{"id":"558c072d-95b5-4578-b505-014788be3fc7","arxiv_id":"1908.02692","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A new point score, the moment, derived from metric space magnitude, allows approximating a data set's convex hull by keeping only high-moment points.","lead":"This paper uses metric space magnitude, a measure of effective number of points, to score each data point and keep only those with high scores when approximating the convex hull. The authors prove a magnitude-based inclusion-exclusion formula and show on synthetic Gaussian data that a few high-moment points recover most hull volume.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section 3.3's hull-volume guarantee has the wrong inequality direction: the error is an integral of a time-dependent maximum, while Algorithm 1 thresholds on the maximum of the integrals, which can understate the true error arbitrarily.","rationale":"The reader's verdict was already CONDITIONAL and identified Section 3.3 as the weakest point. This stress test confirms that diagnosis and sharpens it: the substitution of the maximum of integrals for the integral of the maximum is not merely unproved, it has the wrong direction, so the stated Algorithm 1 guarantee is unsupported. Lemma 1 and Proposition 2 appear correct and constitute a genuine contribution, and the experiments are suggestive but lack baselines and code. Because the mathematical core is sound but the main algorithmic guarantee fails as written, the paper should neither be accepted as is nor rejected outright; it needs a corrected or explicitly heuristic treatment of Section 3.3. This leaves the reader's CONDITIONAL verdict unchanged, with the technical basis made more precise.","tokens_in":10585,"tokens_out":9185,"duration_ms":101103,"concrete_test":"Run a numerical quadrature check on a configuration designed to make the argmax switch, e.g. X = {(−1,0),(1,0),(0,0.1),(0,10)} with P the two lowest-moment points. On a logarithmic grid t ∈ [10^{-3}, 10^3], solve ζ_{tX} w = 1 for each t, record m(t) = max_{y∈P} w_t(y)^2, and compute E = ∫_0^∞ e^{-t} m(t) dt and M = max_{y∈P} ∫_0^∞ e^{-t} w_t(y)^2 dt by quadrature. If E > M, take ε = |P| n |X| M (so the Section 3.3 criterion holds) and verify whether Vol(Conv(X\\P)) < Vol(Conv(X)) − ε. Repeat over 100 random 4- and 5-point configurations to confirm the gap is not a measure-zero artifact. This directly settles whether the max-moment substitution is valid as a hull-volume guarantee.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central algorithmic claim is the Section 3.3 chain ending in the criterion µ0(ˆx) ≤ ε/(|P| n |X|). The formal Proposition 2 gives, at each scale t, |tX\\P| ≥ |tX| − |P| w_t(x_{β_t})^2 (Eq. 12), where β_t is an argmax in P at scale t. Integrating against e^{-t} produces the true error term E = ∫_0^∞ e^{-t} max_{y∈P} w_t(y)^2 dt. The paper then substitutes µ0(ˆx) = max_{y∈P} ∫_0^∞ e^{-t} w_t(y)^2 dt for E. But for every t, max_y f_y(t) ≥ f_y(t) for each fixed y, so E ≥ µ0(ˆx), with equality only if the same point maximizes for almost every t. Thus bounding µ0(ˆx) does not bound E; the true error can be much larger whenever the maximizing point changes with t. This is not a missing technical hypothesis but a false implication in the derived guarantee. The further replacement of |tX| by Vol(Conv(X)) t^n/(n!Vol(B^n)) for all t is also unsupported: Theorem 6 is a t→∞ statement, while the integral runs over (0,∞), and for a finite set |tX| stays bounded (tending to 1 as t→0 and to N as t→∞), whereas the volume term grows without bound. Consequently the claimed guarantee that Algorithm 1 keeps the hull volume within ε is not established; the informal discussion cannot carry the weight of the algorithm's central promise.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the magnitude of finite subsets of Euclidean space and introduces the zeroth moment mu_0(x) = integral_0^infty e^{-t} w_t(x)^2 dt as a per-point geometric score. The main formal results are Lemma 1, expressing |X\\Y| in terms of the weight vector of X and a Schur complement, and Proposition 2, giving upper and lower bounds on |X\\Y| in terms of the extreme point weights. Based on an informal asymptotic discussion in Section 3.3, the paper proposes Algorithm 1, which removes low-moment points and claims to approximate the convex hull volume, and reports experiments on synthetic Gaussian data in dimensions 2-5.","tokens_in":10948,"tokens_out":7877,"duration_ms":82078,"significance":"If the algorithmic guarantee were established, the paper would offer a novel bridge between metric magnitude and computational geometry, and the zeroth moment would be a valuable intrinsic point-ordering tool. The rigorous parts, Lemma 1 and Proposition 2, appear correct and give an explicit accounting of how deleting points affects the magnitude of a finite Euclidean set. The experiments in Section 4.3 show an empirical phenomenon worth exploring. However, the theoretical argument underpinning the central convex-hull claim contains a reversed inequality and an invalid asymptotic substitution, so the advertised guarantee is not currently supported. The paper is honest in calling Section 3.3 a discussion, but the flaw is not a matter of missing technical hypotheses; it is a logically reversed bound that cannot be repaired by epsilon-chasing within the present algorithmic framework.","major_comments":[{"comment":"The substitution of mu_0(hat x) for the integral of the time-dependent maximum has the wrong inequality direction. Since wt(x_{beta_t})^2 >= wt(x)^2 for every fixed x and every t, integrating gives integral_0^infty e^{-t} wt(x_{beta_t})^2 dt >= mu_0(x) for each x and in particular for hat x. The actual error term in the lower bound for mu_0(|tX\\P|) is proportional to that integral, so bounding mu_0(hat x) by epsilon/(|P| n |X|) does not bound the true error; the true error can be arbitrarily larger when the maximizing point changes with t. Thus the claim that Algorithm 1 keeps the magnitude within a predefined threshold is not established.","section":"Section 3.3, Eq. (12) and following paragraph"},{"comment":"The replacement of |tX| by the volume-derived expression is unsupported. For finite X, Theorem 3 gives |tX| -> N as t -> infinity, so the finite-set magnitude is asymptotically constant, whereas Vol(Conv(X)) t^n / (n! Vol(B^n)) grows without bound. Theorem 6 is a statement about t Conv(X), not tX, and only in the t -> infinity limit. Because the integral defining the moment is significant over large t, the substitution used to derive the volume error bound is invalid, and the conclusion that Vol(Conv(X\\P)) is within epsilon of Vol(Conv(X)) does not follow.","section":"Section 3.3, replacement of |tX| by Vol(Conv(X)) t^n / (n! Vol(B^n))"},{"comment":"The criterion in Algorithm 1 is written as mu_0(x_i) <= epsilon/(d_i |X|), but d_i is never defined; Section 3.3 uses epsilon/(|P| n |X|) with n unexplained. More substantively, the algorithm thresholds a per-point moment, whereas Proposition 2's error bound involves the integral of the time-dependent maximum wt(x_{beta_t})^2 over the whole set P being removed. Since the per-point moment is always a lower bound for that error integral, the algorithm's threshold cannot certify the stated magnitude or volume error, even after correcting the notation.","section":"Section 4.1, Algorithm 1, Step 3 and the final paragraph of Section 3.3"}],"minor_comments":[{"comment":"The sentence claiming that a sequence {X_k} of independently sampled finite subsets of Conv(X) has monotonically increasing magnitudes |X_k| converging to |Conv(X)| is false without the subsets being nested; Theorems 4 and 5 give convergence for Hausdorff-convergent sequences, not monotonicity for arbitrary uniform samples.","section":"Section 3.3, first paragraph"},{"comment":"The displayed difference mu_0(x3) - mu_0(x2) is written with the same integrand wt(x3)^2 in both integrals; the second integrand should be wt(x2)^2.","section":"Corollary 1, proof"},{"comment":"The line 'proving w(x3) >= w(x3)' appears to be a typo; the intended conclusion is w(x3) >= w(x2).","section":"Proposition 1, proof"},{"comment":"The runtime analysis counts one matrix inversion as O(n^omega), but the algorithm computes mu_0(x) for every point, which requires solving the n x n system zeta_{tX} w = 1 for many scales t and then numerically integrating over t. The analysis omits the cost of this integration and the choice of quadrature, so the stated complexity of the preprocessing step is incomplete.","section":"Section 4.2, runtime analysis"},{"comment":"The symbol |X| is used for both the magnitude of X and the number of points in X, and Section 3.3 further proposes using |X| in place of Vol(Conv(X)). This triple overloading makes the inequalities in Section 3.3 and Algorithm 1 ambiguous.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is candid that Section 3.3 is informal, but the problem is not a missing epsilon: the displayed inequality direction is logically reversed, and the volume substitution is invalid for finite sets. The authors should either (a) replace the guarantee with an explicit conjecture and reframe the contribution as an empirical heuristic with supporting experiments, or (b) find an algorithmic criterion that actually controls the integral of the time-dependent maximum rather than the per-point moment. As it stands, the advertised convex-hull guarantee is not proved, though the rigorous parts (Lemma 1 and Proposition 2) are worth retaining."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nTwo things you should know about arXiv:1908.02692. First, the linear algebra is genuine: Lemma 1 gives an explicit Schur-complement formula for the magnitude of X\\Y in terms of the weight vector of X, and Proposition 2 gives a clean two-sided bound using the max and min of squared weights. That part is correct and is a real addition to the magnitude literature. Second, the advertised application — approximating convex hull volume with an ε guarantee — does not hold as stated.\n\nThe problem lives entirely in Section 3.3. The true error when removing P is E = ∫ e^{-t} max_{y∈P} w_t(y)^2 dt, because the worst point changes with t. The algorithm instead thresholds on μ0(·x) = max_y ∫ e^{-t} w_t(y)^2 dt, the maximum of the integrals. Since E ≥ μ0(·x), bounding μ0(·x) does not bound E; the inequality runs the wrong way. This is not a missing technical hypothesis. It is a false implication in the central promise of the paper.\n\nThe volume passage is equally unsupported. The paper replaces |tX| with Vol(Conv(X)) t^n / (n! Vol(B^n)) over the whole integration range, citing a t→∞ asymptotic for compact sets. But for finite X, |tX| is bounded between 1 and N, so the volume term is not a legitimate approximation inside the integral. That substitution is what converts magnitude differences into hull-volume differences, so the claimed guarantee is doubly unproven.\n\nCredit where it is due: the corrected inclusion-exclusion formula and the moment definition are genuinely new, and the three-point ordering results are fine. The numerical experiments, though, are thin — only Gaussian data, no baselines, no code or data — and they test the heuristic directly, not the stated guarantee.\n\nMy verdict: the formal core deserves a serious referee, but the hull approximation section needs to be rewritten, either as an honest heuristic with strong empirical validation or with a different, actually proven bound. As written, the paper is a solid lemma plus a speculative application. I would engage with it and cite the lemma, but not the algorithm.","headline":"Lemma 1 and the moment ordering are real contributions, but the convex-hull guarantee in Section 3.3 fails because it conflates an integral of a max with the max of integrals.","tokens_in":11422,"tokens_out":3436,"would_cite":true,"duration_ms":35328,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Metric-space magnitude assigns each point a \"moment,\" and the paper claims that deleting low-moment points alters the total magnitude little and leaves the convex-hull volume nearly unchanged.","keywords":["metric space magnitude","weight vector","zeroth moment","convex hull approximation","finite Euclidean point sets","similarity matrix","magnitude function"],"falsifier":"Take a finite set in $\\mathbb{R}^3$ consisting of many interior Gaussian points plus a sparse layer of extreme vertices, run Algorithm 1, and compare the volume of the returned hull to the claimed $\\epsilon$ bound; if a small-magnitude-error deletion causes a large volume drop, the Section 3.3 substitution is false. A cleaner check is to compute, for each $t$, the point $x_{\\beta_t}$ maximizing $w_t(x)^2$ and compare it with the point $\\hat{x}$ of maximal $\\mu_0$; if they differ on a set where the bound is tight, the replacement of the integrated worst-case by $\\mu_0(\\hat{x})$ fails.","tokens_in":10402,"feed_emoji":"📐","tokens_out":7650,"duration_ms":68237,"temperature":0.7,"pith_summary":"This paper tries to establish that the magnitude of a finite point set—an invariant often read as the effective number of points—can be decomposed into contributions of individual points. It defines for each point a zeroth moment $\\mu_0(x)=\\int_0^\\infty e^{-t} w_t(x)^2\\,dt$, built from the weight vector of the similarity matrix of the scaled set $tX$. The central claim is that points with small moment can be removed with only a small change in magnitude, and that repeatedly removing such points yields a small subset whose convex hull has almost the same volume as the original hull. This would matter because it gives an intrinsic, pre-geometric filter for convex hull computation and other point-set reductions.","feed_headline":"A point's moment predicts whether it can be dropped from a hull","feed_subtitle":"The metric-space magnitude of a point set yields an intrinsic ordering that filters points while preserving hull volume.","key_machinery":"The machinery is the similarity matrix $\\zeta_X(i,j)=\\exp(-\\|x_i-x_j\\|)$ for a finite set $X\\subset\\mathbb{R}^n$, the weight vector $w=\\zeta_X^{-1}\\mathbf{1}$, and the one-parameter family of scaled spaces $tX$. Lemma 1 expresses the magnitude of $X\\setminus Y$ as the original magnitude minus a quadratic form in the deleted weights, with the Schur complement $\\zeta_X/\\zeta_{X\\setminus Y}$ as the middle matrix; Proposition 2 turns that identity into a bound that only involves the number of deleted points and their largest squared weight. The zeroth moment $\\mu_0(x)=\\int_0^\\infty e^{-t}w_t(x)^2\\,dt$ removes the scale parameter and supplies the ordering used by Algorithm 1.","core_discovery":"The paper's central claim is that the per-point entries of the weight vector $w=\\zeta_X^{-1}\\mathbf{1}$ carry geometric information, and that this information can be integrated over scale to form an ordering. Lemma 1 gives the exact formula $|X\\setminus Y| = |X| - w_X|_Y^T (\\zeta_X/\\zeta_{X\\setminus Y}) w_X|_Y$, and Proposition 2 bounds the loss by $|X| \\ge |X\\setminus Y| \\ge |X| - N_Y \\max_{y\\in Y} w_X(y)^2$. The paper then defines the moment $\\mu_0(x)$ as the scale-averaged version of $w_t(x)^2$ and argues, in the informal discussion of Section 3.3, that conditioning on the largest moment in the removed set controls the change in magnitude and, through the volume asymptotic of Theorem 6, controls the change in $\\mathrm{Vol}(\\mathrm{Conv}(X\\setminus Y))$ relative to $\\mathrm{Vol}(\\mathrm{Conv}(X))$.","pith_inferences":["The volume guarantee in Section 3.3 is not fully proved, since it substitutes the volume asymptotics for the magnitude at all scales and replaces the worst-case point with the point of maximal moment; a numerical comparison against random subsampling would show whether the moment ordering is doing real work.","The moment of a point is computed from the full set's weight vector, so it changes when points are removed; Algorithm 1 does not recompute moments after deletion, and an iterative reweighting variant could behave differently.","Higher moments $\\mu_n(x)=\\int_0^\\infty t^n e^{-t}w_t(x)^2\\,dt$ suggest a multiscale description of each point, and could be tested as sharper filters for boundary detection."],"forward_implications":["If the moment ordering is reliable, any exact convex hull algorithm can be preceded by a filtering step that keeps only high-moment points, cutting the input size.","Proposition 2 gives a quantitative guarantee on the magnitude loss: a subset $Y$ with $\\max_{y\\in Y} w_X(y)^2 \\le \\epsilon/N_Y$ can be deleted while keeping $|X\\setminus Y|$ within $\\epsilon$ of $|X|$.","Because the moment is intrinsic to the metric structure rather than to the hull's faces, the method works before the hull is known and could be used for dynamic hull maintenance.","The experiments suggest the filter is strong: on Gaussian clouds in $\\mathbb{R}^2$ through $\\mathbb{R}^5$, the average number of points needed to reach 90% of the full hull volume ranges from about 4 to about 80."],"supporting_citations":[{"why":"Defines magnitude and the weight vector, and proves the similarity matrix is positive definite; these are the base definitions used throughout.","marker":"[5]"},{"why":"Supplies the volume asymptotic ratio used to connect magnitude to convex-hull volume in Section 3.3.","marker":"[2]"},{"why":"Gives the Hausdorff-continuity result used to treat a finite set as approximating its convex hull.","marker":"[8]"},{"why":"Provides the monotonicity of magnitude under inclusion, used to approximate the magnitude of a convex hull by finite subsets.","marker":"[6]"},{"why":"Supplies the inverse-of-a-submatrix identity used in the proof of Lemma 1.","marker":"[9]"},{"why":"Provides the baseline convex-hull algorithm whose worst-case complexity motivates the filtering approach.","marker":"[1]"}],"fun_headline_variants":["Moment of a point reveals its convex hull role","Metric magnitude orders points for hull approximation","Scale-averaged weight predicts hull membership","Hull shortcut via per-point geometric moment","Moment-based filter trims convex hulls"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument needs the unproved step that the volume of the scaled convex hull can be substituted for the magnitude at every scale $t$, and that the single highest-moment point can stand in for the worst-case point that changes with $t$.","fun_headline_variants_meta":{"raw":{"variants":["Moment of a point reveals its convex hull role","Metric magnitude orders points for hull approximation","Scale-averaged weight predicts hull membership","Hull shortcut via per-point geometric moment","Moment-based filter trims convex hulls"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000179,"raw_usage":{"total_tokens":1282,"prompt_tokens":907,"completion_tokens":375,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":523,"completion_tokens_details":{"reasoning_tokens":307}},"tokens_in":523,"tokens_out":375,"duration_ms":5011,"temperature":1.0,"reasoning_tokens":307,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:39:07.423140+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a finite set in $\\mathbb{R}^3$ consisting of many interior Gaussian points plus a sparse layer of extreme vertices, run Algorithm 1, and compare the volume of the returned hull to the claimed $\\epsilon$ bound; if a small-magnitude-error deletion causes a large volume drop, the Section 3.3 substitution is false. A cleaner check is to compute, for each $t$, the point $x_{\\beta_t}$ maximizing $w_t(x)^2$ and compare it with the point $\\hat{x}$ of maximal $\\mu_0$; if they differ on a set where the bound is tight, the replacement of the integrated worst-case by $\\mu_0(\\hat{x})$ fails.","supporting_citations":[{"cited_title":"The magnitude of metric spaces","cited_arxiv_id":null,"evidence_quote":"Defines magnitude and the weight vector, and proves the similarity matrix is positive definite; these are the base definitions used throughout."},{"cited_title":"On the magnitudes of compact sets in Euclidean spaces","cited_arxiv_id":null,"evidence_quote":"Supplies the volume asymptotic ratio used to connect magnitude to convex-hull volume in Section 3.3."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the Hausdorff-continuity result used to treat a finite set as approximating its convex hull."},{"cited_title":"The magnitude of a metric space: From category theory to geometric measure theory","cited_arxiv_id":null,"evidence_quote":"Provides the monotonicity of magnitude under inclusion, used to approximate the magnitude of a convex hull by finite subsets."},{"cited_title":"Ju \\'a rez Ruiz, R","cited_arxiv_id":null,"evidence_quote":"Supplies the inverse-of-a-submatrix identity used in the proof of Lemma 1."},{"cited_title":"Bradford Barber, David P","cited_arxiv_id":null,"evidence_quote":"Provides the baseline convex-hull algorithm whose worst-case complexity motivates the filtering approach."}],"review_version":1}