---

Subject and independence

Role: independent reviewer in a fresh context, commissioned for refutation and given only the assignment. The reviewer took no part in writing the page or any page of its folder and had not read the page or the source before this commission. No computation was used; the review is a reading of the page against the artifact with every step re-derived.

Subject: path wiki/research/erdos_354/yu_chen_bg_reconstruction.md as it stood at 2026-09-28T05:03:27Z (the page), read in full as of that time.

Artifact: the seventeen-page PDF (138,329 bytes) held by the library card Yu and Chen (2026). Physical pp. 13--14 (printed 13--14: Section 11 "The bounded-spacing contradiction (BG)", Subsections 11.1--11.3, display (11.1)) were read in full, sentence by sentence, in the text layer and on page images rendered at 130 dpi; every displayed formula on those pages was checked on the images. Physical pp. 11--12 (Section 10, for the statement of (10.2) and the definitions of good rationals, $H$ and $\delta_i$) were read in full in the text layer and on page images. Physical pp. 2--3 (Section 1, for the conventions on layers, conversions, events, $K_n$ and $\mathbb N$) were read in the text layer, and physical p. 7 (Sections 6--7, for the $R=4$ spacing bound) in the text layer and on a page image. Page images rendered: pp. 7, 11, 12, 13, 14. The physical and printed page numbers agree throughout.

Allowed material actually read: the frozen page; the normalization page and the windows page of the same folder as of the same time, read in full (their Definitions and Statement sections were needed; their proofs were read at the same time but no verdict below rests on them); the theorem page as of the same time through its Statement section; the provenance paragraph of the library card; the Statement paragraph of the problem page wiki/problems/additive_bases/E0354/_index.md; the Erdos-specific sections "Whole-claim report" and "Audit checklist" of docs/verification.md, together with the shared "Audit checklist" section; "Source fidelity" of docs/evidence.md; and docs/math_authoring.md in full.

Exposures, disclosed: (1) printing the head of the problem page showed, past its Statement and Formulation paragraphs, the frontmatter desc and the opening lines of its Status paragraph, which name a site-accepted proof; (2) printing the head of the library card showed, beyond the provenance paragraph, the card's desc, its theorem link row, its Read status paragraph and the start of its Overview; (3) the theorem page's Source and Standing paragraphs precede its Statement and were seen, including a standing sentence about a site-accepted proof; (4) directory listings of the research folder and of its evidence/ folder showed file names only. None of this was used. No evidence/ file, folder _index.md, Current assessment, Known results, other review or web search was read.

Restatement

Setting, inherited from the normalization page and the windows page. A normalized pair $\alpha,\beta>0$ has $N=\lfloor\beta\rfloor<M=\lfloor\alpha\rfloor<2N$ with $N\ge2$; $\theta=\alpha/\beta$ is irrational, so $1<\theta<2$. Layers are the indices $i\ge0$, with weights $a_i=\lfloor2^i\alpha\rfloor$, $b_i=\lfloor2^i\beta\rfloor$ and conversions $u_i=a_{i+1}-2a_i$, $v_i=b_{i+1}-2b_i$, both in $\{0,1\}$. An event is a position $t\ge1$ with $(u_{t-1},v_{t-1})\ne(0,0)$, so an event at $t$ records the conversion at index $t-1$; $K_n$ counts the events in $[1,n]$. The conventions are $0\in\mathbb N$ and natural $\log$. "Incomplete" is taken in the sense of (10.2): the set of weights $\{a_i,b_i:i\ge0\}$ is not complete, that is, infinitely many integers are not sums of distinct weights.

Bounded event spacing: there are an integer $R\ge2$ and a threshold $n_0$ such that every integer $n\ge n_0$ has an event in $(n,Rn]$.

Claim (BG): for a normalized pair with irrational $\theta$, incompleteness and bounded event spacing cannot both hold. Equivalently, under incompleteness, for every integer $R\ge2$ and every threshold $n_0$ there is an integer $n\ge n_0$ with no event in $(n,Rn]$.

Premise (10.2), as the windows page states it: there is a constant $L>0$, depending on $\alpha,\beta$ only, such that for every $T_0\ge1$ and every $\varepsilon>0$ there are an integer $T\ge T_0$ and a reduced rational $p/q$ with $1<p/q<2$ and $|p/q-\theta|<\varepsilon$ whose binary height $H=\lceil\log_2(p+q+1)\rceil$ satisfies $H^2\le4T$, with $K_T\le L\log T$ and $|\delta_i|<2^H$ for $0\le i\le T$, where $\delta_i=qa_i-pb_i$.

Order of choices in the proof, none of which depends on a later one: $R$ and $n_0$ from the hypothesis; $B=2R+1$; $L$ from (10.2); an integer $k\ge\max(1,8L\log B)$; a radius $\varepsilon_k>0$ with $(\theta-\varepsilon_k,\theta+\varepsilon_k)\cap\mathcal C_k=\emptyset$; a lower bound $T_0$ determined by $n_0$, $L$ and $B$; then one window $(T,p/q)$ from (10.2) with this $T_0$ and $\varepsilon_k$.

Checklist

• Quantifiers and scope: pass. The hypothesis quantifies over every integer $n\ge n_0$ and is applied only at the integers $z_i\ge n_0+1$. The window quantifiers are used in the order listed above; $L$ precedes $T_0$ and $\varepsilon$ in (10.2), as the proof needs for fixing $k$. Boundary cases checked: $K_T=0$ gives $r=1$ and the argument still produces two exact layers and one return; $H>T$ makes (11.1) trivial; the last interval $i=r$ is covered by $2B^rx\le T$. The one boundary point not spelled out is the integrality of $x$ (F3), which does not affect the conclusion.
• Circularity: pass. Neither (BG) nor an equivalent is assumed; the contradiction is between the event count $K_T$ of one window and the lower bound $rk>K_T$ derived from the hypotheses.
• Model and convention changes: pass. The event and position conventions match the source (p. 2: the conversion at index $i$ produces the event at position $i+1$) and the page's Step 1 translates between them correctly; the objects counted are the actual events of the pair, not a relaxed system.
• Finite and statistical overreach: inapplicable. No finite verification, averaging or heuristic is used anywhere in the argument.
• Uniformity: pass. $\mathcal C_k$ depends on $k$ alone, so the radius $\varepsilon_k$ depends on $k$ and $\theta$ only; the "more than $k$ events" conclusion is uniform over the exact layers $a<b$ and the multiplier $c$, as the source and the page both say; $L$ is a constant of the pair. The thresholds on $T$ are finitely many fixed conditions.
• Extremal conclusions: inapplicable. No infimum, supremum or sharpness claim is made.
• Consequences and composition: pass, with notes. Each "hence" was re-derived (Weakest steps). The page consumes (10.2) and the digit property $u_i,v_i\in\{0,1\}$ from sibling pages at their author-recorded standing (F5, F6); the Scope paragraph under-reports where irrationality is used (F2).
• Computation: inapplicable. The page has no computation.
• Reproduction: inapplicable. The page states no rerun command or coverage claim.
• Source and verdict fidelity: pass, with one suggested correction. The statement, definitions and all three subsections match pp. 13--14 clause by clause; the locator "physical pp. 13--14" and the label (11.1) are right; the Standing paragraph claims author-recorded status only; the step headings carry the subsection numbers in a form that reads as display labels (F1).

Weakest steps

1. Density of exact layers, display (11.1). Fix the window $(T,p/q,H)$. For $0\le i\le T-H$, if $(u_j,v_j)=(0,0)$ for $j=i,\dots,i+H-1$, iterating $a_{j+1}=2a_j+u_j$ and $b_{j+1}=2b_j+v_j$ gives $a_{i+H}=2^Ha_i$ and $b_{i+H}=2^Hb_i$, so $\delta_{i+H}=2^H\delta_i$. Since $i+H\le T$, (10.2) gives $|\delta_{i+H}|<2^H$, so $|\delta_i|<1$ and the integer $\delta_i$ is $0$. Contrapositive: a nonexact layer $i\le T-H$ has some $j\in[i,i+H-1]$ with $(u_j,v_j)\ne(0,0)$, that is, an event at $t=j+1\in[i+1,i+H]\subseteq[1,T]$. Assign each such $i$ one such $t$. The layers assigned to a given $t$ lie in $\{t-H,\dots,t-1\}$, at most $H$ of them, and there are $K_T$ event positions in $[1,T]$; so at most $HK_T$ nonexact layers lie in $[0,T-H]$, and the remaining layers $T-H+1,\dots,T$ number $H$. Hence $\#\{0\le i\le T:\delta_i\ne0\}\le HK_T+H=E$, and by (10.2) $E=H(K_T+1)\le2\sqrt T\,(L\log T+1)$. Composition: $E$ enters only through $x=\max(n_0+1,E+1)$ and the pigeonhole of Step 3, where the bound makes $E+1\le T^{3/4}$ for large $T$.

2. Cost of a nontrivial return, Subsection 11.2. Let $a<b$ be exact for $p/q$ with an event in $(a,b]$ and suppose $(a,b]$ has at most $k$ events. Put $U=\sum_{j=a}^{b-1}2^{b-1-j}u_j$ and $V$ likewise with $v_j$; unrolling the recurrences, $a_b=2^{b-a}a_a+U$ and $b_b=2^{b-a}b_a+V$, so $\delta_b=2^{b-a}\delta_a+qU-pV$ and exactness at both ends gives $qU=pV$. An event at $t\in(a,b]$ is a nonzero $(u_{t-1},v_{t-1})$ with $t-1\in[a,b-1]$, so $(U,V)\ne(0,0)$; as $p,q>0$, $qU=pV$ forces $U,V>0$ and $U/V=p/q\in(1,2)$. Because $u_j,v_j\in\{0,1\}$, the binary ones of $U$ and of $V$ sit at the indices of the nonzero conversions, so each has at most $k$ ones. Since $U>V>0$, $J=\lfloor\log_2U\rfloor\ge\lfloor\log_2V\rfloor$ is the exponent of the largest power of two in either word; $U/2^J\in[1,2)$ is a sum of at most $k$ distinct terms $2^{-m}$ with $m\ge0$, each in $\mathcal A$ (here $1=2^{-0}\in\mathcal A$ needs $0\in\mathbb N$), padded with $0\in\mathcal A$ to $k$ summands, so $U/2^J\in\mathcal S_k$; likewise $V/2^J\in\mathcal S_k$, and $V/2^J=(q/p)(U/2^J)>1/2$. Hence $p/q\in\mathcal C_k$. Now $\mathcal A\subset[0,1]$ is closed (its only limit point $0$ belongs to it) and bounded; $\mathcal S_k$ is the image of $\mathcal A^k$ under the continuous addition map; $\mathcal C_k$ is the image of the compact set $\mathcal S_k\times(\mathcal S_k\cap[1/2,k])$ under the continuous map $(x,y)\mapsto x/y$; so $\mathcal C_k$ is compact and consists of rationals, and the irrational $\theta$ has positive distance $\varepsilon_k$ from it. Contrapositive: if $|p/q-\theta|<\varepsilon_k$, every pair of exact layers with an event between them has more than $k$ events between them. Composition: $\varepsilon_k$ is the tolerance handed to (10.2); nothing about $a$, $b$, $T$ or the multiplier $c$ enters $\varepsilon_k$.

3. Geometric capacity, Subsection 11.3. With $K=K_T$, $x=\max(n_0+1,E+1)$ and $r=\lfloor K/k\rfloor+1$: $B^r\le B^{K/k+1}$, and $K/k\le L\log T/(8L\log B)=\log T/(8\log B)$ gives $B^{K/k}\le T^{1/8}$, so $B^r\le BT^{1/8}$. Once $T^{3/4}\ge\max(n_0+1,\,2\sqrt T(L\log T+1)+1)$ we have $x\le T^{3/4}$, and once also $T\ge(2B)^8$, $2B^rx\le2BT^{7/8}\le T$. For $0\le i\le r$ the interval $[B^ix,2B^ix]$ lies in $[1,T]$ and holds $B^ix+1\ge E+2>E$ integers (for integer $x$; at least $\lfloor B^ix\rfloor\ge E+1$ integers otherwise), so by (11.1) it contains an exact layer $z_i$. Then $z_{i+1}\ge B^{i+1}x=(2R+1)B^ix>2RB^ix\ge Rz_i$ and $z_i\ge x\ge n_0+1$, so bounded spacing places an event in $(z_i,Rz_i]\subseteq(z_i,z_{i+1}]$. The $r$ intervals $(z_i,z_{i+1}]$, $0\le i<r$, are pairwise disjoint subsets of $[1,T]$, each with more than $k$ events by step 2, so $K\ge r(k+1)>rk=(\lfloor K/k\rfloor+1)k>K$. Composition: this is the contradiction that proves (BG); the only inputs are (11.1), step 2 at the window's $p/q$, and the spacing hypothesis at the integers $z_i$.

Strongest attack

The strongest attack aimed at the uniformity of step 2, which is the place where a hidden dependence on the window would break the proof. The attempt: make the neighborhood of $\theta$ that step 2 needs shrink with the window, so that no single $\varepsilon$ could be handed to (10.2) before $T$ is chosen. Concretely, one tries to build, for a fixed $k$ and $p/q$ arbitrarily close to $\theta$, a return $(a,b]$ with at most $k$ events whose words escape $\mathcal S_k\times[1/2,\infty)$ after normalization: put the $k$ ones of $V$ far below those of $U$, so that $V/2^J<1/2$ and the ratio $p/q$ is not certified to lie in $\mathcal C_k$. The attack fails because $qU=pV$ with $1<p/q<2$ forces $V=(q/p)U>U/2\ge2^{J-1}$, so $V/2^J>1/2$ whatever the digit pattern; the membership $p/q\in\mathcal C_k$ then depends on the reduced ratio alone, and $\mathcal C_k$ depends on $k$ alone, so the radius $\varepsilon_k=\operatorname{dist}(\theta,\mathcal C_k)$ is fixed before the window. A variant, placing the top digit in $V$ rather than $U$, fails for the same reason: $U>V$ forces the largest power of two into $U$, so the word that is at least $1$ is the numerator and the denominator exceeds $1/2$. A second attack on the count (11.1), pushing nonexact layers into the last $H$ positions where no event inside $[1,T]$ need witness them, is absorbed by the separate term $H$ in $E$. A third attack on the capacity step, trying to make the last interval $[B^rx,2B^rx]$ leave $[0,T]$ or to make consecutive returns overlap, fails because $2B^rx\le2BT^{7/8}\le T$ for $T\ge(2B)^8$ and because $z_{i+1}>Rz_i\ge2z_i$. No defect was found.

Premises

• (10.2), from the windows page of the same folder, an author-recorded reconstruction; the source statement on physical p. 12 was read in full and agrees with the windows page's interface: there is a constant $L>0$ such that for every $T_0\ge1$ and $\varepsilon>0$ there are an integer $T\ge T_0$ and a reduced $p/q\in(1,2)$ with $|p/q-\theta|<\varepsilon$, $H^2\le4T$, $K_T\le L\log T$ and $|\delta_i|<2^H$ for $0\le i\le T$. The page uses every item of this interface and nothing beyond it; in particular it needs $L$ fixed before $T_0$ and $\varepsilon$, which the windows page's statement provides. Its proof was not verified here.
• Definitions and the digit property, from the normalization page, author-recorded: layers, weights, conversions $u_i,v_i\in\{0,1\}$ (item 4), the event set, $K_n$ and $0\in\mathbb N$. Item 4 is used in step 2 to read the binary digits of $U$ and $V$ as conversions; the source states it on physical p. 2 ("where $u_i,v_i\in\{0,1\}$"). The page cites the normalization page for the objects but not item 4 by name (F5).
• Bounded event spacing is the hypothesis being refuted, taken for a general integer $R\ge2$; the page's Scope paragraph attributes $R=4$ to the source's Section 6 (physical p. 7, read: "every integer $n\ge n_0$ has an arrival event in $(n,4n]$"). The proof of (BG) does not consume $R=4$.
• No external theorem is imported by the page directly; Dirichlet's theorem enters only inside the windows page. The page's own standing sentence names the page author-recorded, which is all it is entitled to.
• Explicit assumptions: the pair is normalized; $\theta$ is irrational (used in (10.2) and directly in step 2); the sequence is incomplete (used only through (10.2)); bounded spacing (assumed for contradiction).

Findings

F1. Severity: suggested. Location: the headings "Step 2: nontrivial returns cost many events (11.2)" and "Step 3: geometric capacity (11.3)". Defect: the parenthesized numbers have the form of display labels, but Section 11 of the source has one numbered display, (11.1) on physical p. 13; "11.2" and "11.3" are the subsection numbers printed on pp. 13 and 14 ("11.2 A uniform lower bound on nontrivial return cost", "11.3 Geometric capacity"). The page's own Source paragraph separates "Subsections 11.1--11.3" from "display (11.1)", so a reader of the headings looks for displays that do not exist. Witness: physical pp. 13--14. Proposed replacement: "Step 1: exact layers are dense (Subsection 11.1)", "Step 2: nontrivial returns cost many events (Subsection 11.2)", "Step 3: geometric capacity (Subsection 11.3)", keeping the tag (11.1) on the display.

F2. Severity: suggested. Location: Scope, "The argument uses the windows of (10.2), so it needs incompleteness and irrationality; bounded spacing enters only through". Defect: the sentence accounts for irrationality only through (10.2), but Step 2 uses it directly ("The irrational $\theta$ is not in the closed set $\mathcal C_k$"), and that use is the one that makes the return cost uniform. Witness: the page's Step 2, and physical p. 13, "A fixed irrational $\theta$ consequently has a neighbourhood disjoint from $\mathcal C_k$" (the source's spelling). Proposed replacement: "The argument uses the windows of (10.2), so it needs incompleteness and irrationality; irrationality is used again directly in Step 2, where $\theta\notin\mathcal C_k$ gives the uniform return cost; bounded spacing enters only through the choice of $x>n_0$ and the events in $(z_i,Rz_i]$."

F3. Severity: note. Location: Definitions, "an integer $R\ge2$ and a threshold $n_0$", and Step 3, "contains $B^ix+1\ge x+1>E$ layers". Defect: the threshold is not declared an integer, and the exact count $B^ix+1$ of integers in $[B^ix,2B^ix]$ presumes that $x=\max(n_0+1,E+1)$ is an integer. The conclusion survives either way, since an interval $[y,2y]$ with $y\ge E+1$ holds at least $\lfloor y\rfloor\ge E+1$ integers, and the hypothesis quantifies over integers $n$, so a real threshold may be replaced by its ceiling. Witness: physical p. 14 says only "contains more than $E$ layers". Proposed replacement: "an integer $R\ge2$ and an integer threshold $n_0$".

F4. Severity: note. Location: Definitions, "$\mathcal A=\{0\}\cup\{2^{-j}:j\in\mathbb N\}$". Defect: the convention $0\in\mathbb N$, inherited silently from the normalization page, is load-bearing here: Step 2's "one of them is at least $1$" needs $1=2^{-0}\in\mathcal A$, and with $\mathbb N$ starting at $1$ the normalized words would not lie in $\mathcal S_k$. Witness: physical p. 13, "at least one is at least one". Proposed replacement: append "(with $0\in\mathbb N$, so $1\in\mathcal A$)".

F5. Severity: note. Location: Step 2, "nonnegative integers whose binary digits are the conversions". Defect: this reading of $U$ and $V$ uses $u_j,v_j\in\{0,1\}$, item 4 of the normalization page, which the page consumes without naming. Witness: physical p. 2, "where $u_i,v_i\in\{0,1\}$". Proposed replacement: "nonnegative integers whose binary digits are the conversions, since $u_j,v_j\in\{0,1\}$ (item 4 of the normalization page)".

F6. Severity: note. Location: Proof, "All windows below come from (10.2) on the windows page, which is available under these hypotheses." Defect: (10.2) is consumed as a premise, but its standing is not named at the point of use; the page's Standing paragraph speaks for the page only. Proposed replacement: "All windows below come from (10.2) on the windows page, consumed here as a premise at that page's author-recorded standing; it is available under these hypotheses."

F7. Severity: note. Location: Statement, "If the sequence is incomplete". Defect: neither this page nor the windows page says what "the sequence" is or what its incompleteness means; the normalization page defines completeness for a set. The intended reading, that the set of weights $\{a_i,b_i:i\ge0\}$ is not complete, is the one the whole cluster uses, and the hypothesis enters this page only through (10.2), so nothing mathematical turns on it. Proposed replacement: in Definitions, "The sequence is incomplete if the set of weights $\{a_i,b_i:i\ge0\}$ is not complete in the sense of the normalization page."

Verdict

Source fidelity: faithful. The statement, the definitions of exact layers, bounded spacing, $\mathcal A$, $\mathcal S_k$ and $\mathcal C_k$, the display (11.1), and the three subsections of Section 11 on physical pp. 13--14 are reproduced with their hypotheses, quantifiers and constants unchanged; the page's added reasons (the compactness of $\mathcal C_k$, the explicit bounds on $E$ and $B^r$, the ordering of the $z_i$) expand the source without altering or strengthening it, and the standing sentence claims author-recorded status only. The suggested corrections F1 and F2 concern labels and the scope account, not the mathematics.

The argument as reconstructed: sound. Every deduction of Steps 1--3 was re-derived above and composes as the page says, with the constants chosen in an order that no later choice disturbs.

Limitations: (10.2) and the normalization page's item 4 are consumed at their author-recorded standing and were checked here only as statements against the source's pp. 2 and 12, not re-proved; the source's Section 6 bound $R=4$ was read but not verified; the meaning of "incomplete" was taken from the cluster's usage (F7); the review is a reading and used no computation.

This focused review assigns no tier and changes no status.