Ball-Shaped Ambiguity Sets

• Ball-shaped ambiguity sets are defined by bounded distance constraints around a nominal object, allowing uncertainty to be parameterized over varied ambient spaces such as distributional control, robust MDPs, and kernel methods.
• They enable tractable reformulations and regularized designs by linking moment or transition uncertainty with explicit norm-based penalties and convex optimization frameworks.
• Applications span distributionally robust control, Bayesian safety in MDPs, probabilistic motion planning, and infinite-dimensional settings, illustrating both their practical benefits and theoretical limitations.

Searching arXiv for recent and relevant papers on ball-shaped ambiguity sets to ground the article. Ball-shaped ambiguity sets are uncertainty regions defined by a radius constraint around a nominal object, but the ambient space in which the “ball” is drawn varies by problem class. In distributionally robust control, the ball may be centered at a nominal mean–covariance pair; in robust MDPs, it is often an $L_1$-ball in the probability simplex around a nominal transition law; in kernel methods, it is an MMD ball in an RKHS over embedded distributions; in PDE-based uncertainty propagation, it is frequently a Wasserstein ball around an empirical distribution; and in infinite-dimensional probability theory, it can denote a translate of a fixed-radius ball with maximal probability content. Across these settings, the common role of the ball is to parameterize uncertainty by a center, a radius, and a geometry, thereby converting poorly specified distributional uncertainty into a tractable design variable or a measurable uncertainty region (Taha et al., 11 Dec 2025).

1. Geometric forms and ambient spaces

The expression “ball-shaped ambiguity set” is not tied to a single metric or representation. What is shared is a constraint of the form “distance to a nominal object is bounded,” where the nominal object may be a probability distribution, a moment pair, an embedded measure, or a geometric center, and the distance may be induced by a norm, a divergence, or an RKHS metric.

Setting	Representative ball constraint	Representative use
Moment-based control	$$\|\mu-\bar\mu\|_2^2\le r_1,\ \|\Sigma-\bar\Sigma\|_p\le r_2$$	Distributionally robust regret-optimal control
Robust MDPs	$$\{p\in\Delta^S:\|p-\bar p_{s,a}\|_1\le \psi_{s,a}\}$$	Safe policy evaluation and synthesis
Gaussian planning	$$\|\mu_{\mathbb P}-\hat\mu\|_2 + c\sqrt{\lambda_{\max}(\Sigma_{\mathbb P})}\le \hat r$$	Chance-constrained reachability to a Euclidean ball
Kernel dynamics	$\{\mu: MMD(\mu,\hat\mu)\le \rho\}$	Multistep ambiguity propagation in RKHS
PDE uncertainty propagation	$\{F: W_1(\widehat F,F)\le \rho\}$	Data-driven ambiguity propagation under hyperbolic dynamics
Banach-space maximal balls	$B_r(x_0)=x_0+B_r$ maximizing $\mu(B_r(x))$	Fixed-shape sets of maximal probability

This taxonomy indicates that ball geometry is often chosen to match the representation used by the underlying model. In moment-based DRO, the geometry is over first and second moments; in robust RL, over transition probabilities; in kernelized dynamics, over kernel mean embeddings; and in probabilistic planning, over coupled mean–covariance descriptions of Gaussian beliefs. This suggests that “ball-shaped” is best understood as a structural template rather than a commitment to Euclidean balls in state space (Petrik et al., 2019, Rose et al., 6 Oct 2025, Zhu, 2023, Boso et al., 2020, Schmidt, 2023).

2. Moment balls in distributionally robust control

A recent control-theoretic formulation studies finite-horizon, linear-quadratic stochastic control when the disturbance distribution is unknown but its mean and covariance are only known approximately. The disturbance trajectory is stacked as $w=(x_0,w_0,\dots,w_{T-1})\in\mathbb R^n$, the dynamics admit the compact representation $x=Fu+Gw$, and the quadratic cost is

$$\|\mu-\bar\mu\|_2^2\le r_1,\ \|\Sigma-\bar\Sigma\|_p\le r_2$$0

with $$\|\mu-\bar\mu\|_2^2\le r_1,\ \|\Sigma-\bar\Sigma\|_p\le r_2$$1 and $$\|\mu-\bar\mu\|_2^2\le r_1,\ \|\Sigma-\bar\Sigma\|_p\le r_2$$2. The controller class is strictly causal affine disturbance-feedback,

$$\|\mu-\bar\mu\|_2^2\le r_1,\ \|\Sigma-\bar\Sigma\|_p\le r_2$$3

with $$\|\mu-\bar\mu\|_2^2\le r_1,\ \|\Sigma-\bar\Sigma\|_p\le r_2$$4 block lower triangular in the appropriate strictly causal sense. The regret is measured relative to the best noncausal controller $$\|\mu-\bar\mu\|_2^2\le r_1,\ \|\Sigma-\bar\Sigma\|_p\le r_2$$5, where

$$\|\mu-\bar\mu\|_2^2\le r_1,\ \|\Sigma-\bar\Sigma\|_p\le r_2$$6

and simplifies to

$$\|\mu-\bar\mu\|_2^2\le r_1,\ \|\Sigma-\bar\Sigma\|_p\le r_2$$7

The resulting design problem is

$$\|\mu-\bar\mu\|_2^2\le r_1,\ \|\Sigma-\bar\Sigma\|_p\le r_2$$8

The ambiguity set is ball-shaped around a nominal moment pair $$\|\mu-\bar\mu\|_2^2\le r_1,\ \|\Sigma-\bar\Sigma\|_p\le r_2$$9: $$\{p\in\Delta^S:\|p-\bar p_{s,a}\|_1\le \psi_{s,a}\}$$0 where $$\{p\in\Delta^S:\|p-\bar p_{s,a}\|_1\le \psi_{s,a}\}$$1, $$\{p\in\Delta^S:\|p-\bar p_{s,a}\|_1\le \psi_{s,a}\}$$2, and $$\{p\in\Delta^S:\|p-\bar p_{s,a}\|_1\le \psi_{s,a}\}$$3 is the Schatten $$\{p\in\Delta^S:\|p-\bar p_{s,a}\|_1\le \psi_{s,a}\}$$4-norm. The paper identifies $$\{p\in\Delta^S:\|p-\bar p_{s,a}\|_1\le \psi_{s,a}\}$$5 with a nuclear norm ball, $$\{p\in\Delta^S:\|p-\bar p_{s,a}\|_1\le \psi_{s,a}\}$$6 with a Frobenius norm ball, and $$\{p\in\Delta^S:\|p-\bar p_{s,a}\|_1\le \psi_{s,a}\}$$7 with a spectral norm ball. Because only moments are constrained, the formulation allows temporal correlation and non-Gaussianity (Taha et al., 11 Dec 2025).

The central technical result is an exact convex reformulation. Writing $$\{p\in\Delta^S:\|p-\bar p_{s,a}\|_1\le \psi_{s,a}\}$$8 and $$\{p\in\Delta^S:\|p-\bar p_{s,a}\|_1\le \psi_{s,a}\}$$9, the optimal policy has the form

$$\|\mu_{\mathbb P}-\hat\mu\|_2 + c\sqrt{\lambda_{\max}(\Sigma_{\mathbb P})}\le \hat r$$0

where $$\|\mu_{\mathbb P}-\hat\mu\|_2 + c\sqrt{\lambda_{\max}(\Sigma_{\mathbb P})}\le \hat r$$1 solves

$$\|\mu_{\mathbb P}-\hat\mu\|_2 + c\sqrt{\lambda_{\max}(\Sigma_{\mathbb P})}\le \hat r$$2

with $$\|\mu_{\mathbb P}-\hat\mu\|_2 + c\sqrt{\lambda_{\max}(\Sigma_{\mathbb P})}\le \hat r$$3 the dual exponent of $$\|\mu_{\mathbb P}-\hat\mu\|_2 + c\sqrt{\lambda_{\max}(\Sigma_{\mathbb P})}\le \hat r$$4. Mean uncertainty therefore induces a spectral-norm penalty $$\|\mu_{\mathbb P}-\hat\mu\|_2 + c\sqrt{\lambda_{\max}(\Sigma_{\mathbb P})}\le \hat r$$5, while covariance uncertainty induces a Schatten-$$\|\mu_{\mathbb P}-\hat\mu\|_2 + c\sqrt{\lambda_{\max}(\Sigma_{\mathbb P})}\le \hat r$$6-norm penalty $$\|\mu_{\mathbb P}-\hat\mu\|_2 + c\sqrt{\lambda_{\max}(\Sigma_{\mathbb P})}\le \hat r$$7. The paper interprets this as a regularized version of nominal LQ stochastic control, with $$\|\mu_{\mathbb P}-\hat\mu\|_2 + c\sqrt{\lambda_{\max}(\Sigma_{\mathbb P})}\le \hat r$$8 recovering the ambiguity-free case, $$\|\mu_{\mathbb P}-\hat\mu\|_2 + c\sqrt{\lambda_{\max}(\Sigma_{\mathbb P})}\le \hat r$$9 corresponding to robust regret control, and $\{\mu: MMD(\mu,\hat\mu)\le \rho\}$0 yielding classical LQR-like behavior. This establishes a direct connection between ball radii and robustness regularization (Taha et al., 11 Dec 2025).

The same work shows that the convex program is SDP-representable, but emphasizes that interior-point methods scale poorly with horizon and system size because of heavy memory and second-order computational cost. It therefore derives a dual projected subgradient method on a product of covariance balls,

$\{\mu: MMD(\mu,\hat\mu)\le \rho\}$1

with update

$\{\mu: MMD(\mu,\hat\mu)\le \rho\}$2

If $\{\mu: MMD(\mu,\hat\mu)\le \rho\}$3, this reduces to projected gradient ascent. The projection onto Schatten balls is performed by eigendecomposition followed by projection of eigenvalues onto an $\{\mu: MMD(\mu,\hat\mu)\le \rho\}$4-ball, preserving positive semidefiniteness. In the reported experiments on a damped double integrator with AR(1) disturbances, the proposed Nuc-Regret, Frob-Regret, and Spec-Regret methods show that performance first improves as the radius increases from zero and then becomes over-conservative at large radii; Spec-Regret matches the oracle causal controller at its best tuned radius in the uncorrelated case $\{\mu: MMD(\mu,\hat\mu)\le \rho\}$5; and at $\{\mu: MMD(\mu,\hat\mu)\le \rho\}$6 the SDP takes over 16 minutes on average while the first-order method finishes in under 8 seconds (Taha et al., 11 Dec 2025).

3. $\{\mu: MMD(\mu,\hat\mu)\le \rho\}$7-balls in robust MDPs and Bayesian tightening

In robust MDPs, ball-shaped ambiguity sets are typically sets of transition probabilities in the simplex. The standard rectangular construction at each state–action pair takes the form

$\{\mu: MMD(\mu,\hat\mu)\le \rho\}$8

and enters the robust Bellman operator

$\{\mu: MMD(\mu,\hat\mu)\le \rho\}$9

This geometry underlies both frequentist and Bayesian safe-policy constructions. The Hoeffding-style confidence ball reviewed in the Bayesian robust-MDP papers is

$\{F: W_1(\widehat F,F)\le \rho\}$0

with the guarantee

$\{F: W_1(\widehat F,F)\le \rho\}$1

The stated limitation is that such sets are conservative because they enforce safety uniformly over all policies and all value functions (Petrik et al., 2019).

The Bayesian Confidence Interval or Bayesian Credible Region construction retains the $\{F: W_1(\widehat F,F)\le \rho\}$2-ball shape but centers it at the posterior mean,

$\{F: W_1(\widehat F,F)\le \rho\}$3

and chooses the radius as the smallest posterior credible radius satisfying

$\{F: W_1(\widehat F,F)\le \rho\}$4

The implementation described in the paper is based on posterior sampling and a quantile computation, and the safety guarantee remains

$\{F: W_1(\widehat F,F)\le \rho\}$5

The article-length significance of this result is not a new geometry but a tighter Bayesian calibration of an existing ball geometry (Petrik et al., 2019, Russel et al., 2018).

RSVF further relaxes the requirement that the ambiguity set be a confidence interval. For a fixed value function $\{F: W_1(\widehat F,F)\le \rho\}$6, it defines the safety-sufficient halfspace

$\{F: W_1(\widehat F,F)\le \rho\}$7

where

$\{F: W_1(\widehat F,F)\le \rho\}$8

The key geometric statement is that safety is guaranteed if the ambiguity set intersects each relevant halfspace. RSVF then fits a ball

$\{F: W_1(\widehat F,F)\le \rho\}$9

by optimizing both center and radius: $B_r(x_0)=x_0+B_r$0 with

$B_r(x_0)=x_0+B_r$1

The resulting robust return remains safe, but the paper’s conceptual point is that the ball is no longer a generic posterior confidence region; it is a value-adapted uncertainty set whose location and size are chosen to guarantee only the decision-relevant lower bound (Petrik et al., 2019, Russel et al., 2018).

This line of work also clarifies a common misconception: ball-shaped ambiguity sets in robust RL need not be centered at the empirical or posterior mean. A ball in the simplex can instead be centered at a point chosen by an optimization over supporting hyperplanes induced by relevant value functions. Conversely, recent work on robust $B_r(x_0)=x_0+B_r$2-learning for finite ambiguity sets explicitly positions itself as going beyond ball-based uncertainty by allowing a finite, user-specified set of transition measures

$B_r(x_0)=x_0+B_r$3

rather than a neighborhood around one nominal model. This contrast shows that ball-shaped sets are only one modeling choice among several in robust sequential decision making (Decker et al., 2024).

4. Coupled mean–covariance balls for Gaussian motion planning

A distinct use of ball-shaped ambiguity sets appears in probabilistic motion planning for linear Gaussian systems with Euclidean ball goals. Here the ambiguity set is a set of Gaussian distributions whose mean and covariance jointly satisfy a single coupled inequality. For center $B_r(x_0)=x_0+B_r$4, radius $B_r(x_0)=x_0+B_r$5, and risk level $B_r(x_0)=x_0+B_r$6, the Gaussian-quantile version is

$B_r(x_0)=x_0+B_r$7

and the $B_r(x_0)=x_0+B_r$8-based version is

$B_r(x_0)=x_0+B_r$9

The paper states the associated guarantee

$\mu(B_r(x))$0

This couples the allowable mean displacement and covariance spread through a shared radius budget (Rose et al., 6 Oct 2025).

The paper contrasts this geometry with earlier ellipsoidal, Wasserstein/Gelbrich, and MAXCOVAR-style formulations. The earlier ellipsoidal ambiguity set constrains the mean to lie in a fixed ellipsoid independent of covariance, while MAXCOVAR fixes the mean and bounds covariance. The new ball-shaped ambiguity set allows tradeoff: a distribution may have a mean farther from the goal center if its covariance is small enough. The paper’s stated interpretation is direct: if the mean is far from the goal center, then the covariance must be small; if the covariance is larger, the mean must be closer. This coupling is presented as a tighter representation of the Gaussian distributions that truly satisfy a ball-reaching chance constraint (Rose et al., 6 Oct 2025).

This ambiguity set is the basic object in the MAX-COV-BALL backward reachable tree. The synthesis problem maximizes the radius of the initial ball-shaped ambiguity set,

$\mu(B_r(x))$1

subject to the mean, covariance, and shape recursions

$\mu(B_r(x))$2

together with chance constraints and a terminal inclusion condition. The paper proves a maximum-coverage theorem in the absence of process noise and state constraints: $\mu(B_r(x))$3 and also proves dominance over MAXCOVAR with strict containment under mild conditions. It further extends the construction to a Minkowski-sum target

$\mu(B_r(x))$4

through an ellipsoidal generalization

$\mu(B_r(x))$5

which permits the center of an inner ball to vary over an outer ellipsoid (Rose et al., 6 Oct 2025).

5. Ball propagation under learned and physical dynamics

Ball-shaped ambiguity sets are also used when uncertainty must be propagated forward in time. In nonlinear data-driven dynamics models based on kernel conditional mean embeddings and Koopman operators, the uncertainty object is the embedded distribution $\mu(B_r(x))$6, and the ambiguity region is an MMD ball: $\mu(B_r(x))$7 or equivalently

$\mu(B_r(x))$8

Since

$\mu(B_r(x))$9

the ambiguity set is literally a norm ball in RKHS geometry. The paper’s multistep propagation result bounds the error between the empirical and true embedded distributions after $w=(x_0,w_0,\dots,w_{T-1})\in\mathbb R^n$0 steps, and the recursive algorithm updates both center and radius, outputting an ambiguity tube $w=(x_0,w_0,\dots,w_{T-1})\in\mathbb R^n$1 together with radii $w=(x_0,w_0,\dots,w_{T-1})\in\mathbb R^n$2. The stated conceptual claim is that these MMD balls are the natural geometric structure for learned kernel dynamics because the dynamics themselves are represented through RKHS operators (Zhu, 2023).

A different propagation problem arises for hyperbolic conservation laws with uncertain inputs. There, the ambiguity sets are 1-Wasserstein balls around empirical distributions of uncertain initial and boundary data. At each spatial point or boundary point, the paper constructs empirical distributions from $w=(x_0,w_0,\dots,w_{T-1})\in\mathbb R^n$3 i.i.d. samples and uses a concentration-of-measure bound

$w=(x_0,w_0,\dots,w_{T-1})\in\mathbb R^n$4

with $w=(x_0,w_0,\dots,w_{T-1})\in\mathbb R^n$5, transferring the radius through globally Lipschitz input maps. The pointwise ambiguity balls are written as

$w=(x_0,w_0,\dots,w_{T-1})\in\mathbb R^n$6

with an analogous boundary construction. The propagation mechanism is the CDF PDE obtained by the method of distributions: $w=(x_0,w_0,\dots,w_{T-1})\in\mathbb R^n$7 For general nonlinear hyperbolic equations, the paper propagates upper and lower CDF envelopes, leading to ambiguity bands. For linear dynamics, it derives a closed PDE for the Wasserstein radius itself,

$w=(x_0,w_0,\dots,w_{T-1})\in\mathbb R^n$8

where

$w=(x_0,w_0,\dots,w_{T-1})\in\mathbb R^n$9

The reported guarantee is that the propagated ambiguity balls still contain the true distributions with confidence $x=Fu+Gw$0 (Boso et al., 2020).

Taken together, these two lines of work show that the tractability of ball propagation depends strongly on the geometry used. In the RKHS setting, the ball is compatible with linear operator propagation in embedding space; in the hyperbolic PDE setting, Wasserstein balls can be propagated exactly only for linear dynamics, while nonlinear dynamics require envelope propagation. This suggests that dynamic ambiguity propagation is easiest when the uncertainty geometry is aligned with the evolution operator acting on the chosen representation (Zhu, 2023, Boso et al., 2020).

6. Infinite-dimensional existence theory, nonexistence, and modeling limits

Ball-shaped ambiguity sets also arise as fixed-shape sets of maximal probability content. In a Banach space $x=Fu+Gw$1, the question is whether for every radius $x=Fu+Gw$2 there exists a center $x=Fu+Gw$3 such that

$x=Fu+Gw$4

where $x=Fu+Gw$5. The relevant existence theorem states that if $x=Fu+Gw$6 is the separable dual of a Banach space and $x=Fu+Gw$7 is a bounded weak$x=Fu+Gw$8-closed convex set with non-empty interior, then for every Borel probability measure $x=Fu+Gw$9 on $$\|\mu-\bar\mu\|_2^2\le r_1,\ \|\Sigma-\bar\Sigma\|_p\le r_2$$00 there exists $$\|\mu-\bar\mu\|_2^2\le r_1,\ \|\Sigma-\bar\Sigma\|_p\le r_2$$01 such that

$$\|\mu-\bar\mu\|_2^2\le r_1,\ \|\Sigma-\bar\Sigma\|_p\le r_2$$02

Taking $$\|\mu-\bar\mu\|_2^2\le r_1,\ \|\Sigma-\bar\Sigma\|_p\le r_2$$03 yields existence of a ball of radius $$\|\mu-\bar\mu\|_2^2\le r_1,\ \|\Sigma-\bar\Sigma\|_p\le r_2$$04 with maximal $$\|\mu-\bar\mu\|_2^2\le r_1,\ \|\Sigma-\bar\Sigma\|_p\le r_2$$05-mass. Immediate positive cases include separable reflexive Banach spaces and $$\|\mu-\bar\mu\|_2^2\le r_1,\ \|\Sigma-\bar\Sigma\|_p\le r_2$$06 (Schmidt, 2023).

The same paper proves that existence can fail in natural infinite-dimensional settings. In $$\|\mu-\bar\mu\|_2^2\le r_1,\ \|\Sigma-\bar\Sigma\|_p\le r_2$$07 with the sup norm and product exponential measure

$$\|\mu-\bar\mu\|_2^2\le r_1,\ \|\Sigma-\bar\Sigma\|_p\le r_2$$08

the map $$\|\mu-\bar\mu\|_2^2\le r_1,\ \|\Sigma-\bar\Sigma\|_p\le r_2$$09 has no maximizer for any radius $$\|\mu-\bar\mu\|_2^2\le r_1,\ \|\Sigma-\bar\Sigma\|_p\le r_2$$10, although the supremum is explicit: $$\|\mu-\bar\mu\|_2^2\le r_1,\ \|\Sigma-\bar\Sigma\|_p\le r_2$$11 A second counterexample constructs a measure on $$\|\mu-\bar\mu\|_2^2\le r_1,\ \|\Sigma-\bar\Sigma\|_p\le r_2$$12 that is absolutely continuous with respect to a Gaussian measure and still has no maximal ball. The paper then gives two nonexistence examples on Wiener space, including conditioned Wiener measure above a moving boundary. The stated conceptual takeaway is a sharp dichotomy: in structurally favorable Banach spaces, maximal-probability fixed-shape convex sets always exist; in other natural infinite-dimensional settings, even very regular measures may fail to admit any ball of maximal probability (Schmidt, 2023).

This existence theory clarifies a potential misconception in applications: the intuitive idea that a ball-shaped uncertainty region can always be centered where the measure “concentrates most” is not valid in full generality. The robust-optimization literature often treats the center and radius as optimization variables or data-driven estimates, but the Banach-space results show that the existence of an actual maximizing ball is a structural property of the ambient space. A related modeling limit appears in robust RL, where recent work argues that ball-shaped ambiguity sets are too restrictive when uncertainty is scenario-based rather than distance-based, and replaces them with finite, user-specified ambiguity sets. A plausible implication is that ball geometries are most effective when uncertainty is naturally described as perturbations around a nominal object, and less effective when the admissible models are intrinsically asymmetric, multimodal, or discrete (Decker et al., 2024).

Ball-shaped ambiguity sets therefore occupy a broad but technically heterogeneous position in contemporary uncertainty modeling. Their importance comes from the fact that they often preserve tractability, admit explicit regularization interpretations, or align with the geometry of the underlying model representation. Their limitations arise when the chosen geometry decouples quantities that should be coupled, when dynamic propagation is incompatible with the model class, or when the ambient space does not support existence of a maximal-probability ball.