WorksMathematics
The three parts of Erdős Problem 12: counting functions and reciprocal sums
Abstract
We give a unified account of the three numbered questions in Erdős Problem 12 for sets in which no element divides the sum of two distinct larger elements. The positive answer to the square-root counting question and the negative answer to a universal power-saving bound are prior AlphaProof Nexus results of Google DeepMind, presented here with attribution. For the reciprocal-sum question, the new argument proves that one absolute constant bounds the reciprocal sum for every such set, finite or infinite. It combines a restriction estimate for sifted integers, congruence restrictions on prime fibres, and a predictable sum over random prime-factor prefixes. A companion Lean development checks the three original questions and the uniform reciprocal bound.
The three questions and their attribution
A set has Property if there are no pairwise distinct elements such that
In particular, equality of the two larger summands is permitted by the formulation and does not constitute a forbidden triple. Write . Erdős Problem 12 asks three separate questions about the counting function and reciprocal sum of such an infinite set.
Theorem 1.1 (Part (i), AlphaProof Nexus). There is an infinite set with Property such that
Theorem 1.2 (Part (ii), AlphaProof Nexus). There is no absolute such that every infinite set with Property satisfies for infinitely many . More precisely, for each there are an infinite set with Property and an integer such that
The set and the threshold may depend on .
Theorem 1.3 (Part (iii), with a uniform bound). There is an absolute constant such that every set with Property satisfies
The same constant applies to finite and infinite sets.
The answers to the three numbered questions are therefore, respectively, yes, no, and yes. The first two answers are prior results of the Google DeepMind AlphaProof Nexus work of Tsoukalas, Kovsharov, Shirobokov et al. [11]. They are included here with attribution, exposition, and adaptations of the released Lean proofs [4, 5]; they are not contributions first obtained in this manuscript. The reciprocal-sum argument of Theorem 1.3 is the argument developed in the original manuscript by the four named authors. The present version brings the three statements into one account and one checked Lean source file.
The quantifier order in (1) matters: . We use precisely this parameterwise assertion. The construction below does not assert that one fixed set works for every . The lower bounds in part (ii) are compatible with the uniform reciprocal bound in part (iii): the onset and the set itself vary with the exponent.
Erdős and Sárközy [3] proved that every infinite set with Property P has density zero, and formulated the reciprocal-sum problem and its stronger uniform version. For the counting questions, § 2 follows the congruence-block constructions in the released AlphaProof Nexus Lean files. Those are the concrete proof instances incorporated in the companion formalization. The constructions are presented here directly, rather than by relying on admitted statements in the Formal Conjectures catalog.
The proof of part (iii) has three main steps. A quantitative incompatibility between separated blocks of rough elements is combined with sieve estimates to control prime fibres carrying inherited congruence restrictions. We then represent reciprocal sums by random prime-factor prefix weights. An adapted rule for selecting the restrictions gives a uniform total expected weight. The restriction estimate is a special case of Bera–Viswanadham [2]; § 3 includes a direct proof in the form needed here. The Mertens prefix weights are classical; see Alexeev et al. [1]. Earlier work of Plasma AI [8] contains the reduction to odd primitive sets, restricted reciprocal-sum estimates, and the problem of controlling accumulated congruence restrictions. These attributions apply to the reciprocal-sum argument, independently of the provenance of parts (i) and (ii).
Notation and conventions
Notation and conventions. All logarithms are natural unless a base is indicated. We write , and for distance from a real number to the nearest integer. The constants implicit in and are absolute unless a dependence is indicated. An integer is -rough if no prime factor is at most , and -smooth if every prime factor is at most . The integer satisfies both conventions. We write for the largest prime factor of , with . A set is primitive if no two distinct members divide one another. Property P is inherited by subsets and by division by a common divisor. The paper uses positive integers; the Lean statements use natural numbers, with . The constructions below consist of positive integers.
The counting constructions: parts (i) and (ii)
This section is an attributed exposition of the released proofs [4, 5]. We isolate the common congruence argument and then make the two choices of scale. Omitting finitely many initial blocks simplifies the estimates and does not affect the asserted asymptotic statements. The Lean implementations verify explicit initial bounds as well.
Lemma 2.1 (Congruence separation). Let be finite sets of positive integers, with every element of smaller than every element of when . Suppose that there are pairwise distinct primes such that elements of are (mod ) and (mod ) for every . If each lies in an interval with , then has Property P.
Proof. Suppose , , and . If either larger element lies in a later block, then is or (mod ), whereas (mod ). This is impossible because . Otherwise both larger elements lie in , and , which precludes divisibility.
A positive square-root lower limit
Choose increasing primes , starting with , so that . Bertrand’s postulate gives such a choice, and induction yields . Put
The Chinese remainder theorem supplies such that and . Define
Then . Also , while . Consequently
Thus, after removing finitely many initial blocks, each lies in . The scales are strictly separated because . The residues of have the form in Lemma 2.1, since for . Their union therefore has Property P.
For large ,
If falls inside the short interval containing , the preceding complete block contributes elements. If falls in the following gap, block contributes elements. These two cases give
for every sufficiently large . To verify that the lower limit is finite as well, note that , so , up to the omitted finite prefix. Since ,
This proves Theorem 1.1. In particular, the lower limit is bounded away from zero and bounded above; this agrees with the real-valued lower-limit convention used by the Lean statement.
Failure of a universal power saving
Fix . Choose distinct odd primes with , and put . Then . Choose an integer with , and define
Choose a residue that is and for , and set
The intervals are disjoint and ordered because . They satisfy the strict width inequality in Lemma 2.1. Once again their union has Property P, and it is infinite because the blocks are nonempty for all sufficiently large .
Counting a residue class in an interval gives
The logarithm to base 2 of is at least , whereas
The difference has leading term , whose coefficient is positive. Therefore, for all sufficiently large ,
For , the complete block is already counted in . This proves (1). If a universal saving existed, apply the construction with . For sufficiently large , , contradicting the required infinitely many strict inequalities. This proves Theorem 1.2.
A restriction estimate for sifted integers
Let be any set of primes. Define a multiplicative function supported on squarefree products of primes in by
and put for all other positive integers . For , set
Using strict inequalities at the sieve cutoff removes endpoint ambiguities.
Lemma 3.1. Let , and let be a finite -separated set. For every fixed and every complex sequence supported on ,
The constant is uniform in .
Proof. We give the enveloping-sieve argument in the form needed here. For a parameter , let , write , and put . For , define
The mean of modulo is zero. Its nonzero additive Fourier coefficients all equal , and
Consequently the nonnegative periodic function
equals on . Orthogonality and the Chinese remainder theorem show that its constant Fourier coefficient is .
A nonconstant Fourier coefficient can occur only at a reduced fraction , where is squarefree and supported on . Its absolute value is at most
where is the number of distinct prime factors. Indeed, in a term , a prime dividing has at most three membership patterns in , and contributes a coefficient of magnitude at most . A prime outside must occur in both factors to contribute a constant term, which is . Summing those common factors costs at most . Finally, for all sufficiently large primes, and the finitely many remaining primes contribute an absolute factor. Separating the prime factors below gives
For a nonzero vector , put
Choose a nonnegative integrable function majorizing on , with Fourier transform supported on , integral , and decay . One construction is an absolute multiple of the convolution of with
Poisson summation implies
Use to majorize the dual square sum. The constant coefficient of contributes only when . For a fixed ordered pair , at most one reduced nonzero fraction , , can satisfy , since distinct such fractions have distance at least . Every nonzero denominator in (5) is at least . We obtain
If , then and the second term is absent. Since , , and (6) holds, this gives
Write
The one-frequency case of (7) gives . For , choose unit complex phases realizing the absolute values in the pairing with . Cauchy–Schwarz and (7) imply
whenever . Thus on the bounded range where the level set is nonempty. Integrating this estimate against proves (3). □
Lemma 3.2. There are absolute constants and with the following properties. If , , and , then
If , then
Proof. Put , where is a sufficiently small absolute constant. In the Euler product
normalize the weight of each squarefree product to a probability distribution. A prime is then selected with probability . The expected logarithm of the product is
Markov’s inequality shows that at least half the total weight is carried by products smaller than , after is fixed sufficiently small. Mertens’ estimates give
Choose small enough that for all sufficiently large . Since , this proves (9). The same argument with all primes at most proves (8). □
Interaction between separated rough blocks
Lemma 4.1. Let , let be powers of two with , and put . Assume . Suppose that and are -rough subsets of one set with Property P. With
one has
Proof. The one-frequency case of Lemma 3.1, together with (8), gives . We may assume both sets are nonempty.
Work on , where , with normalized Fourier transform
Function-space norms are normalized by , while Fourier-space -norms use counting measure. Set
Let , , and . Define
All the products in this definition are smaller than . Its Fourier transform and mean satisfy
Every prime factor of is at most . The -rough part of is therefore exactly . In particular the representation determines uniquely, and
The support of avoids all primes in . For , put
Then has mean , is everywhere, and is at least on the even integers in . Indeed, , and this latter interval is contained in the doubled interval for every . Write .
A uniform Fourier comparison
We claim that
Fix a frequency , and put . Throughout this comparison take .
First suppose that for every . The frequencies , , are -separated. Lemma 3.1, applied on and normalized by , gives an -norm on these frequencies. Hölder’s inequality then yields
The geometric average over also gives .
Otherwise, for some , reduce the nearby rational approximation to obtain
If , every is a unit modulo , so
The geometric average over therefore gives .
If , divide into consecutive chunks of length at most . Within a chunk the frequencies are separated by at least , since . Apply Lemma 3.1 to each chunk and then average over . This gives
In both cases with , the interval average defining is bounded by
This term is smaller than the right side of (12), after decreasing if needed.
It remains to consider . Use the lift of with , and put . If , geometric summation gives
If , couple the uniform measure on to the grid by intervals . The endpoint adjustment has total variation , and the phase changes by inside each interval. Hence
These bounds prove (12), since and .
The forbidden correlation
By (11),
Apply Lemma 3.1 to , using the prime band , the full -spaced frequency grid, and (9). The result is
Since , Parseval gives , and thus . Interpolating with (12), with , gives
The full-prime restriction estimate likewise gives . Fourier inversion and Hölder’s inequality, with exponents and , now show that replacing by in
changes the expression by at most .
Every element of is odd. Thus is an even integer in whenever , and the expression with is at least . The expression with vanishes. A nonzero term would give the integer equality with and in the underlying set. Distinct contradict Property P. If , then , which is impossible because is -smooth and is -rough. Therefore , which proves (10). □
Prime sets with inherited congruence restrictions
We use the analytic large sieve and the Brun–Titchmarsh inequality in their classical forms [6, 7]. In particular, if a set of primes occupies at most reduced residue classes modulo , then, for a fixed , , , and , partial summation gives
Indeed, Brun–Titchmarsh bounds the counting function by for .
Lemma 5.1. Every set of primes with Property P satisfies , with one absolute constant.
Proof. Remove the possible prime , which costs at most . If the remaining set is nonempty, let be its least element. The reciprocal sum over is by the ordinary upper bound for primes and partial summation.
Enumerate the primes of greater than as , and set
For every earlier prime modulus , the later primes occupy at most one residue in each pair of nonzero residues modulo . Otherwise two distinct later primes have sum divisible by . The Chinese remainder theorem therefore gives a fraction at most of the reduced classes modulo . Moreover . Equation (14) bounds the mass in by . These intervals cover all remaining primes of . For a finite list, its last prime is smaller than the final product squared, so the same argument applies. The resulting geometric series proves the claim. □
Lemma 5.2. There are absolute constants and with the following property. Let be a set of primes with Property P. Let be distinct odd primes, and set
with if . Suppose and, for each , the residues of occupy at most one member of every pair of opposite nonzero classes modulo . Then
Proof. For bounded , the condition bounds and . Lemma 5.1 then suffices. We henceforth take larger than a fixed absolute threshold. The choice of , made below in terms of , is fixed once and for all.
The Chinese remainder theorem puts all primes of in at most reduced residue classes modulo . For any dyadic , Brun–Titchmarsh gives
Here , and . For , use the ordinary prime upper bound. We will also use
Call the dyadic bin at light if its cardinality is at most . All light bins together contribute to . If there is no other bin, the proof is complete.
Let , , be the first bin that is not light, and put . Choose
distinct primes from that bin. This is possible for all sufficiently large , by (17). Let be times their product, and set
Choose sufficiently large that, uniformly for ,
for all sufficiently large . This follows from (17). The same range gives .
The near range
Take . In dyadic bins meeting , every prime is -rough. The normalized count
is , by (16). If two such counts are at least , then Lemma 4.1 bounds their ratio of scales by , unless the scales differ by a factor less than 16. The hypotheses of that lemma hold throughout the range by (18). It follows that
Integrating this bound over gives
The middle range
Let , where . For every prime , the set occupies at most classes modulo . Orthogonality and Cauchy–Schwarz give
The reduced fractions belonging to different such primes are distinct and have spacing at least . The analytic large sieve therefore gives
Summing over bins meeting , including the at most two boundary bins, yields
For the first inequality use and . For the second, divide by and use
uniformly in the permitted range, by (17).
The far range
On , use the external and the smallest selected internal primes as moduli. Their product is at most , hence at most for sufficiently large . The occupied fraction of reduced classes is at most . Equation (14) therefore bounds the mass in this range by .
Enumerate the primes of above as , and now put , with . The primes in are at least , and occupy at most the fraction of the reduced classes modulo . With
one has . Another application of (14) gives mass . As in Lemma 5.1, these intervals also cover every term of a finite list. Altogether the far range contributes
Combine this with the light-bin bound, (19), and (20). Since and , we obtain (15). □
Prime-factor prefixes
Let be an odd primitive set with Property P. If , then , so suppose . For each positive integer , define
Each has Property P, and all its primes are odd. Every element of has a unique representation in which is its largest prime factor. Hence
The term at is , by Lemma 5.1.
Independently for all primes , let have geometric distribution
Form the ordered list consisting of copies of , with primes in increasing order. A prefix product is the product of an initial finite segment of this list.
Lemma 6.1. For , the probability that is a prefix product is
Proof. Put . For primes , the valuation must equal . At , the condition is , which has probability . Independence gives the exact formula, and Mertens’ product estimate gives the comparison. □
Assign to a prefix the score
Tonelli’s theorem and (22) show that it is enough to prove
Lemma 5.2 with gives . More precisely, if and , then
We reserve for this least terminal prime and use for the prime coordinate of the prefix.
We may omit scores produced by the second or a later copy of a prime. Their total expectation is bounded by
The omitted copies remain in the list and in all later prefix products. Scores at primes below any fixed absolute threshold also have bounded total expectation. Thus only the first copy at each sufficiently large prime coordinate needs to be considered.
Lemma 6.2. Suppose and are prefix products on one path, occurs earlier, , and . Then the residues of occupy at most one class from each opposite pair of nonzero classes modulo . The terminal primes from distinct nonempty prefixes on a path are distinct.
Proof. We have . If , then divides every , . These are distinct elements because , contrary to primitivity. Thus . All primes in exceed . For distinct , the congruence would now imply
with both summands larger than , contradicting Property P. This proves the first assertion. If two distinct nested prefixes had the same terminal prime , then the distinct elements would contradict primitivity. □
An adapted summation of prefix scores
We process every prime coordinate in increasing order, including coordinates for which . At coordinate , reveal , retain all its copies in the prefix product, and score the first copy if . Let be the exponent in Lemma 5.2. The following scheduling rule uses selected terminal primes as congruence moduli, which we call guards.
After guards have been activated, write for their product and . A ready stage starts at the first unprocessed prime at least , where is determined by the previously revealed coordinates. Initially and is a fixed sufficiently large absolute threshold. A stage starts with no pending guard.
At a scored prefix with , schedule , retaining the smallest pending prime. If the newly scheduled minimum equals the current coordinate, activate it immediately after scoring. A guard already pending when its coordinate is reached is activated before revealing that coordinate, so that the coordinate belongs to the delay described below. The stage ends on activation.
Every finite pending prime is eventually reached, since all prime coordinates are processed. A ready stage that never activates therefore has no nonempty scored prefix. The guard selected in a completed stage arose from a previously visited prefix. Lemma 6.2 shows that the old guards impose the required opposite-class restrictions on every nonempty fibre in that ready stage. They are distinct and smaller than its current coordinate.
At a nonempty scored prefix with , Lemma 5.2, applied with , gives
The harmless factors in (24) are absorbed by the fixed lower threshold for .
Lemma 7.1. *Conditional on the history at the start of a ready stage with active guards, the expected total score in that stage is .
Proof. Call a change of the pending minimum large if its logarithm decreases by a factor of at least 2. The first finite pending value is counted as large. At consecutive large changes, the ratio at least doubles, since increases and the pending logarithm halves. Every such ratio is at most 1 before activation. Thus the sum of these ratios over all large changes is at most 2 on every path. Their total score is bounded by , by (25).
For all other nonempty scored prefixes, let be the pending prime known before revealing . Then
Interpret the right side as zero when no finite prime is pending. In that case the first nonempty prefix has already been assigned to the large-change class.
Both readiness at and are measurable before is revealed. A first-copy score at can occur only if , an event of conditional probability . If is the prime eventually activated, every ready coordinate is at most , and every previous finite pending value is at least . Consequently, pathwise,
For a stage without activation the sum is zero. Taking conditional expectation before each coordinate, and then applying Tonelli, proves the same bound for all remaining scores. □
After activating , put . Process the remaining coordinates below as a delay, then start the next ready stage at the first unprocessed prime at least that threshold. The old guards remain valid throughout the delay.
Lemma 7.2. Conditional on the activation history, the expected total score in the delay is .
Proof. Equation (25) bounds every scored prefix during the delay by , using and the old guards. Moreover , so
Conditional on the activation history, all strictly later prime coordinates retain their independent geometric laws. Their expected number of first copies before the next ready stage is at most
by Mertens’ estimate. If was pending before its coordinate, that coordinate is also in the delay and contributes at most one additional score. If the guard was scheduled at its own coordinate, that score was already counted in the ready stage. This proves the claim. □
Proposition 7.3. Every odd primitive set with Property satisfies , with one absolute constant.
Proof. Stage-start and activation histories are measurable at their respective prime coordinates. The conditional expectation statements in Lemmas 7.1 and 7.2 can therefore be applied on each event specifying these coordinates. All such coordinates are countable, and the rule never conditions on a future valuation.
Each completed ready stage adds one new guard. Summing the two lemmas over , and restoring the omitted small-prime and repeated-copy scores, gives
Equation (23), Lemma 6.1, and the decomposition (21) complete the proof.
Completion of the proof
Proof of Theorem 1.3. It is enough to bound uniformly for finite sets with Property . Let consist of the elements minimal under divisibility. Every has a proper divisor in . Assign to one such divisor . No receives two distinct elements, since it would divide their sum and both would exceed it. Every assigned is at least . Hence
The set is primitive.
For , define
Each is an odd primitive set with Property . Proposition 7.3 supplies one constant bounding all their reciprocal sums, including the possible singleton . Therefore
This bound is independent of . Apply it to the finite truncations of an arbitrary , and use monotone convergence to obtain (2).
An explicit finite prefix bound
We give a finite formulation of the argument in § 7, with explicit bounds for the total expected score. Fix an odd primitive set with Property . In this section, the fibres and scores are defined using in place of . Independently for each prime , choose with
For , summing the geometric series gives . Thus, for every , the event that is a prefix has exactly the probability in Lemma 6.1. Every expectation in this section is a finite sum.
Apply the scheduling rule of § 7, processing integer times and changing the state only at prime coordinates. Charge an activation coordinate using its incoming set of guards. Let be the number of guards activated before time , and put
Since and , the quantity is at an activation and zero otherwise. In particular,
Choose uniformly in the prime-tail estimate so that and, for a ready first-copy prefix with ,
Increase the fixed initial threshold if necessary.
The ready coordinates
If the pending terminal prime before time is , define
Set both quantities to zero when no prime is pending. A pending prime is at least the current coordinate. Between activations its value can only decrease.
Write for the first-copy score charged at a ready time, and set it to zero at other times. Then
where the indicator is zero at a composite time. To verify this inequality, first suppose that no activation occurs. If a nonempty prefix is scored, the new pending prime is at most its terminal prime , and consequently
On this event the first three terms in parentheses in (29) sum to . If no score occurs, . At an activation, , , and , so the activation term covers both the lost potential and the score. During a delay there is no pending prime and these quantities are zero.
The state before time depends only on earlier coordinates. Independence therefore gives, for prime ,
Choose such that for . The weighted prime sum and the monotonicity of the pending minimum imply
Indeed, without activation the increment of the numerator accounts for , and a decrease of the denominator increases the potential. At an activation, the numerator through time is at most , so the lost potential together with the current summand is at most .
Every terminal prime of a nonempty fibre is at most . Hence no prime is pending after time , and both potentials begin and end at zero. Summing (30) and using (28) gives
Summation of (29) now bounds the expected ready contribution by
The delayed coordinates
Suppose that the latest activation was at a prime , with guards active before that activation and old product . Put . The inequality gives . Mertens’ estimate supplies an absolute constant , depending only on the fixed , such that
for every possible activation prime . Define the accumulated mass after this activation by
Then . For , so that , set
Before the first activation set . These quantities are nonnegative. At a delayed prime coordinate the increment of is , so decreases by that amount. At a ready coordinate without activation it is constant.
At a new activation, the previous remaining term is nonnegative. The new value satisfies
If the activation coordinate belongs to the delay, the new already includes . In this case the first inequality above holds with the stronger left side , so the potential pays this coordinate as well. Thus the total predictable delayed cost is at most . Each delayed first-copy score is at most , and its occurrence has conditional probability . Its total expected contribution is therefore at most
The finitely many initial coordinates contribute an absolute constant. The repeated-copy contribution is bounded using
Combining these estimates with (31) and (32) yields a bound for a prefix independent of and . Lemma 6.1 and (21) then give the uniform reciprocal bound for finite odd primitive sets. § 8 extends it to all sets with Property P.
The single-file formalization and provenance
The companion file Erdos12AllParts.lean imports only Mathlib. It contains the local dependency closure of the reciprocal-sum argument and the two attributed AlphaProof Nexus proof developments. Its public interface is as follows.
| Entry point | Mathematical statement |
part_i | A good infinite set has positive lower limit at square-root scale. |
part_ii | The universal power-saving assertion is false. |
part_iii | Every good infinite set has a convergent reciprocal sum. |
all_parts | The conjunction of the three preceding answers. |
parameterwise_<br>dense | The set and eventual counting threshold may depend on the exponent. |
uniform_<br>reciprocal_bound | One constant bounds the reciprocal sum of every set with Property P. |
Table 1.
All entry points belong to the namespace Erdos12AllParts. The definition IsGood retains the quantifiers and strict inequalities of the original Formal Conjectures statement.
The check uses Lean 4.35.0-rc2 and Mathlib commit 065356127b1dc0016f66b7283ce0ce2c4055aa55. The prior proof files were taken from Google DeepMind’s alphaproof-nexus-results repository at commit 0647711a71183c1ea492ad60860776617ce1ea88. They are distributed under the Apache License, Version 2.0, with copyright 2025 Google LLC. The companion file retains these notices and the license. Their original environment used Lean 4.27.0; the adaptation isolates their namespaces, removes redundant options and imports, updates the image-cardinality API, and expands the interval-residue counting lemma into a direct proof.
The analytic formalization also retains the Euler–Maclaurin and Mertens developments from PrimeNumberTheoremAnd, obtained via chen_theorem, and the vendored large-sieve development from goldbach-lean [9, 10]. Their recorded source revisions and attributions are preserved in the Lean header and the validation/provenance files; these components are likewise recorded under Apache License 2.0 in the supplied project.
The final checks compile the combined source, check each public entry point, and inspect its axiom dependencies. The only axioms reported are propext, Classical.choice, and Quot.sound. The file introduces no admitted theorem and does not obtain the first two results by importing placeholder catalog declarations. A separate check compares all three endpoints with the pinned original catalog statements. The validation receipt records the source digest and commands.
Acknowledgments and contribution statement
Parts (i) and (ii), including the constructions and the released formal proofs adapted here, are credited to the AlphaProof Nexus work of Google DeepMind and Tsoukalas, Kovsharov, Shirobokov et al. [11]. The present manuscript supplies a unified exposition and the recorded adaptation to the pinned Lean environment. The four named authors retain responsibility for the reciprocal-sum argument, the integration, and the final mathematical claims.
ChatGPT and Codex, developed by OpenAI, assisted with the mathematical development of the reciprocal-sum argument, checking calculations, literature searches, source integration, and manuscript preparation. The earlier reductions and analytic ingredients used in part (iii) are credited in the introduction and body of the proof.
References
References
- [1]B. Alexeev, K. Barreto, Y. Li, J. D. Lichtman, L. Price, J. I. Shah, Q. Tang and T. Tao, Primitive sets and von Mangoldt chains: Erdős Problem #1196 and beyond, preprint (2026), arXiv:2605.00301v1.arxiv.org/abs/2605.00301
- [2]T. Bera and G. K. Viswanadham, Restriction estimates with sifted integers, preprint (2026), arXiv:2512.21640v3.arxiv.org/abs/2512.21640
- [3]P. Erdős and A. Sárközy, On the divisibility properties of sequences of integers, Proc. London Math. Soc. (3) 21 (1970), 97–101. doi:10.1112/plms/s3-21.1.97.DOI
- [4]Google DeepMind, AlphaProof Nexus results: Erdős Problem 12(i), Lean source (2026), Apache License, Version 2.0. Repository commit 0647711a71183c1ea492ad60860776617ce1ea88. https://github.com/google-deepmind/alphaproof-nexus-results/blob/0647711a71183c1ea492ad60860776617ce1ea88/APNOutputs/ErdosProblems/erdos_12_parts_i.lean.
- [5]Google DeepMind, AlphaProof Nexus results: Erdős Problem 12(ii), Lean source (2026), Apache License, Version 2.0. Repository commit 0647711a71183c1ea492ad60860776617ce1ea88. https://github.com/google-deepmind/alphaproof-nexus-results/blob/0647711a71183c1ea492ad60860776617ce1ea88/APNOutputs/ErdosProblems/erdos_12_parts_ii.lean.
- [6]H. L. Montgomery and R. C. Vaughan, The large sieve, Mathematika 20 (1973), 119–134.
- [7]H. L. Montgomery and R. C. Vaughan, Multiplicative Number Theory I. Classical Theory, Cambridge Studies in Advanced Mathematics, vol. 97, Cambridge University Press, Cambridge, 2007.
- [8]Plasma AI, Composite-modulus large sieve for Property P, research notes, Erdős Problems (2026), online edition, accessed 4 November 2026.
- [9]PrimeNumberTheoremAnd, formalization project, https://github.com/AlexKontorovich/PrimeNumberTheoremAnd, commit c6c73610b406689c4b58325cbe1342ecca25d755; retrieved through https://github.com/lixiang90/chen_theorem, commit 11dcca8fc3ab340358503278c46d72aef072a6e7.
- [10]goldbach-lean, large-sieve formalization, https://github.com/subfish-zhou/goldbach-lean, commit 9d4c3a4010d42aedd9e7231bf94386019cfffccc.
- [11]G. Tsoukalas, A. Kovsharov, S. Shirobokov et al., Advancing mathematics research with AI-driven formal proof search, preprint (2026), arXiv:2605.22763v2, Appendix B.4.arxiv.org/abs/2605.22763