A METACALIBRATION PROOF OF THE ISOPERIMETRIC PROBLEM IN n-SPACE AND ON THE n-SPHERE

JAMES DILTS ADVISED BY DR. GARY LAWLOR

Abstract. We solve the isoperimetric problem both in Euclidean n-space and on the surface of a standard n-sphere, using a generalization of calibration techniques which we call metacalibration.

1. Introduction The isoperimetric problem is a classical problem of geometry. The basic question asks, among all gures with the same perimeter (iso-perimetric means same perimeter), which has the greatest area. The original problem is deceptively difcult to prove, and thus escaped a rigorous proof until Weierstrass in 1879 [9]. Isoperimetric problems have a long history. The rst example we have in history is a story given about Dido, Queen of Cartage in Africa. She sailed to northern Africa from Phoenicia, and the locals agreed to give her as much land as she could enclose with the skin of an ox. Being clever, she cut the oxhide into very thin strips, and thereby had enough length to encircle an entire nearby hill. By using the shape of a circle, she maximized her gains from the deal [8]. As with most geometric problems, the ancient Greeks also worked on this problem. The earliest result was given by Zenodorus in a now lost text. Other ancient authors cite him though, letting us know that he proved that the circle has greater area than any polygon with the same perimeter [1]. This is an impressive result, considering the tools available to him. New and powerful tools for optimization problems in general, and thus the isoperimetric problem as well, were discovered with the invention of calculus by Newton and Leibniz. Their work provided one of the rst systematic ways to prove minimization. But one of the main problems with the methods was that it was very dicult to prove that a optimizing gure actually exists. (For example, remember the classic calculus example, f (x) = x3 x, which has a local minimum at 31/2 but has no global minimum.) For instance, Steiner in 1838 gave an argument that seemed to show that the circle maximized area [6]. His argument showed that the maximizer must be both convex and symmetric in all directions. The only gure satisfying this is the circle, but that does not show that the circle is actually the maximizer. It only showed that if such a maximizer did exist, it must be the circle. Weierstrass,
1

JAMES DILTS ADVISED BY DR. GARY LAWLOR

in 1879, was nally able to show that the circle was in fact the true maximizer [9]. Up till recently, most approaches to this type of problem employed indirect methods such as calculus of variations. This method is very similar to the idea of nding an extremum by taking the rst derivative. However, the complexity involved is a serious barrier in many problems. Also, there is the worry of nding a local rather than a global maximum and the diculty of proving that there is in fact a global maximum at all. Direct methods, such as calibration, work dierently. Rather than look at other shapes close to the original gure (like a derivative), it picks the gure it thinks is best and then directly compares it in some way to every other gure we are considering. Calibration cannot handle isoperimetry directly, and so we usually look at the dual problem of nding the gure of constant area with minimized perimeter. Calibration uses a closed vector eld to show that for a measurement function P , P () =

n=

n P ().

where is the proposed minimizer, and is an element in the class of competitors for the problem. (For the isoperimetric problem, this class of competitors is likely the set of manifolds enclosing a particular area, and P measures the total perimeter of the gure.) If this set of inequalities can be set up, then minimizes P . Calibration has been very successful at solving problems involving xed boundaries due to some of the properties of vector elds and their integration, but so far has been stumped by problems of xed area, such as the dual of the isoperimetric problem mentioned above. Metacalibration, developed by Dr. Gary Lawlor of Brigham Young University, employs a similar methodology to calibration but introduces new tools allowing it to solve new types of problems. One such class of problems are what we call equitent problems, which seek to nd the minimum perimeter or surface area to contain a given area and span a boundary. For example, what shape spans the four vertices of a square and contains a certain area? Some work has already been done on these problems [3]. The basis of the metacalibration proof is the attempt to nd a intermediate function G on the gures such that we can set up the set of inequalities P () = G() G() P (). The similarity of metacalibration to standard calibration can be seen by interpreting G as the integral of a vector eld over the gure. Then the equations become P () =

n P ().

A METACALIBRATION PROOF OF THE ISOPERIMETRIC PROBLEM

which is almost exactly the same as calibration. But notice how changes based on each gure. This small, seemingly insignicant change is actually quite powerful. Metacalibration received its name from this similarity. In this paper, we will use metacalibrations to prove that the circle is perimeter minimizing among gures enclosing a given area. We will then use the same method to prove that the n-ball is perimeter minimizing in Rn , though we will omit many details that are the same as in the proof for the circle. This will be followed by the proof of the isoperimetric problem on the surface of a n-sphere, again with some details omitted. 2. Isoperimetric Problem in R2 First we will give the full proof in R2 to introduce the style of proof. To prove that the circle minimizes perimeter, we need only to show the three (in)equalities, namely P () = G() G() P (). Theorem 2.1. In R2 , for a given area a0 , the gure enclosing a0 that minimizes perimeter is the circle of appropriate area. Let C be the set of competitors, namely the set of 1-dimensional piecewise smooth compact manifolds enclosing area a0 . For any C, we set the bottom of the gure to be y = 0 and the top to be y = h. We can parameterize the gure vertically by t in the sense that at each t, there is an intersection between and the line {y = t}. This intersection may be empty. At t, let A(t) be the area of the interior of below the line {y = t}, and let l(t) be the length of intersection of that line with the interior of . We will suppress notation, and leave out the ts. Lemma 2.2. Given an A and l, there exists a unique circle enclosing area A below a line, intersecting that line with length l. Proof. Construct the circle with center (0, y) that passes through the points (l/2, 0) and (l/2, 0). Note that the center of a circle passing through these two points must be on the y-axis. We can look at the area of this circle below the x-axis as a function of y, A(y). As y increases, the part of the new circle below the axis is contained by the part of the old circle below the axis, and so A(y) is strictly decreasing. Since between any two of these circles there is another, A(y) is also continuous. As y goes to innity, A(y) approaches zero. As y goes to negative innity, A(y) approaches innity. Thus, since A(y) is a continuous, monotonic function, for each area a [0, ], there exists an y such that A(y) = a. Since A(y) is monotonic, y is unique. For each t [0, h], we dene (t) to be the gure described in the previous lemma. An example is shown in gure 1. Let r1 be the radius of (t).

JAMES DILTS ADVISED BY DR. GARY LAWLOR

Figure 1. Constructing (t) Let P (t) be the perimeter of below {y = t}, and G(t) be the perimeter of below {y = t}. Finally, let be the angle between the slicing line and , as shown in the gure 1. Lemma 2.3. For all C and all t [0, h], P 2 t 1+ l /2 t
2

Proof. We take a slice of between y = t and y = t + . Since is arbitrarily small, we can consider the boundary of to be straight lines. If there are four or more edges, we can slide the pieces together to reduce the total perimeter by reducing to two edges. Now suppose the edges of are not symmetric. In gure 2, the outer edges, which represent , are not symmetric. We stack the two edges on top of each other, end to end, and then connect the other two endpoints with a third line, again as in gure 2. This line is shorter than the sum of the lengths of the other two as it is a straight line, and equal if and only if the original gure was already symmetric. The lower bound then follows by computing the length of this line. We will be comparing G(t) and P (t) by their derivatives, G and P , and so t t in order to apply the second fundamental theorem of calculus, we need that both of the derivatives are continuous except on a countable set, and that G and P are indeed antiderivatives of the derivatives. The rst is true since is piecewise smooth. The second requires that P and G are continuous, which is not true when there is a at horizontal portion of perimeter at some t, such as at t = 1 2 in gure 3. Lemma 2.4. Any competitor can be rotated in order to remove any such discontinuities.

A METACALIBRATION PROOF OF THE ISOPERIMETRIC PROBLEM

Figure 2. Minimizing

P t

Figure 3. A gure with P (t) discontinuous at t = 1/2 Proof. Separate all at portions of perimeter in into equivalence classes based on the angle between that segment and a given slicing line. There are uncountably many real valued angles possible. If has positive length in each of these equivalence classes, must have innite perimeter, which is thus not a minimizer. This is because the uncountable sum of positive elements is always innite. Thus, there is at least one angle in which has zero perimeter. Then rotate to make one of these angles horizontal in order to remove all discontinuities in P (t). Since A(h) = a0 and l(h) = 0, (h) is a circle of area a0 for all . Thus P () = G() = G(), where is the circle of area a0 since both functions measure the perimeter of the same gure. Thus, in order to prove minimization, we need only show that G() P (), or with our other denitions for G, and P , that G(h) P (h) for all . Proof of Theorem 2.1. By the chain rule, (1) G G l G A = + t l t A t
l t

We will nd three of these partial derivatives, letting

be denoted l .

JAMES DILTS ADVISED BY DR. GARY LAWLOR

First, since A(t) changes as the length of the cross-section of , A =l t Next, a standard nding of dierential geometry gives us G 1 == A r1 where is the mean curvature of the circle , and r1 is the radius. (See, for example, Frank Morgans Riemannian Geometry: A Beginners Guide [4].)

Figure 4. Extending by

to nd

G l

Next, to nd G , we will extend l by , and bend the boundary of in the new l direction, as shown in gure 4. Again by the chain rule, we have that G G G A = + l l A l Since the increase in area (see gure 4) varies as
3 2

3/2 A A = lim lim =0 0 l 0 l In the limit, is also the angle between the slicing line and the modied . We nd that the perimeter is extended by cos (). When we take the limit as 0

A METACALIBRATION PROOF OF THE ISOPERIMETRIC PROBLEM

we nd that G G cos() = lim = lim = cos () 0 l 0 l By replacement, equation 1 becomes G l = cos ()l + t r1 (2) l l = 2 cos () + 2 2r1

Figure 5. Finding sin() A radius of and half of the slicing line form two sides of a triangle, as shown in gure 5. Trigonometry then shows us that l = cos 2r1 2 Thus equation 2 becomes = sin ()

G l = 2 cos () + sin () t 2 (3) = 2(cos (), sin ()) l ,1 2 2 l ,1 2

by the Cauchy-Schwartz inequality. By combining lemma 2.3 and inequality 3 we get G 2 t l ,1 2 =2 1+ l 2

2

P t

JAMES DILTS ADVISED BY DR. GARY LAWLOR

and so
h

G(h) G(0) =
0

G dt t

h 0

P dt = P (h) P (0) t

Since G(0) = P (0) = 0, we have that G(h) P (h) Thus for a given enclosed area, the circle is perimeter minimizing. 3. Isoperimetric Problem in Rn The proof in n dimensions follows the same outline, but of course with dierent details. We will leave out some of the details that are similar. We work in Rn+1 rather than Rn to simplify several coecients and subscripts, but the proof is exactly same. Theorem 3.1. In Rn , for a given n-volume v0 , the gure enclosing v0 that minimizes (n 1)-dimensional perimeter is the n-ball of appropriate volume. From here on, volume means (n + 1)-dimensional volume, area means ndimensional area, and perimeter means n-dimensional surface area. Let C be the set of competitors, namely the set of n-dimensional piecewise smooth compact manifolds enclosing volume v0 . We use the standard x1 , x2 , xn+1 for our basis. For any C, we set the bottom of the gure to be xn+1 = 0, and the top to be xn+1 = h. We can parameterize the gure by t in the sense that at each t, there is an intersection between and the plane {xn+1 = t}. This intersection may be empty. At t, let V be the volume of the interior of below the plane {xn+1 = t}, and let A be the area of intersection of that plane with the interior of . For each t [0, h], we construct as follows. On some plane, pick a n-ball of area A. We call the radius of this ball r2 . Let (t) be the portion of the (n + 1)-ball below that plane that intersects this n-ball and that also contains volume V below that plane. Let r1 be the radius of (t). That (t) is unique up to translation given V and A is shown by a proof very similar to lemma 2.2, so we do not include the proof. Let P (t) be the perimeter of below {xn+1 = t}, and G(t) be the perimeter of below {xn+1 = t}. Finally, let be the incident angle between the slicing plane and the tangent plane to at any point of their intersection. This is analogous to how we picked in the R2 case. Lemma 3.2. For all C and all t [0, h], P n1 nCn r2 t 1 + r22

A METACALIBRATION PROOF OF THE ISOPERIMETRIC PROBLEM

We omit the proof of this lemma. It is similar in principle to lemma 2.3, though n1 it relies on the lower dimensional case. The nCn r2 comes from calculating the surface area for a solid of revolution, where Cn is the volume of a unit (n 1) ball. We will again be comparing G(t) and P (t) by their derivatives, G and P , t t and so in order to apply the second fundamental theorem of calculus, we need that both of the derivatives are continuous except on a countable set, and that G and P are indeed antiderivatives of the derivatives. The rst is true since is piecewise smooth. The second requires that P and G are continuous, which is not true when there is a at horizontal portion of perimeter at some t, as before. The proof that we can rotate our gure to remove such discontinuities is completely equivalent to lemma 2.4, and so we omit it. Since V (h) = v0 and A(h) = 0, (h) is a (n + 1)-ball for all . Thus P () = G() = G(), since both functions measure the perimeter of the same gure. Thus, in order to prove minimization, we need only show that G() P (), or with our other denitions for G, and P , that G(h) P (h) for all . Proof of Theorem 3.1. By the chain rule, (4) G G r2 G V = + t r2 t V t

We will nd three of these partial derivatives, letting r2 be denoted r2 , where r2 t is the radius of the slice of (t). First, since V (t) changes as the area of the cross-section of , V n = A(t) = Cn r2 t where Cn is the volume of an unit (n 1)-ball. Next, a standard nding of dierential geometry gives us n G == V r1 where is the mean curvature of the (n + 1)-ball , and r1 is the radius. Again, see [4]. G Next, to nd r2 , we proceed as we did before in nding G . By extending r2 l instead of l, we extend a whole ring of , instead of just a line. Thus, the width of the ring is the same as the old G . We then calculate the volume of a solid of l revolution to nd that G n1 = cos ()nCn r2 r2 By replacement, equation 4 becomes G n n1 n = cos ()nCn r2 r2 + Cn r2 t r1

10

JAMES DILTS ADVISED BY DR. GARY LAWLOR

n1 cos ()r2 + = nCn r2

r2 r1 A radius of and of the ball in that intersects the slicing plane form two sides of a triangle, as shown in gure 5, though with r2 instead of l/2. Trigonometry then shows us that r2 = cos = sin () r1 2 Thus equation 5 becomes G n1 = nCn r2 (cos ()r2 + sin ()) t (5) (6)
n1 n1 = nCn r2 (cos (), sin ()) (r2 , 1) nCn r2 (r2 , 1)

by the Cauchy-Schwartz inequality. By combining lemma 3.2 and inequality 6 we get G P n1 n1 1 + r22 nCn r2 (r2 , 1) = nCn r2 t t and so h h P G dt dt = P (h) P (0) G(h) G(0) = 0 t 0 t Since G(0) = P (0) = 0, we have that G(h) P (h) Thus for a given enclosed volume, the (n + 1)-ball is perimeter minimizing. 4. Isoperimetric Problem on a n-Sphere The general principle of this proof is the same as the previous ones, but there are some new pieces to the proof. As before, we will omit some details that are very similar to the previous section. Theorem 4.1. On the surface of a standard n-sphere, for a given n-dimensional volume v0 , the gure enclosing v0 that minimizes perimeter is an n-dimensional geodesic ball. In this section, volume refers to n-dimensional volume, and area refers to (n1) dimensional volume, and perimeter refers to (n 1)-dimensional boundary of a manifold. Let C be the set of competitors, namely the set of (n1)-dimensional piecewise smooth compact manifolds enclosing volume v0 on the surface of a n-sphere. We use standard spherical coordinates, (1 , 2 , n1 , ). For any C, we can parameterize the gure by t in the sense that at each t, there is an intersection between and the line { = t}. This intersection may be empty. We call such lines lines. At t, let V be the volume of below the line { = t}, and let A be the area of intersection of that line with the interior of .

A METACALIBRATION PROOF OF THE ISOPERIMETRIC PROBLEM

11

For each t [0, ], we construct as follows. Along the line { = t}, pick a (n 1) ball of area A. We call the geodesic radius of this ball r3 . If we consider this ball to be in Rn+1 , we can consider the at radius of this ball as well. Call this at radius r2 . Let (t) be the portion of the n ball below that line that intersects this (n 1) ball and that also contains volume V below the line. That (t) is unique up to translation along the line given V and A is shown by a proof again very similar to lemma 2.2. Let P (t) be the perimeter of below { = t}, and G(t) be the perimeter of below { = t}. Finally, let be the incident angle between the line and the tangent plane to at any point of their intersection. This is again analogous to how we picked in the R2 isoperimetric problem. Many of these quantities are shown in gure 6 for the case where n = 3. It is of course very dicult to draw a 4-dimensional sphere on paper, and so this shows just a piece of the problem. For n = 3, the lines are 3 balls, or spheres, as is .

Figure 6. A slice of a sphere on a 4 ball Lemma 4.2. For all C and all t (0, ), P n1 nCn r2 1 + (r3 e)2 t where e is the the change in r3 due to the natural expansion and contraction of lines as increases. Again, we omit the proof of this lemma. It is almost the same as the omitted n1 proof for lemma 3.2, and again relies on the lower dimensional case. The nCn r2 is similarly from calculating the surface area for a solid of revolution, where Cn is the volume of a unit (n 1) ball. We will again be comparing G(t) and P (t) by their derivatives, G and P , t t and so in order to apply the second fundamental theorem of calculus, we need

12

JAMES DILTS ADVISED BY DR. GARY LAWLOR

that both of the derivatives are continuous except on a countable set, and that G and P are indeed antiderivatives of the derivatives. The rst is true since is piecewise smooth. The second requires that P and G are continuous, which is not true when there is a at horizontal portion of perimeter at some t, as before. By a proof very similar to lemma 2.4, we can move the pole of our lines to remove such discontinuities. Since V () = v0 and A() = 0, () is a n-ball for all . Thus P () = G() = G(), since both functions measure the perimeter of the same gure. Thus, in order to prove minimization, we need only show that G() P (), or with our other denitions for G, and P , that G() P () for all . Proof of Theorem 3.1. By the chain rule, (7) dG G d G dr3 G dV = + + dt dt r3 dt V dt

We will set t, and so d = 1. dt We will nd two of these partial derivatives, letting dr3 be denoted r3 . The dt last term here will cancel later, and so we will leave it as it is. G G For r3 , consider a similar gure to that for r2 in the previous section. Very similarly, we can nd the change in perimeter, and take a limit, giving us that G n1 = cos ()nCn r2 r3 Next, for G , instead of calculating directly, we will consider a dierent gure. Create (t), but instead of changing it as usual, extend the n-ball innitesimally, as in gure 7. Another way to think of this would be to have the competitor actually be the n-ball. Call this new extended gure , and its perimeter P (t).

Figure 7. An extension of This gives us a new application of the chain rule. However, several pieces are the same as before, as they are true for all n-balls on a n-sphere. Thus we get (8) dP G G dV n1 dr3 = + cos ()nCn r2 + dt dt V dt

A METACALIBRATION PROOF OF THE ISOPERIMETRIC PROBLEM

13

G G where V dV is the same in both pictures because V is the same for all such dt gures, and dV is just A. Note that G and dr3 are in fact dierent here than dt dt before, but since they will only be used briey, we will use the same notation. n1 If we set t = , then we nd that P = sin() nCn r2 and r3 = (cot()+e, where e represents the change in r3 due to the natural expansion of lines on a n-sphere. This is illustrated in gure 8. If we then take the limit as 0, we nd that dP P 1 n1 = lim = nCn r2 0 dt sin () dr3 r3 = lim = cot () + e 0 dt

Figure 8. Finding

dP dt

and

dr3 dt

Using those equations, and rearranging equation 8 we get 1 G dV G n1 = nCn r2 ( cos ()(cot () + e)) sin () V dt 1 cos2 () G dV ) cos ()e) sin () V dt G dV n1 = nCn r2 (sin () cos ()e) V dt in equation 7, we get
n1 = nCn r2 ((

Replacing

G dV G dV dG n1 n1 = nCn r2 (sin () cos ()e) + nCn r2 cos ()r3 + dt V dt V dt n1 = nCn r2 (sin () + cos ()(r3 e)) If we split up this into two vectors and use the Cauchy-Schwartz inequality we nd an upper bound for dG dt dG n1 = nCn r2 (cos (), sin ()) (r3 e, 1) dt

14

JAMES DILTS ADVISED BY DR. GARY LAWLOR

n1 nCn r2 (r3 e, 1)

(9)

By combining lemma 4.2 and inequality 9, we get dP dG n1 n1 1 + (r3 e)2 nCn r2 (r3 e, 1) = nCn r2 dt dt and so dG dP G() G(0) = dt dt = P () P (0) 0 dt 0 dt Since G(0) = P (0) = 0, we have that G() P () Thus for a given volume on the surface of a n-sphere, the n-ball is perimeter minimizing. 5. Further Research This paper has presented a proof for an already solved problem, but by the new and powerful method of metacalibration. The results prove may not be dicult to solve with traditional methods, but metacalibration has the potential to solve previously unsolved problems. The isoperimetric problem in more generalized symmetric spaces has been studied, but only a few results have been reached [2] [7]. We think that the method presented in this paper can and will be successful in this area. Metacalibration has also been useful in solving other minimization problems, such as bubbles trapped in tetrahedral soap lms [3], the double bubble in Rn [5], and others. A natural next step would be to attempt to prove that the standard triple bubble in the plane is perimeter minimizing, which we are currently working on. References
[1] Viktor Blasjo. The evolution of the isoperimetric problem. The American Mathematical Monthly, 112:526566, 2005. [2] WT. Hsiang. On the uniqueness of isoperimetric solutions and imbedded soap bubbles in non-compact symmetric spaces. Inventiones Mathematicae, 98(1):3958, 1989. [3] Drew Johnson. The tetrahedral soap lm with trapped bubble. Preprint. [4] Frank Morgan. Riemannian Geometry: A Beginners Guide. A K Peters, 1998. [5] Donald Sampson and James Dilts. Metacalibration proof of the double bubble conjecture in Rn . Preprint. [6] Jakob Steiner. Einfache beweise der isoperimetrischen hauptstze. J. Reine Angew. Math., a 18:281296, 1838. [7] T. Thies. A sharp isoperimetric inequality for rank one symmetric spaces. Manuscripta Mathematica, 111(1):97104, 2003. [8] Sir William Thomson. Popular Lectures and Addresses, volume 2. MacMillan and Co., 1894. [9] Karl Weierstrass. Mathematische Werke, volume 7. Mayer and Mller, 1927. u