Professional Documents
Culture Documents
STOCHASTIC CALCULUS
DAVID APPLEBAUM
Abstract. Robin Hudsons work on quantum central limit theorems, quan-
tum Brownian motion, quantum stopping times and formal quantum sto-
chastic calculus is reviewed and reappraised.
1. Introduction
Robin Hudson and K. R. Parthasarathy are the creators of quantum stochastic
calculus. The key paper which contained almost all the basic results was published
in 1984 in Communications in Mathematical Physics [19]. Here we nd, in
particular, the construction of quantum stochastic integrals, quantum Itos formula,
the existence and uniqueness of linear quantum stochastic dierential equations
(QSDEs), necessary and sucient conditions for unitarity of solutions and the
dilation of quantum dynamical semigroups (at least for one degree of freedom). In
this article I will focus on the pre-history of quantum stochastic calculus with a
particular emphasis on Robin Hudsons contribution. This will cover work that was
published in the period 1971-84. As can be seen by a quick journey to MathSciNet,
this was a highly productive period for Robin. I am not going to survey all his
papers from this period in this article or even all of those that he wrote in the area
of quantum probability. What I will present is four key milestones - the quantum
central limit theorem, quantum Brownian motion, quantum stopping times and
the heuristic version of quantum stochastic calculus that preceded its rigorous
development in [19].
A brief comment on the title. In the ancient Chinese philosophy of Taoism, the
mysterious tao is often described (at least in contemporary English translation) as
a pathless path. Of course in quantum theory, a particle does not have a path
in the usual sense and consequently path space techniques are inappropriate tools
for studying quantum processes.
Notation. If H is a complex Hilbert space then B(H) is the algebra of all
bounded linear operators on H. If T is a densely dened closeable linear operator
dened on H with adjoint T
.
Received 2010-4-12; Communicated by the editors.
2000 Mathematics Subject Classication. Primary 81S25; Secondary 81Q12, 81S05, 81R15,
81R30.
Key words and phrases. Quantum central limit theorem, quantum Brownian motion, quan-
tum stop times, Fock space, quantum stochastic calculus, quantum diusions.
481
Serials Publications
www.serialspublications.com
Communications on Stochastic Analysis
Vol. 4, No. 4 (2010) 481-491
482 DAVID APPLEBAUM
2. Quantum Central Limit Theorem
Robins rst paper on quantum probability was joint work with his PhD student
Clive Cushen [9]. The opening words of the introduction to this paper are almost a
clarion call for quantum probability: In recent years there has been an increasing
awareness that the foundations of quantum mechanics lie in a non-commutative
analogue of axiomatic probability theory. In order to formulate a quantum central
limit theorem (CLT), Cushen and Hudson rst needed to decide what should a
quantum random variable be and how could a sequence of these be identically
distributed and independent? The basic (bosonic) quantum random variable is
a canonical pair (q, p) of linear self-adjoint operators acting in a complex Hilbert
space H and satisfying the Heisenberg commutation relation
[p, q] := pq qp = iI,
on a suitable dense domain. The distribution of the pair (q, p) is determined by
a mixed state which is identied with its density operator . Now consider an
innite sequence ((q
n
, p
n
), n N) of such canonical pairs satisfying
[q
i
, q
j
] = [p
i
, p
j
] = 0; [p
i
, q
j
] = i
ij
I,
for each i, j N and equipped with the state . They are said to be independent
if all nite subsets are independent in the sense that if A N is nite with
A = {i
1
, . . . , i
N
} then
A
is unitarily equivalent to
i1
iN
. Here
A
is the
reduced density operator dened on L
2
(R
N
) by
tr(
A
T) = tr((T)),
for all T B(L
2
(R
N
)) and is the canonical embedding of B(L
2
(R
N
)) into B(H)
that is determined by von Neumanns uniqueness theorem. The reduced states
i1
, . . . ,
iN
are similarly obtained by taking A = {i
1
}, . . . , {i
N
} (respectively).
The sequence ((q
n
, p
n
), n N) comprises identically distributed quantum random
variables if
i1
= =
iN
for every nite set A N. Now we average. For each
n N, dene
p
n
=
1
n
(p
1
+ +p
n
), q
n
=
1
n
(q
1
+ +q
n
).
It is easy to see that (q
n
, p
n
) form a canonical pair and the main result of the
paper is to prove the quantum central limit theorem:
lim
n
tr(
n
T) = tr(
T), (2.1)
for all T B(L
2
(R)). Here
n
is the reduced density operator corresponding
to the canonical pair (q
n
, p
n
) and
n
), n N) which satisfy the canonical anticommutation relations
(CARs):
{a
i
, a
j
} = {a
i
, a
j
} = 0; {a
i
, a
j
} =
ij
I,
for i, j N, where {A, B} := AB + BA is the anticommutator. Once again we
get a new representation of the CARs by averaging:
a
#
n
:=
1
n
(a
#
1
+ +a
#
n
),
and the fermionic central limit theorem yields convergence of corresponding re-
duced states to a fermionic quasi-free state (i.e. a fermionic Gaussian). In this
work, the key tool was cumulants rather than quasi-characteristic functions. The
development of the appropriate fermionic version of the latter required the use of
Grassman algebra techniques [2].
The Cushen-Hudson work is without doubt a landmark paper, and not just for
its inuence on quantum probability. The study of quantum central limit theorems
is a ourishing enterprise in its own right which has attracted the attention of a
large number of authors since the early 1970s, see e.g. [11], [10], [12] and [21]. In-
deed the recent article [21] lists seven distinct areas in which quantum central limit
theorems have been developed and applied including quantum information theory,
graph theory and combinatorics. K.R.Parthasarathy told the author
1
that read-
ing this paper for the rst time deeply inuenced his subsequent mathematical
life (sic.)
3. Quantum Brownian Motion
The rst paper to study quantum Brownian motion was published by Robin
together with another of his PhD students
2
Anne Cockcroft in 1977 [7]. In this
paper, the passage of time is modelled by the closed interval [0, 1] , but the authors
could just as easily used R
+
:= [0, ) and this became the standard choice in later
work. A quantum Brownian motion
3
is a pair (P(t), Q(t), t [0, 1]) of self-adjoint
operator-valued functions acting in a complex Hilbert space H together with a
distinguished vector to determine expectations such that:
(i) [P(s), P(t)] = [Q(s), Q(t)] = 0; [P(s), Q(t)] = is t
for all s, t [0, 1].
(ii) P(0) = Q(0) = 0.
1
E-mail communication in April 2010
2
Robin has always been extremely generous in sharing his ideas with others and many PhD
students, including the author, have been beneciaries of this largesse.
3
In [7] the authors used the terminology quantum Wiener process, but quantum Brownian
motion became the preferred usage amongst practitioners.
484 DAVID APPLEBAUM
(iii) For := (a, b] [0, 1], dene the canonical pair (p
, q
) by
p
:=
p(b) p(a)
b a
, q
:=
q(b) q(a)
b a
.
For arbitrary pairwise disjoint (
n
, n N), the sequence ((p
n
), q
n
)), n
N) consists of independent and identically distributed (i.i.d.) canonical
pairs having quantum Gaussian distributions with mean 0 and variance
2
1 in the state determined by .
Note that each of (Q(t), t 0) and (P(t), t 0) are separately equivalent to
the probabilists Brownian motion but the non-trivial commutation relation in (i)
ensures that these are not simultaneously diagonisable. The key probabilistic input
of the denition is in (iii) which extends to a non-commutative framework the fact
that a classical Brownian motion (B(t), t 0) has stationary and independent
increments with each random variable
B(t)B(s)
ts
N(0,
2
).
The main result of this paper is to show that any quantum Brownian motion is
unitarily equivalent to the pair of co-ordinate and momentum eld operators
which are indexed by the indicator function 1
[0,t)
and are associated to a cyclic
representation of the extremal universally invariant quasi-free state
on the Weyl
CCR algebra over H dened on Weyl operators W(f) by
(W(f)) = e
1
2
2
||f||
2
for each f H. The case = 1 is the Fock state.
In later years, the development of quantum stochastic calculus, made it more
convenient to identify quantum Brownian motion with the annihilation/creation
operator valued process (A(t), A
(t), t 0) dened by
A(t) =
1
2
(Q(t) +iP(t)), A
(t) =
1
2
(Q(t) iP(t)),
which satisfy the commutation relations (CCRs):
[A(s), A(t) = [A(s)
, A(t)
] = 0, [A(s), A
(t)] = s tI,
for all s, t 0. The Cockroft-Hudson theory then tells us that every quantum
Brownian motion for which = 1 is unitarily equivalent to
(a(1
[0,t)
), a
(1
[0,t)
), t 0),
where a(), a
(1
[0,t)
), A
(t) = a
(1
[0,t)
) I +I a(1
[0,t)
),
where
2
2
= 1,
2
+
2
=
2
. The distinguished vector is . The non-Fock
quantum Brownian motions have a deep and beautiful mathematical structure
(see e.g. [16]) and it may well be that they havent yet been exploited to their full
potential.
The Cockcroft-Hudson paper is justly celebrated as marking the birth of quan-
tum Brownian motion. Perhaps less well-known is the follow-up paper [8] by the
ROBIN HUDSONS PATHLESS PATH 485
same authors (but now augmented with S.Gudder) which directly followed it in the
same volume of the Journal of Multivariate Analysis. Here the authors establish
a functional central limit theorem for quantum Brownian motion. The set-up is
as follows. Suppose that we have a sequence of canonical pairs ((p
n
, q
n
), n N)
that are i.i.d. with respect to a state for which
(q
n
) = (p
n
) = ({p
n
, q
n
}) = 0; (q
2
n
) = (p
2
n
) =
2
,
for all n N, where
2
1. For each t [0, 1] consider the operators dened by
P
n
(t) =
1
n
(p
1
+ +p
[nt]
+ (nt [nt])p
[nt]+1
),
Q
n
(t) =
1
n
(q
1
+ +q
[nt]
+ (nt [nt])q
[nt]+1
).
It follows that for each s, t [0, 1]
[P
n
(s), P
n
(t)] = [Q
n
(s), Q
n
(t)] = 0, [P
n
(s), Q
n
(t)] = i(s t +r
n
),
where lim
n
r
n
= 0. The authors demonstrate that the sequence ((P
n
, Q
n
),
n N) converges weakly to quantum Brownian motion of variance
2
. A large
part of the paper grapples with the question of what weak convergence might
mean in this context. The authors build an elaborate technical apparatus which
inter alia requires the compact uniform closure of a C
, 0)
where N
:= {P(t), Q(t), 0 t }
0
dE() which is such
that E() N
for all 0.
5
At least formally the random time-shifted quantum
Brownian motion should be
P
T
(t) :=
0
(P(t +) P())dE(), Q
T
(t) :=
0
(Q(t +) Q())dE()
for each t 0. In order to give these formal expressions a rigorous meaning, Robin
denes them indirectly as innitesimal generators of the unitary operator-valued
spectral integrals dened for each x R by
U
P(t)
(x) : =
0
e
ixP(t+)
e
ixP()
dE(),
V
Q(t)
(x) : =
0
e
ixQ(t+)
e
ixQ()
dE() (4.1)
so (U
P(t)
(x), x R) and (V
Q(t)
(x), x R) are each strongly continuous one-
parameter unitary groups and we have
U
P(t)
(x) = e
ixP(t)
, V
Q(t)
(x) = e
ixQ(t)
,
for each t 0, x R. Much of the technical work in the paper involves the
construction of integrals of the type considered in (4.1) as limits of Riemann sums
in the strong operator topology.
Before we can state the quantum strong Markov property, we need the concepts
of pre-T and post-T von Neumann algebras which well denote as N
T]
and N
[T
respectively. These are dened by
N
T]
:= {A N, AE() = E()A N
.
The strong Markov property states that N
T]
and N
[T
are independent in the
state
6
and that ((P
T
(t), Q
T
(t)), t 0) is a quantum Brownian motion of vari-
ance
2
.
5
Later literature on quantum stop times usually dened these directly in terms of projection-
valued measures - see e.g. [29].
6
i.e. , AB = , A, B for all A N
T]
, B N
[T
.
ROBIN HUDSONS PATHLESS PATH 487
A corresponding strong Markov property was established by the author for
fermion Brownian motion in [2]. A key later development of quantum stop-
ping times was the paper [29] by Parthasarathy and Sinha in which the tenso-
rial factorisation of Fock space over L
2
(R
+
) corresponding to the splitting f
f1
[0,t)
+f1
[t,)
was extended to the case where t is replaced by a quantum stopping
time. Another paper worth mentioning (which sadly has never been followed up in
the literature) is a very interesting study of rst exit times by J.-L.Sauvageot [31].
Although there has continued to be sporadic work on quantum stopping times (see
e.g. [15] for a recent survey article by Robin) it seems that a breakthrough is still
needed to forge it into a tool that is of similar value in quantum probability to its
commutative counter-part.
5. Formal Quantum Stochastic Calculus
In the last part of this survey I will focus on work carried out during the early
1980s. A great deal of the standard conceptual structure of quantum stochastic
calculus was developed by Robin and co-workers (principally K.R.Parthasarathy
and R.F.Streater) from a heuristic viewpoint. The rigorous development came a
lot later. In many ways these formal calculations (which are quite satisfactory to
most physicists) constitute the essence of the subject.
At this time, the basic quantum processes were understood to be the annihila-
tion/creation pair (A(t), A
(t), t 0) (where A
#
(t) = a
#
(1
[0,t)
) acting in boson
Fock space (L
2
(R
+
)) and equipped with the vacuum vector to determine ex-
pectations. The ltration was induced by the canonical isomorphism between
(L
2
(R
+
)) and (L
2
([0, t))) (L
2
([t, )) which maps each exponential vector
e(f) to e(f1
[0,t)
) e(f1
[t,)
). In order to dene formal quantum stochastic inte-
grals we write dA
#
(t) := a
#
(1
[t,t+dt)
). Everything follows from the eigenrela-
tion:
A(t)e(f) =
t
0
f(s)ds
e(f),
for each f L
2
(R
+
). Taking a deep breath, we then nd that formal dierentiation
yields:
dA(t)e(f) = f(t)e(f)dt, (5.1)
and so for suitable operator-valued processes (F(t), t 0) we can dene the quan-
tum stochastic annihilation integral
t
0
F(s)dA(s) by its action on exponential
vectors:
t
0
F(s)dA(s)
e(f) =
t
0
F(s)f(s)ds
e(f).
The creation integral is obtained by formal adjunction:
e(f),
t
0
G(s)dA
(s)
e(g)
t
0
G
(s)dA(s)
e(f), e(g)
t
0
f(s)e(f), G(s)e(g),
for each f, g L
2
(R
+
). The celebrated quantum Ito formula is summarised in the
following table:
488 DAVID APPLEBAUM
dA
(t) dA(t) dt
dA(t) dt 0 0
dA
(t) 0 0 0
dt 0 0 0
These formal relations are suggested by the following type of calculation using
the CCRs and (5.1):
e(f), dA(t)dA
(t)dA(t)e(g)
= dte(f), e(g) +dA(t)e(f), dA(t)e(g)
= (dt +o(dt))e(f), e(g)
The preceeding heuristic calculations were all given a precise rigorous meaning
in the seminal paper [19].
The non-trivial Ito correction term dA(t)dA
j
c
j
f
j
(A(t)
, t)g
j
(A(t), t),
where each c
j
C and f
j
and g
j
are smooth. Dene formal partial derivatives by
M(t)
A(t)
:=
j
c
j
1
f
j
(A(t)
, t)g
j
(A(t), t),
M(t)
A(t)
:=
j
c
j
f
j
(A(t)
, t)
1
g
j
(A(t), t)
and
M(t)
t
:=
j
c
j
[
2
f
j
(A(t)
, t)g
j
(A(t), t) +f
j
(A(t)
, t)
2
g
j
(A(t), t)],
where for i = 1, 2,
i
denotes partial dierentiation with respect to the ith variable.
The authors then show that
dM(t) =
M(t)
A(t)
dA
(t) +
M(t)
A(t)
dA(t) +
M(t)
t
dt.
In the two papers [17] and [18], Hudson and Parthasarathy investigate quantum
diusions. These are prototypes for the quantum stochastic processes (in the sense
of [1]) that eventually became known as quantum stochastic ows or Evans-Hudson
ows (see e.g. [25] and [27]). The authors work in the space h := h
0
(L
2
(R
+
))
where h
0
is a complex, separable Hilbert space which carries a representation of
the CCRs. So we have a pair (a, a
t
] = 1 for
all t 0. These are required to be adapted to the Fock ltration in that each
ROBIN HUDSONS PATHLESS PATH 489
a
#
t
= a
#
1
(t) I where a
#
1
(t) operates non-trivially on h
0
(L
2
([0, t))). The form
of the perturbation is given by
da
t
= F(t)dA(t) +G(t)dA
(t) +H(t)dt,
and applying quantum Itos formula to the CCRs yields the restraint equations
(which are to be read pointwise in t):
[F, a
] = [a, G
] = 0,
[H, a
] + [a, H
] = F
F GG
.
Furthermore they obtain formal conditions for the dynamics to be induced by
a unitary operator-valued process (U(t), t 0). Indeed the unitarity requirement
implies the form:
dU(t) = U(t)
LdA
(t) L
dA(t) +
iH
1
2
L
dt
,
with U(0) = I, where L and H are (ampliations of) linear operators acting on h
0
with H being formally self-adjoint. In order to obtain
a
#
(t) = U(t)(a
#
I)U(t)
], H = i[H, a]
1
2
(L
La 2L
aL +aL
L),
so H = L(a) where L is the Lindblad generator. Indeed it is precisely the generator
of the quantum dynamical semigroup (T
t
, t 0) dened by
T
t
(X) = E
0
(U(t)(X I)U(t)
),
for each X B(h
0
), where E
0
denotes the vacuum conditional expectation. A
fermionic version of some of these ideas was developed in [4].
As was pointed out above, these ideas were made fully rigorous in [19] which
also introduced the conservation process
7
into quantum stochastic calculus and
thus completed the trio of basic integrators. The theory developed therein has
been described and extended in a number of monographs and surveys (see e.g.
[27], [25], [5], [23]) so the reader will surely forgive me if I stop at this point.
In conclusion, the period 1971-1984 saw a remarkable period of activity from
Robin and his collaborators which led from the quantum central limit theorem to
quantum Brownian motion and then to the development of quantum stochastic
calculus. It is perhaps a little unfair to compare this to the gap between the
rst use of the central limit theorem by Abraham de Moivre in 1733 and the
discovery of stochastic calculus by Kiyosi Ito in the 1940s (see [22] for a concise
historical account of developments leading to the birth of the latter), nonetheless
it is certainly a considerable achievement.
7
Sometimes called the gauge or number process
490 DAVID APPLEBAUM
References
1. Accardi, L., Frigerio, A, and Lewis, J. T.: Quantum stochastic processes, Publ. Res. Inst.
Math. Sci. (Kyoto Univ.) 18 (1982) 97133.
2. Applebaum, D.: The strong Markov property for fermion Brownian motion, J. Funct. Anal.
65 (1986) 273291.
3. Applebaum, D.: A Levy-Cielsielski expansion for quantum Brownian motion and the con-
struction of quantum Brownian bridges, J. Appl. Anal. 13 (2007) 275290.
4. Applebaum, D. and Hudson, R. L.: Fermion Diusions, J. Math. Phys. 25 (1984) 851861.
5. Biane, P.: Calcul stochastique non-commutatif, in Lectures on Probability Theory: Ecole
dEte de Probabilites de Saint Flour XXIII, P.Bernard ed. Lecture Notes in Math. 1608
(1995) 196 Springer-Verlag Berlin.
6. Bo zejko, M. and Speicher, R.: An example of a generalized Brownian motion, Commun.
Math. Phys. 137 (1991) 519531.
7. Cockroft, A. M. and Hudson, R. L.: Quantum mechanical Wiener processes, J. Mult. Anal.
7 (1977) 107124.
8. Cockroft, A. M., Gudder, S. P., and Hudson, R.L.: A quantum mechanical functional central
limit theorem, J. Mult. Anal. 7 (1977) 125149.
9. Cushen, C. D. and Hudson, R. L.: A quantum-mechanical central limit theorem, J. Appl.
Prob. 8,(1971) 454469.
10. Dorlas, T. C.: A non-commutative central limit theorem, J. Math. Phys. 37 (1996) 4662
4682.
11. Giri, N. and von Waldenfels, W.: An algebraic version of the central limit theorem, Z.
Wahrsch. verw. Geb. 42 (1978) 129134.
12. Goderis, D., Verbeure, A., and Vets, P.: Non-commutative central limits, Probab. Th. Rel.
Fields 82 (1989) 527544.
13. Hudson, R. L.: A quantum-mechanical central limit theorem for anti-commuting observables,
J. Appl. Prob. 10 (1973) 502509.
14. Hudson, R. L.: The strong Markov property for canonical Wiener processes, J. Funct. Anal.
34 (1979) 266281.
15. Hudson, R. L.: Stop times in Fock space quantum probability, Stochastics 79 (2007) 383391.
16. Hudson, R. L. and Lindsay, J. M.: A non-commutative martingale representation theorem
for non-Fock quantum Brownian motion, J. Funct. Anal. 61 (1985) 202221.
17. Hudson, R. L. and Parthasarathy, K. R.: Quantum diusions, in Theory and application
of random elds (Bangalore, 1982), Lecture Notes in Control and Inform. Sci., 49 (1983)
111121 Springer, Berlin.
18. Hudson, R. L. and Parthasarathy, K. R.: Construction of quantum diusions, in Quantum
probability and applications to the quantum theory of irreversible processes (Villa Mondrag-
one, 1982), Lecture Notes in Math., 1055 (1984) 173198 Springer, Berlin.
19. Hudson, R. L. and Parthasarathy, K. R.: Quantum Itos formula and stochastic evolution,
Commun. Math. Phys 93 (1984) 301323.
20. Hudson, R. L. and Streater, R. F.: Itos formula is the chain rule with Wick ordering, Phys.
Lett. 86A (1981) 277279.
21. Jacksic, V., Pautrat, Y., and Pillet, C.-A.: A quantum central limit theorem for sums
of independent, identically distributed random variables, to appear in J. Math. Phys 51,
(2010).
22. Jarrow, R. and Protter, P.: A short history of stochastic integration and mathematical
nance; the early years 1880-1970, in A Festschrift for Herman Rubin, Institute of Mathe-
matical Statistics Lecture Notes- Monograph Series Volume 45 (2004) 7591.
23. Lindsay, J. M., Quantum stochastic analysis in Quantum Independent Increment Processes I:
From Classical Probability to Quantum Stochastic Calculus, M Schurmann, U. Franz (Eds.),
Lecture Notes in Mathematics Vol.1865 (2005) 181-271 Springer-Verlag Berlin Heidelberg.
24. Mansuy, R. and Yor, M.: Aspects of Brownian Motion, Springer-Verlag Berlin Heidelberg,
2008.
25. Meyer, P. A.: Quantum Probability for Probabilists, (second edition), Lecture Notes in Math-
ematics Vol 1538 Springer-Verlag, Berlin, Heidelberg, 1995.
ROBIN HUDSONS PATHLESS PATH 491
26. Muraki, N.: Noncommutative Brownian motion in monotone Fock space, Commun. Math.
Phys. 183 (1997) 557570.
27. Parthasarathy, K. R.: An Introduction to Quantum Stochastic Calculus, Birkhauser Verlag
Basel, 1992.
28. Parthasarathy, K. R.: What is a Gaussian state?, Communications on Stochastic Analysis
4 (2010) 143160.
29. Parthasarathy, K. R. and Sinha, K. B.: Stop times in Fock space stochastic calculus, Prob.
Theory Rel. Fields 73 (1987) 317349.
30. Rogers, L. C. G. and Williams, D: Diusions, Markov Processes and Martingales, vol. 1,
Foundations, Wiley 1979, 1994; Cambridge University Press, 2000.
31. Sauvageot, J-L.: First exit time: a theory of stopping times in quantum processes. Quantum
probability and applications, III (Oberwolfach, 1987), Lecture Notes in Math. Vol.1303
(1988) 285299 Springer, Berlin.
32. Speicher, R.: A new example of independence and white noise, Prob. Theory Rel. Fields
84 (1990) 141159.
David Applebaum: School of Mathematics and Statistics, University of Sheffield,
Sheffield S3 7RH England.
E-mail address: d.applebaum@sheffield.ac.uk
URL: http://www.applebaum.staff.shef.ac.uk/
ON THE CHARACTERISTIC FUNCTION OF RANDOM
VARIABLES ASSOCIATED WITH BOSON LIE ALGEBRAS
LUIGI ACCARDI AND ANDREAS BOUKAS
Abstract. We compute the characteristic function of random variables de-
ned as self-adjoint linear combinations of the Schrodinger algebra generators.
We consider the extension of the method to innite dimensional Lie algebras.
We show that, unlike the second order (Schrodinger) case, in the third order
(cube of a Gaussian random variable) case the method leads to a nonlinear
innite system of ODEs whose form is explicitly determined.
1. Introduction
It is known (see [5] and the bibliography therein) that, while the usual Heisen-
berg algebra and sl(2; R) admit separately a continuum limit with respect to the
Fock representation, the corresponding statement for the Lie algebra generated by
these two, i.e. the Schrodinger algebra is false.
Stated otherwise: there is a standard way to construct a net of C
algebras
each of which, intuitively speaking, is associated to all possible complex valued
step functions dened in terms of a xed nite partition of R into disjoint in-
tervals (see [6]), but there are obstructions to the natural extension of the Fock
representation to this net of C
] := a a
a = 1
then
(i) the Heisenberg algebra 1 is the Lie algebra generated by a, a
, 1
(ii) the sl(2) algebra is the Lie algebra generated by
a
2
, a
2
, a
a +
1
2
, a
a +
1
2
, 1
(iv) the Schrodinger algebra o is the Lie algebra generated by
a, a
, a
2
, a
2
, a
a +
1
2
, 1
(v) the universal enveloping Heisenberg algebra | (containing 1, sl(2), O and o
as sub-algebras) is the Lie algebra generated by
a
n
a
k
; n, k = 0, 1, 2, ...
ON THE CHAR. FUNCTION OF RVS ASSOCIATED WITH BOSON LIE ALGEBRAS 495
Dening duality by (a)
= a
a, a
] = a
; [a, a
a] = a
[a
2
, a
2
] = 2 + 4 a
a ; [a
2
, a
a] = 2 a
2
; [a
a, a
2
] = 2 a
2
and
[a
2
, a
] = 2 a ; [a, a
2
] = 2 a
Proof. This is a well known fact that can be checked without diculty.
Proposition 2.4. (Commutation relations in o) Using the notation S
1
0
= a
,
S
0
1
= a, S
2
0
= a
2
, S
0
2
= a
2
, S
1
1
= a
a +
1
2
, S
0
0
= 1 we can write the commutation
relations among the generators of o as
[S
n
k
, S
N
K
] = (k N K n) S
n+N1
k+K1
(2.1)
with (S
n
k
)
= S
k
n
, for all n, k, N, K 0, 1, 2 with n +k 2 and N +K 2.
Proof. The proof follows by using Lemma 2.3.
Remark 2.5. The concise formula (2.1) for the commutation relations among the
generators of o, in particular the inclusion of [S
0
2
, S
2
0
] = 4 S
1
1
, was the reason
behind the inclusion of the additive
1
2
term in the denition of the generators of
o.
To describe the commutation relations in |, using the notation
_
y,z
x
_
=
_
y
x
_
z
(x)
,
n,k
= 1
n,k
where
n,k
is Kroneckers delta, x
(y)
= x(x 1) (x y + 1) for
x y, x
(y)
= 0 for x < y, x
(0)
= 1, we dene
L
(N, K; n, k) = H(L 1)
_
K,0
n,0
_
K, n
L
_
k,0
N,0
_
k, N
L
__
(2.2)
where H(x) is the Heaviside function (i.e., H(x) = 1 if x 0 and H(x) = 0
otherwise). Notice that if L exceeds (K n) (k N) then
L
(N, K; n, k) = 0.
Proposition 2.6. (Commutation relations in |) Using the notation B
x
y
= a
x
a
y
,
for all integers n, k, N, K 0
[B
N
K
, B
n
k
] =
L1
L
(N, K; n, k) B
N+nL
K+kL
where
L
(N, K; n, k) is as in (2.2).
Proof. The proof follows from a discretization of Lemma 2.3 of [5] (i.e. by elimi-
nating the time indices t and s and replacing the Dirac delta function by 1) and
is a consequence of the General Leibniz Rule, Proposition 2.2.2. of [11] .
496 LUIGI ACCARDI AND ANDREAS BOUKAS
Denition 2.7. (The Heisenberg Fock space) The Heisenberg Fock space T is
the Hilbert space completion of the linear span c of the set of exponential vectors
y() = e
a
n
a
k
y() =
k
n
n
[
=0
y( +) ; n, k = 0, 1, 2, ...
If s R and X is a self-adjoint operator on the Heisenberg Fock space T
then, through Bochners theorem, , e
i s X
can be viewed as the characteristic
function (i.e. the Fourier transform of the corresponding probability measure) of
a classical random variable whose quantum version is X.
The dierential method (see [11], [10] and Lemma 3.3 below) for the compu-
tation of , e
i s X
relies on splitting the exponential e
s X
into a product of
exponentials of the Lie algebra generators. In the case of a Fock representation,
where the Lie algebra generators are divided into creation (raising) operators, an-
nihilation (lowering) operators that kill the vacuum vector , and conservation
(number) operators having as an eigenvector, the exponential e
s X
is split into
a product of exponentials of creation operators.
The coecients of the creation operators appearing in the product exponentials
are, in general, functions of s satisfying certain ordinary dierential equations
(ODEs) and vanishing at zero. In the case of a nite dimensional Lie algebra (see
section 3) the ODEs can be solved explicitly and the characteristic function of the
random variable in consideration can be explicitly determined.
For innite dimensional Lie algebras such as, for example, the sub-algebra of
| generated by a
3
and a
3
, the situation is dierent. As illustrated in section 5,
the ODEs form an innite non-linear system which, in general, is hard (if at all
possible) to solve explicitly.
3. Schrodinger Stochastic Processes
In this section, using the notation of Proposition 2.4, we compute the charac-
teristic function , e
i s V
, s R, of the quantum random variable
V = a S
1
0
+ a S
0
1
+b S
2
0
+
b S
0
2
+ S
0
0
+ S
1
1
where a, b C with b ,= 0 (the case b = 0 is discussed in the remarks following
Proposition 3.4) and , R with 4 [b[
2
2
,= 0.
The above form of V is the extention of the forms considered in [2]-[4] to the
full Schrodinger algebra. The splitting lemmas for the Heisenberg and sl(2) Lie
algebras can be found in [11]. An analytic proof of the splitting lemma for sl(2)
can be found in [8].
Remark 3.1. Here, and in what follows, the terms splitting and disentanglement
refer to the expression of the exponential of a linear combination of operators as
a product of exponentials.
ON THE CHAR. FUNCTION OF RVS ASSOCIATED WITH BOSON LIE ALGEBRAS 497
Lemma 3.2. For all a, b C,
(i)
S
0
2
e
a S
2
0
= 4 a
2
S
2
0
e
a S
2
0
+ 4 a e
a S
2
0
S
1
1
+e
a S
2
0
S
0
2
(ii)
S
0
2
e
b S
1
0
= b
2
e
b S
1
0
(iii)
S
0
1
e
a S
2
0
= 2 a S
1
0
e
a S
2
0
+e
a S
2
0
S
0
1
(iv)
S
1
1
e
a S
1
0
= a S
1
0
e
a S
1
0
+e
a S
1
0
S
1
1
(v)
S
1
1
e
a S
1
0
=
_
a S
1
0
+
1
2
_
e
a S
1
0
(vi)
S
1
1
e
b S
2
0
= 2 b S
2
0
e
b S
2
0
+e
b S
2
0
S
1
1
(vii)
S
1
1
e
b S
2
0
=
_
2 b S
2
0
+
1
2
_
e
b S
2
0
Proof. The proof of (i) through (iii) can be found in [2]. The proof of (iv) and (vi)
follows, respectively, from Propositions 2.4.2 and 3.2.1 of [11] with, in the notation
of [11] , D = S
0
1
, X = S
1
0
, X D = S
1
1
1
2
, H = t 1, and f(X) = e
a X
for (iv), and
S
2
0
= 2 R, S
0
2
= 2 , S
1
1
= , and f(R) = e
b R
for (vi). Finally, (v) and (vii) follow
from the fact that S
1
1
=
1
2
.
The following lemma is the extention of the splitting lemmas of [1]-[3] to the
full Schrodinger algebra.
Lemma 3.3. For all s C
e
s V
= e
w1(s) S
2
0
e
w2(s) S
1
0
e
w3(s)
where, letting
K =
_
4 [b[
2
2
; L = arctan
_
K
_
; =
a
2
b
=
a
K
+ tan L ; = ( cos L + sin L)
we have that
w
1
(s) =
K tan (K s +L)
4
b
(3.1)
w
2
(s) = tan (K s +L) + sec (K s +L) + (3.2)
and
w
3
(s) = c
0
+c
1
s +c
2
ln (cos (K s +L))
+c
3
ln (sec (K s +L) + tan (K s +L))
+c
4
tan (K s +L) +c
5
sec (K s +L) (3.3)
498 LUIGI ACCARDI AND ANDREAS BOUKAS
where
c
1
= + a +
b (
2
2
)
c
2
=
_
a + 2
b
K
+
1
2
_
c
3
=
a + 2
b
K
c
4
=
b (
2
+
2
)
K
c
5
=
2
b
K
c
0
= (c
2
ln (cos L) +c
3
ln (sec L + tan L)
+c
4
tan L +c
5
sec L)
Proof. We will show that w
1
(s), w
2
(s) and w
3
(s) satisfy the dierential equations
w
1
(s) = 4
b w
1
(s)
2
+ 2 w
1
(s) +b
w
2
(s) = (4
b w
1
(s) +) w
2
(s) +a + 2 a w
1
(s)
w
3
(s) = + a w
2
(s) + 2
b w
1
(s) +
b w
2
(s)
2
+
2
with w
1
(0) = w
2
(0) = w
3
(0) = 0, whose solutions are given by (3.1), (3.2) and
(3.3) respectively. We remark that the dierential equations dening w
1
(s) and
w
2
(s) are, respectively, of Riccati and linear type. Moreover, unlike the innite
dimensional case treated in section 5, the ODE for w
1
does not involve w
2
and w
3
.
So let
E := e
s (b S
2
0
+
b S
0
2
+a S
1
0
+ a S
0
1
+S
0
0
+ S
1
1
)
(3.4)
Then, by the disentanglement assumption,
E = e
w1(s) S
2
0
e
w2(s) S
1
0
e
w3(s)
(3.5)
where w
i
(0) = 0 for i 1, 2, 3. Then, (3.5) implies that
s
E =
_
w
1
(s) S
2
0
+w
2
(s) S
1
0
+w
3
(s)
_
E (3.6)
and, since by Lemma 3.2,
S
0
2
E = (4 w
1
(s)
2
S
2
0
+ 4 w
1
(s) w
2
(s) S
1
0
(3.7)
+2 w
1
(s) +w
2
(s)
2
) E
S
0
1
E =
_
2 w
1
(s) S
1
0
+w
2
(s)
_
E (3.8)
and
S
1
1
E =
_
2 w
1
(s) S
2
0
+w
2
(s) S
1
0
+
1
2
_
E (3.9)
from (3.4) we also obtain
s
E = ((b + 4
b w
1
(s)
2
+ 2 w
1
(s)) S
2
0
(3.10)
+(a + 2 a w
1
(s) + 4
b w
1
(s) w
2
(s) + w
2
(s)) S
1
0
+ + a w
2
(s) + 2
b w
1
(s) +
b w
2
(s)
2
+
2
) E
ON THE CHAR. FUNCTION OF RVS ASSOCIATED WITH BOSON LIE ALGEBRAS 499
From (3.6) and (3.10), by equating coecients of S
2
0
, S
1
0
and 1, we obtain (3.1),
(3.2) and (3.3) thus completing the proof.
Proposition 3.4. (Characteristic Function of V ) In the notation of Lemma 3.3,
for all s R
, e
i s V
= exp (c
0
+i c
1
s +c
2
ln (cos (K i s +L))
+c
3
ln (sec (K i s +L) + tan (K i s +L))
+c
4
tan (K i s +L) +c
5
sec (K i s +L))
Proof. Using the fact that for any z C
e
z S
0
2
= e
z S
0
1
=
by Lemma 3.3 we have
, e
s V
= , e
w1(s) S
2
0
e
w2(s) S
1
0
e
w3(s)
= e
w2(s) S
0
1
e
w1(s) S
0
2
, e
w3(s)
= , e
w3(s)
= e
w3(s)
,
= e
w3(s)
= exp (c
0
+c
1
s +c
2
ln (cos (K s +L))
+c
3
ln (sec (K s +L) + tan (K s +L))
+c
4
tan (K s +L) +c
5
sec (K s +L))
and the formula for the characteristic function of V is obtained by replacing s by
i s where s R.
Remark 3.5. In the well known (see [13] and [11]) Gaussian case a = 1, b = =
= 0, the dierential equations for w
1
(s), w
2
(s) and w
3
(s) in the proof of Lemma
3.3 are greatly simplied and yield
w
1
(s) = 0 ; w
2
(s) = s ; w
3
(s) =
s
2
2
which implies that
, e
i s V
= , e
i s (S
1
0
+S
0
1
)
= e
s
2
2
(3.11)
i.e. V = S
1
0
+S
0
1
is a Gaussian random variable.
Remark 3.6. In the case when a = 0 we nd that = = = 0, c
1
= , c
2
=
1
2
,
c
3
= c
4
= c
5
= 0, c
0
=
1
2
ln (cos L) and so
w
3
(s) =
1
2
ln (cos L)
1
2
ln (cos (K s +L)) +s
Therefore, by Proposition 3.4, assuming that cos L ,= 0, for all s C such that
cos (i K s +L) ,= 0, the characteristic function of V = b S
2
0
+
b S
0
2
+ S
0
0
+ S
1
1
is
, e
i s V
= (cos L)
1/2
(sec (i K s +L))
1/2
e
i s
which, in the case = 0, is the characteristic function of a continuous binomial
random variable (see [11]).
500 LUIGI ACCARDI AND ANDREAS BOUKAS
Remark 3.7. If b = 0 then the ODEs for w
1
, w
2
, w
3
in the proof of Lemma 3.3 take
the form
w
1
(s) = 2 w
1
(s)
w
2
(s) = w
2
(s) +a + 2 a w
1
(s)
w
3
(s) = + a w
2
(s) +
2
with w
i
(0) = 0 for i 1, 2, 3. Therefore, for ,= 0
w
1
(s) = 0
w
2
(s) =
a
(e
s
1)
w
3
(s) =
[a[
2
2
(e
s
1) +
_
+
2
+
[a[
2
_
s
Therefore, the characteristic function of V = a S
1
0
+ a S
0
1
+ S
0
0
+ S
1
1
is
, e
i s V
= e
|a|
2
2
(e
i s
1)+i
2
+
|a|
2
s
which, for +
2
+
|a|
2
n
a
k
, we consider quantum
random variables of the form
W =
n,k0
_
c
n,k
B
n
k
+ c
n,k
B
k
n
_
where c
n,k
C.
Extending the framework and terminology of [11] and [12] to the innite di-
mensional case, the group element (in terms of coordinates of the rst kind)
e
s W
= e
s
n,k0
(c
n,k
B
n
k
+ c
n,k
B
k
n
)
can be put, through an appropriate splitting lemma, in the form of coordinates of
the second kind
e
s W
=
n,k0
e
f
n,k
(s) B
n
k
e
f
n,k
(s) B
k
n
for some functions f
n,k
(s). Since B
n
k
= 0 for all k ,= 0, after several commuta-
tions, we nd that
e
s W
=
n0
e
wn(s) B
n
0
)
3
= (B
0
1
+B
1
0
)
3
By (3.11) W is the cube of a Gaussian random variable. Using a a
= 1 + a
a
we nd that
(a +a
)
2
= a
2
+a
2
+ 2 a
a + 1
and
(a +a
)
3
= a
3
+ 3 a
+ 3 a
2
a + 3 a + 3 a
a
2
+a
3
Thus
W = a
3
+ 3 a
+ 3 a
2
a + 3 a + 3 a
a
2
+a
3
Lemma 5.1. For all analytic functions f and for all w
i
C, i = 1, 2, 3, 4
[a, f(a
)] = f
(a
) (5.1)
[f(a), a
] = f
(a) (5.2)
Moreover, if a = 0 then for n 1
a
n
f(a
) = f
(n)
(a
) (5.3)
where f
(n)
denotes the n-th derivative of f.
Proof. The proof of (5.1) and (5.2) can be found in [11]. To prove (5.3) we notice
that, by Proposition 2.6, for n 1 and k 0
[a
n
, a
k
] =
l1
_
n
l
_
k
(l)
a
kl
a
nl
Thus, assuming that f(a
) =
k0
c
k
a
k
, we have that
a
n
f(a
) = [a
n
, f(a
)]
=
k0
c
k
[a
n
, a
k
]
=
k0
c
k
l1
_
n
l
_
k
(l)
a
kl
a
nl
k0
c
k
k
(n)
a
kn
= f
(n)
(a
)
= G
n
(a
) e
g(a
where
G
1
(a
) = g
(a
)
G
2
(a
) = g
(a
) +g
(a
)
2
G
3
(a
) = g
(a
) + 3 g
(a
) g
(a
) +g
(a
)
3
Proof. The proof follows from Lemma 5.1 by taking f(a
) = e
g(a
)
.
Corollary 5.3. Let g(a
) =
k=0
w
k
a
k
. Then, for n 1, 2, 3,
a
n
e
g(a
)
= G
n
(a
) e
g(a
where G
1
(a
) =
k=0
(k + 1) w
k+1
a
k
and
G
2
(a
) =
k=0
(k + 1) (k + 2) w
k+2
a
k
+
k,m=0
(k + 1) (m+ 1) w
k+1
w
m+1
a
k+m
G
3
(a
) =
k=0
(k + 1) (k + 2) (k + 3) w
k+3
a
k
+3
k,m=0
(k + 1) (m+ 1) (m+ 2) w
k+1
w
m+2
a
k+m
+
k,m,=0
(k + 1) (m+ 1) ( + 1) w
k+1
w
m+1
w
+1
a
k+m+
Proof. The proof follows directly from Corollary 5.2.
Lemma 5.4. Let s C. Then
e
s W
=
n=0
e
wn(s) a
n
=
n=0
e
wn(s) B
n
0
where w
n
(0) = 0 for all n 0 and the w
n
satisfy the nonlinear innite system of
disentanglement ODEs
w
0
(s) = w
1
(s)
3
+ 3 w
1
(s) (1 + 2 w
2
(s)) + 6 w
3
(s)
w
1
(s) = 3 w
1
(s)
2
(1 + 2 w
2
(s)) + 18 w
3
(s) w
1
(s)
+12 w
2
(s)
2
+ 12 w
2
(s) + 24 w
4
(s) + 3
w
2
(s) = 12 w
2
(s)
2
w
1
(s) + (54 w
3
(s) + 12 w
1
(s)) w
2
(s)
+3 w
1
(s) + 27 w
3
(s) + 36 w
1
(s) w
4
(s)
+9 w
1
(s)
2
w
3
(s) + 60 w
5
(s)
w
3
(s) = 54 w
3
(s)
2
+ (30 w
1
(s) w
2
(s) + 18 w
1
(s)) w
3
(s)
+24 w
4
(s) w
2
(s) + 60 w
1
(s) w
5
(s) + 72 w
2
(s) w
4
(s)
+12 w
1
(s)
2
w
4
(s) + 120 w
6
(s) + 6 w
2
(s) + 48 w
4
(s)
+12 w
2
(s)
2
+ 1
ON THE CHAR. FUNCTION OF RVS ASSOCIATED WITH BOSON LIE ALGEBRAS 503
and for n 4,
w
n
(s) = (n + 1) (n + 2) (n + 3) w
n+3
(s)
+3 (n 1) w
n1
(s) + 3 (n + 1)
2
w
n+1
(s)
+3
k,m0
k+m=n
(k + 1) (m+ 1) (m+ 2) w
k+1
(s) w
m+2
(s)
+
k,m,0
k+m+=n
(k + 1) (m+ 1) ( + 1) w
k+1
w
m+1
w
+1
+3
k,m0
k+m=n1
(k + 1) (m+ 1) w
k+1
(s) w
m+1
(s)
Proof. As in the proof of Lemma 3.3, let
E := e
s (a
3
+3 a
+3 a
2
a+3 a+3 a
a
2
+a
3
)
. (5.4)
Then by the disentanglement assumption for e
s W
, we also have that
E =
n=0
e
wn(s) a
n
(5.5)
where w
n
(0) = 0 for all n 0. Then, (5.5) implies that
s
E =
n=0
w
n
(s) a
n
E (5.6)
and, using Corollary 5.3, from (5.4) we also obtain
s
E = (a
3
+ 3 a
+ 3 a
2
a + 3 a + 3 a
a
2
+a
3
) E
=
k0
(k + 1) (k + 2) (k + 3) w
k+3
(s) a
k
+3
k,m0
(k + 1) (m+ 1) (m+ 2) w
k+1
(s) w
m+2
(s) a
k+m
+
k,m,=0
(k + 1) (m+ 1) ( + 1) w
k+1
w
m+1
w
+1
a
k+m+
+3 a
+ 3
k0
(k + 1) w
k+1
(s) a
k+2
+3
k0
(k + 1) w
k+1
(s) a
k
+ 3
k0
(k + 1) (k + 2) w
k+2
(s) a
k+1
+3
k,m0
(k + 1) (m+ 1) w
k+1
(s) w
m+1
(s) a
k+m+1
+a
3
E (5.7)
and the dierential equations dening the w
n
s are obtained from (5.6) and (5.7)
by equating coecients of the powers of a
.
504 LUIGI ACCARDI AND ANDREAS BOUKAS
Remark 5.5. As in the nite dimensional case, w
0
is determined by straight-
forward integration. However, unlike the nite-dimensional (Schrodinger) case
of Lemma 3.3, in the innite-dimensional case the disentanglement ODEs are cou-
pled, with the ODE for each w
n
depending on w
1
, ..., w
n+3
. The ODEs for w
1
, w
2
,
and w
3
are of pseudo (due to coupling)-Riccati type.
Proposition 5.6. (Characteristic function of W) For all s R
, e
i s W
= e
w0(i s)
where w
0
is as in Lemma 5.4.
Proof. The proof is similar to that of Proposition 3.4.
References
1. Accardi, L., Boukas, A.: Fock representation of the renormalized higher powers of white noise
and the VirasoroZamolodchikovw Lie algebra, J. Phys. A Math. Theor., 41 (2008).
2. Accardi, L., Boukas, A.: Random variables and positive denite kernels associated with
the Schroedinger algebra, Proceedings of the VIII International Workshop Lie Theory and
its Applications in Physics, Varna, Bulgaria, June 16-21, 2009, pages 126-137, American
Institute of Physics, AIP Conference Proceedings 1243.
3. Accardi, L., Boukas, A.: Central extensions and stochastic processes associated with the Lie
algebra of the renormalized higher powers of white noise, with Luigi Accardi, Proceedings
of the 11th workshop: non-commutative harmonic analysis with applications to probability,
Bedlewo, Poland, August 2008, Banach Center Publ. 89 (2010), 13-43.
4. Accardi, L., Boukas, A.: The Fock kernel for the Galilei algebra and associated random
variables, submitted.
5. Accardi, L., Boukas, A., Franz, U.: Renormalized powers of quantum white noise, Innite
Dimensional Analysis, Quantum Probability, and Related Topics, Vol. 9, No. 1 (2006), 129
147.
6. Accardi, L., Dhahri, A.: Polynomial extensions of the Weyl C
OSTLER*
Abstract. We derive the form of the Belavkin-Kushner-Stratonovich equa-
tion describing the ltering of a continuous observed quantum system via
non-demolition measurements when the statistics of the input eld used for
the indirect measurement are in a general coherent state.
1. Introduction
One of the most remarkable consequences of the Hudson-Parthasarathy quan-
tum stochastic calculus [21] is V. P. Belavkins formulation of a quantum theory of
ltering based on non-demolition measurements of an output eld that has inter-
acted with a given system [4, 6, 7, 8]. Specically, we must measure a particular
feature of the eld, for instance a eld quadrature, or the count of the eld quanta,
and this determines a self-commuting, therefore essentially classical, stochastic
process. The resulting equations have structural similarities with the classical
analogues appearing in the work of Kallianpur, Striebel, Kusnher, Stratonovich,
Zakai, Duncan and Mortensen on nonlinear ltering, see [16, 23, 24, 28]. This
showed that the earlier models of repeated quantum photon counting measure-
ments developed by Davies [14, 15] could be realized using a concrete theory: this
was rst shown by taking the pure-jump process limit of diusive quantum ltering
problems [3].
There has been recent interest amongst the physics community in quantum
ltering as an applied technique in quantum feedback and control [1, 2, 10, 11, 12,
17, 19, 22, 26, 27]. An additional driver is the desire to go beyond the situation
of a vacuum eld and derive the lter for other physically important states such
as thermal, squeezed, single photon states, etc. In this note we wish to present
the lter for non-demolition quadrature and photon-counting measurements when
the choice of state for the input eld is a coherent state with intensity function
. The resulting lters are a deformation of the vacuum lters and reduce to the
latter when we take 0, this is perhaps to be expected given that the coherent
states have a continuous-in-time tensor product factorization property. We derive
the lters using the reference probability approach, as well as the characteristic
function approach.
Received 2010-6-14; Communicated by D. Applebaum.
2000 Mathematics Subject Classication. Primary 15A15 ; Secondary 15A09, 15A23.
Key words and phrases. Quantum ltering and estimation, coherent state.
* This research is supported by the Engineering and Physical Sciences Research Council,
project grant EP/G039275/1.
505
Serials Publications
www.serialspublications.com
Communications on Stochastic Analysis
Vol. 4, No. 4 (2010) 505-521
506 JOHN E. GOUGH AND CLAUS K
OSTLER
1.1. Classical Non-linear Filtering. We consider a state based model where
the state X
t
evolves according to a stochastic dynamics and we make noisy obser-
vations Y
t
on the state. The dynamics-observations equations are the SDEs
dX
t
= v (X
t
) dt +
X
(X
t
) dW
proc
t
, (1.1)
dY
t
= h(X
t
) dt +
Y
dW
obs
t
(1.2)
and we assume that the process noise W
proc
and the observation noise W
obs
are
uncorrelated multi-dimensional Wiener processes. The generator of the state dif-
fusion is then
/ = v
i
i
+
1
2
ij
XX
2
ij
,
where
XX
=
X
X
. The aim of ltering theory to obtain a least squares estimate
for the state dynamics. More specically, for any suitable function f of the state,
we would like to evaluate the conditional expectation
t
(f) := E[f (X
t
) [ T
Y
t]
],
with T
Y
t]
being the -algebra generated by the observations up to time t.
1.1.1. Kallianpur-Striebel Formula. By introducing the Kallianpur-Striebel like-
lihood function
L
t
(x[y) = exp
_
t
0
_
h(x
s
)
dy
s
1
2
h(x
s
)
h(x
s
) ds
_
for sample state path x = x
s
: 0 s t conditional on a given sample observa-
tion y = y
s
: 0 s t, we may represent the conditional expectation as
t
(f) =
_
C
x
0
[0,t]
f (x
t
) L
t
(x[y) P[dx]
_
C
x
0
[0,t]
L
t
(x[y) P[dx]
y=Y ()
=
t
(f)
t
(1)
where P is canonical Wiener measure and
t
(f) () =
_
C
x
0
[0,t]
f (x
t
) L
t
(x[Y ()) P[dx] .
1.1.2. Duncan-Mortensen-Zakai and Kushner-Stratonovich Equations. Using the
Ito calculus, we may obtain the Duncan-Mortensen-Zakai equation for the un-
normalized lter
t
(f), and the Kushner-Stratonovich equation for the normalized
version
t
(f). These are
d
t
(f) =
t
(/f) dt +
t
_
fh
_
dY
t
,
d
t
(f) =
t
(/f) dt +
_
t
_
fh
t
(f)
t
_
h
_
dI
t
,
where (I
t
) are the innovations:
dI
t
:= dY
t
t
(h) dt, I (0) = 0.
We note that there exist variants of these equations for more general processes
than diusions (in particular for point processes which will be of relevance for
photon counting), and for the case where the process and observation noises are
correlated.
QUANTUM FILTERING IN COHERENT STATES 507
1.1.3. Pure versus Hybrid Filtering Problems. We remark that we follow the tra-
ditional approach of adding direct Wiener noise W
obs
to the observations. We could
of course consider a more general relation of the form dY
t
= h(X
t
) dt +
Y
dW
obs
t
but for constant coecients
Y
a simple rearrangement returns us to the above
setup.
The situation where we envisage dY
t
= h(X
t
) dt +
Y
(X
t
) dW
obs
t
, with
Y
a
known function of the unobserved state, must be considered as being too good to
be true since we can then obtain information about the unobserved state by just
examining the quadratic variation of the observations process, since we then have
dY (t) dY (t)
=
Y
(X
t
)
Y
(X
t
)
OSTLER
we met so called non-demolition conditions which guarantee that quantum mea-
surement process itself does not destroy the statistical features which we would
like to infer. We will now describe these elements in more detail below.
2.1. Quantum Estimation. We shall now describe the reference probability
approach to quantum ltering. Most of our conventions following the presentation
of Bouten and van Handel [12]. For an alternative account, including historical
references, see [9].
Let A be a von Neumann algebra and E be a normal state. In a given experiment
one may only measure a set of commuting observables Y
: A. Dene the
measurement algebra to be the commutative von Neumann algebra generated by
the chosen observables
M= vNY
: A A.
We may estimate an observable X A from an experiment with measurement
algebra M if and only if
X M
:= A A : [A, Y ] = 0, Y M,
That is, if it is physically possible to measure X in addition to all the Y
. There-
fore the algebra vNX, Y
M
dened by
E[ E[X [ M] Y ] E[XY ], Y M. (2.1)
In contrast to the general situation regarding conditional expectations in the
non-commutative setting of von Neumann algebras [25], this particular denition is
always nontrivial insofar as existence is guaranteed. Introducing the norm|A|
2
:=
E[A
A], we see that the conditional expectation always exists and is unique up to
norm-zero terms. It moreover satises the least squares property
|X E[X [ M]| |X Y |, Y M.
As the set vNX, Y
_
`
?
,
,
X
E[X[M]
M
M
Figure 1. Quantum Conditional Expectation E[[M] as a least
squares projection onto M from its commutant.
The following two lemmas will be used extensively, see [12] and [13].
Lemma 2.2 (Unitary rotations). Let U be unitary and dene
E[X] := E
_
U
XU
and let
M= U
MU. Then
E[U
XU [
M] = U
E[X [ M] U.
Here we think of going from the Schrodinger picture where the state E is xed
and observables evolve to U
with E
_
F
= 1 and set
E
F
[X] := E
_
F
XF
. Then
E
F
[X [ M] =
E
_
F
XF [ M
E[F
F [ M]
.
Proof. For all Y M,
E
_
E
_
F
XF [ M
= E
_
F
XFY
= E
F
[XY ], since [F, Y ] = 0,
= E
F
[E
F
[X [ M] Y ]
= E
_
F
FE
F
[X [ M] Y
, since F M
= E
_
E
_
F
F [ M
E
F
[X [ M] Y
!
2.2. Quantum Stochastic Processes. We begin by reviewing the theory of
quantum stochastic calculus developed by Hudson and Parthasarathy [21] which
gives the mathematical framework with which to generalize the notions of the
classical Ito integration theory.
We take R
+
= [0, ). We shall denote by L
2
symm
(R
n
+
) the space of all square-
integrable functions of n positive variables that are completely symmetric: that
510 JOHN E. GOUGH AND CLAUS K
OSTLER
is, invariant under interchange of any pair of its arguments. The Bose Fock space
over L
2
(R
+
) is then the innite direct sum Hilbert space
F :=
n=0
L
2
symm
(R
n
+
)
with the n = 0 space identied with C. An element of F is then a sequence
= (
n
)
n=0
with
n
L
2
symm
(R
n
+
) and ||
2
=
n=0
_
[0,)
n
[
n
(t
1
, , t
n
) [
2
dt
1
dt
n
< . Moreover, the Fock space has inner product
[ =
n=0
_
[0,)
n
n
(t
1
, , t
n
)
(t
1
, , t
n
) .
Th physical interpretation is that = (
n
)
n=0
F describes the state of a quan-
tum eld consisting of an indenite number of indistinguishable (Boson) particles
on the half-line R
+
. A simple example is the vacuum vector dened by
:= (1, 0, 0, )
clearly corresponding to no particles. (Note that the no-particle state is a genuine
physical state of the eld and is not just the zero vector of F!) An important class
of vectors are the coherent states () dened by
[()]
n
(t
1
, , t
n
) := e
2 1
n!
(t
1
) (t
n
),
for L
2
(R
+
). (The n = 0 component understood as e
2
.) The vacuum then
corresponds to (0).
For each t > 0 we dene the operators of annihilation B(t), creation B
(t) and
gauge (t) by
[B(t) ]
n
(t
1
, , t
n
) :=
n + 1
_
t
0
n+1
(s, t
1
, , t
n
) ds,
[B
(t) ]
n
(t
1
, , t
n
) :=
1
n
n
j=1
1
[0,t]
(t
j
)
n1
_
t
1
, ,
t
j
, , t
n
_
,
[(t) ]
n
(t
1
, , t
n
) :=
n
j=1
1
[0,t]
(t
j
)
n
(t
1
, , t
n
) .
The creation and annihilation process are adjoint to each other and the gauge is
self-adjoint. We may dene a eld quadrature by
Q(t) = B(t) +B
(t)
and this yields a quantum stochastic process which is essentially classical in the
sense that it is self-adjoint and self-commuting, that is [Q(t) , Q(s)] = 0 for
all t, s 0. We remark that for each [0, 2) we may dene quadratures
Q
(t) = e
i
B(t) + e
i
B
0
k(t)dQ(t)
= e
1
2
0
k(t)
2
dt
.
QUANTUM FILTERING IN COHERENT STATES 511
Table 1. Quantum Ito Table
dB d dB
dt
dB 0 dB dt 0
d 0 d dB
0
dB
0 0 0 0
dt 0 0 0 0
.
We also note that (t) : t 0 is also an essentially classical process and for
the choice of a coherent state yields a non-homogeneous Poisson process: for real
k ()
() [ e
i
0
k(t)d(t)
() = exp
_
0
[ (t) [
2
_
e
ik(t)
1
_
dt.
We consider a quantum mechanical system with Hilbert space h being driven by
an external quantum eld input. the quantum eld will be modeled as an idealized
Bose eld with Hilbert space
_
L
2
(R
+
, dt)
_
which is the Fock space over the one-
particle space L
2
(R
+
, dt). Elements of the Fock space may be thought of as vectors
=
n=0
n
where
n
=
n
(t
1
, , t
n
) is a completely symmetric functions with
n=0
_
[0,)
n
[
n
(t
1
, , t
n
) [
2
dt
1
dt
n
< .
The Hudson-Parthasarathy theory of quantum stochastic calculus gives a gen-
eralization of the Ito theory of integration to construct integral processes with
respect to the processes of annihilation, creation, gauge and, of course, time. This
leads to the quantum Ito table 1.
We remark that the Fock space carries a natural ltration in time obtained
from the decomposition F
= F
t]
F
(t
into past and future subspaces: these are
the Fock spaces over L
2
[0, t] and L
2
(t, ) respectively.
2.3. Continuous-Time Quantum Stochastic Evolutions. On the joint space
h F, we consider the quantum stochastic process V () satisfying the QSDE
dV (t) =
_
(S I) d(t) +L dB
(t)
L
S dB(t) (
1
2
L
L +iH) dt
_
V (t),
with V (0) = 1, and where S is unitary, L is bounded and H self-adjoint. This
specic form of QSDE may be termed the Hudson-Parthasarathy equation as the
algebraic conditions on the coecients are necessary and sucient to ensure uni-
tarity (though the restriction for L to be bounded can be lifted). The process
is also adapted in the sense that for each t > 0, V (t) acts non-trivially on the
component h F
t]
and trivially on F
(t
.
2.3.1. The Heisenberg-Langevin Equations. For a xed system operator X we set
j
t
(X) := V
(t))
+j
t
(/
01
X) dB(t) +j
t
(/
00
X) dt (2.3)
512 JOHN E. GOUGH AND CLAUS K
OSTLER
input
system
output
Figure 2. Input - Output component
where the Evans-Hudson maps /
XS X,
/
10
X = S
[X, L],
/
01
X = [L
, X]S
/
00
X = /
(L,H)
and in particular /
00
takes the generic form of a Lindblad generator:
/
(L,H)
=
1
2
L
[X, L] +
1
2
[L
out
(t) := V
S)dB
(t) +j
t
(S
L)dB(t) +j
t
(L
L)dt. (2.6)
2.3.3. The Measurement Algebra. We wish to consider the problem of continu-
ously measuring a quantum stochastic process associated with the output eld.
we shall chose to measure an observable process of the form
Y
out
(t) := V (t)
[I Y
in
(t)]V (t) (2.7)
which corresponds to a quadrature of the eld when
Y
in
(t) = Q(t) = B
(t) +B(t),
or counting the number of output photons when
Y
in
(t) = (t).
We introduce von Neumann algebra
Y
in
t]
= vN
_
Y
in
(s) : 0 s t
_
,
and dene the measurement algebra up to time t to be
Y
out
t]
= vN
_
Y
out
(s) : 0 s t
_
V (t)
Y
in
t]
V (t) . (2.8)
Note that both algebras are commutative:
_
Y
out
(t) , Y
out
(s)
= V (T)
_
I
_
Y
in
(t) , Y
in
(s)
_
V (T) = 0
QUANTUM FILTERING IN COHERENT STATES 513
for T = ts. The family
_
Y
out
t]
: t 0
_
then forms an increasing family (ltration)
of von Neumann algebras.
2.3.4. The Non-Demolition Property. The system observables may be estimated
from the current measurement algebra
j
t
(X)
_
Y
out
t]
_
. (2.9)
The proof follows from the observation that for t s
_
j
t
(X) , Y
out
(s)
= V (t)
_
X I, I Y
in
(s)
V (t) = 0.
2.4. Constructing The QuantumFilter. The ltered estimate for j
t
(X) given
the measurements of the output eld is then
t
(X) := E
_
j
t
(X) [ Y
out
t]
_
.
Let
E
t
[X] = E[j
t
(X)], then by lemma 2.2
t
(X) = E
_
j
t
(X) [ Y
out
t]
_
= V (t)
E
t
_
X[Y
in
t]
_
V (t) .
2.4.1. Reference Probability Approach. Suppose that there is an adapted process
F () such that F (t)
_
Y
in
t]
_
and
E
t
[X[] = E
_
F (t)
(X 1) F (t)
_
for all system
operators X, then by lemma 2.3
t
(X) = E
_
j
t
(X) [ Y
out
t]
_
= V (t)
E
t
_
X [ Y
in
t]
_
V (t)
= V (t)
E
_
F (t)
(X 1) F (t) [ Y
in
t]
_
E
_
F (t)
F (t) [ Y
in
t]
_ V (t)
This is essentially a non-commutative version of the Girsanov transformation from
stochastic analysis. The essential feature is that the transformation operators F (t)
giving the change of representation for the expectation lie in the commutant of
the measurement algebra up to time t.
We therefore obtain an operator-valued Kallianpur-Striebel relation
t
(X) =
t
(X)
t
(1)
, (2.10)
which may be called the quantum Kallianpur-Striebel, where
t
(X) := V (t)
E
_
F (t)
(X 1) F (t) [ Y
in
t]
_
V (t) . (2.11)
514 JOHN E. GOUGH AND CLAUS K
OSTLER
3. Coherent State Filters
We shall consider the class of states
E
[ ] =
(3.1)
of the form
= () (3.2)
where is a normalized vector in the system Hilbert space and () is the coherent
state with test function L
2
[0, ). We note that
dB(t) () = (t) dt () ,
d(t) () = (t) dB
(t) () .
We see that
E
[dj
t
(X)] = E
[j
t
_
/
(t)
X
_
]dt (3.3)
where
/
(t)
X = /
00
X + (t)
/
10
X + (t) /
01
X +[ (t) [
2
/
11
X (3.4)
The generator is again of Lindblad form and in particular we have
/
(t)
/
(L
(t)
,H)
with
L
(t)
= S (t) +L. (3.5)
e may dene a parameterized family of density matrices on the system by setting
tr
h
t
X = E
[j
t
(X)], in which case we deduce the master equation
t
= /
(t)
() ,
where the adjoint is dened through the duality tr
h
/X =tr
h
/
X.
From the input-output relation for the eld
dB
out
= j
t
(S) dB(t) +j
t
(L) dt
we obtain the average
E
[dB
out
] = j
t
(S) (t) +j
t
(L) dt
= j
t
_
L
(t)
_
dt.
3.1. Quadrature Measurement. We take Y
in
(t) = B(t) + B
(t) which is a
quadrature of the input eld. Setting (t) = V (t) we have that
d (t) = ((S 1) (t) +L) dB
(t) (t) (L
S (t) +
1
2
L
_
L
S (t) +
1
2
L
L
t
dY
in
(t) (t) +
K
t
dt (t) ,
where
L
t
= L + (S I) (t) = L
(t)
(t) ,
K
t
= L
S (t)
1
2
L
L iH L
(t)
(t) + (t)
2
.
It follows that (t) F (t) where F (t) is the adapted process satisfying the
QSDE
dF (t) =
L
t
dY
in
(t) F (t) +
K
t
dt F (t) , F (0) = I.
Moreover F (t) is in the commutant of Y
in
t]
and therefore allows us to perform the
non-commutative Girsanov trick.
From the quantum Ito product rule we then see that
d [F
(t) XF (t)] = F
(t)
_
X
L
t
+
L
t
X
_
F (t) dY
in
(t)
+F
(t)
_
t
X
L
t
+X
K
t
+
K
t
X
_
F (t) dt
and this leads to the SDE for the un-normalized lter
d
t
(X) =
t
_
X
L
t
+
L
t
X
_
dY
out
(t)
+
t
_
t
X
L
t
+X
K
t
+
K
t
X
_
dt.
After a small bit of algebra, this may be written in the form
d
t
(X) =
t
_
X
L
t
+
L
t
X
_
_
dY
out
(t)
_
(t) + (t)
_
dt
+
t
_
/
(t)
X
_
dt.
This is the quantum Zakai equation for the lter based on continuous measurement
of the output eld quadrature.
To obtain the quantum Kushner-Stratonovich equation we rst observe that
the normalization satises the SDE
d
t
(I) =
t
_
L
t
+
L
t
_
_
dY
out
(t)
_
(t) + (t)
_
dt
t
(I)
=
t
_
L
t
+
L
t
_
t
(I)
2
_
dY
out
(t)
_
(t) + (t)
_
dt
t
_
L
t
+
L
t
_
2
t
(I)
3
dt.
The product rule then allows us to determine the SDE for
t
(X) =
t
(X)
t
(I)
:
d
t
(X) =
t
_
/
(t)
X
_
+dt
_
t
_
X
L
t
+
L
t
X
_
t
(X)
t
_
L
t
+
L
t
__
dI (t) ,
516 JOHN E. GOUGH AND CLAUS K
OSTLER
where the innovations process satises
dI (t) = dY
out
(t)
_
t
_
L
t
+
L
t
_
+ (t) + (t)
_
dt.
We note that the innovations is martingale with respect to ltration generated
by the output process for the choice of probability measure determined by the
coherent state.
3.1.1. The Quadrature Measurement Filter for a Coherent state. In it convenient
to write the ltering equations in terms of the operators L
(t)
. The result is the
Belavkin-Kushner-Stratonvich equation for the ltered estimate based on optimal
estimation of continuous non-demolition eld-quadrature measurements in a co-
herent state ().
d
t
(X) =
t
_
/
(t)
X
_
dt
+
_
t
_
XL
(t)
+L
(t)
X
_
t
(X)
t
_
L
(t)
+L
(t)
__
dI
quad
(t) ,
(3.7)
with the innovations
dI
quad
(t) = dY
out
(t)
t
_
L
(t)
+L
(t)
_
dt. (3.8)
3.2. Photon Counting Measurement. For convenience we shall derive the
lter based on measuring the number of output photons under the assumption
that the function is bounded away from zero. We discuss how this restriction
can be removed later. We now set Y
in
(t) = (t). We again seek to construct an
adapted process F (t) in the commutant of Y
in
t]
such that (t) = F (t) . We start
with 3.6 again but now note that
dB
(t) (t) =
1
(t)
d(t) (t)
and making this substitution gives
d (t) =
1
(t)
L
t
dY
in
(t) (t)
_
1
2
L
L +iH +L
S (t)
_
dt (t) .
We are then lead to the Zakai equation
d
t
(X) =
1
[ (t) [
2
t
_
t
X
L
t
+(t)
L
t
+
L
t
X (t)
_
dY
out
(t)
t
_
1
2
XL
L +
1
2
L
LX +i [X, H] XL
S (t) (t)
LX
_
dt
which may be rearranged as
d
t
(X) =
t
_
/
X
_
dt
+
1
[ (t) [
2
t
_
t
X
L
t
+(t)
L
t
+
L
t
X (t)
_
_
dY
out
(t) [ (t) [
2
dt
.
QUANTUM FILTERING IN COHERENT STATES 517
To determine the normalized lter, we note that
d
t
(I) =
1
[ (t) [
2
t
_
t
L
t
+(t)
L
t
+
L
t
(t)
_
_
dY
out
(t) [ (t) [
2
dt
and that
d
1
t
(I)
=
t
_
t
L
t
+(t)
L
t
+
L
t
(t)
_
t
(I)
2
dt
t
_
t
L
t
+(t)
L
t
+
L
t
(t)
_
t
(I)
_
t
_
t
L
t
+(t)
L
t
+
L
t
(t)
_
+[ (t) [
2
_dY
out
(t) .
Applying the Ito product formula yields the quantum analogue of the Kushner-
Stratonovich equation for the normalized lter;
d
t
(X) =
t
_
/
X
_
dt +
1
t
_
t
L
t
+(t)
L
t
+
L
t
(t)
_
+[ (t) [
2
dI (t)
t
(
t
X
L
t
+(t)
L
t
+
L
t
X (t))
t
(X)
t
(
t
L
t
+(t)
L
t
+
L
t
(t))
_
and the innovations process is now
dI (t) = dY
out
(t)
_
t
_
t
L
t
+(t)
L
t
+
L
t
(t)
_
+[ (t) [
2
_
dt.
We note that the innovations is again a martingale with respect to ltration gen-
erated by the output process for the choice of probability measure determined by
the coherent state.
The derivation above relied on the assumption that (t) ,= 0, however this is
not actually essential. In the case of a vacuum input, it is possible to apply an
additional rotation W (t) satisfying dW (t) = [z
dB(t) zdB
(t)
1
2
[z[
2
dt]W (t) ,
with W (0) = I, and apply the reference probability technique to the von Neumann
algebra generated by N (t) = W (t) (t) W (t)
t
_
L
(t)
XL
(t)
_
t
_
L
(t)
L
(t)
_
t
(X)
_
dI
num
(t) , (3.9)
with the innovations
dI
num
(t) = dY
out
(t)
t
_
L
(t)
L
(t)
_
dt. (3.10)
518 JOHN E. GOUGH AND CLAUS K
OSTLER
4. Characteristic Function Approach
As an alternative to the reference probability approach, we apply a method
based on introducing a process C (t) satisfying the QSDE
dC (t) = f (t) C (t) dY (t) , (4.1)
with initial condition C (0) = I. Here we assume that f is integrable, but otherwise
arbitrary. This approach is a straightforward extension of a classical procedure
and as far as we are aware was rst used in the quantum domain by Belavkin [4].
The technique is to make an ansatz of the form
d
t
(X) = T
t
(X) dt +1
t
(X) dY (t) (4.2)
where we assume that the processes T
t
(X) and 1
t
(X) are adapted and lie in Y
t]
.
These coecients may be deduced from the identity
E[(
t
(X) j
t
(X)) C (t)] = 0
which is valid since C (t) Y
t]
. We note that the Ito product rule implies I +
II +III = 0 where
I = E[(d
t
(X) dj
t
(X)) C (t)] ,
II = E[(
t
(X) j
t
(X)) dC (t)] ,
III = E[(d
t
(X) dj
t
(X)) dC (t)] .
We illustrate how this works in the case of quadrature and photon counting in
a coherent state. For convenience of notation we shall write S
t
for j
t
(S), etc.
4.1. Quadrature Measurement. Here we have
dY (t) = S
t
dB(t) +S
t
dB(t)
+ (L
t
+L
t
) dt
so that
I = E
[T
t
(X) C (t) +1
t
(X) (S
t
t
+S
t
+L
t
+L
t
) C (t)] dt
E
__
(/
00
X)
t
+ (/
01
X)
t
t
+ (/
10
X)
t
t
+ (/
11
X)
t
[
t
[
2
_
C (t)
_
dt,
II = E
[(
t
(X) X
t
) f (t) C (t) (S
t
t
+S
t
+L
t
+L
t
)] dt,
III = E
[1
t
(X) (/
01
X)
t
S
t
(/
11
X)
t
S
t
f (t) C (t)] dt.
Now from the identity I +II +III = 0 we may extract separately the coecients
of f (t) C (t) and C (t) as f (t) was arbitrary to deduce
t
((
t
(X) X
t
) (S
t
t
+S
t
+L
t
+L
t
))
+
t
(1
t
(X) (/
01
X)
t
S
t
(/
11
X)
t
S
t
) = 0,
t
_
T
t
(X) +1
t
(X) (S
t
t
+S
t
+L
t
+L
t
)
_
/
(t)
X
_
t
_
= 0.
QUANTUM FILTERING IN COHERENT STATES 519
Using the projective property of the conditional expectation (
t
t
=
t
) and the
assumption that T
t
(X) and 1
t
(X) lie in Y
t]
, we nd after a little algebra that
1
t
(X) =
t
_
XL
(t)
+L
(t)
X
_
t
(X)
t
_
L
(t)
+L
(t)
_
,
T
t
(X) =
t
_
/
(t)
X
_
1
t
(X)
t
_
L
(t)
+L
(t)
_
,
so that the equation (4.2) reads as
d
t
(X) =
t
_
/
(t)
X
_
dt +1
t
(X)
_
dY (t)
t
_
L
(t)
+L
(t)
_
dt
_
.
4.2. Photon Counting Measurement. We now have
dY (t) = d(t) +L
t
S
t
dB(t) +S
t
L
t
dB(t)
+L
t
L
t
dt
so that
I = E
__
T
t
(X) +1
t
(X)
_
[
t
[
2
+L
t
S
t
t
+S
t
L
t
t
+L
t
L
t
__
C (t)
_
dt
E
__
_
/
(t)
X
_
t
_
C (t)
_
dt
II = E
_
(
t
(X) X
t
) f (t) C (t)
_
[
t
[
2
+L
t
S
t
t
+S
t
L
t
t
+L
t
L
t
__
dt,
III = E
_
1
t
(X)
_
[
t
[
2
+L
t
S
t
t
+S
t
L
t
t
+L
t
L
t
_
f (t) C (t)
_
dt
E
[(/
01
X)
t
S
t
L
t
+ (/
01
X)
t
t
f (t) C (t)] dt
E
__
(/
11
X)
t
[
t
[
2
+ (/
11
X)
t
S
t
L
t
t
_
f (t) C (t)
_
dt.
This time, the identity I +II +III = 0 implies
t
_
(
t
(X) X
t
) L
(t)
t
L
(t)
t
+1
t
(X) L
(t)
t
L
(t)
t
_
t
_
(/
01
X)
t
S
t
L
(t)
t
+ (/
11
X)
t
S
t
L
(t)
t
t
_
= 0,
t
_
T
t
(X) +1
t
(X) L
(t)
t
L
(t)
t
_
/
(t)
X
_
t
_
= 0.
Again, after a little algebra, we nd that
1
t
(X) =
t
_
L
(t)
XL
(t)
_
t
_
L
(t)
L
(t)
_
t
(X) ,
T
t
(X) =
t
_
/
(t)
X
_
1
t
(X)
t
_
L
(t)
L
(t)
_
,
so that the equation (4.2) reads as
d
t
(X) =
t
_
/
(t)
X
_
dt +1
t
(X)
_
dY (t)
t
_
L
(t)
L
(t)
_
dt
_
.
In both cases, the form of the lter is identical to what we found using the
reference probability approach.
520 JOHN E. GOUGH AND CLAUS K
OSTLER
5. Conclusion
Both the quadrature lter (3.7) and the photon counting lter (3.9) take on the
same form as in to the vacuum case and of course reduce to these lters when we
set 0. In both cases it is clear that averaging over the output gives
E
[d
t
(X)] = E
_
j
t
_
/
(t)
X
__
dt
which is clearly the correct unconditioned dynamics in agreement with (3.3), and
we obtain the correct master equation.
It is worth commenting on the fact that the pair of equations now replacing
the dynamical and observation relations (1.1,1.2) are the Heisenberg-Langevin
equation (2.3) and the appropriate component of the input-output relation (2.6).
The process and observation noise have the same origin however the nature of
the quantum ltering based on a non-demolition measurement scheme results in
a set of equations that resemble the uncorrelated classical Kushner-Stratonovich
equations.
5.1. Is Quantum Filtering still a Pure Filtering Problem? The form of the
input-output relations (2.6) might suggest that it is possible to learn something
about the system dynamics by examining the quadratic variation of the output
process, however, this is not the case! We in fact have an enforced too good to be
true situation here as the output elds satisfy the same canonical commutation
relations as the inputs with the result that the quantum Ito table for the output
processes B
out
, B
out
and
out
has precisely the same structure as table 1. Therefore
we always deal with a pure ltering theory in the quantum models considered here.
Acknowledgment. The authors gratefully acknowledge the support of the UK
Engineering and Physical Sciences Research Council under grant EP/G039275/1.
References
1. Armen, M. A., Au, J. K, Stockton, J. K., Doherty A. C., and Mabuchi, H.: Adaptive
homodyne measurement of optical phase. Phys. Rev. Lett., 89:133602, (2002).
2. Barchielli, A.: Direct and heterodyne detection and other applications of quantum stochastic
calculus to quantum optics. Quantum Opt., 2 (1990) 423441.
3. Barchielli, A., Belavkin, V. P.: Measurements continuous in time and posterior states in
quantum mechanics, J. Phys. A, Math Gen, 24 (12), 14951514 (1991)
4. Belavkin, V. P.: Quantum ltering of Markov signals with white quantum noise. Radiotech-
nika i Electronika, 25 (1980) 14451453.
5. Belavkin, V. P.: Stochastic calculus of input-output processes and non-demolition lter-
ing, Reviews of the Newest Achievements in Science and technology, Current Problems of
Mathematics VINITI, Ed. A.S. Holevo, 36, 2967, (1989)
6. Belavkin, V. P.: Quantum continual measurements and a posteriori collapse on CCR. Com-
mun. Math. Phys., 146 (1992), 611635.
7. Belavkin, V. P.: Quantum stochastic calculus and quantum nonlinear ltering.Journal of
Multivariate Analysis, 42 (1992) 171201.
8. Belavkin, V. P.: Quantum stochastic positive evolutions: Characterization, construction,
dilation. Commun. Math. Phys., 184 (1997) 533566.
9. Belavkin, V. P., Edwards, S. C.: Quantum Filtering and Optimal Control, in Quantum
Stochastics and Information, Eds. V.P. Belavkin and M. Gut a, pp. 143205, World Scientic
(2008)
QUANTUM FILTERING IN COHERENT STATES 521
10. Bouten, L. M, Edwards, S. C., and Belavkin, V. P.: Bellman equations for optimal feedback
control of qubit states. J. Phys. B, At. Mol. Opt. Phys., 38:151160, 2005.
11. Bouten, L. M., Gut a, M. I., and Maassen, H.: Stochastic Schrodinger equations.J. Phys. A,
37 (2004) 31893209.
12. Bouten, L., van Handel, R.: Quantum ltering: a reference probability approach,
arXiv:math-ph/0508006.
13. Bouten, L., van Handel, R., and James, M. R.: An introduction to quantum ltering SIAM
J. Control Optim. 46 (2007) 21992241.
14. Davies, E. B.: Quantum stochastic processes. Commun. Math. Phys. 15 (1969) 277304.
15. Davies, E. B.: Quantum Theory of Open Systems. Academic Press, London New-York San
Francisco, 1976.
16. Davis, M. H. A, and Marcus, S. I.: An introduction to nonlinear ltering. In M. Hazewinkel
and J. C. Willems, editors, Stochastic Systems: The Mathematics of Filtering and Identi-
cation and Applications, pages 5375. D. Reidel, 1981.
17. Doherty, A. C., Habib, S., Jacobs, K., Mabuchi, H., and Tan, S. M.: Quantum feedback and
classical control theory. Phys. Rev. A, (2000) 62:012105.
18. Geremia, J. M., Stockton, J. K., Doherty, A. C., and Mabuchi, H.: Quantum Kalman ltering
and the Heisenberg limit in atomic magnetometry. Phys. Rev. Lett., 91:250801, (2003).
19. Gough, J. E., Belavkin, V. P., and Smolyanov, O. G.: Hamilton-Jacobi-Bellman equations
for quantum ltering and control. J. Opt. B: Quantum Semiclass. Opt. 7 (2005) S237S244.
20. Holevo, A. S.: Quantum stochastic calculus. J. Soviet Math., 56 (1991) 26092624. Transla-
tion of Itogi Nauki i Tekhniki, Ser. Sovr. Prob. Mat. 36 (1990), 328.
21. Hudson, R. L., and Parthasarathy, K. R.: Commun.Math.Phys. 93 (1984) 301323.
22. James, M. R.: Risk-sensitive optimal control of quantum systems. Phys. Rev. A, 69:032108,
(2004).
23. Kushner, H. J.: Jump-diusion approximations for ordinary dierential equations with wide-
band random right hand sides. SIAM J. Control Optim. 17 (1979)729744, .
24. Kushner, H. J.: Diusion approximations to output processes of nonlinear systems with
wideband inputs and applications. IEEE Trans. Inf. Th. 26 (1980) 715725.
25. Takesaki, M.: Conditional expectations in von Neumann algebras. J. Funct. Anal. 9 (1971)
306321.
26. Van Handel, R., Stockton, J. K. and Mabuchi, H.: Feedback control of quantum state
reduction. IEEE Transactions on Automatic Control, 50 (2005) 768780, .
27. Wiseman, H. M., and Milburn, G. J.: Quantum theory of eld-quadrature measurements.
Phys. Rev. A, 47( 1993) 642662.
28. Zakai, M.: On the optimal ltering of diusion processes. Z. Wahrsch. th. verw. Geb., 11
(1969) 230243, .
John E. Gough: Institute of Mathematics and Physics, Aberystwyth University,
SY23 3BZ, Wales, United Kingdom
E-mail address: jug@aber.ac.uk
Claus K ostler: Institute of Mathematics and Physics, Aberystwyth University,
SY23 3BZ, Wales, United Kingdom
E-mail address: cck@aber.ac.uk
MARKOV CHAINS AND DYNAMICAL SYSTEMS:
THE OPEN SYSTEM POINT OF VIEW
ST
EPHANE ATTAL*
Abstract. This article presents several results establishing connections be-
tween Markov chains and dynamical systems, from the point of view of open
systems in physics. We show how all Markov chains can be understood as
the information on one component that we get from a dynamical system on a
product system, when losing information on the other component. We show
that passing from the deterministic dynamics to the random one is character-
ized by the loss of algebra morphism property; it is also characterized by the
loss of reversibility. In the continuous time framework, we show that the solu-
tions of stochastic dierential equations are actually deterministic dynamical
systems on a particular product space. When losing the information on one
component, we recover the usual associated Markov semigroup.
1. Introduction
This article aims at exploring the theory of Markov chains and Markov processes
from a particular point of view. This point of view is very physical and commonly
used in the theory of open systems. Open systems are physical systems, in classical
or in quantum mechanics, which are not closed, that is, which are interacting with
another system. In general the system we are interested in is small (for example,
it has only a nite number of degrees of freedom), whereas the outside system is
very large (often called theenvironment, it may be a heat bath typically).
This is now a very active branch of research to study such systems coupled to an
environment. In classical mechanics they are used to study conduction problems
(Fouriers law for example, see [3], [5]) but more generally out of equilibrium dy-
namics (see [14], [4]). In quantum mechanics, open systems appear fundamentally
for the study of decoherence phenomena (see [8]), but also it is the basis of quan-
tum communication (see [11]). Problems of dissipation, heat conduction, out of
equilibrium dynamics in quantum mechanics (see [9], [10]) lead to very important
problems which are mostly not understood at the time we write this article.
The aim of this article is to make clear several ideas and connections between
deterministic dynamics of closed systems, eective dynamics of open systems and
Markov processes.
Received 2010-7-9; Communicated by D. Applebaum.
2000 Mathematics Subject Classication. Primary 37A50, 60J05, 60J25, 60H10; Secondary
37A60, 82C10.
Key words and phrases. Markov chains, dynamical systems, determinism, open systems, sto-
chastic dierential equations, Markov processes.
* Work supported by ANR project HAM-MARK N
ANR-09-BLAN-0098-01.
523
Serials Publications
www.serialspublications.com
Communications on Stochastic Analysis
Vol. 4, No. 4 (2010) 523-540
524 ST
EPHANE ATTAL
Surprisingly enough these ideas are made rather clear in the literature when
dealing with the quantum systems, but not that much with classical ones! Indeed,
it is common in quantum mechanics to consider a bipartite system on which one
component is not accessible (it might be an environment which is too complicated
to be described, or it might be Bob who is sharing the photons of a correlated
pair with Alice, in Quantum Information Theory). It is well-known that, tracing
out over one component of the system, the unitary Schrodinger dynamics becomes
described by completely positive maps, in discrete time, or completely positive
semigroups, in continuous time.
In [2], for example, the authors show how every discrete time semigroup of
completely positive maps can be described by a realistic physical system, called
Repeated Quantum Interactions. They show that in the continuous time limit
these Hamiltonian dynamics spontaneously converge to a dynamics described by
a quantum Langevin equation.
In this article we establish many similar results in the context of classical dy-
namical systems and Markov chains. The article is structured as follows. In
Section 2, we show that Markov chains appear from any dynamical system on a
product space, when averaging out one of the two components. This way, Markov
chains are interpreted as what remains on one system when it interacts with some
environment but we do not have access to that environment. The randomness
appears directly for the determinism, solely by the fact that we have lost some
information. We show that any Markov chain can be obtained this way. We also
show two results which characterize what properties are lost when going from a
deterministic dynamical system to a Markov chain: typically the loss of algebra
morphism property and the loss of reversibility.
In Section 3 we explore the context of classical Markov process in the continuous
time setup. We actually concentrate on stochastic dierential equations. Despite
their random character, we show that stochastic dierential equations are ac-
tually deterministic dynamical systems. They correspond to a natural dynamical
system which is used to dilate some Markov processes into a deterministic dy-
namics. The role of the environment is played by the canonical probability space
(here the Wiener space), the action of the environment is the noise term in the
stochastic dierential equation.
2. Markov Chains and Dynamical Systems
2.1. Basic Denitions. Let us recall some basic denitions concerning dynam-
ical systems and Markov chains.
Let (E, c) be a measurable space. Let
T be a measurable function from E to
E. We then say that
T is a dynamical system on E. Such a mapping
T induces a
natural mapping T on /
(E) dened by
Tf(x) = f(
Tx) .
Note that this mapping clearly satises the following properties (proof left to the
reader).
MARKOV CHAINS AND DYNAMICAL SYSTEMS 525
Proposition 2.1.
(1) T is a -homomorphism of the -algebra /
(E),
(2) T(1l
E
) = 1l
E
,
(3) |T| = 1.
What is called dynamical system is actually the associated discrete-time semi-
group (
T
n
)
nN
, when acting on points, or (T
n
)
nN
, when acting on functions.
When the mapping
T is invertible, then so is the associated operator T. The
semigroups (
T
n
)
nN
and (T
n
)
nN
can then be easily extended into one-parameter
groups (
T
n
)
nZ
and (T
n
)
nZ
, respectively.
Let us now recall basic denitions concerning Markov chains. Let (E, c) be a
measurable space. A mapping from E c to [0, 1] is a Markov kernel if
(a) x (x, A) is a measurable function, for all A c,
(b) A (x, A) is a probability measure, for all x E.
When E is a nite set, then is determined by the quantities
P(i, j) = (i, j)
which form a stochastic matrix, i.e. a square matrix with positive entries and sum
of each row being equal to 1.
In any case, such a Markov kernel acts on /
(E) as follows:
f(x) =
_
E
f(y) (x, dy) .
A linear operator T on /
1
2
(x, A) =
_
E
2
(y, A)
1
(x, dy) .
This kernel represents the Markov kernel resulting from making a rst step fol-
lowing
1
and then another step following
2
.
A Markov chain with state space (E, c) is a discrete-time stochastic process
(X
n
)
nN
dened on a probability space (, T, P) such that each X
n
: E is
measurable and
E[f(X
n+1
) [ X
0
, X
1
, . . . , X
n
] = E[f(X
n+1
) [ X
n
]
526 ST
EPHANE ATTAL
for all bounded function f : E R and all n N. In particular, if T
n
denotes
the -algebra generated by X
0
, X
1
, . . . , X
n
, then the above implies
E[f(X
n+1
) [ T
n
] = L
n
f(X
n
)
for some function L
n
f. The Markov chain is homogeneous if furthermore L
n
does
not depend on n. We shall be interested only in this case and we denote by L this
unique value of L
n
:
E[f(X
n+1
) [ T
n
] = Lf(X
n
) . (2.1)
Applying successive conditional expectations, one gets
E[f(X
n
) [ T
0
] = L
n
f(X
0
) .
If (x, dy) denotes the conditional law of X
n+1
knowing X
n
= x, which coincides
with the conditional law of X
1
knowing X
0
= x, then is a Markov kernel and
one can easily see that
Lf(x) =
_
E
f(y) (x, dy) = f(x) .
Hence L is the Markov operator associated to .
With our probabilistic interpretation we get easily that f(x) is the expecta-
tion of f(X
1
) when X
0
= x almost surely. The measure P is the distribution
of X
1
if the distribution of X
0
is P.
We complete this section with the following nal denition. A Markov kernel
is said to be deterministic if for all x E the measure (x, ) is a Dirac mass.
This is to say that there exists a measurable mapping
T : E E such that
(x, dy) =
T(x)
(dy) .
In other words, the Markov chain associated to is not random at all, it maps
with probability 1, each point x to
T(x): it is a dynamical system.
2.2. Reduction of Dynamical Systems. Now consider two measurable spaces
(E, c) and (F, T), together with a dynamical system
T on E F, equipped with
the product -eld. As above, consider the lifted mapping T acting on /
(EF).
For any bounded measurable function f on E, we consider the bounded (mea-
surable) function f 1l on E F dened by
(f 1l)(x, y) = f(x) ,
for all x E, y F.
Assume that (F, T) is equipped with a probability measure . We shall be
interested in the mapping L of /
(E) dened by
Lf(x) =
_
F
T(f 1l)(x, y) d(y) =
_
F
(f 1l)
(E F) /
(E F)
1l
_
/
(E) /
(E) .
L
528 ST
EPHANE ATTAL
In more physical language, what we have obtained here can be interpreted in two
dierent ways. If we think of the dynamical system
T rst, we have emphasized
the fact that losing the information of a deterministic dynamics on one of the
components creates a random behavior on the other component. The randomness
here appears only as a lack of knowledge of deterministic behavior on a larger
world. A part of the universe interacting with our system E is inaccessible to us
(or at least we see a very small part of it: an average) which results in random
behavior on E.
In the converse direction, that is, seen from the Markov kernel point of view,
what we have obtained is a dilation of a Markov transition kernel into a dynamical
system. Consider the kernel L on the state space E. It does not represent the
dynamics of a closed system, it is not a dynamical system. In order to see L as
coming from a true dynamical system, we have enlarged the state space E with
an additional state space F, which represents the environment. The dynamical
system
T represents the true dynamics of the closed system E+environment.
Equation (2.2) says exactly that the eective pseudo-dynamics L that we have
observed on E is simply due to the fact that we are looking only at a subpart of
a true dynamical system and an average of the F part of the dynamics.
These observations would be even more interesting if one could prove the con-
verse: every Markov transition kernel can be obtained this way. This is what we
prove now, with only a very small restriction on E.
Recall that a Lusin space is a measurable space which is homeomorphic (as a
measurable space) to a Borel subset of a compact metrisable space. This condition
is satised for example by all the spaces R
n
.
Theorem 2.3. Let (E, c) be a Lusin space and a Markov kernel on E. Then
there exists a measurable space (F, T), a probability measure on (F, T) and a
dynamical system
T on E F such that the Markov kernel L associated to the
restriction of
T to E is equal to .
Proof. Let (x, dz) be a Markov kernel on (E, c). Let F be the set of functions
from E to E. For every nite subset = x
1
, . . . , x
n
E and every A
1
, . . . , A
n
c consider the set
F(x
1
, . . . , x
n
; A
1
, . . . , A
n
) = y F ; y(x
1
) A
1
, . . . , y(x
n
) A
n
.
By the Kolmogorov Consistency Theorem (which applies for E is is a Lusin space!)
there exists a unique probability measure on F such that
(F(x
1
, . . . , x
n
; A
1
, . . . , A
n
)) =
n
i=1
(x
i
, A
i
) .
Indeed, it is easy to check that the above formula denes a consistent family of
probability measures on the nitely-based cylinders of F, then apply Kolmogorovs
Theorem.
Now dene the dynamical system
T : E F E F
(x, y) (y(x), y) .
MARKOV CHAINS AND DYNAMICAL SYSTEMS 529
With the same notations as in the proof of Theorem 2.2, we have X(x, y) = y(x)
in this particular case and hence
(y F ; X(x, y) A) = (y F ; y(x) A) = (x, A) .
This proves our claim by (2.3).
Note that in this dilation of L, the dynamical system T has no reason to be
invertible in general. It is worth noticing that one can construct a dilation where
T is invertible.
Proposition 2.4. Every Markov kernel , on a Lusin space E, admits a dilation
= E F. Let x
0
be a xed element of E and dene the mapping
T
on
E F
by
_
(x, (x
0
, y)) = (y(x), (x, y)),
is a bijection of E F
is invertible and
dilates the same Markov kernel as
T.
2.3. Iterating the Dynamical System. We have shown that every dynamical
system on a product set gives rise to a Markov kernel when restricted to one of the
sets. We have seen that every Markov kernel can be obtained this way. But one
has to notice that our construction allows the dynamical system
T to dilate the
Markov kernel L as a single mapping only. That is, iterations of the dynamical
system T
n
do not in general dilate the semigroup L
n
associated to the Markov
process. Let us check this with a simple counter-example.
Put E = F = 1, 2. On F dene the probability measure (1) = 1/4 and
(2) = 3/4. Dene the dynamical system
T on E F which is the anticlockwise
rotation:
EPHANE ATTAL
In particular
L
2
= L.
Let us compute
T
2
. We get
T
2
(1, 1) = (2, 2),
T
2
(2, 1) = (1, 2),
T
2
(2, 2) = (1, 1),
T
2
(1, 2) = (2, 1) .
Hence the associated X-mapping, which we shall denote by X
2
, is given by
X
2
(1, 1) = 2, X
2
(2, 1) = 1, X
2
(2, 2) = 1, X
2
(1, 2) = 2 .
This gives the Markovian matrix
L
2
=
_
0 1
1 0
_
,
which is clearly not equal to L
2
.
It would be very interesting if one could nd a dilation of the Markov kernel L
by a dynamical system
T such that any power T
n
would also dilate L
n
. We would
have realized the whole Markov chain as the restriction of iterations of a single
dynamical system on a larger space.
This can be performed in the following way (note that this is not the only way,
nor the more economical). Let L be a Markov operator on a Lusin space E with
kernel and let T be a dynamical system on E F which dilates L. Consider
the set
F = F
N
. The elements of
F are sequences (y
n
)
nN
in F. Put
S : E
F E
F
(x, y) (X(x, y
1
), (y))
where X is as in the the proof of Theorem 2.2 and is the usual shift on
F:
(y) = (y
n+1
)
nN
.
Then
S can be lifted into a morphism S of /
(E
F), as previously. Fur-
thermore, any function f in /
(E
F), with
(f 1l)(x, y) = f(x).
Theorem 2.5. For all n N
(E) we have
_
F
S
n
(f 1l)(x, y) d (y) = (L
n
f)(x) .
Proof. Recall that we noticed in the proof of Theorem 2.2, that the mapping Y
associated to
T played no role in the proof of this theorem, only the mapping X
was of importance. In particular this implies that Theorem 2.5 is true for n = 1,
for the dynamical systems
T and
S share the same X-mapping.
By induction, let us assume that the relation
_
F
S
k
(f 1l)(x, y) d (y) = (L
k
f)(x)
MARKOV CHAINS AND DYNAMICAL SYSTEMS 531
holds true for all f /
F
S
n+1
(f 1l)(x, y) d (y) =
=
_
F
S
n
(f 1l) (X(x, y
1
), (y)) d (y)
=
_
F
_
F
[2
S
n
(f 1l) (X(x, y
1
), y) d
[2
(y) d(y
1
) .
Put x = X(x, y
1
), the above is equal to
_
F
_
F
[2
S
n
(f 1l) ( x, y) d
[2
(y) d(y
1
)
=
_
F
L
n
(f)( x) d(y
1
) (by induction hypothesis)
=
_
F
L
n
(f)(X(x, y
1
)) d(y
1
)
= L
n+1
(f)(x) .
With this theorem and with Theorem 2.3, we see that every Markov chain on E
can realized as the restriction on E of the iterations of a deterministic dynamical
system
T acting on a larger set.
The physical interpretation of the construction above is very interesting. It
represents a scheme of repeated interactions. That is, we know that the result
of the deterministic dynamics associated to
T on E F gives rises to the Markov
operator L on E. The idea of the construction above is that the environment is
now made of a chain of copies of F, each of which is going to interact, one after the
other, with E. After, the rst interaction between E and the rst copy of F has
happened, following the dynamical system
T, the rst copy of F stops interacting
with E and is replaced by the second copy of F. This copy now interacts with
E following
T. And so on, we repeat these interactions. The space E keeps the
memory of the dierent interactions, while each copy of F arrives independently
in front of E and induces one more step of evolution following
T.
As a result of this procedure, successive evolutions restricted to E correspond
to iterations of the Markov operator L. This gives rise to behavior as claimed: an
entire path of the homogeneous Markov chain with generator L.
2.4. Defect of Determinism and Loss of Invertibility. We end up this sec-
tion with some algebraic characterizations of determinism for Markov chains. The
point is to characterize what exactly is lost when going from the deterministic
dynamics T on E F to the Markov operator L on E.
Theorem 2.6. Let (E, c) be a Lusin space. Let (X
n
) be a Markov chain with
state space (E, c) and with transition kernel . Let L be the Markov operator on
532 ST
EPHANE ATTAL
/
(E) associated to :
Lf(x) =
_
E
f(y) (x, dy) .
Then the Markov chain (X
n
) is deterministic if and only if L is a -homomorphism
of the algebra /
(E).
Proof. If the Markov chain is deterministic, then L is associated to a dynamical
system and hence it is a -homomorphism (Proposition 2.1).
Conversely, suppose that L is a -homomorphism. We shall rst consider the
case where (E, c) is a Borel subset of a compact metric space.
Take any A c, any x E and recall that we always have
(x, A) = L(1l
A
)(x) .
The homomorphism property gives
L(1l
A
)(x) = L(1l
2
A
)(x) = L(1l
A
)
2
(x) = (x, A)
2
.
Hence (x, A) satises (x, A)
2
= (x, A). This means that (x, A) is equal to 0
or 1, for all x E and all A c.
Consider a covering of E with a countable family of balls (B
i
)
iN
, each of
which with diameter smaller than 2
n
(this is always possible as E is separable).
From this covering one can easily extract a partition (S
i
)
iN
of E by measurable
sets, each of which with diameter smaller than 2
n
. We shall denote by o
n
this
partition.
Let x E be xed. As we have
ES
n (x, E) = 1 we must have (x, E) = 1
for one and only one E o
n
. Let us denote by E
(n)
(x) this unique set. Clearly,
the sequence (E
(n)
(x))
nN
is decreasing (for otherwise there will be more than
one set E o
n
such that (x, E) = 1). Let A =
n
E
(n)
(x). The set A satises
(x, A) = 1, hence A is non-empty. But also, the diameter of A has to be 0, for it
is smaller than 2
n
for all n. As a consequence A has to be a singleton y(x), for
some y(x) E. Hence we have proved that for each x E there exists a y(x) E
such that n(x, y(x)) = 1. This proves the deterministic character of our chain.
The case where E is only homeomorphic to a Borel subset E
of a compact met-
ric space is obtained by using the homeomorphism to transfer suitable partitions
o
n
of E
to E.
The result above is quite amazing. It gives such a clear and neat characterization
of the dierence between a true Markov operator and a deterministic one! One can
even think of several applications of this characterization, for example one may be
able to measure the level of randomness of some Markov operator by evaluating
for example
sup
_
_
T(f
2
) T(f)
2
_
_
; f /
(E)
associated to a Markov chain (X
n
). If L is invertible in the category of Markov
operators then (X
n
) is deterministic.
Proof. Recall that a Markov operator L maps positive functions to positive func-
tions. Hence, in the same way as one proves Cauchy-Schwarz inequality, we always
have
L(f) = L(f)
and
L([f[
2
) L(
f)L(f)
(hint: write the positivity of T((f +g)(f +g)) for all C).
Let M be a Markov operator such that ML = LM = I. We have
[f[
2
=
ff = M L(
ff) M(L(
f)L(f)) M L(
f) M L(f) =
ff = [f[
2
.
Hence we have equalities everywhere above. In particular
M L(
ff) = M(L(
f)L(f)) .
Applying L to this equality, gives
L(
ff) = L(
f)L(f) ,
for all f /
(E).
By polarization it is easy to prove now that L is a homomorphism. By Theorem
2.6 it is the Markov operator associated to a deterministic chain.
The result above is more intuitive than the one of Theorem 2.6, from the point
of view of open systems. If the dynamical system
T on the large space E F is
invertible, this invertibility is always lost when projecting on E. The fact we do
not have access to one component of the coupled system means that we lose all
chance of invertibility.
3. Continuous Time
We now leave the discrete-time set-up to concentrate on continuous-time dy-
namical systems. We aim to show that stochastic dierential equations are actu-
ally a particular kind of continuous-time dynamical systems. In particular they
are deterministic. The type of dynamical system we shall obtain this way is a
continuous-time version of the construction of Theorem 2.5.
3.1. Preliminaries. Let us consider the d-dimensional Brownian motion W on
its canonical space (, T, P). This is to say that = C
0
(R
+
; R
d
) is the space
of continuous functions on R
+
with values in R
d
and which vanish at 0, equiped
with the topology of uniform convergence on compact sets, the -eld T is Borel
-eld of and the measure P is the Wiener measure, that is, the law of a d-
dimensional Brownian motion on . The canonical Brownian motion (W
t
) is
dened by W
t
() = (t), for all and all t R
+
. This is to say, coordinate-
wise: W
i
t
() =
i
(t), for i = 1, . . . , d.
534 ST
EPHANE ATTAL
We dene for all s R
+
the shift
s
as a function from to by
s
()(t) = (t +s) (s) .
We dene the shift operator
s
as follows. If X is any random variable on we
denote by
s
(X) the random variable X
s
, whatever is the state space of X.
In particular we have
s
(W
t
) = W
t+s
W
s
.
As the process Y
t
= W
t+s
W
s
, t R
+
, is again a d-dimensional Brownian mo-
tion, this implies that the mapping
s
preserves the measure P. As a consequence
s
is an isometry of L
2
((, T, P); R
d
).
Lemma 3.1. If H is a predictable process in R
d
, then, for all xed s R
+
, the
process K
t
=
s
(H
ts
), t s is also predictable.
Proof. The process K as a mapping from [s, +[ to R
d
is the composition of
H with the mapping (, t) = (
s
(), t s) from [s, +[ to R
+
. We just
need to check that is measurable for the predictable -algebra T.
Consider a basic predictable set A]u, v], with u < v and A T
u
, then
1
(A]u, v]) =
1
s
(A)]u +s, v +s] .
We just need to check that
1
s
(T
u
) T
u+s
. The -algebra T
u
is generated by
events of the form (W
i
t
[a, b]), for t u and i = 1, . . . , d. The set
1
s
(W
i
t
[a, b])
is equal to (W
i
t+s
W
i
s
[a, b]), hence it belongs to T
u+s
.
One needs also to note that
1
(A0) =
1
s
(A) s T
s
s
for all A T
0
. We have proved the predictable character of K.
In the following, the norm ||
2
is the L
2
((, T, P); R
d
)-norm. For a R
d
-valued
predictable process H we put
_
t
0
H
s
dW
s
=
d
i=1
_
t
0
H
i
s
dW
i
s
.
Lemma 3.2. Let H be a predictable process in R
d
such that
_
t+s
0
|H
u
|
2
2
du < .
Then we have
s
__
t
0
H
u
dW
u
_
=
_
t+s
s
s
(H
us
) dW
u
. (3.1)
Proof. If H is an elementary predictable process then the identity (3.1) is obvious
from the fact that
s
(FG) =
s
(F)
s
(G) for any scalar-valued F and G. A
general stochastic integral
_
t
0
H
s
dW
s
is obtained as a limit in the norm
_
_
_
_
_
t
0
H
s
dW
s
_
_
_
_
2
2
=
_
t
0
|H
s
|
2
2
ds ,
of stochastic integrals of elementary predictable processes. As
s
is an isometry,
it is clear that Equation (3.1) holds true for any stochastic integral.
MARKOV CHAINS AND DYNAMICAL SYSTEMS 535
We are now ready for the main result of this section. Beforehand recall the
following result on stochastic dierential equations (cf [13], Chapter V). Let f be
a locally bounded Lipschitz function from R
n
to R
n
and g a locally bounded Lips-
chitz function from R
n
to M
nd
(R). Consider the stochastic dierential equation
X
x
t
= x +
_
t
0
f(X
x
u
) du +
_
t
0
g(X
x
u
) dW
u
,
which is a shorthand for
(X
x
t
)
i
= x
i
+
_
t
0
f(X
x
u
)
i
du +
d
j=1
_
t
0
g(X
x
u
)
i
j
dW
j
u
,
for all i = 1, . . . , n. Then this equation admits a solution X
x
and this solution
is unique, in the sense that any other process on (, T, P) satisfying the same
equation is almost surely identical to X
x
.
Theorem 3.3. Let W be a d-dimensional Brownian motion on its canonical space
(, T, P). Let f be a locally bounded Lipschitz function from R
n
to R
n
and g a
locally bounded Lipschitz function from R
n
to M
nd
(R). Denote by X
x
the unique
stochastic process (in R
n
) which is a solution of the stochastic dierential equation
X
x
t
= x +
_
t
0
f(X
x
u
) du +
_
t
0
g(X
x
u
) dW
u
.
Then, for all s R
+
, for almost all , we have, for all t R
+
,
X
X
x
s
()
t
(
s
()) = X
x
s+t
() .
Remark 3.4. Let us be clear about the sentence for all s R
+
, for almost all
, we have, for all t R
+
above. It means that for all s R
+
, there exists a
null-set A
s
such that for all w A
s
we have, for all t R
+
...
Proof. Let s be xed. Dene, for all
Y
x
u
() =
_
_
X
x
u
() if u s ,
X
X
x
s
()
us
(
s
()) if u > s .
Then Y
x
s+t
satises
Y
x
s+t
() = X
X
x
s
()
t
(
s
())
= X
x
s
() +
__
t
0
f(X
X
x
s
u
) du
_
(
s
()) +
__
t
0
g(X
X
x
s
u
) dW
u
_
(
s
())
= x +
__
s
0
f(X
x
u
) du
_
() +
__
s
0
g(X
x
u
) dW
u
_
()+
+
_
t
0
f(X
X
x
s
u
)(
s
()) du +
__
s+t
s
s
_
g(X
X
x
s
us
)
_
dW
u
_
()
by Lemma 3.2.
536 ST
EPHANE ATTAL
Now, coming back to the denition of Y we get
Y
x
s+t
() = x +
_
s
0
f(Y
x
u
)() du +
__
s
0
g(Y
x
u
) dW
u
_
() +
_
t
0
f(Y
x
u+s
)() du+
+
__
s+t
s
g(Y
x
u
) dW
u
_
()
= x +
__
s
0
f(Y
x
u
) du +
_
s
0
g(Y
x
u
) dW
u
+
_
s+t
s
f(Y
x
u
) du+
+
_
s+t
s
g(Y
x
u
) dW
u
_
()
= x +
__
s+t
0
f(Y
x
u
) du +
_
s+t
0
g(Y
x
u
) dW
u
_
() .
This shows that Y
x
is solution of the same stochastic dierential equation as X
x
.
We conclude easily by uniqueness of the solution.
3.2. Stochastic Dierential Equations and Dynamical Systems. We are
now ready to establish a parallel between stochastic dierential equations and
dynamical systems. Recall how we dened discrete time dynamical systems
T in
Section 2 and their associated semigroups (
T
n
). In continuous time the denition
extends in the following way.
A continuous-time dynamical system on a measurable space (E, c) is a one-
parameter family of measurable functions (
T
t
)
tR
+ on E such that
T
s
T
t
=
T
s+t
for all s, t. That is,
T is a semigroup of functions on E.
Each of the mappings
T
t
can be lifted into an operator on /
(E), denoted by
T
t
and dened by
T
t
f(x) = f(
T
t
x) .
The following result is now a direct application of Theorem 3.3.
Corollary 3.5. Let W be a d-dimensional Brownian motion on its canonical space
(, T, P). Let f be a locally bounded Lipschitz function from R
n
to R
n
and let g be
a locally bounded Lipschitz function from R
n
to M
nd
(R). Consider the stochastic
dierential equation (on R
n
)
X
x
t
= x +
_
t
0
f(X
x
u
) du +
_
t
0
g(X
x
u
) dW
u
.
Then the mappings
T
t
on R
n
dened by
T
t
(x, ) = (X
x
t
(),
t
())
dene a continuous time dynamical system on R
n
, in the sense that there
exists a null set A such that for all A, for all x R
n
and for all
s, t R
+
we have
T
t
T
s
(x, ) = T
s
T
t
(x, ) = T
s+t
(x, ) .
MARKOV CHAINS AND DYNAMICAL SYSTEMS 537
Proof. The null set A
s
appearing in Theorem 3.3 also depends on the initial point
x R
n
. Let us denote by A
x,s
this set, instead. Let A
x,s,t
be the null set
A
x,s
A
x,t
. Finally put
A =
_
xQ
n
_
s,tQ
+
A
x,s,t
.
Then A is a null set and for all A the relations
T
t
T
s
(x, ) = T
s
T
t
(x, ) = T
s+t
(x, )
hold true for all x Q
n
and all s, t Q
+
, by Theorem 3.3.
The solution X
x
t
() is continuous in t, except for a null set A
of s. Hence,
by continuity, the relations above remain true for all s, t R
+
, if (A A
).
In the same way, as the solution X
x
t
depends continuously in x, we conclude
easily.
This is to say that, apart from this minor restriction to the complement of a null
set in , a stochastic dierential equation is nothing more than a deterministic
dynamical system on a product set R
n
, that is, it is a semigroup of point
transformations of this product set.
when this dynamical system is restricted to the R
n
-component. But before
establishing this result, we need few technical lemmas. In the following
t]
denotes
the space of continuous functions from [0, t] to R
d
. For all we denote by
t]
the restriction of to [0, t]. Finally P
t]
denotes the restriction of the measure P
to (
t]
, T
t
).
Lemma 3.6. The image of the measure P under the mapping
t]
(
t]
,
t
())
is the measure P
t]
P.
Proof. Recall that (s) = W
s
() and
t
()(s) = W
t+s
() W
t
(). If A is a nite
cylinder of
t]
and B a nite cylinder of , then the set
; (
t]
,
t
()) AB
is of the form
; W
t1
() A
1
, . . . , W
tn
() A
n
, (W
s1
W
t
)() B
1
, . . .
. . . , (W
s
k
W
t
)() B
k
for some t
1
, . . . , t
n
t and some s
1
, . . . , s
k
> t. By the independence of the
Brownian motion increments, the probability of the above event is equal to
P
t]
(
t]
; W
t1
() A
1
, . . . , W
tn
() A
n
)
P( ; (W
s1
W
t
)() B
1
, . . . , (W
s
k
W
t
)() B
k
) .
This is to say,
P( ; (
t]
,
t
()) AB) =
= P
t]
(
t]
t]
;
t]
A) P( ;
t
() B) .
538 ST
EPHANE ATTAL
This is exactly the claim of the lemma for the cylinder sets. As the measures P and
P
t]
P are determined by their values on the cylinder sets, we conclude easily.
Lemma 3.7. Let g be a bounded measurable function on
t]
. Then we have
_
t]
g(,
) dP
t]
() dP(
) =
_
g(
t]
,
t
()) dP() .
Proof. This is just the Transfer Theorem for the mapping of Lemma 3.6.
Theorem 3.8. Let (, T, P) be the canonical space of a d-dimensional Brownian
motion W. Let f be a locally bounded Lipschitz function from R
n
to R
n
and let g be
a locally bounded Lipschitz function from R
n
to M
nd
(R). Consider the stochastic
dierential equation (on R
n
)
X
x
t
= x +
_
t
0
f(X
x
u
) du +
_
t
0
g(X
x
u
) dW
u
and the associated dynamical system
T
t
(x, ) = (X
x
t
(),
t
()) .
For any bounded function h on R
n
consider the mapping
P
t
h(x) = E[ T
t
(h 1l)(x, )] =
_
h(X
x
t
()) dP() .
Then (P
t
)
tR
+ is a Markov semigroup on R
n
with generator
A =
n
i=1
f
i
(x)
x
i
+
1
2
n
i,j=1
d
=1
g
i
(x)g
j
(x)
2
x
i
x
j
.
Proof. The fact that each P
t
is a Markov operator is a consequence of Theorem
2.2. Let us check that they form a semigroup.
First of all note that, since X
x
is a predictable process, the quantity X
x
t
()
depends only on
t]
and not on the whole of . We shall denote by X
x
t
(
t]
) the
associated function of
t]
.
By denition of P
t
we have
P
t
(P
s
h)(x) = E[T
t
(P
s
h 1l)(x, )]
=
_
P
s
h(X
x
t
()) dP()
=
_
h
_
X
X
x
t
()
s
(
)
_
dP(
) dP()
=
_
t]
_
h
_
X
X
x
t
()
s
(
)
_
dP(
) dP
t]
()
=
_
h
_
X
X
x
t
(
t]
)
s
(
t
())
_
dP() (by Lemma 3.7)
=
_
h
_
X
x
s+t
()
_
dP() (by Theorem 3.3)
= P
s+t
h(x) .
MARKOV CHAINS AND DYNAMICAL SYSTEMS 539
We have proved the semigroup property.
The rest of the proof comes from the usual theory of Markov semigroups and
their associated generators (see for example [12], Chapter VII).
We have proved the continuous time analog of Theorem 2.5. Every Markov
semigroup, with a generator of the form above, can be dilated on a larger set (a
product set) into a deterministic dynamical system. What is maybe more surpris-
ing is that the deterministic dynamical system in question is a stochastic dieren-
tial equation. Theorem 3.8 and Corollary 3.5 show that a stochastic dierential
equation can actually be seen as a particular deterministic dynamical system.
Theorem 3.8 above again gives an open system point of view on Markov pro-
cesses: Markov processes are obtained by the restriction of certain types of dy-
namical systems on a product space, when one is averaging over one inaccessible
component. The role of the environment is now played by the Wiener space and
the role of the global dynamics on the product space is played by the stochastic
dierential equation.
In this section we have developed the Brownian case only. But it is clear that all
this discussion extends exactly in the same way to the case of the Poisson process.
Indeed, the arguments developed above are mostly only based on the independent
increment property.
We have said that stochastic dierential equations are particular dynamical sys-
tems which are continuous analogues of those of Section 2.3: repeated interactions.
In the article [7], the convergence of discrete-time repeated interactions models to
stochastic dierential equations is proved.
Remark 3.9. We do not pretend that all the results presented in this article are
new. Let us be clear about that. The fact that restrictions of dynamical systems
can give rise to Markov chains is rather well-known among specialists of dynamical
systems. The results of Subsection 2.4 are adaptations to the classical context of
similar results on completely positive maps for quantum systems. The fact that
stochastic dierential equations give rise to deterministic dynamical systems is also
not new and can be found for example in [6] (see also [1] for more general noises).
The originality of our article lies more in its survey character, in the way we
put all these results together, in the connection we make with repeated interaction
systems and in the physical point of view we adopt.
Acknowledgment. The author is very grateful to the referee of this article for
his very careful reading, his remarks and suggestions.
References
1. Applebaum, D.: Levy Processes and Stochastic Calculus, Second edition, Cambridge Studies
in Advanced Mathematics, 116. Cambridge University Press, Cambridge, 2009.
2. Attal, S. and Pautrat, Y.: From repeated to continuous quantum interactions,Annales Henri
Poincare (Theoretical Physics) 7 (2006) 59104.
3. Bernardin, C. and Olla, S.: Fouriers law for a microscopic heat conduction model,Journal
of Statistical Physics 121 (2005) 271289.
4. Bodineau,T. and Giacomin, G.: From dynamic to static large deviations in boundary driven
exclusion particle systems,Stochastic Processes and Applications 110 (2004) 6781.
540 ST
EPHANE ATTAL
5. Bonetto, F., Lebowitz, J. L., and Rey-Bellet, L.: Fourier Law: A challenge to theorists, In:
Mathematical Physics 2000, A. Fokas, A. Grigoryan, T. Kibble, and B. Zegarlinski (Eds.)
Imp. Coll. Press, London 2000.
6. Carverhill, A.: Flows of stochastic dynamical systems: ergodic theory,Stochastics 14 (1985)
273317.
7. Deschamps, J.: Continuous time limit of classical repeated interaction systemspreprint.
8. Haake, F. and Spehner, D.: Quantum measurements without macroscopic superposi-
tionsPhysical Review A 77, 052114 (2008), 24 pp.
9. Jaksic, V., Ogata, Y., and Pillet, C.-A. : The Green-Kubo formula and the Onsager reci-
procity relations in quantum statistical mechanics,Communications in Mathematical Physics
265, 3 (2006), 721738.
10. Karevski, D. and Platini, T.: Quantum non-equilibrium steady states induced by repeated
interactions,Phys Rev. Letter, to appear.
11. Preskill, J.: Quantum Information and Quantum Computation, Course Files on the Web:
http://www.theory.caltech.edu/people/preskill/ph229
12. Revuz, D. and Yor, M.: Continuous Martingales and Brownian Motion, Springer Verlag,
Grundlehren der mathematischen Wissenschaften 293 (2005).
13. Rogers, L. C. G. and Williams, D.: Diusions, Markov Processes and Martingales, Volume
2, Cambridge University Press (2000).
14. Ruelle, D.: A departure from equilibrium,Nature 414 no. 6861 (2001), 263264.
St ephane Attal: Universit e de Lyon, Universit e de Lyon 1, C.N.R.S., Institut
Camille Jordan, 21 av Claude Bernard, 69622 Villeubanne cedex, France
E-mail address: attal@math.univ-lyon1.fr
URL: http://math.univ-lyon1.fr/~attal
COSINE AND GAUSSIAN TRANSFORMS*
CARLOS LIZAMA AND ROLANDO REBOLLEDO
Abstract. This paper introduces and investigates probabilistic properties
of a class of Gaussian processes connected with cosine transforms, which
have been used to describe non Markovian classical open systems in physics.
Several examples of these processes are considered, namely, the fractional
Brownian motion with Hurst coecient > 1/2 and other processes appearing
in viscoelasticity models.
1. Introduction
This paper concerns the construction of Gaussian non Markovian processes
obtained from integration with respect to a Brownian motion on the frequency
domain. These processes are motivated by the non Markovian approach to Open
System Dynamics.
Indeed, consider rst a classical mechanical system which can be observed and
will be referred as the main system. This system is immersed in a reservoir or heat
bath composed of a great number of harmonic oscillators which collide with the
main system, where the frequencies vary over the positive real line. To describe the
action of the reservoir on the main system one supposes rst that the interaction
is produced at discrete frequencies jh, where j N and h > 0. Variations on the
momentum of the main system are described by a memory kernel and a random
force. The latter depends on the variation of initial conditions of each harmonic
oscillator. This is the setup used in a number of recent papers in physics (see eg.
[5]).
This paper is aimed at investigating probabilistic properties of non Markovian
Gaussian processes generated by the above approach to open system dynamics.
Moreover, we enlarge the class of such processes by tracing a parallel with the
theory of cosine transform. Cosine transforms appear naturally when considering
classical heat bath made of an innite number of harmonic oscillators.
In the next section we start studying positive cosine transforms, to follow in
section 3 with the denition of a Gaussian transform. In section 4, these Gaussian
transforms are represented as stochastic integrals. The case of discrete Gaussian
transforms and their limits are considered in section 5. We end the paper by
Received 2010-5-21; Communicated by the editors.
2000 Mathematics Subject Classication. 60G15; 60G22.
Key words and phrases. Cosine transform, non Markovian Gaussian processes, fractional
Brownian motion.
* This research has been supported by the Bicentennial Foundation grant PBCT-ADI13
Laboratorio de Analisis Estocastico and FONDECYT project 1100485.
541
Serials Publications
www.serialspublications.com
Communications on Stochastic Analysis
Vol. 4, No. 4 (2010) 541-551
542 CARLOS LIZAMA AND ROLANDO REBOLLEDO
providing a number of examples. Applications to open systems in physics will be
considered in a separate paper.
2. Positive Cosine Transforms
Let v L
1
(R
+
) be a continuous positive convex function, therefore its second
derivative in distributions sense is positive. Dene
R(t) =
_
t
0
(t u)v(u)du, (2.1)
for all t 0. Note that R(t) is a positive function. We begin with the following
result.
Proposition 2.1. The function
K(t, s) = R(t) +R(s) R(t s) (2.2)
is positive denite.
Proof. Since v is convex, the function
2
_
0
v(t) cos(xt)dt (2.3)
is positive as proved in [8]. Therefore, dene
f(x) =
_
0
v(t) cos(xt)dt, (2.4)
for all x 0. This is a function in L
2
(R
+
). Moreover, v(t) is the cosine transform
of the function f
2
, that is
v(t) =
_
0
f
2
(x) cos(xt)dx.
As a result, R can be written
R(t) =
_
0
_
f(x)
x
_
2
(1 cos(xt))dt.
Therefore, K can be written as
K(s, t) =
_
0
_
f(x)
x
_
2
[(1 cos(xt))(1 cos(xs)) + sin(xt) sin(xs)] dx. (2.5)
Now, take any nite set I of positive numbers and z(t) C, t I, then
s,tI
z(t)z(s)K(s, t) =
_
0
_
f(x)
x
_
2
tI
z(t)(1 cos(xt))
2
dx
+
_
0
_
f(x)
x
_
2
tI
z(t) sin(xt)
2
dx
0.
sR
a(s)K(s, ),
where a : R C is a function with nite support. Dene,
g, h =
s,t
a(s)b(t)K(s, t), (2.8)
where g =
s
a(s)K(s, ) and h =
t
b(t)K(t, ), a and b with nite support. So
that,
g, K(t, ) = g(t), (t R).
Thus, in particular, K(s, ), K(t, ) = K(s, t), s, t R
+
. The space h(K) is
then the completion of 1(K) for the scalar product (2.8).
Notice that, K
1
(s, ), K(t, ) = K
1
(s, t) = K
1
(s, ), K
1
(t, ). Therefore,
K
1
(s, ), K
2
(t, ) = K
1
(s, ), K(t, ) K
1
(t, )
= K
1
(s, ), K(t, ) K
1
(s, ), K
1
(t, )
= K
1
(s, t) K
1
(s, t)
= 0.
As a result we have the following easy consequence.
Corollary 2.3. Under the above notations, it holds:
h(K) = h(K
1
) h(K
2
). (2.9)
3. Gaussian Transforms
We keep the notations of the previous section. So that to each v L
1
(R
+
)
continuous, positive and convex function, we associate a positive denite kernel K
which is decomposed in two kernels K
1
and K
2
.
Applying Thm. 2.4 in [6] one easily obtains the following Proposition.
Proposition 3.1. There exist a probability space (, T, P) and two independent,
centered Gaussian processes X
i
= (X
i
(t); t 0) with covariances K
i
, i = 1, 2.
Moreover, a Hilbert space h(X
i
) L
2
(, T, P) is associated to each process X
i
by means of an isomorphism of Hilbert spaces dened through
s
a(s)K
i
(s, )
s
a(s)X
i
(s), where a has nite support, i = 1, 2.
544 CARLOS LIZAMA AND ROLANDO REBOLLEDO
Remark 3.2. As a result, if one denes X = (X
1
, X
2
) and Z = X
1
+X
2
, they are
centered Gaussian processes as well and it holds:
h(X) = h(X
1
) h(X
2
) (3.1)
h(Z) = h(X
1
) h(X
2
). (3.2)
Moreover, h(X) is isomorphic to h(K
1
) h(K
2
) and h(Z) is isomorphic to h(K).
The above processes are all continuous as applications from [0, [ into the space
L
2
(, T, P). The continuity of their trajectories will be obtained in Theorem 3.5.
Denition 3.3. Let 1 denote the class of all functions v L
1
(R
+
) which are
continuous, positive and convex. To each v 1 the previous Proposition associates
the centered Gaussian process Z, denoted from now on B
v
, whose covariance is K
given by (2.2) (equivalently by (2.5)). We say that the distribution of B
v
is the
Gaussian transform of the function v.
It is worth noticing that the process Z constructed in Proposition 3.1 is not
unique. However, its distribution is uniquely determined by K, so that the Gauss-
ian transform is dened as a probability distribution on the space of trajectories
of the process Z. To alleviate the notations we will abuse the language speaking
of the process B
v
as the Gaussian transform of v. We need the following Lemma.
Lemma 3.4. The function R given by (2.1) satises
_
1
_
2R(e
y
2
)dy (2 |v|
1
)
1/2
_
1
e
x
2
/2
dx < . (3.3)
Proof. Let x [1, [, then
R(e
x
2
) =
_
e
x
2
0
(e
x
2
u)v(u)du
e
x
2
_
e
x
2
0
v(u)du
e
x
2
|v|
1
.
Therefore,
_
1
_
2R(e
y
2
)dy (2 |v|
1
)
1/2
_
1
e
x
2
/2
dx < .
1
(y, t)dW
1
(y) +
_
x
0
2
(y, t)dW
2
(y), (4.1)
(x, t) R
+
[0, T].
The following is one of the main results of this paper.
Theorem 4.1. The process Z dened by (4.1) is a centered Gaussian process with
trajectories in the space C(R
+
[0, T]) and covariance given by
K((y, s), (x, t)) =
_
xy
0
_
f(r)
r
_
2
[(1 cos(rt))(1 cos(rs)) + sin(rt) sin(rs)] dr.
(4.2)
Moreover, Z(x, t) converges in L
2
to a limit Z(, t), and this convergence is
uniform in t [0, T]. The distribution of the process (Z(, t); t [0, T]) coincides
with that of B
v
.
Proof. Z is clearly a Gaussian process since it is the sum of two independent sto-
chastic integrals of deterministic functions with respect to the Brownian motions
W
1
and W
2
.
The covariance is the sum of the covariances of each integral since they are
independent. These covariances are, respectively:
_
xy
0
1
(r, t)
1
(r, s)dr;
_
xy
0
2
(r, t)(r, s)dr.
This yields formula (4.2).
Moreover, notice that for k = 1, 2, sup
t[0,T]
_
x
2
k
(r, t)dr tends to 0 as x
since the functions t
2
k
(r, t)1
[x,[
(r) are continuous and decreasing to 0 on the
compact interval [0, T], thus they converge to 0 uniformly by Dinis lemma. As a
546 CARLOS LIZAMA AND ROLANDO REBOLLEDO
result, Z(, t) exists and t Z(, t) is a continuous centered Gaussian process
given by
Z(, t) =
_
0
_
f(x)
x
_
(1cos(xt))dW
1
(x) +
_
0
_
f(x)
x
_
sin(xt)dW
2
(x). (4.3)
Finally, the covariance of Z(, ) is
E(Z(, t)Z(, s)) =
_
0
_
f(x)
x
_
2
[(1 cos(xt))(1 cos(xs)) + sin(xt) sin(xs)] dx,
so that it coincides with K, the covariance of the Gaussian transform B
v
. There-
fore, the distribution of Z(, ) coincides with that of B
v
. This completes the
proof.
The following two Corollaries, provide concrete examples. The rst one, is
concerned with fractional Brownian motion. We recall that fractional Brownian
motion was dened some years ago to model diusive, subdifussive and superdi-
fussive transport processes, also known as anomalous difussion processes.
Corollary 4.2. The fractional Brownian motion with Hurst coecient H = 1
2
,
where 0 < < 1 corresponds to the Gaussian transform
_
(2 )(1 )B
v
=
_
2H(2H 1)B
v
of the function v(t) = t
, (t > 0).
Proof. Notice that v is convex on ]0, [, and one obtains
(f(x))
2
= 2
1
sin
2
(1 )x
1
,
for all x [0, [. Moreover,
R(t) =
_
t
0
_
s
0
u
duds =
1
(2 )(1 )
t
2
.
As a result, for all s, t 0,
K(s, t) = R(t) +R(s) R([t s[) =
1
(2 )(1 )
_
[t[
2H
+[s[
2H
[t s[
2H
_
,
where H = 1 /2.
Remark 4.3. The relationship between the cosine transform of f
2
(x) and v(t) helps
to explain, in a wide sense, the fundamental choice done by Kupfermann in [5, eq.
(2.3), p.295]. In fact, consider the linear integral convolution equation
Q(t) = Q
0
_
t
0
R(t s)Q(s)ds. (4.4)
In the case H = 1/2, we obtain R(t) = 1 and the above equation has solution
Q(t) = Q
0
e
t
which corresponds to the standard diusive model. For general
H = 1 /2, the equation is equivalent to the fractional dierential equation
D
t
Q(t) = Q(t),
COSINE AND GAUSSIAN TRANSFORMS 547
where D
t
denotes the Riemann-Liouville fractional derivative of order (0, 1)
(equivalently, H (1/2, 1)). The solution is given by:
Q(t) = Q
0
t
1
E
,
(t
),
where E
,
denotes the Mittag-Leer function. Observe that the choice of the
memory kernel R(t) (equivalently, v(t)) is arbitrary (v convex) in our context.
The choice of an exponential function v yields another example.
Corollary 4.4. Let v(t) = e
t
for t 0. Then the Gaussian transform of v is
represented as
B
v
(t) =
_
0
1
x
2
(
2
+x
2
)
(1 cos(xt))dW
1
(x) (4.5)
+
_
0
1
x
2
(
2
+x
2
)
sin(xt)dW
2
(x).
Proof. Note that v is a convex function and,
(f(x))
2
=
2
_
0
e
t
cos(xt)dx =
2
(
2
+x
2
)
. (4.6)
Consequently,
R(t) =
1
2
e
t
+
1
t
1
2
, (4.7)
so that
K(s, t) =
1
2
_
e
t
+e
s
e
|ts|
_
+
1
(t +s [t s[)
1
2
. (4.8)
Therefore, the Gaussian transform of v is represented as (4.5).
5. Discrete Gaussian Transforms
The representation of B
v
as a stochastic integral suggests to introduce the fol-
lowing family of processes, based on approximations of Brownian motions by means
of random walks. We start by constructing a canonical probability space where
two mutually independent sequences (
n
)
nN
and (
n
)
nN
of A(0, 1)random vari-
ables are dened. This is done by a well-known standard procedure on the space
G = R
N
R
N
of all sequences g = (g
n
)
nN
which have values in R R, so
that g
n
= (g
1
n
, g
2
n
) and call
n
(g) = g
1
n
R and
n
(g) = g
2
n
R, n N. By
means of Kolmogorov Extension Theorem, one can built up a probability measure
P
G
on (G, (), where ( is the product of Borel elds B(R)
N
B(R)
N
such
that (
n
)
nN
, (
n
)
nN
be mutually independent identically distributed sequences
of random variables each of them with distribution A(0, 1).
Denote D = D([0, [, R) the space of functions w = (w(t); t 0) dened in
[0, [ with values in R, which have left-hand limits (w(t) = lim
st, s<t
w(s)) and
are right-continuous (w(t+) = lim
st, s>t
w(s) = w(t)) at each t 0 (with the
convention w(0) = w(0)). Thus D represents all possible trajectories followed
by states of particles of the main system. D is endowed with the -algebra T
generated by the maps w w(t) for all t 0.
548 CARLOS LIZAMA AND ROLANDO REBOLLEDO
Now, dene = DG, and T = T(. We extend the denition of the variables
n
,
n
to by writing
n
() =
n
(g),
n
() =
n
(g), for each = (w, g) . and
the probability P
G
is extended as well to the space through P(AB) = P
G
(B),
A T, B (.
Consider h > 0 and let us dene
Z
h
(x, t) =
h
[x/h]
k=1
f(kh)
kh
(sin(kht)
k
+ (1 cos(kht))
k
) , (5.1)
for (x, t) R
+
[0, T].
Introducing the -algebras T
h
x
generated by (
j
,
j
) j [x/h], the processes
W
h
1
(x) =
h
[x/h]
j=1
j
(5.2)
and
W
h
2
(x) =
h
[x/h]
j=1
j
(5.3)
become two independent martingales. It is well known that the distribution of
W
h
= (W
h
1
, W
h
2
) weakly converges to the distribution of a two-dimensional Wiener
process W = (W
1
, W
2
).
We start by studying the convergence in distribution of the family of processes
(Z
h
).
Lemma 5.1. Let a real separable Banach space b endowed with its Borel -eld,
be given and let L
2
([0, [, b) be a locally bounded function. Then, the process
I
h
() :=
_
0
(u)dW
h
1
(u) =
h
[/h]
j=1
(jh)
j
,
with trajectories in D(R
+
, b), converges in distribution to
I() :=
_
0
(u)dW
1
(u),
as h 0. Moreover, the process I() is a square integrable martingale.
Proof. This Lemma is an easy extension to the Banach space setting of a well-
known consequence of limit theorems for the convergence in distribution of sto-
chastic integrals. Here we state a more direct and self-contained proof.
Since, b is separable, the elds of Borel and that generated by cylinders
coincide when they are completed by null sets under any tight probability mea-
sure. On the other hand, the space D([0, [, b) of discontinuous functions with
no oscillatory discontinuities becomes a Polish space with the Skorokhod topology.
Introduce the family of D([0, [, b)-modulus as follows:
For any element D([0, [, b), > 0 and each N N, let
w
N
(, ) := inf
{xi}
max
0<ir
sup |(u) (v)| : u, v [x
i
, x
i+1
[ , (5.4)
COSINE AND GAUSSIAN TRANSFORMS 549
where the inmum extends over the nite sets of points x
i
satisfying
_
0 = x
0
< x
1
< . . . < x
r
= N,
x
i
x
i1
> , i = 1, 2, . . . , r.
(5.5)
According to Prokhorovs Theorem, a family of processes (
h
)
h>0
with trajec-
tories in D([0, [, b) is tight if and only if the following two conditions below are
satised:
(T1) For all N N,
lim
a
sup
h>0
P
_
sup
x[0,N]
_
_
h
(x)
_
_
> a
_
= 0;
(T2) For all N N, for all > 0 it holds that
lim
0
sup
h>0
P
_
w
N
(
h
, ) >
_
= 0.
A straightforward computation shows that
P
_
sup
x[0,N]
_
_
I
h
(x)
_
_
> a
_
1
a
2
h
N/h
j=1
|(jh)|
2
N
a
2
_
0
|(u)|
2
du, (5.6)
which implies (T1) and
w
N
(I
h
, ) sup
u[0,N]
|(u)| w
N
(W
h
1
, ). (5.7)
Since (W
h
1
)
h>0
is a D([0, [, R)tight (even better, it is C([0, [, R)tight), from
the above inequality one obtains easily that (I
h
)
h>0
satises the hypothesis (T2)
and so is D([0, [, b)tight.
Finally, to nish the proof, one needs to prove that for any functional , in
the dual b
m
i=1
i
, I
h
(x
i
) to
m
i=1
i
, I(x
i
)
holds (Cramer-Wold device). However,
m
i=1
i
, I
h
(x
i
) =
_
maxi xi
0
_
i=1
i
, (u)1
[0,xi[
(u)
_
dW
h
1
(u), (5.8)
so that it suces to show that for any measurable and bounded real function g,
one has the convergence in distribution of (g W
h
1
)
x
=
_
x
0
g(u)dW
h
1
(u) towards (g
W
1
)
x
=
_
x
0
g(u)dW
1
(u) for any xed x [0, [. This is an easy consequence of the
550 CARLOS LIZAMA AND ROLANDO REBOLLEDO
Central Limit Theorem for Local Martingales. Indeed, the associated increasing
processes of the martingales g W
h
1
are given by
x h
[x/h]
j=1
(g(jh))
2
,
and this family converges to
_
x
0
(g(u))
2
du,
which is the associated increasing process of the continuous Gaussian martingale
g W
1
. Therefore, by the Martingale Central Limit Theorem [7], g W
h
1
converges
in distribution, as a process, to g W
1
. As a result g W
h
1
(x) converges to g W
1
(x)
for all x [0, [ and the proof is complete. Moreover, since L
2
([0, [, b) the
process I is a square-integrable martingale. As a result, there exists the terminal
variable I() b.
Consider the Banach space C of real-valued continuous functions dened on
[0, T], with the uniform norm ||. Denote S
x
(t) = sin(xt), C
x
(t) = cos(xt), for
all x R
+
. Call D(R
+
, C) the Polish space of all cadlag functions from R
+
to C
endowed with Skorokhods topology.
Writing Z
h
(x, ) =
[x/h]
k=1
f(kh)
kh
(
k
S
kh
+
k
(1 C
kh
)) one notices that each
process x Z
h
x
= Z
h
(x, ) has trajectories in D(R
+
, C). Moreover,
Z
h
x
= M
h
1
(x) +M
h
2
(x), (5.9)
where M
h
1
(x) =
[x/h]
k=1
f(kh)
kh
k
S
kh
and M
h
2
(x) =
[x/h]
k=1
f(kh)
kh
k
(1 C
kh
)
are two independent Cvalued martingales.
Proposition 5.2. Let f be dened by (2.4). Then the family ((M
h
1
, M
h
2
))
h>0
converges in distribution as h 0 to a Gaussian C C-valued martingale M =
(M
1
, M
2
) represented as
M(x) =
__
x
0
f(u)
u
S
u
dW
1
(u),
_
x
0
f(u)
u
(1 C
u
)dW
2
(u)
_
and Z
h
converges in distribution as h 0 to the C-valued process Z given by
Z(x, ) =
_
x
0
f(u)
u
S
u
dW
1
(u) +
_
x
0
f(u)
u
(1 C
u
)dW
2
(u). (5.10)
Proof. The proof follows from Lemma 5.1: take
1
(u) =
f(u)
u
S
u
and
2
(u) =
f(u)
u
(1 C
u
), so that
M
h
=
__
0
1
(u)dW
h
1
(u),
_
0
2
(u)dW
h
2
(u)
_
.
Each component having a limit in distribution with continuous trajectories, the
couple converges in distribution to
M =
__
0
1
(u)dW
1
(u),
_
0
2
(u)dW
2
(u)
_
,
and the process Z
h
converges to Z =
_
0
1
(u)dW
1
(u) +
_
0
2
(u)dW
2
(u).
COSINE AND GAUSSIAN TRANSFORMS 551
Corollary 5.3. Under the above hypotheses, each martingale M
j
(j = 1, 2) has a
nal variable M
j
(), and
Z(, ) =
_
0
f(u)
u
S
u
dW
1
(u) +
_
0
f(u)
u
(1 C
u
)dW
2
(u) (5.11)
coincides with the canonical version of B
v
.
References
1. Dudley, R.: Gaussian processes on several parameters, The Annals of Mathematical Statistics
36 (1965) 771788.
2. Fernique, X.: Des resultats nouveaux sur les processus gaussiens, Seminaire de probabilites
de Strasbourg IX, Lecture Notes in Math. 465 (1975) 318335.
3. Fernique, X.: Sur la regularite de certaines fonctions aleatoires dOrnstein-Uhlenbeck. An-
nales de linstitut Henri Poincare Probab. Statist. 26 (1990) 399417.
4. Fernique, X.: Regularite de fonctions aleatoires gaussiennes, stationnaires, Probability The-
ory and Related Fields 88 (1991) 521536.
5. Kupferman, R.: Fractional kinetics in KacZwanzig heat bath models, Journal of Statistical
Physics 114 (2004) 291326.
6. Neveu, J.: Processus Aleatoires Gaussiens. Seminaire de Mathematiques Superieures. Presses
de lUniversite de Montreal, 1968 (Ed. polycopie. Cours Universite de Paris VI, 1972).
7. Rebolledo, R.: Central limit theorems for local martingales, Z. Wahrsch. Verw. Gebiete 51
(1980) 269286.
8. Tuck, E. O.: On positivity of Fourier transforms, Bull. Austral. Math. Soc., 74 (2006) 133
138.
Carlos Lizama: Centro de An alisis Estoc astico y Aplicaciones and Departamento
de Matem atica, Universidad de Santiago de Chile, Casilla 304, Santiago, Chile
E-mail address: carlos.lizama@usach.cl
Rolando Rebolledo: Centro de An alisis Estoc astico y Aplicaciones, Facultad de
Matem aticas, Universidad Cat olica de Chile, Casilla 306, Santiago 22, Chile
E-mail address: rrebolle@uc.cl
TRANSFORMATION OF QUANTUM L
EVY PROCESSES
ON HOPF ALGEBRAS
MICHAEL SCH
(G) and E by
L
(E)) then X
s,t
: E G will give the algebra homomorphism j
s,t
mapping a
function f on G to the function j
s,t
(f) = fX
s,t
on E. The j
s,t
form a commutative
process in the sense that they are dened on a commutative algebra. Replacing
the monoid G by a bialgebra and the classical probability space E by what is
called a quantum probability space, the notion of a quantum Levy process (QLP)
on a bialgebra over a quantum probability space has been introduced in L.
Accardi, M. Sch urmann, and W. von Waldenfels [1].
Received 2010-7-12; Communicated by D. Applebaum.
2000 Mathematics Subject Classication. Primary 46L53, 16T10, 60G51; Secondary 81S25,
60J25, 60B15, 46L55.
Key words and phrases. Quantum probability, noncommutative processes with independent
increments, Levy processes, Hopf algebras in quantum theory, quantum stochastic dierential
equations.
* Supported by research funds of the Italian MIUR and of the University of Molise.
553
Serials Publications
www.serialspublications.com
Communications on Stochastic Analysis
Vol. 4, No. 4 (2010) 553-577
554 MICHAEL SCH
j
tn1,tn
(b)
(with the convolution with respect to replaced by the convolution with respect
to
) form a Cauchy net over the partitions of the interval [s, t]. From this it
is easy to show that their limits, which we denote by k
s,t
(b) determine on their
linear hull a unique QLP k over (B,
-bialgebras (cf. J.M. Lindsay and A. Skalski [7]). For compact quantum groups
there should not be a problem, because these unital C
on V satisfying (v
= b
= a
1
a
2
.
A complex vector space ( is a coalgebra if there are linear maps : ( ( (
and : ( C, called coproduct and counit respectively, satisfying
(id) = (id ) (coassociativity)
( id) = id = (id ) (counit property).
Following Sweedler we frequently use the notation c
(1)
c
(2)
for (c) suppressing
both summation and indices. Let
0
:= , and for n 1 dene
n
= (
n1
id) .
Sweedlers notation extends to writing c
(1)
c
(2)
c
(n)
for
n
(c), n 1.
Sometimes we shall need to equip also the conjugate vector space ( with a
coalgebra structure. Note that the canonical bijection i = i
1
: c c from ( to ( is
an anti-linear isomorphism. The same is true for the canonical bijections i
n
from
the nfold tensor power of ( to the nfold tensor power of (. We may write
i
n
(c
1
c
n
) = c
1
. . . c
n
= c
1
. . . c
n
.
Note that i
n
i
m
= i
n+m
(where the tensor product of antilinear mappings is well-
dened). By i
0
we denote complex conjugation of C. It is, then, easy to convince
oneself that := i
0
i
1
1
and := i
2
i
1
2
make ((, , ) a coalgebra.
We shall also need the tensor product ((
1
(
2
, , ) of two coalgebras ((
1
,
1
,
1
)
and ((
2
,
2
,
2
) where :=
1
2
and := (id id) (
1
2
) and denotes
the ip c d d c.
A bialgebra (B, , ) is a coalgebra which is also a unital algebra, and in
such a way that and are algebra homomorphisms. If / is a unital algebra
556 MICHAEL SCH
a) 0 for all
a /. Let (/, ) be a quantum probability space, that is, a unital algebra
with a state (a normalized positive linear functional : / C). A quantum
stochastic process j =
_
j
i
_
iI
, indexed by some index set I, is a family of quantum
random variables j
i
(that is, of unital algebra homomorphisms j
i
: B /). By
i
:= j
i
we denote the distribution of j
i
. The notion of independence used for
quantum Levy processes on bialgebras in this paper is the tensor independence.
A (stationary) quantum Levy process on B over / is a quantum stochastic
process j =
_
j
s,t
_
, indexed by s, t R
+
, s t, satisfying the following four
conditions.
(LP1) The increments j
s,t
of disjoint intervals are tensor independent in ,
that is,
_
j
s1,t1
(b
1
) j
sn,tn
(b
n
)
_
=
s1,t1
(b
1
)
sn,tn
(b
n
)
for all n N, b
k
B, (s
1
, t
1
], . . . , (s
n
, t
n
] mutually disjoint intervals
of R
+
, and
[j
s
k
,t
k
(b
1
), j
s
l
,t
l
(b
2
)] = 0
for all k ,= l and all b
1
, b
2
B
(LP2) The increments are stationary, that is,
s,t
=
0,ts
for all 0 s t.
(LP3) The process is weakly continuous in , that is, lim
t0
0,t
(b) = (b)
for all b B.
(LP4) The j
s,t
are increments under convolution, that is, j
r,s
j
s,t
= j
r,t
for
all 0 r s t and j
t,t
(b) = (b)1 for all 0 t < .
(For a topological extension of this notion of a QLP to compact quantum groups
and operator space coalgebras see [7].) We observe that by (LP1) and (LP4) every
Levy process fullls the condition
(LP4)
r,s
s,t
=
r,t
for all 0 r s t and
t,t
= .
Therefore, by (LP2) and (LP3) the states
t
:=
0,t
form a weakly continuous
semigroup under convolution. By (LP1), (LP2) and (LP4) this convolution semi-
group determines all joint moments (that is exactly all expressions of the form
of the left-hand side of the rst equation of (LP1), even if we drop the condition
that the (s
k
, t
k
] are mutually disjoint). In other words, two Levy processes are
stochastically equivalent, if and only if they have the same convolution semigroup.
We can associate a generator with a convolution semigroup through
t
= e
t
for
all t 0. Essentially, this follows from the Fundamental Theorem on Coalgebras;
TRANSFORMATION OF QUANTUM L
b) 0 for all
b in the kernel of the counit and that (b
t
_
tR+
there is (up to unitary equiv-
alence) at most one cyclic Levy process. (Unitary equivalence is much stronger
than stochastic equivalence.) Eectively, if j is a cyclic process on (D, ) which
fullls (LP1) - (LP3) and (LP4), then it is not dicult to show that also (LP4)
holds. M. Sch urmann [11, Proposition 1.9.5] shows that for every convolution
semigroup of states on a bialgebra there is a (unique up to unitary equiva-
lence) cyclic Levy process (even without continuity). This construction involves
the GNS-construction of all
t
, their tensor products and an inductive limit over
558 MICHAEL SCH
)) +
t
((b) (b)) + A
t
((b)) + (b)t (2.3)
denotes the generator process of j
s,t
with A
t
,
t
, A
t
the annihilation, preservation
and creation processes on Boson Fock space over
s
(L
2
(R
+
, K)) =
s
(L
2
(R
+
)K
where K denotes the completion of the index space K. For quite a long time it
was an open problem, to decide whether Fock space and dierential equation can
be set in such a way that the Fock vacuum is cyclic for the resulting Levy process.
Only quite recently and simultaneously, U. Franz, M. Sch urmann, and M. Skeide
came up, not with just one, but with a whole bunch of proofs for the armative
answer.
The proof due to M. Skeide (see U. Franz [3, Theorem 1.21]) uses in an essential
way the representation on the Fock space and equation (2.2) and shows that for
every b B with (b) = 1 the vectors
j
t0,t1
(b) j
tn1,tn
(b), (2.4)
s = t
0
t
1
. . . t
n1
t
n
= t, converge over the interval partitions of
(s, t] to an exponential vector of the form exp(k1
(s,t]
) where k K is a vector
depending on b. (Cyclicity is, then, a simply consequence of Skeides proof in [13]
of a result due to K.R. Parthasarathy and V.S. Sunder [9].) Immediately, from
this construction, the idea emerged to construct an explicit isomorphism from the
space of the abstract Levy process of [11, Proposition 1.9.5] to the Fock space of
the Levy process obtained via [11, Theorem 2.5.3]. Namely, if in (2.4) we replace
j and with the abstract process
j and its cyclic vector
, we know from [3,
Theorem 1.21] that they converge. Sending the limit to exp(k1
(s,t]
) establishes a
unitary from the abstract representation space
D to the Fock space. If we can
manage to do this without using [3, Theorem 1.21], then we will obtain a direct
proof of representability of the Levy process as a cyclic process on the Fock space.
The idea for a transformation of a (cyclic) Levy process originates in the fol-
lowing observation. Let us put B
1
:= b B: (b) = 1. Suppose the element
b B
1
is group-like, that is, (b) = b b. (Note that b B being group-like, the
counit property forces b = 0 or b B
1
.) Then
j
t0,t1
(b) j
tn1,tn
(b) = j
t0,t1
j
tn1,tn
(b) = j
s,t
(b)
so that the limit is over a constant and gives back what j
s,t
(b) does to the cyclic
vector. In general, there need not be group-like elements in B
1
, and if, then they
need not generate B. However, if we were able to dene a dierent comultiplication
on B for which all elements in B
1
are group-like, then
k
s,t
(b) = limj
t0,t1
(b) j
tn1,tn
(b)
TRANSFORMATION OF QUANTUM L
j
tn1,tn
with respect to the new comultiplication. Now we do just the opposite
and look at the limit of
k
t0,t1
k
tn1,tn
(b) (2.5)
for the original comultiplication. If this reverse transformation gives back j, then,
knowing that the representation space of the intermediate group-like process k is
isomorphic to a Fock space, we will know that also the representation space of j
is a Fock space. Technically, in general, it is not possible to equip B directly with
a comultiplication that makes the elements of B
1
group-like.
However, it is possible to associate with every bialgebra B its Weyl bialgebra
CB
1
. The vector space CB
1
contains the set B
1
as a basis consisting entirely
of group-like elements. And the k
s,t
(b) dened on elements of B
1
determine a
unique Levy process on CB
1
. But now the k
st
do no longer dene a linear mapping
B L
a
(D). (They do dene a linear mapping CB
1
L
a
(
D) where
D is the linear
span in D of what the k
s,t
(b) generate from .) So the convolutions in (2.5) with
respect to the comultiplication of B do no longer have a meaning. The problem is
solved if we associate again with B a special kind of bialgebra; see example 3.2.
We will equip this tensor bialgebra with a certain comultiplication, so that the
convolutions in (2.5) are dened with respect to this comultiplication.
3. Statement of Results and Applications
In Section 3.1 we state the main result of this paper (Theorem 3.4) on the
transformation of QLPs. We introduce two bialgebra structures on the tensor
algebra over the kernel of the counit of a bialgebra. Moreover, we show that
lifted generators give rise to QLPs on the tensor bialgebra which by restriction
lead back to a version of the original QLP (Proposition 3.3). Section 3.2 is on
the reversion of the transformation which is always possible if the transformation
is surjective. In Section 3.3 we treat applications of our results to classical Levy
processes, to realizations of QLPs by their Weyl process and by their generator
process, to the passage from the Wiener process to Azema martingales, and to
Trotter product formulae for QLPs.
3.1. Transformation of QLPs. Let (B, , ) and ((, , ) be two bialgebras.
A transformation of B is a unital algebra homomorphism : ( B satisfying
= . (3.1)
This means that preserves the counit. Since (1) = 1 it is easy to see that (3.1)
is equivalent to the condition ((
0
) B
0
where (
0
= ker , B
0
= ker . In the
560 MICHAEL SCH
nN
V
n
where V
n
denotes the n-fold tensor product of V with itself, V
0
= C, with unit
element (1, 0, . . . ) and the multiplication given by
(v
1
v
n
)(w
1
w
m
) = v
1
v
n
w
1
w
m
for n, m N, v
1
, . . . , v
n
, w
1
, . . . , w
m
V . The tensor algebra satises the following
universal property. There exists a vector space embedding : V T(V ) of V into
T(V ) such that for any linear mapping f from V into an algebra / there is a unique
algebra homomorphism T(f): T(V ) / such that T(f)(v) = f(v) for all v V .
Then, any algebra homomorphisms g : T(V ) / is uniquely determined by its
restriction to V . In a similar way, an involution on V gives rise to a unique
extension as an involution on T(V ). Thus, for a vector space V we can form the
tensor algebra T(V ) over V . This algebra becomes a bialgebra if we extend
the mappings : V T(V ) T(V ), v v 1+1v and T(0): V (, v 0 as
algebra homomorphisms to dene the comultiplication and the counit on T(V ).
The elements v V are so-called primitive elements of the coalgebra T(V ). By
extending v v as an algebra anti-homomorphism, we obtain an antipode so
that T(V ) becomes a Hopf algebra. We call T(V ) the tensor Hopf algebra
over V .
Let (B, , ) be any bialgebra. The set B
0
= b B: (b) = 0 is a ideal of
B. The tensor Hopf algebra (T(B
0
), , T(0)) is called the generator Hopf algebra
of B.
We obtain a pair of bialgebras by taking for the rst bialgebra B itself, and
for the second one the generator Hopf algebra of B. The role of is played by the
counit preserving algebra homomorphism dened by (b
1
b
n
) = b
1
b
n
for b
1
, . . . , b
n
B
0
. We call the multiplication map and denote it by M.
Example 3.2. (Induced tensor bialgebra associated with a bialgebra)
Let (B, , ) and (T(B
0
), , T(0)) be as in example 3.1. We can dene another
coalgebra structure on T(B
0
). Denote by
E: B
0
B
0
(B
0
B
0
) T(B
0
) T(B
0
)
the canonical embedding coming from the identications of B
0
with B
0
1 and
1 B
0
respectively and of B
0
B
0
T(B
0
) T(B
0
). Moreover, consider the
restriction
0
of to B
0
. Then
0
: B
0
B
0
B
0
(B
0
B
0
)
and (T(B
0
), T(E
0
), T(0)) is a bialgebra. We can understand this bialgebra
as a kind of big version of B and (T(B
0
), T(
0
), T(0)) is called the induced ten-
sor bialgebra associated with B. In the context of the algebraic set-up the rst
bialgebra is (T(B
0
), , T(0)) and the second bialgebra is (T(B
0
), T(
0
), T(0)).
TRANSFORMATION OF QUANTUM L
, B
(b
1
b
n
)
= , J
s,t
(b
1
b
n
)
Using the properties of a QLP, we obtain from that
, J
s,t
(b
1
) J
s,t
(b
n
) = , J
s,t
(b
1
b
n
)
for all , D
J
which proves (3.3).
A generator of a Levy process on B is lifted via to a generator of a
Levy process on (. The question arises, what is the relationship between the two
Levy processes? We will show how the second process can be computed from the
rst one and vice versa.
In the sequel, Z
st
denotes the set of all partitions of an interval [s, t] R
+
. Let
= s = t
0
< t
1
< < t
n1
< t
n
= t be a partition of [s, t] and dene
|| = maxt
j+1
t
j
[ 0 j n 1.
We turn Z
st
into a directed set by writing
1
2
:
1
2
.
Theorem 3.4. Let (B, , ) be a bialgebra and let (j
s,t
)
0st
be the unique
cyclic Levy process over (D
j
, ) whose convolution semigroup is given by a gener-
ator . Let ((, , ) be another bialgebra and let : ( B be a transformation
of B. Denote by H
k
the Hilbert subspace of D
j
dened by
H
k
:= span
_
(j
t0,t1
)(c
1
) (j
tn1,tn
)(c
n
):
n N, c
1
, . . . , c
n
(, 0 s t, s = t
0
t
1
t
n1
t
n
= t
_
.
562 MICHAEL SCH
(c)
_
Zst
converges in norm to
an element in H
k
where
(c) = (j
t0,t1
) (j
tn1,tn
)(c). (3.4)
Moreover, setting
k
s,t
(c) := lim
(c)
determines a unique cyclic Levy process k =
_
k
s,t
_
0st
on ( over a dense subspace
(D
k
, ) of H
k
. The convolution semigroup of this process has generator .
The proof will be given in Section 4. We call k the transform of j.
We formally will describe the construction of k
s,t
out of j
s,t
in the above theorem
by the short hand writing
(j
s,t
) = k
s,t
. (3.5)
We may call k
s,t
the innitesimal convolution product of j
s,t
.
Remark 3.5. In more detail, Theorem 3.4 says that, under the assumptions of
the theorem, the following L
2
-type construction holds. If
k
s,t
is a concrete cyclic
Levy process over (
D,
),
D a pre-Hilbert space with cyclic vector
, on ( with
generator , then
k
s,t
(c)
k
s,t
(c)
k
s,t
(c) | = | k
s,t
(c) (3.6)
for all c (. Notice that the formal writing (3.5) can always be given a mathe-
matical meaning by (3.6). Of course, when the process
k
s,t
consists of bounded
operators, equations (3.6) make sense on H
k
and
H where
H denotes the Hilbert
space which is the completion of
D. However, boundedness does not always hold
in the applications; cf. Examples 3.10, 3.12, 3.13.
3.2. Reversion of the transformation. The reverse transformation of a Levy
process on ((, , ) into a Levy process on (B, , ) requires a counit preserving
algebra homomorphism which, roughly speaking, is the inverse of . The
construction of assumes the surjectivity of . This is equivalent to ((
0
) = B
0
and the existence of an injective linear mapping
: B
0
(
0
such that = id
B
.
The linear mapping is not unique. Its existence follows from the existence of a
self-adjoint basis
_
b
i
_
iI
of the vector space B
0
, I some index set. Choose c
i
(
self-adjoint such that (c
i
) = b
i
. This is possible since is surjective. Dene the
linear map by (b
i
) = c
i
. In view of the universal property of tensor algebras
we extend the linear map to a algebra homomorphism
= T(): T(B
0
) (
TRANSFORMATION OF QUANTUM L
_
Zst
converges in norm to an
element in D
k
where
(b) := (k
t0,t1
)
T(
0
)
T(
0
)
(k
tn1,tn
)(b).
Moreover, setting
j
s,t
(b) := lim
(b)
determines a unique cyclic Levy process j =
_
j
s,t
_
0st
on B over a dense subspace
(D
j
, ) of D
k
. The convolution semigroup of this process has generator .
The following result says that the transform j in Theorem 3.6 is the reversion
of the transform k in Theorem 3.4.
Theorem 3.7. Let be a surjective transformation of the bialgebra B. and let
be a generator on B. An application of Theorem 3.6 to the process
_
k
s,t
_
0st
of Theorem 3.4 gives back the original Levy process on (B, , ). Moreover, we
have H
k
= D
j
= D
k
.
For a proof of Theorems 3.6 and 3.7 see Section 4.
3.3. Applications.
Example 3.8. (Group-like bialgebras)
For a set S the vector space generated by S is the vector space CS consisting of all
functions f : S C with nite support. Assume in addition that S is a monoid
with identity e S. Since S is a basis, the multiplication map S S S induces
a map m: CS CS CS that turns CS into an algebra with identity element
e S CS. Since S is a basis of CS the mapping m induces an algebra structure
on CS with unit element e.
The vector space generated by a set satises the following universal property.
There exists an embedding : S CS such that any mapping from S to some
vector space V can be uniquely extended to a linear mapping
: CS V such that
=
. This can be used to dene a coalgebra structure on CS. We understand S
as a set of group-like elements. We extend the mappings : S CSCS, (s) =
s s and : S C, (s) = 1 to linear mappings on CS. We will denote the
564 MICHAEL SCH
b ) =
b
b and (
b ) = 1 for
b C(B
1
). B
1
is equal to the set of group-like elements in CB
1
, i.e. B
1
= 0 ,=
b
CB
1
: (
b ) =
b ) = b for b B
1
and (b) =
b +1
1 for b B
0
. Now we are able to express
the reverse transformation (2.5) by (k
t0,t1
)
T(
0
)
T(
0
)
(k
tn1,tn
)(b)
for b B
0
.
Example 3.9. (Construction of quantum Levy processes I)
We apply Theorem 3.4 to Example 3.8. Let (B, , ) be some bialgebra and let
_
j
s,t
_
0st
be a cyclic Levy process on B over (D
j
, ) with generator . In view
of Theorem 3.4 we have that
k
s,t
(
b) := lim
j
t0,t1
(b) . . . j
tn1,tn
(b),
b B
1
, denes a cyclic Levy process
_
k
s,t
_
0st
on the Weyl bialgebra (CB
1
, , )
of B over (D
k
, ) where D
k
is a linear subspace of D
j
. Thus, for each pair
k
s,t
(
b), k
s,t
( c) for b, c B
1
and 0 s t < we have
k
s,t
(
b), k
s,t
( c) = e
(ts)(b
c)
.
The generator denes a coboundary by (2.1). Thus, we compute
e
(ts)(b)
k
s,t
(
b), e
(ts)(c)
k
s,t
( c) = e
(ts)((b
)(c)+(b
c))
= e
(ts)(b),(c)
= E((b) 1
[s,t]
), E((c) 1
[s,t]
)
where : B
1
K is the canonical mapping to a dense linear subspace K of a
Hilbert space K and E(() 1
[s,t]
) denotes the exponential vector of () 1
[s,t]
in the Boson Fock space
s
(L
2
([s, t], K)). Here () 1
[s,t]
denotes the function in
L
2
([s, t], K) which is a constant equal to () on the interval [s, t] and 0 elsewhere.
The space K is obtained from by the GNS construction mentioned in Section
2. Hence,
k
s,t
(b)
= e
(ts)(b)
E((b) 1
[s,t]
)
s
(L
2
([s, t], K))
where b B
1
, (b) C and (b) K. In other words, the vectors k
s,t
(b) behave
like exponential vectors in the Boson Fock space
s
(L
2
([s, t], K)) and the k
s,t
(b) act
TRANSFORMATION OF QUANTUM L
=
s
(L
2
([s, t], K)) and thus D
k
=
s
(L
2
(R
+
, K)). The-
orem 3.6 states that the vectors k
s,t
(b), b B
1
, are total in D
js,t
D
j
as well,
i.e.
D
j
= D
k
=
s
(L
2
(R
+
, K)).
We proved that each cyclic quantum Levy process on a bialgebra can be realized
on a Boson Fock space
s
(L
2
(R
+
, K)).
Example 3.10. (Construction of quantum Levy processes II)
In the situation of Example 3.9, an application of Theorem 3.6 allows to reconstruct
j
s,t
from the process k
s,t
on the Weyl bialgebra of B. The realization of the latter
on Fock space can simply be written down. In the present example we describe a
realization of a Levy process on a Bose Fock space that parallels the construction
in [4] and [11] with the help of quantum stochastic calculus in the sense of R.L.
Hudson and K.R. Parthasarathy.
Applying our result to the situation of Example 3.1 and 3.2 with = id there are
two possibilities. If we put the rst bialgebra B equal to the induced bialgebra
and ( equal to the generator Hopf algebra, then for b B
0
we have
(b) =
n
i=1
j
ti1,ti
(b) (3.7)
and Theorem 3.4 tells us that 3.7 converges to
I
s,t
(b) =
_
A
s,t
((b
)) +
s,t
((b)) + A
s,t
((b)) + (b) (t s)
_
in norm where A
s,t
,
s,t
, A
s,t
denote the annihilation, preservation and creation
operators of the interval [s, t] on Boson Fock space
s
(L
2
(R
+
, K)); see the preced-
ing section. For arbitrary b B we nd
I
s,t
(b) = (b)1 + A
s,t
((b
)) +
s,t
((b) (b)) + A
s,t
((b)) + (b) (t s).
I
s,t
is called the generator process of the Levy process j
s,t
.
We may construct j
s,t
out of I
s,t
if we take the generator Hopf algebra for the
rst bialgebra B and the induced one for the second bialgebra (. Then by Theorem
3.4 we obtain a QLP J
s,t
on T(B
0
) as the limit
J
s,t
=
T(0)
T(I
s,t
)
of the convolution products of the generator process where now, of course, convo-
lution is with respect to the original comultiplication of B. The limit is to be
understood in the sense of the remark after Theorem 3.4; see equations (3.5) and
(3.6).
The QLP J
s,t
has the generator M and by Proposition 3.3 the restriction of
its cyclic version to B is a version of a QLP with generator . So our procedure
566 MICHAEL SCH
s,t
on Boson Fock
space. In fact, if dt is small, then in all relevant formula one may substitute
j
t,t+dt
with I
t,t+dt
. We nd, in a heuristic sense,
j
s,t+dt
j
s,t
= j
s,t
j
t,t+dt
j
s,t
j
s,t
I
t,t+dt
j
s,t
= j
s,t
(I
t,t+dt
1).
If we put dI
t
= I
s,t+dt
I
s,t
(independent of s < t), this gives a heuristical meaning
to
j
s,t
= 1 +
_
t
s
j
s,r
dI
r
as a quantum stochastic integral. We remark that this interpretation as an integral
is not limited to the above choice. Whenever k is a transformed process obtained
from j via (3.5), we formally may write
k
s,t
= 1 +
_
t
s
k
s,r
(dj
t
),
where dj
t
:= j
t,t+dt
1.
Example 3.11. (Classical Levy processes and unitary evolutions)
Let G be a topological group and denote by !(G) the space of all coecient
functions of continuous nite-dimensional representations of G. Then f !(G)
i there are n N and continuous complex-valued functions f
1
, . . . f
n
, g
1
, . . . g
n
on
G such that
f(xy) =
n
i=1
f
i
(x) g
i
(y) x, y G.
!(G) is a commutative algebra. By setting
f =
n
i=1
f
i
g
i
, f = f(e)
!(G) becomes a commutative Hopf algebra. In various cases (e.g., when G is
compact or locally compact abelian) the group G is uniquely determined by !(G).
Let us assume that G is compact. Then !(G) is the Kren algebra of G.
A classical Levy process X
t
on G gives rise to a quantum Levy process j
t
on !(G) by putting j
t
(f) = f X
t
. Here j
t
= j
0.t
and j
s,t
= (j
s
S) j
t
where S is the antipode of !(G). Let us specialize further to the case when G
is the group |
d
of unitary d d-matrices. Then !(G) equals the Hopf algebra
C[x
kl
, x
kl
; k, l = 1, . . . d] divided by the ideal generated by the elements which are
the entries of the matrices xx
1 and x
x1 where we put x = (x
kl
)
k,l=1,...d
. The
comultiplication is given by x
kl
=
d
i=1
x
ki
x
il
and the counit by x
kl
=
kl
.
The antipode is given by S(x
kl
) = x
lk
. By replacing the commuting indeterminates
x
kl
by non-commuting indeterminates, we dene a non-commutative bialgebra
Cx
kl
, x
kl
; k, l = 1, . . . d/xx
= 1, x
x = 1
which we denote by |d; cf. [11]. (It is easy to see that |d is not a Hopf
algebra.)
TRANSFORMATION OF QUANTUM L
)
kl
+ i H
kl
;
cf. [11] and [3]. The generator process (cf. the previous example) is given by
matrices I
s,t
M
d
(C) (L
2
(R
+
, K)) with
(I
s,t
)
ij
= A
s,t
((W
L)
ji
) +
s,t
((W 1)
ij
) +A
s,t
(L
ij
) +(i H
1
2
(LL
))
ij
(t s)
Theorem 3.4 says that
I
t0,t1
I
t1,t2
. . . I
tn1,tn
converges, again in the sense of the remark after Theorem 3.4, to the Levy process
U
s,t
which is the unitary process on C
d
(L
2
(R
+
, K)) given by (U
s,t
)
ij
= j
s,t
(x
ij
).
This is a generalization of a construction already given by W. von Waldenfels [16].
A classical Levy process on |
d
is a special case of a QLP on |d.
Example 3.12. (Azema martingales)
Consider the algebra Cx, x
) y
k
) =
_
1 if M(x, x
) = xx
0 otherwise
M(x, x
) / a monomial in x and x
, k N
0
, then K = C, (x
) = 1, (x) = 0,
(x) = 0 and (y) = q. The linear functional is the generator of the quantum
q-Azema martingale (X
t
, X
t
, Y
t
) if we consider the Azema bialgebra, and it
generates the process (A
t
, A
t
, Y
t
) in the case of the primitive/group-like structure
of / where Y
t
is the second quantization of multiplication by q 1
[0,t]
. The process
X
t
satises the quantum stochastic dierential equation
dX
t
= (q 1)X
t
d
t
+ dA
t
, X
0
= 0;
see [8, 10].
An application of Theorem 3.4 yields the formula
W
t
= lim
n1
j=0
Z
tj,tj+1
568 MICHAEL SCH
2
respectively. This can be done in the framework of transformation in at least
three dierent ways.
1. For two bialgebras B
1
and B
2
we can form the tensor product B
1
B
2
as
the bialgebra with comultiplication coming from the coalgebra tensor product
(cf. Section 2) and the usual algebra structure on B
1
B
2
(i.e. (b
1
b
2
)(c
1
c
2
) =
b
1
c
1
b
2
c
2
and (b c)
= b
). For generators
1
,
2
on B
1
, B
2
we form the
generator
:=
1
2
+
1
2
on B
1
B
2
. Indeed, is hermitian with (1) = 0, and if
b B
1
B
2
, (
1
2
)(
b) = 0
b) = (
1
2
)(
b) + (
1
2
)(
b)
=
1
_
(id
2
)(
)(id
2
)(
b)
_
+
2
_
(
1
id)(
)(
1
id)(
b)
_
0
since
1
(id
2
)(
b) = 0 =
2
(
1
id)(
b)
so that
is conditionally positve.
For QLPs j
(1)
s,t
, j
(2)
s,t
on B
1
, B
2
over (D
1
,
1
), (D
2
,
2
) with generators
1
,
2
we
form the process
j
s,t
on B
1
B
2
j
s,t
(b
1
b
2
) = j
(1)
s,t
(b
1
) j
(2)
s,t
(b
2
).
Then
j
s,t
is a QLP on B
1
B
2
over (D
1
D
2
,
1
2
) and consists of two
independent components j
(1)
s,t
and j
(2)
s,t
. Moreover, the convolution semigroup of
j
s,t
is given by
t
(b
1
b
2
) =
(1)
1
(b
1
)
(2)
2
(b
2
)
TRANSFORMATION OF QUANTUM L
)(b
1
b
n
) =
_
(b
1
1 +1 b
1
) (b
n
1 +1 b
n
)
_
=
A{1,...,n}
(b
A
b
A
c )
(3.8)
where for a subset i
1
< < i
l
of 1, . . . , n we put b
A
= b
i1
b
i
l
, b
= 1.
Since b
1
, . . . , b
n
are in the kernel of the counit we have that (3.8) is equal to
1
(b
1
b
n
) +
2
(b
1
b
n
) and
= (
1
+
2
) M. Denote by J
s,t
the QLP
on T(B
0
) with generator (
1
+
2
) M. By Proposition 3.3 the restriction j
s,t
of
J
s,t
to B T(B
0
) is a QLP on B with generator
1
+
2
. Equation (3.5) becomes
(b B
0
)
j
s,t
(b) = lim
_
j
(1)
t0,t1
(b
(1)
) 1 +1 j
(2)
t0,t1
(b
(1)
) (b
(1)
)1 1
_
_
j
(1)
tn1,tn
(b
(n)
) 1 +1 j
(2)
tn1,tn
(b
(n)
) (b
(1)
)1 1
_
= lim
_
(j
(1)
t0,t1
)(b
(1)
) 1 +1 j
(2)
t0,t1
(b
(1)
)
_
_
(j
(1)
tn1,tn
)(b
(n)
) 1 +1 j
(2)
tn1,tn
(b
(n)
)
_
= lim
_
j
(1)
t0,t1
(b
(1)
) 1 +1 (j
(2)
t0,t1
)(b
(1)
)
_
_
j
(1)
tn1,tn
(b
(n)
) 1 +1 (j
(2)
tn1,tn
)(b
(n)
)
_
.
If we take equidistant partitions we obtain
, j
s,t
(b) = lim
n
_
(1)
ts
n
+
(2)
ts
n
_
n
(b)
= e
(ts)(1+2)
(b)
which, of course, also follows from Theorem 3.4.
2. There is a multiplicative version of the above construction. Dene the
algebra homomorphism : CB
1
B B by
(b) = b b
for b B
1
. We use the notation of Example 3.8. One proves that =
: T(B
0
) B B is a transformation of B B, and that
= M. This
time equation (3.5) can be written (b B
1
)
j
s,t
(b) = lim
i
_
j
(1)
t0,t1
(b
1,i
) j
(2)
t0,t1
(b
1,i
) j
(1)
tn1,tn
(b
n,i
) j
(2)
tn1,tn
(b
n,i
)
_
with
n
b =
i
b
1,i
b
n,i
, b
l,i
B
1
,
i
C.
570 MICHAEL SCH
j
(1)
t0,t1
(b
(1)
) j
(2)
t0,t1
(b
(2)
) j
(1)
tn1,tn
(b
(2n1)
) j
(2)
tn1,tn
(b
(2n)
).
In the case of Example 3.11 this gives exactly the formula of [6] for a nite di-
mensional initial space. If B is cocommutative (in which case is a coalgebra
homomorphism!) we obtain j
s,t
= (j
(1)
s,t
j
(2)
s,t
) , and the Trotter formula be-
comes trivial.
The three constructions of a quantum Levy process with generator
1
+
2
depend on the choice of the rst bialgebra and the transformation . The most
natural seems to be the one of 3. where is the comultiplication itself whereas
in cases 1. and 2. the transformation does not depend on . In 1. and 2.
the two original processes are rst put together in an additive and group-like way
respectively. Case 3. is a real Trotter formula for exponentials given by the
comultiplication .
4. Proof of Theorems
In principle, Theorem 3.4 is proved if we show that the nets in (3.4) are Cauchy.
To show this, in Section 4.1 we prove a lemma about innitesimal products in
Banach algebras (an extension of ideas in V. Liebscher and M. Skeide [5]) and
a coalgebra version, appealing to the Fundamental Theorem on Coalgebras; see
L. Accardi, M. Sch urmann, and W. von Waldenfels [1]. These lemmas plus the
algebraic Proposition 4.3 allow to prove Proposition 4.4, which is the analytic
heart of the proof of Theorem 3.4.
4.1. Preparatory lemmas. We start with a lemma that imitates, like in [5],
proofs of the Trotter product formula.
Lemma 4.1. Let / be a Banach algebra. Suppose we have a constant R > 0 and
a family
_
A
()
_
M
of functions (M being some index set)
r A
()
r
= I + rG +S
()
r
/
on R
+
where G / and S
()
r
satises
_
_
_S
()
r
_
_
_ r
2 C
2
2
for some constant C
not depending on M and all r R. Then for all intervals [s, t] R
+
, all
partitions = s = t
0
< t
1
< < t
n1
< t
n
= t (n N) of [s, t] with || R,
and an arbitrary choice of elements
1
, . . . ,
n
of M, we have
_
_
_A
(1)
t1t0
A
(n)
tntn1
e
(ts)G
_
_
_ || (ts)e
(ts) max(G,C)
C
2
+|G|
2
e
G
2
.
Proof. By assumption
_
_
_A
(
k
)
r
_
_
_ 1 + r |G| + r
2 C
2
2
e
r max(G,C)
, and thus
_
_
_A
(
)
t
t
1
A
(
k
)
t
k
t
k1
_
_
_ e
(t
k
t
1
) max(G,C)
TRANSFORMATION OF QUANTUM L
j=1
A
(1)
t1t0
A
(j1)
tj1tj2
_
A
(j)
tjtj1
e
(tjtj1)G
_
e
(tj+1tj)G
e
(tntn1)G
.
We have
_
_
A
(j)
tjtj1
e
(tjtj1)G
_
_
_
_
A
(j)
tjtj1
I (t
j
t
j1
)G
_
_
+
_
_
I + (t
j
t
j1
)Ge
(tjtj1)G
_
_
(t
j
t
j1
)
2
C
2
+|G|
2
e
(tjtj1)G
2
.
From this estimate, from the estimate preceding it, and from
n
j=1
(t
j
t
j1
)
2
||
n
j=1
(t
j
t
j1
) = || (t s)
the statement follows.
There is a coalgebra version of Lemma 4.1 deduced from the Fundamental
Theorem on Coalgebras which yields that the coalgebra generated by a nite subset
of a coalgebra is nite dimensional. In the sequel, L(V, W) denotes the vector space
of linear maps between vector spaces V and W. We put L(V, V ) = L(V ). Let
((, , ) be a coalgebra and let L((, C) = (
n=0
T()Cc
n!
converges in any norm. By the
Fundamental Theorem on Coalgebras for every c ( such a (
c
exists. We deduce
that the series
e
(c) :=
n=0
n
n!
(c) = e
T()
(c) (4.1)
converges for all (
R
()
r
(c)
r
2
D
c
for some constant
D
c
> 0, depending on c ( but not on , and all r R. Then there exist
constants C
c
> 0 and
c
> 0 such that for all intervals [s, t] R
+
, all partitions
572 MICHAEL SCH
n
= s = t
0
< t
1
< < t
n1
< t
n
= t (n N) of [s, t] with || R, and an
arbitrary choice of elements
1
, . . . ,
n
of M, we have
f
(1)
t1t0
f
(n)
tntn1
(c) e
(ts)
(c)
|| (t s)e
(ts) max(c,Cc)
C
2
c
+
2
c
e
c
2
.
Proof. Choose b ( and x a nite-dimensional sub-coalgebra (
b
of ( containing
b. Fix a norm on (
b
. From the weak estimates
R
()
r
(c)
r
2
D
c
we easily con-
clude the strong estimate
_
_
_R
()
r
_
_
_ r
2
D for a suitable constant D for the linear
functionals R
()
r
on (
b
. (Just take your favorite elementary proof of the Uniform
Boundedness Principle for nite-dimensional Banach spaces.) Consider the linear
operator
A
()
r
:= T(f
()
r
) (
b
on (
b
, so
A
()
r
= I + rG +S
()
r
where G := T() (
b
and S
()
r
= T(R
()
r
) (
b
.
L((
b
) is a Banach algebra with respect to the operator norm. Since T is a
bijection from (
b
onto T((
b
) L((
b
), and since all norms on nite-dimensional
spaces are equivalent, S
()
r
satises
_
_
_S
()
r
_
_
_ r
2 C
2
2
for some constant C. In view
of lemma 4.1 we obtain the claimed statement if we choose C
c
= C
_
|| |c| and
c
= |G|
_
|| |c|.
4.2. Proof of Theorem 3.4. Consider the Hilbert subspaces (0 s t)
H
s,t
=span
_
j
t0,t1
(b
1
) j
tn1,tn
(b
n
):
n N, s = t
0
t
1
t
n
= t, b
1
, . . . , b
n
B
_
of D
j
where H
0
= C. Put H
t
= H
0,t
. Using the shift and the unit vector , we
dene mappings U
s,t
: H
s
H
t
H
s+t
by
U
s,t
(j
s0,s1
(b
1
) j
sn1,sn
(b
n
) j
t0,t1
(c
1
) j
tm1,tm
(c
m
))
= j
s0,s1
(b
1
) j
sn1,sn
(b
n
)j
t0+s,t1+s
(c
1
) j
tm1+s,tm+s
(c
m
)
where U
s,t
( ) = and b
1
, . . . , b
n
, c
1
, . . . , c
m
B, n, m N. Indeed, the
mappings U
s,t
are unitary. The shift is isometric and the unit vector is cyclic
which ensures surjectivity. Therefore, we may think of the family of Hilbert spaces
(H
t
)
t0
as a tensor product system in the sense of W. Arveson [2]; see M. Skeide
[14]. In fact, it is of type I which means that it comes from a Boson Fock space;
see Examples 3.10 and 3.9.
Let 0 s = t
0
t
1
t
n1
t
n
= t. Using the unitary isomorphism
H
t0,t1
H
t1,t2
H
tn1,tn
= H
s,t
, in the sequel, we identify
j
t0,t1
(b
1
) j
tn1,tn
(b
n
) = j
t0,t1
(b
1
) j
tn1,tn
(b
n
). (4.2)
TRANSFORMATION OF QUANTUM L
(c),
(d) e
(ts)
(c
d)
(n)
= s
n1
= t
(n)
0
< t
(n)
1
< < t
(n)
kn1
< t
(n)
kn
= s
n
(c),
(d). Then, by
Proposition 4.3,
L
,
= L
{s0,s1},
(1) . . . L
{s
l1
,s
l
},
(l) .
In the expression for L
{sn1,sn},
(n) we have
j
sn1,sn
(c) = (j
t
(n)
0
,t
(n)
1
. . . j
t
(n)
kn1
,t
(n)
kn
) (c)
574 MICHAEL SCH
on B ( by
M
(b c) := j
t0,t1
. . . j
tm1,tm
(b),
(c)
then, again by Proposition 4.3,
L
{sn1,sn},
(n) (c d) = M
{t
(n)
0
,t
(n)
1
}
. . . M
{t
(n)
kn
,t
(n)
kn1
}
((c) d).
For [0, ||] we dene L
(n)
:= L
{sn1,sn1+},
(n)
()
, where
(n)
() :=
_
[s
n1
, s
n1
+ ]
(n)
_
_
s
n1
+
_
.
(Roughly speaking, if s
n
s
n1
, then
(n)
() concides with the part of
(n)
up to s
n1
+ , and otherwise it adds another interval to the partition.)
We dene the linear functionals M
r
:= M
{,+r}
on B (. Note that these do
not depend on 0. We nd
M
r
(b c) = M
{,+r}
(b c)
= j
,+r
(b), j
,+r
(c) =
r
(b
(c)) =
_
( ) +rG+R
r
_
(b c),
where G(b c) := (b
(c)) and R
r
fullls the condition of Lemma 4.2. For xed
[s, t], it follows that for every c d ( ( there exists a constant C
c,d
such that
L
(n)
(c d) e
G
((c) d)
|
(n)
()| C
c,d
2
C
c,d
for all partitions
(n)
of [s
n1
, s
n
]. (The constant C
c,d
might depend on [s, t].)
From this it is routine to conclude that the L
(n)
d).
By taking (nite!) linear combinations, we obtain suitable constants D
for every
( (. From this the statement follows.
Corollary 4.5. The net
_
(c)
_
Zst
is a Cauchy net.
Proof. We have to show that for > 0 there is a such that , Z
st
, ~
and ~ , implies |
(c)
(c),
(c) e
(ts)
(c
c)
<
2
16
. (4.4)
So, for ~ we have
|
(c)
(c)|
2
=
(c),
(c) +
(c),
(c)
(c),
(c)
(c),
(c)
2
4
.
TRANSFORMATION OF QUANTUM L
(c)
(c)| |
(c)
(c)| +|
(c)
(c)|
which nishes the proof.
The limit of the Cauchy net
_
(c)
_
Zst
in D
j
will be denoted by
s,t
(c).
Remark 4.6. Taking the limit of (4.3) over ~ for xed , we nd the same
estimate for
(c),
s,t
(c). The fact that (4.3) does not depend on the precise
form of but only on its width || and computations similar to the proof of the
corollary, show that |
(c)
s,t
(c)| is small, whenever || is suciently small.
In particular, it follows that
lim
n
n
(c) =
s,t
(c)
for each sequence
n
in Z
st
with lim
n
[[
n
[[ = 0.
To conclude the proof of Theorem 3.4, we start by observing that
s,t
(c) =
t0,t1
. . .
tn1,tn
. (4.5)
(To see this, simply take the limit of
t0,t1
(c
1
) . . .
tn1,tn
(c
n
): c
1
, . . . , c
n
(
_
.
By (4.5),
s,t
(c) =
t0,t1
. . .
tn1,tn
it follows ~ = D
k
D
k
. We
put D
ks,t
:=
D
k
. Of course, [s
, t
] [s, t] = D
k
s
,t
D
ks,t
. We put
D
kt,
:=
tr<s
D
kr,s
and D
k
:= D
k0,
. On D
ks,t
we dene an operator by
setting
t0,t1
(c
1
)
tn1,tn
(c
n
)
t0,t1
(c
(1)
c
1
)
tn1,tn
(c
(n)
c
n
).
To see that this is well-dened, we simply observe that the operator has a formal
adjoint on that domain, namely, simply the operator whith c replaced by c
. (By
taking joint renements, if necessary, we may assume that the two vectors we
choose to check the adjoint condition are in the same D
k
.)
We extend this operator by amplication to an operator k
s,t
(c) on D
k
= D
k0,s
D
ks,t
D
kt,
. Clearly, c k
s,t
(c) is multiplicative, so that the k
s,t
dene a family
of homomorphisms.
A simple application of coassociativity (and, once more, (4.5)) shows that k
r,s
k
s,t
= k
r,t
for r < s < t. Therefore, the family of mappings k
s,t
forms a Levy
process on ( over (D
k
, ) with generator . That D
k
is dense in H
k
, will
follow from the proof of Theorem 3.6.
4.3. Proof of Theorems 3.6 and 3.7. By Theorem 3.4 we know that for B
T(B
0
)
(B) = (k
t0,t1
)
T(0)
T(0)
(k
tn1,tn
)(B)
converges in norm and denes a cyclic QLP J
s,t
on D
J
. We have ,B = id
and ( ) = M is the generator of J
s,t
. by Proposition 3.3 the restriction
of J
s,t
to B is a QLP
j
s,t
with generator . For the proof of
j = j it suces to
show that |
j
s,t
(b)|
2
0. So it remains to be shown
576 MICHAEL SCH
(b), j
s,t
(d) = e
(ts)
(b
d).
Proof. Let = s = t
0
< t
1
< . . . < t
n1
< t
n
= t and write j
s,t
= j
t0,t1
. . .
j
tn1,tn
. Then, as in the proof of Proposition 4.4, from Proposition 4.3 we nd
(b), j
s,t
(d) = L
t1t0
. . . L
tntn1
(b d),
where we dene the linear functionals L
r
(bd) := k
0,r
(b), j
0,r
(d) on BB.
We are done, if we show that the L
r
fulll the conditions of Lemma 4.2 with
the correct linear term. In fact, if in (4.3) we insert = 0, r (so that || = r)
and perform the limit over , the estimate remains valid for
{0,r}
(d), k
0,r
(b) = L
r
(b d).
This ends also the proof of Theorem 3.7. We have
Corollary 4.8. The vectors k
s,t
, c (, generate D
j
in the sense that
D
j
= D
k
= span
_
k
t0,t1
(c
1
) k
tn1,tn
(c
n
):
n N, 0 s t < , s = t
0
t
1
t
n1
t
n
= t, c
1
, . . . , c
n
(
_
.
Acknowledgments. We wish to thank the referee for a number of very insightful
comments that helped a lot to improve the presentation.
References
1. Accardi, L., Sch urmann, M., and von Waldenfels, W.: Quantum independent increment
processes on superalgebras, Math. Z. 198 (1988) 451477.
2. Arveson, W.: Noncommutative Dynamics and E-Semigroups, Monographs in Mathematics,
Springer, New York Berlin Heidelberg, 2003
3. Franz, U.: Levy processes on quantum groups and dual groups, in: M. Sch urmann, U.
Franz, editors, Quantum Independent Increment Processes II, Lect. Notes Math. 1866 161
257 (2006), Springer, New York Berlin Heidelberg.
4. Hudson, R. L. and Parthasarathy, K. R.: Quantum Itos formula and stochastic evolutions,
Commun. Math. Phys. 93 (1984) 301323.
5. Liebscher, V. and Skeide, M.: Constructing units in product systems, Proc. Amer. Math.
Soc. 136 (2008) 989997.
6. Lindsay, J. M. and Sinha, K. B.: A quantum stochastic Lie-Trotter product formula, Indian
J. Pure Appl. Math. 41 (2010) 313325.
7. Lindsay, J. M. and Skalski, A.: Quantum stochastic convolution cocycles II, Commun. Math.
Phys. 280 (2008) 575610.
8. Parthasarathy, K. R.: Azema martingales and quantum stochastic calculus, in: R.R. Ba-
hadur, editor, Proc. R.C. Bose Memeorial Symposium (1990), Wiley Eastern, New Delhi.
9. Parthasarathy, K. R. and Sunder, V. S.: Exponentials of indicator functions are total in
the boson Fock space (L
2
[0, 1]), in: R. L. Hudson and J. M. Lindsay, editors, Quantum
Probability Communications X (1998) 281284, World Scientic, Singapore.
10. Sch urmann, M.: Quantum q-white noise and a q-central limit theorem, Commun. Math.
Phys. 140 (1991) 589615.
11. Sch urmann, M.: White Noise on Bialgebras, Lect. Notes Math., vol. 1544, Springer, New
York Berlin Heidelberg, 1993.
12. Sch urmann, M. and Vo, S.: Positivity of free convolution semigroups, in: J.C. Garcia, R.
Quezada, S.B. Sontz, editors, Quantum Probability and Related Topics, Proceedings of the
28th Conference (2008) 225241, World Scientic, Singapore.
TRANSFORMATION OF QUANTUM L
t
is completely positive for all t 0.
The notion of domination or subordination goes back at least as far as work
of Arveson. More recent generalisations include subordination in the theory of
completely positive denite kernels (i.e. maps k : S S B(/; B) satisfying a
suitable positivity requirement, where S is a set and /, B are C
-algebras) and
semigroups of these ([4], Section 3.3). A characterisation of the subordinates of a
given CP map is the content of Arvesons Radon-Nikod ym Theorem ([2], The-
orem 1.4.2). In this paper we will look only at subordination of one E-semigroup
by another, which gives the opportunity to provide the following short proof of
Arvesons result in this setting.
Lemma 1.3. Let H be a Hilbert space and let , be a pair of endomorphisms on
B(H). The following are equivalent:
(i) ;
(ii) (I)
_
B(H)
_
_
I
S
_
=
_
I S
S
S
_
E-SEMIGROUPS SUBORDINATE TO CCR FLOWS 581
and since the dierence is 2-positive we have
0
_
(I) 0
0 I
_
( )
(2)
__
I S
S
S
___
(I) 0
0 I
_
=
_
(I)(I)(I) (I)
3
(I)
_
(S) (S)
_
_
(S
) (S
)
_
(I) (S
S) (S
S)
_
=: R.
But (I) (I), and these operators are projections, hence
(I)(I)(I) = (I) = (I)
3
.
Thus the top-left hand corner of the positive 2 2 operator matrix R is zero.
Writing this as R = T
T for T = [
A B
C D
], where A, B, C, D B(H), it follows that
A = C = 0, and so the o-diagonal entries of R are also zero. Hence (I)(S) =
(S) = (S)(I) and thus (i) implies (ii).
That (ii) implies (i) follows from the nal part of the lemma concerning the
map , and this follows since for such projections P we have
(S) (S) = P
(S)P
,
so that is CP.
Remark 1.5. (a) In proving that (i) implies (ii) we do not use the full force of
the assumption that is CP, but in fact only require that this dierence be
2-positive (and can then conclude that the dierence is in fact CP). Such a feature
is also found in Bhats generalisation of this lemma ([5], Proposition 4.2) where he
starts with a unital endomorphism and a CP map , and, assuming only that
is positive, shows that the dierence is CP, with (S) = (I)(S) where
(I)
_
B(H)
_
is a positive contraction.
(b) In the above, if and are endomorphisms with , then is determined
by the value of the projection (I). On the other hand, unless (I) = I, dierent
projections in
_
B(H)
_
_
B(H)
_
, then
1
=
2
if and only if there are projections p (I) and q
1
, q
2
(I)
such that P
1
= p +q
1
and P
2
= p +q
2
.
Example 1.6. Let H be a Hilbert space and L B(H) an isometry. Let (S) :=
LSL
. The corresponding
subordinates of are and 0 respectively.
To apply the lemma to E-semigroups in Theorem 1.9 below (which is a special
case of Theorem 4.3 of [5]) we require a little more terminology.
Denition 1.7. Let H be a Hilbert space and an E-semigroup on B(H). A
family of contraction operators X = (X
t
)
t0
on H is a left -cocycle if
(i) X
0
= I;
(ii) X
s+t
= X
s
s
(X
t
) for all s, t 0;
(iii) the maps t X
t
and t X
t
are continuous in the strong operator
topology.
582 STEPHEN J. WILLS
If, instead, (ii) is replaced by
(ii)
X
s+t
=
s
(X
t
)X
s
for all s, t 0
then we speak of a right -cocycle. A local -cocycle is a cocycle for which X
t
t
_
B(H)
_
t
(S) = B
t
t
(S) for all S B(H).
Moreover, if C = (C
t
)
t0
is any projection-valued, local -cocycle, then
t
(S) :=
C
t
t
(S) denes an E-semigroup subordinate to .
Proof. Using Lemma 1.3, the proof is now just a matter of checking that denitions
hold. For example if (i) holds, then each B
t
=
t
(I)
t
_
B(H)
_
and is projection-
valued, with
B
s+t
=
s+t
(I) =
s
_
t
(I)
_
=
s
(I)
s
_
t
(I)
_
= B
s
s
(B
t
).
Since B
s
B
t
whenever s t, and t B
t
is weak operator continuous, it is
strong operator continuous and thus a local -cocycle.
Remark 1.10. (a) The lack of injectivity for single maps noted after Lemma 1.3
does not arise for E-semigroups and cocycles: if P is a projection-valued, local
-cocycle, then
P
t
= P
t
t
(P
0
) = P
t
t
(I) so P
t
t
(I).
(b) It follows from this result that the only E
0
-semigroup subordinate to an
E
0
-semigroup is itself.
Example 1.11. Let H be a Hilbert space and L = (L
t
)
t0
a strongly continuous
semigroup of isometries on H. Put
t
(S) := L
t
SL
t
, then is an E-semigroup. The
only projection-valued, local -cocycle is (
t
(I))
t0
, so that the only subordinate
E-semigroups of is itself.
E-SEMIGROUPS SUBORDINATE TO CCR FLOWS 583
2. Cocycles of CCR Flows
The most understood class of E-semigroups are the CCR and CAR ows; these
comprise the type I examples, with the work of Powers et al. ([12, 1, 11]) designed
to yield new examples of type II semigroups. Here we revert to a consideration of
the CCR ow, bringing stochastic methods to bear.
Let h and k be a pair of Hilbert spaces, called the initial space and noise
dimension space respectively. For each measurable set I [0, [ let T
I
denote
the symmetric/bosonic Fock space over L
2
(I; k), with T
R+
abbreviated to T, so
that
T
= T
I
T
I
c .
This isomorphism is conveniently eected via the correspondence
(f) (f[
I
) (f[
I
c )
dened in terms of the useful total set of exponential vectors:
c
S
:= (f) : f S, where (f) := (1, f, (2!)
1/2
f f, (3!)
1/2
f f f, . . .),
and where S is any suciently large subset of L
2
(R
+
; k). For the purposes of this
paper we shall write c for c
S
in the case that S is the set of (right-continuous) step
functions in L
2
(R
+
; k). The natural unitary isomorphism L
2
(R
+
; k)
= L
2
([t, [; k)
gives rise to the identication
T
= (0) T
[t,[
T
[0,t[
T
[t,[
= T,
and thence the right-shift map
k
t
on B(T) that has image I
[0,t[
B(T
[t,[
). This
map is a normal
-homomorphism. The CCR ow determined by h and k is then
h,k
=
_
h,k
t
:= id
B(h)
k
t
_
t0
.
[We now clarify tensor product notation: denotes the algebraic tensor product,
is used for the tensor product of Hilbert spaces and the spatial tensor product,
whereas denotes the ultraweak tensor product.]
In general we will drop the dependence of the CCR ow on h and k in what
follows, referring, for example, to -cocycles.
Denition 2.1. A -cocycle X is:
(a) adapted if X
t
B(h T
[0,t[
) I
[t,[
for all t 0,
(b) Markov-regular if the expectation semigroup T
t
:= E[X
t
] on h is norm-
continuous, where the vacuum conditional expectation E is dened by
u, E[S]v = u (0), Sv (0), u, v h.
Remark 2.2. A -cocycle X is local if X
t
I
h
B(T
[0,t[
)I
F
[t,[
for each t, which
is a strictly stronger condition than adaptedness whenever dimh > 1.
The following combines Theorem 6.6 of [10] with Theorem 7.5 of [9]. We dene
k := C k, and := I
h
P, where P is the projection
k 0 k.
Theorem 2.3. There is a one-to-one correspondence between the set of Markov-
regular, contraction, left, adapted -cocycles and
C
0
(h, k) := F B(h
k) : F +F
+F
F 0,
584 STEPHEN J. WILLS
under which cocycles X are associated to their stochastic generatorF through
X X
F
, where X
F
is the unique solution of the left Hudson-Parthasarathy
quantum stochastic dierential equation
X
t
= I +
_
t
0
X
s
(F I
F
) d
s
, (2.1)
with
X
s
denoting the ampliation of X
s
B(h T) to all of B(h
k T).
The theorem above says that every such cocycle satises the equation (2.1),
and that all (contractive) solutions of this equation are indeed left cocycles. In the
case when h is nite dimensional or the cocycle is local, the fact that we assumed
that t X
t
is strongly continuous is enough to guarantee that the cocycle is
Markov-regular, but this condition is a nontrivial requirement otherwise. Our
basic reference for quantum stochastic calculus is [8], where the same notations
are used.
Locality of -cocycles is also easily characterised.
Lemma 2.4. Let F C
0
(h, k) and let X = X
F
be the associated left -cocycle.
Then X is local if and only if F I
h
B(
k).
Proof. Since
t
_
B(hT)
_
= I
h
B(T
[0,t[
) I
[t,[
, the result follows from Corol-
lary 6.5 of [10].
Hudson and Parthasarathy focused on determining those F B(h
k) that give
rise to unitary solutions of (2.1). (Co)isometry and contractivity were considered
later, with positive contraction-valued and projection-valued adapted -cocycles
studied in [14] (although for the former see [5] for the case when h = C). In
Section 5.2 of [4] the morphisms of time ordered Fock modules are characterised;
Theorem 4.4.8 of the same paper gives a one-to-one correspondence between en-
domorphisms of product systems of Hilbert modules and local cocycles of an E
0
-
semigroup on a related C
= h (h k),
so that
F =
_
A B
C D
_
for A B(h), B B(h k; h), C B(h; h k) and D B(h k).
Proposition 2.5. Let F C
0
(h, k). The following are equivalent:
(i) X = X
F
is projection-valued;
(ii) F A B(
F = 0;
(iii) F =
_
L
L L
L PI
A B(
.
In this case projection-valued, adapted -cocycles, which are now all local, are
indexed by pairs (P, u), where P B(h) is a projection, and u Ker P is a vector.
The new result in this section is the following.
Theorem 2.7. Let X
F
and X
G
be a pair of Markov-regular, projection-valued,
adapted -cocycles with generators
F =
_
L
L L
L P I
_
and G =
_
M
M M
M QI
_
. (2.2)
(a) The following are equivalent:
(i) X
F
t
X
G
t
for all t 0, equivalently X
F
t
X
G
t
= X
F
t
for all t 0.
(ii) G+GF = 0.
(iii) P Q and M = Q
L.
(b) Suppose that F, G A B(
k)
_
C
0
(h, k) such that
X
F
t
= (X
H
t
)
X
H
t
and X
G
t
= X
H
t
(X
H
t
)
D and Q = DD
.
In this case any H for which (2.3) holds has the form H =
_
A B
C DI
with D as
above and where
C = M +E, B = L
D, and A =
1
2
(E
E +BB
+C
C) +iK
for some E A [k satisfying Q
E = 0 and some K = K
A.
Remark 2.8. The condition E A [k eectively means that it can be written
in block form as a column with entries taken from A.
Proof. (a) Let =
i
u
i
(f
i
), =
j
v
j
(g
j
) hc. Then the rst and
second fundamental formulae of quantum stochastic calculus ([8], Theorems 3.13
and 3.15) give that
X
F
t
, X
G
t
, X
F
t
is equal to
_
t
0
_
X
F
s
(s), (GI
F
)
X
G
s
(s) +(F I
F
)
X
F
s
(s),
X
G
s
(s)
+(F I
F
)
X
F
s
(s), (GI
F
)
X
G
s
(s) (s), (F I
F
)
X
F
s
(s)
_
ds
(2.4)
=
_
t
0
_
(s),
X
F
s
[(F +G+FG) I
F
]
X
G
s
(s) (s), (F I
F
)
X
F
s
(s)
_
ds
586 STEPHEN J. WILLS
where (s) =
i
u
i
f
i
(s) (f
i
) and similarly for (s), with
d :=
_
1
d
_
for any
d k. Thus if (i) holds, then the integral above is zero, and since the integrand is
continuous in a neighbourhood of 0 we get
(0), [(F +G+FG) I
F
](0) (0), (F I
F
)(0) = 0
from which (ii) follows since (0) and (0) range over a total subset of h
k T.
Let A be a commutative von Neumann algebra as in Proposition 2.5. To
see that (ii) implies (i), note that in (2.4) the term in square brackets equals
F I
F
(A B(
k)) I
F
, which commutes with
X
F
s
(A I
k
)B(T). Con-
sequently X
F
X
G
and X
F
are both solutions of the right version of the Hudson-
Parthasarathy QSDE (2.1) for the same coecient F, and so uniqueness of solu-
tions gives (i).
Finally, (ii) is equivalent to the following four equations being satised:
M
L = M
M, M
P = 0, M = Q
L and PQ = P.
That all four together are equivalent to just the last two is a consequence of the
fact that QM = 0, equivalently Q
M = M, and also PL = 0.
(b) This time, for any H
_
A B(
k)
_
C
0
(h, k), it follows that (X
H
)
X
H
and
X
H
(X
H
)
+
H
H and H +H
+HH
+H
H = F and H +H
+HH
= G.
Writing H in block form as in the statement, this becomes
_
A+A
+C
C B +C
D
B
+D
C D
D I
_
=
_
L
L L
L P I
_
(2.5)
and
_
A+A
+BB
+BD
C +DB
DD
I
_
=
_
M
M M
M QI
_
. (2.6)
This shows immediately that (i) implies (ii).
To see that (ii) implies (i), we have to satisfy a total of eight equations when
comparing components in (2.5) and (2.6). However, (ii) is merely the statement
that it is possible to nd D A B(k) such that the bottom-right components
are equal in both equations. Comparing bottom-left components we need
B
+D
C = L and C +DB
= M. (2.7)
But note that DD
= Q and D
)C = M DL so Q
C = M.
E-SEMIGROUPS SUBORDINATE TO CCR FLOWS 587
Since Q
= L D
M D
E so B = L
D.
Now the bottom-left components of (2.5) are equal by construction, hence so are
the top-right components. For the o-diagonal components in (2.6) we have
C +DB
= M +E +DL DD
E = M +DL +Q
E = M
as required.
Finally, to ensure that the top-left components in (2.5) and (2.6) are equal, it is
clear that twice the real part of A must equal both L
LC
C and M
MBB
,
with the imaginary part unconstrained. That these two conditions can be satised
simultaneously follows since
L
L C
C = L
L (M +E)
(M +E)
= L
L M
M E
E M
E E
M
= L
L M
M E
E,
since M
E = (Q
M)
E = M
E = 0, and similarly
M
M BB
= M
M (L
D)(L
D)
= M
M L
L E
QE +E
DL + (DL)
E
= L
L M
M E
E.
Thus, if (ii) holds, that is if P and Q are equivalent in A B(k), then it is pos-
sible to ll out the matrix for H in a way such that (2.5) and (2.6) both hold,
hence (i) holds. Moreover, the computations above reveal all possible solutions to
the problem.
Remark 2.9. (a) Part (a) above is in some sense more satisfactory that part (b)
since we must have F /B(
k) and G A B(
+H
H +HH +HH
+HH
H = 0.
588 STEPHEN J. WILLS
The keen reader is invited to check directly that an H satisfying the structure
relations in part (b) does indeed satisfy this equation.
3. Relations on the Subordinates of an E-semigroup
For a given E-semigroup on some B(H), let Sub() denote the set of E-
semigroups that are subordinate to . We now turn our attention to possible
natural relations that can be dened on the set of all E-semigroups on B(H), or
perhaps just on Sub() for some such semigroup .
The most obvious is clearly the relation of subordination itself. It is imme-
diate from the denition that it is a partial order on the set of all E-semigroups,
hence on each subset Sub().
Lemma 3.1. Let be an E-semigroup and , Sub(). Then if and
only if
t
(I)
t
(I) for all t 0.
Proof. If , then in particular (
t
t
)(I) 0, and so
t
(I)
t
(I). Con-
versely, suppose that we have
t
(I)
t
(I) for each t. Now by Theorem 1.9, the
families (
t
(I))
t0
and (
t
(I))
t0
are projection-valued, local -cocycles, and also
each
t
(I)
t
(I) is a projection, with
t
(S)
t
(S) =
_
t
(I)
t
(I)
_
t
(S) =
_
t
(I)
t
(I)
_
t
(S)
_
t
(I)
t
(I)
_
for all S B(H), so that
t
t
is CP as required.
Remark 3.2. The temptation to adjust the hypotheses above and deal with all
E-semigroups rather than those from Sub() should be resisted. Whilst it is true
that given any two E-semigroups and , if , then
t
(I)
t
(I), the
converse is not true: if and are E
0
-semigroups, then obviously
t
(I) =
t
(I),
but we need not have = .
Let us apply this to the set Sub() for the CCR ow =
h,k
. By Theorem 1.9
any E-semigroup subordinate to is determined by a projection-valued, local -
cocycle (X
t
=
t
(I))
t0
, and from Lemma 2.4, Theorem 2.3 and Proposition 2.5,
this in turn is specied uniquely by a pair (P
, u
) where P
B(k) is a projection,
and u
Ker P
, through
X
= X
F
for F = I
h
_
|u
|
2
u
[
[u
_
. (3.1)
So, from part (a) of Theorem 2.7, we get the following.
Proposition 3.3. Suppose that , Sub() with associated projection-valued,
local -cocycles X
and X
and
(ii) u
= u
+u
for some u
(k) P
(k).
Example 3.4. Let h = C and k = L
2
[0, 1]. For each r [0, 1] let P
r
denote the
projection P
r
f = f1
[0,r]
on k. Then dene
(r)
to be the E-semigroup associated
to the projection-valued, local -cocycle with stochastic generator
F
r
=
_
0 0
0 P
r
I
_
.
E-SEMIGROUPS SUBORDINATE TO CCR FLOWS 589
It follows that for all r s in [0, 1] we have
(r)
(s)
, that is, Sub() contains an
uncountable, linearly ordered subset. This can only happen since we have taken
an innite dimensional k, although the noise dimension space k is separable. In
the case of a nite dimensional k the maximum number of distinct semigroups in
a chain in Sub() is 1 + dimk.
In the theory of E-semigroups, a trick to overcome such features of relations is
to consider semigroups up to some form of equivalence.
Denition 3.5. E-semigroups and on B(H) are cocycle conjugate if there is
a left -cocycle U such that
(i)
t
(S) = U
t
t
(S)U
t
for all t 0, S B(H), and
(ii)
t
(I) = U
t
U
t
.
This is denoted , or
U
if we want to highlight the particular cocycle.
Most of the literature in this area deals solely with E
0
-semigroups, when a
seemingly dierent denition of cocycle conjugacy is found: E
0
-semigroups and
are called cocycle conjugate if there is a unitary, left -cocycle U such that
condition (i) above holds. However, note that if and are E
0
-semigroups above,
then conditions (i) and (ii) together imply that the U
t
are unitaries. Since we are
forced to deal in this paper with general E-semigroups (recall Remark 1.10 (b)),
we need the version of cocycle conjugacy given in Denition 3.5, where the U
t
are
partial isometries courtesy of (ii), but not necessarily unitaries.
Proposition 3.6. The relation is an equivalence relation on the set of all E-
semigroups on B(H).
Proof. We have
U
for U
t
=
t
(I). If
U
, then
U
s+t
=
s
(U
t
)U
s
=
s
(I)
s
(U
t
)U
s
= U
s
U
s
s
(U
t
)U
s
= U
s
(U
t
),
so that U
t
(I)U
t
= (U
t
U
t
)
2
= U
t
U
t
, and
U
t
t
(S)U
t
=
t
(S), so that
U
. Similar computations show that if
U
and
V
, then
V U
, where V U = (V
t
U
t
)
t0
is a left -cocycle.
The following relation was then discussed in [13].
Denition 3.7. Let H be a Hilbert space and let , be E-semigroups on B(H).
Write if
for which
.
Proposition 3.8. The relation is reexive and transitive.
Proof. Reexivity is immediate: and , so that . To establish
transitivity assume that and . So there are E-semigroups
and
,
V
and
.
Set V
t
= V
t
t
(I), then, using Theorem 1.9,
V
s+t
= V
s
s
(V
t
)
s
_
t
(I)
_
= V
s
s
(V
t
)
s
_
t
(I)
_
s
(I) = V
s
(V
t
),
so that V
is a left
-cocycle. Moreover
(V
t
)
t
=
t
(I)V
t
V
t
t
(I) =
t
(I)
t
(I)
t
(I) =
t
(I).
590 STEPHEN J. WILLS
Consequently, if we dene
t
(S) = V
t
(S)(V
t
)
,
then
is an E-semigroup, with
by construction. Also
t
(S)
t
(S) = V
t
t
(S)V
t
V
t
(S)(V
t
)
= V
t
(
t
t
)(S)V
t
,
so that
. But then
and
, showing that .
However, as noted in [13], the problem is proving that is antisymmetric on
the set of all E-semigroups considered up to cocycle conjugacy. That is, if
and do we necessarily have ? New results of Liebscher suggests that
it is not antisymmetric ([7]).
Inspired by the results of Section 2, we instead introduce an alternative relation
on the subset Sub(), rather than deal with all E-semigroups simultaneously.
Denition 3.9. Let be an E-semigroup on B(H), and let , , Sub(). We
dene
t
(I) = U
t
U
t
and
t
(I) = U
t
U
t
for all t 0.
We write
U
to highlight the particular -cocycle U. In addition we dene
if there is some
with
.
These relations behave similarly to the versions without the superscript .
Proposition 3.10. The relation
U
, so
that
U
and
V
, then
V U
.
For the putative partial order
and
and
.
Set V
t
= V
t
t
(I), a pointwise product of local -cocycles, hence itself a local
-cocycle. Moreover
Ran
t
(I) Ran
t
(I) = initial space of V
t
,
so that V
t
(I). Let P
t
=
V
t
(V
t
)
Sub() be the E-
semigroup determined by this P through Theorem 1.9. Then
t
(I) = P
t
=
V
t
(V
t
)
, hence
. Moreover,
t
(I) = V
t
t
(I)V
t
V
t
V
t
=
t
(I)
so that
. Thus
as required.
For the CCR ow the missing piece of the puzzle is provided by Theorem 2.7.
E-SEMIGROUPS SUBORDINATE TO CCR FLOWS 591
Proposition 3.11. For the CCR ow , if , Sub() with
and
, then
.
Proof. Let , Sub() with associated projection-valued, local -cocycles X
and X
and F
, where
F
= I
h
_
|u
|
2
u
[
[u
_
, F
= I
h
_
|u
|
2
u
[
[u
_
,
in the notation (3.1). By part (b) of Theorem 2.7 we have
if and only if
P
and P
with
,
and
with
and X
, we have, now using part (a) of Theorem 2.7 as well as Lemma 3.1,
that
P
, P
, P
and P
.
That is, P
and P
, and so
as required.
Example 3.12. Let
(r)
: r [0, 1] Sub(
h,k
) be the E-semigroups from
Example 3.4 where h = C, k = L
2
[0, 1]. Now P
r
I for each r ]0, 1] (since
P
r
= T
r
T
r
and T
r
T
r
= I where (T
r
f)(s) =
rf(rs)), and so it follows that
(r)
.
Acknowledgment. Comments from Martin Lindsay, Michael Skeide and the
referee received during the writing of this paper resulted in several useful improve-
ments and clarications.
References
1. Alevras, A., Powers, R. T., and Price, G. L.: Cocycles for one-parameter ows of B(H), J.
Funct. Anal. 230 (2006) 164.
2. Arveson, W.: Subalgebras of C
,0
V
t
L
(t)
(dt), V
0
= 1
h
, (1.1)
(where the coecients L
(t) are constants ([6, 8, 10]). In particular, in the absence of the conservation
martingale, the equation takes the form
dV
t
=
j
V
t
L
j
(t)a
(dt) V
t
L
j
(t)a(dt) +V
t
G(t)dt
with the formal unitarity condition:
j
L
j
(t)L
j
(t) + 2Re G(t) = 0
for almost every t 0, in analogy with the case when L
Partially supported by Mid-career Researcher Program through NRF grant funded by the
MEST (No. R01-2008-000-10843-0).
* Supported by the UK- India Education and Research Initiative (UKIERI) project RA2007.
j
: j 1, respectively. For each A B(h 1) and u, v h we dene a linear
operator A(u, v) B(1) by
1
, A(u, v)
2
= u
1
, A v
2
,
1
,
2
1
and read o the following properties (for the proof, see Lemma 2.1 in [11]):
Lemma 2.1. Let A, B B(h 1). Then for any u, v, u
i
, v
i
h (i = 1, 2) we
have
(i) A(, ) : h h B(1) is a jointly continuous sesqui-linear map, and if
A(u, v) = B(u, v) for all u, v h, then A = B,
(ii) A(u, v) = F
u
AF
v
, |A(u, v)| |A||u||v| and A(u, v)
= A
(v, u),
(iii) A(u
1
, v
1
)B(u
2
, v
2
) = [A([v
1
>< u
2
[ 1
H
) B] (u
1
, v
2
),
(iv) AB(u, v) =
j1
A(u, e
j
)B(e
j
, v), where the series converges strongly,
(v) 0 A(u, v)
A(u, v) |u|
2
A
A(v, v),
UNITARY PROCESSES WITH INDEPENDENT INCREMENTS 595
(vi) for any
1
,
2
1 we have
A(u
1
, v
1
)
1
, B(u
2
, v
2
)
2
=
j1
u
2
j
, [B([v
2
>< v
1
[ [
2
><
1
[)A
] u
1
j
= v
1
1
, [A
([u
1
>< u
2
[ 1
H
)B] v
2
2
.
For each A B(h 1) and Z
2
= 0, 1, we dene an operator A
()
B(h 1) by
A
()
:=
_
A if = 0,
A
if = 1.
For 1 k n, we dene a unitary exchanging map P
k,n
: h
n
1 h
n
1
by
P
k,n
(u
1
u
n
) := u
k,n
(1)
u
k,n
(n)
on product vectors, where
k,n
is the permutation k, k + 1, , n, 1, . . . , k 1
of 1, 2, , n. Let = (
1
,
2
, ,
n
) Z
n
2
. Consider the ampliation of the
operator A
(
k
)
in B(h
n
1) given by
A
(n,
k
)
:= P
k,n
(1
h
n1 A
(
k
)
)P
k,n
.
Now we dene the operator
A
()
:=
n
k=1
A
(n,
k
)
:= A
(n,1)
A
(n,n)
as in B(h
n
1). Note that as here, through out this article, the product symbol
n
k=1
stands for product with the ordering from 1 to n. For product vectors
u, v h
n
one can see that
A
()
(u, v) =
_
n
i=1
A
(n,i)
_
(u, v) =
n
i=1
A
(i)
(u
i
, v
i
) B(1), (2.2)
moreover, for 1 m n, we see that
_
m
i=1
A
(n,i)
_
(u, v) =
m
i=1
A
(i)
(u
i
, v
i
)
n
i=m+1
u
i
, v
i
B(1). (2.3)
When = 0 Z
n
2
, for simplicity we shall write A
(n,k)
for A
(n,
k
)
and A
(n)
for A
()
.
3. Unitary Processes with Independent Increments
Let U
s,t
: 0 s t < be a family of unitary operators in B(h 1) with
U
s,s
= 1 for any s 0 and be a xed unit vector in 1. Let us consider the
family of unitary operators U
()
s,t
in B(h 1) for Z
2
given by U
(0)
s,t
= U
s,t
and U
(1)
s,t
= U
s,t
. As in Section 2, for xed n 1, Z
n
2
and each 1 k n, we
dene the families of operators U
(n,
k
)
s,t
and U
()
s,t
in B(h
n
1). By identity
(2.2), for product vectors u, v h
n
and Z
n
2
, we have
U
()
s,t
(u, v) =
n
i=1
U
(i)
s,t
(u
i
, v
i
).
596 UN CIG JI, LINGARAJ SAHU, AND KALYAN B. SINHA
We assume the following on the family of unitary U
s,t
B(h 1).
Assumption A:
(A1) (Evolution)
1
For any 0 r s t < , U
s,t
U
r,s
= U
r,t
and U
s,s
= 1,
(A2) (Independence of increments) for any 0 s
i
t
i
< (i = 1, 2) such
that [s
1
, t
1
) [s
2
, t
2
) = ,
(i) U
s1,t1
(u
1
, v
1
) commutes with U
s2,t2
(u
2
, v
2
) and U
s2,t2
(u
2
, v
2
) for any
u
i
, v
i
h (i = 1, 2).
(ii) For pairs (u
i
, v
i
) and (p
j
, w
j
) h(i = 1, 2, . . . , n; j = 1, 2, . . . , k) and
[a, b) and [r, s) disjoint intervals,
_
,
n
i=1
U
(i)
a,b
(u
i
, v
i
)
k
j=1
U
(
j
)
r,s
(p
j
, w
j
)
_
=
_
,
n
i=1
U
(i)
a,b
(u
i
, v
i
)
_ _
,
k
j=1
U
(
j
)
r,s
(p
j
, w
j
)
_
.
Assumption B: (Regularity) for any > t s 0,
sup [, (U
s,t
1)(u, v)[ : |u| = |v| = 1 C[t s[
for some positive constant C independent of s, t.
Remark 3.1. Similar sets of assumptions of independence can also be found in
the analysis of Levy processes([4]).However here,unlike in [11, 12], the stationarity
condition is not assumed.
As in [11, 12], we need further assumptions for Gaussianity and minimality:
Assumption C: (Gaussianity) for each t s 0 and any u
k
, v
k
h,
k
Z
2
(k = 1, 2, 3),
lim
ts
1
t s
_
,
_
3
k=1
(U
(
k
)
s,t
1)(u
k
, v
k
)
_
_
= 0. (3.1)
Assumption D: (Minimality) the set
o
0
=
_
U
s,t
(u, v) :
s = (s
1
, s
2
, , s
n
), t = (t
1
, t
2
, , t
n
); 0 s, t < ,
s
j
t
j
; u =
n
k=1
u
k
, v =
n
k=1
v
k
h, n 1
_
is total in 1.
Remark 3.2. The Assumption D is not really a restriction, one can as well work
by replacing 1 by 1
0
, the closure of the linear span of o
0
. In fact, it is easy to see
that U
s,t
leaves h1
0
invariant and that its restriction to h1
0
is an isometry.
For the unitarity of the restriction, it will be necessary to dene o
0
as the span
of U
()
s,t
(u, v)[s, t; u, v; so that the restriction of U
s,t
to h 1
0
is an isometry.
However, as can be seen in the sequel, we only use the isometry of U
s,t
in this
article.
1
It may be noted that the evolution equation here is from right to left instead of left to right
as was the case in [11], [12]. This is done in order to be in conformity with the notation of [9]
enabling us to use the results there (see Appendix) with minimal changes.
UNITARY PROCESSES WITH INDEPENDENT INCREMENTS 597
3.1. Vacuum Expectation. Let us look at the various evolutions associated
with the U
s,t
. Dene a two parameter family of operators T
s,t
on h by
u, T
s,t
v := , U
s,t
(u, v) , u, v h.
For each t s 0, since U
s,t
is unitary, T
s,t
is a contractions.
Remark 3.3. The Assumption B implies |T
s,t
1| C[t s[. In particular,
lim
ts
T
s,t
= 1 uniformly in s.
Lemma 3.4. Under the Assumptions A and B, the family T
s,t
of contrac-
tions satises
(i) for any r s t < , T
s,t
T
r,s
= T
r,t
and T
s,s
= 1
h
(ii) for any t
t s 0, |T
s,t
T
s,t
| C[t
t[.
Proof. (i) The evolution and independent increment property of U
s,t
and the
denition of T
s,t
gives the result.
(ii) By (i), for a xed s 0 and any t
t s, we have
|T
s,t
T
s,t
| = | (T
t,t
1) T
s,t
| |T
s,t
||T
t,t
1| C[t
t[.
Then we have the following result about the evolutions of the type T
s,t
by
corollary 6.2 in the Appendix:
There exists G L
loc
(R
+
, B
s
(h)) (denition is given in Appendix) such that
T
s,t
1 =
_
t
s
G()T
s,
d (3.2)
and lim
h0
T
t,t+h
I
h
= G(t) in the strong operator topology for almost every t.
We shall need the following observation (see Equation (6.2) in [11]):
k1
|(U
s,t
1) (
k
, w)|
2
= w, (1 T
s,t
) w +(1 T
s,t
) w, w (3.3)
for any w h, where
k
is an complete orthonormal basis of h.
Lemma 3.5. (i) Under the Assumption C, for any s 0 and n 3,
u, v h
n
and Z
n
2
,we have
lim
ts
1
t s
_
,
_
n
k=1
__
U
(
k
)
s,t
1
_
(u
k
, v
k
)
_
_
_
= 0, (3.4)
(ii) assume B and C.Then for u, v h, product vectors p, w h
n
and Z
2
,
Z
n
2
,we have
lim
ts
1
t s
_
(U
s,t
1)
()
(u, v),
_
U
(
)
s,t
1
_
(p, w)
_
(3.5)
= (1)
lim
ts
1
t s
_
(U
s,t
1) (u, v),
_
U
(
)
s,t
1
_
(p, w)
_
.
598 UN CIG JI, LINGARAJ SAHU, AND KALYAN B. SINHA
Proof. (i) The proof is a simple modication of the proof of Lemma 6.6 in [11].
(ii) The idea here is similar to that in the proof of Lemma 6.7 in [11]. For = 0,
it is obvious. To see this for = 1, put
=
_
U
(
)
s,t
1
_
(p, w)
and consider the following
lim
ts
1
t s
_
U
s,t
+U
s,t
2
_
(u, v),
_
(3.6)
= lim
ts
1
t s
__
U
s,t
1
_
(U
s,t
1)
(u, v),
_
= lim
ts
1
t s
k1
(U
s,t
1) (e
k
, v), (U
s,t
1) (e
k
, u) .
On the other hand, we have
1
t s
k1
(U
s,t
1) (e
k
, v), (U
s,t
1) (e
k
, u)
_
_
k1
1
t s
|(U
s,t
1) (e
k
, v)|
2
_
_
_
_
k1
1
t s
|(U
s,t
1) (e
k
, u)|
2
_
_
.
By (3.3) and (iv) in Lemma, the above quantity is equal to
2Re
_
v,
1 T
s,t
t s
v
_
1
t s
,
__
U
s,t
1
_
(U
s,t
1)
(u, u)
_
= 2Re
_
v,
1 T
s,t
t s
v
_
1
t s
,
_
2 U
s,t
U
s,t
_
(u, u)
_
.
Since by Assumption B, [
_
v,
1Ts,t
ts
v
_
[ C|v|
2
for any v h and since by the
part(i) of this lemma,
lim
ts
1
t s
,
_
2 U
s,t
U
s,t
_
(u, u)
_
= 0,
we obtain by (3.7) that lim
ts
1
ts
_
U
s,t
+U
s,t
2
_
(u, v),
_
=
lim
ts
1
t s
k1
_
(U
s,t
1) (e
k
, u), (U
s,t
1) (e
k
, v)
_
U
(
)
s,t
1
_
(p, w)
_
= 0,
which implies (3.6).
For each s 0 and for vectors u, v, p, w h the identity (3.5) gives
lim
ts
1
t s
_
(U
s,t
1)
()
(u, v), (U
s,t
1)
(
)
(p, w)
_
(3.7)
= (1)
+
lim
ts
1
t s
(U
s,t
1) (u, v), (U
s,t
1) (p, w) .
UNITARY PROCESSES WITH INDEPENDENT INCREMENTS 599
We now introduce the partial trace Tr
H
which is a linear map from B
1
(h 1)
to B
1
(h) dened by
u, Tr
H
(B)v :=
j1
u
j
, Bv
j
, u, v h
for B B
1
(h 1). In particular, Tr
H
(B) = Tr(B
2
) B
1
for B = B
1
B
2
. Then
we dene a family of operators Z
s,t
0st
on the Banach space B
1
(h) by
Z
s,t
() = Tr
H
_
U
s,t
( [ >< [) U
s,t
, B
1
(h). (3.8)
Thus, for any u, v, p, w h, we have
p, Z
s,t
([w >< v[)u := U
s,t
(u, v), U
s,t
(p, w) . (3.9)
For B
1
(h), by the denition of Z
s,t
and trace norm (see page no. 47 in [5]),
we have
|Z
s,t
()|
1
=
_
_
Tr
H
[U
s,t
( [ >< [) U
s,t
]
_
_
1
= sup
,: ons of h
k1
k
, Tr
H
_
U
s,t
( [ >< [) U
s,t
k
_
sup
,: ons of h
j,k1
k
j
, U
s,t
( [ >< [) U
s,t
k
j
_
_
_
U
s,t
( [ >< [) U
s,t
_
_
1
||
1
.
Thus Z
s,t
is contractive. For any u, v h,
|U
s,t
(u, v)|
2
= u, Z
s,t
([v >< v[)u
and positivity of Z
s,t
is clear.
Lemma 3.6. Under the Assumptions A and B, Z
s,t
is a family of positive
contractive map on B
1
(h) satisfying
(i) for any 0 r s t < , Z
s,t
Z
r,s
= Z
r,t
, Z
s,s
= 1
(ii) for any t
t s 0, |Z
s,t
Z
s,t
|
1
4C[t
t[,
(iii) For any B
1
(h), Tr(Z
s,t
) = Tr().
Proof. (i)To prove evolution property of Z
s,t
it is enough to show that
U
r,t
(u, v), U
r,t
(p, w) = p, Z
r,t
([w >< v[)u = p, Z
s,t
Z
r,s
([w >< v[)u
for any u, v, p, w h. This can be checked by using the evolution and independent
increment properties of the unitary family U
s,t
.
(ii) For any rank one operator = [w >< v[, w, v h, we have
600 UN CIG JI, LINGARAJ SAHU, AND KALYAN B. SINHA
|(Z
s,t
1)([w >< v[)|
1
= sup
{},{} ONB of h
k1
[
k
, (Z
s,t
1)([w >< v[)
k
[
= sup
,
k1
[U
s,t
(
k
, v), U
s,t
(
k
, w)
k
, v
k
, w[
sup
,
k1
[(U
s,t
1)(
k
, v), (U
s,t
1)(
k
, w)[
+ sup
,
k1
[
k
, v, (U
s,t
1)(
k
, w)[
+ sup
,
k1
[, (U
s,t
1)(
k
, v)
k
, w[
sup
,
_
_
k1
|(U
s,t
1)(
k
, v)|
2
_
_
1/2
_
_
k1
|(U
s,t
1)(
k
, w)|
2
_
_
1/2
+ sup
,
_
_
k1
[
k
, v[
2
_
_
1/2
_
_
k1
[
k
, (T
s,t
1)w[
2
_
_
1/2
+ sup
,
_
_
k1
[
k
, w[
2
_
_
1/2
_
_
k1
[
k
, (T
s,t
1)v[
2
_
_
1/2
.
Hence by identity (3.3) and Assumption B we obtain
|(Z
s,t
1)([w >< v[)|
1
2|(T
s,t
1)||w||v| +|(T
s,t
1)w||v| +|(T
s,t
1)v||w|
4C[t s[|w| |v|.
Now any for =
k
k
[
k
><
k
[ B
1
(h), where
k
and
k
are two
orthonormal bases of h and we have
|Z
s,t
() |
1
4C
_
k
[
k
[
_
[t s[ 4C||
1
[t s[
and hence
|Z
s,t
1| 4C[t s[. (3.10)
By evolution property and contractivity of Z
s,t
|Z
s,t
Z
s,t
| = | (Z
t,t
1) Z
s,t
| |Z
s,t
||Z
t,t
1| 4C[t
t[.
(iii) It can be proved as in lemma 6.5 in [11]
The Corollary 6.2 in the Appendix leads to following result for the evolution
Z
s,t
.: Under the Assumptions A and B there exists / L
loc
(R
+
, B
s
(B
1
(h)))
UNITARY PROCESSES WITH INDEPENDENT INCREMENTS 601
(see Appendix for denition) such that
Z
s,t
1 =
_
t
s
/()Z
s,
d, lim
h0
Z
t,t+h
I
h
= /(t). (3.11)
4. Construction of Noise Space
Consider the algebra M generated by the tuples (u, v, ) with multiplication
structure given by (u, v, ) (p, w,
) = (u w, v z,
). For each s 0 we
dene a scalar valued map K
s
on M M by setting, for (u, v, ), (p, w,
) M,
K
s
((u, v, ), (w, z,
)) := lim
ts
1
t s
__
U
()
s,t
1
_
(u, v),
_
U
s,t
1
_
(p, w)
_
if the limit exists.
Theorem 4.1. For almost every s 0
(i) the map K
s
is a positive denite kernel on M,
(ii) there exists a unique (up to unitary equivalence) separable Hilbert space
k
s
, an embedding
s
: M k
s
such that
s
(u, v, ) : (u, v, ) M is total in k
s
, (4.1)
s
(u, v, ),
s
(p, w,
) = K
s
((u, v, ), (p, w,
)) , (4.2)
(iii) for any (u, v, ) M, u =
n
i=1
u
i
, v =
n
i=1
v
i
and = (
1
, ,
n
)
s
(u, v, ) =
n
i=1
k=i
u
k
, v
k
s
(u
i
, v
i
,
i
), (4.3)
(iv)
s
(u, v, 1) =
s
(u, v, 0) for any u, v h,
(v) for xed u, v, p, w h, the map s K
s
((u, v), (p, w)) =
s
(u, v),
s
(p, w)
is Lebesgue measurable and locally bounded in R
+
.
Proof. (i) The proof is exactly same as the proof of Lemma 7.1 in [11]. By Lemma
3.5, for elements (u, v, ), (p, w,
) M, Z
m
2
and
Z
n
2
, we have
K
s
((u, v, ), (p, w,
)) (4.4)
= lim
ts
1
t s
__
U
()
s,t
1
_
(u, v),
_
U
(
)
s,t
1
_
(p, w)
_
=
1im, 1jn
k=i
u
k
, v
k
l=j
p
l
, w
l
lim
ts
1
t s
_
(U
s,t
1)
(i)
(u
i
, v
i
), (U
s,t
1)
(
j
)
(p
j
, w
j
)
_
.
Since
(U
s,t
1) (u, v), (U
s,t
1) (p, w)
= U
s,t
(u, v), U
s,t
(p, w) u, vp, w
u, v , (U
s,t
1) (p, w) , (U
s,t
1) (u, v)p, w
= p, (Z
s,t
1) ([w >< v[)u u, v p, (T
s,t
1) w u, (T
s,t
1) vp, w,
602 UN CIG JI, LINGARAJ SAHU, AND KALYAN B. SINHA
the existence of the limits on the right hand side of (4.4) follows from the identity
(3.5) and by the equations (3.2) and (3.11), K
s
is given as
K
s
((u, v, ), (p, w,
)) (4.5)
= (1)
+
lim
ts
__
p,
Z
s,t
1
t s
([w >< v[)u
_
u, v
_
p,
T
s,t
1
t s
w
__
(1)
+
lim
ts
_
u,
T
s,t
1
t s
v
_
p, w
= (1)
+
_
p, /(s)([w >< v[)u u, v p, G(s)w u, G(s)vp, w
_
.
(ii) For each s 0, the Kolmogorovs construction [10] to the pair (M, K
s
)
provides a Hilbert space k
s
as the closure of the span of
s
(u, v, ) : (u, v, ) M.
(iii) Again as in [11], for any (p, w,
s
(u, v, ),
s
(p, w,
) = K
s
((u, v, ), (p, w,
))
=
n
i=1
k=i
u
k
, v
k
s
(u
i
, v
i
,
i
),
s
(p, w,
) .
Since
s
(p, w,
) : (p, w,
) M is a total subset of k
s
, (4.3) follows.
(iv) By (3.5), we have
s
(u, v, 1),
s
(p, w,
) =
s
(u, v, 0),
s
(p, w,
)
and hence
s
(u, v, 1) =
s
(u, v, 0).
By parts (iii) and (iv) of this theorem, it is clear that k
s
is spanned by the
family
s
(u, v) : u, v h, where we have written
t
(u, v) for
t
(u, v, 0).
Since G L
loc
(R
+
, B
s
(h)) and / L
loc
(R
+
, B
s
(B
1
(h))) it follows from (4.5)
that
s
(., .) : h h k
s
is sesquilinear and continuous and thus separability of
k
s
follows from that of h.
(v) This follows similarly as for (iv).
For any two orthonormal bases
k
,
k
of h, the collection of vectors
s
(
k
,
l
) : k, l 1
is a countable total family in k
s
and
s
s
(u, v),
s
(p, w) = K
s
((u, v), (p, w))
is a Lebesgue measurable function. Thus s
s
(u, v),
s
(
k
,
l
) is measurable
and therefore the family k
s
: s 0 spanned by
s
(u, v) : s 0, u, v h, is a
measurable eld of Hilbert spaces (Chapter 8, [3]).
For any T 0, dene K
T
((u, v), (p, w)) =
_
T
0
K
s
((u, v), (p, w))ds
=
_
T
0
p, /(s)([w >< v[)u u, v p, G(s)w u, G(s)vp, wds.
Since each K
s
is positive denite it can be seen that K
T
is a positive denite kernel.
Let the associated Hilbert space k
T
. There exists a family of vectors
T
(u, v) which
UNITARY PROCESSES WITH INDEPENDENT INCREMENTS 603
spans the Hilbert space k
T
such that
T
(u, v),
T
(p, w) = K
T
((u, v), (p, w))
=
_
T
0
K
s
((u, v), (p, w))ds =
_
T
0
s
(u, v),
s
(p, w)ds.
Comparing the two expressions for K
T
, it follows that
t
(u, v),
t
(p, w) = p, /(t)([w >< v[)[G(t)w >< v[[w >< G(t)v[u. (4.6)
In k
T
there exists a bounded self adjoint operator A with absolutely continuous
simple spectrum such that A
T
(u, v)(s) = s
s
(u, v) for almost every s [0, T] and
k
T
is the direct integral
_
[0,T]
k
s
ds ( [3]). There is natural isometric embedding of
k
T
in k
T
for T T
by setting
T,T
s
(u, v) =
T
s
(u, v) for all 0 s T and 0 for
s (T, T
].
Remark 4.2. The integral
_
R+
K
s
((u, v), (u, v))ds =
_
R+
|
s
(u, v)|
2
ds need not
exist and therefore
_
R+
k
s
ds may not be dened.
Lemma 4.3. Under the hypothesis of Theorem 4.1, we have the following:
(i) There exists a unique strong measurable family of bounded operators L(t) :
h h k
t
such that
|L(t)v|
2
= 2Re v, G(t)v , v h.
(ii) The map t L(t) is locally norm bounded.
Proof. (i) By the identity (4.5), for any u, v h, we have for almost every t 0
|
t
(u, v)|
2
= u, /(t)([v >< v[)u u, v u, G(t)v u, G(t)vu, v.
and thus
k
|e
k
t
(e
k
, v)|
2
=
k
|
t
(e
k
, v)|
2
=
k
_
e
k
, /(t)([v >< v[)e
k
e
k
, v e
k
, G(t)v e
k
, G(t)v e
k
, v
_
= Tr (/(t)([v >< v[)) v, G(t)v v, G(t)v.
Moreover, since Z
s,t
is trace preserving it follows that Tr (/(t)([v >< v[)) = 0.
Therefore
k
|e
k
t
(e
k
, v)|
2
= 2Re v, G(t)v . This implies that
k
e
k
t
(e
k
, v) is convergent in norm and in fact for almost every t it denes a bounded
operator L(t) : h h k
t
given by L(t)v =
k
e
k
t
(e
k
, v) with
|L(t)v|
2
= 2Re v, G(t)v . (4.7)
The strong measurability of t L(t) follows from the denition.
The part (ii) follows from the local norm boundedness of G(.).
604 UN CIG JI, LINGARAJ SAHU, AND KALYAN B. SINHA
5. Hudson-Parthasarathy (HP) Evolution Systems and Equivalence
5.1. HP Evolution Systems. In order to simplify the discussion of the existence
and uniqueness of the solution of HP type quantum stochastic dierential equation
in
sym
(
_
R+
k
s
ds) and to be able to refer to the existing literature, it is convenient
to introduce the following point of view which allow us to embed the process in
the standard Fock space =
sym
(L
2
(R
+
, k)) where k = l
2
(N).
Note that for almost every t 0, k
t
is a complex separable Hilbert space.
Setting d(t) = the dimension of k
t
, d : R
+
N is measurable and dening
n
= t : d(t) = n, R
+
can be written as disjoint union
n=1
n
of measurable
sets. Let us consider the Hilbert space l
2
(N) with a xed orthonormal basis E
j
:
j 0. Now for t
n
, n < we embed k
t
as the n dimensional subspace
SpanE
j
: 1 j n of k and for t
, k
t
identied with k. Then the direct
integral
_
R+
k
t
dt =
n1
L
2
(
n
, C
n
)
L
2
(
, k). If
= , then
_
R+
k
t
dt is
isometrically embedded in L
2
(R
+
, k).
For any subset D L
2
(R
+
, k), let c(D) be the subspace of which is spanned
by the set e(f) : f D of exponential vectors dened as:
e(f) :=
n0
f
n
n!
.
For 0 s < t < and f / = L
2
(R
+
, k), we denote the functions 1
[0,s]
f,
1
(s,t]
f and 1
[t,)
f by f
s]
, f
(s,t]
and f
[t
, where 1
A
is the indicator function of
A [0, ). Then the Hilbert spaces / and can be decomposed as / = /
s]
/
[s,t)
/
[t
and =
s]
[s,t)
[t
via the unitary isomorphism given by:
e(f) e(f
s]
) e(f
(s,t]
) e(f
[t
)
s]
[s,t)
[t
,
where /
s]
= L
2
([0, s), k), /
[s,t)
= L
2
([s, t), k), /
[t
= L
2
([t, ), k) and
s]
=
(/
s]
),
[s,t)
= (/
[s,t)
),
[t
= (/
[t
).
Let us consider the Hudson-Parthasarathy (HP) type equation on h :
V
s,t
= 1
h
+
,0
_
t
s
L
()V
s,
(d). (5.1)
Here the coecients L
(t) =
_
_
t1
h
for (, ) = (0, 0),
a
_
1
[0,t]
E
j
(t)
_
for (, ) = (j, 0),
a
_
1
[0,t]
E
k
(t)
_
for (, ) = (0, k),
_
1
[0,t]
[E
k
(t) >< E
j
(t)[
_
for (, ) = (j, k),
(5.2)
where E
j
(t) = E
j
for j 1, 2, d(t) and E
j
(t) = 0 otherwise. With respect to
the orthonormal basis E
j
(t) we have bounded operators L
j
(t) : t 0, j 1 in
h such that
u, L
j
(t)v = E
j
,
t
(u, v) = u E
j
, L(t)v, u, v h. (5.3)
For the details about quantum stochastic calculus see [10, 6]).
Now, let us state the main result of this article.
UNITARY PROCESSES WITH INDEPENDENT INCREMENTS 605
Theorem 5.1. (i) The HP equation
V
s,t
= 1
h
+
,0
_
t
s
L
()V
s,
(d) (5.4)
on h
sym
(/) with coecients L
(t) given by
L
(t) =
_
_
G(t) for (, ) = (0, 0),
L
j
(t) for (, ) = (j, 0),
L
k
(t)
V
s,t
, 0 s t < . (5.6)
Proof. (i) The existence of the strong solution V
s,t
of the equation (5.4) follows
exactly as in Proposition 27.5 of ([10]) since for any h , we have
j=1
|(L
j
() I)|
2
=
i
|(L
j
()v
i
) E
i
|
2
=
j
|L
j
()v
i
|
2
=
i
|L()v
i
|
2
sup
0T
(|L()|
2
)
i
|v
i
|
2
= sup
0T
(|L()|
2
)||
2
,
where we have written =
v
i
E
i
,E
i
an ONB in .
The isometry of V
s,t
follows easily as in the proof of the theorem 27.8 of ([10]).
On the other hand for the proof of co-isometry of V
s,t
we proceed as in Theorem
5.3.3 of ([6]) and for f, g L
2
(R
+
, k) dene Y
g,f
(t) : B(h) B(h) by
Y
g,f
(t)X = [
j
g
j
(t)L
j
(t), X][
j
f
j
(t)L
j
(t)
, X]+
j
L
j
(t)
XL
j
(t)+XG(t)+
G(t)
X, so that if we setX
g,f
(s, t) = . e(g), (V
s,t
V
s,t
). e(f), then X
g,f
(s, .)
satisfy the equation
X
g,f
(s, t) = e(g), e(f)I
h
+
_
t
s
Y
g,f
()X
g,f
(s, )d. (5.7)
By the equation (4.7) ,we note that e(g), e(f)I
h
is a solution of the linear equation
(5.7) and hence by the uniqueness of the solution of the B(h)-valued initial value
problem we have that X
g,f
(s, t) = e(g), e(f)I
h
or V
s,t
is a co-isometry, leading
to the unitarity of the same. We postpone the proof of part (ii) to the next two
subsections.
For = (
1
,
2
, ,
n
) Z
n
2
, we dene V
()
s,t
B(h
n
) by setting V
()
s,t
B(h ) by
V
()
s,t
=
_
V
s,t
for = 0,
V
s,t
for = 1.
The next result veries the properties of Assumption A for the family V
s,t
with = e(0) .
606 UN CIG JI, LINGARAJ SAHU, AND KALYAN B. SINHA
Lemma 5.2. The unitary solution V
s,t
of HP equation (5.4) satises
(i) for any 0 r s t < , V
r,t
= V
s,t
V
r,s
,
(ii) Assumption A holds for the family V
s,t
with the distinguished vector
e(0) in ,
(iii) for any 0 s t < ,
e(0), V
s,t
(u, v)e(0) = u, T
s,t
v , u, v h.
Proof. (i) For xed 0 r s t < , we set W
r,t
= V
s,t
V
r,s
. Then by (5.4), we
have
W
r,t
= V
r,s
+
,0
_
t
s
L
()V
s,
V
r,s
(d)
= W
r,s
+
,0
_
t
s
L
()W
r,
(d),
since W
r,s
= V
r,s
V
s,s
= V
r,s
. Thus the family W
r,t
of unitary operators also
satises the HP equation (5.4) for V
r,t
. Hence by the uniqueness of the solution of
this quantum stochastic dierential equation, W
r,t
= V
r,t
for any 0 r s t <
, and the result follows.
(ii) For any 0 s t < , the solution V
s,t
B(h
[s,t]
). Therefore, for
p, w h, V
s,t
(p, w) B(
[s,t]
) and the Assumptions A2(i) and A2(ii) are veried
by the property of the continuous tensor-factorization of the Fock space.
(iii) Let us dene
_
u,
T
s,t
v
_
:= e(0), V
s,t
(u, v)e(0) , u, v h.
Then
T
s,t
is a contractive family of operators and by (5.4), we have that
T
s,t
= 1 +
_
t
s
G()
T
s,
d. (5.8)
Thus
T
s,t
T
s,t
satises the dierential equation
T
s,t
T
s,t
=
_
t
s
G()(
T
s,
T
s,
)d.
Since G() is locally norm bounded, an iteration of this equation will lead to
T
s,t
= T
s,t
for almost all s, t and therefore by continuity also for all s, t.
Consider the family of operators
Z
s,t
dened by
Z
s,t
() = Tr
_
V
s,t
( [e(0) >< e(0)[)V
s,t
, B
1
(h).
As for Z
s,t
, it can be seen that
Z
s,t
is a contractive family of maps on B
1
(h) and,
in particular, for any u, v, p, w h,
_
p,
Z
s,t
([w >< v[)u
_
= V
s,t
(u, v)e(0), V
s,t
(p, w)e(0) .
Lemma 5.3. The family
Z
s,t
is a uniformly continuous evolution of contraction
on B
1
(h) and
Z
s,t
= Z
s,t
, where Z
s,t
is given as in (3.8).
UNITARY PROCESSES WITH INDEPENDENT INCREMENTS 607
Proof. By (5.4) and Itos formula, for u, v, p, w h
_
p,
_
Z
s,t
1
_
([w >< v[)u
_
= V
s,t
(u, v)e(0), V
s,t
(p, w)e(0) u, v p, w
=
_
t
s
V
s,
(u, v)e(0), V
s,
(G()
p, w)e(0) d
+
_
t
s
V
s,
(G()
u, v)e(0), V
s,
(p, w)e(0) d
+
j
_
t
s
V
s,
(L
j
()
u, v)e(0), V
s,
(L
j
()
p, w)e(0) d
=
_
t
s
_
p, G()
Z
s,
([w >< v[) +
Z
s,
([w >< v[)G()
j1
L
j
()
Z
s,
([w >< v[L
j
()
u
_
d.
Thus by identity (5.3) for L
j
(t) and (4.6), we have that
_
p,
_
Z
s,t
1
_
()u
_
=
_
t
s
_
p, /()
Z
s,
()u
_
d, (5.9)
where = [w >< v[. Thus the family
Z
s,t
satises the equation
Z
s,t
() = +
_
t
s
/()
Z
s,
()d, B
1
(h).
Therefore, proceeding as in the proof of Lemma 5.2 (iii) we can conclude that
Z
s,t
= Z
s,t
.
5.2. Minimality of HP Evolution Systems. In this section we shall show the
minimality of the HP evolution system V
s,t
discussed in Section 5.1 which will
be needed to prove (ii) in Theorem 5.1, i.e., to establish unitary equivalence of
U
s,t
and V
s,t
. We shall prove here that the subset
o
=
_
V
s,t
(u, v)e(0) :
s = (s
1
, s
2
, , s
n
), t = (t
1
, t
2
, , t
n
); 0 s, t < ,
s
j
t
j
; u =
n
i=1
u
i
, v =
n
i=1
v
i
h, n 1
_
is total in the symmetric Fock space (/) (L
2
(R
+
, k)), where
V
s,t
(u, v)e(0) := V
s1,t1
(u
1
, v
1
) V
sn,tn
(u
n
, v
n
)e(0).
Let T 0 be xed and as in ([11]), we note that for any 0 s < t T, u, v h,
1
t s
[V
s,t
1] (u, v)e(0) = (s, t, u, v) +(s, t, u, v) +(s, t, u, v) +(s, t, u, v),
(5.10)
608 UN CIG JI, LINGARAJ SAHU, AND KALYAN B. SINHA
where these vectors in the Fock space are given by
(s, t, u, v) :=
1
t s
j1
_
t
s
u, L
j
()v a
j
(d) e(0),
(s, t, u, v) :=
1
t s
_
t
s
u, G()v d e(0),
(s, t, u, v) :=
1
t s
j1
_
t
s
(V
s,
1) (L
j
()
u, v) a
j
(d) e(0),
(s, t, u, v) :=
1
t s
_
t
s
(V
s,
1) (G()
u, v) de(0).
Note that any can be written as =
(0)
(1)
, where
(n)
is in the
n-fold symmetric tensor product L
2
(R
+
, k)
sn
L
2
(
n
) k
n
. Here
n
is the
n-simplex t = (t
1
, t
2
, , t
n
) : 0 t
1
< t
2
< t
n
< .
Lemma 5.4. Let u, v h and let C
T
= 4e
T
sup|L()|
2
+|G()|
2
: 0 t.
Then for any 0 s t T,
(i)
|(V
s,t
1)ve(0)|
2
C
T
[t s[|v|
2
. (5.11)
(ii) |(V
s,t
1)(u, v) e(0)|
2
C
T
|u|
2
|v|
2
[t s[.
(iii) For any u h
|
j1
_
t
s
V
s,
(u, L
j
()v)a
j
(d)e(0)|
2
C
T
|u|
2
|v|
2
[t s[.
Proof. (i) By estimates of quantum stochastic integration (Proposition 27.1, [10])
|(V
s,t
1)ve(0)|
2
= |
j1
_
t
s
V
s,
L
j
()a
j
(d) ve(0) +
_
t
s
V
s,
G()d ve(0)|
2
2e
T
_
t
s
j1
|L
j
()v|
2
+|G()v|
2
d
2e
T
|v|
2
_
t
s
|L()|
2
+|G()|
2
d
= |v|
2
C
T
[t s[.
(ii) By Lemma 2.1 (ii), we have (V
s,t
1)(u, v)e(0) = F
u
(V
s,t
1)ve(0) and
therefore the result follows from (i).
UNITARY PROCESSES WITH INDEPENDENT INCREMENTS 609
(iii) By lemma 2.1,
|
j1
_
t
s
V
s,
(L
j
()
u, v)a
+
j
(d)e(0)|
2
= |
j
_
t
s
F
u
(L
j
() I
)V
s,
a
+
j
(d)ve(0)|
2
2e
T
|u|
2
|v|
2
sup
0T
|L()|
2
[t s[
C
T
|u|
2
|v|
2
[t s[,
where we have used the standard estimate of a quantum stochastic integral.
Lemma 5.5. For any u, v h, 0 s t T,
(i) sup|(s, t, u, v)|
2
: 0 s t T C
2
T
|u|
2
|v|
2
and |(s, t, u, v)|
C
T
[t s[
1/2
|u||v|.
(ii) For any (L
2
(R
+
, k)), lim
ts
, (s, t, u, v) = 0 and
lim
ts
, (s, t, u, v) =
j1
u, L
j
(s)v
(1)
j
(s) =
(1)
(s),
s
(u, v), a.e. s 0.
Proof. (i) By Lemma 5.4, part (iii), we have
|(s, t, u, v)|
2
=
1
[t s[
2
|
j1
_
t
s
(V
s,
1)(L
j
()
u, v)a
j
(d) e(0)|
2
= [t s[
2
|
j
_
t
s
F
u
(L
j
() I
)(V
s,
I)a
+
j
(d)ve(0)|
2
2e
T
|u|
2
[t s[
2
sup
|L()|
2
_
t
s
|(V
s,
I)ve(0)|
2
d
C
2
T
|u|
2
|v|
2
,
where we have used the estimate (5.10). Similarly,
|(s, t, u, v)| =
1
[t s[
|
_
t
s
(V
s,
1)(G()
u, v)d e(0)|
= [t s[
1
|
_
t
s
F
u
(G() I
)(V
s,
I)ve(0)d|
|u|[t s[
1
sup
(|G()|)
_
t
s
|(V
s,
I)ve(0)|d
C
T
[t s[
1/2
|u||v|.
(ii) For any f L
2
(R
+
, k),
610 UN CIG JI, LINGARAJ SAHU, AND KALYAN B. SINHA
e(f), (s, t, u, v) = e(f),
1
t s
j1
_
t
s
(V
s,
1)(L
j
()
u, v)a
j
(d) e(0)
=
1
t s
j1
_
t
s
f
j
()e(f), (V
s,
1)(L
j
()
u, v) e(0)d
=
1
t s
_
t
s
R(s, )d,
where R(s, ) =
j1
f
j
()e(f), (V
s,
1)(L
j
()
j1
_
t
s
u, L
j
()v
(1)
j
()d. (5.12)
Since
[
j1
u, L
j
()v
(1)
j
()[
2
|u|
2
j1
|L
j
()v|
2
[
(1)
j
()[
2
C
|v|
2
|
(1)
()|
2
,
the function
j1
u, L
j
()v
(1)
j
() is in L
2
and hence locally integrable. Thus
we get
lim
ts
, (s, t, u, v) =
j1
u, L
j
(s)v
(1)
j
(s) =
(1)
(s),
s
(u, v) a.e. s 0.
n
k=1
M(s
k
, t
k
, u
k
, v
k
) e(0) = 0, where
M(s
k
, t
k
, u
k
, v
k
) =
(V
s
k
,t
k
1)
t
k
s
k
(u
k
, v
k
) (s
k
, t
k
, u
k
, v
k
) (s
k
, t
k
, u
k
, v
k
)
and lim
ts
means t
k
s
k
for each k.
(ii) lim
ts
,
n
k=1
(s
k
, t
k
, u
k
, v
k
) =
(n)
(s
1
, s
2
, , s
n
),
s1
(u
1
, v
1
)
sn
(u
n
, v
n
).
Proof. (i) First note that M(s, t, u, v)e(0) = (s, t, u, v) + (s, t, u, v). So by the
above observations in Lemma 5.5, M(s, t, u, v)e(0) is uniformly bounded in
UNITARY PROCESSES WITH INDEPENDENT INCREMENTS 611
s, t and lim
ts
e(f), M(s, t, u, v)e(0) = 0, f L
2
(R
+
, k). Since the intervals
[s
k
, t
k
)s are disjoint for dierent ks,
e(f),
n
k=1
M(s
k
, t
k
, u
k
, v
k
) e(0) =
n
k=1
e(f
[s
k
,t
k
)
), M(s
k
, t
k
, u
k
, v
k
) e(0)
and thus lim
ts
e(f),
n
k=1
M(s
k
, t
k
, u
k
, v
k
) e(0) = 0.
Since
n
k=1
M(s
k
, t
k
, u
k
, v
k
) e(0) is uniformly bounded in s
k
, t
k
requirement
follows for .
(ii) It can be proved similarly as part (iii) of the previous Lemma.
Lemma 5.7. Let be such that
, = 0, o
. (5.13)
Then we have
(i)
(0)
= 0 and
(1)
= 0,
(ii) for any n 0,
(n)
= 0,
(iii) the set o
j1
u, L
j
(s)v
(1)
j
(s)
=
(1)
(s),
s
(u, v).
So
(1)
(s),
s
(u, v)
_
= 0 for any u, v h for almost every s. Since
s
(u, v) :
u, v h is total in k
s
, it follows that
(1)
(s) = 0 k
s
for almost every 0 s ,
i.e,
(1)
= 0.
(iii) We prove this by induction. The result is already proved for n = 0, 1. For
n 2, assume as induction hypothesis that for all m n 1,
(m)
(s) = 0, for
almost every s
m
(s
i
for i = 1, 2, , m). To show that
(n)
= 0, we note
that by a similar argument as in [11],
_
(n)
(s
1
, s
2
, , s
n
),
s1
(u
1
, v
1
)
sn
(u
n
, v
n
)
_
= 0.
for almost every s
n
(s
i
). Since
s
(u, v) : u, v h is total in k
s
, it follows
that
(n)
(s
1
, s
2
, , s
n
) = 0 k
s1
k
sn
for almost every (s
1
, s
2
, , s
n
)
n
.
612 UN CIG JI, LINGARAJ SAHU, AND KALYAN B. SINHA
5.3. Unitary Equivalence. We shall now prove (ii) in Theorem 5.1 that the
unitary evolution U
s,t
on h 1 is unitarily equivalent to the unitary solution
V
s,t
of HP equation (5.4). To prove this we need the following two results.
Lemma 5.8. Let U
s,t
(u, v) and U
s
,t
(p, w) be in o, where v, z h
n
. Then
there exist an integer m 1, a = (a
1
, a
2
, , a
m
), b = (b
1
, b
2
, , b
m
) with
0 a
1
b
1
a
m
b
m
< , partition R
1
R
2
R
3
= 1, , m with
[R
i
[ = m
i
, family of vectors x
k
l
, g
ki
h and y
k
l
, h
ki
h for l R
1
R
2
and
i R
2
R
3
such that
U
s,t
(u, v) =
lR1R2
U
a
l
,b
l
(x
k
l
, y
k
l
),
U
s
,t
(p, w) =
lR2R3
U
a
l
,b
l
(g
k
l
, h
k
l
).
Proof. It follows from the evolution hypothesis of U
s,t
that for r [s, t] and a
complete orthonormal basis f
j
h we can write
U
s,t
(u, v) =
j1
U
s,r
(u, f
j
)U
r,t
(f
j
, v).
,t
(p, w) o, we have
_
U
s,t
(u, v), U
s
,t
(p, w)
_
=
_
V
s,t
(u, v)e(0), V
s
,t
(p, w)e(0)
_
. (5.14)
Proof. The proof of (5.14) is very similar to that in [11]. In fact, for
0 s t < , U
s,t
(u, v), U
s,t
(p, w) = p, Z
s,t
([w >< v[)u
while
V
s,t
(u, v)e(0), V
s,t
(p, w)e(0) =
_
p,
Z
s,t
([w >< v[)u
_
but
Z
s,t
= Z
s,t
.
Now dening a map : 1 by sending U
s,t
(u, v) o to V
s,t
(u, v)e(0) o
,
as in [11], we can establish unitary equivalence of HP evolution V
s,t
with the
evolution U
s,t
we started with.
6. Appendix
Let X be a complex separable Banach space with the Radon Nikodym
property, i.e., every f Lip(R, X) f : R X[|f(t)f(s)| C[ts[ for some
0 < C < is dierentiable almost everywhere. In such a case, f
loc
(R, X)
and
f(t) f(s) =
_
t
s
f
()d. (6.1)
UNITARY PROCESSES WITH INDEPENDENT INCREMENTS 613
It is known [1] that separable reexive Banach spaces and separable dual Banach
spaces have the Radon-Nikodym property. Thus the cases relevant to our problem
in which X = h and X = B
1
(h) qualify as spaces with Radon-Nikodym property.
We shall denote by B
s
(X) the linear space B(X) equipped with strong operator
topology.
Let
S
s,t
[s, t R, s t be a contractive evolution acting on a complex sepa-
rable Banach space X, i.e., |
S
s,t
| 1 and
S
r,t
=
S
s,t
S
r,s
,
S
s,s
= 1 for r s t.
Then we have the following theorem [9] characterizing such evolution.
Theorem 6.1. Let the Banach space X have the Radon-Nikodym property and let
the evolution
S
s,t
satisfy uniform Lipshitz condition: |
S
s,t
1| C[ts[ for s, t
R and s t. Then there exists an operator valued function
G L
loc
(R, B
s
(X))
such that
S
s,t
= 1 +
_
t
s
g()
S
s,
d.
This theorem is proven in [9]. We need to adapt this for the evolutions (viz.,
T
s,t
and Z
s,t
) that we have constructed earlier where s, t R
+
.
Given a contractive evolution S
s,t
on R
+
, we can extend it to dene a contractive
evolution
S
s,t
on R as follows:
S
s,t
=
_
_
_
S
s,t
if 0 s t
1 if s t 0
S
0,t
if s 0 t.
It is easy to check that this
S
s,t
is a contractive evolution on R. Furthermore, it
is clear that
S
s,t
satises Lipshitz condition on R if S
s,t
does the same on R
+
.
Corollary 6.2. Let X be either h or B
1
(h) and let T
s,t
and Z
s,t
be contractive
evolutions on R
+
respectively. Then there exist operator valued functions G
L
loc
(R
+
, B
s
(h)) and / L
loc
(R
+
, B
s
(B
1
(h))) respectively such that
T
s,t
= 1 +
_
t
s
G()T
s,
d
and
Z
s,t
= 1 +
_
t
s
/()Z
s,
d.
Acknowledgment. The authors thank the referee for a careful reading of the
manuscript and making numerous suggestions for the improvement of the article,
in particular, for bringing the attention to the reference [9].
References
1. Arendt, W., Batty, C. J. K., Hieber, M., and Neubrander, F.: Vector-valued Laplace Trans-
forms and Cauchy problems, Monographs in Mathematics, 96, Birkhauser Verlag,Basel,2001.
2. Bhat, B. V. R. and Sinha, K. B.: A stochastic dierential equation with time-dependent and
unbounded operator coecients; J. Funct. Anal. 114 (1993) 1231.
3. Dixmier, J. : C
dA
+
= Idt, dA
dN = dA
but A
+
= A
,
they determine three linear-independent self-adjoint combinations
B
0
= N, B
1
=
1
2
_
A
+ A
+
_
, B
2
=
i
2
_
A
A
+
_
(1.2)
with a noncommutative multiplication table induced from (1.1). Each operator-
valued function B
i
(t) B
t
i
represents on the vacuum state vector
T a
real classical Levy process B
t
i
=
t
i
with zero expectation given by a Gelfand
QUANTUM DYNAMICS AND GIRSANOV TRANSFORMATION 617
transformation
t
i
() = (
t
i
) of Fock vector
t
i
= B
t
i
P
i
(g) = 1
[ V
i
(g) 1
[V
i
(g)] i = 0, 1, 2
on V
i
(g) = exp
_
i
_
g (t) dB
t
i
is such that B
t
1
and B
t
2
are classical Brownian mo-
tions identical in distribution, P
1
= P
2
, but B
t
0
is even nonstochastic martingale
identical in distribution P
0
=
0
to zero. However, due to mutual noncommut-
ativity [B
i
, B
k
] ,= 0 they do not have joint probability law
P(g
) coinciding with
E
[V(g
)] for
V(g
) = exp
_
i
i
_
g
i
(t) dB
t
i
_
.
Therefore, these classical martingales cannot be simultaneously represented as a
vector-valued stochastic process
t B
t
() =
_
B
t
0
, B
t
1
, B
t
2
_
() =
t
()
in any classical Kolmogorovian probability space (, F, P). If they could, they
would have a commutative multiplication table, but due to the noncommutativity
they do not have even joint spectrum which is required for the Kolmogorovian
functional representation. They are quantum Levy noises as the quantum mar-
tingales with independent increments with respect to the conditional expectations
E
t
: A A
t
on an operator algebra A = L() of multiple quantum stochastic in-
tegrals generated by V(g
t>0
T
t
of
the vacuum state E[V] = 1
[ V1
A and the
functional of mathematical expectation E : A C dened as the scalar product
1
.
1.2. HP Ito algebra and matrix notations. Now we introduce the convenient
tensor notation A
= B
i
indexing by pairs i = (, ) of two-valued indices
, + J
+
and , J
= B
and A
= B
0
such that
A
+
(t) = tI, A
= A
, A
= N, A
+
= A
+
. (1.3)
Identifying + (+, ), 0 (, ), (, ) and (+, ), the Einstein sum-
mation convention holds for B
i
B
iJ+J
K
i
B
i
=
J+,,J
K
.
Note that the rule of identication K
,
K
is contravariant to B
,
B
.
This allows to introduce also the covariant notation B
i
= B
in terms of the
transposition (, ) = (, ) mapping J
= J
+
J
onto J
= J
J
+
, with the
inverse mapping j i of j = (, ) also denoted by j = i.
618 V. P. BELAVKIN
In these notation the pseudo-Hermiticity
(B
, B
, B
0
, B
+
)
= (B
, B
, B
0
, B
+
) =
_
B
, B
+
, B
0
, B
_
of the canonical integrators B
j
= A
is simply written as B
j
= B
j
in terms of
B
j
:= B
j
with respect to the involution j
J
+
of and such that B
= B
.
The convenience of this matrix notation allows to express the HP multiplication
table (1.1) for the canonical basis B
= A
dA
dA
as
suggested in [3][14]. However, it is a particular case of the universal noncommut-
ative Ito multiplication table dB
j
dB
k
=
m
j,k
dB
m
dening in a pseudo-Hermitian
basis B
i
= B
i
= B
i
the associative QS Ito covariations
dX
t
dY
t
= K
i
i,l
L
l
dt +
n=
K
i
i,,l
n
L
l
dB
n
t
= K
i
i,l
n
L
l
dB
n
t
(1.4)
in terms of the structural constants
i,l
n
=
n
l,i
for the product of quantum-
stochastic dierentials dX
t
= K
i
dB
i
t
and dY
t
= L
l
dB
l
t
. Here
j
,l
k
= (
l
,j
k
)
are
-Hermitian structural coecients which are not necessarily symmetric,
i,l
n
,=
l,i
n
if dB
i
dB
l
,= dB
l
dB
i
, but satisfying the associativity condition corresponding to
(dX
t
dY
t
) dZ
t
= dX
t
(dY
t
dZ
t
) .
Moreover, these complex coecients should satisfy a positivity condition corres-
ponding to the semi-positivity dXdX
0, where dX
t
= K
i
dB
i
t
is dened by
K
i
= K
i
, such that the covariance matrix = [
,k
] with
,k
dt = E
_
dB
j
dB
k
j
= B
j
.
The HP table (1.1) in the canonical basis (1.3) corresponds to the case
i,l
n
:=
m
j,k
j = (, ) , k = (, ) , m = (, )
of matrix units multiplication table e
j
e
k
=
in terms of 3 3-matrices
e
=
_
e
,
,
=,,+
=,,+
, e
,
,
=
and e
k
e
e
,
,
. The set of these units indexed by J
, , 0, + forms the
basis for the HP Ito -algebra canonically represented by the triangular matrices
K = [K
] = K
= K
i
i,l
n
L
l
= K
K
i
and L
L
l
. Thus, the quantum Ito product
(1.4) is represented in terms of the matrix product L
=
L
= (L
, L
+
, L
0
, L
) of L
=
(L
, L
, L
0
, L
+
) by matrix elements
_
L
= L
i
=
L
i
for i = (, ) J
+
J
i
=
L
of
L
:=
L
j
L
j
.
QUANTUM DYNAMICS AND GIRSANOV TRANSFORMATION 619
1.3. Matrix representation of classical and Ito algebras. The deterministic
Newton-Leibniz dierential calculus based on the nilpotent multiplication rule
(dt)
2
= 0 can be easily extended from smooth to continuous trajectories x(t)
having right, say C-valued, derivatives (t). The dierentials dx = x(t + dt)
x(t) = dt form obviously one-dimensional -algebra with the trivial products
= 0 representing [dx[
2
= 0. This formal algebra of dierential calculus was
generalized by Ito [23] to non-smooth continuous diusions having conditionally
independent increments dx
of the
intensity (t) > 0 dening its expectation by E(dn) = dt and =
1/2
. These
rules are known respectively as the dierential multiplications for the standard
Wiener process w(t) and for the forward dierentials of the standard Poisson
process n(t) compensated by its mean value t.
The linear complex spans b
0
= d + e
w
of formal increments d, e
w
, having
the only nonzero product e
2
w
:= e
w
e
w
= d given by a nilpotent in second order
element e
w
= e
w
, is two-dimensional commutative -algebra called the Wiener-Ito
algebra a
0
with the state l (b) = dening the drift E(dx
0
) = dt. Another basic
Ito -algebra is the direct sum b
1
= da of the nilpotent d and the unital -algebra
a = Ce
n
with usual multiplication of complex numbers rescalling the unit jumps
e
n
= e
n
e
n
e
2
n
dn of the standard Poisson process n(t) = (m(t) + t/) /
corresponding to = 1. It can be also written in the form of spans b
1
= d+e
m
,
where is dened by the drift E(dx
1
) = dt of dx
1
= dt +dm, with the formal
increments e
m
= e
n
d/ dm of the compensated Poisson process m(t) having
the products
e
2
m
= d + e
m
=
2
p, e
m
d = 0 = d e
m
,
where p = p p
n
. It is also associative
and commutative algebra called the Ito-Poisson algebra with the state l (b) = .
As in the case of Newton algebra d = Cd, the Ito -algebras b
0
and b
1
have
no Euclidean operator realization, but they can be represented by nonunital -
algebras of triangular 3 3-matrices b
0
= d+e
w
, b
1
= d+e
m
in the matrix
pseudo-Hermitian basis
d =
_
_
0 0 1
0 0 0
0 0 0
_
_
, e
w
=
_
_
0 1 0
0 0 1
0 0 0
_
_
, e
m
=
_
_
0 1 0
0 1
0 0 0
_
_
,
620 V. P. BELAVKIN
d
:= Id
I = d, e
w
:= Ie
w
I = e
w
, e
m
:= Ie
m
I = e
m
.
with respect to the falling identity (Ito) metric I =
_
=,,+
=,,+
, not to be
mixed with the identity I = [
] indexed by , = , + with
b
= 0 if = + or = and involution b b
represented by b
=
_
=
b
,
where
b
= b
+
describes the vacuum state
on the maximal Ito algebra b (k) of quadruples (b
)
=,
=+,
with independent values
b
+
C, b
+
k, b
and b
[bv
) := v
bv
= b
+
,
induced in the canonical representation by the vacuum vector v
=
_
+
_
with the
covector v
= (1, 0, 0) = v
b
_
+
= 0 b = 0
on b
k generated by bv
[bv
) =
_
b
b
_
+
b
+
b
+
0 b b
.
Let k be separable and J
= J
i = j
i
= e
i
b
i, j J
.
A basis e
J
= (e
j
)
j=
in b
i,j
:= e
i
e
j
= l (e
i
e
j
)
i
,j
(1.5)
for the basic vectors e
i
= (e
i
)
+
k
has quasiinverse =
as its -conjugation
i,j
=
,j
, the complex conjugation of
= [
i,j
] corresponding to the Her-
mitian adjoint covariance matrix
= [
i
,j
] =
.
QUANTUM DYNAMICS AND GIRSANOV TRANSFORMATION 621
The Gram matrix of a modular nonstandard basis has quasiinverse with a
common -real support = = dened as the minimal -real orthoprojec-
tion matrix = [
i
j
] =
such that
= , = ,
2
= =
. (1.6)
If the basis is not overcomplete, then = 1 such that for the standard basis
, otherwise
= , where =
1
2
_1
2
1
2
of
e
j
= l
_
e
i
e
j
_
i
,j
for adjoint covectors e
i
=
_
e
i
_
such that e
i
=
i,j
e
j
is the modular -
conjugation e
i
= e
i
represented by e
i
= J
e
i
J in terms of an antilinear isometry
J
= sJ = J
1
, where s = s
= s is any signature s
2
= 1 commuting with
b. The dual -Hermitian matrix =
_
i,j
by (1.5), where e
i
= e
i
represented by e
i
= J
i
J.
In the classical case we have always
J
in k
, or, if
the classical standard basis is overcomplete, satises the projectivity condition
= = corresponding to the normality of overcomplete basis in k
. Thus,
the classical standard basis is always selfdual in the sense that the completeness
condition
e
j
e
j
= I
= e
j
e
j
(1.7)
in the Hilbert space k
is satised for e
i
= e
i
i.e. e
i
= e
i
implying symmetricity
e
i
= e
i
of the matrix representation b b (k
cannot be selfdual,
unless the Ito state l is tracial,
l (e
i
e
k
) = l (e
k
e
i
) i.k J
.
which does not imply the commutativity e
i
e
k
= e
k
e
i
. However, one can
prove that the dual Ito algebra b =
b generated in b (k
) by the basis e
J
of the
transposed matrices e
i
= e
i
, commutes in the standard representation with the
modular Ito algebra b since e
i
e
k
= e
k
e
i
due to
i
k
=
i
k
for
i
k
=
_
e
i
e
k
_
+
= e
i
k
and m
i
k
= m
i
k
for the martingale part m
i
k
= e
i
e
k
i
k
d of e
i
e
k
in the following
multiplication table
e
i
e
k
=
i,k
d +
j
i,k
e
j
, e
i
e
k
=
i
k
d +m
i
k
, e
i
e
k
=
i,k
d +
i,k
j
e
j
. (1.8)
622 V. P. BELAVKIN
Here
i,k
=
,i
such that =
, =
_
i
k
= is -symmetric orthoprojector,
j
i,k
=
j
,i
. Note that
i
=
i,j
j
, including
as the case
= for i = , where
with
=
_
j
k,i
_
=
.
The components of the quantized vector-processes e
j
(t) = (t, e
j
) and e
j
(t) =
_
t, e
j
_
, dened on the Fock-Guichardet space F = (K) over K = kL
2
(R
+
)
L
2
k
as in [11][14], in general do not commute with their adjoints but commute with
their tensor-independent increments as in the classical Levy case. However, the
dual quantum standard Levy processes e
J
(t), e
J
(t) do not coincide as in the
classical case but are maximally entangled having mutually commuting maximally
correlated -Hermitian components e
i
= e
i
, e
i
= e
i
. They generate the dual
W*-ltration of maximal mutually commuting W*-algebras B
t
0
and
B
t
0
= J
B
t
0
J on
the Fock ltration over K
t
0
= L
2
[0, t] which are antiisomorphic by the transposition
B = B
L
t
as preduals of their L
1
-completions L
r
L
t
with respect to the symmetric
pairings q p
r
=
r
(q p) of B
r
= and L
r
induced by the reference weights
r
=
t
[L
r
. An adapted B-process with respect to the ltration (B
t
) is described by
a function Q(t) = q
t
B
t
mapping a totally ordered set T into the corresponding
-subalgebras of B := B
t
. The embeddings
t
(r) : L
r
L
t
are assumed to be
regular such that they are adjointable,
t
(r) :=
r
(t)
:= q
p
t
t
(r, q) [p
_
t
q B
r
, p L
t
with respect to the symmetric pairing given by . Then the adjoint comorphisms
r
(t) :=
t
(r)
r
(s)
s
(t) =
r
(t) r < s < t,
such that
r
=
r
(t)
t
for all localizations
t
: L L
t
at t r dened on L = B
, then, obviously
t
= [B
t
, and the
localizations are simply restrictions
r
(t) = E
r
[L
t
of the conditional expectations
E
r
: L L
r
as the positive projections onto L
r
embedded into the L
1
-completion
L of B B
r
.
An L-adapted process Z(t) = z
t
L
t
is called L
1
-martingale if
r
_
t, z
t
_
= z
r
(or E
r
_
z
t
= z
r
) r < t,
QUANTUM DYNAMICS AND GIRSANOV TRANSFORMATION 623
and is called global martingale if z
t
=
t
(z) (or z
t
= E
t
[z]). Every positive
normalized L
1
-martingale t p
t
denes a state
t
(q) = q p
t
on B which is
called global if p
t
is the global martingale given by a positive normalized p L.
2.2. Quantum adapted processes and martingales. From now on we assume
the invariance of all B
t
B with respect to left (and right) modular involutions
with respect to the symmetric pairing. Then the weight (p) = 1 p on L admits
the compatible conditional expectations E
t
: L L
t
as positive projections such
that E
s
E
t
= E
s
r, t R
+
onto L
t
= M
t
L:
E
s
_
1
t
_
= 1
s
, E
t
(a
pa) = a
E
t
(p) a p L, a B
t
.
Let B
t
denote the -algebras of adapted C-processes as continuous functions
Q : r Q(r) B
r
mapping r [0, t[ into the modular subalgebras B
r
B
t
such
that B
s
_ B
t
if s < t. Then the dual space / = B
to B =
t>0
B
t
with
respect to the integral pairing
Q P
:=
_
Q(t) P(t) dt Q B, P /
ds
for each t < .
2.3. Quantum Ito semimartingales. We consider adapted quantum Ito L
1
-
processes Z dened as the special semimartingales by
Z(t) Z(r) =
_
t
r
K(s) dB(s)
t
r
(K) .
Here K = (K
i
)
iJ
are adapted integrands indexed by a separable set J
with an
isolated point J
such that J
= J
for J
= J
and
t
r
(K) =
_
t
r
K
i
(s) dB
i
(s) +
_
t
r
K
(s) ds,
where B
i
(t) L
t
are not necessarily canonical QS martingales with QS dieren-
tials forming a basis for an Ito algebra for each t. The nite variation part
_
t
r
K
(s) ds :=
_
t
r
s
(dZ(s))
t
r
[Z(t) Z(r)]
624 V. P. BELAVKIN
is given by the deterministic integrator B
(t) = t1 = B
(t)
dening it as abso-
lutely continuous process
t
0
[Z] (t) = Z
0
+
_
t
0
K
(s) ds
for K
i
(t) := B
i
(t, )
= B
i
(t) i J
= (K
), where K
(t) =
K
i
(t)
, and therefore Z
(t) = Z
0
+
t
0
(K
).
Moreover, we shall assume the associativity
[X [Y Z]] = [[X Y] Z]
of QS integrals in terms of quantum stochastic covariation
[Y Z]
t
0
:=
_
t
0
[d (YZ) (dY) Z Y(dZ)] (s)
such that it can be written in terms of an Ito product
(K dB) (L dB) = (K L) dB
of noncommuting dY = K dB and dZ = L dB as
[Y Z] (t) =
_
t
0
dYdZ =
_
t
0
(K L) dB.
Thus, we assume that quantum Ito semimartingales form an integral -algebra
with respect to the bracket [Y Z] given by an associative quantum Ito -algebra
of the corresponding QS integrands K(t) dened as the QS derivatives D
Z
(t) of
Z at t.[13].One can prove similar as in [17][18] that every integral Ito -algebra
has also the canonical triangular block matrix representation in the B*-algebra
b (/) of a two-sided Hilbert-Schmidt module in place of the Hilbert space k for the
Levy-Ito algebra at each time t.
3. Quantum Hidden Dynamics and QS Filtering
3.1. Quantum object-output stochastic processes. Let M
t]
0
= M
o
M
t
0
de-
note the W*-product algebra generated on the Hilbert product H
t]
0
= hH
t
by the
present (quantum object) algebra M
o
B (h), say the matrix algebra as operator
algebra on h = C
d
, and the past (classical output) algebra M
t
0
B (H
t
0
), say gener-
ated by multiplication operators Y
t
0
: g
t
0
y
t
0
g
t
0
for adapted C-valued functionals
y
t
0
M
t
0
, g
t
0
H
t
0
of classical Levy process trajectories
t
0
= (r) : r [0, t) such
that M
t
0
L
Q
(, B), H
t
0
L
2
Q
(, B) on the Levy probability space (, B, Q).
The latter example has the split property M
t+s
0
= M
t
0
M
s
t
on H
t+s
0
= H
t
0
H
s
t
for every t and s > 0, and we shall assume this innite divisibility not only for the
QUANTUM DYNAMICS AND GIRSANOV TRANSFORMATION 625
Abelian algebras M
t
0
M
t
0
but also for the general W*-algebras M
t
0
, say generated
by the dual quantum Brownian or Levy processes
e
j
(t) := (t, e
j
) B
j
(t) , e
j
(t) :=
_
t, e
j
_
B
j
(t) j J
(s) Z(s)
) dA
(s)
t
r
(J
Z
ZI)
of the QS derivative K(t) = D
t
Z
dening the QS germ-matrix J
t
Z
= Z(t) +D
t
Z
I,
where Z := I Z ZI is the ampliation [Z
Z
r
(t) = V
r
(t) Z(t) V
r
(t)
r
(t, Z(t)) ,
where V
r
(t) : H
t]
0
F
t
H
r]
0
F
r
is a unitary or an isometric evolution
V
r
(t) = I+
t
r
(VL) such that
r
(Z
Z) =
r
(Z)
r
(Z) for any Z M
t]
0
. We shall
call the isometric W*-monomorphisms
t0
(t) semimorphisms as they preserve all
algebraic operations except,.maybe, the identity I
t]
0
M
t]
0
which is mapped into the
decreasing orthoprojectors P
r
(t) = V
r
(t) V
r
(t)
P
t+s]
r
s > 0 of the quantum
object survival from r up to t which is said to be stable only if P
r
(t) = I
r]
0
for all
t and r < t.
The input output semiunitary transformations V = V
r
(t) : r t form a
linear hemigroup
V
t0
(t) = V
t0
(r) V
r
(t) t
0
< r < t.
as the representation of a small category of the totally ordered set R
+
. Since the
seminal paper [22] it has been main object of study in quantum stochastics as a
resolving family for the Hudson-Parthasarathy (HP) QS dierential equation
dV(t) = V(t) (dt, S I) V(t) = L(t) A(dt)
written in the canonical Hudson-Parthasarathy basis A =
_
A
=+,
=,
as
dV(t) = V(t)
_
L
+
dt + L
dA
+ L
dA
+ L
+
dA
+
_
.
Here L = [L
]
=,
=+,
LI is the quadruple of QS logarithmic derivatives dening
the QS germ G(t) = V(t) S(t) J
t
V
for the Hemigroup V by S = I +L. The fol-
lowing theorem gives necessary HP unitarity conditions in a compact germ form.
They are also sucient for the existence and uniqueness of the semiunitary solu-
tions under the appropriate integrability conditions [11][12] for the operator-valued
adapted function L(t).
626 V. P. BELAVKIN
Theorem 3.1. The solution of HP equation is unitary, V
= V
1
(isometry,
V
= G
1
(pseudo-isometry G
G = I),
i.e. i the logarithmic germ S(t) = V(t)
J
t
V
is pseudo-unitary (pseudo-isometry).
This implies the relations
S
= I
, S
+
+ S
= 0, S
+
S
+ S
= 0, S
+
+ S
+
S
+
+ S
+
= 0.
for S
= I
+ L
, S
+
= L
+
, S
= L
, S
+
= L
+
.
3.2. Quantum stochastic ows and evolution equation. The QS Heisen-
berg equation for the maximal W*-algebra N
s
t
= B (F
s
t
) is generated by the vacuum
noise increments A
(t
) A
(t), t
[t, t
s
) of the canonical HP basis B
= A
. It
may actually be driven by a smaller Ito -subalgebra than b (k), corresponding to
a product W*-algebra N
t
0
B (F
s
t
) of quantum noise on Fock space F
s
t
= (K
s
t
)
for K
s
t
= L
2
k
(t, t
s
], say generated by the classical Langevin forces
f
k
(t
) given by
an Abelian Ito -subalgebra for each t
s
= t + s > t on F
s
t
. It is usually described
by the Langevin QS equation
d
Z
r
(t) =
_
dt,
r
_
t, J
t
Z
_
Z
r
(t)
_
,
Z
r
(r) = Z(r) .
where the QS germ (t) = J
t
= (t) for with (J) = SJS
r
_
t, J
t
Z
_
= G
r
(t) J
t
Z
G
r
(t)
=
r
_
t,
J
t
Z
_
,
is obtained by applying quantum Ito product formula
d (VZV
) =
_
dt, GJ
t
Z
G
VZV
_
.
Theorem 3.2. [14] The solution
Z(t) = (t, Z(t)) of the QS Heisenberg equation
is homomorphic (unital,
r
(t, I) = I) i all germs
r
(t) are pseudo-homomorphic
(unital):
r
_
t, J
J
_
=
r
(t, J)
r
(t, J) , (
r
(t, I) = I) .
In particular, J
t
X
= I
t
x X(t) gives Langevin equation, and the case J
t]
Y
= J
t
Y
1
with
r
_
t, J
t
Yt
_
=
Y
r
(t) I +
r
(t, D
t
Y
) gives the output equation:
d
X =
_
dt,
_
t,
X
_
X(t)
_
, d
Y =
_
dt,
_
t, D
t
Y
__
.
3.3. Quantum quasi-Markov hemigroups and master equation. Let M
t]
=
A
t]
B
t]
be generated by the input W*-algebras A
t]
= A
o
A
t
on the Hilbert spaces
H
t]
= h H
t
with an initial cyclic state-vector
t]
0
=
t
0
and an output mod-
ular algebras B
t]
= B
o
B
t
from the commutants
A
t]
= J
A
t]
J of A
t]
on H
t]
. We
assume that (A
t
, B
t
) is increasing W*-product system on the components H
t
= F
t
0
of Fock space F
0
= F
t
0
F
t
and take the vacuum
t
F
t
0
as an initial product
state vector
t
0
separating B
t
. A QS adapted dynamics
r
(t) : M
t]
M
r]
N
r
is called A
t]
-Markov on vacuum state vectors
F
t
for the decreasing future
quantum noise W*-algebras N
t
B (F
t
) if
t
(t
s
, Z) = E
t
(t
s
)
t
(t
s
, Z) E
t
(t
s
) A
t]
t > r, Z A
t
M
s]
t
, (3.1)
QUANTUM DYNAMICS AND GIRSANOV TRANSFORMATION 627
where E
t
(t
s
) =
s
:
t]
t]
s
of
t
H
t]
into H
t]
F
s
t
. Then
the family of CP maps
t
(t
s
) is output compatible,
_
t
(t
s
, Z) , B
t]
= 0, and can
be lifted to a hemigroup of CP maps M
t+s]
M
t]
satisfying the B
t
-modularity
condition
t
(t
s
, Y
t
ZY
t
) = Y
t
(t
s
, Z) Y
Y B
t
, Z M
t+s]
, Y
t
= Y I
s]
t
.
It is resolving family for the generalized Lindblad equation
d
dt
r
(t, Z(t)) =
r
_
t,
_
t, J
t
Z
__
,
r
(r, Z) = Z M
r]
. (3.2)
dened by the form-generator (t, J) :=
+
(t, J) commuting with B
t]
on the germs
J G
o
A
t
of the process Z(t
s
) A
t
M
s]
t
at each t. Here G
o
= M
o
g is the initial
germ algebra dened as the M
o
-envelope of the -semigroup g = 1 m unitizing
the Ito -algebra M(t) = a b m of the mutually commutative Ito subalgebras
a, b b (k) generating respectively A
t
=
t
0
(A
) and B
t
=
t
0
(B
) by multiple
QS integrals
t
0
(M
) of M
n=0
M
n
as dened for M =
_
t>0
M(t) dt in
[11][12]. Every such completely bounded generator (t) can be uniquely lifted to
the B
t]
-modular map G
o
M
t]
M
t]
.
Proposition 3.3. Let J(t) = J
t
Z
be the germ of a QS process Z : t M
t
with
J
+
(t) = 0 for all t. Then the A
t]
-Markov generator (t) for a QS hemigroup of
r
(t, Z(t)) = V
r
(t) Z(t) V
r
(t)
+
on G
o
M
t]
written in terms of the logarithmic QS derivatives L
= S
of the QS evolution V
r
(t) as
_
t, J
t
Z
_
=
_
L
+
Z + ZL
+
+ L
+
+ L
+ J
_
(t) . (3.3)
Note that the condition J
+
= 0 corresponds to the property of Z(t) to be a local
vacuum martingale M(t): E
r
(t) M(t) E
r
(t) = M(r), and (t, J
t
Z
) = (t, J
t
M
) +
J
+
(t) for any vacuum semimartingale Z(t) = M(t) +
_
t
0
J
+
dr. In particular, if
D = J X = 0, reads simply as the Lindbladian on X = Ix,
(X) = L
+
X + XL
+
+ L
XL
(x) . (3.4)
However, it denes the A
o
-Markov dynamics only if (t, J) A
o
for J G
o
, e.g.
if L
(t) = IL
(t) with L
(t) A
o
.
For the output processes Z(t) = Y
t
1 having the QS derivative D = 1b(t) Y
t
with b(t) B t 0,
_
J
t
Yt
_
=
_
(1) + L
+
+ L
+ b
+ b
+
_
Y
t
.
Here Y
t
B
t]
is a local QS semimartingale, e.g. a local QS martingale charac-
terized by b
+
= 0, say Y
t
= W
t
(b) y with any y B of zero expectation and
a QS exponential martingale W
t
(b) B
t
as the normal-ordered Weyl exponents
W
t
(b) =: exp [
t
0
(b)] :. The compatibility of (J) with B
t]
on the QS logarithmic
germs J = J
t
Y
Y
1
t
= 1 +b(t) is ensured by L
A
t]
.
In the following we assume that b = a is dual to a modular Ito B*-algebra in
the standard representation a b (k).
628 V. P. BELAVKIN
3.4. Filtering of quantum hidden quasi-Markov dynamics. Let us consider
a QS hemigroup
F
r
(t) of propagators H
r]
F
r
H
t]
F
t
satisfying the QS
evolution equation
d
F
r
(t) =
F
r
(t)
_
L
+
(t) dt + L
k
(t) d e
k
t
_
=
_
dt,
F
r
(t) L(t)
_
F
r
(r) = I
r]
. (3.5)
with L
k
= L
k
and e
k
(t) =
_
t, e
k
_
. Applying the QS Ito formula to
r
(t, Z(t)) =
F
r
(t) Z(t)
F
r
(t)
,
we obtain a QS evolution equation
d
(t, Z(t)) =
_
dt,
_
t,
_
t, J
t
Z
_
X(t)
__
Z(t) M
t]
(3.6)
describing in terms of (t, J) = C(t) JC(t)
a QS entangling process
(t) : M
t]
M
o
A
t
by the hemigroup
r
(t) of simple CP transformations
r
(t) : M
t]
M
r]
A
tr
r
.
They are dened by the logarithmic QS germ C(t) = I + L(t) of
F
r
(t) given by
the logarithmic QS derivative
L(t) = L
+
(t) d +
kJ
L
k
(t) e
k
. (3.7)
In general the QS transformations
r
(t) map M
t]
not into a W*-algebra but into
the kernels [15] which are aliated to M
r]
A
tr
r
since, as it follows from the next
Lemma, they satisfy the quasi-Markovianity condition
t
(t
s
, Z) = E
t
(t
s
)
t
(t
s
, Z) E
t
(t
s
) A
t]
t > r, Z A
t
M
s]
t
following from the condition (3.1) for
r
(t, Z). This quasi-Markovianity, dened
as the Markovianity with respect to an increasing algebras A
t]
in place of a single
object algebra A
o
, and also the uniqueness of the solutions to (3.5), it is sucient to
assume that only the operator-functions L
+
(t) and L
(t) A
t]
as
aliated to A
t]
.
Lemma 3.4. Let the A
t]
-Markov hemigroup
r
(t) be dened as the unique
solution of the generalized Lindblad equation (3.2) on the increasing M
t]
with the
generator (3.3) (J) =
+
(J). Then it is unraveled by the A
t]
-Markov QS hemig-
roup of simple CP transformation
r
(t) in the sense that it is the conditional
expectation
r
(t, Z(t)) =
r
_
r
(t, Z(t))
:= E
r
r
(t, Z(t)) E
r
corresponding to the vacuum-induced white noise thermostates
r
[X] =
r
X
r
on
the algebras A
r
generated by the independent increments of the input quantum
Levy noise e (t) e (r).
QUANTUM DYNAMICS AND GIRSANOV TRANSFORMATION 629
Proof. Since E
r
(t) E
t
= E
r
= E
t
E
r
(t), the vacuum expectation
r
=
r
(t)
t
of
d
r
(Z) (t) is dened by the expectation
r
_
t,
+
(J
t
Z
)
_
dt of d
r
(Z) (t) written by
modularity property as
t
_
d
r
(Z)
(t) =
F
r
(t)
t
_
_
dt,
_
t, J
t
Z
_
Z(t)
_
F
r
(t)
r
_
t,
+
(t, J)
_
dt,
where
+
(J) dt =
t
_
_
dt, C(t) JC(t)
__
with J
+
= 0 is a generalized Lindblad
generator
( (J) Z)
+
= C
+
Z + ZC
+
+ C
+
+ C
+ J
+
(J) . (3.8)
In fact, since C
+
= L
+
and C
k
L
k
(t) e
k
= L
k
e
k
with e
k
, this
+
(t) coincides with the generator (t) =
+
(t)
of the A
t]
-Markov backward master equation vacuum induced by the isometric or
unitary evolution V
r
(t) on M
t]
. Thus, expectation
r
r
(t) of the QS bracket
evolution
r
(t) driven by the white thermonoise e (t) must coincide by the unique-
ness argument with the A
t]
-Markov evolution
r
r
(t) vacuum induced by the
Heisenberg evolution
r
(t).
The following theorem denes the structure of the generators for QS master
equations induced by the general QS dynamics with respect to any output quantum
Levy process given by a modular Ito B*-algebra. This result extends the semi-
quantum ltering theory [14] determined by any classical output Levy process to
fully quantum case. The noncommutative ltering in the A
o
-Markovian case was
outlined for quantum nite-dimensional Wiener temperature noise corresponding
to a Wiener-Ito modular algebra in [10].
Theorem 3.5. Let b be a modular output Levy-Ito B*-algebra with the symmetric
basis e
J
not necessarily complete in a. Then the A
t]
-Markov QS evolution
(t, x) =
F
0
(t) x
F
0
(t)
, t 0
restricted to the object algebra A
o
and ltered with respect to B
t
satises the Heis-
enberg QS equation
d
r
(t, x) +
(t, (t, x)) dt =
i,kJ
_
t,
j
(t, x)
_
d e
j
(t) , (3.9)
where (x) = Kx + xK
i,kJ
L
i
x
i,k
L
k
= (x) denes the Lindbladian
(3.4) in the Hermit-symmetric basis with K = L
+
, L
k
= L
k
and
j
(t, x) = L
j
(t) x + xL
j
(t) +
i,kJ
L
i
(t) x
i,k
j
L
k
(t) ,
are the uctuating coecients =
_
j
_
jJ
with
i,k
j
=
j
k,i
.
Proof. By QS product Ito formula applied to
F
0
(t) x
F
r
(t)
+
(x) = L
+
x + xL
+
+ L
i
x
_
e
i
e
k
_
+
L
k
= (x) ,
630 V. P. BELAVKIN
since
_
e
i
e
k
_
+
=
i,k
, and by e
i
e
k
=
i,k
j
e
j
we obtain
(x)
j
e
j
= L
i
e
i
x + xe
k
L
k
+ L
i
xe
i
e
k
L
k
=
j
(x) e
j
L
t
q[x
t
_
t
:=
_
q
x
t
_
= (q
x
s
)
q[
s
_
x
t
__
s
s t, q D
s
.
It can be equivalently described by the process X (t) = x
t
in (D, ) with respect
to the generalized Cameroon-Martin density
m
t
_
1
t
_
L
t
, where
B
t
can also be represented by a density s
t
_
r
t
_
as
_
x
t
_
=
_
_
x
t
_
r
t
_
=
_
x
t
_
r
t
__
=
_
x
t
s
t
_
.
The noncommutative Girsanov theorem makes the process Z(t) into a (B, )-
martingale with respect to a transformed state . If Z(t) = (x
t
) given by an
injective transformation of a martingale x
t
in the modular -algebra (D, )
(
(1
t
) r
t
in
(B, ) as the density of with respect to such that
(Q(t) Z(t) (t)) = (t) [Q(t) Z(t)
=
_
(Q(t)) x
t
_
Q(t) B
t
.
Any other state on D dened by the martingale density q
t
A
t
is also repres-
ented in (B, ):
_
x
t
_
=
_
(x
t
)
(q
t
)
_
=
_
x
t
q
t
_
.
If = = , this can be simply written in the form of the transformation
identity
_
q
t
_
[
_
x
t
__
q
t
[x
t
_
. x
t
D
t
, q
t
A
t
.
QUANTUM DYNAMICS AND GIRSANOV TRANSFORMATION 631
4.2. Standard Ito pairing and marginalization. Let A be a C*-algebra with
a vector state (A) = ([A)
H generated by A
.
Let (a, l) be a B*-Ito -algebra with
l (a) = (v
[av
) v
av
= a
+
in the canonical matrix representation a b (k) of the Hilbert space k generated by
the action of a on the vector v
=
_
+
_
, and b a
=
_
c
+
d +c
_
with the dual subalgebra
b b (k
)
of J
J = a
+
d+a
represented in b (k
k generated
by a
on v
+
. We denote by R = r
A and S = s
)-pairing
R, S := (v
+
[RSv
+
) =
_
[R
, S
+
_
R R, S R
, (4.1)
induced by the state l , where v
+
= (
on the completion R
= R
of the commutant R
represented on k
as
the dual space to the normed B*-algebra R. Note the decomposition R
= A
corresponding to R = R
+
R
+
d+R
and S = YI+
S
can be written as
_
R
,
S
_
+
+
, Y
_
=
X, P
+
_
+
_
R
,
S
_
+
+
, Y
_
=
X, P
+
_
+
_
R, S
_
.
Here R
= X
R
is the germ-martingale, (R
+
= 0, of X, uniquely dened
by the decomposition S = S
+
d +S
.
It can be easily seen that the standard pairing has the modularity property
_
CRC
, S
_
=
_
v
+
[CRSC
v
+
_
=
R, C
SC
_
S R
, R R
uniquely dened by
v
+
RC
= v
+
CR = v
+
C
R R R
as C
:= J
J C
= J
1
. Here denotes
the left involution
_
X C
Z
_
=
_
XL
Z
_
with respect to the modular pairing
(X
Y) = X, Y for
X = X
, Y R
= X =
J
XJ:
(X Z) :=
_
Xv
+
[Zv
+
_
=
_
Zv
+
[Xv
+
_
(Z X) X, Z R.
A positive normalized element A
:= A
A) = 0 (A
A) = 0. A simple conditionally
CP transformation (R) = (CRC)
+
from R into A such that (
S) B := B
transforms B
into S
by
R,
( ) = (R) = (R) , R R, B
632 V. P. BELAVKIN
if it is given by the transposable C R. One can show that
is also conditionally
CP such that each
by the decomposition
( ) = I + ( )
S
, where
( ) = L
+ L
+L
(4.2)
in terms of L = CI is decomposed as ( ) =
( ) d +
( ).
4.3. Dual reduced and unravelling QS L-dynamics. Let A
t
be the increas-
ing W*-algebras generated by quantum Ito-Levy noises e
j
(r) for r [0, t) and
j J on Fock space F =
_
L
2
k
_
. The modular subalgebras
B
t
= J
B
t
J are deed
on F
=
_
L
2
k
_
as the dual to B
t
generated by e
i
= (e
i
) for i J
with the
Weyl operator basis W
t
(b) : b b, where b is the modular Ito algebra spanned
by e
and e
J
. It is given by the solutions to
dW
t
(b) = W
t
(b) (dt, b) , W
0
(b) = I.
This solution can be written in the normal form in terms of the logarithmic matrix
l
= ln (I
+ b
) as.
W
t
(b) = e
t
0
b
i
+
(r)dA
+
i
e
t
0
l
i
k
(r)dA
k
i
e
t
0
b
k
(r)dA
k
t
0
b
+
(r)dr
.
If the QS dynamics is a A
t]
-Markov process on the state vector =
hF
with the vacuum
F
t
, the dual state dynamics
t
(r) form a forward hemigroup
deed on the predual space L = A
r]
by the CP maps
r
(t)
:
X
t
,
t
(r, )
t
:=
r
(t, X) ,
r
_
X
t
,
r
(t)
_
t
X A
t]
, A
r]
.
Here and below we shall take the vacuum (A
t]
, A
t]
)-pairings
X, :=
[X
_
t
__
X,
t
, X A
t]
, A
t]
.
Let B
t]
B
t
of
A
t]
generated an initial modular subalgebra
B
o
A
o
and the Weyl basis W
t
(a),
a
b a. If the QS dynamics is also Markov with respect to B
t]
A
t]
for all t,
the dual dynamics can be reduced to a hemigroup of
t
(r) : B
r]
B
t]
. It is given
by the marginalization
t
(r) as the preduals to the restrictions
r
(t) :=
r
(t) [B
t]
t
(r)
.
In particular, the reduced evolution
t
=
t
( ) from the initial state on B
o
A
o
is dened by
X
t
,
t
( )
=
0
(t, X
t
) ,
, X
t
B
t]
, B
o
.
The reduced states
t
can be described on X
t
= W
t
(a) x with x B by the
characteristic function
C
t
(a, x) = W
t
(a) x,
t
=
0
(t, X
t
) ,
satisfying the equation
dC
t
(a, x) = d X
t
(a) ,
t
for C
t
(a, x) = X
t
(a) ,
t
.
QUANTUM DYNAMICS AND GIRSANOV TRANSFORMATION 633
Note that since
t
(r) mapping B
r]
b, x
B
o
by the equation
dC
t
(a, x) = d X
t
(a) ,
t
for C
t
(a, x) = X
t
(a) ,
t
,
where d X
t
(a) ,
t
= 0 in the martingale case a
+
= 0, X
t
, d
t
=
X
t
,
(
t
)
_
dt
and
dX
t
, d
t
=
iJ
X
t
a,
i
( ) e
i
_
dt X
t
a,
( ) dt a
b. (4.3)
4.4. Noncommutative QS master equation. The following theorem denes
the structure of the generators for QS master equations induced by the general
QS dynamics with respect to any output quantum Levy process given by a modu-
lar Ito B*-algebra. This result extends the semi-quantum unraveling dynamics [5]
determined by any classical output Levy process to the quantum Levy-Ito temper-
ature process. The proof of the theorem in the general case is similar to the proof
of the semi-quantum ltering theorem [7] for the commutative processes e
i
= e
i
but noncommutative initial algebra B. It was outlined in [10] for fully quantum
diusive case of noncommuting self-adjoint Wiener processes e
i
= e
i
correspond-
ing to quantum nite-dimensional Wiener-Ito nilpotent in second order modular
algebra.
Theorem 4.1. (Main) Let
r
(t) be the quantum hemigroup dynamics on the
increasing W*-subalgebras B
t]
A
t]
generated by
B
o
A
o
and the modular Levy-
Ito B*-algebra
b a with a symmetric basis d, e
i
: i J
indexed by J
= J
+
(t). Then the reduced states
(t) =
t
( ) B
t]
jJ
j
(t, (t)) d e
j
(t) (0) B
o
(4.4)
with QS modular noises e
j
(t) =
_
t, e
j
_
and uctuating coecients given by
j
(t, ) = L
j
(t) + L
j
(t) +
i,kJ
L
k
(t)
j
i,k
L
k
(t) .
including
= with L
= K = L
+
and L
i
= (L
e
i
= L
+
e
i
.
Proof. The predual generator : B
t]
S
t]
for the generalized Lindbladian writ-
ten for an A
t]
-martingales Z = X in the form (3.3) is found by modularity in (4.2)
for the A
t]
-adapted transposable C = I + L such that C
iJ
e
i
L
i
, L
iJ
e
i
L
i
,
634 V. P. BELAVKIN
where L
i
= L
+
e
i
, L
i
= e
i
+
are determined by the completeness as
e
j
, L
_
=
_
X[
_
e
j
L
+
_
=
_
e
j
e
i
_
+
_
X[L
i
_
=
j
i
e
i
_
X[L
_
=
_
X[L
j
_
since
i
j
=
_
e
i
e
j
_
+
= e
i
j
is either
i
j
or a an idempotent kernel such that
i
j
e
j
=
e
i
_
in k
=
_
L
e
j
_
+
and
e
j
, L
_
=
_
X[
_
L
e
j
L
+
_
=
_
e
i
e
j
e
k
_
+
_
X[L
i
L
k
_
=
_
X[L
i
j
i,k
L
k
_
.
since
_
e
i
e
j
e
k
_
+
=
_
e
j
e
i
e
k
_
+
=
m
i,k
_
e
j
e
m
_
+
=
j
m
m
i,k
=
m
i,k
. Therefore,
j
( ) = L
j
+ L
j
+ L
i
j
i,k
L
k
in (4.2) determined for j J
b spanned by e
J
.
This formula also determines
( ) with
i,k
=
i.k
by
_
X
t
[
(
t
)
_
=
_
X
t
[
_
L
+ L
+L
+
_
with (L
+
= L
,
_
L
+
= L
given by L
= (L
+
= L
+
. This complets
the proof.
Note that if L(t)
S
t]
corresponding to the span L
= e
i
L
i
of (L
+
by the
family e
J
of Ito basis e
J
with B
t]
-adapted L
i
(t), the B
t]
-Markov QS evolution
(4.4) is completely unravelled by the hemigroup of QS propagators
V
r
(t) satisfying
the QS unraveling equation
d
V
r
(t) +
V
r
(t) K (t) dt =
V
r
(t) L
i
(t) d e
i
,
V
r
(t) = I.
This can be easily seen by applying the Ito formula to the solution of (4.4) for t > r
with (r) = in the form
t
(t, ) =
V
r
(t)
V
r
(t)
corresponding to A
t]
-Markovianity of the dual dynamics
= . Such unraveling
is predetermined by hemigroup Markovianity on the W*-algebras B
t]
. If the basis
e
J
of b is not complete in k
= J. Note that
such ravelling is not unique and can be chosen classical by taking a commutative
Ito algebra c b
_
k
_
which is always modular with the basis e
j
= e
j
in the
orthogonal complement k
= k k
independent of e
J
= e
J
. The incompletely unravelled
QUANTUM DYNAMICS AND GIRSANOV TRANSFORMATION 635
B
t]
-Markov dynamics
r
(t, Z)].
If is unital, the density operators (t) are normalized on B
t]
to the positive
martingale (t) = I, (t) describing the statistics of the complementary pro-
cesses y
j
(t) = V
0
(t) e
(t) V
0
(t)
making sense in
(pq
) =
_
q
p
_
, (q
p) =
_
pq
_
p, q A
with the right modular involution dened as
=
_
q
p
_
or right q[p
+
=
_
pq
_
modular pairings
such that p
[q
= (qp) = q
[p
+
, but usually a symmetric pairing
q
[p (q p) = (q p) p
[q
is preferred for use, satisfying the modular conditions
aq[p =
q[a
p
_
, qa[p =
_
q[pa
_
a, q, p A
with respect to another left involution and
:= (q(p)) (q p) ,
where is a root of the modular automorphism
= of A uniquely determined
by the polar decomposition a
= (a)
a) = (aa
.
5.2. Regular states and Radon-Nikodym densities. The noncommutative
L
1
-space L = L
is dened by the L
1
-completion A
= A
:= sup
a1
[a[b
[ =
1
2
sup
a1
(a
b + b
a) ([b[
)
636 V. P. BELAVKIN
Without changing notation we can extend by weak continuity the bilinear form
(b a) to the natural pairing ML C of the the norm-adjoint algebra M = L
as the weak
closure of A L
M equipped with
|a|
= inf c : a
a c1 |a| .
Note that Mis unital -algebra called W*-algebra as having the preadjoint Banach
space L = M
= L, M
[p = q[apa
q M, p L, b = a
A
A linear positive functional : A C of the form
(a) = (a p) = p[a
a A,
dening by a positive normalized p L
+
, (p) = 1 an expectation on A, is
referred as the regular state, and as the normal regular state when it is extended
on M = L
b) = 0 b I, then
the induced state = is absolutely continuous with respect to in the sense
(b
b) = 0 (a
a) = 0 b a, a A.
It has density r =
(p) B
as a CP contraction
: A
a) t(b
: L L
in
d[
(p)
= (d) [p p A
, d D.
It is easy to see that the comorphism, intertwining the left involution
with
=
_
(d)
p
_
= d
(p) p L, d D
and
(pa) =
is a CP map and
(L) I
:= d D : (d) = 0
1
(0) .
If : A D is an injection inverted by and having the -modularity property
_
(d) a(d)
_
= d (a) d
a A, d D,
then (A) I
A, and
of from A onto L = A
= O and
therefore cannot be unital if I
,= O.
More generally, a completely positive -contraction : B
A quasiinverse in
the sense = to a -morphism of D
A into B is called quasimorphism
(quasiexpectation inverting ) if it satises the -modularity condition
(b (d)) = (b) d b B, d D
in a modular -algebra (
= Osince I
:=
1
(0) =
e
D for e
=
_
d
_
for all d D due to
_
d
_
= (d)
by
[B =
i = , where =
. In the
most important case = of injective this denes
on B by marginalization
i [D = .
5.4. Semiconditioning and semi-Markovianity. A comorphism
for =
as a modular map
_
d
p
_
= (d)
(p) p
A
, d D
inverting a comorphism
[D of the
on
A. Then the composition E =
, given by the
comorphism
and
, is a positive projection L
pd) = d
E(p) d p L
, d D.
The restriction E[
A =
and
qd) = d
(q) d q
A, d D.
This all remains true if
1
, a
1
, d
1
) =
_
1
(d)
a
1
1
(d
1
)
_
where a
1
is from a central -subalgebra A
1
B
1
.
A contractive CP kernel : D
1
A
1
D
1
L
r
(V
t
), r, t R
+
,
he asked what other ways may such cocycles be generated are they all of this
type? Armed with various quantum martingale representation theorems ([21, 23]),
the latter obtained in collaboration with Parthasarathy at the 1984-5 Warwick
Symposium on Stochastic Dierential Equations, Robin and I were able to pro-
vide an immediate answer. Our answer was a qualied yes, if the process of ran-
domisation is broadened to involve quantum Brownian motion ([10]) along with
Received 2010-9-21; Communicated by D. Applebaum.
2000 Mathematics Subject Classication. Primary 46L53, 81S25; Secondary 47D06.
Key words and phrases. Noncommutative probability, quantum stochastic cocycle, E
0
-
semigroup, CCR ow, holomorphic semigroup.
641
Serials Publications
www.serialspublications.com
Communications on Stochastic Analysis
Vol. 4, No. 4 (2010) 641-660
642 MARTIN LINDSAY
the Poisson-type process derived from dierential second quantisation, if adapt-
edness to the corresponding ltration of operator algebras is imposed, and if one
assumes sucient regularity for the cocycle. The setting is Wiener space L
2
(),
with its semigroup of shifts inducing the semigroup (
t
)
t0
on B(L
2
()), now
called CCR ow. This is translated to symmetric Fock space T over L
2
(R
+
) via
the Wiener-Segal-Ito isomorphism (see, for example [27]).
Theorem 1.1 ([22]). Let V = (V
t
)
t0
be a unitary quantum stochastic cocycle
on h, and suppose that V is Markov-regular. Then there is a unique operator
F B(h h) such that V satises the quantum stochastic dierential equation
dV
t
= V
t
d
F
(t), V
0
= I
hJ
.
(All the terms used here will be dened in Section 2). By the quantum Ito
formula, the coecient F necessarily satises the Hudson-Parthasarathy unitarity
conditions q(F) = 0 = r(F), where
q(F) := F
+F +F
F and r(F) := F +F
+FF
,
and denotes the quantum Ito projection
_
0
I
h
L/2 L
C
L C I
h
_
,
with C being unitary, H selfadjoint and L arbitrary, and the quantum stochastic
dierential equation reads
dV
t
= V
t
_
LI
J
dA
+(C I
h
) I
J
dN
t
L
C I
J
dA
t
+(iHL
L/2) I
J
dt
_
,
with (A
t
)
t0
, (N
t
)
t0
and (A
t
)
t0
being respectively the creation, preservation
(number, exchange or guage) and annihilation processes.
The result was proved as follows. Letting K denote the generator of the expec-
tation semigroup of V , then K is bounded (by Markov-regularity) and
X :=
_
V
t
I
hJ
_
t
0
V
s
(K I
J
)ds
_
t0
denes a quantum martingale: E
s
[X
t
] = X
s
(s t), satisfying X
0
= 0, which may
be shown to be regular in the sense of Parthasarathy and Sinha ([41]) so that
X
t
=
t
(G) =
_
t
0
L
t
t
dA
+
_
t
0
(C
t
t
I
hJ
) dN
t
+
_
t
0
M
t
t
dA
t
,
where G =
__
0 M
t
L
t
C
t
I
hF
_ _
t0
for bounded processes L
t
, M
t
and C
t
. The proof
is completed by verifying that each of the processes
(V
t
L
t
t
)
t0
, (V
t
M
t
t
)
t0
and (V
t
C
t
t
)
t0
is (a.e.) constant. If the cocycle is instead adapted to the ltration of a non-
minimal variance quantum Brownian motion then the same result holds (with no
preservation integral) with similar proof, but using the martingale representation
theorem for martingales with respect to this ltration ([21]).
DIFFERENTIATING QUANTUM STOCHASTIC COCYCLES 643
Journe then gave a qualied no to Arvesons question. Along with an analysis
of quantum stochastic cocycles V , assumed only to be strongly continuous, he
gave an example to show that in general V will not satisfy a quantum stochastic
dierential equation. The essential point is that we may associate an operator
L to V and, in favourable circumstances also an operator C, but the domains
of L and L
= I
h
[ : h h h, respectively E
= I
h
[ : h
t
h
t
h
t
(1.1)
for vectors and
t
from Hilbert spaces h and h
t
. Ultraweak tensor products are
denoted and purely algebraic tensor products by . For an operator space V
and Hilbert space h,
V
h
denotes the ampliation x V x I
h
.
644 MARTIN LINDSAY
Fix now, and for the rest of the paper, two Hilbert spaces h and k. Set
k := Ck
and, for c k, set c :=
_
1
c
_
k; also, for f L
2
(R
+
; k) set
f(t) :=
f(t). We often
make the identication
h
k = h (h k),
and employ the quantum Ito projection := [
0
I
] B(h
k) = B(h (h k)).
2. Quantum Stochastic Cocycles
For 0 r < t , the symmetric Fock space over L
2
([r, t[; k) is denoted by
T
[r,t[
and the identity operator on T
[r,t[
by I
[r,t[
; T
R+
is abbreviated to T. We use
normalised exponential vectors:
(f) := e
|f|
2
/2
(f), f L
2
(R
+
; k),
where (f) is the exponential vector (1, f, (2!)
1/2
f
2
, ) T. As is well-known,
the family (f) : f L
2
(R
+
; k) is linearly independent and total in T. The
following considerable strengthening of the latter property, due to Parthasarathy-
Sunder for one dimensional k and Skeide for general k, has proved very useful
(see [27] for a proof). For a subset S of k dene
c
S
:= Lin(f) : f S
S
where S
S
:= f L
2
(R
+
; k): f is an S-valued step function (with the convention
that we always take the right-continuous versions), and abbreviate c
k
to c.
Proposition 2.1. Let T be a subset of k containing 0. Then the following are
equivalent:
(i) T is total in k;
(ii) c
T
is dense is T.
Write W(f) (f L
2
(R
+
; k)) for the unitary (Weyl) operator on T determined
by
W(f)(g) = e
i Imf,g)
(f +g), g L
2
(R
+
; k),
R
t
(t R
+
) for the unitary (time-reversal) operator on T determined by
R
t
(g) = (r
t
g) where (r
t
g)(s) =
_
g(t s) s [0, t[,
g(s) s [t, [,
(g L
2
(R
+
; k)),
S
t
(t R
+
) for the isometric (shift) operator on T determined by
S
t
(g) = (s
t
g) where (s
t
g)(r) =
_
0 r [0, t[,
g(r t) r [t, [,
(g L
2
(R
+
; k)),
t
(t R
+
) for the *-homomorphic (shift) operator on B(h T) determined by
u(f),
t
(T)v(g)
_
= (f
[0,t[
), (g
[0,t[
)
u(s
t
f), T v(s
t
g)
_
,
(T B(h T), f, g L
2
(R
+
; k), u, v h), and E for the completely positive and
contractive (expectation) map
E = id
B(h)
(0)
: B(h T) B(h).
DIFFERENTIATING QUANTUM STOCHASTIC COCYCLES 645
By a quantum stochastic contraction cocycle on h with noise dimension space k,
we mean a family of contraction operators V = (V
t
)
t0
on h T satisfying
s V
s
is strongly continuous,
V
t
B(h T
[0,t[
) I
[t,[
,
V
r+t
= V
r
r
(V
t
) and V
0
= I
hJ
, r, t R
+
.
Let QS
c
C(h, k) denote the collection of such cocycles. For V QS
c
C(h, k),
P
0,0
:=
_
E[V
t
]
_
t0
denes a contractive C
0
-semigroup on h called the expectation semigroup of V .
The semigroup property follows from the identities
E = E E
r
and E[V
r
(T I
J
)] = E[V
r
]T, r R
+
, T B(h),
for the conditional expectation maps
E
r
: B(h T) B(h T), X
_
id
B(hJ
[0,r[
)
(0
[r,[
)
_
(X) I
[r,[
, r 0.
The continuity of P
0,0
corresponds precisely to the strong (equivalently, weak op-
erator) continuity assumption on V ([37], Lemma 1.2). The expectation semigroup
is just one of the family of associated semigroups of V :
P
c,d
:=
_
(id
B(h)
(c
[0,t[
),(d
[0,t[
)
)(V
t
)
_
t0
, c, d k,
which together determine V through the semigroup representation ([34], Propo-
sition 6.2). In fact any family P
c,d
: c T
t
, d T where T and T
t
are each
total in k and contain 0 suces, thanks to Proposition 2.1. This is important
since, for example, if k is one-dimensional then one may take T = 0, 1 and V is
then determined by its expectation semigroup together with just three of its other
associated semigroups.
Write QS
c
C
Mreg
(h, k) for the subclass of Markov-regular QS contraction cocy-
cles, that is those whose associated semigroups are all norm-continuous. Contrac-
tivity of V actually implies that Markov-regularity is equivalent to norm-continuity
of just the expectation semigroup ([34], Theorem 6.6).
Example 2.2 (Weyl cocycles). For c k,
W
c
:=
_
I
h
W(c
[0,t[
)
_
t0
denes a Markov-regular QS contraction cocycle; E[W
c
t
] = e
t|c|
2
/2
I
h
.
Two useful constructions of new cocycles from old are as follows. Let V
QS
c
C(h, k). Then the dual cocycle
V is given by
V
t
:=
_
R
t
V
t
R
t
_
t0
;
its associated semigroups are given, in terms of the associated semigroups P
c,d
:
c, d k of V , by
P
c,d
t
= (P
d,c
t
)
; in particular,
V is Markov-regular if and only if
V is. The associated cocycles of V are dened by
V
c,d
:=
_
(W
c
t
)
V
t
W
d
t
_
t0
, c, d k.
646 MARTIN LINDSAY
It is easily veried that these are indeed QS contraction cocycles; they are all
Markov-regular if V is, and the (c, d)-associated semigroup of V is precisely the
expectation semigroup of the (c, d)-associated cocycle of V .
Example 2.3. Let U = (U
x
= e
ixH
)
xR
be a strongly continuous one-parameter
unitary group on h and let (B
t
)
t0
be the standard Wiener process. Then
(V
t
F)() = U
Bt()
F(), F L
2
(; h) = h L
2
(),
denes a family of unitaries V = (V
t
)
t0
on h L
2
() or, under the Wiener-
Segal-Ito isomorphism, on h T (with k = C) which comprises a QS contraction
cocycle with expectation semigroup (e
tH
2
/2
)
t0
. This is the example highlighted
by Arveson; it is Markov-regular if and only if H is bounded, in other words U is
norm-continuous.
Processes other than Brownian motion may be used and multidimensions can
easily be incorporated. Here we are interested in the general structure of subclasses
of QS
c
C(h, k) and this involves quantum processes, as we have seen. For many
examples of cocycles arising from quantum optics and classical probability see [6,
18, 36], and references therein.
So far we have only considered contraction operator cocycles on a Hilbert space.
Given V QS
c
C(h, k),
k
V
t
: x B(h) V
t
(x I
J
)V
t
(h T), t R
+
, (2.1)
denes a family of completely positive contractions k
V
= (k
V
t
)
t0
enjoying the
cocycle relation
k
r+t
=
k
r
r
k
t
, r, t R
+
,
where
k
r
is the natural extension of k
r
to a map
Ran
r
= B(h) I
[0,r[
B(T
[r,[
) B(h) B(T
[0,r[
) B(T
[r,[
) = B(h T).
Note that the induced cocycle k
V
is unital if and only if V is coisometric, and
is homomorphic if V is partially isometric with V
t
V
t
I
h
B(T) (t 0), in
particular if V is isometric. Partially isometric and projection-valued QS cocycles
are analysed in [W
1,2
]. Note also that isometry for V is equivalent to unitality for
k
V
.
In generalising the above class of cocycles, in particular to non-inner cocycles,
it is convenient to drop the contractivity/boundedness condition. Fortunately the
cocycle identity is easily expressible in terms of certain slices of the maps (k
t
)
t0
as follows:
f,g
r+t
=
f,g
r
S
r
f,S
r
g
t
, r, t R
+
, (2.2)
where
f,g
t
(x) :=
_
id
B(h)
,
_
k
t
with = (f
[0,t[
) and = (g
[0,t[
), and this
requires only that each k
t
(x) is an operator whose domain includes suciently
many vectors of the form u(h) and that the resulting operators
f,g
t
(x) are
bounded.
A concrete operator space is a closed subspace V of B(H; K) for some Hilbert
spaces H and K; we speak of V being an operator space in B(H; K), or B(H; K)
being the ambient (full operator) space of V. The adjoint operator space V
is
the operator space x
: V
, x
(x)
. A
basic notion is that of complete boundedness for a linear map between operator
spaces : V W. This means that the norms of the matrix liftings
(n)
: [v
i,j
] [(v
i,j
)], (2.3)
in which M
n
(V) has the norm inherited from B(H; K), are uniformly bounded:
||
cb
:= sup
nN
|
(n)
| < .
There are several excellent texts on operator space theory ([5, 16, 42, 43]).
Now let V be an operator space in B(h; h
t
). For any Hilbert spaces h and h
t
,
the h-h
t
-matrix space over V is dened as follows:
V
M
B(h; h
t
) :=
_
T B(h h; h
t
h
t
) :
B(h;h
)
(id
B(h;h
)
)(T) V
_
([35]). Recalling the E-notations (1.1), a convenient characterisation, in terms of
any total subsets S and S
t
of h and h
t
, is as follows:
V
M
B(h; h
t
) =
_
T B(h h; h
t
h
t
) :
S,
S
E
TE
V
_
.
Thus V
M
B(h; h
t
) is an operator space in B(h h; h
t
h
t
).
Warning. For a C
)
: V
M
B(h; h
t
) W
M
B(h; h
t
)
satisfying
E
(
M
id
B(h;h
)
_
(T)E
= (E
TE
), T V
M
B(h; h
t
), h,
t
h
t
,
moreover,
M
id
B(h;h
)
is a completely bounded operator with cb-norm at most
||
cb
([35]). If B(h; h
t
) is nite dimensional then need only be bounded for
M
id
B(h;h
)
to exist; in this case the matrix lifting is a bounded operator. To
handle both situations conveniently we coined the term h-boundedness for a linear
map : V W, meaning bounded/completely bounded according as h is nite-
dimensional/innite-dimensional, and write h-B(V; W) for this class of maps.
A QS cocycle on V with noise dimension space k and exponential domain c
T
(where T is a total subset of k containing 0) is a family k = (k
t
)
t0
of linear maps
k
t
: V L(h c
T
; h
t
T)
= L
_
c
T
; L(h; h
t
T)
_
satisfying
E
(f)
k
t
(x)E
(g)
V,
the weak cocycle relation (2.2) and the adaptedness condition
u
t
(f), k
t
(x)u(g) = (f
[t,[
), (g
[t,[
)u
t
(f
[0,t[
), k
t
(x)u(g
[0,t[
)
(x V, u h, u
t
h
t
, g S
T
, f S, t R
+
). The collection of such cocycles
is denoted QSC(V: c
T
) (we shall impose the appropriate continuity condition in
648 MARTIN LINDSAY
t shortly). The subclass of cocycles k QSC(V : c
T
) having an adjoint cocycle
k
QSC(V
: c
T
), so that
E
(x
)E
= (E
k(x)E
, x V, c
T
,
t
c
T
, t R
+
,
is denoted QSC
(V: c
T
, c
T
). A cocycle k has h-bounded columns if, in the notation
k
t,
:= k
t
()E
: V L(h; h
t
T), it satises
k
t,
(V) V
M
[T, and
k
t,
is h-bounded V V
M
[T, t R
+
, c
T
.
We say that a cocycle k is h-bounded if it satises the stronger conditions
k
t
(V) V
M
B(T), and
k
t
is h-bounded V V
M
B(T), t R
+
.
Mapping cocycles k QSC(V: c
T
) also have associated semigroups:
T
c,d
:=
_
E
(c
[0,t[
)
k
t
()E
(d
[0,t[
)
_
t0
, c k, d T,
and, as for operator cocycles, the collection of associated semigroups determines
the cocycle thanks to Proposition 2.1. For a thorough investigation of the re-
constructability of cocycles from compatible families of semigroups see [38], which
was inspired by [1].
We now introduce the continuity condition which plays the central role in Sec-
tion 3. Let QSC
h-bHc
(V: c
T
) denote the class of cocycles k QSC(V: c
T
) having
h-bounded columns and such that
t k
t,
is Holder
1
2
-continuous R
+
h-B(V; V
M
[T) at t = 0,
and let QSC
h-bHc
(V: c
T
, c
T
) denote the subclass
_
k QSC
(V: c
T
, c
T
) QSC
h-bHc
(V: c
T
) : k
QSC
h-bHc
(V
: c
T
)
_
.
We refer to elements of QSC
k-bHc
(V : c
T
) as Holder cocycles. The reason for
highlighting this class of cocycle is hinted at in the observation (3.2) and fully
justied in Theorem 3.1. In this connection, note the elementary estimate
|W
c
t,(0)
W
c
r,(0)
| = |c|(t r)
1/2
+O(t r), c k, (2.4)
as t r in a bounded interval.
We again call cocycles k QSC(V : c
T
) Markov-regular (respectively, cb-
Markov-regular) when each associated semigroup is norm-continuous (respectively,
cb-norm-continuous), and we have the inclusions:
QSC
h-bHc
(V: c
T
) QSC
Mreg
(V: c
T
), and QSC
cbHc
(V: c
T
) QSC
cbMreg
(V: c
T
).
Note that the prescription
k
t
([u) = V
t
_
[u I
J
_
(2.5)
(u h, c, t R
+
) denes a completely contractive mapping cocycle k
QSC
cbHc
([h : c, c), as follows from (2.4).
3. Holder Continuous Cocycles
For this section we x an operator space V in B(h; h
t
). We rst discuss the
generation of QS cocycles on V by means of QS dierential equations, and then we
characterise various classes of cocycle so-generated. The main result (Theorem 3.1)
is the converse to the existence theorem established in [35]. Further details and
full proofs will apear in the forthcoming paper [28].
Let F(1
C
T;
k-B(V; V
M
[
(t) k
0
=
V
J
, (3.1)
has a unique (c
T
, c
T
)-weakly regular, (c
T
, c
T
)-weak solution, denoted k
.
Usually coecients of QS dierential equations are assumed to be linear:
L(
D;
k-B(V; V
M
[
u
t
(f), k
s
_
E
f(s)
g(s)
(x)
_
u(g)
_
(u
t
h
t
, f S
T
, x V, u h, g S
T
); (c
T
, c
T
)-weakly regular means
s E
k
s
()E
is locally bounded R
+
B(V), c
T
,
t
c
T
.
The unique solution enjoys the following further property:
k
QSC
k-bHc
(V: c
T
), (3.2)
and if (1
C
T) CB(V; V
M
[
k) then k
QSC
cbHc
(V: c
T
). Moreover, if
has an adjoint map
F
_
1
C
T
t
;
k-B(V
; V
M
[
k)
_
satisfying
E
c
d
(x
) =
_
E
c
(x)
_
, x V, c T, d T
t
,
then k
QSC
k-bHc
(V: c
T
, c
T
) with adjoint cocycle k
. In this case,
= [
1
C
]T
650 MARTIN LINDSAY
for a unique map in L
D;
k-B(V; V
M
[
k)
_
, the space of linear maps :
D
k-B(V; V
M
[
:
D
t
k-B(V
; V
M
[
k), where
t
= k
t
[
L
T
, t R
+
.
The above facts are essentially contained in [35], supplemented by [32]; the minor
modications needed for the present generality are explained in [28].
The situation is summarised as follows. The map
F
_
1
C
T;
k-B(V; V
M
[
k)
_
QSC
k-bHc
(V: c
T
), k
,
is injective and restricts to maps
F
_
1
C
T; CB(V; V
M
[
k)
_
QSC
cbHc
(V: c
T
)
L
D;
k-B(V; V
M
[
k)
_
QSC
k-bHc
(V: c
T
, c
T
)
L
D; CB(V; V
M
[
k)
_
QSC
cbHc
(V: c
T
, c
T
),
where D = Lin T and D
t
= Lin T
t
.
The following theorem extends Theorem 5.6 of [32], where V is assumed to be
nite dimensional.
Theorem 3.1. Each of the above four maps is bijective.
The idea of the proof is to dierentiate directly, in other words, to show that
for k QSC
k-bHc
(V: c
T
) and c T, the family
c,t
:= t
1
_
E
(0)
E
1
[0,t[
_
_
k
t,(c
[0,t[
)
V
(c
[0,t[
)
_
() (t > 0)
converges in
k-B(V; V
M
[
k) to a map
c
, as t 0
+
, in the point-W.O. topology:
,
c,t
(x) ,
c
(x), x V, h
t
, h,
and then to show that the resulting map F
_
1
C
T;
k-B(V; V
M
[
k)
_
satises k = k
have
the same associated semigroup generators. The rest then follows easily. Here
E
1
[0,t[
:= I
h
k
1
[0,t[
[ = I
h
_
I
k
1
[0,t[
[
_
B(h
t
T; h
t
k) = B(h
t
) B(T; k), (3.3)
by means of the inclusion k L
2
(R
+
) = L
2
(R
+
; k) T, as one-particle subspace,
and
V
(x) := x [ for c
T
, x V.
For bounded QS cocycles we can say more. Note rst the natural inclusions
h-B(V; V
M
B(
k)) L
_
k; h-B(V; V
M
[
k)
_
,
for any Hilbert space h, arising from the identication
c
= ()E
c
(c k).
Theorem 3.2. Let k QSC
k-bHc
(V : c
T
, c
T
) be bounded, with locally uniform
bounds. Then k = k
[
L
T
for a unique map
k-B(V; V
M
B(
k)). Moreover, if
k QSC
cbHc
(V: c
T
, c
T
) then CB(V; V
M
B(
k)).
DIFFERENTIATING QUANTUM STOCHASTIC COCYCLES 651
The idea of the proof of this is to let L
D;
k-B(V; V
M
[
k)
_
be the
mapping arising from Theorem 3.1, to set
= E
0
(), =
0
() and =
_
0
()
_
t
:= t
1
E
1
[0,t[
k
t
()E
1
[0,t[
(t > 0)
to a map B(V; V
M
B(k)), in the point-W.O. topology. The resulting map
:= [
], where =
V
k
, lies in B(V; V
M
B(
d
= E
c
d
(), d D, c D
t
;
the rest then follows easily.
The following two known results ([35, 34]) may be easily deduced from the above
theorem.
Corollary 3.3. Let k QSC
Mreg
(A : c) be completely positive and contractive,
on a C
.
This is proved by judicious use of the operator Schwarz inequality which reveals
that, under the given hypotheses, k QSC
cbHc
(A: c, c).
The precise form that must take for k
k) such that V = V
F
, that is V is the unique (weak, contractive) solution
of the QS dierential equation dV
t
= V
t
d
F
(t), V
0
= I
hJ
.
This is deduced from Theorem 3.1 via the correspondences (2.5). The quantum
Ito formula provides the following necessary and sucient condition on an operator
F B(h
k) for V
F
to be contractive:
q(F) 0 where q(F) := F
+F +F
F, (3.4)
equivalently, r(F) 0 where r(F) := F +F
+FF
k) : q(F) 0 QS
c
C
Mreg
(h, k), (3.5)
extending the bijection
F B(h
k) : q(F) = 0 = r(F) QS
u
C
Mreg
(h, k)
given in Theorem 1.1, where QS
u
C
Mreg
(h, k) denotes the collection of Markov-
regular unitary cocycles.
In the following section we describe a signicant extension of this result, to a
class of strongly continuous operator cocycles, which necessarily goes beyond the
realm of QS dierential equations.
652 MARTIN LINDSAY
4. Holomorphic Cocycles
For this section we again x Hilbert spaces h and k. We introduce the class
of holomorphic QS contraction cocycles on h with noise dimension space k, and
consider their innitesimal characterisation. We then describe the connection to
minimal quantum dynamical semigroups. The main results of this section are The-
orems 4.5 and 4.11. Further details and full proofs will appear in the forthcoming
paper [30].
Let X(h) denote the class of operators K on h which are densely dened and
dissipative: Reu, Ku 0 (u DomK) with no dissipative extension, in other
words maximal dissipative. It is well-known that such operators are precisely
the generators of contractive C
0
-semigroups on h ([15], Theorem 6.4). We are
interested in the collection X
hol
(h) of quadratic forms q on h which are accretive
and semisectorial:
Re q[u] 0 and
Imq[u]
C|u|
2
+
, u Q,
for some C 0, where
Q := Domq and |u|
2
+
:=
_
Re q[u] +|u|
2
_
1/2
,
as well as being densely dened and closed:
Q is dense in h and Q is complete w.r.t. the norm | |
+
.
The term semisectorial is nonstandard, but avoids the potentially confusing
term continuous favoured by some experts. Sectorial, which is standard, is
the strengthening [ Imq[u][ C Re q[u] (u Q) for some C 0.
For q X
hol
(h) there is a unique operator K on h satisfying
DomK =
_
u Q : the (conjugate-linear) fnl. v Q q(v, u) is bounded
_
,
u, Ku = q[u], u Q,
where Q = Domq and q( , ) is the sesquilinear form associated with q[] via
polarisation, moreover K X(h). The operators K X(h) that arise from forms
q X
hol
(h) are precisely those generators of contractive C
0
-semigroups which are
semisectorial: [ Imu, Ku[ C
_
Reu, Ku + |u|
2
_
1/2
(u DomK) for some
C 0. In this way we view X
hol
(h) as a subset of X(h).
By a holomorphic contraction semigroup on h we mean a contractive C
0
-semi-
group P = (P
t
)
t0
for which there is an angle ]0, /2] such that (e
t
P
t
)
t0
ex-
tends to a contraction-valued holomorphic function
B(h), where
denotes
the open sector of the complex plane z C 0 : [ arg z[ < .
Warning. Denitions in the literature vary. With our denition, the holomor-
phic contraction semigroups are precisely those whose generators are in X
hol
(h)
(see e.g. [40]).
Denition 4.1. A cocycle V QS
c
C(h, k) is holomorphic if its expectation semi-
group is holomorphic. Write QS
c
C
hol
(h, k) for the resulting class of cocycles.
For V QS
c
C
hol
(h, k) we write
V
for the form-generator of its expectation
semigroup and Q
V
for Dom
V
. The dual of a holomorphic QS contraction cocycle
DIFFERENTIATING QUANTUM STOCHASTIC COCYCLES 653
V is holomorphic with
V
= (
V
)
:= u
V
[u] and Q
V
= Q
V
.
Since norm-continuous semigroups are holomorphic, we have the inclusion
QS
c
C
hol
(h, k) QS
c
C
Mreg
(h, k). (4.1)
Katotype relative boundedness arguments for quadratic forms ([26]) yield our
rst two consequences of the holomorphic assumption.
Proposition 4.2. Let V QS
c
C
hol
(h, k). Then the following hold:
(a) All of the associated semigroups of V are holomorphic and their form-
generators have equal domain Q
V
.
(b) V is nonsingular, that is, with E
1
[0,t[
:=
_
E
1
[0,t[
_
B(h k; h T),
t
1
E
1
[0,t[
V
t
E
1
[0,t[
C
V
in the weak operator topology as t 0
+
,
for a contraction operator C
V
on h k.
Note that, in view of the identity R
t
E
1
[0,t[
= E
1
[0,t[
, it follows that, for V
QS
c
C
hol
(h, k),
C
V
= (C
V
)
.
We come now to the central denition of this section.
Notation. Let X
(4)
hol
(h, k) denote the collection of quadruples F = (, L,
L, C I),
where X
hol
(h), L and
L are operators from h to hk with domain Q := Dom
and C is a contraction in B(h k), satisfying
|F|
2
2 Re [] (4.2)
for the operator F and quadratic form on h (h k) = h
k given by
Dom = DomF := Q(h k), F :=
_
0 0
L C I
_
, and
[] := [u]
_
, Lu +
Lu, +, (C I)
_
for =
_
u
_
Dom.
Remark 4.3. (i) The relation (4.2) contains the inequalities
|Lu|
2
2 Re [u], |
Lu|
2
2 Re [u], u Q, (4.3)
(ii) If is bounded, so that Q = h, then L and
L are bounded, [] = , F
( h
k) where F :=
_
K M
L CI
in which M =
L
t
E
(0)
u
L
V
u (weakly), as t 0
+
, u Q
V
.
(b) The quadruple F
V
:= (
V
, L
V
,
L
V
, C
V
I) belongs to X
(4)
hol
(h, k).
(c) L
V
=
L
V
so
L
V
= L
V
and thus
F
V
=
_
(
V
)
,
L
V
, L
V
, (C
V
)
I
_
.
This is proved using abstract Ito integration in Fock space. Relative bound-
edness arguments with Yosida-type approximation, at the quadratic form level,
now combine with the Markov-regular theory and the semigroup characterisation
of QS cocycles to yield the central result.
Theorem 4.5. The map V F
V
is a bijection from QS
c
C
hol
(h, k) to X
(4)
hol
(h, k).
The inverse of this map is naturally denoted F V
F
and referred to as the
quantum stochastic generation map for holomorphic QS cocycles. It extends the
bijection F V
F
given in (3.5), by Remark 4.3 (ii).
What is achieved in the above result is the innitesimal characterisation of a
large class of QS contraction cocycles which includes all those whose expectation
semigroup is symmetric, as well as the Markov-regular cocycles previously char-
acterised. Whereas cocycles in the latter class are all governed by QS dierential
equations, it is not hard to construct examples in the former class which are not
so-governed. Here is the simplest example.
Example 4.6. Let A be a nonnegative, selfadjoint, unbounded operator on h
with dense range, let k = C, let P B(h) be the orthogonal projection with
range Cv for a vector v in h DomA, and set F =
_
|A |
2
/2, PA, PA, 0
_
.
Then F X
(4)
hol
(h, k), however the holomorphic QS contraction cocycle V
F
is not
governed by a QS dierential equation on any exponential domain. The reason for
this is that (PA)
= AP = 0[
v]
, so that the domain on which V satises a QS
dierential equation ([37], Theorem 4.2) fails to be dense in h.
We may now put some esh on Part (a) of Proposition 4.2.
Theorem 4.7. Let V QS
c
C
hol
(h, k), and let =
V
be the quadratic form on
h
c,d
[u] = (uc, u
d) +(c, d)|u|
2
, u Q
V
,
where (c, d) := (|c|
2
+|d|
2
)/2 c, d.
This chimes with Remark 4.3 (ii) and also with (4.1) in [37], the corresponding
identity for the case where V is governed by a QS dierential equation with
reasonable block matrix operator as coecient.
DIFFERENTIATING QUANTUM STOCHASTIC COCYCLES 655
The next result shows that to each holomorphic QS contraction cocycle on h
with noise dimension space k corresponds a holomorphic contraction semigroup on
h
k.
Theorem 4.8. Let F = (, L,
L, C I) X
(4)
hol
(h, k) and let be the associated
quadratic form on h
Lu, +, (C I)
_
, for =
_
u
_
Dom.
Then X
hol
(h
k).
Proof. Since Q is dense in h and
Re []
1
2
|F|
2
, Dom = Q(h k),
where F :=
_
0 0
L CI
_
_
Re [u] +|u|
2
+||
2
_
1/2
on Dom = Q (h k), and the fact that the accretive form is closed and
semisectorial itself.
In view of Remark 4.3 (iii), it follows that there is a bijective correspondence
between the collection QS
c
C
hol
(h, k) and a class of holomorphic contraction semi-
groups on h
k. The latter semigroups are quite dierent from the global semi-
groups associated with cocycles considered in [38], all of which are specically
coordinate-dependent and enjoy a Schur-action.
We have not so far discussed the question of coisometry and unitarity of holo-
morphic QS contraction cocycles V . These questions are of interest for their
(unitality, i.e. identity-preserving, and multiplicativity) implications for induced
cocycles on B(h), and von Neumann subalgebras thereof. It is not hard to verify
that a necessary condition for isometry is that equality holds in the inequality
governing F
V
:
2 Re
V
[] = |F
V
|
2
, Q(h k),
the holomorphic counterpart to the condition q(F) = 0 for Markov-regular cocy-
cles. Whilst this condition is also sucient in the case of Markov-regular cocycles,
sucient conditions are trickier in the holomorphic case. We quote a general result
from [38] (see also [30]). Recall the denition (2.1).
Theorem 4.9. Let V QS
c
C(h, k) and let T
c,d
: c, d k be the associated
semigroups of the induced cocycle k
V
on B(h). Then, for any total subset T of k
containing 0, the following are equivalent:
(i) V is coisometric (equivalently, k
V
is unital ).
(ii) T
c,c
is unital for all c T.
656 MARTIN LINDSAY
Thus, in case k is nite dimensional, (ii) need only involve the verication of
unitality of (1 + dimk) semigroups.
The above result frames a theorem of Fagnola to the eect that, in case V
satises a QS dierential equation on a core for the generator of its expectation
semigroup, unitality of just the expectation semigroup T
0,0
of k
V
suces ([18],
Theorem 5.23).
We next explain the connection to minimal quantum dynamical semigroups.
By a quantum dynamical semigroup on B(h) is meant a semigroup T = (T
t
)
t0
of
normal, completely positive contractions on B(h) which is continuous in the point-
ultraweak topology. Let X
(2)
(h, k) denote the collection of pairs (K, L) consisting
of a C
0
-semigroup generator K X(h) and operator L from h to h k satisfying
DomL DomK and |Lu|
2
2 Reu, Ku, u DomK.
For x B(h) dene the quadratic form /(x) = /
(K,L)
(x) on h by
Dom/(x) := DomK,
/(x)[u] := Ku, xu +u, xKu +Lu, (x I
k
)Lu.
A quantum dynamical semigroup T on B(h) is minimal for the pair (K, L)
X
(2)
(h, k) if it satises:
(i) for all x B(h), u DomK and t R
+
,
u, T
t
(x)u = u, xu +
_
t
0
ds /
(K,L)
(T
s
(x))[u];
(ii) if T
t
is another quantum dynamical semigroup satisfying (i) then
T
t
t
(x)) T
t
(x) for all x B(h)
+
, t R
+
.
Minimal quantum dynamical semigroups are the noncommutative counterparts of
Fellers minimal solutions of the Fokker-Planck equation.
Theorem 4.10 ([14]). Let (K, L) X
(2)
(h, k). Then there is a unique minimal
quantum dynamical semigroup T
(K,L)
associated with (K, L).
Note that if T
(K,L)
is conservative, that is unital: T
(K,L)
t
(I) = I (t R
+
), then
/
(K,L)
(I) = 0:
|Lu|
2
+ 2 Reu, Ku = 0, u DomK.
In the spirit of the inclusion X
hol
(h) X(h) we may consider the subclass
X
(2)
hol
(h, k) of X
(2)
(h, k) consisting of pairs (, L) where X
hol
(h) and (K, L)
X
(2)
(h, k), K being the operator corresponding to . In this case we write T
(,L)
for the associated minimal semigroup on B(h). In view of the relation
|Lu|
2
2 Re [u], u Dom,
for F = (, L,
L, C I) X
(4)
hol
(h, k), any such F truncates to a pair (, L)
X
(2)
hol
(h, k). Conversely, any pair (, L) X
(2)
hol
(h, k) dilates to a quadruple
F = (, L, C
L, C I) X
(4)
hol
(h, k)
DIFFERENTIATING QUANTUM STOCHASTIC COCYCLES 657
by choosing a unitary, or contraction, C in B(h; k); the choices C = I, respec-
tively C = 0, or C = I, being notable ones. The next result underpins these
considerations.
Theorem 4.11. Let V = V
F
for F = (, L,
L, C I) X
(4)
hol
(h, k). Then the
expectation semigroup of the induced cocycle on B(h) of the dual semigroup
V =
V
F
coincides with the minimal semigroup associated with the truncation (, L)
X
(2)
hol
(h, k) of F, equivalently,
E
_
V
t
(x I
J
)V
t
= T
(,L)
(x), x B(h), t R
+
.
This is the holomorphic counterpart to a corresponding result for QS contrac-
tion cocycles governed by a QS dierential equation ([18], Theorem 5.22). The
relationship between Fagnolas analysis which uses nite particle vectors (rather
than exponential vectors) in a crucial way, and builds on earlier work of Mohari,
Sinha and others and our analysis, remains an intriguing one.
To end I shall briey describe a new tool for the construction and analysis of
QS cocycles the QS Trotter product formula ([29]). In brief what this achieves
is the representation of the QS cocycle V
F
, where F has the form
F = F
1
F
2
:=
_
_
K
1
+K
2
M
1
M
2
L
1
C
1
I
1
L
2
C
2
I
2
_
_
B(h
k) = B(hhk
1
hk
2
)
with k = k
1
k
2
and I
i
denoting the identity operator on hk
i
(i = 1, 2), in terms
of the cocycles V
F1
and V
F2
, where
F
i
=
_
K
i
M
i
L
i
C
i
I
i
_
B(h
k
i
) = B(h h k
i
) for i = 1, 2.
It is easily veried that, if F
1
C
0
(h, k
1
) and F
2
C
0
(h, k
2
), then F C
0
(h, k).
In the quantum control literature the composition , known as the concatena-
tion product, has application to quantum networks, particularly in combination
with the series product ([20]). The latter corresponds to the Evans-Hudson per-
turbation formula ([17, 13, 19, 4]) specialised to the case where the free QS ow
is implemented by a (Markov-regular) QS unitary cocycle V ; the perturbed QS
ow is implemented by the unitary cocycle whose generator is (up to the choice
of parameterisation) the series product of the stochastic generator of V and the
perturbation coecient.
The QS Trotter product formula extends to both mapping cocycles and to
holomorphic contraction operator cocycles ([31]). In the latter case this is given
by
F = F
1
F
2
= (, L,
L, C I) X
(4)
hol
(h, k)
where, for (
i
, L
i
,
L
i
, C
i
I
i
) X
(4)
hol
(h, k
i
) (i = 1, 2),
=
1
+
2
, L =
_
L
1
L
2
_
,
L =
_
L
1
L
2
_
, C I =
_
C
1
I
1
C
2
I
2
_
,
658 MARTIN LINDSAY
yielding V
F
in terms of V
F1
and V
F2
. The only constraint for forming this concate-
nation product is that the intersection Dom
1
Dom
2
is dense in h since the sum
of two closed, accretive, semisectorial forms is closed, accretive and semisectorial,
and we have the following identities, by Pythagoras:
[] =
1
[
1
] +
2
[
2
] and |F| = |F
1
1
| +|F
2
2
|
where, for =
_
u
_
and =
_
1
2
_
,
1
:=
_
u
1
_
and
2
:=
_
u
2
_
. In the case of mapping
cocycles, the homomorphic property of cocycles so-constructed on an operator
algebra is investigated in [11].
In all of this work, the key ingredients are Trotter products of the associated
semigroups of the constituent cocycles.
Acknowledgment. The joint work with Kalyan Sinha described in Section 4 is
supported by the UKIERI Research Collaboration Network grant Quantum Prob-
ability, Noncommutative Geometry & Quantum Information.
References
1. Accardi, L. and Kozyrev, S. V.: On the structure of Markovian ows, Chaos Solitons and
Fractals 12 (2001) no. 14-15, 26392655.
2. Applebaum, D., Bhat, B. V. R., Kustermans, J., and Lindsay J. M.: Quantum Independent
Increment Processes, Vol. I: From Classical Probability to Quantum Stochastics, (eds. U.
Franz & M. Sch urmann), Lecture Notes in Mathematics 1865, Springer, Heidelberg, 2005.
3. Arveson, W.: Continuous analogues of Fock space, Mem. Amer. Math. Soc. 80 (1989), no.
409.
4. Belton, A. C. R., Lindsay, J. M., and Skalski, A. G.: Quantum Feynman-Kac perturbations,
in preparation.
5. Blecher, D. P. and Le Merdy, C.: Operator Algebras and their Modulesan Operator Space
Approach, Oxford University Press, Oxford, 2004.
6. Bouten, L., van Handel, R., and Silberfarb, L.: Approximation and limit theorems for quan-
tum stochastic models with unbounded coecients, J. Funct. Anal. 254 (2008) no. 12,
31233147.
7. Bradshaw, W. S.: Stochastic cocycles as a characterisation of quantum ows, Bull. Sci.
Math. (2) 116 (1992), 134.
8. Chebotarev, A. M. and Fagnola, F.: Sucient conditions for conservativity of minimal quan-
tum dynamical semigroups, J. Funct. Anal. 153 (1998), 382404.
9. Christensen, E. and Evans, D. E.: Cohomology of operator algebras and quantum dynamical
semigroups, J. London Math. Soc. (2) 20 (1979) no. 2, 358368.
10. Cockfoft, A. M. and Hudson, R. L.: Quantum mechanical Wiener processes, J. Multivariate
Analysis 7 (1977) no. 1, 107124.
11. Das, B., Goswami, D., and Sinha, K. B.: Trotter-Kato product formula for quantum sto-
chastic ows, Preprint, 2010.
12. Das, B. K. and Lindsay, J. M.: Quantum stochastic analysis in Banach space, in preparation.
13. Das, P. K. and Sinha, K. B.: Quantum ows with innite degrees of freedom and their
perturbations, in: Quantum Probability and Related Topics VI, (eds. L. Accardi & W. von
Waldenfels), 109123, World Scientic, Singapore, 1992.
14. Davies, E. B.: Quantum dynamical semigroups and the neutron diusion equation, Rep.
Math. Phys. 11 (1977) no. 2, 169188.
15. Davies, E. B.: One-Parameter Semigroups, London Mathematical Society Monographs 15,
Academic Press, London, 1980.
16. Eros, E. G. and Ruan, Z-J.: Operator Spaces, Oxford University Press, 2000.
17. Evans, M. P. and Hudson, R. L.: Perturbations of quantum diusions, J. London Math. Soc.
(2) 41 (1990) no. 2, 373384.
DIFFERENTIATING QUANTUM STOCHASTIC COCYCLES 659
18. Fagnola, F.: Quantum Markov semigroups and quantum ows, Proyecciones 18 (1999) no.
3, 1144.
19. Goswami, D., Lindsay, J. M., and Wills, S. J.: A stochastic Stinespring theorem, Math. Ann.
319 (2001) no. 4, 647673.
20. Gough, J. and James, M. R.: The series product and its application to quantum feedforward
and feedback networks, IEEE Trans. Automat. Control 54 (2009) no. 11, 25302544.
21. Hudson, R. L. and Lindsay, J. M.: A noncommutative martingale representation theorem
for non-Fock quantum Brownian motion, J. Funct. Anal. 61 (1985) no. 2, 202221.
22. Hudson, R. L. and Lindsay, J. M.: On characterizing quantum stochastic evolutions, Math.
Proc. Camb. Phil. Soc. 102 (1987) no. 2, 363369
23. Hudson, R. L., Linday, J. M., and Parthasarathy, K. R.: Stochastic integral representation
of some quantum martingales in Fock space, in: From Local Times to Global Geometry,
Control and Physics, (ed. K.D. Elworthy), Pitman Research Notes in Mathematics 150,
121131, Longman Sci. Tech., Harlow, 1986.
24. Hudson, R. L. and Parthasarathy, K. R.: Quantum Itos formula and stochastic evolutions,
Commun. Math. Phys. 93 (1984) no. 3, 301323.
25. Journe, J.-L.: Structure des cocycles markoviens sur lespace de Fock, Probab. Theory Related
Fields 75 (1987) no. 2, 291316.
26. Kato, T.: Perturbation Theory for Linear Operators, Springer, Berlin, 1966.
27. Lindsay, J. M.: Quantum stochastic analysis an introduction, in: [2].
28. Lindsay, J. M.: Holder-continuous quantum stochastic cocycles on operator spaces, in prepa-
ration.
29. Lindsay, J. M. and Sinha, K.B.: A quantum stochastic Lie-Trotter product formula, Indian
J. Pure Appl. Math. 41 (2010) no. 1, 313325.
30. Lindsay, J. M. and Sinha, K. B.: Holomorphic quantum stochastic contraction cocycles, in
preparation.
31. Lindsay, J. M. and Sinha, K. B.: Trotter product formulae for quantum stochastic cocycles,
in preparation.
32. Lindsay, J. M. and Skalski, A. G.: Quantum stochastic dierential equations, J. Math. Ann.
Appl. 330 (2007) no. 2, 10931114.
33. Lindsay, J. M. and Wills, S. J.: Existence, positivity and contractivity for quantum stochastic
ows with innite dimensional noise, Probab. Theory Related Fields 116 (2000) no. 4, 505
543.
34. Lindsay, J. M. and Wills, S. J.: Markovian cocycles on operator algebras, adapted to a Fock
ltration, J. Funct. Anal. 178 (2000) no. 2, 269305.
35. Lindsay, J. M. and Wills, S. J.: Existence of Feller cocycles on a C