Let a≥3, let U0=0, U1=1, Un+1=aUn−Un−1, and let δ=±1. We prove that 1 and the numbers ∑n≥1δn/Unl, l=1,3,…,A−1, span a Q-vector space of dimension at least (π/12+π2+o(1))A. So infinitely many of these sums with l odd are irrational, for instance infinitely many of ∑n1/F2nl. With the known linear independence of 1, ∑nδn/Un and ∑nδn/Un3, one of ∑nδn/Unl, l=5,7,…,23, lies outside the span of these three numbers. The proof adapts the author's construction for odd q-zeta values at real quadratic units, with a weighted bilateral sum. No single sum such as ∑n1/Un5 is decided, and sequences of norm −1, such as the Fibonacci numbers, are not covered.
Introduction
Fix an integer a≥3 and let U0=0, U1=1, Un+1=aUn−Un−1. Then Un=(ηn−η−n)/(η−η−1) with η=(a+a2−4)/2>1, a unit of norm +1 of K=Q(a2−4), and every unit η>1 of norm +1 of a real quadratic field arises in this way, with a=η+η−1. Examples are Un=F2n (a=3), Un=Fkn/Fk for even k (a=Lk) and Un=P2n/2, where Pn are the Pell numbers (a=6). Let δ∈{1,−1} and, for even A≥4 and 1≤r<A/2,
τ(A,r)=12A+π2(A+4r2)4π2r(A−2r)−12A.
Theorem 1.1. For every even A≥4 and every integer r with 1≤r<A/2,
dimQQ+1≤l≤A−1lodd∑Qn≥1∑Unlδn≥1+τ(A,r).
As A→∞ through even integers, this dimension is at least (π/12+π2+o(1))A=(0.67178…+o(1))A. In particular, at least ⌈1+τ(A,r)⌉−1 of the numbers ∑nδn/Unl with odd l≤A−1 are irrational, and infinitely many of the numbers ∑nδn/Unl, l odd, are irrational.
Key words and phrases. Lucas sequences, Fibonacci numbers, reciprocal sums, Lambert series, irrationality, linear independence, Nesterenko’s criterion.
The function τ is that of the author’s dimension bound for odd q-zeta values at real quadratic units [12 Theorem 1.1]. The bound depends neither on a nor on δ.
Corollary 1.2. Let Ak be the smallest even A with maxr(1+τ(A,r))>k−1, so that Theorem 1.1 gives dim≥k for the sums with odd l≤Ak−1. For 2≤k≤11 the values are those of Table 1, attained at r=k−1; at every even A<Ak and every r, 1+τ(A,r)≤k−1.
k
Ak
r
1+τ(Ak,r)
k
Ak
r
1+τ(Ak,r)
2
4
1
1.2438
7
88
6
6.0650
3
12
2
2.1601
8
118
7
7.0518
4
24
3
3.0948
9
152
8
8.0296
5
40
4
4.0358
10
190
9
9.0012
6
62
5
5.0635
11
234
10
10.0097
Table 1. Thresholds of Corollary 1.2.
For the first new cases we use the following result, proved with a Lean formalisation in [13]: the numbers 1, ∑nδn/Un and ∑nδn/Un3 are linearly independent over Q. Let V3 be their Q-span, so dimQV3=3.
Corollary 1.3. (1) At least one of the numbers ∑nδn/Unl with l∈{5,7,…,23} lies outside V3; in particular it is irrational. (2) At least two of the numbers ∑nδn/Unl, l∈{5,7,…,39}, are linearly independent modulo V3. (3) In general, the ∑nδn/Unl with l odd, 5≤l≤Ak−1, span a space of dimension at least k−3 modulo V3.
At A=12 Theorem 1.1 gives dim≥3, which is contained in [13]; part (1), from A=24, is the first finite statement that goes beyond it.
Corollary 1.4. Theorem 1.1 and Corollaries 1.2 and 1.3 hold for ∑n1/F2n and ∑n(−1)n/F2n, for ∑nδn/Fkn with k≥2 even, and for ∑nδn/P2n. For instance, one of ∑n1/F2n, l∈{5,7,…,23}, lies outside Q+Q∑n1/F2n+Q∑n1/F3n.
Theorem 1.5. Let d≥2 be squarefree, K=Q(d), and let ν∈OK× have norm +1 and 0<∣ν∣<1. For every even A≥4 and every integer r with 1≤r<A/2,
dimQQ+1≤j≤A−1jodd∑QdWj(ν)≥1+τ(A,r).
Theorems 1.1 and 1.5 are equivalent by an elementary change of basis (Lemma 2.1, Remark 2.2). Theorem 1.5 does not follow from [12 Theorem 1.1], which concerns 1,dζq(3),…,dζq(A−1) at one unit q: the Wj(ν) mix the units ν and ν2, and j=1 occurs.
The method
In [12] the rational function Rn(T;q) of Krattenthaler, Rivoal and Zudilin [11 §3] is summed over qk and q−n−k, k≥1, with opposite signs. This gives linear forms in the odd ζq(l) whose coefficients are invariant or anti-invariant under q↦1/q, up to a power of q. At a unit of norm +1, where Galois conjugation is q↦1/q, integer forms in the numbers dζq(l) are then small at both real places, and Nesterenko’s criterion applies. We keep the rational function and its partial-fraction coefficients, at q=ν2, and change only the sum:
Sn(ν)=k≥1∑νkR(ν2k)−k≥1∑ν−n−kR(ν−2n−2k).
The weights turn the q-zeta values into the Wj(ν), the minus sign removes the even j, and j=1 survives (Lemma 4.1); the reciprocity under ν↦1/ν holds for all odd j, j=1 included (Lemma 4.2). Coefficients, decay and denominators change only by factors eO(n), so τ is unchanged.
We cite [12] for the properties of Rn and prove what is new. Nothing is formalised. Besides [12] the proofs use Nesterenko’s criterion [16, 8, 11] and the asymptotic ∑ℓ≤nϕE(ℓ)=3n2/π2+O(nlogn)[9 Theorem 330]; Corollary 1.3 also uses [13].
The change of basis
Lemma 2.1. Let η>1 be a unit of norm +1 of a real quadratic field, with trace a and Lucas sequence (Un), let δ∈{1,−1}, ν=δ/η, and let d be the squarefree part of a2−4. For odd l≥1,
where Pl(m)=∏i(m2−i2)/(2l−1(l−1)!), the product over odd i with 1≤i≤l−2. Hence Vl=∑j≤ljoddal,jWj(ν), over odd j≤l, with al,j=[mj−1]Pl∈Q and al,l=0; and for every odd L≥1,
Proof. Since 1/Unl=(η−η−1)lη−nl(1−η−2n)−l and l is odd, δnη−nl=νnl and η−2n=ν2n. Expanding (1−ν2n)−l and summing over n (absolutely convergent), Vl=∑i≥0(l−1i+l−1)νl+2i/(1−νl+2i). With m=l+2i, (l−1(m−l)/2+l−1)=∏t=1l−1(m−l+2t)/(2l−1(l−1)!)=Pl(m), since the factors m−l+2t are the numbers m±i, i odd, i≤l−2. As Pl vanishes at m=1,3,…,l−2, the sum may run over all odd m. Pl is even of degree l−1 with leading coefficient 1/(2l−1(l−1)!), which gives the triangular expansion. Finally (η−η−1)2=a2−4, so (η−η−1)l=(a2−4)(l−1)/2fd with f∈Q×. □
For example V3=(W3−W1)/8 and V5=(9W1−10W3+W5)/384.
Remark 2.2. By Lemma 2.1, Theorem 1.1 for (a,δ) is Theorem 1.5 at ν=δ/η. Conversely, for ν as in Theorem 1.5, η=1/∣ν∣>1 is a unit of norm +1, its trace a=η+η−1 is an integer >2, and ν=δ/η with δ the sign of ν. Since η∈K∖Q, Q(a2−4)=K, so the squarefree d of Lemma 2.1 is that of Theorem 1.5.
What is used from [Kri26c]
From now on A≥4 is even, 1≤r<A/2, n≥1 and e=(A−2r)n/2∈Z. Write (x;q)m=∏i=0m−1(1−xqi), Φℓ for the ℓ-th cyclotomic polynomial, dn(q)=∏ℓ≤nΦℓ(q), ϕE for Euler’s function and T(n)=∑ℓ≤nϕE(ℓ). Let
be the rational function [12 (2.1)], with partial fractions [12 (3.1)]
R(T)=j=0∑ns=1∑A(1−qjT)scs,j(q),cs,j∈Q(q).
We use the following statements of [12] verbatim. Items (a)–(c) are identities for an indeterminate q; (d) and (e) hold for every complex q with 0<∣q∣<1, where κ=∣q∣; (f) holds for every unit q of norm +1 of a real quadratic field with 0<∣q∣<1.
(f) [12 Lemma 7.1, Remark 7.2, and §9 before (9.2)]: σ(dn(q))=dn(1/q)=−q−T(n)dn(q), log∣dn(q)∣=O(n), and T(n)=3n2/π2+O(nlogn)[9 Theorem 330].
We also use the trace argument of [12 Lemma 9.3(a)]: TrK/Q(x/d)∈Z for x∈OK, d squarefree; and Nesterenko’s criterion as stated in [12 §9, proof of Theorem 1.1], from [16], [8 Theorem A], [11 Proposition 2]: if ϑ1,…,ϑN are real and there are integers pj,n with n21log∑jpj,nϑj→α<0 and limsupnn21log∣pj,n∣≤β for every j, then dimQ∑jQϑj≥1−α/β.
We apply all of this at q=ν2 with ν a unit of norm +1, 0<∣ν∣<1; then q is a unit of norm +1 with 0<q<1 and Q(q)=K. The sums Sn and coefficients Pl, P0 of [12] are not used; ours are defined below. (In [12 §9] the letter ν denotes log(1/∣q∣); here we write λ=log(1/κ).)
Let Rn=Q[ν,ν−1,dn(ν2)−1] for an indeterminate ν. By (c), cs,j(ν2)∈Rn. Since the identity dn(1/q)=−q−T(n)dn(q) of (f) holds in Q(q) (see also [12 §4]), Rn is stable under ν↦1/ν, and evaluation at ν∈C with 0<∣ν∣=1 is a ring homomorphism Rn→C, since the roots of dn(ν2) are roots of unity.
The weighted bilateral forms
Let ν∈C, 0<∣ν∣<1, q=ν2 and κ=∣q∣, and let Sn=Sn(ν) be as in the introduction. No argument is a pole, and by (a), ν−n−kR(q−n−k)=ν−n−kqA(n/2+k)R(qk)=νkν(A−1)(n+2k)R(qk), so by (d) both series converge absolutely. For s≥1 let ps(μ)=∏t=1s−1(μ+2t−1)/(2s−1(s−1)!), so that ps(2i+1)=(s−1i+s−1); ps has degree s−1, nonnegative coefficients, and ps(1)=1. For odd l with 1≤l≤A−1 put, with cs,j=cs,j(ν2) and sums over 1≤s≤A, 0≤j≤n,
and ps(μ)+ps(−μ)=2∑λeven[μλ]ps⋅μλ with λ≤s−1≤A−1, so λ≤A−2. Summing over μ gives ∑lPlWl with l=λ+1. All rearrangements are of absolutely convergent double series. □
No relation such as ∑s,jcs,j=0 is used, since the weights make every series converge. The term λ=0 gives W1; a one-sided weighted sum would keep the even W’s as well.
Lemma 4.2 (reciprocity). In Rn: Pl(1/ν)=νn−2ePl(ν) for every odd l, l=1 included, and P0(1/ν)=−νn−2eP0(ν).
Proof. The substitution ν↦1/ν is q↦1/q, so cs,j(ν−2)=ν−2ecs,n−j(ν2) by (b). With j′=n−j,
If ν is a unit of norm +1 in K, then σν=1/ν, so σ(f(ν))=f(1/ν) for f∈Rn, and
σ(Pl(ν))=νn−2ePl(ν),σ(P0(ν))=−νn−2eP0(ν).(2)
Heuristically, Wj(1/ν)=−Wj(ν)−(1−2j−1)ζ(1−j) formally, and the constant vanishes for every odd j, including j=1 (where $1-2^0=0): all Wj behave as anti-invariant, which is why dWj is the right twist.
Proof. Write dn=dn(ν2). Since 2s−1(s−1)!ps∈Z[μ] and 2s−1(s−1)! divides FA for s≤A, we have FA[μl−1]ps∈Z, and [μl−1]ps=0 unless λ≤s−1. So only s≥l contributes to Pl, and then FAdnA−ldn,cs,jν−j[μl−1]ps lies in Z[ν±1] by (c). For P0: 1−qt divides dn for 1≤t≤n by (c), and (1−q−t)−s=(−qt)s(1−qt)−s, so dnsBs,j∈Z[ν±1], and dnAcs,jν−jBs,j=(dnA−scs,j)(dnsBs,j)ν−j.
Coefficients and decay
Lemma 5.1 (coefficients). Letν∈C, 0<∣ν∣<1, κ=∣ν∣2, ρ=(1−κ)/2 and Γn=∑j=0n∑s=1A∣ν∣−j∣cs,j(ν2)∣. For every oddlwith1≤l≤A−1, ∣Pl(ν)∣≤2Γnand∣P0(ν)∣≤n∣ν∣(1−κ)−AΓn, and
Proof. Apply (e) at q=ν2. Since ρ,κ<1 and s≥1, ρ−A(j+1)≤ρ−A(n+1) and κ−E(j)≤κ−(A+4r2)n2/8−rn/2; with ∣ν∣−j≤∣ν∣−n and (n+1)A terms this gives the bound for Γn. The coefficients of ps are nonnegative with sum ps(1)=1, so ∣[μl−1]ps∣≤1 and ∣Pl∣≤2Γn. For 1≤t≤n, ∣νt(1−qt)−s∣≤∣ν∣(1−κ)−A and ∣ν−t(1−q−t)−s∣=∣ν∣t(2s−1)∣1−qt∣−s≤∣ν∣(1−κ)−A; as Bs,j has n terms, this gives the bound for P0. All factors except κ−(A+4r2)n2/8 are eO(n), since ∣ν∣−n=κ−n/2.
Lemma 5.2 (decay). Fixν∈Cwith0<∣ν∣<1. For all largen, Sn(ν)=0and
Proof. By (d) and the computation at the start of § 4, Sn=∑k≥rn+1νk(1−ν(A−1)(n+2k))R(qk). For large n, ∣ν∣(A−1)(n+2rn+2)≤21, so each factor 1−ν(A−1)(n+2k) has modulus in [21,23], and ∣νkR(qk)∣∈[C−1,C]κk(e+1/2) by (d). So the term k=rn+1 has modulus at least (2C)−1κ(rn+1)(e+1/2), and the others have total modulus at most 23Cκ(rn+2)(e+1/2)/(1−κe+1/2). The ratio of these bounds tends to 0 as e→∞, so ∣Sn∣∈[(4C)−1,4C]κ(rn+1)(e+1/2) for large n; and (rn+1)(e+21)=21r(A−2r)n2+O(n). □
Proof of the theorems
Let d, K, ν be as in Theorem 1.5, q=κ=ν2, λ=log(1/κ), and Dn=FAdn(q)A∈Z[ν±1]⊂OK as in Lemma 4.3. By (f), ∣σDn∣=κ−AT(n)∣Dn∣ and log∣Dn∣=O(n), so
log∣NDn∣=log∣DnσDn∣=π23An2λ+O(nlogn).(4)
Balancing (as in [12 (9.3)]). Put ϱn=∣ν∣n−2e∣σDn∣/∣Dn∣, and let m=mn∈Z be the unique integer with
2≤∣ν∣2mϱn<2∣ν∣−2.(5)
Put Dn′=ν−mDn∈OK and Mn=∣ν∣n−2e∣σDn′∣. Since σν=1/ν, Mn/∣Dn′∣=∣ν∣2mϱn∈[2,2∣ν∣−2); so ∣Dn′∣≤Mn/2 and Mn2=(Mn/∣Dn′∣)∣ν∣n−2e∣NDn∣, whence by (4)
logMn=2π23An2λ+O(nlogn).(6)
For odd l with 1≤l≤A−1 let ℓ0,n=TrK/Q(Dn′P0(ν)), ℓl,n=TrK/Q(Dn′Pl(ν)/d) and Ln=ℓ0,n+∑lℓl,ndWl(ν).
Lemma 6.1. (a) ℓ0,n,ℓl,n∈Z. (b) Ln=(Dn′−νn−2eσDn′)Sn(ν). (c) For large n, Ln=0 and
Proof. (a) By Lemma 4.3, Dn′P0(ν),Dn′Pl(ν)∈OK; use Tr(OK)⊂Z and the trace argument of Section 3. (b) Write Tr(z)=z+σz. The terms z add up to Dn′Sn by Lemma 4.1. By (2) and σd=−d, the conjugate terms are σ(Dn′P0)=−νn−2eσ(Dn′)P0 and σ(Dn′Pl/d)dWl=−νn−2eσ(Dn′)PlWl, which add up to −νn−2eσ(Dn′)Sn. (c) ∣Dn′−νn−2eσDn′∣∈[Mn−∣Dn′∣,Mn+∣Dn′∣]⊂[21Mn,23Mn]; use Lemma 5.2 and (6). (d) ∣ℓ0,n∣≤(∣Dn′∣+Mn)∣P0∣≤23Mn∣P0∣, and likewise ∣ℓl,n∣≤23Mn∣Pl∣/d. Use (3) and (6). □
Proof of Theorem 1.5. Apply Nesterenko’s criterion to the A/2+1 numbers 1 and dWj(ν), j=1,3,…,A−1, and to the forms Ln, which are defined for every n≥1. By Lemma 6.1(c), (d), whose first estimate is two-sided because Lemma 5.2 and the asymptotic in (f) are,
α=(2π23A−2r(A−2r))λ,β=(2π23A+8A+4r2)λ>0.
If α≥0 then τ(A,r)≤0 and there is nothing to prove. Otherwise the dimension is at least
Proof of Theorem 1.1. By Lemma 2.1 at ν=δ/η, the span in Theorem 1.1 is that of Theorem 1.5, with d the squarefree part of a2−4. The asymptotic is [12 proof of Corollary 1.2]: with t=12+π2/(2π) and r=⌈tA⌉ (admissible for A≥6), τ(A,r)/A→4π2t/(12+π2+4π2t2)=π/12+π2. If at most u of the sums with odd l≤A−1 were irrational, the span would have dimension at most u+1; this gives the count, and maxrτ(A,r)→∞.
Proof of Corollaries 1.3 and 1.4. Let V be the span of Theorem 1.1 at A=Ak, so dimV≥k by Corollary 1.2. As V is V3 plus the span of the sums with 5≤l≤Ak−1, and dimV3=3 by [13 Theorem 1.1] (its case (P,Q)=(a,1), for both δ; in its Lean statement (a,E)=(a,−1)), dimV/V3≥k−3; (1) and (2) are k=4 and k=5. For Corollary 1.4: F2n is the sequence with a=3; for even k, Fkn=FkUn with η=ϕk and a=Lk, so ∑nδn/Fknl=Fk−l∑nδn/Unl; and P2n=2Un with a=6.
Thresholds and computations
Proof of Corollary 1.2. Since 12A+π2(A+4r2)>0, the inequality 1+τ(A,r)>k−1 is equivalent to
π2[4r(A−2r)−(k−2)(A+4r2)]>12A(k−1).
For 2≤k≤11, every even A≤Ak and every 1≤r<A/2, this inequality was decided in exact rational arithmetic from 3.14159265358979<π<3.14159265358980 (no case was left undecided), with the results stated; for example maxr(1+τ(22,r))=2.9499<3.
From k=7 (A7=88) on, two values of r reach the threshold at Ak (at A=234, r=11 gives 10.0208), and A12=282. The ratio maxr(1+τ(A,r))/A is 0.6475, 0.6635, 0.6691, 0.6709 at A=102,…,105.
Computations
Computations. The ancillary files contain scripts (python-flint, mpmath) and logs in two independent sets, the author’s (anc/) and those of the AI referee pass described under Provenance, written without the author’s code (anc/referee/). Only the threshold certification (manyu_opt.py, Arb balls; referee/ref_opt.py, exact rationals) enters a proof. The other checks confirm Lemma 2.1 (a≤7, l≤23), Lemmas 4.2 and 4.3 exactly in Q(ν) (35 triples (A,r,n), A≤24), and, at five units such as ±φ−2 and 2−3 with (A,r) up to (24,3) (120 rows, 8000 bits), the integrality of the ℓ’s, Lemmas 4.1, 5.1, 5.2 and 6.1(b); the referee repeated the last group in 44 rows and computed the denominator profiles of Remark 8.1(c).
Remarks and limits
Remark 8.1. (a) Norm −1. For Fn, Pn, and Fkn with k odd, the unit has norm −1, Galois conjugation acts as ν↦−1/ν, and Lemma 4.2 gives no relation between σ(Pl(ν)) and Pl(ν). These sequences are not covered; for them the results we know concern l≤3[1, 15, 14, 19, 13].
(b) Single values. The theorems decide no individual sum such as ∑n1/Un5, and do not say which l give irrational values. We see no way to eliminate W1 and W3 from the forms Ln without multiplying their coefficients by factors ecn2 with c>0.
(c) Sharper thresholds. At q=1/p the bounds of [11] were improved by saving one power of dn[10] and by a refinement of Nesterenko’s criterion [8]; we have not adapted either. A full power of dn(ν2) cannot be saved in Lemma 4.3: in every case we computed exactly (in Q(ν), A≤24) the denominator of P0 contains Φ2n(ν)A, so dn(ν2)A−1P0∈/Z[ν±1] there. In these cases the exponent A occurs only at Φ2ℓ(ν) with ℓ in roughly (2n/3,n] and at a few small indices, so a partial saving is not excluded.
(d) Odd A. We have not proved a version of the construction for odd A (then e is a half-integer for odd n). Such a version, with the same function τ, would give nothing new. For fixed r≥1,
τ(A,r)=(12+π2)A+4π2r2(4π2r−12)A−8π2r2
is a Möbius function of A whose denominator is positive for A>0 and whose determinant (4π2r−12)⋅4π2r2+8π2r2(12+π2) is positive, so it is strictly increasing in A. Moreover r<A/2 implies r<(A+1)/2. Hence τ(A,r)<τ(A+1,r), and Theorem 1.5 at the even number A+1 concerns the same numbers dW1,dW3,…,dWA.
Related work
André-Jeannin proved that ∑nxn/wn is irrational for a class of binary recurrences that includes every Un(P,±1) with x=±1, in particular ∑nqn/Un∈/Q[1]. Tachiya [18] proved irrationality results for weight-one Lambert series at algebraic integers of Pisot type; Bundschuh and Väätänen [2] proved linear independence results for such series, with applications to reciprocal sums of Fibonacci and Lucas numbers, and [3] treats mixed weights one and two at rational bases; Duverney, Suzuki and Tachiya [5] proved linear independence results for reciprocal sums of Fibonacci and Lucas numbers along certain sequences of indices. For both signs of the norm the even powers ∑nUn−2s are transcendental [4], and their algebraic relations are known [6, 7]; these are values of quasi-modular forms, to which Nesterenko’s theorem [17] applies. Wang [24] expresses ∑nnpUn−s through derivatives of Eisenstein series when p≡s(mod2) (norm +1) and when p,s are even (norm −1); the sums ∑nUn−l, l odd, of this note are outside these parity classes, and we know of no such expression for them or for their alternating versions.
At q=1/p with p∈Z∖{0,±1}, Krattenthaler, Rivoal and Zudilin proved a lower bound of order A for the dimension of the span of 1,ζq(3),ζq(5),…,ζq(A−1)[11], refined in [10, 8]. The analogue at real quadratic units for dζq(l) is [12], of which this note is a variant.
For weight three, the author proved ∑n1/Fn3∈/Q and ∑n1/Un3∈/Q for all real quadratic units [15, 14], and the linear independence of 1,∑nδn/Un,∑n(δε)n/Un3, with ε the norm of the unit [13], all with Lean formalisations. The Hexagon contributor twinheterodox posted neighbouring results of weight at most three: the alternating sums ∑n(−1)n−1/Un3[19], the weight-three Lambert sum ∑nn2/(bn−1) at Pisot, Salem and other algebraic bases [22, 20], a weighted reciprocal Lucas sum [23] and the reciprocal Mersenne cubes [21]; these preprints are described by their author as primarily AI-generated, have no formal proofs, and were not checked by us. None of these works treats an odd power l≥5, and in our searches of Hexagon, arXiv, GitHub and Zenodo (November 2026) we found no earlier result on ∑nδn/Unl for an odd l≥5.
Provenance
The construction was found by varying the summation in [12 §9], choosing weights so that the resulting values are those of Lemma 2.1. The proof was checked line by line in a separate whole-chain referee pass, which rebuilt the exact computations of § 7 with independent code and found no error affecting a statement. Nothing in this note is formalised. Its external inputs are those of [12 Theorem 1.1] and the lemmas listed in § 3; the Lean-verified result of [13] enters only Corollary 1.3. The referee pass was itself carried out by an AI model (a separate Claude session, without access to the author’s code); it is not a peer review.
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 proofs, to run the computations and to draft this text. The author directed the project and is responsible for its content.
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), no. 19, 539–541.
[2]P. Bundschuh and K. Väanänen, Linear independence of certain Lambert and allied series, Acta Arith. 120 (2005), no. 2, 197–209.DOI
[3]———, Linear independence of q-analogues of certain classical constants, Results Math. 47 (2005), no. 1-2, 33–44.DOI
[4]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 Math. Sci. 73 (1997), no. 7, 140–142.DOI
[5]D. Duverney, Y. Suzuki, and Y. Tachi, Linear independence results for sums of reciprocals of Fibonacci and Lucas numbers, Acta Math. Hungar. 162 (2020), no. 1, 375–392, arXiv:1908.07290.arxiv.org/abs/1908.07290
[6]C. Elsner, S. Shimomura, and I. Shiokawa, Algebraic relations for reciprocal sums of Fibonacci numbers, Acta Arith. 130 (2007), 37–60.DOI
[7]———, Algebraic relations for reciprocal sums of even terms in Fibonacci numbers, J. Math. Sci. (N.Y.) 180 (2012), no. 5, 650–671, translated from Fundam. Prikl. Mat. 16 (2010), no. 5, 173–200.
[8]S. Fischler and W. Zudilin, A refinement of Nesterenko’s linear independence criterion with applications to zeta values, Math. Ann. 347 (2010), 739–763.
[9]G. H. Hardy and E. M. Wright, An introduction to the theory of numbers, sixth ed., Oxford University Press, Oxford, 2008, Revised by D. R. Heath-Brown and J. H. Silverman.
[10]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
[11]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), no. 1, 53–79, arXiv:math/0311033.DOI
[12]S. Krisman, Irrationality of odd q-zeta values at real quadratic units, Hexagon 2611.00693v1, https://hexagonmath.org/abs/2611.00693v1, 2026.
[13]———, Linear independence of reciprocal and reciprocal-cube sums of Lucas sequences, Hexagon preprint; Lean formalisation at https://github.com/telaviv/lucas-reciprocal-independence, 2026.hub.valency.io/works/W_dyu427fe
[14]———, The reciprocal cube sums of Lucas sequences of quadratic units are irrational, Hexagon 2611.00491v1, https://hexagonmath.org/abs/2611.00491v1; Lean formalisation at https://github.com/telaviv/lucas-zeta3-irrationality, 2026.
[15]———, The sum of the reciprocal cubes of the Fibonacci numbers is irrational, Hexagon 2611.00473v1, https://hexagonmath.org/abs/2611.00473v1; Lean formalisation at https://github.com/telaviv/fibonacci-zeta3-irrationality, 2026.
[16]Yu. V. Nesterenko, On the linear independence of numbers, Moscow Univ. Math. Bull. 40 (1985), no. 1, 69–74.
[17]———, Modular functions and transcendence questions, Sb. Math. 187 (1996), no. 9, 1319–1348.DOI
[18]Y. Tachiya, Irrationality of certain Lambert series, Tokyo J. Math. 27 (2004), 75–85.DOI
[19]twinheterodox, The alternating sum of reciprocal Fibonacci cubes is irrational, Hexagon 2611.00613v1, https://hexagonmath.org/abs/2611.00613v1, 2026, Described by its author as primarily AI-generated; no formal proof.
[20]———, A height criterion for irrational weight-three Lambert sums, Hexagon 2611.00638v1, https://hexagonmath.org/abs/2611.00638v1, 2026, Described by its author as primarily AI-generated; no formal proof.
[21]———, The sum of the reciprocal Mersenne cubes is irrational, Hexagon 2611.00643v1, https://hexagonmath.org/abs/2611.00643v1, 2026, Described by its author as primarily AI-generated; no formal proof.
[22]———, The weight-three Lambert sum at the golden ratio is irrational, Hexagon 2611.00623v1, https://hexagonmath.org/abs/2611.00623v1, 2026, Described by its author as primarily AI-generated; no formal proof.
[23]———, A weighted reciprocal Lucas sum lies outside ℚ(√5), Hexagon 2611.00639v1, https://hexagonmath.org/abs/2611.00639v1, 2026, Described by its author as primarily AI-generated; no formal proof.
[24]W. Wang, Evaluation of reciprocal sums of hyperbolic functions using quasimodular forms, J. Math. Anal. Appl. 549 (2025), no. 2, Paper No. 129526, arXiv:2312.16929.DOI