Probabilistic Tube Bounds in Stochastic Systems

• Probabilistic Tube Bounds are rigorous constructs that enclose all possible trajectories of a system within a high-probability tube to quantify uncertainty.
• They leverage sampling methods, concentration inequalities, and convex optimization to compute reachable sets across dynamical systems, stochastic processes, and random fields.
• These bounds underpin applications in safety verification, robust control, learning-based forecasting, and geometric estimation in contexts like neural ODEs and SDEs.

A probabilistic tube bound refers to a mathematically rigorous enclosure of all possible trajectories of a dynamical system, random field, or stochastic process, such that the trajectories remain within a "tube" or region around a reference path, set, or manifold with high probability. These constructions quantify uncertainty and robustness, yielding finite-sample or asymptotic probability guarantees for verification, control, learning, and geometric estimation in contexts ranging from stochastic differential equations to reachability analysis and prediction interval regression.

1. Core Mathematical Formulations

Probabilistic tube bounds come in several canonical forms, but always involve the high-probability containment of random trajectories within a time-/space-varying set, typically parameterized by a tube width $r(t)$ or set $\Omega_t$, and measured against a reference trajectory or manifold.

General Template

For a dynamical system $X_t$ (with or without stochasticity), the aim is to certify

$$\Pr\left\{ X_t \in \Omega_t \text{ for all } t\in[0,T] \right\} \geq 1-\delta.$$

Key settings include:

• Continuous-depth ODEs (Neural ODEs):
  • For $x'(t) = f(x(t)), x(t_0) \in B_0 = B(x_0, r_0)$, the reachable set at time $t$ is $R(t; B_0) = \{\phi(t;x) : x\in B_0\}$.
  • A probabilistic tube is constructed so that $R(t_j; B_0) \subseteq B(\phi(t_j; x_0), \hat M_j)$ holds simultaneously for all $j$ with total probability at least $1-\delta$ (Gruenbacher et al., 2021).
• Stochastic Differential Equations:
  • For $\Omega_t$0 and reference $\Omega_t$1 from $\Omega_t$2,

  $\Omega_t$3

  where $\Omega_t$4 is linked to system noise, contraction, and time horizon via martingale concentration (Liu et al., 5 Mar 2025).

• Target Tube Reachability:

  • For stochastic/deterministic systems $\Omega_t$5 and a "target tube" $\Omega_t$6,

  $\Omega_t$7

  is maximized or underapproximated over initial states (Vinod et al., 2018, Vlahakis et al., 2024, Gao et al., 6 May 2025).

• Gaussian/Random Fields:

  • The tube method (Euler characteristic heuristic) estimates $\Omega_t$8 by integrating the probability of hitting threshold $\Omega_t$9 over tubes around $X_t$0; the relative error decays exponentially in the subexponential tail case (Kuriki et al., 15 Jul 2025).

2. Probabilistic Tube Bounds in Reachability and Verification

Sampling-Based and Optimization Approaches

• GoTube Algorithm (ODEs and continuous-depth NNs):
  • At each time $X_t$1, draws $X_t$2 i.i.d. initial conditions, simulates trajectories, and computes the maximal deviation $X_t$3. A concentration-based padding $X_t$4 (from DKW inequality) is added:

  $X_t$5

  so that

  $X_t$6 - A union bound over time steps gives a full-horizon tube with overall confidence $X_t$7 (Gruenbacher et al., 2021).

• Convex Optimization (Discrete-Time):

  • For control systems, a Bellman recursion or open-loop convex program yields polytopic underapproximations of the set of initial states from which staying in a target tube is achievable with level $X_t$8. The method leverages log-concavity for set interpolation and convexity of reachable regions (Vinod et al., 2018).
• Set-Erosion in Stochastic Safety:
  • A "probabilistic tube" $X_t$9 quantifies deviation due to stochastic noise. Eroding the safety set by $$\Pr\left\{ X_t \in \Omega_t \text{ for all } t\in[0,T] \right\} \geq 1-\delta.$$0 reduces stochastic verification to a deterministic one, with the tube bound

  $$\Pr\left\{ X_t \in \Omega_t \text{ for all } t\in[0,T] \right\} \geq 1-\delta.$$1

  where typically $$\Pr\left\{ X_t \in \Omega_t \text{ for all } t\in[0,T] \right\} \geq 1-\delta.$$2 for suitable system-dependent $$\Pr\left\{ X_t \in \Omega_t \text{ for all } t\in[0,T] \right\} \geq 1-\delta.$$3 (Liu et al., 5 Mar 2025).

Techniques for Multi-Agent and MPC Settings

• Probabilistic Reachable Tubes for Multi-Agent Systems:

  • The error components of each agent evolve as decoupled processes, for which confidence regions (ellipsoidal, via Chebyshev) are propagated through time. By recursively composing one-step probabilistic reachable sets, a high-probability tube for the entire network state is achieved, with an explicit lower bound on joint probability via Boole's inequality (Vlahakis et al., 2024).
• Learning-Based Tube MPC:
  • Tube boundaries are learned via quantile regression or "Tube Loss" objectives to encapsulate trajectory uncertainty, yielding tubes that (with prescribed coverage) contain a fraction $$\Pr\left\{ X_t \in \Omega_t \text{ for all } t\in[0,T] \right\} \geq 1-\delta.$$4 of all possible regression targets or system states (Anand et al., 2024, Fan et al., 2020, Gao et al., 6 May 2025).

3. Probabilistic Tube Bounds in Hypoelliptic and Geometric Settings

Degenerate Diffusions and Non-Isotropic Tubes

• Diffusions under Hörmander Condition:
  • For SDEs where the diffusion matrix may be degenerate, but the Lie algebra generated by diffusion and bracket terms is full-rank, the tube is defined via a non-isotropic norm $$\Pr\left\{ X_t \in \Omega_t \text{ for all } t\in[0,T] \right\} \geq 1-\delta.$$5 reflecting different propagation speeds in direct and bracket directions. Two-sided exponential bounds for trajectory containment probabilities are established:

  $$\Pr\left\{ X_t \in \Omega_t \text{ for all } t\in[0,T] \right\} \geq 1-\delta.$$6

  where $$\Pr\left\{ X_t \in \Omega_t \text{ for all } t\in[0,T] \right\} \geq 1-\delta.$$7 captures local geometry and degeneracy (Bally et al., 2016, Bally et al., 2012).

Tube Volume and Random Geometry

• Volume of Tubular Neighborhoods (Algebraic Geometry):

  • For a manifold or variety $$\Pr\left\{ X_t \in \Omega_t \text{ for all } t\in[0,T] \right\} \geq 1-\delta.$$8 and tube of radius $$\Pr\left\{ X_t \in \Omega_t \text{ for all } t\in[0,T] \right\} \geq 1-\delta.$$9, the volume (and thus probability that a random point is within $x'(t) = f(x(t)), x(t_0) \in B_0 = B(x_0, r_0)$0 of $x'(t) = f(x(t)), x(t_0) \in B_0 = B(x_0, r_0)$1) can be bounded in terms of the variety's algebraic complexity (degree, codimension), via Weyl's tube formula, Crofton formula, and Bézout-type degree bounds. These estimates provide upper bounds for the probability of small-distance proximity in high-dimensional spaces (Lotz, 2012).
• Random Cantor Sets and Tube Intersections:
  • In geometric measure theory, random fractal sets can satisfy uniform upper bounds on Hausdorff measure within tubes, i.e.,

  $x'(t) = f(x(t)), x(t_0) \in B_0 = B(x_0, r_0)$2

  for all $x'(t) = f(x(t)), x(t_0) \in B_0 = B(x_0, r_0)$3, directly quantifying how mass concentrates within arbitrarily thin tubes (Chen, 2014).

4. Practical Construction and Computation of Probabilistic Tubes

Algorithmic Elements

• Sampling, Padding, and Concentration:

  • Sampling-based methods (e.g., GoTube) utilize random or quasi-random exploration of initial condition balls, updating tube radii with statistical concentration inequalities (DKW/Hoeffding) to maintain probabilistic validity (Gruenbacher et al., 2021).
• Convex Optimization:
  • Reach set underapproximations leverage epigraphical and line search convex programs, iteratively computing polytopic boundaries and interpolating coverage levels to obtain anytime probabilistic certificates (Vinod et al., 2018).
• Receding Horizon and Learning-Based Updates:
  • In control, tube cross-sections are refined via recursive propagation or online data-driven approaches (linear programs or neural quantile regressors), providing adaptive, scenario-based guarantees on containment and recursive feasibility (Vlahakis et al., 2024, Gao et al., 6 May 2025, Fan et al., 2020).

Parameterization and Tightness

• Control of Tightness and Level:
  • Parameters such as the tightness factor $x'(t) = f(x(t)), x(t_0) \in B_0 = B(x_0, r_0)$4, number of samples $x'(t) = f(x(t)), x(t_0) \in B_0 = B(x_0, r_0)$5, or tube width regularization in loss objectives allow practitioners to trade off between statistical conservatism and tube width. Probability mass can be allocated across time or state dimensions using the union bound (Gruenbacher et al., 2021, Anand et al., 2024).
• Explicit Formulas:
  • In certain linear/stochastic settings, closed-form expressions relate tube radius explicitly to system noise, contraction rates, time horizon, and target failure rates (Liu et al., 5 Mar 2025).

5. Theoretical Guarantees and Limitations

Asymptotic and Finite-Sample Guarantees

• Exponential Decay and Convergence:
  • As the number of samples $x'(t) = f(x(t)), x(t_0) \in B_0 = B(x_0, r_0)$6, statistical padding $x'(t) = f(x(t)), x(t_0) \in B_0 = B(x_0, r_0)$7 vanishes, and the coverage probability approaches the target level for both GoTube and convex polytope reach set methods (Gruenbacher et al., 2021, Vinod et al., 2018).
  • For random fields and tube methods, the relative error $x'(t) = f(x(t)), x(t_0) \in B_0 = B(x_0, r_0)$8 of tube formula approximation decays exponentially for subexponential tails but not for regularly varying heavy-tailed marginals, where the method may remain biased (Kuriki et al., 15 Jul 2025).
• Non-Isotropic, Degenerate Cases:
  • In hypoelliptic diffusions, tube bounds account for different scaling in tangent and bracket directions; the result is strictly stronger than results obtained through isotropic (Euclidean) norm tubes and is locally equivalent to the intrinsic control (Carathéodory) metric (Bally et al., 2016, Bally et al., 2012).

Limitations

• Sharpness and Failure Regimes:
  • In geometric measure settings, bounds on the mass in tubes are sharp up to the threshold $x'(t) = f(x(t)), x(t_0) \in B_0 = B(x_0, r_0)$9; above this, tube-nullity intervenes and upper bounds cannot hold (Chen, 2014).
  • For random field tube methods, regularly varying tails invalidate tail-accuracy, requiring alternative Bonferroni-type bounds or correction terms (Kuriki et al., 15 Jul 2025).
  • Volume-of-tube bounds become ineffective for varieties with high complexity or in high dimension due to exponential constants, although they remain polynomial in degree for fixed $t$0 (Lotz, 2012).

6. Applications and Impact

• Safety Verification & Model Checking: Probabilistic tubes are used for formal safety verification in neural ODEs, nonlinear stochastic systems, and controller synthesis, allowing tractable reduction of chance constraints to deterministic tightening (Gruenbacher et al., 2021, Liu et al., 5 Mar 2025, Vlahakis et al., 2024, Gao et al., 6 May 2025).
• Learning and Forecasting: Tube Loss and quantile-based algorithms deliver direct construction of optimal prediction intervals and forecast tubes in neural networks, achieving correct empirical coverage and minimal prediction width efficiently (Anand et al., 2024, Fan et al., 2020).
• Geometric Probability and Statistics: Probabilistic tube bounds control rare event probabilities, estimate small-volume neighborhoods in algebraic geometry, and enable precise calculation of tail exceedance probabilities for high-threshold random fields (Lotz, 2012, Kuriki et al., 15 Jul 2025).
• Random Graphs and Percolation Theory: Fluctuation lower bounds for first-passage percolation use tube-slicing and anti-concentration arguments to provide the first polynomial variance lower bounds in high-dimensional geometries (Damron et al., 2022).
• Geometric Measure Theory: The intersection behavior of random fractals with tubular neighborhoods answers open questions on mass concentration, applying probabilistic tube bounds in measure-theoretic and fractal contexts (Chen, 2014).

7. Summary Table: Main Approaches to Probabilistic Tube Bounds

Context/Setting	Construction/Method	Probabilistic Guarantee/Bound
Neural ODEs, Continuous Depth	GoTube: Sampling + DKW Padding	$t$1
Stochastic Control, Reachability	Convex Polytope, Dynamic Prog.	$t$2
SDEs, Nonlinear Stochastic Systems	Martingale/Concentration Analysis	$t$3
Learning-Based Forecasting	Tube Loss/Quantile Regression	Empirical coverage $t$4, width minimized in a single objective
Diffusions, Hypoelliptic Settings	Non-Isotropic Norm Tubes	Two-sided exponential probability bounds in bracket-adapted norm
Algebraic Geometry, Random Fields	Tube/Euler Characteristic Formula	Excursion probability, with relative error $t$5 (subexp. tails)

Probabilistic tube bounds thus provide a unifying principle for the explicit, computable, and probability-quantified enclosure of system trajectories, spanning domains from stochastic verification, robust control, learning-based forecasting, to geometric analysis of high-dimensional or fractal sets. The mathematical techniques combine sampling, convex optimization, concentration inequalities, martingale theory, geometric measure, and curvature integration, yielding robust guarantees in both classical and contemporary high-dimensional regimes.