1. Introduction
Since its inception, the mathematical treatment of quantum theory has generated many problems in measure and integration theory, some of which are still being worked out. At the forefront is the spectral theory of self-adjoint operators in Hilbert space, where the decomposition of a Hermitian matrix with respect to the orthogonal projections onto the eigenspace of the eigenvalues of A is replaced by the spectral decomposition with respect to the self-adjoint spectral measure P associated with the self-adjoint linear operator T. The spectrum of T is the complement of the set of all numbers for which is a bounded linear operator on , so any eigenvalue of T belongs to .
For a quantum system in a state
, the conventional interpretation of quantum measurement suggests that the number
is the probability that an observation of the quantity represented by the self-adjoint operator
T has its value in the Borel set
. If
ψ is also an eigenvector of
T for the eigenvalue
, then:
that is, in the state
ψ, an observation of
T yields the value
λ with certainty, so explaining such facts as the quantisation of energy levels in an atom. Clearly, the operator-valued spectral measure uniquely associated with a quantum observable is a fundamental concept in quantum theory.
Another problem of integration theory arising from quantum physics is the Feynman-Kac formula:
The left-hand side represents the dynamics of a quantum system described by a free Hamiltonian
perturbed by a potential
V where
is the operator of multiplication by
V. The finitely-additive operator valued set functions
,
, are manufactured from the free evolution
,
, and the spectral measure
Q associated with the configuration or position operators; see [
1,
2,
3,
4] and the extensive references in these monographs. Although we shall need to consider integration with respect to certain finitely additive set functions manufactured from a pair of spectral measures, the operator valued integrals considered in the present work are orders of magnitude more tractable than the singular path integral on the right hand side of equation (
1). Even at the basic level of integration theory considered here, Grothendieck’s inequality [
26] provides essential insights.
A number of problems in scattering theory, spectral theory and their applications in the context of a Hilbert space
are treated by considering integrals of the form:
for spectral measures
E and
F,
and operator valued functions
and scalar functions
φ. Such integrals are
bilinear in the functions Φ and operator valued measures
E and
F. Unfortunately, the well-developed theories of bilinear integration of Bartle [
5] and Dobrakov [
6,
7,
8,
9] do not apply in this situation due to the variational properties of spectral measures acting on a Hilbert space.
The 2-variation
for a spectral measure
E and a vector
is always finite. The supremum is taken over all finite partitions
into Borel sets. However, for the spectral measure of multiplication by characteristic functions say, the
total variation of the
-valued measure
,
, is infinite on each set of positive measure where
is nonzero, which leads to difficulties interpreting integrals like (2) using the classical theory of bilinear integration utilising semivariation [
5].
An integral of the first form in (2) arises in treating the connection between time-dependent scattering theory and stationary state scattering theory [
10,
11]. Bilinear integrals of this nature have recently been handled by a “decoupling” method in [
12]. In this technique, an auxiliary tensor product
is defined, and the tensor product integral:
is an element of
for each
, in the fashion of [
13]. The bilinear evaluation map
,
,
, uniquely defines a continuous linear map
for which
for every
and
, so that:
By this means, the variational properties of the spectral measure E play no role in the definition of the first integral in (2).
Similar difficulties arise in the theory of stochastic integration, which is now thoroughly understood. For a Brownian motion process
with respect to a probability measure
P, there exists a unique
-valued orthogonally scattered measure
W given by
,
. The multiplication map
for random variables
is actually continuous into
on the closure of
in
as
X runs over all adapted simple processes; see [
14], Theorem 5.3. For an adapted process
X, the stochastic integral
may be viewed as an example of a bilinear integral
. The two-variation of
W on a Borel set
is the Lebesgue measure
of
B, but
has infinite variation on any set of positive Lebesgue measure: the variation of a Brownian motion process
on any interval is infinite
P-a.e.
The last type of integral in (2) is a double operator integral studied in a series of papers by Birman and Solomyak [
15,
16,
17,
18,
19,
20]. With the choice of the function:
for a sufficiently smooth function
, the equality:
holds if
are densely-defined self-adjoint operators in the Hilbert space
with spectral measures
E and
F, respectively, and
belongs to the operator ideal
of all trace class operators on
. The finitely-additive spectral measure
acts on the operator ideal
.
Integration with respect to finitely-additive spectral measures is studied in considerable detail in [
21],
Section 2, where for an algebra
of subsets of a non-empty set Ω and a complex Banach space
X, a finitely-additive set function
is called a finitely-additive spectral measure if
for all
and
, the identity operator on
X. In the case that
is actually a
σ-algebra of subsets of Ω and
M is countably additive for the strong operator topology of
, the operator-valued measure
M is simply called a spectral measure. The spectral theorem for a self-adjoint operator
T acting in a Hilbert space asserts the existence of a unique spectral measure
P whose values are self-adjoint projection operators, such that
. The finitely-additive spectral measure
is defined on the algebra
generated by measurable rectangles
by the formula:
for
,
. It is only in trivial cases that
is countably additive for the strong operator topology of
; see Proposition 19 below. The finitely-additive spectral measure
is similarly defined in the case that 𝔖 is a symmetrically-normed operator ideal. The commutative Banach *-algebra
of all equivalence classes of
-integrable functions is discussed in Proposition 15 below.
An application of the formula above leads to the expression:
for a finite Borel measure
Ξ on
. It turns out that
Ξ is absolutely continuous with respect to the Lebesgue measure, and its density
ξ with respect to the Lebesgue measure is Krein’s spectral shift function with respect to the pair
. If
is the scattering operator associated with the self-adjoint operators
A and
B, then the remarkable formula:
holds ([
22], Chapter 8).
Examples of the other integrals of the form (2) arise in the spectral theory of block operators, resonance and optimal control theory, numerical analysis and the theory of Krein’s spectral shift function in scattering theory and non-commutative geometry. The authoritative treatment of the applications of the decoupling approach to bilinear integration is given in the papers referred to in the sections to follow.
Section 2 deals with solutions of operator equations
where
A is a self-adjoint operator,
B is a closed linear operator and
Y is bounded. Representations for the solution
X given in
Section 3 in terms of an integral of the second form in (2) leads to estimates for the norm of
X in terms of the spectral separation between the linear operators
A and
B. As in the case of the connection between time-dependent and stationary state scattering theory [
10,
11], such integrals have been previously referred to as strong operator-valued Stieltjes integrals in [
23,
24]. A brief account of the connection with the spectral analysis of block operator matrices considered in [
23,
24] is also given. In case
B is also a self-adjoint operator, the solution
X of the operator equation above can be expressed as a double operator integral described in
Section 4 using decoupled integrals. For spectral measures
E and
F acting on a Hilbert space
, a double operator integral is an integral with respect to a finitely-additive spectral measure
acting on a symmetrically normed operator ideal
. The space
of
-integrable functions has been characterised by Peller in [
25] for the cases
and
. An elementary proof of Peller’s characterisation is given in Theorem 16 of
Section 5 by appealing to Pisier’s recent account [
26] of Grothendieck’s theorem. Peller’s representation facilitates an explicit formula given in [
20] for the trace of the integral
for
in the case
. The existence of Krein’s spectral shift function is established in Theorem 22 of
Section 6 using double operator integrals and Fourier transforms.
2. Linear Operator Equations
The analysis of the equation
for linear operators
A,
B,
X and
Y acting in a Hilbert space
has many applications in operator theory, differential equations and quantum physics; see [
27] for a relaxed discussion with numerous examples.
Starting with the case of scalars, the equation
has a unique solution provided that
. For the case of diagonal matrices
and
, for any matrix
, there exists a unique solution
X of the equation
if and only if
for
, and then, the solution
is given by:
The operator version is called the Sylvester–Rosenblum theorem in [
27], although earlier versions are due to Krein and Daletskii ([
27], p. 1). For a continuous linear operator
A on a Banach space
, the spectrum
of
A is the set of all
for which
is not invertible.
Theorem 1 (Sylvester–Rosenblum theorem). Let be a Banach space and let A and B be continuous linear operators on for which . Then, for each operator , the equation has a unique solution .
As a taster for applications of the Sylvester–Rosenblum Theorem, suppose that
A and
B are bounded normal operators on a Hilbert space
with spectral measures
and
, respectively. Then, there exists
, such that for any two Borel subsets
and
of
separated by a distance:
the projections
,
, satisfy the norm estimate:
The norm represents the angle between the subspaces and . Such estimates are useful in numerical computations. Even in finite dimensional Hilbert spaces, the Sylvester–Rosenblum theorem leads to eigenvalue estimates for matrix norms independent of dimension.
Theorem 2 ([
28], Theorem 5.1a)
. Let A and B be two normal matrices with eigenvalues and ,
respectively, counting multiplicity. With the same constant c mentioned above, if ,
then there exists a permutation π of the index set ,
such that:for .
The Sylvester–Rosenblum theorem also comes with a representation of the solution
X of the equation
if
. Suppose that the contour Γ is the union of closed contours in the plane, with total windings one around
and zero around
. Then:
Other representations of the solution are possible by utilising the spectral properties of the operators
A and
B; see [
27], Section 9.
In the present paper, we are concerned with solutions
X of the operator equation
when
A is an unbounded self-adjoint or normal operator acting in a Hilbert space
and
B is a closed unbounded operator. If the spectra
and
are a positive distance apart, then we hope to construct the solution
X of
by the formula:
in place of (
3) with respect to the spectral measure
of
A. The operator-valued measure
acts on the values of the operator-valued function
. As in the case of scattering theory considered in [
12], for
, the vector
often has the representation:
where the
-valued function
,
, is
-integrable in the tensor product space
and
is the continuous linear extension of the composition map
,
,
.
If the operator
B is itself a bounded linear operator, then the simpler representation (
3) may be employed with the contour Γ winding once around
and zero times around
.
Because we shall be dealing with unbounded operators
A and
B, we have to be careful about domains when interpreting the equation
. We follow the treatment in [
23],
Section 2. Applications of Equation (
4) to perturbation theory and the spectral shift function may also be found in [
23] and at the end of the next section.
3. Integral Solutions of Operator Equations
Definition 1. Let
and
be Hilbert spaces. Suppose that
and
are closed and densely-defined linear operators with domains
and
. Given
, a continuous linear operator
is said to be a weak solution of the equation:
if for every
and
, the equality:
holds with respect to the inner product
of
.
The domain of the adjoint of A is the set of all elements k of , such that the linear map , , is the restriction to of , , for an element and then .
A strong solution
of (
5) has the property that:
and:
The existence of strong solutions of the operator Equation (
5) is discussed in [
29] under the assumption that
A and
are the generators of
-semigroups, a situation that arises in delay or partial differential equations and control theory. Strong solutions of (
5) may not exist in this setting, even when the spectra
and
are separated by a vertical strip ([
29], Example 9).
In the case that
A and
B are both self-adjoint operators, the following result is a consequence of [
28], Theorem 4.1; (see [
23], Theorem 2.7).
Theorem 3. Let and be Hilbert spaces. Suppose that and are self-adjoint operators whose spectra and are a distance apart. Then, Equation (5) has a unique weak solution:for any function ,
continuous on ,
such that: Moreover .
The integral representing the solution X is a Pettis integral for the strong operator topology.
We now turn to the tensor product topology
τ mentioned above. Let
be Banach spaces. For
, we have:
for all
and
,
and all
. Hence, if we let:
over all representations
,
, of
, then the inequality
holds for the product map
by the Hahn–Banach theorem. The completion of the linear space
with respect to the norm
is written as
.
For a self-adjoint operator
A in a Hilbert space
and a closed, densely-defined operator
B in a Hilbert space
, the domains
and
are endowed with the respective graph norms associated with the closed operators
B and
A. Suppose also that
τ is the topology on the tensor product
defined by Formula (
9) with
, and let
be the completion of the tensor product with the norm topology
τ. As in [
12], Proposition B.11, the collection
of continuous linear functionals on the Banach space
E separates points of
E, and the composition map:
has a continuous linear extension
. The following definition of a bilinear integral is suggested by [
13].
Definition 2. Let be a Hilbert space. A function is said to be m-integrable in for an operator valued measure ,
if for each ,
the scalar function is integrable with respect to the scalar measure and for each ,
there exists an element of E, such that:for every .
If f is m-integrable in E, then is defined for each by: We also denote by or .
In the present context, the representation of solutions of Equation (5) via bilinear integration is analogous to the treatment in [
12],
Section 3, for scattering theory.
Example 3. Suppose that A is a bounded self-adjoint operator defined on a Hilbert space , such that for some . Let be the generator of a uniformly-bounded -semigroup , , on the Hilbert space .
We can employ (3) in this situation to represent the weak solution of Equation (5), but it is instructive to see how the integral (4) converges with the assumptions above.
Let
be the projective tensor product of the Hilbert space
with the space
of bounded linear operators on
with the uniform norm (see [
30], Section III.6). Then,
belongs to the tensor product
for each
and
, and the function
,
, is continuous in
, because
A is assumed to be bounded, so:
converges in
uniformly for
t in any bounded interval. The inequalities:
ensure that
converges as a Bochner integral in the projective tensor product
and:
belongs to
, too. Then:
defines a continuous linear operator:
belonging to
with norm bounded by:
In order to deal with unbounded operators, we replace the projective tensor product topology π by the topology τ defined by Formula (9).
Lemma 4. Let and be Hilbert spaces. Suppose that is a self-adjoint operator with spectral measure and is a densely-defined, closed linear operator, such that .
Let .
For each ,
the -valued function:is -integrable in on every compact subset of .
Furthermore, there exist -valued -simple functions:such that for each ,
in as for -almost all and for each compact subset of K of ,
as .
Proof. For a closed and densely-defined operator
T, the resolvent
is defined for all complex numbers
λ belonging to the resolvent set
. Suppose that
is non-empty. Then, the resolvent equation:
for
ensures that
,
, is a holomorphic operator-valued function for the uniform operator topology. It follows that for each
, the function:
is continuous in the projective tensor product
for the uniform norm on
. For a compact subset
K of
, let
be the part of
A on
K. Then, for a contour
with a winding of one around
K and zero around the closed set
, the integral:
is bounded by
, so the function:
converges as a Bochner integral in
. For every Borel subset
S of the set
K and
, an application of Cauchy’s integral formula yields:
so according to Definition 2 (replacing the topology
τ by the stronger projective topology
π), the function
is
-integrable in
on the set
K and:
as an element of the projective tensor product
for each Borel subset
S of
K.
Because the operator-valued function
,
, is uniformly continuous on the compact set
K, for each
, there exists an
-valued
-simple function
, such that:
so that:
as
for each
. According to the identity (12), it follows that:
as
. Because the spectral measure
is inner regular on compact sets, the simple functions
,
, can be pieced together from the simple functions
,
, on each compact set
K. ☐
If both operators
A and
B are self-adjoint, then Theorem 3 ensures that a weak solution
X of Equation (
5) exists and gives a norm estimate for
X. If just one operator is self-adjoint, the following result is applicable.
Theorem 5. Let and
be Hilbert spaces. Suppose that is a self-adjoint operator with spectral measure ,
and is a densely-defined, closed linear operator, such that .
Let .
- (i)
Equation (5) has a strong solution if and only if there exists an operator valued measure ,
such that:for each compact subset K of .
The operator valued measure M exists if and only if:for every .
Then,
is the unique strong solution of Equation (5). - (ii)
If for each ,
the function given by Formula (11) is -integrable in on ,
then the map ,
defines a continuous linear operator ,
and the operator:is the unique strong solution of Equation (5). Let .
The function is -integrable in on if and only if:
Proof. The proof of (i) is similar to the proof of (ii), which we now give. Suppose that for each
, the function
is
-integrable in
on
. Then, for
, we have:
because
for all
, and by Formula (
3), the operator
is the unique solution of the equation:
The case of unbounded
B is mentioned in [
23], Lemma 2.5. Because
and
A and
commute, we have
for all
. Now,
converges in
as
for each
, so
belongs to
by the uniform boundedness principle. Suppose that
. Then,
, so
belongs to the closure of
A restricted to the subspace:
Hence, , and X is therefore a strong solution of Equation (5). On the other hand, if (5) does have a strong solution X, it can be written as with uniformly bounded over compact sets .
Conversely, suppose that the bound (14) holds for every
. There exists an increasing sequence of compact subsets
,
, of
, such that:
for every
and
, because the spectral measure
is a regular operator valued Borel measure. Let
. Then,
is
-null, and
are pairwise disjoint.
For each
,
and
:
If the bound (14) holds, then:
The projective tensor product
is associated with the trace class operators on
via the embedding
defined by
. Then:
for
and the bound:
holds for each
and
by [
31], Proposition I.1.11. It follows from the weak sequential completeness of the Hilbert space
and the Orlicz–Pettis theorem ([
31], Corollary I.4.4) that the sum
converges unconditionally in
for each
and:
is a bounded linear operator whose norm is bounded by
. According to the non-commutative Fatou lemma (see
Section 4),
belongs to
and:
for each
. Hence, the function
is
-integrable in
on
and
. The uniform boundedness principle and the Vitali–Hahn–Saks theorem ensures that the formula
defines an
-valued measure
M for the strong operator topology, so that (i) applies. ☐
Remark 4. The operator-valued measure
is called a strong operator-valued Stieltjes integral in [
23,
24]. According to Lemma 4, for each compact subset
K of
, the operator
can be written as a Stieltjes integral:
for
-simple function
,
, which may be chosen to be step functions based on finite intervals, restricted to the spectrum
of
A.
Example 5. The solution
X in Theorem 3 is actually a strong solution. If
A and
B are self-adjoint and
, then:
belongs to
for each
and
. To see this, let
. Then, the integral:
converges in
because
,
, is continuous in
and
, so:
and
. Then, by an appeal to Theorem 5 (ii), the operator:
is the unique strong solution of Equation (
5). It is shown in [
24], Lemma 4.2, that there is actually no distinction between weak and strong solutions of the Sylvester–Rosenblum Equation (
5) because the bound (
13) follows from the boundedness in the weak operator topology.
Example 6. If
A is self-adjoint,
B is densely defined and closed,
and there exists
and a sector:
that is contained in
, then according to [
29], Theorem 15:
The application of the integral representation of solutions of the Sylvester–Rosenblum Equation (
5) to the spectral analysis of block operator matrices is discussed in detail in [
23,
24]. It is worthwhile to mention the background concerning self-adjoint operator block matrices:
acting in the orthogonal sum
of separable Hilbert spaces
and
. Then,
H can also be written as
for the operator matrices:
with
A self-adjoint and
B bounded. A strong solution
Q of the equation:
having the form:
determines a reducing subspace for the original block operator matrix operator
H, so that:
Consequently, if U is the unitary operator associated with the polar decomposition , then is the block diagonalization of H for self-adjoint operators , similar to the operators and , respectively.
The Equation (
15) is called Riccati’s equation. It also arises in optimal control theory when the operator entries may not be self-adjoint; see [
23] for a list of references. Equation (
15) is quadratic in
Q, and provided that
with respect to the distance
between the spectra of
U and
V, a fixed point argument produces a unique strong solution
Q of the associated operator equation
([
23], Theorem 3.6) in the case that
U and
V are bounded self-adjoint operators; see also [
24],
Section 5. When
and
are separated and
are small perturbations, solutions of (
15) are constructed in [
23], Theorems 7.4, 7.6 and 7.7. In the analysis of resonances between scattering channels, the situation where
also arises [
32].
Conditions for which the equation
(mod
) is valid almost everywhere are given in [
23], Theorem 6.1, for the spectral shift function
ξ with respect to the pair
and the spectral shift function
with respect to the self-adjoint pair
,
. Actually, the almost sure decomposition
can be deduced from [
23], Lemma 7.10, and Equation (
28) below, where the distinguished Birman–Solomyak representation is chosen for Krein’s spectral shift function by employing double operator integrals.
4. Double Operator Integrals
As mentioned in Example 5 above, if
and
are self-adjoint operators,
and
, then for each
, the function
,
, is
-integrable in
, and
is the unique strong solution of Equation (
5). Because
B is self-adjoint, we can rewrite the solution
X as an iterated integral:
with respect to the spectral measures
,
associated with
A and
B.
An application of the Fubini strategy sees the expression:
as a representation of the strong solution of the operator equation:
in the case that both
A and
B are self-adjoint operators.
Integrals like (16) have been studied extensively in the case that
is a Hilbert–Schmidt operator and, more generally, when
Y belongs to the Schatten ideal
in
for some
, where they are called double operator integrals [
20].
Following [
33], Section III.2, a subspace
of the collection
of all bounded linear operators on a separable Hilbert space
is called a symmetrically-normed ideal with norm
if
is a Banach space and:
for , , we have and ;
if S has rank one, then ; and
the closed unit ball of is sequentially closed in the weak operator topology of , that is if with and if in the weak operator topology of , then and .
For
, the Schatten ideal
consists of all compact operators
T whose singular values
belong to
with the norm
for
and
. The singular values
are the eigenvalues of the positive operator
. For
,
is a symmetrically-normed ideal. Condition c. is often called the non-commutative Fatou lemma. It fails for the compact operators
, but
is itself a symmetrically-normed (improper) ideal with the uniform norm. The symmetrically-normed ideal
of trace class operators on
may be identified with the projective tensor product
([
30], III.7.1).
For a bounded linear operator
T on a Hilbert space
, the expression:
is a double operator integral if
E is an
-valued spectral measure on the measurable space
and
F is an
-valued spectral measure on the measurable space
. The function
is taken to be uniformly bounded on
. In Formula (
16),
, so that
is bounded by
for
when the spectra
and
are a positive distance
δ apart.
The map
,
, is continuous into the space
of Hilbert–Schmidt operators and:
so that the map
,
, is actually a spectral measure acting on
, and the equality:
holds for all bounded measurable functions
([
20], Section 3.1).
The situation is more complicated if the space
of Hilbert–Schmidt operators (with the Hilbert–Schmidt norm) is replaced by the Schatten ideal
in
for some
not equal to two or as in the case of Formula (
16), by
itself, because the map
,
,
, only defines a finitely-additive set function
acting on elements
, so that:
For a bounded function
, the double operator integral
may be viewed as a continuous generalisation of a classical Schur multiplier:
for an infinite matrix
, with respect to the matrix units
corresponding to an orthonormal basis
of
. If
denotes the orthogonal projection onto the linear space span
for each
, then:
for the operators
acting on the infinite matrix
for
.
To be more precise, let
be a symmetrically-normed ideal in
. The linear map
is defined by
for
and
. In the language of [
20],
Section 4, the element
of
is the transformer on
associated with
. The tensor product
is defined by completion with respect to the norm (9).
Definition 7. Let and be measurable spaces and a separable Hilbert space. Let be an operator valued measure for the strong operator topology and be a -valued measure.
An
-measurable function
is said to be
-integrable in
if for every
, the function
φ is integrable with respect to the scalar measure
and for every
, there exists
, such that:
for every
.
If
φ is
-integrable in
and
is the multiplication map, then:
The following observation is useful for treating double operator integrals.
Proposition 6. Let ,
and be as in Definition 7. If ,
then there exists a unique vector measure:such that for each and .
Consequently, every bounded -measurable function is -integrable in and: Proof. If
T is a trace class operator on
, then there exists orthonormal sets
,
and a summable sequence
of scalars, such that
for every
. For each
, the total variation of the product measure:
is bounded by
for every
. Here,
and
denote the semi-variation of
m and
n, respectively ([
31], p. 2). It follows that
admits a unique countably-additive extension
whose semi-variation with respect to the norm (9) is bounded by
and
converges in
uniformly on
. ☐
Corollary 7. Let and be measurable spaces and a separable Hilbert space. Let and be operator valued measures for the strong operator topology. Then, there exists a unique operator valued measure:such that: Proof. It is easy to check that for
and
, the formula:
defines a linear operator
on
whose operator norm is bounded by
, and
,
, is countably additive in
for the strong operator topology for each
. ☐
Given , the expression , and , is the restriction to product sets of the -valued measure .
The following notation gives an interpretation of Formula (
16) in the case that the operator
Y belongs to the symmetrically-normed ideal
,
or
. The collection
of trace class operators is a linear subspace of
in each case.
Let
be the finitely-additive set function defined by:
that is
is finitely additive on the algebra
of all finite unions of product sets
for
,
.
Suppose that the function
is integrable with respect to the measure
with values in
. If for
and
, the linear map:
is the restriction to
of a continuous linear map
, then we write:
for for the continuous linear map
and we say that
φ is
-integrable if:
for every
and
.
To check that the operator is uniquely defined, observe that is norm dense in for . In the case , the closure in the ultra-weak topology can be taken.
Although
is only a finitely-additive set function, the
-valued set function:
of an
-integrable function
φ defines a finitely-additive
-valued set function on the algebra generated by all product sets
for
and
.
Corollary 7 tells us that for an
-integrable function
, the
-valued set function:
is countably additive in the strong operator topology for each
. The following simple observation describes the situation for other operator ideals
.
Proposition 8. Suppose that is an -integrable function. For each ,
the set function:is separately σ-additive in the strong operator topology of ,
that is,
for all pairwise disjoint ,
and all pairwise disjoint ,
.
The following result was proven by Birman and Solomyak ([
20], Section 3.1).
Theorem 9. Let and be measurable spaces and a separable Hilbert space. Let and be spectral measures. Then, there exists a unique spectral measure ,
such that for all and:for every bounded -measurable function .
Moreover, For spectral measures
P and
Q, the formula:
holds for each
,
and
, so it is only necessary to verify that
in order to show that
φ is
-integrable.
The following observation gives an interpretation of Formula (
16) as a double operator integral. The Fourier transform of
is the function
defined by
for
.
Theorem 10. Let be a separable Hilbert space. Let and be spectral measures on .
Let for some or .
Suppose that and for all .
Then,
and: Proof. For
and Borel subsets
of
, by Fubini’s theorem, we have:
The right-hand side is a Bochner integral in the strong operator topology of
because:
are continuous unitary groups in the strong operator topology. Moreover,
is continuous in the norm of
for compact subsets
of
, because
is a symmetrically-normed ideal in
; so, the Bochner integral converges in
itself, and we obtain:
For increasing to , the inclusion and the bound (18) is now a consequence of the non-commutative Fatou lemma. ☐
Corollary 11. Let be a separable Hilbert space, and let be self-adjoint operators with spectral measures and ,
respectively. Let for some or .
If the spectra of A and B are separated by a distance ,
then and: In particular, Equation (5) has a unique strong solution for given by the double operator integral:
so that .
Although the Heaviside function
is not the Fourier transform of an
-function, the following result of Gohberg and Krein ([
34], Section III.6) holds, in case
. The general case is outlined in [
20], Theorem 7.2.
Theorem 12. Let be a separable Hilbert space. Let and be spectral measures on .
Then:for every .
The following recent result of Sukochev and Potapov [
35] settled a long outstanding conjecture of Krein for the index
p in the range
.
Theorem 13. Let be a separable Hilbert space. Let and be spectral measures on .
Suppose that is a continuous function for which the difference quotient:is uniformly bounded. Then, for every ,
and there exists ,
such that: Such a function f is said to be uniformly Lipschitz on and .
Corollary 14. Suppose that is a uniformly Lipschitz function. Then, for every ,
there exists ,
such that:for any self-adjoint operators A and B on a separable Hilbert space .
Proof. Let
and
be the spectral measures of
A and
B, respectively, and suppose that
. Then, according to [
20], Theorem 8.1 (see also [
21], Corollary 7.2), the equality:
holds, and the norm estimate follows from Theorem 13. ☐
5. Traces of Double Operator Integrals
In this section, let and be given measurable spaces, a separable Hilbert space and , spectral measures. Let for some or . The Banach space of P-integrable functions is isomorphic to the C*-algebra of P-essentially bounded functions. The analogous result for -integrable functions follows.
Proposition 15. For an -measurable function ,
let be the equivalence class of all functions equal to φ -almost everywhere. Let:with the pointwise operations of addition and scalar multiplication with the norm: Then,
,
and is a commutative Banach *-algebra under pointwise multiplication. If ,
then:is a commutative -algebra. Furthermore, the Banach *-algebras:are isometric, where is the uniformly-closed subspace of consisting of compact operators on .
Remark 8. The analogy of double operator integrals with multiplier theory in harmonic analysis is fleshed out in [
21], Example 2.13, as follows.
If Λ is a locally-compact abelian group with Fourier transform
, the spectral measure
Q is defined by multiplication by characteristic functions on
and
is the spectral measure of the “momentum operator” on Λ, then for
, the space
of Fourier multipliers on
coincides with the commutative Banach *-algebra
for the finitely-additive set function
defined as in [
21], Example 2.13, by the spectral measure
P acting on on
. For example, when
, the operator
is the Hilbert transform for
.
It is only in the case
that
. One might argue that multiplier theory in commutative harmonic analysis is devoted to the study of the commutative Banach *-algebra
for
. The analysis of the commutative Banach *-algebra
for general spectral measures
E and
F and symmetric operator ideal
has many applications to scattering theory and quantum physics [
20].
The commutative Banach *-algebra
is characterised by a result of Peller [
25].
Theorem 16. Let be a uniformly-bounded function. Then, if and only if there exist a finite measure space and measurable functions and ,
such that and: The norm of with the representation (19) satisfies:for Grothendieck’s constant .
Moreover,
is the infimum of all numbers on the right-hand side of the inequality (20) for which there exists a finite measure ν, such that the representation (19) holds for φ.
Formula (19) is to be interpreted in the sense that φ is a special representative of the equivalence class . It is worthwhile to make a few remarks on the significance of Formula (19) in order to motivate its proof below.
If the functions
α and
β in the representation (19) have the property that
,
and
,
, are strongly
ν-measurable in
and
, respectively, then the function
,
, is strongly measurable in the projective tensor product
, and:
Hence, the function
,
, is Bochner integrable in
, that is
. However, it is only assumed
α is
-measurable and
β is
-measurable, so this conclusion is unavailable.
Let
be a finite measure, such that
for
and
for all
with
. Such a measure exists by the Bartle–Dunford–Schwartz Theorem ([
31], Corollary I.2.6) or, more simply,
for some orthonormal basis
of
. Let
be a finite measure corresponding to
Q. Then,
and
.
There is a bijective correspondence between elements
of the projective tensor product
and nuclear operators
, such that for each
,
for
almost all
, in the sense that for functions with:
the kernel
corresponds to the nuclear operator:
Nuclear operators between Banach space are discussed in [
30], Section III.7.
In the case that
and
are projections onto the standard basis vectors, then
is the classical Schur multiplier operator (17) and Grothendieck’s inequality ensures that
; see Proposition 18 below and [
26], Theorem 3.2. In this case, the measure
ν in Formula (
19) is the counting measure on
, and there is no difficulty with strong
ν-measurability in an
-space.
The passage from the discrete case to the case of general spectral measures P and Q sees the nuclear operators from to replaced by one-integral operators from to , which leads to the Peller representation (19).
5.1. Schur Multipliers and Grothendieck’s Inequality
If
E is any
-valued spectral measure and
, the identity:
ensures that the
-valued measure
has bounded
-semi-variation in
, the Hilbert space tensor product
with norm
. It follows from [
5] that for any essentially bounded functions
and
and
, the
-valued function
f is
-integrable in
, and the
-valued function
g is
-integrable in
. Then, there exist operator-valued measures
and
, such that:
There is a simple sufficient condition for
Observe first that the linear map
defined by:
has a continuous linear extension to a contraction
corresponding to taking the trace in the discrete index. The formula:
for
,
and
defines a finitely-additive set function:
with values in
, because
can be identified with
for any Hilbert space
. Moreover,
Then, the operator:
is an element of
, that is
belongs to
, and we have the representation:
where
and
. Moreover, the bound:
holds. The same argument works if the spectral measures
P,
Q are replaced by any two operator valued measures
and
by appealing to the metric form ([
26], Theorem 2.4) of Grothendieck’s inequality, so that:
Alternatively, for each
, the linear operator:
can be realised as the operator associated with the bounded sesquilinear form:
See ([
20], Theorem 4.1).
A remarkable consequence of Grothendick’s inequality is that for
, Peller’s representation (19) is necessary
almost everywhere. The analysis of Pisier [
26] leads the way.
The projective tensor product
is the completion of the tensor product
with respect to the norm:
Another distinguished norm on
is given by:
where the infimum runs over all possible representation
for
,
and
. Then,
may also be viewed as the norm of factorisation through a Hilbert space:
where the infimum runs over all Hilbert spaces
and all
for which
has the finite representation
with respect to the standard basis
of
. Another way of viewing
is:
over representations
,
, because:
Proposition 17. Let be a function that defines a Schur multiplier ,
that is in matrix notation .
The following conditions are equivalent.- (i)
.
- (ii)
There exists a Hilbert space and functions , with values in the closed unit ball of , such that , .
- (iii)
For all finite subsets of ,
the bound:
holds.
Proof. Suppose first that
φ is zero off a finite set
. Then, the bound (i) is equivalent to the condition that:
for all linear maps
with norm
and matrix
with respect to the standard basis and all
,
, that is
φ belongs to the polar
of the set
of all matrices
with
a,
α,
β as described. According to [
26], Remark 23.4, the set
is itself the polar
of the set
of all matrices
with
. Then (i) holds if and only if
φ belongs to
, which is exactly Condition (iii). Conditions (ii) and (iii) are equivalent by the the definition of the norm
. The passage to all of
follows from a compactness argument. ☐
Remark 9. (a) The argument above uses the factorisation of the norm
dual to
described in [
26], Proposition 3.3 and Remark 23.4; this only relies on the Hahn–Banach Theorem.
(b) The representation (21) is the measure space version of the implication (ii) ⇒ (i) above. The necessity of the condition (21) in the general measure space setting is proven using complete boundedness arguments in [
36], Theorem 3.3; see also [
37,
38].
One version of Grothendieck’s inequality from [
26] is that the norm
and the projective tensor product norm are equivalent on
with:
The constant is Grothendieck’s constant. The projective tensor product version of Proposition 17 follows, with the same notation.
Proposition 18. Let be finite subsets of ,
and let be a function vanishing off .
Then: Passing to infinite sets, a bounded function
with
necessarily has a representation:
with
, as in Peller’s representation (
19).
5.2. Schur Multipliers on Measure Spaces
We first note that for any choice of finite measures
,
equivalent to
P and
Q, respectively, the Banach algebra
is isometrically isomorphic to the set of multipliers of the projective tensor product
, that is
if and only if for every
, the function
is equal
-a.e. to an element of
and
is equal to the norm of the linear map:
If and are another pair of such equivalent measures, then the operator of multiplication by is a unitary map from to and similarly for , so that multiplication by is an isometric isomorphism from the space onto .
Proposition 19. Let ,
be finite measures equivalent to the spectral measures P,
Q,
respectively. Then,
is isometrically isomorphic to the set of multipliers of the projective tensor product and the identity:holds.
Proof. Let be a sequence of vectors in with , such that is an orthogonal set of vectors in for each . Such a sequence of vectors can always be manufactured by taking any vectors with and for a measure equivalent to P, the sets where . Then, , , will do the job. Let be the corresponding vectors for Q.
As noted above, the norm of
is invariant under a change of equivalent measures, so we may as well assume that:
so that the mappings
,
, and
,
, define a unitary equivalences
,
between
and
and
, respectively.
The map
with integral kernel
is the trace class. Let
be the corresponding trace class operator
. Then:
Let
. Then,
and:
It follows that
is the kernel of the trace class operator
, such that:
the equality
holds and:
According to the identities above,
for all
, so if:
for all
satisfying
, then
, the identity (24) holds and:
The equality (23) follow from the identities:
for
,
and the unitary equivalence
defined above. The analogous identities hold for the spectral measure
Q. ☐
Proof of Theorem 16 We proceed by reduction to the -case considered in Proposition 18. A Lusin μ-filtration of a σ-finite measure space is an increasing family of σ-algebras, such that for a set of full measure, , and each element of is the countable union of sets belonging to a countable partition of into sets of finite positive μ-measure and such that each set in is contained in an element of , for .
Let , be finite measures equivalent to P, Q, respectively. Because both and are isomorphic to the separable Hilbert space , for the purpose of obtaining the representation (19), we may suppose that the underlying σ-algebras are countably generated.
Let
be a Lusin
-filtration, and let
be a Lusin
-filtration. Suppose that
,
is the
n-th partition associated with
and
is the
n-th partition associated with
. The corresponding projection operators
and
are defined by:
The conditional expectation is defined for any measurable function that is integrable over any set , .
It is easy to verify that
for the matrix:
and the operator
with kernel
.
Moreover, for every finite rank operator
, the bound:
Suppose that there exists , such that for all
Then,
tr
, and taking
, the martingale convergence theorem shows that the bound:
holds for every finite rank operator
. It follows from [
39], Theorem 6.16, that
belongs to the Banach ideal
of one-integral operators from
to
. Because
is a dual space, [
39], Corollary 5.4, ensures that
enjoys the factorisation:
for some bounded linear operators
T1 and
T2 and finite measure space (
T;
;
ν). The given factorisation also follows by the original 1954 Grothendieck argument with the choice
,
in [
30], Section IV.9.2.
Every bounded linear operator
u from
to
is an integral operator with a bounded kernel, because
defines a continuous linear functional on
(see [
40], Lemma 2.2, for a compactness argument), so there exist bounded measurable functions
and
, such that:
The representation (19) and the associated bounds follow if we can take:
We know from the bounds (18) that:
The norm
defined on
is the norm of factorisation through a Hilbert space. For any bounded linear operator
between Banach spaces
X and
Y,
where the infimum runs over all Hilbert spaces
and all possible factorisations:
of
through
with
. Taking
and
, the bound (22) says that:
with respect to the completion
of
in the norm
,
The norm estimates:
follow from the definition of
and the contractivity of the conditional expectation operators
.
According to Proposition 19, the norm of the linear operator:
associated with multiplication by
φ on
is equal to:
The equality
is proven in [
36], Theorem 3.3, using complete boundedness techniques, but this can be established in a more elementary way by noting that if
, then the martingale convergence theorem ensures that
in the strong operator topology of:
as
and also:
as
. Then,
by duality. The equality:
follows for each
from Proposition 17 by replacing
in (iii) by
for
. The final assertion of Theorem 16 follows from the equalities:
☐
Remark 10. (a) The original proof of Peller [
25,
40], Theorem 2.2, factorises the finite rank operator
instead, so the constant
appears in place of
in the bound associated with (19).
(b) Let
be the closure of the linear space of all
in the uniform norm of the space of operators
corresponding to the compact linear operators from
to
. By [
30], Section IV.9.2, the function
in Formula (
19) is
ν-integrable in the space of one-integral operators:
and
.
(c) The proof above shows that operator
is (strictly) one-integral in the sense of [
39], p. 97, and [
30], Section IV.9.2, if and only if
. The reason that we may have
for some
, that is the function
associated with the representation (19) fails to be
ν-integrable in
, so that
thereby fails to be a nuclear operator, is that the vector measure
associated with a continuous linear map
u from
to
has a weak*-density, but not necessarily a strongly-measurable density in
.
For any
and
, the operator:
is the trace class. Moreover, the expression
,
, is a complex measure
on the
σ-algebra
, such that
. As indicated in [
20], Section 9.1, the identity:
holds. In the case that
is a finite rank operator, together with the polarisation, the bound (23) shows that the operator
with integral kernel
φ is the trace class and:
The same bound holds for all
. The identity:
ensures that
for
with
and
bounded on Λ. Then, the equality:
holds, because both sides are continuous for
.
The representation (21) converges in
, and there exists a set
of full
-measure, such that:
for all
, where the right-hand sum converges absolutely. The expression above constitutes a distinguished element of the equivalence class
. Consequently, Formula (25) is valid because:
As in the proof of Theorem 16, for any Lusin
-filtration
of Λ, for each
, the conditional expectation operators
with respect to the
σ-algebra
and the finite measure
have the property that:
by the martingale convergence theorem. Consequently, setting:
wherever the limit exists, the equality
holds for
-almost all
.
Remark 11. There is a representative function
φ of the equivalence class
that is continuous for the so-called
ω-topology of [
37], Proposition 9.1, so Formula (25) may also be derived from the trace formula for a trace class operator with a continuous integral kernel. In fact, Peller’s representation (19) can be deduced directly from Proposition 18 by employing the
ω-continuity of
φ rather than the martingale convergence theorem; see [
37], Remark p. 139.
6. The Spectral Shift Function
The following perturbation formula of Birman and Solomyak ([
20], Theorem 8.1) was mentioned in the proof of Corollary 14. The operator ideal
is taken to be
for
or
for a given Hilbert space
.
Theorem 20. Let be a separable Hilbert space, and let A and B be self-adjoint operators with the same domain, such that .
Let and be the spectral measures on associated with A and B, respectively. Suppose that is a continuous function for which the difference quotient:is uniformly bounded and .
Then:and: If
, then we would like to calculate the trace of
. The method of the preceding section is unavailable with different spectral measures
,
, so we can try to invoke the Daletskii–Krein formula ([
20], Equations (9) and (10)). For a sufficiently smooth function
f, this takes the form:
with
,
and
,
. At each point
, the same spectral measure
is involved, so from Formula (25), we can expect that:
for the complex measure
,
, with
. It turns out that is absolutely continuous with respect to the Lebesgue measure on
from which the formula:
is obtained. The function
is Krein’s spectral shift function.
We now turn to establishing the validity of Formula (26) for a restricted class of functions
f. Better results are known, for example, from [
25,
41,
42,
43], but our purpose is to describe applications of singular bilinear integrals, such as double operator integrals to problems in the perturbation theory of linear operators. The approach of Boyadzhiev [
44] best suits the purpose.
Setting
, we first note that
for each
and
, because the perturbation series:
converges in the norm of
and
is norm differentiable in
. Moreover,
The following result is straightforward, but it depends on some measure theoretic facts. It establishes that Ξ is a complex measure.
Lemma 21. The function ,
,
is strongly measurable in for each and .
There exists a unique operator-valued measure ,
σ-additive for the strong operator topology, such that the equality:
holds for each
. For each
, the set function
,
, is a complex measure, and we have:
Proof. If
is the Fourier transform of a finite measure
μ, then:
as a Bochner integral and by dominated convergence
,
, is continuous in
for each
. By a monotone class argument,
,
, is strongly measurable for all bounded Borel measurable functions
f.
For each
, the set function
is nonnegative and finitely additive, and the algebra
is generated by product sets
for
and
, so
,
. The set function
is separately countably additive with respect to Borel sets, so it is inner regular with respect to compact product sets and, so, countably additive (countable additivity may fail without inner-regularity; see [
45]).
Denoting the extended measure by the same symbol, for all . The -valued measure is weakly countable additive by polarity and, so, norm countably additive by the Orlicz–Pettis theorem.
For each
and orthonormal basis
of
, the bound:
holds and:
by the Beppo–Levi convergence theorem, because:
so Equation (28) holds. ☐
An application of Fubini’s Theorem for disintegrations of measures shows that:
for each
. The identity:
follows for each
by polarisation. Because:
for any orthonormal basis
of
, the Fourier transform of the measure
Ξ is:
We need to establish that the inverse Fourier transform
of the uniformly bounded, continuous function:
belongs to
. Then,
is the spectral shift function. Clearly, the value of Φ at zero is irrelevant.
It suffices to show that there exists
, such that:
with
and
,
for every finite positive measure
μ, because then
as elements of the space
of Schwartz distributions on
. Therefore, we consider the class of functions
for which
and
and, consequently,
.
Theorem 22. Let be a separable Hilbert space, and let A and B be self-adjoint operators with the same domain, such that .
Then, there exists a function ,
such that:
for every function
for which there exists a finite positive Borel measure
μ on
, such that:
- (a)
.
- (b)
.
- (c)
If , then a.e.
- (d)
ξ is zero a.e. outside of the interval .
Proof. The proof is set out in considerable detail in [
44]. Here, we review the salient points.
The estimate
follows from the bound (27) and the calculation:
obtained from an application of Fubini’s theorem with respect to
and
on
for
. Then:
An expression for the spectral shift function
ξ may be obtained from Fatou’s theorem ([
46], Theorem 11.24). Suppose that
ν is a finite measure on
and:
is the Cauchy transform of
ν. Then,
ν is absolutely continuous if:
The jump function
defined for almost all
is then the density of
ν with respect to the Lebesgue measure. For
, if the representation:
were valid, then we would expect that
has the representation:
where the arctan function may be expressed as:
In the case that
for
and
,
, a calculation, given explicitly in [
44], shows that the function:
is harmonic and uniformly bounded in the upper half-plane. By Fatou’s theorem ([
46], Theorem 11.23), the boundary values
are defined for almost all
and satisfy:
for every
, so in the case that
has rank one, Formula (
30) is valid.
For an arbitrary self-adjoint perturbation:
with
, the function
may be defined in a similar fashion for
,
, so that
in
as
from which it verified that
. ☐
The representation
obtained above may be viewed as the Fourier transform approach. In the case of a rank one perturbation
, the Cauchy transform approach is developed by Simon [
47] with the formula:
for
for some
, established in [
47], Theorem 1.9, by computing a contour integral. Here, the boundary value
is expressed as:
for almost all
with respect to the Cauchy transform:
The Cauchy transform approach is generalised to Type II von Neumann algebras in [
41].
Many different proofs of Krein’s Formula (29) are available for a wide class of functions
f, especially in a form that translates into the setting of non-commutative integration [
41,
42,
43]. As remarked in [
20], p. 163, an ingredient additional to double operator integrals (such as complex function theory) is needed to show that the measure is absolutely continuous with respect to the Lebesgue measure on
. Krein’s original argument uses perturbation determinants from which follows the representation Det
for the scattering matrix
for
A and
B ([
22], Chapter 8).