{"id":"e8d9aa50-e909-4df3-a5d7-e04cff7cbb69","arxiv_id":"1908.04658","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"For finite p, the periodic r-smooth fixed volume L_p discrepancy of the Fibonacci point set with b_n points is O(sqrt(log(b_n v)) / b_n^r); for p=∞ it is O(log(b_n v)/b_n^r).","lead":"This paper proves new upper bounds for the periodic smooth fixed volume discrepancy of Fibonacci point sets in the unit square, improving the known rate by a square root of a logarithm in finite L_p norms. The result shows that the boxes producing large discrepancy cannot be arbitrarily small, which matters for numerical integration and dispersion problems.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1.2 applies Lemma 2.2 outside its stated u ≤ 1/2 hypothesis when r=1 and v is large; the proof needs an extension of the dyadic-sum bound to u in (0,1]^d.","rationale":"The single most load-bearing step is the imported dyadic-sum bound: without Lemma 2.2, the high-frequency tail in (2.5) and (2.9) does not have the required decay, and the sqrt(log(b_n v)) improvement over the p=infinity bound would not follow. The reader correctly identified this lemma as the weak point. My check of the manuscript text shows an additional precision: the lemma is quoted with u in (0,1/2]^d, which is not satisfied for r=1 and v close to 1, so the proof as written has a condition gap in part of the claimed parameter range. I found no other issues: the Littlewood-Paley step (2.6), the Fourier characterization of the Fibonacci lattice (2.2), the dyadic block count (2.10), and the final series convergence are all sound, and the lower bounds cited are consistent with the stated sharpness claims. Because the gap is narrow and likely fixable by a short extension argument, the appropriate disposition is conditional acceptance rather than rejection or unqualified acceptance.","tokens_in":7225,"tokens_out":15345,"duration_ms":142297,"concrete_test":"Independently verify Lemma 2.2 for d=2 and u in (0,1]^d under the condition p_r(u) ≥ 2^{-t}. Concretely, provide a proof by splitting the index set into {j: 2^{s_j} u_j ≤ 1} and {j: 2^{s_j} u_j > 1}; for the second set, bound the product by (2^{s_j} u_j)^{-r/2} and sum over compositions of t to obtain O(t^{d-1})(2^t p_r(u))^{-r/2}. If the split proof succeeds, the r=1 gap closes and Theorem 1.2 is fully justified; if it fails, identify a numerical counterexample (e.g., d=2, r=1, u=(1,1), t in [1,1000]) to refute the extension.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central estimate (2.5) is derived from Lemma 2.2, which is stated for u in (0,1/2]^d. But Theorem 1.2 is claimed for all r in N and all v ≥ c/b_n. For r=1, a box B in [0,1)^2 with volume close to 1 has u_j up to 1, so Lemma 2.2 as stated does not apply in the regime where the high-frequency tail is estimated. The paper does not supply an alternative argument for u_j > 1/2. Thus the proof as written does not cover the r=1, large-v case of the central claim. The extension is likely routine: for u_j > 1/2 the min factor becomes (2^{s_j} u_j)^{-r/2}, and the corresponding contribution to the dyadic sum is of order (t+1)^{d-1} (2^t p_r(u))^{-r/2}, which still matches the claimed upper bound whenever p_r(u) ≥ 2^{-t}. But this step needs to be stated and proved in the manuscript.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper studies the periodic r-smooth fixed volume L_p discrepancy of the Fibonacci point sets in dimension two. The authors introduce a discrepancy functional obtained by averaging over shifts of a smooth hat function whose support is a box of prescribed volume v, and they prove (Theorem 1.2) that for every r in N and 1 <= p < infinity there are constants c,C > 0 such that for v >= c/b_n the discrepancy of the Fibonacci set F_n is at most C sqrt(log(b_n v)) / b_n^r. A weaker sup-norm bound with an additional logarithm is given in Theorem 1.3 for r >= 2. The proofs combine the Fibonacci lattice's exact Fourier cancellation on hyperbolic crosses with Littlewood-Paley decomposition and a dyadic sum estimate imported from the authors' earlier work [16].","tokens_in":7447,"tokens_out":35075,"duration_ms":314426,"significance":"If the theorem is correct, it gives a near-optimal, up to the square root of a logarithm, L_p fixed-volume discrepancy bound for Fibonacci sets, improving on the p = infinity log-rate and supporting the paper's interpretation that 'bad boxes' cannot be too small. The use of shift-averaging over hat functions is a genuine new ingredient, and the main derivation is explicit and largely self-contained. However, two load-bearing issues in the written proof must be repaired before the result is fully established: the stated hypothesis of Lemma 2.2 does not cover all boxes relevant to the r = 1 case, and the counting bound in (2.10) appears inconsistent with the subsequent series estimate.","major_comments":[{"comment":"Equation (2.5) is asserted for all r >= 1 and all t with v >= r^d 2^{-t+1}, but it is derived from Lemma 2.2, which is stated only for u in (0,1/2]^d. For r = 1, the boxes B subset [0,1)^2 allowed in Definition 1.2 may have side lengths u_j > 1/2, for example when the volume is close to 1, so Lemma 2.2 does not apply to those boxes. Since Theorem 1.2 includes r = 1 and the proof's final series requires only r > 1 - 1/p, the r = 1 case is genuinely in scope. The manuscript must supply an extension of Lemma 2.2 to u in (0,1]^d, or an alternative argument for coordinates u_j in [1/2,1], and must state the resulting condition on p_r(u); as written, the central upper bound for r = 1 is not proved.","section":"§2, equation (2.10) and the following estimate"},{"comment":"The counting bound in (2.10) is stated as #(rho(s) cap L(n)) <= C 2^{2t-t0}, but the next displayed inequality uses the factor 2^{2(t-t0)(1-1/p)}, which corresponds to # <= C 2^{t-t0} rather than # <= C 2^{2t-t0}. If (2.10) were used as printed, the exponent in the series would be 2(2t-t0)(1-1/p) - 2rt; for the central case r = 1, p = 2 this gives a term of order 2^{-t0} log(2^t v), whose sum over t diverges. The proof therefore requires either correcting (2.10) to the sharper per-block count # <= C 2^{t-t0}, with a justification from the Fibonacci lattice structure, or providing a two-regime argument. This issue is load-bearing for Theorem 1.2 and also affects the one-sentence proof of Theorem 1.3.","section":"§2"}],"minor_comments":[{"comment":"The derivation of (2.5) applies Lemma 2.2 with smoothness parameter 2r rather than r, because H_B^r(s)^2 contains the square of the min-factor; this is not stated and should be explained for readability.","section":"§2"},{"comment":"There is a typo: 'Parselval's identity' should read 'Parseval's identity'.","section":"Text near (2.9)"},{"comment":"The proof of Theorem 1.3 is only sketched; the claim that 'we need r > 1 for the last series' is not demonstrated in the text and should be expanded, especially in view of the counting issue in (2.10).","section":"Theorem 1.3 proof"}],"recommendation":"major_revision","confidential_remarks":"For the editor: the paper is within scope for a numerical analysis or discrepancy theory journal. The reliance on [16, Lemma 6.1] is acceptable in principle, but the extension needed for the r = 1 case should be proved in this paper rather than deferred to a citation. I would also ask the authors to expand the proof of Theorem 1.3 and to correct the counting estimate in (2.10) before the manuscript is reconsidered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear Colleague,\n\nThe short version: this paper proves a better upper bound for the periodic smooth fixed volume discrepancy of Fibonacci point sets, replacing the log(b_n v) from the p=∞ case by sqrt(log(b_n v)) for 1 ≤ p < ∞. The change of setting is a shift-average over hat functions instead of a supremum, and the proof is a clean Fourier / Littlewood–Paley argument. Theorem 1.2 is new and, as far as I can tell, correct in its intended scope.\n\nWhat I like: the paper carefully separates the periodic L_p setting from the earlier non-periodic and p=∞ results. It is honest about what is not known, especially that the bounds are not claimed to be optimal for every v. The dependence on Lemma 2.2 from the authors' prior work is standard practice, and the lower bounds cited from [17,18] are not used to prove the new upper bound.\n\nThe weak spot is a small but real gap. Lemma 2.2 is stated for u ∈ (0,1/2]^d, but the proof of Theorem 1.2 applies it for boxes with u_j up to 1 when r=1. For r=1 and large boxes the side parameters u_j can exceed 1/2, and the displayed estimate (2.5) does not follow from the lemma as stated. The stress-test note flags this exactly, and the concern lands. The extension is likely routine: for u_j > 1/2 the min factors saturate and the dyadic sum has the same order. But it needs to be written out. This is a minor revision, not a fatal flaw.\n\nThe citation pattern is acceptable. The paper leans on the authors' own prior work, but the new contribution is clearly separated, and the constants and Fourier estimates are explicit.\n\nThe paper is for people working in discrepancy theory, quasi-Monte Carlo, or hyperbolic cross approximation. It is a specialized result, but it strengthens a known connection between dispersion and discrepancy, and I expect it to be cited in later work on Fibonacci lattices.\n\nIf asked to referee: I would accept after a minor revision, asking for the dyadic-sum lemma to be extended to u ∈ (0,1]^d or for an explanation of why the r=1 case can be reduced. The main argument holds up; the gap is at the edge of the parameter range.\n\nBottom line: this deserves a serious referee. Accept with minor revision.","headline":"Genuinely new sqrt(log) bound for L_p fixed-volume discrepancy of Fibonacci sets, with a real but minor gap in the r=1 large-box case.","tokens_in":7958,"tokens_out":5493,"would_cite":true,"duration_ms":47834,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["11K38","65D30"],"pacs":[],"model":"deepseek-v4-flash","headline":"Fibonacci point sets reach sqrt(log) rate for periodic L_p fixed-volume discrepancy.","keywords":["fixed volume discrepancy","Fibonacci point set","periodic discrepancy","L_p discrepancy","Littlewood-Paley decomposition","hyperbolic cross","cubature formulas","discrepancy theory"],"falsifier":"Compute $\\sigma_u^r(t)$ numerically for $r=2$, dimension $d=2$, and several $u$ with $p_r(u)\\ge 2^{-t}$, and compare it with $C(\\log(2^{t+1}p_r(u)))^{-1}/(2^t p_r(u))^{r/2}$; if no fixed $C$ works for arbitrarily large $t$, Lemma 2.2 is false and the proof of Theorem 1.2 collapses. A cheaper check is to evaluate $\\tilde D_2^2(F_n,1/b_n)$ for moderate $n$ by direct Fourier summation: the observed rate should show the square-root-of-log factor, not the full log, for the theorem to be right.","tokens_in":7021,"feed_emoji":"📐","tokens_out":8970,"duration_ms":86304,"temperature":0.7,"pith_summary":"The paper studies a discrepancy-type characteristic, the r-smooth fixed-volume discrepancy, for the Fibonacci point set in the unit square. Its central result, Theorem 1.2, says that when test functions are averaged over translations of smooth 'hat' boxes of fixed volume v, the periodic L_p discrepancy for any 1 ≤ p < ∞ is bounded by C $\\sqrt$(log(b_n v)) / b_n^r whenever v ≥ c/b_n. This replaces the p = ∞ bound C log(b_n v)/b_n^r from earlier work, so averaging over shifts purchases a square root of a logarithm. Known lower bounds make the new rate sharp in a certain sense, and the result implies that 'bad boxes' for ordinary discrepancy cannot be arbitrarily small: their volume must be at least $b_n^{{-1+δ}}$ for some δ > 0. A sympathetic reader would care because it shows that a natural geometric averaging trick, rather than a better point set, is enough to close the gap to optimal behavior.","feed_headline":"Shift-averaging improves Fibonacci point discrepancy to sqrt(log)","feed_subtitle":"For integral norms, periodic fixed-volume discrepancy drops from log to sqrt(log) with the same point set.","key_machinery":"The central object is the periodic $r$-smooth fixed-volume $L_p$ discrepancy $\\tilde D_p^r(\\xi,v)$ from Definition 1.2, which replaces the supremum over box shifts by an $L_p$ average over $z$, computed against the periodized convolution hat kernel $\\tilde h_B^r$. The mechanism that carries the argument is a four-step Fourier estimate. First, the Fibonacci cubature formula is exactly zero on the hyperbolic cross $\\Gamma(\\gamma b_n)$, because the frequency lattice $L(n)=\\{k: k_1+b_{n-1}k_2\\equiv0\\pmod{b_n}\\}$ avoids that cross (Lemma 2.1). Second, the tail of the Fourier series is organized into dyadic shells $\\rho(s)$; Lemma 2.2, imported from [16], bounds the sum $\\sigma_u^r(t)$ of the shell coefficients. Third, the number of lattice points in a shell at level $t$ is at most $C 2^{2(t-t_0)}$ for $t$ beyond the critical level $t_0 \\asymp \\log b_n$ (inequality 2.10). Fourth, the Littlewood-Paley inequality (2.6) for $p \\in [2,\\infty)$ converts the $\\ell^2$ sum of shell $L_p$ norms into an $L_p$ bound, producing the square root of the logarithm; for $p<2$ the result follows from $L_2$ because the underlying measure has total mass one.","core_discovery":"For the Fibonacci point set $F_n \\subset [0,1)^2$ with $b_n$ points, the paper defines the periodic $r$-smooth fixed-volume $L_p$ discrepancy $\\tilde D_p^r(\\xi,v)$ by taking the $L_p$ norm over shifts $z$ of the difference between the integral and the sample average of the periodized smoothed box $\\tilde h_B^r(\\cdot - z)$ (Definition 1.2). Theorem 1.2 asserts that for every $r \\in \\mathbb N$ and $1 \\le p < \\infty$ there are constants $c,C>0$ with $\\tilde D_p^r(F_n,v) \\le C \\sqrt{\\log(b_n v)}/b_n^r$ for all $v \\ge c/b_n$. The proof writes the cubature error as a Fourier series over the nonzero lattice points $k \\in L(n)$; Lemma 2.1, the empty hyperbolic-cross property, kills all low frequencies, a dyadic bound imported from [16] controls the tail, and the Littlewood-Paley inequality assembles the dyadic blocks in $L_p$. The same method with the triangle inequality instead of Littlewood-Paley gives only the $p=\\infty$ bound $C\\log(b_n v)/b_n^r$, so the improvement is specifically a consequence of $L_p$ averaging over shifts.","pith_inferences":["The same mechanism plausibly extends to other point sets whose error vanishes on a hyperbolic cross of size comparable to the number of points, such as rank-1 lattices with good generating vectors, giving the same $\\sqrt{\\log m}$ improvement over $p=\\infty$; the only input needed is an empty-cross lemma like Lemma 2.1.","The Littlewood-Paley versus triangle inequality contrast suggests a general principle: any cubature formula whose error has a lacunary Fourier spectrum will show a logarithmic gap between pointwise and $L_p$ discrepancy, with the gap exactly one square root of a logarithm.","One could test whether the restriction to the periodic setting is essential; if the same averaging trick works for the non-periodic hat functions from Definition 1.1, the result would improve Theorem 1.1's $\\log(b_n v)$ bound as well, but the paper leaves this open.","A quantitative version might determine the optimal $\\delta$ in the statement that bad boxes have volume at least $b_n^{-1+\\delta}$ by matching the lower-bound constructions of [18] against the upper-bound proof."],"forward_implications":["For every $1\\le p<\\infty$, the periodic fixed-volume discrepancy of the Fibonacci set has the same order in the worst case over $v$ as the lower bound $m^{-r}(\\log m)^{(d-1)/2}$ from [18] in dimension two: the $\\sqrt{\\log m}$ factor is unavoidable up to constants.","Any box that realizes the discrepancy lower bound must have volume at least $b_n^{-1+\\delta}$ for some fixed $\\delta>0$; boxes of volume only $(\\log b_n)^c/b_n$ cannot be the worst ones.","The $p=\\infty$ rate $m^{-r}(\\log m)^{d-1}$, already known from [17], is sharp in the same sense, so the difference between the $p<\\infty$ and $p=\\infty$ rates is intrinsic to taking the supremum over shifts.","Via the connection established in [16] between fixed-volume discrepancy and dispersion, the improved discrepancy bound yields the corresponding upper bound for the dispersion of Fibonacci point sets in the unit square.","For $p<2$ no additional work is needed: the $L_p$ norm on the unit cube is dominated by the $L_2$ norm, so the $L_2$ bound automatically gives the $L_p$ bound."],"supporting_citations":[{"why":"Supplies the fixed-volume discrepancy definition, the earlier log-rate upper bound, and the imported Lemma 2.2 (its Lemma 6.1) that controls the dyadic shell sums.","marker":"[16]"},{"why":"Introduces the periodic p=∞ discrepancy and the lower bound (1.8) showing the log^{d-1} factor that the p<∞ result avoids.","marker":"[17]"},{"why":"Provides the lower bound (1.7) for optimized periodic discrepancy, establishing that the sqrt(log) factor in dimension two is necessary up to constants.","marker":"[18]"},{"why":"Supplies Lemma 2.1, the empty hyperbolic-cross property of the Fibonacci lattice, which makes the low-frequency part of the cubature error vanish.","marker":"[19]"},{"why":"Roth's classical lower bound for the case r=1, which the new result extends and compares against.","marker":"[10]"},{"why":"Schmidt's strengthened lower bounds for r=1, cited with [10] for the derivation of p>1 lower bounds.","marker":"[12]"}],"fun_headline_variants":["Shift-average hat functions cut Fibonacci discrepancy to sqrt(log)","L_p shift averaging beats sup over shifts for Fibonacci discrepancy","Fixed-volume discrepancy improves to sqrt(log) for Fibonacci sets","Shift-averaging yields sqrt(log) upper bound for fixed-volume discrepancy","Averaging hat-function shifts sharpens Fibonacci discrepancy to sqrt(log)"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof relies on an unproved bound from a previous paper saying that certain sums of Fourier coefficients over dyadic frequency shells decay at a specific rate; if that bound is false, or its condition is not met for the volumes and frequencies used, the central upper bound does not follow.","fun_headline_variants_meta":{"raw":{"variants":["Shift-average hat functions cut Fibonacci discrepancy to sqrt(log)","L_p shift averaging beats sup over shifts for Fibonacci discrepancy","Fixed-volume discrepancy improves to sqrt(log) for Fibonacci sets","Shift-averaging yields sqrt(log) upper bound for fixed-volume discrepancy","Averaging hat-function shifts sharpens Fibonacci discrepancy to sqrt(log)"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000762,"raw_usage":{"total_tokens":3399,"prompt_tokens":980,"completion_tokens":2419,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":596,"completion_tokens_details":{"reasoning_tokens":2334}},"tokens_in":596,"tokens_out":2419,"duration_ms":18957,"temperature":1.0,"reasoning_tokens":2334,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:35:11.171640+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute $\\sigma_u^r(t)$ numerically for $r=2$, dimension $d=2$, and several $u$ with $p_r(u)\\ge 2^{-t}$, and compare it with $C(\\log(2^{t+1}p_r(u)))^{-1}/(2^t p_r(u))^{r/2}$; if no fixed $C$ works for arbitrarily large $t$, Lemma 2.2 is false and the proof of Theorem 1.2 collapses. A cheaper check is to evaluate $\\tilde D_2^2(F_n,1/b_n)$ for moderate $n$ by direct Fourier summation: the observed rate should show the square-root-of-log factor, not the full log, for the theorem to be right.","supporting_citations":[{"cited_title":"Dispersion of the Fibonacci and the Frolov point sets","cited_arxiv_id":"1709.08158","evidence_quote":"Supplies the fixed-volume discrepancy definition, the earlier log-rate upper bound, and the imported Lemma 2.2 (its Lemma 6.1) that controls the dyadic shell sums."},{"cited_title":"Fixed volume discrepancy in the periodic case","cited_arxiv_id":"1710.11499","evidence_quote":"Introduces the periodic p=∞ discrepancy and the lower bound (1.8) showing the log^{d-1} factor that the p<∞ result avoids."},{"cited_title":"Remarks on numerical integration, discrepancy, and diaphony","cited_arxiv_id":"1711.07017","evidence_quote":"Provides the lower bound (1.7) for optimized periodic discrepancy, establishing that the sqrt(log) factor in dimension two is necessary up to constants."},{"cited_title":"Temlyakov, Multivariate approximation, Cambridge Universit y Press, 2018","cited_arxiv_id":null,"evidence_quote":"Supplies Lemma 2.1, the empty hyperbolic-cross property of the Fibonacci lattice, which makes the low-frequency part of the cubature error vanish."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Roth's classical lower bound for the case r=1, which the new result extends and compares against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Schmidt's strengthened lower bounds for r=1, cited with [10] for the derivation of p>1 lower bounds."}],"review_version":1}