Irrational generators in Shapiro’s conjecture

This supplies the irrational case of the ordinary-prime omission theorem. It also gives a count for pairs of integers on the two sides of an irrational dilation.

Theorem. Suppose EE is a set of relative density zero among the primes. Let SES_E be the positive integers whose prime factors lie outside EE, and put W(Y)=∏p∈E,p≤Y(1−1/p)W(Y)=\prod_{p\in E,\ p\le Y}(1-1/p). For a fixed irrational λ>1\lambda\gt1 and the denominators QQ of its continued-fraction convergents, along X=Q2X=Q^2 we have

#{1≤n≤X:n∈SE,⌊λn⌋∈SE}∼XW(X)2.\#\{1\le n\le X:n\in S_E,\ \lfloor\lambda n\rfloor\in S_E\} \sim XW(X)^2.

In particular, the pair (n,m)(n,m) with m=⌊λn⌋m=\lfloor\lambda n\rfloor violates unit separation if a multiplicative semigroup contains SES_E and λ\lambda.

Lattice counts with congruences

At X=Q2X=Q^2 use the area-preserving map

T(n,m)=(n/Q,Q(λn−m)).T(n,m)=(n/Q,\ Q(\lambda n-m)).

The region 0<n≤X0\lt n\le X, 0<λn−m<10\lt\lambda n-m\lt1 maps to a square of side QQ, with the indicated half-open boundary. For convergents a/Qa/Q and a′/Q′a'/Q' of consecutive indices, the vectors T(Q,a)T(Q,a) and T(Q′,a′)T(Q',a') give a basis of Λ=Tℤ2\Lambda=T\mathbb Z^2. Both vectors have coordinates bounded by 1 in absolute value. The bound on the second coordinate uses |λQ−a|<1/Qnext|\lambda Q-a|\lt 1/Q_{\mathrm{next}} and |λQ′−a′|<1/Q|\lambda Q'-a'|\lt1/Q. Their determinant has absolute value 1. Thus the shortest length in Λ\Lambda is bounded below by an absolute constant. In particular the number of nonzero lattice vectors in a ball of radius RR is O(R2)O(R^2) whenever the ball contains a nonzero vector.

We will also use Λh\Lambda_h for the sublattice with h∣nh\mid n, and Λ′h\Lambda'_h for the one with h∣mh\mid m. Either has covolume hh. For any one of these lattices LL, write ℓ(L)\ell(L) for the shortest length of a nonzero vector. A translate of dLdL meets the square in

Xd2detL+O(1+Qdℓ(L))\frac{X}{d^2\det L}+O\left(1+\frac{Q}{d\ell(L)}\right)

points, by the two-dimensional lattice-counting estimate in terms of the first successive minimum. The estimate applies to the translated region as well. Adding or dropping a side of length QQ changes the bound by O(1+Q/(dℓ(L)))O(1+Q/(d\ell(L))).

For L=ΛhL=\Lambda_h use integer coordinates (n/h,m)(n/h,m) on the lattice, and for L=Λ′hL=\Lambda'_h use (n,m/h)(n,m/h). At each sifting prime rr, erase one coordinate’s zero residue or both coordinates’ zero residues. For a squarefree product dd of these primes, the simultaneous erased set occupies ρ(d)≤2ω(d)d\rho(d)\le 2^{\omega(d)}d residue pairs modulo dd, by the Chinese remainder theorem. Each residue pair corresponds to one translate of dLdL. The count for these congruences is

XdetLρ(d)d2+O(2ω(d)(d+Qℓ(L))).\frac{X}{\det L}\frac{\rho(d)}{d^2} +O\left(2^{\omega(d)}\left(d+\frac Q{\ell(L)}\right)\right).

The local retained densities are 1−1/r1-1/r or (1−1/r)2(1-1/r)^2, so Mertens’ theorem gives the uniform dimension-two product bound. With sieve level DD and range z=D1/sz=D^{1/s}, the fundamental lemma bounds the sifted total above and below by

(1+O(e−s))XdetLU(z)+O(Dlog(2D)Qℓ(L)+D2log(2D)),(1+O(e^{-s}))\frac{X}{\det L}U(z) +O\left(D\log(2D)\frac Q{\ell(L)}+D^2\log(2D)\right),

where U(z)U(z) is the product of the local retained densities. We will use the displayed error as a two-sided bound on the difference from the main term.

Small and medium omitted primes

The prime number theorem and partial summation give W(X)=(logX)−o(1)W(X)=(\log X)^{-o(1)} and W(Xc)/W(X)→1W(X^c)/W(X)\to1 for each fixed c>0c\gt0. Fix an error tolerance ε>0\varepsilon\gt0. On Λ\Lambda use level D=X1/8D=X^{1/8} and range XcX^c with c>0c\gt0 sufficiently small depending on ε\varepsilon. Both coordinates are sifted by EE. The count passing this first sieve is (1+O(ε)+o(1))XW(X)2(1+O(\varepsilon)+o(1))XW(X)^2 along X=Q2X=Q^2.

Fix 0<δ<min(c,1/10)0\lt\delta\lt\min(c,1/10). A failure at an omitted prime pp in the range Xc<p≤X1−δX^c\lt p\le X^{1-\delta} lies on Λp\Lambda_p or Λ′p\Lambda'_p. In either case we may use upper sieve level D0=Xδ/10D_0=X^{\delta/10} and a range z0=Xc0z_0=X^{c_0} with a sufficiently small fixed 0<c0<c0\lt c_0\lt c. Both original coordinates can be sifted by the primes in EE up to z0z_0, because pp is invertible at each of those primes. The main terms summed over pp are at most a constant times

XW(z0)2∑Xc<p≤X1−δ,p∈E1p=o(XW(X)2).XW(z_0)^2\sum_{X^c\lt p\le X^{1-\delta},\ p\in E}\frac1p=o(XW(X)^2).

For the remainders, consider the primes in [P,2P)[P,2P) from this range. For R≤PR\le\sqrt P, each prime satisfying ℓ(Λp)≤R\ell(\Lambda_p)\le R divides the nonzero nn coordinate of some lattice vector in the radius-RR ball. The nn coordinate is nonzero because every nonzero vector with n=0n=0 has length at least QQ, while R≤X(1−δ)/2<QR\le X^{(1-\delta)/2}\lt Q. Moreover |n|≤QR≤X|n|\le QR\le X. Each such nn has at most 1/c1/c prime divisors above XcX^c. There are O(R2)O(R^2) vectors, so there are Oc(R2)O_c(R^2) primes satisfying this short-vector condition. The same bound holds for Λ′p\Lambda'_p, since m=0m=0 implies length at least λQ\lambda Q for a nonzero vector and |m|≤λQR+R/Q=Oλ(X)|m|\le\lambda QR+R/Q=O_\lambda(X) otherwise.

Dyadically summing the inverse lengths up to P\sqrt P, then using O(P)O(P) primes for the lengths above P\sqrt P, gives

∑p∈[P,2P)(ℓ(Λp)−1+ℓ(Λ′p)−1)≪c,λP\sum_{p\in[P,2P)}\left(\ell(\Lambda_p)^{-1}+\ell(\Lambda'_p)^{-1}\right) \ll_{c,\lambda}\sqrt P

with the sum restricted to the medium range. Thus the total sieve remainder for medium primes is bounded by a constant times

D0log(2D0)QX(1−δ)/2+D02log(2D0)X1−δ=o(XW(X)2).D_0\log(2D_0)Q X^{(1-\delta)/2} +D_0^2\log(2D_0)X^{1-\delta}=o(XW(X)^2).

Large omitted primes

Suppose n=hpn=hp for a prime p>X1−δp\gt X^{1-\delta}. Then h≤Xδh\le X^\delta, and every survivor of the first sieve has h∈SEh\in S_E. On Λh\Lambda_h, sieve the coordinate n/hn/h by all primes up to z1=Xc1z_1=X^{c_1}, using 0<c1≤c0\lt c_1\le c sufficiently small for level D1=X1/8D_1=X^{1/8}. This upper bounds the requirement that n/h=pn/h=p is prime. Sift mm by EE up to the same range. The retained product is

∏r≤z1(1−1/r)W(z1)≪c1W(X)logX.\prod_{r\le z_1}(1-1/r)\, W(z_1)\ll_{c_1}\frac{W(X)}{\log X}.

The lattice count gives the upper bound Oc1(XW(X)/(hlogX))O_{c_1}(XW(X)/(h\log X)) plus the remainder. If it is mm that has a large prime factor, use Λ′h\Lambda'_h with h≤λXδh\le\lambda X^\delta and exchange the two coordinates. Both lattice minima are bounded below by that of Λ\Lambda. The sum of the sieve remainders for hh in this range is Oλ(Xδ(D1Q+D12)logX)O_\lambda(X^\delta(D_1 Q+D_1^2)\log X), which is o(XW(X)2)o(XW(X)^2).

For Y=λXδY=\lambda X^\delta, Euler’s product gives ∑h≤Y,h∈SEh−1≪W(Y)logY\sum_{h\le Y,\ h\in S_E}h^{-1}\ll W(Y)\log Y. Since W(Y)∼W(X)W(Y)\sim W(X) for each fixed δ\delta, the large primes altogether remove at most Oc1,λ(δXW(X)2)+o(XW(X)2)O_{c_1,\lambda}(\delta XW(X)^2)+o(XW(X)^2). Choose δ\delta small enough after fixing cc and c1c_1 that this contribution is O(εXW(X)2)O(\varepsilon XW(X)^2). The upper and lower estimates then differ from the claimed main term by O(εXW(X)2)+o(XW(X)2)O(\varepsilon XW(X)^2)+o(XW(X)^2). Taking ε\varepsilon arbitrarily small proves the theorem.