{"id":"4cb5037f-f290-4221-b8a8-02e8757c5372","arxiv_id":"2607.17282","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Starting from any separating hyperplane, the algorithm repeatedly rotates it toward the max-margin separator of the nearest points, preserving separation and increasing margin until the global optimum is reached (though finite termination is not proven).","lead":"This paper designs an iterative geometric algorithm that refines an initial separating hyperplane into the maximum-margin separator for binary data, using only local 'active' points at each step. It includes proofs that margin increases each iteration and experiments on MNIST digit pairs showing competitive runtimes on larger datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central guarantee of convergence to the global optimum is unproven: the paper admits finite termination remains open, so the abstract's 'until convergence' is not established.","rationale":"The paper's central claim is that the iterative procedure reaches the global maximum-margin separator. The local-global lemma (Lemma 1) and the margin-monotonicity argument are sound: if the loop exits, the output is correct. However, the loop exit condition is exact equality, and the paper explicitly concedes that finite termination is not proven. A strictly increasing bounded margin sequence alone does not imply convergence of the hyperplanes to the optimum, nor does it imply that the equality H_i == H*_{A_i} is ever attained. This is the single most load-bearing concern because it directly undermines the abstract's advertised guarantee. Other issues—weak baseline comparisons, missing error bars, worst-case active-set size—are secondary; they affect the efficiency claims but not the correctness of the central claim. The reader's weakest_assumption identifies the same gap, and the proposed exhaustive exact search is a concrete way to test whether a counterexample exists. Since the paper is already rated CONDITIONAL on addressing this termination question, my stress test does not change the verdict.","tokens_in":12512,"tokens_out":10598,"duration_ms":107216,"concrete_test":"Run an exhaustive exact-arithmetic search over all linearly separable configurations of n<=6 points with integer coordinates in [-2,2]^2 and all labelings; implement Algorithm 1 with rational arithmetic, trying every valid pivot pair at each step, and stop only when H_i == H*_{A_i}. If any run revisits an active set with a non-increasing margin or fails to terminate within a large bounded number of steps, the central claim is false. If every enumerated instance terminates, the open termination question is not settled but no concrete counterexample exists, and the paper should at minimum replace 'until convergence' with an explicit asymptotic or epsilon-termination guarantee.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Algorithm 1's only stopping rule is H_i == H*_{A_i} (lines 5–6); if this equality never occurs, the loop never exits. Section 3.3 states: 'The existence of a finite upper bound on iter, is not proven in this paper and remains an open question,' and Section 6 repeats this. The proof establishes only that the margin strictly increases at each iteration and is bounded above by the global margin, so the margin sequence converges to some limit; it does not establish that the limit equals the global margin, that the normals converge, or that exact equality is ever attained. The abstract's guarantee of 'convergence to the global optimum' is therefore exactly the unproven part. Since the algorithm uses exact equality as its termination test, even asymptotic convergence would not imply the loop terminates. This is a genuine gap in the central claim, not a cosmetic caveat.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an iterative geometric algorithm for computing the hard-margin SVM separator when an initial separating hyperplane is available. Starting from any separating hyperplane, the algorithm repeatedly computes the current active set, solves the small hard-margin SVM problem on that active set, and rotates the current hyperplane's normal toward the active-set optimum while preserving separation and strictly increasing the margin. The authors prove three auxiliary lemmas (local-global optimality, an angle-margin relation, and a projected-convex-hull optimality condition), report experiments on MNIST digit pairs, and openly state that finite termination of the iteration is not proven.","tokens_in":12716,"tokens_out":15734,"duration_ms":140783,"significance":"If the convergence guarantee could be established, the paper would offer an appealing decomposition: the global SVM problem is replaced by a sequence of much smaller active-set subproblems, and every intermediate iterate is a valid separating hyperplane. The three lemmas are interesting and appear largely correct, and the experiments include reproducible code and honest reporting of cases where the method is slower. The main value is therefore conditional on closing the termination gap or on explicitly repositioning the method as a heuristic with an open convergence question.","major_comments":[{"comment":"The abstract's central claim — 'until convergence to the global optimum' — is not established. The only stopping rule is the exact equality H_i == H*_i (Algorithm 1, lines 5–6), and Section 3.3 explicitly says no finite upper bound on iter is proven. The argument in Section 3.1 shows only that the margin strictly increases and is bounded above by the global margin; it does not prove that the limit equals the global margin, that the normals converge, or that exact equality is ever attained. Since termination uses exact equality, even asymptotic approach to H* would not make the loop exit. This is a load-bearing gap in the main claimed guarantee and must be either proved or explicitly retracted in the abstract and introduction.","section":"Section 3.3 and Algorithm 1"},{"comment":"Equation (13) is incorrect as written. If the minimum is taken over all samples, then any active non-support sample p satisfies w_i^T(p - s_p) = 0 and w*_i^T(p - s_p) > 0, giving alpha_max^+(p) = 0. Thus alpha_actual = 0 at every iteration and no progress is possible. The minimization must be restricted to samples outside the current active set (as the surrounding prose says) and to admissible values in (0,1], ignoring negative values and values greater than 1. The current formula therefore does not describe the algorithm that the experiments presumably run.","section":"Section 3.2, Eq. (13)"},{"comment":"The proof states that 'lambda_i > 0 if and only if x_i is a support vector,' but Section 2 defines support vectors more broadly to include redundant samples with zero KKT multipliers. Under that definition the equivalence is false, and the representation w_KKT = C(P* - N*) with positive convex coefficients may fail. The proof should explicitly restrict S to strict support vectors (positive KKT multipliers) or otherwise justify why the broader definition does not affect the argument. Since Lemma 2 underpins the margin-monotonicity claim, this needs a precise fix.","section":"Appendix, Lemma 2 proof"},{"comment":"The pivot selection is an 'arbitrarily selected' oppositely labeled support vector pair from H*_i. The rotation path, the blocking sample encountered, and hence the whole active-set sequence depend on this choice. The paper gives no argument that the algorithm's behavior — or its termination/convergence — is invariant under the pivot choice. In the absence of a termination proof, the arbitrary pivot is an additional source of uncontrolled behavior; the authors should at least state the precise pivot rule used in the experiments and discuss whether the theoretical claims are meant to hold for every pivot.","section":"Algorithm 1, lines 8–9"}],"minor_comments":[{"comment":"The condition H_i == H*_i is not a meaningful numerical stopping test. A tolerance or a check on margin/KKT residual should be specified, especially since the experiments are carried out in floating point.","section":"Algorithm 1"},{"comment":"If alpha_actual = 1, the algorithm could return H*_i immediately; as written, it performs an extra iteration and relies on the equality test at the next loop. This is harmless mathematically but should be clarified.","section":"Section 3.2"},{"comment":"The column header says 'Speedup' but the text calls it 'Ratio.' Also, the table reports only averages over 10 runs; standard deviations or per-run ranges would help assess the stability of the wall-clock comparisons.","section":"Table 1"},{"comment":"There are several small typographical slips: Eq. (38) writes conv(Π_H(P)) and conv(Π_H(N)) instead of A+ and A−; later lines contain 'p∈A−' and similar notational inconsistencies. These should be corrected.","section":"Appendix, Lemma 3"},{"comment":"The definition of the active set assumes that H is positioned equidistantly between the closest oppositely labeled samples, but the case of ties is not formalized. A precise definition of 'closest' and of the active set under ties would improve reproducibility.","section":"Section 2"},{"comment":"Reference [9] is an arXiv preprint; if it is not yet peer reviewed, the authors should state its status or provide an alternative public source for the linearly separable digit pairs.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is honest about the open termination question, but the abstract and introduction nevertheless assert convergence to the global optimum. I would not reject outright because the geometric idea is sound conditionally and the experiments are reproducible; however, the written algorithm contains a concrete bug in Eq. (13) that must be fixed, and the central convergence claim must either receive a proof or be substantially weakened. The novelty is incremental rather than breakthrough, but the active-set decomposition is a reasonable contribution for a specialized journal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing you should know: this paper is a legitimate geometric iteration scheme for hard-margin SVM, not a fake. The author starts from any separating hyperplane, computes the maximum-margin separator for the current active set, and rotates the normal vector toward it while preserving separation and strictly increasing the margin each iteration. The local–global optimality lemma (Lemma 1) is simple but correct, and the angle–margin identity (Lemma 2) is a neat piece of geometry. The paper also ships code and data on GitHub, and the author is unusually candid about the main gap.\n\nThat gap is real and load-bearing. The algorithm's only stopping rule is exact equality between the current hyperplane and the active-set optimum (lines 5–6). The paper proves the margin strictly increases and is bounded above, so the margin sequence converges to some limit. But it does not prove that limit equals the global margin, that the normals converge, or that exact equality is ever attained. Section 3.3 and Section 6 explicitly state that a finite upper bound on iterations is not proven and remains an open question. Without finite termination, the abstract's claim of 'convergence to the global optimum' is unsupported. The stress-test note is right: this is not a cosmetic caveat, it is the central guarantee.\n\nThe experiments are not strong enough to rescue the argument. They use 20 MNIST digit-pair tasks, linearly separable subsets, compared against a single CVXPY/OSQP baseline on Colab CPU. On the small test sets the method is about ten times slower; on the larger training sets it wins in 6–8 of 10 cases. No error bars, one baseline, and timings that vary wildly. The observation that active sets remain small is nice, but the 'competitive with state-of-the-art' claim is overstated.\n\nTwo other soft spots: the pivot selection is arbitrary (any opposite-labeled support vector pair) with no proof that the choice doesn't affect the trajectory or cause cycles, and the subproblems are solved by calling CVXPY, making the method a wrapper around an existing solver.\n\nWho is this for? Readers interested in geometric interpretations of SVM and active-set decomposition. It deserves a serious referee because the core idea is mathematically interesting and testable, and the termination question is genuinely open. But in its current form, it should not be accepted as a proof of convergence.","headline":"A coherent geometric active-set method with a solid monotone-margin argument, but the central convergence-to-global-optimum claim is explicitly unproven and the experiments are too weak to support the performance claims.","tokens_in":13152,"tokens_out":1922,"would_cite":false,"duration_ms":19445,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C20","68T05"],"pacs":[],"model":"deepseek-v4-flash","headline":"Starting from any separating hyperplane, repeatedly rotating its normal toward the maximum-margin separator of the current active set increases the margin at every step and, the paper claims, converges to the global optimal separating hyper","keywords":["maximum-margin hyperplane","hard-margin SVM","active set","separating hyperplane","geometric optimization","convex quadratic programming","linear separability","iterative refinement"],"falsifier":"Construct a linearly separable point set, start with any valid separating hyperplane, and record the triple (active set, pivot, αactual) at each iteration. If the same triple ever recurs before reaching α=1—or if iterations continue without the margin increasing toward γ*—the claimed convergence to the global optimum is false. Such a recurrence is checkable by simulation on finite data.","tokens_in":12408,"feed_emoji":"📐","tokens_out":6147,"duration_ms":59706,"temperature":0.7,"pith_summary":"This paper takes the hard-margin support vector machine problem—find the hyperplane that separates two labeled point sets with the largest possible gap—and asks whether a merely feasible separator can be polished into the optimal one. It answers with a geometric iteration: identify the samples closest to the current hyperplane (the active set), compute the maximum-margin separator of just those samples, and rotate the current normal vector toward that separator around a pivot, keeping every sample on the correct side. The rotation strictly increases the margin at every step, so every intermediate hyperplane is a usable, better separator. The paper proves a local-to-global optimality criterion and an angle–margin identity, and reports experiments on handwritten-digit pairs suggesting the approach can beat one-shot solvers on larger datasets when a feasible start is available.","feed_headline":"Rotating toward active-set optima reaches the max-margin hyperplane","feed_subtitle":"Each iteration solves a tiny subproblem, and the margin rises every step—so large datasets can beat one-shot solvers.","key_machinery":"The active set and the interpolated normal. Active set A_i is the set of samples at minimal distance from the current hyperplane H_i. The paper forms w_i(α)=(1−α)w_i+αw*_i, rotating the normal from its current direction to the normal of the active-set optimum while pivoting the plane about the midpoint of a positive/negative support-vector pair. The identity γ_i(α)=γ*_i cos θ(α), where θ is the angle between the two normals, converts the geometric rotation into a monotone margin gain, and the maximal admissible α is found in closed form by solving w_i(α)^T(p−s_p)=0 for every outside sample p.","core_discovery":"At the paper's center is a locality principle: a separating hyperplane that is already optimal for its active set—the closest samples from each class—is necessarily globally optimal, so the algorithm can stop when the active-set optimum coincides with the current plane. If it does not, the paper's move is to rotate the normal vector toward the normal of the active set's maximum-margin separator, holding the plane through the midpoint of an oppositely labeled support-vector pair. During that rotation, support vectors stay glued to the margin hyperplanes, non-support active samples drift away, and the margin evolves as γ(α) = γ*_A cos θ(α), making it strictly increase until a new sample blocks","pith_inferences":["A natural next test is to search over low-dimensional linearly separable configurations for a cycle of hyperplanes; the paper itself leaves finite termination open, so exhibiting a cycle would falsify convergence while proving one cannot occur would close the main gap.","Because the blocking calculation is a batch of dot products and the iteration never loses feasibility, the method is a plausible warm-start strategy for streaming or incrementally updated data where a good separator is maintained online.","The closed-form α values could be computed with batched or randomized linear algebra for very large n, making the per-iteration O(nd) cost attractive relative to global solvers on huge datasets."],"forward_implications":["Every iterate is a valid separating hyperplane with a strictly larger margin than the previous one, so the process can be stopped early and still deliver a useful classifier.","The original global quadratic program is replaced by a sequence of much smaller active-set quadratic programs; on the reported roughly 12,000-sample experiments the active sets stayed close to the support-vector count.","Local optimality plus global separation is a certificate of global optimality, giving the algorithm a natural stopping rule.","The method is agnostic to the subroutine used for the active-set subproblem, so any existing SVM optimizer can be dropped in."],"fun_headline_variants":["Rotating hyperplanes beat one-shot solvers on large datasets","Active-set rotation boosts margin to global optimum iteratively","Geometry-guided SVM: tiny subproblems, rising margin each step","From any separator, rotate to the max-margin plane via active sets","Iterative hyperplane tuning: local optima, global gain"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"That the iteration count is finite: Section 3.3 concedes no finite upper bound on iterations is proven and termination remains open; if the process can cycle or fail to reach the active-set optimum, the convergence guarantee fails.","fun_headline_variants_meta":{"raw":{"variants":["Rotating hyperplanes beat one-shot solvers on large datasets","Active-set rotation boosts margin to global optimum iteratively","Geometry-guided SVM: tiny subproblems, rising margin each step","From any separator, rotate to the max-margin plane via active sets","Iterative hyperplane tuning: local optima, global gain"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000153,"raw_usage":{"total_tokens":1036,"prompt_tokens":725,"completion_tokens":311,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":469,"completion_tokens_details":{"reasoning_tokens":225}},"tokens_in":469,"tokens_out":311,"duration_ms":3650,"temperature":1.0,"reasoning_tokens":225,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T18:27:34.053420+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a linearly separable point set, start with any valid separating hyperplane, and record the triple (active set, pivot, αactual) at each iteration. If the same triple ever recurs before reaching α=1—or if iterations continue without the margin increasing toward γ*—the claimed convergence to the global optimum is false. Such a recurrence is checkable by simulation on finite data.","supporting_citations":[],"review_version":1}