{"id":"0e7abad6-fe7d-4a8b-a163-d266f3414542","arxiv_id":"2502.06716","paper_version":1,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"The Kolmogorov n-width of W_1^1[0,1] in L_q[0,1], 2<q<∞, is of order n^{-1/2} log n, settling the last open case of the classical width classification.","lead":"This paper proves the exact rate at which functions with bounded variation can be approximated by n-dimensional subspaces in the L_q norm, for every q greater than 2: roughly one over the square root of n, times a logarithm. It settles the last open case of a classical classification problem in approximation theory.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified: proof is internally consistent; the only gap (regularity of eta_t) is closed by continuity of metric projection in L_q.","rationale":"The reader's verdict is ACCEPT. I independently checked the main estimates. The proof of the Lemma is sound: the orthogonality computation (10), isotropy (11), I2 bound via ∫||H||^2, and I3 bound all match. The only possible gap is the regularity of η_t, but the metric projection in L_q is continuous because L_q is uniformly convex, so the gap is fillable without altering constants. Therefore no significant objection; verdict UNCHANGED.","tokens_in":4998,"tokens_out":25464,"duration_ms":214253,"concrete_test":"Add to the proof of the Theorem, immediately after choosing Q_n, the sentence: 'For each t let η_t denote the metric projection of χ_t onto Q_n; by uniform convexity of L_q this projection is single-valued and continuous in t.' Then re-verify that sup_t ||χ_t−η_t||_{L_q}<δ n^{-1/2}\\log n and that (2) is violated with the same δ. If this step goes through without changing any estimate, the selection concern is fully resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"I examined the lower-bound proof and the Lemma on which it rests. The key steps check out: (5) and (6) are valid for q>2 by the Marcinkiewicz-Paley square-function estimate and orthogonality; the auxiliary functions Z_{k,j} are orthogonal with ∫Z^2=2^k, giving isotropy and hence (11); the estimates I1≍log n, I2≲δ log n, I3^{1/q'}≤C n^γ are consistent once k1=⌊γq log2 n⌋. The choice of δ makes the numerator ∼log n, so (3) follows. In the proof of the Theorem, the only unstated point is that the family η_t can be chosen so that t↦η_t is measurable (indeed continuous). Since χ_t is continuous into L_q and L_q is uniformly convex for 1<q<∞, the metric projection onto the fixed finite-dimensional subspace Q_n is single-valued and continuous; taking η_t=P_{Q_n}χ_t gives the required sup bound. Thus the omitted selection argument is a harmless exposition gap, not a threat to the central claim. I found no circularity, no miscomputed constant, and no invented entity.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proves that for every q in (2,∞) the Kolmogorov n-width of the univariate Sobolev class W_1^1 in L_q satisfies d_n(W_1^1,L_q) ≍ n^{-1/2} log n. The upper bound is cited from Kulanin, and the main contribution is a matching lower bound. The lower bound is obtained through a lemma: for any family {η_t} lying in an n-dimensional subspace, either the average squared L2 distance to the step functions χ_t is at least δ^2 n^{-1} log^2 n, or the average Lq distance is at least n^{-γ}. The lemma is proved via weighted Haar coefficients, a duality/Hölder step, and isotropy of specially constructed functions Z_{k,j}. The proof tracks all constants, with the levels k0 and k1 chosen as floor functions of logarithmic scales.","tokens_in":5197,"tokens_out":9106,"duration_ms":71288,"significance":"If correct, this result completes the study of orders of decay of Kolmogorov widths for classical univariate Sobolev classes of integer smoothness, closing the last open parameter case r=p=1, 2<q<∞. The proof is transparent and largely self-contained, using standard tools (Marcinkiewicz–Paley, Hölder duality, metric projection) and tracking constants explicitly. The lower-bound method, which combines dyadic decomposition with simultaneous control of all levels via duality and averaging, is elegant and likely to be influential. The manuscript also gives credit to prior work and clearly identifies the open gap it closes.","major_comments":[{"comment":"The passage asserting the existence of a family {η_t} with sup_t ‖χ_t−η_t‖_{L_q} < δ n^{-1/2} log n is not justified by the definition of the n-width alone: the width gives for each t a near-optimal approximation, but the Lemma requires a single measurable family in t, and without measurability the integrals in (2) and (3) are undefined. This gap is load-bearing for the lower-bound proof. It is fixable: since the map t↦χ_t is continuous into L_q and L_q is uniformly convex for q∈(2,∞), the metric projection P_{Q_n}χ_t onto the fixed finite-dimensional subspace Q_n is single-valued and continuous; taking η_t=P_{Q_n}χ_t yields the required measurable family. Please add this argument or an appropriate measurable-selection reference.","section":"Proof of Theorem (paragraph following the Lemma)"}],"minor_comments":[{"comment":"The definition of A≍B contains a corrupted symbol: 'A /greaterorsimilarB' should read 'A ≳ B' (or 'A ≥ cB').","section":"Introduction"},{"comment":"In the definition of Z_{k,j}(t), the variable in the first line is written as 'x' but should be 't'; the support condition should be t ∈ [(j−1)2^{−k}, j2^{−k}].","section":"Proof of Lemma (definition of Z_{k,j})"},{"comment":"The final paragraph contains typos: 'is δ is small enough' should be 'if δ is small enough', and 'Denonimator' should be 'denominator'.","section":"Proof of Lemma (final paragraph)"}],"recommendation":"major_revision","confidential_remarks":"The paper closes the last open case in a classical problem, and the proof is sound apart from the measurable-selection gap described in the major comment. The self-citation [Mal21] is contextual and not load-bearing. The manuscript would likely be acceptable after the short addition requested in the major comment; the revision is minor in spirit but touches a load-bearing point, so I recommend major_revision under the stated rubric."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know upfront: this paper resolves the last open parameter case in the classical table of Kolmogorov widths for univariate Sobolev classes. For r=p=1 and 2<q<∞, the sharp order is n^{-1/2} log n. The question had been open since Kulanin's 1983 bounds, which left a logarithmic gap, and KMR18 only removed the epsilon. This is a genuine theorem.\n\nWhat's actually new is the lower-bound proof. The dyadic setup is Kulanin's, but the author controls all dyadic levels simultaneously, using tailored duality and averaging in the spirit of Gluskin. The auxiliary functions Z_{k,j} have an isotropy property that makes the duality estimate sharp, and the I2 term is controlled by projecting Z onto the n-dimensional subspace V_n. I checked the orthogonality, the computation of I1, the bound I2 ≲ δ log n, and the choice k1 = floor(γq log n); they are consistent. There is no circularity and no fitting of constants.\n\nThe proof does have one unstated step. In the Theorem proof, the author passes from the existence of a subspace Q_n with sup_t inf_{η∈Q_n} ||χ_t−η||_{L_q} below the threshold to a family {η_t} with the same uniform bound, and then integrates in t. This requires a measurable selection. The gap is harmless: L_q is uniformly convex for 1<q<∞, the metric projection onto Q_n is single-valued and continuous, and χ_t is continuous into L_q, so taking η_t = P_{Q_n} χ_t works. It is an exposition omission, not a flaw.\n\nNo other substantive objections. The upper bound is cited from Kulanin rather than reproved; acceptable. The constants are existential; fine. The self-citation to the author's Besov paper is contextual, not load-bearing. The paper is short and readable.\n\nThis is for approximation theorists working on widths and Sobolev embeddings. Outside that subfield the impact is modest, but inside it this closes a long-standing case. It deserves a serious referee. Assuming the referee asks for the selection argument to be written out, it should be published.\n\nSend it to peer review.","headline":"Resolves the last open case in the classical Sobolev-width table with a clean, checkable proof; worth refereeing.","tokens_in":5808,"tokens_out":3239,"would_cite":true,"duration_ms":26282,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["41A46","46E35"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proves sharp two-sided estimates for the Kolmogorov n-width of the Sobolev class W_1^1[0,1] in L_q[0,1] for 2<q<∞, establishing that d_n(W_1^1,L_q) ≍ n^{-1/2} log n and completing the classical table of width orders for…","keywords":["Kolmogorov width","Sobolev class W_1^1","L_q approximation","n-widths","dyadic decomposition","Haar system","logarithmic asymptotics","univariate Sobolev classes"],"falsifier":"Fix q=3 and compute or bound from below the quantity d_n($W_1^{1}$[0,1], L_3[0,1]) for increasing n using a discretized linear programming formulation; if $n^{{1/2}}$ d_n / log n tends to 0, the theorem's lower bound fails. Alternatively, for infinitely many n exhibit an n-dimensional subspace whose elements approximate every step function χ_t within o($n^{{-1/2}}$ log n) in L_q norm, which would directly contradict the lemma.","tokens_in":4744,"feed_emoji":"📐","tokens_out":5256,"duration_ms":46455,"temperature":0.7,"pith_summary":"The paper proves that the Kolmogorov n-width of the univariate Sobolev class $W_1^{1}$[0,1] in L_q[0,1] has exact order $n^{{-1/2}}$ log n for every 2<q<∞. The upper bound was already known; the paper supplies the matching lower bound. This closes the last unresolved parameter case in the classical study of sharp order estimates for widths of integer-smoothness Sobolev classes. The proof controls all dyadic frequency levels at once through a duality estimate and an averaging argument, avoiding the earlier loss of a logarithmic factor.","feed_headline":"Kolmogorov width of W_1^1 pinned to n^{-1/2} log n","feed_subtitle":"Sharp two-sided estimate for 2<q<∞ closes the last open case in the classical Sobolev width table.","key_machinery":"The key object is the family of step functions χ_t and their Fourier–Haar coefficients X_{k,j}(t), together with an auxiliary family Z_{k,j}(t) of functions of t that are mutually orthogonal and satisfy ∫ Z_{k,j}(t)^2 dt = 2^k. These Z functions serve as a dual test family: the weighted inner product of X(t) with Z(t), averaged over t, equals (k1-k0+1)/a, a quantity of order log n. The Marcinkiewicz–Paley inequality transfers L_q norms of functions to weighted ℓ_q norms of Haar coefficients, and the isotropy identity ∫⟨v,Z(t)⟩^2 dt = ‖v‖^2 makes the projection of Z onto an n-dimensional subspace have integrated squared norm exactly n, bounding the cross term I2 by about δ log n. The dichotomy in the lemma follows by comparing these estimates.","core_discovery":"The central claim is the two-sided estimate c1(q) $n^{{-1/2}}$ log n ≤ d_n($W_1^{1}$[0,1], L_q[0,1]) ≤ c2(q) $n^{{-1/2}}$ log n for all q>2 and n≥2. The paper contributes the missing lower bound, obtained by showing that any n-dimensional subspace that approximates every step function χ_t (the sign-changing function at t) to within less than δ $n^{{-1/2}}$ log n in L_q must fail to control the Haar coefficients of these functions in a weighted ℓ_q sense. A specially designed family Z_{k,j}(t) of functions of t provides a duality test: its inner product with the Haar coefficient vectors of χ_t is large, of order log n, while the error's L_q norm and the finite-dimensionality of the subspace impose upper bounds of order $n^{{-1/2}}$ and $n^{{γ}}$. The combination forces the lower bound and completes the parameter range r=p=1, 2<q<∞.","pith_inferences":["Beyond the paper's claims, the simultaneous dyadic-control mechanism may extend to lower bounds for average Kolmogorov widths or for related Besov classes, whose Haar-coefficient structure is similar.","The argument suggests a general principle: when a one-parameter family of targets must be approximated by a finite-dimensional subspace, a logarithmic factor arises precisely because no single frequency band can absorb all the error.","A routine but necessary addition to the written proof is a measurable selection argument ensuring that the approximating functions η_t can be chosen to depend measurably on t; without it, the averaging over t in the lemma is not fully justified."],"forward_implications":["The sharp order for r=p=1, 2<q<∞ now matches the pattern expected from the classical table, with exactly a logarithmic factor multiplying n^{-1/2}.","Kulanin's lower bound with log^{1/2-ε} n is improved to a full log n, closing the logarithmic gap that remained after earlier work.","Belinsky's result for subspaces generated by n harmonics has the same order as the unrestricted n-width, so allowing arbitrary approximating subspaces does not improve the rate.","The lemma establishes a quantitative dichotomy: any n-dimensional family approximating the step functions either has L_2-average error at least δ n^{-1/2} log n or L_q error at least n^{-γ} for γ>1/q."],"supporting_citations":[{"why":"Supplies the dyadic decomposition and the already-known upper bound c2(q) n^{-1/2} log n that the theorem adopts.","marker":"[Kul83]"},{"why":"Provides the duality and averaging technique used in inequalities (7), (8), and (11) to control all dyadic levels simultaneously.","marker":"[Gls81]"},{"why":"Records the previous best lower bound that removed ε from Kulanin's exponent but left the logarithmic gap, which this paper closes.","marker":"[KMR18]"},{"why":"Establishes the same order for subspaces generated by n harmonics, showing the trigonometric width is sharp and framing the comparison for arbitrary subspaces.","marker":"[Bel84]"},{"why":"Settled the sharp orders for q>max{2,p} and serves as the historical baseline for the parameter case completed here.","marker":"[Kas77]"},{"why":"Studied closely related Besov classes B^1_{1,θ} and obtained sharp estimates there, but its method did not directly give the W_1^1 bound.","marker":"[Mal21]"}],"fun_headline_variants":["W_1^1 width solved: n^{-1/2} log n","Sharp estimate for W_1^1 Kolmogorov width","Missing lower bound for W_1^1 width filled","n^{-1/2} log n order for W_1^1 widths proven","Complete order for W_1^1 width found"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof assumes that, given an approximation of each step function χ_t by some n-dimensional subspace, the approximating functions η_t can be chosen so that they depend measurably on t, because the central estimate averages over all t in [0,1].","fun_headline_variants_meta":{"raw":{"variants":["W_1^1 width solved: n^{-1/2} log n","Sharp estimate for W_1^1 Kolmogorov width","Missing lower bound for W_1^1 width filled","n^{-1/2} log n order for W_1^1 widths proven","Complete order for W_1^1 width found"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0004,"raw_usage":{"total_tokens":2019,"prompt_tokens":808,"completion_tokens":1211,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":424,"completion_tokens_details":{"reasoning_tokens":1119}},"tokens_in":424,"tokens_out":1211,"duration_ms":9039,"temperature":1.0,"reasoning_tokens":1119,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T14:33:41.041472+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fix q=3 and compute or bound from below the quantity d_n($W_1^{1}$[0,1], L_3[0,1]) for increasing n using a discretized linear programming formulation; if $n^{{1/2}}$ d_n / log n tends to 0, the theorem's lower bound fails. Alternatively, for infinitely many n exhibit an n-dimensional subspace whose elements approximate every step function χ_t within o($n^{{-1/2}}$ log n) in L_q norm, which would directly contradict the lemma.","supporting_citations":[],"review_version":1}