Baringhaus–Gaigall Transformation

• Baringhaus–Gaigall transformation is a mixed-data method that combines Laplace and probability‐generating function kernels to characterize joint independence.
• It underpins test statistics for both bivariate and multivariate settings, offering explicit finite-sum formulations when using exponential-type weights.
• Its Hilbert-space formulation clarifies asymptotic behaviors, with quadratic statistics converging to weighted chi‐squared limits and linear statistics to normal limits.

Searching arXiv for papers on the Baringhaus–Gaigall transformation and mixed-data independence testing. The Baringhaus–Gaigall transformation is a transform-based device for characterizing independence in mixed data that combine a positive absolutely continuous component with a count component. In the formulation developed for independence testing, it is defined for a bivariate random vector $(X,Y)$ by

$$\psi(s,t)=E\!\left(e^{-sX}t^Y\right),\qquad s\ge 0,\; 0\le t\le 1,$$

with associated marginal transforms

$L(s)=E(e^{-sX}),\qquad G(t)=E(t^Y).$

Its central role is that the joint transform $\psi$ characterizes the joint law and, in particular, independence is equivalent to transform factorization: $$X \text{ and } Y \text{ are independent } \iff \psi(s,t)=L(s)G(t)\quad \text{for all } s\ge 0,\; 0\le t\le 1.$$ This factorization principle is the basis for a family of test statistics for bivariate, multivariate, and total independence problems in mixed-type data (Jelić et al., 28 Jul 2025).

1. Definition and probabilistic setting

The bivariate setting considers a random vector $(X,Y)$ such that $X>0$ is a positive absolutely continuous random variable and $Y\in\mathbb{N}_0=\{0,1,2,\dots\}$ is a count random variable. The support is restricted by the convention

$X=0 \iff Y=0.$

Hence $(X,Y)$ takes values in

$$\psi(s,t)=E\!\left(e^{-sX}t^Y\right),\qquad s\ge 0,\; 0\le t\le 1,$$0

The framework also assumes finite first moments: $$\psi(s,t)=E\!\left(e^{-sX}t^Y\right),\qquad s\ge 0,\; 0\le t\le 1,$$1

Within this setting, the Baringhaus–Gaigall transformation is

$$\psi(s,t)=E\!\left(e^{-sX}t^Y\right),\qquad s\ge 0,\; 0\le t\le 1,$$2

The transform is mixed in the precise sense that $$\psi(s,t)=E\!\left(e^{-sX}t^Y\right),\qquad s\ge 0,\; 0\le t\le 1,$$3 is the Laplace kernel for the continuous component, whereas $$\psi(s,t)=E\!\left(e^{-sX}t^Y\right),\qquad s\ge 0,\; 0\le t\le 1,$$4 is the probability-generating-function kernel for the count component. The marginal transforms are recovered as

$$\psi(s,t)=E\!\left(e^{-sX}t^Y\right),\qquad s\ge 0,\; 0\le t\le 1,$$5

This construction is specifically tailored to mixed data. A plausible implication is that its relevance derives from matching each coordinate type with a standard integral-transform device appropriate to its state space: Laplace transforms for positive continuous variables and pgfs for count variables.

2. Characterization of independence

The key characterization result is that the joint transform $$\psi(s,t)=E\!\left(e^{-sX}t^Y\right),\qquad s\ge 0,\; 0\le t\le 1,$$6 characterizes the joint law of $$\psi(s,t)=E\!\left(e^{-sX}t^Y\right),\qquad s\ge 0,\; 0\le t\le 1,$$7, and independence is equivalent to factorization: $$\psi(s,t)=E\!\left(e^{-sX}t^Y\right),\qquad s\ge 0,\; 0\le t\le 1,$$8 (Jelić et al., 28 Jul 2025).

Accordingly, dependence is represented by the deviation

$$\psi(s,t)=E\!\left(e^{-sX}t^Y\right),\qquad s\ge 0,\; 0\le t\le 1,$$9

Under independence this quantity vanishes identically; under dependence it does not. The testing framework built around the Baringhaus–Gaigall transformation therefore reduces independence testing to assessing whether the empirical counterpart of this deviation is sufficiently close to zero over a weighted transform domain.

This characterization extends beyond the scalar bivariate case. For independence between two vectors, the same factorization principle is preserved at the vector-transform level: $L(s)=E(e^{-sX}),\qquad G(t)=E(t^Y).$0 and for total independence the transform must factor into the product of all one-dimensional marginal transforms (Jelić et al., 28 Jul 2025). In this sense, the transformation functions as a characterizing transform for mixed-type dependence structures.

3. Empirical transform and test statistics

Given an i.i.d. sample

$L(s)=E(e^{-sX}),\qquad G(t)=E(t^Y).$1

the empirical transform and empirical marginals are

$L(s)=E(e^{-sX}),\qquad G(t)=E(t^Y).$2

$L(s)=E(e^{-sX}),\qquad G(t)=E(t^Y).$3

The empirical “independence defect” is

$L(s)=E(e^{-sX}),\qquad G(t)=E(t^Y).$4

Two test statistics are then defined as functionals of $L(s)=E(e^{-sX}),\qquad G(t)=E(t^Y).$5. For a weight function $L(s)=E(e^{-sX}),\qquad G(t)=E(t^Y).$6 satisfying

$L(s)=E(e^{-sX}),\qquad G(t)=E(t^Y).$7

the quadratic statistic is

$L(s)=E(e^{-sX}),\qquad G(t)=E(t^Y).$8

and the linear statistic is

$L(s)=E(e^{-sX}),\qquad G(t)=E(t^Y).$9

The paper emphasizes that $\psi$0 is simpler and admits a standard asymptotic normal limit, whereas $\psi$1 leads to a weighted $\psi$2-type limit under the null (Jelić et al., 28 Jul 2025). This division yields two distinct inferential modes: a norm-based quadratic measure of discrepancy and a linear functional of the same transform defect.

For the special weight choice

$\psi$3

both statistics admit explicit finite-sum representations. The quadratic statistic becomes

$\psi$4

The linear statistic becomes

$\psi$5

These formulas show that, with exponential-type weights, the transform-domain integrals reduce to explicit sample sums. This suggests a computational advantage: the tests can be implemented without numerical integration once the weight family is fixed.

4. Hilbert-space formulation and asymptotics

The statistics are placed in the Hilbert space

$\psi$6

equipped with inner product

$\psi$7

In this notation,

$\psi$8

This representation gives an exact geometric interpretation: independence testing becomes testing whether the empirical transform differs from the product of the empirical marginal transforms as an element of a weighted $\psi$9 space. The quadratic statistic is the squared norm of the defect, while the linear statistic is its projection onto the constant function $$X \text{ and } Y \text{ are independent } \iff \psi(s,t)=L(s)G(t)\quad \text{for all } s\ge 0,\; 0\le t\le 1.$$0 (Jelić et al., 28 Jul 2025).

Under finite second moments and the independence setup, the almost sure limits are

$$X \text{ and } Y \text{ are independent } \iff \psi(s,t)=L(s)G(t)\quad \text{for all } s\ge 0,\; 0\le t\le 1.$$1

$$X \text{ and } Y \text{ are independent } \iff \psi(s,t)=L(s)G(t)\quad \text{for all } s\ge 0,\; 0\le t\le 1.$$2

Under $$X \text{ and } Y \text{ are independent } \iff \psi(s,t)=L(s)G(t)\quad \text{for all } s\ge 0,\; 0\le t\le 1.$$3, where $$X \text{ and } Y \text{ are independent } \iff \psi(s,t)=L(s)G(t)\quad \text{for all } s\ge 0,\; 0\le t\le 1.$$4, both limits are $$X \text{ and } Y \text{ are independent } \iff \psi(s,t)=L(s)G(t)\quad \text{for all } s\ge 0,\; 0\le t\le 1.$$5.

The null weak limit of the empirical process is

$$X \text{ and } Y \text{ are independent } \iff \psi(s,t)=L(s)G(t)\quad \text{for all } s\ge 0,\; 0\le t\le 1.$$6

where $$X \text{ and } Y \text{ are independent } \iff \psi(s,t)=L(s)G(t)\quad \text{for all } s\ge 0,\; 0\le t\le 1.$$7 is a centered Gaussian process with covariance kernel

$$X \text{ and } Y \text{ are independent } \iff \psi(s,t)=L(s)G(t)\quad \text{for all } s\ge 0,\; 0\le t\le 1.$$8

Consequently,

$$X \text{ and } Y \text{ are independent } \iff \psi(s,t)=L(s)G(t)\quad \text{for all } s\ge 0,\; 0\le t\le 1.$$9

where $(X,Y)$0 are the eigenvalues of the covariance operator induced by $(X,Y)$1 and $(X,Y)$2, and $(X,Y)$3. For the linear statistic,

$(X,Y)$4

where $(X,Y)$5 and

$(X,Y)$6

For the special weight $(X,Y)$7,

$(X,Y)$8

with ratio-consistent estimator

$(X,Y)$9

and

$X>0$0

5. Multivariate and total-independence extensions

The same transform is generalized to vectors

$X>0$1

For

$X>0$2

the empirical mixed transform is

$X>0$3

Its marginal components are

$X>0$4

$X>0$5

For testing independence of two vectors, the linear and quadratic statistics are

$X>0$6

$X>0$7

With

$X>0$8

the paper derives explicit finite-sum forms for both statistics (Jelić et al., 28 Jul 2025).

Under the null of vector independence,

$X>0$9

The covariance kernel becomes

$Y\in\mathbb{N}_0=\{0,1,2,\dots\}$0

The same transform is then adapted to total independence of all components of

$Y\in\mathbb{N}_0=\{0,1,2,\dots\}$1

In that case the null is that the transform factors into the product of the one-dimensional marginal transforms: $Y\in\mathbb{N}_0=\{0,1,2,\dots\}$2 The empirical statistics $Y\in\mathbb{N}_0=\{0,1,2,\dots\}$3 and $Y\in\mathbb{N}_0=\{0,1,2,\dots\}$4 compare the empirical transform against this fully factorized product. Under total independence,

$Y\in\mathbb{N}_0=\{0,1,2,\dots\}$5

6. Significance, scope, and related considerations

The transformation matters because it is natural for mixed data, characterizing, computationally tractable, asymptotically well-behaved, and flexible in higher dimensions (Jelić et al., 28 Jul 2025). In the paper’s formulation, these properties arise directly from the choice of transform kernels and the factorization criterion.

Several points delimit its scope. First, the construction is not a generic characteristic-function method for arbitrary mixed distributions; it is given for settings that combine count variables with positive, absolutely continuous variables. Second, the support convention

$Y\in\mathbb{N}_0=\{0,1,2,\dots\}$6

is part of the formal setup in the bivariate development. Third, the asymptotic theory is formulated in weighted transform spaces and depends on the choice of weight function $Y\in\mathbb{N}_0=\{0,1,2,\dots\}$7. These are structural assumptions, not merely implementation details.

A common misconception would be to treat the Baringhaus–Gaigall transformation as itself a test statistic. In the cited development, it is instead the transform whose empirical factorization defect generates the test statistics $Y\in\mathbb{N}_0=\{0,1,2,\dots\}$8, $Y\in\mathbb{N}_0=\{0,1,2,\dots\}$9, and their multivariate analogues. Another possible misconception would be to view the quadratic and linear statistics as interchangeable. The paper distinguishes them sharply: the quadratic statistic has a weighted $X=0 \iff Y=0.$0-type null limit, while the linear statistic has a standard normal limit after standardization, and the paper reports strong empirical performance especially for the standardized linear version $X=0 \iff Y=0.$1 (Jelić et al., 28 Jul 2025).

From a broader methodological perspective, the Baringhaus–Gaigall transformation places mixed-data independence testing in the same conceptual class as transform-based methods that detect dependence through failure of factorization. This suggests continuity with a larger literature on characteristic and Laplace-type approaches, although the specific claims above concern the mixed Laplace–pgf construction and its consequences as developed in (Jelić et al., 28 Jul 2025).