In standard statistics, we are taught to find the average of two numbers by adding them and dividing by two. This works because numbers exist in a "flat" Euclidean space where the shortest path between any two points is a straight line. However, much of the data we care about today—the 3D shape of a prostate, the orientation of fibers in the brain, or the rotation of a robotic arm—does not live in a flat space.
When data resides on a curved surface, or manifold, standard arithmetic fails. As Anthony Yezzi and Hamid Krim point out in Statistics and Analysis of Shapes, an arithmetic mean of two valid shapes might result in a "shape" that is mathematically impossible or anatomically nonsensical. To analyze this data correctly, we must move from linear algebra to Riemannian geometry.
The Geometry of the Mean
The most fundamental operation in statistics is finding the "center" of a dataset. On a manifold, we replace the arithmetic mean with the Fréchet mean (also known as the Karcher mean).
Instead of summing coordinates, the Fréchet mean is defined as the point that minimizes the sum of squared geodesic distances to all data points. A geodesic is the generalization of a straight line to curved spaces—like the great-circle path an airplane takes over the Earth.
While calculating this mean traditionally required expensive batch optimization, Jeffrey Ho, Baba C. Vemuri, and colleagues demonstrated in Recursive Karcher Expectation Estimators And Geometric Law of Large Numbers that we can estimate this mean recursively. By moving incrementally along geodesics as new data arrives, we can find the "geometric center" of complex objects like symmetric positive-definite (SPD) matrices orders of magnitude faster than before.
Principal Geodesic Analysis
Once we have a mean, we usually want to understand variation. In flat space, we use Principal Component Analysis (PCA) to find the straight lines along which data varies most. On a manifold, these "lines" become geodesics.
In Prostate Shape Modeling Based on Principal Geodesic Analysis Bootstrapping, Erik Dam and P. Thomas Fletcher introduced Principal Geodesic Analysis (PGA). Instead of projecting data onto a flat plane, PGA projects data onto geodesic submanifolds. This ensures that the "variations" captured by the model remain within the valid geometric space of the object, allowing researchers to capture 98% of prostate shape variation without creating "broken" intermediate shapes.
Why Intrinsic Geometry Matters in Neuroimaging
The need for these tools is most acute in medical imaging. As Hongtu Zhu and colleagues argue in Statistical Learning Methods for Neuroimaging Data Analysis with Applications, brain data often consists of tensors or directional vectors that exist on non-Euclidean manifolds.
If you treat a brain tensor (which must be positive-definite) as a simple vector and apply standard linear regression, you might predict a "negative" tensor—a physical impossibility. By using geodesic regression, researchers like Aditya Kulkarni can model how brain structures change over time while respecting the manifold's curvature (Differentially Private Geodesic Regression).
Similarly, Jörn Schulz, Stephen M. Pizer, and J. S. Marron developed NEUJIVE to analyze how multiple brain structures, like the hippocampus and caudate nucleus, vary together (Non-Euclidean Analysis of Joint Variations in Multi-Object Shapes). By accounting for the non-linear geometry of these shapes, they significantly improved the accuracy of early Autism Spectrum Disorder classification.
Scaling Geometric Statistics
The transition from "flat" to "curved" statistics is computationally demanding. To make these methods accessible, researchers have developed unified frameworks and libraries. Alice Le Brigant, Nina Miolane, and others created Geomstats, a Python library that treats shapes, curves, and surfaces as points in abstract manifolds (Learning from landmarks, curves, surfaces, and shapes in Geomstats).
These frameworks allow for more sophisticated modeling, such as:
- Manifold Normalization: Rudrasis Chakraborty extended deep learning normalization techniques to Riemannian manifolds, ensuring that neural networks can process manifold-valued data without destroying its geometric integrity (ManifoldNorm).
- Anisotropic Clustering: Hengchao Chen and Qiang Sun developed mixtures of geodesic factor analyzers to identify "stretched" clusters on spheres, providing a better way to find clinical subtypes in neuroimaging (Mixtures of geodesic factor analyzers on Riemannian homogeneous spaces).
By shifting the perspective from "data in a box" to "data on a surface," this field provides the rigorous mathematical foundation necessary to analyze the complex, non-linear structures of the natural world.