Inverse Problems An Introduction - kirchgraber-Kirsch-Sto Er
Inverse Problems An Introduction - kirchgraber-Kirsch-Sto Er
Inverse Problems An Introduction - kirchgraber-Kirsch-Sto Er
t
0
e
ps
x(s)ds , t [0, 1] .
Reconstruction Problem: Given a dierentiable function y : [0, 1] R and a
parameter p > 0 determine x : [0, 1] R such that
y(t) =
t
0
e
ps
x(s)ds for all t [0, 1] .
Parameter Identication Problem: Given a dierentiable function y : [0, 1]
R and a continuous function x : [0, 1] R, nd a parameter p > 0 such that
y(t) =
t
0
e
ps
x(s)ds for all t [0, 1] .
The term e
pt
may be considered as a discount factor.
As we know from integration theory, the forward problem can be solved analytically and nu-
merically in a very stable way. Here we are mainly interested in the the two inverse problems.
Clearly, y(0) = 0 is a necessary condition for solvability of the inverse problems.
The reconstruction problem is equivalent to
y
(t) = e
pt
x(t) for all t [0, 1] .
Therefore, we get back the function x by
x(t) = y
(t)e
pt
, t [0, 1] .
Notice that we have to dierentiate the data y.
The parameter identication problem is equivalent to
y
(t) = e
pt
x(t) for all t [0, 1] .
3
Clearly, the parameter is overdetermined we will come back to this classication by this set
of equations: we need just one
^
t (0, 1) where
p =
1
^
t
ln
_
y
(
^
t)
x(
^
t)
_
makes sense. Again, we have to dierentiate the data y.
The prototype of an inverse problem will be an equation of the form
F(x) = y (x X, y Y) (1.1)
with a mapping F from the space X into the space Y . For such an equation, the unknown is
x and the data are usually the right-hand side y. If the stability condition is violated, the
numerical solution of the inverse problem by standard methods is dicult and often yields
instability, even if the data are exact (since any numerical method has internal errors acting like
noise). Therefore, special techniques, so called regularization methods have to be used in order
to obtain a stable approximation of the solution. The appropriate construction and analysis of
regularization methods and subsequently (or simultaneously) of numerical schemes is the major
issue in the solution of inverse problems.
1.2 A list of inverse Problems
Let us start with a very short historical view on topics related to Inverse problems.
The term inverse problem is a well known mathematical term dating at least from the 19
th
century.
- For many centuries people have been searching for hiding places by tapping walls and
analyzing echo; this is a particular case of an inverse problem.
- Platon
1
is talking in his allegory of cave in a philosophical context about the reconstruction
of reality from observations of shadows (distorted blurred images from objects outside the
cave) on a wall. Today we may consider Platons problem as an inverse problem to solve.
- The problem which may be considered as one of the oldest inverse problem in applied
science is the computation of the diameter of the earth by Eratosthenes
2
in 200 b. Chr.;
see [6].
- In 1800 Gauss
3
used the method of least squares to reconstruct the orbit of a comet
from earlier orbit data. This method became fundamental in the applied sciences.
- When astronomer Urbain Le Verrier worked out the mathematics to successfully predict
where the planet Neptune would be discovered in the night sky back in 1846, he was really
solving an inverse problem. By that, he used the observations that had been recorded about
Uranus position in its orbit to infer how that orbit might have been aected by what was
then a hypothetical eighth planet in the solar system.
1
Platon, 427 - 347
2
Eratosthenes, 284 202
3
C. F. Gauss, 1777 1855
4
- In 1917 Radon
4
published a paper entitled by ber die Bestimmung von Funktionen durch
ihre Integralwerte lngs gewisser Mannigfaltigkeiten; see [20]. In this paper a transforma-
tion is studied which is nowadays called the Radon transform and which is the basis of
the X-ray tomography.
- It was Heisenberg
5
who conjectured that quantum interaction was totally characterized
by its scattering matrix which collects information of the interaction (at innity). The
discovery of neutrinos by measuring consequences of its existence is in the spirit of inverse
problems too.
- The weakly electric nocturnal sh Gnathonemus petersii is well known for its ability to use
its electro-sensory system in the dark for active electrolocation of objects; see [4].
Over the past 40 years, the number of publications on inverse problems has grown rapidly.
Nowadays there are several mathematical journals devoted to this topic. Especially, the subject of
image processing and non-invasively, non-destructively mapping became very important (Making
the unseen visible with mathematics).
The following list of problems gives a good impression of the wide variety of applications.
Several of the listed problems may be seen under the motto making the invisible visible.
Estimation of a correlation matrix from data;
the inverse problem of geomagnetic induction;
X-ray tomography, ultrasound tomography, laser tomography;
acoustic scattering, scattering in quantum mechanics;
inverse problem of elastography;
radio-astronomical imaging, image analysis;
locating cracks or mines by electrical prospecting;
seismic exploration, seismic tomography;
the use of electrocardiography and magnetocardiography;
evolution backwards in time, inverse heat conduction;
the inverse problem of potential theory;
can you hear the shape of a drum/manifold?
deconvolution, reconstruction of truncated signals;
compartmental analysis, parameter identication;
data assimilation;
determing the volatility in models for nancial markets;
discrete tomography, shape from probing.
1.3 Ill-posedness/Well-posedness
In a complete analysis of an inverse problems we denote the collection of methods and results
in this eld by inversion theory the questions of existence, uniqueness, stability and
construction of objects have to be considered.
4
J. Radon, 1887 1956
5
W. Heisenberg, 1901 1976
5
1.3.1 Denition of well-posedness
The question of existence and uniqueness is of great importance in testing the assumption behind
any mathematical model. If the answer in the uniqueness question is no, then we know that even
perfect data do not contain enough information to recover the physical quantity to be estimated.
By questioning for stability we have to decide whether the solution depends continuously on the
data. Stability is necessary if we want to be sure that a variation of the given data in a suciently
small range leads to an arbitrarily small change in the solution. This concept was introduced by
Hadamard in 1902 in connection with the study of boundary value problems for partial dierential
equations and he designated unstable problems ill-posed
6
otherwise well-posed. The nature of
inverse problems (irreversibility, causality, unmodelled structures, . . . ) leads to ill-posedness as
a characteristic property of these problems.
When solving ill-posed problems (numerically), we must certainly expect some diculties,
since any error acts as a perturbation on the original equation and so may cause arbitrarily
large variations in the solution. Since errors cannot be completely avoided, there may be a range
of plausible solutions and we have to nd out a reasonable solution. These ambiguities in the
solution of inverse problems which are unstable can be reduced by incorporating some sort of
a-priori information that limits the class of allowable solutions. By a-priori information we
mean an information which has been obtained independently of the observed values of the data.
This a-priori information may be given as a deterministic or a statistical information. We shall
restrict ourselves mainly to deterministic considerations.
As we already have seen, an inverse problem may be formulated as the problem to solve an
equation governed by an operator A(p) . Ill-posedness tells us that the inverse A(p)
1
does not
exist and/or is not continuous. The remedy is regularization. The idea of the regularization
theory is to replace the inverse of such an operator by a one-parameter family of continuous
operators and to choose the best approximation in this family by a clever strategy. Such a
regularization strategy is usually based on a merit function and uses two main ingredients
for solving an inverse problem in a stable way: a-priori information and signal to noise ratio
(SNR). The signal to noise ratio is a quantity which describes the relation of the size of the
true solution (signal, image, . . . ) to the size of the noise contained in the measured quantity
(right-hand side of the equation,. . . ).
The Hilbert space l
2
and operators on l
2
(see Appendix 1.6 for informations concerning these
objects) are used to formulate toy-problems; see the example below. In the course of the lecture
we will see that these toy-problems may be seen at least from the theoretical point of view in
a one-to-one correspondence to applied problems via the singular value decomposition. Let us
present a rst example using this elementary framework.
Example 1.2 Consider the problem of solving the equation
Ax = y
where A is a linear mapping from l
2
into l
2
; see Appendix 1.6. Therefore, x and y are sequences
in l
2
:
x = (x
1
, x
2
, . . . ) , Ax = y = (y
1
, y
2
, . . . ) ;
6
Hadamard believed many mathematicians still do that ill-posed problems are actually incorrectely posed
and articial in that they would not describe physical systems. He was wrong! Nowadays we know that such
problems arise in a fundamental way in the modelling of complex (physical) systems.
6
Consider the specic case
Ax := (x
1
,
1
2
x
2
,
1
3
x
3
, . . . ) , x l
2
.
A is bounded since
|Ax| =
_
k=1
1
k
2
x
2
k
_1
2
k=1
1x
2
k
_1
2
= |x|
and the candidate of the inverse of A is given by
A
1
y := (y
1
, 2y
2
, 3y
3
, . . . ) .
Existence of a solution: For the right-hand side
y := (1,
1
2
,
1
3
, . . . )
no solution (in l
2
) exists!
Uniqueness of a solution: It is easy to see that uniqueness holds.
Stability: For the right-hand side
y
n
:= (0, 0, . . . , 0,
1
n
, 0, 0, . . . )
the solution is given by
x
n
:= (0, 0, . . . , 0,
n, 0, 0, . . . )
and we have
lim
n
|y
n
|
2
= 0 , lim
n
|x
n
|
2
= lim
n
n = .
Therefore A
1
cannot be a bounded operator!
Regularization: Approximate the mapping A by A
of A
is given as follows:
A
1
y = (
1
1 +
y
1
,
2
1 + 2
y
2
,
3
1 + 3
y
3
, . . . ) .
Clearly A
1
abs,U
f(x) := sup
xU
|f(x) f(x)|
|x x|
(1.2)
abs,U
f(x) is called the absolute condition number of f . When x ,= and f(x) ,= then we call
rel,U
f(x) := sup
xU
|f(x) f(x)|
|f(x)|
_
|x x|
|x|
(1.3)
the relative condition number of f .
Remark 1.3 Suppose we want to solve a linear equation in the innite-dimensional case gover-
ned by the operator A:
Ax = y
Then the evaluation problem for the inverse A
1
is unstable when the operator A
1
is unbounded.
As a rule, this is the case when the problem to solve the equation is ill-posed.
Under the assumption that f in the evaluation problem is a dierentiable function dened
on R in the general case we would have to consider partial dierentability we obtain an
innitesimal relative condition number:
cond
rel
f(x) :=
|x||f
(x)|
|f(x)|
(1.4)
This number could also be called the elasticity of the function value with respect to its argument.
From this number the problems of evaluating the function f in x can be read o.
Sensitivity analysis, which characterizes the change in model output due to variations
in model input parameters, is of critical importance in simulation models; see Section 2.1.4.
Sensitivity coecients, dened as the partial derivatives of the model output with respect to
the input parameters, are useful in assessing the reliability of the output from a complex model
with many uncertainty parameters. Thus, we can state that condition numbers, as dened above,
characterize the sensitivity of a mathematical problem to data perturbations.
8
1.3.3 Types of errors
Consider again the problem E(f, x) of evaluating the mapping f in (a neighborhood of) x . From
the numerical point of view we have to take into account the presence of various sources of errors:
Model-error f may be considered as a model for a physical process: instead of f an approxi-
mation
f is available only.
Errors in the data x is known approximately only (due to measurement errors): instead of x
an approximation x is available only.
Truncation errors The value
f( x) can be computed not exactly, arising from the fact that an
innite number of operations has to be replaced by a nite number of operations.
Rounding errors We have to expect errors due to the fact that in the computation process a
nite number of real numbers is available only.
Example 1.4 Consider the evaluation of the function
f(t) := ce
at
+ de
bt
, t R (1.5)
in t. The function f may be considered as a model for a chemical reaction: a, b are certain
reaction constants, c, d describe the mixing relation.
A source of error in the evaluation of f in a point t may be:
a, b, c, d are known approximately: model error.
t is known approximately: error in the data.
The exponential function can be computed via the potential series by truncating the innite
summation process only: truncation error.
The evaluation of an exponential yields an irrational number in general; rounding is neces-
sary.
Example 1.5 As we shall see, the price C = C(S, t) at time t with value S of the underlying for
an European call option in the Black-Scholes model is given by
C(S, t) = S^(d
+
()) Ke
r(Tt)
^(d
()) (1.6)
with
d
() =
ln(
S
K
) + (r
2
2
)(T t)
T t
, 0 ,
where ^ denotes the density of the standard normal distribution:
^(a) :=
a
2
exp(
s
2
2
)ds , a R.
Here r, , K, T are the usual constants.
In the evaluation of this price formula various sources of errors should be regarded. Above all,
the evaluation of the distribution function ^ is a source of errors due to the fact that it contains
the exponential function which is dicult to evaluate in a stable way, especially for large and
small arguments.
9
1.3.4 Over- and underdetermined linear systems
One of the most basic and important problems in science and engineering is the accurate and e-
cient simultaneous solution of a system of linear equations. Moreover, most numerical techniques
in optimization and dierential equations involve repeated linear equation solving.
A system of m linear equations in n unknowns consists of a set of algebraic relations of the
form
n
j=1
a
ij
x
j
= y
i
, i = 1, . . . , m (1.7)
where x
j
are the unknowns, a
ij
are the coecients of the system and y
i
are the components of
the right hand side. System (1.7) can be written more conveniently in matrix form as
Ax = y (1.8)
where we have denoted A = (a
ij
) K
mn
the coecient matrix, by y = (y
i
) K
m
the right
hand side vector and by x = (x
i
) K
n
the unknown vector, respectively; here K is the eld of
real or complex numbers. We call a solution of (1.8) any n-tuple of values x
j
which satises (1.7).
We call a system (1.8) consistent if there exists a solution otherwise inconsistent. The case
that (1.8) is not consistent is not an unimportant situation. This situation occurs in virtually
every branch of science and engineering.
We say that (1.8) is overdetermined
8
if we have more equations than unknowns: m > n.
Normally we do not expect that an overdetermined system has a solution. Clearly, we may decide
by Gaussian elimination applied to the rectangular system whether a solution exists from the
analytical point of view. But from the numerical point of view it may be dicult to decide this
since rounding errors may destroy the solvability and vice versa.
We can reformulate the problem of solving the system (1.8) in another way to ask for the
vector x such that the residual r := Ax y is minimized in a given vector norm. In the case
where we use the euclidean norm, | |
2
, this becomes precisely the important statistical problem
of nding the best least squares solution of (1.8). Under the assumption that the rank of A is
n then this least squares solution x
= (A
t
A)
1
A
t
y (1.9)
Here the matrix (A
t
A)
1
A
t
is the so called Moore-Penrose inverse or pseudoinverse. Notice
that this pseudoinverse exists since A
t
A is a n n-matrix with rank n. As we know from
numerical analysis, this solution x
t
0
x(s)ds , t [0, 1] .
Inverse Problem: Given a dierentiable function y : [0, 1] R determine
x := y
.
11
We are interested in the inverse problem. Since y should be considered as the result of mea-
surements the data y are noisy and we may not expect that the noisy data y is presented by
a continuously dierentiable function. Therefore, the inverse problem has no obvious solution,
especially when y(0) ,= 0 . (In practice we simply subtract y(0) from y.)
Moreover, the problem should not be formulated in the space of continuous functions since
perturbations due to noise lead to functions which are not continuous. But the analysis and the
message of the results is not very dierent from the following in the more elementary case of
continuous perturbations.
Suppose that instead of the continuous function y : [0, 1] R a measured function
y
of y at (0, 1) by
x
,h
() := D
h
y
() :=
y
( + h) y
()
h
where h ,= 0 has to be chosen such that + h [0, 1] . We obtain
|x
,h
() x()| |
y( + h) y()
h
x()| + |
(y
y)( + h) (y
y)()
h
| .
Under the assumption that x is continuously dierentiable we have
y( + h) y()
h
x() =
1
2
y
h
(1.11)
follows. Now it is clear that the best what we can do is to balance the terms on the right hand
side in (1.11). This is done by the choice
h := h
opt
:= 2
_
E
. (1.12)
This gives
|x
,h()
() x()| 2
E . (1.13)
The Diagram 1.3 which is a graphical presentation of the bound (1.11) is typical for approximati-
ons in ill-posed problems: there are two terms in the error estimate, a term due to approximation
of the inverse mapping and a term due to measurement error. The balance of these two terms
gives an optimal reconstruction result. Thus, in contrast to well-posed problems, it is not the
best to discretize ner and ner. One may consider ill-posed problems under the motto When
the imprecise is preciser.
9
Moreover, the bound in (1.13) shows that halving the (measurement)
error does not lead to a halving of the absolute error in the result (as it is usually the case in
direct problems).
9
This is the title of [11].
12
h E
h
h
error bound
Abbildung 1.3: Error balance
The requirement of the bound in (1.10) descri-
bes an information concerning the solution we
want to nd. Therefore this information is cal-
led a priori information/a priori knowledge.
The quantity
SNR :=
E
s
0
x(t)dt = y(s) , s [0, 1]
is the simplest example of a Volterra equation of the rst kind which is an equation of the form
b
a
(t, s)x(t)dt = y(s) , s [c, d] ,
where is a kernel function from [a, b] [c, d] into R.
1.4.2 Numerical dierentiation realized with oating point numbers
We consider the problem of the numerical dierentiation realized on a computer with oating
point numbers. Floating numbers are elements x of
F := F(g, t, e
min
, e
max
) :=
[d
1
. . . d
t
] g
et
|
d
1
, . . . d
t
{0, 1, . . . , g 1}, d
1
,= 0,
e Z, e
min
e e
max
, {+, }
{0} .
These numbers are called normalized since d
1
,= 0 is required. The parameters of this set are:
g N, g 2 base of the numbers
t N number of signicant digits d
i
{0, 1, . . . , g 1} (precision)
m := d
1
. . . d
t
N mantissa of x
e Z exponent of x
[e
min
, e
max
] range of exponents
The consequences of the use of these numbers in the dierentiation is the subject of this subsec-
tion. We restrict us to the case g = 10 .
Consider the approximation of the derivative f
(x) =
f(x + h) f(x)
h
+ O(h) ;
the term O(h) is from of the remainder of the approximation. If the step with h is given as h =
10
s
then the dierence quotient is an approximation of f
(x) =
f(x + h) f(x h)
2h
+ O(h
2
)
with step width h = 10
s
then the precision is
min(2s, t s) , i.e. 2t/3 .
A further improvement results from
f
(x) =
4
3
f(x + h) f(x h)
2h
1
3
f(x + 2h) f(x 2h)
2h
+ O(h
4
)
for a step width h = 10
s
since the precision is now
min(4s, t s) , i.e. 4t/5 .
The observation is that going ahead in the same way the gain-rate of signicant digits is poor.
Each of the formulae above may be used to approximate the derivative. An improvement of
the accuracy can be achieved by halving the stepsize h. Notice that in halving the stepsize the
cost is dierent: one new evaluation for the rst formula, 2 additional evaluation for the second
and 4 evaluations for the third.
1.5 Numerical algorithms
Algorithms are the main tools to formulate methods to solve problems. Here we formulate the
main goals in the analysis of algorithms. In Chapter 3 we shall consider the questions again under
more realistic assumptions.
1.5.1 Algorithms
The solution of a problem by numerical methods is usually formulated by an ordered list of steps
where each of which uses data of the problem and results from an earlier step; the evaluation
problem may be considered as such a problem when we are forced to decompose the mapping f
as follows: f = f
l
f
1
. Such a list of instructions (of length l) is called an algorithm when
certain requirements are fulllled:
An algorithm is a nite set of welldened feasible ordered instructions for accom-
plishing some task which, given an initial state, will terminate after a nite number
of operations in a corresponding recognizable end-state.
Dierent algorithms may complete the same task with a dierent set of instructions in more
or less time, space or eort than others. The analysis of an algorithm is concerned with the
following characteristics:
Accuracy: Accuracy refers to the absolute or relative error of an approximate quantity. The
numerical approximation should be as accurate as possible. This requires the algorithm to be
numerically stable.
14
Robustness/Reliability: The algorithm should solve many problems well and it should warn
the user, if the result is inaccurate. Hence it should be possible to estimate the error.
Eciency/Complexity: The amount of operations and the size of the memory required should
be as small as possible.
Precision: Precision is the accuracy with which the basic arithmetic operations are performed.
In all our analyses there is an implicit assumption that the given arithmetic is of arbitrarily high
precision just in the more principal consideration of numerical algorithms precision may play a
role.
To discuss whether a numerical solution approximates the solution of a problem, we need a
measure of the distance between a numerical solution and the exact solution to the problem and
a method parameter which we can use to vary the numerical method.
A numerical method is convergent if the distance between the numerical solution and the
exact solution goes to zero as the method parameter approaches some limit. In order to obtain
this property, the numerical method must be stable. This means that the error growth factor
is bounded is bounded independent of the method parameter. Stability is purely a property
of the numerical method and is independent of the problem. To have any hope of establishing
convergence, it is necessary to establish some kind of connection between the problem and the
numerical method. This connection is consistency. In the following we will discuss the properties
convergence, stability, consistency in mathematical terms.
1.5.2 Analysis of numerical algorithms
Let X be a space endowed with the metric d. Let x
.
Denition 1.1 The pair (x
, (x
n
)
nN
) is called consistent i
lim
n
x
n
= x
when > 0 .
In Chapter 4 we will discuss methods which are appropriate to extract from the sequence
(x
n,
)
nN
good approximations for x
(K) := {x = (x
k
)
kN
|x
k
K for all k N, sup
kN
|x
k
| < } .
c(K) := {x = (x
k
)
kN
|x
k
K for all k N, (x
k
)
kN
convergent} .
c
0
(K) := {x = (x
k
)
kN
|x
k
K for all k N, (x
k
)
kN
is a null sequence} .
c
c
(K) := {x = (x
k
)
kN
|x
k
K for all k N, x
k
= 0 for almost all k N} .
Obviously,
c
c
(K) c
0
(K) c(K) l
(K) .
Addionally, we set
|x|
:= sup
kN
|x
k
| , x = (x
k
)
kN
l
.
On can show that (l
(K), | |
(K) and c
0
(K) is a
closed subspace of c(K) . This implies that (c(K), | |
), (c
0
(K), | |
kN
|x
k
|
p
< }
and
|x|
p
:=
_
kN
|x
k
|
p
_1
p
, x = (x
k
)
kN
l
p
(K) .
Then (l
p
(K), | |
p
) is a Banach space.
Denition 1.5 Let (X, | |
X
), (Y, | |
Y
) be normed spaces. A mapping f : D Y, D X, is
called continuous in x
0
D i
> 0 > 0 x D(|x x
0
|
X
< = |f(x) f(x
0
)|
Y
< ) .
The mapping f is called continuous i f is continuous in each x
0
D.
17
Example 1.9 Let (X, | |) be a normed space. Then the norm considered as a mapping from X
int K is continuous. This follows from
||x| |y|| |x y| .
Moreover, the mappings
X x x + w X, R a aw X, X x rx X,
where w X and r R are xed elements are continuous too. This follows from
|(x + w) (x
0
+ w)| = |x x
0
| , |aw a
0
w| = |a a
0
||w| , |rx rx
0
| = |r||x x
0
| .
The most interesting mappings in linear algebra are the linear mappings. Let (X, | |
X
), (Y, |
|
Y
) be normed spaces. As we know, a mapping L : X Y is linear i
L(au + bv) = aL(u) +bL(v) for all a, b K, u, v X.
The following theorem is easy to prove.
Theorem 1.6 Let (X, ||
X
), (Y, ||
Y
) be normed spaces and let L : X Y be a linear mapping.
Then the following conditions are equivalent:
(a) L is bounded, i.e. there exists c > 0 with |L(x)|
Y
c|x|
X
for all x X.
(b) L is continuous in each x
0
X.
(c) L is continuous in x
0
:= .
Denition 1.7 Let (X, | |
X
), (Y, | |
Y
) be normed spaces and let L : X Y be continuous.
Then the number |L| := sup{|L(x)|
Y
||x|
X
1} is called the operator norm.
Important is the estimate
|L(x)|
Y
|L| |x|
X
, x X, (1.14)
for all x X. It is easy to prove
Corollary 1.8 Let (X, | |
X
), (Y, | |
Y
), (Z, | |
Z
) be normed spaces and let L : X Y, K :
Y Z be continuous. Then
|K L| |K||L| .
Let (X, | |
X
), (Y, | |
Y
), (Z, | |
Z
) be normed spaces. We set
L(X, Y) := {T : X Y|T linear and continuous} .
The space L(X, Y) endowed with the operator norm is a Banach space.
Denition 1.9 Let X be a vector space with scalar space K {R, C} . A mapping : XX K
is called an inner product on X if the following conditions are satised:
10
10
Here, as usual, z denotes the complex conjugate of z
18
(a) (x, x) R, (x, x) > 0 for all x X\{} ;
(b) (x, y) = (y, x) for all x, y X;
(c) (ax + by, z) = a(x, z) +b(y, z) for all x, y, z X, a, b R.
A vector space X endowed with the inner product is called a pre-Hilbert space.
Let (X, ) be a pre-Hilbert space. Then we have for all x, y X
|(x, y)|
_
(x, x)
_
(y, y) ,
with equality when x, y are linear dependent. This inequality is called the CauchySchwarz
inequality. A consequence of this inequality is the fact that the pre-Hilbert space (X, ) becomes
a normed space since | |
, dened by
|x|
:=
_
(x, x) , x X,
is a norm in X.
Denition 1.10 The pre-Hilbert space (X, ) is called a Hilbert space if the normed space
(X, | |
) is complete.
Let (X, ) be a pre-Hilbert space. In the following we want to use the notation which is
common in physics:
, ) := (, ) .
Cauchy-Schwarz inequality becomes now
|x, y)| |x||y| .
x
y
K
y
y
m
n
Abbildung 1.4: Projection theorem
Here are three further identities in a pre-Hilbert
space (X, , )) .
Parallelogram identity |x+y|
2
+|x
y|
2
= 2|x| + 2|y|
2
, x, y X.
Polarization identity
11
x, y) =
1
4
(|x +y|
2
|x y|
2
) +
i
4
(|x +
iy|
2
|x iy|
2
) , x, y X.
Theorem of Pythagoras |x + y|
2
=
|x|
2
+|y|
2
for x, y X with x, y) = 0 .
A subset K in a vector space X is convex i
ax + (1 a)x K for all x, y K, a [0, 1] .
Theorem 1.11 (Variational inequality) Let X be a pre-Hilbert space and let K X be non
empty and convex. Then for all y K the following conditions are equivalent:
(a)
|x y| = inf
zK
|x z| . (1.15)
11
This is the case of complex scalars.
19
(b)
x y, z y) 0 for all z K. (1.16)
Theorem 1.12 (Projection theorem) Let X be a Hilbert space and let K X be non empty
and convex. Then there exists for each x X a uniquely determined y K with
|x y| = inf
zK
|x z| . (1.17)
The vector y in Theorem 1.12 is called the projection of x with respect to K. We denote this y
as P
K
x . The result of the theorem above is the fact that the mapping
X x P
K
x K
is well dened. This mapping is called the projection of X onto K.
Denition 1.13 Let X be a pre-Hilbert space.
(a) A set O X is called an orthonormal system (in X) i
|e| = 1 for all e O, e, f) = 0 for all e, f O, e ,= f .
(b) An orthonormal system is called an orthonormal basis i there exists no orthonormal
system O
with O O
, O
,= O.
Clearly, every orthonormal system is a linearly independent set. Due to Zorns lemma, each pre-
Hilbert space possesses an orthonormal basis. If the underlying pre-Hilbert space X is separable,
i.e. when X is the closure of a denumerable subset A, then there exists an orthonormal basis in
X which is denumarable. Such an orthonormal system can be constructed with the help of the
GramSchmidt orthogonalization.
1.7 Bibliographical comments
Inverse problems are considered for instance in [2, 8, 12, 15, 21, 24, 25]. Early sources for the
treatment of inverse and ill-posed problems are [5, 7, 14, 13, 16, 19, 22, 23].
A monography concerning the important subject tomography is [17]; see also [18]. The theme
inverse crime is discussed for example in [3, 9, 10, 26]. A comprehensive bibliography for
numerical dierentiation is presented in
file:///C:/Users/gruppe/Desktop/NumericalDiffBib
lnk
3.html
and
http://sourcedb.cas.cn/sourcedb
igg
cas/cn/zjrck/200907/W020100801406250190094.pdf
A useful source for the tools in functional analysis (normed spaces, continuous opperators,
function spaces,. . . ) is [1]. The results concerning Banach and Hilbert spaces in the appendix
can be found in [1] too.
20
1.8 Exercises
1.1 Show that under the assumption
|x
() :=
y
( + h) y
()
h
combined with a step-width rule h := h() leads to the error estimate
|x
,h()
() x()| cE
1/3
2/3
.
Here c is a constant independent of , E.
1.2 Suppose that f : [0, 1] R is three times continuous dierentiable. Show for [0, 1)
f
() =
1
2h
(3f() +4f( + h) f( + 2h)) +O(h
2
) .
O() is the big-O Landau-Symbol.
1.3 Suppose that f : [0, 1] R is twice continuous dierentiable. Show for (0, 1)
lim
h0
|f
()
f( + h) 2f() +f( h)
h
2
| = 0 .
1.4 Suppose that f : [0, 1] R is four times continuous dierentiable. Show for (0, 1)
and h > 0 such that + h, h [0, 1],
|f
()
f( + h) 2f() +f( h)
h
2
|
h
2
12
max
s[0,1]
|f
(4)
(s)| .
O() is the big-O Landau-Symbol.
1.5 A simple model for the growth of a population is given by
u
= qu
where u denotes the size of the population and where q : [0, 1] R is a parameter. Find a
method to reconstruct the function q : [0, 1] R from observations of u : [0, 1] R.
1.6 The distance s of a runner from a xed mark is measured at certain times t:
t 0.0 0.5 1.0 1.5 2.0
s 0.00 3.65 6.80 9.90 12.15
(a) Use central dierences in order to reconstruct the velocity of the runner at times t = 0.5
and t = 1.25 .
(b) Describe a method in order to reconstruct the acceleration of the runner and apply this
method for t = 1.5 .
1.7 Derive a formula for approximating the second derivative based on ve equidistant nodes
symmetric to the evaluation point. What is the error of the approximation?
1.8 Let X be the Hilbert space l
2
and dene the linear operator A : X X by the formula
A(x)
k
=
k
x
k
, k N, where (
k
)
kN
is a given sequence of real numbers.
Show: Then is an eigenvalue of A if and only if there is an index k such that =
k
.
21
1.9 Consider the Volterra integral equation
y(t) +
t
0
sin(t s)y(s)ds = 1 , t R,
with ,= 0 . (Then the integral equation is said to be of the 2
nd
kind and - as opposed to the
case = 0 (integral equation of the 1
st
kind) a well-posed equation.)
(a) Show that this equation can be transformed into an initial value problem of 2
nd
order.
(b) In order to solve the integral equation, solve the initial value problem.
1.10 Let X be the Hilbert space L
2
([0, 1]) of all equivalence classes of square-integrable functions
on the interval [0, 1] . Dene an operator A onto X by declaring A(x)(t) := g(t)x(t) , t [0, 1],
for some bounded and measurable function g.
Show: A possesses no eigenvalues if g is chosen appropriate.
1.11 Consider the problem of calculating the fractional derivative of a function f given in
L
2
(R) :
D
f(x) :=
1
(n + 1 )
d
n+1
dx
n+1
f(t)
(x t)
n
dt
for n N, n < < n+1. Such problems are frequently encountered in many practical contexts.
It is well known that if 0 < 1, then D
u)(x) =
1
()
u(t)
(x t)
1
dt = f(x), < x < .
Compute D
f for
f(x) := e
x
2
, f(x) :=
0 , x 1
1 + x , 1 < x 0
1 x , 0 < x 1
0 1 < x
, f(x) :=
0 , x 1
1 , 1 < x 1
0 1 < x
, f(x) := 0 .
1.12 Can you hear the length of string? Consider the boundary value problem
u
= f , u(0) = u(l) = 0
where f : R R is a given continuous function. Suppose that the solution u and f are known.
Find the length l of the interval.
1.13 Consider the boundary value problem
u
+ u = f , u(0) = u(l) = 0
where f : R R is a given continuous function. Find sucient conditions on f such that the
constant > 0 can be computed from an observation u() for some point (0, l).
1.14 The computation of eigenvalues of a given matrix is a standard problem in (numerical)
linear algebra.
(a) If this problem is considered as a direct problem what should be the formulation of the
inverse problem?
22
(b) Let
1
:= 1 i,
2
:= 1 + i . Compute a matrix A R
2,2
with eigenvalues
1
,
2
.
(c) Is the matrix A in (b) uniquely determined? Give an argument for your answer.
1.15 Consider the real Hilbert space (l
2
, | |
2
) . Let x = (x
k
)
kN
l
2
and let x
= (x
k
)
kN
,
0, be a sequence of real numbers. Suppose
|x
k
x
k
| , k N,
k=1
k
x
2
k
E
2
<
where (
k
)
kN
is a sequence of reals with the properties 0 <
k
k+1
, k N, lim
k
k
= .
Dene for N N, 0, the sequence x
N,
l
2
by
x
N,
k
:=
k
, if k N 1
0 , if k N
.
Show:
(a) |x x
N,
|
2
2
f(N, ) := N
2
+
1
N
E
2
.
(b) Minimize (balance) f(N, ) with respect to N for the choice
k
:= k
2
, k N.
(c) Minimize (balance) f(N, ) with respect to N for the choice
k
:= e
k
, k N.
(d) Compare and comment the results in (b),(c).
1.16 The problem of computing the eigenvalues of given matrix is well known. If this problem
is considered as a direct problem what can be the formulation of the inverse problem?
1.17 Given is a well of depth H. How much time T passes until one hears the impact of a stone
which falls freely into the shaft, starting with zero velocity at time 0? This is the direct problem.
The associated inverse problem is: Is it feasible to determine the depth of the shaft from T?
Due to Galilei, the distance s the stone has fallen after a period t of time is given by
s =
1
2
gt
2
.
Here g denotes the gravitational constant. Let c denote the speed of sound, which has to be
taken into consideration when calculating T.
(a) Determine T in dependence of H.
(b) Determine H in dependence T .
(c) What is the precision in H under the assumption that the time T which has passed can be
measured with a precision of ?
1.18 Let f be a real function dened over the interval [0, ). The Laplace transform L(f) is
dened as the integral
L(f)(s) =
0
e
st
f(t)dt
provided the integral exists.
Given the function g : (0, ) s
1
2s
2
(2 3e
s
+ e
3s
) R. Find a function f with
L(f) = g.
23
1.19 Consider the integral equation
1
0
exp(st)x(t)dt = y(s) , 0 s 1 (1.18)
which admits the solution
x(t) := exp(t) , 0 t 1 ,
when the right-hand side is given by
y(s) :=
(exp(s +1) 1)
(s + 1)
, 0 s 1 .
If we approximate (1.18) by the trapezoidal rule, we get
1
0
exp(st)x(t)dt h
_
_
1
2
x(0) +
n1
j=1
exp(jhs)x(jh) +
1
2
exp(s)x(1)
_
_
, 0 s 1 ,
where h := 1/n. With nodes s = ih we get
h
_
_
1
2
x
0
+
n1
j=1
exp(jhs)x
j
+
1
2
exp(s)x
n
_
_
= y
i
:= y(ih) , i = 0, . . . , n, (1.19)
where the numbers x
j
are to be considered as approximations of x(jh).
Tackle the following problems for the cases n = 4, 8, 16 with the help of Matlab.
(a) Find the matrix A which governs the system of linear equations (1.19). beschreibt.
(b) Determine the LU decomposition of A.
(c) Calculate the approximation x
1+n/2
and compare with the value of the exact solution
x(1/2) .
(d) Calculate a condition number of A.
1.20 Consider the following sets of sequences of real numbers:
l
:= {(x
n
)
nN
||(x
n
)
n
|
:= sup
n
|x
n
| < } ,
l
p
:= {(x
n
)
nN
||(x
n
)
n
|
p
:= (
i=1
|x
n
|
p
)
1
p
< } , 1 p <
c := {(x
n
)
nN
| lim
n
x
n
exists} ,
c
0
:= {(x
n
)
nN
| lim
n
x
n
= 0} .
Show for 1 p < :
l
p
c
0
c l
1.21 Consider the left-shift operator dened for a sequence of real numbers x := (x
n
)
nN
by
A(x) := (x
2
, x
3
, x
4
, . . . ) .
Show:
(a) A is a linear and continuous operator which maps l
p
to l
p
and has operator norm |A| = 1
for each 1 p . Here the norm | |
p
is dened as above.
(b) Is there a sequence e with the property Ae = e other than the sequence (0, 0, )?
24
1.22 Consider the mapping
T : c(K) c
0
(K) , (x
n
)
nN
(y
n
)
nN
with y
n
:=
lim
k
x
k
, if n = 1
x
n
lim
k
x
k
, else
.
(a) Show that T is bijective and continuous.
(b) Compute |T|, |T
1
| .
1.23 In statistics the sample variance of n numbers x
1
, . . . , x
n
is dened as
s
2
n
:=
1
n 1
n
i=1
(x
i
x)
2
where x :=
1
n
n
i=1
x
i
(two-pass-formula) (1.20)
s
2
n
:=
1
n 1
_
_
n
i=1
x
2
i
1
n
_
n
i=1
x
i
_
2
_
_
(one-pass-formula) (1.21)
The computational behavior is very poor in the presence of (rounding-)errors. Instead consider
the updating formulae
M
1
:= x
1
, M
k
:= M
k1
+ (x
k
M
k1
)/k, k = 2, . . . , n,
Q
1
:= 0, Q
k
:= Q
k1
+ (k 1)(x
k
M
k1
)
2
/k, k = 2, . . . , n
(1.22)
(a) Show that the updating formulae (1.22) result in s
2
n
= Q
n
/(n 1) .
(b) Compare the results for s
2
n
in the case
x
1
:= 10000, x
2
:= 10001, x
3
:= 10002, x
4
:= 10003
via the formulae in (1.20), (1.21), (1.22) respectively. Compute in single precision arithmetic.
1.24 Consider the Hilbert matrix of order 3:
H
3
:=
_
_
_
_
_
1
1
2
1
3
1
2
1
3
1
4
1
3
1
4
1
5
_
_
_
_
_
.
(a) Compute the inverse T
3
of H
3
.
(b) Compute the inverse
T
3
of H
3
by Gaussian elimination by using the oating-point number
system F(10, 3, , ). Compare
T
3
with rd(T
3
).
25
Literaturverzeichnis
[1] K. Atkinson and W. Han. Theoretical Numerical Analysis. A Functional Analysis Frame-
work. Texts in applied mathematics, vol. 39. Springer, New York, 2001.
[2] J. Baumeister. Stable Solution of Inverse Problems. Vieweg, Braunschweig, 1987.
[3] D. Colton and R. Kress. Integral Equation Methods in Scattering Theory. Wiley, New York,
1983.
[4] G. Emde. Active electrolocation of objects in weakly electric sh. The Journal of Experi-
mental biology, 202:12051215, 1999.
[5] V.M. Fridman. Method of successive approximations for fredholm integral equations of the
rst kind. Uspekhi Mat. Nauk, 11:233234, 1956.
[6] C.W. Groetsch. Inverse Problems in Mathematical Sciences. Vieweg, Braunschweig, 1993.
[7] J. Hadamard. Sur les problmes aux drives partielles et leur signication physique. Prin-
ceton University Bulletin, pages 4952, 1902.
[8] H.W. Engl, M. Hanke and A. Neubauer. Regularization of Inverse Problems. Kluwer,
Dordrecht, 1996.
[9] J. Kaipio and E. Somersalo. Statistical and Computational Inverse Problems. Springer,
Berlin, 2005.
[10] J. Kaipio and E. Somersalo. Statistical inverse problems: discretization, model reduction
and inverse crimes. J. of Computational and Applied Mathematics, 198:493504, 2007.
[11] A. Kirchgraber, D. Stoer und A. Kirsch. Schlecht gestellte Probleme oder Wenn das
Ungenaue genauer ist. Mathematische Semesterberichte, 51:175206, 2005.
[12] A. Kirsch. An Introduction to the Mathematical Theory of Inverse Problems. Springer, New
York, 1996.
[13] C. Lanczos. Evaluation of noisy data. J. SIAM Numer. Anal., 1:7685, 1964.
[14] L. Landweber. An iteration formula for fredholm integral equations of the rst kind. Amer.
J. Math., 73:615624, 1951.
[15] A.K. Louis. Inverse und schlecht gestellte Probleme. Teubner, Stuttgart, 1989.
[16] R.W. McKelvey. An analysis of approximative methods for fredholms integral equation of
the rst kind. Technical report, 1956.
[17] F. Natterer. The Mathematics of Computerized Tomography. John Wiley, New York, 1986.
26
[18] F. Natterer. Imaging and inverse problems of partial dierential equations.
Jahresbericht der DMV, 109:3148, 2006.
[19] D.L. Phillips. A technique for the numerical solution of certain integral equations of the
rst kind. J. Assoc. Comput. Mach., 9:8497, 1962.
[20] J. Radon. ber die Bestimmung von Funktionen durch ihre Integralwerte lngs gewisser
Mannigfaltigkeiten. Ber. Verh. Schs. Akad. Wiss., 69:262277, 1917.
[21] A. Rieder. Keine Probleme mit Inversen Problemen. Vieweg, Braunschweig, 2003.
[22] A.N. Tikhonov. On the stability of inverse problems. Doklady Akademii Nauk SSSR, 39:195
198, 1943.
[23] A.N. Tikhonov. Solution of incorrectly formulated problems and the regularization method.
Soviet Math. Dokl., 4:10351038, 1963.
[24] A.N. Tikhonov and V.Y. Arsenin. Solutions of Ill-Posed Problems. Winston & Sons, Wa-
shington, D.C., 1977.
[25] L. von Wolfersdorf. Inverse und schlecht gestellte Probleme, 1993. Sitzungsberichte der
schsischen Akademie der Wissenschaften zu Leipzig.
[26] A. Wirgin. The inverse crime. arXiv:math-ph/0401050v1, 2004.
27