JOURNAL OF APPLIED PHYSICS VOLUME 95, NUMBER 7 1 APRIL 2004

Electrical capacitance of the unit cube

Chi-Ok Hwanga)
Computational Electronics Center, Inha University, 253 Yonghyun-Dong Nam-Gu, Incheon 402-751, Korea
Michael Mascagni
Department of Computer Science and School of Computational Science and Information Technology,
Florida State University, Tallahassee, Florida 32306
共Received 8 October 2003; accepted 13 January 2004兲
It is well known that there is no analytical expression for the electrical capacitance of a cube, even
though it has been claimed that one can compute this capacitance numerically to high precision.
However, there have been some disparities between reported numerical results of the capacitance of
the unit cube. In this article, the ‘‘walk on planes’’ 共WOP兲 algorithm 关M. L. Mansfield, J. F. Douglas,
and E. J. Garboczi, Phys. Rev. E 64, 061401 共2001兲兴 is used to compute the capacitance of the unit
cube. With WOP, we remove the error from the ⑀-absorption layer commonly used in ‘‘walk on
spheres’’ computations so that there is no inherent error introduced in these WOP computations
except the intrinsic Monte Carlo sampling error of size O(N 1/2). This WOP technique comes from
the isomorphism, provided by probabilistic potential theory, between the electrostatic Dirichlet
problem on a conducting surface, and the corresponding Brownian motion first-passage expectation.
The numerical result we obtain with WOP, 0.660 678 2⫾1⫻10⫺7 , supports the deterministic
calculations by Read 关F. H. Read, J. Comp. Phys. 133, 1 共1997兲兴, the results by Given et al. 关J. A.
Given, J. B. Hubbard, and J. F. Douglas, J. Chem. Phys. 106, 3721 共1997兲兴 and our previous results
关C.-O. Hwang and M. Mascagni, J. Korean Phys. Soc. 42, L1 共2003兲; M. Mascagni and N. A.
Simonov, J. Comp. Phys. 共in press兲兴. Also, it is noted that the exact value conjectured by Hubbard
and Douglas is inconsistent with our calculations, and that some other calculations fall outside our
error bounds. © 2004 American Institute of Physics. 关DOI: 10.1063/1.1664031兴

I. INTRODUCTION Read,7 who extrapolated the number of subdivisions to infin-

It is well known that there is no analytical expression for
the electrical capacitance of a cube, even though it has been
claimed that one can compute this capacitance numerically
to high precision.1–3 However, as shown in Table I, there
have been some reported disparities between numerical re-
sults of the capacitance of the unit cube. At present, there is
an exact value conjectured by Hubbard and Douglas,4 and
several numerical results for the capacitance.1–3,5–11 The nu-
merical methods used to obtain these results can be catego-
rized into two groups; deterministic and probabilistic 共Monte
Carlo兲 methods. Deterministic methods include surface
charge2,9,11 or boundary element methods 共by Cochran10兲,7,12
and finite-difference methods.5,13 Among the Monte Carlo
methods, there are Brownian dynamics algorithms,3,14 a first-
passage 共FP兲 algorithm,15 a refined Brownian dynamics
algorithm,1 an extrapolation technique16 coupled to the
‘‘walk on spheres’’ 共WOS兲 method,16 and random walk on
the boundary method.8
The surface charge method2,11 divides the surface area of
the unit cube into small subareas. The charge density on each
subarea is then assumed to be constant, and the constant total
electrical potential enables us to obtain a set of linear equa-
tions with unknown charges on the subareas. Solving the set
of linear equations, we can compute the total charge and,
hence, the capacitance. This method was refined later by
a兲
Electronic mail: chwang@hsel.inha.ac.kr
ity. In the finite-difference method,13 the exterior Dirichlet
problem is transformed into an interior Dirichlet problem
containing the origin with the usual finite-difference scheme.
Later, Given et al.15 introduced a FP Monte Carlo algo-
rithm, and as an example numerically computed the capaci-
tance of the unit cube. In the FP Monte Carlo algorithm, they
refined the previous Brownian dynamics algorithm10 to re-
move the ⑀-absorption layer by using a set of Green’s func-
tions to provide exact FP probability distributions to termi-
nate the walks 共see Fig. 1兲. WOS methods employ an
⑀-absorption layer to terminate the random walks.3,17–22 The
boundary, on which random walks terminate, is thickened by
⑀, and when a walker enters this ⑀-absorption layer, the walk
is terminated by choosing the boundary point closest to the
point, as shown in Fig. 2.
Recently, using a simple error-controllable method
coupled to a refined Brownian dynamics algorithm,1 we sup-
ported the calculations by Given et al.,15 by Greenspan
et al.,13 and by Read.7 In the method, the error from the
⑀-absorption layer, one of the two error sources, was con-
trolled by exploiting the linearity of that error so that it is
made much smaller than the statistical sampling error. Also,
it was noted that the exact value conjectured by Hubbard and
Douglas is inconsistent with our calculations, and that some
other calculations fall outside these bounds even though their
claimed error is much smaller. More recently, a bias-free
Monte Carlo method, the ‘‘random walk on the boundary’’

0021-8979/2004/95(7)/3798/5/$22.00 3798 © 2004 American Institute of Physics

Downloaded 26 Mar 2004 to 165.246.34.43. Redistribution subject to AIP license or copyright, see http://jap.aip.org/jap/copyright.jsp
J. Appl. Phys., Vol. 95, No. 7, 1 April 2004 C. Hwang and M. Mascagni 3799

TABLE I. Capacitance of the unit cube in units of 4 ␲ ⑀ 0 .

Reitan–Higginsa Surface charge method 0.6555

Greenspan–Silvermanb Finite difference method 0.661
Cochranc Boundary element method 0.6596
Goto–Shi–Yoshidad Surface charge method 0.660 674 7⫾5⫻10⫺7
Conjectured Hubbard–Douglase Analytic conjecture 0.65946...
Douglas–Zhou–Hubbardf Brownian dynamics algorithm 0.6632⫾0.0003
Given–Hubbard–Douglasg Refined Brownian dynamics algorithm 0.660 675⫾0.000 01
Readh Refined boundary element method 0.660 678 5⫾6⫻10⫺7
Mansfield–Douglas–Garboczii Brownian dynamics algorithm 0.660 69共2兲
Hwang–Mascagnij Modified Brownian dynamics algorithm 0.660 683⫾0.000 005
Bai–Lonngrenk Surface charge method 0.6601
Mascagni-Simonovl Random walk on the boundary 0.660 678 0⫾2.7⫻10⫺7
Our result Walk on planes 0.660 678 2⫾1⫻10⫺7
a g
See Ref. 2. ee Ref. 15.
b
See C. S. Brown 共1990兲 共Ref. 5兲 and Ref. 13. h
See Ref. 7.
c i
See Ref. 10. See Ref. 6.
d j
See Ref. 9. See Ref. 1.
e k
See Ref. 4. See Ref. 11.
f l
See Ref. 10. See Ref. 8.

method was introduced for this problem, and was claimed to
obtain the most accurate estimate to date, 0.660 678 0⫾2.7
⫻10⫺7 .
It is commonly thought that Monte Carlo methods are
most efficient when dealing with complicated geometries, in
very high dimensional calculations, and for problems involv-
ing singularities, for example objects that include folds,
cusps, and corners in electrostatic problems. Problems in-
volving singularities require deterministic methods that are
have either highly refined geometries, or use very high-order
methods. This makes us believe that in computing the ca-
pacitance of the unit cube, Monte Carlo methods are the
efficient algorithms per unit accuracy. Especially, Monte
Carlo methods which do not have bias in the estimate such as
the FP Monte Carlo algorithm, and the random walk on the
boundary method, which are also the most efficient Monte
Carlo algorithms to date. However, it should be noted that
the FP Monte Carlo algorithm uses nontrivial Green’s func-
tions which implies time-consuming function evaluation that
often relies on tabulation and interpolation for efficiency.

FIG. 1. Schematic side view that illustrates an absorbed series of FP jumps
using the FP Monte Carlo algorithm. In this illustration, the ␦-boundary
layer usage is shown: When a Brownian walker is initiated with uniform
probability on the launching sphere, L of radius b, and reaches inside the
␦-boundary layer, it begins to intersect the cube.
FIG. 2. Schematic side view that illustrates an absorbed series of FP jumps
using the modified Brownian dynamics algorithm. In WOS, the boundary is
thickened by ⑀, and when a Brownian walker is initiated with uniform prob-
ability on the launching sphere, L of radius b, and enters this ⑀-absorption
layer, the walk is terminated.

Downloaded 26 Mar 2004 to 165.246.34.43. Redistribution subject to AIP license or copyright, see http://jap.aip.org/jap/copyright.jsp
3800 J. Appl. Phys., Vol. 95, No. 7, 1 April 2004 C. Hwang and M. Mascagni

In this article, we introduce another bias-free Monte

Carlo method for updating the electrical capacitance of the
unit cube. This article is organized as follows. In Secs. 2 and
3, we explain the ‘‘walk on planes’’ 共WOP兲6 Monte Carlo
algorithm and refined Brownian dynamics algorithm with
WOP, respectively, for computing the electrical capacitance
of the unit cube. Section 4 provides our numerical results,
discussion, and conclusions.

II. THE WOP ALGORITHM

In this section, using the isomorphism, provided by
probabilistic potential theory,23,24 between the electrostatic
Poisson problem, and the corresponding Brownian motion
expectation of the first passage, we introduce the WOP6
Monte Carlo method for calculating the electrical capaci-
tance of the unit cube. 共In favor of the cube geometry, we
guess that another bias-free ‘‘walk on rectangles’’
method25,26 can be used for the capacitance of the unit cube
also. However, we should note that the surface Green’s func-
tion for an rectangle is much more complicated than that for
an infinite plane.兲
The isomorphism between the electrostatic induced
charge problem on a conducting surface and the correspond-
ing Brownian motion expectation gives us the FP probability
distribution of a Brownian walker at distance d from an in-
finite plane making first-contact on the plane. The surface
induced charge density on a conducting plane of infinite ex-
tent when a unit point charge at a distance d from the plane
is given as follows:27
d repeated until we either escape to infinity or terminate on the
␴ 共 r 兲 ⫽⫺ . 共1兲
2 ␲ 共 d ⫹r 2 兲 3/2
2
Here, r is the radial distance from the intersection point
where the distant line and the plane meet 共see Fig. 3兲. Then,
the cumulative density function, as a function of r, is given
as:
冕 冕
2␲
FIG. 3. Schematic side view that illustrates a FP jump using the WOP
algorithm. In WOP, there is no ⑀ layer. A Brownian walk is initiated with
uniform probability on the launching sphere, L of radius b, and proceeds by
either striking an infinite plane which includes one of the six surfaces of the
unit cube (a⫽1) and is declared to be absorbed landing on the surface of
the unit cube, or falls on the plane outside of the cube surface, in which case
the walk continues.
ability, ␤, of a Brownian FP trajectory started on a ‘‘launch-
ing sphere’’ and making first-passage on the object in
question3,6,10 共see Fig. 4兲:
C⫽b ␤ . 共4兲
Here, the launching sphere of radius b should completely
enclose the given object. In this refined Brownian dynamics
algorithm, we constantly endeavor to take WOP FP jumps
when the Brownian walker is placed inside or on the launch-
ing sphere, or use the probability of going to infinity28 with-
out going back to the launching sphere when the walker is
placed outside the launching sphere. These sorts of steps are
boundary of the given object.
We calculate ␤ by performing simulations for the capaci-
tance of the unit cube as follows. A Brownian walker is
initially placed randomly on the surface of the launching
sphere. After that, each Brownian walk is constructed as a
series of first-passage jumps from the present position of the

0 0
r
⫺
rd
2 ␲ 共 d 2 ⫹r 2 兲 3/2
dr⫽
1
共 1⫹ 共 r/d 兲 2 兲 1/2
. 共2兲
Brownian walker to a new position on the first-passage plane
which includes one of the six faces of the unit cube 共see Fig.
Thus, we can sample the radial FP point at a distance r when
a Brownian walker starts at the distance d from the plane by
using the following expression:
r
d
⫽ 冑 共 1⫺U 2 兲
U2
U苸 关 0,1兲 , 共3兲

where U is a uniform random number.

III. REFINED BROWNIAN DYNAMICS ALGORITHM

WITH WOP

In this section, for calculating the electrical capacitance

of the unit cube we explain a refined Brownian dynamics
algorithm with WOP.
The capacitance, C, of an arbitrarily shaped conducting
object 共actually the relative capacitance with respect to a
launching sphere兲 can be obtained by computing the prob-
FIG. 4. Schematic view that illustrates a WOP FP jump. In this WOP, the
radial FP jump location r is sampled according to r/d⫽ 冑(1⫺U 2 )/U 2 ,
where U is in 共0, 1兲 and the angle ␾ is in 共0, 2␲兲.

Downloaded 26 Mar 2004 to 165.246.34.43. Redistribution subject to AIP license or copyright, see http://jap.aip.org/jap/copyright.jsp
J. Appl. Phys., Vol. 95, No. 7, 1 April 2004 C. Hwang and M. Mascagni 3801

TABLE II. Capacitances of the unit cube in units of 4 ␲ ⑀ 0 using WOP.

Random number generatora No. of trajectories Capacitance

12
lfg 10 0.660 677 980
lfg 1012 0.660 677 690
lcg 1012 0.660 678 271
lcg 1012 0.660 678 294
lcg 1012 0.660 678 515
mlfg 1012 0.660 678 547
mlfg 1012 0.660 678 066
cmrg 1012 0.660 678 383
cmrg 1012 0.660 678 345
cmrg 1012 0.660 677 704
lcg64 1012 0.660 678 062
lcg64 1012 0.660 677 404
lcg 1013 0.660 678 325
lcg 1013 0.660 678 042
a
See Ref. 29.
FIG. 5. This illustration shows the usage of the replacement distribution
density function ␻共␪,␾兲 when the WOP jump from the launching sphere
ends outside the launching sphere. With the probability, (1⫺ P inf), the infinite plane of a Brownian walker at a distance from the
walker goes back to the launching sphere. plane was given by the isomorphism between the electro-
static induced charge problem on a conducting surface and
the corresponding Brownian motion expectation.
4兲. If the Brownian walker is placed inside the launching
In WOP, there is no ⑀ layer. When a Brownian walker is
sphere but not on one of the six faces of the unit cube,
initiated with uniform probability on the launching sphere, L
another first-passage plane is constructed, and the Brownian
of radius b, it lands on an infinite plane which includes one
walk continues. When the Brownian walker lands outside the
of the six surfaces of the unit cube. It is declared absorbed if
launching sphere, it goes to infinity28 without going back to
it lands on one of the six surfaces of the unit cube. If not
the launching sphere with probability
absorbed, the process continues. If the walker escapes the
b original launching sphere, then, with a known probability it
P inf⫽1⫺ . 共5兲 either goes to infinity and is terminated or returns to the
r
launching sphere.
Here, b is the launching sphere radius and r the distance of
The averaged result from Table II, 0.660 678 2⫾1
the Brownian walker from the center of the launching
⫻10⫺7 , with the total of 3.2⫻1013 trajectories supports the
sphere. If it is determined that the walker goes back to the
results by Read,7 the results by Given et al.,15 and our pre-
launching sphere, we use the replacement distribution
vious results.1,8 The sampling fractional error corresponding
function28
to one standard deviation is given by15
1⫺ ␣ 2
␻共 ␪,␾ 兲⫽
4 ␲ 共 1⫺2 ␣ cos ␪ ⫹ ␣ 2 兲 3/2
共6兲
冑 1⫺ p 1
p 冑N
, 共8兲
to place it on the launching sphere. Here the coordinates 共␪,
␾兲 are defined with respect to the polar axis that joins the where p is the probability of contacting the target object of a
position of the Brownian walker to the launching sphere cen- single random walker, and N is the number of random walks.
ter, and ␣ is b/r 共see Fig. 5兲. This procedure is repeated until These computations are performed on a 16-node PC cluster
the Brownian walker either reaches the unit cube or goes to using message passing interface and scalable parallel random
infinity. After doing the Brownian walk N times, we obtain number generators.29 This computation is yet illustration of
an estimate of the probability, ␤: the simplicity, beauty, and power of Monte Carlo methods in
the problems involving singularities, for example, folds,
Na cusps, and corners in electrostatic problems.
␤⬇ , 共7兲
N
where N a is the number of times the Brownian trajectory ACKNOWLEDGMENT
terminates on the unit cube. The authors wish to acknowledge the financial support
of the Computational Electronics Center, Inha University, In-
IV. NUMERICAL RESULTS, DISCUSSION, cheon, Korea, for this work.
AND CONCLUSIONS
In this article, we introduced another bias-free Monte 1
C.-O. Hwang and M. Mascagni, J. Korean Phys. Soc. 42, L1 共2003兲.
Carlo method for calculating the electrical capacitance of the
2
D. K. Reitan and T. J. Higgins, J. Appl. Phys. 22, 223 共1951兲.
3
H.-X. Zhou, A. Szabo, J. F. Douglas, and J. B. Hubbard, J. Chem. Phys.
unit cube. This method utilized the geometrical properties of 100, 3821 共1994兲.
the unit cube to remove the ⑀-absorption layer in WOS by 4
J. B. Hubbard and J. F. Douglas, Phys. Rev. E 47, R2983 共1993兲.
using WOP. The first-passage probability distribution on the 5
C. S. Brown, Comput. Math. Appl. 20, 43 共1990兲.

Downloaded 26 Mar 2004 to 165.246.34.43. Redistribution subject to AIP license or copyright, see http://jap.aip.org/jap/copyright.jsp
3802 J. Appl. Phys., Vol. 95, No. 7, 1 April 2004 C. Hwang and M. Mascagni

6
M. L. Mansfield, J. F. Douglas, and E. J. Garboczi, Phys. Rev. E 64, 18
A. Haji-Sheikh and E. M. Sparrow, J. Heat Transfer 89, 121 共1967兲.
061401 共2001兲. 19
L. H. Zheng and Y. C. Chiew, J. Chem. Phys. 90, 322 共1989兲.
7
F. H. Read, J. Comput. Phys. 133, 1 共1997兲. 20
S. Torquato and I. C. Kim, Appl. Phys. Lett. 55, 1847 共1989兲.
8
M. Mascagni and N. A. Simonov, J. Comput. Phys. 共in press兲. 21
I. C. Kim and S. Torquato, J. Appl. Phys. 71, 2727 共1992兲.
9
E. Goto, Y. Shi, and N. Yoshida, J. Comput. Phys. 100, 105 共1992兲. 22
T. E. Booth, J. Comput. Phys. 39, 396 共1981兲.
10 23
J. F. Douglas, H.-X. Zhou, and J. B. Hubbard, Phys. Rev. E 49, 5319 M. Freidlin, Functional Integration and Partial Differential Equations
共1994兲. 共Princeton University Press, Princeton, NJ, 1985兲.
11
E.-W. Bai and K. E. Lonngren, Comput. Electr. Eng. 28, 317 共2002兲. 24
K. L. Chung and Z. Zhao, From Brownian Motion to Schrödinger’s Equa-
12
H. Ymeri, B. Nauwelaers, and K. Maex, J. Korean Phys. Soc. 40, 277 tion 共Springer, Berlin, 1995兲.
共2002兲. 25
T. E. Booth, J. Comput. Phys. 47, 281 共1982兲.
13
D. Greenspan and E. Silverman, Proc. IEEE 53, 1636 共1965兲. 26
S. Torquato, I.-C. Kim, and D. Cule, J. Appl. Phys. 85, 1560 共1999兲.
14
Y. Terada and M. Tokuyama, J. Korean Phys. Soc. 38, 512 共2001兲. 27
J. R. Reitz, F. J. Milford, and R. W. Christy, Foundations of Electromag-
15
J. A. Given, J. B. Hubbard, and J. F. Douglas, J. Chem. Phys. 106, 3721 netism Theory 共Addison-Wesley, Reading, MA, 1979兲.
共1997兲. 28
B. A. Luty, J. A. McCammon, and H. -X. Zhou, J. Chem. Phys. 97, 5682
16
C.-O. Hwang, J. Korean Phys. Soc. 共in press兲. 共1992兲.
17
M. E. Müller, Ann. Math. Stat. 27, 569 共1956兲. 29
M. Mascagni and A. Srinivasan, ACM Trans. Math. Softw. 26, 436 共2000兲.

Downloaded 26 Mar 2004 to 165.246.34.43. Redistribution subject to AIP license or copyright, see http://jap.aip.org/jap/copyright.jsp