Low-rank tensor methods for stochastic forward and inverse problems

Post on 08-Feb-2017

63 views 0 download

transcript

Low-rank tensor methods for PDEs withuncertain coefficients and

Bayesian Update surrogate

Alexander Litvinenko

Center for UncertaintyQuantification

Center for UncertaintyQuantification

Center for Uncertainty Quantification Logo Lock-up

http://sri-uq.kaust.edu.sa/

Extreme Computing Research Center, KAUST

Alexander Litvinenko Low-rank tensor methods for PDEs with uncertain coefficients and Bayesian Update surrogate

4*

The structure of the talk

Part I (Stochastic forward problem):1. Motivation2. Elliptic PDE with uncertain coefficients3. Discretization and low-rank tensor approximations4. Tensor calculus to compute QoI

Part II (Bayesian update):1. Bayesian update surrogate2. Examples

13

13

17

17

14

14 17

13

17

14 15

13 13

17 29

13 48

15

13 13

13 13

15 13

13

13 16

23

8 8

13 15

28 29

8

8 15

8 15

8 15

19

18 18

61

57

23

17 17

17 17

23 35

57 60

61 117

17

17 17

17 17

14 14

14

7 7

14 14

34

21 14

17 14

28 28

10

10 13

17 17

17 17

11 11

17

11 11

69

40

17 11

17 11

36 28

69 68

10

10 11

9 9

10 11

9

9 12

14 14

21 21

14

14

11

11 11

42

14

11 11

11 11

14 22

38 36

12

12 13

12 12

10 10

12

10 10

23

12 10

10 10

15 15

13

10 10

15 15

69

97

49

28

16 15

12 12

21 21

48 48

83 132

48 91

16

12 12

13 12

8 8

13

8 8

26

13 8

13 8

22 21

13

13 13

9 9

13 13

9

9 13

49

26

9 12

9 13

26 22

49 48

12

12 14

12 14

12 14

15

9 9

18 18

26

15 15

14 14

26 35

15

14 14

15 14

15 14

16

16 19

97

68

29

16 18

16 18

29 35

65 64

97 132

18

18 18

15 15

18 18

15

15

14

7 7

33

15 16

15 17

32 32

16

16 17

14 14

16 17

14

14 18

64

33

11 11

14 18

31 31

72 65

11

11

8

8 14

11 18

11 13

18

13 13

33

18 13

15 13

33 31

20

15 15

19 15

18 15

19

18 18

53

87

136

64

35

19 18

14 14

35 35

64 66

82 128

61 90

33 62

8

8 13

14 14

17 14

18 14

17

17 18

2917 18

10 10

35 35

19

10 10

13 10

19 10

13

13

10

10 14

70

28

13 15

13 13

29 37

56 56

15

13 13

15 13

15 13

19

19

10

10 15

23

11 11

12 12

28 33

11

11 12

11 12

11 12

18

15 15

115

66

23

18 15

18 15

23 30

49 49

121 121

18

18 18

12 12

18 18

12

12 18

22

11 11

11 11

27 27

11

11 11

11 11

10 10

17

10 10

6222

17 10

17 10

21 21

59 49

13

10 10

18 18

10 10

11 11

10

10 11

27

10 11

10 11

32 21

12

12 15

12 13

12 15

13

13 19

88

115

62

27

13 19

13 14

27 32

62 59

115 121

61 90

10

10 11

14 14

21 14

12 12

14

10 10

12 12

29

14 12

15 12

35 35

14

14 15

11 11

14 15

11

11

8

8 16

69

29

11 18

11 23

28 28

62 62

18

18

8

8 15

15 15

13 13

15

13 13

29

15 13

13 13

33 28

16

13 13

16 13

15 13

18

15 15

135

62

29

18 15

18 15

22 22

69 62

101 101

10

10 11

19 19

15 15

7 7

15

7 7

40

15 7

15 7

40 22

19

19

9

9 13

18 18

19 22

18

18

11

10 10

11 11

62

31

18 20

11 11

31 31

39 39

20

11 11

19 11

12 11

19

12 12

26

12 12

14 12

13 13

12

12 14

13 13

4*

KAUST

I received very rich collaboration experience as a co-organizator of:I 3 UQ workshops,I 2 Scalable Hierarchical Algorithms for eXtreme Computing

(SHAXC) workshopsI 1 HPC Conference (www.hpcsaudi.org, 2017)

4*

My interests and collaborations

4*

Motivation to do Uncertainty Quantification (UQ)

Motivation: there is an urgent need to quantify and reduce theuncertainty in output quantities of computer simulations withincomplex (multiscale-multiphysics) applications.

Typical challenges: classical sampling methods are often veryinefficient, whereas straightforward functional representationsare subject to the well-known Curse of Dimensionality.

My goal is systematic, mathematically founded, development ofUQ methods and low-rank algorithms relevant for applications.

Center for UncertaintyQuantification

Center for UncertaintyQuantification

Center for Uncertainty Quantification Logo Lock-up

-1 / 39

4*

UQ and its relevance

Nowadays computational predictions are used in criticalengineering decisions and thanks to modern computers we areable to simulate very complex phenomena. But, how reliableare these predictions? Can they be trusted?

Example: Saudi Aramco currently has a simulator,GigaPOWERS, which runs with 9 billion cells. How sensitiveare the simulation results with respect to the unknown reservoirproperties?

Center for UncertaintyQuantification

Center for UncertaintyQuantification

Center for Uncertainty Quantification Logo Lock-up

0 / 39

4*

Part I: Stochastic forward problem

Part I: Stochastic Galerkin method to solveelliptic PDE with uncertain coefficients

4*

PDE with uncertain coefficient and RHS

Consider− div(κ(x , ω)∇u(x , ω)) = f (x , ω) in G × Ω, G ⊂ R2,u = 0 on ∂G, (1)

where κ(x , ω) - uncertain diffusion coefficient. Since κ positive,usually κ(x , ω) = eγ(x ,ω).For well-posedness see [Sarkis 09, Gittelson 10, H.J.Starkloff11, Ullmann 10].Further we will assume that covκ(x , y) is given.

Center for UncertaintyQuantification

Center for UncertaintyQuantification

Center for Uncertainty Quantification Logo Lock-up

1 / 39

4*

My previous work

After applying the stochastic Galerkin method, obtain:Ku = f, where all ingredients are represented in a tensor format

Compute maxu, var(u), level sets of u, sign(u)[1] Efficient Analysis of High Dimensional Data in Tensor Formats,

Espig, Hackbusch, A.L., Matthies and Zander, 2012.

Research which ingredients influence on the tensor rank of K[2] Efficient low-rank approximation of the stochastic Galerkin matrix in tensor formats,

Wahnert, Espig, Hackbusch, A.L., Matthies, 2013.

Approximate κ(x , ω), stochastic Galerkin operator K in TensorTrain (TT) format, solve for u, postprocessing[3] Polynomial Chaos Expansion of random coefficients and the solution of stochastic

partial differential equations in the Tensor Train format, Dolgov, Litvinenko, Khoromskij, Matthies, 2016.

Center for UncertaintyQuantification

Center for UncertaintyQuantification

Center for Uncertainty Quantification Logo Lock-up

2 / 39

4*

Typical quantities of interest

Keeping all input and intermediate data in a tensorrepresentation one wants to perform different tasks:

I evaluation for specific parameters (ω1, . . . , ωM),I finding maxima and minima,I finding ‘level sets’ (needed for histogram and probability

density).Example of level set: all elements of a high dimensional tensorfrom the interval [0.7,0.8].

Center for UncertaintyQuantification

Center for UncertaintyQuantification

Center for Uncertainty Quantification Logo Lock-up

3 / 39

4*

Canonical and Tucker tensor formats

Definition and Examples of tensors

Center for UncertaintyQuantification

Center for UncertaintyQuantification

Center for Uncertainty Quantification Logo Lock-up

4 / 39

4*

Canonical and Tucker tensor formats

[Pictures are taken from B. Khoromskij and A. Auer lecture course]

Storage: O(nd )→ O(dRn) and O(Rd + dRn).

Center for UncertaintyQuantification

Center for UncertaintyQuantification

Center for Uncertainty Quantification Logo Lock-up

5 / 39

4*

Definition of tensor of order d

Tensor of order d is a multidimensional array over a d-tupleindex set I = I1 × · · · × Id ,

A = [ai1...id : i` ∈ I`] ∈ RI , I` = 1, ...,n`, ` = 1, ..,d .

A is an element of the linear space

Vn =d⊗`=1

V`, V` = RI`

equipped with the Euclidean scalar product 〈·, ·〉 : Vn ×Vn → R,defined as

〈A,B〉 :=∑

(i1...id )∈I

ai1...id bi1...id , for A, B ∈ Vn.

Center for UncertaintyQuantification

Center for UncertaintyQuantification

Center for Uncertainty Quantification Logo Lock-up

6 / 39

4*

Examples of rank-1 and rank-2 tensors

Rank-1:f (x1, ..., xd ) = exp(f1(x1) + ...+ fd (xd )) =

∏dj=1 exp(fj(xj))

Rank-2: f (x1, ..., xd ) = sin(∑d

j=1 xj), since

2i · sin(∑d

j=1 xj) = ei∑d

j=1 xj − e−i∑d

j=1 xj

Rank-d function f (x1, ..., xd ) = x1 + x2 + ...+ xd can beapproximated by rank-2: with any prescribed accuracy:

f ≈∏d

j=1(1 + εxj)

ε−∏d

j=1 1ε

+O(ε), as ε→ 0

Center for UncertaintyQuantification

Center for UncertaintyQuantification

Center for Uncertainty Quantification Logo Lock-up

7 / 39

4*

Tensor and Matrices

Rank-1 tensor

A = u1 ⊗ u2 ⊗ ...⊗ ud =:d⊗µ=1

Ai1,...,id = (u1)i1 · ... · (ud )id

Rank-1 tensor A = u ⊗ v , matrix A = uvT , A = vuT , u ∈ Rn,v ∈ Rm,Rank-k tensor A =

∑ki=1 ui ⊗ vi , matrix A =

∑ki=1 uivT

i .Kronecker product of n × n and m ×m matrices is a new blockmatrix A⊗ B ∈ Rnm×nm, whose ij-th block is [AijB].

Center for UncertaintyQuantification

Center for UncertaintyQuantification

Center for Uncertainty Quantification Logo Lock-up

8 / 39

4*

Discretization of elliptic PDE

Now let us discretize our diffusion equation withuncertain coefficients

Center for UncertaintyQuantification

Center for UncertaintyQuantification

Center for Uncertainty Quantification Logo Lock-up

9 / 39

4*

Karhunen Loeve and Polynomial Chaos Expansions

Apply bothKarhunen Loeve Expansion (KLE):κ(x , ω) = κ0(x) +

∑∞j=1 κjgj(x)ξj(θ(ω)), where

θ = θ(ω) = (θ1(ω), θ2(ω), ..., ),ξj(θ) = 1

κj

∫G (κ(x , ω)− κ0(x)) gj(x)dx .

Polynomial Chaos Expansion (PCE)κ(x , ω) =

∑α κ

(α)(x)Hα(θ), compute ξj(θ) =∑

α∈J ξ(α)j Hα(θ),

where ξ(α)j = 1κj

∫G κ

(α)(x)gj(x)dx .

Further compute ξ(α)j ≈∑s

`=1(ξ`)j∏∞

k=1(ξ`, k )αk .

Center for UncertaintyQuantification

Center for UncertaintyQuantification

Center for Uncertainty Quantification Logo Lock-up

10 / 39

4*

Final discretized stochastic PDE

Ku = f, where

K:=∑s

`=1K` ⊗⊗M

µ=1∆`µ, K` ∈ RN×N , ∆`µ ∈ RRµ×Rµ ,u:=

∑rj=1 uj ⊗

⊗Mµ=1 ujµ, uj ∈ RN , ujµ ∈ RRµ ,

f:=∑R

k=1 f k ⊗⊗M

µ=1 gkµ, f k ∈ RN and gkµ ∈ RRµ .(Wahnert, Espig, Hackbusch, Litvinenko, Matthies, 2011)

Examples of stochastic Galerkin matrices:

Center for UncertaintyQuantification

Center for UncertaintyQuantification

Center for Uncertainty Quantification Logo Lock-up

11 / 39

4*

Computing QoI in low-rank tensor format

Now, we consider how tofind maxima in a high-dimensional tensor

4*

Maximum norm and corresponding index

Let u =∑r

j=1⊗d

µ=1 ujµ ∈ Rr , compute

‖u‖∞ := maxi:=(i1,...,id )∈I

|ui | = maxi:=(i1,...,id )∈I

∣∣∣∣∣∣r∑

j=1

d∏µ=1

(ujµ)

∣∣∣∣∣∣ .Computing ‖u‖∞ is equivalent to the following e.v. problem.

Let i∗ := (i∗1 , . . . , i∗d ) ∈ I, #I =

∏dµ=1 nµ.

‖u‖∞ = |ui∗ | =

∣∣∣∣∣∣r∑

j=1

d∏µ=1

(ujµ)

i∗µ

∣∣∣∣∣∣ and e(i∗) :=d⊗µ=1

ei∗µ ,

where ei∗µ ∈ Rnµ the i∗µ-th canonical vector in Rnµ (µ ∈ N≤d ).

Center for UncertaintyQuantification

Center for UncertaintyQuantification

Center for Uncertainty Quantification Logo Lock-up

12 / 39

Then

u e(i∗) =

r∑j=1

d⊗µ=1

ujµ

d⊗µ=1

ei∗µ

=r∑

j=1

d⊗µ=1

ujµ ei∗µ

=r∑

j=1

d⊗µ=1

[(ujµ)i∗µei∗µ

]

=

r∑j=1

d∏µ=1

(ujµ)i∗µ

︸ ︷︷ ︸

ui∗=

d⊗µ=1

e(i∗µ) = ui∗e(i∗).

Thus, we obtained an “eigenvalue problem”:

u e(i∗) = ui∗e(i∗).

Center for UncertaintyQuantification

Center for UncertaintyQuantification

Center for Uncertainty Quantification Logo Lock-up

13 / 39

4*

Computing ‖u‖∞, u ∈ Rr by vector iteration

By defining the following diagonal matrix

D(u) :=r∑

j=1

d⊗µ=1

diag((ujµ)`µ

)`µ∈N≤nµ

(2)

with representation rank r , obtain D(u)v = u v .Now apply the well-known vector iteration method (with ranktruncation) to

D(u)e(i∗) = ui∗e(i∗),

obtain ‖u‖∞.[Approximate iteration, Khoromskij, Hackbusch, Tyrtyshnikov 05],

and [Espig, Hackbusch 2010]

Center for UncertaintyQuantification

Center for UncertaintyQuantification

Center for Uncertainty Quantification Logo Lock-up

14 / 39

4*

How to compute the mean value in CP format

Let u =∑r

j=1⊗d

µ=1 ujµ ∈ Rr , then the mean value u can becomputed as a scalar product

u =

⟨ r∑j=1

d⊗µ=1

ujµ

,

d⊗µ=1

1nµ

⟩ =r∑

j=1

d⊗µ=1

⟨ujµ, 1µ

⟩nµ

=

(3)

=r∑

j=1

d∏µ=1

1nµ

( nµ∑k=1

(ujµ)k

), (4)

where 1µ := (1, . . . ,1)T ∈ Rnµ .Numerical cost is O

(r ·∑d

µ=1 nµ)

.

Center for UncertaintyQuantification

Center for UncertaintyQuantification

Center for Uncertainty Quantification Logo Lock-up

15 / 39

4*

Numerical Experiments

2D L-shape domain, N = 557 dofs.Total stochastic dimension is Mu = Mk + Mf = 20, there are|J | = 231 PCE coefficients

u =231∑j=1

uj,0 ⊗20⊗µ=1

ujµ ∈ R557 ⊗20⊗µ=1

R3.

Center for UncertaintyQuantification

Center for UncertaintyQuantification

Center for Uncertainty Quantification Logo Lock-up

16 / 39

4*

Level sets

Now we compute level sets

sign(b‖u‖∞1− u)for b ∈ 0.2, 0.4, 0.6, 0.8.

I Tensor u has 320 ∗ 557 ≈ 2 · 1012 entries ≈ 16 TB ofmemory.

I The computing time of one level set was 10 minutes.I Intermediate ranks of sign(b‖u‖∞1− u) and of rank(uk )

were less than 24.

Center for UncertaintyQuantification

Center for UncertaintyQuantification

Center for Uncertainty Quantification Logo Lock-up

17 / 39

4*

Part II

Part II: Bayesian update

We will speak about Gauss-Markov-Kalman filter for theBayesian updating of parameters in comput. model.

4*

Mathematical setup

Consider

K (u; q) = f ⇒ u = S(f ; q),

where S is solution operator.Operator depends on parameters q ∈ Q,hence state u ∈ U is also function of q:

Measurement operator Y with values in Y:

y = Y (q; u) = Y (q,S(f ; q)).

Examples of measurements:y(ω) =

∫D0

u(ω, x)dx , or u in few points

Center for UncertaintyQuantification

Center for UncertaintyQuantification

Center for Uncertainty Quantification Logo Lock-up

18 / 39

4*

Random QoI

With state u a RV, the quantity to be measured

y(ω) = Y (q(ω),u(ω)))

is also uncertain, a random variable.Noisy data: y + ε(ω),

where y is the “true” value and a random error ε.

Forecast of the measurement: z(ω) = y(ω) + ε(ω).

Center for UncertaintyQuantification

Center for UncertaintyQuantification

Center for Uncertainty Quantification Logo Lock-up

19 / 39

4*

Conditional probability and expectation

Classically, Bayes’s theorem gives conditional probability

P(Iq|Mz) =P(Mz |Iq)

P(Mz)P(Iq) (orπq(q|z) =

p(z|q)

Zspq(q));

Expectation with this posterior measure is conditionalexpectation.

Kolmogorov starts from conditional expectation E (·|Mz),from this conditional probability via P(Iq|Mz) = E

(χIq |Mz

).

Center for UncertaintyQuantification

Center for UncertaintyQuantification

Center for Uncertainty Quantification Logo Lock-up

20 / 39

4*

Conditional expectation

The conditional expectation is defined asorthogonal projection onto the closed subspace L2(Ω,P, σ(z)):

E(q|σ(z)) := PQ∞q = argminq∈L2(Ω,P,σ(z)) ‖q − q‖2L2

The subspace Q∞ := L2(Ω,P, σ(z)) represents the availableinformation.

The update, also called the assimilated valueqa(ω) := PQ∞q = E(q|σ(z)), is a Q-valued RV

and represents new state of knowledge after the measurement.Doob-Dynkin: Q∞ = ϕ ∈ Q : ϕ = φ z, φmeasurable.

Center for UncertaintyQuantification

Center for UncertaintyQuantification

Center for Uncertainty Quantification Logo Lock-up

21 / 39

4*

Numerical computation of NLBU

Look for ϕ such that q(ξ) = ϕ(z(ξ)), z(ξ) = y(ξ) + ε(ω):

ϕ ≈ ϕ =∑α∈Jp

ϕαΦα(z(ξ))

and minimize ‖q(ξ)− ϕ(z(ξ))‖2L2, where Φα are polynomials

(e.g. Hermite, Laguerre, Chebyshev or something else).Taking derivatives with respect to ϕα:

∂ϕα〈q(ξ)− ϕ(z(ξ)),q(ξ)− ϕ(z(ξ))〉 = 0 ∀α ∈ Jp

Inserting representation for ϕ, obtain:

Center for UncertaintyQuantification

Center for UncertaintyQuantification

Center for Uncertainty Quantification Logo Lock-up

22 / 39

4*

Numerical computation of NLBU

∂ϕαE

q2(ξ)− 2∑β∈J

qϕβΦβ(z) +∑β,γ∈J

ϕβϕγΦβ(z)Φγ(z)

= 2E

−qΦα(z) +∑β∈J

ϕβΦβ(z)Φα(z)

= 2

∑β∈J

E [Φβ(z)Φα(z)]ϕβ − E [qΦα(z)]

= 0 ∀α ∈ J .

Center for UncertaintyQuantification

Center for UncertaintyQuantification

Center for Uncertainty Quantification Logo Lock-up

23 / 39

4*

Numerical computation of NLBU

Now, rewriting the last sum in a matrix form, obtain the linearsystem of equations (=: A) to compute coefficients ϕβ: ... ... ...

... E [Φα(z(ξ))Φβ(z(ξ))]...

... ... ...

...ϕβ

...

=

...

E [q(ξ)Φα(z(ξ))]...

,

where α, β ∈ J , A is of size |J | × |J |.

Center for UncertaintyQuantification

Center for UncertaintyQuantification

Center for Uncertainty Quantification Logo Lock-up

24 / 39

4*

Numerical computation of NLBU

We can rewrite the system above in the compact form:

[Φ] [diag(...wi ...)] [Φ]T

...ϕβ...

= [Φ]

w0q(ξ0)...

wNq(ξN)

[Φ] ∈ RJα×N , [diag(...wi ...)] ∈ RN×N , [Φ] ∈ RJα×N .Solving this system, obtain vector of coefficients (...ϕβ...)

T forall β.Finally, the assimilated parameter qa will be

qa = qf + ϕ(y)− ϕ(z), (5)

z(ξ) = y(ξ) + ε(ω), ϕ =∑

β∈JpϕβΦβ(z(ξ))

Center for UncertaintyQuantification

Center for UncertaintyQuantification

Center for Uncertainty Quantification Logo Lock-up

25 / 39

4*

Example: Lorenz 1963 problem (chaotic system of ODEs)

x = σ(ω)(y − x)

y = x(ρ(ω)− z)− yz = xy − β(ω)z

Initial state q0(ω) = (x0(ω), y0(ω), z0(ω)) are uncertain.

Solving in t0, t1, ..., t10, Noisy Measur. → UPDATE, solving int11, t12, ..., t20, Noisy Measur. → UPDATE,...

IDEA of the Bayesian Update (BU):Take qf (ω) = q0(ω).Linear BU: qa = qf + K · (z − y)Non-Linear BU: qa = qf + H1 · (z − y) + (z − y)T · H2 · (z − y).

Center for UncertaintyQuantification

Center for UncertaintyQuantification

Center for Uncertainty Quantification Logo Lock-up

26 / 39

Trajectories of x,y and z in time. After each update (newinformation coming) the uncertainty drops. [O. Pajonk, B. V. Rosic, A.

Litvinenko, and H. G. Matthies, 2012]

Center for UncertaintyQuantification

Center for UncertaintyQuantification

Center for Uncertainty Quantification Logo Lock-up

27 / 39

4*

Example: Lorenz problem

10 0 100

0.10.20.30.40.50.60.70.8

x

20 0 200

0.050.1

0.150.2

0.250.3

0.350.4

0.45

y

0 10 200

0.10.20.30.40.50.60.70.80.9

1

z

xfxa

yfya

zfza

Figure: quadratic BU surrogate, measure the state (x(t), y(t), z(t)).Prior and posterior after one update.

Center for UncertaintyQuantification

Center for UncertaintyQuantification

Center for Uncertainty Quantification Logo Lock-up

28 / 39

4*

Example: Lorenz Problem

10 5 00

0.10.20.30.40.50.60.70.80.9

x

x1x2

15 10 50

0.050.1

0.150.2

0.250.3

0.350.4

0.450.5

y

y1y2

5 10 150

0.10.20.30.40.50.6

z

z1z2

Figure: Comparison of the posterior functions computed by linear andquadratic BU after second update.

Center for UncertaintyQuantification

Center for UncertaintyQuantification

Center for Uncertainty Quantification Logo Lock-up

29 / 39

4*

Example: Lorenz Problem

20 0 200

0.020.040.060.080.1

0.120.140.16

x

50 0 500

0.010.020.030.040.050.060.070.080.09

y

0 10 200

0.020.040.060.080.1

0.120.140.160.18

z

xfxa

yfya

zfza

Figure: Quadratic measurement (x(t)2, y(t)2, z(t)2): Comparison of apriori and a posterior for NLBU

Center for UncertaintyQuantification

Center for UncertaintyQuantification

Center for Uncertainty Quantification Logo Lock-up

30 / 39

4*

Example: 1D elliptic PDE with uncertain coeffs

−∇ · (κ(x , ξ)∇u(x , ξ)) = f (x , ξ), x ∈ [0,1]

+ Dirichlet random b.c. g(0, ξ) and g(1, ξ).3 measurements: u(0.3) = 22, s.d. 0.2, x(0.5) = 28, s.d. 0.3,x(0.8) = 18, s.d. 0.3.

I κ(x, ξ): N = 100 dofs, M = 5, number of KLE terms 35, beta distribution for κ, Gaussian covκ, cov.length 0.1, multi-variate Hermite polynomial of order pκ = 2;

I RHS f (x, ξ): Mf = 5, number of KLE terms 40, beta distribution for κ, exponential covf , cov. length 0.03,multi-variate Hermite polynomial of order pf = 2;

I b.c. g(x, ξ): Mg = 2, number of KLE terms 2, normal distribution for g, Gaussian covg , cov. length 10,multi-variate Hermite polynomial of order pg = 1;

I pφ = 3 and pu = 3

Center for UncertaintyQuantification

Center for UncertaintyQuantification

Center for Uncertainty Quantification Logo Lock-up

31 / 39

4*

Example: updating of the solution u

0 0.5 1-20

0

20

40

60

0 0.5 1-20

0

20

40

60

Figure: Original and updated solutions, mean value plus/minus 1,2,3standard deviations

[graphics are built in the stochastic Galerkin library sglib, written by E. Zander in TU Braunschweig]

Center for UncertaintyQuantification

Center for UncertaintyQuantification

Center for Uncertainty Quantification Logo Lock-up

32 / 39

4*

Example: Updating of the parameter

0 0.5 10

0.5

1

1.5

0 0.5 10

0.5

1

1.5

Figure: Original and updated parameter κ.

Center for UncertaintyQuantification

Center for UncertaintyQuantification

Center for Uncertainty Quantification Logo Lock-up

33 / 39

4*

Future plans and possible collaboration

Future plans and possible collaboration ideas

4*

Future plans, Idea N1

Possible collaboration work with Troy Butler: To develop alow-rank adaptive goal-oriented Bayesian update technique. Thesolution of the forward and inverse problems will be considered as awhole adaptive process, controlled by error/uncertainty estimators.

z

(y - z) q

f ε

forward update

low-rank and adaptive

y

f z

(y - z)

ε

forwardy q.....

low-rank and adaptive

... q update

Stochastic forward spatial discret.

stochastic discret.

low-rank approx.

Inverse problem

Errors

inverse operator approx.

4*

Future plans, Idea N2

Edge between Green functions in PDEs and covariancematrices.Possible collaboration with statistical group, Doug Nychka(NCAR), Havard Rue

Center for UncertaintyQuantification

Center for UncertaintyQuantification

Center for Uncertainty Quantification Logo Lock-up

34 / 39

4*

Future plans, Idea N3

Data assimilation techniques, Bayesian update surrogare.Develop non-linear, non-Gaussian Bayesian updateapproximation for gPCE coefficients.Possible collaboration with Jan Mandel, Troy Butler, Kody Law,Y. Marzouk, H. Najm, TU Braunschweig and KAUST

4*

Collaborators

1. Uncertainty quantification and Bayesian Update: Prof. H.Matthies, Bojana V. Rosic, Elmar Zander, Oliver Pajonkfrom TU Braunschweig, Germany,

2. Low-rank tensor calculus: Mike Espig from RWTH Aachen,Boris and Venera Khoromskij from MPI Leipzig

3. Spatial and environmental statistics: Marc Genton, YingSun, Raphael Huser, Brian Reich, Ben Shaby and DavidBolin.

4. Some others: UQ, data assimilation, high-dimensionalproblems/statistics

4*

Conclusion

I Introduced low-rank tensor methods to solve elliptic PDEswith uncertain coefficients,

I Explained how to compute the maximum, the mean, levelsets,... in low-rank tensor format,

I Derived Bayesian update surrogate ϕ (as a linear,quadratic, cubic etc approximation), i.e. computeconditional expectation of q, given measurement y .

Center for UncertaintyQuantification

Center for UncertaintyQuantification

Center for Uncertainty Quantification Logo Lock-up

34 / 39

4*

Example: Canonical rank d , whereas TT rank 2

d-Laplacian over uniform tensor grid. It is known to have theKronecker rank-d representation,

∆d = A⊗IN⊗...⊗IN +IN⊗A⊗...⊗IN +...+IN⊗IN⊗...⊗A ∈ RI⊗d⊗I⊗d

(6)with A = ∆1 = tridiag−1,2,−1 ∈ RN×N , and IN being theN × N identity. Notice that for the canonical rank we have rankkC(∆d ) = d , while TT-rank of ∆d is equal to 2 for anydimension due to the explicit representation

∆d = (∆1 I)×(

I 0∆1 I

)× ...×

(I 0

∆1 I

)×(

I∆1

)(7)

where the rank product operation ”×” is defined as a regularmatrix product of the two corresponding core matrices, theirblocks being multiplied by means of tensor product. The similarbound is true for the Tucker rank rankTuck (∆d ) = 2.

4*

Advantages and disadvantages

Denote k - rank, d-dimension, n = # dofs in 1D:

1. CP: ill-posed approx. alg-m, O(dnk), hard to computeapprox.

2. Tucker: reliable arithmetic based on SVD, O(dnk + kd )

3. Hierarchical Tucker: based on SVD, storage O(dnk + dk3),truncation O(dnk2 + dk4)

4. TT: based on SVD, O(dnk2) or O(dnk3), stable5. Quantics-TT: O(nd )→ O(d logqn)

4*

How to compute the variance in CP format

Let u ∈ Rr and

u := u − ud⊗µ=1

1nµ

1 =r+1∑j=1

d⊗µ=1

ujµ ∈ Rr+1, (8)

then the variance var(u) of u can be computed as follows

var(u) =〈u, u〉∏dµ=1 nµ

=1∏d

µ=1 nµ

⟨r+1∑i=1

d⊗µ=1

uiµ

,

r+1∑j=1

d⊗ν=1

ujν

=r+1∑i=1

r+1∑j=1

d∏µ=1

1nµ

⟨uiµ, ujµ

⟩.

Numerical cost is O(

(r + 1)2 ·∑d

µ=1 nµ)

.

4*

Computing QoI in low-rank tensor format

Now, we consider how tofind ‘level sets’,

for instance, all entries of tensor u from interval [a,b].

4*

Definitions of characteristic and sign functions

1. To compute level sets and frequencies we needcharacteristic function.2. To compute characteristic function we need sign function.

The characteristic χI(u) ∈ T of u ∈ T in I ⊂ R is for every multi-index i ∈ I pointwise defined as

(χI(u))i :=

1, ui ∈ I,0, ui /∈ I.

Furthermore, the sign(u) ∈ T is for all i ∈ I pointwise definedby

(sign(u))i :=

1, ui > 0;−1, ui < 0;0, ui = 0.

Center for UncertaintyQuantification

Center for UncertaintyQuantification

Center for Uncertainty Quantification Logo Lock-up

36 / 39

4*

sign(u) is needed for computing χI(u)

LemmaLet u ∈ T , a,b ∈ R, and 1 =

⊗dµ=1 1µ, where

1µ := (1, . . . ,1)t ∈ Rnµ .(i) If I = R<b, then we have χI(u) = 1

2(1+ sign(b1− u)).

(ii) If I = R>a, then we have χI(u) = 12(1− sign(a1− u)).

(iii) If I = (a,b), then we haveχI(u) = 1

2(sign(b1− u)− sign(a1− u)).

Computing sign(u), u ∈ Rr , via hybrid Newton-Schulz iterationwith rank truncation after each iteration.

Center for UncertaintyQuantification

Center for UncertaintyQuantification

Center for Uncertainty Quantification Logo Lock-up

37 / 39

4*

Level Set, Frequency

Definition (Level Set, Frequency)Let I ⊂ R and u ∈ T . The level set LI(u) ∈ T of u respect to I ispointwise defined by

(LI(u))i :=

ui ,ui ∈ I ;0,ui /∈ I ,

for all i ∈ I.The frequency FI(u) ∈ N of u respect to I is defined as

FI(u) := # suppχI(u).

Center for UncertaintyQuantification

Center for UncertaintyQuantification

Center for Uncertainty Quantification Logo Lock-up

38 / 39

4*

Computation of level sets and frequency

PropositionLet I ⊂ R, u ∈ T , and χI(u) its characteristic. We have

LI(u) = χI(u) u

and rank(LI(u)) ≤ rank(χI(u)) rank(u).The frequency FI(u) ∈ N of u respect to I is

FI(u) = 〈χI(u),1〉 ,

where 1 =⊗d

µ=1 1µ, 1µ := (1, . . . ,1)T ∈ Rnµ .

Center for UncertaintyQuantification

Center for UncertaintyQuantification

Center for Uncertainty Quantification Logo Lock-up

39 / 39