{"id":"e69e078a-7d43-42c9-84a8-d982b005e56d","arxiv_id":"2411.10262","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"This paper gives LMI conditions for designing interval observers that monitor state safety bounds in neural-network-controlled systems, demonstrated on a lateral vehicle control simulation.","lead":"The paper designs a safety monitor, an interval observer, that computes guaranteed upper and lower bounds on the state of a nonlinear system containing a neural network. It does this by abstracting the activation functions with quadratic constraints and solving linear matrix inequalities.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The interval-inclusion property rests on Lemma 4, but the auxiliary network in (16) as printed appears to pair nonpositive weights with the lower state, which reverses the sign of Φ(x)−Φ(x,x) and would invalidate (12).","rationale":"The reader and I converge on the same load-bearing point: Lemma 4 and the sign conditions (12)-(13) are the mechanism that makes the error system cooperative, and the paper neither proves Lemma 4 nor removes the ambiguity in (16)-(17). I chose this over the practical-stability epsilon gap because the epsilon step is repairable: (37) makes Γ6 negative on the positive orthant, and the neural network is globally Lipschitz, so a uniform ε can be extracted; it does not threaten the inclusion property. The sign condition, by contrast, is existential: if the nonpositive/nonnegative weight pairing is reversed, the lower auxiliary network bounds the wrong side and the interval observer no longer brackets the state. The scalar example makes the issue concrete and testable. Since a corrected formula or a clarifying footnote may fully resolve the concern, and since the simulation is supportive but not disambiguating, the reader's CONDITIONAL verdict remains appropriate. The paper should restate (16)-(17) unambiguously, include a proof or a precise citation of Lemma 4, and ideally release the simulation and a sign-check script.","tokens_in":19478,"tokens_out":21398,"duration_ms":217298,"concrete_test":"Use the scalar counterexample: Φ:R→R is a one-layer network with W=-1, b=0, and identity final activation, so Φ(x)=-x. Construct Φ_lower and Φ_upper exactly as written in (16)-(17), with W_l=-1 and Wbar_l=0. Evaluate at x_lower=0, x=1, x_upper=2. If (16) is read literally, Φ_lower=-0=0 and Φ(x)-Φ_lower=-1-0=-1, violating (12), while Lemma 4 predicts +1. Repeat for random 2-layer tanh networks on a dense grid x_lower≤x≤x_upper and check whether every component of Φ(x)-Φ_lower and Φ_upper-Φ(x) is nonnegative. Also test the corrected formula v_lower = Wbar ω_lower + W_l ω_upper; only this version should pass. This single check distinguishes a typesetting artifact from a genuine mathematical error in the auxiliary-network design.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Proposition 1 and Theorem 2 guarantee x≤x≤x only if (12)-(13) hold, i.e., Φ(x)−Φ(x,x)≥0 and Φ(x,x)−Φ(x)≥0 elementwise. These inequalities are passed to Lemma 4, cited from Xiang (2021), without a proof in this paper. The construction is load-bearing because the positivity of e and e in the error system (9), and hence the entire interval-observer claim, depends on it. In the manuscript, (15) defines a nonpositive part W (entries w_ij<0, else 0) and a nonnegative part Wbar (entries w_ij≥0, else 0). Read literally, the lower auxiliary network in (16) is v_l = W_l ω_l + Wbar_l ωbar_l, i.e., the nonpositive weights multiply the lower state and the nonnegative weights multiply the upper state. For a scalar one-layer network with W=-1, bias 0, no activation in the final layer, W_l=-1 and Wbar_l=0. For x_lower≤x≤x_upper, Φ_lower = -x_lower, so Φ(x)-Φ_lower = x_lower - x ≤0, the opposite of (12). The correct mixed-monotone lower construction needs v_l = Wbar_l ω_l + W_l ωbar_l (nonnegative weights with the lower state, nonpositive weights with the upper state), which gives Φ(x)-Φ_lower = x_upper - x ≥0. If the over/under bars are merely lost in typesetting, the text must be corrected explicitly; if the construction is as printed, Lemma 4 is false and the central safety-monitoring property in Theorem 2 collapses. This is the single most load-bearing concern because all subsequent analysis assumes the cooperative sign property.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an interval-observer-based safety monitor for nonlinear dynamical systems with embedded feedforward neural networks. Two auxiliary neural networks are constructed from a sign decomposition of the original network weights, and the activation functions are abstracted by global sector quadratic constraints. The observer gains are obtained by solving linear matrix inequalities that enforce positivity (Metzler structure) and practical stability of the error dynamics. The method is demonstrated on a lateral vehicle control example, where the simulated state trajectories lie between the estimated upper and lower bounds.","tokens_in":19812,"tokens_out":11969,"duration_ms":116470,"significance":"If the claims hold, the paper offers a tractable, optimization-based design for runtime interval estimation in learning-enabled control systems, which is relevant for safety monitoring. The paper's strengths include a mostly self-contained derivation after borrowing Lemma 4, LMI variables that are genuine optimization variables rather than fitted parameters, a standard quadratic-constraint formulation for activation functions, and a concrete vehicle simulation showing that the estimated bounds contain the trajectories. The contribution is incremental relative to earlier interval-observer work, but the combination of auxiliary-network construction with quadratic-constraint-based gain synthesis is useful and the simulation supports the proof of concept.","major_comments":[{"comment":"The pairing of the sign-split weights with the lower and upper states is reversed relative to what condition (12) requires. On the literal reading of (16), the negative part W^- multiplies the lower-layer state omega^-(l-1), built from x^-, and the positive part W^+ multiplies the upper-layer state omega^+(l-1), built from x^+. For a scalar one-layer network with W = -1 and no bias, this gives Phi^-(x^-, x^+) = -x^-, so Phi(x) - Phi^-(x^-, x^+) = x^- - x <= 0 for x^- <= x, contradicting (12). The lower auxiliary network should instead pair W^+ with x^- and W^- with x^+. Because Lemma 4 is the only support for (12)-(13), and (12)-(13) are necessary for e >= 0 and e^+ >= 0 in Proposition 1 and for the interval property in Theorem 2, this must be corrected or explicitly justified. If it is a typesetting defect, the equations and the statement of Lemma 4 need to be rewritten in unambiguous notation.","section":"Sec. 3, Eqs. (16)-(17) and Lemma 4"},{"comment":"The assertion that strict negativity of the quadratic form in (36) implies existence of epsilon > 0 satisfying (38) is not justified. Strict negativity on a constraint set does not automatically give a uniform margin, and this step is what converts the dissipative inequality into dV/dt <= -epsilon V + c2. The authors should either prove the margin by a compactness argument on normalized nonzero vectors in the feasible set, or exhibit epsilon directly from the LMI by retaining a small negative-definite remainder in (26) before eliminating variables.","section":"Sec. 3.2, Theorem 2 proof, Eqs. (36)-(40)"},{"comment":"Lemma 4 is imported from prior work without proof, but the notation in (15)-(17) does not allow the reader to check its hypotheses. In view of the apparent sign reversal in (16)-(17), a bare citation to Xiang (2021) is insufficient. The paper should either prove Lemma 4 in the present notation or provide an explicit dictionary between the variables used here and those in the cited paper.","section":"Sec. 3, Lemma 4 and Eq. (18)"}],"minor_comments":[{"comment":"The two split weight matrices are denoted by underbar and overbar symbols that are visually almost identical in the typeset text. Use W^- and W^+ or otherwise visually distinct symbols, and add a sentence in words describing which part multiplies which state bound.","section":"Sec. 3, Eq. (15)"},{"comment":"Table 1 is difficult to parse and the block entries appear fragmented in the displayed layout. A cleaner block-matrix definition or a structured listing of the blocks would let readers verify the dimensions and entries of N.","section":"Sec. 3.1, Table 1"},{"comment":"The blocks Ntilde_{Phi x}, Ntilde_{Phi omega}, Ntilde_{v x}, and Ntilde_{v omega} are defined implicitly through a 2x2 block array. Their exact dimensions relative to Q, S, and the state and weight vectors should be stated explicitly, since the LMI (26) cannot be verified otherwise.","section":"Sec. 3.2, Theorem 2"},{"comment":"Remark 6 asserts that e = e^+ = 0 occurs only at t = 0; if this fact is used in the proof, it should be justified, and otherwise the remark should be removed or reformulated.","section":"Sec. 3.2, Remark 6"},{"comment":"The dimensions of B_u and C in (41) should be checked against the problem formulation: both are 4x4, making y four-dimensional, while the text describes two physical outputs e1 and e2. The relation between C and the measured outputs should be clarified.","section":"Sec. 4, Eq. (41)"},{"comment":"Several references are incomplete in the bibliography, for example the Xiang (2021) entry lacks volume and page numbers; please complete all entries.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"For the editor: the central idea is sound and the simulation behavior is consistent with the expected interval-observer property, so I suspect the sign issue in (16)-(17) is a notation or typesetting error rather than a fundamental flaw. However, as submitted, the main theorem is not verifiable because the load-bearing Lemma 4 is stated with equations that appear to contradict it. A careful revision that fixes the auxiliary-network construction, provides or links a verifiable proof of Lemma 4, and tightens the epsilon step would make the paper acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea is worth engaging with, but the manuscript has a load-bearing flaw that needs to be fixed before it can be accepted.\n\nThe genuinely new piece is the LMI-based observer gain design using global sector quadratic constraints, which is a legitimate extension of the interval observer framework in Xiang (2021) and the QC approach in Yin et al. (2022). Theorem 2 is derived in real detail, the Schur complement manipulations check out, and the lateral vehicle simulation is a sensible illustration. That part is fine.\n\nThe problem is the auxiliary neural network construction. Equations (16) and (17) define the lower network with v_l = W_l * omega_l + W_bar * omega_bar, i.e., the nonpositive weights multiply the lower state and the nonnegative weights multiply the upper state. For a scalar weight W = -1, this gives Phi(x) - Phi_lower = x_lower - x <= 0, the opposite of condition (12). The correct mixed-monotone construction needs the nonpositive weights to multiply the upper state. As printed, Lemma 4 is false, and since Proposition 1 and Theorem 2 rest entirely on the sign conditions (12)-(13), the interval-observer guarantee collapses. This may well be a typesetting slip, but it is not a minor typo; it is the load-bearing step, and the paper cites Lemma 4 from Xiang (2021) without reproving it.\n\nA secondary issue: in the proof of Theorem 2, the step where a uniform epsilon > 0 is asserted to exist from the strict inequality (37) is handwavy. It is likely true because of continuity and compactness, but it should be argued, not asserted.\n\nThe paper also does not release code or data for the simulation, which makes it harder to confirm the LMI feasibility and the plotted bounds. That is a reproducibility gap, but a minor one.\n\nWho this is for: researchers working on runtime monitoring and verification of neural-network-enabled control systems. They will get a useful synthesis of QCs with interval observers, provided the sign issue is resolved. The paper deserves a serious referee, but the current version should not be accepted without an explicit correction of (16)-(17), a proof or a precise citation of the corrected Lemma 4, and a clean justification of the epsilon step.","headline":"A plausible interval-observer design for NN systems, but the auxiliary network construction in (16) appears to have the signs swapped, which breaks the central Lemma 4 and the main theorem as printed.","tokens_in":20353,"tokens_out":2261,"would_cite":false,"duration_ms":21153,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"For a nonlinear system driven by a neural network, an interval observer with gains fixed by two linear matrix inequalities and activation functions enclosed by quadratic sector constraints keeps certified upper and lower bounds on the…","keywords":["interval observer","safety monitoring","neural network control systems","quadratic constraints","linear matrix inequalities","positive systems","practical stability","lateral vehicle control"],"falsifier":"Take a small feedforward network with known weights and a monotone Lipschitz activation such as tanh, pick vectors $\\underline{x} \\le x \\le \\overline{x}$, and evaluate the original network (5) together with the two auxiliary networks (16)--(17). If any entry of $\\Phi(x) - \\underline{\\Phi}(\\underline{x},\\overline{x})$ or of $\\overline{\\Phi}(\\underline{x},\\overline{x}) - \\Phi(x)$ is negative, the imported sign lemma is false and the positivity half of Theorem 2 does not follow from the construction as stated.","tokens_in":19243,"feed_emoji":"🚗","tokens_out":14301,"duration_ms":123142,"temperature":0.7,"pith_summary":"Safety monitoring of a nonlinear dynamical system whose dynamics include a neural network component is reframed as an online state-estimation problem. The paper proposes a Luenberger-type interval observer that produces a lower and an upper estimate of the state, and it proves conditions under which the true state is always sandwiched between the two estimates. The route is to split each weight matrix of the neural network into positive and negative parts to form two auxiliary networks, and to abstract every activation function by a global-sector quadratic constraint. Observer gains are found by checking feasibility of one LMI and one matrix inequality; whenever they are feasible, the observer also has practically stable error, so the estimate interval does not diverge. A lateral vehicle control simulation shows the interval bounds tracking the states of a vehicle steered by a neural network controller.","feed_headline":"Two inequalities certify safe bounds for neural-network system states","feed_subtitle":"Two auxiliary networks and an LMI feasibility check give online state bounds with no offline reachability computation.","key_machinery":"The load-bearing construction is the interval observer (8) together with two auxiliary neural networks (16)--(17) built by splitting every weight matrix into a negative part and a positive part so that the auxiliary outputs bound the original network output from above and below. This sign property is what makes the error system cooperative, i.e., its dynamics matrix $A-LC$ is Metzler (all off-diagonal entries nonnegative), so nonnegative initial errors remain nonnegative for all time. The activation functions are then abstracted by global-sector quadratic constraints (Theorem 1), which replace each nonlinearity by an inequality involving the sector bounds $\\alpha$ and $\\beta$; this is what lets Lyapunov analysis be expressed as linear matrix inequalities. The final feasibility conditions (26)--(27) simultaneously enforce the Metzler property and the decrease of the Lyapunov function $V = \\tilde e^{T} Q \\tilde e$, yielding both invariance of the interval and practical stability of the error.","core_discovery":"The central result is Theorem 2. For a Lipschitz nonlinear system of the form (2) with a feedforward neural network $\\Phi$ and an auxiliary-network pair built by the weight-splitting rule (14)--(17), if the matrix inequalities (26) and (27) are feasible, then the error system (9) is positive and practically stable. Consequently, system (8) is an interval observer: the inequalities $\\underline{x}(t) \\le x(t) \\le \\overline{x}(t)$ hold for every $t \\ge 0$ starting from compatible initial bounds, and the estimation error stays bounded with a bound derived from the input uncertainty and the Assumption 3 parameters. The observer gains are recovered as $\\tilde L = Q^{-1}M$, where $Q$ is a diagonal positive definite matrix. The proof works by making the error dynamics cooperative (Metzler), replacing the activation functions with global-sector quadratic constraints, and applying a quadratic Lyapunov function $V(\\tilde e) = \\tilde e^{T} Q \\tilde e$ to convert the design into convex LMI feasibility.","pith_inferences":["Because the quadratic-constraint setup only needs sector bounds $\\alpha$ and $\\beta$, one could replace the global sector with state-dependent local sectors and likely tighten the interval width whenever a smaller operating region is known a priori.","The Metzler condition (positivity) and the Lyapunov inequality (stability) are separate constraints, so a designer could optimize the two observer gains independently, for instance to minimize the asymptotic interval width under the same feasibility conditions.","The online bounds could be used as triggers for supervisory control: approaching an estimated bound would activate a safety controller, turning the interval observer into a runtime safety layer for learning-enabled systems.","The same weight-splitting and sector-constraint machinery should extend to switched or hybrid plants provided the Metzler and sector conditions are re-established per mode, which the paper lists as future work."],"forward_implications":["A feasible solution to (26)--(27) certifies before deployment that the state will remain inside the observer interval for all $t \\geq 0$, so a safety violation is detectable the moment a trajectory touches or crosses a bound.","The method applies to any activation function that is monotone, Lipschitz, and sits in a known sector; ReLU, tanh, sigmoid, and leaky ReLU all satisfy these assumptions, so the design is not restricted to one nonlinearity.","Because gain synthesis is reduced to LMI feasibility, the design is a convex optimization problem solvable by standard numerical packages, avoiding the high computational cost of offline reachability analysis.","The estimation-error bound is computable from the data: the quantities of Assumption 3 and the input uncertainty intervals determine the practical-stability residual, giving a pre-deployment guarantee on the width of the interval.","In the lateral vehicle control example, the interval observer tracks both the lateral position error and the yaw angle error, keeping the simulated trajectories within their upper and lower envelopes over the tested horizon."],"supporting_citations":[{"why":"Provides Lemma 4, the auxiliary-network sign lemma on which the positivity of the error system and the interval property of the observer depend.","marker":"(Xiang, 2021)"},{"why":"Supplies the offset local-sector quadratic constraints that Theorem 1 extends to a global-sector form for the error system.","marker":"(Yin et al., 2022)"},{"why":"States the cooperative-systems lemma used to conclude that nonnegative initial errors stay nonnegative for all time.","marker":"(Efimov & Ra¨ıssi, 2016)"},{"why":"Provides Lemma 3, the Lyapunov condition that turns the derived error bound into practical uniform exponential stability.","marker":"(Ge & Wang, 2004)"},{"why":"Gives the estimation procedure for the Assumption 3 parameters that quantify the nonlinearity bound in the error dynamics.","marker":"(Zheng, Efimov, & Perruquetti, 2016)"},{"why":"Supplies Lemma 1, used in the proof to show that diagonal scaling preserves the Metzler property of the error dynamics.","marker":"(Wang, Li, & Xiang, 2022)"},{"why":"Supplies the lateral vehicle bicycle model, the steady-state-error analysis, and the parameter values used in the simulation.","marker":"(Rajamani, 2011)"},{"why":"Provides the baseline feedback gain whose operating data train the neural network controller in the numerical example.","marker":"(Alleyne, 1997)"}],"fun_headline_variants":["Interval observer with two neural nets certifies safe state bounds","Two auxiliary networks yield online interval bounds via LMI checks","Quadratic constraints make interval observers for neural systems","LMI feasibility guarantees safe interval bounds for neural states","Two-net interval observer certifies safe bounds via LMIs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is the imported lemma that for any state between the observer bounds, the two auxiliary networks built from the weight-splitting formulas (16)--(17) keep both error differences elementwise nonnegative; the paper relies on this lemma without reproving it, and the sign direction in the construction formulas is not reconciled with the lemma's conclusion, so if the lemma fails the error system need not stay positive and the invariant bounds $\\underline{x} \\le x \\le \\overline{x}$ are lost.","fun_headline_variants_meta":{"raw":{"variants":["Interval observer with two neural nets certifies safe state bounds","Two auxiliary networks yield online interval bounds via LMI checks","Quadratic constraints make interval observers for neural systems","LMI feasibility guarantees safe interval bounds for neural states","Two-net interval observer certifies safe bounds via LMIs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001724,"raw_usage":{"total_tokens":6789,"prompt_tokens":890,"completion_tokens":5899,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":506,"completion_tokens_details":{"reasoning_tokens":5821}},"tokens_in":506,"tokens_out":5899,"duration_ms":39123,"temperature":1.0,"reasoning_tokens":5821,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T19:50:34.659412+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a small feedforward network with known weights and a monotone Lipschitz activation such as tanh, pick vectors $\\underline{x} \\le x \\le \\overline{x}$, and evaluate the original network (5) together with the two auxiliary networks (16)--(17). If any entry of $\\Phi(x) - \\underline{\\Phi}(\\underline{x},\\overline{x})$ or of $\\overline{\\Phi}(\\underline{x},\\overline{x}) - \\Phi(x)$ is negative, the imported sign lemma is false and the positivity half of Theorem 2 does not follow from the construction as stated.","supporting_citations":[{"cited_title":", Efimov, D","cited_arxiv_id":null,"evidence_quote":"Gives the estimation procedure for the Assumption 3 parameters that quantify the nonlinearity bound in the error dynamics."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies Lemma 1, used in the proof to show that diagonal scaling preserves the Metzler property of the error dynamics."},{"cited_title":"APACrefauthors \\ 2011","cited_arxiv_id":null,"evidence_quote":"Supplies the lateral vehicle bicycle model, the steady-state-error analysis, and the parameter values used in the simulation."},{"cited_title":"APACrefauthors \\ 1997","cited_arxiv_id":null,"evidence_quote":"Provides the baseline feedback gain whose operating data train the neural network controller in the numerical example."}],"review_version":1}