bibkey: kimihiro2026robin1984 authors: Jonas Whidden year: 2026 title: “Robin1984: the actual Robin, colossally abundant and Lagarias equivalences” doi: null url: https://github.com/kimihiro64/Robin1984/tree/acab1a31f31e0499518a4416b63281e0b4838f9c claim: “The pinned source proves actual Mathlib RH equivalent to the strict actual divisor-sum inequality for every natural n > 5040, with actual standard-three-axiom Comparator, NanoDa and Lean replay evidence; its unchanged arithmetic carrier compiles locally, while the full proof closure is not yet ported.” strata_touched:
- D5/S3/Arith/Robin/PrimePrefixMobiusDirectedAbel
- D5/S3/Arith/Robin/PrimePrefixOriginalA license: Apache-2.0 triage: anchor
Robin’s criterion with upstream proof acceptance
Source repository kimihiro64/Robin1984,
immutable main commit acab1a31f31e0499518a4416b63281e0b4838f9c.
The pinned CITATION.cff
names Jonas Whidden as the formalization author. The repository account name
does not establish an additional personal author. The mathematical Robin
equivalence belongs to Guy Robin (1984), its converse uses Nicolas–Landau,
and the separate elementary criterion belongs to Jeffrey C. Lagarias.
Exact mathematical supply
The pinned Definitions.lean
sets sigmaOneNat n := ArithmeticFunction.sigma 1 n and uses Mathlib’s
Real.eulerMascheroniConstant. Its predicate is exactly
Robin1984.Core.NativeRobinInequalityAll universally quantifies this strict
inequality over every natural n with 5040 < n.
The theorem
Robin1984.riemannHypothesis_iff_nativeRobinInequalityAll in
Equivalence.lean
is the unconditional equivalence between that predicate and Mathlib’s actual
RiemannHypothesis. The imported Mathlib definition concerns zeros of the
actual riemannZeta, excludes the negative even trivial zeros and s = 1,
and requires real part 1/2. There is no project replacement RH predicate.
The forward proof partitions the integer domain at Real.log (n : Real) = 74500.
FiniteComplete.lean
proves the literal Robin inequality unconditionally when 5040 < n and
log n < 74500. Exact startup certificates cover 5040 < n < 720720;
the remaining finite logarithmic interval is covered by certified tangents.
LargeHeightRobin.lean
proves the large range under an explicit actual RH assumption. The reverse
proof in
NicolasLandauRobinBridge.lean
uses the proved actual Nicolas negative oscillation under failure of RH.
The three public proved endpoints in
Solution.lean
are Robin1984.robin_inequality_iff_riemannHypothesis,
Robin1984.riemannHypothesis_iff_colossallyAbundant_robin and
Robin1984.riemann_hypothesis_iff_lagarias_elementary_criterion.
The colossally abundant endpoint restricts the same Robin inequality to the
actual extremizers of the parameterized abundancy ratio. Each endpoint is an
equivalence; it does not establish either side unconditionally.
Actual verification and dependency boundary
The saved authenticated GitHub API results identify
CI run 34748103247
with the exact source commit above and successful Lean build and
statement-against-solution job 103700837549.
The complete saved job log exports the three public endpoints at line 1623,
then records actual NanoDa acceptance, actual Lean default-kernel acceptance
and Your solution is okay! at lines 1624–1628. This is a recorded upstream
run, not a new local replay of that proof closure.
The pinned comparator.json
permits only propext, Quot.sound and Classical.choice, and enables NanoDa.
The pinned
verification script
requires enable_nanoda to be exactly true and fixes Comparator at
68a064109f01c08f47c8edc9f51d6a2bbffaa188.
The actual Comparator implementation was read:
Util.lean
traverses types and proof values, including opaque values;
Axioms.lean
recursively rejects unpermitted axioms reachable from the target theorems;
Main.lean
checks statement/definition agreement and these axioms before NanoDa and Lean
replay. An error aborts that sequence. NanoDa also receives
unpermitted_axiom_hard_error = true.
The saved own-source import closure contains 149 modules (1,517,626 bytes)
for Equivalence.lean and 167 modules (1,742,374 bytes) for Solution.lean.
Their Git blobs and byte hashes were checked. A comment/string-stripped
lexical scan found no proof-hole, custom-axiom or unsafe/native-evaluation
tokens in those own closures. The one opaque is an initialized
CA5040EventsPackage with literal actual events and an rfl equality.
The separate Challenge.lean contains its three intended statement holes
and is not imported by the solution proof.
These are source import closures, not minimal proof-constant closures or a
full manual review of all proof bodies. Their five external PNT module
seeds lead to 99 modules (2,350,044 bytes) in the exact fork
kimihiro64/PrimeNumberTheoremAnd at
f8f58c749d6cde8a641348fcd5e4702993651cd6.
That imported fork closure contains 13 literal sorry declarations in four
modules. The actual target axiom enforcement and dual-kernel acceptance
exclude those placeholders from the three selected theorem proofs; they do
not certify every declaration in those imported modules. No complete
third-party source audit or new local #print axioms of the three upstream
endpoints is claimed.
Local entry and intended consumers
Upstream uses Lean v4.33.1, Mathlib
0df444a360eaa60ab8c11dca51a86af692955474, the PNT fork above, leancert
621a43d7cf21f87872392a01e874f2f1dbddc926, PrimeCert
916ee9c35a57af128d5089a69172415051e700ca and LeanArchitect
d9013cc08bd2b5483e837368dfa4cc7ead92a5c2. The complete manifest is retained
with the source snapshot.
The unchanged Arithmetic/Definitions.lean and a transient Iff.rfl
check of its literal all-n > 5040 statement compiled with actual exit 0
under this repository’s Lean v4.33.0, Mathlib
db584cd6d46c92f209a44c0f1c829460d327499d, --trust=0 and
debug.skipKernelTC=false. The receipt is
robin1984-local-carrier-native36/receipt.json; both upstream and local
definition bytes have SHA256
cac28c82777fc92ea8ec4e18eed1b29d2b2a361e40f3fe3c1f3e4d9c8e5b1b88.
This proves compatibility of the arithmetic statement. The 149-module
equivalence proof and its external closure have not been ported or given
local canonical acceptance. No D5 wrapper was retained for the transient
definitional check.
The strata_touched entries name existing local Robin research targets.
PrimePrefixMobiusDirectedAbel retains the actual harmonic Mobius prefix,
same finite window and signed anchor/terminal, with prefix estimates as
caller hypotheses. PrimePrefixOriginalA proves the original two-piece
Phi constant’s integrability and strict bound below 1/2. Neither currently
imports the external equivalence. Connecting the original signed arithmetic
tail to the literal all-integer Robin predicate remains a mathematical
obligation; a locally accepted equivalence would then supply a precise RH
endpoint. The external unconditional finite certificates offer an additional
concrete source for replacing repeated finite-range work.
Attribution and licence
The complete upstream root
Apache-2.0 LICENSE
was read and retained: 11,357 bytes, Git blob
261eeb9e9f8b2b4b0d119366dda99c6fd7d35c64, SHA256
c71d239df91726fc519c6eb72d318ec65820627232b2f796219e87dcf35d0ab4.
The complete pinned repository tree has no root or proof-path NOTICE file.
The exact PNT fork carries the same root licence, and its added xi-divisor
source retains Matteo Cipollina’s 2026 author/copyright header. Upstream
LICENSING.md
does not relicense mathematical papers or third-party dependencies.
This note distributes no transplanted upstream proof code. Any later source
port must retain the applicable licence, original notices, immutable source
pin and description of modifications.
Local prime-event proof closure and price connection
A later strict local check compiled the unchanged three-module closure
Robin1984.Helpers.Event, Robin1984.Finite.CutoffIndex and
Robin1984.ColossallyAbundant.EventGainFormula on the same repository
Lean v4.33.0 and Mathlib pin, with --trust=0 and
debug.skipKernelTC=false. Each source compilation and the exact price
compatibility check exited 0. The complete receipt is
robin1984-event-compatibility-20261008/receipt.json.
The two actual supplier proofs
FiniteSupport.simplifiedLayerGain_le_of_base_le and
event_gain_eq_caSimplifiedLayerGain print only propext,
Classical.choice and Quot.sound.
The fixed positive layer j has the exact gain
The substantive upstream proof rewrites its rational term as the inverse of
sum_(i<j) q^(i+1), then proves that g_j decreases with the real base q.
Its event bridge identifies this formula with the actual prime-exponent
insertion gain. A transient compiled check identifies the independently
defined event price e.threshold with the existing local
goldenLayerMarginal e.p e.j. This is a genuine coordinate match to the
prime-price route; no renamed D5 price wrapper was retained.
The next actual integer packet source, CAProfile.lean, has a five-module
own-source closure with no PNT fork or certificate imports. It proves that
actual prime-layer packet mass and gain equal log n and
log(sigma(n)/n). A subsequent strict local check compiled the unchanged
five-module CAProfile closure and the additional RobinLogMargin.lean on
this same local pin. The already accepted Definitions and Event oleans were
reused; all four new upstream-source compilations and the axiom check exited 0.
The full receipt is
robin1984-actual-packet-compatibility-20261008/receipt.json.
The proved CA existence theorem, actual packet log-mass and log-abundancy
identities, and literal Robin margin equivalence each print only the standard
three axioms. This establishes local compatibility of the actual integer
packet and margin suppliers, with no renamed D5 wrapper retained.
CAReduction.lean has an 88-module own-source closure, including 59
certificate modules, and is not a three-module consequence. The full
149-module RH equivalence remains unported.
Consuming the compatible gain monotonicity in a new price-tail proof, connecting the actual integral reserve/head/tail to the literal integer Robin margin, and proving its strict signed inequality remain separate mathematical tasks. The small accepted supplier supplies none of those unproved conclusions by itself.