{"id":"7ddf0cfe-a9e8-4c71-91b5-cd09ffe33ffa","arxiv_id":"2508.20039","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"The robust path of a linear robust optimization problem is a Bregman projection of a dual-space curve, and proximal point trajectories of the nominal problem approximate it with a geometry-dependent error bound.","lead":"This paper shows that the family of robust solutions obtained by varying the uncertainty set in a linear robust optimization problem can be described as Bregman projections onto the feasible set, and approximated by the trajectory of a single proximal point run on the deterministic problem. The result offers a way to trace the entire robustness-efficiency tradeoff in one pass instead of solving many separate robust problems.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Main results govern the regularized subpath P′(V), not the robust path P(V); Lemma 7 only proves P′(V)⊆P(V), and the paper's own Figure 1 caption concedes this, so the abstract's characterization of 'the robust path' is unsupported.","rationale":"This is the single load-bearing point because every downstream claim about approximation and recovery is about P′. If P(V) has points outside P′, the algorithm can only approximate a subset; the advertised geometric characterization of 'the robust path' is then a characterization of a different object. The reader's weakest_assumption identifies exactly this. I add a concrete instance showing the subset can be strict under the paper's own Assumptions 1 and 2, so the concern is not merely a missing proof. The paper deserves credit for the clean Bregman-projection machinery and the exactness results on P′; the issue is a scope mismatch, not an internal inconsistency. The condition Π_X(0)=Π_Aff(X)(0) in Corollary 1 and the monotonicity in Proposition 1 are meaningful sufficient conditions, but they concern P′ and the proximal path, not the full P(V). A clarification and a corrected abstract are enough, so the CONDITIONAL verdict stands; no rejection is warranted. No formal verification exists, so the analytical gap is the main risk.","tokens_in":30742,"tokens_out":18085,"duration_ms":173267,"concrete_test":"Evaluate Definitions 1 and 9 on X={t e1:0≤t≤1}⊂R^2, V=unit Euclidean disk, a0=(-1,0), g(t)=t^2. If P(V)=[0,1]e1 and P′(V)=(0,1]e1 for finite ω, then Lemma 7's inclusion is strict; amend the abstract and theorem statements to 'the characterizable subpath', or prove surjectivity of ω↦r(ω)=ω∇g(||x′_R(ω)||_{V°}) onto [0,∞).","verdict_should_be":"UNCHANGED","load_bearing_attack":"Every central theorem (Theorem 1's Bregman projection formula, Theorem 2's sharp bound, Corollary 1's zero-gap case, Theorem 4's proximal-path equivalence) is stated for P′(V) from Definition 9, i.e., minimizers of <a0,x>+ω g(||x||_{V°}). Lemma 7(i) proves only that P′(V)⊆P(V); Lemma 7(ii) attaches a radius to each regularized solution, but it never shows that for every finite r every (or even any) robust solution in Definition 1 is representable as x′_R(ω) for the fixed g. The gap is acknowledged in the Figure 1 caption, which says P′ is 'a subset of P(V)' — yet the abstract and Algorithm 1 are phrased as if the whole robust path were recovered. This is not idle: for X={t e1:0≤t≤1}, V the unit disk, a0=(-1,0), g(t)=t^2, the robust path P(V) contains (0,0) for every r>1 and all of [0,1]e1 at r=1, while P′(V) consists of (0,1]e1 (with 0 only as ω→∞). Thus Theorems 2 and 4 do not cover the robust solution at large radii; only a closure convention could repair this, and no such convention is stated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies linear robust optimization problems with gauge uncertainty sets and introduces the \"robust path\" P(V), the collection of robust optimal solutions as the uncertainty-set radius varies. Its central object is a redefined \"characterizable robust path\" P'(V), whose points are minimizers of the regularized problem argmin_{x in X} <a0,x> + omega g(||x||_{V°}). Section 3 proves a Bregman projection representation for P'(V), the central path, and the proximal path (Theorem 1). Section 4 bounds the divergence between the central path and P'(V) (Theorem 2), identifies zero-gap cases (Corollary 1 and Propositions 1-2), bounds central-versus-proximal path distance (Theorem 3), and gives sufficient conditions under which proximal iterates are exact P'(V) points with a closed-form radius formula (Theorem 4). Algorithm 1 operationalizes the proximal-path recovery, and Section 5 reports portfolio optimization and adversarial deep learning experiments.","tokens_in":1477,"tokens_out":1403,"duration_ms":80517,"significance":"The geometric view is elegant and largely correct: the Bregman projection derivation in Section 3 follows from standard variational properties, uses no fitted parameters, and provides explicit, falsifiable statements. Theorem 2's uniform bound and its sharpness example are valuable, and the two zero-error cases (polyhedral monotonicity and polar pairs) are genuine structural insights. The portfolio experiments support the algorithmic claims, and the deep learning experiment, while outside the assumptions, is honestly labeled as such. The main obstacle is that the theorems characterize P'(V), a subset of the robust path P(V), rather than P(V) itself; the paper's abstract and Algorithm 1 overstate what is proved. If that gap is resolved, or the claims are restricted to the regularized path, the contribution is solid and publishable.","major_comments":[{"comment":"The central claims of the paper are stated for the characterizable path P'(V), not for the robust path P(V) of Definition 1. Lemma 7(i) proves only P'(V) is contained in P(V), and Lemma 7(ii) only attaches a radius r(omega)=omega grad_g(||x'_R(omega)||_{V°}) to each regularized solution. No result shows that every robust solution in P(V), or every radius r, is represented by some x'_R(omega) for the fixed g. The Figure 1 caption concedes this by saying P'(V) is 'a subset of P(V)'. The gap is not harmless: for X={t e1 : 0<=t<=1}, V the unit disk, a0=(-1,0), and g(t)=t^2, P(V) contains the origin for every r>1 and the whole segment [0,1]e1 at r=1, while P'(V) covers only (0,1]e1, with the origin obtained only as omega goes to infinity. Consequently Theorem 2 and Theorem 4 establish approximation or recovery only of the regularized subpath P'(V), not of P(V). The abstract's statement that 'a robust path can be characterized' is therefore not supported. The fix is to prove a surjectivity or representation theorem under additional conditions, to adopt and state an explicit closure convention for P(V), or to restrict the abstract, theorems, and Algorithm 1 to P'(V) and describe it as a subset of robust solutions.","section":"Section 3.2, Definition 9, Lemma 7, Figure 1 caption"},{"comment":"Theorem 4's sufficient condition (Pi^phi_X(0)=Pi^phi_{Aff(X)}(0) plus monotonicity) proves that each proximal iterate equals x'_R(omega_k), and Lemma 7 then shows that x_k solves the robust problem at radius r_k=omega_k grad_g(||x_k||_{V°}). This is a valid statement about a subset of robust solutions, and the closed-form radius formula is useful. However, the theorem does not imply that the proximal path visits all of P(V), and Algorithm 1's description of its output as 'an (approximate) robust path of (RC)' overstates the coverage. The algorithm's output should be described as a regularized path contained in P(V), with approximation bounds measured against P'(V), unless the surjectivity issue identified above is resolved.","section":"Section 4.3, Theorem 4, Algorithm 1"}],"minor_comments":[{"comment":"The phrase 'robust path can be characterized' should be qualified to refer to the characterizable robust path P'(V), since Theorem 1 applies to P'(V) rather than to P(V) as defined in Definition 1.","section":"Abstract and Section 1"},{"comment":"The word 'robsut' in the input line should be 'robust'.","section":"Algorithm 1"},{"comment":"The text says the max-return portfolio is xR twice; the second occurrence should be xE.","section":"Section 5.1.1, item (iii)"},{"comment":"The use of the double polar U°° should be justified explicitly by the closed, convex, origin-containing assumption on U, or the notation should be simplified to avoid an unstated identification.","section":"Appendix E, Proposition 2"},{"comment":"The second inequality in the definition of kappa-expansiveness uses the translated set S+d; it would help to state explicitly that d is finite and that S+d is closed and convex, and that kappa is uniform over the admissible sets S.","section":"Definition 10"}],"recommendation":"major_revision","confidential_remarks":"The P'(V)-versus-P(V) gap is the decisive issue. I do not see it as a reason for rejection, because the regularized-path results are themselves substantive and the gap can likely be closed by a reframing or by added surjectivity conditions. The paper fits the journal's scope well."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is worth reading. The central idea is that for linear robust problems with gauge uncertainty sets, the regularized robust path P′(V) is a Bregman projection of a dual-space ray onto the feasible set, and the same lens describes the central and proximal paths of the deterministic counterpart. That unification is new and genuinely useful. Theorem 1 is the heart; Theorem 2's bound is sharp; the zero-gap corollaries for polyhedrally monotone sets and polar pairs are clean. The portfolio experiments support the theory, and the deep-learning speedup is compelling even if heuristic.\n\nBut the abstract sells a bit more than the paper delivers. The main theorems are all about P′(V), the path of minimizers of ⟨a0,x⟩ + ω g(‖x‖_{V°}), not the robust path P(V) from Definition 1. Lemma 7 shows P′(V) ⊆ P(V) and attaches a radius to each P′ point; it never shows every robust radius is represented. The stress-test example is right: with X a segment along e1, V the disk, a0 = (−1,0), and g(t) = t², the robust path contains the origin for every r > 1, while P′(V) is (0,1]e1 plus 0 only as a limit. So for large radii the approximated path misses the actual robust solution. The Figure 1 caption honestly says P′ is a subset of P, but the abstract, Algorithm 1, and several theorem statements do not carry the caveat. That is the load-bearing soft spot.\n\nIt is fixable. The authors can either state all results for the “characterizable robust path” and justify why that object is the right one, or prove a surjectivity or closure condition that covers the missing radii. The latter would be more useful, because without it the algorithm's output in the example stops being robust after the first proximal step.\n\nMinor issues: Proposition 1's proof has a parameter typo (λ0 where ω_k is intended), and the numerical section has no code, seeds, or error bars — the results are plausible but not reproducible as given. Citation practice seems fair; the self-citation is not load-bearing.\n\nVerdict: send to peer review after a major revision. The geometric core is solid and the examples are instructive; the authors just need to align the claims with the theorems and provide reproducibility for the numerics.","headline":"Genuinely nice Bregman-projection unification of robust and deterministic paths, but the headline result only covers the regularized subpath P′(V) — the abstract overstates it, and the paper should be revised, not rejected.","tokens_in":31567,"tokens_out":4093,"would_cite":true,"duration_ms":40236,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C47","90C25","49N15"],"pacs":[],"model":"deepseek-v4-flash","headline":"One proximal-point run on the deterministic counterpart approximates, and sometimes exactly reproduces, the full robust path of a linear robust optimization problem.","keywords":["robust optimization","uncertainty set calibration","robust path","Bregman projection","proximal point method","regularization paths","Pareto efficient robust solutions","gauge functions"],"falsifier":"Enumerate the robust path by solving the min-max problem over a fine grid of radii $r$ for a concrete instance with a compact strictly convex smooth uncertainty set $V$ and a convex feasible set $X$, and compare it with the regularized path $\\arg\\min_{x\\in X}\\langle a_0,x\\rangle+\\omega g(\\|x\\|_{V^\\circ})$ over a fine grid of $\\omega$. A single robust solution not reproduced by any regularized solution would falsify the claim that the proximal path covers the full robust path.","tokens_in":30487,"feed_emoji":"📈","tokens_out":9880,"duration_ms":79548,"temperature":0.7,"pith_summary":"This paper tries to make uncertainty-set calibration in robust optimization cheap. Instead of solving the robust counterpart many times for different uncertainty-set shapes and radii, the authors define the robust path—all robust solutions as the radius grows—and show that this path is a Bregman projection of a curve in the dual space onto the feasible region. The same projection identity describes the central path and the proximal point path of the deterministic counterpart, so a single proximal-point run initialized at the most robust solution approximates the entire robust path. The approximation error is controlled by the geometry of the feasible region and the uncertainty set, and in two cases—polyhedrally monotone feasible regions, and polar-pair feasible and uncertainty sets—the approximation is exact. The payoff is that entire trade-off curves for portfolio optimization and adversarial training can be produced in one pass instead of many robust solves.","feed_headline":"One proximal run traces the whole robust path","feed_subtitle":"A single optimization trajectory replaces many robust solves by tracing the entire uncertainty trade-off curve.","key_machinery":"The load-bearing object is the Bregman projection $\\Pi_X^\\varphi$ of a dual-space curve onto the feasible set, where $\\varphi=g\\circ\\|\\cdot\\|_{V^\\circ}$ is a Legendre function built from the polar of the uncertainty set. Passing through the bijection $\\nabla\\varphi$, the robust path, the central path, and each proximal iterate are all projections of rays of the form $\\nabla\\varphi(0)-\\omega^{-1}a_0$ or $\\nabla\\varphi(x_R)-\\omega^{-1}a_0$; the paths differ only in where the ray starts. The supporting machinery is the dual characterization of Bregman projections onto affine subspaces and translated cones, the $\\kappa$-expansiveness of the projection, and a face-monotonicity condition that makes proximal iterates coincide with central-path points on polyhedral feasible regions.","core_discovery":"On the paper's own terms, the central discovery is that robust solutions of the min-max problem are, on a characterizable subset, the regularized solutions $\\arg\\min_{x\\in X}\\langle a_0,x\\rangle+\\omega(g\\circ\\|\\cdot\\|_{V^\\circ})(x)$, and this regularized path equals the Bregman projection $P'(V)=\\{\\Pi_X^\\varphi(\\nabla\\varphi^*(\\nabla\\varphi(0)-\\omega^{-1}a_0)):\\omega\\in[0,\\infty)\\}$ with $\\varphi=g\\circ\\|\\cdot\\|_{V^\\circ}$. The central path of the deterministic problem, using the same $\\varphi$ and initialized at the most robust solution $x_R=\\arg\\min_{x\\in X}\\varphi(x)$, satisfies the sharp bound $D_\\varphi(x_{CP}(\\omega),x'_R(\\omega))\\le\\kappa^2 D_\\varphi(\\Pi_X^\\varphi(0),\\Pi_{\\mathrm{Aff}(X)}^\\varphi(0))$. The proximal point path approximates that central path, and under $\\kappa$-expansiveness, $\\Pi_X^\\varphi(0)=\\Pi_{\\mathrm{Aff}(X)}^\\varphi(0)$ and monotonicity on a polyhedral $X$, every proximal iterate $x_k$ is exactly a robust solution with radius $r_k=\\omega_k\\nabla g(\\|x_k\\|_{V^\\circ})$, where $\\omega_k=(\\sum_{j=0}^{k-1}\\lambda_j^{-1})^{-1}$. Thus choosing the uncertainty-set shape is equivalent to choosing the distance-generating function, and choosing the step-size cadence chooses the radii of the robust solutions.","pith_inferences":["The same one-pass recipe would plausibly extend to convex nonlinear objectives through an epigraph reformulation, turning any regularization-path method on the enlarged problem into a robust-path method; the paper gestures at this but does not prove it.","Because the robust path is one-dimensional and traceable without repeated robust solves, the remaining deployment choice—which uncertainty radius to use—could be automated by scanning the recovered path with out-of-sample performance, reducing a high-dimensional calibration problem to a line search.","The bound involving $\\Pi_{\\mathrm{Aff}(X)}^\\varphi(0)$ suggests a cheap pre-computation: if the divergence between the projection of the origin onto $X$ and onto its affine hull is tiny, the proximal path is guaranteed close to the robust path before running any algorithm.","The results imply an exact equivalence between two previously separate notions only when the feasible region is straight enough relative to the origin; curved feasible sets would require an extra quantitative control on that projection gap, which the paper does not provide."],"forward_implications":["A single proximal-point pass on the deterministic counterpart, initialized at the most robust solution, yields an approximate robust path; the practitioner no longer needs to solve the robust min-max problem at many radii.","The uncertainty-set shape $V$ maps to the distance-generating function $\\varphi=g\\circ\\|\\cdot\\|_{V^\\circ}$, and the proximal step sizes $\\{\\lambda_k\\}$ map to the radii $\\{r_k\\}$, so both design levers have exact algorithmic counterparts.","On polyhedrally monotone feasible regions (for example a simplex under ellipsoidal uncertainty) and when $X$ and $V$ are polar pairs up to rescaling, the proximal iterates are exact robust solutions, not merely approximations.","The approximation error is governed by two geometric quantities: the $\\kappa$-expansiveness of the Bregman projection and the divergence between the projections of the origin onto $X$ and onto $\\mathrm{Aff}(X)$, so the bound can be checked before any robust problem is solved.","In adversarial deep learning—where the linear-objective assumption fails—the same one-pass recipe produces clean-accuracy and robust-accuracy trade-offs comparable to repeated adversarial training, with a reported 85% reduction in computational time."],"supporting_citations":[{"why":"Supplies the Legendre-function theory, the gauge/polar duality, and the variational-inequality characterizations used throughout the proofs.","marker":"Rockafellar (1970)"},{"why":"Gives existence, uniqueness, and the dual characterization of Bregman projections used to map paths between primal and dual space.","marker":"Bauschke et al. (1997)"},{"why":"Provides the duality formula for Bregman projections onto translated cones and affine subspaces, the key step in Theorem 2's bound.","marker":"Bauschke (2003)"},{"why":"Frames the uncertainty set via gauge constraints, grounding the dual reformulation of the robust counterpart.","marker":"Freund (1987)"},{"why":"Contributes the face-monotonicity condition used in Proposition 1 and Theorem 4.","marker":"González-Sanz et al. (2025)"},{"why":"Supplies the Two-Fund Theorem benchmark against which the proximal path is compared in portfolio experiments.","marker":"Markowitz (2008)"},{"why":"Formulates adversarial training as a robust optimization problem, the deep-learning setting tested in Section 5.2.","marker":"Madry et al. (2018)"},{"why":"Provides the FGSM adversarial-training baseline whose computation time is compared with the paper's Algorithm 2.","marker":"Wong et al. (2020)"}],"fun_headline_variants":["One proximal trajectory traces the entire robust path","Single optimization run reveals robust trade-off curve","Replacing many solves: one run traces robust path","All robust solutions from a single run","One trajectory, entire robust path"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central guarantee covers the characterizable regularized path $P'(V)$, and the paper proves only $P'(V)\\subseteq P(V)$; if the true robust path contains a solution that no regularized problem with the chosen $g$ reproduces, the approximation bounds do not cover it.","fun_headline_variants_meta":{"raw":{"variants":["One proximal trajectory traces the entire robust path","Single optimization run reveals robust trade-off curve","Replacing many solves: one run traces robust path","All robust solutions from a single run","One trajectory, entire robust path"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000769,"raw_usage":{"total_tokens":3507,"prompt_tokens":1143,"completion_tokens":2364,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":759,"completion_tokens_details":{"reasoning_tokens":2299}},"tokens_in":759,"tokens_out":2364,"duration_ms":17675,"temperature":1.0,"reasoning_tokens":2299,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:49:57.820358+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Enumerate the robust path by solving the min-max problem over a fine grid of radii $r$ for a concrete instance with a compact strictly convex smooth uncertainty set $V$ and a convex feasible set $X$, and compare it with the regularized path $\\arg\\min_{x\\in X}\\langle a_0,x\\rangle+\\omega g(\\|x\\|_{V^\\circ})$ over a fine grid of $\\omega$. A single robust solution not reproduced by any regularized solution would falsify the claim that the proximal path covers the full robust path.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Legendre-function theory, the gauge/polar duality, and the variational-inequality characterizations used throughout the proofs."},{"cited_title":"(1997) Legendre functions and the method of random B regman projections","cited_arxiv_id":null,"evidence_quote":"Gives existence, uniqueness, and the dual characterization of Bregman projections used to map paths between primal and dual space."},{"cited_title":"Journal of Approximation Theory 121(1):1--12","cited_arxiv_id":null,"evidence_quote":"Provides the duality formula for Bregman projections onto translated cones and affine subspaces, the key step in Theorem 2's bound."},{"cited_title":"Mathematical Programming 38:47--67","cited_arxiv_id":null,"evidence_quote":"Frames the uncertainty set via gauge constraints, grounding the dual reformulation of the robust counterpart."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Two-Fund Theorem benchmark against which the proximal path is compared in portfolio experiments."},{"cited_title":"International Conference on Learning Representations","cited_arxiv_id":null,"evidence_quote":"Formulates adversarial training as a robust optimization problem, the deep-learning setting tested in Section 5.2."},{"cited_title":"International Conference on Learning Representations","cited_arxiv_id":null,"evidence_quote":"Provides the FGSM adversarial-training baseline whose computation time is compared with the paper's Algorithm 2."}],"review_version":1}