ISSN 0081-5438, Proceedings of the Steklov Institute of Mathematics, 2014, Vol. 285, pp. 222–232.


c Pleiades Publishing, Ltd., 2014.
Original Russian Text c Yu.N. Orlov, V.Zh. Sakbaev, O.G. Smolyanov, 2014, published in Trudy Matematicheskogo Instituta imeni
V.A. Steklova, 2014, Vol. 285, pp. 232–243.

Feynman Formulas as a Method

of Averaging Random Hamiltonians
Yu. N. Orlov a , V. Zh. Sakbaev b , and O. G. Smolyanov c
Received February 2014

Abstract—We propose a method for finding the mathematical expectation of random un-
bounded operators in a Hilbert space. The method is based on averaging random one-parameter
semigroups by means of the Feynman–Chernoff formula. We also consider an application of
this method to the description of various operations that assign quantum Hamiltonians to the
classical Hamilton functions.
DOI: 10.1134/S0081543814040154

INTRODUCTION
Feynman formulas give a representation of the Schrödinger semigroup exp(−tH), t ≥ 0, or
Schrödinger group exp(itH), t ∈ R, in terms of the limits of finitely multiple integrals over the
Cartesian powers of the configuration or phase space (as the multiplicity tends to infinity) of a
classical Hamiltonian system whose quantization yields the Hamilton operator H (this means that
H is a self-adjoint operator corresponding to the Hamilton function H of the classical system; in
particular, H may be a pseudodifferential operator whose symbol is the Hamilton function H).
When one uses integrals over the products of the configuration space, one speaks of Lagrangian
Feynman formulas; and when integrals over the products of the phase space are used, one speaks
of Hamiltonian Feynman formulas.
In this paper, we introduce a new method of averaging random one-parameter semigroups and
their generators by means of Feynman formulas with the use of Chernoff’s theorem [9]. By a random
semigroup we will mean a vector-valued random variable G with values in the set of one-parameter
semigroups of operators acting in some Banach space. If (E, A, μ) is a probability space on which
a random semigroup G is defined, then the generator of the random semigroup G is a random
variable HG on the same probability space that is defined by the following condition: for every
ε ∈ E, the value HG (ε) of the random variable HG is a generator of the semigroup G(ε). In
this paper, we show that the mathematical expectation of a random semigroup G is an operator-
valued function FG that is Chernoff equivalent (see [11]) to some semigroup UG . It is natural to
think of the generator of UG as the mathematical expectation of the random generator HG of the
random semigroup G. Chernoff equivalence means that the sequence of operator-valued functions
{(FG (t/n))n , t ≥ 0} converges in the strong operator topology to the semigroup UG (t), t ≥ 0,
uniformly on any interval [0, T ] of the half-axis t ≥ 0.
This definition of averaging of generators is an extension of the averaging procedure in the
space of operators, because when all values of a random generator are bounded, its mean value
a Keldysh Institute of Applied Mathematics, Russian Academy of Sciences, Miusskaya pl. 4, Moscow, 125047 Russia.
b Moscow Institute of Physics and Technology (State University), Institutskii per. 9, Dolgoprudnyi, Moscow oblast,
141700 Russia.
c Faculty of Mechanics and Mathematics, Moscow State University, Moscow, 119991 Russia.

E-mail addresses: yuno@kiam.ru (Yu.N. Orlov), fumi2003@mail.ru (V.Zh. Sakbaev),

smolyanov@yandex.ru (O.G. Smolyanov).

222
 FEYNMAN FORMULAS AS A METHOD OF AVERAGING 223

coincides with the ordinary mean value of elements of a Banach space (see Remark 2 to Theorem 2
below). When the values of a random generator are unbounded operators, the natural notion of the
mathematical expectation of the random variable HG may not be well defined, but the definition
introduced here is applicable (see Examples 1 and 2).
The proposed method of averaging random semigroup generators that are unbounded operators
is based on the study of averaging the exponential function of unbounded operators that takes values
in the space of bounded linear operators. The method is analogous to the method of summation
of unbounded operators by considering the sums of their quadratic forms (see [3]). In both cases,
instead of averaged operators themselves, one analyzes the values of some operator-valued function
on these operators. Of course, the method proposed in this paper is much more complicated due
to the application of Chernoff iterations and the notion of Chernoff equivalence of operator-valued
functions, which was introduced in [11].
One of the reasons for ambiguity in the quantum dynamics of a system (see [2]) is the am-
biguity in the choice of the classical Hamiltonian corresponding to a given law of motion of the
system in the coordinate space. For every law of motion of a classical system in the coordinate
space, there is a set of Hamilton functions to which this law of motion in the coordinate space
corresponds (Hamilton functions that define the same law of motion in the coordinate space are
called equivalent in [2]). Every classical Hamiltonian can be assigned a quantum Hamiltonian, but
the quantum Hamiltonians corresponding to equivalent (in the above sense) classical Hamiltonians
are different in the sense that the unitary transformation groups generated by these Hamiltonians
are different (see [2]). Now if some probability measure is defined on the set of equivalent classical
Hamiltonians (and hence on the set of the corresponding quantum Hamiltonians), then random
variables with values in the set of classical (quantum) Hamiltonians and with values in the set
of unitary groups arise. Another (possibly, the main) reason for the ambiguity in quantization is
the noncommutativity of the coordinate and momentum operators (see [1]). The relation between
this ambiguity and the averaging operation introduced above will be discussed in greater detail
below.
Chernoff’s theorem is one of the most general statements on the approximation of one-parameter
semigroups by operator-valued functions; it can be regarded as a far-reaching generalization of the
Euler theorem on the approximation of a solution to a differential equation by polygonal chains to
the infinite-dimensional case. In particular, this theorem is used for constructing approximations
of semigroups by Feynman formulas and approximations of equations with variable coefficients by
equations with piecewise constant coefficients.
One of the first applications of Chernoff’s theorem is (see [10, 6]) the derivation of the rep-
resentation of the unitary group generated by a Hamiltonian by means of the Feynman formula.
Namely, every classical Hamiltonian H is assigned a quantum Hamiltonian H  = H, which is a
pseudodifferential operator in the Hilbert space H = L2 (R) with symbol H. In this case, the
 t ∈ R, which
numerical function exp(itH), t ∈ R, is assigned an operator-valued function exp(itH),
is not a semigroup but, according to Chernoff’s theorem, is equivalent in the sense of the definition
from [11] to the unitary group exp(itH), t ∈ R.
In the present study, we apply a similar procedure in order to construct the mean value of a
random Hamiltonian. We assign the mean value in the sense defined above to every probability
measure on the set of quantum Hamiltonians corresponding to the same classical Hamilton function.
Thus, we define another quantization method, i.e., a method for associating an operator in the
Hilbert space with a classical Hamilton function. A similar situation occurs when the same Hamilton
function is considered as the τ -symbol (see [4]) of a family of pseudodifferential operators indexed
by the variable τ from the interval [0, 1]. In our case, the role of the parameter τ is played by a
measure on a set of operators in a Hilbert space.

PROCEEDINGS OF THE STEKLOV INSTITUTE OF MATHEMATICS Vol. 285 2014

224 Yu.N. ORLOV et al.

Here the following question remains unstudied and is worth examining: What is the relationship
 t ∈ R, the mean value of the ran-
between the mean value of the random operator function exp(itH),

dom group exp(itH), t ∈ R, and the operator function whose symbol is the mean value of the random
function exp(itH), t ∈ R (we conjecture that all three operator functions are Chernoff equivalent).

1. MAIN DEFINITIONS
A random variable is a measurable mapping ξ from a measure space (E, A, μ) to some Banach
space or to a linear topological space Z. In this context we will consider random semigroups; in
this case, the space Z is the Banach space Cs (R+ , B(X)) of strongly continuous mappings from the
half-axis R+ to the Banach space of linear bounded operators in a Banach space X, and the values
of the random variable ξ are one-parameter semigroups.
In particular, we consider the case when the algebra A is the maximal algebra 2E of all subsets of
the space E. If the measurable space E is a Hausdorff topological space and ε0 ∈ E, then we denote
by W (E, ε0 ) the class of all nonnegative normed finitely additive measures on the measurable space
(E, 2E ) that are concentrated in an arbitrary punctured neighborhood of the point ε0 in the sense
that the value of any measure in the class W (E, ε0 ) is equal to one on any punctured neighborhood
of the point ε0 (see [8, 7, 13]). We establish a relation between generalized limits for ε → ε0 of the
random semigroup G as a mapping G : E → Cs (R+ , B(X)) and the mathematical expectations of
the random semigroup G on the probability space (E, 2E , μ) with measures μ in the class W (E, ε0 ).
Random semigroups arose in [8] as a result of regularization that consists in approximating a
maximal symmetric operator which is not a generator of a semigroup by sequences of generators.
The ambiguity in the choice of approximations and the absence of convergence of the sequence of
regularized semigroups lead in this case to the necessity of studying random semigroups.
The source of randomness of the semigroups related to a quantum system is the ambiguity in
the choice of the classical Hamiltonian describing a family of trajectories in the classical coordinate
space (see [2]). By a classical Hamiltonian we mean a continuously differentiable function H on
the phase space Q × P of a system of classical mechanics. A classical Hamiltonian is said to be
nondegenerate if the function H is twice continuously differentiable and its Hessian is uniformly
bounded away from zero.
We say that two classical Hamiltonians are equivalent if the sets of trajectories corresponding
to these Hamiltonians in the coordinate space (i.e., the projections of the trajectories in the phase
space Q × P onto the coordinate space Q) coincide. In [2], it was shown that the class of equivalent
nondegenerate Hamiltonians is fairly wide and that any two representatives of a class of equivalent
Hamiltonians generate identical families of coordinate trajectories but different families of phase
trajectories. Moreover, different equivalent Hamiltonians generate different quantum dynamical
semigroups, i.e., different groups of unitary transformations of the Hilbert space H = L2 (Q).
Then the quantization of classical systems as a mapping from the set of classical Hamiltonians
to the set of quantum Hamiltonians (self-adjoint operators in the Hilbert space L2 (Q)) and then
to the set of unitary groups generated by quantum Hamiltonians is a random semigroup provided
that the set K of equivalent classical Hamiltonians with the algebra 2K of all subsets is equipped
with some nonnegative normed measure.

2. RANDOM SEMIGROUPS ASSOCIATED WITH THE CAUCHY PROBLEM

Let Z = Cs (R+ , B(X)) (or Z = Cw (R+ , B(X))) be the space of strongly (or weakly) continuous
mappings from the half-axis R+ to the Banach space B of linear bounded transformations of a
Banach space X (that has a preadjoint space X∗ ). Consider a Z-valued random variable defined
on a measurable space (E, 2E , μ) with a measure μ ∈ W (E); i.e., consider a random process with
values in B(X).

PROCEEDINGS OF THE STEKLOV INSTITUTE OF MATHEMATICS Vol. 285 2014

 FEYNMAN FORMULAS AS A METHOD OF AVERAGING 225

As the set E, we can take the set of all generators of strongly continuous semigroups G(X)
acting in the space X and equip this set with the topology τ of strong (or weak) graph convergence.
Let ε0 ∈ G(X) be a limit point of the set G(X) in the topology τ . Denote by W (E, ε0 ) the set of all
measures μ ∈ ba(E, 2E ) (i.e., elements of the Banach space of finitely additive measures of bounded
variation defined on the algebra 2E ) that are nonnegative, normed to unity, and concentrated in an
arbitrary punctured neighborhood of the point ε0 of the set E.
On the space Z, we define a family of functionals that act on an arbitrary element z ∈ Z
according to the rule ϕt,A,g (z) = z(t)A, g, t ∈ R+ , A ∈ X, g ∈ X∗ . Consider the topology τm
on Z generated by the family of functionals ϕt,A,g , t ∈ R+ , A ∈ X, g ∈ X∗ . Then the space Z
equipped with the structure of the algebra of Borel subsets is a measurable space, and the mapping
ξ : E → Z is a random variable.
By the mathematical expectation of a random variable ξ as a mapping from the measure space
(E, 2E , μ) to the topological space Z, we will mean the Pettis integral

M ξ = ξε dμ(ε),
E

where M ξ is an element of the space Z such that the equality


M ξ(t)A, g = ξε (t)A, g dμ(ε) (2.1)
E

holds for all t ∈ R+ , A ∈ X, and g ∈ X∗ . Since the latter integral is the Radon integral of a
numerical function with respect to a finitely additive measure, it is well defined and equality (2.1)
defines a linear bounded transformation M ξ(t) ∈ B(X) for any t ∈ R+ .
Theorem 1. Let μ be a nonnegative normed measure on the algebra 2E of subsets of the set E.
If a family of mappings ξ is uniformly bounded and there exists a linear dense subspace D in X such
that the family of mappings ξε (t)A ∈ C(R+ , X), ε ∈ E, is weakly (strongly) uniformly Lipschitz
continuous for every A ∈ D, then M ξ(t) ∈ Cw (R+ , B(X)) (M ξ(t) ∈ Cs (R+ , B(X))).
Proof. The weak uniform Lipschitz continuity means that for any A ∈ D and g ∈ X∗ there
exists a constant L > 0 such that

sup |ξε (t + Δt)A − ξε (t)A, g| ≤ L|Δt|;

t∈R+ , ε∈E

the strong Lipschitz continuity means that for any A ∈ D there exists a constant L > 0 such that

sup ξε (t + Δt)A − ξε (t)AX ≤ L|Δt|.

t∈R+ , ε∈E

The uniform boundedness of a random semigroup ξ means that supε∈E, t∈R+ ξε (t)B(X) ≤ C for
some C > 0.
Therefore, by the uniform boundedness condition, for any t ≥ 0, A ∈ D, and g ∈ X∗ the
function ξε (t)A, g is bounded on the set E and, hence, integrable with respect to measure μ in the
Radon sense (see [8]), and the integral (2.1), as a function of the argument g,
 is a linear continuous
functional on the space X∗ . Therefore, for any A ∈ D, the Pettis integral G(X) ξε A dμ(ε) ∈ X is

defined, and the mapping A → G(X) ξε A dμ(ε) is linear in A in view of the linearity of the Pettis
integral
 and continuous in view of the uniform boundedness of the mapping ξ. Hence, the mapping
A → G(X) ξε A dμ(ε) can be extended by continuity from the linear subspace D to a linear bounded
transformation of the space X.

PROCEEDINGS OF THE STEKLOV INSTITUTE OF MATHEMATICS Vol. 285 2014

226 Yu.N. ORLOV et al.

Therefore, for every t > 0, the mean value M ξ(t) = G(X) ξε (t) dμ(ε) ∈ B(X) is defined. The
function M ξ(t) of the variable t ∈ R+ is weakly (strongly) continuous and even weakly (strongly)
Lipschitz continuous, because the inequality in the condition of weak (strong) Lipschitz continuity
of the mapping ξ is preserved under integration. Theorem 1 is proved.
Thus, under the hypotheses of Theorem 1, the mathematical expectation of the process ξ is also
an element of the space Cw (R+ , B(X)), which will be called an averaged family of transformations
of the space X.
According to the theorem on the perturbation of the generator of a strongly continuous semi-
group by a bounded operator (see [3, Theorem 9.2.1]), the following proposition is valid.
Proposition 1. If the operator Lε ∈ G(X) is a generator of a contraction semigroup in the
space X for every ε ∈ E, then the operator Lε,V = Lε + V, V ∈ B(X), is also a generator of a
strongly continuous semigroup in the space X; moreover, exp(Lε,V t)B(X) ≤ exp(VB(X) t) for
all t ≥ 0, and the following estimate is valid for any T > 0 :
   
sup exp(Lε,V t) − exp(Lε,V t)B(X) ≤ C T, V B(H) V − V B(X) .
t∈[0,T ]

Remark 1. The mathematical expectation of a random semigroup may not be a semigroup.

For instance, in the example in [12], the mathematical expectation is a one-parameter family of
transformations described by a limit integro-differential equation whose solutions u(t), t ≥ t0 , are
not invariant with respect to the choice of the initial time in the Cauchy problem.
An explicit example of violation of the semigroup property under the averaging of semigroups is
given by the following elementary observation. The numerical functions eit , t ≥ 0, and e−it , t ≥ 0,
are unitary strongly continuous semigroups of transformations of the one-dimensional Banach space
X = R; however, their half-sum (eit + e−it )/2 = cos t is not a semigroup of transformations of the
space X, because cos(t + s) = cos t · cos s − sin t · sin s, while the semigroup equality cos(t + s) =
cos t · cos s fails.
Thus, the mathematical expectation of a random semigroup may be an operator-valued function
that is not a semigroup. However, even operator-valued functions that have lost the semigroup prop-
erty can carry information on some semigroup that is uniquely determined by the operator-valued
function via the same iteration procedure as applied in the construction of Feynman approximations
of semigroups on the basis of Chernoff’s theorem (see [9]).
Following [11], we introduce the following equivalence relation on the set Π of strongly continuous
operator-valued functions F : [0, +∞) → B(X) that satisfy the condition F (0) = I and the condition
F (t)B(X) ≤ 1 + ct, t ∈ [0, δ), for some c ≥ 0 and have a derivative F  (0) at zero whose closure
serves as a generator of a strongly continuous semigroup.
Definition 1. Operator-valued functions F, G ∈ Π that act from some closed right half-
neighborhood of zero on the number axis to the Banach space B(X) of bounded linear operators
on a Banach space X are Chernoff equivalent if the derivatives of these functions at zero coincide
in the essential domain of the generator of a semigroup or, equivalently, if the following condition
holds for every T > 0 and every u ∈ X:
  n   n 
 t t 
lim sup   G − F u
 = 0. (2.2)
n→∞ t∈[0,T ] n n

The equivalence relation introduced covers a fairly wide class Π of operator-valued functions.
It is symmetric, reflexive, and transitive on the class Π. Every function in Π is equivalent to some
semigroup; therefore, the class Π is factorized by this equivalence relation to the set of strongly
continuous semigroups of transformations of the space X.

PROCEEDINGS OF THE STEKLOV INSTITUTE OF MATHEMATICS Vol. 285 2014

 FEYNMAN FORMULAS AS A METHOD OF AVERAGING 227

To cover the set of all strongly continuous operator-valued functions satisfying the condition
F (0) = I by the equivalence relation, we introduce the following definition.
Definition 1 . Operator-valued functions F and G acting from some closed right half-neigh-
borhood of zero on the number axis to the Banach space B(X) of bounded linear operators on a
Banach space X are Chernoff equivalent if condition (2.2) holds for every T > 0 and every u ∈ X.

3. APPLICATIONS OF AVERAGED FAMILIES OF TRANSFORMATIONS

TO THE QUANTIZATION OF HAMILTONIAN SYSTEMS
OF CLASSICAL MECHANICS
Let us study the mathematical expectation of a random semigroup that arises due to the ambi-
guity in the choice of the classical Hamiltonian (see [2]) corresponding to a family of trajectories of
a classical system in the coordinate space.
Let H = H(q, p), (q, p) ∈ R2d , be the Hamiltonian of a system of classical mechanics with the
coordinate space Q = Rd and the phase space R2d = Rd × Rd endowed with the standard symplectic
form. Denote by ΓH the set of trajectories of the Hamiltonian system with Hamiltonian H in the
coordinate space Q.
A classical Hamiltonian is a twice continuously differentiable function on the phase space R2d
that is strictly convex with respect to the variables p ∈ Rd . Following the definition in [2], we say that
a Hamiltonian system on the phase space R2d with a Hamiltonian H  is equivalent to a Hamiltonian
system on the same phase space with a Hamiltonian H if ΓH  = ΓH , i.e., if every trajectory of
the system with the Hamiltonian H  in the space Q is also a trajectory of the system with the
Hamiltonian H, and vice versa. The trajectories of different equivalent Hamiltonian systems with
common initial data in the phase space are different; however, their projections onto the coordinate
space coincide. The introduced relation on the set of Hamiltonians on the space R2d is indeed an
equivalence relation. Examples of equivalent Hamiltonians are given  by the Hamiltonians of free

p) = p2/2 + p4/4, H  (q, p) = 1 + p2 , etc.
motion such as H(q, p) = p2/2, H(q,
Let M be the set of classical Hamiltonians equivalent to a Hamiltonian H. Then the trajectories
in the coordinate space Q that correspond to the Hamiltonians in the set M are indistinguishable,
whereas the trajectories in the phase space are different. However, the quantum semigroups gen-
erated by quantum Hamiltonians corresponding to different equivalent classical Hamiltonians are
different. As stressed in [2], this is another reason (in addition to the ordering of noncommuting
operators) for the ambiguity in quantization of classical Hamiltonian systems, even in the case of
systems whose Hamiltonians depend only on the one variable p.
A random semigroup is a mapping from the set M of equivalent classical Hamiltonians on which
a certain measure ν is defined to the linear space of strongly continuous semigroups that act in the
Hilbert space H = L2 (Rd ).
Let the measure ν be a probability measure concentrated on some countable subset {An } ⊂ M ,
and let
 ∞
S = A(m) dν = An μn ,
M n=1
∞
where μn ≥ 0 and n=1 μn = 1. We analyze the dynamic properties of the mean value of the
random semigroup F(t) = M exp(−itA) dμ(A) = n∈N exp(−itAn )μn , t ≥ 0. As pointed out
above, the mean value of a random semigroup may not be a semigroup. However, according to the
following theorem, the mean value of a random semigroup generates a semigroup that is Chernoff
equivalent to this mean value and whose generator is the mean value of the Hamiltonian.
Theorem 2. Let {An } be a sequence of self-adjoint operators in a Hilbert space H. Let {μn }
be a sequence of nonnegative numbers such that the sum of the series of these numbers is equal to

PROCEEDINGS OF THE STEKLOV INSTITUTE OF MATHEMATICS Vol. 285 2014

228 Yu.N. ORLOV et al.

one. Suppose that there exists a number m ∈ N such that for all n ≥ m + 1 the operators An are
bounded and the series ∞ k=m+1 μk Ak  converges. Suppose also that there exists a linear subspace
D ⊂ H that is an essential domain of each of the operators An , n ∈ 1, m, and Sn = nk=1 μk Ak ,
n ∈ 1, m.
Then the mean value of the random semigroup F(t) = n∈N exp(−itA ∞ n )μ(n), t ≥ 0, is Chernoff
equivalent to the unitary group U(t) = exp(−itS), t ∈ R, where S = k=1 μk Ak .

Proof. According to the hypotheses of the theorem, the series ∞ k=1 μk Ak converges in the
operator
m norm topology to the operator S and D is an essential domain of the operators Sm =
k=1 μk Ak for all m ∈ N and, hence, of the operator S.
Let us check that the function F satisfies all the hypotheses of Chernoff’s theorem.
Since exp(−itAn )|t=0 = I for all n ∈ N and μ(N) = 1, it follows that F(0) = I.
Let u ∈ H. In view of the strong continuity of the semigroups Un (t) = exp(−itAn ), t ∈ R,
the functions Un (t)u, t ∈ R, are continuous for all n ∈ N and uniformly bounded by the constant
uH . Since exp(−itA n )B(H) = 1 for all n ∈ N and the measure μ is nonnegative and normed on
the set N, the series n∈N exp(−itAn )uμn converges uniformly in t ∈ R by the Weierstrass test to
a continuous function. Hence, the operator function F(t), t ∈ R, is strongly continuous.
In addition, F(t)B(H) ≤ n∈N exp(−itAn )B(H) μn = μ(N) = 1; hence, the inequality
F(t)B(H) ≤ 1 holds.
Let us show that for any u ∈ D the function F(t)u = k∈N μk Uk (t)u is continuously differen-
tiable on [0, +∞), and find the limit value of the derivative F (+0)u.
Let u ∈ D. Then, for any k ∈ N, the function vk (t, u) = exp(−itAk )u, t ∈ R, is continuously
d
differentiable. Moreover, the series k∈N μk dt vk (t) converges uniformly on [0, T ] for any t > 0,
because
  d
 
  m ∞
μk  v (t)
 dt k 
 = μ k A v
k k (t) = μ k A k u ≤ μ k Ak u + μk Ak  · u < +∞.
k∈N k∈N k∈N k=1 k=m

Hence, for any u ∈ D, the function F(t)u, t > 0, is continuously differentiable on [0, +∞); moreover,
d
its derivative has a limit equal to n∈N μn dt un (+0) = Su as t → +0.
Thus, the function F(t), t ≥ 0, satisfies all the hypotheses of Chernoff’s theorem, and so
limm→+∞ [(F(t/n))n − exp(−itS)]uH = 0; i.e., the function F is Chernoff equivalent to the
semigroup U. Theorem 2 is proved.
The last hypothesis of the theorem—the existence of an essential domain that is common for the
terms of the operator series and for its partial sums—is rather complicated and difficult to verify.
The following statement provides a sufficient condition for this hypothesis of the theorem to hold.
m
Corollary 1. Let A1 , . . . , Am be quantum Hamiltonians, and let j=1 pj = 1 and pj > 0.
If the quadratic form of the operator A1 majorizes the quadratic forms of the operators Ak , k =
definition D(A1 ) is an essential domain for the other operators
2, . . . , m (see [3]), then the domain of
and the averaged Hamiltonian A = m j=1 pj Aj , and the mean value of the semigroups is Chernoff
equivalent to the semigroup generated by the average Hamiltonian.
Remark 2. The assertion of Theorem 2 is nontrivial and interesting even in the particular
case when the operators Aj are bounded and self-adjoint for all j ∈ N.
Remark 3. The common domain of definition D of the operators An is an essential domain
of these operators rather than merely a dense subset of the space H, because the information on
the domain D determines uniquely a unitary semigroup, i.e., a self-adjoint operator, rather than a
symmetric operator that has a variety of self-adjoint extensions.
Remark 4. The assertion of the theorem is also of interest in the case of a finite series. It is
an additive analog of Trotter’s theorem and takes the following form.

PROCEEDINGS OF THE STEKLOV INSTITUTE OF MATHEMATICS Vol. 285 2014

 FEYNMAN FORMULAS AS A METHOD OF AVERAGING 229

Definition 2. We say that a maximal dissipative operator L is the sum of maximal dissipative
operators p1 L1 and p2 L2 for arbitrary p1 ≥ 0 and p2 ≥ 0, p1 + p2 = 1, in the exponential sense if
the operator-valued function G(t) = p1 exp(tL1 ) + p2 exp(tL2 ), t ≥ 0, is Chernoff equivalent to the
semigroup exp(tL).
It is obvious that the sum of operators is commutative and associative, the zero operator is the
zero element, and every maximal dissipative operator has its inverse.
Definition 3. Let L(ε), ε ∈ E, be a function that takes values in the set of maximal dissipative
operators. We say that a maximal dissipative operator L is the integral of the function L(ε), ε ∈ E,
with respect to a nonnegative normed
 measure μ on the algebra 2E in the exponential sense if the
operator-valued function G(t) = E exp(tL(ε)) dμ(ε), t ≥ 0 (where the integral is understood in the
weak operator topology in the sense of Pettis), is Chernoff equivalent to the semigroup exp(tL).
Thus, the introduced procedure of averaging the generators of semigroups generalizes the aver-
aging procedure for bounded linear operators.
Remark 5. It seems impossible to give up the condition of nonnegativity of the numbers μn
∞
and replace it even by the condition of absolute convergence of the series n=1 μn within the
proposed proof method, because then the boundedness condition for the norm of the operator F (t)
may fail for small values of t. Indeed, there exist unbounded self-adjoint operators A1 and A2 in
a Hilbert space H such that if μ1 + μ2 = 1, then, for any t > 0 and any δ > 0, there exists a
unit vector u ∈ H such that the inequality (μ1 exp(−itA1 ) + μ2 exp(−itA2 ))uH ≥ |μ1 | + |μ2 | − δ
holds. As the operators A1,2 , one can take commuting operators with an orthonormal basis of
 (1,2) 
eigenvectors {ek } corresponding to eigenvalues λk , respectively. Therefore,
    (1)   (2) 
μ1 exp(−itA1 ) + μ2 exp(−itA2 ) ek = μ1 exp −itλk + μ2 exp −itλk ek

and, hence,
      (2) (1) 
 μ1 exp(−itA1 ) + μ2 exp(−itA2 ) ek  = μ1 + μ2 exp −it λk − λk .
H

 (1,2)  (2) (1)

The eigenvalues λk can be chosen so that the set of numbers λk − λk , k ∈ N, is dense
in R. This implies that the function F(t) = μ1 exp(−itA1 ) + μ2 exp(−itA2 ) does not satisfy the
hypothesis F(t) ≤ eat , t ∈ (0, δ), of Chernoff’s theorem.

4. EXAMPLES OF MEAN VALUES OF A SEQUENCE OF SELF-ADJOINT

OPERATORS THAT ARISE UNDER REGULARIZATION
OF A HAMILTONIAN WITH DEGENERACY

Example 1 (power regularization of a maximal symmetric operator). Let L be a maximal

symmetric but non-self-adjoint operator in a Hilbert space H. Suppose that {Ln } is a sequence of
self-adjoint operators in the space H that has the following properties:

(1) the linear variety D = D(Lq ) ∩ n∈N D(Ln ) is dense in the space H;
(2) for any n ∈ N, there exists a linear operator Qn from H to H with the domain of definition
D(Qn ) = D(Lq−1 ) that maps the linear variety D(Lq ) ⊂ D(Qn ) to the linear variety D(Ln )
and, for any u ∈ D(Lq ), satisfies the estimate

Qn u − uH + Ln Qn u − Qn LuH ≤ bn (Lq uH + uH ),

where the sequence {bn } is infinitesimal.

PROCEEDINGS OF THE STEKLOV INSTITUTE OF MATHEMATICS Vol. 285 2014

230 Yu.N. ORLOV et al.

Following [8], we will call such sequences {Ln } power regularizations of the symmetric operator L.
Proposition 2 (see [8]). Let a maximal symmetric operator L have a power regulariza-
tion of order q ∈ N. Let μ be a nonnegative normed purely finitely additive measure on the
set N. If the indices (n− , n+ ) of the operator L satisfy the condition n+ = 0, then the se-
quence of semigroups {exp(−itLn )}, t ≥ 0, converges in the strong operator topology to the
isometric semigroup
 exp(−itL), t ≥ 0, uniformly on every finite interval, and the equality
exp(−itL) = N exp(−itLn ) dμ(n), t ≥ 0, holds. If, instead, n− = 0, then the sequence of semi-
groups {exp(−itLn )}, t ≥ 0, converges in the weak operator topology to the contraction semi-
∗ ∗
group exp(−itL ), t ≥ 0, uniformly on every finite interval, and the equality exp(−itL ) =
N exp(−itLn ) dμ(n), t ≥ 0, holds.
Example 2 (a sequence of Hamiltonians that converges in the topology of strong graph con-
vergence to a limit Hamiltonian).
Proposition 3 (see [8]). Let L be a maximal symmetric operator in a Hilbert space H. Let
μ be a nonnegative normed purely finitely additive measure on the set N. Suppose that a sequence
of self-adjoint operators {Ln } is such that its strong graph limit Γ contains the graph ΓL of the
operator L (coincides with it ). If the indices (n− , n+ ) of the operator L satisfy the condition
n+ = 0, then the sequence of semigroups {exp(−itLn )}, t ≥ 0, converges in the strong operator
topology to the isometric semigroup exp(−itL), t ≥ 0, uniformly on any finite interval, and the
equality exp(−itL) = N exp(−itLn ) dμ(n), t ≥ 0, holds. If, instead, n− = 0, then the sequence
of semigroups {exp(−itLn )}, t ≥ 0, converges in the weak operator topology to the contraction
∗ ∗
semigroup exp(−itL ), t ≥ 0, uniformly on every finite interval, and the equality exp(−itL ) =
N exp(−itLn ) dμ(n), t ≥ 0, holds.
Therefore, for any choice of a nonnegative normed purely finitely additive measure μ on the
set N, the following equalities hold in the sense of Definition 3: L = N Ln dμ(n) if n+ = 0, and
L∗ = N Ln dμ(n) if n+ > 0.
Remark 6. If a sequence of self-adjoint operators {Ln } is such that its strong graph limit Γ
contains the graph ΓL of the maximal symmetric operator  L (coincides with it), then this does
not mean that the common domain of definition D = n∈N D(Ln ) is dense in the space H; the
variety D may even be trivial.
Thus, the mean values of random semigroups may be semigroups. There are examples when
the mean values of random strongly continuous semigroups are not semigroups but are strongly
continuous operator functions (see [5, 8, 12]).

5. CONVERSION OF CHERNOFF’S THEOREM

The above-defined equivalence relation on the set of strongly continuous operator-valued func-
tions raises the following topical question: What conditions on an operator-valued function are
sufficient for this function to be equivalent to a semigroup, and what conditions are necessary?
A partial answer to this question is given in the following theorem.
Theorem 3. Suppose that an operator-valued function F : [0, T ) → B(X) is continuous in the
strong operator topology and satisfies the conditions F(0) = I and F(t)B(X) ≤ eat , t ∈ (0, δ), for
some a ∈ R and δ > 0. If the sequence {Gn } of operator-valued functions Gn (t) = (F(t/n))n ,
n ∈ N, t ∈ [0, nT ), converges in the strong operator topology uniformly on any interval [0, t], then
the limit operator-valued function is a strongly continuous one-parameter semigroup of operators in
the space X of type ω ≤ a.
Proof. For all n ∈ N, the functions Gn (t) satisfy the conditions Gn (0) = I and map the interval
[0, nT ) to the space B(X) continuously in the strong operator topology. The uniform convergence
of the sequence {Gn } in the strong operator topology on every interval [0, t] implies the continuity

PROCEEDINGS OF THE STEKLOV INSTITUTE OF MATHEMATICS Vol. 285 2014

 FEYNMAN FORMULAS AS A METHOD OF AVERAGING 231

of the limit function G in the strong operator topology and the equality G(0) = I. Moreover, it
follows from the estimate for the norm of the function F that

G(t)B(X) ≤ eat ∀ t ≥ 0. (5.1)

Let us prove that G(t + s) = G(t)G(s) for all t, s ≥ 0.

1. Let t ≥ 0 and m ∈ N. Then
     N m
mt n t
F = F
n N
provided that n = mN . Since the sequence of operator functions {Gn } converges to the operator
function G in the strong operator topology for any t > 0, it follows that for any m ∈ N the sequence
{([F(t/N )]N )m } converges to the operator (G(t))m in the strong operator topology. Therefore, the
convergence of the sequence {Gn } to the operator function G as n → ∞ implies that G(mt) =
(G(t))m for all t ≥ 0 and m ∈ N. Thus, G(t/m) = (G(t))1/m for all t ≥ 0 and m ∈ N.
2. If s = pt/q for some q ∈ N and p ∈ Z+ , then s + t = (p + q)t/q and, in view of item 1, the
equalities G(t) = (G(t/q))q , G(s) = (G(t/q))p , and G(t + s) = (G(t/q))q+p are valid. Hence, the
function G satisfies the semigroup property G(t + s) = G(t)G(s) on the set of rational numbers.
3. Let t and s be arbitrary nonnegative numbers and u ∈ X be a vector. Since the operator
function G is strongly continuous, for any ε > 0 there exists a τ > 0 such that the inequalities
G(t + ξ)u − G(t)uX ≤ ε/2 and G(t + s + ξ)u − G(t + s)uX ≤ ε/2 hold for all ξ ∈ (−τ, τ )
such that t − τ ≥ 0. For any s ∈ [0, +∞), there exists a t ≥ 0 such that t and s are rationally
commensurable and |t − t| < τ . So, G(t )G(s) = G(t + s). However, G(t)u − G(t )u ≤ ε/2 and
G(t + s)u − G(t + s)u ≤ ε/2; therefore, G(t + s)u − G(t)G(s)uX ≤ ε and, since the number
ε > 0 and the vector u ∈ X are arbitrary, the semigroup equality holds for arbitrary t, s ∈ R+ . The
theorem is proved.
Remark 7. Theorem 3 is also valid for a sequence of iterations of operator-valued functions
whose values are nonlinear mappings of the Banach space X into itself. Indeed, the proof of the
theorem does not use the linearity of the operators G(t), t ≥ 0. To formulate the result on nonlinear
mappings, we need to introduce new spaces. Let N (X) be the Banach space of nonlinear bounded
mappings of the Banach space X into itself that satisfy the condition A(θ) = θ, with the norm
defined as AN (X) = supu=θ (A(u)X /uX ). A mapping F : R+ → N (X) is said to be strongly
continuous if the mapping Fu : R+ → X is continuous for any u ∈ X.
Theorem 4. Suppose that an operator-valued function F : [0, T ) → N (X) is strongly contin-
uous and satisfies the conditions F(0) = I and F(t)N (X) ≤ eat , t ∈ (0, δ), for some a ∈ R and
δ > 0. If the sequence {Gn } of operator-valued functions Gn (t) = (F(t/n))n , n ∈ N, t ∈ [0, nT ),
converges in the strong operator topology uniformly on every interval [0, t], then the limit operator-
valued function is a strongly continuous one-parameter semigroup G of nonlinear transformations
of the space X that satisfies the estimate G(t)N (X) ≤ eat .
Remark 8. For any operator-valued function F : [0, T ) → B(X) satisfying the hypotheses of
Theorem 3, the sequence {Gn } of operator-valued functions Gn (t) = (F(t/n))n , n ∈ N, t ∈ [0, nT ),
is compact in the weak operator topology pointwise on the half-axis R+ . However, the convergence of
the sequence of operator functions {Gn } to the operator function G in the weak operator topology
for all t > 0 does not imply that for all m ∈ N the sequence {([F(t/N )]N )m } converges to the
operator (G(t))m in the weak operator topology. Nevertheless, in the case of a finite-dimensional
space X, the strong and weak operator topologies are equivalent. Therefore, if {TAn } is a sequence
of continuous semigroups of linear transformations of a finite-dimensional
 Banach space X, then the
sequence of Chernoff iterations {Gn } of the mean value G = N TAn dμ(n) of the semigroups TAn of

PROCEEDINGS OF THE STEKLOV INSTITUTE OF MATHEMATICS Vol. 285 2014

232 Yu.N. ORLOV et al.

transformations of the finite-dimensional

 space X has a unique limit point, the semigroup generated
by the averaged operator A = N An dμ(n). Hence, in the finite-dimensional space, the mean value
of any random semigroup is Chernoff equivalent to some semigroup.

ACKNOWLEDGMENTS
This work was supported by the Russian Science Foundation (project no. 14-11-00687) and the
Russian Foundation for Basic Research (project no. 14-01-00516).

REFERENCES
1. F. A. Berezin, The Method of Second Quantization (Nauka, Moscow, 1986) [in Russian]; Engl. transl. of the
1st ed.: The Method of Second Quantization (Academic, New York, 1966), Pure Appl. Phys. 24.
2. V. V. Dodonov, V. I. Man’ko, and V. D. Skarzhinsky, “Nonuniquenesses of the variational description of classical
systems and the quantization problem,” Tr. Fiz. Inst. im. P.N. Lebedeva, Akad. Nauk SSSR 152, 37–89 (1983)
[Proc. Lebedev Phys. Inst. Acad. Sci. USSR 176 (Suppl.), 49–122 (1988)].
3. T. Kato, Perturbation Theory for Linear Operators (Springer, Berlin, 1966; Mir, Moscow, 1972).
4. V. P. Maslov, Operational Methods (Nauka, Moscow, 1973; Mir, Moscow, 1976).
5. J. O. Ogun, Yu. N. Orlov, and V. Zh. Sakbaev, “On a transformation of the initial-data space for a Cauchy
problem with blow-up solutions,” Preprint No. 87 (Keldysh Inst. Appl. Math., Moscow, 2012).
6. Yu. N. Orlov, V. Zh. Sakbaev, and O. G. Smolyanov, “Rate of convergence of Feynman approximations of
semigroups generated by the oscillator Hamiltonian,” Teor. Mat. Fiz. 172 (1), 122–137 (2012) [Theor. Math.
Phys. 172, 987–1000 (2012)].
7. V. Zh. Sakbaev, “Set-valued mappings specified by regularization of the Schrödinger equation with degeneration,”
Zh. Vychisl. Mat. Mat. Fiz. 46 (4), 683–699 (2006) [Comput. Math. Math. Phys. 46, 651–665 (2006)].
8. V. Zh. Sakbaev, Cauchy Problem for a Degenerate Linear Differential Equation and Averaging Its Approximating
Regularizations (Ross. Univ. Druzhby Narodov, Moscow, 2012), Sovrem. Mat., Fundam. Napr. 43.
9. P. R. Chernoff, “Note on product formulas for operator semigroups,” J. Funct. Anal. 2, 238–242 (1968).
10. O. G. Smolyanov, A. G. Tokarev, and A. Truman, “Hamiltonian Feynman path integrals via the Chernoff
formula,” J. Math. Phys. 43 (10), 5161–5171 (2002).
11. O. G. Smolyanov, H. von Weizsäcker, and O. Wittich, “Chernoff’s theorem and discrete time approximations of
Brownian motion on manifolds,” Potential Anal. 26 (1), 1–29 (2007).
12. L. Tartar, “Memory effects and homogenization,” Arch. Ration. Mech. Anal. 111 (2), 121–133 (1990).
13. K. Yosida and E. Hewitt, “Finitely additive measures,” Trans. Am. Math. Soc. 72, 46–66 (1952).

Translated by I. Nikitin

PROCEEDINGS OF THE STEKLOV INSTITUTE OF MATHEMATICS Vol. 285 2014