🔢 mathematics

Poisson-Sampled Fréchet Means on Gaussian Information Manifolds

This paper establishes a rigorous finite-window theory for Poisson-sampled Fréchet means on Gaussian information manifolds, deriving exact statistical properties such as consistency, central limit theorems, and error decompositions for spatial networks with distribution-valued marks, while specializing the results to covariance-varying Gaussian models and Wasserstein geometry.

Original authors: Gourab Ghatak

Published 2026-07-23

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Gourab Ghatak

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

The Science of Averaging the Un-Averagable

Imagine you are trying to find the "average" location of a flock of birds, but these birds aren't just points in space; they are carrying entire weather maps, probability charts, or complex 3D shapes in their beaks. This is the world of information geometry, a branch of science where data isn't just a list of numbers, but a shape living on a curved surface. Think of a flat sheet of paper versus a crumpled ball of paper or a saddle shape. On a flat sheet, the average of two points is just the spot right in the middle. But on a curved surface, the "middle" might be a completely different place, and the straight line you'd draw to get there might actually curve around the shape.

Now, imagine these birds appear randomly, like raindrops hitting a window, governed by a Poisson process. This is a fancy way of saying they show up at random times and places, and sometimes, by pure bad luck, no birds show up at all in the area you are looking at. The big question scientists have been wrestling with is: How do you calculate a true, reliable average (called a Fréchet mean) when your data is both curved and randomly sparse? If you just try to average everything together without accounting for the randomness of how many birds showed up, your math breaks, and your "average" might be a ghost that doesn't exist. This paper tackles that exact puzzle, providing a rigorous rulebook for how to find the center of a crowd when the crowd itself is a mystery.

The Paper's Big Idea: Counting the Raindrops to Find the Center

The authors, led by Gourab Ghatak, have built a precise mathematical framework to solve the problem of averaging these complex, shape-carrying data points when they are sampled randomly. They realized that previous methods often made a dangerous mistake: they assumed the number of data points was fixed or ignored the fact that sometimes the window is empty.

The "Zero-Count" Problem and the Magic Formula
The paper starts by fixing a fundamental flaw in how we usually think about averages. If you look at a small patch of sky and count the birds, you might get zero. If you get zero, you can't calculate an average. The authors insist that we must condition our math on the fact that we did see at least one bird. They introduce a special "magic formula" (a factor called $H(m)$ ) that corrects the average based on how likely it was to have a small or large number of birds.

Here is the kicker: The paper proves that the simple guess of "1 divided by the average number of birds" ( $1/m$ ) is wrong. It's only a rough guess for when you have a huge number of birds. When the number of birds is small, the correction factor is much larger. For example, if you expect 5 birds on average, the simple guess says the correction is 0.2, but the paper's exact math shows it's actually about 0.258. This difference matters a lot when you are dealing with rare events or small sample sizes.

The "Correlation Floor": Why More Data Doesn't Always Help
One of the most fascinating discoveries in the paper is what happens when the birds aren't just random individuals but are part of a single, connected weather system (a "spatially correlated field"). Imagine the birds are all reacting to the same wind gust.

The authors show that if you keep adding more birds to your window (increasing the density), you eventually hit a "correlation floor." This is a hard limit on how accurate your average can get. No matter how many birds you count, you can't average out the fact that they are all moving together. The error in your average stops shrinking and stays stuck at a specific level determined by how connected the birds are.

However, if you make your observation window bigger (looking at a larger area of sky) instead of just packing more birds into the same spot, you can break through this floor. The paper provides exact formulas showing that expanding the window reduces the error, while just densifying the same spot does not.

The "Thinning" Trick
The paper also explores what happens if you randomly throw away some of your data (a process called "thinning"), like keeping only every second bird that lands. They found that if you compare the average of the birds you kept to the average of all the birds (including the ones you threw away), the error between them is surprisingly small and predictable. This is because both averages are looking at the same underlying weather pattern. The "correlation floor" cancels out in this comparison, meaning the two averages stay very close to each other, even if you throw away half the data.

Where This Works: Curved Spaces and Moving Covariances
The authors tested their theory on two specific types of curved spaces where data lives:

The Univariate Gaussian Manifold: This is where data is just a bell curve with a changing width (variance). The paper shows that the "average" of two bell curves with the same width but different centers isn't just a bell curve in the middle with the same width. The average actually has a wider width. This is a counter-intuitive result that only appears when you respect the true curvature of the space.
The Covariance Manifold: This is for complex, multi-dimensional data where the relationships between variables (the covariance matrix) change. The paper handles cases where these matrices don't "play nice" (they don't commute), which is a common headache in real-world data. They proved that even with these messy, non-commuting matrices, their formulas for the average and the error hold true.

What the Paper Rules Out
The authors are very careful to say what their theory doesn't cover. They explicitly rule out the idea that you can just use simple, flat-space math (like a standard arithmetic average) for these problems. They show that restricting data to a "fixed covariance" (keeping the width of the bell curve the same) removes the interesting geometric content and leads to wrong answers if you try to apply it to the full, curved space. They also clarify that their results for "Wasserstein" geometry (a different way of measuring distance between shapes) only work in very specific, simple cases and don't apply to the general curved spaces they study.

How Sure Are They?
The paper is not just a guess or a simulation. The authors have derived exact mathematical proofs for their main formulas. They have shown that their "magic formula" for the count correction is mathematically precise, not an approximation. They also ran computer simulations (Monte Carlo trials) to double-check their math, and the numbers matched their exact formulas perfectly, down to the tiny decimal places. For instance, in a test with non-commuting matrices, their exact formula predicted a risk of 0.118711, and the simulation gave 0.118578, a difference so small it's likely just computer rounding noise.

The Takeaway
In short, this paper gives us a new, rigorous way to find the "center" of a crowd of complex, shape-shifting data points when the crowd size is random and the points are connected. It teaches us that you can't just count heads and divide; you have to account for the randomness of the count itself and the hidden connections between the data points. If you ignore these factors, your average might be a mirage. But with the paper's new tools, we can calculate the true average, understand the limits of our accuracy, and know exactly how much error to expect, whether we are looking at a few data points or a massive, expanding window.

Technical Summary: Poisson-Sampled Fréchet Means on Gaussian Information Manifolds

Problem Statement

The paper addresses the challenge of averaging distribution-valued marks (e.g., Gaussian posteriors) attached to the locations of a spatial network modeled by a Poisson point process. This problem sits at the intersection of stochastic geometry and information geometry. Three primary difficulties arise when combining these theories:

Infinite Integrals: A stationary Poisson process on $\mathbb{R}^s$ has infinitely many points, causing the unnormalized integral of squared distances (the standard Fréchet objective) to diverge in non-degenerate models.
Random Sample Size: A bounded observation window contains a Poisson-distributed number of marks, which is zero with positive probability. Standard fixed-sample Fréchet mean theory does not directly apply to this random-count scenario.
Convergence of Minimizers: While Campbell's formula handles the expectation of random sums, it does not inherently justify the convergence of random minimizers or the interchange of minimization and expectation.

Furthermore, restricting Gaussian marks to a common covariance matrix flattens the geometry, removing the intrinsic information-geometric content. The paper argues that to genuinely interact with information geometry, non-trivial covariance variation must be retained, even though the set of common-covariance matrices is not geodesically closed in the full Fisher–Rao manifold.

Methodology

The authors develop a rigorous finite-window theory to resolve the divergence and random-count issues.

Finite-Window Population Target: Instead of an unnormalized objective over the whole space, the population target is defined as the barycenter of the normalized window mark law ( $Q_B$ ). This ensures the objective is well-posed.
Conditional Sampling Representation: Conditional on the Poisson count $N_B = n > 0$ , the marked points form an ordinary independent and identically distributed (i.i.d.) sample from $Q_B$ . This allows the transfer of fixed-sample results to the random-count setting.
Exact Poissonization Operator: The paper introduces a zero-truncated Poissonization operator, $P_m[\cdot]$ , which exactly transforms fixed-sample risks, tail bounds, and limit theorems into their random-count counterparts. A key finding is that the reciprocal-count factor is not $1/m$ , but rather $H(m) = \mathbb{E}[1/N \mid N > 0]$ , which involves the exponential integral and Euler–Mascheroni constant.
Geodesic Reduction: For marks supported on a single geodesic, the manifold problem reduces to a one-dimensional problem in arclength coordinates. This allows for closed-form solutions even in curved spaces.
Stochastic Geometry Tools: The analysis utilizes Slivnyak's theorem for Palm distributions (distinguishing reduced vs. non-reduced estimators) and Poisson splitting for thinning operations.

Key Contributions and Results

1. Exact Poissonization Calculus

The paper establishes an exact interface between fixed-sample Fréchet mean theory and Poisson sampling.

Risk and Tails: Any fixed-sample risk bound $r_n$ or tail bound $\alpha_n(u)$ can be converted to the Poisson case via $P_m$ . For example, if $r_n \le C/n$ , the Poisson risk is bounded by $C H(m)$ .
Central Limit Theorems (CLT): An intrinsic CLT for fixed samples transfers to the random-index case. If $\sqrt{n} \text{Log}_b(\hat{b}_n) \Rightarrow Z$ , then $\sqrt{m} \text{Log}_b(\hat{b}_{N_m}) \Rightarrow Z$ conditional on $N_m > 0$ .

2. Geodesic-Supported Marks and Spatial Correlation

When marks lie on a geodesic (e.g., log-scale Gaussians or covariance matrices along a geodesic path), the authors derive exact mean-square errors (MSE).

Independent Marks: The MSE is $v H(m)$ , where $v$ is the variance of the underlying coordinate.
Spatially Correlated Fields: For a common random field, the MSE decomposes exactly into two terms:
$\text{MSE} = C_B + (V_B - C_B)H(m_B)$
where $C_B$ $C_{B}$ is the integrated field covariance (the "correlation floor") and $V_B$ $V_{B}$ is the local variance.
- Densification vs. Window Growth: Increasing point density ( $\lambda \to \infty$ ) on a fixed window reduces the Poisson term but leaves the correlation floor $C_B$ intact. Expanding the window ( $|B| \to \infty$ ) reduces both terms, with the asymptotic constant involving the integrated covariance of the field.
Spatial CLT: For stationary Gaussian fields, a spatial CLT is proven where the limiting variance is the sum of the integrated field covariance and the Poisson diagonal contribution.

3. Palm Distributions and Thinning

Palm Laws: Using Slivnyak's theorem, the paper derives exact risk formulas for both reduced (excluding the typical point) and non-reduced (including the typical point) Palm estimators. The non-reduced estimator involves different count factors ( $G_1, G_2$ ) and retains the typical point's correlation with the field.
Thinning: The paper provides exact formulas for the distortion between a full barycenter and a thinned (retained) barycenter. Crucially, when both are computed from the same field realization, the correlation floor $C_B$ cancels out in the difference, a dependence absent when comparing independent field copies.

4. Gaussian Information-Geometric Specializations

The theory is specialized to two curved Gaussian manifolds:

Univariate Fisher–Rao Manifold: Modeled as a scaled hyperbolic half-plane. The paper highlights that the geodesic midpoint of two equal-variance Gaussians generally has a larger variance, demonstrating that fixed-variance submanifolds are not geodesically closed.
Affine-Invariant Covariance Manifold: The theory applies to multivariate covariance matrices with fixed mean, accommodating arbitrary non-commuting matrices. The population barycenter satisfies a non-commutative Karcher equation.
Wasserstein Geometry: The paper briefly notes that while 1D Wasserstein barycenters have a simple form, the general Bures–Wasserstein geometry is not globally CAT(0), so the Hadamard space results do not apply there.

Significance and Scope

The paper claims its significance lies in providing a mathematical interface rather than a universal semantic application. It explicitly avoids proposing bandit policies or comparing estimators to baselines that discard covariance information.

Rigorous Finite-Window Framework: It resolves the divergence issues of infinite Poisson processes by defining a well-posed population target on bounded windows.
Exactness: The results are not asymptotic approximations for finite $m$ ; they provide exact formulas for risks, tails, and distortions involving specific Poisson count factors ( $H(m), G_1(m)$ , etc.).
Separation of Effects: The framework cleanly separates the effects of random spatial sampling (Poisson count), spatial dependence (correlation floor), and intrinsic geometry (curvature).
Limitations: The exact second-order formulas require marks to lie on a single geodesic. The paper does not treat arbitrary correlated fields wandering in multiple tangent directions of a curved manifold, nor does it treat the full multivariate Gaussian manifold with simultaneous mean and covariance variation as a Hadamard space, as closed-form distances are unavailable there.

The work concludes that while the theory is narrower than a multi-application framework, it is "materially stronger" because every target is well-defined, zero counts are explicitly handled, and curvature assumptions are strictly matched to the specific Gaussian submanifolds where they hold.