betaFeedback

WorksMathematics

Pointwise representation counts for Eisenstein restricted sums

Abstract

Let AA consist of the positive integers supported on primes 1 mod 31\bmod 3, including 11. We count distinct ordered representations n=a+bn=a+b for n≡2 mod 6n\equiv2\bmod6 and n=a+3bn=a+3b for n≡4 mod 6n\equiv4\bmod6, with a,b∈Aa,b\in A. In either family prescribe an interval of length HH centered at xx for one actual summand. Put M=min⁡(x,n−x)M=\min(x,n-x). For every fixed 0<ε<10<\varepsilon<1, if H≤MH\le M and H3≥n1+εMH^3\ge n^{1+\varepsilon}M, the count has order HJ(n)/log⁡nHJ(n)/\log n for every sufficiently large admissible nn. Here J(n)J(n) is the product of (p−1)/(p−2)(p-1)/(p-2) over the odd primes p≡2 mod 3p\equiv2\bmod3 dividing nn. In particular, central windows of width nθn^\theta are allowed for 2/3<θ<12/3<\theta<1, and windows at a prescribed small summand of size comparable to their width are allowed for 1/2<θ<11/2<\theta<1. The lower bound combines angular Hecke estimates with a sieve that prescribes small split factors in both summands.

The pointwise count

Let A={a∈Z>0:p∣a⇒p≡1 mod 3}\mathcal{A}=\{a\in\mathbb{Z}_{>0}:p\mid a\Rightarrow p\equiv1\bmod3\}, where pp always denotes a prime. Write 3⋅A={3a:a∈A}3\cdot\mathcal{A}=\{3a:a\in\mathcal{A}\}. We count ordered pairs (s1,s2)(s_1,s_2) of positive integers summing to nn, with

(s1,s2)∈{A×A,n≡2(mod6),A×(3⋅A) or (3⋅A)×A,n≡4(mod6).(s_1,s_2)\in\begin{cases} \mathcal{A}\times\mathcal{A}, & n\equiv2\pmod6,\\ \mathcal{A}\times(3\cdot\mathcal{A})\ \text{or}\ (3\cdot\mathcal{A})\times\mathcal{A}, & n\equiv4\pmod6. \end{cases}

In the second line one orientation is prescribed in advance. For 0<x<n0<x<n let r(n;x,H)r(n;x,H) count these pairs with ∣s1−x∣≤H/2\lvert s_1-x\rvert\le H/2. In particular, if the first summand is 3a3a, it is 3a3a that lies in the interval.

Theorem 1.1. Fix 0<ε<10<\varepsilon<1 and put M=min⁡(x,n−x)M=\min(x,n-x). Uniformly for every sufficiently large admissible nn and every real x,Hx,H with

0<x<n,1≤H≤M,H3≥n1+εM,0<x<n,\qquad1\le H\le M,\qquad H^3\ge n^{1+\varepsilon}M,

one has, in each of the prescribed families,

r(n;x,H)≍εHJ(n)log⁡n,J(n)=∏p∣n, p≥5p≡2 (3)p−1p−2.(1)r(n;x,H)\asymp_\varepsilon\frac{HJ(n)}{\log n},\qquad J(n)=\prod_{\substack{p\mid n,\ p\ge5\\ p\equiv2\ (3)}}\frac{p-1}{p-2}. \tag*{(1)}

The constants in the two bounds and the threshold for nn depend only on ε\varepsilon.

The product has a direct interpretation. At p=5p=5, for example, the two conditions 5∤s15\nmid s_1 and 5∤n−s15\nmid n-s_1 forbid two residues when 5∤n5\nmid n, and one when 5∣n5\mid n. The ratio of the allowed proportions is (4/5)/(3/5)=4/3(4/5)/(3/5)=4/3. The same calculation explains every factor of J(n)J(n) and yields an upper bound by an ordinary sieve on integers.

The lower bound begins with a larger family that can be counted spectrally. Write N(z)=zzˉN(z)=z\bar{z} for the Eisenstein norm on Z[ω]\mathbb{Z}[\omega], with ω2+ω+1=0\omega^2+\omega+1=0. The equation N(z)+N(w)=nN(z)+N(w)=n is a quaternion norm equation. At an inert prime p≡2 mod 3p\equiv2\bmod3, p∣N(z)p\mid N(z) forces z∈pZ[ω]z\in p\mathbb{Z}[\omega]. It is therefore possible to remove the large inert primes by a lattice count within the prescribed window. A separate issue is that many coordinate pairs can give the same two norm values. Sifting out small split primes bounds this multiplicity, with a baseline coordinate count of size H/(log⁡n)2H/(\log n)^2 before the target factors. To gain a factor of log⁡n\log n we allow small split primes to occur to exponent one in one summand, or in both summands when they divide the target. For instance, 7∤n7 \nmid n permits placing 77 in either summand. If 7∣n7 \mid n, its use as a common factor reduces the target to n/7n/7. The split-prime local factors for these two situations supply the same logarithmic gain. The prescribed factors can be read off the resulting values, and each pair then has bounded weighted coordinate multiplicity.

The angular estimate costs M/H\sqrt{M/H} in the Hecke remainder. Multiplying by the square-root error from the norm shell gives nM/H\sqrt{nM/H}. Comparing this with the main-term width HH explains H3≥n1+εMH^3 \ge n^{1+\varepsilon}M. The proof below treats the finite level and the angular variable together, which lets us prescribe the mixed orientation at any center.

Proof overview. Section 22 estimates the angularly weighted shell with finite congruence tests. Section 33 handles primes dividing the target, with a worked example at 727^2. Section 44 sums over the small split factors to recover the logarithm. Section 55 removes the remaining inert primes, bounds the weight of each distinct pair, and gives the matching integer-sieve upper bound.

Relation to earlier work. Dietmann and Elsholtz [2] prove eventual binary existence for the Gaussian prime restriction p≡1(mod4)p \equiv1 \pmod{4} and obtain ≫δn1−δ\gg_\delta n^{1-\delta} pairs for every δ>0\delta> 0 without restricting the summand sizes. Howe, Huffman, Kohlmeyer and Kruer [1 Theorem 1.1] prove existence for every positive target 2(mod4)2 \pmod{4}, using quaternionic norm counts and a sieve on the representing vectors. The prime sets used here and in the Gaussian problem are incomparable, as 77 and 1717 show. We use the same norm-counting principle for the Eisenstein ring. The new estimates needed here are angular localization with prescribed residue conditions, the densities when a sieving prime divides the target, and a lower bound that counts distinct value pairs at the integer-sieve scale. Ramaré and Viswanadham [4 Introduction and Theorem 1.6] discuss the split-prime Eisenstein set in their sieve framework and prove a ternary additive theorem for primitive Gaussian values.

The Frobenian correlation theorem of Loughran and Matthiesen [9 Theorem 1.9] requires pairwise independent linear parts. Fixing the total gives the one-variable forms t,n−tt,n-t, with proportional linear parts. In the mixed case, restricting to the necessary residue class and dividing one form by 33 leaves the same dependence. Freeing the total gives a two-variable sum, which averages over the slices that our theorem estimates individually.

Iwaniec’s half-dimensional sieve [10 Theorem 1] uses a half-dimensional local product and a remainder from counts in residue classes. Sieving both summands directly excludes two residues per inert prime away from the target divisors. As the inert primes have density 1/21/2, this gives sieve dimension one. A half-dimensional sieve on one summand with the other already restricted would require progression estimates for that restricted sequence in translated windows, uniformly in the target. Greaves [11 main theorem] obtains a bound of order at least n/(log⁡n)5/2n/(\log n)^{5/2} for prime pairs whose complementary value is a sum of two squares. The cutoff p,q≤n/2p,q \le\sqrt{n}/2 in that Gaussian input leaves the split of the complementary sum of two squares free. Here the angular Hecke estimate supplies the window-dependent distribution, and the split-factor sum reaches the scale of distinct values.

Corollary 1.2. The count in every fixed interval of positive length for s1/ns_1/n inside (0,1)(0,1) is of order nJ(n)/log⁡nnJ(n)/\log n. In particular the unrestricted count has this order. At any fixed interior share, windows of width nθn^\theta have count of order nθJ(n)/log⁡nn^\theta J(n)/\log n for 2/3<θ<12/3 < \theta< 1. For 1/2<θ<11/2 < \theta< 1 and fixed 0≤a<b0 \le a < b, the same count holds with s1∈[anθ,bnθ]s_1 \in[an^\theta,bn^\theta].

The results give two-sided order bounds. A leading constant and a numerical onset are open here, as are the central and small-summand endpoint exponents 2/32/3 and 1/21/2.

An angular estimate on a quaternion shell

Fix nn, xx, HH as in Theorem 1.1. We first count coordinate representations with an angular weight and prescribed local conditions.

The norm shell

Let R=Z[ω]R = \mathbb{Z}[\omega] and K=Q(ω)K = \mathbb{Q}(\omega). The Eisenstein lattice is norm Euclidean, since every complex number is within squared distance 1/31/3 of a lattice point. Write B=K+KjB = K + Kj, where j2=−1j^2 = -1 and jz=zjˉjz = z\bar{j}, and put O=R+Rj\mathcal{O} = R + Rj. Its reduced norm is nrd⁡(z+wj)=N(z)+N(w)\operatorname{nrd}(z + wj) = N(z) + N(w). The trace pairing on 1,ω,j,ωj1,\omega,j,\omega j has determinant −9-9, so the order has reduced discriminant 33. The algebra is definite and the number of ramified places is even, so it is ramified at 33 and infinity. Thus O\mathcal{O} is maximal.

Rounding both complex coordinates gives a remainder of reduced norm at most 2/32/3. In any nonzero integral right ideal, take a least positive norm element α\alpha and divide on the left by α\alpha. The remainder in that ideal has norm smaller than nrd⁡(α)\operatorname{nrd}(\alpha) unless it vanishes. Every right ideal is therefore principal, as are the fractional ideals by scaling. The unit group UU consists of the twelve elements of R×∪R×jR^\times\cup R^\times j. For 3∤m3 \nmid m, let Qm={α∈O:nrd⁡(α)=m}Q_m = \{\alpha\in\mathcal{O} : \operatorname{nrd}(\alpha) = m\}. At each ℓe∥m\ell^e \mathrel{\|} m, integral right ideals correspond locally to index-ℓe\ell^e sublattices of Zℓ2\mathbb{Z}_\ell^2, of which there are 1+ℓ+⋯+ℓe1 + \ell+ \cdots+ \ell^e. Each ideal has twelve generators of norm mm, giving

∣Qm∣=12σ(m),σ(m)=∑d∣md.(2)|Q_m| = 12\sigma(m), \qquad\sigma(m) = \sum_{d \mid m} d. \tag*{(2)}

We use this all-ideal convention throughout [3 Propositions 41.3.1 and 41.3.6]. The number of coordinates of a positive norm aa is 6f(a)6f(a), where ff is multiplicative. Its local values are f(pe)=e+1f(p^e) = e + 1 at split primes p≡1(mod3)p \equiv1 \pmod{3}, f(pe)=1f(p^e) = 1 at 33, and f(pe)=12∣ef(p^e) = 12_{\mid e} at inert primes p≡2(mod3)p \equiv2 \pmod{3}. Indeed a split prime has two conjugate factors in RR, and a choice of their exponents gives each possibility up to a unit. The ramified prime in RR above 33 is generated by 1−ω1 - \omega and has norm 33. In particular, f(a)≤τ(a)f(a) \le\tau(a), where τ\tau is the divisor function. At an inert prime, R/pRR/pR is a field, so norm divisibility by pp is equivalent to coordinate divisibility by pp. For even mm, both component norms have the same parity. Both are even precisely on the shell scaled by 22 from m/4m/4. The odd-component shell therefore has size

Ro(m)=12{σ(m)−σ(m/4)},Ro(m)m=18∏pe∥mp≥5σ(pe)pe,(3)R_o(m) = 12\{\sigma(m) - \sigma(m/4)\}, \qquad\frac{R_o(m)}{m} = 18 \prod_{\substack{p^e \mathrel{\|} m \\ p \ge5}} \frac{\sigma(p^e)}{p^e}, \tag*{(3)}

where σ(m/4)=0\sigma(m/4) = 0 if 4∤m4 \nmid m.

Finite tests and the Hecke bound

Use the finite group GG obtained by multiplying the fixed factor G3=(O/3O)×/F3×G_3 = (\mathcal{O}/3\mathcal{O})^\times/\mathbb{F}_3^\times and any factors PGL⁡2(Z/paZ)\operatorname{PGL}_2(\mathbb{Z}/p^a\mathbb{Z}) at p≥5p \ge5, with a=1a = 1 or 22. The Hecke index in this subsection is coprime to the level. Write Γ\Gamma for the image of UU in GG. The kernel on UU is {±1}\{\pm1\}, as is seen at 33. Each group has a norm squareclass map, using determinant at split quaternion places. Both squareclasses occur. At 33, for example, the norms of 11 and 1+j1 + j give the two classes. They are detected modulo pp at every odd level prime.

For k≥0k \ge0, let ρk=Sym⁡2kC2\rho_k = \operatorname{Sym}^{2k}\mathbb{C}^2 denote the unitary representation of SU⁡(2)\operatorname{SU}(2) on that space. For a real quaternion of nonzero norm use uα=α/nrd⁡(α)∈SU⁡(2)u_\alpha= \alpha/\sqrt{\operatorname{nrd}(\alpha)} \in\operatorname{SU}(2), under the usual identification of unit quaternions with SU⁡(2)\operatorname{SU}(2). The action on order units descends to Γ\Gamma. Use the space of functions F:G→Sym⁡2kC2F : G \to\operatorname{Sym}^{2k}\mathbb{C}^2 satisfying F(γg)=ρk(γ)F(g)F(\gamma g) = \rho_k(\gamma)F(g), with squared norm summed once per left coset. The all-ideal Hecke operator is

(TmF)(g)=112∑α∈Qmρk(uα)−1F(αfing).(4)(T_mF)(g) = \frac{1}{12}\sum_{\alpha\in Q_m} \rho_k(u_\alpha)^{-1}F(\alpha_{\mathrm{fin}}g). \tag*{(4)}

where αfin\alpha_{\mathrm{fin}} is projective reduction in GG. Let K\mathcal{K} be the product of the projective order-unit groups at the finite places, and K0\mathcal{K}_0 the kernel of reduction to GG. At 33 we use order units, with reduction modulo 3O33\mathcal{O}_3. A residue unit lifts to an order unit since its reduced norm is a 33-adic unit. At each p≥5p \ge5 in the level, the corresponding reduction of GL2(Zp)\mathrm{GL}_2(\mathbb{Z}_p) is surjective modulo pap^a. Thus K/K0=G\mathcal{K}/\mathcal{K}_0=G.

Class number one gives PB×(Af)=PB×(Q)K\mathrm{PB}^{\times}(\mathbb{A}_f)=\mathrm{PB}^{\times}(\mathbb{Q})\mathcal{K}, where Af\mathbb{A}_f denotes the finite adeles. A rational element stabilizing K\mathcal{K} has a representative locally equal to a scalar times an order unit. A single rational scalar cancels the valuations of the local scalars, giving an element of UU. Consequently the finite adelic quotient is Γ\G\Gamma\backslash G. The projective rational stabilizer at any level coset is trivial. Indeed, K0\mathcal{K}_0 is normal in K\mathcal{K} and only the units ±1\pm1 lie in the kernel of reduction. With coefficients, rational left equivariance gives the stated action of Γ\Gamma [8 Section 2, (4)–(6) and Proposition 2.5]. Each orbit has ∣Γ∣|\Gamma| elements and hence the same finite measure. At infinity Schur orthogonality for a unit test vector gives the factor 1/(2k+1)1/(2k+1). The resulting automorphic norm differs from the counting norm by a common scalar.

To see the direction and normalization in (4), lift gg to g~∈K\widetilde{g}\in\mathcal{K} supported at level primes. A right-ideal representative aa in the correspondence of norm mm can be chosen supported at primes of mm, so g~a=ag~\widetilde{g}a=a\widetilde{g}. If the ideal is αO\alpha\mathcal{O}, write a=αua=\alpha u with u∈Ku\in\mathcal{K} projectively. At level primes uu reduces to αfin−1\alpha_{\mathrm{fin}}^{-1}. Hence the summand is ρk(uα)F(αfin−1g)\rho_k(u\alpha)F(\alpha_{\mathrm{fin}}^{-1}g). Changing α\alpha by a right order unit cancels by equivariance. Averaging over all twelve generators implements the sum once per ideal. Replacing α\alpha by αˉ\bar{\alpha} gives (4). Conjugation also proves self-adjointness.

We use the bound

∥Tm∥≤τ(m)m(5)\lVert T_m\rVert\le\tau(m)\sqrt{m} \tag*{(5)}

on the whole space for k>0k>0, and on the complement of the joint norm-squareclass characters for k=0k=0. Here are the relevant hypotheses for cuspidal transfer. A finite-dimensional smooth irreducible local factor at a split place ℓ\ell is trivial on the norm-one group, since its open normal kernel contains the elementary unipotents by diagonal conjugation. Such a factor cannot occur at positive weight. In fact O[1/ℓ]1\mathcal{O}[1/\ell]^1 has dense real image in SU(2)\mathrm{SU}(2). Equation (2) gives infinitely many norm-one elements α/ℓa\alpha/\ell^a of exact denominator ℓa\ell^a by taking nrd⁡(α)=ℓ2a\operatorname{nrd}(\alpha)=\ell^{2a} and omitting ℓO\ell\mathcal{O}. There are at most 12(2a+1)12(2a+1) single-component points of that norm, so these elements extend outside the Eisenstein torus normalizer. A proper infinite closed subgroup of SU(2)\mathrm{SU}(2) normalizes its one-dimensional identity component. It contains the order-three unit ω\omega, which must then lie in that torus and determines it as the Eisenstein torus. This contradicts the points outside its normalizer. For a finite adelic representative gg, impose the congruences g−1γgg^{-1}\gamma g belonging to the level away from ℓ\ell. They cut out a finite-index subgroup of O[1/ℓ]1\mathcal{O}[1/\ell]^1, still dense by connectedness. At ℓ\ell the discrepancy is absorbed by invariance under the local norm-one group. Rational left invariance then forces the coefficient at gg to be fixed by that dense subgroup, which is impossible in the nontrivial irreducible representation.

For k=0k=0, suppose the factor at ℓ\ell is finite-dimensional and let hh be a finite adelic norm-one element. Strong approximation for B1B^1 away from {∞,ℓ}\{\infty,\ell\} lets us choose a rational norm-one element γ\gamma approximating ghg−1ghg^{-1} outside ℓ\ell [3, Main Theorem 28.5.3]. In ghg−1(γghg−1)ghg^{-1}(\gamma ghg^{-1}), the final factor can be taken in the right stabilizer of the vector away from ℓ\ell. At ℓ\ell it has norm one and acts trivially. Thus the exceptional constituents are invariant under the entire finite adelic norm-one group. On K\mathcal{K} they depend only on the joint norm squareclasses in GG. Indeed, at the odd level primes a square norm can be scaled to one, and outside the level the order units already act trivially. Conversely, each product of local signs comes from a quadratic Dirichlet character composed with reduced norm. This is projectively well-defined since scalar norms are squares, and it has trivial real type since real reduced norms are positive. By (4) its eigenvalue is σ(m)\sigma(m) times its value on the norm class of mm. These sign functions span the scalar exceptional space.

The remaining constituents have infinite-dimensional factors at split places. Jacquet–Langlands gives cuspidal GL2\mathrm{GL}_2 representations of holomorphic weight 2k+22k+2 and trivial central character, unramified at every ℓ∣m\ell\mid m [5 Theorem 14.4 and Proposition 5.20]. Let αℓ,βℓ\alpha_\ell,\beta_\ell be their unitary Satake parameters. Their product is one. The classical Hecke roots are ℓk+1/2αℓ,ℓk+1/2βℓ\ell^{k+1/2}\alpha_\ell,\ell^{k+1/2}\beta_\ell, so Deligne’s bound gives ∣αℓ∣=∣βℓ∣=1|\alpha_\ell|=|\beta_\ell|=1 [6 Theorems 4.9 and 5.1] [7 Theorem 1.6].

For the operator at ℓa\ell^a the all-ideal right cosets are represented by upper triangular matrices with diagonal ℓi,ℓa−i\ell^i,\ell^{a-i} and upper right entry modulo ℓi\ell^i, for 0≤i≤a0\leq i\leq a. In normalized unramified induction [5 Section 3], the spherical vector has value ℓa/2−iαℓiβℓa−i\ell^{a/2-i}\alpha_\ell^i\beta_\ell^{a-i} there. Summing the ℓi\ell^i choices of the upper right entry gives the eigenvalue

ℓa/2∑i=0aαℓiβℓa−i.\ell^{a/2}\sum_{i=0}^{a}\alpha_\ell^i\beta_\ell^{a-i}.

This includes imprimitive ideals and equals the classical coefficient at ℓa\ell^a divided by ℓak\ell^{ak}. In particular, the recurrence is Tℓa=TℓTℓa−1−ℓTℓa−2T_{\ell^a}=T_\ell T_{\ell^{a-1}}-\ell T_{\ell^{a-2}} for a≥2a\geq2. The eigenvalue has modulus at most (a+1)ℓa/2(a+1)\ell^{a/2}. Multiplicativity now gives (5) with the same bound at every weight and level.

A finite test with a prescribed direction

For a test b:G→Cb:G\to\mathbb{C} and a unit vector vv in degree kk, define

F(g)=1∣Γ∣∑γ∈Γρk(γ)vb(γ−1g).F(g)=\frac{1}{|\Gamma|}\sum_{\gamma\in\Gamma}\rho_k(\gamma)v b(\gamma^{-1}g).

This is the orthogonal projection of b(g)vb(g)v onto equivariant functions in the full counting norm on GG. Its coset norm is at most ∥b∥ℓ2(G)/∣Γ∣\|b\|_{\ell^2(G)}/\sqrt{|\Gamma|}. For another unit vector ww, reindexing the shell by the left unit lift of each γ\gamma gives

12⟨w,(TmF)(1)⟩=∑α∈Qmb(αfin)⟨w,ρk(uα−1)v⟩.(6)\frac{1}{2}\langle w,(T_mF)(1)\rangle=\sum_{\alpha\in\mathcal{Q}_m}b(\alpha_{\mathrm{fin}})\langle w,\rho_k(u_\alpha^{-1})v\rangle. \tag*{(6)}

Here the inner product is linear in its second variable. The averaging factor cancels because for a lift γ0\gamma_0 the substitution α=γ0α′\alpha=\gamma_0\alpha' cancels ρk(γ)\rho_k(\gamma) against the archimedean multiplier. This formula allows bb to distinguish the two components.

Writing α=z+yj\alpha=z+yj, put tα=(N(z)−N(y))/mt_\alpha=(N(z)-N(y))/m. For v=wv=w the unit vector proportional to XkYkX^kY^k, the matrix coefficient in (6) equals Pk(tα)P_k(t_\alpha), where PkP_k is the degree-kk Legendre polynomial with Pk(1)=1P_k(1)=1. Indeed, the polynomial action gives the binomial formula

Pk(tα)=∑j=0k(−1)j(kj)2(N(z)/m)k−j(N(y)/m)j.P_k(t_\alpha)=\sum_{j=0}^{k}(-1)^j\binom{k}{j}^2\left(N(z)/m\right)^{k-j}\left(N(y)/m\right)^j.

Passing to uα−1u_\alpha^{-1} leaves the coefficient unchanged.

In degree zero, projecting onto the sign characters replaces bb by its average conditioned on the joint norm squareclasses of mm. Indeed these characters form an orthogonal basis for functions of the squareclasses, and the eigenvalue signs evaluate them at mm. The averaging defining FF preserves these means. For a product of local tests the conditional mean is the product of their conditional means. At each positive degree, evaluation at one coset has norm at most one, so (6) is bounded in absolute value by 12τ(m)m∥b∥ℓ2(G)\frac{1}{2}\tau(m)\sqrt{m}\|b\|_{\ell^2(G)}. The same error bound applies to the scalar cuspidal part.

The cost of a translated window

Choose a fixed smooth 0≤ϕ≤10\leq\phi\leq1 of compact support in (−1/2,1/2)(-1/2,1/2) and integral Iϕ>0I_\phi>0. Put h=H/nh=H/n, b=M/nb=M/n, and λ=x/n\lambda=x/n. We use the bump

W(t)=ϕ((1+t)/2−λh).W(t)=\phi\left(\frac{(1+t)/2-\lambda}{h}\right).

As h≤bh\leq b, on its support 1−t2≍b1-t^2\asymp b, and the support has length at most 2h2h. The constant Legendre coefficient of WW is IϕhI_\phi h.

Lemma 2.1. The Legendre expansion W=∑k≥0ckPkW=\sum_{k\geq0}c_kP_k converges uniformly and absolutely, with ∑k≥1∣ck∣≪ϕb/h\sum_{k\geq1}|c_k|\ll_\phi\sqrt{b/h}.

Proof. We use ∣Pk(t)∣≪k−1/2(1−t2)−1/4|P_k(t)|\ll k^{-1/2}(1-t^2)^{-1/4} for −1<t<1-1<t<1. For a short proof, let Y(u)=sin⁡uPk(cos⁡u)Y(u)=\sqrt{\sin u}P_k(\cos u) on 0<u<π0<u<\pi. The Legendre equation gives Y′′+QY=0Y''+QY=0 with Q=(k+1/2)2+1/(4sin⁡2u)Q=(k+1/2)^2+1/(4\sin^2u). The derivative of Y2+(Y′)2/QY^2+(Y')^2/Q is −Q′(Y′)2/Q2≥0-Q'(Y')^2/Q^2\geq0 up to π/2\pi/2. At π/2\pi/2 this energy is O(k−1)O(k^{-1}), by P2j(0)=(−1)j4−j(2jj)P_{2j}(0)=(-1)^j4^{-j}\binom{2j}{j}, P2j+1(0)=0P_{2j+1}(0)=0 and Pk′(0)=kPk−1(0)P_k'(0)=kP_{k-1}(0). Reflection gives the bound on both halves.

Write L=−ddt((1−t2)ddt)\mathcal{L}=-\frac{d}{dt}\left((1-t^2)\frac{d}{dt}\right). In the rescaled coordinate v=(t−2λ+1)/(2h)v=(t-2\lambda+1)/(2h), the operator is b/h2b/h^2 times a differential operator of order two with coefficients and derivatives bounded on the support of ϕ\phi. Indeed its two coefficients in that scale are −(1−t2)/(4b)-(1-t^2)/(4b) and th/bth/b. Integrating by parts twice using LPk=k(k+1)Pk\mathcal{L}P_k=k(k+1)P_k, and also using the direct integral, gives

∣ck∣≪ϕhk b−1/4min⁡(1,(bh2k2)2).|c_k|\ll_\phi h\sqrt{k}\,b^{-1/4}\min\left(1,\left(\frac{b}{h^2k^2}\right)^2\right).

The factor k\sqrt{k} includes the coefficient normalization (2k+1)/2(2k+1)/2. Summing below and above K=b/hK=\sqrt{b}/h gives Oϕ(hb−1/4K3/2)=Oϕ(b/h)O_\phi(hb^{-1/4}K^{3/2})=O_\phi(\sqrt{b/h}). Since ∣Pk∣≤1|P_k|\leq1 on [−1,1][-1,1], the series converges uniformly. Its Legendre coefficients identify the sum with WW by polynomial density. □\square

Combining this lemma with (6), a finite product test of modulus at most one has weighted shell count

12σ(m)Iϕh∏plocal conditional mean+Oϕ(∣G∣ τ(m)mM/H).(7)12\sigma(m)I_\phi h\prod_p \text{local conditional mean}+O_\phi\left(\sqrt{|G|}\,\tau(m)\sqrt{m}\sqrt{M/H}\right). \tag*{(7)}

The same WW is used when the target is divided by a common factor. Both component norms scale together and tαt_\alpha is unchanged.

Local densities at all target valuations

On a shell at mm, put A1=N(z)A_1=N(z) and A2=N(y)A_2=N(y). The bump WW restricts both norms to be positive. At a variable prime p≥5p\geq5 write gm(p)g_m(p) for the density of the bad test p∣A1A2p\mid A_1A_2. For p∤mp\nmid m this is the conditional mean in (7). At p∣mp\mid m we derive its analogue by reducing to coprime Hecke indices.

Coprime target and level

At 33 the ordinary family requires no test, since m≡2(mod3)m\equiv2\pmod3 forces both norms to be units modulo 33. In the mixed family m≡1(mod3)m\equiv1\pmod3, and we require the prescribed component to have 33-adic valuation exactly one. Among the nine elements of R/3RR/3R there are six units, two nonzero nonunits, and zero. The nonzero nonunits have norm valuation one, since 1−ω1-\omega has norm 33 and its square is associated to 33. Norm one modulo 33 requires one unit and one nonunit, giving 2⋅6⋅3=362\cdot6\cdot3=36 raw quaternion residues. Exactly 6⋅26\cdot2 have a unit in one prescribed component and a nonzero nonunit in the other. Passing to projective classes preserves the ratio, so either mixed orientation has mean 1/31/3.

At an inert p∤mp\nmid m, by anisotropy a bad point has one component zero modulo pRpR. Its image in PGL⁡2(Fp)\operatorname{PGL}_2(\mathbb{F}_p) lies in the normalizer of the nonsplit torus (R/pR)×/Fp×(R/pR)^\times/\mathbb{F}_p^\times. This is the union of that torus and its translate by jj, with 2(p+1)2(p+1) elements out of p(p2−1)p(p^2-1). The norms of the torus elements cover the squareclasses equally, so gm(p)=2/[p(p−1)]g_m(p)=2/[p(p-1)] in either class.

At a split p∤mp\nmid m, the two factors of the Eisenstein subalgebra occupy the diagonal of M2(Fp)M_2(\mathbb{F}_p). The other Eisenstein component occupies the off-diagonal. Their norms are the products of the respective entries, up to sign. For a fixed nonzero determinant, the first row is uniform on nonzero rows, each with pp completions. Thus a given entry is zero with probability 1/(p+1)1/(p+1). Two zero entries are possible only when they form an opposite pair. For either such pair there are p−1p-1 matrices of the given determinant, giving probability 1/[p(p+1)]1/[p(p+1)]. Inclusion-exclusion gives

1−gm(p)=Bp:=1−4p+1+2p(p+1)=(p−1)(p−2)p(p+1).(8)1-g_m(p)=B_p:=1-\frac{4}{p+1}+\frac{2}{p(p+1)}=\frac{(p-1)(p-2)}{p(p+1)}. \tag*{(8)}

For the exact condition vp(Ai)=1v_p(A_i)=1 at either one prescribed component, exactly one of its two entries is zero modulo pp, and this entry must be nonzero modulo p2p^2. All matrix lifts of an invertible residue are invertible, and determinant squareclasses are determined modulo pp. The conditional probability of the last lift condition is (p−1)/p(p-1)/p. Therefore the mean of vp(Ai)=1v_p(A_i)=1 at level p2p^2 is

tp=2(p−1)2p2(p+1),ap:=tp2Bp=p−1p(p−2).(9)t_p=\frac{2(p-1)^2}{p^2(p+1)},\qquad a_p:=\frac{t_p}{2B_p}=\frac{p-1}{p(p-2)}. \tag*{(9)}

These means hold on projective groups as well, since all tests are unchanged by a unit scalar.

Primes dividing the target

For inert p∣mp\mid m, divisibility of one norm by pp forces the other to be divisible by pp as well, so both components are in pRpR. The bad set is empty if vp(m)=1v_p(m)=1. Otherwise it is obtained by multiplying both coordinates by pp from the shell at m/p2m/p^2. This preserves WW and the other prime tests. Hence for e=vp(m)e=v_p(m) the scalar density is

gm(p)=0 (e=1),gm(p)=σ(pe−2)σ(pe) (e≥2).(10)g_m(p)=0\ (e=1),\qquad g_m(p)=\frac{\sigma(p^{e-2})}{\sigma(p^e)}\ (e\ge2). \tag*{(10)}

Now let p=ππˉp=\pi\bar{\pi} be split in RR, and fix rr prime to pp. At target perp^e r, let UeU_e be the weighted count with no restriction at pp, and GeG_e the count with both norms units at pp. Carry along WW and the other prime tests. At target rr, let L0L_0 count with the additional test

12(1π∣z+1πˉ∣z+1π∣y+1πˉ∣y).\frac{1}{2}\left(\mathbf{1}_{\pi\mid z}+\mathbf{1}_{\bar{\pi}\mid z}+\mathbf{1}_{\pi\mid y}+\mathbf{1}_{\bar{\pi}\mid y}\right).

It has mean 2/(p+1)2/(p+1) by the single-entry calculation. If one norm has valuation v>0v>0, the v+1v+1 allocations of its two prime factors are equally frequent. The average of the test at that pair of norms is v/(v+1)v/(v+1). The conditions at other primes are unchanged on reallocating these factors. Indeed, dividing out or inserting factors above pp acts by units at every other prime, preserving all the norm valuations and the individual split-prime divisibilities used in the tests.

Classify a pair of norms on the shell perp^e r by the minimum jj of its two pp-adic valuations. If j<ej<e, both valuations are jj; removing pjp^j from each norm gives a Ge−jG_{e-j} pair with allocation multiplier (j+1)2(j+1)^2. If j=ej=e, the residual target is rr. When both residual norms are units the multiplier is (e+1)2(e+1)^2; if one has valuation v>0v>0, the ratio of allocation counts is (e+1)(v+e+1)/(v+1)=(e+1)2−e(e+1)v/(v+1)(e+1)(v+e+1)/(v+1)=(e+1)^2-e(e+1)v/(v+1). In either reduction the normalized norm coordinate stays fixed. This proves

Ue=∑j=0e−1(j+1)2Ge−j+(e+1)2U0−e(e+1)L0(e≥1).(11)U_e=\sum_{j=0}^{e-1}(j+1)^2G_{e-j}+(e+1)^2U_0-e(e+1)L_0\qquad(e\ge1). \tag*{(11)}

The generating series for (j+1)2(j+1)^2 is A(T)=(1+T)/(1−T)3A(T)=(1+T)/(1-T)^3, and that for e(e+1)e(e+1) is C(T)=2T/(1−T)3C(T)=2T/(1-T)^3. Thus

∑e≥1GeTe=∑e≥0UeTeA(T)−U0+C(T)A(T)L0.\sum_{e\ge1}G_eT^e=\frac{\sum_{e\ge0}U_eT^e}{A(T)}-U_0+\frac{C(T)}{A(T)}L_0.

The coefficients of 1/A(T)1/A(T) have absolute value at most eight, and those of C(T)/A(T)C(T)/A(T) at most two. The bad count Ue−GeU_e-G_e is therefore a combination of unrestricted counts at lower pp-exponents and the L0L_0 count at the unit target, with coefficient absolute sum at most 8e+28e+2. Applying the same identity to the scalar prediction from (7) uses Ue=σ(pe)U_e=\sigma(p^e), U0=1U_0=1, and L0=2/(p+1)L_0=2/(p+1) after factoring out the other local factors. Substitution gives ∑e≥1GeTe=(p−1)2p+1T1−pT\sum_{e\geq1}G_eT^e=\frac{(p-1)^2}{p+1}\frac{T}{1-pT} for those scalar values. In particular

1−gm(p)=(p−1)2pe−1(p+1)σ(pe)(e=vp(m)≥1, p split).(12)1-g_m(p)=\frac{(p-1)^2p^{e-1}}{(p+1)\sigma(p^e)} \qquad(e=v_p(m)\geq1,\ p\ \text{split}). \tag*{(12)}

Example at a split square. Take m=72rm=7^2r with 7∤r7\nmid r. The first two cases of (11) give

U2−G2=4G1+9U0−6L0=4U1−7U0+2L0.U_2-G_2=4G_1+9U_0-6L_0=4U_1-7U_0+2L_0.

The term at 7r7r is unrestricted at 77, and the only remaining 77-level test is L0L_0 at rr. For the main term, factor out the other primes and insert U2=57U_2=57, U1=8U_1=8, U0=1U_0=1, L0=1/4L_0=1/4. The resulting main term for G2G_2 is 63/263/2, giving 1−gm(7)=(63/2)/571-g_m(7)=(63/2)/57. Including the shell factor σ(72)/72=57/49\sigma(7^2)/7^2=57/49 gives 9/149/14. This is B7⋅(6/5)B_7\cdot(6/5), since B7=15/28B_7=15/28 at a unit target. In the cofactor sum of § 4.2 this gives weight 6/56/5 without a common factor 77, and (6/5)/7(6/5)/7 with that common factor, since the residual target 7r7r still has positive valuation at 77. In general (12) gives Ep=(p−1)/(p−2)E_p=(p-1)/(p-2) at every positive valuation.

The joint estimate

Fix squarefree coprime uu, vv made of split primes coprime to mm. Place the condition vp(A1)=1v_p(A_1)=1 at p∣up\mid u and vp(A2)=1v_p(A_2)=1 at p∣vp\mid v. Give each coordinate pair weight W(t)/τ(uv)W(t)/\tau(uv). Here and below mm is even and prime to 33. Keep the odd-component condition and the prescribed condition at 33. For any squarefree ℓ\ell coprime to 6uv6uv, let Sm(ℓ)S_m(\ell) be the weighted count additionally bad at every prime dividing ℓ\ell, and put

Xm=IϕhRo(m)βm,βm={1,m≡2(mod6),1/3,m≡4(mod6).X_m=I_\phi hR_o(m)\beta_m,\qquad \beta_m= \begin{cases} 1, & m\equiv2\pmod6,\\ 1/3, & m\equiv4\pmod6. \end{cases}

Extend gmg_m multiplicatively on squarefree products.

Lemma 3.1. For m≤nm\leq n as above and any fixed ζ>0\zeta>0,

Sm(ℓ)=Xm∏p∣uv(tp/2) gm(ℓ)+Oϕ,ζ((uv)3/2n1/2+ζM/H).(13)S_m(\ell)=X_m\prod_{p\mid uv}(t_p/2)\,g_m(\ell)+O_{\phi,\zeta}\left((uv)^{3/2}n^{1/2+\zeta}\sqrt{M/H}\right). \tag*{(13)}

Proof. First omit oddness. If the level is coprime to mm, apply (7). A factor at p∣uvp\mid uv has size p3∣PGL⁡2(Fp)∣<p6p^3|\operatorname{PGL}_2(\mathbb{F}_p)|<p^6 and a factor at p∣ℓp\mid\ell has size less than p3p^3. Thus the norm cost is O((uv)ℓ3/2)O((uv)^{\ell^3/2}), the fixed factor at 33 being bounded.

For p∣(ℓ,m)p\mid(\ell,m), eliminate the test by inert scaling or (11). Each time, the remaining test at the eliminated prime is present only at a reduced target coprime to it. The allocation calculation works with any of the other tests still imposed. In particular the test at a split prime produced by an earlier elimination is an average of individual divisibilities, all invariant under changes of allocation at another prime. The coprime conditional means are independent of the nonzero determinant class at those other primes. Extracting split powers and inert squares does not change the target class modulo 33, and the prescribed orientation is preserved. Hence the scalar terms factor, with local factors (10) and (12). The total coefficient variation in the elimination is at most τ(m)4\tau(m)^4, because 8e+2≤(e+1)48e+2\leq(e+1)^4. Each resulting target divides mm, and the remaining level divides the original level. The product of these losses with the divisor factor in (7) is absorbed by Oζ(nζ)O_\zeta(n^\zeta).

Finally subtract the shell at m/4m/4 if it exists. Rescaling its coordinates by 22 preserves the normalized angle, the odd-prime tests and the orientation at 33. The valuations at all odd primes agree with those for mm, so the scalar σ(m)\sigma(m) is replaced by σ(m)−σ(m/4)\sigma(m)-\sigma(m/4). This gives (13). ∎

The lower sieve and the split factors

Fix κ,ξ>0\kappa,\xi> 0 with 4κ+52ξ<ε/24\kappa+ \frac{5}{2}\xi< \varepsilon/2. We use a divisor level D=nξD = n^{\xi} and prime bounds Z=nηZ = n^{\eta} and Y=nλ0Y = n^{\lambda_0}, with 0<λ0<η0 < \lambda_0 < \eta chosen below. Consider triples of pairwise coprime squarefree integers d,u,vd,u,v, supported on split primes at most YY, satisfying

d∣n,(uv,n)=1,duv≤nκ.(14)d \mid n,\qquad(uv,n)=1,\qquad duv \le n^{\kappa}. \tag*{(14)}

On the shell m=n/dm=n/d use the prescriptions of Lemma 3.1 at u,vu,v, and require p∤A1A2p \nmid A_1A_2 at every other prime 5≤p≤Z5 \le p \le Z, including the primes of dd. Since d≡1(mod6)d \equiv1 \pmod6, the target mm belongs to the same family. The weight on the shell is W(t)=ϕ((dA1−x)/H)W(t)=\phi((dA_1-x)/H), so after rescaling the norm values by dd the original window is enforced.

Finite inclusion-exclusion

We give the sieve weights together with the level they require. Partition the sieving primes into blocks Z2−j−1<p≤Z2−jZ^{2^{-j-1}} < p \le Z^{2^{-j}} for j≥0j \ge0. For the main-term probabilities gm(p)g_m(p), the sum in each block is bounded by an absolute CC and vj=∏p in block j(1−gm(p))v_j=\prod_{p\text{ in block }j}(1-g_m(p)) is bounded below by an absolute c>0c>0. In fact the formulas in Section 3 lie uniformly below one and are O(1/p)O(1/p). Prime harmonic Mertens gives a bounded sum of 1/p1/p in each block. Omitting primes, including the primes of uvuv, preserves these bounds. We also use the progression version of Mertens modulo 3, ∑p≤Tp≡a (3)1/p=12log⁡log⁡T+O(1)\sum_{\substack{p\le T\\p\equiv a\ (3)}}1/p=\frac{1}{2}\log\log T+O(1) for a=1,2a=1,2; see [9 Lemma 2.4].

In block jj truncate inclusion-exclusion at consecutive orders kj=2⌈(K+j)/2⌉k_j=2\lceil(K+j)/2\rceil and kj−1k_j-1, where KK is a fixed large integer. Denote the resulting polynomials in the bad-prime indicators by Uj,LjU_j,L_j. Bonferroni gives 0≤1j≤Uj0\le1_j\le U_j and Lj≤1jL_j\le1_j, where 1j1_j indicates avoidance in the block. Telescoping ∏jUj−∏j1j\prod_j U_j-\prod_j1_j gives the lower polynomial

∏j1j≥∏jUj−∑j(Uj−Lj)∏i≠jUi.(15)\prod_j 1_j \ge\prod_j U_j-\sum_j(U_j-L_j)\prod_{i\ne j}U_i. \tag*{(15)}

Under independent bad events with probabilities gm(p)g_m(p),

EUj≥vj,E(Uj−Lj)≤Ckjkj!.\mathbb{E}U_j\ge v_j,\qquad\mathbb{E}(U_j-L_j)\le\frac{C^{k_j}}{k_j!}.

Choose KK so that the sum over jj of Ckj/(ckj!)C^{k_j}/(ck_j!) is at most 1/21/2. The expectation of the lower polynomial is at least 12∏jvj\frac{1}{2}\prod_jv_j. Also 0≤EUj−vj≤E(Uj−Lj)0\le\mathbb{E}U_j-v_j\le\mathbb{E}(U_j-L_j), so the expectation of the upper polynomial ∏jUj\prod_jU_j is O(∏jvj)O(\prod_jv_j). We will use this upper estimate later with integer-sieve probabilities.

Put S=∑j≥0kj2−j<∞S=\sum_{j\ge0}k_j2^{-j}<\infty and η=ξ/S\eta=\xi/S. Every divisor in either polynomial is at most ZS=DZ^S=D. The upper coefficients have absolute value at most one, and the lower coefficients at most one plus the number of nonempty blocks, which is O(log⁡log⁡(3n))O(\log\log(3n)). Apply (13) to the lower polynomial with the nonnegative shell weights. For each triple the main contribution is at least

12XmVm(Z)∏p∣uvap,Vm(Z)=∏5≤p≤Z(1−gm(p)).(16)\frac{1}{2}X_mV_m(Z)\prod_{p\mid uv}a_p,\qquad V_m(Z)=\prod_{5\le p\le Z}(1-g_m(p)). \tag*{(16)}

Here ap=tp/(2Bp)a_p=t_p/(2B_p) accounts for replacing the ordinary sieve factor at a prime of uvuv by the exact-valuation factor. Summing the remainders over all divisors uses ∑ℓ≤Dℓ3/2≪D5/2\sum_{\ell\le D}\ell^{3/2}\ll D^{5/2}. There are at most nκ+o(1)n^{\kappa+o(1)} triples because a fixed product duvduv has at most τ(duv)2\tau(duv)^2 decompositions. The total absolute remainder, for any ζ>0\zeta>0, is

Oϕ,ζ(n1/2+4κ+(5/2)ξ+ζM/H)=Oε(H/log⁡n).O_{\phi,\zeta}\left(n^{1/2+4\kappa+(5/2)\xi+\zeta}\sqrt{M/H}\right)=O_{\varepsilon}(H/\log n).

The equality uses nM/H≤Hn−ε/2\sqrt{nM/H}\le Hn^{-\varepsilon/2} and the strict choice of κ,ξ\kappa,\xi, with ζ\zeta small enough. All of these estimates are uniform in x,Hx,H.

The logarithm gained from small factors

The target-independent product

V∗(Z)=∏5≤p≤Zp splitBp∏5≤p≤Zp inertbp,bp=1−2p(p−1)V_\ast(Z)=\prod_{\substack{5\le p\le Z\\p\ \mathrm{split}}}B_p\prod_{\substack{5\le p\le Z\\p\ \mathrm{inert}}}b_p,\qquad b_p=1-\frac{2}{p(p-1)}

satisfies V∗(Z)≍(log⁡Z)−2V_\ast(Z)\asymp(\log Z)^{-2}, since Bp=1−4/p+O(p−2)B_p=1-4/p+O(p^{-2}) and the inert factors have a convergent positive product. To compare this with Vm(Z)V_m(Z), include the shell factor Ro(m)/mR_o(m)/m from (3). For pe∥mp^e\mathbin{\Vert}m split and p≤Zp\le Z,

σ(pe)pe(1−gm(p))=(p−1)2p(p+1)=BpEp,Ep=p−1p−2.\frac{\sigma(p^e)}{p^e}(1-g_m(p))=\frac{(p-1)^2}{p(p+1)}=B_pE_p,\qquad E_p=\frac{p-1}{p-2}.

For pe∥mp^e\mathbin{\Vert}m inert and p≤Zp\le Z the combined factor is 1+1/p=bpEp1+1/p=b_pE_p, by (10), including e=1e=1. The inert primes dividing mm and nn are the same. For the inert primes dividing mm beyond ZZ, the shell factor is at least 1+1/p=bpEp1+1/p=b_pE_p, and the product of the extra factors bpb_p is bounded below. All split enhancements are at least one. Using hm=(H/n)m=H/dh_m=(H/n)m=H/d, we get

XmVm(Z)≫ϕHdJ(n)V∗(Z)∏p≤Y, p splitp∣mEp.(17)X_mV_m(Z)\gg_\phi\frac{H}{d}J(n)V_\ast(Z)\prod_{\substack{p\le Y,\ p\ \mathrm{split}\\p\mid m}}E_p. \tag*{(17)}

Temporarily drop the product bound in (14). The sum of the cofactor weights in (16) and (17) factors over split p≤Yp\le Y. For p∤np\nmid n the three options of absence, inclusion in uu, or inclusion in vv have weights 11, apa_p, apa_p. If pe∥np^e\mathbin{\Vert}n, the two options of absence or inclusion in dd have weights EpE_p, Ep′/pE'_p/p, where Ep′=1E'_p=1 for e=1e=1 and Ep′=EpE'_p=E_p for e≥2e\ge2. The compensation at target divisors is the identity

Ep+Ep′p≥Ep+1p=1+2ap.E_p+\frac{E'_p}{p}\ge E_p+\frac{1}{p}=1+2a_p.

Thus the full cofactor sum is at least ∏p≤Y, p split(1+2ap)≍log⁡Y\prod_{p\le Y,\ p\ \mathrm{split}}(1+2a_p)\asymp\log Y, since ap=1/p+O(p−2)a_p=1/p+O(p^{-2}). Normalize the local weights to independent probabilities. At p∤np\nmid n the probability of p∣duvp\mid duv is 2ap/(1+2ap)≪p−12a_p/(1+2a_p)\ll p^{-1}, and at p∣np\mid n it is at most 1/p1/p. Hence Elog⁡(duv)≤C′log⁡Y\mathbb{E}\log(duv)\le C'\log Y, uniformly in nn, by ∑p≤Y(log⁡p)/p≪log⁡Y\sum_{p\le Y}(\log p)/p\ll\log Y. Take 0<λ0<min⁡(η,κ/(3C′))0<\lambda_0<\min(\eta,\kappa/(3C')). Markov’s inequality then retains at least half the weight on duv≤nκduv\le n^\kappa. Inserting (17) into (16) and subtracting the aggregate remainder proves that the total retained weight is

≫εHJ(n)log⁡Y(log⁡Z)2≫εHJ(n)log⁡n.(18)\gg_\varepsilon HJ(n)\frac{\log Y}{(\log Z)^2}\gg_\varepsilon\frac{HJ(n)}{\log n}. \tag*{(18)}

From coordinates to distinct pairs

For each dd fix ρd∈R\rho_d\in R of norm dd, obtained by multiplying one factor above each split prime of dd. Embed the residual shell at m=n/dm=n/d in the original shell by (z,y)↦(ρdz,ρdy)(z,y)\mapsto(\rho_dz,\rho_dy). This preserves the fractional norm coordinate, the prescribed orientation at 33, and inert divisibility above ZZ.

Each original value pair recovers its triple d,u,vd,u,v. At split primes up to YY, the only possible positive valuation is one, occurring in both norms at the primes of dd, only in the first at uu, and only in the second at vv. The images for distinct triples are consequently disjoint. This makes it possible to bound the remaining losses on the original shell without summing an error over the triples.

Large inert primes

If an inert prime p>Zp > Z divides the first component norm, that component is pβp\beta. The norm of β\beta belongs to [(x−H/2)/p2,(x+H/2)/p2][(x-H/2)/p^2,(x+H/2)/p^2]. For the other component use n−xn-x in place of xx. Each resulting annulus has squared-radius width H/p2H/p^2 and outer radius at most n/p\sqrt{n}/p. Covering each lattice point by a translated fundamental cell, of fixed area and bounded radius, gives O(H/p2+n/p+1)O(H/p^2+\sqrt{n}/p+1) possible β\beta. Indeed the cells lie in the annulus obtained by moving the two boundaries by a fixed distance, truncating the inner radius at zero. The complementary positive norm has at most 6max⁡a≤nτ(a)=no(1)6\max_{a\le n}\tau(a)=n^{o(1)} coordinate choices.

Since all weights are at most one, summing on the original shell over the two components and Z<p≤nZ<p\le\sqrt{n} bounds the lost weight by

no(1)(HZ+n(1+log⁡n))=oε(H/log⁡n).n^{o(1)}\left(\frac{H}{Z}+\sqrt{n}(1+\log n)\right)=o_{\varepsilon}(H/\log n).

Here ZZ is a fixed power of nn, and the hypothesis H3≥n1+εMH^3\ge n^{1+\varepsilon}M with M≥HM\ge H gives H≥n(1+ε)/2H\ge n^{(1+\varepsilon)/2}. Thus (18) persists after removing every inert prime. Oddness and the condition at 3 now give the required norm values.

The weight of one value pair

On the residual shell write A1=uk1A_1=uk_1 and A2=vk2A_2=vk_2. All split prime factors of k1,k2k_1,k_2 exceed Z=nηZ=n^\eta. Each of these two numbers has at most 1/η1/\eta split factors counted with multiplicity, and the possible single factor 3 has no effect on coordinate multiplicity. The split primes in uvuv occur to valuation one. By the multiplicity formula in § 2, the weighted coordinate multiplicity, after division by τ(uv)\tau(uv), is at most

36f(k1)f(k2)≤36⋅22/η.36f(k_1)f(k_2)\le36\cdot2^{2/\eta}.

Each original pair determines its triple and residual pair. Dividing the surviving weight by this constant proves the lower bound of Theorem 1.1.

The upper bound on integers

For the upper bound, write s=s1s=s_1 in the interval of length HH. At every inert p≥5p\ge5 we must exclude the residues 0,n(modp)0,n\pmod p. Let ν(p)\nu(p) be their number, namely one if p∣np\mid n and two otherwise, and extend ν\nu multiplicatively to squarefree products of these primes. The Chinese remainder theorem and the count in a residue class give Hν(ℓ)/ℓ+O(ν(ℓ))H\nu(\ell)/\ell+O(\nu(\ell)) integers in all the excluded sets for p∣ℓp\mid\ell, uniformly in the interval location and endpoint convention.

Use the upper polynomials of § 4 with probabilities ν(p)/p\nu(p)/p, sieving to Z′=nc1Z'=n^{c_1} for a sufficiently small fixed c1>0c_1>0. The block probability bounds hold as before. Choose c1c_1 so that the corresponding divisor level is D′≤n1/4D'\le n^{1/4}. The total error is O(D′log⁡(2D′))=o(H/log⁡n)O(D'\log(2D'))=o(H/\log n), since ν(ℓ)≤τ(ℓ)\nu(\ell)\le\tau(\ell) and ∑ℓ≤D′τ(ℓ)≪D′log⁡(2D′)\sum_{\ell\le D'}\tau(\ell)\ll D'\log(2D'). The upper main term is at most a constant times HH times

∏5≤p≤Z′p inert(1−2p)∏5≤p≤Z′p∣n, p inertp−1p−2≪J(n)log⁡n.\prod_{\substack{5\le p\le Z'\\ p\ \mathrm{inert}}}\left(1-\frac{2}{p}\right) \prod_{\substack{5\le p\le Z'\\ p\mid n,\ p\ \mathrm{inert}}}\frac{p-1}{p-2} \ll\frac{J(n)}{\log n}.

This proves the upper bound and completes Theorem 1.1. The argument for this upper bound in fact works in any interval whose length is at least a fixed positive multiple of n\sqrt{n}.

Proof of Corollary 1.2. In a fixed interval of shares, take a smaller fixed interval of length at most the distance of its center from the nearest endpoint of [0,1][0,1]. This satisfies the hypotheses with H≍nH\asymp n for large nn, giving the lower bound. Use the upper sieve on the full interval and on [0,n][0,n] for the unrestricted count. At x=λnx=\lambda n with 0<λ<10<\lambda<1, a width H=nδH=n^\delta satisfies the condition for any fixed 0<ε<3θ−20 < \varepsilon< 3\theta- 2. Finally, inside [anθ,bnθ][an^\theta, bn^\theta] take a subinterval of width comparable to nθn^\theta and at most its midpoint. Its distance to the nearer endpoint of [0,n][0,n] is comparable to nθn^\theta, so one can use 0<ε<2θ−10 < \varepsilon< 2\theta- 1. The upper sieve on the full interval gives the matching bound, also for a=0a = 0. □\square

AI assistance and review. ChatGPT (OpenAI) assisted with mathematical arguments, literature checks, exposition, proof review and verification code.

References

  1. [1]W. Howe, G. Huffman, J. Kohlmeyer, and L. Kruer, Euler’s 1747 conjecture on sums of integers with restricted prime factors, Hexagon 2611.00370v1, 12 November 2026. https://hexagonomath.org/2611.00370v1.hub.valency.io/works/W_p3mybhqd
  2. [2]R. Dietmann and C. Elsholtz, Sums of two numbers having only prime factors congruent to one modulo four, Banach Center Publ. 118 (2019), 177–179. https://doi.org/10.4064/bc118-11.
  3. [3]J. Voight, Quaternion Algebras, Graduate Texts in Mathematics 288, Springer, 2021, Chapters 28 and 41. https://doi.org/10.1007/978-3-030-56694-4.
  4. [4]O. Ramaré and G. K. Viswanadham, The trigonometric polynomial on sums of two squares, an additive problem and generalisation, arXiv:2509.23260v1, 2025. https://arxiv.org/abs/2509.23260v1.
  5. [5]H. Jacquet and R. P. Langlands, Automorphic Forms on GL(2), Lecture Notes in Mathematics 114, Springer, 1970, Theorem 14.4 and Section 5. https://publications.ias.edu/sites/default/files/automorphic-forms-on-gl2_rpl_9.pdf.
  6. [6]P. Deligne, Formes modulaires et représentations ℓ-adiques, Séminaire Bourbaki, exposé 355 (1968/69), published 1971, 139–172. https://www.numdam.org/item/SB_1968-1969__11__139_0.pdf.
  7. [7]P. Deligne, La conjecture de Weil. I, Publications Mathématiques de l’IHÉS 43 (1974), 273–307. https://www.numdam.org/item/PMIHES_1974__43__273_0.pdf.DOI
  8. [8]M. Greenberg and J. Voight, Lattice methods for algebraic modular forms on classical groups, in Computations with Modular Forms, Contrib. Math. Comput. Sci. 6, Springer, 2014, 147–179. https://jvoight.github.io/articles/orthohecke-240123.pdf.arxiv.org/abs/1209.2460
  9. [9]D. Loughran and L. Matthiesen, Frobenian multiplicative functions and rational points in fibrations, J. Eur. Math. Soc. 26 (2024), 4779–4830. https://doi.org/10.4171/jems/1374.
  10. [10]H. Iwaniec, The half dimensional sieve, Acta Arith. 29 (1976), 69–95. https://doi.org/10.4064/aa-29-1-69-95.
  11. [11]G. Greaves, On the representation of a number in the form x² + y² + p² + q² where p, q are odd primes, Acta Arith. 29 (1976), 257–274. https://doi.org/10.4064/aa-29-3-257-274.

Paper details

vlnc:W_cx7qk2j4
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 10

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

OpenAI Codex, OpenAI ChatGPT

View statement
Looking for related papers…
Contents