📊 statistics

The Spatial Cram'{e}r--von Mises Test of Independence under $\beta$ -Mixing: Asymptotic Theory and Python Implementation

This paper establishes the asymptotic theory for a spatial Cramér–von Mises test of independence under polynomial $\beta$ -mixing dependence in stationary random fields and provides a Python implementation that demonstrates the test's superior power, particularly with Anderson–Darling weights, compared to traditional methods like Mantel and cross- $K$ tests.

Original authors: Marco Mandap

Published 2026-05-20

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Marco Mandap

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 a detective trying to solve a mystery in a crowded city. Your job is to figure out if two groups of people, let's call them "Team X" and "Team Y," are moving completely independently of each other, or if they are secretly coordinating their steps.

In a normal city where everyone walks randomly and doesn't talk to their neighbors (what statisticians call "independent data"), it's easy to check this. You just look at the crowd and see if the patterns match up.

But in the real world, people often influence their neighbors. If one person stops to tie a shoe, the person next to them might stop too. If a whole neighborhood moves in a wave, everyone in that block moves together. This is what the paper calls spatial dependence. When data points are "neighbors," standard detective tools often get confused and think there's a secret connection when there isn't one, or they miss a real connection because the "noise" of the neighborhood is too loud.

This paper introduces a new, super-smart detective tool called the Spatial Cramér–von Mises Test. Here is how it works, explained simply:

1. The Problem: The "Noisy Neighborhood"

The authors are looking at data spread out on a map (like temperature readings or stock prices in different cities). They want to know: Is the pattern of X related to the pattern of Y?
The catch is that the data is "sticky." A value at one spot is likely similar to the value right next to it. This "stickiness" (called $\beta$ -mixing) breaks the rules of the old, classic tests. If you use the old tests on this sticky data, you might get false alarms.

2. The Solution: The "Inner-Form" Trick

Instead of trying to map the entire city's movement at once (which is mathematically messy and requires very strict rules), the authors use a clever shortcut. They break the problem down into two smaller, simpler puzzles.

The Analogy: Imagine you have a giant, tangled ball of yarn representing the relationship between Team X and Team Y. Instead of trying to untangle the whole ball at once, they cut the yarn into two separate strands: one for Team X and one for Team Y.
The Math: They reformulate the test so that it treats the two teams separately first, then multiplies the results. This is called a product kernel. Because they separated the strands, they can use a well-known mathematical shortcut (called a U-statistic) that ignores the messy "neighborhood noise" and focuses only on the core relationship.

3. The "Magic Formula" (The Limit Distribution)

Once they separate the strands, they need to know what a "normal" result looks like to decide if the teams are actually connected.

In the old days, you'd need a super-complex computer simulation to guess this.
The authors found a magic formula (an eigenvalue factorization). It turns out that the "score" for the whole city is just the product of the "scores" for Team X and Team Y.
This allows them to calculate a critical value (a threshold score) using a simple weighted sum of random numbers (specifically, a mix of $\chi^2$ variables). If your test score is higher than this threshold, you can confidently say, "Yes, Team X and Team Y are coordinating!"

4. The "Weight" Choice: The Magnifying Glass

The test uses a "weight function" to decide how much attention to pay to different parts of the data. The authors tested three different "magnifying glasses":

Uniform: Looks at everything equally.
Optimal Normal: Tries to be the most efficient but turns out to be too sensitive to noise, making it hard to detect weak connections.
Anderson–Darling: This is the paper's star player. It acts like a magnifying glass that zooms in on the edges and tails of the data. The simulations showed this one is the best at spotting secret connections, even when the data is very "sticky" or the connection is weak.

5. What They Found (The Results)

The authors wrote a Python program to test their new detective tool against the old ones.

The Old Tools (Mantel, Cross-K): These tools are like detectives who only look at the distance between people. They got fooled by the "neighborhood noise." They couldn't tell if the teams were actually coordinating or just living in the same neighborhood. They had almost zero power to find the truth in these complex, sticky situations.
The New Tool (Spatial CvM): This tool successfully ignored the neighborhood noise and found the secret connections. It worked perfectly even when the data was very dependent on its neighbors.
The Winner: The Anderson–Darling version of the new test was the most powerful, finding connections that the others missed.

Summary

The paper says: "We built a new way to test if two things are related when they are living in a crowded, sticky neighborhood. We found a mathematical trick to separate the two problems, derived a simple formula for the results, and proved that our new tool is much better than the old ones at finding the truth without getting confused by the crowd."

They also provided the Python code (the "blueprint") so anyone can use this new detective tool immediately.

Technical Summary: The Spatial Cram´er–von Mises Test of Independence under $\beta$ -Mixing

Problem Statement
The paper addresses the challenge of testing bivariate independence in stationary random fields on $\mathbb{R}^2$ . While the classical Cram´er–von Mises (CvM) test is well-established for independent and identically distributed (i.i.d.) observations, its application to spatially dependent data is non-trivial due to the complex dependence structure. The null hypothesis $H_0$ posits that the marginal components $X_s$ and $Y_s$ of a bivariate random field $Z_s = (X_s, Y_s)$ are independent, implying the joint cumulative distribution function (CDF) factors as $H(x, y) = F_1(x)F_2(y)$ . The authors aim to derive the asymptotic distribution of the spatial CvM statistic under polynomial $\beta$ -mixing dependence and provide a reproducible Python implementation.

Methodology
The authors propose a theoretical framework and computational pipeline that bypasses the need for a functional central limit theorem (FCLT). The methodology rests on three primary theoretical ingredients:

Davydov-type Covariance Bounds: The authors establish integrability of the spatial covariance kernel under a polynomial $\beta$ -mixing condition where the decay rate $\theta$ satisfies $\theta > 2(2 + \delta)/\delta$ . This ensures the spatial covariance is absolutely summable, a prerequisite for limit theorems.
U-Statistic Reformulation: Following De Wet (1980), the spatial CvM statistic is reformulated as a degenerate U-statistic of order 2 with a product kernel $Q = G_1 \otimes G_2$ . This "inner-form" approach utilizes empirical marginal CDFs, allowing the two-dimensional spectral problem to be reduced to two one-dimensional problems.
Limit Theorem Extension: By combining Gregory's (1977) limit theorem for degenerate U-statistics with Yoshihara's (1976) extension to $\beta$ -mixing sequences, the authors derive the asymptotic distribution without proving weak convergence in function spaces (e.g., $C([0, 1])$ ).

Key Contributions

Asymptotic Theory: The paper derives the limit distribution of the centered and normalized spatial CvM statistic. Under the null hypothesis, the statistic converges to a weighted sum of correlated $\chi^2_1$ variables. Crucially, the eigenvalues of the limiting operator factorize as products of marginal eigenvalues ( $\gamma_{jk} = \lambda^{(1)}_j \cdot \lambda^{(2)}_k$ ).
Small-Bandwidth Limit: In the limit where the kernel bandwidth $h \to 0$ , the correlation between the limiting variables vanishes, and the distribution reduces to a weighted sum of independent $\chi^2_1$ variables, recovering the classical i.i.d. form.
Explicit Critical Values: The authors provide explicit formulas for eigenvalues corresponding to three common weight functions: Uniform, Optimal Normal, and Anderson–Darling. These allow for the computation of critical values without numerical integration.
Software Implementation: A complete Python package is documented, featuring:
- Generation of Mat´ern random fields via circulant embedding.
- Computation of the CvM statistic using inner-form kernel decomposition.
- Evaluation of asymptotic critical values via Monte Carlo simulation of the limit distribution.
- Permutation-based inference for finite-sample p-values.
- Implementation of competing tests (Mantel, cross-K, distance covariance) for power comparison.

Results
Simulation experiments conducted on $m \times m$ lattices yield the following findings:

Size Calibration: The test maintains acceptable size (rejection rates within 1–2 percentage points of nominal levels) for sample sizes $n \ge 400$ under moderate to strong mixing. At $n=100$ , the test is slightly oversized under weak mixing ( $\theta=3$ ).
Power Performance:
- The Anderson–Darling weight function consistently achieves the highest power across Gaussian and t-copula alternatives, particularly at weak dependence levels.
- The Optimal Normal weight exhibits near-zero power at weak alternatives due to a divergent centering constant that inflates critical values.
- The Uniform weight performs competitively but is generally outperformed by Anderson–Darling.
Comparison with Competing Tests:
- The Mantel and Cross-K tests demonstrate no power against cross-dependence in spatially correlated fields, failing to distinguish spatial autocorrelation from true cross-dependence.
- The Distance Covariance test performs well but is generally outperformed by the Anderson–Darling CvM test at weak alternatives.
Robustness: The test size and power are stable across variations in the Mat´ern effective range parameter, confirming the robustness of the small-bandwidth limit implementation.

Significance and Claims
The paper claims that its primary significance lies in providing a rigorous asymptotic justification for using the CvM test on spatially dependent data without resorting to the more demanding functional central limit theorem. By utilizing the U-statistic route, the authors avoid tightness arguments (Arzel´a–Ascoli) and bracketing entropy calculations, requiring only scalar convergence.

The authors position the spatial CvM test as a robust tool for detecting cross-dependence in random fields where traditional spatial tests (Mantel, Cross-K) fail. The work serves as a "robustness result," justifying the use of standard i.i.d. critical values for spatially dependent data under the small-bandwidth limit, provided the mixing conditions are met. The paper does not claim to solve all spatial independence problems but offers a specific, theoretically grounded solution for bivariate independence testing under $\beta$ -mixing, accompanied by a verified software implementation for immediate application.