{"id":"8d28df97-69c5-4a22-bc17-11fe462cde55","arxiv_id":"2505.07287","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"Co-identifying qLPV models and robust invariant sets with interval-bound-propagation tightening gives larger invariant sets than previous concurrent synthesis, with no loss of prediction accuracy on the test oscillator.","lead":"This paper trains a quasi-LPV model and, at the same time, finds a robust control invariant set that the model guarantees, using interval bounds on the scheduling function to shrink the uncertainty. The result is a controller for constrained nonlinear plants with less conservative safety sets than the authors' earlier approach.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Plant-level guarantee rests on unverified κ=1.01 in W; if [1, Prop. 4]'s lower bound is larger, the reported reduced-conservativeness RCI sets are not certified for the oscillator.","rationale":"The reader's weakest_assumption and my own analysis converge on the same load-bearing concern: the validity of the data-driven disturbance set W in (12) for the true plant. All plant-level guarantees in the paper pass through Proposition 1, and the numerical demonstration of reduced conservativeness is only meaningful if W contains C E ⊕ V. The paper cites [1, Prop. 4] for a lower bound on κ but neither states nor verifies that bound for the identified model. I considered other potential concerns, such as the quantifier structure in Corollary 1 (which as written allows a to depend on (i,j), potentially loosening ˜S beyond what Proposition 3 requires), but Algorithm 1 uses a single a from BoundProp, so the numerical results remain valid at the model level. The W issue is more fundamental because it undermines the connection to the physical oscillator, which is the stated purpose of the framework. This does not change the reader's CONDITIONAL verdict: the mathematical construction appears sound, but the headline claim requires the authors to either verify κ = 1.01 for the identified model or state and use the provable bound from [1, Prop. 4].","tokens_in":10855,"tokens_out":12048,"duration_ms":114399,"concrete_test":"Re-run the Section 5.0.1 experiment, but for the identified Θ_0.07 compute the minimal invariant set E for the observer error dynamics (as in [1, Prop. 4]) and verify C E ⊕ V ⊆ W with κ = 1.01. If this fails, raise κ to the provable lower bound from [1, Prop. 4], recompute all d_ζ values, and check whether the reported reduction below d_base = 5.0936 persists; also rerun the Figure 3 closed-loop simulation with the corrected W.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline benefit—RCI sets with reduced conservativeness that certify closed-loop regulation of the oscillator—flows through Proposition 1, which requires the data-driven disturbance set W in (12) to satisfy (7), W = C E ⊕ V, for a certified invariant observer error set E and output error set V. Section 2.2.1 sets W as a κ-inflated box and simply assumes κ = 1.01 is adequate by citing [1, Prop. 4], without stating the required lower bound or checking it for the identified model. If the true required κ exceeds 1.01, the RCI set X(q̂_k) certified for (21) is not certified for the physical plant, and the low d_ζ values in Figure 1 are only a property of an uncertified auxiliary model. Since Section 5 is the only numerical evidence for the central claim, the claim is conditional on an unverified inclusion. The paper even acknowledges the need for 'sufficiently large κ' in Section 5.0.2 but never quantifies it.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a concurrent identification framework for qLPV models and robust control invariant (RCI) sets, extending the authors' earlier work [1]. The main idea is to replace the full simplex-valued uncertainty set of the scheduling-dependent model with a tightened convex hull \\tilde\\Delta, obtained by using lower bounds a_i on each scheduling function p_i(z) over the candidate RCI set. This yields a new regularization function r(\\Theta) defined as the optimal value of a nonlinear robust optimization problem, which is then evaluated by an iterative algorithm that alternates interval bound propagation (IBP) with QP solves. The resulting differentiable surrogate is embedded in a gradient-based training loop for joint model identification and RCI-set construction. A numerical oscillator example reports that the obtained RCI sets have markedly smaller size values d_\\zeta than two benchmarks, while BFR scores remain around 90.7-90.9.","tokens_in":11115,"tokens_out":8529,"duration_ms":88467,"significance":"The tightening idea is sound and potentially useful: Proposition 3 and Corollary 1 give a clean construction in which a lower-bound vector a on the scheduling functions over X(q) yields a less conservative uncertain LTI enclosure, and the observation that a=0 recovers the previous formulation guarantees no additional conservativeness in that branch. The reported numerical reductions (d_\\zeta significantly below d_seq=22.6924 and d_base=5.0936) are encouraging, and the release of code is a positive feature. The strength of the paper is the derivation itself; its main weaknesses are that the plant-level guarantee depends on an unverified disturbance-set inclusion and that the algorithm used to evaluate the central regularization function has no stated convergence or approximation guarantees.","major_comments":[{"comment":"The numerical demonstration of the central claim is conditional on an unverified disturbance-set inclusion. Proposition 1 requires W = CE \\oplus V in (7); Section 2.2.1 defines W in (12) as a \\kappa-inflated box and refers to [1, Prop. 4] for lower bounds on \\kappa. Section 5.0.1 fixes \\kappa = 1.01 without reporting whether that value satisfies the required lower bound for the identified model. If the true required \\kappa exceeds 1.01, the RCI sets certified for (21) and the d_\\zeta values in Figure 1 are not certified for the oscillator plant, and the closed-loop trajectories in Figure 3 lack the claimed guarantee. The statement in Section 5.0.2 that 'For sufficiently large values of \\kappa>1' Problem (27) is recursively feasible only reinforces the need to verify or quantify this assumption. Please verify the lower bound for the identified model or explicitly restrict the numerical claims to the auxiliary uncertain model.","section":"Sec. 5.0.1, Eq. (12) and Proposition 1"},{"comment":"The paper replaces the nonlinear robust optimization problem (23) by a finite sequence of QPs over \\hat S(q_k,a_k) and uses the output r_{\\hat k} as the regularization value, but no convergence, monotonicity, or even feasibility guarantee is provided for Algorithm 1; the text states that theoretical properties are future research. Consequently, the d_\\zeta numbers in Figure 1 are upper bounds that depend on the initialization q_0, on \\zeta, and on \\hat k (here \\hat k=200), rather than solutions of (23). Since r(\\Theta;q_0) is then used as a differentiable objective in Algorithm 2, the training procedure optimizes a surrogate whose relationship to the exact r(\\Theta) is uncharacterized. Please provide at least a feasibility/monotonicity statement for Algorithm 1 or report the results as feasible RCI-set sizes rather than as values of the regularizer.","section":"Sec. 3.2 and Algorithm 1"},{"comment":"The definition of \\tilde S is formally ill-posed. It reads '\\exists a \\in [0,p(z)], \\forall z\\in X(q)' inside the set description, which if interpreted literally allows a to depend on z. However, Proposition 3 and the proof of Corollary 1 require a single vector a satisfying 0 \\le a_i \\le p_i(z) for all z\\in X(q). Please restate the existential quantifier outside the universal quantification over z, i.e., 'there exists a such that 0\\le a_i\\le p_i(z) for all z\\in X(q)', before the vertex constraints.","section":"Corollary 1, definition of \\tilde S"}],"minor_comments":[{"comment":"The proof divides by 1 - \\sum_j a_j, which vanishes when \\sum_j a_j = 1; this case is not treated separately. The conclusion is still true (in that case p(z)=a on X(q) and \\tilde\\Delta is a singleton), but the proof should handle it explicitly or by a continuity argument.","section":"Proposition 3 proof"},{"comment":"In Proof 2, the phrase 'observing that S is RCI for (21)' should read 'observing that X(q) is RCI for (21)'.","section":"Corollary 1 proof"},{"comment":"The componentwise nature of \\bar w := \\max_{w\\in W} w and \\underline w := \\min_{w\\in W} w should be stated explicitly, since W is a set of vectors.","section":"Sec. 2.2.1, Eq. (12)"},{"comment":"The caption does not identify the dashed horizontal lines; please label the d_seq and d_base values and distinguish them from the plotted d_\\zeta curve.","section":"Figure 1 caption"},{"comment":"The claim that choosing \\hat k>1 'often results in Algorithm 2 converging to suboptimal points' is unsupported; either provide evidence or soften this statement.","section":"Sec. 5.0.1"},{"comment":"Reference [19] contains a typographical error in the journal name: 'IIEEE' should be 'IEEE'.","section":"Reference list"}],"recommendation":"major_revision","confidential_remarks":"The manuscript relies heavily on the authors' own companion paper [1] for the baseline regularizer, the characterization of W, the initialization, and the configuration-constraint machinery. The novel tightening is clean and publishable in principle, but the verification of \\kappa and the missing analysis of Algorithm 1 should be addressed before the numerical claims can be accepted at face value."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core step here is Proposition 3, and it is correct: by using per-state lower bounds on the softmax scheduling functions, the qLPV model is enclosed in a tighter multiplicative uncertainty set, and the set-inclusion proof is clean. The observation that a = 0 recovers the earlier QP formulation means the new condition is no more conservative than the old one—that is exactly the right kind of extension. Building the new regularization on interval bound propagation is also a legitimate and practical refinement, and the released code plus the differentiability argument make it usable.\n\nThe numerical demonstration does show the claimed effect: for the oscillator example, the d-values fall well below both d_seq and d_base while the BFR stays essentially flat. So the method appears to work on the test problem.\n\nThe soft spots are real but not fatal. The most important is the disturbance set W in (12). The paper cites [1, Prop. 4] for a lower bound on κ that would certify W = CE ⊕ V, then simply sets κ = 1.01 in the experiments without stating or checking that bound. If the required κ is larger, then the RCI sets certified for the uncertain model are not certified for the physical oscillator. This makes the central claim conditional, and the paper itself only says 'sufficiently large κ' in the controller-synthesis section without quantifying it. That needs to be fixed.\n\nSecond, Algorithm 1 has no convergence proof. The authors acknowledge this and show one monotonic run in Figure 2, but a single trajectory is not evidence of general behavior. Third, the numerical study is one plant, one random dataset, no seed variation, and the comparison with d_seq is somewhat apples-to-oranges because the sequential benchmark uses a maximal RCI set while the concurrent benchmarks use configuration-constrained polytopes. The comparison with d_base is fair, and that alone carries the point.\n\nThe reliance on the authors' prior work is heavy but honest—[1] contains the base architecture, the W characterization, and the initialization. Nothing here is hidden.\n\nThis paper is for researchers working on LPV identification with formal control guarantees. It deserves a serious referee, not a desk reject. I would recommend acceptance after the authors verify the κ condition on their example, add at least a seed variation or a second example, and either prove or more carefully qualify the convergence of Algorithm 1.","headline":"A correct and practically useful tightening of the authors' own concurrent qLPV identification scheme, but the advertised reduction in conservativeness is verified only for one simulation and rests on an unverified disturbance-set inclusion.","tokens_in":11635,"tokens_out":1284,"would_cite":false,"duration_ms":14751,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93B30","93D09","90C47","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that identifying a qLPV model together with a robust control invariant set can be made less conservative by shrinking the model's multiplicative uncertainty with interval bound propagation on the softmax scheduling…","keywords":["quasi-LPV models","robust control invariant sets","concurrent synthesis","control-oriented regularization","interval bound propagation","multiplicative uncertainty","softmax scheduling","system identification"],"falsifier":"Collect a long validation set from the true plant, simulate the observer (5) with the identified model, and form the samples $w_t=y_t-Cz_t^w$; if any sample lies outside the box $\\mathcal W=\\{w: |w-c_w|\\le 1.01\\,\\epsilon_w\\}$ used at identification time, the RCI certificate is invalid for the plant. Equivalently, run the closed-loop controller (27) beyond the training horizon; a single trajectory leaving $Y$ disproves the guarantee.","tokens_in":10643,"feed_emoji":"🛡️","tokens_out":11109,"duration_ms":107030,"temperature":0.7,"pith_summary":"The paper tries to establish that a quasi-linear parameter-varying (qLPV) model learned from input-output data can be certified, at the same time as it is identified, to admit a robust control invariant (RCI) set, meaning a set of states from which a controller can keep the plant inside given output and input constraints despite disturbances, and that the certified set can be considerably less conservative than in previous work. The mechanism is a tightened uncertain linear model: instead of letting the scheduling vector range over the whole simplex, the paper uses lower bounds on each softmax scheduling function over the candidate invariant set to shift the vertex matrices and shrink the multiplicative uncertainty. The lower bounds are computed by interval bound propagation through the scheduling network, so the whole pipeline remains differentiable and can be embedded in gradient-based system identification. If the claim is right, concurrent identification of models and safety guarantees becomes less conservative without degrading prediction accuracy.","feed_headline":"Tighter qLPV uncertainty shrinks safe invariant sets below baseline","feed_subtitle":"New regularization certifies smaller safe sets for learned qLPV models without losing fit accuracy.","key_machinery":"The load-bearing object is the tightened multiplicative-uncertainty hull $\\tilde\\Delta \\subseteq \\Delta$ in (21)-(22), together with the interval-bound-propagation step that produces valid lower bounds $a_i \\le p_i(z)$ on an inflated bounding box $B(q)$ around the candidate invariant set. Proposition 3 shows that invariance with respect to $\\tilde\\Delta$ is sufficient for invariance of the qLPV model, and Corollary 1 turns that into the constraint set $\\tilde S$. Algorithm 1 alternates bound propagation (step 2) with a differentiable quadratic program over $\\hat S(q_k,a)$ (step 3), so the regularization value is differentiable and can serve as a term in the identification objective. The configuration-constrained polytope parameterization $X(q)=\\mathrm{CH}\\{V_j q,\\ j\\in I_v^1\\}$ is what allows the RCI conditions to be written as finite linear inequalities.","core_discovery":"The central claim is Proposition 3: if, on the candidate configuration-constrained polytope $X(q)$, each scheduling component satisfies $p_i(z)\\ge a_i$, then the qLPV system is contained in an uncertain LTI system $\\tilde\\Delta$ whose vertices are the shifted matrices $\\tilde A_i=(1-\\sum_{j=1}^{n_p} a_j)A_i+\\sum_{j=1}^{n_p} a_j A_j$ (and similarly for $B$ and $L$). Consequently any robust control invariant set certified for $\\tilde\\Delta$ is automatically invariant for the qLPV model. The paper defines the new regularization $r(\\Theta)=\\inf_{(q,v)\\in\\tilde S} d(\\bar A,\\bar B,C,q)$, where $\\tilde S$ is the constraint set from Corollary 1; since $\\tilde S$ contains the earlier set $S$ (recovered by setting $a=0$), the new value is always at least as good and is typically strictly better. Algorithm 1 evaluates this regularization by alternating interval bound propagation with a differentiable quadratic program, which is what allows it to be used inside the identification loop. On the oscillator example the identified models achieve $d$-values significantly below both $d_{\\mathrm{seq}}=22.6924$ for sequential identification and $d_{\\mathrm{base}}=5.0936$ for the prior concurrent approach, with best-fit rates near 90.7 percent.","pith_inferences":["Editorial inference: the same lower-bound tightening could be applied with exact polytope bounds on the scheduling network instead of an inflated box, which may produce even tighter $\\tilde\\Delta$ and smaller $d$-values at higher computational cost.","Editorial inference: the comparison between $\\zeta$ values is metric-dependent; a metric based on the volume of $X(q)$ rather than the tracking-distance $d$ could rank the identified models differently, so the reported gain should be read as specific to the tracking formulation.","Editorial inference: a natural testable extension is to make $\\zeta$ adaptive during training, shrinking the bounding box as the model improves, which could remove the need to grid over $\\zeta$ and improve convergence behavior.","Editorial inference: the approach's reliance on monotone activations is a practical limitation; replacing interval bound propagation with a differentiable verifier for general activations is the obvious next step and is flagged by the authors as future work."],"forward_implications":["Any qLPV model returned by Algorithm 2 carries the certificate that an RCI set exists, so a tracking controller that keeps the plant output inside $Y$ while respecting input limits can be synthesized from the identified model.","Because $S\\subseteq\\tilde S$, the improved regularization never produces a worse RCI-size score than the earlier QP-based regularization on the same model, and it yields strictly better scores whenever the scheduling lower bounds are informative.","On the oscillator test problem, the approach achieves $d$-values below both benchmarks (22.6924 and 5.0936) while the best-fit rate stays in the narrow band 90.73-90.74 on the training data, so the conservativeness reduction does not come at the cost of model quality.","Running Algorithm 1 for one iteration inside the identification loop is enough for the reported gains, and the value $r_k$ is observed to converge monotonically in the post-hoc evaluation of Figure 2.","Because the proof of Proposition 3 only uses $p(z)\\in\\mathcal P$, the same tightening applies when the scheduling function also depends on the input, as pointed out in Remark 1."],"supporting_citations":[{"why":"introduces the concurrent-synthesis QP regularization and the softmax qLPV parameterization that this paper tightens.","marker":"[1]"},{"why":"gives the standard definition of robust control invariance that the certified sets must satisfy.","marker":"[12]"},{"why":"supplies the configuration-constrained polytope parameterization and the vertex-map inequalities used in Proposition 2 and Corollary 1.","marker":"[13]"},{"why":"supports the observer-error invariant-set argument that turns output data into the disturbance set W in Proposition 1.","marker":"[14]"},{"why":"supplies interval bound propagation, the method used to compute the scheduling lower bounds a_i.","marker":"[17]"},{"why":"provides the differentiable QP solver that lets Algorithm 1's output be differentiated inside the identification loop.","marker":"[23]"},{"why":"gives the initial qLPV model and the Best Fit Rate score used to evaluate prediction quality and initialize the algorithm.","marker":"[24]"},{"why":"computes the maximal robust control invariant set used as the sequential benchmark d_seq.","marker":"[25]"}],"fun_headline_variants":["qLPV bound shrinks safe sets below baseline","Uncertain LTI wrap cuts qLPV conservativeness","Tighter invariant sets from bound-propagation","Reduced conservativeness for qLPV robust invariant sets","Smaller certified safe sets for qLPV models"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The data-driven disturbance box $\\mathcal W$, inflated by $\\kappa=1.01$, is assumed to contain all disturbances of the true plant; if it does not, the certified invariant set says nothing about the physical system.","fun_headline_variants_meta":{"raw":{"variants":["qLPV bound shrinks safe sets below baseline","Uncertain LTI wrap cuts qLPV conservativeness","Tighter invariant sets from bound-propagation","Reduced conservativeness for qLPV robust invariant sets","Smaller certified safe sets for qLPV models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000329,"raw_usage":{"total_tokens":1841,"prompt_tokens":956,"completion_tokens":885,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":572,"completion_tokens_details":{"reasoning_tokens":807}},"tokens_in":572,"tokens_out":885,"duration_ms":8516,"temperature":1.0,"reasoning_tokens":807,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:20:14.477398+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Collect a long validation set from the true plant, simulate the observer (5) with the identified model, and form the samples $w_t=y_t-Cz_t^w$; if any sample lies outside the box $\\mathcal W=\\{w: |w-c_w|\\le 1.01\\,\\epsilon_w\\}$ used at identification time, the RCI certificate is invalid for the plant. Equivalently, run the closed-loop controller (27) beyond the training horizon; a single trajectory leaving $Y$ disproves the guarantee.","supporting_citations":[{"cited_title":"Combined learning of linear parameter-varying models and robust control invariant sets,","cited_arxiv_id":null,"evidence_quote":"introduces the concurrent-synthesis QP regularization and the softmax qLPV parameterization that this paper tightens."},{"cited_title":"Blanchini and S","cited_arxiv_id":null,"evidence_quote":"gives the standard definition of robust control invariance that the certified sets must satisfy."},{"cited_title":"Configuration-constrained tube MPC,","cited_arxiv_id":null,"evidence_quote":"supplies the configuration-constrained polytope parameterization and the vertex-map inequalities used in Proposition 2 and Corollary 1."},{"cited_title":"Robust output feedback model predictive control of constrained linear systems,","cited_arxiv_id":null,"evidence_quote":"supports the observer-error invariant-set argument that turns output data into the disturbance set W in Proposition 1."},{"cited_title":"An L-BFGS-B approach for linear and nonlinear system identification under ℓ1 and group-lasso regularization,","cited_arxiv_id":null,"evidence_quote":"gives the initial qLPV model and the Best Fit Rate score used to evaluate prediction quality and initialize the algorithm."},{"cited_title":"Infinite time reachability of state-space regions by using feedback control,","cited_arxiv_id":null,"evidence_quote":"computes the maximal robust control invariant set used as the sequential benchmark d_seq."}],"review_version":1}