Sublinear Expectation Structure

• Sublinear expectation is a framework that generalizes classical expectations by aggregating multiple linear expectations via supremum/infimum to handle model uncertainty.
• It adheres to key axioms like monotonicity, constant preservation, positive homogeneity, and subadditivity, ensuring consistent risk and ambiguity evaluation.
• Applications span robust stochastic calculus, quantitative finance, and statistical inference, offering practical methods to deal with uncertain dynamics.

A sublinear expectation structure is a generalization of the classical linear expectation used in probability theory, designed to accommodate model uncertainty and ambiguity in probabilistic models. Instead of specifying a single probability measure, sublinear expectations consider a (possibly non-dominated) family of probability measures and aggregate by taking a supremum (worst-case) or infimum (best-case), thereby accommodating uncertainty in mean, variance, or the full distributional structure. This framework leads to robust stochastic calculus, generalized limit theorems, and risk measures suitable for applications in quantitative finance, statistics, and the theory of stochastic processes under ambiguity.

1. Defining Sublinear Expectation and Representation

A sublinear expectation is a functional $$\mathbb{E} : \mathcal{H} \to \mathbb{R}\cup\{+\infty\}$$ on a space of real-valued random variables (typically a vector lattice or linear space containing constants) satisfying:

• Monotonicity: $X \le Y$ pointwise $\implies \mathbb{E}[X] \le \mathbb{E}[Y]$
• Constant preservation: For all $c\in\mathbb{R}$, $\mathbb{E}[c]=c$
• Subadditivity: $\mathbb{E}[X+Y] \le \mathbb{E}[X] + \mathbb{E}[Y]$
• Positive homogeneity: For all $\lambda\ge0$, $\mathbb{E}[\lambda X] = \lambda\mathbb{E}[X]$

A fundamental result, established in both static and dynamic settings, is that any sublinear expectation $\mathbb{E}$ can be represented as the supremum over a collection of linear expectations: $\mathbb{E}[X] = \sup_{P\in\mathcal{P}} E_P[X]$ where $\mathcal{P}$ is a convex (possibly non-dominated) family of probability measures or, more generally, linear functionals that are dominated by $\mathbb{E}$ on the given space (Cohen, 2011, 0803.2656, Huang et al., 2024, Fang et al., 2017).

In this context, uncertainty is modeled not by a single "true" measure but by varying over all $P\in\mathcal{P}$, with the sublinear expectation acting as a worst-case aggregator.

2. Key Examples and Basic Structures

Sublinear expectations subsume a broad array of structures:

• Upper/lower expectation: $$\overline{\mathbb{E}}[X]:=\sup_{P\in\mathcal{P}} E_P[X]$$, $$\underline{\mathbb{E}}[X]:=-\,\sup_{P\in\mathcal{P}} E_P[-X]$$
• Capacity (non-additive set function): $V(A):=\sup_{P\in\mathcal{P}} P(A)$
• G-expectation, G-normal distributions, and G-Brownian motion: As formalized by Peng, the G-expectation models volatility and mean uncertainty via a sublinear operator linked to fully nonlinear PDEs and stochastic control (0803.2656, Nutz, 2010).

Several foundational results carry over from the linear case:

• Law of large numbers (LLN): Under suitable notions of independence and moment conditions, empirical averages converge quasi-surely to intervals determined by upper/lower means (Fang et al., 2017, Masasila et al., 20 Feb 2026, Huang et al., 2024).
• Central limit theorem (CLT): Normalized sums converge in distribution to a G-normal law, with limits characterized by a range of variances and solved using nonlinear PDEs (the G-heat equation).

Discrete time and finite state: In finite-state settings, sublinear expectations can be computed explicitly via repeated maximization over compact convex parameter sets (Yang et al., 2024). Discrete-time sublinear expectation and martingale theory parallels classical probability, with optional stopping, up/downcrossing inequalities, and convergence (Cohen et al., 2011).

Path space and Markovian settings: On $D([0,\infty),X)$ for countable $X$, downward-continuous sublinear expectations can be extended from finitely-dependent functions to a large class of measurable functionals (Erreygers, 2023). Sublinear Markov semigroups with bounded rates induce robust expectations and correspond to upper-envelope generators (Erreygers, 2023).

3. Conditional, Dynamic, and Time-Consistent Sublinear Expectation

Conditional sublinear expectations are defined analogously to classical conditional expectations but using essential suprema over conditional expectations under each $P\in\mathcal{P}$:

$\mathbb{E}[X|\mathcal{G}](\omega) := \esssup_{P\in\mathcal{P}} E_P[X|\mathcal{G}](\omega)$

When time-consistency (a dynamic programming principle) is imposed, one constructs a time-indexed family $\{\mathbb{E}_t\}$ satisfying the "tower property": $$\mathbb{E}_s(\mathbb{E}_t[X]) = \mathbb{E}_s[X], \qquad 0\le s\le t$$ The dynamic theory underlies robust stochastic control, risk measures, and BSDEs under ambiguity (Cohen, 2011, Nutz, 2010, Huang et al., 2024, Nutz et al., 2012, Cohen et al., 2011).

A major technical advance is the precise construction of pathwise versions of conditional sublinear expectations in nondominated models, requiring analytic set theory and quasi-sure analysis to resolve measurability and uniqueness issues, especially for Borel or upper semicontinuous payoffs (Nutz et al., 2012, Nutz, 2010).

4. Sublinear Expectations for Random Sets and Multivariate Structures

In higher dimensions and for random closed sets, sublinear expectations generalize to set-valued functionals mapping $L^p$-integrable random convex sets to deterministic convex sets. The axioms are adapted to set addition (Minkowski sum), set inclusions, and random convex geometry (Molchanov et al., 2019):

• Monotonicity: $X \subset Y$ a.s. $\implies \mathcal{E}(X)\subset \mathcal{E}(Y)$
• Positive homogeneity and subadditivity hold via set-valued operations
• Dual representation: Support functionals $h(\mathcal{E}(X),u)$ admit dual representations over cones of weights, generalizing the supremum over linear functionals in the scalar case

Minimal and maximal extensions, primal/dual constructions, and exactness of expectations are addressed using convex-analytic techniques.

5. Independence, Pseudo-independence, and Limit Theorems

Classical probabilistic independence splits into several nonequivalent notions under sublinear expectations. For random variables $X$ and $Y$, one says $Y$ is independent of $X$ under $\mathbb{E}$ if, for any test function $\varphi$,

$$\mathbb{E}[\varphi(X,Y)] = \mathbb{E}\bigl[\mathbb{E}[\varphi(x,Y)]_{x=X}\bigr]$$

Pseudo-independence is a related relaxation and is necessary for certain limit theorems (Li, 2021).

Strong law of large numbers and central limit theorems have been established for both i.i.d. and $\alpha$-mixing sequences under sublinear expectations, with convergence to intervals or maximal distributions rather than single points, reflecting mean or variance uncertainty. Explicit rates of convergence and ergodic results have been developed for invariant and mixing systems, including robust versions of Birkhoff's theorem (Masasila et al., 20 Feb 2026, Huang et al., 2024, Fang et al., 2017).

In path-dependent, stochastic control, or volatility-uncertain environments, sublinear expectations encode robust laws of large numbers and G-CLTs, thus subsuming both classical and fully nonlinear stochastic limit behavior (Nutz, 2010, 0803.2656, Masasila et al., 20 Feb 2026).

6. Extension, Duality, and Aggregation Techniques

The extension of sublinear expectations from finitely dependent (cylinder) functions to wider classes relies on downward (monotone) continuity conditions, convex duality, and Daniell–Stone arguments. For countable-state path spaces, downward continuity ensures unique, expectation-preserving extensions to measurable functionals (Erreygers, 2023).

Duality theorems show that every convex downward-continuous expectation on the finitary domain $D$ admits a supremal integral representation over a closed set of dominating measures: $$\widehat{E}[F] = \sup_{\substack{P:\ P[f]\le E[f]\ \forall f\in D}} \int F\,dP$$ with minimal penalties and uniqueness on appropriate subspaces (Erreygers, 2023).

The problem of aggregation (finding pathwise versions agreeing with all prior models) is subtle; measurability obstacles and non-uniqueness can arise except for very regular choices of functionals and sets of measures (Nutz et al., 2012).

7. Applications and Connections

Sublinear expectation structures support robust risk measures, stochastic calculus under volatility and drift uncertainty, nonlinear PDE representations (notably in stochastic control and finance), and inference in the presence of ambiguity or model misspecification.

Canonical applications include:

• Mathematical finance: pricing and hedging under model uncertainty, robust portfolio optimization, martingale and BSDE approaches with Knightian uncertainty
• Stochastic processes: G-Brownian motion, robust stochastic integration, control with uncertain dynamics
• Statistical inference: learning under ambiguity, worst-case law of large numbers and adaptive estimation (Huang et al., 2024, Fang et al., 2017, Huang et al., 2024)

Moreover, ergodic theory, invariant sublinear systems, and mixing phenomena are integrated into the framework, extending Birkhoff-type theorems and LLNs to ambiguous and nonlinear settings with applications to SDEs and capacity theory (Huang et al., 2024).

Table: Core Axioms and Representations

Property	Scalar Sublinear Expectation	Set-Valued Sublinear Expectation
Monotonicity	$X\le Y \implies \mathbb{E}[X]\le\mathbb{E}[Y]$	$$X\subset Y \implies \mathcal{E}(X)\subset\mathcal{E}(Y)$$
Constant preservation	$\mathbb{E}[c]=c$	$\mathcal{E}(F)\supset F$ (constant sets)
Positive homogeneity	$\mathbb{E}[\lambda X]=\lambda\mathbb{E}[X]$	$\mathcal{E}(\lambda X)=\lambda\mathcal{E}(X)$
Subadditivity	$\mathbb{E}[X+Y]\le\mathbb{E}[X]+\mathbb{E}[Y]$	$$\mathcal{E}(X+Y)\subset \mathcal{E}(X) + \mathcal{E}(Y)$$

The sublinear expectation structure thus unifies diverse robust methodologies in probability through a precise axiomatic and dual framework, aligning large deviation risk, robust inference, and nonlinear stochastic analysis under a single formalism (Huang et al., 2024, Masasila et al., 20 Feb 2026, 0803.2656, Fang et al., 2017, Molchanov et al., 2019, Erreygers, 2023).