Weak Variance Alpha-Gamma Process

• Weak variance alpha-gamma process is a multivariate Lévy process that generalizes the variance gamma model via weak subordination, allowing dependent Brownian components and a full covariance structure.
• It constructs the process using a common gamma factor paired with idiosyncratic gamma components, resulting in VG marginals with an enhanced range of dependence and adjustable kurtosis.
• Key analytical properties include an explicit characteristic exponent, calibration through MOM, MLE, DME approaches, and self-decomposability under zero drift conditions.

{"query":"Weak variance alpha-gamma process arXiv WVAG weak subordination selfdecomposability calibration", "max_results": 10} The weak variance alpha-gamma process, usually abbreviated WVAG and also called the weak variance-$\alpha$-gamma or weak VaG process, is a multivariate Lévy process obtained by weakly subordinating an $n$-dimensional Brownian motion by an $n$-dimensional alpha-gamma subordinator. In the cited literature, it is presented as a multivariate generalization of the variance gamma framework that relaxes the classical requirement of independent Brownian components, thereby allowing an arbitrary covariance matrix $\Sigma$ while retaining the Lévy property (Buchmann et al., 2016). It is also a principal example of the broader class of weak variance generalized gamma convolution processes, $WVGG^n$, obtained by weak subordination with multivariate Thorin subordinators (Buchmann et al., 2017).

1. Definition and origin

Weak subordination was introduced by Buchmann, Lu, and Madan as an extension of both univariate and multivariate classical subordination. In the classical multivariate setting, the pathwise operation $X\circ T$ generally preserves the Lévy property only in two special situations: when the components of $T$ are indistinguishable, or when the components of $X$ are independent. Weak subordination replaces literal pathwise composition by a law-based construction $X\odot T$ that still yields a Lévy process even when $X$ has dependent components (Buchmann et al., 2016).

Within that framework, let $n$0, let $n$1 be an $n$2-dimensional Brownian motion with drift $n$3 and arbitrary covariance matrix $n$4, and let $n$5 be an $n$6-dimensional alpha-gamma subordinator. Then

$n$7

is called a weak variance alpha-gamma process provided

$n$8

Another paper writes the parameter ordering as $n$9; both notational conventions occur in the cited literature (Buchmann et al., 2017).

The alpha-gamma subordinator is built from independent gamma subordinators $n$0. With $n$1, $n$2, and $n$3, define

$n$4

take

$n$5

independently, and set

$n$6

This gives each component a common gamma factor $n$7 together with an idiosyncratic gamma factor $n$8 (Buchmann et al., 2017).

The direct predecessor is the strong variance-$n$9-gamma process of Luciano and Semeraro, obtained by traditional subordination with an alpha-gamma clock and a Brownian motion with independent components. The weak model removes the restriction that $\Sigma$0 be diagonal. When $\Sigma$1 is diagonal, the weak model reduces to the classical strong $\Sigma$2 model (Buchmann et al., 2018).

2. Characteristic exponent, decomposition, and dependence structure

For $\Sigma$3, the cited calibration paper gives the characteristic exponent

$\Sigma$4

This formula exhibits a decomposition into a common multivariate variance-gamma-type term and coordinatewise corrections (Buchmann et al., 2018).

The same paper states that $\Sigma$5 has the decomposition in law

$\Sigma$6

where $\Sigma$7 is multivariate $\Sigma$8 and the $\Sigma$9 are univariate $WVGG^n$0 terms. In that sense, the model is a sum of one common multivariate variance-gamma component and several idiosyncratic univariate variance-gamma components. A later paper on energy spread options makes this representation explicit in a closely related notation, writing the WVAG process as a sum of independent $WVGG^n$1 processes plus deterministic drift (Leung et al., 15 Jul 2025).

Each marginal is univariate $WVGG^n$2: $WVGG^n$3 Thus the model preserves variance-gamma marginals while allowing multivariate dependence beyond the strong $WVGG^n$4 setup (Buchmann et al., 2018).

A central feature is the covariance structure. For $WVGG^n$5,

$WVGG^n$6

The cited paper emphasizes that the term $WVGG^n$7 gives $WVGG^n$8 a broader dependence range than $WVGG^n$9. Dependence can enter through Brownian correlation $X\circ T$0, drift interaction $X\circ T$1, and the shared gamma-clock effect via $X\circ T$2 and the $X\circ T$3 (Buchmann et al., 2018).

The model is therefore characterized in the literature by common and idiosyncratic time changes, VG marginals with possibly different kurtoses, full support jump measure, and a wider range of dependence than $X\circ T$4 (Buchmann et al., 2016).

3. Thorin representation and the $X\circ T$5 framework

The weak variance alpha-gamma process is a special case of a weak variance generalized gamma convolution process. In the general $X\circ T$6 construction, if $X\circ T$7 and $X\circ T$8, then

$X\circ T$9

For the alpha-gamma case, the Thorin measure is particularly simple (Buchmann et al., 2017).

Specifically, the Thorin measure of a driftless alpha-gamma subordinator is

$T$0

Thus the weak variance alpha-gamma process is a weakly subordinated Brownian motion by a Thorin subordinator whose Thorin measure is supported on finitely many rays or points (Buchmann et al., 2017).

In the general $T$1 setting, the characteristic exponent is

$T$2

and in the driftless case $T$3 this simplifies to

$T$4

For $T$5, this specialization becomes a finite linear combination of logarithmic terms corresponding to the common factor $T$6 and the idiosyncratic gamma components $T$7 (Buchmann et al., 2017).

The Lévy-measure interpretation developed in the weak-subordination paper is consistent with this structure. Jumps occur in two ways: a common jump driven by $T$8 affecting all components together, and idiosyncratic jumps driven by $T$9 affecting component $X$0 individually. This decomposition is the hallmark of the weak VaG construction (Buchmann et al., 2016).

4. Self-decomposability

A major structural question for $X$1 is self-decomposability. The general result for $X$2 states that if the Brownian motion subordinate is driftless, then the process is self-decomposable: $X$3 More precisely, for $X$4,

$X$5

This extends an earlier strong-subordination result to weak subordination (Buchmann et al., 2017).

For the weak variance alpha-gamma process, the specialization is especially sharp. Corollary $X$6 in the cited paper states that if

$X$7

then

$X$8

while if

$X$9

then

$X\odot T$0

Accordingly, for the weak variance alpha-gamma process, self-decomposability is completely characterized by driftlessness, provided the Brownian covariance is invertible (Buchmann et al., 2017).

The negative direction in the general $X\odot T$1 theory depends on moment conditions for the Thorin measure. A principal sufficient condition is

$X\odot T$2

If $X\odot T$3, $X\odot T$4, $X\odot T$5, and this condition holds, then

$X\odot T$6

For measures of the form

$X\odot T$7

the key integrability condition is equivalent to

$X\odot T$8

In the alpha-gamma case, because the Thorin measure is discrete on rays, these conditions become transparent and can be checked explicitly (Buchmann et al., 2017).

The cited paper also emphasizes an important subtlety. The moment conditions are sufficient for non-self-decomposability, but not always necessary in a naive form. An example is constructed of a $X\odot T$9 process that is still self-decomposable even though the Brownian motion subordinate has nonzero drift. This example is not the alpha-gamma process itself. For $X$0, however, the result is complete under invertibility of $X$1: driftlessness is exactly the criterion (Buchmann et al., 2017).

5. Calibration and statistical inference

A dedicated calibration study compares three estimation methods for the $X$2 parameters $X$3: method of moments (MOM), maximum likelihood estimation (MLE), and digital moment estimation (DME). MOM is based on least-squares matching of theoretical and empirical moments. MLE uses a numerically computed likelihood via Fourier inversion. DME fits the model by matching empirical and theoretical quantiles or probabilities (Buchmann et al., 2018).

Because the density is not known in closed form, MLE requires Fourier inversion of the characteristic function: $X$4 provided $X$5. For $X$6 and $X$7, assuming $X$8 is invertible, the cited paper derives the sufficient condition

$X$9

which implies

$n$00

This is the Fourier invertibility condition used to justify numerical likelihood evaluation (Buchmann et al., 2018).

The same paper reports two simulation regimes. For a bivariate $n$01 model with $n$02, the condition is satisfied and MLE gives the best overall fit. For $n$03, the condition is violated and DME gives the best fit, while MLE still works reasonably well and MOM again performs worst (Buchmann et al., 2018).

In an empirical application to daily log-returns of the S&P500 and FTSE100 over five years, the paper finds that the WVAG model fits better than VAG across fit diagnostics, that for WVAG DME is best among the three estimation methods, and that the likelihood-ratio test strongly rejects the VAG restriction $n$04: $n$05 The estimated drift vector is reported to be very close to zero, and the authors note that, using the self-decomposability result for $n$06, they do not reject self-decomposability at the $n$07 level (Buchmann et al., 2018).

6. Ornstein–Uhlenbeck embeddings, Esscher transform, and applications

The $n$08 process has also been embedded into multivariate Lévy-driven Ornstein–Uhlenbeck dynamics in two distinct ways. In WVAG-OU, the stationary distribution of the OU process is $n$09. In OU-WVAG, the background driving Lévy process (BDLP) itself is $n$10. For the OU recursion at equally spaced times $n$11,

$n$12

where

$n$13

This innovation term is the basis of the likelihood and simulation theory in both models (Lu, 2020).

For the WVAG-OU model, the cited paper derives an explicit BDLP representation and shows that the BDLP is a compound Poisson process with drift. The corresponding innovation law is a discrete-continuous mixture, which yields exact simulation and a modified likelihood based on a dominating mixture measure. For the OU-WVAG model, the paper proves that the innovation term $n$14 is absolutely continuous, so the standard Lebesgue likelihood applies, although the exponent lacks the closed form available in the $n$15-OU case (Lu, 2020).

A later application uses $n$16 as the BDLP in an OU model for energy prices, with

$n$17

That paper emphasizes that $n$18 is a flexible multivariate jump model with a natural dependence structure, and studies forward pricing, Carr–Madan call pricing, and Hurd–Zhou FFT pricing for spread options (Leung et al., 15 Jul 2025).

A central structural result in that setting is that $n$19 is not closed under the Esscher transform. Although the transformed process remains a sum of independent $n$20 components, it is generally not a $n$21 process of the original form, because the transformed marginal components are no longer jointly parameterized in the WVAG structure. The paper identifies this non-closure as a key difference between $n$22 and the multivariate $n$23 class (Leung et al., 15 Jul 2025).

The literature situates $n$24 in several finance-facing contexts. The self-decomposability paper notes applications in instantaneous portfolio theory, multivariate stock return modeling, and calibration work in finance, and stresses that weak subordination is attractive because it allows general covariance matrices $n$25, unlike strong subordination models that often require diagonality or independent components (Buchmann et al., 2017). The energy paper extends that line to energy spread options under variance gamma-driven OU dynamics (Leung et al., 15 Jul 2025).

A related but indirect development is the study of gamma-driven stochastic differential equations of the form

$n$26

where $n$27 is a gamma process. That work does not introduce a process called the weak variance alpha-gamma process, but it is described as a mathematically relevant analogue for gamma-driven, monotone, volatility-modulated models with weak solution theory and likelihood-ratio formulas (Belomestny et al., 2021). A plausible implication is that the broader analytical toolkit around gamma subordinators, absolute continuity of laws, and likelihood-based inference continues to reinforce the mathematical setting in which $n$28 is studied.