Chapter 7: The Mathematician at the Wall

Thesis: Mathematics presents the unverifiable in its purest form (intractable, and sometimes undecidable). The responses that work are to verify a finite slice and prove a bound (a certificate), to swap the target for an equivalent but more tractable statement (proxy substitution), and to use probabilistic methods that accept an $\varepsilon$ of error (the probabilistic method).

At the Wall

What makes a hard problem hard is often not how difficult it is, but that you cannot measure how far away from it you still are.

A climber can see the summit. Someone debugging a program gets error messages. Both at least know whether they are heading the right way, whether they are getting closer or drifting farther. Proving a mathematical conjecture gives you none of this. You might be one idea away from the answer, or a century away, and you have no instrument that tells you which. In his eight-page paper of 1859⁹, Riemann set down the conjecture later named after him almost as an aside. He then added that one would of course wish for a rigorous proof, but that after a few fleeting, futile attempts he had provisionally set the matter aside, since it was not necessary for the immediate object of his inquiry. It has stayed set aside for more than a hundred and sixty years.

This chapter looks at what mathematicians actually do in front of this wall. I chose mathematics as a setting because it offers the unverifiable in its purest form. There is no hidden state here, no opponent out to deceive you, no excuse that time ran short. A statement is either true or false, with nothing in between. And yet even in this cleanest of places, verification stays systematically out of reach. If we can see clearly how skilled people act here, we gain a point of reference for what happens in the other, messier settings of this part: the person at the console, the agent let loose, and the organization that cannot see itself.

The Verification Gap

First, let us pin down the shape of the wall.

Checking a proof is easy; finding one is hard. Hand someone a fully written-out formal derivation, and checking it line by line against the axioms and rules of inference is mechanical work. In principle a machine could do it, and it is guaranteed to return a yes or a no in finitely many steps. Finding that derivation is another matter. This asymmetry is the foundation of the whole chapter.

The asymmetry has a precise logical form. In 1936 Church and Turing each proved that the decision problem (the Entscheidungsproblem) has no solution. No algorithm can decide, for an arbitrary first-order statement, whether it is logically valid, or equivalently (by Gödel's completeness theorem) whether it is provable. Take a system that is sufficiently rich, recursively axiomatizable, and consistent, such as Peano arithmetic or ZFC. The set of its theorems is recursively enumerable but not recursive. You can list all of its proofs, one after another, yet no program can reliably tell you when a statement is not a theorem. Checking is decidable, theoremhood is undecidable, and the difference between the two is the gap.

One layer deeper (skippable): The qualifier "sufficiently rich" is doing real work. Some theories really are decidable: Presburger arithmetic (the natural numbers with addition only), and Tarski's theory of real closed fields. In those worlds a decision procedure exists, and every statement can be settled mechanically. Undecidability is not the universal fate of logic. It is the price you pay once expressive power crosses a certain threshold. Analytic number theory, where the Riemann hypothesis (RH) lives, sits far above that threshold.

Sharper still, it is not only proof search that is hard. Even the question "am I close?" has no decision procedure. That is what makes RH so tormenting: it resists not only solution but any estimate of progress. Among the five faces of Chapter 2, this chapter stands on the border between undecidable and intractable. Some problems have no procedure even in principle. Others have one, but it would cost too much to run within the lifetime of the universe. You cannot see what lies behind the wall, so a good mathematician stops chiseling straight at it and changes stance. The three stances below will come back in other chapters of this book, under other names.

Certificates and Bounds

The first stance: stop trying to verify the whole. Verify only a slice, and prove a guaranteed bound for it.

Go back to the $\zeta$ function. Riemann took the series Euler had linked to the primes, extended it to a function on the whole complex plane, and wrote down its completed form

$$\xi(s) = \tfrac{1}{2}\,s(s-1)\,\pi^{-s/2}\,\Gamma!\left(\tfrac{s}{2}\right)\zeta(s),\qquad \xi(s)=\xi(1-s).$$

This symmetric functional equation folds the critical strip onto itself, left onto right. The conjecture says that all the nontrivial zeros of $\zeta$ lie on the critical line $\operatorname{Re}(s)=\tfrac12$ (the trivial zeros sit at the negative even integers). The quantifier "all" is where the unverifiability lives.

Finitely many zeros, though, can be verified, and here two things that often get run together have to be kept apart. In 2004 Gourdon computed the first $10^{13}$ zeros and found them all on the line¹⁷. That was numerical computation, done in high-precision floating point. It inspires extremely strong confidence, but it is not a certificate, because it does not rigorously bound the rounding error. What Platt did in 2017 was something else¹⁸. Using interval arithmetic, he rigorously confined every zero with imaginary part up to about $3.06\times10^{10}$ to the critical line, and in 2021 Platt and Trudgian pushed that height to $3\times10^{12}$. Only the last two results are certificates: bounded, local, mechanically recheckable guarantees. Such a guarantee is true, but it is not a theorem. However high the verified zeros go, they never cross the threshold of "all."

This stance has respectable precedents in mathematics. In 1896 Hadamard and de la Vallée Poussin each proved the prime number theorem $\pi(x)\sim x/\ln x$. They relied on a bound much weaker than RH but within reach: $\zeta$ has no zeros on the line $\operatorname{Re}(s)=1$. If you cannot prove that all the zeros sit at $\tfrac12$, first prove that none of them sit on the line $1$. This trades a weak statement you can prove for a stretch of real progress on the road to the strong one.

Carry the same stance over into software and it looks familiar at once. A type system does not prove a program "entirely correct." It proves only one property (that an integer will not be dereferenced as a pointer), and in exchange it gets a check that is decidable. Formal verification goes further. In 2017 Hales's team published a machine-checkable proof of the Kepler conjecture²⁵, compressing an argument that human referees had fought over for years into a certificate that can be verified line by line. The price is always the same. A certificate buys certainty on a slice and pays for it in generality, and a slice is not a theorem. So some people turn instead to moving the target itself.

Proxy Substitution

The second stance: stop defending the original statement to the last, and swap it for an equivalent statement that is easier to handle.

RH has an astonishing number of equivalent reformulations. In 1997 Xian-Jin Li gave a criterion¹⁴: RH holds if and only if the real numbers $\lambda_n$ are nonnegative for all $n\ge1$, where

$$\lambda_n=\sum_{\rho}\left[1-\left(1-\tfrac{1}{\rho}\right)^{n}\right],$$

with the sum taken over the nontrivial zeros. A geometric statement about where the zeros lie becomes a statement about the positivity of a sequence of numbers. Nyman and Beurling gave another: RH is equivalent to the indicator function $\chi_{(0,1)}$ lying in the $L^2(0,1)$ closure of the span of a family of dilated fractional-part functions. That turns the zero problem into an approximation problem. In 2003 Báez-Duarte tightened this into a sequence version that uses only integer dilations¹⁶, corresponding to a sequence of distances $d_n\to0$. And in 2002 Lagarias gave an equivalent statement elementary enough to write on a postcard³²: for all $n$, $\sigma(n)\le H_n+\exp(H_n)\ln H_n$, where $H_n$ is the harmonic number and $\sigma$ is the sum of divisors.

I have walked part of this road myself. I carried RH into Li's criterion, into the Nyman-Beurling-Báez-Duarte approximation framework, and then into the language of operator spectra and stochastic processes. Each time I nursed the same hope: in a new language, the difficulty might show a handle in the new coordinates. And each time I reached the same conclusion. The equivalence was real, and the difficulty had not shrunk by a single ounce. I had not unlocked the problem. I had only renamed it and dressed it in new clothes.

This is the standard way proxy substitution fails in mathematics, and it deserves an accurate name: a faithful but no-easier proxy. The equivalence guarantees that it still points at the same true target (faithful), yet it is not one bit more tractable than the original (no easier). Whether the move succeeds depends entirely on getting faithfulness and greater tractability at the same time, and getting both is extremely rare. It is so rare that this is where the whole craft lies.

Lay these two dimensions out as a table, and a thread running through this chapter comes into view:

	Easier	No easier
Faithful	The ideal proxy (rare; this is where the whole craft lies)	Mathematics' equivalent reformulations: you have only renamed the difficulty (this chapter)
Unfaithful	Goodhart: you optimize the proxy, and the true target rots (Chapters 8, 11)	Useless; nobody would want it

The mathematician runs aground at the top-right end of the diagonal: faithful but no easier. Later, in the chapter on organizations, we will run aground at the other end, with a proxy that is easier but unfaithful. You optimize it with all your might while the thing you actually care about goes bad. It is the same move, failing in two opposite directions. Chapter 11 will formally join these two ends. For now, remember one thing. Swapping the target is neither cheating nor a way out. It is a stance, and whether it works is a separate question.

The Probabilistic Method

The third stance goes most against the grain of mathematics. Stop demanding a two-valued verdict. Hold a calibrated probability instead, and accept a bounded risk of error in order to act.

Primality testing is the cleanest example. Deterministic algorithms for deciding whether a large number is prime are costly, so Miller-Rabin changes the question. If $n$ is composite, a randomly chosen base lets it slip through with probability at most $1/4$, and $k$ independent rounds drive the chance of a wrong verdict down to $\le (1/4)^k$. Solovay and Strassen had earlier given a version with error $\le (1/2)^k$²⁰ (in 1977); Rabin's version dates from 1980¹⁹. "Prime with probability $1-\varepsilon$" is a fundamentally different kind of knowledge from "proven prime." But it is good enough for engineering, and you can make it as sharp as you like: just add a few more rounds.

What matters is to see clearly what has been given up. What is given up is the kind of certainty, not the rigor. The bound $(1/4)^k$ is itself a theorem, and its proof is airtight. You have not lowered the standard. You have swapped it for another standard, one that can be delivered within budget. The probabilistic method has a place in pure mathematics as well. Erdős's probabilistic method (which Alon and Spencer wrote up as a classic²⁶) can prove that an object exists by proving that the probability of its appearing is positive, without handing you the object itself. Existence is proved; the construction is absent.

This stance reaches all the way to the belief in RH itself. In 1973 Montgomery studied the pair correlation of the zeros³⁰. He derived, and conjectured, that the normalized pair-correlation function of the zeros takes the form

$$R_2(u)=1-\left(\frac{\sin \pi u}{\pi u}\right)^{2}.$$

At Princeton, Dyson saw at a glance that this is the pair correlation of the eigenvalues in the Gaussian unitary ensemble (GUE) of random matrices. In 1987 Odlyzko ran the numbers on a vast number of zeros³³ and confirmed the agreement to a breathtaking degree. None of this is a proof, but it is extremely strong evidence. It leads mathematicians to believe RH, and to believe it in a way no different in essence from the way a physicist believes a law that has not yet been falsified. Mathematics, it turns out, has grown its own method for forming beliefs where verification is out of reach.

How Mathematicians Judge

This brings us to the most human part of the chapter. When the oracle never comes, what do people with good judgment rely on to hold a belief and to decide where to push?

Mathematics is deductive on the outside and plausible on the inside. Pólya wrote Mathematics and Plausible Reasoning² to describe how mathematicians, lacking a proof, size up a statement through analogy, induction, and special cases. Hadamard surveyed the psychology of mathematical invention³ and recorded its rhythm of incubation and sudden insight. Poincaré left us his famous moment⁵: the instant his foot touched the step of an omnibus, the connection between Fuchsian functions and non-Euclidean geometry came to him without warning. None of these replaces proof. They are what a person actually relies on along the stretch of road before the proof arrives, where there are no instruments.

Why do mathematicians believe RH before anyone has proved it? Because the evidence piles up from every direction, and each strand supports the others. A vast number of zeros have been verified on the line. There are many equivalent forms, and none has collapsed. An analogue has already been conquered (the Riemann hypothesis for varieties over finite fields, proved by Deligne). And the statistics of the zeros match the predictions of random matrices with great precision. No single one of these is a proof, but together they make up a disciplined belief.

The mathematical community has also reflected on the standing of this kind of belief. Thurston's 1994 essay "On Proof and Progress in Mathematics"²¹ argues that what mathematics advances is human understanding, not just a stock of formal proofs. The 1993 debate between Jaffe and Quinn over "theoretical mathematics"²² asked how far conjecture-driven, evidence-first work counts as mathematics. Tao asked what good mathematics is²⁹, and not one item in his answer is "already proved." Put these side by side, and the ability to judge that "this statement is most likely true, and worth investing in" turns out to be a calibrated belief in its own right, which is what Part IV will name directly. People who handle this well are not paralyzed when the oracle is absent, and they do not pretend to certainty either. They hold a belief with its degree of confidence marked, and then act all the same.

What Comes Next

Up against the unverifiable in its purest form, mathematicians never got a decision procedure. What they got instead was this: certificates (verify a slice, prove a bound); proxy substitution (swap the target for an equivalent statement, and admit that it is often faithful but no easier); probabilistic acceptance (give up the two-valued verdict, hold a calibrated probability, and act); and the judgment to hold a belief before the proof.

None of these is an emergency measure peculiar to mathematics. Caging untrusted code in a sandbox, auditing an organization, reading the unspoken preferences of a user from behind an interface: in each case, what people reach for is the same small set of things, in different jargon. Part III will take each move out of the field where it grew, give it its own name, and set it side by side across fields. That comparison table is the heart of the whole book.

One more thing needs saying plainly. I cannot verify this book's own central claim either, the claim that "responses converge on the same small set." My belief in it is the same kind of thing as a mathematician's belief in RH: a belief built on evidence from many fields, with its degree of confidence marked, but without a proof. Chapter 14 will return to this, and there the book will do for itself the very thing it has been describing all along.

Next chapter: 8. The Organization Blind to Itself →← 6. The Agent Released

References

Waypoints: 1. historical scientific judgment; 2. theoretically studied material; 3. how science progresses; 4. how to live in an unverifiable world. This section was checked source by source.

G. Pólya (1945). How to Solve It: A New Aspect of Mathematical Method. Princeton University Press. doi:10.1515/9781400828678 [1] Pólya breaks mathematical problem-solving into four stages: understanding the problem, devising a plan, carrying it out, and looking back, and systematically lays out heuristic strategies such as analogy, special cases, working backward, and auxiliary problems. What he writes is not the proof of theorems but the instrumentless process of exploration before a proof is found, exactly what the section "How Mathematicians Judge" cares about.
G. Pólya (1954). Mathematics and Plausible Reasoning (2 vols.). Princeton University Press. Google Books [1][4] The two volumes treat, respectively, induction and analogy in mathematics and the logical structure of plausible reasoning, arguing that before they reach a rigorous proof mathematicians form their belief in a proposition by observing special cases, inducing patterns, and weighing evidence. The main text borrows from it to point out that "mathematics is deductive on the outside and plausible on the inside," and it is the source reading for understanding the concept of plausible reasoning.
J. Hadamard (1945). An Essay on the Psychology of Invention in the Mathematical Field. Princeton University Press. Google Books [1] Hadamard surveyed the creative psychology of mathematicians and distilled the rhythm of discovery into preparation, incubation, illumination, and verification, stressing subconscious work and the unheralded flash of inspiration. It provides a first-hand psychological study for this chapter's account of Poincaré-style insight, showing that much of mathematical judgment happens outside consciousness and outside proof.
H. Poincaré (1902). La Science et l'Hypothèse. Flammarion. Google Books [1] In this classic of the philosophy of science, Poincaré discusses the standing of mathematical hypotheses, conventions, and geometry, arguing that many foundational choices are not imposed by experience but adopted out of convention and convenience. It shows how a leading mathematician reflects on the epistemological foundations of his own discipline, in tune with this chapter's concern with how to hold a belief where verification is unavailable.
H. Poincaré (1908). Science et Méthode. Flammarion. Google Books [1] The passage in which inspiration arrives as he steps onto the running board of an omnibus is the most often cited first-hand record in the psychology of mathematical discovery, by which Poincaré dissects the role of intuition, choice, and the subconscious in creation. The main text uses this moment directly, to show how insight connects scattered threads without any prior sign.
G. H. Hardy (1940). A Mathematician's Apology. Cambridge University Press. Google Books [1] Hardy defends the value of pure mathematics, holding that good mathematics lies in its seriousness, depth, and inevitable beauty, not in utility. As the introspection of a great number theorist on his own craft, it defines on what grounds a mathematician judges whether a piece of work is worth doing, of a piece with the question that closes this chapter, "what is good mathematics."
E. P. Wigner (1960). "The Unreasonable Effectiveness of Mathematics in the Natural Sciences." Communications on Pure and Applied Mathematics, 13(1). doi:10.1002/cpa.3160130102 [2][3] Wigner marvels that abstract mathematical concepts can describe the physical world with such precision, calling this fit a strange gift we neither understand nor deserve. The puzzle this short essay poses has no agreed answer to this day; for this chapter it exemplifies how a belief in a deep regularity can be taken seriously in the absence of proof.
I. Lakatos (1976). Proofs and Refutations: The Logic of Mathematical Discovery. Cambridge University Press. doi:10.1017/cbo9781139171472 [1][3] Lakatos, taking Euler's polyhedron formula as his example, re-enacts through a fictional classroom dialogue how definition, proof, and counterexample mutually revise one another and together advance mathematics. It overturns the stereotype of mathematics as once-and-for-all deduction, presenting knowledge as growing crookedly through conjecture and refutation, exactly suited to this chapter's concern with the true shape of mathematical progress.
B. Riemann (1859). "Über die Anzahl der Primzahlen unter einer gegebenen Größe." Monatsberichte der Berliner Akademie. link [2][3] Riemann's eight-page paper continues the $\zeta$ function onto the complex plane, gives the functional equation, and links the distribution of primes to the nontrivial zeros of $\zeta$; the conjecture about the location of the zeros written down there in passing is the later Riemann hypothesis. It is the source text of the whole chapter, and the line quoted at this chapter's opening about "set aside provisionally after futile attempts" comes from here.
H. M. Edwards (1974). Riemann's Zeta Function. Academic Press. Google Books [2] Edwards's monograph develops around Riemann's original 1859 paper, gradually laying out the theory of the $\zeta$ function, the prime number theorem, and the origins of the Riemann hypothesis, attending to both historical context and technical detail. It is the classic introduction to the $\zeta$ function and RH, providing reliable background for the concepts this chapter touches, such as zeros and the critical line.
E. C. Titchmarsh, rev. D. R. Heath-Brown (1986). The Theory of the Riemann Zeta-function (2nd ed.). Oxford University Press. Google Books [2] This is the standard advanced monograph on the analytic theory of the $\zeta$ function, systematically treating the distribution of zeros, zero-density estimates, and mean-value theorems on the critical line, with Heath-Brown's revision adding more recent advances. It represents the sum of the technical results rigorously established around RH, and is the specialist background reference when this chapter discusses "certificates and bounds."
E. Bombieri (2000). "Problems of the Millennium: The Riemann Hypothesis." Clay Mathematics Institute. link [2][4] This is the official problem statement of RH written for the Clay Mathematics Institute's Millennium Prize Problems, in which Bombieri concisely sets out the conjecture's origins, precise formulation, and weight in number theory. It is the authoritative entry point for understanding why RH is ranked among the problems of the century, and this chapter's assessment of RH's standing finds its warrant here.
J. B. Conrey (2003). "The Riemann Hypothesis." Notices of the American Mathematical Society, 50(3). link [2][3] Conrey's survey, aimed at a broad readership, marshals the various kinds of evidence supporting RH, including the numerical verification of zeros, the agreement with random matrix theory, and the proven analogue over function fields. It gathers in one place the evidence on which this chapter's three stances rely, and is a convenient read for understanding why the mathematical community believes RH.
X.-J. Li (1997). "The Positivity of a Sequence of Numbers and the Riemann Hypothesis." Journal of Number Theory, 65(2). doi:10.1006/jnth.1997.2137 [2][4] Xian-Jin Li proves that RH is equivalent to a sequence of real numbers $\lambda_n$ defined from the zeros being nonnegative for all $n$, translating the geometric statement about the location of the zeros into a positivity criterion for a sequence. The main text takes it as the foremost example of proxy substitution, showing how an equivalent rewriting can be faithful yet not necessarily easier.
E. Bombieri & J. C. Lagarias (1999). "Complements to Li's Criterion for the Riemann Hypothesis." Journal of Number Theory, 77(2). doi:10.1006/jnth.1999.2392 [2][4] The two authors point out that Li's criterion is in fact a special case of a family of general inequalities for an arbitrary multiset of complex numbers, not specific to the $\zeta$ function, and, using the Guinand-Weil explicit formula, give an arithmetic expression for $\lambda_n$, connecting it to Weil's criterion for RH. It deepens the understanding of Li's criterion and shows how the same equivalent proposition is rewritten again and again across different languages, an extension of this chapter's theme of proxy substitution.
L. Báez-Duarte (2003). "A Strengthening of the Nyman-Beurling Criterion for the Riemann Hypothesis." Atti della Accademia Nazionale dei Lincei, Rendiconti Lincei Mat. Appl., 14(1). arXiv:math/0202141 [2][4] Báez-Duarte tightens the Nyman-Beurling approximation criterion into a sequence version using only integer dilations, making RH equivalent to a sequence of approximation distances $d_n$ tending to zero. It is yet another equivalent rewriting listed in this chapter, carrying the zero problem into the framework of $L^2$ approximation, and likewise confirms that a faithful proxy is often no easier to solve.
X. Gourdon (2004). "The 10^13 First Zeros of the Riemann Zeta Function, and Zeros Computation at Very Large Height." Online technical report (numbers.computation.free.fr). link [2][4] Gourdon, using high-precision floating-point computation with the Odlyzko-Schönhage algorithm, verified that the first $10^{13}$ zeros all fall on the critical line. This chapter deliberately sets it against Platt: it gives extremely strong numerical confidence but does not rigorously bound the rounding error, and so it is a numerical result rather than a mechanically recheckable certificate.
D. J. Platt (2017). "Isolating Some Non-trivial Zeros of Zeta." Mathematics of Computation, 86(307). doi:10.1090/mcom/3198 [2][4] Platt uses interval arithmetic to isolate the zeros rigorously on the critical line, giving the error a provable upper bound and thereby raising numerical verification to a mechanically recheckable certificate. This chapter uses it to exemplify the stance of "certificates and bounds": prove not the whole, but a guaranteed bound for a finite slice.
M. O. Rabin (1980). "Probabilistic Algorithm for Testing Primality." Journal of Number Theory, 12(1). doi:10.1016/0022-314x(80)90084-0 [2][4] Rabin gives the Miller-Rabin primality test: if $n$ is composite, a randomly chosen base slips past it with probability at most $1/4$, and $k$ independent rounds drive the chance of misjudgment down to $(1/4)^k$. This chapter uses it as the cleanest example of the probabilistic method, showing that the error bound itself is a rigorously proved theorem, and what is given up is the kind of certainty, not the rigor.
R. Solovay & V. Strassen (1977). "A Fast Monte-Carlo Test for Primality." SIAM Journal on Computing, 6(1). doi:10.1137/0206006 [2][4] Solovay and Strassen earlier proposed a probabilistic primality test based on the Jacobi symbol, with a single-round chance of misjudgment of at most $1/2$, one of the founding works of randomized algorithms. This chapter places it alongside Rabin's version, showing that trading the probabilistic method for a decision deliverable within budget has long had precedents in computational number theory.
W. P. Thurston (1994). "On Proof and Progress in Mathematics." Bulletin of the American Mathematical Society, 30(2). doi:10.1090/s0273-0979-1994-00502-6 [1][3] Thurston argues that what mathematics truly advances is human understanding of mathematics, not merely the inventory of formal proofs, and that proof is only a socialized means by which the community transmits and confirms understanding. This chapter cites it at the close to challenge the narrowing view that "mathematics equals proven theorems," and it is essential reading for reflecting on the standing of proof.
A. Jaffe & F. Quinn (1993). "'Theoretical Mathematics': Toward a Cultural Synthesis of Mathematics and Theoretical Physics." Bulletin of the American Mathematical Society, 29(1). doi:10.1090/s0273-0979-1993-00413-0 [1][3] Jaffe and Quinn propose distinguishing "theoretical mathematics" from rigorous mathematics, suggesting that conjecture-driven, not-yet-rigorously-proved work be explicitly labeled so as not to erode the reliability of mathematics, which set off a widely noted debate in the community. This chapter borrows from that debate to ask to what extent evidence-first work counts as mathematics, cutting right to the whole book's concern with verification and belief.
J. von Neumann (1947). "The Mathematician." In R. B. Heywood (ed.), The Works of the Mind. University of Chicago Press. Google Books [1][3] In this essay von Neumann reflects on the nature of mathematics, on how mathematics travels back and forth between abstraction and empirical sources, how it chooses its direction by aesthetic standards, and why drifting too far from the empirical source carries the risk of degeneration. From the vantage of a master who ranged across many fields, it shows the weight of aesthetics and taste in mathematical judgment, echoing this chapter's discussion of how mathematicians decide where to push.
K. Appel & W. Haken (1977). "Every Planar Map Is Four Colorable, Part I: Discharging." Illinois Journal of Mathematics, 21(3). doi:10.1215/ijm/1256049011 [2][3] Appel and Haken, with the aid of a computer-checked set of unavoidable configurations, proved the four color theorem, the first famous mathematical proof to depend essentially on a computer. It raised a debate that continues to this day: whether a proof no human can read line by line still counts as a proof, directly bearing on this chapter's discussion of certificates and mechanically recheckable guarantees.
T. Hales et al. (2017). "A Formal Proof of the Kepler Conjecture." Forum of Mathematics, Pi, 5. doi:10.1017/fmp.2017.1 [2][3][4] Hales's team, in the Flyspeck project, used the HOL Light and Isabelle proof assistants to complete a fully formalized, mechanically checkable proof of the Kepler conjecture, settling the unresolved status the original proof had been left in because human referees could not fully check it. This chapter uses it to show how formal verification compresses a contested argument into a line-by-line verifiable certificate.
N. Alon & J. H. Spencer (1992). The Probabilistic Method. Wiley. Google Books [2][4] This classic systematically presents the probabilistic method pioneered by Erdős: to prove that some combinatorial object exists, prove that the probability of its appearing at random is positive, and thereby conclude that it must exist, while often being unable to construct it explicitly. This chapter borrows from it to point out the feature of the probabilistic method in proving existence in pure mathematics: existence proved, construction absent.
P. J. Davis & R. Hersh (1981). The Mathematical Experience. Birkhäuser. Google Books [1][3] Davis and Hersh, starting from the actual experience of mathematicians, discuss the existential status of mathematical objects, the role of proof, and the philosophical situation of mathematics, presenting a practitioner's perspective different from formalist dogma. It provides this chapter with a close-to-the-ground reflection for understanding how mathematicians hold beliefs in practice and regard truth.
W. T. Gowers (2000). "The Two Cultures of Mathematics." In Mathematics: Frontiers and Perspectives. American Mathematical Society. link [1][3] Gowers distinguishes two cultures within mathematics: theory-builders and problem-solvers, the former solving problems for the sake of understanding, the latter understanding for the sake of solving problems, with algebraic geometry, the Langlands program, and combinatorial number theory as contrasts. It shows that mathematicians can hold different measures of what is deep and what is good work, echoing this chapter's discussion of the standards of mathematical judgment.
T. Tao (2007). "What Is Good Mathematics?" Bulletin of the American Mathematical Society, 44(4). doi:10.1090/s0273-0979-07-01168-8 [1][3] Tao enumerates the many mutually distinct dimensions of good mathematics, from rigor, depth, and beauty to richness of application and the power to open new directions, arguing that no single standard exists and that, over the long run, these qualities often pull on one another. This chapter borrows from it to show that judging whether a piece of work is worth investing in is itself an ability, and that not one of its standards is "already proved."
H. L. Montgomery (1973). "The Pair Correlation of Zeros of the Zeta Function." In Analytic Number Theory, Proc. Sympos. Pure Math., XXIV. American Mathematical Society. doi:10.1090/pspum/024/9944 [2][3] Montgomery studies the normalized pair correlation of the zeros of $\zeta$, deriving and conjecturing its form, and Dyson at once recognized this as exactly the pair correlation of the eigenvalues of the random-matrix Gaussian unitary ensemble. This linkage opened the deep connection between number theory and random matrix theory, and is one of the key pieces of evidence when this chapter discusses on what the belief in RH is built.
P. Sarnak (2004). "Problems of the Millennium: The Riemann Hypothesis." Clay Mathematics Institute. link [2][3][4] Sarnak's note for the Clay Institute emphasizes the generalized forms of RH and its central role in analytic number theory, explaining why so many other results take it as a premise. From the vantage of an expert active in the connection between zero statistics and random matrices, it rounds out this chapter's understanding of RH's importance and its web of evidence.
J. C. Lagarias (2002). "An Elementary Problem Equivalent to the Riemann Hypothesis." The American Mathematical Monthly, 109(6). doi:10.1080/00029890.2002.11919883 [2][4] Lagarias gives an elementary inequality using only the harmonic numbers and the sum-of-divisors function, proving that it holds for all $n$ if and only if RH holds, translating the abstruse zero problem into an arithmetic statement that could almost be written on a postcard. This chapter uses it to show that proxy substitution can be elementary on the surface while the difficulty remains undiminished.
A. M. Odlyzko (1987). "On the Distribution of Spacings Between Zeros of the Zeta Function." Mathematics of Computation, 48(177). doi:10.1090/s0025-5718-1987-0866115-0 [2][4] Odlyzko, with a vast amount of high-precision computation, examines the distribution of spacings between the zeros of $\zeta$ and finds it in astonishing agreement with the prediction of the random-matrix Gaussian unitary ensemble, providing strong numerical support for the Montgomery-Dyson conjecture. This chapter uses it to show that this statistical fit, though not a proof, is extremely strong evidence leading mathematicians to believe RH.

← 6. The Agent Released8. The Organization Blind to Itself →