{"id":"6c8d5062-e68f-470c-9bec-245c44d83970","arxiv_id":"2411.09237","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"An unsupervised PINN is used to learn nonlinear observer gains by enforcing the contraction matrix inequality, with an exponential ISS bound claimed for the resulting observer.","lead":"The authors propose training a physics-informed neural network to compute the correction gain of a contraction-based nonlinear observer, with a loss that penalizes violations of a matrix inequality. They report exponential convergence and noise robustness on two oscillator examples, but the loss formula contains a sign error that undermines the method as stated.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"For y=x1 and any C1 gain with k(x,h(x))=0, the (2,2) diagonal of the MPDI is forced to Jf22+2λ, which is positive in both numerical examples; hence no P=I gain satisfying (7) exists and the learned observer cannot inherit Theorem 1/Theorem 3 guarantees.","rationale":"The reader's rejection is correct, but the weakest assumption is not just that the trained network is ε-close to an ideal gain; for the chosen examples the ideal gain itself cannot exist under the paper's P=I restriction. This is a stronger, internal inconsistency: the exact boundary condition k(x,h(x))=0 plus C1 regularity forces the unmeasured-direction diagonal entry of the MPDI to be positive. Therefore the central mechanism — training a loss that enforces (7) to obtain an exponentially convergent observer with P=I — is impossible for the very systems used for validation. I agree with the reader's verdict (REJECT), with partial agreement on the stated weakest assumption; the unverified ε-closeness is a real concern, but the underlying premise of an ideal P=I contraction gain fails a priori for the examples. The leading-principal-minor criterion in Theorem 2 is also not sufficient for negative-semidefiniteness, but the non-existence of a P=I gain is the more decisive problem.","tokens_in":9304,"tokens_out":21867,"duration_ms":248927,"concrete_test":"Analytical check: for y=x1, impose k(x1,x2,x1)=0 for all x1,x2 and differentiate with respect to x2 to obtain ∂k2/∂x̂2=0 on the diagonal. Substituting into (7) gives D22=∂f2/∂x2+2λ; this equals 2λ for system (20) and 1+2λ at x1=0 for system (19), contradicting D≤0. To make it empirical, use automatic differentiation on the trained network to evaluate the eigenvalues of D on a fine grid restricted to the diagonal y=h(x̂) (e.g. 10^4 points over X). If the maximum eigenvalue is positive — which the analytic argument shows is unavoidable for any C1 gain satisfying the boundary condition — the learned gain does not satisfy the MPDI (7), and the unsupervised loss is not actually enforcing the contraction condition on the zero-output-error manifold.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV.B trains the network against (7) with P=I, and Theorem 3 starts from an ideal k satisfying (4) with P=I and λ>0. For the two examples, h(x)=x1. The boundary condition k(x,h(x))=0 in (4) implies k(x1,x2,x1)=0 for all x1,x2. Since k is C1, differentiating in x2 gives ∂k2/∂x̂2=0 at every diagonal point (x̂,h(x̂)). Consequently the (2,2) entry of D=He(∂f/∂x̂+∂k/∂x̂)+2λI is D22=∂f2/∂x2+2λ. For the reverse Duffing (20), ∂f2/∂x2=0, so D22=2λ>0. For the Van der Pol (19), ∂f2/∂x2=1−x1², so D22=1+2λ>0 at x1=0. A symmetric matrix with a positive diagonal entry cannot be negative semidefinite, so (7) has no C1 solution with P=I in either example. Hence Assumption 3(A2) — existence of an ideal k with contraction rate λ — is vacuous for the systems used in the experiments, and the loss (6)-(8), even if perfectly minimized, cannot certify the global exponential convergence claimed in Theorem 1/Theorem 3. The boundary loss (9) does not fix this: it penalizes the value of k at the diagonal, not the diagonal derivative that makes D22 positive.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an unsupervised physics-informed neural network (PINN) approach for designing the correction term of a nonlinear observer for autonomous systems. The observer gain is learned by minimizing a loss that penalizes violation of the contraction matrix partial differential inequality (MPDI) from Theorem 1, with the metric fixed to the identity, plus a boundary-condition loss. The authors also state a robustness result (Theorem 3) bounding the estimation error under measurement noise and neural-network approximation error, and they report simulations on a Van der Pol oscillator and a reverse Duffing oscillator. The central claim is that the learned gain enforces the contraction conditions and therefore yields a globally exponentially convergent observer whose robustness is quantified by Theorem 3.","tokens_in":9678,"tokens_out":11133,"duration_ms":114237,"significance":"The problem is relevant: numerically solving the contraction-based observer MPDI is genuinely challenging, and a learning-based unsupervised formulation is a reasonable idea to explore. I credit the authors for avoiding dependence on labeled trajectory data, for using automatic differentiation to compute the required Jacobians, and for attempting an explicit robustness analysis with noise and approximation error. However, the formal core of the paper is not sound as written. The loss function does not correctly encode negative semidefiniteness, and, more seriously, for the two numerical examples there exists no C1 gain satisfying the MPDI with P=I, so the assumptions underlying Theorem 3 are vacuous in the very systems used for validation. These are load-bearing issues that invalidate the main claims. The paper would need substantial reworking—correcting the loss, choosing a feasible metric and/or examples, and repairing Theorem 3—before its contribution could be credited.","major_comments":[{"comment":"The loss for even principal minors has the wrong sign. Since (7) requires D to be negative semidefinite, every even principal minor must be nonnegative, so the penalty should grow when Δ_i is negative. Writing l_i = min(0, Δ_i) for even i makes the loss decrease as Δ_i becomes more negative, thereby actively encouraging violation of the contraction condition. Independently, Theorem 2 as stated is false: for symmetric matrices, alternation of the leading principal minors is necessary but not sufficient for negative semidefiniteness; for example, diag(0,1) has leading minors Δ1=0 and Δ2=0, satisfying (−1)^1Δ1≥0 and (−1)^2Δ2≥0, yet the matrix is not negative semidefinite. A correct criterion requires all principal minors, not only the leading ones. Thus minimizing (6) does not enforce (7).","section":"Section IV.B, Eq. (8)"},{"comment":"With P=I and h(x)=x1, no C1 gain satisfying (7) exists for either test system. The boundary condition k(x,h(x))=0 in (4), together with C1 regularity, forces ∂k₂/∂x̂₂=0 at every point of the form (x̂₁,x̂₂,x̂₁). For the reverse Duffing oscillator (20), ∂f₂/∂x₂=0, so the (2,2) entry of D in (7) is 2λ>0. For the Van der Pol oscillator (19), at x₁=0 that entry is 1+2λ>0. A symmetric matrix with a strictly positive diagonal entry cannot be negative semidefinite, and the condition (7) is required to hold for all (x̂,y), including y=h(x̂). Therefore Assumption 3(A2) is vacuous in the examples, and the learned observer cannot inherit the guarantees of Theorem 1 or Theorem 3. The boundary loss (9) does not repair the obstruction, because it penalizes the value of k at the diagonal, not its derivative along x̂₂.","section":"Sections IV.B and VI, Eqs. (7), (19), (20)"},{"comment":"The bound in (14) does not follow from the proof. From (17) and the claimed inequality (18), the differential inequality is ˙V ≤ −(λ−2)V + (1/2)∥ε∥² + (L²/2)∥v∥². The solution of this inequality yields an estimation bound with decay rate (λ−2)/2 and asymptotic gain proportional to (∥ε̄∥² + L²∥v̄∥²)^{1/2}/(λ−2)^{1/2}, not the stated e^{−ηt} factor and the term (ε̄+Lv̄)/(2√η). The step from (17) to (18) is also insufficiently justified: the contraction condition gives, at best, e^T[f(x)−f(x̂)−k(x̂,y)] ≤ −λ∥e∥², and the additional ∥e∥² term from Young's inequality changes the effective rate; the citation to [31, Theorem 4.3, p. 231] is too terse. Since Theorem 3 is presented as the formal robustness guarantee, this gap is significant.","section":"Section V, Theorem 3 and Eq. (14)"}],"minor_comments":[{"comment":"The manuscript inconsistently uses both “MPDI” and “MDPI” for the matrix partial differential inequality; please standardize.","section":"Throughout"},{"comment":"“Principle minors” should be “principal minors” in the text describing Eq. (8).","section":"Section IV.B"},{"comment":"The training algorithm is said to be “provided upon request,” but for a methods paper the pseudocode should be fully self-contained, including batch size and the exact stopping criteria, or the code should be released.","section":"Section IV.C"},{"comment":"Figure 3 is captioned “Inverse Duffing” while the text and Eq. (20) call the system the “reverse Duffing” oscillator; the nomenclature should be consistent.","section":"Section VI"},{"comment":"The dataset region is written as “(X , Y) = ([ −1, 1]², [−1, 1])”; this notation is ambiguous and should be clarified, since X is a subset of R² and Y is a subset of R.","section":"Section VI"},{"comment":"The “identity” from [31] that converts the path term into the contraction-rate term should be stated explicitly, including its assumptions, because the proof depends on it.","section":"Section V, proof of Theorem 3"}],"recommendation":"reject","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Thanks for the report. I agree with the reader's verdict, and the second-pass stress test lands: the central design problem is infeasible for both examples. The paper's core idea—train an unsupervised PINN so the observer gain satisfies the contraction MPDI—is genuinely new, and the authors know the relevant literature. The two-condition setup (contraction inequality plus boundary k(x,h(x))=0) is the right thing to try. The robustness analysis is mostly standard ISS with a Lipschitz gain, so it is not a selling point, but at least it is clearly framed.\n\nThe problems are load-bearing. First, the principal-minor loss has a sign error: for even i, l_i = min(0, Δ_i) rewards negative minors, exactly the wrong direction for negative semidefiniteness. That alone would make training fail to enforce (7) even when a solution exists. Second, no such solution exists for the examples with P=I. Since h(x)=x1, the boundary condition k(x1,x2,x1)=0 for all x1,x2, and C1 regularity forces ∂k_2/∂x_2=0 on the diagonal. For the reverse Duffing, the (2,2) entry of D is then 0+2λ>0; for Van der Pol at x1=0 it is 1+2λ>0. A symmetric negative semidefinite matrix cannot have a positive diagonal entry. So (7) is infeasible in the training region itself, and Theorem 3's assumption of an ideal gain with P=I is vacuous for both experiments. The boundary loss does not repair this.\n\nThere is also a smaller factor inconsistency in Theorem 3's proof (the contraction term and the extra ||e||² seem to give rate λ-1, not λ-2), but that is secondary. Reproducibility is weak: no code, no data, \"training algorithm provided upon request,\" no baselines, and no verification that the trained network is close to an ideal gain.\n\nWhat is good: the paper is honest about the difficulty of solving the MPDI directly, and the idea of using principal minors in a PINN loss is worth exploring. With a corrected even-minor penalty and a flexible P—or different examples—it could become a real method.\n\nWho should read this: people working on learning-based contraction observers. As a paper, this version is not acceptable. My recommendation: let a serious referee see it—the flaw is specific and fixable, and a good report could save the useful part. But the current version should not be published without major revision.","headline":"The PINN idea is new, but the training target is infeasible for the paper's own examples, and the loss function's sign error makes the central claim unsupported.","tokens_in":10187,"tokens_out":4851,"would_cite":false,"duration_ms":55616,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93B53","93C10","68T07","93D30"],"pacs":[],"model":"deepseek-v4-flash","headline":"An unsupervised physics-informed neural network can learn contraction-based observer gains that guarantee global exponential convergence of the estimation error for autonomous nonlinear systems.","keywords":["physics-informed neural networks","nonlinear observers","contraction analysis","matrix partial differential inequality","unsupervised learning","exponential stability","input-to-state stability","autonomous systems"],"falsifier":"Take the trained gain and evaluate the largest eigenvalue of $D(\\hat{x},y) = \\operatorname{He}\\{\\frac{\\partial f}{\\partial \\hat{x}} + \\frac{\\partial \\hat{k}_\\theta}{\\partial \\hat{x}}\\} + 2\\lambda I$ on a dense grid over the domain $\\mathcal{X}\\times\\mathcal{Y}$; if any eigenvalue is positive at a grid point, the MPDI (7) is violated and the claimed exponential convergence guarantee is not applicable to that gain.","tokens_in":9111,"feed_emoji":"🧠","tokens_out":5034,"duration_ms":49083,"temperature":0.7,"pith_summary":"Contraction-based observer design has long been held back by the need to solve a matrix partial differential inequality (MPDI) for the observer's correction gain. This paper replaces that analytic step with an unsupervised physics-informed neural network (PINN) that learns the gain by penalizing violations of the inequality in its loss function. The authors claim the learned gain enforces the contraction condition enough to guarantee global exponential convergence of the estimation error, and they prove an exponential input-to-state-stability bound under bounded measurement noise and approximation error. If true, this gives a practical, data-free route to nonlinear observers with strong guarantees.","feed_headline":"Neural net learns observer gains with exponential convergence","feed_subtitle":"Trained on physics alone, a PINN solves the matrix inequality that blocked contraction-based observer design.","key_machinery":"The central object is the symmetric matrix $D(\\hat{x},y)$ from the contraction condition (7), whose negative semi-definiteness is enforced via Sylvester's criterion using its leading principal minors. The loss function $L = \\mu_1 L_{\\text{MPDI}} + \\mu_2 L_{\\text{BC}}$ penalizes the signed principal minors (positive for odd indices, negative for even indices) and the boundary condition $\\hat{k}_\\theta(\\hat{x}, h(\\hat{x}))=0$, with collocation points sampled uniformly over the domain of interest. Automatic differentiation provides the Jacobian $\\partial \\hat{k}_\\theta / \\partial \\hat{x}$ needed to form $D$.","core_discovery":"The paper's central claim is that the loss function (6)-(8), built from the leading principal minors of the matrix $D(\\hat{x},y)=\\operatorname{He}\\{\\frac{\\partial f}{\\partial \\hat{x}} + \\frac{\\partial \\hat{k}_\\theta}{\\partial \\hat{x}}\\} + 2\\lambda I$, enforces the contraction MPDI (7) on the training domain, so the learned gain $\\hat{k}_\\theta$ yields a globally exponentially convergent observer for the autonomous nonlinear system (1). Theorem 3 further states that under a Lipschitz gain and bounded noise and approximation error, the estimation error satisfies the exponential input-to-state bound (14) whenever the contraction rate $\\lambda > 2$. The claim is supported by simulation on the Van der Pol and reverse Duffing oscillators, where increasing $\\lambda$ reduces the mean squared estimation error under 15% measurement noise.","pith_inferences":["A natural extension is to verify post-training whether the learned gain actually satisfies the MPDI on a dense grid; if not, the theoretical guarantee is void even if simulations look good. This check is not performed in the paper.","The method could be extended to non-autonomous systems by making the gain depend explicitly on time, though the paper leaves this as future work.","The choice $P=I$ simplifies the MPDI, but allowing a non-constant metric $P(x)$ could widen the class of acceptable systems; the same PINN framework might be adapted to learn both $P$ and $k$ together.","The ISS bound depends on the Lipschitz constant $L$ of the gain and the bounds of noise and error; a post-training computation of these constants could turn the qualitative guarantee into a computable error certificate."],"forward_implications":["The approach removes the need to solve the MPDI analytically; any smooth autonomous system with a known Jacobian can be plugged into the training loop directly.","The learned observer is robust to bounded measurement noise and network approximation error, with the steady-state error bound shrinking as the contraction rate $\\lambda$ increases.","Because the loss is unsupervised (no state trajectories required), the method needs only collocation points, not data, to design an observer.","The exponential input-to-state bound (14) gives a quantitative guarantee that the estimation error remains within a neighborhood determined by noise level and approximation error.","Higher $\\lambda$ improves noise rejection but makes it harder to satisfy the MPDI, revealing a design trade-off in choosing the contraction rate."],"supporting_citations":[{"why":"Supplies Theorem 1, the contraction-based observer design condition that the loss function enforces.","marker":"[31]"},{"why":"Introduces contraction analysis, the differential stability framework underlying the whole construction.","marker":"[11]"},{"why":"Provides Sylvester's criterion for negative semi-definiteness, which the loss uses via leading principal minors.","marker":"[32]"},{"why":"Establishes the physics-informed neural network methodology that the unsupervised training is built on.","marker":"[21]"},{"why":"Documents the difficulty of solving the MPDI, the long-standing obstacle this paper targets.","marker":"[20]"}],"fun_headline_variants":["PINN solves contraction inequality for observer design","Neural observer via contraction guarantees exponential error decay","Unsupervised PINN designs nonlinear observers without labeled data","Contraction-based observer gains learned by physics-informed network","Physics-informed network solves matrix inequality for exponential observers"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole guarantee rests on the unverified Assumption 3(A2) that the trained network's gain is uniformly close to some ideal gain that satisfies the strict contraction inequality with $P=I$ and $\\lambda>2$; if no such ideal gain exists or the training is not $\\varepsilon$-close to it, the exponential bounds in Theorem 3 do not follow.","fun_headline_variants_meta":{"raw":{"variants":["PINN solves contraction inequality for observer design","Neural observer via contraction guarantees exponential error decay","Unsupervised PINN designs nonlinear observers without labeled data","Contraction-based observer gains learned by physics-informed network","Physics-informed network solves matrix inequality for exponential observers"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00033,"raw_usage":{"total_tokens":1787,"prompt_tokens":843,"completion_tokens":944,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":459,"completion_tokens_details":{"reasoning_tokens":871}},"tokens_in":459,"tokens_out":944,"duration_ms":9008,"temperature":1.0,"reasoning_tokens":871,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T20:53:38.491187+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the trained gain and evaluate the largest eigenvalue of $D(\\hat{x},y) = \\operatorname{He}\\{\\frac{\\partial f}{\\partial \\hat{x}} + \\frac{\\partial \\hat{k}_\\theta}{\\partial \\hat{x}}\\} + 2\\lambda I$ on a dense grid over the domain $\\mathcal{X}\\times\\mathcal{Y}$; if any eigenvalue is positive at a grid point, the MPDI (7) is violated and the claimed exponential convergence guarantee is not applicable to that gain.","supporting_citations":[{"cited_title":"Observer design for continuous-time dynamical systems","cited_arxiv_id":null,"evidence_quote":"Supplies Theorem 1, the contraction-based observer design condition that the loss function enforces."},{"cited_title":"Lohmiller and Slotine","cited_arxiv_id":null,"evidence_quote":"Introduces contraction analysis, the differential stability framework underlying the whole construction."},{"cited_title":"Positive definite matrices and sylvester’s criterion","cited_arxiv_id":null,"evidence_quote":"Provides Sylvester's criterion for negative semi-definiteness, which the loss uses via leading principal minors."},{"cited_title":"Review on contraction analysis and computation of contraction metrics","cited_arxiv_id":null,"evidence_quote":"Documents the difficulty of solving the MPDI, the long-standing obstacle this paper targets."}],"review_version":1}