{"id":"1e4a761c-f002-4414-8145-604c357ae7c6","arxiv_id":"2607.13762","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"No finite NPA level reproduces the quantum value of the doubly-tilted CHSH functional in any neighbourhood of the critical tilt; each level strictly overshoots with at least quadratic gap.","lead":"This paper proves that no finite level of the NPA hierarchy is exact for the doubly-tilted CHSH functional near the critical tilt: for every level there is an interval where the level value strictly exceeds the quantum value. The result settles an open question about the convergence of the standard hierarchy of semidefinite relaxations used to bound nonlocal quantum correlations.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 4.4's witness identity rests on Lemmas 4.8–4.9 and a 1.4M-pair enumeration; a missing fine regime or failed grid coverage would collapse the proof. An independent exhaustive check of regime totality and grid coverage would settle it.","rationale":"The Reader's weakest_assumption is exactly the correctness of the finite enumeration that proves Theorem 4.4: Lemma 4.9's grid-coverage bound and Lemma 4.8's regime classification. After reading the proof chain, I agree this is the most load-bearing point. The rest of the argument — the structural laws (Section 3), the small-λ tangent program (Section 5), and the exact-arc lemma (Section 6) — is either fully analytic or uses the witness identity as a black box; if Theorem 4.4 holds, the rest appears sound. I considered whether the external quantum expansion (1) is a greater risk, but it is derived symbolically from a published sextic and is not the novel step. I considered whether Lemma 6.1's bounds could hide an error, but the lemma's proof checks out and is applied with strict margins (t0 > ρ_k). Thus the witness identity is the keystone. The symbolic verifier weakens the concern but does not eliminate it, because totality of the regime classifier is a code property. The recommended test — an independent re-implementation that checks both the grid coverage and E=0 at grid points — would settle whether the concern lands. Pending that, the Reader's CONDITIONAL verdict is appropriate; I do not move it.","tokens_in":23048,"tokens_out":30656,"duration_ms":259089,"concrete_test":"Write an independent, freshly-written program that (1) re-implements the kernel bases (T and D), the word-reduction rules of Lemma 4.5, and y* of Eq. (2) directly from the paper; (2) enumerates all fine regimes exactly as in Lemma 4.8 and, for each regime, verifies that the pairs with side lengths ≤20 realize the full {0,1,2}-grid in every free coordinate (checking the minimal 'large' values and the bound 6+4+6+4=20); (3) evaluates E(K1,K2) at every grid point and checks E=0. If both (2) and (3) pass for all regimes, the witness identity is certified for all levels; any uncovered regime or nonzero E pinpoints the flaw.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Theorem 1.1) depends on the exact witness identity N_k^T Γ(y*) N_k = u_k u_k^T at every level (Theorem 4.4). Its proof is a finite exact-integer verification of 1,413,721 kernel-element pairs with side lengths ≤20, made conclusive by Lemma 4.8 (within each 'fine regime', the error E(K1,K2) is a polynomial of degree ≤2 in each of at most 4 free length coordinates) and Lemma 4.9 (every fine regime realizes a full {0,1,2}-grid of those coordinates with all side lengths ≤20). If a fine-regime classification omits a possible configuration, or the grid-coverage bound fails for some regime (e.g., minimal 'large' values exceeding 6), then E is not proved to vanish for arbitrarily long kernel pairs, and the arc construction of Sections 5–6 — and with it Theorem 1.1 — collapses. The symbolic per-regime verifier (verify_witness_symbolic.py) reduces but does not remove this risk: its 'regime classifier is total' assertion is a property of the code, not a hand-proof, and both verifiers share the same regime definitions. The text's derivation of Lemma 4.9 (6+4+6+4=20) is terse and not independently checked. This is the same load-bearing assumption identified by the Reader.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper claims that for the symmetric doubly-tilted CHSH functional B_s, no finite level of the NPA hierarchy is exact on any neighbourhood of the critical point s=0. For every level k≥2 it constructs explicit rational g_k>0 and s*_k>0 with c_k(s) ≥ 4−s+g_k s^2 on (0,s*_k], while the quantum value is 4−s+s^3/6+O(s^4), yielding a uniform non-exactness gap. The proof is a primal construction: an exactly feasible moment arc y(s)=y_0+s y_1+s^2 y_2 at each level, built from a level-independent signed witness y* satisfying N_k^T Γ(y*)N_k = u_k u_k^T, a uniform Slater point, and a no-repair positivity lemma that makes the arc feasible without o(s^2) corrections. The argument is computer-assisted in the strict sense, with exact-integer enumerations, a fully symbolic per-regime verification, and an independent clean-room re-implementation.","tokens_in":23431,"tokens_out":30786,"duration_ms":268419,"significance":"If correct, this is a substantial result: it answers in the negative an open question of Gigena et al. and demonstrates a genuine uniform failure of finite-level NPA exactness at an attained, self-tested critical point. The proof is unusually rigorous for a computer-assisted argument: exact rational certificates, a level-uniform Slater point, a no-repair lemma with certified interval endpoints, machine-checked enumeration, a symbolic per-regime proof, and an adversarial audit with a clean-room implementation. The mechanism is also conceptually interesting: the quadratic overshoot is shown to live in the non-quantum part of the NPA tangent cone, realized by a single level-independent signed witness. If the central verification step (Theorem 4.4) is correct, the paper establishes a sharp and surprising result about the NPA hierarchy.","major_comments":[{"comment":"The conclusive force of the witness identity N_k^T Γ(y*)N_k = u_k u_k^T rests on the completeness of the fine-regime classification and the grid-coverage lemma. The hand proof of Lemma 4.9 is a single arithmetical sentence ('6+4+6+4=20'), and the totality of the regime classifier is asserted only as a property of the verification code. Because the ≤20 exact-integer enumeration and the symbolic per-regime verifier share the same regime definitions, a missing regime or a failure of the degree/grid bound would invalidate both. This is the single most load-bearing step: without E≡0 for all kernel pairs, the feasible-arc construction of Sections 5–6, and with it Theorem 1.1, collapses. Please expand the proof of Lemma 4.8 to a complete case analysis (including the treatment of empty side words, for which 'first letter' is undefined), give a detailed derivation of the bounds in Lemma 4.9, and","section":"§4.3, Lemmas 4.8–4.9 and completion of Theorem 4.4"}],"minor_comments":[{"comment":"The numerical bounds a_2>1/39, a_3>1/188, a_4>1/641 are stated as unconditional but rely on certified ladder jets from the unpublished companion note [3]. Please state this dependency explicitly, or include the jets in the manuscript so the claims can be verified without access to the companion.","section":"§A / Theorem 1.1"},{"comment":"The swap-symmetrization soundness is cited to companion note [4]. Since it is a one-line averaging argument (any maximizer can be averaged with its party-swap), include the proof for self-containedness rather than referencing an unpublished companion.","section":"§2.1"},{"comment":"The symbolic verifier is described as using 'clamp 3' while Lemma 4.8 states a clamp at 4. Clarify the relationship between these two thresholds and why the different clamp does not affect the totality of the classification.","section":"§4.3"},{"comment":"The treatment of empty side words is not explicit. In the fine-regime data, the 'first letter' of an empty word is undefined. Specify how the regime representation handles the empty word and how Lemma 4.5 (iii) applies when one of the base words is empty.","section":"§4.3 / Lemma 4.5"},{"comment":"The 'germ stabilization law' is described but not formally stated or proven. Since it is not used in the proof, please label it as a heuristic remark or provide a precise statement and proof.","section":"Remark 4.10"},{"comment":"The value c_1(0)≈4.73 is stated without derivation or citation; add a footnote or reference.","section":"§1.1"}],"recommendation":"major_revision","confidential_remarks":"The central claim appears likely correct, and the paper is unusually thorough. The main concern is the completeness of the fine-regime classification in the proof of Theorem 4.4, which is load-bearing for the computer-assisted proof. I recommend requesting a more detailed hand-verifiable derivation of Lemmas 4.8 and 4.9, or an independent regime-classifier implementation, before acceptance. The secondary dependency on the companion notes for the numerical a_k bounds should also be disclosed. If the authors can address the regime-completeness concern, the paper would be a strong candidate for acceptance in this journal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I read Pakhunov's proof that no finite NPA level is exact for the doubly-tilted CHSH functional near the critical tilt. If correct, this settles a concrete open question from Gigena et al. in the negative, and it does so with a genuinely interesting construction: a level-independent signed witness that compresses to a rank-one form at every level, an exactly feasible arc with a cubic objective, and a no-repair lemma that makes the guarantee uniform in k. That is a real result, not a repackaging.\n\nWhat earns credit: the proof chain is unusually explicit. The witness identity is checked by exact integer arithmetic over a finite enumeration, with degree and coverage lemmas that turn the finite check into a proof. The independent re-verification—including a symbolic per-regime proof and a clean-room implementation—is a serious attempt to catch self-deception. The one external input is the published quantum value, cross-checked symbolically. This is how computer-assisted proofs in quantum information should be packaged.\n\nThe soft spots are real but not fatal, in my reading. The biggest is the pair of companion notes [3] and [4], same author, unpublished, which supply the almost-quantum expansion and the swap-symmetrization soundness lemma. The main theorem does not need everything in those notes, but the upper-bound half of the exponent statement does, and that part is explicitly conditional. More centrally, the proof of Theorem 4.4 depends on Lemma 4.9: every fine regime realizes a full {0,1,2}-grid within side length 20. That lemma is argued in a few lines, and the 1.4M-pair enumeration is the load-bearing step. I did not run the code, and the stress-test note has a fair point: the totality of the regime classifier is a property of the verifier, not a hand-proof. An independent exhaustive check of regime totality and grid coverage would remove most of my residual worry. That is a request for more evidence, not a claim that the proof is wrong.\n\nWho should read this: anyone working on SDP hierarchies, Bell nonlocality, or sum-of-squares convergence. It deserves a serious referee—ideally someone willing to actually run the verification scripts and comb through Lemma 4.8–4.9. I would not desk-reject it. I would send it to peer review and ask for a high bar on the enumeration coverage before acceptance.","headline":"A serious, carefully built proof that no finite NPA level is exact near the critical tilt; the central argument looks sound, but the load-bearing finite enumeration deserves independent scrutiny before full acceptance.","tokens_in":23835,"tokens_out":1428,"would_cite":true,"duration_ms":37728,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["81P40","90C22"],"pacs":["03.65.Ud","03.67.-a"],"model":"deepseek-v4-flash","headline":"No finite level of the NPA hierarchy is exact for the doubly-tilted CHSH functional near its critical tilt; every finite level overshoots the quantum value on a small interval.","keywords":["NPA hierarchy","doubly-tilted CHSH","quantum value","semidefinite programming","exactness","critical tilt","computer-assisted proof","tangent cone"],"falsifier":"Run the exact verifier against the closed-form witness: compute E(K₁,K₂)=Σ c_p c'_q y*(m(p,q))−u(K₁)u(K₂) for all 1,413,721 kernel-element pairs with side lengths ≤20; any nonzero value refutes the witness identity and the quadratic lower bound. Independently, fix a level k, compute c_k(s) at s=10⁻⁶ and 10⁻⁷ with high-precision semidefinite programming; if (c_k(s)−4+s)/s² tends to 0 as s→0, the claimed quadratic gap is false.","tokens_in":22951,"feed_emoji":"🎯","tokens_out":7188,"duration_ms":69973,"temperature":0.7,"pith_summary":"The paper tries to prove that the standard semidefinite relaxation hierarchy for quantum correlations fails, at every finite level, to reproduce the true quantum value of the doubly-tilted CHSH functional in any neighbourhood of its critical point. It constructs, for each level k≥2, an explicit feasible moment curve whose objective is 4−s+g_k s² while the true quantum value is 4−s+O(s³), so the level-k value strictly exceeds the quantum value on a small interval. The proof rests on a closed-form, level-independent witness moment function that produces a rank-one compression on the relation lattice at every level, together with exact-integer computer verification of a finite enumeration. If correct, the result answers an open question from the recent determination of the quantum maximum and shows the failure is genuinely non-quantum: the overshoot direction cannot come from any state or smooth family of quantum models.","feed_headline":"No finite NPA level is exact at the tilted CHSH critical point","feed_subtitle":"Every finite level overshoots the quantum value on a tiny interval, by at least a quadratic gap.","key_machinery":"The proof is carried by three interacting objects: the relation lattice N_k, the closed-form witness y*, and a no-repair positivity lemma. N_k consists of integer vectors forced into the kernel of every optimal moment matrix at the critical point—trivial reductions and dressings—and has dimension 2k². The witness y* is a signed class function with the exact rank-one compression N_kᵀΓ(y*)N_k = u_k u_kᵀ at every level, where u_k is an explicit level-uniform vector. The no-repair lemma shows that once the first-order compressed form is rank-one and the second-order block has a strictly positive margin, the quadratic arc y_0+s y_1+s² y_2 is exactly feasible, giving the quadratic lower bound with","core_discovery":"The central claim is that for every NPA level k≥2 there are explicit rational constants g_k>0 and s*_k>0 such that c_k(s) ≥ 4−s+g_k s² for all s in (0,s*_k]. Since the quantum value has the cubic expansion c_Q(s)=4−s+s³/6+O(s⁴), every finite level strictly overshoots the quantum value on some interval (0,ε_k]; no finite level is exact on any neighbourhood of the critical point. The lower bound is proven by a primal construction: a quadratic moment arc y_0+s y_1+s² y_2 that is exactly feasible at each level, built from a level-independent signed witness y* satisfying N_kᵀΓ(y*)N_k = u_k u_kᵀ. The author also shows that for k≥3 no quantum state and no smooth curve of quantum models can realize","pith_inferences":["If the proof is correct, the non-exactness is a structural property of the restricted certificate class: any relaxation that truncates monomial word length the same way will inherit the same uniform gap near this critical point.","The signed, level-independent witness suggests a template for other Bell functionals whose quantum maximum has cubic contact at a boundary point: look for a class-function witness with rank-one compression and use the same no-repair arc to prove finite-level inexactness.","A quantitative next step would be to measure the decay of a_k or of t_0(k): a uniform-in-k lower bound on g_k, or a level-to-level inequality, would turn positivity into a divergence rate for the required level.","Because the exponent-two upper bound rests on the companion note's conditional expansion, a clean verification of that dual arc would make the Θ(s²) statement unconditional—a sharper target than re-checking the whole chain."],"forward_implications":["For every fixed level k≥2 there is an explicit interval (0,ε_k] on which the level-k relaxation value strictly exceeds the quantum value; no finite level is exact in any neighbourhood of the critical tilt.","The level of the hierarchy required to reach the quantum value diverges as the tilt approaches the critical point: D(s)→∞ as s→0⁺.","The overshoot is at least quadratic for every k≥2, with unconditional numerical constants a₂>1/39, a₃>1/188, a₄>1/641; granted the companion's dual expansion, the exponent is exactly two and bounded above by 3/64.","The gain direction cannot be realized by any state or by any smooth family of quantum models, locating the finite-level failure in the non-quantum part of the NPA tangent cone.","All constants in the construction—g_k, s*_k, ε_k—are explicit rationals, and the proof is accompanied by exact-integer verification programs that can be rerun."],"fun_headline_variants":["No finite NPA level nails tilted CHSH near critical tilt","Finite NPA levels always overshoot critical CHSH quantum bound","Doubly-tilted CHSH: every finite NPA level fails near criticality","NPA hierarchy has no exact level for critical tilted CHSH","Even high NPA levels overshoot quantum value at critical tilt"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The proof leans on the exhaustive enumeration of kernel-pair configurations in Lemma 4.9—that every fine regime is covered by a {0,1,2} grid within side length 20, so checking 1,413,721 pairs settles the witness identity—and on the published quantum-value expansion c_Q(s)=4−s+s³/6+O(s⁴) to convert the quadratic lower bound into strict overshoot.","fun_headline_variants_meta":{"raw":{"variants":["No finite NPA level nails tilted CHSH near critical tilt","Finite NPA levels always overshoot critical CHSH quantum bound","Doubly-tilted CHSH: every finite NPA level fails near criticality","NPA hierarchy has no exact level for critical tilted CHSH","Even high NPA levels overshoot quantum value at critical tilt"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000158,"raw_usage":{"total_tokens":1194,"prompt_tokens":1012,"completion_tokens":182,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":756,"completion_tokens_details":{"reasoning_tokens":90}},"tokens_in":756,"tokens_out":182,"duration_ms":2956,"temperature":1.0,"reasoning_tokens":90,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T03:46:53.644159+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the exact verifier against the closed-form witness: compute E(K₁,K₂)=Σ c_p c'_q y*(m(p,q))−u(K₁)u(K₂) for all 1,413,721 kernel-element pairs with side lengths ≤20; any nonzero value refutes the witness identity and the quadratic lower bound. Independently, fix a level k, compute c_k(s) at s=10⁻⁶ and 10⁻⁷ with high-precision semidefinite programming; if (c_k(s)−4+s)/s² tends to 0 as s→0, the claimed quadratic gap is false.","supporting_citations":[],"review_version":1}