Set-Membership Identification Overview

• Set-membership identification is a methodology that defines the set of all parameter values consistent with observed data and known error bounds.
• It employs geometric representations like polyhedra, zonotopes, and ellipsoids to capture uncertainty and support robust control and state estimation.
• Algorithms range from LP-based recursions to SOS relaxations, providing both deterministic and probabilistic performance guarantees.

Searching arXiv for the cited set-membership identification papers to ground the article in current literature. Set-membership identification (SMI) is a family of estimation methods in which the unknown model is characterized by the set of all parameters consistent with a model structure, the available measurements, and known bounds on uncertainties, rather than by a single unconstrained estimate. In bounded-error settings this yields a Feasible Solution Set (FSS) or Feasible Parameter Set (FPS) that is maintained, updated, intersected, or outer-approximated over time; in newer stochastic formulations the same logic is extended to high-probability uncertainty sets under sub-Gaussian noise with unbounded support. Recent work places SMI at the intersection of linear and nonlinear identification, errors-in-variables estimation, event-triggered state estimation, robust control synthesis, and experiment design through universal inputs (Wang et al., 2017, Brändle et al., 1 Apr 2026, Shakouri et al., 1 Jul 2026).

1. Core formulation and problem classes

At its most general, SMI exploits the known boundedness of uncertainty to constrain model parameters to the values that are consistent with observed data. In the switched linear setting with bounded process noise, a constrained estimator can be written as

$$\widehat{A} \in \arg\min_{A \in \mathbb{R}^{d\times d}} \frac{1}{n}\sum_{i=1}^{n} l(X_i,Y_i,A) \quad \mathrm{s.t.}\quad Y_i-AX_i\in W,\ i=1,\dots,n,$$

and when $l(\cdot)\equiv 0$ the problem reduces to pure feasibility (Hespanhol et al., 2019). The same set-valued logic appears in online linear time-varying errors-in-variables identification, where the estimator returns parameter uncertainty intervals

$$\underline{\theta}_{k}(t)=\min_{\theta(t)\in D_\theta(t)} \theta_k(t),\qquad \overline{\theta}_{k}(t)=\max_{\theta(t)\in D_\theta(t)} \theta_k(t),$$

rather than a single point estimate (Fosson et al., 2021).

A canonical bounded-noise linear-regressor formulation is the one-step-ahead model

$y_k=\phi_k^\top \theta+e_k,$

or, in the notation of a linear discrete-time LTI SISO predictor,

$y(k+1)=\theta^\top \hat\phi(k)+\xi(k)+d(k+1),$

with $|d(k)|\le \bar d$ and $\theta\in\Omega$ (D'Amico et al., 2022). The smallest residual bound compatible with the data is

$$\underline\lambda= \min_{\theta\in\Omega,\ \lambda\ge 0}\lambda \quad \text{s.t.}\quad |y(k+1)-\theta^\top \hat\phi(k)|\le \lambda+\bar d,$$

and the corresponding inflated FPS is

$$\Theta(\alpha)= \left\{ \theta\in\Omega: |y(k+1)-\theta^\top \hat\phi(k)|\le \alpha\underline\lambda+\bar d,\ \forall k \right\},$$

or, equivalently,

$$\Theta(\alpha)=\{\theta\in\Omega:\|Y-\Phi\theta\|_\infty\le \epsilon\},\qquad \epsilon=\alpha\underline\lambda+\bar d$$

(D'Amico et al., 2022).

The same conceptual template extends beyond deterministic bounded support. For nonlinear systems represented in lifted coordinates,

$l(\cdot)\equiv 0$0

with i.i.d. zero-mean isotropic sub-Gaussian noise, the finite-sample uncertainty set is defined as

$l(\cdot)\equiv 0$1

thereby replacing deterministic support bounds with a high-probability sample-covariance constraint (Brändle et al., 1 Apr 2026). This broadens SMI from a purely worst-case bounded-error formalism to a hybrid deterministic-probabilistic uncertainty calculus.

2. Set representations and geometry

The geometry of SMI is determined by how consistency constraints are encoded. For bounded additive noise with linear regressors, each absolute-value inequality induces a slab, and their intersection yields a convex polyhedron. For other problem classes, one finds zonotopes, ellipsoids, quadratic-matrix-inequality regions, and interval products.

Representation	Defining form	Representative use
Polyhedral FPS	$l(\cdot)\equiv 0$2	Linear bounded-noise regression and robust LMI design (D'Amico et al., 2022)
Zonotopic AFSS	$l(\cdot)\equiv 0$3	Recursive outer-approximation under additive and multiplicative uncertainty (Wang et al., 2017)
Parameter uncertainty intervals	$l(\cdot)\equiv 0$4	Online LTV EIV identification (Fosson et al., 2021)
QMI uncertainty set	Convex QMI centered at $l(\cdot)\equiv 0$5	Stochastic SME for nonlinear lifts (Brändle et al., 1 Apr 2026)
Ellipsoidal state set	$l(\cdot)\equiv 0$6	Event-triggered set-membership state estimation (Zheng et al., 2022)

For the polyhedral case, each constraint

$l(\cdot)\equiv 0$7

is equivalent to

$l(\cdot)\equiv 0$8

so that

$l(\cdot)\equiv 0$9

When $$\underline{\theta}_{k}(t)=\min_{\theta(t)\in D_\theta(t)} \theta_k(t),\qquad \overline{\theta}_{k}(t)=\max_{\theta(t)\in D_\theta(t)} \theta_k(t),$$0 is itself polyhedral, convexity is preserved, and vertex enumeration provides a bridge to polytopic robust control design (D'Amico et al., 2022).

Zonotopic representations arise when exact FSS propagation is intractable. In the linear regression model with additive and multiplicative uncertainties, the FSS is recursively updated as

$$\underline{\theta}_{k}(t)=\min_{\theta(t)\in D_\theta(t)} \theta_k(t),\qquad \overline{\theta}_{k}(t)=\max_{\theta(t)\in D_\theta(t)} \theta_k(t),$$1

and an Approximated FSS is maintained as a zonotope

$$\underline{\theta}_{k}(t)=\min_{\theta(t)\in D_\theta(t)} \theta_k(t),\qquad \overline{\theta}_{k}(t)=\max_{\theta(t)\in D_\theta(t)} \theta_k(t),$$2

because zonotopes are closed under affine maps and Minkowski sums and admit efficient outer-approximation rules for intersections with strips or polyhedra (Wang et al., 2017).

In stochastic SME, the geometry becomes QMI-based. Substituting $$\underline{\theta}_{k}(t)=\min_{\theta(t)\in D_\theta(t)} \theta_k(t),\qquad \overline{\theta}_{k}(t)=\max_{\theta(t)\in D_\theta(t)} \theta_k(t),$$3 into the sample-covariance constraint yields a convex uncertainty set centered at $$\underline{\theta}_{k}(t)=\min_{\theta(t)\in D_\theta(t)} \theta_k(t),\qquad \overline{\theta}_{k}(t)=\max_{\theta(t)\in D_\theta(t)} \theta_k(t),$$4,

$$\underline{\theta}_{k}(t)=\min_{\theta(t)\in D_\theta(t)} \theta_k(t),\qquad \overline{\theta}_{k}(t)=\max_{\theta(t)\in D_\theta(t)} \theta_k(t),$$5

which is ellipsoidal in $$\underline{\theta}_{k}(t)=\min_{\theta(t)\in D_\theta(t)} \theta_k(t),\qquad \overline{\theta}_{k}(t)=\max_{\theta(t)\in D_\theta(t)} \theta_k(t),$$6 (Brändle et al., 1 Apr 2026).

State-estimation variants use ellipsoidal feasible sets rather than parameter polytopes. In event-triggered observability, the observer fuses a prediction ellipsoid with an ellipsoid inferred from transmitted measurements and no-event conditions, producing a posterior set $$\underline{\theta}_{k}(t)=\min_{\theta(t)\in D_\theta(t)} \theta_k(t),\qquad \overline{\theta}_{k}(t)=\max_{\theta(t)\in D_\theta(t)} \theta_k(t),$$7 that outer-approximates the intersection and is optimal in the sense of trace at each step (Zheng et al., 2022).

3. Deterministic, probabilistic, and asymptotic guarantees

A central reason for using SMI is that its guarantees are typically explicit. In the switched linear autonomous setting with bounded noise and nonsequential measurements, the feasibility estimator

$$\underline{\theta}_{k}(t)=\min_{\theta(t)\in D_\theta(t)} \theta_k(t),\qquad \overline{\theta}_{k}(t)=\max_{\theta(t)\in D_\theta(t)} \theta_k(t),$$8

is strongly consistent: $$\underline{\theta}_{k}(t)=\min_{\theta(t)\in D_\theta(t)} \theta_k(t),\qquad \overline{\theta}_{k}(t)=\max_{\theta(t)\in D_\theta(t)} \theta_k(t),$$9 The same strong consistency extends to minimizing any continuous loss over the feasible set. This result is notable because classical proofs for unstable linear systems with sequential measurements do not apply when grouped measurements are drawn from a single trajectory with arbitrary switching (Hespanhol et al., 2019).

For finite-data bounded-noise regression, deterministic consistency is often combined with probabilistic coverage. In the scenario-based inflation procedure for $y_k=\phi_k^\top \theta+e_k,$0, one chooses $y_k=\phi_k^\top \theta+e_k,$1 so that

$y_k=\phi_k^\top \theta+e_k,$2

Then, with probability at least $y_k=\phi_k^\top \theta+e_k,$3 over the sampled scenarios,

$y_k=\phi_k^\top \theta+e_k,$4

In the reported examples with $y_k=\phi_k^\top \theta+e_k,$5, $y_k=\phi_k^\top \theta+e_k,$6, $y_k=\phi_k^\top \theta+e_k,$7, and $y_k=\phi_k^\top \theta+e_k,$8, the method produced $y_k=\phi_k^\top \theta+e_k,$9 with empirical violation rate $y(k+1)=\theta^\top \hat\phi(k)+\xi(k)+d(k+1),$0 in a minimum-phase example and $y(k+1)=\theta^\top \hat\phi(k)+\xi(k)+d(k+1),$1 with empirical violation rate $y(k+1)=\theta^\top \hat\phi(k)+\xi(k)+d(k+1),$2 in a non-minimum-phase example, both below the $y(k+1)=\theta^\top \hat\phi(k)+\xi(k)+d(k+1),$3 target (D'Amico et al., 2022).

Under sub-Gaussian noise with unbounded support, the guarantee becomes finite-sample and high-probability. If

$y(k+1)=\theta^\top \hat\phi(k)+\xi(k)+d(k+1),$4

then

$y(k+1)=\theta^\top \hat\phi(k)+\xi(k)+d(k+1),$5

Moreover, with probability at least $y(k+1)=\theta^\top \hat\phi(k)+\xi(k)+d(k+1),$6, if

$y(k+1)=\theta^\top \hat\phi(k)+\xi(k)+d(k+1),$7

then

$y(k+1)=\theta^\top \hat\phi(k)+\xi(k)+d(k+1),$8

whereas variance overestimation yields a non-singleton limit set containing $y(k+1)=\theta^\top \hat\phi(k)+\xi(k)+d(k+1),$9, and variance underestimation can make the limit set empty (Brändle et al., 1 Apr 2026). This introduces a diagnostic role for emptiness that is absent from classical bounded-support SME.

Guarantees also exist at the level of simulation accuracy and state-set boundedness. For one-step-ahead linear models with bounded measurement noise, the estimated finite-horizon error bound $|d(k)|\le \bar d$0 leads to an explicit infinite-horizon simulation bound. If

$|d(k)|\le \bar d$1

then

$|d(k)|\le \bar d$2

and the identified ARX model is asymptotically stable (Lauricella et al., 2020). In event-triggered state estimation, if $|d(k)|\le \bar d$3, the posterior ellipsoids satisfy an asymptotic bound of the form

$|d(k)|\le \bar d$4

so the estimation sets remain asymptotically bounded (Zheng et al., 2022).

4. Algorithms and computational realizations

SMI has produced a diverse algorithmic toolbox, ranging from LP-based online recursions to sparse semidefinite relaxations. In scenario-inflated polyhedral FPS construction, each scenario requires one open-loop simulation of length $|d(k)|\le \bar d$5 and one LP with variables $|d(k)|\le \bar d$6 and $|d(k)|\le \bar d$7, followed by constraint removal and vertex enumeration. The total cost is $|d(k)|\le \bar d$8, with $|d(k)|\le \bar d$9 chosen by the binomial bound; in the reported examples $\theta\in\Omega$0, while vertex counts grew to around $\theta\in\Omega$1 (D'Amico et al., 2022).

Recursive zonotopic SMI replaces exact feasible-set propagation with guaranteed outer approximations. The CAZI algorithm processes one measurement at a time by intersecting a zonotope with support strips obtained via LPs, whereas PAZI processes mini-batches by intersecting a zonotope with a strip-polyhedron and selecting the outer approximation through an LMI that contracts a $\theta\in\Omega$2-radius metric. Both algorithms preserve the containment $\theta\in\Omega$3, with CAZI generally lighter computationally and PAZI potentially tighter when batch information is informative (Wang et al., 2017).

In online LTV EIV identification, the main difficulty is that bounded input and output noise create bilinear terms such as $\theta\in\Omega$4 and $\theta\in\Omega$5. Introducing auxiliary product variables and replacing bilinear equalities by McCormick envelopes yields LPs that are globally optimal for the target PUI bounds. The method solves $\theta\in\Omega$6 LPs per time step, and the reported average CPU times are approximately $\theta\in\Omega$7 ms per recursion in a first-order example and approximately $\theta\in\Omega$8 ms in a second-order example (Fosson et al., 2021).

Continuous-time MIMO SMI via Tustin discretization attacks the derivative-estimation problem by replacing

$\theta\in\Omega$9

and lifting the resulting sampled-data constraints into a polynomial optimization problem. The relaxation is then handled by sparse SOS/SDP hierarchies, specifically correlative sparsity and term sparsity through CS-TSSOS. In a second-order SISO example, the Tustin-based method matched bounds from a model-transformation approach within $$\underline\lambda= \min_{\theta\in\Omega,\ \lambda\ge 0}\lambda \quad \text{s.t.}\quad |y(k+1)-\theta^\top \hat\phi(k)|\le \lambda+\bar d,$$0 while running in $$\underline\lambda= \min_{\theta\in\Omega,\ \lambda\ge 0}\lambda \quad \text{s.t.}\quad |y(k+1)-\theta^\top \hat\phi(k)|\le \lambda+\bar d,$$1 s versus $$\underline\lambda= \min_{\theta\in\Omega,\ \lambda\ge 0}\lambda \quad \text{s.t.}\quad |y(k+1)-\theta^\top \hat\phi(k)|\le \lambda+\bar d,$$2 s, with approximately $$\underline\lambda= \min_{\theta\in\Omega,\ \lambda\ge 0}\lambda \quad \text{s.t.}\quad |y(k+1)-\theta^\top \hat\phi(k)|\le \lambda+\bar d,$$3 memory reduction. In a third-order SISO example, the Chebyshev center achieved average relative error $$\underline\lambda= \min_{\theta\in\Omega,\ \lambda\ge 0}\lambda \quad \text{s.t.}\quad |y(k+1)-\theta^\top \hat\phi(k)|\le \lambda+\bar d,$$4 versus $$\underline\lambda= \min_{\theta\in\Omega,\ \lambda\ge 0}\lambda \quad \text{s.t.}\quad |y(k+1)-\theta^\top \hat\phi(k)|\le \lambda+\bar d,$$5 for SRIVC, while the experimental TISO example reported FIT values in the $$\underline\lambda= \min_{\theta\in\Omega,\ \lambda\ge 0}\lambda \quad \text{s.t.}\quad |y(k+1)-\theta^\top \hat\phi(k)|\le \lambda+\bar d,$$6 range (Cerone et al., 26 Aug 2025).

These algorithmic realizations illustrate a recurrent pattern: exact feasible-set computation is often replaced by a computationally tractable surrogate that preserves containment, while additional structure—scenario sampling, strip intersections, McCormick envelopes, or sparsity-exploiting SOS relaxations—is used to recover practical scalability.

5. Integration with robust control and experiment design

A major contemporary role of SMI is not only to estimate models but to mediate between data and robustness. In VRFT-based controller synthesis for linear discrete-time SISO systems, the set-membership step constructs a polyhedral uncertainty set $$\underline\lambda= \min_{\theta\in\Omega,\ \lambda\ge 0}\lambda \quad \text{s.t.}\quad |y(k+1)-\theta^\top \hat\phi(k)|\le \lambda+\bar d,$$7, maps it into a polytopic state-space model, and uses it solely to impose robust stability constraints during controller synthesis. Performance optimization remains fully data-driven via VRFT, while robust quadratic stability is enforced by LMIs over all vertices of the identified polytope (D'Amico et al., 2022).

In adaptive learning-based MPC for interconnected systems, SMI is used online to learn the uncertain coupling strengths $$\underline\lambda= \min_{\theta\in\Omega,\ \lambda\ge 0}\lambda \quad \text{s.t.}\quad |y(k+1)-\theta^\top \hat\phi(k)|\le \lambda+\bar d,$$8. At each time step, subsystem $$\underline\lambda= \min_{\theta\in\Omega,\ \lambda\ge 0}\lambda \quad \text{s.t.}\quad |y(k+1)-\theta^\top \hat\phi(k)|\le \lambda+\bar d,$$9 solves $$\Theta(\alpha)= \left\{ \theta\in\Omega: |y(k+1)-\theta^\top \hat\phi(k)|\le \alpha\underline\lambda+\bar d,\ \forall k \right\},$$0 LPs to update the interval bounds $$\Theta(\alpha)= \left\{ \theta\in\Omega: |y(k+1)-\theta^\top \hat\phi(k)|\le \alpha\underline\lambda+\bar d,\ \forall k \right\},$$1 and $$\Theta(\alpha)= \left\{ \theta\in\Omega: |y(k+1)-\theta^\top \hat\phi(k)|\le \alpha\underline\lambda+\bar d,\ \forall k \right\},$$2, producing a monotone shrinkage

$$\Theta(\alpha)= \left\{ \theta\in\Omega: |y(k+1)-\theta^\top \hat\phi(k)|\le \alpha\underline\lambda+\bar d,\ \forall k \right\},$$3

The learned disturbance set is then embedded into tube-based robust MPC, where the prestabilizing controller, rigid tube, tightened constraints, and terminal ingredients are adapted online; recursive feasibility and closed-loop ISS are established under the paper’s passivity and diagonal-dominance conditions (Aboudonia et al., 2024).

A related but distinct development embeds the set-membership equations directly inside adaptive MPC through strong duality. For

$$\Theta(\alpha)= \left\{ \theta\in\Omega: |y(k+1)-\theta^\top \hat\phi(k)|\le \alpha\underline\lambda+\bar d,\ \forall k \right\},$$4

suprema over $$\Theta(\alpha)= \left\{ \theta\in\Omega: |y(k+1)-\theta^\top \hat\phi(k)|\le \alpha\underline\lambda+\bar d,\ \forall k \right\},$$5 are replaced by dual variables, and predicted set updates are written exactly inside the MPC problem. This exact reformulation enables a predicted worst-case cost defined over a predicted state tube, so that the optimizer can trade control performance against future information gain while preserving robust constraint satisfaction and recursive feasibility (Parsi et al., 2022).

Experiment design pushes SMI one step earlier, from estimation after data collection to input construction before data collection. An input is called universal for identification if, when applied to any system complying with the prior knowledge, it yields data suitable for accurate identification. For controllable priors, persistently exciting inputs of order $$\Theta(\alpha)= \left\{ \theta\in\Omega: |y(k+1)-\theta^\top \hat\phi(k)|\le \alpha\underline\lambda+\bar d,\ \forall k \right\},$$6 recover the familiar fundamental-lemma route. However, other forms of prior knowledge permit non-PE universal inputs with markedly lower sample requirements. In the relative orbital motion example, a two-impulse input with horizon $$\Theta(\alpha)= \left\{ \theta\in\Omega: |y(k+1)-\theta^\top \hat\phi(k)|\le \alpha\underline\lambda+\bar d,\ \forall k \right\},$$7 is universal for exact identification although PE of order $$\Theta(\alpha)= \left\{ \theta\in\Omega: |y(k+1)-\theta^\top \hat\phi(k)|\le \alpha\underline\lambda+\bar d,\ \forall k \right\},$$8 would require $$\Theta(\alpha)= \left\{ \theta\in\Omega: |y(k+1)-\theta^\top \hat\phi(k)|\le \alpha\underline\lambda+\bar d,\ \forall k \right\},$$9. In a network example with $$\Theta(\alpha)=\{\theta\in\Omega:\|Y-\Phi\theta\|_\infty\le \epsilon\},\qquad \epsilon=\alpha\underline\lambda+\bar d$$0 identical systems, a noisy $$\Theta(\alpha)=\{\theta\in\Omega:\|Y-\Phi\theta\|_\infty\le \epsilon\},\qquad \epsilon=\alpha\underline\lambda+\bar d$$1-accuracy design was achieved with $$\Theta(\alpha)=\{\theta\in\Omega:\|Y-\Phi\theta\|_\infty\le \epsilon\},\qquad \epsilon=\alpha\underline\lambda+\bar d$$2, contrasted with $$\Theta(\alpha)=\{\theta\in\Omega:\|Y-\Phi\theta\|_\infty\le \epsilon\},\qquad \epsilon=\alpha\underline\lambda+\bar d$$3 samples for a robust fundamental-lemma-based design. Exact identification under noise is possible only when the prior set is uniformly discrete (Shakouri et al., 1 Jul 2026).

6. Scope, limitations, and recurrent misconceptions

A recurrent misconception is that SMI is synonymous with deterministic bounded-support noise. That view matches much of the classical literature, including bounded additive noise, multiplicative uncertainty, and polyhedral consistency sets, but it is no longer exhaustive. Stochastic SME for nonlinear systems with i.i.d. sub-Gaussian isotropic noise replaces deterministic support bounds with high-probability covariance constraints while preserving the uncertainty-set perspective (Brändle et al., 1 Apr 2026).

Another misconception is that SMI is intrinsically a point-estimation methodology. In fact, the primary object is usually a set: an FPS, an FSS, a PUI family, a zonotopic AFSS, or an ellipsoidal state set. Point estimates such as the Chebyshev center, continuous-loss minimizers over the feasible set, or central interval estimates are secondary summaries extracted from that set (Fosson et al., 2021, Cerone et al., 26 Aug 2025).

The principal limitations are structural and computational. Many formulations assume known or estimable order, valid a priori bounds, and sufficient excitation. In bounded-noise FPS construction, finite data can underestimate the true residual bound, motivating inflation parameters and scenario calibration (D'Amico et al., 2022). In guaranteed-simulation SMI, conservative inflation factors $$\Theta(\alpha)=\{\theta\in\Omega:\|Y-\Phi\theta\|_\infty\le \epsilon\},\qquad \epsilon=\alpha\underline\lambda+\bar d$$4 and $$\Theta(\alpha)=\{\theta\in\Omega:\|Y-\Phi\theta\|_\infty\le \epsilon\},\qquad \epsilon=\alpha\underline\lambda+\bar d$$5 enlarge feasible sets and error bounds, while underestimated noise bounds can invalidate the guarantees (Lauricella et al., 2020). In continuous-time Tustin-based identification, inaccurate discretization-error bounds can destroy feasibility (Cerone et al., 26 Aug 2025). In adaptive MPC for interconnected systems, no persistent excitation is required for safety, but excitation still affects shrink speed rather than soundness (Aboudonia et al., 2024).

Computationally, exact vertex enumeration may become prohibitive; PAZI’s LMI can become infeasible when conditioning is poor or uncertainty is large; common quadratic Lyapunov certificates are conservative; and sparse SOS relaxations, although greatly reduced, remain heavy for high-order or large-scale problems (Wang et al., 2017, D'Amico et al., 2022, Cerone et al., 26 Aug 2025). Experiment design adds a conceptual limitation: persistent excitation is not universally necessary, but for single-input open controllable priors it is necessary and sufficient in the noise-free case, whereas exact identification under noise requires uniformly discrete prior sets (Shakouri et al., 1 Jul 2026).

Taken together, these developments portray set-membership identification not as a single algorithm but as a methodological family organized around consistency sets, outer approximations, and explicit guarantees. Its distinctive feature is that uncertainty is represented geometrically and manipulated directly, whether the objective is strong consistency under nonsequential switching, online EIV tracking, probabilistically calibrated robust control, event-triggered state estimation, continuous-time identification from corrupted samples, or finite-horizon experiment design.