Li's criterion is the sharpest known equivalence between the Riemann hypothesis and a concrete sequence of real numbers. Define, over the nontrivial zeros ρ of ζ,
Then RH holds if and only if λₙ ≥ 0 for every n ≥ 1 (Li 1997). The constants are logarithmic derivatives of the completed zeta function at s = 1 — equivalently, coefficients of the Keiper generating function log ξ(1/(1−z)) = −log 2 + Σ λₙ zⁿ/n — and each one individually measures how far the zeros collectively are from leaving the critical line.
This post reports the current state of a small but rigorous effort: the first three coefficients of that sequence are now certified end to end — numerically in our own pipeline, symbolically in Lean 4 + mathlib, with every quantitative claim machine-checked and every identification documented. The chain is strictly increasing: λ₁ ≈ 0.0231 < λ₂ ≈ 0.0923 < λ₃ ≈ 0.2076.
Every λₙ passes through the same three-stage pipeline before it enters Lean.
Gate 1 — a certified interval from on-line zeros. For a zero ρ = 1/2 + iγ (upper half plane), set θ = 2·arctan(1/(2γ)) — an exact identity, not an approximation. The conjugate pair contributes 2(1 − cos nθ) ≥ 0 to λₙ, so every partial sum S_K is a rigorous lower bound. A standard Riemann–von Mangoldt tail estimate supplies the matching upper bound: λₙ ∈ [S_K, S_K + T_K] with T_K = n²(ln(γ_K/2π) + 2)/γ_K + n²/γ_K². With K = 2000 verified zeros (γ_K = 2515.29), these intervals are about 0.028–0.031 wide.
Gate 2 — a closed form, derived rather than recalled. The literature's closed forms for λₙ are easy to mis-transcribe (we caught two convention mismatches along the way), so each one is derived: compute λₙ to 80 decimal places via an independent route — the derivative formula λₙ = 1/(n−1)! · dⁿ/dsⁿ [sⁿ⁻¹ log ξ(s)]|₁ — then run PSLQ integer-relation detection against the basis {1, γ, γ², γ³, γ₁, γγ₁, γ₂, ln 2, ln π, π², ζ(3)}. The relation comes out with small exact integer coefficients, and the result is then confirmed three independent ways: against OEIS, against the derivative route (agreement to ~10⁻⁵⁰), and against the certified sandwich interval from Gate 1.
Gate 3 — an exact replay of every Lean inequality. The Lean proofs bound constants by decimals (more below). Before writing a single line of Lean, a Python script re-verifies every witness inequality in exact rational arithmetic (fractions.Fraction), so the certificates in the theorem statements are known to be non-vacuous to the last digit.
| n | certified interval (K = 2000) | closed form | value |
|---|---|---|---|
| 1 | [0.022961, 0.023908] (exact interval arith.) | 1 + γ/2 − ln 2 − (ln π)/2 | 0.0230957090… |
| 2 | [0.090576382823, 0.103286923022] | 1 + γ − γ² − 2γ₁ − 2 ln 2 − ln π + π²/8 | 0.09234573522804667… |
| 3 | [0.203657877722, 0.232256593171] | 1 + (3/2)γ − 3γ² + γ³ − 6γ₁ + 3γγ₁ + (3/2)γ₂ − 3 ln 2 − (3/2) ln π + (3/8)π² − (7/8)ζ(3) | 0.20763892055432480… |
The λ₃ closed form is the newest member. PSLQ returned the integer relation and the result was confirmed against OEIS A104540, against the derivative route (agreement 5.0 × 10⁻⁵¹), and against a five-point central third difference of the Keiper generating function (agreement 1.7 × 10⁻¹⁰). The new constants entering are the second Stieltjes constant γ₂ = −0.0096903631928723… and Apéry's constant ζ(3) = 1.2020569031595942853… — a pattern that will continue: λₙ is a polynomial in γ of degree n plus Stieltjes constants up to γₙ₋₁, and for n ≥ 4 the zeta values ζ(4), ζ(5), … join.
The Lean side (Riemann/LiCriterion.lean) contains no numerics — it contains interval theorems:
theorem liLambda1_mem : 0.0054 < liLambda1 ∧ liLambda1 < 91 / 1000
theorem liLambda2_mem : (0.07 : ℝ) < liLambda2 ∧ liLambda2 < 0.15
theorem liLambda3_mem : (0.15 : ℝ) < liLambda3 ∧ liLambda3 < 0.3
theorem liLambda1_lt_liLambda2 : liLambda1 < liLambda2
theorem liLambda2_lt_liLambda3 : liLambda2 < liLambda3
The interesting question is how a machine proves an inequality about γ, γ², γ³ and γγ₁ when the only facts available are a box γ ∈ (0.5604, 0.594) and certified decimals for the Stieltjes constants. Three tricks do all the work:
Certified decimals, zero axioms. Each transcendental constant enters as a plain def holding a 31-digit decimal — stieltjes1Certified, stieltjes2Certified, zeta3Certified. A def is not an axiom: it creates a new constant equal to its decimal literal, by rfl. The identification of that decimal with the true Stieltjes constant is external mathematical content, documented in the audit file — exactly how a paper cites a table.
The concavity-product trick. The γ-atoms of λ₃ are the cubic γ³ − 3γ² + (3/2)γ, which is strictly decreasing on the box. Instead of bounding γ³ and γ² separately (lossy), the proof uses the exact algebraic identity
proved by ring, with both factors shown positive by elementary means (γ² < 0.594γ and the box bounds). This is one algebraic degree higher than the corresponding identity used for λ₂ — the same trick, scaled with the coefficients.
Boxing the cross term. The atom γγ₁ is handled by pure sign reasoning: since γ₁ < 0, multiplying the box γ ∈ (0.5604, 0.594) by γ₁ flips the inequalities, boxing γγ₁ between the two constants 0.594·γ₁ and 0.5604·γ₁.
Every load-bearing declaration is audited by #print axioms. All nine new declarations bill exactly [propext, Classical.choice, Quot.sound] — no sorryAx, no native_decide, no custom axioms.
Li's criterion needs λₙ ≥ 0 for all n — infinitely many. A certified chain λ₁ < λ₂ < λ₃ proves nothing about n = 4 directly; it proves that the sequence's head is positive, growing, and machine-checked at every step. The value is methodological: the pipeline (sandwich → PSLQ → exact replay → Lean) now has zero remaining manual steps, and each new n is a bounded increment — new Stieltjes decimals, a new cubic identity, a new interval. The mathematics remains open exactly where it was: the single frontier sorry ("no zeros in (1/2, 7/8]") in our build is untouched, and closing it would subsume this entire ledger, since the quasi-RH ingredient above 7/8 plus any proof of the sliver would give the zero structure that makes positivity a theorem rather than a computation.
There is also a quiet epistemological point. The λₙ are doubly determined — once by the zeros (the sandwich) and once by the Stieltjes constants (the closed form). That our two pipelines agree on all three values to high precision is a consistency check across two independent derivations of the same constants, and the kind of check that has historically caught real errors.
All artifacts are in the research repository:
scripts/routes/li_criterion_certified.py, li_lambda2_certified.py, li_lambda3_certified.py — the certification pipelinesdata/routes/rt_lc_lambda1.json, rt_lc_lambda2.json, rt_lc_lambda3.json — certified intervals, closed forms, gate verdictslean/Riemann/LiCriterion.lean — the Lean formalization (commits 89c9429, 304e6cb)research/QUASI_RH_CERT_AUDIT.md — the audit trail, Parts 1–5References: Li, X.-J., J. Number Theory 65 (1997) 325–333; Keiper, J. B., Math. Comp. 58 (1992) 765–773; Coffey, M. W., J. Math. Phys. 46 (2005) and Proc. R. Soc. A 464 (2008) 711–731.