betaFeedback

WorksMathematics

Asymptotically full minimal signed unit-fraction identities

Abstract

Let f(N)f(N) be the largest size of a subset of {1,…,N}\{1,\ldots,N\} admitting a signed zero sum of its reciprocals with no proper nonempty signed zero subsum. We prove f(N)=N−o(N)f(N)=N-o(N) using the subset-sum theorem of Liu and Sawhney. The construction joins finitely many large-prime groups by one CRT-chosen unit fraction. The resulting positive sum can be arbitrarily large, although the density cost of excluding those primes is arbitrarily small. A dense coprime-denominator expansion of that sum provides the opposite side and preserves minimality. Earlier public work claims this asymptotic and the protected-prime construction. This version gives an alternative approach through Liu and Sawhney, with an accompanying Lean formalization.

Statement

For a finite set B⊂NB \subset\mathbb{N}, write R(B)=∑n∈B1/n\mathcal{R}(B)=\sum_{n\in B}1/n. Call A⊆{1,…,N}A\subseteq\{1,\ldots,N\} primitive if there are signs δn∈{−1,1}\delta_n\in\{-1,1\} satisfying

∑n∈Aδnn=0,∑n∈Cδnn≠0for every ∅≠C⊊A.\sum_{n\in A}\frac{\delta_n}{n}=0,\qquad\sum_{n\in C}\frac{\delta_n}{n}\ne0\quad\text{for every }\varnothing\ne C\subsetneq A.

Let f(N)f(N) be the maximum of ∣A∣|A| over these sets.

Theorem 1. For every ε>0\varepsilon>0 and every sufficiently large NN, there is a primitive A⊆{1,…,N}A\subseteq\{1,\ldots,N\} with ∣A∣≥(1−ε)N|A|\ge(1-\varepsilon)N. Consequently

lim⁡N→∞f(N)N=1.\lim_{N\to\infty}\frac{f(N)}{N}=1.

The lower bound (1−1/e+o(1))N(1-1/e+o(1))N follows from a dense expansion of 11 and the singleton {1}\{1\} [3].

Earlier public candidate proofs of the density-one asymptotic are recorded in [4, 5]; their existence precludes treating the present result as an uncontested new solution. Our positive side is a fixed, possibly enormous, finite set once ε\varepsilon is chosen. Only the opposite side grows with NN. The result is asymptotic and supplies no practical threshold for NN.

A finite positive set with only one coprime subsum

For a squarefree integer DD, call a rational number DD-integral if its reduced denominator is coprime to DD. This includes 00 and all integers. The DD-integral rationals form an additive subgroup of Q\mathbb{Q}.

Lemma 2. Let K≥1K \ge1, L=lcm⁡(1,…,K)L = \operatorname{lcm}(1,\ldots,K), and B=LHKB = LH_K, where HK=∑m=1K1/mH_K = \sum_{m=1}^{K} 1/m. Let P\mathcal{P} be a finite set of at least two primes, each greater than BB, and put D=∏p∈PpD = \prod_{p \in\mathcal{P}} p. There exists a positive integer tt coprime to DD such that the set

E={pm:p∈P, 1≤m≤K}∪{Dt}E = \{pm : p \in\mathcal{P},\ 1 \le m \le K\} \cup\{Dt\}

has the following properties:

  1. All its denominators are distinct, and all have a prime factor in P\mathcal{P}.

  2. r:=R(E)r := \mathcal{R}(E) is DD-integral and

r=HK∑p∈P1p+1Dt.r = H_K \sum_{p \in\mathcal{P}} \frac{1}{p} + \frac{1}{Dt}.
  1. The only subsets C⊆EC \subseteq E for which R(C)\mathcal{R}(C) is DD-integral are ∅\varnothing and EE.

Proof. Since p>B≥Kp > B \ge K, both LL and BB are invertible modulo every p∈Pp \in\mathcal{P}. Choose tt by the Chinese remainder theorem so that

tB(D/p)+L≡0(modp)(p∈P).(1)tB(D/p) + L \equiv0 \pmod{p} \qquad(p \in\mathcal{P}). \tag*{(1)}

The prescribed residue −L(B(D/p))−1-L(B(D/p))^{-1} is nonzero modulo pp, so gcd⁡(t,D)=1\gcd(t,D) = 1. Take its positive representative, or add a multiple of DD.

The groups {pm:m≤K}\{pm : m \le K\} are disjoint: a denominator in the pp-group has no other prime factor in P\mathcal{P}. The connector DtDt has all these prime factors, so is distinct from every group denominator.

For a subset C⊆EC \subseteq E, let

Jp={m≤K:pm∈C},bp=∑m∈JpL/m.J_p = \{m \le K : pm \in C\}, \qquad b_p = \sum_{m \in J_p} L/m.

and let h∈{0,1}h \in\{0,1\} indicate whether Dt∈CDt \in C. Every bpb_p is an integer in [0,B][0,B], with bp=0b_p = 0 exactly for an empty JpJ_p, and bp=Bb_p = B exactly for a full JpJ_p. Multiplying R(C)\mathcal{R}(C) by the common denominator DtLDtL gives the integer

UC=t∑p∈P(D/p)bp+hL.(2)U_C = t \sum_{p \in\mathcal{P}} (D/p)b_p + hL. \tag*{(2)}

Each p∈Pp \in\mathcal{P} occurs to exponent exactly one in DtLDtL. Thus R(C)\mathcal{R}(C) is DD-integral if and only if every such pp divides UCU_C.

If h=0h = 0, reduction modulo pp gives bp≡0(modp)b_p \equiv0 \pmod p. As 0≤bp≤B<p0 \le b_p \le B < p, this forces bp=0b_p = 0 for every pp, hence CC empty. If h=1h = 1, subtracting (1) from the condition p∣UCp \mid U_C gives bp≡B(modp)b_p \equiv B \pmod p. The same bound forces bp=Bb_p = B for every pp, hence C=EC = E. Finally, (1) directly makes UEU_E divisible by every pp, so rr is DD-integral. The formula for rr is immediate.

Lemma 3. For every 0<δ<1/40 < \delta< 1/4 and R>0R > 0, there are a squarefree DD and a finite set EE as in Lemma 2 such that

ρ:=φ(D)D≥1−2δ,r=R(E)≥R.\rho:= \frac{\varphi(D)}{D} \ge1 - 2\delta,\qquad r = \mathcal{R}(E) \ge R.

Proof. Choose KK so that HK≥R/δH_K \ge R/\delta, and then choose primes p>max⁡{LHK,1/δ}p > \max\{L H_K,1/\delta\} until

δ≤∑p∈P1p<2δ.\delta\le\sum_{p\in\mathcal{P}} \frac{1}{p} < 2\delta.

Such a finite choice exists because the sum of reciprocal primes diverges. Each summand is less than δ\delta, so at least two primes are chosen. The finite product inequality gives

ρ=∏p∈P(1−1/p)≥1−∑p∈P1/p>1−2δ.\rho= \prod_{p\in\mathcal{P}}(1-1/p) \ge1-\sum_{p\in\mathcal{P}}1/p > 1-2\delta.

Lemma 2 gives r≥HKδ≥Rr \ge H_K\delta\ge R.

Dense expansions avoiding a fixed set of primes

The only deep input is Proposition 5.2 of Liu–Sawhney [1]. For transparency, its hypotheses are recorded in the appendix. The following application also explains why fixed prime factors in the target denominator cause no problem.

Lemma 4. Fix a squarefree DD and a positive DD-integral rational rr. Put ρ=φ(D)/D\rho=\varphi(D)/D. For every 0<κ<min⁡(r/2,1/2)0<\kappa<\min(r/2,1/2), set c=exp⁡(−(r+κ)/ρ)c=\exp(-(r+\kappa)/\rho). For all sufficiently large NN, there is a set

FN⊆{n∈[cN,N]:gcd⁡(n,D)=1}F_N \subseteq\{n\in[cN,N]:\gcd(n,D)=1\}

with R(FN)=r\mathcal{R}(F_N)=r and

∣FN∣≥(ρ(1−c)−κ+o(1))N.(3)|F_N|\ge\bigl(\rho(1-c)-\kappa+o(1)\bigr)N. \tag*{(3)}

In particular,

lim inf⁡N→∞max⁡∣FN∣N≥ρ(1−e−r/ρ),\liminf_{N\to\infty}\frac{\max|F_N|}{N}\ge\rho(1-e^{-r/\rho}),

where the maximum allows all coprime denominators at most NN.

Proof. Let WN={n∈[cN,N]:gcd⁡(n,D)=1}W_N=\{n\in[cN,N]:\gcd(n,D)=1\}. Since D,cD,c are fixed, elementary counting in residue classes gives

∣WN∣=ρ(1−c)N+OD(1),R(WN)=ρlog⁡(1/c)+o(1)=r+κ+o(1).(4)|W_N|=\rho(1-c)N+O_D(1),\qquad\mathcal{R}(W_N)=\rho\log(1/c)+o(1)=r+\kappa+o(1). \tag*{(4)}

Set

M0=min⁡(c/2,10−5)N,u=(log⁡N)4/5,K0=M0e−u,S=Ne−6u,η=(log⁡log⁡N)−2.M_0=\min(c/2,10^{-5})N,\qquad u=(\log N)^{4/5},\qquad K_0=M_0e^{-u},\qquad S=Ne^{-6u},\qquad\eta=(\log\log N)^{-2}.

The endpoint can be rounded to an integer. Remove from WNW_N all numbers that are not SS-smooth and all numbers with Ω(n)>5log⁡log⁡N\Omega(n)>5\log\log N. Lemmas 2.2–2.3 of [1] show that this removes O(N/(log⁡N)1/5)+O(N/(log⁡N)3)=o(N)O(N/(\log N)^{1/5})+O(N/(\log N)^3)=o(N) elements. Because these numbers are at least cNcN, their reciprocal mass is o(1)o(1).

Next, while a prime power qq occurring in the current set satisfies qR(Wq)<ηq\mathcal{R}(W_q)<\eta, remove all its multiples. Each prime power is used at most once, so the total reciprocal mass removed is at most

η∑q≤N1q=O(1/log⁡log⁡N)=o(1).\eta\sum_{q\le N}\frac{1}{q}=O(1/\log\log N)=o(1).

Call the final set WN∗W_N^*, and put Q=lcm⁡(WN∗)Q=\operatorname{lcm}(W_N^*). We have

R(WN∗)=r+κ+o(1),qR((WN∗)q)≥η(5)\mathcal{R}(W_N^*)=r+\kappa+o(1),\qquad q\mathcal{R}((W_N^*)_q)\ge\eta \tag*{(5)}

for every prime power occurring in WN∗W_N^*.

Write r=a/br=a/b in lowest terms. For each prime power q∣bq\mid b, gcd⁡(q,D)=1\gcd(q,D)=1. Again by residue-class counting,

R((WN)q)=ρqlog⁡(1/c)+o(1)>0.\mathcal{R}((W_N)_q)=\frac{\rho}{q}\log(1/c)+o(1)>0.

Here qq is fixed. Since the entire cleanup removes only o(1)o(1) reciprocal mass, some multiple of qq remains in WN∗W_N^*. Therefore b∣Qb\mid Q. In particular,

x=Q(R(WN∗)−r)x=Q\bigl(\mathcal{R}(W_N^*)-r\bigr)

is an integer, positive for large NN, and x/Q=κ+o(1)x/Q=\kappa+o(1).

Apply Proposition 5.2 with d=1/5d=1/5, e=1/100e=1/100, and the displayed M0,K0,S,ηM_0,K_0,S,\eta. All its size conditions hold for large NN. For the last condition, the second term in GG gives

G≫c(log⁡N)3/5(log⁡log⁡N)−9,G\gg_c(\log N)^{3/5}(\log\log N)^{-9},

and hence

G2(log⁡N)2e(log⁡(N/M0)+(log⁡N)4/5)≫c(log⁡N)19/50(log⁡log⁡N)18⟶∞.\frac{G^2}{(\log N)^{2e}(\log(N/M_0)+(\log N)^{4/5})} \gg_c\frac{(\log N)^{19/50}}{(\log\log N)^{18}}\longrightarrow\infty.

The bounds on SS hold because its factor e−6ue^{-6u} dominates the e−2ue^{-2u} and polylogarithmic factors in those bounds; S≥N.99S\ge N^{.99} eventually. The target condition follows from (5) and κ>0\kappa>0:

(1+1/log⁡N)x/Q≤R(WN∗)≤(log⁡N)x/Q.(1+1/\log N)x/Q\le\mathcal{R}(W_N^*)\le(\log N)x/Q.

The proposition supplies GN⊆WN∗G_N\subseteq W_N^* with reciprocal sum x/Q=R(WN∗)−rx/Q=\mathcal{R}(W_N^*)-r. Take FN=WN∗∖GNF_N=W_N^*\setminus G_N, so that R(FN)=r\mathcal{R}(F_N)=r. This applies the proposition to a small complementary target; its sampling proportion is less than $1/2forlargefor largeN. Finally every term in WN∖FNW_N\setminus F_N is at most NN, so

∣WN∖FN∣≤NR(WN∖FN)=N(κ+o(1)).|W_N\setminus F_N|\le N\mathcal{R}(W_N\setminus F_N)=N(\kappa+o(1)).

Together with (4) this proves (3). Taking the lower limit in NN and then letting κ↓0\kappa\downarrow0 gives the final assertion. No uniformity as D,r,κD,r,\kappa vary is required.

Primitivity and completion of the asymptotic

Proof of Theorem 1. Fix 0<δ<1/40<\delta<1/4 and use Lemma 3 with R=log⁡(1/δ)R=\log(1/\delta). This gives fixed D,E,rD,E,r and ρ≥1−2δ\rho\ge1-2\delta, r≥log⁡(1/δ)r\ge\log(1/\delta). By Lemma 4, choose FNF_N with R(FN)=r\mathcal{R}(F_N)=r and all denominators coprime to DD. For large NN, every element of the fixed set EE is at most NN. Also E∩FN=∅E\cap F_N=\varnothing, since every element of EE has a prime factor dividing DD.

Give EE positive signs and FNF_N negative signs. The total is zero. Suppose C⊆E∪FNC\subseteq E\cup F_N is a signed zero subset, and put E′=C∩EE'=C\cap E, F′=C∩FNF'=C\cap F_N. Then

R(E′)=R(F′).\mathcal{R}(E')=\mathcal{R}(F').

The right side is DD-integral. Lemma 2 therefore forces E′=∅E'=\varnothing or E′=EE'=E. In the first case positivity forces F′=∅F'=\varnothing. In the second, R(F′)=r=R(FN)\mathcal{R}(F')=r=\mathcal{R}(F_N) and positivity forces F′=FNF'=F_N. Thus the signed identity is primitive.

Its cardinality and Lemma 4 give

lim inf⁡N→∞f(N)N≥ρ(1−e−r/ρ)≥(1−2δ)(1−δ)≥1−3δ.\liminf_{N\to\infty}\frac{f(N)}{N} \ge\rho(1-e^{-r/\rho}) \ge(1-2\delta)(1-\delta) \ge1-3\delta.

The middle inequality uses ρ≤1\rho\le1 and r/ρ≥r≥log⁡(1/δ)r/\rho\ge r \ge\log(1/\delta). Let δ↓0\delta\downarrow0, and combine this with f(N)≤Nf(N) \le N. This proves the asserted limit and the quantified first statement.

For a small concrete example, K=2K=2 and P={5,7,11}\mathcal{P}=\{5,7,11\} give D=385D=385, t=73t=73, and the primitive identity

15+110+17+114+111+122+128105=95146=12+18+139+122776.\frac{1}{5}+\frac{1}{10}+\frac{1}{7}+\frac{1}{14}+\frac{1}{11}+\frac{1}{22}+\frac{1}{28105}=\frac{95}{146}=\frac{1}{2}+\frac{1}{8}+\frac{1}{39}+\frac{1}{22776}.

The four denominators on the right are coprime to 385. This illustrates minimality, not the asymptotic density estimate.

Verification and attribution

The protected-prime construction, the amplification of its reciprocal mass at small prime-density cost, and the final primitivity argument already appear in the public manuscript of g-01234, dated 16 July 2026 [4]. A later submission by peilinliu66-dev, opened on 3 October 2026, claims a complete Lean proof using a thinned-prime inverse-sum completion route [5]. That submission was pending when checked on 16 October 2026; we have not independently reproduced it. These earlier claims must be considered in any assessment of priority or prize eligibility. The present dense-expansion lemma instead applies Liu–Sawhney’s Proposition 5.2. No novelty claim is made for the protected-prime idea. The Fourier analysis is due to Liu–Sawhney and their predecessors. The attached exact audit checks small global CRT examples exhaustively and checks a larger example locally at every selected prime. These finite checks do not replace the proof of the asymptotic or certify the imported analytic theorem.

The accompanying file Erdos319.lean proves the limit and explicit fixed-sign witnesses without additional mathematical hypotheses. Its analytic estimates include ports of existing proofs from plby/lean-proofs, commit 8822f7ddef30fadbd92e1c6ab4ed897af356af5e; those background proofs retain their separate attribution. The checked environment is Lean 4.35.0-rc2 with Mathlib commit 065356127b1dc0016f66b7283ce0ce2c4055aa55. The main theorem is

Erdos319Formal.erdos_problem_319.

The original 1980 statement was checked at printed p. 43 [2]: the same signs are used in the whole sum and in every subsum. The conclusion determines the leading asymptotic, not the lower-order deficit or the exact maximum for each finite NN.

The listed team prepared this AI-assisted proof presentation and formalization for review. Team credit does not assert that each member manually wrote every proof step. No independent human review, prize acceptance, or entitlement to payment is asserted.

The precise Liu–Sawhney input

Here is Proposition 5.2 of [1], restated with renamed parameters. For W⊆[1,N]W \subseteq[1,N], let Wq={n∈W:q∣n}W_q=\{n\in W:q\mid n\} and let QW=lcm⁡(W)Q_W=\operatorname{lcm}(W). There is an absolute CC such that, for fixed d∈(0,1)d\in(0,1), ϵ∈(0,1/10)\epsilon\in(0,1/10) and sufficiently large NN, the following holds. Suppose η≥1/log⁡N\eta\ge1/\log N and

N.99≤S≤K0≤M0≤N/104,K0≤M0exp⁡(−(log⁡N)1−d),N^{.99}\le S\le K_0\le M_0\le N/10^4,\qquad K_0\le M_0\exp\left(-(\log N)^{1-d}\right),
S≤min⁡{M02CN,ηM0K02N2(log⁡N)3}.S \le\min\left\{\frac{M_0^2}{CN},\frac{\eta M_0K_0^2}{N^2(\log N)^3}\right\}.

Define

G=max⁡{η(log⁡N)d(log⁡log⁡N)3,η2(log⁡N)1−2dlog⁡(N/M0)2(log⁡log⁡N)5}.G=\max\left\{\frac{\eta}{(\log N)^d(\log\log N)^3},\frac{\eta^2(\log N)^{1-2d}}{\log(N/M_0)^2(\log\log N)^5}\right\}.

Assume

G2(log⁡N)2e(log⁡(N/M0)+(log⁡N)1−d)≥C.\frac{G^2}{(\log N)^{2e(\log(N/M_0)+(\log N)^{1-d})}}\ge C.

Let W⊆[M0,N]W\subseteq[M_0,N] satisfy: every prime power dividing an element of WW is at most SS; Ω(n)≤5log⁡log⁡N\Omega(n)\le5\log\log N for n∈Wn\in W; and qR(Wq)≥ηq\mathcal{R}(W_q)\ge\eta for every prime power occurring in WW. For every positive integer xx with

(1+1/log⁡N)x/QW≤R(W)≤(log⁡N)x/QW,(1+1/\log N)x/Q_W\le\mathcal{R}(W)\le(\log N)x/Q_W,

there is W′⊆WW'\subseteq W such that R(W′)=x/QW\mathcal{R}(W')=x/Q_W.

In the dense-expansion lemma all parameters other than NN are fixed or explicitly specified functions of NN. In particular, the proof does not use the inconsistent displayed sign of KK in the arXiv version’s proof of Theorem 1.3; it uses the proposition’s stated K0≤M0exp⁡(−(log⁡N)1−d)K_0\le M_0\exp(-(\log N)^{1-d}) condition directly.

References

  1. [1]Y. P. Liu and M. Sawhney, On further questions regarding unit fractions, arXiv:2404.07113v1 (2024), Proposition 5.2 and Lemmas 2.2–2.3. International Mathematics Research Notices (2026), rnaf382; published article.arxiv.org/abs/2404.07113
  2. [2]P. Erdős and R. L. Graham, Old and new problems and results in combinatorial number theory (1980), p. 43; original scan.
  3. [3]T. F. Bloom, Erdős Problem 319.
  4. [4]g-01234, A Candidate Proof That c(N)/N → 1 in Erdős Problem 319, circulation draft, 16 July 2026; pinned manuscript.
  5. [5]peilinliu66-dev, JSP-000265: complete Erdős 319 fixed-sign density-one Lean proof, submission opened 3 October 2026; awards PR 4669; selected proof revision.

Paper details

vlnc:W_jegm4pep
Version
v1 · latest
Published
Fresh — published within the last 30 days
License
CC BY 4.0
Usage

views · 11

downloads · 0

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

  • Mathematics

AI assistance

View statement
Looking for related papers…
Contents