{"id":"fe4af81d-e089-4798-a840-c16ef82c0198","arxiv_id":"2608.08318","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":8.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"For MIMO LTI plants, a backward-difference PID implementation is stable at all sufficiently small sampling periods exactly when the ideal loop is stable and the derivative channel matrix CBK_D is Schur, away from boundary spectra.","lead":"Backward-difference derivative action in a digital MIMO PID controller can make a stable continuous-time design unstable no matter how fast the loop samples. The paper proves an exact condition: the ideal PID loop must be stable and the matrix formed from the plant output-input path and derivative gain must have all eigenvalues inside the unit circle.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified.","rationale":"The reader's weakest_assumption pointed to the boundary-spectra exclusions and the E-nonsingularity condition. I agree these are the only places where the theorem's hypotheses could fail, but they are explicit and do not create a load-bearing concern for the conditional equivalence stated in Theorem 2(3). The sufficiency direction requires no boundary exclusions beyond what Hurwitz and Schur already imply, and the necessity direction is valid because any violation of the two conditions yields an eigenvalue branch that exits the unit disk for all sufficiently small h. I independently re-derived the key algebraic steps, including the lifted matrix, the characteristic determinant for Proposition 1, the semisimple bases V and W, the first-order reduced matrix, and the boundary drift formula in Proposition 2; all matched the paper's expressions. Hence no correctness risk is identified that would change the reader's ACCEPT verdict.","tokens_in":9646,"tokens_out":20010,"duration_ms":180796,"concrete_test":"Use a symbolic algebra system to recompute W A1 V from the paper's definitions for a random instance (e.g., n=3, p=m=2); verify that it equals Aid exactly. Then numerically confirm that, when MD has an eigenvalue with |nu|>1, the lifted matrix Ah has an eigenvalue with modulus greater than 1 for h in {1e-3, 1e-4, 1e-5}, matching the fast-mode instability direction.","verdict_should_be":"UNCHANGED","load_bearing_attack":"No significant objection identified. The central claim (Theorem 2(3)) was checked step-by-step: the lifted matrix (8) is derived correctly from the controller updates and zero-order hold plant solution; Proposition 1's spectrum computation via Sylvester's identity is correct; the cluster at 1 is semisimple with first-order reduction W A1 V = Aid; and the fast limits -spec(MD) follow from the block structure of A0. The necessity direction is sound because, under the stated spectral exclusions, a non-Hurwitz Aid or a non-Schur MD produces an eigenvalue branch outside the unit disk for all sufficiently small h. The only caveats, E nonsingular and the exclusion of boundary spectra, are explicitly stated in the theorem and partially treated in Proposition 2; they do not undermine the conditional iff. I find no internally inconsistent step or hidden assumption that would threaten the headline result.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This note studies the fast-sampling limit of a sampled-data MIMO PID controller in which the derivative term is realized by a backward difference. The authors derive the exact lifted discrete-time model (8) and show that its state matrix has n+p eigenvalues near 1 with first-order drift given by the ideal continuous-time PID matrix A_id, plus p fast eigenvalues converging to the negated spectrum of M_D = C B K_D (Proposition 1, Theorem 1). The main result, Theorem 2, states that away from boundary spectra the sampled loop is Schur stable for all sufficiently small h if and only if A_id is Hurwitz and M_D is Schur. Minimal counterexamples illustrate the obstruction, a boundary-drift formula is provided for simple fast eigenvalues on the unit circle, and the relative-degree case C B = 0 is shown to be obstruction-free.","tokens_in":9696,"tokens_out":16611,"duration_ms":142943,"significance":"The result is a precise and useful caution for sampled PID implementation: a backward-difference realization can be unstable for arbitrarily small h even when the ideal continuous-time loop is exponentially stable, unless the simple matrix condition rho(C B K_D) < 1 holds. The paper's method is exact lifting followed by standard perturbation theory, and the appendix proofs are self-contained and checkable. The main theorem is honestly stated with explicit spectral exclusions, and Proposition 2 partially resolves the simple unit-circle boundary case. The scalar and MIMO examples are minimal and clearly isolate the mechanism. The contribution is appropriately scoped: Remark 5 correctly notes that the condition is realization-specific, not a universal property of all digital PID implementations.","major_comments":[],"minor_comments":[{"comment":"The author block contains 'Y . Wu' with a space before the period; this should read 'Y. Wu'.","section":"Authors and affiliations"},{"comment":"Several expressions involving h are typeset ambiguously, for example 'h−1Γh' should be read and typeset as h^{-1} Γ_h; please ensure all such expressions use clear superscript notation.","section":"Section II-B, Eq. (8)"},{"comment":"In Eq. (13) and the surrounding text, the terms written as 'h2kiλ' should be typeset as h^2 k_i λ; the missing superscripts appear repeatedly and should be corrected.","section":"Section IV-C, Eq. (13)"},{"comment":"The notation 'q, ℓ∈ Cp' should read 'q, ℓ \\in C^p', and similarly 'ℓ∗' should be typeset as ℓ^* for clarity.","section":"Proposition 2"},{"comment":"In the proof of Proposition 1, the Schur-complement calculation is performed for λ \\neq 0 and then extended by polynomial continuation; this is correct, but a sentence explicitly noting that the extension is justified by continuity of the determinant would improve readability.","section":"Appendix A"}],"recommendation":"minor_revision","confidential_remarks":"For the editor: the mathematical core is sound and the paper is a well-executed short note. The required changes are editorial (typos and typesetting) rather than substantive. I would be happy to see the paper accepted after a minor revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: the paper is correct and worth publishing. It proves that the standard MIMO backward-difference PID realization has fast sampled eigenvalues converging to -spec(CBK_D), so ideal-loop stability plus Schur stability of CBK_D is necessary and sufficient for small-h stability, away from boundary spectra. The mechanism is real: the previous output sample is controller memory, and it does not disappear as h -> 0. This also explains why relative-degree-two treatments with CB=0 miss the obstruction, since then the fast spectrum collapses to zero. That is a genuine new result relative to the cited artificial-delay/LMI literature.\n\nStrengths. I checked the lifted matrix (8), the spectral factorization in Proposition 1, the semisimple reduction W A1 V = Aid, and the boundary drift formula (12). They are consistent. The MIMO counterexample with D0 is well chosen because it shows the condition is spectral, not entrywise. Proposition 2 is a nice extra: it gives first-order drift for a simple fast limit on the unit circle and honestly notes that second-order terms decide when the drift vanishes. The paper does not oversell the criterion; Remark 5 correctly says it is realization-specific.\n\nSoft spots are mostly at the boundaries, and the paper labels them. The iff in Theorem 2(3) excludes imaginary-axis Aid eigenvalues and unit-circle MD eigenvalues. Proposition 2 handles only simple unit-circle fast limits; repeated eigenvalues or vanishing first-order drift are left open. That is a limitation, but it is explicit and does not weaken the interior result. E nonsingular is a standard well-posedness condition for derivative feedback. The only thing I would push on is the novelty claim: 'first such characterization' is credible for MIMO, but the authors should do a more careful SISO prior-art search, since scalar backward-difference integrator examples may be folklore. That is a revision request, not a substantive flaw. Citations are appropriate; no self-citation is load-bearing.\n\nWho benefits: sampled-data and PID control researchers, especially people who teach backward-difference derivative action as a harmless implementation. It deserves a serious referee and likely acceptance after a modest literature pass. I would cite it and would bring it to reading group.","headline":"The exact fast-mode criterion rho(CBK_D)<1 for MIMO backward-difference PID is correct, clearly proved, and worth publishing, with only its explicitly stated boundary exclusions limiting the otherwise clean iff.","tokens_in":10290,"tokens_out":2542,"would_cite":true,"duration_ms":24303,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93C05","93C35","93C57","93D20"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proves that the standard backward-difference PID derivative can destabilize a stable multivariable loop at arbitrarily fast sampling, and gives an exact spectral test for when it does.","keywords":["derivative feedback","fast modes","finite differences","multivariable PID control","sampled-data control","singular perturbation","backward difference","spectral criterion"],"falsifier":"If the claimed instability criterion were false, the scalar integrator $\\dot x=u$, $y=x$ with $k_p,k_i>0$ and $k_d>1$ would be a counterexample: the ideal matrix $A_{\\rm id}$ is Hurwitz while $|k_d|>1$, and direct computation of the eigenvalues of the lifted matrix $A_h$ for $h=0.1,0.01,0.001$ gives $\\rho(A_h)>1$ in every case, confirming the prediction that no sufficiently small sampling period restores stability.","tokens_in":9386,"feed_emoji":"⚙️","tokens_out":12951,"duration_ms":109159,"temperature":0.7,"pith_summary":"The paper asks whether the most common digital way to implement derivative action—replacing $\\dot y(t)$ by the backward difference $(y_k-y_{k-1})/h$—keeps a well-designed continuous-time MIMO PID loop stable as the sampling period $h$ shrinks. The answer it proves is 'not necessarily': the previous output sample becomes a controller state, and in the exact lifted discrete model it creates fast eigenvalues that converge to $-\\operatorname{spec}(CBK_D)$ rather than to the continuous-time spectrum. Consequently, if any eigenvalue of $CBK_D$ lies outside the unit disk, the sampled loop is unstable for every sufficiently small $h$, no matter how stable the ideal loop is. Away from boundary spectra the criterion is exact: stability for all small sampling periods holds if and only if the ideal closed-loop matrix $A_{\\rm id}$ is Hurwitz and $CBK_D$ is Schur. This matters because it turns a practical rule of thumb—'sample faster to get closer to continuous time'—into a checkable spectral condition on the implementation itself.","feed_headline":"Backward-difference PID can be unstable at any fast sampling rate","feed_subtitle":"The stored previous output sample creates modes fast sampling cannot move; a simple spectral test decides when.","key_machinery":"The engine is the exact lifted discrete-time matrix $A_h$ from (8), whose state includes the stored previous output sample $r_k=y_{k-1}$. As $h\\to0^+$, $A_h$ tends to $A_0$ with spectrum $\\{1\\}^{n+p}\\cup(-\\operatorname{spec}(M_D))$, where $M_D=CBK_D$; the factor $h^{-1}\\Gamma_h\\to I_n$ keeps the derivative-memory block alive. A first-order perturbation reduction, applied to the semisimple cluster at 1, shows the slow eigenvalues move as $1+h\\mu+o(h)$ with $\\mu\\in\\operatorname{spec}(A_{\\rm id})$, while the simple-eigenvalue perturbation formula gives the boundary drift of a fast eigenvalue at $-\\nu$, $|\\nu|=1$, in Proposition 2. This split turns a sampled-data problem into two finite-dimensional spectral checks: Hurwitz stability of $A_{\\rm id}$ and Schur stability of $M_D$.","core_discovery":"For the plant $\\dot x=Ax+Bu$, $y=Cx$ with the output PID law $u=-K_P y-K_I z-K_D \\dot y$, $\\dot z=y$, the paper's central discovery is a spectral split of the sampled closed-loop dynamics. With the exact lifted transition matrix $A_h$ over one sampling interval, $n+p$ eigenvalues track the ideal loop, $\\lambda_j(h)=1+h\\mu_j+o(h)$ with $\\mu_j\\in\\operatorname{spec}(A_{\\rm id})$, while the remaining $p$ eigenvalues converge to $-\\operatorname{spec}(CBK_D)$. Under $E=I_n+BK_DC$ nonsingular, no imaginary-axis eigenvalue of $A_{\\rm id}$, and no unit-circle eigenvalue of $CBK_D$, Theorem 2 states that $A_h$ is Schur for all sufficiently small $h$ if and only if $A_{\\rm id}$ is Hurwitz and $\\rho(CBK_D)<1$. If $A_{\\rm id}$ has a positive-real-part eigenvalue or $CBK_D$ has a modulus-greater-than-one eigenvalue, the sampled loop is unstable for all small $h$; the relative-degree case $CB=0$ makes the extra condition automatic.","pith_inferences":["Extension: For alternative discrete differentiators—forward differences, filtered derivatives, or predictor-based estimators—the same lifted-state construction would likely replace $M_D=CBK_D$ by a realization-dependent fast matrix, so an analogous spectral condition should be re-derived for each implementation rather than inherited.","Extension: The criterion opens a design route the paper does not pursue: when $CB\\neq0$, choose $K_D$ so that $CBK_D$ is Schur, turning the derivative gain into a tuning knob for the sampling-induced fast mode.","Extension: For nonlinear or time-varying plants there is no fixed matrix $CBK_D$, so the fast-sampling limit may show a similar but state-dependent obstruction; a testable prediction is that very small-$h$ simulations of a plant with direct feedthrough in $\\dot y$ would exhibit persistent high-frequency oscillations even as $h\\to0$."],"forward_implications":["Any PID design for a plant with $CB\\neq0$ must verify $\\rho(CBK_D)<1$; exponential stability of the ideal continuous-time law alone is not enough to certify the backward-difference implementation.","Decreasing $h$ improves the slow approximation—$\\rho(A_h)=1+h\\alpha_{\\rm id}+o(h)$ when $M_D$ is Schur—but cannot move the fast eigenvalues, so smaller sampling periods cannot rescue an unstable derivative channel.","For plants with relative degree at least two ($CB=0$), the obstruction vanishes: $M_D=0$ and every Hurwitz ideal loop yields a stable fast-sampled backward-difference PID.","In the SISO case the fast-sampling check reduces to $|CB\\,k_D|<1$, whereas ideal well-posedness only requires $1+CB\\,k_D\\neq0$, so the sampled constraint is genuinely stronger.","When a fast limit sits on the unit circle ($|\\nu|=1$), Proposition 2 gives the first-order drift that decides inward or outward crossing, and both outcomes can occur."],"supporting_citations":[{"why":"Supplies the standard well-posedness condition for derivative feedback: $E=I_n+BK_DC$ (equivalently $\\det(I_p+CBK_D)\\neq0$) must hold for the ideal loop to be an ODE.","marker":"[5]"},{"why":"Represents the artificial-delay/LMI results that guarantee fast-sampled stability for relative-degree-at-least-two systems, the benchmark the paper contrasts with its $CB=0$ Proposition 3.","marker":"[7]"},{"why":"Provides the first-order perturbation formula for a semisimple eigenvalue cluster, used to prove the slow eigenvalues move as $1+h\\mu_j+o(h)$ with $\\mu_j\\in\\operatorname{spec}(A_{\\rm id})$.","marker":"[11]"},{"why":"Gives the relative-degree-2 sampled PID analysis in output-derivative coordinates, which the paper's lift identifies as the $CB=0$ special case.","marker":"[16]"},{"why":"Supplies the simple-eigenvalue perturbation formula used in Proposition 2 to compute the first-order drift of a fast eigenvalue whose limit lies on the unit circle.","marker":"[17]"}],"fun_headline_variants":["Backward-difference PID unstable at every sampling rate","Fast sampling cannot tame backward-difference PID modes","PID with backward differences: instability persists at tiny h","MIMO PID: the stored sample that breaks fast sampling","Exact spectral criterion for backward-difference PID instability"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The exact if-and-only-if criterion holds only when the ideal derivative feedback is well posed (a certain matrix is nonsingular) and when neither the ideal loop nor the derivative-memory matrix has a borderline eigenvalue; at the boundary, higher-order terms decide stability, so the simple criterion is silent.","fun_headline_variants_meta":{"raw":{"variants":["Backward-difference PID unstable at every sampling rate","Fast sampling cannot tame backward-difference PID modes","PID with backward differences: instability persists at tiny h","MIMO PID: the stored sample that breaks fast sampling","Exact spectral criterion for backward-difference PID instability"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000921,"raw_usage":{"total_tokens":3912,"prompt_tokens":867,"completion_tokens":3045,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":483,"completion_tokens_details":{"reasoning_tokens":2969}},"tokens_in":483,"tokens_out":3045,"duration_ms":22997,"temperature":1.0,"reasoning_tokens":2969,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T00:09:36.905433+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"If the claimed instability criterion were false, the scalar integrator $\\dot x=u$, $y=x$ with $k_p,k_i>0$ and $k_d>1$ would be a counterexample: the ideal matrix $A_{\\rm id}$ is Hurwitz while $|k_d|>1$, and direct computation of the eigenvalues of the lifted matrix $A_h$ for $h=0.1,0.01,0.001$ gives $\\rho(A_h)>1$ in every case, confirming the prediction that no sufficiently small sampling period restores stability.","supporting_citations":[{"cited_title":"Quasi feedback forms for differential-algebraic systems,","cited_arxiv_id":null,"evidence_quote":"Supplies the standard well-posedness condition for derivative feedback: $E=I_n+BK_DC$ (equivalently $\\det(I_p+CBK_D)\\neq0$) must hold for the ideal loop to be an ODE."},{"cited_title":"Stabilization by using artificial delays: An LMI approach,","cited_arxiv_id":null,"evidence_quote":"Represents the artificial-delay/LMI results that guarantee fast-sampled stability for relative-degree-at-least-two systems, the benchmark the paper contrasts with its $CB=0$ Proposition 3."},{"cited_title":"Kato, Perturbation Theory for Linear Operators , 2nd ed","cited_arxiv_id":null,"evidence_quote":"Provides the first-order perturbation formula for a semisimple eigenvalue cluster, used to prove the slow eigenvalues move as $1+h\\mu_j+o(h)$ with $\\mu_j\\in\\operatorname{spec}(A_{\\rm id})$."},{"cited_title":"Sampled-data implementation of derivative-dependent control using artificial delays,","cited_arxiv_id":null,"evidence_quote":"Gives the relative-degree-2 sampled PID analysis in output-derivative coordinates, which the paper's lift identifies as the $CB=0$ special case."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the simple-eigenvalue perturbation formula used in Proposition 2 to compute the first-order drift of a fast eigenvalue whose limit lies on the unit circle."}],"review_version":1}