Skip to content

Section 3.15 · PDF pages 49–52

The character boxes and the thin estimate on a host

The notation transcription is provisional. Check it against the printed mathematics when citing or reusing a formula.

3.15 The character boxes and the thin estimate on a host

Fix a host field as in Lemma 3.20, with modulus Q_{S}, distinct primes q, d, and finite abelian frame A. Write Q_{odd} for the odd part of Q_{S}. A character of the host means the primitive Dirichlet character attached to a character of its Galois group. Its conductor divides Q_{S}qd: the latter character is obtained from the unit group modulo Q_{S}qd, and passing to its primitive character divides this modulus.

Lemma 3.33 (Real-character decomposition and host boxes). Suppose M > 0, Q_{odd} | M, every prime dividing Q_{S} divides 2M, and PairGood(M, ϵ, Y, X, q, d) holds. For each real primitive host character χ whose conductor f_{χ} is divisible by q or d, put j = 1 if d | f_{χ}, and j = 0 otherwise. Then

Selectable PDF notation • transcription unchecked
f_{χ} ≤8Q_{odd}(e^{Y ϵ/5}e^{jY ϵ/5(log Y )2}, )
L(s, χ) ̸= 0 ℜs > 399/400, |ℑs| ≤8Me^{Y ϵ/5}e^{jY ϵ/5(log Y )2} .

In particular every such real L-function has no real zero greater than 399/400. The six quadratic boxes and the three QDq(X, ϵ, 4) bounds used in HostGood follow from the same pair.

Proof. A real Dirichlet character takes the values 1, −1 on units. At an odd prime p, the subgroup 1+pZ/p^{v}Z has odd order, so maps trivially to {1, −1}; its local conductor exponent is at most one. At 2, a quadratic character factors through modulus 8. Indeed the units modulo 2^{v}, for v ≥ 3, are generated by −1, 5; the elements congruent to 1 mod 8 are exactly the even powers of 5, so every quadratic character kills them. The cyclic description follows from v_{2}(5^{2r}−1) = r+2, proved by the factorization 5^{2r+1}−1 = (5^{2r}−1)(5^{2r}+1). The conductors at levels below 8 satisfy the same bound.

By the Chinese remainder theorem, the local factor at either q or d is consequently trivial or the unique quadratic character at that prime. The remaining local factors give a real primitive character ρ of conductor dividing 8Q_{odd}. Thus there are i, j ∈{0, 1} with

Selectable PDF notation • transcription unchecked
f_{χ} = f_{ρ}q^{i}d^{j}, χ(a) = ρ(a)(a/q)^{i}(a/d)^{j}. (63)

The equality includes nonunits: a positive local exponent gives zero on a nonunit at that prime. Each nontriv-ial quadratic factor at a new prime is primitive, and a product of primitive characters of coprime conductors

is primitive. To see the latter assertion, if the product factored through a proper divisor, fixing all but one CRT coordinate to 1 would force that local primitive factor to factor through a proper divisor as well. This also proves the exact conductor equality. In particular j is exactly the indicator of d | f_{χ}; and i, j cannot both be zero under the hypothesis of the lemma.

Every integer coprime to 2Mqd is coprime to f_{χ}, since every prime factor of Q_{S} divides 2M. Therefore (63) is an inducing relation in precisely the sense used in the definition of PairGood. Also f_{ρ} | 8M. Its real-character clause gives the stated zero box. The pair’s upper bounds

Selectable PDF notation • transcription unchecked
q ≤e^{Y ϵ/5}, d ≤e^{Y ϵ/5(log Y )2}

and f_{ρ} ≤8Q_{odd} give the conductor bound. At a real argument the height condition is automatic. This proves the first two assertions.

For completeness, the six small boxes use ρ = 1 or the primitive quadratic character χ_{8}(a) = (2/a), and (i, j) = (1, 0), (0, 1), (1, 1). On odd a, quadratic reciprocity gives

(−q/a) = (a/q), (−d/a) = (a/d) (q, d ≡3 mod 4).

The conductor of χ_{8} is exactly 8: if it factored through a proper divisor, it would factor through 4, whereas 5 ≡ 1 mod 4 and χ_{8}(5) = −1. The corresponding conductors therefore have the same product form as above, and primitive uniqueness identifies any proposed inducing character with this product. The three bounds at X instead use ρ = 1 and these same nonzero (i, j). The pair gives ‖𝒢_{χ}(X)‖≤ X^{1/2+ϵ/2} and |L^{′}(1, χ)/L(1, χ)| ≤4 log X. Since X ≥Y > 1, the first bound is at most X^{1/2+ϵ}, exactly the required QDq bound. These are the arguments of CharacterBoxes/RealSplit.lean:43--160, RealBox.lean:13--112 and PairRows.lean:9--215,264--320.

Proposition 3.34 (A uniform thin estimate). Fix B ≥1 and 0 < ϵ < 1/52. There is x_{0} = x_{0}(B, ϵ) such that every host with

Q_{odd} ≤B, log(Q_{S}qd) ≤2x^{ϵ/5}(log x)^{2}

and the real zero exclusion of Lemma 3.33 satisfies, for every primitive host character χ and every y ≥

Selectable PDF notation • transcription unchecked
^{1}_{3}e^{xϵ/3},
|𝒢_{χ}(y)| ≤ye^{−xϵ/8} (x ≥x_{0}). (64)

Thus the constant c_{6} in the host construction can be taken to be 1, independent of B and ϵ.

Proof. First give a uniform positive gap from 1 for all real nonprincipal host characters. Those whose conductors contain q or d have the fixed gap 1/400. For all others the local conductor computation in the preceding proof gives f_{χ} | 8Q_{odd}, hence f_{χ} ≤8⌊B⌋. There are only finitely many Dirichlet characters at these bounded conductors. Each nonprincipal L(s, χ) is continuous and nonzero at 1, so is nonzero on some real interval about 1. Take the minimum of these finitely many positive radii and 1/400, obtaining c_{B} > 0. Every real nonprincipal host character therefore has no real zero in (1 −c_{B}, 1). This is the finite-minimum argument in MeanSquareBounds/SmallZeros.lean:11--45.

Here is the analytic estimate used after this reduction. For a primitive character χ of conductor f, write ℓ= max(1, log f) and

Selectable PDF notation • transcription unchecked
ψ_{χ}(t) =∑Λ(n)χ(n), ψ_{χ}^{0}(t) ={^{ψχ(t),χ ̸= 1,}
n≤t^{ψ(t) −t,f = 1, χ = 1.}

The relative weighted progression theorem (61), summed against χ(a) over the units modulo f, gives abso-lute C, a > 0 such that, if t ≥f^{12},

Selectable PDF notation • transcription unchecked
{ √ }
|ψ_{χ}^{0}(t)| ≤Ct e^{−a log t/ℓ}+ e^{−alog t} + t^{−cB} + t^{−1/4} . (65)

We detail why this application is uniform. The primary main term cancels for a nonprincipal character by finite character orthogonality. The exceptional term also cancels unless χ equals the real exceptional character: the sum of χ(a)χ_{∗}(a) is zero otherwise. In that remaining case the established gap gives t^{β∗}/β_{∗}≤ 2t^{1−cB}. The sum of the absolute main terms H_{a}(t) is at most 3t; this bounds the summed relative errors_{√} without a factor f. The contribution of proper prime powers is at most C t(log t)^{2}, which is at most Ct^{3/4} after an absolute initial range. Finally the 3/5 exponent in the weighted theorem implies the square-root exponential written in (65). For the principal primitive character f = 1, apply that theorem modulo 2 to the odd primes; the omitted prime 2 and proper prime powers have the same elementary error bound. This gives (65) also in the principal case. There is no growing-modulus exception omitted in this step.

The exact tail identity for the functions used here is

Selectable PDF notation • transcription unchecked
∫ ∞_{ψ}0_{(t)}
𝒢_{χ}(y) = −y_{y}^{χ}_{t2} dt. (66)
∫ y

For χ ̸= 1, finite summation by parts gives R_{χ}(y) = y_{1} ψ_{χ}(t)t^{−2}dt. The estimate just proved makes the full integral absolutely convergent; Abelian passage from −L^{′}/L(s, χ) as s ↓ 1 identifies it with −L^{′}(1, χ)/L(1, χ). Subtracting gives (∫ 66). For the principal character, the Laurent expansion −ζ^{′}/ζ(s) =

Selectable PDF notation • transcription unchecked
∞

1/(s −1) −γ + O(s −1) gives _{1} (ψ(t) −t)t^{−2}dt = −1 −γ. Indeed (−ζ^{′}/ζ)(s)/s −1/(s −1) tends to −1 −γ, and the same absolutely convergent estimate permits passage to the limit. This agrees with the correction (1 + γ)y −y log y in 𝒢_{1}. This is the integration step of MeanSquareBounds/GIntegral.lean and Thin.lean:10--49.

Set z = log y. Integration of (65), using v = log t, bounds the integral in (66) by

Selectable PDF notation • transcription unchecked
{_{ ℓ √} √ }
C_{a}e^{−az/ℓ}+ (1 + z) C_{a}e^{−az} + c^{−}_{B}^{1}e^{−cBz} + 4e^{−z/4} .

For the square-root term substitute w =^{√}v; the resulting integral of 2we^{−aw} is an absolute constant depend-ing only on a times (1 +^{√}z)e^{−a√z}. The conductor hypothesis gives ℓ≤3x^{ϵ/5}(log x)^{2} eventually. Also z ≥x^{ϵ/3}− log 3 ≥^{1}_{2}x^{ϵ/3}. Consequently

Selectable PDF notation • transcription unchecked
_{azax}2ϵ/15_{√a}
≥ a z ≥ x^{ϵ/6}.

ℓ 6(log x)^{2 ,} 2

Both dominate x^{ϵ/8}, with arbitrarily large fixed multiplicative margin, because 2/15 > 1/8 and 1/6 > 1/8. The factors ℓ, 1 +^{√}z can be absorbed by using half their exponential exponents. The remaining exponents c_{B}z, z/4 dominate it as well. Increase x_{0}(B, ϵ) so that the sum of all four terms is at most e^{−xϵ/8}. Each bound persists for every larger z; the functions involved are decreasing after their fixed initial ranges. This proves (64) simultaneously for all the stated y, χ, with c_{6} = 1. It is the conductor split and integral estimate used in MeanSquareBounds/GaussSumFunctionEq.lean:55--107.

Lemma 3.35 (Polynomially bounded auxiliary primes). There is an absolute A ≥1 such that every integer m ≥2 has a prime p ≡1 mod m with p ≤m^{A}. Consequently there is an absolute C_{1} ≥1 such that every N ≥30 admits three distinct primes p, p^{′}, q_{0}, each at most N^{C1}, satisfying

p, p^{′}≡1 mod 15N, q_{0} ≡1 mod N.

Proof. Use the checked weighted progression theorem (61) at X = m^{A}, now for every m ≥ 2. Choose A ≥12 large enough that its absolute relative error is less than 1/2, simultaneously for these moduli. The first exponential is e^{−dA}. The second is uniformly small as A →∞: the function v^{3/5}/(log v)^{1/5} is increasing after an absolute initial range, and log X = A log m ≥A log 2. Thus one absolute A suffices. The main term

for residue 1 is strictly positive. If there is an exceptional pair, write δ = 1−β_{∗}> 0. Then − log β_{∗}≤δ/β_{∗}, so X^{−δ}/β_{∗}< 1 once log X > 1/β_{∗}. Here β_{∗}≥1 −1/(13 log(3m)) > 7/8 and A log m ≥12 log 2 > 8/7. In its absence positivity is immediate. Hence ϑ(X; m, 1) > 0, which means there is a prime p ≤X in that progression. This proves the exact assertion SmallOne; no ineffective zero has been removed.

For the second assertion, apply this result successively to 15N, 15Np, and Npp^{′}. As 15N ≤N^{2}, the chosen primes satisfy

Selectable PDF notation • transcription unchecked
_{p ≤N}2A_{,p}′_{≤N}(2A+2)A_{,q0 ≤N}[1+2A+(2A+2)A]A_{.}

The new congruence p^{′} ≡ 1 mod p forces p^{′} ̸= p, and q_{0} ≡ 1 mod p, p^{′} forces both other inequalities. Reduction of their congruences gives the displayed residues. The final exponent is at least the preceding two when A ≥1; take it for C_{1}. This is the construction in ShortNorms/ExistsThreePrimesOneModLeRpow.lean:10--70.