Posterior and Likelihood Sensitivity in Bayesian Distributionally Robust Optimization

Technical Summary: Posterior and Likelihood Sensitivity in Bayesian Distributionally Robust Optimization

Problem Statement
Bayesian models provide a framework for learning from data by combining a likelihood function $L_\theta(\cdot)$ and a prior distribution $\rho$ to form a predictive distribution. However, decision-makers (DMs) often face concerns regarding model misspecification. Specifically, the expected cost of a decision may be highly sensitive to deviations from the assumed predictive distribution. While existing literature addresses robustness in empirical optimization, there is a need to quantify the sensitivity of Bayesian decisions specifically to deviations in two distinct components: the posterior distribution (representing parameter uncertainty) and the likelihood function (representing the structural model of the data generation process). The paper addresses the problem of characterizing and managing the trade-off between performance (expected cost) and robustness against these specific sources of uncertainty.

Methodology
The authors introduce the concepts of worst-case posterior sensitivity and worst-case likelihood sensitivity. These are defined within a Bayesian Distributionally Robust Optimization (DRO) framework.

1. Uncertainty Sets: The methodology defines uncertainty sets for the posterior $\eta$ and the likelihood $Q_\theta$ using deviation measures $d(\cdot|\cdot)$.

   • $P(\epsilon) = \{ \eta \mid d(\eta|\rho) \le \epsilon \}$ represents deviations from the nominal posterior.
   • $L_\theta(\delta) = \{ Q \mid d(Q|L_\theta) \le \delta \}$ represents deviations from the nominal likelihood for each parameter $\theta$.

2. Worst-Case Expected Cost: The authors formulate the worst-case expected cost $V(\epsilon, \delta)$ as a nested maximization problem:
   $$V(\epsilon, \delta) = \max_{\eta \in P(\epsilon)} \mathbb{E}_\eta \left[ \max_{Q_\theta \in L_\theta(\delta)} \mathbb{E}_{Q_\theta}[f(\theta, Y)] \right]$$
   Here, the inner maximization accounts for worst-case likelihood deviations conditional on $\theta$, and the outer maximization accounts for worst-case posterior deviations over the parameter space.

3. Regularization and Sensitivity Expansion: Drawing on techniques from robust empirical optimization (specifically referencing works [8, 9, 10]), the paper derives a local expansion of the worst-case cost for small uncertainty set sizes ($\epsilon, \delta$). Under Assumption 4.1 (regarding the behavior of ambiguity costs), the worst-case cost is shown to be approximately the nominal expected cost plus a weighted sum of sensitivity terms:
   $$V(\epsilon, \delta) \approx \mathbb{E}_\rho[\mathbb{E}_{L_\theta}[f]] + g(\delta)\mathbb{E}_\rho[S_{L_\theta}(f)] + g(\epsilon)S_\rho(\mathbb{E}_{L_\theta}[f])$$
   where $g(\cdot)$ is a function determined by the deviation measure, $S_{L_\theta}$ is the likelihood sensitivity, and $S_\rho$ is the posterior sensitivity.

4. Specific Deviation Measures: The paper explicitly derives these sensitivities for smooth $\phi$-divergences. In this case, the sensitivity measures correspond to standard deviations:

   • Posterior Sensitivity: Proportional to the standard deviation of the conditional expected cost under the posterior, $\sigma_\rho(\mathbb{E}_{L_\theta}[f])$.
   • Likelihood Sensitivity: Proportional to the posterior expectation of the conditional standard deviation of the cost, $\mathbb{E}_\rho[\sigma_{L_\theta}(f)]$.
     The framework also allows for "mix and match" scenarios where different deviation measures (e.g., $\phi$-divergence for the posterior and CVaR-based measures for the likelihood) are used.

Key Contributions

• Novel Sensitivity Metrics: The paper introduces distinct quantitative measures for the sensitivity of Bayesian decisions to posterior misspecification versus likelihood misspecification.
• Decomposition of Robustness: It demonstrates that a Bayesian DRO problem can be approximated as a regularized nominal problem where the regularizer is a weighted sum of posterior and likelihood sensitivities. This establishes a direct link between distributional robustness and sensitivity analysis.
• Asymptotic Behavior: A critical theoretical finding is the differential behavior of the two sensitivities as learning occurs. The paper proves that posterior sensitivity vanishes when the posterior distribution degenerates to a point mass (i.e., when parameter uncertainty is eliminated through learning). In contrast, likelihood sensitivity does not vanish even when parameters are known, as it reflects the inherent uncertainty in the data distribution $Y$ given $\theta$.
• Pareto Trade-offs: The authors show that solving a Bayesian DRO problem traces a nearly Pareto-optimal frontier between performance (expected cost) and robustness (posterior and likelihood sensitivity).

Results
The paper validates these theoretical findings through a pricing application where demand is normally distributed with unknown parameters $a$ and $b$.

• Learning Effects: Numerical experiments show that as the variance of the posterior decreases (simulating data accumulation), the posterior sensitivity decreases and eventually vanishes. However, the likelihood sensitivity remains constant, confirming that learning reduces parameter uncertainty but not model structural uncertainty.
• Frontier Analysis: The reward-sensitivity frontiers illustrate that DRO allows a decision-maker to reduce sensitivity significantly with minimal cost to expected reward, particularly when initial parameter uncertainty is high.
• Independence: In the specific pricing example, the reward-likelihood sensitivity frontier was found to be independent of the posterior variance, while the reward-posterior sensitivity frontier was independent of the likelihood variance (a property specific to the linear-quadratic structure of the example).

Significance and Claims
The paper claims to provide a systematic approach for decision-makers concerned with the fragility of their assumptions. By distinguishing between posterior and likelihood sensitivity, the framework allows DMs to identify which component of their model (parameter estimation vs. structural form) is driving robustness concerns.

• Robustness vs. Performance: The primary significance lies in formalizing the "price of robustness" as a trade-off between expected cost and sensitivity.
• Guidance for System Design: The authors suggest that these sensitivity measures can guide the selection of uncertainty sets and identify when the cost of robustness is high, potentially prompting system redesign.
• Limitations of Learning: A key insight is that while data learning eliminates posterior sensitivity, it does not eliminate likelihood sensitivity. Therefore, a robust Bayesian decision must account for the likelihood model's potential misspecification even after parameters are well-estimated.

The paper concludes that a distributionally robust formulation offers a near-Pareto-optimal mechanism to balance performance and robustness against both types of model uncertainty.