{"id":"d61effd4-fea6-436f-b867-2d4d16807494","arxiv_id":"2412.03609","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"OPIDMD combines online proximal gradient descent with physics-informed matrix constraints to learn time-varying linear models of dynamical systems, claiming state-of-the-art short-term prediction on noisy benchmarks.","lead":"An online variant of Dynamic Mode Decomposition that adds physics-style matrix constraints through convex optimization, tested on noisy simulations of fluid, quantum, and chaotic systems. It reports better short-term forecasts than standard DMD methods, notably R2 = 0.991 on a noisy Lorenz system, which could matter for real-time monitoring and control.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed convergence to a unique global optimum is not proven and is likely false: Algorithm 1 minimizes a time-varying objective with a fixed step size, and no regret or dynamic regret bound is supplied.","rationale":"I read the paper in good faith: the convex reformulations of several matrix-constrained DMD problems and the Bayesian/MAP equivalence are correct, and the empirical demonstration is a reasonable engineering contribution. The decisive condition for the headline claim is that the online proximal gradient iterate converges to the global optimum of the cumulative objective. That condition is not established: Algorithm 1 uses one-point gradients on a moving objective, and Appendix A.2 provides no theorem, only an assertion that convexity plus sequential data implies convergence to a unique optimum, which is false for fixed-step online gradient descent. The empirical section's choice of per-system constraints, single R2 values without error bars or hyperparameters, and absence of code further weaken the 'best prediction performance' claim, but they are secondary to the theoretical gap. The reader's weakest_assumption identifies the same issue, and the conditional verdict remains appropriate: the paper should either add a convergence or regret analysis, or substantially weaken the optimality guarantee it advertises.","tokens_in":58282,"tokens_out":4180,"duration_ms":41545,"concrete_test":"Construct a static ground-truth system (n=3, fixed A*, x_k ~ N(0,I), y_k = A*x_k + Gaussian noise) and run Algorithm 1 with the paper's fixed step size for the same number of steps as the Lorenz experiment, 149,850. Compute the exact minimizer A_batch of the full cumulative objective (same lambda and constraint, e.g., tridiagonal projection) and report ||A_k - A_batch||_F. If this gap does not decrease to zero (say below 1e-6) as k grows, the claimed convergence to a unique global optimum is falsified; if it does, the general claim still needs a proof, but the specific claim would be supported in at least one relevant setting.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central theoretical claim is that OPIDMD 'ensures convergence to a unique global optimum' (Abstract) because each subproblem is convex. The load-bearing step is Algorithm 1, which updates A_{k+1} = prox_{t_k R}(A_k - t_k grad g_new(A_k)) using only the gradient of the newly arrived term g_new(A) = ||y_new - A x_new||^2, not the cumulative objective sum_{i=1}^k ||y_i - A x_i||^2 + lambda R(A). This is online gradient descent on a sequence of changing objectives. Appendix A.2 asserts that 'the inherent nature of convex problems guarantees... convergence to this unique global optimum,' but standard online convex optimization provides regret bounds, not convergence to a fixed optimum; with a fixed step size the iterates generally track a moving minimizer or oscillate in a neighborhood of it. No regret or dynamic regret bound is given. Uniqueness is also not guaranteed for l1 or trace-norm penalties when the quadratic term is not strictly convex, e.g., a zero data point leaves all A optimal. Since the abstract's global-optimum guarantee rests on this unsupported assertion, the central claim is vulnerable.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes OPIDMD, an online extension of physics-informed Dynamic Mode Decomposition. It casts piDMD as a convex composite optimization problem, gives a Bayesian/MAP interpretation of common regularizers, and applies online proximal gradient descent with projections or soft-thresholding for eight constraint types. Experiments on advection, Schrödinger, advection-diffusion, cylinder flow, Lorenz, and a five-degree-of-freedom mechanical system report R² values and claim superior short-term forecasting, most notably R²=0.991 on the noisy Lorenz system. The paper also provides a bias-variance analysis for the l2-regularized estimator and releases an open-source Python implementation.","tokens_in":58526,"tokens_out":5166,"duration_ms":52757,"significance":"If the theoretical guarantee were established, OPIDMD would be a useful general-purpose real-time solver for physics-constrained DMD. The paper has concrete strengths: explicit algorithms for each constraint, a mostly correct derivation of constrained-versus-penalized equivalence under Slater's condition, and an open-source implementation that aids reproducibility. However, the central online-convergence claim is not supported by the current theory, and the headline Lorenz result is weakened by constraint selection after inspection of the true operator and by single-run evaluations without error bars. The paper is therefore better described as a promising engineering contribution whose current claims outrun its evidence.","major_comments":[{"comment":"The claim that Algorithm 1 'ensures convergence to a unique global optimum' is not established. Algorithm 1 performs an online proximal gradient step using only the gradient of the newly arrived loss g_new(A) = ||y_new - A x_new||^2, whereas the cumulative objective in Eq. (26) changes at every iteration. Standard online convex optimization provides regret bounds, not convergence to a fixed minimizer, and with a fixed step size the iterates typically track or oscillate around a moving minimizer. The assertion in Appendix A.2 that 'the inherent nature of convex problems guarantees' convergence to a unique global optimum is therefore incorrect as stated. In addition, uniqueness is not guaranteed for the l1 and trace-norm penalized problems when the quadratic data term is not strictly convex, e.g., when a zero snapshot arrives. Please replace the abstract's guarantee with a concrete regret or dynamic-regret analysis, or modify the algorithm so that it actually minimizes the cumulative objective.","section":"Section 4.2, Algorithm 1, and Appendix A.2"},{"comment":"The Lorenz R²=0.991 result is not an independent prediction of the method. The tridiagonal constraint is selected after inspecting the True DMD matrix: the text states that the local dependency 'leads the True DMD matrix to exhibit a tridiagonal structure. Hence, we verify the physical validity of the model by imposing the tridiagonal matrix as a physical constraint.' Since the constraint family is chosen using the noise-free target operator, the experiment demonstrates that the chosen family contains the true structure; it does not demonstrate that OPIDMD discovers that structure from noisy data alone. Please provide an a priori physical justification for the constraint, or use a constraint-selection protocol based only on training/validation data, and evaluate on held-out systems.","section":"Section 5.3 and Table 4"},{"comment":"The abstract's claim that OPIDMD 'achieves the best prediction performance in short-term forecasting' is contradicted by the paper's own results. In Table 4, for Cylinder Flow, Exact DMD achieves R²=0.938 while OPIDMD achieves 0.935, and Section 5.2 explicitly states that 'Exact DMD proved optimal as a global linear model and achieved the best predictive performance.' The claim in the abstract should be restricted to the settings in which it actually holds, or the comparison should be re-run and reported consistently.","section":"Abstract and Table 4"},{"comment":"The numerical evaluation lacks uncertainty quantification. All R² values are point estimates from a single noisy realization, with no standard errors, confidence intervals, or multiple noise seeds. Several reported differences between methods are small (e.g., 0.935 versus 0.938 for Cylinder Flow), so without repeated trials the claimed advantages are not statistically substantiated. Please report mean ± standard deviation over multiple noise realizations and state the number of runs used.","section":"Section 5, Tables 4 and 5"}],"minor_comments":[{"comment":"The indicator function I_C(x) is defined without formally introducing the set C in the surrounding text; please clarify that C denotes the constraint set in matrix space.","section":"Equation (24)"},{"comment":"Algorithm 8 is titled 'Proximal Operator for unclear Norm Constraint' and repeatedly says 'unclear norm'; this should be 'nuclear norm' throughout.","section":"Algorithm 8"},{"comment":"The noise model is described as '25% non-stationary Gaussian white noise,' but 'non-stationary' is not defined; please specify how the noise variance or distribution changes across time and whether the same model is used for every example.","section":"Section 5, opening paragraph"},{"comment":"The ODMD R² entries are astronomically negative (e.g., -9.73 × 10^196), which suggests numerical overflow or a degenerate fit; please report these values in a more informative way, such as on a clipped scale or as 'not meaningful.'","section":"Table 4"},{"comment":"The comparison for Cylinder Flow uses all eigenvalues for piDMD but only the first 50 eigenvalues for Exact DMD, OGD, and OPIDMD; this asymmetry should be justified, since it could affect the reported R² values.","section":"Section 5.2"},{"comment":"The complexity table lists 'Streaming DMD' and 'Online DMD' but the main text does not clearly define the exact versions used; please give precise references or pseudocode for the compared online algorithms.","section":"Appendix D"}],"recommendation":"major_revision","confidential_remarks":"The paper is best positioned as an engineering/methods contribution rather than a theory paper. The main revision should remove or properly qualify the global-optimum guarantee, add a regret analysis or change the algorithm to a cumulative-objective solver, and restructure the Lorenz experiment to avoid circular constraint selection. If the authors address these points, the contribution may be publishable as an applied DMD methodology paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know: this paper adapts piDMD to an online setting by running proximal gradient descent on one new data point at a time, and that combination is genuinely new. The convexity table, the Slater-condition equivalence between constrained and penalized forms, and the proximal/projection operators are all correct. The Bayesian MAP interpretation of piDMD is standard but cleanly laid out, and the bias-variance analysis for the l2 case is a real addition. The numerical work is broad and shows OPIDMD beating Exact DMD and piDMD on several noisy short-term forecasting tasks. That is real engineering value.\n\nBut the headline claim that the algorithm 'ensures convergence to a unique global optimum' does not hold up. Algorithm 1 is online gradient descent on a sequence of changing objectives, and the appendix simply asserts that convexity guarantees convergence. Standard online convex optimization gives regret bounds, not convergence to a fixed optimum, and with a fixed step size the iterates can track a moving minimizer or oscillate. No regret or dynamic regret bound is supplied. Uniqueness also fails for l1 or trace-norm penalties when the quadratic term is not strictly convex, e.g., with a zero data point. The abstract's central theoretical selling point is therefore not established.\n\nTwo more soft spots. The Lorenz R2 of 0.991 is partly circular: the tridiagonal constraint was chosen after inspecting the True DMD matrix in Section 5.3, so that result is a verification that the chosen family contains the true structure, not an independent prediction. And the per-system constraint selection in Table 5 is done to maximize performance, which weakens the 'best prediction performance' claim. The paper also reports single R2 values without error bars, gives no hyperparameters, and mentions a Github implementation but provides no link. All of these are fixable.\n\nThe citation pattern looks fine; the authors engage with piDMD, online DMD, Bayesian DMD, and streaming DMD. This is not a fundamental advance, but it is a competent combination that practitioners in structural dynamics, fluids, or control could use for real-time modeling of slowly varying systems.\n\nRecommendation: send it to peer review, but require the authors to either supply a proper dynamic regret analysis or soften the convergence claim, report multiple seeds with error bars, and release the code. The core idea is salvageable and worth referee time.","headline":"Useful online extension of piDMD, but the global-convergence guarantee is unsupported and the Lorenz result is partly self-fulfilling.","tokens_in":59032,"tokens_out":1950,"would_cite":false,"duration_ms":22051,"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":"OPIDMD recasts physics-constrained DMD as convex online optimization and reports an R² of 0.991 on a noisy Lorenz forecast.","keywords":["dynamic mode decomposition","physics-informed machine learning","online learning","proximal gradient descent","convex optimization","Bayesian DMD","time-varying dynamical systems","short-term forecasting"],"falsifier":"Train OPIDMD with a fixed step size on a stream generated by a linear system whose true operator switches between two matrices at a known time, then compare the final iterate $A_k$ against the exact minimizer of the full-batch convex problem using all data seen so far; if they differ materially, or if the per-step objective does not settle, the global-optimality claim is falsified.","tokens_in":1527,"feed_emoji":"🌀","tokens_out":2069,"duration_ms":60912,"temperature":0.7,"pith_summary":"This paper introduces Online Physics-informed DMD (OPIDMD), an algorithm that solves physics-constrained Dynamic Mode Decomposition as a convex optimization problem and updates the learned linear operator one data point at a time. The authors' central claim is that this formulation converges to a unique global optimum, and that the resulting time-varying operator gives the best short-term forecasting among Exact DMD, Online DMD, and piDMD across the tested systems, including an R² of 0.991 on a noisy Lorenz system. The wider promise is a real-time solver for physics-informed DMD that scales to streaming data and resists noise.","feed_headline":"Physics-constrained DMD goes online, hitting R² 0.991 on noisy Lorenz","feed_subtitle":"A convex proximal-gradient solver updates the DMD operator point-by-point, outpacing Exact DMD, piDMD, and Online DMD on short-term…","key_machinery":"The load-bearing object is the online proximal gradient step $A_{k+1} = \\operatorname{prox}_{tR}\\big(A_k - t\\nabla g(A_k)\\big)$ for the composite objective $f(A) = \\|Y_k - A_k X_k\\|_F^2 + \\lambda R(A_k)$, where $\\operatorname{prox}_{tR}$ is the proximal mapping associated with the physical constraint. Each structural constraint receives its own proximal operator: projection onto circulant, symmetric, upper-triangular, or tridiagonal matrices; singular-value soft thresholding for the trace norm; elementwise soft thresholding for the $\\ell^1$ norm; and plain gradient descent for the $\\ell^2$ norm. The same convex framework is interpreted probabilistically through Bayesian MAP estimation, with the regularizer playing the role of a matrix prior. A bias-variance calculation for the $\\ell^2$ case shows that stronger regularization shrinks the eigenvalues of the estimator and reduces variance at the cost of added bias.","core_discovery":"The paper's core discovery is that the physics-informed DMD problem can be reformulated as minimizing $\\|Y - AX\\|_F^2 + \\lambda R(A)$, with physical constraints represented as convex regularizers or projection sets, and then solved online by proximal gradient descent. Under Slater's condition the constrained and penalized forms are equivalent, and the Bayesian MAP interpretation shows that the penalty term acts as a prior. Because the data arrives sequentially, each update only needs the latest snapshot pair, and the method can accommodate implicit, circulant, symmetric, triangular, tridiagonal, low-rank, $\\ell^1$, and $\\ell^2$ constraints. The paper reports that OPIDMD's time-varying linear operator acts as a local linearization, which is why short-term forecasting on the chaotic Lorenz system reaches an $R^2$ of 0.991 while Exact DMD, Online DMD, and piDMD stay near 0.06.","pith_inferences":["The paper's global-optimality claim is stated for the online objective, not for the time-varying cumulative loss; a natural test is to measure dynamic regret against the best fixed matrix in hindsight, which the paper does not present.","The tridiagonal constraint is chosen because the Lorenz system's true DMD matrix is nearly tridiagonal; an extension would be to select the constraint adaptively via online cross-validation rather than by inspection.","The bias-variance analysis is carried out only for the $\\ell^2$ penalty; analogous shrinkage and sparsity-bias statements for the trace norm and $\\ell^1$ norm would require separate calculations.","Since OGD's implicit regularization is invoked as the foundation, a precise characterization of that implicit bias for online least squares would strengthen the explanation of the smooth matrices observed in the experiments."],"forward_implications":["Physics-informed DMD can be maintained online in $O(n^2)$ per new snapshot for most constraints, instead of re-solving the full-batch problem at every data arrival.","The learned operator is time-varying, so each step's $A_k$ acts as a local linear approximation of a nonlinear system; this is the mechanism behind the reported Lorenz result.","Slater's condition plus convexity makes the constrained piDMD problem and the penalized MAP problem equivalent, so a single convex solver covers the eight physical constraints considered.","Physical constraints act as a bias-variance knob: they reduce forecast variance on noisy data, and overly strong constraints add bias when they do not match the true dynamics.","Even without explicit constraints, online gradient descent's implicit regularization gives noise robustness and simple, low-rank-like matrices."],"supporting_citations":[{"why":"Supplies the piDMD manifold-constrained formulation that OPIDMD extends into an online convex solver.","marker":"[23]"},{"why":"Gives the Exact DMD rank-constrained optimization problem that the paper reinterprets in convex and probabilistic form.","marker":"[13]"},{"why":"Provides the Bayesian DMD probabilistic framework used for the MAP equivalence and bias-variance analysis.","marker":"[29]"},{"why":"Defines the online DMD scenario for slowly time-varying systems that OPIDMD is designed to improve.","marker":"[33]"},{"why":"Supplies the online proximal gradient descent method adopted as the core solver.","marker":"[60]"},{"why":"Offers the streaming DMD algorithm used as a baseline in the computational time comparison.","marker":"[38]"},{"why":"Provides the DMD monograph and the cylinder flow dataset used in the numerical experiments.","marker":"[9]"}],"fun_headline_variants":["Online physics-informed DMD: real-time, R^2 0.991 on Lorenz","Proximal-gradient DMD goes online, surpasses Exact and piDMD","Physics-constrained DMD online, R^2 0.991 on noisy Lorenz","OPIDMD: online physics-informed DMD, R^2 0.991 short-term","Real-time DMD with physics constraints, R^2 0.991 on Lorenz"],"cache_read_input_tokens":61184,"weakest_assumption_plain":"The claim that the method converges to a unique global optimum rests on the assumption that online proximal gradient descent with a fixed step size converges to the optimum of the entire accumulated objective even though the objective is rebuilt from a new data point at every step, and no regret or dynamic-regret bound is supplied.","fun_headline_variants_meta":{"raw":{"variants":["Online physics-informed DMD: real-time, R^2 0.991 on Lorenz","Proximal-gradient DMD goes online, surpasses Exact and piDMD","Physics-constrained DMD online, R^2 0.991 on noisy Lorenz","OPIDMD: online physics-informed DMD, R^2 0.991 short-term","Real-time DMD with physics constraints, R^2 0.991 on Lorenz"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001208,"raw_usage":{"total_tokens":4989,"prompt_tokens":970,"completion_tokens":4019,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":586,"completion_tokens_details":{"reasoning_tokens":3906}},"tokens_in":586,"tokens_out":4019,"duration_ms":26091,"temperature":1.0,"reasoning_tokens":3906,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T22:35:42.638276+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train OPIDMD with a fixed step size on a stream generated by a linear system whose true operator switches between two matrices at a known time, then compare the final iterate $A_k$ against the exact minimizer of the full-batch convex problem using all data seen so far; if they differ materially, or if the per-step objective does not settle, the global-optimality claim is falsified.","supporting_citations":[{"cited_title":"In: Proceedings of the Twenty-Sixth International Joint Conference on Artificial Intelligence, pp","cited_arxiv_id":null,"evidence_quote":"Provides the Bayesian DMD probabilistic framework used for the MAP equivalence and bias-variance analysis."},{"cited_title":"In: Proceedings of COMPSTAT’2010: 19th International Conference on Computational StatisticsParis France, August 22-27, 2010 Keynote, Invited and Contributed Papers, pp","cited_arxiv_id":null,"evidence_quote":"Supplies the online proximal gradient descent method adopted as the core solver."},{"cited_title":"Accessed 2023-12-12","cited_arxiv_id":null,"evidence_quote":"Offers the streaming DMD algorithm used as a baseline in the computational time comparison."},{"cited_title":"SIAM, ??? (2016)","cited_arxiv_id":null,"evidence_quote":"Provides the DMD monograph and the cylinder flow dataset used in the numerical experiments."}],"review_version":1}