{"id":"8a9e38fc-8aac-4c77-bb5c-e15df0aa78e4","arxiv_id":"2608.07716","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The cosine measure of a function at a point is defined as the cosine measure of its non-descent directions, giving a sharp threshold for guaranteed descent and algorithms for finite-max and l1-norm functions.","lead":"The paper defines the cosine measure of a function at a point, a number that tells how good a set of search directions must be to guarantee it contains a downhill direction, even when the function is not smooth or the gradient is zero. It gives algorithms to compute this number for common nonsmooth functions and a closed-form value for the l1 norm.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 10 is sound, but Algorithm 2's proof (Theorem 29) has an unproven projection claim; the claim is true for the optimal sine vector, so the gap is fixable and does not invalidate the algorithm.","rationale":"The central claim of the paper, Theorem 10, is correct: cm(f;x) = cm(D^c) is the sharp threshold, and the proof via the cosine vector set is valid. The genuinely delicate point is Algorithm 2's correctness proof in Theorem 29, exactly where the reader located the weakest assumption. I agree that the projection assertion is not proved in the paper, and that the formula cm(D^c) = min_i sqrt(1-(t_i^*)^2) depends on it. However, the assertion is not false: for the optimal sine vector u^*, the projection onto any active facet lies inside the cone because p_l^\\top(u^* - t^* p_j) \\ge t^*(1-p_j^\\top p_l) \\ge 0 for every facet normal p_l. Thus the concern lands as a missing proof, not as a counterexample to Algorithm 2. The finite-max section also states the set of non-descent directions as a union involving active gradients; strictly speaking, higher-order terms can make directions with zero active-gradient value descent, but the closure of the true non-descent set is the stated union, and cm/cV depend only on the closure, so the algorithmic conclusions survive. I therefore see no reason to change the reader's conditional verdict; the condition should be that the projection argument in Theorem 29 is supplied before publication.","tokens_in":20662,"tokens_out":27746,"duration_ms":287687,"concrete_test":"Settle the projection claim by an independent derivation plus a numerical check. Analytically: prove that if u^* maximizes \\min_{p\\in P} p^\\top u with t^*>0 and p_j^\\top u^* = t^*, then q = u^* - (u^{*\\top} p_j) p_j satisfies p_l^\\top q \\ge 0 for every facet normal p_l, so q lies in the cone. Numerically: for a random pointed 4-facet cone in R^3 (e.g., p1=e1, p2=e2, p3=e3, p4=normalize(-0.2,-0.2,1)), compute u^* by the SOCP \\max\\{t : \\|u\\|\\le 1,\\ p_i^\\top u \\ge t\\}, verify the projection membership for each active facet, and compare Algorithm 2's value \\min_i \\sqrt{1-(t_i^*)^2} with a brute-force minimization of \\max_{d\\in\\delta C} d^\\top u over a fine triangulation of the boundary; the two values should match.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The weakest load-bearing step is in Theorem 29 (Section 4.2). After reducing cm((S_i^\\circ)^c) to min_{u\\in S_i^\\circ} max_{d\\in \\delta\\operatorname{pspan}(S_i)} d^\\top u, the proof asserts \\max_{d\\in\\delta C} d^\\top u_i^* = \\max_{p\\in P_i} \\sqrt{1-(p^\\top u_i^*)^2}, which requires that the projection of u_i^* onto the facet attaining the maximum actually lies inside that facet of the cone. The text only notes that the projection 'need not lie in pspan(S_i)' and asserts that the maximum is attained at a closest facet, without proof. This is a genuine gap: if the projection fell outside the facet, the maximum over that facet could be attained on a lower-dimensional face, and the formula could overestimate cm. A good-faith check shows the assertion is nevertheless true. Let u^* solve \\max_{\\|u\\|\\le 1} \\min_{p\\in P} p^\\top u, set t^* = \\min_p p^\\top u^* > 0, and let p_j be active with p_j^\\top u^* = t^*. For any facet normal p_l, p_l^\\top (u^* - t^* p_j) = p_l^\\top u^* - t^*(p_j^\\top p_l) \\ge t^*(1 - p_j^\\top p_l) \\ge 0. Hence q = u^* - (u^{*\\top} p_j) p_j lies in the cone and is the required boundary point; inactive facets have smaller projection norms. So the disputed equality is true, and Algorithm 2's formula is correct. The paper should include this argument before acceptance.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces the cosine measure of a function at a point, cm(f;x), defined as the cosine measure of the set D^c of all unit non-descent directions of f at x. The central result, Theorem 10, states that any closed set S of unit vectors with cm(S)>cm(f;x) must contain a descent direction of f at x. The paper proves basic properties of cm(f;x), gives examples showing that every value in [-1,1] can occur when the gradient vanishes, and develops three algorithms: Algorithm 1 for the case where cl D^c is a finite union of polyhedral cones contained in a closed half-space, Algorithm 2 for the complementary positive case via the spherical complement, and Algorithm 3 specialized to finite-max functions. It closes with a closed-form expression for the cosine measure of the l1 norm at a point with m zero coordinates, cm = sqrt(m/n).","tokens_in":21038,"tokens_out":20548,"duration_ms":196282,"significance":"If the results are correct, cm(f;x) provides a natural worst-case threshold for guaranteeing a descent direction in settings where the classical gradient condition fails, and it extends cosine-measure theory from finite positive spanning sets to infinite and nonsmooth directional sets. The l1-norm formula in Theorem 34 is elegant and immediately usable in derivative-free optimization analyses. The paper contains several strengths: Theorem 10 has a short, correct proof; the examples in Section 3.1 collectively demonstrate attainability of all values in [-1,1]; the algorithms are deterministic and reduce to second-order cone programs; and the exposition is generally clear. The main weakness is that the correctness proof of Algorithm 2 contains an unproven geometric assertion, and both Algorithms 1 and 2 rely on an unpublished companion result. These issues are fixable but currently block verification of a load-bearing part of the paper.","major_comments":[{"comment":"The proof of Theorem 29 asserts, without proof, that max_{d in delta-pspan(S_i)} d^T u_i^* equals max_{j_i} ||Proj_{H_{j_i}} u_i^*||, and that the maximum is attained at a facet realizing the minimal distance from u_i^*, for which the projection lies in the cone. This assertion is load-bearing: if the projection onto the closest facet fell outside that facet, the maximum could be attained on a lower-dimensional face and the formula cm((S_i^o)^c) = max_{p in P_i} sqrt(1-(p^T u_i^*)^2) would not follow. The text only notes that the projection 'need not lie in pspan(S_i)' and does not prove the required containment for the optimal sine vector. Please insert the missing argument: for an active normal p_j with p_j^T u_i^* = t_i^* = min_{p in P_i} p^T u_i^*, the vector q = u_i^* - (u_i^{*T} p_j) p_j satisfies p_l^T q = p_l^T u_i^* - t_i^*(p_j^T p_l) >= t_i^*(1 - p_j^T p_l) >= 0 for every p_l in P_i, so q lies in the cone and on the boundary; this establishes the asserted equality and the correctness of the algorithm's formula.","section":"Section 4.2, Theorem 29"},{"comment":"The correctness proofs of Algorithms 1 and 2 depend on Lemma 33 of reference [2], which is cited as an unpublished manuscript with overlapping authorship. This lemma is used in the proof of Theorem 25 to justify the exactness of the convex relaxation for nonpositive cosine measure, and it is invoked again at the end of Theorem 29 to justify the convex relaxation in the positive case. Because this lemma is exactly what licenses replacing the nonconvex cosine-measure problem by the SOCPs in Steps (1.1) and (1.2), the paper should either state and prove the needed lemma in a self-contained appendix or give the precise statement of Lemma 33 and explicitly identify the dependency. As written, the reader cannot verify a load-bearing step without consulting a separate unpublished paper.","section":"Section 4.1 and 4.2"}],"minor_comments":[{"comment":"The examples in Section 5 (Examples 33 and 34) exercise only Algorithm 3, which is a simplified version of Algorithm 2 for a single cone. A small hand-worked or numerical instance of Algorithm 2 with k > 1 cones (for example, the configuration of Example 11) would materially help validate Step 2 and the assumption that pspan(S_i) cap pspan(S_j) = {0}.","section":"Section 4.2"},{"comment":"Definition 1 uses the same symbol for the positive span and for the normalized positive span, and Algorithms 1, 2, and Theorem 21 rely on the normalized version. Since the two notions are used interchangeably in the displayed text, please introduce a distinct symbol for the normalized cone (for example, overline-pspan) to avoid ambiguity.","section":"Definition 1 and Section 4"},{"comment":"There are several typographical slips that should be corrected in a revision: 'the the cosine measure' in the abstract, 'authours' in the acknowledgments, 'particulary' and 'in particualar' in Section 5, and 'The set D^c for f_alpha at x_0 = 0 is represented by the red spherical cap' in Example 16 where the figure is not included in the text.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper's main idea is sound and the l1-norm result is attractive, but the proof of Algorithm 2 has a genuine gap that needs to be filled, and the reliance on the unpublished companion paper [2] should be made explicit or the lemma proved internally. I believe the gap is fixable, so I recommend major revision rather than rejection. The editor may also wish to confirm that reference [2] is publicly available and that its Lemma 33 matches the usage in this manuscript."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing: the central idea is sound and worth knowing. The paper defines cm(f;x) as the cosine measure of the set of all unit non-descent directions of f at x, and proves the threshold theorem (Theorem 10): a closed set of search directions S with cm(S) > cm(f;x) must contain a descent direction. This is a natural, genuinely new way to quantify how good a search set has to be when f is nonsmooth or the gradient is zero. The examples in Section 3.1 showing cm(f;x) can take every value in [-1,1] are clean. The closed form for the l1 norm (Theorem 34), cm(||·||_1; x0) = sqrt(m/n) with m the number of zero entries, is a concrete result people will cite.\n\nThe soft spots are in Section 4.2. Theorem 29 asserts that the maximum over the boundary of a polyhedral cone equals the maximum over facet-normal projections of the optimal vector, but it does not prove that the projection onto the closest facet lies inside that facet. The text only says the projection need not lie in pspan(S_i). That is a real gap. Good-faith inspection shows the assertion is true for the optimal sine vector—the stress-test note has the argument—so the algorithm survives, but the authors should insert those lines before publication. A referee should not let this slide.\n\nThe other issue is dependence on the companion paper [2] for Lemma 33, which underlies Algorithm 1's exactness. That is an unpublished preprint with overlapping authors. Given the central role, the editor should ask for the companion to be posted or the lemma proved in-line. It does not affect Algorithm 2 or the main theorems.\n\nNo numerical experiments back Algorithm 2. For a theory paper that is acceptable, but a couple of low-dimensional examples would have caught the missing projection argument earlier.\n\nOverall: solid, honest paper, with one fixable gap and one dependency to clean up. It deserves a serious referee, and I'd accept conditionally. The cosine measure of a function is a new organizing concept for nonsmooth DFO, and the l1 formula is a small but useful payoff. Reading group: worth a session if people care about direct-search convergence.","headline":"Sound new concept, fixable proof gap in Algorithm 2's theorem, useful l1 closed form; accept with revisions.","tokens_in":21538,"tokens_out":2668,"would_cite":true,"duration_ms":23660,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C56","90C30","49J52","52A20"],"pacs":[],"model":"deepseek-v4-flash","headline":"Every function and point carry a threshold — the cosine measure of its non-descent directions — and any search set whose cosine measure exceeds it must contain a descent direction, even without differentiability or a nonzero gradient.","keywords":["cosine measure","descent directions","infinite sets of directions","finite-max function","ℓ1 norm","nonsmooth optimization","second-order cone program","positive spanning sets"],"falsifier":"Take a pointed polyhedral cone in $\\mathbb{R}^3$ with facet normals $\\{p_1,\\dots,p_m\\}$, solve the SOCP of Algorithm 2 for its sine vector $u^*$, and compare $\\sqrt{1-(t^*)^2}$ with $\\max_{d \\in \\partial C \\cap \\mathbb{S}^2} (u^*)^{\\top}d$ computed by dense boundary sampling; if the two differ for any cone whose optimal vector's projection onto its nearest facet falls outside that facet, the identity behind Theorem 29 is refuted. A direct geometric check is simpler: for the computed $u^*$, verify that its orthogonal projection onto every facet hyperplane realizes the maximum of $(u^*)^{\\top}d$ over the cone's boundary only for facets containing that projection — if the projection is outside the facet for the minimizing facet, the formula is unsupported.","tokens_in":20484,"feed_emoji":"🎯","tokens_out":20117,"duration_ms":143670,"temperature":0.7,"pith_summary":"The paper introduces the cosine measure of a function at a point, $\\mathrm{cm}(f;\\mathbf{x})$, defined as the cosine measure of the set of all unit non-descent directions of $f$ at $\\mathbf{x}$. Its central theorem states that any set of search directions whose cosine measure exceeds $\\mathrm{cm}(f;\\mathbf{x})$ is guaranteed to contain a descent direction of $f$ at $\\mathbf{x}$ — a guarantee that survives nondifferentiable functions and points where the gradient vanishes. The authors show by explicit examples that at a zero-gradient point the cosine measure of a smooth function can take any value in $[-1,1]$, so the threshold genuinely quantifies how hard a point is for direction-based descent. Algorithms compute the threshold exactly through second-order cone programs, including for infinite sets of directions; for finite-max functions the computation reduces to one SOCP built from active gradients, and for the $\\ell^1$ norm it yields the closed formula $\\mathrm{cm}(\\lVert \\mathbf{x}\\rVert_1;\\mathbf{x}_0)=\\sqrt{m/n}$, where $m$ is the number of zero entries of $\\mathbf{x}_0$. If the paper is right, direct-search methods get a per-point, per-function budget: a polling set whose cosine measure beats this value is provably assured of a descent direction, even where classical smooth analysis says nothing.","feed_headline":"Exceed this cosine value and a descent direction is guaranteed","feed_subtitle":"The guarantee now extends to nondifferentiable points and zero gradients, with a closed formula for the ℓ1 norm.","key_machinery":"The central object is the set $D^c(f;\\mathbf{x})$ of unit non-descent directions of $f$ at $\\mathbf{x}$, together with its cosine measure $\\mathrm{cm}(f;\\mathbf{x}) = \\mathrm{cm}(D^c) = \\min_{\\lVert u\\rVert=1} \\sup_{d\\in D^c} u^{\\top}d$. The load-bearing identity is the threshold theorem: $\\mathrm{cm}(S) > \\mathrm{cm}(f;\\mathbf{x})$ forces $S$ to contain a descent direction, with the proof running through a separating vector $u \\in \\mathrm{cV}(f;\\mathbf{x})$ and the observation that a maximizer of $u^{\\top}d$ over $S$ cannot lie in $D^c$. For computation, the mechanism is a second-order cone program: in the nonpositive case, $\\mathrm{cm}(D^c) = \\min\\{t : \\lVert u\\rVert \\le 1,\\ u^{\\top}d \\le t \\text{ for all } d\\in S\\}$, using the exactness of the convex relaxation when the cosine measure is nonpositive; in the positive case, the spherical complement of $\\mathrm{cl}\\,D^c$ is written as a disjoint union of interiors of pointed polyhedral cones $\\mathrm{pspan}(S_i)$, and for each cone the algorithm computes its sine vector $u_i^*$ — the unit vector inside the cone farthest in angle from every boundary facet — via the program $\\max\\{t : \\lVert u\\rVert \\le 1,\\ p^{\\top}u \\ge t \\text{ for all facet normals } p\\}$, from which $\\mathrm{cm}(D^c) = \\min_i \\sqrt{1-(t_i^*)^2}$. For finite-max functions, the facet normals of the relevant cone are exactly the normalized negative active gradients, collapsing the computation to a single SOCP; the sine vector of a cone coincides with the incenter of a proper cone.","core_discovery":"The central discovery is that every function $f$ and point $\\mathbf{x}$ carry a number $\\mathrm{cm}(f;\\mathbf{x})$ — the cosine measure of the function at the point — equal to the cosine measure of the set $D^c(f;\\mathbf{x})$ of all unit non-descent directions of $f$ at $\\mathbf{x}$. Theorem 10 shows this number is a sharp threshold: every closed set $S$ of unit vectors with $\\mathrm{cm}(S) > \\mathrm{cm}(f;\\mathbf{x})$ contains a descent direction of $f$ at $\\mathbf{x}$, and the same holds for every orthogonal rotation of $S$; the threshold cannot be improved, because the set $D^c$ itself has cosine measure exactly $\\mathrm{cm}(f;\\mathbf{x})$ and contains no descent direction. The paper demonstrates that equality cases are richly attainable: at a point where $\\nabla f(\\mathbf{x}) = \\mathbf{0}$, the value $\\mathrm{cm}(f;\\mathbf{x})$ can be anything in $[-1,1]$, so a vanishing gradient imposes no restriction on the threshold. It gives deterministic algorithms that compute $\\mathrm{cm}(f;\\mathbf{x})$ exactly — a second-order cone program when the closure of $D^c$ lies in a closed half-space, and a per-cone sine-measure computation over the facet normals of pointed polyhedral cones otherwise — and applies them to finite-max functions, where the non-descent directions are governed by the active gradients. For the $\\ell^1$ norm this yields the closed form $\\mathrm{cm}(\\lVert\\mathbf{x}\\rVert_1;\\mathbf{x}_0) = \\sqrt{m/n}$, where $m$ is the number of zero entries of $\\mathbf{x}_0$, with the consequence that any set $E$ of directions satisfying $\\mathrm{cm}(E) > \\sqrt{m/n}$ must contain a descent direction of the $\\ell^1$ norm at $\\mathbf{x}_0$.","pith_inferences":["I read $\\mathrm{cm}(f;\\mathbf{x})$ as a per-point measure of descent difficulty: tabulating it for other standard nonsmooth functions — maximum eigenvalue, nuclear norm, hinge loss, $\\ell^1$ of $A\\mathbf{x}-\\mathbf{b}$ — along the lines of Theorem 34 would produce a practical difficulty map for direction-based methods (the paper lists these functions as future work, not as a difficulty map).","The threshold suggests an adaptive polling strategy the paper does not develop: a derivative-free method could estimate $\\mathrm{cm}(f;\\mathbf{x})$ at the current iterate and construct a polling set whose cosine measure provably exceeds it, turning a static guarantee into a per-iteration dynamic one.","Because the sine vector of a cone is the incenter of a proper cone, existing incenter algorithms and bounds could supply estimates or upper bounds for $\\mathrm{cm}(f;\\mathbf{x})$ in high dimensions, where enumerating cone facets is exponential."],"forward_implications":["Any direct-search polling set whose cosine measure exceeds $\\mathrm{cm}(f;\\mathbf{x})$ is guaranteed to contain a descent direction of $f$ at $\\mathbf{x}$, with the guarantee now valid for nondifferentiable functions and zero-gradient points, not only for $\\mathcal{C}^1$ functions with nonzero gradient.","The threshold is sharp: because $D^c(f;\\mathbf{x})$ itself has cosine measure exactly $\\mathrm{cm}(f;\\mathbf{x})$ and contains no descent direction, no condition based only on the cosine measure of a set can force a descent direction below this value; when $\\mathrm{cm}(f;\\mathbf{x}) > -1$, a set can always be built at equality that fails to descend.","For the $\\ell^1$ norm at a point with $m$ zero entries, any set $E$ of unit directions with $\\mathrm{cm}(E) > \\sqrt{m/n}$ must contain a descent direction at that point, giving a closed-form, coordinate-readable condition.","A vanishing gradient carries no information about the threshold: smooth functions at zero-gradient points realize every value of $\\mathrm{cm}(f;\\mathbf{x})$ in $[-1,1]$, and the value $0$ occurs exactly when the non-descent directions fit in some closed half-space.","The cosine measure of infinite sets of directions becomes computable exactly whenever the closure of the non-descent set is a finite union of finitely generated cones: one SOCP in the half-space case, one SOCP per pointed cone in the positive case."],"supporting_citations":[{"why":"Supplies the extended cosine-measure definition for possibly infinite sets of unit vectors, the positive-spanning and zero-measure characterization results (used in Proposition 4 and Theorem 13), and the exactness of the convex relaxation for nonpositive measures that Algorithm 1 relies on.","marker":"[2]"},{"why":"Provides the orthogonal-invariance theorem used in Lemma 9, which lets Theorem 10 extend its descent guarantee to every rotated copy of the search set.","marker":"[34]"},{"why":"Gives the original definition of cosine measure and the classical smooth, nonzero-gradient descent guarantee that this paper generalizes to the nondifferentiable and zero-gradient cases.","marker":"[26]"},{"why":"Book-level reference for direct-search convergence theory based on positive spanning sets with cosine measure bounded away from zero; it frames the applied question the new threshold answers.","marker":"[7]"},{"why":"Supplies the deterministic cosine-measure algorithm for finite positive spanning sets and the active-set basis result used in the paper's computational adaptation to infinite sets.","marker":"[14]"},{"why":"Provides the frame-reduction theorem for polyhedral cones used to preprocess generating sets in Algorithms 1 and 2, including the uniqueness of frames for pointed cones.","marker":"[33]"}],"fun_headline_variants":["A sharp cosine threshold for descent directions","Cosine measure for nondifferentiable functions","Zero gradient? Cosine measure still sets the bar","The number that guarantees a descent direction","Exact threshold: any set above it has a descent"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is a geometric assertion the proof leaves open in the positive-cosine case: that the direction most deeply hidden inside each pointed cone of search directions always sees its widest angle to the boundary at a facet whose containing hyperplane it projects onto from inside, so the projection never lands outside the cone; if that fails for some cone, the algorithm's returned value for the cosine measure is not established.","fun_headline_variants_meta":{"raw":{"variants":["A sharp cosine threshold for descent directions","Cosine measure for nondifferentiable functions","Zero gradient? Cosine measure still sets the bar","The number that guarantees a descent direction","Exact threshold: any set above it has a descent"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000288,"raw_usage":{"total_tokens":1827,"prompt_tokens":1220,"completion_tokens":607,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":836,"completion_tokens_details":{"reasoning_tokens":540}},"tokens_in":836,"tokens_out":607,"duration_ms":5547,"temperature":1.0,"reasoning_tokens":540,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T00:24:45.342957+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a pointed polyhedral cone in $\\mathbb{R}^3$ with facet normals $\\{p_1,\\dots,p_m\\}$, solve the SOCP of Algorithm 2 for its sine vector $u^*$, and compare $\\sqrt{1-(t^*)^2}$ with $\\max_{d \\in \\partial C \\cap \\mathbb{S}^2} (u^*)^{\\top}d$ computed by dense boundary sampling; if the two differ for any cone whose optimal vector's projection onto its nearest facet falls outside that facet, the identity behind Theorem 29 is refuted. A direct geometric check is simpler: for the computed $u^*$, verify that its orthogonal projection onto every facet hyperplane realizes the maximum of $(u^*)^{\\top}d$ over the cone's boundary only for facets containing that projection — if the projection is outside the facet for the minimizing facet, the formula is unsupported.","supporting_citations":[],"review_version":1}