{"id":"f67c67d8-231e-4499-be21-a6da4b55970d","arxiv_id":"2509.06534","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"A closed-form sensitivity bound is proposed for how estimation error in parameterized linear systems responds to parameter perturbations, extending prior model-reduction robustness results to systems with control inputs.","lead":"The paper proposes a robustness metric for data-driven estimation of parameterized linear systems, measuring how estimation error changes as model parameters vary, and derives upper bounds on this sensitivity. If valid, it would let engineers compare estimation methods by their parametric sensitivity without simulating trajectories.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1's bound is dimensionally inconsistent: the LHS is a squared L2 norm (quadratic in input amplitude) while the RHS's forced terms are linear in ||B̄u||_∞, so the inequality fails for large inputs.","rationale":"The reader's verdict is REJECT, and the scaling argument in their rationale is decisive. I independently re-derived the structure of the proof. Equation (14) is the standard convolution expression for ȳ. In the proof, 'the square of 2-norm of ∂ȳ/∂θ_i' is written as ∫_0^∞ ... dt, so the LHS is the squared L2 norm. The derivative ∂ȳ/∂θ_i satisfies a linear ODE driven by (∂Ā/∂θ_i)x̄ plus a term from initial-condition sensitivity; since x̄ itself is affine in u, ∂ȳ/∂θ_i is affine in u. Consequently the LHS contains a term ∫||L(u)(t)||^2 dt, homogeneous of degree 2 in u. The proposed bound has K1 (constant), K2 and K3 terms linear in ||B̄u||_∞; no term is quadratic in the input amplitude. Thus the inequality cannot be a valid universal bound. The proof's a1b2 computation is where this manifests: it correctly identifies a1b2 as the squared L2 norm of the forced-response sensitivity, then bounds it by a single power of ||B̄u||_∞, losing the second power. Appendix A's a2b2 also ends up linear in ||B̄u||_∞ for a cross term that should be linear (one u factor), which is consistent, but the pure forced term is mishandled. The u=0 special case is genuinely unaffected, and Theorem 2 (initial-condition sensitivity) is a different statement. My concern differs from the reader's formal weakest_assumption (Lemma 2/3 uniformity); I find the scaling mismatch more fundamental, though the Lemma 2/3 issue about uniformity for e^{-Aτ} and finite horizons is also present. The concrete test with input scaling cleanly separates these and would settle the theorem.","tokens_in":11379,"tokens_out":9316,"duration_ms":97310,"concrete_test":"Set x(0)=0 in the Section V example (so only the forced response remains, Case 2). Take a constant input u(t)=c·1, compute the true sensitivity norm squared L(c)=∫_0^N ||∂ȳ/∂θ_1(t)||^2 dt by numerical integration/finite differences for c=1,10,100 (N fixed, e.g., N=5), and evaluate the Theorem 1 bound R(c)=K3 N^2 ||∂Ā/∂θ_1||^2 ||B̄u||_∞. Since L(c) grows as c^2 while R(c) grows as c, for large c the inequality L(c)≤R(c) will be violated; record the crossover c. This directly isolates the input-scaling inconsistency and avoids initial-condition terms.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central theorem's LHS is, per the proof (eq. following (14)), the squared L2 norm ∫_0^∞ ||∂ȳ/∂θ_i(t)||^2 dt. The map u↦∂ȳ/∂θ_i is affine (linear in the forced-response part plus a constant from initial conditions), so scaling u→c u scales the LHS by c^2 (plus lower-order terms). The RHS of (13), however, has K2 and K3 terms that are linear in ||B̄u||_∞, with K2, K3 independent of u. For any fixed parameter, initial state, and input direction with nonzero forced response, taking c large makes RHS/LHS → 0, so (13) cannot hold. The same scaling error appears in the a1b2 bound (15), where ∫||∫C∂e^{A(t-τ)}/∂θ B u dτ||^2 dt is majorized by a term linear in ||B̄u||_∞, and in the clubbed a1b1+a2b2 bound (17). The u=0 special case sidesteps this because all u-terms vanish, explaining why the unforced result can be correct. A secondary mismatch: the proof integrates to ∞ while K3 contains N^2 with no finite-horizon subscript on the LHS, but the scaling flaw is already decisive.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a sensitivity-based robustness metric for the estimation error of parameterized linear dynamical systems with control inputs. The central object is the augmented error system of Eqs. (5)-(7), and the robustness metric R in Eqs. (4)/(9) is defined through the normalized sensitivity of the estimation error. The main theoretical contribution is Theorem 1 (Eq. (13)), which claims a trajectory-free upper bound on the squared norm of the sensitivity ∂ȳ/∂θ_i in terms of ‖∂Ā/∂θ_i‖, the input norm ‖B̄u‖∞, and constants K1,K2,K3 involving the initial condition, the log-norm of Ā, C̄, and a horizon N. Theorem 2 handles parameter-dependent initial conditions. The paper also includes an LPV mass-spring-damper example and a comparison with a Gramian-based bound. The u=0 special case (Eq. (20)) reproduces a bound from the author's prior model-reduction work.","tokens_in":11704,"tokens_out":11887,"duration_ms":129674,"significance":"If Theorem 1 were correct, it would give a useful, easily computable certificate of parametric robustness for data-driven estimators, avoiding repeated simulation. A notable strength is the explicit claim of a closed-form bound and the availability of code. The u=0 reduction to existing results provides a partial consistency check. However, the forced-input extension, which is the paper's main novelty, is unsound: the claimed inequality has the wrong scaling in input amplitude, and the quantity bounded in the proof is an L2 norm that is not connected to the pointwise robustness metric R defined in Section II. Since these are foundational to the paper's central claim, the contribution is not established.","major_comments":[{"comment":"The LHS of Eq. (13) is defined in the proof as ∫_0^∞ ‖∂ȳ/∂θ_i(t)‖^2 dt, i.e., the squared L2 norm of the sensitivity. Differentiating Eq. (14) shows u ↦ ∂ȳ/∂θ_i is affine, with forced-response part linear in u. Replacing u by c u scales the forced part of the LHS by c^2, while the RHS's K2 and K3 terms scale only by c. For any input direction with nonzero forced sensitivity, the RHS/LHS ratio tends to zero as c→∞, so Eq. (13) cannot hold. The same defect appears in Eq. (15), where a1b2 is a squared norm but is bounded by a term linear in ‖B̄u‖∞. This is not a question of conservatism; it is a scaling contradiction.","section":"Section IV.A, Theorem 1, Eq. (13)"},{"comment":"The metric R in Eqs. (4) and (9) is defined pointwise in time through the ratio ‖∂err/∂θ_i(t)‖ / ‖err(t)‖. Theorem 1 instead bounds the time-integrated quantity ∫_0^∞ ‖∂ȳ/∂θ_i(t)‖^2 dt. The paper does not prove that this L2 bound controls the pointwise ratios appearing in R, nor does it define the norm in Eq. (13) consistently with the pointwise norm used in the metric. Thus, even setting aside the scaling issue, the theorem does not establish the advertised robustness metric.","section":"Section II.B and Section IV.A"},{"comment":"The LHS of Eq. (13) is an integral over [0,∞), while K3 contains N^2, where N is described as the maximum time for which inputs are applied. If a finite horizon was intended, the theorem should state the LHS as ‖∂ȳ/∂θ_i‖_{2,[0,N]}^2 and all intermediate bounds (e.g., Eq. (15)) should be derived on [0,N]; if an infinite horizon was intended, the N^2 term is undefined. The proof of Eq. (15) integrates to ∞ and then introduces N ad hoc, so the theorem's time horizon is not well defined.","section":"Section IV.A, Eqs. (13)-(15)"},{"comment":"The derivation of the a2b2 term replaces the inner integral ∫_0^t C̄ ∂e^{Ā(t-τ)}/∂θ_i B̄u(τ)dτ by ∫_0^t ∂e^{-Āτ}/∂θ_i B̄u(τ)dτ. These are not equal: ∂e^{A(t-τ)}/∂θ_i is not the derivative of e^{-Aτ}. The subsequent use of Lemma 3 is therefore applied to a different integral, and the K2 contribution in Eq. (17) is not established. A demonstration that the replacement is a valid conservative bound is needed.","section":"Appendix A, a2b2 bound"}],"minor_comments":[{"comment":"The typesetting of d_R is garbled; the placement of absolute-value bars and the division structure are unclear. Please rewrite the definition with clear parentheses.","section":"Section II.B, Eq. (3)"},{"comment":"The statement that D=0 is 'without loss of generality' needs a concrete explanation of how the D term is incorporated. As written, Theorems 1 and 2 assume D=0, but the augmented system definition (6) includes D̄.","section":"Section II.A"},{"comment":"Lemma 1 states a bound for e^{At} without specifying the time domain. Lemma 3 applies the same type of estimate to negative-time arguments; a two-sided log-norm bound and its domain of validity should be stated explicitly.","section":"Section III, Lemma 1"},{"comment":"The figures lack axis labels and units. The statement that the proposed bound is 'closer to true' would be easier to assess if the ground-truth norm, the baseline Gramian norm, and the time horizon were clearly specified.","section":"Section V, Figures 1-2"}],"recommendation":"reject","confidential_remarks":"The forced-input extension is the paper's main claimed novelty, but the central inequality has an elementary scaling flaw: the proved LHS is quadratic in input amplitude while the RHS is linear. This is not a borderline issue of conservatism, and a revision would require redefining the metric, the theorem, and the proof. The paper also connects a time-integrated bound to a pointwise robustness metric without a supporting argument. In my view the manuscript is not suitable for publication without a fundamentally different formulation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper up front. One: the unforced (u=0) case is sound and matches the author's own earlier model-reduction sensitivity bound, which is a good sanity check. Two: the forced-input theorem, which is the advertised contribution, cannot be true as stated. The LHS is the squared L2 norm of the output sensitivity; scaling the input u by c scales the forced response linearly, so the LHS grows like c^2. The RHS has terms linear in ||B̄u||_∞, with constants independent of u, so for large c the inequality fails. That is decisive.\n\nWhat is actually new? The extension of the sensitivity metric from model reduction to data-driven estimation with control inputs, plus the augmented-system formulation and the LPV example with code. The metric itself is sensible, and the comparison against a Gramian-based baseline is a reasonable idea. The u=0 case being a clean reduction to prior work is a legitimate validation of that special case.\n\nThe soft spots beyond the scaling flaw are several. In the proof, the cross-term a1b2 is written as a product of two different convolution integrals and then declared equal to the squared norm of one of them—that step hides the quadratic input dependence. The proof also integrates the LHS to infinity while the bound introduces a finite horizon N in K3; those two things are never reconciled. Lemma 2 and 3 are individually fine, but applying them inside convolution integrals over an infinite horizon produces terms like ∫ t e^{|µ|t} dt that diverge unless you already assumed a finite horizon, which the theorem does not. So the issue is not just a loose constant; the derivation does not support the stated inequality.\n\nWho gets value from this? Someone building parameter-specific robustness tools for LPV or grey-box estimation might find the augmented-system setup and the unforced bound useful. But the main result as written is not usable, and the paper would need substantial rework—either restrict the claims to the unforced case or fix the forced-input analysis with the correct scaling and a genuinely finite-horizon bound.\n\nFor peer review: I would send it to a referee rather than desk-reject, because the flaw is instructive and the framework is salvageable, but I would expect the referee to come back with the same scaling objection. It is not a publishable paper in its current form.","headline":"The paper has a clean framework and a correct unforced special case, but the headline forced-input bound fails on a simple input-scaling argument and the proof mixes norms and horizons.","tokens_in":12139,"tokens_out":4068,"would_cite":false,"duration_ms":46081,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93B35","93C05","93B30"],"pacs":[],"model":"deepseek-v4-flash","headline":"The sensitivity of estimation error to each parameter is bounded by an explicit formula, so robustness can be checked without simulation.","keywords":["parametric robustness","sensitivity analysis","data-driven estimation","linear dynamical systems","parameter uncertainty","estimation error","matrix exponential derivative bound","augmented system"],"falsifier":"Pick a Hurwitz but poorly conditioned A(θ), compute ||∂e^{At}/∂θ|| numerically by finite differences over a grid of t up to horizon N, and check whether it ever exceeds ||∂A/∂θ|| t e^{-|mu|t}. Alternatively, simulate the mass-spring-damper LPV example with the paper's parameters, compute ∂ȳ/∂θ₁ by finite differences, and compare its squared 2-norm to the Theorem 1 bound; any violation of the bound at a single t would refute the central claim.","tokens_in":11254,"feed_emoji":"⚙️","tokens_out":7520,"duration_ms":78390,"temperature":0.7,"pith_summary":"This paper tries to establish a practical guarantee: for a parametrized linear dynamical system with control inputs, the sensitivity of the estimation error to each parameter can be bounded by a closed-form expression, so an engineer can judge how robust an estimator is without simulating trajectories. The bound is stated as Theorem 1 and separates three sources of sensitivity — initial conditions, forced inputs, and their interaction — into three terms with explicit constants K1, K2, K3 that depend on the augmented system, its log-norm decay rate, and the time horizon. If the bound is correct, it gives data-driven estimation a parameter-specific robustness metric: rather than asking whether an estimator works on average, one can ask which parameters the estimate is fragile to. The paper also handles parameter-dependent initial conditions via a Lyapunov/observability-Gramian bound (Theorem 2), and checks the unforced-input special case against a known model-reduction robustness result.","feed_headline":"Parametric error sensitivity bounded without simulation","feed_subtitle":"Ranks candidate estimators by parameter robustness, no trajectory runs.","key_machinery":"The load-bearing object is the augmented error system, x̄ = [x; x̃], Ā = diag(A, Ã), B̄ = [B; B̃], C̄ = C − C̃, whose output ȳ is exactly the estimation error. The proof machinery is a pair of derivative bounds for matrix exponentials: Lemma 2 bounds ‖∂e^{At}/∂θ_i‖ by ‖∂A/∂θ_i‖ t e^{-|µ|t} for Hurwitz A, and Lemma 3 bounds the negative-time exponential e^{-As} similarly. These turn the sensitivity integral into moments of t against decaying exponentials, which produce the explicit constants 1/(4|µ|³), 2/|µ|⁵, and 2/|µ|, and hence K1, K2, K3. The log-norm µ = λmax(Ā+Āᵀ)/2 encodes the stability margin, so the constants reflect how quickly the augmented error system forgets both initial conditi","core_discovery":"The central claim is Theorem 1: for the augmented error system ẋ̄ = Āx̄ + B̄u, ȳ = C̄x̄ with Ā block diagonal containing A and the estimated Ã, the squared 2-norm of the sensitivity ∂ȳ/∂θ_i is bounded by K1 ‖∂Ā/∂θ_i‖² + K2 ‖∂Ā/∂θ_i‖³ ‖B̄u‖∞ + K3 N² ‖∂Ā/∂θ_i‖² ‖B̄u‖∞, with K1, K2, K3 given explicitly in terms of the augmented output matrix, initial condition, log-norm, and horizon. The bound is assembled from the convolution solution of the augmented system; the key move is to pass derivatives through the matrix exponential and apply Lemmas 2 and 3, which bound ‖∂e^{At}/∂θ_i‖ ≤ ‖∂A/∂θ_i‖ t e^{-|µ|t} (and its negative-time analogue) for Hurwitz A. Each of the four products from expanding the s","pith_inferences":["The proof is written for parameter dependence in A and x(0); the same convolution expansion with ∂B̄/∂θ_i and ∂C̄/∂θ_i terms would extend Theorem 1 to fully parameter-dependent B and C, an extension the paper leaves implicit.","Because K1 and K2 scale as 1/|µ|³ and 1/|µ|⁵, near-marginal systems will make the bound conservative; a finite-horizon analogue might trade tightness for practical usability.","One can test the metric's utility directly: on a benchmark parameterized system, compare the Theorem 1 sensitivity ranking across estimators with rankings from finite-difference ground truth; the paper motivates but does not run that benchmark.","Reading R as a cost function opens a design loop — choose estimated parameters to maximize R with respect to nuisance parameters while keeping sensitivity to parameters of interest high; the paper names this direction as future work."],"forward_implications":["Estimator selection can be made parameter-aware: for each θ_i, compute the Theorem 1 bound on the augmented system and rank candidate estimators on sensitivity to that parameter.","The robustness metric R can be restricted to a subset of physically relevant parameters, isolating, say, motor constants or aerodynamic coefficients in a rotor system while ignoring others.","The unforced-input special case reproduces the known model-reduction robustness bound, and the zero-initial-condition case gives K3N²‖∂Ā/∂θ_i‖²‖B̄u‖∞; both validate the theorem as an extension rather than a separate result.","For parameter-dependent initial conditions, Theorem 2 gives a bound through the maximum eigenvalue of the observability Gramian P, again computable without trajectory simulation.","The LPV mass-spring-damper example indicates the new bound is tighter than a finite-horizon observability-Gramian bound, making it usable as a worst-case guarantee rather than merely a formal one."],"supporting_citations":[{"why":"Supplies the classical bound ||e^{At}|| ≤ e^{-|mu|t} for Hurwitz A that Lemmas 2 and 3 use to get exponential decay in the derivative bounds.","marker":"[22]"},{"why":"Defines the logarithm norm mu(A)=lambda_max(A+A^T)/2, the stability-margin quantity appearing in K1, K2, K3.","marker":"[23]"},{"why":"Provides the convolution formula for forced linear responses, the starting point of the Theorem 1 proof.","marker":"[24]"},{"why":"Introduces the robustness-distance formulation and the Lyapunov/Gramian proof idea reused for Theorem 2.","marker":"[21]"},{"why":"Gives the unforced-input special-case bound that Theorem 1 reduces to, used to validate the new metric.","marker":"[26]"},{"why":"Characterizes the observability Gramian via the Lyapunov equation, used in Theorem 2 and the finite-horizon baseline.","marker":"[25]"}],"fun_headline_variants":["New metric ranks estimators by parameter robustness","Explicit bound on parameter sensitivity in linear system estimation","Analytic bound ties parameter error to control input and dynamics","Theorem quantifies estimation robustness without simulation runs","Parameter uncertainty bound guides estimator selection"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is that Lemma 2's uniform bound ||∂e^{At}/∂θ_i|| ≤ ||∂A/∂θ_i|| t e^{-|mu|t} holds for every matrix argument and every time in the convolution integrals, including negative-time exponentials; the K1, K2, K3 constants come from integrating that decay, so a counterexample to the bound would break the formula.","fun_headline_variants_meta":{"raw":{"variants":["New metric ranks estimators by parameter robustness","Explicit bound on parameter sensitivity in linear system estimation","Analytic bound ties parameter error to control input and dynamics","Theorem quantifies estimation robustness without simulation runs","Parameter uncertainty bound guides estimator selection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000207,"raw_usage":{"total_tokens":1225,"prompt_tokens":724,"completion_tokens":501,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":468,"completion_tokens_details":{"reasoning_tokens":433}},"tokens_in":468,"tokens_out":501,"duration_ms":6392,"temperature":1.0,"reasoning_tokens":433,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T23:29:23.820817+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Pick a Hurwitz but poorly conditioned A(θ), compute ||∂e^{At}/∂θ|| numerically by finite differences over a grid of t up to horizon N, and check whether it ever exceeds ||∂A/∂θ|| t e^{-|mu|t}. Alternatively, simulate the mass-spring-damper LPV example with the paper's parameters, compute ∂ȳ/∂θ₁ by finite differences, and compare its squared 2-norm to the Theorem 1 bound; any violation of the bound at a single t would refute the central claim.","supporting_citations":[{"cited_title":"G. dahlquist, stability and error bounds in the numerical integration of ordinary differential equations. 85 s. stockholm 1959. k. tekniska h ¨ogskolans handlingar,","cited_arxiv_id":null,"evidence_quote":"Supplies the classical bound ||e^{At}|| ≤ e^{-|mu|t} for Hurwitz A that Lemmas 2 and 3 use to get exponential decay in the derivative bounds."},{"cited_title":"On logarithmic norms,","cited_arxiv_id":null,"evidence_quote":"Defines the logarithm norm mu(A)=lambda_max(A+A^T)/2, the stability-margin quantity appearing in K1, K2, K3."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the convolution formula for forced linear responses, the starting point of the Theorem 1 proof."},{"cited_title":"Robustness guarantees for structured model reduction of dynamical systems with applications to biomolec- ular models,","cited_arxiv_id":null,"evidence_quote":"Introduces the robustness-distance formulation and the Lyapunov/Gramian proof idea reused for Theorem 2."},{"cited_title":"Robustness guarantees for structured model reduction of dynamical systems,","cited_arxiv_id":null,"evidence_quote":"Gives the unforced-input special-case bound that Theorem 1 reduces to, used to validate the new metric."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Characterizes the observability Gramian via the Lyapunov equation, used in Theorem 2 and the finite-horizon baseline."}],"review_version":1}