arXiv:1104.

5242v2 [quant-ph] 9 Feb 2012

Open Quantum Systems.

An Introduction

Angel
Rivas1,2 and Susana F. Huelga1
1

Institut f
ur Theoretische Physik, Universitat Ulm, Ulm
D-89069, Germany.
Departamento de Fsica Teorica I, Universidad Complutense,
28040 Madrid, Spain.
Abstract
We revise fundamental concepts in the dynamics of open quantum
systems in the light of modern developments in the field. Our aim is to
present a unified approach to the quantum evolution of open systems
that incorporates the concepts and methods traditionally employed by
different communities. We present in some detail the mathematical
structure and the general properties of the dynamical maps underlying
open system dynamics. We also discuss the microscopic derivation of
dynamical equations, including both Markovian and non-Markovian
evolutions.

Contents
1 Introduction

2 Mathematical tools
2.1 Banach spaces, norms and linear operators
2.2 Exponential of an operator . . . . . . . . .
2.3 Semigroups of operators . . . . . . . . . .
2.3.1 Contraction semigroups . . . . . .
2.4 Evolution families . . . . . . . . . . . . . .
3 Time evolution in closed quantum systems

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

4
4
6
8
11
14
18

4 Time evolution in open quantum systems

4.1 Dynamical maps . . . . . . . . . . . . . .
4.2 Universal dynamical maps . . . . . . . . .
4.3 Universal dynamical maps as contractions
4.4 The inverse of a universal dynamical map .
4.5 Temporal continuity. Markovian evolutions

.
.
.
.
.

21
23
26
29
31
33

.
.
.
.

35
35
37
41
45

.
.
.
.
.
.
.

51
52
56
57
61
61
65
67

.
.
.
.
.
.
.

71
72
73
75
77
77
80

case
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .

82
83
84
86

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

5 Quantum Markov process: mathematical structure

5.1 Classical Markovian processes . . . . . . . . . . . . .
5.2 Quantum Markov evolution as a differential equation
5.3 Kossakowski conditions . . . . . . . . . . . . . . . . .
5.4 Steady states of homogeneous Markov processes . . .

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

6 Microscopic description: Markovian case

6.1 Nakajima-Zwanzig equation . . . . . . . . . . . . . . . . . .
6.2 Weak coupling limit . . . . . . . . . . . . . . . . . . . . . . .
6.2.1 Decomposition of V . . . . . . . . . . . . . . . . . . .
6.2.2 Correlation functions . . . . . . . . . . . . . . . . . .
6.2.3 Davies theorem . . . . . . . . . . . . . . . . . . . . .
6.2.4 Explicit expressions for k and Sk . . . . . . . . . .
6.2.5 Typical examples . . . . . . . . . . . . . . . . . . . .
6.2.6 Some remarks on the secular approximation and positivity preserving requirement . . . . . . . . . . . . .
6.2.7 Failure of the assumption of factorized dynamics . . .
6.2.8 Steady state properties . . . . . . . . . . . . . . . . .
6.3 Singular coupling limit . . . . . . . . . . . . . . . . . . . . .
6.4 Extensions of the weak coupling limit . . . . . . . . . . . . .
6.4.1 Weak coupling for multipartite systems . . . . . . . .
6.4.2 Weak coupling under external driving . . . . . . . . .
7 Microscopic description: non-Markovian
7.1 Integro-differential models . . . . . . . .
7.2 Time-convolutionless forms . . . . . . . .
7.3 Dynamical coarse graining method . . .
8 Conclusion

90

Introduction

To write an introduction to the dynamics of open quantum systems may seem

at first a complicated, albeit perhaps unnecessary, task. On the one hand, the
2

field is quite broad and encompasses many different topics which are covered
by several books and reviews [137]. On the other hand, the approach taken
to study the dynamics of an open quantum system can be quite different
depending of the specific system being analyzed; for instance, references in
quantum optics (e.g. [1,10,11,15,17,26]) use tools and approximations which
are usually quite different from the studies focused on condensed matter or
open systems of relevance in chemical physics (e.g. [8, 13, 27, 30, 33]). Moreover, very rigorous studies by the mathematical physics community were carried out in the 1970s, usually from a statistical physics point of view. More
recently, and mainly motivated by the development of quantum information
science, there has been a strong revival in the study of open many body systems aimed at further understanding the impact of decoherence phenomena
on quantum information protocols [38].
This plethora of results and approaches can be quite confusing to novices
in the field. Even within the more experienced scientists, discussions on
fundamental properties (e.g. the concept of Markovianity) have often proven
not to be straightforward when researches from different communities are
involved.
Our motivation in writing this work has been to try to put black on white
the results of several thoughts and discussions on this topic with scientists of
different backgrounds during the last three years, as well as putting several
concepts in the context of modern developments in the field. Given the extension of the topic and the availability of many excellent reviews covering
different aspects of the physics of open quantum systems, we will focus here
on one specific issue and that is embedding the theory as usually explained
in the quantum optical literature within the mathematical core developed in
the mathematical physics community. Finally, we make a connection with
the current view point of the dynamics of open quantum systems in the light
of quantum information science, given a different, complementary, perspective to the meaning of fundamental concepts such that complete positivity,
dynamical semigroups, Markovianity, master equations, etc.
In order to do this, we will restrict the proof of important mathematical results to the case of finite dimensional systems; otherwise this study
would be much more extensive and potentially interested people who are not
expert in functional analysis or operators algebras, could not easily benefit from it. However, since the most important application of the theory of
open quantum system nowadays is arguably focused on the field of controlled
quantum technologies, including quantum computation, quantum communication, quantum metrology, etc., where the most useful systems are finite (or
special cases of infinite systems), the formal loss of generality is justified in
this case. The necessary background for this work is therefore reduced to
3

familiarity with the basics of quantum mechanics and quantum information

theory.
Furthermore, we have tried to incorporate some novel concepts and techniques developed in the last few years to address old problems such that those
exemplified by the use of dynamical maps, Markovian and non-Markovian
evolutions or microscopic derivations of reduced dynamics. We have organized the presentation as follows. In section 2 we introduce the mathematical
tools required in order to ensure a minimum degree of self-consistency in the
study; a mathematical experienced reader may skip this part. In section 3 we
review very succinctly the concept of closed quantum system and discuss
its time evolution. We explain in section 4 the general features of dynamics
of open quantum systems and its mathematical properties. Section 5 deals
with the mathematical structure of a very special important dynamics in
open quantum systems, which are the Markovian quantum processes. The
microscopic derivation of Markovian quantum dynamics is explained in section 6, whereas some methods for obtaining non-Markovian dynamics are
briefly introduced in section 7. Finally, we conclude the review by making a
summary of the main ideas we have discussed, and the topics that we have
left out, in the section 8.

2
2.1

Mathematical tools
Banach spaces, norms and linear operators

In different parts within this work, we will make use of some properties of
Banach spaces. For our interest it is sufficient to restrict our analysis to finite
dimensional spaces. In what follows we revise a few basic concepts [3941].
Definition 2.1. A Banach space B is a complete linear space with a norm
k k.
Recall that a space is complete if every Cauchy sequence converges in it.
An example of Banach space is provided by the set of self-adjoint trace-class
linear operators on a Hilbert space H. This is, the self-adjoint operators
whose trace norm is finite k k1 ,

kAk1 = Tr A A = Tr A2 , A B.
Now we focus our attention onto possible linear transformations on a
Banach space, that we will denote by T (B) : B 7 B. The set of all of them
forms the dual space B .
4

Proposition 2.1. The dual space of B, B is a (finite) Banach space with

the induced norm
kT k =

kT (x)k
= sup kT (x)k,
xB,x6=0 kxk
xB,kxk=1
sup

T B .

Proof : The proof is basic.

Apart from the triangle inequality, the induced norm fulfills some other
useful inequalities. For example, immediately from its definition one obtains
that
kT (x)k kT kkxk, x B.
(1)
Moreover,

Proposition 2.2. Let T1 and T2 be two linear operators in B , then

kT1 T2 k kT1 kkT2 k.

(2)

Proof. This is just a consequence of the inequality (1); indeed,

kT1 T2 k =

sup
xB,kxk=1

kT1 T2 (x)k

sup
xB,kxk=1

kT1 kkT2 (x)k = kT1 kkT2 k.

Q.E.D.

The next concept will appear quite often in the forthcoming sections.
Definition 2.2. A linear operator T on a Banach space B, is said to be a
contraction if
kT (x)k kxk, x B,
this is kT k 1.

Finally, we briefly revise some additional basic concepts that will be featured in subsequent sections.
Definition 2.3. For any linear operator T on a (finite) Banach space B, its
resolvent operator is defined as
RT () = (1 T )1 , C.
It is said that C belongs to the (point) spectrum of T , (T ), if RT ()
does not exist; otherwise C belongs to the resolvent set of T , (T ).
As a result (T ) and (T ) are said to be complementary sets.
The numbers (T ) are called eigenvalues of T , and each element in
the kernel x ker(1T ), which is obviously different from {0} as (1T )1
does not exist, is called the eigenvector associated with the eigenvalue .
5

2.2

Exponential of an operator

The operation of exponentiation of an operator is of special importance in

relation to linear differential equations.
Definition 2.4. The exponential of a linear operator L on a finite Banach
space is defined as

X
Ln
L
e =
.
(3)
n!
n=0
The following proposition shows that this definition is indeed meaningful.
Proposition 2.3. The series defining eL is absolutely convergent.
Proof. The following chain of inequalities holds

X
n
L
X kLn k X kLkn

L

= ekLk .
ke k =


n!
n!
n!
n=0

n=0

n=0

Since ekLk is a numerical series which is convergent, the norm of eL is upper

bounded and therefore the series converges absolutely.
Q.E.D.
Proposition 2.4. If L1 and L2 commute with each other, then e(L1 +L2 ) =
eL1 eL2 = eL2 eL1 .
Proof. By using the binomial formula as in the case of a numerical series, we
obtain

n  
X
(L1 + L2 )n X 1 X n k nk
(L1 +L2 )
e
=
=
L1 L2
n!
n!
k
n=0
n=0
=

X
n
X
n=0 k=0

k=0
X

1
Lk Lnk =
k!(n k)! 1 2

p=0 q=0

1 p q
L L = eL1 eL2 = eL2 eL1 .
p!q! 1 2

Q.E.D.
The above result does not hold when L1 and L2 are not commuting operators, in that case there is a formula which is sometimes useful.
Theorem 2.1 (Lie-Trotter product formula). Let L1 and L2 be linear operators on a finite Banach space, then:
 L1 L2 n
e(L1 +L2 ) = lim e n e n
.
n

Proof. We present a two-step proof inspired by [42]. First let us prove that
for any two linear operators A and B the following identity holds:
n

A B =

n1
X
k=0

Ak (A B)B nk1.

(4)

We observe that, in the end, we are only adding and subtracting the products
of powers of A and B in the right hand side of (4), so that
An B n = An + (An1 B An1 B) B n
= An + (An1 B An1 B + An2 B 2 An2 B 2 ) B n =
= An + (An1 B An1 B + An2 B 2 An2 B 2 + . . .
+ AB n1 AB n1 ) B n .
Now we can add terms with positive sign to find:
An + An1 B + . . . + AB n1 =

n1
X

Ak+1 B nk1,

k=0

and all terms with negative sign to obtain:

n1

B . . . AB

n1

B =

n1
X

Ak B nk .

k=0

Finally, by adding both contributions one gets that

n

A B =

n1
X

k+1

nk1

k=0

A B

nk

n1
X
k=0

Ak (A B)B nk1,

as we wanted to prove.
Now, we denote
Xn = e(L1 +L2 )/n ,

Yn = eL1 /n eL2 /n ,

and take A = Xn and B = Yn in formula (4),

Xnn

Ynn

n1
X
k=0

Xnk (Xn Yn )Ynnk1 .

On the one hand,

kXnk k = k e(L1 +L2 )/n

k

k ke(L1 +L2 )/n kk ekL1 +L2 kk/n

e(kL1 k+kL2 k)k/n ekL1 k+kL2 k ,

7

k n.

Similarly, one proves that

kYnk k ekL1 k+kL2 k ,

k n.

So we get
kXnn

Ynn k

n1
X
k=0

n1
X
k=0

kXnk (Xn

Yn )Ynnk1 k

n1
X
k=0

kXnk kkXn Yn kkYnnk1 k

kXn Yn ke2(kL1 k+kL2 k) = nkXn Yn ke2(kL1 k+kL2 k) .

On the other hand, by expanding the exponential we find that


k X
 k  j

X
1 L1 + L2
1
L1
L2


lim nkXn Yn k = lim n

n
n

k!
n
k!j!
n
n
k=0
k,j=0

 
1
1
= 0.
lim n
[L
,
L
]
+
O
2
1
n 2n2
n3

Therefore we finally conclude that:

n


lim eL1 +L2 eL1 /n eL2 /n = lim kXnn Ynn k = 0.
n

Q.E.D.

Using a similar procedure one can also show that for a sum of N operators,
the following identity holds:
!n
N
Y
PN
Lk
.
(5)
e k=1 Lk = lim
en
n

2.3

k=1

Semigroups of operators

Definition 2.5 (Semigroup). A family of linear operators Tt (t 0) on a

finite Banach Space forms a one-parameter semigroup if
1. Tt Ts = Tt+s , t, s
2. T0 = 1.

Definition 2.6 (Uniformly Continuous Semigroup). A one-parameter semigroup Tt is said to be uniformly continuous if the map
t 7 Tt
is continuous, this is, limts kTt Ts k = 0, s. This kind of continuity is
normally referred to as continuity in the uniform operator topology [39],
and it is sufficient to analyze the finite dimensional case, which is the focus
of this work.
Theorem 2.2. If Tt forms a uniformly continuous one-parameter semigroup,
then the map t 7 Tt is differentiable, and the derivative of Tt is given by
dTt
= LTt ,
dt
with L =

dTt
| .
dt t=0

Proof. Since Tt is uniformly continuous on t, the function V (t) defined by

Z t
V (t) =
Ts ds, t 0,
0

is differentiable with dVdt(t) = Tt (to justify this, it is enough to consider

the well-known arguments used for the case of R, see for example [43], but
substituting the absolute value by the concept of norm). In particular,

V (t)
dV (t) 
V (t) V (0)
lim
= T0 = 1,
= lim
=
t0
t0
t
t
dt t=0

this implies that there exist some t0 > 0 small enough such that V (t0 ) is
invertible. Then,
Z t0
Z t0
1
1
1
Tt = V (t0 )V (t0 )Tt = V (t0 )
Ts Tt ds = V (t0 )
Tt+s ds
0
0
Z t+t0
1
= V (t0 )
Ts ds = V 1 (t0 ) [V (t + t0 ) V (t)] ,
t

since the difference of differentiable functions is differentiable, Tt is differentiable. Moreover, its derivative is
dTt
Tt+h Tt
Th 1
= lim
= lim
Tt = LTt .
h0
h0
dt
h
h
Q.E.D.
9

Theorem 2.3. The exponential operator T (t) = eLt is the only solution of
the differential problem
 dT (t)
= LT (t), t R+ ,
dt
T (0) = 1.
Proof. From the definition of exponential given by Eq. (3), it is clear that
T (0) = 1, and because of the commutativity of the exponents we find


deLt
eL(t+h) eLt
eLh 1 Lt
e .
= lim
= lim
h0
h0
dt
h
h
By expanding the exponential one obtains
eLh 1
1 + Lh + O(h2 ) 1
= lim
= L.
h0
h0
h
h
lim

At this point, it just remains to prove uniqueness. Let us consider another

function S(t) and assume that it also satisfies the differential problem. Then
we define
Q(s) = T (s)S(t s), for t s 0,
for some fixed t > 0, so Q(s) is differentiable with derivative

dQ(s)
= LT (s)S(t s) T (t)LS(t s) = LT (s)S(t s) LT (t)S(t s) = 0.
ds
Q(s) is therefore a constant function. In particular Q(s) = Q(0) for any
t s 0, which implies that
T (t) = T (t)S(t t) = Q(t) = Q(0) = S(t).
Q.E.D.
Corollary 2.1. Any uniformly continuous semigroup can be written in the
form Tt = T (t) = eLt , where L is called the generator of the semigroup and
it is the only solution to the differential problem
 dTt
= LTt , t R+
dt
(6)
T0 = 1.
Given the omnipresence of differential equations describing time evolution
in physics, it is now evident why semigroups are important. If we can extend
the domain of t to negative values, then Tt forms a one-parameter group
whose inverses are given by Tt , so that Tt Tt = 1 for all t R.
There is a special class of semigroups which is important for our proposes,
and that will be the subject of the next section.
10

2.3.1

Contraction semigroups

Definition 2.7. A one-parameter semigroup satisfying kTt k 1 for every

t 0, is called a contraction semigroup.
Now one question arises, namely, which properties should L fulfill to
generate a contraction semigroup? This is an intricate question in general
that is treated with more detail in functional analysis textbooks such as [44
46]. For finite dimensional Banach spaces, the proofs of the main theorems
can be drastically simplified, and they are presented in the following. The
main condition is based in properties of the resolvent set and the resolvent
operator.
Theorem 2.4 (Hille-Yoshida). A necessary and sufficient condition for a
linear operator L to generate a contraction semigroup is that
1. {, Re > 0} (L).
2. kRL ()k (Re )1 .
Proof. Let L be the generator of a contraction semigroup and a complex
number. For a b we define the operator
Z b
L(a, b) =
e(1+L)t dt.
a

Since the exponential series converges uniformly (to see this one uses again
tools of the elementary analysis, specifically the M-test of Weierstrass [43], on
the Banach space B) it is possible to integrate term by term in the previous
definition
Z b

Z bX

X
(1 + L)n tn
(1 + L)n
n
L(a, b) =
dt =
t dt
n!
n!
a n=0
a
n=0



X
an+1
(1 + L)n bn+1
.

=
n!
n
+
1
n
+
1
n=0
Thus by multiplying L(a, b) by (1 L) we get


(1 L)L(a, b) = L(a, b)(1 L) = e(1+L)a e(1+L)b .
Now assume Re > 0 and take the limit a 0, b , then as
lim ke(1+L)b k = lim ke1b eLb k

11

lim ke1b kkeLb k

lim ke1b k = 0,

and
lim e1a eLa = 1,

a0

we obtain
(1 L)

Z

 Z
e dt =

1t Lt

e
0


e dt (1 L) = 1.

1t Lt

So if Re > 0 then (L) and from the definition of the resolvent operator
we conclude
Z
RL () =
e1t eLt dt.
0

Moreover,

Z
Z
Z


1t Lt
1t Lt

kRL ()k =
e
e dt
ke
e kdt
ke1t kkeLt kdt
0
0
Z
Z 0
Z
1

ke1t kdt =
|et |dt =
e Re()t dt =
.
Re()
0
0
0
Conversely, suppose that L satisfies the above conditions 1 and 2. Let
be some real number in the resolvent set of L, which means it is also a real
positive number. We define the operator
L = LRL (),
and it turns out that L L as . Indeed, it is enough to show that
RL () 1. Since (0, ) (L) there is no problem with the resolvent
operator taking the limit, so by writing RL () = LRL () + 1 we obtain
kLk
= 0,

lim kRL () 1k = lim kLRL ()k lim kLkkRL ()k lim

where we have used 2. Moreover,

2 R ()1]t
L

keL t k = keLRL ()t k = ke[

2 R ()t
L

k et ke

2 kRL ()kt

et e

et et = 1,

and therefore
eLt = lim eL t

is a contraction semigroup.
Q.E.D.

12

Apart from imposing conditions such that L generates a contraction semigroup, this theorem is of vital importance in general operator theory on arbitrary dimensional Banach spaces. If the conditions of the theorem 2.4 are
fulfilled then L generates a well-defined semigroup. Note that for general
operators L the exponential cannot be rigourously defined as a power series, and more complicated tools are needed. On the other hand, to apply
this theorem directly can be complicated in many cases, so one needs more
manageable equivalent conditions.
Definition 2.8. A semi-inner product is an operation in which for each pair
x1 , x2 of elements of a Banach space B there is an associated complex number
[x1 , x2 ], such that
[x1 , x2 + x3 ] = [x1 , x2 ] + [x1 , x3 ],
[x1 , x1 ] = kx1 k2 ,
|[x1 , x2 ]| kx1 kkx2 k.

(7)
(8)
(9)

for any x1 , x2 , x3 B and , C.

Definition 2.9. A linear operator L on B is called dissipative if
Re[x, Lx] 0 x B.

(10)

Theorem 2.5 (Lumer-Phillips). A necessary and sufficient condition for a

linear operator L on B (of finite dimension) to be the generator of a contraction semigroup is that L be dissipative.
Proof. Suppose that {Tt , t 0} is a contraction semigroup, then just by
using the properties of a semi-inner product we have that:
Re[x, (Tt 1)x] = Re[x, Tt x] kxk2 |[x, Tt x]| kxk2
kxkkTt xk kxk2 kxk2 kTt k kxk2 0,
as kTt k 1. So L is dissipative
1
Re[x, (Tt 1)x] 0.
t0 t

Re[x, Lx] = lim

Conversely, let L be dissipative. Considering a point within its spectrum

(L) and a corresponding eigenvector x ker(1 L). Since [x, (1
L)x] = 0 we have
Re[x, (1 L)x] = Re()kxk2 Re[x, Lx] = 0,
13

given that Re[x, Lx] 0, Re() 0. Thus any complex number Re() > 0
is in the resolvent of L. Moreover, let (L), again from the properties of
a semi-inner product we have that
Re()kxk2 = Re()[x, x] Re([x, x] [x, Lx]) = Re[x, (1 L)x]
|[x, (1 L)x]| kxkk(1 L)xk,
so k(1 L)xk Re()kxk, x 6= 0. Then, by setting x = (1 L)
x in the
expression for the norm of RL ():
kRL ()k =

kRL ()xk
=
kxk
xB,x6=0
sup

k
xk
k(1 L)
xk
x
B,
xker(1L)
/
sup

1
k
xk
=
,
Re()k
xk
Re()

and therefore L satisfies the conditions of the theorem 2.4.

Q.E.D.

2.4

Evolution families

As we have seen semigroups arise naturally in the context of problems involving linear differential equations, as exemplified by Eq. (6), with timeindependent generators L. However some situations in physics require solving
time-inhomogeneous differential problems,
 dx
= L(t)x, x B
dt
x(t0 ) = x0 ,
where L(t) is a time-dependent linear operator. Given that the differential
equation is linear, its solution must depend linearly on the initial conditions,
so we can write
x(t) = T(t,t0 ) x(t0 ),
(11)
where T(t,t0 ) is a linear operator often referred to as the evolution operator.
Obviously, T(t0 ,t0 ) = 1 and the composition of two consecutive evolution
operators is well defined; since for t2 t1 t0
x(t1 ) = T(t1 ,t0 ) x(t0 ) and x(t2 ) = T(t2 ,t1 ) x(t1 ).
The evolution operator between t2 and t0 is given by
T(t2 ,t0 ) = T(t2 ,t1 ) T(t1 ,t0 ) ,
14

(12)

and the two-parameter family of operators T(t,s) fulfills the properties

T(t,s) = T(t,r) T(r,s) ,
T(s,s) = 1.

trs

(13)

These operators are sometimes called evolution families, propagators, or fundamental solutions. One important fact to realize is that if the generator
is time-independent, L, the evolution family is reduced to a one-parameter
semigroup in the time difference T(t,s) = T =ts . However, in contrast to general semigroups, it is not enough (even in finite dimension) to require that
the map (t, s) 7 T(t,s) be continuous to ensure that it is differentiable [47], although in practice we usually assume that the temporal evolution is smooth
enough and families fulfilling Eq. (13) will also considered to be differentiable. Under these conditions we can formulate the following result.
Theorem 2.6. A differentiable family T(t,s) obeying the condition (13) is the
only solution to the differential problems
 dT(t,s)
= L(t)T(t,s) , t s
dt
T(s,s) = 1,
and

dT(t,s)
ds

T(s,s)

= T(t,s) L(t),
= 1.

ts

Proof. If the evolution family T(t,s) is differentiable

dT(t,s)
T(t+h,s) T(t,s)
T(t+h,t) 1
= lim
= lim
T(t,s) = L(t)T(t,s) .
h0
h0
dt
h
h
Analogously one proves that
dT(t,s)
= T(t,s) L(t).
ds
To show that this solution is unique, as in theorem 2.3, let us assume that
there exist some other S(t,s) solving both differential problems. Then, for
every fixed t and s, we define now Q(r) = T(t,r) S(r,s) for t r s. It is
= 0 and therefore
immediate to show that dQ(r)
dr
T(t,s) = Q(s) = Q(t) = S(t,s) .
Q.E.D.

15

Again opposed to the case of semigroups, to give a general expression for

an evolution family T(t,s) in terms of the generator L(t) is complicated. To
illustrate why let us define first the concept of time-ordering operator.
Definition 2.10. The time-ordered operator T of a product of two operators
L(t1 )L(t2 ) is defined by
T L(t1 )L(t2 ) = (t1 t2 )L(t1 )L(t2 ) + (t2 t1 )L(t2 )L(t1 ),
where (x) is the Heaviside step function. Similarly, for a product of three
operators L(t1 )L(t2 )L(t3 ), we have
T L(t1 )L(t2 )L(t3 ) =
+
+
+
+
+

(t1 t2 )(t2 t3 )L(t1 )L(t2 )L(t3 )

(t1 t3 )(t3 t2 )L(t1 )L(t3 )L(t2 )
(t2 t1 )(t1 t3 )L(t2 )L(t1 )L(t3 )
(t2 t3 )(t3 t1 )L(t2 )L(t3 )L(t1 )
(t3 t1 )(t1 t2 )L(t3 )L(t1 )L(t2 )
(t3 t2 )(t2 t1 )L(t3 )L(t2 )L(t1 ),

so that for a product of k operators L(t1 ) L(tk ) we can write:

X
T L(t1 ) L(tk ) =
[t(1) t(2) ] [t(k1) t(k) ]L[t(1) ] L[t(k) ],

(14)
where is a permutation of k indexes and the sum extends over all k! different
permutations.
Theorem 2.7 (Dyson Series). Let the generator L(t ) to be bounded in the
interval [t, s], i.e. kL(t )k < , t [t, s]. Then the evolution family T(t,s)
admits the following series representation
T (t, s) = 1 +

Z tZ
X

m=1

t1

tm1

L(t1 ) L(tm )dtm dt1 ,

(15)

where t t1 . . . tm s, and it can be written symbolically as

Rt

T (t, s) = T e

L(t )dt

Z t
Z t

X
1
1+

T L(t1 ) L(tm )dt1 dtm .

m! s
s
m=1
(16)

16

Proof. To prove that expression (15) can be formally written as (16) note
that by introducing the time-ordering definition Eq. (14), the k th term of
this series can be written as
Z
Z t
1 t

T L(t1 ) L(tk )dt1 dtk

k! s
s
Z Z
Z t(k1)
1 X t t(1)
=

L[t(1) ] L[t(k) ]dt1 dtk

k! s s
s
Z t Z t1
Z tk1
L(t1 ) L(tk )dtk dt1 ,
=

where in the last step we have used that t t(1) . . . t(k) s, and that
the name of the variables does not matter, so the sum over is the addition
of the same integral k! times.
To prove Eq. (15), we need first of all to show that the series is convergent,
but this is simple because by taking norms in (16) we get the upper bound
!m
"
#

X
|t s|m
kT (t, s)k
sup kL(t )k
= exp sup kL(t )k|t s| ,

m!
t [s,t]
t [s,t]
m=0
which is convergent as a result of L(t ) being bounded in [t, s]. Finally, the
fact that (15) is the solution of the differential problem is easy to verify by
differentiating the series term by term since it converges uniformly and so
does its derivative (to see this one uses again the M-test of Weierstrass [43]
on the Banach space B).
Q.E.D.
This result is the well-known Dyson expansion, which is widely used for
instance in scattering theory. We have proved its convergence for finite dimensional systems and bounded generator. This is normally difficult to prove
in the infinite dimensional case, but the expression is still used in this context
in a formal sense.
Apart from the Dyson expansion, there is an approximation formula for
evolution families which will be useful in this work.
Theorem 2.8 (Time-splitting formula). The evolution family T(t,s) can be
given by the formula
T(t,s) =

lim

max |tj+1 tj |0

0
Y

e(tj+1 tj )L(tj )

j=n1

for t = tn . . . t0 = s,

(17)

where L(tj ) is the generator evaluated in the instantaneous time tj (note the
descending order in the product symbol).
17

Proof. The idea for this proof is taken from [48], more rigorous and general
proofs can be found in [44, 45]. Given that the product in (17) is already
temporally ordered, nothing happens if we introduce the temporal-ordering
operator
T(t,s) =
=

lim

max |tj+1 tj |0

lim

0
Y

j=n1
Pn1

Te

max |tj+1 tj |0

j=0

e(tj+1 tj )L(tj ) ,
(tj+1 tj )L(tj )

where one realizes that the time-ordering operator is then making the work
of ordering the terms in the exponential series in the bottom term in order to
obtain the same series expansion as in the top term. Now it only remains to
recognize the Riemann sum [43] in the exponent to arrive at the expression
Rt

T(t,s) = T e
which is Eq. (16).

L(t )dt

,
Q.E.D.

In analogy to the case of contraction semigroups, one can define contraction evolution families. However, since we have now two parameters, there
is not a unique possible definition, as illustrated below.
Definition 2.11. An evolution family T(t,s) is called contractive if kT(t,s) k 1
for every t and every s such that t s.
Definition 2.12. An evolution family T(t,s) is called eventually contractive
if there exist a s0 such that kT(t,s0 ) k 1 for every t s0 .
It is clear that contractive families are also eventually contractive, but
in this second case there exist a privileged initial time s0 , such that one can
have kT(t,s) k > 1 for some s 6= s0 . This difference is relevant and will allow
us later on to discriminate universal dynamical maps given by differential
equations which are Markovian from those which are not.

Time evolution in closed quantum systems

Probably time evolution of physical systems is the main factor in order to

understand their nature and properties. In classical systems time evolution
is usually formulated in terms of differential equations (i.e. Euler-Lagranges
equations, Hamiltons equations, Liouville equation, etc.), which can present

18

very different characteristics depending on which physical system they correspond to. From the beginning of the quantum theory, physicists have been
often trying to translate the methods which were useful in the classical case to
the quantum one, so was that Erwin Schrodinger obtained the first quantum
evolution equation in 1926 [49].
This equation, called Schrodingers equation since then, describes the
behavior of an isolated or closed quantum system, that is, by definition, a
system which does not interchanges information (i.e. energy and/or matter
[50]) with another system. So if our isolated system is in some pure state
|(t)i H at time t, where H denotes the Hilbert space of the system,
the time evolution of this state (between two consecutive measurements) is
according to
d
i
|(t)i = H(t)|(t)i,
(18)
dt
~
where H(t) is the Hamiltonian operator of the system. In the case of mixed
states (t) the last equation naturally induces
d(t)
i
= [H(t), (t)],
(19)
dt
~
sometimes called von Neumann or Liouville-von Neumann equation, mainly
in the context of statistical physics. From now on we consider units where
~ = 1.
There are several ways to justify the Schrodinger equation, some of them
related with heuristic approaches or variational principles, however in all of
them it is necessary to make some extra assumptions apart from the postulates about states and observables; for that reason the Schrodinger equation
can be taken as another postulate. Of course the ultimate reason for its
validity is that the experimental tests are in accordance with it [5153]. In
that sense, the Schrodinger equation is for quantum mechanics as fundamental as the Maxwells equations for electromagnetism or the Newtons Laws
for classical mechanics.
An important property of Schrodinger equation is that it does not change
the norm of the states, indeed




d|(t)i
dh(t)|
d
|(t)i + h(t)|
=
h(t)|(t)i =
dt
dt
dt
= ih(t)|H (t)|(t)i ih(t)|H(t)|(t)i = 0, (20)

as the Hamiltonian is self-adjoint, H(t) = H (t). In view of the fact that the
Schrodinger equation is linear, its solution is given by an evolution family
U(t, r),
|(t)i = U(t, t0 )|(t0 )i,
(21)
19

where U(t0 , t0 ) = 1.
The equation (20) implies that the evolution operator is an isometry (i.e.
norm-preserving map), which means it is a unitary operator in case of finite
dimensional systems. In infinite dimensional systems one has to prove that
U(t, t0 ) maps the whole H onto H in order to exclude partial isometries, this
is easy to verify from the composition law of evolution families [54] and then
for any general Hilbert space
U (t, t0 )U(t, t0 ) = U(t, t0 )U (t, t0 ) = 1 U 1 = U .

(22)

On the other hand, according to these definitions for pure states, the
evolution of some density matrix (t) is
(t1 ) = U(t1 , t0 )(t0 )U (t1 , t0 ).

(23)

We can also rewrite this relation as

(t1 ) = U(t1 ,t0 ) [(t0 )],

(24)

where U(t1 ,t0 ) is a linear (unitary [55]) operator acting as

U(t1 ,t0 ) [] = U(t1 , t0 )[]U (t1 , t0 ).

(25)

Of course the specific form of the evolution operator in terms of the

Hamiltonian depends on the properties of the Hamiltonian itself. In the simplest case in which H is time-independent (conservative system), the formal
solution of the Schodinger equation is straightforwardly obtained as
U(t, t0 ) = ei(tt0 )H/~ .

(26)

When H is time-dependent (non-conservative system) we can formally write

the evolution as a Dyson expansion,
Rt

U(t, t0 ) = T e

t0

H(t )dt

(27)

Note now that if H(t) is self-adjoint, H(t) is also self-adjoint, so in

this case we have physical inverses for every evolution. In fact when the
Hamiltonian is time-independent, U(t1 , t0 ) = U(t1 t0 ). So the evolution operator is not only a one-parameter semigroup, but a one-parameter
group, with an inverse for every element U 1 ( ) = U( ) = U ( ), such
that U 1 ( )U( ) = U( )U 1 ( ) = 1. Of course in case of time-dependent
Hamiltonians, every member of the evolution family U(t, s) has the inverse
U 1 (t, s) = U(s, t) = U (t, s), being this one also unitary.
20

If we look into the equation (24), we can identify as a member of the

Banach space of the set of trace-class self-adjoint operators, and U(t1 ,t0 ) as
some linear operator acting on this Banach space. Then everything considered for pure states concerning the form of the evolution operator is also
applicable for general mixed states, actually the general solution to the equation (19) can be written as
Rt

(t) = U(t,t0 ) (t0 ) = T e

t0

Lt dt

(t0 ),

(28)

where the generator is given by the commutator Lt () = i[H(t), ]. This is

sometimes called the Liouvillian, and by taking the derivative one immediately arrives to (19).

Time evolution in open quantum systems

Consider now the case where the total Hilbert space is decomposed into two
parts, in the form H = HA HB , each subspace corresponding to a certain
quantum system; the question to answer is then, what properties does the
time evolution of each subsystem have?.
First of all, one must establish the relation between the total density
matrix H and the density matrix of a subsystem, say A, which is given
by the partial trace over the other subsystem B:
A = TrB ().

(29)

There are several ways to justify this; see for example [38, 56]. Let us focus
first in the description of a measurement of the system HA viewed from the
whole system H = HA HB . For the system HA a measurement is given by
a set of self-adjoint positive operators {Mk } each of them associated with a
possible result k. Suppose that we are blind to the system B, and we see the
state of system A as some A . Then the probability to get the result k when
we make the measurement {Mk } will be
pA (k) = Tr(Mk A ).
Now one argues that if we are viewing only the part A through the measurement Mk , we are in fact observers of an extended system H through some
k such that, for physical consistency, for any state of the
measurement M
composed system compatible with A which is seen from A (see figure 1),
the following holds
k ).
pA (k) = Tr(M
21

 k is a genuine
But assuming this is true for every state A , we assume that M
measurement operator on H, which implicitly implies that is independent of
the state of the whole system on which we are making the measurement.
Now consider the special case where the state of the whole system has the
form = A B , then the above equation is clearly satisfied with the choice
k = Mk 1,
M
pA (k) = Tr(Mk A ) = Tr[(Mk 1)A B ] = Tr(Mk A ) Tr(B ) = Tr(Mk A ).
k is independent of the
Actually this is the only possible choice such that M
particular states A and B (and so of ). Indeed, imagine that there exist
k , independent of , which is also a genuine measure on
another solution, M
H, then from the linearity of the trace we would get
k ) Tr[(M
k Mk 1)] = 0,
pA (k) = Tr(Mk A ) = Tr(Mk 1) = Tr(M
for all A and so for all . This means
k Mk 1)] = 0,
Tr[(M
for any self-adjoint operator , therefore as the set of self-adjoint operators
k Mk 1 = 0 and then M
k = Mk 1.
forms a Banach space this implies M

Figure 1: On the left side we perform a measure over the combined system A
k , but we have disconnected the information
and B, which we denote by M
provided by B. If one assume that this situation is equivalent to a measure
k = Mk 1
Mk only the system A, this is illustrated on the right side, then M
(see discussion on the text).
Once established this relation, one can wonder about the connection between A and . If {|jA i|B i} is a basis of H then
X
hjA |hB |(Mk 1)|jA i|B i
pA (k) = Tr[(Mk 1)] =
j

X
j,

hjA |Mk hB ||B i|jA i = Tr(Mk A ),

22

hence A is univocally given by the partial trace

X
A =
hB ||B i TrB ().

One important property of the partial trace operation is that, even if

= |ih| was pure, A can be mixed, this occurs if |i is entangled.
Given the total density matrix (t0 ), by taking the partial trace in (23),
the state of A at time t1 , A (t1 ), is given by
A (t1 ) = TrB [U(t1 , t0 )(t0 )U (t1 , t0 )].

(30)

Provided that the total evolution is not factorized U(t1 , t0 ) = UA (t1 , t0 )

UB (t1 , t0 ), both quantum subsystems A and B are interchanging information
with each other (they are interacting), and so they are open quantum systems.

4.1

Dynamical maps

Let us set A as our system to study. A natural task arising from the equation
(30) is trying to rewrite it as a dynamical map acting on HA which connects
the states of the subsystem A at times t0 and t1 :
E(t1 ,t0 ) : A (t0 ) A (t1 ).
The problem with this map is that in general it does not only depend on the
global evolution operator U(t1 , t0 ) and on the properties of the subsystem B,
but also on the system A itself.
In order to clarify this, let us write the total initial state as the sum of
two contributions [57]:
(t0 ) = A (t0 ) B (t0 ) + corr (t0 ),

(31)

where the term corr (t0 ) is not a quantum state; satisfies

TrA [corr (t0 )] = TrB [corr (t0 )] = 0,
and contains all correlations (classical as well as quantum) between the two
subsystems. The substitution of (31) in (30) provides
A (t1 ) = TrB {U(t1 , t0 )[A (t0 ) B (t0 ) + corr (t0 )]U (t1 , t0 )}
X
=
i TrB {U(t1 , t0 )[A (t0 ) |i ihi |]U (t1 , t0 )}
i

+ TrB [U(t1 , t0 )corr (t0 )U (t1 , t0 )]

X
=
K (t1 , t0 )A (t0 )K (t1 , t0 ) + (t1 , t0 )

E(t1 ,t0 ) [A (t0 )],

23

(32)


where = {i, j}, K{i,j}(t1 , t0 ) = i h
Pj |U(t1 , t0 )|i i and we have used
the spectral decomposition of B (t0 ) = i i |i ihi |. Apparently the operators K (t1 , t0 ) depend only on the global unitary evolution and on the
initial state of the subsystem B, but the inhomogeneous part (t1 , t0 ) =
TrB [U(t1 , t0 )corr (t0 )U (t1 , t0 )] may no longer be independent of A because
of the correlation term corr [58]. That is, because of the positivity requirement of the total density matrix (t0 ), given some A (t0 ) and B (t0 ) not any
corr (t0 ) is allowed such that (t0 ) given by (31) is a positive operator; in the
same way, given A (t0 ) and corr (t0 ) not any B (t0 ) is allowed.
A paradigmatic example of this is the case of reduced pure states. Then
the following theorem is fulfilled.
Theorem 4.1. Let one of the reduced states of a bipartite quantum system
be pure, say A = |iA h|, then corr = 0.
Proof. Indeed, let us consider first that =
P |ih| is pure as well. By
using the Schmidt decomposition of |i =
i i |ui , vi i, where for all i in
the sum i > 0, and {|ui i} and {|vi i} are orthonormal basis of HA and HB
respectively, we obtain
"
#
X
X
A = TrB
i j |ui , vi ihuj , vj | =
2i |ui ihui |.
(33)
i,j

It is elementary to show that a pure state cannot be written as non-trivial

convex combination of another two states [59, 60]; then one coefficient, say
i must be equal to one, and i = 0 for all i 6= i . This implies that has
the product form
= |ui ihui | |vi ihvi | |iA h| |iB h|.
If is mixed, it can be written as convex combination of pure states
X
=
pk |k ihk |,
k

(k)

if A = TrA (|k ihk |), the state of A is

X
(k)
A =
pk A ,
k

but as A is pure, we have only two possibilities, either some pk = 1 and

(k)
pk = 0 for k 6= k or A is the same state A for every k. In the first
24

case is pure obviously, for the second option

P let us introduce the Schmidt
decomposition again for each state |k i = i i,k |ui,k , vi,k i:
X
(k)
A = A = TrA (|k ihk |) =
2i,k |ui,k ihui,k |, k
i

so both i,k and |ui,k i can be taken to be independent of k, and

X
A =
2i |ui ihui |.
i

This equation is actually the same as (33) and so one i is equal to one, and
i = 0 for all i 6= i . With these requirements we get |k i = |ui , vi ,k i, and
finally
=

X
k

pk |k ihk | =

X
k

pk |ui , vi ,k ihui , vi ,k |
= |ui ihui |

X
k

pk |vi ,k ihvi ,k | |iA h| B ,

and hence corr = 0.

Q.E.D.
As a result, the positivity requirement forces each term in the dynamical
map to be interconnected and dependent on the state upon which they act;
this means that a dynamical map with some values of K (t1 , t0 ) and (t1 , t0 )
may describe a physical evolution for some states A (t0 ) and an unphysical
evolution for others.
Given this fact, a possible approach to the dynamics of open quantum
systems is by means of the study of dynamical maps and their positivity
domains [6166], however that will not be our aim here.
On the other hand, the following interesting result given independently
by D. Salgado et al. [67] and D. M. Tong et al. [68] provides an equivalent
parametrization of dynamical maps.
Theorem 4.2. Any kind of time evolution of a quantum state A can always
be written in the form
X
A (t1 ) =
K (t1 , t0 , A ) A (t0 )K (t1 , t0 , A ) ,
(34)

where K (t1 , t0 , A ) are operators which depend on the state A at time t0 .

25

Proof. We give here a very simple proof based on an idea presented in [64].
Let A (t0 ) and A (t1 ) be the states of a quantum system at times t0 and t1
respectively. Consider formally the product A (t0 ) A (t1 ) HA HA , and
a unitary operation USWAP which interchanges the states in a tensor product

USWAP |1 i |2 i = |2 i |1 i, so USWAP (A B)USWAP

= B A. It is
evident that the time evolution between A (t0 ) and A (t1 ) can be written as
i
h

= A (t1 ),
E(t1 ,t0 ) [A (t0 )] = Tr2 USWAP A (t0 ) A (t1 )USWAP

where Tr2 denotes the partial trace with respect to the second member of
the composed state. By taking the spectral decomposition of A (t1 ) in the
central term of the above equation we obtain an expression of the form (34).
Q.E.D.
Note that the decomposition of (34) is clearly not unique. For more details
see references [67] and [68].
It will be convenient for us to understand the dependence on A of the
operators K (t1 , t0 , A ) in (34) as a result of limiting the action of (32). That
is, the above theorem asserts that restricting the action of any dynamical map
(32) from its positivity domain to a small enough set of states, actually to
one state A if necessary, is equivalent to another dynamical map without
inhomogeneous term (t1 , t0 ). However this fact can also be visualized as
a by-product of the specific nonlinear features of a dynamical map as (32)
[69, 70].
As one can already guess, the absence of the inhomogeneous part will
play a key role in the dynamics. Before we move on to examine this in the
next section, it is worth remarking that C. A. Rodrguez et al. [71] have
found that for some classes of initial separable states (t0 ) (the ones which
have vanishing quantum discord), the dynamical map (34) is the same for a
set of commuting states of A . So we can remove the dependence on A of
K (t1 , t0 , A ) when applying (34) inside of the set of states commuting with
A .

4.2

Universal dynamical maps

We shall call a universal dynamical map (UDM) to a dynamical map which is

independent of the state it acts upon, and so it describes a plausible physical
evolution for any state A . In particular, since for pure A (t0 ), corr (t0 ) = 0,
according to (32) the most general form of a UDM is given by
X
A (t1 ) = E(t1 ,t0 ) [A (t0 )] =
K (t1 , t0 ) A (t0 )K (t1 , t0 ) ,
(35)

26

in addition, the equality

X

K (t1 , t0 ) K (t1 , t0 ) = 1

(36)

is fulfilled because Tr[A (t1 )] = 1 for any initial state A (t0 ). Compare (34)
and (35).
An important result is the following (see figure 2):
Theorem 4.3. A dynamical map is a UDM if and only if it is induced from
an extended system with the initial condition (t0 ) = A (t0 ) B (t0 ) where
B (t0 ) is fixed for any A (t0 ).
Proof. Given such an initial condition the result is elementary to check.
Conversely, given a UDM (35) one can always consider the extended state
(t0 ) = A (t0 ) |iB h|, with |iB HB fixed for every state A (t0 ), and
so K (t1 , t0 ) = B h |U(t1 , t0 )|iB , where | iB is a basis of HB . Since
K (t1 , t0 ) only fixes a few partial elements of U(t1 , t0 ) and the requirement
(36) is consistent with the unitarity condition
X

B h|U (t1 , t0 )| iB h |U(t1 , t0 )|iB = B h|U (t1 , t0 )U(t1 , t0 )|iB

= B h|1 1|iB = 1,

such a unitary operator U(t1 , t0 ) always exists.

Q.E.D.

Figure 2: Schematic illustration of the result presented in theorem 4.3.

Moreover UDMs possess other interesting
properties. For instance
P
P they
(i)
are linear maps, that is for A (t0 ) = i i A (t0 ) (with i 0 and
i i = 1),
a straightforward consequence from definition of UDM is
X
(i)
E(t1 ,t0 ) [A (t0 )] =
i E(t1 ,t0 ) [A (t0 )].
i

27

In addition, UDMs are positive operations (i.e. the image of a positive operator is also a positive operator).
Furthermore, a UDM holds another remarkable mathematical property.
Consider the following Gedankenexperiment; imagine that the system AB
is actually part of another larger system ABW , where the W component
interact neither with A nor with B, and so it is hidden from our eyes, playing
the role of a simple bystander of the motion of A and B. Then let us consider
the global dynamics of the three subsystems. Since W is disconnected from A
and B, the unitary evolution of the whole system is of the form UAB (t1 , t0 )
UW (t1 , t0 ), where UW (t1 , t0 ) is the unitary evolution of the subsystem W . If
the initial state is ABW (t0 ) then

ABW (t1 ) = UAB (t1 , t0 ) UW (t1 , t0 )ABW (t0 )UAB

(t1 , t0 ) UW
(t1 , t0 ).

Note that by taking the partial trace with respect to W we get

h
i

AB (t1 ) = UAB (t1 , t0 ) TrW UW (t1 , t0 )ABW (t0 )UW (t1 , t0 ) UAB
(t1 , t0 )
h
i

= UAB (t1 , t0 ) TrW UW

(t1 , t0 )UW (t1 , t0 )ABW (t0 ) UAB
(t1 , t0 )

= UAB (t1 , t0 )AB (t0 )UAB

(t1 , t0 ),

so indeed the presence of W does not perturb the dynamics of A and B.

Since we assume that the evolution of A is described by a UDM, according
to the theorem 4.3 the initial condition must be AB (t0 ) = A (t0 ) B (t0 ),
and then the most general initial condition of the global system which is
compatible with that has to be of the form ABW (t0 ) = AW (t0 ) B (t0 ),
where B (t0 ) is fixed for any AW (t0 ). Now let us address our attention to
the dynamics of the subsystem AW :
AW (t1 ) = TrB [UAB (t1 , t0 ) UW (t1 , t0 )AW (t0 ) B (t0 )

UAB
(t1 , t0 )

UW
(t1 , t0 )

by using the spectral decomposition of B (t0 ), as we have already done a

couple of times, one obtains
X

AW (t1 ) =
K (t1 , t0 ) UW (t1 , t0 )AW (t0 )K (t1 , t0 ) UW
(t1 , t0 )

= E(t1 ,t0 ) U(t1 ,t0 ) [AW (t0 )] .

Here E(t1 ,t0 ) is a dynamical map on the subsystem A and the operator U(t1 ,t0 ) [] =

UW (t1 , t0 )[]UW
(t1 , t0 ) is the unitary evolution of W . As we see, the dynamical map on the system AW is the tensor product of each individual dynamics.
28

So if the evolution of a system is given by a UDM, E(t1 ,t0 ) , then for any unitary evolution U(t1 ,t0 ) of any dimension, E(t1 ,t0 ) U(t1 ,t0 ) is also a UDM; and
in particular is a positive-preserving operation. By writing



E(t1 ,t0 ) U(t1 ,t0 ) = E(t1 ,t0 ) 1 1 U(t1 ,t0 ) ,

as 1 U(t1 ,t0 ) is a unitary operator, the condition for positivity-preserving

asserts that E(t1 ,t0 ) 1 is positive. Linear maps fulfilling this property are
called completely positive maps [72].
Since we can never dismiss the possible existence of some extra system
W , out of our control, a UDM must always be completely positive. This may
be seen directly from the decomposition (35). The reverse statement, i.e. any
completely positive map can be written as (35), is a famous result proved
by Kraus [7]; thus the formula (35) is referred as Kraus decomposition of a
completely positive map. Therefore another definition of a UDM is a (tracepreserving) linear map which is completely positive, see for instance [21, 38].
Complete positivity is a stronger requirement than positivity, the best
known example of a map which is positive but not completely positive is the
transposition map (see for example [38]), so it cannot represent a UDM. It
is appropriate to stress that, despite the crystal-clear reasoning about the
necessity of complete positivity in the evolution of a quantum system, it is
only required if the evolution is given by a UDM, a general dynamical map
may not possess this property [6164, 67, 69, 70]. Indeed, if (t1 , t0 ) 6= 0,
E(t1 ,t0 ) given by (32) is not positive for every state, so certainly it cannot be
completely positive. In addition note that if ABW (t0 ) 6= AW (t0 ) B (t0 )
we can no longer express the dynamical map on AW as E(t1 ,t0 ) U(t1 ,t0 ) , so
the physical requirement for complete positivity is missed.

4.3

Universal dynamical maps as contractions

Let us consider now the Banach space B of the trace-class operators with
the trace norm (see comments to the definition 2.1). Then, as we know, an
operator B is a quantum state if it is positive-semidefinite, this is
h||i 0,

|i H,

and it has trace one, which is equal to the trace norm because the positivity
of its spectrum,
B is a quantum state 0 and kk1 = 1.

+ B, actually the Banach space

We denote the set of quantum states by B
+ ; in other words, any other
B the smallest linear space which contains B
+ also contains B.
linear space containing B
29

We now focus our attention in possible linear transformations on the

Banach space E(B) : B 7 B, and in particular, as we have said in section
2.1, the set of these transformations is the dual Banach space B , in this
case with the induced norm


kE()k1
kEk1 = sup
.
kk1
B,6=0
Particularly we are interested in linear maps which leave invariant the
+ , i.e. they connect any quantum state with another quantum
subset B
state, as a UDM does. The following theorem holds
+ if and only if it
Theorem 4.4. A linear map E B leaves invariant B
preserves the trace and is a contraction on B
kEk1 1.

+,
Proof. Assume that E B leaves invariant the set of quantum states B
then it is obvious that it preserves the trace; moreover, it implies that for
every positive the trace norm is preserved, so kE()k1 = kk1 . Let B
+ , then by the spectral decomposition we can split = + ,
but
/B
where
(
P
+ = j |j ihj |, for i 0,
P
= j |j ihj |, for j < 0,
here j are the eigenvalues and |j ihj | the corresponding eigenvectors. Note
that + and are both positive operators. Because
|j ihj | are orthogonal
P
projections we get (in other words as kk1 = j |j |):
kk1 = k + k1 + k k1 ,

but then
kE()k1 = kE( + )E( )k1 kE( + )k+kE( )k1 = k + k1 +k k1 = kk1 ,
so E is a contraction.
Conversely, assume that E is a contraction and preserves the trace, then
+ we have the next chain of inequalities:
for B
kk1 = Tr() = Tr[E()] kE()k1 kk1 ,

+ if and only if
so kE()k1 = kk1 and Tr[E()] = kE()k1 . Since B
+
for any B
+.
kk1 = Tr() = 1, the last equality implies that E() B
Q.E.D.
This theorem was first proven by A. Kosakowski in the references [73,74],
later on, M. B. Ruskai also analyzed the necessary condition in [75].
30

4.4

The inverse of a universal dynamical map

Let us come back to the scenario of dynamical maps. Given a UDM, E(t1 ,t0 ) ,
one can wonder whether there is another UDM describing the evolution in
the time reversed sense E(t0 ,t1 ) in such a way that
E(t0 ,t1 ) E(t1 ,t0 ) = E(t11 ,t0 ) E(t1 ,t0 ) = 1.
Note, first of all, that it is quite easy to write a UDM which is not bijective
(for instance,
take {|n i} to be an orthonormal basis of the space, then define:
P
E() = n |ihn ||n ih| = Tr()|ih|, being |i an arbitrary fixed state),
and so the inverse does not always exist. But even if it is mathematically
possible to invert a map, we are going to see that in general the inverse is
not another UDM.
Theorem 4.5. A UDM, E(t1 ,t0 ) , can be inverted by another UDM if and only
if it is unitary E(t1 ,t0 ) = U(t1 ,t0 ) .
Proof. The proof is easy (assuming Wigners theorem) but lengthy. Let us
remove the temporal references to make the notation less cumbersome, so we
write E E(t1 ,t0 ) . Since we have shown that UDMs are contractions on B,
kE()k1 kk1 ,

B.

Let us suppose that there exists a UDM, E 1, which is the inverse of E; then
by applying the above inequality twice
kk1 = kE 1 E()k1 kE()k1 kk1 ,
and we conclude that
kE()k1 = kk1 ,

B.

(37)

Next, note that such an invertible E would connect pure stares with pure
states. Indeed, suppose that |ih| is a pure state, and E(|ih|) is not,
then we can write
E(|ih|) = p1 + (1 p)2 ,

1 > p > 0,

for some 1 6= 2 6= E(|ih|). But after taking the inverse this equation
yields
|ih| = pE 1 (1 ) + (1 p)E 1 (2 ),

and E 1 is a (bijective) UDM, E 1 (1 ) 6= E 1(2 ) are quantum states; so such

a decomposition is impossible as |ih| is pure by hypothesis; and a pure
31

state cannot be decomposed as a non-trivial convex combination of another

two states.
Now consider = 12 (|1 ih1 | |2 ih2 |) for two arbitrary pure states
|1 ih1 | and |2 ih2 |. We can calculate its eigenvalues and eigenvectors
by restring the computation to the two dimensional subspace spanned by
|1 i and |2 i; because any other state |i can be decomposed as a linear
combination of |1 i, |2 i and | i, such that | i = 0. Thus the eigenvalue
equation reads
(|1 i + |2 i) = (|1 i + |2 i),
for some , and . By taking the scalar product with respect to h1 | and
h2 |, and solving the two simultaneous equations one finds
=
so the trace norm kk1 =

1p
1 |h2 |1 i|2 ,
2

|j | turns out to be

kk1 =

1 |h2 |1 i|2 .

(38)

However, since E connects pure states with pure states,

E() =

1
1
[E(|1 ih1 |) E(|2 ih2 |)] = (|1 ih1 | |2 ih2 |),
2
2

and therefore from equations (37) and (38) we conclude that

|h2 |1 i| = |h2 |1 i|.
Invoking Wigners theorem [54, 59], the transformation E have to be implemented by
E = V V ,
where V is a unitary or anti-unitary operator. Finally since the effect of an
anti-unitary operator involves the complex conjugation [54], that is equivalent
to the transposition of the elements of a trace-class self-adjoint operator,
which is not a completely positive map; we conclude that V has to be unitary,
E(t1 ,t0 ) = U(t1 ,t0 ) .
Q.E.D.
From another perspective, the connection between the failure of an inverse for
a UDM to exist just denotes the irreversibility of the universal open quantum
systems dynamics.

32

4.5

Temporal continuity. Markovian evolutions

One important problem of the dynamical maps given by UDM is related with
their continuity in time. This can be formulated as follows: assume that we
know the evolution of a system between t0 and t1 , E(t1 ,t0 ) , and between t1 and
a posterior time t2 , E(t2 ,t1 ) . From the continuity of time, one could expect that
the evolution between t0 and t2 was given by the composition of applications,
E(t2 ,t0 ) = E(t2 ,t1 ) E(t1 ,t0 ) . However, the exact meaning of this equation has to
be analyzed carefully.
Indeed, let the total initial state be (t0 ) = A (t0 ) B (t0 ), and the unitary evolution operators U(t2 , t1 ), U(t1 , t0 ) and U(t2 , t0 ) = U(t2 , t1 )U(t1 , t0 ).
Then the evolution of the subsystem A between t0 and any subsequent t is
clearly given by,
E(t1 ,t0 ) [A (t0 )] = TrB [U(t1 , t0 )A (t0 ) B (t0 )U (t1 , t0 )]
X
=
K (t1 , t0 ) A (t0 )K (t1 , t0 ) ,

and
E(t2 ,t0 ) [A (t0 )] = TrB [U(t2 , t0 )A (t0 ) B (t0 )U (t2 , t0 )]
X
=
K (t2 , t0 ) A (t0 )K (t2 , t0 ) .

The problem arises with E(t2 ,t1 ) :

E(t2 ,t1 ) [A (t1 )] = TrB [U(t2 , t1 )(t1 )U (t2 , t1 )]
X
=
K (t2 , t1 , A ) A (t1 )K (t2 , t1 , A ) ,

being (t1 ) = U(t1 , t0 )[A (t0 ) B (t0 )]U (t1 , t0 ). Since (t1 ) is not a tensor
product in general, and it depends on what initial states were taken for both
subsystems, E(t2 ,t1 ) has not the form of an UDM (see figure 3).
Let us to illustrate this situation from another point of view. If the
application E(t1 ,t0 ) is bijective it seems that we can overcome the problem by
defining
E(t2 ,t1 ) = E(t2 ,t0 ) E(t11 ,t0 ) ,
(39)

which trivially satisfies the relation E(t2 ,t0 ) = E(t2 ,t1 ) E(t1 ,t0 ) . However, as we
have just seen, E(t11 ,t0 ) is not completely positive unless it is unitary, so in
general the evolution E(t2 ,t1 ) is still not a UDM [76].
The same idea can also be used to define dynamical maps beyond UDM
[77]. That is, given the global system in an arbitrary general state (t0 ),
33

Figure 3: General situation describing the dynamics of open quantum systems in terms of a UDM. At time t0 the total state is assumed to be factorized, (t0 ) = A (t0 ) B (t0 ), and the dynamical maps from this point are
UDMs. However it is not possible to define a UDM from some intermediate
time t1 > t0 , because generally at this time the total state can no longer be
written as a tensor product.
if this is not a tensor product, we know that at least in some past time
tinitial , before the subsystems started interacting, the state was factorized
(tinitial ) = A (tinitial ) B (tinitial ). If we manage to find a unitary operation
U 1 (t0 , tinitial ) which maps (t0 ) to the state (tinitial ) = A (tinitial )B (tinitial )
for every reduced state A (t0 ) compatible with (t0 ), having B (tinitial ) the
same for all A (t0 ), we can describe any posterior evolution from t0 to t1 as
the composition of the inverse map and the UDM from tinitial to t1 as in (39).
However to find such an unitary operator can be complicated, and it will not
always exist.
On the other hand, as a difference with the dynamics of closed systems,
it is clear that these difficulties arising from the continuity of time for UDMs
make it impossible to formulate the general dynamics of open quantum systems by means of differential equations which generate contractive families
on B (i.e. families of UDMs). However, sometimes one can write down a
differential form for the evolution from a fixed time origin t0 , which is only a
valid UDM to describe the evolution from that time t0 . In other words, the
differential equation generates an eventually contractive family from t0 on B
(see definition 2.12).
To illustrate this, consider a bijective UDM, E(t,t0 ) , then by differentiation
(without the aim to be very rigorous here)
dE(t,t0 )
dE(t,t0 ) 1
dA (t)
=
[A (t0 )] =
E(t,t0 ) [A (t)] = Lt [A (t)],
dt
dt
dt
dE

(40)

1
0)
E(t,t
. This is the ultimate idea behind the time convoluwhere Lt = (t,t
dt
0)
tionless method which will be briefly described in section 7.2. If the initial

34

condition is given at t0 , then, by construction, the solution to the above

equation is A (t) = E(t,t0 ) [A (t0 )]. Nevertheless, there is not certainty that
the evolution from any other initial time is a UDM. Indeed, as we may guess,
it is not difficult to write solutions given the initial condition (t1 ) (t1 > t0 )
as (39).
Definition 4.1. We will say that a quantum systems undergoes a Markovian
evolution if it is described by a contractive evolution family on B and thus
we recover the composition law for the UDMs
E(t2 ,t0 ) = E(t2 ,t1 ) E(t1 ,t0 ) ,

(41)

which is the quantum analogue to the Chapman-Kolmogorov equation in

classical Markovian process (see next section 5.1). This property is sometimes
also referred as divisibility condition.
Typically, the dynamics of an open quantum system is not Markovian,
because as we have already said both systems A and B develop correlations
corr during their evolution and so the reduced dynamical map is not a UDM
for some intermediate starting time t1 . However if the term which goes with
corr does not affect so much the dynamics, a Markovian model can be a
good approximate description of time evolution. We will analyze this deeply
in section 6.

Quantum Markov process: mathematical

structure

Before studying under which conditions can an open quantum system be approximately described by a Markovian evolution, we examine in this section
the structure and properties of quantum Markovian process.

5.1

Classical Markovian processes

To motivate the adopted definition of a quantum Markov process, let us recall

the definition of classical Markov processes. More detailed explanations can
be found in [14,21,78,79]; here we just sketch the most interesting properties
for our purposes without getting too concern with mathematical rigor.
On a probability space, a stochastic process is a family of random variables
{X(t), t I R}. Roughly speaking, a stochastic process is specified by
a random variable X depending on a parameter t which usually represents
time. Assume that I is a countable set labeled by n, then a stochastic process
35

is a Markov process if the probability that the random variable X takes a

value xn at any arbitrary time tn , is conditioned only by the value xn1 that
it took at time tn1 , and does not depend on the values at earlier times. This
condition is formulated in terms of the conditional probabilities as follows:
p(xn , tn |xn1 , tn1 ; . . . ; x0 , t0 ) = p(xn , tn |xn1 ),

tn I.

(42)

This property is sometimes expressed in the following way: a Markov process

does not have memory of the history of past values of X. They are so-named
after the Russian mathematician A. Markov.
Consider now this process on a continuous interval I, then tn1 is time
infinitesimally close to tn . So for any t > t just from the definition of
conditional probability
p(x, t; x , t ) = p(x, t|x , t )p(x , t ),
where p(x, t; x , t ) is the joint probability that random variable X takes value
x at time t and x at time t . By integrating with respect to x we find the
relation between unconditional probabilities
Z
p(x, t) = dx p(x, t|x , t )p(x , t ),

and let us write it as

Z
p(x, t) = dx K(x, t|x , t )p(x , t ) with K(x, t|x , t ) = p(x, t|x , t ). (43)

The Markov process is called homogeneous if K(x, t|x , t ) is only a function

of the difference between the two time parameters involved K(x, t|x , t ) =
Ktt (x|x ).
Next, take the joint probability for any three consecutive times t3 > t2 >
t1 and apply again the definition of conditional probability twice
p(x3 , t3 ; x2 , t2 ; x1 , t1 ) = p(x3 , t3 |x2 , t2 ; x1 , t1 )p(x2 , t2 ; x1 ; t1 )
= p(x3 , t3 |x2 , t2 ; x1 , t1 )p(x2 , t2 |x1 , t1 )p(x1 , t1 ).
The Markov condition implies that p(x3 , t3 |x2 , t2 ; x1 , t1 ) = p(x3 , t3 |x2 , t2 ),
then by integrating over x2 and dividing both sides by p(x1 , t1 ) one gets
Z
p(x3 , t3 |x1 , t1 ) = dx2 p(x3 , t3 |x2 , t2 )p(x2 , t2 |x1 , t1 ),

which is called Chapman-Kolmogorov equation. Observe that with the notation of (43) this is
Z
K(x3 , t3 |x1 , t1 ) = dx2 K(x3 , t3 |x2 , t2 )K(x2 , t2 |x1 , t1 ).
(44)
36

Thinking of K(x, t|x , t ) as propagators of the time evolution between t and

t (see Eq. 43), equation (44) express that they form evolution families with
that composition law. Moreover, since K(x, t|x , t ) are conditional probabilities they connect any probability p(x , t ) with another probability p(x, t) so
in this sense they are universal (preserve the positivity of p(x , t ) and the
normalization).
We can now draw a clear parallelism with a quantum setting. In the
definition of quantum Markov process (definition 4.1), the role played by the
probabilities p(x, t) is played by the density operators (t) and the role of
conditional probabilities p(x, t|x , t ) is played by UDMs E(t,t ) , such that (41)
is the quantum analog to the classic Chapman-Kolmogorov equation (44).

5.2

Quantum Markov evolution as a differential equation

For positive , consider the difference

(t+)(t) = [E(t+,0) E(t,0) ](0) = [E(t+,t) 1]E(t,0) [(0)] = [E(t+,t) 1](t),
provided that the limit 0 is well defined (we assume that time evolution
is smooth enough). We can obtain a linear differential equation for (t)
(called master equation)
[E(t+,t) 1]
d(t)
(t + ) (t)
= lim
= lim
(t) = Lt (t),
0
0
dt

(45)

where by definition the generator of the evolution is

[E(t+,t) 1]
.
0

Lt = lim

If some quantum evolution is Markovian according to our definition 4.1,

we may wonder what is the form of these differential equations. The answer
is given by the following theorem, which is the main result of this section.
Theorem 5.1. A differential equation is a Markovian master equation if and
only if it can be written in the form


X
1
d(t)

= i[H(t), (t)] +
k (t) Vk (t)(t)Vk (t) {Vk (t)Vk (t), (t)} ,
dt
2
k
(46)
where H(t) and Vk (t) are time-dependent operators, with H(t) self-adjoint,
and k (t) 0 for every k and time t.
37

Proof. Let us show the necessity first (for this we follow a generalization of
the argument given in [9] and [21] for time-independent generators). If E(t2 ,t1 )
is a UDM for any t2 t1 it admits a decomposition of the form
X
E(t2 ,t1 ) [] =
K (t2 , t1 )K (t2 , t1 ).

Actually, let {Fj , j = 1, . . . , N } be a complete orthonormal basis with respect to the Hilbert-Schmidt inner product (Fj , Fk )HS = Tr(Fj Fk ) = jk , in

such a way that we chose FN 2 = 1/ N and the rest of operators are then
traceless. Now the expansion of the Kraus operators in this basis yields
X
E(t2 ,t1 ) [] =
cjk (t2 , t1 )Fj Fk ,
j,k

where the elements

cjk (t2 , t1 ) =

(Fj , K (t2 , t1 ))HS (Fk , K (t2 , t1 ))HS

form a positive semidefinite matrix c(t2 , t1 ). Indeed, for any N 2 -dimensional

vector v we have that

!2

X  X
X

v
(F
,
K
(t
,
t
))
vj cjk (t2 , t1 )vk =
(v, c(t2, t1 )v) =

k
k
2 1 HS  0.


j,k

Since this holds for any t2 t1 , let us take t1 = t and t2 = t + , then the
generator reads

X cjk (t + , t)Fj F 1
1 cN 2 N 2 (t + , t) N
k
Lt () = lim
= lim

0
0 N

j,k
2


N 1 
cN 2 j (t + , t)
1 X cjN 2 (t + , t)
Fj +
Fj
+

N j=1
#
2 1
N
X
cjk (t + , t)
+
Fj Fk .

j,k=1

Now, we define the time-dependent coefficients ajk (t) by

cN 2 N 2 (t + , t) N
,
0

cjN 2 (t + , t)
ajN 2 (t) = lim
, j = 1, . . . , N 2 1,
0

cjk (t + , t)
ajk (t) = lim
, j, k = 1, . . . , N 2 1,
0

aN 2 N 2 (t) = lim

38

(47)

the following operators,

2

and

N 1
1 X
F (t) =
ajN 2 (t)Fj ,
N j=1

1
aN 2 N 2
1 + F (t) + F (t) ,
G(t) =
2N
2


1 
F (t) F (t) ,
2i
which is self-adjoint. In terms of these operators the generator (47) can be
written like
H(t) =

Lt () = i[H(t), ] + {G(t), } +

2 1
N
X

ajk (t)Fj Fk .

j,k=1

Since the UDMs preserve trace for every density matrix ,

# )
("
2 1
N
X
ajk (t)Fk Fj ,
0 = Tr [Lt ()] = Tr
2G(t) +
j,k=1

and therefore we conclude that

2

N 1
1 X
G(t) =
ajk (t)Fk Fj .
2
j,k=1

P
Note that this is nothing but the condition K (t2 , t1 )K (t2 , t1 ) = 1 expressed in differential way. Thus the generator adopts the form
2 1


N
X
1

Lt () = i[H(t), ] +
ajk (t) Fj Fk {Fk Fj , } .
2
j,k=1
Finally note that because of the positive semidefiniteness of cjk (t + , t), the
matrix ajk (t), j, k = 1, . . . , N 2 1 is also positive semidefinite, so for any t
it can be diagonalized by mean of a unitary matrix u(t); that is,
X
umj (t)ajk (t)unk (t) = m (t)mn ,
j,k

where each eigenvalue is positive m (t) 0. Introducing a new set of operators Vk (t),
2 1
2 1
N
N
X
X

Vk (t) =
ukj (t)Fj , Fj =
ukj (t)Vk (t),
j=1

k=1

39

we obtain the desired result



X
1

Lt () = i[H(t), ] +
k (t) Vk (t)Vk (t) {Vk (t)Vk (t), } ,
2
k

with k (t) 0, k, t.
Assume now that some dynamics is given by a differential equation as
(46). Since it is a first order linear equation, there exists a continuous family
of propagators satisfying
E(t2 ,t0 ) = E(t2 ,t1 ) E(t1 ,t0 ) ,
E(t,t) = 1,
for every time t2 t1 t0 . It is clear that these propagators are trace preserving (as for any , Tr(Lt ) = 0). So we have to prove that for any times
t2 t1 the dynamical maps E(t2 ,t1 ) are UDMs, or equivalently completely positive maps. For that we use the approximation formulas introduced in section
2; firstly the time-splitting formula (17) allow us to write the propagator like
E(t2 ,t1 ) =

0
Y

lim

max |tj+1 tj |0

Lt (tj+1 tj )

(48)

j=n1

where t2 = tn tn1 . . . t0 = t1 . Next, take any element of this product

Lt

= t+1 t 0,

and since the constants k (t) are positive for every time, we can write this
generator evaluated at the instant t as

X
1

Vk (t )()Vk (t ) {Vk (t )Vk (t ), ()} ,

Lt () = i[H(t ), ()] +
2
k
where Vk (t ) =

k (t )Vk (t ). Now we split this generator in several parts,

L =
t

M
X
k=0

Lk ,

where
L0 () = i[H(t ), ()]
Lk () = Vk (t )()Vk (t ),

M
X
1

40

{Vm (t )Vm (t ), ()},

k = 1, . . . , M,

(49)
(50)

and here M is the upper limit in the sum over k in (46). On one hand,
differentiating with respect to one can easily check that

eL0 = e{[iH(t )

k 2 Vk (t )Vk (t )

] } e{[iH(t )

k 2 Vk (t )Vk (t )

] }

which has the form OO , so it is completely positive for all . On the other
hand, for 0

X
m m m
Lk
e =
V (t )V (t ),
m! k k
m=0
P

has also the Kraus form m Om Om

, so it is completely positive for all k.
Now everything is ready to invoke the Lie-Trotter product formula (5), since
!n
N
Y
PM

L
k
L t

L
en
e = e( k=0 k ) = lim
n

k=1

is the composition (infinite product) of completely positive maps eLk /n , it

is completely positive for all . Finally, because this is true for every instantaneous t the time-splitting formula (48) asserts that E(t2 ,t1 ) is completely
positive for any t2 and t1 , because it is just another composition of completely
positive maps [80].
Q.E.D.
This result can be taken as a global consequence of the works of A. Kossakowski [73, 74, 81] and G. Lindblad [82], who analyzed this problem for the
case of time-homogeneous equations, that is, when the UDMs depend only
on the difference = t2 t1 , E(t2 ,t1 ) = E and so they form a one-parameter
semigroup (not a group because the inverses are generally not a UDM, see
theorem 4.5). As we know, under the continuity assumption, E = eL can be
defined and the generator L is time-independent, moreover it has the form

X 
1

(51)
k Vk (t)Vk {Vk Vk , (t)} .
L(t) = i[H, (t)] +
2
k
This is the result proven by Gorini, Kossakowski and Sudarshan [81] for finite
dimensional semigroups and by Lindblad [82] for infinite dimensional systems
with bounded generators, that is, for uniformly continuous semigroups [46].

5.3

Kossakowski conditions

It is probably worthwhile to dedicate a small section to revise the original

idea of Kossakowskis seminal work [73, 74, 81], as it has clear analogies with
41

some results in classical Markov processes. The route that A. Kossakowski

and collaborators followed to arrive at (51) was to derive some analogue to
the classical Markovian conditions on the generator L. That is, for classical
Markov process in finite dimensional systems, a matrix Q is the generator
of an evolution stochastic matrix eQ for any > 0 if only if the following
conditions in its elements are satisfied:
Qii 0, i,
Qij 0, i 6= j,
X
Qij = 0, i.

(52)
(53)
(54)

The proof can be found in [78]. To define similar conditions in the quantum
case, one uses the Lumer-Phillips theorem 2.5, as E = eL has to be a contraction semigroup in the Banach space of self-adjoint matrices with respect
to the trace norm. For that aim, consider the next definition.
P
Definition 5.1. Let B, with spectral decomposition = j j Pj . We
define the linear operator sign of as
X
sgn() =
sgn(j )Pj ,
j

where sgn(j ) is the sign function of each

1,
sgn(j ) =
0,

1,

eigenvalue, that is
j > 0
j = 0
j < 0.

Definition 5.2. We define the following product of two elements , B,

[, ]K = kk Tr[sgn()].

(55)

Proposition 5.1. The product (55) is a semi-inner product. In addition it

is a real number.
Proof. The proof is quite straightforward. It is evident that for the product
(55), (7) and (8) hold. FinallyP(9) also follows easily. By introducing the
spectral decomposition of = j j Pj ,


X

X


|[, ]K | = kk 
sgn(j ) Tr(Pj ) kk
|sgn(j )||Tr(Pj )|


j
j
X
kk
|Tr(Pj )|;
j

42

P
and we also use the spectral decomposition of = j j Pj, to obtain


 X
X
X X


Tr(P
P
)
|k ||Tr(Pj Pk )|
|Tr(Pj )| =

k
j k 


j
j
k
k,j
X
X
X

=
|k | Tr(Pj Pk ) =
|k | Tr(Pk) =
|k | = kk.
k,j

The reality of [, ]K is clear from the fact that both sgn() and are selfadjoint.
Q.E.D.
Now we can formulate the equivalent to the conditions (52), (53) and (54) in
the quantum case.
Theorem 5.2 (Kossakowski Conditions). A linear operator L on B (of finite
dimension N) is the generator of a trace preserving contraction semigroup
if and only if for every resolution of the identity P = (P1 , P2 , . . . , PN ), 1 =
P
j Pj , the relations
N
X

Aii (P) 0
Aij (P) 0

(i = 1, 2, . . . , N),
(i 6= j = 1, 2, . . . , N),

(56)
(57)

Aij (P) = 0

(j = 1, 2, . . . , N)

(58)

i=1

are satisfied, where

Aij (P) = Tr[Pi L(Pj )].
Proof. The condition (58) is just the trace preserving requirement Tr[L()] =
0, B. We have
N
X
i=1

Aij (P) =

N
X
i

Tr[Pi L(Pj )] = Tr[L(Pj )] = 0,

Pj P

Tr[L()] = 0,

B.

On the other hand, the theorem of Lumer-Phillips 2.5 asserts that

[, L()]K 0;

(59)

taking to be some projector P we obtain

Tr[P, L(P )] 0,
43

(60)

so (56) is necessary. Furthermore (59) can be rewritten as

X
X
[, L()]K =
sgn(k )j Tr[Pk L(Pj )] =
|j | Tr[Pj L(Pj )]
j

k,j

X
k6=j

sgn(k )j Tr[Pk L(Pj )] 0,

(61)

P
where we have used the spectral decomposition =
j j Pj again. By
splitting (58) in a similar way
X
Tr[Pk L(Pj )], for each j = 1, . . . , N.
Tr[Pj L(Pj )] =
k6=j

Here note that on the right hand side the sum is just over the index k (not
over j). Inserting this result in (61) yields
X
kj Tr[Pk L(Pj )] 0,
(62)
k6=j

where the sum is over the two indexes, with

kj = |j |[1 sgn(k ) sgn(j )] 0.
Thus for trace-preserving semigroups the condition (61) is equivalent to (56)
and (62). Therefore if (56), (57) and (58) are satisfied L is the generator of
a trace preserving contraction semigroup because of (60) and (62). It only
remains to prove the necessity of (57), that can be seen directly from (62).
However, it is also possible to show it from the definition of generator:
Tr[Pk (E 1)Pj ]
Tr[Pk E (Pj )]
= lim
0,
0
0

since E (Pj ) 0 as E is a trace preserving and completely positive map.

Q.E.D.
Tr[Pk L(Pj )] = lim

Finally from this result, one can prove (51). For that one realizes (similarly to the necessity proof of theorem 5.1) that a trace and self-adjoint
preserving generator can be written as
2 1


N
X
1

ajk Fj Fk {Fk Fj , } ,
L = i[H, ] +
2
j,k=1
where H = H and the coefficients ajk form a Hermitian matrix. E is
completely positive (not just positive) if and only if L 1 (in the place of L)
satisfies the conditions (56), (57) and (58). It is easy to see [81] this implies
that ajk is positive semidefinite, and then one obtains (51) by diagonalization.
The matrix ajk is sometimes referred to as Kossakowski matrix [28, 83].
44

5.4

Steady states of homogeneous Markov processes

In this section we give a brief introduction to the steady state properties of

homogeneous Markov processes, i.e. completely positive semigroups. This is
a complicated topic and still an active area of research, so we only expose
here some simple but important results.
Definition 5.3. A semigroup E is relaxing if there exist a unique (steady)
state ss such that E (ss ) = ss for all and
lim E () = ss ,

for every initial state .

Of course if one state ss is a steady state of a semigroup E , it is an
eigenoperator of the operator E with eigenvalue 1,
E (ss ) = eL ss = ss ,
and, by differentiating, it is an eigenoperator of the generator with zero
eigenvalue,
L(ss ) = 0.
(63)
If this equation admits more than one solution then the semigroup is obviously not relaxing, and the final state will depend on the initial state, or
might even oscillate inside of the set of convex combinations of solutions.
However, the uniqueness of the solution of equation (63) should not presumably be a guarantee for the semigroup to be relaxing. First of all let us point
out the next result.
Theorem 5.3. A completely positive semigroup E = eL on a finite dimensional Banach space B has always at least one steady state, this is, there
exists at least a state ss which is a solution of (63).
Proof. We prove this by construction. Take any initial state (0); we calculate the average state of its evolution by (this is sometimes called ergodic
average)
Z
1 T L
= lim
e (0)d .
(64)
T T 0

It is simple to show that the integral converges just by taking norms

Z T

Z
1

1 T L
L


lim
e (0)d lim
ke kk(0)kd 1.
T T 0
T T 0
45

Because eL is completely positive, it is obvious that is a state. Moreover,

is clearly a steady state
eL = .
Q.E.D.
The easiest way to show the steady state properties of a finite dimensional
semigroup is by writing the generator and the exponential in matrix form.
In order to do that, let us consider again the Hilbert-Schmidt inner product
on B and a set of mutually orthonormal basis {Fj , j = 1, . . . , N 2 }, that is
(Fj , Fk )HS = Tr(Fj Fk ) = jk . Then the matrix elements of the generator L
in that basis will be
Ljk = (Fj , LFk )HS = Tr[Fj L(Fk )].
To calculate the exponential of Ljk one uses the Jordan block form (in general
Ljk is not diagonalizable). That is, we look for a basis such that Ljk can be
written as a block-diagonal matrix, S1 LS = L(1) L(2) . . . L(M ) , of M
blocks where S is the matrix for the change of basis, and each block has the
form

i 1 0 . . . 0
.

..
. ..
0 i 1

. .. .. ..
.
.
L(i) =
.
.
.
0
.

.
.
.
.
.
.
.
.
. 1
0 . . . . . . 0 i
Here i denotes some eigenvalue of L and the size of the block coincides
with its algebraic multiplicity ki . It is a well-know result (see for example
references [46, 84]) that the exponential of L can be explicitly computed as
(1)

(2)

(M )

eL = SeL eL . . . eL S1






= S e1 eN1 e2 eN2 . . . eM eNM S1 ,

(65)

where the matrices Ni = L(i) i 1 are nilpotent matrices, (Ni )ki = 0, and
thus their exponentials are easy to deal with. If L is a diagonalizable matrix,
Ni = 0 for every eigenvalue.
Now we move on to the principal result of this section.
Theorem 5.4. A completely positive semigroup is relaxing if and only if
the zero eigenvalue of L is non-degenerate and the rest of eigenvalues have
negative real part.

46

Proof. For the only if part. The condition of non-degenerate zero eigenvalue of L is trivial, otherwise the semigroup has more than one steady state
and is not relaxing by definition. But even if this is true, it is necessary
that the rest of eigenvalues have negative real part. Indeed, consider some
eigenvalue with purely imaginary part L() = i, where is the associated
eigenvector. By using the fact that eL is trace preserving


eL () = ei Tr eL () = Tr() = ei Tr() Tr() = 0.

So we can split as a difference of positive operators = + , given by

its spectral decomposition as in the proof of theorem 4.4. Since Tr() = 0,
Tr(+ ) = Tr( ) = c, and we can write /c = + , where are now
quantum states. The triangle inequality with the steady state ss reads
keL (/c)k = keL (+ )k keL (+ ss )k + keL ( ss )k
= keL (+ ) ss k + keL ( ) ss k.
Thus if the semigroup is relaxing we obtain lim keL (/c)k = 0, but this
is impossible as eL () = ei and
keL (+ )k = keL (/c)k = |ei |k(/c)k = k/ck > 0,

Conversely, let us denote 1 = 0 the non-degenerate zero eigenvalue of

L, since the remaining ones have negative real part, according to (65) in the
limit the matrix form of the semigroup will be






E = lim eL = lim S e1 e2 eN2 . . . eM eNM S1

= S (1 0 . . . 0) S1 = SD1 S1 ,

where

D1 =

0 0
0 0

.. . . ..
.
.
.
0 0 0

(66)

1
0
..
.

coincides with the projector on the eigenspace associated with the eigenvalue
1. From here it is quite simple to see that the application of E on any
initial state written as a vector in the basis of {Fj , j = 1, . . . , N 2 } will give
us just the steady state. More explicitly, if ss is the steady state, its vector
representation will be
vss = [(F1 , ss )HS , . . . , (FN 2 , ss )HS ]t ,
47

and the condition eL ss = ss is of course expressed in this basis as

eL vss = vss .

(67)

On the other hand, as we already have said, D1 is the projector on the

eigenspace associated with the eigenvalue 1, which is expanded by vss , so we
can write D1 = vss (vss , ) or D1 = |vss ihvss | in Dirac notation.
As a particular case ( ) of (67), we have E vss = vss , by making a
little bit of algebra with this condition
E vss = SD1 S1 vss = vss Svss (vss , S1 vss ) = vss
Svss = (vss ,S11 vss ) vss ,

(68)

therefore vss is eigenvector of S with eigenvalue (vss ,S11 vss ) .

Finally consider any initial state written as a vector v in the orthonormal basis {Fj , j = 1, . . . , N 2 }. We have
lim eL v = E v = SD1 S1 v = Svss (vss , S1 v ),

thus using (68),

E v =

(vss , S1 v )
vss = C()vss ,
(vss , S1 vss )

C()

(vss , S1 v )
,
(vss , S1 vss )

and in the notation of states and operators this equation reads

lim eL = C()ss ,

for all .

One can compute C() explicitly and find that C() = 1 for every , however
this follows from the condition that eL is a trace preserving semigroup.
Q.E.D.
This theorem can be taken as a weaker version of the result by Schirmer
and Wang [85], which assert that if the steady state is unique then the semigroup is relaxing (there are not purely imaginary eigenvalues). A particular
simple case is if the unique steady state is pure ss = |ih|. Then since a
pure state cannot be expressed as a nontrivial convex combination of other
states, the equation (64) implies that the semigroup is relaxing. This was
also stated for an arbitrary UDM in theorem 8 of reference [86].
On the other hand, it is desirable to have some condition on the structure of the generator which assures that the semigroup is relaxing as well
as a characterization of the steady states; there are several results on this
topic (see [6, 85, 8799] and references therein). Here we will only present an
important result given by Spohn in [89].
48

Theorem 5.5 (Spohn). Consider a completely positive semigroup, E = eL ,

in B with generator

X 
1

k Vk Vk {Vk Vk , } ,
L = i[H, ] +
2
kI
for some set of indexes I. Provided that the set {Vk , k I} is self-adjoint
(this is, the adjoint of every element of the set is inside of the set) and the
only operators commuting with all of them are proportional to the identity,
which is expressed as {Vk , k I} = c1, the semigroup E is relaxing.
Proof. Since {Vk , k I} is self-adjoint we can consider a self-adjoint orthonormal basis with respect to the Hilbert-Schmidt product {Fj
, j = 1, . . . , N 2 },
such that, as in the proof of theorem 5.1, we chose FN 2 = 1/ N, and the
remaining elements are traceless. Expanding
Vk =

2 1
N
X

m=1

we rewrite the generator as

vk,N 2
vkm Fm + 1,
N

2 1
N
X



1
] +
L = i[H + H,
amn Fm Fn {Fn Fm , } ,
2
m,n=1

(69)

(70)

is the self-adjoint operator

where H
2

N 1


k X

vk,N
vk,N 2 vkm
H=
2 vkm Fm ,
2i N m=1
kI
P

and the Hermitian matrix is amn = kI k vkm vkn

. Now note that we can
trivially chose amn to be strictly positive; for that it is enough to diagonalize
again amn and ignore the members in the sum of (70) which goes with zero
eigenvalue as they do not contribute, and write again the less dimensional
matrix amn in terms of the numbers of elements of the basis strictly necessary,
say p N 2 1,


p
X
1

amn Fm Fn {Fn Fm , } .
L = i[H + H, ] +
2
m,n=1

Next, consider some > 0 to be smaller than the smallest eigenvalue of the
strictly positive amn , and split the generator in two pieces L = L1 + L2 with

p 
X
1
Fm Fm {Fm Fm , } .
L2 =
2
m=1
49

It is clear that L2 is the generator of a completely positive semigroup, and

the same for L1 since amn mn is still a positive semidefinite matrix. Next,
for some B we compute
(, L2 ())HS = Tr[L2 ()] =
p

p
X

m=1

Tr[(Fm )2 2 Fm2 ]

p


X
X
Tr ([Fm , ][Fm , ]) =
Tr [Fm , ] [Fm , ] 0,
=
2 m=1
2 m=1

so L2 has only real eigenvalues smaller or equal to zero. Moreover for the
eigenvalue 0 we have Tr[L2 ()] = 0 which requires [Fm , ] = 0 for m =
1, ..., p. But, because of the equation (69), it implies that [Vk , ] = 0 for
k I and therefore = c1 by presupposition.
L1 and L2 denote the restriction of L, L1 and L2 to the subNow, let L,
space generated by {Fj , j = 1, . . . , N 2 1}, this is the subspace of traceless
self-adjoint operators. Then since L1 is the generator of a contracting semigroup all the eigenvalues of L1 have real part smaller or equal to zero. On the
other hand from the foregoing we conclude that the spectrum of L2 is composed only of strictly negative real numbers, this implies that L = L1 + L2
L1
has only eigenvalues with strictly negative real part. Indeed, writing L,
L
1 and L
2 , we can split them in their
and L2 in its matrix representation, L,
Hermitian and anti-Hermitian parts, e.g.
=L
A + iL
B,
L





where LA = L + L /2 and LB = L L /2i are Hermitian matri 1 and L

2 . It is quite evident that the real part of the
ces. Similarly for L

A is a negative definite
spectrum
negative if and only if L
 of L is strictly

A v < 0 for any Span{Fj , j = 1, . . . , N 2 1}. But, since
matrix, v , L




A = L
A1 + L
A2 where L
A1 = L
1 + L
/2 and L
A2 = L
2 + L
/2, we
L
1
2
obtain

 
 

A v = v , L
A1 v + v , L
A2 v < 0,
v , L




A1 v 0 and v , L
A2 v < 0. Thus all the eigenvalues of L
as v , L
have strictly negative real part.
Finally since the eigenvalues of L are just the ones of L plus one (this is
obvious because the dimension of the subspace of traceless operators is just
one less than the whole state dimension N 2 ), and this extra eigenvalue has
to be equal to 0 because theorem 5.3 asserts that there exists always at least
50

one steady state. Therefore L fulfills the conditions of theorem 5.4 and the
semigroup eL is relaxing.
Q.E.D.
For checking if some semigroup satisfies the conditions of this theorem
it is sometimes useful to explore if the operators {Vk , k I} correspond to
some irreducible representation of a group. Then we may invoke the Schurs
lemma [100].

Microscopic description: Markovian case

As we have already pointed out, the dynamics of an open quantum system is

notably different from that of a closed system; both from practical and fundamental points of view. Essential questions as the existence of universally
valid dynamics or the continuity in time of these, become non-trivial matters
for the evolution of open quantum systems.
Generally it would be necessary to describe the evolution in the extended
closed system, and trace out the state at the end of the evolution. Unfortunately, many times, one of the systems is out of our control, and/or it has
an infinite number of degrees of freedom, such that the task of calculating
the associated Kraus operators can be difficult and impractical. That is the
case for example of a finite N-dimensional system coupled with an infinite
dimensional one; according to Eq. (35), we would obtain, in principle, infinite terms in the sum. However it can be easily shown that any completely
positive map on a N-dimensional system can be written with at most N 2
Kraus operators (of course; since the dimension of B is N 2 , so the use of
the infinite terms is unnecessary.
For that reason other techniques apart from the brute force strategy
will become very useful and they will be the subject of this section.
FirstLY, we analyze some conditions under for a UDM to be approximately be approximately represented by a quantum Markov process. We
will start from a microscopic Hamiltonian and apply some approximations
leading to UDMs which fulfill the divisibility property (41). Note that, in
principle, as we have already said, the reduced evolution of a quantum system
is never exactly Markovian. This is because, as discussed in section 4.5, there
is always a privileged time, say t0 , where the both systems A and B start
to interact being then the global state a product = A B and defining
a UDM for A . Since generally the global state can no longer be factorized,
and (generally) the dynamical map from there is not a UDM, the divisibility
property (41) is violated.
51

6.1

Nakajima-Zwanzig equation

Let assume that A is our open system while B plays the role of the environment; the Hamiltonian of the whole system is given by
H = HA + HB + V,

(71)

where HA and HB are the Hamiltonians which govern the local dynamics of
A and B respectively, and V stands for the interaction between both systems.
We will follow the approach initroduced by Nakajima [101] and Zwanzig
[102] using projection operators, which is also explained in [3, 20, 21] among
others. In this method we define in the combined Hilbert space system plus
environment H = HA HB , two orthogonal projection operators, P and
Q, P 2 = P, Q2 = Q and PQ = QP = 0, given by
P = TrB () B ,
Q = (1 P),

(72)
(73)

where H is the whole state of the system and environment and B HB

is a fixed state of the environment. In fact, we choose B to be the physical
initial state of the environment. Strictly speaking this is not necessary but
the method becomes much more complicated if one chooses another state.
In addition, we assume that the system B is initially in thermal equilibrium,
this is


B = th = eHB [Tr eHB ]1 ,

where = 1/T . Note that P give all necessary information about the
reduced system state A ; so the knowledge of the dynamics of P implies
that one knows the time evolution of the reduced system.
If we start from the von Neumann equation (19) for the whole system
d(t)
= i[HA + HB + V, (t)],
dt

(74)

and take projection operators we get

d
P(t) = iP[HA + HB + V, (t)],
dt
d
Q(t) = iQ[HA + HB + V, (t)].
dt

(75)
(76)

As usual in perturbation theory we shall work in interaction picture with

respect to the free Hamiltonians
(t) = ei(HA +HB )t (t)ei(HA +HB )t .
52

Since


TrB [
(t)] = TrB ei(HA +HB )t (t)ei(HA +HB )t


= eiHA t TrB eiHB t (t)eiHB t eiHA t = eiHA t TrB [(t)]eiHA t = A (t).
the projection operators are preserved under this operation, and equations
(75) and (76) can be written in interaction picture just like
d
P (t)

= iP[V (t), (t)],

dt
d
Q
(t) = iQ[V (t), (t)].
dt

(77)
(78)

Let us use the notation V(t) i[V (t), ]; then by introducing the identity 1 = P + Q between V(t) and (t) the equations (77) and (78) may be
rewritten as
d
P (t)

= PV(t)P (t)
+ PV(t)Q
(t),
dt
d
Q
(t) = QV(t)P (t)
+ QV(t)Q
(t).
dt
The formal integration of the first of these equations gives
Z t
Z t
P (t) = P(0) +
dsPV(s)P (s)
+
dsPV(s)Q
(s),
0

(79)
(80)

(81)

where we have set 0 as the origin of time without losing generality. On the
other hand, the solution of the second equation can be written formally as
Z t
Q
(t) = G(t, 0)Q
(0) +
dsG(t, s)QV(s)P (s).
(82)
0

This is nothing but the operational version of variation of parameters formula

for ordinary differential equations (see for example [103, 104]), where the
solution to the homogeneous equation,
d
Q
(t) = QV(t)Q
(t),
dt
is given by the propagator
Rt

G(t, s) = T e
53

dt QV(t )

Inserting (82) in (81) yields

Z t
Z t
P (t) = P(0) +
dsPV(s)P (s)
+
dsPV(s)G(s, 0)Q
(0)
0
0
Z t Z s
+
ds
duPV(s)G(s, u)QV(u)P (u).
(83)
0

This is the integrated version of the so-called generalized Nakajima-Zwanzig

equation
d
P (t) = PV(t)P (t)
+ PV(t)G(t, 0)Q
(0)
dt
Z t
+
duPV(t)G(t, u)QV(u)P (u).

(84)

Next, two assumptions are usually made. First one is that PV(t)P = 0,
this means,
h


i
PV(t)P = i TrB [V (t), A B ] B = i TrB V (t)B , A B = 0,



for every and then every A , which implies TrB V (t)B = 0. If this is not
fulfilled (which is not the case in the most practical situations), provided that
B commutes with HB one can always define a new interaction Hamiltonian
with a shifted origin of the energy for A [10, 34]. The change
V = V TrB [V B ] 1,

and HA = HA + TrB [V B ] 1,

(85)

makes the total Hamiltonian to be the same and

h
i
h
i
+H )t
+H )t

i(HA
i(HA
B
B

TrB V (t)B = TrB e

V e
B

= eiHA t TrB [eiHB t V eiHB t B ]eiHA t

eiHA t TrB [V B ]eiHA t TrB [eiHB t B eiHB t ]

= eiHA t TrB [V B ]eiHA t eiHA t TrB [V B ]eiHA t = 0,

as we wished.
The second assumption consists in accepting that initially (0) = A (0)
B (0), which is the necessary assumption to get a UDM. Of course some
skepticism may arise thinking that if the system B is out of our control, there
is not guaranty that this assumption is fulfilled. Nonetheless the control on
the system A is enough to assure this condition, for that it will be enough to
prepare a pure state in A, for instance by making a projective measurement
54

as argued in [105], to assure, by theorem 4.1 that at the initial time the global
system is in a product state.
These two assumptions make to vanish the second and the third term in
equation (83), and the integro-differential equation yields
Z t
d
P (t)
=
duK(t, u)P (u),
(86)
dt
0
where
K(t, u) = PV(t)G(t, u)QV(u).
If this kernel is homogeneous K(t, u) = K(tu), which is the case for stationary states B (as exemplified in next section), this equation can be formally
solved by a Laplace transformation [4], but this usually turns out into an
intractable problem in practice.
In order to transform this integro-differential equation in an Markovian
master equation one would wish that the integral on the right hand side of
(86) turns into a generator of the form of (46). For that aim, the first intuitive
guess is to try that K(t, u) behaves as a delta function with respect to (u).
For that to be true the typical variation time A of (u) (which is only due
to the interaction with B because we have removed the rest of the dynamics
by taking integration picture) has to be much larger than some time B ,
which characterizes the speed at which the kernel K(t, u) is decreasing when
|t u| 1. Of course this kind of approximations intrinsically involve some
assumptions on the size of B and the strength of the Hamiltonians, and
there is in principle no guaranty that the resulting differential equation has
the form of (46). However, several works [106110] studied carefully two
limiting procedures where A /B so that this turns out to be the case.
These two limits are
The weak coupling limit. Let us rewrite V V , where accounts
for the strength of the interaction. Since for small the variation of
(u) is going to be slow, in the limit 0 one would get A ,
and so A /B for every fixed B . Of course for = 0 there
is not interaction, and so we will need to rescale the time parameter
appropriately before taking the limit. This is discussed in more detail
in next section 6.2.
The singular coupling limit. This corresponds to somehow the opposite case; here we get B 0 by making K(t, u) to approach a delta
function. We will discuss this limit in the forthcoming section 6.3.

55

We should stress that one can construct Markovian models ad hoc which
try to describe a system from a microscopic picture without involving these
procedures; see for example [111, 112]. In this sense, these two cases are
sufficient conditions to get a Markovian evolution, but not necessary.

6.2

Weak coupling limit

As we said above, if we assume the weak coupling limit, the change of (u)
is negligible within the typical time B where the kernel K(t, u) is varying.
To perform this limit, consider again Eq. (83) under the assumptions of
PV(t)P = 0 and (0) = A (0) B (0):
Z t Z s
2
P (t) = P(0) +
ds
duPV(s)G(s, u)QV(u)P (u),
(87)
0

where we have redefined V V (so V V). Since in the interaction

picture the whole state satisfies a von Neumann equation of the form
d
(t) = i[V (t), (t)],
dt

(88)

we can connect the whole state at times s and u (note u s) by the unitary
solution
(s, u) = U(s,
u)
(s) = U (s, u)
(u)U
(u)
(s, u)
u) U(u,
s)
(u) = U
(u)U(s,
(s).
By introducing this operator in the last term of (87) we find
Z t Z s
2
s)
ds
P (t) = P(0) +
duPV(s)G(s, u)QV(u)P U(u,
(s).
0

Now the term inside the double integral admits an expansion in powers of
given by the time-ordered series of
Rt

G(s, u) = T e s dt QV(t ) and,

h
i h
i
R
R
u) = T e us dt V(t ) = T e us dt V (t ) ,
U (u, s) = U(s,

(89)

where the adjoint is taken over every element of the expansion and T denotes
the antichronological time-ordering operator. This is defined similarly as T

56

in (14) but ordering in the opposite sense, this is, the operators evaluated at
later times appear first in the product. Thus, at lowest order we get
Z t Z s
2
P (t) = P(0) +
ds
duPV(s)QV(u)P (s) + O(3 )
Z0 t Z0 s
= P(0) + 2
ds
duPV(s)V(u)P (s) + O(3 ).
0

where we have used again that PV(t)P = 0. After a change of variable

u su, writing explicitly the projector P and the operators V, we conclude
that
Z t Z s
h
i
2
A (t) = A (0)
ds
du TrB V (s), [V (s u), A (s) B ] + O(3 ).
0

6.2.1

(90)

Decomposition of V

In order to continue with the derivation, we write the interaction Hamiltonian

as
X
Ak Bk ,
(91)
V =
k

Ak

Bk

with
= Ak and
= Bk . Note that since V = V , one can always chose
Ak and Bk to be self-adjoint. Indeed, assume that V is a sum of products of
any general operators Xk and Yk , it will be written then like
n
X
V =
Xk Yk + Xk Yk .
(92)
k=1

Any operator can be decomposed as

(a)

(b)

Xk = Xk + iXk ,
where

i(Xk Xk )
(Xk + Xk )
(b)
, Xk =
2
2
(a)
are self-adjoint. This, and the same kind of decomposition for Yk = Yk +
(b)
iYk inserted in equation (92) gives
n 

 
X
(a)
(b)
(a)
(b)
Xk iXk Yk + iYk
V =
(a)

Xk =

k=1

 

(a)
(b)
(a)
(b)
+ Xk + iXk Yk iYk

= 2

n
X
k=1

(a)

(a)

Xk Y k

57

(b)

(b)

+ Xk Y k ,

so one can make the choice

(a)

Ak = 2Xk ,

(a)

Bk = 2Yk ,

(b)

(b)

Ak = 2Xkn ,

Bk = 2Ykn ,

for k = 1, . . . , n
for k = n + 1, . . . , 2n

to arrive at (91).
Now we focus on the operators Ak , let us assume for the moment that
the spectrum of HA is discrete, and let | i be an eigenstate associate with
the eigenvalue . We define
X
Ak () =
| ih |Ak | ih |,
=

where the sum is over every and such that their difference is . The
operator so defined is an eigenoperator of LA = i[HA , ] with eigenvalue i;
indeed,
X
LA Ak () = i[HA , Ak ()] = i
( )| ih |Ak | ih | = iAk ().
=

As a result,

LA Ak () = iAk (),

with

Ak () = Ak (),

and this holds for every k. In addition, it is straightforward to verify that

[HA , Ak ()A ()] = 0.

(93)

On the other hand, to write Ak () and Ak () in the interaction picture

with respect to HA is quite easy:
A

eL t Ak () = eiHA t Ak ()eiHAt = eit Ak (),

A

eL t Ak () = eiHA t Ak ()eiHAt = eit Ak ().

Moreover, note that
X

Ak () =

| ih |Ak | ih | = Ak ,

Ak () = Ak .

and similarly that

58

Thus, by introducing these two decompositions for Ak in (91) and taking the
interaction picture, one gets
X
X
k (t) =
(t).
V (t) =
eit Ak () B
eit Ak () B
(94)
k
,k

,k

Now, coming back to the equation (90) the use of the above first equality
with V (s u) and the second one with V (s) yields
Z t XX

2
ds
A (t) = A (0) +
ei( )s sk ()[A ()
A (s), Ak ( )]
0

, k,

i( )s

+ e

s
A (s)Ak ()] + O(3 ).
k ()[A ( ),

Here we have introduced the quantities

Z s
h
i
s
iu

k () =
due Tr Bk (s)B (s u)B
Z0 s
h
i
k (u)B B ,
=
dueiu Tr B

(95)

(96)

where the last step is justified because B commutes with eiHB t . Next, we
take a rescaled time = 2 t and = 2 s, obtaining
A (t) = A ( /2 )
Z
= A (0) +

i( )/2

+ e

XX

/2

A (/2 ), Ak ( )]
ei( )/ k ()[A ()

, k,

/2

()[A ( ), A (/2 )Ak ()] + O().

From this equation we see that in the limit 0 (keeping finite and )
the time scale of change of A ( /2 ) is just . This is merely because the
dependency with 2 of A ( /2 ) enters only in the free evolution which we
remove by taking the interaction picture. The remaining time scale due to
the interaction with the environment is , as expressed in the above equation
(the integral runs from 0 to ), so we write
2

lim A (t) = lim eiHA / A ( /2 )eiHA / A ( ),

59

and thus
A ( ) = A (0) + lim

i( )/2

XX

/2

A (/2 ), Ak ( )]
ei( )/ k ()[A ()

, k,

/2

k ()[A ( ), A (/2 )Ak ()] + O()

Z
XX

2 /2
A (), Ak ( )]
= A (0) + lim
d
ei( )/ k ()[A ()
+ e

, k,

/2

+ ei( )/ k

()[A ( ), A ()Ak ()].

(97)

Before going further in the derivation, a mathematical result which will

be particularly useful from now on is the following.
Proposition 6.1 (Riemann-Lebesgue lemma). Let f (t) be integrable in [a, b]
then
Z b
lim
eixt f (t)dt = 0.
x

Proof. There are several methods to prove this, here we define g(t) = f (t)[(t
a) (t b)] and so
Z b
Z
ixt
e f (t)dt =
eixt g(t)dt = g(x),
a

where g(x) denotes the Fourier transform of g(t). Now it is evident that g(t)
is square integrable, then as the Fourier transform preserve norms [45] (this
is sometimes referred to as Parsevals theorem):
Z
Z
2
|g(t)| dt =
|
g (x)|2 dx,

which implies that g(x) is also square integrable and then limx g(x) = 0.
Q.E.D.
Because of this result (applied on the Banach space B, to be concise)
in the limit 0 the oscillatory factors in (97) with 6= are going
to vanish. This so-called secular approximation is done in the same spirit
that well known rotating wave approximation in the context of quantum
optics [20, 21, 26, 28, 34]. We obtain
Z
XX

A (), Ak ()]
A ( ) = A (0) +
d
k ()[A ()
0

k,

A ()Ak ()]
k ()[A (),
Z

A (0) +

60

d L [
A ()] .

(98)

Here the coefficients are given by the one-sided Fourier transform

Z
h
i

k (u)B B .
k () =
dueiu Tr B
0

6.2.2

Correlation functions

In order to make hconvergent the

i above integrals, the environmental correlak (u)B B have to decrease as u is increasing. However,
tion functions Tr B
if we use the corresponding eigenoperator decomposition of Bk for the Hamiltonian HB , we find
h
i X
k (u)B B =
Tr B
eiu Tr [Bk ()B B ] ,

which shows that the correlation functions are periodic in u and their integral extended to infinity will diverge. The only possible loophole to break
the periodicity is the assumption that the system B has infinite degrees of
freedom, in such a way that HB has a continuous spectrum. Then everything is essentially the same but the sums are substituted by integrals, i.e.
the decomposition in eigenoperators will be
Z a
Bk =
dBk (),
(99)
a

where a is the maximum eigenfrequency (it can be infinite). So the correlation

functions become
h
i Z a
k (u)B B =
Tr B
deiu Tr [Bk ()B B ] ,
a

and of course now they are not periodic as

h
i
k (u)B B = 0,
lim Tr B
u

because of the proposition (6.1).

6.2.3

Davies theorem

We have just seen that a necessary condition for the weak limit to exist is that
the environmental system B has infinite degrees of freedom. However this is
not a sufficient condition because the decreasing behavior of the correlation
functions may be not fast enough for the one-sided Fourier transform to exist.
A sufficient condition was pointed out by Davies.
61

Theorem 6.1 (Davies). If there exist some > 0 such that

Z  h
i



dt Tr Bk (t)B B  (1 t) < ,

(100)

then the weak coupling is strictly well defined (for a bounded interaction V )
and

lim 2 kE(,0) [A (0)] eL [A (0)] k1 = 0

0, = t

for all A (0) and uniformly in every finite interval [0, 0 ], where E(,0) denotes
the exact reduced dynamics in the interaction picture.
Proof. To prove the condition (100) is beyond the scope of this work, it was
developed in [106, 107]. Under (100) the convergency to the above result is
clear from equation (98) which is the integrated version of
d
A ( )

A ( ) A ( ) = eL
[
A (0)] .
= L
d
Q.E.D.
On one hand, note that the method works because there are no eigenfrequencies, , arbitrary close to one another, as a result of the discreteness
of the spectrum of HA , otherwise the secular approximation is jeopardized.
For the infinite dimensional case where this is not satisfied, recently D. Taj
and F. Rossi [113] have proposed another approximation method which substitutes the secular approximation and also leads to a completely positive
semigroup.
On the other hand, an important consequence of the secular approximation is that it guaranties that the generator L has the correct form of
a Markovian homogenous master equation (theorem 5.1). To see this, let
us decompose the matrices
k () in sum of Hermitian and anti-Hermitian
parts
1
(101)

k () = k () + iSk (),
2
where the coefficients
1

Sk () = [
k () k ()],
2i
and
k () =
() +
k ()
Z k
h
i Z
h
i
iu

k (u)B th
=
due Tr Bk (u)B th +
dueiu Tr B
0
Z0
h
i
k (u)B th ,
=
dueiu Tr B
(102)

62

form Hermitian matrices. In terms of these quantities, the generator L can

be written as
L [
A ( )] = i[HLS , A ( )]


XX
1

+
k () A ()
A ( )Ak () {Ak ()A (), A ( )} ,
2
k,
where the Hamiltonian part is given by
XX
HLS =
Sk ()Ak ()A ().

(103)

k,

In order to prove that L has the form of a generator of a Markovian homogenous evolution, it is required just to show that the matrix k () is positive
semidefinite for all . This is a consequence of the Bochners theorem [40,45],
which asserts that the Fourier transform of a function of positive type is a
positive quantity. Note that a function f (t) is of positive type if for any tm
and tn the matrix constructed as fmn = f (tm tn ) is positive semidefinite.
For any vector v we have
Z
h
i
X
X

k (u)B th v
vk Tr B
vk k ()v =
dueiu
(v, v) =

k,

where C =

dueiu

k,

Bk vk .

k,



vk Tr eiHB u Bk eiHB u B th v



dueiu Tr eiHB u C eiHB u Cth ,



Proposition 6.2. The function f (t) = Tr eiHB t C eiHB t Cth is of positive type and then the matrix k is positive semidefinite (v, v) 0.
Proof. The positivity of (v, v) 0 follows from the Bochners theorem if
f (t) is of positive type because it is its Fourier transform. For the proof of
the first statement, let us take the trace in the eigenbasis of HB (remember
that the spectrum is taken to be continuous):
Z
Z
iHB t iHB t
f (t) = dh |e
C e
Cth | i = deit p h |C eiHB t C| i,
here we have used that
th | i =

e
| i p | i,
Tr [e(HB ) ]
63

withRp 0 since th is a positive operator. By introducing the identity

1 = d | ih |,
Z
f (t) =
ddeit p h |C eiHB t | ih |C| i
Z

=
ddei( )t p |h |C| i|2 .
Finally we take another arbitrary vector w and the inner product
X

(w, fw) =
wm
f (tm tn )wn
m,n

dd

X
m,n

i( )(tm tn )
wm
e
wn p |h |C| i|2


2
X



2
i( )tm 
wm
e
dd 
 p |h |C| i| 0,


m

because the integrand is positive on the whole domain of integration.

Q.E.D.
As a conclusion of these results, if for a small, but finite , we substitute
2

the exact reduced dynamics E(t,0) by the Markovian one eL = e Lt , the error
which we make is bounded and tends to zero as decrease provided that the
assumptions of the Davies theorem are fulfilled.
Of course one can write 2 L and immediately go back to Schrodinger
picture by using the unitary operator eiHA t to obtain
dA (t)
= i[HA + 2 HLS , A (t)]
(104)
dt


XX
1

2
+
k () A ()A (t)Ak () {Ak ()A (), A (t)} .
2
k,
Note that because of (93), the Hamiltonian HLS commutes with HA , so it
just produces a shift in the energy levels of the system A. Thus this is
sometimes called Lamb shift Hamiltonian. However, note that HLS is not
only influenced by an environment in the vacuum state (zero temperature)
but also by thermal fluctuations.

64

6.2.4

Explicit expressions for k and Sk

We can give explicit expressions for k and Sk by using the decomposition

of the operators Bk in eigenoperators. For the decay rates
Z
h
i
k (u)B th
k () =
dueiu Tr B
Z
Z
a

=
d
duei( )u Tr [Bk ( )B th ]
a
Z a
= 2
d ( )Tr [Bk ( )B th ]
a

= 2Tr [Bk ()B th ] ,

(105)

provided that (a, a). So k () is just proportional to the correlation

function of the eigenoperator Bk () with the same frequency , unless this
frequency is outside of the spectrum of [HB , ], then k () will be zero.
To get an explicit expression for the shifts Sk is a little bit more complicated. From (101) we have
i
Sk () = i
k () + k ().
2
The substitution of the decomposition in eigenoperators in
k, () and the
previous result (105) for k,() yields
Z a
Z

Sk () = i
d
duei( )u Tr [Bk ( )B B ] + iTr [Bk ()B th ] .
a

(106)
R

Note that the integral 0 duei( )u in not well-defined. Similarly it hapR

pens with duei( )u , but we know this integral is proportional to the

delta function. The mathematical theory which provides a meaningful interpretation to these integrals is the theory of distributions or generalized
functions. Here we will not dwell on it, references [40, 41] provide introductions and further results can be found in [114, 115]. For us it is enough to
show that these integrals can have a meaning if we define them as a limit of
ordinary functions. For instance if we define
Z 0
Z
Z

i( )uu
i( )u
duei( )u+u ,
due
+ lim
due
= lim+

65

by making the elementary integrals,

Z
i
i

duei( )u = lim+
+ lim

0 ( ) + i
0 i ( )

i
i
+
= lim

0 ( ) + i
i ( )
2
= lim 2
2( ).
0 + ( )2

We identify the well-known family of functions f ( ) = [2+(

)2 ] which
tend to the delta function as 0 [117] because for any (regular enough)
integrable function g()
Z b
lim
d f ( )g( ) = g(), (a < < b).
0

So,
do the same procedure but just with the one-sided Fourier transform
R if wei(
)u
due
, we obtain in equation (106)
0
Z a
1
Tr [Bk ( )B B ] + iTr [Bk ()B th ] .
Sk () = lim+
d

0
( ) + i
a
Multiplying the integral by the complex conjugate of the denominator we
obtain
Z a
Z a

Tr [Bk ( )B B ]
( )
d
lim+
d
=
lim
Tr [Bk ( )B B ]
( )2 +2

+
0
0
( ) + i
a
a
Z
a
lim+ i
d ()2 +2 Tr [Bk ( )B B ]
0
Z aa
( )

d (
= lim+
)2 +2 Tr [Bk ( )B B ]
0

iTr [Bk ()B B ] .

Therefore the last term is canceled in the shifts, which read

Z a
( )

Sk () = lim+
d (
)2 +2 Tr [Bk ( )B B ] .
0

Now we decompose the integration interval (a, a) in three subintervals

(a, ), ( , + ) and ( + , a) and take the limit 0
Z
( )

d (
Sk () = lim+
)2 +2 Tr [Bk ( )B B ]
,0
Zaa
( )

d (
+ lim+
)2 +2 Tr [Bk ( )B B ]
,0
+
Z +
( )

+ lim+
d (
)2 +2 Tr [Bk ( )B B ] .
,0

66

The last integral just goes to zero because for any as small as we want
the integrand is an integrable function without singularities. Whereas, since
( )
1
lim0+ (
)2 +2 = ( ) , the limiting process in is just the definition of
the Cauchy principal value of the integral [117]:
Z

Z a

Tr [Bk ( )B B ]
Tr [Bk ( )B B ]
Sk () = lim+
d
+
d
0
( )
( )
a
+
Z a
Tr [Bk ( )B B ]
= P.V.
d
.
(107)
( )
a
In theory of distributions, this result is sometimes referred as Sochozkis
formulae [115].
6.2.5

Typical examples

It is not worth to dwell on examples of the application of the above results,

as in the books are plenty of them. Just for sake of comparison, we sketch
some of the most common cases.
Two-level system damped by a bath of harmonic oscillators. The total Hamiltonian is
H = Hsys + Hbath + V
Z max
Z max


0

dh() + a + a ,
da a +
z +
2
0
0

where z is the corresponding Pauli matrix, = (x iy )/2, and

a accounts for the continuous bosonic operators [a , a ] = (
); where is the frequency of each mode. The upper limit max is
the maximum frequency present in the bath of harmonic oscillators,
it could be infinity provided that the coupling function h() decreases
fast enough. We have taken the continuous limit from the beginning,
however it is common to do it in the course of the computation.
The interaction Hamiltonian may be written in the form V = A1
B1 + A2 B2 , where it is easy to check that the system operators and
their eigendecompositions are given by
A1 = x = + + , A1 (0 ) = , A1 (0 ) = + ,
A2 = y = i(+ ), A2 (0 ) = i , A2 (0 ) = i+ .
Similarly for the bath operators,
Z
Z max
(a +a )
=
dh() 2
B1 =

max

max

67

dB1 (),

with B1 () =

h()a
,
2

B1 () =

h()a
,
2

and
Z
Z max
i(a a )
=
dh() 2
B2 =
ih()a
,
2

dB2 (),

max

with B2 () =

max

B2 () =

ih()a
.
2

The decay rates are given by k,(0 ) = 2 Tr[Bk (0 )B B ], particularly

11 (0 ) = 2 Tr[B1 (0 )B1 B ]
Z max

dh() Tr[a0 (a + a )B ]
h(0 )
=
2
0

2
h (0 ) Tr[a0 a0 B ] = J(0 )[
n(0 ) + 1].
=
2
2
Here J(0 ) = h2 (0 ) is the so-called spectral density of the bath (which
accounts for the strength of the coupling per frequency) at frequency
0 and n
(0 ) is the mean number of bosons in the thermal state B of
the bath with frequency 0 ; i.e. the expected number of particles of
the Bose-Einstein statistics

1
n
(0 ) = e(0 /T ) 1
,
where T is the temperature of the bath. For the rest of the elements
one obtains essentially the same



1 i
(0 ) = J(0 )[
n(0 ) + 1]
.
i 1
2

Similarly for k, (0 ),

n(0 )
(0 ) = J(0 )
2

1 i
i 1

Let us make the analog for the shifts Sk ; the first element is given by
Z max
Tr [B1 ( )B1 B ]
d
S11 (0 ) = P.V.
(0 )
max
Z
max
Tr [Bk ( )B B ]
d
= P.V.
(0 )
Z0 max
Tr [Bk ( )B B ]
d
+ P.V.
(0 + )
0


Z max
1
n
( ) + 1
n
( )

=
d J( )
,
P.V.
+
4
(0 ) (0 + )
0
68

and the rest of the elements are

S22 (0 ) = S11 (0 ),
S12 (0 ) =

S21
(0 )

i
= P.V.
4

max
0


n
( )
n
( ) + 1
.
d J( )

(0 ) (0 + )

For the shifts with 0 :



Z max
1
n
( ) + 1
n
( )

S11 (0 ) = S22 (0 ) =
d J( )
,
P.V.
+
4
(0 + ) (0 )
0


Z max
i
n
( )
n
( ) + 1

S12 (0 ) = S21 (0 ) =
.
d J( )
P.V.
+
4
(0 + ) (0 )
0
Therefore according to (103) the shift Hamiltonian is
Z max
Z max

n( )
J( )
J( )
d
[
,

]
+
P.V.
+
d
HLS = P.V.
+

(0 )
(0 )
0
0
Z max
Z max

n( )
J( ) (z + 1)
J( )
d

+
P.V.
d
= P.V.
z
(0 )
(0 )
2
0
0



+
z ,
2
where is the integral depending on n
( ) which leads to a shift due
to the presence of the thermal field (Stark-like shift), and is the
second integral independent of n
( ) which lead to a Lamb-like shift
effect. Of course the identity 1 does not contribute to the dynamics as
it commutes with any , so we can forget it.
For the total equation we immediately obtain



dsys (t)
0 +

= i
+ z , sys (t)
dt
2


1
+ [
n(0 ) + 1] sys (t)+ {+ , sys (t)}
2


1
+
n(0 ) + sys (t) { + , sys (t)} ,
2
where = 2J(0 ).
Harmonic oscillator damped by a bath of harmonic oscillators. This case
is actually quite similar to the previous one, but the two-level system
69

is now substituted by another harmonic oscillator:

H = Hsys + Hbath + V
Z
Z max

da a +
0 a a +
0

max
0

dh() a a + aa .

The bath operators B1 and B2 are the same as above and the system
operators are now
A1 = a + a, A1 (0 ) = a, A1 (0 ) = a ,
A2 = i(a a), A2 (0 ) = ia, A2 (0 ) = ia .
Thus by mimicking the previous steps we obtain
dsys (t)
= i [(0 + ) z , sys (t)]
dt


1
+ [
n(0 ) + 1] asys (t)a {a a, sys (t)}
2


1

+
n(0 ) a sys (t)a {aa , sys (t)} .
2

Note in this case that, on one hand, the Stark-like shift does not
contribute because of the commutation rule of a and a ; and, on the
other hand, the dimension of the system is infinite and the interaction
with the environment is mediated by an unbounded operator. However as long as domain problems do not arise one expects that this is
the correct approximation to the dynamics (and actually it is, see for
instance [36]).
Pure dephasing of a two-level system. In this case the interaction commutes with the system Hamiltonian and so the populations of the initial
system density matrix remain invariant,
H = Hsys + Hbath + V
Z max
Z max


0

dh()z a + a .

da a +
z +
2
0
0

This is an interesting case to explore because it involves some subtleties,

but it is exactly solvable (see for example chapter 4 of [21], [116], and
references therein). The interaction Hamiltonian is decomposed as V =
A B, where A = A( = 0) = z (z is eigenoperator of [z , ] with
zero eigenvalue!) and

Z max
B() = h()a ,
dB(), with
for > 0.
B=
B() = h()a ,
max
70

So some problem arises when we calculate the decay rates because

B( = 0) is not well-defined. The natural solution to this is is to
understand = 0 as a limit in such a way that
(0) = 2 lim Tr[B()BB ],
0

but then there is a further problem. Depending on which side we approach to 0 the function Tr[B()BB ] takes the value 2J()[
n() + 1]
for > 0 or 2J(||)
n(||) for < 0. Thus the limit in principle would not be well-defined, except in the high temperature regime
where n
(||) n
(||) + 1. However both limits give the same result
provided that one assumes lim0 J(||) = 0. This is a natural physical assumption because a mode with 0 frequency does not have any
energy and influence on the system. Thus if lim0 J(||) = 0 we find
(0) = 2 lim J(||)
n(||).
0

Actually, since n
(||) goes to infinity as 1/|| when decreases, if
J(||) does not tend to zero linearly in (this is called Ohmic-type
spectral density [33]) this constant (0) will be either 0 or infinity. This
fact restricts quite a lot the spectral densities which can be treated
for the pure dephasing problem in the weak coupling framework with
significant results (which is not a major problem because this model is
exactly solvable as we have already pointed out). On the other hand,
shifts do not occur because HLS z z = 1, and finally the master
equations is
h
i
dsys (t)
0
= i
z , sys (t) + (0) [z sys (t)z sys (t)] .
dt
2
6.2.6

Some remarks on the secular approximation and positivity

preserving requirement

As we have seen, the weak coupling limit provides a rigorous mathematical

procedure to derive Markovian master equations. Nonetheless, in solid state
or chemical physics the interactions are typically stronger than for electromagnetic environments, and it is sometimes preferred to use the perturbative treatment without performing the secular approximation. However, this
method jeopardizes the positivity of the dynamics [118] (see also [105, 119])
and several tricks have been proposed to heal this drawback. For example, one possibility is taking in consideration just the subset of the whole
possible states of the system which remain positive in the evolution [120],
71

or describing the evolution by the inclusion of a slippage operator [121].

These proposals can be useful just in some situations; in this regard it seems
they do not work well for multipartite systems [122, 123].
We have already explained in detail the requirement of the complete positivity for a universal dynamics. The option to proceed without it may be
useful but risky, and actually the risk seems to be quite high as the positivity
of the density matrix supports the whole consistency of the quantum theory.
We believe the secular approximation should be used as it is mathematically
correct in the weak coupling limit. In case of the failure of a completely
positive semigroup to describe correctly the dynamics of the physical system, it is probably more appropriate the use of non-Markovian methods as
the dynamical coarse graining (see section 7.3) which preserves complete
positivity.
6.2.7

Failure of the assumption of factorized dynamics

In some derivations of the weak coupling limit it is common to use the following argument. One makes the formal integration of the von Neumann
equation in the interaction picture (88)
d
(t) = i[V (t), (t)],
dt
to give
(t) = (0) i

ds[V (s), (s)].

By iterating this equation twice one gets

Z t
Z t Z s
2

(t) = (0) i
ds[V (t), (0)]
ds
du[V (s), [V (u), (u)]].
0

After taking partial trace, under

h the usuali aforementioned assumptions of
(0) = A (0) B (0), and TrB V (t)B (0) = 0, this equation is simplified
to
Z t Z s
2
A (t) = A (0)
ds
du TrB [V (s), [V (u), (u)]].
0

Now, one argues that the system B is typically much larger than A and so,
provided that the interaction is weak, the state of B remains unperturbed by
the presence of the A, and the whole state can be approximated as (u)
A (u)B (u). By making this substitution in the above equation one arrives
at equation (90) (modulo some small details which one can consult in most
of the textbooks of the references), and henceforth proceed to derive the
72

Markovian master equation without any mention to projection operators,

etc.
This method can be considered as an effective model in order to obtain
equation (90), however it does not make any sense to assume that the physical
state of the system factorizes for all times. Otherwise how would it be possible
the interaction between both subsystems?. Moreover, why not making the
substitution (u) A (u) B (u) directly in the von Neumann equation
(88) instead of in its integrated and iterated version?.
The validity of the substitution (u) A (u) B (u) is justified because
of the factorized initial condition and the perturbative approach of the weak
coupling limit. Of course we pointed out previously that the weak coupling
approach makes sense only if the coupling is small and the environment has
infinite degrees of freedom. This fits in the above argument about the size
of B, but it can dismissed even by numerical simulations (see [36]) that
the real total state (u) is close to A (u) B (u). It should be therefore
be considered as an ansatz to arrive at the correct equation rather than a
physical requirement on the evolution.
6.2.8

Steady state properties

In this section we shall analyze the steady state properties of the Markovian master equation obtained in the weak coupling limit (104). The most
important result is that the thermal state of the system
th
A =

e(HA /T )
,
Tr[e(HA /T ) ]

with the same temperature T as the bath is a steady state. To see this one
notes that the bath correlation functions satisfies the so-called Kubo-MartinSchwinger (KMS) condition [124], namely
 iH (u+i)

1
iHB u
B
Tr
e
B
e
B
k

Tr (eHA )


1
iHB (u+i)
iHB (u+i) HB
=
Tr
B
e
B
e
e

k
Tr (eHA )


k (u + i)i. (108)
= Tr B eiHB (u+i) Bk eiHB (u+i) th = hB B

h
i

hBk (u)B i = Tr th Bk (u)B =

73

k in eigenoperators B( ) of [HB , ], we obtain, similarly

By decomposing B
to (105), the following relation between their Fourier transforms
Z
h
i
k (u)B th
k () =
dueiu Tr B
Z
h
i

k (u + i)th
=
dueiu Tr B B
Z
Z
a

=
d e
duei( )u Tr [B Bk ( )th ]
a

Z a

= 2
d e ( )Tr [B Bk ( )th ]
a
n h
io

Tr Bk ()B th
= 2e Tr [B Bk ()th ] = 2e
= 2e {Tr [Bk ()B th ]}

= e k
() = e k ().

(109)

On the other hand, since A( ) is an eigenoperator of [HA , ], one easily

obtains the relations

th
th
A Ak () = e Ak ()A ,

th
Ak ()th
A Ak () = e
A.

(110)
(111)

dth

Now one can check that dtA = 0 according to equation (104). Indeed, it is
obvious that th
A commutes with the Hamiltonian part, for the remaining one
we have


XX
1
dth
A
th
th
=
k () A ()A Ak () {Ak ()A (), A }
dt
2
k,
h
i
XX

th
=
k () e A ()Ak ()th

A
()A
()

A
A
k

k,

i
XXh
k ()e A ()Ak () k ()Ak ()A () th
=
A

k,

i
XXh

k ()A ()Ak () k ()Ak ()A () th

=
A = 0,

k,

as the sum in goes from max to max for some maximum difference
between energies max . This proves that th
A is a steady state. Of course this
does not implies the convergency of any initial state to it (an example is the
pure dephasing equation derived in section 6.2.5). Sufficient conditions for
that were given in the section 5.4.
74

6.3

Singular coupling limit

As explained in section 6.1, the singular coupling limit is somehow the complementary situation to the weak coupling limit, since in this case we get
B 0 by making the correlation functions to approach a delta function.
However, as a difference with the weak coupling limit, the singular coupling
limit is not so useful in practice because as its own name denotes, it requires
some singular situation. For that reason and for sake of comparison, we
sketch here the ideas about this method from the concepts previously defined.
For further details the reader will be referred to the literature.
There are several ways to implement the singular coupling limit, one
consists in rescaling the total Hamiltonian as [9, 21, 28]
H = HA + 2 HB + 1 V,
and perform the singular limit by taking 0. Starting from the equation
(83) under the assumptions of PV(t)P = 0 and (0) = A (0) B (0) we have
Z t Z s
1
P (t) = P(0) + 2
ds
duPV(s)G(s, u)QV(u)P (u).
(112)
0
0
For a moment, let us take a look to the expression at first order in the
expansion of G(s, u):
Z t Z s
1
P (t)
= P(0) + 2
ds
duPV(s)V(u)P (u) + O(3 ),
0
0
which we rewrite as
1
A (t) = A (0) 2

ds

h
i
du V (s)[V (u)
A (u) B ] + O(3 ).

By introducing the general form of the interaction Eq. (91),

Z t Z s
h
i

A (t) = A (0) +
ds
duCk (s, u) A (u)
A (u), Ak (s)
0
0
h
i

+ Ck (s, u) Ak (s), A (u)A (u) + O(3 ),

(113)

here the correlation functions are

Ck (s, u) =

1
k (s)B
(u)B ].
Tr[B
2

If we assume that B is again some eigenstate of the free Hamiltonian we get

Ck (s u) =

1
k (s u)B B ],
Tr[B
2
75

(114)

and by using the eigendecomposition (99) of Bk and taking into account the
factor 2 in the free Hamiltonian HB , we find
Z a
i(su)
1
Ck (s u) = 2
de 2 Tr[Bk ()B B ].
a
Because the proposition 6.1, in the limit 0 the above integral tends to
zero as a square integrable function in x = 2 (s u). It implies that it
tends to zero at least as 2 and so we expect this convergence to be faster
than the prefactor 12 in the correlation functions. However in the peculiar
case of s = u the correlation functions go to infinity, so indeed in this limit
they approach a delta function Ck (s u) (s u). Next, by making again
the change of variable u s u, equation (113) reads
Z t Z s
h
i

A (t) = A (0) +
ds
duCk(u) A (s u)
A (s u), Ak (s)
0
h0
i

+ Ck (u) Ak (s), A (s u)A (s u) + O(3 ).

Therefore in the limit of 0, we can substitute u by 0 in the A operators
and extend the integration to infinite without introducing an unbounded
error
Z t
h
i

A (t) = A (0) +
dsk A (s)
A (s), Ak (s)
0
h
i

+ k A (s), A (s)Ak (s) + O(3 ),

where

k =

duCk(u).

R
We may split this matrix in Hermitian k = 0 duCk (u) and anti-Hermitian
part iSk as in equation (101), and differentiate to obtain
d
A (t)
LS (t), (t)]
= i[H
dt

n
o
1
+ k A (t)
A (t)Ak (t)
Ak (t)A (t), A (t) + O(3 ),
2
P
LS (t) =

where H
k, Sk Ak (t)A (t). Since k has the same expression as
in (102) with = 0 it is always a positive semidefinite matrix and up to
second order the evolution equation has the form of a Markovian master
equation. However, from the former it is not clear what happens with the
76

higher order terms; a complete treatment of them can be found in [108110],

where it is proven that they vanish. The idea behind this fact is that by
using the property that for a Gaussian state (which B is assumed to be)
the correlation functions of any order can be written as a sum of products
of two time correlation functions (114). Particularly odd order correlation
functions vanish because of the condition PV(t)P = 0, and we obtain a
product of deltas for even order. Actually the expansion is left only with
those products in which some time arguments appear in overlapping order
such as
(t1 t3 )(t2 t4 ),
for t4 t3 t2 t1 , and because the time-ordered integration in the
expansion of G(s, u) they do not give any contribution.
Finally by coming back to Schrodinger picture we obtain that in the
singular coupling limit
dA (t)
= i[HA + HLS , (t)]
dt


1
+ k A A (t)Ak {Ak A , A (t)} ,
2
P
note that HLS = k, Sk Ak A does not commute in general with HA .
Despite the singular coupling limit is not very realistic, it is possible to
find effective equations for the evolution in which self-adjoint operators Ak
also appear. For example under interactions with classic stochastic external
fields [108] or in the study of the continuous measurement processes [21,125].

6.4

Extensions of the weak coupling limit

Here we will discuss briefly some extensions of the weak coupling limit for
interacting systems, the reader can consult further studies as [36] for example.
6.4.1

Weak coupling for multipartite systems

Consider two equal (for simplicity) quantum systems and their environment,
such that the total Hamiltonian is given by
H = HA1 + HA2 + HE + V.
Here HA1 (HA2 ) is the free Hamiltonian of the subsystem 1 (2), HE is the
free Hamiltonian of the environment and V denotes the interaction between
the subsystems andPthe environment. This interaction term is assumed to
(1)
(2)
(1)
(2)
have the form V = k Ak Bk + Ak Bk , where Ak (Ak ) are operators
77

acting on the subsystem 1 (2), and Bk on the environment. So that the

weak coupling limit for the whole state of both subsystems 12 leads to a
Markovian master equation like
d12 (t)
= i[HA1 + HA2 + 2 HLS , 12 (t)]
(115)
dt


XX
1

2
+
k () A ()12 (t)Ak () {Ak ()A (), 12(t)} ,
2

k,

(1)

(2)

where Ak () are linear combinations of Ak and Ak decomposed in eigenoperators of [HA1 +HA2 , ]. Of course the form of these operators depend on the
form of the coupling. For instance, suppose thatP
we have two identical envi(1)
(1)
(2)
(2)
ronments acting locally on each subsystem V = k Ak Bk + Ak Bk ,
then we will obtain
2

X
d12 (t)
(j)
= i
[HA1 + HA2 + 2 HLS , 12 (t)]
dt
j=1


X X (j)
1 (j)
(j)
(j)
(j)
2
+
k () A ()12 (t)Ak () {Ak ()A (), 12 (t)} .
2
k,
Consider now the case in which we perturb the free Hamiltonian of the
subsystems by an interaction term between them of the form V12 (where
accounts for the strength). Instead of using the interaction picture with
respect to the free evolution HA1 +HA2 +HB , the general strategy to deal with
this problem is to consider the interaction picture including the interaction
between subsystems HA1 + HA2 + V12 + HB , and then proceeds with the
weak coupling method.
However if is small, one can follow a different route. First, taking the
interaction picture with respect to HA1 + HA2 + HB , the evolution equation
for the global system is
d
(t) = i[V12 (t), (t)] i[V (t), (t)] V12 (t)
(t) + V(t)
(t).
dt
Similarly to section 6.1, we take projection operators
d
P (t)

= PV12 (t)
(t) + PVSB (t)
(t),
dt
d
Q
(t) = QV12 (t)
(t) + QVSB (t)
(t),
dt

78

(116)
(117)

and find the formal solution to the second equation

Z t
Q
(t) = G(t, 0)Q
(0) +
dsG(t, s)QV12 (s)P (s)
0
Z t
+
dsG(t, s)QV(s)P (s),

(118)

where

Rt

G(t, s) = T e

dt Q[V12 (t )+V(t )]

Now the procedure is as follows, we introduce the identity 1 = P + Q just

in the last term of equation (116),
d
P (t) = PV12 (t)
(t) + PV(t)P (t) + PV(t)Q
(t),
dt
whose formal integration yields
Z t
Z t
P (t) = P(0) +
dsPV12 (s)
(s) +
dsPV(s)Q
(s),
0

(119)

here we have used the condition PV(t)P = 0. Next we insert the formal
solution (118) into the last term. By assuming again an initial factorized
state subsystemsenviroment (Q(t0 ) = 0) we find
Z t
P (t)
= P(0) +
dsPV12 (s)
(s)
0
Z t Z s
Z t Z s
+
ds
duK1 (s, u)P (u) +
ds
duK2 (s, u)P (u),
0

where the kernels are

K1 (s, u) = PV(s)G(s, u)QV12 (u)P = 0,
K2 (s, u) = 2 PV(s)G(s, u)QV(u)P.
The first one vanishes because V12 (s) commutes with P and QP = 0, and
the second kernel up to second order in and becomes
K2 (s, u) = 2 PV(s)QV(u)P + O(3 , 2 ) = 2 PV(s)V(u)P + O(3 , 2 ).
Therefore the integrated equation of motion reads
Z t
P (t)
= P(0) +
dsPV12 (s)
(s)
0
Z t Z s
2
+
ds
duPV(s)V(u)P (u) + O(3 , 2).
0

79

If we consider small intercoupling such that & , we can neglect the

higher orders and the weak coupling procedure of this expression (c.f. section
6.2) will lead to the usual weak coupling generator with the Hamiltonian part
V12 added. This is in Schrodinger picture
d12 (t)
= i[HA1 + HA2 + V12 + 2 HLS , 12 (t)]
(120)
dt


XX
1

2
k () A ()12 (t)Ak () {Ak ()A (), 12(t)} .
+
2
k,
Compare equations (115) and (120).
Surprisingly, this method may provide good results also for large under
some conditions, see [36].
6.4.2

Weak coupling under external driving

Let us consider now a system A subject to an external time-dependent perturbation Hext (t) and weakly coupled to some environment B, in such a
way that the total Hamiltonian is
H = HA + Hext (t) + HB + V.
In order to derive a Markovian master equation for this system we must take
into account of a couple of details. First, since the Hamiltonian is timedependent the generator of the master equation will also be time-dependent,
dA (t)
= Lt A (t),
dt
whose solution defines a family of propagators E(t2 ,t1 ) such that
S (t2 ) = E(t2 ,t1 ) S (t1 ),
E(t3 ,t1 ) = E(t3 ,t2 ) E(t2 ,t1 ) .
This family may be contractive (i.e. Markovian inhomogeneous) or just eventually contractive (see section 2.4). Secondly, there is an absence of rigorous
methods to derive at a Markovian master equation (i.e. contractive family)
in the weak coupling limit when the system Hamiltonian is time-dependent,
with the exception of adiabatic regimes of external perturbations [126, 127].
Fortunately if the Hext (t) is periodic in t, it is possible to obtain Markovian
master equations, even though the complexity of the problem increases.

80

On one hand, if is small, in the spirit of the previous section we can

expect that the introduction of Hext (t) just in the Hamiltonian part of the
evolution would provide a good approximation to the motion,
dA (t)
= i[HA + Hext (t) + 2 HLS , A (t)]
(121)
dt


XX
1

2
k () A ()A (t)Ak () {Ak ()A (), A (t)} ,
+
2
k,
which has the form of a Markovian master equation (46).
On the other hand, for larger the only loophole seems to work in the
interaction picture generated by the unitary propagator
U(t1 , t0 ) = T ei

R t1
t0

[HA +Hext (t )]dt

Taking t0 = 0 without loss of generality, the time evolution equation for

(t) = U (t, 0)(t)U(t, 0) is
d
(t)
= i[V (t), (t)].
(122)
dt
Following an analogous procedure which was used in section 6.2 for timeindependent generators, one immediately deals with the problem that it is
not clear whether there exists something similar to the eigenoperator decomposition of V (t) = U (t, 0)V U(t, 0) as in (94). Note however that since the
dependency of the operator Hext (t) with t is periodical, by differentiation of
U(t, 0) one obtains a differential problem for each Ak in (91)
dAk (t)
= i[Ak (t), HA + Hext (t)].
(123)
dt
with periodic terms. This kind of equations can be studied with the wellestablished Floquet theory (see for example [103, 104]). Particularly it is
possible to predict if its solution is a periodic function. In such a case, the
operator P
in the new picture will have a formal decomposition similar to (94),

Ak (t) = Ak ()eit , where now Ak () are some operators which do not

necessarily satisfy a concrete eigenvalue problem. However note that the
importance of such a decomposition is that the operators Ak () are timeindependent. This allows us to follow a similar procedure to that used for
time-independent Hamiltonians with the secular approximation, and we will
obtain
d
A (t)
= i[2 HLS , A (t)]
dt


XX
1

2
+
k () A ()
A (t)Ak () {Ak ()A (), A (t)} ,
2
k,
81

P P

where HLS =

k, Sk ()Ak ()A () of course. By coming back to

Schrodinger picture,

d
A (t)
= i[HA + Hext (t) + 2 HLS (t), A (t)]
dt


XX
1

2
+
k () A (, t)
A (t)Ak (, t) {Ak (, t)A (, t), A (t)} ;
2
k,

here Ak (, t) = U(t, 0)Ak ()U (t, 0). For these manipulations it is very
useful to decompose (when possible) the unitary propagator as product of

non-commuting exponential operators U(t, 0) = eiH0 t eiHt , where the trans removes the explicit time-dependence of the
formation eiH0 t H(t)eiH0 t = H
Hamiltonian.
A more systematic approach to this method based on the Floquet theory
can be found for instance in the work of H. -P. Breuer and F. Petruccione
[21, 128], and in [129].
There is a very important thing to note here. As a difference to the
usual interaction picture with respect to time-independent Hamiltonians,
the change of picture given by a time-dependent Hamiltonian can affect the
strength of V in a non-trivial way. If there exist some time, t = tc
such that kHext (tc )k = , the validity of the weak coupling limit may be
jeopardized. For example one may obtain things like 1/0 inside of V (t).

Microscopic description: non-Markovian case

The microscopic description of non-Markovian dynamics is much more involved than the Markovian one, and developing efficient methods to deal
with it is actually an active area of research nowadays. One of the reasons
for this difficulty is that the algebraic properties like contraction semigroups
and/or evolution families are lost and more complicated structures arise (e.g.
evolution families which are only eventually contractive). We have already
mentioned the advantages of being consistent with the UDM description if
we start from a global product state. According to section 4.3 if the evolution family E(t,s) describes the evolution from s to t and the initial global
product state occurs at time t0 , then it must be eventually contractive from
that point, i.e. kE(t,t0 ) k 1 for t t0 . However it is possible that kE(t,s) k > 1
for arbitrary t and s. Up to date, there is not a clear mathematical characterization of evolution families which are only eventually contractive and
this makes it difficult to check whether some concrete model is consistent.
There are non-Markovian cases where the environment is made of a few
degrees of freedom (see for example [130132], and references therein). This
82

allows one to make efficient numerical simulations of the whole closed system
and after that, just tracing out the environmental degrees of freedom, we
obtain the desired evolution of the open system.
However when the environment is large, a numerical simulation is completely inefficient, unless the number of parameters involved in the evolution
can be reduced (for example in case of Gaussian states [36] or using numerical renormalization group procedures [133, 134]). For remaining situations,
the choice of a particular technique really depends on the context. For concreteness, we sketch just a few of them in the following sections.

7.1

Integro-differential models

As we have seen in section 6.1 the exact evolution of a reduced system can
be formally written as an integro-differential equation (86),
Z t
d
P (t)
=
duK(t, u)P (u),
dt
0
with a generally very complicated kernel
K(t, u) = PV(t)G(t, u)QV(u).
Typically this kernel cannot be written in a closed form and, furthermore,
the integro-differential equation is not easily solvable. On the other hand, it
is possible to perform a perturbative expansion of the kernel on the strength
of V(t) and proceed further than in the Markovian case, where the series
is shortened at the first non-trivial order (for an example of this see [135]).
Of course, by making this the complete positivity of the reduced dynamics
is not usually conserved; and one expects a similar accuracy for the timeconvolutionless method (at the same perturbative order) explained in the
next section, which is simpler to solve.
As an alternative to the exact integro-differential equation, several phenomenological approaches which reduce to Markovian evolution in some limits has been proposed. For example, if L is the generator of a Markovian
semigroup, an integro-differential equation can be formulated as
Z t
d(t)
=
k(t t )L[(t )],
(124)
dt
0
where k(t t ) is some function which accounts for memory effects, and
simplifies the complicated Nakajima-Zwanzig kernel. Of course, in the limit
k(tt ) (tt ) the Markovian evolution is recovered. The possible choices
83

for this kernel which assure that the solution, E(t,0) , is a UDM have been
studied in several works [136138], being the exponential ansazt k(t t )

geg(tt ) the most popular.

Another interesting phenomenological integro-differential model is the socalled post-Markovian master equation, proposed by A. Shabani and D. A.
Lidar [139],
Z t
d(t)

=L
k(t t )eL(tt ) [(t )].
(125)
dt
0
It also approaches the Markovian master equation when k(t t ) (t t ).
Further studies about the application of these models and the conditions to
get UDMs can be found in [140142].
Apart from these references, the structure of kernels which preserve complete positivity has been analyzed in [143].
We should note however that, it has been recently questioned whether
these phenomenological integro-differential equations reproduce some typical
features of non-Markovian dynamics [144].

7.2

Time-convolutionless forms

The convolution in integro-differential models is usually an undesirable characteristic because, even when the equation can be formulated, the methods
to find its solution are complicated. The idea of the time-convolutionless
(TCL) forms consist of removing this convolution from the evolution equation to end up with an ordinary differential equation. Behind this technique
lies the property already expressed in section 4.5 that a UDM, E(t,t0 ) , can be
expressed as the solution of the differential equation (40)
dA (t)
= Lt [A (t)],
dt
dE

1
0)
with generator Lt = (t,t
E(t,t
. In general this equation generates an evoludt
0)
tion family that is not contractive, i.e. kE(t2 ,t1 ) k > 1, for some t1 and t2 and
so non-Markovian. However, it generates an eventually contractive family
from t0 , kE(t,t0 ) k 1, as the map starting from t0 , E(t,t0 ) , is usually assumed
to be a UDM by hypothesis.
To formulate this class of equations from a microscopic model is not
simple, although it can been done exactly in some cases such as, for a damped
harmonic oscillator [145,146] or for spontaneous emission of a two level atom
[147,148]. In general, the recipe is resorting to a perturbative approach again.
Details of this treatment are carefully explained in the original work [149]
or in some textbooks (see [20, 21] for instance). The key ingredient is to

84

 t) where s t defined in (89),

insert the backward unitary operator U(s,
t)
(s) = U(s,
(t), into the formal solution of the Q
(t), Eq. (82),
Z t
Q
(t) = G(t, 0)Q
(0) +
dsG(t, s)QV(s)P U (s, t)
(t).
0

t) and (t), and

Now, by introducing the identity 1 = P + Q between U(s,
after considering a factorized initial condition we obtain
[1 (t)]Q
(t) = (t)P (t),
here
(t) =

t),
dsG(t, s)QV(s)P U(s,

where we have rewritten V as V.

Next, note that (0) = 0 and (t)|=0 = 0, so that the operator [1(t)]
may be inverted at least for not too large couplings and/or too large t. In
such a case we would obtain
Q
(t) = [1 (t)]1 (t)P (t).

Introducing this in the second member of equation (79), the first term vanishes as PV(t)P = 0, and we arrive at the differential equation
dP(t)
= Lt [P(t)],
dt
where

Lt = PV(t)[1 (t)]1 (t)P.

We can formally write the inverse operator as a geometric series

[1 (t)]1 =

n (t),

n=0

and then the generator as

Lt = PV(t)

n (t)P.

n=1

On the other hand, n (t) can be expanded in powers of by introducing the

t), thus we can construct successive approxiexpansions of G(t, s) and U(s,
mations to Lt in powers of ,
X
(n)
Lt =
n Lt .
n=2

85

Here the sum starts from n = 2 because at lowest order (t)

and
Z t
Z t
(2)

Lt =
dsPV(t)QV(s)P =
dsPV(t)V(s)P,
0

Rt
0

dsQV(s)P,

which leads actually to the differentiated version of equation (90), as expected

at lowest order. After a little bit of algebra, the next term in the expansion
turns out to be
Z t Z t
(3)
Lt =
ds
dt1 PV(t)V(t1 )V(s)P,
0

and further terms can be calculated in the same fashion.

In addition, D. Chrusci
nski and A. Kossakowski [150] have recently shown
the equivalence between integro-differential equations and TCL equations.
They connect both generators by means of a Laplace transformation showing
some kind of complementarity between both methods. That is, if the integral
kernel is simple, the TCL generator is highly singular, and viceversa.
On the other hand, given a TCL equation and an initial condition at
time t0 , the problem of knowing whether its solution E(t,t0 ) is a UDM, i.e. a
completely positive map, is an open question for the non-Markovian regime.
Note that the perturbative expansion at finite order of the TCL generator
can generally break the complete positivity (in the same fashion as a finite
perturbative expansion of a Hamiltonian dynamics breaks the unitary character). So the classification of TCL generators which are just eventually
contractive from a given time t0 is an important open problem in the field
of dynamics of open quantum systems. Partial solutions have been given in
the qubit case [151] and in the case of commutative generators Lt [152, 153].
For this latter one [Lt1 , Lt2 ] = 0, for all t1 and t2 , so that the time-ordering
operator has no effect and the solution is given by
Rt

E(t,t0 ) = e

t0

Lt dt

Therefore, certainly E(t,t0 ) is completely positive if and only if

the standard form (51) for every final t.

7.3

Rt

t0

Lt dt has

Dynamical coarse graining method

The dynamical coarse-graining method was recently proposed by G. Schaller

and T. Brandes [154, 155]. Thiswas suggested as away to avoid the secular
approximation under weak coupling while keeping the complete positivity in
the dynamics. Later on, this method has been used in some other works
[156, 157]. Here we briefly explain the main idea and results.
86

Consider the total Hamiltonian (71) and the usual product initial state
(0) = A (0) B , where B is a stationary state of the environment. In the
interaction picture the reduced state at time t is


A (t) = TrB U(t, 0)A (0) B U (t, 0) ,
(126)
where

U(t, 0) = T ei

Rt
0

V (t )dt

is the unitary propagator, and again accounts for the strength of the interaction. Next we introduce the expansion of the propagator up to the first
non-trivial order in equation (126)
Z t
Z t
h
h
ii
2

A (t) = A (0) T
dt1
dt2 TrB V (t1 ), V (t2 ), A (0) B + O(3 ),
2
0
0
here we have assumed again that the first order vanishes Tr[V (t)B ] = 0 like
in (85). Let us take a closer look to structure of the double time-ordered
integral
Z t
Z t
h
h
ii
T
dt1
dt2 TrB V (t1 ), V (t2 ), A (0) B
Z0 t
Z0 t
h
h
ii
=
dt1
dt2 (t1 t2 ) TrB V (t1 ), V (t2 ), A (0) B
Z0 t
Z0 t
h
h
ii
+
dt1
dt2 (t2 t1 ) TrB V (t2 ), V (t1 ), A (0) B

0
0
(2)
L1 [A (0)]

(2)

(2)

+ L2 [A (0)] + L3 [A (0)],

where by expanding the double commutators we have found three kind of

(2)
(2)
(2)
terms, L1 , L2 and L3 . The first one appears twice, it is
(2)

L1
[A (0)] =
2

h
i
dt2 (t1 t2 ) TrB V (t1 )A (0) B V (t2 )
0
0
h
i
+(t2 t1 ) TrB V (t1 )A (0) B V (t2 )
Z t
Z t
=
dt1
dt2 [(t1 t2 ) + (t2 t1 )]
0
0
h
i
TrB V (t1 )A (0) B V (t2 )
Z t
Z t
h
i
=
dt1
dt2 TrB V (t1 )A (0) B V (t2 ) ,
0

dt1

87

because obviously the step functions cancel each other. Taking in account
the other analogous term (that is the corresponding to interchange t1 t2
in the double commutators) we can write the total contribution as
(2)

L1 [A (0)] = 2 TrB [A (0) B ] ,

Rt
where = 0 V (t )dt .
The second term goes like
Z t
Z t
h
i
(2)
L2 [A (0)] =
dt1
dt2 (t1 t2 ) TrB V (t1 )V (t2 )A (0) B
0
0
h
i
+(t2 t1 ) TrB V (t2 )V (t1 )A (0) B
Z t
Z t
=
dt1
dt2 [(t1 t2 ) + (t2 t1 )]
0
0
h
i
TrB V (t2 )V (t1 )A (0) B
n
o
+(t1 t2 ) TrB [V (t1 ), V (t2 )]A (0) B


= TrB 2 A (0) B
Z t
Z t
n
o
+
dt1
dt2 (t1 t2 ) TrB [V (t1 ), V (t2 )]A (0) B .
0

And similarly the remaining term can be expressed as



(2)
L3 [A (0)] = TrB A (0) B 2
Z t
Z t
n
o

dt1
dt2 (t1 t2 ) TrB A (0) B [V (t1 ), V (t2 )] .
0

Thus, everything together gives A (t) A (0) + Lt [A (0)] + O(3 ) with


2  (2)
(2)
(2)
L1 [A (0)] + L2 [A (0)] + L3 [A (0)]
L [A (0)] =
2
t
i[HLS
, A (0)] + D t [A (0)].
t

t
Here the self-adjoint operator HLS
is
Z
Z t
n
o
2 t
t
HLS =
dt1
dt2 (t1 t2 ) TrE [V (t1 ), V (t2 )]B ,
2i 0
0

and the dissipative part is given by




1 2
t
2
D [A (0)] = TrB A (0) B
, A (0) B .
2
88

By using the spectral decomposition of B it is immediate to check that for

any fixed t, Lt has the Kossakowski-Lindblad form (51).
Since at first order
t

A (t) = (1 + Lt )A (0) + O(3 ) eL A (0),

(127)

the dynamical coarse graining method consists of introducing a coarse

graining time and defining a family of generators of semigroups by
L
L
.

And the corresponding family of differential equations

A (t) = L A (t),
which solution

A (t) = etL A (0),

(128)

will obviously coincide with (128) up to second order when taking = t, this
=t
is A
(t) = A (t).
Therefore we have found a family of semigroups (127) which have the
Kossakowski-Lindblad form for every and give us the solution to the evolution for a particular value of ( = t). This implies that complete positivity
is preserved as we desire. In addition, one may want to write this evolution
as a TCL equation, i.e. to find the generator Lt such that the solution of
d
A (t) = Lt [A (t)]
dt
fulfills A (t) = tA (t). That is given as in (40),




d Lt Lt
d Lt t Lt t
e
=
e
e
e .
Lt =
dt
dt
Particularly for large times t , this generator Lt approach the usual
weak coupling generator (104) (in the interaction picture). This is expected
because the upper limit of the integrals in the perturbative approach will

approach to infinity and the non-secular terms ei( )t will disappear for
large t (because of proposition 6.1). This is actually equivalent to what we
obtain by taking the limit 0 in the rescaled time. Details are found
in [154].

89

Conclusion

We have reviewed in some detail the mathematical structure and the main
physical features of the dynamics of open quantum systems, and presented
the general properties of the dynamical maps underlying open system dynamics. In addition, we have analyzed some mathematical properties of quantum
Markovian processes, such as their differential form or their steady states.
After that we have focused our attention on the microscopic derivations of
dynamical equations. First of all we have analyzed the Markovian case, recalling the weak and singular coupling procedures, as well as some of their
properties; and secondly, we have discussed briefly some methods which go
beyond Markovian evolutions.
Several interesting topics have not been tackled in this article. Examples
are influence functional techniques [21,23,33,158,159], stochastic Schrodinger
equations [15, 26, 160163], or important results such as the quantum regression theorem, which deals with the dynamics of multi-time time correlation
functions [15, 23, 26, 164, 165]. Moreover, recently several groups have addressed the important problem of introducing a measure to quantify the
non-Markovian character of quantum evolutions [166173]. However, despite
the introductory nature of this work, we hope that it will be helpful to clarify
certain concepts and derivation procedures of dynamical equations in open
quantum systems, providing a useful basis to study further results as well as
helping to unify the concepts and methods employed by different communities.

Acknowledgments
We thank Ramon Aguado, Fatih Bayazit, Tobias Brandes, Heinz-Peter Breuer,
Daniel Burgarth, Alex Chin, Dariusz Chrusci
nski, Animesh Datta, Alberto
Galindo, Pinja Haikka, Andrzej Kossakowski, Alfredo Luis, Sabrina Man
iscalco, Laura Mazzola, Jyrki Piilo, Miguel Angel
Rodrguez, David Salgado, Robert Silbey, Herbert Spohn, Francesco Ticozzi, Shashank Virmani
and Michael Wolf for fruitful discussions. And we are specially grateful to
Bruce Christianson, Adolfo del Campo, Ines de Vega, Alexandra Liguori,
and Martin Plenio for careful reading and further comments. This work
was supported by the STREP project CORNER, the Integrated project QESSENCE, the project QUITEMAD S2009-ESP-1594 of the Consejera de
Educacion de la Comunidad de Madrid and MICINN FIS2009-10061.

90

References
[1] W. H. Louisell, Quantum Statistical Properties of Radiation, John Wiley
& Sons Inc, New York, 1973.
[2] G. S. Agarwal, Master equation methods in quantum optics, in: E. Wolf
(ed.), Progress in Optics Vol. 11, Elsevier, Amsterdam, 1973, 1.
[3] F. Haake, Statistical treatment of open systems by generalized master
equations. Springer Tracts in Modern Physics 66 (1973), 98.
[4] E. B. Davies, Quantum Theory of Open Systems, Academic Press, London, 1976.
[5] V. Gorini, A. Frigerio, M. Verri, A. Kossakowski and E. C. G. Sudarshan,
Rep. Math. Phys. 13 (1978), 149.
[6] H. Spohn, Rev. Mod. Phys. 53 (1980), 569.
[7] K. Kraus, States, Effects, and Operations Fundamental Notions of
Quantum Theory, Lect. Notes Phys. 190, Springer, Berlin, 1983.
[8] A. J. Leggett, S. Chakravarty, A. T. Dorsey, M. P. A. Fisher, A. Garg
and W. Zwerger, Rev. Mod. Phys. 59 (1987), 1.
[9] R. Alicki and K. Lendi, Quantum Dynamical Semigroups and Applications, Lect. Notes Phys. 286, Springer, Berlin, 1987.
[10] C. Cohen-Tannoudji, J. Dupont-Roc and G. Grynberg, Atom-Photon
Interactions, John Wiley & Sons, New York, 1992.
[11] H. Carmichael, An Open Systems Approach to Quantum Optics,
Springer, Berlin, 1993.
[12] A. Bohm, H. -D. Doebner and P. Kielanowski (eds.), Irreversibility and
causality. Semigroups and rigged Hilbert spaces, Lect. Notes Phys. 504,
Springer, Berlin, 1996.
[13] K. Blum, Density Matrix Theory and Applications, Plenum Press, New
York, 1996.
[14] C. W. Gardiner, Handbook of Stochastic Methods, Springer, Berlin, 1997.
[15] H. J. Carmichael, Statistical Methods in Quantum Optics (2 vols),
Springer, Berlin, 1999.
91

[16] R. Zwanzig, Nonequilibrium Statistical Mechanics, Oxford University

Press, New York, 2001.
[17] R. R. Puri, Mathematical Methods of Quantum Optics, Springer, Berlin,
2001.
[18] R. Alicki and M. Fannes, Quantum Dynamical Systems, Oxford University Press, New York, 2001.
[19] G. Lindblad, Non-Equilibrium Entropy and Irreversibility, Reidel Publishing Company, Dordrecht, 2001.
[20] B. Fain, Irreversibilities in Quantum Mechanics, Kluwer Academic Publishers, New York, 2002.
[21] H. -P. Breuer and F. Petruccione, The Theory of Open Quantum Systems, Oxford University Press, Oxford, 2002.
[22] B.-G. Englert and G. Morigi, Five Lectures on Dissipative Master Equations, in: A. Buchleitner and K. Hornberger, (eds.), Coherent Evolution
in Noisy Environments, Lect. Notes Phys. 611, 55, Springer, Berlin,
2002.

[23] T. Brandes, Quantum Dissipation http://wwwitp.physik.tu-berlin.de/brandes/public html/p

(2003).
[24] F. Benatti and R. Floreanini (eds.), Irreversible Quantum Dynamics,
Lect. Notes Phys. 622, Springer, Berlin, 2003.
[25] E. Joos, H. D. Zeh, C. Kiefer, D. J. W. Giulini, J. Kupsch, I. -O. Stamatescu, Decoherence and the Appearance of a Classical World in Quantum Theory, Springer, Berlin, 2003.
[26] C. W. Gardiner and P. Zoller, Quantum Noise, Springer, Berlin, 2004.
[27] V. May and O. Kuhn, Charge and Energy Transfer Dynamics in Molecular Systems, Wiley-VCH, Weinheim, 2004.
[28] F. Benatti and R. Floreanini, Int. J. Mod. Phys. B 19 (2005), 3063.
[29] S. Attal, A. Joye and C. -A. Pillet (eds.), Open Quantum Systems (3
vols.), Springer, Berlin, 2006.
[30] A. Nitzan, Chemical Dynamics in Condensed Phases: Relaxation,
Transfer, and Reactions in Condensed Molecular Systems, Oxford University Press, New York, 2006.
92

[31] Y. Tanimura, J. Phys. Soc. Jpn. 75 (2006), 082001.

[32] M. A. Schlosshauer, Decoherence and the Quantum-to-Classical Transition, Springer, Berlin, 2007.
[33] U. Weiss, Quantum Dissipative Systems, World Scientific, Singapore,
2008.
[34] S. Kryszewski and J. Czechowska-Kryszk, Master equation - tutorial
approach, arXiv:0801.1757 (2008).
[35] K. Hornberger, Introduction to Decoherence Theory, in: A. Buchleitner,
C. Viviescas and M. Tiersch, (eds.), Entanglement and Decoherence,
Lect. Notes Phys. 768, 221, Springer, Berlin, 2009.
[36] A. Rivas, A. D. K. Plato, S. F. Huelga and M. B. Plenio, New J. Phys.
12 (2010), 113032.
[37] C. H. Fleming and B. L. Hu, Ann. Phys. (N.Y.) 327 (2012), 1238.
[38] M. A. Nielsen and I. L. Chuang, Quantum Computation and Quantum
Information, Cambridge University Press, Cambridge, 2000.
[39] E. Kreyszig, Introductory Functional Analysis with Applications, Wiley,
New York, 1978.
[40] M. Reed and B. Simon, Methods of Modern Mathematical Physics I,
Academic Press, San Diego, 1980.
[41] N. Bocccara, Functional Analysis: An Introduction for Physicists, Academic Press, 1990.
[42] A. Galindo and P. Pascual, Problemas de Mecanica Cuantica, Eudema,
Madrid, 1989.
[43] M. Spivak, Calculus, Cambridge University Press, Cambridge, 1994.
[44] K. Yosida, Functional Analysis, Springer, Berlin, 1980.
[45] M. Reed and B. Simon, Methods of Modern Mathematical Physics II,
Academic Press, San Diego, 1975.
[46] K. -J. Engel and R. Nagel, One-Parameter Semigroups for Linear Evolution Equations, Springer, New-York, 2000.

93

[47] The most simple example is in one dimension, take nowhere differentiable function f (t) and define T(t,s) = f (t)/f (s) for f (s) 6= 0.
[48] F. Schwabl, Quantum Mechanics, Springer, Berlin, 2007.
[49] E. Schrodinger, Phys. Rev. 28 (1926), 1049.
[50] One point to stress here is that a quantum system that interchanges
information with a classical one in a controllable way (such as a pulse
of a laser field on an atom) is also considered isolated.
[51] R. Gahler, A. G. Klein and A. Zeilinger, Phys. Rev. A 23 (1981), 1611.
[52] S. Weinberg, Phys. Rev. Lett. 62 (1989), 485.
[53] J. J. Bollinger, D. J. Heinzen, Wayne M. Itano, S. L. Gilbert and D. J.
Wineland, Phys. Rev. Lett. 63 (1989) , 1031.
[54] A. Galindo and P. Pascual, Quantum Mechanics (2 vols.), Springer,
Berlin, 1990.
[55] It can be shown it is the only linear (isomorphic) isometry of trace-class
operators. Actually the proof of theorem 4.5 is similar.
[56] S. Virmani, Entanglement Quantification & Local Discrimination, PhD
Thesis, Imperial College London, 2001.

[57] P. Stelmachovi
c and V. Buzek, Phys. Rev. A 64 (2001), 062106.
[58] Leaving aside the case of factorized global dynamics U(t1 , t0 ) =
UA (t1 , t0 ) UB (t1 , t0 ), see: D. Salgado and J. L. Sanchez-Gomez,
Comment on Dynamics of open quantum systems initially entangled with environment:
Beyond the Kraus representation
arXiv:quant-ph/0211164 (2002); H. Hayashi, G. Kimura and Y. Ota,
Phs. Rev. A 67 (2003), 062109.
[59] A. Peres, Quantum Theory: Concepts and Methods, Kluwer Academic
Publishers, Dordrecht, 1995.
[60] L. E. Ballentine, Quantum Mechanics: A Modern Development, World
Scientific, Singapore, 1998.
[61] T. F. Jordan, A. Shaji and E. C. G. Sudarshan, Phys. Rev. A 70 (2004),
052110.
[62] A. Shaji and E. C. G. Sudarshan, Phys. Lett. A 341 (2005), 48.
94

[63] A. Shaji, Dynamics of initially entangled open quantum systems, PhD

Thesis, University of Texas at Austin, 2005.

[64] H. A. Carteret, D. R. Terno and K. Zyczkowski,

Phys. Rev. A 77 (2008),
042113.
[65] K. Modi and E. C. G. Sudarshan, Phys. Rev. A 81 (2010), 052119.
[66] C. A. Rodrguez-Rosario, K. Modi and A. Aspuru-Guzik, Phys. Rev. A
81, 012313 (2010).
[67] D. Salgado, J. L. Sanchez-Gomez and M. Ferrero, Phys. Rev. A 70
(2004), 054102.
[68] D. M. Tong, L. C. Kwek, C. H. Oh, J. -L. Chen and L. Ma, Phys. Rev.
A 69 (2004), 054102.
[69] P. Pechukas, Phys. Rev. Lett. 73 (1994), 1060.
[70] A. J. van Wonderen and K. Lendi, J. Phys. A: Math. Gen. 33 (2000),
5557.
[71] C. A. Rodrguez, K. Modi, A. Kuah, A. Shaji and E. C. G. Sudarshan,
J. Phys. A: Math. Gen. 41 (2008), 205301.
[72] W. Stinespring, Proc. Am. Math. Soc. 6 (1955), 211.
[73] A. Kossakowski, Rep. Math. Phys. 3 (1972), 247.
[74] A. Kossakowski, Bull. Acad. Polon. Sci. Math. 20 (1972), 1021.
[75] M. B. Ruskai, Rev. Math. Phys. 6 (1994), 1147.
[76] Note that the composition two completely positive maps is also a completely positive map. However the composition of two arbitrary operators can be completely positive regardless of whether they are individually completely positive or not.
[77] C. A. Rodriguez-Rosario and E. C. G. Sudarshan, Non-Markovian Open
Quantum Systems, arXiv:0803.1183 (2008).
[78] J. R. Norris, Markov Chains, Cambridge University Press, Cambridge,
1997.
[79] S. N. Ethier and T. G. Kurtz, Markov Processes: Characterization and
Convergence, Wiley, New Jersey, 2005.
95

[80] Another proof of sufficiency can be founded in Appendix E of [154].

[81] V. Gorini, A. Kossakowski and E. C. G. Sudarshan, J. Math. Phys. 17
(1976), 821.
[82] G. Lindblad, Commun. Math. Phys. 48 (1976), 119.
[83] F. Benatti, A. M. Liguori and A. Nagy, J. Math. Phys. 49 (2008),
042103.
[84] R. A. Horn and C. R. Johnson, Topics in Matrix Analysis, Cambridge
University Press, Cambridge, 1991.
[85] S. G. Schirmer and X. Wang, Phys. Rev. A 81, 062306 (2010).
[86] D. Burgarth and V. Giovannetti, New J. Phys. 9 (2007), 150.
[87] E. B. Davies, Commun. Math. Phys. 19 (1970), 83.
[88] H. Spohn, Rep. Math. Phys. 10 (1976), 189.
[89] H. Spohn, Lett. Math. Phys. 2 (1977), 33.
[90] D. E. Evans, Commun. Math. Phys. 54 (1977), 293.
[91] A. Frigerio, Lett. Math. Phys. 2 (1977), 79.
[92] A. Frigerio, Commun. Math. Phys. 63 (1978), 269.
[93] A. Frigerio and H. Spohn, Stationary states of quantum dynamical semigroups and applications, in: L. Accardi, V. Gorini, G. Paravicini, (eds.),
Proceedings of Mathematical Problems in the Theory of Quantum Irreversible Processes, pp. 115-135, Laboratoria di Cibernetica del CNR,
1978.
[94] B. Baumgartner, H. Narnhofer and W. Thirring, J. Phys. A: Math.
Theor. 41 (2008), 065201.
[95] B. Baumgartner and H. Narnhofer, J. Phys. A: Math. Theor. 41 (2008),
395303.
[96] F. Ticozzi and L. Viola, IEEE T. Automat. Contr. 53 (2008), 2048.
[97] F. Ticozzi and L. Viola, Automatica 45 (2009), 2002.
[98] F. Ticozzi, S. G. Schirmer and X. Wang, IEEE T. Automat. Contr. 55
(2010), 2901.
96

[99] F. Ticozzi, R. Lucchese, P. Cappellaro, L. Viola, Hamiltonian Control of

Quantum Dynamical Semigroups: Stabilization and Convergence Speed,
arXiv:1101.2452 (2011).
[100] L. Jansen and H. Boon, Theory of Finite Groups. Application in
Physics, North-Holland, Amsterdam, 1967.
[101] S. Nakajima, Progr. Theor. Phys. 20 (1958), 984.
[102] R. Zwanzig, J. Chem. Phys. 33 (1960), 1338.
[103] C. Chicone, Ordinary Differential Equations with Applications,
Springer, New York, 2006.
[104] E. L. Ince, Ordinary Differential Equations, Dover, New York, 1956.
[105] R. S. Whitney, J. Phys. A: Math. Theor. 41 (2008), 175304.
[106] E. B. Davies, Commun. Math. Phys. 39 (1974), 91.
[107] E. B. Davies, Math. Ann. 219 (1976), 147.
[108] V. Gorini and A. Kossakowski, J. Math. Phys. 17 (1976), 1298.
[109] A. Frigerio and V. Gorini, J. Math. Phys. 17 (1976), 2123.
[110] A. Frigerio, C. Novellone and M. Verri, Rep. Math. Phys. 12 (1977),
279.

[111] M. Ziman, P. Stelmachovi

c and V. Buzek, Open Syst. Inf. Dyn. 12
(2005), 81.
[112] M. Ziman and V. Buzek, Phys. Rev. A 72 (2005), 022110.
[113] D. Taj and F. Rossi, Phys. Rev. A 78 (2008), 052113.
[114] L. Schwartz, Theorie des Distributions, Hermann, Paris, 1973.
[115] V. S. Vladimirov, Methods of the Theory of Generalized Functions,
Taylor & Francis, London, 2002.
[116] R. Doll, D. Zueco, M. Wubs, S. Kohler and P. Hanggi, Chem. Phys.
347, 243 (2008).
[117] G. B. Arfken and H. J. Weber, Mathematical Methods for Physicists,
Academic Press, San Diego, 1985.
97

[118] R. D
umcke and H. Spohn, Z. Phys. B34 (1979), 419.
[119] Y. Zhao and G. H. Chen, Phys. Rev. E 65 (2002), 056120.
[120] A. Suarez, R. Silbey and I. Oppenheim, J. Chem. Phys 97 (1992), 5101.
[121] P. Gaspard and M. Nagaoka, J. Chem. Phys. 111 (1999), 5668.
[122] F. Benatti, R. Floreanini and M. Piani, Phys. Rev. A 67 (2003),
042110.
[123] F. Benatti, R. Floreanini and S. Breteaux, Laser Physics 16 (2006),
1395.
[124] R. Kubo, J. Phys. Soc. Jpn. 12 (1957), 570; P. C. Martin and J.
Schwinger, Phys. Rev. 115 (1959), 1342.
[125] C. M. Caves and G. J. Milburn, Phys. Rev. A 36 (1987), 5543.
[126] E. B. Davies and H. Spohn, J. Stat. Phys. 19 (1978), 511.
[127] R. Alicki, J. Phys. A: Math. Gen. 12 (1979), L103.
[128] H. -P. Breuer and F. Petruccione, Phys. Rev. A 55 (1997), 3101.
[129] S. Kohler, T. Dittrich and P. Hanggi, Phys. Rev. E 55 (1999), 300.
[130] V. V. Dobrovitski and H. A. De Raedt, Phys. Rev. E 67 (2003), 056702.
[131] E. Paladino, M. Sassetti, G. Falci, U. Weiss, Phys. Rev. B 77 (2008),
041303(R).
[132] N. P. Oxtoby, A. Rivas, S. F. Huelga and R. Fazio, New J. Phys. 11
(2009), 063028.
[133] R. Bulla, H. -J. Lee, N. -H. Tong and M. Vojta, Phys. Rev. B 71 (2005),
045122; R. Bulla, T. Costi and T. Pruschke, Rev. Mod. Phys. 80 (2008),
395.
[134] J. Prior, A. W. Chin, S. F. Huelga and M. B. Plenio, Phys. Rev. Lett.
105 (2010), 050404.
[135] B. Vacchini and H. -P. Breuer, Phys. Rev. A 81 (2010), 042103.
[136] S. M. Barnett and S. Stenholm, Phys. Rev. A 64 (2001), 033808.

98

[137] S. Daffer, K. Wodkiewicz, J. D. Cresser and J. K. McIver, Phys. Rev.

A 70 (2004), 010304.
[138] H. -P. Breuer and B. Vacchini, Phys. Rev. Lett. 101 (2008), 140402.
[139] A. Shabani and D. A. Lidar, Phys. Rev. A 71 (2005), 020101(R).
[140] S. Maniscalco and F. Petruccione, Phys. Rev. A 73 (2006), 012111.
[141] S. Maniscalco, Phys. Rev. A 75 (2007), 062103.
[142] X. L. Huang and X. X. Yi, EPL 82 (2008), 50001.
[143] A. Kossakowski and R. Rebolledo, Open Syst. Inf. Dyn. 16 (2009), 259.
[144] L. Mazzola, E. -M. Laine, H. -P. Breuer, S. Maniscalco and J. Piilo,
Phys. Rev. A 81 (2010), 062120.
[145] B. L. Hu, J. P. Paz and Y. Zhang, Phys. Rev. D 45 (1992), 2843.
[146] R. Karrlein and H. Grabert, Phys. Rev. E 55 (1997), 153.
[147] B. M. Garraway, Phys. Rev. A 55 (1997), 2290.
[148] H. -P. Breuer, B. Kappler and F. Petruccione, Phys. Rev. A 59 (1999),
1633.
[149] F. Shibata, Y. Takahashi and N. Hashitsume, J. Stat. Phys. 17 (1977),
171.
[150] D. Chrusci
nski and A. Kossakowski, Phys. Rev. Lett. 104 (2010),
070406.
[151] M. J. W. Hall, J. Phys. A: Math. Theor. 41 (2008), 269801.
[152] D. Chrusci
nski, A. Kossakowski, P. Aniello, G. Marmo and F. Ventriglia, A class of commutative dynamics of open quantum systems,
arXiv:1004.1388 (2010).
[153] D. Chrusci
nski and A. Kossakowski, General form of quantum evolution, arXiv:1006.2764 (2010).
[154] G. Schaller and T. Brandes, Phys. Rev. A 78 (2008), 022106.
[155] G. Schaller and T. Brandes, Phys. Rev. A 79 (2009), 032110.
[156] F. Benatti, R. Floreanini, U. Marzolino, EPL 88 (2009), 20011.
99

[157] F. Benatti, R. Floreanini, U. Marzolino, Phys. Rev. A 81 (2010),

012105.
[158] R. P. Feynman and F. L. Vernon, Ann. Phys. (N. Y.) 24 (1963), 118.
[159] A. O. Caldeira and A. J. Leggett, Physica 121A (1983), 587.
[160] N. Gisin and I. C. Percival, J. Phys. A: Math. Gen. 25 (1992), 5677.
[161] L. Diosi, N. Gisin, J. Halliwell and I. C. Percival, Phys. Rev. Lett. 74
(1995), 203.
[162] M. B. Plenio and P. L. Knight, Rev. Mod. Phys. 70 (1998), 104.
[163] I. de Vega, D. Alonso and P. Gaspard, Phys. Rev. A 71 (2005), 023812.
[164] M. Lax, Phys. Rev. 129 (1963), 2342.
[165] D. Alonso and I. de Vega, Phys. Rev. Lett. 94 (2005), 200403.
[166] M. M. Wolf, J. Eisert, T. S. Cubitt, and J. I. Cirac, Phys. Rev. Lett.
101 (2008), 150402.
[167] H. -P. Breuer, E. -M. Laine and J. Piilo, Phys. Rev. Lett. 103 (2009),
210401.
[168] A. Rivas, S. F. Huelga and M. B. Plenio, Phys. Rev. Lett. 105 (2010),
050403.
[169] X. -M. Lu, X. Wang and C. P. Sun, Phys. Rev. A 82 (2010), 042103.
[170] A. R. Usha Devi, A. K. Rajagopal and Sudha, Phys. Rev. A 83 (2011),
022109.
[171] E. Andersson, J. D. Cresser, M. J. W. Hall, Canonical form of master equations and characterization of non-Markovianity, arXiv:1009.0845
(2010).
[172] S. C. Hou, X. X. Yi, S. X. Yu and C. H. Oh, An alternative nonMarkovianity measure by divisibility of dynamical map, arXiv:1102.4659
(2011).
[173] D. Chruscinski, A. Kossakowski and A. Rivas, Phys. Rev. A 83 (2011),
052128.

100