Introduction
Stable distributions are increasingly being used in both theoretical and empirical investigations in applied literature. Starting with a well cited article by Mandelbrot [ [21 ] , there has been empirical support for thick tail laws in general and stable laws in particular. McFarland and Sung [1982] used symmetric stable laws to study returns from foreign exchange market; as a continuation of this paper, So [ [33 ] observed that return distribution follow negatively skewed stable laws. Buckle [ [6 ] in his analysis of return of a share price observed negatively skewed stable laws. Tsionas [ [34 ] used sub-Gaussian univariate stable laws in a stable regression setup. Carr and Wu [ [8 ] used negatively skewed stable laws in analyzing persistence of volatility smirk across maturity spectrum in the context of S&P 500 Index options; they observed that a skewed univariate stable law does a better job of fitting option prices as compared to stochastic volatility models. Rachev and Mittnik [ [29 ] has extensively used thick tails distributions. Thus, skewed univariate stable law has been receiving the attention of researchers in empirical finance.
There has been increased attention towards multivariate stable distributions in the last two decades. Here again, majority of the applications that we are aware of are from applied finance, mainly with the objective of modeling financial returns from a portfolio. Commonly studied examples are from foreign exchange markets and stock markets, such as, joint movement of a basket of currencies or a set of stock indices. Also, there are occasions where distribution of the basket as a whole is necessary; computation of Value at Risk (VaR) of a portfolio is one such example. Most computations of VaR are currently based on multivariate Gaussian laws, which is known to underestimate losses. Limitations of multivariate Gaussian laws led researchers to practical alternatives, such as historical simulation methods [14 , 5 ] , and copula functions [5 , 9 ] .
An alternative strand of literature approaches this problem using sub-Gaussian models; see, for instance, Qiou and Ravishanker [ [28 ] . However, as elucidated by Samorodnitsky and Taqqu [ [32 ] , sub-Gaussian models are ’not very far away from Gaussian models’ in the sense of Proposition 2.5.7 [32 ] . Moreover, not all symmetric stable laws are sub-Gaussian.
If individual return distribution of each item in a basket is likely to have thick tail, as most research with univariate laws suggests, it is more natural to use a multivariate law with this property. Due to generalised central limit theorem, multivariate stable laws are ideal candidates, but are also notorious for their mathematical intractability. Multivariate stable laws are implicitly described through its characteristic function which involves a measure on the unit sphere in R d \mathbb{R}^{d} R d , known as spectral measure. Basic properties and tail dependence structures of multivariate stable laws are expressed in terms of the spectral measure.
Despite its importance in understanding multivariate stable laws, statistical estimation of spectral measure continues to be a major challenge. Attempts in estimating spectral measure so far have been based on a discrete approximation result by Byczkowski et al. [ [7 ] and results on univariate projections by Abdul-Hamid and Nolan [ [1 ] . Research by Nolan and his collaborators has given us reasonable understanding of bivariate stable laws; see, for instance, Nolan and Panorska [ [26 ] for bivariate stable densities in foreign exchange rates. Pivato and Seco [ [27 ] approached the problem of estimation of spectral measure through spherical harmonic analysis. More recently, Tsionas [ [35 ] developed Bayesian inference using one-dimensional projections of Matsui and Takemura [ [23 ] .
In the present paper, we approach the problem using a LePage-type representation [20 , 19 ] with the von Mises – Fisher distribution as the spectral measure. This approach is motivated by five factors.
von Mises through Brownian motion: Gordon and Hudson [ [12 ] showed that the von Mises distribution arises as the hitting distribution of a two-dimensional Brownian motion with drift; see also Gordon and Hudson [ [13 ] . The pole-seeking Brownian motion of Kendall [ [15 ] gives a vivid physical example through bird navigation. To our knowledge, the present work is the first to apply the Gordon–Hudson characterisation to motivate the spectral measure of a multivariate stable distribution. Gatto [ [11 ] extended the Gordon–Hudson result to higher dimensions.
Regular variation: For d d d -dimensional stable laws, the tail probability and directional behaviour in any cone are jointly determined by the normalised spectral measure [Resnick, 1987, [3 ] . The von Mises – Fisher distribution provides an explicit parametric form for this measure, with the mean direction η \eta η identifying the dominant tail direction and the concentration κ \kappa κ measuring tail anisotropy.
LePage-type representation: Damien et al. [ [10 ] developed an approximation theorem for simulating one-dimensional infinitely divisible distributions. We extend their approach to the multivariate setting, identifying a latent structure within stable laws — radial, spectral, and Poisson components — that yields a representation theorem in the spirit of Le Page et al. [ [19 ] and forms the basis of the Gibbs sampler (§ 3.2 ).
Truncation error through Gaussian: The LePage series is truncated at m m m terms for numerical implementation, introducing an approximation error. We model this error through a Gaussian component, which yields an explicit likelihood function for all parameters jointly — a step that makes full Bayesian inference tractable for d d d -dimensional stable distributions.
Radial component as latent variable: The classical approach integrates out the radial component analytically, yielding a characteristic function that requires special treatment at the Cauchy boundary α = 1 \alpha= 1 α = 1 . In the present paper, the radial component is retained as an explicit latent variable in the Gibbs sampler, following Kuelbs [ [18 ] . This has two consequences; the likelihood is smooth in α ∈ ( 0 , 2 ) \alpha\in(0, 2) α ∈ ( 0 , 2 ) with no special case at α = 1 \alpha= 1 α = 1 ; and the radial draws carry direct information about α \alpha α through a sufficient statistic, independently of the spectral parameters. This is a fundamental departure from existing approaches to multivariate stable inference.
Use of von Mises – Fisher distribution as a model for normalized spectral measure allows us to explore tools of directional statistics in an altogether different area of d d d -dimensional stable laws. In particular, terminologies of directional statistics elicit natural and lucid interpretation of the mean direction of the spectral mass. Through von Mises – Fisher distribution, asymmetry of the spectral measure is controlled by its concentration parameter. We believe that this approach eliminates some of the limitations of discrete approximations and also seems to enhance the horizon of understanding of multivariate stable laws.
A perhaps surprising finding is that the Gaussian limit of the bivariate stable distribution with von Mises spectral measure has an explicit covariance matrix Σ Γ = Γ ∞ ( I 2 + R 2 ( κ ) A ( η ) ) \Sigma_\Gamma= \Gamma_\infty(I_2 + R_2(\kappa)A(\eta)) Σ Γ = Γ ∞ ( I 2 + R 2 ( κ ) A ( η )) , where R 2 ( κ ) = I 2 ( κ ) / I 0 ( κ ) R_2(\kappa) = I_2(\kappa)/I_0(\kappa) R 2 ( κ ) = I 2 ( κ ) / I 0 ( κ ) is the second-order mean resultant length and Γ ∞ \Gamma_\infty Γ ∞ is the total spectral mass. This connects the tail anisotropy of the stable distribution directly to the Gaussian covariance structure at the boundary. For general result in d d d -dimension, see § 2.3 .
The paper is organised as follows. Section 2 motivates the von Mises spectral measure through its connection to the hitting distribution of Brownian motion on the unit circle. Section 3 develops the LePage-type representation and complete-data likelihood. Identifiability of the stability index α \alpha α is established in § 4 . § 5 describes the Markov chain initialisation, with full details in the online Appendix. Section 6 presents the simulation design and results. Section 7 concludes.
Directional structure of multivariate stable distributions
Gordon–Hudson–Gatto Theorem
Gordon and Hudson [ [12 ] showed that the von Mises distribution arises as the hitting distribution of a drifted Brownian motion on S 1 S^1 S 1 . Gatto [ [11 ] extended this result to general dimension d ≥ 2 d \ge2 d ≥ 2 ; the von Mises–Fisher distribution vMF ( μ , κ ) \operatorname{vMF}(\mu,\kappa) vMF ( μ , κ ) on S d − 1 S^{d-1} S d − 1 is the hitting distribution of a d d d -dimensional Brownian motion with constant drift ν = κ μ \nu= \kappa\mu ν = κ μ on the unit sphere S d − 1 S^{d-1} S d − 1 .
Theorem 1 (Gordon–Hudson–Gatto). Let { B t } t ≥ 0 \{B_t\}_{t\geq0} { B t } t ≥ 0 be a d d d -dimensional standard Brownian motion starting at the origin, with constant drift vector ν = κ μ ∈ R d \nu= \kappa\mu\in\mathbb{R}^d ν = κ μ ∈ R d , where μ ∈ S d − 1 \mu\in S^{d-1} μ ∈ S d − 1 and κ ≥ 0 \kappa\geq0 κ ≥ 0 . That is, B t = W t + κ μ t B_t = W_t + \kappa\mu t B t = W t + κ μ t , with W t W_t W t a standard Brownian motion in R d \mathbb{R}^d R d . Let τ = inf { t > 0 : ∥ B t ∥ = 1 } \tau= \inf\{t > 0 : \lVert B_t\rVert= 1\} τ = inf { t > 0 : ∥ B t ∥ = 1 } be the first hitting time of S d − 1 S^{d-1} S d − 1 , and let S = B τ / ∥ B τ ∥ ∈ S d − 1 S = B_\tau/\lVert B_\tau\rVert\in S^{d-1} S = B τ / ∥ B τ ∥ ∈ S d − 1 be the hitting direction. Then S S S has the von Mises–Fisher distribution:
S ∼ vMF ( μ , κ ) , S \sim\operatorname{vMF}(\mu,\kappa), S ∼ vMF ( μ , κ ) , with density
f vMF ( s ∣ μ , κ ) = C d ( κ ) exp ( κ μ T s ) , s ∈ S d − 1 , (1) f_{\operatorname{vMF}}(s \mid\mu,\kappa) = C_d(\kappa)\exp(\kappa\mu^\mathsf{T}s), \qquad s \in S^{d-1}, \tag*{(1)} f vMF ( s ∣ μ , κ ) = C d ( κ ) exp ( κ μ T s ) , s ∈ S d − 1 , (1) where C d ( κ ) = κ d / 2 − 1 ( 2 π ) d / 2 I d / 2 − 1 ( κ ) C_d(\kappa) = \frac{\kappa^{d/2-1}}{(2\pi)^{d/2}I_{d/2-1}(\kappa)} C d ( κ ) = ( 2 π ) d /2 I d /2 − 1 ( κ ) κ d /2 − 1 is the normalising constant and I p ( ⋅ ) I_p(\cdot) I p ( ⋅ ) is the modified Bessel function of the first kind of order p p p .
At d = 2 d = 2 d = 2 , (1) reduces to the von Mises density (2) :
f vM ( θ ∣ η , κ ) = e κ cos ( θ − η ) 2 π I 0 ( κ ) , θ ∈ [ 0 , 2 π ) , (2) f_{\operatorname{vM}}(\theta\mid\eta,\kappa) = \frac{e^{\kappa\cos(\theta-\eta)}}{2\pi I_0(\kappa)}, \qquad\theta\in[0,2\pi), \tag*{(2)} f vM ( θ ∣ η , κ ) = 2 π I 0 ( κ ) e κ c o s ( θ − η ) , θ ∈ [ 0 , 2 π ) , (2)
where I 0 ( κ ) I_0(\kappa) I 0 ( κ ) is the modified Bessel function of the first kind of order zero.
This result motivates our choice of spectral measure for X ∼ Stable ( α , Γ ) X \sim\operatorname{Stable}(\alpha,\Gamma) X ∼ Stable ( α , Γ ) in R d \mathbb{R}^d R d . Let Γ ~ = Γ / Γ ∞ \widetilde{\Gamma} = \Gamma/\Gamma_\infty Γ = Γ/ Γ ∞ be the normalised version of the spectral measure Γ \Gamma Γ , where Γ ∞ = Γ ( S d − 1 ) \Gamma_\infty= \Gamma(S^{d-1}) Γ ∞ = Γ ( S d − 1 ) . By setting Γ ~ ∼ vMF ( μ , κ ) \widetilde{\Gamma} \sim\operatorname{vMF}(\mu,\kappa) Γ ∼ vMF ( μ , κ ) , the spectral measure acquires a natural Brownian motion interpretation—it is the distribution of exit directions of an independent, drifted Brownian motion from the unit sphere. The mean direction μ \mu μ identifies the dominant tail direction of the stable distribution and the concentration κ \kappa κ measures the degree of tail anisotropy. To the best of our knowledge, this connection between the vMF spectral measure and the drifted Brownian motion hitting distribution has not previously appeared in the stable distribution literature. Table 1 provides the parameter-level linkage for general d d d .
Table 1. Linkage between the vMF spectral measure parameters and the first hitting direction of a d d d -dimensional drifted Brownian motion on S d − 1 S^{d-1} S d − 1 [12 , 11 ] .
Regular Variation Characterisation
The connection between the spectral measure Γ \Gamma Γ and the tail behaviour of X ∼ Stable ( α , Γ ) X \sim\operatorname{Stable}(\alpha,\Gamma) X ∼ Stable ( α , Γ ) in R d \mathbb{R}^{d} R d is made precise by the regular variation characterisation [30 , 3 ] . With Γ ~ ∼ vMF ( μ , κ ) \widetilde{\Gamma} \sim\operatorname{vMF}(\mu,\kappa) Γ ∼ vMF ( μ , κ ) , for any Borel set A ⊂ S d − 1 A \subset S^{d-1} A ⊂ S d − 1 , Γ ( ∂ A ) = 0 \Gamma(\partial A) = 0 Γ ( ∂ A ) = 0 , we have
lim x → ∞ P ( X / ∥ X ∥ ∈ A , ∥ X ∥ > x ) P ( ∥ X ∥ > x ) = Γ ~ ( A ) = ∫ A f v M F ( s ∣ μ , κ ) d s . (3) \lim_{x \to\infty} \frac{P(X/\lVert X\rVert\in A,\ \lVert X\rVert> x)}{P(\lVert X\rVert> x)} = \widetilde{\Gamma}(A) = \int_A f_{\mathrm{vMF}}(s \mid\mu,\kappa)\,\mathrm{d}s. \tag*{(3)} x → ∞ lim P (∥ X ∥ > x ) P ( X / ∥ X ∥ ∈ A , ∥ X ∥ > x ) = Γ ( A ) = ∫ A f vMF ( s ∣ μ , κ ) d s . (3)
Thus, conditional on lying in the far tail, the probability that X X X points into direction set A A A converges to the normalised spectral measure evaluated at A A A . The spectral measure is therefore the limiting angular distribution of extreme observations. For the vMF spectral measure, (3) provides an explicit parametric form for this limiting distribution, valid in any dimension d ≥ 2 d \ge2 d ≥ 2 , controlled by the mean direction μ \mu μ and concentration κ \kappa κ .
It is also worth mentioning that the von Mises–Fisher distribution vMF ( μ , κ ) \operatorname{vMF}(\mu,\kappa) vMF ( μ , κ ) on S d − 1 S^{d-1} S d − 1 is the unique maximum entropy distribution subject to a fixed mean resultant vector E [ s ] = s ˉ \mathbb{E}[s] = \bar{s} E [ s ] = s ˉ (fixed mean direction μ ∈ S d − 1 \mu\in S^{d-1} μ ∈ S d − 1 and fixed mean resultant length R ˉ = ∥ s ˉ ∥ ∈ [ 0 , 1 ] \bar{R} = \lVert\bar{s}\rVert\in[0,1] R ˉ = ∥ s ˉ ∥ ∈ [ 0 , 1 ] ) [22 ] .
The Gaussian Boundary
At the Gaussian boundary, we show that all spectral measures induces a specific covariance structure. In the case of vMF, the covariance gets a closed form through its second moment matrix.
Theorem 2 (Gaussian boundary, general d d d ). Let X ∼ Stable ( α , Γ ) X \sim\operatorname{Stable}(\alpha, \Gamma) X ∼ Stable ( α , Γ ) in R d \mathbb{R}^{d} R d , α ∈ ( 0 , 2 ) \alpha\in(0,2) α ∈ ( 0 , 2 ) , with finite spectral measure Γ \Gamma Γ on S d − 1 S^{d-1} S d − 1 and total mass Γ ∞ = Γ ( S d − 1 ) < ∞ \Gamma_{\infty} = \Gamma(S^{d-1}) < \infty Γ ∞ = Γ ( S d − 1 ) < ∞ . As α → 2 \alpha\to2 α → 2 ,
X → d N d ( 0 , Σ Γ ) , Σ Γ = 2 Γ ∞ ∫ S d − 1 s s T Γ ~ ( d s ) . (4) X \xrightarrow{d} N_{d}(0,\Sigma_{\Gamma}), \qquad\Sigma_{\Gamma}=2\Gamma_{\infty}\int_{S^{d-1}}ss^{\mathsf{T}}\widetilde{\Gamma}(\mathrm{d}s). \tag*{(4)} X d N d ( 0 , Σ Γ ) , Σ Γ = 2 Γ ∞ ∫ S d − 1 s s T Γ ( d s ) . (4)
The limiting covariance Σ Γ \Sigma_{\Gamma} Σ Γ is determined entirely by the second moment of the normalised spectral measure Γ ~ \widetilde{\Gamma} Γ and the total spectral mass Γ ∞ \Gamma_{\infty} Γ ∞ .
Thus, as per Theorem 4, at the boundary, information on second moment of the spectral measure is retained and all higher-order information is lost.
Corollary 1 (von Mises–Fisher spectral measure). Under the conditions of Theorem 4, let Γ = Γ ∞ Γ ~ \Gamma=\Gamma_{\infty}\widetilde{\Gamma} Γ = Γ ∞ Γ where Γ ~ \widetilde{\Gamma} Γ has density f v M F ( ⋅ ; μ , κ ) f_{\mathrm{vMF}}(\cdot;\mu,\kappa) f vMF ( ⋅ ; μ , κ ) with respect to the uniform measure on S d − 1 S^{d-1} S d − 1 , μ ∈ S d − 1 \mu\in S^{d-1} μ ∈ S d − 1 is the mean direction, and κ ≥ 0 \kappa\geq0 κ ≥ 0 is the concentration parameter. Then
Σ Γ = 2 Γ ∞ [ b ( κ , d ) I d + c ( κ , d ) μ μ T ] , (5) \Sigma_{\Gamma}=2\Gamma_{\infty}\left[b(\kappa,d)I_{d}+c(\kappa,d)\mu\mu^{\mathsf{T}}\right], \tag*{(5)} Σ Γ = 2 Γ ∞ [ b ( κ , d ) I d + c ( κ , d ) μ μ T ] , (5) where
a ( κ , d ) = 1 d + d − 1 d I d / 2 + 1 ( κ ) I d / 2 − 1 ( κ ) , b ( κ , d ) = 1 − a ( κ , d ) d − 1 , c ( κ , d ) = a ( κ , d ) − b ( κ , d ) . (6) \begin{aligned}
a(\kappa,d)&=\frac{1}{d}+\frac{d-1}{d}\frac{I_{d/2+1}(\kappa)}{I_{d/2-1}(\kappa)},\\
b(\kappa,d)&=\frac{1-a(\kappa,d)}{d-1},\\
c(\kappa,d)&=a(\kappa,d)-b(\kappa,d). \tag*{(6)}
\end{aligned} a ( κ , d ) b ( κ , d ) c ( κ , d ) = d 1 + d d − 1 I d /2 − 1 ( κ ) I d /2 + 1 ( κ ) , = d − 1 1 − a ( κ , d ) , = a ( κ , d ) − b ( κ , d ) . (6)
The matrix Σ Γ \Sigma_{\Gamma} Σ Γ has eigenvector μ \mu μ with eigenvalue 2 Γ ∞ a ( κ , d ) 2\Gamma_{\infty}a(\kappa,d) 2 Γ ∞ a ( κ , d ) and every direction orthogonal to μ \mu μ with eigenvalue 2 Γ ∞ b ( κ , d ) 2\Gamma_{\infty}b(\kappa,d) 2 Γ ∞ b ( κ , d ) . The anisotropy ratio a ( κ , d ) / b ( κ , d ) a(\kappa,d)/b(\kappa,d) a ( κ , d ) / b ( κ , d ) increases monotonically from 1 (isotropic, κ = 0 \kappa=0 κ = 0 , uniform spectral measure) to ∞ \infty ∞ , which follows from Amos [ [2 ] .
Proposition 1 (Anisotropy under von Mises–Fisher spectral measure). Under the conditions of Corollary 3, the anisotropy ratio
ρ ( κ , d ) = a ( κ , d ) b ( κ , d ) = ( d − 1 ) a ( κ , d ) 1 − a ( κ , d ) \rho(\kappa,d)=\frac{a(\kappa,d)}{b(\kappa,d)}=\frac{(d-1)a(\kappa,d)}{1-a(\kappa,d)} ρ ( κ , d ) = b ( κ , d ) a ( κ , d ) = 1 − a ( κ , d ) ( d − 1 ) a ( κ , d ) satisfies the following:
(i) ρ ( 0 , d ) = 1 \rho(0,d)=1 ρ ( 0 , d ) = 1 : when κ = 0 \kappa=0 κ = 0 the normalised spectral measure Γ ~ \widetilde{\Gamma} Γ is uniform and Σ Γ ∝ I d \Sigma_{\Gamma}\propto I_{d} Σ Γ ∝ I d (isotropic);
(ii) ρ ( κ , d ) \rho(\kappa,d) ρ ( κ , d ) is strictly increasing in κ \kappa κ on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) , so greater spectral concentration along μ \mu μ produces strictly greater anisotropy;
(iii) lim κ → ∞ ρ ( κ , d ) = ∞ \lim_{\kappa\to\infty}\rho(\kappa,d)=\infty lim κ → ∞ ρ ( κ , d ) = ∞ : as the spectral measure Γ ~ \widetilde{\Gamma} Γ concentrates on { μ } \{\mu\} { μ } the distribution becomes degenerate along μ \mu μ .
Hence κ \kappa κ is a pure anisotropy parameter: it controls the eigenvalue ratio of the Gaussian limit covariance Σ Γ \Sigma_{\Gamma} Σ Γ monotonically, independently of α \alpha α and Γ ∞ \Gamma_{\infty} Γ ∞ .
Remark 1. Regular variation equation (3) reveals that the anisotropy of extreme observations is governed entirely by κ \kappa κ : the limiting angular distribution of tail realisations is vMF ( μ , κ ) \operatorname{vMF}(\mu,\kappa) vMF ( μ , κ ) for every α ∈ ( 0 , 2 ] \alpha\in(0,2] α ∈ ( 0 , 2 ] . The anisotropy ratio ρ ( κ , d ) \rho(\kappa,d) ρ ( κ , d ) of Proposition 1, derived at the Gaussian boundary α = 2 \alpha= 2 α = 2 , therefore reflects a geometric property that is in fact exact in the tail at every α \alpha α .
Corollary 2 (Bivariate case, d = 2 d = 2 d = 2 ). At d = 2 d = 2 d = 2 , with μ = ( cos η , sin η ) T \mu= (\cos\eta,\sin\eta)^{\mathsf{T}} μ = ( cos η , sin η ) T and R 2 ( κ ) = I 2 ( κ ) / I 0 ( κ ) R_{2}(\kappa) = I_{2}(\kappa)/I_{0}(\kappa) R 2 ( κ ) = I 2 ( κ ) / I 0 ( κ ) , Corollary 3 reduces to
Σ Γ = Γ ∞ ( I 2 + R 2 ( κ ) A ( η ) ) , A ( η ) = ( cos 2 η sin 2 η sin 2 η − cos 2 η ) . (7) \Sigma_{\Gamma} = \Gamma_{\infty}(I_{2} + R_{2}(\kappa)A(\eta)), \qquad A(\eta) = \begin{pmatrix} \cos2\eta& \sin2\eta\\ \sin2\eta& -\cos2\eta\end{pmatrix}. \tag*{(7)} Σ Γ = Γ ∞ ( I 2 + R 2 ( κ ) A ( η )) , A ( η ) = ( cos 2 η sin 2 η sin 2 η − cos 2 η ) . (7)
Remark 2 (Special cases). (i) Isotropic case (κ = 0 \kappa= 0 κ = 0 , uniform spectral measure): a ( 0 , d ) = 1 / d a(0,d) = 1/d a ( 0 , d ) = 1/ d , b ( 0 , d ) = 1 / d b(0,d) = 1/d b ( 0 , d ) = 1/ d , c ( 0 , d ) = 0 c(0,d) = 0 c ( 0 , d ) = 0 , giving Σ Γ = ( 2 Γ ∞ / d ) I d \Sigma_{\Gamma} = (2\Gamma_{\infty}/d)I_{d} Σ Γ = ( 2 Γ ∞ / d ) I d .
(ii) Concentrated case (κ → ∞ \kappa\to\infty κ → ∞ , point mass at μ \mu μ ): a → 1 a \to1 a → 1 , b → 0 b \to0 b → 0 , c → 1 c \to1 c → 1 , giving Σ Γ → 2 Γ ∞ μ μ T \Sigma_{\Gamma} \to2\Gamma_{\infty}\mu\mu^{\mathsf{T}} Σ Γ → 2 Γ ∞ μ μ T .
(iii) Generality : Theorem 4 holds for any finite spectral measure Γ \Gamma Γ — discrete, continuous, or mixed.
LePage-type Representation and Gaussian Truncation Error
The classical benchmark for representing stable distributions is the LePage series representation [20 , 19 ] and its generalisation [31 ] . Let { X i } i = 1 n \{X_i\}_{i=1}^{n} { X i } i = 1 n be stable random vectors in R d \mathbb{R}^{d} R d . Let s i j ∈ S d − 1 s_{ij} \in\mathbb{S}^{d-1} s ij ∈ S d − 1 , i = 1 , … , n i = 1,\ldots,n i = 1 , … , n , j = 1 , 2 , … j = 1,2,\ldots j = 1 , 2 , … , be independent identically distributed draws from the normalised spectral measure Γ ~ \widetilde{\Gamma} Γ . The LePage representation is
X i = ∑ j = 1 ∞ L i j ( α ) s i j , i = 1 , … , n , (8) X_i = \sum_{j=1}^{\infty} L_{ij}(\alpha)s_{ij}, \qquad i = 1,\ldots,n, \tag*{(8)} X i = j = 1 ∑ ∞ L ij ( α ) s ij , i = 1 , … , n , (8)
where L i j ( α ) L_{ij}(\alpha) L ij ( α ) are weights derived from the arrival times of a unit-rate Poisson process on R + \mathbb{R}_{+} R + , independent of { s i j } \{s_{ij}\} { s ij } .
Direct implementation of (8) is computationally demanding; the series is infinite and sorting Poisson arrival times is expensive. Instead, we use a modified version of the infinitely divisible representation of Damien et al. [ [10 ] , truncated at m m m terms, with truncation error captured through a Gaussian component:
X i = ∑ j = 1 m L i j ( α ) s i j + ϵ i , i = 1 , … , n , (9) X_i = \sum_{j=1}^{m} L_{ij}(\alpha)s_{ij} + \epsilon_i, \qquad i = 1,\ldots,n, \tag*{(9)} X i = j = 1 ∑ m L ij ( α ) s ij + ϵ i , i = 1 , … , n , (9)
where ϵ i ∼ N d ( 0 , ( σ 2 / m ) I d ) \epsilon_i \sim N_d(0,(\sigma^2/m)I_d) ϵ i ∼ N d ( 0 , ( σ 2 / m ) I d ) and σ 2 > 0 \sigma^2 > 0 σ 2 > 0 is an unknown parameter that absorbs the truncation bias. In the modified representation, L i j ( α ) L_{ij}(\alpha) L ij ( α ) is a function of independent Poisson and inverted-gamma variates.
A LePage-type representation theorem
Various authors developed methods to draw from univariate infinitely divisible distributions. We adopt a univariate result by Damien et al. [ [10 ] for stable distributions.
Theorem 3 (Damien et al. [ [10 ] , modified for stable distributions). Let s 1 , … , s m s_1,\ldots,s_m s 1 , … , s m be independent and identically distributed from vMF ( η , κ ) \operatorname{vMF}(\eta,\kappa) vMF ( η , κ ) . Let r 1 , … , r m r_1,\ldots,r_m r 1 , … , r m be independent and identically distributed as inverted gamma with parameter α \alpha α and density
g ( r ) = 1 Γ ( α ) r − 1 − α e − 1 / r , r > 0. (10) g(r)=\frac{1}{\Gamma(\alpha)}r^{-1-\alpha}e^{-1/r},\quad r>0. \tag*{(10)} g ( r ) = Γ ( α ) 1 r − 1 − α e − 1/ r , r > 0. (10) For j = 1 , … , m j=1,\ldots,m j = 1 , … , m , let y j y_j y j be drawn independently from Poisson ( λ j ) \operatorname{Poisson}(\lambda_j) Poisson ( λ j ) , where λ j = Γ ∞ Γ ( α ) e 1 / r j / m \lambda_j=\Gamma_\infty\Gamma(\alpha)e^{1/r_j}/m λ j = Γ ∞ Γ ( α ) e 1/ r j / m . Define
X ( m ) = ∑ j = 1 m r j { y j − λ j 1 + r j 2 } s j = ∑ j = 1 m L j s j , (11) X^{(m)}=\sum_{j=1}^{m}r_j\left\{y_j-\frac{\lambda_j}{1+r_j^2}\right\}s_j=\sum_{j=1}^{m}L_js_j, \tag*{(11)} X ( m ) = j = 1 ∑ m r j { y j − 1 + r j 2 λ j } s j = j = 1 ∑ m L j s j , (11) where L j = r j { y j − λ j / ( 1 + r j 2 ) } L_j=r_j\left\{y_j-\lambda_j/(1+r_j^2)\right\} L j = r j { y j − λ j / ( 1 + r j 2 ) } . Let ψ m ( t ) = E [ e i ⟨ t , X ( m ) ⟩ ] \psi_m(t)=\mathbb{E}[e^{i\langle t,X^{(m)}\rangle}] ψ m ( t ) = E [ e i ⟨ t , X ( m ) ⟩ ] , t ∈ R d t\in\mathbb{R}^d t ∈ R d , denote the characteristic function of X ( m ) X^{(m)} X ( m ) . Then ψ m ( t ) → ψ ( t ) \psi_m(t)\to\psi(t) ψ m ( t ) → ψ ( t ) as m → ∞ m\to\infty m → ∞ , where ψ ( t ) \psi(t) ψ ( t ) is the characteristic function of the Stable( α , Γ ) (\alpha,\Gamma) ( α , Γ ) distribution.
Figure 1 illustrates the bivariate stable distribution generated by the LePage construction (Theorem 3) with a von Mises spectral measure vM ( μ , κ ) \operatorname{vM}(\mu,\kappa) vM ( μ , κ ) . Each panel shows n = 500 n=500 n = 500 independent realisations of X i ( m ) X_i^{(m)} X i ( m ) , with m = 200 m=200 m = 200 terms, together with the polar density of the spectral measure (red curve) centred at the origin. The mean direction μ = π / 4 \mu=\pi/4 μ = π /4 is fixed throughout; the concentration parameter κ \kappa κ increases across columns (κ = 0.01 , 2 , 5 \kappa=0.01,2,5 κ = 0.01 , 2 , 5 ), and the stability index α \alpha α varies across rows (α = 1 \alpha=1 α = 1 , Cauchy; α = 1.99 \alpha=1.99 α = 1.99 , near-Gaussian). The directional structure of the LePage sum is governed jointly by the vM parameters ( μ , κ ) (\mu,\kappa) ( μ , κ ) through the spectral measure and by α \alpha α through the radial components r i j r_{ij} r ij .
Figure 1. Bivariate stable samples generated from the LePage construction (Theorem 3) with von Mises spectral measure vM ( μ , κ ) \operatorname{vM}(\mu,\kappa) vM ( μ , κ ) , μ = π / 4 \mu=\pi/4 μ = π /4 and Γ ∞ = 1 \Gamma_{\infty}=1 Γ ∞ = 1 . Each panel shows n = 500 n=500 n = 500 independent realisations of X i ( m ) X_i^{(m)} X i ( m ) with m = 200 m=200 m = 200 terms, truncated to r ∈ [ 0.1 , 10 6 ] r\in[0.1,10^6] r ∈ [ 0.1 , 1 0 6 ] . The red curve is the polar density of the vM spectral measure, scaled for display. Rows: stability index α = 1 \alpha=1 α = 1 (Cauchy, top) and α = 1.99 \alpha=1.99 α = 1.99 (near-Gaussian, bottom). Columns: concentration κ = 0.01 \kappa=0.01 κ = 0.01 (near-uniform), κ = 2 \kappa=2 κ = 2 (moderate), κ = 5 \kappa=5 κ = 5 (concentrated). Points beyond the 95th percentile of the radial component r i j r_{ij} r ij are excluded from display.
Construction of a Likelihood Function
Bayesian inference for Stable ( α , Γ ) \operatorname{Stable}(\alpha,\Gamma) Stable ( α , Γ ) requires a tractable likelihood. Existing density representations [36 , 1 ] do not factor as a joint density for ( X , latent variables ) (X,\text{latent variables}) ( X , latent variables ) , precluding the data-augmentation approach that Buckle [ [6 ] used successfully for univariate stable laws. We resolve this via a multivariate extension of Damien et al. [ [10 ] , combined with the characteristic function of Kuelbs [ [18 ] and the Gaussian error component in (9) , yielding a likelihood that supports full Bayesian inference.
Let { X 1 , … , X n } \{X_1,\ldots,X_n\} { X 1 , … , X n } be independently and identically distributed Stable ( α , Γ ) \operatorname{Stable}(\alpha,\Gamma) Stable ( α , Γ ) random vectors. Construct X i ( m ) X_i^{(m)} X i ( m ) using r i j r_{ij} r ij , y i j y_{ij} y ij , s i j s_{ij} s ij , j = 1 , … , m j=1,\ldots,m j = 1 , … , m , as in Theorem 3 and define X i ( m ) X_i^{(m)} X i ( m ) as in equation (11) , i = 1 , … , n i=1,\ldots,n i = 1 , … , n . Define ϵ i = X i − X i ( m ) \epsilon_i=X_i-X_i^{(m)} ϵ i = X i − X i ( m ) , i = 1 , … , n i=1,\ldots,n i = 1 , … , n . Given r i j r_{ij} r ij , y i j y_{ij} y ij , s i j s_{ij} s ij , j = 1 , … , m j=1,\ldots,m j = 1 , … , m , i = 1 , … , n i=1,\ldots,n i = 1 , … , n , we assume that ϵ 1 , … , ϵ n \epsilon_1,\ldots,\epsilon_n ϵ 1 , … , ϵ n follow independent bivariate Gaussian law with mean vector zero and dispersion matrix ( σ 2 / m ) I d (\sigma^2/m)I_d ( σ 2 / m ) I d ; that is,
ϵ i ∼ N d ( 0 , σ 2 m I d ) , i = 1 , … , n , (12) \epsilon_i \sim N_d\left(0,\frac{\sigma^2}{m}I_d\right),\quad i=1,\ldots,n, \tag*{(12)} ϵ i ∼ N d ( 0 , m σ 2 I d ) , i = 1 , … , n , (12)
where σ 2 \sigma^2 σ 2 is unknown and I d I_d I d is the identity matrix.
Deriving the likelihood function is now straightforward. By conditional independence of the latent variables ( r i j , y i j , s i j ) (r_{ij},y_{ij},s_{ij}) ( r ij , y ij , s ij ) given the parameters, the complete-data likelihood factors as
L ( α , η , κ ) = { ∏ i = 1 n ∏ j = 1 m 1 Γ ( α ) e − 1 / r i j r i j 1 + α } × { ∏ i = 1 n ∏ j = 1 m e − λ i j λ i j y i j y i j ! } × { ∏ i = 1 n ∏ j = 1 m f v M F ( s i j ∣ η , κ ) } × { ∏ i = 1 n f ( ϵ i ∣ α , η , κ , { r i j } , { y i j } , { s i j } ) } . (13) \begin{aligned}
L(\alpha,\eta,\kappa)={}&\left\{\prod_{i=1}^{n}\prod_{j=1}^{m}\frac{1}{\Gamma(\alpha)}\frac{e^{-1/r_{ij}}}{r_{ij}^{1+\alpha}}\right\}\times\left\{\prod_{i=1}^{n}\prod_{j=1}^{m}\frac{e^{-\lambda_{ij}}\lambda_{ij}^{y_{ij}}}{y_{ij}!}\right\}\\
&\times\left\{\prod_{i=1}^{n}\prod_{j=1}^{m}f_{\mathrm{vMF}}(s_{ij}\mid\eta,\kappa)\right\}\times\left\{\prod_{i=1}^{n}f(\epsilon_i\mid\alpha,\eta,\kappa,\{r_{ij}\},\{y_{ij}\},\{s_{ij}\})\right\}. \tag*{(13)}
\end{aligned} L ( α , η , κ ) = { i = 1 ∏ n j = 1 ∏ m Γ ( α ) 1 r ij 1 + α e − 1/ r ij } × { i = 1 ∏ n j = 1 ∏ m y ij ! e − λ ij λ ij y ij } × { i = 1 ∏ n j = 1 ∏ m f vMF ( s ij ∣ η , κ ) } × { i = 1 ∏ n f ( ϵ i ∣ α , η , κ , { r ij } , { y ij } , { s ij }) } . (13)
where λ i j = Γ ∞ Γ ( α ) e 1 / r i j / m \lambda_{ij}=\Gamma_\infty\Gamma(\alpha)e^{1/r_{ij}}/m λ ij = Γ ∞ Γ ( α ) e 1/ r ij / m .
Our approach towards estimating the spectral measure Γ \Gamma Γ is based on the likelihood function derived in equation (13) ; derivation of the likelihood function is given in Section C as online material. Inference based on approximate likelihood function is not uncommon especially in the absence of an explicit density function form. For instance, likelihood-based inference in frequency domain in time series is generally based on the Whittle likelihood which is the asymptotic distribution of the periodogram; see Choudhuri, Ghosal, and Roy (2003) and references therein for this example.
Let Ψ f = Γ ∞ Γ ( α ) \Psi_f=\Gamma_\infty\Gamma(\alpha) Ψ f = Γ ∞ Γ ( α ) denote the combined scale parameter. Based on Theorem 3 and the von Mises–Fisher spectral density, the complete-data log-likelihood decomposes as
ℓ c ( α , Ψ f , κ , η ) = ℓ P o i s ( Ψ f ) + ℓ G a u s s ( Ψ f , κ , η ) + ℓ I G ( α ) + ℓ v M F ( κ , η ) . (14) \ell_c(\alpha,\Psi_f,\kappa,\eta)=\ell_{\mathrm{Pois}}(\Psi_f)+\ell_{\mathrm{Gauss}}(\Psi_f,\kappa,\eta)+\ell_{\mathrm{IG}}(\alpha)+\ell_{\mathrm{vMF}}(\kappa,\eta). \tag*{(14)} ℓ c ( α , Ψ f , κ , η ) = ℓ Pois ( Ψ f ) + ℓ Gauss ( Ψ f , κ , η ) + ℓ IG ( α ) + ℓ vMF ( κ , η ) . (14)
Note that Γ ∞ \Gamma_\infty Γ ∞ enters ℓ c \ell_c ℓ c only through Ψ f \Psi_f Ψ f . For simulation studies, we use specific form of (14) with d = 2 d=2 d = 2 . In this case we can express s s s in terms of the directional angle θ \theta θ . With d = 2 d=2 d = 2 , individual components of the log-likelihood function in (14) are the following.
Poisson term: ℓ P o i s ( Ψ f ) = ∑ i , j [ y i j log λ i j − λ i j − log ( y i j ! ) ] , λ i j = Ψ f m exp ( 1 r i j ) \ell_{\mathrm{Pois}}(\Psi_f)=\sum_{i,j}[y_{ij}\log\lambda_{ij}-\lambda_{ij}-\log(y_{ij}!)],\ \lambda_{ij}=\frac{\Psi_f}{m}\exp\left(\frac{1}{r_{ij}}\right) ℓ Pois ( Ψ f ) = ∑ i , j [ y ij log λ ij − λ ij − log ( y ij !)] , λ ij = m Ψ f exp ( r ij 1 ) .
Gaussian term: ℓ G a u s s = − ( m / ( 2 σ 2 ) ) ∑ i = 1 n ∥ X i − X i ( m ) ∥ 2 \ell_{\mathrm{Gauss}}=-(m/(2\sigma^2))\sum_{i=1}^{n}\lVert X_i-X_i^{(m)}\rVert^2 ℓ Gauss = − ( m / ( 2 σ 2 )) ∑ i = 1 n ∥ X i − X i ( m ) ∥ 2 , where the series approximation is X i ( m ) = ∑ j = 1 m L i j s ( θ i j ) X_i^{(m)}=\sum_{j=1}^{m}L_{ij}s(\theta_{ij}) X i ( m ) = ∑ j = 1 m L ij s ( θ ij ) , L i j = r i j ( y i j − λ i j / ( 1 + r i j 2 ) ) L_{ij}=r_{ij}(y_{ij}-\lambda_{ij}/(1+r_{ij}^2)) L ij = r ij ( y ij − λ ij / ( 1 + r ij 2 )) , s ( θ ) = ( cos θ sin θ ) T s(\theta)=(\cos\theta\ \sin\theta)^{\mathsf{T}} s ( θ ) = ( cos θ sin θ ) T .
Inverted-gamma on r i j r_{ij} r ij : ℓ I G ( α ) = − n m log C δ ( α ) − ( 1 + α ) ∑ i , j log r i j − ∑ i , j ( 1 / r i j ) \ell_{\mathrm{IG}}(\alpha)=-nm\log C_\delta(\alpha)-(1+\alpha)\sum_{i,j}\log r_{ij}-\sum_{i,j}(1/r_{ij}) ℓ IG ( α ) = − nm log C δ ( α ) − ( 1 + α ) ∑ i , j log r ij − ∑ i , j ( 1/ r ij ) .
Von Mises on θ i j \theta_{ij} θ ij : ℓ v M ( κ , η ) = − n m log I 0 ( κ ) + κ ∑ i , j cos ( θ i j − η ) \ell_{\mathrm{vM}}(\kappa,\eta)=-nm\log I_0(\kappa)+\kappa\sum_{i,j}\cos(\theta_{ij}-\eta) ℓ vM ( κ , η ) = − nm log I 0 ( κ ) + κ ∑ i , j cos ( θ ij − η ) .
IG term ℓ I G ( α ) \ell_{\mathrm{IG}}(\alpha) ℓ IG ( α ) is C ∞ C^{\infty} C ∞ in α \alpha α since Γ ( α ) \Gamma(\alpha) Γ ( α ) is C ∞ C^{\infty} C ∞ on ( 0 , 2 ) (0,2) ( 0 , 2 ) . The Gaussian term depends on α \alpha α only through Γ ( α ) \Gamma(\alpha) Γ ( α ) in λ i j = Ψ f e 1 / r i j / m \lambda_{ij}=\Psi_f e^{1/r_{ij}}/m λ ij = Ψ f e 1/ r ij / m , which is likewise C ∞ C^{\infty} C ∞ . The practical consequence is that the MCMC sampler for α \alpha α requires no special treatment at α = 1 \alpha=1 α = 1 and the chain crosses this boundary freely. A stronger result — C ∞ C^{\infty} C ∞ smoothness of the characteristic function itself — is established in Theorem A1 of the Appendix.
A truncated version
Let 0 < r min < r max < ∞ 0<r_{\min}<r_{\max}<\infty 0 < r m i n < r m a x < ∞ be the lower and upper truncation thresholds for the radial jump sizes. The truncated inverted-gamma distribution IG ( α , r min , r max ) \operatorname{IG}(\alpha,r_{\min},r_{\max}) IG ( α , r m i n , r m a x ) has density
g ( r ∣ α , r min , r max ) = r − α − 1 e − 1 / r C r min , r max ( α ) , r ∈ ( r min , r max ) , (15) g(r\mid\alpha,r_{\min},r_{\max})=\frac{r^{-\alpha-1}e^{-1/r}}{C_{r_{\min},r_{\max}}(\alpha)},\qquad r\in(r_{\min},r_{\max}), \tag*{(15)} g ( r ∣ α , r m i n , r m a x ) = C r m i n , r m a x ( α ) r − α − 1 e − 1/ r , r ∈ ( r m i n , r m a x ) , (15)
where the normalising constant is given by
C r min , r max ( α ) = ∫ r min r max r − α − 1 e − 1 / r d r = Γ ( α ) [ P ( α , 1 / r min ) − P ( α , 1 / r max ) ] , (16) C_{r_{\min},r_{\max}}(\alpha)=\int_{r_{\min}}^{r_{\max}}r^{-\alpha-1}e^{-1/r}\,\mathrm{d}r=\Gamma(\alpha)\left[P(\alpha,1/r_{\min})-P(\alpha,1/r_{\max})\right], \tag*{(16)} C r m i n , r m a x ( α ) = ∫ r m i n r m a x r − α − 1 e − 1/ r d r = Γ ( α ) [ P ( α , 1/ r m i n ) − P ( α , 1/ r m a x ) ] , (16)
with P ( α , x ) = ∫ 0 x t α − 1 e − t d t / Γ ( α ) P(\alpha,x)=\int_{0}^{x}t^{\alpha-1}e^{-t}\,\mathrm{d}t/\Gamma(\alpha) P ( α , x ) = ∫ 0 x t α − 1 e − t d t /Γ ( α ) , the lower incomplete gamma function.
The truncated radial density at (15) is a computational regularisation of the full inverted gamma density (10) : the lower truncation at r min > 0 r_{\min}>0 r m i n > 0 removes radial components that are negligibly small in magnitude, while the upper truncation at r max < ∞ r_{\max}<\infty r m a x < ∞ guards against numerical overflow from the rare but large radial components in the upper tail of the inverted gamma density. As r min → 0 r_{\min}\to0 r m i n → 0 and r max → ∞ r_{\max}\to\infty r m a x → ∞ , the normalising constant C r min ( α ) → Γ ( α ) C_{r_{\min}}(\alpha)\to\Gamma(\alpha) C r m i n ( α ) → Γ ( α ) and the characteristic function of the truncated LePage sum converges to that of the untruncated sum (Proposition 4 in the Appendix). The stable limit ψ ( t ) \psi(t) ψ ( t ) of Theorem 3 is therefore recovered in the joint limit, confirming that truncation introduces no asymptotic bias. In practice, r min r_{\min} r m i n and r max r_{\max} r m a x are set as functions of the sample size m g e n m_{\mathrm{gen}} m gen . We use r min = 1 / log ( m g e n ) r_{\min}=1/\log(m_{\mathrm{gen}}) r m i n = 1/ log ( m gen ) and r max = m g e n 1 / α ^ r_{\max}=m_{\mathrm{gen}}^{1/\hat{\alpha}} r m a x = m gen 1/ α ^ , where α ^ \hat{\alpha} α ^ is the Koutrouvelis initialisation estimate of α \alpha α (§ 5.1 ). Nuances relating to m g e n m_{\mathrm{gen}} m gen are discussed in § 6.2 .
The Psi Reparametrization
For computational convenience, Γ ∞ \Gamma_{\infty} Γ ∞ is reparametrized as
Ψ = Γ ∞ C r min , r max ( α ) = mean Poisson intensity per term e 1 / r / m . (17) \Psi=\Gamma_{\infty}C_{r_{\min},r_{\max}}(\alpha)=\frac{\text{mean Poisson intensity per term}}{e^{1/r}/m}. \tag*{(17)} Ψ = Γ ∞ C r m i n , r m a x ( α ) = e 1/ r / m mean Poisson intensity per term . (17)
We can easily shift between Γ ∞ \Gamma_{\infty} Γ ∞ and Ψ \Psi Ψ through the conversion formula as Γ ∞ = Ψ / C r min , r max ( α ) \Gamma_{\infty}=\Psi/C_{r_{\min},r_{\max}}(\alpha) Γ ∞ = Ψ/ C r m i n , r m a x ( α ) .
Remark 4 (Notation). Throughout, Ψ f = Γ ∞ Γ ( α ) \Psi_f=\Gamma_{\infty}\Gamma(\alpha) Ψ f = Γ ∞ Γ ( α ) refers to the Poisson intensity of the full untruncated series (Theorem 3), while Ψ = Γ ∞ C r min , r max ( α ) \Psi=\Gamma_{\infty}C_{r_{\min},r_{\max}}(\alpha) Ψ = Γ ∞ C r m i n , r m a x ( α ) refers to the truncated version used in the likelihood. Implementation uses Ψ \Psi Ψ .
Remark 5 (Why Ψ \Psi Ψ is natural). The parameter Ψ \Psi Ψ appears directly in the Poisson intensities λ i j = ( Ψ / m ) e 1 / r i j \lambda_{ij}=(\Psi/m)e^{1/r_{ij}} λ ij = ( Ψ/ m ) e 1/ r ij and is identified by ℓ P o i s ( Ψ ) \ell_{\mathrm{Pois}}(\Psi) ℓ Pois ( Ψ ) independently of C r min , r max ( α ) C_{r_{\min},r_{\max}}(\alpha) C r m i n , r m a x ( α ) . The reparametrization from Γ ∞ \Gamma_{\infty} Γ ∞ to Ψ \Psi Ψ avoids the need to compute C r min , r max ( α ) C_{r_{\min},r_{\max}}(\alpha) C r m i n , r m a x ( α ) in the Poisson likelihood update, making the Gibbs step for Ψ \Psi Ψ a standard log-normal slice sampler.
Anatomy of construction
For d = 2 d=2 d = 2 , our approximate series representation is
X i ( m ) = ∑ j = 1 m L i j s ( θ i j ) , (18) X_i^{(m)}=\sum_{j=1}^{m}L_{ij}s(\theta_{ij}), \tag*{(18)} X i ( m ) = j = 1 ∑ m L ij s ( θ ij ) , (18)
where r i j ∼ IG ( α ) r_{ij}\sim\operatorname{IG}(\alpha) r ij ∼ IG ( α ) , y i j ∼ Poisson ( λ i j ) y_{ij}\sim\operatorname{Poisson}(\lambda_{ij}) y ij ∼ Poisson ( λ ij ) , θ i j ∼ vM ( η , κ ) \theta_{ij}\sim\operatorname{vM}(\eta,\kappa) θ ij ∼ vM ( η , κ ) , and λ i j = ( Ψ / m ) exp ( 1 / r i j ) \lambda_{ij}=(\Psi/m)\exp(1/r_{ij}) λ ij = ( Ψ/ m ) exp ( 1/ r ij ) . The first term r i j y i j r_{ij}y_{ij} r ij y ij is the raw jump: the product of the radial scale r j r_j r j and the discretised Poisson count y j y_j y j . The second term removes the expectation to ensure the series is centred, analogous to the centering in the classical LePage series for α ≥ 1 \alpha\geq1 α ≥ 1 . This is a truncated LePage series with four key modifications.
(i) Radial component r i j r_{ij} r ij : In the classical LePage series, the radial jumps are Γ j − 1 / α \Gamma_j^{-1/\alpha} Γ j − 1/ α , where Γ j = γ 1 + ⋯ + γ j \Gamma_j=\gamma_1+\cdots+\gamma_j Γ j = γ 1 + ⋯ + γ j are the arrival times of a unit-rate Poisson process (γ i ∼ i i d Exp ( 1 ) \gamma_i\mathrel{\mathop{\sim}^{\mathrm{iid}}}\operatorname{Exp}(1) γ i ∼ iid Exp ( 1 ) ). In our representation, r i j r_{ij} r ij plays the role of Γ j − 1 / α \Gamma_j^{-1/\alpha} Γ j − 1/ α ; its marginal distribution is the truncated inverted-gamma density at (15) .
(ii) Poisson bin counts y i j y_{ij} y ij : The classical LePage series sums over all Poisson arrivals individually. Theorem 3 partitions the intensity measure into m m m bins of equal width 1 / m 1/m 1/ m , and y i j ∼ Poisson ( λ i j ) y_{ij}\sim\operatorname{Poisson}(\lambda_{ij}) y ij ∼ Poisson ( λ ij ) counts the arrivals in the j j j -th bin for observation i i i . The expectation λ i j = ( Ψ / m ) e 1 / r i j \lambda_{ij}=(\Psi/m)e^{1/r_{ij}} λ ij = ( Ψ/ m ) e 1/ r ij gives the expected number of arrivals in that bin, connecting the bin width 1 / m 1/m 1/ m to the total jump intensity Ψ \Psi Ψ .
(iii) Centering λ i j / ( 1 + r i j 2 ) \lambda_{ij}/(1+r_{ij}^{2}) λ ij / ( 1 + r ij 2 ) : The classical LePage series requires centering for α ≥ 1 \alpha\geq1 α ≥ 1 to ensure convergence. The term − r i j λ i j / ( 1 + r i j 2 ) -r_{ij}\lambda_{ij}/(1+r_{ij}^{2}) − r ij λ ij / ( 1 + r ij 2 ) provides this centering. Each term L i j s ( θ i j ) L_{ij}s(\theta_{ij}) L ij s ( θ ij ) in equation (18) is a centred directional jump in R 2 \mathbb{R}^{2} R 2 .
(iv) The direction s ( θ i j ) s(\theta_{ij}) s ( θ ij ) : s ( θ i j ) = ( cos θ i j , sin θ i j ) T s(\theta_{ij})=(\cos\theta_{ij},\sin\theta_{ij})^{\mathsf{T}} s ( θ ij ) = ( cos θ ij , sin θ ij ) T is a unit vector on S 1 S^{1} S 1 drawn from the spectral measure vM ( η , κ ) \operatorname{vM}(\eta,\kappa) vM ( η , κ ) . It represents the tail direction of the j j j -th jump — the direction in which the j j j -th component pulls the sum.
Geometric interpretation: Each term L i j s ( θ i j ) L_{ij}s(\theta_{ij}) L ij s ( θ ij ) is a random vector in R 2 \mathbb{R}^{2} R 2 pointing in direction θ i j \theta_{ij} θ ij with signed magnitude L i j L_{ij} L ij . The sum ∑ j = 1 m L i j s ( θ i j ) \sum_{j=1}^{m}L_{ij}s(\theta_{ij}) ∑ j = 1 m L ij s ( θ ij ) accumulates these directional jumps, and by the convergence of the LePage series, the distribution of this sum approaches the bivariate stable distribution Stable ( α , Γ ) \operatorname{Stable}(\alpha,\Gamma) Stable ( α , Γ ) as m → ∞ m\to\infty m → ∞ . The Gaussian approximation error term ε i ∼ N 2 ( 0 , ( σ 2 / m ) I 2 ) \varepsilon_i\sim N_2(0,(\sigma^2/m)I_2) ε i ∼ N 2 ( 0 , ( σ 2 / m ) I 2 ) captures the contribution of the unrepresented tail beyond m m m terms.
Identification
From the inverted-gamma component of the likelihood function, the score for α \alpha α is
∂ ℓ c ∂ α = ∂ ℓ IG ∂ α = − n m C r min , r max ′ ( α ) C r min , r max ( α ) − ∑ i , j log r i j . (19) \frac{\partial\ell_c}{\partial\alpha}
=
\frac{\partial\ell_{\operatorname{IG}}}{\partial\alpha}
=
-nm\frac{C'_{r_{\min},r_{\max}}(\alpha)}{C_{r_{\min},r_{\max}}(\alpha)}
-
\sum_{i,j}\log r_{ij}. \tag*{(19)} ∂ α ∂ ℓ c = ∂ α ∂ ℓ IG = − nm C r m i n , r m a x ( α ) C r m i n , r m a x ′ ( α ) − i , j ∑ log r ij . (19)
The sufficient statistic for α \alpha α is ∑ i , j log r i j \sum_{i,j}\log r_{ij} ∑ i , j log r ij , identified from the radial draws { r i j } \{r_{ij}\} { r ij } independently of ( κ , η , Ψ ) (\kappa,\eta,\Psi) ( κ , η , Ψ ) . The shape of the truncated IG ( α , r min , r max ) \operatorname{IG}(\alpha,r_{\min},r_{\max}) IG ( α , r m i n , r m a x ) distribution changes with α \alpha α , and this shape change is precisely what makes α \alpha α identifiable from { r i j } \{r_{ij}\} { r ij } , as quantified by the strictly positive Fisher information in Proposition 2.
Proposition 2 (α \alpha α identified for all α ∈ ( 0 , 2 ) \alpha\in(0,2) α ∈ ( 0 , 2 ) ). The complete-data Fisher information for α \alpha α from the radial draws { r i j } \{r_{ij}\} { r ij } is
F α = n m [ C r min , r max ′ ′ ( α ) C r min , r max ( α ) − ( C r min , r max ′ ( α ) C r min , r max ( α ) ) 2 ] > 0 , α ∈ ( 0 , 2 ) . (20) \mathcal{F}_{\alpha}=nm\left[\frac{C^{\prime\prime}_{r_{\min},r_{\max}}(\alpha)}{C_{r_{\min},r_{\max}}(\alpha)}-\left(\frac{C^{\prime}_{r_{\min},r_{\max}}(\alpha)}{C_{r_{\min},r_{\max}}(\alpha)}\right)^2\right]>0,\qquad\alpha\in(0,2). \tag*{(20)} F α = nm [ C r m i n , r m a x ( α ) C r m i n , r m a x ′′ ( α ) − ( C r m i n , r m a x ( α ) C r m i n , r m a x ′ ( α ) ) 2 ] > 0 , α ∈ ( 0 , 2 ) . (20)
The strict positivity follows from the exponential family structure of the truncated IG ( α , r min , r max ) \operatorname{IG}(\alpha,r_{\min},r_{\max}) IG ( α , r m i n , r m a x ) distribution (Corollary 6, Appendix). Table 2 reports F α / n m \mathcal{F}_{\alpha}/nm F α / nm at selected values of α \alpha α , with r min = 0.1 r_{\min}=0.1 r m i n = 0.1 and r max = 10 6 r_{\max}=10^{6} r m a x = 1 0 6 .
Table 2. Complete-data Fisher information per observation F α / n m = Var α [ log r ] \mathcal{F}_{\alpha}/nm=\operatorname{Var}_{\alpha}[\log r] F α / nm = Var α [ log r ] at selected α \alpha α , with r min = 0.1 r_{\min}=0.1 r m i n = 0.1 , r max = 10 6 r_{\max}=10^{6} r m a x = 1 0 6 .
Remark 6 (Near-Gaussian identifiability). Table 2 shows that F α / n m \mathcal{F}_{\alpha}/nm F α / nm decreases from 244.949 at α = 0.25 \alpha=0.25 α = 0.25 to 0.060 at α = 1.99 \alpha=1.99 α = 1.99 , with the steepest drop between α = 1.0 \alpha=1.0 α = 1.0 and α = 1.5 \alpha=1.5 α = 1.5 . Beyond α = 1.75 \alpha=1.75 α = 1.75 the values cluster near zero, consistent with Theorem 4: as α → 2 \alpha\to2 α → 2 the stable law converges to a Gaussian, and α \alpha α becomes difficult to identify from the radial components alone. Simulation confirms wider posterior intervals for α ≥ 1.75 \alpha\ge1.75 α ≥ 1.75 .
The complete-data Fisher information F α \mathcal{F}_{\alpha} F α is decreasing in α \alpha α (Table 2 ), but the effective information reaching the sampler is not monotone: at small α \alpha α , the large radial components that carry most information about α \alpha α lie beyond r max r_{\max} r m a x and are not directly observed, so the sampler must impute them from a truncated distribution.
Initialisation and diagnostics
Initialisation
The initial value α ^ 0 \hat{\alpha}_0 α ^ 0 is obtained by applying the characteristic function regression of Koutrouvelis [ [17 ] to four univariate projections of the data (at 0 ∘ 0^\circ 0 ∘ , 45 ∘ 45^\circ 4 5 ∘ , 90 ∘ 90^\circ 9 0 ∘ , 135 ∘ 135^\circ 13 5 ∘ ), taking the median of valid estimates. The Koutrouvelis regression linearises log ( − log ∣ ϕ ^ n ( t ) ∣ 2 ) \log(-\log|\hat{\phi}_n(t)|^2) log ( − log ∣ ϕ ^ n ( t ) ∣ 2 ) in log ∣ t ∣ \log|t| log ∣ t ∣ to read off α \alpha α as a slope. This performs well when the data are genuinely heavy-tailed, so that the empirical characteristic function decays at a rate clearly distinct from the Gaussian. Near the boundary α → 2 \alpha\to2 α → 2 , however, the slope flattens and the estimate becomes unreliable, coinciding with a collapse of the Fisher information for α \alpha α ; see Table 2 . We therefore supplement the Koutrouvelis step with a near-normality check: if the sample excess kurtosis of the pooled projections falls below 1.0 — a threshold inconsistent with α < 1.8 \alpha< 1.8 α < 1.8 at n = 2000 n = 2000 n = 2000 — the regression is bypassed and α ^ 0 = 1.95 \hat{\alpha}_0 = 1.95 α ^ 0 = 1.95 is used instead. All other parameters are initialised from the same projections using moment-based estimates.
The spectral parameters ( η , κ ) (\eta, \kappa) ( η , κ ) are initialised from the unit-vector projections u i = X i / ∥ X i ∥ \mathbf{u}_i = \mathbf{X}_i/\lVert\mathbf{X}_i\rVert u i = X i / ∥ X i ∥ . The mean direction η ^ 0 \hat{\eta}_0 η ^ 0 is the sample circular mean and the concentration κ ^ 0 \hat{\kappa}_0 κ ^ 0 is obtained via the piecewise Best and Fisher [ [4 ] approximation to A − 1 ( R ^ 0 ) A^{-1}(\hat{R}_0) A − 1 ( R ^ 0 ) , where R ^ 0 \hat{R}_0 R ^ 0 is the sample mean resultant length. The total mass Γ ∞ \Gamma_\infty Γ ∞ is initialised by applying the Koutrouvelis scale estimator to each of eight equally spaced projection directions, computing Γ ^ ∞ ( k ) = ( γ ^ ( k ) ) α ^ 0 / I ^ ( u k ; α ^ 0 , η ^ 0 , κ ^ 0 ) \hat{\Gamma}_\infty^{(k)} = (\hat{\gamma}^{(k)})^{\hat{\alpha}_0}/\hat{I}(\mathbf{u}_k;\hat{\alpha}_0,\hat{\eta}_0,\hat{\kappa}_0) Γ ^ ∞ ( k ) = ( γ ^ ( k ) ) α ^ 0 / I ^ ( u k ; α ^ 0 , η ^ 0 , κ ^ 0 ) per direction, and taking the median of valid candidates. Full details of all initialisations are given in Appendix § 16 .
Convergence Diagnostics
We monitor convergence using two complementary diagnostics computed at each iteration. SSE ratio. The residual sum of squares S S E = ∑ i ∥ X i − X i ( m ) ∥ 2 \mathrm{SSE} = \sum_i \lVert\mathbf{X}_i - \mathbf{X}_i^{(m)}\rVert^2 SSE = ∑ i ∥ X i − X i ( m ) ∥ 2 , normalised by 2 n σ ^ 2 2n\hat{\sigma}^2 2 n σ ^ 2 , should equal 1 at convergence, since σ 2 \sigma^2 σ 2 is updated by a conjugate inverse-gamma step that absorbs the current residual variance. A value near 1 confirms internal self-consistency of the Gaussian approximation, but does not guarantee convergence of ( α , Ψ , η , κ ) (\alpha,\Psi,\eta,\kappa) ( α , Ψ , η , κ ) : if these are poorly estimated, σ 2 \sigma^2 σ 2 simply inflates to compensate.
logr gap. The sufficient statistic for α \alpha α under the truncated I G ( α ) \mathrm{IG}(\alpha) IG ( α ) distribution is log r ‾ = ∑ i , j log r i j / ( n m ) \overline{\log r} = \sum_{i,j}\log r_{ij}/(nm) log r = ∑ i , j log r ij / ( nm ) . At stationarity this should agree with its expected value ref_logr ( α ) = E I G ( α , δ , r u p p e r ) [ log r ] \operatorname{ref\_logr}(\alpha) = E_{\mathrm{IG}(\alpha,\delta,r_{\mathrm{upper}})}[\log r] ref_logr ( α ) = E IG ( α , δ , r upper ) [ log r ] under the current draw of α \alpha α . We monitor the gap
logr_gap = log r ‾ − ref_logr ( α ) , (21) \operatorname{logr\_gap} = \overline{\log r} - \operatorname{ref\_logr}(\alpha), \tag*{(21)} logr_gap = log r − ref_logr ( α ) , (21)
which serves as a real-time diagnostic for α \alpha α convergence; a persistent non-zero gap indicates that the chain has not found the correct stability index; also see Theorem 7 in the online material. Simulation showed that α \alpha α convergence does not automatically imply convergence of κ \kappa κ . We use n = 2000 n = 2000 n = 2000 observations with a chain of 50,000 iterations and a burn-in of 20,000 to ensure adequate mixing across all parameters. Correct convergence requires both SSE ratio ≈ 1 \approx1 ≈ 1 and logr-gap ≈ 0 \approx0 ≈ 0 simultaneously; SSE ratio ≈ 1 \approx1 ≈ 1 with logr-gap ≠ 0 \ne0 = 0 signals that σ 2 \sigma^2 σ 2 has inflated to absorb model misfit while α \alpha α remains biased.
Prior Distributions
All priors are uniform, reflecting broad uncertainty while excluding degenerate values:
α ∼ Uniform ( 10 − 4 , 1.999 ) , η ∼ Uniform ( 0 , 2 π ) , κ ∼ Uniform ( 0.1 , 50 ) , Γ ∞ ∼ Uniform ( 0.1 , 50 ) . \begin{aligned}
\alpha&\sim\operatorname{Uniform}(10^{-4}, 1.999),\\
\eta&\sim\operatorname{Uniform}(0, 2\pi),\\
\kappa&\sim\operatorname{Uniform}(0.1, 50),\\
\Gamma_{\infty} &\sim\operatorname{Uniform}(0.1, 50).
\end{aligned} α η κ Γ ∞ ∼ Uniform ( 1 0 − 4 , 1.999 ) , ∼ Uniform ( 0 , 2 π ) , ∼ Uniform ( 0.1 , 50 ) , ∼ Uniform ( 0.1 , 50 ) .
The upper bound α u p p e r = 1.999 \alpha_{\mathrm{upper}} = 1.999 α upper = 1.999 excludes the Gaussian boundary; the lower bound α l o w e r = 10 − 4 \alpha_{\mathrm{lower}} = 10^{-4} α lower = 1 0 − 4 excludes numerical degeneracy. The reparametrized mass Ψ = Γ ∞ C r min , r max ( α ) \Psi= \Gamma_{\infty} C_{r_{\min},r_{\max}}(\alpha) Ψ = Γ ∞ C r m i n , r m a x ( α ) inherits a uniform prior on ( 0.1 , 50 ) (0.1,50) ( 0.1 , 50 ) . The approximation variance σ 2 \sigma^2 σ 2 is assigned an informative inverse-gamma prior with mean near unity, reflecting the expectation that the Gaussian truncation error is of moderate size.
Simulation Study
Purpose
The simulation study has two objectives. First, to assess recovery of the stability index α \alpha α across the full range ( 0 , 2 ) (0,2) ( 0 , 2 ) , including the boundaries α → 0 \alpha\to0 α → 0 and α → 2 \alpha\to2 α → 2 . Second, misspecification with respect to m m m and r max r_{\max} r m a x .
Adaptive Truncation Parameters
The LePage series requires a lower truncation bound r min r_{\min} r m i n (minimum jump size) and an upper bound r max r_{\max} r m a x (maximum jump size). The series structure changes across α \alpha α . Near α = 0 \alpha= 0 α = 0 , the distribution is dominated by a few very large jumps requiring a large r max r_{\max} r m a x . Near α = 2 \alpha= 2 α = 2 , the Gaussian structure arises from accumulation of many small jumps requiring a large number of terms. For data generation, we set m g e n m_{\mathrm{gen}} m gen adaptively.
m g e n = max ( 30 , min ( 10 4 , ⌊ 30 + 5 α 2 + 15 ( 2 − α ) 2 ⌋ ) ) (22) m_{\mathrm{gen}} = \max\left(30,\min\left(10^4,\left\lfloor30+\frac{5}{\alpha^2}+\frac{15}{(2-\alpha)^2}\right\rfloor\right)\right) \tag*{(22)} m gen = max ( 30 , min ( 1 0 4 , ⌊ 30 + α 2 5 + ( 2 − α ) 2 15 ⌋ ) ) (22)
with truncation bounds r min = max ( 0.1 , 1 / ln ( m g e n ) ) r_{\min} = \max\left(0.1,1/\ln(m_{\mathrm{gen}})\right) r m i n = max ( 0.1 , 1/ ln ( m gen ) ) , r max = min ( m g e n 1 / α , 10 6 ) r_{\max} = \min\left(m_{\mathrm{gen}}^{1/\alpha},10^6\right) r m a x = min ( m gen 1/ α , 1 0 6 ) . The lower bound r min r_{\min} r m i n prevents numerical overflow in the Poisson intensity λ i j = ( Ψ / m ) e 1 / r i j \lambda_{ij} = (\Psi/m)e^{1/r_{ij}} λ ij = ( Ψ/ m ) e 1/ r ij . The adaptive value r min = 1 / ln ( m g e n ) r_{\min} = 1/\ln(m_{\mathrm{gen}}) r m i n = 1/ ln ( m gen ) is the smallest for which e 1 / r min ≈ m g e n e^{1/r_{\min}} \approx m_{\mathrm{gen}} e 1/ r m i n ≈ m gen remains within double precision. The upper bound r max = 10 6 r_{\max} = 10^6 r m a x = 1 0 6 with lower bound r min = 0.1 r_{\min} = 0.1 r m i n = 0.1 retains 98.97% of the IG(0.25) mass at α = 0.25 \alpha= 0.25 α = 0.25 , the most demanding case; for α ≥ 0.5 \alpha\ge0.5 α ≥ 0.5 the retained mass exceeds 99.9%. The small truncation loss at α = 0.25 \alpha= 0.25 α = 0.25 induces a minor positive bias in Γ ^ ∞ \widehat{\Gamma}_{\infty} Γ ∞ , consistent with the simulation results in Table 3 .
Table 3 : Simulation results for the bivariate stable–von Mises model (n = 2000 n = 2000 n = 2000 , m = 30 m = 30 m = 30 , κ 0 = 2 \kappa_{0} = 2 κ 0 = 2 , Γ ∞ , 0 = 1 \Gamma_{\infty,0} = 1 Γ ∞ , 0 = 1 , η 0 = π / 4 \eta_{0} = \pi/4 η 0 = π /4 ; 10 replications). Bias = posterior mean minus truth (circular wrapping for η \eta η ). Row marked κ 4 \kappa^{4} κ 4 use the kurtosis-based initialisation (§ 5.1 ); all other rows use Koutrouvelis ECF initialisation. Row BVN: bivariate normal data with covariance matrix at (7) .
Table 3. Simulation results for the bivariate stable–von Mises model.
Remark 7 (m g e n m_{\mathrm{gen}} m gen versus m m m ). m g e n m_{\mathrm{gen}} m gen is used solely for data generation and varies with α \alpha α . The MCMC uses a fixed m = 30 m = 30 m = 30 terms for all the runs, with r min r_{\min} r m i n and r max r_{\max} r m a x fixed at the values corresponding to the true α \alpha α .
Simulation Designs
Main simulation. LePage series data with n = 2000 n = 2000 n = 2000 observations, m = 30 m = 30 m = 30 terms, across α ∈ { 0.25 , 0.50 , 0.75 , 1.00 , 1.25 , 1.50 , 1.75 , 1.99 } \alpha\in\{0.25, 0.50, 0.75, 1.00, 1.25, 1.50, 1.75, 1.99\} α ∈ { 0.25 , 0.50 , 0.75 , 1.00 , 1.25 , 1.50 , 1.75 , 1.99 } , with Γ ∞ = 1 \Gamma_\infty= 1 Γ ∞ = 1 , η = π / 2 \eta= \pi/2 η = π /2 , κ = 2 \kappa= 2 κ = 2 , σ 2 = 1 \sigma^2 = 1 σ 2 = 1 . Ten replications per α \alpha α value, each with 50,000 MCMC iterations, 20,000 burn-in, thinning by 5, giving 6,000 post-burnin samples.
BVN comparison. At α = 1.99 \alpha= 1.99 α = 1.99 , ten additional runs use exact bivariate Gaussian data X i ∼ N 2 ( 0 , Σ Γ ) X_i \sim N_2(0, \Sigma_\Gamma) X i ∼ N 2 ( 0 , Σ Γ ) with Σ Γ = Γ ∞ ( I 2 + R 2 ( κ ) A ( η ) ) \Sigma_\Gamma= \Gamma_\infty(I_2 + R_2(\kappa)A(\eta)) Σ Γ = Γ ∞ ( I 2 + R 2 ( κ ) A ( η )) , bypassing the LePage series entirely. This comparison isolates inference algorithm performance from data generation approximation error.
Simulation Results
Table 3 summarises posterior performance across the α \alpha α grid (n = 2000 n = 2000 n = 2000 , m = 30 m = 30 m = 30 , 10 replications).
Stability index α \alpha α . For α ∈ [ 0.75 , 1.75 ] \alpha\in[0.75, 1.75] α ∈ [ 0.75 , 1.75 ] the bias satisfies ∣ α ^ − α ∣ ≤ 0.036 |\hat{\alpha} - \alpha| \leq0.036 ∣ α ^ − α ∣ ≤ 0.036 and coverage ranges from 0.5 to 0.9, consistent with the moderate Fisher information in this region (Table 2 ). At α = 0.25 \alpha= 0.25 α = 0.25 there is a small upward bias of +0.036; at α = 1.99 \alpha= 1.99 α = 1.99 the downward bias reaches − 0.105 -0.105 − 0.105 , reflecting the near-zero Fisher information near the Gaussian boundary (§ 4 ).
Mean direction η \eta η . The mean direction is the best-recovered parameter: bias ∣ η ^ − η 0 ∣ < 0.022 |\hat{\eta} - \eta_0| < 0.022 ∣ η ^ − η 0 ∣ < 0.022 for all α ∈ [ 0.25 , 1.99 ] \alpha\in[0.25, 1.99] α ∈ [ 0.25 , 1.99 ] , confirming that the von Mises mean direction is identifiable independently of the stability index.
Concentration κ \kappa κ . For α ≤ 0.50 \alpha\leq0.50 α ≤ 0.50 the posterior mean of κ \kappa κ is substantially below the true value (κ ^ ≈ 1.6 \hat{\kappa} \approx1.6 κ ^ ≈ 1.6 vs. κ 0 = 2 \kappa_0 = 2 κ 0 = 2 ). This reflects the difficulty of identifying the spectral concentration when the distribution is dominated by a few very large jumps. For α ∈ [ 0.75 , 1.99 ] \alpha\in[0.75, 1.99] α ∈ [ 0.75 , 1.99 ] estimation improves markedly, with ∣ Bias κ ∣ ≤ 0.079 |\operatorname{Bias}_{\kappa}| \leq0.079 ∣ Bias κ ∣ ≤ 0.079 .
Mass parameter Γ ∞ \Gamma_\infty Γ ∞ . For α ∈ [ 0.75 , 1.75 ] \alpha\in[0.75, 1.75] α ∈ [ 0.75 , 1.75 ] the bias in Γ ^ ∞ \hat{\Gamma}_\infty Γ ^ ∞ is negligible (∣ Bias Γ ∞ ∣ ≤ 0.025 |\operatorname{Bias}_{\Gamma_\infty}| \leq0.025 ∣ Bias Γ ∞ ∣ ≤ 0.025 ). At α = 0.25 \alpha= 0.25 α = 0.25 there is a systematic upward bias of +0.831, consistent across all ten replications (range +0.43 to +1.34). As shown in § 6.2 , the upper truncation bound r max = 10 6 r_{\max} = 10^6 r m a x = 1 0 6 retains 98.97% of the IG ( 0.25 ) \operatorname{IG}(0.25) IG ( 0.25 ) mass; the residual truncation loss causes the sampler to overestimate Ψ = Γ ∞ Γ ( α ) \Psi= \Gamma_\infty\Gamma(\alpha) Ψ = Γ ∞ Γ ( α ) to compensate. This bias attenuates rapidly and is negligible by α = 0.75 \alpha= 0.75 α = 0.75 .
Near-Gaussian boundary check (BVN). As a boundary check, we apply the sampler to data generated from a bivariate normal distribution (α = 1.99 \alpha= 1.99 α = 1.99 , κ 0 = 2 \kappa_0 = 2 κ 0 = 2 ). With ECF based initialisation for α \alpha α , the posterior mean α ^ = 1.940 \hat{\alpha} = 1.940 α ^ = 1.940 is close to 1.99 with coverage 0.9, confirming that the sampler does not catastrophically misidentify Gaussian data; kurtosis based initialisation recovers alpha. However, κ \kappa κ is severely under-estimated (κ ^ = 0.184 \hat{\kappa} = 0.184 κ ^ = 0.184 , Bias κ = − 1.816 \operatorname{Bias}_{\kappa} = -1.816 Bias κ = − 1.816 ) and Γ ∞ \Gamma_\infty Γ ∞ is biased downward (Γ ^ ∞ = 0.474 \hat{\Gamma}_\infty= 0.474 Γ ^ ∞ = 0.474 ). The large bias in η ^ \hat{\eta} η ^ (+0.847) is expected: when κ ≈ 0 \kappa\approx0 κ ≈ 0 the von Mises distribution is nearly uniform and η \eta η is unidentifiable.
Sensitivity with respect to m m m and r max r_{\max} r m a x . Table 4 reports posterior means and biases for α = 0.25 \alpha= 0.25 α = 0.25 and α = 1.99 \alpha= 1.99 α = 1.99 across six combinations of m ∈ { 30 , 50 } m \in\{30, 50\} m ∈ { 30 , 50 } and r max ∈ { 10 4 , 10 6 , 10 8 } r_{\max} \in\{10^4, 10^6, 10^8\} r m a x ∈ { 1 0 4 , 1 0 6 , 1 0 8 } . The directional parameter η \eta η is essentially insensitive to both tuning choices, with biases no larger than 0.021 across all twelve cells. For α \alpha α and κ \kappa κ , increasing r max r_{\max} r m a x from 10 4 10^4 1 0 4 to 10 6 10^6 1 0 6 yields a noticeable reduction in bias at α = 0.25 \alpha= 0.25 α = 0.25 : the bias in α ^ \hat{\alpha} α ^ falls from +0.102 to +0.036 and Γ ^ ∞ \hat{\Gamma}_\infty Γ ^ ∞ improves substantially, reflecting the importance of an adequate radial truncation for κ 4 \kappa^{4} κ 4 If the sample excess kurtosis of the pooled projections falls below 1.0, the Koutrouvelis regression is bypassed and α ^ 0 = 1.95 \hat{\alpha}_{0} = 1.95 α ^ 0 = 1.95 is used. For the BVN row without this guard, one replication received fallback option with α ^ 0 = 0.70 \hat{\alpha}_{0} = 0.70 α ^ 0 = 0.70 and failed to converge (α ^ = 1.530 \hat{\alpha} = 1.530 α ^ = 1.530 after 50,000 iterations); the guard eliminates this failure, giving α ^ = 1.993 \hat{\alpha} = 1.993 α ^ = 1.993 .
Table 4. Sensitivity of posterior means to the number of LePage terms m m m and the radial truncation r max r_{\max} r m a x (n = 2000 n = 2000 n = 2000 , κ 0 = 2 \kappa_0 = 2 κ 0 = 2 , Γ ∞ , 0 = 1 \Gamma_{\infty,0} = 1 Γ ∞ , 0 = 1 , η 0 = π / 2 \eta_0 = \pi/2 η 0 = π /2 ; 10 replications). Bias = posterior mean − truth. The standard setting used in Table 3 is m = 30 m = 30 m = 30 , r max = 10 6 r_{\max} = 10^{6} r m a x = 1 0 6 (marked †).
heavy-tailed data. Further increasing r max r_{\max} r m a x to 10 8 10^{8} 1 0 8 gives only marginal additional improvement. At α = 1.99 \alpha= 1.99 α = 1.99 the dependence on r max r_{\max} r m a x is weaker and non-monotone, consistent with the lighter tails near the Gaussian boundary. Increasing m m m from 30 to 50 produces small and inconsistent changes throughout: biases shift by at most 0.010 in either direction with no systematic gain. The standard setting m = 30 m = 30 m = 30 , r max = 10 6 r_{\max} = 10^{6} r m a x = 1 0 6 therefore offers a reasonable balance between accuracy and computational cost and is used throughout the main analysis.
Smooth transition at the Cauchy boundary. Figure 2 displays the MCMC trace for one replication at true α = 1.00 \alpha= 1.00 α = 1.00 (n = 2,000 n = 2{,}000 n = 2 , 000 , m = 30 m = 30 m = 30 ). The upper panel traces the sampled α \alpha α values across 50,000 iterations; the chain crosses the Cauchy boundary α = 1 \alpha= 1 α = 1 a total of 190 times in the post-burn-in period, confirming that the sampler explores both sides of this boundary freely and does not become trapped in a sub-region of the parameter space. The lower panel shows the gap log r t − ref log r ( α t ) \log r_t - \operatorname{ref}_{\log r}(\alpha_t) log r t − ref l o g r ( α t ) , where log r t \log r_t log r t is the mean log-radial component at iteration t t t and ref log r ( α ) = E α [ log r ] \operatorname{ref}_{\log r}(\alpha) = \mathbb{E}_{\alpha}[\log r] ref l o g r ( α ) = E α [ log r ] is its expected value under the truncated inverse-gamma distribution at the current α \alpha α . Because ref log r ( α ) \operatorname{ref}_{\log r}(\alpha) ref l o g r ( α ) is monotone increasing in α \alpha α , the two panels move in opposition: a downward excursion of α \alpha α below the truth reduces the reference value by less than it reduces log r ‾ \overline{\log r} log r , driving the gap negative, while an upward excursion drives it positive; see also Theorem 7 in the online material.
Figure 2. MCMC trace for replicate 201 (true α = 1.00 \alpha= 1.00 α = 1.00 , n = 2,000 n = 2{,}000 n = 2 , 000 , m = 30 m = 30 m = 30 ). Top: sampled α \alpha α ; dashed line is the truth; vertical line separates burn-in (20,000 iterations). Bottom: gap log r t − ref log r ( α t ) \log r_t - \operatorname{ref}_{\log r}(\alpha_t) log r t − ref l o g r ( α t ) ; see text for interpretation.
Figure 3 plots the sample excess kurtosis of one marginal projection of LePage-generated data against α \alpha α , based on 50 replications at each grid point with n = 2000 n = 2000 n = 2000 and adaptive m m m . Across the entire range α ∈ { 0.25 , 0.50 , … , 1.99 } \alpha\in\{0.25, 0.50, \ldots, 1.99\} α ∈ { 0.25 , 0.50 , … , 1.99 } the median excess kurtosis remains far above both the Gaussian benchmark of zero and the guard threshold of 1.0, the smallest median value being approximately 20 at α = 1.50 \alpha= 1.50 α = 1.50 . This confirms that LePage samples at finite m m m do not reproduce the near-Gaussian tail behaviour of truly Gaussian data even at the boundary α = 1.99 \alpha= 1.99 α = 1.99 . The upturn in both the median and the band width at α = 1.99 \alpha= 1.99 α = 1.99 reflects increasing variability of the LePage sum near the boundary.
Figure 3. Sample excess kurtosis of one marginal projection of LePage-generated bivariate stable data, as a function of the stability index α \alpha α (n = 2000 n = 2000 n = 2000 , m m m adaptive, 50 replications). Points show the median across replications; the shaded band spans the 10th to 90th percentiles. The dashed line marks the Gaussian benchmark (excess kurtosis = 0) and the dotted line marks the guard threshold of 1.0 used to detect near-Gaussianity.
Discussion
We have developed a full Bayesian inference framework for α \alpha α -stable distributions with von Mises–Fisher spectral measure in arbitrary dimension d d d , grounded in the LePage series representation with latent radial, spectral, and Poisson components. The likelihood, Gibbs sampler, and logr_gap diagnostic are derived for general d d d ; the bivariate case d = 2 d = 2 d = 2 with the circular von Mises measure serves as the simulation testbed.
The sampler recovers all four parameters reliably across the full range of α \alpha α , including the heavy-tailed extreme α = 0.25 \alpha= 0.25 α = 0.25 . Continuity of the likelihood at the Cauchy boundary α = 1 \alpha= 1 α = 1 is established theoretically and confirmed empirically by free chain crossing in both directions. Near the Gaussian boundary α = 2 \alpha= 2 α = 2 , Fisher information for α \alpha α collapses and ECF-based initialisation can fail; the kurtosis-based guard provides a data-adaptive remedy that eliminates catastrophic non-convergence without affecting inference elsewhere.
Extension to d > 2 d > 2 d > 2 requires no structural change: the radial and Poisson draws are dimension-free, and the vMF spectral measure on S d − 1 \mathbb{S}^{d-1} S d − 1 replaces the circular von Mises distribution directly. Only the projection integral requires numerical evaluation for $d > 2.
We have shown that the von Mises–Fisher spectral measure is supported by three independent lines of argument. The regular variation characterisation (3) shows that κ \kappa κ governs the angular concentration of extreme observations independently of α \alpha α , giving the vMF family a geometrically transparent parametrisation of tail anisotropy. Proposition 1 shows that the same κ \kappa κ determines the eigenvalue ratio of the Gaussian-boundary covariance, linking spectral concentration monotonically to the shape of probability contours. Finally, the hitting distribution of Brownian motion on S d − 1 S^{d-1} S d − 1 from an interior point is exactly vMF [15 , 11 ] , and the associated pole-seeking diffusion [15 ] has vMF as its stationary distribution, providing a classical probabilistic foundation for spectral concentration. Taken together, these connections place the vMF spectral measure at the intersection of multivariate stable modelling and directional statistics.
Disclaimer
Views expressed are purely that of the author and do not reflect that of the organisation with which author is associated with.
AI Policy Statement by the authors
We have used Generative AI tools of Claude (Anthropic), version Claude Sonnet 4.6, accessed via claude.ai. The AI assistant was used for LaTeX formatting, equation typesetting, document preparation, grammar and style checking, R code debugging, formatting of simulation scripts, and, generation of initial drafts of some remarks and captions. All scientific content, theoretical results, simulation design, interpretation of results, and final decisions on all aspects of the manuscript are solely the work of the authors. The AI tool did not contribute to the intellectual content of the paper.
Proof of Gaussian Boundary Theorem
Theorem 4 (Gaussian boundary, general d d d ). Let X ∼ Stable ( α , Γ ) X \sim\operatorname{Stable}(\alpha,\Gamma) X ∼ Stable ( α , Γ ) in R d \mathbb{R}^d R d , α ∈ ( 0 , 2 ) \alpha\in(0,2) α ∈ ( 0 , 2 ) , with finite spectral measure Γ \Gamma Γ on S d − 1 S^{d-1} S d − 1 and total mass Γ ∞ = Γ ( S d − 1 ) < ∞ \Gamma_\infty= \Gamma(S^{d-1}) < \infty Γ ∞ = Γ ( S d − 1 ) < ∞ . As α → 2 \alpha\to2 α → 2 ,
X → d N d ( 0 , Σ Γ ) , Σ Γ = 2 Γ ∞ ∫ S d − 1 s s T Γ ~ ( d s ) . (23) X \xrightarrow{d} N_d(0,\Sigma_\Gamma), \qquad\Sigma_\Gamma= 2\Gamma_\infty\int_{S^{d-1}} ss^\mathsf{T}\widetilde{\Gamma}(ds). \tag*{(23)} X d N d ( 0 , Σ Γ ) , Σ Γ = 2 Γ ∞ ∫ S d − 1 s s T Γ ( d s ) . (23) The limiting covariance Σ Γ \Sigma_\Gamma Σ Γ is determined by the second moment of the normalised spectral measure Γ ~ \widetilde{\Gamma} Γ and the total spectral mass Γ ∞ \Gamma_\infty Γ ∞ ; no further properties of Γ \Gamma Γ are required.
Proof. Step 1. Characteristic function. The characteristic function of an α \alpha α -stable distribution in R d \mathbb{R}^d R d with spectral measure Γ \Gamma Γ is
log φ X ( t ) = − ∫ S d − 1 ∣ t T s ∣ α [ 1 − i sgn ( t T s ) tan ( π α 2 ) ] Γ ( d s ) , t ∈ R d , (24) \log\varphi_X(t) = -\int_{S^{d-1}} |t^\mathsf{T}s|^\alpha\left[1-i\operatorname{sgn}(t^\mathsf{T}s)\tan\left(\frac{\pi\alpha}{2}\right)\right]\Gamma(ds), \qquad t \in\mathbb{R}^d, \tag*{(24)} log φ X ( t ) = − ∫ S d − 1 ∣ t T s ∣ α [ 1 − i sgn ( t T s ) tan ( 2 π α ) ] Γ ( d s ) , t ∈ R d , (24)
[32 Theorem 2.3.1] , valid for all α ∈ ( 0 , 2 ) \alpha\in(0,2) α ∈ ( 0 , 2 ) and any finite spectral measure Γ \Gamma Γ .
Step 1b. Skewness term vanishes as α → 2 \alpha\to2 α → 2 . Since ∣ sgn ( t T s ) ∣ ≤ 1 |\operatorname{sgn}(t^\mathsf{T}s)| \le1 ∣ sgn ( t T s ) ∣ ≤ 1 for all s ∈ S d − 1 s \in S^{d-1} s ∈ S d − 1 and ∣ t T s ∣ α ≤ ∥ t ∥ α |t^\mathsf{T}s|^\alpha\le\|t\|^\alpha ∣ t T s ∣ α ≤ ∥ t ∥ α on S d − 1 S^{d-1} S d − 1 ,
∣ ∫ S d − 1 ∣ t T s ∣ α sgn ( t T s ) Γ ( d s ) ∣ ≤ ∥ t ∥ α Γ ∞ < ∞ . (25) \left|\int_{S^{d-1}} |t^\mathsf{T}s|^\alpha\operatorname{sgn}(t^\mathsf{T}s)\Gamma(ds)\right| \le\|t\|^\alpha\Gamma_\infty< \infty. \tag*{(25)} ∫ S d − 1 ∣ t T s ∣ α sgn ( t T s ) Γ ( d s ) ≤ ∥ t ∥ α Γ ∞ < ∞. (25)
Since tan ( π α / 2 ) → tan ( π ) = 0 \tan(\pi\alpha/2) \to\tan(\pi) = 0 tan ( π α /2 ) → tan ( π ) = 0 as α → 2 \alpha\to2 α → 2 , the imaginary part of log φ X ( t ) \log\varphi_X(t) log φ X ( t ) satisfies
tan ( π α 2 ) ∫ S d − 1 ∣ t T s ∣ α sgn ( t T s ) Γ ( d s ) ⟶ 0 as α → 2. (26) \tan\left(\frac{\pi\alpha}{2}\right)\int_{S^{d-1}} |t^\mathsf{T}s|^\alpha\operatorname{sgn}(t^\mathsf{T}s)\Gamma(ds) \longrightarrow0 \qquad\text{as } \alpha\to2. \tag*{(26)} tan ( 2 π α ) ∫ S d − 1 ∣ t T s ∣ α sgn ( t T s ) Γ ( d s ) ⟶ 0 as α → 2. (26)
Consequently, the limit of log φ X ( t ) \log\varphi_X(t) log φ X ( t ) as α → 2 \alpha\to2 α → 2 is determined by the real part:
log φ X ( t ) ∼ − ∫ S d − 1 ∣ t T s ∣ α Γ ( d s ) as α → 2. (27) \log\varphi_X(t) \sim-\int_{S^{d-1}} |t^\mathsf{T}s|^\alpha\Gamma(ds) \qquad\text{as } \alpha\to2. \tag*{(27)} log φ X ( t ) ∼ − ∫ S d − 1 ∣ t T s ∣ α Γ ( d s ) as α → 2. (27)
Step 2. Pointwise limit of the integrand. Fix t ∈ R d t \in\mathbb{R}^d t ∈ R d and s ∈ S d − 1 s \in S^{d-1} s ∈ S d − 1 , and write u = ∣ t T s ∣ ≥ 0 u = |t^\mathsf{T}s| \ge0 u = ∣ t T s ∣ ≥ 0 . Since u α = exp ( α log u ) u^\alpha= \exp(\alpha\log u) u α = exp ( α log u ) is continuous in α \alpha α for each fixed u ≥ 0 u \ge0 u ≥ 0 ,
∣ t T s ∣ α ⟶ ∣ t T s ∣ 2 = ( t T s ) 2 as α → 2 , (28) |t^\mathsf{T}s|^\alpha\longrightarrow|t^\mathsf{T}s|^2 = (t^\mathsf{T}s)^2 \qquad\text{as } \alpha\to2, \tag*{(28)} ∣ t T s ∣ α ⟶ ∣ t T s ∣ 2 = ( t T s ) 2 as α → 2 , (28)
pointwise for every s ∈ S d − 1 s \in S^{d-1} s ∈ S d − 1 .
Step 3. Dominated convergence. For α ∈ [ 1 , 2 ] \alpha\in[1,2] α ∈ [ 1 , 2 ] , since ∥ s ∥ = 1 \|s\| = 1 ∥ s ∥ = 1 for all s ∈ S d − 1 s \in S^{d-1} s ∈ S d − 1 ,
∣ t T s ∣ α ≤ ∣ t T s ∣ + ∣ t T s ∣ 2 ≤ ∥ t ∥ + ∥ t ∥ 2 . (29) |t^\mathsf{T}s|^\alpha\le|t^\mathsf{T}s| + |t^\mathsf{T}s|^2 \le\|t\| + \|t\|^2. \tag*{(29)} ∣ t T s ∣ α ≤ ∣ t T s ∣ + ∣ t T s ∣ 2 ≤ ∥ t ∥ + ∥ t ∥ 2 . (29)
This bound is integrable with respect to Γ \Gamma Γ since Γ \Gamma Γ is finite:
∫ S d − 1 ( ∣ t T s ∣ + ∣ t T s ∣ 2 ) Γ ( d s ) ≤ ( ∥ t ∥ + ∥ t ∥ 2 ) Γ ∞ < ∞ . (30) \int_{S^{d-1}} \left(\lvert t^{\mathsf{T}}s\rvert+ \lvert t^{\mathsf{T}}s\rvert^{2}\right)\Gamma(ds) \le\left(\lVert t\rVert+ \lVert t\rVert^{2}\right)\Gamma_{\infty} < \infty. \tag*{(30)} ∫ S d − 1 ( ∣ t T s ∣ + ∣ t T s ∣ 2 ) Γ ( d s ) ≤ ( ∥ t ∥ + ∥ t ∥ 2 ) Γ ∞ < ∞. (30)
By the Dominated Convergence Theorem,
lim α → 2 ∫ S d − 1 ∣ t T s ∣ α Γ ( d s ) = ∫ S d − 1 ( t T s ) 2 Γ ( d s ) . (31) \lim_{\alpha\to2}\int_{S^{d-1}}\lvert t^{\mathsf{T}}s\rvert^{\alpha}\Gamma(ds)=\int_{S^{d-1}}\left(t^{\mathsf{T}}s\right)^{2}\Gamma(ds). \tag*{(31)} α → 2 lim ∫ S d − 1 ∣ t T s ∣ α Γ ( d s ) = ∫ S d − 1 ( t T s ) 2 Γ ( d s ) . (31)
Step 4. Evaluate the limiting integral. Write ( t T s ) 2 = t T s s T t \left(t^{\mathsf{T}}s\right)^{2}=t^{\mathsf{T}}ss^{\mathsf{T}}t ( t T s ) 2 = t T s s T t and interchange integration and the bilinear form:
∫ S d − 1 ( t T s ) 2 Γ ( d s ) = t T ( ∫ S d − 1 s s T Γ ( d s ) ) t = t T M t , (32) \int_{S^{d-1}}\left(t^{\mathsf{T}}s\right)^{2}\Gamma(ds)=t^{\mathsf{T}}\left(\int_{S^{d-1}}ss^{\mathsf{T}}\Gamma(ds)\right)t=t^{\mathsf{T}}Mt, \tag*{(32)} ∫ S d − 1 ( t T s ) 2 Γ ( d s ) = t T ( ∫ S d − 1 s s T Γ ( d s ) ) t = t T M t , (32)
where
M = ∫ S d − 1 s s T Γ ( d s ) (33) M=\int_{S^{d-1}}ss^{\mathsf{T}}\Gamma(ds) \tag*{(33)} M = ∫ S d − 1 s s T Γ ( d s ) (33)
is a d × d d\times d d × d positive semidefinite matrix (positive definite whenever Γ \Gamma Γ is not supported on a proper subspace of R d \mathbb{R}^{d} R d ).
Step 5. Identify the limiting characteristic function. Combining Steps 1b–4,
lim α → 2 log φ X ( t ) = − t T M t . (34) \lim_{\alpha\to2}\log\varphi_X(t)=-t^{\mathsf{T}}Mt. \tag*{(34)} α → 2 lim log φ X ( t ) = − t T M t . (34)
For Y ∼ N d ( 0 , Σ ) Y\sim N_d(0,\Sigma) Y ∼ N d ( 0 , Σ ) , log φ Y ( t ) = − 1 2 t T Σ t \log\varphi_Y(t)=-\frac{1}{2}t^{\mathsf{T}}\Sigma t log φ Y ( t ) = − 2 1 t T Σ t , so (34) is the log characteristic function of N d ( 0 , 2 M ) N_d(0,2M) N d ( 0 , 2 M ) .
Step 6. Convergence in distribution. Since φ X ( t ) → exp ( − t T M t ) \varphi_X(t)\to\exp(-t^{\mathsf{T}}Mt) φ X ( t ) → exp ( − t T M t ) pointwise for all t ∈ R d t\in\mathbb{R}^{d} t ∈ R d , and the limit is the characteristic function of N d ( 0 , 2 M ) N_d(0,2M) N d ( 0 , 2 M ) , which is continuous at t = 0 t=0 t = 0 , Lévy’s continuity theorem gives
X → d N d ( 0 , Σ Γ ) , Σ Γ = 2 M = 2 ∫ S d − 1 s s T Γ ( d s ) . (35) X \xrightarrow{d} N_d(0,\Sigma_{\Gamma}),\qquad\Sigma_{\Gamma}=2M=2\int_{S^{d-1}}ss^{\mathsf{T}}\Gamma(ds). \tag*{(35)} X d N d ( 0 , Σ Γ ) , Σ Γ = 2 M = 2 ∫ S d − 1 s s T Γ ( d s ) . (35)
Writing Γ = Γ ∞ Γ ~ \Gamma=\Gamma_{\infty}\widetilde{\Gamma} Γ = Γ ∞ Γ with Γ ~ ( S d − 1 ) = 1 \widetilde{\Gamma}(S^{d-1})=1 Γ ( S d − 1 ) = 1 , we have Σ Γ = 2 Γ ∞ ∫ S d − 1 s s T Γ ~ ( d s ) \Sigma_{\Gamma}=2\Gamma_{\infty}\int_{S^{d-1}}ss^{\mathsf{T}}\widetilde{\Gamma}(ds) Σ Γ = 2 Γ ∞ ∫ S d − 1 s s T Γ ( d s ) , confirming that the limit is determined by Γ ∞ \Gamma_{\infty} Γ ∞ and the second moment of the normalised spectral measure Γ ~ \widetilde{\Gamma} Γ . □
Corollary 3 (von Mises–Fisher spectral measure). Let Γ \Gamma Γ have density Γ ∞ f v M F ( ⋅ ; μ , κ ) \Gamma_{\infty}f_{\mathrm{vMF}}(\cdot\,;\mu,\kappa) Γ ∞ f vMF ( ⋅ ; μ , κ ) with respect to the uniform surface measure on S d − 1 S^{d-1} S d − 1 , μ ∈ S d − 1 \mu\in S^{d-1} μ ∈ S d − 1 is the mean direction, and κ ≥ 0 \kappa\ge0 κ ≥ 0 is the concentration parameter. Then
Σ Γ = 2 Γ ∞ [ b ( κ , d ) I d + c ( κ , d ) μ μ T ] , (36) \Sigma_{\Gamma}=2\Gamma_{\infty}\left[b(\kappa,d)I_d+c(\kappa,d)\mu\mu^{\mathsf{T}}\right], \tag*{(36)} Σ Γ = 2 Γ ∞ [ b ( κ , d ) I d + c ( κ , d ) μ μ T ] , (36) where
a ( κ , d ) = 1 d + d − 1 d I d / 2 + 1 ( κ ) I d / 2 − 1 ( κ ) , (37) a(\kappa,d)=\frac{1}{d}+\frac{d-1}{d}\frac{I_{d/2+1}(\kappa)}{I_{d/2-1}(\kappa)}, \tag*{(37)} a ( κ , d ) = d 1 + d d − 1 I d /2 − 1 ( κ ) I d /2 + 1 ( κ ) , (37) b ( κ , d ) = 1 − a ( κ , d ) d − 1 , (38) b(\kappa,d)=\frac{1-a(\kappa,d)}{d-1}, \tag*{(38)} b ( κ , d ) = d − 1 1 − a ( κ , d ) , (38) c ( κ , d ) = a ( κ , d ) − b ( κ , d ) , (39) c(\kappa,d)=a(\kappa,d)-b(\kappa,d), \tag*{(39)} c ( κ , d ) = a ( κ , d ) − b ( κ , d ) , (39) and I p ( ⋅ ) I_p(\cdot) I p ( ⋅ ) denotes the modified Bessel function of the first kind of order $p.
Corollary 4 (Bivariate case, d = 2 d = 2 d = 2 ). At d = 2 d = 2 d = 2 , with μ = ( cos η , sin η ) T \mu= (\cos\eta,\sin\eta)^{\mathsf{T}} μ = ( cos η , sin η ) T and R 2 ( κ ) = I 2 ( κ ) / I 0 ( κ ) R_2(\kappa) = I_2(\kappa)/I_0(\kappa) R 2 ( κ ) = I 2 ( κ ) / I 0 ( κ ) , Corollary 3 reduces to
Σ Γ = Γ ∞ ( I 2 + R 2 ( κ ) A ( η ) ) , A ( η ) = ( cos 2 η sin 2 η sin 2 η − cos 2 η ) . (40) \Sigma_{\Gamma} = \Gamma_{\infty}(I_2 + R_2(\kappa) A(\eta)), \qquad A(\eta) = \begin{pmatrix} \cos2\eta& \sin2\eta\\ \sin2\eta& -\cos2\eta\end{pmatrix}. \tag*{(40)} Σ Γ = Γ ∞ ( I 2 + R 2 ( κ ) A ( η )) , A ( η ) = ( cos 2 η sin 2 η sin 2 η − cos 2 η ) . (40) Proof. At d = 2 d = 2 d = 2 , I d / 2 + 1 ( κ ) / I d / 2 − 1 ( κ ) = I 2 ( κ ) / I 0 ( κ ) = R 2 ( κ ) I_{d/2+1}(\kappa)/I_{d/2-1}(\kappa) = I_2(\kappa)/I_0(\kappa) = R_2(\kappa) I d /2 + 1 ( κ ) / I d /2 − 1 ( κ ) = I 2 ( κ ) / I 0 ( κ ) = R 2 ( κ ) , so a ( κ , 2 ) = 1 2 ( 1 + R 2 ( κ ) ) a(\kappa,2) = \frac{1}{2}(1 + R_2(\kappa)) a ( κ , 2 ) = 2 1 ( 1 + R 2 ( κ )) , b ( κ , 2 ) = 1 2 ( 1 − R 2 ( κ ) ) b(\kappa,2) = \frac{1}{2}(1 - R_2(\kappa)) b ( κ , 2 ) = 2 1 ( 1 − R 2 ( κ )) , c ( κ , 2 ) = R 2 ( κ ) c(\kappa,2) = R_2(\kappa) c ( κ , 2 ) = R 2 ( κ ) . Substituting into Corollary 3 and using 2 μ μ T = I 2 + A ( η ) 2\mu\mu^{\mathsf{T}} = I_2 + A(\eta) 2 μ μ T = I 2 + A ( η ) (since 2 cos 2 η − 1 = cos 2 η 2\cos^2\eta- 1 = \cos2\eta 2 cos 2 η − 1 = cos 2 η , 2 sin 2 η − 1 = − cos 2 η 2\sin^2\eta- 1 = -\cos2\eta 2 sin 2 η − 1 = − cos 2 η , 2 cos η sin η = sin 2 η 2\cos\eta\sin\eta= \sin2\eta 2 cos η sin η = sin 2 η ) gives the result.
Monotonicity of the anisotropy ratio
The matrix Σ Γ \Sigma_{\Gamma} Σ Γ has eigenvector μ \mu μ with eigenvalue 2 Γ ∞ a ( κ , d ) 2\Gamma_{\infty}a(\kappa,d) 2 Γ ∞ a ( κ , d ) and all directions orthogonal to μ \mu μ with eigenvalue 2 Γ ∞ b ( κ , d ) 2\Gamma_{\infty}b(\kappa,d) 2 Γ ∞ b ( κ , d ) ; the anisotropy ratio λ max / λ min = a ( κ , d ) / b ( κ , d ) \lambda_{\max}/\lambda_{\min} = a(\kappa,d)/b(\kappa,d) λ m a x / λ m i n = a ( κ , d ) / b ( κ , d ) increases monotonically from 1 1 1 (isotropic, κ = 0 \kappa= 0 κ = 0 ) to ∞ \infty ∞ (rank-1 degenerate, κ → ∞ \kappa\to\infty κ → ∞ ).
Proposition 3. Let
a ( κ , d ) = 1 d + d − 1 d I d / 2 + 1 ( κ ) I d / 2 − 1 ( κ ) a(\kappa,d) = \frac{1}{d} + \frac{d-1}{d}\frac{I_{d/2+1}(\kappa)}{I_{d/2-1}(\kappa)} a ( κ , d ) = d 1 + d d − 1 I d /2 − 1 ( κ ) I d /2 + 1 ( κ ) and
b ( κ , d ) = 1 − a ( κ , d ) d − 1 , b(\kappa,d) = \frac{1-a(\kappa,d)}{d-1}, b ( κ , d ) = d − 1 1 − a ( κ , d ) , where d ≥ 2 d \ge2 d ≥ 2 and I ν I_{\nu} I ν denotes the modified Bessel function of the first kind of order ν \nu ν . Define the anisotropy ratio by
ρ ( κ , d ) = a ( κ , d ) b ( κ , d ) = ( d − 1 ) a ( κ , d ) 1 − a ( κ , d ) . \rho(\kappa,d) = \frac{a(\kappa,d)}{b(\kappa,d)} = \frac{(d-1)a(\kappa,d)}{1-a(\kappa,d)}. ρ ( κ , d ) = b ( κ , d ) a ( κ , d ) = 1 − a ( κ , d ) ( d − 1 ) a ( κ , d ) . Then ρ ( κ , d ) \rho(\kappa,d) ρ ( κ , d ) is strictly increasing in κ \kappa κ on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) . Moreover,
ρ ( 0 , d ) = 1 \rho(0,d) = 1 ρ ( 0 , d ) = 1 and
lim κ → ∞ ρ ( κ , d ) = ∞ . \lim_{\kappa\to\infty}\rho(\kappa,d) = \infty. κ → ∞ lim ρ ( κ , d ) = ∞. Proof. Define
R ( κ , d ) = I d / 2 + 1 ( κ ) I d / 2 − 1 ( κ ) . R(\kappa,d) = \frac{I_{d/2+1}(\kappa)}{I_{d/2-1}(\kappa)}. R ( κ , d ) = I d /2 − 1 ( κ ) I d /2 + 1 ( κ ) . *Step 1: R ( κ , d ) R(\kappa,d) R ( κ , d ) is strictly increasing in κ . ∗ \kappa.* κ . ∗ Amos [2 ] shows that κ ↦ I ν + 2 ( κ ) / I ν ( κ ) \kappa\mapsto I_{\nu+2}(\kappa)/I_{\nu}(\kappa) κ ↦ I ν + 2 ( κ ) / I ν ( κ ) is strictly increasing on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) for all ν ≥ 0 \nu\ge0 ν ≥ 0 . Setting ν = d / 2 − 1 \nu= d/2 - 1 ν = d /2 − 1 , which satisfies ν ≥ 0 \nu\ge0 ν ≥ 0 since d ≥ 2 d \ge2 d ≥ 2 , gives R ( κ , d ) = I ν + 2 ( κ ) / I ν ( κ ) R(\kappa,d) = I_{\nu+2}(\kappa)/I_{\nu}(\kappa) R ( κ , d ) = I ν + 2 ( κ ) / I ν ( κ ) , so R ( κ , d ) R(\kappa,d) R ( κ , d ) is strictly increasing on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) .
*Step 2: a ( κ , d ) a(\kappa,d) a ( κ , d ) is strictly increasing and takes values in ( 0 , 1 ) (0,1) ( 0 , 1 ) for κ ∈ ( 0 , ∞ ) . ∗ \kappa\in(0,\infty).* κ ∈ ( 0 , ∞ ) . ∗ Since
a ( κ , d ) = 1 d + d − 1 d R ( κ , d ) a(\kappa,d) = \frac{1}{d} + \frac{d-1}{d}R(\kappa,d) a ( κ , d ) = d 1 + d d − 1 R ( κ , d )
and ( d − 1 ) / d > 0 (d-1)/d > 0 ( d − 1 ) / d > 0 , the function a ( κ , d ) a(\kappa,d) a ( κ , d ) is strictly increasing in κ \kappa κ . For the bounds, note that I ν ( κ ) > 0 I_\nu(\kappa) > 0 I ν ( κ ) > 0 for all κ > 0 \kappa> 0 κ > 0 and ν ≥ 0 \nu\ge0 ν ≥ 0 , so R ( κ , d ) > 0 R(\kappa,d) > 0 R ( κ , d ) > 0 and hence a ( κ , d ) > 1 / d > 0 a(\kappa,d) > 1/d > 0 a ( κ , d ) > 1/ d > 0 . For the upper bound, the modified Bessel function I ν ( κ ) I_\nu(\kappa) I ν ( κ ) is strictly decreasing in its order ν \nu ν for fixed κ > 0 \kappa> 0 κ > 0 and ν ≥ 0 \nu\ge0 ν ≥ 0 [25 §10.37] , so I d / 2 + 1 ( κ ) < I d / 2 − 1 ( κ ) I_{d/2+1}(\kappa) < I_{d/2-1}(\kappa) I d /2 + 1 ( κ ) < I d /2 − 1 ( κ ) and therefore R ( κ , d ) < 1 R(\kappa,d) < 1 R ( κ , d ) < 1 . Consequently,
a ( κ , d ) < 1 d + d − 1 d = 1. a(\kappa,d) < \frac{1}{d}+\frac{d-1}{d}=1. a ( κ , d ) < d 1 + d d − 1 = 1.
Thus a ( κ , d ) ∈ ( 0 , 1 ) a(\kappa,d) \in(0,1) a ( κ , d ) ∈ ( 0 , 1 ) for all finite κ > 0 \kappa> 0 κ > 0 , and ρ ( κ , d ) \rho(\kappa,d) ρ ( κ , d ) is well-defined.
Step 3: ρ ( κ , d ) \rho(\kappa,d) ρ ( κ , d ) is strictly increasing. Consider
g ( a ) = ( d − 1 ) a 1 − a , 0 < a < 1. g(a)=\frac{(d-1)a}{1-a}, \qquad0<a<1. g ( a ) = 1 − a ( d − 1 ) a , 0 < a < 1.
Its derivative
g ′ ( a ) = d − 1 ( 1 − a ) 2 > 0 g'(a)=\frac{d-1}{(1-a)^2}>0 g ′ ( a ) = ( 1 − a ) 2 d − 1 > 0
shows that g g g is strictly increasing on ( 0 , 1 ) (0,1) ( 0 , 1 ) . Since ρ ( κ , d ) = g ( a ( κ , d ) ) \rho(\kappa,d)=g(a(\kappa,d)) ρ ( κ , d ) = g ( a ( κ , d )) and a ( κ , d ) a(\kappa,d) a ( κ , d ) is strictly increasing in κ \kappa κ , the composition ρ ( κ , d ) \rho(\kappa,d) ρ ( κ , d ) is strictly increasing on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) .
Step 4: Boundary values. For the lower boundary, the small-argument expansion I r ( κ ) ∼ ( κ / 2 ) r / Γ ( r + 1 ) I_r(\kappa)\sim(\kappa/2)^r/\Gamma(r+1) I r ( κ ) ∼ ( κ /2 ) r /Γ ( r + 1 ) as κ → 0 + \kappa\to0^+ κ → 0 + gives
lim κ → 0 + R ( κ , d ) = lim κ → 0 + ( κ / 2 ) d / 2 + 1 / Γ ( d / 2 + 2 ) ( κ / 2 ) d / 2 − 1 / Γ ( d / 2 ) = lim κ → 0 + ( κ / 2 ) 2 Γ ( d / 2 ) Γ ( d / 2 + 2 ) = 0. \lim_{\kappa\to0^+}R(\kappa,d)
=\lim_{\kappa\to0^+}\frac{(\kappa/2)^{d/2+1}/\Gamma(d/2+2)}{(\kappa/2)^{d/2-1}/\Gamma(d/2)}
=\lim_{\kappa\to0^+}\frac{(\kappa/2)^2\Gamma(d/2)}{\Gamma(d/2+2)}
=0. κ → 0 + lim R ( κ , d ) = κ → 0 + lim ( κ /2 ) d /2 − 1 /Γ ( d /2 ) ( κ /2 ) d /2 + 1 /Γ ( d /2 + 2 ) = κ → 0 + lim Γ ( d /2 + 2 ) ( κ /2 ) 2 Γ ( d /2 ) = 0.
Hence a ( 0 , d ) = 1 / d a(0,d)=1/d a ( 0 , d ) = 1/ d and b ( 0 , d ) = ( 1 − 1 / d ) / ( d − 1 ) = 1 / d b(0,d)=(1-1/d)/(d-1)=1/d b ( 0 , d ) = ( 1 − 1/ d ) / ( d − 1 ) = 1/ d , so ρ ( 0 , d ) = 1 \rho(0,d)=1 ρ ( 0 , d ) = 1 .
For the upper boundary, the large-argument expansion I r ( κ ) ∼ e κ / 2 π κ I_r(\kappa)\sim e^\kappa/\sqrt{2\pi\kappa} I r ( κ ) ∼ e κ / 2 π κ as κ → ∞ \kappa\to\infty κ → ∞ gives lim κ → ∞ R ( κ , d ) = 1 \lim_{\kappa\to\infty}R(\kappa,d)=1 lim κ → ∞ R ( κ , d ) = 1 , so
lim κ → ∞ a ( κ , d ) = 1 d + d − 1 d = 1 , \lim_{\kappa\to\infty}a(\kappa,d)=\frac{1}{d}+\frac{d-1}{d}=1, κ → ∞ lim a ( κ , d ) = d 1 + d d − 1 = 1 ,
and therefore lim κ → ∞ b ( κ , d ) = 0 \lim_{\kappa\to\infty}b(\kappa,d)=0 lim κ → ∞ b ( κ , d ) = 0 . Consequently,
lim κ → ∞ ρ ( κ , d ) = lim κ → ∞ ( d − 1 ) a ( κ , d ) 1 − a ( κ , d ) = ∞ . \lim_{\kappa\to\infty}\rho(\kappa,d)
=\lim_{\kappa\to\infty}\frac{(d-1)a(\kappa,d)}{1-a(\kappa,d)}
=\infty. κ → ∞ lim ρ ( κ , d ) = κ → ∞ lim 1 − a ( κ , d ) ( d − 1 ) a ( κ , d ) = ∞.
This completes the proof.
Derivation of the likelihood function in Section 3.2
L ( α , η , κ ) ∝ f ( { X i } , { r i j } , { y i j } , { s i j } ∣ α , η , κ ) ∝ f ( { r i j } ∣ α , η , κ , { X i } ) × f ( { y i j } ∣ α , η , κ , { X i } , { r i j } ) × f ( { s i j } ∣ α , η , κ , { X i } , { r i j } , { y i j } ) × f ( { X i } ∣ α , η , κ , { r i j } , { y i j } , { s i j } ) = f ( { r i j } ∣ α ) × f ( { y i j } ∣ α , { r i j } ) × f ( { s i j } ∣ η , κ ) × f ( { X i } ∣ α , η , κ , { r i j } , { y i j } , { s i j } ) = f ( { r i j } ∣ α ) × f ( { y i j } ∣ α , { r i j } ) × f ( { s i j } ∣ η , κ ) × f ( { ϵ i } ∣ α , η , κ , { r i j } , { y i j } , { s i j } ) = { ∏ i = 1 n ∏ j = 1 m f ( r i j ∣ α ) } × { ∏ i = 1 n ∏ j = 1 m f ( y i j ∣ α , r i j ) } × { ∏ i = 1 n ∏ j = 1 m f ( s i j ∣ η , κ ) } × { ∏ i = 1 n f ( X i ∣ α , η , κ , { r i j } , { y i j } , { s i j } ) } = { ∏ i = 1 n ∏ j = 1 m 1 Γ ( α ) e − 1 / r i j r i j 1 + α } × { ∏ i = 1 n ∏ j = 1 m e − λ i j λ i j y i j y i j ! } × { ∏ i = 1 n ∏ j = 1 m f Z ( s i j ∣ η , κ ) } × { ∏ i = 1 n f ( ϵ i ∣ α , η , κ , { r i j } , { y i j } , { s i j } ) } . (41) \begin{aligned}
L(\alpha,\eta,\kappa) &\propto f(\{X_i\},\{r_{ij}\},\{y_{ij}\},\{s_{ij}\}\mid\alpha,\eta,\kappa)\\
&\propto f(\{r_{ij}\}\mid\alpha,\eta,\kappa,\{X_i\})\times f(\{y_{ij}\}\mid\alpha,\eta,\kappa,\{X_i\},\{r_{ij}\})\\
&\quad\times f(\{s_{ij}\}\mid\alpha,\eta,\kappa,\{X_i\},\{r_{ij}\},\{y_{ij}\})\times f(\{X_i\}\mid\alpha,\eta,\kappa,\{r_{ij}\},\{y_{ij}\},\{s_{ij}\})\\
&=f(\{r_{ij}\}\mid\alpha)\times f(\{y_{ij}\}\mid\alpha,\{r_{ij}\})\times f(\{s_{ij}\}\mid\eta,\kappa)\\
&\quad\times f(\{X_i\}\mid\alpha,\eta,\kappa,\{r_{ij}\},\{y_{ij}\},\{s_{ij}\})\\
&=f(\{r_{ij}\}\mid\alpha)\times f(\{y_{ij}\}\mid\alpha,\{r_{ij}\})\times f(\{s_{ij}\}\mid\eta,\kappa)\\
&\quad\times f(\{\epsilon_i\}\mid\alpha,\eta,\kappa,\{r_{ij}\},\{y_{ij}\},\{s_{ij}\})\\
&=\left\{\prod_{i=1}^{n}\prod_{j=1}^{m}f(r_{ij}\mid\alpha)\right\}\times\left\{\prod_{i=1}^{n}\prod_{j=1}^{m}f(y_{ij}\mid\alpha,r_{ij})\right\}\\
&\quad\times\left\{\prod_{i=1}^{n}\prod_{j=1}^{m}f(s_{ij}\mid\eta,\kappa)\right\}\times\left\{\prod_{i=1}^{n}f(X_i\mid\alpha,\eta,\kappa,\{r_{ij}\},\{y_{ij}\},\{s_{ij}\})\right\}\\
&=\left\{\prod_{i=1}^{n}\prod_{j=1}^{m}\frac{1}{\Gamma(\alpha)}\frac{e^{-1/r_{ij}}}{r_{ij}^{1+\alpha}}\right\}\times\left\{\prod_{i=1}^{n}\prod_{j=1}^{m}\frac{e^{-\lambda_{ij}} \lambda_{ij}^{y_{ij}}}{y_{ij}!}\right\}\\
&\quad\times\left\{\prod_{i=1}^{n}\prod_{j=1}^{m}f_Z(s_{ij}\mid\eta,\kappa)\right\}\times\left\{\prod_{i=1}^{n}f(\epsilon_i\mid\alpha,\eta,\kappa,\{r_{ij}\},\{y_{ij}\},\{s_{ij}\})\right\}. \tag*{(41)}
\end{aligned} L ( α , η , κ ) ∝ f ({ X i } , { r ij } , { y ij } , { s ij } ∣ α , η , κ ) ∝ f ({ r ij } ∣ α , η , κ , { X i }) × f ({ y ij } ∣ α , η , κ , { X i } , { r ij }) × f ({ s ij } ∣ α , η , κ , { X i } , { r ij } , { y ij }) × f ({ X i } ∣ α , η , κ , { r ij } , { y ij } , { s ij }) = f ({ r ij } ∣ α ) × f ({ y ij } ∣ α , { r ij }) × f ({ s ij } ∣ η , κ ) × f ({ X i } ∣ α , η , κ , { r ij } , { y ij } , { s ij }) = f ({ r ij } ∣ α ) × f ({ y ij } ∣ α , { r ij }) × f ({ s ij } ∣ η , κ ) × f ({ ϵ i } ∣ α , η , κ , { r ij } , { y ij } , { s ij }) = { i = 1 ∏ n j = 1 ∏ m f ( r ij ∣ α ) } × { i = 1 ∏ n j = 1 ∏ m f ( y ij ∣ α , r ij ) } × { i = 1 ∏ n j = 1 ∏ m f ( s ij ∣ η , κ ) } × { i = 1 ∏ n f ( X i ∣ α , η , κ , { r ij } , { y ij } , { s ij }) } = { i = 1 ∏ n j = 1 ∏ m Γ ( α ) 1 r ij 1 + α e − 1/ r ij } × { i = 1 ∏ n j = 1 ∏ m y ij ! e − λ ij λ ij y ij } × { i = 1 ∏ n j = 1 ∏ m f Z ( s ij ∣ η , κ ) } × { i = 1 ∏ n f ( ϵ i ∣ α , η , κ , { r ij } , { y ij } , { s ij }) } . (41)
Smoothness of the Characteristic Function
Continuity of the Lévy-Khintchine integral
Following Kuelbs, we express the characteristic function using the Lévy–Khintchine integral. We show that this integral is smooth in the stability index α ∈ ( 0 , 2 ) \alpha\in(0,2) α ∈ ( 0 , 2 ) , including at the Cauchy boundary α = 1 \alpha=1 α = 1 . Define
K ( v , α ) = ∫ 0 ∞ [ e i r v − 1 − i r v 1 + r 2 ] d r r 1 + α , v ∈ R , α ∈ ( 0 , 2 ) . (42) \mathcal{K}(v,\alpha)=\int_{0}^{\infty}\left[e^{irv}-1-\frac{irv}{1+r^2}\right]\frac{dr}{r^{1+\alpha}},\qquad v\in\mathbb{R},\quad\alpha\in(0,2). \tag*{(42)} K ( v , α ) = ∫ 0 ∞ [ e i r v − 1 − 1 + r 2 i r v ] r 1 + α d r , v ∈ R , α ∈ ( 0 , 2 ) . (42)
Lemma 1 (Smoothness of the Lévy–Khintchine integral in α \alpha α ). For every v ∈ R v\in\mathbb{R} v ∈ R , the integral K ( v , α ) \mathcal{K}(v,\alpha) K ( v , α ) in (42) is finite for every α ∈ ( 0 , 2 ) \alpha\in(0,2) α ∈ ( 0 , 2 ) . Moreover, for each fixed v ∈ R v\in\mathbb{R} v ∈ R , the map α ⟼ K ( v , α ) \alpha\longmapsto\mathcal{K}(v,\alpha) α ⟼ K ( v , α ) is C ∞ C^\infty C ∞ on ( 0 , 2 ) (0,2) ( 0 , 2 ) . In particular, it is smooth at α = 1 \alpha=1 α = 1 , and no separate treatment of the Cauchy case is required.
Proof. Write
K ( v , α ) = A ( v , α ) + i B ( v , α ) , \mathcal{K}(v,\alpha)=A(v,\alpha)+iB(v,\alpha), K ( v , α ) = A ( v , α ) + i B ( v , α ) ,
where
A ( v , α ) = ∫ 0 ∞ [ cos ( r v ) − 1 ] r − 1 − α d r (43) A(v,\alpha)=\int_{0}^{\infty}\left[\cos(rv)-1\right]r^{-1-\alpha}\,dr \tag*{(43)} A ( v , α ) = ∫ 0 ∞ [ cos ( r v ) − 1 ] r − 1 − α d r (43)
B ( v , α ) = ∫ 0 ∞ [ sin ( r v ) − r v 1 + r 2 ] r − 1 − α d r . (44) B(v,\alpha)=\int_{0}^{\infty}\left[\sin(rv)-\frac{rv}{1+r^2}\right]r^{-1-\alpha}\,dr. \tag*{(44)} B ( v , α ) = ∫ 0 ∞ [ sin ( r v ) − 1 + r 2 r v ] r − 1 − α d r . (44)
If v = 0 v=0 v = 0 , then the integrand in (42) vanishes identically, so that
K ( 0 , α ) = 0. \mathcal{K}(0,\alpha)=0. K ( 0 , α ) = 0.
Thus, assume for the moment that v ≠ 0 v\ne0 v = 0 .
As r → 0 + r\to0^{+} r → 0 + , Taylor expansion gives
cos ( r v ) − 1 = − r 2 v 2 2 + O ( r 4 ) \cos(rv)-1=-\frac{r^2v^2}{2}+O(r^4) cos ( r v ) − 1 = − 2 r 2 v 2 + O ( r 4 )
and
sin ( r v ) − r v 1 + r 2 = O ( r 3 ) . \sin(rv)-\frac{rv}{1+r^2}=O(r^3). sin ( r v ) − 1 + r 2 r v = O ( r 3 ) .
Consequently, near zero, the integrands in (43) and (44) behave respectively as
O ( r 1 − α ) and O ( r 2 − α ) . O(r^{1-\alpha}) \qquad\text{and}\qquad O(r^{2-\alpha}). O ( r 1 − α ) and O ( r 2 − α ) .
Both are integrable near zero because α < 2 \alpha<2 α < 2 .
As r → ∞ r\to\infty r → ∞ ,
∣ cos ( r v ) − 1 ∣ ≤ 2 , \lvert\cos(rv)-1\rvert\le2, ∣ cos ( r v ) − 1 ∣ ≤ 2 ,
and
∣ sin ( r v ) − r v 1 + r 2 ∣ ≤ 1 + ∣ v ∣ r 1 + r 2 ≤ 1 + ∣ v ∣ r . \left\lvert\sin(rv)-\frac{rv}{1+r^2}\right\rvert\le1+\frac{\lvert v\rvert r}{1+r^2}\le1+\frac{\lvert v\rvert}{r}. sin ( r v ) − 1 + r 2 r v ≤ 1 + 1 + r 2 ∣ v ∣ r ≤ 1 + r ∣ v ∣ .
Therefore, both integrands are bounded at infinity by a constant multiple of
r − 1 − α , r^{-1-\alpha}, r − 1 − α ,
which is integrable because α > 0 \alpha>0 α > 0 . This proves that K ( v , α ) \mathcal{K}(v,\alpha) K ( v , α ) is finite for every
( v , α ) ∈ R × ( 0 , 2 ) . (v,\alpha)\in\mathbb{R}\times(0,2). ( v , α ) ∈ R × ( 0 , 2 ) .
It remains to prove smoothness in α \alpha α . Fix v ∈ R v\in\mathbb{R} v ∈ R and let
[ α 0 , α 1 ] ⊂ ( 0 , 2 ) [\alpha_0,\alpha_1]\subset(0,2) [ α 0 , α 1 ] ⊂ ( 0 , 2 )
be compact. For every integer k ≥ 0 k\ge0 k ≥ 0 ,
∂ k ∂ α k [ ( cos ( r v ) − 1 ) r − 1 − α ] = ( − log r ) k ( cos ( r v ) − 1 ) r − 1 − α . \frac{\partial^k}{\partial\alpha^k}\left[(\cos(rv)-1)r^{-1-\alpha}\right]=(-\log r)^k(\cos(rv)-1)r^{-1-\alpha}. ∂ α k ∂ k [ ( cos ( r v ) − 1 ) r − 1 − α ] = ( − log r ) k ( cos ( r v ) − 1 ) r − 1 − α .
For 0 < r < 1 0<r<1 0 < r < 1 , the inequality
∣ cos ( r v ) − 1 ∣ ≤ v 2 r 2 2 \lvert\cos(rv)-1\rvert\le\frac{v^2r^2}{2} ∣ cos ( r v ) − 1 ∣ ≤ 2 v 2 r 2
gives
∣ ( − log r ) k ( cos ( r v ) − 1 ) r − 1 − α ∣ ≤ v 2 2 ∣ log r ∣ k r 1 − α 1 . \left|(-\log r)^k(\cos(rv)-1)r^{-1-\alpha}\right| \le\frac{v^2}{2}|\log r|^k r^{1-\alpha_1}. ( − log r ) k ( cos ( r v ) − 1 ) r − 1 − α ≤ 2 v 2 ∣ log r ∣ k r 1 − α 1 .
Since α 1 < 2 \alpha_1<2 α 1 < 2 , the function
∣ log r ∣ k r 1 − α 1 |\log r|^k r^{1-\alpha_1} ∣ log r ∣ k r 1 − α 1
is integrable on ( 0 , 1 ) (0,1) ( 0 , 1 ) .
For r ≥ 1 r\ge1 r ≥ 1 , using ∣ cos ( r v ) − 1 ∣ ≤ 2 |\cos(rv)-1|\le2 ∣ cos ( r v ) − 1∣ ≤ 2 , we obtain
∣ ( − log r ) k ( cos ( r v ) − 1 ) r − 1 − α ∣ ≤ 2 ( log r ) k r − 1 − α 0 . \left|(-\log r)^k(\cos(rv)-1)r^{-1-\alpha}\right| \le2(\log r)^k r^{-1-\alpha_0}. ( − log r ) k ( cos ( r v ) − 1 ) r − 1 − α ≤ 2 ( log r ) k r − 1 − α 0 .
This is integrable on ( 1 , ∞ ) (1,\infty) ( 1 , ∞ ) because α 0 > 0 \alpha_0>0 α 0 > 0 .
Thus, differentiation under the integral sign is justified, and
∂ k A ∂ α k ( v , α ) = ∫ 0 ∞ ( − log r ) k ( cos ( r v ) − 1 ) r − 1 − α d r . \frac{\partial^k A}{\partial\alpha^k}(v,\alpha)=\int_0^\infty(-\log r)^k(\cos(rv)-1)r^{-1-\alpha}\,dr. ∂ α k ∂ k A ( v , α ) = ∫ 0 ∞ ( − log r ) k ( cos ( r v ) − 1 ) r − 1 − α d r .
The same argument applies to the imaginary part. Indeed,
∂ k B ∂ α k ( v , α ) = ∫ 0 ∞ ( − log r ) k [ sin ( r v ) − r v 1 + r 2 ] r − 1 − α d r . \frac{\partial^k B}{\partial\alpha^k}(v,\alpha)=\int_0^\infty(-\log r)^k\left[\sin(rv)-\frac{rv}{1+r^2}\right]r^{-1-\alpha}\,dr. ∂ α k ∂ k B ( v , α ) = ∫ 0 ∞ ( − log r ) k [ sin ( r v ) − 1 + r 2 r v ] r − 1 − α d r .
Near zero, the integrand is dominated by a constant multiple of
∣ log r ∣ k r 2 − α 1 , |\log r|^k r^{2-\alpha_1}, ∣ log r ∣ k r 2 − α 1 ,
and near infinity by a constant multiple of
( log r ) k r − 1 − α 0 . (\log r)^k r^{-1-\alpha_0}. ( log r ) k r − 1 − α 0 .
Both functions are integrable on their respective domains.
Therefore, K ( v , α ) \mathcal{K}(v,\alpha) K ( v , α ) is C ∞ C^\infty C ∞ as a function of α \alpha α on ( 0 , 2 ) (0,2) ( 0 , 2 ) . Since 1 ∈ ( 0 , 2 ) 1\in(0,2) 1 ∈ ( 0 , 2 ) , this smoothness includes α = 1 \alpha=1 α = 1 .
The Double Integral
For t ∈ R d t\in\mathbb{R}^d t ∈ R d , write t = τ u t=\tau u t = τ u where τ = ∥ t ∥ ≥ 0 \tau=\|t\|\ge0 τ = ∥ t ∥ ≥ 0 and u = t / ∥ t ∥ ∈ S d − 1 u=t/\|t\|\in S^{d-1} u = t /∥ t ∥ ∈ S d − 1 (for t = 0 t=0 t = 0 the integral vanishes trivially). For a d d d -dimensional stable model with vMF spectral density f v M F ( s ∣ μ , κ ) f_{\mathrm{vMF}}(s\mid\mu,\kappa) f vMF ( s ∣ μ , κ ) on S d − 1 S^{d-1} S d − 1 , the relevant double integral is
∫ S d − 1 ∫ 0 ∞ [ e i r ⟨ t , s ⟩ − 1 − i r ⟨ t , s ⟩ 1 + r 2 ] d r r 1 + α f v M F ( s ∣ μ , κ ) d s . (45) \int_{S^{d-1}}\int_0^\infty\left[e^{ir\langle t,s\rangle}-1-\frac{ir\langle t,s\rangle}{1+r^2}\right]\frac{dr}{r^{1+\alpha}}f_{\mathrm{vMF}}(s\mid\mu,\kappa)\,ds. \tag*{(45)} ∫ S d − 1 ∫ 0 ∞ [ e i r ⟨ t , s ⟩ − 1 − 1 + r 2 i r ⟨ t , s ⟩ ] r 1 + α d r f vMF ( s ∣ μ , κ ) d s . (45)
Theorem 5 (Smoothness in the stability index, general d d d ). Fix t ∈ R d t\in\mathbb{R}^d t ∈ R d , μ ∈ S d − 1 \mu\in S^{d-1} μ ∈ S d − 1 , and κ ∈ [ 0 , ∞ ) \kappa\in[0,\infty) κ ∈ [ 0 , ∞ ) . Then the double integral (45) is C ∞ C^\infty C ∞ as a function of α ∈ ( 0 , 2 ) \alpha\in(0,2) α ∈ ( 0 , 2 ) , including at α = 1 \alpha=1 α = 1 .
Proof. For each s ∈ S d − 1 s \in S^{d-1} s ∈ S d − 1 , set v ( s ) = ⟨ t , s ⟩ v(s) = \langle t, s\rangle v ( s ) = ⟨ t , s ⟩ . By the Cauchy–Schwarz inequality,
∣ v ( s ) ∣ = ∣ ⟨ t , s ⟩ ∣ ≤ ∥ t ∥ ∥ s ∥ = ∥ t ∥ , |v(s)| = |\langle t, s\rangle| \le\|t\| \|s\| = \|t\|, ∣ v ( s ) ∣ = ∣ ⟨ t , s ⟩ ∣ ≤ ∥ t ∥∥ s ∥ = ∥ t ∥ ,
so v ( s ) v(s) v ( s ) is bounded uniformly in s ∈ S d − 1 s \in S^{d-1} s ∈ S d − 1 . By Lemma 1, the inner integral equals K ( v ( s ) , α ) \mathcal{K}(v(s), \alpha) K ( v ( s ) , α ) for each s s s , so (45) equals
∫ S d − 1 K ( v ( s ) , α ) f v M F ( s ∣ μ , κ ) d s . \int_{S^{d-1}} \mathcal{K}(v(s), \alpha) f_{\mathrm{vMF}}(s \mid\mu, \kappa)\,ds. ∫ S d − 1 K ( v ( s ) , α ) f vMF ( s ∣ μ , κ ) d s .
If t = 0 t = 0 t = 0 , then v ( s ) = 0 v(s) = 0 v ( s ) = 0 for every s s s and K ( 0 , α ) = 0 \mathcal{K}(0,\alpha) = 0 K ( 0 , α ) = 0 , so the integral vanishes identically and is trivially C ∞ C^\infty C ∞ .
Now suppose t ≠ 0 t \ne0 t = 0 . Since ∣ v ( s ) ∣ ≤ ∥ t ∥ |v(s)| \le\|t\| ∣ v ( s ) ∣ ≤ ∥ t ∥ uniformly in s s s , the dominating functions in the proof of Lemma 1 may be taken with ∣ v ∣ |v| ∣ v ∣ replaced by ∥ t ∥ \|t\| ∥ t ∥ , giving bounds that are uniform in s s s . Specifically, for α ∈ [ α 0 , α 1 ] ⊂ ( 0 , 2 ) \alpha\in[\alpha_0,\alpha_1] \subset(0,2) α ∈ [ α 0 , α 1 ] ⊂ ( 0 , 2 ) , the k k k th α \alpha α -derivative of the inner integrand in (45) is dominated, near r = 0 r = 0 r = 0 , by a constant multiple of
∣ log r ∣ k r 1 − α 1 + ∣ log r ∣ k r 2 − α 1 , |\log r|^k r^{1-\alpha_1} + |\log r|^k r^{2-\alpha_1}, ∣ log r ∣ k r 1 − α 1 + ∣ log r ∣ k r 2 − α 1 ,
and, for r ≥ 1 r \ge1 r ≥ 1 , by a constant multiple of
( log r ) k r − 1 − α 0 , (\log r)^k r^{-1-\alpha_0}, ( log r ) k r − 1 − α 0 ,
both integrable over ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) and independent of s s s .
Since f v M F ( ⋅ ∣ μ , κ ) f_{\mathrm{vMF}}(\cdot\mid\mu,\kappa) f vMF ( ⋅ ∣ μ , κ ) is a probability density on S d − 1 S^{d-1} S d − 1 ,
∫ S d − 1 f v M F ( s ∣ μ , κ ) d s = 1 , \int_{S^{d-1}} f_{\mathrm{vMF}}(s \mid\mu,\kappa)\,ds = 1, ∫ S d − 1 f vMF ( s ∣ μ , κ ) d s = 1 ,
the same bounds dominate the s s s -integrated expression. By Fubini’s theorem and differentiation under both integrals, (45) is C ∞ C^\infty C ∞ in α ∈ ( 0 , 2 ) \alpha\in(0,2) α ∈ ( 0 , 2 ) . In particular, it is smooth at α = 1 \alpha= 1 α = 1 .
Approximation with truncation
Proposition 4 (Truncation approximation). Let X ( r min , r max ) ( m ) X_{(r_{\min},r_{\max})}^{(m)} X ( r m i n , r m a x ) ( m ) denote the LePage sum of Theorem 3 with radial components restricted to ( r min , r max ) (r_{\min},r_{\max}) ( r m i n , r m a x ) , 0 < r min < r max < ∞ 0 < r_{\min} < r_{\max} < \infty 0 < r m i n < r m a x < ∞ , and with normalising constant
C r min ( α ) = ∫ r min r max r − α − 1 e − 1 / r d r C_{r_{\min}}(\alpha) = \int_{r_{\min}}^{r_{\max}} r^{-\alpha-1}e^{-1/r}\,dr C r m i n ( α ) = ∫ r m i n r m a x r − α − 1 e − 1/ r d r in place of Γ ( α ) \Gamma(\alpha) Γ ( α ) . Then, for every fixed m m m and every t ∈ R d t \in\mathbb{R}^d t ∈ R d ,
ψ m ( r min , r max ) ( t ) ⟶ ψ m ( t ) as r min → 0 and r max → ∞ , (46) \psi_{m}^{(r_{\min},r_{\max})}(t) \longrightarrow\psi_m(t)
\qquad\text{as } r_{\min} \to0 \text{ and } r_{\max} \to\infty, \tag*{(46)} ψ m ( r m i n , r m a x ) ( t ) ⟶ ψ m ( t ) as r m i n → 0 and r m a x → ∞ , (46) where ψ m ( t ) \psi_m(t) ψ m ( t ) is the characteristic function of X ( m ) X^{(m)} X ( m ) in Theorem 3. Consequently,
ψ m ( r min , r max ) ( t ) ⟶ ψ ( t ) as m → ∞ , δ → 0 , r max → ∞ , (47) \psi_{m}^{(r_{\min},r_{\max})}(t) \longrightarrow\psi(t)
\qquad\text{as } m \to\infty,\ \delta\to0,\ r_{\max} \to\infty, \tag*{(47)} ψ m ( r m i n , r m a x ) ( t ) ⟶ ψ ( t ) as m → ∞ , δ → 0 , r m a x → ∞ , (47) where ψ ( t ) \psi(t) ψ ( t ) is the characteristic function of Stable ( α , Γ ) \operatorname{Stable}(\alpha,\Gamma) Stable ( α , Γ ) .
Proof. Step 1. Convergence of the normalising constant. Since r − α − 1 e − 1 / r ≥ 0 r^{-\alpha-1}e^{-1/r}\ge0 r − α − 1 e − 1/ r ≥ 0 and
∫ 0 ∞ r − α − 1 e − 1 / r d r = Γ ( α ) < ∞ for α ∈ ( 0 , 2 ) , \int_{0}^{\infty} r^{-\alpha-1}e^{-1/r}\,dr=\Gamma(\alpha)<\infty\quad\text{for }\alpha\in(0,2), ∫ 0 ∞ r − α − 1 e − 1/ r d r = Γ ( α ) < ∞ for α ∈ ( 0 , 2 ) ,
the monotone convergence theorem gives
C r min ( α ) = ∫ r min r max r − α − 1 e − 1 / r d r ⟶ Γ ( α ) as r min → 0 , r max → ∞ . (48) C_{r_{\min}}(\alpha)=\int_{r_{\min}}^{r_{\max}}r^{-\alpha-1}e^{-1/r}\,dr \longrightarrow\Gamma(\alpha) \quad\text{as }r_{\min}\to0,\ r_{\max}\to\infty. \tag*{(48)} C r m i n ( α ) = ∫ r m i n r m a x r − α − 1 e − 1/ r d r ⟶ Γ ( α ) as r m i n → 0 , r m a x → ∞. (48)
Step 2. Convergence of the truncated Lévy–Khintchine integral. For v ∈ R v\in\mathbb{R} v ∈ R , define
K δ , r max ( v , α ) = ∫ δ r max [ e i r v − 1 − i r v 1 + r 2 ] d r r 1 + α . (49) \mathcal{K}_{\delta,r_{\max}}(v,\alpha)=\int_{\delta}^{r_{\max}}\left[e^{irv}-1-\frac{irv}{1+r^2}\right]\frac{dr}{r^{1+\alpha}}. \tag*{(49)} K δ , r m a x ( v , α ) = ∫ δ r m a x [ e i r v − 1 − 1 + r 2 i r v ] r 1 + α d r . (49)
The integrand satisfies
∣ e i r v − 1 − i r v 1 + r 2 ∣ r − 1 − α ≤ { v 2 2 r 1 − α r ≤ 1 , C r − 1 − α r > 1 , \left|e^{irv}-1-\frac{irv}{1+r^2}\right|r^{-1-\alpha}\le
\begin{cases}
\frac{v^2}{2}r^{1-\alpha} & r\le1,\\
Cr^{-1-\alpha} & r>1,
\end{cases} e i r v − 1 − 1 + r 2 i r v r − 1 − α ≤ { 2 v 2 r 1 − α C r − 1 − α r ≤ 1 , r > 1 ,
for a constant C C C depending only on v v v and α \alpha α . Both bounds are integrable on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) for α ∈ ( 0 , 2 ) \alpha\in(0,2) α ∈ ( 0 , 2 ) . By the dominated convergence theorem,
K r min , r max ( v , α ) ⟶ K ( v , α ) as r min → 0 , r max → ∞ . (50) \mathcal{K}_{r_{\min},r_{\max}}(v,\alpha)\longrightarrow\mathcal{K}(v,\alpha)\quad\text{as }r_{\min}\to0,\ r_{\max}\to\infty. \tag*{(50)} K r m i n , r m a x ( v , α ) ⟶ K ( v , α ) as r m i n → 0 , r m a x → ∞. (50)
Step 3. Convergence of the characteristic function. The log characteristic function of X r min , r max ( m ) X^{(m)}_{r_{\min},r_{\max}} X r m i n , r m a x ( m ) takes the form
log ψ m ( r min , r max ) ( t ) = Γ ∞ ∫ S d − 1 K r min , r max ( ⟨ t , s ⟩ , α ) Γ ~ ( d s ) . \log\psi_m^{(r_{\min},r_{\max})}(t)=\Gamma_\infty\int_{S^{d-1}}\mathcal{K}_{r_{\min},r_{\max}}(\langle t,s\rangle,\alpha)\,\widetilde{\Gamma}(ds). log ψ m ( r m i n , r m a x ) ( t ) = Γ ∞ ∫ S d − 1 K r m i n , r m a x (⟨ t , s ⟩ , α ) Γ ( d s ) .
By (50) and the fact that ∣ K δ , r max ( ⟨ t , s ⟩ , α ) ∣ \left|\mathcal{K}_{\delta,r_{\max}}(\langle t,s\rangle,\alpha)\right| ∣ K δ , r m a x (⟨ t , s ⟩ , α ) ∣ is bounded uniformly in s ∈ S d − 1 s\in S^{d-1} s ∈ S d − 1 and in ( r min , r max ) (r_{\min},r_{\max}) ( r m i n , r m a x ) by an integrable function (from Step 2 with ∣ v ∣ ≤ ∥ t ∥ |v|\le\lVert t\rVert ∣ v ∣ ≤ ∥ t ∥ ), a further application of the dominated convergence theorem gives
log ψ m ( r min , r max ) ( t ) ⟶ log ψ m ( t ) as r min → 0 , r max → ∞ . \log\psi_m^{(r_{\min},r_{\max})}(t)\longrightarrow\log\psi_m(t)\quad\text{as }r_{\min}\to0,\ r_{\max}\to\infty. log ψ m ( r m i n , r m a x ) ( t ) ⟶ log ψ m ( t ) as r m i n → 0 , r m a x → ∞.
The second statement follows by combining this with Theorem 3.
The logr_gap Diagnostic
This appendix derives the theoretical properties of the logr_gap diagnostic and proves the sign theorem used in § 6 .
The Truncated IG Distribution
The radial draws r i j r_{ij} r ij follow the truncated inverted-gamma distribution IG ( α , r min , r max ) \operatorname{IG}(\alpha,r_{\min},r_{\max}) IG ( α , r m i n , r m a x ) with density:
g ( r ∣ α , r min , r max ) = r − α − 1 e − 1 / r C r min , r max ( α ) , r ∈ ( r min , r max ) , (51) g(r \mid\alpha,r_{\min},r_{\max})=\frac{r^{-\alpha-1}e^{-1/r}}{C_{r_{\min},r_{\max}}(\alpha)},\qquad r\in(r_{\min},r_{\max}), \tag*{(51)} g ( r ∣ α , r m i n , r m a x ) = C r m i n , r m a x ( α ) r − α − 1 e − 1/ r , r ∈ ( r m i n , r m a x ) , (51)
where the normalising constant is:
C r min , r max ( α ) = ∫ r min r max r − α − 1 e − 1 / r d r . (52) C_{r_{\min},r_{\max}}(\alpha)=\int_{r_{\min}}^{r_{\max}}r^{-\alpha-1}e^{-1/r}\,dr. \tag*{(52)} C r m i n , r m a x ( α ) = ∫ r m i n r m a x r − α − 1 e − 1/ r d r . (52)
The reference log-r is:
ref_logr ( α ) = E IG ( α , r min , r max ) [ log r ] = ∫ r min r max log ( r ) ⋅ r − α − 1 e − 1 / r d r C r min , r max ( α ) , (53) \operatorname{ref\_logr}(\alpha)=E_{\operatorname{IG}(\alpha,r_{\min},r_{\max})}[\log r]=\frac{\int_{r_{\min}}^{r_{\max}}\log(r)\cdot r^{-\alpha-1}e^{-1/r}\,dr}{C_{r_{\min},r_{\max}}(\alpha)}, \tag*{(53)} ref_logr ( α ) = E IG ( α , r m i n , r m a x ) [ log r ] = C r m i n , r m a x ( α ) ∫ r m i n r m a x log ( r ) ⋅ r − α − 1 e − 1/ r d r , (53)
and the logr_gap diagnostic is:
logr_gap = log r ‾ c h a i n − ref_logr ( α t r u e ) , (54) \operatorname{logr\_gap}=\overline{\log r}_{\mathrm{chain}}-\operatorname{ref\_logr}(\alpha_{\mathrm{true}}), \tag*{(54)} logr_gap = log r chain − ref_logr ( α true ) , (54)
where log r ‾ c h a i n \overline{\log r}_{\mathrm{chain}} log r chain is the post-burnin mean of log r i j \log r_{ij} log r ij .
Exponential Family Structure
Proposition 5 (IG as exponential family). The density (51) belongs to the one-parameter exponential family with natural parameter η e f = − α \eta_{\mathrm{ef}}=-\alpha η ef = − α , sufficient statistic T ( r ) = log r T(r)=\log r T ( r ) = log r , and log-partition function A ( η e f ) = log C r min , r max ( − η e f ) A(\eta_{\mathrm{ef}})=\log C_{r_{\min},r_{\max}}(-\eta_{\mathrm{ef}}) A ( η ef ) = log C r m i n , r m a x ( − η ef ) .
Corollary 5 (ref_logr as log-partition derivative). By the standard exponential family identity:
ref_logr ( α ) = − d d α log C r min , r max ( α ) = − C r min , r max ′ ( α ) C r min , r max ( α ) . (55) \operatorname{ref\_logr}(\alpha)=-\frac{d}{d\alpha}\log C_{r_{\min},r_{\max}}(\alpha)=-\frac{C'_{r_{\min},r_{\max}}(\alpha)}{C_{r_{\min},r_{\max}}(\alpha)}. \tag*{(55)} ref_logr ( α ) = − d α d log C r m i n , r m a x ( α ) = − C r m i n , r m a x ( α ) C r m i n , r m a x ′ ( α ) . (55)
Proof. Under the substitution t = 1 / r t=1/r t = 1/ r :
C r min , r max ( α ) = ∫ 1 / r max 1 / r min t α − 1 e − t d t , C r min , r max ′ ( α ) = ∫ 1 / r max 1 / r min t α − 1 e − t log t d t , (56) C_{r_{\min},r_{\max}}(\alpha)=\int_{1/r_{\max}}^{1/r_{\min}}t^{\alpha-1}e^{-t}\,dt,\qquad C'_{r_{\min},r_{\max}}(\alpha)=\int_{1/r_{\max}}^{1/r_{\min}}t^{\alpha-1}e^{-t}\log t\,dt, \tag*{(56)} C r m i n , r m a x ( α ) = ∫ 1/ r m a x 1/ r m i n t α − 1 e − t d t , C r m i n , r m a x ′ ( α ) = ∫ 1/ r m a x 1/ r m i n t α − 1 e − t log t d t , (56)
and ∫ r min r max log ( r ) r − α − 1 e − 1 / r d r = − C r min , r max ′ ( α ) \int_{r_{\min}}^{r_{\max}}\log(r)r^{-\alpha-1}e^{-1/r}\,dr=-C'_{r_{\min},r_{\max}}(\alpha) ∫ r m i n r m a x log ( r ) r − α − 1 e − 1/ r d r = − C r m i n , r m a x ′ ( α ) (since log r = − log t \log r=-\log t log r = − log t ). Hence ref_logr ( α ) = − C r min , r max ′ ( α ) / C r min , r max ( α ) = − ( d / d α ) log C r min , r max ( α ) \operatorname{ref\_logr}(\alpha)=-C'_{r_{\min},r_{\max}}(\alpha)/C_{r_{\min},r_{\max}}(\alpha)=-(d/d\alpha)\log C_{r_{\min},r_{\max}}(\alpha) ref_logr ( α ) = − C r m i n , r m a x ′ ( α ) / C r m i n , r m a x ( α ) = − ( d / d α ) log C r m i n , r m a x ( α ) .
Monotonicity and Fisher Information
Theorem 6 (Strict monotonicity). ref_logr ( α ) \operatorname{ref\_logr}(\alpha) ref_logr ( α ) is strictly decreasing in α \alpha α on ( 0 , 2 ) (0,2) ( 0 , 2 ) :
d d α ref_logr ( α ) = − Var IG ( α , r min , r max ) [ log r ] < 0. (57) \frac{d}{d\alpha}\operatorname{ref\_logr}(\alpha)=-\operatorname{Var}_{\operatorname{IG}(\alpha,r_{\min},r_{\max})}[\log r]<0. \tag*{(57)} d α d ref_logr ( α ) = − Var IG ( α , r m i n , r m a x ) [ log r ] < 0. (57)
Proof. From the exponential family representation, the variance of the sufficient statistic T ( r ) = log r T(r) = \log r T ( r ) = log r equals the second derivative of the log-partition function:
Var I G [ log r ] = d 2 d α 2 log C r min , r max ( α ) > 0. (58) \operatorname{Var}_{\mathrm{IG}}[\log r] = \frac{d^2}{d\alpha^2} \log C_{r_{\min},r_{\max}}(\alpha) > 0. \tag*{(58)} Var IG [ log r ] = d α 2 d 2 log C r m i n , r m a x ( α ) > 0. (58)
Differentiating (55) :
d d α ref_logr ( α ) = − d 2 d α 2 log C r min , r max ( α ) = − Var I G [ log r ] < 0. (59) \frac{d}{d\alpha}\operatorname{ref\_logr}(\alpha) = -\frac{d^2}{d\alpha^2}\log C_{r_{\min},r_{\max}}(\alpha) = -\operatorname{Var}_{\mathrm{IG}}[\log r] < 0. \tag*{(59)} d α d ref_logr ( α ) = − d α 2 d 2 log C r m i n , r m a x ( α ) = − Var IG [ log r ] < 0. (59)
Corollary 6 (Complete-data Fisher information). The Fisher information for α \alpha α from n m nm nm iid draws from IG ( α , r min , r max ) \operatorname{IG}(\alpha,r_{\min},r_{\max}) IG ( α , r m i n , r m a x ) is:
I ( α ) = n m ⋅ Var I G [ log r ] = n m d 2 d α 2 log C r min , r max ( α ) > 0 , (60) I(\alpha) = nm \cdot\operatorname{Var}_{\mathrm{IG}}[\log r] = nm\frac{d^2}{d\alpha^2}\log C_{r_{\min},r_{\max}}(\alpha) > 0, \tag*{(60)} I ( α ) = nm ⋅ Var IG [ log r ] = nm d α 2 d 2 log C r m i n , r m a x ( α ) > 0 , (60) strictly positive for all α ∈ ( 0 , 2 ) \alpha\in(0,2) α ∈ ( 0 , 2 ) .
Proof of Proposition 2
Proof. From the exponential family representation of IG ( α , r min , r max ) \operatorname{IG}(\alpha,r_{\min},r_{\max}) IG ( α , r m i n , r m a x ) (Proposition 5), the natural parameter is η e f = − α \eta_{\mathrm{ef}}=-\alpha η ef = − α and the sufficient statistic is T ( r ) = log r T(r)=\log r T ( r ) = log r . By the standard exponential family identity, the variance of the sufficient statistic equals the second derivative of the log-partition function A ( α ) = log C r min , r max ( α ) A(\alpha)=\log C_{r_{\min},r_{\max}}(\alpha) A ( α ) = log C r m i n , r m a x ( α ) :
Var I G ( α , r min , r max ) [ log r ] = d 2 d α 2 log C r min , r max ( α ) = C r min , r max ′ ′ ( α ) C r min , r max ( α ) − ( C r min , r max ′ ( α ) C r min , r max ( α ) ) 2 > 0 , (61) \operatorname{Var}_{\mathrm{IG}(\alpha,r_{\min},r_{\max})}[\log r] = \frac{d^2}{d\alpha^2}\log C_{r_{\min},r_{\max}}(\alpha) = \frac{C''_{r_{\min},r_{\max}}(\alpha)}{C_{r_{\min},r_{\max}}(\alpha)} - \left(\frac{C'_{r_{\min},r_{\max}}(\alpha)}{C_{r_{\min},r_{\max}}(\alpha)}\right)^2 > 0, \tag*{(61)} Var IG ( α , r m i n , r m a x ) [ log r ] = d α 2 d 2 log C r m i n , r m a x ( α ) = C r m i n , r m a x ( α ) C r m i n , r m a x ′′ ( α ) − ( C r m i n , r m a x ( α ) C r m i n , r m a x ′ ( α ) ) 2 > 0 , (61)
where the inequality holds since IG ( α , r min , r max ) \operatorname{IG}(\alpha,r_{\min},r_{\max}) IG ( α , r m i n , r m a x ) is non-degenerate for 0 < r min < r max < ∞ 0<r_{\min}<r_{\max}<\infty 0 < r m i n < r m a x < ∞ . The Fisher information for α \alpha α from n m nm nm independent radial draws is therefore
F α = n m [ C r min , r max ′ ′ ( α ) C r min , r max ( α ) − ( C r min , r max ′ ( α ) C r min , r max ( α ) ) 2 ] > 0 , \mathcal{F}_{\alpha}=nm\left[\frac{C''_{r_{\min},r_{\max}}(\alpha)}{C_{r_{\min},r_{\max}}(\alpha)}-\left(\frac{C'_{r_{\min},r_{\max}}(\alpha)}{C_{r_{\min},r_{\max}}(\alpha)}\right)^2\right]>0, F α = nm [ C r m i n , r m a x ( α ) C r m i n , r m a x ′′ ( α ) − ( C r m i n , r m a x ( α ) C r m i n , r m a x ′ ( α ) ) 2 ] > 0 ,
which establishes (20) .
Connection to Score Equation
The complete-data IG log-likelihood is:
ℓ I G ( α ) = − n m log C r min , r max ( α ) − ( 1 + α ) ∑ i , j log r i j − ∑ i , j r i j − 1 . (62) \ell_{\mathrm{IG}}(\alpha)=-nm\log C_{r_{\min},r_{\max}}(\alpha)-(1+\alpha)\sum_{i,j}\log r_{ij}-\sum_{i,j}r_{ij}^{-1}. \tag*{(62)} ℓ IG ( α ) = − nm log C r m i n , r m a x ( α ) − ( 1 + α ) i , j ∑ log r ij − i , j ∑ r ij − 1 . (62)
Setting ∂ ℓ I G / ∂ α = 0 \partial\ell_{\mathrm{IG}}/\partial\alpha=0 ∂ ℓ IG / ∂ α = 0 and using (55) :
ref_logr ( α ^ ) = log r ‾ . (63) \operatorname{ref\_logr}(\hat{\alpha})=\overline{\log r}. \tag*{(63)} ref_logr ( α ^ ) = log r . (63)
The MCMC posterior mode satisfies this exactly; the posterior mean satisfies it approximately, with the approximation improving as the posterior concentrates.
The Sign Theorem
Theorem 7 (logr_gap sign theorem). The posterior mean α ^ \hat{\alpha} α ^ satisfies the approximate score equation (63) . Therefore:
sign ( α ^ − α t r u e ) = − sign ( l o g r _ g a p ) , (64) \operatorname{sign}(\hat{\alpha}-\alpha_{\mathrm{true}})=-\operatorname{sign}(\mathrm{logr\_gap}), \tag*{(64)} sign ( α ^ − α true ) = − sign ( logr_gap ) , (64) where l o g r _ g a p = log r c h a i n ‾ − r e f _ l o g r ( α t r u e ) \mathrm{logr\_gap}=\overline{\log r_{\mathrm{chain}}}-\mathrm{ref\_logr}(\alpha_{\mathrm{true}}) logr_gap = log r chain − ref_logr ( α true ) .
Proof. From (63) : r e f _ l o g r ( α ^ ) ≈ r e f _ l o g r ( α t r u e ) + l o g r _ g a p \mathrm{ref\_logr}(\hat{\alpha})\approx\mathrm{ref\_logr}(\alpha_{\mathrm{true}})+\mathrm{logr\_gap} ref_logr ( α ^ ) ≈ ref_logr ( α true ) + logr_gap . If l o g r _ g a p > 0 \mathrm{logr\_gap}>0 logr_gap > 0 , then r e f _ l o g r ( α ^ ) > r e f _ l o g r ( α t r u e ) \mathrm{ref\_logr}(\hat{\alpha})>\mathrm{ref\_logr}(\alpha_{\mathrm{true}}) ref_logr ( α ^ ) > ref_logr ( α true ) , and by strict decrease of r e f _ l o g r \mathrm{ref\_logr} ref_logr (Theorem 6), α ^ < α t r u e \hat{\alpha}<\alpha_{\mathrm{true}} α ^ < α true (negative bias). The case l o g r _ g a p < 0 \mathrm{logr\_gap}<0 logr_gap < 0 gives positive bias by the same argument. If l o g r _ g a p = 0 \mathrm{logr\_gap}=0 logr_gap = 0 then α ^ = α t r u e \hat{\alpha}=\alpha_{\mathrm{true}} α ^ = α true exactly.
Remark 8 (Real-time diagnostic). Theorem 7 provides an immediate convergence check and bias predictor during MCMC:
Table 8.
The diagnostic is independent of ( Ψ , η , κ , σ 2 ) (\Psi,\eta,\kappa,\sigma^2) ( Ψ , η , κ , σ 2 ) and holds under correct specification and spectral misspecification alike.
Numerical Computation
In practice, r e f _ l o g r ( α ) \mathrm{ref\_logr}(\alpha) ref_logr ( α ) is computed via the change of variable t = 1 / r t=1/r t = 1/ r :
r e f _ l o g r ( α ) = D ( α , t u p ) − D ( α , t l o w ) γ ( α , t u p ) − γ ( α , t l o w ) , (65) \mathrm{ref\_logr}(\alpha)=\frac{D(\alpha,t_{\mathrm{up}})-D(\alpha,t_{\mathrm{low}})}{\gamma(\alpha,t_{\mathrm{up}})-\gamma(\alpha,t_{\mathrm{low}})}, \tag*{(65)} ref_logr ( α ) = γ ( α , t up ) − γ ( α , t low ) D ( α , t up ) − D ( α , t low ) , (65)
where t l o w = 1 / r max t_{\mathrm{low}}=1/r_{\max} t low = 1/ r m a x , t u p = 1 / r min t_{\mathrm{up}}=1/r_{\min} t up = 1/ r m i n , γ ( α , x ) = ∫ 0 x t α − 1 e − t d t \gamma(\alpha,x)=\int_{0}^{x}t^{\alpha-1}e^{-t}\,\mathrm{d}t γ ( α , x ) = ∫ 0 x t α − 1 e − t d t is the lower incomplete Gamma function, and D ( α , x ) = ∂ γ ( α , x ) / ∂ α = ∫ 0 x t α − 1 e − t log ( t ) d t D(\alpha,x)=\partial\gamma(\alpha,x)/\partial\alpha=\int_{0}^{x}t^{\alpha-1}e^{-t}\log(t)\,\mathrm{d}t D ( α , x ) = ∂ γ ( α , x ) / ∂ α = ∫ 0 x t α − 1 e − t log ( t ) d t . Both quantities are available in standard numerical libraries.
MCMC initialization
Overview
The Gibbs sampler requires starting values for four parameters: ( α , κ , η , Γ ∞ ) (\alpha,\kappa,\eta,\Gamma_{\infty}) ( α , κ , η , Γ ∞ ) . All initialisations are derived from the observed data { X i } i = 1 n \{X_i\}_{i=1}^{n} { X i } i = 1 n using method-of-moments or projection-based estimators, without using the latent variables { r i j , y i j , s i j } \{r_{ij},y_{ij},s_{ij}\} { r ij , y ij , s ij } which are not yet available at initialisation.
Initialisation of α \alpha α
Koutrouvelis ECF Estimator on Four Projections
The stability index α \alpha α is initialised using the Koutrouvelis [ [17 ] empirical characteristic function (ECF) estimator, applied independently to four univariate projections of the bivariate data { X i } i = 1 n \{X_i\}_{i=1}^n { X i } i = 1 n , X i = ( X i 1 , X i 2 ) T X_i=(X_{i1},X_{i2)^{\mathsf{T}}} X i = ( X i 1 , X i 2 ) T :
Z i ( 1 ) = X i 1 , Z i ( 2 ) = X i 2 , Z i ( 3 ) = X i 1 + X i 2 2 , Z i ( 4 ) = X i 1 − X i 2 2 . (66) Z_i^{(1)}=X_{i1},\qquad Z_i^{(2)}=X_{i2},\qquad Z_i^{(3)}=\frac{X_{i1}+X_{i2}}{\sqrt{2}},\qquad Z_i^{(4)}=\frac{X_{i1}-X_{i2}}{\sqrt{2}}. \tag*{(66)} Z i ( 1 ) = X i 1 , Z i ( 2 ) = X i 2 , Z i ( 3 ) = 2 X i 1 + X i 2 , Z i ( 4 ) = 2 X i 1 − X i 2 . (66)
These correspond to projections onto unit vectors at angles 0 ∘ 0^\circ 0 ∘ , 90 ∘ 90^\circ 9 0 ∘ , 45 ∘ 45^\circ 4 5 ∘ , and 135 ∘ 135^\circ 13 5 ∘ respectively. Each projected series { Z i ( k ) } \{Z_i^{(k)}\} { Z i ( k ) } is univariate α \alpha α -stable with the same stability index α \alpha α .
For each projection k = 1 , … , 4 k=1,\ldots,4 k = 1 , … , 4 , the Koutrouvelis estimator regresses log ( − log ∣ ψ ^ n ( t j ) ∣ 2 ) \log(-\log|\widehat{\psi}_n(t_j)|^2) log ( − log ∣ ψ n ( t j ) ∣ 2 ) against log ∣ t j ∣ \log|t_j| log ∣ t j ∣ at a grid of frequencies { t j } \{t_j\} { t j } :
log ( − log ∣ ψ ^ n ( t j ) ∣ 2 ) = log ( 2 σ α ) + α log ∣ t j ∣ + ϵ j , (67) \log(-\log|\widehat{\psi}_n(t_j)|^2)=\log(2\sigma^\alpha)+\alpha\log|t_j|+\epsilon_j, \tag*{(67)} log ( − log ∣ ψ n ( t j ) ∣ 2 ) = log ( 2 σ α ) + α log ∣ t j ∣ + ϵ j , (67)
where ψ ^ n ( t ) = n − 1 ∑ i = 1 n e i t Z i ( k ) \widehat{\psi}_n(t)=n^{-1}\sum_{i=1}^n e^{\mathrm{i}tZ_i^{(k)}} ψ n ( t ) = n − 1 ∑ i = 1 n e i t Z i ( k ) and the slope gives α ^ K o u t ( k ) \widehat{\alpha}_{\mathrm{Kout}}^{(k)} α Kout ( k ) .
Remark 9 (Choice of projection directions). The four directions 0 ∘ 0^\circ 0 ∘ , 45 ∘ 45^\circ 4 5 ∘ , 90 ∘ 90^\circ 9 0 ∘ , 135 ∘ 135^\circ 13 5 ∘ are chosen for three reasons. First, they form two orthogonal pairs covering the semicircle [ 0 ∘ , 180 ∘ ) [0^\circ,180^\circ) [ 0 ∘ , 18 0 ∘ ) at equal spacing of 45 ∘ 45^\circ 4 5 ∘ — projections beyond 180 ∘ 180^\circ 18 0 ∘ are redundant since ⟨ − u , X ⟩ \langle-u,X\rangle ⟨ − u , X ⟩ has the same stable index as ⟨ u , X ⟩ \langle u,X\rangle ⟨ u , X ⟩ . Second, every possible mean direction η ∈ [ 0 ∘ , 360 ∘ ) \eta\in[0^\circ,360^\circ) η ∈ [ 0 ∘ , 36 0 ∘ ) is within 22.5 ∘ 22.5^\circ 22. 5 ∘ of one of the four directions, ensuring that at least one projection aligns reasonably well with the dominant tail direction regardless of η \eta η . Third, the projections reduce to X i 1 X_{i1} X i 1 , X i 2 X_{i2} X i 2 , ( X i 1 + X i 2 ) / 2 (X_{i1}+X_{i2})/\sqrt{2} ( X i 1 + X i 2 ) / 2 , and ( X i 1 − X i 2 ) / 2 (X_{i1}-X_{i2})/\sqrt{2} ( X i 1 − X i 2 ) / 2 , which require no trigonometric computation. The median over valid estimates provides robustness against any single poorly-aligned projection.
Robustness Steps
Each individual estimate α ^ K o u t ( k ) \widehat{\alpha}_{\mathrm{Kout}}^{(k)} α Kout ( k ) is validated before use:
Extreme values of { Z i ( k ) } \{Z_i^{(k)}\} { Z i ( k ) } are trimmed to the [ 0.1 % , 99.9 % ] [0.1\%,99.9\%] [ 0.1% , 99.9% ] quantile range before applying Kout.
At least 50 observations and 20 distinct values are required; otherwise the estimate is discarded.
The estimate is accepted only if finite and within [ α l o w e r , α u p p e r ] [\alpha_{\mathrm{lower}},\alpha_{\mathrm{upper}}] [ α lower , α upper ] ; otherwise the fallback α ^ 0 = 0.70 \widehat{\alpha}_0=0.70 α 0 = 0.70 is substituted.
The final initialisation is the median of all valid estimates:
α ^ 0 = median { α ^ K o u t ( k ) : k ∈ V } , (68) \widehat{\alpha}_0=\operatorname{median}\{\widehat{\alpha}_{\mathrm{Kout}}^{(k)}:k\in\mathcal{V}\}, \tag*{(68)} α 0 = median { α Kout ( k ) : k ∈ V } , (68)
where V ⊆ { 1 , 2 , 3 , 4 } \mathcal{V}\subseteq\{1,2,3,4\} V ⊆ { 1 , 2 , 3 , 4 } is the set of valid projections (∣ V ∣ ≥ 2 |\mathcal{V}|\ge2 ∣ V ∣ ≥ 2 required; fallback α ^ 0 = 0.70 \widehat{\alpha}_0=0.70 α 0 = 0.70 if ∣ V ∣ < 2 |\mathcal{V}|<2 ∣ V ∣ < 2 ). The result is clipped to [ α l o w e r , α u p p e r ] [\alpha_{\mathrm{lower}},\alpha_{\mathrm{upper}}] [ α lower , α upper ] .
Initialisation of κ \kappa κ and η \eta η
Unit-Vector Projections
The spectral parameters ( κ , η ) (\kappa,\eta) ( κ , η ) are initialised from the directions of the observed data. For each observation X i ∈ R 2 X_i \in\mathbb{R}^2 X i ∈ R 2 with ∥ X i ∥ > 0 \lVert X_i\rVert> 0 ∥ X i ∥ > 0 , compute the unit vector:
u i = X i ∥ X i ∥ = ( cos θ ^ i sin θ ^ i ) , θ ^ i = atan2 ( X i 2 , X i 1 ) ∈ [ 0 , 2 π ) . (69) u_i = \frac{X_i}{\lVert X_i\rVert} = \begin{pmatrix} \cos\widehat{\theta}_i \\ \sin\widehat{\theta}_i \end{pmatrix}, \qquad\widehat{\theta}_i = \operatorname{atan2}(X_{i2},X_{i1}) \in[0,2\pi). \tag*{(69)} u i = ∥ X i ∥ X i = ( cos θ i sin θ i ) , θ i = atan2 ( X i 2 , X i 1 ) ∈ [ 0 , 2 π ) . (69)
Observations with ∥ X i ∥ = 0 \lVert X_i\rVert= 0 ∥ X i ∥ = 0 or non-finite norms are excluded.
Initial η \eta η : Circular Mean Direction
Compute the mean cosine and sine of the projected directions:
C ‾ = 1 n 0 ∑ i ∈ U cos θ ^ i , S ‾ = 1 n 0 ∑ i ∈ U sin θ ^ i , (70) \overline{C} = \frac{1}{n_0}\sum_{i\in\mathcal{U}}\cos\widehat{\theta}_i, \qquad\overline{S} = \frac{1}{n_0}\sum_{i\in\mathcal{U}}\sin\widehat{\theta}_i, \tag*{(70)} C = n 0 1 i ∈ U ∑ cos θ i , S = n 0 1 i ∈ U ∑ sin θ i , (70)
where U \mathcal{U} U is the set of valid observations and n 0 = ∣ U ∣ n_0 = \lvert\mathcal{U}\rvert n 0 = ∣ U ∣ . The mean resultant length and initial mean direction are:
R ^ 0 = C ‾ 2 + S ‾ 2 , (71) \widehat{R}_0 = \sqrt{\overline{C}^{\,2}+\overline{S}^{\,2}}, \tag*{(71)} R 0 = C 2 + S 2 , (71)
η ^ 0 = atan2 ( S ‾ , C ‾ ) m o d 2 π . (72) \widehat{\eta}_0 = \operatorname{atan2}(\overline{S},\overline{C}) \bmod2\pi. \tag*{(72)} η 0 = atan2 ( S , C ) mod 2 π . (72)
Initial κ \kappa κ : Approximate Inversion of R ˉ \bar{R} R ˉ
The concentration κ \kappa κ is initialised by inverting R ˉ ( κ ) = I 1 ( κ ) / I 0 ( κ ) \bar{R}(\kappa)=I_1(\kappa)/I_0(\kappa) R ˉ ( κ ) = I 1 ( κ ) / I 0 ( κ ) at R ^ 0 \hat{R}_0 R ^ 0 . Rather than solving numerically, the following piecewise approximation of Best and Fisher [ [4 ] is used:
κ ^ 0 = { κ l o w e r R ^ 0 < 10 − 8 , 2 R ^ 0 + R ^ 0 3 + 5 6 R ^ 0 5 R ^ 0 < 0.53 , − 0.4 + 1.39 R ^ 0 + 0.43 1 − R ^ 0 0.53 ≤ R ^ 0 < 0.85 , 1 R ^ 0 3 − 4 R ^ 0 2 + 3 R ^ 0 R ^ 0 ≥ 0.85 , (73) \hat{\kappa}_0 =
\begin{cases}
\kappa_{\mathrm{lower}} & \hat{R}_0 < 10^{-8},\\
2\hat{R}_0+\hat{R}_0^3+\frac{5}{6}\hat{R}_0^5 & \hat{R}_0 < 0.53,\\
-0.4+1.39\hat{R}_0+\frac{0.43}{1-\hat{R}_0} & 0.53 \le\hat{R}_0 < 0.85,\\
\frac{1}{\hat{R}_0^3-4\hat{R}_0^2+3\hat{R}_0} & \hat{R}_0 \ge0.85,
\end{cases} \tag*{(73)} κ ^ 0 = ⎩ ⎨ ⎧ κ lower 2 R ^ 0 + R ^ 0 3 + 6 5 R ^ 0 5 − 0.4 + 1.39 R ^ 0 + 1 − R ^ 0 0.43 R ^ 0 3 − 4 R ^ 0 2 + 3 R ^ 0 1 R ^ 0 < 1 0 − 8 , R ^ 0 < 0.53 , 0.53 ≤ R ^ 0 < 0.85 , R ^ 0 ≥ 0.85 , (73)
clipped to [ κ l o w e r , κ u p p e r ] = [ 0.1 , 50 ] [\kappa_{\mathrm{lower}},\kappa_{\mathrm{upper}}]=[0.1,50] [ κ lower , κ upper ] = [ 0.1 , 50 ] .
Remark 13 (Approximation quality). The unit-vector projection u i = X i / ∥ X i ∥ u_i=X_i/\lVert X_i\rVert u i = X i / ∥ X i ∥ extracts the direction of each observation but ignores its magnitude. For heavy-tailed data (α \alpha α small), the direction of X i X_i X i is dominated by the largest radial component r i j r_{ij} r ij and thus approximates the corresponding spectral angle θ i j \theta_{ij} θ ij . For α \alpha α near 2, the data directions are more diffuse and the approximation is less accurate, but the Gibbs sampler corrects this during burn-in.
Initialisation of Γ ∞ \Gamma_\infty Γ ∞
Eight-Direction Projection Estimator
The total spectral mass Γ ∞ \Gamma_\infty Γ ∞ is initialised from the scale of univariate projections across D = 8 D=8 D = 8 equally spaced directions. Let:
ϕ k = ( k − 1 ) π D , k = 1 , … , D , (74) \phi_k=\frac{(k-1)\pi}{D},\qquad k=1,\ldots,D, \tag*{(74)} ϕ k = D ( k − 1 ) π , k = 1 , … , D , (74)
with unit vectors u k = ( cos ϕ k , sin ϕ k ) T u_k=(\cos\phi_k,\sin\phi_k)^\mathsf{T} u k = ( cos ϕ k , sin ϕ k ) T . For each direction k k k , project the median-centred data:
Z i ( k ) = ⟨ u k , X i − X ~ ⟩ , i = 1 , … , n , (75) Z_i^{(k)}=\langle u_k,X_i-\tilde{X}\rangle,\qquad i=1,\ldots,n, \tag*{(75)} Z i ( k ) = ⟨ u k , X i − X ~ ⟩ , i = 1 , … , n , (75)
where X ~ = ( X ~ 1 , X ~ 2 ) T \tilde{X}=(\tilde{X}_1,\tilde{X}_2)^\mathsf{T} X ~ = ( X ~ 1 , X ~ 2 ) T is the componentwise median.
Per-Direction Scale Estimates
For each direction k k k , apply the Koutrouvelis estimator to { Z i ( k ) } \{Z_i^{(k)}\} { Z i ( k ) } to obtain α ^ ( k ) \hat{\alpha}^{(k)} α ^ ( k ) and scale γ ^ ( k ) \hat{\gamma}^{(k)} γ ^ ( k ) (the univariate stable scale parameter). The stable scale satisfies:
γ ^ ( k ) = ( Γ ∞ I ( u k ; α , η , κ ) ) 1 / α , (76) \hat{\gamma}^{(k)}=\left(\Gamma_\infty\mathcal{I}(u_k;\alpha,\eta,\kappa)\right)^{1/\alpha}, \tag*{(76)} γ ^ ( k ) = ( Γ ∞ I ( u k ; α , η , κ ) ) 1/ α , (76)
giving the per-direction Γ ∞ \Gamma_\infty Γ ∞ candidate:
Γ ^ ∞ ( k ) = ( γ ^ ( k ) ) α ^ u s e I ^ ( u k ; α ^ 0 , η ^ 0 , κ ^ 0 ) , (77) \hat{\Gamma}_\infty^{(k)}=\frac{\left(\hat{\gamma}^{(k)}\right)^{\hat{\alpha}_{\mathrm{use}}}}{\hat{\mathcal{I}}(u_k;\hat{\alpha}_0,\hat{\eta}_0,\hat{\kappa}_0)}, \tag*{(77)} Γ ^ ∞ ( k ) = I ^ ( u k ; α ^ 0 , η ^ 0 , κ ^ 0 ) ( γ ^ ( k ) ) α ^ use , (77)
where α ^ 0 \hat{\alpha}_0 α ^ 0 is the estimate from Section G.2, and η ^ 0 \hat{\eta}_0 η ^ 0 , κ ^ 0 \hat{\kappa}_0 κ ^ 0 are from Section G.3.
Numerical Evaluation of I \mathcal{I} I
The scale integral for direction u k = ( cos ϕ k , sin ϕ k ) T \boldsymbol{u}_k = (\cos\phi_k, \sin\phi_k)^{\mathsf{T}} u k = ( cos ϕ k , sin ϕ k ) T is evaluated numerically on a grid of G = 4096 G = 4096 G = 4096 equally spaced angles θ g = 2 π g / G \theta_g = 2\pi g/G θ g = 2 π g / G , g = 0 , … , G − 1 g = 0, \ldots, G - 1 g = 0 , … , G − 1 :
I ^ ( u k ; α , η , κ ) = ∑ g = 0 G − 1 ∣ ⟨ u k , s ( θ g ) ⟩ ∣ α f v M ( θ g ; η , κ ) ⋅ 2 π G , (78) \widehat{\mathcal{I}}(\boldsymbol{u}_k; \alpha, \eta, \kappa) = \sum_{g=0}^{G-1} \left|\langle\boldsymbol{u}_k, s(\theta_g)\rangle\right|^\alpha f_{\mathrm{vM}}(\theta_g; \eta, \kappa) \cdot\frac{2\pi}{G}, \tag*{(78)} I ( u k ; α , η , κ ) = g = 0 ∑ G − 1 ∣ ⟨ u k , s ( θ g )⟩ ∣ α f vM ( θ g ; η , κ ) ⋅ G 2 π , (78)
where s ( θ ) = ( cos θ , sin θ ) T s(\theta) = (\cos\theta, \sin\theta)^{\mathsf{T}} s ( θ ) = ( cos θ , sin θ ) T and f v M ( θ ; η , κ ) = e κ cos ( θ − η ) / ( 2 π I 0 ( κ ) ) f_{\mathrm{vM}}(\theta; \eta, \kappa) = e^{\kappa\cos(\theta-\eta)}/(2\pi I_0(\kappa)) f vM ( θ ; η , κ ) = e κ c o s ( θ − η ) / ( 2 π I 0 ( κ )) . Noting ⟨ u k , s ( θ g ) ⟩ = cos ( ϕ k − θ g ) \langle\boldsymbol{u}_k, s(\theta_g)\rangle= \cos(\phi_k - \theta_g) ⟨ u k , s ( θ g )⟩ = cos ( ϕ k − θ g ) , this simplifies to:
I ^ ( u k ; α , η , κ ) = 1 G I 0 ( κ ) ∑ g = 0 G − 1 ∣ cos ( ϕ k − θ g ) ∣ α exp ( κ cos ( θ g − η ) ) , (79) \widehat{\mathcal{I}}(\boldsymbol{u}_k; \alpha, \eta, \kappa) = \frac{1}{G I_0(\kappa)} \sum_{g=0}^{G-1} \left|\cos(\phi_k - \theta_g)\right|^\alpha\exp\bigl(\kappa\cos(\theta_g - \eta)\bigr), \tag*{(79)} I ( u k ; α , η , κ ) = G I 0 ( κ ) 1 g = 0 ∑ G − 1 ∣ cos ( ϕ k − θ g ) ∣ α exp ( κ cos ( θ g − η ) ) , (79)
a Riemann sum with G G G points giving negligible approximation error.
Final Estimate
Valid candidates are those with γ ^ ( k ) > 0 \widehat{\gamma}^{(k)} > 0 γ ( k ) > 0 , I ^ ( k ) > 0 \widehat{\mathcal{I}}^{(k)} > 0 I ( k ) > 0 , and Γ ^ ∞ ( k ) ∈ ( 0 , ∞ ) \widehat{\Gamma}^{(k)}_{\infty} \in(0,\infty) Γ ∞ ( k ) ∈ ( 0 , ∞ ) . The final initialisation is the median over valid directions:
Γ ^ ∞ , 0 = median { Γ ^ ∞ ( k ) : k ∈ V } , (80) \widehat{\Gamma}_{\infty,0} = \operatorname{median}\left\{\widehat{\Gamma}^{(k)}_{\infty}: k \in\mathcal{V}\right\}, \tag*{(80)} Γ ∞ , 0 = median { Γ ∞ ( k ) : k ∈ V } , (80)
where V \mathcal{V} V is the set of valid directions, clipped to [ z l o w e r , z u p p e r ] = [ 0.1 , 50 ] [z_{\mathrm{lower}}, z_{\mathrm{upper}}] = [0.1, 50] [ z lower , z upper ] = [ 0.1 , 50 ] .
Remark 14 (Robustness of the eight-direction estimator). Using D = 8 D = 8 D = 8 directions and taking the median provides robustness against poor scale estimates in any single projection direction. If fewer valid directions are available, the median is taken over those that are valid. Errors in ( α ^ u s e , η ^ 0 , κ ^ 0 ) (\widehat{\alpha}_{\mathrm{use}}, \widehat{\eta}_0, \widehat{\kappa}_0) ( α use , η 0 , κ 0 ) propagate to Γ ^ ∞ , 0 \widehat{\Gamma}_{\infty,0} Γ ∞ , 0 through I ^ \widehat{\mathcal{I}} I , but are corrected during the 10,000-iteration burn-in.