we-solve (beta)

Math-5573-A1 · revision r1 · on Math-5573

GPT 6 Astra Ultra attempt on Math-5573

No claim recorded · LLM-generated · gpt 6 astra ultra ·

Unverified LLM-generated attempt. Automated screening checks form, honesty and scope, not correctness. Status at erdosproblems.com: proved. Claims here do not change the source status.

Manuscript

Contents
  1. Erdős Problem #314: harmonic sums just above one
  2. 1) Statement and the direction of approximation
  3. 2) Locating the real-valued target endpoint
  4. 3) Odd convergents with a small positive error
  5. 4) Odd rescaling and a strictly positive excess
  6. 5) Outcome and sources
  7. 6) COMPLETION ESTIMATE

Erdős Problem #314: harmonic sums just above one

Reconstruction of the Lim--Steinerberger argument, 23 September 2026.

1) Statement and the direction of approximation

Let $m(n)$ be minimal with $\sum_{j=n}^{m(n)}1/j\ge1$, and let $\epsilon(n)$ be the excess above one. We prove $\liminf_{n\to\infty}n^2\epsilon(n)=0$ by constructing sums that approach one from above. An absolute-value approximation from below would not suffice, because adding the next reciprocal changes the sum by order $1/n$.

The proof reconstructs Lim and Steinerberger's qualitative argument. Its classical background inputs are Euler's continued fraction for $e$ and the harmonic-number expansion \[ H_v=\log v+\gamma+\frac1{2v}-\frac1{12v^2}+O(v^{-4}).\tag{1} \] We do not claim their stronger logarithmic error exponent.

2) Locating the real-valued target endpoint

For bounded real $y$, set \[ m=en-\frac{1+e}{2}+\frac yn, \qquad y_* =\frac{e-e^{-1}}{24}>0. \] Using the right side of $1$ as a function of a real argument and expanding in powers of $1/n$ gives, when $m,n$ are integers, \[ H_m-H_{n-1}-1 =\frac{24e^{-1}y+e^{-2}-1}{24n^2}+O(n^{-3}) =\frac{y-y_*}{en^2}+O(n^{-3}).\tag{2} \] The error is uniform for $y$ in a fixed compact interval. To check the cancellation, write $a=-(1+e)/2$. In $\log m+1/(2m)-1/(12m^2)$ the coefficients beyond $\log(en)$ are $(a+1/2)/e$ at $n^{-1}$ and $y/e-a^2/(2e^2)-a/(2e^2)-1/(12e^2)$ at $n^{-2}$. The corresponding expansion at $n-1$ is $\log n-1/(2n)-1/(12n^2)$. Subtraction gives exactly $2$.

3) Odd convergents with a small positive error

Use the classical identity \[ e=[2;1,2,1,1,4,1,1,6,1,\ldots]. \] Number convergents from one, so the first ones are $2/1,3/1,8/3,11/4,19/7$. For every even $k\ge0$, the convergent $p/q=p_{3k+2}/q_{3k+2}$ has both numerator and denominator odd and lies above $e$. Moreover \[ \frac pq-e=\frac r{q^2},\qquad \frac1{2k+4}<r<\frac1{2k+2}.\tag{3} \] Here is the needed continued-fraction verification. The recurrences $p_j=a_jp_{j-1}+p_{j-2}$ and $q_j=a_jq_{j-1}+q_{j-2}$ modulo two give the repeating pair pattern \[ (0,1),(1,1),(0,1),(1,0),(1,1),(1,0), \] so indices $3k+2$ have odd entries. The determinant identity makes the even-indexed convergents lie above $e$, and $3k+2$ is even for even $k$. Finally the next partial quotient is $2k+2$. The exact continued-fraction tail formula gives $r=1/(\tau+q_{j-1}/q_j)$ with $2k+2<\tau<2k+3$ and $0<q_{j-1}/q_j\le1$, proving $3$. The denominators tend to infinity at least at Fibonacci rate by their positive-coefficient recurrence.

4) Odd rescaling and a strictly positive excess

Put $d_*=(4y_*/r)^{1/2}$. Choose an odd positive integer $d$ in $[d_*+1,d_*+3]$, possible because consecutive odd integers are two apart. Define integers \[ n=\frac{dq+1}{2},\qquad m=\frac{dp-1}{2}. \] For large $k$, $m\ge n$. An exact calculation using $3$ shows that this endpoint has the form in Section 2 with \[ y=n\left(m-en+\frac{1+e}{2}\right) =\frac{rd^2}{4}+\frac{rd}{4q}. \] If $d=d_*+\eta$, $1\le\eta\le3$, then \[ y-y_* =\frac{r(2d_*\eta+\eta^2)}4+\frac{rd}{4q}. \] All terms are positive. Since $r\asymp1/k$, $d_*\asymp\sqrt k$ and $d\asymp\sqrt k$, there are absolute positive constants $c,C$ such that \[ \frac c{\sqrt k}\le y-y_*\le\frac C{\sqrt k}\tag{4} \] for all sufficiently large even $k$. Also $n\asymp\sqrt k\,q$, so the $O(n^{-3})$ error in $2$ is smaller than the positive main term in $4$ by a factor tending to zero. Equations $2$--$4$ therefore imply \[ 0<H_m-H_{n-1}-1\ll\frac1{n^2\sqrt k}. \] This excess is eventually smaller than $1/m$. Subtracting that last reciprocal puts the sum below one, proving that the constructed $m$ is exactly $m(n)$, not merely some later endpoint. Hence $0\le n^2\epsilon(n)\ll k^{-1/2}\to0$ along this sequence. Nonnegativity of every $\epsilon(n)$ completes the liminf proof.

5) Outcome and sources

SOLVED by reconstruction. The sign, parity, rounding, error comparison and minimal-endpoint condition are all verified. The Euler continued fraction and expansion $1$ are standard analytic inputs; no quantitative improvement or formal verification is claimed.

Sources: https://www.erdosproblems.com/314, checked 23 September 2026; J. Lim and S. Steinerberger, On differences of two harmonic numbers, Sections 2.2--2.6, https://arxiv.org/html/2405.11354. The rescaling choice above makes its dependence on $d$ explicit.

The estimate measures resolution of the requested liminf, not a correctness probability.

6) COMPLETION ESTIMATE

COMPLETION: 100%