# Decision-Dependent Stochastic Optimization: The Role of Distribution Dynamics

**Authors:** Zhiyu He, Saverio Bolognani, Florian Dörfler, Michael Muehlebach  
**arXiv:** [2503.07324](https://arxiv.org/abs/2503.07324)  
**Date:** 2025-03-10  
**Summary page:** https://www.summarizepaper.com/en/paper/2503.07324v1/

## Key points

- Authors: Zhiyu He, Saverio Bolognani, Florian Dörfler, Michael Muehlebach
- Decision-maker's choices can influence the environment and alter data-generating distribution
- Feedback loop formalized as endogenous distribution shift with nonlinear dynamics
- Online algorithm developed to adapt to and regulate dynamic distribution for optimal decision-making outcomes
- Helps characterize distribution dynamics and evaluate algorithm performance
- Opinion dynamics context: opportunistic party maximizes affinity in dynamic polarized population
- Recommender system scenario: performance optimization with discrete distributions in probability simplex
- Outperforms traditional methods by actively regulating distribution dynamics within complex problem structures
- Shows proposed algorithm has faster convergence rates, lower optimality gaps, and closer proximity to optimal solutions compared to vanilla algorithms

## Summary

In their paper "Decision-Dependent Stochastic Optimization: The Role of Distribution Dynamics," authors Zhiyu He, Saverio Bolognani, Florian Dörfler, and Michael Muehlebach examine the impact of distribution shifts on decision-making processes. They introduce the concept of decision dependence, where a decision-maker's choices can influence the environment and alter the data-generating distribution. This feedback loop is formalized as an endogenous distribution shift characterized by nonlinear dynamics that connect the evolving distribution with the decision. The authors explore stochastic optimization in this dynamic regime to understand how dynamics shape problem structures. They develop an online algorithm designed to adapt to and regulate the dynamic distribution, ultimately achieving optimal decision-making outcomes. By adopting a distributional perspective throughout their analysis, they demonstrate how this viewpoint aids in characterizing distribution dynamics and evaluating the performance of their proposed algorithm. The theoretical results are showcased through two scenarios: an opinion dynamics context where an opportunistic party maximizes affinity within a dynamic polarized population, and a recommender system scenario focusing on performance optimization with discrete distributions in the probability simplex. The authors highlight how their algorithm outperforms traditional methods by actively regulating distribution dynamics while considering complex problem structures. Additionally, Figure 3 illustrates how different algorithms interact with dynamics and solve specific problems, showcasing that the proposed algorithm exhibits faster convergence rates, lower optimality gaps, and closer proximity to optimal solutions compared to vanilla algorithms. This superior performance is attributed to its ability to adapt and regulate distribution dynamics within complex problem structures.