You are on page 1of 12

1

Introduction and the model

Begin by introducing the Hamiltonian. Since our interest is in creating the


most general expression possible, we may find ourselves suffering from some
index overload! The model consists of two subsystems, labelled by coordinates
q and x respectively, with an interaction between them. While the q subsystem
is completely arbitrary, we stipulate that the xm subsystem consists of a selfinteracting harmonic bath, characterised by a dynamical matrix . The ith
particle has coordinate xi = xi , momentum pi = pi and mass Mij = mi ij .
Finally the interaction between the systems is bilinear between bath coordinates
and some arbitrary function g(q) of the q coordinates. More concisely, we have
the Hamiltonian:
1
1
Htot = Hq + pTx M 1 px + xT M 1/2 x gT (q)x
|2
{z 2
} | {z }
HB

(1)

HI

Our aim is to express the dynamical evolution of the q sub-system in terms


of a c-number equation for the reduced density matrix. The first step towards
this is decoupling the bath, to present Htot in a more streamlined form. First,
consider the eigenvectors of , chosen such that they form an orthonormal basis
eTn en = nn0 , with eigenvalues en = n2 en . This then allows for a set of normal
coordinates n to be defined:
n = eTn x
(2)
gn (q) = eTn g(q)

(3)

With these substitutions, the Hamiltonian is decoupled and the mass of


particles is rescaled to unity:
X mn
Htot = Hq +
(2 n2 n2 ) + gn (q)n
(4)
2 n
n

The density matrix in the path integral formalism

The density matrix is a complete description of a quantum system. In the case


of two interacting systems (as above), it is given by:
(Q, X; Q0 , X 0 ) =


pi hQ, X|i i hi |Q0 X 0 i = (Q, X) (Q0 , X 0 ) av

all states

(5)
Where Q and X describe the collective coordinates of the two systems. More
concretely, the density matrix is used to generate expectation values of observables with:
hAi = Tr(A)
1

(6)

So that for an operator acting only on Q variables:


ZZ
hAi =
(Q, X; Q0 , X)A(Q, Q0 )dQ dQ0 dX

(7)

If one is interested only in the properties of one of these systems, then the
reduced density matrix at time T can be expressed as:
Z
(QT ; Q0T ) = (QT , XT ; Q0T XT )dXT
(8)
i.e. A trace over the X system.
The time evolution of a density matrix is governed by the quantum Liouville
equation:

= [H, ]
(9)
t
This differential equation can be equally well expressed in an integral form,
using the propagator G:
ih

(QT , XT ; Q0T , XT0 )

Z
=

dQdQ00 dXdX 0 G(QT , XT ; Q, X)


(Q, Q00 ; X, X 0 )G (Q0T XT0 ; Q00 X 0 )

(10)

q,t

Dq( ) exp (iS[q( )]/h)

G(q, t; q ) =

(11)

q 0 ,0

The substitution of (11) and (10) into (8) allows for a path integral representation of the reduced density matrix at some time T :

(QT , Q0T ) =

Z Z

(XT XT0 ) dXT dXT0 dX dX 0 DX(t) DX 0 (t)


exp (i/
h [SX (X) SX (X 0 ) + SI (Q, X) SI (Q0 , X 0 )])(Q; Q0 ; X, X 0 )
exp (i/
h [SQ (Q) SQ (Q0 )])DQ(t)DQ0 (t)dQ dQ00
(12)
The question of how to progress this formal expression into an evolution
equation for the model Hamiltonian in the previous section is what will occupy
the rest of this piece(?). A critical factor in this is in the treatment of the initial
density matrix.

2.1

To partition or not to partition

In treatments of interacting quantum systems, it is frequently (almost universally) almost universally assumed that before the interaction is turned on at
t = 0, the density matrix of the system can be expressed as:
tot
Q
X
0 =
0

(13)

While this simplifies the computation of an influence functional, in most


systems of interest the interaction between a particle and its environment is an
integral part of the system and not within our ability to control.
There is perhaps an analogy for this assumption- much like the Sykes-Picot
agreement, partitioning makes life easier, but may not accurately reflect the real
states and will almost certainly come to grief later.
Of course the corollary of this is that peace in the middle east will only be
achieved when we can fully understand and characterise quantum thermodynamics.
Moving on...
An alternative approach is to consider the full interacting system once it
has been allowed to equilibrate (requiring a time-independent potential). The
initial state would then be described by the canonical density matrix:
=

1 H
e
Z

(14)

Where = kB1T
Then there is a class of realisable initial density matrices of the form:
X
j O
j0
O
(15)
tot
0 =
j

Where the operators act only on the Fock space of the q system and leave
the bath coordinates unchanged. For example, any dynamical measurement of
the q system would have this form. Equally, one can imagine either turning on
(or off) an external potential at time 0 and examining the dynamics. A simple
example would be a single charged particle in a bath of electrically neutral
particles with a switch on (off) of an electric field.
An expression for the initial density matrix in the coordinate representation
( (Q; Q0 ; X, X 0 ) = hQ, X|
|Q0 , X 0 i) by inserting fat unities between operators:
0
0
tot
0 (Q; Q ; X, X ) =

XZ

Q
0 hQ|O
j |
j0 |Q
0i
dQd
q i (Q; Q0 ; X, X 0 )hQ0 |O

(16)

j
0 0 0
Q
0 ) = P hQ|O
j |QihQ

Defining a preparation function (Q, Q0 , Q,


|Oj |Q i
j
this becomes:
Z
tot
0
0
Q
0 (Q, Q0 , Q,
Q
0 ) (Q;
Q
0 ; X, X 0 )
0 (Q; Q ; X, X ) = dQd
(17)

2.2

The density matrix as a path integral

As noted by Feynman, it is possible to express the equilibrium density matrix


of a system in the path integral formalism, performing a Wick rotation t it
and then integrating all paths through imaginary time from q(0) = Q0 up to the
q(h) = Q, with similar constraints for x. More formally:

1
(Q, Q ; X, X ) =
Z
0

q(
h)=Q;
Z x(h)=X

1
DqDx exp ( S E [q, x])
h

(18)

q(0)=Q0 ; x(0)=X 0

Here Z is the partition function for equilibrium system, and S E is the


Euclidean action. Combining this with eqn. (17) one can express the initial
density matrix as a path integral:

0
0
tot
0 (Q; Q ; X, X )

X 1
=
Z
j
Z

q(
h)=Q;
Z x(h)=X

DqDx
q(0)=Q0 ; x(0)=X 0


1 E
0
0 0

dQdQ (Q, Q , Q, Q ) exp ( S [q, x])


h

(19)

The next step is to evolve this through time. In the language of propagators,
this can be expressed as (appending time labels to coordinates):

(QT , XT ; Q0T , XT0 )

Z
=

dQ0 dQ00 dX0 dX00 G(QT , XT ; Q0 , X0 )


(Q0 , Q00 ; X0 , X00 )G (Q0T XT0 ; Q00 X00 )

(20)

If the path integral representation for the propagators is also used, and the
density matrix is traced over closed paths in the X system, then we arrive at
an expression for the reduced density matrix:
Z 
1
Q
0
Q
0 (Q0 , Q0 , Q,
Q
0)
T (Q; Q ) =
dQ0 dQ00 dQd
0
Z



i
1 E
0
0
0

DQDQ DQ exp
(SQ [Q] SQ [Q ]) SQ [Q] F[Q, Q , Q]
(21)
h
h

=
F[Q, Q0 , Q]

1
ZB

X0 , X00 ]F [Q0 , XT , X00 ]


dXT dX0 dX00 F [Q, XT , X0 ]F E [Q,

(22)
Here F describes the influence functional for the system, decomposed into
three, smaller influence functionals. Lacking a better term for them, they will
be referred to as the forwards, backwards and Euclidean funzionina.
4

This requires some (a lot) of unpicking, given the number of coordinates


that must be... well, coordinated. Firstly it is important to explain how the
three Q system coordinates arise. The Q and Q0 coordinates describe trajectories through real time, integrated over initial positions in the phase space.
coordinates have arisen as a consequence of projecting the
In contrast the Q
density matrix into some arbitrary coordinate space. The result of this is that
in the path integral representation these coordinates are the terminals for all of
0 = Q(
= 0) and
the trajectories in imaginary time as in eqn. (18). That is, Q

Q = Q( = h
).
The influence functional can be considered as the evaluation of all path
integrals over the bath coordinates. This amounts to a separate path integration
using the interacting and bath-part of the action in both the propagator and
density matrix. As the reduced density matrix requires a trace over the X
coordinates at time T , XT = XT0 , F involves an integration over only three
fixed points, which together form a continuous contour through complex time.
Explicitly, if B and C are the start and end points of a path y( ) :
y(T
Z )=B


i
(SB [y(t)] + SI [y(t), A])
Dy(t) exp
h

F [A, B, C] =

(23)

y(0)=C

While the Euclidean influence functional is simply a Wick rotated version of


the same:

B,
C]
=
F [A,
E

y(
hZ)=B



1 E

[
y ( )] + SIE [
y ( ), A])
D
y ( ) exp (SB
h

(24)

y(0)=C

In addition to this, the normalising factor has been split such that Z =
ZB Z. Where ZB is the partition function for the bath Hamiltonian. This is
done so that in the limit of no interactions between the Q and X subsystems,
the influence functional is identically 1. The term this prefactor is attributed
to is a matter of taste, with some sources preferring to incorporate it into eqn.
(24) rather than eqn. (22) in order for the Euclidean influence functional to
maintain the same structure as eqn. (18).

Calculating the Influence Functional

Having established the influence functional formalism, it is worth returning to


the decoupled Hamiltonian from the introduction, deriving an analytic expression for the systems influence functional.
As the influence functional is decomposed into funzionina in eqn. (23), the
first step is to explicitly calculate this for both the real and Euclidean actions.

Taking the forward-path, with a trajectory described by the vector of bath


coordinates x(t), with terminals x0 and xT , the funzionina is given by:
x(TZ)=xT

F [q, xT , x0 ] =

"

#
Z T
i
1 T
T
1/2
T
Dx(t) exp
(
dt (x x M p x M x) g (q)x)
h 0
2

x(0)=x0

Y
n

n (TZ)=n,T

"

i
Dn (t) exp
(
h

#
Y
mn 2
dt
(n n2 n2 ) + gn (q)n ) =
Fn [q, n,T , n,0 ]
2
n

Z
0

n (0)=n,0

(25)
Where the second and third equalities follow from the decoupling of the
Hamiltonian. The game then is to evaluate Fn . There are a great many standard
techniques for this, exploiting the Gaussian nature of the integral. Skipping to
the end, we obtain:
r
mn n
1
(26)
exp [ n [q, n,T , n,0 ]]
Fn [q, n,T , n,0 ] =
2ih sin(n T )
h
|
{z
}
Pn

The prefactor Pn will be an important check later, and with the backwards
and Euclidean prefactors denoted by Pn and PnE respectively. The phase factor
contains the classical action, given by:

n [q, n,T , n,0 ] = i


|

mn n cos(n T ) 2
mn n
2
(n,T + n,0
)+i
n,0 n,T
2 sin(n T )
2 sin(n T )
{z
}
|
{z
}
A

1
n,0 i
sin(n T )
|

1
+i
mn n sin(n T )
|

Z
dt

dt sin(n (T t))gn (q(t))


{z
}

1
n,T i
sin(n T )
|
Z

dt sin(n (t))gn (q(t))


{z
}

dt0 sin(n (T t) sin(n t0 )gn (q(t))gn (q(t0 )) (27)


{z
}
E

For completeness, there should be five constants for each of the n oscillators,
but given the decoupled nature of the Hamiltonian, it is assumed that the precise
argument of these constants can be determined from context (. no mixed terms
of form n,0 An+j , j 6= 0).

For the Euclidean funzionina, the phase will be given by:


mn n
h , n,0 ] = mn n cos(n h) (2
2
E
n,0 n,h
n [q, n,
n,
h + n,0 )
2 sinh(n h)
2 sinh(n h)
|
{z
}
|
{z
}

n,0

1
sinh(n h)
|

n,h

1
mn n sinh(n
h)
|

d sinh(n ( h))gn (
q ( ))
{z
}

1
sinh(n h)
|

d sinh(n )gn (
q ( ))
{z
}

d
0

d 0 sinh(n (h ) sinh(n 0 )gn (q( ))gn (q( 0 ))


{z
}

(28)
and the prefactor:
PnE

r
=

m
2h sinh(h)

(29)

Substituting this back into eq. (22), combined with the continuity of the
0
contour n,h = n,0 , n,0 = n,0
gives the overall influence functional as:
Z
1 Y
2 E
0
F[q, q , q] =
|Pn | Pn
dn,0
dn,T dn,0
ZB n


1
i
0
0
(n [q, n,T , n,0 ] n [q, n,0
, n,T ]) E
[q,

]
exp
n,0 n,0
h

h n
0

(30)

Despite the somewhat intimidating appearance of these phase factors, more


sense can be made of this when we consider that the phases are a second order
polynomial in the three integration variables... Linear algebra to the rescue!

3.1

Calculating the Prefactor

Rather than attempting to integrate eqn. (30) one variable at the time, it is
worth noting that the expression can be compacted using vector-matrix notation
with the following definitions:

n,0
0
yn = n,0
(31)
n,T

A + A
B
B

= B
A A B
B
B
0

C C0

b = (C + D)
D D0

(32)

(33)

c = E E0 E

(34)

Having done this, the influence functional is now explicitly a multivariate


Gaussian. Completing the square yields:
Z
1 Y
2 E

|Pn | Pn
dyn
F[Q, Q , Q] =
ZB n
{z
}
|
0

Ptot

1
exp
h

"

b
yn +
2

T

b
yn +
2

#)
T

b b + c

(35)




1
1
F[q, q 0 , q] = (h)3N/2 Ptot ||1/2 exp
bT b + c
h
4

(36)

Integrating this is trivial, producing:

It is now worth returning to the issue of prefactors. If this expression for


the influence functional is correct, then as previously noted, it should become 1
in the absence of interactions. This demands that the overall prefactor is unity.
In order to evaluate this, the value of ZB is needed. Fortunately, the partition
function for the path Hamiltonian is (relatively) straightforward to calculate,
and gives:
ZB =

2 sinh( 12 n h)
r
Y
mn n
mn n
1
=
2 sinh( n h)
2
h
sin(
T
)
2
h
sinh(
h

)
2
n
n
n

(37)

Ptot

(38)

Now all that remains is to calculate the determinant (appending n labels to


coefficients):
Y
n An )
||1 =
2Bn2 (B
(39)
n

A useful identity at this juncture is given by:


(
n cosh2 ( n h ) +
2B
2

An Bn =
n sinh2 ( n h ) 2B
2
8

(40)

Substituting the appropriate values, together with judicious use of the doubleangle trigonometric formulae, it is possible to verify:
(h)3/2 ||1/2 = Ptot 1

(41)

Meaning the overall the influence functional is a pure (complex) phase. A


good sign! The final step is now to rearrange this in a way

3.2

Rearranging the phase

Dropping the n subscript, the phase contribution from a single bath oscillator
is given by:
1
(42)
= (bT b + c)
h
We can begin to evaluate this by inverting the matrix :

B 2
B 2
B(A A B)
A A)
= ||
(43)
B 2
B 2
B(B
2
B(B
A A) A (A2 + B 2 )
B(A A B)
While some effort has been made to massage the phase into the most userfriendly form possible, this is as far as such an approach can take us. The g
interaction can take three possible arguments from the Q system (q, q 0 , q), there
are six possible terms involving combinations of these coordinates. Calculating
the overall phase is simply a case of considering these term by term in the matrix
multiplication, reinserting the explicit form for the matrix entries.
1
= ([g(q(t))g(q(t0 ))] + [g(q 0 (t))g(q 0 (t0 ))]+
h

[g(q(t))g(q 0 (t0 ))]+[g(


q ( ))g(q(t))]+[g(
q ( ))g(q 0 (t))]+[g(
q ( ))g(
q ( 0 ))])
(44)
As an example, consider the g(
q ( ))g(
q ( 0 )) term. This will be appear in
2 2

any term involving C , D , C D and E. Adding these terms up generates:


[g(
q ( ))g(
q ( 0 ))] =

2
(C + D)

E
sinh2 ( h )
16B
2

Now define the factor:


Kp =

1
2m sinh( h2 )

This is somewhat post-hoc but this factor appears as a common multiplication of all six terms. Also useful is the following transformation for time ordering
integrals:

Z Z1

Z Z

d1 d2 (f (1 , 2 ) + f (2 , 1 ))

d1 d2 f (1 , 2 ) =
0

(45)

Using this and substituting in the integral expressions:

Zh

[g(
q ( ))g(
q ( ))] = Kp

Z
d

d 0 g(
q ( ))g(
q ( 0 ))

T1
sinh((h )) sinh( 0 )
+
2
h

h
4 cosh( 2 ) sinh ( 2 )
cosh( h2 )

)
(46)

T1 = [sinh((
h )) sinh((
h 0 ))
+ sinh( ) sinh( 0 )
+ sinh( ) sinh((
h 0 ))
+ sinh((
h )) sinh( 0 )]

(47)

This is a nightmare!
BUT, note that, for T1 that it is possible to simplify the expression with a
0
transformation of the variable + h
2 and doing the same for . This is
valid as long as the inverse transformation is applied after this simplification.
Call this transformed term T10 :
T10




h

h
h
h
0
0
= sinh(
+ ) + sinh(
)
sinh(
+ ) + sinh(
)
2
2
2
2
(48)
h
) cosh( ) cosh( 0 )
(49)
2
Now perform the inverse transformation on and 0 , to re-express T1 as:
= 4 sinh2 (

T1 =
h

h
) cosh( ) cosh( 0 ) + sinh2 (
) sinh( ) sinh( 0 )
2
2
h

h
cosh(
) sinh(
) (cosh( ) sinh( 0 ) + cosh( 0 ) sinh( ))
2
2
cosh2 (

(50)

of

Finally, reinsert this into the integrand and expand the second term in terms
The hyperbolic trigonometric terms in the integrand will become:

h

2 .

10

1
h
) cosh( ) cosh( 0 ) +
sinh2 (
) sinh( ) sinh( 0 )
h

2
2
cosh( 2 )
h

h
sinh(
) cosh( ) sinh( 0 ) sinh(
) cosh( 0 ) sinh( )
2
2
h

+ 2 sinh(
) cosh( ) sinh( 0 ) cosh(a) sinh( ) sinh( 0 )
2
1
h
sinh2 (

) sinh( ) sinh( 0 )
2
cosh( h
)
2

cosh(

(51)
Comparing with the full expansion for cosh(a + 0 ):

cosh(a) cosh( ) cosh( 0 ) cosh(a) sinh( ) sinh( 0 )


sinh(a) cosh( 0 ) sinh( ) + sinh(a) cosh( ) sinh( 0 )
(52)
This establishes the hyperbolic terms in the integrand can be expressed as:
Kp cosh(

h
+ 0)
2

(53)

Other terms can similarly be calculated in this extraordinarily tedious manner. A more efficient approach however is to employ a numerical symbolic manipulation program. Some care must be taken here- implied integration and time
ordering by hand are necessary to obtain the correct expression, but eventually
it is possible to obtain the final expression for the overall influence functional:
1
F[q, q 0 , q] = exp( [q, q 0 , q])
h

[q, q , q] =

X

Z
dt

Z
d

Z
d

(54)

d 0 Kn (i 0 i )gn (
q ( ))gn (
q ( 0 ))

dt Kn (t i )gn (
q ( ))(gn (q(t)) gn (q 0 (t)))


dt0 (gn (q(t))gn (q 0 (t)))[Kn (tt0 )gn (q(t0 ))K (tt0 )gn (q 0 (t0 ))]

(55)

11

defining:
Kn () =

cosh[n ( h2 i)]
2mn n sinh( h2n )

12

(56)

You might also like