{"id":"fd36b1bd-1583-470f-955b-9fe53b7bdfdc","arxiv_id":"2504.21832","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"The paper derives LMI-based sufficient conditions for a nonlinear dynamic controller to stabilize interval observers of bounded Jacobian discrete-time systems, claiming tighter bounds than static feedback in one simulation.","lead":"This paper designs a dynamic controller for uncertain nonlinear discrete-time systems by stabilizing an interval observer, a pair of upper and lower state bounds, instead of the unknown true state. The method adds a nonlinear feedback term and a separation principle so the observer and controller gains can be designed separately, with simulations suggesting tighter bounds than a static version.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lemma 1's separation proof is invalid: the comparison matrix is not block triangular and the comparison step requires a cooperativity property that is neither stated nor satisfied, so Theorem 1's ISS guarantee lacks a proved foundation.","rationale":"The reader correctly identified the comparison principle in Lemma 1 as the weakest assumption. My stress-test agrees and sharpens the concern: the manuscript's own displayed blocks contradict the block-triangularity claim, and the comparison inequality requires a monotonicity property that is absent. This is not a matter of disagreement with prevailing consensus; it is an internal gap in the central proof. The false block-triangularity claim in particular is easy to verify from equations (21)-(22), so the concern is concrete rather than merely speculative.\n\nI do not recommend changing the overall conditional verdict. The framework is plausible, the LMI machinery is standard once the comparison system is accepted, and the paper points to prior work for observer-gain synthesis. The defects are repairable in principle: the authors could add explicit cooperativity conditions on \\tilde A, restrict the design so that all negative off-diagonal couplings are eliminated, or prove ISS directly for the actual augmented system without the comparison detour. Until one of these is supplied, the separation principle and the H∞/ISS guarantees are not established. The simulation in Section V is supportive but cannot validate the general theorem, and it does not test the comparison inequality or the block-triangularity assertion.\n\nThus the appropriate status remains conditional: the paper's central claim should be accepted only if the missing comparison/monotonicity argument is supplied. My read does not move the reader's verdict.","tokens_in":12941,"tokens_out":13701,"duration_ms":141323,"concrete_test":"Take the numerical data in Section V (A, C, B, D, L, F_φ, F_ψ, α, ε, and the reported controller gains). Reconstruct \\tilde A from Lemma 1 and run two algorithmic checks: (a) attempt all block permutations of the five n-blocks and verify whether any permutation makes \\tilde A block triangular; the coupling between x_c and x̄, x̲ will show that no such permutation exists. (b) Test monotonicity of the comparison map z ↦ \\tilde A z + λ(z) by computing its Jacobian; cooperativity requires all off-diagonal entries to be nonnegative. In particular, evaluate \\tilde A_{21} = (A−LC)^⊖ + F_φ − |L|F_ψ for the reported L and F_ψ. Since a positive matrix |L|F_ψ is subtracted, \\tilde A_{21} will have negative entries, so the comparison system is not cooperative.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Lemma 1 is the load-bearing step: Theorem 1 and the SDP (23) stabilize the comparison system (22), and only through (22) do the framers (18) and errors (20) inherit ISS and the bound (24). The proof of Lemma 1 fails on two connected points.\n\n(1) The matrix \\tilde A displayed in Lemma 1 is asserted to be block lower triangular, but it is not. Row 3 is the x_c dynamics and contains nonzero blocks \\tilde A_{34}, \\tilde A_{35} multiplying x̄, x̲, while rows 4 and 5 couple back to x_c via \\tilde A_{43}, \\tilde A_{53} and to e,e via \\tilde A_{41}, \\tilde A_{42}, \\tilde A_{51}, \\tilde A_{52}. There is a dependency cycle x_c → x̄ → x_c, so no reordering of the state vector makes \\tilde A block triangular. The spectral argument cannot justify separate design.\n\n(2) Even if (22) were Schur stable, a componentwise inequality z_{k+1} ≤ \\tilde A z_k + λ(z_k) + Λη_k only bounds the true augmented system if the right-hand side is cooperative, i.e., monotone nondecreasing in z. No such property is proved, and it is false in general: \\tilde A_{21} = (A−LC)^⊖ + F_φ − |L|F_ψ subtracts the positive matrix |L|F_ψ, and \\tilde A_{45} = −(A−LC)^⊖ − BK_d is sign-indefinite. With negative comparison coefficients, boundedness of the comparison system does not imply boundedness of the true framer/error system.\n\nConsequently the separation principle, the ISS claim, and the H∞ bound (24) are not established by the manuscript. The gap is concrete and central, though possibly repairable by imposing explicit cooperativity/Metzler conditions or by proving ISS directly for the actual augmented system.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript considers discrete-time nonlinear systems with bounded Jacobians, nonlinear measurements, and bounded state/measurement noise. Rather than stabilizing the uncertain plant directly, it constructs an interval framer/observer (13) whose upper and lower trajectories bound the true state, and then designs a nonlinear dynamic output-feedback controller (16) for the framer system. The main claims are a separation principle (Lemma 1), an LMI-based synthesis procedure (Theorem 1) that guarantees input-to-state stability of the closed-loop comparison system and the H-infinity bound (24), and a numerical example showing tighter intervals than a static controller from the author's prior work.","tokens_in":13353,"tokens_out":3242,"duration_ms":36695,"significance":"If the main results were fully established, the paper would be a useful extension of the author's earlier static interval-observer control method to dynamic nonlinear control, with the advertised separation principle enabling separate design of observer and controller gains. The manuscript has clear strengths: it uses mixed-monotone decomposition tools in a nontrivial way, the problem formulation is explicit about noise and nonlinear observations, and the numerical example reports concrete matrices and achieved attenuation levels. However, the central proof of the separation principle is incomplete in a load-bearing way, and the current draft does not establish the ISS claims that the LMI synthesis is meant to guarantee. The contribution is therefore conditional on a repairable but nontrivial revision of Lemma 1 and the associated comparison argument.","major_comments":[{"comment":"The proof of Lemma 1 asserts that the matrix \\tilde A in (22) is block lower triangular and uses this to separate the design of L from the design of the controller gains. This assertion is incorrect: the displayed blocks include \\tilde A_{34} = K_b + K_x^\\nu F_\\varphi and \\tilde A_{35} = -(K_b + K_x^\\nu F_\\varphi) in row 3, while rows 4 and 5 contain \\tilde A_{43} = B C_c, \\tilde A_{53} = B C_c, as well as nonzero couplings to e and \\bar{e} such as \\tilde A_{41}, \\tilde A_{42}, \\tilde A_{51}, \\tilde A_{52}. These entries create a dependency cycle x_c -> \\bar{x},\\underline{x} -> x_c, so no reordering of the state vector makes \\tilde A block triangular. The spectral argument in the proof therefore does not support separate design of the observer gain and the dynamic controller gains as stated.","section":"Section IV-B, Lemma 1, Eq. (22)"},{"comment":"The proof transfers ISS of the comparison system (22) to ISS of the original closed-loop framer and error systems. Such a transfer requires the comparison dynamics to be cooperative, i.e. the right-hand side must be monotone nondecreasing in the comparison state z_k. No monotonicity property is proved, and it is in fact not satisfied by the displayed entries: \\tilde A_{21} = (A-LC)^\\ominus + F_\\varphi - |L|F_\\psi subtracts the nonnegative matrix |L|F_\\psi, and \\tilde A_{45} = -(A-LC)^\\ominus - B K_d is sign-indefinite in general. With negative comparison coefficients, boundedness (or ISS) of the comparison system does not imply boundedness (or ISS) of the actual framer/error system. The comparison principle used in the proof of Lemma 1 is therefore not valid without additional explicit cooperativity or monotonicity conditions.","section":"Section IV-B, Lemma 1, Eq. (21)-(22)"},{"comment":"The claim that F_\\varphi can 'without loss of generality' be assumed invertible by increasing its diagonal elements is not harmless. The Lipschitz constant used in Theorem 1 is \\gamma = \\|F_\\varphi\\|_\\infty, and the bound (5) is used with this F_\\varphi in the comparison inequalities (21). Increasing diagonal entries changes \\gamma and changes the comparison matrix \\tilde A through the terms involving F_\\varphi. The manuscript does not show that the modified F_\\varphi preserves the required decomposition bound with the same \\gamma, nor that the resulting comparison system remains a valid upper bound for the original framer/error dynamics. This point needs a proof or an explicit restatement of the assumptions under which the modification is valid.","section":"Section III, paragraph after Eq. (9)"},{"comment":"The proof of Theorem 1 is abbreviated: it invokes [18, Lemma 3] and then applies similarity transformations and changes of variables without showing that the transformed LMIs are equivalent to the original ones, especially with respect to the block-diagonal congruence transformations and the definition of \\Gamma. In particular, the relation diag_5(\\tilde K^*) = (Q^{-1} \\Theta^*)^\\top is stated without derivation, and the dimensions of the matrices in (23) and (26)-(27) are not checked against the stated 20n/\\hat{n} block sizes. Since the ISS guarantee depends entirely on this LMI step, the proof needs to be written out in full or the missing equivalence argument must be provided.","section":"Section IV-C, Theorem 1, proof of (26)-(27)"}],"minor_comments":[{"comment":"In the matrix A displayed in Section V, the entry shown as '01384' appears to be a typo and should likely read '0.1384'.","section":"Section V, Example data"},{"comment":"The signal \\tilde{w}_k in the H-infinity bound (24) is never defined; the reader must infer that it denotes the augmented noise vector \\eta_k from (22). Please define it explicitly.","section":"Equation (24)"},{"comment":"The proof states that 'by applying the last inequality in (21), it is straightforward to see that (20) admits a linear comparison system with state matrix \\tilde A_u'. The derivation is not shown, and given the coupling terms introduced in (18), this step is not immediate; it should be expanded or the relevant inequalities should be written out.","section":"Lemma 1 proof"},{"comment":"Reference [2] contains a typo ('Teansaction'), and several key results are cited as 'follows the lines of [8]' or 'accepted' ([15]); please update the reference list with complete publication data and identify precisely which parts of the proofs rely on these prior results.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is an incremental extension of the author's prior work [15] from static to dynamic interval-observer-based control, and it relies heavily on [8], [15], and [16] for decomposition bounds, the observer SDP, and the static comparator. This is not disqualifying if the extension is nontrivial, but the editor may wish to verify the publication status of [15] and [8] and consider whether the novelty is sufficient for the target journal. The main technical issue is the unproven comparison step in Lemma 1; the separation principle as stated is not established, and the LMI synthesis in Theorem 1 therefore currently lacks a sound stability guarantee. A revision that either proves the required cooperativity or replaces the comparison argument with a direct stability analysis could make the paper publishable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The dynamic controller with the two nonlinear compensation terms is a sensible generalization of your earlier static interval-observer controller, and the LMI synthesis in Theorem 1 is concrete enough to implement. But the separation principle in Lemma 1 is not proved: the proof steps from the actual augmented system to a componentwise comparison system (22) and then claims that ISS of the comparison system transfers back. That transfer requires the comparison dynamics to be cooperative (monotone) in the state. The matrix \\tilde A is not nonnegative—entries like \\tilde A_{21} contain -|L|F_\\psi and \\tilde A_{45} contains -BK_d—so a larger component of z can make the bound for another component smaller. Stability of \\tilde A therefore does not bound the true framer and error systems. The fix is to impose explicit Metzler/nonnegativity conditions on \\tilde A or to prove ISS directly for the actual augmented system. Until that is done, Theorem 1's ISS claim and the H-infinity bound (24) rest on an unproven comparison step.\n\nOne factual note: the stress-test claim that \\tilde A is not block lower triangular is wrong. In the z ordering used in the paper, the top-right 2x3 block is zero, so the matrix is block lower triangular, and the eigenvalue argument for separating the L-design from the control-gain design is valid for the linear part. The problem is not the triangularity; it is the missing monotonicity.\n\nOn the positive side, the problem setup is clean, the mixed-monotone decomposition machinery is appropriate, and extending the static controller to a dynamic one with nonlinear compensation is a natural and useful step. The H-infinity formulation is standard, and the simulation shows the dynamic controller yields tighter intervals than the static one, though a single example is not enough to support \"significantly outperforms\"—that needs quantitative Monte Carlo comparisons.\n\nThe proof of Theorem 1 is also abbreviated: the change of variables from (26)-(27) to (23) is only sketched, and the reader has to trust that the bilinearities are eliminated correctly. The invertibility fix for F_phi (adding epsilon_0 to the diagonal) is fine because increasing a nonnegative F preserves the bound (5).\n\nWho is this for? Control researchers working on interval observers, positive systems, and robust output feedback for discrete-time nonlinear systems. The idea is worth pursuing, but the central gap is load-bearing. I would send it to peer review and ask the author either to add explicit cooperativity conditions on \\tilde A or to prove ISS directly for the actual augmented system. If the gap is repaired, this becomes a solid contribution; as written, it is conditionally acceptable at best.","headline":"A sensible dynamic extension of the author's static interval-observer controller, but Lemma 1's comparison step needs a cooperativity condition that is never imposed, so the main stability and H∞ guarantees are not established.","tokens_in":13855,"tokens_out":9514,"would_cite":false,"duration_ms":81488,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93C55","93B53","93D25","93C10"],"pacs":[],"model":"deepseek-v4-flash","headline":"Nonlinear dynamic feedback control stabilizes uncertain discrete-time systems by stabilizing an interval observer around the state, with a separation principle for gain design and LMI conditions giving ISS and an H-infinity bound.","keywords":["interval observers","bounded Jacobian nonlinear systems","dynamic output feedback","separation principle","H-infinity control","linear matrix inequalities","mixed-monotone decomposition","discrete-time nonlinear systems"],"falsifier":"Fix a system satisfying Assumption 1, choose $L$ from the SDP in Proposition 3, solve SDP (23), and simulate the full closed loop (8), (13), (16) from the vertices of the initial interval with noise sequences $w_k,v_k$ held at their extreme values. If the true state ever leaves $[\\underline{x}_k,\\overline{x}_k]$, if the framer error $\\varepsilon_k$ becomes negative, or if $\\lVert z_k\\rVert_2^2 > \\mu^* \\lVert \\tilde{w}_k\\rVert_2^2$ for any $k$, the claimed ISS property and bound (24) are false.","tokens_in":12717,"feed_emoji":"🎛️","tokens_out":9627,"duration_ms":97672,"temperature":0.7,"pith_summary":"The paper asks whether a discrete-time nonlinear system with bounded Jacobians, nonlinear measurements, and unknown-but-bounded process and measurement noise can be stabilized without ever measuring its state. The proposed answer is to build an interval observer—upper and lower trajectories guaranteed to contain the true state—and then to design a nonlinear dynamic output-feedback controller that drives the interval tight around the state. The paper claims a separation principle, in which the observer gain is designed first and the dynamic controller gains second, and an LMI synthesis guaranteeing that the closed-loop comparison system is input-to-state stable with an $H_\\infty$ disturbance-attenuation bound. If these claims are right, interval-observer control extends to a broad nonlinear discrete-time class with noisy nonlinear outputs, and the added controller terms yield tighter state enclosures than the static feedback design, as the paper's unstable five-state example shows.","feed_headline":"Dynamic control beats static feedback on interval observers","feed_subtitle":"A separation principle plus LMIs keeps unstable nonlinear systems bounded with tighter state enclosures.","key_machinery":"The central object is the interval framer (13), a pair of coupled recursions for upper and lower state bounds built from tight mixed-monotone decomposition functions $\\varphi_d$ and $\\psi_d$; the same decompositions give matrices $F_\\varphi$ and $F_\\psi$ that bound the nonlinear increments via inequality (5). The framer error system (14) is a positive system once $L$ stabilizes $|A-LC|$. The argument then moves to the comparison augmented system (22), whose matrix $\\tilde{A}$ is block lower triangular; this block structure is what lets observer and controller gains be designed separately. The SDP (23) is obtained by applying congruence transformations and a change of variables to known Lipschitz-nonlinear ISS conditions, eliminating bilinear products of the decision variables with $B$.","core_discovery":"On the paper's own terms, the discovery is that the interval-framer idea can be turned into a dynamic control design for bounded-Jacobian discrete-time systems without measuring the state. Writing $f(x)=Ax+\\varphi(x)$ and $g(x)=Cx+\\psi(x)$ with $\\varphi,\\psi$ Jacobian-sign-stable, the framer equations (13) produce bounds $\\underline{x}_k \\le x_k \\le \\overline{x}_k$, and the framer error evolves as $\\varepsilon_{k+1}=|A-LC|\\varepsilon_k + \\delta^\\varphi_k + |L|\\delta^\\psi_k + |LV|\\delta_v + |W|\\delta_w$. Lemma 1 claims a separation principle: the augmented system of framers, tracking errors, and controller states has a comparison system (22) whose state matrix is block lower triangular, so a stabilizing observer gain $L$ for the framer error system plus subsequently chosen dynamic gains stabilize the whole closed loop. Theorem 1 converts this into an SDP (23), with gains recovered from (25), and states that the closed-loop system is ISS and obeys $\\lVert z_k\\rVert_2^2 \\le \\mu^* \\lVert \\tilde{w}_k\\rVert_2^2$. The paper notes in Remark 2 that the separation is not 'full,' since $L$ remains an input to the controller synthesis.","pith_inferences":["Editorial extension: the comparison step in Lemma 1 would be fully justified if $\\tilde{A}$'s off-diagonal entries and $\\lambda(z)$ were component-wise nonnegative; checking this property on a given example is a direct test, and when it holds, simpler linear comparison tools could replace the Lipschitz-based SDP.","Editorial extension: since Lemma 1's separation is one-way, jointly optimizing $L$ and the controller gains is a natural next step that the paper itself flags in Remark 2.","Editorial extension: the same set-valued state estimate could drive event-triggered or model-predictive controllers; the conclusion names switched and hybrid systems and MPC as future work."],"forward_implications":["The true state remains inside the computed interval at every time step, so the same controller can certify state bounds while regulating the system.","Observer and controller gains are synthesized in two separate convex steps, which keeps the design tractable as the state dimension grows.","The synthesis produces a numerical $H_\\infty$ attenuation level $\\mu^*$, allowing a designer to trade interval tightness against worst-case disturbance amplification.","Nonlinear measurement functions and both state and measurement noise are handled, going beyond earlier interval-observer control formulations restricted to linear, noise-free outputs.","The example indicates that the dynamic controller's extra gains yield substantially tighter closed-loop intervals than the static feedback design, not just stability."],"supporting_citations":[{"why":"Supplies the observer-gain SDP and framer structure that Proposition 3 uses to make (13) an interval observer.","marker":"[8]"},{"why":"The static feedback design that the dynamic controller generalizes and outperforms in the simulation comparison.","marker":"[15]"},{"why":"Yields Propositions 1 and 2, namely the mixed-monotone decomposition and the tight $F_\\varphi,F_\\psi$ bounds used throughout the framer and error dynamics.","marker":"[16]"},{"why":"Provides the Lipschitz-nonlinear ISS/LMI conditions that the SDP (23) adapts through similarity transformations and variable changes.","marker":"[18]"},{"why":"Justifies perturbing $F_\\varphi$ to a diagonally dominant invertible matrix so the invertibility assumption can be made without losing the bound (5).","marker":"[17]"}],"fun_headline_variants":["Dynamic control tightens interval bounds over static feedback","Interval observers get dynamic control via separation principle","Optimal dynamic control for uncertain discrete-time systems","Dynamic gains outperform static in interval observer control","Separation principle enables dynamic interval observer control"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the component-wise bounding comparison system (22) can stand in for the real closed-loop dynamics—stability of that bounding system is taken to imply stability of the true interval and error systems—but the transfer is assumed rather than proved, and the comparison matrix $\\tilde{A}$ contains sign-indefinite entries.","fun_headline_variants_meta":{"raw":{"variants":["Dynamic control tightens interval bounds over static feedback","Interval observers get dynamic control via separation principle","Optimal dynamic control for uncertain discrete-time systems","Dynamic gains outperform static in interval observer control","Separation principle enables dynamic interval observer control"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00049,"raw_usage":{"total_tokens":2409,"prompt_tokens":940,"completion_tokens":1469,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":556,"completion_tokens_details":{"reasoning_tokens":1401}},"tokens_in":556,"tokens_out":1469,"duration_ms":9573,"temperature":1.0,"reasoning_tokens":1401,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:53:19.296937+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fix a system satisfying Assumption 1, choose $L$ from the SDP in Proposition 3, solve SDP (23), and simulate the full closed loop (8), (13), (16) from the vertices of the initial interval with noise sequences $w_k,v_k$ held at their extreme values. If the true state ever leaves $[\\underline{x}_k,\\overline{x}_k]$, if the framer error $\\varepsilon_k$ becomes negative, or if $\\lVert z_k\\rVert_2^2 > \\mu^* \\lVert \\tilde{w}_k\\rVert_2^2$ for any $k$, the claimed ISS property and bound (24) are false.","supporting_citations":[{"cited_title":"Khajenejad and S.Z","cited_arxiv_id":null,"evidence_quote":"Supplies the observer-gain SDP and framer structure that Proposition 3 uses to make (13) an interval observer."},{"cited_title":"Optimal Feedback Stabilizing Control of Bounded Jacobian Discrete-Time Systems via Interval Observers","cited_arxiv_id":"2410.09222","evidence_quote":"The static feedback design that the dynamic controller generalizes and outperforms in the simulation comparison."},{"cited_title":"Khajenejad, F","cited_arxiv_id":null,"evidence_quote":"Yields Propositions 1 and 2, namely the mixed-monotone decomposition and the tight $F_\\varphi,F_\\psi$ bounds used throughout the framer and error dynamics."},{"cited_title":"Abbaszadeh and Horacio J","cited_arxiv_id":null,"evidence_quote":"Provides the Lipschitz-nonlinear ISS/LMI conditions that the SDP (23) adapts through similarity transformations and variable changes."},{"cited_title":"Sootla, Y","cited_arxiv_id":null,"evidence_quote":"Justifies perturbing $F_\\varphi$ to a diagonally dominant invertible matrix so the invertibility assumption can be made without losing the bound (5)."}],"review_version":1}