In many domains—from radar signal processing to 3D computer vision—data does not live in a flat, Euclidean space. Instead, it resides on curved surfaces known as Riemannian manifolds. For instance, the covariance matrices used in radar live on the manifold of Hermitian Positive-Definite (HPD) matrices, while object orientations live on rotation manifolds.
When we need to find the "average" of such data, the standard arithmetic mean fails because it ignores the underlying geometry. While the Fréchet mean (the manifold equivalent of an average) respects this geometry, it inherits a fatal flaw from its Euclidean cousin: it is hypersensitive to outliers. A single "spiky" interference signal or a misaligned 3D scan can pull the mean far away from the true center of the data.
To solve this, researchers have turned to the Fréchet median (or geometric median), a robust estimator that identifies the central tendency of data on a manifold while remaining remarkably resilient to noise and corruption.
The Geometry of Robustness
The intuition behind the Fréchet median lies in how we define "distance" to a center. In Euclidean space, the arithmetic mean minimizes the sum of squared distances to all points. Because the distances are squared, a single outlier far from the cluster exerts a massive "pull" on the result.
The Fréchet median, by contrast, minimizes the sum of the distances themselves (the L1-norm). On a manifold, this means minimizing the sum of geodesic distances—the shortest paths along the curved surface. As Frederic Barbaresco and Marc Arnaudon demonstrate in Medians and Means in Riemannian Geometry: Existence, Uniqueness and Computation, this shift from squared distances to linear distances grants the estimator a high breakdown point. In practice, this means the median remains stable even if nearly 50% of the data is corrupted by outliers.
Filtering Radar Clutter and Interference
This robustness is a game-changer for signal detection. In radar systems, a target signal is often buried in "clutter"—unwanted echoes from waves or rain—and intermittent interference. Traditional detectors that use arithmetic averaging to estimate background noise are easily fooled by these "spiky" outliers, leading to false alarms or missed targets.
By treating covariance matrices as points on an HPD manifold, Xiaoqiang Hua shows in MIG median detectors with manifold filter that the geometric median can effectively "ignore" these spikes. Because the median focuses on the most representative state of the signal, it prevents leakage from high-power interferers into the background estimate. Similarly, Yusuke Ono argues in Towards a median signal detector through the total Bregman divergence and its robustness analysis that this approach allows detectors to maintain accuracy even in non-stationary environments where noise levels fluctuate rapidly.
Robustness in Vision and Symmetry
The need for geometric medians extends to 3D scene understanding and object orientation. When aligning 3D models or discovering symmetries, algorithms must deal with multimodal distributions—cases where there might be several "plausible" orientations, but only one is correct.
In RECON: Robust symmetry discovery via Explicit Canonical Orientation Normalization, Alonso Urbano and colleagues use a Tukey-Fréchet mean to identify a canonical orientation that is resilient to outliers. This ensures that a model's "reference frame" isn't skewed by a few poorly aligned samples.
In the realm of 3D Gaussian Splatting, Xinpeng Zhao introduces a Huberized geodesic distance in Rh-3DGS. This acts as a hybrid: it behaves like a mean for small errors (to maintain precision) but acts like a median for large errors (to provide robustness). This allows the model to distill clean semantic information even when the input multi-view data is noisy or inconsistent.
Computing the Median on Complex Spaces
Calculating a median on a curved surface is computationally harder than a simple average. It usually requires iterative algorithms like Iteratively Reweighted Least Squares (IRLS) or stochastic gradient descent on the manifold.
Researchers are expanding these techniques to increasingly complex mathematical spaces:
- Flag Manifolds: Nathan Mankovich and Tolga Birdal developed algorithms for Chordal Averaging on Flag Manifolds, which represent nested sequences of subspaces. Their work shows that median-based statistics on these manifolds can handle cases where nearly half the dataset is pure noise.
- Essential Manifolds: In computer vision, Roberto Tron and Kostas Daniilidis have applied robust L1-norm estimation to the Essential Manifold, enabling more stable 3D reconstructions from noisy camera poses.
By moving from "flat" arithmetic to "curved" robust statistics, these methods allow machines to find the signal within the noise, respecting the intrinsic structure of the data while ignoring the outliers that would otherwise distort our understanding of it.