{"id":"fe84c688-dec7-4945-8b03-ed62b5c6b1ac","arxiv_id":"2411.18422","paper_version":2,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A Tikhonov-regularized inertial dynamical system for convex multiobjective optimization converges fast to weak Pareto optimal points, with strong convergence to a minimum-norm solution in the main regime.","lead":"This paper introduces a second-order differential equation that combines vanishing damping with Tikhonov regularization for multiobjective optimization, proving fast decay of a merit function and convergence to weak Pareto optimal points, often the minimum-norm one. It extends single-objective acceleration theory to the multiobjective setting and lays groundwork for fast iterative algorithms with strong convergence guarantees.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The strong-convergence/minimum-norm claim is conditional on A3, and Example 2.3 shows A3 is genuinely restrictive and not implied by convexity and Lipschitz gradients.","rationale":"The paper is a mathematically serious extension of single-objective Tikhonov-regularized inertial dynamics to multiobjective problems. The main chain from the Lyapunov estimates in Propositions 4.2-4.3 through the convergence rates in Theorem 4.7 to the strong convergence statement in Theorem 4.8 is logically coherent under the stated standing assumptions. I found no circularity and no gap in the argument that would make Theorem 4.8 false under (A1)-(A3). The genuinely load-bearing weakness is assumption (A3), exactly as the Reader identified: the continuity of the minimum-norm selection map z0 is needed in the proof of Theorem 2.1, and Example 2.3 demonstrates that this property can fail for convex, differentiable objectives with Lipschitz gradients. Because A3 is explicitly disclosed and the theorem is conditional, I do not regard this as a reason to reject or downgrade the paper; it is a scope restriction that should be borne in mind when applying the result. A separate minor issue is that Proposition 2.6(ii) states ||x-z||^2 <= t^p phi_t(x)/beta, whereas strong convexity with modulus beta/t^p gives ||x-z||^2 <= 2 t^p phi_t(x)/beta; the extra constant factor does not change any asymptotic rate in the paper. Overall, the reader's ACCEPT verdict remains appropriate.","tokens_in":43572,"tokens_out":24656,"duration_ms":217283,"concrete_test":"Reproduce Example 2.3 and integrate (MTRIGS) numerically for a parameter regime with p < q+1, e.g. p = 0.5, q = 0.2, alpha = 4, beta = 1/2, using a sufficiently fine finite-difference scheme and recording x(t), z(t), and phi(x(t)) up to large T. If x(t) has multiple accumulation points or stays away from the minimum-norm weak Pareto point while z(t) oscillates, then A3 is essential for Theorem 4.8's conclusion. If x(t) nevertheless converges to a weak Pareto point despite z(t) oscillating, then A3 is only an artifact of the current proof technique and a weaker replacement such as continuity along q(t) = f(x(t)) should be sought.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The proof of Theorem 4.8 reduces strong convergence of x(·) to convergence of the generalized regularization path z(·) to z0(f*) in Theorem 2.1, together with the decay ||x(t)-z(t)|| -> 0 from Theorem 4.7. Theorem 2.1 relies essentially on assumption (A3): continuity of z0(q) = proj_{S(q)}(0). In the proof, A3 is used to bound z_k by z0(q_k) and to force every weak cluster point to be the minimum-norm element z0(q*). Without this continuity, a weak cluster point of z(·) can lie in S(q*) without being the minimum-norm element. Example 2.3 constructs convex, differentiable objectives with Lipschitz gradients for which z0 is discontinuous at q = 0 and an oscillating continuous path q(t) -> 0 makes z(t) oscillate in its second coordinate, never approaching z0(0) = (0,2). Thus A3 is not a consequence of the standing convexity/smoothness hypotheses, and it excludes natural finite-dimensional problems with T-shaped weak Pareto sets. The paper does disclose A3 as a standing assumption, so this is not an internal contradiction; nevertheless, it is the load-bearing condition determining whether the advertised strong convergence and minimum-norm selection actually hold, and the proof mechanism collapses without it.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a second-order inertial multiobjective gradient-like dynamical system with asymptotically vanishing damping and vanishing Tikhonov regularization, (MTRIGS), defined in a Hilbert space for convex differentiable objectives. It proves existence of trajectory solutions in finite dimensions, derives Lyapunov-type energy estimates, and obtains merit-function decay rates O(t^{-2q}) or O(t^{-p}) depending on the parameter regime. In the regime q in (0,1), p < q+1, it establishes strong convergence of bounded trajectories to a weak Pareto optimal point that is the minimum-norm element of the intersection of the lower level sets of the objectives (Theorem 4.8). In other regimes it proves weak convergence via Opial's argument, and it supports the theory with numerical experiments. The standing assumptions (A1)-(A3) include the new regularity condition (A3), which requires continuity of the projection onto the argmin sets S(q).","tokens_in":43841,"tokens_out":11236,"duration_ms":95507,"significance":"The results, if correct, provide a multiobjective analogue of the single-objective Tikhonov-regularized inertial gradient theory, combining fast merit-function decay with a minimum-norm selection principle. The proofs are detailed and carefully parameterized, with a clear summary table of rates and a useful new tool in the generalized regularization path z(t). The paper is transparent about the restrictive standing assumption (A3), and Example 2.3 convincingly shows that (A3) is not a consequence of convexity and Lipschitz smoothness. This limits the scope of the advertised strong-convergence result but does not invalidate the conditional theorems. The numerical experiments illustrate the predicted behavior, although they are not a substitute for the analytical proofs.","major_comments":[{"comment":"Assumption (A3) is load-bearing for the paper's central strong-convergence claim. In the proof of Theorem 2.1, continuity of z0 is used both to bound the sequence {z_k} and to force every weak cluster point to equal z0(q*); without it, a cluster point can lie in S(q*) without being the minimum-norm element. Example 2.3 demonstrates that for convex, differentiable functions with Lipschitz gradients in R^2, z0 can be discontinuous and the regularization path z(t) can oscillate without converging to z0(0). Consequently, Theorem 4.8's strong convergence and minimum-norm selection are genuinely conditional on (A3), which is a restrictive condition beyond the standard convexity and smoothness hypotheses. The manuscript should state this limitation prominently in the abstract and in the statement of Theorem 4.8, and it should point the reader to the sufficient condition in Definition 1.8/Theorem 1.10 for verifying (A3).","section":"§1.2.2 and Theorem 4.8"},{"comment":"In inequality (4.23), the notation μ_q(t) appears, but the subsequent integrating factor M_r(t) is defined with μ_r(t) = λ/t^r - 2r/t. This is presumably a typographical error: for a fixed r in the theorem, the differential inequality should use μ_r(t), not μ_q(t). The mistake does not affect the argument, but it should be corrected for readability.","section":"§4.2, proof of Theorem 4.7"}],"minor_comments":[{"comment":"The row for the regime p < q+1 reports exponents involving max(q, p-q), but Theorem 4.7 is stated for arbitrary r in [q,1) ∩ [p-q,1). The table should explicitly state that the displayed exponents correspond to the choice r = max(q, p-q), since this is the sharpest choice among the admissible values.","section":"Table 1"},{"comment":"The notation \"projC(x(t))+ β/tp x(t)+¨x(t)(0)\" should read \"proj_{C(x(t)) + β/t^p x(t) + ddot{x}(t)}(0)\" to make clear that the projection is onto the set C(x(t)) + β/t^p x(t) + ddot{x}(t). The current typesetting obscures the intended expression.","section":"Definition 3.1 and Proposition 3.4"},{"comment":"In the continuity proof, several displayed formulas have missing or misplaced parentheses, e.g., \"max_i (fi(z(t) - qi(t))\" and the subsequent term \"max_i (fi(z(t) - qi(t)) - β/(2t^p) ∥z(t)∥²\". The intended expressions are clear from context, but the formulas should be typeset correctly.","section":"Proposition 2.5"},{"comment":"The notation \"r := q+1/2\" in the proof is ambiguous; it should be \"r := (q+1)/2\" to match the endpoint used in the argument leading to ∫ s ∥dot{x}(s)∥² ds < ∞.","section":"Theorem 4.12 and Remark 4.13"}],"recommendation":"minor_revision","confidential_remarks":"The paper is technically sound and the conditional results are proved in detail. The main reason I am not recommending accept outright is that the central strong-convergence claim depends essentially on (A3), which is explicitly disclosed but is a strong regularity condition not implied by the other assumptions. This is a scope limitation rather than a mathematical error. The authors can address it by qualifying the abstract and Theorem 4.8 and by adding a short discussion of when (A3) is expected to hold, referencing Example 2.3 and the sufficient growth condition in Section 1.2.2. The remaining issues are typographical and presentational."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read it carefully because the stress-test flagged A3; I think the flag is right and it is the main thing to know. The paper is a genuine extension: it builds the multiobjective Tikhonov-regularized inertial system (MTRIGS), proves fast merit-function decay O(t^{-p}), and in the p<q+1 regime proves strong convergence to a weak Pareto point characterized as minimum-norm in the intersection of lower level sets. The generalized regularization path z(t) defined via adaptive Pascoletti-Serafini scalarization is new, and the Lyapunov machinery is adapted competently. Existence is proved in finite dimension, the parameter table is useful, and the numerics support the theory. The authors also deserve credit for including Example 2.3, which honestly shows the limits of their key assumption.\n\nThe soft spot is real: Assumption (A3), continuity of z0(q)=proj_{S(q)}(0), is load-bearing. Theorem 4.8 reduces strong convergence of x(t) to convergence of z(t) to z0(f*), and Theorem 2.1 needs A3 to ensure that weak cluster points of z(t) are not just in S(q*) but are the minimum-norm point. Without A3, z(t) can oscillate among elements of S(q*) with larger norm, as Example 2.3 shows for convex differentiable functions with Lipschitz gradients on R^2. So the advertised minimum-norm selection is not a consequence of convexity and smoothness; it holds only under a genuine additional regularity condition. The paper is transparent about this, so it is not a hidden flaw, but it does narrow the scope of the main theorem. The same example also shows that Theorem 2.1 fails in natural finite-dimensional problems, so this is not a technicality.\n\nThe other limitations are disclosed too: existence is finite-dimensional, uniqueness is open, and (A2) is strong. The proofs are long and some rely on prior results [33,34], but the reuse is legitimate and the derivation is coherent. The numerical section is illustrative rather than a benchmark, which is fine for a continuous-dynamics paper.\n\nFor a reader working on multiobjective inertial dynamics this deserves a serious read and a serious referee. I would send it to review. I would ask the authors to add a short discussion or sufficient conditions under which (A3) holds, since as written the main strong-convergence theorem is conditional on a condition that is easy to violate.","headline":"A technically careful multiobjective analogue of TRIGS with a genuinely new regularization path, but the headline strong-convergence result is conditional on a restrictive continuity assumption (A3) that can fail for simple convex smooth problems.","tokens_in":44377,"tokens_out":2366,"would_cite":true,"duration_ms":23169,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C29","90C30","90C25","91A12","91B55","34G20","34E10","37L05"],"pacs":[],"model":"deepseek-v4-flash","headline":"For convex multiobjective problems, a damped inertial flow with vanishing Tikhonov regularization achieves $O(t^{-p})$ merit decay and, in the main regime, strong convergence to the minimum-norm weak Pareto point.","keywords":["multiobjective optimization","Pareto optimality","Tikhonov regularization","inertial dynamics","asymptotic vanishing damping","strong convergence","merit function","Lyapunov analysis"],"falsifier":"Take the two convex functions defined in Example 2.3 with the explicit regularization path (2.12); as $t\\to+\\infty$ the second coordinate of $z(t)$ oscillates between $2.25$ and $2.75$, so $z(t)$ does not converge to the minimum-norm weak Pareto point $(0,2)$, demonstrating that without the continuity assumption (A3) the path-convergence mechanism of Theorem 2.1 — and therefore the strong convergence in Theorem 4.8 — fails.","tokens_in":43362,"feed_emoji":"🎯","tokens_out":9477,"duration_ms":75942,"temperature":0.7,"pith_summary":"This paper studies a second-order inertial dynamics for convex differentiable multiobjective problems, in which both the damping $\\alpha/t^q$ and the Tikhonov regularization $\\beta/t^p$ vanish as $t\\to+\\infty$. Its main assertion is that, for $q\\in(0,1)$ and $p<q+1$, every bounded trajectory converges strongly to a weak Pareto optimal point, and this limit is exactly the minimum-norm element of the intersection of the lower level sets at the limit values of the objective functions. At the same time the multiobjective merit function decays like $O(t^{-p})$, which can be chosen arbitrarily close to the Nesterov-type rate $O(t^{-2})$. Extending the single-objective theory to the multiobjective setting matters because it provides a selection principle: the dynamics not only approaches the Pareto set, it identifies a distinguished point on it, and it offers a template for accelerated gradient and proximal methods with strong iterate convergence.","feed_headline":"Inertial multiobjective flow picks the least-norm Pareto point","feed_subtitle":"A generalized regularization path drives trajectories to a weak Pareto point with least norm.","key_machinery":"The machinery is a Lyapunov analysis built around the generalized Tikhonov regularization path. For $\\lambda>0$ and $r\\in[q,1]$, the paper defines energies $$E^r_{\\$\\lambda$}(t)=$t^{{2r}}$\\varphi_t(x(t))+\\tfrac12\\|\\$\\lambda$(x(t)-z(t))+t^r\\dot x(t)\\|^2+\\tfrac{\\$\\lambda$}{2}\\bigl(r $t^{{r-1}}$+\\$\\alpha$ $t^{{r-q}}$-2\\$\\lambda$\\bigr)\\|x(t)-z(t)\\|^2,$$ where $\\varphi_t$ is the merit function of the regularized problem and $z(t)$ is the generalized path. Estimates on the derivative of this energy, combined with an integrating factor and an auxiliary integrability lemma, yield the decay rates and the key estimate $\\|x(t)-z(t)\\|\\to0$. Strong convergence follows once the continuity assumption (A3) guarantees that $z(t)$ itself converges to the minimum-norm element of the limiting lower-level set.","core_discovery":"The central claim is Theorem 4.8: under Assumptions (A1)-(A3), if $q\\in(0,1)$ and $p<q+1$, every bounded trajectory solution $x(\\cdot)$ of (MTRIGS), $$\\ddot x(t)+\\frac{\\$\\alpha$}{t^q}\\dot x(t)+\\operatorname{proj}_{C(x(t))+\\frac{\\$\\beta$}{t^p}x(t)+\\ddot x(t)}(0)=0,\\qquad C(x)=\\operatorname{conv}\\{\\nabla f_i(x):i=1,\\dots,m\\},$$ converges strongly to a weak Pareto optimal point $x^*$ that minimizes the norm on $\\bigcap_{i=1}^m L(f_i,f_i(x^*))$, and the merit function satisfies $\\varphi(x(t))=O(t^{-p})$ (Theorem 4.7). The proof proceeds by attaching to each trajectory a generalized regularization path $z(t)=\\arg\\min_z \\max_i(f_i(z)-f_i(x(t)))+\\frac{\\beta}{2t^p}\\|z\\|^2$, showing $z(t)\\to x^*$ and then $\\|x(t)-z(t)\\|\\(\\to0$). The paper also establishes weak convergence in the regimes $q+1<p$ and the boundary case $p=2$, $\\beta\\geq q(1-q)$, and the rate $\\varphi(x(t))=O(t^{-2q})$ when $2q<p$.","pith_inferences":["A practical consequence is that the strong-convergence guarantee should only be claimed for problems where the auxiliary map $z_0(\\cdot)$ is continuous; verifying this condition, or using regularization schemes that recover it, becomes part of algorithm design, and the paper's Example 2.3 is a concrete test case.","The excluded boundary $p=q+1$ is the natural place to look next: the single-objective theory exhibits a threshold phenomenon in this regime, and one would expect a similar phase transition for the multiobjective system.","The same Lyapunov construction should translate to discrete-time multiobjective proximal or fast gradient algorithms, predicting $O(1/k^2)$-type decreases of the merit function and strong convergence of the iterates to the minimum-norm weak Pareto point.","Because the continuity assumption (A3) is about the geometry of the weak Pareto set in value space, an empirical check is possible: compute $z_0(q)$ on a grid of $q$ vectors and look for jumps; problems with kinked Pareto fronts are the likely failure cases."],"forward_implications":["In the regime $q\\in(0,1)$ with $p<q+1$, any bounded trajectory of (MTRIGS) converges strongly to a weak Pareto optimal point that is the minimum-norm element of the intersection of lower level sets at the limit function values.","The merit function values decay as $O(t^{-p})$; since $p$ can be chosen below but arbitrarily close to $q+1<2$, the value convergence can be made arbitrarily close to the $O(t^{-2})$ rate typical of Nesterov-type accelerated dynamics.","In the regime $q+1<p$ and in the boundary case $p=2$, $\\beta\\geq q(1-q)$, bounded trajectories converge weakly to a weak Pareto optimal point, with the velocity satisfying $\\int^{\\infty} t\\|\\dot x(t)\\|^2\\,dt<+\\infty$.","These results lay the groundwork for fast gradient and proximal point methods in multiobjective optimization that carry strong convergence guarantees for the iterates.","Numerical experiments indicate that the Tikhonov term actively steers the trajectory to the minimal-norm Pareto point, in contrast to the unregularized inertial multiobjective system."],"supporting_citations":[{"why":"Supplies the single-objective template: inertial dynamics with vanishing Tikhonov regularization converge strongly to the minimum-norm solution, which this paper extends to multiobjective problems.","marker":"[2]"},{"why":"Establishes rates $\\varphi(x(t))=O(t^{-p})$ and strong convergence for the single-objective TRIGS system, providing the framework adapted here.","marker":"[27]"},{"why":"Introduces the multiobjective inertial gradient-like system with asymptotic vanishing damping (MAVD) whose merit-function analysis and energy estimates are the starting point.","marker":"[33]"},{"why":"Introduces the base inertial multiobjective system (IMOG') and its properties used to motivate (MTRIGS).","marker":"[34]"},{"why":"Provides the ODE model for Nesterov acceleration that determines the choice of the time-varying damping term.","marker":"[35]"},{"why":"Gives the classical properties of the Tikhonov regularization path, in particular the strong convergence of $x_\\varepsilon$ to $\\mathrm{proj}_{\\arg\\min f}(0)$, used in defining and bounding the generalized path.","marker":"[9]"},{"why":"Supplies the merit function $\\varphi$ and the characterization of weak Pareto points ($\\varphi(x)=0$) used throughout the analysis.","marker":"[39]"},{"why":"Underpins the adaptive scalarization view of the generalized regularization path as a Pascoletti-Serafini scalarization.","marker":"[21]"},{"why":"Provides the Mosco convergence theorem connecting continuity of the projection $z_0(\\cdot)$ to set-valued convergence, the content of assumption (A3).","marker":"[10]"}],"fun_headline_variants":["Inertial multiobjective flow converges to least-norm Pareto","Pareto-optimal least-norm selected via inertial Tikhonov","Multiobjective inertial convergence to least-norm Pareto","Least-norm Pareto reached by inertial multiobjective flow"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is assumption (A3): the map that sends a reference vector $q$ to the minimum-norm solution $z_0(q)$ of $\\arg\\min_z\\max_i(f_i(z)-q_i)$ must be continuous, and the paper's Example 2.3 shows this continuity can fail for perfectly convex functions with Lipschitz gradients; if it fails, the regularization path can oscillate and the strong-convergence conclusion collapses.","fun_headline_variants_meta":{"raw":{"variants":["Inertial multiobjective flow converges to least-norm Pareto","Pareto-optimal least-norm selected via inertial Tikhonov","Multiobjective inertial convergence to least-norm Pareto","Least-norm Pareto reached by inertial multiobjective flow"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000854,"raw_usage":{"total_tokens":3736,"prompt_tokens":996,"completion_tokens":2740,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":612,"completion_tokens_details":{"reasoning_tokens":2684}},"tokens_in":612,"tokens_out":2740,"duration_ms":19734,"temperature":1.0,"reasoning_tokens":2684,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:13:20.268041+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the two convex functions defined in Example 2.3 with the explicit regularization path (2.12); as $t\\to+\\infty$ the second coordinate of $z(t)$ oscillates between $2.25$ and $2.75$, so $z(t)$ does not converge to the minimum-norm weak Pareto point $(0,2)$, demonstrating that without the continuity assumption (A3) the path-convergence mechanism of Theorem 2.1 — and therefore the strong convergence in Theorem 4.8 — fails.","supporting_citations":[{"cited_title":"Attouch, A","cited_arxiv_id":null,"evidence_quote":"Supplies the single-objective template: inertial dynamics with vanishing Tikhonov regularization converge strongly to the minimum-norm solution, which this paper extends to multiobjective problems."},{"cited_title":"Sonntag and S","cited_arxiv_id":null,"evidence_quote":"Introduces the multiobjective inertial gradient-like system with asymptotic vanishing damping (MAVD) whose merit-function analysis and energy estimates are the starting point."},{"cited_title":"Sonntag and S","cited_arxiv_id":null,"evidence_quote":"Introduces the base inertial multiobjective system (IMOG') and its properties used to motivate (MTRIGS)."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the ODE model for Nesterov acceleration that determines the choice of the time-varying damping term."},{"cited_title":"Beer , On Mosco convergence of convex sets , Bulletin of the Australian Mathematical Society, 38 (1988), pp","cited_arxiv_id":null,"evidence_quote":"Provides the Mosco convergence theorem connecting continuity of the projection $z_0(\\cdot)$ to set-valued convergence, the content of assumption (A3)."}],"review_version":1}