Optimization Under Performative Distribution Shifts

444 concepts213 papers2 proposals

In classical machine learning, we treat the world as a static observer. We collect a dataset, train a model, and deploy it under the assumption that the underlying data distribution remains unchanged. However, when machine learning models are deployed in social, economic, or clinical systems, the predictions themselves often trigger behavioral changes in the subjects being measured.

Consider a bank deploying a credit scoring model. If the model determines that having fewer open credit card accounts increases creditworthiness, individuals will strategically close accounts to improve their scores. The data distribution has shifted, not due to an external environmental change, but as a direct consequence of the model's deployment.

This phenomenon is known as performative prediction. When a model's deployment alters the very data it aims to predict, it creates a feedback loop that breaks the foundational independent and identically distributed (i.i.d.) assumption of statistical learning theory.

The Distribution Map

To mathematically formalize this feedback loop, Perdomo et al. introduced the concept of a distribution map $\mathcal{D}(\theta)$ in their foundational paper Performative Prediction. Instead of drawing data from a fixed distribution $\mathcal{D}$, the data distribution becomes an explicit function of the deployed model parameters $\theta$.

When a model $\theta$ is deployed, the risk we observe is not the standard risk over a static dataset, but the performative risk:

$$\text{PR}(\theta) = \mathbb{E}_{z \sim \mathcal{D}(\theta)} [\ell(\theta; z)]$$

where $\ell(\theta; z)$ is the loss function evaluated on a data point $z$ sampled from the model-induced distribution $\mathcal{D}(\theta)$. This formulation reveals a fundamental tension in reactive environments: should we optimize our model to perform well on the current distribution, or should we optimize it to guide the environment toward a new, more favorable distribution?

Performative Stability vs. Performative Optimality

The most intuitive way to handle a shifting distribution is to react to it. If a model is deployed and the data shifts, a practitioner will typically collect new data and retrain the model. This iterative process is called Repeated Risk Minimization (RRM).

At each step $t$, RRM solves a decoupled optimization problem, finding a model that minimizes the loss on the distribution induced by the previous model $\theta_t$:

$$\theta_{t+1} = \arg\min_{\theta} \mathbb{E}_{z \sim \mathcal{D}(\theta_t)} [\ell(\theta; z)]$$

If this process converges, it reaches a performatively stable point $\theta^*_{\text{stable}}$, where the model is optimal for the distribution it creates:

$$\theta^*_{\text{stable}} = \arg\min_{\theta} \mathbb{E}_{z \sim \mathcal{D}(\theta^*_{\text{stable}})} [\ell(\theta; z)]$$

Mendler-Dünner et al. provided the first non-asymptotic convergence guarantees for stochastic gradient descent (SGD) under RRM in Stochastic Optimization for Performative Prediction, proving that simple retraining strategies can reliably find stable fixed points if the distribution shift is sufficiently bounded (i.e., if the distribution map is Lipschitz continuous with a small constant).

However, as Miller et al. argue in Outside the Echo Chamber: Optimizing the Performative Risk, performative stability is fundamentally myopic. A stable point is merely a fixed point of a reactive loop; it does not necessarily minimize the true performative risk $\text{PR}(\theta)$. The true performative optimum $\theta^*_{\text{opt}}$ is the global minimizer of $\text{PR}(\theta)$:

$$\theta^*_{\text{opt}} = \arg\min_{\theta} \mathbb{E}_{z \sim \mathcal{D}(\theta)} [\ell(\theta; z)]$$

The gap between stability and optimality arises because RRM ignores the causal influence that changing $\theta$ has on the distribution $\mathcal{D}(\theta)$.

Repeated Risk Minimization (Myopic / Reactive):
[ Deploy θ_t ] ---> [ Population Shifts to D(θ_t) ] ---> [ Retrain θ_t+1 on D(θ_t) ]

Performative Gradient Descent (Strategic / Proactive):
[ Evaluate θ_t ] -> [ Anticipate how D(θ) changes ] -> [ Update to θ_t+1 via Total Gradient ]

To find the true performative optimum, we must compute the total gradient of the performative risk with respect to $\theta$. By applying the chain rule, the total gradient splits into two distinct components:

$$\nabla \text{PR}(\theta) = \underbrace{\mathbb{E}_{z \sim \mathcal{D}(\theta)} [\nabla_{\theta} \ell(\theta; z)]}_{\text{Idealized Gradient}} + \underbrace{\nabla_{\theta} \left( \mathbb{E}_{z \sim \mathcal{D}(\phi)} [\ell(\theta; z)] \right) \Big|_{\phi=\theta}}_{\text{Performative Gradient}}$$

  1. The Idealized Gradient: This measures how the loss changes as the model parameters change, keeping the data distribution fixed. This is what standard machine learning algorithms compute.
  2. The Performative Gradient: This measures how the loss changes because the data distribution itself shifts in response to the model parameters.

Computing the performative gradient is exceptionally difficult because the distribution map $\mathcal{D}(\theta)$ represents complex human behavior and is generally a black box. If we ignore the performative gradient, we are blind to how our decisions steer the environment.

To overcome this, Izzo et al. developed Performative Gradient Descent (PerfGD) in How to Learn when Data Reacts to Your Model: Performative Gradient Descent. PerfGD explicitly estimates the derivative of the data distribution by perturbing the deployed models and observing the resulting shifts in data. By incorporating this derivative, PerfGD actively steers the system toward the performative optimum rather than settling for a suboptimal stable point.

When explicit distribution derivatives are entirely unavailable, derivative-free optimization can be deployed. For example, Hikima introduces Guided Zeroth-Order methods in Guided Zeroth-Order Methods for Stochastic Non-convex Problems with Decision-Dependent Distributions, which leverages known structural components of the loss function to drastically reduce the sample complexity required to estimate performative shifts.

Memory and Gradual Adaptation

Early formulations of performative prediction assumed that the population adapts instantaneously to any newly deployed model. In practice, human behavior adapts gradually over time due to habit, information friction, or physical constraints.

When the population has memory, the data distribution at time $t$ depends on both the newly deployed model $\theta_t$ and the historical state of the population. Zou et al. addressed this stateful challenge in How to Learn when Data Gradually Reacts to Your Model by introducing Stateful Performative Gradient Descent (Stateful PerfGD). This framework allows an algorithm to "simulate" and optimize for the long-term steady state of a population without needing to wait periods of time for the population to fully equilibrate.

Complementing this, Li et al. analyzed state-dependent performative prediction through the lens of stochastic approximation in State Dependent Performative Prediction with Stochastic Approximation. They proved that even when data samples are heavily biased by historical population states, carefully tuned step sizes allow stochastic gradient updates to track and converge to stable solutions.

Multi-Agent Competition and Social Dynamics

The scope of performative prediction naturally expands from a single centralized model to decentralized ecosystems where multiple models interact with the same population. For instance, competing ride-sharing platforms or banks deploy distinct models that simultaneously influence a shared pool of users.

Narang et al. formalized this multi-agent setting in Learning in Stochastic Monotone Games with Decision-Dependent Data and Multiplayer Performative Prediction: Learning in Decision-Dependent Games. In these decision-dependent games, the global data distribution is modeled as a joint function of all players' parameters: $\mathcal{D}(\theta_1, \dots, \theta_N)$. The authors demonstrate that agents who actively model this joint feedback loop avoid the severe revenue and performance losses suffered by myopic agents who treat data as static.

Furthermore, these feedback loops have profound societal implications, particularly regarding algorithmic fairness. As Zhang et al. demonstrate in Achieving Long-Term Fairness in Sequential Decision Making, ignoring performative distribution shifts can cause a model that appears fair during initial training to inadvertently alter group participation rates over time, compounding historical disparities and degrading long-term fairness guarantees.

By shifting the optimization paradigm from passive observation to active, causal anticipation, the performative prediction framework provides the mathematical tools necessary to design stable, optimal, and socially responsible models in an inherently reactive world.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 242 papers
  2. Version 1 · 2026-10-04 · 326 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_performative_distribution_shift_79ef5125,
  title = {performative distribution shift},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/performative-distribution-shift/rdr_79ef5125ca914704ad5dad8bd6e953e2},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 2, frozen text):

@misc{lacuna_performative_distribution_shift_79ef5125_v2,
  title = {performative distribution shift},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/performative-distribution-shift/rdr_79ef5125ca914704ad5dad8bd6e953e2/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake