Let μ denote the irrationality exponent. Bruiltet proved in 2002 that μ(Γ(1/4))≤10143 and μ(Γ(1/3))≤10151. We prove μ(Γ(1/4))≤3.24⋅1026 and μ(Γ(1/3))≤1.74⋅1026, with explicit thresholds h(p/q)≥5.4⋅107 and h(p/q)≥4.7⋅107 respectively, against 1075 and 1079 in the published statements. The proof keeps the architecture of Bruiltet's argument – Philippon's transcendence criterion together with a zero estimate in commutative algebraic groups – and replaces three of its ingredients. First, an approximate Schwarz principle based on Hermite interpolation replaces the gravitation estimate of Reyssat; on a lattice point set its cost per node is governed by a lattice energy and drops from about 3.7 to about 1.0. Second, a summation lemma for the failure of Diaz's approximation lemma sharpens the passage from an approximation measure to an irrationality measure. Third, the auxiliary point is taken at a torsion point of order 4, respectively 3, which lowers the degree of the base field from 8d to 2d. All numerical constants are produced by a verifier run in interval arithmetic, supplied as ancillary files.
Introduction
For an irrational real number θ, the irrationality exponent μ(θ) is the supremum of the set of real μ such that 0<∣θ−p/q∣<max(∣p∣,∣q∣)−μ has infinitely many solutions. Finiteness of μ(Γ(1/4)) and μ(Γ(1/3)) follows from the algebraic independence of π/ω and η/ω for a lattice with algebraic invariants, due to Chudnovsky, together with the quantitative refinements of Philibert and Philippon. The first explicit values are due to Bruiltet [1]:
h(p/q)≥1075⟹Γ(1/4)−qp>(qe1)10143,
h(p/q)≥1079⟹Γ(1/3)−qp>(qe1)10151.
These remain, to our knowledge, the only published bounds for these two numbers; Cohen and Zudilin [2] obtain reasonable measures for Chowla–Selberg gamma quotients and describe the above as the state of the art for the individual values.
Our result is the following.
Theorem 1. For every rational p/q with h(p/q)≥5.4⋅107,
Γ(1/4)−qp>(qe1)3.24⋅1026,
and for every rational p/q with h(p/q)≥4.7⋅107,
Γ(1/3)−qp>(qe1)1.74⋅1026.
In particular μ(Γ(1/4))≤3.24⋅1026 and μ(Γ(1/3))≤1.74⋅1026.
The architecture is that of [1]: an approximation measure for the point θ=(1:π/ω:η/ω)∈P2(C) is obtained from Theorem 1 of [7], the rank of the interpolation matrix being controlled by Theorem 9 of [6]; the approximation measure is then converted into an irrationality measure for Γ. We keep this skeleton and change three of its ingredients, in § 3, § 5 and § 6.
Throughout, PROVED marks a statement proved here or quoted from the literature, and VERIFIED marks a numerical statement produced by the certified verifier of § 8, together with the range over which it was checked.
Notation
Ω=Zω1+Zω2 is a lattice of C with ℑ(ω2/ω1)>0 and algebraic invariants g2,g3; ℘, ζ, σ are the associated Weierstrass functions, η1,η2 the quasi-periods, and Vol(Ω) the area of a fundamental domain. For a nonzero period ω we write η for the associated quasi-period, and
θ1=π/ω,θ2=η/ω,θ=(1:θ1:θ2).
For x∈Qm, d(x) is the degree of Q(x) and h(x) the absolute logarithmic height; M(P) is the Mahler measure of P∈Z[X] and μ(P)=logM(P). For f analytic and r>0, M0(f,r)=sup∣z∣≤r∣f(z)∣. Finally Dist is the projective distance of [8].
We use throughout the exact growth constant
B∗=2Vol(Ω)π(1)
which is the value of Bruiltet’s B(Ω)=21(∣η1/ω1∣+2π/Vol(Ω)) when the Legendre relation is used to eliminate η1/ω1 in the lemniscatic and hexagonal cases.
An approximate Schwarz principle by Hermite interpolation
Theorem 1 of [7] rests on an approximate Schwarz principle in which the quantity Gr,r′(S) measures the “gravitation” of the interpolation nodes. In the elliptic applications of [8] and [1], G is bounded by the estimate of Reyssat [9], namely (66R/(δM))N−1, which is valid for an arbitrary separated set of nodes. When the nodes form a piece of a lattice, the following substitution is available and is substantially sharper.
For 0<r′<r and a finite set S={y1,…,yN} of pairwise distinct points of ∣z∣≤r′−1, put, following [7 §2.a],
where (ℓk,t)1≤k≤N,0≤t<T is the Hermite basis attached to S and T.
Proof. Let H be the Hermite interpolant of f at the yk to order T, so that H=∑k,tt!f(t)(yk)ℓk,t and M0(H,ρ)≤ε∑k,tM0(ℓk,t,ρ) for every ρ. Put φk(z)=r(z−yk)/(r2−ykz) and B=∏kφkT, so that ∣B∣=1 on ∣z∣=r. The function (f−H)/B is analytic on ∣z∣≤r, because f−H vanishes to order at least T at each yk. By the maximum principle, ∣f−H∣≤∣B∣(M0(f,r)+M0(H,r)) on ∣z∣≤r. Finally, for ∣z∣=r′ and each k, the function z↦∣z−yk∣/∣r2−yˉkz∣ attains its maximum at z=−r′yk/∣yk∣, whence
and therefore ∣B∣≤Er,r′(S)T on ∣z∣=r′. Combining the three displays gives the statement.
Remark 3. The product in (2) carries a + sign in the denominator precisely because the maximum above is attained at the antipodal point; this is the point at which Lemma 2 matches the quantity Er,r′ already present in [7].
We now bound Γ when S is a finite subset of a lattice Ω, contained in a disc of radius R0 and such that N=#S. Write λ=r/r′ and r′=R0+1.
Proposition 4 (PROVED). Let d1≤d2≤⋯ be the moduli of the nonzero vectors of Ω arranged in increasing order, and set L=∑i≤N−1logdi. Then
where, for ρ∈{r,r′}, Xρ=∑y∈Slog(ρ+∣y∣)−L, where ς=d1/2, and where S1=∑i≤N−1di−1, S2=∑i≤N−1di−2.
Proof. Fix k. For every j=k the point yk−yj is a nonzero vector of Ω, and these N−1 vectors are pairwise distinct; hence ∑j=klog∣yk−yj∣≥L by rearrangement. Consequently, for ∣z∣=ρ,
uniformly in k. The Hermite basis element ℓk,t is this product times (z−yk)t/t! times the truncation at order T−t of the reciprocal of the product, expanded at yk; bounding the latter by Cauchy’s formula on the circle of radius ς=d1/2 centred at yk, which encloses no other node, together with −log(1−x)≤x+x2 for 0≤x≤1/2, legitimate since ς/di≤1/2, produces the three correction terms log(1/ς), ςS1 and ς2S2. Summation over k and t<T contributes at most log(NT)≤log(2r) to each of the two circles, whence the factor 2log(2r).
Remark 5. In the verifier, L, S1 and S2 are evaluated by exact enumeration of the lattice vectors of modulus at most 200ω1; the tail is bounded in closed form by di≥(i+1)Vol(Ω)/π−diam(Ω). The sum ∑ylog(ρ+∣y∣) is bounded cell by cell on a 1000×1000 grid; each cell is evaluated at its unfavourable corner, and its point count is bounded by its area thickened by diam(Ω), divided by Vol(Ω).
The approximation measure
Let z0 be a torsion point of C/Ω, say nz0=ω′∈Ω, and let ω be a nonzero period. Set θ1=π/ω, θ2=η/ω.
Theorem 6 (PROVED; constants verified). With the notation of §2 there is a real A>0, effectively computable from Ω, ω and z0, such that for every α=(1:α1:α2)∈P2(Q),
The proof applies Theorem 1 of [7] in its original form. We give the data, the verification of its constraints, and the resulting inequality.
Data
Put K0=Q(α1,α2,g2,℘(z0),℘′(z0)) and a=[K0:Q]. For the two lattices of §5 one has Q(g2,℘(z0),℘′(z0))=Q(i), so that a≤2d(α); at the 3-division point of [1] the 3-division polynomial 3x4−6x2−1 is irreducible over Q and ℘′ adds a square root, so that a≤8d(α) there. Set
f1=σ2℘,f2=σ⋅(ζ−ωηz),f3=σ,
three entire functions. For D1,D2∈N∗ take the monomials
Every monomial fh is σ2D1+D2℘h1(ζ−(η/ω)z)h2, so the three functions enter with bounded exponents and a constant total weight in σ.
The point set is G⊂Ω, the intersection of Ω with a disc of radius ρ centred at the origin, truncated by the strip ∣ℑz∣≤ρsinφ parallel to the rows Zω; M denotes the number of rows met, N=#G, and R0=sups∈G∣z0+s∣. With integers T0<T we take
and ϖ2(r)=ϖ3(r)+log(1+∣η/ω∣r+∣σ(z0)∣−1) bounds f2. The exponents of D give (6) below.
The matrix of approximations
Following the device of Chudnovsky used in [8], one expands z and ζ in the local parameter u=(℘−℘(z0))/℘′(z0) at z0: with γ−1=ζ(z0)−(η/ω)z0 and γ0=0,
with H(γ−1,…,γi)≤c3i. The quasi-periodicity and the Legendre relation give, for s=s1ω1+s2ω2 and ω=a1ω1+a2ω2,
ζ(z+s)−ωη(z+s)=ω2i(a1s2−a2s1)π+ζ(z)−ωηz,
so that the columns indexed by s and s′ with a1(s2−s2′)=a2(s1−s1′) coincide: the number of distinct columns is the number M of rows. Each entry of the matrix is a polynomial Qh,(t,s) in θ1,θ2 with coefficients in Q(g2,℘(z0),℘′(z0)), of degree at most D2 and of height at most c4(D1+TlogD2+D2log(TM)). The vectors of approximations are obtained by specialising θ at α, with λi+1,⋅=σ(z0+s)2D1+D2 and the Leibniz coefficients μt′,t,s of [8 (5)], and MD denotes the resulting matrix.
Verification of the constraints
Write δ(i)=[Ki+1:K0]s=1 for all i, which is Cδ. Put P∞=max(1,∣℘′(z0)∣) and h(P)=logP∞.
Cr,r′: ∣Si+1∣≤R0<r′=R0+1≤r<R=+∞, and Hr′(Si+1)=max(1,(r′−∣y∣)−∣t∣)=1 because r′−∣y∣≥1 on the support. This is why r′ is taken one unit larger than R0.
CN,r: cardSi′≤T0N and ∣Si′∣≤R0≤r.
CE: since n=1 the parasitic factor of [7] disappears and E(i)≥Er,r′(Si′)Hr′(Si+1)=Er,r′(G)T0, so that by (2)
the sum being bounded cell by cell as in Proposition 4, and N from below by the row-by-row count. The gravitation Gr,r′(Si′) is bounded by Lemma 2 and Proposition 4, which is where the present argument departs from [9].
Cϕ: by the height estimate of the Qh,(t,s) above,
and ϕ(0) is the same expression with T replaced by 1, since the index 0 involves only derivatives of order <T0.
Cϱ: with cℓ=−log∣σ(z0)∣ and ℓp=log30+21logP∞, ℓq=log4+logP∞,
ϱ(j)=max(Tℓp+D1ℓq,2D1cℓ),
the second branch coming from ∣λi+1,⋅∣−1≤∣σ(z0)∣−(2D1+D2), for which one uses that ∣σ(z0+s)∣≥∣σ(z0)∣ for every s∈G, a consequence of the functional equation of σ on the quadrant containing G.
U: the passage from U=−logDist(θ,α) to the quantity U(j) of CU costs
V1=2D1(A0+B∗(R0+1)2)+(c6+log2)T+Tℓp+D1ℓq,
where c6=log(2∣z0∣−1max(24∣℘′(z0)∣−1∣z0∣−2,1)) bounds the Leibniz coefficients μt′,t,s, together with logΓ from Lemma 2.
Rank of the matrix
If MD were not of rank #D, a nontrivial linear combination of its rows would give a nonzero P∈C[X,Y] of bidegree ≤(D1,D2) vanishing to order ≥T along the analytic subgroup expG(z,α2z) of the extension G of E by Ga, at all the points of Σ, where Σ is indexed by the rows. By Theorem 9 of [6] there is a proper algebraic subgroup G′ with
all the terms of Q being independent of h once the D2h(α) part of ϕ has been set aside. The minimum is then realised by U(j), whence
U≤T0(T0−logE−Dϖ(r)+T0V1+logΓ)=Ah+O(1),
with A obtained by optimising M,D2,rL,λ,ϕ at each degree. Replacing h by h+log(1+d) and tracking the dependence in d through a≤2d gives the shape d3/2log(1+d) of the statement. The O(1) is made explicit in § 7.
The two instances
Γ(1/4)
Take y2=4x3−4x, so that Ω=Zω1+Ziω1 with ω1=Γ(1/4)2/8π, and η1=π/ω1. Take ω=ω1 and the 4-torsion point z0=(ω1+ω2)/4. Then verified
Writing ω′=4z0 and γ′−1=ζ(z0)−η′/4, the duplication formula ζ(2z)=2ζ(z)+℘′′(z)/(2℘′(z)) and ζ(ω′/2)=η′/2 give γ′−1=−℘′′(z0)/(4℘′(z0))=(1−i)/2, and the Legendre relation gives
γ−1=ζ(z0)−ω1η1z0=21−i−2iθ1.(8)
The algebraic part (1−i)/2 has minimal polynomial 4z4+1, hence height 41log4=log2, and lies in K=Q(i,α); the term −2iθ1 shifts the integer index of 2iπ(a1s2−a2s1)/ω by one half, and clearing that denominator costs D2log2. The base field is K=Q(i,α1,α2), of degree at most 2d(α) instead of 16d(α).
Γ(1/3)
Take y2=4x3−1, so that Ω=Zω1+Zρω1 with ρ=e2iπ/3, ω1=Γ(1/3)3/(2π) and η1=2π/(3ω1). Take ω=ω1 and the 3-torsion point z0=(2ω1+ω2)/3. Then verified
as follows from ω1=Γ(1/4)2/8π, η1=π/ω1 in the first case and ω1=Γ(1/3)3/(2π), η1=2π/(3ω1) in the second, together with the Legendre relation. In both displays Γ occurs to the first power only. Approximating y (resp. Y) by an algebraic β of degree δ and height τ, and θ by p/q, gives α∈P2(Q) with d(α)≤δ and
h(α)≤cδτ+h+O(1),c=4(Γ(1/4)),c=3(Γ(1/3)),
c being the largest power of y, resp. Y, occurring in (10). This is the constant c of Lemma 8.
Lemma 7 (Dirichlet, real case; PROVED). Let 0<β<1, let d≥1, put S=∑i≤d∣θ∣i and assume
t≥1−βlog(2d+1)+d1logS.
Then there is a nonzero P∈Z[X] with degP≤d, logM(P)≤t and log∣P(θ)∣≤−βdt.
Proof. Put H=⌊et/d+1⌋. The (H+1)d+1 polynomials with coefficients in {0,…,H} take at θ values lying in a real interval of length at most HS, so two of them, say P1=P2, take values at distance at most HS/((H+1)d+1−1)≤SH−d. Put P=P1−P2. Then ∥P∥2≤d+1H≤et, so M(P)≤∥P∥2≤et by Landau’s inequality, and log∣P(θ)∣≤logS−dlogH≤−βdt by the hypothesis on t.
Only the one-dimensionality of the pigeonhole is used; this is what allows β up to 1, where the complex case of [4 Lemme 1] is limited to β<1/2.
We next feed Lemma 7 into Lemma 2 of [4], which splits P either into a power of an irreducible factor carrying at least 2β/3 of the weight, or into two coprime factors. On the second branch we use Proposition 3.5 of [10] directly, with the sum constraints degF+degG≤d and μ(F)+μ(G)≤t supplied by Lemma 2 of [4]; since ∣Res(F,G)∣≥1, 2mn≤2d2/4 and nμ(F)+mμ(G)≤dt, this yields, for every integer s≥1,
logρ≤−s24(3βs−1−4log2td)dt,(11)
ρ being the distance from θ to the roots of FG. On the first branch we use the lemma of Diaz [3], in the form of Proposition 6 of [1].
The following lemma is what makes the two branches quantitative. Let S1 and S2 denote the running maxima, over degrees at most D, of the slope of Theorem 6 and of that slope divided by the degree; both are nondecreasing and satisfy S1(d)≤dS2(d). Write t=mh and let c be the coefficient of h in the conversion (c=4 for Γ(1/4), c=3 for Γ(1/3)).
Lemma 8 (PROVED). Let ∏jQjej be a group of factors, the Qj irreducible of degree δj and height τj, with ∑jejδj≤xD and ∑jejτj≤yt. If the lemma of Diaz fails on every Qj, then the normalised weight ϖ=(∑jejwj)/(Dt), where wj=−log∣Qj(θ)∣, satisfies
ϖ≤Φ(x,y)=s≥1min{sxy+2s+1S2(xD)(mx+Dcy)}.
Proof. Failure at Qj for a given s reads wj≤δjτj/s+2s+1(hS1(δj)+cτjS2(δj)). Multiply by ej, sum over j, and use
the last two because S2 is nondecreasing and δj≤xD. Dividing by Dt and taking the minimum over s gives the statement.
The configuration of factors is then controlled cell by cell on the square 0≤x,y≤1, each cell being evaluated at its unfavourable corner, using the three constraints of Lemma 2 of [4]: ϖF≥β/3, ϖG≥β/3 and ϖF+ϖG≥β on the coprime branch, and ϖQ≥2β/3 together with ϖQ≥β−Φ of the complement on the power branch.
Threshold and effective statements
Theorem 6 is used in its asymptotic form in h, in which D1 and T are proportional to h while D2, M and the shape parameters are fixed. The terms discarded are therefore exactly those independent of h. Since the exponents of a monomial of D satisfy h3=2(D1−h1)+D2−h2, the growth term is
where ϖ1,ϖ2,ϖ3 bound f1,f2,f3; only the terms carrying D1 are linear in h. The discarded part of (12) is D2max(ϖ2,ϖ3)(r), and since QT0=aD2(1+q)h at the optimum, D2 cancels and its relative weight is
a(1+q)hmax(ϖ2,ϖ3)(r).
The remaining discarded terms are ac8D2log(TM) coming from ϕ, with c8=7+logc1+h(P), that is c8=8.04 for Γ(1/4) and c8=7 for Γ(1/3) by § 5; c9D2logT coming from ϱ; the terms in M2 of U; and Δj+Δ0. All of these are smaller than the previous one at the optimum. Requiring the total to be at most 0.25%, which inflates the exponent by about 1%, and adding the hypothesis D2≤D1 of § 4(iv), the threshold of Lemma 7 and the term 4tlog2 of (11), gives
H0=5.4⋅107(Γ(1/4)),H0=4.7⋅107(Γ(1/3)).
the binding constraint being in both cases the D2 part of (12). The other constraints are inactive by at least two orders of magnitude; those of Lemma 7 and of (11) by more than ten.
Certification
The numerical part consists of two stages.
The slope tables of Theorem 6 are computed over 3489 cells covering 1≤d≤1030, geometrically spaced with ratio 1.02. Within each cell the free parameters – number of rows, D2, the ratio L/#S1, the radius ratio λ and the shape parameter of the region – are optimised in floating point, and the resulting value is then re-evaluated in interval arithmetic; only the interval upper bound is retained, inflated by a further factor 1+10−12. Each table entry is the running maximum, which is what Lemma 8 requires.
The conversion is verified in extended precision (80-bit, ϵ≈10−19) with an outward relative margin of 10−9, which makes every test strictly harder and dominates the accumulated rounding error; the same verification in double precision returns identical verdicts. The final value is assembled in interval arithmetic. Minimality is checked by confirming that the binding criterion fails at 0.99999 times the optimum.
The ancillary files contain the verifier and the tables. VERIFIED outputs:
μ(Γ(1/4))≤3.2011⋅1026,μ(Γ(1/3))≤1.7158⋅1026,
to which the 1% inflation of § 7 is applied in Theorem 1.
Acknowledgements
Technical assistance: Claude (Anthropic).
References
[1]S. Bruiltet, D’une mesure d’approximation simultanée à une mesure d’irrationalité: le cas de Γ(1/4) et Γ(1/3), Acta Arith. 104 (2002), 243–281.
[2]H. Cohen, W. Zudilin, Continued fractions and irrationality measures for Chowla–Selberg gamma quotients, arXiv:2510.00215.arxiv.org/abs/2510.00215
[3]G. Diaz, Une nouvelle propriété d’approximation diophantienne, C. R. Acad. Sci. Paris Sér. I 324 (1997), 969–972.DOI
[4]M. Laurent, D. Roy, Criteria of algebraic independence with multiplicities and approximation by hypersurfaces, Ann. Inst. Fourier 49 (1999), 27–55.
[5]D. W. Masser, Sharp estimates for Weierstrass elliptic functions, J. Anal. Math. 90 (2003), 257–302.DOI
[6]P. Philippon, Nouveaux lemmes de zéros dans les groupes algébriques commutatifs, Rocky Mountain J. Math. 26 (1996), 1069–1088.DOI
[7]P. Philippon, Une approche méthodique pour la transcendance et l’indépendance algébrique de valeurs de fonctions analytiques, J. Number Theory 64 (1997), 291–338.DOI
[8]P. Philippon, Mesures d’approximation de valeurs de fonctions analytiques, Acta Arith. 88 (1999), 113–127.DOI
[9]E. Reyssat, Approximation algébrique de nombres liés aux fonctions elliptiques et exponentielle, Bull. Soc. Math. France 108 (1980), 47–79.DOI
[10]D. Roy, M. Waldschmidt, Approximation diophantienne et indépendance algébrique de logarithmes, Ann. Sci. École Norm. Sup. 30 (1997), 753–796.DOI