---

Subject and independence

The reviewer is an independent examiner working in a fresh context from the commissioned assignment alone; the reviewer took no part in writing the page or any page in its folder, and the charge is refutation. The subject is path wiki/research/erdos_354/yu_chen_fe_reconstruction.md as it stood at 2026-09-28T05:03:27Z, read whole.

The artifact is the seventeen-page manuscript PDF held as the folder-name PDF beside the library card. Physical pp. 7--9 (printed 7--9), holding Section 8 with displays (8.1)--(8.5), (FE) and (FE-R) and Subsections 8.1--8.3, were read line by line in the text layer and again as page images rendered at 130 dots per inch, every display checked against the image. Physical pp. 1--3 (the theorem and scope, Section 1 with Subsection 1.1, the start of Section 2) were read in the text layer for the definitions of the normalized pair, the digit recurrences, the event set and count, the prefix objects, and gap and span. The top of p. 7 (Sections 6 and 7) was read incidentally with Section 8. Pages 4--6 and 10--17 were not read.

Allowed material read: the normalization page as of the same time, its Definitions and Statement in full and its Proof of items 4 and 6 (the deductions under check consume the digit facts, the gap bound and the identity for $L_n-S_n$); the library card's provenance paragraph; the Statement of Problem 354; docs/verification.md "Whole-claim report" and both "Audit checklist" sections; docs/evidence.md "Source fidelity"; and docs/math_authoring.md whole.

Exposures: the library card was displayed whole, so its Read status, Overview, Standing and Bears-on sections reached the reviewer; the normalization page's own Standing paragraph reached the reviewer with the rest of that page; the opening lines of the problem page's Formulation paragraph came with its Statement; and the file names under the folder's evidence/ directory, including other reviews, were seen in directory listings, with no content under evidence/ read. No Current assessment, no other review, and no web search reached the reviewer. Nothing in the exposed text was used.

Restatement

Setting. A normalized pair is $\alpha,\beta>0$ with $N=\lfloor\beta\rfloor<M=\lfloor\alpha\rfloor<2N$ and $N\ge2$; no irrationality and no incompleteness is assumed anywhere. For $i\ge0$ the weights are $a_i=\lfloor2^i\alpha\rfloor$, $b_i=\lfloor2^i\beta\rfloor$ with digits $u_i=a_{i+1}-2a_i$, $v_i=b_{i+1}-2b_i$ in $\{0,1\}$ and $w_i=u_i+v_i$. $P_n$ is the set of subset sums of the $2n$ weights of indices below $n$, each used at most once, the empty sum included, so $0,S_n\in P_n\subseteq[0,S_n]$ with $S_n=\sum_{i<n}(a_i+b_i)$; $L_n=a_n+b_n$; $K_n$ is the number of indices $i\in[0,n-1]$ with $(u_i,v_i)\ne(0,0)$. $Q_n=L_n-|P_n|$ is the number of integers of $[0,L_n-1]$ outside $P_n$ (well defined as $S_n<L_n$), $e_n=S_n+1-|P_n|$ the number of integers of $[0,S_n]$ outside $P_n$, and $R_n$ the largest $b-a$ over integer intervals $[a,b]$ whose integers all lie in $P_n$. Cardinalities count distinct values.

(FE). For every $n\ge0$,

$$Q_n\le2(M+N+10)\,2^n\exp\Bigl(-\frac{K_n}{64N}\Bigr).$$

(FE-R). For every $n\ge1$,

$$R_n+2\ge\frac{M+N}{2\bigl(2(M+N+10)+1\bigr)}\exp\Bigl(\frac{K_n}{64N}\Bigr).$$

Both statements are asserted for every normalized pair, rational ratio included, with constants depending on $M$ and $N$ only. The auxiliary convention: $f_n$ is the $L_n$-periodic indicator of the positions of $[0,L_n-1]$ outside $P_n$, and $J_t(f)$ sums $|f(x+t)-f(x)|$ over one period.

Checklist

• Quantifiers and scope. Pass. (FE) is proved for all $n\ge0$ (the block bound holds for all $m,d\ge0$ and $m=0$ is taken) and (FE-R) for $n\ge1$, the only place $n\ge1$ is used being $S_n\ge2^{n-1}(M+N)$. The boundary $n=0$ of (8.2) was checked separately: $[1,S_0-1]$ is empty, there is no internal run, and the bound reads $M+N-1\le N+M+N$. The scope "every normalized pair" holds: no step uses irrationality or incompleteness, and the consumed normalization items 4 and 6 are stated there for every normalized pair. The cases $Q_n=0$ and $B_n=1$ mentioned on the page never occur, since $Q_n\ge B_n-1$ and $B_n\ge M+N\ge5$; they are harmless (F3).
• Circularity. Pass. (8.3) at layer $n+1$ is the same statement proved for every index by one argument, not an induction hypothesis; nothing equivalent to (FE) is assumed.
• Model and convention changes. Pass. The periodic extension $f_n$ replaces the finite window, and the transfer is explicit: (8.4) keeps the old period, pays the $w$ padded positions, and the two case arguments restrict to positions inside $[0,L'-1]$ before invoking it. The reduction of $P_n$ modulo $L_n$ is injective on $[0,L_n-1]$.
• Finite and statistical overreach. Inapplicable. No finite check or heuristic stands in for a proof; the two numerical inequalities are proved for all $n\ge0$ and all $N\ge2$.
• Uniformity. Pass. $C_0$, $a$ and $c_0$ depend on $M,N$ only; the potential inequality is verified at its worst case $n=0$, $N=2$ with margin $567/64-35/4=7/64$; the bound $(n+55)/32\le2(n+1)$ holds for all $n\ge0$.
• Extremal conclusions. Pass. $R_n$ is a maximum over a finite nonempty family ($[0,0]$ qualifies), and the run of maximal size in Step 7 exists because the runs are finitely many and nonempty.
• Consequences and composition. Pass. Every "hence" was re-derived (Weakest steps below). The consumed clauses are normalization items 4 and 6, supplied at the strength used; the inequalities $b_n<a_n<2b_n$ are used in Step 5 without citation (F4). Step 7 supplies $a_{n-1}+b_{n-1}\ge2^{n-1}(M+N)$, which follows from $a_{i+1}\ge2a_i$, $b_{i+1}\ge2b_i$.
• Computation. Inapplicable. The page runs no computation and cites no evidence program.
• Reproduction. Inapplicable. There are no rerun commands or coverage claims.
• Source and verdict fidelity. Pass with notes. Every display, constant, label, subsection title and physical page on the page matches pp. 7--9; the Standing paragraph claims author-recorded standing only. The reading of the source's "width" is unmarked (F2), and the page's desc describes $Q_n$ loosely (F1).

Weakest steps

(8.2), the run decomposition. Suppose $Q_n>0$. Since $0\in P_n$, $f_n(0)=0$ and $f_n$ is not constant, so over one period the transitions of $f_n$ come in pairs, one at each end of every maximal cyclic run of $1$s, and $J_1(f_n)=2r$ with $r$ the number of such runs. Every position of $[1,L_n-1]$ outside $P_n$ lies in exactly one run. Because $S_n\in P_n$, a run either lies inside $[1,S_n-1]$ or equals the terminal arc $[S_n+1,L_n-1]$, which is a full run of $1$s bounded by the $0$s at $S_n$ and at $L_n\equiv0$; it has $B_n-1\ge M+N-1$ positions. An internal run is $[p+1,q-1]$ for consecutive elements $p<q$ of $P_n$, so its length is $q-p-1\le\operatorname{gap}(P_n)-1\le N-1$ for $n\ge1$; for $n=0$ there is none. Adding, $Q_n\le(r-1)(N-1)+B_n-1\le(N/2)J_1(f_n)+B_n$. This is the only place the gap bound enters; the terminal run, which can be long, is paid separately by $B_n$, and the potential later absorbs $B_n<3N+2n$.

(8.5) in the case $u=1$. Put $a'=2a+1$, $L'=2L+1+v$. From (8.3) at layer $n+1$, $J_{a'}(f_{n+1})\le2G_{n+1}$; the $2b+v$ residues $x\in[0,2b+v-1]$ are distinct modulo $L'$ and satisfy $a'\le x+a'\le2L+v=L'-1$, so the partial sum over them of $|f_{n+1}(x+a')-f_{n+1}(x)|$ is at most $2G_{n+1}$ and every position involved lies in $[0,L'-1]$. By the triangle inequality, $|f_n(x+a')-f_n(x)|$ is at most $|f_{n+1}(x+a')-f_{n+1}(x)|$ plus $|f_{n+1}(x+a')-f_n(x+a')|$ plus $|f_{n+1}(x)-f_n(x)|$; the positions $x+a'$ are distinct, as are the positions $x$, so each of the two correction sums is a partial sum of the left side of (8.4) and is at most $G_n+w$. Dropping the $v$ terms with $x\ge2b$ (they are nonnegative) and using $\tau(x+2a)\le|f_n(x+2a+1)-f_n(x)|+|f_n(x+2a)-f_n(x)|$, where the last terms over the $2b<L$ distinct residues $x$ sum to at most $J_{2a}(f_n)\le4G_n$, gives $\sum_{x=0}^{2b-1}\tau(x+2a)\le2G_{n+1}+6G_n+2w$. The residues $x+2a$ run over $I=[L-2b,L-1]$ because $2a\equiv-2b$; $I$ and $I+b$ together contain the $3b$ consecutive residues from $L-2b$, which is the whole circle since $3b\ge a+b=L$, that is $a\le2b$. With $\tau(x+b)\le\tau(x)+|\tau(x+b)-\tau(x)|$ and $\sum_{x\bmod L}|\tau(x+b)-\tau(x)|\le2J_b(f_n)\le4G_n$ (the inequality $\bigl||A|-|B|\bigr|\le|A-B|$ followed by the triangle inequality and the substitution $x\mapsto x+1$), $J_1(f_n)\le2\sum_I\tau+4G_n$, which is at most $4G_{n+1}+16G_n+4w$. The case $(u,v)=(0,1)$ uses the $L$ residues $x\in[2b+1,2b+L]$, which lie in $[0,L'-1]$ because $b\le a$, whose shifts $x+2a\ge L'$ wrap to $x-2b-1\in[0,L-1]$; the same two corrections give $J_{2a-1}(f_n)\le2G_{n+1}+2G_n+2$, and $J_1=J_{-1}\le J_{2a-1}+J_{-2a}$ closes with $J_{-2a}=J_{2a}\le4G_n$. Both cases feed Step 6 only through (8.5) with $w\le2$.

The potential and the block partition. At a nonzero conversion, (8.2) and (8.5) give $Q_n\le8NG_n+2NG_{n+1}+2Nw_n+B_n$, hence $G_n+G_{n+1}\ge(Q_n-B_n-4N)/(8N)$ as $G_{n+1}\ge0$. Two applications of (8.1) give $Q_{n+2}=4Q_n+2w_n+w_{n+1}-2G_n-G_{n+1}$, which is at most $4Q_n+6-(G_n+G_{n+1})\le(4-1/(8N))Q_n+B_n/(8N)+13/2$. Dividing by $2^{n+2}$ and using $B_n/(8N)<3/8+n/(4N)\le3/8+n/8$, $z_{n+2}\le\rho z_n+2^{-n}(n+55)/32\le\rho z_n+2(n+1)2^{-n}$. For $V_n=z_n+9(n+1)2^{-n}$ the needed inequality $2(n+1)+9(n+3)/4\le9\rho(n+1)$ reads $4.25n+8.75\le9\rho(n+1)$; at $n=0$, $N=2$ the right side is $567/64=8.859\ldots$, and the slope $9\rho>4.25$ makes larger $n$ easier, so $V_{n+2}\le\rho V_n\le\sigma^2V_n$ because $\sigma^2=\rho+(64N)^{-2}$ and $V_n\ge0$. At a zero conversion $Q_{n+1}\le2Q_n$ gives $V_{n+1}\le z_n+9(n+2)2^{-(n+1)}\le V_n$, and always $z_{n+1}\le z_n+2^{-n}\le V_n$. Scanning the indices $m,\ldots,m+d-1$ from the left, a zero conversion is a one-step block with factor $1$ and no event, a nonzero conversion at $i\le m+d-2$ is a two-step block $\{i,i+1\}$ with factor $\sigma^2\le\sigma^{K_{i+2}-K_i}$ because it holds at most two events and $0<\sigma<1$, and a nonzero conversion at $m+d-1$ is unpaired; chaining the paired blocks bounds $V$ at the end of the last paired block by $V_m$ times $\sigma$ to the number of events in paired blocks, and the unpaired case costs one event and one factor $1/\sigma\le2$ through $z_{m+d}\le V_{m+d-1}$. With $m=0$, $z_0=M+N-1$, $V_0=M+N+8$ and $\sigma^{K_n}\le e^{-K_n/(64N)}$, so $Q_n\le2(M+N+8)2^ne^{-aK_n}$, inside the source's $C_0=2(M+N+10)$.

Strongest attack

The attack aimed at the change of period in (8.5) when $w=2$. There $L'=2L+2$, and the shifted positions $x+a'$ for $x\le2b$ reach $2L$ and $2L+1$, the two padded positions where the word $f_nf_n1^2$ is $1$ while the old periodic extension gives $f_n(2L)=f_n(0)=0$ (and $f_n(2L+1)=f_n(1)=1$, since no weight is below $N\ge2$). The hope was that replacing $f_{n+1}$ by $f_n$ there costs more than (8.4) allows, or that a hole filled by a translate at one of these positions is charged twice. It fails: for every $y\in[0,L'-1]$, $|f_{n+1}(y)-f_n(y)|$ is at most $|f_{n+1}(y)-\text{word}(y)|$ plus $|\text{word}(y)-f_n(y)|$, the first term is nonzero at exactly the $G_n$ filled holes and the second only on the $w$ padded positions, so the sum over any set of distinct positions of $[0,L'-1]$ is at most $G_n+w$ whatever the overlap; the two correction sums in Step 5 are over distinct positions each, so each is charged $G_n+w$ once, exactly as the page says. A second attempt tested whether (8.3) could undercount: an integer $y+a_n$ with residue outside $P_n\bmod L_n$ cannot lie in $P_n+L_n$, whose residues are those of $P_n$, so the injection into the $G_n$ new elements stands. A third attempt tried to break the potential at its tightest point, $n=0$ and $N=2$, where the elementary inequality has margin $7/64$; it holds, and at $n=0$ the bound $B_0/(8N)+13/2\le55/8$ also holds since $B_0=M+N<3N$. No attack produced a counterexample or a gap.

Premises

• Normalization page (same folder, as of the same time, held). Interface used: item 4, $u_i,v_i\in\{0,1\}$ and $b_i<a_i<2b_i$ for all $i\ge0$; item 6, $\operatorname{gap}(P_n)\le N$ for $n\ge1$ and $S_n<L_n$ for all $n$, with the identity $L_n-S_n=M+N+\sum_{i<n}(u_i+v_i)$ from its proof. Read depth: Definitions and Statement in full, Proof of items 4 and 6 re-derived. Its own Standing paragraph describes it as author-recorded; no other standing was read.
• The source, Section 8 (held, pp. 7--9 read in full, with Section 1 on pp. 2--3 for definitions). The page reconstructs it rather than importing it; the statement interface is (FE) and (FE-R) as restated above, with the source's own constants.
• Elementary supplied facts: $a_i\ge2^iM$ and $b_i\ge2^iN$ from the recurrences, used for $S_n\ge2^{n-1}(M+N)$ in Step 7; $K_n\le n$; $1-x\le e^{-x}$; $\bigl||A|-|B|\bigr|\le|A-B|$.
• No external theorem is imported and no assumption beyond the normalized pair is made; neither irrationality nor incompleteness is used.

Findings

F1. Severity: suggested. Location: frontmatter desc, "the number of unrepresented positions below the next weight decays exponentially in the number of events". Defect: $Q_n$ counts the positions of $[0,L_n-1]$ with $L_n=a_n+b_n$, the next period and the sum of the two next weights, not the positions below one weight; and it is the doubling-normalized count $Q_n/2^n$ that decays, $Q_n$ itself being at most $L_n<2^n(M+N+2)$. Witness: p. 7, the definition $Q_n=L_n-|P_n|$ with $L_n=a_n+b_n$ (p. 2), and (FE) on p. 9. Proposed replacement: "Reconstructs the estimate that the number of unrepresented positions below the next period, divided by the period's doubling, decays exponentially in the number of events, through a boundary-variation bound at nonzero conversions and a two-step potential, and the contiguous-run lower bound it implies."

F2. Severity: note. Location: Definitions, "let $R_n$ be the largest $b-a$ over integer intervals $[a,b]$". Defect: the source (p. 9, Subsection 8.3) says "the largest width of an integer interval contained in $P_n$" and does not define width; the page fixes the reading $b-a$ without marking it as a reading. The reading is corroborated by the source's display $R_n+2\ge(S_n+2)/(e_n+1)$, which is exactly what the $b-a$ reading yields, while the element-count reading would give $R_n+1$ on the left and only strengthen (FE-R). Proposed replacement: append "(the source's 'width', read here as $b-a$; reading it as the number of elements would only strengthen (FE-R))".

F3. Severity: note. Location: Step 2, "each of length at most $N-1$ because $\operatorname{gap}(P_n)\le N$ (normalization page, item 6)" and "(empty when $B_n=1$)". Defect: item 6 states the gap bound for $n\ge1$, while (8.2) is asserted for every $n\ge0$ and used at $n=0$ in Step 6 when $w_0>0$; the bound is still true there because $[1,S_0-1]$ is empty and no internal run exists, but the page does not say so. The parenthetical describes an impossible case: $B_n\ge M+N\ge5$, so the terminal run is never empty, and likewise $Q_n\ge B_n-1\ge4$ makes the case "$Q_n=0$" void. Witness: p. 7, $B_n=M+N+\sum_{i<n}(u_i+v_i)$, with $M\ge3$, $N\ge2$. Proposed replacement: "each of length at most $N-1$ because $\operatorname{gap}(P_n)\le N$ for $n\ge1$ (normalization page, item 6), there being none when $n=0$" and "(never empty, as $B_n\ge M+N$)".

F4. Severity: note. Location: Step 5, "because $2b<a+b=L$", "that is $a\le2b$", "that is $b\le a$". Defect: the inequalities $b_n<a_n<2b_n$ are consumed three times without a citation; the source cites "$b<a<2b$" at the covering step (p. 8). They are available from normalization item 4. Proposed replacement: at the first use write "because $2b<a+b=L$ (normalization page, item 4)".

Verdict

Source fidelity: faithful. The statement, constants, labels, subsection titles and physical pages match Section 8 of the held manuscript, and the page alters or strengthens nothing the source proves; the notes F2--F4 concern marking and citation, and F1 the page's own summary line.

The argument as reconstructed: sound. Every deduction from (8.1) to (FE-R) was re-derived, including the two case arguments of (8.5), the numerical inequalities of the potential at their tightest parameters, and the block partition.

Limitations: pages 4--6 and 10--17 of the manuscript were not read, so the Scope paragraph's closing sentence about what the later digit-budget and window arguments exploit is unchecked here; the manuscript's Lean formalization was not consulted; the review covers the page's mathematics against the source and not the standing of any consumer.

This focused review assigns no tier and changes no status.