Professional Documents
Culture Documents
Analysis
This page intentionally left blank
F R O N T I E R S IN
APPLIED MATHEMATICS
The SIAM series on Frontiers in Applied Mathematics publishes monographs dealing
with creative work in a substantive field involving applied mathematics or scientific
computation. All works focus on emerging or rapidly developing research areas that
report on new techniques to solve mainstream problems in science or engineering.
The goal of the series is to promote, through short, inexpensive, expertly written
monographs, cutting edge research poised to have a substantial impact on the solutions
of problems that advance science and technology. The volumes encompass a broad
spectrum of topics important to the applied mathematical areas of education,
government, and industry.
EDITORIAL BOARD
H.T. Banks, Editor-in-Chief, North Carolina State University
Stanley, Lisa G. and Stewart, Dawn L, Design Sensitivity Analysis: Computational Issues of
Sensitivity Equation Methods
Vogel, Curtis R., Computational Methods for Inverse Problems
Lewis, F. L; Campos,].; and Selmic, R., Neuro-Fuzzy Control of Industrial Systems with Actuator
Nonlinearities
Bao, Gang; Cowsar, Lawrence; and Masters, Wen, editors, Mathemot/co/ Modeling in Optical
Science
Banks, H.T.; Buksas, M.W.; and Lin,T., Electromagnetic Material Interrogation Using Conductive
Interfaces and Acoustic Wavefronts
Oostveen, Job, Strongly Stabilizable Distributed Parameter Systems
Griewank, Andreas, Evaluating Derivatives: Principles andTechniques of Algorithmic
Differentiation
Kelley, C.T., Iterative Methods for Optimization
Greenbaum,Anne, Iterative Methods for Solving Linear Systems
Kelley, C.T., Iterative Methods for Linear and Nonlinear Equations
Bank, Randolph E., PLTMG:A Software Package for Solving Elliptic Partial Differential Equations.
Users'Guide 7.0
More, Jorge J. and Wright, Stephen J., Optimization Software Guide
Riide, Ulrich, Mathematical and Computational Techniques for Multilevel Adaptive Methods
Cook, L. Pamela, Transonic Aerodynamics: Problems in Asymptotic Theory
Banks, H.T., Control and Estimation in Distributed Parameter Systems
Van Loan, Charles, Computational Frameworks for the Fast Fourier Transform
Van Huffel, Sabine and Vandewalle, Joos, The Total Least Squares Problem: Computational
Aspects and Analysis
Castillo, Jose E., Mothematical Aspects of Numerical Grid Generation
Bank, R. E., PLTMG: A Software Package for Solving Elliptic Partial Differential Equations.
Users' Guide 6.0
McCormick, Stephen F., Multilevel Adaptive Methods for Partial Differential Equations
Grossman, Robert, Symbolic Computation: Applications to Scientific Computing
Coleman,Thomas F. and Van Loan, Charles, Handbook for Matrix Computations
McCormick, Stephen F., Multigrid Methods
Buckmasterjohn D., The Mathematics of Combustion
Ewing, Richard E., The Mothematics of Reservoir Simulation
Design Sensitivity
Analysis
Computational Issues of
Sensitivity Equation Methods
Lisa G. Stanley
Montana State University
Bozeman, Montana
Dawn L. Stewart
Air Force Institute of Technology
Wright Patterson AFB, Ohio
siam
Society for Industrial and Applied Mathematics
Philadelphia
Copyright © 2002 by the Society for Industrial and Applied Mathematics.
1098765432 I
All rights reserved. Printed in the United States of America. No part of this book may be
reproduced, stored, or transmitted in any manner without the written permission
of the publisher. For information, write the Society for Industrial and Applied Mathematics,
3600 University City Science Center, Philadelphia, PA 19104-2688.
TA174.S7622002
620'.0042-dc2l
2002075757
is a registered trademark.
Contents
List of Figuresh xi
Foreword xvii
Preface xix
1 Introduction 1
3 Model Problems 17
3.1 Heat in a Thin Rod with a Parameter-Dependent Forcing Term . . . . 17
3.2 Heat in a Thin Rod with a Parameter-Dependent Boundary 19
3.2.1 State Equation 20
3.2.2 Sensitivity Equation 20
3.3 Nonlinear Model 22
3.3.1 State Equation 22
3.3.2 Sensitivity Equation 23
4 Computational Algorithms 25
4.1 The Method of Mappings 25
4.1.1 Transformation Techniques 25
4.1.2 An Algebraic Transformation 27
vii
viii Contents
4.2 SEMs 28
4.2.1 Hybrid SEM 28
4.2.2 Abstract Version of the Semianalytic Method 31
4.2.3 Applying the H-SEM to the Nonlinear Model 33
4.3 Approximation Framework 35
4.3.1 Variational Formulations 35
4.3.2 Piecewise Linear Finite Elements 37
5 Numerical Results 43
5.1 Linear Model 43
5.1.1 State Approximations 43
5.1.2 H-SEM 44
5.1.3 A-SAM 47
5.2 Nonlinear Model 49
5.2.1 Convergence of the State and the Sensitivity 49
5.2.2 Sensitivities in Optimal Design 52
5.3 State Gradient Approximations 60
5.3.1 A Global Projection Scheme 61
5.3.2 A Local Projection Scheme 62
5.3.3 Numerical Results 62
Bibliography 135
Index 139
This page intentionally left blank
List of Figures
xi
xii List of Figures
xv
This page intentionally left blank
Foreword
Sensitivity analysis consists of a set of tools that can be utilized in the context of optimization,
optimal design, or simply system analysis to assess the influence of parameters on the state of
the system. Continuous sensitivity methods construct a sensitivity equation whose solution
provides the sensitivity of the state variable with respect to a parameter. To utilize such
information, computational methods are typically applied. This book discusses certain
pitfalls that can arise when sensitivities are computed for systems whose sensitivities exist
in a function space. In addition, it carefully presents the theory and motivating examples
for avoiding these pitfalls.
As computing power has progressed, computational methods have opened the way to
making sensitivity analysis a tractable design tool. It is this aspect of sensitivity analysis that
this book addresses. A first (and perhaps obvious) step to computing sensitivities might be to
approximate the state equation and then to use this approximation to compute sensitivities.
Although this seems like a benign order of steps (and one often applied), the authors show
that errors can arise that are sometimes catastrophic with respect to computational accuracy.
The book presents illuminating examples of what can go wrong in computing sensi-
tivities when the domain is a function space. Although the examples are for spatial variables
in one and two dimensions, they give insight that should lead to further research in more
complicated domains. This work is application unspecific; the authors provide concrete
example problems to point out some of the problems that can arise and what improve-
ments can be obtained. There are some simple one-dimensional partial differential equation
problems and some more sophisticated two-dimensional steady state flow problems. The
authors consider how to obtain accurate state gradient calculations, and why accuracy may
be compromised if one computes sensitivities of computational models without consider-
ing the underlying distributed parameter system. The authors present examples in which
computing the sensitivity by approximating the continuous sensitivity equation and solving
can be different from computing the sensitivity of the approximation to the state equation.
A major advantage of continuous sensitivity methods as presented in this book is the great
increase in computational speed.
The intended audience for this work is engineers, scientists, and applied mathemati-
cians who want to use continuous sensitivity equations in a computational setting. An
undergraduate knowledge of analysis and numerical analysis is assumed along with a fa-
miliarity with finite element methods. Necessary concepts and results for Hilbert spaces are
provided or referenced for the reader who is unfamiliar with these ideas.
The editorial board of this book series believes this edition is timely and will prove
most helpful to the reader who wants to use sensitivity analysis in a computational context.
xvii
This page intentionally left blank
Preface
This book provides an introduction to the computational aspects of continuous sen-
sitivity equation methods (CSEMs) for partial differential equations. It is intended for
advanced undergraduate and graduate students in the areas of numerical analysis and ap-
plied and computational mathematics as well as other scientists and engineers who have a
particular interest in modeling, design, control, and optimization of physical systems.
With the significant advances in computer technology during the twentieth century, in-
dustries that design and manufacture high-performance products are increasingly interested
in exploiting the advantages of computer-aided design, numerical analysis, and optimal
design methods. For example, a detailed analysis of aerodynamic systems, automated man-
ufacturing processes, and casting or molding processes can be performed using a software
package prior to time-consuming, expensive, and labor-intensive physical experiments.
During the last decade, considerable effort has been devoted to the development of com-
putational tools that allow designers to mathematically analyze such physical processes.
Sensitivity analysis plays an important role in this effort. Many simulation tools can be
enhanced by using local information provided by state sensitivities to obtain nearby solu-
tions. Hence, one is able to develop a qualitative picture of how small changes in design
parameters might affect the performance of a physical system.
This book is intended to give the reader an overview of applications and computational
issues regarding sensitivity calculations performed using CSEMs. The focus is on the
construction and analysis of algorithms for computing sensitivities. Consequently, adjoint
methods and other methods used primarily in the context of optimization applications are
omitted from this text. Myriad algorithms have been developed for the computation of
sensitivities. In Chapter 1, a number of these methods are briefly discussed. The methods
are addressed in sufficient detail to illustrate their respective advantages and disadvantages.
A number of schemes fall under the heading of CSEMs, and these methods have been applied
to a variety of problems. The term sensitivity equation methods refers to a broad class of
computational methods that compute a sensitivity by deriving and solving an equation
known as the sensitivity equation. This text focuses on the use of CSEMs, which refers to
the formal mathematical differentiation of governing equations (usually partial differential
equations) to derive the sensitivity equations (which also take the form of partial differential
equations). Chapter 1 gives a general overview of this approach and Chapter 2 provides
the appropriate definitions to answer the question of what exactly is meant by the term
sensitivity. Essentially, the sensitivity is interpreted as a partial derivative with respect to a
design parameter. Chapters 2 and 6 present the mathematical framework developed by the
authors for the purpose of rigorous mathematical justification of the formal differentiation
xix
xx Preface
techniques that form the base of the CSEM concept. These chapters are very mathematical
and are intended for readers with a background in real analysis and function analysis. A
general framework is given for a class of linear elliptic boundary value problems in Chapter
2, while Chapter 6 provides an abstract framework for Navier-Stokes equations.
Chapter 3 introduces the one-dimensional model problems used throughout the book,
to provide the reader with a relatively simple platform for absorbing the concepts presented.
Furthermore, these model problems allow the authors to demonstrate numerical techniques
in a simple format and to illustrate some subtle computational issues in subsequent chapters.
The models are constructed in detail and are related to the real-world applications on which
they are based. In the course of outlining these problems, many of the fundamental char-
acteristics of CSEMs are discussed. For example, an inherent by-product of CSEMs is the
coupling of the sensitivity equation (partial differential equation) to the original state equa-
tion (partial differential equation). This coupling can take the form of sensitivity boundary
conditions that require state gradient information, or, in the case of nonlinear state equations,
the sensitivity equation may include terms that involve the state variable or the gradient of
the state variable.
Chapters 4 and 5 focus on the computational algorithms used for sensitivity calcu-
lations throughout the book. Computational schemes are developed for both linear and
nonlinear elliptic problems. The method of mappings is incorporated into the CSEM as an
approach to computing shape sensitivities for a class of heat transfer applications. Numeri-
cal results for such algorithms are explored and discussed in significant detail in Chapter 5.
To develop a framework for fluid flow applications for both high and low Reynolds number
situations and for applications involving boundary layers, considerable effort is concentrated
on the analysis of a nonlinear model problem. In the case of nonlinear state equations, terms
involving state gradients often appear as coefficients in the sensitivity equations. This phe-
nomenon produces a special set of obstacles when the goal is to quickly and cheaply obtain
accurate sensitivity approximations. Indeed, one of the model problems illustrates that ob-
taining accurate state gradient approximations is of fundamental importance for sensitivity
computations in the presence of a boundary layer. Hence, an entire section of Chapter 5 is
devoted to the topic of extracting and improving state gradient approximations. Projection
techniques first developed for a posteriori error estimators in finite element codes are used
to improve the accuracy of these state gradient approximations. In addition, Chapter 5
contains a brief section that explores the usefulness of sensitivity approximations within the
framework of an optimal design problem.
Finally, Chapters 7 and 8 investigate the validity of the CSEM approach for some
examples of two-dimensional fluid flow problems. In Chapter 7, the continuous sensitivity
equation method is shown to be numerically efficient for computing sensitivities for flow
around a cylinder and for flow over a bump. However, Chapter 7 points out that, in certain
instances, the accuracy of the sensitivity calculations can be greatly improved by taking
advantage of adaptive refinement strategies that may already be available within the flow
solver producing the state approximations. Chapter 8 presents an example that combines
adaptive refinement with projection techniques in order to provide exceptionally accurate
sensitivity approximations for nominal extra computational expense.
The authors would like to take this opportunity to thank the many people and agencies
who have made this book possible. Lisa Stanley would like to express her appreciation
to the Air Force Office of Scientific Research for the AASERT fellowship that provided
Preface xxi
funding for research and travel during 1997-1999. Major Dawn Stewart is also grateful to
the Air Force for granting its permission for this book to be written. The Interdisciplinary
Center for Applied Mathematics provided both authors with the computer facilities, the
office facilities, and, most important, the mentoring necessary to carry out the research
component of their dissertations. These mentors include the thesis advisor Dr. John A. Burns
and other friends and colleagues, such as Dr. Jeff T. Borggaard, Dr. Gene M. Cliff, and
Dr. Belinda B. King. A special note of thanks to Dr. Dominique Pelletier for allowing the
authors the opportunity to use and modify the code used to generate steady state solutions
and sensitivities for Navier-Stokes equations. The authors are also extremely grateful to
everyone at SLAM, especially Dr. Tom Banks, Editor-in-Chief of the Frontiers series, and
Marianne Will, Acquisitions Editor, for providing the opportunity to publish work in the
well-respected Frontiers series. Lisa Stanley would also like to acknowledge the Burns
Telecommunication Center, in particular Ritchie Boyd, on the campus of Montana State
University for the computer facilities needed for some graphics editing.
Finally, this section would not be complete without the recognition of our families for
their constant encouragement, support, patience, and never-ending love and understanding
throughout the years.
Lisa G. Stanley
Dawn L. Stewart
This page intentionally left blank
1 Introduction
1
2 Chapter 1. Introduction
be exploited to produce the best overall computational tool for a specific application. For
example, if sensitivity calculations are used to enhance a simulation tool, accuracy may
be more important than speed. However, when used for gradient computations within an
optimization algorithm, speed may be more crucial than high accuracy. Hence, the scientific
motivation behind the sensitivity calculation can often influence the choice of computational
algorithm.
Many numerical methods have been developed for the purpose of sensitivity compu-
tation. The first, and maybe the best known, is the finite difference technique. The method is
very simple to implement and has a background that is fundamentally established in mathe-
matics. However, for complex fluid flow problems where each state approximation requires
considerable computational resources, this method requires two flow solves, each at differ-
ent parameter values, to form a difference quotient. In large aerospace flow problems, this
is sometimes an extremely severe requirement, since mesh generation itself can take weeks
to months. In addition, a step-size sufficiently small to ensure accurate approximations is
not generally known a priori. Such obstacles can make finite difference techniques imprac-
tical in many situations and difficult to incorporate into optimization schemes. A procedure
for estimating an optimal step-size is described in [36]. The optimal step-size attempts
to balance truncation error and round-off error. Unfortunately, since the discretization is
generally nonuniform, the truncation error varies from point to point so that the results must
be used in combination with some other decision algorithm.
Another early numerical technique was the "discretize-then-differentiate" approach.
This scheme approximates sensitivities by first employing some discretization scheme to
approximate the solution to a partial differential equation and then implicitly differentiating
this result to obtain a sensitivity approximation scheme. Recent advances in this area
include the development of automatic differentiation packages (see [25], [47]), such as
ADIFOR. Given source code for state calculations, this package generates source code for
sensitivity calculations. Once a computer code for generating a finite-dimensional state
approximation has been constructed, the code is viewed as a sequence of compositions
of elementary functions, and differentiation is achieved by clever and repeated use of the
chain rule. Automatic differentiation has been applied to shape optimization problems and
optimal design problems (see [16] and [34]), and the ideas have been considered and refined
for a variety of other applications; see [26], [27], [33]. For more references and a general
discussion comparing this method with other shape optimization algorithms, see [22]. A
disadvantage of the discretize-then-differentiate technique is that in cases in which the mesh
is parameter dependent, as is the case in shape-optimization problems, then differentiation
of the discrete PDF leads to mesh sensitivities on the right-hand side. Although much work
has been done in recent years to get a handle on these quantities (see [53]), calculating
mesh derivatives is still not well understood, particularly in cases in which the meshes are
prescribed adaptively.
Several new techniques have been developed in an attempt to alleviate some of the dis-
advantages of the two approaches mentioned above. The idea is to employ a "differentiate-
then-discretize" scheme, or a so-called CSEM, which consists of implicitly differentiating
a partial differential equation to obtain a sensitivity equation and choosing an appropriate
discretization scheme based on that equation's structure. Then both the partial differential
equation and the sensitivity equation are discretized to obtain finite-dimensional equations
for numerical approximations to the partial differential equation and the sensitivity equation.
Chapter 1. Introduction 3
As demonstrated in this work, a number of schemes fall under the heading of CSEMs. Early
use of the sensitivity equation for optimal design purposes is seen in [38], and theoretical
framework describing sensitivity functions can be found, e.g., [29] and [42]. Variants of
sensitivity equation methods have been applied to a variety of problems. In [23], Godfrey
investigates several uses of sensitivity equation methods (SEMS) for the analysis of chemi-
cally reacting flows in aerodynamic applications. SEMs are used to compute aerodynamic
stability derivatives in [24] and [45]. Optimal control of fluid flows and shape optimization
algorithms using these methods can be found in [7], [8], [9], [10], [11], [12], [14], [28].
The references given here are by no means comprehensive or exhaustive, but they serve as
a starting point for the interested reader.
As is the case with most new research, the CSEM approach provides its own set of
advantages and disadvantages. Although the CSEM has a very mathematical foundation,
derivation of the sensitivity equation can be difficult at times. Parameters can enter a
model in a variety of ways, and much of the analysis still must be done on an individual
problem basis for many cases. Identifying the correct boundary conditions for the sensitivity
equation can be tricky. The mathematical constructions in Chapters 2 and 6 certainly provide
a platform for further analysis. However, developing a comprehensive mathematical theory
for CSEMs will require much more sophisticated tools than those presented here. Even for a
relatively small class of elliptic partial differential equations, there is no blanket approach to
sharply characterizing the smoothness of the sensitivity. During the course of our research,
we have seen that smoothness may be gained, lost, or maintained when one compares the
smoothness of the state to that of the sensitivity. Once a firm mathematical foundation is
developed, the advantage of the CSEM approach is that it will provide insight into situations
where the sensitivity equation must be interpreted in a weak sense. This may be the case
with discontinuous coefficients or forcing functions when the discontinuities are somehow
dependent on the parameter of interest. Another area where CSEMs may be useful is in
optimal design. CSEMs can provide cheap and accurate sensitivity approximations for
gradient calculations. However, the issue of consistent gradients is far from resolved; see
[7]. In addition, optimal shape design is an area in which the authors would like to make
a contribution. This topic has long been linked to variational problems through study in
the area of the calculus of variations; see [39], [40], and [41] for applications focusing
on structural design optimization. Optimal shape design encompasses a mathematically
challenging set of problems as the domains of differential operators become parameter
dependent. In this setting, the concept of differentiation of an operator becomes quite
difficult to characterize in certain instances.
Clearly, the advantages, disadvantages, and potential applications of CSEMs are still
revealing themselves as the research continues. This book is a compilation of the authors'
dissertation work while attending Virginia Polytechnic Institute and State University. Some
of the numerical results in this book were generated using MATLAB on various UNIX plat-
forms, including SUN workstations and Silicon Graphics Indigo and Indigo 2 workstations.
The two-dimensional fluid flow results were generated using an adaptive finite element code
developed by Dominique Pelletier and other researchers at Ecole Polytechnique de Mon-
treal. The code was modified to calculate sensitivities at Virginia Polytechnic Institute and
State University by Dr. Jeff Borggaard and Dr. Dawn Stewart.
This page intentionally left blank
2 Mathematical Framework for
Linear Elliptic Problems
This chapter surveys the mathematical background needed to analyze the regularity of a
particular class of linear elliptic partial differential equations, which includes some of the
model problems introduced in Chapter 3. In particular, the state equation takes the form of
a linear elliptic boundary value problem defined on a parameter-dependent domain. First,
we give the general formulation of the class of state equations to be considered. We then
present the necessary notation and framework to provide the existence and uniqueness of
both strong and weak solutions for the general form. Finally, we discuss the implications
this framework has for the derivation of sensitivity equations. Readers who are interested
in an overview of the general approach of SEMs may choose to skip this chapter on a first
reading; it contains a number of technical mathematical details and is not essential reading
for those who seek a general exposure to SEMs. However, readers who are already familiar
with the topic may find that the details of this chapter help to solidify their understanding
of the mathematical concepts used in the derivation of sensitivity equations.
Let Q c Rm denote an open domain of design parameters. For each q Q, assume
that C Rn, n = 1,2, or 3, is a domain satisfying the cone condition with the
boundary (q) such that locally, (q) lies on one side of (q). Consider the elliptic
boundary value problem
where each of the data K(x , q), f ( x , q), and g(x, q) may depend on the design parameter
q. Throughout this chapter, we assume that the coefficient K ( x , q) is a real-valued function
satisfying 0 < m1 K ( x , q) m2 for all x (q), q Q. The long-term research goal
is to mathematically characterize the dependence of the state w(x, q) on small changes in
the design described by q.
2.1 Preliminaries
This section introduces some preliminary concepts and results that are used throughout
the following three chapters. The trace theorems for Sobolev spaces are presented in the
5
6 Chapter 2. Mathematical Framework for Linear Elliptic Problems
context needed for this work. Variational tools such as bilinear forms, Gelfand triples, and
their related properties are reviewed, and the results from the theory of partial differential
equations relevant to this work are surveyed. The authors concede that the material in
this chapter gives only an abbreviated introduction to the relevant mathematical concepts,
and the reader should be cautioned that some of the material relies on knowledge of basic
concepts from functional analysis, such as linear operators, inner-product spaces, and duality
pairings. Additional reading for the related functional analysis topics includes [43], [52],
and [55].
2.1.1 Notation
Most of the notation used throughout the book is consistent with that of Wloka [54]. We
begin with some general geometric assumptions. Sobolev spaces as well as their inner
products and related norms are described. Let Q represent a bounded, connected, open
subset of R", n = 1, 2, or 3, satisfying the cone condition. Denote the boundary of by
We assume that is piecewise smooth and that, locally, lies on one side of
Unless otherwise stated, the variable x denotes the spatial variable or vector of spatial
variables. The state and sensitivity variables are denoted by boldface letters such as u, w, and
s to indicate that this variable may represent a vector of scalar variables. We assume a finite
number of design parameters, and the notation q denotes the vector of design parameters.
The design space is denoted by Q. Partial derivatives with respect to either the spatial
variable or to a parameter are denoted as
The derivatives referred to in this work are Frechet derivatives unless otherwise specified.
The following definition is given as a reminder and to preface the mathematical framework
constructed in later sections: Let (X, \\ • \\ X ), (Y, || • ||y), and (Q, \\ • \\Q) be normed linear
spaces, and denote the set of all bounded linear operators mapping X to Y by B(X, Y).
Let G B(X, Y). Then G is said to be Frechet differentiable at x0 X if there exists a
continuous linear operator D x G(x 0 ) B(X, 7) such that for every h X with XQ + h X,
Frechet derivatives along with other notions of differentiation are discussed in [46]. If X,
Y, and Z are normed linear spaces, then the product space X x Q has norm defined by
and
respectively.
Let the space of infinitely differentiable functions defined on be denoted by C
Likewise, the subset of these functions that have compact support in is denoted by
It is well known that the set C Hm is dense in Hm and the completion of
C with respect to the norm \\ • \\m is defined to be the set Hm0 . The dual space of
Hm0 , that is, the set of all bounded linear functionals defined on H0 m is denoted as
H-m . For a general Hilbert space, V, we denote its dual space as V*. Recall that the
norm on the dual space is the operator norm given by
to depict the duality pairing. In general, we use the notation {•, •} for both inner products
and duality pairings. The subscripts are used only when this notation could be confused
with that of an L2 inner product, and we remark on the specific meaning of the notation
whenever the usage is ambiguous.
8 Chapter 2. Mathematical Framework for Linear Elliptic Problems
The domain is said to be of class C if the region is (k, )-smooth. That is,
for each point x € , there exists a neighborhood Ux of x that can be transformed by a
(k, )-diffeomorphism to the unit ball in Rn in such a way that the boundary is in one-
to-one correspondence with the central plane xn = 0 and so that Ux lies on one side
of the central plane while Ux lies on the other side of the central plane. The notation
denotes the complement of the closure of . One can think of this diffeomorphism
as a mapping that maintains an orientation with respect to the boundary . Now that the
groundwork has been laid, we assume that the domain belongs to the class C . The
appropriate values of k and are assigned as we consider each class of solutions in the
following sections.
We summarize the trace theorems taken from section 8 of [54] in the following results.
Theorem 2.2 (trace theorem). Suppose that is (k, )-smooth, and let l — m > , with
m a nonnegative integer and l + 1 < k + . Then there exists a continuous linear trace
operator
For the case that we are only interested in the mapping T0 ,we can relax the restriction
l + 1 < k + to l < k + X. One can also characterize the kernel of the trace operator
introduced above. We condense this theorem to concentrate only on the particular result
needed for the analysis in the following sections, and we also need the inverse theorem
found in [54].
Theorem 2.3. Suppose that is (0, 1)-smooth. IfT0: Hl(ti) -> H1/2( is the trace
operator given above, then
Theorem 2.4 (inverse theorem). Let bea(k, ) -smooth domain, and let
Then there exists a continuous, linear extension operator
More detailed results concerning Sobolev spaces, trace operators, and related subjects
can be found in [1].
Definition 2.5. Let V and H be Hilbert spaces. Suppose that the mapping i: V —> H is a
continuous, injective embedding with Im(i') dense in H. Then the embedding i': H —> V*
described by
is continuous and injective and Im(i') is dense in V*. The preceding construction
The assumptions on V can be relaxed to a reflexive Banach space, but for the following
sections, the restriction to Hilbert spaces is sufficient. Since Im(i') = i'(H) is dense in V*,
we can uniformly approximate each linear functional (on the unit ball) in V* by the inner
product on H. That is, for each / V*, we can write
The completeness of H and the continuity of i' imply the existence of h1 H such that
10 Chapter 2. Mathematical Framework for Linear Elliptic Problems
and
Definition 2.7. The bilinear form, a(•, •), is V-elliptic if the following conditions hold:
and
If a: V x V —> R is a continuous bilinear form, that is, if (2.7) holds, then a uniquely
determines a continuous linear mapping L: V —> V by
and || a || = ||L||. The proof of this statement is a consequence of the Riesz representation
theorem, and a more general version can be found in Theorem 17.8 of [54]. Moreover, let
(h 1 , i(v))H denote the representation (2.4) of an arbitrary linear functional l V*. Then by
the Riesz representation theorem, we have the relation
where R is the Riesz isomorphism R: V* V. Using this relation along with (2.9), we
obtain
And A is said to be V-elliptic if the bilinear form a(•, •) is. The construction given above
leads us to the Lax-Milgram theorem, which provides the existence and uniqueness of weak
solutions to elliptic boundary value problems in the following sections.
2.2. Regularity of Elliptic Boundary Value Problems 11
Theorem 2.8 (Lax-Milgram theorem). If the bilinear form a(x, y) is V-elliptic, then the
corresponding operator A: V —> V* given in (2.11) is a linear topological isomorphism
between the spaces V and V*. The norms are bounded in the following way:
where c\ and c2 are the constants given in Definition 2.7. Or, equivalently, for each f V*,
the variational equation Ax = / has a unique solution in V, and this solution depends
continuously on f.
Many forms of this theorem are given in the literature; see [5] and [21]. This particular
version is found in [54]. Once these fundamental results have been introduced, we are ready
to move to the regularity results for the general form of the state equation given in (2. l)-(2.2).
As noted above, we assume for the moment that the domain is independent of the parameter
q. In section 2.3 we describe how these results can be used, in conjunction with mapping
techniques, to treat differential equations defined on parameter-dependent domains. Specific
assumptions on the coefficient k(k, q) and on the domain and the boundary are given
in each section. To simplify the discussion further, we first transform (2.13)-(2.14) into a
problem with homogeneous boundary conditions. Assume that wg(•) H1 satisfies
where T0 is the trace operator from section 2.1.2. Observe that if g then the trace
theorem implies the existence of such that (2.15) holds. If
then the trace theorem implies that there exists satisfying (2.15). In the
more general setting, if then the precise meaning of (2.15) requires some
discussion that is beyond the scope of this book. In the case of sections 2.2.1 and 2.2.2, the
preceding justification is sufficient for the formulation of the homogeneous problem. Let
z = w — wg; then z satisfies (at least formally)
12 Chapter 2. Mathematical Framework for Linear Elliptic Problems
To say that z is a solution to (2.20) is equivalent to asserting that (2.16)-(2.17) has a strong
solution.
forms a Gelfand triple, with the identity (inclusion) being the map from
For a fixed value of the parameter q in (2.1), define the bilinear form
2.2. Regularity of Elliptic Boundary Value Problems 13
R by the following:
The Cauchy-Schwarz and Poincare inequalities can be used to show that this bilinear form
is continuous and -elliptic. In particular, there exist constants m2 and c such that
and
Here, m2 is the upper bound placed on the coefficient k(x, q) for all x and for all
q Q. The constant c involves the lower bound on the coefficient k(k, q) as well as the
constant prescribed by the Poincare inequality. See [1], [13], and [54] for details regarding
these results as well as many other theoretical results related to variational formulations of
partial differential equations. We are now ready to define the term weak solution.
Using the bilinear form defined above, one can construct the operator A: D(A)
where and
Note that the variational equation given in (2.22) is equivalent to the operator equation (in
This work focuses on state equations with Dirichlet boundary conditions; however,
section 21 of Chapter 3 in [54] contains several examples of elliptic equations with various
types of boundary conditions treated using variational techniques. The reader may also
14 Chapter 2. Mathematical Framework for Linear Elliptic Problems
choose to consult several other classical texts, such as [2], [20], and [37], for general
expositions on the theory of partial differential equations.
In the previous sections, we constructed a mathematical framework for analyzing
the regularity of elliptic boundary value problems on a fixed domain. However, we want to
apply these concepts to equations defined on parameter-dependent domains in order to derive
sensitivity equations for shape parameters. With this goal in mind, we now transition to the
method of mappings that provides a mechanism for transforming equations on parameter-
dependent domains to equations defined on fixed domains, thereby allowing us to use the
results of this chapter to analyze sensitivity equations.
Theorem 2.13 (implicit function theorem). Assume Z, Y, and Q are Hilbert spaces, and
let G: D(G) C Z x Q —> Y be continuous on a neighborhood U of the point ( Z 0 , q0)
int[D(G)]. If
1. G(z0,q0) = 0,
A more general version of this theorem can be found in Zeidler's book [56].
2.3. Mappings and Implications to Sensitivity Analysis 15
where C: Q (Z, Y) is a continuous linear operator, Z and Y are Banach spaces, and
F: Q —> Y. One can further generalize this equation by defining the operator G: Zx Q
Y as follows:
With each value of the design parameter q, we associate the corresponding state that solves
(2.26) to form the pair (z(q), q) Z x Q such that
That is, for a fixed q Q, we characterize the state z(•, q) as the solution to the operator
equation (2.28). This induces a natural mapping z: Q Z described by
then it is straightforward to show that L (q0) is a bounded linear operator from Q to Y. Further
suppose that D q F(q 0 ): Q Y exists in the Frechet sense. Then the partial derivative of
G with respect to q at the point (z0, q0) exists, and [ G(z0, q 0 ) ] : Q Y is given by
Theorem 2.14. Let q0 int( Q) be fixed. Suppose that there exists a unique Z 0 (q 0 ) int(Z)
so that G(z0(g0), q0) = 0. Further suppose that [L(q 0 )] -1 exists in B(Y, Z). If the Frechet
derivatives Dq£(q0) and D q F(q 0 ) exist in B(Q, B(Z, F)) and B(Q, 7), respectively, then
the sensitivity p = Dqz(+qo): Q —> Z exzste and satisfies the operator equation
Proof. Equation (2.33) follows from applying the operator equation given in (2.25) of
Theorem 2.13 to equations (2.30) and (2.32) given above.
Theorem 2.14 implies that for a given parameter q0, the sensitivity p(., q0) belongs
to the function space Z, and it is the solution to the operator equation (2.33). At the
beginning of the following chapter, we present an example to demonstrate the derivation
of the sensitivity equation. It is important to note that the choice of function spaces Z and
Y is nontrivial. Moreover, one can construct relatively simple state equations for which
the preceding theorem does not apply. In the next chapter, we present some examples and,
where appropriate, relate the examples with the framework presented here.
3 Model Problems
This chapter describes three examples of state equations used throughout the book to explore
the concept of a sensitivity and to illustrate several of the mathematical issues involved in
its computation. For clarity of presentation, we have chosen state equations whose spatial
domains are one-dimensional. In addition to their simplicity, these examples are convenient
for comparing numerical results since both the state and the sensitivity equations have
analytical solutions. They also provide the reader insight into several aspects of sensitivity
computation that are inherent to our approach. The first example is used to illustrate the
use of the mathematical framework outlined in the previous chapter. A state equation is
introduced, the operator formulation is given, and an operator sensitivity equation is derived.
The remaining model problems—one is linear and one is a nonlinear boundary value
problem—provide us with one-dimensional examples of state equations with parameter-
dependent domains. The state and sensitivity equations are outlined along with their respec-
tive closed-form solutions. In subsequent chapters, several aspects pertaining to sensitivity
analysis and the implementation of CSEMs are explored for each of these problems. Com-
putational methods for numerically approximating the solutions to the sensitivity equations
are described, and numerical results are given.
17
18 Chapter 3. Model Problems
Using results from section 2.2.2, it can be shown that the boundary value problem (3. l)-(3.2)
possesses a strong solution in H 2 (0, 1) for each q Q = (1,2). Define the sensitivity
The mathematical framework described in the previous chapter can be used to derive the
corresponding sensitivity equation for p(., q). Here, we illustrate the importance of choosing
the function spaces Z and Y correctly, and we note that a formal derivation can often provide
the insight necessary for this choice.
The first approach might be to let Z == H2(0, 1) H10(0, 1) and Y = L2(0, 1). In
this case, the operator £: Q -> B(H2(0, 1) H10(0, 1), L2(0, 1)) is defined by
for each q Q, where A0 is defined in (2.19) in section 2.2.1. In this case, £ = £(q 0 ) = A0
is constant with respect to the parameter q. Furthermore, the operator F: Q L2(0, 1) is
defined by
Using this framework, the operator G: Z x Q L2(0, 1), and the operator equation given
by
G(w,q) = £(q)w-F(q) = 0
is equivalent to (3.1)-(3.2). Clearly, the operator £ ( q ) - l = [A0]-1 exists in (L2(0, 1),
H 2 (0, 1) H10(0, 1)) as it is a special case of the differential operator A0 given in sec-
tion 2.2.1. The structure of L implies that its Frechet derivative is the zero operator in
B(Q, B(Z, Y)). However, the derivative DqF(q) does not exist in the L2-sense, and Theo-
rem 2.14 is not applicable for this construction.
However, if one chooses the spaces Z = H 1 0 (0, 1)and Y = H - l ( ), then the operator
£(q): Q B(H 1 0 (0,1), H-l is defined by
The A operator is defined in (2.23). It follows from the framework presented in section 2.2.2
that A-l = [£(q) - 1 exists in B ( H - 1 H 1 0 (0, 1)), and the Frechet derivative D q L(q 0 )
1 -l
is the zero operator in B(H 0 (0, 1), H Furthermore, one now defines the operator
F: by
G(w,q) = £(q)w-F(q) = 0
3.2. Heat in a Thin Rod with a Parameter-Dependent Boundary 19
is equivalent to (3.1)-(3.2). Using this construction, one can now represent the Frechet
derivative of F with respect to q in the appropriate fashion. In particular, for a given qo,
DqF(qo): Q H - l ( ) exists and is given by
for all u(•) H10(0, 1) and for each h Q. Here, is the Dirac delta function
with mass at x = The hypotheses of Theorem 2.14 are now satisfied. Therefore, the
sensitivity p exists (in H 1 0 (0, 1)) and satisfies the operator equation (in
-l
H (0,1)) given by
That is, for any qo Q, the sensitivity equation is well defined and is given by the variational
equation
Furthermore, the sensitivity p(•, q0) H10(0, 1) is the unique solution to (3.5). From the
derivation of the operator sensitivity equation, one observes that the differential operator
describing the state equation, A in the above paragraphs, is also the differential operator
associated with the sensitivity equation. Hence, the operator sensitivity equation is derived
with the appropriate mathematical rigor, and a corresponding differential equation, our
sensitivity equation, is constructed using the operator expression. In particular, p( , q)
satisfies the following differential equation
where is the Dirac delta function with mass at The boundary conditions for
1
are clear since p(•, q) H 0(0,1).
In the following sections, we consider other examples of state equations and their
associated sensitivity equations. Once the examples are introduced, we examine specific
mathematical and computational issues related to sensitivity analysis for these types of
problems. Although they are somewhat simple in nature, these model problems illustrate
many of the fundamental issues addressed in this book.
The goals are to solve (3.7)-(3.8) for the state, w(x, q), for a given value of q and to determine
the sensitivity of the state to small changes in the parameter. The forcing function, f(•), is
discontinuous on (0, 2) and belongs to the space L2(0, 2). For a fixed q (1, 2), the state
w(•, q) is a function belonging to H2(0, q) DH10(0,q). In particular, one can verify that
the analytical solution to (3.7)-(3.8) is
The sensitivity equation for this problem is discussed in the following section.
From (3.10), we see that for each q Q = (1, 2), the sensitivity is a linear function of x
and has the following form:
Note that for a fixed q Q, s(x, q) is a C function on (0, q). We now derive a differential
equation for which the function given in (3.12) is a solution.
Beginning with the state equation defined on the interval (q), one can implicitly
differentiate (3.7)-(3.8) to obtain a sensitivity equation. At this stage, it is important to
note that this differentiation is rather formal. In general, the partial derivatives w(x, q)
and w(x, q) need to be continuous to interchange the order of differentiation. In this
case w(x, q) is discontinuous, but one can verify by hand that the sensitivity s(x, q) =
w ( x , q ) satisfies the differential equation
3.2. Heat in a Thin Rod with a Parameter-Dependent Boundary 21
The sensitivity equation (3.13)-(3.14) is a linear elliptic boundary value problem with
Dirichlet boundary conditions. The boundary conditions should be derived with care as
the right endpoint of the domain (0, q) depends explicitly on the parameter q. Taking the
total derivative of the right boundary condition in (3.8) with respect to the parameter q leads
to the correct boundary condition for the sensitivity
For each q > 1, the solution to this boundary value problem is given by
where w(x, q) is the solution of (3.15)-(3.16) and the integer p represents the number of
data points. In gradient-based optimization one needs the derivative
to define the sensitivity. One approach to the evaluation of this gradient at a given q is to
compute the state, and the sensitivity, and to form the computation
(3.19). This involves first solving (3.15)-(3.16) for w(x, q) and then computing s(x, q).
3.3. Nonlinear Model 23
On the other hand, we can implicitly differentiate the boundary value problem (3.15)-(3.16)
and obtain a boundary value problem for the sensitivity s(x, q). It follows that s(x) satisfies
the linear differential equation
As in the previous example, the boundary conditions in (3.22) require the application of the
chain rule and should be treated with care.
This page intentionally left blank
4 Computational Algorithms
This chapter describes two computational methods for solving sensitivity equations. Note
that the model problems given in Chapter 3 are defined on parameter-dependent domains.
We address this issue here, and suggestions for dealing with this situation in the context
of numerical sensitivity approximations are given. In particular, mapping techniques are
discussed, and we show how these techniques can be blended with SEMs to numerically ap-
proximate sensitivities. We also show that there are some pitfalls associated with combining
these techniques.
25
26 Chapter 4. Computational Algorithms
and
are equivalent for any an integer). Furthermore, the trace operator, T0,
commutes with the operation of transforming; see page 131 in [54]. Finally, the theory
of elliptic differential operators from [54] also gives the important result that under the
appropriately smooth diffeomorphism, (strongly) elliptic operators defined on transform
to (strongly) elliptic operators defined on . This is also true for hyperbolic and parabolic
operators as noted in [32]. A general form for the relationship between the original operator
and the transformed operator is given on page 143 of [54]. We now address some of the
practical aspects of using transformations.
From a computational standpoint, transforming can be a complex process for partial
differential equations with spatial domains in R2 or R3. In addition to the issue of regular-
ity, one is concerned with loss of accuracy. The mappings are implemented by defining a
new coordinate system that maps to a fixed computational space, . Transformations
that produce coordinate systems that are orthogonal at boundaries are preferred. Other-
wise, accuracy decreases as orthogonality declines (see [53, pp. 3-5]). A two-dimensional
coordinate system on takes the form
Aside from the issue of orthogonality, T must also satisfy other mathematical properties. In
particular, T is required to be one-to-one, and the mapping should maintain some smoothness
in the distribution of grid points. Ideally, T should also provide some mechanism for
clustering grid points once a discretization has been constructed.
For problems involving complex geometries, determining the mappings can be a
complicated process. The transformations can be constructed using conformal mapping
methods, partial differential equation methods, or algebraic methods. Conformal mapping
methods are based on the theory of complex variables and are used only when working with
a physical domain given in two spatial dimensions. Partial differential equation methods
are also widely used. With this approach, the computational domain is rectangular with a
uniform grid, and the location of corresponding points in the physical domain is determined
through the solution of a system of linear, or nonlinear, partial differential equations. Elliptic
grid generators are the most commonly used. However, there are methods that use hyperbolic
or parabolic equations to determine the mappings. To determine a mapping, these equations
are usually solved numerically. Consequently, this method may introduce numerical errors
into the grid-generation aspect of the overall computational scheme. As we will see later, the
spatial derivatives of the transformations are needed to construct the transformed equations.
Hence, questions of invertibility and differentiability become more tedious to resolve when
the mappings are computed numerically. For this reason, computing the transformations
4.1. The Method of Mappings 27
numerically can be a drawback. However, for very complicated geometries, this approach
may be the only one suitable for numerical computation.
Finally, algebraic methods determine an algebraic expression which relates points
in with those in . Interpolation schemes are used to represent the interior points of
the domain in terms of the points along the boundaries. Implementation of these methods
is often simple and fast. Another advantage of algebraic methods is that the derivatives
of the transformations can be computed analytically, thereby reducing computational time
and avoiding the introduction of additional numerical inaccuracies into the computational
method. However, these methods may be difficult to implement for very complex ge-
ometries. An algebraic approach is used to construct the transformations presented in this
chapter.
Once the transformation is determined, the process of transforming the differential
equation begins. We remark here that the crucial step is to derive the relationship between
the differential operator defined on and the corresponding operator defined on £2. This
requires the computation of the derivatives of the mappings with respect to the spatial vari-
ables. An example of this process is given in the following section, and the details of
this relationship will be presented in the context of the example. For equations posed on
parameter-dependent domains, the mappings are also parameter dependent. In order to con-
struct the transformed equations on the computational domain, derivatives of the mapping
with respect to the spatial variables are computed, and these derivatives will also involve
the parameter. Hence, the design parameter often appears explicitly in the transformed
differential equation although the domain of definition no longer depends on the parameter.
This can be seen in the example given in the following sections.
One important issue for sensitivity calculations is that the choice of the computational
method can result in the need to also compute the partial derivative of the mapping with
respect to the parameter. Usually referred to as a "mesh sensitivity," this derivative is
often difficult or even impossible to compute in two-dimensional and three-dimensional
problems. In particular, this derivative is difficult to obtain when the transformations are
computed numerically or when adaptive algorithms are used. The mesh sensitivity appears
in the example presented in the following section. We comment further on this issue as we
examine the following example.
is the inverse of T and is commonly referred to as the mesh map. As noted earlier, trans-
forming can be a complex process for two-dimensional and three-dimensional problems.
For the one-dimensional examples given here, transforming is straightforward. Some of the
difficulties that occur in two-dimensional and three-dimensional problems, such as holes
and corners in the domain, are not present in this case.
4.2 SEMs
This section presents two examples of SEMs used for the numerical approximation of
sensitivities. We give a detailed description of the schemes used to obtain the numerical
calculations presented in subsequent chapters. The transformation techniques described in
section 4.1.1 are used to develop two computational methods for sensitivity approximation.
The computational methods presented in sections 4.2.1 and 4.2.2 use the linear model prob-
lem given in (3.7)-(3.8) as a platform for investigating the implementation of the mapping
techniques, in combination with SEMs, for the numerical approximation of sensitivities. In
section 4.2.3, we demonstrate how one of the SEMs can be applied to the nonlinear model
givenin(3.15)-(3.16).
As illustrated in this chapter, there are several ways to implement SEMs, and these
variations can yield algorithms with different convergence properties. We consider two
specific SEMs used in conjunction with the transformation techniques described in section
4.1. The first is based on transforming both the state and the sensitivity equations to the
computational domain, solving the transformed equations, and mapping these solutions back
to the physical domain. The second approach transforms the state equation and then derives
its sensitivity equation. Once the state and sensitivity systems are solved, the solutions are
mapped back to the physical domain. There are benefits and drawbacks to each method.
Indeed, it is not always obvious which scheme is best for a given problem. Many questions
need to be answered before a complete theory can be developed, and we address some of
these issues with the computational schemes and numerical results presented here. We now
describe each of the SEMs in general and in the context of the linear model problem.
a sensitivity equation that is also posed on the physical domain. This step is completed in
section 3.2.2. Once the state and sensitivity equations are described, the method of mappings
is applied. For the sake of clarity, the equations are given here. The state equation is given
by
with
The transformations defined in section 4.1.2 are used to construct the transformed
functions. The transformed state is denoted by (-, •); likewise, the notation (-, •) refers to
the transformed sensitivity. For and q (1,2), define
and
To determine the transformed boundary value problems on , one must also determine
the action of the forcing function (•) under the mapping. Once transformed, the original
forcing function becomes
which now depends explicitly on the parameter q. Using the above definitions and the chain
rule, the spatial derivatives of the original functions and those of the transformed functions
are related by
and
30 Chapter 4. Computational Algorithms
As noted in the previous section, the differential operators acting on the respective domains
are related through algebraic expressions involving the spatial derivatives of the mapping. As
a result, the transformed equations can be algebraically more complicated than their original
counterparts. The identities developed above are used to derive transformed boundary value
problems for both the transformed state and the transformed sensitivity defined in (4.2). The
transformed state equation is given by the differential equation
Note that equations (4.7)-(4.8) and (4.9)-(4.10) remain weakly coupled through the bound-
ary conditions of the transformed sensitivity equation, (4.10).
Once the transformations are constructed and the transformed equations are derived,
one can proceed to numerically approximate the solutions to these equations. The discretiza-
tion for this particular example is discussed in section 4.3. The H-SEM is summarized below
to illustrate the key components of the approach.
H-SEM
Step 1. Solve the transformed state equation (4.7)-(4.8) for
Step 2. Solve the transformed sensitivity equation (4.9)-(4.10) for
Step 3. Map back to the physical domain to obtain the sensitivity given by
Observe that by using (3.10), (3.12), and the definitions of the transformed functions given
in (4.2), the closed-form solutions to the transformed state and transformed sensitivity
equations are given by
and
4.2. SEMs 31
respectively.
We take a moment here to address the issue of regularity for the transformed equations.
Some of the notation and language used in this paragraph is clarified in Chapter 2. For a fixed
value of (1, 2), it follows from Theorem 2.10 that the transformed state equation given
in (4.7)-(4.8) has a unique strong solution Theorem 2.2 implies that the
trace of the function at the boundary can only be interpreted as a function
belonging to the function space Hence, the boundary data (4.10) of the transformed
sensitivity equation only satisfy the requirements of Theorem 2.12, and the transformed
sensitivity equation is only guaranteed to possess a weak solution for a
fixed value of the parameter. In this one-dimensional example, the transformed sensitivity
is actually a strong solution to the differential equation. In particular,
However, in higher dimensions, the appearance of the trace of the gradient of the state
within the boundary conditions of the sensitivity equation is an issue that presents not only
theoretical but also computational dilemmas. The computational issues are addressed in
this as well as in later chapters. First, we describe an alternative approach for obtaining a
sensitivity equation.
we apply the framework given in Chapter 2, and the sensitivity, p( , q), satisfies the differ-
ential equation
where is the Dirac delta function with mass at The reader is referred to section
3.1 for the details.
For this example, can be calculated directly from (4.11). In particular,
Clearly, the transformed sensitivity is not equal to the sensitivity of the transformed state;
that is, Observe that the sensitivity of the transformed state,
is less smooth than the transformed sensitivity in (4.12). Furthermore, the
relationship between the transformed sensitivity and the sensitivity of the transformed state
can be derived using (4.5)-(4.6) and the chain rule. Beginning with the definition of the
sensitivity of the transformed state, it follows that
The relationship between s(x, q} and can be obtained by using (4.2) and the definition
of the transformation T. Direct computation yields
Observe the appearance of the mesh sensitivitv. and the spatial derivative of
in this equation. For this example, can be calculated analytically using
algebraic techniques. However, for two-dimensional and three-dimensional problems, the
4.2. SEMs 33
transformations are often constructed using numerical algorithms such as the partial differ-
ential equation method discussed earlier. Obtaining derivatives of these maps can be very
difficult, and as shown in the following sections, the computation of the spatial derivative
can also affect the quality of the sensitivity approximation. Below is an outline of
the key steps for this method.
A-SAM
Step 1. Solve the transformed state equation (4.7)-(4.8) for
Step 2. Solve (4.15)-(4.16) for the sensitivity of the transformed state
Step 3. Map back to the physical domain and obtain the sensitivity s(x, q) using
Observe that in order to calculate s(x, q) using the A-SAM approach, one generally
needs to compute not only the sensitivity of the transformed state but also the spatial deriva-
tive of the transformed state and the mesh derivative As mentioned, the mesh
derivative is calculated analytically for this example.
Turning to the mathematical issues of existence and regularity of these sensitivities,
we note that (4.18) provides an avenue with which to pursue theoretical issues regarding the
original sensitivity s(-, •) by means of the sensitivity of the transformed state p(-, •)• That
is, we may analyze the existence and regularity of s(-, •) by first examining these issues
for p(-, •) and then applying those results to s(-, •) through the equation (4.18). In some
situations, such as the current example, this can be advantageous. Recall that the existence
of and the rigorous derivation of the corresponding sensitivity equation are discussed
in section 3.1. Before moving to the numerical solution of the state and sensitivity equations
for each of the methods previously detailed, we first illustrate how the H-SEM approach
can be applied to the nonlinear model problem given in section 3.3.1.
After applying this transformation, we obtain the transformed state equation defined on the
computational domain and given by the Dirichlet problem
34 Chapter 4. Computational Algorithms
and
and
respectively. When applying the inverse transforms to the numerical solutions, numerical
errors can be induced. In particular, the map T(x, q) is often constructed numerically for
practical CFD problems. Moreover, in (4.26) the presence of the derivative z(l, q) at the
boundary can introduce additional errors. These are practical issues that are important to
address in more complex problems. Later in this chapter, some of these practical issues are
presented, and suggestions are given in section 5.3 for dealing with the issue of numerically
approximating state gradient information.
See Chapter 2 and the references therein for a precise definition and further discussion.
Much of the same notation is used for both the linear and nonlinear models.
Linear Model
We begin by considering the transformed state equation in (4.7)-(4.8). Multiplying by an
arbitrary function n V and integrating by parts, we have the following integral equation:
This equation, along with the bilinear form given in section 2.1.3 (with K ( x , q) = 1) and
the L2 inner product defined in section 2.1.1, produces the weak form of the transformed
state equation. In particular, (4.7)-(4.8) is equivalent to the following variational equation.
Find such that
for all n(-) V. See [5] and [54] for further details regarding the equivalence of the
variational formulation to the boundary value problem.
The H-SEM method requires us to solve the transformed sensitivity equation given in
(4.9)-(4.10). Note that the boundary conditions (4.10) are nonhomogeneous. To simplify
36 Chapter 4. Computational Algorithms
Now that s*(-, q) has been chosen, the corresponding variational equation is defined in V.
Find v(.) V such that
a(., •) is the bilinear form defined in section 2.2.2 with K(X, q) = 1. The bounded linear
functional is given by
Nonlinear Model
Turning our attention to the nonlinear problem in section 4.2.3, recall that one seeks to
approximate the weakly coupled system defined by the transformed state equation
4.3. Approximation Framework 37
and
respectively. Observe that we have homogenized boundary conditions prior to any varia-
tional formulation for this example. To numerically approximate the solutions to (4.35)-
(4.38), we first obtain the weak formulations using an approach similar to that outlined
in the sections above. Note that for each and exist and belong to
Then for each V the transformed state z satisfies the variational
equation
for all V. We now move to a brief description of the finite element algorithm that is
used to obtain numerical results.
Linear Model
For the finite element approximation of the variational equations, we begin by constructing
the grid. Recall the variational form of the transformed state equation given in (4.28). Note
that the function (., q) is discontinuous at the point in the computational domain.
Hence, a grid point of the mesh is placed at that point. We partition the domain into
subintervals where Let
the width of the ith subinterval be denoted by for We
choose the finite-dimensional subspace of (0, 1) to be the space spanned by N piecewise
linear basis functions denoted
38 Chapter 4. Computational Algorithms
and
and
for i, j = 1,2,... , N, and it is sometimes referred to as the stiffness matrix. The vectors
on the right-hand side of the equations are given as follows:
and
4.3. Approximation Framework 39
Once vN( ) has been obtained, the approximation of the transformed sensitivity is
computed using
Note that the subscript notation used on refers to the use of the H-SEM approach for
the sensitivity calculation. The approximation to the original sensitivity is obtained using
the mapping T and the relation Likewise, we use the notation
to denote the sensitivity approximation obtained using the abstract version of the
SAM method. For this approach, the approximation and the relationship
are used to calculate the sensitivity approximation. Note that the spatial derivative and the
sensitivity of the mapping M are hardwired into the relation. Therefore, no error is incurred
from the mapping terms when we recover using the expression above.
Nonlinear Model
Although there are several possible choices for finite element spaces, we limit our discussion
to the simplest (convergent) scheme. Recall that the nonlinear model motivates the algorithm
we apply to two-dimensional flow problems in later chapters; hence, we note that in more
complex problems one must choose these spaces with care to ensure the algorithm satisfies
the appropriate convergence criteria (inf-sup conditions, etc.).
For this example, we construct a finite element scheme that allows for more flexibility
than in the previous linear model. In particular, the weak coupling of the transformed state
and transformed sensitivity equations allows one to choose different mesh sizes for each
of these numerical calculations. We remark that such an option is certainly viable for the
linear model; however, it is not explored in this work. In this section, N denotes the number
of basis functions used for the finite element solution of the transformed state equation, and
M defines the number of basis functions used for the transformed sensitivity calculation.
For each of the numerical calculations, the grid points are equally spaced within the domain
40 Chapter 4. Computational Algorithms
and
denote Galerkin approximations of the pair z(.) and u(.). Observe that and
respectively. This leads us to the variational equations
Also note that the coefficients of the unknowns involve the computation of several forms
of inner products dictated by (4.53). With a certain amount of algebra and patience, these
coefficients can be determined analytically. This was our approach; however, these coeffi-
cients can also be generated numerically if one prefers. The nonlinear system given above
can be solved using a black box nonlinear solver, or one can generate one's own code as an
exercise in numerical methods. The approximations given later in the book were generated
using Newton's method with a line search. Although it might seem a bit impervious, the
equation in (4.54) yields an N x N linear system of equations. Once ZN is known, then the
terms involving can be used to determine the matrix and right-hand-side vectors for
this equation.
Note that UM (•) depends on ZN (•) and its spatial derivative (•) on To empha-
size the dependence, we let u N,M (.) denote the solution of (4.54), given that z N (.) obtained
from (4.53) is used in (4.54). At this point, two important observations play a key role in
the construction of accurate numerical sensitivities:
• The freedom to choose separate finite element spaces for the transformed state z( , q)
and the transformed sensitivity u( , q) allows for the development of schemes that
4.3. Approximation Framework 41
simultaneously converge to the state and the sensitivity. In addition, both h-refinement
(mesh refinement) and p-refinement (the selection of higher-order elements) can be
combined to construct numerical solutions of ZN ( , q) and uN,M ( , q) so that the error
in u N,M ( , q) is sufficiently small to ensure convergence of optimal design algorithms
based on the SEM (see [7], [9], and [10]).
• The solution U N , M ( ) depends not only on zN( ) but also on its derivative .
Moreover, since zN( ) is piecewise linear, is a piecewise constant (discon-
tinuous) function. However, the actual transformed sensitivity u( ) is smooth, and
one might expect to lose at least one order of accuracy in u N,M ( ). In fact, things can
be much worse unless special care is exercised.
There are two obvious fixes with which to address these issues. One could use higher-
order splines for the sensitivity variable u( ). However, this method will be more expensive,
and it is not reasonable to expect great improvements unless higher-order schemes are also
used for the state equation. The other obvious fix is to use mesh refinement in M (assuming
accuracy in N). A third approach makes use of smoothing projections. The idea is similar
to the method used to obtain a posteriori error estimators for adaptive mesh generation (see
[11], [30], [57], and [58]). This approach is outlined in section 5.3 and is applied to the
model problem and to a two-dimensional fluid flow problem.
This page intentionally left blank
5 Numerical Results
This chapter contains numerical calculations associated with the computational algorithms
outlined in Chapter 4. Numerical results are presented for the two SEMs applied to the
linear model. This is followed by a section containing numerical approximations for the
nonlinear model using the Hybrid method.
43
44 Chapter 5. Numerical Results
5.1.2 H-SEM
In this section, we present sensitivity calculations obtained by applying the H-SEM algo-
rithm. Figure 5.4 shows the convergence of the finite element approximations to ( , 1.5).
Observe that the entire error results from the approximation of the boundary condition
Since the transformation M is smooth, the
5.1. Linear Model 45
only error in (x, 1.5) is due to this approximation. Figure 5.5 shows the slow conver-
gence of (x, 1.5) to (x, 1.5). Although numerical results are given only for q = 1.5, the
error in the approximate boundary condition and the qualitative behavior of the convergence
are similar over the entire parameter range. We take a moment to clarify this issue in the
following paragraph.
Recall that ( ) is obtained using a piecewise linear approximation. Thus, the
46 Chapter 5. Numerical Results
finite element spatial derivative is a piecewise constant function. This function is used
to approximate the spatial derivative at the boundary point £ = 1. Figure 5.6 shows
a piecewise constant approximation used to obtain an approximate boundary condition
. Hence, the error in (l) results in sensitivity errors that can be attributed to the
poor approximation of this boundary condition. There are techniques that can be used to
obtain better approximations to the spatial derivative along the boundary. Higher-order
5.1. Linear Model 47
elements can be used in the transformed state calculation, but this can be costly for two-
dimensional and three-dimensional problems. As an alternative, projection techniques have
been developed to enhance the accuracy of the spatial derivative for nominal expense. This
technique is explored in detail in section 5.3 and the references therein.
5.1.3 A-SAM
We turn our attention to numerical results obtained by using the A-SAM algorithm for
sensitivity calculations. First, recall that (x, q) is constructed from and
using the relationship.
Note that the subscript A is used to denote the sensitivity approximation obtained using the
A-SAM approach. The finite element approximations to p( , 1.5) are shown in Figure 5.7
for various values of N. Since the sensitivity equation (4.33) is linear, the approximations
pN( , 1.5) converge as expected. When constructing (x, 1.5) from (5.1), the piecewise
constant approximation of produces discontinuities in , as shown in Figure
5.8. These discontinuities occur at points of the physical domain that correspond to mesh
nodes of the computational domain lying in the interval . Note that the expressions
for the mesh derivatives in (5.1) are hardwired, continuous functions, and the finite element
approximations to p( , q) are continuous. It follows that
as . However, one does not get convergence in the energy norm since
does not belong to . Even if one computes the Hl error of the sensitivity
approximation over each individual element of the mesh and computes the total error by
48 Chapter 5. Numerical Results
summing these local errors, we see that this error tends to a nonzero constant with mesh
refinement; see Figure 5.9.
The numerical results presented in the previous sections indicate that each SEM im-
plemented here suffers from computational difficulties. Both algorithms require accurate
gradient information from the state approximation to accurately approximate the sensitiv-
5.2. Nonlinear Model 49
ity. The H-SEM method requires accurate gradient information only along the boundary,
whereas the A-SAM approach relies on accurate gradient approximations over the entire
spatial domain. We also see that the use of mapping techniques in conjunction with SEMs
must be done with care, and the mathematical analysis of the sensitivity equations can
be useful in determining approximation errors related to the choice of numerical scheme.
Moreover, the choice of computational scheme may also depend on the needs of the designer.
The hybrid method provides sensitivity approximations that converge in the energy norm,
and this issue may be important to a designer who is using the sensitivity approximations
to analyze the influence of the design parameter on the underlying mathematical model.
However, if the designer is approximating sensitivities for use in an optimization algorithm,
then the L2 accuracy of the sensitivity approximations may be sufficient.
where WN(X, q) is obtained using (4.25). Notice that the gradient has the form
It is important for the reader to observe that the H-SEM applied to the optimization problem
replaces , the sensitivity of the discrete solution, with an approximation to
, in particular , from (4.26).
5.2. Nonlinear Model 53
Figure 5.17. L2 error of the solution and sensitivity approximations atq = 1.2.
5.2. Nonlinear Model 55
p =4
q=2 q = 1.4
Perturbation Perturbation
0.1250 0.4495 0.4516 0.0021 0.8248 0.7431 -0.0817
0.2500 0.8284 0.7722 -0.0563 1.3541 1.3875 0.0335
0.5000 1.4641 1.4080 -0.0561 2.1394 1.9848 -0.1546
0.7500 2.0000 1.8338 -0.1662 2.7553 2.5758 -0.1795
p = 16
q =2 q = 1.4
Perturbation Perturbation
0.0312 0.1213 0.1126 -0.0088 0.2677 0.2690 0.0013
0.0625 0.2361 0.2207 -0.0154 0.4806 0.4480 -0.0326
0.0938 0.3452 0.3670 0.0218 0.6629 0.6375 -0.0254
0.1250 0.4495 0.4372 -0.0123 0.8248 0.7563 -0.0685
0.1562 0.5495 0.5916 0.0421 0.9720 0.9772 0.0052
0.1875 0.6458 0.6611 0.0154 1.1079 1.0335 -0.0744
0.2188 0.7386 0.7651 0.0265 1.2347 1.2820 0.0473
0.2500 0.8284 0.8687 0.0403 1.3541 1.4106 0.0565
0.3125 1.0000 0.9091 -0.0909 1.5749 1.4821 -0.0928
0.3750 1.1623 1.2674 0.1051 1.7768 1.6265 -0.1503
0.4375 1.3166 1.2486 -0.0680 1.9641 1.9460 -0.0181
0.5000 1.4641 1.6086 0.1445 2.1394 2.0504 -0.0890
0.5625 1.6056 1.6483 0.0427 2.3048 2.0950 -0.2098
0.6250 1.7417 1.9067 0.1651 2.4619 2.3973 -0.0646
0.6875 1.8730 1.9668 0.0938 2.6117 2.8040 0.1922
0.7500 2.0000 2.0554 0.0554 2.7553 2.9296 0.1743
58 Chapter 5. Numerical Results
Figure 5.22. The cost function and its approximations for p = 16 and q* ~ 2.
Figure 5.23. The cost function and its approximations for p = 16 and q* ~ 1.4.
5.2. Nonlinear Model 59
Table 5.2. Optimization results for p = 16, qopt ~ 2, qinit = 1.2, PWC derivatives.
Table 5.3. Optimization results for p = 16, qopt ~ 1.4, qinit = 2.0, PWC derivatives.
The standard Gauss-Newton algorithm is used to approximate q*; see [18] for details.
The algorithm solves a least-squares problem at each iteration and proceeds as described
in Figure 5.19. The data to be matched, denoted by in (5.2), are indicated by pluses in
Figures 5.20 and 5.21. This data set was generated by randomly perturbing the value of
w( ) in (3.17) using q = 2 and q — 1.4 with p = 4,16 data points. Table 5.1 shows the
numerical values of the data for comparison purposes.
The exact cost functional, J(q), and several approximations to it, JN(q), are plotted
60 Chapter 5. Numerical Results
in Figures 5.22 and 5.23 for the case where p = 16 and q* ~ 2 and q* ~ 1.4, respectively.
Although several simulations were constructed with various data sets, we present the
results of two of these:
• Case 1. The true state was calculated for q = 2.0 and the noise vector in Table 5.1 was
added to obtain data for optimization. Here, the optimal value of the design parameter
is approximately q* ~ 2. The optimization algorithm was started at qinit = 1.2.
• Case 2. In this case, the true state was calculated for q = 1.4 and the noise vector
in Table 5.1 was again added to obtain data for optimization. Here, the optimal value
of the design parameter is approximately q* ~ 1.4. Here, the optimization algorithm
was started at qinit = 2.0.
The scheme was considered converged when the norm of the gradient of the cost functional
was less than 10~7. Notice that the simulations were performed using sensitivities calculated
using the natural piecewise constant finite element gradient approximations. Tables 5.2 and
5.3 show the results of these simulations as N, M ranged from 2 to 128. The runs were
measured in seconds and were performed on a Silicon Graphics Onyx2. Notice the effect of
the inaccurate sensitivity approximations on the convergence of the optimization scheme for
Case 1 when N = M. As expected, larger values of N, M were required for convergence of
the optimization algorithm for Case 2. In the following section, we return to this problem
and use smoothing projections to enhance sensitivity computations and convergence.
where 7 = 1, 2 , . . . , M and
5.3. State Gradient Approximations 61
Remark. Even in higher dimensions, the finite element derivatives are often discontinuous
across element faces. It is possible to construct higher-order basis functions that provide
continuous derivatives across element faces. However, from the standpoint of CFD, this
construction also must satisfy the inf-sup conditions in order to guarantee convergence, and
the computational cost of using higher-order basis functions often outweighs the benefits
of improved accuracy. The projection techniques surveyed here provide the scientist with
the convenience of low-order basis functions along with improved accuracy for a relatively
little computational work.
where gi, and (•) are the "hat functions" defined in section 4.3. Observe that
and contains functions with nonzero trace on the boundary of [0, 1].
Define gN( ) to be the orthogonal projection of gN( ) onto SN. In particular,
where the vector a = (a 1 ,a 2 ) T contains the coefficients of the basis functions. These
coefficients are determined by solving the normal equation system
With higher-order finite elements and in higher dimensions, one must resolve the
value of at a nonvertex node. An averaging technique is generally used. The local
projection technique is described in more detail in Chapter 8.
Derivative Approximations
Figure 5.24 shows the exact spatial derivative along with the finite element derivative and
its local and global projections for the case when q = 2.0 and N = 4. Figure 5.25 shows a
similar set of functions and approximations for the case when q = 1.2 and N = 8. Clearly,
the global and local projections give different results. Since we have an exact solution for the
nonlinear example, and thus its spatial derivative, we can calculate element by element L2
errors as well as overall L2 errors over the entire domain for each derivative approximation.
Figures 5.26 and 5.27 show the L2 element errors for the two cases q = 2.0, N = 4 and
q = 1.2, N = 8, respectively. The L2 errors over the spatial domain for each of these
cases are summarized in Table 5.4. Figures 5.26 and 5.27 demonstrate that the errors for the
local projection technique are highest near the boundary, but away from the boundary the
local projection technique actually has less error than the global projection technique for
these two cases. We now analyze these techniques in calculating sensitivities for varying
discretizations (N, M) and parameter values (q) to gain a better understanding of how the
improved derivative approximations affect our numerical sensitivities.
Sensitivity Approximations
Figures 5.28,5.29, and 5.30 display the sensitivity approximations using the three different
state gradient approximations against the exact sensitivities for q — 2.0, 1.4, and 1.2,
respectively (with varying mesh sizes). It is clear that the use of a projection technique
greatly improves the accuracy of our sensitivity approximations, especially as q —> 1.
Recall that at q = 1.4 and q = 1.2 our sensitivity approximations obtained using PWC
derivatives were extremely inaccurate so that the approximations do not even show up on
64 Chapter 5. Numerical Results
L2 errors
N = 4 and q = 2 W = 8 and q = 1.2
PWC 0.4330 1.1432
Global 0.2165 0.8760
Local 0.2954 1.0501
the graphs in Figures 5.29 and 5.30 for some values of N. When compared with the original
discontinuous state gradient approximations, the local projections clearly decrease the L2
error of the sensitivities for a given N and q, as shown in Figure 5.31. Moreover, the most
promising result is that the projections stabilize the calculations over the parameter range.
The local projections also give slightly more accurate sensitivity approximations for some
values of q when compared to the global projection technique. Figure 5.32 compares the
L2 error of sensitivities calculated using local projections with the error calculated using
global projections. Note that as N increases, the global projections do a better job than local
projections over a wider range of q. This is not surprising since the global projection is the
"best" least-squares linear approximation (using the finite element basis) of the piecewise
continuous finite element derivative in the limit. Next, we briefly examine the effect that
the improved sensitivity approximations have on the optimization problem considered in
section 5.2.2.
66 Chapter 5. Numerical Results
Figure 5.31. Model problem—L2 error of sensitivity approximations (PWC and local).
68 Chapter 5. Numerical Results
Figure 5.32. Model problem—L2 error of sensitivity approximations (global and local).
Optimization Results
We evaluated the same two cases considered in section 5.2.2: qinit = 1.2 with q* ~ 2
(Case 1) and qinit = 2.0 with q* ~ 1.4 (Case 2). This time the simulations were performed
using the piecewise linear derivative approximations. Tables 5.5-5.8 show the results of
these simulations as N, M range from 2 to 128. Note that the use of the piecewise linear
derivative approximations clearly improve the results of the optimization algorithm for
Case 1 and that the global and local projection schemes provided very similar results. The
improvement was not as marked for Case 2; however, the scheme did converge for the case
N — M = 16 with the improved gradient approximations and did not with the piecewise
constant derivatives. In addition, the local scheme required slightly less time to converge
than the global projection scheme (see Tables 5.7 and 5.8).
6 Mathematical Framework for
Navier-Stokes Equations
This chapter presents the mathematical framework used to assert the existence of the sen-
sitivity for Navier-Stokes equations. The necessary preliminaries, such as function spaces
and notation, are given, and the Dirichlet problem is treated in detail. We use, throughout,
the following nondimensional form of the Navier-Stokes equations for the two-dimensional
steady state flow of an incompressible, viscous fluid in a bounded domain with a
Lipschitz-continuous boundary :
Recall that v = where Re is the Reynolds number, for the nondimensional form of the
equations. See [44] and [51] for a more detailed derivation of the Navier-Stokes equations.
To discuss existence and uniqueness and to introduce the variational form of the problem,
we introduce some standard function spaces as well as some necessary bilinear and trilinear
forms.
Recall that the trace theorem proves the existence of a trace operator, and, as noted in [54],
the range of the trace operator is . The norm for functions g belonging to
71
72 Chapter 6. Mathematical Framework for Navier-Stokes Equations
can be defined by
where
Also, let
and
Recall that the trilinear form, has some nice properties described in the
following lemma.
In order to discuss the uniqueness of the solutions (u, p) to (6.3), we introduce the
norm of the trilinear form a1 ( • ; - , - ) , denoted , and defined by
We also set
The variational form of the nonhomogeneous partial differential equation (6.1) with (6.9)
is obtained using standard weak formulations. In particular, we consider the following
problem.
74 Chapter 6. Mathematical Framework for Navier-Stokes Equations
To prove existence for the nonhomogeneous problem, we need the following technical
result, due to Hopf (see Lemma 2.3 in [21]).
Lemma 6.4. Let satisfy (6.10). For any there exists a function
such that
Theorem 6.5. Let and satisfying (6.10). There exists at least one
pair which is a solution of (6.11).
Before stating the uniqueness result again, we make a few definitions. For any function
6.3. An Abstract Framework for Navier-Stokes 75
, define
and
where
Define as in by
The basic uniqueness result for (6.11) is found in [21]. We state it below for conve-
nience.
Remark. Although the results above provide existence and uniqueness for the basic Navier-
Stokes problems, they do not address the continuity and differentiability of these solutions
with respect to the parameter q. For example, in the problems considered below we have
, and/or g = g(q) so that v0 = = v0( ), where
q is some parameter defining the flow. We need to establish the smoothness of these
mappings to address the existence and uniqueness of solutions to the sensitivity equations.
This is the subject of the following sections.
Moreover, we suppose that these solutions are nonsingular in the sense that
As an immediate consequence of (6.22), it follows from the implicit function theorem (see,
e.g., [56]) that q u(q) is a Cp function from A into X.
and we set
Lemma 6.8. There exists a continuous linear function D : V H1 ( ) such that for each
g V, we have that w = D(g) satisfies
and
for each
and is defined by
It is also clear that (u, p) satisfies the homogeneous Navier-Stokes equations, (6.1) and
(6.2), with right-hand-side f if and only
The implicit function theorem implies that (u, p) is a continuously differentiable
function of f if the linear map is an isomorphism. But this is equivalent to
the condition that the homogeneous sensitivity equation has a unique solution in y for each
f H-1( ). We show in the following that the sensitivity equation does indeed have a
unique solution in y so that we have (u, p) is a C1 function of f. We then return to the
smoothness of solutions to the parameter-dependent Navier—Stokes equations.
recall that in some cases . As in the case of the one-dimensional model problem,
the sensitivity equations are obtained by implicitly differentiating the flow equations and
their associated boundary conditions. The sensitivity equations fit into the following general
form. Given and satisfying (6.10) and (u, p) a solution of (6.1)
with (6.9), find a pair (s, r) H1 x such that (s, r) satisfies
for all z
Theorem 6.11. Assume the hypotheses of Theorem 6.6. If(u, p) is the unique solution of
Variational Problem 6.2, then there exists a unique solution, (s, r), to Variational Problem
6.3.
Proof. To check that the equations (6.41) have a unique solution, it is sufficient to prove
that the bilinear form
Now, assume that v > V0, where V0 is defined by (6.17). Then, there exists a function
such that
Since
we obtain
Theorem 6.12. Assume that q is such that v > vo . There exists a neighborhood
of such that for all q the solution of the variational, parameter-
dependent Navier-Stokes equations (6.11), with f = f(q) and g = g(q), exists and is
unique. Moreover, the solution is a Cl function of q.
Proof. We consider the map F defined by (6.31). We have shown that there exists Rn
so that we have a branch of solutions {(q, u(q)); q }, i.e., that both (6.20) and (6.21)
hold for q , for both cases presented in section 6.3.3.
We now turn our attention to the Frechet derivative of F, Du F. We have
Again, the fact that DuF is an isomorphism can be shown to be equivalent to the fact
that the homogeneous sensitivity equation has a unique solution for all q . We showed
in section 6.4.2 that the sensitivity equation does indeed have a unique solution. Then, by
the implicit function theorem, we have that q (u(q), p(q)) is a C1 function from into
X.
7Two-DimensionalFlow
Problems
We now return to analyzing design sensitivities, and we focus on two specific parameter-
dependent flow problems in CFD. We present flow equations for these problems and use
formal implicit differentiation to derive the continuous sensitivity equations. The reader
should observe that the coupling of the sensitivity equations to the state equations through
the appearance of state and state gradient terms can be seen in both examples. Hence, many
of the numerical issues addressed for the nonlinear model of section 3.3 are also addessedfor
the examples introduced in this chapter. An adaptive finite element technique is described
and used to solve the state and sensitivity equations. The numerical results are presented to
investigate the convergence of the adaptive grids.
81
82 Chapter 7. Two-Dimensional Flow Problems
follows:
The sensitivity equations are obtained by implicitly differentiating the system (6.1) and its
associated boundary conditions with respect to the parameter q. Assuming the order of
differentiation can be interchanged, we obtain the following:
The sensitivity boundary conditions at the inflow and outflow are obtained by differ-
entiating (7.1), producing
After differentiation, the no-penetration, no-slip conditions for u imply no-penetration, no-
slip conditions for s. In section 6.4 we present the weak form of these sensitivity equations
and prove an existence result.
for 0 y 3, where = 0.5 was a constant parameter describing the strength of the
inflow. Again, no-penetration and no-slip conditions are applied on the top and bottom
channel sidewalls as well as on the bump.
7.2. Flow over a Bump 83
For this application, we examine a shape sensitivity. Here, the shape of the bump is
a cubic spline determined by a parameter, q. We examine two cases: q = q1 R1 and
q = (q1, q2, q3)T R3. We seek to find the sensitivity of the flow in the channel to changes
in q, which we denote
We must derive and solve a separate set of linear sensitivity equations for each sensi-
tivity. The sensitivity equations are
For q R3, q1, q2, and q3 specify the height of the spline at x = 1.5, 2.0, and 2.5,
84 Chapter 7. Two-Dimensional Flow Problems
Error Estimation
The error estimation is performed using an approach introduced by Zhu and Zienkiewicz
[3], [57], [58] and involves the postprocessing of stresses and strains. Recall that the energy
7.3. A Finite Element Formulation 85
norm of u is
Note that the energy norm has a form very similar to the H1 seminorm. In fact, it can be
shown that they are equivalent norms. Using the energy norm for the velocity, define the
so-called Stokes norm of the solution as
As pointed out in [50], the use of the energy norm over the H1 seminorm offers some
advantages, especially to the engineering community. Note that both the velocity and the
pressure norms are expressed in terms of surface forces, which are the quantities of prime
interest in engineering fluid mechanics. Second, errors computed in these norms can be
interpreted as errors in the stresses, which can then easily be related to errors in global
quantities such as lift and drag.
The Zhu-Zienkiewicz approach uses the Stokes norm to measure the error, e(u, p) =
(uex — uh, pex — ph), where (uex, pex) is the exact solution of the flow problem and (uh, ph)
86 Chapter 7. Two-Dimensional Flow Problems
Remeshing Strategy
Once error estimates are obtained for each element, say, ei, a new mesh density (or element
size), d, is calculated which requires equidistribution of the element errors across all the
elements. For example, if we wish to reduce the error in each element by a factor of y, then
the target error, eT, for an element in the new mesh can be given by
where e is the error over the entire domain and N is the number of elements. If one assumes
that the finite element method is of order k, then it is reasonable to write
where h is the current element size and d is the ideal element size we seek. Clearly, we
have assumed that we are in the asymptotic range of the finite element method and that the
7.4. Some Numerical Results 87
convergence constant, c, is the same for both meshes. This may or may not be the case.
Nevertheless, the system (7.18)-(7.19) can be solved for the new mesh density, d, obtaining
The new element size computation is done for each of the dependent variables in the problem,
e.g., velocity, pressure, and an ideal mesh density is obtained for each. Finally, the minimum
of these is used for the generation of the new mesh. The details of the actual redefinition of
the mesh are omitted here; we refer the interested reader to [30] and [31].
Despite the rather major assumptions made above, this adaptive remeshing strategy
works remarkably well. The strategy has been verified in a series of numerical experiments
for a variety of flow types (see [30], [31], [35], [48], [49]). We use this adaptive strategy in
the problems investigated below.
The outflow boundary condition is modified, however, since applying the Dirichlet boundary
condition at the outflow causes numerical instability (see [17]). A free boundary condition
is used that requires
where is the outward normal at the end of the channel. No-flow and no-slip conditions
are applied on the top, bottom, and cylinder sidewalls.
As noted earlier, we wish to approximate the sensitivity of the solution with respect
to the inflow parameter q. The sensitivity equations are given by (7.3), and the sensitivity
boundary conditions at the inflow are obtained as before, with
The computational outflow boundary conditions for the sensitivity equations become
88 Chapter 7. Two-Dimensional Flow Problems
Numerical results for this flow problem were generated using the approximation tech-
niques outlined in section 7.3. The Reynolds number for the calculations is Re = 100, the
length of the channel is 8 (L = 6), and the sensitivity parameter value is q = 0. Contour
plots of the u, v-velocity fields as well as the u, u-velocity sensitivities are given in Figures
7.5 through 7.8. In Figure 7.4, the initial and adapted grids are shown. It is clear that the
mesh is refined around the cylinder and in areas of large velocity gradients. This gives
improved approximations of the velocity field, as can be seen in Figures 7.5 and 7.6. Since
the mesh refines on the velocity field, it is convenient that, for this problem, the sensitivity
flow field is similar to the velocity field, so that as the mesh refines we obtain improved
sensitivity approximations as well (see Figures 7.7 and 7.8). It is important to note that this
is not always the case, as is shown in [11]. The code was modified by Jeff Borggaard to
adapt on the sensitivity field as well, and some results for this are shown later.
The Reynolds number for the calculations is Re = 100 and the length of the channel L = 8.
The initial and adapted grids are shown in Figure 7.9. Contour plots for u, v velocities and
sensitivities are displayed in Figures 7.10 and 7.11. Note that the mesh refines in the area of
the bump and the elements are allowed to become larger downstream in the channel where
the flow is again quadratic. The bump problem is similar to the cylinder problem in that the
sensitivities are largest in the same areas where the mesh refines. Thus we obtain improved
7.4. Some Numerical Results 89
This chapter describes a local projection technique (analogous to that given in Chapter 5.3)
for a two-dimensional case. We demonstrate that this projection technique, which provides
improved state gradient approximations, can be combined with adaptive grid refinement,
on both the state and sensitivities, to improve the accuracy of sensitivity approximations.
A case study is presented to evaluate the use of state and sensitivity approximations for the
evaluation of cost functionals and their gradients in optimization.
95
96 Chapter 8. Adaptive Mesh Refinement Strategies
as
where the vectors ax, ay R6 contain the coefficients of the basis functions. These coeffi-
cients are obtained by solving the following least-squares problems:
Thus, for each component of g*, we solve the following 6x6 system of linear equations
fora x and ay:
The finite element fluxes, and , are obtained in the usual manner by differentiating the
finite element basis functions. Note that the left-hand matrix is independent of the quantity
being projected and thus can be viewed as the projection matrix for node, , and can be
used for obtaining locally projected derivative approximations for all the dependent variables
(e.g., u and v) as long as the projection basis, P, is not changed. Once the linear systems
are solved, we have a quadratic expression for the locally projected derivative, g*, at each
8.1. A Local Projection for Higher Dimensions 97
vertex. Let denote the expression for g* obtained by solving the systems (8.3)-(8.4),
where is the subdomain associated with element vertex, . Define and similarly
for the remaining element vertices, and . We need a unique definition of g* for any
point, inside the element (see Figure 8.2). For quadratic elements, there are several ways
to do this. We describe the technique employed in the current version of the code.
Unique nodal values of g* at the element vertices, which we denote and ,
are simply defined as follows:
Nodal values for the midside nodes are obtained by averaging the values of the polynomial
expressions for g* at the endpoints of element side. For example,
Then, at any point in the element, the value of the locally projected derivative g* at is
where the Nj are the quadratic basis functions for the finite element space and the are
the nodal values of the locally projected derivative obtained as described above. In the
following section, we apply the described local projection technique to obtain improved
sensitivity approximations for two flow problems.
98 Chapter 8. Adaptive Mesh Refinement Strategies
interpolated onto the first adapted mesh. Note that the difference between the approxima-
tions using the projected gradients and those using the natural finite element derivatives
is no longer as apparent. Again, a node-by-node comparison was made. These errors are
shown in Figures 8.9(a) and (b) and 8.10 (a) and (b) and are plotted on the same scales used
for the initial mesh for easy comparison. We also present nodal errors for the two schemes
in the cases where we adapt on the flow and the sensitivities (see Figures 8.9(c) and (d) and
8.10(c) and (d)). At this Reynolds number, there seems not to be a significant advantage
gained by adapting on both the flow and sensitivity errors.
To get an overall evaluation of the errors, we used the node-by-node errors to calculate
an L2 error over the entire domain for the flow and the sensitivities (see Table 8.1). This
more clearly shows the 50% reduction of the error for the local projection scheme on the
initial mesh. Note that the errors for u, v on the first adapted meshes are the same for the
unprojected and projected schemes on the mesh that was refined only on the flow, but they
are slightly different (since the meshes are slightly different) when the meshes were also
refined on the sensitivity. Note that adapting on the flow and sensitivity errors not only
improved the sensitivity approximations over those obtained by adapting on the flow alone
but also improved the numerical approximations for the flow.
We now turn to the results for Re = 350. We begin by comparing the state and
sensitivity approximations for L = 6 and L = 15 to ascertain whether the length of the
channel affects our results. Figure 8.11 shows the initial meshes for the two channel lengths
as well as the u, v velocity contours. It is clear that the flow more nearly returns to the
inflow for the longer channel. However, the results for the shorter channel are very similar
to those for the longer channel in the areas where they overlap. In fact, Figures 8.12 and
8.13 show that the dramatic difference between the sensitivities obtained with unprojected
derivatives and those obtained using locally projected derivatives occurs in both the long
and the shorter channels. This gives us reasonable confidence that the length of the channel
is not affecting our results.
We now return to complete the same error analysis for Re = 350 that we did for
Re = 100. The initial and first adapted meshes are shown in Figure 8.14. Note that
the difference between the three first adapted meshes is definitely greater at this Reynolds
number. This is due, at least in part, to the greater discrepancy between the flow error,
the unprojected sensitivity error, and the projected sensitivity error. For Re = 350, the
differences between the sensitivities calculated with the two different derivative schemes is
now dramatic, as can be seen in Figures 8.15 and 8.16. The node-by-node error analysis
(see Figure 8.17) shows that locally projected derivatives are definitely better, although at
this Reynolds number, both approximations contain fairly large errors. Using the locally
projected derivatives to calculate sensitivities reduces the overall L2 error by about 600%
(see Table 8.2, initial mesh section).
The same analysis is done on the first adapted meshes (See Figures 8.18 through 8.21).
Here, adapting on the flow as well as the sensitivity makes a greater difference in the error
reduction from the initial mesh to the first adapted mesh. This is easily seen in the overall
L2 errors for the approximate flow and sensitivities displayed in Table 8.2. It is interesting
to note that the errors for all the quantities are less for Projected (Flow & Sensitivity) than
for Unprojected (Flow & Sensitivity), although the Unprojected (Flow & Sensitivity) mesh
is quite a bit finer.
106 Chapter8. Adaptive Mesh Refinement Strategies
Initial Mesh
Unproj Proj
u 1.7397E-01 1.7397E-01
V 9.0160E-02 9.0160E-02
su 5.0463E-01 2.8227E-01
sv 2.2453E-01 1.3050E-01
First Adapted Meshes
Unproj (Flow) Proj (Flow) Unproj (Flow & Sens) Proj (Flow & Sens)
u 1.0313E-01 1.0313E-01 7.2637E-02 5.3305E-02
V 4.5997E-02 4.5997E-02 3.3000E-02 2.8115E-02
su 3.4173E-01 3.5800E-01 2.5988E-01 2.7273E-01
sv 1.4487E-01 1.1039E-01 9.1075E-02 7.6292E-02
Figure 8.11. Initial meshes and u, v-velocity contours for L = 6, 15 and Re = 350.
8.2. Numerical Results for Two-Dimensional Problems 109
Initial Mesh
Unproj Proj
u 1.4808E-00 1.4808E-00
V 4.1024E-01 4.1024E-01
su 1.9438E+01 3.6791E-00
Sv 3.0582E-00 5.8348E-01
First Adapted Meshes
Unproj (Flow) Proj (Flow) Unproj (Flow & Sens) Proj (Flow & Sens)
u 9.6580E-01 9.6580E-01 3.5985E-01 2.8674E-01
V 2.6044E-01 2.6044E-01 1.0257E-01 7.7747E-02
su 3.8080E-00 3.1614E-00 1.1145E-00 1.0423E-00
Sv 1.5339E-00 8.7266E-01 3.4686E-01 2.9430E-01
To conclude, we observe that using the locally projected derivatives clearly stabilizes
the calculations over a larger range of Reynolds numbers, as it did in section 5.3.3 for the
one-dimensional model problem. Plots of the overall L2 sensitivity errors on the initial
mesh are shown in Figures 8.22 and 8.23. In addition, the mesh refinement very effectively
reduces the errors for the flow and the sensitivities. At most Reynolds numbers, after the
mesh is refined once, there is little difference between the sensitivities calculated with the
unprojected and the locally projected techniques. At higher Reynolds numbers, however,
using locally projected derivatives to calculate sensitivities remains advantageous. We now
turn our attention to obtaining numerical approximations to the design sensitivities for the
flow over a bump problem.
Figure 8.25. u-velocity contours and sensitivity vectors, Re = 500, flow over a bump.
8.2. Numerical Results for Two-Dimensional Problems 123
Figure 8.26. Flow over a bump, initial and adapted meshes for Re = 1000, L = 20.
124 Chapter8. Adaptive Mesh Refinement Strategies
gradient approximations improve in the second and third adapted meshes, the values of
the sensitivities become much more accurate, not only in the local area of the bump but
downstream as well. Note also that the secondary whirlpool, which appears in the Re = 500
case, is seen on the initial mesh in this case also. As the mesh is refined, however, the size of
the whirlpool is diminished until it is almost gone, as can be seen in Figure 8.31(c). Clearly,
having accurate state gradient approximations on the boundary is necessary for accurate
sensitivity approximations, especially shape sensitivities. This is an area requiring future
research.
Finally, we turn to the issue of using numerical approximations of the state and sensitivities
to approximate cost functionals and their gradients. A comparison of our results with a
problem considered by Burkardt in Chapter 11 of [14] is given. Burkardt presents what he
refers to as a discretized sensitivity failure. We evaluate his results and show that the process
of adaptive mesh refinement is key to obtaining good cost function and gradient evaluations
in order to prevent inaccurate results from an optimization code.
We examine the numerical experiment carried out in section 11.2 of [14]. We evaluate
the following cost functional:
Here, P is the number of matching points and for the results we present herein P was fixed
at 15. The matching points, (3, yi), i = 1, 2 , . . . , P, were evenly distributed along the line
x = 3, y [0, 3]. Also, the target profile, uTarg(3, y), was obtained by computing a finite
element solution with the shape parameter, qTarg = (0.375, 0.5,0.375)T, and = 0.5, and
then interpolating to obtain the values for uT(3, yi), i = 1, 2,... , P.
The gradient of the cost function with respect to q is expressed as
The SEM uses the discrete approximation of the continuous sensitivity equation as before
to approximate the value of the cost gradient. We denote this approximation of the cost
gradient by qJh; thus we have
Our results are displayed in Table 8.3 and Figures 8.32 and 8.33. Burkardt reported
a local maximum at S = 5. As seen in Figure 8.32, we do not get the same results on the
initial mesh. To determine if the difference was due to not having sufficient accuracy on the
initial mesh, we also present the results for meshes which were refined on the flow field.
It is possible that the difference between our results and those of Burkardt could be due
to our having fixed the value of = 0.5. It is also possible that the difference is a result
of numerical inaccuracies due to discretization differences. We perform another numerical
experiment similar to the one described above with one exception: we now use Re = 100.
Figures 8.34 and 8.35 show the cost function and gradient along the same line explored
above. Note that adapting the mesh is even critical to obtaining accurate, smooth cost and
gradient approximations at this Reynolds number.
8.2. Numerical Results for Two-Dimensional Problems 131
s q1 q2 q3 Jh.103 ( qJ
h.
S).103 Jh.103 (qJh - ).103
Initial Mesh 03 Mesh
0 -0.117 0.419 -0.149 16.9 -0.99 12.1 -0.56
1 -0.097 0.422 -0.128 16.3 -1.03 11.7 -0.57
2 -0.077 0.425 -0.107 15.5 -1.00 11.3 -0.58
3 -0.057 0.428 -0.086 14.6 -0.94 10.8 -0.60
4 -0.038 0.432 -0.065 13.7 -0.91 10.3 -0.61
5 -0.018 0.435 -0.044 12.9 -0.92 9.95 -0.62
6 0.001 0.438 -0.023 12.2 -1.04 9.49 -0.63
7 0.020 0.441 -0.002 11.4 -1.00 9.02 -0.65
8 0.040 0.444 0.018 10.6 -0.96 8.56 -0.67
9 0.060 0.448 0.039 9.72 -0.99 8.00 -0.68
10 0.079 0.451 0.060 9.02 -1.15 7.48 -0.70
11 0.099 0.454 0.081 8.35 -1.07 6.95 -0.72
12 0.119 0.457 0.102 7.46 -1.08 6.39 -0.74
13 0.138 0.461 0.123 6.21 -0.77 5.76 -0.75
14 0.158 0.464 0.144 5.62 -0.78 5.17 -0.76
15 0.178 0.467 0.165 5.00 -0.79 4.56 -0.77
16 0.197 0.470 0.186 4.38 -0.78 3.94 -0.76
17 0.217 0.474 0.207 3.70 -0.76 3.27 -0.75
18 0.237 0.477 0.228 3.08 -0.74 2.66 -0.72
19 0.256 0.480 0.249 2.47 -0.70 2.07 -0.68
20 0.276 0.483 0.270 1.89 -0.64 1.51 -0.62
21 0.296 0.487 0.291 1.33 -0.56 0.99 -0.53
22 0.316 0.490 0.312 0.72 -0.50 0.57 -0.42
23 0.335 0.493 0.333 0.32 -0.35 0.25 -0.29
24 0.355 0.496 0.354 0.11 -0.21 0.05 -0.13
25 0.375 0.500 0.375 0.00 0.00 0.01 0.05
26 0.394 0.503 0.396 0.06 0.16 0.16 0.28
27 0.414 0.506 0.417 0.31 0.38 0.53 0.53
28 0.434 0.509 0.438 0.81 0.63 1.16 0.81
29 0.453 0.513 0.459 1.63 0.92 2.08 1.13
30 0.473 0.516 0.480 2.72 1.22 3.30 1.48
132 Chapter 8. Adaptive Mesh Refinement Strategies
135
136 Bibliography
[13] D. Braess. Finite Elements. Cambridge University Press, Cambridge, UK, 1997.
[14] J. Burkardt. Sensitivity Analyses and Computational Shape Optimization for Incom-
pressible Flows. Ph.D. thesis, Virginia Polytechnic Institute and State University,
Blacksburg, VA, 1995.
[15] P. Ciarlet. The Finite Element Method for Elliptic Problems. North-Holland, Amster-
dam, 1978.
[16] T. Coleman, F. Santosa, and A. Verma. Semi-automatic differentiation. In Compu-
tational Methods for Optimal Design and Control, Progress in Systems and Control
Theory, Birkhauser, Boston, 1998, pp. 113-126.
[17] C. Cuvelier, A. Segal, and A. A. van Steenhoven. Finite Element Methods and Navier-
Stokes Equations. Reidel, Dordrecht, The Netherlands, 1986.
[18] J. E. Dennis, Jr., and R. B. Schnabel. Numerical Methods for Unconstrained Op-
timization and Nonlinear Equations. Classics in Applied Mathematics 16, SIAM,
Philadelphia, 1996.
[19] D. A. Field and V. Komkov, eds. Geometric Aspects of Industrial Design, SIAM,
Philadelphia, PA, 1992.
[20] A. Friedman. Partial Differential Equations. Robert E. Krieger, Huntington, NY, 1976.
[21] V. Girault and P. Raviart. Finite Element Methods for Navier-Stokes Equations. Com-
putational Mathematics 5, Springer-Verlag, New York, 1987.
[22] R. Glowinski and J. He. On shape optimization and related issues. In Computational
Methods for Optimal Design and Control, Birkhauser, Boston, 1998, pp. 151-179.
[23] A. Godfrey. Using sensitivities for flow analysis. In Computational Methods for
Optimal Design and Control, Progress in Systems and Control Theory, Birkhauser,
Boston, 1998, pp. 181-196.
[24] A. Godfrey and E. Cliff. Direct calculation of force derivatives: A sensitivity-equation
approach. In 36th AIAA Aerospace Sciences Meeting and Exhibit, Reno, NV, 1998.
[25] A. Griewank and G. F. Corliss, eds. Automatic Differentiation of Algorithms. SIAM,
Philadelphia, 1991.
[26] A. Griewank. On automatic differentiation. In Mathematical Programming, KTK
Scientific, Tokyo, 1989, pp. 83-107.
[27] A. Griewank. Evaluating Derivatives: Principles and Techniques of Algorithmic
Differentiation. SIAM, Philadelphia, 2000.
[28] M. Gunzburger. Sensitivities in computational methods for optimal flow control. In
Computational Methods for Optimal Design and Control, Progress in Systems and
Control Theory, Birkhauser, Boston, 1998, pp. 197-227.
Bibliography 137
[30] J.-F. Hetu and D. Pelletier. Adaptive remeshing for viscous incompressible flows.
AIAA Journal, 30(8): 1986-1992, 1992.
[31] J.-F. Hetu and D. Pelletier. Fast adaptive remeshing for viscous incompressible flows.
AIAA Journal, 30(ll):2677-2682, 1992.
[35] F. Ilinca and D. Pelletier. A unified approach for adaptive solutions of compressible
and incompressible flows. In AIAA 35th Aerospace Sciences Meeting and Exhibit,
1992.
[36] J. lott, R. Haftka, and H. Adelman. Selecting Step Sizes in Sensitivity Analysis by
Finite Differences. Technical Report 86282, NASA, 1985.
[37] F. John. Partial Differential Equations. 4th ed. Springer-Verlag, New York, 1991.
[38] J. B. Keller. The shape of the strongest column. Archive for Rational Mechanics and
Analysis, 5(4):275-285, 1960.
[39] R. V. Kohn and G. Strang. Optimal design and relaxation of variational problems. I.
Communications on Pure and Applied Mathematics, 39(1): 113-137, 1986.
[40] R. V. Kohn and G. Strang. Optimal design and relaxation of variational problems. II.
Communications on Pure and Applied Mathematics, 39(2): 139-182, 1986.
[41] R. V. Kohn and G. Strang. Optimal design and relaxation of variational problems. III.
Communications on Pure and Applied Mathematics, 39(3):353-377, 1986.
[43] E. Kreyszig. Introductory Functional Analysis with Applications. John Wiley & Sons,
New York, 1989.
[45] A. Limache and E. Cliff. Aerodynamic sensitivity theory for rotary stability derivatives.
In AIAA Atmospheric Flight Mechanics Conference, Portland, OR, 1999.
138 Bibliography
mass matrix, 61
mathematical frameworks, 5-16, 71-80
method of mappings, 1, 25-28
139