{"id":"7bc5c7a6-60d0-409d-abd9-f2591df30092","arxiv_id":"2506.03650","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"With fast sampling, a state variable filter least squares method yields parameter estimation error variance O(h) in the sampling interval, even with colored noise, offsets, and closed-loop operation.","lead":"This paper shows that identifying continuous-time models from very fast sampled data, using a simple state variable filter least squares fit, makes parameter error shrink in proportion to the sampling interval. That makes fast sampling a practical lever for hard identification problems such as unstable closed-loop systems and sensors with offsets.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The O(h) covariance claim rests on treating the SVF regressor D_F as deterministic; the correlation between filtered noise in D_F and Y_A is never bounded, so Theorem 1 does not yet apply to Eq. (42).","rationale":"The reader's weakest-assumption identifies the load-bearing gap: the SVF estimator (42) is not the deterministic-regressor regression (11), and Section III-D.1's phrase 'we can treat D_F as effectively deterministic' is an assertion, not a rate. Lemma 1 gives O(h) per-entry filtered-noise variance, but the least-squares estimator involves sums of O(1/h) such entries, so the normalized cross-covariance between noise in D_F and noise in Y_A must be shown to vanish at O(h); this is precisely the classical source of SVF bias. The paper's numerical experiments, especially the parameter-error and nu-gap convergence plots, provide genuine empirical support for the O(h) scaling, so the claim is plausible and the correct disposition remains conditional rather than a rejection. The concern does not change the reader's CONDITIONAL verdict; it does specify the missing proof step that should be supplied before the theoretical claim is accepted as fully general.","tokens_in":18026,"tokens_out":10783,"duration_ms":137442,"concrete_test":"Re-derive the bias/covariance of (42) by expanding D_F = D_0 + Delta, Y_A = Y_{A0} + e, where Delta and e are the filtered-noise parts, and compute the leading h-order of the normalized cross-terms (h/T_f) E[Delta^T e] and (h/T_f) E[Delta^T Delta] for the P1 closed-loop setup with the paper's zero-order-hold noise model. If both are O(h) or smaller and the resulting covariance matches Theorem 1's O(h) bound, the 'effectively deterministic' step is justified; if either saturates at O(1), the theorem does not cover the actual SVF estimator.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III-D.1 jumps from Lemma 1 (each filtered noise component has variance O(h)) to 'we can treat D_F as effectively deterministic' and then invokes Theorem 1 with Phi_h = D_F and Y_h = Y_A. The estimator (42) is not the regression model (11): D_F contains the same filtered w and eta that produce the equation error in Y_A, so the regressor is stochastic and, in closed loop, structurally correlated with the noise. To apply Theorem 1 one must prove, not assert, that the normalized cross-terms vanish at a known rate, e.g. that (h/T_f) E[Delta^T e] is O(h), where Delta and e are the filtered-noise parts of D_F and Y_A. This is exactly the classical SVF bias: E[(D_F^T D_F)^{-1} D_F^T (noise)] is nonzero when regressor and equation-error noise are correlated. Per-entry variance O(h) does not by itself rule out an O(1) normalized bias, because the sums defining the least-squares estimator contain O(1/h) terms. The numerical results are consistent with the claimed scaling, but they do not supply the missing rate; until that calculation is provided, the actual SVF estimator is not covered by Theorem 1.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper claims that using a state variable filter (SVF)-like least squares method for continuous-time system identification, the variance of the parameter estimation error scales as O(h) with the sampling interval h, and that this scaling holds even with colored noise, noise correlations, closed-loop operation, measurement offsets, and unstable MIMO plants. The theoretical basis is Theorem 1, which gives an O(h) covariance bound for least squares with a deterministic regressor and bounded total noise energy. The authors then apply Theorem 1 to the SVF estimator by asserting that the noisy regressor matrix D_F can be treated as deterministic in the limit h -> 0. Numerical experiments on several plants, including an unstable MIMO system, show error decreasing as h decreases, seemingly confirming the O(h) scaling.","tokens_in":18279,"tokens_out":10104,"duration_ms":115839,"significance":"If the central claim is correct, the result is practically important: it would mean that fast sampling plus a simple, classical SVF least squares fit can solve closed-loop identification, offset rejection, and unstable MIMO identification without controller knowledge or specialized algorithms. The paper is clearly written and the numerical study is extensive, including a meaningful metric (ν-gap) that shows improvement of SVF over discrete-time ARX/SSARX as h decreases. However, the central theoretical step—applying a theorem about deterministic regressors to an estimator with a stochastic, noise-correlated regressor—is not justified in the manuscript, and the noise model underlying the O(h) scaling is too narrow to support the abstract's 'independent of noise color' claim. These issues are load-bearing, so the paper needs substantial revision before the claim is established.","major_comments":[{"comment":"The proof that the SVF least squares estimator obeys the O(h) covariance bound is missing. Theorem 1 requires the regressor Φ_h to be deterministic, but in (42) the regressor D_F is constructed from filtered noisy measurements and is therefore stochastic and, in closed loop, correlated with the equation error in Y_A because both contain filtered versions of the same noises w and η. The sentence 'we can treat D_F as effectively deterministic' is an assertion, not a derivation; the classical SVF bias arises precisely from this correlation. To invoke Theorem 1 one must prove, not assert, that the normalized cross terms vanish at a known rate, for example that (h/T_f) E[D_F^T (Y_A - D_F θ*)] = O(h) and that (h/T_f) D_F^T D_F converges to a positive definite limit in a suitable sense. Until this calculation is supplied, the O(h) variance claim for the actual SVF estimator is unproven.","section":"Section III-D.1, Eq. (42)"},{"comment":"The O(h) variance scaling in Lemma 1 is derived for a discrete-time white-noise input whose per-sample variance is fixed. For bandlimited or continuous-time colored noise, the per-sample variance of the filter output does not decrease once h is below the noise correlation time; then the total noise energy over a fixed interval T_f scales as O(1/h), violating Assumption (A2) and invalidating the O(h) covariance conclusion. The paper's claim that the scaling holds 'independent of noise color' is therefore not supported. The numerical experiments avoid this regime because the noise is generated by sampling and holding white noise with a fixed 5×10^-6 s update interval, and h is never smaller than 10^-5 s. Please either restrict the claim to the discrete-time white-noise model with per-sample variance independent of h, or provide an analysis of bandlimited noise showing when Tr(Σ_h)=O(1) remains valid.","section":"Section II-B, Lemma 1 and Section II-C, Assumption (A2)"},{"comment":"The offset-removal procedure relies on the filtered version of a constant signal decaying as t → ∞, and the paper states that discarding data before t ≥ 15 s suffices. However, this is a heuristic choice: the decay rate depends on the chosen filter and the value of the offset, and the discarded interval is not part of the formal problem statement. The paper should either specify how the settling time is chosen or note that the method requires a user-selected truncation that is not covered by the O(h) analysis.","section":"Section III-D.2, offset handling"}],"minor_comments":[{"comment":"In Remark 2, the text says 'Tr(Cov(θ̂)) = Σ E[v(kh)^2]', but the quantity being constrained by Assumption (A2) is the total noise variance Tr(Σ_h), not the covariance of the estimate. Please correct the notation.","section":"Section II-C, Remark 2"},{"comment":"The proof of Lemma 1 uses '≈' without explicit error bounds. Please state the regularity assumptions on F(s) (stable, strictly proper, with minimal realization) and give an asymptotic expansion such as ||F_h(z)||_2^2 = h ||F(s)||_2^2 + O(h^2) as h → 0, with a brief justification.","section":"Section II-B, Lemma 1"},{"comment":"The symbol E is used both for the expectation operator and for the residual vector in (37)-(38). This is confusing; please use a different symbol for the residual, such as R or ε.","section":"Section III-C, Eqs. (37)-(38)"},{"comment":"The indexing in the definition of the held signals ̅u(t) and ̅y(t) is introduced only here; please define it earlier in Section III and ensure the round/ceil notation is consistent with the sampling instants t=k h used elsewhere.","section":"Section IV-B, Eq. (70)"},{"comment":"The conclusion states the scaling is valid 'even with the presence of colored noise or the noise correlations between variables', but the theoretical and numerical sections only treat additive noise that is white at the sampling instants with independent MIMO channels. Please soften this sentence or provide supporting analysis for correlated variables.","section":"Section V, Conclusions"}],"recommendation":"major_revision","confidential_remarks":"The core idea is attractive and the simulations are convincing, but the theoretical gap in Section III-D.1 is central: the paper applies a deterministic-regressor theorem to a stochastic, noise-correlated regressor without proof. This is fixable—a direct bias-variance calculation may well yield the claimed O(h) rate—but it must be written out. The noise-model issue in Lemma 1/A2 is also important and may require restricting the claim in the abstract. The self-citation [30] is not used circularly; it only notes an equivalence that is not load-bearing. I recommend a major revision rather than rejection, since the numerical evidence indicates the phenomenon is real and the missing proof is within reach."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Good paper to know about. The claim is that sampling much faster than the usual rule turns the old, biased SVF method into a workhorse for hard identification problems: closed-loop, unstable, offset-ridden, MIMO, all with one least-squares fit. The experiments support this: they sweep h over four decades, show the error going down as O(h) to O(h^2), and compare fairly against ARX/SSARX. The nu-gap metric for unstable plants is sensible. I buy the empirical phenomenon.\n\nWhat is genuinely new is the observation that fast sampling, by shrinking the variance of filtered noise while keeping the signal-to-noise ratio fixed, removes the bias that has kept SVF out of mainstream use. The scaling theorem itself (Theorem 1) is a few lines of norm inequalities and is fine. The problem is in the application. In Section III-D.1 the authors need to go from 'each filtered noise component has variance O(h)' to 'we can treat D_F as effectively deterministic.' That step is asserted, not proved. D_F contains the same filtered noise that appears in the equation error Y_A; in closed loop the correlation is structural. The classical SVF bias is exactly the non-zero expectation of (D_F^T D_F)^{-1} D_F^T e. Per-entry variance O(h) does not rule out an O(1) normalized cross-term, because the sums contain O(1/h) terms. So Theorem 1 does not currently cover the actual estimator in Eq. (42). The numerical evidence is consistent with the missing bound holding, and a competent referee could likely extract the right rate from the authors, but as written the proof is incomplete.\n\nThere is also an untested caveat: the O(h) law relies on total noise energy staying O(1) as h shrinks, which holds for aliased white noise without anti-aliasing filters, but real bandlimited noise would stop delivering O(h) gains at some point. The authors do not discuss that saturation.\n\nBottom line: this is a serious paper, worth refereeing and probably worth publishing after the proof gap is closed. The applications are real and the experiments are honest. I would read it carefully and ask for the missing correlation bound.\n\nWho it is for: anyone working in continuous-time identification or closed-loop identification. The idea is simple enough to be useful in practice. Recommendation: send to peer review. The gap is real but repairs are plausible.","headline":"A serious empirical paper with a real proof gap: the O(h) variance claim for the actual SVF estimator is asserted, not proven, because the regressor D_F is stochastic and correlated with the equation error.","tokens_in":18803,"tokens_out":5342,"would_cite":true,"duration_ms":60294,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Fast sampling makes state-variable-filter identification errors vanish","keywords":["Continuous-time models","Closed-loop identification","Estimation error variance","Fast sampling","State variable filter","System identification","Aliasing noise","MIMO systems"],"falsifier":"Simulate the paper's closed-loop unstable plant P1 with a colored disturbance concentrated in the passband of the SVF filter, then compute the mean squared parameter error for $h = 1$ ms, $0.1$ ms, and $0.01$ ms; if the log-log slope departs from $-1$ or flattens, the deterministic-regressor assumption is the bottleneck.","tokens_in":17791,"feed_emoji":"⚡","tokens_out":6968,"duration_ms":73114,"temperature":0.7,"pith_summary":"The paper argues that the classical rule of sampling about ten times the system bandwidth is not a fundamental limit: when identification is done in continuous time with a state variable filter, the variance of the least-squares parameter estimate is proportional to the sampling interval $h$, so making $h$ smaller monotonically improves accuracy. The scaling is claimed to hold even when the noise is colored, correlated with the input, or carries constant offsets, and it applies in closed-loop operation with an unknown, possibly nonlinear controller. A sympathetic reader should care because this turns a single linear regression, with no solver sophistication and no controller information, into a method that can identify unstable MIMO plants under heavy noise. The paper supports the claim with simulations in stable, unstable, offset-ridden, and MIMO closed-loop settings.","feed_headline":"Fast sampling makes state-variable-filter errors vanish","feed_subtitle":"A least-squares SVF fit improves as the sampling interval shrinks, handling noise, offsets, and unstable closed-loop plants.","key_machinery":"The machine is the SVF-like least-squares estimator: pass the measured input $u$ and output $y$ through a bank of strictly proper filters $F(p)$, form $Y_A$ from the filtered $n$-th derivative of $y$ and $D_F$ from the filtered lower derivatives of $u$ and $y$, and compute $\\hat{\\theta} = (D_F^\\top D_F)^{-1} D_F^\\top Y_A$. Its work is to make the regression asymptotically deterministic: the filter $H_2$-norm lemma gives filtered noise variance $O(h)$, while the deterministic signal component is preserved, so Theorem 1 applies. A numerator factor $p$ in $F(p)$ also nullifies constant offsets by making the filter's steady-state response to a constant vanish.","core_discovery":"The central claim is Theorem 1: for the regression $Y_h = \\Phi_h \\theta^\\star + V_h$, if the regressor matrix $\\Phi_h$ is deterministic and well conditioned as $h \\to 0$ and the total noise energy $\\operatorname{Tr}(\\Sigma_h)$ is $O(1)$, then the least-squares estimate satisfies $\\operatorname{Tr}(\\operatorname{Cov}(\\hat{\\theta})) = O(h)$. The paper's application claim is that the SVF regressors inherit exactly these conditions because they are outputs of continuous-time filters, so they neither lose rank, unlike discrete-time regressors as $h \\to 0$, nor carry noise whose variance fails to shrink. Lemma 1 supplies the mechanism: for any strictly proper filter, the variance of filtered white noise is approximately $h$ times the continuous-time $H_2$ norm. The paper concludes from this that fast sampling plus the SVF-like method is a general-purpose cure for closed-loop identification, offset rejection, and unstable MIMO identification, requiring only one least-squares solve.","pith_inferences":["The $O(h)$ argument should extend to any continuous-time estimator built from filtered regressors whose filters are strictly proper, such as instrumental-variable variants, but the paper only proves it for plain least squares with an effectively deterministic regressor.","A practical design rule suggested by the paper is the opposite of the classical one: sample as fast as hardware and conditioning allow, and use continuous-time filtering in place of effortful pre-sampling anti-aliasing.","The theorem's deterministic-regressor assumption leaves the classical SVF bias unevaluated; whether the $O(h)$ slope survives in practice may depend on how fast the noise-regressor correlation decays, which the filter's roll-off controls.","A hardware experiment without anti-aliasing filters would be the sharpest test, because the aliasing-noise mechanism the paper identifies should appear as the dominant $O(h)$ floor."],"forward_implications":["As the sampling interval $h$ is decreased, SVF least-squares identification improves without a lower limit, unlike discrete-time ARX or SSARX methods, which deteriorate once $h$ is too small.","Closed-loop identification of unstable plants no longer requires knowledge of the controller, the reference signals, or any nonlinearity in the feedback path.","Constant offsets in input disturbance and output measurement are rejected by choosing a filter with a zero at $p=0$, with no need to estimate the offset sizes.","The result carries over to MIMO systems with a common-denominator polynomial model, requiring only one least-squares fit over all channels."],"supporting_citations":[{"why":"Establishes the 'ten times the bandwidth' sampling rule and the discrete-time pole-clustering problem the paper argues against.","marker":"[1]"},{"why":"Supplies the early rationale behind the classical sampling-rate guideline.","marker":"[2]"},{"why":"Documents numerical deterioration of discrete-time models under fast sampling via the delta operator.","marker":"[3]"},{"why":"Provides the continuous-time identification setting and sampled-data machinery the SVF method builds on.","marker":"[4]"},{"why":"Shows the fixed-pole observer method, equivalent to SVF, can identify unstable plants in closed loop.","marker":"[30]"},{"why":"Defines the ν-gap metric used to compare identified models of unstable plants.","marker":"[31]"}],"fun_headline_variants":["Fast sampling shrinks SVF errors at O(h)","SVF least squares: error drops linearly with sample time","Beat the ten-times rule: SVF handles fast sampling","Closed-loop ID solved by fast sampling + SVF","SVF beats bandwidth limits: error O(h) under fast sampling"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The result assumes the noisy regressor matrix can be treated as deterministic as the sampling interval goes to zero; the actual SVF bias comes from the correlation between that noise and the output error, and the paper does not derive the rate at which that correlation vanishes.","fun_headline_variants_meta":{"raw":{"variants":["Fast sampling shrinks SVF errors at O(h)","SVF least squares: error drops linearly with sample time","Beat the ten-times rule: SVF handles fast sampling","Closed-loop ID solved by fast sampling + SVF","SVF beats bandwidth limits: error O(h) under fast sampling"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000526,"raw_usage":{"total_tokens":2527,"prompt_tokens":922,"completion_tokens":1605,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":538,"completion_tokens_details":{"reasoning_tokens":1537}},"tokens_in":538,"tokens_out":1605,"duration_ms":13838,"temperature":1.0,"reasoning_tokens":1537,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:59:32.347559+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate the paper's closed-loop unstable plant P1 with a colored disturbance concentrated in the passband of the SVF filter, then compute the mean squared parameter error for $h = 1$ ms, $0.1$ ms, and $0.01$ ms; if the log-log slope departs from $-1$ or flattens, the deterministic-regressor assumption is the bottleneck.","supporting_citations":[{"cited_title":"Ljung, System Identification: Theory for the User , ser","cited_arxiv_id":null,"evidence_quote":"Establishes the 'ten times the bandwidth' sampling rule and the discrete-time pole-clustering problem the paper argues against."},{"cited_title":"On the choice of sampling rates in parametric identification of time series,","cited_arxiv_id":null,"evidence_quote":"Supplies the early rationale behind the classical sampling-rate guideline."},{"cited_title":"Improved finite word length charac- teristics in digital control using delta operators,","cited_arxiv_id":null,"evidence_quote":"Documents numerical deterioration of discrete-time models under fast sampling via the delta operator."},{"cited_title":"Garnier and L","cited_arxiv_id":null,"evidence_quote":"Provides the continuous-time identification setting and sampled-data machinery the SVF method builds on."},{"cited_title":"Direct closed-loop identification of continuous- time systems using fixed-pole observer model,","cited_arxiv_id":null,"evidence_quote":"Shows the fixed-pole observer method, equivalent to SVF, can identify unstable plants in closed loop."},{"cited_title":"Frequency domain uncertainty and the graph topology,","cited_arxiv_id":null,"evidence_quote":"Defines the ν-gap metric used to compare identified models of unstable plants."}],"review_version":1}