ETH Z · Dank DievorliegendeArbeitist unterderLeitung vonHerrnProf. ChristophSchwabentstanden....
Transcript of ETH Z · Dank DievorliegendeArbeitist unterderLeitung vonHerrnProf. ChristophSchwabentstanden....
Research Collection
Doctoral Thesis
Sparse perturbation algorithms for elliptic PDE's with stochasticdata
Author(s): Todor, Radu Alexandru
Publication Date: 2005
Permanent Link: https://doi.org/10.3929/ethz-a-005151392
Rights / License: In Copyright - Non-Commercial Use Permitted
This page was generated automatically upon download from the ETH Zurich Research Collection. For moreinformation please consult the Terms of use.
ETH Library
Diss. ETH No. 16192
Sparse Perturbation Algorithmsfor Elliptic PDE's with Stochastic Data
A dissertation submitted to the
SWISS FEDERAL INSTITUTE OF TECHNOLOGY ZURICH
for the degree of
Doctor of Mathematics
presented by
RADU ALEXANDRU TODOR
Dipl. Math. University of Bucharest
born November 12, 1973
citizen of Romania
accepted on the recommendation of
Prof. Dr. Christoph Schwab, examiner
Prof. Dr. Ralf Hiptmair, co-examiner
Prof. Dr. Reinhold Schneider, co-examiner
2005
Dank
Die vorliegende Arbeit ist unter der Leitung von Herrn Prof. Christoph Schwab entstanden.
Ich danke ihm ganz herzlich für seine stete Unterstützung, seine hilfreichen und motivieren¬
den Ratschläge, und seine immer sehr engagierte Betreuung meiner Doktorarbeit.
Herrn Prof. Ralf Hiptmair und Herrn Prof. Reinhold Schneider danke ich für ihre kon¬
struktiven Bemerkungen und die Übernahme des Koreferats.
Ferner gilt mein Dank meiner Familie, die mich immer liebevoll unterstützt hat, sowie
meinen Kolleginnen und Kollegen aus dem Seminar für Angewandte Mathematik, die mir
den Aufenthalt an der ETH angenehm gemacht haben.
Financial support by the European Union (contract number HPRN-CT-2002-00286) and
the Swiss Federal Office for Science and Education (grant BBW 02.0418), within the IHP
network 'Breaking Complexity', is gratefully acknowledged.
i
Seite Leer /
Blank leaf
Abstract
In this work we investigate elliptic partial differential equations with stochastic data. We
develop fully deterministic algorithms using statistics of the data (expectation and/or
higher order moments) as input, for the computation of similar statistics of the stochastic
solution. We introduce a sparsified higher order moment representation which, in the con¬
text of a classical perturbation method enables us to formulate efficient algorithms for the
computation of solution moments, using sparse grids and wavelet techniques. Assuming
analytic spatial regularity of the data fluctuation, we prove almost linear complexity for
these algorithms.
Choosing stationary diffusion in a bounded domain as a model problem, we discuss in
Chapter 1 the setting, basic notations and definitions, plus the main results.
We devote Chapter 2 to the case of a stochastic source term and deterministic diffusion
coefficient. We derive deterministic moment equations and use sparse grids discretization
to preserve almost optimal convergence rates. We propose a new solution algorithm for
the resulting linear system given by a well-conditioned (due to the use of a wavelet basis),
yet fully populated matrix with a tensor product block structure, to achieve log-linear
complexity (in the number of degrees of freedom) despite the high dimensionality of the
moment problem.
In Chapter 3 we address the more general case of a stochastic diffusion coefficient and
stochastic source term. We show that the expectation of the stochastic solution can be
computed starting from the higher order moments of the data. We introduce approximate
sparsified representations of these moments and show that using them as input in the com¬
putation of the solution expectation results in an algorithm of almost linear complexity,
under the analyticity assumption mentioned above. We finally combine this algorithm with
the one developed in Chapter 2 to obtain similar, efficient algorithms for the computation
of higher order moments of the stochastic solution.
We conclude by discussing in Chapter 4 possible further theoretical developments and
extensions of the results presented in this thesis.
m
Seite Leer /
Blank leaf
Kurzfassung
Gegenstand dieser Arbeit ist die Untersuchung elliptischer partieller Differentialgleichungenmit stochastischen Daten. Basierend auf Datenstatistiken (Erwartungswert, höhere Mo¬
mente) werden rein deterministische Algorithmen zur Berechnung von ähnlichen Statistiken
der stochastischen Lösung entwickelt. Effiziente Darstellungen der höheren Momente eines
Zufallsfeldes werden eingeführt, und zusammen mit dünnen Gittern und wavelet Techniken
im Rahmen einer Störungsmethode verwendet, um schnelle Verfahren zur Berechnung der
Lösungsmomente zu erhalten. Im Falle der im Ort analytischen Datenfluktuationen wird
fast lineare Gesamtkomplexität gezeigt.
In Kapitel 1 wird das Modelproblems (stationäre Diffusion im beschränkten physikalischen
Gebiet), die grundlegenden Konzepte und Definitionen eingeführt, sowie die Hauptresul¬tate erläutert.
Der Fall einer stochastischen Quelle mit deterministischem Diffusionskoeffizient wird in
Kapitel 2 behandelt. Hoch dimensionale, deterministische Momentgleichungen werden
hergeleitet, mit Hilfe von dünnen Gittern und Waveletbasen diskretisiert, was zu gut kon¬
ditionierten, allerdings voll besetzten linearen Systemen führt. Ein Lösungsverfahren wird
entwickelt, das die Tensorproduktstruktur des ursprünglichen Momentproblems ausnutzt,
um log-lineare Gesamtkomplexität zu erreichen.
Für das vollständig stochastische Diffusionsproblem (sowohl Quelle als auch Diffusionsko¬
effizient werden als Zufallsfelder modelliert) wird in Kapitel 3 ein Algorithmus zur Berech¬
nung dos Erwartungswertes entwickelt. Höhere Momente der stochastischen Daten wer¬
den algorithmisch in eine Approximation des Erwartungswertes der stochastischen Lösung
umgesetzt. Im Falle einer im Ort regulären Fluktuation der Daten und unter Verwendungder effizienten Darstellungen der Momente als Eingabe wird erneut fast lineare Gesamtkom¬
plexität bewiesen. Ferner wird dieser Algorithmus mit dem in Kapitel 2 beschriebenen Ver¬
fahren kombiniert, was auch eine effiziente Berechnung der höheren Momente ermöglicht.
Im letzten Kapitel werden Möglichkeiten zur Verallgemeinerung der zuvor entwickelten
Theorie aufgezeigt.
v
VI
Contents
Dank i
Abstract iü
Kurzfassung v
1 Introduction 1
1.1 Problem Formulation 1
1.2 Main Results 4
1.3 Notations and Definitions 8
2 Stochastic Source Term 11
2.1 Generalities 11
2.2 The Moment Equation 13
2.3 Discretization with Sparse FEM 14
2.4 Iterative Solution and Complexity 17
3 Stochastic Coefficient 27
3.1 Random Fields 27
3.2 Discretization 30
3.3 Neumann Series Decompsition 31
3.4 Discrete Mixed Moment Problem 33
3.4.1 Mixed Moment Algorithm 36
3.4.2 Sparsification 41
3.4.3 Complexity 43
3.5 Mean Field Computation 44
3.6 Computation of Higher Order Moments 45
3.6.1 Discretization 46
3.6.2 Algorithms 48
4 Conclusions and Open Issues 57
vii
5 Appendix 59
5.1 A Trace Lemma 59
5.2 A Fast Decaying Sequence 61
Bibliography 63
Curriculum Vitae 67
viii
Chapter 1
Introduction
1.1 Problem Formulation
The rapid development of (adaptive) algorithms, hardware and software in recent years
has made the accurate numerical solution of elliptic partial differential equations a routine
matter in many (but not all) engineering applications. Despite this, results of accurate finite
element (FE) computations often deviate significantly from the responses of the physicalsystem under consideration. Having eliminated the discretization error, and assumingthat the modelling error inherent in the selected partial differential equations (pde's) is
negligible (i.e. that the adopted pde's precisely describe the physics of the system under
consideration), the gap between simulation and observation must be due to uncertainty in
the input data.
One task for applied mathematics and scientific computing is therefore (see e.g. [GS91],[ER03]) to develop tools for uncertainty processing, i.e. the systematic and quantitativenumerical representation of uncertainty in input data and its propagation to the outputof a finite element simulation. This requires new developments in several areas of appliedmathematics and engineering: input parameters are replaced by random variables resp.
by random fields with known or estimated statistics, and traditional deterministic finite
element methods are reformulated to allow for randomness in input data and solution.
In this work, we focus on this latter aspect, i.e. the formulation, design and analysis of
deterministic finite element solutions of elliptic pde's with stochastic data.
We consider as a model problem diffusion in a bounded domain D C Rd, where the data
(source term / and diffusion coefficient a) and therefore the solution u too, are allowed to
depend randomly on a new parameter w £ Q,
-div(a(-,u;)Vu(-,aO) = /(-,£*;) in H~l{D) P-a.e. w e ft. (1.1)
Here we assume the sample set Ü to be equipped with a probability measure P quantifyingthe uncertainty and giving the probability of occurance for each event a in a set E C 2U
of admissible events (that is, (ft, £, P) probability space).
1
2 CHAPTER 1. INTRODUCTION
Identifying the data a, f as well as the solution u with random vector fields we rewrite (1.1)as
Affl(w)«(w) = /M in H~l{D) P-a.e. eu G ft. (1.2)
Both (1,1) and (1.2) are to be understood here and throughout this work variationally in
H£{D), that is,
/ a(x, uj)Vu(x) Vu(x) da; = (/(•, w), u(-)>H-i(o),Hä(D) Vu H^(D) P-a.e. weft.
D
Under the uniform ellipticity assumption (A denotes here the Lebesgue measure)
a_ < a(x,uj) < a+ Ax P-a.e. (x,u) e D x ft,
(1.1) is well-posed in L°(Ü,II^(D)) for data / L°(Ü, H-^D)) and a G L°(D x ft) (hereL° denotes measurability). For equation (1.1) we then formulate the following
Problem. Given statistical information w.r.t. uj about the data (a, /), compute statistical
information w.r.t. u) about u solution to (1.1) (statistical information comprises first and
higher order moments, variance, probability distribution function etc.).
Since for a problem with stochastic data knowing all joint probability densities (i.e. com¬
plete statistical information) is in practice hardly the case, reasonable assumptions can be
made only on some data statistics. Defining for any ieN+ the moment of order k of a
measurable u : D x ft — R by
Mk(u) :Dk :=DxDx---xD-+R
Mk(u)(x1,x2,..., xfc) := / u(xi,u)u(x2,ui) u(xk,u) dP(uj)n
whenever the integral exists, the moment version of the problem above reads,
Moment Problem. Given moments w.r.t. uj of the data (a,/), compute moments w.r.t.
üj of u solution to (1.1) (most relevant are first and second order moments).
Several approaches have been proposed in the literature to treat these two problems numer¬
ically, of which we mention and briefly discuss here the Monte Carlo (MC), the perturbation
and the stochastic Galerkin (SG) methods.
The MC simulation is probably the simplest and most natural approach, based on samplingthe coefficients and solving deterministic problems by a standard (e.g. Galerkin) finite el¬
ement method (see, e.g. [BTZ04]). The computation is fully parallelizable (therefore easy
to code), but requires knowledge of data distribution everywhere in the physical domain,
1.1. PROBLEM FORMULATION 3
a very strong assumption. Besides, the method is in general considered to converge slowlydue to the statistical error generated by sampling, although the overall complexity can be
easily shown to behave polynomially in inverse accuracy (see Definition 1.1).
The perturbation approach is widely used in engineering applications (sec [KH92] and ref¬
erences therein). There are several versions, of which we mention here only the Neumann
expansion (see [Kel64]), the first order second moment (FOSM) (see [DW81]), and the
Wiener/generalized polynomial chaos (W/gPC) (see [Wie38], [SchOO], [XK02]) methods.
In one of its mathematically rigorous forms the perturbation method consists in invert¬
ing the operator with random coefficients as a Neumann series around its deterministic
mean and solving then sequentially a large number of deterministic problems with differ¬
ent source terms. Its good performance has been demonstrated in practice (at least for
small fluctuations), but only superalgebraic complexity estimates have been obtained in
theory (see e.g. [BC02]). In this work we also adopt the Neumann perturbation approach
and, after carrying out a careful error analysis, we propose an improved version, involvingalso sparse grids and wavelet techniques, and leading, under additional but reasonable reg¬
ularity assumption on the fluctuation, to moment computation algorithms of subalgebraic,
nearly optimal complexity (see Definition 1.1 below). Note that the fluctuation regularity
assumption allows the control of the data truncation error and acts here therefore as a
substitute for the standard moment closure hypothesis.
The parametrization of uncertainty is one of the key points in the numerical treatment of
problems with stochastic data. The number of parameters used to describe the random
fields has in general a huge impact on the algorithm complexity. A random field expansion
of Karhunen-Löeve type separating the deterministic and stochastic variables at an opti¬
mal approximation rate (see e.g. [Loè77], [Loè78]) has become in recent years a standard
procedure to transform the original stochastic problem into a parametric deterministic one,
where the parameter involved belongs to a high dimensional space.
The SG method interprets then the transformed problem variationally in both the param¬
eter and the physical variable and uses a standard (Galerkin) FEM to produce a numerical
solution (note the need for numerical integration schemes in high dimensional domains).Backward substitution gives the numerical solution to the original stochastic problem and
postprocessing is required to obtain the desired statistical information. Just as in the case
of a MC simulation, very detailed information on the data distribution is needed, and the
finite elements in use today to treat the transformed problem can only ensure superalge¬
braic complexity raten (see e.g. [BTZ04], [FST05], [MKÛ5]).An interesting and promising alternative to the SG method would be to essentially ap¬
ply the same complexity reduction method used here (see section 3.4.2) in the context of
reconstruction techniques (collocation, interpolation) for the solution of the transformed
4 CHAPTER 1. INTRODUCTION
problem, by taking into account its analytic parameter dependence. This leads to results
comparable with those presented in this work - nearly optimal complexity under qualita¬
tively similar assumptions.
Definition 1.1. We say that an algorithm for the computation of M.k{u) with u solution
to (1.1) has nearly optimal complexity if, assuming that the deterministic diffusion problem
in D with data given by the expectations of a and / can be solved with an accuracy e using
0(e~/3) operations and memory (for some ß > 0, asymptotically ase-> 0), the algorithm
requires at most 0{e~^~°^>) operations and memory for an ^-accurate computation of
Mk(u).
In other words, an algorithm for the computation of Mk(u) (where u solves a stochastic
model problem) is nearly optimal if the computational effort required is equivalent, asymp¬
totically as the prescribed accuracy e goes to 0, to the effort needed to approximately solve
one deterministic model problem. For example, if the deterministic model problem can be
solved in linear complexity (/? — 1), e.g. by a multigrid method, then a nearly optimal
algorithm for Mk{u) has (nearly) linear complexity, too.
Example 1.2. An algorithm for the computation of M}(u) with u solution to (1.1), which
amounts to solving e~°(1) deterministic diffusion problems in D for an accuracy of e — 0
and with data as regular as the expectations of a, f, has nearly optimal complexity.
We mention that the mean field computation algorithm we are going to present is of the
type mentioned in the example above and will be developed around some very classical
ideas of the perturbation approach. Besides, it is almost fully parallelizable and its imple¬
mentation can be based on any préexistent FE solver of the deterministic model (diffusion)
problem in D (to be called with many different r.h.s.'s).
1.2 Main Results
We give in the following a summary of the main results contained in this work. We empha¬
sise that the main novelty consists in improved perturbation algorithms of nearly optimal
complexity (thus more efficient than MC, standard perturbation or SG methods, under
similar assumptions).
The case of a deterministic diffusion coefficient a in (1.1) (that is, depending only on the
physical variable x, a(x,u) = a(x) \fx G D) with a stochastic source term / (that is,
depending on both the physical and the stochastic variables x and ui respectively) is a key
step in the development of solution algorithms for (1.1) using a perturbation approach.
1.2. MAIN RESULTS 5
This analysis has been carried out in [ST03a], [ST03b]. We briefly review next the main
results obtained in [ST03b].
Defining the solution operator to the model problem, Aa :— —div(aV) : Hq(D) —> H~1(D),and
Aa ® Aa ® • • ® Aa : Hl(Dk) -> H~\Dk)
where
Hs{Dk) := HS(D)®HS(D)®---®HS(D) VsG[-l,oo[
Ht(Dk) := Hl{D)®Hl{D)®---®Hl{D)
are the anisotropic Sobolev spaces of mixed highest derivatives (e.g. H1(D2) ~ {w(x,y) G
L2(D x L>) | Vxv, Vyv, VxVyv E £2}), a deterministic equation for Mk(u) in terms of
Mk(f) is
(Aa ® Aa ® • • • ® Aa)A**(u) - Mk{f) in H~1{Dk).
This deterministic moment problem for A^fc(w) in terms of Mk{f) shows that we can trade
randomness in the original pde at the expense of high dimension in the problem domain.
Employing a sparse tensor product discretization and a wavelet preconditioning procedure,
the corresponding discrete moment problem is shown to lead to a well-conditioned linear
system. Further, a solution algorithm for the resulting system (given in general by a fully
populated matrix, but having a special blockwise tensor product structure) is developed
to ensure log-linear (in N, number of dof's in D) number of operations and memory re¬
quirements. The high dimensional fc-th moment problem is therefore solvable numerically
in work essentially proportional to that of the mean field equation, i.e. for k — 1, up to
logarithmic terms. In other words, nearly optimal complexity is achieved by the algorithm
presented in [ST03b] in the case of a deterministic diffusion coefficient.
In chapter 3 we take up the case of a stochastic coefficient. The main difficulty is now that,
in contrast to the case of a deterministic coefficient, there is no obvious pde with Mk(u) as
a solution. We start with the simplest such problem, that of computing Ml{u), the mean
field (or expectation) of u solution to (1.1). The idea is, as in the widely used perturbation
approach, to use a representation (decomposition)
a(x, lü) — e(x) + r(x, uj) — 'deterministic expectation' + 'random fluctuation'
where the fluctuation is assumed to be smaller than the expectation in the sense that
\\r(x> ')\\L°°(m0 < sup
" v
/",(U)
< 1.
XD e(x)
Choosing e appropriately, it is easily seen that this condition is (almost) equivalent to the
ellipticity of Aa.
6 CHAPTER 1. INTRODUCTION
We further assume the fluctuation r to be analytic in the physical domain D C [— 1, l]d,but emphasize that this condition can be relaxed to only piecewise finite Sobolev regularity,
r<=A([-l,l]d,L°°(Sl)).
The analyticity assumption, which is satisfied e.g. if the covariance of a is Gaussian,
ensures the existence of a fluctuation expansion as a fast (quasi-exponential) convergent
scries separating the deterministic and stochastic variables,
r(x,u)=J2<i>(x)xm(u)> \\<t>®Xm\\L~(Dxfl)<e-°(ml/dî VmGN+. (1.3)
Such a representation is clearly not unique - one can choose e.g. the Karhuneri-Loeve
(KL) expansion of r, or (4>m)me^+ to be the Legendre polynomials in [— 1, l]d. In both
cases the deterministic part (4>m)meN+ of the expansion (1.3) is available (by an additional
eigenvalue computation for the integral operator associated to the kernel M2(r), in the
KL case) and statistical information on the stochastic part (Xm)mepj+ follows by testing r
against (0m)m6N+)
Xm(u)) = / r(x, u)c/>m(x) dP(uj) Vm G N+.
n
Choosing a truncation order M for the fluctuation series,
M
rjif := 53 4>m{x)Xm(uj)m=l
and a FE discretization level L (corresponding to the FE space VL of dimension NL), we
define recursively, with Ae — —div(eV), the sequence (vj^jm by
Aev0Jj(-,uj) = f(-,u)
in VI P-a.e. we!],
Aei>j,i(-,o;) = div(rM(',a;)^_i,L(-,w)) Vj G N+
The Neumann series Y1jçnvj,l *s then easily shown to converge in Hq(D), uniformly in
to, like a geometric series. Truncating it a level n G N and assuming the data e, f to be
sufficiently regular in D as to ensure a FE convergence rate $(Nl) uniformly in lo G ft, we
have
H^V) - E-**W)lku» £ e-°(Mi/d) + *(Nl) + e~°M. (1.4)3=0
In order to balance the contributions of the three discretization steps taken so far (trun¬cation of fluctuation series, FE in D, truncation of Neumann series) and to achieve an
accuracy of order s > 0 we choose
M~|log£|d, L s.t. $(NL) < e, n~\\oge\. (1.5)
1.2. MAIN RESULTS 7
However, the computation of Ml(v]iL) for 0 < j < n in (1.4) is not an obvious task. We
show that Ail(vjtL) can be computed exactly using any deterministic solver in D at FE
discretization level L, starting from the mixed moment
Mb'l){rM,f)(K,x) ~ I rM(xi,Lj)rM(x2,uj) • -rM(xj,uj)f(x,Lü) dP(uj) xG D3,x G D
a
via the following composition of j + 1 bounded operators defined in terms of the div, V
operators in D, the trace operator on the diagonal set in D x D and the FE solution
operator A~^ of the deterministic diffusion problem in D with coefficient e, at FE level L,
mA(D*,H-\D)) ld^®A<\ A^,VL) - A{ÎP-\VL) - •••
MM(TM,f)L~iD3) e,L> M^(rM,v0Jj) - M^(rM,Vl,L)
... _> Vl
- M\v3,L)
The alternative representation of the (j, 1) mixed moment of (rM,f),
M^\rM,f)= J2 | / Xmi (u)--- Xmj (u)f(x, u) dP(uj) ) 0mi ® • ® 0mj. (1.6)
^»a 02v
1/ £L°°(D3)G P"1^)
is processed by the diagram (f) and transformed into M1(vjtL). The mean field computa¬
tion algorithm consists therefore in running (f) for all 0 < j < n and with (1.6) as input.
The total cost is shown to be essentially as high as the computational effort needed to
run the first step in (f) for j — n (definition of the following j operators in (j) thus omit¬
ted). Taking into account the mixed moment symmetry w.r.t. permutations of the index
set (mi,m2,... ,m3), we deduce that the algorithm requires solving as many deterministic
diffusion problems in D at FE discretization level L as there are terms in the alternative
representation (1.6) of the mixed moment of (r^, /), that is at least
Mnjn\ ~ £-°(1)|log£|°(((*-1)|log£|) deterministic problems in D with accuracy e.
This superalgebraic (for d > 1) overall complexity (which renders the algorithm rather
inefficient, compared to a standard Monte Carlo simulation of quadratic complexity, see
discussion below) can be substantially reduced by noting that, due to (1.4), (1.5), not exact
but only approximate, £-accurate knowledge of M^iy^i) for 0 < j < n is required. Based
on the fast decay (1.3) of the fluctuation series, we show that many terms in the alternative
representation (1.6) of the mixed moment of (rM, f) can be discarded without affecting the
0(e) computation accuracy. Using the resulting sparsified mixed moment representation
8 CHAPTER 1. INTRODUCTION
as input in (f), the algorithm complexity is shown to become nearly optimal, consisting in
solving
~ | log£|°(|log£l+1 ) < e~°W deterministic problems in D with accuracy e. (1.7)
We conclude the presentation of the mean field computation algorithm by a comparison
with the Monte Carlo method, in terms of overall complexity. Note first that sampling
a, f over a finite sample set fto requires knowledge of the data distribution everywhere in
D. Averaging the discrete (FE level L) solutions (ui(cj))ueQ0 of the sampled diffusion
problems we obtain the accuracy estimate
It follows that the sampling can not distinguish between a rough and a smooth (in the
physical variable) fluctuation, and that the Monte Carlo method amounts in this case to
solving
~ e~2 deterministic problems in D with accuracy e, (1.8)
which is the desired quadratic complexity estimate.
Summarizing, in the first part of chapter 3 we propose a perturbation mean field compu¬
tation algorithm and prove that additional information on the data (smoothness in the
physical variable) allows a nearly optimal complexity estimate, asymptotically as e —> 0.
The algorithm is thus shown to be less expensive than the Monte Carlo simulation ((1.7)vs. (1.8)) and all other known algorithms (SG, W/gPC, FOSM etc.).In the second part of chapter 3 we combine the perturbation algorithm presented above
for the mean field computation with the sparse grid solution algorithm for the higher or¬
der moment problem discussed in chapter 2. The result is a new algorithm also of nearly
optimal complexity for the computation of Mk(u) for k > 2.
1.3 Notations and Definitions
We conclude this introductory part by setting up basic notations and definitions that will
be constantly used throughout this work. We consider a bounded open set D C Md (the
physical domain) with Lipschitz boundary F = dD, we denote by A the Lebesgue com¬
pletion of the Borelian er-algebra on D, and use the standard notation A for the Lebesgue
measure on D. We model the uncertainty in the data through a probability space which
we denote by (ft, £,P), Equipped with the product a-algebra A x S, the space D x ft
becomes a a-algebra.
Definition 1.3. If n G N+, a measurable function g : D x ft — W1 is called a random
vector field on D. If n — 1, g is simply a random field on D.
1.3. NOTATIONS AND DEFINITIONS 9
In the following, mathematical objects (sets, functions, spaces, etc.) defined in terms of
the physical domain D will be referred to as deterministic, whereas those depending on ft
will be termed stochastic.
To formulate a minimal regularity requirement on the random fields we define the Bochner
spaces of vector-valued measurable functions.
Definition 1.4. For an arbitrary Banach space (P, || • \\B), we define the space Lgtep(ft, P)of all B-valued E-measurable step-functions on ft by
£steP(^> B) '= {9 Œ -> B I 3 (fti)ieN+ C S partition of ft, (k)ieN+ C B s.t. g = ^ XnM,
where xn* denotes for i G N+ the indicator function of the set ft» C ft.
The Bochner space of B-valued ^-measurable functions on ft is then given by
L°(ft, B) := {g : ft - B | 3 (gn)n^+ C P°ep(ft, P) s.t. gn > g P-a.e. in ft}.n—too
Clearly then, if (P, || • ||) is a space of measurable (vector valued) functions on D, the
corresponding Bochner space consists of random (vector) fields.
The notion of positivity for a stochastic diffusion coefficient reads
Definition 1.5. A random field a G L°(ft, L°°(D)) is called strictly positive if a(-) is P-a.e.
strictly positive, i.e. there exists AT G S such that P(N) = 0 and
Vw G ft \ N 3a_iW, a+fLJ G R+ s.t. a_,w < a(cu) < a+iU1 A-a.e. in D. (1.9)
The existence of a unique stochastic solution to an elliptic problem with stochastic data
follows then from the well-posedness of the corresponding deterministic problem (see also
[Bab61]). In other words, the well-posedness is preserved by tensoring with measurable
functions.
Proposition 1.6. If a e P0(ft,L°°(r;)) is strictly positive and f G L°(ft, H~1(D)), then
there exists a unique u G L°(Ü,H^(D)) such that
-div(o(w)V«(w)) =/(w) inir1(D) P-a.e. in ft. (1.10)
Proof. Since the uniqueness part follows immediately from the well-posedness of the de¬
terministic diffusion problem ((1.10) pointwise in ft), we only check here the existence of
a (measurable) solution to (1.10). To this end, we consider two sequences (an)n>i) (fn)n>i
as in Definition 1.4, as well as the corresponding representations an — X)»eN Xnnian,i
and fn — 5jjeN+Xn»,</n,i respectively (w.l.o.g. we assume them subordinate to the same
partition of ft). With the notations in (1.9) we set for any n G N+
In '•— {i G N+ I a_ u/2 < an^i < 2a+ilil A-a.e. in D for some u G ftn,i}. (1.11)
10 CHAPTER 1. INTRODUCTION
From (1.9) and the P-a.e. convergence of the sequence (an)„e^+ to a we deduce that 2n / 0
for n large enough. Clearly, for any i G In the diffusion problem in D with data (on>i, fnti)has a unique solution un^ Pi\(D). We then set
un~ E X<W
• un,i + Xn\uieInnnii• 0 G Ls°tep(ft, H^D)) (1.12)
and prove next that (un)n£^+ converges P-a.e. in Hq(D). The limit, denoted by u, is in
view of Definition 1.4 an element of L°(ft, H^(D)) and will be seen to satisfy (1.10).
Let A' G S be a null set such that the pointwise convergence of the sequences (an)n>!,
(fn)n>\ as well as (1.9) hold on ft \ N. Fixing an arbitrary to G ft \ N, it follows that there
exists n^ > 1 such that for any n > nw there exists i^ G Xn with the property u G ft^,^.
This implies for any n > n^ that un(oj) — Un^ solves the deterministic diffusion problem
in D with data (aniiul, fn,^)- The ellipticity condition (1.11), uniform in n > n^, of all these
problems ensures via a standard Strang estimate the Hq(D) convergence of un(u) to u(ui)
solution to the deterministic diffusion problem in D with data (a(ui), f (uj)), as n — oo,
which concludes the proof. D
Remark 1.7. Under the assumptions in Proposition 1.6, the Doob-Dynkin Lemma (see
e.g. [Fed69]) ensures the existence of a Borel measurable mapping
h : Raïï^L°°(D) x H~1(D) - P0X(P>) (1.13)
such that
u(uj) = h(a(u),f(uj)) P-a.e. in ft. (1.14)
Troughout this work we assume that problem (1.10) satisfies the ellipticity condition (1.9)
uniformly in ft, that is it holds
Assumption 1.8. There exist a_,a+ G M+ such that
0 < a_ < a(u) < a+ A-a.e. in D, P-a.e. in ft. (1-15)
Note that for numerous engineering applications assumption (1.15) is satisfied (see e.g.
[GS91]). However, (1.15) is clearly less general than (1.9) and does not cover e.g. the
case of a log-normal diffusion coefficient, a — exp(F) with Y a gaussian random field
(for details and applications of log-normally distributed random fields in groundwater and
polution monitoring problems see [Gil87]).
Chapter 2
Stochastic Source Term
Throughout this section the diffusion coefficient is assumed to be deterministic, that is, a
depends only on the physical variable x G D. (1.15) becomes in this case
0 < a_ < a(x) < a+ < co A-a.e. in D. (2.1)
The source term / of (1.10) is a random field with finite first (or higher) order moments
fLk(n,H-\D)) &GN+. (2.2)
Under these assumptions we study the existence and computability of second and higher
order moments of u solution to (1.10), if the corresponding moments of the data / are
known.
2.1 Generalities
We begin with an introduction of the appropriate functional spaces which will help us
define the moments associated to a random field. The standard reference for the material
we briefly review in the following is [Yos95].
Definition 2.1. If k > 1 is an integer and H is an arbitrary separable Hilbcrt space, we
define the space of Lk, H-valued Bochner integrable functions (see e.g. [Yos95]) on ft by
Lk(Q; H) := i f : ft - H | /measurable, f \\f(cü)\\kH dP(w) < oo > / ~
ii/iiW) := j \\m\\kH dp(u),
where we use the same notation for a P-a.e. equivalence (denoted by ~) class and one of
its members.
A useful characterisation of these spaces follows from
11
12 CHAPTER 2. STOCHASTIC SOURCE TERM
Theorem 2.2 (Bochner). / G Lk(Û; II) if and only if there exists a sequence of H-valued
step functions (fj)jeN such that
fj-*fP-a.e.onSl and / \\fj - f\\kH -» 0, as j - oo. (2.3)
n
The case k = 1 is particularly important, since the elements of L1(ft; II) can be integrated.
Theorem 2.3. If f Ll(Çt; H), then the vector-valued integral
l f(uj)dP(u)) GP (2.4)
o
is well-defined by means of a sequence of II-valued step functions (fj)jen satisfying (2.3)
for k — 1, by
[ f(uj) dP(u) := lim / fJu) dP(u) G H. (2.5)J j-HX Jü Ü
The role of the abstract Hilbert space H will be played in the context of the moments
associated to random fields by the anisotropic Sobolev spaces.
Definition 2.4. If k G N* and v = (vi,v2,..., Vk) G Rk, we denote by
k
Pv(P)fc):=(g)P^(P) (2.6)3=1
the anisotropic Sobolev space of order v on Dk. Similarly we define the homogeneous
anisotropic Sobolev space H%(Dk) (Convention: for any s G R, s :— (s, s,..., s) G Rk).
Definition 2.5. If k > 1 is an integer, s G R and the random field f on D satisfies
/ePfc(ft,Ps(D)),then
/() ® /(•) <g> • • • ® /() G L\Q, Hs(Dk)), (2.7)s
v/
k times
so that the k-th order moment M.k(f) of the random field f (sometimes called k-point
correlation) is defined via Theorem 2.3 by
Mk(f) := J f{u) <g> /H ® • • • ® /(w) dP(cu) G Hs(Dk). (2.8)
n
Remark 2.6. Under the assumptions in Definition 2.5 it is easy to see that if s > 0, then
the A;-th order moment of / given by (2.8) satisfies
Mk(f)(xx,... ,xk) = J f(Xl, w) f(x2, Lu)--- f(xk, lu) dP(u), (2.9)
for any (xi, x2,..., xk) G Dk.
2.2. THE MOMENT EQUATION 13
2.2 The Moment Equation
Here we derive a deterministic PDE for the k-th order moment Mk(u), establish its well-
posedness in anisotropic Sobolev scale in Dk, and study the solution regularity.
Theorem 2.7. If f e Pfc(ft, P"1^)), then u solution to (1.10) satisfies
ueLk(Çl,H*(D)), (2.10)
so that Mk(u) G H$(Dk).
Proof. The conclusion follows at once from Definition 2.5 and the well-posedness of the
deterministic diffusion problem in D,
IMw)Hffi(D) <ca\\f(uj)\\h-hd) P-a.e. w G ft, (2.11)
where the constant ca > 0 depends only on the diffusion coefficient a. G
We prove next that M.k(u) satisfies the deterministic equation in Dk obtained by tensorising
(1.10) with itself k times. To this end we introduce the following notations,
k
a®,k ._ ǧa£L°"(Dk)i=i
k
V®-fc := (g)V GBftf1 (I?*), ®J=1£2(£>)**)3=1
where we denote by B(X, Y) the space of bounded linear operators between the Hilbert
spaces X and Y, with B(X) := B(X,X).
Theorem 2.8. Mk(u) is the unique solution in Hl(ük) of the variational problem
(a®>kV®>kMk(u),V®'kM)LHD)M = {Mk(f),M)H-i{Dk)^{Dk) \JM G Hl(Dk). (2.12)
Proof. The existence and uniqueness of a solution to (2.12) are easily proved using the Lax-
Milgram Lemma in appropriate anisotropic Sobolev spaces, noting that tensor products
of bounded positive homeomorphisms between Hilbert spaces induce homeomorphisms
between corresponding tensor products spaces.
Now, since / G Lk(ü, H~l(D)), there exists a sequence (/„)„6n, (hn)nçN of P-valued step
functions on ft satisfying (2.3) with H :— H~1(D). By dominated convergence and (2.5),
Urn Ma(fn) - Ma(f) in H~1(Dk). (2.13)71—K»
We write /„ — J2qejn fn^-nq,„ where ln,,n denotes the indicator function of ft9i„ G S and,
f% G H~l(D) for any q,n (for each n the family (ftgiTs)(7ejn is a partition of ft). To the
14 CHAPTER 2. STOCHASTIC SOURCE TERM
deterministic data /* we associate the solution u^ G H^(D) of the corresponding diffusion
problem,
(aVulVv)LHD)ä = (K,v)H-HD)iHiiD) Vt; G H^D), (2.14)
and set un :~ J2qejn un^q,n- The continuous dependence (2.11) on the data of the solution
of the deterministic problem and (2.3) for / imply
lim un — u P-a.e. on ft, lim \\\u(u) — un(u))\\k„i(TT.dP(u]) = Q. (2.15)n—>oo n—*oo J -"ov^
From Definition 2.5, (2.15) and (2.5) we deduce that
lim Mh(un) = Mk(u) in P1^). (2.16)71—t-OO
Choosing in (2.14) & different deterministic test functions v\,v%,..., vk, taking the product
of the resulting k equations and summing over q with weights P(ft9>n), we obtain that
Mk(un) solves the deterministic problem
(a®'kV^kMk(un), V®'kM)LHD)ak = {Ma(fn),M)H-1(Dk)Mèm WW G H^(Dk) (2.17)
(use here that tensor products of total sets in Hilbert spaces are total in product spaces).The desired equation for Mk(u) follows then from (2.13) and (2.16) by letting n — oo in
(2.17). D
The regularity of Mk(u) follows naturally from that of the data Ma(f), and the result,
as well as its proof, is analogous to the one presented in [ST03a] for k — 2 (bounded
invertibility of operators acting in Hilbert spaces is preserved by tensorization). We only
state it, as follows. Recall that a deterministic diffusion problem in D is said to satisfy the
shift theorem at order s > 0 if H~1+s regularity of the data implies H1+s regularity of the
solution.
Theorem 2.9. // the deterministic diffusion problem in D with coefficient a satisfies the
shift theorem at order s, then (2.12) satisfies a shift theorem at order s too, in the sense
that ifMk(f) G H~1+s(Dk) implies Mk(u) G H1+s(Dk).
Remark 2.10. In the case of a polygon or polyhedron D, a shift theorem at order s >
0 holds in weighted spaces Hlß+8'2(D) (see [BG88]). The proof of Theorem 2.9 can be
correspondingly adapted to deduce then a shift theorem for the moment equation (2.12)in an anisotropic weighted Sobolev scale.
2.3 Discretization with Sparse FEM
We investigate now the numerical approximation of M,k(u), using the Finite Element
Method for the deterministic elliptic problem (2.12). We start by defining general FE
spaces, so let
V0 C Vi C ... C VL C ... C Hl{D) (2.18)
2.3. DISCRETIZATION WITH SPARSE FEM 15
be a dense hierarchical sequence of finite dimensional subspaces of Pq(P), with Nl :=
dim(Vi) < oo for all L.
Assumption 2.11. The following approximation property holds,
min ||W - v\\HuD) < $(NL, s)\\u\\H.+i(D), Vu G Hs+l(D) n Hl0(D), (2.19)
where Q(N, s) —» 0 for s > 0 as N —» oo denotes the convergence rate of the given FE
space sequence.
For regular solutions the usual FE spaces based on quasiuniform, shape regular meshes are
suitable.
Example 2.12. If T — {7i}LeN is a nested sequence of regular triangulations of D of
meshwidth hi = h^-i/l, we choose Vl to be the space of all continuous piecewise polyno¬
mials of degree p on 72 vanishing on dD. Then Assumption 2.11 holds with
NL~cT2d-L, $(N,s) = cTN-s, 6:=mm{p,s}/d. (2.20)
Since the A;-th order moment M.k(u) of u solves the elliptic problem (2.12) on Dk, we
construct FE spaces in Dk starting from the FE spaces {VlJ^o in D. Full tensor product
spaces,k
Vi:=(g)Vi3=1
are natural candidates, but, due to efficiency reasons, we use here the sparse tensor product
spaces that are defined by (see [Zen91], [BG99])
VL := Span I (g) Vi} | 0 < h + i2 + ... + ik < L I. (2.21)
Another description of the sparse tensor space (2.21) can be given in terms of the hierarchic
excess Wi at level L > 0 of the scale {Vl}l>q, which is chosen to be an arbitrary algebraic
complement of Vl-i in Vl (set V_i :— {0}). Since then
Vl = © Wu (2.22)0<i<L
it is easily seen that VL admits the direct algebraic (not necessarily L2 or H1 orthogonal!)
decomposition
VL:= 0 <g)WtjC 0 <g)W,=(g)VL. (2.23)0<u+i2+...+ik<L j'=l Q<h,i2,...,ik<L j=l j=l
The discretized version of (2.12) using the sparse FE space VL reads
{a®'kV®>kMkL(u\ V®'kML)L2{D)M = (Mk(f)ML)w^),Him VML G VL, (2.24)
16 CHAPTER 2. STOCHASTIC SOURCE TERM
where we denote by MkL(u) G Vl the discrete solution of (2.12). The approximation
property (2.19) allows an estimate of the discretization error in terms of the functional <E>,
as follows.
Proposition 2.13. If M.L(u) is the solution to (2.24), L > k — 1 and the approximation
property (2.19) holds, then
k
\\Mk(u)-MkL(u)\\2H1(Dk)<ca,k-J2c^^- E ll-^HIlWw (2-25)3=1 JCO, ,fc}
Card(J)^j
where ej G {0, l}k,ej(j) = 1 iff j G J and
3-i
cO'.*) = E E (HNh,s).^(Nl2,s)--^(Nlm,s).HN0,s))2+m=l li+...+lm=L~m+l
£ $(Nll,s)2.Q>(Nh,s)2---<S>(Nh,s)2 (2.26)h+...+l3=L-j+l
Besides, the constant caik depends exponentially on k, that is, ca,k ~ ct-
Proof. The estimate (2.25) follows using the quasioptimality of the FE solution to (2.24),
the approximation property (2.19) and the description (2.23) of the sparse tensor space with
Wl defined as the orthogonal complement of Vl-\ in Vj, w.r.t the standard scalar product
(•,) in Hq(D). More precisely, we consider the following orthogonal decomposition in
P^o(P'fc) equipped with the tensor product Hilbert structure induced by <S)j=i(v) (hereand in the following orthogonality in H^(Dk) is to be understood w.r.t. this Hilbert
structure).k
Mk(u) - PÈL(M\u)) = Y. ®PLMk(u) (2.27)
at>0,l<t<k
where P^ denotes the orthogonal projection on Wa w.r.t. (-,-), acting in the i-th dimension
of Dk. As the notation suggests, PyL denotes the Po(Pfc)-orthogonal projection on Vl-
Further we consider Q%a to be the projection on Va acting in the i-th direction of Dk and
note that the sum in the decomposition (2.27) is Po(£>fe)-orthogonal, since the excesses
(M/l)lsn are pairwise Po(P)-orthogonal. We rewrite the r.h.s. of (2.27), pointing out
those dimensions 1 < j < k for which ct3 — 0 (coarsest approximation),
E (®^W(«) = E E E (®^,®^W(«),£L«,>i+i \*=i / p=i jc{i,2 ,k} £]eja]>L+i \jeJ 3%j j
a,>0CardJ=p aj>!
\ /
2.4. ITERATIVE SOLUTION AND COMPLEXITY 17
and we cast the first inner sum of projections for J — {ji,j2, ,jp} in the form
E fé)^:<8)po ) =(id-«(g)(id-prt(g)Poja„>l
+ E p£®(Id - <ft-j <8> (Id - pon) (8) po"l<£ n>3 jéj
2
Ql'a2^1
+ • + E ^^-«-ET-iJ®^ (2-28)
Here the /-th sum in the decomposition (2.28) consists of those terms in the l.h.s. corre¬
sponding to indices ai,a2,...ap > 1 with £)n=i ctn > L + 1 and for which / < p is the
smallest integer with the property J2„=1 an > L + 1.
The conclusion follows from (2.28) using the trivial estimate ||P„|| < ||Id —QL-iH (operatornorm in Pß(Pfe))and the approximation property (2.19). D
Specializing Proposition 2.13 by choosing the FE spaces from Example 2.12 we obtain
Corollary 2.14. For the sparse tensor product FE space based on the construction given
in Example 2.12 and for any L > k — 1 it holds
\\Mk(u)-MkL(u)\\H,{Dk) < c^AlogNLf-^NL'-WM^W^^= caAT,u(^ëNLYk-^2N£s (2.29)
and
dim Vh < CkrQo&Ntf^NL, (2.30)
where s — (s,s,..., s), Ô — minjj), s}/d, and all the constants depend exponentially on k.
Note that the full tensor product space would require 0(NL\) degrees of freedom for a
relative tolerance 0(N£5).
2.4 Iterative Solution and Complexity
Sparse FE spaces discussed in section 2.3 allow an important reduction of the number of
degrees of freedom needed to compute a discrete solution to the moment problem (2.12) up
to a prescribed accuracy. After choosing a basis in the sparse FE space Vl and computing
the corresponding stiffness matrix SL, finding the discrete solution of (2.24) amounts to
solving the linear system
8LMk(u) = Mk(f). (2.31)
18 CHAPTER 2. STOCHASTIC SOURCE TERM
In order to be able to solve (2.31) efficiently using an iterative method (e.g. the conjugate
gradient (CG) method), we must ensure well-conditioning and sparsity of the involved
matrix SL. While the first property can be ensured by a wavelet preconditioning procedure,
the latter (which does not hold, in fact) will be replaced by a proper use of the anistropic
structure of the problem, which will show that computation and storage of SL is not
necessary.
Concerning notations, T will be in the following a family of double indices running in
W x Nd.
Assumption 2.15. There exist a family * — (ipj,i)(j,i)eT C Hq(D) and constants cu c2 > 0
such that any u G 111(D) can be expanded as a convergent series,
« = E ci^j,i G I$(D) (2.32)
and the following stability condition is fulfilled
ci>* E ic^i2-
il E ci^iMii(u) ^ c2,* E ic^i2- (2-33)
Constructions of families \[> satisfying Assumption 2.15 are known for intervals (D =]0,1[),
hypercubes (D —]0, l[d) or polygonal domains (see e.g. [DS99]).
Example 2.16 (hierarchical basis on interval). For D —]0,1[, let <fi be the hat function
on E, piecewise linear, taking values 0,1,0 at 0,1/2,1 and vanishing outside ]0,1[. Setting
? '•= {0'> *) I 0 < j, l<i< 2j} and
^i(x) := 2'j/24>(2jx -Ï + 1) Vx G]0,1[,
the family \& — (ipj,i)(j,%)eF satisfies Assumption 2.15.
Example 2.17 (prewavelets on interval). With D, T and </> as above, we define on R
the function ijj, piecewise linear, taking values (1, -6,10, -6,1) at (1/2,1,3/2, 2,, 5/2) and
vanishing outside ]0,3[. Similarly, ijj1 take (9, —6,1) at (1/2,1,3/2) and V>r assumes values
(1,-6,9) at (1/2,1,3/2). Further, we define the scaling function by ipo,i :— <j> and the
boundary wavelets,
^,i(x) := 2~3^l(23x)
*/)j#(x) - 2~j/2ipT(2jx - 2j + 1) Vx G]0,1[, Vj > 1.
The interior wavelets are given by
ipj,i(x) := 2-j/2ip(2jx -z + 2) Vx G]0,1[, V2 < 2 < 2j - 1J > 2. (2.34)
The family \I> = (Tpj.ufjtier satisfies then Assumption 2.15.
2.4. ITERATIVE SOLUTION AND COMPLEXITY 19
For further examples we refer the reader to [Coh03], [Dah97] and the references therein.
Example 2.18 (wavelets on hypercube). In D =]0, l[d choose T := {(j, i) G Nd x Nd | 0 <
i> 1 < i < 23} (inequalities involving multi-indices are understood componentwise). Ten-
sorizing the family constructed in Example 2.17 by setting
d
1PiAx) = nii>U„U)(.xq) Vl = Wi<«<Jei) (2-35)9=i
we obtain (after rescaling in PX(P>)) a family ^ = (ipj,i)(j,i)eF which satisfies Assumption
2.15 (see e.g. [G095]).
Formally, an increasing FE space sequence in D C Rd can be defined in terms of the family
VL := Span{V^| 0 < [j^ < L} (2.36)
(j is in general a multi-index, so that \j\oo :— maxi<g<d jq).We consider the algebraic complement WL of VL~i in VL given by
WL := Span{^ | tfU = L}, (2.37)
and obtain, via (2.23), the following representation of the sparse tensor space Vl,
k k
VL = Span{^,i := <g) VjM,iM I E U(^)loo < ^}, (2-38)f=l v=\
where \(v) denotes the z^-th line of the k x d matrix j and similarly for i.
The algebraic excess Wl of the sparse tensor scale (Vl)l>o is then given by
k k
WL = Span{V>j,i := <g) ^M)iH | £ |j(i/)|«, = P}, (2.39)
and can be further decomposed as
Wl=@Wi_ with WL = Spaiityj,, | |j(^)U = /„}, (2.40)
|I|=L
where
|/| := I, + ;2 + ... + lk V/ G Nfc.
For further use we collect, for P > 0, the basis functions in the definition (2.37) of WL in
a vector denoted ^l- Similarly, for / G Nfc we consider \E^ to be the vector containing all
basis functions of Wi_ as defined in (2.40).
Concerning the properties of the stiffness matrix SL that are of interest for solving (2.31)
(well-conditioning and sparsity), it holds
20 CHAPTER 2. STOCHASTIC SOURCE TERM
Proposition 2.19. The matrices (Sl)l>o have uniformly bounded condition number
K(SL)<ca!T,*,k VP>0, (2.41)
where the upper bound cati^tk depends exponentially on k.
Proof. From (2.33) it follows that the basis (ipj,i)(j,i)eT gives a homeomorphism of Hilbert
spaces between £2 and Hq(D), or that
« = E °i.^ — \< - E M2 (2-42)
defines an equivalent norm on Hq(D). The same holds then for the basis ijj^\ introduced in
(2.38). It follows that for M := (-Mj,i)j,i G M^ with NL := dim t>L, X := Ej,i^J,i^J,iis an element of Vl and
(SLM,M)^L = (a®>kV®>kM,V®'kM)LHD)<* ~ \\M\\2Hèm ~ J] l^wl2 ^ H^llk-
D
Proposition 2.20. For Examples 2.16, 2.17, 2.18 above as well as for similar wavelet
constructions the matrix SL is not sparse, in the sense that (compare (2.30)j
nnz(SL) > O(Nl) VP > 0. (2.43)
Proof. It is easily seen that the entries of SL corresponding to the indices i,j,i',j' with
j(l)=j'(2)= (P, P,..., L) are in general nonzero, implying the desired lower bound. D
The nonsparsity makes the storage and use of SL rather costly, and shows that ideas be¬
yond multilevel preconditioning are needed for an efficient solution of (2.31). The key
observation is the special structure of the discrete operator (or, equivalently, of SL), which
inherits (blockwise) the tensor product structure of the continuous operator (see (2.12)).A matrix-vector multiplication algorithm of log-linear complexity will be derived in the
following, exploiting the special tensor product structure of the problem. It will be shown
that, in terms of memory, only (blockwise) storage of the matrix SL corresponding to the
case k = 1 is necessary. Relating SL to SL will show that one step of CG algorithm can
be performed by manipulating only sparse blocks of SL. Moreover, an algorithm of log-
linear complexity in Ni for performing the matrix-vector multiplication required by the
CG method will be derived.
Storage of the load vector is necessary too, but, due to (2.30), the corresponding memory
requirement grows only log-linearly in NL, as P — oo.
We start with the derivation of the relation between SL and SL. To this end we denote by
(•, •)„, the scalar product associated with the norm (2.42). (•, -)w is obviously equivalent to
2.4. ITERATIVE SOLUTION AND COMPLEXITY 21
the standard scalar product in Hl(D) and (^j,i){j,i)^r is an orthonormal basis of Hq(D)
equipped with (•, -)w. Let PL and QL be the orthogonal projections in Pq(P) w.r.t. (-, •)„„on VL and WL given by (2.36) and (2.37) respectively. Clearly then,
i=o
Correspondingly, we denote by PL and QL the orthogonal projections on VL and WL (see
(2.38), (2.39)) w.r.t. the scalar product on Hl(Dk) obtained by tensorization of (•, -)w.On account of (2.38), (2.39) we obtain the multilevel decomposition
L
Pl = Y,Ql, (2.44)1=0
as well as the detail decomposition
Ql = E Qb (2-45)
wherek
Ql-=®Qi„ (2-46)j/=i
is the projection on the space Wi_ introduced in (2.40).We further denote by Qk the bilinear form of the moment problem (2.12),
Qk:=Q® Q®...®Q, (2.47)
where
Q(u,v):=(AVu,Vv)L2(D) Vu,v G H^(D). (2.48)
The discrete problem in Vl is then given by the bilinear form
QkL(u,v) := Qk(pLU,PLv) \/u,ve VL C Ht(Uk), (2.49)
or, inserting (2.44) and (2.45) in (2.49), by
L
QkL(u,v)=J2 E Qk(QLu,Qnv) Vu,v G VL. (2.50)U'=0 üeN*
|i|»i,IÜ=i'
Recalling that #/ is the vector containing the basis functions of W\ given in (2.40), we
write
Qi_u = u[ */, (2.51)
with real vector coefficients ui and similarly for v.
Using (2.51) in (2.50), we obtain
L
Ql(«,v)=£ E ul-êLli-v^ (2-52)l,l'=0 i,i^nk
\L\=i,\lL\=i'
22 CHAPTER 2. STOCHASTIC SOURCE TERM
where the matrix Sfo, is given by evaluating the bilinear form on the basis functions,
But, in view of (2.47) and (2.39), we have
k k
§& = Qk(Vi, *r) = ® e(*z„, H) = ® SU> (2-53)
where
S\'y ~ Q&t, *J') V0<l,l'<L (2.54)
are the blocks of the stiffness matrix SL corresponding to the deterministic diffusion prob¬
lem in D with coefficient a (or, equivalently, to the case A; = 1).The representation formulae (2.52) and (2.53) show that
Q£(«,«)=E E «M®SWtW. (2-55)1,1'=0 i,i£eN* \f=l /
111=1,11:1=!'
that is the stiffness matrix SL of the k-th moment problem computed w.r.t. the basis
(2.38) of the FE space Vl has a block structure
^ = \^l,l') LlleN* )"~
\l\=l<LIÜ=('<L
and each block is a tensor product of certain blocks of the stiffness matrix SL.
Moreover, SL is almost sparse, once for the basis (^,i)(j,t)e.F the following local support
assumption holds true. We remark that Examples 2.16, 2.17, 2.18 as well as similar wavelet-
type constructions are in this category.
Assumption 2.21. There exists p G N+ such that for all 1 < i < 23 G Nd and f G Nd,the set supp (ipj,i) n supp (Vyy) has nonempty interior for at most pd - TJ^=1 max(l, 23<i~:>'')values of i!.
Remark 2.22. From Assumption 2.21 it follows by a simple counting argument that
nnz(Sj,) ^ Pd • (min(*>l') + i)*"1 • 2*max^-''> V0 <l,l' < L. (2.56)
To formulate the matrix-vector multiplication algorithm we consider also, for each pair
I — (lv)ï=vlL ~ (0*=i> a reordering (permutation) gijj_ of {1,2,..., k} such that
Q k f k k }
E l"(») + E U») ^ max i E ^ E ï» \ v 1 ^ i ^k- (2-57)
The existence of such a permutation a is easy to prove, by choosing x„ — l„, yu — l'v for
any 1 < v < k in the following result.
2.4. ITERATIVE SOLUTION AND COMPLEXITY 23
Lemma 2.23. If (xv)\<u<k and (yv)i<v<k are two families of positive real numbers, then
there exists a permutation a of the set {1,2,..., k} such that
q k ( k k \
E x»w+ E y<v) -
maxi Exv' E y» \ v * ^ <? < fc- (2-58)
v=l u-q+l \v=l v=l )
Proof. We use induction on k. Since for k = 1 the claim is trivial, we assume that it holds
also for some k > 1. Consider (xv)i<.v<k+\ and (yu)i<u<k+i two families of positive real
numbers and define zv :— xv for 1 < v < k — 1 and Zk :— Xk + Xk+\, as well as tv :— yv for
1 < v < k — 1 and tk :— yk + yk+i- The induction assumption ensures the existence of a
permutation r of {1, 2,..., k} such that
q k ( k k }
E Zr(y)+ E M") ^ max ] EXi" EVv \ 'v : -q -k- (2-59)
We set then a(u) := t(v) for all i/ < r~1(k) and a(z^) :— t(v—1) for all ^ > r-1(A;)+l. Now,
if 2/jt + Xfe+i < Xk + 2/&+1 holds, we set <7(t_1(ä;)) := k, a(T_1(k) + 1) :— k + 1. Otherwise,
that is if ji/fc + xfc+i > xk + yk+l, we define <7(t_1(ä;)) := k + 1 and a(r~1(k) + 1) :~ A;. With
this choice for a one can easily check the inequalities (2.58). D
In order to simplify the exposition of the algorithm we introduce, for any 1 < q < k, an
arbitrary index pair (/, J/_) and a permutation a = <7y/ associated to it in the sense explained
above, the following tensor product matrices
k
Tt^-®U^ (2-6°)i>=i
where
v e{<7(l),<7(2),...,<r(</-l)}
U»:=< SL^ v =<& (2-61)
v £{a(q+l),a(q + 2),...,a(k)}
and Id/,/ denotes for / > 0 the identity matrix of size dimW/. With these notations, each
block in (2.55) can be expressed as a product of matrices of the type introduced in (2.60),
k
OC Siv,vv = Tijs,k ' ^iz,k-i ' ' ' ^i£,v (2.62)i>=i
For later use we remark that (2.56) implies the following sparsity estimate for Ttj, Q.
Remark 2.24. It holds
5-1 k
m<TULq) ^ C?W n&W + l)d • (min(^(9), C(9)) + I)""1 • II (W) + 1)"v—L u~q+l
. 2d-Œï=î/"(")+max^(«)'C(9)}+Ef=<I+i^(1,))i (2.63)
24 CHAPTER 2. STOCHASTIC SOURCE TERM
Proof. The estimate (2.63) follows from the obvious equality
/i-l k
nnz(Thu) = TTdimW/, >
" nnz(S/' ,, ) • TT dimWr ,
Q=l 5=/i+l
the asymptotic estimate dimW/ < cT(l + l)d • 2d' and (2.56). D
Based on the factorization formula (2.62), we derive now the multiplication algorithm of
the matrix S by a vector x.
input : blockwise storage of SL — (ä'/lf)o</,/'<l (sparse), x — (xJh+h+-h<L
output: SLx
l for / satisfying X^„=1 lv < L do
2 initialize (SLx)i :— 0
3 for U_ satisfying £^t=i /[, < P do
4 compute 2// := T^k T^h-i • • ?&i •
^
5 update (SLx)L := (5Lx)^ + j«
6 end
7 end
Algorithm 1: Sparse tensor matrix-vector multiplication
Remark 2.25. The order in the product on line 5: of Algorithm 1 is essential for the
efficiency of the multiplication.
Remark 2.26. Due to (2.60), (2.61), the multiplication of 7}jj, by a vector x is done by
multiplying the blocks of x by Sf v .
Wc estimate next the complexity of Algorithm 1.
Theorem 2.27. Algorithm 1 performs the matrix-vector multiplication x — SLx using at
most 0((log NL)kd+2k~2Nl) floating point operations, and it requires only storage of the
stiffness matrix SL corresponding to the case k = 1 (mean field problem) and of x.
Proof. Due to (2.55), (2.62) we write
k
(SLx)L = £ (g) Stu a* = E T&j, 2&fc-i • • T^ •
xL.V_ v=\ a
U1\<L \H\<L
The multiplication under the summation can be performed using at most
k
#/,,:= £nnz(?^) (2.64)9=1
2.4. ITERATIVE SOLUTION AND COMPLEXITY 25
floating point operations. From (2.63) we obtain that
k
#UL < CTA* E(k(D + V* ' * (^-A) + ^ (min(W C(,)) + I)'"1'?=i
• (Cfa+i) + l)d- ' • (C(fc) + !)d • 2*(Sî'*<">+{'*t«)'W+£î=«+i£<.o).
Using the defining property (2.57) of er = ery/ we deduce that for P > 1,
#U' < or,*.* (max (HI, |f I})"*"1 . 2*"<W ^.
The block (SLx)i can be computed using £^, #/(// operations. Finally, the number of
operations needed to perform x — SLx (collect all blocks (8Lx)i for all V) admits the
upper estimate
I H\L\<L \i±\<L
(I+ k — 1\
J < (/ + l)fc_1 solutions
l G Nfe, we conclude
#flops(x -, SLx) < cw £(/ + l)dk+2k~2 2dl < cw(\ogNL)kd+2k-2NL.
D
Due to Proposition 2.19, the number of steps required by the CG algorithm to compute
the discrete solution up to a prescribed accuracy is bounded once we use the solution at
level P — 1 as initial guess of the solution at level P. Thus it holds
Theorem 2.28. The deterministic problem (2.12) for the k-th order moment Mk(u) G
Hs+1(Dk) n H^(Dk) of the random solution u to (1.10) is solvable by a sparse tensor
product FEM with NL dof's in D and for any L > k — 1 at a cost of
cr,kA^gNL)kd+2k-2NL (2.65)
floating point operations, with a memory
CT,k(l°&NL)k-lNL, (2.66)
for a relative accuracy of
ca,u,TdlogNLyk-lV2NZs, (2.67)
where ö — min{p, s}/d.
Note that, up to logarithmic terms, the estimates (2.65), (2.66), (2.67) are asymptotically
as L —> oo similar to those of the case k = 1 (mean field problem). Further, all constants
involved depend exponentially on k.
26 CHAPTER 2. STOCHASTIC SOURCE TERM
Chapter 3
Stochastic Coefficient
Under the uniform positivity Assumption 1.8 we investigate equation (1.10),
Aa(w)u(w) = f(uj) in H'1(D) P-a.e. u G Q
using a perturbation argument. We describe the stochastic diffusion coefficient a as a
random fluctuation r around a deterministic expectation (mean field) e and separate the
deterministic and stochastic parts in r by a tensorial representation. Under regularity
assumptions on r(x,u>) in the physical variable x (but not on the size of r, apart from
the standard one ensuring the well-posedness of (1.10)) we prove that arbitrary (weighted)moments of the solution u to (1.10) can be computed in essentially the same complexity
as the solution of one deterministic diffusion problem in physical domain P. Besides, the
resulting algorithm requires only a deterministic solver for diffusion problems in D and it
is partially parallelizable.
3.1 Random Fields
We consider a splitting of the diffusion coefficient into a deterministic expectation e and a
random fluctuation r,
a(x,Lü) = e(x) +r(x,u) V(x,w) G D x Ü (3.1)
with a positive e G L°°(D) (not necessarily equal to / a(.,Lü)dP(uj)),
a
0 < e_ < e(x) < e+ < oo Vx G P. (3.2)
It follows that the random fluctuation r G L°°(D x Ü) too, and we require that the
fluctuation be pointwise smaller than the expectation.
Assumption 3.1. For the representation (3.1) it holds
0 < „ := esssup^11^"'-/11.^^ < 1. (3.3)
27
28 CHAPTER 3. STOCHASTIC COEFFICIENT
Remark 3.2. The constant expectation choice
e(x) := (o_ + a+)/2 Vx G D (3.4)
satisfies Assumption 3.1 with v < (a+ — a_)/(a+ + a_) < 1.
The more natural (from a statistical point of view) choice
e(x):= I a(x,u))dP(uj) Vx G D (3.5)
n
satisfies (3.3) if the density function of r(x, •) is even for any x G D (positive and negative
fluctuations occur with equal probabilities).
Concerning the fluctuation term r we formulate also a modelling assumption as well as a
condition of regularity in the physical variable.
Assumption 3.3. The random fluctuation r can be represented in Lœ(D x SI) as a con¬
vergent seriesoo
r = E t ® Xm (3-6)m—1
with known <pm G C°°(D) (deterministic), Xm G P°°(S1) (stochastic).
The representation formula (3.6) describes the tensor product nature of the random field
r, satisfying therefore the separation ansatz in the deterministic and stochastic variables
x G D and wë!] respectively. A standard representation of this type is the Karhunen-
Loève expansion (see e.g. [Loè77], [Loè78]), and a further example is given below (see
Example 3.6).
The regularity of the random field r is quantified by the convergence rate of the series
(3.6).
Assumption 3.4. The random fluctuation r admits a representation (3.6) for which there
exist constants cr, Ci>r, a > 0 such that
\\4>m ® A"m||LOO(SxS2) < c,. exp(-chrma) Vm G N+. (3.7)
The regularity Assumption 3.4 holds for instance for random fields r which are analytic
in the physical variable, with values in L°°(Q). This follows from standard approximation
theory of analytic functions (see e.g. [Dav63]).
Definition 3.5. For an arbitrary Banach space P and an open set U Ç Cn we denote by
A(U, B) the space of P-valued complex analytic functions on U which are also continuous
on the closure of U. Note that A(U, B) equipped with
MU(t/,i?) := sup ||w(x)||ß V22 G A(U,B)xeu
becomes a Banach space.
3.1. RANDOM FIELDS 29
Example 3.6. If D C [-1, l]d and r G A([-l, l]d, P°°(f2)), then a representation (3.6)exists with (0m)meN+ the Legendre polynomials in [—1, l]^ (tensor products of standard
Legendre polynomials in [—1,1], scaled to have P2-norm equal to 1) and
Xm(uj) := / r(x, Lü)4>m(x) dx P-a.e. lü G Ü, Vm G N+. (3.8)
H,x]d
Moreover, (3.7) holds with a — \jd and c^T depending on the size of the analyticity domain
of r in a complex neighbourhood of [—1, l]d.
Remark 3.7. By knowing Xm for m G N+ we understand that any mixed moment of any
finite subset of the random variable family (ATm)m6N+ is (computationally) available. For
instance, for the construction presented in Example 3.6, the analyticity assumption on r
ensures that the integral on the r.h.s. of (3.8) can be estimated using a quadrature rule
(xq,wq)qQ in [-1, l]d. It follows
Xm(-) ^ Er(X9' ')<t)rn(Xq)wq. (3.9)geQ
Given the realisations (a(-,a;))i„eno) an arbitrary k-th order mixed moment of the family
(Xm)mçN+ is then estimated by the mean
fxmi(u;)Xm2(uj)...Xmk(uj)dP(u;) ~ -L J^ fl (E7^'^"»»« ) . (3.10)
Remark 3.8. The following arguments, proofs and algorithms carry over to the case of a
random fluctuation r satisfying Assumption 3.3 only locally on D (i.e. if a representation
formula (3.6) holds locally on a finite covering of D). Moreover, a simple scaling argument
shows that we can assume w.l.o.g. from now on that D C [—1, l]d.
Since computations can handle only finite data sets, we truncate the infinite expansion
(3.6) of r and introduce for any M G NU {oo} (convention: r0 — 0) the M-term truncated
fluctuationM
Tm :=E^"® Xm> aM -=e + rM. (3.11)
The error introduced thereby in the solution of (1.10) is then controlled due to the Strang
Lemma.
Proposition 3.9. 7/ Assumption 3.4 holds, then there exists MT G N such that for any
M > Mr the problem of finding uM G L°(Q, H^(D)) such that
AaMMwMH-/(w) inH~l(D) P-a.e. LueÜ (3.12)
is well-posed and
\\u(lo) - um(u)\\hI(d) < ce,rex.p(-chTMa)\\f(cu)\\H-y(D) P-a.e. u G Ü. (3.13)
30 CHAPTER 3. STOCHASTIC COEFFICIENT
Remark 3.10. If (Xrn)mG^+ is an independent family of random variables with vanishing
mean (e.g. for the (3.5) choice of e), then Mr can be chosen equal to 0, due to the fact that
problem (3.12) becomes elliptic uniformly in M G N. This can be easily seen by taking
the conditional expectation of (1.15) w.r.t. the <r-algebra generated by X\, X2,..., Xm-
Throughout this section we will constantly make use of the mixed moments of vector valued
random fields, which are defined as follows.
Definition 3.11. If H is a separable Hilbert space, j G N+ U {oo}, p = (pi,p2, ,Pj) G
Lk(ü, H)3 and n — (nj, n2,..., n3) G W with ||n||i :— 72i + n2 H \-n3 — k, then
Mn(p) := pi(Lü)®--.®p1(u)®...®pj(uj)®---®pj(u;) dP(uj) G II ® ® II (3.14)J "
v' *>
vS
"V
'
nni times n3 times k times
is called the mixed moment of p of order n.
Further notations will be 13 := (1,1,..., 1) and, for m G N^_, Xm :— (Xmi,Xm2,..., Xmj).
j times
3.2 Discretization
Considering the hierarchical FE space sequence (Vl)lêN given by (2.18) and satisfying As¬
sumption 2.11, we formulate for any M G NU {oo} and any FE discretization level P G N
the discrete counterpart of (3.12),
VL : Find uMiL G L°(Ü, VL) such that
ABM(w)UW,L(w) =/(w) in VI P-a.e. ue Ü. (3.15)
From Assumption 3.4, PL is well-posed for M > Mr.
Assuming that the conditions ensuring a shift theorem at level 5 for the diffusion problem in
D with coefficient a(-, lo) and data f(-,oj) are satisfied uniformly in lo G Ü, the discretization
error satisfies the standard estimate.
Assumption 3.12. There exists s > 0 such that a G L°°(Q,CS(D)) (following e.g. from
e G CS(D) and r G A(D, L°°(U))), and dD G Cs+1, so that the deterministic diffusion
problem in D with coefficient o satisfies a shift theorem at level s > 0 (see, e.g. [GT01]).
Proposition 3.13. /// G P0(^,P"1+S(P)) and Assumptions 2.11, 3.4, 3.12 hold, then
\\um(lü) - uMiL(uj)\\hi(D) < caJ\f(uj)\\H-i+siD)$(NL, s) P-a.e. u G Ü, (3.16)
and for any M > Mr, P G N. Here the functional $ quantifies the approximation property
of the scale (VL)LeN (see also Example 2.12).
3.3. NEUMANN SERIES DECOMPSITION 31
3.3 Neumann Series Decompsition
Considering / G P°(fi, H~1(D)) and the hierarchical FE space sequence (2.18), we formu¬
late for any M G N U {oo}, M > Mr and any FE discretization level P G N the following
sequence of well-posed problems,
VL : Find um,l e P°(fi, VL) such that
&aM(u)UM,L(u) = f(u) in VI P-a.e. lo G fi, (3.17)
V0il Find v0tL G P°(fi, VL) such that
AeVo,LM = /H in VI P-a.e. w G fi, (3.18)
and for jf G N+
^,l : Find v3tL G P°(fi, KL) such that
Aevj<L(uj) = diY(rM(oj)Vv3-ltL(uj)) in Vt P-a.e. u G fi. (3.19)
Using basic P1 estimates for the diffusion problem in D we show next the exponential
decay of vjiL, the solution to (3.19), in the energy norm as j —> oo, and the exponential
convergence of the series with general term v3iL to um,l solution to (3.17). To this end, we
introduce, for any e G C,
aw.« := e + erm G P°°(P> x fi) (3.20)
and study the diffusion problem with stochastic coefficient üm,<:-
Proposition 3.14. Under Assumptions 3.1, 3.4, there exists 8U > 0 such that the problem
of finding uM,L,t G P°(fi, Vl ® C) with
~Km^)Um,l,z(u) = /H m^ P-a. e.wGfi (3.21)
is well-posed uniformly in M > Mr, P G N and e G S where S denotes the complex strip
S:=[-l-<5„,l + <S,,]xzRcC. (3.22)
Proof. For notational ease we drop u from all notations. The real and imaginary parts
^um,l,(. G Vl and ^Sum,l, Vl should solve P-a.e. a; G fi the system equivalent to (3.21),
f &maMßuM,L* ~ A$aMlßUM,L,e = f m V£ .
23.
\ AaaM£SRuM,£,£ + AsROM)eöuM,L,£ = 0 in V£
Choosing an ONB in Vl (equipped with the Hq(D) norm), (3.23) becomes a linear system
whose matrix A has the form
32 CHAPTER 3. STOCHASTIC COEFFICIENT
with Ai G RNlxNl the real-valued symmetric matrix associated to the bilinear form of
AsßaMt
in Vl, and A2 G RNlxNl the real-valued symmetric matrix associated to the bilinear
form of AoaMe in Vl- Noting that
5RaM,e = e + (3fc)rM (3.25)
is bounded away from zero for e in 5 given by (3.22) with 5V small enough we deduce
that Ai has (positive) spectrum, bounded away from 0 uniformly in lo G fi, P G N and
M >Mr. Since
\ ( °ß ) ' ( °ß ) ) = {Aia,a)RNL + ^l/3'^ (3.26)
we obtain that the spectrum of A is bounded away from 0 uniformly in lo G fi, P G N and
M >Mr. D
Proposition 3.15. Under Assumptions 3.1, 3.4 ,the mapping
S 3 e ^ %,v(w) G H^(D) (3.27)
imï/i Mm,i,£ solution to (3.21) is holomorphic P-a.e. in ft for any M > Mr, P G N, with Mr
as in Proposition 3.9.
Proof. Fixing e0 G S, we show that UM,L,e is holomorphic as a function of e in a neighbour¬
hood of eo by explicitly constructing the solution of (3.21) as a Taylor expansion around
e0. To this end we consider the following sequence of well-posed problems
AaMieoMuM,L,eoH = /(<*>) in VI P-a.e. u) G ft, (3.28)
and for j G N+
A«Af,.0M^,L,eo(w) = div(rM(w)Vvi-i,L,eo(w)) in V£ P-a.e. well, (3.29)
where by convention v0,l^0 :~ uM,L,e0- From (3.29) with v3tL,e as test function we deduce
that, P-a.e. in fi it holds,
||V^,e»||L2(Z)) < ^^^llVvj.^MllLHo) VjGN+. (3.30)
Note that the existence of c£0 > 0 is ensured by Proposition 3.14. From (3.30) we obtain
the exponential decay of |p|,7'||wj),L,eo(u;)llfl'1(z>) with j —> oo, uniformly P-a.e. in fi for p in
a small complex neighbourhood of 0 depending on e0 and ||r||x,»(£,xn). In particular the
series
E^^'oM (3-31)
converges in E\(D), uniformly P-a.e. in fi for p in a small complex neighbourhood of 0.
Moreover, the sum solves the diffusion problem with stochastic coefficient a^tü+p, as one
can easily see by summing (3.28) and (3.29) for j G N+. The value of (3.31) is therefore
equal to UM.L.eo+p^) -P-a.e. in fi, which concludes the proof. D
3.4. DISCRETE MIXED MOMENT PROBLEM 33
The domain of analyticity of the mapping (3.27) contains therefore the closed disc of center
0 and radius 1 in the complex plane, which ensures that the distance between eo — 0 and
the singularities of (3.27) is strictly larger than 1. As a consequence, the Taylor expansion
of (3.27) around eo = 0 converges exponentially in p — 1.
Proposition 3.16. If f L°(ft,H~1+s(D)) and Assumptions 3.1, 3.4 hold, there exists
cli0 > 0 such that
\\vj<L(uj)\\Hi{D)<ca\\f(uj)\\H-i(D)exp(-clAj) P-a.e. lu G fi, (3.32)
and
n
\\um,l(^) - E'^L^Hffo(ß) - cJ/Mlliï-H-D) exp(-CiiQn) P-a.e. lu G fi, (3.33)3=0
and for any M > Mr, j,n,L G N.
Proof. Taylor expansion of (3.27) around eo = 0 gives,
um,l,p(uj) = ^p'v3;l,o(uj) (3.34)
so that
v3,l,o(uj) = j$um,l,z(u)\c=0. (3.35)
Noting that VjtL,o — 'v3\l for any j G N+ and um,l,i = um,l, all we have to check is the
exponential convergence in Pq(P>) of the series on the l.h.s. of (3.34) for p = 1. But this is
equivalent to the exponential decay of Vjtii0 as j —» oo in the Hq(D) norm, which in turn
follows from (3.35) and the Cauchy representation formula
diuM,LMl=» = ^ï] z3+ldz V? G N
r
applied on an arbitrary circle F of center 0 and radius strictly larger than 1, contained in
the strip S. D
3.4 Discrete Mixed Moment Problem
Since our primary aim is the development and the analysis of an algorithm based on the
Neumann expansion (3.18), (3.19) for the computation of the simplest (first/second order)
statistics of it solution to (1.10), let us collect now in the next result the errors introduced by
the discretization steps of the previous sections (modelling, FE discretization and Neumann
truncation errors, see (3.13), (3.16) and (3.33)).
34 CHAPTER 3. STOCHASTIC COEFFICIENT
Proposition 3.17. /// G ^(ft^^+^D)) and Assumptions 3.1, 3.4, 3.12 hold, then
M") ~^2vjiL(uj)\\Hi{D) <
3=0
< ca,s\\f(uj)\\H-i+s(D)(^NL,s) + exp(-chTMa) + exp(-chan)) P-a.e. u G fi, (3.36)
for any M > Mr, n, L G N.
Note that, after averaging (3.36) over lj G fi and balancing the truncation orders M,n
and the discretization level P, the computation of the mean fields M.1(v3,l) for 0 < j < n
produces also an approximation of M.l(u). However, computing the moments M1(vj!l) for
0 < j < n is not an obvious task. Under additional regularity assumptions on the random
fluctuation r, we develop next an algorithm which makes this computation possible in nearly
optimal complexity (compare Definition 1.1). It turns out that this computation has to be
based on knowledge of higher order mixed moments of the pair (tm , /) • Assuming these
are available, we derive a general algorithm for the computation of the mixed moments
associated to the pair (rjh, u) (and not only of A^1(m)). This algorithm will be later used
in the context of second (and higher) order moment computation.
Assumption 3.18. There exists an open complex neighbourhood U Ç Cd of D such that
Assumption 3.4 is satisfied and the decay estimate (3.7) holds in L°°(U x fi) (in particular,
4>m G A(U) for all m G N+).
Remark 3.19. Assumption 3.18 implies r G A(U, P°°(fi)) and the convergence of the
series (3.6) in A(U, L°°(ft)). Reciprocally, if r G A(U', P°°(fi)) with V c U' Q Cd (which
holds e.g. for a diffusion coefficient a with Gaussian covariance), then Assumption 3.18 is
satisfied with (3.6) the Legendre/Karhunen-Loève expansion of r (compare Example 3.6).
Definition 3.20. If A" is a bounded linear operator between two Banach spaces B\ and
P2, and U Ç Cn is an open set, let A(U, Pj) be the Banach space of P^-valucd functions
which are continuous on U and analytic in U (i = 1,2). We define Idu <S> X to be the
bounded linear operator given by
Idu ® X : A(U, Pi) -> A(U, B2), ((ldu®X)h)(z)--X(h(z)) Vz g V.
Remark 3.21. The direct sum of d copies of the trace operator S defined in Lemma 5.1
induces by restriction to functional spaces on U a bounded linear operator denoted for
notational ease again by S,
S : A(U, L2(D)d) - L2(D)d, (ß(u)){x) - u(x, x) A-a.e. x G D. (3.37)
We formulate next for j G N+ arbitrary the problem to which the mixed moment of
(rMi'Vj,L) of order (j, 1) is solution. To this end we introduce the notation
A^ : H~\D) -> Vl (3.38)
3.4. DISCRETE MIXED MOMENT PROBLEM 35
for the bounded FE solution operator of the homogeneous diffusion problem with coefficient
e corresponding to the FE space Vl equipped with the Hq(D) norm.
Further, for any U Ç Cd and any j G N+ we define the product domain
U3:=UxUx-..xUc C3d.s
v'
j times
Definition 3.22. For any j G N+ and L G N we define the operator T3j4 as the following
composition of linear operators,
A(U^ Vl) ^22îZ A(W, L\D)d) Id"~10S> A(W-\ I?(D)d) Id^-l8div1
Id--10diV> A(U3~\ H-\D)) Id"-l8Aä, A{U3-^ yLy (3.39)
Similarly, for any j G N and P G N we define the operator Sj,l given by
Id,,., ® A~L
A(U3, H~\D)) — ^ A(U3, VL). (3.40)
It then holds
Proposition 3.23. The operator T^l defined by (3.39) is bounded uniformly in j, L and
maps M^'^^m^Vj^l) into M^~l,l\rM, Vj'+i.-l) for any f G N.
Similarly, the operator S3tL defined by (3.40) is bounded uniformly in j, P and maps M^(rM, f)
into M^l\rM,VQiL).
Proof. Note first that the mixed moments have been introduced in Definition 3.11. Observe
next that the second operator in (3.39) is bounded uniformly in j due to Lemma 5.1 (see
Appendix) and Definition 3.20, while the others are trivially continuous, uniformly in P.
The mapping property follows by taking the tensor product of Pj'+i,l (see (3.19)) with
I'm(w) ® • • ® rju(w) (j — 1 times) and averaging the resulting equation over lo G fi-
The mapping property of S3tL follows then analogously, by taking the tensor product of V0,l
(see (3.18)) with rM(u) ® ® rM(Lo) (j — 1 times) and averaging the resulting equation
over a; G fi.
Proposition 3.23 shows in particular that the first order moment (expectation) A41(v3>l) =
M^^(rM-,Vjti) follows by successive composition of operators of type (3.39) from the
higher order moment M^'^(rM,v0,L)- This in turn is obtained by mapping M(3'^(rM,f)via SjtL- The discrete problem solved by Jv^^Uj^l) is therefore given by
Corollary 3.24. For any j,f G N it holds
M{3'>l\rM, vJtL) = (Tr+1,L o Tr+2,L o • • • o T3,+3,l o Sr+3tL)(M^+^(rM, /)). (3.41)
In particular, for j' = 0,
M\vjtL) = (TltL oT2iLo..-o Tj,L o S3j)(M^(rM, /)). (3.42)
36 CHAPTER 3. STOCHASTIC COEFFICIENT
Note that (3.41) can be viewed as a recursion of parametric deterministic diffusion problems
in D where after each recursion step taking the trace via S reduces the dimension of the
parameter space by d.
3.4.1 Mixed Moment Algorithm
From Corollary 3.24 and Proposition 3.17 it follows that the computation of the exact
mean field Ml(u) requires knowledge of the mixed moments jM^'^rM, /) f°r j G N. The
more such moments are available, the higher the truncation order of the Neumann series
and the better the expected approximation of Ml(u).
Concerning the mixed moments M.^3'^(rM, f) we have the following representation formula,
and estimates following from Assumption 3.18.
Proposition 3.25. If j G N and r satisfies Assumption 3.18, then for
M^l)(rM, f) = JrM(uj)®...®rM(Lo)®f(Lü) dP(uj) G A(W, H~l(D)) (3.43)
ü j times
it holds
M^(rMJ)= J2 M^^X^f)®^, (3.44)men',
where for any index m = (mi, m2,..., m3) G N+, UniH^ :— max^^ \mk\,
</>m := <t>my <8> (pm2 <S> • • ® <i>mj G A(U3), (3.45)
and (in view of Definition 3.11)
M^(XmJ) = JXmi(Lo)Xm2(to) Xrnj(Lo)f(uj)dP(uj) G H-\D).n
Further, the general term of the expansion (3.44) satisfies
||A^lj+1(^m,/)®0m|U(w,H-»(D)) < cjexp(-ci,r||m||S)||/|Ui(n,H-i(D)) (3-46)
with ||m||a :— (m" + rn2 + • + rap1/".
Equation (3.44) and the definitions of all operators involved in (3.42) enable us to derive
the following algorithms for the computation of the mixed moments of the pairs (tm,v3^l)and (rM,UM,L) respectively, once a solver for the deterministic diffusion problem in D is
available.
3.4. DISCRETE MIXED MOMENT PROBLEM 37
input : • M, j,f, L
• approximation of M.^'+3'^(rM, f) of the form
M%£j*\rM,f)= J2 h^^ with hm G H-\D)
% e.g. hm := M1i'+>+1(Xm, f) for exact representation of M^'^'^rM, /)• solver of diffusion problem in D with deterministic coefficient e and
FE space VL
output: • approximation of M^'^^m,^^) of the form
M%£)(rM,VjiL) = Yl Srnfa
l|m|l»<M
with gm G Vl
1 compute for all m G N3++3, Hm^ < M
gm- ^Lhm G VL
2 for k — j' + j : — 1 : f + 1 do
3 compute for all m G N+ x, HmH^ < M
9m := A"i I div ^ </>mV^(m,m) 1 G Vl
M
m=l
4 end
5 compute and return
T
Algorithm 2: M%pX)(rM,v3tL) = Algorithm 2 MJ,f,L,M%£j,1)(rM,f)
38 CHAPTER 3. STOCHASTIC COEFFICIENT
input : • M, n,f, L
• approximations of M^'^'^^m, f) for all 0 < j < n, of the form
Mïp^W/) = E ^- with hmeH~1(D)
% e.g. hm :— M1''+'+1(Xm,f) for exact representation of M{-3'+3,1\rM, f)
• solver of diffusion problem in D with deterministic coefficient e and
FE space Vl
output: • approximation of M^3 '^(r^, um,l) of the form
M%£l(rM, uMiL) = Y 9ui(S>r.
||m||00<M
with gm G Vl
l compute for all j — 0 : 1 : n
M^W^l) = Algorithm 2 [m, j,/,!^^1'^,/)
2 compute and return
M^^M, UMtL) := J2M%p](rM, HU3=0
Algorithm 3: M%pl(rM,uM,L) = Algorithm 3 M, n,f, L, (Xip^J,1)(rM, f))0<3<
Proposition 3.26. For Algorithm 2 it holds
M%V(rM, v3tL) = (Ty+ltL o Ty+2,L o • • o Ty+3tL o S3,+3,L)(M%;3>lXrM, /)), (3.47)
so that
||A4(/,1W vhL) ~ M%vl)(rM,v3,L)\\A{w',ul{D)) <
< Ca\\M«+3»(rM, f) - M%^\rM, f)\\A{U3'+J,u-m). (3.48)
Proof. Estimate (3.48) is a direct consequence of (3.47) and Proposition 3.23, so that it
suffices to prove (3.47). To this end, we simply note that, in view of (3.40), applying S3>+3,l
to A4^ppJ,1(rM, /) corresponds to step 1: of the algorithm, whereas (3.39) ensures that step
3: consists of successive applications of the operators T3>+]jl, 2y+:?_li/J,..., Ty+])/-,. D
3.4. DISCRETE MIXED MOMENT PROBLEM 39
Proposition 3.27. For Algorithm 3 it holds
II.M^VM.UM.i) - MU£l(rM,uMtL)\\A{uriHèm <
<ca||/||Li(n,/f-i(D))lkM||i00(i7xn)exp(-ci,an)+n
+ EàLWMWWHm, f) ~ M%^l\rM, f)\\A{ui'+itH-HD)). (3.49)3=0
Proof. The l.h.s. of (3.49) can be estimated from above using the triangle inequality by
n n
\\MW(rM,uM,L - J>,l)|| + E \\M{j''l)(rM,v3jj) - M%P(rMivi%L)\\ (3.50)3=0 3=0
(all norms in A(U3 , Pq(P))). The conclusion follows then from the error estimates in
Propositions 3.16, 3.26.
One important property of the mixed moments which can be immediately used to reduce
the complexity of Algorithms 2, 3 is their symmetry w.r.t. m. To describe it, we define on
N+ the equivalence relation
m ~ m' <é=^ 3cj- permutation of {1,2,..., j} s.t. m'k = ma(k) VI < k < j, (3.51)
so that for any m G N^_ the orbit 0(m) of m is given by
ö(m) :— {m! G NJ+ | 3a permutation of {1, 2,... ,j} s.t. m'k — m^k) VI < k < j}.
Due to Definition 3.11 it then clearly holds
Remark 3.28. The mixed moment Mlj+1(Xm, f) is symmetric in m, that is,
m~m' = M1^(Xm,f) = M1^(Xmr,f)eH~1(D). (3.52)
Remark 3.28 shows that reasonable approximations of the mixed moments of the pair
(rM, f) used as input for Algorithm 2 should also satisfy the symmetry condition w.r.t. m,
m - m' =* hm = h^ G H'l(D). (3.53)
A simple backward induction argument on k shows then that this symmetry is preserved
by step 1: and the recursion in step 3: of Algorithm 2.
Proposition 3.29. If the input data (hm) m6NJ'+^ is symmetric w.r.t. m, then gm givenIMI»<M
by steps 1: and 3: of Algorithm 2 is also symmetric in m,
m~m' =» 9r* = 9m>ZVL<zHl{D). (3.54)
40 CHAPTER 3. STOCHASTIC COEFFICIENT
Remark 3.30. The complexity of the mean field computation using Algorithm 3 with
f — 0 and exact input data hm — M.lj+1(Xm, f) is in general superalgebraic. This follows
by a simple comparison error estimate (3.36) vs. total number of deterministic problems
which are to be solved in steps 1: and 3: of Algorithm 2. After choosing
M ~ P1A\ n ~ L (3.55)
in order to balance the contributions of the discretization steps to the total error estimate
(3.36), Algorithm 3 requires solving (even after taking into account the symmetry of the
t/m's) at least
^OKl/a-DIog^+l) (356)Mn
deterministic diffusion problems in D at level P (already in step 1:, Algorithm 2 with
j = n) for an accuracy of 0(N£S). Recalling that a = 1/d for analytic regularity of
the perturbation r (see also Example 3.6), (3.56) shows the superalgebraic complexity of
Algorithm 3 in dimension d > 2. In view of the algebraic (quadratic) complexity of the
Monte Carlo method, Algorithm 3 becomes then too expensive to use. Note however that
a subexponential complexity upper bound can be easily shown for Algorithm 2 by counting
arguments similar to those loading to (3.56).
The high complexity of Algorithm 3 can be significantly reduced by taking into account
the analyticity assumption on r, through the estimate (3.46). We show that representation
(3.44) can be sparsified, that is, many terms can be neglected (due to their small size as
estimated by (3.46)), without affecting the total discretization error estimate (3.36). We
consider thus for any j G N an index set (the explicit construction will follow in section
3.4.2)
T3 C {m G W+ | llmlloo < M} (3.57)
(convention: N^ — {0},T0 :— {0}) and apply Algorithms 2, 3 using as input the approxi¬
mation of MS1' +3'^(rM, f) given by
M<£*1KruJ):=M%$l\ru,f)= £ hm ® <f>m, (3.58)
which amounts to simply setting the input coefficients hm :— 0 for m ^ Ty+J-.
Defining
M%^(rM, v3,L) := (7>+i,l o 7>+2,L o •. o 7>+,,L o Sr+^)(M%^1](rM, /)), (3.59)
we rewrite Algorithms 2, 3 corresponding to the truncation index sets (Tj)j£n (sec Algo¬
rithms 4, 5).
Concerning notations, we introduce for any j, j' G N with j > f the projection operator
onto the first f coordinates, U3i3> : N3+ — N3+, and for any index set r C N3+ we put
U.jy(T) := |J njy(m) = {m' G W+ \ 3m G T s.t. m'k = mk VI < k < j'} Ç Nj.mer
3.4. DISCRETE MIXED MOMENT PROBLEM 41
Remark 3.31. The approximation error estimates for Algorithms 4, 5 are the same as
those given in Propositions 3.26 and 3.27, with approximate moments MaVP replaced by
sparsified approximate moments Ma.pP}r-
By sparsifying the moment representation using a convenient index family (Tj)je^ we gain
a substantial complexity reduction of Algorithm 3 (number of deterministic problems to
be solved in step 1:, see also Remark 3.30), without deteriorating the error estimate (3.49).
This will be shown in section 3.5, after the explicit construction of the index sets (T3)3£n-Here we only observe that if the index sets (Y3)jem are compatible with the equivalence
relation ~, that is,
meïj^ 0(m) Ç T3 \fj G N, (3.60)
and with the projection operators IT, that is,
n.yOf,-)^ Vj,j'N,j>f, (3.61)
then the following complexity estimate for Algorithm 4 is a consequence of Proposition
3.29, which applies also in the case of sparsified input data.
Proposition 3.32. If the index family (T3)je^ satisfies the compatibility conditions (3.60),
(3.61); then Algorithm 4 requires solving \T3'+3/ ~ | + \T3i+j_i/ ~ | + |Tj/+J_2/ ~ | H (-
IT?'/ ~ I deterministic diffusion problems in D at FE discretization level L.
3.4.2 Sparsification
For a fixed M G N and any j, q G N we define the sparsification index sets
Tjjq :— {m G N+ | ||m||oo < M and m has less than q different entries }. (3.62)
Proposition 3.33. Under Assumption 3.18 there exists c2<r > 0 such that for any j, q G N
it holds
|| J2 M^(Xm, f) ® Kh^WM-HD)) < 4eM-C2,rq1+a)\\f\\mn,H-HD). (3.63)
l|m|loo<M
Proof. In view of (3.46) it suffices to estimate (convention: empty sum equals 0)
Sj,q-=4 E exP(-<Vllmlla)- (3-64)
l|m|!oo<M
To each m G N3+ with HmH^ < M we associate the vector n— (rii,n2,... ,Um) G NM
defined by
VI < m < M nm:= p <=$> m has p entries equal to m. (3.65)
42 CHAPTER 3. STOCHASTIC COEFFICIENT
Then ||n||i := ni + n2 + • + Um = j and for any n with ||n||i — j there are exactlyfi
indices m G N+ with ||m||oo < M to which n is associated via (3.65).72i!n2!- .-riM
Moreover, we have
m G Tjj9 <=$> nm > 1 for less than q different values of m taken from {1,..., M}.
(3.66)With this notation, the sum S3tq in (3.64) can be written
M
SJ,Q = 4 E ,/' rexp(-ci,r Y]nmma),
neN-M
llnlli^j.lsuppfn)!^«
77,1!n2!
• -nM!nW« L ^
m=l
or, parametrizing the indices n through their support and the values they assume there,
min{j,M}.,
i
5M=4 E E E \3..n i^kE^O' (3-67)î=9 l<mi<-<m,<M neN^,||n||i=j
mi '
'
*"*' fc=l
supp(n)={m1, ,«!,}
The general term of the sum (3.67) can be estimated from above by
(nmi-l)!.">mt-l)!eXp(-Cl'-E(^ " ^TJ^xp^g^). (3.68)
Using (3.68) in (3.67) and performing the innermost sum via the multinomial formula gives
Sj,q < 4^ E (Eexp(~Cl>rmfcM P »!expKrE<)i>q l<mi<--<m,<M \k=l ) ^ ' k=\
* ^E^_i(i)<! E «?(-<*,, £>g). (3.69)«>9 l<mi<-"<m2<M k—\
Due to m/t > k for any 1 < k < i we absorb the factorial in the exponential under the
inner summation and obtain (with any 0 < c2<v < ci,r)
min{j,M} ,
.^s
s3« < 4 E *\M E cxP(-C2,rx;ot=g
^ 'l<mi<---<m,<M fc=l
< c3. max V) exp(-c2,rVm£). (3.70)l<mi<-<m,<M A;=l
The conclusion follows then from Lemma 5.2 (see Appendix). D
Proposition 3.33 shows that the sparsification error in the mixed moment representation
(3.44) with the index set TJ>g (3.62) is just as large as the largest neglected term, up to
the constants involved.
A further obvious property of the index sets (Tm)j£n is their consistency with the equiv¬
alence relation ~ and with the projection operators II in the sense of (3.60), (3.61). The
number of equivalence classes in TJt9 w.r.t. ~ can be bounded using a simple counting
argument.
3.4. DISCRETE MIXED MOMENT PROBLEM 43
Proposition 3.34. The number of equivalence classes in Tji9 satisfies the bound
rw~i<(M+i)«o'+i)'. (3-71)
Proof. Since every equivalence class contains exactly one element with non-decreasing entry
sequence, we need to count (via (3.65)) the distinct solutions n = (rii, n2,..., Um) G NM
of ri\ + 77-2 + ... + riM — j with nm > 1 for less than q values of 1 < m < M. We obtain
iw-i-E(T)(2-î)-(M+iro'+irD
3.4.3 Complexity
The computational effort required by Algorithm 4 with sparsification index sets given by
(3.62),
r3-T3,q V;GN, (3.72)
and with q G N to be chosen later follows directly from Propositions 3.32 and 3.34.
Proposition 3.35. Algorithm 4 consists in solving at most (j + j' + 1)9+1(M + l)9 deter¬
ministic diffusion problems at FE discretization level L.
Combining Propositions 3.27, 3.33, 3.35 we obtain the following error and complexity
estimate for Algorithm 5 below with exact input data,
hm := .MV-m+1 (Xm, /) Vm G r3,+j, VO < j < n. (3.73)
and sparsification index sets given by (3.72). We denote in the following the output of
Algorithm 5 below with exact input data (3.73) by
M{4'V(rM,uM,L). (3.74)
Proposition 3.36. /// G P1^, P-1(P0) and Assumptions 3.1, 3.18 hold, then for the
solution um,l G Ll(ft, Vl) to (3.17),
^aM{uj)UM,L(u) ~ /(w) in Vl P-a-e. lo £ ft (3.75)
it holds (w.l.o.g. ca > 1)
\\Mb'>l)(rM,UM,L) -M^^M^M^hiW^iD)) <
< cJ/||Li(n,H-i(D))(lkM|li'oo(c/x0) exp(-c1)a72) + c{+nexp(-c2trq1+a)). (3.76)
Moreover, M4'n (rM,UM,L) can be computed starting from M^J(rM, f) for all 0 <j<
n via Algorithm 5 below at the cost of solving at most (j' + 72 + l)q+2(M +l)q deterministic
diffusion problems in D at FE level L.
44 CHAPTER 3. STOCHASTIC COEFFICIENT
3.5 Mean Field Computation
From Propositions 3.17 and 3.36 with f — 0 we obtain
Proposition 3.37. /// G P1^, H~1(D)) and Assumptions 3.1, 3.18 hold, then
\\M}(uMiL) - M^^Um^WhUd) < Ca||/IUi(n,H-i(D))(exp(-Ci,an) + c^exp(-c2,ro1+a)).
(3.77)
//, in addition, f G LS(ft, H~1+S(D)) and Assumption 3.12 holds too, then
\\MHu) ~ M^n(uMiL)\\HèiD) <
< c0,a||/||Li(n,H-1+'(D))(exp(-c1,rMa) + $(NL, s) + exp(-ci,„n) + c>xp(-c2,rg1+Q))
(3.78)
for any M > MT, n,L,qe N.
Moreover, M\^(um,l) can be computed starting from M^ (^m,/) for all 0 < j < n
via Algorithm 5 at the cost of solving at most (n + l)q+2(M + l)q deterministic diffusion
problems in D at FE level P.
Note that the four terms on the r.h.s. of (3.78) correspond to the four discretization steps
(modelling assumption, FEM in D, truncation of Neumann series, and mixed moment
sparsification for the pair (r^f,/) respectively). Balancing the discretization errors in
(3.78) by the (unique up to multiplicative constants) appropriate choice of parameters
M~P1/a, n~L, <7~P1/(1+a) (3.79)
and comparing with the computational effort estimate obtained in Proposition 3.37 we
deduce (assuming also exact integration for all r.h.s.'s in step 3: of Algorithm 4),
Theorem 3.38. Under Assumptions 3.1 (well-posedness), 3.12 (uniform data regular¬
ity), 3.18 (analytic perturbation), the mean field of u random solution to (1.10) with
f G Ll(ft, H~l+S(D)) is numerically computable using Algorithm 5 at nearly the same
cost as one deterministic diffusion problem in D.
If the available deterministic solver in D needs 0(e~@) floating point operations and mem¬
ory to reach an accuracy e for one deterministic diffusion problem with coefficient e and
Hs regular source term (usually ß = l/«5 + o(l) with Ô = min(s,p)jd), then Algorithm 5 re¬
quires at most 0(e~@~°^) operations and memory to achieve the same accuracy £ —» 0 for
M1^) in the case of stochastic data. Algorithm 5 has therefore nearly optimal complexity
in the sense of Definition 1.1.
Proof. Choosing the discretization parameters M,n,q as in (3.79), the number of deter¬
ministic diffusion problems in D to be solved at FE level P (and given in Proposition 3.37)
admits then (up to a multiplicative constant) the upper bound
Lo(LVU+*>) < exp(0((logiVL)1/(1+^ loglogNL)) < N°L(1) as P -> oo, (3.80)
3.6. COMPUTATION OF HIGHER ORDER MOMENTS 45
which concludes the proof. D
Algorithms 4, 5 turn out to be also amenable to (almost full) parallelization.
Remark 3.39. The computation in step 1: of Algorithm 4 is fully parallelizable w.r.t.
m G Tj'+j. The computation in step 3: of Algorithm 4 is also fully parallelizable w.r.t.
m G Ilj'+3tk_i(Yji+j), but sequential w.r.t. k = / + j : — 1 : f + 1 (that is, j serial steps
are to be taken). Note that, due to (3.79), j < n ~ P ~ — loge, where e is the prescribed
accuracy to be reached in the mean field/higher order moment computation.
Further, the computation in step 1: of Algorithm 5 is also fully parallelizable w.r.t. j —
0 : 1 : n.
Remark 3.40. All results and algorithms derived for the mean field computation can
be easily generalized to the case of arbitrary first order statistics (weighted expectations,
obtained by replacing the probability measure P by a new one Q such that dQ/dP exists
in P°°(f2) and it is known).
3.6 Computation of Higher Order Moments
Based on the results obtained in Chapter 2 (stochastic source term), on the mean field
computation Algorithm 5 and under the regularity Assumption 3.18 we show that also
higher order moments of the solution u to (1.10) can be computed in almost the same
complexity as the solution of one deterministic diffusion problem in D.
We formulate first a continuous moment problem which we subsequently discretize and
decompose as direct sum of several mean field problems. These are in turn solved using
Algorithm 5.
Remark 3.41. In the following we discuss (for simplicity and since correlation and variance
kernels are often more relevant than higher order moments) only the computation of the
second order moment (k — 2). Corresponding results hold for all higher order moments
and can be derived in a similar manner.
For / G L°(ft, H~X(D)) we recall that the diffusion problem (1.10) with stochastic data
consists in finding u G L°(ft, Hq(D)) such that
Aa(lJ)«(w) - f(oo) in H~\D) P-a.e. toZft. (3.81)
After truncating the fluctuation expansion (3.6) at level M (see also (3.11)) we are left
with the problem (3.12) of finding uM G L°(ft,H^(D)) such that
Km{.)Um(^) = /M in H~l(D) P-a.e. a; G £1 (3.82)
From Proposition 3.9 we immediately obtain
46 CHAPTER 3. STOCHASTIC COEFFICIENT
Proposition 3.42. If Assumption 3.4 holds, then for any M > MT
\\u(lo)®u(lo)-um(lo)®um(uj)\\h^d^) < Ce,r,a-ex-P(-ci,rMa)\\f(Lo)\\2H-HD) P-a.e. ueft,
(3.83)
so that
\\M2(u) - .M2(22M)IUw ^ c*,r,a- exp(-c1<rMa)\\f\\h{n<H-HD)y (3.84)
We consider also the diffusion problem with constant coefficient 1 and stochastic right hand
side / G L°(ft, H~l(D)), whose solution we denote by / G L°(ft, H£(D)),
Af(u) = f(u) in P-'(P) P-a.e. lo G ft. (3.85)
We formulate next a continuous variational problem, to which um(-)®um(-) G P°(^, H^(D2))
is the unique solution. This problem will be in turn fully discretized. By Id#i(£)) we denote
in the following the identity operator acting in Hq(D).
Proposition 3.43. P-a.e. in ft it holds,
(A®A)(f(Lo)®f(uj)) = f(Lo)®f(uj) inH-\D)®H~l(D), (3.86)
(AaM(w) <g> ldm(D))(uM(u>) ® f(uj)) = (A ® IdHm)(f(Lo) ® J(lo)) in II~l(D) ® H^D),
(3.87)
(IdHi(D)®AaM(ij))(uM(Lo)®uM(üo)) = (ïdHi{D)®A)(uM(iu)®f(Lv)) in H^(D)®H~\D).
(3.88)
Proof. All three equations follow directly from (3.82) and (3.85) by tensorization. D
3.6.1 Discretization
We introduce the sparse tensor product spaces in D x D by
Vl := Span {Vt®Vv | 0 </ + /'< L} c H^(D2). (3.89)
Under Assumption 2.11, the approximation property of the scale (Vl)lsn reads (see e.g.
[Zen91], [ST03b])
min ||u - v\\HHD2} < $(NL, a)||«|U-i+.(Da) Vu G H~1+S(D2) n I$(D2), (3.90)
where the functional l> can be explicitly written in terms of <£. For instance we have
ê(NL) := cT(\ogNL)1/2N£s with 5 := min{p, $}/d for the FE spaces discussed in Example
2.12.
We further assume the existence of a wavelet family in the physical domain D, consistent
with the FE space scale (Vl)l£N-
3.6. COMPUTATION OF HIGHER ORDER MOMENTS 47
Assumption 3.44. There exist a family * = (ipki) im C Hq(D) and two constantsfcëiAj
£%*, c-2,® > 0 such that any u Hq(D) can be expanded as a convergent series,
u=5>k,^M ^Ho(D) (3.91)
fceA;
and the following three conditions are fulfilled
1. stability,
ci,* ^2 lCfc-'|2-
Il S cuM\hï(d) ^ c2,* X] IcmI2- (3-92)lew (N leN
fceA; feeA; fceA;
2. local support,
|A;| ~ 2di and for any /, /', the set supp (ipk,i) H supp (V>fc',Z') has nonempty interior for at
most ~(l + l' + l)^-1 • 2max^-r> values of (k, k') G A, x Ar.
3. consistency,
V"L - span{V>M | 0 < / < L, k G Vi}.
Constructions of families \P satisfying Assumption 3.44 are known for intervals (D =]0,1[),
hypercubes (D =]0, l[d) or polygonal domains (see e.g. [DS99]). Note that the stability
condition ensures the existence of a scalar product (•, •)* in H^(D) which is equivalent to
the standard one and has the property that the wavelet family * :— (ipk,i) ie« is an ONB
in (IP0(D), (.,-)*).
We consider the discrete counterparts of (3.86), (3.87), (3.88) associated to the sparse
FE space hierarchy (Vljlcn, and consisting in finding (//)l, (W)m,l, (uu)m,l G L°(ft, VL)
such that
(A®A)((ff)L(Lo)) = f(u;)®f(u3) in V£, (3.93)
(AffiMM ® ldHm)((uf)MiL(u)) = (A ® IdHi(D))((//)L(a;)) in V£, (3.94)
(Idffi(D) ® AaM(Lj))((uu)M,L(uj)) = (ldHi{D) <» A)((u/)MiL(o;)) in 0L*. (3.95)
More precisely, by (3.94) we understand the variational formulation which in the second
variable is associated to the scalar product (bilinear form) (•,•)*, and similarly for (3.95).
Proposition 3.45. Under Assumption 3.12, it holds, P-a.e. in ft and for all M > Mr
(here Id := Idjjx^)),
\\UM(UJ) ® UM(W) - (uu)jlf,LM||ffi(D3) < Ca,*(||(Id ~ ^X/M ® /(^)) IU01(^) +
+ ||(Id - PyL)(uM(uo) ® /»)||Wol(D=) + ||(Id - PyL)(uM(uj) ® «mM)||hi(^)), (3-96)
so t/iai z/ also Assumption 3.12 is satisfied, then P-a.e. lo G ft
\\um(uj) ® %(w) - (u«)M,L(w)||iii(Da) < c^,**(iVL)||/(w)||^-i+.(£)) (3.97)
48 CHAPTER 3. STOCHASTIC COEFFICIENT
and
\\M2(uM) - Ml((uu)MtL)\\HHDi) < ca,*e(NL)\\f\\h(n,H-i+sm. (3.98)
Recall that $(NL) :— cT(log NlY^N^5 with 5 :— min{p, s}/d for the FE spaces discussed
in Example 2.12.
Proof. The P-a.e. uniform ellipticity of the bilinear forms giving (the l.h.s.'s of) (3.86),
(3.87), (3.88) for M > Mr ensures the quasi-optimality of the corresponding FE solutions
in (3.93), (3.94), (3.95). We thus have, P-a.e. in ft,
||«m(w) ®«Af(w) - (MU)M,LM||jfi(£)2) < Caty(\\(ld - PyL)(uM(u) ® Um(uj))\\H1(D>)+
+ \\um(u) ® f(co) - (uJ)m,l(u)\\hUd*)) (3-99)
\\UM(UJ) ® /(w) - K/)m,lM||tf0i(L>2) < Ca,*(II(Id - PvL){uM{w) ® /H)||fl-i(£,a)+
+ ll/>) «8 /» " (/7)lMII*oW) (3-100)
||/» ® /(W) - (//)L(a;)||ffi(D3) < ||(Id - P^X/M ® /("))||*oW. (3-101)
Inserting (3.101) in (3.100) and then the resulting estimate (3.100) in (3.99) we obtain
(3.96). Further, (3.97) follows from (3.96) and the approximation property (3.90) of the
sparse tensor product spaces. Integrating (3.97) over ft we obtain (3.98).
3.6.2 Algorithms
We start by decomposing equations (3.94), (3.95) into simpler, mean field problems using
the wavelet family * := (ipk,i) '^ Since (ff)L, (uJ)ml, (uu)m,l £ L°(ft, VL), we can write
(/7)l = J2 ^,i®^k,L-i (3-102)0<1<L
fceAL_;
(uf)u,L = J2 ßk,i®^k,L-i (3-103)
= J2 ^L-i®-yk,i (3-104)0<1<L
keAL_t
(uu)M;L = Yl V'fc.L-Z®^ (3.105)0<1<L
*eAt_;
where ak,i, ßk,i, lk,u h,i G L°(ft, V) for all 0 < / < L, k G AL-i-
For future use we note here that the family (7^/) o<i<l can be obtained from the familyfceAL_,
(ßki) o<i<l by a rearrangement of its coefficients in the Vl basis (^/)o<kz,. The proof'
fcAL„,'
fceA;
follows by simple manipulations of (3.103) and (3.104).
3.6. COMPUTATION OF HIGHER ORDER MOMENTS 49
Proposition 3.46. It holds
7W(üj) = ^2 (ßk\L-i'(u),ipk,L-i)vipk>,i> P-a.e. lü eft. (3.106)0<l'<l
fc'eAj,
Using the decompositions (3.102), (3.103), (3.104), (3.105) in the variational formulation
of (3.94), (3.95) we obtain
Proposition 3.47. Equation (3.94) is equivalent to
AaM{Lj)ßktl(Lo) = Aakil(uj) in Vf V0 < I < P, k G A^ P-a.e. oo G ft. (3.107)
Similarly, equation (3.95) is equivalent to
AaM(^ök,i(Lo) = A-ïk,i(u) in VC V0 < I < L, k G AL-i P-a.e. u G ft. (3.108)
Keeping in mind the results of section 3.4, let us explain how the correlation computation
algorithm based on the discretization step given by (3.93), (3.94), (3.95) should work.
The correlation kernel A/(1((wu)m,l) can be obtained using (3.105) from the mean fields
•M1^,;), which in turn follow in view of (3.108) from the mixed moments of (rM,7fc,;)-
By rearrangement via (3.106), the mixed moments of (rM,lk,i) are expressed in terms of
the mixed moments of (r^, ßk,i), which follow then from the mixed moments of (rM, ctk,i)-
Finally, the mixed moments of (rM,ctk,i) can be obtained, due to (3.102), from the mixed
moments of (rM, (//)l) (or of (rM, f ® /)). The algorithm will therefore have to perform
all these steps in reverse order, starting from the available information, the mixed moments
of the pair (rM,f®f)-
In the following we use the notation
(A®A)^ : H~\D) ® H~\D) - VL (3.109)
for the discrete solution operator associated to the bilinear form of A <S> A (coercive on
Hq(D2)) and the sparse FE space Vl- Further, we denote by
iW.HÏM^C, i*,i,(0 :=(Ç,il>)v VÇeHl(D) (3.110)
the contraction in the Hilbert space (Hq(D), (, •)$) with vector ip G Hq(D).
With these notations, simple algebra shows the connection between the mixed moments of
the pair (rM, f ® f) and those of (rM, ak<t).
Proposition 3.48. In A(Uj, H^(D)) it holds
M$'1\rM,0[k,l)= Y2 9m,akzl ® <f>m (3.111)meTj
where
0m,aW = (IdHi(D)®z*^,L_!)(A®A)Z1X^«(Xm,/®;E/) G Vt (3.112)
satisfy the symmetry condition w.r.t. m (invariance under entry permutations).
50 CHAPTER 3. STOCHASTIC COEFFICIENT
Proof. Using (3.102) and the linearity of the mixed moments of order (j, 1) in the last
argument we write
M%'1)(rM,(ff)L)= J2 M^1](rM,ctk,i)®^k,L-i. (3.113)
fceAL_,
But (3.93) shows that
M^\rM, (ff)L) = Y, (A^A)^1^^, / ®x f) ® <j>m, (3.114)
so that the conclusion follows by applying \dHi^ ® z*,^fcL_, ® Id^«^) to (3.114) and
(3.113). D
Noting that, due to (3.93) and (3.102), / G L2(ft,II^(D)) ensures ak,i G L\ft, H^(D))
boundedly, we apply Proposition 3.36 to the equation (3.107).
Proposition 3.49. Under Assumptions 3.1, 3.18, M^^'(rM,ßkti) can be computed star-
rting from My3'
(rM,ctkj) for all 0 < j < n via Algorithm 5 at the cost of solving at
most (j' + n + l)q+2(M + l)q deterministic diffusion problems in D at FE level I. The
error estimate reads
<
< Ca||aw||Li(Q^i(Z3))(||rM||{00(i7xn)cxp(-clian) + c{+"exp(-c2]7.<71+a)). (3.115)
The next step of the algorithm, passing from the mixed moments of the pair (rM,ßkj) to
the mixed moments of (rM, 7k,i), can be analyzed using Proposition 3.46 and the linearity
of the approximate mixed moment M^1' in the last argument (see (3.59)).
Proposition 3.50. //
M(4'»(rM, ßk,i) = J2 SmA, ® &n, (3-116)
mT3,
with #m,Ä,i e Vi, then
M^\rM,lk,i) = J2 0m,7fcll ® ^m (3.U7)
meTj,
where #m,7M G VJ is given by
9m,yk,i = /] (9m,ßk,iL_l,,lI>k,L-l)*ll1k',l'- (3.118)
fc'eA,,
Moreover, the following error estimate holds,
IIA^W.Tw) - M^(rM,lkMA{WM1(D))<
< Ca,*||/||22(n>jy-ip))(lkM|lioc([/xn)exp(-Cl,an) + c3a+nexp(-c2ir(? +a)). (3.119)
3.6. COMPUTATION OF HIGHER ORDER MOMENTS 51
Proof. From (3.106) we obtain
M{4£\'rM,lk,l) = Y M?,n\rM,(ßk',L-l'(-),lJ1k,L~l)<i>'lpk>,l')0<l'<l
fc'eA,,
= J^ (IdL-(^') ® **,^.l-i ® IdHHD))M%t£\rM, ßk',L-l>) ® ^fc',J'
0<l'<l meïjifc'eA,,
from which (3.118) follows by interchanging the summations. To check (3.119) we first
note that jk,i solves, in view of (3.106),
AttM(w)7fc.iM = Y (ßk<,L-i'W),i>k,L-i)ykaM{w)'il)ki,i> h\VL* P-a.e. uj ft. (3.120)0</'</
fc'A"J,
Next we observe that the ^(ft, H~1(D)) norm of the r.h.s. in (3.120) can be estimated
from above by
Ca>* /( XI W^'.L-l'^Wjji^Y^dP^) < C,*||(u/)jlf,L||Li(n,/rçp»)) < Ca,*|l/llï,2(n,H-i(£>))»o
0<i'<i"
fc'eA,,
so that (3.119) follows from Proposition 3.36 applied to the equation (3.120). O
Finally we investigate the computation of ök,i solution to (3.108). We apply the T version
of Proposition 3.27 (see also Remark 3.31) to equation (3.108) using as input data for
Algorithm 5 the approximate moments
M%p%(rM,Alk,i) := M^i\rM, A7*,i) (3-121)
and obtain
Proposition 3.51. Under Assumptions 3.1, 3.18, an approximate mean field
Ml^^A^i) := Algorithm 5 [m, n', 0, (M^\rM, A7w))0<,-,<n,] (3.122)
can be computed at the cost of solving at most (n' + \)q+2(M + l)q deterministic diffusion
problems in D at FE level I and for an accuracy (w.l.o.g. ca, \\rM\\L°°(Uxü) > lj
\\MH6u) - MlnwVMmD) £
< ca\\f\\2mn,H-i(D))(eM-Ci,an') + ||rM ||2»([,xn)exp(-cli0n) + <'+nexp(-c2,^1+a)).(3.123)
Due to the orthogonal decomposition (3.105) we then have, with NL := diml4>
52 CHAPTER 3. STOCHASTIC COEFFICIENT
Corollary 3.52. Under Assumptions 3.1, 3.18 it holds
\\M1((uu)M,l) ~ Yl ^k,L-l®Mlpp,ytntnl(SkMHZ{D*)<0<1<L
fceAj,.,
<ca\\f\\2LHnM-imÛl/2(exp(-chan')^rM\\^(3.124)
We summarize more formally in Algorithm 6 the second moment computation we have
developed and analyzed in this section. The total error estimate follows by collecting all
discretization errors, as given by Propositions 3.42, 3.45 and Corollary 3.52.
Proposition 3.53. It holds
II^'H" Y, V'fc,L-/®Xapp,T,n,n'(^)ll//01(Û2)^0<1<L
fceAL_,
< Ca,r||/||ia(n,H-ip))(exp(-clirMa) + *(NL) + A>i/3(exp(-Cl,an')+
+ lkM||2Vx«)exp(-Cl,an) + c;?'+"exp(-c2,7y+Q))). (3.125)
The complexity of Algorithm 6 can be estimated using Propositions 3.49 and 3.51.
Proposition 3.54. Step 1 of Algorithm 6 consists in solving |Tn/+„/ ~ | + |T„<+n_i/ ~
\-\ 1-1 To/ ~ | second moment diffusion problems in D x D with constant coefficient and
in the sparse FE space Vl.
Steps 2 and 4 of Algorithm 6 consist in solving at most 2|A/-/_z|(n' + n + \)q+2(M + l)q
deterministic diffusion problems in D with coefficient e and in the FE space Vi for all
0 < / < P.
Step 3 of Algorithm 6 consists only of a sorting procedure applied to the coefficient sets of
(9m,ßki) °s'<l in the wavelet basis ^ (compare (3.118)/
Proof. The complexity estimate of step 1 follows from the moment symmetry w.r.t. m,
whereas those of steps 2 and 4 are direct consequences of Proposition 3.36.
Note that a more explicit complexity estimate of step 1: in Algorithm 6 follows from the
main result in Chapter 2, that we recall next.
Theorem 3.55. Under Assumption 2.15 and if f G L2(ft, H~1+S(D)), each problem in
step 1: of Algorithm 6 is numerically solvable for any L > 1 at a cost of
cr,*(logiVL)2d+2AL (3.126)
floating point operations, with
CT(logNL)NL (3.127)
memory, for an accuracy of
ca,u,r(logNL)1/2NZô, (3.128)
where Ô = min{p, s}/d.
3.6. COMPUTATION OF HIGHER ORDER MOMENTS 53
Balancing the discretization errors in (3.125) by the choice of parameters
M-P1/", n'~L, n~P, q ~ Lll{l+a) (3.129)
and comparing with the computational effort estimate obtained in Proposition 3.54 we
deduce (assuming also exact integration for the r.h.s. of all deterministic problems to be
solved in Algorithm 6),
Theorem 3.56. Under Assumptions 3.1 (well-posedness), 3.12 (uniform data regularity),
3.18 (analytic perturbation) and 2.15 (existence of Hq(D)-stable wavelet family), the sec¬
ond order moment (correlation) of u random solution to (1.10) with f G L'2(ft,H~1+s(D))
is computable using Algorithm 6 at nearly the same cost as one deterministic diffusion
problem in D.
More precisely, if the available wavelet based solver in D needs 0(e~ß) floating point op¬
erations and memory to reach an accuracy £ for one deterministic diffusion problem with
coefficient e and Hs regular source term (usually ß — I/o + o(l) with ö = min(s,p)/d),
then Algorithm 6 requires at most 0(e~@~0^) operations and memory to achieve the same
accuracy e —> 0 for M2(u) in the case of stochastic data. Algorithm 6 has therefore nearly
optimal complexity in the sense of Definition 1.1.
Proof Due to Propositions 3.34, 3.54 and Theorem 2.28, the computational effort of step
1 consists in at most NL+ operations as P —> oo, by the same argument used in (3.80).
This argument shows also that the cost of steps 2 and 4 admits (up to a multiplicative
constant) the upper bound
Y\AL-i\N°LWNl+0{1) < NL+0{1) asP-oo, (3.130)
for an accuracy of &(Nl) ~ N£ (compare Theorem 3.55). D
Remark 3.57. All solution operations contained in steps 1:, 2:, 3:, 4: of Algorithm 6 are
fully parallelizable w.r.t. the indices l,j,j' and k G Al-i-
54 CHAPTER 3. STOCHASTIC COEFFICIENT
input : • M, j,f, L
• approximation of M^ +3,1\rM, f) of the form
M%^1}(rM,f) = Y h^ with hmeH~1(D)
% e.g. hm :— M1y+i+1(Xm, f) for exact sparsified representation of
M{4'+jA)(rMJ) (see (3.73))• solver of diffusion problem in D with deterministic coefficient e and
FE space VL
output: • approximation of M^'^^tmjVj^) of the form
,0M)M.%£t(rM,vjjL)= Y 9m^
tn&n3'+3,j'(Tj'+3^
with gm G VL
1 compute for all m g Yy+3
2 for k = f + j : -1 : j' + 1 do
3
gm := A~^m G Vl
compute for all m G Uj'+j^-iC^f+j)
(
9m -- A-L div y ^m'fcVp,, G Vi
4 end
5 return
M%p^(rM,v3,L) := 5Z #m<^
men/+j,3'(T/+3)
Algorithm 4: ^^(rM,^) = Algorithm 4 MJ,/, P, A<VM,/)<(j'-H,l)/
3.6. COMPUTATION OP HIGHER ORDER MOMENTS 55
input : • M, n,f, L
• approximations of M^'+3'l\rM, f) f°r all 0 < j < n, of the form
M%^l)(rM,f)= Y h<t> with hmeH~1(D)meTj'+l
% e.g. hm := M1^'+^1(Xm, f) for exact sparsified representation of
M%'+lA\rMJ) (see (3.73))
• solver of diffusion problem in D with deterministic coefficient e and
FE space VL
output: • approximation of M^'^^m^m^) of the form
Mp'Ä,n(rM, um,l) = Y 9m<^m
with gm G Vl
l compute for j — 0 : 1 : n
<0M)Mi;/r(rM)vJtL):= Algorithm 4
M,j,j',L,M^'(rM,f)Aj'+3,y)t
2 compute and return
M%p\n(rM,uM,L) :=YMZ&(rM,v3,L)j=0
,0',»Algorithm 5:
M^^n(rM, uM,L) = Algorithm 5 M, n,j>, L, (M^'^m, f))0<3<nA3'+3,1) I
56 CHAPTER 3. STOCHASTIC COEFFICIENT
input : • M, n, n', L
• approximations of M^'+3'1\tm, f ®x f) for all 0 < j < n, 0 < / < n'
of the form
M%+jA\rM,f<8>f)= Y h^-lW+3,1),
;i• m,j w j )=
with hm := A^V+J+i(Xm, / ®x /) H-\D2)
• solver of deterministic diffusion problem in D with coefficient e
on FE space scale (V/)o<kl
• ^-wavelet based deterministic solver of second moment diffusion problem
in D x D with sparse FE space Vl
output: • approximation Miw^tnn,((uu)M,L) £ VL of M2(u)
1 compute for all m G T^+j-, 0 < j < n, 0 < f < n' and all 0 < / < P, k G AL-i
9m,akil := (Idffi(I3) ® i*,^iL_,)(A®A)£1/iin G Vi
2 compute for all 0 < f < n! and all 0 < I < L, k G Al_j,
M{4'A)(rM,ßk,i)= Y VA,i®^m
by(i'+^i)/
A^T.n (rM, &.i) - Algorithm 5 \M,n,j', I, (M% +3'l>(rM, Aaw))0<,<
3 compute for all 0 < f < n' and all 0 < / < P, A; G Al_/ via (3.118),
M$;1){TM,'1kÙ= Y 3m,lkil ® (pm
meYji
4 compute for all 0 < I < L, k G Al-/,
MSÎï.^^M, 4,/) - Algorithm 5 [m, n', 0, /, (M^l)(rM, Mk,i))o<r<n'
5 compute and return
MlpPtr,n,n>((UU)M,L) •= Y ^-' ® ^W.T.n.n'OW)
fceAL_,
Algorithm 6: (M\w T n n,(£M)) °<<<^ = Algorithm 6[M, n, n', P, (A4?'+j,1)(rM, / ®
/)) 0<j<n ]0<j'<Ji'
Chapter 4
Conclusions and Open Issues
Classical perturbation algorithms for the moment computation of the solution to an ellip¬
tic problem with stochastic data have been further developed and investigated in terms of
complexity. Statistical information about the data, needed as algorithm input, has been
explicitly determined. It has been shown that, under essentially only one additional as¬
sumption, the spatial regularity of the fluctuation, the algorithms are, asymptotically in
work vs. accuracy, more efficient than the standard MC/SG/WPC methods, and a priori
nearly optimal (in the sense of Definition 1.1) complexity bounds have been proven. As in
the case of a standard MC simulation, the computation is almost fully parallelizable and
the implementation requires essentially only one routine to call any available, préexistent
FE code for the corresponding deterministic elliptic problem and run it successively with
different source terms.
We conclude by enumerating several possible extensions of the results presented in this
thesis, along with some further open problems for the solution of which the techniques and
ideas used here might be, in our opinion, relevant.
1. Relaxation of the fluctuation regularity condition r G A(D, L°°(ft)) to only finite
Sobolev regularity. We further conjecture that, in order that perturbation algorithms
retain nearly optimal complexity, the stochastic fluctuation must be at least piecewise
continuous in the physical domain D.
More generally, it could be of interest the construction of a regularity space scale
(Xs)o<s<2 interpolating between smooth and rough data corresponding to 6 — 0 and
5 = 2 respectively, such that r G Xs ensures the complexity rate e~5+of-^ for the
first/higher order moment algorithms, or for some combination of perturbation and
Monte Carlo algorithms (quasi-Monte Carlo).
2. Development of parallel results for the stochastic Galerkin (SG) method which is
known to deliver, under stronger assumptions (known joint distributions of the ran¬
dom variables (Xm)me^+ in the fluctuation expansion), complete statistical informa-
57
58 CHAPTER 4. CONCLUSIONS AND OPEN ISSUES
tion (i.e. the probability distribution function) on the stochastic solution u.
3. Study of the perturbation algorithm scalability w.r.t. the correlation length of the
diffusion coefficient a and the case of a small correlation length, equivalent to a slow
(still exponential) convergence of the fluctuation expansion due to unfavourable val¬
ues of the constants involved in the decay estimates, or to a large pre-asymptotic
domain for all results presented in this thesis. This pre-asymptotic analysis should
also give some insight into the case of a rough stochastic fluctuation, including white
noise, and might require new random field representation techniques, in order to
parsimoniously parametrize the data uncertainty, plus an investigation of the appli¬
cability of perturbation algorithms in this new context.
4. Generalization and adaptation of perturbation algorithms to the case of unbounded
stochastic diffusion coefficient with bounded samples (lower and upper bounds are
uniform in the physical domain, but are sample dependent, e.g. log-normally dis¬
tributed a).
5. Extension to general second order elliptic/parabolic/hyperbolic problems with stochas¬
tic data (advection-diffusion, linear elasticity, heat, wave equations, Stokes problems
etc.).
Chapter 5
Appendix
5.1 A Trace Lemma
Lemma 5.1. If s,t,u are real numbers such that
t>s>0, u>s>0 and t + u~s>d/2, (5.1)
then the mapping (the trace on the diagonal set in Rd x Rd)
S(R2d) 3 0 -=* Z(<p)(x) = (f>(x,x) G S(Rd) (5.2)
has a unique linear and continuous extension from P'(IRd) ® Hu(Rd) to Ps(Rd).
Proof. Denoting by Td (and by hat) the Fourier transformation (and transform) in Rd, we
compute the Fourier transform of the mapping S (denote by dx the normalized Lebesgue
measure in Rd, dx = (2ir)-d/2d\),
S$(p) := (ft o H o f£(4>))(p) = J e~^(E o F£$)){x) dx
= f e~^x f eix<t+r>^(Ç,r])dÇdridx
= J e-irxe0î>{tO-Ç)dïMdx = j fa,p-i)d£. (5.3)
R3rf
Since the uniqueness of an extension follows by a density argument, we only show that
S : <S(M2d) — <S(Md) computed in (5.3) has a linear and continuous extension from Kt(Md)®
Ku(Rd) to Ks(Rd) if s,t,u satisfy (5.1) and
Kr(Rd) := L2(Rd, (x)2rdx) with (x)r := (1 + |x|2)r/'2 Vr > 0. (5.4)
59
60 CHAPTER 5. APPENDIX
We estimate therefore
\m\\h = J(p)2s-(jktP-odA dp
2
(pY
(^(p~0u(0*(p-0"-^p-0^ dp
<
RdRd NR"* / \r<* /
r (pY3- il y /g\2t/p _ ^
^lU"^,^) • iMk*®*«- (5.5)
All we have to check is that (5.1) ensures the finiteness of the P°° norm in (5.5). To this
end, we note first that it is enough to prove the uniform boundedness of the integral in
(5.5) for p outside a compact set in Rd, e.g. for \p\ > 1. Let 1/^/2 < c < 1 be an arbitrary
real number. We split the integral in two parts, coresponding to the integration domains
D\ :— {£ | (p — 0 > c- (p)} and D2 :— {£ | (p — £) < c- (p)} and estimate the corresponding
integrals as follows
f (p)2s ! f 1
J (02t(p~t)2udli -PJ (02t(p-02{u-s)d^ (5'6)
But it is easy to see that (p — £) > c • (p) with l/y/2 < c < 1 and |p| > 1 implies
(p — 0 > c' • (£) for any 0 < d < y/2c2 — 1. Inserting in (5.6) we obtain (w > 5)
f (P)2'dC<
1 /• 1g (57)
y (£)2*(p _ ^2u"< -
C2S . cJ2(u-s) J ß)2(t+u-s)"Ç ^' >
-Dl Krf
and the r.h.s. of (5.7) is finite, due to t + u — s > d/2, and does not depend on p.
To estimate the integral over D2, we note first that D2 is a ball of center p and radius
y/c2 • (p)2 — 1, so that (£) > c' (p) with some positve d > 0 depending only on c < 1. It
follows(^)2S
„ ^
! / (P)2(s-t]f (P)2Sd: <
J_/
J (02t(p-02uç -
d*'J {p-02uD2 Di
di
r^j
< <p)2^- j (r)"2"^"1 dr
0
(p)2(S-t)-2u+d if u < rf/2
(p)2(s-*) • log(p) if u^d/2 (5.8)
(p)2(s-t) if u> d/2
with constants independent of p. The uniform boundedness follows in all cases from the
assumptions t > s and t + u — s > d/2.
5.2. A FAST DECAYING SEQUENCE 61
5.2 A Fast Decaying Sequence
Here we prove that, if y,a > 0 and j G N+, the sum of the series with general term
exp(—y 2~Di=i m%) indexed over 1 < mi < ... < m3 < oo is, qualitatively and uniformly in
j G N+, just as large as the leading term, corresponding to mk — k for all 1 < k < j. More
precisely, it holds
Lemma 5.2. If a > 0, and x > y > z > 0, then there exist ca]X>2/, caiytZ > 0 such that
1J
1
c^expf-x--—jl+a) < Y T\eM-yrnk)<ca,y,zexp(-z-^j1+a) (5'9)l<mi<...<mj<oo k—1
VjGN+.
Proof. For y > 0 and j, J G N+ U {oo} with j < J we set
s*- e n««pHK)- (5-iq)l<mi<...<raj<J k=l
We check first the lower bound in (5.9). To this end we note that the sum in (5.10) contains
the term corresponding to mk — k for all 1 < k < j, therefore we have
S3<J>exp(--yYka). (5-11)fc=i
The conclusion follows by noting that
} 1
Yk° ^ a + !)1+a / x<X(ix = 0' + i)1^!^- (5-12)
It remains to prove the upper bound of the sum in (5.9). From the definition (5.10) it
follows that the sequence (Sj>3)3^+ is rapidly decaying, that is
Sj>3 < ca,y,0ß3' VjGN+,V/3>0. (5.13)
From (5.10) we derive the recursive formula
j-i
Sj,J+1 = S3,j + exV(~y(J+ir) Y Y[eM-y<)l<mi<...<mj_i<J fc=l
= Sjj + expi-viJ + mSj-uj (5-14)
From (5.14) it follows immediately by induction on j that
Sjtj < S3,oo = Jim Sj,j < oo Vj G N+) (5.15)
62 CHAPTER 5. APPENDIX
and that
Sjt00 < Sjj + Y eM-yk°) ^-1>00. (5.16)fc=J+l
Now, for an arbitrary 7 G (0,1) we have, for j large (j > j0, with j0 depending on y, a, 7),
00
Y exp(-yfc*)<7, (5-17)fc=j+i
which ensures
S/,00 < Sj-j + 7-SVi.oo, Vi > jo- (5-18)
From (5.13) and (5.18) we deduce that
Sjt00 < ca,vß((7 + ß)j + 7J-J0+15i0-i,oo), (5.19)
which shows that Sjj0O — 0 as j —> 00, by choosing ß such that 7 + /? < 1. The sequence
(Âj,0O)jeN+ is in particular bounded, that is
S;,«, < Cc,tf VjGN+. (5.20)
Since this inequality holds for any y > 0, the conclusion follows then from (5.20) upon
replacing y by y — z. d
Bibliography
[Bab61] Ivo Babuska. On randomised solutions of Laplace's equation. Casopis Pest. Mat,
86:269-276, 1961.
[BC02] Ivo Babuska and Panagiotis Chatzipantelidis. On solving elliptic stochastic partial
differential equations. Comput. Methods Appl. Mech. Engrg., 191(37-38):4093-
4122, 2002.
[BG88] Ivo Babuska and Ben Qi Guo. Regularity of the solution of elliptic problems with
piecewise analytic data. I. Boundary value problems for linear elliptic equation
of second order. SIAM J. Math. Anal, 19(l):172-203, 1988.
[BG99] Hans-Joachim Bungartz and Michael Griebel. A note on the complexity of solv¬
ing Poisson's equation for spaces of bounded mixed derivatives. J. Complexity,
15(2):167-199, 1999.
[BTZ04] Ivo Babuska, Raul Tempone, and Georgios E. Zouraris. Galerkin finite element
approximations of stochastic elliptic partial differential equations. SIAM J. Nu-
mer. Anal, 42(2):800-825 (electronic), 2004.
[Coh03] Albert Cohen. Numerical analysis of wavelet methods, volume 32 of Studies in
Mathematics and its Applications. North-Holland Publishing Co., Amsterdam,
2003.
[Dah97] Wolfgang Dahmen. Wavelet and multiscale methods for operator equations. In
Acta numerica, 1997, volume 6 of Acta Numer., pages 55-228. Cambridge Univ.
Press, Cambridge, 1997.
[Dav63] Philip J. Davis. Interpolation and approximation. Blaisdell Publishing Co. Ginn
and Co. New York-Toronto-London, 1963.
[DS99] Wolfgang Dahmen and Rob Stevenson. Element-by-element construction of
wavelets satisfying stability and moment conditions. SIAM J. Numer. Anal,
37(l):319-352 (electronic), 1999.
63
64 BIBLIOGRAPHY
[DW81] Mike D. Dettinger and J.L. Wilson. First order analysis of uncertainty in numer¬
ical models of groundwater flow, 1, mathematical development. Water Resour.
Res., 17(1):149-161, 1981.
[ER03] Isaac Elishakoff and Yongjian Ren. Finite element methods for structures with
large stochastic variations, volume 7 of Oxford Texts in Applied and Engineering
Mathematics. Oxford University Press, Oxford, 2003.
[Fed69] Herbert Fédérer. Geometric measure theory. Die Grundlehren der mathematis¬
chen Wissenschaften, Band 153. Springer-Verlag New York Inc., New York, 1969.
[FST05] Philipp Frauenfelder, Christoph Schwab, and Radu A. Todor. Finite elements
for elliptic problems with stochastic coefficients. Comput. Methods Appl. Mech.
Engrg., 194(2-5):205-228, 2005.
[GÜ87] Richard 0. Gilbert. Statistical methods for environmental pollution monitoring.
John Wiley & Sons, New York, 1987.
[G095] Michael Griebel and Peter Oswald. Tensor product type subspace splittings and
multilevel iterative methods for anisotropic problems. Adv. Comput. Math., 4(1-
2):171~206, 1995.
[GS91] Roger G. Ghanem and Pol D. Spanos. Stochastic finite elements: a spectral
approach. Springer-Verlag, New York, 1991.
[GT01] David Gilbarg and Neil S. Trudinger. Elliptic partial differential equations of
second order. Classics in Mathematics. Springer-Verlag, Berlin, 2001. Reprint of
the 1998 edition.
[Kel64] Joseph B. Keller. Stochastic equations and wave propagation in random me¬
dia. In Proc. Sympos. Appl. Math., Vol. XVI, pages 145-170. Amer. Math. Soc,
Providence, R.I., 1964.
[KH92] Michael Kleiber and Tran Duong Hien. The stochastic finite element method.
John Wiley & Sons Ltd., Chichester, 1992.
[Loè77] Michel Loève. Probability theory. I. Springer-Verlag, New York, fourth edition,
1977. Graduate Texts in Mathematics, Vol. 45.
[Loè78] Michel Loève. Probability theory. II. Springer-Verlag, New York, fourth edition,
1978. Graduate Texts in Mathematics, Vol. 46.
[MK05] Hermann G. Matthies and Andreas Keese. Galerkin methods for linear and non¬
linear elliptic stochastic partial differential equations. Comput. Methods Appl.
Mech. Engrg., 194(12-16):1295-1331, 2005.
BIBLIOGRAPHY 65
[SchOO] Wim Schoutens. Stochastic processes and orihogonal polynomials, volume 146 of
Lecture Notes in Statistics. Springer-Verlag, New York, 2000.
[ST03a] Christoph. Schwab and Radu A. Todor. Sparse finite elements for elliptic problems
with stochastic loading. Numer. Math., 95(4):707-734, 2003.
[ST03b] Christoph Schwab and Radu A. Todor. Sparse finite elements for stochastic
elliptic problems—higher order moments. Computing, 71(l):43-63, 2003.
[Wie38] Norbert Wiener. The homogeneous chaos. Amer. J. Math., 60:897-936, 1938.
[XK02] Dongbin Xiu and George Em Karniadakis. The Wiener-Askey polynomial chaos
for stochastic differential equations. SIAM J. Sei. Comput, 24(2):619-644 (elec¬
tronic), 2002.
[Yos95] Kösaku Yosida. Functional analysis. Classics in Mathematics. Springer-Verlag,
Berlin, 1995. Reprint of the sixth (1980) edition.
[Zen91] Christoph Zenger. Sparse grids. In Parallel algorithms for partial differential
equations (Kiel, 1990), volume 31 of Notes Numer. Fluid Mech., pages 241 251.
Vieweg, Braunschweig, 1991.
66 BIBLIOGRAPHY
«a*«».
Curriculum Vitae
Name Radu Alexandru Todor
Date of Birth November 12, 1973
Place of Birth Brasov, Romania
Nationality Romanian
Education
1980-1992
1988, '89, '90
1991
1992
1992-1997
1997-1999
1999
2000-present
Primary School/Secondary School/College
"Andrei Saguna" Brasov, Romania
First Prize, Romanian National Mathematical Olympiad
Bronze Medal, International Mathematical Olympiad, Sigtuna, Sweden
Gold Medal, Jury Prize, Balkan Mathematical Olympiad, Athens, Greece
Silver Medal, International Mathematical Olympiad, Moscow, Russia
Study of Mathematics, Bucharest University
B. Sc. Diploma Thesis: Morse Inequalities on Covering Manifolds
Advisor: Prof. V. Iftimie
Post Graduate Studies, Bucharest University
Master Program: Nonlinear Equations and Modelling
Master Thesis: Spectral Theory for Uniformly Propagative Systems
Advisor: Prof. V. Iftimie
(April-June) Research Stay at Department of Mathematics,
Université Cergy-Pontoise, Paris, France
Ph. D. Student and Research Assistant, ETH Zurich
Advisor: Prof. Ch. Schwab
67