{"id":"e7f8385f-caf6-4011-9705-85552ac02c2a","arxiv_id":"2510.24981","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A function class whose sublevel sets are star-shaped at a minimizer unifies several generalized convexity conditions and yields linear convergence of first-order methods, including a new proximal point result on star-shaped sets.","lead":"This paper introduces star quasiconvexity, a broad condition on a function's sublevel sets that includes convex, quasiconvex, star-convex and quasar-convex functions as special cases. It proves linear convergence of gradient-type methods and, more novelly, of the proximal point algorithm on star-shaped, nonconvex sets.","discovery_kind":"unification","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section 4's heavy-ball/Nesterov convergence for star quasiconvexity is asserted by citing proofs that use pairwise quasiconvexity (10); the proof transfer is not established.","rationale":"The reader's weakest assumption, the star-shapeness of K in Theorem 33, is a correct observation about the theorem's scope, but the PPA contraction proof itself appears internally sound: inequality (22) follows from the strong star quasiconvexity hypothesis on the segment [xbar, x*], and the algebra leading to (34) is valid. The more serious risk to the central claim is Section 4. The paper's own abstract and contribution list present heavy-ball and Nesterov accelerations as part of the unified first-order treatment, yet the theorems are not proved and the cited external proofs assume strong quasiconvexity, a property strictly stronger than strong star quasiconvexity in pairwise form. Since strong star quasiconvexity is equivalent only to the restricted secant inequality (24), one must verify that the external Lyapunov arguments never use (10) for pairs not on a ray from x*. The typo in Theorem 30, which states 'strongly quasiconvex' rather than 'strongly star quasiconvex', reinforces this concern. I therefore keep the reader's CONDITIONAL verdict: Theorem 33 is well supported, but the broader claim of a unified linear-convergence framework for first-order methods needs a proof that the heavy-ball/Nesterov results transfer to star quasiconvexity.","tokens_in":16817,"tokens_out":38781,"duration_ms":332791,"concrete_test":"Re-derive the Lyapunov inequality of [17, Theorem 13] using only (24) in place of (10), and identify whether any step compares h at two points not collinear with x*. If such a step occurs, Theorem 27 does not follow. As a numerical check, construct a smooth 1D function satisfying (24) but not (10), e.g. h'(x)=0.6γ for 0<x<1, h'(x)=γx/2 for x<0 and x>1, smoothed near 0 and 1, and run update (27) with α=0.5 and β=1/(2L). If the heavy-ball Lyapunov decrease fails or the claimed rate 1−ρ/σ is not observed, the Section 4 transfer collapses.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract advertises linear convergence of gradient-type methods beyond convexity, and Section 4 claims heavy-ball and Nesterov acceleration for strongly star quasiconvex functions. But Theorem 27 is disposed of as 'Same proof than [17, Theorem 13] with θ=0', and Theorem 30 as 'Same proof than [17, Proposition 16 and Theorem 17] with θ=0', with no derivation. The transfer is not automatic. By Proposition 24, strong star quasiconvexity is equivalent to the restricted secant inequality ⟨∇h(y), y−x*⟩ ≥ (γ/2)||y−x*||². The proof of [17] for strongly quasiconvex functions uses the pairwise implication (10): h(x)≤h(y) ⇒ ⟨∇h(y), x−y⟩ ≤ −(γ/2)||x−y||². Star quasiconvexity gives no such pairwise control when x and y are not on the same ray from x*; a 1D function can satisfy (24) while failing (10) at pairs (x,y) with x<0<y. Indeed, Theorem 30's statement itself says 'strongly quasiconvex', not 'strongly star quasiconvex', which is either a typo or an indication that the claimed extension was not actually proved. Until the Lyapunov proof in [17] is checked against only (24), the paper's claim to unify heavy-ball/Nesterov linear convergence for star quasiconvex functions is unsupported. This is more load-bearing than the acknowledged practical difficulty of prox on star-shaped sets, because it affects the advertised reach of the framework, not just its implementation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces the class of (strongly) star quasiconvex functions, defined by inequality (15), in which every point y is compared with a fixed minimizer xbar along the segment [xbar,y]. It claims this class unifies convex, star-convex, quasiconvex, quasar-convex, and positively homogeneous functions. The main theoretical contributions are: a geometric characterization via star-shaped sublevel sets (Theorem 9), a characterization as quasiconvexity along rays through the minimizer (Theorem 12), a gradient characterization via the restricted secant inequality (24) in the differentiable case (Proposition 24), derived properties such as quadratic growth, 2-supercoercivity, and the Polyak- Lojasiewicz inequality, and a linear convergence result for the proximal point algorithm on closed star-shaped domains (Theorem 33). The paper also claims linear convergence for gradient, heavy-ball, and Nesterov accelerated methods, with proofs delegated to a companion paper [17] by setting a parameter to zero.","tokens_in":17158,"tokens_out":8782,"duration_ms":76468,"significance":"If the main claims hold, the paper provides a genuinely unifying framework: Theorem 33 is, to the authors' knowledge, the first linear convergence result for PPA on nonconvex star-shaped sets, and the characterization results in Section 3 are clean and potentially useful for subsequent work. The PPA proof in Theorem 33 is self-contained, short, and algebraically sound after a few notation fixes, and the paper is honest about limitations: Remark 36 acknowledges that computing proximity operators on star-shaped sets is currently impractical, and Remark 35(c) notes that no function-value convergence rate is available. These are real strengths. However, the paper's advertised reach extends beyond PPA to gradient-type methods, and that portion of the manuscript is not supported by the present text; the main inclusion example also contains a miscalculation. These issues are load-bearing for the paper's central claim as stated in the abstract, but they are local and fixable, so major revision rather than rejection is appropriate.","major_comments":[{"comment":"The proofs of the heavy-ball and Nesterov acceleration results are not provided; each is disposed of with a sentence of the form 'Same proof than [17, Theorem 13] with θ=0' or 'Same proof than [17, Proposition 16 and Theorem 17] with θ=0'. This is not an automatic transfer. The Lyapunov arguments in [17] are built on the pairwise quasiconvex implication (10), which controls arbitrary pairs x,y with h(x)≤h(y), whereas strong star quasiconvexity gives only the restricted secant inequality (24), which controls each point against the fixed minimizer xbar. The manuscript does not show that (24) implies the pairwise control needed by the cited proofs, and the two conditions are not equivalent in general. Moreover, Theorem 30's hypothesis reads 'strongly quasiconvex' rather than 'strongly star quasiconvex', so either the statement is a typo or the claimed extension is not actually proved. Since the abstract and Section 4 advertise linear convergence of heavy-ball and Nesterov accelerations for strongly star quasiconvex functions, this gap is load-bearing and must be fixed, either by supplying the missing argument or by explicitly restricting the claims to the proven cases.","section":"Section 4, Theorems 27 and 30"},{"comment":"The definition of phi(x) in Example 8 is phi(x)=||x||/||x||_p h(||x||). For z=(1/2,1/2), one has ||z||=1/sqrt(2), ||z||_p=alpha/beta, and h(||z||)=h(1/sqrt(2))=beta, so the displayed formula gives phi(z)=beta^2/(sqrt(2) alpha), which is less than alpha and therefore does not demonstrate non-quasiconvexity. The displayed relation phi(z)=(alpha/beta)/(1/sqrt(2)) beta = sqrt(2) alpha > alpha uses the reciprocal ratio ||z||_p/||z|| instead of the stated ||z||/||z||_p. This example is the one used to establish that the inclusion of strongly star quasiconvex functions in the union of quasiconvex and quasar-convex functions is proper, so the claim is currently unsupported as written. The formula and computation must be made consistent, and the star quasiconvexity and non-quasiconvexity of the corrected function must be re-verified.","section":"Section 3.1, Example 8"}],"minor_comments":[{"comment":"The title contains the typo 'an Unified Approach'; it should be 'a Unified Approach'.","section":"Title"},{"comment":"The statement says S_delta(h) for all alpha in R, but the sublevel notation uses delta. Use the same symbol throughout.","section":"Theorem 9"},{"comment":"The formula for E_1 contains the double equality symbol ':==': it should read E_1 := h(x_k) - h^* + (alpha^2/beta)||x_k - x_{k-1}||^2.","section":"Corollary 28"},{"comment":"The sentence 'taking z=x^k, x^*=x^{k+1} and x=x^k in (22)' is inconsistent with Proposition 23, where the third argument is the minimizer xbar. The displayed inequality (33) is correct if x is read as xbar, but the notation must be corrected and the minimizer should be denoted consistently, for instance by xbar, throughout the proof.","section":"Theorem 33 proof"},{"comment":"If the intended extension is to strongly star quasiconvex functions, the hypothesis should read 'strongly star quasiconvex' rather than 'strongly quasiconvex'; as written, the theorem does not state the claimed result.","section":"Theorem 30 statement"},{"comment":"Remark 21(i) cites 'Corollary 18' for 2-supercoercivity, but the result is Proposition 18.","section":"Remark 21"}],"recommendation":"major_revision","confidential_remarks":"The paper relies heavily on two preprints or submitted works, [17] and [34], for results that are load-bearing in Section 4. The authors should clarify the status of these references and either include the transferred proofs or narrow the claims. The central PPA result is sound and novel, but the Section 4 claims and Example 8 must be repaired before the paper is acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know up front: the proximal point theorem is the genuine contribution, and Section 4's heavy-ball/Nesterov claims do not hold up as written. The class itself is not new—Remark 13 admits [34] studied it under a different name, and Proposition 24 is the restricted secant inequality of [44]—but Theorem 33, linear convergence of PPA on closed star-shaped sets, appears to be new and the proof is clean.\n\nWhat works: the sublevel-set characterization (Theorem 9), quasiconvexity along rays (Theorem 12), the fixed-point criteria for the proximity operator (Propositions 22–23), and the 2-supercoercivity result are all solid and useful. The PPA proof is self-contained and algebraically correct. I checked the contraction step; it goes through.\n\nSoft spots, in rough order of severity.\n\nFirst, the transfer from [17] to star quasiconvexity is not established. Theorems 27 and 30 are disposed of with \"Same proof than [17, Theorem 13] with θ=0\" and \"Same proof than [17, Proposition 16 and Theorem 17] with θ=0.\" But [17]'s proofs use the pairwise quasiconvexity implication (10), which gives you control at y for every x with h(x) ≤ h(y). Star quasiconvexity only gives you control along the ray from x̄ to y, i.e., the restricted secant inequality (24). You cannot just set θ=0 and keep the Lyapunov argument unless you show the argument only ever calls (10) with one point at x̄. The paper does not show that. And Theorem 30's statement literally assumes \"strongly quasiconvex,\" not \"strongly star quasiconvex,\" which is either a typo or a sign that the advertised extension was not actually proved. This is load-bearing because the abstract sells gradient-type linear convergence.\n\nSecond, Example 8 has a concrete inconsistency. The definition of φ uses ∥x∥/∥x∥_p, but the displayed computation of φ(z) uses the reciprocal ratio times h(∥z∥_p). As printed, the example does not demonstrate what it claims. It is probably fixable, but it needs fixing.\n\nThird, a minor point: the paper honestly notes in Remark 36 that prox on star-shaped sets is not practical yet. That is a limitation, not a flaw.\n\nVerdict: the PPA result deserves a serious referee. The paper should not be accepted in its current form because the Section 4 claims are unsupported and Example 8 is wrong as written. But the PPA part is real, the exposition is clear, and the authors are candid about prior work. I'd send it out with a request to fix those two issues and re-examine the gradient-method claims.\n\nFor your own reading: read Theorem 33 and Propositions 22–23; skip Section 4 or read it with suspicion.","headline":"The PPA-on-star-shaped-sets theorem is real and clean; the heavy-ball/Nesterov claims in Section 4 are not actually proved.","tokens_in":17662,"tokens_out":5674,"would_cite":true,"duration_ms":45776,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C26","90C25","90C30"],"pacs":[],"model":"deepseek-v4-flash","headline":"Strongly star quasiconvex functions unify convex, quasiconvex, star-convex, and quasar-convex classes, and the proximal point algorithm converges linearly to their unique minimizer on closed star-shaped sets.","keywords":["star quasiconvexity","linear convergence","proximal point algorithm","star-shaped sets","generalized convexity","restricted secant inequality","gradient dominance","nonconvex optimization"],"falsifier":"Run Algorithm 1 on a closed, bounded star-shaped set that is not convex (for example, a four-leaf clover domain) with a strongly star quasiconvex function such as the one constructed in Example 8, and record the ratio $\\|x_{k+1}-\\bar{x}\\|^2/\\|x_k-\\bar{x}\\|^2$. If at any step this ratio exceeds $1/(1+\\beta'\\gamma)$, the contraction (34) and Theorem 33 are false.","tokens_in":16629,"feed_emoji":"📉","tokens_out":14258,"duration_ms":112877,"temperature":0.7,"pith_summary":"This paper introduces one inequality as a common cause for linear convergence guarantees that were previously proven separately for several generalized convexity classes: a function $h$ with minimizer $\\bar{x}$ is strongly star quasiconvex with modulus $\\gamma>0$ if $h(\\lambda\\bar{x}+(1-\\lambda)y)\\le h(y)-\\lambda(1-\\lambda)\\frac{\\gamma}{2}\\|y-\\bar{x}\\|^2$ for every $y$ and $\\lambda\\in[0,1]$. The class contains convex, quasiconvex, star-convex, quasar-convex, and positively homogeneous functions, and it is proper: the paper's Example 8 gives a strongly star quasiconvex function that is neither quasiconvex nor quasar-convex. Star quasiconvexity is characterized geometrically by sublevel sets that are star-shaped at a minimizer, and in the differentiable case strong star quasiconvexity is exactly the restricted secant inequality. The main algorithmic consequence is that the proximal point algorithm, on a closed and star-shaped but not necessarily convex feasible set, converges linearly to the unique minimizer with rate at least $1/(1+\\beta'\\gamma)$. A sympathetic reader should care because the paper offers a single structural condition under which several known linear-convergence results become corollaries of one argument.","feed_headline":"One inequality unifies linear convergence beyond convexity","feed_subtitle":"Star quasiconvexity covers convex, quasiconvex, star-convex, and quasar-convex cases with one linear rate.","key_machinery":"The load-bearing object is the star quasiconvexity inequality (15), together with the geometric fact it encodes: every sublevel set $S_\\delta(h)=\\{x:h(x)\\le\\delta\\}$ is star-shaped at the minimizer $\\bar{x}$, meaning that the segment $[\\bar{x},x]$ lies inside the sublevel set for every $x$ in it. In the differentiable case the same inequality is equivalent to the restricted secant condition $\\langle\\nabla h(y), y-\\bar{x}\\rangle\\ge \\frac{\\gamma}{2}\\|y-\\bar{x}\\|^2$, and the proximal point analysis runs through inequality (22), which bounds the prox step by comparing it with the minimizer along the segment inside the star-shaped set $K$. That segment is what turns a one-dimensional ray argument into a global squared-distance contraction for Algorithm 1.","core_discovery":"The paper's central claim is that the class of strongly star quasiconvex functions is the right umbrella for linear convergence beyond convexity. After fixing the unique minimizer $\\bar{x}$, the defining inequality says that moving from any point $y$ toward $\\bar{x}$ cannot increase $h$ by more than the quadratic penalty $\\lambda(1-\\lambda)\\frac{\\gamma}{2}\\|y-\\bar{x}\\|^2$. Theorem 9 proves that this is equivalent, in the non-strong case, to every sublevel set being star-shaped at $\\bar{x}$; Theorem 12 proves equivalence with quasiconvexity along each ray emanating from $\\bar{x}$; Proposition 24 identifies the differentiable form as the restricted secant inequality $\\langle\\nabla h(y), y-\\bar{x}\\rangle\\ge \\frac{\\gamma}{2}\\|y-\\bar{x}\\|^2$, and Proposition 26 shows Lipschitz-smooth strongly star quasiconvex functions satisfy a gradient-dominance inequality with constant $\\gamma^2/(2L)$. Theorem 33 then derives the contraction $\\|x_{k+1}-\\bar{x}\\|^2\\le \\|x_k-\\bar{x}\\|^2/(1+\\beta'\\gamma)$ for the proximal point algorithm on a closed star-shaped set, giving linear convergence to the unique solution; the same rate is exactly as good as in the strongly convex case.","pith_inferences":["Because strong star quasiconvexity is equivalent to quasiconvexity along every ray from the minimizer, one-dimensional convergence proofs for strongly quasiconvex functions should in principle lift to star-shaped domains, provided the domain is star-shaped at the minimizer; this could simplify future algorithm analysis without reproving contraction estimates from scratch.","The gradient-dominance inequality obtained in Proposition 26 is a known sufficient condition for linear convergence of stochastic and incremental gradient methods, so the framework may extend to randomized first-order algorithms under the same star-shaped sublevel-set condition; the paper itself treats only deterministic methods.","Remark 36 concedes that prox computations on star-shaped sets are largely unexplored. Finding tractable star-shaped sets with computable proximity operators, such as unions of convex cones or balls sharing a common center, would turn Theorem 33 from a theoretical statement into a practical algorithm.","The conclusion's suggested connection to prospect theory is testable: one could check whether utility functions whose sublevel sets are star-shaped at a reference point reproduce reference-dependent behavior such as loss aversion, since star quasiconvexity fixes the minimizer as the reference point."],"forward_implications":["Every convex, quasiconvex, star-convex, quasar-convex, and positively homogeneous function considered in the paper falls into the star quasiconvex framework, so the linear-convergence results for those classes become special cases of a single inequality; the inclusion is proper, as shown by Example 8.","A function is star quasiconvex with respect to a minimizer if and only if all its sublevel sets are star-shaped at that minimizer, giving a purely geometric certificate that can be checked without computing gradients (Theorem 9).","In the differentiable case, strong star quasiconvexity coincides with the restricted secant inequality, and Lipschitz-smooth strongly star quasiconvex functions satisfy a gradient-dominance inequality with constant $\\gamma^2/(2L)$ (Propositions 24 and 26).","The proximal point algorithm on a closed, star-shaped feasible set converges linearly to the unique minimizer with rate at least $1/(1+\\beta'\\gamma)$, matching the strongly convex rate (Theorem 33).","Gradient descent, the heavy-ball method, and the momentum-accelerated gradient method all converge linearly for strongly star quasiconvex functions, with the same rates previously known for strongly quasiconvex functions (Theorems 27 and 30)."],"supporting_citations":[{"why":"Supplies the original star quasiconvex definition and the first-order characterization adopted as Proposition 24.","marker":"[34]"},{"why":"Provides existence and uniqueness results, quadratic growth, and proximal algorithms for strongly quasiconvex functions, the base cases the paper extends.","marker":"[25]"},{"why":"Establishes linear convergence of the proximal point method for quasar-convex functions, the result Theorem 33 extends to star-shaped sets.","marker":"[8]"},{"why":"Defines star-convexity, one of the generalized convexity classes shown to be a special case of star quasiconvexity.","marker":"[33]"},{"why":"Defines quasar-convex functions and gives the differentiable characterization used for gradient-type methods.","marker":"[20]"},{"why":"Supplies the heavy-ball and accelerated-gradient convergence proofs that carry over with a parameter set to zero.","marker":"[17]"},{"why":"Gives the strongly quasiconvex characterizations and quadratic growth facts behind Corollary 16 and Proposition 17.","marker":"[18]"},{"why":"Provides the gradient-method and gradient-dominance analysis for strongly quasiconvex functions mirrored in Section 4.","marker":"[26]"},{"why":"Names the restricted secant condition, the differentiable form of strong star quasiconvexity used in Proposition 24.","marker":"[44]"}],"fun_headline_variants":["Star quasiconvexity: one umbrella for linear convergence","Linear convergence beyond convexity? Meet star quasiconvexity","One inequality to unify linear convergence beyond convexity","Star-shaped sublevel sets: the key to linear convergence","Star quasiconvexity: generalized convexity, linear rates"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is geometric: the feasible set $K$ must be star-shaped at the unique minimizer $\\bar{x}$, so that every segment $[\\bar{x}, x^*]$ used in the proof lies inside $K$; nothing in the function class alone guarantees this, and the paper notes that prox steps on star-shaped sets are still impractical.","fun_headline_variants_meta":{"raw":{"variants":["Star quasiconvexity: one umbrella for linear convergence","Linear convergence beyond convexity? Meet star quasiconvexity","One inequality to unify linear convergence beyond convexity","Star-shaped sublevel sets: the key to linear convergence","Star quasiconvexity: generalized convexity, linear rates"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000933,"raw_usage":{"total_tokens":4024,"prompt_tokens":1009,"completion_tokens":3015,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":625,"completion_tokens_details":{"reasoning_tokens":2933}},"tokens_in":625,"tokens_out":3015,"duration_ms":18625,"temperature":1.0,"reasoning_tokens":2933,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:41:45.620661+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Algorithm 1 on a closed, bounded star-shaped set that is not convex (for example, a four-leaf clover domain) with a strongly star quasiconvex function such as the one constructed in Example 8, and record the ratio $\\|x_{k+1}-\\bar{x}\\|^2/\\|x_k-\\bar{x}\\|^2$. If at any step this ratio exceeds $1/(1+\\beta'\\gamma)$, the contraction (34) and Theorem 33 are false.","supporting_citations":[{"cited_title":"Nguyen, T.Q","cited_arxiv_id":null,"evidence_quote":"Supplies the original star quasiconvex definition and the first-order characterization adopted as Proposition 24."},{"cited_title":"Nesterov, B.T","cited_arxiv_id":null,"evidence_quote":"Defines star-convexity, one of the generalized convexity classes shown to be a special case of star quasiconvexity."},{"cited_title":"Hinder, A","cited_arxiv_id":null,"evidence_quote":"Defines quasar-convex functions and gives the differentiable characterization used for gradient-type methods."},{"cited_title":"Lara, R.T","cited_arxiv_id":null,"evidence_quote":"Provides the gradient-method and gradient-dominance analysis for strongly quasiconvex functions mirrored in Section 4."}],"review_version":2}