--- title: Empirical Bernstein Copula url: https://www.emergentmind.com/topics/empirical-bernstein-copula type: topic --- # Empirical Bernstein Copula Searching arXiv for the cited papers and closely related work on empirical Bernstein copulas. The empirical Bernstein copula is a smooth nonparametric copula estimator obtained by applying multivariate Bernstein polynomials to the empirical copula on a regular grid of \([0,1]^d\). In the formulation used by Segers, Sibuya, and Tsukahara, if \(C_n\) denotes the usual empirical copula and \(m=(m_1,\dots,m_d)\in\mathbb N^d\) is a multi-degree, then \[ \widehat C_{n,m}(u) :=B_m(C_n)(u) = \sum_{s_1=0}^{m_1}\cdots\sum_{s_d=0}^{m_d} C_n\!\Bigl(\tfrac{s_1}{m_1},\dots,\tfrac{s_d}{m_d}\Bigr) \prod_{j=1}^d \binom{m_j}{s_j}u_j^{\,s_j}(1-u_j)^{m_j-s_j}, \qquad u\in[0,1]^d. \] Equivalently, it is the empirical copula averaged against a product-Binomial smoothing law, which makes the estimator polynomial, smooth, and amenable to derivative-based inference [1607.04430]. In the bivariate notation used for symmetry testing, \[ C_{n,m}(u,v) = \sum_{k=0}^m\sum_{\ell=0}^m \widehat C_n(k/m,\ell/m)\, P_{m,k}(u)\,P_{m,\ell}(v), \] where \(P_{m,k}(u)=\binom mku^k(1-u)^{m-k}\) [2202.12787]. Across the literature, the estimator functions both as a direct smooth substitute for the stepwise empirical copula and as a building block for resampling, hypothesis testing, conditional distribution estimation, vine copulas, and semiparametric Bernstein-copula models [2301.05495]. ## 1. Definition, rank construction, and copula structure The starting point is the rank-based empirical copula. For i.i.d. observations \(X_1,\dots,X_n\in\mathbb R^d\) with continuous margins, let \(R_{i,j}^{(n)}\) be the rank of \(X_{i,j}\) among \(X_{1,j},\dots,X_{n,j}\), and define \[ C_n(u) = \frac1n\sum_{i=1}^n\prod_{j=1}^d \mathbf1\bigl\{R_{i,j}^{(n)}/n\le u_j\bigr\}. \] The empirical Bernstein copula replaces these hard indicators by Bernstein basis weights indexed on a finite grid [1607.04430]. In bivariate form, with pseudo-observations \[ \widehat U_i=n^{-1}\sum_{j=1}^n\mathbf1\{X_j\le X_i\}, \qquad \widehat V_i=n^{-1}\sum_{j=1}^n\mathbf1\{Y_j\le Y_i\}, \] the usual empirical copula is \[ \widehat C_n(u,v)=\tfrac1n\sum_{i=1}^n\mathbf1\{\widehat U_i\le u,\;\widehat V_i\le v\}, \] and the empirical Bernstein copula of order \(m\) is the Bernstein smoothing of \(\widehat C_n\) on the grid \(\{0,1/m,\dots,1\}^2\) [2202.12787]. A central structural issue is whether Bernstein smoothing preserves the copula property. Segers, Sibuya, and Tsukahara give necessary and sufficient coefficient conditions for a Bernstein polynomial \(B_m(a)\) to be a copula. If the coefficient array satisfies groundedness, uniform-margin constraints, and nonnegativity of the full forward difference \(\Delta_1\cdots\Delta_d a\), then the Bernstein polynomial is a copula; groundedness and marginal conditions are also necessary [1607.04430]. This result explains why some Bernstein smoothers are genuine copulas and why others are not. A related formulation replaces direct smoothing of \(C_n\) by smoothing of the empirical checkerboard copula \(C_n^\#\). Lu and Ghosh define the multivariate empirical checkerboard Bernstein copula (ECBC) \[ C_{m,n}^\#(u_1,\dots,u_d) = \sum_{k_1=0}^{m_1}\cdots\sum_{k_d=0}^{m_d} \tilde\theta_{k_1,\dots,k_d} \prod_{j=1}^d \binom{m_j}{k_j}u_j^{k_j}(1-u_j)^{m_j-k_j}, \] with \(\tilde\theta_{k_1,\dots,k_d}=C_n^\#(k_1/m_1,\dots,k_d/m_d)\). Because \(C_n^\#\) is a copula and the Bernstein basis preserves the copula properties for any \(m_j\), \(C^\#_{m,n}\) is itself a genuine copula [2112.10351]. ## 2. Relation to the empirical beta copula and other smooth copula estimators The most important special case is obtained by setting all Bernstein degrees equal to the sample size. Segers, Sibuya, and Tsukahara show that when \(m_1=\cdots=m_d=n\), the empirical Bernstein copula becomes the empirical beta copula, \[ C_n^\beta(u)=B_{(n,\dots,n)}(C_n)(u), \] which is therefore a particular case of the empirical Bernstein copula [1607.04430]. In the equivalent representation, \[ C_n^\beta(u) = \frac1n\sum_{i=1}^n\prod_{j=1}^d F_{n,R_{i,j}^{(n)}}(u_j), \] each coordinate is smoothed by the c.d.f. of a \(\mathrm{Beta}(r,n+1-r)\) kernel [1607.04430]. The same relationship is reiterated in broader classes of smooth estimators: the empirical beta copula is the special case obtained when the smoothing law is product-Binomial with degree \(n\) [2106.10726]. The literature distinguishes the empirical Bernstein copula from several neighboring constructions. The empirical copula is piecewise constant and not differentiable [2202.12787]. The checkerboard copula \(C_n^\#\) is a multilinear extension and is a genuine copula even for finite \(n\) [2112.10351]. ECBC smooths \(C_n^\#\) rather than \(C_n\), and Lu and Ghosh use this to obtain a genuine copula estimator with data-dependent degrees [2112.10351]. A further distinction concerns the parametric or semiparametric “Bernstein copula” of Dou et al. Their Bernstein copula is a finite-mixture copula \[ C_{\bm m}(u_1,\dots,u_d;\bm p) = \sum_{k_1=0}^{m_1}\cdots\sum_{k_d=0}^{m_d} p_{k_1,\dots,k_d} \prod_{j=1}^d B_{k_j,m_j}(u_j), \] where the weights satisfy marginal-uniformity constraints [1301.2677]. This is not the same object as the empirical Bernstein copula, although the empirical estimator can be viewed as a specific rank-based smoothing device within the same Bernstein basis. The main related estimators are summarized below. | Estimator | Construction | Relation to empirical Bernstein copula | |---|---|---| | Empirical copula \(C_n\) | Rank-based step function | Unsmoothed baseline [1607.04430] | | Empirical Bernstein copula \(\widehat C_{n,m}\) | Bernstein smoothing of \(C_n\) | General smooth polynomial estimator [1607.04430] | | Empirical beta copula \(C_n^\beta\) | Case \(m_j=n\) | Special case of empirical Bernstein copula [1607.04430] | | ECBC \(C_{m,n}^\#\) | Bernstein smoothing of checkerboard copula \(C_n^\#\) | Closely related genuine copula estimator [2112.10351] | This suggests a useful conceptual distinction: empirical Bernstein copulas are primarily defined by how the rank-based copula is smoothed, whereas the empirical beta copula and ECBC are particular structurally advantageous instances within the same Bernstein-polynomial ecosystem. ## 3. Large-sample properties, bias, variance, and smoothing rates A basic result is that the empirical Bernstein copula is asymptotically equivalent to the empirical copula process under standard smoothness assumptions. Segers, Sibuya, and Tsukahara prove that if the minimum degree \(m_*=\min_j m_j\to\infty\) and \(\liminf m_*/n>0\), then \[ G_{n,m} = \sqrt n\{\widehat C_{n,m}-C\} = G_n+o_P(1) \quad\Longrightarrow\quad G \quad\text{in }\ell^\infty([0,1]^d), \] where \(G_n=\sqrt n(C_n-C)\) and \(G\) is the same tight centered Gaussian limit as for the empirical copula [1607.04430]. The proof decomposes the process into a stochastic equicontinuity term and a smoothing bias term [1607.04430]. For the bivariate setting studied in symmetry testing, the empirical Bernstein copula process \[ B_{n,m}(u,v)=\sqrt n\{C_{n,m}(u,v)-C(u,v)\} \] converges in \(\ell^\infty([0,1]^2)\) to \[ B_C(u,v) = \mathbb C(u,v) -\dot C_1(u,v)\,\mathbb C(u,1) -\dot C_2(u,v)\,\mathbb C(1,v), \] where \(\mathbb C\) is the \(C\)-Brownian bridge with covariance \[ \mathrm{Cov}\{\mathbb C(u,v),\mathbb C(s,t)\} = C(u\wedge s,v\wedge t)-C(u,v)C(s,t), \] provided \(C\) has continuous first-order partials and \(m=c\,n^\alpha\) with \(c>0\) and \(\alpha\ge1\); the same source notes that \(\alpha>3/4\) suffices under stronger second-derivative control [2202.12787]. Uniform consistency is established under mild smoothness. In the symmetry-testing paper, if \(m=c\,n^\alpha\) with \(c>0\) and \(\alpha>3/4\), then \[ \sup_{(u,v)\in[0,1]^2}|C_{n,m}(u,v)-C(u,v)|\to0 \] almost surely and in probability [2202.12787]. More precisely, \[ \sup_{(u,v)}|C_{n,m}(u,v)-C_n(u,v)| = O\!\bigl(n^{-1/2}(\log\log n)^{1/2}\bigr), \] while \[ \sup_{(u,v)}|E[C_{n,m}(u,v)]-C(u,v)| = O(m^{-1/2}), \] so that choosing \(\alpha>3/4\) makes the overall error vanish [2202.12787]. Other sources state slightly different rate prescriptions because they work with different loss criteria or asymptotic regimes. Ouimet and Susam report the conditions \[ \frac{n}{m\log\log n}\to c\in[0,\infty) \] for uniform strong consistency and \[ \sqrt n\,m^{-1}\to0 \] for the same limit distribution as the empirical copula. They also state that balancing \(O(m^{-1})\) bias against variance reduction of order \(n^{-1}m^{-1/2}\) leads to the rule-of-thumb \(m\asymp n^{2/3}\), implemented in practice as \(m=\lfloor n^{2/3}\rfloor\) [2506.08857]. In contrast, the two-sample testing paper states that, under mild regularity and continuous bounded first partial derivatives, the pointwise mean-squared error is \[ O\bigl(m^{-2}+1/(nm)\bigr), \] which yields the familiar rule \(m\asymp n^{1/3}\) for minimizing an integrated MSE [2303.02510]. The vine-copula paper, working with a contingency-table version of the estimator, reports bias \(O(1/M)\) and stochastic error \(O_p(\sqrt{(M+1)^2/n})\), motivating growth such as \(M\sim n^{1/3}\) [1210.2043]. These differing prescriptions are not contradictory on their face; they arise from different estimands, approximations, and risk criteria. A plausible implication is that “the” optimal Bernstein degree is context-dependent: testing, tail functionals, generic copula estimation, and checkerboard-based constructions need not select the same rate. ## 4. Smoothness, derivatives, and resampling theory A defining advantage of the empirical Bernstein copula is differentiability. Since \(C_{n,m}\) is a polynomial in the arguments, it is infinitely differentiable in the interior [2202.12787]. Under \(m=c\,n^\alpha\) with \(\alpha>3/4\), one has, uniformly for \(u\in[b_n,1-b_n]\), \(v\in[0,1]\), and \(b_n\to0\) slowly, \[ \left|\frac{\partial C_{n,m}}{\partial u}(u,v)-\frac{\partial C}{\partial u}(u,v)\right| = O\!\bigl(m^{1/2}n^{-1/2}(\log\log n)^{1/2}\bigr)\to0 \] [2202.12787]. This is crucial for multiplier bootstrap constructions, because the empirical copula itself is not differentiable [2202.12787]. The symmetry-testing paper develops a full multiplier bootstrap Bernstein process. With i.i.d. weights \(\xi_1^{(h)},\dots,\xi_n^{(h)}\) satisfying \(E= \mathrm{Var}=1\), one first constructs the Bernstein-smoothed bootstrap process \[ \overline B_{n,m}^{(h)}(u,v) = \sqrt n\,\frac1n\sum_{i,k,\ell} (\xi_i^{(h)}-\bar\xi^{(h)}) \mathbf1\{\widehat U_i\le k/m,\widehat V_i\le \ell/m\} P_{m,k}(u)P_{m,\ell}(v), \] and then corrects for estimated margins by subtracting the partial-derivative terms, \[ B_{n,m}^{(h)}(u,v) = \overline B_{n,m}^{(h)}(u,v) -\left(\frac{\partial C_{n,m}}{\partial u}\right)(u,v)\overline B_{n,m}^{(h)}(u,1) -\left(\frac{\partial C_{n,m}}{\partial v}\right)(u,v)\overline B_{n,m}^{(h)}(1,v). \] Jointly, \[ (B_{n,m},B_{n,m}^{(1)},\dots,B_{n,m}^{(H)}) \Rightarrow (B_C,B_C^{(1)},\dots,B_C^{(H)}) \] in \(\ell^\infty\) [2202.12787]. Kojadinović and collaborators place empirical Bernstein copulas inside a broader class of smooth, possibly data-adaptive empirical copulas \(C_{k:l}^\nu\). Under mild smoothness of the true copula and weak mixing, the smooth sequential process \(C_n^\nu\) differs from the classical sequential empirical copula process by \(o_P(1)\) uniformly, so both have the same weak limit [2106.10726]. The same program yields valid smooth resampling procedures. In the i.i.d. case, a smooth bootstrap based on drawing from the fitted smooth estimator is asymptotically valid [2301.05495]. In the time-series setting, a smooth extension of the sequential dependent multiplier bootstrap is also asymptotically valid [2301.05495]. The same line of work gives a Stute-type representation for empirical Bernstein copula processes. Under first-order smoothness, second-order control, and a variance condition on the smoothing law, one has \[ \sqrt n\{C_n^B(\mathbf u)-C(\mathbf u)\} = \widetilde C_n(\mathbf u)+R_n(\mathbf u), \] with an almost-sure remainder bound whose leading term is \(O(n^{-1/6})\) in the Bernstein case \(\gamma=1\) [2204.11240]. This places the empirical Bernstein copula on the same asymptotic footing as the classical empirical copula while retaining finite-sample smoothness. ## 5. Statistical testing and inferential uses A major inferential use of empirical Bernstein copulas is hypothesis testing based on smooth empirical processes. For bivariate symmetry, the process \[ S_{n,m}(u,v)=B_{n,m}(u,v)-B_{n,m}(v,u) \] is used to construct three statistics, and corresponding bootstrap versions \[ S_{n,m}^{(h)}(u,v)=B_{n,m}^{(h)}(u,v)-B_{n,m}^{(h)}(v,u) \] have the same weak limit [2202.12787]. The study reports that simulations under Gaussian, Clayton, Gumbel, and Frank copulas at various \(\tau\), with \(m\approx n^{4/5}\), show that Bernstein-based tests adhere more closely to nominal level, often under-rejection occurs with empirical-copula tests, and power is higher, especially for the sup-norm statistic [2202.12787]. For implementation, \(H\approx200\)–\(500\) multiplier replicates and a \(20\times20\) integration grid are reported as sufficient for the double-integral statistics [2202.12787]. The same methodology extends to two-sample equality testing. If \(C_{n_1,m_1}\) and \(D_{n_2,m_2}\) are empirical Bernstein copulas from two independent samples, then under \(\mathscr H_0:C\equiv D\), the process \[ F_{n,m}(\mathbf u) = \sqrt{1-\lambda}\,C_{n_1,m_1}(\mathbf u)-\sqrt{\lambda}\,D_{n_2,m_2}(\mathbf u) \] has a Gaussian limit, and three statistics are proposed: an \(L^2\) statistic \(R_{n,m}\), a weighted \(L^2\) statistic \(S_{n,m}\), and a sup-statistic \(T_{n,m}\) [2303.02510]. Under \(\mathscr H_1\) these statistics diverge to infinity, guaranteeing consistency [2303.02510]. The same work studies both multiplier bootstrap and a subsampling Bernstein version, reporting that Bernstein tests outperform tests based on the empirical copula [2303.02510]. Inference on functionals is another major theme. Since the smooth estimator shares the same first-order limit law as the empirical copula under standard conditions, Hadamard-differentiable functionals inherit asymptotic normality through the functional delta method [2204.11240]. This suggests why the empirical Bernstein copula has been used for dependence measures, confidence intervals, and change-point procedures in the later smooth-copula literature [2301.05495]. A recurring practical message is that smoothness is not only cosmetic. It permits direct plug-in estimation of partial derivatives and copula functionals, and it stabilizes resampling procedures that are awkward or unavailable for the nonsmooth empirical copula [2202.12787]. ## 6. Applications, extensions, and methodological variants The empirical Bernstein copula has been adapted well beyond unconditional copula estimation. In tail-focused concordance estimation, Ouimet and Susam define a Bernstein-smoothed lower-tail Spearman’s rho estimator using \(C^B_{n,m}\equiv C_{m,n}\) and establish strong consistency and asymptotic normality under mild regularity conditions [2506.08857]. Their pointwise expansions state \[ \mathrm{Bias}[C_{m,n}(u,v)] = b(u,v)/m + o(m^{-1}), \] \[ \mathrm{Var}[C_{m,n}(u,v)] = \sigma^2(u,v)/n - V(u,v)/(n\,m^{1/2}) + o(n^{-1}m^{-1/2}), \] so smoothing lowers variance at the cost of an \(O(m^{-1})\) bias [2506.08857]. For lower-tail Spearman’s rho on \([0,p]^2\), the reported Monte Carlo experiment with the Farlie–Gumbel–Morgenstern copula, \(n\in\{50,200\}\), \(p\in\{0.1,0.5,1.0\}\), \(m=\lfloor n^{2/3}\rfloor\), and \(K=10\,000\) replications shows MSE reductions up to approximately \(70\%\) in deep-tail settings at \(n=50\) under weak to moderate dependence [2506.08857]. In high-dimensional dependence modeling, Weiss and Scheffer use nonparametric Bernstein copulas as pair-copulas in C-vine and D-vine decompositions [1210.2043]. Their construction starts from pseudo-observations, computes grid counts, and solves a quadratic program to enforce uniform margins and nonnegativity, using the Goldfarb–Idnani dual QP-solver [1210.2043]. The resulting pair-copulas are smooth and nonparametric, and the algorithms of Aas et al. for density evaluation and sequential simulation carry over once the Bernstein c.d.f.s, densities, and \(h\)-functions are implemented [1210.2043]. In conditional copula estimation, Lu and Ghosh employ the empirical checkerboard Bernstein copula estimator to build a fully nonparametric estimator of conditional copulas without selecting a parametric family [2311.02808]. They derive closed-form estimators of conditional Kendall’s \(\tau\) and Spearman’s \(\rho\), and prove large-sample consistency under absolute continuity of the true three-dimensional copula and Lipschitz continuity of \(\partial C/\partial v\) [2311.02808]. Their algorithm first adjusts the univariate conditional margins through lower-dimensional ECBC fits, then constructs a three-dimensional ECBC and differentiates in the covariate direction [2311.02808]. A related 2026 development uses empirical checkerboard/Bernstein approximations to estimate conditional distributions via Sklar’s theorem and then to obtain consistent nonparametric estimators of mean, quantile, and expectile regression functions [2602.01144]. In that setting, with \(m(n)=\lfloor n^s\rfloor\), \(s\in(0,1/2)\), uniform conditional convergence of the estimated Markov kernel is established almost surely under the assumption that the true copula allows a continuous Markov kernel [2602.01144]. A distinct semiparametric direction is the finite-mixture Bernstein copula estimated by EM. Dou et al. propose EM algorithms for estimating the Bernstein copula weights \(\bm p\), prove local linear convergence of the constrained M-step adjustment, and establish \(\sqrt n\)-consistency and asymptotic normality of the resulting semiparametric estimator \(\widehat{\bm p}\) under regularity conditions [1301.2677]. Although this model is different from the rank-smoothed empirical Bernstein copula, it demonstrates how the Bernstein basis also supports likelihood-based copula estimation. ## 7. Practical guidance, common misconceptions, and open issues The choice of the Bernstein degree \(m\) is the main tuning problem. Several prescriptions appear in the literature. For symmetry testing, a rule of thumb \(m\approx c\,n^\alpha\) with \(\alpha\approx4/5\) and \(m