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 uncertain coefficients, with 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.
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).
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 for column stacking, for symmetric blocks, for the Euclidean or spectral norm, and for canonical basis vectors.
Plant, data, and controller memory
Consider the real, discrete-time plant
where , , and is an integer; are initial conditions. The undelayed realization is minimal and its order is known. Define
Minimality makes and coprime. The transfer function is , with relative degree . By Cayley–Hamilton, for arbitrary initial states and ,
For only nonnegative input indices occur. Put and
What the error bound means
At selected times , let
Here and collect the columns and , respectively. Columns may be stacked from several experiments, with valid lag histories in each. The error is an equation error in (3). The bound 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
It deliberately allows nonminimal ARX models: retaining them gives a convex uncertainty description that contains the physical plant. If has full row rank, let
Least-squares orthogonality gives
Thus rejects the candidate, while 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 and , store
where . Let have ones on its first subdiagonal and zeros elsewhere, and set , , and . Define by : its first rows select the first output entries with a minus sign, and its last rows select input entries . Consequently,
The first row computes the next stored output; the other rows only shift measured values. We use , with
The smallest memory used here is . This is a property of this implementation, not a lower bound on the order of every possible controller.
For example, take , and . Then , and the open-loop history matrices are
Only four entries are unknown. The first-row zero on 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 is stabilizable. If is Schur, then converge exponentially to zero once the history is available and feedback is activated.
Proof. For , let be a left PBH eigenvector orthogonal to , so . Its output part is nonzero and obeys the companion recursion, hence and . With , the input recursion is with terminal condition . Therefore , contradicting coprimeness. Extra memory adds only stable zero modes. Schur stability gives exponential decay of stored outputs and inputs. Observability reconstructs from the outputs and required inputs in . ■
This history also contains the information needed for predictor feedback: reconstruct from the stored outputs, then propagate (1) using the stored inputs to obtain . Hence belongs to the class , 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 and one such that
This is the dual form of a common quadratic Lyapunov certificate; the Lyapunov matrix is . It is stronger than asking each compatible model to be stable separately.
Theorem 1. Suppose . A certificate (12) requires rank . Under this rank condition, it exists if and only if there are and such that
Then . If the fitted pair is stabilizable, define
The nonempty sets admitting (12) are exactly those with .
Proof. If with , any compatible generates the line . The first row of (11) is then . Its squared -norm is unbounded, whereas (12) bounds it by . Since has full row rank, this is impossible.
For full rank, write and . By (8), the uncertain matrix is , . A Schur complement and the strict Petersen lemma [6] give the equivalent condition
For , scale by , then take a Schur complement. This gives
Eliminating in (13) and (16) gives the same matrix, by (7). At , nominal stability implies (16) by scaling a nominal Lyapunov matrix sufficiently small; the reverse implication is immediate. Finally, increasing enlarges the set, and strict feasibility persists under a small increase in . 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 , are held fixed while 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 .
Corollary 1: Fix a stabilizable nominal coefficient vector , delay, and memory. Let be the supremum of radii for which the ellipsoid admits (12). For , is positive, finite, positively homogeneous, monotone in the semidefinite order, and concave. For the fitted nominal model, .
Proof. Use (16) with fixed . Convex combinations of feasible tuples prove concavity; scaling all blocks proves homogeneity; increasing preserves feasibility. Nominal stability gives a positive radius. For finiteness, the ellipsoid contains with . Their closed-loop matrices differ by . If both are strict contractions in the same quadratic norm, their half-difference is also a strict contraction, so . Thus . ■
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 . If is persistently exciting (PE) of order , then . Here PE of order means full row rank of the depth- input Hankel matrix. Thus is necessary for (12), while permits this sufficient PE condition, independently of .
An unstructured order- ARX model instead needs rank . At equal memory , every common-quadratic certificate for its compatible set is also a certificate for (6).
Proof. Put . The regressor is an invertible linear image of : its diagonal blocks are a reversed observability matrix and an input reversal matrix. The state/input rank lemma [16 Cor. 2], at depth , gives rank from PE of order . The stated input sequence has length , so its depth- Hankel matrix can have rank only if .
The unstructured regressor contains output and input lags. Its full-rank necessity follows as in Theorem 1. The embedding 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 , the structured LMI order is ; its uncertainty still has dimension . Factoring the data costs versus , 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 , spanning sampling instants from time zero.
Unknown delay: retain, then stabilize
Let contain the true constant delay, and let . Every candidate uses the same columns, with , and the common memory . Retain
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 is nonsingular and for the true physical plant. The candidate sets (6) remain
| Quantity | Structured | Unstructured |
| Unknown coefficients | ||
| Necessary columns | ||
| History | ||
| LMI order |
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 . If is PE of order , then all have rank , and
The fit at every retained delay is unique: . It satisfies , so these candidates describe the same input–output behavior. The PE condition requires .
Proof. Set and , . Each output and candidate regressor is a linear function of
The state/input rank lemma [16], now at depth , makes the data matrix of full row rank . Thus any linear identity on the data holds for arbitrary .
Set its input entries to zero. Since is invertible, is arbitrary; observability then makes the output part of any relation vanish. The remaining input entries are distinct independent coordinates, so . This proves the rank claim. Likewise, an exact candidate fit first forces . After canceling these output terms, arbitrary inputs give
The nonzero extreme terms at lags and require and , hence . Conversely, shifting the leading zero numerator coefficients gives exactly the stated , identically for every history. The PE length count gives .
For example, , allows descriptions with delays and : the sole nonzero input term occurs at lag in both. For , the shifted numerator is 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 , an earlier delay can also fit.
Proposition 2. On noiseless data, let and fit an unstructured ARX model of order . Then . 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 of degree , the pair gives an exact order- ARX equation. Its coefficients form an injective affine family of dimension , proving the rank bound. Choose a real root . Both fitted polynomials vanish at ; the PBH recursion in Lemma 1, read in reverse, gives an uncontrollable eigenvalue 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 has rank . Form and as before.
Theorem 3. Suppose there are shared matrices , and, for every , matrices and scalars satisfying
Then is nonsingular and stabilizes every for every retained . Each is a common quadratic Lyapunov matrix within its coefficient set. With one candidate, feasibility is equivalent to Theorem 1.
Proof. The middle block implies , so is nonsingular. Petersen’s lemma transforms (19) into for every compatible . Since
replacing the lower block by preserves positivity. A Schur complement gives . For one candidate, choose and in (15) for the converse. ▪
The shared variables enforce one implementable gain; the 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 , 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 denote the candidate set of its noiseless reference record. During collection, let the ARX recursion have equation error with total energy at most over all generated samples. There exists such that, for every and every such error, all candidate regressors have full row rank, , 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 , uniformly over these errors. Full row rank therefore persists. Every false noiseless candidate has a strictly positive least-squares residual; continuity and finiteness of keep all such squared residuals above for small . Every equivalent candidate remains consistent because on arbitrary histories: its equation error is the same . For each survivor, and uniformly. Their nominal history matrices converge to one stabilizable matrix. Choose a strict nominal certificate using Lemma 1, and set , , . After eliminating the last block of (19), its perturbation tends to zero, so strict positivity persists. At , use sufficiently small positive . The singleton argument needs only nominal stabilizability and full rank.
Using a sensor-noise bound
Sensor noise perturbs both and the output rows of . Its induced equation error is and therefore depends on unknown coefficients. For sensor error alone, a prior and gives . 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 , where contains the output lags and the candidate input lags. Inputs are exact. Collect every distinct sensor-error sample used by the record into one vector , with , and let select the errors at lag , . The same sample has the same coordinate wherever it reappears. Allow additional selected equation errors with norm at most . Let denote the Moore–Penrose inverse and define
The projection removes the part of the output histories explained by the input histories; measures the independent information that remains.
Proposition 3: Suppose has full row rank, so , and . Write . If , the candidate is inconsistent with the sensor and equation-error bounds. Otherwise, every compatible coefficient vector belongs to
Equivalently, this is (6) for the measured record with .
Proof: Put . Projecting onto gives . The measured equation identity is , whose norm is at most . Least-squares orthogonality then gives . Solving this scalar quadratic proves both claims.
The condition 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 . If full rank and hold for every unrejected candidate, use (19) with . The enclosure may be conservative even though subsequent ellipsoid certification is exact at a fixed delay.
Numerical examples
Both examples use and : 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 for an matrix . This floating-point verification uses no interval arithmetic; solver statuses alone are not taken as certificates.
Robustness at equal controller memory
For each , twelve length-40 experiments use independent standard Gaussian inputs and equation errors uniform in . The same columns give for both fits. Both controllers have ; the structured fit estimates four coefficients and the unstructured fit estimates . 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 ; their smallest relative eigenvalue exceeds .
| Ratio lower bound | ||||
| 1 | 4.377 | 1.879 | 3.547 | 1.22 |
| 3 | 2.277 | 27.7 | ||
| 5 | 1.319 | 349.5 |
Table 2. Matched-memory margins: numerical SDP values, independent upper bounds , and certified ratio lower bounds.
For an independent upper bound, construct an unstructured coefficient vector whose numerator and denominator share . 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 upward; dividing the structured strict lower budget by gives the certified ratio. The separation therefore does not depend on boundary optimality or solver failure.
Three genuinely different delays remain plausible
Now , , and . 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 give . Errors uniform in give the predetermined bound .
All three candidates have rank four, condition number , 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 . A secondary quadratic gain surrogate retains half the maximum strict margin at this fixed scale. The final minimum ratio , for the LMI matrices , is . This certifies the entire three-ellipsoid family. Figure 1 illustrates the same gain on three exact compatible plants: , at , and at . With zero past inputs, observability supplies a physical state for each plant having , . All activate at . These simulations illustrate the certificate.
Figure 1. One controller stabilizes three genuinely different compatible delayed plants. All use and activate at , with identical controller gains.
Conclusion
Known delay registers separate feedback memory from a fixed -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]M. Krstić, Delay Compensation for Nonlinear, Adaptive, and PDE Systems. Boston, MA: Birkhäuser, 2009.DOI
- [2]E. Fridman, Introduction to Time-Delay Systems: Analysis and Control. Cham: Birkhäuser, 2014.DOI
- [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]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]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]A. Bisoffi, C. De Persis, and P. Tesi, “Data-driven control via Petersen’s lemma,” Automatica, vol. 145, p. 110537, 2022.DOI
- [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]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]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]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]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]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]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]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]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]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]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]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
