teal-sea / zeta-labstate of record · compiled 14 Aug 2026 · revision 9ebdea0 · source

Library · docs/12-how-hard-problems-die.md

How Hard Problems Die: A Kill Board

3,583 words · 357 lines · source

docs/08-why-it-is-hard.md catalogued what fails. docs/09-new-ontologies.md described the one time a rebuild worked. This document widens the sample: eight famous problems, the mechanism that actually killed each one, and an honest scoring of which mechanisms RH's live formulations touch.

The short version

Hard problems do not die of effort. They die of a mechanism — a structural move converting the statement into something a known tool can bite. There are not many mechanisms. Seven case studies below: rebuild the ontology (Weil conjectures), bridge to a richer world (Fermat), reduce infinity to a finite certified list (Four Colour, Kepler), squeeze an effective threshold down to meet a computation coming up (weak Goldbach), deform the object by a flow and control the singularities (Poincaré), let insight make the computation irrelevant (Catalan), build an equivalence web wide enough that an outsider's toolkit lands on it (Kadison–Singer). Scoring RH: it touches flow (de Bruijn–Newman, docs/05), equivalence web (docs/07, zeta/criteria.py) and real-rootedness (Li/Jensen, zeta/li.py) — and provably does not fit finite reduction or the effective-threshold squeeze, which is the structural reason verification to any height decides nothing (docs/08 §3). Labels are THEOREM / CONJECTURE / PROGRAM; where I could not check a date or attribution I say so.

Two cautions. This is a survivorship sample — we tabulate the problems that died, not the ones still alive. And in every case the closing input came from outside the problem's own language: the board's most reliable regularity, and bad news for anyone hoping RH falls to a better estimate in the language of docs/01–docs/04.

1. ONTOLOGY REBUILD — the Weil conjectures

Mechanism: when the objects named in the statement cannot carry an argument, rebuild what the objects are until they can.

Count solutions of a polynomial equation over a finite field. For E : y² = x³ − x at p = 1009 — that is zeta.finitefield.hasse_check(-1, 0, 1009), the module taking y² = x³ + ax + b — you get N_1 = 1040, a_p = p + 1 − N = −30, against the Hasse bound 2√p = 63.5295…; re-run while writing this. Package the counts over all extensions into a zeta function and the √p is the 1/2 of Re s = 1/2 — literally: critical_line_check(-1, 0, 1009) returns re_s = [0.5, 0.5] with max_deviation_from_half = 0.0.

THEOREM (Hasse, 1930s, elliptic curves; Weil, published 1948, all curves). The bound holds. Weil then CONJECTURED (1949) the analogue in every dimension.

Twenty-five years followed. Dwork proved rationality in 1960 by p-adic analysis — a technique with no home; it did not touch the RH part. What worked was Grothendieck's rebuild (roughly 1958–1970): schemes, so a scatter of residues is a curve and Spec Z is a space; topoi, so "place" means "category of sheaves obeying axioms"; and étale cohomology (Grothendieck with M. Artin), attaching to the variety honest finite-dimensional vector spaces on which Frobenius x ↦ x^p acts as an honest matrix. Lefschetz's fixed-point formula converts counting into traces:

    N_{q^n}  =  Σ_i  (−1)^i · Tr( Frob^n | H^i_et )

Read that as modal analysis (docs/10 §1): a messy time-domain record becomes a finite sum over eigenvalues. Rationality and the functional equation fall out immediately, and RH becomes "those eigenvalues have modulus exactly q^{i/2}" — pole placement. You cannot place a pole you have no coordinates for; the rebuild supplied the coordinates.

THEOREM (Deligne, La conjecture de Weil. I, Publ. Math. IHÉS 43 (1974), 273–307). The eigenvalue bound. The closing input is positivity: amplification over even tensor powers — Rankin's trick from modular forms, at bottom "squares of real numbers are non-negative" — plus monodromy from Lefschetz pencils (sweep the variety by a one-parameter family of hyperplane slices and track how the cohomology twists as you go round the family; the twisting group has to be big). (Confident in that shape and the citation, not in reciting the chain of lemmas; Kowalski's and Tony Feng's notes on Weil I are the standard entry points.) Grothendieck's own route, the standard conjectures on positivity for algebraic cycles, remains CONJECTURE; Deligne went around it.

Plain-words recap. Nobody found a cleverer estimate. They rebuilt geometry until Frobenius was a matrix, at which point "all zeros on the line" became "this matrix has eigenvalues of a fixed size," and positivity finished it. Symmetry never sufficed; symmetry plus positivity did.

2. BRIDGE BETWEEN WORLDS — Fermat's Last Theorem

Mechanism: transport the problem into a world with far more structure, via a chain that makes a soft-looking conjecture in the second world lethal to the first.

A hypothetical solution a^p + b^p = c^p is one equation with no visible structure. Frey (1984–86) attached to it the elliptic curve y² = x(x − a^p)(x + b^p) — the Frey curve — whose discriminant is so implausibly a perfect p-th power that the curve looks impossible. Serre's ε-conjecture (level lowering) made "looks impossible" precise.

THEOREM (Ribet; proved summer 1986, published Invent. Math. 100 (1990), 431–476). The ε-conjecture. Consequence: the Frey curve cannot be modular — modular meaning its L-function is that of a modular form, an object from complex analysis, not Diophantine algebra. So FLT reduced to CONJECTURE (Taniyama, 1955; Shimura; Weil, 1967): every elliptic curve over Q is modular.

THEOREM (Wiles, Ann. of Math. 141 (1995), 443–551; with Taylor, ibid. 553–572). Modularity for semistable elliptic curves — those whose reduction mod every prime degenerates only in the mildest allowed way — which is enough for FLT, because the Frey curve is one. The history matters for this repo's temperament: Wiles announced in Cambridge in June 1993; a gap was found during refereeing (Nick Katz located it) later that year; it was closed with Richard Taylor in September 1994. Peer review worked exactly as advertised. Full modularity is THEOREM (Breuil–Conrad–Diamond–Taylor, J. Amer. Math. Soc. 14 (2001), 843–939).

The machinery invented on the way is now the subject's standard toolkit: Galois representations (a curve's arithmetic as a matrix-valued symmetry action), Mazur's deformation theory (all lifts of a mod-p representation, parametrised by a universal ring R), and the R = T theorem — that ring equals a Hecke algebra built from modular forms — via Taylor–Wiles patching. An impedance match: two circuits that could not talk, joined by a transformer that was itself the hard part.

A connection this repo owns: modular forms are the same family as our theta function. docs/02 §6 identifies Jacobi's Θ(τ) = Σ q^{n²} as a modular form of weight 1/2 for Γ₀(4), and the √s in Jacobi's identity — hence the 1/2 in Re s = 1/2 — is that weight. Elliptic curves land at weight 2: same rulebook, different weight.

3. FINITE REDUCTION + MACHINE GRIND — Four Colour and Kepler

Mechanism: prove a reduction theorem converting "for all of an infinite set" into "for all of this explicit finite list," then check the list mechanically.

The mathematics is entirely in the reduction; the grind is bookkeeping, at a scale needing its own correctness argument.

Four Colour. Kempe's 1879 proof stood eleven years before Heawood found the error in 1890 — worth remembering. THEOREM (Appel–Haken, announced 1976, published Illinois J. Math. 21 (1977), in two parts, the reducibility half with Koch). Their reduction produced an unavoidable set of 1936 reducible configurations: unavoidable means every planar map contains one, reducible means a map containing one shrinks to a smaller counterexample — so a minimal counterexample cannot exist if every configuration checks out, which is what the computer did. Cut to 633 configurations by Robertson–Sanders–Seymour–Thomas (1997) and formalised end-to-end in Coq by Georges Gonthier (2005, with Benjamin Werner) — reduction and checking, machine-verified.

Kepler (no packing of equal spheres beats face-centred cubic, π/√18 ≈ 74.05%). THEOREM (Hales, announced 1998; published Ann. of Math. 162 (2005), 1065–1185). The reduction turns the infinite packing problem into a finite optimisation over local configurations, closed with interval arithmetic and on the order of a hundred thousand linear programs. The referees famously reported roughly 99% confidence and declined to certify the computational part — this mechanism's honest scar. Hales answered with Flyspeck, completed August 2014 (HOL Light and Isabelle; Forum of Mathematics, Pi 5 (2017)).

The repo-relevant point: interval arithmetic is what separates a computation from a theorem. docs/08 §3.1 draws the same line between Platt–Trudgian and earlier zero-verification announcements, and zeta/rigor.py sits on that seam.

4. ANALYTIC-COMPUTATIONAL SQUEEZE — weak (ternary) Goldbach

Mechanism: prove "true for all n > N₀" with N₀ effective, then drag N₀ down while certified computation climbs up, until the two ranges overlap.

THEOREM (Vinogradov, 1937). Every sufficiently large odd integer is a sum of three primes. The original statement was asymptotic; Borozdkin later extracted an explicit threshold around 3^(3^15), roughly seven million digits (widely quoted, usually dated 1956; check the source). Liu and Wang (2002) got it to about e^3100 ≈ 10^1346 (conversion confirmed: 3100/log 10 = 1346.3). Hopeless.

THEOREM (Helfgott, arXiv:1312.7748, December 2013). Threshold 10^27. Combined with Helfgott–Platt, "Numerical verification of the ternary Goldbach conjecture up to 8.875·10^30" (Experimental Mathematics 22 (2013), 406–409), the ranges overlap and the conjecture is closed. (Accepted by the community; my understanding is that the complete proof has circulated mainly as arXiv preprints and a book manuscript rather than one refereed journal article — verify the current publication status before citing it as such.)

Two things to take. The margin: 10^27 against 8.875·10^30 is nearly four orders of slack — squeezes close because somebody buys overlap, not a knife-edge. And the ingredient: Helfgott's major-arc analysis leans on rigorous interval-arithmetic verification of zeros of Dirichlet L-functions — the same species of certified data zeta/rigor.py produces for ζ itself, one family over. The cleanest existing case of the unglamorous work in docs/08 §6 item 1 becoming a theorem: not evidence, an input.

5. FLOW + SINGULARITY CONTROL — the Poincaré conjecture

Mechanism: do not attack the object. Deform it by a flow that provably simplifies it, and do the real work at the places where the flow blows up.

PROGRAM → THEOREM. Hamilton (J. Differential Geom. 17 (1982), 255–306) introduced Ricci flow, ∂g/∂t = −2·Ric(g): a heat equation for the metric itself. Curvature diffuses, bumps flatten, and a manifold left alone wants to become one of a short list of round model geometries. Hamilton's programme was to run the flow and read off the classification — with one obstruction everybody could name: the flow develops singularities in finite time (necks pinch off), and the argument dies.

THEOREM (Perelman, three arXiv preprints: math/0211159, November 2002; math/0303109, March 2003; math/0307245, July 2003). Singularity control, via a monotone entropy functional (a Lyapunov function — a quantity that can only move one way, so the flow cannot cycle or hide), a no-local-collapsing theorem, a classification of singularities up close, and surgery: cut at the pinch, cap it, restart, with uniform bounds guaranteeing you do this only finitely often on any bounded time interval. Independent expositions followed (Kleiner–Lott; Cao–Zhu; Morgan–Tian).

Emphasise this, because the repo has a stake in it: a heat-type flow closed a Millennium problem. docs/05-de-bruijn-newman.md runs a heat flow on Ξ — the family H_t with H_0 ∝ Ξ — and the shape is identical: the flow smooths (for t large enough all zeros are real, de Bruijn 1950), the obstruction is singularities (zeros colliding and leaving the real axis, docs/05 §4), and everything turns on the critical time.

The asymmetry matters too. Ricci flow ran forward, with time to spare. RH asks about H_t at exactly t = 0, and by THEOREM (Rodgers–Tao, 2018/2020), Λ ≥ 0, so Λ = 0 is RH: no smoothing budget at all. docs/07 §11's blunt corollary — whatever proves RH must be exactly tight, which rules out every argument with room to spare.

6. INSIGHT LAPPING THE COMPUTERS — Catalan's conjecture

Mechanism: the honest counterweight to §3 and §4 — a squeeze can be fully set up, grinding, and still be the wrong strategy, because structural insight can make the computation irrelevant.

CONJECTURE (Catalan, 1844 — the date usually given for his note to Crelle; I have also seen 1842 cited). The only consecutive perfect powers are 8 and 9: the only solution of x^p − y^q = 1 in integers > 1 is 3² − 2³ = 1 (yes, 9 − 8 = 1).

THEOREM (Tijdeman, 1976). Solutions are effectively bounded — precisely a §3-style finite reduction, and celebrated as such. The bound was astronomical, so a twenty-six-year §4-style squeeze began: narrow the exponent range analytically from above (Mignotte's bounds, commonly quoted as p < 7.15·10^11, q < 7.78·10^16 — widely cited figures, worth checking at source) while computation pushed up from below. The gap stayed many orders wide.

THEOREM (Mihăilescu, announced April 2002; "Primary cyclotomic units and a proof of Catalan's conjecture", J. reine angew. Math. 572 (2004), 167–195). Closed with the algebra of cyclotomic fields — Galois modules over group rings, Stickelberger, Thaine — and no computation at all.

Twenty-six years of the mainstream strategy were bypassed. Keep this next to §3 and §4 as a permanent corrective: "the computers are closing in" describes one strategy's progress, not the problem's.

7. EQUIVALENCE WEB + OUTSIDER TOOLKIT — Kadison–Singer

Mechanism: a problem restated a dozen times across unrelated fields becomes searchable by all of them; the killing tool arrives from a field that never cared about the original question.

Kadison–Singer (1959, Amer. J. Math. 81): does every pure state on the diagonal subalgebra of bounded operators on ℓ² extend uniquely to the whole algebra? (A state is a normalised positive linear functional — an abstract "expected value" assigning a number to each observable; pure means it is not an average of two different ones, i.e. it carries no residual uncertainty of its own. Roots in Dirac: do commuting observables determine a state?) Fifty-four years open, acquiring equivalent formulations in operator algebras (Anderson's paving conjecture, 1979), Banach space theory (Bourgain–Tzafriri restricted invertibility), signal processing (the Feichtinger conjecture on frames) and discrepancy theory — Weaver's KS_r (Discrete Math. 278 (2004)), about partitioning vectors so no part is spectrally heavy.

THEOREM (Marcus–Spielman–Srivastava, arXiv:1306.3969, June 2013; Ann. of Math. 182 (2015), 327–350). Weaver's KS₂, hence Kadison–Singer. The method: interlacing families of polynomials and mixed characteristic polynomials, top eigenvalue controlled by a barrier argument resting on real stability (Borcea–Brändén). Stripped to the mechanism: they bounded eigenvalues by proving a polynomial had all real roots and then bounding the largest one. All three authors came from theoretical computer science; the machinery had just been built in the companion paper (Interlacing Families I) to construct Ramanujan graphs, which nobody in operator algebras was watching.

The RH parallel is exact in form. docs/07 catalogues RH's own web — error terms, Mertens, Riesz, Li, Nyman–Beurling, Robin/Lagarias, Weil positivity, Speiser, de Bruijn–Newman — and zeta/criteria.py makes several computable; RH also has a real-rootedness face, the subject of zeta/li.py. But read docs/07 §11 before celebrating: abundance of equivalences is evidence RH is true and deep, and weak evidence it is nearly proved. Kadison–Singer's web paid after fifty-four years, and only because one restatement happened to be in the native language of a tool invented for something else.

8. Scoring RH against the board

8.1 The mechanisms RH's live formulations touch

Flow (§5) — yes, and a real research frontier. docs/05 and zeta/heatflow.py: H_t, the de Bruijn–Newman constant, RH ⟺ Λ = 0 with Λ ∈ [0, 0.2] pinned from both sides (Rodgers–Tao below, Platt–Trudgian above — docs/05 §3). By docs/07 §11's productivity filter — an equivalence pays only if it yields something without proving RH — this is one of only three items in that catalogue that has paid, because Rodgers–Tao's Λ ≥ 0 is a genuine new theorem. The missing piece is docs/09 §7's: our flow acts on a function, not a space. Ricci flow deforms a geometry; there is no geometry under H_t.

Equivalence web (§7) — yes, extensively. docs/07 plus zeta/criteria.py, with §11's standing question against each new entry: where did the hard part go? There is always an answer.

Real-rootedness (§7's actual weapon) — yes, and computable. The closest structural contact on the board, worth pinning with numbers.

THEOREM (Pólya, 1927). For an entire function of the appropriate growth class, having only real zeros is equivalent to hyperbolicity (all roots real) of every associated Jensen polynomial. Applied to Ξ, this makes RH exactly a real-rootedness statement.

Derived, not remembered, in the repo's own normalisation, so nothing depends on the literature's conventions. zeta.heatflow.Phi satisfies H₀(z) = ∫₀^∞ Φ(u) cos(zu) du = Ξ(z/2)/8; expanding the cosine with moments b_n = ∫₀^∞ Φ(u) u^{2n} du gives H₀(z) = g(−z²), g(x) = Σ (b_n/(2n)!) x^n, so RH ⟺ g has only real negative zeros. Writing g(x) = Σ γ(n) x^n/n! fixes γ(n) = n!·b_n/(2n)! and J^{d,n}(X) = Σ_{j≤d} C(d,j)·γ(n+j)·X^j; degree 2 is hyperbolic exactly when the Turán inequality γ(n+1)² ≥ γ(n)·γ(n+2) holds. (zeta/li.py computes γ(n) in the Griffin–Ono–Rolen–Zagier normalisation, 8·ξ(½+z) = Σ γ(n) z^{2n}/n!, which is this one times 64·4ⁿ — checked, not assumed. A positive factor per n rescales X and the polynomial as a whole, so hyperbolicity and the ratios below are the same in either convention.) Measured at dps = 40, the ratio γ(n+1)²/(γ(n)γ(n+2)) for n = 0…7:

    1.0748  1.0627  1.0543  1.0481  1.0432  1.0393  1.0362  1.0335

Above 1 — which is exactly the Turán inequality holding — and shrinking toward 1. J^{d,n} for d = 3,4,5 at n = 0,2,4 also came out hyperbolic, largest |Im(root)| at machine-zero (measured 2.3e-132 at dps = 30, and the exact Sturm count in ℚ[X] agrees in every case). The known theory: THEOREM (Csordas–Norfolk–Varga, Trans. AMS 296 (1986), 521–541) — the Turán inequalities (d = 2) hold unconditionally; THEOREM (Dimitrov–Lucas, Proc. AMS 139 (2011), 1013–1022) — d = 3; THEOREM (Griffin–Ono–Rolen–Zagier, PNAS 116 (2019), 11103–11110) — for each degree d, J^{d,n} is hyperbolic for all sufficiently large n, which leaves a finite unresolved set for every d, and RH needs all of them. Same shape as "a positive proportion of zeros on the line" (docs/08 §1.3): a real theorem on a different axis.

Note also what Marcus–Spielman–Srivastava had that we do not: a finite, explicitly constructed polynomial family. RH's Jensen family is doubly infinite, and by docs/07 §4 its coefficients are as hard as the thing you wanted — computing them honestly needs the zeros. zeta/li.py computes them anyway, because watching the margin shrink is the point.

8.2 The mechanisms RH does not fit — stated plainly

Finite reduction (§3) — no. That mechanism needs a reduction theorem: a proof that the infinite claim follows from a finite explicit list. RH has none. There is no theorem "every zero with |γ| > T lies on the line" for any T, and nobody expects one — the zero-free regions of docs/08 §1.1 all shrink with height, the opposite of what a reduction needs. Without one there is nothing to grind, however large the machine.

Effective-threshold squeeze (§4) — no, for that reason plus a second. Vinogradov's theorem has the form "true above N₀," so lowering N₀ is measurable progress; RH has no N₀-shaped statement to lower, hence no front coming down to meet the computation. And the computation is barely moving: by docs/08 §3.2 the fluctuation carrying the risk is S(T) = (1/π)·arg ζ(1/2 + iT), of typical size √(log log T / 2π²) — about 0.415 at T = 10^13, 0.446 at T = 10^22; doubling it needs a height with roughly 350,000 decimal digits. Weak Goldbach closed with four orders of margin; here the deficit is unbounded. So verification to any height decides nothing — docs/08 §3, now with a structural reason rather than an assertion, and with Littlewood and Mertens (§3.3–3.4 there) as proof that unanimous numerics in this subject have in fact been wrong.

One qualification, cutting the other way: certified verification height is useless as evidence and valuable as an input. Feeding Platt–Trudgian's 3·10^12 back into the same machinery is what let them sharpen Polymath15's Λ ≤ 0.22 to Λ ≤ 0.2 — their own result, not the collaboration's (docs/05 §3) — and Platt's certified L-function zeros are an ingredient in Helfgott's theorem (§4). That is the whole justification for zeta/rigor.py: not a squeeze on RH, a supplier to other people's theorems.

Bridge to a richer world (§2) — partial at best. The Selberg class and Langlands organise ζ into a family (docs/09 §3, §6), a genuine bridge — but no formulation of functoriality is known to imply RH, and there is no Ribet.

Ontology rebuild (§1) — yes, the field's own leading bet. docs/09, docs/10, docs/11: fifty-plus years with the blueprint in hand, and still no space, no operator with a home, no positivity. Insight lapping the computers (§6) — the null case: there is nothing to lap.

Where to go next