Introduction
Quantifying the similarity between probability distributions is a foundational requirement in control theory, essential for tasks such as steering a set of particles to a target shape or ensuring a robot's path covers an area according to a specific density. While metrics like Kullback–Leibler (KL) divergence or Wasserstein distance are popular, they often become computationally prohibitive or difficult to differentiate in high-dimensional spaces. This paper introduces a unified, differentiable family of distances for comparing high-dimensional probability distributions by averaging the discrepancies between their one-dimensional linear projections.
Methodology: Sliced CDF Distance
The core innovation is the Sliced CDF Distance. Instead of comparing high-dimensional distributions directly, the method projects them onto a random unit vector $q$ (a "slice"). In one dimension, comparing Cumulative Distribution Functions (CDFs) is mathematically simple and interpretable.
To make this distance useful for control, the authors replace the standard step-function CDF with a smooth sigmoid approximation. This allows for the calculation of gradients with respect to sample positions, enabling the use of standard gradient-based optimizers like Adam or SGD. This turns complex distribution matching problems into standard loss-minimization tasks.
Experimental Results
The paper evaluates the method using synthetic benchmarks designed to simulate real-world control challenges, including multivariate Gaussians and non-linear robotics models.
Key results from the evaluation include:
- Two-Sample Testing: The metric clearly distinguished between identical and non-identical distributions in high dimensions (up to 50D), matching or outperforming standard baselines.
- Distribution Steering: The algorithm successfully steered 3,000 particles following non-linear unicycle dynamics to match a target Gaussian Mixture Model (GMM).
- Ergodic Control: In a long-horizon task, a single agent's trajectory was optimized so that its time-averaged occupancy matched a spatial target distribution, with the loss decreasing consistently over time.
Conclusion
The Sliced CDF Distance offers a computationally efficient, interpretable, and gradient-friendly framework for distribution matching. By providing theoretical guarantees and a differentiable implementation, it enables the application of simple optimization techniques to complex high-dimensional control problems.
Summary updated