Size-Location Correlation for Set-Valued Processes: Theory, Estimation, and Laws of Large Numbers under $\rho$-Mixing

This paper introduces a variational framework based on the even-odd decomposition of support functions to define geometrically interpretable size-location correlation measures for set-valued processes, establishing their statistical properties under $\rho$-mixing and enabling the analysis of dependence structures that are inaccessible to traditional point-based or selection-based methods.

The Gist

Imagine you are trying to understand the weather patterns of a mysterious, shifting island. In traditional statistics, we usually treat the island as a single point on a map (like a lighthouse) to track its movement. But what if the island isn't just a point? What if it's a whole shape that can grow, shrink, rotate, and change its center all at once?

This paper proposes a new way to study these "shifting islands" (which mathematicians call random sets). Instead of squashing them down to a single point, the authors treat them as full, complex shapes and break them down into two distinct parts: Size and Location.

Here is the story of their discovery, explained simply:

1. The Problem: The "Blurry" View

Imagine you have a balloon that is floating around a room.

• Sometimes it gets bigger (inflates).
• Sometimes it moves to the corner (drifts).
• Sometimes it does both.

Old methods tried to track this balloon by looking at its "center of gravity" (like the Steiner point mentioned in the paper). But this is like trying to understand a storm by only watching the eye. You miss the wind, the rain, and the fact that the storm is expanding.

If the balloon is perfectly round and centered, old methods say, "It's not moving!" But in reality, the balloon might be pulsating (changing size) wildly, even if its center stays still. The old methods were blind to this "size" movement. They confused the two, making it impossible to tell if a change was due to the object moving or just changing shape.

2. The Solution: The "Even-Odd" Magic Trick

The authors invented a mathematical "laser cutter" that slices the data into two perfect, non-overlapping pieces. They call this the Even-Odd Decomposition.

Think of a shape as a song.

• The "Even" Part (Size): This is the rhythm. It tells you how big the shape is and how wide it is, regardless of where it is in the room. It's symmetric (the same on both sides).
• The "Odd" Part (Location): This is the melody. It tells you where the shape is leaning or drifting. It's the part that changes if you slide the shape across the room.

The Magic: In the old world, these two parts were tangled together like headphones in a pocket. In this new framework, they are perfectly separated. You can measure the "rhythm" (size) without hearing the "melody" (location), and vice versa.

3. The New Tools: Measuring the Dance

Once they separated the size and location, the authors created new tools to measure how these shapes interact with each other over time.

• The "Size Correlation": Do two shapes tend to grow and shrink together? (e.g., "When the first balloon inflates, the second one does too.")
• The "Location Correlation": Do they tend to drift in the same direction?
• The "Total Correlation": A mix of both.

They also created a new rule called $\rho$-mixing. Think of this as a "forgetting speed." If two shapes are far apart in time, do they still remember each other? The authors proved that even if the shapes are complex, we can predict how fast they "forget" their past, which is crucial for making long-term forecasts.

4. The Big Discovery: The "Invisible" Dependence

The most exciting part of the paper is a simulation they ran. They created a scenario where two shapes were completely dependent on each other, but in a way that old methods couldn't see.

• The Scenario: Two shapes shared the exact same "pulsing" rhythm (size), but their centers moved randomly and independently.
• The Old Method: Looked at the centers, saw they were random, and said, "These two shapes have zero connection."
• The New Method: Looked at the "rhythm" (size) and said, "These two shapes are dancing in perfect sync!"

This is like two people walking in a park. If you only track their feet (location), they seem to be walking randomly. But if you track their heartbeats (size), you realize they are walking in perfect rhythm. The old method missed the heartbeat; the new method caught it.

5. Why This Matters (The "So What?")

This isn't just about math puzzles. This framework helps us understand real-world uncertainty where things aren't just points.

• Finance: Instead of just tracking the average stock price (a point), we can track the range of possible prices (a shape). We can see if the volatility (size) of two stocks is linked, even if their average prices aren't.
• Robotics: When a robot's sensors are fuzzy, it doesn't know exactly where it is; it knows it's somewhere inside a shape. This math helps robots understand if their "uncertainty shape" is growing or shrinking in sync with other sensors.
• Climate Science: Instead of tracking the average temperature of a region, we can track the "shape" of the temperature range. We can see if the extremes (size) of weather patterns in two different cities are linked, even if the averages aren't.

Summary

The authors built a new pair of glasses. Before, we looked at shifting, growing, and shrinking shapes and only saw their centers. Now, we can see the size and location as separate, independent forces. This allows us to detect hidden connections, predict future behavior more accurately, and finally understand the true "dance" of uncertainty in our world.

In a nutshell: They stopped treating complex, shifting shapes like simple dots, and instead learned how to listen to their rhythm and melody separately.

Technical Summary

Here is a detailed technical summary of the paper "Size–Location Correlation for Set-Valued Processes: Theory, Estimation, and Laws of Large Numbers under ρ-Mixing" by Luc T. Tuyen.

1. Problem Statement

Set-valued random variables (random sets), such as random convex bodies or interval-valued data, are fundamental in stochastic geometry, symbolic data analysis, and robust optimization. A primary challenge in analyzing these objects is defining and estimating dependence structures (variance, covariance, correlation) that respect their geometric nature.

Existing approaches suffer from significant limitations:

• Aggregation Bias: Classical methods often rely on Aumann expectations or selections (e.g., the Steiner point), which aggregate the support function over all directions. This conflates distinct sources of variability: size (shape/width) and location (translation).
• Degeneracy: In cases of centrally symmetric sets, the location component vanishes identically, causing classical correlation measures to degenerate or fail to detect dependence.
• Loss of Directional Information: Finite-dimensional summaries (like the Steiner point) discard the rich directional information contained in the full support function, making it impossible to detect "directional location dependence" that does not manifest as a shift in the center.

The paper addresses the need for a variational framework that separates size and location, provides non-degenerate dependence measures, and establishes asymptotic laws for weakly dependent set-valued processes.

2. Methodology

The core of the proposed methodology is the canonical even–odd decomposition of the support function on the unit sphere $S^{d-1}$.

A. The Even–Odd Decomposition

For a convex compact random set $X$, its support function $h_X(u) = \sup_{x \in X} \langle x, u \rangle$ is decomposed into:

• Size Component (Even): $W_X(u) = \frac{1}{2}(h_X(u) + h_X(-u))$. This encodes width and shape information and is invariant under translation.
• Location Component (Odd): $C_X(u) = \frac{1}{2}(h_X(u) - h_X(-u))$. This encodes location information.

Key Mathematical Property: In the Hilbert space $L^2(S^{d-1}, \sigma)$ (where $\sigma$ is the normalized surface measure), the even and odd subspaces are exactly orthogonal. This induces an additive variance-covariance structure:
$$\text{Var}_{\text{tot}}(X) = \text{Var}_{\text{size}}(X) + \text{Var}_{\text{loc}}(X)$$
There are no cross-terms between size and location.

B. Dependence Measures

The author defines component-wise covariance and correlation indices:

• Covariance: $\text{Cov}_{\bullet}(X, Y) = \int_{S^{d-1}} \text{Cov}(\Phi_{\bullet}^X(u), \Phi_{\bullet}^Y(u)) d\sigma(u)$, where $\bullet \in \{\text{size}, \text{loc}, \text{tot}\}$.
• Steiner-Centered Residual: To isolate higher-order location fluctuations beyond simple translation, the paper introduces a residual location component $C^{\text{res}}_X(u) = C_X(u) - \langle s(X), u \rangle$, where $s(X)$ is the Steiner point. This allows detection of directional dependence invisible to the Steiner point.

C. Mixing and Asymptotics

• $\rho$-Mixing Coefficients: A new compatible $\rho$-mixing coefficient ($\rho_{\max}$) is defined for set-valued processes. It is the supremum of the maximal correlation (Hirschfeld-Gebelein-Rényi) between the support function projections (size or location) at lag $k$, taken over all directions $u$.
• Limit Theorems: Using the orthogonality of the decomposition and the $\rho_{\max}$ mixing condition, the paper establishes:
  • Weak and Strong Laws of Large Numbers (WLLN/SLLN): Convergence of Minkowski averages in the $L^2(\sigma)$ norm.
  • Central Limit Theorem (CLT): Convergence to a Gaussian process in the Hilbert space $H = L^2(S^{d-1}, \sigma)$.
  • Functional CLT (FCLT): Convergence to an $H$-valued Brownian motion.
  • Law of the Iterated Logarithm (LIL): Characterization of the almost sure limit set.

3. Key Contributions

1. Variational Framework: Introduction of a rigorous size–location decomposition for random sets that yields an intrinsic, additive variance structure impossible to recover from point-based representations.
2. Non-Degenerate Dependence Measures: Definition of size, location, and total correlation indices that remain informative even for centrally symmetric sets or when the Steiner point is uninformative.
3. Compatible Mixing Theory: Development of $\rho_{\max}$-mixing coefficients tailored to set-valued processes, which are geometrically interpretable and strictly weaker (more tractable) than classical $\sigma$-algebra mixing coefficients.
4. Asymptotic Stability: Proof of WLLN, SLLN, CLT, FCLT, and LIL for weakly stationary $\rho$-mixing set-valued sequences in the $L^2(\sigma)$ support-function norm.
5. Numerical Realization: Demonstration that these quantities can be estimated via directional Monte Carlo and spherical designs, with proven consistency and convergence rates.

4. Results

• Theoretical Results:
  • The decomposition ensures that size and location fluctuations are orthogonal, allowing for separate analysis.
  • The total covariance is the sum of size and location covariances without cross-terms.
  • Under the summability condition $\sum \rho_{\max}(k) < \infty$, the Minkowski average converges almost surely to the Aumann expectation.
  • The limiting distribution of the normalized partial sums is a Gaussian process with a block-diagonal covariance operator (separating even and odd subspaces).
• Simulation Studies:
  • Scenario S1 (Sign Flip): Correctly identifies perfect size correlation ($+1$) and perfect negative location correlation ($-1$).
  • Scenario S2 (Shared Center, Independent Shape): The Steiner-point correlation is near 1 (false positive for dependence), while the proposed residual location correlation correctly identifies low dependence.
  • Scenario S3 (Shared Shape, Independent Center): The Steiner-point correlation is near 0 (false negative), while the proposed framework detects strong size dependence and significant residual location dependence.
  • Scenario S4 (Sensitivity): Demonstrates that the proposed residual location index captures directional dependence that varies continuously with a mixing parameter, whereas the Steiner point remains uninformative (correlation $\approx 0$).

5. Significance

This work fundamentally advances the statistical theory of random sets by:

• Resolving Structural Limitations: It eliminates the degeneracy issues inherent in classical approaches where size and location are conflated.
• Geometric Interpretability: The dependence measures are directly tied to the geometry of the sets (width vs. translation) rather than abstract scalar summaries.
• Enabling New Applications: The framework provides the theoretical foundation for:
  • Interval-valued regression: Showing that midpoint-radius regression is a special case of support-based risk minimization.
  • Robust Optimization: Separating systematic shifts (location) from uncertainty inflation (size) in robust linear programming.
  • Stochastic Control: Decoupling location tracking and size allocation in decision-making under uncertainty.
• Bridging Geometry and Probability: It establishes that dependence in set-valued processes is not merely a scalar extension but possesses unique probabilistic structures (e.g., block-diagonal long-run covariance) driven by the underlying geometry.

In conclusion, Tuyen provides a unified, geometrically sound, and asymptotically rigorous framework for analyzing dependent set-valued data, offering tools that are both theoretically robust and practically estimable.