Section 3.12 · PDF pages 34–42
Simultaneous analytic selection of the prime pair
The notation transcription is provisional. Check it against the printed mathematics when citing or reusing a formula.
3.12 Simultaneous analytic selection of the prime pair
We spell out the selection argument, since a pointwise estimate at each y would not suffice for (33). For Z > 0, a bounded sequence κ : ℕ →ℂ, and an odd prime q, put
Selectable PDF notation • transcription unchecked
∑ (n)
S_{Z}(y; q, κ) = Λ(n) min(1, y/n) κ(n) .
Selectable PDF notation • transcription unchecked
1≤n≤Z^{q}For a finite prime set P, write MS(C, Z, P; [Y, X]) for the assertion
Selectable PDF notation • transcription unchecked
∑
|S_{Z}(y; q, κ)|^{2}≤Cy log y |P| (Y ≤y ≤X, ∀n, |κ(n)| ≤1) . (41)
Selectable PDF notation • transcription unchecked
q∈P
In particular the bound holds for truncated sequences depending on y; this is needed for the derivative esti-mate below.
Lemma 3.21 (A maximal bound from the mean square). For C ≥0, τ > 0, and η > 0, there is a threshold Y_{0}(C, τ, η), independent of X, Z, P, κ, such that (41) implies
Selectable PDF notation • transcription unchecked
||||{q ∈P : ∃y ∈[Y, X], |S_{Z}(y; q, κ)| >^{1}_{2}y^{1/2+τ}}|||| ≤η|P|whenever Y ≥Y_{0} and |κ(n)| ≤1.
Proof. Fix q, κ. The finite sum S_{Z}(y) = S_{Z}(y; q, κ) is continuous and piecewise linear. Its right derivative is
Selectable PDF notation • transcription unchecked
D_{Z}(v; q, κ) = ∑^{Λ(n)}κ(n)(^{n}).
1≤_{v<n}n≤Z^{nq}Both D_{Z} and |D_{Z}|^{2} are bounded measurable step functions on a compact interval, and
Selectable PDF notation • transcription unchecked
∫ b
S_{Z}(b) −S_{Z}(a) =_{a} D_{Z}(v) dv.Furthermore, with κ_{v}(n) = 1_{v<n}κ(n), one has vD_{Z}(v; q, κ) = S_{Z}(v; q, κ_{v}). Since |κ_{v}| ≤ 1, the mean square yields
Selectable PDF notation • transcription unchecked
∑
_{|DZ(v; q, κ)|2≤C} log v
Selectable PDF notation • transcription unchecked
|P| (Y ≤v ≤X, v > 0). (42)
q∈P^{v}First restrict to A ≤ y ≤ W ≤ 2A, with A ≥ 2. The fundamental theorem of calculus, the triangle inequality and Cauchy–Schwarz give, for any choice y_{q} ∈[A, W],
Selectable PDF notation • transcription unchecked
∫ W
|S_{Z}(y_{q}; q, κ)|^{2}≤2|S_{Z}(A; q, κ)|^{2}+ 2A_{A} |D_{Z}(v; q, κ)|^{2}dv.Sum over q ∈ P, interchange the finite sum and integral, and use (42). On this interval (log v)/v ≤ log(2A)/A, so ∑
|S_{Z}(y_{q}; q, κ)|^{2}≤4CA log(2A)|P|. (43)
Selectable PDF notation • transcription unchecked
q∈P
For each prime having an exceptional y in this block, choose one such y_{q}; choose y_{q} = A for all other primes. Each exceptional term exceeds A^{1+2τ}/4. It follows that their number is at most
16C A^{−2τ} log(2A)|P| ≤^{32C}A^{−τ}|P|.
Selectable PDF notation • transcription unchecked
τ
Here log(2A) ≤2 log A ≤2A^{τ}/τ; the second inequality follows from log u ≤(u^{τ}−1)/τ, u ≥1.
Now choose one exceptional y_{q} ∈[Y, X] for each exceptional prime and assign it a dyadic block
2^{j}Y ≤y_{q} ≤ min(2^{j+1}Y, X).
Such a block exists by taking the integer part of log_{2}(y_{q}/Y ); there are only finitely many assigned indices because P is finite. Apply the preceding estimate to the primes assigned to each block. The mean square there is the restriction of the original one. Summing the estimates gives
Selectable PDF notation • transcription unchecked
∑ 32C
|{exceptional primes}| ≤^{32C}Y^{−τ}|P| 2^{−jτ} =_{−τ} Y^{ −τ}|P|.
Selectable PDF notation • transcription unchecked
τ_{j≥0} τ(1 −2 )Choose Y_{0} ≥2 so that its coefficient is at most η. This choice is independent of the upper endpoint X, the truncation Z, and the prime set. This proves every step of MeanSquareBounds/BadShare.lean:9--318.
For clarity, the quantitative zero-density input used by the source is the following precise statement:
Selectable PDF notation • transcription unchecked
∃C_{D} ≥0 ∀Q ≥3 :_{(44)}#{χ primitive : N_{χ} ≤Q, ∃ρ, L(ρ, χ) = 0, ℜρ > 399/400, |ℑρ| ≤Q} ≤C_{D}Q^{1/4}.
Only the number of characters is counted; a character with several such zeros is counted once. A finite-set formulation suffices if this set is not first constructed as a finite set. The source proves it in JacobiCharacterSums/CountZAll.lean:13--15 through its zero-detection and mollifier argument. At referee level it has the following verified classical justification.
Proposition 3.22 (The required zero-density estimate). Assertion (44) holds.
Proof. Use Jutila’s Corollary, equation (1.6), on printed page 56 of Zero-density estimates for L-functions, Acta Arithmetica 32 (1977), 55–62, doi:10.4064/aa-32-1-55-62. It bounds the sum of zero counts over all primitive characters of conductor at most Q, for T ≥ 2 and 3/4 ≤ α ≤ 1. Take its integer k = 2, α = 399/400, T = Q, and its positive error exponent 1/100. The same corollary gives A_{1}(α) = 2 for α ≥21/26. Thus its bound is
Selectable PDF notation • transcription unchecked
_{(Q}2_{Q)}2/400+1/100_{= Q}9/200_{≤Q}1/4_{.}Every character counted here contributes at least one zero to that sum. The Euler product excludes zeros with ℜs > 1; the principal character of conductor 1 is included, with its pole at 1 excluded. The corollary omits neither real characters nor exceptional real zeros.
Lemma 3.23 (From bad characters to bad primes). Assume (44). Let B be a finite set of primes q > F such that, for each q ∈B, some primitive character of conductor fq, where f ≤F and fq ≤Q, fails ZB(χ, Q). Then |B| ≤C_{D}Q^{1/4}.
Proof. Choose one such character χ_{q} for each q. The map q →→ χ_{q}, including its conductor, is injective. Indeed, equality at q, q^{′} implies fq = f^{′}q^{′}. Since q is prime, it divides f^{′} or q^{′}. The first is impossible because 0 < f^{′} ≤ F < q; the second gives q = q^{′}. All chosen characters are in the set counted in (44), proving the bound. This is the argument of MeanSquareBounds/CountBadCore.lean:8--59.
Here is the precise progression range in which the remaining analytic estimates are used. Put
P(m, a, U) = {p : p prime, U^{1/2}< p ≤U, p ≡a (mod m)}.
A progression is admissible at (ϵ, Y, U) when
m > 0, 8 | m, (a, m) = 1,_{4}^{1}(log Y )^{2} log m ≤ log U,
Selectable PDF notation • transcription unchecked
(45)
Y^{ϵ/6}≤ log U ≤Y^{ϵ/5}(log Y )^{2}.
The three analytic estimates needed in this range are:
Selectable PDF notation • transcription unchecked
|P(m, a, U)| ≥U^{9/10},
( ) (46)MS C_{m}, (Y^{K})^{400}, P(m, a, U); [Y, Y^{K}] ,
the finite-sum-to-𝒢 comparison stated below.
The first is required eventually for every fixed ϵ > 0. The second is required with one absolute C_{m} > 0, eventually for each fixed 0 < ϵ ≤1, K ≥1, and with the additional condition
p | m =⇒ p ≤Y^{ϵ/5} or p > (Y^{K})^{400} (p prime). (47)
The comparison requires one absolute C_{B} > 0. For each 0 < ϵ ≤1/4, τ > 0, K ≥1, it is to hold eventually, uniformly for every primitive nonprincipal χ of conductor at most Q, where
Y^{ϵ/6}≤ log Q ≤Y^{ϵ/4}, ZB(χ, Q).
If χ(n) = κ(n)(n/q) for every n, and
Selectable PDF notation • transcription unchecked
|S_{(Y} K_{)}400(y; q, κ)| ≤^{1}_{2}y^{1/2+τ} (Y ≤y ≤Y^{K}),then
|L^{′}(1, χ)/L(1, χ)| ≤C_{B} log Y, |𝒢_{χ}(y)| ≤y^{1/2+τ} (Y ≤y ≤Y^{K}). (48)
These are exact quantitative supplier statements, with their uniform constants specified. The next proof uses them conditionally; the zero-density assertion has already been established in Proposition 3.22.
Lemma 3.24 (One simultaneous selection stage). Assume the three estimates just specified. For every fixed 0 < ϵ ≤1/4, τ > 0, K ≥1, η > 0, and integer R ≥0, there is a threshold Y_{0} with the following property. Let (45) and (47) hold, let 1 ≤F ≤U^{1/2}, and let Ξ be a set of at most R primitive characters of conductors at most F. Then P = P(m, a, U) is nonempty and at most η|P| of its primes fail the following condition: every primitive character inducing ξ(n)(n/q), for any ξ ∈Ξ, is nonprincipal, has conductor at most FU, satisfies ZB(χ, FU), and satisfies (48). Induction here may be tested on any prescribed nonzero modulus on which those values are specified.
Proof. Choose the thresholds for the three progression estimates. Apply Lemma 3.21 with exceptional frac-tion η/[2(R + 1)]. Enlarge the threshold so that
Selectable PDF notation • transcription unchecked
3 η
_{2}Y^{ ϵ/5}(log Y )^{2} ≤Y^{ ϵ/4}, C_{D}U^{3/8}≤ U^{ 9/10}
Selectable PDF notation • transcription unchecked
2
for every allowed U, and U ≥ 4. These are eventual inequalities because (log Y )^{2} = o(Y^{ϵ/20}), U^{3/8} = o(U^{9/10}), and log U ≥Y^{ϵ/6}→∞. Put Q = FU. As 1 ≤F ≤U^{1/2},
Y^{ϵ/6}≤ log Q ≤^{3}_{2} log U ≤Y^{ ϵ/4}.
The first progression estimate gives |P| ≥U^{9/10}> 0.
For q ∈ P, q > U^{1/2} ≥ F; thus q is an odd prime not dividing the conductor f_{ξ} of any ξ ∈ Ξ. The product of the primitive character ξ with the nontrivial quadratic character modulo q is primitive of conductor f_{ξ}q. This follows prime by prime: at the q-factor it has the full nontrivial quadratic conductor, and at each prime dividing f_{ξ} restriction to that unit factor recovers ξ, whose local conductor cannot decrease. It is consequently nonprincipal and has conductor f_{ξ}q ≤FU.
Let B_{1} consist of primes for which one of these primitive products fails ZB(χ, Q). Lemma 3.23 gives
Selectable PDF notation • transcription unchecked
|B_{1}| ≤C_{D}Q^{1/4}≤C_{D}U^{3/8}≤(η/2)|P|.For each ξ ∈Ξ, let B_{2,ξ} consist of primes having some y ∈[Y, Y^{K}] at which the finite sum exceeds ^{1}_{2}y^{1/2+τ}. Since a character has absolute value at most 1, the maximal mean-square lemma gives
Selectable PDF notation • transcription unchecked
|B_{2,ξ}| ≤_{2(R}^{η}_{ + 1)}|P|.Their union therefore has at most (η/2)|P| elements. Outside the two exceptional sets the comparison (48) applies.
Finally a second primitive character agreeing with the product away from a nonzero modulus is the same primitive character. Inflate both to a common multiple of their conductors and the testing modulus; they agree on all units there. Equality of the inflated characters gives the same minimal conductor and the same primitive character. Thus a failure of the condition for any inducing character is already detected by B_{1} or one of the B_{2,ξ}. The union bound proves the assertion. This is MeanSquareBounds/StagePow.lean:41--170, including its separation of the two exceptional causes.
We now apply the stage twice. Put z = Y^{ϵ/5} and, temporarily, let
z(log Y )^{2}≤z^{′}≤K^{2}z(log Y )^{2}.
The first prime set is P(8, 3, e^{z}). For a selected first prime q, put
Selectable PDF notation • transcription unchecked
M_{odd} = q ∏ p.
p≤zp odd prime
Let the second prime set consist of primes
Selectable PDF notation • transcription unchecked
′ ′
e^{z/2}< d ≤e^{z}, d ≡7 (mod 16), d ≡−1 (mod M_{odd}).
If q ≤z the repeated prime in M_{odd} does not change its prime support; the congruences remain equivalent. The Chinese remainder theorem identifies this second set with P(16M_{odd}, a_{q}, e^{z′}) for a residue class a_{q} coprime to 16M_{odd}.
Proposition 3.25 (Two stages produce a good pair). Assume the three analytic estimates (46), including their stated comparison, and the following logarithmic-derivative improvement:
Selectable PDF notation • transcription unchecked
∀0 < ϵ < 1/52 ∀A > 0 ∃Y_{0} ∀Y ≥Y_{0} :
(49)Y^{ϵ/6}≤ log Q ≤AY^{ϵ/5}(log Y )^{2}, N_{χ} ≤Q, χ ̸= 1, ZB(χ, Q)
=⇒ |L^{′}(1, χ)/L(1, χ)| ≤4 log Y.
Then PairEx holds.
Proof. Fix M ≥ 1, 0 < ϵ < 1/52, A ≥ 1. Use the selection stage first with parameter ϵ and then with
6ϵ/5 < 1/4. The finite sets of starting primitive characters will have at most #(ℤ^{̂}/8M)^{×}, respectively twice this number, elements. Use τ = ϵ/2, K = A, and exceptional fraction 1/2 in both stages. All thresholds are chosen before Y and q.
Here are the checks allowing those applications. In the first stage take m = 8, a = 3, U = e^{z}, F = 8M. For large Y , (log Y )^{2}≤z/2, z ≥16M, and log Y ≥1. Since log 8 < 2.1, these give ^{1}_{4}(log Y )^{2} log 8 ≤z; the other bounds in (45) follow from ϵ/6 ≤ϵ/5. The only prime in m is 2 ≤z, and 8M ≤e^{z/2}. Replacing each character modulo 8M by its primitive character does not increase the size of this finite set, and preserves its values on units of that modulus. The stage therefore leaves a nonempty set of first primes for which all twists by ( · /q) have the required boxes and estimates. Every such q exceeds 8M, eventually.
For the second stage, the elementary primorial bound ∏_{p≤z}p ≤4^{z} and q ≤e^{z} give
log(16M_{odd}) ≤4z
for large Y . Hence ^{1}_{4}(log Y )^{2} log(16M_{odd}) ≤z^{′}. Its parameter ϵ^{′} = 6ϵ/5 has ϵ^{′}/6 = ϵ/5; thus Y^{ ϵ′/6} ≤z^{′}. Also K^{2}≤Y^{ϵ/25} eventually, so
z^{′}≤K^{2}Y^{ϵ/5}(log Y )^{2}≤Y^{ϵ′/5}(log Y )^{2}.
Its modulus primes other than q are at most max(2, z) ≤Y^{ϵ′/5}. For large Y , 800K log Y ≤z; since q > e^{z/2}, this gives q > (Y^{K})^{400}. This verifies (47).
The starting characters here are the primitive characters of ξ( cdot )( · /q)^{i}, for all ξ mod 8M, i = 0, 1. Their conductors are at most 8Mq ≤ F = 8Me^{z}. As z^{′} ≥ z(log Y )^{2} and z →∞, log(8M) + z ≤ z^{′}/2 eventually; hence F ≤ e^{z′/2}. The threshold remains inde-pendent of q, because its size bound was uniform and the number of starting characters was fixed. At least half the second prime set is therefore good. Choose q from the first good set and d from its second good set, and take z^{′}= z(log Y )^{2}. The congruences and exponential ranges give exactly the window-pair conditions; moreover q < d, and (8M, qd) = 1.
For a twist with j = 0, necessarily i = 1, the first stage gives its box of size 8Me^{z}. One may remove d from the testing modulus: equality away from 2Mqd identifies the same primitive character as equality away from 8Mq, by the uniqueness proved in the stage. For j = 1, the second stage gives the box of size 8Me^{z+z′}, for either i. Inflating a character modulo M, or a real primitive character of conductor dividing 8M, to modulus 8M preserves all the needed values on units. Thus these two stages cover every character quantified in PairGood. The finite-sum convention agrees with 𝒢_{χ} for a nonprincipal character, since n ≤y makes max(y/n −1, 0) = y/n −1.
Finally, for j = 0, 1 write
Selectable PDF notation • transcription unchecked
′
Q_{j} = 8Me^{z}e^{jz}, log Q_{j} = log(8M) + z + jz^{′}.
For Y ≥ exp(1 + log(8M)),
Y^{ϵ/6}≤ log Q_{j} ≤3Y^{ϵ/5}(log Y )^{2}.
Apply (49) with its constant A = 3. It gives the specific 4 log Y bound, hence 4 log y for y ≥Y . The two stages already give the bounds |𝒢_{χ}(y)| ≤y^{1/2+ϵ/2} and the boxes at the exact heights Q_{j}. This proves every condition of PairGood(M, ϵ, Y, Y^{A}; q, d). The proof follows MeanSquareBounds/Stage1Succ.lean:36--95, CharOf.lean:61--217, PairShare.lean:26--52, PairExists.lean:9--49 and pairEx.lean:11--169.
Lemma 3.26 (The zero-free-box contour estimate). For 0 < ν ≤ 1/800 there is K_{ν} ≥ 0 such that, for every primitive nonprincipal χ with N_{χ} ≤Q, Q ≥2, and ZB(χ, Q), and all z ≥1, 0 < η ≤1,
Selectable PDF notation • transcription unchecked
|R_{χ}(z + zη) −R_{χ}(z) −zη D_{χ}(1)|
( ) (50)≤K_{ν} log Q log Q η z^{1−1/400+ν}+^{z2} , D_{χ} = −L^{′}/L.
Selectable PDF notation • transcription unchecked
Q
The constant depends on ν, not on the character or the three real variables.
Proof. We first justify the local logarithmic-derivative bound used on the contour. The partial sums of a nonprincipal character modulo N are bounded by N, by splitting into complete periods. Summation by parts therefore represents L(s, χ), for ℜs > 0, as
Selectable PDF notation • transcription unchecked
∫ ∞(∑ )
L(s, χ) = s ( χ(n))u^{−s−1}du.
^{1} n≤uOn the disk of radius 11/20 centered at 21/20+it this gives a polynomial upper bound CN(|t|+2), with an absolute C, since the real part is at least 1/2. At the center the Euler product gives the absolute lower bound 1/ζ(21/20). Normalize by that central value and call the resulting analytic function f. Its central value is 1, and log max |f| = O(log(N(|t| + 2))).
For completeness, the local zero subtraction used here can be seen directly from Jensen’s formula. The total multiplicity of zeros in the concentric disk of one third the radius is at most log max |f|/ log 3. Remove
their finite disk Blaschke product B; it has modulus 1 on the outer circle, so h = f/B still has modulus at most max |f|. It is zero free in the inner disk. Choose the analytic logarithm of h there, with its central imaginary part zero. Its central real part is nonnegative, because |B(0)| ≤ 1. The Cauchy estimate for a holomorphic function with bounded real part (equivalently, apply the disk positive-real-part transform and Schwarz’s lemma) bounds h^{′}/h on the disk of one tenth the outer radius by an absolute constant times log max |f|. The derivative of a disk Blaschke factor equals 1/(s −ρ) plus a term bounded by a constant times the reciprocal outer radius there. Hence
Selectable PDF notation • transcription unchecked
L^{′} ∑ m_{ρ} ( )(s, χ) = + O log(N(|t| + 2))
Selectable PDF notation • transcription unchecked
L s −ρ
ρ in the inner disk
for the real parts used below, namely 1 −1/400 + ν ≤ℜs < 11/10, ℑs = t. Indeed their distance from 21/20 + it is at most 21/400 < (11/20)/10. If |t| ≤Q −1, all the subtracted zeros have imaginary part of absolute value at most Q. The zero-free box forces their real parts to be at most 1 −1/400; thus |s −ρ| ≥ν. Since N(|t| + 2) ≤Q^{3}, the displayed formula gives
|D_{χ}(s)| ≤K_{B}(ν) log Q (ℜs ≥1 −1/400 + ν, |ℑs| ≤Q −1) . (51)
For ℜs ≥ 11/10, the same bound follows from the absolutely convergent Euler logarithmic derivative. This is the local zero-subtraction argument of JacobiCharacterSums/TailBox.lean:35--119; the numerical local constant is immaterial here.
Set σ_{1} = 1 −1/400 + ν, T = Q/2. The absolutely convergent Perron identity on ℜw = 2 is
Selectable PDF notation • transcription unchecked
_{1} ∫ 2+i∞_{y}w
R_{χ}(y) =_{2πi2−i∞} D_{χ}(w)_{w(w −1)}dw.To verify it, insert the absolutely convergent series D_{χ}(w) =∑Λ(n)χ(n)n^{−w}, interchange sum and inte-gral, and use the elementary kernel integral
Selectable PDF notation • transcription unchecked
_{1} ∫ 2+i∞_{a}w_{w(w −1)}dw = max(a −1, 0) (a > 0).
Selectable PDF notation • transcription unchecked
2πi_{2−i∞}Its two residues at 0, 1 give the value for a > 1; closing the contour in the other direction gives zero for a < 1, and continuity gives the value at a = 1. The quadratic denominator makes the vertical tails integrable.
Move the finite segment at height T to ℜw = σ_{1}. The box excludes all poles of D_{χ} in this rectangle. Only the kernel pole at w = 1 is crossed, with residue yD_{χ}(1); w = 0 remains outside. Do this at y = z(1 + η) and y = z, and subtract. For 0 ≤ℜw ≤2,
Selectable PDF notation • transcription unchecked
|(z(1 + η))^{w}−z^{w}| ≤4η|w|z^{ℜw}.Indeed factor out z^{w}, write the remaining power as exp(w log(1 + η)), and integrate its derivative on the unit segment; use log(1 + η) ≤η and (1 + η)^{ℜw}≤4. On the left vertical side,
Selectable PDF notation • transcription unchecked
|w −1| ≥_{2}^{1}(1/800 + |ℑw|),so its reciprocal integral is at most 4 log(1+800T) ≤40 log Q. Together with (51), this bounds the left-side difference by 160K_{B}(ν)(log Q)^{2}ηz^{σ1}.
The two horizontal segments for either y have length at most 2, |w|, |w −1| ≥T, and y^{ℜw}≤4z^{2}. Their individual integrals are at most 16K_{B}(ν) log Q z^{2}/Q. On the omitted right-line tails, the Euler series bounds |D_{χ}(2 + it)| by an absolute constant, so integration of t^{−2} contributes at most an absolute multiple of z^{2}/Q. Adding the four horizontal terms, two right tails and the left difference, and dividing by 2π, proves (50) after enlarging K_{ν}. These are precisely the contour and kernel estimates in TailKernel.lean:9--125, TailContourA.lean:20--35 and TailContour.lean:14--135.
Proposition 3.27 (The finite-sum comparison and 4 log Y ). The comparison (48) and the improvement (49) hold, with all the uniformities stated there.
Proof. Write
Selectable PDF notation • transcription unchecked
P_{χ}(z) = ∑^{Λ(n)χ(n)}, T_{χ}(z) = D_{χ}(1) −P_{χ}(z).
1≤n≤z^{n}The elementary difference of the ramp sums gives, for h > 0, z ≥1,
Selectable PDF notation • transcription unchecked
h(h + 1)
|R_{χ}(z + h) −R_{χ}(z) −hP_{χ}(z)| ≤ log(z + h).
Selectable PDF notation • transcription unchecked
z
In fact the summands with n ≤z cancel exactly. At most h + 1 integers remain in z < n ≤z + h; each contributes at most log(z + h) h/z. Combining this estimate at h = zη with (50) bounds
Selectable PDF notation • transcription unchecked
( )
_{|Tχ(z)| ≤} K_{ν} log Q log Q ηz^{1−1/400+ν}+ z^{2}/Q + η(zη + 1) log(z + zη)
Selectable PDF notation • transcription unchecked
_{zη} . (52)First take x = Y^{K}, z = x^{400}, η = x^{−2} and ν = ϵ/(1600K). Multiplication of the right side by x simplifies to
log Q x^{403} x^{398} + 1 K_{ν}(log Q)^{2}Y^{ϵ/4}+ K_{ν} +_{399} log(x^{400}+ x^{398}).
Selectable PDF notation • transcription unchecked
Q x
The last term is at most 802: for x ≥3, x^{400}+ x^{398}≤x^{401}, log x ≤x, and x ≤x^{399}. The first term is at most K_{ν}Y^{3ϵ/4}, since log Q ≤Y^{ϵ/4}. The second is at most
K_{ν}Y^{404K} exp(−Y^{ϵ/6}) −→0,
because log Q ≤x and Q ≥ exp(Y^{ϵ/6}). These bounds, with ϵ, K fixed before Y , give
Selectable PDF notation • transcription unchecked
|Y^{K}T_{χ}((Y^{K})^{400})| ≤Y^{ϵ} (53)eventually, uniformly in Q, χ.
There is an absolute C_{0} > 0 with ∑_{n≤z}Λ(n)/n ≤ C_{0} log z for z ≥ 2. Taking K = 1 in (53) gives |T_{χ}(Y^{400})| ≤1 when ϵ ≤1. Consequently
|L^{′}(1, χ)/L(1, χ)| ≤400C_{0} log Y + 1 ≤(400C_{0} + 1) log Y
for Y ≥3. This proves the first estimate in (48), with an absolute constant. For its other estimate, the exact identity, for Z ≥y,
Selectable PDF notation • transcription unchecked
𝒢_{χ}(y) = −S_{Z}(y; q, κ) −yT_{χ}(Z) (χ(n) = κ(n)(n/q))follows by splitting the finite sums at n = y. Use Z = (Y^{K})^{400}, (53), and y ≤Y^{K}. Since ϵ ≤1/4 and
Selectable PDF notation • transcription unchecked
Y ≥16,
Y^{ϵ}≤^{1}_{2}y^{1/2+τ}.The assumed half-size bound for S_{Z} proves the second estimate in (48).
To obtain the coefficient 4, use the same contour with fixed ν = (2/5)/1600. Let l = log Q, s = l^{1/4}, x = s^{10}, z = x^{400}, and η = x^{−2}. For large l, s ≥3 and l^{1006}≤Q = e^{l}. The same simplification of (52), now without multiplying by x, shows that its first two terms are at most K_{ν} each:
Selectable PDF notation • transcription unchecked
l^{2}x^{−1+(2/5)/4}= s^{8}s^{−9}= s^{−1}≤1, lx^{402}/Q ≤1.The latter uses lx^{402}= (s^{4})^{1006}≤Q. Its elementary third term is at most 802. Thus |T_{χ}(z)| ≤2K_{ν} + 802.
Here one may use the more precise elementary Mertens bound ∑_{n≤z}Λ(n)/n ≤ log z + 2. For integer a ≥1, the factorial identity log(a!) =∑_{n≤a}Λ(n)⌊a/n⌋ and log(a!) ≤a log a −a + 1 + log a, together with the verified bound ∑_{n≤a}Λ(n) < 1.03883a used above, give it: (1 + log a)/a ≤1. Taking a = ⌊z⌋ gives the real-variable version. It follows that
|L^{′}(1, χ)/L(1, χ)| ≤1000 log log Q + C, C = 2K_{ν} + 804.
Indeed log z = 400 log x = 1000 log l. If log Q ≤AY^{ϵ/5}(log Y )^{2}, then
log log Q ≤ log A +_{5}^{ϵ} log Y + 2 log log Y.
The coefficient 1000ϵ/5 is less than 50/13 < 4 when ϵ < 1/52. The remaining 2000 log log Y +1000 log A+ C is o(log Y ); the lower bound log Q ≥Y^{ϵ/6} ensures all preceding large-l requirements. This proves (49). The calculations are those of MeanSquareBounds/TailKExp.lean:8--145, kappaC.lean:9--62, BoxOfS.lean:10--60 and Kappa4Loglog.lean:9--153.