Professional Documents
Culture Documents
Editors:
J.-M. Morel, Cachan
F. Takens, Groningen
B. Teissier, Paris
Charles W. Groetsch
Stable Approximate
Evaluation of
Unbounded Operators
ABC
Author
Charles W. Groetsch
The Traubert Chair
School of Science and Mathematics
The Citadel
Charleston, SC 29409
USA
e-mail: charles.groetsch@citadel.edu
DOI 10.1007/3-540-39942-9
This work is subject to copyright. All rights are reserved, whether the whole or part of the material is
concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting,
reproduction on microfilm or in any other way, and storage in data banks. Duplication of this publication
or parts thereof is permitted only under the provisions of the German Copyright Law of September 9,
1965, in its current version, and permission for use must always be obtained from Springer. Violations are
liable for prosecution under the German Copyright Law.
Springer is a part of Springer Science+Business Media
springer.com
°c Springer-Verlag Berlin Heidelberg 2007
The use of general descriptive names, registered names, trademarks, etc. in this publication does not imply,
even in the absence of a specific statement, that such names are exempt from the relevant protective laws
and regulations and therefore free for general use.
Typesetting by the author and SPi using a Springer LATEX package
Cover design: WMXDesign GmbH, Heidelberg
Printed on acid-free paper SPIN: 11856795 VA41/3100/SPi 543210
In Memory of Joaquin Bustoz, Jr.
1939-2003
Preface
The key idea of von Neumann’s proof involves regarding the graph of the
operator as a subspace embedded in a product Hilbert space (in this sense
the proof is truly Cartesian). We also show that this notion, combined with
von Neumann’s alternating projection theorem, applied in the product space,
can be used to give an alternate non-spectral proof of one of the best known
operator stabilization methods.
Our intent is for the monograph to be reasonably self-contained. We begin
with a fairly informal introductory chapter in which a number of model inverse
problems leading to the evaluation of unbounded operators are introduced.
The next chapter fills in background material from functional analysis and
operator theory that is helpful in the sequel. We hope that this approach
will make the monograph a useful source of collateral reading for students
in graduate courses in functional analysis and related courses in analysis and
applied mathematics.
Much of the work that is reported here was originally carried out in col-
laboration with my friends Otmar Scherzer of the University of Innsbruck,
Austria and Martin Hanke-Bourgeois of the University of Mainz, Germany.
With both Otmar and Martin I had the happy experience of open and friendly
collaborations in which my benefits exceeded my contributions.
While writing these notes I learned of the tragic death of my earliest
colleague and coauthor, Joaquin Bustoz, Jr., of Arizona State University.
Besides his many research papers on summability theory and special functions,
Joaquin’s important and lasting contributions to the mathematical education
of disadvantaged youth made his passing a great loss to the profession, to say
nothing of the personal sense of loss felt by his friends and colleagues. This
monograph is fondly dedicated to Joaquin’s memory.
I am grateful to the Charles Phelps Taft Research Center for providing
support, in the form of a Faculty Fellowship, during the preparation of this
work.
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
1
Some Problems Leading to Unbounded
Operators
His majesty then turned to me, and requested me to explain the reason
why such great effects should proceed from so small a cause ...
The Adventures of Hajji Baba of Isfahan
J. Morier
1.1 Introduction
the given data into the data space of the solution operator and subsequent
approximate evaluation of the solution operator. Of course these approxima-
tions must be carried out very carefully in order to preserve stability. A key
point in this process is the choice of parameters defining the approximate
evaluation in such a way as to control the instabilities.
It is hoped that this monograph which explores these ideas, and is
expressed in rigorous operator theory, will be interesting to mathematicians
and will also be potentially useful to those in the scientific community who
confront unstable ill-posed problems in their work. Our intent is also to pro-
vide a source for collateral study for students in courses in applied functional
analysis and related courses. With this in mind the next chapter provides
brief details of some of the operator theory that will be used in the sequel.
Before proceeding to this background material we give a semi-formal intro-
duction to some problems from application areas that lead to the evaluation
of unbounded operators.
We remind the reader that a linear operator T : X → Y from a normed
linear space X into a normed linear space Y is called bounded if the number
T x
T = sup
x=0 x
u(0, t) = 0, u(1, t) = 0
of the standing wave could then be observed directly. The density coefficient
ρ(x) satisfies
y (x)
ρ(x) = − 2 .
ω y(x)
That is, the required coefficient ρ(x) is obtained by evaluating the (nonlinear)
operator
y
L(y) = − 2
ω y
defined, for example, on the set of functions y on [0, 1] having absolutely
continuous first derivative and for which L(y) lies in an appropriate function
space, say L2 [0, 1]. The unbounded nature of the second derivative operator
can then mean that exceedingly small experimental errors in the measurement
of y can lead to unacceptably large errors in the computation of y and hence
the evaluation of L(y) is unstable.
and hence
∂u
= |n|fˆ(n) exp(inθ).
∂ν
n∈Z
1.2 Some Unbounded Operators 5
n∈Z
u(x, y) → 0, as x → ∞.
Here π
2
an = f (η) sin nηdη
π 0
and the unbounded operator L is defined on the domain
∞
D(L) = f ∈ L [0, π] :
2
n |an | < ∞ .
2 2
n=1
Instability may arise in more subtle ways in the analysis of partial differential
equations. A well-known example occurs in potential theory. Consider the
6 1 Some Problems Leading to Unbounded Operators
model Cauchy problem for Laplace’s equation on an infinite strip. Here the
unstable mapping takes one boundary distribution into another. Specifically,
suppose u(x, y) is a harmonic function in the strip
∆u = 0 in Ω,
u(0, y) = f (y),
ux (0, y) = 0,
u(1, y) = g(y)
d2
û = ω 2 û,
dx2
û(0) = fˆ,
d
û(0) = 0
dx
giving
ĝ(ω) = û(1, ω) = fˆ(ω) cosh (ω).
Therefore the linear operator connecting f to g is given by
a condition that says roughly that high frequency components of f must decay
very rapidly. In particular, L is defined on band-limited functions and hence
is densely defined in L2 (R). However, L is unbounded. Indeed, if
1.2 Some Unbounded Operators 7
fˆn = χ[n−1/2,n+1/2]
where {cn } are certain complex constants. Suppose boundary values g(θ) =
u(1, θ) on the unit circle are to be given in terms of values f (ϕ) = u(h, ϕ) on
the interior circle r = h, where g has the complex Fourier expansion
g(θ) = cn einθ
n∈Z
Then
f (ϕ) = cn h|n| einϕ
n∈Z
and
D(L) = f ∈ L2 [0, 2π] : h−2|n| |an |2 < ∞
n∈Z
where 2π
1
an = f (ϕ)e−inϕ dϕ.
2π 0
The operator L is densely defined in L2 [0, 2π] and unbounded since, if fm (ϕ) =
eimϕ , then fm is bounded, while Lfm → ∞ as m → ∞. Note that L may
8 1 Some Problems Leading to Unbounded Operators
Ax = {
x, qn }
For example, if H is the real space L2 [0, 1] and pk (s) = sk , these are called
the Hausdorff moments. These moments have a mechanical interpretation if
we think of x as representing the mass density distribution of a bar, for then
1 1 1
x(s)ds, sx(s)ds, s2 x(s)ds
0 0 0
1.3 Unbounded Operators and the Heat Equation 9
represent the total mass of the bar and its moment and moment of inertia,
respectively, about its left end. The mapping from the moments to the dis-
tribution is unbounded. In fact, let L : D(L) ⊂ l2 → L2 [0, 1] be the linear
operator defined on
∞
D(L) = {µk } ∈ l : 2
µk s ∈ L [0, 1] .
k 2
k=0
for k > p/2 and note that µ(p)k ≤ 1/(k − p/2) and hence {µ(p)k } ∈ l2 and
µ(p) is bounded independently of p. Now let xp (s) = s−(1+p ln s)/2 and note
that 1 0
1
xp 2 = s−(1+p ln s) ds = e−t(1+pt) et dt = π/p
0 −∞ 2
and hence xp ∈ L2 [0, 1]. Also note that L({µ(p)}) = xp , and, as previously
observed µ(p) is bounded, but
1
xp 2 = π/p → ∞, as p→0
2
showing that L is unbounded.
In direct problems for partial differential equations the goal is to find a solu-
tion, or features of a solution, given the differential operator, including coeffi-
cients, and certain ancillary information in the form of initial values, boundary
values or source terms. The differential operator may be viewed as a mapping
from certain combinations of coefficients, initial values, boundary values and
source terms to solutions. Inverse problems for partial differential equations
typically require the recovery of distributed coefficients, initial values, bound-
ary values or source terms given sufficient information about the solution.
Since the solution enters the problem through some process of differentiation
it is reasonable to expect that the resolution of inverse problems for par-
tial differential equations will involve the evaluation of unbounded operators.
In this section we provide several examples of inverse problems for the one-
dimensional heat equation that give rise to unbounded solution operators.
The examples treat the reconstruction of initial distributions, source terms,
coefficients and boundary values. In each example we proceed formally and
identify the unbounded solution operator for the inverse problem.
10 1 Some Problems Leading to Unbounded Operators
and hence
∞
2 n2 π
g(x) = u(x, 0) = e f (y) sin ny sin nxdy.
n=1
π 0
In other words, the solution g of the inverse problem is obtained from the
data f via the unbounded operator L defined on functions f in the set
∞ π
2m2 2
D(L) = f ∈ L [0, π] :
2
e am < ∞, am = f (y) sin mydy
m=1 0
Here the instability is apparent: small (in L2 norm) perturbations of the data
2
f can, thanks to the factors en , translate into very large changes in the
solution g.
1.3 Unbounded Operators and the Heat Equation 11
which, after a couple of integrations by parts, noting that w = −g, is equiv-
alent to
2 π
−n2 bn = g(s) sin nsds.
π 0
Therefore,
∞
g(x) = (−n2 bn ) sin nx. (1.5)
n=1
However,
∞ −n2
f (x) = u(x, 1) = n=1 bn e sin nx + w(x)
∞ −n2 2
π
= n=1 bn e + π 0
w(s) sin nsds sin nx
∞ −n2
= n=1 (bn e − bn ) sin nx, by (1.4),
12 1 Some Problems Leading to Unbounded Operators
and hence
2 π
bn (e−n − 1) =
2
f (s) sin nsds.
π 0
By (1.5) one is then led to the explicit representation
∞ π
2 n2
g(x) = Lf (x) = sin nx f (s) sin nsds.
π n=1 1 − e−n2 0
We note that D(L) is dense in L2 [0, π] since all finite linear combinations
of the eigenfunctions fn (s) = sin ns are contained in D(L). Moreover, L is
unbounded since fn 2 = π/2 while
πn2
Lfn 2 = →∞
2(1 − e−n2 )
as n → ∞.
and therefore the coefficient g(t) is related to the profile h(t) = u(.5, t) by
h (t)
g(t) = − 2
π h(t)
That is, the coefficient g(t) is related to the processed profile
f (t) = (− ln |h(t)|)/π 2
d
by the unbounded operator L = , namely, g = Lf .
dt
1.3 Unbounded Operators and the Heat Equation 13
An unbounded operator also plays a role in a simple model problem for the
two-dimensional steady state heat equation. Suppose that the temperature
distribution g(x) = u(x, 1) on the top edge of a rectangle [0, π] × [0, 1] is
desired, where the temperature u(x, y) on the rectangle satisfies
∆u = 0 in (0, π) × (0, 1)
and boundary data on [0, π] are over-specified by the boundary flux condition
∂u
(x, 0) = f (x), for x ∈ [0, π].
∂y
Elementary separation of variables techniques then lead to the representation
g = Lf , where the operator
∞
2 sinh n π
(Lf )(x) = f (s) sin nsds sin nx
π n=1 n 0
Note that if fn (x) = sin nx, then fn 2 = π/2, while Lfn 2 → ∞, and hence
the operator L : D(L) ⊆ L2 [0, π] → L2 [0, π] is unbounded.
∂u ∂2u
= 2 , 0 < x, 0<t
∂t ∂ x
u(x, 0) = 0, 0≤x
∂u
(0, t) = g(u(0, t)) − f (t)
∂x
where g is some unknown function satisfying g(0) = 0. The goal is to estimate
the unknown function g which governs the thermal transfer at the surface,
given the ambient temperature history f (t) and the surface temperature his-
tory ϕ(t) = u(0, t), both of which can be measured by thermal probes. This is
an inverse problem of the classic type of determining physics from measure-
ments and leads, via a Laplace transform analysis, to the equation
t
g(ϕ(t)) = f (t) + ϕ (τ )k(t − τ )dτ, t > 0
0
√
where the kernel k is given by k(t) = 1/ πt. From this equation values of
the unknown function g may be determined in terms of measured values of
ϕ(t) := u(0, t) and the unbounded operator
t
(Lϕ)(t) = ϕ (τ )k(t − τ )dτ. (1.6)
0
Similar problems arise for other simple geometries and boundary condi-
tions. We illustrate the identification of a nonlinear boundary condition in
a simple heat transfer problem for a hollow ball. Consider a spherical shell
b ≤ r ≤ a, which is initially at temperature zero, whose inner surface r = b
is constantly at temperature zero, and whose outer surface r = a is heated
under the influence of a spatially homogeneous ambient temperature f (t). In
suitable units the temperature u(r, t) satisfies
2
ut = ∆u = urr + ur , b < r < a, t > 0
r
u(r, 0) = 0, b ≤ r ≤ a
u(b, t) = 0, t > 0
r2 U + 2rU − sr2 U = 0
U (b) = 0, U (a) = G
where G is the Laplace transform of g(u(a, t)) − f (t) and primes signify differ-
entiation with respect to r. It is a routine, though tedious, matter to show that
√ √
a2 G e sr /r − e− s(r−2b) /r
U= √ √ √ √
(a s − 1)ea s + (a s + 1)e−(a−2b) s
We see that identifying the unknown function g which controls the heat trans-
fer at the surface involves computing values of the function
t
(Lϕ)(t) = ϕ (τ )k(t − τ )dτ.
0
1.4 Overview
In this chapter we have acted a bit cavalierly in our use of mathematics as
our chief concern was to provide a variety of examples of unbounded linear
operators arising in application areas, particularly in inverse problems. The
remainder of this monograph has a more rigorous tone. In the coming chap-
ters certain aspects of rigorous operator approximation theory are developed
and applied to the general problem of stabilizing the evaluation of a closed
unbounded linear operator defined on a dense subspace of a Hilbert space.
In the next chapter, some basic results of the theory of Hilbert spaces
and operator theory are briefly reviewed. Among the tools that shall prove
indispensable for the later development are certain weak compactness and
convergence notions, the spectral theorem for a self-adjoint bounded linear
operator, von Neumann’s theorem on the boundedness of a type of resolvent
of a closed unbounded operator, and his theorem on the convergence of the
alternating projection algorithm. These topics and related ideas will be put
to use in the general approximation theory developed in later chapters.
Chapter 3 presents an abstract general scheme for stabilized evaluation of
a closed densely defined linear operator. In the general scheme von Neumann’s
theorem is teamed up with the spectral theorem for bounded self-adjoint linear
operators to produce “smoothed” data that converge in operator graph norm
to the true data vector. Some general convergence results are proved beginning
with the case of error-free data. For the case of rough, approximately specified
data, convergence is proved under certain a priori assumptions relating the
stability parameter to the error level in the data and to certain general features
of the approximation scheme. Various special instances of the general scheme,
including methods in which the stability parameter assumes a continuous
range of values as well as methods in which an iteration number assumes the
role of the stability parameter, are briefly discussed.
The best-known stabilized evaluation method, the Tikhonov-Morozov
method, is taken up in Chapter 4. Some special features of this method,
1.5 Notes 17
1.5 Notes
for all x ∈ H. With this norm a Hilbert space is a complete normed linear
space, that is, a sequence {xn } ⊂ H which is Cauchy in this norm (that is,
xn − xm → 0 as n, m → ∞) converges to some vector in H. A typical exam-
ple of a Hilbert space is the space L2 (Ω) consisting of Lebesgue measurable
square integrable real-valued functions on a domain Ω ⊆ Rn with the inner
product
f, g = f (x)g(x)dx
Ω
20 2 Hilbert Space Background
We remind the reader that any inner product space may be completed to
form a Hilbert space by a process of endowing the space of certain equivalence
classes of Cauchy sequences with an appropriate inner product (see, e.g., [50]).
The inner product in a Hilbert space satisfies the Cauchy-Schwarz
inequality
|
x, y| ≤ xy
and hence the inner product is continuous with respect to the associated norm.
Since x2 =
x, x, one quickly finds that the norm associated with an inner
product satisfies the Parallelogram Law:
where at most countably many terms in the sum are nonzero. This is known as
Bessel’s inequality. An orthonormal set {xα }α∈A is called complete if
x, xα =
0 for every α ∈ A implies that x = 0. If {xα }α∈A is a complete orthonormal
set, then every x ∈ H has a convergent Fourier expansion
x=
x, xα xα
α∈A
and
x2 = |
x, xα |2 .
α∈A
S ⊥ = {y ∈ H :
x, y = 0, for all x ∈ S}.
ym − yn 2 = ym − x + x − yn 2
x2 2 = x − x1 2 = d2 ≤ x − (x1 +
x2 , yy)2 = x2 2 − |
x2 , y|2
and hence
x2 , y = 0, that is, x2 ∈ S ⊥ .
If the subspace S contains a complete orthonormal set {xn }, then PS may
be computed by use of the Fourier projection:
PS x =
x, xn xn
n
for each x ∈ H.
If S is a subspace of H, then S ⊥⊥ = S, the (norm) closure of S. In fact,
⊥⊥
S = (S ⊥ )⊥ is a closed subspace of H containing S and therefore, S ⊆ S ⊥⊥ .
On the other hand, if x ∈ S ⊥⊥ , then x = x1 + x2 where x1 ∈ S ⊆ S ⊥⊥ and
⊥
x2 ∈ S = S ⊥ . Hence, x2 = x − x1 ∈ S ⊥⊥ and so x2 ∈ S ⊥⊥ ∩ S ⊥ = {0}.
That is, x = x1 ∈ S, and hence S ⊥⊥ ⊆ S.
Each y ∈ H gives rise to a real-valued linear functional l : H → R defined
by l(x) =
x, y. The Cauchy-Schwarz inequality guarantees that this linear
functional is continuous (with respect to the norm). The Riesz Representation
Theorem asserts that every continuous linear functional on H has this form.
Indeed, if l : H → R is a nonzero continuous linear functional, then the
nullspace of l, defined by
N (l) = {x ∈ H : l(x) = 0}
x − l(x)z, z = 0,
of the space C02 (Ω) of twice continuously differentiable functions with com-
pact support contained in the interior of the bounded domain Ω ⊆ Rn with
piecewise smooth boundary ∂Ω. If u, v ∈ C02 (Ω), then u and v vanish on the
boundary ∂Ω and Green’s identity gives
(−∆u)vdx = ∇u · ∇vdx =
u, v
Ω Ω
for all v ∈ C02 (Ω). On the other hand, for any fixed f ∈ L2 (Ω), the linear
functional l : H01 (Ω) → R defined by
l(v) = f (x)v(x)dx
Ω
for all v ∈ H01 (Ω). This u ∈ H01 (Ω) is called the weak solution of the elliptic
boundary value problem (2.1). “Weak” notions play an important role in the
sequel.
x − xn 2 =
x − xn , x − xn = x2 − 2
xn , x + x2 → 0,
that is, weak convergence along with convergence of the norms implies strong
convergence. The Cauchy-Schwarz inequality also implies that the norm is
weakly lower semi-continuous, namely, if xn x, then
and hence x ≤ lim inf xn . The uniform boundedness principle implies
that every weakly convergent sequence is bounded (see [26], p. 183, for an
elementary proof that does not employ the uniform boundedness principle).
The next basic result is a partial converse and plays an important role in later
developments.
Theorem 2.1. Each bounded sequence in a Hilbert space has a weakly con-
vergent subsequence.
v → lim
zn , v
n
|
zn − zm , v| = |
zn − zm , v − yN +
zn − zm , yN |
≤ M + |
zn − zm , yN | < (M + 1)
l(v) = lim
zn , v
n
24 2 Hilbert Space Background
then one sees that l is a linear functional and l(v) ≤ M v and hence l is
continuous. By the Riesz representation theorem there is a w ∈ H such that
l(v) =
w, v, that is
lim
zn , v =
w, v
n
N (T ) = {x ∈ H1 : T x = 0}
and the range
R(T ) = {T x : x ∈ H1 }.
There is a useful relationship between these subspaces for an operator and its
adjoint. We see that w ∈ R(T )⊥ if and only if
0 =
T x, w =
x, T ∗ w
2.3 Bounded Linear Operators 25
N (T ∗ ) = R(T )⊥ , N (T ∗ )⊥ = R(T )
N (T ) = R(T ∗ )⊥ , N (T )⊥ = R(T ∗ ).
m
Tx = lk (x)vk
k=1
for certain vectors {u1 , . . . , um } ⊂ H1 . Note that every finite rank operator is
continuous, but more is true. If {xn } is a weakly convergent sequence in H1 ,
say xn w, then
m
m
T xn =
xn , uk vk →
w, uk vk = T w,
k=1 k=1
and hence a finite rank operator T maps weakly convergent sequences into
strongly convergent sequences. Finite rank operators are a special case of an
important class of linear operators that enjoy this weak-to-strong continuity
property.
Note that every finite rank operator is compact and every compact oper-
ator is a fortiori continuous. Also, limits of finite rank operators are compact
(the proof is a standard “/3” argument). More precisely, if {Tn } is a sequence
of finite rank operators converging in operator norm to an operator T , then
T is compact.
Compact operators acting on a Hilbert space have a particularly sim-
ple structure expressed in terms of certain characteristic subspaces known as
eigenspaces. The eigenspace of a linear operator T : H → H associated with
a scalar λ is the subspace
N (T − λI) = {x ∈ H : T x = λx}.
In general, an eigenspace may be trivial, that is, it may consist only of the
zero vector. If N (T − λI) = {0}, we say that λ is an eigenvalue of T . If T is
self-adjoint, then the eigenvalues of T are real numbers and vectors in distinct
eigenspaces are orthogonal to each other. If T is self-adjoint, compact, and
of infinite rank, then the eigenvalues of T form a sequence of real numbers
{λn }. This sequence converges to zero, for taking an orthonormal sequence
{xn } with xn ∈ N (T − λn I) we have λn xn = T xn → 0 since xn 0 (a
consequence of Bessel’s inequality). Since xn = 1, it follows that λn → 0.
The fact that T xn → 0 for a sequence of unit vectors {xn } is an abstract
version of the Riemann-Lebesgue Theorem.
The prime exemplar of a compact self-adjoint operator is the integral
operator on the real space L2 [a, b] generated by a symmetric kernel k(·, ·) ∈
L2 ([a, b] × [a, b]):
b
(T f )(s) = k(s, t)f (t)dt.
a
This operator is bounded since for f ∈ L2 [a, b] and k(·, ·) ∈ L2 ([a, b] × [a, b])
the Cauchy-Schwarz inequality gives
2.3 Bounded Linear Operators 27
b
b
T f 2 = a
| a k(s, t)f (t)dt|2 ds
b
b
b
≤ ( |k(s, t)|2 dt a |f (t)|2 dt)ds
a a
b
b
= f 2 a a
|k(s, t)|2 dtds.
It is also compact because it is the limit in operator norm of the finite rank
operators
b
TN f = cn,m f (t)φm (t)dtφn
n.m≤N a
where {φn }∞ 2
1 is a complete orthonormal sequence in L [a, b] and
∞
k(s, t) = cn,m φn (s)φm (t)
n,m=1
is the Fourier expansion of k(·, ·) relative to the orthonormal basis{φn (s)φm (t)}
for L2 ([a, b] × [a, b]).
Operators with weakly singular kernels form another class of compact
operators on the space L2 [a, b]. These operators are important in certain ap-
plication areas, particularly potential theory. As an illustration we consider
the one-dimensional case. Given a constant α with 0 < α < 1 and a function
k(·, ·) ∈ L2 ([a, b] × [a, b]), the integral operator T defined by
b
k(s, t)
(T f )(s) = f (t)dt
a |s − t|α
is said to have a weakly singular kernel. Such an operator is bounded on
L2 [a, b] since if f ∈ L2 [a, b] and if M is an essential bound for |k(·, ·)|, then
by the Cauchy-Schwarz inequality
b b 2
k(s, t)
T f =
2
f (t)dt ds
a a |s − t|α
b b
1 b
|f (t)|2
≤M 2
dt dt ds
a a |s − t|α a |s − t|α
M 2 (b − a)(1−α) b b
|f (t)|2
≤ dtds
1−α a a |s − t|α
M 2 (b − a)(1−α) b b
= |f (t)|2 |s − t|−α dsdt
1−α a a
M 2 (b − a)2(1−α)
≤ f 2 .
(1 − α)2
28 2 Hilbert Space Background
where ⎧
⎨ |s − t|, |s − t| ≥ 1
n
rn (s, t) =
⎩
1/n, |s − t| < 1
n.
b 2
k(s, t)
= (1 − n |s − t| )f (t)dt ds
α α
|s − t|α
a In
2
b
|k(s, t)|
≤C |f (t)|dt ds
a In |s − t|α
b
dt |f (t)|2
≤C dt ds
a In |s − t|α In |s − t|α
C b
|f (t)|2 C
≤ dtds ≤ 1−α f 2
n1−α a In |s − t|α n
vj = 1,
vi , vj = 0, for i = j, and T vj = λj vj .
(here the eigenvalues are repeated in accordance with the dimension of the
associated eigenspace). The closure of the range of T is the closure of the span
of this sequence of eigenvectors. Since T is self-adjoint, N (T )⊥ = R(T ); the
decomposition theorem then gives the representation
∞
w = Pw +
w, vj vj
j=1
This result is known as the Spectral Theorem. The spectral theorem enables
the development of a remarkably useful functional calculus of operators. Given
a real-valued function which is defined and bounded on the set of eigenvalues
{λj }∞
j=1 of a compact self-adjoint operator T : H → H, the operator f (T ) :
H → H is defined by
∞
f (T )x = f (λj )
x, vj vj
j=1
where {vj } are orthonormal eigenvectors associated with the respective (re-
peated as necessary) eigenvalues. Since f is bounded, Bessel’s inequality en-
sures that this series converges, that is, f (T ) : H → H is well-defined. It is a
routine matter to verify that the operator f (T ) so defined is linear, bounded
(in fact, f (T ) = max |f (λj )|) and self-adjoint.
j
The spectral theorem for compact self-adjoint linear operators can be
phrased in another way that is suggestive of its generalization to bounded
self-adjoint linear operators. Denote by Pλj the orthogonal projection
onto the
eigenspace N (T − λj I), in particular, P0 = PN (T ) . Now let Eλ = λj ≤λ Pλj .
Then Eλ is a projection operator satisfying Eλ = O for λ < λmin and Eλ = I
for λ ≥ λmax , where [λmin , λmax ] is the smallest closed interval containing
the eigenvalues of T . The spectral representation of T may now be written in
terms of the family of projections {Eλ } as
30 2 Hilbert Space Background
b
Tx = λj (Eλj − Eλj−1 )x = λdEλ
j a
for any a ≤ λmin and b > λmax , in which the operator-valued integral is to be
interpreted in the Riemann-Stieljes sense. We note that the projection-valued
function Eλ is right continuous in the sense that limλ→λ+ Eλ x = Eλ0 x for
0
each x ∈ H.
If T is a bounded linear operator, then the spectrum of T , denoted σ(T ), is
the set of all complex numbers λ such that (T − λI) does not have a bounded
inverse. The spectrum of a bounded self-adjoint linear operator is a nonempty
compact subset of real numbers contained in the interval [mT , MT ] where
mT = min {
T x, x : x = 1}
and
MT = max {
T x, x : x = 1}.
The spectral theorem asserts that there exists a family {Eλ }λ∈R of projection
operators on H, called a resolution of the identity generated by T , having the
properties enunciated in the previous paragraph, such that
b
Tx = λdEλ
a
f (T ) = max |f (λ)|.
λ∈σ(T )
A particularly important consequence of this result is the fact that if {pn (λ)}
is a sequence of easily computable functions (e.g., polynomials) which are
continuous and uniformly convergent to f on a closed interval [a, b] containing
the spectrum of T , then
pn (T ) − f (T ) → 0
as n → ∞.
Lx, y =
x, y ∗
for some vector y ∗ ∈ H1 and all x ∈ D(L). The vector y ∗ is then uniquely
defined and we set L∗ y = y ∗ . Then the operator L∗ : D(L∗ ) ⊆ H2 → H1 is
linear. If L is both densely defined and closed, then L∗ is also densely defined
(see below).
As a simple example, we determine the adjoint of the differentiation oper-
ator on H = L2 [0, 1]. Let
Lf, g =
f, g ∗
1
for all f ∈ D(L). In particular, for f ≡ 1 we find that g ∗ (t)dt = 0. Now let
0
t
h(t) = − g ∗ (s)ds.
0
Lf, h − g = 0 for all f ∈ D(L). But R(L) contains all continuous functions
and hence g = h ∈ D∗ . We conclude that D(L∗ ) = D∗ and L∗ g = −g .
A linear operator L : D(L) ⊆ H1 → H2 is called closed if its graph G(L) =
{(x, Lx) : x ∈ D(L)} is a closed subspace of the product Hilbert space H1 ×H2 .
This means that if {xn } ⊂ D(L), xn → x ∈ H1 , and Lxn → y ∈ H2 , then
32 2 Hilbert Space Background
(x, y) ∈ G(L), that is, x ∈ D(L) and Lx = y. For example, the differentiation
operator defined in the previous paragraph is closed. For suppose {fn } ⊆ D(L)
and fn → f and fn → g, in each case the convergence being in the L2 [0, 1]
norm. Since x
fn (x) = fn (0) + fn (t)dt
0
we see that the sequence of constant functions {fn (0)} converges in L2 [0, 1]
and hence the numerical sequence {fn(0)} converges to some real number C.
x
Now define h ∈ D(L) by h(x) = C + g(t)dt. Then, for any t ∈ [0, 1], we
0
have by use of the Cauchy-Schwarz inequality
x
|fn (x) − h(x)| = |fn (0) − C + 0 fn (t) − g(t)dt|
1
≤ |fn (0) − C| + 0
|fn (t) − g(t)|dt
∂u ∂2u
= + g(s), 0 < s < π, 0 < t < 1,
∂t ∂s2
where u(s, t) is subject to the boundary and initial conditions
one wishes to reconstruct the source distribution g(s) from the spatial temper-
ature distribution x(s) = u(s, 1), one is led by formal separation of variables
techniques, as was seen in Chapter 1, to the representation
∞
n2
g(s) = (Lx)(s) = an sin ns
n=1
1 − e−n2
where π
2
an = x(s) sin nsds.
π 0
That is, g = Lx, where L is the linear operator on L2 [0, π] with domain
∞ π
2
D(L) = x ∈ L2 [0, π] : m4 a2m < ∞, am = x(s) sin msds
m=1
π 0
2.4 Unbounded Operators 33
which is defined above. Note that D(L) is dense in L2 [0, π] since all finite
linear combinations of the eigenfunctions φn (s) = sin ns are contained in
D(L). Moreover, L is unbounded since φn 2 = π/2, while
πn2
Lφn 2 = →∞
2(1 − e−n2 )
as n → ∞. Finally, L is a closed operator, that is, its graph is a closed subspace
of L2 [0, π]×L2 [0, π]. Indeed, if {xk } ⊆ D(L) and xk → x in L2 [0, π] as k → ∞,
while Lxk → g in L2 [0, π], then
g − Lxk , φn → 0 as k → ∞ for each n. Also,
n2 n2
g − Lxk , φn =
g, φn − −n2
xk , φn →
g, φn −
x, φn .
1−e 1 − e−n2
Therefore,
n2
g, φn =
x, φn
1 − e−n2
and hence
∞
∞
2
m2
m4 |
x, φm |2 ≤ |
g, φm |2 = g2 .
m=1 m=1
1 − e−m2
by
Lf = F −1 {fˆ(ω) cosh ω}
where F is the Fourier transform operator and F{f } = fˆ. If {fn } ⊆ D(L)
and Lfn → g in L2 (R), then
and hence {fˆn (ω) cosh ω} converges to ĝ and to fˆ cosh ω. Therefore, fˆ cosh ω ∈
L2 (R), that is, f ∈ D(L), and
Therefore, L is closed.
The next proposition shows that closed linear operators are common.
34 2 Hilbert Space Background
Lu, y = limn→∞
Lu, yn
= limn→∞
u, L∗ yn =
u, z
The four fundamental equations relating the ranges and nullspaces that we
discussed previously for bounded linear operators continue to hold for closed
densely defined linear operators. For example, if w ∈ R(L)⊥ , then
Lx, w = 0 =
x, 0
for all x ∈ D(L) and hence w ∈ D(L∗ ) and L∗ w = 0, that is, R(L)⊥ ⊆ N (L∗ ).
On the other hand, if w ∈ D(L∗ ) and L∗ w = 0, then for all x ∈ D(L)
Lx, w =
x, L∗ w = 0
and hence N (L∗ ) ⊆ R(L)⊥ . Therefore, R(L)⊥ = N (L∗ ). It then follows that
The example just given shows that a symmetric linear operator is not
necessarily bounded. The Hellinger-Toeplitz Theorem gives a sufficient con-
dition for a symmetric operator to be bounded: a symmetric linear operator
whose domain is the entire space is bounded. This theorem is a special case
of the next result because a symmetric linear operator is, as is every adjoint
operator, necessarily closed (see Theorem 2.4).
If a linear operator L : D(L) ⊆ H1 → H2 is closed, then D(L) is a Hilbert
space when endowed with the graph inner product [·, ·] defined by:
[x, y] =
x, y +
Lx, Ly for x, y ∈ D(L).
Indeed, if {xn } ⊆ D(L) is Cauchy in the graph norm, then {xn } is Cauchy
in H1 and {Lxn } is Cauchy in H2 . Therefore, there is an x ∈ H1 such that
xn → x and a y ∈ H2 such that Lxn → y. Since L is closed, x ∈ D(L) and
y = Lx and hence {xn } converges to x ∈ D(L) in the graph norm, that is,
D(L) is complete with respect to the graph norm.
If L is closed and everywhere defined, i.e., D(L) = H1 , then since the
graph norm dominates the norm on H1 , we find, by the corollary to Banach’s
theorem, that the norm in H1 is equivalent to the graph norm. In particular,
the operator L is then bounded:
Theorem 2.5 (Closed Graph Theorem). A closed everywhere defined lin-
ear operator is bounded.
The arguments developed by von Neumann in his study of unbounded
operators may properly be called Cartesian as they rely on clever use of the
graph of the linear operator L : D(L) ⊆ H1 → H2 , considered as a subspace
of the product Hilbert space H1 × H2 , along with like treatment of the adjoint
operator L∗ . We now present an important theorem of von Neumann [47]
concerning the operators
:= (I + L∗ L)−1
L
and LL
H1 × H2 = G(L) + G(L)⊥ .
However,
G(L)⊥ = {(−L∗ y, y) : y ∈ D(L∗ ) ⊆ H2 }. (2.1)
−L∗ y, x +
y, Lx =
−L∗ y, x +
L∗ y, x = 0
and hence (−L∗ y, y) ∈ G(L)⊥ . On the other hand, if (u, v) ∈ G(L)⊥ , then for
all x ∈ D(L)
x, u +
Lx, v = 0
that is,
Lx, v =
x, −u
(u, v) = (−L∗ v, v)
for some x ∈ D(L) and y ∈ D(L∗ ). But then y = −Lx and h = x − L∗ y, that
is, x ∈ D(L∗ L) and h = (I + L∗ L)x. Hence (I + L∗ L)−1 is defined on all of
H1 and, since h2 ≥ x2 ,
= (I + L∗ L)−1 ≤ 1.
L
We also have
2 =
L(I + L∗ L)−1 x, L(I + L∗ L)−1 x =
Lx,
LLx x − Lx
≤ x2
is self-adjoint.
and hence L
2.4 Unbounded Operators 37
The ideas of this proof may also be used to provide a result of independent
interest regarding the adjoint. Suppose the operator T : H2 × H1 → H1 × H2
is defined by
T (y, x) = (−x, y).
Then, as we have seen in the proof of the last theorem (see 2.1),
L∗ y, 0 −
y, z = 0
L∗ y, x =
y, u
for all y ∈ D(L∗ ). But this is the defining condition for L∗∗ , that is,
and L
We point out a few basic properties of the operators L that will be
useful later.
= D(LL∗ ) and R(L)
Lemma 2.8. R(L) = D(L∗ L).
This result extends easily to positive integral powers of the operator. In-
⊆ D(LL∗ ), it follows im-
deed, if n is a positive integer, then, since R(L)
∗ n
mediately that R(L ) ⊆ D((LL ) ). On the other hand, supposing that
n
k ),
(I + LL∗ )x = x + LL∗ x ∈ D((LL∗ )k ) ⊆ R(L
that is
k ) = R(L
x ∈ (I + LL∗ )−1 R(L k+1 ).
Proof. The first equality follows from the fact that L∗ Lx = 0 if and only if
= x.
x = (I + L∗ L)x, that is, if and only if Lx
Suppose x ∈ D(L L) and let w = x + L∗ Lx. Then (I − L)w
∗ = L∗ Lx and
∗
hence R(L L) ⊆ R(I − L). On the other hand, if y = z − Lz for some z ∈ H1 ,
then
∈ D(L∗ L)
z − y = Lz
⊆ R(L∗ L).
and (I + L∗ L)(z − y) = z. That is, y = L∗ L(z − y), i.e., R(I − L)
N (LL∗ ) = N (I − L)
and R(LL∗ ) = R(I − L).
1/2 x = L
LL 1/2 Lx
(a more general result is proved in Theorem 3.2 of the next chapter). Using
this we have:
1/2 : H1 → H2 is bounded and
Lemma 2.10. LL
1/2 )∗ (LL
(LL 1/2 ) = I − L
2.4 Unbounded Operators 39
(LL 1/2 )x, y =
L
1/2 )∗ (LL
1/2 Lx, L∗ LLx, y.
But,
L∗ LLx = (I − L)x,
= L∗ LLx
1/2 )∗ (LL
(LL 1/2 )x, y =
(I − L)x,
y
(LL 1/2 ) = I − L.
1/2 )∗ (LL
L∗ Lx2 ≥
L∗ Lx, x = Lx2
< 1.
and hence I − L
We also note the following fact about the operator LL∗
and L∗ y = −y . Therefore,
cosh (t − 1) cosh s/ sinh 1, s ≤ t
G(s, t) =
cosh (s − 1) cosh t/ sinh 1, t < s
where
n2
ϕn (x) = sin nx and cn = .
1 − e−n2
It is easy to see that
∞
=
Lf (1 + c2n )−1
f, ϕn ϕn
n=1
2.5 Pseudo-inversion 41
2.5 Pseudo-inversion
We are interested in generalized solutions of the equation Lx = y where
L : D(L) ⊆ H1 → H2 is a closed densely defined linear operator and y ∈ H2 .
Of course the equation has a solution if and only if y ∈ R(L), however we can
associate generalized solutions with any y in the dense subspace R(L)+R(L)⊥
of H2 . Indeed, if y ∈ R(L) + R(L)⊥ , then P y ∈ R(L) where P is the projector
of H2 onto R(L), the closure of the range of L. In this case the equation
Lx = P y
has solutions and we call any such solution a least squares solution of the
original equation. If x ∈ D(L) is a least squares solution then
which accounts for the “least squares” terminology. Among the possibly infi-
nitely many least squares solutions there can be at most one lying in N (L)⊥ .
In fact, if u and v are least squares solutions, then u − v ∈ D(L) and
L(u − v) = Lu − Lv = P y − P y = 0
and hence if u and v lie in N (L)⊥ , then u − v ∈ N (L) ∩ N (L)⊥ = {0}. On the
other hand, if x is a least squares solution and x = x1 + x2 ∈ N (L) + N (L)⊥ ,
42 2 Hilbert Space Background
LL† y = PR(L) y
L† Lx = PN (L)⊥ x
PN (L)⊥ x = x2 = x − x1 ∈ D(L).
Also,
L† Lx = L† Lx2 .
However, x2 ∈ N (L)⊥ ∩ D(L) and Lx2 = Lx = PR(L) Lx and hence, x2 =
L† Lx. That is, L† Lx = PN (L)⊥ x. We summarize these results in the following
theorem:
Theorem 2.12. If L : D(L) ⊆ H1 → H2 is a closed densely defined linear
operator, then L† : D(L† ) = R(L) + R(L)⊥ → D(L) ∩ N (L)⊥ is closed and
densely defined. Also
We now give a result that relates the Moore-Penrose inverse of the generally
unbounded operator L∗ L to that of the bounded linear operator I − L.
Theorem 2.13. If L is closed and densely defined, then
− L)
(L∗ L)† = L(I †.
is bounded
Proof. First note that from Lemma 2.9 and the fact that L
+ R(I − L)
= R(I − L) ⊥
† ) = D(L(I
= D((I − L) − L)
† ).
− L)
L(I † z = L(I
− L)
† L∗ Lw
− L)
= L(I † (I + L∗ L − I)w
− L)
= L(I
† (I − L)(I + L∗ L)w
= LP ∗
)⊥ (I + L L)w.
N (I−L
⊥ , i.e.,
therefore, (I + L∗ L)w ∈ N (I − L)
∗ ∗
PN (I−L
)⊥ (I + L L)w = (I + L L)w.
Hence
− L)
L(I † z = L(I
+ L∗ L)w = w = (L∗ L)† z.
2.6 Optimization
We begin with a fairly general situation in which there may be multiple op-
tima. Suppose S ⊆ H is weakly closed and x ∈ H. Then there is a y ∈ S such
that y − x = dist(S, x), where dist(S, x) is the distance from x to S:
d = inf{z − x : z ∈ S},
then there is a sequence {zn } ⊆ S such that lim zn − x = d. The sequence
{zn } then has, by Theorem 2.1, a weakly convergent subsequence, say znk y.
Since S is weakly closed we then have y ∈ S and, by the weak lower semi-
continuity of the norm,
d ≤ y − x ≤ inf znk − x = d.
Note that as a consequence of the continuity of the norm, dist(S, x) also exists
if S is (strongly) closed. The set of all points in S nearest to x is called the
projection of x onto S and is denoted PS (x):
u − v2 = u − x − (v − x)2
≤ 4 × dist(S, x) − 4 × dist(S, x) = 0.
The purely geometric fact that a weakly closed convex set has the nearest
point property has many important applications in optimization theory. As
an illustration we provide a very simple example of optimal control of a one-
dimensional dynamical system. Suppose a unit point mass moves on the real
line and is steered from the origin with initial velocity 1 by a control (i.e.,
external force) u ∈ L2 ([0, 1]). Our interest is in a control that returns the
particle to a soft landing at the origin in unit time while expending minimal
effort. As a measure of effort we take
1
u2 = |u(t)|2 dt.
0
where the dots indicate temporal derivatives. The motion of the particle is
then governed by
Let S be the set of all functions (controls) u ∈ L2 ([0, 1]) for which the equa-
tions (2.3) are satisfied for some trajectory x ∈ H 2 [0, 1]. Note that S is clearly
a convex set. Also, S is weakly closed in L2 ([0, 1]). Indeed, suppose {un } ⊂ S
and un u ∈ L2 ([0, 1]). Let
t s
x(t) = u(τ )dτ ds
0 1
and note that x ∈ H 2 [0, 1], ẍ = u, x(0) = 0, and ẋ(1) = 0. Also, if xn is the
trajectory corresponding to the control un , then since um u,
0 0
1 = ẋn (0) = un (τ )dτ =
−1, un →
−1, u = u(τ )dτ = ẋ(0)
1 1
and hence the set S is weakly closed. Therefore S contains a unique control
that is nearest to the origin, and this of course is the minimal effort control
which solves the optimal control problem.
A simple sketch convinces one that the nearest point in a weakly closed
convex set is characterized by the following useful inequality.
Theorem 2.14. If S is weakly closed and convex, then x0 = PS (x) if and
only if x0 ∈ S satisfies
x − x0 , y − x0 ≤ 0 for all y ∈ S.
Proof. If x0 satisfies the stated inequality, then for any y ∈ S,
= y − x0 2 + 2
y − x0 , x0 − x + x0 − x2 ≥ x − x0 2
= x − x0 2 − 2t
x − x0 , y − x0 + t2 x0 − y2 .
46 2 Hilbert Space Background
So if
x − x0 , y − x0 > 0 then x − y(t) < x − x0 for t sufficiently small,
contradicting the fact that x0 = PS (x).
=
PS x, y −
PS x, PS y =
PS x, y − PS y = 0
and therefore PS is self-adjoint.
T 1 = P1 , T2 = P2 P1 , T3 = P1 P2 P1 , etc.
that is, ⎧
⎨ (P2 P1 )n/2 , n even
Tn =
⎩
P1 (P2 P1 )(n−1)/2 , n odd
One sees easily that
⎧
⎨ Tm+n , m+n odd
∗
Tm Tn =
⎩
Tm+n−1 , m + n − 1 odd
and hence
∗ ∗
Tm x − Tn x2 =
Tm Tm x, x −
Tm Tn x, x −
Tn∗ Tm x, x +
Tn∗ Tn x, x
=
T2m−1 x, x − 2
Tq(m,n) x, x +
T2n−1 x, x
(2.5)
where q(m, n) = m + n or m + n − 1, whichever is odd. Now for any positive
integer j we have
T2j−1 x, x = Tj x2 ≤ Tj−1 x2 =
T2j−3 x, x
and therefore {
T2j−1 x, x} is a decreasing sequence of non-negative numbers
and hence
a(x) = lim
T2j−1 x, x
j→∞
P1 P x = lim P1 Tn x = lim Tn x = P x
n→∞ n→∞
In particular we see (setting w = z) that PS(z) (x) = PS(z) (y) if and only
if x − y ∈ S ⊥ .
Lemma 2.20. If A1 = S1 (z) and A2 = S2 (y) are non-disjoint affine sets then
PA2 (PA1 (x)) − PA1 ∩A2 (x) = PS2 PS1 (x − PA1 ∩A2 (x)).
= PA1 ∩A2 (x), then by the previous lemma, we have, since x
Proof. Let x ∈
A1 ∩ A2 :
PA2 (PA1 (x)) − x
= PS2 PA1 (x) + PS2⊥ x
−x
−x
= PS2 (PS1 x + PS1⊥ x )
Having dispensed with these preliminaries we may now prove the conver-
gence of the alternating projection algorithm for two affine sets. From the
previous two lemmas we have
(PA2 PA1 )2 (x) − PA1 ∩A2 (x) = PA2 PA1 (PA2 PA1 (x)) − PA1 ∩A2 (x)
= PA2 PA1 (PA2 PA1 (x)) − PA1 ∩A2 (PA2 PA1 (x))
= PS2 PS1 (PA2 PA1 (x) − PA1 ∩A2 (PA2 PA1 (x)))
2.7 Notes
The literature on Hilbert spaces and linear operators is wide and deep. Our
treatment in this chapter is close in spirit to works on “applied” or “numer-
ical” functional analysis such as [3], [49], [50], [33] and [36]. The classical
masterpiece by pioneers in the subject [44] is still hard to beat and the work
[35] by a modern master is highly recommended. A proof of Banach’s Theo-
rem (also called the Open Mapping Theorem) can be found, for example, in
[13] p. 57. The usual proof of the uniform boundedness principle relies on a
topological “category” argument. Halmos [26] (Problem 20) called for a more
elementary non-topological proof. Such was supplied by Hennefeld [29] whose
proof turned out to be a rediscovery of a classical argument of Hausdorff [28].
Lucid presentations of fundamental results on closed unbounded opera-
tors, including the closed graph theorem, von Neumann’s Theorem (originally
presented in [47]) and the Hellinger-Toeplitz theorem can be found in [44].
Lemma 2.8 can be extended to positive powers of the operators L and L
(see [21]).
An account of the Moore-Penrose inverse of bounded linear operators,
with special emphasis on approximation methods, can be found in [12]. The
most authoritative source for best approximation in Hilbert space is [7]. The
Alternating Projection theorem for closed subspaces was presented by von
Neumann in his Princeton lectures of 1933 and published in [48]. The identity
in Theorem 2.13 and related results may be found in [17].
Successive projection onto finitely many special affine sets, namely hyper-
planes, is the basis of Kaczmarz’s iterative method for underdetermined linear
systems, which in turn is the foundation of the ART (algebraic reconstruction
technique) method in computed tomography (see, e.g., [16] for a discussion
and additional references). The method has been widely extended to cover
the case of successive projection onto convex sets (see, e.g., [7]) and has been
applied to various nonlinear optimization problems (see, e.g. [46]).
3
A General Approach to Stabilization
Sα(δ) xδ − Lx → 0 as δ → 0.
= (I + L∗ L)−1
L
⊆ [0, 1] and the operator LL
is bounded and self-adjoint with σ(L) : H1 → H2
is bounded with LL ≤ 1. In the same way the operator L : H 2 → H2
defined by
54 3 A General Approach to Stabilization
= (I + LL∗ )−1
L
: H2 → H1 is bounded.
is bounded and self-adjoint and the operator L∗ L
Before proceeding we notice a simple property of these operators.
Theorem 3.1. If f ∈ C[0, 1] and x ∈ D(L), then f (L)x ∈ D(L) and
f (L)Lx = Lf (L)x.
∈ D(L∗ L) and (I +L∗ L)Lx
Proof. For any x ∈ H1 we have Lx = x. Therefore,
if x ∈ D(L), then
= (I + LL∗ )LLx.
Lx = L(I + L∗ L)Lx
We then have
LLx
= (I + LL∗ )−1 Lx = LLx
and from this it follows that
p(L)Lx
= Lp(L)x
f (L)Lx
= lim pn (L)Lx
= lim Lpn (L)x.
n n
f (L)Lx, y
y = limn
Lpn (L)x,
L∗ y
= limn
pn (L)x,
L∗ y.
=
f (L)x,
and
= f (L)Lx.
Lf (L)x
∗ x = L∗ f (L)x
In the same way one finds that f (L)L for all x ∈ D(L∗ ).
√
is bounded
Corollary 3.2. If f (t) = tg(t) where g ∈ C[0, 1], then Lf (L)
and
≤ g∞
Lf (L)
3.1 A General Method 55
∗ (Lf (L))x,
(Lf (L)) Lf (L)z
z =
Lf (L)x,
=
f (L)Lx,
f (L)Lz
=
L
1/2 g(L)Lx, 1/2 g(L)Lz
L
L)Lx,
=
Lg( L)Lz
Lg(
=
LLg( Lg(L)z
L)x,
=
L∗ LLg( g(L)z.
L)x,
But
=I −L
L∗ LL
and therefore,
∗ (Lf (L))x,
(Lf (L)) L)x,
z =
(I − L)g( g(L)z.
and
|tTα (t)| is uniformly bounded for α > 0, t ∈ [0, 1]. (3.3)
Note that, since LL are both (by von Neumann’s theorem) bounded
and Tα (L)
linear operators, the approximations yα given by (3.1) are defined for all
x ∈ H1 , not just for x ∈ D(L), and the mapping x → yα is stable. In particular
this means that the approximations
α (L)x
yαδ = LLT δ
for Lx then consists of a choice of a family {Tα } satisfying (3.2) and (3.3)
matched with a parameter choice strategy α = α(δ) designed to ensure that
δ
yα(δ) → Lx as δ → 0. Before treating some specific cases we establish basic
convergence and stabilization results. We consider first the case of error-free
data x.
Theorem 3.3. Suppose L : D(L) ⊆ H1 → H2 is a closed densely defined lin-
ear operator and {Tα } is a family of continuous real-valued functions defined
on [0, 1] satisfying (3.2) and (3.3).
α (L)x
(i) For all x ∈ H1 , xα = LT → x as α → 0.
But by (3.2) and (3.3), the function 1 − tTα (t) converges in a pointwise and
uniformly bounded manner to the function
1, t = 0
ϕ(t) =
0, t ∈ (0, 1]
then
The spectral theorem applied to the bounded self-adjoint operator L
gives
x − xα → PN (L
) x = 0, as α → 0.
By Lemma (3.1), if x ∈ D(L), then
α (L))x
Lx − yα = L(I − LT α (L))Lx.
= (I − LT
then
The spectral theorem applied to the bounded self-adjoint operator L
gives
Lx − yα → PN (L
) Lx = 0, as α → 0.
To establish the final assertion, note that if {yα } has a bounded sequence,
then it has a weakly convergent subsequence, say yαn w, for some sequence
αn → 0. Now yαn = Lxn where xn = LT α (L)x
and by the properties of {Tα }
n
and the Spectral Theorem,
xn → x − PN (L
) x = x as n→∞
But since the graph of L is closed and convex, and hence weakly closed,
xn → x and Lxn w, and we have x ∈ D(L) and Lx = w. So if x ∈ / D(L),
then {yα } must be unbounded.
3.1 A General Method 57
=
(I − L)T
α (L)(x α (L)(x
− xδ ), LT − xδ )
α (L)
≤ δ 2 ||(I − L)T LT α (L).
α (L)x
Theorem 3.6. If x ∈ D(L∗ L), then Lx − LLT = O(ω(α, 1/2)).
Proof. Write
Lx − LLT = LSα (L)x
α (L)x
where Sα (t) = 1 − tTα (t). Then, on setting w = x + L∗ Lx, we find that
α (L)x
Lx − LLT 2 =
L∗ LSα (L)x,
Sα (L)x
+ Sα (L)w,
=
−Sα (L)x
Sα (L)x
≤
Sα (L)w, Lw
Sα (L)
L
= Sα (L) 1/2 w2 ≤ ω(α, 1/2)2 w2 .
n∈Z
defined by
(Lf )(eiθ ) = |n|fˆ(n) exp(inθ).
n∈Z
3.2 Some Cases 59
where 2π
1
fˆ(n) = f (t)e−int dt .
2π 0
then one sees easily that
|n|
yαδ (eiθ ) = L(I + αL L)∗ −1 δ
f (exp iθ) = fδ (n) exp (inθ).
1 + αn2
n∈Z
∂u ∂2u
= + g(x), 0 < x < π, 0 < t,
∂t ∂x2
and π
∞
2 n2
Lf (x) = sin nx f (s) sin nsds.
π n=1 1 − e−n2 0
where 2
n2 n2
αn = / 1+α .
1 − e−n2 1 − e−n2
Returning to the general situation, we find that since
(I + αL∗ L)−1 = ((1 − α)I + α(I + L∗ L))−1 = L((1 + αI)−1
− α)L
60 3 A General Approach to Stabilization
Note that this class of functions satisfies criteria (3.2) and (3.3) and hence
then we have:
Corollary 3.7. If x ∈ D(L) and x − xδ ≤ δ, then
√
Lxα − Lxδα ≤ δ/ α.
which is clearly not greater than 1 for 0 < ν ≤ 1 and hence for 0 < ν ≤ 1 and
0 < α < 1 inequality (3.8) holds and hence so does the bound (3.7). We may
therefore take
ω(α, ν) = αν
0 < . . . ≤ λn+1 ≤ λn ≤ . . . ≤ λ2 ≤ λ1 .
Then
Lxα = LL(αI −1 x
+ (1 − α)L)
= L(αI −1 Lx
+ (1 − α)L)
∞ λj
=
Lx, uj uj
j=1
α + (1 − α)λj
and hence
∞ (1 − λj )2
Lx − Lxα 2 = α2 |
Lx, uj |2
j=1
(α + (1 − α)λj )2
∞
≥ α2 (1 − λ1 )2 j=1 (α + (1 − α)λj )−2 |
Lx, uj |2 .
for some constant C and all α ∈ (0, 1]. In particular, all of the partial sums of
the above series are uniformly bounded by C. Letting α → 0+ in each of the
individual partial sums shows that
n
λ−2
j |
Lx, uj | ≤ C
2
j=1
∞
for each n and hence the series j=1 λ−2
j |
Lx, uj | is convergent. The vector
2
∞
z= λ−1
j
Lx, uj uj
j=1
Theorem 3.9. If x ∈ D(L) and Lx − Lxα = O(α), then x ∈ D(LL∗ L).
and hence
Lxα = (I + αLL∗ )−1 Lx ∈ D(LL∗ ).
and
Lxα − Lx = −αLL∗ Lxα .
Therefore, by the hypothesis
By Theorem 2.11 we know that LL∗ is closed. This can be seen in another
way by using Theorem 2.7. Indeed, if {yn } ⊆ D(LL∗ ) satisfies, yn → y and
LL∗ yn → p, then, using Theorem (2.7), we have for any u ∈ D(LL∗ ) =
D(L∗∗ L∗ )
LL∗ u, y = limn→∞
LL∗ u, yn = limn→∞
u, L∗∗ L∗ yn
= limn→∞
u, LL∗ yn =
u, p.
3.2 Some Cases 63
Therefore, y ∈ D(LL∗ ) and LL∗ y = p, that is, LL∗ is closed. Hence the
graph of LL∗ is closed and convex, and hence weakly closed. By (3.9) there is
then a sequence αn → 0 with
LL∗ Lxαn w
for some w. But Lxαn → Lx, by Theorem 3.3 (ii). Since the graph of LL∗ is
weakly closed, it follows that Lx ∈ D(LL∗ ) and LL∗ Lx = w. In particular,
x ∈ D(LL∗ L) as claimed.
Cδ 2/(2ν+1) , then
yα(δ)
δ
− Lx = O(δ 2ν/(2ν+1) ).
Note that this theorem gives a best rate of O(δ 2/3 ) for the case ν = 1. In the
next chapter we shall show that this rate is essentially best possible.
1 1
n−1
1 − (1 − t)n
= ≈ (1 − t)j = =: Tn (t).
t 1 − (1 − t) j=0
t
It is easy to see that the family {Tn (t)} satisfies (3.2) and (3.3). Furthermore,
and hence one is led via the general spectral approach to the iterative method
n−1 + Lx
xn = (I − L)x
or equivalently
This gives, for each n, a stable approximation yn = Lxn to the value Lx.
The method (3.10) is a special case (for α = 1, see below) of the iterated
Tikhonov-Morozov method.
Approximation orders that are arbitrarily near to the order of error in the
data, O(δ), are achievable by use of iteration methods. One such method is the
64 3 A General Approach to Stabilization
(I + αL∗ L)xα = x.
n = Lx
(αI + (1 − α)L)x + α(I − L)x
n−1 .
The definition Tn (0) = 0 extends these functions continuously to [0, 1]. Note
that |tTn (t)| ≤ 1 for all n and Tn (t) → 1/t as n → ∞ for each t ∈ (0, 1],
therefore {Tn } satisfies (3.2) and (3.3) and hence the approximations satisfy
Lxn → Lx as n → ∞.
The general stability estimate requires a bound r(n) for the function
1 1−s
(1 − t)Tn (t) = (1 − (1 − s)n )
α s
3.2 Some Cases 65
1−s n
(1 − (1 − s)n ) = (1 − s)j ≤ n
s j=1
where again s = t/(α + (1 − α)t) ∈ [0, 1]. We may assume that α < 1 and then
ν
s
αν (1 − s)n ≤ (1 − s)n sν
(1 − s) + sα
n ν
n ν
≤ = O(n−ν ).
n+ν n+ν
δ −2/(2ν+1) we have
yn(δ)
δ
− Lx = O(δ 2ν/(2ν+1) ).
Note that the restriction ν ≤ 1 is not imposed here and hence rates arbitrarily
close to O(δ) are in principle achievable by the iterated Tikhonov-Morozov
method. In the next chapter we will study the nonstationary Tikhonov-
Morozov method in which the value of the constant α may change from one
iteration to the next.
Of course there are many other possible choices for the family {Tα } that
lead to stabilization methods. For example, another arises from interpolatory
function theory. Let Tn be the polynomial of degree not greater than n that
66 3 A General Approach to Stabilization
It is evident that |tTn (t)| ≤ 1 and therefore {Tn } satisfies (3.3). Also,
⎛ ⎞
n n n
0≤ (1 − βj t) ≤ e−tβj = exp ⎝− βj t⎠ → 0 as n → ∞
j=1 j=1 j=1
for t ∈ (0, 1], and hence {Tn } satisfies (3.2). We therefore immediately obtain
from Theorem 3.3 that the iteratively defined sequence
" #
yn+1 = βn+1 LLx + I − βn+1 L yn , n = 0, 1, . . . y0 = 0 (3.13)
and hence
n
max |(1 − t)Tn (t)| ≤ βj =: σn .
t∈[0,1]
j=1
From the general stability result of the previous section we then obtain
ynδ − yn ≤ δσn1/2 .
However, the function on the right of the inequality above achieves for t ∈ [0, 1]
a maximum value of ν ν (eσn )−ν and hence we may use
" ν #ν
ω(n, ν) = σn−ν .
e
From Theorem 3.5 we obtain:
ν ) for some ν > 0, then
Corollary 3.13. If x ∈ D(L) and Lx ∈ R(L
δ −2/(2ν+1) we have
yn(δ)
δ
− Lx = O(δ 2ν/(2ν+1) ).
Under appropriate conditions one can achieve rates that are arbitrarily
close to optimal for this method by use of an a posteriori choice of the iteration
parameter rather than with an a priori choice as in Theorem 3.14. In fact,
note that ⎛ ⎞
n " #
yn = L ⎝I − ⎠ x = Lxn
I − βj L (3.14)
j=1
where ⎛ ⎞
n "
#
xn = ⎝I − ⎠ x,
I − βj L x0 = 0 (3.15)
j=1
and {xδn } is defined in the same way with x replaced by xδ . The approxima-
tions {xδn } can be compared with the available data xδ in order to monitor
the convergence of {ynδ } to Lx. First note that xδn → xδ as n → ∞ and
$" # $
$ (xδ − xδ )$
xδ − xδ = $ I − βn L
n $ ≤ xδ − xδ .
n−1 n−1
We assume that for a given constant τ > 1, the signal-to-noise ratio is not
less than τ , i.e., we assume that
There is then a first value n = n(δ) ≥ 1 of the iteration index for which
Lemma 3.15. If xn(δ) is given by (3.15) where n(δ) satisfies (3.16), then
x − xn(δ) ≤ (τ + 1)δ.
Proof. From (3.15), using the approximate data xδ in one case and “clean”
data x ∈ D(L) in the other, we have
⎛ ⎞
n(δ)
xδn(δ) − xn(δ) = ⎝I − ⎠ (xδ − x).
(I − βj L)
j=1
Now
x − xn(δ) = xδ − xδn(δ) + x − xδ + xδn(δ) − xn(δ)
"% #
= xδ − xδn(δ) +
n(δ) (x − xδ ).
− βj L)
j=1 (I
≤ 1, we have by (3.16)
Since I − βj L
x − xn(δ) ≤ τ δ + x − xδ ≤ (τ + 1)δ.
"
#1/(2µ+1) "
#2µ/(2µ+1)
1 1
≤ 0
1dEλ z2 0
λ2µ+1 dEλ z2
" #2µ/(2µ+1)
µ+1/2 z2
= z2/(2µ+1) L .
"% #
=
n(δ) L
− βj L) µ Lw = L
µ zn(δ)
j=1 (I
"% #
where zn(δ) =
n(δ) Lw and hence zn(δ) ≤ Lw.
− βj L)
j=1 (I
3.2 Some Cases 69
≤ 1,
However, since x − xn(δ) ≤ (τ + 1)δ and LL
− xn(δ) ), Lx − Lxn(δ)
=
LL(x
≤ (τ + 1)δLx − Lxn(δ) .
that is,
Lx − Lxn(δ) = O(δ µ/(µ+1) ).
then
Lxδn(δ) − Lx = O(δ min((2µ−1)/(2µ),µ/(µ+1)) ).
≥ τ δ − δ = (τ − 1)δ
µ w, then
If x = L
$ $
$ n−1 # $
$ µ" $
xn−1 − x = $ w$ = O(σ −µ ).
I − βj L
$L $ n−1
$ j=1 $
70 3 A General Approach to Stabilization
−µ
However, σn /σn−1 → 1 as n → ∞ and hence σn−1 = O(σn−µ ), therefore
−µ
xn(δ)−1 − x = O(σn(δ) ).
We note that this result says that rates arbitrarily close to optimal may in
principle be obtained by use of the iteration number choice criterion (3.16).
dw
L∗ L = −w
dα
or since w(α) = xδ − z(α)
dz
L∗ L = xδ − z(α). (3.21)
dα
If we approximate the solution of the differential equation by the simple
implicit forward difference method
zn − zn−1
L∗ L = xδ − zn , z0 = 0
h
with step size h, we find on setting β = 1/h and rearranging that
or, on setting β = 1/(αn+1 − αn ), and using the right hand rule on the last
integral, we are led to
zn+1 + βL∗ Lzn+1 = xδ + βL∗ Lzn
which is the iterated Tikhonov-Morozov method. This approach suggests the
use of other closed quadrature rules on the integral above. While we do not
suggest that such rules will lead to methods that offer any computational
advantage over the Tikhonov-Morozov method, it is instructive to see how
the general theory applies to another method. For example, if the trapezoidal
rule is used we are led to the approximation
βL∗ L(zn+1 − zn ) = xδ − (zn+1 + zn )/2
or, setting γ = 2β > 0,
(I + γL∗ L)zn+1 = 2xδ + γL∗ Lzn − zn .
Equivalently,
n (L)x
zn = LT δ
where T0 (t) = 0 and
2 γ(1 − t) − t
Tn+1 (t) = + Tn (t), n = 0, 1, . . . .
γ(1 − t) + t γ(1 − t) + t
One finds immediately that
n
1 γ(1 − t) − t
Tn (t) = 1−
t γ(1 − t) + t
and from this it follows that
2
|(1 − t)Tn (t)| ≤ n.
γ
Therefore, one may take r(n) = O(n) in the general stability estimate. How-
ever, for this method the convergence analysis given previously does not
automatically apply since
γ(1 − t) − t
−1 ≤ <1
γ(1 − t) + t
for t ∈ (0, 1] with the equality holding at t = 1. Therefore, the convergence of
the spectral approximation may fail if the resolution of the identity generated
by L has a jump discontinuity at t = 1. This is equivalent to the condition
that {0} = N (I − L) = N (L) (see Lemma 2.9).
3.3 Notes 73
3.3 Notes
Lardy [34] was the first to exploit von Neumann’s theorem in applications
to series representations for the Moore-Penrose inverse of a closed unbounded
linear operator. The general spectral approach to stabilized approximate eval-
uation based on von Neumann’s theorem was introduced in [15]. Another
approach to general stabilization theory, based on the theory of regulariza-
tion and Theorem 2.10 is suggested in [21].
The best known specific instance of the general method is the Tikhonov-
Morozov method. This method was developed by V.A. Morozov and his
co-workers and is summarized in [39]. A much more extensive treatment of
this method, based on our spectral approach, emerges in the following chap-
ters. See also [40] and [22] for further developments. The line of reasoning in
the proof of Theorem 3.9 is inspired by an argument of Neubauer [41]. The
iterative stabilization method that is motivated by functional interpolation
appears in [18]; the techniques of that paper owe a lot to [27]. It would appear
that other stabilization methods based on numerical integration techniques for
initial value problems for ordinary differential equations could be developed.
In appropriate circumstances the Tikhonov-Morozov method can be
adapted to stably evaluate certain nonlinear operators A. We outline the the-
ory of Al’ber [1] for accomplishing this. Suppose that A : D(A) ⊆ H → H
is a nonlinear monotone operator defined on a subset D(A) of a real Hilbert
space H, that is
Ax − Ay, x − y ≥ 0
x − xδ ≤ δ
A(u + tv) Au as t → 0+
xδα + αAxδα = xδ
is a stable operation. Let xα be the solution of the same equation using the
“clean” data x:
xα + αAxα = x.
The goal is to show that Axδα → Ax if α = α(δ) → 0 in some appropriate
sense as δ → 0. First, we show that Axα → Ax as α → 0.
Note that, by the monotonicity of A
0 ≤
Axα − Ax, xα − x
= −α
Axα − Ax, Axα
≤ −αAxα 2 + αAxAxα
and hence
Axα ≤ Ax.
From the definition of xα we then have
and hence xα → x as α → 0.
Suppose v ∈ H is arbitrary and t ≥ 0. Then
0 ≤
Axα − A(x + tv), xα − (x + tv)
=
Axα , xα − x − t
Axα , v −
A(x + tv), xα − x − tv.
Axαn y.
Therefore, taking limits as αn → 0 above, and using the fact that xαn → x,
we arrive at
0 ≤ −
y, v +
A(x + tv), v =
A(x + tv) − y, v.
0 ≤
Ax − y, v
3.3 Notes 75
Axα Ax as α → 0.
However, since Axα ≤ Ax, it follows from the weak lower semicontinuity
of the norm that Axα → Ax and hence
Axα → Ax as α → 0.
and
Axδα − Axα = α−1 (xδ − xδα ) + (xα − x)
(3.22)
≤ δ/α + α−1 xα − xδα .
But, since
xδα − xα + α(Axδα − Axα ) = xδ − x,
one finds, using the monotonicity of A,
xδα − xα 2 ≤ xδα − xα 2 +
Axδα − Axα , xδα − xα
=
xδ − x, xδα − xα ≤ δxδα − xα
Axδα − Axα → 0.
Axδα → Ax.
4
The Tikhonov-Morozov Method
= minz∈D(L) Φα (z; xδ )
+ α
Lxδα + tLz, Lxδα + tLz
and hence the previous condition is equivalent to
xδα − xδ , z = −α
Lxδα , Lz
for all z ∈ D(L). Therefore, Lxδα ∈ D(L∗ ), that is, xδα ∈ D(L∗ L), and
(I + αL∗ L)xδα = xδ .
We note that (I + αL∗ L)√has a bounded inverse by von Neumann’s theorem
(applied to the operator αL). Therefore, the unique minimizer of Φα (·; xδ )
is given by
xδα = (I + αL∗ L)−1 xδ .
Besides providing additional insight into the Tikhonov-Morozov method,
this theorem suggests the possibility of computing the Tikhonov-Morozov ap-
proximation by a finite element method. This theme will be developed later.
In Theorem 3.8 it was shown that with exact data x ∈ D(L) a convergence
rate of O(α) is achievable with the Tikhonov-Morozov method. We now show
that, except in a trivial case, this rate is best possible. We will denote the
Tikhonov-Morozov approximation with exact data x ∈ D(L) by xα . That is,
xα = (I + αL∗ L)−1 x.
First we note that the order of approximation O(α) is essentially best possible.
4.1 The Tikhonov-Morozov Method 79
Proof. Define eα by
eα = Lx − Lxα = Lx − LL[αI −1 x
+ (1 − α)L]
= {I − L[αI −1 }Lx,
+ (1 − α)L]
and hence
α = α(I − L)Lx.
[αI + (1 − α)L]e
This result raises the question whether the rate O(δ 2/3 ) obtained in Corol-
lary 3.10 is also best possible when dealing with inexact data. It transpires
that this is the case, at least for the very important class of operators L for
which L∗ L has compact resolvent, that is, those operators L for which L is
compact.
To see that the rate O(δ 2/3 ) can not be improved, we begin with an
estimate for the stabilization parameter.
& 2'
Lemma 4.3. If x ∈ / N (L) and Lxδα − Lx = o δ 3 for all xδ satisfying
& 2 '
x − xδ ≤ δ, then α = o δ 3 .
Proof. To see this, let xδ = x − δu, where u is a unit vector and let eδα =
Lxδα − Lx. Then
" #
δ = [αI + (1 − α)L]
[αI + (1 − α)L]e LL(αI
+ (1 − α) −1 x − Lx
L)
α
L(αI
− δ[αI + (1 − α)L]L −1 u
+ (1 − α)L)
− I)Lx − δLLu.
= α(L
& 2'
Since eδα = o δ 3 , by assumption, and since
& '
≤ δLL
δLLu = o δ 23 ,
80 4 The Tikhonov-Morozov Method
we find that
α − I)Lx → 0, as δ → 0.
2 (L
δ3
" #
But, by assumption, x ∈ − I)L) and hence α = o δ 23 .
/ N (L) = N ((L
"We2
#now show that for a wide class of operators the order of convergence
O δ 3 can not be improved. We only consider the important class of opera-
tors L∗ L which have a divergent sequence of eigenvalues. Such is the case if
L is the derivative operator, when −L∗ L is the Laplacian operator, or more
generally whenever L is a differential operator for which L is compact, i.e.,
∗
when L L has compact resolvent.
Theorem 4.4. Suppose x ∈ D(L) and L is compact. If Lx − xδ = o(δ 2/3 )
α
for all xδ ∈ H1 satisfying x − xδ ≤ δ, then x ∈ N (L).
has a sequence {λn } of eigenvalues
Proof. The spectral theorem implies that L
∗
satisfying λn → 0 and hence
" L# L has a sequence of eigenvalues µn = 1/λn −
2
1 → ∞. If Lxδα −Lx = o δ 3 for all xδ with x−xδ ≤ δ, then we will show
& 2'
that x ∈ N (L). Indeed, if x ∈
/ N (L), then, as we have just shown, α = o δ 3 .
Let eδα = Lxδα − Lx, then
and by hypothesis
4
Lxα − Lx2 /δ 3 → 0
as δ → 0 (since xδ = x satisfies x − xδ ≤ δ). Therefore, we must have
2
Lxα − Lx, Lxδα − Lxα + Lxδα − Lxα 2
4 → 0 as δ → 0.
δ3
Suppose that {un } are orthonormal eigenvectors of L∗ L associated with {µn }.
associated with the eigenvalues
Then {un } are eigenvectors of L
λn = 1/(1 + µn )
Lxδα − Lxα 2
= δ 2
L(αI −1 un , L∗ LL(αI
+ (1 − α)L) −1 un
+ (1 − α)L)
|
Lxα − Lx, Lxδαn − Lxα | Lxα − Lx Lxδαn − Lxα
4 ≤ 2 2 → 0.
δn
3
δn3 δn3
This contradiction establishes that x ∈ N (L) and hence that the order O(δ 2/3 )
is essentially best possible.
The convergence rates discussed thus far for the Tikhonov-Morozov method
were based on a priori choices for the parameter α. We now investigate an
a posteriori method of choosing α, due to V.A. Morozov, which depends on
monitoring the quantity xδα −xδ . The basic philosophy of the method is that
this “discrepancy” quantity should be of the same order as the error level in
the data. Throughout we make the reasonable assumption that there is more
signal than noise in the data, that is, that the signal-to-noise ratio, xδ /δ, is
greater than one:
x − xδ ≤ δ < xδ . (4.1)
As the discrepancy method requires observation of xδα − xδ , we first note
that
xδα − xδ = L(αI −1 xδ − xδ
+ (1 − α)L)
(4.2)
−1 [α(L
= (αI + (1 − α)L) − I)xδ ].
f (α) = xδα − xδ 2
82 4 The Tikhonov-Morozov Method
Also,
α2 (1 − λ)2
→ 1 as α→∞
(α + (1 − α)λ)2
and therefore,
lim f (α) = xδ 2 > δ 2
α→∞
xδα(δ) − xδ = δ. (4.3)
The a posteriori choice of the parameter α by the criterion (4.3) is called the
choice by the discrepancy method.
Lemma 4.5. If x ∈
/ N (L) and α is chosen by the criterion (4.3), then α(δ) =
O(δ).
Proof. By (4.2)
δ − xδ ) = α(δ)(L
[αI + (1 − α)L](x − I)xδ
α(δ)
and hence
Lxδα(δ) ≤ Lx. (4.4)
4.2 The Discrepancy Criterion 83
To establish this it suffices to show that every sequence {δn } of positive num-
bers converging to zero has a subsequence, which for notational simplicity we
designate {δm }, satisfying
xδα(δ
m
m)
→ x and Lxδα(δ
m
m)
→ Lx.
Lxδα(δ
m
m)
y
for some y ∈ H2 . However, G(L) is weakly closed (being closed and convex)
and hence y = Lx, i.e.,
xδα(δ
m
m)
→ x and Lxδα(δ
m
m)
Lx.
But then,
Lxδα(δ
m
m)
, Lx → Lx2
and hence by the Cauchy-Schwarz inequality and (4.4),
Under appropriate conditions on the true data x, one can prove a conver-
gence rate for the Tikhonov-Morozov method with parameter chosen by the
discrepancy principle.
Theorem 4.7. If x ∈ D(L∗ L), and if α(δ) is chosen by (4.3), then
√
Lxδα(δ) − Lx = O( δ).
84 4 The Tikhonov-Morozov Method
≤ 2
Lx − Lxδα(δ) , Lx
= 2
x − xδα(δ) , L∗ Lx
≤ 4δL∗ Lx.
=
L∗ L[L(αI −1 (u1 + δn un ) − u1 ],
+ (1 − α)L)
L(αI −1 (u1 + δn un ) − u1
+ (1 − α)L)
=
(λ1 µ1 /(α + (1 − α)λ1 ) − µ1 )u1 + (δn λn µn /(α + (1 − α)λn ))un ,
This result shows that the rate in Theorem 4.7 can not be generally
improved, but if we make a more specific requirement on x and modify
the choice criterion (4.3) appropriately, then the best possible rate for the
Tikhonov-Morozov method is achievable by use of a modified discrepancy
principle .
We suggest the following modified discrepancy principle: choose α = α(δ)
to satisfy
xδα(δ) − xδ = δ 2 /α2 . (4.5)
Proof. To simplify the typography we simply write α for α(δ) in this proof.
Since
δα − xδ ) = α(L
[αI + (1 − α)L](x − I)xδ ,
we have
− I)xδ ≤ xδ − xδ = δ 2 /α2
α(L α
− I)xδ → (L
and since (L − I)x > 0 as δ → 0, we find that α → 0 as
δ → 0, and hence xα − x → 0. We then have
≤ δ + xα − x → 0 as δ → 0.
as δ → 0. Also,
L∗ L(xδα − xα ) = (I − L)(αI −1 (xδ − x)
+ (1 − α)L)
86 4 The Tikhonov-Morozov Method
and hence
L∗ L(xδα − xα ) ≤ I − Lδ/α → 0 as δ → 0.
δ2
= L∗ Lxδα → L∗ Lx
α3
as δ → 0.
We can now prove the promised order of approximation result for the
modified discrepancy principle. Recall from Lemma 2.8 that the condition
x ∈ D(LL∗ L) is equivalent to x ∈ D(L) and Lx ∈ R(L).
Proof. By the lemma, α(δ) ∼ δ 2/3 , so from Theorem 3.4 and Corollary 3.8 we
have
Lxδα(δ) − Lx ≤ Lxδα(δ) − Lxα(δ) + Lxα(δ) − Lx
≤ δ/ α(δ) + O(α(δ)) = O(δ 2/3 ).
δ = Lx
(αI + (1 − α)L)x δ + α(I − L)x
δ , xδ0 = 0
n n−1
or equivalently
(I + αL∗ L)xδn = xδ + αL∗ Lxδn−1 . (4.7)
We have seen that the ordinary Tikhonov-Morozov method has a varia-
tional interpretation. In a similar way the iterated Tikhonov-Morozov method
may be viewed as a multi-stage optimization procedure in which each iterate
serves to stabilize the next iterate. Specifically, it is easy to see that xδn is the
unique minimizer of
which, by the same line of argument used in the analysis of the ordinary
Tikhonov-Morozov method, is equivalent to (4.7).
Again it is instructive to interpret the iterated method in the product
Hilbert space H = H1 × H2 with the norm | · | satisfying
One sees immediately that (xδn , Lxδn ) is the point in the graph G(L) ⊆ H
which is nearest (in the norm | · |) to the point (xδ , Lxδn−1 ) ∈ H.
The basic convergence theory of the iterated Tikhonov-Morozov method
was developed in the previous chapter. The basis of that theoretical develop-
ment was von Neumann’s theorem and the spectral theorem. We now present
an alternative convergence proof grounded on an entirely different idea – the
alternating projection theorem. To do so we formulate the iterated method
in the product Hilbert space H. We begin the discussion by assuming that
the data is the error-free vector x ∈ D(L). Let V1 = {x} × H2 ⊆ H and
V2 = G(L) ⊆ H. Note that both of the sets V1 and V2 are closed in H and
affine (V2 is of course a closed subspace of H since L is closed). Also, the
condition x ∈ D(L) is equivalent to V1 ∩ V2 = ∅.
The iterated method with exact data x ∈ D(L) generates a sequence {xn }
with x0 = 0 and
In particular
that is,
(x1 , Lx1 ) = P2 (x, 0) = P2 P1 (0, 0)
where P2 is the projection of H onto V2 = G(L) and P1 is the projection of
H onto the closed affine set V1 = {x} × H2 . Then P1 (x1 , Lx1 ) = (x, Lx1 ).
Furthermore,
x2 = argminz∈D(L) |(z, Lz) − (x, Lx1 )|2
and hence
(x2 , Lx2 ) = P2 (x, Lx1 ) = P2 P1 (x1 , Lx1 ).
Continuing in this manner we see that
xn → x and Lxn → Lx as n → ∞.
then, if x ∈ D(L)
Also,
≤ δ 2 + αLxδn−1 − Lxn−1 2
and therefore
Lxδn − Lxn 2 ≤ nδ 2 /α.
We have now reproduced, from a different perspective, the result of
Corollary 3.11, namely
Lxδn(δ) → Lx if δ n(δ) → 0 as δ → 0.
4.4 The Nonstationary Method 89
(I + αn L∗ L)xn = x + αn L∗ Lxn−1
or equivalently
n = Lx
(αn I + (1 − αn )L)x + αn (I − L)x
n−1 . (4.8)
βj (t) = αj + (1 − αj )t.
Note that βj (t) ∈ [t, 1] for all t ∈ [0, 1]. With this notation
and hence
t αn (1 − t)
tTn (t) = + (tTn−1 (t))
βn (t) βn (t)
t t
= + 1− (tTn−1 (t)).
βn (t) βn (t)
Since t/βn (t) ∈ [t, 1] ⊆ [0, 1] we find that
for all n and all t ∈ [0, 1]. Therefore, {Tn (t)} satisfies condition (3.2).
Also, since
Tn (t) = 1/βn (t) + (1 − t/βn (t))Tn−1 (t)
one finds that
1 t 1
Tn (t) − = 1− Tn−1 (t) −
t βn (t) t
and hence n
Tn (t) − 1 1 t
= 1−
t t j=1 βj (t)
for t ∈ (0, 1]. Since {αj } is bounded, there is a subsequence {αjk } converging
to some number α ∈ [0, 1]. It then happens that
1 t
→ = 0
βjk (t) α + (1 − α)t
as n → ∞, for all t ∈ (0, 1]. Therefore, {Tn (t)} satisfies condition (3.3).
As a consequence of Theorem (3.3) we have:
Corollary 4.10. If {αj } ⊆ (0, 1] and x ∈ D(L), then Lxn → Lx, where {xn }
is the sequence generated by the nonstationary method (4.8).
Suppose now that x−xδ ≤ δ and denote by {xδn } the sequence generated
by nonstationary method with data xδ , that is,
δn = Lx
(αn I + (1 − αn )L)x δ + αn (I − L)x
δn−1
Since
1−t t
(1 − t)Tn (t) = + 1− ((1 − t)Tn−1 (t))
βn (t) βn (t)
and 0 ≤ (1 − t)/βj (t) ≤ 1/αj for t ∈ [0, 1], we have
and hence
n
|(1 − t)Tn (t)| ≤ σn := αj−1 .
j=1
which gives %n
(1 − tTn (t))tν = tν j=1 (1 − t/βj (t))
%n αj
= (1 + s)−ν sν j=1
αj + s
%n αj
≤ sν j=1
αj + s
where s = t/(1 − t) ∈ [0, ∞). We now need some numerical estimates that
result from a close analysis of the functions
n
αj
f (s) = sν , s ∈ [0, ∞)
α
j=1 j
+s
n
1
g(t) := =ν (4.11)
1 + αk t
k=1
and hence the equation (4.11) has a unique positive solution which we will
denote by t1 . Also, since f (0) = 0 = f (∞), we have
n
n−1
− tj ≥ αj−1 = σn−1 . (4.13)
j=2 j=1
But the sum of the roots of a monic polynomial is the negative of the next-
to-highest order coefficient, and hence
n
n
n
1 −1 1−ν
ti = − αi−1 + α = σn .
i=1 i=1
ν k ν
k=1
" #ν
Lemma 4.13. If 0 < ν < 1, then maxs∈[0,∞) f (s) ≤ ν
1−ν σn−ν .
for all n sufficiently large. Note that this is certainly the case for the stationary
method (αj = α), the geometric choice of parameters (αj = q j−1 α, q < 1),
and the harmonic choice (αj = 1/j) and many other parameter sequences.
While this condition may appear ad hoc, it is in fact a necessary condition for
an estimate of the form
when ν > 1.
Theorem 4.14. If ν > 1 and f (s) ≤ cν σn−ν for some cν > 0 and all n, then
{αj } satisfies (4.14) for some c > 0.
Proof. Let
k
αj
pk (s) =
j=1
αj + s
for s ≥ 0. Then, pn−1 (s) > 0, pn−1 (0) = −σn−1 , and pn−1 (0) = 1. Therefore,
And hence,
s
cν σn−ν ≥ sν pn (s) = αn sν−1 pn−1 (s)
αn + s
αn ν
≥ s (1 − σn−1 s), for s ≥ 0.
αn + s
−1
In particular, we find on setting s = 12 σn−1 that
αn σn−1 ν −ν
cν σn−1 ≥ (1/2) σn−1
1 + 2αn σn−1
and therefore,
ν ν
αn σn−1 σn−1 αn σn−1
≤ 2ν cν ν
= 2 cν .
1 + 2αn σn−1 σn 1 + αn σn−1
We then have
ν−1
1 αn σn−1 αn σn−1
< ≤ cν ≤ 2ν cν (αn σn−1 )ν−1
2 1 + 2αn σn−1 1 + αn σn−1
and hence
1
≤ cν1/(ν−1) 2(ν+1)/(ν−1) σn−1 ,
σn
giving the result.
94 4 The Tikhonov-Morozov Method
which is a contradiction.
We now remove the restriction ν < 1 used in Lemma 4.13 by imposing the
condition (4.14).
where cν = (2ν(c + 1))ν for 0 < ν ≤ 1, while cν = (2ν(c + 1)ν )ν for ν > 1.
and the stated result follows from Lemma 4.13. On the other hand, if 1/2 <
ν ≤ 1, then by Lemma 4.12 and inequality (4.13),
n
t1 ≥ − ti ≥ σn−1
i=2
f (t−1 −ν
1 ) ≤ cν σn (4.16)
holds for all ν with 0 < ν ≤ ν0 and some ν0 ≥ 1. Suppose ν ∈ (ν0 , ν0 + 1] and
n > ν. We will show that (4.16) holds for ν. By Lemma 4.12 and inequality
4.4 The Nonstationary Method 95
f (t−1 −ν ν −ν ν −ν −ν
1 ) ≤ t1 ≤ (2ν) σn−1 ≤ (2ν(c + 1)) σn ≤ cν σn .
On the other hand, if αn < 2(ν − 1)σn−1 , then by (4.16) and (4.15) (recall that
n − 1 > ν − 1 by assumption),
αn t1 αi t1 αi t1
n−1 n−1
−(ν−1)
f (t−1 −ν
1 ) = t1 ≤ αn t1
1 + αn t1 i=1 1 + αi t1 i=1
1 + αi t1
−(ν−1)
≤ αn cν−1 σn−1 ≤ 2(ν − 1)(c + 1)ν cν−1 σn−ν .
The geometric choice of parameters αn = α×q n−1 where α > 0 is fixed and
0 < q < 1 is an attractive option for the choice of the sequence of stabilization
parameters. In this case
1 1−n 1 − q n 1
σn = q ≥ q 1−n = q/αn+1
α 1−q α
and hence (4.14) holds for this choice. Also, by Corollary 4.16
and a linear rate of convergence results which is faster the smoother (i.e., the
larger ν) the true data x.
We now consider an a posteriori choice of the iteration index in the non-
stationary Tikhonov-Morozov method. The analysis follows closely that for
the iterative method based on functional interpolation that was treated in
the previous chapter. Again we assume that the signal-to-noise ratio of the
data is strictly bounded above one:
where τ > 1. We take xδ0 = 0 and again use the criterion (3.16), that is we
denote by n(δ) the first value of the iteration index that satisfies
This, along with Theorem 3.3, guarantees that there is a first value n = n(δ)
of the iteration index that satisfies (4.18). Now,
n(δ) (L)(x
xδn(δ) − xn(δ) = LT δ − x)
and hence
n(δ) (L)(x
x − xn(δ) = xδ − xδn(δ) + LT − xδ ).
But since
n
t
|1 − tTn (t)| = 1− ≤1
j=1 βj (t)
we have
x − xn(δ) ≤ τ δ + δ = (τ + 1)δ, (4.19)
just as in Lemma 2.8.6.
µ w for some w ∈ D(L), then
If x = L
µ (I − LT
Lx − Lxn(δ) = L n(δ) (L))Lw
µ zn(δ)
=L
n(δ) (L))Lw
where zn(δ) = (I − LT and hence
zn(δ) ≤ Lw.
The same argument that was used in the proof of Lemma 3.17 of the previous
chapter now gives:
Lemma 4.18. If x = L µ w for some µ > 0 and some w ∈ D(L) and if n(δ)
is chosen according to (4.18), then
Now,
n−1 (L))(x
(xn−1 − xδn−1 ) − (x − xδ ) = −(I − LT − xδ )
n−1 (L))
and I − LT ≤ 1 and hence by (4.18),
µ w, then
However, if x = L
−µ
xn−1 − x = O(σn−1 ).
xn(δ)−1 − x = O(σn−µ )
4.5 Notes
The general theory of the Tikhonov-Morozov method was initiated by V.A.
Morozov and his colleagues in the late sixties. An account of the theory, in a
more general context than that of the present monograph, can be found in [39]
(see also [40]). The proof that the order of convergence O(δ 2/3 ) is essentially
best possible for the Tikhonov-Morozov method and related results, including
the sharp order of convergence for the method when the parameter is chosen
according to the discrepancy criterion, can be found in [22]. The analysis
98 4 The Tikhonov-Morozov Method
xδα − xδ , z − xδα + α
Lxδα , L(z − xδα ) ≥ 0 (4.22)
is the inner product on H1 × H2 that generates the norm | · |. But one sees
easily that (4.23) is equivalent to (4.22). That is, the minimizer xδα of (4.21),
exists since D(L) is weakly closed, and is unique since D(L) is convex, and is
characterized by the variational inequality (4.22).
If we denote by xα the minimizer of Φα (z; x), where x ∈ D(L), then we
obtain as above
xα − x, z − xα + α
Lxα , L(z − xα ) ≥ 0 (4.24)
4.5 Notes 99
for all z ∈ D(L). Setting z = xα in (4.22), and z = xδα in (4.24), and adding,
one obtains
From this one obtains the key inequality given in the previous case in Corollary
3.7. Indeed, from (4.25) it follows that
xδα − xα ≤ δ
and hence √
L(xα − xδα ) ≤ δ/ α.
By the variational property of Φα (·; x), we have
Φα (xα ; x) ≤ Φα (x; x)
and hence
xα − x2 + αLxα 2 ≤ αLx2 .
In particular, xα → x as α → 0 and Lxα ≤ Lx. Since the graph of L is
closed and convex, and therefore weakly closed, one concludes in the standard
way that xα → x in the graph norm and hence Lxα → Lx as α → 0. One
then has √
Lx − Lxδα ≤ Lx − Lxα + δ/ α
√
and hence, if α = α(δ) satisfies α/ α → 0 as δ → 0, then Lxδα → Lx.
5
Finite-Dimensional Approximations
V1 ⊆ V2 ⊆ . . . ⊆ D(L) ∩ N (L)⊥
and
H1 ⊆ ∪∞
j=1 Vj .
to perturbations in x since the operator LPn has finite rank and hence is a
bounded linear operator.
As a minimal requirement we assume that the projection method is con-
vergent, that is we assume that
LPn L† ≤ C. (5.3)
yn − y = LPn x − Lx
Therefore, by (5.4)
yn(δ)
δ
− y ≤ LPn(δ) δ + (C + 1)(I − Qn(δ) )Lx → 0 (5.6)
as δ → 0.
γn ≥ I − Pn
xδ − xδn ≤ τ γn δ
Mn z → 0 as n→∞
and therefore
γn−1 xδ − xδn ≤ δ + Mn x ≤ τ δ
for n sufficiently large.
104 5 Finite-Dimensional Approximations
(note that this choice depends on the particular value selected for τ and on the
upper bound γn accepted for I − Pn ). The convergence of the projection
method with this discrepancy principle will depend on the approximating
power of the subspaces and the size of the stability bound LPn . We shall
assume that
LPn dist(x, Vn−1 ) → 0 as n → ∞ (5.8)
where
dist(x, Vn−1 ) = inf{x − v : v ∈ Vn−1 }.
Theorem 5.3. Let τ > 1 and suppose n = n(δ) is the first value of n satisfy-
ing (5.7). If (5.8) holds, then Lxδn(δ) → Lx as δ → 0.
Proof. In the interest of shortening the notation we shall write n for n(δ). For
any v ∈ Vn−1 we find that
and therefore
1
δ≤ dist(x, Vn−1 ). (5.9)
τ −1
From (5.6) we then have
δ δ
x − xn0 = limj xδj − xnj0 = limj xδj − xn(δ
j
j)
*
∞ −1/2 −1/2
= j=n λ−1
j
Lx, λj Lvj λj Lvj ,
∞ +
−1 −1/2 −1/2
k=n λk
Lx, λk Lvk λk Lvk
≤ λ−1
n dist(Lx, L(Vn−1 ))
2
and hence
LPn dist(x, Vn−1 ) ≤ dist(Lx, L(Vn−1 )) → 0
as n → ∞, which verifies the hypothesis (5.8) and shows the applicability of
the discrepancy principle (5.7) in this case.
For this particular example we may also give a bound for the error when
the discrepancy method is used. The following theorem shows that for this
example orders of approximation that are arbitrarily near to the optimal order
may be achieved in principle. Before proving the theorem we establish a basic
inequality that we will need.
106 5 Finite-Dimensional Approximations
Lemma 5.4. If L∗ L has compact resolvent and z ∈ D((L∗ L)ν+1 ) for some
ν > 0, then
L∗ Lz ≤ z ν+1 (L∗ L)ν+1 z ν+1 .
ν 1
where 2ν 2
ak = |
z, vk | ν+1 and bk = λ2k |
z, vk | ν+1 .
Using p = (ν + 1)/ν and q = ν + 1, we see that {ak } ∈ lp and {bk } ∈ lq and
hence by Hölder’s inequality
∞ 1 ∞ q q1
L∗ Lz2 ≤ ( k=1 apk ) p ( k=1 bk )
Proof. Throughout this proof we use n = n(δ) where n(δ) is given by (5.7).
Also, since the projectors are orthogonal, we have γn = 1. Let z = (L∗ L)ν+1 x,
then $∞ $
$ $
$ −ν−1 $
dist(x, Vn−1 ) = $ λk
z, vk vk $ ≤ λ−ν−1 z
$ $ n
k=n
and
dist(x, Vn−1 ) = O(λ−ν−1
n ) = O(δ).
Also,
dist(Lx, L(Vn ))2 ≤ L(x − xn )2
=
L∗ L(x − xn ), x − xn
and hence
But,
(L∗ L)ν+1 (x − xn ) = (I − Pn )z ≤ z
and hence
dist(Lx, L(Vn )) ≤ O(x − xn (2ν+1)/(2ν+2) ).
However, by (5.7)
x − xn = (I − Pn )x
= (I − Pn )(x − xδ ) + xδ − Pn xδ
≤ δ + τ δ = (1 + τ )δ,
and hence
=
x, L∗ li
(m) (m) (m)
ym , li =
Lx, li , i = 1, . . . , n(m). (5.14)
and
n(m)
(m)
Im c = cj lj ,
j=1
respectively, and let Gm : Rn(m) → Rn(m) be the linear operator whose matrix
representation relative to the standard basis is the Gramian matrix
n(m) n(m)
[Gm ]ij = [
li , lj ].
If
n(m)
(m)
ym = cj lj = Im c,
j=1
Lx − Lm(δ) xδ → 0
n(m)
(m)
Lm x = cj lj
j=1
where Gm c = Bm x we have
(m) (m)
Lm x2 =
Im c, Im c = i,j ci cj
li , lj = cT Gm c
= (G−1 T −1
m Bm x) Bm x ≤ Gm Bm x
2 2
and hence 2
Lm ≤ G−1
m Bm (5.16)
where
γm := (I − Pm )L∗† → 0 as m → ∞.
The model inverse problem of reconstructing a spatially distributed source
term g(x) given a measured version of the temperature distribution f (x) =
u(x, 1) at a later time in the heat problem
∂u ∂2u
= + g(x), 0 < x < π, 0<t
∂t ∂x2
u(0, t) = u(π, t) = u(x, 0) = 0,
provides an illustration of the result of this section. Formal separation of
variable techniques lead to the representation
∞ π
2 n2
g(x) = (Lf )(x) = f (s) sin nsds sin nx.
π n=1 1 − e−n2 0
Let lk (s) = sin ks, then the orthogonal projector of L2 [0, π] onto
Vm = span{l1 , . . . , lm }
is
∞
2
Pm φ =
φ, lj lj .
π n=1
Also, Gm = π
2 I, where I is the identity operator on Rm and hence λm = π/2.
k2
Since L∗ lk = lk , we have
1 − e−k2
m 2
k2
Bm g2 = |
f, lk |2 ≤ O(m4 )f 2
1 − e−k2
k=1
2 n2
Lf − Pm Lf =
f, ln ln ,
π n>m 1 − e−n2
5.2 Finite Elements 111
Φα (z; xδ ) = z − xδ 2 + αLz2
(see Theorem 4.1) suggests the possibility of using finite element methods to
effectively compute the approximations. To this end, suppose {Vm }∞ m=1 is a
sequence of finite-dimensional subspaces of H1 satisfying
V1 ⊆ V2 ⊆ · · · ⊆ D(L) and ∪∞
m=1 Vm = H1 .
n(m)
(m) (m)
xα,m = cj ϕj
j=1
n(m)
(m) (m) (m) (m) (m) (m)
[
ϕi , ϕj + α
Lϕi , Lϕj ]cj =
x, ϕi , i = 1, . . . , n(m).
j=1
112 5 Finite-Dimensional Approximations
Lxα,m − Lxδα,m .
The stability bound turns out to be the same as that found for the approxi-
mation in infinite dimensional space (see Corollary 3.7). We will employ the
inner product [·, ·] defined on D(L) by
[u, w] =
u, w + α
Lu, Lw
where α is a fixed positive number, and the associated norm |u| = [u, u].
Note that, since L is closed, D(L) is a Hilbert space when endowed with this
inner product.
Lemma 5.7. If x ∈ D(L) ⊆ H1 and xδ ∈ H1 satisfies x − xδ ≤ δ, then
√
Lxα,m − Lxδα,m ≤ δ/ α.
xα,m − x, v + α
Lxα,m , Lv = 0, for all v ∈ Vm (5.17)
and similarly
xδα,m − x, v + α
Lxδα,m , Lv = 0, for all v ∈ Vm .
The condition (5.17) may be expressed in terms of the inner product [·, ·] in
the following way:
[xα,m − x, v] =
xα,m − x, v + α
L(xα,m − x), Lv = −α
Lx, Lv (5.18)
[xδα,m − x, v] =
xδα,m − x, v + α
Lxδα,m − Lx, Lv
=
xδ − x, v +
xδα,m − xδ , v + α
L(xδα,m − x), Lv
=
xδ − x, v − α
Lx, Lv
5.2 Finite Elements 113
and therefore,
[xδα,m − xα,m , v] =
xδ − x, v (5.19)
for all v ∈ Vm . In particular, setting v = xδα,m − xα,m in (5.19), and applying
the Cauchy-Schwarz inequality, one obtains
=
xδ − x, xδα,m − xα,m
≤ δxδα,m − xα,m
Therefore,
xδα,m − xα,m ≤ δ
and hence
αLxδα,m − Lxα,m 2 ≤ δ 2
giving the result.
√
We see from this lemma that the condition δ/ α → 0, combined with a
condition that ensures that Lxα,m − Lx → 0, will together guarantee the
convergence of the stabilized finite element approximations to Lx.
The remaining development requires an analysis of the difference between
xα,m and the infinite dimensional smoothed approximation xα using exact
data x ∈ D(L), which is characterized by
3 4
xα = argminz∈D(L) z − x2 + αLz2 .
This is equivalent to
0 =
xα − x, v + α
Lxα , Lv = [xα − x, v] + α
Lx, v
for all v ∈ D(L). The corresponding finite element approximation xα,m satis-
fies (5.18), that is
0 = [xα,m − x, v] + α
Lx, v
for all v ∈ Vm . Subtracting, we find that
We can express this in a geometrical way by saying that xα,m is the [·, ·] -
orthogonal projection of xα onto the finite-dimensional subspace Vm . That is,
xα,m = Pm xα (5.21)
βm = (I − Pm )L
and
γm = L(I − Pm )L.
LPm and L
Note that since LL, are all bounded linear operators, both of these
quantities is finite. We begin with a result that requires relatively modest
assumptions on the true data x.
Theorem 5.8. If x ∈ D(L∗ L), then
where w =
Proof. First we note that x ∈ D(L∗ L) if and only if x = Lw
x + L∗ Lx. From the characterization (5.21) we have
= |xα − Pm xα |2 ≤ |xα − Pm xα |2
But
xα = (I + αL∗ L)−1 x = L(αI −1 x
+ (1 − α)L)
= L(αI −1 Lw
+ (1 − α)L)
−1 L
Also, (αI + (1 − α)L) ≤ 1. Therefore,
2 + αL(I − Pm )L
αLxα − Lxα,m 2 ≤ ((I − Pm )L 2 )w2
that is,
Lxα − Lxα,m 2 ≤ (βm
2 2
/α + γm )w2 .
Since B is closed and convex, it is weakly closed and also bounded. Therefore,
by Theorem 2.1 there is a subsequence {xnk } with xnk → x ∈ B. As K is
compact, one then has Kxnk → Kx. By the uniform boundedness principle
there is a constant C with An ≤ C for all n. We then have
as k → ∞, which is a contradiction.
βn → 0 as n → ∞.
If we assume that LPn x → Lx for all x ∈ D(L∗ L), that is, we assume condition
(5.1), then we can also apply the lemma to the operator An = L(I − Pn ) to
find that
γn → 0 as n → ∞.
We may now give a basic convergence result for the stabilized finite element
approximations.
Theorem 5.10. Suppose L is compact, condition (5.1) holds, and x ∈ D(L∗ L).
α = α(δ) → 0 as δ → 0 and m = m(α) → ∞ as α → 0 in such a way that
If √
δ/ α → 0 and βm 2
/α → 0, then Lxδα,m → Lx.
Remark. If we are willing to assume more on the true data, namely that
x ∈ R(L ν ) for some ν ≥ 1, then minor modifications of the argument above
gives the bound
√
Lxδα,m − Lx ≤ δ/ α + O( βm (ν)2 /α + γm (ν)2 ) + Lxα − Lx (5.23)
116 5 Finite-Dimensional Approximations
where
ν
βm (ν) = (I − Pm )L
and
ν .
γm (ν) = L(I − Pm )L
From Corollary 3.8 we known that if x ∈ D(LL∗ L), then Lxα − Lx = O(α).
We therefore obtain the following corollary:
Corollary 5.11. If, in addition to the hypotheses of Theorem 5.10, x ∈
D(LL∗ L), γm = O(α), βm = O(α3/2 ) and α ∼ δ 2/3 , then
Vm ⊆ Vm+1 and ∪∞
m=1 Vm = H1 .
Φα (z; xδ ) = z − xδ 2 + αLz2 ,
z − xδ 2 = z − Pm xδ 2 + (I − Pm )xδ 2
and hence
xδm,α = argminz∈Vm z − Pm xδ 2 + αLm z2 .
Therefore, as previously seen, we have
n(m) 2
αλj
xδm,α − Pm x =
δ 2
|
Pm xδ , φj |2 . (5.25)
j=1
1 + αλj
xδm,α − Pm xδ = τ δ.
Proof. From (5.25) we see that for such fixed m = m(δ) the function
d(α) = xδm,α − Pm xδ
We call the choice of α according to Theorem 5.12 the choice by the finite-
dimensional discrepancy principle.
Proof. Since
xδm,α = argminz∈Vm Φα (z; Pm xδ )
we have
Φα (xδm,α ; Pm xδ ) ≤ Φα (Pm x; Pm xδ )
118 5 Finite-Dimensional Approximations
and hence
But then,
τ 2 δ 2 + αLxδm,α 2 ≤ δ 2 + αLPm x2
giving
Lxδm,α ≤ LPm x
since τ > 1.
for all x ∈ D(L∗ L). For a given δ > 0 it then follows that there is an Mδ such
that
1. LPm x2 ≤ δ + Lx2
2. (I − Pm )xδ ≤ δ
3. τ δ ≤ Pm xδ
for all m ≥ Mδ . Furthermore, for each such m = m(δ), there is a unique
α = α(m, δ) satisfying the criterion of Theorem 5.12.
Having set the stage, we may now prove the convergence of the approxi-
mations provided by the finite dimensional discrepancy principle.
Theorem 5.14. Suppose that x ∈ D(L), and that m = m(δ) is large enough
so that (1)-(3) hold. Let α = α(m, δ) be the stabilization parameter determined
by Theorem 5.12. Then
Proof. To simplify notation we write xδm,α for xδm(δ),α(m(δ),δ) . First note that
≤ (τ + 2)δ
≤ LPm x2 − 2
Lxδm,α , Lx + Lx2
≤ δ + Lx2 − 2
Lx − Lxδm,α , Lx + Lx2
= δ + 2
Lx − Lxδm,α , Lx
= δ + 2
x − xδm,α , L∗ Lx
=
x − xδ + xδ − Pm xδ + Pm xδ − xδm,α , L∗ Lx
+2
Pm xδ − xδm,α , L∗ Lx
5.3 Notes
The material on direct projection in Section 5.1 follows [19]. The choice of
subspaces in D(L∗ ) is closely related to the method of dual least squares
(see, e.g., Engl, Hanke and Neubauer [8]) for solution of first kind operator
equations by projection.
References
15. Groetsch, C.W. (1992): Spectral methods for linear inverse problems with
unbounded operators. Journal of Approximation Theory, 70, 16-28
16. Groetsch, C.W. (1993): Inverse Problems in the Mathematical Sciences. Vieweg,
Braunschweig
17. Groetsch, C.W. (1995): Inclusions and identities for the Moore-Penrose inverse
of a closed linear operator, Mathematische Nachrichten, 171, 157-164
18. Groetsch, C.W. (2006): An iterative stabilization method for the evaluation of
unbounded operators, Proceedings of the American Mathematical Society 134,
1173-1181
19. Groetsch, C.W., Hanke, M. (1995): Regularization by projection of unbounded
operators arising in inverse problems, Inverse Problems and Applications in Geo-
physics, Industry, Medicine and Technology (D.D. Ang et al. Eds.), pp. 61-70,
Publications of the HoChiMinh City Mathematical Society, Vol. 2, HoChiMinh
City, Viet Nam
20. Groetsch, C.W., Hanke, M. (1995): Regularization of some one-dimensional in-
verse problems for identification of nonlinear surface phenomena. Proceedings
of the 1995 Design Engineering Technical Conferences, Vol. 3C, pp. 903-908,
American Society of Mechanical Engineers, New York
21. Groetsch, C.W., Hanke, M. (1996): A general framework for regularized evalua-
tion of unstable operators, Journal of Mathematical Analysis and Applications,
203, 451-463
22. Groetsch, C.W., Scherzer, O. (1993): The optimal order of convergence for stable
evaluation of differential operators. Electronic Journal of Differential Equations,
3, 1-12
23. Groetsch, C.W., Scherzer, O. (2000): Nonstationary iterated Tikhonov-Morozov
method and third order differential equations for the evaluation of unbounded
operators, Mathematical Methods in the Applied Sciences, 23, 1287-3000
24. Groetsch, C.W., Scherzer, O. (2002): Iterative stabilization and edge detection.
Contemporary Mathematics, 313, 129-142
25. Guenther, R.B., Lee, J.W. (1988): Partial Differential Equations of Mathemat-
ical Physics and Integral Equations. Prentice-Hall, Englewood Cliffs, N.J.
26. Halmos, P. (1967): A Hilbert Space Problem Book. Van Nostrand, Princeton
27. Hanke, M., Groetsch, C.W. (1998): Nonstationary iterated Tikhonov regular-
ization, Journal of Optimization Theory and Applications, 98, 37-53
28. Hausdorff, F. (1932): Zur Theorie der linearen metrischen Räume, Journal für
die Reine und angewandte Mathematik, 167, 294-311
29. Hennefeld, J. (1980): A nontopological proof of the uniform boundedness theo-
rem, American Mathematical Monthly, 87, 217
30. Kato, T. (1966): Perturbation Theory for Linear Operators. Springer-Verlag,
Berlin
31. Keller, J.B., Olmstead, W.E. (1972): Temperature of a nonlinearly radiating
semi-infinite body. Quarterly Journal of Applied Mathematics, 30, 559-566
32. Kirsch, A. (1996): An Introduction to the Mathematical Theory of Inverse Prob-
lems. Springer-Verlag, New York
33. Kress, R. (1989): Linear Integral Equations. Springer-Verlag, Berlin
34. Lardy, L.J. (1975): A series representation of the generalized inverse of a closed
linear operator. Atti Accad. Naz. Lincei (Ser. VIII), 58, 152-157.
35. Lax, P.D. (2002): Functional Analysis. Wiley, New York
36. Lebedev, L.P., Vorovich, I.I., Gladwell, G.M.L. (1996): Functional Analysis:
Applications in Mechanics and Inverse Problems, Kluwer, Dordrecht
References 123
37. Louis, A. (1989): Inverse und schlecht gestellte Probleme. Teubner, Stuttgart
38. Mann, W.R., Wolf, F. (1951): Heat transfer between solids and gases under
nonlinear boundary conditions, Quarterly Journal of Applied Mathematics, 9,
163-184
39. Morozov, V.A. (1984): Methods for Solving Incorrectly Posed Problems,
Springer-Verlag, New York
40. Morozov, V.A. (1999): Some aspects of restoration of signals by a regularization
method. Recent Advances in Numerical Methods and Applications II, O.P. Iliev
et al., (Eds.). World Scientific, Singapore
41. Neubauer, A. (1985): Tikhonov-regularization of ill-posed linear operator equa-
tions on closed convex sets, dissertation, Linz
42. Popov, D.A., Sushko, D.V. (1994): Computation of singular convolutions,
pp. 43-127 in Applied Problems of Radon Transform. S. Gindikin, Ed., American
Mathematical Society Translations, Series 2, Vol. 162, Providence
43. Pöschel, J., Trubowitz, E. (1987): Inverse Spectral Theory, Academic Press,
Orlando
44. Riesz, F., Sz.Nagy, B. (1955): Functional Analysis, Ungar, New York
45. Sylvester, J., Uhlmann, G. (1990): The Dirichlet to Neumann map and ap-
plications, in Inverse Problems in Partial Differential Equations (D. Colton,
R. Ewing, W. Rundell, Eds.), 101-139. SIAM, Philadelphia
46. Youla, D., (1987): Mathematical theory of image restoration by the method of
convex projection, in Image Recovery: Theory and Application, H. Stark, Ed.,
Academic Press, New York
47. von Neumann, J. (1932): Über adjungierte Funktionaloperatoren. Annals of
Mathematics, 33, 294-310
48. von Neumann, J. (1950): Functional Operators - Vol. II. The Geometry of
Orthogonal Spaces, Annals of Mathematics Studies #22. Princeton University
Press, Princeton (This is a reprint of memeographed lecture notes first distrib-
uted in 1933.)
49. Zeidler, E. (1990): Nonlinear Functional Analysis and Its Applications. Springer-
Verlag, Berlin
50. Zeidler, E. (1995): Applied Functional Analysis: Main Principles and Their
Applications. Springer-Verlag, Berlin
Index
eigenspace, 26
Banach’s theorem, 25
eigenvalue, 26
band-limited function, 6
elliptic boundary value problem, 22
Bessel’s inequality, 20
Engl, H., 119
best approximations, 47
error-free data, 56
best possible rate, 85
boundary condition, nonlinear, 14 finite dimensional discrepancy principle,
bounded linear operator, 2 118
Bustoz, Jr., J., VIII finite element methods, 111
finite rank operator, 25
Cannon, J., 17 Fourier coefficients, 8
Cauchy problem, 6 Fourier expansion, 20
Cauchy-Schwarz inequality, 20 Fourier projection, 21
characteristic function, 7 Fourier transform, 3, 6
Closed Graph Theorem, 35 frequencies of vibration, 3
closed linear operator, 32 functional interpolation, 73
closed quadrature rules, 72 fundamental subspaces, 25
compact operator, 26
compact resolvent, 41, 79 geometric choice of parameters, 93
complete orthonormal eigensystem, 61 Gramian matrix, 108
complete orthonormal set, 20 graph inner product, 35
complex Fourier expansion, 7 graph norm, 35
126 Index
Vol. 1701: Ti-Jun Xiao, J. Liang, The Cauchy Problem of Vol. 1723: J. P. Croisille, C. Lebeau, Diffraction by an Im-
Higher Order Abstract Differential Equations (1998) mersed Elastic Wedge (1999)
Vol. 1702: J. Ma, J. Yong, Forward-Backward Stochastic Vol. 1724: V. N. Kolokoltsov, Semiclassical Analysis for
Differential Equations and Their Applications (1999) Diffusions and Stochastic Processes (2000)
Vol. 1703: R. M. Dudley, R. Norvaiša, Differentiabil- Vol. 1725: D. A. Wolf-Gladrow, Lattice-Gas Cellular Au-
ity of Six Operators on Nonsmooth Functions and p- tomata and Lattice Boltzmann Models (2000)
Variation (1999) Vol. 1726: V. Marić, Regular Variation and Differential
Vol. 1704: H. Tamanoi, Elliptic Genera and Vertex Oper- Equations (2000)
ator Super-Algebras (1999) Vol. 1727: P. Kravanja M. Van Barel, Computing the Zeros
Vol. 1705: I. Nikolaev, E. Zhuzhoma, Flows in 2-dimen- of Analytic Functions (2000)
sional Manifolds (1999) Vol. 1728: K. Gatermann Computer Algebra Methods for
Vol. 1706: S. Yu. Pilyugin, Shadowing in Dynamical Sys- Equivariant Dynamical Systems (2000)
tems (1999) Vol. 1729: J. Azéma, M. Émery, M. Ledoux, M. Yor (Eds.)
Vol. 1707: R. Pytlak, Numerical Methods for Optimal Séminaire de Probabilités XXXIV (2000)
Control Problems with State Constraints (1999) Vol. 1730: S. Graf, H. Luschgy, Foundations of Quantiza-
Vol. 1708: K. Zuo, Representations of Fundamental tion for Probability Distributions (2000)
Groups of Algebraic Varieties (1999) Vol. 1731: T. Hsu, Quilts: Central Extensions, Braid Ac-
Vol. 1709: J. Azéma, M. Émery, M. Ledoux, M. Yor tions, and Finite Groups (2000)
(Eds.), Séminaire de Probabilités XXXIII (1999) Vol. 1732: K. Keller, Invariant Factors, Julia Equivalences
Vol. 1710: M. Koecher, The Minnesota Notes on Jordan and the (Abstract) Mandelbrot Set (2000)
Algebras and Their Applications (1999)
Vol. 1733: K. Ritter, Average-Case Analysis of Numerical
Vol. 1711: W. Ricker, Operator Algebras Generated by Problems (2000)
Commuting Projećtions: A Vector Measure Approach
Vol. 1734: M. Espedal, A. Fasano, A. Mikelić, Filtration in
(1999)
Porous Media and Industrial Applications. Cetraro 1998.
Vol. 1712: N. Schwartz, J. J. Madden, Semi-algebraic
Editor: A. Fasano. 2000.
Function Rings and Reflectors of Partially Ordered
Vol. 1735: D. Yafaev, Scattering Theory: Some Old and
Rings (1999)
New Problems (2000)
Vol. 1713: F. Bethuel, G. Huisken, S. Müller, K. Stef-
Vol. 1736: B. O. Turesson, Nonlinear Potential Theory and
fen, Calculus of Variations and Geometric Evolution Prob-
Weighted Sobolev Spaces (2000)
lems. Cetraro, 1996. Editors: S. Hildebrandt, M. Struwe
(1999) Vol. 1737: S. Wakabayashi, Classical Microlocal Analysis
Vol. 1714: O. Diekmann, R. Durrett, K. P. Hadeler, P. in the Space of Hyperfunctions (2000)
K. Maini, H. L. Smith, Mathematics Inspired by Biology. Vol. 1738: M. Émery, A. Nemirovski, D. Voiculescu, Lec-
Martina Franca, 1997. Editors: V. Capasso, O. Diekmann tures on Probability Theory and Statistics (2000)
(1999) Vol. 1739: R. Burkard, P. Deuflhard, A. Jameson, J.-L. Li-
Vol. 1715: N. V. Krylov, M. Röckner, J. Zabczyk, Stochas- ons, G. Strang, Computational Mathematics Driven by In-
tic PDE’s and Kolmogorov Equations in Infinite Dimen- dustrial Problems. Martina Franca, 1999. Editors: V. Ca-
sions. Cetraro, 1998. Editor: G. Da Prato (1999) passo, H. Engl, J. Periaux (2000)
Vol. 1716: J. Coates, R. Greenberg, K. A. Ribet, K. Ru- Vol. 1740: B. Kawohl, O. Pironneau, L. Tartar, J.-P. Zole-
bin, Arithmetic Theory of Elliptic Curves. Cetraro, 1997. sio, Optimal Shape Design. Tróia, Portugal 1999. Editors:
Editor: C. Viola (1999) A. Cellina, A. Ornelas (2000)
Vol. 1717: J. Bertoin, F. Martinelli, Y. Peres, Lectures on Vol. 1741: E. Lombardi, Oscillatory Integrals and Phe-
Probability Theory and Statistics. Saint-Flour, 1997. Edi- nomena Beyond all Algebraic Orders (2000)
tor: P. Bernard (1999) Vol. 1742: A. Unterberger, Quantization and Non-
Vol. 1718: A. Eberle, Uniqueness and Non-Uniqueness holomorphic Modular Forms (2000)
of Semigroups Generated by Singular Diffusion Opera- Vol. 1743: L. Habermann, Riemannian Metrics of Con-
tors (1999) stant Mass and Moduli Spaces of Conformal Structures
Vol. 1719: K. R. Meyer, Periodic Solutions of the N-Body (2000)
Problem (1999) Vol. 1744: M. Kunze, Non-Smooth Dynamical Systems
Vol. 1720: D. Elworthy, Y. Le Jan, X-M. Li, On the Geom- (2000)
etry of Diffusion Operators and Stochastic Flows (1999) Vol. 1745: V. D. Milman, G. Schechtman (Eds.), Geomet-
Vol. 1721: A. Iarrobino, V. Kanev, Power Sums, Goren- ric Aspects of Functional Analysis. Israel Seminar 1999-
stein Algebras, and Determinantal Loci (1999) 2000 (2000)
Vol. 1722: R. McCutcheon, Elemental Methods in Ergodic Vol. 1746: A. Degtyarev, I. Itenberg, V. Kharlamov, Real
Ramsey Theory (1999) Enriques Surfaces (2000)
Vol. 1747: L. W. Christensen, Gorenstein Dimensions Vol. 1777: E. García-Río, D. N. Kupeli, R. Vázquez-
(2000) Lorenzo, Osserman Manifolds in Semi-Riemannian
Vol. 1748: M. Ruzicka, Electrorheological Fluids: Model- Geometry (2002)
ing and Mathematical Theory (2001) Vol. 1778: H. Kiechle, Theory of K-Loops (2002)
Vol. 1749: M. Fuchs, G. Seregin, Variational Methods Vol. 1779: I. Chueshov, Monotone Random Systems
for Problems from Plasticity Theory and for Generalized (2002)
Newtonian Fluids (2001) Vol. 1780: J. H. Bruinier, Borcherds Products on O(2,1)
Vol. 1750: B. Conrad, Grothendieck Duality and Base and Chern Classes of Heegner Divisors (2002)
Change (2001) Vol. 1781: E. Bolthausen, E. Perkins, A. van der Vaart,
Vol. 1751: N. J. Cutland, Loeb Measures in Practice: Re- Lectures on Probability Theory and Statistics. Ecole d’
cent Advances (2001) Eté de Probabilités de Saint-Flour XXIX-1999. Editor: P.
Vol. 1752: Y. V. Nesterenko, P. Philippon, Introduction to Bernard (2002)
Algebraic Independence Theory (2001) Vol. 1782: C.-H. Chu, A. T.-M. Lau, Harmonic Functions
Vol. 1753: A. I. Bobenko, U. Eitner, Painlevé Equations in on Groups and Fourier Algebras (2002)
the Differential Geometry of Surfaces (2001) Vol. 1783: L. Grüne, Asymptotic Behavior of Dynamical
Vol. 1754: W. Bertram, The Geometry of Jordan and Lie and Control Systems under Perturbation and Discretiza-
Structures (2001) tion (2002)
Vol. 1755: J. Azéma, M. Émery, M. Ledoux, M. Yor Vol. 1784: L.H. Eliasson, S. B. Kuksin, S. Marmi, J.-C.
(Eds.), Séminaire de Probabilités XXXV (2001) Yoccoz, Dynamical Systems and Small Divisors. Cetraro,
Vol. 1756: P. E. Zhidkov, Korteweg de Vries and Nonlin- Italy 1998. Editors: S. Marmi, J.-C. Yoccoz (2002)
ear Schrödinger Equations: Qualitative Theory (2001) Vol. 1785: J. Arias de Reyna, Pointwise Convergence of
Vol. 1757: R. R. Phelps, Lectures on Choquet’s Theorem Fourier Series (2002)
(2001) Vol. 1786: S. D. Cutkosky, Monomialization of Mor-
Vol. 1758: N. Monod, Continuous Bounded Cohomology phisms from 3-Folds to Surfaces (2002)
of Locally Compact Groups (2001) Vol. 1787: S. Caenepeel, G. Militaru, S. Zhu, Frobenius
Vol. 1759: Y. Abe, K. Kopfermann, Toroidal Groups and Separable Functors for Generalized Module Cate-
(2001) gories and Nonlinear Equations (2002)
Vol. 1760: D. Filipović, Consistency Problems for Heath- Vol. 1788: A. Vasil’ev, Moduli of Families of Curves for
Jarrow-Morton Interest Rate Models (2001) Conformal and Quasiconformal Mappings (2002)
Vol. 1761: C. Adelmann, The Decomposition of Primes in Vol. 1789: Y. Sommerhäuser, Yetter-Drinfel’d Hopf alge-
Torsion Point Fields (2001) bras over groups of prime order (2002)
Vol. 1762: S. Cerrai, Second Order PDE’s in Finite and Vol. 1790: X. Zhan, Matrix Inequalities (2002)
Infinite Dimension (2001) Vol. 1791: M. Knebusch, D. Zhang, Manis Valuations and
Vol. 1763: J.-L. Loday, A. Frabetti, F. Chapoton, F. Goi- Prüfer Extensions I: A new Chapter in Commutative Al-
chot, Dialgebras and Related Operads (2001) gebra (2002)
Vol. 1764: A. Cannas da Silva, Lectures on Symplectic Vol. 1792: D. D. Ang, R. Gorenflo, V. K. Le, D. D. Trong,
Geometry (2001) Moment Theory and Some Inverse Problems in Potential
Vol. 1765: T. Kerler, V. V. Lyubashenko, Non-Semisimple Theory and Heat Conduction (2002)
Topological Quantum Field Theories for 3-Manifolds with Vol. 1793: J. Cortés Monforte, Geometric, Control and
Corners (2001) Numerical Aspects of Nonholonomic Systems (2002)
Vol. 1766: H. Hennion, L. Hervé, Limit Theorems for Vol. 1794: N. Pytheas Fogg, Substitution in Dynamics,
Markov Chains and Stochastic Properties of Dynamical Arithmetics and Combinatorics. Editors: V. Berthé, S. Fer-
Systems by Quasi-Compactness (2001) enczi, C. Mauduit, A. Siegel (2002)
Vol. 1767: J. Xiao, Holomorphic Q Classes (2001) Vol. 1795: H. Li, Filtered-Graded Transfer in Using Non-
Vol. 1768: M.J. Pflaum, Analytic and Geometric Study of commutative Gröbner Bases (2002)
Stratified Spaces (2001) Vol. 1796: J.M. Melenk, hp-Finite Element Methods for
Vol. 1769: M. Alberich-Carramiñana, Geometry of the Singular Perturbations (2002)
Plane Cremona Maps (2002) Vol. 1797: B. Schmidt, Characters and Cyclotomic Fields
Vol. 1770: H. Gluesing-Luerssen, Linear Delay- in Finite Geometry (2002)
Differential Systems with Commensurate Delays: An Vol. 1798: W.M. Oliva, Geometric Mechanics (2002)
Algebraic Approach (2002) Vol. 1799: H. Pajot, Analytic Capacity, Rectifiability,
Vol. 1771: M. Émery, M. Yor (Eds.), Séminaire de Prob- Menger Curvature and the Cauchy Integral (2002)
abilités 1967-1980. A Selection in Martingale Theory Vol. 1800: O. Gabber, L. Ramero, Almost Ring Theory
(2002) (2003)
Vol. 1772: F. Burstall, D. Ferus, K. Leschke, F. Pedit, U. Vol. 1801: J. Azéma, M. Émery, M. Ledoux, M. Yor
Pinkall, Conformal Geometry of Surfaces in S4 (2002) (Eds.), Séminaire de Probabilités XXXVI (2003)
Vol. 1773: Z. Arad, M. Muzychuk, Standard Integral Ta- Vol. 1802: V. Capasso, E. Merzbach, B.G. Ivanoff, M.
ble Algebras Generated by a Non-real Element of Small Dozzi, R. Dalang, T. Mountford, Topics in Spatial Sto-
Degree (2002) chastic Processes. Martina Franca, Italy 2001. Editor: E.
Vol. 1774: V. Runde, Lectures on Amenability (2002) Merzbach (2003)
Vol. 1775: W. H. Meeks, A. Ros, H. Rosenberg, The Vol. 1803: G. Dolzmann, Variational Methods for Crys-
Global Theory of Minimal Surfaces in Flat Spaces. Mar- talline Microstructure – Analysis and Computation (2003)
tina Franca 1999. Editor: G. P. Pirola (2002) Vol. 1804: I. Cherednik, Ya. Markov, R. Howe, G. Lusztig,
Vol. 1776: K. Behrend, C. Gomez, V. Tarasov, G. Tian, Iwahori-Hecke Algebras and their Representation Theory.
Quantum Comohology. Cetraro 1997. Editors: P. de Bar- Martina Franca, Italy 1999. Editors: V. Baldoni, D. Bar-
tolomeis, B. Dubrovin, C. Reina (2002) basch (2003)
Vol. 1805: F. Cao, Geometric Curve Evolution and Image Vol. 1829: E. Altman, B. Gaujal, A. Hordijk, Discrete-
Processing (2003) Event Control of Stochastic Networks: Multimodularity
Vol. 1806: H. Broer, I. Hoveijn. G. Lunther, G. Vegter, and Regularity. XIV, 313 p, 2003.
Bifurcations in Hamiltonian Systems. Computing Singu- Vol. 1830: M. I. Gil’, Operator Functions and Localization
larities by Gröbner Bases (2003) of Spectra. XIV, 256 p, 2003.
Vol. 1807: V. D. Milman, G. Schechtman (Eds.), Geomet- Vol. 1831: A. Connes, J. Cuntz, E. Guentner, N. Hig-
ric Aspects of Functional Analysis. Israel Seminar 2000- son, J. E. Kaminker, Noncommutative Geometry, Martina
2002 (2003) Franca, Italy 2002. Editors: S. Doplicher, L. Longo (2004)
Vol. 1808: W. Schindler, Measures with Symmetry Prop- Vol. 1832: J. Azéma, M. Émery, M. Ledoux, M. Yor
erties (2003) (Eds.), Séminaire de Probabilités XXXVII (2003)
Vol. 1809: O. Steinbach, Stability Estimates for Hybrid Vol. 1833: D.-Q. Jiang, M. Qian, M.-P. Qian, Mathemati-
Coupled Domain Decomposition Methods (2003) cal Theory of Nonequilibrium Steady States. On the Fron-
tier of Probability and Dynamical Systems. IX, 280 p,
Vol. 1810: J. Wengenroth, Derived Functors in Functional
2004.
Analysis (2003)
Vol. 1834: Yo. Yomdin, G. Comte, Tame Geometry with
Vol. 1811: J. Stevens, Deformations of Singularities
Application in Smooth Analysis. VIII, 186 p, 2004.
(2003)
Vol. 1835: O.T. Izhboldin, B. Kahn, N.A. Karpenko, A.
Vol. 1812: L. Ambrosio, K. Deckelnick, G. Dziuk, M. Vishik, Geometric Methods in the Algebraic Theory of
Mimura, V. A. Solonnikov, H. M. Soner, Mathematical Quadratic Forms. Summer School, Lens, 2000. Editor: J.-
Aspects of Evolving Interfaces. Madeira, Funchal, Portu- P. Tignol (2004)
gal 2000. Editors: P. Colli, J. F. Rodrigues (2003) Vol. 1836: C. Nǎstǎsescu, F. Van Oystaeyen, Methods of
Vol. 1813: L. Ambrosio, L. A. Caffarelli, Y. Brenier, G. Graded Rings. XIII, 304 p, 2004.
Buttazzo, C. Villani, Optimal Transportation and its Ap- Vol. 1837: S. Tavaré, O. Zeitouni, Lectures on Probabil-
plications. Martina Franca, Italy 2001. Editors: L. A. Caf- ity Theory and Statistics. Ecole d’Eté de Probabilités de
farelli, S. Salsa (2003) Saint-Flour XXXI-2001. Editor: J. Picard (2004)
Vol. 1814: P. Bank, F. Baudoin, H. Föllmer, L.C.G. Vol. 1838: A.J. Ganesh, N.W. O’Connell, D.J. Wischik,
Rogers, M. Soner, N. Touzi, Paris-Princeton Lectures on Big Queues. XII, 254 p, 2004.
Mathematical Finance 2002 (2003) Vol. 1839: R. Gohm, Noncommutative Stationary
Vol. 1815: A. M. Vershik (Ed.), Asymptotic Combina- Processes. VIII, 170 p, 2004.
torics with Applications to Mathematical Physics. St. Pe- Vol. 1840: B. Tsirelson, W. Werner, Lectures on Probabil-
tersburg, Russia 2001 (2003) ity Theory and Statistics. Ecole d’Eté de Probabilités de
Vol. 1816: S. Albeverio, W. Schachermayer, M. Tala- Saint-Flour XXXII-2002. Editor: J. Picard (2004)
grand, Lectures on Probability Theory and Statistics. Vol. 1841: W. Reichel, Uniqueness Theorems for Vari-
Ecole d’Eté de Probabilités de Saint-Flour XXX-2000. ational Problems by the Method of Transformation
Editor: P. Bernard (2003) Groups (2004)
Vol. 1817: E. Koelink, W. Van Assche(Eds.), Orthogonal Vol. 1842: T. Johnsen, A.L. Knutsen, K3 Projective Mod-
Polynomials and Special Functions. Leuven 2002 (2003) els in Scrolls (2004)
Vol. 1818: M. Bildhauer, Convex Variational Problems Vol. 1843: B. Jefferies, Spectral Properties of Noncom-
with Linear, nearly Linear and/or Anisotropic Growth muting Operators (2004)
Conditions (2003) Vol. 1844: K.F. Siburg, The Principle of Least Action in
Vol. 1819: D. Masser, Yu. V. Nesterenko, H. P. Schlick- Geometry and Dynamics (2004)
ewei, W. M. Schmidt, M. Waldschmidt, Diophantine Ap- Vol. 1845: Min Ho Lee, Mixed Automorphic Forms, Torus
proximation. Cetraro, Italy 2000. Editors: F. Amoroso, U. Bundles, and Jacobi Forms (2004)
Zannier (2003) Vol. 1846: H. Ammari, H. Kang, Reconstruction of Small
Vol. 1820: F. Hiai, H. Kosaki, Means of Hilbert Space Op- Inhomogeneities from Boundary Measurements (2004)
erators (2003) Vol. 1847: T.R. Bielecki, T. Björk, M. Jeanblanc, M.
Vol. 1821: S. Teufel, Adiabatic Perturbation Theory in Rutkowski, J.A. Scheinkman, W. Xiong, Paris-Princeton
Quantum Dynamics (2003) Lectures on Mathematical Finance 2003 (2004)
Vol. 1848: M. Abate, J. E. Fornaess, X. Huang, J. P. Rosay,
Vol. 1822: S.-N. Chow, R. Conti, R. Johnson, J. Mallet-
A. Tumanov, Real Methods in Complex and CR Geom-
Paret, R. Nussbaum, Dynamical Systems. Cetraro, Italy
etry, Martina Franca, Italy 2002. Editors: D. Zaitsev, G.
2000. Editors: J. W. Macki, P. Zecca (2003)
Zampieri (2004)
Vol. 1823: A. M. Anile, W. Allegretto, C. Ringhofer, Vol. 1849: Martin L. Brown, Heegner Modules and Ellip-
Mathematical Problems in Semiconductor Physics. Ce- tic Curves (2004)
traro, Italy 1998. Editor: A. M. Anile (2003)
Vol. 1850: V. D. Milman, G. Schechtman (Eds.), Geomet-
Vol. 1824: J. A. Navarro González, J. B. Sancho de Salas, ric Aspects of Functional Analysis. Israel Seminar 2002-
C ∞ – Differentiable Spaces (2003) 2003 (2004)
Vol. 1825: J. H. Bramble, A. Cohen, W. Dahmen, Mul- Vol. 1851: O. Catoni, Statistical Learning Theory and Sto-
tiscale Problems and Methods in Numerical Simulations, chastic Optimization (2004)
Martina Franca, Italy 2001. Editor: C. Canuto (2003) Vol. 1852: A.S. Kechris, B.D. Miller, Topics in Orbit
Vol. 1826: K. Dohmen, Improved Bonferroni Inequal- Equivalence (2004)
ities via Abstract Tubes. Inequalities and Identities of Vol. 1853: Ch. Favre, M. Jonsson, The Valuative Tree
Inclusion-Exclusion Type. VIII, 113 p, 2003. (2004)
Vol. 1827: K. M. Pilgrim, Combinations of Complex Dy- Vol. 1854: O. Saeki, Topology of Singular Fibers of Dif-
namical Systems. IX, 118 p, 2003. ferential Maps (2004)
Vol. 1828: D. J. Green, Gröbner Bases and the Computa- Vol. 1855: G. Da Prato, P.C. Kunstmann, I. Lasiecka,
tion of Group Cohomology. XII, 138 p, 2003. A. Lunardi, R. Schnaubelt, L. Weis, Functional Analytic
Methods for Evolution Equations. Editors: M. Iannelli, R. Vol. 1880: S. Attal, A. Joye, C.-A. Pillet, Open Quantum
Nagel, S. Piazzera (2004) Systems I, The Hamiltonian Approach (2006)
Vol. 1856: K. Back, T.R. Bielecki, C. Hipp, S. Peng, Vol. 1881: S. Attal, A. Joye, C.-A. Pillet, Open Quantum
W. Schachermayer, Stochastic Methods in Finance, Bres- Systems II, The Markovian Approach (2006)
sanone/Brixen, Italy, 2003. Editors: M. Fritelli, W. Rung- Vol. 1882: S. Attal, A. Joye, C.-A. Pillet, Open Quantum
galdier (2004) Systems III, Recent Developments (2006)
Vol. 1857: M. Émery, M. Ledoux, M. Yor (Eds.), Sémi- Vol. 1883: W. Van Assche, F. Marcellàn (Eds.), Orthogo-
naire de Probabilités XXXVIII (2005) nal Polynomials and Special Functions, Computation and
Vol. 1858: A.S. Cherny, H.-J. Engelbert, Singular Stochas- Application (2006)
tic Differential Equations (2005) Vol. 1884: N. Hayashi, E.I. Kaikina, P.I. Naumkin, I.A.
Vol. 1859: E. Letellier, Fourier Transforms of Invariant Shishmarev, Asymptotics for Dissipative Nonlinear Equa-
Functions on Finite Reductive Lie Algebras (2005) tions (2006)
Vol. 1860: A. Borisyuk, G.B. Ermentrout, A. Friedman, D. Vol. 1885: A. Telcs, The Art of Random Walks (2006)
Terman, Tutorials in Mathematical Biosciences I. Mathe- Vol. 1886: S. Takamura, Splitting Deformations of Dege-
matical Neurosciences (2005) nerations of Complex Curves (2006)
Vol. 1861: G. Benettin, J. Henrard, S. Kuksin, Hamil- Vol. 1887: K. Habermann, L. Habermann, Introduction to
tonian Dynamics – Theory and Applications, Cetraro, Symplectic Dirac Operators (2006)
Italy, 1999. Editor: A. Giorgilli (2005) Vol. 1888: J. van der Hoeven, Transseries and Real Differ-
Vol. 1862: B. Helffer, F. Nier, Hypoelliptic Estimates and ential Algebra (2006)
Spectral Theory for Fokker-Planck Operators and Witten Vol. 1889: G. Osipenko, Dynamical Systems, Graphs, and
Laplacians (2005) Algorithms (2006)
Vol. 1863: H. Fürh, Abstract Harmonic Analysis of Con- Vol. 1890: M. Bunge, J. Frunk, Singular Coverings of
tinuous Wavelet Transforms (2005) Toposes (2006)
Vol. 1864: K. Efstathiou, Metamorphoses of Hamiltonian Vol. 1891: J.B. Friedlander, D.R. Heath-Brown, H.
Systems with Symmetries (2005) Iwaniec, J. Kaczorowski, Analytic Number Theory, Ce-
Vol. 1865: D. Applebaum, B.V. R. Bhat, J. Kustermans, J. traro, Italy, 2002. Editors: A. Perelli, C. Viola (2006)
M. Lindsay, Quantum Independent Increment Processes I. Vol. 1892: A. Baddeley, I. Bárány, R. Schneider, W. Weil,
From Classical Probability to Quantum Stochastic Calcu- Stochastic Geometry, Martina Franca, Italy, 2004. Editor:
lus. Editors: M. Schürmann, U. Franz (2005) W. Weil (2007)
Vol. 1866: O.E. Barndorff-Nielsen, U. Franz, R. Gohm, B. Vol. 1893: H. Hanßmann, Local and Semi-Local Bifur-
Kümmerer, S. Thorbjønsen, Quantum Independent Incre- cations in Hamiltonian Dynamical Systems, Results and
ment Processes II. Structure of Quantum Levy Processes, Examples (2007)
Classical Probability, and Physics. Editors: M. Schür- Vol. 1894: C.W. Groetsch, Stable Approximate Evaluation
mann, U. Franz, (2005) of Unbounded Operators (2007)
Vol. 1867: J. Sneyd (Ed.), Tutorials in Mathematical Bio-
sciences II. Mathematical Modeling of Calcium Dynamics
and Signal Transduction. (2005)
Vol. 1868: J. Jorgenson, S. Lang, Posn (R) and Eisenstein
Sereies. (2005)
Vol. 1869: A. Dembo, T. Funaki, Lectures on Probabil-
ity Theory and Statistics. Ecole d’Eté de Probabilités de
Saint-Flour XXXIII-2003. Editor: J. Picard (2005)
Vol. 1870: V.I. Gurariy, W. Lusky, Geometry of Müntz
Spaces and Related Questions. (2005)
Vol. 1871: P. Constantin, G. Gallavotti, A.V. Kazhikhov,
Y. Meyer, S. Ukai, Mathematical Foundation of Turbulent
Viscous Flows, Martina Franca, Italy, 2003. Editors: M.
Cannone, T. Miyakawa (2006)
Vol. 1872: A. Friedman (Ed.), Tutorials in Mathemati-
cal Biosciences III. Cell Cycle, Proliferation, and Cancer
(2006)
Vol. 1873: R. Mansuy, M. Yor, Random Times and En-
largements of Filtrations in a Brownian Setting (2006)
Vol. 1874: M. Yor, M. Émery (Eds.), In Memoriam Paul-
André Meyer - Séminaire de Probabilités XXXIX (2006)
Vol. 1875: J. Pitman, Combinatorial Stochastic Processes.
Ecole d’Eté de Probabilités de Saint-Flour XXXII-2002. Recent Reprints and New Editions
Editor: J. Picard (2006)
Vol. 1876: H. Herrlich, Axiom of Choice (2006) Vol. 1618: G. Pisier, Similarity Problems and Completely
Vol. 1877: J. Steuding, Value Distributions of L- Bounded Maps. 1995 – Second, Expanded Edition (2001)
Functions(2006) Vol. 1629: J.D. Moore, Lectures on Seiberg-Witten Invari-
Vol. 1878: R. Cerf, The Wulff Crystal in Ising and Percol- ants. 1997 – Second Edition (2001)
ation Models, Ecole d’Eté de Probabilits de Saint-Flour Vol. 1638: P. Vanhaecke, Integrable Systems in the realm
XXXIV-2004. Editor: Jean Picard (2006) of Algebraic Geometry. 1996 – Second Edition (2001)
Vol. 1879: G. Slade, The Lace Expansion and its Vol. 1702: J. Ma, J. Yong, Forward-Backward Stochas-
Appli- cations, Ecole d’Eté de Probabilités de Saint-Flour tic Differential Equations and their Applications. 1999. –
XXXIV-2004. Editor: Jean Picard (2006) Corrected 3rd printing (2005)