0% found this document useful (0 votes)
133 views9 pages

Mathematical Models in Philosophy of Science

1. This document discusses mathematical models used in science and some of the fundamental questions philosophers of science ask about them, such as what models are, how they represent real-world systems, and how they are validated. 2. It outlines five basic questions philosophers of science ask about mathematical models, including what they are, how they represent target systems, how representation works, how models can change over time, and how they provide information about the systems they represent. 3. The document seeks to summarize different perspectives philosophers of science have taken in answering these questions about the nature and role of mathematical models in scientific knowledge.
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
133 views9 pages

Mathematical Models in Philosophy of Science

1. This document discusses mathematical models used in science and some of the fundamental questions philosophers of science ask about them, such as what models are, how they represent real-world systems, and how they are validated. 2. It outlines five basic questions philosophers of science ask about mathematical models, including what they are, how they represent target systems, how representation works, how models can change over time, and how they provide information about the systems they represent. 3. The document seeks to summarize different perspectives philosophers of science have taken in answering these questions about the nature and role of mathematical models in scientific knowledge.
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 9

Mathematical Models in Philosophy of Science

Zoltan Domotor, University of Pennsylvania, Philadelphia, PA, USA


 2015 Elsevier Ltd. All rights reserved.

Abstract

This article presents an overview of the most important foundational aspects of mathematical models used in science. The
importance of mathematical models is widely recognized for their essential role in the acquisition and organization of
scientific knowledge, and their success in representing the structure, mechanism, and behavior of real-world systems is
undeniable. However, what are these models, in virtue of what do they represent, what is the basis for choosing one particular
model over another, and how are models validated? These are some of the conundrums contemporary philosophy of science
tries to solve. Answers to these and other problems are given in the general frameworks of equationally vs. structurally conceived
mathematical models and their symmetry relationships. We illustrate several major types of models, their modifications and
structural enrichments on a concrete extended example, known for its significant intuitive content.

On the most fundamental level, the basic questions contem- represent such changes. To illustrate the pluralism in
porary philosophy of science asks and tries to answer about dynamical model types, recall the choices between
mathematical models used in science are as follows: deterministic and stochastic mathematical models, contin-
uous-time and discrete-time models, process-based and
1. What are mathematical models? There are many issues
agent-based models, classical and relativistic models, and so
buried in this question, known as the ‘ontological puzzle’:
forth. Applications of mathematics in systematic studies of
are models equations, formal systems, set-theoretic or
stereotype empirical systems tend to involve parametrized
category-theoretic structures, fictional entities, all of the
families, hierarchies, or networks of models.
above, or something else (e.g., ‘mediators’ between theory
4. How are models validated, what is the basis for choosing one
and the real world, autonomous tools of investigation,
particular model over another, and when is a model acceptable?
reconstructions of or formal stand-ins for so-called target
Validation is an empirical procedure intended to determine
systems)?
the degree to which a mathematical model under consid-
2. In virtue of what do mathematical models represent the chosen
eration is an accurate representation of its target system
aspects of a natural system, phenomenon, mechanism, or process
within its domain of intended application. As products of
under study – henceforth referred to as the target system – and
model developers’ cognitive work, mathematical models
how do model builders know whether or not a mathematical
are not independent of human understanding. Neverthe-
model correctly represents? These are very hard semantic and
less, their validity rests solely on their ability to predict
epistemic questions, since the success of mathematical
experimental outcomes to within antecedently specified
modeling in science is undeniable, but the nature of link-
errors (irrespective of agreement with common sense and
ages between models and systems, underlying the models’
intuition) and on the degree of coherence with other
success, is difficult to pin down. Einstein (1954, page 233)
accepted models.
puts his finger on this central problem in his famous
5. In virtue of what do mathematical models provide information
remark: “As far as the propositions of mathematics refer to
about their target systems? This question concerns the epis-
reality, they are not certain; and as far as they are certain, they
temology of models. We know that models usually repre-
do not refer to reality.” Applied mathematicians, scientists,
sent their target systems in ways that uphold the basic
and systems engineers have always held the view that
principles of knowledge acquisition about them. However,
most propositions of an adequate mathematical model
according to the philosophers of several different realist
refer to reality in some way, because without such
persuasions, a mathematical model can serve as a reliable
reference, the presence of mathematics in scientific
source of information about the target system only if it
representations would be unintelligible. A major problem
correctly represents the target system’s structure, function,
is that a mathematical model-based ‘representation’ is
or behavior under consideration, that is, if researchers
generally not a variant of copying or mirroring.
want to extract information from a mathematical model,
3. In what ways can mathematical models represent? Mathematical
they have to come to terms with how the model
methods form a huge part of scientific and engineering
represents.
practice. A natural system under study may be represented
by several different mathematical models, and conversely, Having identified the basic questions about models
a given mathematical model often represents many studied by philosophers of science, we next summarize
different (possibly unrelated) systems, both within and several corresponding answers. We gather them in one place
across various scientific disciplines. A large part of largely to standardize the terminology. It should come as no
contemporary natural and social science is concerned with surprise to the reader that there are several important ques-
investigating the temporal and spatial changes in target tions to which philosophers of science do not have satisfac-
systems. There are many ways in which models can and do tory answers.

International Encyclopedia of the Social & Behavioral Sciences, 2nd edition, Volume 14 http://dx.doi.org/10.1016/B978-0-08-097086-8.43057-6 791
792 Mathematical Models in Philosophy of Science

Debates Over Mathematical Models in Philosophy enclosed in a vessel. The real-valued macroscopic quan-
of Science tity P in the equation denotes gas pressure, quantity V
stands for the vessel’s volume, quantity T refers to the
Recently, philosophers of science have developed a large variety absolute temperature of gas in the vessel, and R is the gas
of approaches and solutions to the questions itemized earlier. constant, ensuring the equation’s dimensional homoge-
We begin by reviewing some of them here and set the scene neity. For another example of a model from the
for a more detailed study of mathematical models, their perspective of the standard conception consider the
construction, and application. now well-known Lotka–Volterra differential equations
dx dy
¼ ax  xy and ¼ xy  by of predator–prey interac-
dt dt
Mathematical Models tions. A simple regression model in statistics is defined
by a linear matrix equation of the form Y ¼ AX + Z, where
The question ‘What are mathematical models?’ has two
Y is a (normally distributed) random variable, A denotes
prominent answers, which considerably simplify the infla-
a coefficient matrix comprising constants, and Z stands
tionary use of the term ‘model’ in science:
for an independent error random variable with normal
1. In the first instance, there is the syntactic equational answer, distribution. In theoretical physics, systems theory, and
advocated by logical empiricists in the 1920s and retained mathematical economics, the most common examples of
by most philosophers of science as the so-called received mature mathematical models (in the received view sense)
view or standard conception until the mid-1960s. The idea is are special families of ordinary, partial differential, or
to define mathematical models in terms of systems of (most stochastic differential equations, involving selected real-
commonly equational) axioms of some sort, modulo valued quantities and parameters, and usually derived
logical equivalence, stated in a (semi) formal language. from first principles and idealized target system specifi-
According to the standard conception, equations together cations. In some cases, it is possible to solve these
with chosen constraints and initial conditions are used in equations analytically, and in analytically intractable
deriving numerous predictions that must then be situations suitable discretization algorithms are necessary
compared with independently obtained evidential data, to obtain approximate computer-generated solutions.
gained through measurement and experimentation. 2. There is an alternative semantic or structuralist answer to the
Repeated prediction–data comparisons form an integral question about the nature of mathematical models, first
part of model verification and confirmation. If the calcu- proposed by Suppes (1957, last Chapter) and Beth (1961),
lated predictions of a model under study and indepen- and further developed by those who followed them. The
dently obtained data agree to within a reasonable error, semantic view treats mathematical models as set-
then model users can be said to have a proper grounding theoretic structures of some sort – typically sets with
for their growing degrees of confidence that the model some kind of additional structure (e.g., relations and
correctly represents the target system. A model is confirmed operations thereon).
when its observable predictions are borne out, and only In the heydays of semantic model theory, Suppes (1988)
highly confirmed models should be accepted by model and others argued that a scientific theory should be specified
users. From an empiricist perspective, a mathematical by the class of those set-theoretic structures in which the
model is empirically adequate provided that all of its (equational) axioms of the theory are true. Accordingly,
predictions are borne out. Unfortunately, a model user whenever the classical view specifies a model by a body of
may never be in a position to know whether the model nonlogical axioms or other descriptive means, Suppes
under study is empirically adequate. An inadequate model introduces a corresponding set-theoretical predicate whose
is usually reformulated by appropriately changing the extension coincides with the class of those set-theoretic
assumptions under which it was constructed. models that satisfy the axioms. The choice of set theory as
The logical essence of real-world applications of models the best formal tool for representing theories is amply
in the empiricist received view sense is succinctly caught in the justified by its universal character. On this view, scientific
following inference scheme: models are nothing but appropriate classes of set-theoretic
structures of the same similarity type. In fact, one can
proceed a step further and consider hierarchies of models,
starting from low-level data models and closing with high-
level theoretical models. All this can be extended to
Simply, empirically or contingently true equations networks of models in which the traditional notion of
combined with initial conditions together with mathemat- a scientific theory (i.e., a deductively closed system of
ically sound calculations of predictions from these equa- empirical propositions) disappears. For Suppes, the
tions and initial conditions yield empirically true meaning of the concept of model is the same in both
predictions that can be compared with experimental data. mathematics and the empirical sciences.
The model is viewed as inadequate if the agreement Scientific models tend to involve intricate mixtures of
between prediction and independently obtained data lies linguistic and mathematical elements. Typically, equations or
outside of a reasonable estimate of error. other stylized syntactic items are used to describe the
The basic example of a model in the received view modeler’s set-theoretic structure, intended to represent the
sense is the state equation P$V ¼ R$T of an ideal gas target system in some idealized or approximate manner.
Mathematical Models in Philosophy of Science 793

According to the structural empiricists, a workable concep- A structural realist view of the model-target system relation-
tion of validation of models of target systems relies on (1) con- ship is outlined in the following block diagram:
structing data models, i.e., mathematical structures (matrices,
histograms, trajectories, signal curves, etc.) that encode raw
data from target system measurements in an idealized and Equational Description System
rectified manner and (2) structure-preserving embeddings of specification model
these data models into the system’s model. Accordingly,
a model correctly represents its target system provided that all Isomorphism
available data models are faithfully embeddable into it. We
summarize the structural empiricist idea of data-based model System s
validation by the following block diagram: structure

Equational Description System For fairly obvious reasons (stemming from various ideali-
specification model zations, abstractions, and approximations), mathematical
models constructed and used in science and engineering
Embedding seldom meet the unreasonably strong structuralist definition of
representation; the isomorphism requirement is usually
relaxed in several ways by passing to a partial isomorphism or
Data
homomorphism of some sort.
model
In contemporary philosophy of science, structuralism is
a leading approach to models and modeling, and it has several
The great classical example of a set-theoretical predicate is competing variants. The so-called ontic structuralists hold the
given by the class of Kolmogorov’s probability models. A set- view that mind-independent structure is all there is ‘out there.’
theoretic triple hU; B ; Pi is said to be a probability model In more detail, according to French and Ladyman (2003),
provided that (1) U is a nonempty set, called the sample space; natural systems, phenomena, and processes should be
(2) B is a Boolean sigma algebra of subsets of U, interpreted conceptualized in purely structural terms. Any nonstructural
as the algebra of events; and (3) P is a countably additive understanding of the nature of empirical objects is
probability measure on B . Here the customary domain of metaphysical and unwarranted by the physical science itself.
application of a theory is taken care of by the choice of Since according to ontic structuralists systems and
a concrete probability model hU; B ; Pi. Differential mani- phenomena do possess structure but they themselves are not
folds of relativity theory, directed graphs constructed for structure, models can be meaningfully thought of as
kinship relations, and normal-form game-theoretic structures representing their target systems via partial isomorphisms.
provide additional examples of structuralist-type models. Another approach, called epistemic structuralism, defends
the view that there is more to the world than its structure (e.g.,
objects, particles, events, and minds with natures and
Models and Representations
essences), but only structures are knowable. Structural empiricism
In response to the question ‘In virtue of what do mathematical argues that the structure of data and observables can be known,
models represent?’ an often-discussed metaphysical answer is but denies that science should aim at knowing more. In brief,
that mathematical models represent in virtue of nature itself a target system’s mathematical model is correct provided that
being essentially mathematical and logical. This goes well with the system’s data structures are faithfully embeddable into the
the contention that causal models represent causal relationships model. Thus, the model of a target system is empirically adequate
realized by the target system in a one-to-one fashion. The se- if it correctly captures the structure of all system-generated data,
mantic view, held (among others) by Van Fraassen (1994), offers made accessible via embedding maps.
the following explicit structuralist account of representation: Giere (1988, Chap. 3) and his followers argue that none of
the above-mentioned mappings between models and target
systems can capture the notion of representation. In their
A mathematical model M represents the target system S provided view, a model represents a target system just in case it is
that M is structurally isomorphic to S and M is actually intended by
the model builder to represent S.
similar to the system. This substantially weakened
requirement allows models to be only approximately the
same as their target system. Unfortunately, a similarity
A model M is said to be isomorphic to the object of relation-based account of representation does not fare any
modeling (target system) S if the following conditions hold: better than the structuralist isomorphism requirement, since
(1) each basic constituent of S stands in a one-to-one rela- in real-world modeling situations the notion of similarity
tionship with a uniquely associated element in M and of applies to practically anything, until the modeler provides an
course conversely, (2) to every quantitative property of the explicit list of all its relevant aspects and degrees – an
target system S there corresponds a unique real number- unworkable task in practice. More importantly, mathematical
valued function ear-marked by the model, and (3) each models tend to possess extra structure that their target systems
quantitative property relationship in S has a unique counter- do not have (e.g., the assumption of infinitely many agents
part relationship in the model M. in economic models or infinitely many gas atoms in
794 Mathematical Models in Philosophy of Science

thermodynamic models), and conversely, for reasons of trac- available only in degrees of probability. Unthoughtful model
tability, not all structures instantiated by a natural system are acceptance may to lead to two kinds of errors: (1) rejecting
included in its model. This issue comes up also in the context of a good model, prompted by validation based on incorrect or
isomorphism, discussed earlier. Most philosophers share the biased data, or (2) accepting a bad model, endorsed by
view that the problem of representation requires further study. incomplete validation. Fourth, validation is often limited by
the available class of observational data that are far from
sufficient to showing that the model is correct in its entire
Pluralism in Model Construction
range of application. Last, it is not true that good predictions
Aside from technical tractability and approximation issues, can only be obtained from a model that is causally sound.
there is no limit to the ways in which mathematical models Even near-perfect prediction–observation agreements over
may be constructed. Excluding the cases of unmanageable a wide range of observations may turn out to be a mere
complexity and uninformative simplicity, models may employ coincidence. Conversely, even good models can make bad
any mathematical structure whatsoever, including determin- predictions, as illustrated by Batitsky and Domotor (2007) in
istic, probabilistic, and stochastic variants and involving discussing models that are intended to represent the chaotic
discrete, computable, measurable, continuous, or smooth real- behaviors of simple target systems. In addition to robustness
or complex-valued quantities. and sensitivity analyses and establishing as many
Conventional top-down model building (e.g., in physics and connections between the model under study and its target
systems engineering) is characteristically based on first princi- system as possible, validation should also investigate the
ples of an established scientific theory and constraints provided compatibility relations between the model and the target
by the target system’s idealized specification. Often from the system’s other accepted models.
simple ‘initial’ models more complex ones can be developed by In applications, it is far from sufficient to know that
gradually removing certain idealizations and abstractions. a dynamical model is statistically valid. The reason is that
However, small increments in de-idealization or modeling a statistically valid theoretical model may turn out to be
realism are often paid for by significant increases of hard-to- dynamically (or qualitatively) invalid, meaning that it may fail
manage mathematical complexity. to represent correctly the dynamical invariants of the target
Since in biology, economics, and the social sciences in system (including equilibrium states, periodic, aperiodic,
general, there is limited access to universal laws and empirical chaotic, or strange attractors, and their basins) under various
regularities, the styles of model construction vary widely. In choices of parameter values. Note that even if a dynamical
particular, bottom-up or data-driven (e.g., time series generated) model provides highly accurate predictions in its tested regions
models tend to represent a relatively narrow spectrum of of intended use, it does not necessarily follow that the model is
phenomena, while broadly intended models often suffer from dynamically valid, since it may include spurious dynamics in
low degrees of realism. In these contexts, the point of using untested regions that are also of interest. The target system’s
a model is mostly to reveal its implications, explore its limits, spectrum of actual behaviors remains largely unknown from
and treat it as a guide in the development of epistemically more the perspective of its representing model. Further observation
powerful models. and theoretical research may be needed to obtain a better
knowledge about the dynamical invariants and characteristics
of the system. Additional problems arise when the model fails
Validation of Mathematical Models
to be valid in most parts of its domain of application. Defi-
Although some mathematical models are built for general ciencies in the model may be traced to at least three starting
scientific research purposes (e.g., fictional, simulation, and ‘toy’ conditions: (1) incorrectly derived equations (e.g., exclusion of
models), applications demand that models be validated by decisive higher degrees of nonlinearity), (2) a coarse-grained
comparisons of their predictions with measurement data and choice of time and/or space discretization parameters that
results of other (valid) models. Model validation is a process of misses their pathological sensitivity to infinitesimal changes,
determining the degree to which a model under study is an and (3) reliance on data samples that are contaminated with
accurate representation of its target system from the perspec- gross errors.
tives of its intended use and domain of application.
There are several problems that occupy the center stage in
Learning from Models
philosophical discussions of validation. First, since in view of
idealizations a mathematical model is rarely correct in all Models can serve as reliable vehicles of learning about target
subregions of its theoretically stipulated domain of applica- systems only if they correctly represent these systems. Thus,
tion, in general, mathematical models cannot be proved valid. model-based knowledge acquisition about systems is inti-
Second, in the case of complex models it would be too costly mately tied to the quality of representation. What is of partic-
and time-consuming to determine whether or not the model is ular interest here to philosophers of science is how a transfer of
valid over its entire intended domain of application. Verifi- knowledge from a correctly representing model to its target
cations are usually conducted until there is sufficient confi- system is achieved.
dence that the model works well within its envisioned domain We have now surveyed several answers to the philosophers’
of application. Third, determining that the model is adequate basic questions about models. After these preliminaries,
in the context of several empirical tests does not guarantee that we now turn to a more detailed study of different types of
it is valid everywhere in its applicable domain. Model users’ mathematical models, furnished with a wide variety of alge-
confidence that the model correctly represents the system is braic and geometric structures. The underlying objective is
Mathematical Models in Philosophy of Science 795

a characterization of the fundamental duality between equa- the motions of many other physical systems. However, the
tional and structuralist conceptions of mathematical models accompanying physical interpretations are different.
that most modern philosophers of science failed to recognize, The first term in the equation above encodes inertia, and the
and instead they continue to defend the idea that mathematical second term stands for the impact of gravity. Recall that the
models are basically set-theoretic or category-theoretic struc- pendulum’s arm is suspended from a pivot point around which
tures. Unfortunately, the idea that mathematical models in it oscillates or rotates in a vertical plane without surface resis-
science are structures that exist independently of and are tance and forcing, so that there are no extra additive terms in the
unrelated to equational characterizations does not sit well with equation for these causes. Recall that generally a model is
several contemporary mathematical results. a simplified mathematical representation of a real-world
system of interest for designated scientific purposes. Although
the target system has many important features, not all of them
Symmetry Relationship Between Equational and can and should be included in the model for reasons of trac-
Structuralist Conceptions of Mathematical Models tability and limited epistemic import. Those features that are
included often involve drastic idealizations known to be
In the first part of the previous section, we drew the reader’s empirically false.
attention to two opposing views about the nature of mathe- Solutions to the foregoing nonlinear pendulum equation
matical models. There is, however, a different approach to are continuously twice differentiable, time-dependent, real-
models that is even more important. In this alternative valued functions of the form q : T/S (intended to capture
approach, the main idea is unification via symmetry or duality, the pendulum bob’s time-varying angular positions in all
based on a less familiar one-to-one correspondence between regimes), where T denotes a designated continuum-time
the classical (algebraic differential) equation-based and domain, usually isomorphic to the additive group of reals,
(geometric) structure-based conceptions of models. This and the less familiar notation S ¼ df R=2pZ ¼ R mod 2p is
unified view of models is attractive, because it is respectful of reserved for the circle group of reals, whose underlying set is
the algebraic and geometric aspects of models in a comple- isomorphic to the half-open interval [0,2p]. In this way,
mentary fashion. Although this type of algebraic–geometric a given solution q assigns to each time instant t a unique
correspondence can be formalized in an abstract category- angle qðtÞ ¼ x in S, interpreted as the bob’s angular position
theoretic setting, we prefer a concrete illustration furnished by at t. Technically, all exact (nonelementary) solutions of the
various pendulum models. We choose pendulum models as pendulum equation are specified analytically by Jacobian
our running example because (1) most readers are intuitively elliptic sine functions, and the period is expressed in terms of
familiar with the physical behavior of pendulums, (2) they are the so-called complete elliptic integral of the first kind.
of historic and basic importance to philosophers of science, (3) Therefore, for any given initial condition the corresponding
it is possible to extend a pendulum model’s domain of appli- solution can only be obtained numerically, albeit with
cation by gradual de-idealization and structural enrichment, arbitrary accuracy, using rapid computer calculations. An
and (4) pendulum models are available for various physical alternative approach is a passage to discrete-time difference
modifications of classical pendulums, including coupled and equations of the form
double pendulums, extensible and forced pendulums, pendu- g
lums in a viscous fluid, relativistic and quantum pendulums, qnþ1  2qn þ qn1 þ ðDtÞ2 sin qn ¼ 0
l
and so forth.
where Dt is the length of the time interval, serving as the
Recall that a simple undamped planar pendulum is classi-
elementary step of the difference scheme, and for any
cally modeled by the autonomous, deterministic, second-order
natural number n, value qn ¼ df q(nDt) refers to the bob’s
nonlinear differential equation
approximate angular position at time nDt. To improve the
d2 q g accuracy of calculations, modelers should choose very small
þ sin q ¼ 0
dt 2 [ values of Dt.
To link the classical pendulum equation and its solutions to
in which the time-dependent unknown quantity q represents the
a corresponding structural description of the pendulum’s
pendulum’s continuously varying angular displacement,
dynamical behavior, we must reduce the original equation
measured from its downward vertical to the arm pivoting
to a pair of first-order autonomous differential equations of
without friction on a shaft, to which a bob of tiny size with mass
the form
m is attached at its swinging end. Constant g denotes the
homogeneous local gravitational acceleration acting on the dq
¼ y [1]
pendulum’s bob downward and parameter [ captures the length dt
of the pendulum’s idealized weightless and perfectly rigid
arm. These coefficients are needed for individuating the target dy g
¼  sin q [2]
pendulum (modulo duplicate relationships) in its classical dt [
gravitational ambience. The foregoing nonlinear pendulum Evidently, their solutions come in pairs of functions that
equation is derived from Newton’s second law of motion and are conveniently combined by set-theoretic pairing into
the pendulum’s accompanying idealizing scenario, involving a single solution of the form q; y : T/S  R that assigns to
significant idealizations of friction, torque, resistance, and each instant of time t a unique angular position-angular
elasticity. We mention in passing that in addition to the velocity pair [q;y](t) ¼ (q(t),y(t)), interpreted as the pendu-
pendulum, the nonlinear equation above also represents lum’s state of motion at t. We now have the mathematical
796 Mathematical Models in Philosophy of Science

means at hand to encode the pendulum’s physical behavior operator A. For additional details, see Engel, Nagel (2000), and
geometrically. In particular, the pendulum’s two-dimensional the references therein.
cylinder-shaped topological state space S  R (generated by Structuralists argue that the particular system of classical
the circle S) and certain smooth curves therein, corresponding equations that generates the dynamical transition map is of
to state-parametrized solutions and forming the target secondary concern. In a dynamical model, comprising a state
system’s so-called phase portrait, provide a complete geometric space and dynamical maps, each state space point is an
as opposed to an algebraic understanding of the pendulum’s abstract, information-bearing encoding of the target system’s
possible behaviors. physical mode of being at a given time. Assuming that the
We turn now from the solutions of the foregoing equations system under consideration has a state structure, its adequate
to a uniquely associated continuously differentiable dynamical state space model comes with a state space comprising
transition map of the form points that are presumed to contain complete information
about the past history of the system, relevant to its future
T : T  ðS  RÞ/S  R behavior. Thus, if the target system’s instantaneous physical
defined by Tt ðx; vÞ¼df hqx;v ðtÞ; yx;v ðtÞi for all time instants t, state – encoded by a point in its representing state space – is
angles x, and velocities v. We may regard Tt ðx; vÞ as a formal known, then the system’s subsequent temporal evolution
encoding of the pendulum’s ‘final’ state into which the can be predicted with the help of the model’s dynamical
‘initial’ state ðx; vÞ evolves after time t. Here it is important maps, without any additional knowledge of what has
to notice that, although there is a plentiful supply of solu- happened to the system. The dynamical map need not be
tions to the pendulum equations, they have to be individ- continuous or smooth. It can also be measurable, comput-
uated by states. For example, in view of the equality able, discrete, and local (i.e., only partially defined). The
hqx;v ð0Þ; yx;v ð0Þi ¼ ðx; vÞ, the state ðx; vÞ singles out the foregoing structuralist definition of a deterministic dynam-
solution qx;v ; yx;v (used in the definition above) and therefore ical model fits well into the conceptual framework of cate-
the trajectory or orbit fTt ðx; vÞjt ˛Tg of ðx; vÞ in S  R. gory theory that provides a formal adjointness account of
Trajectories densely fill the entire cylindrical state space and interplay between dynamical models and received-view
serve as prominent geometric tools in the study of equations.
pendulum motions. An obvious question now arises: what is duality in the case
It is easy to verify that the induced time-indexed dynamical of probability models? Thanks to the so-called Riesz and
maps spectral representation theorems, the question is not hard to
answer. Each probability model hU; B ; Pi comes with its von
Tt : S  R/S  R Neumann algebra LN ðU; B ; PÞ of bounded real-valued
random variables of the form X : U/R, modulo P-measure
satisfy the following functional equations for all time instants t
zero – meaning that random variables X and X 0 are viewed as
and t 0 in T:
equal whenever the equality PðfujXðuÞsX 0 ðuÞgÞ ¼ 0 holds.
1. Identity property: T0 ¼ I (identity transformation) and And every abstractly given commutative von Neumann algebra
2. Group property: Ttþt 0 ¼ Tt $Tt 0 . is isomorphic to an algebra of this type. Thus, as common in
statistics, one can start with a probability model and then
Dynamical maps completely determine the time evolution
switch to the induced algebra of random variables, their
of the pendulum’s motion in the following way: if we know
expectations, and equations between them or proceed in the
that the pendulum is in state ðx; vÞ at time t ¼ 0, then it is in
reverse direction.
state Tt ðx; vÞ at time t. Needless to add, in general, knowledge
We digress briefly to consider data models. Since exper-
of Tt is only approximate. As the reader may have already
imental data and numerical analysts’ calculations tend to
realized, we now have a canonical structuralist dynamical
come only with limited degrees accuracy (due to measure-
model hS  R; Ti of the classical pendulum – a semantic
ment errors and round-off operations), it is necessary to
counterpart or dual of the pendulum equations. In general,
consider selected finite configurations of states of a parent
dynamical models are given by an underlying (topological,
dynamical model that are conveniently viewed as the
measurable, etc.) state space and a time-indexed family of
underlying discrete state spaces of data models. In the spirit of
transformations (dynamical maps) thereon, satisfying the
state space discretization algorithms, a region of interest
functional equations.
(e.g., a basin of attraction of the model’s phase portrait) in
An impressively large class of deterministic dynamical
the model’s state space is subdivided into cells of appro-
models arises from autonomous systems of ordinary or partial
priate mesh size. Each cell comes with a designated repre-
differential equations by simple state-parametrizations of their
sentative state, usually calculated by suitable averaging.
solution spaces. Conversely, since Tt ðx; vÞ is a state-parame-
Dynamical maps on discrete state spaces are then deter-
trized solution of the pendulum equations, it automatically
mined by temporal transitions between cell representatives.
determines these equations. In general, the (vector) differential
Discretized dynamical models associated with state space
equation
regions provide the formal meeting ground for measurement
dTt results and calculated predictions. From a structural empir-
¼ ATt icist point of view, data models are finitary variants of dis-
dt
cretized dynamical models, intended to validate the parent
corresponding to a dynamical model hS  R; Ti is reconstructed model. Large families of compatible data models can be
from the analytic properties of Tt and those of a suitable used to ‘cover’ the parent dynamical model’s domain of
Mathematical Models in Philosophy of Science 797

intended application, as illustrated in the following simpli- 2. If E > 2 mg[, then the bob makes complete unidirectional
fied block diagram: rotations ad infinitum. These large energy values individ-
uate open orbits.
3. If E ¼ 2 mg[, then the pendulum’s bob has exactly the
Evolution Solutions Dynamical energy needed to reach the top vertical position, and it will
equations model remain in this unstable position until the slightest pertur-
bation sends it into one of the two trajectories intersecting at
Covering this point. The associated ‘border’ trajectory fðx; vjEðx; vÞÞ ¼
system of 2mglg is called a separatrix, since it separates the two
previous regimes (oscillations and rotations) and accord-
Data Data ingly divides the state space into disjoint regions of closed
model model and open orbits. From a physical standpoint, separatrix is
the trajectory in S  R that encodes the bob’s motion with
exactly enough energy for it to reach the top unstable
(hyperbolic) fixed point, taking an infinite amount of time
If the covering system of data models obtained from various to get there.
measurements locally agrees with the parent dynamical
model’s domain of application, that is to say, each data model These kinds of considerations lead to the conclusion that
in the covering system is embeddable into the parent model to equations in a differential paradigm are necessary for formal
within antecedently specified errors, then the parent model is derivations and calculations, while their semantic counterparts
said to be empirically adequate relative to the cover. Anyone who are crucial in geometric interpretations and embodiment of
is familiar with the chaotic behavior of certain systems quickly physical content.
realizes that covering chaotic regimes by deterministic data Before leaving this section, we make a few technical
models is ineffective and a passage to carefully designed remarks about observables and their dynamical models. It is
statistical data models (e.g., built from histograms) becomes easy to see that the bob’s velocity y : S  R/R is also an
necessary. observable quantity, if viewed as the second projection map,
At this stage, we widen the structuralist point of view defined by yðx; vÞ ¼ v. A very similar approach applies also
somewhat. Models in contemporary mathematical physics are to angular displacement q. Since these and other observables
based on the notion of observable, intended to encode a target can be added and multiplied, they automatically form an
system’s quantitative property of interest. In simple situations, algebra. Let CðS  RÞ be the algebra of all continuous
model builders assume that (1) observables are real-valued observables on the cylindric topological state space S  R.
functions, possessing values that can be approximately ascer- Conceptually, the topological structure of the state space is
tained by reading an appropriately chosen measuring instru- now traded for an algebraic structure in a unified manner.
ment and (2) their algebraic nature is best expressed by Since angular position, angular velocity, and other measure-
a suitable quantity algebra. Thus, in our running example ments are seldom perfectly accurate, we need a second
instead of describing the pendulum’s possible motions structural component that handles all uncertainties associ-
geometrically in terms of state space trajectories, we turn to ated with information acquisition about quantities. Let SðS
a more effective quantitative description of pendulum RÞ be the convex space of all normalized positive linear
dynamics, using an associated algebra of observables. The all- functionals of the form 3 : CðS  RÞ/R. Since for any
important observable quantity is the pendulum’s total energy quantity Q the value 3ðQÞ of 3 is the expectation of Q, it makes
(Hamiltonian) function E : S  R/R, defined by the sum of good sense to interpret the abstract functional 3 as the target
kinetic and potential energy functions system’s statistical state. All this is based on the so-called
m[2 2 spectral theorem that states that each statistical state has
Eðx; vÞ ¼ df v þ mg[ð1  cosxÞ a unique expectation integral representation relative to
2
a probability measure on some measurable space. In sum,
at each pendulum state ðx; vÞ. The total energy function is truly
quantity algebras and associated convex spaces of statistical
significant because (1) the pendulum equation can be recon-
states allow model builders to construct mathematical
structed from it by Hamiltonian differentiation and (2) each
models in an admittedly abstract albeit highly unified
curve of the phase portrait corresponds to a different energy
manner. More details on applications of quantity algebras to
value and thus to a different initial condition. Furthermore, it is
measurement theory may be found in Domotor and Batitsky
well-known that contingently on the possible values E of the
(2009).
energy function E with respect to its critical level 2mg[ (required
In particular, each structuralist topological state space dynam-
for the bob to reach the top point on the vertical), there are
ical model hS  R; Ti induces an algebraic dynamical model
three major types of pendulum motion that the experimenter
hCðS  RÞ; T i of observables, defined by the operator
can observe:
composition Tt ðQÞ ¼ df Q+Tt for all quantities Q and time
1. If E < 2mg[, then there is insufficient energy available and instants t. Concretely, in view of Et ðx0 ; v0 Þ ¼ Eðxt ; vt Þ, where
the pendulum’s bob is performing simple oscillations ðxt ; vt Þ ¼ Tt ðx0 ; v0 Þ, the energy quantity can now be considered
around the minimum. Larger energy values E specify as time dependent and evolving. In the same spirit, we can also
correspondingly larger closed (quasi) elliptical orbits in the consider a statistical dynamical model induced on statistical
pendulum’s state space. states SðS  RÞ.
798 Mathematical Models in Philosophy of Science

The passage from an underlying state space dynamical by constant or periodic external forces and their combina-
model to the algebraic dynamical model of quantities tions have mathematical significance in adding extra
becomes necessary when the modeler can no longer ignore the complexity to model construction. The question now arises as
ways in which measurements are performed on the systems to whether or not less idealized, and hence more complete
under study. When an observable is measured, its values are and more accurate models are decisively better to work with
obtained only with a certain probability. Unfortunately, the than simpler, less complicated models. As we have already
structuralists’ state space dynamical model represents indicated, model accuracy and overall adequacy is not
the target system’s behavior in itself, without providing the achievable by a simple increase in structural complexity. The
necessary conceptual ingredients for physical measurement reasons are buried in inaccuracies in background conditions,
and accompanying statistical properties. Discretized algebraic limited knowledge of too many parameter values, and esca-
dynamical models of quantities and associated algebraic data lating round-off errors in calculating predictions based on
models are based on algebraic dynamical models induced by complex models. Since complex models tend to lead to
discrete state space dynamical models. Topological state space resource-bounded and hence significantly less accurate results
dynamical models suffer from a fundamental inadequacy in than those obtained from simpler models, their complexity
the case of chaotic target systems, since they focus only on the must be commensurate with background information avail-
deterministic side of behavior and the statistical aspects are able in their applications.
ignored. However, in these complex scenarios only statistical In many applications, the same basic pendulum differen-
methods are effective. We shall not give any further details. The tial equation drops out of a wholly different idealizing
technical aspects of these scenarios are studied by Lasota and scenario of, say, a coupled mass-spring system, consisting of
Mackey (1994). a (point) mass attached to a spring at one end, where the other
end of the spring is tied to a fixed frame. Other prominent
examples of systems, whose dynamical behavior is also
Structural Enrichments, Perturbations, and Other modeled by ‘pendulum equations’, include electric circuits,
Modifications of Mathematical Models chemical reactions, and interacting biological populations
with oscillatory behaviors. Of course, in each application, the
In applications of mathematics to real-world problems, there indeterminate and individuating parameters are interpreted
is a frequent requirement of significant flexibility in mutating differently.
models by various combinations, de-idealizations, and struc- A whole new field of mathematical model construction is
tural enrichments. As an illustrative background, we shall opened up with the introduction of relativistic and quantum
continue to use the running example of pendulums. aspects of a target system. In the special case of a simple
We start with two examples involving a combination and pendulum, we already know that the solution space of the
mutation of nonlinear pendulum models. First, consider nonlinear pendulum equation includes many relativistically
a double pendulum. It consists of one pendulum attached to forbidden solutions that encode physically impossible behav-
another, so that it has two arms and two balls, where the second iors – individuated by the pendulum’s superluminal angular
arm is attached to the first arm’s ball. In the case of large velocities. The key idea here is that in the absence of complete
displacements, a double pendulum is famously chaotic, but for knowledge of the model’s empirical domain of applicability,
small oscillations it behaves in simple predictable ways. Double a researcher may be in an epistemic danger of trying to physi-
pendulum equations are derived from Newton’s second law of cally interpret (in terms of behavior) some of the solutions –
motion and simple background conditions, dictated by the countenanced by the model, which are actually meaningless
idealizing scenario. Structuralists represent the double in the physical world.
pendulum by the product dynamical model hS2  R2 ; T5T0 i To remedy the situation, the foregoing classical simple
with state space S2  R2 yðS  RÞ  ðS  RÞ and a product pendulum model must be extended to the relativistic case
dynamical map T5T0 obtained from the constituent dynamical (studied, e.g., by Erkal (2000)), having the form
maps of each pendulum and the coupling relation, responsible !2 !
for chaotic regimes. d2 q [ dq g
þ 1 sin q ¼ 0
Our other example addresses de-idealization cases of dt 2 c dt [
pendulum damping and forcing by various mechanisms.
Elementary treatments usually assume that the damping torque where parameter c denotes the speed of light. Because formal
is directly proportional to the angular velocity. This so-called models usually possess limited empirical domains of appli-
linear damping is captured quite simply by adding an extra term cability, in comprehensive treatments of target system
to the original nonlinear pendulum equation, so that now we behavior several different, closely related models may become
have necessary.
Real-world systems tend to have many representing models,
d2 q dq g
þ a þ sin q ¼ 0 and these models often possess a surplus content, which
dt 2 dt [ supports their mutations and extensions in unexpected ways.
where the damping coefficient a in the linear damping term is For example, in the presence of randomness or ‘noise’, a clas-
assumed to be positive. In order to achieve a better repre- sical stochastic pendulum equation provides the most appropriate
sentation of friction, quadratic and other terms may be added representation of randomly perturbed motion, observed (say)
to the initial pendulum equation, making the computation of on a pendulum attached to a vibrating ceiling. A typical model
approximate solutions increasingly harder. Pendulums driven of the behavior of simple pendulums affected by ‘noise’ is
Mathematical Models in Philosophy of Science 799

presented by the following pair of stochastic differential clear already that the duality between equations and solution-
equations: induced structures has significantly deepened our under-
standing of the nature of mathematical models, their
dQ
¼ Y [3] construction, and validation.
dt

dY g dW
¼  sin Q þ eY [4]
dt [ dt
See also: Axiomatic Theories; Logical Positivism and Logical
where Q denotes the ‘unknown’ stochastic angular position, i.e., Empiricism; Mathematical Psychology; Measurement Theory:
a time-dependent angle-valued random variable; Y symbolizes Conjoint Analysis Applications; Structuralism.
the ‘unknown’ stochastic angular velocity, i.e., a time-dependent
real-valued random variable; the noise parameter e > 0 is
intended to capture the strength of random noise; and W is
a stochastic (e.g., Wiener) process, encoding the pendulum’s
random erratic motion on a microscale. In this model, realiza- Bibliography
tions of trajectories, representing the pendulum’s motion in
a sample, are jagged in proportion to the value of noise Batitsky, V., Domotor, Z., 2007. When good theories make bad predictions. Synthese
parameter 3. Putting the earlier described constructions together 157, 79–103.
Beth, E., 1961. Semantics of physical theories. In: Freudenthal, H. (Ed.), The Concept
we envision the construction of stochastic dynamical models for and the Role of the Model in Mathematics and Natural and Social Sciences. Reidel,
double pendulums, relativistic pendulums, and so forth. Dordrecht.
We can also consider the case in which the stochastic Domotor, Z., Batitsky, V., 2009. Measurement, information channels, and dis-
component has a parametric perturbation form, captured by cretization: exploring the links. Measurement Science Review 9, 134–161.
Einstein, A., 1954. Ideas and Opinions. Crown Publishers, New York.
the stochastic differential equation
Engel, K.-J., Nagel, R., 2000. One-Parameter Semigroups for Linear Evolution
d 2 Q g  Equations: Graduate Texts in Mathematics 194. Springer, New York.
þ þ eW sin Q ¼ 0 Erkal, C., 2000. The simple pendulum: a relativistic revisit. European Journal of
dt 2 [ Physics 21, 377–384.
in which the ‘noise’ affects the period of oscillation. French, S., Ladyman, J., 2003. Remodeling structural realism: quantum physics and
the metaphysics of structure. Synthese 136, 31–65.
Importantly, although in recent years efforts have been
Giere, R.N., 1988. Explaining Science. Chicago University Press, Chicago, IL.
made to come to grips with the conceptual side of structuralist Lasota, A., Mackey, M.C., 1994. Chaos, Fractals and Noise: Stochastic Aspects of
dynamical models, practically all stochastic models are pre- Dynamics, second ed. Springer-Verlag, New York.
sented in an equational form. Suppes, P., 1957. Introduction to Logic. Van Nostrand, Princeton, NJ.
In conclusion, in the absence of other options it is too early Suppes, P., 1988. Scientific Structures and Their Representation. Preliminary Version.
Stanford University, CA.
to tell which of the two conceptions (i.e., equational vs. Van Fraassen, B.C., 1994. Interpretation in science; science as interpretation. In:
structural) of mathematical models will prove more important Hilgevoord, J. (Ed.), Physics and Our View of the World. Cambridge, New York,
in the future development of models in science. However, it is pp. 169–187.

You might also like