1307.3020v2derivation of A True (T ! 0+) Quantum Transition-State Theory. II. Recovery of The Exact Quantum Rate in The Absence of Recrossing
1307.3020v2derivation of A True (T ! 0+) Quantum Transition-State Theory. II. Recovery of The Exact Quantum Rate in The Absence of Recrossing
1307.3020v2derivation of A True (T ! 0+) Quantum Transition-State Theory. II. Recovery of The Exact Quantum Rate in The Absence of Recrossing
+
) quantum transition-state theory.
II. Recovery of the exact quantum rate in the absence of recrossing
Stuart C. Althorpe
i=1
q
i1
i1
/2|e
N
H
|q
i
+
i
/2
_
q
i
+
i
/2|e
i
Ht/
|z
i
_
z
i
|e
i
Ht/
|q
i
i
/2
_
(1)
where N is the number of polymer beads,
N
= /N,
with = 1/k
B
T, and q {q
1
, . . . , q
N
}, with z and
similarly dened. The function f(q) is the ring-polymer
dividing surface, which is invariant under cyclic permuta-
tions of the polymer beads (i.e. of the individual q
i
), and
thus becomes invariant to imaginary-time translation in
the limit N . The operator
F[f(q)] gives the ux
perpendicular to f(q), and is given by
F[f(q)] =
1
2m
N
i=1
_
p
i
f(q)
q
i
[f(q)] +[f(q)]
f(q)
q
i
p
i
_
(2)
Note that we employ here a convention introduced in
Part I, that the rst term inside the curly brackets is in-
serted between e
N
H
q
i
+
i
/2
_
and
q
i
+
i
/2
e
i
Ht/
in Eq. (1), and the second term between e
i
Ht/
q
i
+
i
/2
_
and
q
i
+
i
/2
N
H
. This is done to empha-
sise the form of C
[N]
fs
(t); [Eq. (1) is written out in full in
Part I].
We can regard C
[N]
fs
(t) as a generalized Kubo-
transformed time-correlation function, since it correlates
an operator (in this case
F[f(q)]) on the (imaginary-
time) Feynman paths at t = 0 with another operator (in
this case h[f(z)]) at some later time t, and would re-
duce to a standard Kubo-transformed function if these
operators were replaced by linear functions of position or
momentum operators. The advantage of C
[N]
fs
(t) is that
it allows both the ux and the side dividing surface to
be made the same function of ring-polymer space (i.e.
f), which is what makes C
[N]
fs
(t) non-zero in the limit
t . One can show
1
that the invariance of f(q) to
imaginary time-translation in the limit N ensures
that C
[N]
fs
(t) is positive-denite in the limits t 0
+
and
N . This allows us to dene the quantum TST rate
k
Q
()Q
r
() = lim
t0+
lim
N
C
[N]
fs
(t) (3)
where
k
Q
()Q
r
() = lim
N
1
(2)
N
_
dq
_
dP
0
[f(q)]
_
B
N
(q)
P
0
m
h(P
0
)
_
2
N
2
m
e
P
2
0
N
/2m
N
i=1
q
i1
|e
N
H
|q
i
_
(4)
3
Comparison with refs. 68 shows that k
Q
() is identical
to the RPMD-TST rate. The terms quantum TST and
RPMD-TST are therefore equivalent (and will be used
interchangeably throughout the article).
For quantum TST to be applicable, one must be able
to assume that real-time coherences have only a small
eect on the rate. It then follows that (a good approxi-
mation to) the optimal dividing surface f(q) is the one
that maximises the free energy of the ring-polymer en-
semble. If the reaction barrier is reasonably symmetric,
48
or if it is asymmetric but the temperature is too hot for
deep tunnelling, then a good choice of dividing surface is
f(q) = q
0
q
(5)
where
q
0
=
1
N
N
i=1
q
i
(6)
is the centroid. (This special case of RPMD-TST was in-
troduced earlier
2024
and referred to as quantum TST;
to avoid confusion we refer to it here as centroid-TST
26
)
If the barrier is asymmetric, and the temperature is be-
low the cross-over to deep tunnelling, then a more com-
plicated dividing surface should be used which allows the
polymer to stretch.
6
As mentioned above, f(q) must be
invariant under cyclic permutation of the beads so that
it becomes invariant to imaginary time-translation in the
limit N , and thus gives positive-denite quantum
statistics.
It is assumed above, and was stated without proof in
Part I, that the RPMD-TST rate gives the exact quan-
tum rate in the absence of recrossing, and is thus a good
approximation to the exact rate if the amount of recross-
ing is small. The remainder of this article is devoted to
deriving this result.
III. LONG-TIME LIMITS
A. Hybrid ux-side time-correlation function
To analyze the t limit of C
[N]
fs
(t), we will nd
it convenient to consider the t limit of the closely
related hybrid ux-side time-correlation function:
C
[N]
fs
(t) =
_
dq
_
dz
_
d
F[f(q)]h(z
1
q
i=1
q
i1
i1
/2|e
N
H
|q
i
+
i
/2
_
q
i
+
i
/2|e
i
Ht/
|z
i
_
z
i
|e
i
Ht/
|q
i
i
/2
_
(7)
Note that we could equivalently have inserted any one
of the other z
i
into the side-function, and also that we
could simplify this expression by collapsing the identities
_
dz
i
e
i
Ht/
z
i
_
z
i
e
i
Ht/
, i = 1 [but we have not done
so in order to emphasise the relation with C
[N]
fs
(t)].
The function C
[N]
fs
(t) does not give a quantum TST,
except in the special case that N = 1 and f(q) = q
1
. In
this case, C
[N]
fs
(t) is identical to C
[1]
fs
(t), whose t 0
+
limit was shown in Part I to be identical to the quan-
tum TST introduced on heuristic grounds by Wigner in
1932.
49
For N > 1, the ux and side dividing surfaces
in C
[N]
fs
(t) are dierent functions of ring-polymer space,
with the result that C
[N]
fs
(t) tends smoothly to zero in the
limit t 0
+
.
1
By taking the t limit of the equivalent side-ux
time-correlation function C
[N]
sf
(t), we show in Appendix
A that
k
Q
()Q
r
() = lim
t
C
[N]
fs
(t) (8)
where k
Q
() is the exact quantum rate, and this expres-
sion holds for all N 1. For N = 1, we have thus proved
that the ux-side time-correlation function that gives the
Wigner form of quantum TST (see above) also gives the
exact rate in the limit t .
50
For N > 1, which
is our main concern here, C
[N]
fs
(t) has the same limits
as the Miller-Schwarz-Tromp
3
ux-side time-correlation
function, tending smoothly to zero as t 0
+
, and giving
the exact quantum rate as t . We can also evaluate
the t limit of C
[N]
fs
(t) directly [i.e. not via C
[N]
sf
(t)].
We apply rst the relation
lim
t
_
dz
x|e
i
Kt/
|z
_
h(z q
z|e
i
Kt/
|y
_
=
_
dp
x|p
_
h(p)
p|y
_
(9)
where
K is the kinetic energy operator and
x|p
_
=
(2)
1/2
exp (ipx); this converts Eq. (7) into a form that
involves applications of the Mller operator
51
lim
t
e
i
Ht/
e
i
Kt/
(10)
onto momentum states
p
i
_
. We then use the relation
p
_
=
p
_
(11)
where
p
_
is the (reactive) scattering wave function
with outgoing boundary conditions,
52
to obtain
lim
t
C
[N]
fs
(t) =
_
dpA
N
(p)h(p
1
) (12)
with
A
N
(p) =
_
dq
_
d
F[f(q)]
i=1
q
i1
i1
/2|e
N
H
|q
i
+
i
/2
_
q
i
+
i
/2|
pi
_
pi
|q
i
i
/2
_
(13)
4
FIG. 1: Representation of the momentum integrals in
Eqs. (12) and (17) for N = 3. The axes (a) are positioned such
that the origin is at the centre of each of the cubes, which are
cut by (b) the centroid dividing surface h(p
0
) (blue), and (c)
the dividing surface h(p1) (blue). The red arrow represents
the centroid axis. This picture can be generalized to N > 3,
by replacing the cubes with N-dimensional hypercubes.
B. Representation of the ring-polymer momentum
integral
To analyze the properties of Eq. (12) (and of Eq. (17)
given below), we will nd it helpful to represent the
space occupied by the intregrand as an N-dimensional
hypercube,
53
whose edges are the axes p
max
< p
i
<
p
max
, i = 1 . . . N, in the limit p
max
. We assume
no familiarity with the geometry of hypercubes, and in
fact use this terminology mainly to indicate that once a
property of A
N
(p) has been derived for N = 3 (where
the hypercube is simply a cube and thus easily visualised
as in Fig. 1) it generalizes straightforwardly to higher N.
The only formal properties of hypercubes that we need
are, rst that a hypercube has 2
N
vertices, second that
one can represent the hypercube by constructing a graph
showing the connections between its vertices, and third
that the graph for a hypercube of dimension N can be
made by connecting equivalent vertices on the graphs
of two hypercubes of dimension N 1. Figure 2 illus-
trates this last point, showing how the graph for a cube
(N = 3) can be made by connecting equivalent vertices
on the graphs for two squares (N = 2). Figure 2 also
introduces the (self-evident) notation that we will use to
label vertices; e.g. (1, 1, 1) refers to the vertex on an
N = 3 hypercube (i.e. a cube) located at p
1
= p
max
,
p
2
= p
3
= p
max
.
These properties allow one to build up a hypercube by
adding together its subcubes in a recursive sequence. By
subcube we mean that each p
i
is conned to either the
positive or negative axis; there are therefore 2
N
subcubes,
each corresponding to a dierent vertex of the hypercube
(so we can label the subcubes using the vertex notation
introduced above). Figure 3 shows how one can build up
an N = 3 hypercube (i.e. a cube) by adding its subcubes
together recursively, joining rst two individual subcubes
along a line, then joining two lines of subcubes in the form
of a square, and nally joining two squares of subcubes
to give the entire cube. The analogous sequence can be
used to build up a hypercube of any dimension N from
its subcubes, and will be useful in Sec. IV.B.
We now dene the energies
E
i
E
(p
i
) =
p
2
i
2m
+V
prod
p
i
> 0
=
p
2
i
2m
+V
reac
p
i
< 0 (14)
and introduce the notation p
i
, such that
p
i
=
_
p
2
i
+ 2m(V
prod
V
reac
) p
i
> 0
p
i
= +
_
p
2
i
+ 2m(V
reac
V
prod
) p
i
< 0 (15)
where V
reac
and V
prod
are the asymptotes of the potential
V (x) in the reactant (x ) and product (x )
regions; i.e. the tilde has the eect of converting a prod-
uct momentum to the reactant momentum corresponding
to the same energy E
i
, and vice versa. Note that we will
not need to interconvert between the reactant and prod-
uct momenta if one or other of them is imaginary, and
hence the square roots in Eq. (15) are always real.
For a symmetric barrier, it is clear that p
i
= p
i
, and
from this it is easy to show that
A
N
( p) = A
N
(p) for symmetric barriers (16)
where p ( p
1
, p
2
, . . . , p
N
); i.e. A
N
(p) is antisymmetric
with respect to inversion through the origin. Clearly this
antisymmetry ensures that the integration of A
N
( p) over
the entire hypercube (i.e. with the side function omitted)
gives zero. This integral is also zero for an asymmetric
barrier, but there is then no simple cancellation of A
N
(p)
with A
N
( p).
Finally, we note that A
N
(p) is symmetric with respect
to cyclic permutations of the p
i
, and thus has an N-fold
axis of rotational symmetry around the diagonal of the
hypercube on which all p
i
are equal. We will refer to this
diagonal as the centroid axis, since displacement along
this axis measures the displacement of the momentum
centroid p
0
=
N
i=1
p
i
/N.
FIG. 2: Diagram showing how a cube can be built up by
connecting the equivalent vertices on two squares. One can
similarly build up an N-dimensional hypercube by connecting
the equivalent vertices on two (N1)-dimensional hypercubes.
This gure also illustrates the notation used in the text to
label the vertices of a hypercube.
5
FIG. 3: Diagram showing how a cube can be built up recur-
sively in three steps from its eight subcubes. One can similarly
build up an N-dimensional hypercube in N steps from its 2
N
subcubes.
C. Ring-polymer ux-side time-correlation
function
It is straightforward to modify the above derivation
to obtain the t limit of the ring-polymer ux-side
time-correlation function C
[N]
fs
(t). The only change nec-
essary is to replace the side function h(z
1
) by h[f(z)],
which gives
lim
t
C
[N]
fs
(t) =
_
dpA
N
(p)h[f(p)] (17)
where A
N
(p) is dened in Eq. (13), and f(p) is dened
by
lim
t
h[f(pt/m)] = h[f(p)] (18)
i.e. f(p) is the limit of f(p) at very large distances. In the
special case that f(q) = q
0
, we obtain f(p) = p
0
= f(p);
but in general f(p) = f(p). A time-independent limit
of Eq. (18) is guaranteed to exist, since otherwise f(q)
would not satisfy the requirements of a dividing surface.
Whatever the choice of f(q), it is clear that the (per-
mutationally invariant) h[f(p)] encloses a dierent part
of the hypercube than does h(p
1
). For example, if
f(q) = q
0
and N = 3, then h[f(p)] = h(p
0
) cuts out
the half of the cube on the positive side of the hexagonal
cross-section shown in Fig. 1b, whereas h(p
1
) cuts o the
top half of the cube on the p
1
axis (Fig. 1c). Thus we
cannot in general expect the t limits of C
[N]
fs
(t) and
C
[N]
fs
(t) to be the same, unless A
N
(p) satises some spe-
cial properties in addition to those just mentioned. We
will show in the next two Sections that A
N
(p) does sat-
isfy such properties if there is no recrossing of any surface
orthogonal to f(q) in ring-polymer space.
IV. RING-POLYMER MOMENTUM
INTEGRALS
A. Structure of AN(p)
One can show using scattering theory (see Appendix
B) that A
N
(p) consists of the terms
A
N
(p) = a
N
(p)
_
N1
i=1
(E
i+1
E
i
)
_
+r
N
(p) (19)
where a
N
(p) is some function of p, and r
N
(p) satises
r
N
(p
1
, . . . , p
j
, . . . , p
N
) =
p
j
p
j
r
N
(p
1
, . . . , p
j
, . . . , p
N
)
(20)
(where the dots indicate that all the p
i
except p
j
take
the same values on both sides of the equation). Equa-
tion (20) is equivalent to stating that r
N
(p) alternates in
sign between adjacent subcubes (i.e. subcubes that dier
in respect of just one axis), or that r
N
(p) takes oppo-
site signs in even and odd subcubes (where a subcube is
dened to be even/odd if it has an even/odd number of
axes for which p
i
< 0). Note that r
N
(p) = 0 if any p
i
,
i = 1 . . . N, is imaginary (see Appendix B).
The rst term in Eq. (19) describes a set of 2
N
-
function spikes running along all the lines in the hyper-
cube for which the energies E
i
, i = 1 . . . N, are equal.
There is one such line in every subcube. Two of these
lines point in positive and negative directions along the
centroid axis (i.e. the diagonal of the hypercube). The
other 2
N
2 o-diagonal spikes radiate out from this axis.
If the barrier is symmetric, then each o-diagonal spike
is a straight line joining the centre of the hypercube to
one of its vertices. If the barrier is asymmetric, the o-
diagonal spikes are hyperbolae [on account of Eq. (15)].
The o-diagonal spikes are distributed with N-fold rota-
tional symmetry about the centroid axis because of the
invariance of A
N
(p) under cyclic permutations; e.g. for
N = 3, the spikes (1, 1, 1), (1, 1, 1), (1, 1, 1) (where
this notation identies each spike by the subcube that it
runs through) rotate into one another under cyclic per-
mutation of the beads; see Fig. 4.
B. Cancellation of the term rN(p)
We now show that r
N
(p) in Eq. (19) contributes zero
to C
[N]
fs
(t) and C
[N]
fs
(t) in the limits t, N , and may
6
FIG. 4: Plot of the o-diagonal spikes in AN(p) for N = 3,
obtained by looking down the centroid axis (the red arrow in
Fig. 1b).
therefore be ignored when discussing whether C
[N]
fs
(t)
gives the exact quantum rate in these limits. This prop-
erty is easy to show for a symmetric barrier, for which
Eqs. (16) and (20) imply that r
N
(p) is zero for all even
N, and thus that the contribution to the integral from
r
N
(p) tends to zero in the limit N . For an asym-
metric barrier, r
N
(p) is in general non-zero. However, we
now show that the alternation in sign between adjacent
subcubes [Eq. (20)] causes r
N
(p) to cancel out in both
C
[N]
fs
(t) and C
[N]
fs
(t) in the limits t, N .
This cancellation is easy to demonstrate for C
[N]
fs
(t):
One simply notes that the side-function h(p
1
) encloses an
even number of subcubes, which can be added together
in adjacent pairs. For example, if we add together the
adjacent subcubes (1, . . . , 1, 1) and (1, . . . , 1, 1) (where
the dots indicate that the intevening values of 1 and 1
are the same for the two subcubes), we obtain
_
0
dp
1
. . .
_
0
dp
N1
_
0
dp
N
r
N
(p)h(p
1
)
+
_
0
dp
1
. . .
_
0
dp
N1
_
0
dp
N
r
N
(p)h(p
1
) (21)
(where the dots indicate that the integration ranges for
p
i
, i = 2 . . . N 2 are the same in both terms). We
can change the limits on the last integrand to 0
by transforming the integration variable from p
N
to p
N
,
and using the relation p
i
dp
i
= p
i
d p
i
[see Eq. (15)]. Equa-
tion (20) then ensures that the two terms in Eq. (21)
cancel out. Hence the contribution from r
N
(p) cancels
out in the t limit of C
[N]
fs
(t) (for any N > 0).
Using similar reasoning, we can show that the contri-
bution from r
N
(p) to C
[N]
fs
(t) cancels out in the limits
t, N . For nite N, this cancellation is in general
54
only partial, because the function h[f(p)] encloses dif-
ferent volumes in any two adjacent subcubes. However,
one can show that the total mismatch in the volumes en-
closed in the even subcubes and the odd subcubes tends
rapidly to zero as N . The trick is to build up the
hypercube recursively, by extending to higher N the se-
quence shown in Fig. 3 for N = 3. The jth step in this
sequence can be written
S(N) =
_
dp
1
. . .
_
dp
j
_
0
dp
j+1
. . .
_
0
dp
N
r
N
(p)h[f(p)]
=
_
dp
1
. . .
_
dp
j1
_
0
dp
j
. . .
_
0
dp
N
r
N
(p)h[f(p)]
+
_
dp
1
. . .
_
dp
j1
_
0
dp
j
_
0
dp
j+1
. . .
_
0
dp
N
r
N
(p)h[f(p)] (22)
(where the rst set of dots in each term indicates that
the intervening integration ranges are < p
i
< ,
and the second set that they are 0 < p
i
< ). Because
each subcube in the second term is adjacent to its coun-
terpart in the third term, there is an almost complete
cancellation in the r
N
(p) terms. All that is left is the
residue,
S(N) =
_
dp
1
. . .
_
dp
j1
_
0
dp
j
. . .
_
0
dp
N
r
N
(p)
_
h[f(p
1
, . . . , p
j
, . . . , p
N
)]
h[f(p
1
, . . . , p
j
, . . . , p
N
)]
_
(23)
which occupies the volume sandwiched between the two
heaviside functions. Appendix C shows that this volume
is a thin strip on the order of N times smaller than the
volume occupied by r
N
(p) in each of the two terms that
were added together in Eq. (22). Now, each of these
terms was itself the result of a similar addition in the
j 1 th step, which also reduced the volume occupied
by r
N
(p) by a factor on the order of N, and so on. As
a result, the volume occupied by r
N
(p) after the Nth
(i.e. nal) step is on the order of N
N
times smaller than
the volume of a single subcube. The mismatch in volume
between the even and odd subcubes thus tends rapidly
to zero in the limit N , with the result that r
N
(p)
cancels out completely
55
in C
[N]
fs
(t) in the limits t, N
.
C. Comparison of -function spikes
We have just shown that only the rst term in Eq. (19)
contributes to C
[N]
fs
(t) and C
[N]
fs
(t) in the limits t, N
. Any dierence between these quantities can thus be
accounted for by comparing which spikes are enclosed
by the side functions h(p
1
) and h[f(p)]. It is clear that
both h(p
1
) and h(p
0
) enclose the spike that runs along the
centroid axis in a positive direction, and exclude the spike
7
that runs in a negative direction. A little thought shows
that this property must hold for any choice of h[f(p)]
(since the positive spike corresponds to all momenta p
i
travelling in the product direction as t , and vice
versa for the negative spike).
Any dierence between the t, N limits of C
[N]
fs
(t)
and C
[N]
fs
(t) can therefore be explained in terms of which
o-diagonal spikes are enclosed by h(p
1
) and h[f(p)].
These functions will enclose dierent sets of spikes. For
example, for a symmetric barrier, with N = 3, the
function h(p
0
) encloses the o-diagonal spikes (1, 1, 1),
(1, 1, 1) and (1, 1, 1), whereas h(p
1
) encloses (1, 1, 1),
(1, 1, 1) and (1, 1, 1).
We have therefore obtained the result that the t, N
limit of C
[N]
fs
(t) is identical to that of C
[N]
fs
(t) (and thus
gives the exact quantum rate) if the contribution from
each o-diagonal spike to A
N
(p) is individually zero. We
make use of this important result in the next Section.
V. EFFECTS OF RECROSSING
The results just obtained show that quantum TST will
give the exact quantum rate if two conditions are satis-
ed. First, there must be no recrossing of the cyclically
invariant dividing surface f(q) (by which we mean sim-
ply that C
[N]
fs
(t) is time-independent). Second, each of
the o-diagonal spikes [in the rst term of Eq. (19)] must
contribute zero to C
[N]
fs
(t) in the long-time limit. We
now show that this last condition is satised if there is
no recrossing of any dividing surface orthogonal to f(q)
in ring-polymer space.
A. Orthogonal dividing surfaces
A dividing surface g(q) orthogonal to f(q) satises
N
i=1
g(q)
q
i
f(q)
q
i
= 0 (24)
When f(q) = q
0
, the surface g(q) can be any function
of any linear combination of polymer beads orthogonal
to q
0
. For a more general (cyclically permutable) f(q),
g(q) will also take this form close to the centroid axis
(where, by denition, all degrees of freedom orthogonal
to the centroid vanish), and will assume a more general
curvilinear form away from this axis.
By no recrossing of g(q), we mean that the time-
correlation function
M
[N]
fs
(t) =
_
dq
_
dz
_
d
F[f(q)]h[g(z)]
i=1
q
i1
i1
/2|e
N
H
|q
i
+
i
/2
_
q
i
+
i
/2|e
i
Ht/
|z
i
_
z
i
|e
i
Ht/
|q
i
i
/2
_
(25)
is time-independent. We know from Part I that the
t 0
+
limit of M
[N]
fs
(t) is zero, since the ux and side di-
viding surfaces are dierent. Hence no recrossing of g(q)
implies that M
[N]
fs
(t) is zero for all time t, indicating that
there is no net passage of ux from the initial distribu-
tion on f(q) through the surface g(q). Taking the t
limit (using the same approach as in Sec. III), we obtain
lim
t
M
[N]
fs
(t) =
_
dpA
N
(p)h[g(p)]
= 0 if no recrossing of g(q) (26)
where A
N
(p) is dened in Eq. (13), and g(p) is dened
analogously to f(p), i.e.
lim
t
h[g(pt/m)] = h[g(p)] (27)
In the N limit, the contribution of r
N
(p) to
M
[N]
fs
(t) cancels out (for the same reason that it can-
cels out in C
[N]
fs
(t)see Sec. IV.B). Equation (26) is thus
equivalent to stating that the total contribution to A
N
(p)
from the spikes enclosed by h[g(p)] is zero if there is no
recrossing of g(q).
B. Eect of no recrossing orthogonal to f(q)
If there is no recrossing of any g(q) orthogonal to f(q),
we can use Eq. (26) to generate a set of equations giving
constraints on the spikes. Let us see what eect these
constraints have in the simple case that N = 3 and
f(q) = q
0
.
56
We can construct dividing surfaces g(q)
orthogonal to f(q) by taking any function of the normal
mode coordinates
Q
x
=
1
6
(2q
1
q
2
q
3
)
Q
y
=
1
2
(q
2
q
3
) (28)
Let us take
g
r
(q) =
_
Q
2
x
+Q
2
y
r
g
F
(q) = F[(Q
x
, Q
y
)] (29)
where r
q
i
+
i
/2
_
q
i,1
+
i,1
/2, . . . , q
i,F
+
i,F
/2
_
(33)
in Eq. (1), and integrating over the multi-dimensional co-
ordinates (q, z, ). The dividing surface f(q) is now in-
variant under collective cyclic permutations of the coordi-
nates q, and is thus a permutationally invariant function
of the replicas {
1
(q
1
), . . . ,
N
(q
N
)} of a (classical)
reaction coordinate (q
1
, . . . , q
F
).
It is straightforward to analyze the t behaviour
of C
[N]
fs
(t) by combining the analysis of Secs. III-V with
centre-of-mass-frame scattering theory. All we need to
note is that the relative motion of the reactant or prod-
uct molecules can be described by a one-dimensional scat-
tering coordinate, with all other degrees of freedom be-
ing described by channel functions
51
(which include the
rovibrational states of the scattered molecules, and also
specify whether the system is in the reactant or product
arrangement). We will denote the momentum of the ith
replica along the scattering coordinate as
i
, using the
convention that
i
is negative in the reactant arrange-
ment and positive in the product arrangement. Since all
other internal degrees of freedom are bound, it follows
that
lim
t
h[
i
(p
i
t/m)] = h(
i
) (34)
This last result allows us to construct a multi-
dimensional generalisation of the hybrid function C
[N]
fs
(t)
by replacing h[f(q)] in C
[N]
fs
(t) by h[
i
(q
i
)]. One can
show (by generalizing Appendix A) that the multi-
dimensional C
[N]
fs
(t) gives the exact quantum rate in the
limit t . We then take the t limits of C
[N]
fs
(t)
and C
[N]
fs
(t) by using the scattering relation
i
_
v
i
_
=
i,vi
_
(35)
where
i
_
is a momentum eigenstate,
v
i
_
is a reactant or
product channel function, and
i,vi
_
is a scatter-
ing eigenstate satisfying outgoing boundary conditions.
As in one-dimension, we obtain integrals over an N-
dimensional hypercube:
lim
t
C
[N]
fs
(t) =
_
dA
N
()h(
i
) (36)
lim
t
C
[N]
fs
(t) =
_
dA
N
()h[f()] (37)
where {
1
, . . . ,
N
}, and A
N
() is a generalisa-
tion of A
N
(p), obtained by making the replacements
of Eq. (33) in Eq. (13), replacing
pi
_
by
i,vi
_
,
and summing over v
i
. The function f() is a multi-
dimensional generalisation of f(p), and satises
lim
t
h[f(pt/m)] = h[f()] (38)
(which is equivalent to stating that f(q) separates cleanly
the reactants from the products in the limit t ).
The derivation of Appendix B generalizes straightfor-
wardly to multi-dimensions, with the result that A
N
()
has the analogous structure to A
N
(p) in Eq. (19). Fol-
lowing Sec. IV and Appendix C, one can show that only
the -function spikes [corresponding to the rst term in
Eq. (19)] contribute to C
[N]
fs
(t) and C
[N]
fs
(t) in the lim-
its t, N . There are many more of these spikes in
multi-dimensions than in one-dimension, since there is a
spike for every possible pair of (open) reactant or prod-
uct channels. However, it is possible to isolate each o-
diagonal spike by constructing angular functions F (see
Sec. V and Appendix D) in the space orthogonal to f().
It then follows that each o-diagonal spike in A
N
() con-
tributes zero to C
[N]
fs
(t) in the limits t, N , if there is
no recrossing of surfaces orthogonal to f(q) in the space
.
Hence we have obtained the same result in multi-
dimensions as in one-dimension: that the RPMD-TST
rate is equal to the exact quantum rate if there is no
recrossing of the dividing surface, nor of any surface or-
thogonal to it in an (N1)-dimensional space orthogonal
to f(q), which describes (in the t 0
+
limit) the uctu-
ations in the polymer-bead positions along the reaction
coordinate (q
1
, . . . , q
F
). It is impossible to recross these
orthogonal surfaces in the classical (i.e. high-temperature
limit), where RPMD-TST thus reduces to classical TST.
VII. CONCLUSIONS
We have shown that quantum TST (i.e. RPMD-TST)
is related to the exact quantum rate in the same way
that classical TST is related to the exact classical rate;
i.e. quantum TST is exact in the absence of recrossing.
Recrossing in quantum TST is more complex than in
classical TST, since, in addition to recrossing of the ring-
polymer dividing surface, one must also consider recross-
ing through surfaces that describe uctuations in the po-
10
sitions of the polymer beads along the reaction coordi-
nate. Such additional recrossing disappears in the classi-
cal and parabolic barrier limits, and thus becomes impor-
tant only at temperatures below the cross-over to deep
tunnelling. Previous RPMD-TST calculations
6
indicate
that the resulting loss in accuracy increases slowly as the
temperature is reduced below cross-over, such that quan-
tum TST remains within a factor of two of the exact rate
at temperatures down to below half the cross-over tem-
perature. However, it is clear that further work will be
needed in order to predict quantitatively how far one can
decrease the temperature below cross-over before quan-
tum TST breaks down (which will always happen at a
suciently low temperature).
Just as with classical TST, quantum TST will not
work for indirect reactions, such as those involving long-
lived intermediates, or diusive dynamics (e.g. the high-
friction regimes of the quantum Kramers problem
61
).
However, this leaves a vast range of chemical reactions
for which quantum TST is applicable, and for which it
will give an excellent approximation to the exact quan-
tum rate. The ndings in Part I and in this article
thus validate the already extensive (and growing) body
of results from RPMD rate-simulations
719
(which give
a lower bound to the RPMD-TST rate), as well as re-
sults obtained using the older centroid-TST method
2025
(which is a special case of RPMD-TST
6,26
).
Acknowledgments
TJHH is supported by a Project Studentship from the
UK Engineering and Physical Sciences Research Council.
APPENDIX A: Long-time limit of the hybrid
ux-side time-correlation function
Here we derive Eq. (8), which states that C
[N]
fs
(t) gives
the exact quantum rate in the t limit. We use the
property that
C
[N]
fs
(t) = C
[N]
sf
(t) =
d
dt
C
[N]
ss
(t) (A1)
where C
[N]
sf
(t) and C
[N]
ss
(t) are the side-ux and side-side
time-correlation functions corresponding to C
[N]
fs
(t). We
then write C
[N]
sf
(t) as
C
[N]
sf
(t) =
_
dq
_
d
1
h[f(q)]
q
1
1
/2|e
N
H
|q
2
_
i=3
q
i1
|e
N
H
|q
i
_
q
N
|e
N
H
|q
1
+
1
/2
_
q
1
+
1
/2|e
i
Ht/
F(q
)e
i
Ht/
|q
1
1
/2
_
(A2)
where
F(q
) is the ux operator
3
F(q
) =
1
2m
_
p(q q
) +(q q
) p
(A3)
and insert identities of the form e
i
Ht/
e
i
Ht/
to obtain
C
[N]
sf
(t) =
_
dq
_
d
1
h[f(q)]
q
1
1
/2|e
i
Ht/
e
N
H
e
i
Ht/
|q
2
_
i=3
q
i1
|e
i
Ht/
e
N
H
e
i
Ht/
|q
i
_
q
N
|e
i
Ht/
e
N
H
e
i
Ht/
|q
1
+
1
/2
_
q
1
+
1
/2|e
i
Ht/
F(q
)e
i
Ht/
|q
1
1
/2
_
(A4)
This allows us to take the t limit of C
[N]
sf
(t) by
using Eq. (9) together with the relation
p
i
_
=
+
pi
_
(A5)
where
+
lim
t
e
i
Ht/
e
i
Kt/
(A6)
and
+
pi
_
is a (reactive) scattering wave function with
incoming boundary conditions.
51
We then obtain
C
[N]
sf
(t) =
_
dp
_
dp
1
h
_
f
_
p
1
+p
1
2
, p
2
, . . . , p
N
__
+
p
1
|e
N
H
|
+
p2
_
i=3
+
pi1
|e
N
H
|
+
pi
_
+
p
N
|e
N
H
|
+
p1
_
+
p1
|
F(q
)|
+
p
1
_
(A7)
From the orthogonality of the scattering eigenstates,
51
we obtain
+
p
|e
N
H
|
+
p
_
= e
p
2
N
/2m
(p p
) (A8)
11
We also know that
h
_
f(p, p, . . . , p)
= h(p) (A9)
(since otherwise f(q) would not correctly distinguish be-
tween reactants and products in the limit t ). We
thus obtain
lim
t
C
[N]
sf
(t) =
_
dp e
p
2
/2m
h(p)
+
p
|
F(q
)|
+
p
_
(A10)
which is the t limit of the Miller-Schwarz-Tromp
ux-side time-correlation function,
3
from which we ob-
tain Eq. (8).
APPENDIX B: Derivation of the structure of AN(p)
Here we derive Eq. (19) of Sec. IV. We rst dene a
special type of side-side time-correlation function,
P
[N]
l
(E, t) =
_
dq
_
dz
_
d
h[f(q)]
_
_
N
i=1,i=l
h(z
i
q
)
_
_
i=1
q
i1
i1
/2|e
N
H
|q
i
+
i
/2
_
q
i
+
i
/2|e
i
Ht/
(
H E
i
)|z
i
_
z
i
|e
i
Ht/
|q
i
i
/2
_
(B1)
where E {E
1
, E
2
, . . . , E
N
}, and then consider its t
time-derivative,
Q
[N]
l
(E) = lim
t
d
dt
P
[N]
l
(E, t) (B2)
We can obtain two equivalent expressions for Q
[N]
l
(E),
by evaluating it as either a ux-side or a side-ux time-
correlation function. The ux-side version is
Q
[N]
l
(E) =
_
dq
_
dp
_
d
F[f(q)]
_
_
N
i=1,i=l
h(p
i
)
_
_
i=1
q
i1
i1
/2|e
N
H
|q
i
+
i
/2
_
q
i
+
i
/2|(
H E
i
)|
pi
_
pi
|q
i
i
/2
_
(B3)
which gives
|p
l
|
1
A
N
(p
1
, . . . , p
l
, . . . , p
N
)
+| p
l
|
1
A
N
(p
1
, . . . , p
l
, . . . , p
N
)
=
Q
[N]
l
(E)
m
N
N
i=1,i=l
|p
i
| if p
l
real (B4)
and
|p
l
|
1
A
N
(p
1
, . . . , p
l
, . . . , p
N
) =
Q
[N]
l
(E)
m
N
N
i=1,i=l
|p
i
| if p
l
imaginary (B5)
The side-ux version is
Q
[N]
l
(E) =
_
ds
_
ds
h[f(s +s
)]
_
N
i=1
+
s
i1
|e
N
H
|
+
si
_
_
+
s
l
|(
H E
l
)|
+
s
l
_
j=1,j=l
+
sj
|(
H E
j
)
F(q
)|
+
s
j
_
i=1,i=l
i=j
+
si
|(
H E
i
)
h(q
)|
+
s
i
_
(B6)
The second to fourth lines in this expression contain the
-functions,
(s
l
s
l
)
N
i=1
(s
i1
s
i
)[E
+
(s
i
) E
i
] (B7)
where E
+
(s
i
) is dened the other way round to E
(p
i
)
of Eq. (14), and where the -functions in s
i
and s
i
result
from the orthogonality of the scattering states
+
s
_
.
51
Integrating over s
i
and s
i
, we obtain
Q
[N]
l
(E) = b
N
(p)(E
l+1
E
l
) (B8)
where b
N
(p) is some function of p (which we do not
need to know explicitly). Substituting this expression
into Eqs. (B4) and (B5), we obtain
|p
l
|
1
A
N
(p
1
, . . . , p
l
, . . . , p
N
)
+| p
l
|
1
A
N
(p
1
, . . . , p
l
, . . . , p
N
)
= c
N
(p)(E
l+1
E
l
) if p
l
real (B9)
and
|p
l
|
1
A
N
(p
1
, . . . , p
l
, . . . , p
N
)
= c
N
(p)(E
l+1
E
l
) if p
l
imaginary (B10)
where c
N
(p) is some function of p. This derivation was
obtained for the case that p
i
> 0, i = l, but can clearly
be repeated for all combinations of positive and negative
p
i
[by replacing various h(z
i
q
) by h(z
i
+q
)]. Hence
Eqs. (B9) and (B10) holds for all p.
Now, we can obtain alternative expressions for the
righthand side of Eqs. (B9) and (B10) by adding and
subtracting sequences of terms that correspond to fol-
lowing dierent paths through the hypercube. Consider,
12
for example (for the case that p
i
, p
j
are both real), the
sequence
|p
j
|
1
A
N
(p
1
, . . . , p
i
, . . . , p
j
, . . . , p
N
)
+| p
j
|
1
A
N
(p
1
, . . . , p
i
, . . . , p
j
, . . . , p
N
)
= X
N
(p)(E
j+1
E
j
)
|p
i
|
1
A
N
(p
1
, . . . , p
i
, . . . , p
j
, . . . , p
N
)
+| p
i
|
1
A
N
(p
1
, . . . , p
i
, . . . , p
j
, . . . , p
N
)
= Y
N
(p)(E
i+1
E
i
)
| p
j
|
1
A
N
(p
1
, . . . , p
i
, . . . , p
j
, . . . , p
N
)
+|p
j
|
1
A
N
(p
1
, . . . , p
i
, . . . , p
j
, . . . , p
N
)
= Z
N
(p)(E
j+1
E
j
) (B11)
where each of X
N
(p), Y
N
(p), Z
N
(p) is some (dierent)
function of p. Combining these expressions, we obtain
|p
i
|
1
A
N
(p
1
, . . . , p
i
, . . . , p
N
)
+| p
i
|
1
A
N
(p
1
, . . . , p
i
, . . . , p
N
)
= P
N
(p)(E
i+1
E
i
) +Q
N
(p)(E
j+1
E
j
)
(B12)
where P
N
(p) = |p
j
|| p
j
|
1
Y
N
(p), and Q
N
(p) =
|p
j
|
_
|p
i
|
1
X
N
(p) +| p
i
|
1
Z
N
(p)
.
59
We can repeat this
procedure for each of the N 1 dierent values of j = i.
Because the resulting set of coecients P
N
(p) and Q
N
(p)
are linearly independent
62
, it follows that
|p
i
|
1
A
N
(p
1
, . . . , p
i
, . . . , p
N
)
+| p
i
|
1
A
N
(p
1
, . . . , p
i
, . . . , p
N
)
= d
N
(p)
N1
i=1
(E
i+1
E
i
) if p
i
real
(B13)
and
|p
i
|
1
A
N
(p
1
, . . . , p
i
, . . . , p
N
)
= d
N
(p)
N1
i=1
(E
i+1
E
i
) if p
i
imaginary
(B14)
where d
N
(p) is some function of p. From this, we obtain
Eq. (19) of Sec. IV.
APPENDIX C: Cancellation of the term rN(p) in
the limit N
Because the function f(p) must vary smoothly with
imaginary time and converge in the limit N , it can
be rewritten as a function of a nite number K of the
linear combinations
P
i
=
N
j=1
T
ij
p
j
i = 1, . . . , K (C1)
in which T
ij
N
1
(i.e. P
i
is normalised such that it
converges in the limit N ; e.g. T
0j
= N
1
corre-
sponds to the centroid). It then follows that f(p)/p
j
N
1
, and hence that
lim
N
f(p
1
, . . . , p
j
, . . . , p
N
) = f(p) + ( p
j
p
j
)
f(p)
p
j
(C2)
provided the range of p
j
p
j
is nite [which it is because
r
N
(p) contains Boltzmann factors]. We can therefore
write the N limit of Eq. (23) as
lim
N
S(N) =
_
dp
1
. . .
_
dp
j1
_
0
dp
j
. . .
_
0
dp
N
r
N
(p)
( p
j
p
j
)
f(p)
p
j
[f(p
1
, . . . , p
j
, . . . , p
N
)]
(C3)
which shows that the volume sandwiched between the
two heaviside functions becomes a strip of width ( p
j
p
j
)f(p)/p
j
N
1
in the limit N .
APPENDIX D: Isolating the o-diagonal spikes for
N > 3
It is straightforward to generalize the result obtained
for N = 3 and f(q) = q
0
in Sec. V.B to general N and
to any (cyclically invariant) choice of f(q).
We consider rst the special case of a centroid dividing
surface [f(q) = q
0
]. The space orthogonal to q
0
can be
represented by orthogonal linear combinations Q
i
, i =
1, . . . N 1 of q
i
, analogous to Q
x
and Q
y
in Sec. V.B.
We can then dene a generalized radial dividing surface
g
r
(q) =
_
N1
i=1
Q
2
i
r
(D1)
(which is invariant under cyclic permutation of the q
i
)
and generalized angular dividing surfaces
g
F
(q) = F[(Q
X
, Q
Y
)] (D2)
with
(Q
X
, Q
Y
) = arctan(Q
Y
/Q
X
) (D3)
where (Q
X
, Q
Y
) can be chosen to be any mutually or-
thogonal pair of linear combinations of the Q
i
. From
Eq. (27), the t limits of g
r
(q) and g
F
(q) are
g
r
(p) = lim
0
_
N1
i=1
P
2
i
(D4)
and
g
F
(p) = F[(P
X
, P
Y
)] (D5)
13
where P
i
and (P
X
, P
Y
) are the linear combinations of
p
i
analogous to Q
i
and (Q
X
, Q
Y
). We can then pro-
ceed as for the N = 3 example in Sec. V.B. Substitut-
ing g
r
(p) into Eq. (26), we obtain the constraint that
the spikes along the centroid axis contribute zero (since
g
r
(p) encloses these spikes only). This leaves us free
to construct angular dividing surfaces g
F
(q) in various
planes (Q
X
, Q
Y
) (which need not be mutually orthogo-
nal) in order to enclose individual o-diagonal spikes.
60
Equation (26) then gives a set of constraints, each stating
that the contribution to A
N
(p) from one of these spikes
is zero if there is no recrossing of any surface orthogonal
to f(q).
This reasoning can be applied with slight modication
to a general (cyclically invariant) dividing surface f(q).
By construction, such a surface reduces to a function of
just the centroid near the centroid axis, and hence there
exists a radial coordinate in the (N1)-dimensional curvi-
linear space orthogonal to f(q) which reduces to g
r
(q)
close to the centroid axis. We therefore obtain the con-
straint that the spikes along the centroid axis contribute
zero, and are then free to isolate each of the o-diagonal
spikes by using curvilinear generalisations of the angles
, which sweep over curvilinear surfaces that are orthog-
onal to f(q), and which reduce to the form of Eq. (D3)
close to the centroid axis.
1
T.J.H. Hele and S.C. Althorpe, J. Chem. Phys. 138,
084108 (2013).
2
D. Chandler and P.G. Wolynes, J. Chem. Phys. 74, 4078
(1981).
3
W.H. Miller, S.D. Schwartz and J.W. Tromp, J. Chem.
Phys. 79, 4889 (1983).
4
W.H. Miller, J. Chem. Phys. 61, 1823 (1974).
5
E. Pollak and P. Talkner, Chaos 15, 026116 (2005).
6
J.O. Richardson and S.C. Althorpe, J. Chem. Phys. 131,
214106 (2009).
7
I.R. Craig and D.E. Manolopoulos, J. Chem. Phys. 122,
084106 (2005).
8
I.R. Craig and D.E. Manolopoulos, J. Chem. Phys. 123,
034102 (2005).
9
R. Collepardo-Guevara, I.R. Craig and D.E. Manolopou-
los, J. Chem. Phys. 128, 144502 (2008).
10
R. Collepardo-Guevara, Y.V. Suleimanov and
D.E. Manolopoulos, J. Chem. Phys. 130, 174713
(2009); J. Chem. Phys. 133, 049902 (2010).
11
Y.V. Suleimanov, R. Collepardo-Guevara and
D.E. Manolopoulos, J. Chem. Phys. 134, 044131
(2011).
12
R. Perez de Tudela, F.J. Aoiz, Y.V. Suleimanov and D.E.
Manolopoulos, J. Phys. Chem. Lett. 3, 493 (2012).
13
S. Habershon, D.E. Manolopoulos, T.E. Markland and
T.F. Miller III, Annu. Rev. Phys. Chem. 64, 387 (2013).
14
Y.V. Suleimanov, J. Phys. Chem. C 116, 11141 (2012).
15
N. Boekelheide, R. Salom on-Ferrer and T.F. Miller III,
Proc. Natl. Acad. Sci. 108, 16159 (2011).
16
A.R. Menzeleev, N. Ananth and T.F. Miller III, J. Chem.
Phys. 135, 074106 (2011).
17
J.S. Kretchmer and T.F. Miller III, J. Chem. Phys. 138,
134109 (2013).
18
T. Stecher and S.C. Althorpe, Mol. Phys. 110, 875 (2012).
19
Y. Li, Y.V. Suleimanov, M. Yang, W.H. Green and H.
Guo, J. Phys. Chem. Lett. 4, 48 (2013).
20
M.J. Gillan, Phys. Rev. Lett. 58, 563 (1987).
21
M.J. Gillan, J. Phys. C 20, 3621 (1987).
22
G.A. Voth, D. Chandler and W.H. Miller, J. Chem. Phys.
91, 7749 (1989).
23
G.A. Voth, D. Chandler and W.H. Miller, J. Phys. Chem.
93, 7009 (1989).
24
S. Jang, C.D. Schwieters and G.A. Voth, J. Phys. Chem.
A 103, 9527 (1999).
25
E.M. McIntosh, K.T. Wikfeldt, J. Ellis, A. Michaelides and
W. Allison, J. Phys. Chem. Lett. 4, 1565 (2013).
26
To avoid confusion, note that we use the term centroid-
TST to refer to the quantum TST method introduced
in refs. 2023, and not to any application of the cen-
troid molecular dynamics (CMD) method of Voth and co-
workers [such as discussed in, e.g., S. Jang and G.A. Voth,
J. Chem. Phys. 112, 8747 (2000)].
27
W.H. Miller, J. Phys. Chem. A 105, 2942 (2001).
28
T. Yamamoto, H. Wang and W.H. Miller, J. Chem. Phys.
116, 7335 (2002).
29
J. Liu and W.H. Miller, J. Chem. Phys. 131, 074113
(2009).
30
W.H. Miller, Y. Zhao, M. Ceotto and S. Yang, J. Chem.
Phys. 119, 1329 (2003).
31
J. Vancek, W.H. Miller, J.F. Castillo and F.J. Aoiz, J.
Chem. Phys. 123, 054108 (2005).
32
W.H. Miller, Faraday Discuss. Chem. Soc. 62, 40 (1977).
33
W.H. Miller, R. Hernandez, N.C. Handy, D. Jayatilaka and
A. Willetts, Chem. Phys. Lett. 172, 62 (1990).
34
T.L. Nguyen, J.F. Stanton and J.R. Barker, Chem. Phys.
Lett. 499, 9 (2010).
35
W.H. Miller, J. Chem. Phys. 62, 1899 (1975).
36
C.G. Callan and S. Coleman, Phys. Rev. D 16, 1762
(1977).
37
V.A. Benderskii, D.E. Makarov and C.A. Wight, Chemi-
cal Dynamics at Low Temperatures, Adv. Chem. Phys. 88
(Wiley, New York, 1994).
38
S. Andersson, G. Nyman, A. Arnaldsson, U. Manthe and
H. J onsson, J. Phys. Chem. A 113, 4468 (2009).
39
Z. Smedarchina, W. Siebrand and A. Fern andez-Ramos, J.
Chem. Phys. 127, 174513 (2007).
40
J.B. Rommel and J. Kastner, J. Chem. Phys. 134, 184107
(2011).
41
J.B. Rommel, T.P.M. Goumans and J. K astner, J. Chem.
Theor. Comput. 7, 690 (2011).
42
S.C. Althorpe, J. Chem. Phys. 134, 114104 (2011).
43
D. Chandler, Introduction to Modern Statistical Mechanics
(Oxford University Press, New York, 1987).
44
D. Frenkel and B. Smit, Understanding Molecular Simula-
tion (Academic Press, London, 2002).
45
To clarify, note that we do not mean here the extended
(and ctitious) classical dynamics in ring-polymer space;
we mean applications of the exact quantum time-evolution
operator exp(i
Ht/).
46
D. Chandler, J. Chem. Phys. 68, 2959 (1978).
14
47
An alternative route to deriving these results could be to
use linear response theory [as was done a long time ago
for the Kubo-transformed version of the ux-ux time-
correlation function; see T. Yamamoto, J. Chem. Phys.
33, 281 (1960)].
48
A centroid dividing surface will break down also for a sym-
metric barrier at very low temperatures, when it becomes
necessary to include other (even) normal modes of the ring-
polymer in the dividing surface; see ref. 6.
49
E. Wigner, Z. Phys. Chem. B 19, 203 (1932).
50
This supplies the necessary proof of Eq. (23) of Part I,
which was deferred to the present article.
51
J.R. Taylor, Scattering Theory (Dover, New York, 2006).
52
Meaning that
x|
x|
x|p
+ R(p)
x| p