Abstract
The ratio of two probability density functions is a fundamental quantity that appears in many areas of statistics and machine learning, including causal inference, reinforcement learning, covariate shift, outlier detection, independence testing, importance sampling, and diffusion modeling. Naively estimating the numerator and denominator densities separately using, e.g., kernel density estimators, can lead to unstable performance and suffers from the curse of dimensionality as the number of covariates increases. For this reason, several methods have been developed for estimating the density ratio directly based on (a) Bregman divergences or (b) recasting the density ratio as the odds in a probabilistic classification model that predicts whether an observation is sampled from the numerator or denominator distribution. Additionally, the density ratio can be viewed as the Riesz representer of a continuous linear map, making it amenable to estimation via (c) minimization of the so-called Riesz loss, which was developed to learn the Riesz representer in the Riesz regression procedure in causal inference. In this paper we show that all three of these methods can be unified in a common framework, which we call Bregman-Riesz regression. We further show how data augmentation techniques can be used to apply density ratio learning methods to causal problems, where the numerator distribution typically represents an unobserved intervention. We show through simulations how the choice of Bregman divergence and data augmentation strategy can affect the performance of the resulting density ratio learner. A Python package is provided for researchers to apply Bregman-Riesz regression in practice using gradient boosting, neural networks, and kernel methods.
Introduction
In causal inference and machine learning, estimating the ratio of two probability density functions is a critical task. For instance, in inverse probability weighting (IPW), this ratio serves as an importance weight to translate expectations from observed data to a target intervention distribution. However, naive estimation—calculating the numerator and denominator separately and then dividing them—is notoriously unstable, especially in high-dimensional settings.
While specialized methods like Bregman divergence minimization or Riesz regression exist, they have historically been treated as fragmented literatures. Furthermore, causal inference presents a unique challenge: the numerator usually represents a counterfactual intervention (e.g., "what if everyone received the treatment?"). Because we cannot directly sample from this unobserved alternate reality, standard direct density ratio learners cannot be applied out-of-the-box.
The Bregman-Riesz Regression Framework
This paper introduces Bregman-Riesz regression, a mathematical framework that unifies Riesz regression, Bregman divergence minimization (such as KLIEP and uLSIF), and classification-based odds estimation. The authors prove that these disparate methods are mathematically equivalent to minimizing an empirical "Bregman-Riesz Risk."
Central to this framework is the choice of the Bregman divergence, which determines how errors are weighted across the range of density ratios. The framework supports several common divergences, as detailed in the table below:
A key insight of this unification is that standard probabilistic classification (e.g., using a logistic regression classifier to estimate odds) is mathematically equivalent to minimizing a specific Bregman divergence, termed the "negative binomial" divergence.
Different divergences offer different robustness properties. For example, when distributions have poor overlap, true density ratios can become extremely large. Standard least-squares loss applies a constant weight across all values, which often leads to overfitting on these extreme outliers. In contrast, divergences like Negative-Binomial and Itakura–Saito naturally down-weight errors on large density ratios, making them significantly more robust to overlap violations.
Data Augmentation for Causal Estimands
To bridge the gap between standard density ratio learning and causal inference, the authors propose estimand-specific data augmentation strategies to handle the missing "numerator" data:
- Modified Treatment Policies / Shift Effects: The hypothetical policy is applied directly to observed covariates to generate synthetic samples for the numerator distribution.
- Stabilized Weights (for Dose-Response): The dependence between treatment and covariates is broken using permutations, derangements, or resampling to simulate a randomized treatment assignment.
The study found that "dynamic augmentation"—where the augmented data is resampled via permutations at each training epoch or batch—yielded the most stable performance compared to using fixed, pre-computed datasets.
Experimental Results
The framework was evaluated using synthetic datasets with 20-dimensional covariates to test various causal estimands. The results showed that Bregman-Riesz regression methods, particularly when implemented with Multilayer Perceptrons (MLPs) and Gradient Boosted Machines (GBMs), consistently outperformed traditional propensity score-based methods.
The choice of divergence proved critical: in scenarios prone to large density ratios, the Negative-Binomial and Itakura–Saito divergences dramatically outperformed Least-Squares. For example, an MLP trained with the Itakura–Saito divergence achieved the lowest absolute bias (0.034) for the average policy effect.
Conclusion and Future Work
Bregman-Riesz regression provides a powerful, unified approach to density ratio estimation in causal contexts. By combining flexible machine learning models with robust loss functions and tailored data augmentation, it overcomes many limitations of traditional separate density estimation.
One current limitation is handling Riesz representers that are not strictly positive, such as the Average Treatment Effect (ATE), which involves a signed difference between density ratios. Future work will focus on extending the framework to handle these cases and exploring methods to learn the optimal divergence directly from data.
Summary updated