Bourgain–Fuchs: review with singular-series follow-up

Updated 20 October 2026. The original 15 October panel report is retained below.

New finding: the intended uniform singular-series bound is false

A user-supplied Web GPT Pro argument has now been checked against the preserved author edition, independently audited by a separate agent and reproduced with exact integer enumeration. For a family in one fixed permitted strip packing, the right-hand arithmetic expression in author-copy Lemma 3.6 stays at 4 while the singular series equals $160(2t+1)/(9\pi^2)$ and grows without bound.

The full counterexample and checking record give the packing, forms, local calculations, source qualification and short-interval selection caveat. The JAMS typesetting was not inspected. This is new, separately attributed evidence; it is not retrospectively counted as a completed original panel assessment. The positive-density theorem remains unresolved by this review, and no complete repair of the supplied proof is claimed.

The audit now also explains the equal-weight valuation layers highlighted in the response relayed by Nikolay Bogachev. It distinguishes the valid first-level finite-field bound from the invalid lifting comparison, and lists the remaining obligations for the proposed divisor-weight replacement. No general repair or new comparison with the JAMS typesetting is claimed.

The decorated graph records the singular-series formula as refuted in that stated scope, and the earlier short-interval proof step as defective with its statement unresolved. Most other nodes remain unadjudicated. The older roadmap's edges are candidate dependencies, so the displayed terminal markers do not certify a complete proof hierarchy or propagate a headline verdict.

Original report: 15 October 2026

Bourgain–Fuchs: a substantive gap in the supplied argument

Review date: 15 October 2026. Jean Bourgain and Elena Fuchs, A proof of the positive density conjecture for integer Apollonian circle packings. Primary reviewed edition: arXiv 1001.3894v1, 21 January 2010, 20 pages.

Disposition and scope

Red: a substantive defect in the supplied proof, with no checked repair. Sol identified an unsupported short-interval counting deduction; Gemini independently confirmed the objection against the complete reviewed source and found the same deduction in the separately supplied author-hosted copy. The affected proof input is the congruence-average lemma labelled sievebound. It is Lemma 3.7 in the arXiv source index and Lemma 3.8 in the author copy.

The headline positive-density theorem remains unresolved by this review. A gap in this supplied argument is not a counterexample to its conclusion. No repair was attempted in this campaign, and no claim of irreparability is made. This disposition concerns the exact manuscript editions described below; the journal text was not inspected.

The campaign obtained 15 accepted local Sol assessments and two accepted Gemini arbitration assessments. Opus's packet attempts supplied no reports. Findings outside the two returned arbitration cases remain unadjudicated. This is a source-backed, AI-assisted review with independent arbitration, not a human referee judgment or formal proof certificate.

The argument being examined

The manuscript parametrizes tangent-circle curvatures by shifted binary quadratic forms. It chooses a family of shifts, obtains a lower bound for the combined sizes of represented-value sets, and seeks an upper bound for their pairwise overlaps. A quaternary-form counting argument and congruence averages control those overlaps; inclusion–exclusion then supplies the proposed positive-density conclusion.

The short-interval congruence bound is used in the overlap estimate, which in turn enters the final density argument. The objection therefore concerns a proof input to the headline result, rather than only an unused explicit constant.

The separate reading page and dependency map present this structure without panel conclusions. Native source labels and quoted passages take priority over inferred printed numbers; model reports sometimes use the comparison copy's numbering.

Main finding: a cumulative estimate does not give the required short-interval bound

The source defines a cumulative progression count $B(x,q,r)$ and quotes the estimate

The proof of sievebound then invokes this estimate for intervals starting at $x=2^k$ with length $h=\eta 2^k/\sqrt{k}$. Its claimed upper bound has scale $\eta(c_q/q)2^k/k$.

Subtracting the available cumulative estimates at the two endpoints leaves an error allowance of scale

For a fixed admissible modulus, this allowance is larger than the required short-interval scale by a factor proportional to $k^{5/14}/\eta$, which grows with $k$. The pointwise error estimate supplies no control on the difference of its endpoint errors. This does not show that the actual counting error is that large; it shows that the displayed estimate does not justify the displayed deduction.

The exact source transition is: “By Lemma~\ref{friedithm}, the inner sum is bounded above (up to a constant) by”, followed by the claimed short-interval expression. It occurs at reviewed-source lines 552–559, after the cumulative estimate at lines 538–541. In the author-hosted copy, the corresponding deduction appears in the proof of Lemma 3.8 on page 22.

What would resolve this: a valid uniform short-interval estimate or an upper-bound sieve argument with the necessary interval lengths, residue conditions and parameter dependence. The manuscript mentions an upper-bound sieve as a possible route, but that remark does not supply the missing estimate. We have not independently established such a replacement.

Additional issues in the same deduction

Gemini also confirmed the following objections to this local proof, retaining them separately in the arbitration record:

• The selected curvatures are shifted form values, $a=f_{a_0}(x,y)-a_0$. Passing to sums of two squares requires tracking that shift and the multiplier in the completed-square identity. The displayed comparison does not carry those transformations through the intervals and residue classes.
• The quoted cumulative estimate restricts the residue class, including coprimality. The congruence-average lemma is stated without those restrictions and is then applied to divisible residues.
• The displayed error factor $(\log q/\log\log q)^{1/7}$ is unbounded, so the subsequent uniform absorption into a constant does not follow as written.
• The literal factor $\log\log q$ also needs a small-modulus convention at $q=2$. This last issue was classified as cosmetic, separately from the substantive short-interval gap.

The arbiter found these passages in the supplied author-hosted copy as well. This is an assessment of those displayed arguments, not an inspection of the journal edition or of every possible route to the estimates.

A separate overbroad counting statement

The source's density lemma gives an explicit leading-constant formula for a positive definite integral binary quadratic form without requiring that the form be primitive. Sol's scaling objection was independently accepted by Gemini.

Take $f_1=x^2+y^2$ and $f_2=2x^2+2y^2$. Their represented-value counts satisfy $B_{f_2}(X)=B_{f_1}(X/2)$, so positive leading constants must satisfy $c(f_2)=c(f_1)/2$. The displayed formula instead assigns the same positive constant to discriminants $-4$ and $-16$: their ramified prime sets and the relevant Kronecker characters coincide. Thus the standalone statement and its explicit formula cannot both hold under all its printed hypotheses.

The same formulation appears as Theorem 2.2 in the author copy, with a changed attribution. This finding does not disprove the underlying results of James or Bernays. The downstream argument only needs an appropriate positive leading constant for the relevant fixed form, so this defect alone does not defeat the headline theorem.

Objections narrowed or rejected

The arbiter rejected one alleged missing-definition issue: the relevant relation and quadratic form are present in the full manuscript, immediately before the quoted counting lemma. Their absence was a packet-context artifact. It is not a manuscript gap.

Earlier source checks also identified a missing $a_0$ in a quadratic-form coefficient and a wrong entry in the displayed group homomorphism. Both expressions are corrected in the supplied author-hosted copy. They are not the basis of the red disposition. Their broader surrounding proofs did not receive a completed arbitration assessment in this pass.

These distinctions are why local reviewer objections pass through a separate arbitration stage before a curated disposition is assigned.

Editions and source record

The reviewed PDF, reviewed LaTeX and source manifest preserve the exact arXiv v1 input. Original native bytes are also retained. Identifying front matter was withheld line-for-line; mathematical content was not corrected. The figure PDFs and captions are preserved, but the text-only panel did not receive figure pixels.

The separate 25-page author-hosted PDF was supplied to arbitration for version checks. Its PDF creation metadata is 20 February 2011; that metadata is not a verified manuscript revision date. References to the “2011 revision” in raw model assessments identify this supplied copy, not an independently verified publication date.

The article has a JAMS journal reference, recorded in the manifest, but the journal text was not inspected. Persistence in the supplied author copy must not be restated as a verified defect in the journal typesetting. No separate author response was supplied, and author agreement is not inferred.

Panel coverage and failures

There were 40 generation attempts, including failures. The original Sol request supplied neither a usable report nor usage data. Initial Opus, Fable and Gemini calls exhausted their allowances. Every bounded Opus attempt used its 16,000-token allowance for reasoning without delivering a report. One bounded Sol response was rejected because a quotation did not match the packet; it remains outside the accepted findings.

Gemini completed arbitration of density and sievebound. Its lowbound and niederheath arbitration attempts exhausted their 16,384-token allowances and supplied no verdicts. Eleven additional queued targets were not submitted for arbitration. Missing or failed assessments are not votes for or against the paper.

All mathematical body text was assigned across 11 named-claim tasks and five narrative tasks, but assignment is not successful proof coverage. The shared map remains a candidate dependency map; external theorem texts, global dependency closure and human mathematical grading remain incomplete. See the local reports and attempt ledger.

Cost

Conservative inference accounting is $8.0285795, against the $10 campaign allowance: $3.452580 for the initial batch, $3.6076035 for bounded reviews, and $0.968396 for arbitration. This includes failed calls and the original Sol request's full $0.605168 reservation because its charge is unknown. It is not an invoice reconciliation. Infrastructure, storage and this interactive session are outside those inference-token figures.

Publication reuses the collected evidence and starts no new reviewer calls. The cost record preserves the calculation and its limits.