Geometric statistics and its link to statisticians working outside it

Question

As an Amazon Associate, I earn from qualifying purchases. For more practice problems like this, see Schaum’s Outline of Statistics, 6th Edition.

1. Restatement of the Question (Plain Language)

The author works on statistics for data that live on curved spaces (Riemannian manifolds, stratified spaces, etc.).
They would like to know:

When does the geometry of the sample space actually change the statistical conclusions?
In other words, can we point to concrete situations where curvature, topology, or singularities materially affect things such as confidence intervals, hypothesis tests, or estimators – not just change the formulas in a cosmetic way?

The answer should give specific examples (both asymptotic and finite‑sample) that illustrate the impact of geometry, and explain why these effects matter to practitioners who normally work in Euclidean settings.

2. Step‑by‑Step Explanation

Below we walk through the main ways geometry can influence statistical inference, illustrated with concrete models and references.
Each point is presented as a logical “step” that builds on the previous one.

Step 1 – Define the Fréchet Mean and Its Euclidean Counterpart

Euclidean case: For i.i.d. points (X_1,\dots ,X_n\in\mathbb R^d) with mean (\mu), the sample mean (\bar X) minimises the squared Euclidean distance
[ \bar X = \arg\min_{p\in\mathbb R^d}\frac1n\sum_{i=1}^n|X_i-p|^2 . ]
Manifold / metric‑space case: With a metric (d(\cdot,\cdot)) on a space (\mathcal M), the (population) Fréchet mean(s) are the minimisers of the Fréchet functional
[ F(p)=\mathbb E\big[ d^2(X,p) \big] . ] The sample Fréchet mean (\hat\mu_n) minimises the empirical version (F_n(p)=\frac1n\sum_{i=1}^n d^2(X_i,p)).

If (\mathcal M=\mathbb R^d) with the Euclidean metric, both definitions coincide, and the familiar properties of the ordinary mean follow.

Step 2 – Asymptotic Distribution: The Role of the Hessian

On a smooth Riemannian manifold ((\mathcal M,g)) with unique Fréchet mean (\mu),

[ \sqrt{n}\,\log_\mu(\hat\mu_n) \;\xrightarrow{d}\; \mathcal N!\bigl(0,\; \Sigma\bigr), \qquad \Sigma = H^{-1} \, \Gamma \, H^{-1}, ]

where

(\log_\mu) is the Riemannian log map (tangent‑space coordinate).
(H = \nabla^2 F(\mu)) is the Hessian (the matrix of second derivatives of the Fréchet function at (\mu)).
(\Gamma = \operatorname{Cov}\bigl(\operatorname{grad} d^2(X,\mu)\bigr)) is the covariance of the Riemannian gradients.

Why geometry matters:

In (\mathbb R^d), (H = 2I_d) (the identity) and (\Sigma = \frac14 \Gamma); the limiting covariance depends only on the data distribution.
On a curved space, (H) contains curvature information (e.g., sectional curvature appears in the second‑order expansion of squared distance).
If curvature makes (H) singular (e.g., at cut‑locus points or on manifolds with non‑positive curvature), the limiting covariance can blow up or collapse in certain directions, changing the shape of confidence ellipsoids.

Concrete illustration – Spherical data:

Let (X_i) be i.i.d. on the unit sphere (S^{k}) with a von Mises–Fisher distribution (\mathrm{vMF}(\mu,\kappa)).
The Fréchet mean is (\mu).
The Hessian (H = \kappa \, (I - \mu\mu^{!\top})) has eigenvalue (\kappa) in the tangent space; as (\kappa\to 0) (the uniform distribution) the Hessian tends to zero, making the CLT variance diverge.
This explains why inference near the uniform distribution on a sphere is dramatically less precise than in Euclidean space with the same concentration parameter.

Step 3 – Stickiness on Stratified / Singular Spaces

A stratified space is a union of manifolds glued together along lower‑dimensional pieces (e.g., a tree space, a space of phylogenetic trees, or the space of symmetric positive‑definite matrices with rank constraints).

Phenomenon:
When the population Fréchet mean lies on a lower‑dimensional stratum, the sample Fréchet means eventually stick to that stratum with probability 1, even if the data are perturbed off it.

Why it is a statistical effect:

The usual Euclidean intuition that the sample mean can wander arbitrarily close to any point in the convex hull fails.
Confidence regions built assuming a full‑dimensional Gaussian limit are wrong: the true limiting distribution lives on a lower‑dimensional linear subspace (often a mixture of a Gaussian on the stratum and a point mass).

Example – BHV Tree Space:

The Billera–Holmes–Vogtmann (BHV) space of phylogenetic trees is a CAT(0) space consisting of orthants glued along faces.
If the true tree is a star tree (all leaf edges of length zero), the population Fréchet mean lies at the cone point, a singularity of codimension (>0).
Empirically, the sample Fréchet means converge to a point on the cone, sticking there after finitely many samples (see Hotz et al., 2013; Barden et al., 2016).
Consequently, a Wald test that ignores stickiness will have inflated Type I error.

Step 4 – Finite‑Sample Smeariness (Curvature‑Dependent Bias & Variance)

Even when the limiting CLT looks Euclidean, curvature can cause finite‑sample deviations that are not negligible for realistic (n).

Mechanism:

The squared‑distance function on a positively curved manifold is convex only up to the injectivity radius.
For samples that occasionally fall outside that radius, the empirical Fréchet function becomes flattened, biasing the minimiser toward the centre of curvature.

Quantitative result on the sphere (S^{k}):

Let (X_i) be i.i.d. from a von Mises–Fisher distribution with concentration (\kappa).
For moderate (n) (e.g., (n=30)), the bias of (\hat\mu_n) is approximately
[ \operatorname{Bias}\bigl(\hat\mu_n\bigr) \;\approx\; \frac{c(k)}{\kappa\, n}\,\mu, ] where (c(k)=\frac{k-1}{2}).
The variance inflation factor is (1 + \frac{c(k)}{\kappa^2 n}).
Both corrections depend on curvature through the factor (c(k)) (which is zero in Euclidean space).

Practical implication:

Wald‑type confidence regions that use the Euclidean variance underestimate the true variability unless they are re‑calibrated by a curvature‑dependent factor (e.g., the “curvature correction” of Bhattacharya & Patrangenaru, 2005).
In neuroimaging, diffusion‑tensor imaging (DTI) data are SPD matrices; ignoring the manifold’s curvature leads to under‑coverage of confidence ellipsoids for mean tensors.

Step 5 – Impact on Regression, KDE, and Hypothesis Testing

Method	Euclidean formulation	Geometric modification	Statistical effect
Geodesic regression (Müller & others)	Linear model (Y = \beta_0 + \beta_1 X + \varepsilon)	Replace straight lines by geodesics; parameters live in the tangent space at a base point; the exponential map is used to map back.	Curvature causes bias in slope estimates; asymptotic covariance contains the Riemannian curvature tensor.
Kernel density estimation (KDE) on (\mathcal M)	( \hat f(x)=\frac{1}{nh^d}\sum K\bigl(\|x-X_i\|/h\bigr))	Use the geodesic distance and the volume element (\sqrt{\det g}).	Bandwidth selection depends on Ricci curvature; in high curvature regions the effective bandwidth is larger, leading to oversmoothing if Euclidean bandwidth is used.
Two‑sample test for equality of means	(T = n\|\bar X_1-\bar X_2\|^2 / \hat\sigma^2)	Replace Euclidean norm by Riemannian distance, and use Hessian‑adjusted covariance.	Critical values change; ignoring curvature inflates Type I error (demonstrated for SPD matrices in Lenglet et al., 2020).

Thus geometry permeates every standard statistical tool when the data live off (\mathbb R^d).

Step 6 – Summary of When Geometry Is Material

Situation	Geometry‑induced statistical effect	Typical magnitude
Unique smooth mean, moderate curvature	Hessian scaling of asymptotic variance; bias of order (1/n)	Often noticeable for (n\lesssim 100) (e.g., directional data on (S^2))
Mean at or near cut‑locus / singularity	Singular Hessian → inflated variance or degenerate limit; stickiness	Can dominate even for large (n) (tree space star‑tree case)
High curvature (small injectivity radius)	Finite‑sample smeariness (bias + variance inflation)	Requires explicit curvature correction; under‑coverage up to 20 % if ignored
Manifold with boundary or stratification	Sample means may lie on lower‑dimensional strata → reduced effective dimension	Affects hypothesis tests and confidence sets; leads to non‑Gaussian limiting distributions
Non‑Euclidean group actions (e.g., rotations)	Quotient spaces produce orbifold singularities → multimodal Fréchet functions	Can cause spurious multiple local minima; bootstrap may fail

Whenever any of the above rows applies, the geometry is material: it changes the shape of confidence regions, the power of tests, or even the consistency of estimators.

3. Final Answer (Key Take‑aways)

Geometry matters whenever the Fréchet (or intrinsic) mean is influenced by curvature or singularities.
The Hessian of the Fréchet function incorporates curvature; a non‑identity Hessian rescales the asymptotic covariance and can become singular, leading to degenerate CLTs.
Stickiness on stratified spaces forces sample means onto lower‑dimensional strata, producing non‑Gaussian limits and invalidating Euclidean Wald tests.
Finite‑sample smeariness shows that curvature can cause noticeable bias and variance inflation for realistic sample sizes; curvature‑dependent recalibration of Wald‑type procedures is essential.
These phenomena appear in concrete applied domains: directional statistics, diffusion‑tensor imaging (SPD matrices), shape analysis, phylogenetic tree spaces, and more.
Practitioners outside the “geometric statistics” community need to adjust standard tools (regression, KDE, two‑sample tests) whenever data live on manifolds with non‑trivial curvature or singularities; otherwise inference can be misleading.

4. Common Mistakes When Dealing With Geometry‑Based Inference

Mistake	Why It Happens	How to Avoid It
Treating the Fréchet mean as if it were the Euclidean mean (ignoring the Hessian).	The definition looks similar; the extra curvature term is easy to overlook.	Explicitly compute or approximate the Hessian (H = \nabla^2F(\mu)) and use (H^{-1}\Gamma H^{-1}) for the asymptotic variance.
Assuming a normal limit in all cases (using standard Wald tests).	The CLT formula is often quoted without its regularity

Original question: Geometric statistics and its link to statisticians working outside it on Cross Validated (Stats Stack Exchange), licensed CC BY-SA.