Brill Academic Publishers

P.O. Box 9000, 2300 PA Leiden,

The Netherlands
Lecture Series on Computer
and Computational Sciences
Volume 1, 2005, pp. 1-16
Grid Enabled Molecular Dynamics: classical and quantum
algorithms
S. C. Farantos
a,b,1
, S. Stamatiadis
a,b
, L. Lathouwers
c
and R. Guantes
d
a
Institute of Electronic Structure and Laser, Foundation for Research and Technology-Hellas,
Iraklion 711 10, Crete, Greece.
b
Department of Chemistry, University of Crete, Iraklion 711 10, Crete, Greece.
c
Department of Wiskunde-Informatica, University of Antwerp, Groenenborgerlaan 171, B2020,
Antwerp, Belgium.
d
Instituto de Matematicas y Fisica Fundamental, Consejo Superior de Investigaciones Cienticas,
Serrano 123, 28006 Madrid, Spain.
Received 1 August, 2005; accepted in revised form 10 August, 2005
Abstract: Molecular simulations have become a powerful tool in investigating the micro-
scopic behavior of matter as well as in calculating macroscopic observable quantities. The
predictive power and the accuracy of the methods used in molecular calculations are closely
related to the current computer technology. Thus, the rapid advancement of Grid com-
puting, i.e. the utilization of geographically distributed computers connected by relatively
high latency networks, is expected to inuence extensively the progress of computational
sciences, provided algorithms which can utilize the hundreds and even thousands of the
available computers in the Grid exist. In this lecture we review some of our computa-
tional methods, classical and quantum, used in small molecules, which seem promising for
studying large scale in time and molecular size problems. In particular, in classical molec-
ular simulations we are searching for specic trajectories connecting two regions of phase
space (rare events) by solving two-point boundary value problems with multiple shoot-
ing techniques. In quantum dynamics we argue that using variable order nite dierence
methods for solving the Schr odinger equation in a cartesian coordinate system result in
sparse Hamiltonian matrices which can make large scale problem solving feasible.
Keywords: Quantum molecular dynamics, classical molecular dynamics, nite dierences,
multiple shooting methods, Grid computing.
Mathematics Subject Classication: Here must be added the AMS-MOS or PACS Numbers
PACS: Here must be added the AMS-MOS or PACS Numbers
1 Introduction
Computational Chemistry is a topic whose progress is closely related to the current technology of
computers. Molecular simulations are important in designing new materials, pharmaceuticals, in
Biological Chemistry and Physics. In general, Molecular Computational Sciences face two chal-
lenges in the twenty rst century. The rst has to do with the size of the systems; studies which
cover molecules from two atoms to nanostructures and the macroscopic states of matter are re-
quired. The second challenge is related to time and the dynamics of the systems; phenomena
1
Corresponding author. E-mail: [email protected]
2 S.C. Farantos
which last from femtoseconds to seconds are important in material and biological sciences. Grid
(distributed) computing could be the solution in bridging the gaps in time and space. A Grid of
computers is envisioned as a seamless, integrated, computational and collaborative environment
embracing dierent categories of distributed systems. Following the classication introduced by
Foster and Kesselman (The Grid: the blueprint for a new computing infrastructure) [1], com-
putational Grids are categorized into ve major classes of applications. In summary, these classes
identify a specic context of applications such as supercomputing applications, high-throughput
computing, on-demand computing, data-intensive computing and collaborative computing. This
new paradigm in scientic computing is rapidly developing. Several Grid testbeds have been de-
ployed around the world among which the European Grid. The ENACTS project [2] with its well
planned and in depth studies has shown that Grids of computers are well developed not only at
the national level but also as a European multi-national integration.
In our days, the most successful computational Grid model is the one based on internet by
accessing thousands of PCs via running the programs as screensavers. The rst such applica-
tion was SETI@home to analyze the data from radio telescopes looking for signs of extrater-
restrial life. Taking its inspiration from SETI, a protein folding project called Folding@Home
(http://www.stanford.edu/group/pandegroup/folding) has been in operation at Stanford Univer-
sity for several years. Another active academic project is Predictor@home (http://predictor.
scripps.edu/) which is aiming to structures prediction. This is a pilot of the Berkeley Open Infras-
tructure for Network Computing (BOINC), a software platform for distributed computing using
volunteer computer resources. Another action on the human protein folding problem has been taken
by the World Community Grid (WCG) (http://www.worldcommunitygrid.org/) which is supported
by United Devices (UD). UD also runs the life-science research hub (http://www.grid.org/) to
search for drug candidates to treat the smallpox virus.
In order to use successfully a worldwide distributed computing environment of hundreds or even
thousands of heterogeneous processors such as the Grid, communications among these processors
should be minimum. There are not many molecular simulations using the Grid environment such as
to allow us to point out the strategy one should adopt in writing codes for molecular applications.
A practical rule is to allow each processor to work independently, even though calculations are
repeated, and only if something important happens to one of them, then they communicate. The
common ingredient of the above successful applications in Internet Grid Computing is the com-
putational algorithm employed, which guarantees minimum and asynchronous communications.
Obviously, there are few problems which can be solved with such algorithms. Most of the interest-
ing problems of Computational Chemistry require solutions of a large number of linear equations
that can not be solved without signicant communication among the computer nodes. On the other
hand, parallelized codes written for large parallel machines may be of no good use when thousands
of computers should be exploited connected by high latency networks. Thus, new algorithms and
new programming paradigms suitable for distributed computing should be investigated in order to
exploit Computational Grids [3].
In this lecture we review some of our methods used in classical and quantum dynamics of
highly excited small molecules and we examine their potential to be utilized in a Grid computa-
tional environment. In quantum dynamics we shall investigate the combination of time evolution
of wave packets in cartesian coordinate representations with angular momentum projections and
the realization of this technique via generalizable time evolution schemes and variable order -
nite dierencing methods for computing the derivatives. With respect to classical dynamics we
shall explore the idea of searching specic trajectories which connect regions of phase space that
correspond to dierent conformations of the molecule by formulating the problem as a two-point
boundary value and using multiple shooting techniques.
Grid Enabled Molecular Dynamics 3
2 Computational Methods
Our understanding of basic molecular phenomena is strongly based on our ability to successfully
solve the Schr odinger equation for realistic molecular systems. This is the aim of Quantum Molec-
ular Dynamics (QMD) and it is the way to obtain reliable quantitative predictions. However, very
useful physical insight can be gained from the use of a causal time dependent picture where, for
instance, the path of a molecular encounter or a vibrational motion can be traced in conguration
space. Therefore, a lot of eorts have been directed towards the development of methods which
combine classical description with quantum corrections. Accurate wave packet propagation, to-
gether with knowledge of the proper classical orbits of the system, is one of the most complete
views that we can have for a molecular process because exact numerical results are supplemented
by a very clear physical picture. Such a program is accomplished by locating families of periodic
orbits which then can be used to determine initial congurations for the wave packets [4, 5].
In the next two subsections we describe our approaches to QMD and Classical Molecular
Dynamics (CMD) which are based on discretizing the conguration space and time respectively.
The grid representation of space and time make the algorithms suitable for using them in a Grid
of computers.
2.1 Quantum Molecular Dynamics
Grid methods for the solution of the Schr odinger equation are nowadays one of the most powerful
and exploited tools both for the time independent (known as Discrete Variable Representation
(DVR)) and time dependent pictures. Concerning the last one, it was the introduction of the
Fourier Pseudospectral (PS) method which provided the necessary accuracy and computational
eciency to compete with the traditional variational techniques and to make feasible the task
(for an excellent review see Ref.[6]). In the time dependent picture or wave packet propagation,
the wave packet is advanced in time by an evolution operator which, if the Hamiltonian

H is time
independent, is an exponential function

U(t;

H) = e
i

Ht
(we put = 1). This can be approximated
by, for instance, a polynomial expansion [7]. The basic operation is therefore reduced to the
evaluation of the action of the Hamiltonian operator onto the wave packet,

H. In long time
propagation this operation has to be repeated many times, and thus, its eciency will generally
determine the computational cost of the problem. In practice, the wave function is often discretized
in a grid using a collocation method and the Hamiltonian operator is represented by a matrix. If
we considered only local operators this matrix would be diagonal, but the Hamiltonian includes the
kinetic energy operator which is non local in the coordinate domain. If the number of collocation
points in each coordinate is n

, the Hamiltonian matrix will contain of the order of N n

nonzero elements, where N is the total number of collocation points N =

and n =

.
In the time independent picture one is concerned with the diagonalization of the Hamiltonian
matrix constructed in the same way to obtain the eigenvalues and eigenvectors. Usually, an iterative
procedure such as the Lanczos method and variants [8] is used for ecient diagonalization, which
again involves products of the Hamiltonian matrix with vectors as the basic step. It is clear that
important computational savings are possible when the Hamiltonian matrix is sparse or its action
on the wave function can be calculated eciently.
A way to increase signicantly the sparsity of the Hamiltonian matrix is by the use of local
methods. Here, local means that the action of the kinetic energy operator (or the Laplacian)
on the wave function is approximated by using only local information or neighboring grid points.
Although the derivative of a function is a local property, the wave function is dened on the
whole conguration space and a piecewise representation of a function by a local polynomial
approximation, generally, converges more slowly than a spectral representation.
In this work we review some recent advances in Finite Dierence algorithms which may result
4 S.C. Farantos
in sparse Hamiltonian matrices. However, developing algorithms for QMD suitable to use the
Grid computing technologies we must rst decide the coordinate system to express the molecular
Hamiltonian.
2.1.1 The angular momentum projection method
In their 1996 review of Quantum Molecular Dynamics [9] R.E. Wyatt and J.Z.H. Zang state it is
very important to develop and apply methods that can be extended to moderately sized polyatomic
molecules (5 to 12 atoms). In spite of the diversity of theoretical approaches and the ever growing
computational facilities it is safe to say that the calculation of vibration-rotation spectra or state
to state cross sections for ve or more atomic systems is all but routine. We will rst examine the
reasons for this state of aairs and then propose a scheme that can satisfy the above ambitions.
The standard way of describing a polyatomic system in the absence of external elds is to
introduce the center of mass, three Euler angles and 3N6 internal coordinates. The Hamiltonian
in mass weighted cartesian coordinates x
km
is written

H =
1
2

k,m

2
x
2
km
+V (R), (1)
where, k = 1, 2, 3 denotes the three cartesian coordinates, m = 1, . . . , N the atoms and V (R) the
potential energy as a function of the internal coordinates R.
The transformation of the original cartesian Hamiltonian to the above general curvilinear co-
ordinates yields the generic form

H =

H
0
_
q,

q
_
+

k
A
k
_
q,

q
_
J
k
+

k,l
1
2
I
1
kl
(q)J
k
J
l
. (2)
J
k
denote the components of angular momentum and I
kl
are the components of the moment of
inertial tensor.
In the above equation we recognize the vibrational, rotational and the vibrational-rotational
coupling parts of the Hamiltonian. We will refer to the derivation of Eq. (2) as frame transfor-
mation (FT). The problem with Eq. (2) is twofold. First, there is the analytical derivation of
the Hamiltonian parts in curvilinear coordinates q
1
, . . . , q
3N6
, and Euler angles , , or angular
momentum operators. Not only this is a cumbersome task for more than three atoms but this
procedure has to be repeated for every choice of curvilinear coordinates and angles. Secondly, the
general expression for the vibrational Hamiltonian reads

H
0
_
q,

q
_
=
1
2

i,j

q
i
_

gg
ij

q
j
_
+V (q), (3)
where g
ij
is the metric tensor of the curvilinear coordinates. The appearance of cross derivatives
implies that the number of partial dierential operators necessarily exceeds (3N 6)(3N 5)/2.
Therefore, the computational eort scales as N
2
.
The question then arises whether an alternative theoretical approach can be devised that avoids
the inherent problems of frame transformation and that can readily be extended to ve or more
atoms. Our basic starting point is to not separate rotations at the operator level but to use
projection techniques to conserve total angular momentum. An alternative version of Eq. (2) is to
introduce N Jacobi vectors the last one of which is the center of mass. The Hamiltonian can then
be written as

H =
1
2

k,m

2
u
2
km
+V (R(u)), (4)
Grid Enabled Molecular Dynamics 5
which contains the 3N 3 kinetic energies of the Jacobi coordinates u
km
, k = 1, 2, 3 and m =
1, . . . , N 1.
We rst observe that all molecular properties such as spectra, cross sections, etc., can be
obtained by Fourier transforming the appropriate time-correlation function. For specied total
angular momentum, J, the vibration-rotation spectra require the calculation of the autocorrelation
function
C
J
(t) =< (0)| exp(i

H
J
t)|(0) >, (5)
while for other properties the appropriate operator has to be inserted in the matrix element.
Consider the rotations about the center of mass which constitute the rotation group of the system.
The corresponding operators on the system wave functions are
R() = exp(iJ
z
) exp(iJ
y
) exp(iJ
z
), (6)
while, the associated irreducible representations are the Wigner functions
D
J
MK
() = exp(iM)d
J
MK
() exp(iK). (7)
From the group elements and the irreducible representations one can construct the group pro-
jection operators
P
J
MK
=
2J + 1
8
2
_
dD
J
MK
()R(). (8)
Their properties have been listed and proven elsewhere [10]. The ones we need in the present
context are

J
P
J
= 1, with P
J
=

K
P
J
KK
, (9)
i.e., one can construct the projector onto the space of total angular momentum J as the K sum
of the operators P
J
KK
. The operators P
J
, being true projection operators, are idempotent and
hermitian and, because the Hamiltonian is rotationally invariant, they commute with

H
P
2
J
= P
J
, P

J
= P
J
, [

H, P
J
] = 0. (10)
Since, the operator

H
J
appearing in Eq.(5) is really P
J

HP
J
, one can rewrite the formula for the
autocorrelation function as
C
J
(t) =< (0)|P
J
exp(i

Ht)P
J
|(0) > . (11)
This expression is the key to the method. Indeed, it shows that evaluating C
J
(t) can be regarded
as a three stage process: the projection of the initial state onto the subspace J, the time evolution
of the initial state under the total Hamiltonian and the calculation of the overlap of the projected
and time evolved states.
The projection of the initial state is a one time numerical integration and is computationally
irrelevant versus the time evolutions. The time evolution of the 3N 6 variables involves cross
derivatives in a 2J +1 dimensional matrix. The alternative propagation under Eq. (4) uses 3N3
strictly separated variables. One can make a rough estimate of the ratio of the computational
eort of calculating the action of Eq. (2) versus Eq. (4). Assuming that all variables are mapped
on a grid of P points we nd (2J + 1)(3N 6)P
3
.
However there is an extra savings device for the evolution under Eq. (4). Indeed, the wave
function will not extend over the entire Jacobi hyperesphere but rather it is conned to the corre-
sponding hypersphere.
6 S.C. Farantos
2.1.2 Variable order nite dierence method
Now we proceed to compute the action of the Hamiltonian operator onto the wave function.
In previous papers [11, 12], we reported results from the application of a variable order Finite
Dierence (FD) method to approximate the action of a Hamiltonian operator on the wave function
in the time-dependent Schr odinger equation or the Hamiltonian matrix elements in the time-
independent picture. One, two and three dimensional model potentials in radial and angular
coordinates were used to investigate the accuracy and the stability of these methods, whereas
in a companion paper [13], the time-dependent Schr odinger equation was solved for the van der
Waals system Ar
3
. The impetus for this project was given by recent advances in high order Finite
Dierence approximations. We mainly refer to the limit of innite order Finite Dierence formulas
with respect to global Pseudospectral methods (PS) investigated by Fornberg [14], and Boyds
work [15] which views Finite Dierence methods as a certain sum acceleration of Pseudospectral
techniques.
Finite Dierence approximations of the derivatives of a function F(x) can be extracted by
interpolating F(x) with Lagrange polynomials, P(x). This allows one to calculate the derivatives
analytically at arbitrarily chosen grid points and with a variable order of approximation. The
Lagrange fundamental polynomials of order N 1 are dened by
L
k
(x) =
N

j=1

(x x
j
)/
N

j=1

(x
k
x
j
), k = 1, 2, ..., N, (12)
where the prime means that the term j = k is not included in the products. The values of L
k
(x
j
)
are zero for j = k and one for j = k by construction. The function can then be approximated as
F(x) P
N
(x) =
N

k=1
F(x
k
)L
k
(x). (13)
P
N
is a polynomial of order N 1. In Ref.[11] we discussed how FD is related to the Sinc-DVR
method by taking the limit in the two above mentioned senses:
i) An innite order limit of centered FD formulas on an equi-spaced grid yields the Discrete
Variable Representation result when we use as a basis set the Sinc functions (Sinc(x) sin(x)/x)
[14, 17]. Although, this limit is dened formally as N, the number of grid points used in the
approximation, tends to innity, some theoretical considerations [14] as well as numerical results
[11] lead us to expect that the accuracy of the FD approximation is the same to that of the DVR
method as we approach the full grid to calculate the FD coecients.
ii) FD can also be viewed as a sum acceleration method which improves the convergence of the
Pseudospectral approximation [15]. The rate of convergence is, however, non-uniform in the wave
number, giving very high accuracy for low wave numbers and poor accuracy for wave numbers
near the aliasing limit [18]. However, this does not cause a severe practical limitation, since, by
increasing the number of grid points in the appropriate region we can have an accurate enough
representation of the true spectrum in the range of interest. This is one property which makes
FD useful as an alternative to the common DVR [19] and other PS methods such as Fast Fourier
Transform techniques (FFT) [6, 7].
The current interest in Finite Dierence methods is fully justied when solutions of the Schr odinger
equation are required for multidimensional systems such as polyatomic molecules. The present
most popular methods employed in quantum molecular dynamics are the Fast Fourier Transform
and the Discrete Variable Representation techniques. FFT generally uses hypercubic grid domains
which result in wasted conguration space sampling. A large number of the selected conguration
points correspond to high potential energy values, which do not contribute to the eigenstates that
Grid Enabled Molecular Dynamics 7
we are seeking. Global DVR methods allow us to choose easily the conguration points which are
relevant to the states we want to calculate, but still, we must employ in each dimension all grid
points. Local methods such as FD have the advantages of DVR but also produce sparse matrices
which make them more economical in computer memory and time provided the PS accuracy is
achieved at lower order than the high order limit.
There are some other benets for FD with respect to global Pseudospectral methods. Conver-
gence can be examined not only by increasing the number of grid points but also by varying in
a systematic way the order of approximation of the derivatives. Finite Dierence methods may
incorporate several boundary conditions and choose the grid points without necessarily relying on
specic basis functions. The topography of the multidimensional molecular potential functions is
usually complex. The ability of using non equi-spaced grids is as important as keeping the grid
points in accordance to the chosen energy interval. The computer codes for a FD representation
of the Hamiltonian can be eciently parallelized, since the basic operation is the multiplication of
a vector by a sparse matrix. The ongoing, active research in the relevant eld produced various
approaches to the ecient distribution of a sparse matrix, especially a regular one, i.e., one
formed through the discretization of a dierential equation on a grid, and the subsequent multipli-
cation with an appropriately distributed vector. The developed methods tackle, in various degrees,
the requirements of load balancing and minimal communication cost across a processor grid. The
overview presented in [20] is a starting point for the available techniques.
Figure 1: Pattern of a Hamiltonian matrix for a system of three degrees of freedom with the kinetic
energy operator evaluated by nite dierence approximation using three point stencils.
In Figure 1 we show schematically the Hamiltonian matrix when we employ a 3-point dierence
8 S.C. Farantos
method to calculate the second derivatives in a three dimensional system. The high sparsity of the
matrix is inherent, and this is the best for distributed computing.
As an example we will study a two dimensional system employed in several investigations in the
past, mainly in the connection between classical and quantum dynamics. One of us [21] carried
out extensive studies of the periodic orbit structure of this system and its relation to quantum
mechanics. The system is described by the Hamiltonian:
H =
1
2
(p
2
x
+p
2
y
) +
1
2
(
2
x
x
2
+
2
y
y
2
) x
2
y, (14)
where the parameters are
2
x
= 0.9,
2
y
= 1.6 and = 0.08. The time dependent Schr odinger
equation was solved numerically using a Chebychev expansion [7] for the propagation in time,
while the action of the Hamiltonian operator on the wave function was evaluated with the FFT
method and with matrix-vector multiplication using FD in order to study the convergence and
compare the computation times. The resulting vibrational spectrum was obtained as usual with
the Fourier transform of the autocorrelation function of the wave packet
I(E) =
1
2
_

exp(iEt) < (x, 0)|(x, t) > dt. (15)

We propagated a Gaussian wave packet initially localized on a 1:2 resonance periodic orbit at
a high energy (E = 23 a.u., see Fig. 12b in Ref.[21]). It is interesting to plot the wave packet in
conguration space and compare the solution of the Schr odinger equation using the two methods.
In both cases we used a rectangular grid with 64 points in each dimension, and x = 0.3175. We
wanted to obtain a high resolution for the spectrum (E 0.056), and thus, we propagated the
wave packet for 1024 time steps. At t = 1/4 of the total time we took a snapshot of the wave
function. In Fig. 2a we show the wave function obtained with the FFT method (the potential
energy contour is superimposed on the same plot at the mean energy of the wave packet). In Figs
2b, c and d we show the same wave packet obtained with a second, fourth and seventh order Finite
Dierence approximation, respectively. The order of approximation (M) yields stencils with 2M+1
points. From Fig. 2 we can see that convergence of the wave function is approached with the FD
method, even for a wave packet quite spread in conguration space. Finer details reproduced by
even higher order FD approximations, however, will not aect very much the spectrum, since, this
is an average of the propagated wave function over the conguration space.
The resulting spectra are shown in Fig. 3 (here we omit the second order FD for the sake
of clarity, although the comparison is very poor as expected). Even with order M = 7, the only
appreciable dierence is in the intensities at high energies.
We recall that, a FD approximation is equivalent to a Taylor expansion series of the kinetic
energy spectrum, therefore, in the limit of x 0 we recover the exact spectrum irrespectively
of the order of the approximation. It is worth to investigate then how the FD converges to the
Fourier method as we increase the order as well as we decrease the grid spacing. We examine
the dierences of the central eigenvalue (E = 22.18) from that obtained with the FFT method
with 64 points in each dimension. When this dierence is less than the resolution in the power
spectrum the results are considered identical. Of course, for higher eigenvalues we should increase
further the order or decrease the grid spacing to converge to the desired resolution, but the general
behavior is seen in Fig. 4 where we used 64, 80, 100 and 120 points (x = 0.3175, 0.2532, 0.2020
and 0.1681, respectively). It is seen that convergence in both directions is quite fast, although it
is computationally cheaper on a single computer to increase the order of the approximation and
take less grid points than to increase the number of grid points (by increasing the order by one we
should add two more grid points to evaluate the Laplacian, but we should increase N by about 20
points at low order to get the same reduction in the error).
Grid Enabled Molecular Dynamics 9
(a)
(b)
(c)
(d)
FFT M=2
M=4 M=7
Figure 2: Snapshots of the wave packet at t = 28 a.u. (1/4 of total propagation time) with a
rectangular grid and 64 grid points in each dimension. (a) FFT method; (b) FD with M = 2
(second order); (c) FD 4
th
order; (d) FD 7
th
order. The stencils are with 2M + 1 points.
2.2 Classical Molecular Dynamics
The term Molecular Dynamics usually implies a Monte Carlo method which involves the integration
of classical trajectories in phase space and a random selection for their initial conditions. Dealing
with large molecules convergence is not an easy task and it usually requires proper sampling of
phase space as well as the integration of the equations of motion for very long times or the sampling
of a large number of trajectories. Therefore, it is useful if we know the type of trajectories we want
to sample, for example those which visit the regions in conguration space of reactants and products
in chemical reactions or trajectories which originate and end to particular phase space structures
like stable and unstable manifolds in phase space. In those cases where the initial and nal states of
a trajectory satisfy a known relation, it is preferable to solve a two-point boundary value (2PBV)
problem than randomly sampling initial conditions for integrating the equations of motion and
accepting or rejecting trajectories according to the nal state of the trajectory. A typical 2PBV
problem is the location of periodic orbits (PO).
2.2.1 The 2PBV problem
A 2PBV problem is formulated in the following way. The phase space of a molecule is described with
the generalized coordinates q
i
(t), i = 1, . . . , N and their conjugate momenta p
i
(t), i = 1, . . . , N
which are functions of time t. For the simplication of the mathematical equations we dene the
10 S.C. Farantos
12 17 22 27
E(a.u.)
0.0000
0.0002
0.0004
0.0006
I
n
t
e
n
s
i
t
y
Figure 3: Power spectra obtained from the correlation function of the wave packet shown in Fig.
2. FFT(solid line), FD 4
th
order (dashed line) and FD 7
th
order (dotted line). The stencils are
with 2M + 1 points.
column vector
x = ( q, p)
T
. (16)
Using x we can write Hamilton equations in the form
dx(t)
dt
=

x(t) = JH[x(t)] (0 t T), (17)
where H is the Hamiltonian function, and J is the symplectic matrix
J =
_
0
N
I
N
I
N
0
N
_
. (18)
0
N
and I
N
are the zero and unit N N matrices respectively. JH(x) is a vector eld, and J
satises the relations,
J
1
= J and J
2
= I
2N
. (19)
We want to nd trajectories whose the initial, x(0), and nal point, x(T), satisfy the equation

B(x(0), x(T); T) = 0. (20)

Grid Enabled Molecular Dynamics 11
0 1 2 3 4 5 6
order(M)
0.5
0.5
1.5
2.5
3.5
e
r
r
o
r
resolution
N=64
N=80
N=100
N=120
Figure 4: Dierences in energy with respect to the central eigenvalue (E = 22.18) in the FFT
spectrum of Fig. 3 as a function of the order of approximation. Rectangular grids with 64, 80, 100
and 120 points were used.
Treating the time T as a parameter we are searching for families of trajectories which are solutions
of the Hamilton equations of motion and simultaneously satisfy the two-point boundary conditions,
Eq. (20). For example, trajectories with known distance of the initial and nal points
x(T) x(0) =

, (21)
the boundary conditions are

B(x(0), x(T)) = x(T) x(0)

= 0. (22)
= 0 yields the boundary conditions for periodic orbits.
The common procedure to solve a 2PBV problem is to cast it into an initial value problem
through an iterative method (Ref.[5] and references therein). We consider the initial values of the
coordinates and momenta, s, of an approximate solution
s = x(0), (23)
as independent variables of the nonlinear functions

B(s; T) =

B(s, x(T; s)). (24)
12 S.C. Farantos

B parametrically depends on the period T. We denote the roots of boundary equations as s

, i.e.,

B(s

; T) = 0. (25)
Hence, if s is a nearby value to the solution s

we can compute the functions

B(s) by integrating
Hamilton equations for the time interval [0, T]. By appropriately modifying the initial values s we
hope to converge to the solution, that is, s s

and B 0.
A common procedure to nd the roots of Eqs (25) is the Newton-Raphson method. This is an
iterative scheme and at each iteration, k, we update the initial conditions of the orbit
s
k+1
= s
k
+ s
k
. (26)
The corrections s
k
are obtained by expanding Eqs (25) in a Taylor series up to the rst order

B(s
k+1
; T) =

B(s
k
+ s
k
; T)

B(s
k
; T) +

B
s
k
s
k
= 0, (27)
or

B(s
k
; T) +
_

B
s
k
Z
k
(0) +

B
x
k
(T)
Z
k
(T)
_
s
k
= 0. (28)
The matrix (Jacobian)
Z
k
(t) =
x
k
(t; s
k
)
s
k
, (29)
is the Fundamental Matrix, which is evaluated by integrating the variational equations

(t) = A(t)

(t) (0 t T), (30)

where the second derivatives of the Hamiltonian with respect to coordinates and momenta are
needed
A(t) = J
2
H[x(t)]. (31)
These equations calculate the linearized part of the dierence of two initially neighboring trajec-
tories in time,

= x

x. The Fundamental Matrix is also a solution of the variational equations

as can be seen by taking the time derivatives of Eq. (29)

Z(t) = A(t)Z(t). (32)

Thus, to perform the k
th
iteration in the Newton-Raphson method we rst integrate for time [0, T]
the dierential equations

x
k
(t) = JH[x
k
(t)]

Z
k
(t) = A
k
(t)Z
k
(t),
with initial conditions
x
k
(0) = s
k
Z
k
(0) = I
2N
.
Then, we solve the linear algebraic equations
_

B
s
k
Z
k
(0) +

B
x
k
(T)
Z
k
(T)
_
s
k
=

B(s
k
; T), (33)
in order to nd the initial conditions for the (k + 1)
th
iteration (Eq. (26)).
Grid Enabled Molecular Dynamics 13
For a periodic orbit the boundary conditions become
[Z
k
(T) I
2N
]s
k
=

B(s
k
; T). (34)
The Fundamental Matrix at period T is called Monodromy Matrix, M = Z(T), and its eigenvalues
are used for the stability analysis of the periodic orbit. Details can be found in Ref.[22].
Once we have located one member of the family of the trajectories that we are looking for, we
can apply continuation techniques [23] to nd trajectories at dierent times T. This is done by
using T as the control parameter. Usually, for small increments of the time T linear extrapolation
methods are sucient [5].
2.2.2 The multiple node 2PBV algorithm
Let us assume that we divide the time T in (m1) time intervals by choosing m nodes. The idea
of the multiple shooting method is to integrate m1 trajectories -one for each time sub-interval-
and to check if the nal values of each trajectory coincide with the initial conditions of the next.
To formulate this idea we introduce the scaled time, = t/T, (0 1),
0 =
1
<
2
< <
m1
<
m
= 1. (35)
Thus, for the simple shooting method m = 2.
t = 0
C
B
x( ; s )
s
s
s
1
s
x( ; s )
x( ) ; s

x(
C
x( s
t T =

m-2
1
1
1
2
3
4
3
2
C
C
2
3
4
2
3
s
m-1
m-2
)
m-1
;
m-1
)
,
m
m
; s
Figure 5: A schematical representation of the multiple shooting procedure applied to periodic
orbits.
In this section we drop the index for the iterations k, and we use the index j to denote the
nodes in time. If the initial conditions of the trajectory at each node j is s
j
at time
j
, and the nal
14 S.C. Farantos
value of the trajectory at time
j+1
is denoted by x(
j+1
; s
j
), then (m 2) continuity conditions
should be satised (for a graphical representation see Fig. 5)

C
j
(s
j
, s
j+1
) = x(
j+1
; s
j
) s
j+1
= 0, j = 1, 2, , m2, (36)
together with the boundary conditions

B(s
1
, s
m1
) = 0. (37)
Now, we have to solve (m1) initial value problems, and for that we adopt the Newton-Raphson
method

C
j
(s
j
+ s
j
, s
j+1
+ s
j+1
)

C
j
(s
j
, s
j+1
) +

C
s
j
s
j
+

C
s
j+1
s
j+1
= 0. (38)
These equations become

C
j
(s
j
, s
j+1
) +Z
j
(
j+1
)s
j
s
j+1
= 0, 1 j m2, (39)
where,
Z
j
(
j+1
) =
x(
j+1
; s
j
)
s
j
. (40)
Linearizing the boundary conditions we get

B(s
1
+ s
1
, s
m1
+ s
m1
) =

B(s
1
, s
m1
) +

B
s
1
Z
1
(0)s
1
+

B
s
m1
Z
m1
(1)s
m1
= 0, (41)
Hence, Eqs (39,41) provide a linear system with 2N(m1) unknown variables. For periodic orbit
boundary conditions these equations in matrix form are written as
_

_
Z
1
I
2N
0 0 0
0 Z
2
I
2N
0 0

0 0 0 Z
m2
I
2N
I
2N
0 0 0 Z
m1
_

_
_

_
s
1
s
2

s
m2
s
m1
_

_
=
_

C
1

C
2

C
m2

B
_

_
(42)
Although each block Z
i
is not sparse the total matrix is sparse and it can be solved with special
routines for distributed computing [24, 25, 26].
3 Conclusions
The algorithms described in the previous sections were used extensively in the past mainly with
triatomic molecules [27, 28, 29]. Currently, we are developing the computer codes to apply the
classical and quantum algorithms to large molecules with biological interest.
Acknowledgment
Support from the Greek Ministry of Education and European Union through the postgraduate
program EPEAEK, Applied Molecular Spectroscopy, is gratefully acknowledged. LL thanks the
Institute of Electronic Structure and Laser-FORTH for the hospitality during his visits to Crete.
Grid Enabled Molecular Dynamics 15
References
[1] I. Foster and C. Kesselman, (eds.). The Grid: Blueprint for a new Computing Infrastructure,
Morgan Kaufmann (1999).
[2] European Network for Advanced Computing Technology for Science (ENACTS),
http://www.epcc.ed.ac.uk/enacts/.
[3] http://tccc.iesl.forth.gr/general/intro/pdf/108.pdf
[4] S. C. Farantos, Exploring Molecular Vibrations with Periodic Orbits, Int. Rev. Phys. Chem.
15, 345374 (1996).
[5] S. C. Farantos, POMULT: A Program for Computing Periodic Orbits in Hamiltonian Systems
Based on Multiple Shooting Algotithms, Comp. Phys. Comm. 108, 240258 (1998).
[6] R. Koslo, Quantum Molecular Dynamics on Grids, Dynamics of Molecules and Chemical
Reactions, Editors: R. E. Wyatt and J. Z. H. Zhang, Marcel Dekker Inc., N.Y., 185230
(1996).
[7] R. Koslo, Propagation Methods for Quantum Molecular Dynamics, Ann. Rev. Phys. Chem.
45, 145178 (1994).
[8] (a) C. Lanczos, J. Res. Nat. Bur. Stand. 45, 58 (1950); (b) C. Iung and C. Leforestier, Direct
calculation of overtones: Application to the CD3H molecule, J. Chem. Phys. 102, 84538461
(1995); (c) R. E. Wyatt, Computation of high-energy vibrational eigenstates: Application to
C6H5D, J. Chem. Phys. 103, 84338443 (1995); (d) R. B. Lehoucq, D. C. Sorensen and C.
Yang, Solution of Large Scale Eigenvalue Problems with Implicitly Restarted Arnoldi Methods,
(SIAM, Philadelphia (1998)).
[9] Dynamics of Molecules and Chemical Reactions, Editors: R. E. Wyatt and J. Z. H. Zhang,
Marcel Dekker, Inc., N.Y. (1996).
[10] J. Broeckhove and L. Lathouwers, Qunatum Dynamics and Angular Mimentum Projection,
Numerical Grid Methods and their Applications to Schr odinger Equation, Editor: C. Cerjan,
Kluwer Academic Publishers, 4956 (1993).
[11] R. Guantes and S. C. Farantos, High Order Finite Dierence Algorithms for Solving the
Schr odinger Equation in Molecular Dynamics, J. Chem. Phys. 111, 1082710835 (1999).
[12] R. Guantes and S. C. Farantos, High Order Finite Dierence Algorithms for Solving the
Schr odinger Equation in Molecular Dynamics. II. Periodic variables, J. Chem. Phys. 113,
1042910437 (2000).
[13] R. Guantes, A. Nezis and S. C. Farantos, Periodic Orbit - Quantum Mechanical Investigation
of the Inversion Mechanism of Ar
3
, J. Chem. Phys. 111, 1083610842 (1999).
[14] (a) B. Fornberg, A Practical Guide to Pseudospectral Methods, Cambridge Monographs on
Applied and Computational Mathematics (Cambridge Univ. Press, 1 (1998)); (b) B. Fornberg
and D. M. Sloan, Acta Numerica, 203-267 (1994).
[15] J. P. Boyd, A fast algorithm for Chebyshev, Fourier, and sinc interpolation onto an irregular
grid, J. Comp. Phys. 103, 243257 (1992).
[16] J. P. Boyd, Sum-accelerated pseudospectral methods: Finite dierences and sech-weighted
dierences, Comp. Methods Appl. Mech. Engrg. 116, 111 (1994).
16 S.C. Farantos
[17] D. T. Colbert and W. H. Miller, A novel discrete variable representation for quantum me-
chanical reactive scattering via the S-matrix Kohn method, J. Chem. Phys. 96, 19821991
(1992).
[18] W. H. Press, B. P. Flannery, S. A. Teukolsky and W. T. Vetterling, Numerical Recipes,
Cambridge University Press (1986).
[19] (a) J. V. Lill, G. A. Parker and J. C. Light, Discrete variable representations and sudden
models in quantum scattering theory, Chem. Phys. Lett. 89, 483489 (1982); (b) J. C. Light, I.
P. Hamilton and J. V. Lill, Generalized discrete variable approximation in quantum mechanics,
J. Chem. Phys. 82, 14001409 (1985); (c) S. E. Choi and J. C. Light, Determination of
the bound and quasibound states of Ar???HCl van der Waals complex: Discrete variable
representation method, ibid. 92, 21292145 (1990).
[20] Brendan Vastenhouw and Rob H. Bisseling, A Two-Dimensional Data Distribution Method
for Parallel Sparse Matrix-Vector Multiplication, SIAM Review 47, 6795 (2005) .
[21] M. Founargiotakis, S. C. Farantos, G. Contopoulos and C. Polymilis, Periodic Orbits, Bifur-
cations and Quantum Mechanical Eigenfunctions and Spectra, J. Chem.Phys. 91, 13891402
(1989).
[22] R. Seydel, From Equilibrium to Chaos: Practical bifurcation and stability analysis, Elsevier
(1988).
[23] E. L. Allgower and K. Georg, Numerical Continuation Methods, Springer series in computa-
tional mathematics, 13, Berlin:Springer-Verlag (1993).
[24] Satish Balay, Kris Buschelman, William D. Gropp, Dinesh Kaushik, Matthew G. Knepley, Lois
Curfman McInnes, Barry F. Smith and Hong Zhang, PETSc, http://www.mcs.anl.gov/petsc
(2001).
[25] Satish Balay, Kris Buschelman, Victor Eijkhout, William D. Gropp, Dinesh Kaushik, Matthew
G. Knepley, Lois Curfman McInnes, Barry F. Smith and Hong Zhang, PETSc Users Manual,
ANL-95/11 - Revision 2.1.5, Argonne National Laboratory (2004).
[26] Satish Balay, Victor Eijkhout, William D. Gropp, Lois Curfman McInnes, Barry F. Smith,
Ecient Management of Parallelism in Object Oriented Numerical Software Libraries, in
Modern Software Tools in Scientic Computing, Editors: E. Arge, A. M. Bruaset and H. P.
Langtangen, 163202, Birkh auser Press (1997).
[27] H. Ishikawa, R. W. Field, S. C. Farantos, M. Joyeux, J. Koput, C. Beck and R. Schinke, HCP
- CPH Isomerization: Caught in the Act, Annual Review of Physical Chemistry 50, 443-484
(1999).
[28] M. Joyeux, S. C. Farantos and R. Schinke, Highly Excited Motion in Molecules: Saddle-Node
Bifurcations and their Fingerprints in Vibrational Spectra, J. Phys. Chem. 106, (feature
article) 54075421 (2002).
[29] M. Joyeux, S. Yu. Grebenshchikov, J. Bredenbeck, R. Schinke and S. C. Farantos, Intramolec-
ular Dynamics Along Isomerization and Dissociation Pathways, Geometrical Structures of
Phase Space in Multi-Dimensional Chaos, I. Prigogine, Stuart Rice (Eds): Advances in
Chemical Physics 130, 267303 John Wiley & Sons,(2005).