty percent.

We can write this mathematically for any number of tosses,
n, to determine the number of heads, m, and tails, n − m
4.1.1 Sample Statistics
  We are often asked what the outcome of an uncertain event
n is likely to be, such as today’s expected high temperature.
P (H = m, T = n − m) = P (H)m P (T )n−m
m The expected value – equivalent to the mean or average –
  is calculated using:
n
= pn (4.5) Z n
m ∞
1X
E(x) = x̄ = x f (x) dx = xi (4.12)
where −∞ n i=1
 
n n!
= (4.6) where n is the number of observations, and f (x) is the
m m! (n − m)! frequency distribution of x. The frequency distribution
represents the likelihood of individual observation. Just
is the combinatorial operator that accounts for the number
as each observation is weighted by 1/n, the distribution of
of opportunities for getting the same outcome, and where
individual observations will have a weight associated with
n! = n × (n − 1) × (n − 2) × · · · × 1 is the factorial of n.
them that corresponds to their frequency.
For example, we can calculate the probability of get-
The variance of individual observations (when the ex-
ting exactly 5 heads and 5 tails in ten tosses:
pected value is unknown) is calculated using:
   5  5
10 1 1 Z ∞
P (H = 5, T = 5) =
5 2 2 e2=
V (x) = x (x − x̄)2 f (x) dx
−∞
   10
10 1 X n
= = 0.246 (4.7) 1
5 2 = (xi − x̄)2 (4.13)
n − 1 i=1
which means that we have a chance of only about 1 in 4 p
of getting an equal number of heads and tails. where xe = V (x) is the standard deviation of the in-
If events are not independent of each other, then we dividual observations. If the expected value is known,
can still calculate their probability using: E(x) = µ, then we have instead:
n
P (A ∪ B) = P (A) + P (B) − P (A ∩ B) (4.8) n−1 1 X
V (x) = σ 2 = x
e 2
= (xi − µ)2 (4.14)
n n i=1
For example, if the probability of a rainy day is thirty
percent, P (A) = 0.3, the probability of snow is ten per- A case with a known mean is when a coin is tossed - the
cent, P (B) = 0.1, and the probability of getting both rain mean number of heads is always 1/2.
CHAPTER 4. STATISTICAL MODELS 21

The coefficient of variation is the ratio of the standard

Table 4.2: Variation in Student Heights, cm
deviation to the mean, c(x) = x e/x̄. An example would be
Year
the variation in height with age - is there more variation 1990 1991 1992 1993 1994 1995 1996
as people grow taller? Plotting the coefficient of variation 180 182 172 169 158 177 173
would indicate how much variation there is as people age. 190 175 172 165 175 167 182
Can you guess which age has the greatest variation? 166 175 172 173 169 156 184
A hydrologic example of the coefficient of variation is 167 178 174 176 178 162 169
streamflow; low-flow variability is probably much smaller 174 169 187 160 188 179 158
than under flood conditions, but their coefficient of varia- 161 174 175 176 159 165 162
tion may be similar. 172 170 150 175 162 186 192
The covariance between two variables is found using: 157 177 182 188 186 169 152
ZZ ∞ 189 167 175 188 181 165 183
174 170 162 188 160 191 167
C(x, y) = (x − x̄) (y − ȳ) f (x, y) dx dy 195 169 170 162 182 174 171
−∞
172 180 171 184 196 167 174
Xn 169 186 163 162 159 170 161
1
= (xi − x̄) (yi − ȳ) (4.15) 156 175 172 167 178 178 161
n − 1 i=1
156 169 178 186 175 177 169
where f (x, y) is the joint frequency distribution between 174 171 156 155 164 183 177
x and y. The correlation between two variables is: 169 178 172 177 184 187 173
171 173 182 176 181 188 183
C(x, y) 189 170 182 168 187 178 161
r(x, y) = p 179 182 165 178 183 178 196
V (x) V (y)
176 177 168 174 182 183 184
Xn 173 154 179 168 177 179 190
(xi − x̄) (yi − ȳ)
= (4.16) 152 183 181 167 180 179 181
i=1
x
e ye 181 171 179 182 182 172 177
163 166 173 152 199 178 180
Note that if x = y, then C(x, y) = C(x, x) = V (x) and x̄ 172.1 173.7 172.5 172.7 177.0 175.6 174.4
r(x, y) = r(x, x) = 1 x
e 11.3 6.7 8.6 10.1 11.4 8.8 11.3
• x̄ = 174.0
4.1.2 Higher Moments
ē = 1.78
• x
Data is skewed when there are unbalanced high and low
• x
e = 9.7
observations. For example, a stream may have an average √
discharge of 10 L/s, an extreme low flow of 1 L/s (9 L/s • x
e/ n = 1.94
below the mean), and an extreme high flow of 100 L/s (90
L/s above the mean). This is an example of a positive
skew, in that the larger observations are farther from the The fourth-moment about the mean, or kurtosis, is used
mean than the smaller observations. The skew - also called to describe the frequency of low-probability events - both
third moment about the mean - is calculated using: extremely high and low:
Z ∞ Z ∞
Skew(x) = (x − x̄)3 f (x) dx Kurtosis(x) = (x − x̄)4 f (x) dx
−∞ −∞

n
X n
1 1 X
= (xi − x̄)3 (4.17) = (xi − x̄)4 (4.19)
n−1 i=1
n − 1 i=1
Like the coefficient of variation, the skew coefficient,
G(x), is normalized, but rather than using the mean, we 4.1.3 Variation of Sample Statistics
instead use the estimated standard deviation:
The sample statistics calculated above were for individual
n Skew(x) observations. Once we have calculated these statistics,
G(x) =
n−2 e3
x they may still be uncertain. For example, we may have
n
calculated the mean and variance of daily rainfall for each
n X
= (xi − x̄)3
(4.18) month of the year. Yet the rainfall in each month varies
e3 i=1
(n − 1) (n − 2) x from year to year. Clearly, there is still a variation in
monthly averages.
This adjusts the skew by how variable the data are - one
For groups of events, the mean of the group, E(x̄)
needs a greater skew when there is a greater variability to
should equal the mean of individual events within the
arrive at the same skew coefficient.
CHAPTER 4. STATISTICAL MODELS 22

group, E(x):

E(x̄) = E(x) (4.20)

but is not true if the calculated mean is biased.

Bias occurs when the average of a distribution does
not converge to the true mean, usually because certain
outcomes are more likely measured than others. Bias is
relatively common in hydrology because we tend to sample
more heavily in good weather, and avoid taking measure-
ments when the weather is awful.
The variance of the sample mean will be different from
the variance of individual events because the average be-
havior is commonly much less variable than the individual Figure 4.2: Normal Distribution
events.
V (x) x
e For example, let’s say that a river flooded following a
V (x̄) = and ē = √
x (4.21) timber cut. We observe that the flood is far worse than
n n
any observed flood, and conclude that the timber harvest
The variation in the variance is described using the caused the flood. We may then be making a Type 1 error.
χ2 (pronounced ki squared ) statistic, which is mathemat- On the other hand, there may actually be adverse flooding
ically equivalent to a gamma distribution (below), with from timber harvests, but the effects were too small to
β = 2 and α = ν/2, and where ν (pronounced nu) are notice with everything else going on. This is the Type 2
the degrees of freedom, equal to the number of observa- error.
tions minus the number of statistics (such as the mean and In statistical testing, we can assume that observations
variance) that are being estimated. This is equivalent to: are normally distributed - the familiar bell-shaped curve.
If the observation is too far from the mean, then we might
2−ν/2 xν/2−1 e−x/2
2
χ (x, ν) = (4.22) think it is fundamentally different from the other observa-
Γ(ν/2) tions. To check, we first find the standard normal variable:
x − x̄
4.1.4 Hypothesis Testing z= (4.23)
x
e
One reason for estimating statistics is to be able to make and then use this variable to make a decision. Using the
a definitive statement about a situation. Did this timber normal distribution, we can calculate the likelihood of this
harvest cause this flooding? Did this dam destroy this variable. If the probability is too small, then we might
habitat? decide that it does not belong.
To answer a question definitively, we need some pro-
cess to decide whether a chance occurrence is sufficiently
unlikely that we can safely say that it is improbable. For 4.2 Regression
example, we might think the number of heads and tails
should be approximately equal, yet we sometimes observe We normally use a linear regression equation of the form:
twenty heads in twenty tosses. We can calculate the prob-
ability as 0.520 = 10−6 , or one chance in a million. We y = yo + a x (4.24)
might suspect that this coin is not fair. where x is the column vector, (n × 1), of independent
Furthermore, we might say that any coin that has a observations, y is the column vector, (n × 1) of dependent
rare outcome (say, less than one chance per thousand, observations, yo is the intercept, and a is the slope.
10−3 ) should not be used. Using this rejection statistic, Surprisingly, this equation is not a linear function. A
we recommend that this coin should be rejected as being linear function, y = f (x), has the property that:
unfair. In other words, outcomes that are beyond nor-
mal expectation are rejected, while outcomes that behave f (2x) = 2 f (x) (4.25)
normally are not rejected.
There are two errors when this method is used. Even which means that if you double the input, you should dou-
a fair coin has a slight chance of yielding a rare outcome. ble the output. To check to see if our equation, above, is
Thus, by rejecting the coin, we may be making a mistake. linear, we have:
We call the rejection of a fair coin a Type 1 Error. On the
other hand, an unfair coin may still give normal results 2 y = 2 (yo + a x) = 2yo + 2ax 6= yo + a (2x) (4.26)
and not be detected. We call the failure to reject an unfair so that it is clearly nonlinear ! To linearize this function
coin a Type 2 Error. we note that:
CHAPTER 4. STATISTICAL MODELS 23

Matrices can be added and subtracted if they have the

y = E(y) = E(yo + ax) same size. That is, if they have the same number of rows
= yo + aE(x) = yo + ax (4.27) and columns.
 
We make a linear equation by subtracting this second u11 + v11 u12 + v12
equation from the original equation: u + v =  u21 + v21 u22 + v22  (4.34)
u31 + v31 u32 + v32
Y = y − y = yo + ax − (yo + ax)
where the size of both u and v is (3, 2). Matrix addition
= a(x − x) = aX (4.28) is also possible if one matrix has a lower dimension, For
where the new variables, Y = y − ȳ and X = x − x̄, are example, we can perform the operation:
just the original data with their means subtracted.  
u11 + v1 u12 + v1
For multiple independent variables, then we can estab- u + v =  u21 + v2 u22 + v2  (4.35)
lish the multiple regression equation: u31 + v3 u32 + v3
y = ao + a1 x1 + a2 x2 + · · · + am xm (4.29) where the size of v is (m) = (3). Note that it is also
possible to add a vector with a size of (n) = (2) by taking
4.2.1 Matrices the transpose:
 
A matrix is an organized set of numbers. For example, a u11 + v1 u12 + v2
simple, two-dimensional matrix may look like: u + v T =  u21 + v1 u22 + v2  (4.36)
  u 31 + v1 u 32 + v2
u11 u12
u =  u21 u22  (4.30) Matrix multiplication is achieved by multiplying every row
u31 u32 of the left-hand matrix by every column of the right-hand
where the double underlines indicate that there are two matrix. In this case, the number of columns of the left-
subscripts, uij , which refer to the elements in each row hand matrix must equal the number of rows of the right-
and column, respectively. In this case, there are three hand matrix, and the result has a size equal to the number
rows and two columns. of rows of the left-hand matrix and the number of columns
We indicate the size of the matrix using the notation, of the right-hand matrix:
(m, n), where m is the number of rows, and n is the num- (m, n) × (n, p) = (m, p) (4.37)
ber of columns. For our case, the size is (3, 2). A three-
dimensional matrix would have a size of (m, n, p), where    
p represents the number of layers in the matrix. u11 u12   u11 v1 + u12 v2
v1
The transpose of a matrix reverses the orientation of u v =  u21 u22  × =  u21 v1 + u22 v2 
v2
the matrix. That is, the columns are exchanged with the u31 u32 u31 v1 + u32 v2
rows. Using the previous matrix as an example, the trans- (4.38)
pose of u is:
  For this problem, the multiplication is (3, 2) × (2, 1) =
u11 u21 u31
uT = (4.31) (3, 1). The identity matrix, I, is defined as the matrix
u12 u22 u32
which when multiplied by another matrix yields the orig-
where the superscript is used to indicate the transpose. inal matrix:
T

The apostrophe, ′ , is another symbol that is commonly uI = u (4.39)

to indicate the transpose, u′ = uT . The size of uT is
(n, m), or (2, 3) for this example. Note that one only has where the sizes of the u and I matrices are (m, n) and
to switch the location of the indices, uTij = uji , to obtain (n, n), respectively. The identity matrix takes the form:
the transpose.  
Adding, subtracting, multiply, and dividing by a con- 1 0 0
stant, c, to a matrix is obtained by performing the opera- I= 0 1 0  (4.40)
tion to every element of the matrix: 0 0 1
 
c + u11 c + u12 for n = 3. Note that I is always a square matrix, where
c + u =  c + u21 c + u22  (4.32) the diagonal elements all equal one and all the off-diagonal
c + u31 c + u32 elements all equal zero.
  Matrix division is not possible, but we can obtain the
cu11 cu12 desired solution by using the matrix inverse:
cu =  cu21 cu22  (4.33)
cu31 cu32 w = uv (4.41)
CHAPTER 4. STATISTICAL MODELS 24

u−1 w = u−1 u v (4.42) and

 
u−1 w = I v (4.43) C1Y
 C2Y 
v = u−1 w (4.44)  
T  C3Y 
X Y = (n − 1)   (4.52)
where  .. 
 . 
CmY
u−1 u = I (4.45)
where CiY is the covariance between Xi and Y.
4.2.2 Ordinary Least Squares Continuing with our effort to estimate the unknown
regression coefficients, β̂, we now pre-multiply both sides
In regression, we normally have a large number of obser-
by the inverse of the (XT X) matrix, (XT X)−1 , yielding:
vations of x and y from which we wish to estimate the
regression coefficients. To do this easily we can make a set β̂ = (XT X)−1 (XT Y) (4.53)
of vector equations:
The residuals are the difference between the estimated and
Y=Xβ (4.46) observed dependent variables:

where Y is a (n×1) column vector of n observations of the e=Y−Y e (4.54)

dependent variable, X is a n × m matrix of n observations
of each of the m variables, and β is the (m × 1) column The mean of the residuals is zero, E(e) = e = 0, and the
vector of unknown regression coefficients. This is the same fitting error is the standard error of the residuals:
as:
     V (e) = ee2 = eT e (4.55)
Y1 X11 X12 X13 · · · X1m β1
 Y2   X21 X22 X23 · · · X2m   β2 
     The variance-covariance matrix of β̂ is:
 Y3   X31 X32 X33 · · · X3m   β3 
 =    (4.47)
 ..   ..   ..  e2 (XT X)−1
V (β̂) = e (4.56)
 .   .  . 
Yn Xn1 Xn2 Xn3 · · · Xnm βm
The variance-covariance matrix is very important because
Common multiple regression uses ordinary least squares, it contains information about how your model parameters
OLS, to estimate the regression coefficients. To begin, we are correlated. For example, if rainfall and temperature
first pre-multiply both sides by the transpose of X, XT : are inversely correlated, then a negative covariance is ex-
pected. Variables that are highly correlated are called col-
T T
X Y=X Xβ (4.48) inear, and result in large values in the variance-covariance
matrix. Multicolinearity is the general problem of using
where (XT X) is a square matrix, diagonally dominant, variables that are highly correlated to each other, causing
and symmetric, and is closely related to the variance- large prediction uncertainties.
covariance matrix: Predicted outcomes, Yp , for specific prediction condi-
Note that the variance of xj is: tions, Xp (as opposed to observed conditions), is given
by:
Xn
1 2
Vj = Xij (4.49) Yp = Xp β̂ (4.57)
n−1 i=1

and the covariance between Xj and Xk is: and the prediction error is:

Xn V (Y − Yp ) = V (e) + Xp V (β̂) XTp

1
Cjk = Xij Xik (4.50)
n − 1 i=1 e2 + eep 2
=e (4.58)

so that the diagonal of the (XT X) matrix is related to which means that the total prediction error depends on
the variance of each of the variables and the off-diagonal the fitting errors, e
e, along with errors due to parameter
elements are the covariances. errors, e
e p .
  The confidence intervals about Yp are given by:
V1 C12 C13 · · · C1m
 C21 V2 C23 · · · C2m  Yp ± t(n, α) V (Y − Yp )1/2 (4.59)
 
XT X = (n−1)  .  (4.51)
 . . 
Cm1 C2m C3m ··· Vm
CHAPTER 4. STATISTICAL MODELS 25

 Pn 2
Pn Pn Pn 
Pni=1 X1i Pni=1 X1i X2i Pni=1 X1i X3i ··· Pni=1 X1i Xmi
 2 
 i=1 X2i X1i i=1 X2i i=1 X2i X3i ··· i=1 X2i Xmi 
XT X =  . .. .. 
.
 . .P .P 
Pn Pn n n 2
i=1 Xmi X1i i=1 Xmi X2i i=1 Xmi X3i ··· i=1 Xmi

4.2.3 Transformations Polynomial Model

Ordinary least squares, the primary approach for form- A general polynomial model is of the form:
ing a regression equation, assumes that the data are ho-
moscedastic, meaning that the errors are constant, and y = ao + a1 x + a2 x2 + a3 x3 + · · · + an xn (4.66)
are independent of the magnitude of the observation. In
To make this linear, we create new variables, x1 = x,
reality, many stream discharge measurements are variable
x2 = x2 , etc., and then use these instead:
- a function of the magnitude of the observation. These
are heteroscedastic errors, which must be eliminated by y = ao + a1 x1 + a2 x2 + a3 x3 + · · · + an xn (4.67)
transforming the data prior to performing a regression.
and estimate the coefficients using OLS.
Exponential Model
Power Model
A linear reservoir drains according to the following expo-
nential equation: A commonly used multiplicative, or power, model is:

y = yo exp(−kt) = yo e−kt (4.60) y = ao xa1 1 xa2 2 xa3 3 · · · xamm (4.68)

It is common that any errors in tail are much smaller than To linearize this, we again use the logarithmic transform,
errors near peak - that is, the magnitude of any errors are a Y = ln y, X1 = ln x1 , etc., yielding:
function of the magnitude of the observation. This implies
that the errors are heteroscedastic. Recall that Y = Ao + a1 X1 + a2 X2 + a3 X3 + · · · am Xm (4.69)
For this exponential model, taking the log of the data
where Ao = ln ao .
eliminates heteroscedasticity:

ln(y) = ln(yo ) − kt (4.61) Examination of Residuals

One method for determining whether a transformation is
which is now a linear model between ln(y) and t:
needed is to examine the residuals. To begin, we first
Y = Yo − kt (4.62) make a prediction using a linear model, ŷ = βx, and then
calculate the residuals, e = ŷ − y, using the difference
where Y = ln y and yo = eYo . between the predicted and observed dependent variables.
We then plot these residuals against the observed depen-
Log-Log Model dent variable, y vs e. Finally, we check to see if there
is any correlation between y and e. If so, then we must
A non-linear reservoir model, such as a weir, has the gen- transform of y until the correlation with the residuals is
eral form: eliminated.

y = yo xn (4.63)
Example: Watershed Characteristics
which can be linearized by taking the logarithm of both Empirical models can be used, such as the Benson model,
sides: can be used to estimate peak flows, Qp :
ln y = ln yo + n ln x (4.64) Qp = a Ab S c Std Ibe Lf (4.70)

Introducing new variables, Y = ln y and X = ln x, yields where A is the watershed area, S is the main channel
the general linear equation: slope, St is the area of lakes and ponds, Ib is the max-
imum 24-hour, 10-year precipitation depth, and L is the
Y = Yo + n X (4.65) main channel length.
Another model, proposed by Scott, is:
where Yo = ln yo or yo = eYo .
Qp = a Ab Em
c
Std Pse I f T g (4.71)
CHAPTER 4. STATISTICAL MODELS 26

where Em is the mean altitude, Ps is the average May to and

September precipitation depth, I is the maximum 24-hour,
∆Ji
2-year precipitation depth, and T is the mean January gi = − (4.77)
temperature. ∆β
And finally, the Borland model is: This approach means that we update the parameters by
b c d e f g
Qp = a A Em Sh S St Pa Ps I La Loh j j k
(4.72) examining what the prediction errors are and coupling
that with coefficients that tell which parameters are most
where Sh is the watershed shape, Pa is the average October useful for minimizing the error.
to April precipitation depth, La is the latitude, and Lo is
the longitude. Kalman Filtering. Kalman Filtering is another means
These models are formed by using peak stormflows at for updating your model. In this case we have system
gages where watershed information is available. Normally, outputs, Y , and state variables, S. We say that we can
these models are regional - in that they can only be applied predict the outputs based upon predictions of the state
in the area where they were developed. variables, which are only approximately known, using:
To form you own model, logarithmic transforms are
used to linearize the equations: Yi,p = H Si,p (4.78)
y = yo + b x1 + c x2 + d x3 + · · · (4.73) We predict the state variables for the next time step based
where y = log Qp , yo = log a, x1 = log(A), etc. Note on the state variables from the previous time step using:
that any number of factors can be added, thus improving
Xi,p = F Si−1,p (4.79)
the fitting or calibration error. Multicollinearity is ma-
jor problem with this approach, however, causing large Once we have an observation of output, Y , we find the
i
prediction errors. One must always check the prediction error between that and our prediction:
uncertainties to make sure that each added parameter fur-
ther decreases the prediction error. ei = Yi,p − Yi (4.80)

4.2.4 Factor Analysis We update the state variables using:

Factor analysis, and its cousin Principal Components Anal- Si = Si,p + K ei (4.81)
ysis (PCA), are commonly used to identify the relation-
ships between variables using the covariance or correlation where K is the Kalman gain matrix:
matrix. If C is the variance-covariance matrix, then we  −1
can form the equation: Ki,p = Pi,p H T R + H Pi,p H T (4.82)

λC = DC (4.74) where R = C[e] and E = C[η] in which e = yi,p − yi ,

η = Si,p − Si , and P is the covariance matrix associated
where λ are the eigenvalues of C and D are the eigenvec-
with the estimation error of the states, which is dynamic
tors, with one eigenvector per eigenvalue. The eigenvalues
and is predicted using:
correspond to the relative importance of a linear set of
combinations of each of the original variables in explain- Pi,p = F P + i − 1 F T + Q (4.83)
ing the correlation between these variables.
The final value of this covariance matrix is:
4.2.5 Model Updating  −1
Pi = Pi,p − Pi,p H T R + H Pi,p H T H P i,p (4.84)
We can establish the regression equation Y = X β between
model outputs, Y , and model inputs, X. Our P objective in
(yo − yp ) . 4.3 Time Series Analysis
2
regression is to minimize the errors, J =
Note that the predicted values of y, yp , are a function of
model parameters, β. A time-series model can be used to forecast variables that
A recursive estimation can be used to update the model change with time. One can use previous observations of
parameters, β, as the system changes (such as happens a time series to predict future values, called an autore-
when floods change the cross-sectional area of the channel, gressive model. Adding previous prediction errors to the
a dike breaks, etc.): model can help improve forecasts, called a moving average
model. Other variables, called external inputs often help
∆βi = (pi )−1 gi (4.75) to predict the forecast. And finally, converting the ob-
served variables by either differencing or integrating them
where
also can improve the prediction accuracy.
pi = pi−1 + (X T X) (4.76)
CHAPTER 4. STATISTICAL MODELS 27

4.3.1 Autoregressive (AR) 4.3.5 Integrated (I)

The autoregressive model uses previous observations to Yet other processes might be described using cumulative
predict future observations: observations, such as using cumulative inflows to predict
reservoir volumes:
ŷ(t) = yo + a1 y(t − 1) + a2 y(t − 2) + . . . + an y(t − n)
n
X Y (t) = Y (t − 1) + y(t) (4.90)
= yo + ai y(t − i) (4.85)
i=1 The resulting time-series model could be a IAR(n), IARMA(n,m)
where n is the memory of the autoregressive system. This or IARMAX(n,m,p), which would just use the Y instead
is often written using AR(n). of y in the prediction equations.

4.3.2 Moving Average (MA) Partial Correlations. Partial correlations refers to the
marginal contribution of the next coefficient. For example,
The moving-average model adds the previous prediction if you have an AR(1) model, find the net correlation of the
errors, meaning that if one is off by a certain amount on next autoregressive term. If there is no partial correlation,
the last time step, then the error, e(t) = ŷ(t) − y(t), is and if r1 = 0.9, then r2 = r12 = 0.9 × 0.9 = 0.81 . If we
added to the next prediction: find r2 > r12 , then there is a positive partial correlation.
We can also say that if r2 < r12 , then we have a negative
ŷ(t) = yo + b1 e(t − 1) + b2 e(t − 2) + . . . + bm e(t − m)
partial correlation.
Xm
In general:
= yo + bi e(t − i) (4.86)
i=1 rab − rbc rac
rab|c = p p
2
(4.91)
where n is the memory of the moving-average error. This 2
1 − rac 1 − rbc
is often written as MA(m), or ARMA(n,m) if the moving-
average model is combined with the autoregressive model. For example, let rab = 0.79 and rac = rbc = 0.90, then
0.79 − (0.90) × (0.90)
4.3.3 External Inputs (X) rab|c = √ √ = − 0.10 (4.92)
1 − 0.902 1 − 0.902
Another time-series approach is to use external inputs,
such as precipitation or upstream flows, to predict down- 4.3.6 Example: Salt River Project
stream events. This takes the form:
Our objective is to determine how much water to release
ŷ(t) = yo + ho x(t) + h1 x(t − 1) + · · · + hp x(t − p) from three reservoirs. Data are collected every six hours
p
and sent within a few minutes to the central office. A
X decision must be made whether to open up the spillways or
= yo + hi x(t − i) (4.87)
i=1
to adjust power production. We want a prediction model
for prediction horizons of 6, 12, 24 and 96-hours ahead.
where p is the memory of the external input. This is
We have data from ten raingages, 16 stream gages, and
the same as the convolution operator, where hi is the
three reservoirs.
unit response function. This is written as X(p) or AR-
Based on a time-series analysis, we find that:
MAX(n,m,p) when combined with the ARMA(n,m) model:
Xn 1. The nearest stream gage upstream of each reservoir
ŷ(t) = yo + ai y(t − i) is the best predictor for 6-hour ahead forecast.
i=1
2. Using the most upstream stream gage plus precipi-
m
X p
X tation is found to provide the best 96-hour forecast.
+ bi e(t − i) + hi x(t − i) (4.88)
i=1 i=1 3. One M A term was sufficient, two AR terms were
generally necessary.
4.3.4 Differenced (D)
4. Recession flows are easier to predict than peaks.
Some processes are better described using their changes,
such as the change in water level in a wetland. In this 5. Poor forecasting during cold weather often results
case, we take the first difference of the data: due to lack of temperature information, which makes
it difficult to distinguish between rain and snow pre-
∆y(t) = y(t) − y(t − 1) (4.89) cipitation.
The resulting time-series model could be a DAR(m), DARMA(n,m),
or DARMAX(n,m,p), which would just use the ∆y instead
of y in the prediction equations.
CHAPTER 4. STATISTICAL MODELS 28

4.4 Problems
1. Regression Analysis
(a) Find and download water quality data for a
minimum of five variables.
(b) Calculate the means, standard deviations, stan-
dard errors of the means, coefficient of varia-
tion, and correlation and covariance matrices.
(c) Select one of the variables as the dependent
variable and find the regression relationship us-
ing both the original and log-transformed val-
ues. Plot the observed and predicted values
against each other.
(d) Calculate the eigenvectors and eigenvalues. Plot
the two most significant eigenvectors against
each other.
2. Time Series Analysis
(a) Download three time-series from the U.S. Ge-
ological Survey water data website, preferably
three stations within the same watershed
(b) For the most upstream location, find the AR(2)
model using both the original and log-transformed
discharges. Plot the predicted and observed
discharges.
(c) For the most downstream location, find the ARX(2,p)
model using the other two stations. Plot the
predicted and observed discharges.
Chapter 5

Differential Equations

Walk through the woods or down a street. Is every-

Table 5.1: General differentiation rules.
thing the same? Do you see the clouds move? The streams
flow? The fountains spout? Describing the variation in Function Derivative
our world has always been done using words. In fact, po- u = f (x) u′ = du/dx
ets have mastered the task of putting meaning into these
c 0
words.
cx c
Besides poetry, another way of describing our environ-
cu c u′
ment uses mathematical expressions. Just as rhyming and
xc c xc−1
meter are important to the poet, differential equations c
u c uc−1 u′
are the mathematical tools we use for precisely defining
uv u v ′ + v u′
change. These differential equations are frequently used to
u/v (v u′ − u v ′ )/v 2
formulate mathematical descriptions of our world. While
ln u u′ /u
algebraic equations are commonly used to relate variables
exp u u′ exp u
to other observed values, differential equations are used to
sin u u′ cos u
incorporate changes in observed values, such as changes in
cos u −u′ sin u
time and space.
An ordinary differential equation, ODE, describes how
one or more variables change as a function of one other In both cases, the water levels only change as a function
variable. In the previous section, the velocity is written of time or position, but not both.
as an ordinary differential equation, which means that the If the water level varies with both time and position,
change in distance is a function of a single variable, in this then we must write a partial differential equation, PDE:
case time.
For example, water levels, h, may change in a linear ∂h ∂h
=m (5.6)
manner with respect to time, t: ∂x ∂t
h = ct + b (5.1) where the partial derivative means that the variable is
held constant for all other variables than the one used for
where c and b are constants. The derivative, or change in
taking the derivative. In this case, the left-hand side is
water levels, with respect to time would be:
calculated for a specific time, ti while the right-hand side
dh is calculated at a specific position, xj :
ḣ = =c (5.2)
dt  
If, instead of time, water levels change with respect to (h′ )ti = m ḣ (5.7)
xj
distance, x, we have:
h = mx + b (5.3) A few basic differential rules are provided in Table 5.1.
where m and b are again constants. The derivative, in this
case, would be: 5.1 System Specification
dh
h′ = =m (5.4) 5.1.1 Types of Differential Equations
dx
If water levels in more than one aquifer are observed, then The general, one-dimensional equation for steady flow in
we might have multiple derivatives with respect to time or saturated geologic media is given by Darcy’s law:
position, i.e.,
∂h
ḣ1 = m ḣ2 or h1 ′ = m h2 ′ (5.5) ~q = −Kh′ = −K (5.8)
∂x

29
CHAPTER 5. DIFFERENTIAL EQUATIONS 30

which, when combined with the conservation of mass equa- 5.1.2 Initial and Boundary Conditions
tion,
Regardless of the type of differential equation, specific ini-
~q = −θ̇ (5.9) tial and boundary conditions must be prescribed to make
the system uniquely defined. Failure to completely specify
yields: these conditions may make the system ill-posed, meaning
that more than one solution may be possible for the prob-
∂θ
K∇2 h = (5.10) lem.
∂t Initial conditions refer to the values of the observed
variable(s) at a prescribed time, such as h = ho at t =
where θ is the water content of the aquifer. This is the
to . Boundary conditions refer to the specification of the
same as:
observed variable(s) at a prescribed location, so that h =
∂2h ∂h hi at x = xi .
K 2 = Ss (5.11)
∂x ∂t There is great flexibility in establishing initial and bound-
ary conditions. Some of the many types of boundary con-
because Ss = dθ/dh. This can also be written as: ditions include:
Dh′′ = ḣ (5.12) Prescribed Head: Used to set a single value of head, h,
for the boundary, such as h = hi at x = xi .
where D = K/S is the hydraulic diffusivity.
For steady (and also for incompressible) flow this is Prescribed Flux: Used to specify the flux, q = −Kh′ ,
just: for the boundary, such as q = qi at x = xi . These
can be either constant or change with each time step.
Kh′′ = 0 (5.13)
Prescribed Gradient: Used to specify the gradient, h′ ,
For higher dimensions, we have: such as h′ = h′i at x = xi .
 
∇ · ~q = ∇ · −K∇h = −Ss ḣ (5.14) Radiation: Used to set the jump, ∆h, across a boundary,
such as ∆h = ∆hi at x = xi .
where K is a second-order tensor. If K is isotropic and
stationary, this reduces to: Mixed: Combines the prescribed head and flux boundary
conditions, a hi + b h′i = c.
 2 2

∂ h ∂ h ∂h
K + 2 = Ss (5.15) It is important to note that these initial and boundary
∂x2 ∂y ∂t
conditions can be changed for each boundary segment for
in two dimensions. each time step, or they can remain constant.
The general one-dimensional wave equation is:
5.1.3 Material Properties
∂2h ∂2h
2
= 2 (5.16) The selection of the appropriate model parameters is part
∂x ∂t
of the conceptual site model development process. There
These various equations can be written in general form are a wide range of material properties that must be de-
as: termined when developing the model. Specific properties
include the roughness, hydraulic conductivity, transmis-
ah′′ii + 2bh′′ij + ch′′jj + dh′i + eh′j + f h + g = 0 (5.17) sivity, leakance, conductance, which are descriptors of the
permeability of the system. The storage terms are also
where hi = ∂h/∂xi , etc. We can determine the type of
important for transient problems, such the specific yield,
equation using the value of λ = b2 − ac, where:
specific storages, storativity, water content, porosity, and
Hyperbolic, λ > 0: A wave equation with the form hii = specific water capacity. Combined metrics such as the
hjj . aquifer diffusivity are also important.

Parabolic, λ = 0: An unsteady flow equation with the

form hii = hj .
5.1.4 Example: Nash Model
The Nash model is commonly used to relate stream dis-
Elliptic, λ < 0: A two-dimensional flow equation with the charge, Q, to precipitation, P , or to account for flood
form hii + hjj = 0. routing between two stations in a channel, Q1 and Q2 .
We first start by defining a storage volume, S = A h,
where h is the depth of water, and A is the area of the
CHAPTER 5. DIFFERENTIAL EQUATIONS 31

storage. We can see that the volume of water in storage

Table 5.2: General integration rules.
increases linearly with the depth of water in storage:
dS Function RIntegral
=A (5.18) f (x) f (x) dx
dh
c cx
The change in storage over time, Ṡ = ∂S/∂t, depends cx c x2 /2
upon the relative magnitude of inflows, I, and outflows, xc (cR+ 1) xc+1
O: cu c u dx
ln x x ln x − x
Ṡ = I − O (5.19) exp x exp x
sin x − cos x
Combining these last two equations yields:
cos x sin x
A ḣ = I − O (5.20)
For the case where there are no inflows, I = 0, we have:
as long as the area does not change with depth and time,
A 6= f (h, t). dh
We can go a step further and assume that the outflow A = −k h (5.24)
dt
from the reservoir depends upon the height of water in
Moving terms yields:
storage:
dh k
O = kh (5.21) = − dt (5.25)
h A
which implies that the rate of outflow increases as the Integrating both sides is:
water depth in storage increases. Substitution yields: Z Z
dh k
=− dt (5.26)
k I h A
ḣ + h = (5.22)
A A which is the same as:
which is an ordinary differential equation. k
ln h = c − t (5.27)
A
5.2 Analytic Solution Methods where c is the constant of integration. Taking the expo-
nent of each side yields:
Many simple differential equations can be solved using a  
k
range of methods, including direct integration as well as a h = exp c − t = C e−αt (5.28)
range of transform methods. Yet, solving PDEs for a range A
of initial and boundary conditions and material properties where C = exp c and α = k/A. To identify the value of C,
can be a challenge. This is especially true if material prop- we must have a value of h for a specific time, t . In this
i i
erties vary spatially or with time. This section focuses on case, we can use h = h when t = 0, which is an initial
o
analytic methods for solving relatively simple problems, condition. This results in the final form:
while the following section introduces numerical methods
that are more appropriate for more complex systems. h = ho e−αt and O = Oo e−αt (5.29)

where O = A h and Oo = A ho is the initial outflow.

5.2.1 Integration For the case where the inflow is a non-zero constant,
The purpose of integration is to convert a differential equa- we have:
tion into an algebraic equation that can be directly solved. dh I
A general list of integration rules are proved as Table 5.2. + αh = (5.30)
dt A
An excellent website for obtaining integrations, if possible,
can be found at integrals.wolfram.com. where α = k/A. A common technique for solving ODEs of
We return to the Nash model to demonstrate how a this kind is to multiply both sides by an integrating factor,
αt
simple differential equation can be solved using integra- e :
tion. Recall that the linear reservoir can be described dh
using: eαt + α eαt h = I eαt (5.31)
dt
dh We use the integrating factor, because the chain rule is:
A = I −kh (5.23)
dt d(u v) dv du
=u +v (5.32)
dt dt dt
CHAPTER 5. DIFFERENTIAL EQUATIONS 32

which, for our case, is:

Table 5.3: Laplace transforms
d(h eαt ) dh αt
= h α eαt + e (5.33) f (x) F (s) = L (f (x))
dt dt
which is exactly what we need! Substitution yields: δ(t) 1
1
d(h eαt ) I 1
= eαt (5.34) s
dt A 1
t s2
so that:
tn−1 1
I Γ(n) sn
αt
d(h e ) = eαt dt (5.35)
A eat 1
s−a
Integration yields: 1 at 1
a (e − 1) s(s−a)
1 I αt
h eαt = c + e (5.36) f ′ (t) s F (s) − f (0+ )
α A
f ′′ (t) s2 F (s) − s f (0+ ) − f ′ (0+ )
where c is the constant of integration. Dividing both sides
by the integrating factor, substituting Aα = k, and insert-
ing initial conditions yields: Returning to the Nash model, we see that:

I I
h = ho e−αt + 1 − e−αt (5.37) h′ (t) + αh(t) = (5.43)
k A
5.2.2 Laplace Transform Methods is the same as:
I
An alternative method for solving ordinary differential [sH(s) − ho ] + α H(s) = (5.44)
equations is to employ a transform method that converts As
differential equations into algebraic equations. One of the so that:
most straight-forward types of transforms is the Laplace I
method. H(s) (s + α) = ho + (5.45)
As
The Laplace transform is defined as:
Z ∞ and:
−st ho I
F (s) = L (f (t)) = e f (t) dt (5.38)
o H(s) = + (5.46)
s + α A s (s + α)
While one may calculate the transform directly, it is usu-
Taking the inverse for each term yields:
ally easier to consult a table of transforms that already
exist. Table 5.3 presents Laplace transforms of some basic I

h(t) = ho e−α t − e−α t − 1 (5.47)
functions. Aα
A very useful feature of the transform is their linearity,
which is equivalent to the result obtained earlier.
i.e.:

c1 f1 (t) + c2 f2 (t) = c1 F1 (s) + c2 F2 (s) (5.39) Laplace Transform Example

and the fact that derivatives are converted into algebraic For a single linear reservoir, we have:
equations: I
h′ (t) + αh(t) = (5.48)
A
f (n) (t) = sn F (s)−sn−1 f (0+ )−sn−2 f ′ (0+ )−· · ·−f n−1 (0+ )
which is the same as:
(5.40)
I
Note that the initial conditions are explicitly incorporated [sH(s) − ho ] + α H(s) = (5.49)
As
during the transform. For integrals, we have:
so that:
Z ∞
F (s) I
f (τ ) dτ = (5.41) H(s) (s + α) = ho + (5.50)
o s As
and for convolution: and:
Z ∞
ho I
f (τ ) g(t − τ ) dτ = F (s) G(s) (5.42) H(s) = + (5.51)
o s + α A s (s + α)
CHAPTER 5. DIFFERENTIAL EQUATIONS 33

Taking the inverse for each term yields: If the outflow from reservoir 1 is the inflow to reservoir 2,
then we have:
I

h(t) = ho e−α t − e−α t − 1 (5.52)
Aα O2 = H2 ⋆ O1 (5.66)
which is equivalent to the result obtained earlier. or, in general:
Let us consider the dirac (impulse) input:
 Oi = Hi ⋆ Oi−1 (5.67)
∞ t = to
δ(to ) = (5.53)
0 t 6= to which is the same as:

where the dirac has the additional constraint: Oi = Hi ⋆ Hi−1 ⋆ · · · ⋆ H1 ⋆ I1 (5.68)

Z ∞
δ(to ) dt = 1 (5.54)
−∞ 5.3 Numerical Solution Methods
For a dirac input, we have:
5.3.1 Finite Difference Methods
O1 = H1 ⋆ δ(to ) = H1 (5.55) Numeric models are popular because they provide the flex-
ibility for solving a broader class of problems where simple
Taking the Laplace Transform of the single reservoir with
analytic answers may not be available because of varia-
a dirac input yields:
tions in material properties or boundary conditions.
Oi = Hi ⋆ Hi−1 ⋆ · · · ⋆ H1 (5.56) Numeric methods using subdivide the problem into
small, homogeneous units which are individually or glob-
δ(to ) ally solved, and then added back together to provide the
h′ (t) + αh(t) = (5.57)
A answer for the whole system.
which is the same as: The finite difference is, intuitively, a straightforward
approach for solving differential equations. We approx-
1
[sH(s) − ho ] + α H(s) = (5.58) imate the solution using a linear approximation to the
A derivative:
so that: ∂h ∆h
ḣ(t) = ≈ (5.69)
1 ∂t ∆t
H(s) (s + α) = ho + (5.59)
A ∂h ∆h
h′x = ≈ (5.70)
and: ∂x ∆x
ho + 1/A ∂h ∆h
H(s) = (5.60) h′y = ≈ (5.71)
s+α ∂y ∆y
For problems in which the second derivative is needed, we
Taking the inverse transform yields: have:
 

1 ∂2h ∆ ∆h ∆2 h
h(t) = ho + e−α t + C (5.61) ′′
hxx = ≈ ∆x
= (5.72)
A ∂x2 ∆x (∆x)2
where C is the constant of integration. Setting h(t) = ho  
∆h
when t = 0 yields: ∂ 2
h ∆ ∆y ∆2 h
′′
hyy = ≈ = (5.73)
∂y 2 ∆y (∆y)2
1
C=− (5.62) Note that these approximations are better where the de-
A
gree of curvature in the function is small. Clearly, taking
which yields: large steps to measure the circumference of a circle is worse
1
than taking very small steps.
h(t) = ho e−α t + 1 − e−α t (5.63) We will use the notation:
A
For a cascade of linear reservoirs, we have the response h(xi , yj , tk ) = hkij (5.74)
for the first reservoir:
to identify the time, t, and location, (x, y), where h is
O1 = H1 ⋆ I1 (5.64) observed. We further define the change in h with respect
to x while holding y and t constant using:
and for each subsequent reservoir:
∆h hki,j − hki−1,j
Oi = Hi ⋆ Ii (5.65) h′i = = (5.75)
∆x xi − xi−1
CHAPTER 5. DIFFERENTIAL EQUATIONS 34

and
∆h hki,j − hki,j−1
h′j = = (5.76)
∆y yj − yj−1
k−1
k ∆h hki,j − hi,j
ḣ = = k (5.77)
∆t t − tk−1
The second derivative is calculated by taking the difference
in first derivatives:
hk k
i+1,j −hi,j hk k
i,j −hi−1,j
∆2 h xi+1 −xi − xi −xi−1
h′′i = = (5.78)
(∆x)2 (xi+1 −xi )+(xi −xi−1 )
2

which corresponds to the change in the change in the func- Figure 5.1: Linear Reservoir Definition Sketch.
tion. For a constant increment, ∆x, this reduces to:
hki+1,j − 2 hki,j + hki−1,j 5.3.2 Finite Element Method
h′′i = (5.79)
(∆x)2
5.3.3 Boundary Element Method
Returning, yet again, to the Nash model, we can ap-
proximate the problem: 5.3.4 Analytic Element Method

ḣ + α h =
I
(5.80) 5.4 Example: Linear Reservoir
A
using: The linear reservoir model is a simple model of water flow
in channels. Consider a simple tank reservoir with a vol-
hk+1 − hk i ume of water, V = h A, equal to the depth, h, and the
+ α hk = (5.81)
t k+1 −t k A area of the tank, A. Let us suppose that there is a valve
that releases water from the bottom of the tank, and the
This is called the forward model because we are predicting
outflow from the tank Q = α h, is a function of how far
the first derivative using existing values of h. Solving for
open the valve is, α, and how high the water level is in the
hk yields:
tank, h. As the water level drops, the outflow from the
 
I
tank should decrease.
hk+1 = hk + − α hk tk+1 − tk (5.82) We can also say that the outflow must be due to a
A
release of water from storage within the tank, or:
which, for a constant time step, ∆t, simplifies to:
dV dh
k+1 k I Q=− = −A (5.88)
h = h (1 − α ∆t) + ∆t (5.83) dt dt
A
which just means that the outflow must equal the time
We could just as easily have defined the problem using:
rate of change of the volume of water in the tank, which
hk+1 − hk I is related to the rate of change in water level in the tank.
+ α hk+1 = (5.84) Combining both equations yields:
∆t A
which yields a slightly different equation: dh
Q = α h = −A (5.89)
I dt
k+1 hk + A ∆t
h = (5.85)
1 + α ∆t Rearranging terms yields:
This form is called the backward model because we are α dh
predicting using the future values of h. dt = − (5.90)
A h
The final, and most accurate form, is the time-centered
model: Integration yields:
 
hk+1 − hk hk+1 + hk I αt
+α = (5.86) + C = − ln h (5.91)
∆t 2 A A
which results in: where C is the constant of integration. This can be sim-
α ∆t

I
plified to:
hk 1 − 2 + ∆t
hk+1 = A
(5.87)
1 + α 2∆t h = ho exp(−βt) (5.92)
CHAPTER 5. DIFFERENTIAL EQUATIONS 35

Figure 5.3: Linear reservoir with a constant input.

Figure 5.2: Cascade of linear reservoirs.
A common technique for solving ODEs of this kind is
where ho = h(t = 0) and β = α/A. Substituting into the to multiply both sides by an integrating factor, eαt :
equation above yields: I αt
ḣ eαt + α h eαt = e (5.98)
Q = α h = Qo exp(−βt) (5.93) A
We use the integrating factor because the chain rule is:
where Qo = Q(t = 0). These are exponential decay equa-
tions, which means the values diminish at a slowing rate d(u v) du dv
over time. Such a curve is typical of first-order streams, = v+u (5.99)
dt dt dt
i.e., streams that are the smallest permanent streams in a
landscape. which, for our case, is:
We can imagine a second order stream as being a sec-
d(h eαt )

ond reservoir located below the first reservoir. In this case, = ḣ eαt + h αeαt (5.100)
dt
the outflow from the first reservoir is the inflow to the sec-
ond, and the outflow from the second is the inflow to the Substitution yields:
third. This series of reservoirs is called a cascade of linear
reservoirs, or Nash model. The equation that describes d(h eαt ) I αt
= e (5.101)
these can be readily found using the Laplace transform dt A
method, giving a gamma function that looks just like many
so that:
flood waves:
I αt
τ n−1 β n exp(−βτ ) d(h eαt ) = e dt (5.102)
h(τ ) = (5.94) A
(n − 1)!
Integrating both sides yields:
where n is the number of reservoirs in series, or, equiva-
lently, the expected response in an nth -order stream. 1 I αt I
h eαt = e = eαt + C (5.103)
For a number of linear reservoirs in parallel, we have: α A k
Dividing both sides by the integrating factor, and inserting
Q = Q1 exp(−β1 t) + Q2 exp(−β2 t) (5.95)
the initial condition, h = ho at t = 0, yields:
+Q3 exp(−β3 t) + · · · + Qn exp(−βn t) I
h = ho e−αt + (1 − e−αt ) (5.104)
We can also consider the case when a constant input, k
I, such as upstream inflows, or precipitation, are present
dV
5.5 Example: Channel Routing
= A ḣ − I (5.96)
dt 5.5.1 General Routing Equation
where ḣ = dh/dt, so that Rather than use simple conceptual models, we can formu-
late a fundamental physical model that describes channel
A ḣ + k h = I (5.97) flows. Once this physical model has be constructed, we can
then apply it for the routing flood flows through streams.
CHAPTER 5. DIFFERENTIAL EQUATIONS 36

The first equation we use is the conservation of mass

equation:
∆S
∆Q = O − I = − (5.105)
∆t
Where ∆Q is the change in discharge along the channel,
O is the channel outflow, I is the channel inflow, S is the
amount of water in storage in the channel section, and
∆S/∆t is the change in storage over time. This equation
means that when inflows exceed outflows, the amount of
water in storage must increase over time.
We can write this as a one-dimensional partial differ- Figure 5.4: Sketch diagram for Muskingum storage.
ential equation (PDE):
∂Q ∂A
=− (5.106) 5.5.2 Nonlinear Outflow-Storage Equation
∂x ∂t
where the left-hand side represents the change in flow per One approach is to use a nonlinear relationship between S
unit length of channel, and the right hand side represents and O, which would be consistent with the weir equation,
the change in channel cross-sectional area, A, per unit O = C S n , where C is the weir coefficient, and n = 3/2
time. for a rectangular weir. Solving for S yields the nonlinear
We can approximately this equation using finite differ- storage-outflow relationship, S = (O/C)2/3 .
ences: Substituting the storage-outflow relationship and solv-
ing for the future outflow yields:
∆Q ∆A
=− (5.107)
∆x ∆t O(t + ∆t) = I(t + ∆t) + I(t) (5.114)
so that:  m 
2 O(t + ∆t)
∆S −O(t) − − S(t)
∆Q = − (5.108) ∆t C
∆t
This must be solved iteratively, using an initial estimate
where ∆Q = O − I and ∆S = ∆A ∆x. Note that this is of O(t + ∆t) on the right-hand side to obtain a revised
identical to the original equation.
estimate on the left-hand side, which is then used on the
Using the original equation, we can define inflows, out-
right-hand side until the value converges.
flows, and storages as the averages between two times:
O(t + ∆t) + O(t) 5.5.3 Muskingum Method
Ō = (5.109)
2
An alternative, early approach for estimating the storage
I(t + ∆t) + I(t)
I¯ = (5.110) function was an empirical method that splits the storage
2 into two parts, a steady, prism storage, and a transient,
∆S = S(t) − S(t + ∆t) (5.111) wedge storage:
Substitution yields:
S = Sprism + Swedge (5.115)
O(t) + O(t + ∆t)
(5.112) where
2
I(t) + I(t + ∆t) Sprism = τ O (5.116)
−
2
S(t) − S(t + ∆t) Swedge = x τ (I − O) (5.117)
=
∆t where τ is the flood wave travel time and x is a weighting
If we are monitoring streamflows, we have the current ob- factor, 0.1 < x < 0.5, that indicates the amount of flood
servations of I(t), O(t), and S(t). We can also obtain attenuation. Substitution yields:
predictions for the upstream inflows, I(t + ∆t). Solving
for the future downstream outflow yields: S = τ O + x τ (I − O) = τ [x I + (1 − x) O] (5.118)

O(t + ∆t) = I(t + ∆t) + I(t) (5.113) Substituting this into the finite difference equation yields:
2 O(t + ∆t) = C1 I(t + ∆t) + C2 I(t) + C3 O(t) (5.119)
−O(t) − [S(t + ∆t) − S(t)]
∆t
But we are still have one unknown on the right-hand where
side of the equation, S(t + ∆t). To eliminate this variable, ∆t − 2τ x
we require a relationship between S and O. C1 = (5.120)
∆t + 2τ (1 − x)
CHAPTER 5. DIFFERENTIAL EQUATIONS 37

∆t + 2τ x
C2 = (5.121) and so on. Substituting these into the original Muskingum
∆t + 2τ (1 − x) equation yields:
−∆t + 2τ (1 − x) n
X
C3 = (5.122)
∆t + 2τ (1 − x) O(t + ∆t) = Ki I(t − i) (5.130)
i=o
Note that C1 + C2 + C3 = 1. The value of x can be
determined using regression. = Ko I(t) + K1 I(t − 1) + K2 I(t − 2) + ...
where
5.5.4 Muskingum-Cunge Method
We can directly obtain the value of x used in the Musk- Ko = C1 (5.131)
ingum equation by 1) adding lateral inflows, qL , and 2)
K1 = C1 C3 + C2
substituting the kinematic velocity, c = dQ/dA, into the
original PDE, yielding: K2 = C3 K1
∂Q 1 ∂Q Kn = C3 Kn−1
+ + qL = 0 (5.123)
∂x c ∂t A problem with this method is that it ignores accumulat-
where qL = QL /∆x is the lateral inflow rate, QL , per unit ing errors that might have been corrected by incorporating
length of stream channel, ∆x = xb − xa . This is the same previous outflows.
as: Yet another method is the SCS Convex method that
is a simplification where C1 = 0, C3 = 1 − C2 , so that:
∂Q ∂Q
c + + c qL = 0 (5.124)
∂x ∂t O(t) = I1 + C2 [I(t − 1) − O(t − 1)] (5.132)
This yields a physically-based routing equation that can The assumption that C1 = 0 implies that ∆t = 2τ x. This
be solved to yield: is clearly the case when x = 0.5 and ∆t = τ . One might
select a time step which satisfies this condition.
O(t + ∆t) = C1 I(t + ∆t) + C2 I(t) (5.125)

+C3 O(t) + C4 5.6 Problems

where C1 , C2 , and C3 are identical with the Muskingum
Method, but now: 1. Write the Partial Differential Equations (PDEs) for:

qL τ ∆t (a) Transient, surface water flow.

C4 = (5.126)
∆t + 2τ (1 − x) (b) Transient, saturated, groundwater flow.
(c) Transient, unsaturated, groundwater flow.
and
 qL  (d) Transient, saturated, groundwater solute trans-
x = 0.5 1− (5.127) port.
c ∆h
where ∆h = hb −ha is the head drop between the upstream 2. Solve the Nash Model for a series of Linear Reser-
and downstream locations. voirs, Q = k H:
(a) Use Laplace Transforms to solve for a series of
5.5.5 Muskingham Variants five linear reservoirs.
A variant of the method, called the Muskingum Crest Seg- (b) Use a spreadsheet to solve the equivalent sys-
ment Routing method, uses only past inflows to the system tem using finite differences.
by replacing O(t) with previous predictions, i.e.: (c) Compare your numeric solution to the analytic
solution.
O(t) = C1 I(t) + C2 I(t − ∆t) (5.128)
3. Solve the Nash Model for a series of Non-Linear
+C3 O(t − ∆t) Reservoirs, Q = k H 1.5 :
and (a) Adjust your spreadsheet for the equivalent non-
linear system.
O(t − 1) = C1 I(t − 1) + C2 I(t − 2 ∆t) (5.129)
(b) Compare your nonlinear solution to the linear
+C3 O(t − 2 ∆t) solution.
4. Write the Finite Difference Equations (FDEs) for:
CHAPTER 5. DIFFERENTIAL EQUATIONS 38

(a) Transient, surface water flow.

(b) Transient, saturated, groundwater flow.
(c) Transient, unsaturated, groundwater flow.
(d) Transient, saturated, groundwater solute trans-
port.
5. Solve a one-dimensional flow equation using finite-
differences.
(a) In a spreadsheet, create a time index in Column
A. Label the top of the Column (Cell A1) with
Time. Start with a value of zero in Cell A2,
and increment each cell by 1 time step.
(b) Label the second column (cell B1) with x =
0. Label the columns C1:Z1 using a distance
increment of 10.
(c) Put constant values of flow, say Q = 10, in cells
B2:B50. Put different values of flow, say Q =
5, in cells B51:B100. These are the upstream
boundary condition, which are prescribed head
boundary conditions.
(d) Place a constant value in each of the cells in
row 2, i.e., Q = 10 in cells B2:Y2. These are
the initial conditions, which are also prescribed
head boundary conditions.
(e) Use a prescribed head downstream boundary con-
dition, say Q = 0, for Column Z, cells Z2:Z100.
(f) Place the time-centered flow equation in cells
C3:Y100.
Chapter 6

Vectors, Complex Variables, and

Quaternions

6.1 Vectors Multiplication or division of a constant, c, is also per-

formed element by element:
A vector is an oriented list of numbers. A vector with two
elements would look like: c · ~u = [cu1 , cu2 ] (6.8)

~v = [v1 , v2 ] (6.1) ~u h u1 u2 i
= , (6.9)
c c c
We can plot these elements on two perpendicular axes, Multiplication of two vectors is more complicated. In fact,
such as the x- and y-axes of a cartesian plot. We say there are two types of multiplication, the dot product, ~u ·~v ,
these axes are orthogonal, or perpendicular, to each other. and the cross product, ~u × ~v . Two dimensional multipli-
The vector is oriented because, like a pointer, it has a base, cation is defined using:
located at the origin, and the tip, located at the position
indicated by the list of numbers. ~u · ~v = u1 v1 + u2 v2 (6.10)
Another way of expressing a vector mathematically is:
~u × ~v = u1 v2 − u2 v1 (6.11)
~v = [v1 , v2 ] = i v1 + j v2 (6.2)
Three dimensional multiplication yields:
where the symbols i and j are just mathematical symbols
to indicate the x- and y-axes. A three-dimensional vector ~u · ~v = u1 v1 + u2 v2 + u3 v3 (6.12)
would be:
~u × ~v = [u2 v3 − u3 v2 , u3 v1 − u1 v3 , u1 v2 − u2 v1 ] (6.13)
~v = [v1 , v2 , v3 ] = i v1 + j v2 + k v3 (6.3) which is a vector. Vector multiplication is commutative
for the dot product, but not for the cross-product:
where there are now three axes, corresponding to the x-, y-
, and z-directions. Higher order vectors are also possible, ~u · ~v = ~v · ~u (6.14)
although not commonly used.
~u × ~v = −~v × ~u (6.15)
6.1.1 Algebra Vector division is not defined.
Vectors addition and subtraction are performed for each
element, independent of the other elements: 6.1.2 Geometry
~u + ~v = [u1 , u2 ] + [v1 , v2 ] A vector can be represented using a pointer, with the base
of the pointer resting at the origin and the tip of the
= [u1 + v1 , u2 + v2 ] (6.4) pointer resting at the value of the vector. Another way
of representing the vector is to define the vector magni-
~u − ~v = [u1 , u2 ] − [v1 , v2 ] tude, r, of the vector using:
= [u1 − v1 , u2 − v2 ] (6.5) q
r = abs{~u} = u21 + u22 (6.16)
These operations are associative and commutative:
and a rotation angle, θ, using:
(~u + ~v ) + w
~ = ~u + (~v + w)
~ (6.6)
u2
~u + ~v = ~v + ~u (6.7) θ = arg{~u} = arctan (6.17)
u1

39
CHAPTER 6. VECTORS, COMPLEX VARIABLES, AND QUATERNIONS 40

The addition of two vectors, w ~ = ~u + ~v , can be visualized

Table 6.1: Vector Identities
by placing the base of the second vector at the tip of the
first vector. The resulting vector is the pointer connecting
the base of the first pointer to the tip of the second pointer. ∇A =a
q
rw = abs{w}~ = w12 + w22 ∇(AB) = Ab + Ba
p ∇(a · b) = (b · ∇)a + (a · ∇)b
= (u1 + v1 )2 + (u2 + v2 )2 (6.18)
+ b × (∇ × a) + a × (∇ × b)
u2 + v2
θw = arg{w}~ = arctan (6.19) ∇2 A = ∇ · ∇A
u1 + v1
so that: =∇·a
∇ · (Ab) = b · a + A∇ · b = b · a + A∇2 B
u1 = r cos θ (6.20)
∇ × ∇A =∇×a
u2 = r sin θ (6.21) =0
The geometry of a three-dimensional vector is commonly ∇ × (Ab) = a × b + A(∇ × b)
written using two angles, θ and λ:
=a×b
u1 = r cos θ cos λ (6.22) ∇ · (a × b) = b · (∇ × a) − a · (∇ × b) = 0
u2 = r sin θ cos λ (6.23) ∇ × (a × b) = (b · ∇)a − (a · ∇)b + (∇ · b)a − (∇ · a)b
R R
u3 = r sin θ sin λ (6.24) V
a dV = V ∇A dV
R
= S A dS
R R
6.1.3 Calculus ∇ · a dV = S a · dS (Divergence theorem)
RV R
Calculus is used to describe how a function changes in ∇ × a dV = − S a × dS
RV H
space or time. Imagine an apple falling from a tree, as S
(∇ × a) · dS= C aḋC (Stokes′ s theorem)
Newton did. He is credited with creating a branch of
mathematics that describes the motion of a falling body.
We also use this mathematics to describe moving liquids. 6.1.4 Tensors
We first define the derivative of spatially variable func-
tion, U = f (x, y, z), which yields a vector: A tensor is used to incorporate the variation in material
  properties. For example, Darcy’s law is used to account
∂U ∂U ∂U
~u = ∇U = , , = [u1 , u2 , u3 ] for the flow of groundwater through porous media in the
∂x ∂y ∂z
subsurface. For spatially homogeneous media we have:
= i u1 + j u2 + k u3 (6.25)
~q = −K∇H = −K~h (6.33)
where ∇ is the gradient, or grad, operator. For two-
dimensional problems, this would be: where q is the flux of water, K is the hydraulic conduc-

∂U ∂U
 tivity, and ~h is the groundwater gradient. In this simple
~u = ∇U = , = [u1 , u2 ] = i u1 + j u2 (6.26) example, the pointer representing the hydraulic gradient
∂x ∂y
is oriented in the same direction as the groundwater flux.
Just as there are two types of vector multiplication, the In some cases, however, the flux direction is not in the
dot (·) and cross (×) products, there are also two types of same direction as the hydraulic gradient. This often oc-
vector derivatives: curs when there is preferential flow in one direction, such
∂u1 ∂u2 ∂u as when there is bedding or geologic layers, or when joints
∇ · ~u = + + (6.27)
∂x ∂y ∂z or fractures tend to direct flow away from the direction
∂u2 ∂u1 of the hydraulic gradient. In these cases, the hydraulic
∇ × ~u = − (6.28) conductivity is represented using a tensor:
∂x ∂y  
and for two-dimensional vectors: K11 K12
K= (6.34)
∂u1 ∂u2 K21 K22
∇ · ~u = + (6.29)
∂x ∂y so that
 
∂u3 ∂u2 ∂u1 ∂u3 ∂u2 ∂u1
∇ × ~u = − , − , − (6.30) ~q = K~h = [K11 h1 + K12 h2 , K21 h1 + K22 h2 ] (6.35)
∂y ∂z ∂z ∂x ∂x ∂y
Note that if ~u = ∇U , then for two-dimensional flow problems. For flow in three di-
mensions, the tensor would be:
∇ · ~u = ∇ · ∇U = ∇2 U (6.31)  
K11 K12 K13
∇ × ~u = ∇ × ∇U = 0 (6.32) K =  K21 K22 K23  (6.36)
because ∂(∂U/∂x)/∂y = ∂(∂U/∂y)/∂x, etc. K31 K32 K33
CHAPTER 6. VECTORS, COMPLEX VARIABLES, AND QUATERNIONS 41

Note that K12 = K21 , etc., so that the hydraulic conduc- Note that any complex variable, u, multiplied by its com-
tivity tensor is symmetric. plex conjugate, u, yields:

uu = (uo +i u1 )(uo −i u1 ) = (u2o +u21 )+i (uo u1 −u1 uo ) = abs{u}2

6.2 Complex Variables (6.48)
A complex variable is a number composed of two parts,
one called the real part and a second called the imaginary which is a real number.
part. Division of complex numbers is not strictly possible,
but we can get around this using the multiplication of the
u = uo + iu1 (6.37) complex conjugate rule. Given that we know w = uv and
u, and wish to find v, we first multiply both sides by the
where i is used to indicate the imaginary part. We denote complex conjugate of u, and then divide by their product:
the real and imaginary parts using the notation:
w = uv (6.49)
ℜ{u} = uo (6.38)
uw = uuv (6.50)
ℑ{u} = u1 (6.39) uw
v= (6.51)
The complex conjugate of an imaginary number is defined uu
using: This is possible because division by a real value is allowed.
u = uo − i u1 = ℜ{u} − ℑ{u} (6.40)
6.2.2 Geometry
6.2.1 Algebra The complex number, u, can be written using:
Adding two complex variables is accomplished by adding u = r(cos θ + i sin θ) (6.52)
the real and imaginary parts separately:
which is similar to the two-dimensional vector definition.
u + v = (uo + i u1 ) + (vo + i v1 ) = (uo + vo ) + i (u1 + v1 ) It can also be shown that:
(6.41)
u = rei θ (6.53)
Subtraction is equivalently defined using:
where:
u − v = (uo + i u1 ) − (vo + i v1 ) = (uo − vo ) + i (u1 − v1 )
ei θ = cos θ + i sin θ (6.54)
(6.42)
Note that ei π = −1. Using the fact that multiplication of
Note that u + u = 2ℜ{u} and u − u = 2ℑ{u}. exponential functions is equivalent to the addition of the
Multiplication and division with a real value, c, where exponents yields:
ℑ(c) = 0, is allowed:


w = uv = ru ei θu rv ei θv = ru rv ei (θu +θv ) (6.55)
cu = cuo + i cu1 (6.43)
u uo u1 which means that rw = ru rv and θw = θu + θv .
= +i (6.44) One can obtain a sense of what the imaginary number,
c c c
i, represents by noting that multiplication is a rotation
Multiplication of two complex variables is achieved us- about the origin. Multiplying a vector by −1 represents a
ing term-by-term expansion: rotation of 180◦ about the origin.
A rotation of 360◦ involves two rotations of 180◦ , so
uv = (uo +i u1 )(vo +i v1 ) = (uo vo +i2 u1 v1 )+i (uo v1 +u1 vo ) the rotation requires two successive multiplications of −1,
(6.45) or (−1) × (−1) = 1, which places the vector back in it’s
original position.
The convention is that whenever the imaginary part is With this example in mind, note that a multiplication
multiplied by itself, the result is a real value with a nega- by i results in a rotation of 90◦ . This is because two
tive sign, i.e.: successive multiplications (or rotations) yields i2 , which
is the same as 180◦ , which is just −1, so that:
i2 = −1 (6.46)
√
i = −1 (6.56)
so that:

uv = (uo vo − u1 v1 ) + i (uo v1 + u1 vo ) (6.47)

CHAPTER 6. VECTORS, COMPLEX VARIABLES, AND QUATERNIONS 42

6.2.3 Calculus (i, j, k). The real, ℜ, and imaginary, ℑ, components of

the quaternion are defined, respectively, using:
The derivative of a spatially variable (real) function, U =
f (z) is found using: ℜ(u) = uo and ℑ(u) = u1 i + u2 j + u3 k (6.67)
∂U ∂U ∂U ∂U ∂U
∇U = = + = −i = u (6.57) to yield:
∂z ∂x ∂i y ∂x ∂y
u = ℜ(u) + ℑ(u) (6.68)
where z = x + i y is the complex position, u = uo + i u1 ,
and uo = ∂U/∂x and uy = ∂U/∂y. Conventional complex variables are defined so that the
The derivative of a complex variable is handled some- quaternion real axis corresponds to the real x-axis, the
what differently. We require that the derivative of an ana- quaternion i-axis corresponds to the imaginary y-axis, and
lytic function be independent of the path of the derivative, the quaternion j and k axes are ignored. An important
so that: advantage of quaternion variables is their ability to extend
∂u ∂u complex variable theory to include two additional dimen-
∇u = = (6.58) sions.
∂x ∂i y
The pure quaternion is defined as a quaternion with
This can only be true if: only imaginary components, i.e., ℜ(u) = 0, and is remark-
ably similar to a three-dimensional vector, ~u.
∂u ∂uo ∂u1
= +i (6.59)
∂x ∂x ∂x ~u = ℑ(u) = [u1 , u2 , u3 ] = u1 i + u2 j + u3 k (6.69)
∂u ∂uo ∂u1 ∂uo ∂u1
= +i = −i + (6.60) 6.3.1 Algebra
∂i y ∂i y ∂i y ∂y ∂y
which requires that: If u and v are both quaternion, then addition is performed
by adding like quantities:
∂uo ∂u1
= (6.61)
∂x ∂y w =u+v
= (uo + u1 i + u2 j + u3 k) + (vo + v1 i + v2 j + v3 k)
∂u1 ∂uo
=− (6.62) = (uo + vo ) + (u1 + v1 )i + (u2 + v2 )j + (u3 + v3 )k
∂x ∂y
which are called the Cauchy-Riemann conditions. It is = wo + w1 i + w2 j + w3 k
clear that for the case of u = ∇U , we have: (6.70)

∂2U ∂2U where

+ =0 (6.63)
∂x2 ∂y 2 wo = uo + vo
∂2U ∂2U w1 = u1 + v1
− =0 (6.64) (6.71)
∂x∂y ∂y∂x w2 = u2 + v2
which are just ∇2 U = 0 and ∇ × ∇U = 0. These cor- w3 = u3 + v3
respond to the divergence and curl of the gradient of U
It is readily shown that addition is both commutative and
being equal to zero.
associative, respectively:

u+v =v+u (6.72)

6.3 Quaternions
and
The quaternion is defined as a number containing four
parts: u + (v + w) = (u + v) + w (6.73)
u = [uo , u1 , u2 , u3 ] (6.65) The conjugate quaternion, u is defined using:

which forms a four-dimensional vector in a coordinate sys- u = uo − u1 i − u2 j − u3 k = ℜ(u) − ℑ(u) (6.74)

tem with four orthogonal axes, equivalent to:
The real and imaginary parts of u can be found using:
u = uo + u1 i + u2 j + u3 k (6.66)
ℜ(u) = 12 (u + u) and ℑ(u) = 21 (u − u) (6.75)
In this notation, u is the vector of the distances from
the origin along one real axis and three imaginary axes The quaternion multiplication system are defined us-
ing the rules shown in Table 6.2. Note that the first row
CHAPTER 6. VECTORS, COMPLEX VARIABLES, AND QUATERNIONS 43

(6.84)
Table 6.2: Quaternion Multiplication Table
so that
Pre- Post-Multiplier  
Multiplier 1 i j k 0 − u1 v1 − u2 v2 − u3 v3 1
1 1 i j k 0 + 0 + u2 v3 − u3 v2 

~u ~v = −(~u·~v )+(~u×~v ) =  i
i i - 1 k -j 0 − u1 v3 + 0 + u3 v1  j
j j -k - 1 i 0 + u1 v2 − u2 v1 + 0 k
k k j -i - 1 (6.85)

where the zeros result from setting uo = 0 and vo = 0.

and column are symmetric, but the remaining rows and One advantage of using quaternion variables lies in their
columns are Hermitian. ability to explicitly incorporate both the dot and cross
These rules are equivalent to the right-hand rule of product multiplications.
vector multiplication for a pure quaternion. Employing Division in the quaternion system is needed if we wish
these rules, quaternion multiplication is: to solve for the quaternion u given two known quaternions,
w and v, defined by:
uv = (uo + u1 i + u2 j + u3 k) (vo + v1 i + v2 j + v3 k)
  w = uv (6.86)
uo vo − u1 v1 − u2 v2 − u3 v3 1
uo v1 + u1 vo + u2 v3 − u3 v2  i
=uo v2 − u1 v3 + u2 vo + u3 v1  j
 Post-multiplication by the conjugate of v, v, yields:

uo v3 + u1 v2 − u2 v1 + u3 vo k w v = (u v) v = u (v v) = u |v|2 (6.87)
(6.76)
where |v| is norm, or scalar length, of the quaternion. Be-
It can be readily shown that quaternion multiplication cause |v| is a scalar, we divide both sides by the squared
is both associative and distributive, respectively: norm of v to yield:
wv wv
u (v w) = (u v) w (6.77) u= = 2 (6.88)
vv |v|
u (v + w) = u v + u w (6.78)
6.3.2 Geometry
But quaternion multiplication is not commutative:
The spatial quaternion, u, can be represented using an
u v 6= v u (6.79) angle quaternion, φ:

Instead, it can be shown by expanding terms that: φ = φo + φ1 i + φ2 j + φ3 k (6.89)

uv = vu (6.80) with:

so that u = eφ = exp(φo + φ1 i + φ2 j + φ3 k)
(6.90)
= exp(φo ) exp(φ1 i) exp(φ2 j) exp(φ3 k)
uv = vu (6.81)
Note that φo is a scalar function, equal to the natural
As noted previously, the pure quaternion takes the logarithm of the norm of u:
form of the traditional vector, i.e., the real part of the
pure quaternion equals zero, ℜ(u) = 0. It is readily shown φo = ln |u| = ln r (6.91)
that pre- and post-multiplication of a pure quaternion, ~u,
by any quaternion, v, always yields a pure quaternion: where
q
w
~ = v ~u v (6.82) r = |u| = u2o + u21 + u22 + u23 (6.92)

The vector dot product corresponds to the negative so that:

of the real result of quaternion multiplication while the
u = r û (6.93)
vector cross product corresponds to the imaginary result,
such that for two pure quaternion, ~u and ~v : where û is a unit quaternion, with:
~u · ~v = −ℜ(~u ~v ) = u1 v1 + u2 v2 + u3 v3 (6.83) ~
û = exp(φ1 i + φ2 j + φ3 k) = exp(φ) (6.94)
and and
~
~u×~v = ℑ(~u ~v ) = (u2 v1 − u3 v2 ) i+(u3 v1 − u1 v3 ) j+(u1 v2 − u2 v1 ) kφ = ℑ(φ) = φ1 i + φ2 j + φ3 k (6.95)
CHAPTER 6. VECTORS, COMPLEX VARIABLES, AND QUATERNIONS 44

Conjugate Angles where

The quaternion conjugate angle, φ, is the angle of the v = r cos φ1 cos φ2 cos φ3 (6.105)
spatial quaternion conjugate, u:
This formulation provides four equations with four un-
~
u = exp φ = r exp φ (6.96) knowns, which has been solved using MAPLE to yield:
 u2 −u2 −u2 +u2 +β 
Addition of angle quaternion is associative and distribu- o 1 2 3
i
tive, but not commutative because spatial quaternion mul-  2(u2 u3 −uo u1 ) 
 2 2 2 2 
tiplication is not commutative. This restriction requires φ

~ = arctan  uo +u1 +u2 +u3 +β j   (6.106)
the angle quaternion conjugate be defined using:  2(uo u2 −u1 u3 ) 
 
u2o +u21 −u22 −u23 +β
~
φ = − (φ3 k + φ2 j + φ1 i) (6.97) 2(u1 u2 −uo u3 ) k

φ = φo − (φ3 k + φ2 j + φ1 i) (6.98) where

p
so that β=± r4 − 4(uo u2 + u1 u3 )2 (6.107)

~(i) = [φ1 , φ2 , φ3 ](i) , with i ∈ 1, . . . , 5, are

φ+φ = φo +(φ1 i + φ2 j + φ1 k)+φo −(φ3 k + φ2 j + φ1 i) = 2φoFive solutions, φ
(6.99) possible, due to the variable sign of β. Setting the first
solution equal to the positive root, then the following, re-
lated solutions are obtained for i ∈ 2, . . . , 5:
Half-Angles
 π (1) π (1) π (1)

Multiplication employing pure quaternion results in half- − 2 + φ1 , 2 − φ2 , 2 + φ3
 π (1) (1) (1) 
angle quaternion. Recall that the pure quaternion algebra ~(i) =  2 + φ1 , − π2 − φ2 , π
2 + φ3 
is: φ  (1)  (6.108)
 π2 + φ(1)
1 , π
2 − φ
(1)
2 , − π
2 + φ 3

(1) (1) (1)
~ = v ~u v
w (6.100) − π2 + φ1 , − π2 − φ2 , − π2 + φ3

so that the angles of v are half of a standard quaternion. Hyberbolic Identities

The standard hyberbolic trigonometric functions for a com-
Exponential and Trigonometric Forms plex angle, θ = θ1 + iθ2 , are:
We can decompose any unit quaternion into the following


exponential products: cosh θ = 12 eθ + e−θ and sinh θ = 21 eθ − e−θ
(6.109)
~ = exp(φ1 i) exp(φ2 j) exp(φ3 k)
û = exp(φ) (6.101)
~ are the same as:
which for a unit angle quaternion, φ,
which can be expanded, term-by-term, into a trigonomet-
ric form:    
~ = 1 eφ~ + e−φ~
cosh φ and ~ = 1 eφ~ − e−φ~
sinh φ
2 2

û = [cos(φ1 ) + sin(φ1 )i] [cos(φ2 ) + sin(φ2 )j] [cos(φ3 ) + sin(φ3 )k] (6.110)
(6.102)
We have already shown that:
which is the same as:  
  cos φ1 cos φ2 cos φ3 − sin φ1 sin φ2 sin φ3 1
cos φ1 cos φ2 cos φ3 − sin φ1 sin φ2 sin φ3 1 ~ sin φ1 cos φ2 cos φ3 + cos φ1 sin φ2 sin φ3  i
sin φ1 cos φ2 cos φ3 + cos φ1 sin φ2 sin φ3  i û = eφ =  
cos φ1 sin φ2 cos φ3 − sin φ1 cos φ2 sin φ3  j
û =  
cos φ1 sin φ2 cos φ3 − sin φ1 cos φ2 sin φ3  j (6.103)
cos φ1 cos φ2 sin φ3 + sin φ1 sin φ2 cos φ3 k
cos φ1 cos φ2 sin φ3 + sin φ1 sin φ2 cos φ3 k
(6.111)

Angle Determination and it is readily shown that:

The quaternion angles can be determined by performing  
the following substitutions: cos φ1 cos φ2 cos φ3 + sin φ1 sin φ2 sin φ3 1
~ − sin φ1 cos φ2 cos φ3 + cos φ1 sin φ2 sin φ3  i
  e −φ
= 
− cos φ1 sin φ2 cos φ3 − sin φ1 cos φ2 sin φ3 
1 − tan φ1 tan φ2 tan φ3 1 j
 tan φ1 + tan φ2 tan φ3  i − cos φ1 cos φ2 sin φ3 + sin φ1 sin φ2 cos φ3 k
u = v 
 tan φ2 − tan φ1 tan φ3  j (6.104)
(6.112)
tan φ3 + tan φ1 tan φ2 k
CHAPTER 6. VECTORS, COMPLEX VARIABLES, AND QUATERNIONS 45

Performing the addition and subtraction yields: respect to s as:

   
cos φ1 cos φ2 cos φ3 −u ′ φ du
sin
∆u duo du1 du2 du3
1 = 1 sin=φ2 sin
limφ3 1= + i+ j+ k (6.117)
  ds φ ∆s→0
 sin φ1 cos ∆s
 ds ds ds ds
~ =  cos φ1 sin φ2 sin φ3 
cosh φ
i
and ~=
sinh φ 2 cos φ3  i
− sin φ1 cos φ2 sin φ3  j  cos φThe φ2 cos φ3 ofj the product of a quaternion func-
1 sinderivative
sin φ1 sin φ2 cos φ3 k cos φ1 u(s),
tion, cos φ2and
sin φa3scalar
k constant, c, where ℑ(c) = 0, with
(6.113) respect to s is:
d(cuo ) d(cu1 ) d(cu2 ) d(cu3 )
which is just: (c u)′ = + i+ j+ k
ds ds ds ds
~ = cosh φ
û = exp φ ~ + sinh φ
~ (6.114) duo du1 du2 du3 (6.118)
=c +c i+c j+c k
ds ds ds ds
which is analogous to the polar form of complex numbers. = c u′
The unit quaternion can also be expanded to:
The derivative of the product of a quaternion function,
û = [cosh(φ1 i) + sinh(φ1 i)] [cosh(φ2 j) + sinh(φ2 j)] [cosh(φ3u(s), and a 3quaternion
k) + sinh(φ k)] constant, q = [qo , q1 , q2 , q3 ], can
(6.115) also be shown to be:

which can be evaluated using quaternion hyperbolic iden- (q u)′ = q u′ (6.119)

tities, shown in Table 6.3, to again yield: as long as q ′ = dq/ds = 0.
~ + sinh φ
~ The derivative of the product of two quaternion func-
û = cosh φ (6.116)
tions, u(s) and v(s), both a function of s, can be shown
to be:

Table 6.3: Quaternion Hyperbolic Identities (u v)′ = u′ v + u v ′ (6.120)

Partial Derivatives

cosh(φ) = (eφ + e−φ )/2 sinh(φ) = (eφ − e−φ )/2 Partial derivatives are defined similarly. Let r and s be
scalar variables and let u(r, s) be a quaternion that is a
cosh(φ1 i) = cos(φ1 ) sinh(φ1 i) = sin(φ1 )i continuously differentiable, single-valued function of r and
cosh(φ2 j) = cos(φ2 ) sinh(φ2 j) = sin(φ2 )j s. We define the partial derivatives of u with respect to r
cosh(φ3 k) = cos(φ3 ) sinh(φ3 k) = sin(φ3 )k and s as:
∂u ∆u ∂ uo ∂ u1 ∂ u2 ∂ u3
cos(φ1 i) = cosh(φ1 ) sin(φ1 i) = sinh(φ1 )i u′(r) = = lim = + i+ j+ k
∂r ∆r→0 ∆r ∂r ∂r ∂r ∂r
cos(φ2 j) = cosh(φ2 ) sin(φ2 j) = sinh(φ2 )j
(6.121)
cos(φ3 k) = cosh(φ3 ) sin(φ3 k) = sinh(φ3 )k
and
cosh(φ1 + βi) = cosh(φ1 ) cos(β) + sinh(φ1 ) sin(β)i
cosh(φ2 + βj) = cosh(φ2 ) cos(β) + sinh(φ2 ) sin(β)j ∂u ∆u ∂ uo ∂ u1 ∂ u2 ∂ u3
u′(s) = = lim = + i+ j+ k
cosh(φ3 + βk) = cosh(φ3 ) cos(β) + sinh(φ3 ) sin(β)k ∂s ∆s→0 ∆s ∂s ∂s ∂s ∂s
(6.122)
sinh(φ1 + βi) = sinh(φ1 ) cos(β) + cosh(φ1 ) sin(β)i
The derivative rules presented previously also hold for par-
sinh(φ2 + βj) = sinh(φ2 ) cos(β) + cosh(φ2 ) sin(β)j
tial derivatives.
sinh(φ3 + βk) = sinh(φ3 ) cos(β) + cosh(φ3 ) sin(β)k

exp(φ1 i) = cos(φ1 ) + sin(φ1 )i = cosh(φ1 i) + sinh(φ1 i) Scalar Derivative

exp(φ2 j) = cos(φ2 ) + sin(φ2 )j = cosh(φ2 j) + sinh(φ2 j) Let H(x) be a continuously differentiable, single-valued,
exp(φ3 k) = cos(φ3 ) + sin(φ3 )k = cosh(φ3 k) + sinh(φ3 k) scalar function of the quaternion position, x = xo + x1 i +
x2 j + x3 k. The gradient of this scalar function is defined
using:
dH ∆H
h = ⋄H = = lim
dx ∆x→0 ∆x
6.3.3 Calculus ∂H ∂H ∂H ∂H
= + + + (6.123)
Let s be a scalar variable and also let u(s) = [uo (s), u1 (s), u2 (s), u3 (s)] ∂ xo ∂ x1 i ∂ x2 j ∂ x3 k
be a quaternion that is a continuously differentiable, single- ∂H ∂H ∂H ∂H
valued function of s. Define the derivative of u(s) with = − i− j− k
∂ xo ∂ x1 ∂ x2 ∂ x3
CHAPTER 6. VECTORS, COMPLEX VARIABLES, AND QUATERNIONS 46

which is the same as: (6.133)

dH dx dH dx
h= = (6.124) where:
dx dx |dx|2  
∂H ∂H ∂H ∂H
This leads to the definition of the gradient operator: h= , , , (6.134)
  ∂ xo ∂ x1 ∂ x2 ∂ x3
d (·) ∂ (·) ∂ (·) ∂ (·) ∂ (·)
⋄(·) = = ,− ,− ,− (6.125) is the quaternion conjugate of the gradient of H, and
dx ∂ xo ∂ x1 ∂ x2 ∂ x3
 
The gradient of the product of a constant, c, and a scalar ∂ ∂∂ xHi
function, H(x), is: hij = (6.135)
∂ xj
⋄(cH) = c⋄H = c h (6.126)
Note that if the order of differentiation is unimportant,
The gradient of the product of two scalar functions, G(x) i.e., h = h , then ⋄2 H is a scalar:
ij ji
and H(x), where ℑ(G) = 0 and ℑ(H) = 0, is:
⋄(G H) = G ⋄H + H ⋄G = G h + H g (6.127) ⋄2 H = h00 + h11 + h22 + h33 (6.136)

which implies that ℑ(⋄2 H) = [0, 0, 0].

Quaternion Derivative
The derivative of a spatially variable quaternion, u(x), is Angle Derivative
obtained using:
It is possible to show using term-by-term expansion that
∂u ∂u ∂u ∂u the gradient of the Cartesian quaternion, u, is related to
⋄u = + + + (6.128)
∂ xo ∂ x1 i ∂ x2 j ∂ x3 k the gradient of the angle quaternion:
which is the same as: ⋄u = ⋄(eφ ) = (⋄φ) u (6.137)
 ∂u ∂ u1 i ∂ u2 j ∂ u3 k

∂ xo + ∂ xo + ∂ xo + ∂ xo
o

  and
 ∂ uo 
 ∂ x i + ∂∂ ux1 ii + ∂∂ ux2 ij + ∂∂ ux3 ik 
 1 1 1 1  (⋄u) u
⋄u =   (6.129) ⋄φ = = ⋄ ln(u) (6.138)
 ∂ uo + ∂ u1 i + ∂ u2 j + ∂ u3 k  uu
 ∂ x2 j ∂ x2 j ∂ x2 j ∂ x2 j 
 
∂ uo ∂ u1 i ∂ u2 j ∂ u3 k which is consistent with the definition of the natural log-
∂ x3 k + ∂ x3 k + ∂ x3 k + ∂ x3 k arithm:
or, equivalently:
 ∂ uo ln(u) = ln(eφ ) = φ (6.139)
∂ u1 ∂ u2 ∂ u3 
∂ xo + ∂ x1 + ∂ x2 + ∂ x3 1
 
 ∂ u1 − ∂ uo
+ ∂ u3
− ∂ u2 
i Hyperbolic Derivatives
 ∂ xo ∂ x1 ∂ x2 ∂ x3 
⋄u = 
 ∂ u2

 (6.130) Note that the derivatives of the hyperbolic functions are:
∂ u3 ∂ uo
∂ x − − + ∂∂ ux13  j
 o ∂ x1 ∂ x2 
∂ u3 ∂ u2 ∂ u1
~
∂ cosh φ ~
∂ cosh φ
∂ xo + ∂ x1 − ∂ x2 − ∂∂ uxo k =− ~
i = sinh φ (6.140)
3
∂(φ1 i) ∂φ1
The gradient of the product of a scalar, H(x), and a
quaternion, u(x), is: ~
∂ sinh φ ~
∂ sinh φ
=− ~
i = cosh φ (6.141)
∂(φ1 i) ∂φ1
⋄(H u) = h u + H ⋄u (6.131)
and
The gradient of the product of two quaternion, u(x) and
v(x), is: ~
∂ cosh φ ~
∂ cosh φ
= −k ~
= sinh φ (6.142)
⋄(u v) = (⋄u) v + (u ⋄) v = (⋄u + u ⋄) v (6.132) ∂(φ3 k) ∂φ3

Let H(x) be a continuous-differentiable, single-valued, ~

∂ sinh φ ~
∂ sinh φ
= −k ~
= cosh φ (6.143)
scalar function of the quaternion position, x. The quater- ∂(φ3 k) ∂φ3
nion derivative is defined using the gradient of the quater-
but the derivatives with respect to φ2 j do not provide the
nion conjugate of the gradient of h:
desired result. Also note that the derivatives w.r.t. φ1
 
+ h00 + h11 + h22 + h33require 1 post-multiplication of the result by i, while the

 − h + h + h − h 
derivatives
i w.r.t. φ2 require pre-multiplication by k.
⋄2 H = ⋄h = ⋄ ⋄H =  01 10
− h02 − h13 + h20 + h31  j
23 32 

− h01 + h12 − h21 + h30 k

CHAPTER 6. VECTORS, COMPLEX VARIABLES, AND QUATERNIONS 47

Integration 6.3.4 Relationship to Vectors

Let s again be a scalar variable and also let u(s) be a The vector gradient operator of a scalar, ∇H, is related
quaternion that is a continuously differentiable, single- to the quaternion gradient, ⋄H, using:
valued function of s. The integral of u with respect to
s is defined using: ∇H = [h1 , h2 , h3 ] = −ℑ(⋄H) = ℑ(h) (6.150)
Z Z
sb sb The vector divergence, ∇ · ~h, and vector curl, ∇ × ~h, are
u ds = [uo + u1 i + u2 j + u3 k] ds related to the quaternion gradient using:
sa s
Z axs Z sb Z sb Z sb
= uo ds + u1 i ds + u2 j ds + u3 k ds ℜ(⋄~h) = ∇ · ~h = h11 + h22 + h33 (6.151)
sa sa sa sa  
Z xs Z sb Z sb Z sb 
h23 − h32 i
= uo ds, u1 ds, u2 ds, u3 ds ℑ(⋄~h) = ∇ × ~h = h31 − h13  j (6.152)
sa sa sa sa
h12 − h21 k
(6.144)
where ~h = [h1 , h2 , h3 ] is a pure quaternion. The quater-
The integral of a constant, c, is:
nion gradient can therefore be related to the divergence
Z sb Z sb and curl derivatives using:
c ds = c ds = c (sb − sa ) (6.145) h i
sa sa
⋄~h = ∇ · ~h, ∇ × ~h (6.153)
The integral of the gradient of a scalar, h = ⋄H(x), is:
Z xb Z xb Z xb Z xb Note that the quaternion gradient explicitly incorporates
dH both )types of vector derivative functions (i.e., both the
h dx = ⋄H dx = dx = dH = H(xb )−H(x a
xa xa xa d x xa divergence and curl) with the additional ability to incor-
(6.146) porate the derivative of the real part of h, ho .
The real part of the quaternion integration corresponds
where the path of integration is immaterial. The integral to Gauss’ divergence theorem:
of the gradient of a quaternion, ⋄u(x), is: Z  Z Z
Z xb ℜ ⋄~u dR3 = (∇ · ~u) dR3 = ~ 2 (6.154)
~u · dR
⋄u dx = u(xb ) − u(xa ) (6.147) R 3 R 3 R2
xa
while the imaginary part corresponds to Stokes’ Theorem:
where the path is again immaterial. The integral of the Z  Z I
product of a constant, c, and the gradient of a quaternion, ~ ~ 1 (6.155)
ℑ ⋄~u dR2 = (∇ × ~u)· dR2 = ~u · dR
⋄u(x) is: R2 R2 R1
Z xb
c ⋄u dx = c [u(xb ) − u(xa )] (6.148) 6.4 The Dual
xa

The quaternion integral formulation is equivalent to All boundary value problems share the need to define a
combining Gauss’ divergence theorem with Stokes’ theo- governing equation (either differential or integral), domain
rem: properties (extent and material properties), boundary con-
Z Z Z ditions, and also initial conditions for transient problems.
⋄2 u dRn = ⋄u dRn−1 = u dRn−2 = U A wide range of solution methodologies exist for using this
Rn Rn−1 Rn−2 information to provide forward predictions of hydraulic
(6.149) head or flux, or inverse predictions of domain properties
and boundary conditions. Because boundary conditions
where: are a vital part of the problem definition, it is critical that
they are correctly formulated. While most models require
• R3 is the boundary three-dimensional volume sur-
that the normal component of flux boundary conditions
rounding the R4 hypervolume;
be specified, many users fail to recognize this requirement.
• R2 is the boundary two-dimensional surface surround- Instead, the flux magnitude is mistakenly used instead of
ing the R3 volume; the more-appropriate, normal component. The tangential
component of flux relative to the boundary should always
• R1 is the boundary one-dimensional line surrounding be discarded unless a complex (or, alternatively, a dual or
the R2 surface; and vector) formulation is used.
• Ro are the boundary points terminating the R1 line
segment.
CHAPTER 6. VECTORS, COMPLEX VARIABLES, AND QUATERNIONS 48

6.4.1 Problem Formulation

Figure 6.1: Diagram illustrating boundary flux compo-
In ground-water hydrology, Darcy’s law is used in com- nents.
bination with the conservation equation to specify the
governing equation, domain properties are defined using
where the superscripts are used to identify neighboring
homogeneous or heterogeneous distributions, and bound-
boundaries (j 6= k).
ary and initial conditions are assigned using knowledge
It should be emphasized that that normal fluxes (i.e.,
of the regional flow behavior. These equations are com-
qn ) are always used instead of flux magnitudes (i.e., |q|)
monly solved using finite-difference or finite-element meth-
for flux boundary conditions. Figure 6.1 presents what we
ods, and can be used for one-, two-, and three-dimensional
mean by the normal flux and the flux magnitude. Note
flow problems.
that the magnitude is the pythagorean
p sum of the normal
For steady flow, the conservation equation can be writ-
and tangential fluxes, i.e., |q| = qn2 + qτ2 .
ten in differential form using the divergence of the flux
Also note that the tangential component of the bound-
vector, ~q:
ary flux should be discarded. While ignoring important
∇ · ~q = 0 (6.156) attributes of the system seems problematic, it arises di-
rectly from the way in which the problem is formulated.
This can be combined with Darcy’s law:

~q = −K ∇h (6.157) 6.4.2 Dual Formulation

to obtain: The dual formulation provides a practical alternative to
the classical formulation that incorporates the tangential
∇ · K∇h = 0 (6.158) component of the boundary flux. For anisotropic media,
the problem can be posed in terms of the stream function,
where h is the hydraulic head, and K is the hydraulic s, the gradient of which is everywhere perpendicular to
conductivity tensor. For homogeneous and isotropic K, the hydraulic gradient, i.e.:
this can be simplified to just ∇2 h = 0.
Transient flow requires the inclusion of a storage term, ∇h · ∇s = 0 (6.164)
resulting in:
which is equivalent to:
∂h
∇ · K∇h = Ss (6.159)
∂t ∂h ∂s ∂h ∂s
+ =0 (6.165)
where Ss is the specific storage coefficient and t is time. ∂x ∂x ∂y ∂y
This can also be written for homogenous Ss using: and lead to the Cauchy-Riemann conditions:
∇ · D∇h = ḣ (6.160) ∂h ∂s ∂h ∂s
= and =− (6.166)
where D = K/Ss is the hydraulic diffusivity tensor and ∂x ∂y ∂y ∂x
ḣ = ∂h/∂t. The governing equation in this case is the curl, rather than
External boundary conditions take the general form: the divergence, of the flux vector:
φb h(Γ) + ψb qn (Γ) = 1 (6.161)
∇ × ~q = 0 (6.167)
where Γ is the boundary, qn is the normal component of
q , and the constants (φb and ψb ) are pre- so that:
the flux vector, ~
scribed using:
∇2 s = 0 (6.168)
P rescribed head (Dirichlet) φb 6= 0 ψb = 0 The dual boundary conditions are written in terms of the
P rescribed f lux (N eumann) φb = 0 ψb 6= 0 tangential components of boundary flux because:
M ixed (Cauchy) φb 6= 0 ψb 6= 0 ∂h ∂s ∂h ∂s
= and =− (6.169)
(6.162) ∂n ∂τ ∂τ ∂n

The final requirement is the specification of material which implies that the tangential change in head equals
properties, which can have distributed values across a finite- the negative of the normal change in stream potential.
difference grid or finite-element mesh. For a finite-element For isotropic media, the problem must be formulated
formulation, the head and fluxes across each internal (element-in terms of the flux vector, ~q, instead of the hydraulic
element) boundary are specified using: gradient, ∇h:

hj = hk and qnj = −qnk (6.163) ~q · ∇s = 0 (6.170)

CHAPTER 6. VECTORS, COMPLEX VARIABLES, AND QUATERNIONS 49

which is equivalent to: • Rasmussen TC, GQ Yu, 2006, “Determination of

groundwater flownets, fluxes, velocities, and travel
∂s ∂s times using the complex variable boundary element
qx + qy =0 (6.171)
∂x ∂y method”, Engr. Analysis with Boundary Elements
30(12):1030-1044.
and lead to the Cauchy-Riemann conditions:
• Rasmussen TC, Yu GQ, 2003, “A complex variable
∂s ∂s
qx = and qy = − (6.172) boundary-element strategy for determining ground-
∂y ∂x water flownets and travel times”, Advances in Water
Resources, 26(4):395-406.
q = ∇2 s = 0 and:
which again leads to ∇ × ~
• Yu GQ, Rasmussen TC, 1997, “Application of the
∂s ∂s
qn = and qτ = − (6.173) Ordinary Least Squares Approach for Solution of
∂τ ∂n Complex Variable Boundary Element Problems”, Int.
J. of Num. Methods in Engineering, 40(7):1281-
6.4.3 Cauchy Integral Formulation 1293.
Alternatively, the complex variable boundary integral for-
mulation can be used to solve steady, two-dimensional flow
problems by means of the Cauchy Integral equation:
6.5 Problems
I 1. Vectors
1 w(ξ)
w(z) = dξ (6.174)
2πi Γ z − ξ (a) Show the results of addition, subtraction, mul-
tiplication and division of two three-dimensional
where w is the complex potential: vectors, ~u = [u1 , u2 , u3 ] and ~v = [v1 , v2 , v3 ]
w = h + is (6.175) (b) Find the gradient, ∇U , of the function, U =
√ x2 + 2xy + y 2. Also find the divergence, ∇2 U =
where i is the imaginary number ( −1), and z is the ∇ · ∇U , and the curl, ∇ × ∇U .
complex potential: (c) Find the gradient, ∇U , of the function, U =
x2 − 2xy + y 2. Also find the divergence, ∇2 U =
z = x + iy (6.176) ∇ · ∇U , and the curl, ∇ × ∇U .
For heterogeneous media, each homogeneous region must 2. Complex Numbers
be enclosed by a boundary and linked to surrounding ho-
mogeneous regions using matching boundary conditions: (a) Determine whether any or all of the mathemat-
ical operations (i.e., addition, subtraction, mul-
wj = wk and q j = −q k (6.177) tiplication, and division) of complex variables is
associative, commutative, and distributive.
where the superscripts are again used to identify neighbor- (b) Calculate and plot the complex function u =
ing boundaries (j 6= k), and q is the complex flux, defined x2 , where x = xo + ix1 .
using: (c) Calculate and plot the complex function u =
dw log(x), where x = xo + ix1 .
q = qx + iqy = −K (6.178) (d) If i represents a rotation of 90◦ , find the com-
dz
plex number which represents a rotation of 45◦ .
where w is the complex conjugate of w. The flux can also
be found using the Hadamard integral formula (Chen and 3. Quaternions
Chen, 2000): (a) Determine whether any or all of the mathemat-
I ical operations (i.e., addition, subtraction, mul-
1 w(ξ)
q(z) = dξ (6.179) tiplication, and division) of quaternions is as-
2πi Γ (z − ξ)2
sociative, commutative, and distributive.
(b) Calculate and plot the function u = x2 , where
6.4.4 References
x = xo + ix1 + jx2 + kx3 .
• Chen JT, YW Chen, 2000, “Dual boundary element (c) Calculate and plot the function u = log(x).
analysis using complex variables for potential prob-
(d) If i, j and k represent a rotation of 90◦ , what
lems with or without a degenerate boundary”, Engi-
do each of these rotations mean?.
neering Analysis with Boundary Elements 24(9):671-
684, doi:10.1016/S0955-7997(00)00025-4. (e) What is the difference between quaternions and
latitudes-longitudes?
Chapter 7
Additional Reading
7.1 Mathematical References
• Abramowitz, M. and I.A. Stegun, 1972, ”Handbook
of Mathematical Functions”, Dover Press.
• Carslaw and Jaeger, 1986, ”Conduction of Heat in
Solids”, Second Edition, Oxford University Press.
• Law, A.M. and W.D. Kelton, 1991, ”Simulation Mod-
eling and Analysis”, Second Edition, McGraw-Hill,
759 pp.
• Pinder, G.F. and W.G. Gray, 1977, ”Finite Element
Simulation in Surface and Subsurface Hydrology”,
Academic Press, 295 pp.
7.2 Management Models
• Bachmat, Y., J. Bredehoft, B. Andrews, D. Holz,
and S. Sebastian, ”Groundwater Management: The
Use of Numerical Models”, 1980, American Geo-
physical Union, Washington, DC, 127 pp.
• National Academy of Sciences, 1990, ”Ground Wa-
ter Models: Scientific and Regulatory Applications”,
National Academy Press, 303 pp.
7.3 Systems/Statistical Models
• Box, G.P. and G.M. Jenkins, 1976, ”Time Series
Analysis: Forecasting and Control”, Holden-Day, 575
pp.
• Dooge, J.C.I., 1973, ”Linear Theory of Hydrologic
Systems”, USDA, ARS, Technical Bulletin 1468, 327
pp (Out of Print).
• McCuen, R.H. and W.M. Snyder, 1986, ”Hydrologic
Modeling: Statistical Methods and Applications”,
Prentice-Hall, 568 pp.
• Salas, J.D., J.W. Delleur, V. Yevjevich and W.L.
Lane, 1980, ”Applied Modeling of Hydrologic Time
Series”, Water Resources Publications, Littleton, CO,
484 pp.
50
• Singh, Vijay P., 1988, ”Hydrologic Systems: 1. Rainfall-
Runoff Modeling”, Prentice-Hall, ($67).
7.4 Watershed/Surface Water Mod-
els
• Haan, C.T., H.P. Johnson, and D.L. Brakensiek,
1982, ”Hydrologic Modeling of Small Watersheds”,
ASAE, Monograph 5, St. Joseph, MI, 533 pp.
• Kraijenhoff, D.A. and J.R. Moll, 1986, ”River Flow
Modelling and Forecasting”, D. Reidel Publishing
Company, 372 pp, (Out of Print).
7.5 Groundwater Models
• Bear, J. and A. Verruijt, 1987, ”Modeling Ground-
water Flow and Pollution”, Reidel, 414 pp.
• Haitjema, H.M., 1995, ”Analytic Element Modeling
of Groundwater Flow”, Academic Press, 394 pp.
• Istok, J.D., 1989, ”Groundwater Modeling by the Fi-
nite Element Method”, American Geophysical Union,
495 pp.
• Liggett, J.A. and P.L-F. Liu, 1982, ”The Boundary
Integral Equation Method for Porous Media Flow”,
Allen and Unwin, 255 pp.
• Remson, I., G.M. Hornberger, and F.J. Molz, 1971,
”Numerical Methods in Subsurface Hydrology, Wiley-
Interscience, 389 pp.
• Ségol, Geneviève, 1994, ”Classic Groundwater Simu-
lations: Proving and Improving Numerical Models”,
Prentice Hall, 531 pp.
• Wang, H.F. and M.P. Anderson, 1982, ”Introduction
to Groundwater Modeling: Finite Difference and Fi-
nite Element Methods”, Freeman, 237 pp.
Bibliography
[1] Abraham, B. and J. Ledolter, 1983, ”Statistical
Methods for Forecasting”, John Wiley and Sons,
New York, 445 pp.
[2] Abramowitz, M. and I.A. Stegun, 1972, ”Handbook
of Mathematical Functions”, Dover Press.
[3] Adamus, Paul R., Lauren T. Stockwell, Ellis J.
Clairain, Jr., Michael E. Morrow, Lawrence P.
Rozas, and R. Daniel Smith, Wetland Evaluation
Technique (WET), Volume 1: Literature Review
and Evaluation Rationale, Wetlands Research Pro-
gram Technical Report WRP-DE-2, U.S. Army
Corps of Engineers, Waterways Experiment Station,
Vicksburg MS, 1991.
[4] Azous, Amanda L. and Richard R. Horner (Eds.),
Wetlands and Urbanization: Implications for the
Future, Lewis Publishers, Boca Raton FL, 2001.
[5] Bachmat, Y., J. Bredehoft, B. Andrews, D. Holz,
and S. Sebastian, ”Groundwater Management: The
Use of Numerical Models”, 1980, American Geo-
physical Union, Washington, DC, 127 pp.
[6] Bear, J. and A. Verruijt, 1987, ”Modeling Ground-
water Flow and Pollution”, Reidel, 414 pp.
[7] Behl, E. and R.D. Jones, 1995, ”The Use and Limi-
tation of Modeling in the Pesticide Registration Pro-
cess at USEPA”, AGU Spring Meeting, Baltimore
MD, May.
[8] Black, Peter E., Watershed Hydrology, Second Edi-
tion, Lewis Publishers, Washington DC, 1996.
[9] Box, G.P. and G.M. Jenkins, 1976, ”Time Series
Analysis: Forecasting and Control”, Holden-Day,
575 pp.
[10] Bredehoeft, J.D. and L.F. Konikow, 1993, ”Ground-
Water Models: Validate or Invalidate”, Ground Wa-
ter, 31(2):178-179.
[11] Brown, V., D.I. Smith, R. Wiseman, and J. Hand-
mer, 1995, ”Risks and Opportunities: Managing En-
vironmental Conflict and Change”, Earthscan Pub-
lications Ltd., London, 354 pp.
[12] Buxton, B.E. (ed), 1989, ”Geostatistical, Sensitiv-
ity, and Uncertainty Methods for Ground-Water
51
Flow and Radionuclide Transport Modeling”, Bat-
telle Press, Columbus OH, 670 pp.
[13] Carslaw and Jaeger, 1986, ”Conduction of Heat in
Solids”, Second Edition, Oxford University Press.
[14] Carslaw, H.S. and J.C. Jaeger, Conduction of Heat
in Solids, Second Edition, Clarendon Press, Oxford
University Press, New York, 1959.
[15] Carter, Virginia, M.S. Bedinger, Richard P.
Novitzki, and W.O. Wilen, ”Water Resources and
Wetlands”, in Greeson, Phillip E., John R. Clark,
and Judith E. Clark (Eds.), Wetland Functions
and Values: The State of Our Understanding, Pro-
ceedings of the National Symposium on Wetlands,
held in Disneyworld Village, Lake Buena Vista FL,
November 7-10, 1978, American Water Resources
Association, Minneapolis MN, p. 344-376, 1979.
[16] Chabreck, Robert H., Coastal Marshes: Ecology
and Wildlife Management, University of Minnesota
Press, Minneapolis MN, 1988.
[17] Cohrssen, J.J. and V.T. Covello, 1989, ”Risk Anal-
ysis: A Guide for Principles and Methods for Ana-
lyzing Health and Environmental Risks”, Executive
Office of the President, Council of Environmental
Quality, Washington DC, 407 pp.
[18] Cole, C., 1994, ”Models in Decision Making and
Regulatory Policy: Model Evaluation, Model Use
and Validation”, Workshop on Coupling Geosphere
Transport Models with Regulatory Policies, Septem-
ber 22 and 23, 1994, Athens, GA.
[19] Cowardin, Lewis M., Virginia Carter, Francis C. Go-
let, and Edward T. LaRoe, Classification of Wet-
lands and Deepwater Habitats of the United States,
U.S. Fish and Wildlife Service, Washington DC,
FWS/OBS-79/31, 1979.
[20] Daniel, Charles C., Jr, ”Hydrology, Geology, and
Soils of Pocosins: A Comparison of Natural and
Altered Systems”, in Richardson, Curtis J. (Ed.),
Pososin Wetlands: An Integrated Analysis of
Coastal Plain Freshwater Bogs in North Carolina,
Hutchinson Ross Publishing Company, Stroudsburg
PA, p. 69-108, 1981.
BIBLIOGRAPHY
[21] Davis, P.A. and M.T. Goodrich, 1990, ”A Pro-
posed Strategy for the Validation of Ground-Water
Flow and Solute Transport Models”, GEOVAL-
1990, Symposium on Validation of Geosphere Flow
and Transport Models, Stockholm, May 14-17.
[22] Davis, P.A., N.E. Olague, and M.T. Goodrich,
1991, ”Approaches for the Validation of Models
Used for Performance Assessment of High-Level Nu-
clear Repositories”, NUREG/CR-5537, U.S. Nu-
clear Regulatory Commission, Washington DC, 21
pp.
[23] Davis, P.A., N.E. Olague, and M.T. Goodrich,
1992, ”Application of a Validation Strategy to
Darcy’s Experiment”, Advances in Water Resources,
15:175:180.
[24] Dennison, Mark S. and James F. Berry, Wetlands:
Guide to Science, Law, and Technology, Noyes Pub-
lications, Park Ridge, NJ, 1993.
[25] Dooge, J.C.I., 1973, ”Linear Theory of Hydrologic
Systems”, USDA, ARS, Technical Bulletin 1468, 327
pp (Out of Print).
[26] Dougherty, D.E. and A.C. Bagtzoglou, 1993,
”A Caution on the Regulatory Use of Numer-
ical Solute Transport Models”, Ground Water,
31(6):1001:1010.
[27] Duever, Michael J., ”Hydrology”, in Patten,
Bernard C. (Ed.), Wetlands and Shallow Continen-
tal Water Bodies, Volume 1: Natural and Human
Relationships, SPB Academic Publishing bv, The
Hague, The Netherlands, p. 61-89, 1990.
[28] Fisk, David W. (Ed.), Wetlands: Concerns and Suc-
cesses, Proceedings of conference held in Tampa,
Florida, September 17-22, 1989, American Water
Resources Association, Bethesda MD, 1989.
[29] Galatowitsch, Susan M. and Arnold G. van der
Valk, Restoring Prairie Wetlands: An Ecological
Approach, Iowa State University Press, Ames IO,
1994.
[30] GAO, U.S. General Accounting Office, 1995,
”Global Warming: Limitations of General Cir-
culation Models and Costs of Modeling Efforts”,
GAO/RCED-95-164, Washington, DC, 35 pp.
[31] Glass, R.J. and V.C. Tidwell, 1991, ”Research Pro-
gram to Develop and Validate Conceptual Models
for Flow and Transport Through Unsaturated Frac-
tured Rock”, SAND90-2261, Sandia National Labo-
ratories, Albuquerque, NM, 57 pp.
[32] Greeson, Phillip E., John R. Clark, and Judith E.
Clark (Eds.), Wetland Functions and Values: The
52
State of Our Understanding, Proceedings of the Na-
tional Symposium on Wetlands, held in Disneyworld
Village, Lake Buena Vista FL, November 7-10, 1978,
American Water Resources Association, Minneapo-
lis MN, 1979.
[33] Haan, C.T., H.P. Johnson, and D.L. Brakensiek,
1982, ”Hydrologic Modeling of Small Watersheds”,
ASAE, Monograph 5, St. Joseph, MI, 533 pp.
[34] Haitjema, H.M., 1995, ”Analytic Element Modeling
of Groundwater Flow”, Academic Press, 394 pp.
[35] Hammer, Donald A., Creating Freshwater Wetlands,
Second Edition, CRC Press, Boca Raton, FL, 1996.
[36] Hey, Donald L. and Nancy S. Philippi, A Case for
Wetland Restoration, John Wiley & Sons, New York
NY, 1999.
[37] Hilborn, R. and M. Mangel, 1997, ”The Ecological
Detective, Confronting Models with Data”, Prince-
ton University Press, Monographs in Population Bi-
ology, No. 28, 315 pp.
[38] Hillel, Daniel, Soil Water: Physical Principles and
Processes, Academic Press, New York NY, 1971.
[39] Himmelblau, D.M. and K.B. Bischoff, Process Anal-
ysis and Simulation - Deterministic Systems, John
Wiley & Sons, New York NY, 1968.
[40] Hook, Donal D. (Ed.), The Ecology and Manage-
ment of Wetlands, Volume 1: Ecology of Wetlands,
Timber Press, Portland OR, 1988.
[41] Hook, Donal D. (Ed.), The Ecology and Manage-
ment of Wetlands, Volume 2: Management, Use
and Value of Wetlands, Timber Press, Portland OR,
1988.
[42] Istok, J.D., 1989, ”Groundwater Modeling by the
Finite Element Method”, American Geophysical
Union, 495 pp.
[43] Kadlec, Robert H., R.B. Williams, and R.D. Scheffe,
”Wetland Evapotranspiration in Temperate and
Arid Climates”, in Hook, Donal D. (Ed.), The Ecol-
ogy and Management of Wetlands, Volume 1: Ecol-
ogy of Wetlands, Timber Press, Portland OR, p.
146-160, 1988.
[44] Kadlec, Robert H. and Robert L. Knight, Treatment
Wetlands, Lewis Publishers, Boca Raton FL, 1995.
[45] Kent, Donald M. (Ed.), Applied Wetlands Science
and Technology, Lewis Publishers, Boca Raton FL,
1994.
[46] Kraijenhoff, D.A. and J.R. Moll, 1986, ”River Flow
Modelling and Forecasting”, D. Reidel Publishing
Company, 372 pp, (Out of Print).
BIBLIOGRAPHY
[47] Law, A.M. and W.D. Kelton, 1991, ”Simulation
Modeling and Analysis”, Second Edition, McGraw-
Hill, 759 pp.
[48] Liggett, J.A. and P.L-F. Liu, 1982, ”The Boundary
Integral Equation Method for Porous Media Flow”,
Allen and Unwin, 255 pp.
[49] Madia, W.J., 1998, ”A Call for More Science in EPA
Regulations”, Science, 282(5386):45.
[50] Majumdar, Shyamal K., Robert P. Brooks, Fred J.
Brenner, and Ralph W. Tiner, Jr. (Eds.), Wetlands
Ecology and Conservation: Emphasis in Pennsylva-
nia, The Pennsylvania Academy of Science, Type-
house of Easton, Phillipsburg NJ, 1989.
[51] Maley, Michael and Jeff Peters, ”Hydrogeologic
Evaluation of the Springtown Alkali Sink, Califor-
nia”, in Means, Jeffrey L., and Robert E. Hinchee
(Eds.), Wetlands & Remediation: An International
Conference, Proceedings of conference held in Salt
Lake City, Utah, November 16-17, 1999, Battelle
Press, Columbus OH, p. 79-86, 1999.
[52] Marble, Anne D., A Guide to Wetland Functional
Design, Lewis Publishers, Boca Raton FL, 1992.
[53] Massman, J. and R.A. Freeze, 1987, ”Groundwater
Contamination From Waste Management Sites: The
Interaction Between Risk-Based Engineering Design
and Regulatory Policy”, Water Resources Research,
23(2):351-380.
[54] McCombie, C. and I. McKinley, 1993, ”Validation
- Another Perspective”, Ground Water, 31(4):530-
531.
[55] McCuen, R.H. and W.M. Snyder, 1986, ”Hydrologic
Modeling: Statistical Methods and Applications”,
Prentice-Hall, 568 pp.
[56] McCuen, Richard H., Hydrologic Analysis and De-
sign, Prentice Hall, Upper Saddle River, NJ, 1998.
[57] Meanley, Brooke, Swamps, River Bottoms and
Canebrakes, Barre Publishers, Barre MA, 1972.
[58] Means, Jeffrey L., and Robert E. Hinchee (Eds.),
Wetlands & Remediation: An International Confer-
ence, Proceedings of conference held in Salt Lake
City, Utah, November 16-17, 1999, Battelle Press,
Columbus OH, 1999.
[59] Mitsch, William J., and James G. Gosselink, Wet-
lands, Third Edition, John Wiley & Sons, New York
NY, 2000.
[60] Morgan, M.G., and M. Henrion, 1990, ”Uncertainty:
A Guide to Dealing with Uncertainty in Quantita-
tive Risk and Policy Analysis”, Cambridge Univer-
sity Press, 332 pp.
53
[61] Morton, A., 1993, ”Mathematical Models: Ques-
tions of Trustworthiness”, British Journal of Phi-
losophy and Science, 44:659-674.
[62] Murkin, Henry R., Arnold G. van der Valk, and
William R. Clark (Eds.), Prairie Wetland Ecology,
Iowa State University Press, Ames IO, 2000.
[63] Narasimhan, T.N., 1994, ”New Approaches to
Groundwater Modeling”, Workshop on Coupling
Geosphere Transport Models with Regulatory Poli-
cies, September 22 and 23, 1994, Athens, GA.
[64] National Academy of Sciences, 1990, ”Ground Wa-
ter Models: Scientific and Regulatory Applications”,
National Academy Press, 303 pp.
[65] Nelson, R.W., 1989, ”The Need to Update Ground
Water Pollution Control Strategies - A Technical Ba-
sis and Historical Perspective”, Recent Advances in
Ground-Water Hydrology, p. 458-479.
[66] Nelson, R.W., 1990, ”Needed Change to the Con-
ceptual Basis of Groundwater Quality Requirements
and Subsurface Evaluation Guidelines”, Hydrologi-
cal Science and Technology, 6:44-88.
[67] Nelson, R.W., 1991, ”A Summary Approach to
Better Understanding Groundwater Contamination
Problems and Evaluating Long-Term Environmental
Consequences”, Conference Proceedings on Health
and Ecological Implications of Radioactively Con-
taminated Environments, April 4-5, 1990, Washing-
ton DC, p. 51-64.
[68] Novitzki, Richard P., ”Wetland Hydrology”, in Ma-
jumdar, Shyamal K., Robert P. Brooks, Fred J.
Brenner, and Ralph W. Tiner, Jr. (Eds.), Wetlands
Ecology and Conservation: Emphasis in Pennsylva-
nia, The Pennsylvania Academy of Science, Type-
house of Easton, Phillipsburg NJ, p. 47- 64, 1989.
[69] NRC, National Research Council, 1990, ”Ground
Water Models: Scientific and Regulatory Applica-
tions”, National Academy Press, Washington, DC,
303 pp.
[70] Odum, William E., Eugene P. Odum, and Howard
T. Odum, ”Nature’s Pulsing Paradigm”, Estuaries,
18(4):547-555, 1995.
[71] Orme, Antony R., ”Wetland Morphology, Hydrody-
namics and Sedimentation”, in Williams, Michael
(Ed.), Wetlands: A Threatened Landscape, The
Institute of British Geographers, The Alden Press
Ltd., Osney Mead, Oxford, Great Britain, p. 42-94,
1990.
[72] Patten, Bernard C. (Ed.), Wetlands and Shallow
Continental Water Bodies, Volume 1: Natural and
Human Relationships, SPB Academic Publishing
bv, The Hague, The Netherlands, 1990.
BIBLIOGRAPHY
[73] Pinder, G.F. and W.G. Gray, 1977, ”Finite Element
Simulation in Surface and Subsurface Hydrology”,
Academic Press, 295 pp.
[74] Reinelt, Lorin E., Brian L. Taylor, and Richard R.
Horner, ”Morphology and Hydrology”, in Azous,
Amanda L. and Richard R. Horner (Eds.), Wet-
lands and Urbanization: Implications for the Future,
Lewis Publishers, Boca Raton FL, p. 31-45, 2001.
[75] Remson, I., G.M. Hornberger, and F.J. Molz,
1971, ”Numerical Methods in Subsurface Hydrology,
Wiley-Interscience, 389 pp.
[76] Richardson, Curtis J. (Ed.), Pososin Wetlands: An
Integrated Analysis of Coastal Plain Freshwater
Bogs in North Carolina, Hutchinson Ross Publishing
Company, Stroudsburg PA, 1981.
[77] Salas, J.D., J.W. Delleur, V. Yevjevich and W.L.
Lane, 1980, ”Applied Modeling of Hydrologic Time
Series”, Water Resources Publications, Littleton,
CO, 484 pp.
[78] Ségol, Geneviève, 1994, ”Classic Groundwater Simu-
lations: Proving and Improving Numerical Models”,
Prentice Hall, 531 pp.
[79] Singh, Vijay P., 1988, ”Hydrologic Systems: 1.
Rainfall-Runoff Modeling”, Prentice-Hall, ($67).
[80] Shrader-Frechette, K.S., 1989, ”Idealized Laws, An-
tirealism, and Applied Science: A Case in Hydroge-
ology”, Synthese, 81:329-352.
[81] Shrader-Frechette, K.S., 1990, ”Perceived Risks Ver-
sus Actual Risks: Managing Hazards Through Nego-
tiation”, Risk - Issues in Health and Safety, 1(4):341-
363.
[82] Shrader-Frechette, K.S., 1991, ”Ethical Dilemmas
and Radioactive Waste: A Survey of the Issues”,
Environmental Ethics, 13:327-343.
[83] Shrader-Frechette, K.S., 1992, ”Risk Estimation and
Expert Judgement: The Case of Yucca Mountain”,
Risk - Issues in Health and Safety, 283
[84] Shrader-Frechette, K.S., 1994, ”High-Level Waste,
Low-Level Logic”, The Bulletin of the Atomic Sci-
entists, November-December, p. 40-45.
[85] Shrader-Frechette, K.S., 1994, ”Biodiversity, Biolog-
ical Uncertainty, and Setting Conservation Priori-
ties”, Biology and Philosophy, 9:167:195.
[86] Stuurman, Roelof J. and Perry G.B. de Louw, ”The
Groundwater-Surface Water Interaction in Wetland
Restoration Management in the Netherlands”, in
Means, Jeffrey L., and Robert E. Hinchee (Eds.),
Wetlands & Remediation: An International Confer-
ence, Proceedings of conference held in Salt Lake
54
City, Utah, November 16-17, 1999, Battelle Press,
Columbus OH, p. 87-94, 1999.
[87] van der Valk, Arnold G. (Ed.), Northern Prairie
Wetlands, Iowa State University Press, Ames IO,
1989.
[88] Verry, Elon S. and Don H. Boelter, ”Peatland Hy-
drology”, in Greeson, Phillip E., John R. Clark, and
Judith E. Clark (Eds.), Wetland Functions and Val-
ues: The State of Our Understanding, Proceedings
of the National Symposium on Wetlands, held in
Disneyworld Village, Lake Buena Vista FL, Novem-
ber 7-10, 1978, American Water Resources Associa-
tion, Minneapolis MN, p. 389-402, 1979.
[89] Wang, H.F. and M.P. Anderson, 1982, ”Introduction
to Groundwater Modeling: Finite Difference and Fi-
nite Element Methods”, Freeman, 237 pp.
[90] Weller, Milton W., Freshwater Marshes: Ecology
and Wildlife Management, Third Edition, Univer-
sity of Minnesota Press, Minneapolis, MN, 1994.
[91] Wierenga, P.J. and D. Bachelet (eds), 1988, ”Valida-
tion of Flow and Transport Models for the Unsatu-
rated Zone”, Conference, May 23-26, 1988, Ruidoso,
NM, Research Report 88-SS-04, Dept. of Agronomy
and Horticulture, New Mexico State University, Las
Cruces, NM, 545 pp.
[92] Williams, C., K. Firehock, and J. Vincentz, Save
Our Streams:Handbook for Wetlands Conservation
and Sustainability. Izaak Walton League of America,
Gaithersburg, MD, 1996.
[93] Williams, Michael (Ed.), Wetlands: A Threatened
Landscape, The Institute of British Geographers,
The Alden Press Ltd., Osney Mead, Oxford, Great
Britain, 1990.
[94] Winter, Thomas C., ”The Concept of Hydrologic
Landscapes”, J. American Water Resources Associ-
ation, 37(2):335-349, 2001.
[95] Zedler, Joy B. (Ed.), Handbook for Restoring Tidal
Wetlands, CRC Press, Boca Raton, FL, 2001.
[96] Zimmerman, D.A., K.K. Wahl, A.L. Gutjahr, and
P.A. Davis, 1990, ”A Review of Techniques for Prop-
agating Data and Parameter Uncertainties in High-
Level Radioactive Waste Repository Performance
Assessments Models”, NUREG/CR-5393, U.S. Nu-
clear Regulatory Commission, Washington, DC.