11
Computers in Mathematical
Inquiry
JEREMY AVIGAD

11.1 Introduction
Computers are playing an increasingly central role in mathematical practice.
What are we to make of the new methods of inquiry?
In Section 11.2, I survey some of the ways in which computers are used
in mathematics. These raise questions that seem to have a generally epis-
temological character, although they do not fall squarely under a traditional
philosophical purview. The goal of this article is to try to articulate some

of these questions more clearly, and assess the philosophical methods that
may be brought to bear. In Section 11.3, I note that most of the issues can
be classified under two headings: some deal with the ability of computers
to deliver appropriate ‘evidence’ for mathematical assertions, a notion that is
explored in Section 11.4, while others deal with the ability of computers to
deliver appropriate mathematical ‘understanding’, a notion that is considered
in Section 11.5. Final thoughts are provided in Section 11.6.

11.2 Uses of computers in mathematics

Computers have had a dramatic influence on almost every arena of scientific
and technological development, and large tracts of mathematics have been
developed to support such applications. But this essay is not about the

I am grateful to Ben Jantzen and Teddy Seidenfeld for discussions of the notion of plausibility in
mathematics; to Ed Dean, Steve Kieffer, and Paolo Mancosu, for comments and corrections; and to
Alasdair Urquhart for pointing me to Kyburg’s comments on Pólya’s essay.



computers in mathematical inquiry 303

numerical, symbolic, and statistical methods that make it possible to use the
computer effectively in scientific domains. We will be concerned, rather, with
applications of computers to mathematics, that is, the sense in which computers
can help us acquire mathematical knowledge and understanding.
Two recent books, Mathematics by Experiment: Plausible Reasoning in the
21st Century (Borwein and Bailey, 2004) and Experimentation in Mathematics:
Computational Paths to Discovery (Borwein et al., 2004) provide a fine overview
of the ways that computers have been used in this regard (see also the
associated ‘Experimental Mathematics Website’, which provides additional
links and resources). Mounting awareness of the importance of such methods
led to the launch of a new journal, Experimental Mathematics, in 1992. The
introduction to the first book nicely characterizes the new mode of inquiry:
The new approach to mathematics—the utilization of advanced computing
technology in mathematical research—is often called experimental mathematics.
The computer provides the mathematician with a ‘laboratory’ in which he or
she can perform experiments: analyzing examples, testing out new ideas, or
searching for patterns ... To be precise, by experimental mathematics, we mean
the methodology of doing mathematics that includes the use of computations for:
1. Gaining insight and intuition.
2. Discovering new patterns and relationships.
3. Using graphical displays to suggest underlying mathematical principles.

4. Testing and especially falsifying conjectures.
5. Exploring a possible result to see if it is worth a formal proof.
6. Suggesting approaches for formal proof.
7. Replacing lengthy hand derivations with computer-based derivations.
8. Confirming analytically derived results.

In philosophical discourse it is common to distinguish between discovery and

justification; that is, to distinguish the process of formulating definitions and
conjectures from the process of justifying mathematical claims as true. Both
types of activities are involved in the list above.
On the discovery side, brute calculation can be used to suggest or test
general claims. Around the turn of the 19th century, Gauss conjectured
the prime number theorem after calculating the density of primes among
the first tens of thousands of natural numbers; such number-theoretic and
combinatorial calculations can now be performed quickly and easily. One
can evaluate a real-valued formula to a given precision, and then use an
‘Inverse Symbolic Calculator’ to check the result against extensive databases
to find a simplified expression. Similarly, one can use Neil Sloan’s ‘On-Line
Encyclopedia of Integer Sequences’ to identify a sequence of integers arising
from a particular calculation. These, and more refined methods along these



304 jeremy avigad

lines, are described in Bailey and Borwein (2005). Numerical methods can also
be used to simulate dynamical systems and determine their global properties,
or to calculate approximate solutions to systems of differential equations where
no closed-form solution is available. Graphical representations of data are often
useful in helping us understand such systems.
Computers are also used to justify mathematical claims. Computational
methods had been used to establish Fermat’s last theorem for the first four
million exponents by the time its general truth was settled in 1995. The
Riemann hypothesis has been established for all complex numbers with im-
aginary part less than 2.4 trillion, though the general claim remains unproved.
Appel and Haken’s 1977 proof of the four-color theorem is a well-known
example of a case in which brute force combinatorial enumeration played an
essential role in settling a longstanding open problem. Thomas Hales’ 1998
proof of the Kepler conjecture, which asserts that the optimal density of sphere
packing is achieved by the familiar hexagonal lattice packing, has a similar
character: the proof used computational methods to obtain an exhaustive
database of several thousand ‘tame’ graphs, and then to bound nonlinear and
linear optimization problems associated with these graphs. (This pattern of
reducing a problem to one that can be solved by combinatorial enumeration
and numerical methods is now common in discrete geometry.) Computer
algebra systems like Mathematica or Maple are used, in more mundane ways,

to simplify complex expressions that occur in ordinary mathematical proofs.
Computers are sometimes even used to find justifications that can be checked
by hand; for example, William McCune used a theorem prover named EQP
to show that a certain set of equations serve to axiomatize Boolean algebras
(McCune, 1997), settling a problem first posed by Tarski.
The increasing reliance on extensive computation has been one impetus
in the development of methods of formal verification. It has long been
understood that much of mathematics can be formalized in systems like Zer-
melo–Fraenkel set theory, at least in principle; in recent decades, computerized
‘proof assistants’ have been developed to make it possible to construct formal
mathematical proofs in practice. At present, the efforts required to verify
even elementary mathematical theorems are prohibitive. But the systems are
showing steady improvement, and some notable successes to date suggest that,
in the long run, the enterprise will become commonplace. Theorems that have
been verified, to date, include Gödel’s first incompleteness theorem, the prime
number theorem, the four color theorem, and the Jordan curve theorem (see
Wiedijk, 2006). Hales has launched a project to formally verify his proof of
the Kepler conjecture, and Georges Gonthier has launched a project to verify



computers in mathematical inquiry 305

the Feit–Thompson theorem. These are currently among the most ambitious
mathematical verification efforts under way.
Thus far, I have distinguished the use of computers to suggest plausible
mathematical claims from the use of computers to verify such claims. But in
many cases, this distinction is blurred. For example, the Santa Fe Institute
is devoted to the study of complex systems that arise in diverse contexts
ranging from physics and biology to economics and the social sciences.
Computational modeling and numeric simulation are central to the institute’s
methodology, and results of such ‘experiments’ are often held to be important
to understanding the relevant systems, even when they do not yield precise
mathematical hypotheses, let alone rigorous proofs.
Computers can also be used to provide inductive ‘evidence’ for precise
mathematical claims, like the claim that a number is prime. For example, a
probabilistic primality test due to Robert Solovay and Volker Strassen works
as follows.¹ For each natural number n, there is an easily calculable predicate,
Pn (a), such that if n is prime then Pn (a) is always true, and if n is not prime
then at least half the values of a less than n make Pn (a) false. Thus, one
can test the primality of n by choosing test values a0 , a1 , a2 , ... less than n at
random; if Pn (ai ) is true for a large number of tests, it is ‘virtually certain’ that
n is prime.²
In sum, the new experimental methodology relies on explicit or implicit

claims as to the utility of computational methods towards obtaining, verifying,
and confirming knowledge; suggesting theorems and making conjectures
plausible; and providing insight and understanding. These claims have a patent
epistemological tinge, and so merit philosophical scrutiny.³ For example, one
can ask:
• In what sense do calculations and simulations provide ‘evidence’ for
mathematical hypotheses? Is it rational to act on such evidence?
• How can computers be used to promote mathematical understanding?
• Does a proof obtained using extensive computation provide mathematical
certainty? Is it really a proof?

¹ A probabilistic test later developed by Michael Rabin, based on a deterministic version by Gary
Miller, has similar properties and is now more commonly used.
² This can be made mathematically precise. For example, suppose a 100-digit number is chosen at
random from a uniform distribution. Number-theoretic results show that there is a non-negligible prior
probability that n is prime. If one then chooses a0 , ... , al at random, one can show that the probability
that n is prime given that Pn (ai ) holds for every i approaches 1, quickly, as l increases.
³ The list is not exhaustive. For example, uses of computers in storing, organizing, and communi-
cating mathematical knowledge also raise issues that merit philosophical attention.



306 jeremy avigad

• Is knowledge gained from the use of a probabilistic primality test any less
certain or valuable than knowledge gained from a proof? What about
knowledge gained from simulation of a dynamical system?
• Does formal verification yield absolute, or near absolute, certainty? Is it
worth the effort?
As presented, these questions are too vague to support substantive discussion.
The first philosophical challenge, then, is to formulate them in such a way
that it is clear what types of analytic methods can have a bearing on the
answers.

11.3 The epistemology of mathematics

A fundamental goal of the epistemology of mathematics is to determine
the appropriate means of justifying a claim to mathematical knowledge.
The problem has a straightforward and generally accepted solution: the proper
warrant for the truth of a mathematical theorem is a mathematical proof, that is,
a deductive argument, using valid inferences, from axioms that are immediately
seen to be true. Much of the effort in the philosophy of mathematics has gone
towards determining the appropriate inferences and axioms, or explaining why

knowledge obtained in this way is worth having. These issues will not be
addressed here.
There are at least two ways in which one may wish to broaden one’s
epistemological scope, neither of which denies the correctness or importance
of the foregoing characterization. For one thing, one may want to have a
philosophical account of warrants for mathematical knowledge that takes into
consideration the fact that these warrants have to be recognized by physically
and computationally bounded agents. A formal proof is an abstract object,
albeit one that we may take to be reasonably well instantiated by symbolic
tokens on a physical page. But proofs in textbooks and mathematical journals
are somewhat further removed from this idealization: they are written in a
regimented but nonetheless imprecise and open-ended fragment of natural
language; the rules of inference are not spelled out explicitly; inferential steps
are generally much larger than the usual formal idealizations; background
knowledge is presupposed; and so on. Few can claim to have verified any
complex theorem from first principles; when reading a proof, we accept
appeals to theorems we have learned from textbooks, journal articles, and
colleagues. The logician’s claim is that the informal proof serves to indicate the
existence of the formal idealization, but the nature of this ‘indication’ is never



computers in mathematical inquiry 307

spelled out precisely. Moreover, we recognize that proofs can be mistaken,

and often express degrees of faith depending on the nature of the theorem,
the complexity of proof, the methods that have been used to prove it, and the
reliability of the author or the authorities that are cited. Just as mathematical
logic and traditional philosophy of mathematics provides us with an idealized
model of a perfect, gapless deduction, we may hope to model the notion of
an ‘ordinary’ proof and ask: when is it rational to accept an ordinary proof as
indicating the existence of an idealized one?⁴
To explore this issue, one need not conflate the attempt to provide an
idealized account of the proper warrants for mathematical knowledge with the
attempt to provide an account of the activities we may rationally pursue in
service of this ideal, given our physical and computational limitations. It is such
a conflation that has led Tymozcko (1979) to characterize mathematics as a
quasi-empirical science, and Fallis (1997, 2002) to wonder why mathematicians
refuse to admit inductive evidence in mathematical proofs. The easy answer
to Fallis’ bemusement is simply that inductive evidence is not the right sort of
thing to provide mathematical knowledge, as it is commonly understood. But
when their remarks are taken in an appropriate context, Tymoczko and Fallis
do raise the reasonable question of how (and whether) we can make sense of
mathematics, more broadly, as an activity carried out by agents with bounded
resources. This question should not be dismissed out of hand.

A second respect in which one may wish to broaden one’s epistemological
ambitions is to extend the analysis to value judgments that go beyond questions
of correctness. On the traditional view, the role of a proof is to warrant the
truth of the resulting theorem, in which case, all that matters is that the proof
is correct. But when it comes to proofs based on extensive computation, a far
more pressing concern is that they do not provide the desired mathematical
insight. Indeed, the fact that proofs provide more than warrants for truth
becomes clear when one considers that new proofs of a theorem are frequently
judged to be important, even when prior proofs have been accepted as correct.
We tend to feel that raw computation is incapable of delivering the type of
insight we are after:
... it is common for people first starting to grapple with computers to make
large-scale computations of things they might have done on a smaller scale by
hand. They might print out a table of the first 10,000 primes, only to find that
their printout isn’t something they really wanted after all. They discover by this
kind of experience that what they really want is usually not some collection of
‘answers’—what they want is understanding. ( Thurston, 1994, p. 162)

⁴ For an overview of issues related to the ‘surveyability’ of proofs, see Bassler (2006).



308 jeremy avigad

We often have good intuitions as to the ways that mathematical developments

constitute conceptual advances or further understanding. It is therefore reason-
able to ask for a philosophical theory that can serve to ground such assessments,
and account for the more general epistemological criteria by which such
developments are commonly judged.
In sum, questions about the use of computers in mathematics that seem
reasonable from a pre-theoretic perspective push us to extend the traditional
philosophy of mathematics in two ways: first, to develop theories of mathemat-
ical evidence, and second, to develop theories of mathematical understanding.
In the next two sections, I will consider each of these proposals, in turn.

11.4 Theories of mathematical evidence

We have seen that some issues regarding the use of computers in math-
ematics hinge on assessments of the ‘likelihood’ that a mathematical assertion
is true:
• a probabilistic primality test renders it highly likely that a number is prime;
• numeric simulations can render it plausible that a hypothesis is true;
• formal verification can render it nearly certain that a theorem has a correct

proof.
Since judgments like these serve to guide our actions, it is reasonable to
ask for a foundational framework in which they can be evaluated. Such
a framework may also have bearing on the development of computational
support for mathematics; for example, systems for automated reasoning and
formal verification often attempt to narrow the search space by choosing the
most ‘plausible’ or ‘promising’ paths.
Probabilistic notions of likelihood, evidence, and support have long played
a role in characterizing inductive reasoning in the empirical sciences, and it is
tempting to carry these notions over to the mathematical setting. However,
serious problems arise when one tries to do so. Roughly speaking, this is
because any mathematical assertion is either true, in which case it holds with
probability 1, or false, in which case it holds with probability 0, leaving no
room for values in between.
Put more precisely, classical approaches to probability model ‘events’ as
measurable subsets of a space whose elements are viewed as possible outcomes
of an experiment, or possible states of affairs. The laws of probability dictate
that if an event A entails an event B, in the sense that A ⊆ B, then the



computers in mathematical inquiry 309

probability of A is less than or equal to the probability of B. In particular, if a

property holds of all possible outcomes, the set of all possible states of affairs
that satisfy that property has probability 1. So, to assign a probability other than
1 to an assertion like ‘5 is prime’, one needs to characterize the primality of 5
as a property that may or may not hold of particular elements of a space. But
5 is prime, no matter what, and so it is difficult to imagine what type of space
could reasonably model the counterfactual case. I may declare X to be the set
{0, 1}, label 0 the state of affairs in which 5 is not prime, label 1 the state of
affairs in which 5 is prime, and then assign {0} and {1} each a probability 1/2.
But then I have simply modeled a coin flip; the hard part is to design a space
that can convincingly be argued to serve as an appropriate guide to behavior
in the face of uncertainty.
It is tempting to resort to a Bayesian interpretation, and view probabilities
as subjective degrees of belief. I can certainly claim to have a subjective degree
of belief of 1/2 that 5 is not prime; but such claims cannot play a role in a
theory of rationality until they are somehow linked to behavior. For example,
it is common to take the outward signs of a subjectively held probability to
be the willingness to bet on the outcome of an experiment (or the result of
determining the true state of affairs) with corresponding odds. In that case,
F. P. Ramsey and Bruno de Finetti have noted that the dictates of rationality
demand that, at the bare minimum, subjective assignments should conform to

the laws of probability, on pain of having a clever opponent ‘make book’ by
placing a system of bets that guarantees him or her a profit no matter what
transpires. But such coherence criteria still depend, implicitly, on having a
model of a space of possible outcomes, against which the possibility of book
can be judged. So one has simply shifted the problem to that of locating
a notion of coherence on which it is reasonable to have a less-than-perfect
certainty in the fact that 5 is prime; or, at least, to develop a notion of coherence
for which there is anything interesting to say about such beliefs.
The challenge of developing theories of rationality that do not assume logical
omniscience is not limited to modeling mathematical beliefs; it is just that the
difficulties involved in doing so are most salient in mathematical settings.
But the intuitions behind ascriptions of mathematical likelihood are often so
strong that some have been encouraged to overcome these difficulties. For
example, Pólya (1941) discusses a claim, by Euler, that it is nearly certain that
the coefficients of two analytic expressions agree, because the claim can easily
be verified in a number of specific cases. Pólya then suggested that it might
be possible to develop a ‘qualitative’ theory of mathematical plausibility to
account for such claims. (See also the other articles in Pólya 1984, and Kyburg’s
remarks at the end of that volume.) Ian Hacking (1967), I. J. Good (1977), and,



310 jeremy avigad

more recently, Haim Gaifman (2004) have proposed ways of making sense of
probability judgments in mathematical settings. David Corfield (2003) surveys
such attempts, and urges us to take them seriously.
Gaifman’s proposal is essentially a variant of the trivial ‘5 is prime’ example
I described above. Like Hacking, Gaifman takes sentences (rather than events
or propositions) to bear assignments of probability. He then describes ways
of imposing constraints on an agent’s deductive powers, and asks only that
ascriptions of probability be consistent with the entailments the agent can ‘see’
with his or her limited means. If all I am willing to bet on is the event that 5 is
prime and I am unable or unwilling to invest the effort to determine whether
this is the case, then, on Gaifman’s account, any assignment of probability
is ‘locally’ consistent with my beliefs. But Gaifman’s means of incorporating
closure under some deductive entailments allows for limited forms of reasoning
in such circumstances. For example, if I judge it unlikely that all the random
values drawn to conduct a probabilistic primality test are among a relatively
small number of misleading witnesses, and I use these values to perform a
calculation that certifies a particular number as prime, then I am justified in
concluding that it is likely that the number is prime. I may be wrong about the
chosen values and hence the conclusion, but at least, according to Gaifman,
there is a sense in which my beliefs are locally coherent.
But does this proposal really address the problems raised above? Without

a space of possibilities or a global notion of coherent behavior, it is hard to
say what the analysis does for us. Isaac Levi (1991, 2004) clarifies the issue
by distinguishing between theories of commitment and theories of performance.
Deductive logic provides theories of the beliefs a rational agent is ideally
committed to, perhaps on the basis of other beliefs that he or she is committed
to, independent of his or her ability to recognize those commitments. On that
view, it seems unreasonable to say that an agent committed to believing ‘A’ and
‘A implies B’ is not committed to believing ‘B’, or that an agent committed to
accepting the validity of basic arithmetic calculations is not committed to the
consequence of those calculations.
At issue, then, are questions of performance. Given that physically and
computationally bounded agents are not always capable of recognizing their
doxastic commitments, we may seek general procedures that we can follow to
approximate the ideal. For example, given bounds on the resources we are
able to devote to making a certain kind of decision, we may seek procedures
that provide correct judgments most of the time, and minimize errors. Can
one develop such a theory of ‘useful’ procedures? Of course! This is exactly
what theoretical computer science does. Taken at face value, the analysis of a
probabilistic primality test shows that if one draws a number at random from



computers in mathematical inquiry 311

a certain distribution, and a probabilistic primality test certifies the number as

prime, then with high probability the conclusion is correct. Gaifman’s theory
tries to go one step further and explain why it is rational to accept the result of
a test in a specific case where it provides a false answer. But it is not clear that this
adds anything to our understanding of rationality, or provides a justification
for using the test that is better than the fact that the procedure is efficient and
usually reliable.
When it comes to empirical events, we have no problem taking spaces of
possibilities to be implicit in informal judgments. Suppose I draw a marble
blindly from an urn containing 500 black marbles and 500 white marbles, clasp
it in my fist, and ask you to calculate the probability that the marble I hold is
black. The question presupposes that I intend for you to view the event as the
result of a draw of a ball from the urn. Without a salient background context,
the question as to the probability that a marble I clasp in my fist is black is close
to meaningless.
In a similar fashion, the best way to understand an ascription of likelihood
to a mathematical assertion may be to interpret it as a judgment as to the
likelihood that a certain manner of proceeding will, in general, yield a correct
result. Returning to Pólya’s example, Euler seems to be making a claim as
to the probability that two types of calculation, arising in a certain way, will
agree in each instance, given that they agree on sufficiently many randomly or

deterministically chosen test cases. If we assign a probability distribution to a
space of such calculations, there is no conceptual difficulty involved in making
sense of the claim. Refined analyses may try to model the types of calculations
one is ‘likely’ to come across in a given domain, and the outcome of such an
analysis may well support our intuitive judgments. The fact that the space in
question may be vague or intractable makes the problem little different from
those that arise in ordinary empirical settings.⁵

⁵ Another nice example is given by Wasserman (2004, Example 11.10), where statistical methods
are used to estimate the value of an integral that is too hard to compute. As the discussion after that
example suggests, the strategy of suppressing intractable information is more congenial to a classical
statistician than to a Bayesian one, who would insist, rather, that all the relevant information should be
reflected in one’s priors. This methodological difference was often emphasized by I. J. Good, though
Wasserman and Good draw opposite conclusions. Wasserman takes the classical statistician’s ability to
selectively ignore information to provide an advantage in certain contexts: ‘To construct procedures
with guaranteed long run performance, ... use frequentist methods.’ In contrast, Good takes the classical
statistician’s need to ignore information to indicate the fragility of those methods; see the references to
the ‘statistician’s stooge’ in Good (1983). I am grateful to Teddy Seidenfeld for bringing these references
to my attention.
I have already noted, above, that Good (1977) favors a Bayesian approach to assigning probabilities to
outcomes that are determined by calculation. But, once again, Levi’s distinction between commitment
and performance is helpful: what Good seems to propose is a theory that is capable of modeling



312 jeremy avigad

Along the same lines, the question as to the probability of the correctness
of a proof that has been obtained or verified with computational means is best
understood as a question as to the reliability of the computational methods
or the nature of the verification. Here, too, the modeling issues are not
unlike those that arise in empirical contexts. Vendors often claim ‘five-nines’
performance for fault-tolerant computing systems, meaning that the systems
can be expected to be up and running 99.999% of the time. Such judgments
are generally based on past performance, rather than on any complex statistical
modeling. That is not to say that there are not good reasons to expect that
past performance is a good predictor, or that understanding the system’s design
can’t bolster our confidence. In a similar manner, formal modeling may,
pragmatically, have little bearing on our confidence in computational methods
of verification.
In sum, there are two questions that arise with respect to theories of
mathematical evidence: first, whether any philosophical theory of mathematical
plausibility can be put to significant use in any of the domains in which the
notions arise; and second, if so, whether a fundamentally different concept
of rationality is needed. It is possible that proposals like Pólya’s, Hacking’s,
and Gaifman’s will prove useful in providing descriptive accounts of human
behavior in mathematical contexts, or in designing computational systems
that serve mathematical inquiry. But this is a case that needs to be made.

Doing so will require, first, a clearer demarcation of the informal data that the
philosophical theories are supposed to explain, and second, a better sense of
what it is that we want the explanations to do.

11.5 Theories of mathematical understanding

In addition to notions of mathematical evidence, we have seen that uses of
computers in mathematics also prompt evaluations that invoke notions of
mathematical understanding. For example:
• results of numeric simulation can help us understand the behavior of a
dynamical system;
• symbolic computation can help shed light on an algebraic structure;
• graphical representations can help us visualize complex objects and thereby
grasp their properties (see Mancosu, 2005).

‘reasonable’ behavior in computationally complex circumstances, without providing a normative

account of what such behavior is supposed to achieve.



computers in mathematical inquiry 313

Such notions can also underwrite negative judgments: we may feel that a proof
based on extensive computation does not provide the insight we are after, or
that formal verification does little to promote our understanding of a theorem.
The task is to make sense of these assessments.
But the word ‘understanding’ is used in many ways: we may speak of
understanding a theory, a problem, a solution, a conjecture, an example, a
theorem, or a proof. Theories of mathematical understanding may be taken
to encompass theories of explanation, analogy, visualization, heuristics, con-
cepts, and representations. Such notions are deployed across a wide range of
fields of inquiry, including mathematics, education, history of mathematics,
cognitive science, psychology, and computer science. In short, the subject
is a sprawling wilderness, and most, if not all, of the essays in this collec-
tion can be seen as attempts to tame it. (See also the collection Mancosu
et al., 2005.)
Similar topics have received considerably more attention in the philosophy
of science, but the distinct character of mathematics suggests that different
approaches are called for. Some have expressed skepticism that anything
philosophically interesting can be said about mathematical understanding, and
there is a tradition of addressing the notion only obliquely, with hushed tones
and poetic metaphor. This is unfortunate: I believe it is possible to develop
fairly down-to-earth accounts of key features of mathematical practice, and that

such work can serve as a model for progress where attempts in the philosophy of
science have stalled. In the next essay, I will argue that philosophical theories
of mathematical understanding should be cast in terms of analyses of the types
of mathematical abilities that are implicit in common scientific discourse where
notions of understanding are employed. Here, I will restrict myself to some
brief remarks as to the ways in which recent uses of computers in mathematics
can be used to develop such theories.
The influences between philosophy and computer science should run in
both directions. Specific conceptual problems that arise in computer science
provide effective targets for philosophical analysis, and goals like that of ver-
ifying common mathematical inferences or designing informative graphical
representations provide concrete standards of success, against which the util-
ity of an analytic framework can be evaluated. There is a large community
of researchers working to design systems that can carry out mathematical
reasoning effectively; and there is a smaller, but significant, community try-
ing to automate mathematical discovery and concept formation (see e.g.
Colton et al., 2000). If there is any domain of scientific inquiry for which
one might expect the philosophy of mathematics to play a supporting role,
this is it. The fact that the philosophy of mathematics provides virtually



314 jeremy avigad

no practical guidance in the appropriate use of common epistemic terms

may lead some to wonder what, exactly, philosophers are doing to earn
their keep.
In the other direction, computational methods that are developed towards
attaining specific goals can provide clues as to how one can develop a broader
philosophical theory. The data structures and procedures that are effective
in getting computers to exhibit the desired behavior can serve to direct our
attention to features of mathematics that are important to a philosophical
account.
In Avigad (2006), I addressed one small aspect of mathematical understand-
ing, namely, the process by which we understand the text of an ordinary
mathematical proof. I discussed ways in which efforts in formal verification
can inform and be informed by a philosophical study of this type of under-
standing. In the next essay, I will expand on this proposal, by clarifying the
conception of mathematical understanding that is implicit in the approach,
and discussing aspects of proofs in algebra, analysis, and geometry in light
of computational developments. In focusing on formal verification, I will be
dealing with only one of the many ways in which computers are used in
mathematics. So the effort, if successful, provides just one example of the ways
that a better interaction between philosophical and computational perspectives
can be beneficial to both.

11.6 Final thoughts
I have surveyed two ways in which the philosophy of mathematics may be
extended to address issues that arise with respect to the use of computers
in mathematical inquiry. I may, perhaps, be accused of expressing too much
skepticism with respect to attempts to develop theories of mathematical
evidence, and excessive optimism with respect to attempts to develop theories
of mathematical understanding. Be that as it may, I would like to close here
with some thoughts that are relevant to both enterprises.
First, it is a mistake to view recent uses of computers in mathematics as a
source of philosophical puzzles that can be studied in isolation, or resolved by
appeal to basic intuition. The types of questions raised here are only meaningful
in specific mathematical and scientific contexts, and a philosophical analysis is
only useful in so far as it can further such inquiry. Ask not what the use of
computers in mathematics can do for philosophy; ask what philosophy can do
for the use of computers in mathematics.



computers in mathematical inquiry 315

Second, issues regarding the use of computers in mathematics are best

understood in a broader epistemological context. Although some of the topics
explored here have become salient with recent computational developments,
none of the core issues are specific to the use of the computer per se.
Questions having to do with the pragmatic certainty of mathematical results,
the role of computation in mathematics, and the nature of mathematical
understanding have a much longer provenance, and are fundamental to
making sense of mathematical inquiry. What we need now is not a philosophy
of computers in mathematics; what we need is simply a better philosophy of
mathematics.

Bibliography
Avigad, Jeremy (2006), ‘Mathematical method and proof ’, Synthese, 153, 105–159.
Bailey, David and Borwein, Jonathan (2005), ‘Experimental mathematics: examples,
methods and implications’, Notices of the American Mathematical Society, 52, 502–514.
Bassler, O. Bradley (2006), ‘The surveyability of mathematical proof: A historical
perspective’, Synthese, 148, 99–133.
Borwein, Jonathan and Bailey, David (2004), Mathematics by Experiment: Plausible
Reasoning in the 21st Century (Natick, MA: A. K. Peters Ltd).

Borwein, Jonathan, Bailey, David, and Girgensohn, Roland (2004), Experimentation
in Mathematics: Computational Paths to Discovery (Natick, MA: A. K. Peters Ltd).
Colton, Simon, Bundy, Alan, and Walsh, Toby (2000), ‘On the notion of interesting-
ness in automated mathematical discovery’, International Journal of Human–Computer
Studies, 53, 351–365.
Corfield, David (2003), Towards a Philosophy of Real Mathematics (Cambridge: Cam-
bridge University Press).
Fallis, Don (1997), ‘The epistemic status of probabilistic proof ’, Journal of Philosophy,
94, 165–186.
(2002), ‘What do mathematicians want?: probabilistic proofs and the epistemic
goals of mathematicians’, Logique et Analyse, 45, 373–388.
Gaifman, Haim (2004), ‘Reasoning with limited resources and assigning probabilities
to arithmetical statements’, Synthese, 140, 97–119.
Good, I. J. (1977), ‘Dynamic probability, computer chess, and the measurement of
knowledge’, in E. W. Elcock and Donald Michie (eds.), Machine Intelligence 8 (New
York: John Wiley & Sons), pp. 139–150. Reprinted in Good (1983), pp. 106–116.
(1983), Good Thinking: The Foundations of Probability and its Applications (Min-
neapolis: University of Minnesota Press).
Hacking, Ian (1967), ‘A slightly more realistic personal probability’, Philosophy of
Science, 34, 311–325.



316 jeremy avigad

Levi, Isaac (1991), The Fixation of Belief and its Undoing (Cambridge: Cambridge
University Press).
(2004), ‘Gaifman’, Synthese, 140, 121–134.
Mancosu, Paolo (2005), ‘Visualization in logic and mathematics’, in Mancosu et al.
(2005).
Mancosu, Paolo, Jørgensen, Klaus Frovin, and Pedersen, Stig Andur (2005), Visual-
ization, Explanation and Reasoning Styles in Mathematics (Dordrecht: Springer-Verlag).
McCune, William (1997), ‘Solution of the Robbins problem’, Journal of Automated
Reasoning, 19, 263–276.
Pólya, George (1941), ‘Heuristic reasoning and the theory of probability’, American
Mathematical Monthly, 48, 450–465.
(1984), Collected papers. Vol. IV: Probability; Combinatorics; Teaching and Learning in
Mathematics, Gian-Carlo Rota, M. C. Reynolds, and R. M. Short eds. (Cambridge,
MA: MIT Press).
Thurston, William P. (1994), ‘On proof and progress in mathematics’, Bulletin of the
American Mathematical Society, 30, 161–177.
Tymozcko, Thomas (1979), ‘The four-color problem and its philosophical signifi-
cance’, Journal of Philosophy, 76, 57–83. Reprinted in Tymoczko (1998), pp. 243–266.
(ed.) (1998), New Directions in the Philosophy of Mathematics, expanded edn (Prince-
ton, NJ: Princeton University Press).
Wasserman, Larry (2004), All of Statistics (New York: Springer-Verlag).
Wiedijk, Freek (2006), The Seventeen Provers of the World (Berlin: Springer-Verlag).