III. Heights and the square-root exponent

The arithmetic problem is to bound a weighted prime-power sum by its main term plus a controlled error. The proof compares the slope of the same row lattice with the slopes of a family of column lattices. At a fixed translate, the columns are counted with their actual ranks. The resulting inequalities can then be averaged with equal weights over translates. For the fixed-level extension we use the pointwise version before averaging over the subgroup. All source references are to the release at commit e7c8972c701d3396947808ab20cc16dc14623d33. Word and translate means · Mean and pointwise bounds · Fixed-level comparison

In rStation paper 192 v2, span_of_dilates derives the CM balance identity and explains how a hypothesized bound on the arithmetic gap yields the prime estimate at one host. Its Theorem 4.9 uses the gap bound, Proposition 4.12 gives the per-host implication, and Corollary 4.24 states the window conclusion with the host, compatibility and scale hypotheses of §4.6. The additional derivation begins in §1, where the finite-place indices give the adjusted column slopes. Section 2 compares their rank-weighted mean with the slope of the common dual row lattice through the integral period maps and the frame rank estimates. The signed height estimate cancels the common divisor term, and §3 uses the finite-prime loss bounds to obtain the same bracket. The comparison is conditional on fitted data at the same cutoff. Sections 4–6 use the shared arithmetic readout and square-root normalization, with a uniform mean over translates also available in the ordinary case under mean loss bounds. Paper 192 v2, Theorems 4.7 and 4.9 and Propositions 4.12–4.13 · Rank filtration · Gain comparison · Pointwise gap bound

The integrated refinement in §2 gives a linear spread term and a centered Gram error. With the general size bounds, §§3, 5 and 6 give a one-sided estimate of order x1/2+νε in the ordinary and fixed-level cases, where ν=max⁡(3,2+eL,21/8+eD). The original comparison also appears throughout.

Proposition (Integrated frame bound)

Fix one label α with a finite nonempty word set Wα. Over a totally complex number field K, let Ni and the common target C=R∨ be Hermitian lattices of positive K-ranks di and Dret, respectively. Use the normalized slopes of §1, with μ the degree divided by rank and μmin⁡, μmax⁡ the extremal Harder–Narasimhan slopes. Let Bi be K-linear maps of the ambient spaces with Bi(Ni)⊂C. For each i take a real hi with hi≥∑v∣∞[Kv:R]log⁡∥Bi∥v2/[K:Q] whenever Bi≠0.

At a common embedding suppose there are matrices Ai:Cdi→CDall with ΠAi=Y−1(Bi)CYi, where Π projects onto the retained Dret rows, Y is a single common invertible target change, and the Yi are invertible source changes. Assume Dret≥(1−δα)Dall>0, 0≤δα<1, 0≤eα≤1/4, and ηα≥0. For positive ci, set QA=∑iciAiAi∗ and assume ∑i∥ciAi∗Ai−Idi∥HS2≤eα∑idi,tr⁡(QA2)≤(1+ηα)(tr⁡QA)2Dall, with Hilbert–Schmidt norms in the frame coordinates.

Write λij for the Harder–Narasimhan slopes of Ni minus hi, indexed by 1≤j≤di with each slope repeated according to its graded K-rank. Put s=∑i∑j=1diλij∑idi=∑idi(μ(Ni)−hi)∑idi,u=max⁡i(μmax⁡(Ni)−hi)−s. If H∗>0 satisfies s−μmin⁡(C)≤H∗, then (1−δα)(s−μ(C))≤3u+(ηα+eα1−eα)H∗. This is (F), proved in §2. If q(t) is the fraction of adjusted column slopes strictly below s−t, its integral on [0,H∗] is the rank mean of min⁡{H∗,(s−λij)+}, where y+=max⁡(y,0). The exact mean s makes the mean of the full deficits (s−λij)+ equal to the mean of the excesses (λij−s)+, bounded by u. This also gives zero integrated source deficit when u=0.

1. The column and row lattices

Fix a finite nonempty set U and, for each α∈U, a finite nonempty word set Wα. A translate depends on α. In the full translate calculation U=H, a finite abelian group. In the fixed-level calculation there is a map tr:U→H whose image contains the required subgroup. Write m≥1 for the number of lifts, TT for the set of tops, nT=∣TT∣, at≥0 for their sizes and A=∑tat. The common size parameters are x,ε,a,b,b′,k,L,La,B0,Nc,X′′. Here a is the reference size, b the reference thick weight, b′ the thin scale, k a size parameter for the lattices, and L,La logarithmic bounds. Fix f>0 and positive weights wαjρ for 1≤j≤m and 1≤ρ≤f. Scalar data · Lattice specialization · Translate coverage

At a fixed α, the coefficient field K is totally complex, and each lattice is a finitely generated OK-module spanning its ambient K-space, with a positive Hermitian metric at every infinite place. Put nK=[K:Q]. In the release's normalization the slope μ(M) is 2slope⁡(M)−log⁡2+nK−1log⁡∣disc⁡(K)∣, where slope⁡(M) is the negative log of the Euclidean covolume divided by the rational rank. Apply the same affine change to the largest and smallest Harder–Narasimhan slopes. The normalized degree is deg⁡M=rk⁡K(M)μ(M). For a line az its degree is −2nKlog⁡Na−1nK∑v∣∞[Kv:R]log⁡∥z∥v2. Classical: the determinant-line formula for degree (with the displayed normalization) and duality of Harder–Narasimhan slopes give μ(R∨)=−μ(R) and μmin⁡(R∨)=−μmax⁡(R) with dual metrics. Write C=R∨ for the common row lattice, realized isometrically in the retained row coordinates. Slope and map-height conventions · Duality and row covolume

For each word i∈Wα there is a full column lattice Mifull, its version Mi after restriction at the type prime Ti, and a sublattice Ni0 of a fractional-ideal multiple of Mi. The retained rank di=rk⁡KMi=rk⁡KNi0 is strictly positive, and Di=rk⁡KMifull. For each finite place w, set ℓw=log⁡Npw/nK, let ew be the ramification index, and use lw=ewℓw. The finite set S of contributing places contains every place over each rational prime it meets. Given finitely supported integral τw,σiw, set Iν=∏wpw−νw for either shift and put Ni=IσiNi0,Ni0⊂Iτ−1Mi,λiw=length⁡OK,w((Iτ−1Mi/Ni0)w). The use of powers with a negative exponent in Iν fixes the direction of all the slope shifts. The maps Bi are integral from Ni to C, and hi bounds the height ∑v∣∞[Kv:R] log⁡∥Bi∥v2/nK for Bi≠0. Supplied by the geometric papers: the lattices, the integral maps, and the restriction and local estimates below are for one and the same family. Family and local hypotheses · Division and map height

Put τw∗=τw/ew, siw∗=σiw/ew and λiw∗=(λiw+diτw)/ew. Let gjw be the local gain, Fjw the common divisor term, and Eijw the type prime exponent term. The family supplies Siw=∑j(gjw+Fjw−Eijw),siw∗≤Siw,G=2∑w∈Slw(∑jgjw−τw∗),Fo=2∑w∈Slw∑jFjw,Ei=2∑w∈Slw∑jEijw,Ri=2∑w∈Slw(Siw−siw∗). Here G is the source's Ga, and Ei its Ee. Both G and Fo depend on the translate alone. Write IO=∑p∈p(S)log⁡p ι(p) for the other finite-place charge, with ι(p)≥0, and T=2∑wlwτw∗. The identity ∑w∣plw=log⁡p follows from ∑w∣pewfw=nK, with fw the residue degree. Local definitions · Arithmetic identification

Positive-rank comparison (proved here from the local inputs). A full-rank inclusion over a discrete valuation ring has determinant valuation equal to the length of the quotient. Applying this at every finite place, and using the ideal twist formula for degree, gives μ(Ni)=μ(Mi)+Silat−Bilen,μmax⁡(Ni)≤μmax⁡(Mi)+Silat,Silat=2∑wlw(siw∗−τw∗)=G+Fo−Ei−Ri,Bilen=2di∑wlw(λiw∗−diτw∗). For the maximum slope, each sublattice of Ni lies in IσiIτ−1Mi. At a fixed rank its saturated enlargement in that lattice raises degree by the nonnegative index term, while the ideal twist moves every slope by Silat. This proves the second line. The positive di is also necessary in the first line. For the floor shift σiw=⌊ewSiw⌋, the contribution to Ri is exactly 2ℓw{ewSiw}, including when Siw<0. Length and slope conversion · Local gain identity · Conversion to place sums · Prior height calculation

Supplied by the geometric papers. For all words, Ti≥4f, Ti>2, T≤b2, and τw∗≥0. At w∣Ti one has τw∗=0 and λiw∗≤Diλi, with λi≥0. Off these places, λiw∗≤Di(τw∗+ι(p(w))). The rank identity and the loss in passage from the full lattice are diDi=(1−1Ti−1)f,μ(Mi)=μ(Mifull)−flog⁡(Ti−1)Ti−2,μmax⁡(Mi)≤μmax⁡(Mifull). The estimates also give 4fb2/Ti+flog⁡(Ti−1)/(Ti−2)≤1, 4λi≤cSl1nTm(a/k+b′), Ri≤cSl1X′′, and cSl1≥4. Restrictions and length bounds · Size majorants

Indeed di/Di=1−ξi with 0≤ξi≤f/(Ti−1)≤2f/Ti≤1/2, by Bernoulli's inequality. The part of Bilen at Ti is at most 4λilog⁡Ti. The other part is at most 2IO/(1−ξi)+ξiT/(1−ξi)≤4IO+4fb2/Ti. With the preceding loss of slope this proves the following inequalities, where μM,i,μM,imax⁡ refer to Mifull, and μN,i,μN,imax⁡ to Ni: μN,i≥μM,i+G+Fo−Ei−Loi,μN,imax⁡≤μM,imax⁡+G+Fo−Ei+cSl1mB0,Loi=cSl1IO+cSl1nTm(a/k+b′)log⁡Ti+cSl1X′′+1. The upper inequality holds already with zero in place of cSl1mB0 under the local hypotheses just used. We keep the released scalar allowance. Positive-rank scalar comparison · Transfer of the bounds · Word loss

The selected family packages the same calculation as a place-sum input. In that notation the length contribution at Ti is supplied directly with bound PT,i≤cSl1nTm(a/k+b′)log⁡Ti, the nonnegative rounding charge is at most cSl1X′′, and the weighted sum of the other local allowances before dividing by the retained share is 2IO. The bound 0≤ξi≤2f/Ti≤1/2, the restricted-slope comparison, and the allowance of 1 give exactly the same Loi by the preceding argument. The geometric supply for this route can give these place-sum hypotheses directly. Selected place estimates · Selected scalar bounds

2. The slope bound and the main term

The constants cDiv,cE,cH,cCount,cSlope,cSl1,cLoss,cHH,eDiv,eLoss are nonnegative. Write γ for Euler's constant and κ=Re⁡ℓ(χC) for the centering constant of the first sign character defined in §4. For positive v, set ω=γ+log⁡(π/v)−κ, and set p(at)=atlog⁡at−at, continuously at zero. Define SDiv=m∑t(p(at)−atω),EDiv=cDivnTaLa+cDivm(B0+Nc)xeDivε,J=cCountLa(1+m2/k),Λα=∑j1f∑ρlog⁡(wαjρv/π)+m ℓg−J,PLoss=cLossnTmbxeLossε. The number ℓg is the sum of the logarithms of the two comparison primes g0,g1 in the constructed family. These are common across words, whereas Ti is the prime used for the restriction of word i. The live constants set cDiv=cE=cH, with the values of cCount,cSlope,cSl1,cLoss,cHH furnished by their respective estimates. Main and losses · Prime and count formulas · Choice of constants

Supplied by the geometric papers. At every translate and for every word, the following inequalities hold. In the pairwise inequalities both words belong to the same Wα. μ(C)+hi−Fo≤SDiv−Alog⁡Ti+EDiv,∣Ei−mAℓg∣≤cEmB0,∣(hi+Alog⁡Ti)−(hi′+Alog⁡Ti′)∣≤cH(nTaLa+mB0),Λα≤log⁡Ti≤cCountmL,∣log⁡Ti−log⁡Ti′∣≤cCountlog⁡k,−cSlopenTaLa(1+m/k)≤μM,i,μM,imax⁡≤cSlopenTaLa. The first inequality is a bound for the dual row slope and the map height. In the divisor assembly the upper bound for the base slope contains +2F in the notation of that assembly. The comparison subtracts 2F on the left, and the scalar Fo is precisely that 2F. The same finite sum entered the column comparison with +Fo. At the divisor stage the −∑tatlog⁡Ti term is the type-prime part of the height. The other top contributions are obtained from p(at)−atω once per lift. The geometric papers owe the estimates of the local indices, the Gram determinants, and the archimedean and factorial terms that give these bounds. Divisor assembly · Normalization and transfer · Height and full-lattice slope bounds

The algebra of that first inequality can be seen before the error is absorbed. In the assembly write the height as a type-prime term plus the lift heights, less the archimedean main term and the covariance normalization. The latter normalization cancels the same term in the upper bound for the base dual slope. For each lift and position ρ, the lower estimate of the archimedean term is subtracted from the lift height's factorial term, leaving at most 1ρ∈TTp(aρ) plus its error. The degree multiplying −ω is 1ρ∈TTaρ+k∗wjρ+Δjρ in the assembly, where k∗ is its degree parameter. The corresponding base slope term is +ω pjwjρ. Their difference leaves −1ρ∈TTaρω and costs at most ∣ω∣(∣pj−k∗∣ ∣wjρ∣+∣Δjρ∣). Thus the exact remaining majorant there is the sum of the base-to-row and base errors, the lift-height errors, the archimedean and type-prime errors, the factorial comparison errors, and those two degree discrepancies. The geometric input bounds that sum by EDiv. Assembly with each error · Selected height and count inputs · Prior reading of the signed height

Supplied by the geometric papers. At each translate take matrices Ai:Cdi→CDall and positive scalars ci. The projections of Ai to the retained rows are the matrices of Bi at a common embedding in fixed invertible changes of coordinates. Write Dret for the dimension of this retained space, with Dret≥(1−δα)Dall>0. For QA=∑iciAiAi∗ the frame conditions are ∑i∥ciAi∗Ai−Idi∥HS2≤eα∑idi,tr⁡(QA2)≤(1+ηα) (tr⁡QA)2/Dall. The Hilbert–Schmidt norm is the square root of the sum of squared matrix entries in orthonormal bases. The matrix identity is ΠAi=Y−1(Bi)CYi, with retained projection Π, invertible source changes Yi, and a single invertible target change Y common to the words. These changes can be nonisometric. In particular Dret=rk⁡KC. The parameters δα,ηα,eα lie in [0,δˉ], [0,ηˉ], [0,eˉ], respectively, with δˉ<1 and eˉ≤1/4. Frame operators · Realization and inequalities

Define the rank mean and the spread ⟨Z⟩α=∑i∈WαdiZi∑i∈Wαdi,s=⟨μN,i−hi⟩α,u=max⁡i(μN,imax⁡−hi)−s. Then u≥0. The geometric height input also gives H∗>0 with s−μmin⁡(C)≤H∗≤xcHHb. For t∈[0,H∗], let q(t) be the fraction of all column slope dimensions with adjusted slope <s−t and p(t) the fraction of retained row slope dimensions with slope <s−t. These fractions can be taken in K-dimensions because the same factor nK converts them to rational ranks. Live mean and spread · Scalar height window

Proved here. The high Harder–Narasimhan step of Ni at s−t+hi maps into the high step of C at s−t. Indeed the minimum slope of the domain step is at least s−t+hi, and the maximum slope of the quotient of C by its high step is strictly less than s−t. A nonzero composite would give a positive-rank quotient of the domain step mapping injectively to a saturated sublattice of the target quotient. The two induced metrics make the composite's squared-norm height at most hi, since the subspace uses the restricted metric and the quotient uses the minimum norm of a lift. The determinant inequality bounds the slope of the domain quotient above by the slope of the saturated image plus hi, contradicting the two slope bounds. This uses the classical Harder–Narasimhan characterization of extremal slopes by saturated sublattices and quotients. The inclusion holds also when the high step or the composite map is zero. Slope and height conventions · Threshold inclusion in the release

At the chosen embedding let Pi be the orthogonal projection onto the transported high column step. Put Dcol=∑idi, Qhigh=∑iciAiPiAi∗, and T=tr⁡QA. Cauchy–Schwarz and the Gram bound give T≥(1−eα)Dcol>0. In fact the trace pairing of the Gram deviations with the identities has modulus at most eαDcol. Pairing instead with I−Pi bounds the discarded trace by Dcol(q(t)+eαq(t)). Thus the discarded trace fraction rt=1−tr⁡(Qhigh)/T is at most 2(q(t)+eα). Since 0≤Qhigh≤QA, tr⁡(Qhigh2)≤tr⁡(QA2). The trace Cauchy–Schwarz inequality gives rk⁡(Qhigh)/Dall≥(1−rt)2/(1+ηα). The image of Qhigh is the span of the high column images. Its codimension cannot increase under projection to the retained rows. The image inclusion just proved then gives the dimension estimate (1−δα)p(t)≤ηα+4eα+4q(t). Here the bound on the ambient codimension fraction uses 1−(1−rt)2/(1+ηα)≤ηα+2rt. Threshold comparison

Proved here. The rank mean of the adjusted slopes is s, and every adjusted slope is at most s+u. For t>0 the mean is at most q(t)(s−t)+(1−q(t))(s+u), so q(t)≤u/(u+t). Since every row deficit below s is at most H∗, s−μ(C)≤∫0H∗p(t) dt. Integrating the dimension bound gives a loss at most 4ulog⁡(1+H∗/u)+(ηα+4eα)H∗1−δα. The first term is zero at u=0, when every adjusted slope is s. This implies the source's scalar comparison, which also allows log⁡∣Wα∣ as a nonnegative loss. Let X=xcHH, and suppose u≤U∗. The inequality log⁡z≤z−1 applied to (1+H∗/u)/(1+X) gives ulog⁡(1+H∗/u)≤ulog⁡(1+X)+H∗1+X−uX1+X≤(U∗+b)log⁡(1+X). The last step uses H∗≤Xb and X/(1+X)≤log⁡(1+X). The same bound holds at u=0. Scalar frame comparison · Logarithmic estimate

Integrated refinement of the frame bound

The “Pointwise bracket” section of Error powers for the same ordinary family integrates the actual source deficits. The “Variance and edges” section of Frame energy and the HN slope bound centers the Gram operator at its actual mean trace. The proof of (F) uses both ideas at the filtration thresholds. The high step at a normalized threshold z uses the last grade of slope at least z, with index zero if no grade qualifies. The corresponding Euclidean threshold is (z−cK)/2, with cK=−log⁡2+nK−1log⁡∣disc⁡K∣. Thus ties belong to the high step and both low counts are strict. Pointwise bracket · Variance and edges · Rank bound in the release

Fix t∈[0,H∗]. Use q=q(t), r=rt, and c=codim⁡(im⁡Qhigh)/Dall. Put Gi=ciAi∗Ai, g=T/Dcol>0, e=eα, and η=ηα. Expanding the square and using ∑itr⁡Gi=T gives ∑i∥Gi−gIdi∥HS2=∑i∥Gi−Idi∥HS2−(1−g)2Dcol≤(e−(1−g)2)Dcol. The blocks (Idi−Pi)−qIdi have total squared Hilbert–Schmidt norm Dcolq(1−q). Their trace pairing with the centered Gram blocks is T(r−q), since the sum of the centered Gram traces is zero. Divide Cauchy–Schwarz by T and use e−(1−g)2g2=e1−e−(g−(1−e))2(1−e)g2 to get r≤q+e′q(1−q), where e′=e/(1−e).

Let J⊥ project orthogonally onto the complement of the high frame span. Since Qhigh is supported on that span and Qhigh≤QA, we have tr⁡(QAJ⊥)≤rT. The trace-zero operator QA−(T/Dall)I has squared Hilbert–Schmidt norm at most ηT2/Dall. Pairing it with J⊥−cI, of squared norm Dallc(1−c), shows that c−ηc(1−c)≤r. The high column spaces are transported by Yi−1 before forming their orthogonal projectors. Their images after row projection lie in the high step of C transported by the common Y−1. Scalar extension preserves rank, and row projection reduces the dimension by at most Dall−Dret. Hence p(t)Dret≤cDall. These dimension comparisons hold for the given invertible basis changes.

Use 2ηc≤c+η and 2e′q≤q+e′ to obtain (1−δα)p(t)≤c≤2r+η≤3q+η+e′. An adjusted column slope λ contributes min⁡(H∗,(s−λ)+) to the integral of the strict low indicator. The rank mean of (s−λ)+ equals that of (λ−s)+, which is at most u. The integral of p(t) is the mean positive row deficit, since s−μmin⁡(C)≤H∗. Integrating gives (1−δα)(s−μ(C))≤3u+(ηα+eα1−eα)H∗.(F) This includes zero spread and empty steps.

Subtract hi from the lower column bound. The divisor inequality gives Fo−hi≥μ(C)−SDiv+Alog⁡Ti−EDiv, which cancels the common divisor term. Apply the lower bound for the full-column mean, the upper bound Ei≤mAℓg+cEmB0, and log⁡Ti≥Λα, with A≥0. For the main term use the exact identity Mα=∑t,jat(log⁡at−1−γ+κ−1f∑ρlog⁡wαjρ),SDiv+mAℓg−AΛα=Mα+AJ. In the second line mAℓg cancels the type-prime term, and the two terms involving log⁡(π/v) cancel each other. Averaging over words with the positive weights di gives s≥G+μ(C)−Mα−K0−⟨Lo⟩α,K0=AJ+EDiv+cEmB0+cSlopenTaLa(1+m/k). Main-term identity and lower mean

To bound u, subtract the lower bound for μN,i′−hi′ from the upper bound for μN,imax⁡−hi before taking the mean in i′. Both G and Fo cancel. The variation of Ei+hi is bounded by V=2cEmB0+cH(nTaLa+mB0)+AcCountlog⁡k, by subtracting and adding the common center mAℓg and the two terms Alog⁡Ti. Thus u≤K3+⟨Lo⟩α,K3=cSlopenTaLa(2+m/k)+V+cSl1mB0,⟨Lo⟩α≤cSl1IO+K7,K7=cSl1nTm(a/k+b′)(cCountmL)+cSl1X′′+1. The maximum over i leaves the estimate unchanged because the right side after the mean in i′ is independent of i. Variation and spread · Count bound for the mean loss

Put D∗=(1−δˉ)−1, J∗=4log⁡(1+xcHH) and E∗=(ηˉ+4eˉ)xcHHb. Substitute the bound for u into the row comparison and use the lower estimate for s. The coefficient of μ(C) is one on each side. This gives, for each α, G−Mα≤log⁡∣Wα∣+K0+⟨Lo⟩α+D∗[J∗(K3+⟨Lo⟩α+b)+E∗]. Gain comparison

Let ASKα be the prime-power sum for a translate, defined in §4. Supplied by the geometric papers: the finite-prime loss bounds control ASKα−G and IO each by PLoss, pointwise for the translates used in the fixed-level route. For the mean theorem it suffices that these same bounds hold on their two uniform means over a nonempty set U. Put MLo=cSl1PLoss+K7 and Ubar=K3+MLo. Adding ASKα−G and using the nonnegative coefficient 1+D∗J∗ of ⟨Lo⟩α gives the pointwise inequality ASKα−Mα≤Errα:=K0+log⁡∣Wα∣+PLoss+MLo+D∗[J∗(Ubar+b)+E∗]. Under mean loss hypotheses the same computation gives 1∣U∣∑α(ASKα−Mα)≤Errmean, where Errmean is the displayed expression with ∣U∣−1∑αlog⁡∣Wα∣ in place of log⁡∣Wα∣. This is the uniform translate mean when U=H. In the fixed-level route the comparison is pointwise on U, and each subgroup translate is assigned a preimage in U before averaging. Its weight in that average is independent of the number of preimages and the ranks of their words. Loss hypotheses and errors · Uniform mean · Pointwise consumer

3. Reducing the height error to a size bracket

Supplied by the geometric papers. For the same data, the size conditions used here are x≥1, 0<ε≤1, b≥1, xb≤a, nT≥1, xm≤k, 1≤k≤min⁡(a,xeK), eK≥0, m≤x1/2+ε/2, 0≤L,La≤2xε, kb′≤a, 0≤X′′≤a, and 0≤at≤2a. Also B0,Nc≥0, log⁡∣Wα∣≤(k−3)log⁡2 for each α, δˉ≤1/2, and E∗≤K0′xεnTa. Here K0′≥0 dominates the seven coefficients cDiv,cE,cH,cCount,cSlope,cSl1,cLoss. The trace and Gram errors in E∗ must be controlled after multiplication by xcHHb. The concrete scales are b=exp⁡(xε), a=xb, X′=exp⁡(xε/3), b′=a/X′, and X′′=xnX′, with n=∣H∣. The estimates below use the displayed size inequalities and permit any scales satisfying them. Error windows · Concrete scales

For the integrated refinement put Eint=(ηˉ+eˉ/(1−eˉ))xcHHb. It suffices to assume Eint≤K0′xεnTa for this route. The original budget implies it since eˉ/(1−eˉ)≤4eˉ for eˉ≤1/4. From (F) and the lower bound for s, the gap at a label is at most K0+(ASKα−G)+(1+3D∗)⟨Lo⟩α+3D∗K3+D∗Eint. The nonnegative coefficient of ⟨Lo⟩α and its bound by cSl1IO+K7 allow the two loss hypotheses to be applied pointwise or after uniform averaging. In both regimes the resulting bound is K0+PLoss+(1+3D∗)MLo+3D∗K3+D∗Eint.

Proved here. In this section write y=xε, d=xeDivε, r=xeLossε, h=1+cHH, g=1+eK, K∗=K0′. Take x≥1 with log⁡x≤y, so log⁡k≤gy and D∗≤2. Put Q0=nTa, Q1=Q0m2/k, Q2=m(B0+Nc), and Q=Q0+Q1+Q2. Then m≤min⁡(x,k), mb≤a, A≤2Q0, and Q0≥max⁡(1,b,X′′,log⁡∣Wα∣). Direct substitution gives K0≤10K∗y(Q0+Q1)+K∗(d+1)Q2,PLoss≤K∗rQ0,MLo≤K∗2rQ0+4K∗2yQ1+(K∗+1)Q0,K3≤(8+2g)K∗yQ0+4K∗Q2. In the MLo bound use a/k+b′≤2a/k. In the last bound the contribution of V is bounded by 3K∗Q2+(2+2g)K∗yQ0. For the integrated refinement, 1+3D∗≤7, 3D∗≤6, and D∗Eint≤2K∗yQ0. The four coefficients of yQ0,rQ0,yQ1,(1+d)Q2 are bounded respectively by (67+12g)K∗+7, K∗+7K∗2, 10K∗+28K∗2, and 25K∗. It follows in both loss regimes that the gap is at most Eint=Cgap[(y+r)Q0+yQ1+(1+d)Q2],Cgap=(68+12g)(1+K∗)2.(G)

For the original comparison also impose 200(1+K∗)2hg≤y, with the original budget E∗. Then log⁡(1+xcHH)≤hy. The same estimates give Ubar+b≤16(1+K∗)2gyrQ, and D∗J∗(Ubar+b)≤128(1+K∗)2hg y2rQ. The remaining terms of the pointwise error are bounded by 22(1+K∗)2hg dry2Q, by the four inequalities and D∗E∗≤2K∗yQ0. Absorbing their coefficients with 200(1+K∗)2hg≤y proves both original error bounds Errα, Errmean ≤ x(eDiv+eLoss+3)ε[nTa(1+m2/k)+m(B0+Nc)]. The hypotheses of this estimate are uniform over words and translates. The threshold depends on the size constants and ε. Absorption in the release · Explicit coefficient inequality

Let n2=∣T2∣ for the set of thick positions and Δ2=n2/n. Supplied by the geometric papers: the upper bounds for B0,Nc, on the same lattice and prime data, are B0≤2nmbxε,Nc≤n2kbxεn(nx−ε). If C∗≥eDiv+eLoss+3, an upper bound on either arithmetic gap is consequently E=xC∗εB,B=nTa(1+m2/k)+m(2nmbxε+n2kbxεn(nx−ε)). This substitution uses the identification of tops, weights, centering constants and ASKα with their arithmetic versions. It is valid pointwise on any translates covered by the construction. Tied data and budgets · Bracket and mean identification · Pointwise substitution

4. From the arithmetic gap to the character sum

Let H be finite abelian, n=∣H∣, and G=(Z/2)2×H. Write Φ={((0,b2),v):b2∈Z/2, v∈H}, so f=∣Φ∣=2n. Put s0=((1,0),0), cconj=((1,1),0), and ρt(g)=cconj+g+((0,0),t),χC(((b1,b2),v))=(−1)b1,ψC(((b1,b2),v))=(−1)b2. Translation by cconj+((0,0),t) identifies G∖Φ with Φ. In Paper II's ordinary coordinates χC=Jd and ψC=Jq. A position ρ is thick at z∈H if ρ=((0,1),s+z) for some s∈T2. Coordinates · First sign and partner · Second sign and signed sum

Supplied by Paper II and the host construction. For each prime p there is an inertia subgroup Ip and a lift Frp∈G. The split exceptional set contains 2,q1 and excludes the distinguished prime d. For primes in that set both Ip and Frp lie in Φ. At d, Id={0,s0}. At every other prime inertia is trivial. For a unitary character ξ of G, write ξpr(pi) for the mean of ξ over iFrp+Ip, extended multiplicatively. For the principal character 1 this function is identically one. Let ℓ(ξ) be the primitive logarithmic derivative at 1 for ξ≠1, and use the centered sums Gξ(y)=∑n′≤yΛ(n′) ξpr(n′)(y/n′−1)+y ℓ(ξ)(ξ≠1),G1(y)=Gz(y)=∑n′≤yΛ(n′)(y/n′−1)−ylog⁡y+(1+γ)y. Here n′ ranges over positive integers, Λ is the von Mangoldt function, and Gz is the ordinary notation used in Paper II, independent of the translate z. Host inertia · Primitive coefficients · Ordinary centered sum

For a top t, At=at>0, and weight w:Φ→(0,∞), define ASK(t,At,w)=2∑p≤Np unramifiedFrp∉Φlog⁡p∑1≤i≤NiFrp∉Φ[At/pi−w(ρt(iFrp))]++log⁡d∑i=1N∑g∈(iFrd+Id)∖Φ[At/di−w(ρt(g))]+,MAIN(At,w)=At(log⁡At−1−γ+ℓ(χC)−f−1∑ρ∈Φlog⁡w(ρ)). Here [s]+=max⁡(s,0), and the unramified sum excludes the split exceptional set, whose contribution vanishes. We assume N≥At/w(ρ) for every ρ, so terms with pi>N vanish. At translate z, the sum over t,j of these ASK's is ASKα for tr(α)=z, and the sum of the real parts of the corresponding MAIN's is Mα. Denote this gap by gap(z). Prime-power weights and centering · Identification

Proved here. Character orthogonality gives f−1∑ξ∈G^ξ(g)‾ ξpr(pi)=2∣Ip∣ 1g∈iFrp+Ip. The complementary cosets are empty at split primes and have exactly one term at d for each i. This proves the prime weights in ASK from Fourier inversion, including the coefficients 2log⁡p and log⁡d. For g∉Φ put wg=w(ρt(g)). In the Fourier expansion of ASK use [At/n′−wg]+=wg[At/(wgn′)−1]+ and replace each resulting prime sum by its centered version. The principal character contributes the correction At(log⁡At−log⁡wg−1−γ). Each other character contributes −Atℓ(ξ) before summing over g. The mean of ξ(g)‾ on the complement of Φ is −1 for χC, 1 for the principal character and 0 for every other character. The centering corrections therefore add up to MAIN, including its +ℓ(χC). It follows that ASK(t,At,w)−MAIN(At,w)=f−1∑g∉Φ∑ξξ(g)‾ wgGξ(At/wg). Prime-power expansion · Centering identity

Supplied by the host and weight construction. For every t∈TT, depth j and translate used in the average, put yt=At/b. Assume x≥3, ∣yt−x∣≤1, b,b′,X′>0, b′X′≤xb, ηw,θ≥0. The weight wzj(ρ) is positive, lies in [b,b+ηw+1] at thick positions, and is at most 2b′ elsewhere. The conditions also give the cutoff N above, ∣Gξ(y)∣≤yθ for all ξ and y≥X′/3, and the two cost bounds 2f(x+1)θ≤1/x,6f(x+1)(ηw+1)≤b/x. The weight tolerance ηw and the row frame error ηα occur in different estimates. Node and cutoff hypotheses · Thin and cost bounds · Fixed-level data

Classical input and elementary bounds. Use finite abelian character orthogonality and the identity log⁡n′=∑d′∣n′Λ(d′). The elementary Chebyshev bound Ψ(Y)=∑n′≤YΛ(n′)≤3Y follows by comparing Ψ(2j)−Ψ(j) with log⁡(2jj)≤2jlog⁡2 for integer j and summing at powers of two. Indeed each prime power in (j,2j] contributes one unit to the valuation of the binomial coefficient and all other contributions are nonnegative. For Y≥1, summing the divisor identity gives log⁡(⌊Y⌋!)=∑n′≤YΛ(n′)⌊Y/n′⌋. Bounding the floors and using ∫1⌊Y⌋log⁡u du≤log⁡(⌊Y⌋!)≤Ylog⁡Y gives log⁡Y−2≤∑n′≤YΛ(n′)/n′≤log⁡Y+3. The lower bound uses ∫⌊Y⌋Ylog⁡u du≤log⁡Y. Differentiate the continuous, piecewise smooth ramps between their breakpoints. Since ∣ξpr∣≤1, one obtains for Y,Z≥1 and Cd=3+∣γ∣ ∣Gz(Y)−Gz(Z)∣≤Cd∣Y−Z∣,∣Gξ(Y)−Gξ(Z)∣≤∣Y−Z∣(log⁡max⁡(Y,Z)+Cd+∣ℓ(ξ)∣)(ξ≠1). Also wGξ(At/w) has derivative in w of modulus at most 3At/w. For ξ≠1 the derivative is the negative of the corresponding twisted Ψ(At/w); for the principal character it is At/w−Ψ(At/w). Integration between two positive weights w,w′ gives ∣wGξ(At/w)−w′Gξ(At/w′)∣≤3Atmin⁡(w,w′) ∣w−w′∣. These bounds give a choice of the absolute constant in the movement estimates. Chebyshev and weight comparison · Movement bounds in the release

Proved here. For every thick partner replace wg by b, at cost at most 3(x+1)(ηw+1) per character and position. For a thin partner At/wg≥(x−1)b/(2b′)≥X′/3, so discarding its term costs at most (x+1)bθ. There are f positions and 2f characters with a prefactor 1/f. Thus the modulus of the error in the Fourier sum, for each t,j,z, is at most Enode=6f(x+1)(ηw+1)+2f(x+1)bθ≤2b/x. Keeping the two position counts gives the sharper node bound 2[3n2(x+1)(ηw+1)+(f−n2)(x+1)bθ]≤b/x, as used in Paper II. The coarser allowance Enode gives the same released exponent. The thick sum left over is independent of j. Write e1(v)=((1,0),v). The thick partners have g=e1(s+z−t) for s∈T2, so this sum is Thz,t=bf∑s∈T2∑ξξ(−e1(s+z−t))Gξ(yt). Single-node comparison and constant

Let H0≤H with T2⊂H0. Denote the characters trivial on {0}×H0 by ΨH0. Summing the character in Thz,t over z∈H0 gives zero unless ξ∈ΨH0. For each surviving ξ the factor is ∣H0∣ξ(e1(t)), since s∈H0 and the first bit is of order two. Hence 1∣H0∣∑z∈H0Thz,t=bΔ22∑ξ∈ΨH0ξ(e1(t))Gξ(yt). This gives the 1/2 in the normalizer P=Δ22 nTmb. Whenever 1∣H0∣∑z∈H0gap(z)≤E, the modulus estimate for the nodes gives mbΔ22∑t∈TTRe⁡∑ξ∈ΨH0ξ(e1(t))Gξ(yt) ≤ E+nTmEnode ≤ E+2nTmb/x. The fixed-level hypothesis of pointwise control implies the mean hypothesis in this display. Subgroup orthogonality · Subgroup comparison

In the ordinary case H0=H, the four survivors are 1,ψC,χC,χCψC, with signs +,+,−,− at s0. Suppose the three nonprincipal survivors obey ∣Gξ(x)∣≤x1/2+ε and ∣ℓ(ξ)∣≤CLlog⁡x with CL≥0, as in the weaker form of Paper II's bounds. The movement from yt to x contributes at most Rmove=Cd+3(log⁡(x+1)+Cd+CLlog⁡x). Before using the upper gap inequality, the Fourier calculation gives the two-sided comparison used for the ordinary normalization in Paper II: ∣Gz(x)−∣H∣−1∑z∈Hgap(z)P∣≤∑ξ∈{ψC,χC,χCψC}∣Gξ(x)∣+Rmove+nTmEnodeP. With ε>0 fixed, Rmove≤x1/2+ε eventually. In particular P(Gz(x)−4x1/2+ε)≤E+2nTmb/x. Four-character signs and explicit bound · Ordinary translate identification

5. Division by the normalizer

Supplied by the geometric papers. At the chosen cutoff let a≤2xb and assume the additional size relations nT≥12nx−ε/8,x−ε≤Δ2≤1,x1/2−ε≤m≤x1/2+ε/2,xm≤k≤n2x1/2−ε. Then Qb=nTmb>0 and P=QbΔ2/2>0. Proved here: for any real Vx, if P(Vx−4x1/2+ε)≤xC∗εB+2Qb/x, division gives the exact normalized bound Vx≤4x1/2+ε+2xC∗εΔ2BQb+4xΔ2. In the ordinary case put Vx=Gz(x). With the preceding size relations, the four terms of B/Qb satisfy amb≤2xm≤2x1/2+ε,ambk≤2xmk≤2,2nmxεnT≤4x1/2+13ε/8,n2kxεn2x−εnT≤2x1/2+9ε/8. The last inequality uses n2≤n. Thus B/Qb≤10x1/2+13ε/8. Since 2/x≤1 and 1/Δ2≤xε, the normalized bound implies Vx≤26x1/2+(C∗+21/8)ε≤x1/2+(C∗+3)εif 26≤x3ε/8. The two size costs x/m and m enter before density and logarithmic allowances. At the trial scale m=xτ their exponents are 1−τ and τ, with maximum minimized at τ=1/2. The fitted range for m allows shifts of order ε in these powers. Size hypotheses · Size bound

For the integrated refinement substitute Eint for xC∗εB in the same division. The four bounds on B/Qb apply to the two parts of Q2/Qb and to Q0/Qb,Q1/Qb through the budgets in §3. With Δ2−1≤y, eDiv≤eD, and eLoss≤eL, this gives EintCgapΔ2Qb≤2x1/2+3ε+2x1/2+(2+eL)ε+2x2ε+12x1/2+(21/8+eD)ε. For ν=max⁡(3,2+eL,21/8+eD) and x≥3, we obtain Vx≤(36Cgap+6)x1/2+νε, including the other two terms in the division. In the ordinary case this holds for Vx=Gz(x) once log⁡x≤xε and Rmove≤x1/2+ε, where Rmove is defined in §4.

6. The fixed-level window input

The release also needs a bound for one fixed family of primitive characters, even when the Artin group varies with x. Use the subgroup H0 from §4, of index o=[H:H0] fixed by the level L. Let Kkeep⊂ΨH0 be the kept characters. Supplied by the host comparison: Kkeep contains the principal character, its characters have ξ(s0)=1, and ∣ΨH0∣=4o. Put c=(3o+1)−1 and aξ=1 if ξ is trivial on H, aξ=c otherwise. For each kept character nontrivial on H, assume ∣∑t∈TTξ((0,0),t)−cnT∣≤4o x−1+εnT. The host comparison gives ∣ℓ(ξ)∣≤Ko for nonprincipal kept characters, and ∣ℓ(ξ)∣≤CElog⁡x, ∣Gξ(x)∣≤x1/2+ε for discarded characters in ΨH0. The primitive readout identifies the kept sum with GL(x)=∑ξ∈KkeepaξRe⁡Gξ(x). As the host changes this is the same function, with the primitive characters and weights fixed by L. In particular GL=Gz at the conductor-one level of Paper II. Host character and share hypotheses · Reading the fixed-level sum · Bound for discarded characters

Proved here. Let Σ(x)=∑ξ∈ΨH0∖Kkeep∣Gξ(x)∣. For any nonprincipal character the movement bound of §4 and Gξ(1)=ℓ(ξ) give ∣Gξ(x)∣≤x(log⁡x+Cd+∣ℓ(ξ)∣). For a kept character ξ, first move Gξ(yt) to Gξ(x) at each t. The remaining sum of phases is exactly nT if ξ is trivial on H. In the other case the deviation from cnT is bounded by the share hypothesis. For a discarded character the real part is bounded below by −∣Gξ(x)∣ less the movement error. For the principal character the movement costs at most Cd. Thus the sum on the left of the subgroup comparison in §4, before its factor mbΔ2/2, is at least nT(GL(x)−Σ(x)−4oR(x)), where one can take R(x)=max⁡{Cd, log⁡(x+1)+Cd+Ko+4o xε(log⁡x+Cd+Ko), log⁡(x+1)+Cd+CElog⁡x}. For fixed o,Ko,CE,ε, 4oR(x)≤4x1/2+ε eventually. Apply §5 with Vx=GL(x)−Σ(x). For eDiv≤eD, eLoss≤eL, put C∗=eD+eL+3 with eD,eL≥0. Then GL(x)≤x1/2+(eD+eL+6)ε+4o x1/2+ε≤x1/2+(eD+eL+7)ε, on increasing the threshold so xε≥1+4o. In the ordinary case the direct estimate in §5 gives the stronger exponent 1/2+(eD+eL+6)ε under Paper II's three-character bounds. One-character comparison · Size and discarded-character absorption

For the integrated refinement use the pointwise bound (G) at each subgroup translate in §4. With x≥3, log⁡x≤xε, 4oR(x)≤4x1/2+ε, the integrated estimate of §5 applies to GL(x)−Σ(x). Adding Σ(x)≤4ox1/2+ε gives GL(x)≤(36Cgap+6+4o)x1/2+νε. The cutoff conditions hold beyond one threshold for fixed ε and the uniform host constants, since R(x)=O(xε(1+log⁡x)).

Handoff to Paper II and the analytic window argument. In the notation of Paper II, E(u)=gap(u), E‾=∣H∣−1∑uE(u), and its Q is our P. The aE in its error bracket is the reference size a. The arithmetic input for its ordinary prime comparison is E‾≤xC∗εB for the host used in that comparison. The pointwise form gives inequality (9) of Paper II. For the release's fixed-level route use gap(z)≤xC∗εB for every z∈H0. Fix a level L, a window exponent λ>1, and eD,eL. For every sufficiently small ε>0 and every requested lower cutoff there must be an interval [Y,Yλ] starting above that cutoff. At every real x in that interval one has compatible host and geometric data with bounds uniform in x as above. A different host and lattice family may be selected at each x, with the same L. The release uses λ=3/2, eD=2, eL=0, 0<ε<1/52 along its geometric route, giving the one-sided bound GL(x)≤x1/2+9ε on late intervals. For any 1/2<β<1 choose ε>0 so 9ε≤β−1/2. Taking the lower cutoff beyond the thresholds in §§3–6 then gives GL(x)≤xβ throughout arbitrarily late intervals. That is the hypothesis of the analytic window theorem for the positively weighted family of primitive characters. Paper I proves the analytic implication. The existence of compatible windows is the supply hypothesis in §7. Paper I · Paper II · Window hypothesis · Host supply · Released exponents · One-point selection · Analytic consumer

7. The geometric inputs to be established in prose

The same window quantifiers in §6 apply to the integrated estimate with either the old frame budget or the allowance Eint. The exponent bounds eD,eL and the window exponent λ are fixed first. For each sufficiently small fixed ε>0, the size and host constants are uniform before requesting the lower cutoff Y0. For either ramp let Cramp be its constant from §5 or §6. For any ρ>0, take Y0 beyond the relevant movement and size thresholds and max⁡(1,Cramp)1/(ρε). This gives a unit-coefficient upper bound with exponent 1/2+(ν+ρ)ε at every real cutoff in the supplied interval. Given 1/2<β<1, first choose ε small enough that (ν+ρ)ε≤β−1/2. For eD=2,eL=0 the general size bound has ν=37/8.

Implication proved Supply needed
Local indices give adjusted column slope bounds (§1). Full and restricted lattices with the stated ranks and metrics, and length, twist and gain estimates on the actual sublattices.
Integral comparison and centered integration give (F) (§2). Integral period maps and height allowances, common-embedding frame realization, row retention and Gram and trace bounds, and the row minimum-slope window.
Signed cancellation and size bounds give the gap bound (G) (§§2–3). The signed height estimate with the same Fo, type-prime counts and slope bounds, the frame budget, and both finite-prime loss bounds, in uniform mean for the ordinary route or pointwise on labels covering H0 for the fixed-level route.
The gap gives the prime ramp bounds (§§4–6). The same host labels, tops and weights in the arithmetic identity, thick and thin estimates, size and character bounds, and, at fixed level, subgroup coverage, character shares and a fixed primitive sum.
Supplied prime bounds give the window implication via Paper I, also at discrete cutoffs for ordinary zeta (§7). Compatible data throughout arbitrarily late power intervals, with uniform constants fixed for each ε before choosing windows. For ordinary zeta, the §7 corollary also permits data at cutoffs with its logarithmic covering bound.

For the ordinary ramp or a fixed level L, supply the windows just stated for each sufficiently small ε>0. At every sufficiently large x in those windows, choose compatible data to meet all four requirements below, with the constants of §3 fixed on the family for that ε. The analytic estimates at the host must apply to the same Artin data, weights and tops. Tying the fields and data · Geometric input assembly · Fixed-level interface

  1. Columns, periods and lengths. Construct the row and column lattices, their metrics and the rational period maps. Prove that the common row is the metric dual of R, that the maps on the shifted columns are integral, that the full and retained ranks are positive with the stated restriction, and that every local length, shift and gain satisfies §1. In particular the finite index loss comes from the same sublattice Ni0 used to define Ni. The common divisor term used there must be the one in the height inequality. Place data · Period-integral lattice and shifts

  2. Frame, counts and height. Establish the single-place matrix realization with positive normalizing weights, the trace and Gram estimates with the indicated row and column dimensions, and the height bound for each nonzero map. Prove the quantitative row minimum-slope window, the count and full-column slope estimates, the type-prime compensation and the corrected-height variation in §2. In the divisor estimate use the signed paired local determinant expression of the row covolume and bound its finite and infinite terms to obtain EDiv with eDiv≤eD. This step requires the common-divisor estimate for the selected family, in addition to the formal cancellation of Fo. Frame hypotheses · Finite determinant expression · Global height input

  3. Finite-prime loss. For each translate used in the pointwise argument, establish both ASKα−G≤cLossnTmbxeLossε and IO≤cLossnTmbxeLossε with eLoss≤eL. This entails the estimates on the orbits served at eligible primes, on the net gain ∑jgjw−τw∗, and on the omitted contributions after weighting by log⁡p. It includes the distinguished ramified prime and the length charge IO. The two bounds on uniform means over H suffice for the ordinary mean route in §4. Local gain and served orbits · Pointwise loss interface · Released loss exponent

  4. Uniform supply and compatibility. For all cutoffs in the windows, prove the coverage of the required translates, the shared top and weight identifications, the size inequalities of §§3 and 5, and the thick and thin conditions of §4. In the fixed-level case include the share and primitive readout hypotheses of §6. In particular a comparison for one abstract choice of numerical slopes gives the scalar implication, and the needed existence statement requires slopes of the actual lattices built from the chosen host. Pointwise data and coverage · Compatibility equations · Quantified family supply

The geometric statements are inputs here. The height and scalar estimates above show how those statements would yield the ordinary and fixed-level prime comparison used in the series.

Ordinary zeta at discrete cutoffs

The ordinary-zeta conclusion also follows when the geometric construction supplies data at a discrete set of cutoffs in each interval. This consequence uses the explicit formula and the window theorem of Paper I. Paper I, Theorem 2 and its proof

Corollary. Fix λ>1 and 0<θ<1. Suppose there are constants A,C,δ>0 such that, for arbitrarily large Y, a set SY⊂[Y,Yλ] has both of the following properties. Every x∈[Y,Yλ] has some s∈SY with ∣log⁡(s/x)∣≤Ax−δ, and G1(s)≤Csθ for every s∈SY. Then every nontrivial zero of ζ has real part at most θ.

Proof. Write B=sup⁡ρRe⁡ρ over the nontrivial zeros. Paper I gives, with multiplicities and an absolute remainder, G1(x)=log⁡(2π)+∑ρxρρ(1−ρ)+O(x−1). The classical zero count N(T)≪Tlog⁡(2T) gives ∑ρ∣ρ∣1/2/∣ρ(1−ρ)∣<∞ by dyadic summation. For real h with ∣h∣≤1/2, integration of ρeρt and the bound on ∣eρh∣ give ∣eρh−1∣≪min⁡(1,∣ρh∣)≤∣ρh∣1/2. Thus for x,xeh≥2, ∣G1(xeh)−G1(x)∣≪xB∣h∣1/2+x−1. At a sufficiently late supplied interval choose the node s for each x and take h=log⁡(s/x). The node bound gives G1(x)≪xθ+xB−δ/2+x−1 throughout that interval. If θ<B, choose b strictly between max⁡(θ,B−δ/2) and B. On arbitrarily late windows the maximum of log⁡(1+max⁡(G1(x),0))/log⁡x is then at most b. Paper I proves that these maxima tend to B as Y→∞. This contradiction proves the corollary.

For each fixed positive integer k, one may take all s=nk in the window. Consecutive powers satisfy log⁡((n+1)k/nk)≤k/n. At either end of a sufficiently late window the first or last power inside it gives the same Ok(x−1/k) bound. For the ordinary implication one may therefore impose the common geometric hypotheses at those powers only. For each fixed ε the covering constants and the positive δ may depend on ε, provided they are fixed before choosing arbitrarily late windows. The bound with θ=1/2+νε at the nodes gives B≤1/2+νε whenever θ<1. Taking ε to zero and using the functional equation gives RH under this discrete supply hypothesis. In particular, it suffices to take ε=1/N along integers N→∞ and use s=n2N at that exponent. The two scales sε=n2 and s=nN are then integers. The power 2N is fixed while the starts of the windows tend to infinity.

One can also choose the two even ranks exactly. Fix ε=1/N below the common existence bound and use the nodes xj=(2j)16N. At these cutoffs the lower endpoints of the two rank intervals are the even integers m=(2j)8N−1=xj1/2−ε/16,k=(2j)72N+160=xj9/2+10ε. For a frame of order h with xj2+8ε≤h≤2xj2+8ε, these choices satisfy 2≤m, m+3≤k and m≤2h≤k for large j, by comparison of the displayed powers of xj. Thus the rank construction applies to these very integers. The logarithmic mesh is 16Nlog⁡(1+1/j)=ON(xj−1/(16N)), so the corollary applies on the same power windows. The auxiliary rank formula still contains a ceiling, the degree of the dyadic field must be a power of two, and the period and weight hypotheses must hold together on the chosen host. Construction of the ranks on a given frame