Estimating copula measure using ranks and subsampling: a simulation study

We describe here a new method to estimate copula measure. From N observations of two variables X and Y, we draw a huge number m of subsamples (size n<N), and we compute the joint ranks in these subsamples. Then, for each bivariate rank (p,q) (0<p,q<n+1), we count the number of subsamples such that there exist an observation of the subsample with bivariate rank (p,q). This counting gives an estimate of the density of the copula. The simulation study shows that this method seems to gives a better than the usual kernel method. The main advantage of this new method is then we do not need to choose and justify the kernel. In exchange, we have to choose a subsample size: this is in fact a problem very similar to the bandwidth choice. We have then reduced the overall difficulty.

💡 Research Summary

The paper proposes a novel non‑parametric method for estimating the copula density of two random variables based on ranks and subsampling. Given a full sample of size N, the authors repeatedly draw m subsamples of size n (< N) (with replacement). Within each subsample s, the ranks of the observations in the X‑ and Y‑coordinates are computed, denoted R_{i,s} and S_{i,s} for the i‑th observation of that subsample. For every possible rank pair (p,q) with 1 ≤ p,q ≤ n, the indicator 1_{R_{i,s}=p, S_{i,s}=q} is summed over all i and s, and the total is divided by m·n. The resulting quantity is taken as an estimate of the copula density at the point (p/n, q/n). In effect the method fills the discrete grid of possible rank pairs by repeatedly sampling; with a sufficiently large number of subsamples the grid becomes densely populated, providing a discrete approximation of the underlying continuous copula density.

The authors discuss two levels of theoretical convergence. First, for a fixed subsample, the empirical rank measure β̂ converges weakly to the true distribution of (F_X(X), F_Y(Y)), exactly as the ordinary empirical distribution does. Second, if the subsample size n satisfies n²/N → 0 as N → ∞, the probability that a subsample contains duplicated original observations tends to zero, making the subsampling with or without replacement asymptotically equivalent. Consequently, the average of the β̂ measures over all subsamples, denoted γ̂, also converges weakly to the true copula distribution. Convergence of the density itself (i.e., pointwise or in L¹) is not proved; the authors acknowledge that establishing density convergence on a finite support is technically demanding.

A practical difficulty is the presence of ties in the original data, which makes ranks ambiguous. The authors’ pragmatic solution is to discard any subsample that contains a tie; this is feasible only when ties are rare, which aligns with the assumption of continuous margins underlying rank‑based tests.

The simulation study has two main parts. In the first part, the authors evaluate goodness‑of‑fit testing. They generate samples from mixtures of an F‑rank copula (with parameter θ = 1, 2, 3) and a Student‑t copula with 4 degrees of freedom and correlation 0.95. They also consider Gaussian copulas with correlations 0.17, 0.32, 0.47. For each generated sample they estimate the copula density using the proposed rank‑subsample method and a conventional kernel estimator, then compute the Kullback‑Leibler divergence between the estimated discrete density and a reference density (obtained by a very large simulated sample). Using these divergences they construct test statistics and empirical critical values. The results (Table 1) show that, when the subsample size n is optimally chosen, the rank‑subsample test attains higher power than the kernel‑based test across all scenarios, with the gap narrowing as N increases.

The second part concerns independence testing under eight different dependence structures: linear (y = a x + ε), quadratic (y = a x² + ε), a circular “donut” pattern (x, y) = a (cos 2πu, sin 2πu) + (ε₁, ε₂), and a volatility‑modulated “butterfly” pattern (y = (1 + a|x|) ε). For each structure they consider sample sizes N = 30 and N = 300, and compare three tests: (i) the new rank‑subsample test, (ii) Deheuvels’ Cramér‑von Mises test, and (iii) a “smart” test that exploits knowledge of the functional form (e.g., Pearson correlation for linear dependence). When the optimal subsample size is known (chosen by exhaustive search), the new test consistently outperforms Deheuvels’ test and often rivals the smart test, especially for non‑monotonic dependencies such as the donut and butterfly cases. Because the optimal n is unknown in practice, the authors propose a minimax‑regret policy: choose n that minimizes the worst‑case loss of power across all dependence types. This leads to n = 8 for N = 30 and n = 10 for N = 300. Under this policy, the new test still retains respectable power, though the advantage over Deheuvels’ test diminishes.

Overall, the paper argues that rank‑based subsampling eliminates the need to select a kernel function, reducing the practitioner’s burden to a single tuning parameter (the subsample size) that plays a role analogous to the bandwidth in kernel methods. The method is computationally straightforward, relies only on order statistics, and appears to provide comparable or superior power in a variety of settings. However, limitations remain: (1) the choice of subsample size is still critical and lacks a fully data‑driven solution; (2) the method yields a discrete approximation of the copula density, which may be insufficient when a smooth estimate is required; (3) handling ties robustly is an open problem; and (4) rigorous asymptotic results for the density estimator are still pending.

The authors conclude by noting that a theoretical paper establishing density convergence is in preparation, and they outline future research directions: deriving the limiting distribution (or large‑deviation behavior) of the test statistic, developing a principled minimax strategy for subsample‑size selection, and extending the approach to higher dimensions. The accompanying R and C code is available upon request.