Myth and the Math

A mathematical problem is a quiet object. It has a statement, a set of known partial results, and a landscape of techniques that have been tried and have failed in instructive ways. Nothing about it demands narrative. But once a problem acquires a public story — a prize, a threat, a promise of revelation — the story begins to select which parts of the problem get attention.

This is not a claim that mathematicians are fooled by myths. It is a claim about allocation. Attention, funding, referee time, conference slots and public patience are finite. A myth changes the shape of what looks worth doing, and it does so without ever making a false mathematical statement. It works entirely by emphasis.

The Riemann Hypothesis is the clearest live case. It carries a million-dollar prize, a reputation as the deepest open problem in mathematics, and a widely repeated claim that its proof would break the internet. Two of those three are true. The third is not, and the cost of the third is measured in the hours spent correcting it.

The popular claim is that proving the Riemann Hypothesis would shatter internet security. The claim is mathematically empty, and it is worth stating exactly why rather than merely asserting that it is wrong. RSA rests on the difficulty of factoring a specific large semiprime. The Riemann Hypothesis concerns the global distribution of primes — how many there are below a bound, and how tightly the error term in that count is controlled. A sharper error term in a counting function does not hand anyone a factoring algorithm.

The situation is stronger still. Large parts of analytic number theory already operate on the assumption that the hypothesis is true. There is an entire literature of results proved 'under RH', and cryptographic parameter choices are frequently made using bounds that assume it. Proving the hypothesis would convert a body of conditional results into unconditional ones. It would not surprise the people who set key sizes.

Meanwhile the actual migration is already under way, and it has nothing to do with Riemann. Post-quantum standards built on lattice problems are being deployed because Shor's algorithm threatens factoring and discrete logarithms on a sufficiently large quantum computer. Lattice hardness is unaffected by anything the Riemann Hypothesis says about primes. The real threat model and the mythic one point in different directions.

The cost of the myth is asymmetric. Each restatement takes a sentence; each correction takes a paragraph and a citation. Over years, that asymmetry converts a technical non-issue into a permanent tax on the people best equipped to do the actual work.

A famous problem with a prize attached is a magnet for short proofs. The preprint servers carry a steady flow of claimed solutions to the Riemann Hypothesis, most of them a few pages long, most of them repeating errors that were catalogued decades ago. The volume is not a scandal; it is a predictable consequence of visibility plus reward.

The damage is not that anyone is fooled for long. It is that filtering is expensive and the burden falls on a small number of qualified readers. Every serious analytic number theorist has a policy for handling unsolicited proofs, and every such policy is a compromise between courtesy and time. The prize did not create the problem's depth; it created the queue.

There is a second-order effect worth naming. When a field becomes accustomed to dismissing short claims, it develops a reflex, and reflexes are indiscriminate. A genuinely unconventional approach — the kind that arrives from an adjacent field with unfamiliar notation — pays the same entry cost as the noise it resembles. Myth raises the noise floor, and a raised noise floor buries quiet signals first.

The mythic pull does not only operate on the popular side. Inside mathematical logic there is a recurring speculation that the Riemann Hypothesis might be independent of ZFC — true but unprovable in standard set theory, a Gödel sentence in a suit. The speculation is not incoherent. It is simply, in the practising view, unproductive: the working consensus on MathOverflow and in the literature is that entertaining it usually does not lead anywhere new.

Part of the reason is structural. RH has an arithmetic form — it is equivalent to a Π₁ statement about explicit bounds — so if it were independent of a sound theory, it would have to be true, since a counterexample would be a finite object that the theory could verify. Independence therefore does not offer the dramatic escape it appears to offer. It converts the problem, it does not dissolve it.

The deeper point is about what speculation costs. Undecidability talk has the texture of rigour — it uses real theorems and real vocabulary — while requiring none of the grinding technical work that actual progress demands. That combination is exactly what makes it seductive, and exactly what makes it a myth rather than a programme.

Real movement on this problem has almost never resembled the cinematic version. It has looked like translation: taking a statement about a Dirichlet series and restating it as a statement about matrices, about exponential sums, about spectra, about cancellation in a large sum where the cancellation must be extracted term by term. The work is sideways, not forward.

The recent record bears this out. Improvements to zero-density estimates — bounding how many zeros can lie off the critical line — have come from recasting Dirichlet polynomials in terms that expose hidden cancellation, and from what practitioners frankly call gambits: steps that violate the conventional wisdom about which inequality to apply where, justified only by what they make possible three lemmas later. Nothing about that process is dramatic while it is happening.

This matters for the myth question because the mythic frame demands a moment. It wants a proof, a date, a consequence. The actual epistemic object is a slow accumulation of reformulations, most of which will never be the final one but each of which narrows the space where the answer can hide. A culture trained to expect the moment will systematically undervalue the accumulation.

It is worth being precise about the mechanism. Myths of this kind do not run on ignorance; they run on fear. The internet-collapse story is compelling because it names a loss. The prize is compelling because it names a rescue. Undecidability is compelling because it offers permission to stop. Each version answers an anxiety, and each answer costs something in rigour.

This is the same mechanism documented in the sycophancy work, arriving from a different direction. A sycophantic system optimises for the answer that relieves discomfort rather than the answer that survives scrutiny. A mythologised problem does the same thing at the level of a whole field: the story that circulates is the one that satisfies, not the one that holds.

The defence is identical in both cases, and it is unglamorous. Name the claim. State what would falsify it. Separate what is established from what is licensed and from what is merely asserted. Keep the ledger visible. Fear distorts logic for a while — but only for as long as nobody is writing down which parts of the argument are actually load-bearing.

The demythologising described above is not only a hygiene measure. It is the precondition for a specific working method: a human and a machine reasoning in parallel on the same problem, each doing the part the other does poorly. Myth is what breaks that method, because a machine trained on public text has absorbed the myth at full strength, and a human under time pressure will accept the fluent version of it.

The division of labour is not mysterious. A language model is fast at breadth — recalling which reformulations have been tried, restating a statement in the vocabulary of an adjacent field, generating the twenty translations of a problem that a single researcher would take a month to enumerate. It is unreliable at knowing which of those twenty is load-bearing. A human working the problem has the opposite profile: slow at breadth, but carrying the tacit sense of which step is cheap and which is a genuine obstruction. Neither profile alone solves a hard problem. Run in parallel, with the human adjudicating and the machine enumerating, the search space gets covered differently than either would cover it.

This is exactly the 'sideways translation' of section 05, mechanised on one side and supervised on the other. Progress on a resistant problem comes from restating it until the restatement exposes cancellation that the original form hid. Enumeration of candidate restatements is a task with no prestige and enormous volume — precisely the task a machine should hold. Judging which restatement is worth three months of technical work is a task requiring accountability for being wrong — precisely the task a machine should not hold.

The failure mode is sycophancy, and it is the same failure documented across this body of work. A model that optimises for agreement will confirm the human's preferred direction, and the parallel process collapses into an echo with two voices. What prevents the collapse is the ledger discipline: every claim marked established, licensed or asserted, every proposed route accompanied by what would rule it out. Under that discipline the machine's fluency becomes an asset rather than a hazard, because fluency without status-marking is exactly how a myth propagates.

None of this is a claim that a machine has solved, or will soon solve, a Millennium problem. The claim is narrower and more defensible: that problems which resist a single mind can yield to a division of cognitive labour, and that the division only works when the myth has been cleared out of the shared workspace first. Fear distorts logic for a while. A second reasoner, held to an explicit ledger, is one of the few things that shortens the while.