Log Ch.F. for Stable Distributions General St(α, β, γ, δ): ( iδω − γ|ω|α + iβγ tan πα {|ω|α sgn ω − ω} α 6= 1 2 log φX (ω) = iδω − γ|ω| − π2 iβγω log |ω| α=1 ( πα α iω δ − βγ tan πα − γ|ω| 1 − iβ tan sgn ω α 6= 1 2 2 = iω (δ − 2βγ log |ω|/π) − γ|ω| α=1 0 < α ≤ 2, −1 ≤ β ≤ 1, 0 < γ < ∞, −∞ < δ < ∞. Standard St(α, β, 1, 0): ( {|ω|α sgn ω − ω} α 6= 1 −|ω|α + iβ tan πα 2 log φX (ω) = −|ω| − π2 iβω log |ω| α=1 Symmetric St(α, 0, γ, 0): log φX (ω) = −γ|ω|α Skewed St(α, 1, γ, 0): ( −γ|ω|α + iγ tan πα {|ω|α sgn ω − ω} α 6= 1 2 log φX (ω) = −γ|ω| − π2 iγω log |ω| α=1 Standard Stable Lévy Measure We now turn to computing the Lévy measure for the standard St(α, β, 1, 0) stable distribution. For α < 0 and > 0 set ν+ (u) ≡ u−α−1 e−u on R+ and 1 compute Z ∞ e iωu ν+ (du) = 0 = ∞ Z eiωu u−α−1 e−u du Z0 ∞ u−α−1 e−u(−iω) du 0 = Γ(−α)( − iω)α ω = Γ(−α)(2 + ω 2 )α/2 e−iα atan Z ∞ ν+ (du) = Γ(−α)α 0 Z ∞ Z ∞ −α−1 −u(−i) u e du sin u ν+ (du) = = 0 0 = Γ(−α)=( − i)α = −Γ(−α)(2 + 1)α/2 sin(α atan 1 ). Thus the ch.f. with Lévy measure given on R+ by ν+ (du) ≡ u−α−1 e−u du is Z ∞ log φ+ (ω) = eiωu − 1 − iω sin u ν+ (du) 0 n ω ω = Γ(−α) (2 + ω 2 )α/2 cos(α atan ) − i sin(α atan ) o 1 −α + iω(2 + 1)α/2 sin(α atan ) This is analytic in α (away from the nonpositive integers), so we first take the analytic continuation to the segment 0 < α < 2, then the limit as → 0: n o πα πα − i sin |ω|α sgn ω − ω (as → 0) → Γ(−α) |ω|α cos 2 2 n o −π πα α α = |ω| − i tan |ω| sgn ω − ω . 2αΓ(α) sin πα 2 2 Similarly the ch.f. with Lévy measure ν− (u) ≡ |u|−α−1 e−|u| du on R− is Z 0 eiωu − 1 − iω sin u ν− (du) −∞ n o −π πα α α → |ω| + i tan |ω| sgn ω − ω 2αΓ(α) sin πα 2 2 log φ− (ω) = 2 so for −1 ≤ β ≤ 1 the ch.f. for Lévy measure ν(du) = πα α Γ(α) sin (1 + β sgn u) |u|−α−1 du π 2 (1) is log φ(ω) = Z ∞ −∞ eiωu − 1 − iω sin u ν(du) = −|ω|α + iβ tan πα |ω|α sgn ω − ω , 2 exactly the Cheng and Liu parametrization of the St(α, β, 1, 0) distribution. The Lévy measure and ch.f. for case α = 1 follow from a similar argument or simply from continuity: ν(du) = 1 + β sgn u |u|−2 du π log φ(ω) = −|ω| − 2βi ω log |ω|. π Four-Parameter Stable Distributions Cheng and Liu (1997) take the remaining parameters γ and δ of the fourparameter family of stable distributions X ∼ St(α, β, γ, δ) to be scale and location parameters, respectively, so X = γZ + δ for some Z ∼ St(α,P β, 1, 0); ∗ ∗ with this choice the parameters γ , δ governing the sum X+ ≡ Xi ∼ ∗ ∗ St(α, β, γ , δ ) of independent stable random variables Xi ∼ St(α, β, γi , δi ) is exceedingly awkward, and it is difficult to describe the stable random fields we will need below. We instead take γ to be a rate parameter (and δ location), leading to the general form πα αγ Γ(α) sin (1 + β sgn u) |u|−α−1 du, π Z ∞ 2 eiωu − 1 − iω sin u ν(du) log φ(ω) = iδω + −∞ ( iδω − γ|ω|α + iβγ tan πα |ω|α sgn ω − ω α 6= 1; 2 = iδω − γ|ω| − iβγ π2 ω log |ω| α = 1. ν(du) = (2) (3) (4) P With this choice the sum X+ ≡ Xi of independent stable random variables Xi ∼ St(α, β, γi, δi ) has a stable distribution P X+ ∼ St(α, Pβ, γ+, δ+ ) with parameters given simply by the sums γ+ = γi and δ+ = δi . 3 Location-Scale Family If X is a stable random variable X ∼ St(α, β, γ, δ) and Y ≡ η + σX is in the same location-scale family for some η, σ ∈ R (note σ need not be positive) then Y ∼ St(α∗ , β ∗ , γ ∗ , δ ∗ ) with α∗ = α β ∗ = β sgn σ γ ∗ = γ|σ|α ( η + σδ + βγ tan πα (|σ|α sgn σ − σ) α 6= 1 2 δ∗ = η + σδ − βγ π2 σ log |σ| α=1 It is straightforward to solve these for the parameters η and σ needed to achieve standard form Y ≡ η + σX ∼ St(α, β, 1, 0) with γ ∗ = 1, δ ∗ = 0: ( −σδ + β tan πα (σ 1−α − 1) α 6= 1 2 σ = γ −1/α η= −σδ + β π2 log σ α = 1. Arbitrary Compensators The Lévy-Khinchine representation Z iωu ∗ log φ(ω) = iωδ + e − 1 − iωh(u) ν(du) R may be written with any bounded compensator h(u) = u + o(u2 ) near u ≈ 0. The choice h(u) = sin u leads to δ ∗ = δ, as in Eqn. (3), while any other bounded compensator h(u) = u + o(u2 ) may require a different shift, Z ∗ δ = δ + [h(u) − sin u] ν(du). For example, the function h(u) ≡ u 1{u2 <c2 } that fully compensates only jumps of size |u| < c would lead to Z ∗ δ = δ + [u 1{u2 <c2 } − sin u] ν(du) 1−α αc − Γ(2 − α) sin πα 2βγΓ(α) sin πα 2 2 , α 6= 1 = δ+ π(1 − α) [log c + γe − 1], for α = 1, where and (by L’Hôpital’s rule) δ ∗ = δ + 2βγ π γe ≈ 0.577216 denotes the Euler-Mascheroni constant. 4 Explicit Examples with α = 3/2 Example St(3/2, 1, 1, 0): For example, the fully-skewed Stable distribution with index α = 3/2 has Lévy measure 3 ν(u) = √ u−5/2 , u > 0. 2 2π With the standard compensator h(u) = sin u it has offset δ √ = 0, but with ∗ compensator h(u) = 1{u2 <1} it would have offset δ = 1 − 3/ 2π and, with √ h(u) = 1{u2 <2 } , δ∗ = 1 − 3/ 2π. ILM for St(3/2, 1, 1, 0): A random variable X ∼ St(3/2, 1, 1, 0) may be generated by the ILM algorithm by fixing > 0, constructing a sequence τn of jumps of a standard Poisson process, setting ν + (u) = ν [u, ∞) Z ∞ 3 √ x−5/2 dx = 2 2π u √ −3/2 = u / 2π, ν − (t) = sup[u > 0 : ν + (u) > t] √ = (2π)−1/3 t−2/3 = (t 2π)−2/3 √ υn = ν − (τn ) = (τn 2π)−2/3 Z ∞ X X = υn − h(u) ν(du) + δ ∗ υn > 1 3 u √ u−5/2 du + δ ∗ 2 2π υn > X √ √ = υn − 3(−1/2 − 1)/ 2π + [1 − 3/ 2π] = X υn > = X υn > υn − Z √ υn + 1 − 3/ 2π. 5 √ iid Equivalently, we may take J ∼ Po −3/2 / 2π and draw υj ∼ Pa(3/2, ) from the Pareto distribution with density (3/2)3/2 u−5/2 , u > , and set X = J X n=1 √ υn + 1 − 3/ 2π. Note the series would not converge without compensation; but with compensation, X → X0 ∼ St(3/2, 1, 1, 0) as → 0. A Stable process may be constructed similarly; if we draw σn ∼ Un(0, 1) independently of τn , for example, we can set X 3 −1 X (t) = υn − t √ 2π υn >,σn ≤t and again take → 0. ILM for St(α, β, γ, δ): A random variable X ∼ St(α, β, γ, δ) may be generated by the ILM algorithm too, by fixing > 0, constructing a sequence τn of jumps of a standard Poisson process and a sequence ρn of ±1-valued independent random variables equal 6 to +1 with probability (1 + β)/2, setting ν + (u) = ν (−∞, −u] ∪ [u, ∞) γ u−α = Γ(1 − α) cos πα 2 − + ν (t) = sup[u > 0 : ν (u) > t] = υn X 2Γ(α)γ sin πα 2 πt 1/α 2Γ(α)γ sin πα 2 = ν (τn ) = πτ Z ∞ n X = υn ρn − h(u) ν(du) + δ ∗ 1/α − υn > πα 1−α 2βγ −1 Γ(1 + α) sin π(α − 1) 2 υn > πα Γ(1 + α) πα 2 + sin +δ −βγ tan 2 π 2 α−1 X 2αΓ(α) sin πα πα 2 1−α + tan = υn ρn − βγ +δ π(α − 1) 2 υ > X = υn ρn − n A Stable process may be constructed similarly on [0, T ] by drawing σn ∼ Un(0, T ) independently of τn and setting X 2αΓ(α) sin πα πα 1/α 2 1−α X (t) ≡ T + δt, υn ρn − βγt + tan π(α − 1) 2 υ >,σ ≤t n n 0 ≤ t ≤ T . A d-dimensional Stable RF may be constructed on [0, T ]d by X X [φ] ≡ T d/α υn ρn φ(σn ) υn > Z 2αΓ(α) sin πα πα 2 1−α + tan − βγ −δ φ(s) dd s, π(α − 1) 2 d [0,T ] for integrable φ ∈ L1 [0, T ]d . The distribution can be generalized somewhat by replacing γ and δ by nonnegative measures (on an arbitrary measure space), and α and β by measurable functions. The distribution really ought to be parametrized by γ α , not γ, changing the standardization relation to X = δ + γ 1/α Z, Z ∼ St(α, β, 1, 0). 7 Truncation Error Bounds: For η < < 1 the difference Xη − X is given by πα X α 2Γ(1 + α) sin 2 Xη − X = υn ρn − βγ η 1−α − 1−α , π(α − 1) η<υ ≤ n a mean-zero random variable with variance Z 2 E(Xη − X ) = u2 ν(du) η Z 2γ α Γ(1 + α) sin πα 2 u1−α = π η πα α 2γ Γ(1 + α) sin 2 2−α − η 2−α = π(2 − α) α 2γ Γ(1 + α) sin πα 2 → 2−α as η → 0, π(2 − α) so X converges in L2 as → 0 with the indicated L2 truncation bound. SimilarRbounds apply to Stable processes and fields, with an additional factor of t or φ2 respectively. References Cheng, R. C. H. and Liu, W. B. (1997) A continuous representation of the family of stable law distributions. J. Roy. Stat. Soc. B, 59, 137–145. 8