In classical quantitative finance, risk management relies on a foundational assumption: we know the underlying probability distribution of future asset prices, or at least a single reference measure that defines which events are possible and which are impossible. When we evaluate different parameters—such as a stock's expected drift—the resulting probability measures remain equivalent, meaning they agree on what constitutes a "null set" (an event with zero probability).
However, real-world markets exhibit Knightian uncertainty, where the true probability distribution is unknown. A particularly severe form of this is volatility uncertainty. If we are uncertain about the future volatility of an asset, two different volatility paths generate probability measures that are mutually singular. They disagree entirely on which events are impossible; an asset path that is certain under one model may be mathematically impossible under another.
This scenario is known as nondominated model uncertainty. When the set of possible models is nondominated, classical mathematical finance breaks down. Standard Lebesgue spaces cannot be defined without a reference measure, standard martingale representation theorems fail, and classical functional analysis tools—like the Hahn-Banach theorem used to prove the absence of arbitrage—no longer apply.
To overcome this crisis, researchers have developed a completely new mathematical framework that replaces single-measure probability theory with nonlinear expectations, robust dualities, and second-order stochastic calculus.
Sublinear Expectations and Nonlinear Calculus
To model a world without a reference probability measure, we must abandon the classical expectation operator $\mathbb{E}[\cdot]$. Instead, we introduce a sublinear expectation operator, denoted as $\mathcal{E}[\cdot]$. Intuitively, if $\mathcal{P}$ represents the set of all plausible probability models, the sublinear expectation computes the worst-case scenario across all of them:
$$\mathcal{E}[X] = \sup_{P \in \mathcal{P}} \mathbb{E}_P[X]$$
The core challenge is ensuring that these expectations can be updated consistently over time as new information arrives. Nutz et al., in Constructing sublinear expectations on path space, solved this aggregation problem by leveraging the theory of analytic sets, constructing time-consistent conditional expectations that operate across a family of mutually singular measures without relying on an "almost surely" reference.
To model continuous asset prices within this framework, standard Brownian motion is replaced by $G$-Brownian motion. While a standard Brownian motion has a deterministic quadratic variation $\langle B \rangle_t = t$, a $G$-Brownian motion allows the quadratic variation (volatility) to fluctuate dynamically and adversarially within a specified range $[\underline{\sigma}, \bar{\sigma}]$.
This nonlinear calculus has been extended to handle complex boundary conditions. For instance, Lin in Stochastic differential equations driven by $G$-Brownian motion with reflecting boundary conditions and Soumana Hima in Reflected stochastic differential equations driven by G-Brownian motion in non-convex domains established the existence and uniqueness of solutions for reflected SDEs under volatility uncertainty. These tools allow researchers to model systems with hard constraints—such as price floors or wealth boundaries—without assuming a fixed volatility structure. To simplify these highly complex, oscillating systems, Han et al. in Mixed stochastic differential equations: Averaging principle result proved stochastic averaging principles, demonstrating that multi-valued systems under volatility uncertainty can be rigorously approximated by simpler, averaged systems.
Robust Pricing-Hedging Dualities
If we cannot trust a single model, how do we price a financial derivative? The solution is superreplication: finding the cheapest trading strategy that completely covers the derivative's payoff under every possible model in our uncertainty set.
The mathematical engine that makes this practical is the robust duality theorem. It states that the minimum cost required to physically superreplicate a claim is exactly equal to the highest theoretical price assigned to it by any valid model within the uncertainty set. Neufeld et al., in Superreplication under Volatility Uncertainty for Measurable Claims, proved this duality for highly irregular, measurable claims (such as digital options), showing that the superreplication price matches the supremum of expectations over the nondominated set of volatility models.
This framework allows researchers to interpolate between two historical extremes: purely model-dependent pricing (which is precise but fragile) and entirely model-free pricing (which is robust but yields unpractically wide price ranges).
Model-Specific Finance ───────► Prediction Sets ───────► Model-Free Finance
(Vulnerable to error) (Hou et al., Guyon et al.) (Impractically wide bounds)
To narrow these bounds, Hou et al., in Robust pricing-hedging dualities in continuous time, introduced "prediction sets"—mathematical constraints that restrict the family of measures to those matching specific, macro-level market beliefs.
This perspective directly connects robust finance to optimal transport. Instead of moving physical mass efficiently, we look for the most adversarial martingale probability measure that matches observed market data. For example, Guyon et al., in Bounds for VIX futures given S&P 500 smiles, framed the pricing of volatility futures as a constrained martingale optimal transport problem, yielding sharp, model-free bounds directly from market option prices. To analyze the geometry of these spaces, Li in Generalized Wasserstein distance and weak convergence of sublinear expectations generalized the classical Wasserstein distance to operate over sets of probability measures, providing the geometric metric needed to prove the convergence of sublinear expectations.
To unify the field, Obloj et al., in A unified framework for robust modelling of financial markets in discrete time, established the mathematical equivalence between the two dominant schools of robust finance: the quasi-sure approach (which uses sets of probability measures) and the pathwise approach (which ignores measures entirely and focuses strictly on possible price trajectories).
Dynamic Consistency and Portfolio Optimization
When an investor attempts to optimize a portfolio under model uncertainty, they face the trap of time-inconsistency. As Osogami et al. demonstrated in Time-Consistency of Optimization Problems, many standard risk-constrained optimization formulations are time-inconsistent: a plan that looks optimal today will no longer be considered optimal tomorrow purely because time has passed and new information has arrived.
To resolve this in portfolio selection, researchers developed robust forward performance criteria. Traditional utility maximization fixes a target at a terminal time horizon $T$, which forces the investor to commit to a specific model structure up front. In contrast, forward criteria allow the investment targets to evolve dynamically.
Obloj et al., in Dynamically consistent investment under model uncertainty: the robust forward criteria, successfully integrated model ambiguity into this forward framework, ensuring that the investor's strategies remain dynamically consistent even when the underlying market models are nondominated. Wang, in G-forward performance process and representation of homothetic case via ergodic quadratic G-BSDE, extended this to continuous time under $G$-Brownian motion, deriving optimal trading strategies that remain robust against worst-case volatility and drift scenarios simultaneously.
For general discrete-time settings, Nutz in Utility Maximization under Model Uncertainty in Discrete Time proved that robust optimal portfolios exist for bounded utility functions using measurable selection theory. To handle highly irregular or unbounded scenarios where model domination is completely absent, Bartl et al., in Robust expected utility maximization with medial limits, introduced sophisticated set-theoretic tools—specifically medial limits and Choquet's capacitability theorem—to prove the existence of optimal strategies without requiring absolute continuity with respect to any reference measure.
Overcoming the Curse of Dimensionality
Why does this framework work computationally? In classical finance, pricing and control problems lead to standard Backward Stochastic Differential Equations (BSDEs) or linear partial differential equations (PDEs). Under volatility uncertainty, because the driver must constantly adapt to the worst-case Hessian (the second derivative of the value function), the problems instead yield Second-Order Backward Stochastic Differential Equations (2BSDEs) or fully nonlinear PDEs.
Soner et al., in Superhedging and Dynamic Risk Measures under Volatility Uncertainty, formally proved that the cost of superhedging a claim under volatility uncertainty is precisely the solution to a 2BSDE. However, solving these equations using traditional grid-based numerical methods is computationally prohibitive for portfolios containing more than a few assets due to the curse of dimensionality.
The breakthrough that made these robust models practically applicable came from deep learning. Beck et al., in Machine learning approximation algorithms for high-dimensional fully nonlinear partial differential equations and second-order backward stochastic differential equations, introduced the Deep 2BSDE method.
Fully Nonlinear PDE / 2BSDE ──► Reformulated as a Multi-Stage Optimization Problem
│
▼
Deep Neural Networks approximate spatial gradients
│
▼
Bypasses the Curse of Dimensionality (Up to 100D)
By reformulating the 2BSDE as a multi-stage stochastic optimization problem and using deep neural networks to approximate the high-dimensional spatial gradients at each time step, the authors successfully solved fully nonlinear equations in up to 100 dimensions. This computational bridge allows the theoretical guarantees of nondominated uncertainty modeling to be deployed on real-world, multi-asset financial portfolios.
Key Foundational Work
General uncertainty in portfolio selection: A case-based decision approach (2007) — contributed 3 concept elements to this direction.
Weak approximation of <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" altimg="si1.gif" display="inline" overflow="scroll"><mml:mi>G</mml:mi></mml:math>-expectations (2011) — contributed 1 concept elements to this direction.