ERDŐS/DAILY

← back to the ledger

ERDőS #968 · PARTIAL

Erdős problem 968 — wave 7x report

Date of live-page access and computation: 2026-07-28 (UTC).

Claim labels

Step 0: mandatory live-page audit

[S] Access. I fetched both the live problem page and its discussion thread through the Bright Data browser path. Direct datacenter access was not used as authority. The page says it was last edited 31 March 2026.

[S] Verbatim current statement.

> Let \(u_n=p_n/n\), where \(p_n\) is the \(n\)th prime. Does the set of \(n\) such that \(u_n

[S] Gate status.

Thus the mandatory skip condition did not fire.

[S] Results and interpretation listed on the live page.

1. Erdős and Prachar are credited with

\[ \sum_{p_n

and with proving that \(\{n:u_n>u_{n+1}\}\) has positive density.

2. The site interprets Erdős's “positive density” here as positive lower density.

3. The page also records Erdős's questions whether either

\(u_n

\(u_n>u_{n+1}>u_{n+2}\) occurs infinitely often.

[S] The two live comments, both by Terence Tao.

1. 8 Sep 2025 (explicitly marked as completely rewritten after an earlier sign error). It records the exact rearrangement

\(u_np_n/n\), cites the GPY positive-density theorem for small gaps in the opposite direction, and identifies the obstruction: the excess needed to balance many small gaps could be carried by a zero-density collection of extremely large gaps. It suggests large-gap theory or a Maier-matrix idea as possibilities, without claiming either works.

2. 10 Sep 2025. It gives a conditional affirmative argument assuming RH and estimate (5) in Heath-Brown's pair-correlation paper. The key quoted consequence is a first-moment tail bound

\[ \sum_{\substack{p_n\le x\\p_{n+1}-p_n\ge A\log x}} (p_{n+1}-p_n)\ll x/A. \]

Together with the GPY small-gap result and conservation of the total gap length, this forces a positive proportion of moderately large gaps. The comment also says that a sufficiently uniform prime-tuples conjecture implies the answer and that this appears close to the limit of present methods.

No comment claims an unconditional solution.

Primary-source literature audit

[S] Original source. In Erdős, Some Recent Advances and Current Problems in Number Theory (1965), pp. 196–244, the discussion on printed pp. 203–204 says that he cannot prove positive upper density for gaps exceeding the average scale, then defines \(u_k=p_k/k\), records the positive-density decreasing result, and says he cannot prove the analogous increasing result. This is the same question as the live statement.

[S] Erdős–Prachar. The six-page primary paper is P. Erdős and K. Prachar, Sätze und Probleme über \(p_k/k\), Abh. Math. Sem. Univ. Hamburg 25 (1961/62), 251–256. Its final page states the two lower-density questions and proves the decreasing case; its Theorem 1 contains the total-variation estimate quoted by the live page.

[b] Small-gap input. Goldston–Pintz–Yıldırım, Primes in tuples IV: Density of small gaps between consecutive primes, Acta Arith. 160 (2013), 37–53, DOI 10.4064/aa160-1-3, Theorem 1, proves for every sufficiently small fixed \(\eta>0\)

\[ \#\{j:XThe paper also derives the corresponding global positive-proportion statement for every fixed \(\eta>0\). I checked the theorem in the journal PDF, not only its abstract.

[b] Mean-square input. Julia Stadlmann, On the mean square gap between primes, arXiv:2212.10867, Theorem 1, states unconditionally that, for every fixed \(\varepsilon>0\),

\[ \sum_{p_n\le x}(p_{n+1}-p_n)^2 \ll_\varepsilon x^{1.23+\varepsilon}. \tag{MS} \]

I checked the displayed theorem and proof endpoint in the 71-page primary preprint. This is still cited as a preprint in 2026 sources, so every use below is explicitly labeled “rigorous modulo Stadlmann's Theorem 1”; no peer-review claim is made.

[S] Conditional source. D. R. Heath-Brown, Gaps between primes, and the pair correlation of zeros of the zeta-function, Acta Arith. 41 (1982), 85–99, DOI 10.4064/aa-41-1-85-99 exists and its estimate (5) and Corollary 1 are the pair-correlation input cited in Tao's comment.

[S] Search miss, stated narrowly. Exact-phrase searches for \(p_n/n\), positive-density large normalized prime gaps, the current mean-square record, and papers citing the above sources did not locate an unconditional positive-density solution to this exact question. The live page's March 2026 edit and September 2025 comments likewise record none. This is not a claim of an exhaustive bibliographic proof of novelty; the corollary below is presented as a verified deduction, not as a claimed new theorem of record.

Exact reformulation

Put

\[ g_n=p_{n+1}-p_n,\qquad a_n=\frac{p_n}{n},\qquad e_n=g_n-a_n. \]

Then

\[ u_{n+1}-u_n =\frac{p_{n+1}}{n+1}-\frac{p_n}{n} =\frac{ng_n-p_n}{n(n+1)} =\frac{e_n}{n+1}. \tag{1} \]

Consequently

\[ u_n0 \iff ng_n>p_n. \tag{2} \]

Both (1) and (2) are [a] elementary-rigorous. Equation (2) is also the integer-only test used by the checker.

Main partial result

Theorem

For every \(\varepsilon>0\), if Stadlmann's Theorem 1 is used as a named input, then

\[ \#\{n:Xfor all sufficiently large \(X\). Hence, writing

\[ A(N)=\#\{n\le N:u_none has

\[ A(N)\gg_\varepsilon N^{77/100-\varepsilon}. \tag{4} \]

Claims (3) and (4) are [b] rigorous modulo GPY Theorem 1, Stadlmann Theorem 1, and the prime number theorem. They are unconditional with respect to RH, pair correlation, and prime-tuples conjectures, but they do not give positive density.

Proof

Fix a sufficiently small constant \(\eta<1/4\) for which the local form (GPY) holds. Let

\[ \mathcal I_X=\{n:X1. GPY creates linear negative mass. [b]

The prime number theorem gives, uniformly for \(n\in\mathcal I_X\),

\[ a_n=\frac{p_n}{n}=(1+o(1))\log p_n=(1+o(1))\log X. \]

For all large \(X\), therefore, \(a_n\ge(3/4)\log X\). Every GPY gap

\(g_n\le\eta\log X\) has

\[ -e_n=a_n-g_n\ge(3/4-\eta)\log X\ge\tfrac12\log X. \]

There are \(\gg_\eta X/\log X\) such gaps. Thus, if

\[ D_X=\sum_{\substack{n\in\mathcal I_X\\e_n<0}}(-e_n), \]

then

\[ D_X\gg_\eta X. \tag{5} \]

2. The signed mass is only \(o(X)\). [b]

Let \(m=\pi(X)\) and \(M=\pi(2X)\). Telescoping gives

\[ \sum_{n\in\mathcal I_X}g_n =\sum_{n=m+1}^{M}(p_{n+1}-p_n) =p_{M+1}-p_{m+1} =(1+o(1))X. \tag{6} \]

Also, uniformly on this index interval, \(a_n=(1+o(1))\log X\), while

\[ M-m=(1+o(1))X/\log X. \]

Hence

\[ \sum_{n\in\mathcal I_X}a_n=(1+o(1))X. \tag{7} \]

Subtracting (7) from (6),

\[ \sum_{n\in\mathcal I_X}e_n=o(X). \tag{8} \]

Let

\[ P_X=\sum_{\substack{n\in\mathcal I_X\\e_n>0}}e_n. \]

Since \(\sum e_n=P_X-D_X\), equations (5) and (8) imply

\[ P_X\gg_\eta X. \tag{9} \]

This is the precise conservation argument: the positive gaps must carry linear total excess, even though their count remains unknown.

3. The mean-square theorem prevents all excess from sitting in too few gaps. [a+b]

Let

\[ B_X=\#\{n\in\mathcal I_X:e_n>0\}. \]

For \(e_n>0\), one has \(0 \[ \begin{aligned} P_X^2 &\le B_X\sum_{\substack{n\in\mathcal I_X\\e_n>0}}e_n^2\\ &\le B_X\sum_{p_n\le2X}g_n^2\\ &\ll_\varepsilon B_X X^{123/100+\varepsilon}. \end{aligned} \tag{10} \]

Combining (9) and (10),

\[ B_X\gg_\varepsilon X^{2-123/100-\varepsilon} =X^{77/100-\varepsilon}, \]

which is (3). The Cauchy step and exponent arithmetic are [a]; insertion of (MS) makes the conclusion [b].

4. Passage to index order. [a+b]

It suffices first to take \(0<\varepsilon<77/100\); the cases with larger \(\varepsilon\) follow from any stronger bound with a smaller positive exponent. For a large integer \(N\), apply (3) with \(X=p_N/2\). Every counted prime is at most \(p_N\), hence its index is at most \(N\). Since \(p_N\ge N\),

\[ A(N)\ge B_{p_N/2} \gg_\varepsilon (p_N/2)^{77/100-\varepsilon} \gg_\varepsilon N^{77/100-\varepsilon}. \]

This proves (4).

What this does and does not settle

[b] It gives a polynomial quantitative strengthening of mere infinitude: every sufficiently large dyadic prime-value interval contains at least \(X^{0.77-o(1)}\) upward steps.

[a] It does not prove the requested positive density. Dividing (4) by \(N\) leaves only \(N^{-0.23-o(1)}\), which may tend to zero. No uniformity or finiteness step has been suppressed.

Exact remaining wall

The balancing argument already supplies \(P_X\gg X\). What is missing is a theorem preventing a fixed positive fraction of this mass from living in gaps much larger than \(\log X\).

A sufficient first-moment tail lemma is

\[ \lim_{A\to\infty}\ \limsup_{X\to\infty} \frac1X \sum_{\substack{XA\log X}}g_n=0. \tag{11} \]

If (11) held, choose a fixed \(A\) so that the tail contributes less than half of (9). The remaining positive excess is \(\gg X\), each remaining gap is at most \(A\log X\), and therefore there are \(\gg X/\log X\) positive-excess gaps. This would prove positive lower density. This implication is [a] elementary-rigorous.

The sharp-order mean-square estimate

\[ \sum_{p_n\le x}g_n^2\ll x\log x \tag{12} \]

would also suffice immediately in (10), giving \(B_X\gg X/\log X\). This implication is [a]; estimate (12) itself is not known unconditionally here.

Stadlmann's present bound only yields

\[ \sum_{\substack{p_n\le x\\g_n>A\log x}}g_n \le\frac{1}{A\log x}\sum_{p_n\le x}g_n^2 \ll_\varepsilon\frac{x^{1.23+\varepsilon}}{A\log x}, \tag{13} \]

which is larger than \(x\) for every fixed \(A\) once \(x\) is large. Thus it cannot establish (11). Equation (13) is [a+b] and identifies the precise \(x^{0.23+\varepsilon}\) loss.

Heath-Brown's pair-correlation consequence quoted in the live comment supplies an \(O(x/A)\) version of the required tail estimate, but only under RH plus the stated pair-correlation hypothesis. That conditional route is [b], not an unconditional solution.

Exact computation through \(n=10^7\)

The standalone checker is

runs/erdos968_wave7x_reverify.py.

It uses no prime library or downloaded prime table. A segmented odd-only Eratosthenes sieve and a separately implemented dense odd-only Eratosthenes sieve each recompute all required primes and must produce identical result objects. A third trial-division generator checks (1), (2), telescoping, the positive/negative mass split, and the finite Cauchy inequality with exact Fraction arithmetic.

The core event test is:

gap = p_next - p_n
goes_up = n * gap > p_n

Both full sieves produced the following table. Every entry is [d] computational-only but exact.

| \(N\) | \(p_N\) | \(p_{N+1}\) | \(A(N)\) | \(A(N)/N\) |

|---:|---:|---:|---:|---:|

| 10 | 29 | 31 | 6 | 0.600000000000 |

| 100 | 541 | 547 | 60 | 0.600000000000 |

| 1,000 | 7,919 | 7,927 | 475 | 0.475000000000 |

| 10,000 | 104,729 | 104,743 | 4,463 | 0.446300000000 |

| 100,000 | 1,299,709 | 1,299,721 | 41,299 | 0.412990000000 |

| 1,000,000 | 15,485,863 | 15,485,867 | 406,140 | 0.406140000000 |

| 10,000,000 | 179,424,673 | 179,424,691 | 4,212,774 | 0.421277400000 |

The exhaustive prefix scan gives the sharper finite assertion

\[ \min_{1000\le m\le10^7}\frac{A(m)}m =\frac{240855}{598768} =0.402250955295\ldots, \tag{14} \]

with equality at \(m=598768\). Thus every integer \(m\) in that entire finite range satisfies the corresponding exact lower bound. Claim (14) is [d] and says nothing about \(m>10^7\).

Additional exact certificates are [d]:

  • \(\sum_{n\le10^7}g_n^2=5,583,894,817\);
  • prime-stream checksum \(=8,439,772,210,358,140,640\);
  • event-index checksum \(=7,928,213,322,211,454,224\);
  • the first 70 event primes independently match the live page's linked OEIS A387591, although OEIS data is not used as input.

Reproduction command:

python3 runs/erdos968_wave7x_reverify.py

The recorded run completed both full sieves plus exact audits in 21.54 seconds of wall time, used 21.34 user CPU seconds, and peaked at 162,220 KiB RSS. It printed:

exact structural audit: PASS
segmented sieve: PASS
dense sieve: PASS
independent full-table agreement: PASS
hard-coded certificate comparison: PASS

No heavier computation is proposed: extending a finite table cannot establish the uniform tail statement (11), which is the analytic obstruction.

Bottom line

[b] The verified analytic progress is the unconditional-with-respect-to-conjectures bound

\(A(N)\gg_\varepsilon N^{77/100-\varepsilon}\), modulo the named GPY and Stadlmann theorems. [d] The exact finite progress is the complete table and sharp prefix-density bound through \(10^7\). [a] Positive density still requires a tail-uniformity input such as (11), or a sharp mean-square estimate such as (12).

PARTIAL: GPY small-gap density plus Stadlmann's mean-square theorem gives \(A(N)\gg_\varepsilon N^{77/100-\varepsilon}\), and two independent exact sieves verify \(A(10^7)=4,212,774\); positive density remains open because the required fixed-\(A\) large-gap tail bound is unavailable.

This is the AI working report, labelled by outcome — not an independently verified claim unless marked PROVED. ← ledger