Robust Admissible Set

• Robust admissible sets are defined as the collection of initial states for which a control law exists to keep system constraints satisfied against all admissible disturbances.
• They are characterized by barrier dynamics, Hamiltonian saddle-point conditions, and a min-max framework that underpins their safety and invariance properties.
• Computational methods, including finite-time convergent algorithms and numerical limiters, drive practical applications across control theory, numerical MHD, and decision-making fields.

“Robust admissible set” denotes a family of set-valued constructions used to encode safety, viability, invariance, or admissibility under uncertainty. Its most literal control-theoretic meaning is the set of initial states from which state and input constraints can be satisfied for all times against all admissible disturbances (Rußwurm et al., 23 Sep 2025). Closely related objects include maximal admissible robust positive invariant sets for uncertain linear systems (Dey et al., 2024), admissible and maximal robust positively invariant sets in epidemic and power-system models (Esterhuizen et al., 2021), and, in more combinatorial or statistical settings, admissible sets whose robustness is expressed through shellability, perturbation-resilient projection, or distribution-free coverage guarantees (He et al., 6 Mar 2026).

1. Control-theoretic definition and quantifier structure

For a nonlinear control system

$$\dot{x}(t)=f(x(t),u(t),d(t)), \qquad u\in U,\ d\in D,$$

with constraint set

$$G \triangleq \{x\in \mathbb{R}^n \mid g_i(x)\le 0,\ i=1,\dots,p\},$$

the robust admissible set is defined by

$$R \triangleq \{ \bar{x} \in G \mid \exists \, u \in U : x^{(\bar{x},u,d)}(t) \in G ~ \forall d \in D, \, \forall t \geq 0 \}.$$

Thus an initial state belongs to $R$ if there exists at least one control law such that every resulting trajectory under every admissible disturbance remains in the constraint set forever. The complement relative to $G$ is

$$G\setminus R = \{ \bar{x} \in G \mid \forall u \in U, \exists d \in D, \exists \bar{t}<\infty: x^{(\bar{x},u,d)}(\bar{t})\notin G\},$$

and the same condition admits the min-max characterization

$$\bar{x} \in R \iff \min_{u \in U} \, \sup_{d \in D} \, \sup_{t \in [0,\infty)} \, \max_{i} g_i(x^{(\bar{x},u,d)}(t)) \le 0.$$

Under the additional assumption that the boundary value $0$ is attained,

$$\bar{x} \in \partial R \iff \min_{u \in U} \, \sup_{d \in D} \, \sup_{t \in [0,\infty)} \, \max_i g_i(x^{(\bar{x},u,d)}(t)) = 0$$

(Rußwurm et al., 23 Sep 2025).

Closely related literatures differ mainly by quantifier order. In power-grid transient stability, the admissible set $\mathcal A$ is a “there exists” safety set, while the maximal robust positively invariant set $$G \triangleq \{x\in \mathbb{R}^n \mid g_i(x)\le 0,\ i=1,\dots,p\},$$0 requires safety for every admissible disturbance realization (Aschenbruck et al., 2021). In constrained SIR and SEIR epidemic models, the admissible set $$G \triangleq \{x\in \mathbb{R}^n \mid g_i(x)\le 0,\ i=1,\dots,p\},$$1 consists of initial states from which there exists at least one allowable intervention keeping the infection cap satisfied forever, whereas the MRPI set $$G \triangleq \{x\in \mathbb{R}^n \mid g_i(x)\le 0,\ i=1,\dots,p\},$$2 is the largest set from which the cap is satisfied forever for every allowable control or uncertainty realization (Esterhuizen et al., 2021). This suggests that, in the control literature, robustness is governed primarily by the $$G \triangleq \{x\in \mathbb{R}^n \mid g_i(x)\le 0,\ i=1,\dots,p\},$$3 structure attached to controls, disturbances, and time.

2. Boundary decomposition, barriers, and Hamiltonian characterization

A central geometric result is that the boundary of the robust admissible set decomposes into a usable part on the state-constraint boundary and a barrier lying strictly inside the constraint set. With

$$G \triangleq \{x\in \mathbb{R}^n \mid g_i(x)\le 0,\ i=1,\dots,p\},$$4

the boundary satisfies

$$G \triangleq \{x\in \mathbb{R}^n \mid g_i(x)\le 0,\ i=1,\dots,p\},$$5

where the usable part is $$G \triangleq \{x\in \mathbb{R}^n \mid g_i(x)\le 0,\ i=1,\dots,p\},$$6 and the barrier is $$G \triangleq \{x\in \mathbb{R}^n \mid g_i(x)\le 0,\ i=1,\dots,p\},$$7. At the interface, the boundary must be tangent to the state constraints and separate the interior of the robust admissible set from its complement. The tangency set is

$$G \triangleq \{x\in \mathbb{R}^n \mid g_i(x)\le 0,\ i=1,\dots,p\},$$8

with $$G \triangleq \{x\in \mathbb{R}^n \mid g_i(x)\le 0,\ i=1,\dots,p\},$$9, and the tangent hyperplane acts as a local separator. Barrier trajectories remain on the barrier until they intersect $$R \triangleq \{ \bar{x} \in G \mid \exists \, u \in U : x^{(\bar{x},u,d)}(t) \in G ~ \forall d \in D, \, \forall t \geq 0 \}.$$0, and the barrier is semipermeable in the sense stated in the source (Rußwurm et al., 23 Sep 2025).

The same paper derives a Pontryagin-style necessary condition for barrier trajectories. For every trajectory on the relevant portion of the barrier, there exists a nonzero absolutely continuous adjoint $$R \triangleq \{ \bar{x} \in G \mid \exists \, u \in U : x^{(\bar{x},u,d)}(t) \in G ~ \forall d \in D, \, \forall t \geq 0 \}.$$1 satisfying

$$R \triangleq \{ \bar{x} \in G \mid \exists \, u \in U : x^{(\bar{x},u,d)}(t) \in G ~ \forall d \in D, \, \forall t \geq 0 \}.$$2

with terminal condition

$$R \triangleq \{ \bar{x} \in G \mid \exists \, u \in U : x^{(\bar{x},u,d)}(t) \in G ~ \forall d \in D, \, \forall t \geq 0 \}.$$3

and Hamiltonian saddle-point condition

$$R \triangleq \{ \bar{x} \in G \mid \exists \, u \in U : x^{(\bar{x},u,d)}(t) \in G ~ \forall d \in D, \, \forall t \geq 0 \}.$$4

for almost all $$R \triangleq \{ \bar{x} \in G \mid \exists \, u \in U : x^{(\bar{x},u,d)}(t) \in G ~ \forall d \in D, \, \forall t \geq 0 \}.$$5 (Rußwurm et al., 23 Sep 2025).

Barrier theory also yields exact boundary descriptions in other applications. In power systems, candidate barrier trajectories are generated by adjoint dynamics and ultimate tangentiality conditions; the admissible-set barrier uses

$$R \triangleq \{ \bar{x} \in G \mid \exists \, u \in U : x^{(\bar{x},u,d)}(t) \in G ~ \forall d \in D, \, \forall t \geq 0 \}.$$6

whereas the MRPI barrier uses

$$R \triangleq \{ \bar{x} \in G \mid \exists \, u \in U : x^{(\bar{x},u,d)}(t) \in G ~ \forall d \in D, \, \forall t \geq 0 \}.$$7

(Aschenbruck et al., 2021). In epidemic management, admissible-set barriers are computed by Hamiltonian minimization and MRPI barriers by Hamiltonian maximization, with boundary trajectories obtained by backward integration from points of ultimate tangentiality on the infection cap (Esterhuizen et al., 2021).

3. Computational constructions in control and safety

For discrete-time linear time-varying systems with parametric uncertainties and additive disturbances, the principal robust admissible object is the maximal admissible robust positive invariant set. Under a fixed feedback law $$R \triangleq \{ \bar{x} \in G \mid \exists \, u \in U : x^{(\bar{x},u,d)}(t) \in G ~ \forall d \in D, \, \forall t \geq 0 \}.$$8, the closed-loop system is

$$R \triangleq \{ \bar{x} \in G \mid \exists \, u \in U : x^{(\bar{x},u,d)}(t) \in G ~ \forall d \in D, \, \forall t \geq 0 \}.$$9

and admissibility with respect to state and input constraints is encoded by

$R$0

The MARPI set is characterized by the infinite intersection

$R$1

where the $R$2 are backward reachable sets defined recursively by finitely many vertex systems. The paper proves that a non-empty MARPI set exists if and only if

$R$3

where $R$4 is the minimal robust positively invariant set for the uncertainty/disturbance dynamics, and it gives a finite-time converging algorithm with stopping criterion

$R$5

(Dey et al., 2024).

A neighboring line of work treats saturation-induced mode changes rather than exogenous disturbance. For linear discrete-time systems with input saturation, the maximal input-saturated output-admissible set $R$6 is approximated by a polyhedral set $R$7 obtained from a partition of the state/reference space into saturated and unsaturated regions, propagation of output constraints within each region, and constraint sharing across region boundaries. The resulting set is proven to be polyhedral, safe, positively invariant, and finitely determined, and strictly larger than the maximal output admissible set that would be obtained by treating input saturation as a constraint. The same source states explicitly that this is not a disturbance-robust set in the classical sense of exogenous uncertainty, although it is robust-like with respect to internal switching induced by saturation (Gautam et al., 2023).

Robust admissibility also appears in adaptive safety control. The Robust Safe Set Algorithm constructs a control action from an uncertainty set of dynamics parameters $R$8, a safety index $R$9, and a time-varying augmentation $G$0. The admissible control set is defined by the derivative condition

$G$1

for all admissible uncertain models, and the method selects the minimum-effort control in this robust feasible set. Under the stated positivity assumptions on uncertain Lie derivatives, the paper proves existence of a robust feasible control and forward invariance of the safe set (Noren et al., 2019).

4. Numerical admissibility in ideal MHD

In high-order numerical schemes for ideal magnetohydrodynamics, the admissible set is a state set of physically meaningful conservative variables

$G$2

satisfying positive density and positive internal energy:

$G$3

For computation, the paper uses the stricter convex closed set

$G$4

Preserving $G$5 is tied directly to robust computations: if density or internal energy becomes nonpositive, pressure can become negative, the system may lose hyperbolicity, the discrete problem can become ill-posed, and the computation may crash or generate nonphysical solutions (Liu et al., 11 May 2026).

The paper formulates a cell-average limiter as a strongly convex constrained minimization problem enforcing global conservation and cellwise admissibility. Its key technical step is a decomposition of the admissible set into slices parameterized by the magnetic energy

$G$6

so that projection onto $G$7 reduces to a one-dimensional minimization in $G$8. The magnetic subproblem has an explicit solution, the scalar convex minimization is solved by the Brent method, and the global constrained optimization is handled by Davis--Yin splitting. Pointwise admissibility is then enforced by the Zhang--Shu positivity-preserving limiter. In this setting, robustness refers to preservation of the admissible state set, global conservation, and stability of DG computations under strong nonlinear dynamics (Liu et al., 11 May 2026).

5. Combinatorial admissible sets and structural robustness

In the theory of local models and Iwahori--Weyl groups, the admissible set is the Kottwitz--Rapoport subset

$G$9

with decomposition

$$G\setminus R = \{ \bar{x} \in G \mid \forall u \in U, \exists d \in D, \exists \bar{t}<\infty: x^{(\bar{x},u,d)}(\bar{t})\notin G\},$$0

Because $$G\setminus R = \{ \bar{x} \in G \mid \forall u \in U, \exists d \in D, \exists \bar{t}<\infty: x^{(\bar{x},u,d)}(\bar{t})\notin G\},$$1 has a unique minimal element but several maximal elements, the augmented admissible set

$$G\setminus R = \{ \bar{x} \in G \mid \forall u \in U, \exists d \in D, \exists \bar{t}<\infty: x^{(\bar{x},u,d)}(\bar{t})\notin G\},$$2

is introduced. The 2025 result proves that for any dominant coweight $$G\setminus R = \{ \bar{x} \in G \mid \forall u \in U, \exists d \in D, \exists \bar{t}<\infty: x^{(\bar{x},u,d)}(\bar{t})\notin G\},$$3, $$G\setminus R = \{ \bar{x} \in G \mid \forall u \in U, \exists d \in D, \exists \bar{t}<\infty: x^{(\bar{x},u,d)}(\bar{t})\notin G\},$$4 is dual shellable by constructing an explicit dual EL-labeling, and interprets robustness as uniformity in all reductive groups, validity for all dominant coweights, persistence on refined subposets $$G\setminus R = \{ \bar{x} \in G \mid \forall u \in U, \exists d \in D, \exists \bar{t}<\infty: x^{(\bar{x},u,d)}(\bar{t})\notin G\},$$5, and a canonical order in which irreducible components of the special fiber can be added one by one while preserving Cohen--Macaulayness at each step (He et al., 15 Sep 2025).

The 2026 parahoric-level extension replaces $$G\setminus R = \{ \bar{x} \in G \mid \forall u \in U, \exists d \in D, \exists \bar{t}<\infty: x^{(\bar{x},u,d)}(\bar{t})\notin G\},$$6 by

$$G\setminus R = \{ \bar{x} \in G \mid \forall u \in U, \exists d \in D, \exists \bar{t}<\infty: x^{(\bar{x},u,d)}(\bar{t})\notin G\},$$7

defines

$$G\setminus R = \{ \bar{x} \in G \mid \forall u \in U, \exists d \in D, \exists \bar{t}<\infty: x^{(\bar{x},u,d)}(\bar{t})\notin G\},$$8

and proves that $$G\setminus R = \{ \bar{x} \in G \mid \forall u \in U, \exists d \in D, \exists \bar{t}<\infty: x^{(\bar{x},u,d)}(\bar{t})\notin G\},$$9 is dual EL-shellable for every dominant $$\bar{x} \in R \iff \min_{u \in U} \, \sup_{d \in D} \, \sup_{t \in [0,\infty)} \, \max_{i} g_i(x^{(\bar{x},u,d)}(t)) \le 0.$$0 and spherical $$\bar{x} \in R \iff \min_{u \in U} \, \sup_{d \in D} \, \sup_{t \in [0,\infty)} \, \max_{i} g_i(x^{(\bar{x},u,d)}(t)) \le 0.$$1. The proof is characteristic-free and intrinsic to the structure of admissible sets, handles arbitrary spherical $$\bar{x} \in R \iff \min_{u \in U} \, \sup_{d \in D} \, \sup_{t \in [0,\infty)} \, \max_{i} g_i(x^{(\bar{x},u,d)}(t)) \le 0.$$2, residue characteristic $$\bar{x} \in R \iff \min_{u \in U} \, \sup_{d \in D} \, \sup_{t \in [0,\infty)} \, \max_{i} g_i(x^{(\bar{x},u,d)}(t)) \le 0.$$3, and non-reduced root systems, and yields an explicit shelling that gives an inductive, component-by-component building procedure for the special fibre that preserves Cohen--Macaulayness at each step (He et al., 6 Mar 2026).

A different refinement appears in the face decomposition of the Iwahori-level admissible set. To each face $$\bar{x} \in R \iff \min_{u \in U} \, \sup_{d \in D} \, \sup_{t \in [0,\infty)} \, \max_{i} g_i(x^{(\bar{x},u,d)}(t)) \le 0.$$4 of the coweight polytope $$\bar{x} \in R \iff \min_{u \in U} \, \sup_{d \in D} \, \sup_{t \in [0,\infty)} \, \max_{i} g_i(x^{(\bar{x},u,d)}(t)) \le 0.$$5, the paper associates a subset $$\bar{x} \in R \iff \min_{u \in U} \, \sup_{d \in D} \, \sup_{t \in [0,\infty)} \, \max_{i} g_i(x^{(\bar{x},u,d)}(t)) \le 0.$$6, proves the intrinsic characterization

$$\bar{x} \in R \iff \min_{u \in U} \, \sup_{d \in D} \, \sup_{t \in [0,\infty)} \, \max_{i} g_i(x^{(\bar{x},u,d)}(t)) \le 0.$$7

defines interior, boundary, and a distinguished center element $$\bar{x} \in R \iff \min_{u \in U} \, \sup_{d \in D} \, \sup_{t \in [0,\infty)} \, \max_{i} g_i(x^{(\bar{x},u,d)}(t)) \le 0.$$8, and shows that the Pappas--Rapoport face map has fibers

$$\bar{x} \in R \iff \min_{u \in U} \, \sup_{d \in D} \, \sup_{t \in [0,\infty)} \, \max_{i} g_i(x^{(\bar{x},u,d)}(t)) \le 0.$$9

The face map is therefore surjective, and $0$0 is partitioned by the interiors of its admissible faces (Yu, 15 May 2026).

Not every combinatorial use of “admissible set” introduces a distinct robust variant. In the theory of squarefree-power-like functions, the relevant notion is a $0$1-admissible $0$2-set, defined by a partition $0$3 satisfying the four conditions stated in the source, and the resulting invariant $0$4 gives the lower bound

$0$5

The same source states explicitly that it does not introduce a separate notion called a “robust admissible set”; any such wording would be external interpretation rather than a named concept in that paper (Chau et al., 1 Oct 2025).

6. Statistical, scheduling, and decision-theoretic variants

In robust conformal prediction, an admissible or acceptable region is built from a score that is stable under outliers, heavy tails, or adversarial perturbation. One construction uses the half-mass-radius nonconformity score

$0$6

which equals the distance from $0$7 to its $0$8-nearest neighbour. The conformal region

$0$9

retains the finite-sample guarantee

$$\bar{x} \in \partial R \iff \min_{u \in U} \, \sup_{d \in D} \, \sup_{t \in [0,\infty)} \, \max_i g_i(x^{(\bar{x},u,d)}(t)) = 0$$0

and converges in symmetric-difference probability to the population robust central set

$$\bar{x} \in \partial R \iff \min_{u \in U} \, \sup_{d \in D} \, \sup_{t \in [0,\infty)} \, \max_i g_i(x^{(\bar{x},u,d)}(t)) = 0$$1

defined through the distance-to-a-measure functional $$\bar{x} \in \partial R \iff \min_{u \in U} \, \sup_{d \in D} \, \sup_{t \in [0,\infty)} \, \max_i g_i(x^{(\bar{x},u,d)}(t)) = 0$$2. The same source establishes exponential concentration and tail bounds for the deviation between the empirical conformal region and its population counterpart (Cholaquidis et al., 20 Apr 2026).

A second conformal line studies adversarial robustness under evasion and poisoning. The basic principle is to upper bound the worst-case conformity score over a threat region,

$$\bar{x} \in \partial R \iff \min_{u \in U} \, \sup_{d \in D} \, \sup_{t \in [0,\infty)} \, \max_i g_i(x^{(\bar{x},u,d)}(t)) = 0$$3

and then define the prediction set by thresholding $$\bar{x} \in \partial R \iff \min_{u \in U} \, \sup_{d \in D} \, \sup_{t \in [0,\infty)} \, \max_i g_i(x^{(\bar{x},u,d)}(t)) = 0$$4. The paper derives provably robust sets with adversarially robust $$\bar{x} \in \partial R \iff \min_{u \in U} \, \sup_{d \in D} \, \sup_{t \in [0,\infty)} \, \max_i g_i(x^{(\bar{x},u,d)}(t)) = 0$$5 coverage, develops tighter CDF-aware bounds rather than mean-only smoothed-score bounds, treats both continuous and sparse data, and extends the same logic to feature poisoning and label poisoning of the calibration set (Zargarbashi et al., 2024).

In spatially continuous wireless systems, admissibility is determined by interference feasibility. A subset of users is admissible if simultaneous activity obeys the interference rule encoded by $$\bar{x} \in \partial R \iff \min_{u \in U} \, \sup_{d \in D} \, \sup_{t \in [0,\infty)} \, \max_i g_i(x^{(\bar{x},u,d)}(t)) = 0$$6, and random admissible-set scheduling selects one admissible subset uniformly at random in each time slot. Under a partition-based sufficient condition involving guaranteed sets and coefficients $$\bar{x} \in \partial R \iff \min_{u \in U} \, \sup_{d \in D} \, \sup_{t \in [0,\infty)} \, \max_i g_i(x^{(\bar{x},u,d)}(t)) = 0$$7, the resulting measure-valued Markov chain is positive Harris recurrent; in the symmetric case, stability holds whenever

$$\bar{x} \in \partial R \iff \min_{u \in U} \, \sup_{d \in D} \, \sup_{t \in [0,\infty)} \, \max_i g_i(x^{(\bar{x},u,d)}(t)) = 0$$8

where $$\bar{x} \in \partial R \iff \min_{u \in U} \, \sup_{d \in D} \, \sup_{t \in [0,\infty)} \, \max_i g_i(x^{(\bar{x},u,d)}(t)) = 0$$9 is the maximum admissible cardinality (Bouman et al., 2010).

Decision-theoretic uses of admissible sets shift the object from states to ambiguity sets. In partial identification, the identified set

$\mathcal A$0

is the relevant uncertainty set. The cited note does not define a formal robust admissible set, but it shows that set-coverage confidence regions are robust for maxmin or regret-based policy choice because they preserve the entire ambiguity set with probability $\mathcal A$1, whereas point-coverage regions may exclude adverse models and thereby make robust decisions fail (Henry et al., 2021). In robust utility maximization, admissibility is strategy-based: the robust admissible class

$\mathcal A$2

is the class in which an optimizer exists when utility is finite on the whole real line, and its robustness derives from requiring the supermartingale property under all local martingale measures with finite generalized entropy relative to at least one candidate model in $\mathcal A$3 (Owari, 2011).

Across these literatures, the term retains a stable core idea: an admissible set records precisely those states, strategies, configurations, or labels that remain acceptable under a specified family of constraints. What changes from one field to another is the source of robustness—worst-case disturbances, all allowable controls, saturation-induced switching, shellability under augmentation, perturbation-resistant prediction, or preservation of an ambiguity set.