g-Expectations¶

Introduction¶

The theory of g-expectations represents a fundamental generalization of classical linear expectation to nonlinear expectations that arise naturally in financial mathematics under model uncertainty, transaction costs, and risk aversion. Introduced by Shige Peng in 1997, g-expectations provide a unified framework for:

Nonlinear pricing: Prices that are not simply expectations under a single probability measure
Dynamic risk measures: Time-consistent measures of risk (e.g., dynamic coherent risk measures)
Robust valuation: Pricing under model uncertainty
Backward stochastic differential equations (BSDEs): Solutions to BSDEs define g-expectations

The mathematical foundation lies in the theory of backward stochastic differential equations (BSDEs) and provides a bridge between stochastic analysis, PDEs, and financial mathematics.

Mathematical Foundations¶

1. Classical Expectation¶

Linear Expectation: Given a probability space $(\Omega, \mathcal{F}, P)$ and random variable $\xi \in L^2(\Omega, \mathcal{F}, P)$:

\[ E[\xi] = \int_{\Omega} \xi(\omega) \, dP(\omega) \]

Properties: 1. Linearity: $E[a\xi + b\eta] = aE[\xi] + bE[\eta]$ 2. Monotonicity: $\xi \geq \eta \implies E[\xi] \geq E[\eta]$ 3. Constant preserving: $E[c] = c$ for constants $c$ 4. Tower property: $E[E[\xi|\mathcal{G}]] = E[\xi]$

Conditional Expectation: For sub-$\sigma$-algebra $\mathcal{G} \subseteq \mathcal{F}$:

\[ E[\xi|\mathcal{G}] \]

is the unique $\mathcal{G}$-measurable random variable satisfying:

\[ E[\xi \mathbb{1}_A] = E[E[\xi|\mathcal{G}] \mathbb{1}_A] \quad \text{for all } A \in \mathcal{G} \]

2. Backward Stochastic Differential Equations¶

BSDE Definition: A pair of processes $(Y_t, Z_t)_{t \in [0,T]}$ satisfying:

\[ Y_t = \xi + \int_t^T g(s, Y_s, Z_s) \, ds - \int_t^T Z_s \, dW_s \]

where: - $\xi$: Terminal condition (random variable, $\mathcal{F}_T$-measurable) - $g: [0,T] \times \mathbb{R} \times \mathbb{R}^d \to \mathbb{R}$: Generator (driver) - $W_t$: $d$-dimensional Brownian motion - $Y_t$: Solution process (value process) - $Z_t$: Control process (gradient process)

Differential Form:

\[ dY_t = -g(t, Y_t, Z_t) \, dt + Z_t \, dW_t, \quad Y_T = \xi \]

3. Well-Posedness¶

Theorem (Pardoux-Peng, 1990): Suppose: 1. $g$ is uniformly Lipschitz in $(y, z)$ 2. $\mathbb{E}\left[\int_0^T |g(t, 0, 0)|^2 dt\right] < \infty$ 3. $\mathbb{E}[|\xi|^2] < \infty$

Then there exists a unique adapted solution $(Y_t, Z_t)$ to the BSDE with:

\[ \mathbb{E}\left[\sup_{t \in [0,T]} |Y_t|^2 + \int_0^T |Z_t|^2 dt\right] < \infty \]

Proof Sketch: Use Picard iteration with contraction mapping in appropriate Banach space of adapted processes.

g-Expectation Definition¶

1. Conditional g-Expectation¶

Definition: Given generator $g$ and terminal condition $\xi$, the conditional g-expectation is:

\[ \mathcal{E}_g[\xi|\mathcal{F}_t] := Y_t \]

where $Y_t$ is the solution to the BSDE:

\[ Y_t = \xi + \int_t^T g(s, Y_s, Z_s) \, ds - \int_t^T Z_s \, dW_s \]

g-Expectation: The unconditional version is:

\[ \mathcal{E}_g[\xi] := \mathcal{E}_g[\xi|\mathcal{F}_0] = Y_0 \]

Interpretation: - Generalizes conditional expectation - Depends on entire path through generator $g$ - Nonlinear in general

2. Properties¶

Proposition (Basic Properties): For g-expectation $\mathcal{E}_g$:

Constant Preserving: $\mathcal{E}_g[c] = c$ for constants $c$
Monotonicity: If $\xi \geq \eta$ a.s., then $\mathcal{E}_g[\xi] \geq \mathcal{E}_g[\eta]$
Translation Invariance: $\mathcal{E}_g[\xi + c|\mathcal{F}_t] = \mathcal{E}_g[\xi|\mathcal{F}_t] + c$
Time Consistency:

$$ \mathcal{E}_g[\mathcal{E}_g[\xi|\mathcal{F}_t]|\mathcal{F}_s] = \mathcal{E}_g[\xi|\mathcal{F}_s] \quad \text{for } s \leq t $$

Proof (Time Consistency): Let $Y_t = \mathcal{E}_g[\xi|\mathcal{F}_t]$. Then:

\[ Y_s = Y_t + \int_s^t g(r, Y_r, Z_r) \, dr - \int_s^t Z_r \, dW_r \]

This BSDE with terminal condition $Y_t$ has solution $Y_s$ at time $s$, establishing time consistency.

3. Examples¶

Example 1 (Linear Case): If $g(t, y, z) = 0$, then:

\[ \mathcal{E}_g[\xi|\mathcal{F}_t] = E[\xi|\mathcal{F}_t] \]

the classical conditional expectation.

Example 2 (Exponential Utility): If $g(t, y, z) = -\frac{\gamma}{2} |z|^2$, then:

\[ \mathcal{E}_g[\xi] = -\frac{1}{\gamma} \log E[e^{-\gamma \xi}] \]

the certainty equivalent under exponential utility with risk aversion $\gamma$.

Example 3 (Worst-Case Expectation): If $g(t, y, z) = \alpha |z|$, then:

\[ \mathcal{E}_g[\xi] = \sup_{\mathbb{Q} \in \mathcal{P}_{\alpha}} E_{\mathbb{Q}}[\xi] \]

where $\mathcal{P}_{\alpha}$ is a set of probability measures with bounded density.

Generators and Their Properties¶

1. Lipschitz Generators¶

Definition: Generator $g$ is Lipschitz if there exists $K > 0$ such that:

\[ |g(t, y_1, z_1) - g(t, y_2, z_2)| \leq K(|y_1 - y_2| + |z_1 - z_2|) \]

for all $(t, y_1, z_1), (t, y_2, z_2)$.

Consequence: Lipschitz generators ensure unique solutions to BSDEs and well-defined g-expectations.

2. Convexity¶

Definition: Generator $g(t, y, z)$ is convex in $(y, z)$ if:

\[ g(t, \lambda y_1 + (1-\lambda) y_2, \lambda z_1 + (1-\lambda) z_2) \leq \lambda g(t, y_1, z_1) + (1-\lambda) g(t, y_2, z_2) \]

for all $\lambda \in [0, 1]$.

Implication: Convex generators lead to concave g-expectations:

\[ \mathcal{E}_g[\lambda \xi_1 + (1-\lambda) \xi_2] \geq \lambda \mathcal{E}_g[\xi_1] + (1-\lambda) \mathcal{E}_g[\xi_2] \]

Risk Aversion: Convexity in $z$ corresponds to risk aversion in the associated valuation.

3. Positive Homogeneity¶

Definition: Generator $g(t, y, z)$ is positively homogeneous in $(y, z)$ if:

\[ g(t, \lambda y, \lambda z) = \lambda g(t, y, z) \]

for all $\lambda > 0$.

Consequence: The corresponding g-expectation is positively homogeneous:

\[ \mathcal{E}_g[\lambda \xi] = \lambda \mathcal{E}_g[\xi] \]

for $\lambda > 0$.

4. Subadditivity¶

Definition: Generator $g$ induces subadditive g-expectation if:

\[ \mathcal{E}_g[\xi_1 + \xi_2] \leq \mathcal{E}_g[\xi_1] + \mathcal{E}_g[\xi_2] \]

Sufficient Condition: If $g(t, y, z)$ is convex and positively homogeneous in $(y, z)$, then $\mathcal{E}_g$ is subadditive.

Comparison Theorems¶

1. Comparison of Terminal Conditions¶

Theorem (Comparison): Let $(Y_t, Z_t)$ and $(\bar{Y}_t, \bar{Z}_t)$ be solutions to BSDEs with the same generator $g$ but terminal conditions $\xi$ and $\bar{\xi}$ respectively.

If $\xi \leq \bar{\xi}$ a.s., then:

\[ Y_t \leq \bar{Y}_t \quad \text{a.s. for all } t \in [0, T] \]

Proof: Define $\Delta Y_t = \bar{Y}_t - Y_t$ and $\Delta Z_t = \bar{Z}_t - Z_t$. Then:

\[ d(\Delta Y_t) = -[g(t, \bar{Y}_t, \bar{Z}_t) - g(t, Y_t, Z_t)] \, dt + \Delta Z_t \, dW_t \]

Apply Itô's formula to $(\Delta Y_t)^-$ (negative part) and use Lipschitz property to show $(\Delta Y_t)^- = 0$.

2. Comparison of Generators¶

Theorem: Let $(Y_t, Z_t)$ and $(\bar{Y}_t, \bar{Z}_t)$ be solutions to BSDEs with generators $g$ and $\bar{g}$ respectively, both with the same terminal condition $\xi$.

If $g(t, y, z) \geq \bar{g}(t, y, z)$ for all $(t, y, z)$, then:

\[ Y_t \leq \bar{Y}_t \quad \text{a.s. for all } t \in [0, T] \]

Interpretation: Larger generator → Larger "running cost" → Smaller value process.

Application: If $g_1 \leq g_2$, then:

\[ \mathcal{E}_{g_1}[\xi] \geq \mathcal{E}_{g_2}[\xi] \]

3. Strict Comparison¶

Theorem (Strict Comparison): Under additional regularity (strict inequality on a set of positive measure), strict inequalities hold:

\[ \xi < \bar{\xi} \text{ a.s. on set } A \text{ with } P(A) > 0 \implies Y_t < \bar{Y}_t \text{ a.s.} \]

Representation Theorems¶

1. Minimal and Maximal Representations¶

Theorem (Peng): For convex and positively homogeneous generator $g$, the g-expectation admits representations:

\[ \mathcal{E}_g[\xi] = \sup_{\mathbb{Q} \in \mathcal{Q}} E_{\mathbb{Q}}[\xi] = \inf_{\mathbb{Q} \in \mathcal{Q}^c} E_{\mathbb{Q}}[\xi] \]

where: - $\mathcal{Q}$: Set of "admissible" probability measures - $\mathcal{Q}^c$: Complement set

Construction: The set $\mathcal{Q}$ is characterized through the generator $g$ via Girsanov's theorem.

2. Entropic Representation¶

Exponential Utility Case: For $g(t, y, z) = -\frac{\gamma}{2} |z|^2$:

\[ \mathcal{E}_g[\xi] = \sup_{\mathbb{Q} \ll P} \left\{ E_{\mathbb{Q}}[\xi] - \frac{1}{\gamma} H(\mathbb{Q}|P) \right\} \]

where:

\[ H(\mathbb{Q}|P) = E_{\mathbb{Q}}\left[\log \frac{d\mathbb{Q}}{dP}\right] \]

is the relative entropy (Kullback-Leibler divergence).

Proof: The optimal measure $\mathbb{Q}^*$ has Radon-Nikodym derivative:

\[ \frac{d\mathbb{Q}^*}{dP}\bigg|_{\mathcal{F}_T} = \exp\left(-\gamma \xi + \frac{\gamma^2}{2} \int_0^T |Z_t|^2 dt - \gamma \int_0^T Z_t \, dW_t\right) \]

where $Z_t$ comes from the BSDE solution.

3. Choquet Capacity Representation¶

Non-Additive Measure: For certain generators, g-expectations correspond to integration with respect to Choquet capacities.

Capacity: A set function $\nu: \mathcal{F} \to [0, 1]$ with: 1. $\nu(\emptyset) = 0$, $\nu(\Omega) = 1$ 2. Monotonicity: $A \subseteq B \implies \nu(A) \leq \nu(B)$

Choquet Integral:

\[ \int_{\Omega} \xi \, d\nu = \int_0^{\infty} \nu(\{\xi \geq t\}) \, dt + \int_{-\infty}^0 [\nu(\{\xi \geq t\}) - 1] \, dt \]

Connection: Certain g-expectations can be written as Choquet integrals with appropriately defined capacity.

Dynamic Risk Measures¶

1. Coherent Risk Measures¶

Definition (Artzner et al., 1999): A functional $\rho: L^{\infty} \to \mathbb{R}$ is a coherent risk measure if:

Monotonicity: $X \geq Y \implies \rho(X) \leq \rho(Y)$
Translation invariance: $\rho(X + c) = \rho(X) - c$
Positive homogeneity: $\rho(\lambda X) = \lambda \rho(X)$ for $\lambda > 0$
Subadditivity: $\rho(X + Y) \leq \rho(X) + \rho(Y)$

Example: Average Value-at-Risk (AVaR or CVaR):

\[ \text{AVaR}_{\alpha}(X) = \frac{1}{\alpha} \int_0^{\alpha} \text{VaR}_u(X) \, du \]

2. Dynamic Coherent Risk Measures¶

Definition: A family $\{\rho_t\}_{t \in [0,T]}$ is a dynamic coherent risk measure if each $\rho_t$ is coherent and satisfies time consistency:

\[ \rho_s(\rho_t(X)) = \rho_s(X) \quad \text{for } s \leq t \]

Theorem (Delbaen et al., 2010): Dynamic coherent risk measures correspond to g-expectations with generators that are convex, positively homogeneous, and Lipschitz.

Representation:

\[ \rho_t(X) = \mathcal{E}_g[-X|\mathcal{F}_t] \]

for appropriate generator $g$.

3. Connection to BSDEs¶

Theorem: If $\rho_t$ is a dynamic coherent risk measure, then $Y_t = \rho_t(X)$ satisfies a BSDE:

\[ dY_t = g(t, Y_t, Z_t) \, dt - Z_t \, dW_t, \quad Y_T = X \]

where $g$ is convex and positively homogeneous in $(y, z)$.

Converse: Given such a generator, the g-expectation defines a dynamic coherent risk measure.

Applications to Finance¶

1. Option Pricing Under Model Uncertainty¶

Setup: Uncertain volatility $\sigma \in [\underline{\sigma}, \overline{\sigma}]$.

Robust Price: The seller's (super-replication) price is:

\[ V_0 = \mathcal{E}_g[\Phi(S_T)] \]

where:

\[ g(t, y, z) = \sup_{\sigma \in [\underline{\sigma}, \overline{\sigma}]} \left\{ -\frac{1}{2} \sigma^2 |z|^2 \right\} = -\frac{1}{2} \underline{\sigma}^2 |z|^2 \]

for positive gamma positions.

BSDE:

\[ dY_t = \frac{1}{2} \underline{\sigma}^2 |Z_t|^2 \, dt + Z_t \, dW_t, \quad Y_T = \Phi(S_T) \]

Hedging Strategy: The optimal hedge is $\Delta_t = Z_t$.

2. Utility Indifference Pricing¶

Setup: Agent with exponential utility $u(x) = -e^{-\gamma x}$ and risk aversion $\gamma > 0$.

Indifference Price: Price $p$ such that:

\[ \sup_{\theta} E[-e^{-\gamma(X_T^{\theta, 0})}] = \sup_{\theta} E[-e^{-\gamma(X_T^{\theta, p} + \Phi)}] \]

where $X_T^{\theta, v}$ is terminal wealth from initial capital $v$ and strategy $\theta$.

g-Expectation: The indifference price satisfies:

\[ p = \mathcal{E}_g[\Phi] \]

with generator:

\[ g(t, y, z) = -\frac{\gamma}{2} |z - \sigma S_t|^2 + \frac{\gamma}{2} \sigma^2 S_t^2 \]

3. Optimal Investment with Ambiguity¶

Problem: Maximize worst-case expected utility:

\[ \sup_{\pi} \inf_{\mathbb{Q} \in \mathcal{Q}} E_{\mathbb{Q}}[u(X_T^{\pi})] \]

where $\mathcal{Q}$ is a set of probability measures representing ambiguity.

Solution: The value function satisfies:

\[ V(t, x) = \mathcal{E}_g[u(X_T) | X_t = x] \]

with appropriate generator encoding ambiguity aversion.

Optimal Strategy: Extracted from the $Z$ process in the BSDE solution:

\[ \pi_t^* = f(Z_t, X_t) \]

4. Credit Risk and CVA¶

Credit Valuation Adjustment (CVA): Adjustment for counterparty default risk.

g-Expectation Framework: The CVA can be computed using:

\[ \text{CVA} = \mathcal{E}_g[\text{Loss at default}] \]

with generator reflecting uncertainty about default intensity and recovery rates.

Advantage: Captures model uncertainty and ambiguity in credit modeling.

Numerical Methods¶

1. Discrete-Time Approximation¶

Euler Scheme: Partition $[0, T]$ into $N$ intervals with $\Delta t = T/N$.

Backward Iteration: Starting from $Y_T = \xi$:

\[ Y_{t_i} = Y_{t_{i+1}} + g(t_i, Y_{t_i}, Z_{t_i}) \Delta t - Z_{t_i} \Delta W_{t_i} \]

Z Estimation: Use least-squares projection:

\[ Z_{t_i} = \arg\min_{z} E\left[\left|Y_{t_{i+1}} - Y_{t_i} - g(t_i, Y_{t_i}, z) \Delta t + z \Delta W_{t_i}\right|^2 | \mathcal{F}_{t_i}\right] \]

Convergence: Under Lipschitz conditions:

\[ E[|Y_0 - Y_0^N|] = O(\Delta t^{1/2}) \]

2. Monte Carlo Methods¶

Algorithm: 1. Simulate $M$ paths of Brownian motion: $\{W^{(m)}_t\}_{m=1}^M$ 2. At each time $t_i$, estimate conditional expectations using regression 3. Backward iterate to compute $Y_0$

Regression: Use basis functions $\{\phi_j\}$ to approximate:

\[ E[Y_{t_{i+1}} | \mathcal{F}_{t_i}] \approx \sum_{j=1}^K \alpha_j \phi_j(S_{t_i}) \]

Complexity: $O(MNK)$ where $M$ is paths, $N$ is time steps, $K$ is basis functions.

3. Deep Learning Approaches¶

Neural Network Parameterization: Represent $(Y_t, Z_t)$ using neural networks:

\[ Y_t = f_{\theta_Y}(t, S_t), \quad Z_t = f_{\theta_Z}(t, S_t) \]

Training: Minimize loss function:

\[ \mathcal{L}(\theta_Y, \theta_Z) = E\left[\left|Y_T - \xi\right|^2 + \int_0^T \left|dY_t + g(t, Y_t, Z_t) dt - Z_t dW_t\right|^2\right] \]

Advantages: - Handles high-dimensional problems - Avoids curse of dimensionality - Scales well with complexity

Advanced Topics¶

1. Reflected BSDEs¶

Definition: A triple $(Y_t, Z_t, K_t)$ satisfying:

\[ Y_t = \xi + \int_t^T g(s, Y_s, Z_s) \, ds + K_T - K_t - \int_t^T Z_s \, dW_s \]

with: 1. $Y_t \geq S_t$ (obstacle constraint) 2. $K_t$ is increasing, continuous, $K_0 = 0$ 3. $\int_0^T (Y_t - S_t) \, dK_t = 0$ (Skorokhod condition)

Interpretation: $K_t$ is the minimal pushing force to keep $Y_t$ above obstacle $S_t$.

Application: American option pricing, optimal stopping problems.

2. Mean-Field BSDEs¶

Setup: Large population of agents, each influenced by others.

Mean-Field BSDE:

\[ Y_t^i = \xi^i + \int_t^T g(s, Y_s^i, Z_s^i, \bar{Y}_s, \bar{Z}_s) \, ds - \int_t^T Z_s^i \, dW_s^i \]

where $\bar{Y}_t = \frac{1}{N} \sum_{i=1}^N Y_t^i$ is the empirical mean.

Limit: As $N \to \infty$, the system converges to a McKean-Vlasov type BSDE.

Application: Systemic risk, large portfolio optimization, crowd behavior in markets.

3. Forward-Backward SDEs¶

Coupled System: $(X_t, Y_t, Z_t)$ satisfying:

\[ dX_t = b(t, X_t, Y_t, Z_t) \, dt + \sigma(t, X_t, Y_t, Z_t) \, dW_t \]

\[ dY_t = -g(t, X_t, Y_t, Z_t) \, dt + Z_t \, dW_t \]

with $X_0 = x$ and $Y_T = \Phi(X_T)$.

Application: Stochastic control problems where control affects both forward dynamics and backward valuation.

Well-Posedness: Requires careful analysis; solutions exist under monotonicity or small-time conditions.

4. Quadratic BSDEs¶

Generator: Quadratic growth in $z$:

\[ g(t, y, z) = f(t, y) + \frac{1}{2} z^\top A(t, y) z \]

Challenge: Standard Lipschitz theory doesn't apply directly.

Existence: Proven under specific conditions (Kobylanski, 2000).

Application: - Exponential utility maximization - Risk-sensitive control - Large deviations

Connections to PDEs¶

1. Feynman-Kac Formula for BSDEs¶

Theorem: If $Y_t = v(t, X_t)$ for some function $v$ and state process $X_t$, then $v$ satisfies the PDE:

\[ \frac{\partial v}{\partial t} + \mathcal{L} v + g(t, v, \sigma^\top \nabla v) = 0 \]

with terminal condition $v(T, x) = \Phi(x)$, where $\mathcal{L}$ is the infinitesimal generator of $X_t$.

Proof: Apply Itô's formula to $v(t, X_t)$ and match terms with the BSDE.

Interpretation: BSDEs provide probabilistic representation for nonlinear PDEs.

2. Viscosity Solutions¶

Definition: A function $v$ is a viscosity solution if it satisfies the PDE in the viscosity sense (comparison with smooth test functions).

Theorem (Crandall-Ishii-Lions): Under appropriate conditions, the BSDE solution $Y_t = v(t, X_t)$ where $v$ is the unique viscosity solution to the associated PDE.

Advantage: Viscosity solutions exist even when classical solutions don't (e.g., non-smooth payoffs).

3. Hamilton-Jacobi-Bellman Equations¶

Stochastic Control: Consider:

\[ V(t, x) = \sup_{\alpha \in \mathcal{A}} E\left[\int_t^T f(s, X_s^{\alpha}, \alpha_s) \, ds + \Phi(X_T^{\alpha}) \bigg| X_t = x\right] \]

HJB Equation:

\[ \frac{\partial V}{\partial t} + \sup_{\alpha} \{\mathcal{L}^{\alpha} V + f(t, x, \alpha)\} = 0 \]

BSDE Connection: The value function satisfies a BSDE with generator:

\[ g(t, y, z) = \sup_{\alpha} \{-f(t, x, \alpha) - \mu^{\alpha}(t, x) \cdot \nabla v - \frac{1}{2} \text{tr}[(\sigma^{\alpha})^2 \nabla^2 v]\} \]

Convergence and Stability¶

1. Continuous Dependence¶

Theorem: Let $(Y_t^n, Z_t^n)$ be solutions to BSDEs with generators $g_n$ and terminal conditions $\xi_n$.

If $g_n \to g$ and $\xi_n \to \xi$ in appropriate norms, then:

\[ Y_t^n \to Y_t, \quad Z_t^n \to Z_t \]

in $L^2$ norm.

Proof: Use Lipschitz continuity and Gronwall's inequality.

Application: Justifies numerical approximations and perturbation analysis.

2. Convergence of Discretizations¶

Theorem: The discrete-time approximation converges to the continuous-time BSDE solution:

\[ \sup_{t \in [0,T]} E[|Y_t - Y_t^N|^2] = O(\Delta t) \]

under appropriate regularity conditions.

Higher-Order Schemes: Milstein-type schemes achieve $O((\Delta t)^{3/2})$ or better.

3. Robustness to Model Perturbations¶

g-Expectations Stability: Small changes in generator lead to small changes in g-expectation:

\[ |g_1 - g_2| \leq \epsilon \implies |\mathcal{E}_{g_1}[\xi] - \mathcal{E}_{g_2}[\xi]| \leq C \epsilon \]

for some constant $C$ depending on $T$ and Lipschitz constants.

Summary and Key Insights¶

1. Fundamental Contributions¶

Nonlinear Expectations: g-expectations extend classical expectations to nonlinear frameworks, capturing risk aversion, ambiguity, and model uncertainty.
BSDE Solutions: Provide constructive method to compute g-expectations through backward stochastic differential equations.
Time Consistency: g-expectations are dynamically consistent, essential for multi-period decision-making and risk management.
Representation Theorems: Connect g-expectations to multiple probability measures, capacities, and supremum/infimum operations.
Dynamic Risk Measures: Provide mathematical foundation for coherent, time-consistent risk measures used in finance and insurance.

2. Practical Implications¶

For Pricing: - Robust pricing under model uncertainty - Utility-based valuation - Credit valuation adjustments

For Risk Management: - Dynamic coherent risk measures - Stress testing frameworks - Regulatory capital calculations

For Portfolio Optimization: - Ambiguity-averse portfolio selection - Robust optimal control - Mean-field game equilibria

3. Theoretical Significance¶

g-Expectations unify: - Stochastic Analysis: BSDEs and martingale theory - PDE Theory: Viscosity solutions and Hamilton-Jacobi equations - Optimization: Stochastic control and minimax problems - Finance: Pricing, hedging, and risk management

The theory continues to evolve with extensions to: - Mean-field limits - Infinite-dimensional settings - Non-Markovian frameworks - Machine learning integration

g-Expectations represent a cornerstone of modern mathematical finance, providing rigorous foundations for robust valuation and decision-making under uncertainty.

Exercises¶

Exercise 1. For a linear generator $g(t, y, z) = \mu y + \lambda z$, show that the g-expectation reduces to a standard expectation under a change of measure. Specifically, prove that $\mathcal{E}_g[\xi] = \mathbb{E}_\mathbb{Q}[e^{-\mu T}\xi]$ where $\mathbb{Q}$ is the measure with Girsanov kernel $\lambda$.

Solution to Exercise 1

Goal. Show that for the linear generator $g(t, y, z) = \mu y + \lambda z$, we have $\mathcal{E}_g[\xi] = E_\mathbb{Q}[e^{-\mu T}\xi]$ where $\mathbb{Q}$ is the measure with Girsanov kernel $\lambda$.

Step 1: Write the BSDE. The BSDE is:

\[ -dY_t = (\mu Y_t + \lambda Z_t) \, dt - Z_t \, dW_t, \quad Y_T = \xi \]

Step 2: Change of measure. Define $\mathbb{Q}$ by the Radon-Nikodym derivative:

\[ \frac{d\mathbb{Q}}{dP}\bigg|_{\mathcal{F}_T} = \mathcal{E}\left(\int_0^T \lambda \, dW_s\right) = \exp\left(\lambda W_T - \frac{\lambda^2}{2}T\right) \]

By Girsanov's theorem, $\tilde{W}_t = W_t - \lambda t$ is a $\mathbb{Q}$-Brownian motion. Thus $dW_t = d\tilde{W}_t + \lambda \, dt$, and the BSDE becomes:

\[ -dY_t = (\mu Y_t + \lambda Z_t) \, dt - Z_t(d\tilde{W}_t + \lambda \, dt) = \mu Y_t \, dt - Z_t \, d\tilde{W}_t \]

Step 3: Solve under $\mathbb{Q}$. Define $\hat{Y}_t = e^{\mu t} Y_t$. By the product rule:

\[ d\hat{Y}_t = \mu e^{\mu t} Y_t \, dt + e^{\mu t} \, dY_t = \mu e^{\mu t} Y_t \, dt + e^{\mu t}(-\mu Y_t \, dt + Z_t \, d\tilde{W}_t) = e^{\mu t} Z_t \, d\tilde{W}_t \]

So $\hat{Y}_t$ is a $\mathbb{Q}$-local martingale. Under integrability conditions ($E_\mathbb{Q}[\int_0^T e^{2\mu t}|Z_t|^2 \, dt] < \infty$), it is a true $\mathbb{Q}$-martingale. Taking $\mathbb{Q}$-expectations:

\[ e^{\mu t} Y_t = E_\mathbb{Q}[e^{\mu T} Y_T | \mathcal{F}_t] = E_\mathbb{Q}[e^{\mu T} \xi | \mathcal{F}_t] \]

Therefore:

\[ Y_t = E_\mathbb{Q}[e^{-\mu(T-t)} \xi | \mathcal{F}_t] \]

At $t = 0$:

\[ \mathcal{E}_g[\xi] = Y_0 = E_\mathbb{Q}[e^{-\mu T} \xi] \]

This shows that the g-expectation with a linear generator reduces to a standard discounted expectation under a changed measure. The parameter $\mu$ acts as a discount rate and $\lambda$ as a Girsanov drift adjustment (market price of risk). $\square$

Exercise 2. Prove that the g-expectation $\mathcal{E}_g[\xi] = Y_0$ where $(Y, Z)$ solves $-dY_t = g(t, Y_t, Z_t) \, dt - Z_t \, dW_t$ with $Y_T = \xi$ satisfies monotonicity: if $\xi_1 \leq \xi_2$ a.s., then $\mathcal{E}_g[\xi_1] \leq \mathcal{E}_g[\xi_2]$. Use the comparison theorem for BSDEs.

Solution to Exercise 2

Goal. Prove monotonicity: $\xi_1 \leq \xi_2$ a.s. implies $\mathcal{E}_g[\xi_1] \leq \mathcal{E}_g[\xi_2]$.

Step 1: Set up. Let $(Y^1_t, Z^1_t)$ solve the BSDE with terminal condition $\xi_1$, and $(Y^2_t, Z^2_t)$ solve the BSDE with terminal condition $\xi_2$, both with the same generator $g$:

\[ Y^i_t = \xi_i + \int_t^T g(s, Y^i_s, Z^i_s) \, ds - \int_t^T Z^i_s \, dW_s, \quad i = 1, 2 \]

We need to show $Y^1_0 \leq Y^2_0$.

Step 2: Difference process. Define $\Delta Y_t = Y^1_t - Y^2_t$ and $\Delta Z_t = Z^1_t - Z^2_t$. Then:

\[ \Delta Y_t = (\xi_1 - \xi_2) + \int_t^T [g(s, Y^1_s, Z^1_s) - g(s, Y^2_s, Z^2_s)] \, ds - \int_t^T \Delta Z_s \, dW_s \]

Step 3: Linearization. Since $g$ is Lipschitz, we can write:

\[ g(s, Y^1_s, Z^1_s) - g(s, Y^2_s, Z^2_s) = \alpha_s \Delta Y_s + \beta_s \cdot \Delta Z_s \]

where $\alpha_s$ and $\beta_s$ are bounded adapted processes (by the Lipschitz property, $|\alpha_s| \leq K$ and $|\beta_s| \leq K$). This linearization follows from the mean value theorem applied componentwise.

Step 4: Apply Ito's formula. Consider $(\Delta Y_t)^+ = \max(\Delta Y_t, 0)$. We apply the comparison argument more directly. Define the measure $\tilde{\mathbb{Q}}$ via:

\[ \frac{d\tilde{\mathbb{Q}}}{dP}\bigg|_{\mathcal{F}_T} = \mathcal{E}\left(\int_0^T \beta_s \, dW_s\right) \]

Under $\tilde{\mathbb{Q}}$, define $\tilde{W}_t = W_t - \int_0^t \beta_s \, ds$. The difference satisfies:

\[ d(\Delta Y_t) = -\alpha_t \Delta Y_t \, dt + \Delta Z_t \, d\tilde{W}_t \]

Step 5: Solve. Define $\Phi_t = \exp(\int_0^t \alpha_s \, ds)$. Then $\Phi_t \Delta Y_t$ is a $\tilde{\mathbb{Q}}$-martingale:

\[ d(\Phi_t \Delta Y_t) = \Phi_t \Delta Z_t \, d\tilde{W}_t \]

Taking $\tilde{\mathbb{Q}}$-expectations:

\[ \Delta Y_0 = E_{\tilde{\mathbb{Q}}}\left[\exp\left(\int_0^T \alpha_s \, ds\right)(\xi_1 - \xi_2)\right] \]

Step 6: Conclude. Since $\xi_1 \leq \xi_2$ a.s., we have $\xi_1 - \xi_2 \leq 0$ a.s. The exponential factor $\exp(\int_0^T \alpha_s \, ds) > 0$. Therefore:

\[ \Delta Y_0 = E_{\tilde{\mathbb{Q}}}\left[\Phi_T (\xi_1 - \xi_2)\right] \leq 0 \]

which gives $Y^1_0 \leq Y^2_0$, i.e., $\mathcal{E}_g[\xi_1] \leq \mathcal{E}_g[\xi_2]$. $\square$

Exercise 3. The generator $g(z) = \frac{1}{2}\overline{\sigma}^2 z^+ - \frac{1}{2}\underline{\sigma}^2 z^-$ (where $z^+ = \max(z, 0)$ and $z^- = \max(-z, 0)$) corresponds to the uncertain volatility model. Show that the resulting g-expectation is sublinear: $\mathcal{E}_g[\xi_1 + \xi_2] \leq \mathcal{E}_g[\xi_1] + \mathcal{E}_g[\xi_2]$. What financial interpretation does sublinearity have for derivative pricing?

Solution to Exercise 3

Goal. Show that the g-expectation with generator $g(z) = \frac{1}{2}\overline{\sigma}^2 z^+ - \frac{1}{2}\underline{\sigma}^2 z^-$ is sublinear.

Step 1: Representation as supremum. The generator $g(z) = \frac{1}{2}\overline{\sigma}^2 z^+ - \frac{1}{2}\underline{\sigma}^2 z^-$ can be rewritten as:

\[ g(z) = \sup_{\sigma \in [\underline{\sigma}, \overline{\sigma}]} \frac{1}{2}\sigma^2 z \]

To verify: if $z \geq 0$, the supremum is attained at $\sigma = \overline{\sigma}$, giving $\frac{1}{2}\overline{\sigma}^2 z = \frac{1}{2}\overline{\sigma}^2 z^+$. If $z < 0$, the supremum is attained at $\sigma = \underline{\sigma}$ (minimizing $|\sigma^2 z|$ when $z < 0$), giving $\frac{1}{2}\underline{\sigma}^2 z = -\frac{1}{2}\underline{\sigma}^2 z^-$.

Step 2: Verify sublinearity of $g$. The function $g$ is convex and positively homogeneous:

Convexity: $g$ is the supremum of linear functions $z \mapsto \frac{1}{2}\sigma^2 z$, hence convex.
Positive homogeneity: $g(\lambda z) = \lambda g(z)$ for $\lambda > 0$, since $(\lambda z)^+ = \lambda z^+$ and $(\lambda z)^- = \lambda z^-$.

Step 3: From generator properties to g-expectation sublinearity. We use the representation theorem for g-expectations with convex, positively homogeneous generators. Such a g-expectation admits the dual representation:

\[ \mathcal{E}_g[\xi] = \sup_{\mathbb{Q} \in \mathcal{Q}} E_\mathbb{Q}[\xi] \]

where $\mathcal{Q}$ is the set of probability measures corresponding to volatility choices $\sigma_t \in [\underline{\sigma}, \overline{\sigma}]$.

Step 4: Prove sublinearity from the representation. For any $\xi_1, \xi_2$:

\[ \mathcal{E}_g[\xi_1 + \xi_2] = \sup_{\mathbb{Q} \in \mathcal{Q}} E_\mathbb{Q}[\xi_1 + \xi_2] = \sup_{\mathbb{Q} \in \mathcal{Q}} \{E_\mathbb{Q}[\xi_1] + E_\mathbb{Q}[\xi_2]\} \]

Since the supremum of a sum is at most the sum of suprema:

\[ \sup_{\mathbb{Q} \in \mathcal{Q}} \{E_\mathbb{Q}[\xi_1] + E_\mathbb{Q}[\xi_2]\} \leq \sup_{\mathbb{Q}_1 \in \mathcal{Q}} E_{\mathbb{Q}_1}[\xi_1] + \sup_{\mathbb{Q}_2 \in \mathcal{Q}} E_{\mathbb{Q}_2}[\xi_2] = \mathcal{E}_g[\xi_1] + \mathcal{E}_g[\xi_2] \]

The inequality is strict in general because the worst-case measure for $\xi_1 + \xi_2$ need not be the same as the individual worst-case measures for $\xi_1$ and $\xi_2$ separately.

Financial interpretation of sublinearity. Sublinearity means that the robust price of a portfolio is at most the sum of the robust prices of its components:

\[ V^{\text{sup}}(\xi_1 + \xi_2) \leq V^{\text{sup}}(\xi_1) + V^{\text{sup}}(\xi_2) \]

This has the interpretation of a diversification benefit: when pricing two derivatives together, the worst-case volatility scenario for the combined position may differ from the individual worst cases, leading to a lower combined price. For example, if $\xi_1$ has positive gamma and $\xi_2$ has negative gamma, the combined position may have reduced gamma exposure, narrowing the volatility uncertainty premium.

From a risk measure perspective ($\rho(\xi) = \mathcal{E}_g[-\xi]$), sublinearity of $\mathcal{E}_g$ corresponds to subadditivity of $\rho$, which is one of the axioms of coherent risk measures and expresses the principle that diversification should not increase risk. $\square$

Exercise 4. The conditional g-expectation $\mathcal{E}_g[\xi | \mathcal{F}_t] = Y_t$ satisfies the tower property: $\mathcal{E}_g[\mathcal{E}_g[\xi | \mathcal{F}_s] | \mathcal{F}_t] = \mathcal{E}_g[\xi | \mathcal{F}_t]$ for $t \leq s$. Prove this using the uniqueness of BSDE solutions, and explain why this property is essential for dynamic risk management.

Solution to Exercise 4

Goal. Prove the tower property $\mathcal{E}_g[\mathcal{E}_g[\xi | \mathcal{F}_s] | \mathcal{F}_t] = \mathcal{E}_g[\xi | \mathcal{F}_t]$ for $t \leq s$ using uniqueness of BSDE solutions.

Step 1: Define the processes. Let $(Y_r, Z_r)_{r \in [0,T]}$ be the unique solution to the BSDE:

\[ Y_r = \xi + \int_r^T g(u, Y_u, Z_u) \, du - \int_r^T Z_u \, dW_u \]

By definition, $\mathcal{E}_g[\xi | \mathcal{F}_r] = Y_r$ for all $r \in [0, T]$.

In particular, $\mathcal{E}_g[\xi | \mathcal{F}_s] = Y_s$.

Step 2: Consider the auxiliary BSDE. Now consider the BSDE on $[0, s]$ with terminal condition $Y_s$ at time $s$. Let $(\tilde{Y}_r, \tilde{Z}_r)_{r \in [0,s]}$ solve:

\[ \tilde{Y}_r = Y_s + \int_r^s g(u, \tilde{Y}_u, \tilde{Z}_u) \, du - \int_r^s \tilde{Z}_u \, dW_u \]

By definition of the conditional g-expectation:

\[ \mathcal{E}_g[\mathcal{E}_g[\xi | \mathcal{F}_s] | \mathcal{F}_t] = \mathcal{E}_g[Y_s | \mathcal{F}_t] = \tilde{Y}_t \]

Step 3: Show $\tilde{Y}_r = Y_r$ on $[0, s]$. Observe that the restriction of $(Y_r, Z_r)$ to $r \in [0, s]$ satisfies:

\[ Y_r = Y_s + \int_r^s g(u, Y_u, Z_u) \, du - \int_r^s Z_u \, dW_u \]

This is exactly a BSDE on $[0, s]$ with terminal condition $Y_s$ and generator $g$. By the uniqueness of BSDE solutions (Pardoux-Peng theorem), the solution to this BSDE is unique. Since both $(Y_r, Z_r)_{r \in [0,s]}$ and $(\tilde{Y}_r, \tilde{Z}_r)_{r \in [0,s]}$ solve the same BSDE on $[0, s]$ with the same terminal condition $Y_s$, we conclude:

\[ \tilde{Y}_r = Y_r, \quad \tilde{Z}_r = Z_r \quad \text{for all } r \in [0, s] \]

Step 4: Conclude. In particular, at $r = t$:

\[ \mathcal{E}_g[\mathcal{E}_g[\xi | \mathcal{F}_s] | \mathcal{F}_t] = \tilde{Y}_t = Y_t = \mathcal{E}_g[\xi | \mathcal{F}_t] \]

This completes the proof. $\square$

Why the tower property is essential for dynamic risk management. The tower property (also called time consistency or dynamic consistency) means that the risk assessment at time $t$ obtained by first evaluating risk at an intermediate time $s$ and then evaluating that assessment at $t$ is identical to directly evaluating risk at $t$. This is essential because:

Consistent decision-making: Without time consistency, a risk manager might assess a position as acceptable at time $t$, but after re-evaluating at time $s > t$, find it unacceptable (or vice versa). This leads to contradictory decisions and the possibility of dynamic arbitrage against the risk measure.
Bellman principle: Time consistency is the stochastic analog of Bellman's principle of optimality. It ensures that optimal hedging strategies computed at different times are mutually consistent, which is necessary for dynamic programming approaches.
Regulatory compliance: Regulators require that risk measures used for capital calculations be dynamically consistent --- otherwise, capital requirements could be manipulated by choosing the evaluation time strategically.
Multi-period hedging: In practice, hedging is done over multiple periods. Time consistency ensures that the hedging strategy derived from the g-expectation at each rebalancing time is consistent with the overall objective.

Exercise 5. Consider the quadratic generator $g(t, y, z) = -ry + \frac{\beta}{2}|z|^2$ arising from exponential utility. Solve the BSDE explicitly for a terminal condition $\xi = (S_T - K)^+$ under geometric Brownian motion and show that the g-expectation gives the indifference price of the option. How does $\beta$ affect the price relative to the Black-Scholes price?

Solution to Exercise 5

Goal. Analyze the BSDE with quadratic generator $g(t, y, z) = -ry + \frac{\beta}{2}|z|^2$ for terminal condition $\xi = (S_T - K)^+$ under geometric Brownian motion.

Step 1: Connection to exponential utility. The quadratic generator $g(t, y, z) = -ry + \frac{\beta}{2}|z|^2$ arises from exponential utility indifference pricing with risk aversion parameter $\beta > 0$ (here we use the convention that $\beta = \gamma$, the Arrow-Pratt coefficient of absolute risk aversion). The BSDE is:

\[ -dY_t = \left(-rY_t + \frac{\beta}{2}|Z_t|^2\right) dt - Z_t \, dW_t, \quad Y_T = (S_T - K)^+ \]

Step 2: Exponential transformation. This is a quadratic BSDE that can be solved by the Cole-Hopf transformation. Define:

\[ \hat{Y}_t = e^{-\beta Y_t} \]

By Ito's formula:

\[ d\hat{Y}_t = -\beta e^{-\beta Y_t} \, dY_t + \frac{\beta^2}{2} e^{-\beta Y_t} |Z_t|^2 \, dt \]

Substituting the BSDE dynamics:

\[ d\hat{Y}_t = -\beta e^{-\beta Y_t}\left[\left(rY_t - \frac{\beta}{2}|Z_t|^2\right)dt + Z_t \, dW_t\right] + \frac{\beta^2}{2} e^{-\beta Y_t}|Z_t|^2 \, dt \]

\[ = \beta r Y_t e^{-\beta Y_t} \, dt - \beta e^{-\beta Y_t} Z_t \, dW_t \]

The $|Z_t|^2$ terms cancel (this is the key simplification of the Cole-Hopf transform for quadratic BSDEs).

Step 3: Linear BSDE for $\hat{Y}_t$. Defining $\hat{Z}_t = -\beta e^{-\beta Y_t} Z_t$, we get:

\[ d\hat{Y}_t = \beta r Y_t \hat{Y}_t (-1/\beta \log \hat{Y}_t) \, dt + \hat{Z}_t \, dW_t \]

This simplifies (using $Y_t = -\frac{1}{\beta}\log \hat{Y}_t$) to:

\[ d\hat{Y}_t = -r \log(\hat{Y}_t) \hat{Y}_t \, dt + \hat{Z}_t \, dW_t \]

This is generally nonlinear in $\hat{Y}$. However, the key result is that the g-expectation gives the indifference price, which is the value $p$ satisfying:

\[ \sup_\pi E[-e^{-\beta(X_T^\pi)}] = \sup_\pi E[-e^{-\beta(X_T^\pi - p + (S_T - K)^+)}] \]

Step 4: Indifference price formula. For exponential utility, the indifference price has the well-known representation:

\[ p = -\frac{1}{\beta}\log E_{\mathbb{Q}^*}\left[e^{-\beta (S_T - K)^+}\right] + \frac{1}{\beta}\log E_{\mathbb{Q}^*}[1] \]

where $\mathbb{Q}^*$ is the minimal entropy martingale measure. Under geometric Brownian motion $dS_t = \mu S_t \, dt + \sigma S_t \, dW_t$ with the unique risk-neutral measure $\mathbb{Q}$ (complete market), this simplifies to:

\[ p = \mathcal{E}_g[(S_T - K)^+] = -\frac{1}{\beta}\log E_\mathbb{Q}\left[e^{-\beta(S_T - K)^+}\right] \]

Step 5: Effect of $\beta$ on the price. The indifference price $p$ relates to the Black-Scholes price $C_{\text{BS}} = E_\mathbb{Q}[(S_T - K)^+]$ as follows:

$\beta \to 0$ (risk-neutral): By L'Hopital's rule or Taylor expansion:

\[ p = -\frac{1}{\beta}\log E_\mathbb{Q}[e^{-\beta(S_T-K)^+}] \to E_\mathbb{Q}[(S_T-K)^+] = C_{\text{BS}} \]

$\beta > 0$ (risk-averse seller): By Jensen's inequality, $E[e^{-\beta X}] \geq e^{-\beta E[X]}$, so:

\[ p = -\frac{1}{\beta}\log E_\mathbb{Q}[e^{-\beta(S_T-K)^+}] \leq -\frac{1}{\beta}(-\beta E_\mathbb{Q}[(S_T-K)^+]) = C_{\text{BS}} \]

Wait --- this gives $p \leq C_{\text{BS}}$, which reflects the buyer's indifference price. For the seller's indifference price, one needs to account for the short position, and the price is:

\[ p^{\text{sell}} = \frac{1}{\beta}\log E_\mathbb{Q}[e^{\beta(S_T-K)^+}] \geq C_{\text{BS}} \]

So higher risk aversion $\beta$ increases the seller's price above the Black-Scholes price and decreases the buyer's price below it, creating a bid-ask spread: $[p^{\text{buy}}, p^{\text{sell}}] \supseteq [C_{\text{BS}}, C_{\text{BS}}]$, with the spread widening as $\beta$ increases.

$\beta \to \infty$ (infinite risk aversion): $p^{\text{sell}} \to \sup_\omega (S_T(\omega) - K)^+ = \infty$ (super-replication), reflecting extreme conservatism. $\square$

Exercise 6. A coherent risk measure $\rho(\xi) = \mathcal{E}_g[-\xi]$ can be defined via g-expectations. For the generator $g(z) = \theta|z|$ (entropic-type), compute $\rho(\xi)$ for a Gaussian random variable $\xi \sim N(\mu, \sigma^2)$ and show that $\rho(\xi) = -\mu + \theta\sigma$. Compare this with the CVaR risk measure and discuss the advantages of the g-expectation-based approach for dynamic risk measurement.

Solution to Exercise 6

Goal. For $g(z) = \theta|z|$ and $\xi \sim N(\mu, \sigma^2)$, compute $\rho(\xi) = \mathcal{E}_g[-\xi]$ and show $\rho(\xi) = -\mu + \theta\sigma$.

Step 1: Dual representation. The generator $g(z) = \theta|z|$ is convex and positively homogeneous in $z$. Therefore the g-expectation admits the dual representation:

\[ \mathcal{E}_g[X] = \sup_{\mathbb{Q} \in \mathcal{Q}_\theta} E_\mathbb{Q}[X] \]

where $\mathcal{Q}_\theta$ is the set of probability measures $\mathbb{Q}$ with Girsanov kernel $q_t$ satisfying $|q_t| \leq \theta$ a.s. These are the measures with bounded density process:

\[ \frac{d\mathbb{Q}}{dP}\bigg|_{\mathcal{F}_T} = \mathcal{E}\left(\int_0^T q_t \, dW_t\right), \quad |q_t| \leq \theta \]

Step 2: Compute $\rho(\xi)$. We have $\rho(\xi) = \mathcal{E}_g[-\xi]$. Using the dual representation:

\[ \rho(\xi) = \sup_{\mathbb{Q} \in \mathcal{Q}_\theta} E_\mathbb{Q}[-\xi] \]

Step 3: Gaussian case. Let $\xi = \mu + \sigma W_T / \sqrt{T}$ (so $\xi \sim N(\mu, \sigma^2)$ under $P$ when $W_T \sim N(0, T)$ with $T = 1$ for simplicity). Under $\mathbb{Q}$ with constant kernel $q_t = q$:

\[ W_T = \tilde{W}_T + qT \]

where $\tilde{W}_T$ is a $\mathbb{Q}$-Brownian motion. So:

\[ \xi = \mu + \sigma(\tilde{W}_1 + q) = (\mu + \sigma q) + \sigma \tilde{W}_1 \]

Under $\mathbb{Q}$, $\xi \sim N(\mu + \sigma q, \sigma^2)$. Therefore:

\[ E_\mathbb{Q}[-\xi] = -(\mu + \sigma q) \]

Step 4: Optimize. Maximizing over $|q| \leq \theta$:

\[ \rho(\xi) = \sup_{|q| \leq \theta} \{-\mu - \sigma q\} = -\mu + \sigma \sup_{|q| \leq \theta} \{-q\} = -\mu + \sigma \cdot \theta \]

The supremum is attained at $q = -\theta$ (since we maximize $-\sigma q$ with $\sigma > 0$, we choose $q$ as negative as possible). Therefore:

\[ \boxed{\rho(\xi) = -\mu + \theta\sigma} \]

Step 5: Comparison with CVaR. The Conditional Value-at-Risk (CVaR) at level $\alpha$ for $\xi \sim N(\mu, \sigma^2)$ is:

\[ \text{CVaR}_\alpha(\xi) = -\mu + \sigma \frac{\phi(\Phi^{-1}(\alpha))}{\alpha} \]

where $\phi$ is the standard normal density and $\Phi$ is the standard normal CDF. Setting $\theta_\alpha = \phi(\Phi^{-1}(\alpha))/\alpha$, we see that $\rho(\xi) = \text{CVaR}_\alpha(\xi)$ when $\theta = \theta_\alpha$.

For example, at $\alpha = 5\%$: $\theta_{0.05} = \phi(\Phi^{-1}(0.05))/0.05 = \phi(-1.645)/0.05 \approx 0.1031/0.05 \approx 2.063$.

So the g-expectation risk measure with $\theta \approx 2.063$ coincides with CVaR at the 5% level for Gaussian random variables.

Step 6: Advantages of g-expectation approach for dynamic risk measurement.

Dynamic consistency (tower property): The g-expectation-based risk measure $\rho_t(\xi) = \mathcal{E}_g[-\xi | \mathcal{F}_t]$ automatically satisfies the tower property $\rho_s(\rho_t(\xi)) = \rho_s(\xi)$ for $s \leq t$. CVaR, when naively extended to dynamic settings, generally violates time consistency, meaning that a position deemed acceptable today might be rejected tomorrow, or vice versa. This can lead to regulatory arbitrage.
Built-in coherence: The g-expectation risk measure inherits all coherence axioms (monotonicity, translation invariance, positive homogeneity, subadditivity) directly from the properties of the generator, without needing to verify them separately.
Constructive computation via BSDEs: The risk measure can be computed by solving a BSDE, which provides both the risk value and the hedging strategy (through $Z_t$). CVaR computation typically requires separate optimization and does not naturally produce hedging strategies.
Flexibility: Different generators $g$ produce different risk measures, allowing the framework to capture various attitudes toward risk and ambiguity in a unified manner. $\square$