{"id":"bcd6fd22-e661-4573-b3b6-3c2aacd88551","arxiv_id":"2607.17343","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"A parity-family averaging identity reduces the signed-spectral-radius problem to walk counting and yields claimed ε-versions of Bilu-Linial for dilute graphs, but the main theorems' final constant extraction is arithmetically wrong.","lead":"This paper averages over a special family of edge-signings that makes every short even cycle unbalanced, and claims this converts the Bilu-Linial spectral-radius problem for sparse graphs into a walk-counting problem with near-Ramanujan answers. The main theorems' final step, however, contains an arithmetic error that inflates their claimed accuracy.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"At k=⌈log n⌉, (n(Ck)^c)^{1/(2k)} tends to e^{1/2}, not 1; Props 21/24 thus give an extra constant unless k≫log n, breaking the abstract's (1+o(1)).","rationale":"I agree with the reader's identification. The final lines of Props 21 and 24 miscompute the root of the polynomial prefactor: (n(Ck)^c)^{1/(2k)} at k=⌈log n⌉ is e^{1/2}(1+o(1)), not 1+o(1). This is a clear mathematical error that invalidates the (1+o(1)) in the central claim. The paper's own Corollary 8 makes the obstruction transparent: any trace certificate at k=log n carries at least n^{1/(2k)} → e^{1/2} from the tree term. To get a true (1+o(1)) one must take k ≫ log n, which in turn requires the subcriticality/bicycle-free hypotheses at scale (log n)^{1+ε}. The abstract states 'scale log n', so the theorems as stated are not proven. The flaw is localized and repairable by either strengthening the hypotheses or restating the constants, and the core identities (Props 7, 10), the hypercube certificate (Prop 5), and the interlacing obstruction (Remark 6) appear sound. Given that the central claim is the paper's main advertised contribution, REJECT remains the appropriate verdict. The reader's weakest assumption captures exactly this step, so my read does not change the verdict.","tokens_in":15352,"tokens_out":6639,"duration_ms":59074,"concrete_test":"Recompute the final step of Prop 24 symbolically: for k=⌈log n⌉, show (n(Ck)^c)^{1/(2k)} = exp(1/2 + o(1)). Concretely, set n=e^m, k=m; then n^{1/(2k)} = e^{1/2}, and (Ck)^{c/(2k)} = exp(c log(Cm)/(2m)) → 1. If this evaluation is substituted, the theorem's final bound acquires a factor e^{1/2}, confirming the concern. A numerical spot-check at n=10^12, k=28, c=10 gives (n(Ck)^c)^{1/(2k)} ≈ 1.65.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The decisive step in both Propositions 24 and 21 is the final estimate (n(Ck)^c)^{1/(2k)} = 1+O(log log n/log n) with k=⌈log n⌉. But n^{1/(2k)} = exp((log n)/(2⌈log n⌉)) → e^{1/2} ≈ 1.6487, and (Ck)^{c/(2k)} → 1, so the factor is e^{1/2}(1+o(1)). Hence the proofs deliver ρ(A_σ) ≤ 2√(d−1) e^{1/2} (1+δ)^{4(r0+2)} (1+o(1)) in Prop 24 and a correspondingly inflated constant in Prop 21, not the stated (1+o(1)). To obtain the claimed (1+o(1)) one must take k ≫ log n (e.g., k=(log n)^{1+ε}), which requires the subcriticality/bicycle-free hypotheses to hold at scale (log n)^{1+ε}, not at 'scale log n' as the abstract asserts. This is not a cosmetic constant: it is the conversion of the counting bounds through Corollary 8, and the paper's own Corollary 8 shows the uniform average already carries this n^{1/(2k)} obstruction. The abstract's headline inequality is therefore overclaimed. The error is localized and plausibly repairable by strengthening the scale hypotheses or restating the constant; the core identities and negative results are not affected.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies affine F_2 families of signings of a d-regular graph that make every even cycle of length at most L unbalanced. It proves a master identity (Prop. 7) expressing the family-averaged even trace as a parity-weighted sum over the cycle-span W, and a kernel-averaged Ihara L-function identity (Prop. 10). It then develops counting bounds for totally even and parity-confined non-backtracking walks under two regimes: (r0,δ)-subcritical at scale ℓ and bicycle-free at radius R. These are assembled via Cor. 8 into ε-versions of the Bilu–Linial conjecture: Propositions 21 and 24 claim signings with ρ ≤ 2√(d−1)(1+Cδlog(1/δ))(1+o(1)) (bicycle-free) and ρ ≤ 2√(d−1)(1+δ)^{4(r0+2)}(1+O(log log n/log n)) (dilute). The paper also gives an exact hypercube certificate and a counterexample to real-rootedness of E_σ det(xI−A_σ²). The decisive quantitative step in both theorems is the estimate (n(Ck)^c)^{1/(2k)} = 1+O(log log n/log n) at k=⌈log n⌉, which is false: n^{1/(2k)}→e^{1/2}.","tokens_in":15522,"tokens_out":7726,"duration_ms":75216,"significance":"If the scaling issue is repaired, the paper's structural contributions are significant: a clean conversion of the sign problem to a counting problem, a mechanism showing that uniform averaging cannot certify sub-Kesten signings, a new route to near-Ramanujan signings on dilute/bicycle-free graphs within a constrained parity family, and a decisive obstruction to two-sided interlacing. The exact hypercube certificate, the matched counting bounds, and the extensive exact computational verification (with code provided) are concrete strengths. However, the headline (1+o(1)) consequences are not established at the stated hypothesis scales because of the asymptotic error detailed below.","major_comments":[{"comment":"The proof sets k=⌈log n⌉ and concludes via (n(Ck)^c)^{1/(2k)} = 1+O(log log n/log n). With k=⌈log n⌉, n^{1/(2k)} = exp((log n)/(2⌈log n⌉)) → e^{1/2}, and (Ck)^{c/(2k)} → 1. Thus the factor is e^{1/2}(1+o(1)), not 1+o(1). Corollary 8 therefore yields ρ ≤ 2√(d−1)(1+δ)^{4(r0+2)} e^{1/2}(1+o(1)), not the stated display. This invalidates the (1+o(1)) claim in Prop. 24 and the abstract's dilute-regime consequence. The error is local and repairable: taking k=(log n)^{1+ε} (with corresponding strengthening of the subcriticality scale) or explicitly stating the constant e^{1/2} would fix it.","section":"Prop. 24, proof, final step"},{"comment":"The same asymptotic error appears in the bicycle-free theorem. The constant K defined in Prop. 21 contains a factor n, so (C''K(Ck)^2)^{1/(2k)} includes n^{1/(2k)}→e^{1/2} when k=⌈log n⌉. The sentence \"which is the claim\" is therefore false: the proof delivers an extra constant e^{1/2} times the stated right-hand side. This is load-bearing because this step converts the counting bound into the abstract's headline inequality for the bicycle-free regime.","section":"Prop. 21, proof, final sentence"}],"minor_comments":[{"comment":"The exponent c in (n(Ck)^c) is never defined; it should be the explicit exponent 3(r0+1)^2+2 appearing in the preceding display.","section":"Prop. 24, proof"},{"comment":"The abstract says \"subcritical at scale log n\" while Proposition 24 states the hypothesis at ℓ=2⌈log n⌉. These should be aligned, and the base of logarithms should be specified; the erroneous e^{1/2} estimate is independent of the base, but clarity is needed.","section":"Abstract and Prop. 24"},{"comment":"The fresh-run encoding says run ends are 'visible to a decoder who knows G and the visited set'; since the decoder does not know the visited set at the start, the encoding/decoding procedure should be described more explicitly.","section":"Prop. 16, proof"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern lands exactly: the final asymptotic estimate in Propositions 21 and 24 is wrong at k=⌈log n⌉, giving an extra e^{1/2}. The error is localized and plausibly repairable by strengthening the scale hypotheses or restating the constant, so I do not recommend rejection. The paper also cites an unpublished companion note [4] for exact spectral evaluations used in Section 7; if that claim remains essential, the note should be available or the claim should be proved in the paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, here's my read.\n\nThe paper has real, new machinery. Proposition 7 (the master identity turning the sign problem into a parity-weighted walk sum over W), Proposition 10 (the kernel-averaged Ihara L-function diagonalization), the exact hypercube certificate (Prop 5), and the non-real-rootedness obstruction (Remark 6) all look correct and are genuinely new relative to the cited literature. The experimental section is unusually careful: matched budgets, conditioned-uniform controls, and honest disclosure of the variance-outlier artifact.\n\nThe soft spot is in the two ε-theorems, and it is not cosmetic. Both Prop 21 and Prop 24 conclude by evaluating (n(Ck)^c)^{1/(2k)} at k=⌈log n⌉ as 1+O(log log n/log n). That is false: n^{1/(2k)} = exp((log n)/(2⌈log n⌉)) → e^{1/2} ≈ 1.649. So the proofs as written give ρ ≤ 2√(d−1)·e^{1/2}·(1+δ)^{...}·(1+o(1)), not the stated (1+o(1)). To get the claimed bound you need k≫log n, which means the subcriticality/bicycle-free hypotheses must hold at scale (log n)^{1+ε} rather than “scale log n” as in the abstract. This is load-bearing: it is the step that converts the counting bounds into the headline inequality. I checked the actual line in the proof of Prop 24 — it says exactly what the stress-test flagged. The paper’s own Corollary 8 already shows the uniform average carries this n^{1/(2k)} obstruction, so the mistake is not isolated.\n\nThe error is localized and plausibly repairable — either restate the constant or strengthen the scale hypotheses. The heavy counting lemmas (Props 16, 19, 20, 22) I could not verify line-by-line, but they are plausible and the experimental evidence supports the general picture.\n\nWho is this for? Researchers working on Bilu-Linial, signed spectra, and Ihara zeta functions. The core identities and the negative result (no two-sided interlacing via the expected polynomial) are worth engaging with. The ε-theorems as stated should not be cited without the e^{1/2} caveat.\n\nRecommendation: send to peer review. The paper deserves a serious referee despite the flawed conclusion; the referee should insist on fixing the k-scaling argument before acceptance.","headline":"Real new machinery, but the headline ε-theorems overclaim: the final step evaluates n^{1/(2 log n)} as 1 instead of e^{1/2}.","tokens_in":16220,"tokens_out":2524,"would_cite":true,"duration_ms":23860,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["05C50","05C38"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proves ε-versions of the Bilu-Linial conjecture by averaging over a constrained parity family of signings, showing the spectral-radius bound reduces to counting non-backtracking walks.","keywords":["signing","Bilu-Linial conjecture","Ramanujan graphs","parity family","Ihara L-function","non-backtracking walks","trace certificate","bicycle-free graphs"],"falsifier":"Compute R_F(k) = (Σ_{z∈W} (−1)^{π(z)} N_{2k}(z))^{1/(2k)} at k=⌈log n⌉ for a d-regular graph that satisfies subcriticality at scale log n but has a small dense core reachable by walks of length log n; if R_F(k) exceeds 2√(d−1)(1+Cδ log(1/δ))·e^{1/2}, the proof's conversion step is invalid and the (1+o(1)) in Propositions 21 and 24 does not follow from the stated hypotheses.","tokens_in":15004,"feed_emoji":"🔢","tokens_out":5385,"duration_ms":46290,"temperature":0.7,"pith_summary":"The paper's central claim is an ε-version of the Bilu-Linial conjecture: every d-regular graph that is subcritical at scale log n, and every d-regular graph bicycle-free at radius C log log n / δ, has a signing whose spectral radius is at most 2√(d−1)(1+Cδ log(1/δ))(1+o(1)). The mechanism is to average not over all signings but over the affine F₂ family that makes every short even cycle unbalanced; a master identity then expresses the family-averaged trace as a parity-weighted count of walks confined to the span of the constraint cycles. Averaging the Ihara L-function over this family diagonalizes it: every prime whose parity escapes the span automatically contributes the Ramanujan rate √(d−1), so the whole deviation from Ramanujan is carried by confined primes. The paper proves matched upper and lower bounds on those confined walk counts, using a doubling injection from below and an ear-decomposition/fresh-run encoding from above, plus window and Moore-bound rank lemmas for bicycle-free graphs. If correct, this gives a route to the Bilu-Linial conjecture on locally cycle-rich graphs, while also showing uniform averaging provably cannot certify sub-Kesten signings.","feed_headline":"Counting walks, not sign choices, gets near-Ramanujan signings","feed_subtitle":"Averaging over constrained parity families turns the sign problem into a counting problem, hitting 2√(d−1)(1+ε) on dilute and bicycle-free g","key_machinery":"The master identity (Proposition 7): E_{σ∈F} tr(A_σ^ℓ) = Σ_{z∈W} (−1)^{π(z)} N_ℓ(z), where W is the span of the constraint cycles, π is the parity form, and N_ℓ(z) counts closed walks of parity z. The kernel-averaged L-function identity (Proposition 10) diagonalizes the family-averaged Ihara product, isolating parity-confined primes. The counting engine is the fresh-run/ear decomposition of a non-backtracking walk, whose number of maximal fresh runs equals the cycle rank of its support, plus a window lemma bounding paths in bicycle-free graphs and a Moore-bound rank lemma for irregular graphs.","core_discovery":"On its own terms, the paper establishes that for any d-regular graph satisfying either of two local sparsity hypotheses—subcriticality at scale log n, or bicycle-freeness at radius C log log n/δ—there exists a signing σ in the parity family (all short even cycles unbalanced) with ρ(A_σ) ≤ 2√(d−1)(1+Cδ log(1/δ))(1+o(1)). The proof runs through a trace certificate: at k = ⌈log n⌉, the family-averaged even trace is bounded by (2√(d−1)(1+η))^{2k} times a polynomial factor, and Corollary 8 converts this into a spectral-radius bound for some member of the family. The paper also proves structural results: on the hypercube every solution of the quadrilateral system satisfies A_σ² = nI, giving an exa","pith_inferences":["The proof as written evaluates the trace certificate at k=⌈log n⌉ and uses (n(Ck)^c)^{1/(2k)} = 1+o(1); at k=log n, n^{1/(2k)} = e^{1/2} ≈ 1.649, so the (1+o(1)) factor appears to require the subcriticality/bicycle-free hypotheses to hold at scale (log n)^{1+ε} rather than scale log n as stated in the abstract. If the hypotheses hold only at the stated scale, the argument as written yields an extr","The same parity-averaging idea could be applied to other constraint systems (e.g. only a β_L-fraction of short cycles unbalanced) and would predict that the family's advantage over uniform signings grows with cycle density; this is directly testable on planted-quadrilateral random graphs.","The exact hypercube certificate suggests that graphs in which every 2-path is completed by exactly one 4-cycle admit an entire affine solution family with exact spectral bound; identifying further such two-eigenvalue signed covers would extend the exact regime beyond Q_n."],"forward_implications":["For every d-regular graph satisfying the stated subcriticality hypothesis, the parity family contains a signing with spectral radius within (1+δ log(1/δ))(1+o(1)) of the Ramanujan bound 2√(d−1).","The same conclusion holds for every d-regular graph bicycle-free at radius C log log n/δ, matching the hypothesis scale of the earlier random-signing result but with the conclusion guaranteed inside a constrained family.","Random 2-lift towers satisfy the dilute hypotheses with high probability, so the ε-Bilu-Linial conjecture holds along these sequences.","Uniform averaging over all signings provably cannot certify a spectral radius below the Kesten profile; only the parity-confined average can.","The two-sided interlacing approach via E_σ det(xI−A_σ²) cannot work in general: this polynomial is not real-rooted even for the 4-cycle."],"fun_headline_variants":["Parity averaging: walk counts replace sign search","Subcritical or bicycle-free? Count walks for Ramanujan","Even cycles unbalanced, walks counted, spectra near-Ramanujan","Walk counts certify spectral radius at Ramanujan rate"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the trace certificate is valid at walk length k ≫ log n, while the stated subcriticality/bicycle-free hypotheses are asserted only at scale log n; if they hold only at that shorter scale, the proof's (1+o(1)) factor fails and only an extra e^{1/2} constant is delivered.","fun_headline_variants_meta":{"raw":{"variants":["Parity averaging: walk counts replace sign search","Subcritical or bicycle-free? Count walks for Ramanujan","Even cycles unbalanced, walks counted, spectra near-Ramanujan","Walk counts certify spectral radius at Ramanujan rate"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00045,"raw_usage":{"total_tokens":2217,"prompt_tokens":972,"completion_tokens":1245,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":716,"completion_tokens_details":{"reasoning_tokens":1181}},"tokens_in":716,"tokens_out":1245,"duration_ms":11993,"temperature":1.0,"reasoning_tokens":1181,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T18:23:13.880056+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute R_F(k) = (Σ_{z∈W} (−1)^{π(z)} N_{2k}(z))^{1/(2k)} at k=⌈log n⌉ for a d-regular graph that satisfies subcriticality at scale log n but has a small dense core reachable by walks of length log n; if R_F(k) exceeds 2√(d−1)(1+Cδ log(1/δ))·e^{1/2}, the proof's conversion step is invalid and the (1+o(1)) in Propositions 21 and 24 does not follow from the stated hypotheses.","supporting_citations":[],"review_version":1}