---
title: Power-Divergence Copulas
url: https://www.emergentmind.com/topics/power-divergence-copulas
type: topic
---

# Power-Divergence Copulas

Power-divergence copulas are a one-parameter family of Archimedean copulas obtained by taking the convex functions that generate the classical power divergences and reinterpreting them as Archimedean generators. Under the convention \(\phi(1)=0\), \(\phi'(1)=0\), and \(\phi''(1)>0\), the resulting family is always valid in the bivariate case, exhibits an unusual combination of positive and negative dependence within a single Archimedean class, and has analytically tractable ordering, tail, and singularity properties [2510.06177]. A later result settled the outstanding complete-monotonicity question on the strict negative branch and thereby certified all-dimensional Archimedean validity for the full range \(\lambda\le -1\) [2607.04467].

## 1. Construction as an Archimedean family

The starting point is the \(\phi\)-divergence convention in which a function
\[
\phi:[0,\infty)\to[0,\infty)
\]
is convex and satisfies
\[
\phi(1)=0,\qquad \phi'(1)=0,\qquad \phi''(1)>0.
\]
Under this convention, any such \(\phi\) is also convex, satisfies \(\phi(1)=0\), and is strictly decreasing on \([0,1]\). Hence it is an Archimedean generator in the adopted convention, and the corresponding bivariate copula is
\[
C(u_1,u_2;\phi)=\phi^{[-1]}(\phi(u_1)+\phi(u_2)),
\]
where
\[
\phi^{[-1]}(t)=
\begin{cases}
\phi^{-1}(t), & 0\le t<\phi(0),\\
0, & \phi(0)\le t<\infty.
\end{cases}
\]
If \(\phi(0)=\infty\), the generator is strict and \(\phi^{[-1]}=\phi^{-1}\) on \([0,\infty)\) [2510.06177].

The power-divergence family is obtained by choosing
\[
\phi_\lambda(x)\equiv
\begin{cases}
\frac{1}{\lambda(\lambda+1)}\left(x^{\lambda+1}-x+\lambda(1-x)\right), & \lambda\neq -1,0,\\[1ex]
1-x+x\log x, & \lambda=0,\\[1ex]
x-1-\log x, & \lambda=-1,
\end{cases}
\qquad \lambda\in(-\infty,\infty).
\]
This parameterization is equivalent to the usual power-divergence family after the affine normalization enforcing \(\phi'(1)=0\), and that normalization is essential for the Archimedean link. The associated copulas are
\[
C_\lambda(u_1,u_2)=\phi_\lambda^{[-1]}(\phi_\lambda(u_1)+\phi_\lambda(u_2)).
\]

The generator derivative is
\[
\phi_\lambda'(x)=
\begin{cases}
\lambda^{-1}(x^\lambda-1), & \lambda\neq -1,0,\\
\log x, & \lambda=0,\\
1-x^{-1}, & \lambda=-1.
\end{cases}
\]
At the origin,
\[
\phi_\lambda(0)=
\begin{cases}
1/(\lambda+1), & \lambda>-1,\\
\infty, & \lambda\le -1.
\end{cases}
\]
Accordingly, \(\lambda\le -1\) gives strict generators, whereas \(\lambda>-1\) gives non-strict generators requiring the pseudo-inverse.

Two special cases admit closed inverses via the Lambert \(W\) function. For \(\lambda=-1\),
\[
\phi_{-1}^{-1}(t)=-W_0\!\left(-e^{-(t+1)}\right),
\]
so
\[
C_{-1}(u_1,u_2) = -W_0\!\left(-u_1u_2\,e^{\,1-(u_1+u_2)}\right).
\]
For \(\lambda=0\),
\[
\phi_0^{[-1]}(t)=
\begin{cases}
\exp\!\left\{W_{-1}\!\left(\frac{t-1}{e}\right)+1\right\}, & 0\le t<1,\\
0, & 1\le t<\infty.
\end{cases}
\]
For general \(\lambda\neq -1,0\), inversion is numerical: if \(t=\phi_\lambda(x)\), one solves
\[
0=x^{\lambda+1}-(\lambda+1)x+\lambda-\lambda(\lambda+1)t
\]
for the unique root \(x\in[0,1]\) [2510.06177].

## 2. Dimensional validity and the complete-monotonicity problem

Bivariate validity is immediate from the Archimedean construction above, but extension to \(d\ge 3\) depends on \(d\)-monotonicity of the pseudo-inverse, and extension to all dimensions is equivalent to complete monotonicity of the inverse generator. In the standard Archimedean criterion recalled later in the literature, if \(\phi:(0,1]\to[0,\infty)\) is continuous, strictly decreasing, \(\phi(1)=0\), and \(\phi^{-1}\) is completely monotone on \((0,\infty)\), then \(\phi\) is an Archimedean generator in every dimension [2607.04467].

The dimension theory of the power-divergence family has two stages. The 2025 source paper established the bivariate family for all \(\lambda\), determined several finite-dimensional regimes, proved complete monotonicity explicitly for \(\lambda=-1\) and \(\lambda=-2\), and conjectured complete monotonicity for all \(\lambda\le -1\). The 2026 paper then resolved the remaining strict negative range and concluded all-dimensional validity for the full branch \(\lambda\le -1\) [2510.06177].

| Parameter range | Finite-dimensional status | All-dimensional status |
|---|---|---|
| \(\lambda>0\) | Valid bivariate only; not \(3\)-monotone | No extension to \(d\ge 3\) |
| \(-0.5<\lambda\le 0\) | Valid bivariate only | Restricted to \(d=2\) |
| \(-1<\lambda\le -0.5\) | Trivariate validity holds | No all-dimensional claim |
| \(\lambda\le -1\) | Strict generator; at least \(3\)-monotonicity | Valid in every dimension |

For the negative branch, the later proof rewrites \(\lambda=-\gamma\) with \(\gamma\ge 1\), uses
\[
\phi_{-\gamma}(x)=
\begin{cases}
-\dfrac{x^{1-\gamma}+(\gamma-1)x-\gamma}{\gamma(\gamma-1)}, & \gamma>1,\\[1ex]
x-1-\log x, & \gamma=1,
\end{cases}
\]
and studies the inverse \(u(t)=\phi_{-\gamma}^{-1}(t)\). Differentiation yields
\[
u'(t)=-\frac{\gamma\,u(t)^\gamma}{1-u(t)^\gamma}.
\]
The proof then introduces
\[
H_\gamma(x)=\frac{\gamma x^\gamma}{1-x^\gamma}=\gamma\sum_{m\ge 1}x^{m\gamma},
\qquad
L_\gamma=H_\gamma(x)\frac{d}{dx},
\]
defines \(P_0(x)=x\) and \(P_{n+1}=L_\gamma P_n\), and derives
\[
(-1)^n u^{(n)}(t)=P_n(u(t)).
\]
The crucial step is an explicit generalized power-series expansion
\[
P_n(x)=\gamma^n\sum_{m_1,\dots,m_n\ge 1}
\left(\prod_{j=0}^{n-1}B_j\right)x^{B_n},
\qquad
B_j=1-j+\gamma(m_1+\cdots+m_j),
\]
together with the bound
\[
B_j\ge 1+j(\gamma-1)\ge 1.
\]
This forces positivity of all exponents and coefficients, hence \(P_n(x)>0\) on \((0,1)\), and therefore strict complete monotonicity of \(\phi_\lambda^{-1}\) for every \(\lambda\le -1\). By the all-dimensional Archimedean criterion, \(\phi_\lambda\) is then a valid Archimedean generator in every dimension for all \(\lambda\le -1\) [2607.04467].

## 3. Dependence geometry, ordering, and tail behavior

Power-divergence copulas are negatively ordered:
\[
\lambda_1<\lambda_2
\quad\Longrightarrow\quad
C_{\lambda_1}(u_1,u_2)\ge C_{\lambda_2}(u_1,u_2)
\quad\forall (u_1,u_2)\in[0,1]^2.
\]
Increasing \(\lambda\) therefore moves the copula downward in the concordance order, from strong positive dependence toward strong negative dependence. The family does not include the product copula as a member, but it spans both bivariate Fréchet–Hoeffding bounds as limits:
\[
\lim_{\lambda\to\infty} C_\lambda(u_1,u_2)=\max\{u_1+u_2-1,0\},
\]
\[
\lim_{\lambda\to-\infty} C_\lambda(u_1,u_2)=\min\{u_1,u_2\}.
\]
A single Archimedean family thus ranges from perfect positive to perfect negative dependence in the bivariate case [2510.06177].

A central structural feature is the zero set. For a generic Archimedean copula the zero curve is defined by
\[
\psi(u_1)+\psi(u_2)=\psi(0).
\]
For the power-divergence family, if \(\lambda\le -1\) then \(\phi_\lambda(0)=\infty\), so the zero set reduces to
\[
Z(C_\lambda)=\{(u_1,0),(0,u_2):u_1,u_2\in[0,1]\}.
\]
If \(\lambda>-1\), the zero set has positive area; as \(\lambda\to\infty\), it tends to the triangle with vertices \((0,0)\), \((0,1)\), and \((1,0)\), which is the zero set of the Fréchet–Hoeffding lower bound. This behavior is directly tied to singularity and exclusion of jointly small values.

Absolute continuity is governed by the criterion
\[
-\lim_{s\downarrow 0}\frac{\psi(s)}{\psi'(s)}.
\]
For the power-divergence family, \(\{C_\lambda:\lambda\le 0\}\) is absolutely continuous. For \(\lambda>0\), the copula has a singular component supported on the zero curve with \(C\)-measure
\[
\frac{\lambda}{\lambda+1}\in(0,1).
\]
Thus positive \(\lambda\) yields a mixed copula, partly absolutely continuous and partly singular. For \(\lambda=1\), the singular component carries exactly \(1/2\) of the probability mass.

Kendall’s function is
\[
K_\lambda(s)=
\begin{cases}
\dfrac{\lambda}{\lambda+1}\left(\dfrac{s^{\lambda+1}-1}{s^\lambda-1}\right), & \lambda\neq -1,0,\\[2ex]
\dfrac{s-1}{\log s}, & \lambda=0,\\[2ex]
\dfrac{s\log s}{s-1}, & \lambda=-1.
\end{cases}
\]
Kendall’s tau has the Archimedean representation
\[
\tau(\lambda)=
\begin{cases}
1+\dfrac{2}{\lambda+1}-\dfrac{4\lambda}{\lambda+1}\displaystyle\int_0^1\frac{s-1}{s^\lambda-1}\,ds, & \lambda\neq -1,0,\\[2ex]
3-4\log 2, & \lambda=0,\\[1ex]
7-\dfrac{2\pi^2}{3}, & \lambda=-1.
\end{cases}
\]
Numerically,
\[
\tau(0)\approx 0.227,\qquad \tau(-1)\approx 0.420,
\]
and \(\tau(\lambda)\) is monotone decreasing with
\[
\lim_{\lambda\to-\infty}\tau(\lambda)=1,\qquad
\lim_{\lambda\to\infty}\tau(\lambda)=-1.
\]
This monotonicity makes inversion of Kendall’s tau a one-to-one estimation method.

Tail behavior is highly asymmetric. The lower-tail coefficient is
\[
T_L(\lambda)=
\begin{cases}
2^{1/(\lambda+1)}, & \lambda<-1,\\
0, & \lambda\ge -1,
\end{cases}
\]
while the upper-tail coefficient is constant across the family:
\[
T_U(\lambda)=2-\sqrt{2}\approx 0.5858.
\]
All members therefore have the same moderate, nonzero upper-tail dependence, but only the more negative parameters exhibit lower-tail dependence. For \(\lambda>-1\), lower-tail dependence vanishes and the positive-area zero set further prevents co-occurrence of sufficiently small values [2510.06177].

## 4. Computation, estimation, and the insurance application

For \(\lambda\neq -1,0\), evaluation of \(C_\lambda\) is numerical. One computes
\[
t=\phi_\lambda(u_1)+\phi_\lambda(u_2),
\]
solves
\[
x^{\lambda+1}-(\lambda+1)x+\lambda-\lambda(\lambda+1)t=0
\]
for the unique root \(x\in[0,1]\), and sets \(C_\lambda(u_1,u_2)=x\) if \(t<\phi_\lambda(0)\), otherwise \(0\). The uniqueness proposition in the source paper guarantees that this root-finding step is well defined for every \(\lambda\neq -1,0\) and every \(t\in[0,\phi_\lambda(0))\). Simulation is performed by the conditional distribution method [2510.06177].

The empirical illustration uses Danish fire insurance losses. The raw dataset contains \(2167\) claims from businesses with fire-related losses between 1980 and 1990, recording building, contents, and profits losses in millions of Danish kroner, inflation-adjusted to 1985. The analysis forms material losses as building plus contents and studies dependence between material losses and profit losses. After removing records with zero profit loss, the working sample size is \(n=616\). The rank-transformed data suggest moderate upper-tail dependence and an absence of jointly small values, consistent with a large lower zero-set region.

Estimation is by inversion of Kendall’s tau. The sample value is
\[
\bar\tau_n=0.361,
\]
and monotonicity of \(\tau(\lambda)\) yields a unique solution of
\[
\bar\tau_n-\tau(\hat\lambda)=0.
\]
The fitted parameter is
\[
\hat\lambda=-0.647.
\]
Because this lies in \((-1,-0.5]\), the fitted copula is absolutely continuous and also extendable to three dimensions in principle, although the application is bivariate.

Benchmark copulas included Archimedean, extreme-value, and elliptical models implemented in the R package `copula`: Clayton survival, Frank, Joe, Gumbel, Galambos, Hüsler–Reiss, Tawn, and Gaussian. All models were fitted by inversion of Kendall’s tau, and goodness of fit was evaluated by parametric-bootstrap tests. Only the power-divergence copula passed at the \(5\%\) level. The reported result for the fitted power-divergence copula is
\[
\hat\lambda=-0.647,\qquad \text{statistic }0.036,\qquad p=0.078.
\]
The reported benchmark \(p\)-values are \(p<0.001\) for Clayton survival, Frank, Joe, Gumbel, and Gaussian; \(p=0.014\) for Galambos; \(p=0.007\) for Hüsler–Reiss; and \(p=0.018\) for Tawn. The interpretation given in the source is geometric: the fitted power-divergence copula combines moderate upper-tail dependence, an adequate absolutely continuous density shape, and a zero-set geometry that captures the empirical absence of jointly small material and profit losses [2510.06177].

## 5. Divergence-based copula methodologies under the same label

The expression “power divergence” also appears in copula research in a different sense: not as a named Archimedean family, but as a divergence criterion defined on copulas or copula representations. This literature is adjacent but conceptually distinct from the power-divergence Archimedean family.

One line defines divergences on copula functions themselves. “Minimum Copula Divergence for Robust Estimation” introduces \(\alpha\)-, \(\beta\)-, and \(\gamma\)-copula divergences, all built directly from copula distribution functions \(C_0\) and \(C_1\) on \([0,1]^d\), and proposes the minimum copula divergence estimator
\[
\hat\theta_D=\arg\min_{\theta\in\Theta} D(\hat C,C_\theta).
\]
Its framework is rank-based, margin-free, and aimed at robust estimation under misspecification and contamination rather than at constructing a new Archimedean family [2502.16831].

A second line studies copula-based dependence measures via general \(f\)- or \(\Phi\)-divergences. “Csiszár indices and interpolating copulas” treats the power-divergence generator
\[
f_\alpha(t)=\frac{t^\alpha-\alpha t-(1-\alpha)}{\alpha(\alpha-1)},
\]
and shows that for continuous margins, the associated dependence index is exactly the \(f\)-divergence between the independence copula and the joint copula, whereas with atomic margins one must use interpolating or checkerboard copulas to obtain the dependence-preserving minimum [2603.29884]. “Parametric dependence between random vectors via copula-based divergence measures” develops a broader \(\Phi\)-divergence framework for dependence between random subvectors, with explicit Gaussian formulas for the KL and Hellinger cases and general parametric estimation for copula models [2302.13611].

These works are not alternate derivations of the Pearse–Bondell family. They use “power divergence” or allied \(f\)-divergence terminology to quantify discrepancy or dependence on the copula scale, whereas power-divergence copulas in the strict sense are the one-parameter Archimedean family generated by \(\phi_\lambda\) [2510.06177].

## 6. Adjacent “power” constructions and recurrent confusions

Several nearby literatures use “power” in still different ways. “Relative local dependence of bivariate copulas” proposes a class defined by the PDE
\[
\frac{\partial^2}{\partial u\partial v}\log c(u,v)=\zeta\,c(u,v)^k,
\]
equivalently \(i^c(u,v)=\zeta c(u,v)^k\), so it is a power-law relation between local dependence and copula density. That paper does not define copulas through Cressie–Read, density power divergence, Tsallis divergence, or Rényi divergence [2407.16948]. “Operator Tail Dependence of Copulas” studies operator power scaling \(u^A\) and non-standard tail dependence with possibly distinct marginal tail exponents; its exact terminology is operator tail dependence, not power-divergence copulas [1611.06193]. “Stepwise Variational Inference with Vine Copulas” is direct on copulas and direct on non-KL divergence objectives, but its setting is Rényi-divergence-based variational inference with vine copulas rather than the power-divergence Archimedean family [2603.22959].

A separate potential source of confusion is outer power Archimedean modeling. No relation between outer power transformations and power-divergence copulas can be attributed to the cited 2020 arXiv record, because that document contains no substantive scientific content and does not establish whether outer power transformations are equivalent to, related to, or distinct from power-divergence constructions [2003.13301].

In the strict usage established by the 2025 and 2026 papers, power-divergence copulas therefore denote the Archimedean family generated by \(\phi_\lambda\), with bivariate validity for all \(\lambda\in\mathbb R\), finite-dimensional restrictions on the non-strict branch, and all-dimensional validity on the strict negative branch \(\lambda\le -1\) [2510.06177]. The broader divergence literature is relevant context, but it studies either divergence functionals on copulas, copula-based dependence indices, or divergence-driven inference, rather than the same family of copulas.

Source: https://www.emergentmind.com/topics/power-divergence-copulas