See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/38372636

On a Class of Hilbert-Schmidt Operators

Article in Journal of Integral Equations and Applications · June 2001

DOI: 10.1216/jiea/996986964 · Source: OAI

CITATIONS READS
0 1,417

1 author:

Dario Pierotti
Politecnico di Milano
59 PUBLICATIONS 677 CITATIONS

SEE PROFILE

All content following this page was uploaded by Dario Pierotti on 22 May 2014.

The user has requested enhancement of the downloaded file.

JOURNAL OF INTEGRAL EQUATIONS
AND APPLICATIONS
Volume 13, Number 2, Summer 2001

ON A CLASS OF
HILBERT-SCHMIDT OPERATORS

DARIO PIEROTTI

ABSTRACT. We consider a class of integral operators with

L2 kernel which are often encountered in the resolution of
boundary value problems and in scattering theory. We prove
that, under certain conditions, these operators are contrac-
tions; applications of this result to a nonlinear differential
equation at resonance and to an integral equation of inverse
scattering theory are discussed.

0. Introduction. In the applications of integral equations methods

to differential equations and scattering theory, one often deals with
integral operators K with L2 kernel K of the following type:
(1.1) K(x, y) = U (x) Γ(x, y)V (y), x, y ∈ Ω ⊆ Rn ,
where U , V ∈ L∞ (Ω) and the function Γ(x, y) is the kernel of a
bounded integral operator (not necessarily compact). In this note, we
discuss sufficient conditions on the functions Γ, U and V for K being
a contraction.
Although the result follows easily by applying well known arguments
of Functional Analysis, it may be useful for the solution of certain
integral equations with L2 kernels, related to boundary value problems
and scattering theory. We will discuss two applications: in the first
one, we provide a non-resonance condition for a nonlinear second
order differential equation, which guarantees unique solvability; in the
second example, we derive an alternative proof of the resolubility of
the Marchenko integral equation [6], which allows to solve the inverse
scattering problem for the Schrodinger equation.

1. Statement and proof of the main result. We prove our result

in the following form:

Theorem 1.1. Let K : L2 (Ω) → L2 (Ω) be a Hilbert-Schmidt

operator with kernel K(x, y) = U (x) Γ(x, y)V (y). Assume that: i) U ,
V are functions in L∞ (Ω), with ||U ||∞ ≤ 1, ||V ||∞ ≤ 1.
c
Copyright
2001 Rocky Mountain Mathematics Consortium

1
2 D. PIEROTTI

ii) The integral operator G defined by

(1.2) (Gϕ)(x) = Γ(x, y)ϕ(y)dy,
Ω

2
is bounded in L (Ω) with ||G|| ≤ 1.
iii) For any ϕ ∈ L2 (Ω) such that ||Gϕ|| = ||ϕ||, either ϕ = 0 or (at
least) one of the following relations hold:

m({x ∈ Ω : ϕ(x) = 0} ∩ {x ∈ Ω : |V (x)| < 1}) > 0;

m({x ∈ Ω : (Gϕ)(x) = 0} ∩ {x ∈ Ω : |U (x)| < 1}) > 0,

where m denotes the usual Lebesgue measure in Rn .

Alternatively, assume that i), ii) hold and that:
iii ) For any ψ ∈ L2 (Ω) such that ||G∗ ψ|| = ||ψ|| (G∗ is the adjoint
of G) either ψ = 0 or (at least) one of the following relations hold:

m({x ∈ Ω : ψ(x) = 0} ∩ {x ∈ Ω : |U (x)| < 1}) > 0;

m({x ∈ Ω : (G∗ ψ)(x) = 0} ∩ {x ∈ Ω : |V (x)| < 1}) > 0.

Then, ||K|| < 1.

Proof. By our assumptions, K is a compact operator with ||K|| ≤ 1;

then, K ∗ K is compact and self-adjoint and ||K ∗ K|| = ||K||2 ≤ 1. By
known properties of compact self-adjoint operators [1], [7], the equality
holds if and only if there exist χ ∈ L2 , χ = 0, such that K ∗ K χ = χ;
by taking the scalar product with χ we get

(1.3) ||K χ|| = ||χ||.

By the same reasoning applied to K ∗ and KK ∗ (recall that also K ∗ is

compact with ||K ∗ || ≤ 1) we also find

(1.4) ||K ∗ η|| = ||η||,

for some nontrivial η ∈ L2 . On the other hand, by (1.1) and assump-

tions i) and ii), we have

||K χ|| = ||U GV χ|| ≤ ||GV χ|| ≤ ||V χ|| ≤ ||χ||

 ON A CLASS OF HILBERT-SCHMIDT OPERATORS 3

and

||K ∗ η|| = ||V ∗ G∗ U ∗ η|| ≤ ||G∗ U ∗ η|| ≤ ||U ∗ η|| ≤ ||η||,

where we have denoted by U , V , U ∗ , V ∗ the multiplication operators by

the functions U (x), V (x) and by the complex conjugates U (x), V (x),
respectively. By the above estimates and (1.3), (1.4) we obtain the
equalities
||U GV χ|| = ||GV χ|| = ||V χ|| = ||χ||;
(1.5)
||V ∗ G∗ U ∗ η|| = ||G∗ U ∗ η|| = ||U ∗ η|| = ||η||.
Now, by setting ϕ = V χ and ψ = U ∗ η, it is readily verified that the
first equation in (1.5) is in contradiction with assumption iii) and the
second with assumption iii ).

Remark. For the applications of the above theorem, it may be

useful to point out some simple sufficient conditions for the validity
of assumption iii) or iii ).
Obviously, assumption iii) (or iii )) holds if there is no nontrivial
vector satisfying the relation ||Gϕ|| = ||ϕ|| (or ||G∗ ψ|| = ||ψ||). The
same is true if ϕ (ψ) is nontrivial, but the set {x ∈ Ω : |V (x)| = 1}
({x ∈ Ω : |U (x)| = 1}) has zero Lebesgue measure.
A further property which is sometimes verified in the applications (see
Section 3) is the analyticity of the function (Gϕ)(x). In this case, Gϕ
vanishes on a discrete set in Ω, so that assumption iii) holds if |U | < 1
on a set of positive measure in Ω. The same is true for assumption iii )
if (G∗ ψ)(x) is analytic and |V | < 1 on a set of positive measure.
Another useful sufficient condition is given by the following

Corollary 1.2. Let K be defined as in Theorem 1.1 and such that

assumptions i) and ii) hold. Let us further assume that Ω is bounded,
Γ(x, y) is symmetric and it is the Green function of an elliptic operator
L acting on H01 (Ω). Finally, suppose that either |U | < 1 or |V | < 1 on
a set of positive measure in Ω. Then, ||K|| < 1.

Proof. By our assumptions, the operator G = L−1 is compact and

self-adjoint with eigenvalues µk such that 1 ≥ µ1 > µ2 ≥ . . . ≥
4 D. PIEROTTI

µk ≥ . . . ; then, ||Gϕ|| = ||ϕ|| > 0 if and only if µ1 = 1 and

ϕ is the first eigenfunction of L; hence, |ϕ| > 0 in Ω. Then, if
m({x : V (x) < 1}) > 0, one easily checks that assumption iii) holds.
Finally, since G = G∗ the same conclusion follows for assumption iii )
if m({x : U (x) < 1}) > 0.

2. Application to differential equations. Let us consider the

nonlinear problem: given f ∈ L2 (0, π) and g ∈ C([0, π] × R), find u
such that

(2.1) −u + g(x, u(x)) = f (x); u(0) = u(π) = 0.

We can prove the following result:

Proposition 2.1. Let us suppose that g, gu are bounded and

continuous functions with |gu | ≤ 1. Assume further that |gu (0, 0)| < 1
or |gu (π, 0)| < 1. Then, problem (2.1) has a unique solution in
H01 (0, π).

Proof. Let us define the operator in H01 (0, π):

π
(2.2) Bu = Γ(x, t)[f (t) − g(t, u(t))] dt,
0

where 
(π − x)t, 0 ≤ t ≤ x ≤ π,
Γ(x, t) =
x(π − t), 0 ≤ x < t ≤ π,
is the Green function of the operator −(d2 /dx2 ) in (0, π) with homo-
geneous Dirichlet conditions. Recall that
∞
2  sin(nx) sin(nt)
Γ(x, t) = .
π n=1 n2

Then, the corresponding Green operator G is compact and self-adjoint

in L2 (0, π), with
∞ 
 cn 2
(2.3) (Gϕ)(x) = sin(nx),
n=1
n2 π
 ON A CLASS OF HILBERT-SCHMIDT OPERATORS 5

where cn are the Fourier coefficients of ϕ. Since {cn } ∈ l2 we can

differentiate (2.3) term by term and obtain

∞ 
 cn 2

(2.4) (Gϕ) (x) = cos(nx),
n=1
n π

From (2.3), (2.4) it follows that ||G|| = ||G || = 1 and that ||Gϕ||
 = ||ϕ||
(or ||G ϕ|| = ||ϕ||) if and only if ϕ(x) is proportional to 2/π sin x,
the first (normalized) eigenfunction.
Clearly, the weak form of (2.1) is equivalent to the fixed point
equation

(2.5) Bu = u,

in the space H01 (0, π). Moreover, by the above discussion and the
assumptions on g, we have ||Bu||H01 ≤ M for some positive constant
M . Hence, we will consider (2.5) in the closed ball

BM = {u ∈ H01 (0, π) : ||u||H01 ≤ M }.

For every u, v in BM we have now

(2.6)
π
(Bu−Bv)(x) = Γ(x, t)[g(t, v(t)) − g(t, u(t))] dt

0 π
= Γ(x, t)gu (t, λv(t) + (1−λ)u(t))[u(t) − v(t)] dt,
0

for some λ = λ(t) ∈ (0, 1). By our assumptions, there is δ > 0 such
that |gs (t, s)| < 1 for (t, s) ∈ [0, δ] × [0, δ] or (t, s) ∈ [π − δ, π] × [0, δ].
We now show that there exists β, depending on δ and M , such that
|u(t)| ≤ δ for every u ∈ BM and t ∈ [0, β] or t ∈ [π − β, π]. In fact, if
the above property does not hold, we can find a sequence {un } ∈ BM
satisfying un (1/n) > δ (or un (π − 1/n) > δ) for every n = 1, 2, . . . ,.
However, by applying Hölder inequality we find

1/n 1/2
1/n
√ √
|un (t)|2 dt ≥ n |un (t)| dt ≥ n δ,
0 0
6 D. PIEROTTI

so that {un } cannot be contained in BM . Hence, we obtain

|gu (t, λv(t) + (1 − λ)u(t))| < 1,

for t ∈ [0, min(β, δ)] (or t ∈ [π −min(β, δ), π]), uniformly with respect to
u, v in BM . By denoting with K = K(u, v) the integral operator with
kernel Γ(x, t)gu (t, λv(t) + (1 − λ)u(t)), we get by Corollary 1.2 that K
is a contraction in L2 (0, π), uniformly with respect to u, v in BM . By
(2.4) and the subsequent discussion, it is clear that the same property
holds for K considered as an operator in H01 (0, π). Then, from (2.6)
we get

(2.7) ||(Bu − Bv)||H01 ≤ L||(u − v)||H01 ,

for every u, v ∈ BM , where L < 1. The proposition now follows by the

Banach contraction principle.

Remark. The above result still holds if g is not bounded, but satisfies
an estimate of the type

(2.8) |g(t, s)| ≤ g0 (t) + α(t)|s|, (t, s) ∈ [0, π] × R,

where g0 ∈ L2 (0, π), 0 ≤ α(t) ≤ 1, α(t) ≤ γ < 1 on a set J ⊆ (0, π) of

positive measure and g0 |J is strictly positive. (Clearly, we can always
take g0 with this property in (2.8)).
In order to prove our claim, we first note that by (2.2) and the
properties of the Green operator it follows that

||Bu||H01 ≤ ||f ||L2 + ||g||L2 .

Furthermore, by (2.8) we have

(2.9) ||g||L2 ≤ ||g0 ||L2 + ||α u||L2 ≤ ||g0 ||L2 + ||u||L2 .

Suppose now that {un } is a sequence with ||un ||L2 = 1 and such that
limn→∞ ||α un ||L2 = 1; then

(2.10) lim ||χJ un ||L2 = 0,

n→∞
 ON A CLASS OF HILBERT-SCHMIDT OPERATORS 7

where χJ is the characteristic function of the set J. Let us further

assume that there exists a subsequence {|um |} (which can be chosen
weakly convergent in L2 (0, π)) with

||g0 + α |um | ||L2

(2.11) lim = 1.
m→∞ ||g0 ||L2 + ||α um ||L2

Then, necessarily, α|um | ' λg0 for some λ > 0 (recall that g0 = 0). In
particular, by recalling (2.10), we find



1
g0 = lim α|um | = 0.
J λ m→∞ J

Since g0 is strictly positive on J, we have a contradiction; as a result,

we can refine (2.9) by writing

||g||L2 ≤ C(||g0 ||L2 + ||u||L2 ),

for some C < 1. Finally, we obtain

||Bu||H01 ≤ ||f ||L2 + C(||g0 ||L2 + ||u||L2 ) ≤ ||f ||L2 + C(||g0 ||L2 + ||u||H01 ).

As a consequence, we have B(BM ) ⊆ BM for every M such that

||f ||L2 + C||g0 ||L2

M≥ .
1−C

We recall that growth assumptions of the type (2.8) are sufficient

non-resonance conditions in a large class of boundary value problems
for nonlinear equations (see, e.g., [3] [4]). In Proposition 2.1, we
provide a simple additional condition on the nonlinearity for the unique
resolubility of problem (2.1).

3. Application to inverse scattering. The (one-dimensional)

Marchenko equation for left-going waves scattered by a potential q(x)
supported in the half-space x > 0 can be written [2],

x
(3.1) A(x, y) + R(y + s)A(x, s) ds + R(x + y) = 0, y < x,
−x
8 D. PIEROTTI

where A(x, y) is the unknown kernel and R(x) is the impulse response
function. If the potential does not support bound states, the impulse
response function is the Fourier transform of the reflection coefficient
R̂(k):


1
(3.2) R(x) = e−ikx R̂(k) dk.
2π R

For potentials that decrease sufficiently fast at infinity, the reflection

coefficient is a continuous function on the real axis with the asymptotic
behavior
 
1
(3.3) R̂(k) = O ,
|k|

for |k| → ∞ [5].

Assuming the above properties, it is convenient to consider a frequency-
domain formulation of (3.1). We set


1
(3.4) A(x, y) = − e−iky M (x, k) dk,
2π R

and substitute (3.2), (3.4) in (3.1); by formal calculations we obtain

the integral equation


sin[(k + k )x]
(3.5) M (x, k) + R̂(k) )
M (x, k ) dk = R̂(k)e−ikx .
R π(k + k

For every x, we can regard (3.5) as an operator equation in L2 (R, dk)

for the unknown vector M (x, ·):

(3.6) M (x, k) + (Kx M )(x, k) = R̂(k)e−ikx ,

where


sin[(k + k )x]
(3.7) (Kx M )(x, k) = R̂(k) M (x, k ) dk .
R π(k + k )

The operator Kx is Hilbert-Schmidt since


 sin[(k + k )x] 2 |x|
R̂(k)   dk dk = ||R̂||2L2 ,
R2 π(k + k ) π
 ON A CLASS OF HILBERT-SCHMIDT OPERATORS 9

and the L2 norm of R̂ is finite by (3.3). We further note that for small
enough x the operator Kx is a contraction; actually, this property holds
for every x:

Proposition 3.1. For any x ∈ R the operator (3.7) is a contraction

in L2 (R, dk) and the equation (3.5) has a unique solution.

Proof. As remarked above, Kx is Hilbert-Schmidt; furthermore, by

the properties of the reflection coefficient, we have |R̂(k)| ≤ 1 for every
k. Let us now consider the operator Gx defined by


sin[(k + k )x]
(Gx ϕ)(k) = )
ϕ(k ) dk .
R π(k + k
We may write

(3.8) Gx = F −1 · Px · F · I,

where F is the Fourier transformation, Px the multiplication operator

by the characteristic function of the interval (−x, x) and I is defined by
(If )(k) = f (−k), for every f ∈ L2 (R). It follows that Gx is bounded
(but not compact) in L2 (R, dk), with ||Gx || ≤ 1. Moreover, again
by (3.8), the function (Gx ϕ)(k) is the Fourier (anti) transform of a
function of compact support; hence, for every ϕ and every x, (Gx ϕ)(k)
is an entire function of exponential type in k. By the remark following
Theorem 1.1, the proposition follows if |R̂(k)| < 1 on a set of positive
measure; this is true by equation (3.3).

Remark. The formal calculations leading to (3.5) can be rigorously

justified by approximating R̂ and M (x, ·) with L1 ∩ L2 functions.
Finally, we recall that the potential q(x) of the Schrödinger equation is
recovered from the kernel A(x, y) by the relation

d
q(x) = 2 A(x, x).
dx
Thus, we get by (3.4)


1 d
q(x) = − e−ikx M (x, k) dk.
π dx R
 10 D. PIEROTTI

REFERENCES

1. H. Brezis, Analyse fonctionnelle Théorie et applications, Masson Editeur,

Paris, 1983.
2. R. Burridge, The Gel’fand-Levitan, the Marchenko, and the Gopinath-Sondhi
integral equations of inverse scattering theory, regarded in the context of inverse
impulse-response problems, Wave Motion 2 (1980), 305 323.
3. A. Fonda and J. Mawhin, Quadratic forms, weighted eigenfunctions and
boundary value problems for non-linear second order ordinary differential equations,
Proc. Roy. Soc. Edinburg Sect. A 112 (1989), 145 153.
4. J. Mawhin, J.R. Ward, Jr. and M. Willem, Variational methods and semi-linear
elliptic equations, Arch. Rational Mech. Anal. 95 (1986), 269 277.
5. R.G. Newton, Inverse scattering I. One dimension, J. Math. Phys. 21 (1980),
493 505.
6. , Scattering theory of waves and particles, Springer, New York, 1982.
7. M. Reed and B. Simon, Methods of modern mathematical physics, I: Functional
analysis, Academic Press, New York, 1972.

Dipartimento di Matematica del Politecnico, Piazza L. da Vinci 32, 20133

Milano, Italy
E-mail address: [email protected]

View publication stats