Set-Theoretic Model Reference Adaptive Control

• Set-theoretic MRAC is a design framework where adaptation is governed by explicit set membership conditions to ensure safety, feasibility, and bounded tracking error.
• It includes formulations for continuous LTI systems with constraints, nonlinear safety-critical systems with prescribed performance, and discrete-time data-driven approaches.
• The methodology employs projection, barrier Lyapunov functions, and safety tubes to guarantee forward invariance and convergence under system uncertainties.

Searching arXiv for the cited works and closely related papers on set-theoretic MRAC. arXiv search query: all:"Set-Theoretic Model Reference Adaptive Control" OR id:([2508.21584](/papers/2508.21584)) OR id:([1909.07916](/papers/1909.07916)) OR id:([2109.10611](/papers/2109.10611)) OR id:([2502.21091](/papers/2502.21091)) Set-theoretic model reference adaptive control denotes a family of MRAC formulations in which the control objective is posed as membership in explicitly defined sets rather than as asymptotic tracking alone. In the recent literature, these sets take several forms: safe state and input sets for uncertain linear systems with disturbances, prescribed tubes around nonlinear reference trajectories, compact convex parameter-estimate sets imposed by projection algorithms, and data-consistent model sets used to characterize when adaptive gains can converge to solutions of the matching equations (Ghosh et al., 29 Aug 2025, Arabi et al., 2019, Shahab et al., 2021, Wang et al., 28 Feb 2025). Across these formulations, the central theme is that adaptation is constrained or shaped by set membership, so that forward invariance, feasibility, boundedness, or convergence can be certified directly in set-theoretic terms.

1. Definition, scope, and unifying viewpoint

In MRAC, the plant is required to track a reference model by means of feedback gains that are updated online. The classical matching equations appear repeatedly across the literature. For the continuous-time constrained LTI setting, the plant and reference model are

$$\dot{x}(t)=Ax(t)+Bu(t)+d(t), \qquad \dot{x}_r(t)=A_r x_r(t)+B_r r(t),$$

with tracking error $e(t)\triangleq x(t)-x_r(t)$, and the standard matching assumption is

$A+BK_x=A_r,\qquad BK_r=B_r$

for some constant controller parameters $K_x,K_r$ (Ghosh et al., 29 Aug 2025). In the discrete-time informativity framework, the unknown plant and the known reference model are

$$x(t+1)=A_{\rm s}x(t)+B_{\rm s}u(t),\qquad x_{\rm m}(t+1)=A_{\rm m}x_{\rm m}(t)+B_{\rm m}r(t),$$

and state-feedback MRAC seeks gains $K,L$ satisfying

$$A_{\rm s}+B_{\rm s}K=A_{\rm m},\qquad B_{\rm s}L=B_{\rm m},$$

so that the error evolves as $e(t+1)=A_{\rm m}e(t)$ (Wang et al., 28 Feb 2025).

What makes these approaches “set-theoretic” is not a single controller structure but the use of sets as the primary objects of analysis. In one line of work, the relevant sets are safe state and input constraint sets and an invariant tracking-error set (Ghosh et al., 29 Aug 2025). In another, they are a safe state set $S_s$ and a prescribed performance tube $\mathcal{T}(t)$ around a nonlinear reference trajectory (Arabi et al., 2019). In a third, they are compact convex parameter sets that constrain adaptive estimates through Euclidean projection (Shahab et al., 2021). In a fourth, they are the data-consistent system set $e(t)\triangleq x(t)-x_r(t)$0, the matching set $e(t)\triangleq x(t)-x_r(t)$1, and the feasible parameter set $e(t)\triangleq x(t)-x_r(t)$2 induced by measured trajectories (Wang et al., 28 Feb 2025).

Formulation	Core set object	Principal guarantee
Constrained continuous-time MRAC	$e(t)\triangleq x(t)-x_r(t)$3	Forward invariance, input admissibility, UUB tracking
Nonlinear safety-critical MRAC	$e(t)\triangleq x(t)-x_r(t)$4	$e(t)\triangleq x(t)-x_r(t)$5 and $e(t)\triangleq x(t)-x_r(t)$6
Projection-based discrete-time MRAC	Compact convex $e(t)\triangleq x(t)-x_r(t)$7	Bounded estimates and linear-like closed-loop behavior
Informativity-based MRAC	$e(t)\triangleq x(t)-x_r(t)$8	Gain convergence iff informative time is finite

This breadth matters because the phrase “set-theoretic MRAC” can otherwise be misunderstood as referring only to safe-set invariance. The cited literature uses the term more broadly: sets may encode safety envelopes, admissible parameter regions, or information contained in data (Ghosh et al., 29 Aug 2025, Shahab et al., 2021, Wang et al., 28 Feb 2025).

2. State and input constrained formulations for uncertain LTI plants

A particularly explicit set-theoretic formulation is developed for uncertain LTI plants with unmatched bounded disturbances and user-defined state and input constraints (Ghosh et al., 29 Aug 2025). The state and input sets are

$e(t)\triangleq x(t)-x_r(t)$9

and the safety objectives are stated as forward invariance of $A+BK_x=A_r,\qquad BK_r=B_r$0, input admissibility in $A+BK_x=A_r,\qquad BK_r=B_r$1, and robustness to unmatched disturbances satisfying $A+BK_x=A_r,\qquad BK_r=B_r$2. The controller is optimization-free and combines a certainty-equivalence MRAC auxiliary law,

$A+BK_x=A_r,\qquad BK_r=B_r$3

with a norm-based saturation map $A+BK_x=A_r,\qquad BK_r=B_r$4, and with a barrier Lyapunov function that acts directly on the tracking error.

The key set-theoretic construction is an auxiliary reference model $A+BK_x=A_r,\qquad BK_r=B_r$5, obtained by saturating the reference trajectory inside a smaller ball of radius $A+BK_x=A_r,\qquad BK_r=B_r$6. This induces the error bound

$A+BK_x=A_r,\qquad BK_r=B_r$7

An ellipsoidal safe error set is then defined as

$A+BK_x=A_r,\qquad BK_r=B_r$8

and the barrier Lyapunov function

$A+BK_x=A_r,\qquad BK_r=B_r$9

diverges at the boundary, which is the mechanism used to render the set forward invariant (Ghosh et al., 29 Aug 2025).

The adaptive law uses projection on compact convex parameter sets,

$K_x,K_r$0

with update laws

$K_x,K_r$1

Projection is used as the robustness mechanism against saturation and unmatched disturbances; the construction explicitly states that no $K_x,K_r$2- or $K_x,K_r$3-modification is needed (Ghosh et al., 29 Aug 2025).

The main theorem couples constraint satisfaction and feasibility. Under the matching conditions, parameter bounds, an initial error $K_x,K_r$4, and the sufficient feasibility condition

$K_x,K_r$5

the closed loop guarantees state safety, input admissibility, bounded closed-loop signals, and uniformly ultimately bounded tracking error (Ghosh et al., 29 Aug 2025). The same paper rewrites the condition as

$K_x,K_r$6

where $K_x,K_r$7 and $K_x,K_r$8, thereby turning feasibility into an affine relation between the allowable state and input bounds. This provides an a priori feasibility check before implementation.

This formulation is notable for the specific conjunction of ingredients it achieves. One paper states that, to the best of its authors’ knowledge, it is the first result that considers both state and input constraints for control of uncertain systems with disturbances and provides sufficient feasibility conditions to check for the existence of an admissible control policy (Ghosh et al., 29 Aug 2025).

3. Nonlinear reference models, prescribed performance, and safety tubes

A second major formulation addresses uncertain nonlinear systems by introducing a nonlinear reference model that represents ideal and safe behavior (Arabi et al., 2019). The plant is

$K_x,K_r$9

while the reference system is

$$x(t+1)=A_{\rm s}x(t)+B_{\rm s}u(t),\qquad x_{\rm m}(t+1)=A_{\rm m}x_{\rm m}(t)+B_{\rm m}r(t),$$0

Here the safe set is denoted $$x(t+1)=A_{\rm s}x(t)+B_{\rm s}u(t),\qquad x_{\rm m}(t+1)=A_{\rm m}x_{\rm m}(t)+B_{\rm m}r(t),$$1, and the reference trajectory is designed to evolve in $$x(t+1)=A_{\rm s}x(t)+B_{\rm s}u(t),\qquad x_{\rm m}(t+1)=A_{\rm m}x_{\rm m}(t)+B_{\rm m}r(t),$$2. Safety is enforced by requiring the plant state to remain within a prescribed distance of the reference trajectory.

The induced set-theoretic object is the time-varying performance tube

$$x(t+1)=A_{\rm s}x(t)+B_{\rm s}u(t),\qquad x_{\rm m}(t+1)=A_{\rm m}x_{\rm m}(t)+B_{\rm m}r(t),$$3

If $$x(t+1)=A_{\rm s}x(t)+B_{\rm s}u(t),\qquad x_{\rm m}(t+1)=A_{\rm m}x_{\rm m}(t)+B_{\rm m}r(t),$$4 for all $$x(t+1)=A_{\rm s}x(t)+B_{\rm s}u(t),\qquad x_{\rm m}(t+1)=A_{\rm m}x_{\rm m}(t)+B_{\rm m}r(t),$$5, then $$x(t+1)=A_{\rm s}x(t)+B_{\rm s}u(t),\qquad x_{\rm m}(t+1)=A_{\rm m}x_{\rm m}(t)+B_{\rm m}r(t),$$6, so forward invariance of $$x(t+1)=A_{\rm s}x(t)+B_{\rm s}u(t),\qquad x_{\rm m}(t+1)=A_{\rm m}x_{\rm m}(t)+B_{\rm m}r(t),$$7 is equivalent to safety. When only a reference set $$x(t+1)=A_{\rm s}x(t)+B_{\rm s}u(t),\qquad x_{\rm m}(t+1)=A_{\rm m}x_{\rm m}(t)+B_{\rm m}r(t),$$8 is known, the construction reduces to a constant tube width $$x(t+1)=A_{\rm s}x(t)+B_{\rm s}u(t),\qquad x_{\rm m}(t+1)=A_{\rm m}x_{\rm m}(t)+B_{\rm m}r(t),$$9 (Arabi et al., 2019).

The barrier mechanism is expressed through the mapping

$K,L$0

where $K,L$1 is the Lyapunov function associated with the nominal exponentially stable error dynamics. Maintaining $K,L$2 guarantees $K,L$3, hence $K,L$4. The adaptive control law is

$K,L$5

and the projection-based update law is

$K,L$6

The factor $K,L$7 is the distinctive feature: as $K,L$8, the effective adaptation rate increases, while deeper inside the tube it decreases. The paper presents this as the mechanism that removes ad-hoc tuning of the adaptation rate $K,L$9 for safety enforcement (Arabi et al., 2019).

The corresponding Lyapunov-like function is

$$A_{\rm s}+B_{\rm s}K=A_{\rm m},\qquad B_{\rm s}L=B_{\rm m},$$0

Under the stated assumptions, including the matching structure $$A_{\rm s}+B_{\rm s}K=A_{\rm m},\qquad B_{\rm s}L=B_{\rm m},$$1 and a nominal controller that yields exponential error stability, the closed-loop trajectories are bounded and $$A_{\rm s}+B_{\rm s}K=A_{\rm m},\qquad B_{\rm s}L=B_{\rm m},$$2 for all $$A_{\rm s}+B_{\rm s}K=A_{\rm m},\qquad B_{\rm s}L=B_{\rm m},$$3; equivalently, $$A_{\rm s}+B_{\rm s}K=A_{\rm m},\qquad B_{\rm s}L=B_{\rm m},$$4 for all $$A_{\rm s}+B_{\rm s}K=A_{\rm m},\qquad B_{\rm s}L=B_{\rm m},$$5 (Arabi et al., 2019).

This formulation clarifies an important point about set-theoretic MRAC. The “set” need not be fixed. In this nonlinear setting, it is a moving safety tube around a nonlinear reference trajectory, and the adaptive law is explicitly barrier-shaped to preserve forward invariance of that tube. In the linear special case, the construction reduces to earlier set-theoretic MRAC results of Arabi et al. 2017 and 2018, where generalized restricted potential functions are recovered (Arabi et al., 2019).

4. Projection onto compact convex parameter sets and linear-like behavior

Another use of the set-theoretic idea appears in discrete-time SISO MRAC for LTI systems with input delay, where the defining set is a known compact convex region containing the predictor parameters (Shahab et al., 2021). The plant is modeled as

$$A_{\rm s}+B_{\rm s}K=A_{\rm m},\qquad B_{\rm s}L=B_{\rm m},$$6

with a stable reference model defined by polynomials $$A_{\rm s}+B_{\rm s}K=A_{\rm m},\qquad B_{\rm s}L=B_{\rm m},$$7 and $$A_{\rm s}+B_{\rm s}K=A_{\rm m},\qquad B_{\rm s}L=B_{\rm m},$$8. After predictor-form manipulation, the weighted output obeys

$$A_{\rm s}+B_{\rm s}K=A_{\rm m},\qquad B_{\rm s}L=B_{\rm m},$$9

with regressor $e(t+1)=A_{\rm m}e(t)$0 and predictor parameter vector $e(t+1)=A_{\rm m}e(t)$1 (Shahab et al., 2021).

The set-theoretic restriction is the choice of a compact convex set $e(t+1)=A_{\rm m}e(t)$2 containing the image of the admissible plant parameter set. A hyperrectangle is given as a practical example. The crucial condition is that the $e(t+1)=A_{\rm m}e(t)$3-th coordinate, corresponding to $e(t+1)=A_{\rm m}e(t)$4, is never zero on $e(t+1)=A_{\rm m}e(t)$5; this guarantees that the certainty-equivalence control law is well defined. The parameter update consists of a tentative projection step and an Euclidean projection onto $e(t+1)=A_{\rm m}e(t)$6: $e(t+1)=A_{\rm m}e(t)$7 The vigilance gate

$e(t+1)=A_{\rm m}e(t)$8

turns off estimation when noise dominates (Shahab et al., 2021).

The control law enforces the predictor-form reference equality

$e(t+1)=A_{\rm m}e(t)$9

which yields the explicit MRAC input

$S_s$0

In this setting, “set-theoretic MRAC” refers to the fact that the estimator is confined to $S_s$1, not to the imposition of state constraints (Shahab et al., 2021).

The main result is a linear-like closed-loop bound. For any $S_s$2 and any $S_s$3, there exists $S_s$4 such that

$S_s$5

If $S_s$6, then

$S_s$7

The consequences stated in the paper are exponential stability of transients, bounded $S_s$8 gains for every $S_s$9, and convolution bounds on exogenous inputs (Shahab et al., 2021).

This result addresses a common misconception that projection in MRAC merely prevents parameter blow-up. In this framework, projection onto a compact convex set, together with the vigilance gate, is what permits the derivation of a “good model” with uniformly exponentially stable nominal dynamics and a perturbation analysis based on Kreisselmeier’s lemma. The resulting guarantees are therefore substantially sharper than the bounded-noise bounded-state statements commonly associated with classical MRAC (Shahab et al., 2021).

5. Data informativity and the weakest online-checkable convergence conditions

A further development reframes MRAC through data informativity (Wang et al., 28 Feb 2025). Instead of asking for persistent excitation or initial excitation, the framework characterizes exactly when the observed data are sufficient for convergence of the adaptive gains to a solution of the matching equations. The plant-consistent set is

$\mathcal{T}(t)$0

and the matching set for fixed gains $\mathcal{T}(t)$1 is

$\mathcal{T}(t)$2

Data are informative for model reference control if $\mathcal{T}(t)$3 for some $\mathcal{T}(t)$4 (Wang et al., 28 Feb 2025).

The necessary and sufficient condition for this informativity is the image inclusion

$\mathcal{T}(t)$5

Equivalently, there exist matrices $\mathcal{T}(t)$6 such that

$\mathcal{T}(t)$7

The paper emphasizes that this condition is strictly weaker than informativity for system identification, which would require

$\mathcal{T}(t)$8

Moreover, informativity for system identification is necessary for informativity for model reference control only in the special case $\mathcal{T}(t)$9 and $e(t)\triangleq x(t)-x_r(t)$00 has full column rank (Wang et al., 28 Feb 2025).

The adaptive algorithm introduces the first informative time

$e(t)\triangleq x(t)-x_r(t)$01

and a feasible parameter set

$e(t)\triangleq x(t)-x_r(t)$02

The parameter matrix $e(t)\triangleq x(t)-x_r(t)$03 is updated by normalized gradient steps toward this feasible set, and the gains are constructed as

$e(t)\triangleq x(t)-x_r(t)$04

When informativity has not yet been achieved, the method injects a rank-raising input $e(t)\triangleq x(t)-x_r(t)$05 so that the data span grows (Wang et al., 28 Feb 2025).

The main convergence statement is exact: the gains satisfy

$e(t)\triangleq x(t)-x_r(t)$06

if and only if $e(t)\triangleq x(t)-x_r(t)$07 (Wang et al., 28 Feb 2025). Under the same assumptions, the state and input remain bounded and the tracking error converges to zero. This is presented as a necessary-and-sufficient characterization of gain convergence based solely on online data.

This perspective sharpens the role of excitation in MRAC. Persistent excitation, initial excitation, and finite excitation are sufficient routes to convergence because they imply richer data properties, but they are not the weakest possible requirements. The informativity condition asks only whether the collected state-transition data span the matching map associated with the reference model. Many directions may remain unexcited, provided that this inclusion holds (Wang et al., 28 Feb 2025).

6. Relations to adjacent methods, demonstrations, and limitations

The recent set-theoretic MRAC literature is closely related to, but distinct from, other safe and adaptive control paradigms. In the constrained LTI formulation, the controller is explicitly contrasted with model predictive control and control barrier function methods, which enforce constraints via online optimization and may require conservative tightening and recursive feasibility analysis under uncertainty; the proposed alternative is optimization-free and adaptive (Ghosh et al., 29 Aug 2025). In the nonlinear safety-critical formulation, invariance is achieved through Lyapunov-barrier shaping embedded in the adaptive law rather than through a CBF-CLF QP (Arabi et al., 2019). A reasonable synthesis is that set-theoretic MRAC trades online optimization for offline verifiable set constructions and online parameter adaptation.

Representative demonstrations follow the structure of the corresponding theories. For constrained continuous-time LTI control, a four-state two-input example with disturbance injected at $e(t)\triangleq x(t)-x_r(t)$08 shows that the BLF-saturated MRAC keeps $e(t)\triangleq x(t)-x_r(t)$09 and $e(t)\triangleq x(t)-x_r(t)$10, whereas a robust MRAC baseline with projection and no BLF violates the state constraint and demands higher input magnitudes (Ghosh et al., 29 Aug 2025). For nonlinear safety-critical control, a forced Van der Pol reference model is used with a safe ellipsoidal set $e(t)\triangleq x(t)-x_r(t)$11; the nominal controller alone violates safety, while the adaptive law maintains $e(t)\triangleq x(t)-x_r(t)$12 and keeps the trajectories in $e(t)\triangleq x(t)-x_r(t)$13 for $e(t)\triangleq x(t)-x_r(t)$14 (Arabi et al., 2019). For the discrete-time linear-like formulation, a time-varying plant with bounded coefficients and square-wave reference exhibits accurate tracking in noise-free intervals, temporary degradation under disturbance, and behavior consistent with the stated convolution bounds (Shahab et al., 2021).

The limitations are formulation-dependent. The constrained LTI design is local to the BLF-defined error region $e(t)\triangleq x(t)-x_r(t)$15, assumes full-state feedback and known $e(t)\triangleq x(t)-x_r(t)$16, requires matching conditions, and may be conservative because the feasibility condition is sufficient rather than necessary (Ghosh et al., 29 Aug 2025). The nonlinear safety-critical formulation requires the parametric uncertainty structure $e(t)\triangleq x(t)-x_r(t)$17, the matching form $e(t)\triangleq x(t)-x_r(t)$18, a nominal controller yielding exponential error stability, and a computable tube radius $e(t)\triangleq x(t)-x_r(t)$19; input constraints are not explicitly handled (Arabi et al., 2019). The discrete-time linear-like theory is restricted to SISO LTI plants with known delay and order bounds, minimum phase zeros, and known sign of the high-frequency gain (Shahab et al., 2021). The informativity-based convergence framework is developed for noise-free input-state data and does not extend its formal MRAC convergence guarantees to bounded-noise models (Wang et al., 28 Feb 2025).

Several extensions are identified within the literature itself. Proposed directions include time-varying state and input constraints, nonlinear plants via appropriate regressors and BLFs, output-feedback MRAC using adaptive observers, and integration with set-membership identification to update parameter bounds online and reduce conservatism (Ghosh et al., 29 Aug 2025). The informativity framework suggests another research axis: replacing overly strong excitation assumptions with online-checkable feasibility conditions derived from model-consistency sets (Wang et al., 28 Feb 2025). Taken together, these works suggest that set-theoretic MRAC is not a narrow variant of adaptive control but a general design philosophy in which safety, feasibility, robustness, or convergence are encoded as properties of explicitly constructed sets and verified through Lyapunov, projection, or data-inclusion arguments.