🤖 machine learning

Sharp Concentration Bounds for Bundle-Valued Statistics on Manifolds

This paper establishes non-asymptotic, dimension-free concentration bounds for transported empirical means of bundle-valued statistics on manifolds, revealing a fundamental bias-variance trade-off where curvature-induced holonomy creates an irreducible error floor alongside standard stochastic fluctuations.

Original authors: Swagatam Das, Vaclav Snasel

Published 2026-07-14

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Swagatam Das, Vaclav Snasel

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to draw a map of the wind blowing across the entire Earth. At every single spot on the globe, you have a tiny, flat arrow showing the wind's speed and direction. In math-speak, these arrows live in "fibers"—little personal vector spaces attached to each point on the Earth's surface.

The problem? The Earth is round. If you try to drag all these local wind arrows to one central spot (say, the North Pole) to calculate an "average wind," you run into a geometric headache. Because the Earth curves, the path you take to drag an arrow matters. If you drag an arrow along the equator versus dragging it over the pole, it might end up pointing in a slightly different direction when it arrives, even if the wind was the same. This twisting effect is called holonomy, and it's caused by the curvature of the planet.

For a long time, statisticians and machine learning experts assumed that if you just gathered enough data (more arrows), the "noise" would wash out, and you'd get a perfect average. This paper says: Not so fast.

The Main Discovery: A Two-Part Error

The authors, Swagatam Das and Václav Snášel, prove that when you average these curved-world data points, your error isn't just one thing. It's actually two things stuck together:

The Random Jitter (The Good News): This is the usual "noise" you expect. If you have a small sample, your average is wobbly. But as you collect more data ( $n$ ), this jitter shrinks. Specifically, it gets smaller at a rate of $1/\sqrt{n}$ . If you quadruple your data, this part of the error cuts in half. This behaves exactly like standard statistics on a flat piece of paper.
The Curvature Floor (The Bad News): This is the paper's big reveal. Even if you collect infinite data, there is a stubborn, unshakeable error floor left over. This is the holonomy bias. It's a deterministic offset caused purely by the geometry of the world. No matter how many wind measurements you take, if the Earth is curved and your data is spread out, your "average" will always be slightly twisted away from the truth.

What They Ruled Out

The paper explicitly argues against the idea that "more data solves everything" in curved spaces.

It's not just a lack of data: You cannot fix the curvature error by simply gathering more samples. The paper proves mathematically that this bias is unavoidable for any method that tries to align data to a common spot (a "transport-based estimator").
It's not just a math glitch: This isn't a flaw in their calculation; it's a fundamental property of the universe they are modeling. If the curvature is high and your data is spread over a large area, this error floor is real and permanent.

How Sure Are They?

The authors are extremely confident. They didn't just guess; they proved it.

The Math: They derived sharp, non-asymptotic bounds (mathematical guarantees that hold for any sample size, not just huge ones). They used rigorous inequalities (Hoeffding and Bernstein types) adapted for curved spaces.
The Lower Bounds: They proved that no algorithm can do better than their formula. They showed that the error must be at least the sum of the random jitter and the curvature floor.
The Simulation: To back up the theory, they ran controlled experiments on a sphere (specifically a sphere with radius $r=1$ $r = 1$ ). They simulated data and measured the error.
- The Result: The simulations matched the theory almost perfectly. The "jitter" part shrank exactly as predicted ( $n^{-1/2}$ ), while the "curvature floor" stayed perfectly flat, refusing to shrink even when they increased the sample size to 10,000.
- The Numbers: In their sphere experiment, the theoretical prediction for the error floor was $\Delta_{hol} = 2 \sin(\pi \rho^2 / 2)$ . When they measured it, the results were within 3.7% of the prediction across all tested configurations. For a data spread radius of $\rho = 1.0$ , the theoretical floor was 2.000, and the measured floor was 1.926.

The Takeaway for the Curious Teen

Think of it like trying to average a bunch of compass needles on a globe.

The Jitter: If you only look at 10 needles, your average direction is shaky. If you look at 10,000, it's steady.
The Floor: But because the globe is round, if you drag those needles from different places to one spot, they twist. If you spread your needles over a large patch of the globe (say, a radius $\rho$ ), that twist creates a permanent offset.

The paper tells us that in the world of geometric machine learning (like analyzing shapes, 3D models, or data on curved surfaces), you have to accept that geometry sets a hard limit on accuracy. You can't just throw more data at the problem to fix it. You have to either shrink the area you are looking at (make $\rho$ smaller) or use a special correction to account for the twist.

The authors provide a "recipe" for this:

If you want to be safe: Keep your data in a small, flat-looking bubble (a "normal ball") where the curvature doesn't twist things much.
If you must look at a big area: Accept that there will be a permanent error floor, roughly proportional to the curvature ( $\kappa$ ) times the square of your data's spread ( $D^2$ ).

In short: In a curved world, more data helps with the noise, but it can't fix the twist. And the authors have the math to prove exactly how big that twist is.

Technical Summary: Sharp Concentration Bounds for Bundle-Valued Statistics on Manifolds

1. Problem Statement

Many geometric machine learning and statistical pipelines operate on data that naturally reside in vector bundles rather than a single shared vector space. Examples include tangent vectors, local frames, and gauge fields, where each observation $s(X_i)$ belongs to a fiber $E_{X_i}$ attached to a specific point $X_i$ on a base Riemannian manifold $M$ .

To compute empirical averages (means) of such data, one must transport observations from their respective fibers to a common reference fiber $E_{x_0}$ via parallel transport and then average them. However, this process introduces two distinct challenges absent in classical Euclidean statistics:

Path Dependence: If the shortest path (minimizing geodesic) between a sample point $X_i$ and the reference point $x_0$ is not unique, the choice of transport path introduces a deterministic ambiguity.
Holonomy Bias: Due to the curvature of the manifold and the bundle connection, parallel transport around a closed loop (formed by two different paths) results in a non-trivial rotation (holonomy). This creates a deterministic bias in the transported mean that does not vanish with additional data, unlike stochastic sampling error.

Existing concentration theories (e.g., Hoeffding, Bernstein) apply to linear spaces, while existing manifold statistics literature (e.g., Fréchet means) often focuses on asymptotic results or data points on the base manifold, failing to address the non-asymptotic concentration of bundle-valued sections under parallel transport.

2. Methodology

The paper develops a non-asymptotic concentration theory for transported empirical means. The core methodological steps are:

Geometric Reduction: The authors assume a measurable transport rule (e.g., selecting a unique minimizing geodesic in a normal ball or a Cartan-Hadamard manifold) to map bundle-valued observations $s(X_i) \in E_{X_i}$ to a fixed reference Hilbert space $E_{x_0}$ . This transforms the problem into averaging i.i.d. random vectors in a fixed Hilbert space.
Sharp Inequalities: Leveraging the work of Pinelis [31] and Boucheron et al. [13], the paper applies dimension-free Hoeffding and Bernstein inequalities to the transported empirical mean $\bar{Y}_n$ . Crucially, the variance proxy is defined via the operator norm of the covariance matrix, ensuring the bounds remain independent of the fiber dimension $k$ .
Bias-Variance Decomposition: The total error is decomposed into:
1. Stochastic Term: Decays at the classical $O(n^{-1/2})$ rate.
2. Geometric Bias Term ( $\Delta_{hol}$ ): A deterministic offset caused by transport ambiguity (holonomy), controlled by the bundle curvature and the geometry of the data support.
Holonomy Quantification: The paper derives explicit, sharp formulas for the holonomy bias. For the tangent bundle of a round sphere $S^2_r$ , it provides a closed-form bound dependent on the area of the loop formed by alternative transport paths.
Robustness and Asymptotics: The framework is extended to heavy-tailed distributions using a median-of-means (MoM) estimator and establishes a Central Limit Theorem (CLT) in the reference fiber, provided the holonomy bias becomes negligible relative to sampling noise.

3. Key Contributions

The paper makes the following specific contributions:

Non-Asymptotic Concentration Bounds: It establishes finite-sample, dimension-free Hoeffding and Bernstein-type bounds for transported bundle-valued means. These bounds explicitly separate stochastic fluctuation from deterministic geometric bias.
Quantification of Holonomy Bias: The authors isolate the deterministic error floor introduced by non-unique geodesics. They provide sharp, closed-form formulas for this bias on the sphere (Proposition 1) and general curvature-based bounds (Theorem 3) for arbitrary bundles.
Minimax Lower Bounds: The paper proves that both the $O(n^{-1/2})$ stochastic rate and the curvature-driven holonomy floor are unavoidable within the class of transport-based estimators (Theorem 4). This confirms that no alignment-based algorithm can eliminate the geometric bias without changing the estimator's fundamental structure (e.g., moving to extrinsic methods).
Robust Estimation: A robust median-of-means estimator is established that achieves the optimal $\sigma/\sqrt{n}$ rate even under heavy-tailed distributions, preserving the bias-variance structure.
Central Limit Theorem: A CLT is proven for the transported mean in the reference fiber, valid when the holonomy bias is asymptotically negligible ( $o(n^{-1/2})$ ).

4. Results

Error Decomposition: The deviation of the transported mean $\bar{Y}_n$ from the population mean $m^\star$ is bounded by:
$\|\bar{Y}_n - m^\star\| \lesssim \frac{B}{\sqrt{n}} + \Delta_{hol}$
where $B$ is the uniform bound on the section, and $\Delta_{hol}$ is the holonomy bias.
Crossover Sample Size: The paper identifies a critical sample size $n_\times$ beyond which increasing the sample size $n$ yields no further reduction in total error because the geometric bias $\Delta_{hol}$ dominates. For a sphere of radius $r$ with data supported in a ball of radius $\rho$ , this crossover occurs roughly when $n \sim (r/\rho)^4$ .
Sphere Validation: Controlled experiments on the sphere $S^2$ validate the theoretical predictions. The stochastic error decays as $n^{-1/2}$ , while the error under alternative transport rules plateaus at a constant value matching the theoretical holonomy bound within 3.7%.
Applications: The theory is applied to manifold regression residuals, gauge-equivariant Graph Neural Networks (GNNs), diffusion tensor imaging, and Wasserstein tangent spaces, demonstrating how the bias-variance decomposition applies to these specific pipelines.

5. Significance and Claims

The paper claims to provide a rigorous statistical foundation for geometric deep learning and manifold learning pipelines that aggregate features via bundle connections. Its significance lies in:

Unifying Framework: It unifies Euclidean concentration theory, manifold statistics, and bundle geometry under a single probabilistic structure with explicit finite-sample guarantees.
Identifying Fundamental Limits: It demonstrates that in curved geometries, statistical accuracy is intrinsically limited by the geometry of the space (curvature/holonomy), not just by data scarcity. This challenges the assumption that more data always leads to better estimates in geometric settings.
Practical Guidance: The results offer concrete criteria for practitioners to determine when curvature effects are negligible (local regimes) and when they dominate (large caps), suggesting remedies such as restricting data support to smaller normal balls or applying holonomy corrections.
Modest Scope: The authors explicitly note that their results apply to transport-based estimators. They acknowledge that extrinsic estimators (which ignore fiber structure) might avoid the holonomy floor but at the cost of geometric interpretability. The paper does not claim to solve all geometric learning problems but specifically addresses the statistical behavior of averaging sections in vector bundles.

In summary, the paper establishes that while classical concentration rates ( $n^{-1/2}$ ) hold for the stochastic component of bundle-valued statistics, a curvature-driven error floor exists that cannot be eliminated by data alone, fundamentally altering the convergence behavior of geometric learning algorithms.