ingemar kaj department of mathematics uppsala university ... · department of mathematics uppsala...
TRANSCRIPT
Scaling properties of Poisson germ-grain models
Ingemar Kaj
Department of MathematicsUppsala [email protected]
Reisensburg Workshop, Feb 2007
Purpose
I In spatial stochastic structure,investigate (long-range)dependence over distance
I independently scattered versusfractal dependence
I Use Poisson germ-grain models,to
- generalize known 1-dimresults (M/G/∞)
- unify heavy-tailed andmicroballs models
- understand self-similarityproperties, etc
0 50 100 150 200
20
40
60
80
100
120
140
160
180
200
Outline
I Poisson germ-grain
I grain size distribution with heavy tails/singular intensities
I random fields
I scaling analysis; zoom-in or zoom-outI three different limit regimes
- Gaussian- intermediate, Poisson- (stable)
I self-similarityI examples
- fractional Brownian field- “fractional Poisson motion”
Presentation based on joint works
Kaj, Leskela, Norros, Schmidt: Scaling limits for random fields withlong-range dependence. Ann. Probab. 35:2, 2007.
Bierme, Estrade, Kaj: About scaling behavior of random ballsmodels. 6th Int. Conf. on Stereology, Spatial Statistics andStochastic Geometry, Prague, June 2006.
Poisson point process
[grains = balls, in this presentation]
I Construct family of grains inRd, {Xj +B(0, Rj)}, builtfrom Poisson point process inRd ×R+.
I Model: Poisson measureN(dx, dr) on Rd ×R+ withintensity measuren(dx, dr) = dxF (dr)
I if F (dr) distribution: eachpoint (x, r) is random ballwith center x, radius r.
I if not: Microballs model
Background
The analog for d = 1 is the M/G/∞model.
0 0.2 0.4 0.6 0.8 10
2
4
6
8
10
12
I With heavy-tailed service times such models underaggregation and proper scaling fluctuate as FractionalBrownian Motion (H > 1/2) in the asymptotic limit.
I More generally, three limit regimes: Gaussian, Intermediate,Stable (IK/Gaigalas, Bernoulli 03; IK, Fractals in Eng 05;IK/Taqqu 2007, etc)
I Bierme and Estrade, JAP 2006, consider micro ball model andderives limit which shares properties of FBM with Hurst indexH < 1/2
Model assumptions
Germ-grain intensity n(dx, dr) = dxF (dr), where F (dr) σ-finitemeasure, with∫
R+rd F (dr) < +∞ finite expected “volume”
and, for β 6= d, F (dr) = f(r)dr where
f(r) ∼ Cβ r−β−1 as r → 0d−β =
{r → 0, β < dr →∞, β > d
Yields random configuration of(super-positioned) mass in Rd.Large grains keep the massconfigurations rigid, small grainscause non-trivial noise effects.
Scaling idea
Study fluctuations around the expected mass configuration{zooming-out (many small heavy-tailed grains limit), d < β (< 2d)
zooming-in (not so many large micro-grains limit), β < d
Scaling parameters
Takenλ,ρ(dx, dr) = λdxFρ(dr)
where Fρ is image measure under r → ρr. As{λ→∞, ρ→ 0 zoom-out (β > d)λ→ 0, ρ→∞ zoom-in (β < d)
the relative speed of Poisson intensity λ and scaled radius ρ, determines
the limit regime:
I Gaussian; λρβ →∞I Intermediate; λρβ → constI Stable; λρβ → 0
Interpretation: E(# of balls of radius > 1 which covers a point) ∼ λρβ
Scaling parameters
Takenλ,ρ(dx, dr) = λdxFρ(dr)
where Fρ is image measure under r → ρr. As{λ→∞, ρ→ 0 zoom-out (β > d)λ→ 0, ρ→∞ zoom-in (β < d)
the relative speed of Poisson intensity λ and scaled radius ρ, determines
the limit regime:
I Gaussian; λρβ →∞I Intermediate; λρβ → constI Stable; λρβ → 0
Interpretation: E(# of balls of radius > 1 which covers a point) ∼ λρβ
Random fields
Pointwise field:
W (y) =∫Rd
∫ ∞
0
1{|y−x|<r}N(dx, dr) = # grains containing y
with
EW (y) =∫Rd
∫ ∞
0
1{|y−x|<r} dxF (dr) = |B(y, 1)|∫Rd
rd F (dr) <∞
More general indexing tool: for measures µ, consider
µ 7→W (µ) =∫Rd
∫ ∞
0
µ(B(x, r))N(dx, dr).
Stochastic integral is well defined as limit in prob. of elementary integrals.Which measures µ?
Fluctuation field
The centered fluctuations are described by the compensated stochasticintegral
Yλ,ρ(µ) =∫Rd
∫ ∞
0
µ(B(x, r)) (Nλ,ρ(dx, dr)− λdxFρ(dr))
Limit problem:
Find normalization n(λ, ρ) such that
Yλ,ρ(µ)/n(λ, ρ) d⇒ Y (µ).
What are the possible limits Y ? Properties of Y (µ)?
Finite energy measures
M1: signed measures of finite total variation |µ|.
For α 6= d define subspace of measures
Mα ={µ ∈M1;
∫Rd×Rd
|z − z′|d−α|µ|(dz)|µ|(dz′) < +∞ and
∫Rd×Rd
|z − z′|2kµ(dz)µ(dz′) = 0 all k s.t. 0 ≤ 2k < d− α}
For β ∈ (0, 2d) with β 6= d, if α > β > d or if α < β < d thenMα ⊂Mβ . Define
Mβ =
⋃
α∈(β,2d)
Mα if β > d
⋃α∈(0,β)
Mα if β < d
Remarks:
I Mα = {0}, if α ≥ 2d
I Mα ={µ ∈M1;
∫Rd×Rd |z − z′|d−α|µ|(dz)|µ|(dz′) < +∞
}, if
d < α < 2d
I Condition∫Rd×Rd |z − z′|2kµ(dz)µ(dz′) = 0 can be replaced by∫
Rd
zj11 z
j22 . . . zjd
d µ(dz) = 0
for integers j1, j2, . . . , jd such that 0 ≤ j1 + j2 + . . .+ jd < d− α
(c.f Matheron 1973, where similar spaces of measures are introduced)
Fractional Gaussian limit
Thm: Let F (dr) ∼ Cβr−β−1 be as above with β < 2d such that
β − d 6∈ 2N. If λρβ →∞ as ρ→ 0β−d, then
Yλ,ρ(µ)√Cβλρβ
f.d.d.=⇒ constWβ(µ), µ ∈ Mβ
where Wβ is a Gaussian random field µ→Wβ(µ) with covariancefunctional
E (Wβ(µ)Wβ(ν)) = cβ
∫Rd
∫Rd
|z − z′|d−β µ(dz)ν(dz′),
and the constant cβ ensures that the covariance is of positive type.
Intermediate limit
Thm: Again, take F (dr) ∼ Cβr−β−1 with β < 2d such that β − d 6∈ 2N.
For arbitrary c > 0, if
λρβ → cβ−d as ρ→ 0β−d,
then
Yλ,ρ(µ)f.d.d.=⇒ constYβ(µc),
where µc(A) = µ(cA), A ⊂ Rd, µ ∈ Mβ , and
Yβ(µ) =∫Rd
∫ ∞
0
µ(B(x, r)) (Nβ(dx, dr)− Cβ dx r−β−1dr).
Covariance
In heavy-tailed case d < β < 2d we have
Var(Wβ(µ)) = Var(Yβ(µ)) = const∫Rd
∫Rd
µ(dy)µ(dy′)|y − y′|β−d
.
For microballs case d− 1 < β < d (0 < H < 1/2), the covariancefunctional has the representation
E (Wβ(µ)Wβ(ν))
=12
∫Rd
∫Rd
(|z|d−β + |z′|d−β − |z − z′|d−β
)µ(dz)ν(dz′).
Similarity properties
I the Gaussian field is self-similar with indexH = (d− β)/2 ∈ (−d/2,∞), i.e. for all s > 0,
W (µs)fdd= sHW (µ), where µs(A) = µ(A/s).
I the intermediate limit is “aggregate-similar” in the sense thatthe scaled process passes through all aggregates of i.i.d.copies: there exists a sequence cm →∞ such that
Yβ(µ1/cm)
fdd=
m∑i=1
Y iβ(µ)
Case d = 1
If 1 < β < 2 then H = (1− β)/2 ∈ (−1/2, 0) and
Var(Wβ(µ)) =∫R
∫R
µ(du)µ(dv)|u− v|−2H
.
The choice µt(du) = 1[0,t](u) du is in Mβ , and yields
Var(Wβ(µt)) =∫ t
0
∫ t
0
du dv
|u− v|2(1−H′)= const t2H′
,
with H ′ = H + 1, so BH(t) = Wβ(µt) is ordinary FBM with1/2 < H ′ < 1.
For 0 < β < 1, take
µt(du) = δt(du)− δ0(du) ∈ Mβ .
Then BH(t) = Wβ(µt) is ordinary FBM with 0 < H < 1/2.
Fractional Poisson motion
Consider intermediate limit in Rd, microballs case d− 1 < β < d. Takeµ = δx − δ0, which belongs to Mβ . Then
x→ Yβ(x) = Yβ(δx − δ0)
has a symmetrized Poisson distribution with mean |x|2H and covariance
Cov(Yβ(x), Yβ(y)) =12(|x|2H+|y|2H−|x−y|2H), H =
β − d2
∈ (0, 1/2)
Comparison with short range case
Thm: Assume ER2d <∞. As ρ→ 0, λ→∞,
Yλ,ρ(µ)(λρ2d)1/2E(R2d)1/2
f.d.d=⇒ W (µ), µ(dx) = φ(x) dx, φ ∈ L1 ∩ L2
where W (µ) is the Gaussian functional with covariance∫Rd φ(x)ψ(x) dx (white noise).