{"id":"7382192d-c607-4cf2-b5a4-ad002a9726d3","arxiv_id":"2505.06525","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"APSMC adaptively updates time-varying state-space matrices through Kalman-filtered states and physics-constrained proximal gradient steps, claiming an optimality that the paper does not actually prove.","lead":"An adaptive state-space identification framework combines a Kalman filter with physics-constrained proximal gradient updates to estimate time-varying dynamics of nonlinear structures from noisy measurements. The authors claim convergence to an optimal model and report strong predictive accuracy on simulations and a scaled bridge experiment.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section 3.1's convergence guarantee rests on an unanalyzed coupling between the Kalman-filter state sequence and the online proximal-gradient update; the paper itself later concedes non-convergence for poor initial models and non-convexity of the constrained problem.","rationale":"The paper's central claim is a theoretical convergence-to-optimality guarantee, and the reader's weakest assumption—the unanalyzed joint convergence of the Kalman-filter state sequence and the proximal-gradient parameter update—is indeed the load-bearing concern. My reading of Section 3.1, Section 5.3, and Appendix B.2 confirms this is the soft spot. The claim would require either a contraction or descent proof for the coupled iteration, or a demonstration that the Kalman-filter state sequence is asymptotically independent of the evolving A (e.g., through a separation principle), and neither appears. The paper's own Section 5.3 admits failure cases with poor A0, and Appendix B.2 concedes non-convexity under the bilinear transformation, so the issue is internal to the manuscript rather than merely outside current consensus. The empirical results—notably the 0.398% NMSE and the 19-sequence generalization from a single 10-second update—are genuine supporting evidence that the recipe can work in practice, but they do not establish the theoretical guarantee asserted in the abstract and conclusions. Additional concerns such as per-sequence hand-tuned learning rates in Table 2 and unreported hyperparameters further weaken the generalization claim, but the theoretical gap alone is sufficient to support the reader's REJECT verdict. No change to the verdict is needed.","tokens_in":48771,"tokens_out":3356,"duration_ms":36543,"concrete_test":"Run a synthetic linear time-invariant test that isolates the coupling: choose n=2 or 4, known A_true, C, Q, R, Gaussian white noise, and no physical constraints. Execute Algorithm 2 from several initializations A0 (including one with spectral radius far from A_true) with constant learning rates in the range used in Table 2. Track ||A_k − A_true|| and ||A_k − A_k_batch||, where A_k_batch is the least-squares fit to the Kalman-filter state sequence at iteration k. If A_k does not converge to A_true for any fixed learning rate, or if the limit depends on A0, the claimed lim ΔA_i = 0 fails. Also check innovation whiteness at the converged fixed point; non-white innovations show the Kalman filter is not the 'true' filter for that A.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.1's guarantee (lim_{i→∞} ΔA_i = 0, following Eq. 29) treats Eq. (24)/(26) as a fixed convex objective solved by proximal gradient. It is not fixed: the regressors x̂_k and targets x̂_{k+1} are outputs of a Kalman filter whose gain and innovation covariance are functions of the A being estimated. The iteration is therefore a coupled stochastic approximation with state-dependent noise, and no Lyapunov function or contraction argument is provided. Convexity of each instantaneous loss does not imply convergence of the alternating filter/optimizer loop, and it does not imply convergence to the data-generating A_true: Eq. (26) defines optimality relative to the estimated state sequence, not relative to the true system matrix. The paper itself concedes the load-bearing assumption fails in Section 5.3 ('when the initial model A0 is significantly inaccurate ... the KF fails to provide accurate predictions, preventing the APSMC algorithm from updating to the correct model' and 'it remains theoretically unclear whether initializing A0 as an identity matrix guarantees convergence to the optimal solution'), and Appendix B.2 states that the bilinear-transform constraints make the optimization problem non-convex, so proximal gradient does not guarantee a global optimum. Since the headline claim is exactly this global-optimality and convergence guarantee, the unproved coupling is the critical weakness.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes the Adaptive Physics-Informed System Modeling with Control (APSMC) framework for online identification of time-varying linear state-space models from noisy input–output data. The method alternates between Kalman filtering (or another Bayesian filter) for state estimation and online proximal-gradient updates of the system matrices A and B, subject to physics-based structural constraints. The paper claims a theoretical guarantee, derived from stochastic subspace identification, that as data accumulate the estimates converge to the theoretically optimal solution (Section 3.1), and it validates the method on a Duffing oscillator, a seismic frame response simulation, and impact tests on a scaled bridge model, reporting NMSE values as low as 0.398%. An open-source Python implementation is provided.","tokens_in":49019,"tokens_out":8451,"duration_ms":76377,"significance":"If the central convergence and optimality claims were valid, APSMC would be a notable contribution to online nonlinear structural system identification: it combines state estimation with constrained optimization, provides physically interpretable time-varying system matrices, and is supported by code and by tests in simulation and experiment. The empirical framework is attractive and the reported prediction accuracy on the bridge data is competitive. However, the theoretical guarantee is not established and is contradicted by the paper's own statements; the experimental design involves heavy per-dataset tuning; and one of the interpretability claims rests on an algebraic error. The contribution is therefore currently an interesting heuristic with promising experiments rather than a validated theoretical method.","major_comments":[{"comment":"The claimed guarantee lim_{i→∞} ΔA_i = 0 is not proven. The objective (24) is convex for a fixed state sequence, but the regressors x̂_k and targets x̂_{k+1} are outputs of a Kalman filter whose gain and innovation covariance depend on the A being estimated, so the iteration is a coupled stochastic approximation with state-dependent noise. Eq. (26) defines A* as the least-squares fit to the filtered states, so the target is self-referential rather than the data-generating A_true; no contraction, Lyapunov, or stochastic-approximation argument is supplied. The paper's own Sec. 5.3 concedes that with a poor initial A0 the Kalman filter fails and the algorithm cannot update to the correct model, and that it is theoretically unclear whether identity initialization guarantees convergence. Appendix B.2 additionally concedes that the bilinear-transform constraint is non-convex, so proximal gradient does not guarantee a global optimum. These admissions directly contradict the headline guarantee.","section":"Sec. 3.1 (Eqs. 24-29), Sec. 5.3, App. B.2"},{"comment":"The equality asserted between the Duffing system and the Jacobian-based time-varying linear system is algebraically false. Substituting J(x) from (46) into (47) yields ẋ₂ = (α - 3βx₁²)x₁ - c x₂ + F cos(ωt) = α x₁ - 3βx₁³ - c x₂ + F cos(ωt), whereas the original system (44) has ẋ₂ = α x₁ - β x₁³ - c x₂ + F cos(ωt); the difference is -2βx₁³. Consequently x_{k+1} = J(x_k) x_k + B_k u_k is not a representation of the nonlinear system, and the numerical agreement of A_k with J(x_k) in Figs. 6-8 is not a valid ground-truth comparison for the claim that A_k tracks the Jacobian.","section":"Sec. 4.1 (Eqs. 46-48)"},{"comment":"The experimental evidence does not support the claimed consistent superiority of APSMC. The learning rate t_k is hand-tuned per dataset and varies by four orders of magnitude (0.004 to 100). For sequences 15 and 19, APSMC's NMSE (18.847% and 17.028%) is substantially worse than ERA's (10.065% and 11.164%). The 0.398% figure quoted in the abstract is the minimum over all sensors and datasets, not a typical result. The model order and Hankel dimensions (m = n = 3000) are also determined ad hoc, without sensitivity analysis.","section":"Sec. 5.2 (Table 2)"},{"comment":"The algorithm's reproducibility is compromised by unspecified key hyperparameters. The noise covariances Q, R, and P0 for the Kalman filter, the regularization weight λ in Eq. (37), and the step-size schedule t_k used in the numerical and experimental studies are not reported (except for selected learning rates in Table 2). Without these, the results cannot be replicated, and the claim that the method is 'theoretically optimal' cannot be assessed against the actual implementation.","section":"Secs. 4-5 (Algorithms 1-2)"}],"minor_comments":[{"comment":"Many figures contain extensive text corruption (e.g., strings of '/uni00000013/uni00000018/...'), and numerous figures lack captions; the manuscript must be regenerated cleanly before review.","section":"Throughout (Figs. 5-22)"},{"comment":"Eq. (2) uses squared error E[(x_k - g)^2] for vector states; it should be a vector norm (or sum of squares) to define the MMSE estimator.","section":"Sec. 2.1, Eq. (2)"},{"comment":"The statement that noise 'may follow arbitrary distributions of white noise' is an overclaim: Kalman filtering is optimal only under Gaussianity, and the paper neither analyzes nor tests non-Gaussian noise.","section":"Sec. 3.2"},{"comment":"The analogy between the algorithm and SGD is imprecise: online proximal gradient with sequentially ordered samples is not stochastic gradient descent (which samples randomly), and the convergence properties differ.","section":"Sec. 5.3"},{"comment":"Several key references are unpublished preprints, notably [2] (SSRN) and [31] (arXiv), and results in Secs. 4-5 draw heavily on [31] and [9]; these dependencies should be updated or clarified.","section":"References"},{"comment":"The abstract's claim of '19 consecutive 10-second time series using only a single initial 10-second segment for model updating' is misleading because Table 2 shows large errors for several sequences and the model was updated sequentially, not held fixed; the description should be precise.","section":"Abstract and Sec. 5.2"}],"recommendation":"reject","confidential_remarks":"This manuscript appears to be an unreviewed preprint with corrupted figure text and an unusually heavy reliance on the authors' own unpublished preprints. More importantly, the headline theoretical claim is contradicted by the authors' own Sec. 5.3 and Appendix B.2 admissions, and the central proof in Sec. 3.1 is a circular, self-referential argument rather than a theorem. The experimental evaluation, while rich, involves per-dataset tuning of the learning rate and does not consistently favor APSMC over baselines. These issues go beyond presentation and cannot be fixed by local edits; the paper would need a fundamental reframing and a full reanalysis of the claims and experiments."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know about this one. First, the central theoretical claim — that APSMC provably converges to the optimal state-space model as data accumulates — does not hold up. Second, the empirical work is real and the method may well be useful in practice; the paper is just oversold on the theory.\n\nThe genuinely new part is modest: the authors extend their prior APSM method by adding an explicit input term B_k u_k and allowing arbitrary white noise. That is a natural extension, and the scaled-bridge experiment with 20 impact tests is a legitimate piece of validation. They also ship code, which is appreciated. Credit where due: the paper is honest about several limitations, and the physical-constraint idea — using the known Jacobian structure to guide the update — is sensible and demonstrably helps in the Duffing and frame examples.\n\nBut the load-bearing convergence argument in Section 3.1 is a sketch, not a proof. It treats the Kalman-filtered state sequence as if it were the true one, then claims convexity of the loss implies global convergence. That ignores the coupling: the state estimates depend on the A being estimated, so the objective is not fixed even in principle. The 'optimal solution' A* is the least-squares fit to a self-generated state sequence, not to the underlying data-generating system. The paper itself concedes the point twice: in Section 5.3 it says that with a poor initial A0 the Kalman filter fails and the algorithm cannot update, and in Appendix B.2 it admits the bilinear transformation makes the constrained problem non-convex, so proximal gradient does not guarantee a global optimum. That is a direct contradiction with the abstract's claim of guaranteed convergence.\n\nThe experiments also have weaknesses. The learning rate is hand-tuned per sequence (Table 2), hyperparameters are underreported, and in sequences 15, 16, and 19 APSMC is substantially worse than the ERA baseline — not just a little, but three to four times worse. That pattern suggests the method's stability depends on getting the tuning right, which casts further doubt on the 'guaranteed' language.\n\nWho gains from this paper? Practitioners in structural dynamics who want an online, physics-constrained identification heuristic and are willing to tune it will find the recipe useful. As a research claim about convergence to an optimal estimator, it fails. I would send it to peer review only with the expectation that the theory is removed or replaced by a much more limited statement — say, that the scheme is a heuristic for tracking time-varying linearizations. The empirical content deserves a shot, but not with the current framing.","headline":"A useful online system identification recipe with real experiments, but the advertised convergence guarantee is not established and the paper's own caveats undermine it.","tokens_in":49581,"tokens_out":2403,"would_cite":false,"duration_ms":25916,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93B30","93E11"],"pacs":[],"model":"deepseek-v4-flash","headline":"APSMC claims that online estimation of a structure's time-varying linear model from noisy data converges to the theoretically optimal solution, with the identified matrices tracking the Jacobian of the nonlinear dynamics.","keywords":["nonlinear dynamical systems","adaptive system modeling","stochastic subspace identification","optimal estimation","structural dynamics","Kalman filter","proximal gradient optimization","physics-informed modeling"],"falsifier":"Run the algorithm on a linear time-invariant system with known $A$, initialize $A_0$ far from the truth with the sensor count below the state dimension, and monitor $\\|A_k - A^*\\|$: the paper itself reports that in this regime the Kalman filter fails to provide accurate predictions, so the claimed limit $\\lim_{i\\to\\infty}\\Delta A_i = 0$ would be violated. A second check targets the constrained version: because the bilinear transform used to impose continuous-time constraints makes that subproblem non-convex, exhibiting two different initial matrices that converge to two different stationary solutions would contradict the global-optimality claim.","tokens_in":48482,"feed_emoji":"🌉","tokens_out":14365,"duration_ms":117380,"temperature":0.7,"pith_summary":"The paper seeks to establish that a nonlinear structure's dynamics can be identified online, from noisy measurements, as a time-varying linear system, and that the estimate is provably optimal once enough data has accumulated. The claim matters because real structures such as bridges and buildings are monitored continuously while their dynamic properties drift with damage, aging, and extreme loading, and existing data-driven methods are mostly offline, noise-sensitive, and yield models without physical meaning. The proposed APSMC framework couples a Kalman filter with physics-constrained proximal-gradient updates, so each new sample refines the system matrix pair $(A_k, B_k)$ while enforcing structural constraints on the model. Within the stochastic subspace identification framework, the paper asserts that the noise contribution vanishes as data accumulates and the convex updates converge to the global optimum based on the true Kalman-filtered state sequence, making the estimated $A_k$ equal to the Jacobian of the nonlinear dynamics.","feed_headline":"One 10-second clip predicts 19 bridge tests (0.398% error)","feed_subtitle":"The physics-guided algorithm behind it claims provably optimal online model updates from noisy data.","key_machinery":"The mechanism is a coupled loop of two classical tools: a Kalman filter that produces the optimal state estimate $\\hat{x}_k$ from the current model, and a proximal-gradient step that updates the matrix pair $(A_k, B_k)$ by minimizing $\\|\\hat{x}_{k+1} - A_k \\hat{x}_k - B_k u_k\\|^2$ plus a physical regularizer, which is effectively an online least-squares fit against filtered states. What licenses the optimality claim is a stochastic subspace identity: the noise subspace is statistically orthogonal to the signal subspace, so as data grows the residual noise vanishes and the least-squares objective is evaluated against the true Kalman-filtered sequence. Physical priors enter through the proximal mapping, with hard structural constraints such as the block form of a mechanical system's continuous-time matrix acting as projections and soft constraints as regularizers. The upshot is that the time-varying $A_k$ can be read as the Jacobian of the nonlinear dynamics, which gives the identified matrices their physical interpretability.","core_discovery":"On the paper's own terms, the discovery is that merging stochastic subspace identification with an online proximal-gradient loop turns nonlinear structural estimation into a convex problem with a provable optimum: with infinite data the subspace argument eliminates the noise component, leaving the least-squares target $A^* = \\arg\\min_A \\sum_i \\|\\hat{x}_{i+1} - A\\hat{x}_i\\|^2$ built from the true Kalman-filtered states, and each online step solves a convex subproblem of this form, so the paper asserts $\\lim_{i\\to\\infty}\\Delta A_i = 0$ and convergence to the theoretically optimal model. Because the system matrix is locally the Jacobian of the dynamics, the identified matrices retain physical meaning, and when structural constraints are imposed through the proximal mapping the method tracks the Jacobian even where purely data-driven fits fail. Experimentally, the claim is that a single 10-second impact-test segment updates the model well enough to predict 19 unseen 10-second tests, with a best normalized mean square error (NMSE) of 0.398%.","pith_inferences":["Because the filter and the matrix update feed each other, a fully rigorous convergence guarantee would require an alternating-minimization or expectation-maximization analysis of the coupled problem; that analysis is not in the paper, and it may impose conditions beyond convexity of each subproblem.","The paper shows that under weak excitation the least-squares problem is not strictly convex, which suggests a testable extension: characterizing which input spectra let the constrained estimator recover the physical Jacobian rather than merely a minimum-Frobenius-norm fit.","Nothing in the argument pins the filter to the linear-Gaussian Kalman filter, so the same loop with an unscented or particle filter for non-Gaussian noise is a direct and untested extension.","Because the proximal step is generic, the loop could in principle be applied to other locally linearizable physical systems, such as identifying contact or friction laws, by changing only the constraint set."],"forward_implications":["Continuous monitoring becomes a live model-update loop: each new acceleration sample refines the system matrices without storing or reprocessing the full measurement history.","The identified matrices carry physical meaning, since their block structure approximates the structure's mass, stiffness, and damping, so damage and drift could be read directly from the estimates.","Extremely short calibration data may suffice: a single 10-second impact-test segment predicted 19 unseen 10-second responses with a minimum NMSE of 0.398%.","Physical constraints are what preserve accuracy under noise: on a frame structure with 30% added noise, constrained APSMC held prediction NMSE at 4.35% versus 12.56% for unconstrained APSMC and 34.54% for the global linear baseline.","The framework extends to arbitrary white-noise distributions and explicit external inputs, covering cases the earlier APSM formulation could not handle."],"supporting_citations":[{"why":"Reference-based stochastic subspace identification; its argument that the noise subspace vanishes with infinite data is the theoretical foundation of the optimality claim.","marker":"[29]"},{"why":"Principal component stochastic subspace identification; supplies the optimal-estimation form of the state-space model used in equations (6)-(7) and the scaled bridge model details.","marker":"[31]"},{"why":"The earlier APSM method this work extends; provides the online proximal-gradient formulation and the Jacobian interpretation.","marker":"[2]"},{"why":"Eigensystem Realization Algorithm; used to construct the initial state-space model for the bridge impact tests.","marker":"[32]"},{"why":"Dynamic mode decomposition with control; the baseline method APSMC is compared against on the frame structure.","marker":"[38]"},{"why":"Online physics-informed dynamic mode decomposition; the cited source for proximal mappings of physical matrix constraints.","marker":"[10]"},{"why":"Unscented Kalman filter; the cited alternative when the linear-Gaussian assumptions of the Kalman filter are relaxed.","marker":"[30]"}],"fun_headline_variants":["Adaptive physics-AI matches bridge tests from one 10-sec clip","One 10-sec clip trains model for 19 bridge tests, 0.4% error","Online optimization tracks bridge dynamics from minimal data","Physics-constrained updates nail 19 unseen bridge tests","Single 10-sec clip yields optimal nonlinear bridge model"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument assumes that alternating between Kalman filtering, whose state estimates depend on the current system matrix, and a convex least-squares update of that matrix drives the pair to the joint optimum; the paper never proves this coupled fixed-point convergence, and it concedes in Section 5.3 that a poor initial matrix with fewer sensors than states prevents the filter from updating the model correctly.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive physics-AI matches bridge tests from one 10-sec clip","One 10-sec clip trains model for 19 bridge tests, 0.4% error","Online optimization tracks bridge dynamics from minimal data","Physics-constrained updates nail 19 unseen bridge tests","Single 10-sec clip yields optimal nonlinear bridge model"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000629,"raw_usage":{"total_tokens":2953,"prompt_tokens":1034,"completion_tokens":1919,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":650,"completion_tokens_details":{"reasoning_tokens":1829}},"tokens_in":650,"tokens_out":1919,"duration_ms":13781,"temperature":1.0,"reasoning_tokens":1829,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:40:52.642597+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the algorithm on a linear time-invariant system with known $A$, initialize $A_0$ far from the truth with the sensor count below the state dimension, and monitor $\\|A_k - A^*\\|$: the paper itself reports that in this regime the Kalman filter fails to provide accurate predictions, so the claimed limit $\\lim_{i\\to\\infty}\\Delta A_i = 0$ would be violated. A second check targets the constrained version: because the bilinear transform used to impose continuous-time constraints makes that subproblem non-convex, exhibiting two different initial matrices that converge to two different stationary solutions would contradict the global-optimality claim.","supporting_citations":[{"cited_title":"Peeters, G","cited_arxiv_id":null,"evidence_quote":"Reference-based stochastic subspace identification; its argument that the noise subspace vanishes with infinite data is the theoretical foundation of the optimality claim."},{"cited_title":"Principal Component Stochastic Subspace Identification for Output-Only Modal Analysis","cited_arxiv_id":"2504.00317","evidence_quote":"Principal component stochastic subspace identification; supplies the optimal-estimation form of the state-space model used in equations (6)-(7) and the scaled bridge model details."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The earlier APSM method this work extends; provides the online proximal-gradient formulation and the Jacobian interpretation."},{"cited_title":"Juang, R","cited_arxiv_id":null,"evidence_quote":"Eigensystem Realization Algorithm; used to construct the initial state-space model for the bridge impact tests."},{"cited_title":"Online Physics-Informed Dynamic Mode Decomposition: Theory and Applications","cited_arxiv_id":"2412.03609","evidence_quote":"Online physics-informed dynamic mode decomposition; the cited source for proximal mappings of physical matrix constraints."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Unscented Kalman filter; the cited alternative when the linear-Gaussian assumptions of the Kalman filter are relaxed."}],"review_version":1}