{"id":"9081a383-1f52-48b1-822f-314ca0e55dc8","arxiv_id":"2508.18764","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A unified implicit-gradient-flow framework, Gravidy, encodes orthant, simplex, box, and Stiefel constraints into the dynamics and proves that stationarity recovers KKT conditions.","lead":"This paper introduces Gravidy, a framework for constrained optimization that builds constraints into the dynamics via reparameterizations and implicit discretizations. It derives KKT conditions from stationarity of these flows and provides convergence guarantees for convex, strongly convex, and nonconvex objectives.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Discrete-time KKT for simplex/box is not proven: strict-interior prox steps have zero complementarity multipliers, so limiting KKT inequalities require an unproven log-ratio analysis.","rationale":"The reader's weakest_assumption focused on inexact inner solves and variable-metric drift. My concern is more basic: even for exact solves, the discrete-time KKT derivation for the simplex and box is incomplete because strict interiority forces the finite-step multipliers to zero, and the limit passage is asserted rather than proved. This directly undermines the headline claim that convergent implicit sequences recover KKT. The orthant exponential case (15) has a valid direct argument, but the softmax/KL-prox and two-sided-entropy cases do not. The variable-metric drift gap is also real but concerns convergence rather than the KKT-from-stationarity mechanism itself, so I rank it secondary. The issue is likely fixable by a careful asymptotic analysis of r_{k,i}; hence I do not recommend rejection, only maintaining the conditional verdict.","tokens_in":35390,"tokens_out":21861,"duration_ms":232165,"concrete_test":"Analyze the exact simplex KL-prox update (18). For a convergent sequence x_k→x* with x* on the boundary, compute r_{k,i}=log x_{k+1,i}-log x_{k,i} for each coordinate with x*_i=0. Prove that limsup_k r_{k,i} ≤ 0 for any C^1 f and any initial interior point, or construct a C^1 f and initial point where this limsup is positive. The first outcome completes the missing KKT proof; the second disproves the discrete KKT claim. A computational trace of r_{k,i} from Algorithm 2 on a boundary-minimizer example (e.g., f(x)=x_2 on Δ_3) can serve as a quick indicator.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"In Section 5.2 B, the simplex KL-prox keeps all finite iterates strictly interior, so the nonnegativity multipliers λ_{k+1} in the finite-step KKT system are identically zero. The paper then asserts that the limiting KKT conditions 'in the limit reproduce the continuous-time conditions above' without proof. At a boundary limit x* with x*_i=0, the term r_{k,i}=log x_{k+1,i}-log x_{k,i} need not tend to 0; its asymptotic sign controls the limiting KKT inequality ∇f_i(x*) ≥ -ν*. No argument establishes that limsup r_{k,i} ≤ 0, and no construction of the limiting multipliers is given. The same gap appears for the box in Section 5.2 C, where the two-sided entropy keeps iterates strictly inside the box and the bound multipliers are zero at every finite step. The convergence theorems in Section 7 establish convergence to a critical point of the Bregman prox, not KKT of the constrained problem. Because the implemented algorithms in Section 6 use exactly these strict-interior prox steps, the claimed KKT guarantees for discrete sequences are unsupported for the simplex and box.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a unified geometry-based framework for constrained optimization: feasible sets are encoded via reparameterizations whose Jacobians vanish at the boundary (orthant, box), via the Fisher–Shahshahani/KL geometry (simplex), or via the canonical Riemannian structure (Stiefel). The continuous-time dynamics are shown to be gradient-like, and the paper argues that stationarity of these dynamics implies KKT conditions. Implicit (backward-Euler/Bregman-prox/Cayley) discretizations are introduced, with convergence results in convex, strongly convex, and nonconvex (KL) settings, and numerical experiments compare the resulting algorithms against standard baselines. The paper also states that convergent implicit sequences generated from the interior recover KKT when stepsizes are bounded away from zero.","tokens_in":35667,"tokens_out":7311,"duration_ms":87522,"significance":"If the main claims hold, the paper would provide a useful unifying perspective: a single kinematic mechanism (vanishing Jacobian or rank-deficient Fisher operator) explains complementary slackness across four constraint geometries, and the implicit schemes have strong stability and convergence properties. The Bregman-proximal convergence theory in Sections 7.2–7.5 is standard and appears correct; the continuous-time KKT derivations are elegant, and the Stiefel Cayley construction is a clean implicit-flow interpretation of existing feasible methods. The paper also provides openly available code and reproducible experiments, which is a notable strength. However, two load-bearing parts of the paper are not yet supported: the discrete-time KKT recovery for simplex/box and the variable-metric convergence theorem as stated. These are fixable but require nontrivial additional analysis or explicit conditions.","major_comments":[{"comment":"The discrete-time KKT claim for the simplex and box is not proven. For the KL-prox step (17)–(18), all finite iterates are strictly interior, so the nonnegativity multipliers λ_{k+1} in the stated KKT system are identically zero. The paper asserts that the limit 'reproduces the continuous-time conditions above' (Eq. 18 and following text), but no argument is given for the limiting inequalities ∇f_i(x*) ≥ τ when x*_i = 0. This requires controlling limsup (log x_{k+1,i} − log x_{k,i})/η_k, which need not tend to 0 and whose sign determines the limiting multiplier. The same gap appears in §5.2 C for the two-sided entropy on a box. Since the algorithms in §6 use exactly these strict-interior prox steps, the claimed KKT recovery for convergent implicit sequences is unsupported for the simplex and box.","section":"§5.2 B and C"},{"comment":"The summed variable-metric inequality contains a positive metric-drift term +(Bτ/2η) Σ ||x_{k+1}−x_k||² that must be dominated by the negative term −(m/2η̄) Σ ||x_{k+1}−x_k||² for the conclusion to yield convergence. No such dominance condition is stated or verified; the theorem only assumes m, M, τ and bounded stepsizes. Without e.g. m η > B τ η̄, Eq. (32) is compatible with divergent increments, and the 'o(1)' in Corollary 7.10 does not rescue the argument because summability of ||x_{k+1}−x_k||² is part of what needs to be proved. This gap affects the claimed global convergence guarantees for the variable-metric (reparameterized orthant/box) constructions.","section":"Theorem 7.8, Eq. (32)"},{"comment":"Proposition 7.2 transfers convergence rates to inexact inner solves under a summable-residual condition, but it does not establish that limit points of the inexact implicit updates satisfy KKT. The implemented algorithms in §6 (MGN for orthant/box, Newton–KKT with tolerance for the simplex, Newton–Krylov for Stiefel) compute approximate solves. Even if their residuals are summable, the KKT characterization is a limit statement about the discrete scheme, and no theorem connects the inexact iterates to the KKT system. Thus the practical algorithms, as implemented, do not inherit the KKT guarantee stated in the abstract without an additional argument.","section":"§7.5 and §6 inner solvers"}],"minor_comments":[{"comment":"The arXiv abstract describes 'Gravidy' while the full-text title and abstract describe a different framing; the two versions should be harmonized before publication.","section":"Title/abstract"},{"comment":"Internal references to 'Appendix 8', 'Appendix 14', 'Appendix 7', and 'Appendix 9' do not match the labeled appendices (A–F) in the compiled text.","section":"Appendix references"},{"comment":"The displayed linearization '˙δ = −G ⋆−1H⋆ [G] δ' is garbled; the intended generator/metric structure should be written clearly.","section":"§2.3.2"},{"comment":"In the 1D KKT check for c<0, the paper states 'choose λ⋆ = ab'; with the paper's convention ∇f(x*) − λ* = 0, the correct multiplier is λ* = −ab > 0. This is a sign error in the example.","section":"Appendix A"},{"comment":"Table 1(a) reports a final KKT residual of 5.86e-05 for Gravidy-Pos (MGN), yet Table 1(b) reports a time to tolerance ≤ 1e-8 for the same method. The inconsistency should be resolved or explained.","section":"Table 1"},{"comment":"The Stiefel discrete-time statement says 'if the inner Newton solve drives the residual to zero' the limit satisfies KKT; this quantifier is informal and should be replaced by a precise statement about finite inexact solves.","section":"§5.2 D"}],"recommendation":"major_revision","confidential_remarks":"The core continuous-time geometry and the Bregman-prox convergence machinery are sound and potentially publishable after revision. The discrete-time KKT claims and the variable-metric theorem are currently overclaimed relative to the proofs provided. I would not recommend rejection, since the gaps appear addressable with additional lemmas (e.g., log-ratio asymptotics for KL-prox and a dominance condition in the drift estimate). The paper would also benefit from a clearer separation between exact prox results and inexact-solve results."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a serious paper worth refereeing, but the advertised discrete-time KKT guarantees for simplex and box outrun the proofs. The continuous-time story is clean: vanishing Jacobians or the Fisher–Shahshahani operator make complementarity fall out of stationarity, and the Bregman proximal point machinery gives solid convergence and rate results. The non-pullback metric choice G(u)=Jg(u) is a modest but genuine design idea, and the Stiefel Cayley update as trapezoidal integration is a nice unifying perspective.\n\nWhere it holds up: Sections 7.2–7.5 and the appendix proofs are standard and correct; the convex/strongly convex rates and the KL analysis for nonconvex are fine. The paper is honest about the continuous-time derivations and doesn't oversell novelty. Code and experiments are there, though numerical reporting is sloppy (missing figure captions, Table 1 inconsistencies).\n\nWhere the soft spots are, in order of severity:\n\n1. Discrete-time KKT for simplex and box (Section 5.2 B/C) is not proven. The KL-prox and two-sided entropy steps keep iterates strictly interior, so the nonnegativity multipliers are identically zero at every finite step. The paper handwaves 'in the limit reproduces the continuous-time conditions above.' The stress-test note is right: you need a log-ratio argument showing limsup of log x_{k+1,i} - log x_{k,i} ≤ 0 at boundary limits, and no such argument appears. The same gap applies to the box. This matters because the implemented algorithms use exactly these steps, so the claim 'convergent implicit sequences recover KKT' is unsupported for those cases. The orthant discrete proof via the multiplicative update is informal but fixable; the simplex/box needs real work.\n\n2. Theorem 7.8 (variable metrics) has a gap: the drift term scales as (Bτ/2η)Σ||x_{k+1}-x_k||², which needs m/(2η̄) > Bτ/(2η) to get descent. That condition is neither stated nor verified. Probably fixable, but as written the theorem doesn't establish convergence.\n\n3. Minor: the paper's title and abstract claim more than the theory delivers; the 'Gravidy' naming is a lot of branding for a modest design choice.\n\nWho should read it: people working on implicit methods, mirror descent, and geometry-aware optimization. It's a useful synthesis with a teachable gap.\n\nRecommendation: send to peer review. A good referee can demand the missing discrete KKT proof or a rephrased claim. The continuous-time contribution and the Bregman analysis justify the review.","headline":"Useful synthesis with a real gap: continuous-time KKT from stationarity works, but the discrete-time simplex/box KKT claim is unproven as written.","tokens_in":36123,"tokens_out":1934,"would_cite":true,"duration_ms":19392,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C30","90C25","65K10"],"pacs":[],"model":"deepseek-v4-flash","headline":"Stationary flows satisfy KKT with feasibility built in","keywords":["constrained optimization","implicit gradient flows","KKT conditions","complementary slackness","Bregman proximal point","Stiefel manifold","A-stability","reparameterization"],"falsifier":"On a small NNLS problem, run the orthant method with a deliberately loose inner tolerance for many outer iterations and monitor the KKT residual: if it stalls above zero while the exact-update version drives it to zero, the transfer of the KKT characterization to inexact solvers is false. Separately, build a convex problem with a rapidly oscillating metric (large drift) and check whether iterates converge to a non-stationary point, which would show the drift condition in the variable-metric theorem is genuinely necessary.","tokens_in":35294,"feed_emoji":"📐","tokens_out":9685,"duration_ms":103927,"temperature":0.7,"texified_at":"2026-08-05T19:59:06.539304+00:00","pith_summary":"This paper argues that constraints in continuous-time optimization should be encoded into the geometry of the flow rather than enforced by projection or penalty. For the nonnegative orthant and boxes, it uses componentwise reparameterizations whose Jacobians vanish on the boundary; for the simplex, the rank-deficient Fisher-Shahshahani operator $\\operatorname{diag}(x)-xx^T$; for orthogonality constraints, the canonical Riemannian structure of the Stiefel manifold. The payoff is a kinematic mechanism: a trajectory that comes to rest can only rest at points where the gradient's normal component vanishes, which is exactly complementary slackness, so stationarity of the dynamics implies the Karush-Kuhn-Tucker conditions. The same logic is shown for the implicit (A-stable) discretizations, which are Bregman proximal steps and therefore come with monotone descent, $O(1/k)$ convex rates, linear contraction under relative strong convexity for any positive step size, and Kurdyka-Lojasiewicz convergence in the nonconvex case.","texify_model":"deepseek-v4-flash","texify_usage":{"total_tokens":6064,"prompt_tokens":846,"completion_tokens":5218,"prompt_tokens_details":{"cached_tokens":0},"prompt_cache_hit_tokens":0,"prompt_cache_miss_tokens":846,"completion_tokens_details":{"reasoning_tokens":4446}},"feed_headline":"Stationary flows satisfy KKT with feasibility built in","feed_subtitle":"Backward-Euler, KL-prox and Cayley steps give convergent constrained optimizers with no step cap.","key_machinery":"The central object is the constraint-induced geometry: the Jacobian-induced metric $G(u)=J_g(u)$ on vector domains (chosen over the pullback metric to keep the u-dynamics simple and to let the Jacobian vanish exactly at active faces), the Fisher-Shahshahani operator $G(x)=\\operatorname{diag}(x)-xx^T$ on the simplex, and the canonical metric with skew field $A(X)=\\nabla\\Phi(X)X^T-X\\nabla\\Phi(X)^T$ on the Stiefel manifold. This object does three jobs: it keeps every continuous and discrete trajectory feasible by construction; it turns stationarity into KKT through null-speed and rank-deficiency arguments; and it makes implicit discretizations (backward Euler, KL-prox, Cayley update) into Bregman-proximal or resolv","core_discovery":"At the center is the 'kinematic slackness' mechanism. Feasibility is built into the flow by $x=g(u)$; with the Jacobian-induced metric, the primal velocity is $\\dot{x}=-J_g(u)\\nabla f(x)$. Since $J_g$ vanishes on active faces or has rank n-1 with kernel $\\operatorname{span}\\{1\\}$ on the simplex, vanishing velocity forces the normal component of the gradient to vanish at the limit: for the exponential map this is written as the identity $x_i \\nabla_i f(x) = -\\dot{x}_i$, i.e., complementary slackness is a kinematic statement. The paper proves that the same conclusion holds for convergent implicit sequences whose stepsizes are bounded away from zero; on the orthant the discrete update is exactly the multiplicative update","pith_inferences":["The null-speed mechanism should transfer to any smooth reparameterization whose Jacobian loses rank exactly on the active set; a testable extension is to cones or product sets built from smooth inequality constraints.","The variable-metric theorem's drift condition suggests a practical safeguard: growth of the step size should be gated by the ratio ||G_{k+1}-G_k|| / ||x_{k+1}-x_k||, giving a computable certificate for the no-stepsize-cap regime.","The implicit multiplicative update on the orthant identifies the KKT mechanism of classical multiplicative updates and may extend their convergence theory beyond nonnegative least squares to general smooth objectives.","Viewing the Stiefel Cayley step as a trapezoidal discretization points toward higher-order implicit integrators (BDF-type) as a route to accelerated feasible manifold methods; the paper itself lists inertial and accelerated variants as open directions."],"forward_implications":["On orthant and box domains, backward Euler in the reparameterized variable is a Bregman-prox step, so convex objectives decrease monotonically and converge at O(1/k) with no restriction on step size.","On the simplex, the implicit KL-prox step avoids the rank-deficient Jacobian in the inner solve; the Newton-KKT system is SPD, and convergent interior sequences satisfy the simplex KKT conditions.","On the Stiefel manifold, the implicit Cayley update preserves X^T X=I exactly, and a zero Riemannian-gradient limit satisfies the first-order optimality condition for orthogonality constraints.","When inner solves are inexact but residuals are summable, the convex and nonconvex rates of the exact implicit scheme survive.","For nonconvex objectives satisfying Kurdyka-Lojasiewicz, the Bregman scheme converges to a critical point with the standard finite/linear/sublinear rate trichotomy; under a Bregman-PL condition it is linearly convergent for any positive step size."],"supporting_citations":[{"why":"Supplies the Riemannian optimization background (tangent spaces, retractions) used throughout the Stiefel treatment.","marker":"[1]"},{"why":"Provides the canonical Riemannian gradient formula whose vanishing is taken as the Stiefel stationarity condition.","marker":"[17]"},{"why":"Feasible Cayley-type update with line search; serves as the main Stiefel baseline and as the discrete Cayley transform the paper derives from an implicit flow.","marker":"[42]"},{"why":"Proximal point method; backward Euler and Bregman prox steps are analyzed as resolvents, giving global convergence results.","marker":"[34]"},{"why":"Mirror descent with entropy/KL geometry; supplies the multiplicative-update structure and the entropic baseline on the simplex.","marker":"[6]"},{"why":"Multiplicative updates for nonnegative least squares; the orthant implicit step coincides with this exact update form, and it serves as a baseline.","marker":"[24]"},{"why":"Kurdyka-Lojasiewicz convergence framework for proximal-type methods, used for the nonconvex analysis.","marker":"[2]"},{"why":"Defines A-stability of one-step integrators; justifies choosing backward Euler and Cayley updates for unconditional stability.","marker":"[16]"},{"why":"Modified Gauss-Newton with Levenberg-Marquardt damping; used as the inner solver for orthant/box implicit residuals.","marker":"[31]"}],"fun_headline_variants":["Feasibility built in: stationarity yields KKT","Geometry-aware flows: KKT from kinematics","Constraint-encoded updates: no step caps, KKT guaranteed","Implicit geometry: stationarity recovers KKT","Large steps, built-in feasibility, KKT at limit"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The discrete-time KKT result assumes the inner equations are solved exactly; when they are only solved approximately, the paper shows rates are preserved but does not prove the limit still satisfies KKT, and the variable-metric proof needs the coordinate system to change slowly, a condition it assumes without checking.","fun_headline_variants_meta":{"raw":{"variants":["Feasibility built in: stationarity yields KKT","Geometry-aware flows: KKT from kinematics","Constraint-encoded updates: no step caps, KKT guaranteed","Implicit geometry: stationarity recovers KKT","Large steps, built-in feasibility, KKT at limit"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000215,"raw_usage":{"total_tokens":1329,"prompt_tokens":870,"completion_tokens":459,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":614,"completion_tokens_details":{"reasoning_tokens":381}},"tokens_in":614,"tokens_out":459,"duration_ms":5633,"temperature":1.0,"reasoning_tokens":381,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T16:13:29.058278+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a small NNLS problem, run the orthant method with a deliberately loose inner tolerance for many outer iterations and monitor the KKT residual: if it stalls above zero while the exact-update version drives it to zero, the transfer of the KKT characterization to inexact solvers is false. Separately, build a convex problem with a rapidly oscillating metric (large drift) and check whether iterates converge to a non-stationary point, which would show the drift condition in the variable-metric theorem is genuinely necessary.","supporting_citations":[{"cited_title":"Absil, R","cited_arxiv_id":null,"evidence_quote":"Supplies the Riemannian optimization background (tangent spaces, retractions) used throughout the Stiefel treatment."},{"cited_title":"Edelman, T.A","cited_arxiv_id":null,"evidence_quote":"Provides the canonical Riemannian gradient formula whose vanishing is taken as the Stiefel stationarity condition."},{"cited_title":"Wen and W","cited_arxiv_id":null,"evidence_quote":"Feasible Cayley-type update with line search; serves as the main Stiefel baseline and as the discrete Cayley transform the paper derives from an implicit flow."},{"cited_title":"Monotone operators and the proximal point algorithm.SIAM journal on control and optimization, 14(5):877–898, 1976","cited_arxiv_id":null,"evidence_quote":"Proximal point method; backward Euler and Bregman prox steps are analyzed as resolvents, giving global convergence results."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Kurdyka-Lojasiewicz convergence framework for proximal-type methods, used for the nonconvex analysis."},{"cited_title":"A special stability problem for linear multistep methods.BIT Numerical Mathematics, 3(1):27–43, 1963","cited_arxiv_id":null,"evidence_quote":"Defines A-stability of one-step integrators; justifies choosing backward Euler and Cayley updates for unconditional stability."},{"cited_title":"Modified gauss–newton scheme with worst-case guarantees for its global performance","cited_arxiv_id":null,"evidence_quote":"Modified Gauss-Newton with Levenberg-Marquardt damping; used as the inner solver for orthant/box implicit residuals."}],"review_version":1}