Let ϕ be a fixed polynomial with integer coefficients, positive leading coefficient and degree k≥2. We prove that the values 2ϕ(n), 1≤n≤N, can be written as the sum of two primes, apart from at most Nαk+o(1) exceptional indices, where αk<1. For quadratic polynomials the exponent is 1/2, for cubic polynomials 5/8, for quartic polynomials 3/4, and for k>4 it is 1−1/(k+1). The proof combines a fourth-moment estimate over primes with mean-value estimates over arbitrary subsets of polynomial sequences.
Introduction
Let P denote the set of primes. Althöfer and Pintz proved very recently in a joint work [PA2026] that the exceptional set in Goldbach’s problem,
E(X):=#{n∈N:n≤X,2∣n,n=p1+p2 for all p1,p2∈P},(1)
Keywords and phrases: Goldbach problem. Exceptional Goldbach numbers among polynomial values.
This was improved a little later by H. Li and Z. Liu [4] to
E(X)≪εX1/2+ε.(3)
The analytic arguments in these works use a fixed zero-free half-plane for Dirichlet L-functions. The result of OpenAI quoted as Theorem A below supplies the required hypothesis. We use these external results in the precise forms stated in § 2. Under the Generalized Riemann Hypothesis (GRH) for Dirichlet L-functions, the classical result of Hardy and Littlewood [HL1924] already gives the exponent in (3).
Throughout the paper, ϕ∈Z[x] is fixed, has degree k≥2 and has positive leading coefficient. Put
Ek,ϕ(N)Ek,ϕ(N):={n∈N:n≤N,2ϕ(n)=p1+p2 for all p1,p2∈P},:=#Ek,ϕ(N).(4)
The finitely many indices for which ϕ(n)≤0 are included in this definition and cause no difficulty. Since the values have size Oϕ(Nk), a direct application of (3) gives no nontrivial power bound for (4) when k≥2. Perelli proved in 1996 [Per1996] that, for every $A > 0,
Ek,ϕ(N)≪ϕ,AN(logN)−A.(5)
This was improved by Brüdern, Kawada and Wooley [1] to
Ek,ϕ(N)≪ϕN1−c0/k,(6)
with a small absolute constant c0>0. Our goal is to show a stronger explicit estimate for every k≥2.
Theorem 1. Let ϕ∈Z[x] be a fixed polynomial of degree k≥2 with positive leading coefficient. For every ε>0 and N≥2,
Ek,ϕ(N)≪ϕ,εNαk+ε,
where α2=1/2, α3=5/8, α4=3/4, and αk=1−1/(k+1) for k>4.
We use the real smooth weight w specified in § 2 of [4], chosen in logarithmic coordinates from the Gevrey class of order 2. In particular, w∈Cc∞((0,1)), 0≤w≤1, and w(u)=1 for 1/8≤u≤7/8. We use the circle method with
Here N measures indices and X measures polynomial values. In the proof we work with subsets for which X/2<2ϕ(n)≤X; on an index interval (N/2,N] all the required scales satisfy X≍ϕNk.
Write T=R/Z and define the major and minor arcs on this circle by
The norm denotes distance to the nearest integer. We choose P=Xκ with a fixed κ<1/2, so that the individual major arcs are disjoint for sufficiently large X. We usually omit P from the notation.
For even m∈[X/2,X], our goal is to obtain
R(m):=p+p′=mp,p′≤X∑w(p/X)w(p′/X)logplogp′>0(10)
when m=2ϕ(n), apart from a small exceptional set of indices. Put
In the present weighted setting the corresponding goal is
R1(m)=(1+o(1))S(m)Iw(m),(15)
where
Iw(m):=r∈Z∑w(r/X)w((m−r)/X)≥m−X/4−1≫X.(16)
Indeed, both weights are one for integers r∈[X/8,m−X/8]. For our purposes it is sufficient to show, uniformly on the indicated range of even m, that
R1(m)≥cS(m)m(17)
with a fixed c>0.
Auxiliary results for exponential sums over primes
Let θ denote the supremum of the real parts of the nontrivial zeros of all primitive Dirichlet L-functions, including the Riemann zeta function. We use the following result of OpenAI [5 OAI2026b].
Theorem A.θ≤7/8.
For the weight fixed in § 1 and the arcs (9), Li and Liu [4], Theorem 1 and Proposition 4.1, give the following two results.
Theorem B. If θ<11/12, there is a fixed κ=κ(θ)∈(4/9,9/20) such that, with P=Xκ,
∫m∣Sw(α)∣4dα≪θ,wX5/2(logX)100.(18)
Theorem C. If θ<1, then for every fixed κ∈(0,9/20) there is c>0 such that, with P=Xκ, uniformly for even m∈[X/2,X],
Remark. Pintz’s explicit formula [Pin2023] expresses the major-arc contribution in terms of the pole at one and possible zeros near the line Res=1. A fixed zero-free half-plane excludes the latter zeros when X is sufficiently large. For the precise range κ<9/20 needed here we use Theorem C. We choose the same κ in Theorems B and C.
Since S(m)≥2C0>0, Theorem C and (16) imply (17) for sufficiently large X. The integral R1(m) is real: the arcs are invariant under α↦−α on T, and Sw(−α)=Sw(α). The proof below only requires the hypothesis θ<11/12; Theorem A supplies it.
Auxiliary results for exponential sums over polynomial values
To estimate the cardinality of the exceptional set, we use sums over its subsets. For A⊆{1,…,M} put
KA(α):=m∈A∑e(2ϕ(m)α),Z=#A.(20)
We require moment estimates that are uniform in the choice of A. The first follows from Wooley’s weighted form of Vinogradov’s mean-value theorem [Woo2019], Corollary 1.4.
Theorem D. Let k≥1 and s=k(k+1)/2. For every sequence b1,…,bM of complex numbers and every ε>0,
A nonzero summand has ∣hj∣≤sMj. Once h1,…,hk−1 are fixed, the relation in (24) determines at most one integer hk, because ak=0. Thus there are at most
j=1∏k−1(2sMj+1)≪kM1+⋯+(k−1)=MD
potentially nonzero summands. For k=1 the product is empty and equals one. Bounding each summand by J proves (3.3).
For k=3 and k=4 we use the following result of Hughes and Wooley [3], Theorem 4.1, equation (4.2). It is valid for every degree at least three.
Theorem F. If ψ∈Z[x] has degree at least three, A⊆{1,…,M} and Z=#A, then
∫01m∈A∑e(αψ(m))8dα≪ψ,εM1+εZ4.(25)
For degree two an elementary divisor argument suffices.
Lemma G. If ψ∈Z[x] has degree two, A⊆{1,…,M} and Z=#A, then
∫01m∈A∑e(αψ(m))4dα≪ψ,εMεZ2.(26)
Proof. Write ψ(m)=am2+bm+c, with a=0, and put
r(t)=#{(x,y)∈A2:ψ(x)−ψ(y)=t}.
Orthogonality gives
∫01m∈A∑e(αψ(m))4dα=t∈Z∑r(t)2,t∑r(t)=Z2.(27)
For t=0 the factorization
(x−y)(a(x+y)+b)=t(28)
gives at most 2d(∣t∣) choices for the signed divisor z=x−y. Once z is fixed, the other factor is fixed, and
a(2x−z)+b=t/z(29)
determines at most one x. Since ∣t∣≪ψM2, the divisor bound, with a suitably smaller exponent, gives r(t)≪ψ,εMε for t=0. For t=0, either x=y or a(x+y)+b=0, whence r(0)≤2Z. Consequently,
t∑r(t)2≤4Z2+Oψ,ε(Mε)t=0∑r(t)≪ψ,εMεZ2.
Remark. The diagonal solutions give a lower bound Z2 for the moment in (3.7). Thus its dependence on Z is sharp, apart from the factor Mε.
Proof of Theorem 1
Fix ϕ and let N be sufficiently large in terms of its coefficients. For indices N/2<n≤N, positivity of the leading coefficient gives
c1Nk≤2ϕ(n)≤c2Nk
with constants c1,c2>0 depending only on ϕ. This value interval is covered by Oϕ(1) consecutive dyadic intervals (X/2,X], all with X≍ϕNk. For one such interval put
A={n∈Ek,ϕ(N):N/2<n≤N,X/2<2ϕ(n)≤X},Z=#A.
At this scale choose P=Xκ with the fixed κ of Theorem B. Theorem C applies to the same arcs. For every n∈A, R(2ϕ(n))=0, and therefore
Summing (4.1), interchanging the finite sum and the integral, and taking absolute values gives
ZX≪∫mSw(α)2K(α)dα≤∫m∣Sw(α)∣2∣K(α)∣dα.(32)
Suppose that, for some q≥4 and b≥0, we have
(∫01∣K(α)∣qdα)1/q≪ϕ,εNb/q+εZ1/2.(33)
All results of § 3 apply with ψ=−2ϕ. Lemma G gives q=4, b=0 for k=2. Theorem F gives q=8, b=1 for k=3,4. Corollary E gives q=k(k+1), b=k(k−1)/2=q/2−k in general. We write
∣Sw(α)∣2=(∣Sw(α)∣4)1/q(∣Sw(α)∣2)1−2/q(34)
and apply Hölder’s inequality to these two factors and ∣K∣, with exponents q, q/(q−2) and q, respectively. Parseval’s identity and the elementary estimate ∑p≤Xlogp≪X give
If Z=0 there is nothing to prove. Otherwise divide by XZ1/2 and square. Since X≍ϕNk, and since ε may be chosen arbitrarily small, we obtain
Z≪ϕ,εN(k+2b)/q+ε.(36)
Thus the quantity to minimize among the available moment estimates is (k+2b)/q.
For k=2, the choices q=4, b=0 give α2=1/2. For k=3,4, the choices q=8, b=1 give αk=(k+2)/8. For k>4, the choices q=k(k+1), 2b=q−2k give αk=1−k/q=1−1/(k+1). Summing over the Oϕ(1) value intervals proves the required bound for the exceptional indices in (N/2,N]. Summing next over the index intervals (N/2j+1,N/2j], and absorbing the finitely many initial indices into the implied constant, proves Theorem 1. □
Application to other thin sequences
Theorem 1 gives related thin sequences containing only finitely many Goldbach exceptional numbers. For example, consider the values 2mk with m=⌊a(n+β)r⌋, where a>0 is fixed. For almost every β∈[0,1] there are only finitely many exceptions when r>1/(1−αk). Thus it suffices to take r>2 for k=2, r>8/3 for k=3, r>4 for k=4, and r>k+1 for k>4. No parity restriction on m is needed.
Put
Bk={m∈N:2mk=p+q for all primes p,q}.
Assume, for a fixed 0≤ck≤1, that for every ε>0
#(Bk∩[1,Y])≪k,εYck+ε(Y≥2).(37)
Proposition. Fix a>0 and r>1/(1−ck), with 0≤ck≤1 as in (37). For almost every β∈[0,1], every sufficiently large n satisfies
2⌊a(n+β)r⌋k=pn+qn(pn,qn prime).(38)
Moreover, for Aa,r,β={2⌊a(n+β)r⌋k:n≥1}, the counting asymptotic
#(Aa,r,β∩[1,X])∼(2akX)1/(kr)
holds for every fixed β∈[0,1].
Proof. Put mn(β)=⌊a(n+β)r⌋ and Bk,n={β∈[0,1]:mn(β)∈Bk}. The assumption on r implies r>1. For T≤n<2T, the derivative of a(n+β)r with respect to β is ≍a,rTr−1. Hence, for every integer b,
meas{β∈[0,1]:mn(β)=b}≪a,rT1−r.(39)
The intervals [anr,a(n+1)r] have disjoint interiors. Their lengths tend to infinity since r>1. For sufficiently large T, each unit interval [b,b+1) therefore meets at most two of them, apart from irrelevant endpoints. It follows from (37) and (39) that
T≤n<2T∑meas(Bk,n)≪a,rT1−r#(Bk∩[1,Ca,rTr])
≪a,r,k,εT1−r+r(ck+ε)=T1−r(1−ck−ε).(40)
Choose ε>0 so that r(1−ck−ε)>1. Summation over T=2j gives ∑nmeas(Bk,n)<∞. The first Borel–Cantelli lemma proves (38); no independence assumption is required. Also mn(β)≥1 for all sufficiently large n, so that the omission of 0 from Bk is harmless. Finally, mn(β)∼anr and the values are eventually strictly increasing. Inverting 2mn(β)k∼2aknkr proves the counting asymptotic.
Remark. The measure argument does not provide an explicit admissible β. For example, can one exhibit a particular β∈[0,1] for which
2⌊(n+β)3⌋2(41)
is the sum of two primes for every sufficiently large n? § 5 proves this for almost every β, by taking k=2, a=1 and r=3. The particularly natural choice β=0 would assert that every sufficiently large number 2n6 is Goldbach, a substantially stronger conclusion than the exceptional-set estimates proved in this paper. We do not know any triple (a,k,r) with a>0, integer k>0, r>0 for which we could specify a value β∈(0,1] such that the expression (38) would be a sum of two primes for every value of n.
AI assistance. OpenAI Astra was used for careful auditing. Its checks concerned the proof, references and notation. It also assisted with the revision and with formulating the above remark.
Ingo Althöfer Faculty of Mathematics and Computer Science Friedrich-Schiller University Jena, D-07740 Jena, Germany E-mail: [email protected]
Dijia Chen Department of Mathematics, University of Wisconsin–Madison Madison, WI 53706, USA E-mail: [email protected]
Alberto Perelli Dipartimento di Matematica, Università di Genova Via Dodecaneso 35, 16146 Genova, Italy E-mail: [email protected]
János Pintz HUN-REN Alfréd Rényi Institute of Mathematics Réáltanoda u. 13–15, H-1053 Budapest, Hungary E-mail: [email protected]
References
[1]J. Brüder[n], K. Kawada and T. D. Wooley, Additive representation in thin sequences, II: The binary Goldbach problem, Mathematika 47 (2000), 117–125.
[3]K. Hughes and T. D. Wooley, Discrete restriction for (x, x³) and related topics, International Mathematics Research Notices 2022, no. 20, 15612–15631. doi:10.1093/imrn/rnab113. Theorem 4.1 and Conjecture 4.2.DOI
[4]H. Li and Z. Liu, Fourth moments of prime exponential sums and exceptional sets for sums and differences of primes, arXiv:2610.26756v1, 19 October 2026. Theorem 1, Proposition 4.1 and Section 2.arxiv.org/abs/2610.26756
[5]OpenAI, The Quasi-Riemann Hypothesis: A Zero-Free Half-Plane Re s > 7/8, OpenAI Math Release preprint, 30 September 2026, family 003, Theorem 1.1. Archived repository version at commit fd4aeeb2ee4f.