Consistent Weightings in Empirical Testing

• Consistent weightings of empirical measures are defined by bounded, integrable functions that emphasize discrepancies between empirical and theoretical distributions, ensuring universal test sensitivity.
• Tailored weights such as median, tail, or power-type adjust the methodology to optimize detection of various dependency structures, significantly improving test power.
• Closed-form rank-based formulas and robust asymptotic properties enable efficient computation and reliable statistical inference in copula-based non-parametric testing.

Consistent weightings of empirical measures are a foundational concept in modern statistical hypothesis testing, particularly in non-parametric inference for dependency structure. In the context of the empirical copula process, such weightings refer to the strategic use of bounded, integrable functions to “emphasize” discrepancies between empirical and theoretical distributions in selected portions of the sample space. By appropriately tailoring these weights, one can design functionals that are both consistent (i.e., leading to tests that are universally sensitive to all departures from the null) and optimally sensitive to specific forms of dependence. This approach is epitomized by the class of weighted Cramér‐von Mises statistics for independence testing.

1. Weighted Cramér‐von Mises Functionals

The non-parametric framework introduced leverages distances between the empirical copula $C_n(u)$ and the independence copula $C^{\perp}(u) = \prod_{j=1}^d u_j$ through the weighted Cramér‐von Mises functional: $$W_n = n \int_{[0,1]^d} [C_n(u) - \prod_{j=1}^d u_j]^2 w(u) du$$ where $w(u)$ is any bounded, integrable weight function on $[0,1]^d$.

Key properties of these functionals include:

• Region-Specific Emphasis: $w(u)$ enables targeting regions such as central (median) areas or tails.
• Adaptivity to Alternatives: Modifying $w(u)$ allows construction of tests tuned for detecting particular dependency structures (e.g., tail, central).
• Consistency: For any such $w(u)$, the resulting tests are consistent against all alternatives; that is, the statistics diverge in probability under any fixed alternative to independence.

2. Role and Selection of Weight Functions

The choice of $w(u)$ is pivotal for the statistical power and sensitivity of the test:

• Uniform Weight: $w(u) = 1$, yielding the classical Cramér‐von Mises statistic, gives equal weight to all parts of the distribution.
• Median Weight: $w_m(u) = \prod_{j=1}^d [u_j(1-u_j)]$, enhances sensitivity to central deviations, where many copula alternatives (like Gaussian and Frank) often peak.
• Tail Weights: Upper-tail $w_p(u) = \prod_{j=1}^d u_j^2$, lower-tail $w_l(u) = \prod_{j=1}^d (1-u_j)^2$, focus the test on extreme co-movements.
• Power-Type Weights: $w(u) = \prod_{j=1}^d u_j^{2\beta_j}$ lets exponents $\beta_j$ modulate regional focus. The optimal power—the largest gain in power—is theoretically achieved when $w(u)$ is proportional to the squared true deviation from independence.

Thus, the “tuning” of the weight function controls the test’s efficacy for types of dependence of practical interest.

3. Asymptotic Properties of Weighted Statistics

Under standard conditions (continuity of copula partial derivatives; $w(u)$ bounded, integrable), the weighted CvM statistic obeys: $\sqrt{n}[C_n(u) - C(u)] \rightarrow \mathcal{M}(u)$ in distribution, where $\mathcal{M}(u)$ is a centered Gaussian process. The continuous mapping theorem yields

$$W_n \xrightarrow{d} W = \int_{[0,1]^d} \mathcal{M}(u)^2 w(u) du$$

The limiting law is pivotal; it does not depend on nuisance parameters from the marginals and is determined only by $w(u)$ and the dimension. This feature enables tabulation of asymptotic critical values for each $w(u)$. Furthermore, this weighting does not compromise the consistency or validity of the test.

4. Computation: Closed-Form Expressions and Practicality

The development includes explicit rank-based formulas for fast implementation:

• Define
  • $$\mu_1(w)(a) = \int_{a_1}^1 ... \int_{a_d}^1 w(u_1, ..., u_d) du_1 ... du_d$$
  • $$\mu_2(w)(a) = \int_{a_1}^1 ... \int_{a_d}^1 (u_1...u_d) w(u_1,...,u_d) du_1...du_d$$
  • $$\mu_3(w) = \int_{0}^1 ... \int_{0}^1 (u_1^2 ... u_d^2) w(u_1,...,u_d) du_1...du_d$$
• The statistic admits a formula: $$W_n = \sum_{i=1}^n \left[ \frac{1}{n} \sum_{l=1}^n \mu_1(w)(\hat U_i \vee \hat U_l) - 2\mu_2(w)(\hat U_i) \right] + n \mu_3(w)$$ where $\hat U_i$ are pseudo-observations of percentile ranks.

For power-type weights, closed forms use simple combinations of powers and ranks. Analogous formulas exist for uniform, median, tail, and symmetric tail weightings. These closed formulas facilitate efficient, scalable computation—especially critical for permutation procedures in dependence testing.

5. Simulation Results: Power and Sensitivity

Extensive bivariate simulations (across Gaussian, $t$, Frank, Gumbel, Clayton copulas) reveal:

• Uniform weights perform well only for non-local dependencies.
• For tail dependence, median- or tail-weighted statistics substantially increase power—by up to a factor of four relative to poor choices.
• Symmetric tail weights (balancing upper and lower tails) yield robust power across various alternatives.
• Overall, maximal power gains are realized when the weighting is proportional to the actual deviation from independence.

This empirically validates the principle that, for greatest effectiveness, weights should reflect the alternative dependence structure targeted by the test.

6. Recommendations and Practical Application

Implementational recommendations include:

• Leverage the “margin-free” property: the asymptotics depend only on the copula and weight, allowing for robust application.
• Choose $w(u)$ according to anticipated alternative: median weights for central dependencies; tail weights for financial or risk data where extremes matter.
• Classical Anderson–Darling weights (variance inverses) may lead to integrability issues; such cases require correction or offset in weighting.
• Consider data-driven or adaptive weight selection: preliminary estimation of deviations from independence may enable construction of nearly-optimal, test-specific $w(u)$.
• Closed-form rank-based formulas support practical use in high-dimensional or resampling-intensive procedures.

7. Implications and Future Directions

Weighted empirical measures as introduced via the weighted CvM functional constitute a unified, flexible framework for sensitive independence testing. The structure is sufficiently general to allow adaptation to user-specified alternatives, is robust (with proven consistency and pivotal asymptotics), and is computationally efficient through explicit formulas. Evidence indicates that careful tuning of weights yields substantial practical power gains, with greatest improvement achieved by matching the weighting to the anticipated dependency structure.

A plausible implication is that future data-driven approaches to constructing or learning effective weights could further enhance sensitivity and precision of copula-based inference while retaining all the desirable theoretical properties established for these weighted statistics.

---

This synthesis integrates the analytical, computational, and practical aspects of consistent weightings of empirical measures, contextualized in the empirical copula process and weighted Cramér-von Mises testing, with a focus on the role of the weight function, statistical properties, computation, and applied recommendations.