Characterization and Estimation of the Variations of a Random Convex Set by its Mean n-Variogram : Application to the Boolean Model S.Rahmani, J-C.Pinoli & J.Debayle Ecole Nationale Supérieure des Mines de Saint-Etienne,FRANCE SPIN, PROPICE / LGF, UMR CNRS 5307 28/10/2015 SR (ENSM-SE / LGF-PMDM) GSI 2015 28/10/2015 1 / 22 Geometric Stochastic Modeling and objectives Section 1 Geometric Stochastic Modeling and objectives SR (ENSM-SE / LGF-PMDM) GSI 2015 28/10/2015 2 / 22 Geometric Stochastic Modeling and objectives Stochastic materials Material modelling Material characterization SR (ENSM-SE / LGF-PMDM) GSI 2015 28/10/2015 3 / 22 Geometric Stochastic Modeling and objectives Germ-Grain model [Matheron 1967] Definition Ξ= [ xi + Ξi (1) xi ∈Φ The Ξi are i.i.d. Φ a point process Law of Φ Law of Ξ0 ⇔ ⇔ Spatial distribution granulometry Boolean model ⇒ Φ Poisson point process of intensity λ SR (ENSM-SE / LGF-PMDM) GSI 2015 28/10/2015 4 / 22 Geometric Stochastic Modeling and objectives Objectives and state of the art Geometrical characterization of Ξ0 from measurements in a bounded window Ξ ∩ M No assumption on Ξ0 ’s shape. Describing Ξ0 . State of the art Miles formulae [Miles 1967] Tangent points method [Molchanov 1995] Minimum contrast method[ Dupac & Digle 1980] ⇒ Mean geometric parameter λ, E[A(Ξ0 )], E[U(Ξ0 )] Formula for distribution for model of disk [Emery 2012] SR (ENSM-SE / LGF-PMDM) GSI 2015 28/10/2015 5 / 22 Geometric Stochastic Modeling and objectives Characterization and description of the grain For homothetic grains: E[U(Ξ0 )] 2π E[U(Ξ0 )] = 4 E[A(Ξ0 )] π Disk of radius r : E[r ] = & E[r 2 ] = Square of side x :E[x] & E[r 2 ] = E[A(Ξ0 )] ⇒ Parametric distribution of homothetic factor! For non homothetic grains: rectangle, ellipse... Same mean for area and perimeter (Minkowski densities) ⇒ insufficient to fully characterize Ξ0 ! What about the variations of these geometrical characteristics? SR (ENSM-SE / LGF-PMDM) GSI 2015 28/10/2015 6 / 22 Theoretical aspects Section 2 Theoretical aspects SR (ENSM-SE / LGF-PMDM) GSI 2015 28/10/2015 7 / 22 Theoretical aspects From covariance of Ξ to variation of Ξ0 Mean covariogram: γ̄Ξ0 (u) = E[A(Ξ0 ∩ Ξ0 + u)] Covariance: CΞ (u) = P(x ∈ (Ξ ∩ Ξ + u)) Relationship: CΞ (u) − pΞ2 1 γ̄Ξ0 (u) = log 1 + γ (1 − pΞ )2 (2) In addition: R SR (ENSM-SE / LGF-PMDM) R2 γ̄Ξ0 (u)du = E[A(Ξ0 )2 ] GSI 2015 28/10/2015 8 / 22 Theoretical aspects Stability by convex dilations Ξ Ξ⊕K (a) grain Ξ0 , intensity λ (b) grain Ξ0 ⊕ K , intensity λ Where X ⊕ Y = {x + y |x ∈ X , y ∈ Y } ⇒ The Boolean model is stable under convex dilations SR (ENSM-SE / LGF-PMDM) GSI 2015 28/10/2015 9 / 22 Theoretical aspects The proposed method Consequently, for all r ≥ 0 we can estimate: Z 2 ζ0,K (r ) = E[A(Ξ0 ⊕ rK ) ] = E[γΞ0 ⊕rK (u)]du R2 SR (ENSM-SE / LGF-PMDM) GSI 2015 28/10/2015 10 / 22 Theoretical aspects The proposed method Consequently, for all r ≥ 0 we can estimate: Z 2 ζ0,K (r ) = E[A(Ξ0 ⊕ rK ) ] = E[γΞ0 ⊕rK (u)]du R2 Steiner’s formula (mixed volumes) A(Ξ0 ⊕ rK ) = A(Ξ0 ) + 2rW (Ξ0 , K ) + r 2 A(K ) The polynomial ζ0,K ζ0,K (r ) = E[A20 ] + 4r E[A0 W (Ξ0 , K )] + r 2 (4E[W (Ξ0 , K )2 ] + + 2A(K )E[A0 ]) + 4r 3 A(K )E[W (Ξ0 , K )] + r 4 A(K )2 SR (ENSM-SE / LGF-PMDM) GSI 2015 28/10/2015 10 / 22 Theoretical aspects The proposed method Consequently, for all r ≥ 0 we can estimate: Z 2 ζ0,K (r ) = E[A(Ξ0 ⊕ rK ) ] = E[γΞ0 ⊕rK (u)]du R2 Steiner’s formula (mixed volumes) A(Ξ0 ⊕ rK ) = A(Ξ0 ) + 2rW (Ξ0 , K ) + r 2 A(K ) The polynomial ζ0,K ζ0,K (r ) = E[A20 ] + 4r E[A0 W (Ξ0 , K )] + r 2 (4E[W (Ξ0 , K )2 ] + + 2A(K )E[A0 ]) + 4r 3 A(K )E[W (Ξ0 , K )] + r 4 A(K )2 ⇒ Estimation of E[A20 ], E[A0 W (Ξ0 , K )] and E[W (Ξ0 , K )2 ] SR (ENSM-SE / LGF-PMDM) GSI 2015 28/10/2015 10 / 22 Theoretical aspects Generalization to nth order moments The mean n-variogram Tn−1 (n) For n ≤ 2, γΞ0 (u1 , · · · un−1 ) = E[A( i=1 (Ξ0 − ui ) ∩ Ξ0 )] Relation n-variogram → n point probability function (see proceding) R R (n) Of course R2 · · · R2 γΞ0 (u1 , · · · un−1 )du1 · · · dun−1 = E[A(Ξ0 )n ] Then the development of E[A(Ξ0 ⊕ K )n ] by Steiner’s formula gives: ∀K convex, nth order moments of (A0 , W (Ξ0 , K )) SR (ENSM-SE / LGF-PMDM) GSI 2015 28/10/2015 11 / 22 Theoretical aspects The interpretation of the mixed area Definition For Ξ0 and K convex, W (Ξ0 , K ) = 12 (A(Ξ0 ⊕ K ) − A(K )) For unit ball :W (Ξ0 , B) = U(Ξ) the perimeter For a segment: W (Ξ0 , Sθ ) = HΞ0 (θ) the Féret’s diameter HΞ0 (θ) Ξ0 Ox θ For a polygon W (Ξ, LN i=1 αi Sθi ) = PN i=1 αi HΞ0 (θi ) ⇒ ∀N, ∀(θ1 , · · · θN ) all moments of (HΞ0 (θ1 ), · · · HΞ0 (θN )) ⇒ Characterization of the random process HΞ0 SR (ENSM-SE / LGF-PMDM) GSI 2015 28/10/2015 12 / 22 Theoretical aspects The Féret’s diameter random process HΞ0 trajectory of HΞ0 is the support function of the realization Ξ0 ⊕ Ξ˘0 Ξ0 θ HΞ0 (θ) Feret diameter HΞ0 (θ) 8 7 6 5 4 0 π/2 π 3π/2 2π orientation θ The process HΞ0 describes and characterizes Ξ0 ⊕ Ξ˘0 NB: Ξ0 isotropic ⇔ HΞ0 strong stationary SR (ENSM-SE / LGF-PMDM) GSI 2015 28/10/2015 13 / 22 Practical aspects Section 3 Practical aspects SR (ENSM-SE / LGF-PMDM) GSI 2015 28/10/2015 14 / 22 Practical aspects The simplest cases Estimation of 1st and 2nd-order moments E[A(Ξ0 )] and E[W (Ξ0 , K )] E[A(Ξ0 )2 ], E[A(Ξ0 )W (Ξ0 , K )] and E[W (Ξ0 , K )2 ] Disk Segment E[A(Ξ0 )2 ] E[A(Ξ0 )2 ] E[A(Ξ0 )U(Ξ0 )] E[HΞ0 (θ)2 ] E[U(Ξ0 )2 ] E[A(Ξ0 )HΞ0 (θ)] Parallelogram E[A(Ξ0 )2 ] E[HΞ0 (θ)2 ] E[A(Ξ0 )HΞ0 (θ)] E[HΞ0 (θ1 )HΞ0 (θ2 )] additional quantity of interest SR (ENSM-SE / LGF-PMDM) GSI 2015 28/10/2015 15 / 22 Practical aspects The simplest cases Estimation of 1st and 2nd-order moments E[A(Ξ0 )] and E[W (Ξ0 , K )] E[A(Ξ0 )2 ], E[A(Ξ0 )W (Ξ0 , K )] and E[W (Ξ0 , K )2 ] Disk Segment E[A(Ξ0 )2 ] E[A(Ξ0 )2 ] E[A(Ξ0 )U(Ξ0 )] E[HΞ0 (θ)2 ] E[U(Ξ0 )2 ] E[A(Ξ0 )HΞ0 (θ)] Parallelogram E[A(Ξ0 )2 ] E[HΞ0 (θ)2 ] E[A(Ξ0 )HΞ0 (θ)] E[HΞ0 (θ1 )HΞ0 (θ2 )] additional quantity of interest SR (ENSM-SE / LGF-PMDM) GSI 2015 28/10/2015 15 / 22 Practical aspects Procedure Realizations Ξ(ω) ∩ M r1 , · · · rn K Dilations (Ξ(ω) ⊕ ri K ) ∩ (M ri K ) Covariances CΞ⊕ri K Mean Covariograms γ̄Ξ0 ⊕ri K Integration Polynomial fitting SR (ENSM-SE / LGF-PMDM) 2 E[A(Ξ0 ⊕ ri K ) ] = Z γ̄Ξ0 ⊕ri K (u)du E[A(Ξ0 )2 ], E[A(Ξ0 )W (Ξ0 , K )], E[W (Ξ0 , K )2 ] GSI 2015 28/10/2015 16 / 22 Practical aspects statistical aspects The following estimator for n point probability function is unbiased and strong consistent as M → ∞. (n) ĈΞ,M (x1 , · · · xn ) = A((Ξ ∩ M) {0, x1 − xn , · · · xn−1 − xn }) A(M {0, x1 − xn , · · · xn−1 − xn }) Then it follows a consistent estimator for n-variogram and thus the moments of (A(Ξ0 ), W (Ξ0 , K )), but not necessarily unbiased. ⇒ Small bias for M bigger than Ξ0 .(check by simulation) SR (ENSM-SE / LGF-PMDM) GSI 2015 28/10/2015 17 / 22 Test by simulation Section 4 Test by simulation SR (ENSM-SE / LGF-PMDM) GSI 2015 28/10/2015 18 / 22 Test by simulation Experiments Several realizations of the following Boolean model a ∼ N(40, 10), b ∼ N(30, 10) M : 500 × 500 100 λ= 500 × 500 r = 0, 1, · · · 10 SR (ENSM-SE / LGF-PMDM) GSI 2015 28/10/2015 19 / 22 Test by simulation Results Dilation with a segment Dilation with a disk 30 E[Hθ (Ξ0 )2 ] E[A(Ξ0 )Hθ (Ξ0 )] E[A(Ξ0 )2 ] 20 10 relative error (%) relative error (%) 30 E[U(Ξ0 )2 ] E[A(Ξ0 )U(Ξ0 )] E[A(Ξ0 )2 ] 20 10 5 5 2 2 0 500 0 1,000 1,500 2,000 1,000 number of realizations number of realizations SR (ENSM-SE / LGF-PMDM) 500 GSI 2015 28/10/2015 20 / 22 Test by simulation Conclusions and prospects Conclusions Theoretical estimator for nth order moments of the process HΞ0 Practical estimation of 1st and 2nd-order moments of: t (A(Ξ ), U(Ξ )) and t (A(Ξ ), H (θ)) 0 0 0 Ξ0 ⇒ Characterization of a random particle depending on 2 parameters: rectangle, ellipse... Prospects Describing complex random convex by first and second order characteristics of the process HΞ0 (Ex:Gaussian process). quantifying the anisotropy of the grain. Bias corrector. SR (ENSM-SE / LGF-PMDM) GSI 2015 28/10/2015 21 / 22 Test by simulation Thanks for listening! SR (ENSM-SE / LGF-PMDM) GSI 2015 28/10/2015 22 / 22