Phil. Trans. Royal Soc. Lond. A 361, 1675-1690

“Mathematics and machines have influenced each other for millennia. The advent of the digital computer introduced a powerfully new element that promises to trans form the relation between them . This paper outlines the thesis that the effect of the digital computer on mathematics , already widespread, is likely to be radical and far-reaching. To articulate this claim, an abstract model of doing mathemat ics is introduced based on a triad of actors of which one, the ‘agent’, corresponds to the function performed by the computer. The model is used to frame two sorts of transformation. The first is pragmatic and involves the alterations and progres sive colonization of the content and methods of enquiry of various mathematical fields brought about by digital methods. The second is conceptual and concerns a fundamental antagonism between the infinity enshrined in classical mathematics and physics (continuity, real numbers, asymptotic definitions) and the inherently real and material limit of processes associated with digital computation. An example which lies in the intersection of classical mathematics and computer science, the *P *= *NP *problem, is analysed in the light of this latter issue.

**1. ****Introduction**

Technology is permeated by mathematics. Every aspect of it depends at some level of its conception, design and technical description on mathematical ideas, knowledge, practices and formalisms. Nowhere is this more evident than for those particular localized implementations of technology we call machines, which in extreme cases are little more than fragments of mathematical structures that engineers have found ways to physically realize. The reverse relation, less obviously, is also the case: mathe matics is permeated by technology, by the traces and effects of machines and appara tuses that have surrounded and impinged on it throughout its history. Machines and mathematics are engaged in a two-way traffic that forms a co-evolutionary loop, so that, for example, a machine like the lever depends on the mathematics of ratios and conversely the theory of ratios and rational numbers is consolidated and motivated by the concrete representations levers provide. Of course, for an early machine such as the wheel-and-axle, the traffic would initially be mostly one way: the mathematical ideas of circles, axes, angularity and rotation being rooted in practical engagement with the machine long before such mathematics was applied to the technologies of circular and cyclic movement.

A materially framed history of mathematics would take this co-evolution as a datum. Such a genealogy of the mathematics -machine nexus starting with the wheel and lever and taking in the pump, abacrn;, du<..:k, priuting press, slide rule, loom, steam-engine, camera, electric motor, typewriter, gramophone, radio and com puter would uncover many particular versions of two-way traffic . Because mathematics is augmented and permanently altered by its encounters with machines, whereas machines are simply isolated implementations of this cumulative resource, the machine’s dependence on mathematics would be dominant and the reverse traffic confined to local and minor effects: a mathematical topic connected to the particular machine would get emphasized, certain abstract mathematical relations created from their machinic realizations, and so on.

With the advent of the digital computer, the entire co-evolutionary dynamic under went a phase shift, taking the machine/mathematics traffic to an entirely new, rad ically productive level. The digital computer was spawned by modern mathematics and metamathematics. With the exception of the abacus (its precursor), it is the most uncompromisingly mathematical of all machines ever created, being theoret ically based on a mathematical idealization of a ‘definite procedure’ -the Turing machine -and implemented as an electronic realization of the binary logic and for mal systems of mathematical logic.

The aim of this paper is to suggest that the effect of the computer on mathematics (pure and applied) will be correspondingly far-reaching and radical, that the com puter will ultimately reconfigure the mathematical matrix from which it sprang, and it. will do this not only by affecting changes in content and method over a wide math ematical terrain, but more importantly by altering the practice, the overall nature, and perhaps the very conception we have of mathematics.

To justify such a claim, we need to employ a framework able to articulate the deeper significance of the changes brought into play by the digital computer; we need, in other words, a general characterization or abstract model of what it means to *do *mathematics, to engage in the activity we call mathematics or, which comes to the same thing, what it means to be and function as a ‘mathematician’ .

# 2. A model of mathematical activity

Behind the various construals of mathematics as an activity-investigation of abstract patterns, symbolic reasoning about ideal entities, study of number and space, and so on-lie three distinct, fundamental theoretical discourses that enter into the subject, namely: idea, symbol and procedure. By ‘idea’ we mean the domain of human thought, mentation, and concepts from individual ideas to the most elabo rate narratives and products of the imagination; by ‘symbol’ the domain of language, signs, symbols and communicational and semiotic practices from notational devices to entire linguistic systems; and by ‘procedure’ the domain of action, process and operations from a single calculation, computation or verification to an arbitrarily complex program of instructions.

Mathematics can be understood as a vast and particular complex of activities across these domains. The figure of the ‘mathematician’, the one who carries out these activities, who knows and does mathematics, is on this understanding an assemblage of agencies, each of which implements the characteristic activity of a domain, so that, far from being a monolithic or indivisible actor, the mathematician can be seen as a plurality of parallel activities, a trio of different sources of action as follows.

The actor corresponding to the domain of the idea I call the *person. *This agency has extra-mathematical physical and cultural presence, is immersed in natural lan guage and the subjectivity it makes available, has insight and hunches, and provides motivation for and is the source of intuitions behind concepts and proofs. Abstracted from the person is the agency corresponding to the domain of the symbol, which I call the *subject. *The subject is the source of the inter-subjectivity embedded in mathematical languages, but is without the person’s capacity of self-reference, since it is circumscribed by languages which, by definition, lack indexical markers of time and physical place. Finally, there is the *agent, *the actor associated with the domain of procedure who functions as the delegate for the person through the mediation of the subject; the agent executes a mathematically idealized version of the actions imagined by the person, and it does so formally since it lacks the subject’s access to meaning and significance.

The model presented here, then, comprehends the mathematician as a tripartite being enacting certain kinds of inter-subjective thought experiments, imagined procedures conceived to take place in idealized times and spaces. The person imagines an idea *X *associated with some procedure Y-the intuitive idea of number and the action of counting, for example. The subject thinks *X *and *Y *within mathe matical signs, for example, via a notation system which describes and determines what ‘number’ signifies mathematically and presupposes what is meant by ‘count ing’. The agent, for whom *X *and *Y *are formal symbols and procedures, without meaning or significance, executes the action imagined by the person; for example, the agent carries out a calculation or other operation on a list of numerical expres sions.

One of the activities performed by the person is giving proofs of assertions. The person makes a claim about an imagined task or procedure -counting, inverting a matrix, etc.-that the agent will execute; the task is articulated via an appropriate linguistic apparatus supplied by the subject. Then, by tracking the agent via this apparatus, that is, by shadowing the subject, the person is persuaded as to the truth of the claim.

These three agencies are by no means absolute or fixed. This is obvious in the case of the person who is manifestly a cultural-historical product. But it is also true of the subject and the agent. They are constructs: what constitutes them-the rules and protocols of legitimate sign usage, permitted definitions, legitimate procedures, agreements about valid reasoning and proof, and so on-has been shaped over time by innovation, conceptual reconfiguration, philosophical criticism, explicit programmes of rigour, and the like. Having said this, however, one should note their relative stability: they do not change easily and when they do they are obliged do so in relation to each other in order to preserve the coherence of the mathematician they jointly facilitate.

The latter constraint implies that a major change in one of these agencies is likely to have significant consequences for the others, and hence for the character of the mathematician and mathematics as a whole. This indeed looks to be the case at the present historical juncture. The computer is providing an agent of incomparable

1678

power and flexibility, able to carry out procedures of which the results were previ ously unsuspected. This frees up the person to enlarge the scope and imaginative possibilities of mathematical thought and its significance. This in turn demands a subject together with a suitably rich and powerful mathematical language for com municating between the person and agent. The result, as we shall see, is that a reconfigured triad of agencies able to operate in new and complementary ways is being put in place.

With this triadic model as framework we can return to the digital computer·seffect on classical mathematics. In order to focus the question on essentials, we shall restrict attention to what is essential and inescapable about the computer: it is a rule-governed *machine; *it is a *digital *as opposed to an analog machine; it is a *material *machine, subject to and impinging on physical reality. The machinic aspect will result in local effects on mathematics involving reasoning and proof as well as the nature of iteration. The digital aspect will impinge on the notion of continuity which, in the form of the linear continuum, is the site of analog thinking within classical mathematics. The material aspect will impinge on the supposed purity of pure mathematics, that is, on its separation from what mathematicians call the ‘real world’. Since these notions form the cornerstone of post-Renaissance mathematics, it is likely that any change in their status will be significant.

The remarks about mathematics and computers that follow fall into two parts. The first (see § 3) is descriptive, outlining some of the varied ways the computer is currently altering the logical status, content, practice and evolution of mathematics, the main thrust of which is a displacement of the infinite and the continuous by the discrete and finite along with the openings such displacement bring into play. The second (see § 4) is more discursive, examining a sense in which infinity might be a philosophical or theoretical problem for computer science. Specifically, I look at a contention by cryptographers that computer science’s infinity-based idealization of a ‘feasible’ algorithm is irrelevant to practical, real-world, real-time computational tasks.

# 3. Local and global transformation

*(a) **Local effects: machine reasoning*

As indicated, every machine will have natural and expected local connections with the mathematics contiguous to it, that is those mathematical objects, practices, structures and topics it naturally requires and facilitates and in some cases has to have invented. For the lever it might have been rational numbers, for clocks and calendars counting and the arithmetic of repetition, for the pendulum the concept of a differential equation, and so on. For the computer, the mathematics natural to it ranges widely, since it embraces two core aspects of the subject: the computa tional/ calculational process itself and the notion of proof in the form of mathematical logic and metamathematics.

The local effects that flow from the computational process have to do with math ematical content, with kinds of mathematics that are inspired by, demand, are amenable to, or are directly fostered by computational and calculational procedures.

Many areas of mathematics are covered by such a description including much of graph theory, finite combinatorics, cryptography, and the entire theory and prac tice of cellular automata (of which more below); in addition there are fields such as

*Will the d**·**ig’ital comp’Uter transform class**·**ical mathematics? *1679

chaos theory, nonlinear dynamics and fractal geometry which embryonically existed before computational methods but only flourished with the entry of the computer into mathematics and which are now inseparable from the computational processes permeating them, And, more at the periphery of mainstream mathematics, there are those areas such as the investigation of symbolic systems and formal and arti ficial languages, the theory of algorithms, and finite model theory which , while not computer-driven, relate naturally to the investigation of computational questions . Indeed, a sub-field of one of these-that of algorithmic complexity -gave rise to the *P *= *NP *question, now considered one of the major open problems of contemporary mathematics. We shall examine an aspect of this problem in § 4.

The local effects associated with the logical, metamathematical aspect of comput ers are oriented not to content but to method: to questions of reasoning and syntax, modes of mathematical argument, deduction, inference, proof and validation, and the mechanics of formal systems. One form this has taken has been the development of computer-generated proofs, computer-aided reasoning, and theorem proving by machine. In a sense there is nothing new to such an enterprise. Ever since Pascal’s invention of a calculating machine and Leibniz’s celebrated call to replace all dispu tation and rational argument by formal, symbolic means, there have been numer ous and varied efforts to mechanize reasoning. The contemporary versions of these range from having the computer find new proofs of existing theorems to the more impressive goal of generating proofs or disproofs of open questions as part of efforts to rival traditional, entirely symbolic methods of mathematical proof and deduc tion by computational validation. Some of these have yielded positive results; one such theorem-proving procedure successfully settled a question, the so-called Rob bins conjecture about certain algebras, which had been open for 60 years (McClure 1997).

Less ambitious is the use of machine reasoning to augment rather than replace the traditional symbolic modes of mathematical argument . The main need for augmen tation is the difficulty of managing large volumes of data and information involved in the handling of cases, circumstances, and effects so numerous and heterogeneous as to put them beyond the capacity of an individual mathematician’s cognitive where withal, not to say lifetime, to examine. For example, the theorems underlying the classification of all simple finite groups relies on a computer-compiled database too massive and unwieldy and internally complex for a single mathematician to process (Soloman 1995). Though ostensibly more mundane than using computers to prove theorems outright, this augmentative approach raises a fundamental issue as to the nature and assumed constitution of ‘the mathematician ‘. A famous example is the four-colour problem, which asked whether every map drawn in the plane could be coloured using only four colours. This was settled positively by K. Appel and Haken in 1976 (Appel&; Haken 1978) by a proof which relied on a computer pro gram to check thousands of intricate map configurations too complex to be verified in any other way.1

A different kind of departure from traditional methods of logical validation made possible by the computer occurs with so-called probabilistic proof procedures, which run a series of spot checks on randomly selected fragments of long proofs and estimate to a given degree of confidence the correctness of the whole (Goldreich 1995). The question raised by probabilistic proofs , by the example of finite groups, and more pointedly by the Appel-Haken argument , is simple and profound : what is or should count as a proof? Traditionally, a proof had always been a persuasive argument, a logical narrative which could be checked and assented to, step by step, by an (in principle, *any) *individual mathematician. Evidently, such is not possible for the collection of all simple finite groups nor for the ensemble of maps that arise and need to be examined in relation to colouring. Plainly, these examples of augmented proof call for the idea of ‘the mathematician’ to be made explicit in a way that has not hitherto been necessary. Specifically, two issues arise: ‘who’ or what arc mathematical arguments addressed to and how is this addressee to be ‘persuaded’ by a putative proof? In other words, what is or should be the constitution of the subject, i.e. what cognitive capacities and technologies of inscription and representation should be available to it? How (according to what principles) might persuasion take place, once it is mediated by a computer program and so removed from the familiar (but in truth largely unexamined) orbit of an individual assenting to logical steps? Both these questions are about the couplings between the triad of agencies constituting the ‘mathematician’, with the second requiring the idea of persuasion as a dynamic involving the person and agent to be made explicit (Rotman 1993).

- (
*b) Global effects: simulation not proof*

Local effects, then, represent computer-inflected changes in mathematical content and forms of proof. In terms of our model of mathematical activity, they are princi pally the concern of the mathematical subject and its relation to the computational agent. 1fore far-reaching arc the epistemic changes involving imagining, understand ing, conceiving analogies, and having intuitions, insight and hunches which concern the character of the person and its relation to the agent. However the chief effect here-that of digital *simulation **–**reaches *beyond mathematics to cover the entire field of contemporary technological and scientifico-mathematical practice and is so widely operative and productive that it has been hailed as an entirely new, third method of scientific research, a 20th-century, specifically computational mode dif ferent from and not reducible to the two existing research modes-the experimen tal/ observational mode that founded science in the 17th century and the much older theoretical/deductive mode basic to mathematical reasoning (Kaufmann & Smarr 1993).

To describe the effects and scope of digital simulation it will be useful to note mathematics’ relation to physics. From the time of Galileo until well into the 20th century, the bulk of external problems and questions put to mathematics came from the natural sciences, principally physics. This, coupled with assumptions within physics of the smoothness and continuity of nature, inseparable from the infinite divisibility of space and time, legitimated and fostered a mathematics dominated by an infinite continuum of real numbers, infinite series, and the differential equa tions of calculus allied to these notions. But in order to operate at all, computers need to discretize calculus by converting these equations into finite-difference equa tions. But even without the effects of such discretization, physics’ employment of infinitary, continuum-based mathematics is no longer secure. ·wen before the advent of the computer, quantum physics was born on the overthrow of the assumption of infinite divisibility of the quantity physicists call ‘action’ and its replacement by the formalism of discrete quanta. Moreover, in empirical terms physics is not merely non-infinitistic but boundedly finite: such easily named quantities as meteors, 10-exponent 100 seconds, 10 exponent 100 particles or 10 exponent 100 light years have no physical meaning in the Universe we inhabit. Since the introduction of computational methods, the move against continuum-based mathematics has strengthened: various projects over the past three decades have argued why and how physics needs to be ‘fini tized’ (Greenspan 1973; Feynman 1982; Wheeler 1988; Landauer 1991). One of the most extreme and uncompromising finitization projects is Edward Fredkin’s (1992) hypothesis of ‘finite nature’ according to which ‘at some scale, space and time are discrete and that the number of possible states of every finite volume of space-time is finite’. Along with this goes the insistence that the entire physical Universe be understood, somehow, as a single device: a digital computing machine, the computa tions of which are the processes of the Universe and the software of which constitutes what we call the laws of physics.

With this in mind let us return to simulation-by which we mean a digital mock up or electronic model of a given physical or mathematical process. To mimic a process by computation, it is necessary to create a digital version of a typical state of the process, together with a rule for changing states, with the idea that repeat edly iterating the rule will give a simulation of the original process. To witness the mimicry, the results are visualized, that is, plotted or displayed graphically, so they present a picturable world: a visual, on-screen image whose electronic behaviour can be manipulated by software. A simulation is created, then, in two phases, discretiza tion and visualization, which together constitute a virtual world close enough to the original one (physical or mathematical) to be a valuable investigative and creative tool.

All of the above applies to mathematical as well as physical worlds. However, in the case of mathematics, where there is apparently no external world to be simulated and which is already in the business of creating, representing and studying virtual worlds, the importance and novelty of the simulational mode is easily missed. One reason is that the visualization it offers seems merely an extension (enormously refined and powerful, certainly) of time-honoured methods of plotting and curve drawing which, however useful as cognitive aids, are eliminable-adding nothing substantive to mathematics. Such a viewpoint, held unproblematically by most mathematicians, is misleading for several reasons.

Firstly, the very premise that diagram and curve drawing in mathematics are epiphenomena!, of undoubted psychological and practical value but of no epistemo logical worth, is a mistaken result of programmes of rigour, such as that of Bourbaki’s rewriting of mathematics within a set-theoretical framework, that neither know nor are interested in how mathematics is actually made or changes (Rotman 2000). Sec ondly, plotting and visualization techniques in the past were severely limited in scope and epistemic importance by the difficulty of performing more than a small number of the relevant calculations: a fact which made it easy to naturalize their results into a mere picturing of pre-existing ideas, to be jettisoned once the idea had been symbolized. Quite to the contrary, the computer creates complex topological surfaces, fractal functions, and iteration-based entities which were previously not only invisible but unimagined, unconceived. The consequent increase in mathematical knowledge has been considerable, ranging from the discovery of new topological surfaces to the opening up of novel research topics (Friedhoff & Benzon 1991; Hanson *et al.** *i995). Clearly, it would be a reductive misapprehension to think of simulation as ‘mere’ visualization. Thirdly, the very focus on the visualization phase allows digital simulation to be reduced to the business of picturing and so occludes the radically transformative changes set in train by the prior phase of discretization and the rcal time iteration of a rule it facilitates- both of which engage the finite discretum and not the infinite continuum.

But, leaving visualization aside, the style and character of a simulation-inflected mathematical science-pragmatic, material, experimental-breaks with mathemati cians’ traditional understanding of mathematics as a purely theoretical and deduc tive science. The advent of an electronically modified agent de-stabilizes the triadic assemblage of actors that has held the pure ‘mathematician’ in place since classical times. In the wake of this empirical, that is, impure, agent comes a digitally re figured person who can intuit, imagine and recognize the new sorts of mathematical objects, simulations and iterations, delivered by computation; and at the same time a digitally re-figured subject is required with the language and writing system the appropriate mathematical ‘software’-necessary to mediate the traffic between person and agent.

Traditional mathematics proves the existence of its abstract objects and establishes claims about them by constructing logically controlled narratives of imagined procedures using an agent who is an ideal, infinite possibility. Simulational math ematics substitutes empirical observation for logical validation and exhibits rather than r,roves the existence of its material objects via actual, real-time computations; its agent is a material, finite actuality.

Vhat, then, are the consequences for future mathematics of this difference? As we have already observed, there arc changes in practical methods of mathemat ical enquiry: empirical mathematics frees the person of the burden of proof and thereby liberates other, previously uninteresting or unrecognized, ways of mathe matical thought and activity. This has implications for the different kinds of objects and processes that the two styles can engage. Essentially, mathematical assertions are predictions about what will happen if certain specified operations are carried out. By insisting that its assertions be logically proved, classical pure mathemat ics restricts itself to investigating procedures of which the outcomes, in principle at least, can *be *predicted. It is not equipped to investigate the fate of incompressible processes whose outcomes cannot be foretold; programs that have to be run in order to reveal their final state; sequences of zeros and ones which are only specifiable by a rule equal in length to them. These are precisely the sort of mathematical object simulational-empirical mathematics handles well. And since there are in fact vastly more such incompressible processes and irreducible sequences than there are of the predictable kind, the territory of empirical mathematics is potentially huge com pared to that of the classical variety. This is essentially the claim Stephen vVolfram (2002) makes for his ‘new science’ based on simulational methods (computational iteration of simple programs such as cellular automata) which represents a ‘major generalization of mathematics -with new ideas and methods, and a vast new area to be explored’. Finally, to put the matter ontologically, in terms of what exists and is observable mathematically by virtue of the very materiality of computation, we can note the importance of actual versus ideal counting and iteration procedures: simulational mathematics achieves its difference and power precisely because it has access to effects the actual results of iterating real-world operations-unavailable to the idealized, necessarily internal, that is, rigorously separate from the world, regularities of classical reasoning.

# 4. Infinity in computer science

*(a) **An intractable problem*

As we have observed, for a physics increasingly committed to a discrete, quantized Universe, the use of infinity within its mathematical apparatus, via continuity, unlim ited divisibility and real numbers, raises certain problems. For theoretical computer science, such a consideration seems irrelevant, since its objects and processes and their mathematical formulations are discrete from the start .

But this does not settle the question of theoretical computer science’s relation to infinity: the science rests on the integers which form an infinite progression and this very attribute seems as intrinsic to computing as infinite divisibility was until recently to physics. In light of this, one can legitimately ask whether the presence of infinity within its mathematical formalism is, or might turn out to be, a problem for the science of computing. ·

I shall respond by concentrating on a combinatorial difficulty that goes back to the early days of the subject, which has subsequently given rise to a mathematized and highly generalized metaproblem, a problem about problems, namely the notoriously difficult *P *= *NP *question, ‘considered to be one of the most important problems in contemporary mathematics and theoretical computer science’ (Sipser 1992).

A famous example of the difficulty that *P *= *NP *mathematizes is the ‘travelling salesman problem’. A salesman has to visit each of *N *cities just once and return to his starting city. He has a table of all the inter-city distances and wants to calculate the length of the shortest circuit that will accomplish this. How should he proceed? The problem seems ideal for the mindless number crunching at which computers excel: the salesman tells the computer to find the length of every circuit in turn and output the length of the shortest. But this involves an extravagant computational cost as *N *increases. Thus, for small values, say *N *= 10, the number of circuits to check is *(N ***-1)! **= 9! or 362 880 circuits, a computational flea bite for present-day computers. For a slightly bigger value, say *N *= 20, the number is approximately 1017, a very large collection of circuits but checkable in a few months by an array of state-of the-art computers. For *N *= 50 or so, no present-day computer or any conceivable improvement thereof could complete a search through all possible circuits, while for *N *= 100 the number is outside the realm of the possible.

Though it features distances and circuits through cities, the underlying issue here is not tied to these notions. The difficulty facing the salesman surfaces in a vast web of seemingly unrelated problems, numbering several thousand, that arise naturally in graph theory, scheduling, routing and timetabling, in cryptography, in Boolean logic. and so on, over a wide range of disparate combinatorial situations. Some, like the salesman’s question , are optimization problems: they ask for the largest or smallest value of a numerical parameter, such as circuit length, flow through a network, the size of a subset, etc. Others are in the form of decision problems, which simply demand a yes or no answer, such *as *the ‘Boolean satisfaction problem’: given a Boolean expression in *N *variables, is there an assignment of 1 (true) or 0 (false) to the variables such that the whole expression has value **1? **The two forms are for the most part convertible into each other.

Regardless of their diverse contexts and origins, all these problems can be seen as posing the same basic question: is there a way of finding the optimum value or the yes/no answer that does not resort to the brute-force method of checking through all the possibilities? In computational terms, is there an algorithm accomplishing the relevant task that runs in a feasible amount of time? One that docs not, in other words, require an impossibly large number of computational steps for a small or reasonably small value of its input variable *N.*

To make this question meaningful, theoretical computer science requires a math ematical definition of ‘feasible’. Now, one cannot expect such a definition to address particular numerical facts (the computational leap from *N *= 10 to *N *= 50 we saw above) or, by the same token, to address what ‘reasonably small’ might mean. As a consequence, the question of feasibility is perforce posed generically, in terms of an unrestricted number variable *N, *and takes the form of a comparison between one type of mathematical function of *N *and another for all values of *N. *Thus, since brute force requires an exponential number of computational steps ( (*N *– 1)! for the salesman’s problem), one has to define ‘feasible’ to mean a number of steps bounded above by some function, *f **(N), *which must be guaranteed to be smaller than any exponential function of *N. *The decision by the computer-science community to choose *f *to be a polynomial function was very natural: polynomials are a robust class of functions, closed with respect to addition, multiplication and composition, whose well-known properties include the crucially relevant mathematical fact that any polynomial function is eventually smaller than any exponential function.

Defining feasible as polynomial time was the founding move of the theory of algo rithms, and though it attracted certain criticisms, the choice made possible the devel opment of complexity theory, an ‘elegant and useful theory that says something meaningful about practical computation’ (Papadimitriou 1994). Elegant certainly, but just how meaningfully related to practical computation is, as we shall see, a matter of dispute.

For theoretical purposes the yes/no decision problem formulation rather than the optimization formulation is more convenient. The class of all decision problems solv able in polynomial time is denoted by *P. *Observe that if a problem *D *belongs to *P, *that is, if there exists a feasible algorithm finding a solution *s *answering *D, *then *D *has the additional property that any proposed solution *s *to *D *can be checked for correctness in polynomial time. This is because one can run the algorithm solving *D *and compare the result with *s *to determine whether *s *is or is not a solution to *D, *all of which requires no more than polynomial time. The class of all such decision problems which can be checked polynomially is denoted by *NP. *Our observation amounts to the fact that Pis automatically a subclass of *NP. *The great open ques tion enshrined in the equation *P *= *NP *thus asks for the reverse inclusion: is *NP** *contained in *P? *In other words, if one can *verify *a possible solution to a decision problem *D *in polynomial time can one *find *a solution to *D *in polynomial time? Evidently, the concept of polynomial is the organizing idea behind the *P *= *NP** *problematic at every level; without it the question would collapse. This is so for several reasons. First, the problem itself, as a mathcmatization of specific combina torial situations such as the ‘travelling salesman problem’, depends totally on the polynomial-based definition of feasibility; in fact the problem is an artefact of this definition, brought into being with it and impossible to conceive without it. Second, its encapsulation of the many thousands of diverse concrete instances, of which the salesman’s problem is merely one, into a single meta-problem depends repeatedly on the unbounded freedom that results from polynomial composition, namely, if *f *(·) and *g(·) *are polynomial functions, then so is f*(g(·)). *Last, and by no means least, the formulation would make no sense in terms of its thousands of motivating concrete instances, all of which are resolvable by brute exponential-search routines, were it not for the mathematical fact that any polynomial function is smaller than any exponential function.

But it is this last, quite crucial fact that is precisely the site of a dissatisfaction with complexity theory, the point at which its practical relevance has been challenged and, not coincidentally, as we shall now see, where infinity enters the mathematical understanding of algorithms put in place by computer science.

( *b) Asymptotic growth*

Defining a feasible computation as one with a polynomial run-time rests, then, on a basic inequality-that polynomial functions are smaller than exponential ones. The guarantee behind the inequality lies in an asymptotic definition of ‘smaller than’ which refers to behaviour of functions ‘at infinity’. Specifically, a function *p( *·) is (asymptotically) smaller than a function *q(·) *if there exists some number *M *such

that *p(N) *< *q(N) *for all values of *N *greater than *M. *In other words, the definition

refers, in a way that cannot be eliminated , to *sufficiently large N **. *Thus, only in the limit, as *N *goes to infinity, is the relative size of two functions resolvable, so as to make it true that any polynomial function *p *is smaller than any exponential function *q.*

But suppose allowing passage to the limit, that is, invoking sufficiently large values of *N, *is ruled out? What if the values of *N *are constrained to lie below some fixed bound *B *In that case, the guarantee vanishes, *however large f3 *is: one can, in other words, no longer assume, in the presence of a bound, that polynomials are automat ically smaller than exponentials. In fact, under the condition of boundedness, the reverse can always be made to happen, polynomials can be *larger *than exponentials: for any exponential function *q( *·), it is always possible to find a polynomial *p( *·) such that *p(N) *> *q(N) *for all *N *below B*.*

The definition of feasibility is supposed to ensure achievable computation times. But what, in the real, finite world where the demand for it originates, is to be meant by achievable? For, on any presently conceivable understanding of physical reality, the existence of a bound *f3 *is inescapable. This will follow from local limits, such as those imposed by computer scientists not being able to wait for unspecified periods for algorithms to terminate, to global ones, such as the Universe we inhabit appearing not to permit an arbitrarily large sequence of physical actions to be executed.

The existence of a bound obviously disallows any appeal to an infinite range of values of *N *and so prevents the theory of complexity from operating. Indeed, the presence of a bound brings into relief a clutch of questions: what sense does it make for a science of computing to formulate its fundamental concept of feasibility in terms of an infinitistic criterion that necessarily appeals to numbers outside the reach of its computational processes? Pragmatically, what is the relevance of the *P *= *NP** *question for categorizing (let alone understanding) the possibilities and limits of computation? Why would its solution have anything to say about *real *feasibility, about the possibility of finding or not being able to find algorithms that worked within the limits and freedoms of the real world rather than those of the idealized, boundless universe of infinitistic mathematics?

These questions are far from idle. In cryptography, for example, one has to ensure that an encrypted message is safe from attack, that no method of decryption with

1686

a *practically achievable *running time exists, regardless of how such a method might be characterized mathematically. In such a context, equating ‘practically achievable’ with the existence of a polynomial-time algorithm makes little sense, and the entire *P *= *NP *problem begins to appear beside the point. Unsurprisingly, this is pre cisely the judgement voiced by many cryptographers and engineers more concerned with practicalities than mathematical elegance. Thus, for example, about the central cryptographic task of finding the factors of an integer, Steve Morgan observes even if factoring *is *polynomial, it isn’t necessarily practical. … A poly nomial algorithm of say the order of N20 is essentially intractable even for small values of *N. *Conversely, an algorithm that ran in the order of 2(0.lN) would make factoring billion bit numbers easy… (quoted in Rheinhold 1995).

Moreover, it is not even necessary to invoke high-degree polynomials to make the point; exactly the same objection arises from the possibility of large multiplicative constants being present. Thus, if *k *is such a constant, 1030 say, then an algorithm running in *linear *time, that is, of order *kN, *will be intractable for small values of *N, *whereas an exponential algorithm of the order of 2(N/K) would be tractable for all values of *N *that could possibly occur in practice.

These examples illustrate that not only is excluding exponential running times and identifying feasibility with polynomial time not to the point, but it runs counter to the underlying intuition relating to ‘smallness’, namely that if (the input to) a problem is small, one wants a feasible computation solving the problem to be small or at least near small. But while 10, say, is by common agreement a small value of *N, *one could not claim that 1020 (one hundred billion billion), let alone 1030 , is small or nearly so. These objections have been long known and long shrugged off by complexity theorists as the price of doing (asymptotic) business. If prodded, the only response they can offer is that N20 , though undeniably a polynomial function, is artificial, that in practice, most familiar algorithms are of the order *Nn, *where *n** *is possibly as large as five or six but certainly a number much smaller than 20, and that, likewise, large constants such as *k *instanced above arc simply not natural. But this appeal is hardly in good faith: the concepts of ‘natural’, and ‘in practice’, as well as a notion of ‘scale’ or ‘size’, and the opposition of ‘small’/’large’, are precisely what have been expunged from the theory’s asymptotic account of feasibility.

In fact, the *P *= *NP *question, and along with it the entire apparatus of algorithmic complexity and the asymptotic characterization of ‘feasible’ that engendered it, is not merely irrelevant to cryptography, but has been deemed conceptually deceptive. For Rheinhold, the apparatus is a ‘naked emperor’, whose effect has been that ‘a whole generation of computer scientists wrongly believes that complexity theory illuminates computation and that the *P *= *NP *problem is the missing link to a theoretical basis for cryptography’ (Rheinhold 1995).

Nor is it the case that cryptographers have special, ungeneralizable needs. Their criticism is sharpened perhaps by the financial, security, and military stakes involved in questions of decryption, but their overall objection stems from unavoidable prag matic demands; as such it arises in any technological or engineering arena that looks to computer science for algorithms to solve real-world numerical problems in real world times.

The presence of a bound, then, not only eliminates the possibility of any asymptot ically based, and in particular polynomial, criterion of feasibility, but it offers a basis to start re-thinking the question of feasible computation. Thus, as has been suggested by several researchers, one can divide the integers into two separate regions, H1 and H2, below and above */3, *respectively, and work inside H1 within a limited universe of discourse situated far below *{3. *Thus, Rheinhold (1995), for example, suggesting the for B*, *goes on to comment that ‘the undue attention paid to classical complexity theory arises from an inclination to assume results that are true in H2 must somehow be true in *H*l· But there is neither a theoretical nor practical basis for this belief.’

Indeed, there is not, and working below */3 *takes a first step toward re-thinking the feasibility question. It achieves this by making explicit the counterproductive nature of ‘letting *N *go to infinity’ in a directly practical way. More significantly perhaps, taking */3 *as a dividing point does induce a numerical scale or size into the proceedings and hence the possibility of defining ‘small’ and ‘large’ in relation to it. But, as it

stands, the approach is crude: the bound */3 *is imposed from outside and its value, chosen deliberately to be so large as to be remote from any realizable computation, is for this very reason unconnected to anything real or observable. Any scale thus induced will be extrinsic to such computations and their limits and unable to state (let alone address) the expectation, attached to the *intuitive *notion of feasibility, that computational time of a solution be of the same ‘order of magnitude’ (‘league’, ‘ballpark’, ‘size’, ‘numerical reach’) as the problem; that is, the expectation that, for ‘small’ *N, *the solution time of an N-sized problem should also be small or nearly so, but not in any event ‘large’.

The irrelevance to cryptography of the *P *= *NP *question reveals a gulf between an infinitistic mathematical concept of feasibility and an intuitive, pragmatic one: a gulf between the demands of real-world computation and the idealized model associated with the classical sequence of integers. It will be obvious that the gulf is a reoccur rence, posed in a particularly dramatic and focused form, of the theoretical difference between empirical and classical mathematics identified earlier. It would seem natural, then, to attempt to tackle the question of feasibility and hence perhaps the *P *= *NP *problematic anew using the resources and possibilities put into play by simulational methods. Whether this attempt to use empirical mathematics can yield dividends is surely knowable only by experiment, but it is not clear that the attempt makes much sense as an approach to the *P *= *NP *problem itself: the very status of this question as a pre-eminent unsolved problem of contemporary pure mathematics indicates how deeply folded it is within the world-view of classical mathematics. In which case, the problem might be the baby that empirical mathematics throws out along with the infinitistic bath water.

Granting this, is there a different mathematical framework for thinking ‘feasibility’, one which tries to navigate between the classical and empirical routes? If there is, finding it would seem to entail overcoming a large and difficult obstacle: how to effect a conceptual escape from the great attractor of the classical integers. No easy matter, since it necessitates refusing the core belief of classical mathematics, the time honoured *Dogma of natural numbers *(Rashevskii 1973), namely their ‘naturality’; their objective, non-human, existence as the ‘given’, ‘true’ and only idealization of actual counting and iteration.

1688

The integers embody a formalism and an idea, a syntax and a semantics; one can identify two types of attempt to escape from this dogma, depending on which aspect is given priority. The first makes questions of significance and meaning subsidiary by starting from a symbolic apparatus based on formal definitions and concepts able to be worked to produce mathematical results, and only subsequently does it evaluate the meanings thereby facilitated and imposed. The second reverses the procedure and insists on an intuition first, a convincing alternative picture of what counting and iteration are to mean before any particular formalism is offered. Examples of both can be found. The first approach has been pursued by Vladimir Sazanov, who introduces the inequality log log *N *< 10 as a formal axiom constraining the value of any *N* considered to be a feasible number, and proves various metamathematical theorems on the internal consistency of the effect of such a definition (Sazonov 1995, 1998). The second has been pursued by the present writer, who focuses on the semiotics of the process of repetition inherent to the integers to refuse their ‘naturality’, and outlines an alternative, non-classical understanding of counting and arithmetic (Rotman 1993, 1998). Whether either of these approaches will succeed in overcoming the dogma of their naturality remains to be seen; evidently, the stakes in challenging the notion are high.

# 5. Conclusion

Of course, the computer will transform classical mathematics; surely this is no longer in question. And it is reasonable to predict that it will do so along the same lines machinic, digital, material-which have determined the changes it has so far wrought. Let me summarize what these are.

The computer is a machine for investigating mathematical reality; it is reconfigur ing the mathematical imagining and ‘reasoning’ in relation to repetition as radically as the microscope/telescope reconfigured vision and ‘seeing’ in relation to scale; in its wake, mathematical thought will never be the same. Its machinic ability to repeat the application of a rule far beyond unaided human capacity enables it to tran scend the knowledge and proof available to such capacity, principally by engendering hitherto unknown and unsuspected mathematical objects, thus confirming the fact that repetition of the same can produce difference. In this it appears to replay, in a fantastically more complex arena, the primal creative act of mathematics, namely counting, which produces an endless stream of new, ever different numbers through the recursive iteration of the selfsame operation of adding the unit.

The computer is a machine whose conception, operation, construction and infras tructure are digital: it is antagonistic to the analog and to notions of smoothness and unbroken continuity associated with it. As a result it presents a conceptual and methodological opposition to the real number continuum which is the formalization within classical mathematics of these notions, and, less directly, to the concept of infinity folded into the very concept of real number. A major effect of this opposi tion to continuum-based mathematics, aided by independent moves within physics to quantize all the parameters measuring physical reality, as we have seen, is to push mathematics away from infinite methods in the direction of the discretely finite.

The computer is a material machine, as opposed to the ideal, immaterial object of classical mathematics, the Turing machine, on which it is modelled. There are two consequences of this fact. One, the topic of § 4, concerns a difficulty that arises when this difference is elided and the ideal, potentially infinite mathematics of the model is used to describe and investigate real-world computations. How seriously theoretical computer science takes this difficulty remains to be seen. The second consequence, more pertinent to the future of mathematics, stems from a kind of reversal of the first, namely capitalizing on the difference between real and imagined counting. The reversal works by actually running a program on a material machine and observing the result rather than theoretically running it on an ideal machine and predicting the result, a procedure which delivers into mathematics effects from the so-called real world: effects mathematical objects and relations-available to the mathematical person in no other way. The result is the opening up of mathematics by a powerful new agent in the direction of an empirical science: a development that not only goes against its time-honoured status as a purely theoretical pursuit, but is a dramatic and radical reversal efforts over the last two centuries to eliminate all traces of the physical world from mathematics’ definitions and methods in the name of an abstract programme of mathematical ‘rigour’.

Finally, a theme of this paper has been the impact of the computer on classical mathematics’ core concept of infinity. This should not be misunderstood. Certainly, the computer’s ongoing colonization of mathematics, particularly in the direction of empirical simulations, is bringing about a widespread de-emphasis of the concept. But to say this is not to decry the concept as such or interpret the computer’s impact as destroying, eliminating or deconstructing it. What is involved is not an attack on it, or a repudiation of its mathematical legitimacy and usefulness, or a refusal of the validity of infinitary mathematics pursued within its own terms, but rather a fresh realization or recognition of its deeply ideal status in the light of computational methods. In the light, that is, of an uncompromisingly finite agent who, as a result of this very finitude is able to sanction use of the concept on the level of the subject and person, but as a convenient way of speaking, a permitted abuse of language, that despite its ‘literal’ meaning has no empirical content. However, it would be naive to think that such a recognition will not impinge ultimately on the status and unchallenged cultural prestige of classical, infinitistic mathematics.

It is a pleasure to acknowledge my gratitude to Professor Alistair MacFarlane for his illuminating comments and suggestions on a draft of this paper.

# References

Appel, K. & Haken, W. 1978 The four-color problem. In *Mathematics today *(ed. L. A. Steen), pp. 153-190. Springer.

Feynman, R. 1982 Simulating physics with computers. *Int. **J. **Theor. Phys. ***21, **467-468. Fredkin, E. 1992 Finite nature. (Available at http://www.im.lcs.mit.edu/poc/fredkin/Finite

Nature.)

Friedhoff, R. & Benzon, W. 1991 *Visualization: the second computer revolution. *San Francisco, CA: W. H. Freeman.

Goldreich, 0. 1995 *Probabilistic proof systems. *Electronic Colloquium on Computational Com plexity Lecture Notes Series. (Available at http://www.eccc.uni-trier.de/eccc-local/ECCC LectureNotes/goldreich/oded.html.)

Greenspan, D. 1973 *Discrete models. *Addison-Wesley.

Hanson, A. J., Munzner, T. & Francis, G. 1995 Interactive methods for visualizable geometry.

*Computer ***27, **73-83.

1690 *B. Rotman*

Kaufmann, W. & Smarr, L. 1993 *Supercomputing and the transformation of science. *Scientific American Library, vol. 43. San Francisco, CA: Freeman.

Landauer, R. 1991 Information is physical. *Phys. Today *44, 23-29.

McClure, W. 1997 Solution of the Robbins problem. *J. **Autom. Reasoning ***19, **263-276. Papadimitriou, C. 1994 *Computational complexity. *Addison-Wesley .

Rashevskii, P. K. 1973 On the dogma of natural numbers . *Russ. Math. Suru. ***28, **143-148. Rheinhold, A. 1995 *P *= *NP *does not affect cryptography. Usenet newsgroup posting. (Available

at http:/ /world.std.com/-reinhold/p=np. txt.)

Rotman, B. 1993 *Ad infinitum: the ghost in Turing’s machine. *Stanford, CA: Stanford University Press .

Rotman, B. 1998 Response to Patrick Peccatte. Posting to Foundations of Mathematics List, 5/14/1998 . (Available at ftp://ftp.psu.edu/simpson/fom/postings on subscription.)

Rotman, B. 2000 *Mathematics as sign. *Stanford, CA: Stanford University Press.

Sazonov, V. 1995 On feasible numbers. In *Logic and computational complexity *(ed. D . Leivant).

Lecture Notes in Computer Science, vol. 960, pp. 30-51.

Sazonov, V. 1998 Reply to Rotman . Posting to Foundations of Mathematics List, 8/31/98. (Available at ftp://ftp .psu.edu/simpson/fom/postings on subscription.)

Sipser, M . 1992 The history and status of the *P *versus *NP *question. In *Proc. 24th Ann. ACM** **Symp. on on **TheOT’IJ **of Computing, Victoria, Canada, *pp. 603-618.

Soloman, R. 1995 On finite simple groups and their classification. *Not. Am. Math. Soc. *42, 231-39 .

Wheeler, J. **H****. **1988 World as system self-synthesized by quantum networking. *IBM **J. **Res.*

*Develop. ***32, **4-15.

Wolfram, S. 2002 *A new kind of science. *Champaign, IL: Wolfram Media.