{"id":"041439ae-a779-4ea8-b0bf-68e2daa76f25","arxiv_id":"2608.12520","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A weighted sum equals the lexicographic cascade in convex MPC exactly when the weight supports the upper image at the lexicographic point; such certificates are computable under ℓ1 hinges but expire from tick to tick.","lead":"This paper derives the exact condition under which a single weighted optimization reproduces a lexicographic (priority-ordered) optimum in convex model predictive control, and shows the condition is a geometric membership question. The authors find the resulting weight certificates are valid only for one decision instance, expire within one control tick in their simulations, and require monitoring and selective fallback rather than a fixed tuning.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Closed-loop divergence may be driven by warm-start mismatch between weighted-sum and cascade solves, not by certificate expiry; the pointwise theorem itself appears sound.","rationale":"The support-normal equivalence theorem and the polyhedral scaled-KKT construction survive scrutiny: the proof of Theorem III.3 correctly handles the closure of the upper image, and Proposition III.6's stationarity equivalence is valid for the fixed convex inner problem. The load-bearing weakness is not in the mathematics but in the transfer from the fixed-instance certificate to the closed-loop measurements. The paper itself names warm-start mismatch as a divergence source, but the campaign does not separate that effect from certificate expiry. This does not overturn the theoretical result; it makes the empirical support conditional, exactly as the reader concluded. The reader's weakest assumption identifies the same concern, and the proposed warm-start ablation would settle whether the closed-loop divergence actually tests the certificate theory.","tokens_in":27424,"tokens_out":21396,"duration_ms":215930,"concrete_test":"Re-run the dense held-out episode (and a subset of the C1-vs-C4 cells) under a controlled warm-start protocol: at each tick, solve both the weighted-sum and the cascade from a common warm start (e.g., the previous tick's accepted solution), and also from each mode's native warm start; record the two linearized instances and test w_dep membership in Omega(p*) for each. If the C1/C4 compliance gap and event-level divergence persist when the instances coincide, the certificate-expiry mechanism is supported; if they collapse, warm-start mismatch is the dominant confound and the Section VI-A expiry measurements must be reinterpreted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Theorem III.3 and the scaled-KKT construction are internally consistent for a fixed convex instance, but the closed-loop attribution is insecure. Section V-B applies the certificate to each sequentially linearised inner problem, while Section VIII explicitly lists 'different warm starts produce different locally convexified subproblems' as a source of WS/cascade divergence. The WS run and the cascade reference run maintain their own warm starts, so at the same tick they can be solving different convex problems; Theorem III.3 certifies only the instance at which it is evaluated. The per-tick compliance match and event-level divergence in Section VII therefore may reflect instance mismatch rather than w_dep outside Omega(p*) or certificate expiry. Section VI-A measures expiry on cascade-derived exemplars, but does not verify that the online WS solve's linearized instance is the same one whose certificate was measured. The paper is candid about this limitation, yet no ablation controls warm-start identity while holding the weight fixed, so the central empirical claim for monitored weighted solves remains conditional.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper studies when a single weighted-sum minimization J(z)+Σ_i w_i V_i(z) reproduces the lexicographic optimum of a convex priority-ordered program. The central theoretical result, Theorem III.3, characterizes the admissible weights as the unit-performance slice of the outward normal cone of the upper image at the lexicographic point. For ℓ1 hinge penalties the paper gives a polyhedral description via a scaled-KKT projection (Theorem III.4, Proposition III.6, Algorithm 1); for ℓ2 squared-hinge penalties it proves a local foreclosure result with O(1/w) and O(1/w²) violation rates (Theorem III.9, Proposition III.10). The paper then presents a closed-loop nuPlan study on a 25-rule rulebook: calibrated weights often have near-equal tier components, the deployed heuristic weight typically lies outside the certified region, certificate lifetime is short and tracks active-set churn, and a monitored weighted solve with cascade fallback gives exploratory event-level gains. The authors are candid that certificates are pointwise and that the compliance-pattern monitor has sensitivity 0.55.","tokens_in":27664,"tokens_out":7756,"duration_ms":73149,"significance":"If the claims hold, the paper is a valuable and practical contribution: it converts the folklore 'separation heuristic' question into a well-posed membership problem, supplies a constructive LP-based calibration procedure with a certified margin for ℓ1 penalties, and gives a clean explanation of why squared-hinge penalties cannot be exact at finite weight. The theoretical core appears correct, and the paper supports it with machine-checkable randomized verification on 1000 instances and two worked examples. The authors are also unusually explicit about limitations: empty robust intersections, short certificate lifetimes, and the imperfect monitor. These strengths make the paper worth serious consideration; the main open question is whether the closed-loop attribution, rather than the pointwise theory, is supported by the experiments.","major_comments":[{"comment":"The theorem's statement is an 'if and only if' for z*_lex being a member of the weighted-sum argmin set, but the paper's practical claim throughout the abstract and Section II.B is that a weight 'reproduces' the lexicographic optimum, i.e. that the WS solve returns z*_lex. When the WS minimizer is not unique, w ∈ Ω(p*) does not force every WS minimizer to equal z*_lex; the additional uniqueness hypothesis in the sentence after Theorem III.3 is therefore load-bearing. Because ℓ1 hinges are piecewise linear, J + Σ_i w_i V_i is generically not strictly convex, so non-uniqueness is not a measure-zero corner case in the deployed setting. Please either state all practical claims with the uniqueness qualifier, add a tie-breaking rule that selects the lexicographic point among WS minimizers, or prove a condition under which uniqueness holds; Algorithm 1's verification step currently only checks trajectory equality 'when the WS minimizer is unique'.","section":"Section III-B, Theorem III.3; Eq. (8)"},{"comment":"The certificate is issued for a fixed convex decision instance, namely the sequentially linearised problem produced from a particular warm start. The online WS solve and the cascade reference maintain their own warm starts, so at the same controller tick they can be solving different linearized instances. Theorem III.3 certifies only the instance at which it is evaluated. Consequently, the per-tick compliance match and event-level divergence in Section VII, and the certificate-lifetime measurements in Section VI-A, may be driven by warm-start mismatch rather than by w_dep lying outside Ω(p*) or by certificate expiry. The paper lists this as a source of divergence in Section VIII but does not separate it empirically. Please add an ablation that holds warm-start identity fixed between the WS and cascade solves (or computes the certificate for the exact linearization used by the online WS solve) and re-measures event-level divergence and certificate lifetime; without such a control, the monitored-deployment conclusion remains conditional.","section":"Section V-B; Section VIII, first divergence source"}],"minor_comments":[{"comment":"The sentence 'because [a]_+^p is convex and Non-decreasing convexity for p in {1,2} ensures...' is ungrammatical and incomplete; it should say, for example, 'because [a]_+^p is convex and nondecreasing for p ∈ {1,2}, the composition V_i is convex.'","section":"Section II-B"},{"comment":"The sentence beginning 'A dynamics-agnostic reference implementation reproduces both examples to 10^-6 and implements the scaled-subgradient formulation β_{i,j}=w_i α_{i,j}.10^-6 and implements...' contains a duplicated fragment and should be rewritten.","section":"Section IV-B"},{"comment":"The per-level tolerance is introduced as a scalar τ>0 but then used as a vector τ_ℓ; please define τ_ℓ explicitly before the definition of lex-Pareto dominance.","section":"Section V-F, Definition V.1"},{"comment":"The phrase 'median lifetime is one sampling interval (zero subsequent ticks at the native rate)' conflates the 2.5 s sampled study, whose median lifetime is one 2.5 s interval, with the dense one-episode study, whose median lifetime is zero subsequent 0.1 s ticks; please state the two sampling rates and the sample sizes explicitly when this summary is given.","section":"Abstract and Section VI-A"},{"comment":"The eligibility column entries 'no a' are cryptic in the main text; the table caption should spell out the footnote that the notation refers to infeasible Chebyshev centre.","section":"Table II"}],"recommendation":"major_revision","confidential_remarks":"This is a borderline major_revision. The theoretical core is sound and the paper is unusually honest about its limitations, but the central empirical claim about monitored weighted deployment needs either the warm-start ablation I request in Major Comment 2 or a careful scoping of the claim. I would not reject: with that ablation or a reframed conclusion, the paper would be a solid contribution to the MPC and rulebook literature."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this paper has a sound, mostly standard geometric core wrapped in an honest empirical package that measures how short-lived weighted-sum certificates are in receding-horizon MPC. The headline result is that a weighted sum reproduces the lexicographic optimum iff the augmented weight (w,1) supports the upper image at the lex point, a normal-cone membership condition. As the reader's take notes, that characterization is close to definitional; the novelty and value are in making it computable. Under l1 hinges the equivalent-weight set is a polyhedron via a scaled-KKT projection in beta_{i,j}=w_i alpha_{i,j}, membership is one LP, and a Chebyshev center returns a certified margin. That genuinely reframes the w1>>w2>>... heuristic as a membership question. The l2 foreclosure is also handled correctly: with nonzero limiting multiplier no finite weight is exact, with O(1/w) decay, and it is verified on 1000 randomized instances and two worked examples.\n\nThe most valuable part is the persistence measurement. On held-out receding-horizon problems the robust intersection of pointwise regions is empty in every class, median lifetime is one sampling interval (zero subsequent ticks at native rate), and validity tracks active-set churn. That is a real, non-obvious finding for practitioners, and it motivates monitored weighted solves with selective cascade fallback. The authors are candid about limits, including monitor sensitivity 0.55 and the exploratory label on the guarded-condition dominance result.\n\nSoft spots, in proportion. The near-equal tier ratios in Table II are inflated by the symmetric operator box [1,1e8]^3: many calibrated weights sit at the 5e7 box corner, so 1:1:1 is partly box artifact. The paper does run a box-sensitivity sweep, which helps, but the claim that near-equality is the typical geometry is weaker than it looks. More importantly, the closed-loop attribution of divergence to certificate expiry is confounded by warm-start mismatch. The WS and cascade runs keep separate warm starts and may be solving different linearized convex instances at the same tick; Theorem III.3 certifies only the fixed instance. Section VIII itself lists different warm starts as a divergence source, but no ablation holds warm-start identity fixed. So the per-tick compliance match and event-level divergence could reflect instance mismatch rather than w departing from Omega(p*). This does not overturn the pointwise theorem or the cascade-derived persistence measurement, but it weakens the claim that certificate expiry explains the closed-loop event record. The reproducibility artifact is also referenced but not directly accessible in the text, which is annoying for verification.\n\nWho this is for: people working on rulebook MPC, lexicographic multi-objective optimization, or soft-constrained MPC generally. The geometry and LP construction are worth knowing even if the closed-loop evidence is conditional. The paper deserves a serious referee: the math core is sound, the empirical honesty is unusual, and the persistence finding is valuable. I would engage with it, expecting the closed-loop attribution to need a warm-start-controlled ablation before the monitored-solve architecture claim is solid.","headline":"Sound geometric core with a computable LP for exact weighted-sum weights under l1, honest persistence measurements showing certificates expire at native rate, but closed-loop divergence attribution is confounded by warm-start mismatch.","tokens_in":28188,"tokens_out":3256,"would_cite":true,"duration_ms":27396,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C25","90C29","90C05","93B45"],"pacs":[],"model":"deepseek-v4-flash","headline":"A weighted sum reproduces a convex lexicographic optimum exactly when its weight supports the upper image at the lex point; membership is decidable by linear feasibility.","keywords":["lexicographic optimization","weighted-sum scalarization","model predictive control","rulebooks","convex optimization","normal cone","exact penalty","autonomous driving"],"falsifier":"Take a convex quadratic program with one binding squared-hinge constraint whose limiting multiplier is nonzero, as in the paper's closed-form example, and solve it at increasing finite weights: the theory predicts the constraint violation stays strictly positive at every finite weight, decaying as $\\lambda^{*}/(2w)+O(w^{-2})$. Observing exact satisfaction (violation at machine zero) at any finite weight, or a decay rate that does not match the predicted leading term, would refute the $\\ell^2$ foreclosure theorem.","tokens_in":27251,"feed_emoji":"🚦","tokens_out":22317,"duration_ms":166577,"temperature":0.7,"pith_summary":"This paper establishes when a single weighted-sum optimization can stand in for the $L+1$ sequential solves of a lexicographic cascade in convex model-predictive control, the setting of automated-driving rulebooks that rank rule violations by priority tiers. The answer is geometric: the weighted sum matches the lexicographic optimum exactly when the weight vector, paired with a unit coefficient on the performance cost, supports the upper image of the achievement map at the lexicographic point, and the valid weights form the unit-performance slice of an outward normal cone whose membership is decidable by linear feasibility. Under $\\ell^1$ hinge penalties that slice is a polyhedron built from a scaled-KKT system, so a linear program returns an interior weight with a certified margin; under $\\ell^2$ squared-hinge penalties the paper proves a dichotomy, namely that no finite weight is exact whenever the limiting multiplier is nonzero, with violation decaying only as $O(1/w)$. The practical payoff is a calibration procedure whose certified weights are near-equal across priority tiers in nine of eleven scenario classes, contradicting the standard 100:10:1 separation heuristic and roughly doubling legal-tier event precision in closed-loop simulation. The measured catch is that certificates are pointwise: no single weight survives across the ticks of an episode, median lifetime is one sampling interval, and persistence tracks active-set stability, which motivates monitored weighted solves with selective cascade fallback.","feed_headline":"One cone-membership check replaces the 100:10:1 weight rule","feed_subtitle":"Exactness is decided by a normal-cone slice, not a big ratio; squared hinges stay approximate at any finite weight","key_machinery":"The central object is the upper image of the achievement map, $P=\\operatorname{cl}\\{(y,t):\\exists z\\in Z,\\ V_i(z)\\le y_i\\ \\forall i,\\ J(z)\\le t\\}$, whose lexicographic point $p^{*}=(V^{*}_1,\\dots,V^{*}_L,J^{*})$ is an extreme point under the convexity assumptions. The argument rides on the outward normal cone $N^{+}_{P}(p^{*})$ and its unit-performance slice $\\Omega(p^{*})=\\{w\\in\\mathbb{R}^{L}_{++}:(w,1)\\in N^{+}_{P}(p^{*})\\}$: a weight is exact exactly when the augmented vector $(w,1)$ is a supporting normal at $p^{*}$. For $\\ell^1$ hinges the working construction is the scaled-KKT system, which replaces boundary-hinge subgradient terms $w_i\\alpha_{i,j}$ by variables $\\beta_{i,j}=w_i\\alpha_{i,j}$ with $0\\le\\beta_{i,j}\\le w_i$; projecting this affine system onto $w$ gives the polyhedral region, and a Chebyshev-centre linear program returns an interior certified weight with an inscribed-ball margin. For $\\ell^2$ squared hinges the load-bearing mechanism is the vanishing boundary force $\\nabla[g]^2_{+}=0$ at $g=0$, which drives the foreclosure theorem and the $O(1/w)$ local-minimizer branch. The online apparatus is a compliance-pattern monitor that compares realized per-tier compliance bits against the cached calibration pattern and triggers cascade fallback on mismatch.","core_discovery":"The paper's central claim is that, under convexity of the feasible set, the constraint functions, and the performance objective, the lexicographic optimum $z^{*}_{lex}$ is a minimizer of the weighted sum $J(z)+\\sum_{i=1}^{L} w_i V_i(z)$ for a weight $w\\in\\mathbb{R}^{L}_{++}$ if and only if $w$ lies in $\\Omega(p^{*})$, the unit-performance slice of the outward normal cone of the upper image $P=\\operatorname{cl}\\{(y,t):\\exists z\\in Z,\\ V_i(z)\\le y_i\\ \\forall i,\\ J(z)\\le t\\}$ at the lexicographic image point $p^{*}=\\Phi(z^{*}_{lex})$. Equivalence is therefore a membership question, not a magnitude question: the separation heuristic $w_1\\gg w_2\\gg\\cdots\\gg w_L$ answers the wrong question. In the $\\ell^1$ hinge regime the slice is a polyhedron obtained by projecting a scaled-KKT system in the variables $\\beta_{i,j}=w_i\\alpha_{i,j}$, so a Chebyshev-centre linear program returns an interior weight with a certified margin; in the $\\ell^2$ squared-hinge regime the boundary force vanishes, no finite weight is exact whenever the limiting multiplier is nonzero, and violation along the local minimizer branch decays as $\\Theta(1/w_i)$ for the raw constraint and $\\Theta(1/w_i^2)$ for the violation functional. The same construction yields the multi-instance intersection of pointwise slices, and the empirical campaign shows that intersection is empty across every sampled episode: the certificate is exact where it is issued and expires quickly, with persistence tracking active-set stability.","pith_inferences":["Because cross-tick validity was perfect for the two classes with stable active sets and at most 0.40 for the classes with churning active sets, a cheaper and possibly more sensitive expiry signal than the paper's compliance-pattern monitor (sensitivity 0.55) would be to detect active-set change itself and re-run the scaled-KKT feasibility check only when the active set moves.","The $\\ell^1$/ $\\ell^2$ dichotomy is really a statement about boundary derivatives, so it should extend to any smooth surrogate: any penalty whose derivative vanishes at the constraint boundary (Huber, log-barrier, or $\\ell^p$ with $p>1$) should require infinite weight for exactness, while any penalty keeping a nonzero boundary subgradient should admit finite exact weights, a testable generalizatio","The polyhedral description of $\\Omega(p^{*})$ together with its Chebyshev margin gives a natural convex formulation for inverse optimal control: recover rule priorities from demonstrated trajectories by fitting weights inside the certified region while maximizing the margin, rather than treating weight learning as a black-box regression."],"forward_implications":["Exactness of a weighted-sum surrogate for a lexicographic cascade is decidable by a single linear feasibility check per decision instance, turning weight calibration from heuristic tuning into a convex computation.","The separation heuristic $w_1\\gg w_2\\gg\\cdots\\gg w_L$ is neither necessary nor sufficient: certified weights in nine of eleven scenario classes are near-equal across tiers (ratios near $1:1:1$ or $2:1:1$), and where the region actively constrains the weights the comfort-tier lower bound is about 89, above the heuristic's 10.","Under $\\ell^2$ squared-hinge penalties, exact equivalence at finite weight is impossible whenever a limiting multiplier is nonzero, so smooth penalties can only approximate the priority order, with violation decaying as $O(1/w)$.","Certificates are pointwise and expire quickly, with median lifetime one sampling interval and the multi-instance intersection empty in all fourteen sampled classes, so a deployable controller needs a validity check as the horizon evolves; the paper's monitored construction with selective cascade fallback is the proposed architecture.","Per-tick compliance agreement can saturate at the top priority levels (median 1.000) while event-level divergence persists, so surrogate validation must track event precision and recall, not only tick-wise agreement."],"supporting_citations":[{"why":"Defines the rulebook formalism whose lexicographic priority semantics the paper implements; supplies the 25-rule structure for the empirical campaign.","marker":"[1]"},{"why":"Supplies the supporting-hyperplane and normal-cone machinery on which the equivalence characterization (Theorem III.3) is built.","marker":"[3]"},{"why":"Gives the existence result for equivalent weights in lexicographic linear programs, which the paper recovers as non-emptiness of the unit-performance slice.","marker":"[4]"},{"why":"Provides explicit weight bounds for lexicographic optimization, read as sufficient membership conditions for the equivalence region.","marker":"[5]"},{"why":"Exhibits lexicographic programs that admit no finite weighting due to instance structure, the contrast case for the paper's penalty-driven $\\ell^2$ impossibility.","marker":"[6]"},{"why":"Gives first-order stationarity conditions for lexicographic problems, which the scaled-KKT projection refines into a characterization of the full weight set.","marker":"[7]"},{"why":"Establishes the exact-penalty threshold, a hinge being exact at finite weight above the dual norm of the optimal multiplier, which the paper lifts from one constraint to $L$ priority levels.","marker":"[18]"},{"why":"Provides the classical scalar soft-constraint exactness result in MPC that the paper generalizes from a threshold to a full-dimensional polyhedral region.","marker":"[20]"},{"why":"Supplies the upper-image construction from vector minimization that is the paper's central geometric object.","marker":"[30]"},{"why":"Provides the closed-loop driving simulation benchmark on which the calibration and closed-loop experiments are run.","marker":"[31]"}],"fun_headline_variants":["Exact MPC weights: a cone slice, not a big ratio","Squared-hinge MPC weights never go exact at finite w","Weight certificates: polyhedral slice for l1, none for l2","MPC weight exactness: a normal-cone slice, not a heuristic","Certified weight margins via Chebyshev centers, but pointwise"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The certificate is proved for one fixed convex decision instance, but the closed-loop planner solves a new locally linearized problem every tick, so a weight that is exact at the calibration tick need not be exact at the next, and the campaign does not isolate that expiry from divergence caused by different warm starts.","fun_headline_variants_meta":{"raw":{"variants":["Exact MPC weights: a cone slice, not a big ratio","Squared-hinge MPC weights never go exact at finite w","Weight certificates: polyhedral slice for l1, none for l2","MPC weight exactness: a normal-cone slice, not a heuristic","Certified weight margins via Chebyshev centers, but pointwise"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000645,"raw_usage":{"total_tokens":3102,"prompt_tokens":1222,"completion_tokens":1880,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":838,"completion_tokens_details":{"reasoning_tokens":1787}},"tokens_in":838,"tokens_out":1880,"duration_ms":10806,"temperature":1.0,"reasoning_tokens":1787,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:09:23.545341+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a convex quadratic program with one binding squared-hinge constraint whose limiting multiplier is nonzero, as in the paper's closed-form example, and solve it at increasing finite weights: the theory predicts the constraint violation stays strictly positive at every finite weight, decaying as $\\lambda^{*}/(2w)+O(w^{-2})$. Observing exact satisfaction (violation at machine zero) at any finite weight, or a decay rate that does not match the predicted leading term, would refute the $\\ell^2$ foreclosure theorem.","supporting_citations":[{"cited_title":"Liability, ethics, and culture-aware behavior specification using rulebooks,","cited_arxiv_id":null,"evidence_quote":"Defines the rulebook formalism whose lexicographic priority semantics the paper implements; supplies the 25-rule structure for the empirical campaign."},{"cited_title":"Equivalent weights for lexicographic multi-objective programs: characterizations and computations,","cited_arxiv_id":null,"evidence_quote":"Gives the existence result for equivalent weights in lexicographic linear programs, which the paper recovers as non-emptiness of the unit-performance slice."},{"cited_title":"Explicit weight bounds for lexicographic optimization,","cited_arxiv_id":null,"evidence_quote":"Provides explicit weight bounds for lexicographic optimization, read as sufficient membership conditions for the equivalence region."},{"cited_title":"A theory of lexicographic multi-criteria optimization,","cited_arxiv_id":null,"evidence_quote":"Gives first-order stationarity conditions for lexicographic problems, which the scaled-KKT projection refines into a characterization of the full weight set."},{"cited_title":"Soft constraints and exact penalty functions in model predictive control,","cited_arxiv_id":null,"evidence_quote":"Provides the classical scalar soft-constraint exactness result in MPC that the paper generalizes from a threshold to a full-dimensional polyhedral region."},{"cited_title":"Ehrgott,Multicriteria Optimization, 2nd ed","cited_arxiv_id":null,"evidence_quote":"Supplies the upper-image construction from vector minimization that is the paper's central geometric object."},{"cited_title":"nuPlan: a closed-loop ML- based planning benchmark for autonomous vehicles,","cited_arxiv_id":null,"evidence_quote":"Provides the closed-loop driving simulation benchmark on which the calibration and closed-loop experiments are run."}],"review_version":1}