Skip to content

Section 3.13 · PDF pages 42–46

The progression estimates underlying the mean square

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

3.13 The progression estimates underlying the mean square

We isolate the exact prime-number-theorem formulation, including its exceptional term, before applying∫

Selectable PDF notation • transcription unchecked
X

it. Let li_{2}(X) =_{2} dt/ log t. For a positive modulus m divisible by 8, let 𝔢_{m} be either absent or a pair (χ_{m}, β_{m}) satisfying

Selectable PDF notation • transcription unchecked
χ^{2} = 1, χ_{m} ̸= 1, L(β_{m}, χ_{m}) = 0, 1 −^{1}

^{m} 8 log m ^{< βm < 1.}

Define

Selectable PDF notation • transcription unchecked
{

li_{2}(X)/φ(m), 𝔢_{m} absent,

Selectable PDF notation • transcription unchecked
M_{m}(a, X) =

[li_{2}(X) −χ_{m}(a) li_{2}(X^{βm})]/φ(m), otherwise.

Since χ_{m} is real, this is a real number. The relative progression input is the assertion that there are absolute c_{3} > 0, A_{0}, C_{0} such that one may make these choices for all m, with no exceptional pair when none exists, and, for (a, m) = 1, X ≥2, C_{0} log m log(log m + 2) ≤ log X,

Selectable PDF notation • transcription unchecked
( √ )
|π(X; m, a) −M_{m}(a, X)| ≤|A_{0}| e^{−c3 log X/ log m}+ e^{−c3 log X} M_{m}(a, X). (54)

The source uses the equivalent version without the absolute value on A_{0}, which can only be weaker after this replacement. Two further inputs to this progression argument are: there is an absolute 0 < c ≤1 for which every real nonprincipal character modulo m and real zero β satisfy

Selectable PDF notation • transcription unchecked
1 −β ≥c/m; (55)

and there is an absolute c_{0} > 0 such that two real nonprincipal characters of moduli m_{1}, m_{2} ≤Z having real zeros β_{i} > 1 −c_{0}/ log Z agree on every integer coprime to m_{1}m_{2}. The latter is the Page consequence of uniqueness in the classical zero-free region: inflate to modulus m_{1}m_{2} ≤Z^{2}, where log(2m_{1}m_{2}) ≤3 log Z, and use its unique possible exceptional character. These statements retain all real zeros; no ineffective excep-tion is discarded. Proposition 3.31 below verifies this formulation, including the exceptional pair convention, from an established relative prime number theorem. Its proof also verifies the two zero-free consequences just stated.

Lemma 3.28 (Counting the admissible progression). Assume (54) and (55). The first assertion in (46) holds eventually, uniformly over (45).

Proof. Put L = log Y , u = log U, and Λ = log m. Admissibility gives

Selectable PDF notation • transcription unchecked
u ≥L^{2}Λ/4, u ≥Y^{ϵ/6}, Λ ≤4Y^{ϵ/5}.

Since m ≥8, Λ > 1. For large L,

Selectable PDF notation • transcription unchecked
( )

|C_{0}| log(Λ + 2) ≤|C_{0}| log 6 + (ϵ/5)L ≤L^{2}/4.

Thus the range condition for (54) is satisfied. Its relative error is at most 1/2, since

Selectable PDF notation • transcription unchecked
u/Λ ≥L^{2}/4, u ≥L^{2}/4

make both exponentials tend to zero uniformly.

For u ≥4, we claim

Selectable PDF notation • transcription unchecked
M_{m}(a, U) ≥_{2um}^{cU}_{2} . (56)

If the exceptional pair is absent, use li_{2}(U) ≥ (U −2)/u, φ(m) ≤ m, and m ≥ 8, c ≤ 1. If present, β_{m} ≥7/8, so U^{βm}≥2. The gap bound and u ≥1 give

Selectable PDF notation • transcription unchecked
U^{βm}≤Ue^{−c/m}≤U(1 −c/(2m)).

The last inequality follows by convexity of e^{−v} on 0 ≤v ≤1, using e^{−1}≤1/2. As χ_{m}(a) ≤1,

Selectable PDF notation • transcription unchecked
_{U −U} βm_{cU}
li_{2}(U) −χ_{m}(a) li_{2}(U^{βm}) ≥_{u} ≥_{2mu}.

Divide by φ(m) ≤m to obtain (56).

For large Y , admissibility also gives

80Λ ≤u, log(320/c) ≤u/20.

The first follows from L^{2}/4 ≥80; the second from u ≥Y^{ϵ/6}. Since u ≤40e^{u/40} and m^{2}= e^{2Λ}≤e^{u/40}, these imply

Selectable PDF notation • transcription unchecked
_{2um}^{cU}_{2} ≥4U^{ 9/10}.

Indeed split e^{u} = e^{9u/10}e^{u/20}e^{u/40}e^{u/40}, and use ce^{u/20} ≥320. The relative error bound therefore gives π(U; m, a) ≥ 2U^{9/10}. At most U^{1/2} integers, hence at most that many primes in the progression, are at most U^{1/2}. Removing them leaves

Selectable PDF notation • transcription unchecked
|P(m, a, U)| ≥2U^{9/10}−U^{1/2}≥U^{9/10}.

Every threshold was independent of m, a, U. This is JacobiCharacterSums/MSqCount.lean:86--387.

Lemma 3.29 (One exceptional squarefree twist). Assume the three progression and zero-free inputs stated above. For fixed ϵ, A, B > 0 there is a threshold such that every progression (45) admits an integer s_{∗}≥0 for which ||||||

Selectable PDF notation • transcription unchecked
∑ (^{q})||||||≤Y^{−B}|P(m, a, U)| (57)
q∈P(m,a,U)^{s}

for every odd squarefree 1 < s ≤Y^{A}, coprime to m, except possibly s = s_{∗}.

Proof. Put Z = mY^{A}. Call such an s exceptional if the chosen pair for modulus ms satisfies β_{ms} > 1 −c_{0}/ log Z, and its character restricts to ( · /s) on the units congruent to 1 modulo m. At most one s is exceptional. In fact the Page assertion makes the two exceptional characters agree on common units. If s_{1} ̸= s_{2}, choose a prime dividing just one of these squarefree integers; by the Chinese remainder theorem choose an integer congruent to 1 modulo m, a quadratic nonresidue at that prime, and 1 at all other primes dividing s_{1}s_{2}. Its two Jacobi values differ, a contradiction. Choose that exceptional integer as s_{∗} if it exists, and otherwise put s_{∗}= 0.

Fix a nonexceptional s. Let 𝒞 be the unit residue classes modulo ms congruent to a modulo m. The sum of (b/s) over 𝒞 is zero: multiply the classes by a unit v ≡1 (mod m) with (v/s) = −1, constructed by the same Chinese remainder argument. Likewise the sum ∑_{b∈𝒞}(b/s)χ_{ms}(b) is zero unless χ_{ms} agrees with the Jacobi character on the units v ≡1 (mod m). Indeed multiplication by any such v multiplies the sum by (v/s)χ_{ms}(v); if the two characters do not agree, this factor differs from 1.

In the agreement case nonexceptionality gives β_{ms} ≤1 −c_{0}/ log Z. With u = log U, L = log Y , put

Selectable PDF notation • transcription unchecked
r = 4u e^{−c0u/ log Z}, k = |A_{0}|(e^{−c3u/ log(ms)}+ e^{−√c3u}).

For large Y , r, k ≤1/2, the range condition in (54) holds for ms, and

Selectable PDF notation • transcription unchecked
4r ≤Y^{−B}/8, 2k ≤Y^{−B}/8, e^{−2u/5}≤Y^{−B}/4. (58)

Here are uniform estimates justifying those choices. Admissibility and s ≤Y^{A} give

Selectable PDF notation • transcription unchecked
L^{2}

log Z = log m + AL,_{4} log m ≤u, log m ≤4e^{ϵL/5}, log u ≤(ϵ/5 + 2)L.

The lower bound u ≥Y^{ϵ/6} dominates every fixed multiple of L^{3}. Combining it with the second inequality makes u/ log Z and u/ log(ms) at least L^{2}/8. Thus the first two error expressions are bounded by fixed multiples of

Selectable PDF notation • transcription unchecked
_{e}(ϵ/5+2)L−c_{0}L^{2}/8_{,e}−c_{3}L^{2}/8_{+ e}−√c_{3}u_{,}

each o(e^{−BL}). Also log log(ms) = O_{ϵ,A}(L), so the range condition follows from u ≥ (L^{2}/8) log(ms). These comparisons prove all three inequalities in (58) uniformly.

If the agreement case occurs, monotonicity of the logarithmic integral gives

Selectable PDF notation • transcription unchecked
| li_{2}(U^{βms})| ≤r li_{2}(U).

For example li_{2}(U) ≥U/(2u) for u ≥4, and li_{2}(U^{β}) ≤2U^{β} for 1/2 ≤β < 1; these yield the displayed bound. The exceptional pair’s definition ensures β > 7/8, so it lies in this range. It follows, using the two zero sums proved above, that the absolute Jacobi-weighted sum of the main terms on 𝒞 is at most

Selectable PDF notation • transcription unchecked
2r∑|M_{ms}(b, U)|.
b∈𝒞

In the agreement case this follows from |M_{ms}(b, U)| ≥ li_{2}(U)/(2φ(ms)); otherwise that weighted sum is exactly zero. The relative error bound also implies

Selectable PDF notation • transcription unchecked
∑|M_{ms}(b, U)| ≤2(|P(m, a, U)| + U^{1/2}).
b∈𝒞

Indeed each nonnegative prime count is at least half its main term when k ≤1/2, and the union of the classes counts precisely the progression primes coprime to s. Their number is no larger than all progression primes. Summing the errors and then removing the primes at most U^{1/2} therefore bounds the desired sum by

Selectable PDF notation • transcription unchecked
(4r + 2k)(|P(m, a, U)| + U^{1/2})+ U^{1/2}.

Primes dividing s have Jacobi value zero and do not create an additional term. By Lemma 3.28, U^{1/2} ≤ e^{−2u/5}|P(m, a, U)|. The three bounds (58) now make the last display at most Y^{−B}|P(m, a, U)|. This is the full exceptional-character cancellation argument of MeanSquareBounds/twistK.lean:148--353 and JacobiCharacterSums/MSqTwistB.lean:54--240.

Proposition 3.30 (The required uniform mean square). Assume the progression and zero-free inputs of the preceding subsection. The mean-square assertion in (46) holds with an absolute constant, for all the stated bounded sequences κ.

Proof. Put N = ⌊(Y^{K})^{400}⌋, w = Y^{ϵ/5}, and P = P(m, a, U). For large Y , N^{2} ≤Y^{800K} < U^{1/2}, since log U ≥Y^{ϵ/6} dominates 1600K log Y . Thus every q ∈P exceeds N^{2}. Apply Lemma 3.29 with A = 800K, B = K + 1. It gives cancellation with η = Y^{−(K+1)} for every squarefree odd 1 < s ≤N^{2} coprime to m, with at most one exception s_{∗}. If the exceptional integer is not greater than 1, replace it by 2; this can only enlarge the set of terms on which no cancellation is used.

Fix Y ≤y ≤Y^{K} and write

Selectable PDF notation • transcription unchecked
∑

c(n) = Λ(n) min(1, y/n), s_{p} = c(n).

Selectable PDF notation • transcription unchecked
n≤N

n a power of p

Only prime powers can have c(n) ̸= 0. Let

Selectable PDF notation • transcription unchecked
R_{m} = ∑ c(n).
n=p^{j}≤N

j even or p|m

For an integer v > 0, let s_{out}(v) be the product of primes not dividing m whose exponent in v is odd. Quadratic reciprocity, and the fixed residue a modulo m divisible by 8, give a fixed sign ε_{n,n}′ ∈{±1} (independent of q ∈P) such that

Selectable PDF notation • transcription unchecked
∑(n) (n^{′})_{= ε′}∑( q )_{,} ||||||∑(n) (n^{′})||||||_{=}||||||∑( q )||||||_{.}
_{q∈P} q q^{n,n}_{q∈P} s_{out}(nn^{′})_{q∈P} q q_{q∈P} s_{out}(nn^{′})

To check this, remove even prime exponents from nn^{′}. The remaining factors at primes dividing m, including the factor 2 and the quadratic reciprocity sign for s_{out}(nn^{′}), depend only on q mod m (since 8 | m); their product is a fixed sign ε_{n,n}′ ∈{±1}. All other odd factors become (q/p). No factor vanishes, because nn^{′}≤N^{2}< q. Factoring out ε_{n,n}′ proves the first equality, and taking absolute values proves the second.

Expand the square in the left side of (41), move its finite sums, and use |κ(n)κ(n^{′})| ≤1. For s_{out}(nn^{′}) /∈ {1, s_{∗}}, use the squarefree-twist bound; for the other pairs use |P|. The resulting upper bound is

Selectable PDF notation • transcription unchecked
{{∑ (∑ )2}}
s^{2}_{p}+ R_{m}^{2} + 2B_{0}^{2}+ 2B_{0}R_{m}+ η( c(n)) |P|, (59)
{ p n≤N }

where B_{0} is any common upper bound for all s_{p}. Here is the pair classification behind its first four terms. If s_{out}(nn^{′}) = 1, either the two prime powers have the same underlying prime, contributing at most ∑_{p}s_{p}^{2}, or both belong to R_{m}, contributing at most R_{m}^{2}. For the fixed value s_{out}(nn^{′}) = s_{∗}> 1, if it is prime, either that prime occurs in both powers or one power has trivial outside squarefree part; the two orders give at most 2B_{0}^{2}+ 2B_{0}R_{m}. If it is not prime, its nonzero contributions must come from two distinct primes whose product is s_{∗}. Those primes are uniquely determined, and the two orders contribute at most 2B_{0}^{2}. These cases also cover zero coefficients. They prove (59) without assuming that κ is multiplicative.

We give the scalar estimates uniformly in the cutoff N. For y ≥3, geometric summation over powers of p gives

Selectable PDF notation • transcription unchecked
∑

s_{p} ≤C_{2} log y, s_{p}^{2}≤C_{3}y log y (60)

Selectable PDF notation • transcription unchecked
p

with absolute positive constants. Indeed for p ≤y, the powers at most y contribute at most log y, and the geometric tail at most 2 log p. For p > y the entire sum is at most 2y log p/p ≤ 2 log y. Thus C_{2} = 3 is possible. To square and sum, primes at most ^{√}y give at most 9^{√}y(log y)^{2}≤18y log y. For ^{√}y < p ≤y, use 9(log y)^{2} ≤18 log y log p and ∑_{p≤y} log p ≤3y. For p > y, summation by parts using this same last bound gives

Selectable PDF notation • transcription unchecked
∑ ∫ ∞

(log p)^{2} 2 log v −1_{dv ≤9}log y

Selectable PDF notation • transcription unchecked
_{p>y} p^{2≤3}_{y} v^{2} y^{.}

Its contribution is at most 36y log y. Thus one may take C_{3} = 108. These are the three ranges used in MeanSquareBounds/SPsqLe.lean.

The even prime powers in R_{m} are bounded by

Selectable PDF notation • transcription unchecked
∑ ∑

Λ(j) min(1, y/j^{2}) ≤3 min(1, y/j^{2}) ≤9^{√}y.

Selectable PDF notation • transcription unchecked
j≤N j≤N

The middle inequality is finite summation by parts with the decreasing weight, using ∑_{j≤v}Λ(j) ≤3v. The last follows by splitting at ^{√}y and comparing the j^{−2} tail with its integral. By (47), primes dividing m and at most N are at most w. There are at most w of them, each contributing at most C_{2} log y. Consequently, with an absolute C_{1} > 0,

Selectable PDF notation • transcription unchecked
√

R_{m} ≤C_{1}(^{√}y + w log y) ≤5C_{1} y.

The last step uses ϵ ≤ 1, w ≤ y^{1/5}, and log y ≤ (10/3)y^{3/10}, so w log y ≤ 4^{√}y. This is the even-power/modulus-prime split in MeanSquareBounds/RRLe.lean:9--99. Finally the Mertens estimate above gives, for Z ≥2, ∑

c(n) ≤C_{4}y log Z

Selectable PDF notation • transcription unchecked
n≤Z

with an absolute C_{4} > 0.

Take B_{0} = C_{2} log y. Since log y ≥1 and log y ≤2^{√}y, the first four terms in (59) are at most

(C_{3} + 25C_{1}^{2}+ 4C_{2}^{2}+ 20C_{1}C_{2})y log y.

For large Y , (400K log Y )^{2}≤Y ; hence the final term is at most

Y^{−(K+1)}C_{4}^{2}y^{2}Y ≤C_{4}^{2}y ≤C_{4}^{2}y log y,

because y ≤ Y^{K}. Substitution in (59) proves (41) with C_{m} = C_{3} + 25C_{1}^{2}+ 4C_{2}^{2}+ 20C_{1}C_{2} + C_{4}^{2}. All constants are absolute; only the eventual threshold depends on ϵ, K. This is the finite Gram expansion and scalar conclusion of MeanSquareBounds/MeanSqAbs.lean:8--58 and meanSqK.lean:51--221.