We prove that ζF(3)=∑n≥1Fn−3 is irrational, where Fn are the Fibonacci numbers. To our knowledge the only odd value of the Fibonacci zeta function previously known to be irrational was ζF(1). The proof is a Hankel-determinant construction over Z[φ], φ=(1+5)/2, adapted from a recent unrefereed construction for the q-zeta value ∑n2/(pn−1) at integers p≥2. Four estimates are combined: a real decay of size exp(−92K3logφ), a smallness exp(−K3logφ) at the Galois conjugate, which replaces the base-prime content gain of the integer case, local cyclotomic denominator bounds at the twisted roots of unity that occur, and nonvanishing by rational interpolation. The whole argument is formalised in Lean 4 with Mathlib.
Introduction
Let Fn be the Fibonacci numbers and ζF(s)=∑n≥1Fn−s the Fibonacci zeta function. André-Jeannin proved that ζF(1) is irrational [1], by Padé approximation of a Lambert series. Duverney, Ke. Nishioka, Ku. Nishioka and Shiokawa proved that the even values ζF(2s) are transcendental, and so are ∑n≥1F2n−1−s for every s≥1[2]; Elsner, Shimomura and Shiokawa determined the algebraic relations among these numbers [3, 4]. All of these sums are values of quasi-modular forms, to which Nesterenko’s theorem [11] applies. The odd values ζF(2s+1), s≥1, are not of this kind, and as far as we know nothing was proved about their arithmetic (see the survey [16] and the introduction of [10]).
Theorem 1.1.ζF(3)=∑n≥1Fn31 is irrational.
With t=φ−1=(5−1)/2 and Binet’s formula, Fn=t−n(1−(−t2)n)/5, so
Ξ:=55ζF(3)=N≥1∑(1−(−t2)N)3t3N.(1)
This is a q-series at the quadratic unit t, a relative of the q-zeta value ζq(3)=∑nn2qn/(1−qn)[9]. For q=1/p with p≥2 an integer, the irrationality of ζq(3) remains open in the refereed literature: Krattenthaler, Rivoal and Zudilin proved that one of ζq(3),…,ζq(11) is irrational [9], Jouhet and Mosaki reduced this to ζq(3),…,ζq(9)[7], and Fischler and Zudilin proved linear independence results of the same kind [6]. In October 2026 a proof that ∑n≥1n2/(pn−1) is irrational for every integer p≥2 was claimed in an unrefereed, AI-assisted research draft with a Lean formalisation, posted on GitHub under the pen name uplift-universal-inc [14, 13]. Its construction is a moving-denominator Hankel determinant in the line of Uvarov [15], Krattenthaler [8], Prévost [12], Zudilin [17, 18] and the unrefereed preprint of Fauzan [5].
Expanding (1.1) in Lambert series does not reduce Theorem 1.1 to a statement about ζq(3): one obtains a Q(5)-linear combination of 1, odd parts of ζq(3) and ζq(1) at q=φ−2, and values of holomorphic modular forms of weights 1 and 3 at q=φ−1,φ−2. We therefore treat the series (1.1) directly.
The method
We adapt the construction of [14] to the real quadratic field Q(5). A linear functional L, a second Euler derivative summed against the weights (−t)n at the nodes (−t2)n, has rational moments and pure simple-pole tails of Ξ (Lemma 2.2). The Hankel determinants ΔK(X) of L against the moving denominators DK(s)=∏a≤K(1−(−t2)as) are polynomials of degree ≤K in X with ΔK(Ξ)=det[L(si+j/DK)]. Four estimates are combined:
size (per K3logφ)
real decay ∣ΔK(Ξ)∣
−2/9
Proposition 3.1
conjugate ∣σΔK(−Ξ)∣
−1
Proposition 5.1
denominators log∣NCK(t)∣
≤c∗=1.1954…<11/9
Propositions 4.4, 4.6
nonvanishing
infinitely many K
Proposition 6.1
Table 2.
If ζF(3) were rational, CK(t)(25v)KΔK(Ξ) would be a nonzero element of Z[φ] of norm <1 for large K.
Three ingredients are new compared with [14]. First, the setting: the Fibonacci kernel is a single q-series over Z[φ], handled with a formal parameter τ specialised at τ=t and at its conjugate τ=−φ. Second, since φ is a unit there is no base prime and no coefficient content to extract. Its role is taken by the smallness of the conjugate determinant at the conjugate point σΞ=−Ξ (Proposition 5.1). Its size, K3logφ=21K3logφ2, is the size of the content gain 21K3logp of [14] at p=φ2. Third, the poles of the coefficients sit at two kinds of twisted roots of unity, those ζ for which −ζ2 has small order and those for which −ζ(−ζ2)k=1; we prove local denominator bounds of both types (Section 4). The present paper does not depend logically on [14]: every step is proved here, and the formal proof is self-contained.
Formal verification
The theorem and every step of its proof are formalised in Lean 4 with Mathlib (https://github.com/telaviv/fibonacci-zeta3-irrationality); Section 8 gives the statement, the map from the propositions below to Lean declarations, and what has to be trusted. The formal proof uses the elementary denominator constant
c∗=811β7+31=36754393
of Proposition 4.6, where β7=∏p≤7(1−p−2); it needs no information on the distribution of Euler’s function beyond a sieve by 2, 3, 5, 7.
The functional and the determinants
Notation
Throughout, φ=(1+5)/2 is the golden ratio, t=φ−1, O=Z[φ], and σ is the nontrivial automorphism of Q(5), so σt=−φ. Euler’s function is written φE(d) and the Möbius function μ(d). We put
ρ=−t,w=−t2,αk=ρwk(k≥0),
ψN=(1−wN)3t3N,Sa=N=1∑aψN,
AK=K(K+1)/2, and θ=sdsd. Then ρ2=−w, ∣w∣=t2 and Ξ=∑NψN. We write hm for the complete homogeneous symmetric polynomial of degree m (hm=0 for m<0). Constants in O(⋅) are absolute and K≥2.
Definition 2.1. For g analytic on a neighbourhood of ∣s∣≤t2 put L(g)=∑n≥1ρn[θ(θ+1)g](wn).
The series converges absolutely since θ(θ+1)g=O(s) and ∣ρnwn∣=t3n.
Lemma 2.2. (i) L(sk)=νk:=k(k+1)αk/(1−αk) for k≥0. (ii) For a≥1, L(1/(1−was))=λa(Ξ−Sa) with λa=2ρ−a.
Proof. (i) θ(θ+1)sk=k(k+1)sk; sum the geometric series. (ii) θ(θ+1)(1−cs)−1=2cs/(1−cs)3, so the left side is ∑nρn2wa+n/(1−wa+n)3=2ρ−a∑N>aρNwN/(1−wN)3, and ρNwN=t3N. □
Let DK(s)=∏a=1K(1−was), with poles xa=w−a, ∣xa∣=φ2a. For a polynomial P of degree <2K−1,
Every quantity above is the value at τ=t of a rational function of an indeterminate τ with rational coefficients: replace t by τ, so ρ=−τ, w=−τ2, αk=(−1)k+1τ2k+1 and ψN=τ3N/(1−(−τ2N)3). Write ΔK,τ(X)∈Q(τ)[X] for the resulting polynomial, and w(ζ)=−ζ2, αk(ζ) for the values at τ=ζ. Since σt=−φ, the coefficientwise conjugate σΔK is the value of ΔK,τ at τ=−φ. Both values are defined because no denominator vanishes at a real number that is neither 0 nor a root of unity.
Proof. Let Vj(s)=∏u=1j(s−w⌈u/3⌉), Rj=⌈j/3⌉ and aj=∑u≤j⌈u/3⌉. The change of basis from 1,s,…,sK−1 to V0,…,VK−1 is unitriangular, so ΔK(Ξ)=det[L(ViVj/DK)].
Since θ(θ+1)g(s0) depends only on the 2-jet of g at s0, the m-th term of L(ViVjW), W=1/DK, vanishes for m<R:=max(Ri,Rj): then ViVj has a triple zero at wm. For m≥R and s=wm every root β of Vj satisfies ∣s∣≤∣β∣, so ∣s−β∣≤2∣β∣ and θ(s−β)=s. Hence ∣Vj(s)∣≤2j∣w∣aj and ∣θuVj(s)∣≤ju2j∣w∣aj∣s∣/∣w∣Rj for u=1,2. On ∣s∣≤t2 we have ∣was∣≤t2a+2, so ∣W∣, ∣θW∣/∣s∣ and ∣θ2W∣/∣s∣ are bounded by an absolute constant. Since θ(θ+1)=θ2+θ kills constants, by Leibniz every term of θ(θ+1)(ViVjW) contains at least one θ on some factor, hence a factor ≤∣s∣/∣w∣R. With ∣ρ∣m≤1 and ∑m≥R∣w∣m−R≤(1−t2)−1,
∣L(ViVjW)∣≤c1(i+j+1)24i+j∣w∣ai+aj.
for an absolute constant c1. Pull ∣w∣ai out of row i and column i and apply Hadamard’s inequality: ∣ΔK(Ξ)∣≤∣w∣2∑j<Kaj(c1(2K)242KK)K. Finally 2∑j<Kaj=(K3+3K2−ιK)/9 with ιK∈{0,4,2} according to Kmod3. □
Denominators
The integrality ring and the root types
Lemma 4.1. The coefficients of ΔK,τ(X) lie in Z[τ,τ−1,Φd(τ)−1:d∈DK], where DK is the set of orders of roots of unity ζ such that (W) w(ζ) has order e≤K, or (α) αk(ζ)=1 for some 1≤k≤K−2.
Proof. The inputs are the Laurent monomials wa and λa=2(−τ)−a; the factors 1/(1−wb−a) and (1−wN)−3 (N≤K), where 1−wn=1−(−1)nτ2n is ± a product of cyclotomic polynomials Φd(τ) with w(ζ)n=1; and νk=k(k+1)αk/(1−αk), of which only k≤K−2 occur (the quotient has degree ≤2K−2−K), with 1−αk=1∓τ2k+1=±∏Φd(τ). Division by DK is integral because its leading coefficient ∏a(−wa)=±τK(K+1) is a unit. No rational constant is inverted. □
For d∈DK the type depends only on d. If ζ has order d, then w(ζ) has order e(d)=2d (d odd), d (d≡2(mod4)), d/4 (d≡4(mod8)), d/2 (d≡0(mod8)), and for each e the orders d with e(d)=e satisfy ∑e(d)=eφE(d)=2φE(e). The least k≥1 with αk(ζ)=1 is kmin(d)=(d−1)/2 (d≡3(mod4)), (3d−1)/2 (d≡1(mod4)), (d−2)/4 (d≡2(mod8), d≥10), (3d−2)/4 (d≡6(mod8)), and kmin(2)=2 (the formula (d−2)/4 would give 0); there is none if 4∣d. (Every primitive d-th root is a power ζj with j coprime to 2d; then −ζ2j=(−ζ2)j and αk(ζj)=αk(ζ)j, and powers coprime to the order preserve the order.) All d∈DK satisfy d≤4K: w(ζ)e=1 gives ζ4e=1, and αk(ζ)=1 gives ζ4k+2=1.
Local setting. Fix ζ of order d∈DK, let ϖ=τ−ζ and let v=ordϖ on Q(ζ)((ϖ)), extended to polynomials in X as the Gauss valuation (the minimum over the coefficients; it is multiplicative). Then:
(a) if e(d)=e≤K, then xa(ζ)=xb(ζ) iff e∣a−b, and then v(xa−xb)=1: indeed xa−xb=xa(1−wb−a), and 1−wn has only simple zeros, since dτdwn=0 at τ=0;
(b) ψN has a pole of order exactly 3 if e∣N and is regular otherwise, by (a); so Sa has a pole of order ≤3, and is regular if a<e;
(c) 1−αk=1∓τ2k+1 is separable, so it has at most a simple zero at ζ;
(d) λa and xa are Laurent monomials, hence units.
Local bounds
Put c(m)=2∑j=0m−2max(0,2m+2−j) and c′(m)=2∑j=0m−1max(0,2m+2−j); thus c(1)=0, c(m)=4(m−1) for m≤6, c(m)≤(m+3)2/4, c′(1)=3 and 0≤c′−c≤3.
Proposition 4.2. Let e=e(d)≤K and partition {1,…,K} into the clusters Ir={a≡r(mode)} of sizes mr, 1≤r≤e. Then
with ϵ=1 if ζ is also an α-root and ϵ=0 otherwise.
Proof. Let F(s)=∏a(s−xa) and Fr=∏a∈Ir(s−xa). Fix a cluster r, write m=mr and Ir={a1<⋯<am}.
Newton cluster basis. For 0≤j<m put Nr,j=(F/Fr)∏ℓ=m−j+1m(s−xaℓ); it vanishes at the top j nodes of the cluster. At ϖ=0 these reduce to (F/Fr)(s−yr)j⋅(unit), yr=xr(ζ), which form a basis of the polynomials of degree <K by the Chinese remainder theorem (distinct residues r give distinct points yr). So the change of basis from 1,…,sK−1 has unit determinant, and v(ΔK,τ) is the valuation of G=[LX(NN′/DK)] in this basis.
Entries.G(r,j),(r′,k)=P+∑aresa(NN′)λa(X−Sa), where P is LX of the polynomial part. (i) If r=r′ then F∣Nr,jNr′,k and all resa=0. (ii) If r=r′, then resai=0 unless i≤m−max(j,k), and for such i
Here (F/Fr)(xai) is a unit, each displayed difference has valuation exactly 1 by (a), and the denominator has valuation exactly m−1. So v(resai)=j+k−(m−1). (iii) v(λa)=0, and by (b) the Gauss valuation of X−Sa is ≥−3 if a≥e and ≥0 if a<e. (iv) For r<e the node a1=r is regular and ai is singular for i≥2; for r=e every node is singular. So the proper part of the (j,k) entry has v≥j+k−(m−1)−3 if some singular node has i≤m−max(j,k), and v≥j+k−(m−1) otherwise. The second case occurs exactly when r<e and max(j,k)=m−1.
Weights. Let κj=max(0,2m+2−j) for j≤m−2 and κm−1=0 if r<e, and κj=max(0,2m+2−j) for all j<m if r=e. We claim, following the two cases of (iv):
(A) if the (j,k) entry has a singular node (all pairs when r=e; the pairs with max(j,k)≤m−2 when r<e), then κj+κk≥m+2−j−k;
(B) if r<e and max(j,k)=m−1, then κj+κk≥0≥m−1−j−k.
In case (A) both indices carry the weight max(0,2m+2−j) (for r=e this includes κm−1=max(0,24−m). If both weights are positive the inequality is an equality. If κk=0, then k≥2m+2, so κj≥2m+2−j≥m+2−j−k. If both vanish, then j,k≥2m+2 and the right side is ≤0. Case (B) holds because j+k≥m−1.
Scale row (r,j) by ϖ⌈κj⌉ and column (r,j) by ϖ⌊κj⌋. The required exponents are integers, and a+b≥n∈Z implies ⌈a⌉+⌊b⌋≥n, so by (iv), (A) and (B) every scaled proper entry is integral. The total scaling is ∑(⌈κj⌉+⌊κj⌋)=2∑κj, which is c(m) for r<e and c′(m) for r=e. All exponents are ≥0.
Polynomial part. The quotient of NN′ by DK has degree ≤K−2 and coefficients in Z[τ±1], so P=∑k≤K−2(coeff)νk. At a pure W-root every νk, k≤K−2, is regular, so P is integral, and so is the scaled matrix; this gives the bound with ϵ=0.
At an α-root, put Pα={k≤K−2:αk(ζ)=1} and write νk=ϖ−1uk with uk regular for k∈Pα (by (c)). Split P=Apol+ϖ−1B, where B(f,g)=∑k∈Pαuk(0)[sk](fg/DK)+ (here (⋅)+ is the polynomial part) and Apol collects ϖ−1(uk−uk(0))[sk](⋅)+ for k∈Pα and νk[sk](⋅)+ for k∈/Pα; both Apol and B have regular entries. If degf≤kmin, then deg(fg)≤kmin+K−1, so the quotient has degree <kmin and B(f,⋅)=0. Since the polynomials of degree ≤kmin form a space of dimension kmin+1, rankB≤max(0,K−1−kmin). After scaling, the matrix is A+ϖ−1B′. Here A is the sum of the scaled proper parts, integral by (A) and (B), and the scaled Apol, integral because all exponents are ≥0. The matrix B′ is B in the Newton basis, scaled; its entries are regular and its rank is that of B, since the scalings are invertible over the fraction field. Expand the determinant multilinearly in the columns: a term taking more than max(0,K−1−kmin) columns from ϖ−1B′ vanishes by linear dependence, and every other term has valuation ≥−max(0,K−1−kmin). □
Lemma 4.3. If e(d)>K and ζ is an α-root, then −v(ΔK,τ)≤max(0,K−1−kmin(d)).
Proof. Work in the monomial basis. Since e>K, no difference ∣a−b∣<K is divisible by e, so by (a)–(d) the coefficients resa, the λa and the Sa (a≤K<e) are regular at ζ. Only the νk, k∈Pα, carry poles, and the rank argument of Proposition 4.2 applies with no scaling. □
Proposition 4.4. Let bK,d be the bound of Proposition 4.2 or Lemma 4.3 (0 if d∈/DK) and CK(τ)=∏dΦd(τ)bK,d. Then CK(τ)ΔK,τ(X)∈Z[τ±1][X]; hence CK(t)ΔK(X)∈O[X], and CK(−φ)σΔK(X)∈O[X] is its conjugate.
Proof. By Lemma 4.1 a coefficient is P/(τn∏dΦdnd) with P∈Z[τ]. The bound vζ≥−bK,d at every primitive d-th root forces Φdnd−bK,d∣P in Q[τ], hence in Z[τ] since Φd is monic. Specialise τ↦t, a unit of O, and τ↦−φ. □
The size of CK(t)
Lemma 4.5. For d≥1, ∣log∣Φd(t)Φd(−φ)∣−ϕE(d)logφ∣≤2E1, where
E1=a≥1∑∣log(1−φ−a)∣=2.1136…
Proof.Φd(x)=∏a∣d(xa−1)μ(d/a) and ∑a∣dμ(d/a)a=ϕE(d). For ∣x∣=φ write xa−1=xa(1−x−a), and ∣log∣1−x−a∣∣≤∣log(1−φ−a)∣; for x=t use ∣1−ta∣ in the same way. □
Let β7=∏p≤7(1−p−2)=1225768 and c∗=811β7+31=36754393=1.19537….
Proposition 4.6.log∣NCK(t)∣≤c∗K3logφ+O(K2logK).
Proof. By Lemma 4.5, log∣NCK(t)∣≤logφ∑dbK,dϕE(d)+2E1∑dbK,d. Here ∑dbK,d=O(K2logK): for each e≤K at most four d have e(d)=e, each with a W part ≤4K+K2/(4e)+3 (see below), and the at most 4K orders d≤4K have α parts ≤K. Split bK,d into its W part and its α part.
W part. Since mr≥1, ∑rmr=K and mr<K/e+1, and since c(m)=4(m−1)+δ(m) with 0≤δ(m)≤max(0,m−5)2/4 (for m≥6, c(m)=(m+2)(m+4)/4 or (m+3)2/4 according to the parity of m),
r<e∑c(mr)+c′(me)≤4(K−e)+4emax(0,K−4e)2+3.
Summing over the orders d with e(d)=e gives the weight ∑e(d)=eϕE(d)=2ϕE(e). Bound ϕE(e)≤eη(e) with η(e)=∏p≤7p∣e(1−1/p), a function of period P=210 with 0≤η≤1 and mean β7. So the W part is at most ∑e≤Kη(e)fK(e), where
fK(y)=8y(K−y)+21max(0,K−4y)2+6y.
We use the elementary estimate: if η has period P, 0≤η≤1 and mean β, and f is a function on [0,K] with ∣f∣≤M and total variation V, then
1≤e≤K∑η(e)f(e)−β∫0Kf(y)dy≤2P(M+V).
(Cut [0,K] into blocks of length P. On a full block (e0,e0+P] both ∑η(e)f(e) and β∫f differ from Pβf(e0) by at most P times the oscillation of f on the block; the oscillations add up to at most V, and the last, incomplete block costs at most 2PM.) For fK we have M,V=O(K2), so the W part is at most
since αj(ζ)=1 means ζ2j+1=(−1)j+1, which has 2j+1 solutions. □
Remark 4.7. The following is a sketch and is not used in the proof. Using ∑n≤yϕE(n)=3y2/π2+O(ylogy) in residue classes modulo 8 and the exact kmin, the same bound holds with c∗ replaced by (12I+80/27)/π2≤1.130790, where I=∫01ℓ(x)xdx∈[0.6822834,0.6831236] and ℓ is the piecewise linear function with ℓ(1/m)=c(m)/m. Numerically, ∑dbK,dϕE(d)/K3 is 1.19 at K=20 and 1.16 at K=40, decreasing towards 1.1308.
Smallness at the conjugate
Proposition 5.1.log∣σΔK(−Ξ)∣≤−K2(K+1)logφ+Klog(KMK), where MK=C0(K2/((1−t3)(1−t2))+φ2) and C0=∏i≥1(1−t2i)−1=2.0770906….
Proof. At τ=−φ the data become: w↦σw=w−1, λa↦σλa=2ta, Sa↦σSa=−Sa (as σψN=−ψN) and νk↦σνk=−k(k+1)/(1+αk) (as σαk=−1/αk). Let LXσ be the corresponding formal functional: LXσ(sk)=σνk and LXσ(1/(1−w−as))=2ta(X+Sa). The conjugate poles are ra=wa, ∣ra∣=t2a, and σDK(s)=∏a(1−w−as)=(−1)Kw−AKFσ(s) with Fσ(s)=∏a(s−ra). Hence
Partial fractions give sk/Fσ=∑j=0k−Khk−K−j(r)sj+∑arak/((Fσ)′(ra)(s−ra)). The pole moments at X=−Ξ. Since 1/(1−w−as)=−ra/(s−ra), we have LXσ(1/(s−ra))=−ra−12ta(X+Sa), and at X=−Ξ this is 2tara−1(Ξ−Sa)=2tara−1∑N>aψN. Writing N=a+M, t3N=t3at3M, t4a=ra2 and wa+M=rawM, this equals Γ(ra), where
Bounds. As hm(r1,…,rK)=wmhm(1,w,…,wK−1) and the coefficients of ∏i≥0(1−∣w∣iu)−1=∑mum∏i=1m(1−t2i)−1 are at most C0, we get ∣hm(r)∣≤C0t2m. For j≥1, ∣σνj∣≤j(j+1)/(1−t3), so the first sum is at most C0K2/((1−t3)(1−t2)) for k≤2K−2. For the second, ∑nn(n+1)t2n+1=2t3/(1−t2)3=2 (as 1−t2=t) and 1/(1−t3)=φ2/2, so ∑n∣γn∣≤φ2. Hence ∣Yk(−Ξ)∣≤MK for 0≤k≤2K−2. Hadamard’s inequality and ∣w∣AKK=φ−K2(K+1) finish the proof. □
Remark 5.2. Coefficientwise, the factor 2ta=2∣ra∣1/2 is not analytic in ra; only at the point X=σΞ=−Ξ does it combine with the tail ∑N>aψN≈t3a into ra2 times an analytic function. This is why the conjugate point, rather than the coefficients of σΔK, carries the gain.
Nonvanishing
Proposition 6.1.ΔK(Ξ)=0 for infinitely many K.
Proof. Let C(z)=L(1/(z−s))=∑nρn2wnz/(z−wn)3 for ∣z∣>t2. It continues to C∖{0} with a pole of order exactly 3 at each wn (leading coefficient 2ρnw2n=0), so it is not rational; it is analytic at ∞ and O(1/z). For degQ≤n−1 put Hn,Q(z)=L(Q(s)/((z−s)Dn(s))), analytic for ∣z∣>t2 and in particular at every xa, and
Tn,Q(z)=L((z−s)Dn(s)Q(z)Dn(s)−Dn(z)Q(s)),
a polynomial in z of degree ≤n−1. Then CQ−Tn,Q=DnHn,Q. If Δn(Ξ)=det[L(si+j/Dn)]=0, a kernel vector gives Qn=0 of degree ≤n−1 with L(sjQn/Dn)=0 for j<n. Put Tn=Tn,Qn and En=CQn−Tn. For ∣z∣>t2+ϵ the series (z−s)−1=∑jsjz−j−1 converges, together with its first two s-derivatives, uniformly on ∣s∣≤t2, so L may be applied termwise; hence Hn,Qn(z)=∑j≥nL(sjQn/Dn)z−j−1=O(z−n−1). So En=O(z−1) and En(xa)=0 for a≤n. If Δn(Ξ)=Δn+1(Ξ)=0, then V=Tn+1Qn−TnQn+1=Qn+1En−QnEn+1 is a polynomial, O(z−n−1), vanishing at x1,…,xn; so V=0 and Tn/Qn=Tn+1/Qn+1. If Δn(Ξ)=0 for all n≥N, all Tn/Qn equal one reduced fraction T/Q, so Qn=UnQ and Tn=UnT, and E:=CQ−T=0 since C is not rational. Then UnE=En=O(1/z) forces E=c0z−q(1+o(1)) with c0=0, q≥1, and degUn≤q−1. But E(xa)=0 for all large a, while En(xa)=0; so Un(xa)=0 for all large a≤n, too many zeros for a polynomial of degree <q once n is large. □
Proof of the theorem
Suppose ζF(3)=u/v with u∈Z, v∈Z>0. Since 5=1+2t, (1.1) gives Ξ=u(1+2t)/(25v) and σΞ=−Ξ. For K in the infinite set of Proposition 6.1 put yK=CK(t)(25v)KΔK(Ξ). By Proposition 4.4, yK∈O, and yK=0 because ΔK(Ξ)=0 and CK(t)=0 (t is not a root of unity, so Φd(t)=0). Moreover σyK=CK(−φ)(25v)KσΔK(−Ξ), so ∣N(yK)∣=∣yKσyK∣≥1, and by Propositions 3.1, 5.1 and 4.6
Since c∗−911=811β7−98=−0.02684…<0, this fails for large K. □
Remark 7.1. Not used in the proof. With the sharp constant of Remark 4.7 the margin is 911−1.130790≥0.0914 per K3logφ. At finite K, computed with exact determinants and interval arithmetic, the sum log∣NCK(t)∣+log∣ΔK(Ξ)∣+log∣σΔK(−Ξ)∣ (without the term 2Klog(25v)) is already negative, about −0.12K3logφ for 8≤K≤26; and the local bounds of Section 4 agree with the exact multiplicities of the cyclotomic factors in Q(τ)[X] for small K. The scripts are available from the author.
Formal verification
The Lean development is the repository https://github.com/telaviv/fibonacci-zeta3-irrationality; all file names below refer to it. It (library FibZeta, Lean 4.35.0, Mathlib v4.35.0, about 8500 lines) proves
where zetaF3 is the series ∑n≥01/Fn+13, written with Mathlib’s Nat.fib and a tsum (FibZeta/Final.lean; hasSum_zetaF3 shows that the series converges to it). The proof uses only the axioms propext, Classical.choice and Quot.sound (Axioms.lean); there is no sorry, no native_decide and no added axiom.
What has to be trusted. For the theorem itself, only the definition of zetaF3 (and Mathlib’s Nat.fib, tsum and Irrational) has to be read: the kernel checks everything else. Since a divergent tsum has the value 0, which is rational, the statement cannot hold for a degenerate reason. To see that the formal proof follows the argument of this paper one also reads the definitions in FibZeta/Defs.lean and the intermediate statements in FibZeta/Statements.lean; an internal comparison of these with an earlier draft of this section is in the repository (FIDELITY_REFEREE.md).
Definitions
FibZeta/Defs.lean defines the construction of Section 2 over an arbitrary field T with a parameter τ∈T: alpha, nu, psi, S, the poles node =w−a, the monic polynomial F=∏a(s−xa) (so DK=(−1)KwμKAKF), the residues res, the two coefficients mConst, mLin of LX(sk/F), the normalisation nrm, and the determinant Delta. Specialised at τ=t (tR) and at τ=−φ it gives ΔK and σΔK; at τ=RatFunc.X it gives ΔK,τ. The real functional is Lfun, with θ(θ+1)g=s2g′′+2sg′ written with deriv, and XiR is Ξ. The clearing polynomial is CKpoly =∏d≤4KΦdbK,d, with exponents nExp = nW + nAlpha: the W part (cluster costs cW, cW’ and cluster sizes clusterSize) and the α part. The root types are defined through the chosen root exp(2πi/d) (zetaC, eW, IsAlphaRoot), and FibZeta/Roots.lean proves that they are the same for every primitive d-th root.
The chain
FibZeta/Statements.lean states the ingredients as propositions, and FibZeta.irrational_zetaF3_of (in FibZeta/Main.lean) derives the theorem from them by the argument of Section 7. Instead of norms in O it uses that t and −φ are the two roots of τ2+τ−1: for Q∈Z[X] the product Q(t)Q(−φ) is an integer, nonzero if Q(t)=0 (norm_ne_zero). The docstrings of the Lean files refer to the numbering of the author’s notes; the header of FibZeta/Statements.lean translates it to the numbering of this paper.
Paper
Lean statement
Lean proof
Lemma 2.2, entries
EntryIdentity
FibZeta.Analytic.entryIdentity
Prop. 3.1
DecayBound
FibZeta.Analytic.decayBound
Lemma 4.1–Prop. 4.4
Integrality
FibZeta.Denom.integrality
Lemma 4.5, Prop. 4.6
CostBound
FibZeta.Cost.costBound
Prop. 5.1
ConjBound
FibZeta.Conj.conjBound
Prop. 6.1
Nonvanishing
FibZeta.NonVan.nonvanishing
Section 7
FibZeta.irrational_zetaF3_of
Table 1.
EntryIdentity is used only in the proof of DecayBound, not by the assembly. The formal statements are equivalent to, or slightly weaker than, the propositions, in the form the assembly needs. DecayBound is ∣ΔK(Ξ)∣≤CK2t2bK with bK=2∑j<Kaj (proved with C=4096). ConjBound is ∣σΔK(−Ξ)∣≤CK2tK2(K+1); its proof bounds the complete homogeneous sums hm(r) by φK instead of C0t2m, which still gives entries of size ≤DK. CostBound is log∣CK(t)CK(−φ)∣≤100121K3logφ+CK2(1+logK); its proof follows Proposition 4.6, with cubic constant c∗=36754393 and the constant 12 in place of 2E1 in the upper half of Lemma 4.5, and c∗≤100121<911. Integrality is Proposition 4.4 for every real τ=0 that is not a root of unity, with one integer polynomial for all such τ; this covers t and −φ at once. Its proof works in the local rings of C(τ) at the roots of unity: the Newton cluster basis at W-roots, the monomial basis elsewhere, and the rank correction at α-roots (FibZeta.Denom.local_bound_cluster, FibZeta.Denom.local_bound_monomial, FibZeta.Denom.integrality_of_local_bounds). The proof of Nonvanishing uses the Cauchy transform C directly: its values at the interpolation nodes, C(xa)=2ta(Ξ−Sa), and its expansion C(1/u)=∑kνkuk+1 at ∞ replace the functional L in the interpolation identities.
Provenance
Parts of the formal development are adapted from the Lean formalisation QZeta3[13] of [14] (Apache License 2.0): the localisation, specialisation and descent machinery, the weighted and rank determinant lemmas, the Cauchy-transform and normality arguments for nonvanishing, the decay estimates, the totient sieve and the cyclotomic size bounds. Every adapted file says so in its header, and the repository’s NOTICE file records it.
Use of AI tools
This work was carried out with substantial assistance from AI systems (Anthropic’s Claude models), which were used to search the literature, to develop and check the proof, to write the Lean formalisation and to draft this text. The author directed the project and is responsible for its content. The formal proof is checked by the Lean kernel.
References
[1]R. André-Jeannin, Irrationalité de la somme des inverses de certaines suites récurrentes, C. R. Acad. Sci. Paris Sér. I Math. 308 (1989), 539–541.
[2]D. Duverney, Ke. Nishioka, Ku. Nishioka, and I. Shiokawa, Transcendence of Rogers–Ramanujan continued fraction and reciprocal sums of Fibonacci numbers, Proc. Japan Acad. Ser. A 73 (1997), 140–142.DOI
[3]C. Elsner, S. Shimomura, and I. Shiokawa, Algebraic relations for reciprocal sums of Fibonacci numbers, Acta Arith. 130 (2007), 37–60.DOI
[4]————, Algebraic relations for reciprocal sums of odd terms in Fibonacci numbers, Ramanujan J. 17 (2008), 429–446.DOI
[5]A. Faunzan, ζ(5) is irrational, unrefereed preprint, 17 September 2026, Zenodo record 22826419.
[6]S. Fischler and W. Zudilin, A refinement of Nesterenko’s linear independence criterion with applications to zeta values, Math. Ann. 347 (2010).
[7]F. Jouhet and E. Mosaki, Irrationalité aux entiers impairs positifs d’un q-analogue de la fonction zêta de Riemann, Int. J. Number Theory 6 (2010), 959–988.arxiv.org/abs/0712.1762
[8]C. Krattenthaler, A determinant identity for moments of orthogonal polynomials that implies Uvarov’s formula for the orthogonal polynomials of rationally related densities, 2021, preprint, arXiv:2103.03969.arxiv.org/abs/2103.03969
[9]C. Krattenthaler, T. Rivoal, and W. Zudilin, Séries hypergéométriques basiques, q-analogues des valeurs de la fonction zêta et séries d’Eisenstein, J. Inst. Math. Jussieu 5 (2006), 53–79.DOI
[10]N. K. Meher and D. K. Sahoo, Analytic continuation of the ℓ-generalized Fibonacci zeta function, Integers 24 (2024), #A85.DOI
[11]Yu. V. Nesterenko, Modular functions and transcendence questions, Sb. Math. 187 (1996), 1319–1348.DOI
[12]M. Prévost, A new proof of the irrationality of ζ(2) and ζ(3) using Padé approximants, J. Comput. Appl. Math. 67 (1996), 219–235.
[13]F. Solomon Admasu, QZeta3: Lean 4 formalisation of the irrationality of ζ₁/ₚ(3) (moving-denominator Hankel argument), version 0.1.0, 2026-10-11, Apache License 2.0; directory lean/ of [14].
[14]uplift-universal-inc, Irrationality of the Lambert-series value ζ₁/ₚ(3), unrefereed AI-assisted research draft with Lean formalisation, version 0.3.0-two-proofs, 2026-10-12, https://github.com/uplift-universal-inc/qzeta_3-irrationality (commit d12d56f).
[15]V. B. Uvarov, The connection between systems of polynomials that are orthogonal with respect to different distribution functions, U.S.S.R. Comput. Math. Math. Phys. 9 (1969), no. 6, 25–36.DOI
[16]M. Waldschmidt, Recent Diophantine results on zeta values: a survey, RIMS Kôkyûroku 1710 (2010), 166–179.
[17]W. Zudilin, On the irrationality of generalized q-logarithm, Res. Number Theory 2 (2016), article 15.DOI
[18]———, A determinantal approach to irrationality, Constr. Approx. 45 (2017), 301–310.DOI