{"id":"3450f62c-4c0b-418e-b49c-318e50de635c","arxiv_id":"2510.20511","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The Banach–Mazur distance between arbitrary convex bodies in R^n is at most n·polylog(n), and with 99% volume containment it is only polylog(n).","lead":"Two mathematicians show that any two convex bodies in high-dimensional space can be aligned so that each fits inside a scaled copy of the other with scale at most n times a logarithmic factor, nearly matching the known lower bound. If only 99% of the volume must be contained, the required scaling shrinks to a pure logarithmic factor.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Corollary 3.3's left-hand inequality is false as stated: it claims E F(B_{t/2}) ≤ E F(v_t), but Proposition 3.2 only gives E F((1/2)B_t) ≤ E F(v_t). This invalidates the written proof of Theorem 3.5's lower bound, though the argument is repairable.","rationale":"The reader's weakest assumption focused on Proposition 2.3's lower-eigenvalue bound being sketched. That is a legitimate gap, and I agree the paper should supply the full proof. However, my reading found a more definite, internal error in the main text: Corollary 3.3 states a two-sided comparison E F(B_{t/2}) ≤ E F(v_t) ≤ E F(2B_t), but the left inequality is false as written. The proof via Proposition 3.2 with r=1/4 only yields E F((1/2)B_t) ≤ E F(v_t). Since B_{t/2} is more spread than (1/2)B_t, the stated inequality would imply that a martingale with smaller quadratic variation dominates in convex order, which is impossible. A simple counterexample (A_t ≡ 1/2 Id, F(x)=x^2) confirms this. This matters because Theorem 3.5's lower bound—the new M-bound—relies on (50), which is exactly this false inequality. The good news is the proof is repairable: using the correct bound still gives the same constant 1/6 in (52), so the central claim survives. Nevertheless, the manuscript as submitted has a false statement in a core lemma, and the appendix's lower-eigenvalue proof remains a sketch with a likely typo in the final union bound. These are precisely the kind of issues that justify a conditional verdict: the main theorem is plausible and likely correct, but the written proof needs correction before acceptance. I also note the abstract promises applications to symplectic geometry and Dirichlet eigenvalues that do not appear in the supplied text; this is a presentation issue rather than a correctness issue.","tokens_in":19620,"tokens_out":21400,"duration_ms":158116,"concrete_test":"Counterexample: take n=1, set Σ_t ≡ 1/2Id so v_t = (1/2)B_t, and F(x)=x^2. Then E F(v_t)=t/4 and E F(B_{t/2})=t/2, violating the left inequality of Corollary 3.3. This proves the statement false as written. Then verify the repair: re-run the proof of Theorem 3.5 with the correct bound E||(1/2)B_t|| ≤ E||v_t||; check that the constants still give E||X|| ≥ c/(κ_n√log n) E||G||. Also complete the Appendix lower-eigenvalue proof and correct the final union bound to use P(τ≤T).","verdict_should_be":"UNCHANGED","load_bearing_attack":"In the proof of Theorem 3.5, inequality (50) uses Corollary 3.3's left bound E||B_{t/2}|| ≤ E||v_t||. But Corollary 3.3 is derived from Proposition 3.2 with r=1/4, which gives E F(v_t) ≥ E F((1/2)B_t), not E F(B_{t/2}). Since B_{t/2} has covariance (t/2)I while (1/2)B_t has covariance (t/4)I, the stated inequality is the reverse of what convex-order monotonicity yields: for convex F, larger covariance makes E F larger. Indeed, with n=1 and A_t ≡ 1/2Id, v_t=(1/2)B_t and E[v_t^2]=t/4, while E[B_{t/2}^2]=t/2; thus E[B_{t/2}^2] ≤ E[v_t^2] is false. The proof of Theorem 3.5 therefore contains a false step as written. The repair is straightforward: replace B_{t/2} by (1/2)B_t in (41) and in (50). Then (1/2)E||B_t|| ≤ E||v_t||, and combining with E||v_t|| ≤ E||a_t||+(1/3)E||B_t|| still yields (1/6)E||B_t|| ≤ E||a_t||. So the main conclusion of Theorem 3.5 is unaffected, but the manuscript must be corrected. This is a separate defect from the sketched lower-eigenvalue half of Proposition 2.3 (Appendix), where the final union bound appears to add P(τ≥T) rather than P(τ≤T); that gap also needs a full proof.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the non-symmetric Banach–Mazur distance. Its main theorem (Theorem 1.1) asserts that any two convex bodies in R^n are at Banach–Mazur distance at most C n log^α n, improving Rudelson's n^{4/3} bound and matching Gluskin's lower bound up to polylog. The proof is via a random isotropic position (Theorem 1.2) and a new two-sided comparison, in the gauge order, between isotropic log-concave measures and Gaussian measures (Theorem 3.5 / Theorem 1.3). The right-hand comparison is attributed to Eldan and Lehec; the left-hand comparison is the paper's main new contribution, proved with stochastic localization and a covariance-process bound (Proposition 2.3). From the comparison and Chevet's inequality the authors derive M and M^* bounds in the isotropic position, an M M^* product O(polylog), a 99%-containment result for two isotropic bodies, and a polylog bound for the partial-containment distance d_PC. Applications to symplectic geometry and Dirichlet eigenvalues are also mentioned in the abstract, though the main body focuses on the geometric statements.","tokens_in":20094,"tokens_out":18999,"duration_ms":146024,"significance":"If the proof is repaired, this is a substantial advance: it essentially determines the diameter of the non-symmetric Banach–Mazur compactum up to polylogarithmic factors, matching the centrally-symmetric behavior, and it establishes a polylogarithmic partial-containment distance between arbitrary convex bodies. The new M-bound in the isotropic position is a natural and useful complement to E. Milman's M^*-bound, and the derivation of the geometric consequences from Theorem 3.5 and Chevet's inequality is clean and well organized. The stochastic-localization framework is presented in detail, with the main analytic proof being largely self-contained. The paper's central claims do not rely on fitted parameters or ad hoc entities; the new left-hand comparison is a genuine mathematical contribution. However, two load-bearing points need correction and completion before the argument is fully reliable.","major_comments":[{"comment":"The left inequality in (41), EF(B_{t/2}) ≤ EF(v_t), is false as stated. Proposition 3.2 with r=1/4 yields EF(v_t) ≥ EF((1/2)B_t), not EF(B_{t/2}). The two processes have different covariances: B_{t/2} has covariance (t/2)I while (1/2)B_t has covariance (t/4)I, so for convex F the expectation with the larger covariance dominates; the asserted direction is the reverse. For n=1, A_t≡Id/2 and v_t=(1/2)B_t, we have E[B_{t/2}^2]=t/2 > t/4=E[v_t^2]. This false step is used at (50) in the proof of Theorem 3.5. The repair is local: replace B_{t/2} by (1/2)B_t in (41), and use (1/2)E||B_t|| ≤ E||v_t|| at (50). The subsequent derivation of (52) still goes through with an inconsequential change of the universal constant. The statement and proof must be corrected.","section":"§3, Corollary 3.3 and proof of Theorem 3.5, Eqs. (41) and (50)"},{"comment":"The lower bound λ_min(A_t) ≥ 1/2 is load-bearing: Corollary 3.3's 'Moreover' part uses it to identify v_t with a_t, which is the basis for the new M-bound. The appendix explicitly says 'the argument is nearly identical, we only provide a sketch', and the final union bound is wrong as written. After bounding P(∃t≤T: λ_min(A_{t∧τ})≤1/2), the passage to λ_min(A_t) adds P(τ≥T); the correct term is P(τ≤T), because if no upper violation occurs before T, any lower violation is already a lower violation for the stopped process. Since Lemma A.4 only makes P(τ≤T) exponentially small, the displayed estimate ≤2exp(-c1/T) does not follow from the preceding lines. A complete proof of the lower-eigenvalue half is needed.","section":"Appendix, Proposition 2.3, lower-eigenvalue half"}],"minor_comments":[{"comment":"The probability is written as 'at least 1−Cexp(c/T)'; the exponent should be negative, i.e., '1−Cexp(−c/T)'.","section":"Proposition 2.3 statement"},{"comment":"The sentence 'by Theorem 1.1, for λ=C√n log^2 n' should refer to Theorem 1.2.","section":"Proof of Theorem 1.1"},{"comment":"In the chain 'E||G||_K ≤ Cκ_n √log n ∥X∥_K', the last term should be E||X||_K; as written it compares the expectation to a random variable.","section":"Lemma 4.1 proof"},{"comment":"The identity ∑_i ⟨H_i^2 θ,θ⟩ = Tr[H_θ^2] is not self-evident for arbitrary symmetric matrices. It holds here because the H_i are coordinate slices of the third-moment tensor and the full symmetry of E[X_i X_j X_k] makes the two expressions equal. Adding one explanatory line would improve clarity.","section":"Lemma A.2 proof"}],"recommendation":"major_revision","confidential_remarks":"The two major issues are local and repairable: the Corollary 3.3 misstatement can be fixed by replacing B_{t/2} with (1/2)B_t without changing the main conclusion, and the lower-eigenvalue half of Proposition 2.3 needs a fuller proof and a corrected union bound. The central idea and the overall structure appear sound; I do not see a reason to doubt the main theorem once these fixes are made."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The main theorem is real and the paper deserves a serious referee. It improves Rudelson's n^{4/3} to n log^α n for the non-symmetric Banach–Mazur distance, which is optimal up to polylog, and it proves a genuinely new polylog bound for partial containment. The left-hand gauge comparison between isotropic log-concave vectors and Gaussians is the new mathematical content; the right-hand side is Eldan–Lehec and the M* bound is E. Milman, properly credited. The derivation of Theorems 1.1–1.6 from Theorem 3.5 via Chevet and the orthogonalization step is clean.\n\nThere are two errors in the write-up, both non-load-bearing. First, Corollary 3.3 states E F(B_{t/2}) ≤ E F(v_t), but Proposition 3.2 with r=1/4 gives E F((1/2)B_t) ≤ E F(v_t). Since B_{t/2} has covariance (t/2)I and (1/2)B_t has (t/4)I, the stated inequality goes the wrong way for convex F. The fix is to replace B_{t/2} by (1/2)B_t in (41) and (50); then (1/2)E||B_t|| ≤ E||v_t||, and the rest of Theorem 3.5 goes through with the same constant. Second, in the lower-eigenvalue half of Proposition 2.3 in the appendix, the final union bound adds P(τ≥T) where the correct term is P(τ≤T). Lemma A.4 bounds P(τ≤T), so the intended argument is fine once the sign is fixed. The reader also flagged that this lower half is only sketched ('argument is nearly identical'); that's fair, and a full version should be in the paper. The abstract also promises applications to symplectic geometry and Dirichlet eigenvalues that do not appear in the supplied full text; the authors should either include them or remove the promise.\n\nNone of this touches the central claims. The proof strategy is coherent, the constants are implicit but standard, and the paper is honest about what comes from prior work and what is new. It belongs in a good journal after a revision that fixes the two slips and completes the lower-eigenvalue proof.","headline":"Real near-optimal Banach–Mazur result with two small repairable slips; send to review.","tokens_in":20569,"tokens_out":3588,"would_cite":true,"duration_ms":29376,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["52A20","52A21"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proves that any two n-dimensional convex bodies, with no symmetry assumption, have Banach-Mazur distance at most C n (log n)^α, optimal up to polylog factors.","keywords":["Banach-Mazur distance","convex bodies","isotropic position","stochastic localization","gauge order","partial containment distance","log-concave measures","M-bound"],"falsifier":"Run the stochastic localization SDE numerically from an isotropic log-concave measure in, say, n=1000 dimensions and watch the smallest eigenvalue of A_t: if it dips below 1/2 before time c/(κ_n² log n), the high-probability event that a_t = v_t is false, and the paper's lower Gaussian comparison loses its foundation.","tokens_in":19530,"feed_emoji":"📐","tokens_out":8127,"duration_ms":70853,"temperature":0.7,"pith_summary":"The paper shows that every pair of n-dimensional convex bodies — no symmetry assumed — are within a factor of C n (log n)^α of each other in the Banach-Mazur sense: after an affine change of coordinates and a translation, each scaled body contains the other. This improves the previous n^{4/3} guarantee and is optimal up to a polylogarithmic factor, matching what was already known for symmetric bodies. The proof works in the isotropic position and is driven by a new two-sided comparison between log-concave measures and Gaussian measures in the gauge order, i.e., uniformly over all norm-like functions. As a byproduct, the paper proves that a relaxed distance, where each body must contain only 99% of the other, grows only polylogarithmically, showing that the large Banach-Mazur distance is carried by small-volume outliers.","feed_headline":"Banach-Mazur distance shrinks to n (log n)^α","feed_subtitle":"Improves the old n^4/3 guarantee to optimal up to log factors, with no symmetry assumption needed.","key_machinery":"Stochastic localization: starting from the uniform measure on a convex body, tilt the density by exp(θ·x - t|x|²/2) where θ_t is driven by Brownian motion, and track the barycenter a_t and covariance A_t of the tilted measure. The barycenter is a martingale with differential A_t dB_t, so the original random vector equals ∫ A_t dB_t in law. The argument hinges on a competition between the drift -A_t² dt and noise from third-moment tensors: for time T of order 1/(κ_n² log n), A_t stays between Id/2 and 2Id with high probability. Clipping A_t to that window yields a martingale v_t that is provably sandwiched between Gaussian expectations via two convexity arguments (one using an endpoint Gaussi","core_discovery":"The paper's central claim is that the non-symmetric Banach-Mazur diameter of the space of n-dimensional convex bodies is n^{1+o(1)}: for any two compact convex sets K1,K2 with nonempty interior, d_BM(K1,K2) ≤ C n log^α n, with universal constants and α ≤ 4. Because there exist pairs of symmetric bodies with distance at least c n, this is best possible up to the polylog factor. The proof follows from a new 'isotropic M-bound': for an isotropic log-concave random vector X (the uniform measure on a body in isotropic position, meaning zero mean and identity covariance) and a standard Gaussian G, the expectations E||X|| and E||G|| agree up to the factor κ_n √log n, uniformly over all gauge functi","pith_inferences":["Editorial: the exponent α≤4 is not claimed optimal; since the polylog enters through the current bound ψ_n ≤ C√log n, an improvement of that isoperimetric estimate would automatically improve the distances we can certify.","Editorial: the d_BM ~ n versus d_PC ~ log n gap suggests a clean dichotomy — bulk geometry is polylog-close, and the full-containment metric is inflated only by narrow tails; this could be tested by comparing random rotations of highly anisotropic bodies.","Editorial: the proof currently rests on a sketched lower-eigenvalue bound for the covariance process; turning that sketch into a complete argument is an immediate verification task for anyone using the M-bound.","Editorial: a natural continuation is to quantify how the partial-containment dilation grows as the required fraction β approaches 1; the paper's Markov argument gives β=99/100 only for a fixed universal constant, and a trade-off curve is not written down."],"forward_implications":["The non-symmetric Banach-Mazur diameter is now known to within a polylog factor; the hardest pair of convex bodies is only n (log n)^α apart, matching the symmetric-case order n.","If the isoperimetric constant ψ_n (the KLS parameter) is eventually proved O(1), the same comparison sharpens to a √log n factor between log-concave and Gaussian expectations, and examples in the paper show that is optimal.","The isotropic position, after a random rotation, simultaneously realizes near-optimal containment for every pair of bodies, so it acts as an M-position without symmetry assumptions.","The partial-containment distance is polylogarithmic: any two isotropic bodies contain 99% of each other after scaling by about log^α n, showing the large Banach-Mazur distance is carried by small-volume outliers.","The new comparison has downstream uses in linear symplectic geometry and in bounding the first Dirichlet eigenvalue of the Laplacian."],"fun_headline_variants":["Banach-Mazur distance for all convex bodies: optimal n (log n)^α","99% containment: distance between convex bodies is only log^α n","Non-symmetric convex bodies: distance n^(1+o(1)) without symmetry assumption","Isotropic position yields polylog-optimal Banach-Mazur bound","Optimal up to polylog: distance between any two convex bodies"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The proof's engine is the claim that during stochastic localization the covariance matrix of the tilted measure remains between half and twice the identity for a short time; the upper half is proved in the appendix, while the lower half is only sketched as 'nearly identical', and if that lower bound fails the key equality between the barycenter and the clipped martingale collapses.","fun_headline_variants_meta":{"raw":{"variants":["Banach-Mazur distance for all convex bodies: optimal n (log n)^α","99% containment: distance between convex bodies is only log^α n","Non-symmetric convex bodies: distance n^(1+o(1)) without symmetry assumption","Isotropic position yields polylog-optimal Banach-Mazur bound","Optimal up to polylog: distance between any two convex bodies"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001233,"raw_usage":{"total_tokens":4983,"prompt_tokens":910,"completion_tokens":4073,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":654,"completion_tokens_details":{"reasoning_tokens":3968}},"tokens_in":654,"tokens_out":4073,"duration_ms":27125,"temperature":1.0,"reasoning_tokens":3968,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T08:25:08.463409+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the stochastic localization SDE numerically from an isotropic log-concave measure in, say, n=1000 dimensions and watch the smallest eigenvalue of A_t: if it dips below 1/2 before time c/(κ_n² log n), the high-probability event that a_t = v_t is false, and the paper's lower Gaussian comparison loses its foundation.","supporting_citations":[],"review_version":1}