Uncertainty Tube: Definitions & Applications

• Uncertainty Tube is a set-indexed moving envelope that captures admissible deviations around a nominal trajectory, supporting robust constraint satisfaction and calibrated predictions.
• Various formulations—polytopic, ellipsoidal, and probabilistic—enable dynamic adaptation of tube geometry for real-time control, safety verification, and interpretability.
• Applications span robust MPC, HMM decoding, risk assessment in vision, and motion planning, each tailoring guarantees to specific disturbance and model uncertainties.

“Uncertainty tube” denotes a set-valued construction that surrounds a nominal forecast, trajectory, or decoded path so that uncertainty can be represented, propagated, calibrated, or constrained. In recent arXiv literature, the term does not refer to a single canonical object: it includes polytopic and ellipsoidal state-space tubes in robust MPC, posterior superlevel sets on complete hidden-path space in HMM decoding, spatiotemporal risk intervals for hazardous objects in vision, and superelliptical envelopes for visualizing ensemble particle trajectories (Tranos et al., 2022, Parsi et al., 2022, Nicosia, 15 Jun 2026, Fu et al., 25 Mar 2026, Li et al., 19 Aug 2025). The common structure is an indexed family of sets whose semantics depend on the application: robust constraint satisfaction, credible-region coverage, calibrated prediction, or interpretable visualization.

1. Formal scope and representative definitions

Across domains, an uncertainty tube is typically indexed by time, prediction step, or sequence position, and its cross-sections are defined relative to a nominal object such as a state trajectory, a latent path, or a risk schedule. The geometry varies substantially, but the role is stable: the tube encodes the admissible or plausible deviation around that nominal object.

Setting	Tube definition	Representative paper
Mixed-traffic platoon control	$$\mathcal{T} \triangleq \{\mathcal{E}_k\}_{k=1}^{N_p},\ \mathcal{E}_k = \bar{e}(k) \oplus \mathcal{F}$$	(Feng et al., 2019)
Ellipsoidal tube MPC	$$\mathcal{X}_{l|k} = \{ x : (x-z_{l|k})^\top P (x-z_{l|k}) \le \alpha_{l|k}^2 \}$$	(Parsi et al., 2022)
Tropical Viterbi tube	$$\mathcal{T}_\eta = \{ x_{1:T} : \psi(x_{1:T};y_{1:T}) \ge \psi^\star(y_{1:T}) - \eta \}$$	(Nicosia, 15 Jun 2026)
Vision risk tube	$\widehat{\mathcal{I}_o = [\,\hat t^{\mathrm{start}}_o,\hat t^{\mathrm{end}}_o\,]\cap\{T,\dots,T+H-1\}$	(Fu et al., 25 Mar 2026)

The control literature usually defines the tube in state space. In mixed-traffic platoon control, the tube is a sequence of sets around a nominal tracking-error trajectory, with each section obtained by Minkowski addition of a nominal value and a robust positively invariant error set (Feng et al., 2019). In HMM decoding, by contrast, the tropical Viterbi tube is a posterior superlevel set on complete hidden-path space, parameterized by a log-score tolerance relative to the Viterbi optimum (Nicosia, 15 Jun 2026). In vision-based risk object identification, the tube is neither a state-space corridor nor an error set, but a per-object future time interval during which an object may be risky (Fu et al., 25 Mar 2026).

Taken together, these formulations show that “tube” is best understood as a moving set-valued envelope rather than as a fixed cylindrical geometry. The index set may be physical time, MPC horizon step, decoding position, or future video frame; the ambient space may be Euclidean state space, hidden-path space, or the discrete time axis itself.

2. State-space tubes in robust predictive control

In robust control, uncertainty tubes are most often used to separate nominal prediction from uncertain realization. A standard decomposition writes the state as the sum of a nominal component and an error component, and the tube is then the set of all states compatible with bounded disturbances and model mismatch. The purpose is constraint tightening: if the nominal plan remains in a tightened feasible set, then the actual state remains in the original constraints.

This logic appears in several distinct parameterizations. Self-Tuning Tube-based MPC defines polytopic tube sections

$$X_{k|t} := \{ x \in \mathbb{R}^{d_x} : T x \le \alpha_{k|t} \},$$

and propagates them by robust one-step inclusion constraints over the current parameter set vertices, additive disturbance set, and excitation set (Tranos et al., 2022). Configuration-Constrained Tube MPC uses half-space sections

$$P(Y,y_k) := \{x \in \mathbb{R}^{n_x} : Yx \le y_k\},$$

together with conic configuration constraints $E y_k \le 0$ that guarantee an exact joint facet-vertex representation and permit online optimization of tube geometry and vertex control laws through a convex program (Villanueva et al., 2022). Scalable ellipsoidal tube MPC instead adopts homothetic ellipsoids,

$$\mathcal{X}_{l|k} = \{ x : (x-z_{l|k})^\top P (x-z_{l|k}) \le \alpha_{l|k}^2 \},$$

so that forward containment and terminal invariance can be encoded by LMIs whose online size scales linearly with the system order (Parsi et al., 2022).

These constructions differ in geometry but share the same tightening principle. In the polytopic setting, robust inclusion is enforced with nonnegative multipliers so that all possible next states remain inside the next cross-section (Tranos et al., 2022). In the ellipsoidal setting, tightened constraints are obtained from ellipsoidal support functions, yielding conditions of the form

$$F z_{l|k} + G v_{l|k} + \alpha_{l|k}\,\bar f \le \mathbf{1}$$

for simultaneous state-input inequalities (Parsi et al., 2022). In mixed platoon control, the tightening appears as Pontryagin differences,

$\overline{\mathcal{E}=\mathcal{E}\ominus\mathcal{F},\qquad \overline{\mathcal{U}=\mathcal{U}\ominus K\mathcal{F},$

which ensure that the actual tracking error and control stay inside the original safety, speed, acceleration, and string-stability limits once the nominal plan is feasible in the tightened sets (Feng et al., 2019).

A common misconception is that a tube in MPC is simply a fixed-radius neighborhood. The cited work shows otherwise: tube sections may be rigid, homothetic, configuration-constrained, or fully online adapted, and the disturbance aggregation may be implemented through Minkowski sums, robust multipliers, or support-function bounds rather than through a single geometric inflation rule (Villanueva et al., 2022, Tranos et al., 2022).

3. Adaptive, dynamic, and learned tube geometry

Several papers replace fixed offline tube geometry with online adaptation. In nonlinear Dynamic Tube MPC, the tube geometry is governed by boundary-layer thickness $$\mathcal{X}_{l|k} = \{ x : (x-z_{l|k})^\top P (x-z_{l|k}) \le \alpha_{l|k}^2 \}$$0 and an error bound $$\mathcal{X}_{l|k} = \{ x : (x-z_{l|k})^\top P (x-z_{l|k}) \le \alpha_{l|k}^2 \}$$1, with dynamics

$$\mathcal{X}_{l|k} = \{ x : (x-z_{l|k})^\top P (x-z_{l|k}) \le \alpha_{l|k}^2 \}$$2

and

$$\mathcal{X}_{l|k} = \{ x : (x-z_{l|k})^\top P (x-z_{l|k}) \le \alpha_{l|k}^2 \}$$3

This couples tube size directly to state-dependent uncertainty and to the optimized control bandwidth $$\mathcal{X}_{l|k} = \{ x : (x-z_{l|k})^\top P (x-z_{l|k}) \le \alpha_{l|k}^2 \}$$4, allowing the controller to tighten the tube near obstacles and relax it in open regions (Lopez et al., 2019). The companion performance analysis reports that DTMPC used up to $$\mathcal{X}_{l|k} = \{ x : (x-z_{l|k})^\top P (x-z_{l|k}) \le \alpha_{l|k}^2 \}$$5 less control effort while achieving up to $$\mathcal{X}_{l|k} = \{ x : (x-z_{l|k})^\top P (x-z_{l|k}) \le \alpha_{l|k}^2 \}$$6 higher speeds, and that ADTMPC reduced the feedback control effort by up to another $$\mathcal{X}_{l|k} = \{ x : (x-z_{l|k})^\top P (x-z_{l|k}) \le \alpha_{l|k}^2 \}$$7, while delivering up to $$\mathcal{X}_{l|k} = \{ x : (x-z_{l|k})^\top P (x-z_{l|k}) \le \alpha_{l|k}^2 \}$$8 better trajectory tracking (Morozov et al., 2020).

Adaptive Tube MPC for uncertain LTI systems pushes the same idea into parametric learning. The uncertainty set $$\mathcal{X}_{l|k} = \{ x : (x-z_{l|k})^\top P (x-z_{l|k}) \le \alpha_{l|k}^2 \}$$9 is refined online by a non-falsified update, the disturbance over-approximation $$\mathcal{T}_\eta = \{ x_{1:T} : \psi(x_{1:T};y_{1:T}) \ge \psi^\star(y_{1:T}) - \eta \}$$0 shrinks accordingly, and the tube cross-sections are recomputed from an outer RPI set for the current estimated dynamics and feedback gain (Dey et al., 16 Mar 2026). The paper’s central claim is that the framework does not require the existence of a common quadratically stabilizing linear feedback gain for the entire parametric uncertainty set, because the prediction model, feedback gain, terminal set, and tube cross-sections are all updated online (Dey et al., 16 Mar 2026).

Sensitivity-based tube NMPC offers a different adaptive mechanism. There, the cross-section at each horizon step is a first-order ellipsoidal set generated from parametric state sensitivities $$\mathcal{T}_\eta = \{ x_{1:T} : \psi(x_{1:T};y_{1:T}) \ge \psi^\star(y_{1:T}) - \eta \}$$1 and an uncertainty weighting matrix $$\mathcal{T}_\eta = \{ x_{1:T} : \psi(x_{1:T};y_{1:T}) \ge \psi^\star(y_{1:T}) - \eta \}$$2. For a scalar constraint function $$\mathcal{T}_\eta = \{ x_{1:T} : \psi(x_{1:T};y_{1:T}) \ge \psi^\star(y_{1:T}) - \eta \}$$3, the robust margin is

$$\mathcal{T}_\eta = \{ x_{1:T} : \psi(x_{1:T};y_{1:T}) \ge \psi^\star(y_{1:T}) - \eta \}$$4

and the nominal constraint is tightened by adding this margin (Silano et al., 28 Apr 2026). The resulting tube is local and first-order, rather than globally invariant, but it still operationalizes the same set-based envelope around the nominal prediction.

The learning literature introduces yet another layer. “Deep Learning Tubes for Tube MPC” learns tube dynamics $$\mathcal{T}_\eta = \{ x_{1:T} : \psi(x_{1:T};y_{1:T}) \ge \psi^\star(y_{1:T}) - \eta \}$$5 by deep quantile regression so that

$$\mathcal{T}_\eta = \{ x_{1:T} : \psi(x_{1:T};y_{1:T}) \ge \psi^\star(y_{1:T}) - \eta \}$$6

holds along the horizon, with monotonicity in $$\mathcal{T}_\eta = \{ x_{1:T} : \psi(x_{1:T};y_{1:T}) \ge \psi^\star(y_{1:T}) - \eta \}$$7 enforced by a dedicated penalty and epistemic inflation added through an uncertainty estimator (Fan et al., 2020). This suggests a shift from hand-derived tube propagation to data-driven propagation, while retaining recursive-feasibility and probabilistic-constraint arguments within the MPC loop.

4. Distributional, conformal, and probabilistic tubes

A second major line of work interprets uncertainty tubes probabilistically rather than purely geometrically. In Wasserstein Tube MPC, the tube is a sequence of ambiguity sets in probability space,

$$\mathcal{T}_\eta = \{ x_{1:T} : \psi(x_{1:T};y_{1:T}) \ge \psi^\star(y_{1:T}) - \eta \}$$8

whose supports are the robust state tubes $$\mathcal{T}_\eta = \{ x_{1:T} : \psi(x_{1:T};y_{1:T}) \ge \psi^\star(y_{1:T}) - \eta \}$$9 (Aolaritei et al., 2023). The key result is exact propagation of the Wasserstein ambiguity set through linear dynamics: the center is pushed forward, the radius $\widehat{\mathcal{I}_o = [\,\hat t^{\mathrm{start}}_o,\hat t^{\mathrm{end}}_o\,]\cap\{T,\dots,T+H-1\}$0 is preserved, and the transportation cost is composed with the pseudoinverse of the disturbance-to-error map (Aolaritei et al., 2023). In this formulation, the tube is literally an uncertainty envelope on distributions, not only on trajectories.

Conformal calibration yields a related but distinct construction. In Adaptive Cruise Control under perception uncertainty, the current state is enclosed in a CP-calibrated box

$\widehat{\mathcal{I}_o = [\,\hat t^{\mathrm{start}}_o,\hat t^{\mathrm{end}}_o\,]\cap\{T,\dots,T+H-1\}$1

with coverage at least $\widehat{\mathcal{I}_o = [\,\hat t^{\mathrm{start}}_o,\hat t^{\mathrm{end}}_o\,]\cap\{T,\dots,T+H-1\}$2, and the tube is propagated by linear dynamics through the centroid recursion $\widehat{\mathcal{I}_o = [\,\hat t^{\mathrm{start}}_o,\hat t^{\mathrm{end}}_o\,]\cap\{T,\dots,T+H-1\}$3 and the half-size recursion $\widehat{\mathcal{I}_o = [\,\hat t^{\mathrm{start}}_o,\hat t^{\mathrm{end}}_o\,]\cap\{T,\dots,T+H-1\}$4 (Li et al., 2024). Robust satisfaction of tightened constraints on these propagated boxes implies a probabilistic safety statement for the original ACC constraints with confidence at least $\widehat{\mathcal{I}_o = [\,\hat t^{\mathrm{start}}_o,\hat t^{\mathrm{end}}_o\,]\cap\{T,\dots,T+H-1\}$5, or $\widehat{\mathcal{I}_o = [\,\hat t^{\mathrm{start}}_o,\hat t^{\mathrm{end}}_o\,]\cap\{T,\dots,T+H-1\}$6 when the quantile is optimized online (Li et al., 2024).

Risk-bounded motion planning extends the probabilistic tube concept to stochastic nonlinear systems and uncertain environments. There the tube may be a polynomial sublevel set or a quadratic tube

$\widehat{\mathcal{I}_o = [\,\hat t^{\mathrm{start}}_o,\hat t^{\mathrm{end}}_o\,]\cap\{T,\dots,T+H-1\}$7

and is sized so that

$\widehat{\mathcal{I}_o = [\,\hat t^{\mathrm{start}}_o,\hat t^{\mathrm{end}}_o\,]\cap\{T,\dots,T+H-1\}$8

for each discrete time along a motion primitive (Han et al., 2023). Safety verification is moved to continuous time by SOS certificates that prove tube containment inside deterministic risk contours derived from obstacle moments and Cantelli-type bounds (Han et al., 2023).

These papers make clear that “uncertainty tube” does not always mean worst-case robustness. Depending on the construction, the guarantee may be a Wasserstein-robust CVaR bound, a conformal marginal-coverage bound, or a per-cycle collision-probability bound. This suggests that the semantics of safety, feasibility, and conservatism must be read from the underlying probabilistic machinery rather than inferred from the word “tube” alone.

5. Latent-path, risk-interval, and visualization interpretations

Outside control, the same term is used for structured uncertainty over discrete paths or over future event timing. In HMM decoding, the tropical Viterbi tube is defined as

$\widehat{\mathcal{I}_o = [\,\hat t^{\mathrm{start}}_o,\hat t^{\mathrm{end}}_o\,]\cap\{T,\dots,T+H-1\}$9

that is, the set of complete hidden trajectories whose complete-data log-score lies within a tolerance of the Viterbi optimum (Nicosia, 15 Jun 2026). State, transition, and change-status projections are then computed exactly by max-plus forward-backward recursions, with dense-transition complexity $$X_{k|t} := \{ x \in \mathbb{R}^{d_x} : T x \le \alpha_{k|t} \},$$0 (Nicosia, 15 Jun 2026). The paper emphasizes that these projections show which local features remain compatible with globally near-optimal complete paths, thereby adding a pathwise uncertainty layer absent from single-path Viterbi decoding (Nicosia, 15 Jun 2026).

In uncertainty-aware vision-based risk object identification, the tube is a per-object future interval rather than a geometric corridor:

$$X_{k|t} := \{ x \in \mathbb{R}^{d_x} : T x \le \alpha_{k|t} \},$$1

Conformal Prediction is applied to per-time nonconformity scores, producing category-aware quantiles and a buffer zone

$$X_{k|t} := \{ x \in \mathbb{R}^{d_x} : T x \le \alpha_{k|t} \},$$2

which is used to classify future steps as risk, no risk, or ambiguous (Fu et al., 25 Mar 2026). The resulting calibrated tube captures temporal uncertainty in risk onset and offset rather than uncertainty in object geometry (Fu et al., 25 Mar 2026).

Visualization work uses the term in a still different way. For particle trajectories predicted by neural networks, the uncertainty tube is a surface swept along the mean trajectory, with local cross-sections given by a superellipse

$$X_{k|t} := \{ x \in \mathbb{R}^{d_x} : T x \le \alpha_{k|t} \},$$3

The principal directions $$X_{k|t} := \{ x \in \mathbb{R}^{d_x} : T x \le \alpha_{k|t} \},$$4 come from the covariance of ensemble samples projected onto the plane orthogonal to the local mean tangent, and the paper uses $$X_{k|t} := \{ x \in \mathbb{R}^{d_x} : T x \le \alpha_{k|t} \},$$5 in most results to accentuate anisotropy (Li et al., 19 Aug 2025). Here the tube is explicitly a visualization device rather than a constraint set, and its purpose is to reveal directional uncertainty and orientation changes without the clutter of spaghetti plots (Li et al., 19 Aug 2025).

A plausible implication is that the term has expanded from robust-control geometry into a broader family of spatiotemporal uncertainty summaries. The defining feature is no longer the ambient space, but the use of a moving set to retain structure that scalar uncertainty measures discard.

6. Guarantees, applications, and recurring limitations

The practical reach of uncertainty tubes is broad. In artificial pancreas control, Glycemic Safety Tube Control uses a glucose safety interval

$$X_{k|t} := \{ x \in \mathbb{R}^{d_x} : T x \le \alpha_{k|t} \},$$6

together with tracking funnels for the unmeasured internal states $$X_{k|t} := \{ x \in \mathbb{R}^{d_x} : T x \le \alpha_{k|t} \},$$7 and $$X_{k|t} := \{ x \in \mathbb{R}^{d_x} : T x \le \alpha_{k|t} \},$$8, and proves robust invariance of these tubes under bounded meals, parametric uncertainty, and estimation errors (Singh et al., 11 May 2026). In differential-drive robotics, Spatiotemporal Tubes are circular safe corridors

$$X_{k|t} := \{ x \in \mathbb{R}^{d_x} : T x \le \alpha_{k|t} \},$$9

synthesized by a scenario-based program and enforced by a closed-form funnel controller that keeps the robot inside the tube while satisfying Temporal Reach-Avoid-Stay specifications (Das et al., 5 Dec 2025). In mixed platooning, the tube bounds CAV tracking errors under HDV prediction uncertainty and supports event-triggered replanning rather than per-step replanning (Feng et al., 2019).

Despite this diversity, the same technical themes recur. Tube methods often claim robust constraint satisfaction, recursive feasibility, and stability or ISS in control settings (Tranos et al., 2022, Parsi et al., 2022, Dey et al., 16 Mar 2026). Probabilistic variants instead claim marginal coverage, tail-risk control, or bounded collision probability (Aolaritei et al., 2023, Li et al., 2024, Han et al., 2023). Sequence- and vision-based tubes provide conservative projected bands or interval coverage rather than dynamical invariance (Nicosia, 15 Jun 2026, Fu et al., 25 Mar 2026).

A recurring misconception is that all uncertainty tubes provide the same kind of guarantee. The literature contradicts that view. A robust invariant tube in state space, a Wasserstein ambiguity tube, a conformal state box, a posterior superlevel set on hidden-path space, and a superelliptical visualization surface all qualify as uncertainty tubes in their respective papers, but they certify different objects and under different assumptions. Exchangeability is central for conformal calibration, sub-Gaussian assumptions matter for concentration-based adaptive MPC, disturbance bounds and invariant-set constructions matter in robust control, and projections may be conservative because projection loses pathwise constraints (Li et al., 2024, Tranos et al., 2022, Nicosia, 15 Jun 2026).

Taken together, the cited works portray the uncertainty tube as a transferable design pattern: an indexed family of sets that preserves structural information about uncertainty while remaining compatible with decision-making, optimization, or interpretation. What changes from domain to domain is the space in which the tube lives, the geometry of its cross-sections, and the meaning of its guarantee.