The Semigroup of Exponents for Partly Divisible Distributions

22 concepts3 papers1 proposal

In classical probability theory, a distribution is typically categorized into one of two extremes: it is either infinitely divisible or indecomposable. An infinitely divisible distribution (like the Gaussian or Poisson distribution) can be broken down into an arbitrary number of independent and identically distributed (i.e.d.) components. Conversely, an indecomposable distribution cannot be split at all.

However, this binary framing overlooks a rich mathematical spectrum. What if a distribution can be decomposed into two or three i.i.d. components, but not five? What if it can only be scaled by specific fractional factors? To capture these nuances, researchers developed the framework of partly divisible distributions, which replaces a simple yes/no classification with a continuous "divisibility fingerprint."

The Divisibility Fingerprint

To understand how we can quantify partial divisibility, we must look at the distribution's characteristic function $\phi(u) = \mathbb{E}[e^{iuX}]$. If we express this function in terms of its second characteristic $\psi(u) = \log \phi(u)$, then an infinitely divisible distribution has the property that $e^{t\psi(u)}$ remains a valid characteristic function for every positive real number $t > 0$.

In Partly divisible probability distributions, Albeverio et al. formalize the middle ground between full divisibility and indecomposability by defining the set of valid scaling exponents:

$$\Lambda(\mu) = \{t > 0 : e^{t\psi} \text{ is a valid characteristic function of a probability measure}\}$$

This set $\Lambda(\mu)$ serves as the structural fingerprint of a distribution's decomposability. The authors proved a vital structural property of this set: $\Lambda(\mu)$ is always a closed semigroup containing 1.

To build intuition for why this must be a semigroup, consider what happens when two scaling factors $s$ and $t$ belong to $\Lambda(\mu)$. By definition, both $e^{s\psi}$ and $e^{t\psi}$ correspond to valid, independent probability distributions. If we convolve these two distributions, their characteristic functions multiply, yielding:

$$e^{s\psi} \cdot e^{t\psi} = e^{(s+t)\psi}$$

Because the convolution of any two valid probability measures always yields another valid probability measure, the combined exponent $s+t$ must also belong to $\Lambda(\mu)$.

By constructing specific examples, the authors demonstrated that $\Lambda(\mu)$ does not have to be the entire half-line $(0, \infty)$. Instead, it can manifest as discrete sets or isolated intervals, interpolating smoothly between infinitely divisible distributions (where $\Lambda(\mu) = (0, \infty)$) and indecomposable ones.

Extending Divisibility to Topological Spaces

When we move from standard real-valued random variables to more abstract algebraic structures, defining this exponent semigroup becomes significantly more challenging. In Partly Divisible Probability Measures on Locally Compact Abelian Groups, Albeverio et al. extended the theory of partial divisibility to locally compact Abelian groups (LCA-groups).

On an abstract group, you cannot simply take the logarithm of a Fourier transform without running into topological obstructions; the transform might vanish, or its phase might wrap around precariously. To circumvent this, the authors introduced the concept of admissible probability measures.

An admissible measure is one whose Fourier transform is non-zero and satisfies strict topological conditions that allow it to be uniquely lifted to a universal covering space. The authors demonstrated that if the dual group of the LCA-group is connected, this topological framework behaves beautifully: the set of exponents $\Lambda(\mu)$ remains a well-defined, closed semigroup. However, if the dual group is not connected, the canonical definition of divisibility exponents breaks down entirely. This reveals that the very nature of how a probability distribution can be decomposed is deeply bound to the underlying geometric and topological properties of the space it inhabits.

The Boundary of Divisibility in Practice

Understanding the exact boundaries of these exponent semigroups is not just an exercise in abstract topology; it dictates how flexible distribution families can be applied in stochastic modeling. For instance, constructing a Lévy process—a foundational tool in financial engineering and physics—fundamentally requires an underlying distribution that is infinitely divisible.

In Analytical properties of generalized Gaussian distributions, Dytso et al. explored these boundaries within the Generalized Gaussian (GG) family, which generalizes the standard normal distribution by introducing a shape parameter $\beta$.

When $\beta = 2$, the distribution is a standard Gaussian, meaning its exponent set $\Lambda(\mu)$ is the complete set of positive reals $(0, \infty)$. But what happens when $\beta$ changes? By analyzing the distribution through the lens of the Lévy-Khinchine representation, the authors proved a sharp behavioral divide:

  • For shape parameters $\beta \le 2$, the distribution remains completely infinitely divisible.
  • For shape parameters $\beta > 2$, infinite divisibility fails.

When $\beta > 2$, the distribution crosses the threshold from infinite divisibility into the realm of partial divisibility or indecomposability. This precise transition illustrates why mapping the exact structure of the exponent semigroup $\Lambda(\mu)$ is essential: it outlines the exact mathematical boundaries within which these distributions can safely be used to model complex, continuous-time stochastic processes.

Key Foundational Work

Partly divisible probability distributions (1998) — contributed 12 concept elements to this direction.
Partly Divisible Probability Measures on Locally Compact Abelian Groups (2000) — contributed 8 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_admissible_probability_measures_4e252f0a,
  title = {admissible probability measures},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/admissible-probability-measures/rdr_4e252f0ae37c492eb46d5a7928aed4d4},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_admissible_probability_measures_4e252f0a_v1,
  title = {admissible probability measures},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/admissible-probability-measures/rdr_4e252f0ae37c492eb46d5a7928aed4d4/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake