Wasserstein Distributionally Robust Risk-Sensitive Estimation via Conditional Value-at-Risk

1. Problem Formulation

The paper addresses the problem of estimating an unknown latent signal vector $\mathbf{x} \in \mathbb{R}^n$ from an observed signal vector $\mathbf{y} \in \mathbb{R}^m$. The joint distribution $P$ of the random vector $\mathbf{z} = (\mathbf{x}, \mathbf{y})$ is unknown but is assumed to belong to an ambiguity set $\mathcal{P}$.

• Ambiguity Set: The set $\mathcal{P}$ is defined as a Type-2 Wasserstein ball of radius $r$ centered at a nominal distribution $P_0$ (typically estimated from data):
  $$\mathcal{P} := \{ P \in \mathcal{M}(\mathbb{R}^d) \mid W_2(P, P_0) \leq r \}$$
  where $W_2$ is the Wasserstein distance and $d = n+m$.
• Estimator Class: The authors restrict the search space to affine estimators of the form $\psi(\mathbf{y}) = A\mathbf{y} + \mathbf{b}$.
• Objective: Unlike traditional methods that minimize the Mean Squared Error (MSE), this work minimizes the Conditional Value-at-Risk (CVaR) of the squared estimation error. CVaR at level $\alpha \in (0, 1]$ measures the expected loss in the worst $\alpha$-fraction of cases, making it a risk-sensitive metric.
• Optimization Goal: Find the affine estimator that minimizes the worst-case CVaR over all distributions in the ambiguity set:
  $$\inf_{\psi \in \mathcal{A}} \sup_{P \in \mathcal{P}} \text{CVaR}_{\alpha}^{P} \left( \|\mathbf{x} - \psi(\mathbf{y})\|^2 \right)$$

2. Methodology

The core methodological contribution is the derivation of a tractable convex reformulation for this min-max problem using duality theory in optimal transport.

A. Strong Duality for Worst-Case CVaR

The authors leverage a strong duality result for Wasserstein distributionally robust optimization. They establish that the worst-case CVaR problem can be transformed into a dual minimization problem involving a scalar variable $\tau$ and a Lagrange multiplier $\gamma$.

• Theorem 1: Provides a dual representation for the worst-case CVaR of a generic loss function $\ell(\mathbf{z})$.
• Theorem 2 (Key Technical Result): Specializes the duality result to quadratic loss functions (specifically the squared error). It proves that the worst-case CVaR of a quadratic loss is equivalent to a convex optimization problem involving the nominal distribution $P_0$ and a parameter $\gamma$.
  • Significance: Unlike previous results that required the nominal distribution to be elliptical or absolutely continuous, this theorem holds for any nominal distribution in $\mathcal{M}(\mathbb{R}^d)$.

B. Semidefinite Programming (SDP) Reformulation

When the nominal distribution $P_0$ is finitely supported (e.g., an empirical distribution derived from a finite dataset), the authors show that the problem can be exactly reformulated as a Semidefinite Program (SDP).

• Theorem 3: If $P_0$ is uniform over a set of $N$ points $\{z_1, \dots, z_N\}$, the original min-max problem is equivalent to the following SDP:
  • Variables: Matrix $A$, vector $\mathbf{b}$, scalars $\gamma, \tau$, and auxiliary variables $s_i$.
  • Constraints: Linear Matrix Inequalities (LMIs) derived from the Schur complement condition to handle the quadratic terms and the CVaR definition.
  • Objective: Minimize $\tau + \frac{1}{\alpha}(\gamma r^2 + \frac{1}{N}\sum s_i)$.

This reformulation allows the use of standard convex optimization solvers (e.g., MOSEK, CVX) to compute the optimal risk-sensitive estimator efficiently.

3. Key Contributions

1. Novel Framework: First study of Wasserstein distributionally robust estimation using a CVaR of squared error criterion. Previous works focused on worst-case MSE (risk-neutral) or used different ambiguity sets (e.g., Gelbrich distance).
2. General Duality Result: Theorem 2 establishes a dual reformulation for quadratic worst-case CVaR problems under Type-2 Wasserstein ambiguity without restrictive assumptions on the nominal distribution (unlike prior work requiring elliptical distributions).
3. Exact Tractability: Theorem 3 provides an exact SDP reformulation for finitely supported nominal distributions, making the problem computationally solvable for real-world datasets.
4. Risk-Sensitivity vs. Robustness: The framework unifies risk sensitivity (via CVaR) and distributional robustness (via Wasserstein ball), offering a tunable approach to guard against both model misspecification and tail events.

4. Experimental Results

The authors evaluated their proposed DR-CVaR estimator on a wholesale electricity price forecasting task using PJM market data (May–July 2013).

• Setup:
  • Input: Hourly day-ahead load forecasts ($\mathbf{y}$).
  • Target: Hourly day-ahead energy prices ($\mathbf{x}$).
  • Test Case: The test set included price spikes significantly higher than those in the training set, creating a challenging out-of-sample scenario for tail-risk performance.
• Comparisons:
  • DR-CVaR: Proposed method with risk level $\alpha = 0.01$.
  • DR-MSE: Distributionally robust estimator minimizing worst-case MSE (equivalent to $\alpha = 1$).
• Findings:
  • Tail Risk Reduction: The DR-CVaR estimator consistently achieved a lower out-of-sample CVaR of squared error compared to the DR-MSE estimator across various ambiguity set radii ($r$).
  • Trade-off: While the DR-CVaR estimator was slightly less accurate on days with typical price patterns (due to the risk-averse trade-off), it significantly outperformed the MSE-based estimator during days with extreme price spikes.
  • Robustness: Both robust estimators outperformed their non-robust counterparts as the ambiguity radius $r$ increased, but the risk-sensitive version provided superior protection against the worst-case tail events.

5. Significance and Impact

• Practical Application: The paper demonstrates that explicitly modeling tail risk via CVaR within a distributionally robust framework is crucial for applications where rare, severe errors have catastrophic consequences (e.g., power grid reliability, financial trading).
• Theoretical Advancement: It bridges the gap between optimal transport theory and risk-sensitive control, providing a rigorous mathematical tool to handle uncertainty in both the data distribution and the extreme events within that distribution.
• Computational Feasibility: By reducing a complex min-max problem to a tractable SDP, the authors make advanced risk-sensitive estimation accessible for practical engineering problems involving finite datasets.

In summary, this paper proposes a mathematically rigorous and computationally efficient method for designing estimators that are robust to distributional uncertainty and specifically optimized to minimize the impact of rare, high-magnitude errors.