Plectis

Development chronology

Development index

Rendered from docs/reference/WAVE_INDEX.md in the public Lean repository. View source

All sections

Wave index

Wave labels are development chronology, not Lean import order.

Read order for humans:

  1. docs/orientation.json or docs/ORIENTATION.md — select one mathematical programme, claim, or exact remaining open proposition.
  2. docs/SOURCE_MAP.md — follow that selected intention into a bounded module route.
  3. Current papers — read the selected problem’s short paper, longer research record, and source handles.
  4. this file — recover development chronology only when chronology is the question.
  5. Erdos249257.lean and ErdosProblems.lean — inspect package topology only when import structure is the question.

For an external statement-identity handoff, use the generated Formal Conjectures crosswalk. It binds each of the eight problem rows to its pinned upstream declaration, source hash, and local problem route. For a module or paper handle, use python3 scripts/query_corpus.py --module <module_path_or_sigil> or --paper-anchor <TeX_label_or_source_ref> to recover the corresponding proof/source context before returning to the problem’s exact boundary.

Lean source checked by the pinned Lean kernel is proof authority. Wave order does not strengthen a claim or close an exact open proposition. Erdős #249 and the universal form of #257 remain open.

Chronology map

Band Where it lives Notes
Pre-wave / early kernel mass Erdos249257/CertificateKernel.lean Core #257 support-family and certificate machinery assembled before the separately named #249 modules
Waves 8–16 mainly CertificateKernel.lean Earlier kernel layers retained inside the assembled microkernel
Waves 17–25 separately named modules under Erdos249257/ Reader-facing ladder, detailed below
Wave 26–27 public material inside CertificateKernel.lean Not a separate top-level module
Wave 30 referenced in LcmConeNonflat.lean Small-prime support of the cone-window argument
Generated certificates GeneratedCertificates.lean + GeneratedCertificates/* Generated finite certificate layers, not waves
Carry trunk five post-wave modules under Erdos249257/ Not waves; a shared binary-carry layer over the kernel, see below
Diagonal pincer frontier exact decomposition modules, algebraic support, and finite prime-certificate shards Post-wave #249 work; sharpens the open certificate-supply obligation without closing it

Waves 17–25 (named modules)

Each module is a self-contained step with a header docstring stating its own honest boundary. The import chain is GapFareyBound → MersenneLambertLadder → … → LcmConeNonflat → CertificateKernel ← GeneratedCertificates.

Wave Module
17 GapFareyBound.lean
18 MersenneLambertLadder.lean
19 GeometricCoprimality.lean
20 GcdMomentCalculus.lean
21 TotientTailPeriodKiller.lean
22 CarrySurvivorExtinction.lean
23 LcmDiagonalReduction.lean
24 LcmConeFlatness.lean
25 LcmConeNonflat.lean

Wave 17 — GapFareyBound.lean

Farey-gap denominator bounds. A mediant/Farey argument places any rational m/q inside a forbidden gap, so no q below the bound can represent S. Elementary and per-window; clears q≤2.49×1017 (K=120) and q≤7.96×1034 (K=240).

Wave 18 — MersenneLambertLadder.lean

The Mersenne–Lambert ladder, machine-checked: the rational rungs L(μ)=1/2, L(φ)=2, the positive lift L(A)=S, and the Möbius-square lens S=1/2+∑μ(d)/(2d−1)2. Engine: a signed, linear-growth weighted Lambert rearrangement.

Wave 19 — GeometricCoprimality.lean

S as coprime-pair mass. #{(a,b):a+b=n,0<a,gcd(a,b)=1}=φ(n), so at r=1/2, S−1/2=P(gcd(X,Y)=1) for independent fair-coin waiting times. Base 2 is the unique self-normalising point of the geometric law.

Wave 20 — GcdMomentCalculus.lean

The squared transform L2(f)=∑f(d)/(2d−1)2=E[(f∗ζ)(gcd(X,Y))], since gcd-divisibility factorises across independent coordinates. Yields L2(μ)=S−1/2, the gcd-moment ladder, and Pillai’s gcd-sum function. The exact family return is the probabilistic gcd-geometry route: the totient gcd-moment declaration and the Stern–Brocot cylinder recursion, cylinder remainder bound, and run-stability declaration return to the paper’s probability-coordinate appendix, with the stable records res:directionnormalization, res:sternbrocotcylinders, and res:sternbrocotruns. The cylinder-law family is also directly recoverable through python3 scripts/query_corpus.py --claim stern_brocot_cylinder_law, which returns the probabilistic_gcd_geometry route and its source-fingerprinted resume handoff. These exact identities and combinatorial bounds do not close the open #249 irrationality or unbounded-certificate obligations.

Wave 21 — TotientTailPeriodKiller.lean

Period, not digits. Rationality of S forces the tail-period law RN+h−RN∈Z; missing that integer by a decidable margin is a finite “kill”. Reduction:

irrational_totient_series_of_certificate_supply

Wave 22 — CarrySurvivorExtinction.lean

Multiple-period collapse. Every multiple of a period is a period, so the obligation collapses onto the one-parameter family periodLcm t = lcm(1..t). Adds the carry-survivor orbit certificate (a bounded integer orbit that must escape a narrow strip).

Wave 23 — LcmDiagonalReduction.lean

Diagonal collapse. Standing on the ray (N = periodLcm t) removes the second parameter: #249 follows from one ℕ-indexed decidable sequence ∃ L, certifiedKill (periodLcm t) (periodLcm t) L holding infinitely often.

Wave 24 — LcmConeFlatness.lean

Cone-flatness law. Rationality forces one fractional constant on the whole lcm cone {k·periodLcm t}. Certificate completeness: a kill exists iff the tail difference is a non-integer:

exists_certifiedKill_iff_tail_diff_notMem_int

Plus rank-2 second-difference certificates.

Wave 25 — LcmConeNonflat.lean

Cone non-flatness refuter. Interrogates a whole menu of cone vertices at once: if their one-sided arcs shared a common fractional part, the minimal-deep-tail vertex would be a common endpoint, and its certificate row denies exactly that. Sharper than pairwise; genuinely joint for menus of size ≥ 3.

Carry trunk (post-wave modules)

These five modules are not waves. They form a shared binary-carry layer added after the chronology above, building on CertificateKernel: what rationality of a binary series forces on its integer carry states. Reading order: GenericTailOrbitRigidity → GreedyAchievementSet → BooleanMobiusCarry → RationalSupportCarrySkeleton → SublogDivisorCoverage.

GenericTailOrbitRigidity

Depends on CertificateKernel.

For coefficients c(n)≤n, the binary series ∑c(n)/2n is rational exactly when a tempered integer carry orbit exists (u(N+1)=2⋅u(N)−v⋅c(N+1) with u(N)/2N→0); every such orbit is rigid, equal to the scaled analytic tail v⋅Tc(N). Balanced pulses also prove unbounded exact successor fan-out, ruling out a generic autonomous finite-state decoder.

Main theorems:

temperedBinaryOrbit_eq_scaledTail
binaryCoeffSeries_rational_iff_exists_temperedBinaryOrbit
balancedPulse_no_autonomous_decoder

GreedyAchievementSet

Depends on CertificateKernel.

Greedy geometry for the Mersenne achievement set (values ∑n∈A1/(2n−1)): strict superincreasing tail inequalities, the quantitative gap asymptotic (2/3)⋅4−n+O(8−n), compact/perfect/totally-disconnected/nowhere-dense structure with Lebesgue measure one, exact real and rational greedy recurrences, membership ⇔ all-level greedy survival, sound one-sided finite rational death certificates (the exact level-one certificate excludes 3/4), and uniqueness of normalised support coding. Certificates prove nonmembership only.

Main theorems:

volume_mersenneAchievementSet
isNowhereDense_mersenneAchievementSet
mem_mersenneAchievementSet_iff_greedy_survival
certifiedGreedyMersenneDeath_not_mem

BooleanMobiusCarry

Depends on GenericTailOrbitRigidity.

Boolean–Möbius carry coordinates for support series: the Lambert coefficient fA(n)=#{a∈A:a∣n} satisfies fA=1A∗ζ and μ∗fA=1A on positive integers; rationality of the support series is equivalent to a tempered carry orbit whose carry quotient is exactly fA; normalised nonempty supports with value p/q correspond exactly to quotient-only Boolean Möbius carry certificates. The displayed value 1/2 has a canonical affine-orbit criterion. Worked support {2,3}: value 10/21, period-six orbit 10,20,19,17,13,26.

Main theorems:

erdosSupportSeries_rational_iff_exists_temperedCarry
exists_normalized_support_fraction_iff_exists_booleanMobiusCarry
support_half_iff_affineBinaryOrbit_tempered

RationalSupportCarrySkeleton

Depends on BooleanMobiusCarry.

Residue wraps and reciprocal mass: the binary repetend identity (least positive residues in a complete doubling cycle sum to odd modulus × number of wraps), an algebraic one-wrap classification, the Cesàro identification of mean support tails with the reciprocal mass ∑a∈A1/a, the exact excess-mean identity, the rationality-forced lower bound 1/ordv⁡(2) on reciprocal mass, the dyadic strengthening (mass divergent or >1 for infinite dyadic-rational supports), and global unboundedness of the positive carry state attached to any infinite support with rational value.

Main theorems:

sum_doublingResidue_eq_mul_wrapCount
one_div_oddOrder_le_reciprocalMass_of_support_fraction
dyadic_support_fraction_reciprocalMass_diverges_or_gt_one
exists_unbounded_shifted_odd_tail_nat_state_of_support_fraction

SublogDivisorCoverage

Depends on RationalSupportCarrySkeleton.

Sublogarithmic divisor coverage. If an Erdős support series has a rational value, consecutive zero windows in its divisor-count coefficient fA have length at most εlog2⁡(N+1)+B for every ε>0, with B≥0 obtained for that support and numerator. The proof builds B from ε, c and v, but the support and numerator are bound before the existential, so this declaration does not export one constant uniform across all supports. The proof composes an explicit fixed-power divisor bound (τ(n)k≤(k2k)kn), a binary-tail estimate, and the exact carry recurrence. This constrains support coverage; it does not solve universal #257.

Main theorem:

supportCoeffZeroWindow_length_le_eps_logb_add

None of these modules claims a solution of Erdős #249 or #257. The contribution boundary is theorem-family-specific: the carry recurrence and strict-tail geometry have direct prior art (Han Wang; Kovač–Tao); Möbius inversion, repetend algebra, and divisor averaging are classical; the converse/rigidity, certificate-normal-form, and coupled reciprocal-mass families remain exact-source-comparison candidates. No priority claim is made.

SupportSunflowerDichotomy is an adjacent conditional #257 route. It gives the exact carry conversion for finite-core orthogonal-petal bouquets and records an explicit alternating-core squared-prime-petal support. Its uniform tail-selection hypothesis is not proved, so it adds no unconditional support family.

CompositeDilationDefect supplies the exact finite correction when a composite support element dilates the divisor-count coefficient, and bounds that correction for orthogonal-petal bouquets. It is a local identity, not a correlation or irrationality theorem.

MaximalOmegaLayer extends the one-prime-power identity to two distinct prime-power layers. It records finite coefficient algebra only, with no bounded-Ω or irrationality endpoint.

Diagonal pincer frontier (post-wave #249 band)

This band refines the lcm-diagonal reduction. Its main reading chain is DiagonalPincerDecomposition → SquaredMersenneDiagonalEnclosure → DiagonalFreshLossBridge. The first module makes diagonal integrality exactly equivalent to a foreign-defect term hitting a full-target interval. The second centres the diagonal expression at a Lambert projection and bounds the signed squared-Mersenne tail. The third turns the remaining fresh loss into an exact integer projection and then into binary suffix conditions. Every irrationality theorem in this band retains an explicit unbounded-supply hypothesis.

DiagonalPincerCertificates and the DiagonalPincerCertificatesT* aggregators assemble 28 checked scales through t=64; the many DiagonalPincerPrimeCertificates.* files are finite proof shards, not separate mathematical claims. The denominator/cyclotomic chain RadicalMobiusShadow → RepunitMobiusNumerator → CyclicTensorMobiusShadow → CyclotomicProjectionOfShadow → PrimePowerJumpDynamics → MersenneShadowCyclotomicNoncollapse → MersenneShadowDenominatorGrowth supplies exact rational-denominator information used by the frontier. The remaining named modules provide finite identities or conditional interfaces; their headers state the boundary locally.

Transport and curvature companion band

This post-wave #249 band forms a second, self-contained reading chain: CurvatureCarry → ExponentOnlyTransport → JointExponentTransport, with PrimeJumpMigration → PrimeJumpWindow, ThreeTransportBoundary, FirstHarmonicGap, TropicalCurvatureCarry, and LcmFactorIdealPulseObstruction as adjacent branches. It proves exact dyadic window decompositions, affine old-channel annihilation, new-channel migration, two concrete finite certificates, and two scoped no-go theorems: fixed local precision and finite LCM factor/anchor shift algebra. The latter is a synthetic countermodel rather than a model of actual totient differences. It does not prove an unbounded certificate supply or a first-harmonic estimate, so #249 remains open. The former one-sided companion is retired; its checked declarations remain queryable through the claim registry and source map.

The finite factor-ideal result is one claim inside the reviewed transport_strategy_no_go_boundaries family, alongside the fixed-precision obstruction. Retrieve that family with python3 scripts/query_corpus.py --publication-family transport_strategy_no_go_boundaries, or inspect the factor-ideal claim with python3 scripts/query_corpus.py --claim lcm_factor_ideal_anchor_pulse_no_go. Its exact formal anchors are lcm_factorIdeal_finiteRank_shiftAlgebra_not_sufficient, lcm_factorIdeal_sparseAnchor_not_sufficient, and lcmAnchorPulse_t3_letters. The family remains a synthetic survivor: its forcing letters need not be actual totient differences, nonlinear fresh-divisor arguments remain outside the result, and no unbounded certificate supply follows.

Source-current #1041 frontier (not a wave)

The public #1041 research corpus is a dated source projection, not a new Lean wave and not a reviewed claim-registry surface. Its source checkpoint is 6658deca35adde05f60bd2a19c76da996698bc9a. Read research_corpus/Erdos1041/FRONTIER.md first: it supersedes stale activation rows by naming the 2026-08-29 refutations, the surviving carriers, and the exact open gaps. Then consult STRONGEST_RESULTS.json for the machine-readable premise/consumer map and CORPUS_MANIFEST.json for the published-file and digest inventory.

For a machine-first handoff, run python3 scripts/query_corpus.py --route erdos_1041. The route checks the four public frontier files against their indexed digests and returns the 35-result/open envelope; for a resumable, source-fingerprinted packet use python3 scripts/query_route_memory.py --problem 1041. Neither route enters the reviewed claim registry, Comparator, or Lean proof authority.

Use this compact reading order:

  1. AttachmentAwareReeb.md and GenericSufficiencyClosure.md for the surviving hub-selection carrier and its generic-to-closed extension.
  2. NearFeketeRadialAngularSplit.md and Degree5AssemblyAndSharpenedCuts.md for the new near-regular and degree-five reductions. The origin-spoke and critical-value-envelope rows refuted in FRONTIER.md are not continuation targets.
  3. PartialClusterPreimagePerimeterIdentity.md, TiedNewtonFaceFibreProduct.md, and ExteriorRootProductCovering.md for exact perimeter, moment, and covering premises; each still requires an attachment-compatible selector or metric consumer.
  4. MinimalHubArmBudgetRefutation.md and SeparatrixAggregateReduction.md for the strongest negative evidence and the machinery that survives only as a possible tool for a different functional.

The paper-facing route remains erdos-1041-lemniscate-newton-flow.md, which points to the checked Newton-flow declarations and states the older topology/metric boundary. The current source frontier does not change the open status of #1041, and its machine rows must not be read as Lean claims.

Expansion-problem reading handoff (not a wave)

The five expansion roots have no shared Lean-wave chronology. Their exact paper/source joins and open boundaries live in the complete eight-problem return matrix and the generated problems.json index. The paper notes are the shortest first read:

Problem Paper note Principal source
#68 factorial carries ErdosProblems/Erdos68/FactorialZeroPlateau.lean
#243 reciprocal-tail rigidity ErdosProblems/Erdos243/ReciprocalTailRigidity.lean
#251 prime-gap dyadic tails ErdosProblems/Erdos251/PrimeGapDyadicTail.lean
#269 three-prime running lcm ErdosProblems/Erdos269/ThreePrimeRunningLcm.lean
#1049 rational-base Lambert RationalBaseLambert · ZudilinHeightRegion · RationalPadeArithmetic · HermitePadeNoGo · QAperyDiagonalNonEquivalence · ZudilinConeArithmetic

These notes preserve their own evidence classes and limitations; the source map is the route to the exact continuation obligation, not a chronology or a claim-status authority.

Package shape

  • CertificateKernel.lean (0.85 MiB, 19,278 lines, 845 declarations; 519 theorems and 275 lemmas): the assembled microkernel and headline interfaces.
  • GeneratedCertificates.lean (1.18 MiB, 27,728 lines, 1,174 declarations) plus 0 generated shards: finite certificate instances checked by the Lean kernel.
  • The diagonal-pincer family contains 0 isolated prime-certificate modules and 0 scale aggregators. The shards are indexed through aggregators rather than presented as separate mathematical claims.
  • Entire checked corpus: 1,903 modules, 164,639 declarations, 160,373 theorem-like declarations, and 8,171 manifest-marked generated-certificate declarations (a classification floor, not the generated share). The release gate rejects sorry, admit, custom axiom declarations, and native_decide.

These are generated inventory facts, not mathematical claim counts. The declaration atlas and Lean source remain the drilldown owners.