betaFeedback

WorksEngineering

Robust Data-Driven Output Feedback for Systems With Unknown Input Delays

Abstract

In a history-feedback realization, an input delay increases memory without increasing the number of unknown physical coefficients. We use this structure to design output-feedback controllers for discrete-time systems with one input, one output, and known undelayed order, using noisy input-output data. For a known delay, we derive a necessary and sufficient linear matrix inequality for common-quadratic stabilization of all models consistent with the data. Optimizing its equation-error bound gives a robustness margin for the measured record. For an unknown constant delay in a known interval, we characterize noiseless delay ambiguity and give a sufficient condition for one controller to stabilize all models for every delay consistent with the error bound. Under stated excitation and nondegeneracy assumptions, this procedure is feasible for all sufficiently small equation errors. Numerical examples compare robustness at equal controller memory and show that certified stabilization can succeed while distinct delay models remain compatible with the same data.

Introduction

An input delay requires feedback to account for commands still on their way to the actuator. Predictor feedback and Lyapunov–Krasovskii methods address known-model delay problems [1], [2]. With an unknown plant, an incorrect delay can also be confused with a poor dynamic fit. An unrestricted higher-order fit discards a useful distinction: the delay changes which past inputs enter the output equation, not how many plant coefficients are unknown.

Data-dependent linear matrix inequalities (LMIs) provide stabilizing controllers without explicit model identification [3]. Data informativity studies guarantees over all data-consistent models [4]; matrix S-lemma and Petersen-lemma results give exact quadratic tests for specified bounded-error models under suitable regularity assumptions [5], [6]. Input–state delay methods use Lyapunov–Krasovskii conditions [7], while [8] combines unknown-delay stabilization with least-squares delay assessment. Output-based unknown-delay regulation in [9] uses continuous-time identification and observer design. Here we certify discrete-time stabilization over all delay descriptions retained by bounded-error data.

Output-feedback designs use measured histories [10] or autoregressive descriptions [11]. Li et al. [12] already exploit auxiliary-system structure for robust control from noisy input and output data; structural prior information is also treated in [13]. Here each delay candidate has 2n2n uncertain coefficients, with nn the undelayed order. We use this input-lag restriction to quantify fixed-record robustness and cover unresolved delays with one controller. Our primary bound concerns equation error; the sensor-noise extension gives a conservative outer enclosure.

The paper makes two contributions.

  1. For a known delay, we derive a necessary and sufficient LMI for common-quadratic stabilization of all data-consistent models (Theorem 1). Optimizing its equation-error budget gives a fixed-record robustness margin. Its centered information radius is homogeneous, monotone, and concave (Corollary 1). At equal memory and for the same record, the structured fit inherits every certificate for the unrestricted fit, so its margin is no smaller (Proposition 1).

  1. For an unknown constant delay in a known interval, we characterize noiseless ambiguity (Theorem 2) and give a sufficient condition for one controller to stabilize all retained families (Theorem 3). Under excitation and nondegeneracy assumptions, the complete procedure is feasible for all sufficiently small equation errors (Corollary 2). Unique delay identification [14] need not precede certified stabilization, as the three-delay example illustrates.

The margin holds the measured record fixed, unlike the experiment-wide tolerance in [15]. Supporting results explain the obstruction in an unrestricted overparameterized family (Proposition 2) and enclose sensor errors without a prior coefficient bound (Proposition 3).

We use col⁡\operatorname{col} for column stacking, ∗* for symmetric blocks, ∥⋅∥2\|\cdot\|_2 for the Euclidean or spectral norm, and eie_i for canonical basis vectors.

Plant, data, and controller memory

Consider the real, discrete-time plant

x(k+1)=Ax(k)+bu(k−d),y(k)=cTx(k),(1)x(k+1) = Ax(k) + bu(k-d), \qquad y(k) = c^\mathsf{T}x(k), \tag*{(1)}

where x∈Rnx \in\mathbb{R}^n, u,y∈Ru,y \in\mathbb{R}, and d≥0d \ge0 is an integer; u(−d),…,u(−1)u(-d),\ldots,u(-1) are initial conditions. The undelayed realization (A,b,cT)(A,b,c^\mathsf{T}) is minimal and its order nn is known. Define

D(z)=det⁡(zI−A)=zn+∑i=1nαizn−i,D(z) = \det(zI-A) = z^n + \sum_{i=1}^{n}\alpha_i z^{n-i},
N(z)=cTadj⁡(zI−A)b=∑i=1nβizn−i.(2)N(z) = c^\mathsf{T}\operatorname{adj}(zI-A)b = \sum_{i=1}^{n}\beta_i z^{n-i}. \tag*{(2)}

Minimality makes DD and NN coprime. The transfer function is z−dN(z)/D(z)z^{-d}N(z)/D(z), with relative degree r=min⁡{i:βi≠0}r=\min\{i:\beta_i\ne0\}. By Cayley–Hamilton, for arbitrary initial states and k≥nk\ge n,

y(k)=−∑i=1nαiy(k−i)+∑i=1nβiu(k−d−i).(3)y(k)=-\sum_{i=1}^{n}\alpha_i y(k-i)+\sum_{i=1}^{n}\beta_i u(k-d-i). \tag*{(3)}

For k≥n+dk\ge n+d only nonnegative input indices occur. Put θ=col⁡(α,β)∈R2n\theta=\operatorname{col}(\alpha,\beta)\in\mathbb{R}^{2n} and

φd(k)=col⁡(−y(k−1),…,−y(k−n),u(k−d−1),…,u(k−d−n)).(4)\varphi_d(k)=\operatorname{col}\left(-y(k-1),\ldots,-y(k-n),u(k-d-1),\ldots,u(k-d-n)\right). \tag*{(4)}

What the error bound means

At selected times K={k1,…,kT}\mathcal{K}=\{k_1,\ldots,k_T\}, let

Y=θTΦd+W,∥W∥22≤γ,(5)Y=\theta^{\mathsf T}\Phi_d+W,\qquad\lVert W\rVert_2^2\le\gamma, \tag*{(5)}

Here Y∈R1×TY\in\mathbb{R}^{1\times T} and Φd∈R2n×T\Phi_d\in\mathbb{R}^{2n\times T} collect the columns y(kj)y(k_j) and φd(kj)\varphi_d(k_j), respectively. Columns may be stacked from several experiments, with valid lag histories in each. The error is an equation error in (3). The bound γ\gamma must come from a valid disturbance bound; a small least-squares residual alone does not supply one. All synthesis statements concern the disturbance-free plant after data collection.

The compatible coefficient set is

Σd(γ)={ϑ:∥Y−ϑTΦd∥22≤γ}.(6)\Sigma_d(\gamma)=\{\vartheta:\lVert Y-\vartheta^{\mathsf T}\Phi_d\rVert_2^2\le\gamma\}. \tag*{(6)}

It deliberately allows nonminimal ARX models: retaining them gives a convex uncertainty description that contains the physical plant. If Φd\Phi_d has full row rank, let

Hd=ΦdΦdT,θ^d=Hd−1ΦdYT,ed2=∥Y−θ^dTΦd∥22.(7)H_d=\Phi_d\Phi_d^{\mathsf T},\qquad\hat{\theta}_d=H_d^{-1}\Phi_dY^{\mathsf T},\qquad e_d^2=\lVert Y-\hat{\theta}_d^{\mathsf T}\Phi_d\rVert_2^2. \tag*{(7)}

Least-squares orthogonality gives

Σd(γ)={θ^d+η:ηTHdη≤ρd},ρd=γ−ed2.(8)\Sigma_d(\gamma)=\{\hat{\theta}_d+\eta:\eta^{\mathsf T}H_d\eta\le\rho_d\},\qquad\rho_d=\gamma-e_d^2. \tag*{(8)}

Thus ed2>γe_d^2>\gamma rejects the candidate, while ρd≥0\rho_d\ge0 is the uncertainty budget left after fitting. Formula (7) is analytical; an SVD or QR factorization should be used numerically.

A measured history is the controller state

For p≥np\ge n and q≥n+dq\ge n+d, store

ζ(k)=col⁡(y(k−1),…,y(k−p),u(k−1),…,u(k−q))∈Rm.(9)\zeta(k)=\operatorname{col}\left(y(k-1),\ldots,y(k-p),u(k-1),\ldots,u(k-q)\right)\in\mathbb{R}^{m}. \tag*{(9)}

where m=p+qm=p+q. Let JℓJ_\ell have ones on its first subdiagonal and zeros elsewhere, and set A0=diag⁡(Jp,Jq)A_0=\operatorname{diag}(J_p,J_q), e=e1∈Rme=e_1\in\mathbb{R}^m, and B=ep+1∈RmB=e_{p+1}\in\mathbb{R}^m. Define Sd∈R2n×mS_d\in\mathbb{R}^{2n\times m} by Sdζ=φdS_d\zeta=\varphi_d: its first nn rows select the first nn output entries with a minus sign, and its last nn rows select input entries d+1,…,d+nd+1,\ldots,d+n. Consequently,

ζ(k+1)=(A0+eθTSd)ζ(k)+Bu(k).(10)\zeta(k+1)=(A_0+e\theta^{\mathsf T}S_d)\zeta(k)+Bu(k). \tag*{(10)}

The first row computes the next stored output; the other rows only shift measured values. We use u(k)=Kζ(k)u(k)=K\zeta(k), with

Ad,K(ϑ)=A0+eϑTSd+BK.(11)A_{d,K}(\vartheta)=A_0+e\vartheta^{\mathsf T}S_d+BK. \tag*{(11)}

The smallest memory used here is (p,q)=(n,n+d)(p,q)=(n,n+d). This is a property of this implementation, not a lower bound on the order of every possible controller.

For example, take n=2n=2, d=1d=1 and (p,q)=(2,3)(p,q)=(2,3). Then ζ(k)=col⁡(y(k−1),y(k−2),u(k−1),u(k−2),u(k−3))\zeta(k)=\operatorname{col}(y(k-1),y(k-2),u(k-1),u(k-2),u(k-3)), and the open-loop history matrices are

A0+eθTS1=[−α1−α20β1β210000000000010000010],B=[00100].A_0+e\theta^{\mathsf T}S_1= \begin{bmatrix} -\alpha_1&-\alpha_2&0&\beta_1&\beta_2\\ 1&0&0&0&0\\ 0&0&0&0&0\\ 0&0&1&0&0\\ 0&0&0&1&0 \end{bmatrix}, \qquad B= \begin{bmatrix} 0\\ 0\\ 1\\ 0\\ 0 \end{bmatrix}.

Only four entries are unknown. The first-row zero on u(k−1)u(k-1) records the delay; the new command enters row three. A larger delay lengthens the known shifts, leaving four uncertain coefficients.

Lemma 1. The true history pair (A0+eθTSd,B)(A_0+e\theta^{\mathsf T}S_d,B) is stabilizable. If Ad,K(θ)A_{d,K}(\theta) is Schur, then x,y,ux,y,u converge exponentially to zero once the history is available and feedback is activated.

Proof. For p=n,q=n+dp=n,q=n+d, let wTw^{\mathsf T} be a left PBH eigenvector orthogonal to BB, so wu,1=0w_{u,1}=0. Its output part is nonzero and obeys the companion recursion, hence D(μ)=0D(\mu)=0 and κ=wy,1≠0\kappa=w_{y,1}\ne0. With g=κ[0dT βT]g=\kappa[0_d^{\mathsf T}\ \beta^{\mathsf T}], the input recursion is wu,j+1=μwu,j−gjw_{u,j+1}=\mu w_{u,j}-g_j with terminal condition μwu,q=gq\mu w_{u,q}=g_q. Therefore ∑j=1qμq−jgj=κN(μ)=0\sum_{j=1}^{q}\mu^{q-j}g_j=\kappa N(\mu)=0, contradicting coprimeness. Extra memory adds only stable zero modes. Schur stability gives exponential decay of stored outputs and inputs. Observability reconstructs x(k)x(k) from the nn outputs and required inputs in ζ(k+n)\zeta(k+n). ■

This history also contains the information needed for predictor feedback: reconstruct x(k−n)x(k-n) from the stored outputs, then propagate (1) using the stored inputs to obtain x(k+d)x(k+d). Hence u(k)=Fxx(k+d)u(k)=F_xx(k+d) belongs to the class u=Kζu=K\zeta, although the design does not require measured states or an explicit predictor [1].

Known delay: an exact robustness margin

Fix the delay and memory. We seek one KK and one Q≻0Q\succ0 such that

Q−Ad,K(ϑ)QAd,K(ϑ)T≻0for every ϑ∈Σd(γ).(12)Q-A_{d,K}(\vartheta)QA_{d,K}(\vartheta)^{\mathsf T}\succ0\qquad\text{for every }\vartheta\in\Sigma_d(\gamma). \tag*{(12)}

This is the dual form of a common quadratic Lyapunov certificate; the Lyapunov matrix is Q−1Q^{-1}. It is stronger than asking each compatible model to be stable separately.

Theorem 1. Suppose Σd(γ)≠∅\Sigma_d(\gamma)\ne\varnothing. A certificate (12) requires rank Φd=2n\Phi_d=2n. Under this rank condition, it exists if and only if there are Q=QT≻0Q=Q^{\mathsf T}\succ0 and Z∈R1×mZ\in\mathbb{R}^{1\times m} such that

Md(γ,Q,Z):=[Q+(YYT−γ)eeTA0Q+BZeYΦdT∗Q−QSdT∗∗Hd]≻0.(13)\mathcal{M}_d(\gamma,Q,Z):= \begin{bmatrix} Q+(YY^{\mathsf T}-\gamma)e e^{\mathsf T} & A_0Q+BZ & eY\Phi_d^{\mathsf T}\\ * & Q & -QS_d^{\mathsf T}\\ * & * & H_d \end{bmatrix}\succ0. \tag*{(13)}

Then K=ZQ−1K=ZQ^{-1}. If the fitted pair (A0+eθ^dTSd,B)(A_0+e\hat{\theta}_d^{\mathsf T}S_d,B) is stabilizable, define

Γd=sup⁡{γ≥ed2:∃Q≻0,Z, Md(γ,Q,Z)≻0}.(14)\Gamma_d=\sup\{\gamma\ge e_d^2:\exists Q\succ0,Z,\ \mathcal{M}_d(\gamma,Q,Z)\succ0\}. \tag*{(14)}

The nonempty sets admitting (12) are exactly those with ed2≤γ<Γde_d^2 \le\gamma< \Gamma_d.

Proof. If ηTΦd=0\eta^{\mathsf T}\Phi_d=0 with η≠0\eta\ne0, any compatible ϑ0\vartheta_0 generates the line ϑ0+tη\vartheta_0+t\eta. The first row of (11) is then aT+tηTSda^{\mathsf T}+t\eta^{\mathsf T}S_d. Its squared QQ-norm is unbounded, whereas (12) bounds it by eTQee^{\mathsf T}Qe. Since SdS_d has full row rank, this is impossible.

For full rank, write A^d=A0+eϑ^TSd\widehat A_d=A_0+e\hat\vartheta^{\mathsf T}S_d and Ed=ρdHd−1/2SdE_d=\sqrt{\rho_d}H_d^{-1/2}S_d. By (8), the uncertain matrix is A^d+BK+eδTEd\widehat A_d+BK+e\delta^{\mathsf T}E_d, ∥δ∥2≤1\lVert\delta\rVert_2\le1. A Schur complement and the strict Petersen lemma [6] give the equivalent condition

\begin{bmatrix} Q-\lambda ee^{\mathsf T} & \widehat A_dQ+BZ & 0\\ * & Q & QE_d^{\mathsf T}\\ * & * & \lambda I_{2n} \end{bmatrix}\succ0,\qquad\lambda>0. \tag*{(15)}\end{bmatrix}

For ρd>0\rho_d>0, scale (Q,Z,λ)(Q,Z,\lambda) by ρd/λ\rho_d/\lambda, then take a Schur complement. This gives

\begin{bmatrix} Q-\rho_dee^{\mathsf T} & \widehat A_dQ+BZ & 0\\ * & Q & QS_d^{\mathsf T}\\ * & * & H_d \end{bmatrix}\succ0. \tag*{(16)}\end{bmatrix}

Eliminating HdH_d in (13) and (16) gives the same matrix, by (7). At ρd=0\rho_d=0, nominal stability implies (16) by scaling a nominal Lyapunov matrix sufficiently small; the reverse implication is immediate. Finally, increasing γ\gamma enlarges the set, and strict feasibility persists under a small increase in γ\gamma. The feasible budgets therefore form the stated interval. ■

The significance of (14) is practical: it measures how much total equation error this record can support within the chosen certificate class. The matrices YY, Φd\Phi_d are held fixed while γ\gamma is optimized. This differs from an experiment-wide guarantee in which the measured record itself changes with the disturbance. Replacing strict inequalities by semidefinite ones gives the same supremum: convex combinations with a strict feasible point approximate any point of the closed feasible set. A boundary optimizer need not be a certificate; recover and verify a controller at a budget below Γd\Gamma_d.

Corollary 1: Fix a stabilizable nominal coefficient vector θ0\theta_0, delay, and memory. Let R(H)\mathcal R(H) be the supremum of radii ρ\rho for which the ellipsoid {θ0+η:ηTHη≤ρ}\{\theta_0+\eta:\eta^{\mathsf T}H\eta\le\rho\} admits (12). For H≻0H\succ0, R\mathcal R is positive, finite, positively homogeneous, monotone in the semidefinite order, and concave. For the fitted nominal model, Γd=ed2+R(Hd)\Gamma_d=e_d^2+\mathcal R(H_d).

Proof. Use (16) with fixed A^d\widehat A_d. Convex combinations of feasible tuples prove concavity; scaling all blocks proves homogeneity; increasing HH preserves feasibility. Nominal stability gives a positive radius. For finiteness, the ellipsoid contains θ0±te1\theta_0\pm t e_1 with t=ρ/H11t=\sqrt{\rho/H_{11}}. Their closed-loop matrices differ by 2teeT2te e^{\mathsf T}. If both are strict contractions in the same quadratic norm, their half-difference is also a strict contraction, so t<1t<1. Thus R(H)≤H11\mathcal R(H)\le H_{11}. ■

At a fixed nominal model, more information cannot reduce the certified radius. The concavity statement further quantifies how this radius behaves when information matrices are combined; it does not compare records with different fitted nominal models.

Proposition 1: For a noiseless experiment, use columns k=n+d,…,n+d+T−1k=n+d,\ldots,n+d+T-1. If u(0),…,u(T+n−2)u(0),\ldots,u(T+n-2) is persistently exciting (PE) of order 2n2n, then rank⁡Φd=2n\operatorname{rank}\Phi_d=2n. Here PE of order ℓ\ell means full row rank of the depth-ℓ\ell input Hankel matrix. Thus T≥2nT\ge2n is necessary for (12), while T≥3nT\ge3n permits this sufficient PE condition, independently of dd.

An unstructured order-(n+d)(n+d) ARX model instead needs rank 2(n+d)2(n+d). At equal memory p=q=n+dp=q=n+d, every common-quadratic certificate for its compatible set is also a certificate for (6).

Proof. Put v(t)=u(t−d)v(t)=u(t-d). The regressor is an invertible linear image of col⁡(x(k−n),v(k−n),…,v(k−1))\operatorname{col}(x(k-n),v(k-n),\ldots,v(k-1)): its diagonal blocks are a reversed observability matrix and an input reversal matrix. The state/input rank lemma [16 Cor. 2], at depth nn, gives rank 2n2n from PE of order 2n2n. The stated input sequence has length T+n−1T+n-1, so its depth-2n2n Hankel matrix can have rank 2n2n only if T≥3nT\ge3n.

The unstructured regressor contains n+dn+d output and input lags. Its full-rank necessity follows as in Theorem 1. The embedding θ↦col⁡(α,0d,0d,β)\theta\mapsto\operatorname{col}(\alpha,0_d,0_d,\beta) preserves residuals and closed-loop matrices at equal memory. Its compatible set therefore contains the structured set. ■

Table 1 separates parameter count from memory. At equal memory p=q=n+dp=q=n+d, the structured LMI order is 6n+4d6n+4d; its uncertainty still has dimension 2n2n. Factoring the data costs O(n2T)O(n^2T) versus O((n+d)2T)O((n+d)^2T), whereas SDP cost also grows with the LMI order. Neither delay-independent parameter count nor PE order means delay-independent duration: the final output is at n+d+T−1n+d+T-1, spanning n+d+Tn+d+T sampling instants from time zero.

Unknown delay: retain, then stabilize

Let D={dmin⁡,…,dmax⁡}\mathcal D=\{d_{\min},\ldots,d_{\max}\} contain the true constant delay, and let Δ=dmax⁡−dmin⁡\Delta=d_{\max}-d_{\min}. Every candidate uses the same columns, with kj≥n+dmax⁡k_j\ge n+d_{\max}, and the common memory p=n,q=n+dmax⁡p=n,q=n+d_{\max}. Retain

Dc(γ)={j∈D:min⁡ϑ∥Y−ϑTΦj∥22≤γ}.(17)\mathcal D_c(\gamma)=\left\{j\in\mathcal D:\min_{\vartheta}\lVert Y-\vartheta^{\mathsf T}\Phi_j\rVert_2^2\le\gamma\right\}. \tag*{(17)}

The true delay belongs to this set under (5). The test rejects an explanation only when its best fit already violates the error bound.

What noiseless data can distinguish

For this subsection and the small-error result below, assume additionally that AA is nonsingular and N(0)=βn≠0N(0)=\beta_n\ne0 for the true physical plant. The candidate sets (6) remain

QuantityStructuredUnstructured
Unknown coefficients2n2n2(n+d)2(n+d)
Necessary columnsT≥2nT\ge2nT≥2(n+d)T\ge2(n+d)
History (p,q)(p,q)(n,n+d)(n,n+ d)(n+d,n+d)(n+d,n+d)
LMI order6n+2d6n+2d6(n+d)6(n+d)

Table 1. Known-delay dimensions. Column bounds are necessary; they do not alone guarantee rank or stabilizability.

relaxed ARX sets; these extra restrictions are not imposed on every candidate.

Theorem 2. Use one noiseless experiment with columns k=n+dmax⁡,…,n+dmax⁡+T−1k = n + d_{\max}, \ldots, n + d_{\max} + T - 1. If u(0),…,u(T+n+Δ−2)u(0), \ldots, u(T + n + \Delta- 2) is PE of order 2n+Δ2n + \Delta, then all Φj\Phi_j have rank 2n2n, and

Dc(0)={d,…,d+r−1}∩D,d=min⁡Dc(0).(18)\mathcal{D}_c(0) = \{d, \ldots, d + r - 1\} \cap\mathcal{D}, \quad d = \min\mathcal{D}_c(0). \tag*{(18)}

The fit at every retained delay j=d+sj = d + s is unique: θs=col⁡(α,βs+1,…,βn,0s)\theta_s = \operatorname{col}(\alpha, \beta_{s+1}, \ldots, \beta_n, 0_s). It satisfies θsTSd+s=θTSd\theta_s^{\mathsf T}S_{d+s} = \theta^{\mathsf T}S_d, so these candidates describe the same input–output behavior. The PE condition requires T≥3n+ΔT \ge3n + \Delta.

Proof. Set v(t)=u(t−d)v(t) = u(t - d) and Δ+=dmax⁡−d\Delta_+ = d_{\max} - d, Δ−=d−dmin⁡\Delta_- = d - d_{\min}. Each output and candidate regressor is a linear function of

ξ(k)=col⁡(x(k−n−Δ+),v(k−n−Δ+),…,v(k−1+Δ−)).\xi(k) = \operatorname{col}\bigl(x(k - n - \Delta_+), v(k - n - \Delta_+), \ldots, v(k - 1 + \Delta_-)\bigr).

The state/input rank lemma [16], now at depth n+Δn + \Delta, makes the data matrix of ξ\xi full row rank 2n+Δ2n + \Delta. Thus any linear identity on the data holds for arbitrary ξ\xi.

Set its input entries to zero. Since AA is invertible, x(k−n)x(k-n) is arbitrary; observability then makes the output part of any relation ηTφj=0\eta^{\mathsf T}\varphi_j = 0 vanish. The remaining input entries are distinct independent coordinates, so η=0\eta= 0. This proves the rank claim. Likewise, an exact candidate fit first forces α′=α\alpha' = \alpha. After canceling these output terms, arbitrary inputs give

∑i=1nβiu(k−d−i)=∑i=1nβi′u(k−j−i).\sum_{i=1}^{n}\beta_i u(k-d-i) = \sum_{i=1}^{n}\beta'_i u(k-j-i).

The nonzero extreme terms at lags d+rd+r and d+nd+n require j+1≤d+rj+1 \le d+r and j+n≥d+nj+n \ge d+n, hence d≤j≤d+r−1d \le j \le d+r-1. Conversely, shifting the leading zero numerator coefficients gives exactly the stated θs\theta_s, identically for every history. The PE length count gives T≥3n+ΔT \ge3n + \Delta.

For example, n=2n = 2, r=2r = 2 allows descriptions with delays dd and d+1d+1: the sole nonzero input term occurs at lag d+2d+2 in both. For s>0s > 0, the shifted numerator is zsN(z)z^sN(z) and has a zero at the origin. Such a description is allowed by the relaxed test, even though the true physical plant has no such zero. The smallest retained delay identifies the physical delay under the stated priors. Without N(0)≠0N(0) \ne0, an earlier delay can also fit.

Proposition 2. On noiseless data, let s=dmax⁡−d>0s = d_{\max} - d > 0 and fit an unstructured ARX model of order h=n+dmax⁡h = n + d_{\max}. Then rank⁡Φag≤2h−s\operatorname{rank}\Phi_{\mathrm{ag}} \le2h-s. Moreover, its compatible unreduced history models include a model with an unstable uncontrollable mode. No history feedback stabilizes this entire set.

Proof. For every monic polynomial psp_s of degree ss, the pair (ps(z)zdD(z),ps(z)N(z))(p_s(z)z^dD(z), p_s(z)N(z)) gives an exact order-hh ARX equation. Its coefficients form an injective affine family of dimension ss, proving the rank bound. Choose psp_s a real root μ>1\mu> 1. Both fitted polynomials vanish at μ\mu; the PBH recursion in Lemma 1, read in reverse, gives an uncontrollable eigenvalue μ\mu of the unreduced history pair. Feedback cannot move it.

This obstruction concerns the entire unrestricted fitted family, not the reduced physical plant. Lee et al. [17] select an ARX representative with stable nonminimal modes; such a selection does not certify every member of the unrestricted family.

One gain for all retained candidates

Assume each retained Φj\Phi_j has rank 2n2n. Form A^j=A0+eθ^jTSj\hat A_j = A_0 + e\hat\theta_j^{\mathsf T}S_j and Ej=γ−ej2Hj−1/2SjE_j = \sqrt{\gamma-e_j^2H_j^{-1/2}S_j} as before.

Theorem 3. Suppose there are shared matrices F∈Rm×mF \in\mathbb{R}^{m\times m}, L∈R1×mL \in\mathbb{R}^{1\times m} and, for every j∈Dc(γ)j \in\mathcal{D}_c(\gamma), matrices Pj=PjT≻0P_j = P_j^{\mathsf T} \succ0 and scalars λj>0\lambda_j > 0 satisfying

\begin{bmatrix} P_j-\lambda_je e^{\mathsf T} & \hat A_jF+BL & 0 \\ * & F+F^{\mathsf T}-P_j & F E_j^{\mathsf T} \\ * & * & \lambda_j I_{2n} \end{bmatrix} \succ0. \tag*{(19)}\end{bmatrix}

Then FF is nonsingular and K=LF−1K = LF^{-1} stabilizes every ϑ∈Σj(γ)\vartheta\in\Sigma_j(\gamma) for every retained jj. Each Pj−1P_j^{-1} is a common quadratic Lyapunov matrix within its coefficient set. With one candidate, feasibility is equivalent to Theorem 1.

Proof. The middle block implies F+FT≻Pj≻0F+F^{\mathsf T} \succ P_j \succ0, so FF is nonsingular. Petersen’s lemma transforms (19) into [PjAj,K(ϑ)F;∗F+FT−Pj]≻0[P_j \quad A_{j,K}(\vartheta)F; * \quad F+F^{\mathsf T}-P_j] \succ0 for every compatible ϑ\vartheta. Since

FTPj−1F−(F+FT−Pj)=(F−Pj)TPj−1(F−Pj)⪰0,F^{\mathsf T}P_j^{-1}F-(F+F^{\mathsf T}-P_j)=(F-P_j)^{\mathsf T}P_j^{-1}(F-P_j)\succeq0,

replacing the lower block by FTPj−1FF^{\mathsf T}P_j^{-1}F preserves positivity. A Schur complement gives Pj−Aj,K(ϑ)PjAj,K(ϑ)T≻0P_j-A_{j,K}(\vartheta)P_jA_{j,K}(\vartheta)^{\mathsf T}\succ0. For one candidate, choose F=Pj=QF=P_j=Q and L=ZL=Z in (15) for the converse. ▪

The shared variables enforce one implementable gain; the PjP_j may differ. Thus (19) does not certify arbitrarily switching delays. It also does not claim that infeasibility rules out every simultaneous stabilizing controller. The dilation follows the robust stability principle in [18].

The design procedure is therefore direct. First, fix the delay interval, common data columns, and memory. Second, fit each candidate and reject only those whose residual exceeds the prescribed bound. Third, check full row rank for every survivor and solve (19) jointly; with a single survivor, the exact test (13) is available. Finally, recover KK, verify strict positivity of the numerical certificate, and activate feedback from the stored history. A consistent rank-deficient candidate prevents a certificate of the present type by Theorem 1; it cannot simply be discarded.

Corollary 2. Fix the input, initial histories, and finite experiment in Theorem 2, and let Dc0\mathcal{D}_c^0 denote the candidate set of its noiseless reference record. During collection, let the ARX recursion have equation error with total energy at most γ\gamma over all generated samples. There exists γˉ>0\bar\gamma> 0 such that, for every 0≤γ≤γˉ0 \le\gamma\le\bar\gamma and every such error, all candidate regressors have full row rank, Dc(γ)=Dc0\mathcal{D}_c(\gamma)=\mathcal{D}_c^0, and (19) is feasible. For a known delay, the same conclusion follows from the rank condition in Proposition 1 without the extra nonsingularity assumptions.

Proof: Over a fixed finite horizon, the perturbed record differs from the noiseless one by O(γ)O(\sqrt{\gamma}), uniformly over these errors. Full row rank therefore persists. Every false noiseless candidate has a strictly positive least-squares residual; continuity and finiteness of DD keep all such squared residuals above γ\gamma for small γ\gamma. Every equivalent candidate remains consistent because θsTSd+s=θTSd\theta_s^T S_{d+s}=\theta^T S_d on arbitrary histories: its equation error is the same WW. For each survivor, θ^j=θs+O(γ)\hat{\theta}_j=\theta_s+O(\sqrt{\gamma}) and Ej=O(γ)E_j=O(\sqrt{\gamma}) uniformly. Their nominal history matrices converge to one stabilizable matrix. Choose a strict nominal certificate Q,KQ,K using Lemma 1, and set F=Pj=QF=P_j=Q, L=KQL=KQ, λj=γ\lambda_j=\sqrt{\gamma}. After eliminating the last block of (19), its perturbation tends to zero, so strict positivity persists. At γ=0\gamma=0, use sufficiently small positive λj\lambda_j. The singleton argument needs only nominal stabilizability and full rank.

Using a sensor-noise bound

Sensor noise perturbs both YY and the output rows of Φj\Phi_j. Its induced equation error is v(k)+∑iαiv(k−i)v(k)+\sum_i\alpha_i v(k-i) and therefore depends on unknown coefficients. For sensor error alone, a prior ∥α∥1≤a∗\|\alpha\|_1\le a_* and ∣v(k)∣≤vˉ|v(k)|\le\bar{v} gives γ=T(1+a∗)2vˉ2\gamma=T(1+a_*)^2\bar{v}^2. The following alternative needs no coefficient prior; it gives a conservative outer set for this errors-in-variables problem [12].

Fix a candidate and suppress its subscript. Write the measured regressor as Φ=col⁡(−X,U)\Phi=\operatorname{col}(-X,U), where XX contains the nn output lags and UU the candidate input lags. Inputs are exact. Collect every distinct sensor-error sample used by the record into one vector vv, with ∥v∥2≤δv\|v\|_2\le\delta_v, and let LivL_i v select the errors at lag ii, i=0,…,ni=0,\ldots,n. The same sample has the same coordinate wherever it reappears. Allow additional selected equation errors with norm at most δw\delta_w. Let U†U^\dagger denote the Moore–Penrose inverse and define

s⊥=σmin⁡(X(I−U†U)),κ=∥col⁡(L1,…,Ln)∥2,s_\perp=\sigma_{\min}\left(X(I-U^\dagger U)\right),\quad\kappa=\left\|\operatorname{col}(L_1,\ldots,L_n)\right\|_2,
bv=δw+δv∥L0+∑iα^iLi∥2,τ=κδvs⊥.(20)b_v=\delta_w+\delta_v\left\|L_0+\sum_i\hat{\alpha}_iL_i\right\|_2,\quad\tau=\frac{\kappa\delta_v}{s_\perp}. \tag*{(20)}

The projection I−U†UI-U^\dagger U removes the part of the output histories explained by the input histories; s⊥s_\perp measures the independent information that remains.

Proposition 3: Suppose Φ\Phi has full row rank, so s⊥>0s_\perp>0, and τ<1\tau<1. Write eLS=∥Y−θ^TΦ∥2e_{\mathrm{LS}}=\|Y-\hat{\theta}^T\Phi\|_2. If eLS>bv/1−τ2e_{\mathrm{LS}}>b_v/\sqrt{1-\tau^2}, the candidate is inconsistent with the sensor and equation-error bounds. Otherwise, every compatible coefficient vector belongs to

(ϑ−θ^)TΦΦT(ϑ−θ^)≤ηv2,(\vartheta-\hat{\theta})^T\Phi\Phi^T(\vartheta-\hat{\theta})\le\eta_v^2,
ηv=bvτ+bv2τ2−(1−τ2)eLS21−τ2.(21)\eta_v=\frac{b_v\tau+\sqrt{b_v^2\tau^2-(1-\tau^2)e_{\mathrm{LS}}^2}}{1-\tau^2}. \tag*{(21)}

Equivalently, this is (6) for the measured record with γeff=eLS2+ηv2\gamma_{\mathrm{eff}}=e_{\mathrm{LS}}^2+\eta_v^2.

Proof: Put z=∥ΦT(ϑ−θ^)∥2z=\|\Phi^T(\vartheta-\hat{\theta})\|_2. Projecting onto ker⁡U\ker U gives ∥α−α^∥2≤z/s⊥\|\alpha-\hat{\alpha}\|_2\le z/s_\perp. The measured equation identity is YT−ΦTϑ=w+(L0+∑iαiLi)vY^T-\Phi^T\vartheta=w+(L_0+\sum_i\alpha_iL_i)v, whose norm is at most bv+τzb_v+\tau z. Least-squares orthogonality then gives eLS2+z2≤(bv+τz)2e_{\mathrm{LS}}^2+z^2\le(b_v+\tau z)^2. Solving this scalar quadratic proves both claims.

The condition τ<1\tau<1 checks whether data separate denominator uncertainty from sensor error sufficiently well. Its failure is inconclusive, not grounds for rejecting a delay. Passing the residual test also does not establish physical consistency: (21) is an outer enclosure. Apply Theorem 1 with γeff\gamma_{\mathrm{eff}}. If full rank and τj<1\tau_j<1 hold for every unrejected candidate, use (19) with Ej=ηv,jHj−1/2SjE_j=\eta_{v,j}H_j^{-1/2}S_j. The enclosure may be conservative even though subsequent ellipsoid certification is exact at a fixed delay.

Numerical examples

Both examples use D(z)=(z−1.1)(z−0.8)D(z)=(z-1.1)(z-0.8) and N(z)=0.2z+0.1N(z)=0.2z+0.1: the undelayed plant is minimal and unstable. Least squares and uncertainty factors use an SVD. We solve the LMIs with CVXPY 1.7.5/Clarabel 0.11.1 and independently reconstruct each returned matrix. A certificate is accepted only when its minimum eigenvalue exceeds 100ϵmachmax⁡(1,∥M∥2)100\epsilon_{\mathrm{mach}}\max(1,\|M\|_2) for an N×NN\times N matrix MM. This floating-point verification uses no interval arithmetic; solver statuses alone are not taken as certificates.

Robustness at equal controller memory

For each d∈{1,3,5}d\in\{1,3,5\}, twelve length-40 experiments use independent standard Gaussian inputs and equation errors uniform in [−10−3,10−3][-10^{-3},10^{-3}]. The same columns k=8,…,39k=8,\ldots,39 give T=384T=384 for both fits. Both controllers have p=q=2+dp=q=2+d; the structured fit estimates four coefficients and the unstructured fit estimates 2(2+d)2(2+d). Initial output pairs are standard Gaussian, corresponding to physical initial states by observability. The seed is 439.

Table 2 reports numerical values of (14). These are squared equation-error energies for fixed records, not noise-amplitude tolerances across regenerated experiments. Separate strict certificates are recovered at ed2+0.99(Γd−ed2)e_d^2+0.99(\Gamma_d-e_d^2); their smallest relative eigenvalue exceeds 2.5⋅10−82.5\cdot10^{-8}.

ddΓstr\Gamma_{\mathrm{str}}Γag\Gamma_{\mathrm{ag}}UagU_{\mathrm{ag}}Ratio lower bound
14.3771.8793.5471.22
32.2776.562⋅10−26.562 \cdot10^{-2}8.131⋅10−28.131 \cdot10^{-2}27.7
51.3193.080⋅10−33.080 \cdot10^{-3}3.735⋅10−33.735 \cdot10^{-3}349.5

Table 2. Matched-memory margins: numerical SDP values, independent upper bounds UagU_{\mathrm{ag}}, and certified ratio lower bounds.

For an independent upper bound, construct an unstructured coefficient vector whose numerator and denominator share z=−1z=-1. The PBH argument of Proposition 2 makes its full history realization unstabilizable. Its residual energy bounds the unstructured margin from above. The root identities, PBH witness, and residual energies are checked in exact rational arithmetic on the saved record. Table 2 rounds these upper bounds UagU_{\mathrm{ag}} upward; dividing the structured strict lower budget by UagU_{\mathrm{ag}} gives the certified ratio. The separation therefore does not depend on boundary optimality or solver failure.

Three genuinely different delays remain plausible

Now d=2d=2, D={1,2,3}D=\{1,2,3\}, and (p,q)=(2,5)(p,q)=(2,5). Twelve length-16 experiments use period-two inputs, each alternating between two independent standard Gaussian amplitudes; initial output pairs are also standard Gaussian (seed 806). Selected columns k=5,…,15k=5,\ldots,15 give T=132T=132. Errors uniform in [−0.01,0.01][-0.01,0.01] give the predetermined bound γ=0.0132\gamma=0.0132.

All three candidates have rank four, condition number κ2(Φj)=89.44\kappa_2(\Phi_j) = 89.44, residual energy 0.004189, and centered squared radius 0.009011. Period-two probing deliberately fails the PE assumption of Theorem 2, so unique delay identification is not expected. The three chosen compatible plants are genuinely different: their first nonzero impulse responses occur at samples 2, 3, and 4.

We solve (19) with ∑jtr⁡Pj=3\sum_j \operatorname{tr} P_j = 3. A secondary quadratic gain surrogate retains half the maximum strict margin at this fixed scale. The final minimum ratio λmin⁡(Mj)/∥Mj∥2\lambda_{\min}(M_j)/\lVert M_j\rVert_2, for the LMI matrices MjM_j, is 1.05⋅10−41.05 \cdot10^{-4}. This certifies the entire three-ellipsoid family. Figure 1 illustrates the same gain on three exact compatible plants: α=(−1.9,0.88)\alpha= (-1.9, 0.88), β=(0.2,0.1)\beta= (0.2, 0.1) at d=2d = 2, and β=(0.1,0.2)\beta= (0.1, 0.2) at d=1,3d = 1, 3. With zero past inputs, observability supplies a physical state for each plant having y(−2)=0.9y(-2) = 0.9, y(−1)=1y(-1) = 1. All activate at k=0k = 0. These simulations illustrate the certificate.

Figure 1. One controller stabilizes three genuinely different compatible delayed plants. All use ζ(0)=(1,0.9,0,0,0,0,0)T\zeta(0) = (1, 0.9, 0, 0, 0, 0, 0)^{\mathsf{T}} and activate at k=0k = 0, with identical controller gains.

Conclusion

Known delay registers separate feedback memory from a fixed 2n2n-dimensional uncertainty. This yields an exact common-quadratic fixed-record margin and a certified shared controller over retained constant-delay families. Uniform small-error feasibility holds under explicit excitation assumptions. Matched-memory bounds demonstrate the structural gain; the three-delay experiment shows that delay identification need not precede certified stabilization. Simultaneous synthesis and sensor-noise enclosure remain conservative.

References

References

  1. [1]M. Krstić, Delay Compensation for Nonlinear, Adaptive, and PDE Systems. Boston, MA: Birkhäuser, 2009.DOI
  2. [2]E. Fridman, Introduction to Time-Delay Systems: Analysis and Control. Cham: Birkhäuser, 2014.DOI
  3. [3]C. De Persis and P. Tesi, “Formulas for data-driven control: Stabilization, optimality, and robustness,” IEEE Trans. Autom. Control, vol. 65, no. 3, pp. 909–924, 2020.arxiv.org/abs/1903.06842
  4. [4]H. J. van Waarde, J. Eising, H. L. Trentelman, and M. K. Camlibel, “Data informativity: A new perspective on data-driven analysis and control,” IEEE Trans. Autom. Control, vol. 65, no. 11, pp. 4753–4768, 2020.arxiv.org/abs/1908.00468
  5. [5]H. J. van Waarde, M. K. Camlibel, and M. Mesbahi, “From noisy data to feedback controllers: Nonconservative design via a matrix S-lemma,” IEEE Trans. Autom. Control, vol. 67, no. 1, pp. 162–175, 2022.
  6. [6]A. Bisoffi, C. De Persis, and P. Tesi, “Data-driven control via Petersen’s lemma,” Automatica, vol. 145, p. 110537, 2022.DOI
  7. [7]J. G. Rueda-Escobedo, E. Fridman, and J. Schiffer, “Data-driven control for linear discrete-time delay systems,” IEEE Trans. Autom. Control, vol. 67, no. 7, pp. 3321–3336, 2022.
  8. [8]G. Portilla, A. Seuret, and S. Mondié, “Robust data-driven control for linear discrete-time systems with unknown delay,” IFAC-PapersOnLine, vol. 59, no. 13, pp. 231–236, 2025.DOI
  9. [9]Z. Zhao, J. Wang, and W. Lan, “Data-driven safe output regulation of strict-feedback linear systems with input delay,” 2026, arXiv:2601.14089v2.arxiv.org/abs/2601.14089
  10. [10]M. Alsalti, V. G. Lopez, and M. A. Müller, “Notes on data-driven output-feedback control of linear MIMO systems,” IEEE Trans. Autom. Control, vol. 70, no. 9, pp. 6143–6150, 2025.DOI
  11. [11]H. J. van Waarde, J. Eising, M. K. Camlibel, and H. L. Trentelman, “A behavioral approach to data-driven control with noisy input–output data,” IEEE Trans. Autom. Control, vol. 69, no. 2, pp. 813–827, 2024.arxiv.org/abs/2206.08408
  12. [12]L. Li, A. Bisoffi, C. De Persis, and N. Monshizadeh, “Controller synthesis from noisy-input noisy-output data,” Automatica, vol. 183, p. 112545, 2026.DOI
  13. [13]J. Berberich, C. W. Scherer, and F. Allgöwer, “Combining prior knowledge and data for robust controller design,” IEEE Trans. Autom. Control, vol. 68, no. 8, pp. 4618–4633, 2023.DOI
  14. [14]S. Björklund and L. Ljung, “A review of time-delay estimation techniques,” in Proc. 42nd IEEE Conf. Decis. Control, vol. 3, 2003, pp. 2502–2507.DOI
  15. [15]A. Peregudin and N. T. Dinh, “Minimal experiments for robust stabilization: Information, spectral geometry, and duration,” 2026, arXiv:2610.02187.arxiv.org/abs/2610.02187
  16. [16]J. C. Willems, P. Rapisarda, I. Markovsky, and B. L. M. De Moor, “A note on persistency of excitation,” Syst. Control Lett., vol. 54, no. 4, pp. 325–329, 2005.
  17. [17]J. Lee, N. H. Jo, H. Shim, F. Dörfler, and J. Kim, “Ensuring stability of non-minimal modes in input-output data-driven representation,” 2026, arXiv:2512.01238v2.arxiv.org/abs/2512.01238
  18. [18]M. C. de Oliveira, J. Bernussou, and J. C. Geromel, “A new discrete-time robust stability condition,” Syst. Control Lett., vol. 37, no. 4, pp. 261–265, 1999.DOI

Paper details

vlnc:W_qx2v3xyp
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Engineering

Not recorded for this paper

View statement
Looking for related papers…
Contents