Mixed Fractional Information: Consistency of Dissipation Measures for Stable Laws
William Cook · 2025
Read Mixed Fractional Information: Consistency of Dissipation Measures for Stable Laws by William Cook (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-13423.txt.
Introduction
William Cook published a paper in 2025 defining a workable notion of information for the heavy-tailed stable distributions where the classical one breaks. Symmetric alpha-stable laws, used for wild, jumpy data, have infinite classical Fisher information, so statisticians lacked a finite yardstick for comparing two such laws of different scales. Cook builds one from how fast relative entropy melts away as one law is interpolated toward another.
Two natural formulas for the new measure are proven to agree exactly, tying the entropy melt rate to differences in score functions through a clean identity. The measure is never negative and vanishes only when the scales match, and the Cauchy case gets a closed form. Numerical checks confirm the identity, grounding future work on heavy-tailed estimation.
Catalog record
| CREATOR | William Cook |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 62 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2504-13423
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Mixed Fractional Information: Consistency of Dissipation Measures for Stable Laws William Cook Abstract Symmetric α \alpha - stable (S α \alpha S) laws with α < 2 \alpha<2 model heavy - tailed data but render classical Fisher information infinite, hindering standard entropy - information analyses. Building on Johnson’s fractional de Bruijn framework, we consider the initial rate of relative - entropy dissipation along the interpolation path X t = ( 1 − t ) 1 / α X 0 + t 1 / α Z s ( α ) X_{t}=(1-t)^{1/\alpha}X_{0}+t^{1/\alpha}Z_{s}^{(\alpha)} , where Z s ( α ) Z_{s}^{(\alpha)} is an S α \alpha S random variable with scale s s . This paper defines and analyses Mixed Fractional Information (MFI) specifically for the fundamental case where the initial distribution X 0 X_{0} is also S α \alpha S, having scale parameter v v . We show that, in this S α \alpha S-to-S α \alpha S setting , MFI admits two equivalent calculation methods: 1. A chain - rule expression involving the derivative D ′ ( v ) D^{\prime}(v) of the relative entropy D ( g v ( α ) ∥ g s ( α ) ) D(g_{v}^{(\alpha)}\|g_{s}^{(\alpha)}) with respect to the scale parameter v v . 2. An integral representation 𝔼 g v [ u ( x , 0 ) ( p v F ( x ) − p s F ( x ) ) ] \mathbb{E}_{g_{v}}\bigl[u(x,0)\,\bigl(p^{F}_{v}(x)-p^{F}_{s}(x)\bigr)\bigr] , where u ( x , 0 ) = x ( v − s ) / ( s v ) u(x,0)=x(v-s)/(sv) is the specific MMSE-related score function for this linear scale interpolation path, and p F p^{F} denotes the Fisher score. Our main result is a rigorous proof of the consistency identity D ′ ( v ) = 1 α v 𝔼 g v [ X ( p v F ( X ) − p s F ( X ) ) ] . D^{\prime}(v)=\frac{1}{\alpha\,v}\,\mathbb{E}_{g_{v}}\bigl[X\,\bigl(p^{F}_{v}(X)-p^{F}_{s}(X)\bigr)\bigr]. This identity mathematically confirms the equivalence of the two MFI calculation methods for S α \alpha S inputs, establishing the internal coherence of the measure within this family. We further prove MFI’s non - negativity (vanishing if and only if v = s v=s ), derive its closed - form expression for the Cauchy case ( α = 1 \alpha=1 ), and validate the consistency identity numerically to 10 − 6 10^{-6} accuracy for general α \alpha . Within the symmetric α \alpha -stable family, MFI thus provides a finite, coherent, and calculable analogue of Fisher information, rigorously connecting entropy dissipation rates to score functions and MMSE-related quantities. This work lays a foundation for exploring potential fractional I-MMSE relations and new (weighted) functional inequalities tailored to heavy-tailed settings. Keywords: Stable Distributions, Information Theory, Relative Entropy, Fractional de Bruijn Identity, MMSE Estimation, Score Function, Entropy Dissipation, Fisher Score, Consistency, MMSE Score Function, Weighted Functional Inequalities, Information Geometry, Stable Entropy Power Inequality. 1 Introduction 1.1 Motivation: Extending Information Theory to Stable Laws Classical information theory finds elegant expression in Gaussian settings. There, Fisher information J(h) and the de Bruijn identity precisely quantify entropy dynamics under additive Gaussian noise [ 14 , 1 ] , underpinning numerous applications in science and engineering. This framework’s reach, however, is curtailed when confronting the heavy-tailed phenomena prevalent in diverse fields such as finance, signal processing, and physics. Such phenomena are often modelled effectively by symmetric \alpha -stable (S \alpha S) distributions with stability index \alpha\in(0,2) . For these distributions, the situation changes fundamentally. Their characteristic power-law density decay, g_{s}^{(\alpha)}(x)\sim c_{\alpha}|x|^{-1-\alpha} as |x|\to\infty (where c_{\alpha}>0 is a constant depending on \alpha ), implies infinite variance (for \alpha<2 ) and ensures the integrand defining classical Fisher information, h(x)(p^{F}_{h}(x))^{2} , decays too slowly for convergence over \mathbb{R} . Consequently, J(h)=\int_{\mathbb{R}}h(x)(p^{F}_{h}(x))^{2}\,\mathop{}\!\mathrm{d}x diverges whenever h is an S \alpha S density with \alpha<2 , rendering this cornerstone measure unusable and highlighting a critical gap in the standard toolkit. This divergence is not merely technical; it signifies a fundamental limitation of standard information-theoretic tools in non-Gaussian, heavy-tailed regimes, necessitating new conceptual frameworks. A promising direction stems from the fractional de Bruijn identity framework developed by Johnson [ 9 ] specifically for S \alpha S target densities (crucially, covering the full range \alpha\in(0,2] , including the Gaussian endpoint). Johnson’s work provides structures to analyse entropy dissipation beyond the Gaussian paradigm, using dynamics inherent to stable processes. Building on this foundation, this paper defines and rigorously analyses Mixed Fractional Information (MFI) specifically for the family of symmetric \alpha -stable distributions . Within this context, MFI is derived from the initial rate of relative entropy dissipation along canonical S \alpha S interpolation paths of the form X_{t}=(1-t)^{1/\alpha}X_{0}+t^{1/\alpha}Z_{s}^{(\alpha)} . We focus particularly on the fundamental case where both the initial distribution X_{0} and the target noise Z_{s}^{(\alpha)} correspond to S \alpha S densities (e.g., with scales v and s , respectively), investigating MFI’s properties and internal consistency within this setting . Our analysis establishes MFI as a finite, consistent, and theoretically grounded measure for comparing pairs of S \alpha S distributions , providing a viable alternative exactly where classical Fisher information diverges for this important class of laws. MFI’s relationship to other generalisations and alternative approaches is detailed in Section 1.4 . 1.2 Symmetric \alpha -Stable Distributions An S \alpha S random variable Z_{s}^{(\alpha)} is characterised by its characteristic function: \phi_{s}(k)=\mathbb{E}[e^{ikZ_{s}^{(\alpha)}}]=e^{-s|k|^{\alpha}},\quad k\in\mathbb{R}, (1) where \alpha\in(0,2] is the stability index and s>0 is the scale parameter. We denote its probability density function by g_{s}^{(\alpha)}(x) . We denote the derivative with respect to x as g_{s}^{(\alpha)\prime}(x) . Key properties include [ 12 , 17 ] : • Stability : If Z_{u}^{(\alpha)}\sim g_{u}^{(\alpha)} and Z_{v}^{(\alpha)}\sim g_{v}^{(\alpha)} are independent, then their sum Z_{u}^{(\alpha)}+Z_{v}^{(\alpha)}\sim g_{u+v}^{(\alpha)} . This corresponds to the convolution property g_{u}^{(\alpha)}*g_{v}^{(\alpha)}=g_{u+v}^{(\alpha)} . • Scaling : For any constant a\in\mathbb{R} , aZ_{s}^{(\alpha)}\sim g_{|a|^{\alpha}s}^{(\alpha)} . Explicitly, g_{v}^{(\alpha)}(x)=v^{-1/\alpha}g_{1}^{(\alpha)}(v^{-1/\alpha}x) . Special cases relevant to this work are: • \alpha=2 (Gaussian): g_{s}^{(2)}(x)=(4\pi s)^{-1/2}e^{-x^{2}/(4s)} . This corresponds to a Gaussian distribution with mean 0 and variance 2s . • \alpha=1 (Cauchy): g_{s}^{(1)}(x)=\frac{1}{\pi}\frac{s}{s^{2}+x^{2}} . Symmetry Assumption. Throughout this paper, we focus exclusively on symmetric \alpha -stable (S \alpha S) distributions. Generalisations to asymmetric stable laws are beyond the current scope. See Section 5 for further discussion. Parameter Range. We consider \alpha\in(0,2] and scale parameters s,v>0 . We denote the Fisher score of a density f by p^{F}_{f}=f^{\prime}/f . For S \alpha S densities g_{v}^{(\alpha)} , we may use the shorthand p^{F}_{v}=p^{F}_{g_{v}^{(\alpha)}} when the context is clear. Relative score notation. For later brevity, when comparing two S \alpha S densities g_{v}^{(\alpha)} and g_{s}^{(\alpha)} , we denote their score difference by \delta p^{F}_{v,s}(x)\;:=\;p^{F}_{g_{v}^{(\alpha)}}(x)-p^{F}_{g_{s}^{(\alpha)}}(x). 1.3 Johnson’s Fractional de Bruijn Identity Framework Johnson [ 9 ] introduced an interpolation process between an arbitrary initial random variable X_{0} with density h and an S \alpha S variable Z_{s}^{(\alpha)} : X_{t}=(1-t)^{1/\alpha}X_{0}+t^{1/\alpha}Z_{s}^{(\alpha)},\quad t\in[0,1). (2) The density h_{t} of X_{t} evolves according to the following fractional Fokker–Planck-type equation [ 9 ] of the form: \frac{\partial h_{t}}{\partial t}(x)=\frac{s}{\alpha(1-t)}\frac{\partial}{\partial x}[h_{t}(x)u(x,t)], (3) Here, u(x,t) acts as a score function along the interpolation, encoding the direction of mass transport. It is intimately linked to MMSE estimation, as formalised