Professional Documents
Culture Documents
Abstract
The classical isoperimetric inequality in R3 states that the surface of
smallest area enclosing a given volume is a sphere. We show that the least
area surface enclosing two equal volumes is a double bubble, a surface
made of two pieces of round spheres separated by a
at disk, meeting
along a single circle at an angle of 120o .
1 Introduction
Double, double, toil and trouble, Fire burn and cauldron bubble.
Macbeth Act 4, Scene 1, Line 10
In this paper we nd the unique surface of smallest area enclosing two equal
volumes. The surface is called a double bubble, and is made of two pieces of
round spheres separated by a disk, meeting along a single circle at an angle of
120o . This is the form assumed by two equally sized soap bubbles which are
brought together until their boundaries conglomerate to form a common wall.
Isoperimetric problems, which study maximizing the size of enclosed regions
whose boundary is of xed size, are among the oldest problems in mathematics.
Queen Dido, founder of the city of Carthage, is reported to have purchased all
the land that could be enclosed with the hide of a steer [11]. She cut the hide
into a leather rope, which she formed into a circle to surround a hill upon which
the city of Carthage was founded. For a broad discussion of the isoperimetric
problem see Osserman [30].
The two volume isoperimetric problem in R3 was considered by the Belgian
physicist J. Plateau [31] and appears in C.V. Boys' famous book on soap bubbles.
As Boys wrote,
When however the bubble is not single, say two have been blown in
real contact with one another, again the bubbles must together take
such a form that the total surface of the two spherical segments and
of the part common to both, which I shall call the interface, is the
partially supported by the National Science Foundation.
1
smallest possible surface which will contain the two volumes of air
and keep them separate.
Extensive interest in the problem was revived in recent years by work of
Frank Morgan. The planar case has been solved in [1] by methods special
to two dimensions. In higher dimensions it is not obvious that a solution to
the problem even exists. A fundamental paper of Almgren [3] establishes the
existence of solutions to a great variety of geometric minimization problems,
including multiple component isoperimetric problems in R3 . Despite the fact
that existence was established some time ago, our result is the rst explicit
example of a surface in R3 solving a multiple component isoperimetric problem.
Multiple region isoperimetric problems arise in many elds, including for
example the growth and shape of cells [25], [35]. They were extensively stud-
ied by physical and biological scientists in the 19th century [21],[31]. Plateau
established experimentally that a soap bubble cluster is a piecewise-smooth sur-
face having only two types of singularities. The rst type of singularity occurs
when three smooth surfaces come together along a smooth triple curve at an
angle of 120o . The second type of singularity occurs when six smooth surfaces
and four triple curves converge at a point, with all angles equal. The angles
are equal to those of the cone over the 1-skeleton of a regular tetrahedron. A
mathematical proof that these types of singularities are the only ones possible
in a minimizing bubble in R3 was given by Taylor [34]. However there were
no explicit minimizing bubbles known for any collection of volumes exhibiting
either of these singularities. Thus we have found the rst explicit example of
a closed minimizing surface in R3 known to exhibit some of the singularities
predicted by Plateau.
At the heart of our arguments is a mixture of geometrical analysis and
estimates of geometric quantities obtained by the use of numerical computation.
We perform these calculations in such a way that there are strict estimates on
the accuracy of the computations.
But yet I'll make assurance double sure.
Macbeth Act 4, Scene 1, Line 83
Since it is still somewhat unusual in a mathematical proof to use digital
computers to do calculations involving real numbers, we will say a few words
about the nature of this part of the argument.
Computers are widely known in the mathematical community to have been
used successfully for analysis of discrete and combinatorial problems, such as
the 4-color theorem of Appel and Haken [6]. Problems such as isoperimetric in-
equalities are qualitatively dierent, since they inherently involve real numbers.
Real numbers are represented in digital computers in a
oating point format
which allows exact description of only a nite number of rationals. Simple cal-
culations, like division or even addition, lead to unrepresentable numbers, and
so the computer must round o. For purposes of mathematical proof, the size
of the round o must be accurately bounded throughout the calculation.
2
Methods for strictly estimating solutions of dierential equations exist, but
are not yet widely used in the mathematical community. The vast majority
of numerical work is for approximation and simulation that does not meet the
standards of a mathematical proof. However there are some signicant cases of
oating point numerical methods in traditional mathematical proofs, for exam-
ple Lanford's work on the Feigenbaum Conjectures [24] and Feerman's work
on the stability of matter [14].
Computer calculations are essential to our proof that equal volume double
bubbles minimize area. There are too many to be done by hand.
Within the volume of which time I have seen Hours dreadful.
Macbeth Act 2, Scene 4, Line 2
The proof parameterizes the space of possible solutions by a two-dimensional
rectangle, one dimension corresponding to an angle and the other to a mean
curvature. This rectangle is divided into 22,393 smaller rectangles which are in-
vestigated by calculations involving a total of 90,159 numerical integrals. Every
single calculation is done with strict error bounds, and all results are precise
mathematical statements. All operations conform to the IEEE 754 standard
for computer arithmetic, a widely adopted method of implementing numerical
computations on computers [5].
Our methods indicate that numerical techniques are likely to play an impor-
tant role in future geometrical arguments.
Is there not a double excellency in this?
Merry Wives of Windsor, Act 3, Scene 3, Line 187
The main result is the following, announced in [16]:
Theorem 1 The unique surface of least area enclosing two equal volumes in
R3 is a double bubble.
Corollary 2 For any surface in R3 enclosing two regions, each of volume v,
the area a satises
a3 243v2 ;
with equality if and only if the surface is isomorphic by a rigid motion of R3 to
the double bubble enclosing two regions of volume v.
The rst issue in establishing Theorem 1 is to nd an appropriate category
of surface in which to minimize area.
On my life, my lord, a bubble.
All's Well Act 3, Scene 6, Line 5
3
It suces for our purposes to consider piecewise smooth two-dimensional
surfaces. It is sometimes useful to consider a much more general notion of
surface, the (F; ; ) sets described in [3]. Our arguments actually prove that
the double bubble minimizes in this larger class. Consideration of a larger
class of surfaces is needed primarily in the establishment of the existence and
regularity of a minimizer, carried out in [3] and [34], and we will not need to be
concerned with them in this paper.
Dene a piecewise-smooth curve to be a nite union of smooth curves, with
any two either disjoint or having intersection contained in their boundaries.
Dene a piecewise-smooth surface to be a nite union of smooth surfaces with
piecewise-smooth boundary curves, with any two surfaces either disjoint or in-
tersecting along a piecewise-smooth curve contained in their boundaries. Dene
a bubble to be a piecewise-smooth surface satisfying:
1. Each two dimensional surface has constant mean curvature.
2. The singular set is of the type described by Plateau. It consists of smooth
triple curves along which three smooth surfaces come together at an angle
of 120o and isolated vertices where six smooth surfaces and four triple
curves converge at a point, with all dihedral angles equal. The angles are
equal to those of the cone over the 1-skeleton of a regular tetrahedron.
3. The mean curvatures, around an edge in the singular set where three
surfaces have common boundary, sum up to zero.
The above conditions are necessary for no local perturbation of a piecewise-
smooth surface to decrease the area while preserving the volume in each of its
complementary regions.
A bubble enclosing regions of prescribed volumes is called a minimizing
bubble if it minimizes area among all bubbles enclosing the same volumes. The
regions are not necessarily connected.
Given positive constants v1 and v2 , let a(v1 ; v2 ) denote the inmum of the
area of a piecewise-smooth surface enclosing two regions R1 and R2 in R3 which
are closed bounded sets with disjoint interiors such that volume(R1 ) = v1 and
volume(R2 ) = v2 .
A double bubble enclosing volumes v1 ; v2 is a surface made of three pieces of
round spheres, meeting along a single circle at an angle of 120o , and enclosing
two connected regions having volumes v1 and v2 . We consider the plane to be a
sphere of innite radius in this setting, allowing the interface of a double bubble
to be a
at disk.
It has been conjectured since the work of Plateau that double bubbles give
the most ecient shape for enclosing two given volumes.
Conjecture 3 The double bubble enclosing volumes v1 ; v2 has area equal to
a(v1 ; v2 ).
Theorem 1 solves this conjecture in the case that v1 = v2 .
4
Figure 1: Equal and unequal volume double bubbles in cross-section.
5
Figure 3: Cross-section of a bubble with an empty torus region
congurations, the double bubble and an additional class of possibilities, called
torus bubbles, whose properties will be discussed in Section 4.
In Section 5 we describe the algorithm used in a series of computations
which show that torus bubbles are not minimizers for the two equal volume
isoperimetric problem in R3 .
6
generalized notion dened using the methods of geometric measure theory. For
our purposes it suces that this class includes the piecewise-smooth surfaces.
Theorem 2.1 ([3]) S (v1 ; v2 ) exists and is a smooth surface almost everywhere.
The nature of the singularities of S (v1 ; v2 ) was established by Taylor.
Theorem 2.2 ([34]) S (v1; v2 ) is a piecewise-smooth surface. Its singularities
consist of smooth triple curves along which three smooth surfaces come together
at an angle of 120o and isolated points where four triple curves and six pieces of
surface converge. At these isolated points the asymptotic cone is the cone over
the 1-skeleton of a regular tetrahedron.
Thus Taylor's work established that S (v1 ; v2 ) is a bubble, in our terminology.
We now establish some properties of minimizing bubbles.
Lemma 2.3 Let
be an oriented curve in R3 intersecting a minimizing bubble
B transversely at regular points such that the initial and nal points of
lie in
the interior of the same region. Then the sum of the mean curvatures of all the
points of
\ B , oriented by
, is zero.
Proof: Perturb the curve slightly so that each of its intersections with B be-
comes perpendicular. Consider the innitesimal eect of a deformation of B
which pushes points in R3 near
a uniform distance along the curve. To rst
order this preserves the volume of each region. The derivative of the area, to
rst order, is given by the sum of the mean curvatures over the points of
\ B .
If this is non-zero, a deformation can be dened which decreases area while
preserving the volume of each region.
A special case of the above lemma occurs when
is a simple closed curve
encircling a triple curve of the bubble. The lemma then implies that the sum
of the mean curvatures around the triple curve adds up to zero. This local
minimization condition is built into our denition of a bubble. More generally,
the lemma implies that any two surfaces separating the same pair of regions has
the same mean curvature.
Corollary 2.4 If S1 and S2 are surfaces in a bubble which each separate regions
R1 and R2 , then their mean curvatures are equal.
Proof: Consider an arc which starts in R1 passes through S1 into R2 , then
on through S2 back into R1 . A sub-arc of starts and ends in R2 . Applying
Lemma 2.3 to both and implies the corollary.
Lemma 2.5 A minimizing bubble B is connected.
Proof: If not, we can move a component by an isometry of R3 until it touches
a distinct component. The resulting singularity violates those allowed in Theo-
rem 2.2.
7
The following result was rst observed by B. White and F. Morgan, and rst
written in [15]. Generalized versions are proven in [29] and [20]. See also [16].
Since we will need to refer to the proof, we present a simple argument for the
case of two regions in R3 . A more detailed argument can be found in [29] and
[20]
Here's ne revolution, an we had the trick to see't.
Hamlet, Act 5, Scene 1
Theorem 2.6 S (v1 ; v2 ) is a surface of revolution.
Proof: Given any vector Z on the unit 2-sphere in R3 , there is a plane PZ in R3
perpendicular to Z which bisects the volume v1 + v2 enclosed by S (v1 ; v2 ). This
plane is unique since S (v1 ; v2 ) is connected. Let f (Z ) denote the proportion of
the volume of R1 on the side of PZ to which Z points. Then 0 f (Z ) 1 and
f ( Z ) = 1 f (Z ). Thus on any great circle the Intermediate Value Theorem
implies that there are at least two points where f (Z ) = 1=2 and the plane PZ
bisects the volume of both regions. We x Z to be such a vector.
Consider the intersection of S (v1 ; v2 ) with each of the two half-spaces de-
termined by PZ . Re
ection of the smaller area piece of S (v1 ; v2 ) lying in one
of these half-spaces gives a new surface S1 (v1 ; v2 ) which encloses regions of the
same volumes. Since S1 (v1 ; v2 ) cannot have less area, it must have the same area
as S (v1 ; v2 ). S1 (v1 ; v2 ) has the property that re
ection through PZ preserves
the new regions, which we continue to call R1 and R2 .
There is a great circle of directions in R3 which is perpendicular to Z , so we
can repeat the above argument to nd a plane PW perpendicular to PZ which
bisects the volume of both regions of S1 (v1 ; v2 ). This gives a surface S2 (v1 ; v2 ),
with the same area as S (v1 ; v2 ), for which re
ection through each of the two
perpendicular planes PZ and PW preserves R1 and R2 .
Since composing re
ections through two perpendicular planes gives a rota-
tion of angle , it follows that rotation of angle about L = PZ \ PW preserves
the regions R1 and R2 . Now consider any plane Q containing L. If the inter-
section of Q with S2 (v1 ; v2 ) is not perpendicular to Q, then replacing half of
S2 (v1 ; v2 ) with its re
ected image through Q gives a new minimizing surface
with singularities violating those allowed by Theorem 2.2. Thus S2 (v1 ; v2 ) is
perpendicular to each plane through L and is a surface of revolution around L.
S1 (v1 ; v2 ) and S2 (v1 ; v2 ) may not be identical, but they coincide on a half-space
of R3 . Moreover this half-space was chosen arbitrarily, so that S1 (v1 ; v2 ) is cut
by PW into two pieces, each a half of some surface of revolution. The axis of
each of these surfaces of revolution is L = PZ \ PW , so that they coincide and
S1 (v1 ; v2 ) itself is a surface of revolution. Similar reasoning shows that S (v1 ; v2 )
is also a surface of revolution, and coincides with S2 (v1 ; v2 ) everywhere. If we
make S1 by re
ecting the other half of S across PZ , then the same argument
shows that the other half of S is a surface of revolution about a line L0 in PZ .
Almgren's regularity results, and in particular the unique continuation property
8
of constant mean curvature surfaces, imply that L0 = L, and hence S itself is a
surface of revolution about L.
Remark. This theorem reduces the double bubble problem to a problem
about networks of curves in the upper half-plane. It is possible to apply work of
Morgan on soap bubbles in surfaces [28], together with bounds on topological
complexity proved by Hutchings [20], to give a direct proof of Theorem 2.2 for
the very special case of the two region isoperimetric problem in R3 .
We next summarize some key results obtained by Hutchings in [20]. Recall
that S (v1 ; v2 ) is a minimizing bubble that separates R3 into regions R1 of volume
v1 and R2 of volume v2 and that a(v1 ; v2 ) is the area of S (v1 ; v2 ).
Theorem 2.7 The function a(v1 ; v2 ) is strictly concave on [0; 1) [0; 1).
Corollary 2.8 If v10 > v1 then a(v10 ; v2 ) > a(v1 ; v2 ).
Proof: As v1 ! 1, a(v1 ; v2 ) a(v1 ) ! 1. Concavity of the function a implies
the corollary.
Corollary 2.9 S (v1 ; v2 ) has no empty regions.
Proof: If there is an empty region, add it one of the other regions by removing
a face on its boundary, and apply the previous corollary.
Deeper connectedness results also follow from concavity. In particular, Hutch-
ings deduced that each of R1 and R2 is connected if the volumes are equal.
Theorem 2.10 S (v1 ; v1 ) encloses exactly two connected components.
Proof: See Hutchings [20, Theorem 4.2].
3 Delaunay Surfaces
In this section, we summarize the classication theory of constant mean curva-
ture surfaces of revolution, the Delaunay surfaces, and present some properties
of these surfaces that will be used in our study of bubbles.
Consider a surface of revolution about the x-axis in R3 , with generating
curve contained in the upper half-plane. If the generating curve is a graph,
y = y(x), then the mean curvature h, which we dene to be the sum of the
principal curvatures, is given by
km = (1 + y_y2 )3=2 ;
kp = p 1 2 ;
y 1 + y_
h = km + kp ;
9
where km is the curvature at (x; y) of the generating curve and kp the normal
curvature of the parallel curve. kp is equal to the reciprocal of the distance to
the x-axis along the perpendicular to the generating curve [33].
The sign of the mean curvature depends on a choice of normal to the surface.
We could refer instead to the mean curvature vector eld, equal to the trace of
the second fundamental form of the surface. The mean curvature can then be
obtained by taking the inner product of the mean curvature vector eld with a
unit normal giving the orientation. The above formula gives the correct sign for
the mean curvature when the unit normal vector eld is oriented towards the
x-axis.
Surfaces of revolution having constant mean curvature were rst studied by
Euler, and classied by Delaunay [12].
Theorem 3.1 The surfaces of revolution of constant mean curvature are the
plane, sphere, catenoid, cylinder, unduloid, and nodoid. The generating curves
of nodoids and unduloids are periodic along the x-axis, and have exactly one
local minimum and one local maximum in each period. The generating curve of
an unduloid is a graph over the x-axis. The generating curve of a nodoid has
one local maximum, one local minimum and two vertical points (points where
y_ = 1) in each period.
Proof: Expositions of the classication of Delaunay surfaces can be found
in [13] and [21].
The nodoid and unduloid are not as well known as the other surfaces. Un-
duloids can be obtained by rolling an ellipse along the x-axis and tracing the
path taken by one of the foci. Nodoids can be similarly obtained by rolling a
hyperbola. See Figure 5.
10
The mean curvature of the surface of revolution generated by a C 2 graph
y(x) is given by:
h = km + kp = (1 + y_y2 )3=2 + p 1 2 : (1)
y 1 + y_
We can integrate this ODE once, after multiplying through by 2yy_ .
0 = h+ y p
1
(1 + y_ 2 )3=2 y 1 + y_ 2
= 2hyy_ + 2yy_2y3=2 p 2y_ 2
(1 + y_ ) 1 + y_
df
= dx ( x )
12
Corollary 3.4 For a nodoid, fh > 0. For an unduloid and a cylinder fh < 0.
For a sphere, a catenoid and a plane, fh = 0.
Proof: For a sphere of radius r, we can compute f at a maximum point where
f = hr2 2r:
Since h = 2=r, we have f = 0. We also compute fh for an unduloid or cylinder
at a local maximum. By Lemma 3.3 we know that 0 < h < hS = 2=r. Then
f = hr2 2r < (2=r)r2 2r = 0:
So h > 0 and f < 0 implying that hf < 0.
We compute fh for a nodoid at a vertical point, where y_ = 1 and f = hy2 .
Then fh = h2 y2 > 0.
The next proposition deduces some useful properties of Delaunay surfaces.
Proposition 3.5 Let D be a Delaunay surface with generating curve , mean
curvature h, maximum y-value yM and minimum y-value ym . Then
1. If D is a nodoid or unduloid, the arc-length of one period of is 2=h.
2. If D is an unduloid, the period P of satises
2(yM + ym ) P (yM + ym ):
3. If D is a nodoid,
h = 2=(yM ym ):
If D is an unduloid,
h = 2=(yM + ym ):
4. If D is a nodoid, has non-zero curvature.
Proof:
1. In [21] the equation for the generating curve of a Delaunay surface is given
in terms of arclength, and it is shown that generating curves of nodoids
and unduloids are periodic, with the arc length of one period given by
2=h.
2. If D is an unduloid, the period P of is equal to the perimeter of the
ellipse that is rolled along the x-axis to generate it [13]. The inequalities
follow by comparing to twice the length of the major axis, 2(yM + ym ),
and the perimeter of an enclosing circle whose diameter is the length of
the major axis, (yM + ym ):
13
3. If D is a nodoid, we calculate the force at yM and ym . At yM we get
f = hyM2 2yM :
At ym we get
f = hym2 + 2ym ;
where the sign of the second term changes since the orientation has re-
versed. Setting these equal we get
hyM2 2yM = hym2 + 2ym
h(yM2 ym2 ) = 2(yM + ym )
h(yM ym ) = 2
h = 2=(yM ym ):
For an unduloid, the proof is identical except for one sign change.
4. At a point where has zero curvature,
km = (1 + y_y2 )3=2 = 0 ;
so y = 0, and has an in
ection point.
Equation 5 then gives a value for the mean curvature:
0 = h(1 + y_ 2 )3=2 + 1 +y y_ ;
2
h= p
1
y 1 + y_ 2
Equation 4 gives that the mean curvature of a sphere with slope y_ at
height y is given by:
h S = p 2 2 > 0:
y 1 + y_
This is exactly twice the value of h at a point of zero curvature whose
height is y and slope is y_ . For a nodoid, h > hS or h < 0 by Lemma 3.3,
so we cannot have hS = 2h and there are no points of zero curvature when
the nodoid is a graph. At points where the nodoid is vertical, h = km since
kp = 0. Thus the curvature is non-zero at vertical points except possibly
when h = 0, which would imply that the Delaunay surface is a plane or a
catenoid, rather than a nodoid as assumed.
A constant mean curvature surface is called stable if there is no normal
variation which decreases its area while preserving the volume on each side
of it. Important stability formulae for constant mean curvature surfaces were
developed by J.L. Barbosa and M. do Carmo [9].
14
Take any shape but that, and my rm nerves shall never tremble.
Macbeth Act 3, Scene 4, Line 102
Proposition 3.6 The smooth subsurfaces of a minimizing bubble S (v1 ; v2 ) are
stable subsurfaces of Delaunay surfaces.
Proof: S (v1 ; v2 ) is a piecewise-smooth surface of revolution whose smooth sub-
surfaces have constant mean curvature by Theorem 2.6 and Theorem 2.2. Any
constant mean curvature surface of revolution in R3 is a Delaunay surface. If
one of these subsurfaces is not stable, then there is a small perturbation which
maintains the volume of each region while decreasing the area. The surface is
then not a minimizer.
We next deduce a formula for the horizontal distance between two points on
a Delaunay curve whose y-coordinates are known.
Proposition 3.7 Let (x1 ; y1 ) and (x2 ; y2 ) be two points on a Delaunay curve
with y1 < y2 . Suppose that between the points the curve is a graph x(y) over the
y-axis. Then Z y2
x2 x1 = p
t dy (6)
y1 (2y + t)(2y t)
where t = t(y) = hy2 f .
If x1 x2 and the curve is also a graph over the x-axis, then the volume
underneath the curve between x = x1 and x = x2 is given by:
Z y2
V (x1 ; x2 ) = p
y2 t dy (7)
y1 (2y + t)(2y t)
Proof: We solve Equation 2 for dx=dy, using that dy=dx 6= 0, and then apply
the change of variables formula.
f = hy2 p 2y 2
1 + y_
1 + y_ 2 = 2ty
p
y_ 2 = 4y t2 t
2 2
r
dy = 4y2 t2
dx t2
dx = p t
dy (2y + t)(2y t)
Note that the sign of t = t(y) is positive if and only if the curve is oriented
to the right, i.e. if cos > 0 in Equation 3. Hence dx is the correctly signed
measure of displacement in the x direction.
15
The volume of the region enclosed by a surface of revolution generated by a
graph y(x) between x1 ; x2 with x1 x2 is given by
Z x2
v= y2 dx :
x1
Since we have already deduced a formula for dx, the volume formula follows
immediately.
The integrals of Proposition 3.7 are singular at a local minimum or local
maximum, and we need to apply a change of variables to obtain a formula
which will apply near such points.
Proposition 3.8 Let (x1 ; y1 ) and (x2 ; y2 ) be two points on a Delaunay curve
which is a graph over the x-axis, with exactly one critical point between them with
value ym . Let h; f be the mean curvature and force associated to the Delaunay
curve.
If ym is a minimum, let y = y(z ) = ym + z 2 , and let t = t(y(z )) = hy(z )2 f .
Then Z py2 ym
x2 x1 = p p
2t dz (8)
y1 ym (2 y + t )(2 hym hy)
and Z py2 ym
V (x1 ; x2 ) = p p
2y2 t dz (9)
y1 ym (2y + t)(2 hym hy )
If ym is a maximum and y(z ) = ym z 2 then
Z pym y2
x2 x1 = p p
2t dz (10)
ym y1 (2y + t)(hy 2 + hym )
and Z pym y2
V (x1 ; x2 ) = p p
2y2 t dz (11)
ym y1 (2y + t)(hy 2 + hym )
Proof: These formulae follow from applying a change of variables to y in Equa-
tion 6 and Equation 7. Suppose ym is a minimum. Divide the Delaunay curve
into the two pieces on opposite side of the minimum, and apply Proposition 3.7
and the substitution y = ym + z 2 to each piece.
Near a minimum of y = y(x), the substitution y = ym + z 2 gives the above
formulae when plugged into the integrals of Proposition 3.7.
Note that one of the quadratics in the denominator factors as
2y t = (y ym )(2 hym hy)
since f = hym 2ym . A factor of z from the change of variables cancels the z in
2
the denominator causing the singularity. The integrands are even functions of
z, and y = ym corresponds to z = 0, so the curve portions on opposite sides of
the minimum can be mapped to positive and negative z values, and combined
into one integral.
The maximum case is similar. Near a maximum the substitution y = ym z 2
is used.
16
4 Torus bubbles
We have established in Lemma 3.6 that a minimizing bubble must be obtained
by revolving a union of Delaunay curves contained in the upper half plane. A
key case of such a bubble is the torus bubble, constructed as follows. Take two
circular arcs of the same radius, facing each other, each with one endpoint and
center on the x-axis, and connect the other endpoints with Delaunay curves
meeting at 120 degrees. Rotating around the x-axis, we get a piecewise-smooth
surface surrounding two components, one homeomorphic to a torus, which we
call the torus component T , and one homeomorphic to a ball, which we call the
ball component B . It is not immediately clear whether it is possible to make such
a construction so that the curves meet at 120o angles and the mean curvatures
sum to zero around each triple curve. Such torus bubbles do indeed exist, but
we will show that none of them are minimizers. Figure 6, generated by John
Sullivan, shows a torus bubble.
17
signs chosen so that so that hi is positive if the mean curvature vector points
into B and negative if it points into T , and ho is positive if the mean curvature
vector points into T and negative if it points outside the bubble.
Denote by (x1 ; y1 ) the initial point where the two Delaunay surfaces start,
and (x2 ; y2 ) the point where they rejoin.
To
(x2,y2)
T
(x1,y1)
Ti
S1 B S2
θ1 θ2
18
Proof: If there is such a symmetry, then the bubble is invariant by re
ection
through the x-y plane and by re
ection through a perpendicular plane P . The
argument given in Theorem 2.6 shows that the bubble is a surface of revolution
around the line of intersection of these two planes. But it is also a surface of
revolution around the x-axis, and therefore a surface of revolution around two
perpendicular axis. It must then be a union of spheres, and not a torus bubble.
Proposition 4.5 If a torus bubble contains a nodoid with 2 vertical points, then
it is unstable.
Proof: We can construct a volume preserving Jacobi eld supported on a proper
subset of the To by taking the normal part of the parallel unit vector eld @=@y,
as in [9, Proposition 2.24].
The next proposition establishes that minimizing torus bubbles do not con-
tain Delaunay curves which are longer than a full period.
Proposition 4.6 If i or o contain a full period of a Delaunay curve, then the
torus bubble is not a minimizer.
Proof: If a boundary surface of the torus component is a subsurface of a nodoid
with width more than a period, then the associated generating curve contains
two vertical points, which implies instability by Lemma 4.5.
If the boundary surface of the torus component is a subsurface of an unduloid
then the lemma is less obvious. Note that we cannot necessarily construct a
volume preserving Jacobi eld from the vector eld @=@x, since the Delaunay
curve may not contain two maxima or two minima. We establish it instead with
a re
ection argument. Assume rst that the unduloid is the inner surface Ti .
An arc of an unduloid curve U which is longer than a period contains three
distinct points with the same y value, (x1 ; y1 ); (x2 ; y1 ); (x3 ; y1 ), with x1 < x2 <
x3 . Re
ect the part of U between the planes x = x1 and x = x3 through the
plane x = (x1 + x3 )=2 to get a new boundary surface U 0 . Re
ection preserves
both area and the volume under U between the planes x = x1 and x = x3 . If
U 0 does not intersect o then the volume of the torus bubble is also preserved.
In this case re
ection causes an angle of less than 180o for U 0 along the circles
x = x1 and x = x3 , and U 0 can be perturbed slightly to decrease area while
preserving enclosed volumes, implying that the torus bubble is not a minimizer.
Possibly U 0 does intersect the outer curve o . In that case the total volume
enclosed by the union of U 0 and To between x = x1 and x = x3 is increased.
U 0 does not intersect the spherical caps of the torus bubble, since they do not
meet the region between the planes x = x1 and x = x3 . Since the volume under
U 0 is equal to the volume of B , the volume in the remaining region must be
larger than that of the original torus region T . There is no increase in area,
contradicting Corollary 2.8.
Assume now that the unduloid is the outer surface To . We again re
ect it to
get a new boundary surface U 0 . The total volume underneath U 0 is the same as
that underneath U . Possibly the inner surface Ti intersects the re
ected surface
19
U 0 . In this case the total volume underneath the union of the two surfaces
has increased. Divide the region underneath U 0 so that the region under Ti
is allocated entirely to B , as before, and the other regions are allocated to T .
Then the total volume of B is preserved, that of T is increased, and area is the
same, a contradiction as before.
Finally, a cylinder of radius r and length longer then 2r is shown to be
unstable in [9].
We note that a somewhat similar argument was used by Athanassenas [7]
in studying the stability of minimizing capillary surfaces. See also [36]. Our
situation has additional complications due to the possible intersections resulting
from the re
ection.
Lemma 4.7 If a torus bubble has 1 < 60o and 2 < 60o then it is not mini-
mizing. In particular, 2 60o in a minimizing torus bubble.
Proof: If both 1 and 2 are less than 60o , the outer bubble To is a nodoid that
becomes vertical twice. Lemma 4.5 implies it is unstable. Since 2 is the larger
angle, 2 60o .
Lemma 4.8 A torus bubble with 2 120o is completely contained inside the
smaller region of the double bubble whose larger outer sphere subtends an angle
equal to 2 .
Proof: Given a vector in the upper half-plane, there is a unique circle perpen-
dicular to the x-axis which is tangent to that vector. A Delaunay curve tangent
to the same vector with greater mean curvature than the corresponding sphere
is a nodoid, by Lemma 3.3. A Delaunay curve tangent to the same vector with
smaller, but still positive, mean curvature is an unduloid or cylinder. Consider
the double bubble whose larger outer sphere subtends an angle equal to 2 .. The
smaller outer spherical cap in the double bubble is tangent to To at (x2 ; y2 ). If
the mean curvature of To is smaller than that of the tangent spherical surface
in the double bubble, then both To and Ti are subsurfaces of unduloids, and
they cannot intersect except at the circle of revolution through (x2 ; y2 ). See
Figure 8. If the mean curvature is greater, then To and Ti are subsurfaces of
nodoids, which intersect before leaving the smaller region of a double bubble,
and hence have smaller volume then the smaller region of the double bubble
whose larger sphere subtends angle 2 .
Corollary 4.9 In a minimizing equal volume torus bubble, 2 < 120o .
Proof: Otherwise Lemma 4.8 implies that the volumes are not equal.
Lemma 4.10 In a minimizing equal volume torus bubble, ho > 0.
Proof: We know from Lemma 4.3 that ho 0. If ho = 0 then To is a catenoid
and strictly increasing, implying, in particular, that 2 > 120o .
Knowledge of ho ; 1 ; 2 can be used to give a fairly accurate qualitative pic-
ture of the torus bubble. To specify our angle choice, we orient the generating
curves i and o of Ti and To so that they run from (x1 ; y1 ) to (x2 ; y2 ).
20
To
S2
Ti
Proposition 4.11 A minimizing equal volume torus bubble has the following
properties:
1. i has negative slope at (x1 ; y1 ) if and only if 1 > 30o . Its angle with the
positive x-axis, is 30o 1 .
2. i has positive slope at (x2 ; y2 ) if and only if 2 > 30o . Its angle with the
positive x-axis is 2 30o .
3. o has positive slope at (x1 ; y1 ) if and only if 1 > 60o . Its angle with the
positive x-axis is 150o 1 .
4. o has angle at (x2 ; y2 ) with the positive x-axis equal to 2 150o .
5. i is a graph. It has a unique local minimum if and only if 1 30o . It
never has a local maximum. If Ti is a nodoid then hi < 0.
6. o always has a unique local maximum. It has a vertical tangent on the
left if and only if 1 60o . It never has a vertical tangent on the right. It
never has a local minimum.
7. If 1 30o the local minimum ymin of Ti has value
ymin = 1 + p1f+i f h
i i
where fi is the force associated to Ti .
8. The local maximum ymax of To has value
p
ymax = 1 + 1 + fo ho
ho
where fo is the force associated to To .
Proof: The rst four assertions follow from the fact that the surfaces meet at
120o angles along a triple curve.
21
If i is not a graph then 2 > 120o , violating Corollary 4.9. If 1 30o then
i has non-positive slope at (x1 ; y1 ). Since 2 > 1 the curve must have a local
minimum in this case.
Suppose i has a local maximum. Since 2 > 60o , i has positive slope at
(x2 ; y2 ) and must have a local minimum as well. However 2 > 1 so i must be
longer than a period, contradicting Proposition 4.6. So it has no local maximum.
If hi 0 and Ti is a nodoid, then we can compare Ti and To to a double
bubble whose interface coincides with S1 , and apply Lemma 3.3. Ti is trapped
inside the component to the right of the interface and To is trapped inside the
component to the left of the interface, as in Figure 9. In particular, h0 < 0 in
this case, a contradiction.
To
Ti
S1
23
y4 ymax ymax
yleft y4 y2
y2
yleft
y3
y3 y1 ymin
y1
ymax
y4
y2
y1
y3
ymin
24
Figure 11: Some eliminated torus bubble possibilities.
25
We calculate in the next proposition the value of the forces fo of To and fi
of Ti .
Proposition 4.15 For the torus bubble determined by 1 and ho, the forces fo
of To and fi of Ti are given by:
p
fo = (ho 1)(1 c1 2 ) + c1 3(1 c1 2 ):
and p
fi = (hi 1)(1 c1 2 ) c1 3(1 c1 2 ) = fo ;
where c1 = cos 1 .
Proof: We rst calculate the force fo of To . At (x1 ; y1 ), y1 = sin 1 and
= 5=6 1 is the angle between the positive x-axis and the graph of o,
fo = ho y1 2 2y1 cos
= ho sin2 1 2 sin 1 cos(5=6 1 )
p
= ho (1 cos2 1 ) 2 sin 1 (( 3=2) cos 1 (1=2) sin ( 1 ))
p
= ho (1 c1 2 ) + c1 3(1 c1 2 ) (1 c1 2 )
p
= (ho 1)(1 c1 2 ) + c1 3(1 c1 2 ) :
Similarly,
fi = hi y1 2 2y1 cos
= hi sin2 1 2 sin 1 cos(=6 1 )
p
= hi (1 cos2 1 ) 2 sin 1 (( 3=2) cos 1 (1=2) sin ( 1 ))
p
= hi (1 c1 2 ) c1 3(1 c1 2 ) (1 c1 2 )
p
= (hi 1)(1 c1 2 ) c1 3(1 c1 2 ) :
Since ho 1 = 1 hi = (hi 1) by Lemma 4.3, fo = fi .
Note: The fact that fo = fi is an example of a general \balancing
principle" for constant mean curvature surfaces, due to R. Kusner [23].
Lemma 4.16 For the torus bubble determined by 1 and ho , the angle 2 is one
of the (at most) two solutions of the equation
f (cos 1 ) = f (cos 2 );
where p
f (c) = (ho 1)(1 c2 ) + c 3(1 c2 ):
Proof: We calculate the force fo of To at both (x1 ; y1 ) and (x2 ; y2 ).
At (x1 ; y1 ) p
fo = (ho 1)(1 c1 2 ) + c1 3(1 c1 2 );
26
where y1 = sin 1 and c1 = cos 1 . A similar calculation at 2 shows
p
fo = (ho 1)(1 c2 2 ) + c1 3(1 c2 2 ):
Setting the two forces equal and applying Lemma 4.14 gives that there are at
most two possible values for c2 = cos 2 . Since 0 < 2 < , where the cosine
function is monotonically decreasing, this implies that there are at most two
possible values for 2 .
One solution occurs when the angles are equal. An example of such a sym-
metric torus bubble, depicted in Figure 6, occurs with 1 = 2 = 90o , and
ho approximately 1:9848. However these symmetric solutions never lead to a
minimizing torus bubble.
Proposition 4.17 In a minimizing torus bubble, 1 < 2 .
Proof: Given two points on any Delaunay curve with the same y-value and
one slope equal to the negative of the slope of the other, there is a re
ectional
symmetry of the curve through the plane half-way between the two points. It
follows that if 1 = 2 then the entire torus bubble has a right/left re
ectional
symmetry and is unstable by Proposition 4.4.
Thus for each 1 , ho , there is at most one value of 2 which can potentially
lead to a stable torus bubble.
From now on we will restrict attention to nonsymmetric torus bubbles, and
assume without loss of generality that 1 < 2 .
Proposition p 4.18 If a minimizingp torus bubble has 60o < 1 < 120o then
ho > 1 3 cot 1 and hi < 1 + 3 cot 1 . The same equation holds with 2
replacing 1 .
Proof: We use Lemma 3.3 to compare the torus bubble with a double bubble
whose smaller exterior sphere subtends an angle of 1 and has mean curvature 2.
If hi is larger than the mean curvature of the corresponding spherical component
of the double bubble (the interface between the two components), then Ti is a
nodoid trapped inside the smaller component of the double bubble, as shown
in Lemma 3.3. See Figure 12. It follows that o cannot intersect i except at
(x1 ; y1 ).
Corollary 4.19 To is a nodoid.
Proof: If 60o 1 120o then Proposition 4.18 shows that the mean curvature
of To is greater than that of the sphere which is tangent to it at (x1 ; y1 ), and
hence Lemma 3.3 implies that To is a nodoid. If 1 < 60o then o starts out
heading to the left. It must become vertical, hence To must be a nodoid.
Lemma 4.20 Let N be a subcurve of a generating curve for a nodoid of mean
curvature h > 0, running between a vertical point on the left at (xv ; yv ) and a
vertical point on the right at (x0v ; yv0 ), with a unique local maximum at (xM ; yM ).
Then a circle C of curvature h which is tangent to N at (xv ; yv ) lies underneath
N.
27
Ti To
S1
28
The assumptions ho :2 and 1 5:7o give an upper bound for the force
fo of To . Calculating at height y1 = sin 1 < 1 5:7=180 < :1 gives
< :09:
We estimate the value of yv by solving fo = yv2 ho .
yv2 = fo =ho < :09=ho
so p
yv < :09=ho :
We estimate x1 xv by the Intermediate Value Theorem:
(yv y1 )=(xv x1 ) = y_ (p)
at some point p 2 (xv ; x1 ). p
o is convex between xv and x1 by Lemma 3.5, so jy_ (p)j > jy_ (x1 )j > 1= 3,
and
p p pp p
jxv x1 j = j(yv y1 )=y_ (p)j < 3(yv y1 ) < 3yv < 3 :09=ho = :27=ho:
This implies that
p
xM x1 = (xM xv ) (x1 xv ) > 1=ho :27=ho :
p
For 0 < ho :2, dierentiation shows that the function 1=ho :27=ho
is decreasing
p and its minimum value occurs at h o = : 2, so that x M x1 >
1=(0:2) :27=(0:2) > 3:8. Thus o travels a distance of at least 3:8 to the
right before hitting i . But the period of i is less than 3:5. So either o changes
direction and has two vertical points and is not a minimizer by Lemma 4.7 or
i contains a full period and is not a minimizer by Proposition 4.6.
The next lemma will be used to give bounds on the size of a circle enclosing
a loop of a nodoid.
Lemma 4.22 If a curve
in the positive quadrant is tangent to the y-axis at
the origin and has curvature k > 1=r > 0, then it does not cross C (r), the
semi-circle of radius r centered at (0; r).
Proof: If it does cross C (r), denote by
0 the subcurve of
starting at the
origin and running to the rst point of intersection with C (r), as in Figure 13.
Let
t0 =
0 + (0; t) be a vertical translate of
0 by a vertical distance of t. Let
T = supft :
t0 \ C (r) 6= ;g. Then
T0 meets C (r) at a point P , but not its
interior, and thus is tangent to C (r) at P . Since the curvature of
T0 is greater
than that of C (r), this is a contradiction.
29
Lemma 4.23 If o and i have
p curvature greater than 1=r, then they are con-
tained in a circle of radius r 3=2.
Proof: By Lemma 4.22 we know that the intersection of the torus component
with the upper half-plane is contained in the intersection of two circles of radius
p at an angle of 120o . This intersection is contained in a circle of
1=r meeting
radius r 3=2.
C(r) γ 'T
γ'
'
Figure 13: The torus component generating curves are trapped inside two circles
meeting at 120o .
Proposition 4.24 Suppose that a torus bubble has hi < 0 and that the inter-
section of the torus component with the upper half-plane has minimum y-value
equal to y0 with y0 > 1= hpi . Then the torus component contains volume
v 22 r2 (y0 + r), where r = 3y0 =2( hi y0 1).
Proof: For a generating curve passing through a point (x; y) in the upper
half-plane,
jkm j jhj jkp j
= jhj p 1
y 1 + y_ 2
jhj y1 :
If y y0 , then
jkm j hi y1 : (15)
0
In a torus bubble we always have that ho > hi , so that the curves o and i
each have curvature greater than hi 1=y0 . By Lemma 4.23 the intersection of
the torus component with the upper half-plane is contained in a circle of radius
30
p
r 3y0 =2( hi y0 1). The volume is obtained from the formula for the volume of
a torus of revolution generated by a circle of radius r centered at height y0 + r)
above the x-axis.
Lemma 4.25 If a torus bubble has hi k < 0 and the intersection of the
torus component with the upper half-plane is contained in fy y0 g where y0 >
1=k, then the torus component contains volume
p
v 2:46(y1 + 23 ky y0 1 )( ky y0 1 )2
0 0
The diameter of w is p
d = 3 ky y0 1 :
0
The center of mass of w has distance from the y-axis equal to the sum of y1
and half the vertical displacement of the chord running between its two vertices.
This chord has angle =2 1 with the x-axis. Thus the y-value of the center
of mass of w is given by
y1 + d2 cos(=2 1 ) y1 + d2 :
Pappus' Theorem for the volume of a solid of revolution obtained by rotating
this area around the x-axis gives that
v 2(y1 + d2 )1:23( ky y0 1 )2 ;
p 0
2:46(y1 + 23 ky y0 1 )( ky y0 1 )2 :
0 0
We apply these volume estimates to reduce the range of possible mean cur-
vatures in a minimizing equal volume torus bubble.
Proposition 4.26 In a minimizing equal volume torus bubble, ho 10.
31
Proof: The hypothesis ho 10 is equivalent to hi 8. Suppose to the
contrary that hi < 8 in a minimizing equal volume torus bubble.
Proposition 4.11 states that i is a graph, so that the volume vi of the
ball component is larger than the volume of the region under S2 . Lemma 4.8
and Lemma 4.7 imply that 60o < 2 < 120o . It follows that vi > 5=24,
the volume under a unit radius sphere subtending an angle of 60o . We now
estimate the volume vo of the torus component. The curves generating
p the
boundary of the torus component contain the point (x2 ; y2 ) where 3=2 y2
1. Their maximum and minimum y-values dier by less than 2=jphi j < 0:25 by
Proposition 3.5, and therefore both curves are contained in fy 3=2 0:25g.
Noting
p that y1 1 and applying Lemma 4.25 with k = 8 and y0 = 0:6 <
3=2 0:25 gives that
p
vo < 2:46(1 + 23 ky y0 1 )( ky y0 1 )2
p 0 0
5 Computation
Like a man to double business bound, I stand in pause where I shall rst begin.
Hamlet Act 3, Scene 3, Line 41
32
The computation was performed using double precision
oating point num-
bers. The fundamental data type is the IEEE 754 64-bit real number. See
[5]. This is a binary representation with a 52-bit mantissa (plus an implied
leading bit), an 11-bit binary exponent, and a sign bit. The IEEE standard
species that the add, subtract, multiply, divide, and square root operations be
performed as if done exactly and rounded to the nearby representable number
according to the rounding mode in eect. There are three rounding modes that
can be chosen: up, down, or nearest. The values +1 and 1 are representable
and behave in the obvious way. Floating point exceptions are masked, but
ags
are sticky and available for clearing and inspection. Only the divide-by-zero,
over
ow, and invalid operation
ags are of interest to us. We will use the term
exception to refer only to one of these. Most computers in use today implement
the IEEE standard.
Combined with the methods of interval arithmetic, see Moore [26], the IEEE
standard allows numerical calculation with exact bounds on accuracy. Interval
arithmetic is a method by which a real valued function on the reals can be
extended to an interval valued function of intervals, with the interval returned
by the function containing all possible function values with arguments from the
domain intervals. This property is sometimes called inclusion monotonicity. An
interval is formed from two IEEE reals. Mathematically, it represents the closed
interval between the two reals. The add, subtract, multiply, divide, and square
root functions are extended to intervals by the IEEE operations on reals along
with directed rounding.
In this section, IEEE reals are denoted by lower case, and intervals of IEEE
reals by upper case. Arithmetic operations are interpreted according to IEEE
and interval rules, not by the usual mathematical denitions. The lower and
upper bounds to an interval are denoted with lower and upper bars, so X =
[X; X ].
Other operations on reals are extended to intervals in straightforward ways.
Relations are interpreted positively, so when X and Y are intervals, X < Y
means for any x 2 X and any y 2 Y , x < y. So we have equivalent expressions
X < Y , X < Y , and X 6= Y , X \ Y = ;. The union of intervals X
and Y is the smallest interval containing the usual sets, and is denoted X [Y .
The interval valued functions absolute value, Max, Min, and intersection are
dened in the obvious way, without rounding. There is no IEEE standard for
transcendental functions, so we designed our algorithm to avoid all calls to
trigonometric functions.
NANs (NAN stands for Not A Number) are never generated in the algorithm.
We assure this by monitoring the IEEE exception
ags.
Bounds obtained with interval arithmetic are usually far from sharp, es-
pecially when using wide intervals. (Our intervals are \fat" in the sense of
Feerman [14].) Often the intervals are wide enough to make a perfectly good
formula look like nonsense. For example, consider a formula involving square
roots. Assuming we're not using imaginary numbers, the validity of the formula
presupposes that the argument to the square root is nonnegative. However,
when we pass to an interval extension of the formula, the interval argument to
33
the square root will often include negative values, and the square root of a neg-
ative number would normally trigger an exception. Such an exception would be
annoying, so we dene the interval square root function Sqrt to simply discard
any negative portion of an interval argument. The justication for this appar-
ent sleight-of-hand rests on the validity of the original formula over the reals.
Since our theorems tell us that the quantity whose square root we are taking is
non-negative, we are justied in truncating the interval to exclude negatives.
The interval function which always returns [ 1; 1] can be used to represent
any function, but not very usefully. We use this interval to return the value of
a division by an interval containing
p zero. The empty set is represented by
[1; 1]. The values of and 3 in our algorithm are expanded to narrow
intervals. Other constants used are representable. Thus we use the expression
5Ho 1 rather than Ho 0:2, which involves the non-representable constant
0:2.
The verify statement is used to assure that a particular condition holds. If
the condition fails, then the entire program is stopped (aborted). If no such
failure occurs, the presence of such a verify statement constitutes a proof that
the condition holds for the given inputs.
Remark. Much of the analysis of this section has an alternative interpreta-
tion in terms of fuzzy logic. An interval may be regarded as a fuzzy real number,
i.e. a real number whose true value is uncertain but constrained to lie within
the interval. The denition of an ordinary step function
f (x) := 0 if x < c
f (x) := 1 otherwise
becomes somewhat problematic if x and c are fuzzy, and their intervals overlap.
Then we no longer have the true-or-false dichotomy of Aristotelian logic, and
can use fuzzy logic to represent uncertain logical values. A fuzzy extension of
the above step function is then:
F (X ) := 0 if X < C
F (X ) := 1 if X > C
F (X ) := [0; 1] if X and C overlap
See the procedure Compare below for a more sophisticated step function
of this type.
The algorithm for rejecting torus bubbles is given below, broken down into
procedures. The idea is to consider a domain of torus bubbles representing a
product of small intervals in each of 1 and ho , and to calculate as much about its
geometry as possible. If no property that can be tested rules these torus bubbles
out as potential minimizers, we calculate their volumes and areas and compare
them to those of the double bubble. The accuracy of these calculations depends
on the size of the domain rectangle we start with. We will show that these can
be chosen small enough to get sucient accuracy, but still large enough that a
reasonable number of them cover all the possibilities.
34
A range of hypothetical torus bubbles is specied by intervals 1 and Ho ,
according to Proposition 4.12. (To avoid unnecessary use of trigonometric func-
tions our algorithm in fact uses an equivalent parametrization of the space of
torus bubbles by intervals S1 and Ho , where S1 = sin 1 .) Rather than solve the
mean curvature dierential equation directly to nd the point (X2 ; Y2 ) where
the Delaunay curves i and o intersect, we assume existence of a torus bubble
to derive Y2 , and then deduce X2 from numerical integrals. Area and volume
calculations can be obtained from additional numerical integrals. At each stage
in the calculation, we check whether the torus bubbles can be rejected based on
the instability results of Section 4. If not, we test whether the volumes are equal.
As it turns out all torus bubbles are rejected for other reasons, so area calcu-
lations are not necessary. Thus they have been eliminated from the algorithm.
Most torus bubbles are rejected because either 2 is out of range, or because the
x-displacement of the Delaunay curves dier, or because the volumes enclosed
dier.
The computation of 2 and Y2 is based on an analysis of the force function.
The integration of the Delaunay curves is complicated by the fact that they
may not be graphs, so we have to choose some parameterization. It turns out
that the most convenient parameterization is in terms of y, because the ODE
involves only y, and because the y-coordinates of the endpoints are specied
by 1 and 2 . This allows x-displacement, volume, and area to be expressed
directly as integrals in terms of y, thereby bypassing a complete ODE solution.
The Delaunay curves are generally not graphs as functions of y, so the in-
tegrals we use have singularities at points where y_ = 0. Each integrand has a
denominator containing a factor which is a square root of a fourth degree poly-
nomial in y. Fortunately, the polynomial factors into linear pieces, and only
one of these factors vanishes when y_ = 0. A change of variables resolves the
singularity, as worked out in Proposition 3.8.
Another diculty is that the geometry sometimes degenerates at the bound-
ary of the regions we are studying. For example, we need to exclude torus bub-
bles with 1 arbitrarily close to 0, but some of our formulae become singular
when 1 = 0, and don't make sense there. If the interval 1 contains 0, then
various intervals representing y-coordinates along the Delaunay curves will also
contain 0, and the integrals will each have a nasty singularity. We get around
this problem by using crude estimates in a small zone near the x-axis, and us-
ing integrals only when a safe distance above the x-axis. This approach works
except when ho is also near 0. In that case we rely on Proposition 4.21.
We now describe some interval procedures used in our algorithm.
The procedure Avgwt calculates a number in between two inputs, to be
used to subdivide integrals later. It is a weighted average.
procedure Avgwt
input: X, Y, w
verify X<Y
verify 0 < w and w < 1
35
z := (1 w)X + wY
verify X < z and z < Y
return z
The procedure Compare checks whether the interval X is to the left of the
interval Y , returning A if yes, B if not, and the union if the intervals overlap.
procedure Compare
input: X; Y; A; B
if X < Y then return A
else if X > Y then return B
else return A[B
36
if (H > 0) then
begin
i
for := 0 to 31 do
Y := Y + F (A + iH + (0[H ))
end
return (0[F (A))Width(A) + (0[F (B ))Width(B ) + Y H
Proposition 5.2 For any integrable f and real numbers a; b; with a b, and in-
terval extensions F; A; B , the integral of f from a to b is contained in Integrate(F; A; B ).
Proof: If A B ,
Z b Z A Z B Z b
f dx = f dx + f dx + f dx :
a a A B
The term Z A
f dx
a
is contained in (0[F (A))(A A) and
Z b
f dx
B
is similarly contained in (0[F (B ))(B B ).
The term Z B
f dx
A
is bounded by lower and upper Riemann sums, using 32 equally sized intervals.
The Y H term, by construction, gives an interval containing those Riemann
sums. Otherwise, A B is false, so A and B overlap, and we can decompose
Z b Z B Z b
f dx = f dx + f dx :
a a B
Since B < A, the Y H term vanishes and the result follows from [a; B ] [a; A]
and inclusion monotonicity.
The next several procedures dene functions to be integrated. A Delaunay
curve can be obtained by integrating the ordinary dierential equation given
in Equation 2 in Section 3. Rather than compute the whole curve, we only
compute the x-displacement and the volume beneath the curve, because that is
all we need, and it is much more computationally ecient.
The procedure Dx is a straightforward interval extension of Equation 6 in
Proposition 3.7. The absolute value sign is a computational convenience. Our
Delaunay curves will sometimes go in the negative x direction, and we explicitly
take that into account by nding the turning point and splitting the integral
37
into the positive and negative parts. The volume element Dv is obtained from
Dx and is also nonnegative.
At local minima and maxima, Dx and Dv are singular. This is because
dy=dx = 0 and is not a computational artifact. So near local extrema, we use a
change of variables to reparametrize the curve, as in Proposition 3.8.
procedure Dx
input: Y; H; F
T := HY 2 F
return T=Sqrt((2Y + T )(2Y T ))
procedure Dxmin
input: Z; H; F
Y := Ymin + Z 2
T := HY 2 F
return 2T=Sqrt((2Y + T )(2 HYmin Y H ))
procedure Dxmax
input: Z; H; F
Y := Ymax Z 2
T := HY 2 F
return 2T=Sqrt((2Y + T )(HY 2 + HYmax ))
procedure Dv
input: Y; H; F
return Y 2 Dx( Y; H; F )
procedure Dvmin
input: Z; H; F
return (Ymin + Z 2 )2 Dxmin ( Z; H; F )
procedure Dvmax
input: Z; H; F
return (Ymax Z 2 )2 Dxmax ( Z; H; F )
The next procedure is used for rejecting a range of torus bubbles. Such a
range is represented by intervals C1 and Ho , representing a range of possible
values for cos 1 and ho . The interval of cosine values C1 is used, rather than
38
an interval of 1 values, to eliminate calls to trigonometric functions by the
algorithm. A variety of tests are performed based on the analysis of previous
sections. The various intervals do not necessarily give sharp estimates, so it is
not a priori clear that very many ranges will be rejected. As it turns out, they
are sharp enough.
procedure CheckRectangle
input: C ;H
1 o
1. if 1000 1 C
995 and 5Ho 1
then return REJECT
2. Define
Hi := 2 Ho
Y1 := Sqrt(1 C12 )
Fi := (Hi 1)Y12 C1 Y1 Sqrt(3)
Fo := Fi
3. C := [ C1 [C1 ] \ [ :5[:5]
2
verify C2 is non-empty.
T := (2(Hi 1)Fi + 3)=(3 + (Hi 1)2 ) (1 C1 2 )
if C2 + T 6= 1 return REJECT
2
5. if C :
1 H
5 and o 1 Sqrt(3) 1 C =Y1
then return REJECT
7. Define:
Vends := (1 C1 )2 (2 + C1 )=3 + (1 C2 )2 (2 + C2 )=3
Ymin := Fi =(1+ Sqrt(1 + Fi Hi ))
Ymax := (1+Sqrt(1 + FoHo ))=Ho
8. if Ymin Hi < 1 then
begin
Y := Compare(C1 ;Sqrt(3)=2; Ymin ; Y1 )
R := (Sqrt3=2)=( Hi 1=Y )
V := 22 R2 (Y + R)
if V < Vends then return REJECT
end
9. Define:
39
if minY <Y Y
2 then 3 := Avgwt( min 2 Y ; Y ; :5)
else return NORESULT
Yleft := Sqrt( o o )F =H
Y
if ( left <Y Y
max ) then left := Max( 1 Y ; Yleft )
else return NORESULT
Y Y ;Y
4 := Avgwt( left max 5) ;:
10. Define:
T := Sqrt(Y1 Ymin )
Z1 := Compare(C1 ;Sqrt(3)=2; T; T )
Z2 := Sqrt(Ymax Y2 )
Z3 := Sqrt(Y3 Ymin )
Z4 := Sqrt(Ymax Y4 )
11. if 1000 1 C
998 then
begin
T Y ;Y ; =
:= Avgwt( 1 2 1 16)
i := ( T Y =
1 )206 100+Integrate(Dx( ; Hi ; Fi ); T; Y2 )
T C =C
:= Sqrt(1 12 ) 1
Y Y
o := ( left 1 )Sqrt(3)
T Y ;Y ; =
:= Avgwt( left 4 1 16)
o := o +Integrate(Dx( o o ) ; H ; F ; T; Y
4)
if o >
i then return REJECT
o := o +Integrate(Dxmax ( o o ) 4 ; H ; F ; Z ; Z2 )
if o > i then return REJECT
if 1 2C1 then return NORESULT
end
12. Define
o :=Integrate(Dx(; Ho ; Fo ); Y1 ; Y4 )+
Integrate(Dxmax (; Ho ; Fo ); Z4 ; Z2 )
if Width(o ) > 20 return NORESULT
verify C1 Sqrt(3)=2 or Y1 Y3
i := Integrate(Dxmin (; Hi ; Fi ); Z1 ; Z3 )+
Integrate(Dx(; Hi ; Fi ); Y3 ; Y2 )
40
V 6 V
15. if i = o return REJECT
else return NORESULT
Vo
Vbase
Vends
41
(fi (hi 1)y2 )2 = 3y2 (1 y2 )
This is a quadratic in y2 . The sum of the two roots in y2 is
2(hi 1)fi + 3
3 + (hi 1)2
and the known root is 1 c1 2 , so subtraction gives the other. Thus the interval
T contains the square of the second root y2 . If there is no such solution C2 with
C22 = 1 T then the bubble cannot be a minimizer and it is rejected.
Step 4 continues to examine the value of y2 . The only possible values for y2 are
in [0; 1], so we intersect before taking the square root of T to get an interval Y2
which contains possible values of y2 , if any. The values in Y2 must be consistent
with the apriori bounds for C2 , so we reject a value which is not. Since
p
fi = (hi 1)(1 c2 2 ) c2 3(1 c2 2 )
by Proposition 4.15, and 1 c2 2 = y22 , we can solve to get an expression for c2
in terms of hi ; fi and y2 . Note that this expression is chosen to give us c2 with
the correct sign. We check the consistency of this expression with c2 in the nal
line of step 4.
Step 5 rejects torus bubbles based on Proposition 4.18. It uses the fact that
cot(1 ) = c1 =y1 .
Step 6 passes on a request to subdivide the rectangle if it is too wide. It does
not reject anything, so there is nothing to justify.
Step 7 denes Vends to be the volume of the solid of revolution between the
spherical ends and the x-axis. See Figure 14. This solid is wholly contained
in the ball component. Ymin is the y-coordinate of the local minimum of the
inner Delaunay curve, based on solving the force equation as in Equation 13 of
Proposition 4.11. This minimum is realized in the torus bubble if and only if
1 30o . Similarly Ymax is the outer Delaunay curve maximum, as in Equa-
tion 4. It is always achieved.
Step 8 excludes torus bubbles where an estimate shows that the torus compo-
nent has less than half the total volume. The volume of the ball component is
bounded below by the volume under the spherical caps, whose value is given by
Vends . An upper bound for the torus component volume is given by the formula
deduced in Lemma 4.24.
Step 9 denes Yleft , the y-value of the point where o has a vertical tangency
on the left as in Figure 10. It is computed by taking the square root of the
nonnegative part of Fo =Ho . Note that Corollary 3.4 guarantees that this value
is positive in a torus bubble, so we are justied in truncating negative values
42
which may have appeared in the calculation before taking the square root. The
outer curve is always a nodoid by Corollary 4.19, and if extended to a vertical
point, the y-coordinate of that point is in the interval Yleft . Replacing Yleft
with Max(Yleft ; Y1 ) assures that Yleft contains the y-coordinate of a vertical
point if there is one, and the initial point otherwise. Y3 and Y4 are used as
intermediate points for chopping integrals. A verify statement contained in the
procedure Avgwt assures that Y3 is smaller than Y2 and larger than Ymin .
Step 10 denes Z1 ; Z2 ; Z3 , and Z4 which are later used as bounds of integration.
The square roots discard the negative part of any interval as before, because for
any particular torus bubble the minimum and maximum y-coordinates provide
bounds for the endpoints of the curves. Z1 is given a negative sign if the lower
curve achieves an interior minimum, and a positive sign otherwise. If the inter-
vals don't allow a conclusion one way or the other about the achievement of a
minimum, then we take the union of both values. Z2 is given a positive sign
because the upper curve always achieves an interior maximum. Z3 and Z4 are
similar.
Step 11 excludes torus bubbles with 1 very small by checking that for these
bubbles o travels further in the positive x-direction than i , so that they cannot
intersect at the required height in Y2 . The angles considered small are those
with cos 1 :998, corresponding to 1 0 where 3o < 0 < 4o .
i is an upper bound for the x-displacement of the inner curve, o a lower
bound for the upper curve. If i < o , then the torus bubble is impossible.
Since the integrands used to calculate Dx become singular at 1 = 0, we use
the Intermediate Value Theorem to estimate i and o when 1 has a value
of no more than 4o . The estimates for i and o near the x-axis are justied
by Proposition 3.5 in the case of To and by Lemma 4.27 in the case of Ti . See
Figure 15 for an illustration of the estimates for i and o .
For 1 < 4o , Lemma 4.27 implies that the inner curve is given by a strictly
increasing function, so that the x-displacement is equal to the integral of Dx
from Y1 to Y2 . This can be broken up into the integral from Y1 to T and the
integral from T to Y2 , where T is set to be one sixteenth of the way from Y1
to Y2 . Lemma 4.27 states that the slope of i below height T is smallest at its
initial point (x1 ; y1 ), so dx=dy cot (30 1 ) cot 26o < 2:06 and the rst
integral is bounded above by that quantity times T Y1 . It follows that i is
an upper bound.
For o we rst nd an upper bound on the leftward displacement between y1
and yleft using a similar intermediate value theoremp estimate. In this case the
slope has absolute value larger than tan 30o = 1= 3 and (Yleft Y1 )Sqrt(3)
gives an upper bound for the leftwards displacement to height Yleft . An in-
tegral is calculated which gives a lower bound for the rightwards displacement
of o between Yleft and Y2 is gotten by integrating dx=dy from a point some-
what above Yleft to Y2 . A lower bound for o is gotten by taking the negative
of the upper bound for the leftwards displacement and adding to it this integral.
43
y2
y left
y 1+(y 2 - y 1)/16
y1< sin(4o)
44
Step 12 calculates intervals for the x displacement of the outer and inner curves
in the case where the angle is larger than those treated in the previous step.
In this step o is an interval containing the x-displacement of the outer curve,
i an interval containing the x-displacement of the inner curve. If the width
of o is larger than 20 then the procedure returns NORESULT, calling for a
subdivision of the input rectangle.
If C1 Sqrt(3)=2, then i achieves an interior minimum. In that case, an inte-
gral along i can be broken up into two pieces, one from Y1 , across the minimum,
to Y3 , and a second one from Y3 to Y2 . Otherwise, if an interior minimum is not
achieved, we can still break up the inner curve into a piece from Y1 to Y3 and
a piece from Y3 to Y2 , but our formulas need to have that Y1 Y3 . A verify
statement assures this.
Step 13 discards failed computations. If i and o are disjoint, then the curves'
nal points cannot coincide.
Step 14 calculates volumes, as indicated in Figure 14. Vbase is the volume inside
Ti . Adding Vbase to Vends gives Vi , the volume of the inner ball component. Vo
is slightly more complicated because the overhang on the left involves a sub-
traction, if there is an overhang. Such an overhang occurs when 1 < 60o . The
volume formula is set up so that volume is counted with a negative sign when
0 is oriented to the left, and positive sign when 0 is oriented to the right. This
is what is required to calculate the volume of the overhang.
Step 15 rejects bubbles if the volumes are unequal.
Volumes of report run with these false and most contrarious quests.
Macbeth Act 4, Scene 1, Line 10
Now that we are able to reject ranges of torus bubbles, we put these ranges
together to reject all the possibilities. Since we don't know in advance how
fat the ranges can be, we break 1 ; Ho space into bite-sized chunks, and then
recursively subdivide further if necessary. We monitor IEEE exception
ags at
this level, so that no computation is trusted if it raised an exception.
procedure DivideAndCheckRectangle
input: 1S ;H
i
C1 := Sqrt(1 1)
2 S
call CheckRectangle( 1 C ;H
o)
if result is REJECT then return SUCCESS
S
split 1 in half, into 1a 1b S ;S
split H H H
o in half, into oa , ob
call DivideAndCheckRectangle( 1a S ;H
oa )
call DivideAndCheckRectangle( 1a S ;H
ob )
call DivideAndCheckRectangle( 1b S ;H
oa )
45
call DivideAndCheckRectangle( 1b S ; Hob )
return SUCCESS
Main program
begin
clear exceptions
i
for := 0 to 99 step 1 do
j
for := 0 to 99 step 1 do
begin
S1 := [i; i + 1]=100
Ho := [j; j + 1]=10
Call DivideAndCheckRectangle( 1 S ; Ho )
end
verify no exceptions raised
Print ``All torus bubbles rejected.''
end
46
We are indebted to Frank Morgan for introducing us to this problem, to J.
Sullivan for generously providing us with computer graphics of torus and double
bubbles, and to Morgan, M. Hutchings, W. Kahan and W. Rossman for helpful
discussions.
References
[1] M. Alfaro, J. Brock, J. Foisy, N. Hodges and J. Zimba The standard
double soap bubble in R2 uniquely minimizes perimeter, Pac. J. Math.
159, 47-59 (1993).
[2] G. Alefeld and J. Herzberger, Introduction to interval computations,
Academic Press, 1983.
[3] F.J. Almgren, Existence and regularity almost everywhere of solutions
to elliptic variational problems with constraints, Memoirs Amer. Math.
Soc. 4, 165-199 (1976).
[4] F.J. Almgren and J. Taylor, The geometry of soap lms and soap bubbles,
Sci. Amer. 235,82-93 (1976).
[5] ANSI/IEEE Standard 754-1985 for Binary Floating-Point Arithmetic,
The Institute of Electrical and Electronic Engineers, New York, 1985.
[6] K. Appel and W. Haken, The solution of the four-color-map problem,
Scientic Amer. 237, 108-121, (1977).
[7] M. Athanassenas, A free boundary problem for capillary surfaces,
Manuscripta Math. 76, 5-19, (1992).
[8] M. Berger and B. Gostiaux, Dierential Geometry, Manifolds curves
and surfaces, Springer-Verlag Graduate Texts 115, (1988).
[9] J. L. Barbosa and M. do Carmo, Stability of hypersurfaces with constant
mean curvature, Math. Z. 185, 339-353 (1984).
[10] C.V. Boys, Soap Bubbles, Dover Publ. Inc. NY 1959 (rst edition 1911).
[11] Carthage, Encyclopedia Britannica, 19th Edition, 1929.
[12] C. Delaunay, Sur la surface de revolution dont la courlure moyenne est
constante, J. Math. Pure et App. 16, 309-321 (1841).
[13] J. Eells, The surfaces of Delaunay, Math. Intelligencer 9, 53-57 (1987).
[14] C. Feerman, The N-body problem in quantum mechanics, Comm. Pure
and Applied Math. Supplement 39, S67-S109 (1986).
[15] J. Foisy, Soap bubble clusters in R2 and R3 , undergraduate thesis,
Williams College (1991).
47
[16] J. Hass, M. Hutchings and R. Schla
y, The Double Bubble Conjecture,
ERA-AMS 1, 98-102 (1995).
[17] J. Hass and R. Schla
y, Double Bubbles Minimize II, (in preparation).
[18] J. Herzberger, Editor Topics in Validated Computations, Elsevier (1994).
[19] H. Hopf, Dierential geometry in the large, Lect. Notes Math., Springer-
Verlag, 1000 (1983).
[20] M. Hutchings, The structure of area-minimizing double bubbles, to ap-
pear in J. Geom. Anal.
[21] K. Kenmotsu, Surfaces of revolution with prescribed mean curvature,
Tohoku Math. J. 32, 147-153 (1980).
[22] Lord Kelvin, Phil. Mag. 24, 503 (1887).
[23] N. Korevaar, R. Kusner and B. Solomon, The structure of complete
embedded surfaces with constant mean curvature, J. Dierential Geom.
30, 465-503 (1989).
[24] O. Lanford, Proof of the Feigenbaum Conjectures, Bull. Amer. Math.
Soc. 6, 427-434 (1982).
[25] H.P. McKean, M. Schreiber and G. H. Weiss, Isoperimetric problem with
application to the gure of cells, J. Math. Physics 6, 479-484 (1994).
[26] R. E. Moore, Methods and Applications of Interval Analysis, SIAM,
1979.
[27] J.D. Moore and R. Schla
y, On equivariant isometric embeddings, Math
Z. 173, 119-129 (1980).
[28] F. Morgan, Soap bubbles in R2 and in surfaces, Pac J Math 165 (1994),
347-361.
[29] F. Morgan, Clusters minimizing area plus length of singular curves,
Math. Ann. 299, 697-714 (1994).
[30] R. Osserman, The isoperimetric inequality, Bull. Amer. Math. Soc. 84,
1182-1238 (1978).
[31] J. Plateau, Statique experimentale et theorique des liquides soumis auz
seules forces molecularies, Gathier-Villars, Paris, 1873.
[32] H. Sagan, Introduction to the Calculus of Variations, Dover, 1992.
[33] M. Spivak, A comprehensive introduction to dierential geometry, Vol-
ume III, Second Edition, Publish or Perish Inc, 1979.
48
[34] J. Taylor, The structure of singularities in soap-bubble-like and soap-
lm-like minimal surfaces, Ann. of Math. 103, 489-539 (1976).
[35] D'arcy Thompson, On growth and form, Cambridge Univ. Press, NY,
1959, (rst edition 1917).
[36] T. I. Vogel, Stability of a liquid drop trapped between two parallel planes,
SIAM J. Appl. Math 47 (1987) 516-525.
Joel Hass
Department of Mathematics
University of California
Davis, CA 95616
e-mail: hass@@math.ucdavis.edu
Roger Schla
y
Real Software
PO Box 1680
Soquel, CA 95073
Internet: rschla
y@@attmail.com
49