{"id":"1a2af3fc-d788-4126-ad2b-c2f48f4df8e5","arxiv_id":"2511.10421","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Under weak smoothness of f and prox-regularity of g, the high-order forward-backward envelope is differentiable and its gradient is Hölder continuous near p-calm points of the composite objective.","lead":"This paper proves regularity properties—boundedness, Hölder continuity, subdifferential formulas, differentiability, and weak smoothness—for the high-order forward-backward envelope, a smoothing surrogate for nonconvex composite optimization. It identifies conditions under which the associated splitting map is single-valued and continuous, which is what gradient-based methods need.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 31(b) proof contains an unjustified Hölder estimate for the linearization gradient, so the claimed exponent η is not established by the argument as written.","rationale":"The reader's verdict is CONDITIONAL, with the weakest assumption identified as Assumption 29 (global p-calmness plus prox-regularity). I agree that the theorems are conditional on this assumption and that it is the main scope limitation. However, the most load-bearing internal weakness is the unjustified inequality in Eq. (4.17) in the proof of Theorem 31(b). This estimate is used to establish the claimed Hölder order of the gradient of the HiFBE, which is part of the central claim. The theorem's conclusion may still be true with a weaker exponent or with a corrected proof — the correct local order for the first term appears to be min(μ, ν/2), which would imply the stated η is a valid (possibly suboptimal) Hölder order — but the argument as written is not rigorous. This warrants a conditional verdict rather than acceptance as-is, because the proof of a key theorem must be corrected. The typo in the final display (μ instead of η) should also be fixed. My assessment does not change the reader's CONDITIONAL verdict, but it adds a new technical reason for it.","tokens_in":23178,"tokens_out":31939,"duration_ms":288504,"concrete_test":"Independently re-derive Eq. (4.17). Concretely, take n = 1, f(x) = x² (so f ∈ C^{2,μ} for any μ ∈ (0,1) with any Lμ > 0), and choose x1 = 0, y1 = 0, x2 = ε, y2 = 2ε^{ν/2} — a choice compatible with the Hölder bound (4.8). For μ = 0.9, ν = 0.5, and ε = 10^{-8}, the left side of (4.17) is ≈ 4×10^{-2} while the right side is ≈ (2×10^{-2})^{0.9} ≈ 2.8×10^{-2}, so the displayed inequality fails. Alternatively, verify whether the inequality can be derived from the decomposition above; if not, the proof of Theorem 31(b) is incomplete and needs a corrected estimate or a revised exponent.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The main differentiability result (Theorem 31(b)) depends on Eq. (4.17), which asserts\n\n∥∇²f(x2)(y2−x2) − ∇²f(x1)(y1−x1)∥ ≤ Lμ ∥x2−x1 + y2−y1∥^μ.\n\nThis is not a consequence of f∈C^{2,μ}. The natural decomposition is\n\n(∇²f(x2)−∇²f(x1))(y2−x2) + ∇²f(x1)((y2−y1)−(x2−x1)),\n\nwhich gives terms of order ∥Δx∥^μ and ∥Δy∥+∥Δx∥. Using the ν/2-Hölder continuity of T (4.8), the best local bound for this term is of order min(μ, ν/2), not μν/2 as claimed. A Lipschitz function is locally μ-Hölder only with a constant depending on the diameter; the displayed inequality would require a joint μ-Hölder property of the mapping (x,y) ↦ ∇²f(x)(y−x) that does not follow from the assumptions. This is a concrete proof gap in the central weak-smoothness statement. The final display of Theorem 31(b) also has an exponent typo (μ instead of η), but the main issue is the invalid estimate used to derive that exponent.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a variational-analytic theory of the high-order forward-backward splitting mapping (HiFBS) and high-order forward-backward envelope (HiFBE) for composite objectives φ=f+g in which f is weakly smooth, i.e. ∇f is ν-Hölder, and g is proper lsc (possibly nonconvex). The main results are: boundedness and uniform boundedness of HiFBS, Hölder (and, for p=2, Lipschitz) continuity of HiFBE, explicit Fréchet/limiting subdifferential formulas, a necessary-and-sufficient differentiability characterization, and — under p-calmness and prox-regularity of g at a calm point — local single-valuedness/continuity of HiFBS and C^1 weak smoothness of HiFBE with explicit Hölder exponents. The bridge is the identity ∇φ^p_γ(x)=∇²f(x)(T^p_γφ(x)-x)+γ^{-1}∥x-T^p_γφ(x)∥^{p-2}(x-T^p_γφ(x)), which converts single-valuedness of HiFBS into differentiability of HiFBE. Theorems 30 and 31 are the headline results.","tokens_in":23526,"tokens_out":20577,"duration_ms":186469,"significance":"If the main results hold, the paper substantially extends the classical forward-backward envelope framework from Lipschitz-smooth f to weakly smooth nonconvex composite problems, providing a principled basis for gradient-type algorithms with high-order regularization. The paper is careful about hypotheses, contains explicit constants, and many auxiliary results (Proposition 18, Theorem 24, Theorem 28) are independently valuable. The proofs are largely detailed and rigorous, and the p-calmness/prox-regularity hypotheses are genuine extra assumptions rather than disguised restatements of the conclusions. However, the proof of the central weak-smoothness statement in Theorem 31(b) contains an invalid Hölder estimate; the theorem is likely repairable, but the argument as written does not establish the claimed result.","major_comments":[{"comment":"The displayed estimate ||∇²f(x2)(y2−x2) − ∇²f(x1)(y1−x1)|| ≤ L_μ||x2−x1+y2−y1||^μ is not a consequence of f∈C^{2,μ}. For a quadratic f with constant Hessian A, L_μ=0 while the left side is ||A[(y2−y1)−(x2−x1)]||, which is generally nonzero. The natural decomposition (∇²f(x2)−∇²f(x1))(y2−x2) + ∇²f(x1)((y2−y1)−(x2−x1)) gives a valid local bound of order ||x2−x1||^μ + ||x2−x1||^{ν/2}, i.e. Hölder order min(μ, ν/2). Since the stated exponent η=(ν/2)min{μ,ν} is no larger than min(μ,ν/2), Theorem 31(b) is plausibly salvageable by replacing this step, but as written the proof of the central weak-smoothness claim is not valid.","section":"§4, Theorem 31(b), Eq. (4.17)"}],"minor_comments":[{"comment":"The last display writes ||∇φ^p_γ(x2)−∇φ^p_γ(x1)|| ≤ L_μ||x2−x1||^μ; the exponent and constant should be L_η||x2−x1||^η as stated before the display.","section":"§4, Theorem 31(b), final display"},{"comment":"The symbol L_ν is reused for both the given Hölder constant of ∇f and the newly constructed Hölder constant of φ^p_γ in the proof. Use L̄_ν or another symbol for the envelope constant to avoid confusion.","section":"§3.1, Theorem 20"},{"comment":"The proofs of Theorem 24 and Theorem 30 set x̄=0 and φ(x̄)=0. This is presumably a translation and normalization, but it is not stated explicitly; please clarify that the results hold for a general p-calm point x̄.","section":"§3.2 and §4, Assumption 29"},{"comment":"Several foundational facts used throughout the paper (well-definedness of HOME/HOPE, basic properties of HiFBE/HiFBS) are quoted from the authors' preprints [23,24,26]. Since these are not yet peer-reviewed, the presentation would be more self-contained if those results were stated with proofs or at least with precise pointers to the relevant statements.","section":"Preliminaries, Facts 10, 14, 16"}],"recommendation":"major_revision","confidential_remarks":"The main obstacle is the invalid estimate in Eq. (4.17) of Theorem 31(b). My assessment is that the theorem can be repaired with a standard decomposition and the resulting local Hölder exponent is at least as good as the stated η, so this is a major but not fatal gap. The paper also relies heavily on the authors' own unpublished preprints for definitions and key facts; if the journal's policy is strict on self-citations to preprints, the authors should be asked to supply the necessary statements. No other concerns about novelty or integrity arose."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Thanks for the HiFBE preprint. Quick take: the paper does real work—boundedness, Hölder continuity, subdifferential formulas, differentiability characterization—and most of it checks out. But the weak-smoothness result (Theorem 31b) has a concrete gap in the proof that the authors need to address before I'd trust it.\n\nWhat's new: beyond the authors' earlier preprints, they give the full Fréchet subdifferential of the envelope, the necessary-and-sufficient differentiability condition, and the local single-valuedness/Hölder continuity of the HiFBS under p-calmness plus prox-regularity. I checked the main inequalities in Theorems 20 and 30 and they're fine. The setup p=1+ν is a natural fit for weakly smooth f, and the paper gives a coherent toolkit for gradient-type methods on the envelope.\n\nThe soft spot is Theorem 31(b). The proof's inequality (4.17) claims that ∥∇²f(x2)(y2−x2)−∇²f(x1)(y1−x1)∥ ≤ Lμ∥x2−x1+y2−y1∥^μ. That does not follow from f∈C^{2,μ}. The natural decomposition gives a term of order ∥x2−x1∥^μ plus a term of order ∥y2−y1∥, so the best you can get locally is O(∥Δx∥^{min(μ,ν/2)}), not O(∥Δx∥^{μν/2}). The inequality is actually false for a quadratic f, where Lμ can be zero but the left-hand side isn't. So the claimed Hölder exponent η=ν/2 min{μ,ν} is not established. The final display also writes μ instead of η—minor but confusing.\n\nThis doesn't kill the paper. The differentiability result (Theorem 31a) relies on Theorem 30, which looks solid, and the earlier regularity theorems stand. But the weak-smoothness claim is a headline contribution, so it needs a correct proof or a corrected (probably smaller) exponent. The p-calmness/prox-regularity assumptions are also genuinely restrictive; they're stated, but a referee should ask for a discussion of how often they hold in applications.\n\nWho's this for? People building high-order proximal algorithms for nonconvex composite problems. It deserves a serious referee and a major revision, not a desk reject. I'd send it out, with the request to fix Theorem 31(b) before publication.","headline":"Solid analytic work on high-order forward-backward envelopes, but the weak-smoothness theorem has a genuine proof gap that needs fixing.","tokens_in":23987,"tokens_out":7035,"would_cite":false,"duration_ms":65373,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C26","65K05","49J52","90C30","49M27"],"pacs":[],"model":"deepseek-v4-flash","headline":"For composite functions with weakly smooth f, the high-order forward-backward envelope is differentiable near calm points.","keywords":["nonconvex composite optimization","high-order forward-backward envelope","high-order forward-backward splitting","weakly smooth functions","prox-regularity","p-calmness","Hölder continuity","scaled gradient method"],"falsifier":"Construct a function with f∈C^{1,ν}_{Lν}, g prox-regular at a p-calm point, yet T^p_γφ has multiple selections arbitrarily close to the calm point, or compute ∇φ^p_γ and show it violates the claimed ν/2-Hölder bound on arbitrarily small neighborhoods. Concretely, test f(x)=|x|^{3/2} composite with a prox-regular but nonconvex g, and check numerically whether the residual map x−T^p_γφ(x) is pointwise unique as γ→0.","tokens_in":23075,"feed_emoji":"🧮","tokens_out":3712,"duration_ms":33386,"temperature":0.7,"pith_summary":"This paper is about a smoothing mechanism for nonconvex composite problems φ=f+g where f has a Hölder-continuous gradient rather than a Lipschitz one. The authors study the high-order forward-backward splitting map (HiFBS) and its associated envelope (HiFBE), a p-power generalization of the classical forward-backward envelope. Their central claim: if g is prox-regular and the point of interest is p-calm, then the HiFBS map is locally single-valued and Hölder continuous, and the HiFBE is continuously differentiable with a gradient that is Hölder continuous with an explicitly quantified exponent. Since the HiFBE shares the infimum with φ, this gives a smooth surrogate on which gradient methods can be run, even though the original problem is nonsmooth and nonconvex. The differentiability order is derived explicitly as η=(ν/2)min{µ,ν}.","feed_headline":"Weakly smooth nonconvex problems gain a differentiable envelope","feed_subtitle":"Prox-regularity plus p-calmness makes the high-order forward-backward envelope smooth near calm points.","key_machinery":"The central object is the pair (T^p_γφ, φ^p_γ): the high-order forward-backward splitting map defined by minimizing f(x)+⟨∇f(x), y−x⟩+g(y)+(1/pγ)‖x−y‖^p, and the envelope φ^p_γ(x) defined as that minimum value. The mechanism that carries the argument is the identity connecting the envelope gradient to the residual x−y, together with p-calmness (a local sharpness inequality) and prox-regularity of g, which yield a coercive quadratic bound on the map. This converts the set-valued stability of the splitting map into single-valuedness, Hölder continuity, and ultimately C^1 regularity of the envelope.","core_discovery":"The core discovery is a two-step bridge: local single-valuedness and Hölder continuity of HiFBS imply differentiability of HiFBE. Under f∈C^{1,ν}_{Lν}, with g prox-regular at a p-calm point x̄ and −∇f(x̄)∈∂g(x̄), there exists γ̄>0 such that for γ∈(0,γ̄) the map T^p_γφ is single-valued, continuous and ν/2-Hölder on a neighborhood of x̄; consequently φ^p_γ is C^1 there, and when f∈C^{2,µ} its gradient is Hölder of order η=(ν/2)min{µ,ν}. The gradient formula is ∇φ^p_γ(x̄)=∇²f(x̄)(ȳ−x̄)+γ^{-1}‖x̄−ȳ‖^{p−2}(x̄−ȳ) with ȳ=T^p_γφ(x̄), so smoothness of the envelope is transported from regularity of the splitting map.","pith_inferences":["A testable extension: compute the HiFBE on weakly smooth functions with known ν and µ and verify numerically that the gradient's Hölder exponent matches η; mismatches would reveal constants or assumptions needing refinement.","A natural next step the authors leave implicit is whether prox-regularity can be relaxed to a weaker local growth condition, which would widen the class of nonsmooth g for which the envelope is smooth.","The p-calmness condition is a sharpness/growth requirement; in unconstrained minimization it holds at isolated minimizers, but checking it at non-minimizing points may be delicate—an extension to p-calm neighborhoods rather than points would strengthen the theory."],"forward_implications":["The iterative high-order forward-backward algorithm is equivalent to a scaled gradient method on the HiFBE, so convergence guarantees for gradient methods can be transferred to the nonconvex composite setting.","For the classical p=2 case, the differentiability result recovers the known FBE gradient formula ∇φ_γ=Q_γ(x)R_γ(x) as a special case.","For f=0 the HiFBE reduces to the high-order Moreau envelope, and the gradient method on it has the form of a scaled gradient step, unifying the theory.","The Hölder exponent η=(ν/2)min{µ,ν} gives a concrete, quantitative smoothness certificate that can be checked or exploited in complexity analyses.","If g is prox-regular at x̄, the local minimizer of the HiFBS subproblem is unique near the calm point, so the envelope is well-defined as a single-valued function."],"fun_headline_variants":["Prox-regularity plus p-calmness yields C1 envelope","High-order forward-backward envelope turns differentiable near calm points","Hölder continuity of splitting map gives differentiable envelope","Nonconvex composite optimization: differentiability from Hölder continuity"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The headline differentiability result requires that the nonsmooth term g be prox-regular at the calm point and that the function satisfy a global p-calmness inequality φ(x)+M‖x−x̄‖^p > φ(x̄); if either of these fails, the local single-valuedness of HiFBS—and therefore the C^1 property of the envelope—is no longer guaranteed.","fun_headline_variants_meta":{"raw":{"variants":["Prox-regularity plus p-calmness yields C1 envelope","High-order forward-backward envelope turns differentiable near calm points","Hölder continuity of splitting map gives differentiable envelope","Nonconvex composite optimization: differentiability from Hölder continuity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001972,"raw_usage":{"total_tokens":7558,"prompt_tokens":782,"completion_tokens":6776,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":526,"completion_tokens_details":{"reasoning_tokens":6718}},"tokens_in":526,"tokens_out":6776,"duration_ms":47693,"temperature":1.0,"reasoning_tokens":6718,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T22:27:02.311718+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a function with f∈C^{1,ν}_{Lν}, g prox-regular at a p-calm point, yet T^p_γφ has multiple selections arbitrarily close to the calm point, or compute ∇φ^p_γ and show it violates the claimed ν/2-Hölder bound on arbitrarily small neighborhoods. Concretely, test f(x)=|x|^{3/2} composite with a prox-regular but nonconvex g, and check numerically whether the residual map x−T^p_γφ(x) is pointwise unique as γ→0.","supporting_citations":[],"review_version":1}