{"id":"070d6c65-0cb0-4515-803f-475255415b84","arxiv_id":"2511.08445","paper_version":2,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"New non-abelian amplification bounds bilinear Kloosterman sums for composite moduli, saving c^{-1/12} for products of two primes of similar size.","lead":"The paper proves new power-saving bounds for bilinear sums of Kloosterman sums with composite moduli, using non-abelian Fourier analysis on SL2(Z/cZ) and a new amplification argument. It resolves the previously open cases p^2 and pq at square-root length and yields applications to twisted L-function moments and large sieves for exceptional cusp forms.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified; Proposition 6.4's count survives scrutiny but warrants independent verification.","rationale":"The reader's weakest-assumption pick matches mine: the entire power saving is produced in Section 6's count. On close reading, the argument holds: the potential d=1 undercount is avoided because (6.8) uses H1H2/d; the two counting methods for h2,h3 are correctly bounded by a min; and the final max over d reduces to endpoints, with intermediate d giving smaller products in the tested parameter ranges. The proof is intricate but consistent, and the claim is independently plausible (matches heuristics). No machine-checked verification exists, but the risk is not sufficient to change the ACCEPT verdict. The proposed concrete test — a brute-force check for small moduli — would de-risk the only non-formal step without requiring a full rewrite.","tokens_in":51455,"tokens_out":26970,"duration_ms":206862,"concrete_test":"Brute-force verify Proposition 6.4 for small moduli and small H. For c in {4,8,9,12,16,25,49, p^2, pq} and all 1 ≤ H1 ≤ H2 ≤ 10 (or up to min(√c,10)), compute the exact number of tuples (h1,...,h6) with |h_i| ≤ H1 for odd i and |h_i| ≤ H2 for even i satisfying T^{h1} S T^{h2} S ... T^{h6} S = I in PSL2(Z/cZ), with a1=a2=1. Compare each count against C_ε c^ε (H2^2 + (H1H2)^2/c) for ε=0.05 and a modest constant C_ε (e.g., 10). If any count exceeds the bound, the proposition fails; if all pass, the core counting estimate is supported in the small-modulus regime.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim depends on Proposition 6.4, the q=6 solution count in PSL2(Z/cZ). This count is the only source of power saving, and the proof is a delicate case analysis. I examined the main risk points: (i) the potential undercount for d=1, where h5h6 might naively have H1H2 values rather than H1H2/c; this is resolved because formula (6.8) correctly carries the factor (1+H1H2/d), so for d=1 it gives H1H2, not H1H2/c. (ii) The h2,h3 counting uses a min of two methods — one via the product h2h3, the other via summing over h2 with a gcd-weighted bound — and the resulting expression is consistent with the final bound. (iii) The final maximization over d is compressed but appears correct: for tested parameter ranges (H1=H2=√c, H1=1,H2=c, H1=H2=c) the bound never exceeds the claimed H2^2 + (H1H2)^2/c by more than a constant factor, which is absorbed by c^{o(1)}. No hidden family of solutions was found. The residual risk is human error in the intricate endpoint analysis, since the proof is not machine-checked, but I do not find a concrete flaw.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a new method, based on non-abelian Fourier analysis and a non-abelian amplification argument, for bounding bilinear (Type II) sums of Kloosterman sums with composite moduli. The central result, Theorem 1.1, gives a power saving over the trivial bound c for sequences of length at most c^{1/2+o(1)} for every modulus, with exponents 1/700 in general and 1/276 when one sequence is bounded. The argument proceeds by realizing Kloosterman matrices as Fourier coefficients of a representation ρ_c^∘ of SL_2(Z/cZ), applying an amplifier supported on a congruence subgroup Γ_c(d), and reducing to an elementary count (Proposition 6.4) of solutions in PSL_2(Z/\\tilde d Z). The paper also proves an asymptotic for twisted second moments of modular L-functions (Theorem 1.5) and an improved large sieve for exceptional Maass forms for composite levels (Theorem 9.4 / Corollary 1.6).","tokens_in":51797,"tokens_out":26238,"duration_ms":228236,"significance":"If the main results are correct, this is a substantial advance: it supplies the first nontrivial Type II Kloosterman bound at square-root length for the composite moduli (such as p^2 or pq) that were the remaining barriers for the moment application, and it does so with an explicit, self-contained method that avoids the algebraic-geometric machinery used in the prime-modulus results. The non-abelian amplifier construction is new and likely to be useful independently. A particularly valuable feature is that the power saving is produced by a purely elementary counting argument, with no fitted parameters or circular benchmarks; the savings are measured directly against the trivial bound (1.2), and external results are used only for the near-prime and prime-modulus cases. The paper also honestly acknowledges the simultaneous independent work [29]. The main risk is the intricacy of Proposition 6.4, the sole source of the saving; I checked the main danger points and found no counterexample, but the proof is delicate enough that a fully expanded verification of the endpoint maximization is desirable.","major_comments":[],"minor_comments":[{"comment":"In the sentence 'Since h5h6 ≡ 1 (mod d), there are O(1 + H1H2/c) ways to pick the nonzero integer h5h6', the denominator should be d, not c. The factor (1+H1H2/d) appears correctly in (6.8), but the displayed sentence is internally inconsistent and, read literally, would undercount the h5,h6 possibilities. Please correct and clarify.","section":"§6, Proposition 6.4, Case 2"},{"comment":"The transition from the bound after (6.8) to the final maximum is compressed. It is asserted that after expansion each term is monotone in d, so the maximum occurs at d=1 or d=c; however the displayed algebra also contains an inequality that should be ≤ rather than = when min(H1/(cd),1) is replaced by H1/(cd). Since Proposition 6.4 is the load-bearing combinatorial step, please spell out this maximization and the inequalities in full.","section":"§6, Eq. (6.8)"},{"comment":"In the sum over g in the estimation of the h2,h3 contribution, the congruence for h2' should be h2' ≡ g^{-1} r (mod d), not h2' ≡ gr (mod d). The resulting count is the same, but the displayed congruence is a typo that should be corrected.","section":"§6, proof of Proposition 6.4"},{"comment":"The right-hand side of the amplifier inequality is written as a complex-valued sum involving χ(g1···gq). The proof shows that this quantity arises from a nonnegative expression, but the statement itself would benefit from a remark that the displayed sum is real (or that one takes its real part), to avoid an apparent type mismatch with the real left-hand side.","section":"§5, Proposition 5.1"},{"comment":"The acknowledgment of simultaneous independent work is appropriate and should remain. It may be worth noting explicitly which results overlap and which remain unique to this paper, since the current remark is brief.","section":"§1.2, Remark on [29]"},{"comment":"The remark after Lemma 5.4 states that the bound is expected to hold for all 0≤j≤k. As written, Proposition 5.3 uses only the proved range, so this is not a gap, but the reader would benefit from a sentence clarifying that the unproved extension is not needed for the main theorems.","section":"§A, Lemma 5.4"}],"recommendation":"minor_revision","confidential_remarks":"This is a strong and original paper. I did not find a concrete mathematical error, and the central claim appears sound after scrutiny of the delicate counting step. My recommendation of minor revision is based on the need to correct several local errors and to expand the verification of Proposition 6.4, which is the sole source of the power saving and should be presented in a way that a reader can check without reconstructing the endpoint maximization. If the editor wishes, an independent check of Proposition 6.4 would be prudent, since the casework is intricate and not machine-checked; however, I do not see a basis for rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Let me get straight to it: this paper is the real thing. It proves power savings over the trivial bound for bilinear Kloosterman sums when M,N are about sqrt(c) for all composite moduli, including p^2 and pq, which the q-van der Corput method could not handle. The mechanism—non-abelian Fourier analysis on SL2(Z/cZ) with an amplifier built from the normal subgroup Γ_c(d)—is genuinely new and reduces the problem to a concrete count of solutions in PSL2(Z/dZ). I checked the count in Proposition 6.4 closely because it carries the saving. The stress-test worry about the d=1 case is resolved by the (1+H1H2/d) factor in (6.8), and I do not see a hidden family of solutions. The final maximization over divisors is terse but plausible.\n\nThe paper is also honest: it states Conjecture 6.2 instead of overclaiming, it integrates the prime-modulus results of Kowalski–Michel–Sawin and Blomer–Milićević rather than duplicating them, and it openly flags the simultaneous work of Milićević–Qin–Wu. That is the right scholarly tone.\n\nSoft spots: the proof of Proposition 6.4 is long elementary casework and is not machine-checked, so there is residual risk of an endpoint error. I would trust it provisionally, but I would want a second referee to parse that section carefully. The savings are modest (c^{-1/700}, c^{-1/276}), which is normal for first-generation results in this area. The paper is dense—49 pages—but not bloated.\n\nWho is this for? Anyone working on Kloosterman sums, moments of twisted L-functions, or large sieve inequalities for exceptional cusp forms. It deserves a serious referee, and I recommend sending it out.","headline":"A dense but genuine advance: new power savings for bilinear Kloosterman sums at composite moduli, with the counting lemma surviving close scrutiny; worth a serious referee.","tokens_in":52247,"tokens_out":2311,"would_cite":true,"duration_ms":24909,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["11L05","11F66","11F72"],"pacs":[],"model":"deepseek-v4-flash","headline":"By replacing abelian characters with a non-abelian amplifier on SL2(Z/cZ), this paper proves a power saving over the trivial bound for bilinear Kloosterman sums at square-root length for every composite modulus.","keywords":["Kloosterman sums","bilinear forms","non-abelian Fourier analysis","amplification","SL2(Z/cZ)","moments of L-functions","large sieve","exceptional cusp forms"],"falsifier":"Compute, for c=p^2 and H1=H2=p, the exact number of 6-tuples (h1,...,h6) in [-p,p]^6 satisfying T^{h1}S...T^{h6}S ≡ ±I mod p^2 or ≡ ±I mod p. Proposition 6.4 predicts at most p^{2+o(1)} and p^{3+o(1)} solutions respectively; finding p^{2+δ} or p^{3+δ} for a fixed δ>0 would invalidate the saving.","tokens_in":51379,"feed_emoji":"🧮","tokens_out":7554,"duration_ms":72047,"temperature":0.7,"pith_summary":"The paper establishes a new upper bound for bilinear sums of Kloosterman sums modulo composite integers c: for sequences supported on intervals of length about sqrt(c), the sum is at most ||alpha|| ||beta|| c^{1-1/700+o(1)}, improving on the trivial bound c. Previous power savings of this type were available only for prime moduli or moduli with special factorizations, and the missing cases—squares of primes and products of two comparable primes—are exactly where the new method works best, saving c^{-1/12}. The argument identifies bilinear Kloosterman matrices with Fourier coefficients of functions on SL2(Z/cZ), then amplifies with a normal-subgroup-supported character sum to reduce the problem to counting solutions of a matrix equation. As applications, the paper proves an asymptotic for averaged second moments of twisted cuspidal L-functions for every modulus, and a large-sieve inequality for exceptional cusp forms.","feed_headline":"Saving c^{-1/700} on Kloosterman sums for every composite modulus","feed_subtitle":"The difficult composite cases p^2 and pq, which earlier methods missed, now yield a c^{-1/12} saving.","key_machinery":"The central objects are the permutation representation rho_c of SL2(Z/cZ) on the projective line P1(Z/cZ), given by Möbius transformations, and its 'sifted' subrepresentation rho^o_c, obtained by removing all subrepresentations that factor through reduction mod d for d|c. A Kloosterman matrix is shown to have operator norm bounded by c times the spectral norm of a Fourier coefficient of a short-word function at rho^o_c. The new amplifier, Proposition 5.1, weights the sixth moment of singular values by a character sum over a normal congruence subgroup Gamma_c(d); this turns the problem into counting 6-tuples h1,...,h6 with |hi| at most H satisfying T^{h1} S ... T^{h6} S = I in PSL2(Z/dZ). The","core_discovery":"Let c be any positive integer and let M,N not exceed c^{1/2+o(1)}. The paper proves that the bilinear sum of Kloosterman sums S(am,n;c) over m≤M, n≤N with (m,n,c)=1 is bounded by ||alpha|| ||beta|| c^{1-1/700+o(1)} for arbitrary complex sequences, and by sqrt(M) ||beta|| c^{1-1/276+o(1)} when the alpha coefficients are bounded by 1. This beats the trivial bound c^{1+o(1)} in the hardest balanced regime where MN is about c. The engine is Theorem 1.2: after writing c=dd'e with d'|d and (d,e)=1, the saving is governed by (f/min(c,d^2))^{1/6}, where f is the largest integer with f^2|cd. For c a square of a prime or a product of two primes of comparable size, this yields a c^{-1/12} saving, and c","pith_inferences":["A natural stress test is the q=8 analogue of the counting bound: if the number of 8-tuples obeys Conjecture 6.2 in the range H roughly sqrt(c), the same framework would give nontrivial prime-modulus bounds for lengths beyond p^{3/8+o(1)}, a barrier that also appears in known results; the paper explicitly leaves this open.","The exponents 1/700 and 1/276 are small because a sixth moment is used to control the top singular value; using a higher even moment or refining the character-size lower bound should improve the exponents without changing the mechanism.","The same pattern—choose a large normal subgroup, amplify, then count solutions in PSL2(Z/dZ)—should also bound bilinear sums of other exponential sums with projective-line geometry, such as additive characters of Möbius transformations; the author notes this possibility but does not carry it out.","The moment application is stated for holomorphic cusp forms; extending it to Maass cusp forms would require additional care about the Ramanujan bound, as the paper indicates, so a fully unconditional Maass-level version is not immediate."],"forward_implications":["For every composite modulus c, bilinear Kloosterman sums at length about sqrt(c) now admit a uniform power saving over the trivial bound, including the previously resistant cases p^2 and pq with comparable primes.","For moduli such as p^2 or pq with comparable primes, the saving reaches c^{-1/12} in the balanced range, and the method beats the trivial bound for M,N between c^{5/12+o(1)} and c^{5/8-o(1)}.","The averaged second moment of two twisted cuspidal L-functions over primitive characters mod q satisfies the predicted asymptotic with an error q^{1-1/674+o(1)} for every modulus q.","For composite levels q with a divisor d of size sqrt(q) such that q/d is square-free, the exceptional-spectrum large sieve inequality holds with a factor q^{6θ/5}, improving on the standard q^θ loss when N is about sqrt(q).","The non-abelian amplification lemma applies to arbitrary finite groups and normal subgroups, making the structure reusable for other exponential sums with SL2 or GL2 geometry."],"fun_headline_variants":["Non-abelian amplification saves c^{-1/700} on Kloosterman sums","Composite Kloosterman bilinear sums: c^{-1/700} saving","New bound for Kloosterman sums beats trivial for all moduli","From prime moduli to composite: Kloosterman sums improved","Kloosterman sums: c^{-1/12} saving for products of two primes"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The power saving enters only through the elementary count of six-fold products T^{h1}S...T^{h6}S equal to the identity in PSL2(Z/cZ) (Proposition 6.4); if that count were larger than c^{o(1)}(H2^2 + (H1H2)^2/c), the main theorem's saving would fail.","fun_headline_variants_meta":{"raw":{"variants":["Non-abelian amplification saves c^{-1/700} on Kloosterman sums","Composite Kloosterman bilinear sums: c^{-1/700} saving","New bound for Kloosterman sums beats trivial for all moduli","From prime moduli to composite: Kloosterman sums improved","Kloosterman sums: c^{-1/12} saving for products of two primes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000263,"raw_usage":{"total_tokens":1442,"prompt_tokens":751,"completion_tokens":691,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":495,"completion_tokens_details":{"reasoning_tokens":590}},"tokens_in":495,"tokens_out":691,"duration_ms":5927,"temperature":1.0,"reasoning_tokens":590,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T22:48:33.654312+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute, for c=p^2 and H1=H2=p, the exact number of 6-tuples (h1,...,h6) in [-p,p]^6 satisfying T^{h1}S...T^{h6}S ≡ ±I mod p^2 or ≡ ±I mod p. Proposition 6.4 predicts at most p^{2+o(1)} and p^{3+o(1)} solutions respectively; finding p^{2+δ} or p^{3+δ} for a fixed δ>0 would invalidate the saving.","supporting_citations":[],"review_version":1}