4.12.2 Sklar's Theorem & the Copula Machinery

Basic Prerequisites: 01 · Copulas from Zero and Probability & Measure Theory (CDFs, generalized inverses).

1. Intuition & Practical Objective

Page 01 showed that the marginals do not determine the joint distribution. This page supplies the structure theorem that says exactly what the missing piece is, and gives you the machinery to build and measure it.

Sklar’s theorem (1959) is the decomposition theorem of multivariate risk:

Every joint distribution factors into its margins and a copula, and conversely any copula joined to any margins is a valid joint distribution. For continuous margins the copula is unique.

The “conversely” is the practically explosive part. It says you can build a multivariate model bottom-up: model each risk factor’s marginal distribution separately (where you often have lots of data), choose a dependence structure separately (where you usually have little), and glue them. A credit-risk book can have heavy-tailed loss marginals and a Gaussian dependence structure; an equity book can have normal marginals and a $t$ dependence structure. These are different models, and Sklar’s theorem guarantees both are legitimate.

The two concrete tools this page delivers:

The copula as a rank object - $C (u_{1}, \dots, u_{d}) = Pr [U_{1} \leq u_{1}, \dots, U_{d} \leq u_{d}]$ where $U_{i} = F_{i} (X_{i})$ ; dependence on the quantile scale, invariant under monotone rescaling.
Rank correlations as copula functionals - Kendall’s $τ$ and Spearman’s $ρ_{S}$ depend only on the copula, so they are the correct inputs for calibrating one. For the Gaussian copula they are closed forms in $ϱ$ : $ρ_{τ} = \frac{2}{π} arcsin ϱ$ and $ρ_{S} = \frac{6}{π} arcsin (ϱ /2)$ .

2. Mathematical Ground Truth & Derivations

2.1 The definition and Sklar’s theorem

Definition (copula). A $d$ -dimensional copula is a distribution function on $[0, 1]^{d}$ with standard-uniform margins: $C : [0, 1]^{d} \to [0, 1]$ , $C (u) = 0$ if some $u_{i} = 0$ , $C (1, \dots, u_{i}, \dots, 1) = u_{i}$ , and $C$ is $d$ -increasing (the rectangle inequality).

Theorem (Sklar 1959). Let $F$ be a joint CDF with margins $F_{1}, \dots, F_{d}$ . Then there is a copula $C$ with

F (x_{1}, \dots, x_{d}) = C (F_{1} (x_{1}), \dots, F_{d} (x_{d})), and C (u_{1}, \dots, u_{d}) = F (F_{1}^{\leftarrow} (u_{1}), \dots, F_{d}^{\leftarrow} (u_{d})),

where $F_{i}^{\leftarrow} (u) = in f {x : F_{i} (x) \geq u}$ . If the margins are continuous, $C$ is unique; otherwise it is unique only on $Ran F_{1} \times \dots \times Ran F_{d}$ .

Proof sketch (continuous case). Set $U_{i} = F_{i} (X_{i})$ . The probability transform gives $U_{i} \sim U (0, 1)$ , and $F_{i}^{\leftarrow} (U_{i}) = X_{i}$ a.s. Then

F (x) = Pr [X_{1} \leq x_{1}, \dots, X_{d} \leq x_{d}] = Pr [U_{1} \leq F_{1} (x_{1}), \dots, U_{d} \leq F_{d} (x_{d})] = C (F_{1} (x_{1}), \dots, F_{d} (x_{d})) .

The converse: given any copula $C$ and margins $F_{i}$ , the function $C (F_{1} (x_{1}), \dots, F_{d} (x_{d}))$ is a valid joint CDF with those margins - take $U \sim C$ and set $X_{i} := F_{i}^{\leftarrow} (U_{i})$ .

Invariance. If $T_{1}, \dots, T_{d}$ are strictly increasing, then $(T_{1} (X_{1}), \dots, T_{d} (X_{d}))$ has the same copula. Dependence is a property of ranks, not levels.

Fréchet bounds. $max (\sum_{i} u_{i} + 1 - d, 0) \leq C (u) \leq min (u_{1}, \dots, u_{d})$ , with $W$ (countermonotone) and $M$ (comonotone) attained.

2.2 The fundamental and implicit copulas

Independence $Π (u) = \prod_{i} u_{i}$ : $X$ ‘s are independent iff their copula is $Π$ .
Comonotonicity $M (u) = min (u_{1}, \dots, u_{d})$ : joint CDF of $(U, \dots, U)$ ; perfect positive dependence.
Countermonotonicity $W (u_{1}, u_{2}) = max (u_{1} + u_{2} - 1, 0)$ : joint CDF of $(U, 1 - U)$ ; exists only in $d = 2$ .
Gaussian copula (implicit, from $N_{d} (0, P)$ ):

C_{P}^{G a} (u) = Φ_{P} (Φ^{- 1} (u_{1}), \dots, Φ^{- 1} (u_{d})), C_{ϱ}^{G a} (u, v) = Φ_{2} (Φ^{- 1} (u), Φ^{- 1} (v); ϱ) = \int_{- \infty}^{Φ^{- 1} (u)} \int_{- \infty}^{Φ^{- 1} (v)} \frac{1}{2 π 1 - ϱ ^{2}} exp (- \frac{s _{1}^{2} - 2 ϱ s _{1} s _{2} + s _{2}^{2}}{2 ( 1 - ϱ ^{2} )}) d s_{1} d s_{2} .

$t$ copula (implicit, from $t_{d} (ν, 0, P)$ ): $C_{ν, P}^{t} (u) = t_{ν, P} (t_{ν}^{- 1} (u_{1}), \dots, t_{ν}^{- 1} (u_{d}))$ .

Simulation (Algorithm 7.11). Generate $Z \sim N_{d} (0, P)$ ; return $U = (Φ (Z_{1}), \dots, Φ (Z_{d}))$ . The $t$ copula adds a chi-square mixing: $X = Z ν / W$ with $W \sim χ_{ν}^{2}$ , then $U_{i} = t_{ν} (X_{i})$ .

Meta distributions. Given a copula and arbitrary margins, $F (x) = C (F_{1} (x_{1}), \dots, F_{d} (x_{d}))$ is a meta-Gaussian (or meta- $t$ , meta-Clayton) distribution.

2.3 Rank correlation as a copula functional

Neither Kendall’s $τ$ nor Spearman’s $ρ_{S}$ sees anything but the copula:

ρ_{τ} (X_{1}, X_{2}) = E [sign ((X_{1} - \tilde{X}_{1}) (X_{2} - \tilde{X}_{2}))] = 4 \iint_{[0, 1]^{2}} C d C - 1,

ρ_{S} (X_{1}, X_{2}) = 12 \iint_{[0, 1]^{2}} (C (u_{1}, u_{2}) - u_{1} u_{2}) d u_{1} d u_{2} = ρ (F_{1} (X_{1}), F_{2} (X_{2})) (linear corr. of the copula) .

Both are $0$ under independence, $\pm 1$ under (counter)monotonicity, and invariant under strictly increasing transforms - exactly what Pearson correlation fails to be. For the Gaussian copula they have closed forms:

ρ_{τ} = \frac{2}{π} arcsin ϱ, ρ_{S} = \frac{6}{π} arcsin (\frac{ϱ}{2}),

so a Gaussian copula can be calibrated by inverting an empirical rank correlation: $ϱ = sin (π ρ_{τ} /2)$ .

3. Computational Implementation - Sklar verified numerically

Stdlib only. Three checks: (a) the Gaussian copula computed as a bivariate-normal integral via Sklar’s formula; (b) the probability transform of a simulated bivariate normal reproducing the copula’s empirical joint CDF; (c) the rank-correlation closed forms.

Three verifications land at once. (a) The copula value $C (0.5, 0.5) = 0.3734$ sits strictly between independence ( $0.2500$ ) and comonotonicity ( $0.5000$ ), and at $ϱ = 0$ it collapses to exactly $u v$ - Sklar’s converse with the independence copula. (b) The copula computed by the double integral matches the empirical joint CDF of the probability transform $U = Φ (X), V = Φ (Y)$ to three decimals ( $0.2992$ vs $0.2990$ ) - the proof of Sklar’s theorem, executed. (c) The empirical rank correlations reproduce the closed forms: $τ = 0.4893$ vs $\frac{2}{π} arcsin (0.7) = 0.4936$ and $ρ_{S} = 0.6780$ vs $\frac{6}{π} arcsin (0.35) = 0.6829$ (the gap is finite-sample bias in the $O (n^{2})$ Kendall estimator).

4. Failure Modes & First-Principles Breakdowns

Discrete margins break uniqueness. With non-continuous margins (default indicators! ratings!) the copula is not unique - Example 7.6 of McNeil gives a bivariate Bernoulli where infinitely many copulas fit the same joint and margins. Any inference about “the” copula from discrete credit data is under-identified.
A copula carries no marginal information. Fitting a beautiful $t$ copula to normal-independence data does not make a fat-tailed portfolio; you must model the margins separately, or the joint tail is wrong even with a “correct” copula. This is the meta-distribution point.
Pearson $ρ$ is not a copula parameter. Only under elliptical dependence does linear correlation determine the copula; in general two variables can have $ρ = 0$ and depend strongly. Rank correlations are the invariant, copula-functional measures.
Simulation error hides in the tail. Copula Monte Carlo is easy in the centre and hard where you need it - a $99.99%$ quantile may rest on a handful of paths. Variance reduction (importance sampling) or analytic shortcuts (Vasicek) are not optional at extreme quantiles.
The integral $4 \iint C d C - 1$ is a definition, not a formula. Kendall’s $τ$ for most copulas has no closed form; you estimate it from the pseudo-sample $(U_{i}, V_{i})$ and then invert. Inverting can leave the positive-definite range (the “ $t$ -copula correlation matrix from $τ$ may not be PD” problem, Algorithm 7.57), requiring an eigenvalue repair.

5. References

Sklar, A. (1959), Fonctions de répartition à n dimensions et leurs marges
McNeil, Frey & Embrechts (2015), Quantitative Risk Management
Nelsen, R. B. (2006), An Introduction to Copulas, 2nd ed.
Joe, H. (1997), Multivariate Models and Dependence Concepts
Schweizer, B. & Wolff, E. F. (1981), On nonparametric measures of dependence for random variables
Nelsen (2006) and Embrechts, McNeil & Straumann (2002)

6. Connected Graph Bridges

Back: 01 · From Zero · Index Hub
Continue: 03 · The Gaussian Copula & 2008 · [[pillars/04-quantitative-risk/copulas-and-dependence/04-tail-dependence-and-t-copula|04 · Tail Dependence & the $t$ -Copula]]
Base: Probability & Measure Theory · Linear Algebra (correlation matrices, PD repair)
Sibling: Credit Risk & Merton (the one-factor copula in credit terms) · EVT & Fat Tails

Kwant Atlas

Explorer