Algorithms for the Computation of Solutions and Bifurcations of the

OrnsteinZernike Equation

A. T. Peplow
Marcus Wallenberg Laboratory, Department of Aeronautics and
Vehicle Engineering, KTH, Stockholm, Sweden.; atpeplow@kth.se

R. E. Beardmore
Department of Mathematics, Imperial College, 180 Queens Gate,
South Kensington, London, SW7 2AZ; r.beardmore@ic.ac.uk

F. Bresme
Department of Chemistry, Imperial College, South Kensington, London, SW7 2AZ; f.bresme@ic.ac.uk
(Dated: February 8, 2006)
We perform computations to obtain the correlation functions of simple liquids by solving the the
Ornstein-Zernike (OZ) equation with the hyper-netted chain (HNC) closure. We are able to identify
precisely the values of temperature and density along the envelope of a so-called no-solution region
in parameter space that has been reported to exist in the literature.
This is done by first applying the pseudo arc-length continuation method to locate fold bifurcations
on vapour and liquid solution branches, we then solve the fold bifurcation equation as a system in
its own right in order to compute the aforementioned envelope.
Our analysis shows a rather complicated solution structure, with the existence of real-valued
solutions inside the no-solution region and the existence of hysteresis or multiple fold bifurcations
near the region of zero inverse isothermal compressibility.
As a final remark, our algorithms are always able to find solutions of zero inverse isothermal
compressibility on both liquid and vapour branches.

I. INTRODUCTION
FIG. 1: Illustration of the PAL algorithm applied to the one-
parameter problem f (u, ) = 0: a Newton correction step is
There is a rich history of computational studies of the conducted in a plane normal to the tangential direction of
OrnsteinZernike integral equation with its various clo- the curve C after an initial predictor step has been taken.
sures and it has been conjectured many times that the
system (defined below as equations (2-3)) fails to have
physically relevant solutions in certain parameter regions
of temperature and density. Several analyses appear to
have established that for the hypernetted chain equation
(HNC) closure, a so-called no-solution region exists that
is associated with the existence of multiple non-physical
solutions [15].
The purpose of this paper is to argue that the solu-
tion structure and bifurcation structure of the Ornstein-
Zernike integral equations, and hence the possible oc-
currence of phase transitions, can only be investigated
using pseudo arc-length continuation (PAL) methods [6],
as advocated in [7, 8]. As a result, we have undertaken
the following tasks:

1. implemented a one-parameter PAL continuation

code and used them to locate solutions that change
with respect to temperature and mean density.
This also allows us to find bifurcations of solutions
as a parameter is varied;
2. implemented two-parameter Newton-GMRES
code used to locate bifurcations that change with
respect to two parameters that have been found
using (1) above. This yields diagrams that track
the envelope of the no-solution region.
branches collide, one is a vapour branch with positive
compressibility, but the other solution branch has nega-
tive compressibility.
To explain this, consider the following simple mathe-
matical model for such behaviour given by the normal
form
1 + x(x 2 ) = 0, (1)

One study [4] argues that the no-solution region is
connected to the onset of complex solutions. Although
the direct physical relevance of this is not apparent, we
shall show that the no solution region occurs because of
a transcritical bifurcation that arises when two solution
where x is a measure of the total correlation function,
1 and 2 are physical parameters that play the role
of inverse compressibility and density respectively. This
model contains a transcritical bifurcation, or X-shaped
solution structure, when 1 = 0 and 2 is viewed as the

bifurcation parameter. This geometry breaks-up under
variations in 1 to give two separate solution branches
with fold bifurcations.
We conjecture that there is an isormorphism be-
tween the solutions of (1) and the discretisation of the
Ornstein-Zernike equations (2-3) below. Proving this
will certainly require the application of deep results from
bifurcation theory, but we are led to this conjecture by
our numerical computations. This connection has been
made before [5], but we should like to point out that
while we believe (1) to describe the fold bifurcations in
the discretisation of (2-3), describing the possible spin-
odals in the system (2-3) itself is an altogether different
matter.
Previous numerical studies that employed Picard or
Newton iteration methods to find solutions for fixed den-
sity and temperature would not have been able to locate
many of the solutions described in this paper. The rea-
son for this is that we provide a series of algorithms de-
veloped in Matlab that is available from the authors[15]
that use a PAL continuation method, moreover it is im-
plemented in such a way as to be independent of the
precise closure relation.
Moreover, despite the fact that [9] demonstrates that
fold bifurcations in any discretisation of equations (2-3)
may well represent spurious behaviour, we shall show
that discretisations of (2-3) possess many fold and tran-
scritical bifurcations using our algorithms.
We remark that the continuation method described in
[10] is termed natural parameter continuation in the nu-
merical analysis literature and what is proposed here is
quite different.
II. GOVERNING EQUATIONS
The isotropic Ornstein-Zernike (OZ) equation with
hyper-netted chain (HNC) closure is an infinite-domain
integral equation with an algebraic constraint, namely
Z transform are given in [7].
h(r) c(r) = h(kx yk)c(kyk)dy, (2)
R3
h(r) = 1 + eu(r) eh(r)c(r) . (3) Richardson extrapolation procedure to obtain O(N 4 )
Here and throughout, x, y R3 , r = kxk is the spatial
coordinate and k k is the Euclidean distance function,
is the mean particle density, = 1/(kB T ) is the Boltz-
mann factor, T being the temperature and u(r) is the
pair potential to be defined below. The total (h) and
direct (c) correlation functions yield the indirect corre-
lation function = hc and, using this definition, equa-
tions (2-3) can be re-written as a single integral in the
form
= (f +eu(r) exp1 ())(f +eu(r) exp1 ()), (4) ALGORITHM IMPLEMENTATION
where f (r) = 1 + eu(r) is the Mayer f-function,
denotes the convolution operator and the function exp1
is defined by exp1 (z) = ez 1.
When seeking solutions of (2-3), due to the presence
of a convolution operator, one generally requires h and c
to be absolutely integrable over (0, ) but it is shown in
[9] that this need not hold in the critical region and that
h may only be square integrable. This remark is crucial
2
to the development of the error analysis that is needed
when projecting (2-3) onto a set of finitely many basis
functions; see [9, 11].
In seeking approximate solutions of (2-3), the only el-
ement of the problem that requires discretisation is the
convolution operator. Moreover, since convolution can
be written using the Hankel transform as
a b = H1 (Ha Hb), (5)
where a, b : [0, ) R and H denotes the spherical
Hankel transform, it is standard practise to truncate the
interval of integration when solving (2-3) in the defini-
tion of the Hankel transform to a finite interval [0, R].
As explained in [7], this means that the Newton itera-
tion problem associated with the PAL algorithm for the
truncated version of (2-3) (equation (6) below) can be
solved using the generalised minimum residual algorithm
(GMRES). However, this truncation procedure does not
necessarily lead to good error estimates.
So, if
Z
(Ha)(s) = 4 sinc(rs)r2 a(r)dr
0
and
Z
1
(H1 a)(s) = sinc(rs)r2 a(r)dr,
2 2 0
and we then define HR a = H(R a) and HR a =
1
H (R a) where R is the characteristic function of the
interval [0, R], we arrive at the equation
= HR [HR (f +eu(r) exp1 ())HR (f +eu(r) exp1 ())].
(6)
Equation (6) is called the finite section problem with
cut-off value R associated with (4) and we solve this
system, that is (6), approximately, exploiting the Fast
Fourier Transform in our numerical calculations at every
opportunity. Again, the details of how to implement this
One significant difference with our discretisation and
the one used in [7] is the fact that we use one step of the
error estimates for solutions of (6), where N is the nu-
merical mesh size, rather than the O(N 2 ) estimates
that are given in [7]. The penalty for this is that while
second-order accuracy can be obtain by performing con-
volution with 3 FFT operations, our method requires 5
such operations, but this can be achieved without com-
promising the O(N log(N )) complexity of the convolu-
tion algorithm.
III. PROBLEM STRUCTURE AND
Under discretisation, equation (6) takes the very gen-
eral fixed-point form
= N (). (7)
In the following discussion we will write N (; ) when
we wish to emphasize the dependence of N on , but
shall suppress it otherwise.

Any smooth system of equations of the general form
(7), such as (4) and (6), is amenable to a Picard iteration
method for sufficiently sufficiently small values of >
0. Let us denote the derivative N () as a linear map
acting on v as given by the limit
N ()v = lim (N ( + v) N ())/.
0
When tracking solutions of (7) as changes there are
at least two possibilities to consider that we label (A1)
and (A2).
(A1) all eigenvalues of N () lie inside the open unit
disk in the complex plane D = {z C : |z| < 1},
It is well-known that one can apply Picard iteration
to (7) provided that (A1) holds at a solution and this
condition does hold for small densities. However, Picard
iteration will not generally yield a solution of (7) if (A1)
fails because a bifurcation of solutions is about to occur.
At this point we must consider the following condition.
(A2) the spectrum of N () decomposes as {},
where is an algebraically simple, real eigenvalue
of N (), D and 1.
One must account for the onset of the instability in-
herent in (A2) when solving (7) and the point of the
method described in [6] is that a new system is formed
that contains (7) as a subsystem. Moreover when apply-
ing a Newton method to this new system of equations
it should converge, even at bifurcation points. The next
section briefly describes how to do this.
A. Numerical Continuation
Rather than view (7) as a single equation defining for
fixed or , view it as an underdetermined equation that
jointly defines a solution curve or branch in (, )-space,
with fixed (one can reverse the roles of and in this
discussion if required). Now, suppose that (A1) ceases
to hold along a solution branch that we denote C. In
order to compute as much of C as possible, we adopt the
following well-known strategy due to Keller and others
that is now a common tool in bifurcation analysis [6, 12].
As C is a one-dimensional curve, we may parameterise
it by arclength. We now augment (4) with a constraint
and introduce an arc-length variable s in order to obtain
a well-posed system of equations and seek a solution with
solve and both as functions of s. The system we
therefore wish to solve, again with Newtons method, is
F (, , s) = (0, 0), where
F (, , s) ( N (), g(, , s)). (8) invertible due to the existence of a non-zero null vector
and g is an additional arclength constraint. In principle,
there are many possible choices for the function g but the
one used in practise is the linearised or pseudo arc-length
constraint given by
g(, , s) = s0 ( 0 ) + 0 ( 0 ) (s s0 ), sion of the derivative, however, to compute derivatives
where a dot () denotes the usual inner product of two
vectors, (s0 , 0 ) is the unit tangent vector at a given
point on the solution curve C and F (0 , 0 , s0 ) = 0.
3
The full algorithm thus has the following steps, where
ds assumes a fixed, small value:
(PAL1) Find (0 , 0 ) C using Picard iteration and
set s0 = 0.
(PAL2) Given (k , k ) C at arclength parameter
sk , let sk+1 = sk + ds and find the unit tangent vector
to C by solving the linear system
F (k , k , sk ) + F (k , k , sk ) = s F (k , k , sk )
for and and then define a direction vector (sk , k ) by
sk = (kk2 + 2 )1/2 , k = (kk2 + 2 )1/2 .
(PAL3) Apply Newton iteration to F (, , s) = 0 with
initial guesses = k + ds sk and = k + ds k and
denote the solution thus found (k+1 , k+1 ).
Steps (PAL2) and (PAL3) require a linear solver and
we use the restarted GMRES algorithm. As discussed in
[7], this is appropriate for integral equations like (6) that
are posed on a finite domain where the right-hand side
is a compact mapping. This is because the spectrum of
I N (), and hence the spectrum of the augmented
derivative (,) F (, , s), will cluster around unity in
the complex plane, although the clustering may well de-
grade as R is increased.
Note that in addition to being able to continue solu-
tions with respect to density, one can also hold density
at a fixed value and continue solutions with respect to
temperature, that is . We shall also provide some illus-
trative computations for this problem later in the paper
but the algorithm is almost identical in implementation.
B. Locus of fold bifurcations
Since (2-3) contains the parameters and , the loca-
tion of fold bifurcations themselves may be considered to
vary as a function of either or ; this is the so-called
codimension-two or two-parameter bifurcation problem
that provides an approximation to the phase-diagram
associated with (2-3) and provides the envelope of the
no-solution region. By applying a Newton method to a
discretisation of the following augmented problem
= N (; ), (9)
Z R
k = N (; )k, 1= k(r)dr, (10)
0
we shall be able to locate the locus of fold points param-
eterised by the variable .
Equations (9-10) ask that and satisfy (7), but also
that the linearisation of the problem I N () is not
that appears in (10), so that (I N ())k = 0. The
integral constraint in (10) is present only to ensure that
k is a nonzero vector.
In order to compute the derivative matrix or operator
that appears in (10) we have utilised an analytic ver-
of (10) itself, we have used numerical central difference
formulae.
Again, we may use GMRES as the linear solver for
the Newton iteration method applied to (9-10) because
 4

the part of this system that determines the null vec-
tor k, namely (10), involves a convolution operator that
leads to a linear system of equations with a clustered
set of eigenvalues. Moreover, because we were able to
locate for each candidate value bifurcation point for
in practice, we had no need to implement the PAL algo-
rithm on (9-10). One may have to resort to this slightly
more expensive approach if the two-parameter diagram
itself contains more complex bifurcations than the sim-
ple codimension-two, transcritical bifurcations [16] that
we present at the end of the paper.
According to convention [13], thermodynamic quan-
tities are expressed in reduced units: = 3 , T =
kB T / = 1/(). However, the collision diameter was
set to the value = 1 and the well-depth to = 1, so
that = and T = 1/ throughout.
FIG. 3: Solution branches of (6) continued in showing fold
bifurcations at low temperatures; density is fixed at =
0.1, 0.2, 0.3.

IV. NUMERICAL RESULTS

The purpose of this section is to illustrate the appli-

cability of the algorithms detailed above to locate many
solutions and solution branches of the Ornstein-Zernike
equation with given potentials; namely the Lennard-
Jones (LJ), Double Yukawa (DY) and Triple Yukawa
(TY) where the Yukawa potentials are chosen to fit the
Lennard-Jones potential, as illustrated in Fig. 2.

FIG. 2: Potentials (/), (a) Lennard-Jones, uLJ (solid), 1/T*

Double Yakawa, uDY (dashed), and Triple Yakawa uT Y
(dashed) and (b) relative error for Triple Yakawa and Double
Yakawa approximations against Lennard-Jones potential. Our first computation for a Lennard-Jones fluid is il-
lustrated in Figure 3 and is taken from [9], it shows the
(a) (b)
general bifurcation structure present in (6) as is var-
0.2
Triple Yakawa
LennardJones
2
10
Double Yakawa
Triple Yakawa
ied, for the low cut-off value of R = 10. The plot illus-
Double Yakawa trates a single fold bifurcation that occurs at sufficiently
0
3
10 low temperatures for all three density values that were
chosen. We now perform computations to examine this
4

0.2
10 bifurcation structure more closely.
Our second set of computations is illustrated in Fig-
0.4
1 1.5 2 2.5 ures 4, 5 and 6, where temperature is now fixed and we
/

r/
have used as the bifurcation parameter; we have also
0.6 shown the behaviour of the total correlation function at
several illustrative density values. The three figures il-
0.8 lustrate the search for a critical temperature Tc where
solutions have infinite compressibility. One can see from
1 the three figures that as temperature is reduced, the sin-
gle vapour branch present in Figure 4 increases in com-
pressibility to that shown in Figure 5 (the plots show
1 1.2 1.4 1.6
r/ a reduction in inverse compressibility) and collides with
the region where inverse compressibility is zero. The
temperature value used in Figure 5 is very close to the
A relative error tolerance of 1011 in the maximum critical value. As temperature is reduced further, Fig-
vector-norm was used for all Newton iterations that we ure 6 illustrates the creation of two disconnected solution
performed, including the two-parameter problem. branches; vapour at low densities and liquid at high den-
sities.
Although our next computation has a small numeri-
A. Lennard-Jones potential cal mesh with only N = 1025 points and also has an
artificially low cut-off value of R = 20 with T = 1.2, in-
In this section we use the Lennard-Jones potential teresting conclusions may still be drawn. In Figure 7 the
solution measure on the y-axis is the inverse (reciprocal)
uLJ (r) = 4((/r)12 (/r)6 ), (11) of isothermal compressibility, given by
Z
with a numerical mesh size of N = 212 +1 = 4097 points,
1 (Hc)(0) or 1 4 r2 c(r)dr (12)
unless otherwise stated. We plot several bifurcation di- 0
agrams; plots with a parameter such as or on the
x-axis and a solution measure, such as inverse isother- that we denote T 1 . The correlation function r2 h(r)
mal compressibility, on the y-axis. at a point on the vapour branch before the first fold is
 5

FIG. 4: Inverse isothermal compressibility plotted against FIG. 6: Inverse isothermal compressibility plotted against
density at temperature T = 1.6, (T > Tc ), for the Lennard- density at temperature T = 1.3, (T < Tc ), for the Lennard-
Jones potential, where R = 80 and N = 212 + 1. Three Jones potential, R = 80 and N = 212 + 1. Three correla-
correlation functions, r2 h(r), are presented on each curve at tion functions, r2 h(r), on each curve are presented at various
density values: (a) = 0.1, (b) = 0.270 and (c) = 0.55. density values : (a) = 0.1, (b) NO SOLUTION and (c)
= 0.55.
(a)
3
1 (a)
1 1.5
0.5
h(r) r2 1
0

h(r) r2
2.5 0.5
0.8
0.5 0
0.5
2 (b) 0.6 1 0 1
1 10 10 10
r2 c(r) dr

r2 c(r) dr
0.5
h(r) r2

0.4
1.5 0
14T
0

14T
0
0.5
0.2
1
(c)
(c)
1 0
0.5 1
h(r) r2

0.5

h(r) r2
0 0.5
0.2
0.5 0
0 0.5
0 0.2 0.4 0.6 1 0 1
10 10 10 0.4
r 0.1 0.2 0.3 0.4 0.5 0.6 1 0 1
10 10 10
r

FIG. 5: Inverse isothermal compressibility plotted against

density at temperature T = 1.408, (T veery close to Tc ), FIG. 7: Inverse isothermal compressibility plotted against
for the Lennard-Jones potential, R = 80 and N = 212 + density at temperature T = 1.2 for the Lennard-Jones po-
1. Three correlation functions, r2 h(r), on each curve are tential, where R = 20, N = 210 + 1. Three solutions,
presented at various density values: (a) = 0.1, (b) = r2 h(r), on each curve are presented at various points: (a)
0.270 and (c) = 0.55. = = 0.084, (b) = = 0.419 and (c) = = 0.509.
Solutions for (a) and (c) lie on the upper part of the vapour
and liquid branches respectively.
(a)
2.5
1 (a)
h(r) r2

0.5 0.8
1.5
0 1
2
h(r) r2

0.5 0.6 0.5

0
(b) 0.4 0.5
r2 c(r) dr

1 1
1.5 1 10 10
(b)
0.2
h(r) r2

0.5
r2 c(r) dr

1.5
14T
0

0 1
h(r) r2

0
1 0.5 0.5
14T
0

0
(c) 0.2 0.5
1 1
1 10 (c) 10
0.5 0.4
h(r) r2

0.5
1
0
h(r) r2

0.6 0.5
0.5
0
0
0 0.2 0.4 0.6 1 0 1
10 10 10 0.8 0.5

r
0.1 0.2 0.3 0.4 0.5 1 0 1
10 10 10
r

indicated in Figure 7(a) and the correlation functions

that were computed on other solution branches are also
shown.
The vapour solution branch emanates from the solu-
tion pair (, ) = (0, 0) with unit compressibility and this
branch extends to higher densities, but it then contin-
ues into regions of negative compressibility. It had been
believed previously that T 1 remained finite when ap-
proaching the so-called no-solution region [1, 10] indicat-
ing that the no-solution region is not linked to a phase
transition. However, we see that inverse compressibil-
ity becomes zero at a point somewhere beyond the first
fold bifurcation and so our initial results appear to dis-
agree in part with this; we are led to the conjecture that
compressibility does diverge at some density, provided
temperature is sufficiently low.
In Figure 7 there is also a solution branch that is ob-
tained for higher densities, a liquid branch. The correla-
tion function r2 h(r) on the upperpart of the branch is
 6

shown in (c) with compressibility is positive. Shown in

FIG. 9: Locus of fold bifurcations (a) and inverse compress-
(b) is a correlation function r2 h(r) that can be computed
ibility values (b) for the Lennard-Jones potential for varied
inside the no-solution region and the solution branch is cut-off values, R, and N = 212 + 1 points. Legend: R = 80
indicated in the centre of the main plot. Other branches (solid), R = 40 (dash-dot), and R = 20 (dashed line).
of solutions in this region have been found before, but
they were complex-valued solutions of (4) that are not
physically relevant [3]. The solutions that we have lo-
1.5

cated in this parameter regime are real solutions, but 1.4

because the inverse compressibility is negative, nor are

ours physically relevant, moreover we have not been able 1.3
R = 80

to locate solutions in this region of positive compressibil- 1.2

ity. That is not to say that none are there, but we have R = 40

T*
no systematic method for locating solutions in this re- 1.1

gion.
1
Figure 8 depicts solutions that are obtained using the R = 20

higher number of mesh points, N = 4097, and a larger 0.9

value for the cut-off parameter, R = 120; this plot mim-

ics the left-hand branch in Figure 7. The right-hand 0.8
0 0.1 0.2 0.3 0.4 0.5 0.6
*
plots show clearly the significant changes in solution
structure beyond the first fold bifurcation and that the 0.2

increase in accuracy has resolved a number of bifurca-

0.15
tions that were missed when using too coarse a mesh,
leading to S-shaped solution branches. 0.1
14 r2c(r)dr

FIG. 8: Inverse isothermal compressibility plotted against 0.05

R = 80
density with T = 1.2 for Lennard-Jones potential, R = 120, R = 40
N = 212 + 1. Two solutions, r2 h(r), on each curve are pre- 0

sented at various points on three curves : (a) = = 0.1127, R = 20

(b) = = 0.1122. Solution (a) located at upper circle in 0.05

left-hand plot, and (b) located at lower circle.

0.1
0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
(a) *

1.5
0.8
1
h(r) r2

0.6 0.5
negative value as density increases; this phenomenon is
0
0.4
examined more closely in [9].
0.5
The correlation functions r2 h(r) along the two-
r2 c(r) dr

0 2
0.2
o 10
r
10
parameter parabolic curve from Figure 9 are presented in
14T

Figure 10: (A) represents solutions in the vapour region,

0

(b)
o
0
1.5
(B) is at the critical point or apex from Figure 9 and (C)
1
represents a liquid state. It is clear in Figure 10(C) that
0.2
solutions are particulary sensitive to the cut-off value R.
h(r) r2

0.5

0.4 0

0.5

0.02 0.04 0.06 0.08 0.1 0.12 0

10
2
10 B. Double Yakawa and Triple Yakawa Potentials
r

In this section we consider the Double Yakawa po-

Figure 9 shows a computed codimension-two bifurca- tential (DY) and Triple Yakawa potentials from [14],
tion diagram, as described in section III B, that bears respectively (13) and (14) below. The parameters in
more than a passing resemblance to a phase diagram. (13) are A1 = 1.6438 z1 = 14.7 1 , A2 = 2.03, and
Density is situated on the horizontal and T on the z2 = 2.69 1 are chosen to fit the Lennard-Jones po-
vertical axis and various cut-off values have been used to tential (11) and match the second virial coefficient to
verify the consistency of our results. that of the Lennard-Jones fluid, although we recall that
The pinnacle of the curve represents a codimension- = 1. Similarly, the parameters for the Triple Yakawa
two bifurcation, namely a fold bifurcation point of the potential take the values c1 = 2.351, z3 = 13.446 1 ,
system (9-10) that represents a transcritical bifurcation c2 = 0.910, z4 = 3.482 1 and z5 = 1.317 1 in (14).
of (7), where is the bifurcation parameter. It is im-
In contrast to the LennardJones potential which has
portant to note here that the form of two-parameter
algebraic decay, this section focusses on the Double
bifurcation curve is quite similar for all cut-off values,
Yakawa and Triple Yakawa potentials that decay expo-
R = 20, 40, 80, that we used but the lower cut-off R = 20
nentially and are given as follows:
results in unphysical solutions since inverse compressibil-
ity is negative in the liquid region. For the higher cut-
off R = 40, inverse compressibility also tends towards uDY (r) = A1 ez1 (r) A2 ez2 (r) , (13)
r
 7

c1 z3 (r)
2 uT Y (r) = e c2 ez4 (r) (1 c2 )ez5 (r) (14)
.
FIG. 10: Distribution functions r h(r) at locations along the r
bifurcation locus, Fig. 9 for various cut-off values, R = 20, 40,
and 80 and N = 212 + 1 points, for Lennard-Jones poten-
tial. ( , T )= (0.122,1.23) (A), (0.275,1.407) (B), and (C)
(0.503,1.170). Note that (B) represents solutions at the peak As previously, we have computed the locus of fold bi-
of the locus of fold bifurcations in Fig. 9 furcations for various values of the cut-off parameter R
and the results are shown in Figure 11. It is clear from
R =20
Figure 11(i) and (ii) that the nature of the curves are
1.5
R =40
R =80
similar for small cut-offs, namely when R = 20, illus-
trating that the decay of the potential does not resolve
1
the issue of numerical degradation of solutions in the
vapour or large density region noted in Figure 10(C).
Increasing the cut-off value to R = 80 does resolve this
r2 h(r)

0.5

as in the Lennard-Jones case and as shown in Fig.11(iii)

0 and (iv).
0.5 In Figure 12 we illustrate the correlation function
r2 h(r) computed along the two-parameter bifurcation
1
10
1
10
0
10
1
10
2 curve from Figure 11. The solutions that are computed
(A) r
using DY and TY potentials are qualitatively similar in
1.5
R =20
appearance to those obtained using the Lennard-Jones
R =40
R =80
potential (compare Figure 10 with Figure 12).
1

0.5
r2 h(r)

V. CONCLUSIONS
0.5

1
10
1
10
0 1
10 10
2 We have shown that by employing pseudo arc-length
(B) r
continuation for the one-parameter formulation, and
Newton-GMRES applied to the two-parameter bifurca-
1.5
R = 20 tion problem of the Ornstein-Zernike equation, one can
compute solutions and bifurcations that are difficult or
R = 40
R = 80

1 impossible to find using direct Newton and Picard meth-

ods.
0.5
Moreover, we are always able to compute solutions of
r2 h(r)

zero inverse compressibility on both liquid and vapour

0 branches for potentials with an attractive tail, provided
that temperature is sufficiently low. As noted in [3], the
0.5 vapour solution branch possesses an S-shaped solution
branch at low temperatures with as many as three so-
10
1 0
10 10
1
10
2
lutions for a given density, each with positive compress-
(C) r ibility provided that the cut-off parameter R used in the
discretisation of (6) is sufficiently large.

[1] R. O. Watts, Hypernetted-chain approximation ap-
plied to a modified lennard-jones fluid, J. Chem. Phys.
50 (1358).
[2] J. Brey, A. Santos, Absence of criticality in the hypernet-
ted chain equation for a truncated potential, Mol. Phys.
57 (1986) 149160.
[3] L. Belloni, Inability of the hypernetted chain integral
equation to exhibit a spinodal line, J. Chem. Phys.
98 (10).
[4] E. Lomba, J. Lopez-Martin, On the solutions of the
hypernetted-chain equation inside the gas-liquid coex-
istence region, J. Stat. Phys. 80 (825).
[5] P. D. Poll, N. W. Ashcroft, Critical behaviour of the
hypernetted chain equation, Phys. Rev. a 35 (12) (1987)
5167 5173.
[6] H. B. Keller, Numerical Solution of Bifurcation and Non-
linear Eigenvalue Problems, Academic Press, New York,
ed. P. Rabinowitz, 1977.
[7] C. T. Kelley, B. Mongomery-Pettitt, A fast solver for the
Ornstein - Zernike equations, J. Comp. Phys. 197 (2004)
491 501.
[8] L. J. D. Frink, A. G. Salinger, Rapid analysis of phase
behavior with density functional theory. II. capillary con-
densation in disordered porous media, J. of Chemical
Physics 118 (16) (2003) 7466 7476.
[9] R. E. Beardmore, A. T. Peplow, F. Bresme, A numerical

bifurcation analysis of the ornstein-zernike equation with
hypernetted chain closure, submitted to SIAM J. Sci.
Comput. (January 2006). [13]
[10] D. A. Tikhonov, G. N. Sarkisov, Singularities of solu-
tion of the ornstein-zernike equation within a gas-liquid [14]
transition region, Russian J. Phys. Chem. 74 (3).
[11] J. R. Henderson, Z. A. Sabeur, Liquid-state integral
equations at high density: On the mathematical origin of [15]
infinite-range oscillatory solutions, J. Chem. Phys. 97 (9)
(1992) 67506758. [16]
[12] E. J. Doedel, R. Paffenroth, A. Champneys, T. Fair-
grieve, Y. Kuznetsov, B. Sandstede, X. Wang, Auto
2000: Continuation and bifurcation software for ordinary
differential equations (with homcont), technical Report,
8
Caltech; see also http://indy.cs.concordia.ca/auto/
(2001).
J. Hansen, I. Macdonald, Theory of Simple Liquids, Aca-
demic Press, London, 1976.
E. Scholl-Paschinger, Phase behaviour of simple fluids
and their mixtures, Ph.D. thesis, Intitut fur Theoretische
Physik, Vienna (February 2002).
Send an email to REB or ATP to request a copy of the
code.
A transcritical bifurcation occurs when two solution
branches collide when varying a second parameter; for
example, this bifurcation occurs at the apex of Figure
9(left).
 9

FIG. 11: (i)Locus of fold bifurcations for the (A) Lennard-

Jones (B) Double Yakawa and (C) Triple Yakawa potentials
for cut-off value, R = 20, and N = 212 + 1 points. (ii) Inverse
compressibility values along locus of fold bifurcations for LJ,
DY and TY potentials. (iii) and (iv) as legend for (i) and (ii)
with higher cut-off value R = 80.

0
1.5
B
C
A
1.4

1.3

1.2
T*

1.1

0.9

0.8
0 0.1 0.2 0.3 0.4 0.5 0.6
(i) *

0.3
B
C
0.25 A

0.2

0.15
14 r2c(r)dr

0.1

0.05

0.05

0.1
0.1 0.2 0.3 0.4 0.5 0.6
(ii) *

1.45
A
B
1.4 C

1.35

1.3

1.25
T*

1.2

1.15

1.1

1.05

1
0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
(iii) *

A
B
0.2 C

0.15
14 r2c(r)dr

0.1

0.05

0
0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
(iv) *
 10

FIG. 12: Correlation functions (i) r2 h(r) (ii) r2 c(r) at various

locations along the two-parameter curve for cut-off value R =
80 and N = 212 +1 points for the DY (i and ii) and TY (iii and
iv) potentials: ( , T )= (0.122,1.23) (A), (0.275,1.407) (B),
and (C) (0.480,1.210). Note that (B) represents solutions at
the apex of the curve of fold bifurcations shown in Figure 11
(i) and (iii).

A
B
1.5 C

1
r2 h(r)

0.5

0.5

1 0 1 2
10 10 10 10
(i) r

2
r2 c(r)

4
A
5 B
C

1 0 1
10 10 10
(ii) r

A
B
1.5 C

1
r2 h(r)

0.5

0.5

1 0 1 2
10 10 10 10
(iii) r

2
r2 c(r)

A
5 B
C
6

1 0 1
10 10 10
(iv) r