{"id":"ba3303d6-1974-4abf-bb9d-69d5eade2558","arxiv_id":"2511.10483","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A max-min angle between convex cones defines a valid distance, equals the Pompeiu-Hausdorff distance on cone-sphere intersections, and is computed analytically for subspaces/revolution cones and approximately for polyhedral cones.","lead":"This paper introduces a new way to measure how different two convex cones are: look at the largest possible smallest angle between their directions, then symmetrize. The measure is shown to match a classical set-distance, and comes with formulas for special cones and a cutting-plane algorithm tested on few-shot image classification.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 5.3 conflates the cutting-plane model's subgradient with the projection returned by Algorithm 5.1; with non-unique projections the returned v need not satisfy (3.5), even with exact stationarity.","rationale":"The reader's weakest assumption was that fmincon actually returns stationary points of the nonconvex master program. My concern is more fundamental and independent of the solver: even granting exact stationarity of the master, the proof of Theorem 5.3 does not establish that the projection-based v_{k+1} used by Algorithm 5.1 is the subgradient delivered by the master's KKT system. When Proj_{Q∩S}(u) is not a singleton, the algorithm may select a projection that fails the necessary optimality condition (3.5), so the local method's advertised certificate is not guaranteed. The distance theorem itself (Theorem 2.5) and the closed-form special cases appear correct; the issue is confined to the local-variant convergence guarantee and the numerical claims built on it. A small deterministic example with non-unique projections can settle whether Theorem 5.3 needs a qualification or a modified projection rule. This is a condition for acceptance rather than a rejection of the whole paper: the central distance construction and global cutting-plane theory remain valuable, but the local-solver guarantee as stated is not reliable.","tokens_in":21809,"tokens_out":24375,"duration_ms":255312,"concrete_test":"Evaluate the proof's key assertion ∂\\hat F^k_Q(u_{k+1})∋v_{k+1} on the 3D example with k=1, u_2=−e1, v_2=e2. Starting from u1=e1, the model is \\hat F^1_Q(u)=⟨e1,u⟩=u_x, so ∂\\hat F^1_Q(−e1)={e1}, which does not contain e2; this directly falsifies the proof step. Then run Algorithm 5.1 with exact master solves on P=span{e1,e2}⊂R^3, Q=cone{e1,e2,e3}, and a deterministic projection rule that returns the lowest-index generator whenever argmax is non-unique. If the algorithm stops at (−e1,e2) with gap 0, verify (3.5): v−⟨u,v⟩u=e2 is not in P^*=span{e3}, contradicting Theorem 5.3's conclusion. Repeat with tie-break v=e3; the condition holds, confirming that the issue is the projection choice.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The local-solution guarantee rests on Theorem 5.3. Its proof starts by 'proceeding... with F_Q replaced by \\hat F^k_Q' and concludes P∋u_{k+1}⊥v_{k+1}−⟨u_{k+1},v_{k+1}⟩u_{k+1}∈P^*, ∂\\hat F^k_Q(u_{k+1})∋v_{k+1}. But stationarity of the master problem only gives some subgradient w∈∂\\hat F^k_Q(u_{k+1}); Algorithm 5.1 line 4 sets v_{k+1}∈Proj_{Q∩S}(u_{k+1}), and there is no reason w=v_{k+1}. This is not a solver artifact: it can fail when the projection is non-unique. Example in R^3: P=span{e1,e2}, Q=cone{e1,e2,e3}; then P^*=span{e3}. For u=−e1, J(u)={2,3}, so Proj_{Q∩S}(u)=cone{e2,e3}∩S contains e2. If the projection tie-break returns e2, then v−⟨u,v⟩u=e2∉P^*, so (3.5) fails, even though u is a stationary/global point of the true problem and −e1 is a stationary point of the master model max(u_x,u_y,0). Thus Theorem 5.3 as written is false for arbitrary tie-breaking; it requires v_{k+1} to be chosen as the multiplier/subgradient from the master solve, or a projection selection rule guaranteeing the KKT face. The paper does not specify such a rule, and the numerical section uses fmincon defaults whose returned multipliers are not used to define v_{k+1}. Hence the advertised guarantee that the local method 'yields an angle that satisfies necessary optimality conditions' is not established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a dissimilarity measure between closed convex cones based on a max-min angle Θ(P,Q), defines Dis_r(P,Q) as the ℓ_r norm of the two sine half-angle terms, and proves (Theorem 2.5) that it is a distance on C^n, coinciding with the Pompeiu-Hausdorff distance between the cones' intersections with the unit sphere. It derives closed forms for linear subspaces (Theorem 4.1) and revolution cones (Theorem 4.3). For polyhedral cones, it proposes a specialized cutting-plane method (Algorithm 5.1), proves global convergence when the master problem is solved globally (Theorem 5.2), and claims that when the master is solved only to stationarity the method still produces points satisfying necessary optimality conditions (Theorem 5.3). Numerical experiments on random cones and ETH-80 image-set classification are reported.","tokens_in":22252,"tokens_out":14668,"duration_ms":150814,"significance":"The geometric construction in Theorem 2.5 is elegant and genuinely useful: it connects a natural angle-based cone dissimilarity to the classical Pompeiu-Hausdorff distance, and the subspace and revolution-cone formulas are clean, parameter-free consequences of the framework. The global cutting-plane convergence theorem (Theorem 5.2) is also valuable and the algorithmic design is sensible. The paper is self-contained and the main distance result does not rely on fitted parameters or circular arguments. The image-set application is illustrative rather than a benchmarked contribution. The main weakness is the local-convergence guarantee: Theorem 5.3, as stated and proved, is not correct under the arbitrary projection tie-breaking used in Algorithm 5.1, and the numerical section does not certify that fmincon returns stationary points. These issues are fixable but are load-bearing for the paper's advertised local-solution guarantee.","major_comments":[{"comment":"The proof of Theorem 5.3 assumes that the projection vector v_{k+1} computed in line 4 of Algorithm 5.1 belongs to ∂\\hat F^k_Q(u_{k+1}). Stationarity of the master problem (5.2) only yields some subgradient w∈∂\\hat F^k_Q(u_{k+1}), and there is no reason that w equals the projection of u_{k+1} onto Q∩S. This is not a solver artifact: it fails when the projection is non-unique. Example in R^3: P=span{e1,e2}, Q=cone{e1,e2,e3}; for u=-e1, J(u)={2,3}, so Proj_{Q∩S}(u)=cone{e2,e3}∩S. If the tie-break returns e2, then v-⟨u,v⟩u=e2∉P*=span{e3}, so system (3.5) fails, although -e1 is a global minimizer of the true problem and a stationary point of the master model max(u_x,u_y,0). Thus Theorem 5.3 as written is false for arbitrary tie-breaking. The proof requires v_{k+1} to be chosen as the multiplier/subgradient returned by the master solve, or a projection-selection rule that guarantees the corre","section":"Section 5.1, Theorem 5.3"},{"comment":"The local variant is presented as satisfying necessary optimality conditions by Theorem 5.3, but the implementation uses fmincon with default parameters on the nonconvex equality-constrained master problem (5.2), and the paper never verifies stationarity of the returned points. For a nonconvex problem with the constraint ||u||=1, default fmincon can terminate at nonstationary points, and the associated multipliers are not used to define v_{k+1}. The reported angles and the classification accuracies in Figs. 8–9 therefore rest on an unverified solver guarantee. Additionally, for the 'global' variant, Gurobi reaches its time limit in about 20% of instances, so those labels are not certified global solutions either. Please report KKT residuals or exit-flag distributions for the local solves, and clearly separate certified-global, certified-stationary, and heuristic results.","section":"Section 6"}],"minor_comments":[{"comment":"The definitions in C^n appear to include the zero cone, for which P∩S^n is empty and Θ(P,Q) is undefined. The paper should either state that all cones are assumed nonzero (or at least not the zero cone) or handle the zero cone separately.","section":"Section 2"},{"comment":"The text says the approach 'could correctly classify 7 classes out of 8 on average', while the introduction states 'an average 90% accuracy'. These statements should be reconciled; the aggregate accuracy depends on the norm and on k, so the precise claim should be stated with the corresponding configuration. Also, 'fmiconto' should be 'fmincon'.","section":"Section 6.1.1 and Introduction"},{"comment":"Reference [7] is not the original source of the ETH-80 dataset. Please cite the primary dataset reference (e.g., Leibe and Schiele, 2003) in addition to, or instead of, the image-set-classification paper.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The central distance theorem and the global cutting-plane convergence are sound and make a solid contribution. The local convergence theorem needs a substantive repair: the algorithm must use a subgradient from the master solve, not an arbitrary projection, and the numerical claims need stationarity certification. This is within the scope of a major revision; the issues are not irreparable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nHere's my take on arXiv:2511.10483. The central construction is a genuine addition: the asymmetric max-min angle Θ(P,Q) and the induced distance Dis_r(P,Q) = 2||(sin(Θ/2), sin(Θ/2))||_r on closed convex cones. Theorem 2.5, showing this equals the Pompeiu-Hausdorff distance between the cones truncated by the unit sphere, is clean and correct. The closed forms for linear subspaces and revolution cones are useful, and the connection to the spectral distance for equal-dimensional subspaces is a nice touch. The paper is honest and largely self-contained.\n\nThe main soft spot is Theorem 5.3, the local-solution convergence guarantee. As stated, the theorem is not true. The proof applies the stationarity of the master program and then asserts that the vector v_{k+1} produced in Algorithm 5.1, line 4, is the subgradient appearing in the KKT conditions. But stationarity only yields some subgradient w in ∂(hat F)_Q^k(u_{k+1}); there is no reason w equals the fresh projection onto Q∩S. This is not a solver artifact—it can fail with non-unique projections. The stress-test example with P=span{e1,e2}, Q=cone{e1,e2,e3} and a tie-break returning e2 for u=-e1 gives v−<u,v>u = e2 ∉ P^*, so (3.5) fails even though u is stationary for both the true and master problems. The fix is either to set v_{k+1} as the multiplier from the master solve or to add a projection selection rule that respects the KKT face. Without that, the advertised “satisfies necessary optimality conditions” claim is unsubstantiated.\n\nThe numerical section is also thin: no code or data released, no baselines, no error bars, and fmincon's stationarity on the nonconvex master is never checked. These are addressable and don't undermine the distance theorems; the classification experiments are an application, not a state-of-the-art claim.\n\nOverall, the paper deserves a serious referee. The distance construction and closed forms are solid and citable. Theorem 5.3 needs a real revision—either fix the proof with an appropriate projection rule or weaken the claim. I'd send it to peer review with that explicit request.","headline":"The distance construction and closed forms are solid, but Theorem 5.3's local optimality claim is false as written—the proof conflates a master-program subgradient with the projection returned by the algorithm.","tokens_in":22726,"tokens_out":3401,"would_cite":true,"duration_ms":37199,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["52A40","90C26","90C46","90C47","49M37"],"pacs":[],"model":"deepseek-v4-flash","headline":"Two one-sided max-min angles between convex cones assemble into a true distance, equal to the Pompeiu-Hausdorff distance on their unit-sphere sections.","keywords":["convex cones","max-min angle","Pompeiu-Hausdorff distance","cutting-plane method","polyhedral cones","image-set classification","few-shot learning","principal angles"],"falsifier":"Take two polyhedral cones in R^3 with known generator matrices, compute the max-min angle by exhaustive angular sampling on the unit sphere to high precision, then run the local variant of the cutting-plane algorithm; if the solver's final iterate violates the KKT system (3.5) — a direct algebraic check — the local convergence claim fails.","tokens_in":21682,"feed_emoji":"📐","tokens_out":5619,"duration_ms":57386,"temperature":0.7,"pith_summary":"This paper introduces a new way to quantify how different two convex cones are, built from the max-min angle: for each direction in one cone, take the smallest angle to the other cone, then take the largest of those. The authors prove that the two one-sided max-min angles combine into a genuine distance, Dis_r(P,Q) = 2||(sin(Θ(P,Q)/2), sin(Θ(Q,P)/2))||_r, which coincides with the Pompeiu-Hausdorff distance between the cones intersected with the unit sphere. For linear subspaces and revolution cones the measure has closed forms, and for polyhedral cones a cutting-plane method computes it, with a local-solution variant that still yields points satisfying necessary optimality conditions. Applied to few-shot image-set classification on the ETH-80 dataset, the measure achieves about 90% average accuracy even when training cones are built from only one to five objects. The work establishes that angle-based comparison of convex cones can be an exact metric, not merely a heuristic similarity.","feed_headline":"Max-min angles make convex cones a metric space","feed_subtitle":"Two directional angles yield a true distance, with closed forms for subspaces and a practical algorithm for polyhedral cones.","key_machinery":"The engine is the max-min angle Θ(P,Q), a pair of directional quantities defined by optimizing over unit vectors in each cone; the metric structure rests on the identity Dis_r(P,Q)=2||(sin(Θ(P,Q)/2), sin(Θ(Q,P)/2))||_r, which links these angles to the Pompeiu-Hausdorff distance between the spherical sections P∩S^n and Q∩S^n through the chord-length relation ||u−v||=2 sin(arccos⟨u,v⟩/2). Computationally, Θ(P,Q) is obtained by minimizing the convex support function F_Q(u)=max_{v∈Q∩S^n}⟨u,v⟩ over the nonconvex set P∩S^n. For polyhedral cones, the paper builds a cutting-plane model of F_Q, leading to a nonconvex quadratically constrained master program; subgradients come from projecting onto Q∩S","core_discovery":"The central claim is Theorem 2.5: for closed convex cones P,Q in R^n, the max-min angle Θ(P,Q)=max_{u∈P∩S^n} min_{v∈Q∩S^n} arccos⟨u,v⟩, together with its mirror Θ(Q,P), defines a distance Dis_r(P,Q)=2||(sin(Θ(P,Q)/2), sin(Θ(Q,P)/2))||_r on the space of all such cones, and this distance equals the Pompeiu-Hausdorff-based distance between the truncated sets P∩S^n and Q∩S^n. Consequently, two cones coincide exactly when both one-sided angles vanish. For linear subspaces of equal dimension, Dis_r reduces to a multiple of the spectral distance; for revolution cones an explicit formula is derived. For polyhedral cones, an adapted cutting-plane algorithm computes the angles: solving the master prog","pith_inferences":["The ordered pair (Θ(P,Q), Θ(Q,P)) could itself serve as a two-dimensional interpretable representation of cone dissimilarity, preserving the asymmetry that Dis_r symmetrizes.","Because Dis_r equals the Pompeiu-Hausdorff distance between spherical sections, any future algorithm that approximates Hausdorff distance on spheres would immediately yield a way to compute Dis_r, connecting this work to computational geometry on spheres.","A testable extension would replace the chord-length factor sin(θ/2) by other monotone functions of angle; the proof structure of Theorem 2.5 suggests which functions preserve the metric axioms.","The cutting-plane framework may extend to non-polyhedral cones whenever projecting onto the cone's spherical section is tractable, for instance for certain spectral cones."],"forward_implications":["Dis_r is an angle-based dissimilarity on all closed convex cones that satisfies the metric axioms, so cone-comparison tasks can rely on it without ad hoc heuristics.","For linear subspaces of equal dimension, Dis_r coincides with a multiple of the spectral distance, computable directly from singular values.","For polyhedral cones, the global cutting-plane variant converges to the exact angle, while the local variant still returns a point satisfying necessary optimality conditions, justifying the use of fast local solvers on large instances.","The two one-sided angles Θ(P,Q) and Θ(Q,P) carry different information, so Dis_r captures a directed structure that a single scalar cannot encode.","In few-shot image-set classification on ETH-80, the distance yields about 90% accuracy even when training cones are built from one to five objects per class."],"fun_headline_variants":["New angle metric turns convex cones into a metric space","Max-min angles: a true distance for convex cones","Cone dissimilarity via max-min angles, now a metric","Distance between cones from max-min angles","Max-min angles yield a metric for convex cones"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The local-solver guarantee rests on the untested premise that the nonlinear solver reaches a stationary point of the nonconvex master problem at every iteration; if it fails, the returned angle may not satisfy the necessary optimality conditions.","fun_headline_variants_meta":{"raw":{"variants":["New angle metric turns convex cones into a metric space","Max-min angles: a true distance for convex cones","Cone dissimilarity via max-min angles, now a metric","Distance between cones from max-min angles","Max-min angles yield a metric for convex cones"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000481,"raw_usage":{"total_tokens":2233,"prompt_tokens":776,"completion_tokens":1457,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":520,"completion_tokens_details":{"reasoning_tokens":1383}},"tokens_in":520,"tokens_out":1457,"duration_ms":9954,"temperature":1.0,"reasoning_tokens":1383,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T22:25:39.655657+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take two polyhedral cones in R^3 with known generator matrices, compute the max-min angle by exhaustive angular sampling on the unit sphere to high precision, then run the local variant of the cutting-plane algorithm; if the solver's final iterate violates the KKT system (3.5) — a direct algebraic check — the local convergence claim fails.","supporting_citations":[],"review_version":1}