Simulating with uncertainty : the rough surface scattering problem

Transcription

1 Simulating with uncertainty : the rough surface scattering problem Uday Khankhoje Assistant Professor, Electrical Engineering Indian Institute of Technology Madras Uday Khankhoje (EE, IITM) Simulating with uncertainty 0 / 36

2 A problem often encountered in numerical analysis 1 A simulation, σ(x), x D is computationally expensive 2 The input x or the domain D displays uncertainty, e.g. in simulating the modes of an optical fibre, the refractive index might not be exactly known, or the boundary may be rough In this light, more useful than σ(x) is the expectation σ(x) Std. dev. in σ(x) due to parameter uncertainty also interesting Uday Khankhoje (EE, IITM) Simulating with uncertainty 1 / 36

3 A problem often encountered in numerical analysis 1 A simulation, σ(x), x D is computationally expensive 2 The input x or the domain D displays uncertainty, e.g. in simulating the modes of an optical fibre, the refractive index might not be exactly known, or the boundary may be rough In this light, more useful than σ(x) is the expectation σ(x) Std. dev. in σ(x) due to parameter uncertainty also interesting Uday Khankhoje (EE, IITM) Simulating with uncertainty 1 / 36

4 A problem often encountered in numerical analysis 1 A simulation, σ(x), x D is computationally expensive 2 The input x or the domain D displays uncertainty, e.g. in simulating the modes of an optical fibre, the refractive index might not be exactly known, or the boundary may be rough In this light, more useful than σ(x) is the expectation σ(x) Std. dev. in σ(x) due to parameter uncertainty also interesting Uday Khankhoje (EE, IITM) Simulating with uncertainty 1 / 36

5 Computing the expectation and standard deviation Assumptions Setup: d-dimensional x = [x 1, x 2,..., x d ] Each x i is mutually independent, distributed with known pdf ρ i Construct a multi-variate pdf, ρ(x) = n i=1 ρ i(x i ) Quantities of interest Expectation σ(x) = σ(x)ρ(x)dx Std. dev. σ(x) = D (σ(x) σ(x) ) 2 ρ(x)dx D Uday Khankhoje (EE, IITM) Simulating with uncertainty 2 / 36

6 Computing the expectation and standard deviation Assumptions Setup: d-dimensional x = [x 1, x 2,..., x d ] Each x i is mutually independent, distributed with known pdf ρ i Construct a multi-variate pdf, ρ(x) = n i=1 ρ i(x i ) Quantities of interest Expectation σ(x) = σ(x)ρ(x)dx Std. dev. σ(x) = D (σ(x) σ(x) ) 2 ρ(x)dx D Uday Khankhoje (EE, IITM) Simulating with uncertainty 2 / 36

7 Any problem with σ(x) = D σ(x)ρ(x)dx? Yes! σ(x) is only known numerically. Recall: σ(x) is the output of an (possibly expensive) simulation Monte Carlo method The most commonly used method to estimate expectation. Instantiate several x i s and approximate σ(x) = n i=1 1 n σ(x i) Convergence rate: independent of d, but slow O( 1 n ) Extra: Create a histogram from σ(x i ) values to estimate pdf of σ. Objective of this talk: Discuss alternatives to Monte Carlo Uday Khankhoje (EE, IITM) Simulating with uncertainty 3 / 36

8 Any problem with σ(x) = D σ(x)ρ(x)dx? Yes! σ(x) is only known numerically. Recall: σ(x) is the output of an (possibly expensive) simulation Monte Carlo method The most commonly used method to estimate expectation. Instantiate several x i s and approximate σ(x) = n i=1 1 n σ(x i) Convergence rate: independent of d, but slow O( 1 n ) Extra: Create a histogram from σ(x i ) values to estimate pdf of σ. Objective of this talk: Discuss alternatives to Monte Carlo Uday Khankhoje (EE, IITM) Simulating with uncertainty 3 / 36

9 Any problem with σ(x) = D σ(x)ρ(x)dx? Yes! σ(x) is only known numerically. Recall: σ(x) is the output of an (possibly expensive) simulation Monte Carlo method The most commonly used method to estimate expectation. Instantiate several x i s and approximate σ(x) = n i=1 1 n σ(x i) Convergence rate: independent of d, but slow O( 1 n ) Extra: Create a histogram from σ(x i ) values to estimate pdf of σ. Objective of this talk: Discuss alternatives to Monte Carlo Uday Khankhoje (EE, IITM) Simulating with uncertainty 3 / 36

10 The alternatives to Monte Carlo Broadly, two families of methods will be discussed: 1 Galerkin Polynomial Chaos (gpc) Need to code a new solver, works for small range of d values 2 Stochastic Collocation (SC) Uses an existing solver, works for small medium range of d values Attractive, because compatible with commercial software What about Monte Carlo? Use for benchmarking and for large range of d values Uday Khankhoje (EE, IITM) Simulating with uncertainty 4 / 36

11 The alternatives to Monte Carlo Broadly, two families of methods will be discussed: 1 Galerkin Polynomial Chaos (gpc) Need to code a new solver, works for small range of d values 2 Stochastic Collocation (SC) Uses an existing solver, works for small medium range of d values Attractive, because compatible with commercial software What about Monte Carlo? Use for benchmarking and for large range of d values Uday Khankhoje (EE, IITM) Simulating with uncertainty 4 / 36

12 Start with Stochastic Collocation: Polynomial interpolation For simplicity, consider a 1-D case of the simulation, σ(x) Polynomial interpolation and integration Choose n-points to evaluate σ(x) and construct a n 1 degree polynomial (Lagrange interpolation): σ n 1 (x) = n i=1 σ(x i)l i (x) Recall, L i (x) = n j=1,j i (x x j)/(x i x j ) Compute expectation as σ(x) n i=1 σ(x i) L i (x)ρ(x)dx = n i=1 σ(x i) α i α i can be pre-computed, and σ(x) is accurate to order n 1. Can we do better? Uday Khankhoje (EE, IITM) Simulating with uncertainty 5 / 36

13 Start with Stochastic Collocation: Polynomial interpolation For simplicity, consider a 1-D case of the simulation, σ(x) Polynomial interpolation and integration Choose n-points to evaluate σ(x) and construct a n 1 degree polynomial (Lagrange interpolation): σ n 1 (x) = n i=1 σ(x i)l i (x) Recall, L i (x) = n j=1,j i (x x j)/(x i x j ) Compute expectation as σ(x) n i=1 σ(x i) L i (x)ρ(x)dx = n i=1 σ(x i) α i α i can be pre-computed, and σ(x) is accurate to order n 1. Can we do better? Uday Khankhoje (EE, IITM) Simulating with uncertainty 5 / 36

14 Start with Stochastic Collocation: Polynomial interpolation For simplicity, consider a 1-D case of the simulation, σ(x) Polynomial interpolation and integration Choose n-points to evaluate σ(x) and construct a n 1 degree polynomial (Lagrange interpolation): σ n 1 (x) = n i=1 σ(x i)l i (x) Recall, L i (x) = n j=1,j i (x x j)/(x i x j ) Compute expectation as σ(x) n i=1 σ(x i) L i (x)ρ(x)dx = n i=1 σ(x i) α i α i can be pre-computed, and σ(x) is accurate to order n 1. Can we do better? Uday Khankhoje (EE, IITM) Simulating with uncertainty 5 / 36

15 Stochastic Collocation: Gaussian quadrature Yes! Gaussian quadrature Use the theory of orthogonal polynomials, i.e. find polynomials s.t. p m (x)p n (x)ρ(x)dx = δ m,n Pick the n points, {x i } to be roots of n th degree polynomial, p n (x) This also gives σ(x) n i=1 σ(x i) α i, but now integral is accurate to order 2n 1 Common orthogonal polynomials Legendre, x [ 1, 1], ρ(x) = 1. Jacobi, x ( 1, 1), ρ(x) = (1 x) α (1 + x) β, α, β > 1. Hermite, x (, ), ρ(x) = e x2. Uday Khankhoje (EE, IITM) Simulating with uncertainty 6 / 36

16 Stochastic Collocation: Gaussian quadrature Yes! Gaussian quadrature Use the theory of orthogonal polynomials, i.e. find polynomials s.t. p m (x)p n (x)ρ(x)dx = δ m,n Pick the n points, {x i } to be roots of n th degree polynomial, p n (x) This also gives σ(x) n i=1 σ(x i) α i, but now integral is accurate to order 2n 1 Common orthogonal polynomials Legendre, x [ 1, 1], ρ(x) = 1. Jacobi, x ( 1, 1), ρ(x) = (1 x) α (1 + x) β, α, β > 1. Hermite, x (, ), ρ(x) = e x2. Uday Khankhoje (EE, IITM) Simulating with uncertainty 6 / 36

17 Stochastic Collocation: Gaussian quadrature Yes! Gaussian quadrature Use the theory of orthogonal polynomials, i.e. find polynomials s.t. p m (x)p n (x)ρ(x)dx = δ m,n Pick the n points, {x i } to be roots of n th degree polynomial, p n (x) This also gives σ(x) n i=1 σ(x i) α i, but now integral is accurate to order 2n 1 Common orthogonal polynomials Legendre, x [ 1, 1], ρ(x) = 1. Jacobi, x ( 1, 1), ρ(x) = (1 x) α (1 + x) β, α, β > 1. Hermite, x (, ), ρ(x) = e x2. Uday Khankhoje (EE, IITM) Simulating with uncertainty 6 / 36

18 But our problem is d-dimensional! Solution: Extend one-dimensional Gaussian quadrature to d dimensions With x = [x i, x 2,..., x d ], express σ(x) = d i=1 σ i(x i ) (implicitly) The integral splits up: σ(x)ρ(x)dx = i σi (x i )ρ i (x i )dx i Apply n-point Gaussian quadrature (GQ) in each dimension: σ(x) = d i=1 n j=1 σ i(x i,j )α j. Combine the σ i s to get σ(x k ). An example in 2D (x, y) and a 2-point GQ in each dimension: σ(x, y) = [σ x (x 1 )α 1 + σ x (x 2 )α 2 ][σ y (y 1 )α 1 + σ y (y 2 )α 2 ] = σ(x 1, y 1 )α 1,1 + σ(x 1, y 2 )α 1,2 + σ(x 2, y 1 )α 2,1 + σ(x 2, y 2 )α 2,2 All dimensions are multiplied: called the tensor product rule Uday Khankhoje (EE, IITM) Simulating with uncertainty 7 / 36

19 But our problem is d-dimensional! Solution: Extend one-dimensional Gaussian quadrature to d dimensions With x = [x i, x 2,..., x d ], express σ(x) = d i=1 σ i(x i ) (implicitly) The integral splits up: σ(x)ρ(x)dx = i σi (x i )ρ i (x i )dx i Apply n-point Gaussian quadrature (GQ) in each dimension: σ(x) = d i=1 n j=1 σ i(x i,j )α j. Combine the σ i s to get σ(x k ). An example in 2D (x, y) and a 2-point GQ in each dimension: σ(x, y) = [σ x (x 1 )α 1 + σ x (x 2 )α 2 ][σ y (y 1 )α 1 + σ y (y 2 )α 2 ] = σ(x 1, y 1 )α 1,1 + σ(x 1, y 2 )α 1,2 + σ(x 2, y 1 )α 2,1 + σ(x 2, y 2 )α 2,2 All dimensions are multiplied: called the tensor product rule Uday Khankhoje (EE, IITM) Simulating with uncertainty 7 / 36

20 But our problem is d-dimensional! Solution: Extend one-dimensional Gaussian quadrature to d dimensions With x = [x i, x 2,..., x d ], express σ(x) = d i=1 σ i(x i ) (implicitly) The integral splits up: σ(x)ρ(x)dx = i σi (x i )ρ i (x i )dx i Apply n-point Gaussian quadrature (GQ) in each dimension: σ(x) = d i=1 n j=1 σ i(x i,j )α j. Combine the σ i s to get σ(x k ). An example in 2D (x, y) and a 2-point GQ in each dimension: σ(x, y) = [σ x (x 1 )α 1 + σ x (x 2 )α 2 ][σ y (y 1 )α 1 + σ y (y 2 )α 2 ] = σ(x 1, y 1 )α 1,1 + σ(x 1, y 2 )α 1,2 + σ(x 2, y 1 )α 2,1 + σ(x 2, y 2 )α 2,2 All dimensions are multiplied: called the tensor product rule Uday Khankhoje (EE, IITM) Simulating with uncertainty 7 / 36

21 Visualizing the tensor product rule E.g. function evaluation points in a 2D 5-point GQ (25 evals) : Denote as σ 5,5 Curse of dimensionality is clear: number of function evaluations = n d Can we do better? y x Theorem: (Mysovskikh 1968, Möller 1976) To attain a polynomial exactness equal to m, the (optimal) required number of grid-points has lower and upper bounds given by ( ) ( ) d + m /2 d + m N min = N opt = N max m /2 m Uday Khankhoje (EE, IITM) Simulating with uncertainty 8 / 36

22 Visualizing the tensor product rule E.g. function evaluation points in a 2D 5-point GQ (25 evals) : Denote as σ 5,5 Curse of dimensionality is clear: number of function evaluations = n d Can we do better? y x Theorem: (Mysovskikh 1968, Möller 1976) To attain a polynomial exactness equal to m, the (optimal) required number of grid-points has lower and upper bounds given by ( ) ( ) d + m /2 d + m N min = N opt = N max m /2 m Uday Khankhoje (EE, IITM) Simulating with uncertainty 8 / 36

23 Visualizing the tensor product rule E.g. function evaluation points in a 2D 5-point GQ (25 evals) : Denote as σ 5,5 Curse of dimensionality is clear: number of function evaluations = n d Can we do better? y x Theorem: (Mysovskikh 1968, Möller 1976) To attain a polynomial exactness equal to m, the (optimal) required number of grid-points has lower and upper bounds given by ( ) ( ) d + m /2 d + m N min = N opt = N max m /2 m Uday Khankhoje (EE, IITM) Simulating with uncertainty 8 / 36

24 Sparse Grids: Smolyak (1963) 1 Every σ i,j is an approximation to the actual integral, σ 2 Introduce a level parameter k = (i + j) and let the max level be denoted by l = max{k} Telescope a series of different levels to approximate σ e.g. (Max) level l = 4: σ [ σ 3,1 + σ 1,3 + σ 2,2 ] k=4 [ σ 2,1 + σ 1,2 ] k=3 e.g. (Max) level l = 5: σ [ σ 3,2 + σ 2,3 + σ 1,4 + σ 4,1 ] [ σ 3,1 + σ 1,3 + σ 2,2 ] As the max level increases, approximation becomes better Uday Khankhoje (EE, IITM) Simulating with uncertainty 9 / 36

25 Sparse Grids: Smolyak (1963) 1 Every σ i,j is an approximation to the actual integral, σ 2 Introduce a level parameter k = (i + j) and let the max level be denoted by l = max{k} Telescope a series of different levels to approximate σ e.g. (Max) level l = 4: σ [ σ 3,1 + σ 1,3 + σ 2,2 ] k=4 [ σ 2,1 + σ 1,2 ] k=3 e.g. (Max) level l = 5: σ [ σ 3,2 + σ 2,3 + σ 1,4 + σ 4,1 ] [ σ 3,1 + σ 1,3 + σ 2,2 ] As the max level increases, approximation becomes better Uday Khankhoje (EE, IITM) Simulating with uncertainty 9 / 36

26 Sparse Grids: Smolyak (1963) 1 Every σ i,j is an approximation to the actual integral, σ 2 Introduce a level parameter k = (i + j) and let the max level be denoted by l = max{k} Telescope a series of different levels to approximate σ e.g. (Max) level l = 4: σ [ σ 3,1 + σ 1,3 + σ 2,2 ] k=4 [ σ 2,1 + σ 1,2 ] k=3 e.g. (Max) level l = 5: σ [ σ 3,2 + σ 2,3 + σ 1,4 + σ 4,1 ] [ σ 3,1 + σ 1,3 + σ 2,2 ] As the max level increases, approximation becomes better Uday Khankhoje (EE, IITM) Simulating with uncertainty 9 / 36

27 Sparse Grids: Smolyak (1963) 1 Every σ i,j is an approximation to the actual integral, σ 2 Introduce a level parameter k = (i + j) and let the max level be denoted by l = max{k} Telescope a series of different levels to approximate σ e.g. (Max) level l = 4: σ [ σ 3,1 + σ 1,3 + σ 2,2 ] k=4 [ σ 2,1 + σ 1,2 ] k=3 e.g. (Max) level l = 5: σ [ σ 3,2 + σ 2,3 + σ 1,4 + σ 4,1 ] [ σ 3,1 + σ 1,3 + σ 2,2 ] As the max level increases, approximation becomes better Uday Khankhoje (EE, IITM) Simulating with uncertainty 9 / 36

28 Visualizing the sparse grid rule σ [ σ 3,1 + σ 1,3 + σ 2,2 ] k=4 [ σ 2,1 + σ 1,2 ] k=3 Sparse Grid Points Tensor Grid Points = Substantial savings in higher dims 13 Points (SG) v/s 25 points (TP)! Uday Khankhoje (EE, IITM) Simulating with uncertainty 10 / 36

29 Visualizing the sparse grid rule σ [ σ 3,1 + σ 1,3 + σ 2,2 ] k=4 [ σ 2,1 + σ 1,2 ] k=3 Sparse Grid Points Tensor Grid Points = Substantial savings in higher dims 13 Points (SG) v/s 25 points (TP)! Uday Khankhoje (EE, IITM) Simulating with uncertainty 10 / 36

30 Visualizing the sparse grid rule σ [ σ 3,1 + σ 1,3 + σ 2,2 ] k=4 [ σ 2,1 + σ 1,2 ] k=3 Sparse Grid Points Tensor Grid Points = Substantial savings in higher dims 13 Points (SG) v/s 25 points (TP)! Uday Khankhoje (EE, IITM) Simulating with uncertainty 10 / 36

31 Sparse Grid rule: finer points TP rule: σ = n 1 i 1 =1... n d i d =1 σ(x i 1,..., x id )α i1... α id = (Q n 1 Q n 2... Q n d)[σ] n i point quadrature in the i th dim; n i n d points SG rule: With max level l, and k = k 1 + k k d : σ = ( 1) l k( d 1 l k) (Q k 1 Q k 2... Q k d)[σ] l d+1 k l k i point quadrature in the i th dim; 2 l d l /l! points Nested quadrature rules (e.g. Clenshaw-Curtis, Gauss-Konrad) re-use fn evaluation points between levels. Uday Khankhoje (EE, IITM) Simulating with uncertainty 11 / 36

32 Sparse Grid rule: finer points TP rule: σ = n 1 i 1 =1... n d i d =1 σ(x i 1,..., x id )α i1... α id = (Q n 1 Q n 2... Q n d)[σ] n i point quadrature in the i th dim; n i n d points SG rule: With max level l, and k = k 1 + k k d : σ = ( 1) l k( d 1 l k) (Q k 1 Q k 2... Q k d)[σ] l d+1 k l k i point quadrature in the i th dim; 2 l d l /l! points Nested quadrature rules (e.g. Clenshaw-Curtis, Gauss-Konrad) re-use fn evaluation points between levels. Uday Khankhoje (EE, IITM) Simulating with uncertainty 11 / 36

33 Sparse Grid rule: finer points TP rule: σ = n 1 i 1 =1... n d i d =1 σ(x i 1,..., x id )α i1... α id = (Q n 1 Q n 2... Q n d)[σ] n i point quadrature in the i th dim; n i n d points SG rule: With max level l, and k = k 1 + k k d : σ = ( 1) l k( d 1 l k) (Q k 1 Q k 2... Q k d)[σ] l d+1 k l k i point quadrature in the i th dim; 2 l d l /l! points Nested quadrature rules (e.g. Clenshaw-Curtis, Gauss-Konrad) re-use fn evaluation points between levels. Uday Khankhoje (EE, IITM) Simulating with uncertainty 11 / 36

34 Switching gears: from sampling to projection The two methods (Monte Carlo, Stochastic Collocation) considered so far were of a sampling kind: σ estimated using samples of σ(x). Projection based approach: Overview of Galerkin method Governing equation: Θf(x) = g(x), where Θ is an operator, g is a known function, and f is to be determined. Project f in a known basis, {φ j }: f(x) = j u jφ j (x) Take an inner product with the same basis functions on both sides to get: j φ i(x), Θφ j (x) u j = φ i (x), g(x). BC used to simplify. This is a system of equations; solve for u and get f. Straightforward when x is spatio-temporal. But when stochastic? Uday Khankhoje (EE, IITM) Simulating with uncertainty 12 / 36

35 Switching gears: from sampling to projection The two methods (Monte Carlo, Stochastic Collocation) considered so far were of a sampling kind: σ estimated using samples of σ(x). Projection based approach: Overview of Galerkin method Governing equation: Θf(x) = g(x), where Θ is an operator, g is a known function, and f is to be determined. Project f in a known basis, {φ j }: f(x) = j u jφ j (x) Take an inner product with the same basis functions on both sides to get: j φ i(x), Θφ j (x) u j = φ i (x), g(x). BC used to simplify. This is a system of equations; solve for u and get f. Straightforward when x is spatio-temporal. But when stochastic? Uday Khankhoje (EE, IITM) Simulating with uncertainty 12 / 36

36 Switching gears: from sampling to projection The two methods (Monte Carlo, Stochastic Collocation) considered so far were of a sampling kind: σ estimated using samples of σ(x). Projection based approach: Overview of Galerkin method Governing equation: Θf(x) = g(x), where Θ is an operator, g is a known function, and f is to be determined. Project f in a known basis, {φ j }: f(x) = j u jφ j (x) Take an inner product with the same basis functions on both sides to get: j φ i(x), Θφ j (x) u j = φ i (x), g(x). BC used to simplify. This is a system of equations; solve for u and get f. Straightforward when x is spatio-temporal. But when stochastic? Uday Khankhoje (EE, IITM) Simulating with uncertainty 12 / 36

37 Switching gears: from sampling to projection The two methods (Monte Carlo, Stochastic Collocation) considered so far were of a sampling kind: σ estimated using samples of σ(x). Projection based approach: Overview of Galerkin method Governing equation: Θf(x) = g(x), where Θ is an operator, g is a known function, and f is to be determined. Project f in a known basis, {φ j }: f(x) = j u jφ j (x) Take an inner product with the same basis functions on both sides to get: j φ i(x), Θφ j (x) u j = φ i (x), g(x). BC used to simplify. This is a system of equations; solve for u and get f. Straightforward when x is spatio-temporal. But when stochastic? Uday Khankhoje (EE, IITM) Simulating with uncertainty 12 / 36

38 generalized Polynomial Chaos (gpc): a very brief history Polynomial Chaos (PC): coined by Norbert Wiener studying Gaussian stochastic processes (1938). Used Hermite polynomials as basis. Ghanem (1998) used theory of Wiener-Hermite PC to represent random processes in an orthogonal basis of Hermite polynomials. Xiu, Karniadakis (2002) generalize to non-gaussian using other orthogonal polynomials, wavelets, etc: generalized PC (gpc) Finally, solve the gpc system of equations using galerkin projection: stochastic galerkin (SG) method. Uday Khankhoje (EE, IITM) Simulating with uncertainty 13 / 36

39 generalized Polynomial Chaos (gpc): a very brief history Polynomial Chaos (PC): coined by Norbert Wiener studying Gaussian stochastic processes (1938). Used Hermite polynomials as basis. Ghanem (1998) used theory of Wiener-Hermite PC to represent random processes in an orthogonal basis of Hermite polynomials. Xiu, Karniadakis (2002) generalize to non-gaussian using other orthogonal polynomials, wavelets, etc: generalized PC (gpc) Finally, solve the gpc system of equations using galerkin projection: stochastic galerkin (SG) method. Uday Khankhoje (EE, IITM) Simulating with uncertainty 13 / 36

40 generalized Polynomial Chaos (gpc): a very brief history Polynomial Chaos (PC): coined by Norbert Wiener studying Gaussian stochastic processes (1938). Used Hermite polynomials as basis. Ghanem (1998) used theory of Wiener-Hermite PC to represent random processes in an orthogonal basis of Hermite polynomials. Xiu, Karniadakis (2002) generalize to non-gaussian using other orthogonal polynomials, wavelets, etc: generalized PC (gpc) Finally, solve the gpc system of equations using galerkin projection: stochastic galerkin (SG) method. Uday Khankhoje (EE, IITM) Simulating with uncertainty 13 / 36

41 The basis functions in gpc Start with a RV, call it z as before. Let it have a distribution function, F z (θ) = P (z θ) and a pdf ρ(θ) s.t. df z (θ) = ρ(θ)dθ The generalized Polynomial Chaos basis functions are orthogonal basis functions, ψ i (z), satisfying : ψ i (z)ψ j (z) = ψ i (θ)ψ j (θ)ρ(θ)dθ = γ i δ ij Construct linear space of polynomials of degree at most n: P n (z) Various kinds depending on ρ(θ) Legendre, θ [ 1, 1], ρ(θ) = 1/2. Jacobi, θ ( 1, 1), ρ(θ) = (1 θ) α (1 + θ) β, α, β > 1. Hermite, θ (, ), ρ(θ) = e θ2. Uday Khankhoje (EE, IITM) Simulating with uncertainty 14 / 36

42 The basis functions in gpc Start with a RV, call it z as before. Let it have a distribution function, F z (θ) = P (z θ) and a pdf ρ(θ) s.t. df z (θ) = ρ(θ)dθ The generalized Polynomial Chaos basis functions are orthogonal basis functions, ψ i (z), satisfying : ψ i (z)ψ j (z) = ψ i (θ)ψ j (θ)ρ(θ)dθ = γ i δ ij Construct linear space of polynomials of degree at most n: P n (z) Various kinds depending on ρ(θ) Legendre, θ [ 1, 1], ρ(θ) = 1/2. Jacobi, θ ( 1, 1), ρ(θ) = (1 θ) α (1 + θ) β, α, β > 1. Hermite, θ (, ), ρ(θ) = e θ2. Uday Khankhoje (EE, IITM) Simulating with uncertainty 14 / 36

43 The basis functions in gpc Start with a RV, call it z as before. Let it have a distribution function, F z (θ) = P (z θ) and a pdf ρ(θ) s.t. df z (θ) = ρ(θ)dθ The generalized Polynomial Chaos basis functions are orthogonal basis functions, ψ i (z), satisfying : ψ i (z)ψ j (z) = ψ i (θ)ψ j (θ)ρ(θ)dθ = γ i δ ij Construct linear space of polynomials of degree at most n: P n (z) Various kinds depending on ρ(θ) Legendre, θ [ 1, 1], ρ(θ) = 1/2. Jacobi, θ ( 1, 1), ρ(θ) = (1 θ) α (1 + θ) β, α, β > 1. Hermite, θ (, ), ρ(θ) = e θ2. Uday Khankhoje (EE, IITM) Simulating with uncertainty 14 / 36

44 A scalar example of gpc-sg Consider: a simple equation in one unknown, u: a u = b Now: Let the system parameters a, b have some uncertainty, e.g. a(θ) = a 0 + αθ, where θ is a uniform RV in [ 0.5, 0.5]. In this case what is u or u 2? Expand: solution in the ψ basis u(θ) = n j=1 u jψ j (θ) Do: Galerkin testing with the same basis functions: Get a system of equations, where we solve for u: Au = b, where A ij = ψ i (θ)a 0 ψ j (θ) + ψ i (θ)αθψ j (θ) = a 0 γ i δ ij + α ψ i (θ)θψ j (θ) and b j = ψ i (θ)b(θ) So: u = u(θ)ρ(θ)dθ = i u i ψ i (θ) and u 2 = i u2 i γ i. = std. dev. in u can be computed : u 2 u 2 = uncertainty quantified! Uday Khankhoje (EE, IITM) Simulating with uncertainty 15 / 36

45 A scalar example of gpc-sg Consider: a simple equation in one unknown, u: a u = b Now: Let the system parameters a, b have some uncertainty, e.g. a(θ) = a 0 + αθ, where θ is a uniform RV in [ 0.5, 0.5]. In this case what is u or u 2? Expand: solution in the ψ basis u(θ) = n j=1 u jψ j (θ) Do: Galerkin testing with the same basis functions: Get a system of equations, where we solve for u: Au = b, where A ij = ψ i (θ)a 0 ψ j (θ) + ψ i (θ)αθψ j (θ) = a 0 γ i δ ij + α ψ i (θ)θψ j (θ) and b j = ψ i (θ)b(θ) So: u = u(θ)ρ(θ)dθ = i u i ψ i (θ) and u 2 = i u2 i γ i. = std. dev. in u can be computed : u 2 u 2 = uncertainty quantified! Uday Khankhoje (EE, IITM) Simulating with uncertainty 15 / 36

46 A scalar example of gpc-sg Consider: a simple equation in one unknown, u: a u = b Now: Let the system parameters a, b have some uncertainty, e.g. a(θ) = a 0 + αθ, where θ is a uniform RV in [ 0.5, 0.5]. In this case what is u or u 2? Expand: solution in the ψ basis u(θ) = n j=1 u jψ j (θ) Do: Galerkin testing with the same basis functions: Get a system of equations, where we solve for u: Au = b, where A ij = ψ i (θ)a 0 ψ j (θ) + ψ i (θ)αθψ j (θ) = a 0 γ i δ ij + α ψ i (θ)θψ j (θ) and b j = ψ i (θ)b(θ) So: u = u(θ)ρ(θ)dθ = i u i ψ i (θ) and u 2 = i u2 i γ i. = std. dev. in u can be computed : u 2 u 2 = uncertainty quantified! Uday Khankhoje (EE, IITM) Simulating with uncertainty 15 / 36

47 Random inputs? (More than just a collection of RVs!) Example: A random surface adjacent points are not independent of each other, there is some correlation: y x The Kosambi-Karhunen-Loeve (KL) expansion is widely used to represent random processes: s(x, θ) = s 0 (x) + k=1 ηk f k (x)z k (θ) s 0 (x) is the mean of the random process η, f solve this eigenvalue problem: C(i, j)f k (j)dj = η k f k (i) where C(i, j) = cov(z i, z j ) is the correlation between two RVs, z i, z j z(θ) represents mutually uncorrelated normal RVs ( z k = 0) Expansion truncated to d terms in practice Uday Khankhoje (EE, IITM) Simulating with uncertainty 16 / 36

48 Random inputs? (More than just a collection of RVs!) Example: A random surface adjacent points are not independent of each other, there is some correlation: y x The Kosambi-Karhunen-Loeve (KL) expansion is widely used to represent random processes: s(x, θ) = s 0 (x) + k=1 ηk f k (x)z k (θ) s 0 (x) is the mean of the random process η, f solve this eigenvalue problem: C(i, j)f k (j)dj = η k f k (i) where C(i, j) = cov(z i, z j ) is the correlation between two RVs, z i, z j z(θ) represents mutually uncorrelated normal RVs ( z k = 0) Expansion truncated to d terms in practice Uday Khankhoje (EE, IITM) Simulating with uncertainty 16 / 36

49 Random inputs? (More than just a collection of RVs!) Example: A random surface adjacent points are not independent of each other, there is some correlation: y x The Kosambi-Karhunen-Loeve (KL) expansion is widely used to represent random processes: s(x, θ) = s 0 (x) + k=1 ηk f k (x)z k (θ) s 0 (x) is the mean of the random process η, f solve this eigenvalue problem: C(i, j)f k (j)dj = η k f k (i) where C(i, j) = cov(z i, z j ) is the correlation between two RVs, z i, z j z(θ) represents mutually uncorrelated normal RVs ( z k = 0) Expansion truncated to d terms in practice Uday Khankhoje (EE, IITM) Simulating with uncertainty 16 / 36

50 Random inputs? (More than just a collection of RVs!) Example: A random surface adjacent points are not independent of each other, there is some correlation: y x The Kosambi-Karhunen-Loeve (KL) expansion is widely used to represent random processes: s(x, θ) = s 0 (x) + k=1 ηk f k (x)z k (θ) s 0 (x) is the mean of the random process η, f solve this eigenvalue problem: C(i, j)f k (j)dj = η k f k (i) where C(i, j) = cov(z i, z j ) is the correlation between two RVs, z i, z j z(θ) represents mutually uncorrelated normal RVs ( z k = 0) Expansion truncated to d terms in practice Uday Khankhoje (EE, IITM) Simulating with uncertainty 16 / 36

51 KL expansion for exponential correlation s(x, θ) = s 0 (x) + k=1 ηk f k (x)z k (θ) KL expansion eigenvalues and functions can be analytically calculated in some cases, e.g. exponential correlation function C(i, j) = exp( i j /l), (correlation length l). Decay rate of eigenvalues depends inversely on the correlation length = more RVs for smaller l. Uday Khankhoje (EE, IITM) Simulating with uncertainty 17 / 36

52 KL expansion for exponential correlation s(x, θ) = s 0 (x) + k=1 ηk f k (x)z k (θ) KL expansion eigenvalues and functions can be analytically calculated in some cases, e.g. exponential correlation function C(i, j) = exp( i j /l), (correlation length l). Decay rate of eigenvalues depends inversely on the correlation length = more RVs for smaller l. Uday Khankhoje (EE, IITM) Simulating with uncertainty 17 / 36

53 KL expansion for exponential correlation s(x, θ) = s 0 (x) + k=1 ηk f k (x)z k (θ) KL expansion eigenvalues and functions can be analytically calculated in some cases, e.g. exponential correlation function C(i, j) = exp( i j /l), (correlation length l). Decay rate of eigenvalues depends inversely on the correlation length = more RVs for smaller l. Eigenvalues l = 2 l = 1 l = Eigenvalue number Uday Khankhoje (EE, IITM) Simulating with uncertainty 17 / 36

54 The case of multiple random variables with correlation Consider again the random rough surface... y x given by the KL expansion: s(x, θ) = s 0 (x) + d k=1 ηk f k (x)z k (θ) consists of multiple random variables, z i. Multivariate gpc (i.e. tensor product of univariate gpc) Ψ i ( z) = ψ i1 (z 1 )... ψ id (z d ), 0 d j=1 i j n (max degree) where i = (i 1,..., d): indices, z = (z 1,..., z d ) RVs Belong to the space P d n of dimension w = ( ) n+d n As in univariate case, proceed by Galerkin method and compute mean, variance, etc Uday Khankhoje (EE, IITM) Simulating with uncertainty 18 / 36

55 The case of multiple random variables with correlation Consider again the random rough surface... y x given by the KL expansion: s(x, θ) = s 0 (x) + d k=1 ηk f k (x)z k (θ) consists of multiple random variables, z i. Multivariate gpc (i.e. tensor product of univariate gpc) Ψ i ( z) = ψ i1 (z 1 )... ψ id (z d ), 0 d j=1 i j n (max degree) where i = (i 1,..., d): indices, z = (z 1,..., z d ) RVs Belong to the space P d n of dimension w = ( ) n+d n As in univariate case, proceed by Galerkin method and compute mean, variance, etc Uday Khankhoje (EE, IITM) Simulating with uncertainty 18 / 36

56 The case of multiple random variables with correlation Consider again the random rough surface... y x given by the KL expansion: s(x, θ) = s 0 (x) + d k=1 ηk f k (x)z k (θ) consists of multiple random variables, z i. Multivariate gpc (i.e. tensor product of univariate gpc) Ψ i ( z) = ψ i1 (z 1 )... ψ id (z d ), 0 d j=1 i j n (max degree) where i = (i 1,..., d): indices, z = (z 1,..., z d ) RVs Belong to the space P d n of dimension w = ( ) n+d n As in univariate case, proceed by Galerkin method and compute mean, variance, etc Uday Khankhoje (EE, IITM) Simulating with uncertainty 18 / 36

57 Changing gears... So far... This completes an overview of stochastic computation: 1 Monte Carlo (MC) 2 Stochastic Collocation (SC) 3 Stochastic Galerkin (SG) using generalized Polynomial Chaos Moving on... To make matters more concrete, consider the problem of computing the electromagnetic scattering from a random rough surface: e.g. seen in microwave remote sensing, [MC] a, [SC,SG] b a Khankhoje et al. Computation of radar scattering from heterogeneous rough soil using the finite element method, 2013 IEEE TGRS b Khankhoje et al. Stochastic Solutions to Rough Surface Scattering using the finite element method, 2017 IEEE TAP Uday Khankhoje (EE, IITM) Simulating with uncertainty 19 / 36

58 Changing gears... So far... This completes an overview of stochastic computation: 1 Monte Carlo (MC) 2 Stochastic Collocation (SC) 3 Stochastic Galerkin (SG) using generalized Polynomial Chaos Moving on... To make matters more concrete, consider the problem of computing the electromagnetic scattering from a random rough surface: e.g. seen in microwave remote sensing, [MC] a, [SC,SG] b a Khankhoje et al. Computation of radar scattering from heterogeneous rough soil using the finite element method, 2013 IEEE TGRS b Khankhoje et al. Stochastic Solutions to Rough Surface Scattering using the finite element method, 2017 IEEE TAP Uday Khankhoje (EE, IITM) Simulating with uncertainty 19 / 36

59 Traditional FEM setup for rough surface scattering Random rough surface instance generated, and the domain meshed Based on incident field, Radar cross-section (RCS) computed Above steps repeated for several ( 100) instances Quite inefficient! Uday Khankhoje (EE, IITM) Simulating with uncertainty 20 / 36

60 Traditional FEM setup for rough surface scattering Random rough surface instance generated, and the domain meshed Based on incident field, Radar cross-section (RCS) computed Above steps repeated for several ( 100) instances Quite inefficient! Uday Khankhoje (EE, IITM) Simulating with uncertainty 20 / 36

61 Traditional FEM setup for rough surface scattering Random rough surface instance generated, and the domain meshed Based on incident field, Radar cross-section (RCS) computed Above steps repeated for several ( 100) instances Quite inefficient! Uday Khankhoje (EE, IITM) Simulating with uncertainty 20 / 36

62 Handle the rough surface intelligently Note that the mesh need only change near the surface... A P C E E i Γ h t h b B Q D F Let s(x) define rough surface (e.g. KL expansion) Move each node smoothly within sandwich region: y y + y { s(x)( h t y h y = t ), 0 < y < h t s(x)( y+h b h b ), h b < y < 0 Partition the domain into parts that can move, and those that need not CD will deform to rough surface Zero deformation by the time y = h t or y = h b Uday Khankhoje (EE, IITM) Simulating with uncertainty 21 / 36

63 Handle the rough surface intelligently Note that the mesh need only change near the surface... A P C E E i Γ h t h b B Q D F Let s(x) define rough surface (e.g. KL expansion) Move each node smoothly within sandwich region: y y + y { s(x)( h t y h y = t ), 0 < y < h t s(x)( y+h b h b ), h b < y < 0 Partition the domain into parts that can move, and those that need not CD will deform to rough surface Zero deformation by the time y = h t or y = h b Uday Khankhoje (EE, IITM) Simulating with uncertainty 21 / 36

64 Handle the rough surface intelligently Note that the mesh need only change near the surface... A P C E E i Γ h t h b B Q D F Let s(x) define rough surface (e.g. KL expansion) Move each node smoothly within sandwich region: y y + y { s(x)( h t y h y = t ), 0 < y < h t s(x)( y+h b h b ), h b < y < 0 Partition the domain into parts that can move, and those that need not CD will deform to rough surface Zero deformation by the time y = h t or y = h b Uday Khankhoje (EE, IITM) Simulating with uncertainty 21 / 36

65 Handle the rough surface intelligently Note that the mesh need only change near the surface... A P C E E i Γ h t h b B Q D F Let s(x) define rough surface (e.g. KL expansion) Move each node smoothly within sandwich region: y y + y { s(x)( h t y h y = t ), 0 < y < h t s(x)( y+h b h b ), h b < y < 0 Partition the domain into parts that can move, and those that need not CD will deform to rough surface Zero deformation by the time y = h t or y = h b Uday Khankhoje (EE, IITM) Simulating with uncertainty 21 / 36

66 The standard FEM recipe: deterministic solver 2D vector FE basis functions W A P C E E i Γ h t h b B Q D F Maxwell s equations in weak form: W [ ( 1 ɛ r H) k 2 0 µ r H] ds = 0 First order absorbing boundary conditions on A B F E A Performing Galerkin testing Au = b, A C m m, u, b C m, A pq = Σ e α e,pq ( r e ) + δ pq ν p ( r p ), b p = τ p ( r p ), where r e = (x i, x j, x k, y i, y j, y k ), e th ele r p = (x i, x j, y i, y j ), p th edge Uday Khankhoje (EE, IITM) Simulating with uncertainty 22 / 36

67 The standard FEM recipe: deterministic solver 2D vector FE basis functions W A P C E E i Γ h t h b B Q D F Maxwell s equations in weak form: W [ ( 1 ɛ r H) k 2 0 µ r H] ds = 0 First order absorbing boundary conditions on A B F E A Performing Galerkin testing Au = b, A C m m, u, b C m, A pq = Σ e α e,pq ( r e ) + δ pq ν p ( r p ), b p = τ p ( r p ), where r e = (x i, x j, x k, y i, y j, y k ), e th ele r p = (x i, x j, y i, y j ), p th edge Uday Khankhoje (EE, IITM) Simulating with uncertainty 22 / 36

68 The standard FEM recipe: deterministic solver 2D vector FE basis functions W A P C E E i Γ h t h b B Q D F Maxwell s equations in weak form: W [ ( 1 ɛ r H) k 2 0 µ r H] ds = 0 First order absorbing boundary conditions on A B F E A Performing Galerkin testing Au = b, A C m m, u, b C m, A pq = Σ e α e,pq ( r e ) + δ pq ν p ( r p ), b p = τ p ( r p ), where r e = (x i, x j, x k, y i, y j, y k ), e th ele r p = (x i, x j, y i, y j ), p th edge Uday Khankhoje (EE, IITM) Simulating with uncertainty 22 / 36

69 The standard FEM recipe: deterministic solver 2πr Ef z r E 2 z i Quantity of interest: radar cross-section σ 2D = lim Ez f k ( r) = 0 e i(k 0 r π/4) 8π r ẑ (ˆr M( r ) + Z 0 µ rˆr ˆr J( r ))e ik 0ˆr r dl 1 Start with a blank mesh: rough surface is flat 2 Deform mesh virtually using KL expansion for rough surface 3 Compute σ 2D for this mesh 4 Reset the mesh and go to (2) till convergence ( 100 times) Uday Khankhoje (EE, IITM) Simulating with uncertainty 23 / 36

70 The standard FEM recipe: deterministic solver 2πr Ef z r E 2 z i Quantity of interest: radar cross-section σ 2D = lim Ez f k ( r) = 0 e i(k 0 r π/4) 8π r ẑ (ˆr M( r ) + Z 0 µ rˆr ˆr J( r ))e ik 0ˆr r dl 1 Start with a blank mesh: rough surface is flat 2 Deform mesh virtually using KL expansion for rough surface 3 Compute σ 2D for this mesh 4 Reset the mesh and go to (2) till convergence ( 100 times) Uday Khankhoje (EE, IITM) Simulating with uncertainty 23 / 36

71 The standard FEM recipe: deterministic solver 2πr Ef z r E 2 z i Quantity of interest: radar cross-section σ 2D = lim Ez f k ( r) = 0 e i(k 0 r π/4) 8π r ẑ (ˆr M( r ) + Z 0 µ rˆr ˆr J( r ))e ik 0ˆr r dl RCS (db) SPM FEM Scattering Angle (θ) 1 Start with a blank mesh: rough surface is flat 2 Deform mesh virtually using KL expansion for rough surface 3 Compute σ 2D for this mesh 4 Reset the mesh and go to (2) till convergence ( 100 times) Uday Khankhoje (EE, IITM) Simulating with uncertainty 23 / 36

72 What is deterministic here? Input to the FEM solver: 1 Incidence angle, ɛ(r) for object 2 Set of d normal RVs, {z k }, to use in the KL expansion: s(x, θ) = s 0 (x) + d k=1 ηk f k (x)z k (θ) The FEM is run for a specified surface, hence deterministic solver Monte Carlo process converges as O(1/ n mc ), independent of d. (Recall: d depends on the correlation length of the surface) i.e. σ = n mc i=1 σ( r, z i)/n mc Can we keep the same solver, but have faster convergence? Possibly: Let s try stochastic collocation Uday Khankhoje (EE, IITM) Simulating with uncertainty 24 / 36

73 What is deterministic here? Input to the FEM solver: 1 Incidence angle, ɛ(r) for object 2 Set of d normal RVs, {z k }, to use in the KL expansion: s(x, θ) = s 0 (x) + d k=1 ηk f k (x)z k (θ) The FEM is run for a specified surface, hence deterministic solver Monte Carlo process converges as O(1/ n mc ), independent of d. (Recall: d depends on the correlation length of the surface) i.e. σ = n mc i=1 σ( r, z i)/n mc Can we keep the same solver, but have faster convergence? Possibly: Let s try stochastic collocation Uday Khankhoje (EE, IITM) Simulating with uncertainty 24 / 36

74 What is deterministic here? Input to the FEM solver: 1 Incidence angle, ɛ(r) for object 2 Set of d normal RVs, {z k }, to use in the KL expansion: s(x, θ) = s 0 (x) + d k=1 ηk f k (x)z k (θ) The FEM is run for a specified surface, hence deterministic solver Monte Carlo process converges as O(1/ n mc ), independent of d. (Recall: d depends on the correlation length of the surface) i.e. σ = n mc i=1 σ( r, z i)/n mc Can we keep the same solver, but have faster convergence? Possibly: Let s try stochastic collocation Uday Khankhoje (EE, IITM) Simulating with uncertainty 24 / 36

75 From Monte Carlo to Stochastic Collocation 1 Construct multivariate pdf of the d random normal variables: ρ( z) = d j=1 ρ j(z j ) over domain D d, D = (, ) 2 Express expected value as σ = D d σ( r, z)ρ( z)d z 3 Consider n sc evals of the above integral at predecided quadrature points, z i. 4 Express σ( r, z) in terms of interpolating multivariate polynomials (e.g. Langrage) {P (i) ( z)} d i=1 giving σ( r, z) = n sc i=1 σ( r, z i)p (i) ( z) 5 Finally, σ = n sc i=1 σ( r, z i)α i where α i = D d ρ( z)p (i) ( z)d z Uday Khankhoje (EE, IITM) Simulating with uncertainty 25 / 36

76 From Monte Carlo to Stochastic Collocation 1 Construct multivariate pdf of the d random normal variables: ρ( z) = d j=1 ρ j(z j ) over domain D d, D = (, ) 2 Express expected value as σ = D d σ( r, z)ρ( z)d z 3 Consider n sc evals of the above integral at predecided quadrature points, z i. 4 Express σ( r, z) in terms of interpolating multivariate polynomials (e.g. Langrage) {P (i) ( z)} d i=1 giving σ( r, z) = n sc i=1 σ( r, z i)p (i) ( z) 5 Finally, σ = n sc i=1 σ( r, z i)α i where α i = D d ρ( z)p (i) ( z)d z Uday Khankhoje (EE, IITM) Simulating with uncertainty 25 / 36

77 From Monte Carlo to Stochastic Collocation 1 Construct multivariate pdf of the d random normal variables: ρ( z) = d j=1 ρ j(z j ) over domain D d, D = (, ) 2 Express expected value as σ = D d σ( r, z)ρ( z)d z 3 Consider n sc evals of the above integral at predecided quadrature points, z i. 4 Express σ( r, z) in terms of interpolating multivariate polynomials (e.g. Langrage) {P (i) ( z)} d i=1 giving σ( r, z) = n sc i=1 σ( r, z i)p (i) ( z) 5 Finally, σ = n sc i=1 σ( r, z i)α i where α i = D d ρ( z)p (i) ( z)d z Uday Khankhoje (EE, IITM) Simulating with uncertainty 25 / 36

78 Stochastic Collocation for rough surface scattering SC recipe: σ = n sc i=1 σ( r, z i)α i This just looks like the old MC formula! σ = n mc i=1 σ( r, z i)/n mc So, we can use the same solver for σ( r, z i ) Results How many function evals using sparse grid? 2 l d l /l Compare MC and SC for d = 50, l = 1 Evals: MC=100, SC=101 Uday Khankhoje (EE, IITM) Simulating with uncertainty 26 / 36

79 RCS (db) Stochastic Collocation for rough surface scattering SC recipe: σ = n sc i=1 σ( r, z i)α i This just looks like the old MC formula! σ = n mc i=1 σ( r, z i)/n mc So, we can use the same solver for σ( r, z i ) Sparse Grid Stroud 3 Monte Carlo Scattering Angle (3) Results How many function evals using sparse grid? 2 l d l /l Compare MC and SC for d = 50, l = 1 Evals: MC=100, SC=101 Uday Khankhoje (EE, IITM) Simulating with uncertainty 26 / 36

80 The devil is in the details For the same surface as before, what happens if you reduce d? Compare MC and SC for d = 20, l = 1 Evals: MC=100, SC=41 Uday Khankhoje (EE, IITM) Simulating with uncertainty 27 / 36

81 RCS (db) The devil is in the details For the same surface as before, what happens if you reduce d? Sparse Grid Stroud 3 Monte Carlo Scattering Angle (3) Compare MC and SC for d = 20, l = 1 Evals: MC=100, SC=41 Uday Khankhoje (EE, IITM) Simulating with uncertainty 27 / 36

82 Conclusions about Stochastic Collocation Given surface needs to be accurately represented by the finite KL expansion, i.e. there is an optimal d. For MC, found that 100 iterations give convergence to within 1dB. The cheapest SC would have level l = 1, giving n sc 2d = if the surface requires d > 50, MC is better This critical number may decrease if we employ anisotropic sparse grids. Why? KL eigen values decay and not all dims are as important. Uday Khankhoje (EE, IITM) Simulating with uncertainty 28 / 36

83 Conclusions about Stochastic Collocation Given surface needs to be accurately represented by the finite KL expansion, i.e. there is an optimal d. For MC, found that 100 iterations give convergence to within 1dB. The cheapest SC would have level l = 1, giving n sc 2d = if the surface requires d > 50, MC is better This critical number may decrease if we employ anisotropic sparse grids. Why? KL eigen values decay and not all dims are as important. Uday Khankhoje (EE, IITM) Simulating with uncertainty 28 / 36

84 Conclusions about Stochastic Collocation Given surface needs to be accurately represented by the finite KL expansion, i.e. there is an optimal d. For MC, found that 100 iterations give convergence to within 1dB. The cheapest SC would have level l = 1, giving n sc 2d = if the surface requires d > 50, MC is better This critical number may decrease if we employ anisotropic sparse grids. Why? KL eigen values decay and not all dims are as important. Uday Khankhoje (EE, IITM) Simulating with uncertainty 28 / 36

85 Moving over to Stochastic Galerkin Recall the earlier (scalar) example of a u = b, where we replaced a a 0 + αθ, where θ was a uniform RV. In the case of the FEM, we have a sparse matrix equation Au = b to solve. Now, A pq must be transformed as per the KL expansion for the surface. Recall: A pq = Σ e α e,pq ( r e ) + δ pq ν p ( r p ), b p = τ p ( r p ), where r e = (x i, x j, x k, y i, y j, y k ), e th ele r p = (x i, x j, y i, y j ), p th edge Use our mesh deformation scheme, which transformed y y + y. This leads to A pq Ãpq, b p b p Uday Khankhoje (EE, IITM) Simulating with uncertainty 29 / 36

86 Moving over to Stochastic Galerkin Recall the earlier (scalar) example of a u = b, where we replaced a a 0 + αθ, where θ was a uniform RV. In the case of the FEM, we have a sparse matrix equation Au = b to solve. Now, A pq must be transformed as per the KL expansion for the surface. Recall: A pq = Σ e α e,pq ( r e ) + δ pq ν p ( r p ), b p = τ p ( r p ), where r e = (x i, x j, x k, y i, y j, y k ), e th ele r p = (x i, x j, y i, y j ), p th edge Use our mesh deformation scheme, which transformed y y + y. This leads to A pq Ãpq, b p b p Uday Khankhoje (EE, IITM) Simulating with uncertainty 29 / 36

87 Moving over to Stochastic Galerkin Recall the earlier (scalar) example of a u = b, where we replaced a a 0 + αθ, where θ was a uniform RV. In the case of the FEM, we have a sparse matrix equation Au = b to solve. Now, A pq must be transformed as per the KL expansion for the surface. Recall: A pq = Σ e α e,pq ( r e ) + δ pq ν p ( r p ), b p = τ p ( r p ), where r e = (x i, x j, x k, y i, y j, y k ), e th ele r p = (x i, x j, y i, y j ), p th edge Use our mesh deformation scheme, which transformed y y + y. This leads to A pq Ãpq, b p b p Uday Khankhoje (EE, IITM) Simulating with uncertainty 29 / 36

88 Steps in Stochastic Galerkin 1 Accomplish the perturbations by Taylor expanding to second order: (1 st, 2 nd derivatives are computed by finite differences) Ã pq = β pq + d i=1 β(i) pq z i + d i,j=1 β(i,j) pq z i z j, matrices m m bp = ζ p + d i=1 ζ(i) p z i, vectors m 1 2 Now do Galerkin testing in the basis of orthogonal polynomials corresponding to the PDF of the RVs: Expand each u w i=1 u iψ i ( z) Take inner products with Ψ i ( z) on both sides of Au = b 3 Resulted in another bigger system of equations: F v = g, F C mw mw, v, g C mw Uday Khankhoje (EE, IITM) Simulating with uncertainty 30 / 36

89 Steps in Stochastic Galerkin 1 Accomplish the perturbations by Taylor expanding to second order: (1 st, 2 nd derivatives are computed by finite differences) Ã pq = β pq + d i=1 β(i) pq z i + d i,j=1 β(i,j) pq z i z j, matrices m m bp = ζ p + d i=1 ζ(i) p z i, vectors m 1 2 Now do Galerkin testing in the basis of orthogonal polynomials corresponding to the PDF of the RVs: Expand each u w i=1 u iψ i ( z) Take inner products with Ψ i ( z) on both sides of Au = b 3 Resulted in another bigger system of equations: F v = g, F C mw mw, v, g C mw Uday Khankhoje (EE, IITM) Simulating with uncertainty 30 / 36

90 Steps in Stochastic Galerkin 1 Accomplish the perturbations by Taylor expanding to second order: (1 st, 2 nd derivatives are computed by finite differences) Ã pq = β pq + d i=1 β(i) pq z i + d i,j=1 β(i,j) pq z i z j, matrices m m bp = ζ p + d i=1 ζ(i) p z i, vectors m 1 2 Now do Galerkin testing in the basis of orthogonal polynomials corresponding to the PDF of the RVs: Expand each u w i=1 u iψ i ( z) Take inner products with Ψ i ( z) on both sides of Au = b 3 Resulted in another bigger system of equations: F v = g, F C mw mw, v, g C mw Uday Khankhoje (EE, IITM) Simulating with uncertainty 30 / 36

91 Details of Stochastic Galerkin What is a the structure of: F v = g, F C mw mw, v, g C mw mw m Each block is sparse, multiplied by stochastic inner products like: E ab = Ψ a ( z)ψ b ( z), E (i) ab = Ψ a( z)z i Ψ b ( z), E (i,j) ab = Ψ a ( z) z i z j Ψ b ( z), Needs to be solved only once Once we solve this equation, we get an expression for the far field, and from that we get the RCS as E f ( r) 2. Uday Khankhoje (EE, IITM) Simulating with uncertainty 31 / 36

92 Details of Stochastic Galerkin What is a the structure of: F v = g, F C mw mw, v, g C mw mw m Each block is sparse, multiplied by stochastic inner products like: E ab = Ψ a ( z)ψ b ( z), E (i) ab = Ψ a( z)z i Ψ b ( z), E (i,j) ab = Ψ a ( z) z i z j Ψ b ( z), Needs to be solved only once Once we solve this equation, we get an expression for the far field, and from that we get the RCS as E f ( r) 2. Uday Khankhoje (EE, IITM) Simulating with uncertainty 31 / 36

93 Details of Stochastic Galerkin What is a the structure of: F v = g, F C mw mw, v, g C mw mw m Each block is sparse, multiplied by stochastic inner products like: E ab = Ψ a ( z)ψ b ( z), E (i) ab = Ψ a( z)z i Ψ b ( z), E (i,j) ab = Ψ a ( z) z i z j Ψ b ( z), Needs to be solved only once Once we solve this equation, we get an expression for the far field, and from that we get the RCS as E f ( r) 2. Uday Khankhoje (EE, IITM) Simulating with uncertainty 31 / 36

94 Computational details of Stochastic Galerkin In solving F v = g, F C mw mw, v, g C mw Not a cheap computation: direct matrix solvers, use an iterative method: BiCgStab + block-diagonal, mean-based pre-conditioner The matrix F is never stored explicitly, instead we only need to be able to compute the product of F with another vector for the iterative method to work Compute times and memory demands are still much higher than MC/SC. Scope for lot of research in sparse data structures for F, g SG-gPC 1200s (10 iters), 10 GB RAM MC 950s (100 surfaces), 150 MB RAM Uday Khankhoje (EE, IITM) Simulating with uncertainty 32 / 36

95 Computational details of Stochastic Galerkin In solving F v = g, F C mw mw, v, g C mw Not a cheap computation: direct matrix solvers, use an iterative method: BiCgStab + block-diagonal, mean-based pre-conditioner The matrix F is never stored explicitly, instead we only need to be able to compute the product of F with another vector for the iterative method to work Compute times and memory demands are still much higher than MC/SC. Scope for lot of research in sparse data structures for F, g SG-gPC 1200s (10 iters), 10 GB RAM MC 950s (100 surfaces), 150 MB RAM Uday Khankhoje (EE, IITM) Simulating with uncertainty 32 / 36