Robust Estimation of Location in Matrix Manifolds Using the Projected Frobenius Median

This paper proposes a computationally efficient and robust method for estimating the location of data on various matrix manifolds by computing the Frobenius median in an ambient Euclidean space and projecting it onto the manifold, while establishing its theoretical properties and demonstrating its effectiveness through simulations and real-world earthquake data.

The Gist

Imagine you are trying to find the "center" of a group of things, but these things aren't just numbers on a spreadsheet. They are complex shapes, directions, or 3D structures that live on a curved surface, like the surface of a sphere or a twisted ribbon. In statistics, this is called finding the location on a manifold.

The problem? Real-world data is messy. Sometimes, sensors glitch, or someone enters a wrong number. These "outliers" can completely ruin the average. If you have a group of people standing in a circle and one person runs off to the moon, the average position of the group suddenly jumps halfway to the moon. That's not a good description of where the group actually is.

This paper introduces a new, super-tough way to find the center of these complex shapes, even when the data is full of noise and mistakes. They call it the Projected Frobenius Median (PFM).

Here is the simple breakdown of how it works, using some everyday analogies.

1. The Problem: The "Curved" World

Most of us are used to doing math on flat surfaces (like a piece of paper). If you want the average of points on a flat sheet, you just add them up and divide by the count. Easy.

But many real-world data points (like the orientation of a robot arm, the shape of a protein, or the direction of an earthquake's fault) live on curved surfaces (manifolds).

• The Old Way: To find the center on a curve, statisticians usually try to walk along the curve, measuring the distance to every point, and find the spot that minimizes the total walking distance.
• The Trouble: This is like trying to find the center of a mountain range by hiking every possible path. It's slow, it's hard to calculate, and if you start in the wrong spot, you might get stuck in a local valley (a "local minimum") and think you found the bottom when you didn't. Plus, if one hiker goes off a cliff (an outlier), the whole calculation gets skewed.

2. The Solution: The "Shadow" Trick (The PFM)

The authors propose a clever shortcut. Instead of walking on the curved surface, they do the heavy lifting in the "flat" world surrounding it, and then just drop a shadow back onto the curve.

Think of it like this:

• The Curved Surface: Imagine a shiny, curved mirror (the manifold) floating in a large, empty room (the flat Euclidean space).
• The Data: You have a bunch of glowing marbles sitting on the mirror.
• The Outliers: Some of those marbles are actually fake, glowing red lights placed far away from the group.

The PFM Method:

1. Step 1: The Flat Shadow. Instead of trying to find the center on the mirror, the authors pretend the mirror isn't there. They look at the glowing marbles in the big, flat room. They calculate the Spatial Median (a robust "center" that ignores extreme outliers) in this flat room.
   • Analogy: Imagine the marbles are casting shadows on the floor. You find the center of the shadows on the flat floor. Because the floor is flat, you can use standard, fast, and reliable computer tools to do this. The "median" on the floor naturally ignores the fake red lights because they are too far away to pull the center too much.
2. Step 2: The Projection. Once they have the center point on the flat floor, they simply "project" it straight up (or down) onto the curved mirror.
   • Analogy: You shine a light from the center of the floor straight up to the mirror. Where the light hits the mirror is your new, robust center.

3. Why is this better?

• It's Fast: Calculating the center on a flat floor is easy. Calculating it on a twisted, curved mirror is a nightmare. This method does the hard math on the easy surface.
• It's Robust: Because they use the "median" (the middle value) instead of the "mean" (the average) in the flat room, a few crazy outliers don't drag the center away. The "middle" stays put.
• It's Unique: Sometimes, on curved surfaces, there are multiple "centers" (like the North Pole and South Pole are both centers of the equator). This method almost always gives you one single, clear answer.
• It's Flexible: It works on many different types of shapes, from simple spheres to complex 3D orientations used in earthquake science.

4. Real-World Test: Earthquakes

The authors tested this on real earthquake data. Earthquakes have "moment tensors," which are 3x3 matrices describing how the ground moved.

• The Scenario: They looked at earthquake data from Papua New Guinea. Some of the data points were weird (outliers) due to measurement errors or unusual geological events.
• The Result: The old methods (like the standard average) got pulled toward the weird data, giving a distorted picture of the earthquake's direction. The new Projected Frobenius Median ignored the weird data and correctly identified the true direction of the fault lines, even when up to 40% of the data was "bad."

Summary

Imagine you are trying to find the center of a flock of birds flying in a complex 3D formation, but a few drones are flying wildly off-course.

• Old Method: Try to calculate the center by measuring the distance between every bird and every other bird while they are flying. It's confusing, slow, and the rogue drones mess it up.
• New Method (PFM): Pretend the birds are flying in a flat 2D parking lot. Find the center of the flock there (ignoring the rogue drones). Then, map that center back up to the 3D sky.

The paper proves that this "flat-then-curved" approach is mathematically sound, incredibly fast, and much harder to fool by bad data than previous methods. It's a new, robust tool for statisticians working with complex shapes and directions.

Technical Summary

Here is a detailed technical summary of the paper "Robust Estimation of Location in Matrix Manifolds Using the Projected Frobenius Median."

1. Problem Statement

The paper addresses the challenge of estimating the location (central tendency) of data lying on matrix manifolds (curved spaces of matrices) in the presence of outliers.

• Context: Matrix-valued data arises in diverse fields such as computer vision (Stiefel/Grassmann manifolds), statistical shape analysis (Kendall shape spaces), and geophysics (seismic moment tensors).
• Limitations of Existing Methods:
  • Intrinsic Estimators (e.g., Fréchet Median/Geometric Median): While conceptually appealing, minimizing the sum of intrinsic distances is computationally expensive, often lacks a closed-form solution, suffers from non-uniqueness, and is prone to converging to local minima.
  • Mean Estimators: Standard means are highly sensitive to outliers.
  • Median-of-Means (MoM): Recent extensions to manifolds improve deviation bounds but rely on arbitrary partitioning of data and their robustness/uniqueness in the presence of outliers remains under-explored.
• Goal: Develop a robust, computationally efficient, and unique location estimator for various matrix manifolds that maintains desirable equivariance properties.

2. Methodology: The Projected Frobenius Median (PFM)

The authors propose a two-step extrinsic approach that leverages the geometry of the ambient Euclidean space to simplify computation while preserving robustness.

Step 1: Ambient Space Computation

Instead of minimizing intrinsic distances on the manifold, the method first computes the Frobenius Median in the linear ambient space ($\mathcal{A}$) containing the manifold ($\mathcal{M}$).

• Frobenius Norm: For a matrix $X$, the norm is $\|X\|_F = \sqrt{\text{tr}(X^\top X)}$.
• Spatial Median Connection: The Frobenius median is mathematically equivalent to the spatial median (geometric median) of the vectorized data in the ambient Euclidean space.
  • For Stiefel manifolds ($\mathcal{V}_{k,r}$), vectorization is standard (vec).
  • For Grassmann manifolds ($\mathcal{G}_{k,r}$) and Complex Projective spaces, specific vectorization operators (vech$\sqrt{2}$, vec$\mathcal{H}$) are used to handle symmetry and Hermitian properties, ensuring the Frobenius norm equals the Euclidean norm of the vectorized form.
• Computation: This step utilizes well-established, efficient algorithms for the spatial median (e.g., Weiszfeld's algorithm), avoiding the iterative minimization of intrinsic distances.

Step 2: Projection

The computed ambient median $\hat{A}$ is projected onto the target manifold $\mathcal{M}$ to obtain the Projected Frobenius Median (PFM), denoted $\hat{M} = \pi(\hat{A}; \mathcal{M})$.

• Projection Rules:
  • Stiefel ($\mathcal{V}_{k,r}$): Project via Singular Value Decomposition (SVD). If $\hat{A} = U \Sigma V^\top$, the projection is $UV^\top$.
  • Grassmann ($\mathcal{G}_{k,r}$): Project via Spectral Decomposition. If $\hat{A} = \sum \lambda_i q_i q_i^\top$, the projection is $\sum_{i=1}^r q_i q_i^\top$ (sum of the top $r$ eigenvectors).
  • Complex Projective Space ($\mathcal{CP}^{k-1}$): Similar to Grassmann but for rank-1 Hermitian matrices; projection is the outer product of the leading eigenvector.
  • Projective Stiefel ($\mathcal{PV}_{k,r}$): A more complex two-stage projection involving handling sign ambiguities ($\pm 1$) and nested ambient spaces.

3. Key Contributions

1. Unified Framework: The PFM is applicable to a broad class of matrix manifolds, including Stiefel, Grassmann, Complex Projective, and Projective Stiefel manifolds.
2. Theoretical Guarantees:
   • Uniqueness: The estimator is unique provided the sample data are not colinear in the ambient space.
   • Equivariance: The method possesses natural equivariance properties under orthogonal/unitary transformations (e.g., $X \to QXR^\top$), ensuring the estimator transforms consistently with the data.
   • Asymptotic Normality: The authors derive the Influence Function (IF) and establish the Central Limit Theorem (CLT) for the PFM on these manifolds, characterizing the asymptotic distribution in the relevant tangent spaces.
3. Computational Efficiency: By reducing the problem to a Euclidean spatial median calculation followed by a closed-form projection (SVD/Eigen-decomposition), the method avoids the heavy iterative optimization required by intrinsic Fréchet medians.

4. Results

The paper validates the method through simulations and real-world data analysis.

Simulation Studies

• Planar Shape Space (Complex Projective Space):
  • Setup: Data generated from a complex Bingham distribution with added outliers (up to 45% contamination).
  • Comparison: PFM (labeled EMedian) vs. Intrinsic Fréchet Mean (IMean), Intrinsic Fréchet Median (IMedian), and Median-of-Means (MoM).
  • Findings: EMedian consistently outperformed intrinsic methods. While IMedian showed some robustness, its error increased significantly with contamination, and it frequently converged to local minima. MoM performed well only under mild contamination. EMedian maintained low estimation error and stability even at 45% contamination.
• Projective Stiefel Manifolds:
  • Setup: Data from a frame Watson distribution with outliers.
  • Comparison: PFM vs. a non-robust mean estimator (Arnold & Jupp, 2013).
  • Findings: The mean estimator's error deteriorated rapidly with outliers. The PFM remained stable and accurate even with 40% contamination, demonstrating strong robustness.

Real-World Application: Earthquake Moment Tensors

• Data: Seismic moment tensors (3x3 symmetric, zero-trace matrices) from Papua New Guinea and Solomon Islands (2006–2016). These represent orthogonal axial frames (T, B, P axes).
• Analysis: The authors estimated the central axial frame for a specific region containing suspected outliers.
• Findings:
  • Under normal conditions, the mean and median were similar.
  • When outliers were amplified or symmetry broken, the sample mean shifted significantly toward the outliers.
  • The spatial median (PFM) remained stable, with its bootstrap confidence regions containing the true population frame, whereas the mean was pulled outside these regions.

5. Significance

• Practical Utility: The PFM offers a "plug-and-play" solution for robust statistics on manifolds, utilizing existing software for spatial medians rather than requiring custom manifold optimization code.
• Robustness: It solves the critical issue of outlier sensitivity in non-Euclidean data, which is vital for applications like shape analysis and geophysics where data quality can be variable.
• Theoretical Foundation: By establishing the influence function and asymptotic normality, the paper provides the necessary statistical tools for inference (e.g., hypothesis testing, confidence intervals) in these complex spaces.
• Future Directions: The work opens avenues for $k$-sample hypothesis testing on projective Stiefel manifolds and the development of robust dispersion estimators for compact parameter spaces.

In summary, the paper successfully bridges the gap between robust statistics and differential geometry, providing a method that is theoretically sound, computationally efficient, and empirically superior to existing alternatives for handling outliers in matrix manifold data.