geometric measure theory

Preface

Singular geometry governs the physical universe: soap bubble clusters meet-ing along singular curves, black holes, defects in materials, chaotic turbulence,crystal growth. The governing principle is often some kind of energy mini-mization. Geometric measure theory provides a general framework for under-standing such minimal shapes, a priori allowing any imaginable singularityand then proving that only certain kinds of structures occur.

Jean Taylor used new tools of geometric measure theory to derive thesingular structure of soap bubble clusters and sea creatures, recorded byJ. Plateau over a century ago (see Section 13.9). R. Schoen and S.-T. Yauused minimal surfaces in their original proof of the positive mass conjec-ture in cosmology, recently extended to a proof of the Riemannian PenroseConjecture by H. Bray. David Hoffman and his collaborators used moderncomputer technology to discover some of the first new complete embeddedminimal surfaces in a hundred years (Figure 6.1.3), some of which look justlike certain polymers. Other mathematicians are now investigating singulardynamics, such as crystal growth. New software computes crystals growingamidst swirling fluids and temperatures, as well as bubbles in equilibrium, ason the front cover of this book. (See Section 16.8.)

This year 2000, Hutchings, Morgan, Ritore, and Ros announced a proof ofthe Double Bubble Conjecture, which says that the familiar double soap bubbleprovides the least-area way to enclose and separate two given volumes of air.The planar case was proved by the 1990 Williams College NSF “SMALL”undergraduate research Geometry Group [Foisy et al.]. The case of equalvolumes in R3 was proved by Hass, Hutchings, and Schlafly with the help ofcomputers in 1995. The general R3 proof has now been generalized to R4 by

vii

viii Preface

the 1999 Geometry Group [Reichardt et al.]. The whole story appears in printfor the first time here in Chapters 13 and 14.

This little book provides the newcomer or graduate student with an illus-trated introduction to geometric measure theory: the basic ideas, terminology,and results. It developed from my one-semester course at MIT for graduatestudents with a semester of graduate real analysis behind them. I have includeda few fundamental arguments and a superficial discussion of the regularitytheory, but my goal is merely to introduce the subject and make the standardtext, Geometric Measure Theory by H. Federer, more accessible.

Other good references include L. Simon’s Lectures on Geometric MeasureTheory, E. Guisti’s Minimal Surfaces and Functions of Bounded Variation,R. Hardt and Simon’s Seminar on Geometric Measure Theory, Simon’sSurvey Lectures on Minimal Submanifolds, J. C. C. Nitsche’s Lectures onMinimal Surfaces (now available in English), R. Osserman’s updated Surveyof Minimal Surfaces, H. B. Lawson’s Lectures on Minimal Submanifolds,and A. T. Fomenko’s books on The Plateau Problem. S. Hildebrandt and A.Tromba offer a beautiful popular gift book for your friends, reviewed byMorgan [14, 15]. J. Brothers and also Sullivan and Morgan assembled listsof open problems. There is an excellent Questions and Answers about AreaMinimizing Surfaces and Geometric Measure Theory by F. Almgren [4], whoalso wrote a review [5] of the first edition of this book. The easiest startingplace may be the Monthly article “What is a Surface?” [Morgan 24].

It was from Fred Almgren, whose geometric perspective this book attemptsto capture and share, that I first learned geometric measure theory. I thankmany graduate students for their interest and suggestions, especially BennyCheng, Gary Lawlor, Robert McIntosh, Mohamed Messaoudene, and MartyRoss. I also thank typists Lisa Court, Louis Kevitt, and Marissa Barschdorf.Jim Bredt first illustrated an article of mine as a member of the staff of Link,a one-time MIT student newspaper. I feel very fortunate to have him with meagain on this book. I am grateful for help from many friends, notably TimMurdoch, Yoshi Giga and his students, who prepared the Japanese translation,and especially John M. Sullivan. I would like to thank my new editor, RobertRoss, and my original editor and friend Klaus Peters. A final thank you goesto all who contributed to this book at MIT, Rice, Stanford, and Williams.Some support was provided by National Science Foundation grants, by myCecil and Ida Green Career Development Chair at MIT, and by my DennisMeenan Third Century Chair at Williams.

This third edition includes updated material and references and four newChapters 14–17, including the first appearance in print of sketches of the recentproofs of the Double Bubble and Hexagonal Honeycomb conjectures, andother recent results on immiscible fluids, flows, and isoperimetric inequalities.

Preface ix

Bibliographic references are simply by author’s name, sometimes with anidentifying numeral or section reference in brackets. Following a useful prac-tice of Nitsche [2], the bibliography includes cross-references to each citation.

Frank MorganWilliamstown, [email protected]

CHAPTER 1

Geometric Measure Theory

Geometric measure theory could be described as differential geometry,generalized through measure theory to deal with maps and surfaces that are notnecessarily smooth, and applied to the calculus of variations. It dates from the1960 foundational paper of Herbert Federer and Wendell Fleming on “Normaland Integral Currents,” recognized by the 1986 AMS Steele Prize for a paperof fundamental or lasting importance, and earlier and contemporaneous workof L. C. Young [1, 2], E. De Giorgi [1, 3], and E. R. Reifenberg [1–3] (seeFigure 1.0.1). This chapter will give a rough outline of the purpose and basicconcepts of geometric measure theory. Later chapters will take up these topicsmore carefully.

1.1. Archetypical Problem. Given a boundary in Rn, find the surface ofleast area with that boundary. See Figure 1.1.1. Progress on this problemdepends crucially on first finding a good space of surfaces to work in.

1.2. Surfaces as Mappings. Classically, one considered only two-dimensio-nal surfaces, defined as mappings of the disc. See Figure 1.2.1. Excellentreferences include J. C. C. Nitsche’s Lectures on Minimal Surfaces [2], nowavailable in English, R. Osserman’s updated Survey of Minimal Surfaces, andH. B. Lawson’s Lectures on Minimal Submanifolds. It was not until about1930 that J. Douglas and T. Rado surmounted substantial inherent difficultiesto prove that every smooth Jordan curve bounds a disc of least mappingarea. Almost no progress was made for higher-dimensional surfaces (untilin a surprising turnaround B. White [1] showed that for higher-dimensional

1

2 Geometric Measure Theory

Figure 1.0.1. Wendell Fleming, Fred Almgren, and Ennio De Giorgi, three of thefounders of geometric measure theory, at the Scuola Normale Superiore, Pisa, summer,1965; and Fleming today. Photographs courtesy of Fleming.

Figure 1.1.1. The surface of least area bounded by two given Jordan curves.

Figure 1.2.1. Surface realized as a mapping, f, of the disc.

Geometric Measure Theory 3

surfaces the geometric measure theory solution actually solves the mappingproblem too).

Along with its successes and advantages, the definition of a surface as amapping has certain drawbacks (see Morgan [24]):

(1) There is an inevitable a priori restriction on the types of singularities thatcan occur;

(2) There is an a priori restriction on the topological complexity; and

(3) The natural topology lacks compactness properties.

The importance of compactness properties appears in the direct method descri-bed in the next section.

1.3. The Direct Method. The direct method for finding a surface of leastarea with a given boundary has three steps.

(1) Take a sequence of surfaces with areas decreasing to the infimum.(2) Extract a convergent subsequence.(3) Show that the limit surface is the desired surface of least area.

Figures 1.3.1–1.3.4 show how this method breaks down for lack of compact-ness in the space of surfaces as mappings, even when the given boundary isthe unit circle. By sending out thin tentacles toward every rational point, thesequence could include all of R3 in its closure!

1.4. Rectifiable Currents. An alternative to surfaces as mappings is provi-ded by rectifiable currents, the m-dimensional, oriented surfaces of geometricmeasure theory. The relevant functions f: Rm ! Rn need not be smooth, butmerely Lipschitz, i.e.,

jf�x� � f�y�j � Cjx � yj,

for some “Lipschitz constant” C.

Figure 1.3.1. A surface with area C 1.


Figure 1.3.2. A surface with area C 14 .


Fortunately there is a good m-dimensional measure on Rn, called Hausdorffmeasure, H

m. Hausdorff measure agrees with the classical mapping area ofan embedded manifold, but it is defined for all subsets of Rn.

A Borel subset B of Rn is called (Hm, m) rectifiable if B is a countable union

of Lipschitz images of bounded subsets of Rm, with Hm�B� < 1. (As usual,

we will ignore sets of Hm measure 0.) That definition sounds rather general,

and it includes just about any “m-dimensional surface” I can imagine. Never-theless, these sets will support a kind of differential geometry; for example, itturns out that a rectifiable set B has a canonical tangent plane at almost everypoint.

Finally, a rectifiable current is an oriented rectifiable set with integer multi-plicities, finite area, and compact support. By general measure theory, one canintegrate a smooth differential form ϕ over such an oriented rectifiable set S,and hence view S as a current, i.e., a linear functional on differential forms,

ϕ 7!∫

Sϕ.

Geometric Measure Theory 5


This perspective yields a new natural topology on the space of surfaces, dualto an appropriate topology on differential forms. This topology has usefulcompactness properties, given by the fundamental Compactness Theorembelow. Viewing rectifiable sets as currents also provides a boundary operator∂ from m-dimensional rectifiable currents to (m � 1)-dimensional currents,


defined by�∂S��ϕ� D S�dϕ�,

where dϕ is the exterior derivative of ϕ. By Stokes’s Theorem, this definitioncoincides with the usual notion of boundary for smooth, compact manifoldswith boundary. In general, the current ∂S is not rectifiable, even if S is recti-fiable.

1.5. The Compactness Theorem. Let c be a positive constant. Then the setof all m-dimensional rectifiable currents T in a fixed large closed ball in Rn,such that the boundary ∂T is also rectifiable and such that the area of both Tand ∂T are bounded by c, is compact in an appropriate weak topology.

1.6. Advantages of Rectifiable Currents. Notice that rectifiable currentshave none of the three drawbacks mentioned in Section 1.2. There is certainlyno restriction on singularities or topological complexity. Moreover, the com-pactness theorem provides the ideal compactness properties. In fact, the directmethod described in Section 1.3 succeeds in the context of rectifiable currents.In the figures of Section 1.3, the amount of area in the tentacles goes to 0.Therefore, they disappear in the limit in the new topology. What remains isthe disc, the desired solution.

All of these results hold in all dimensions and codimensions.

1.7. The Regularity of Area-Minimizing Rectifiable Currents. One seri-ous suspicion hangs over this new space of surfaces: The solutions they provideto the problem of least area, the so-called area-minimizing rectifiable currents,may be generalized objects without any geometric significance. The followinginterior regularity results allay such concerns. (We give more precise state-ments in Chapter 8.)

(1) A two-dimensional area-minimizing rectifiable current in R3 is a smoothembedded manifold.

(2) For m � 6, an m-dimensional area-minimizing rectifiable current in RmC1

is a smooth embedded manifold.

Thus in low dimensions the area-minimizing hypersurfaces provided bygeometric measure theory actually turn out to be smooth embedded mani-folds. However, in higher dimensions, singularities occur, for geometric andnot merely technical reasons (see Section 10.7). Despite marked progress,understanding such singularities remains a tremendous challenge.

CHAPTER 2

Measures

This chapter lays the measure-theoretic foundation, including the definitionof Hausdorff measure and covering theory. The general reference is Federer [1,Chapter II].

2.1. Definitions. For us a measure � on Rn will be what is sometimes calledan outer measure: a nonnegative function � on all subsets of Rn (with thevalue C 1 allowed, of course), which is countably subadditive, i.e., if A iscontained in a countable union, [ Ai, then

��A� � ��Ai�.

A set A ² Rn is called measurable if, for all E ² Rn, ��E \ A�C��E \ AC� D ��E�. The class of measurable sets is a �-algebra, i.e., closedunder complementation, countable union, and countable intersection. If A is acountable disjoint union of measurable sets Ai, then ��A� D ��Ai�.

The smallest �-algebra containing all open sets is the collection of Borelsets. A measure � is called Borel regular if Borel sets are measurable andevery subset of Rn is contained in a Borel set of the same measure.

Suppose that � is Borel regular, A is measurable, and ε > 0. If ��A� < 1,then A contains a closed subset C with ��A �C� < ε. If A can be covered bycountably many open sets of finite measure, then A is contained in an openset W with ��W � A� < ε [Federer, 2.2.3].

All Borel sets are measurable if and only if Caratheodory’s criterion holds:(1) Whenever A1, A2 are sets a positive distance apart, then

��A1 [ A2� D ��A1�C ��A2�.

7


2.2. Lebesgue Measure. There is a unique Borel regular, translation invari-ant measure on Rn such that the measure of the unit cube [0, 1]n is 1. Thismeasure is called Lebesgue measure, L

n.

2.3. Hausdorff Measure [Federer, 2.10]. Unfortunately, for general “m-dimensional” subsets of Rn (for m < n), it is more difficult to assign anm-dimensional measure. The m-dimensional area of a C1 map f from adomain D ² Rm into Rn is classically defined as the integral of the JacobianJmf over D. [Computationally, at each point x 2 D, �Jmf�2 equals the sum ofthe squares of the determinants of the m ð m submatrices of Df�x� or, equiv-alently, the determinant of �Df�x��tDf�x�.] The area of an m-dimensionalsubmanifold M of Rn is then defined by calculating it on parameterizedportions of M and proving that the area is independent of choice of parame-terization.

In 1918, F. Hausdorff introduced an m-dimensional measure in Rn whichgives the same area for submanifolds, but is defined on all subsets of Rn.When m D n, it turns out to be equal Lebesgue measure.

DEFINITIONS. For any subset S of Rn, define the diameter of S

diam�S� D supfjx � yj: x, y 2 Sg.

Let ˛m denote the Lebesgue measure of the closed unit ball Bm�0, 1� ² Rm.For A ² Rn, we define the m-dimensional Hausdorff measure H

m�A� by thefollowing process. For small υ, cover A efficiently by countably many setsSj with diam�Sj� � υ, add up all the ˛m�diam�Sj�/2�m, and take the limit asυ ! 0:

Hm�A� D lim

υ!0infA²[Sj

diam�Sj��υ

˛m

(diam�Sj�

2

)m

.

The infimum is taken over all countable coverings fSjg of A whose membershave diameter at most υ. As υ decreases, the more restricted infimum cannotdecrease, and hence the limit exists, with 0 � H

m�A� � 1. In Figure 2.3.1,the two-dimensional area is approximated by �r2. The spiral of Figure 2.3.2illustrates one reason for taking the limit as υ ! 0, since otherwise a spiralof great length could be covered by a single ball of radius 1.

Countable subadditivity follows immediately from the definition. The mea-surability of Borel sets follows easily from Caratheodory’s criterion 2.1(1).

To see that each A ² Rn is contained in a Borel set B of the same measure,note first that each Sj occurring in the definition of H

m�A� may be replaced byits closure, so that [ Sj is Borel. If fS�k�j g is a countable sequence of coverings

Measures 9

Figure 2.3.1. The Hausdorff measure (area) of a piece of surface A is approximatedby the cross-sections of little balls which cover it.

Figure 2.3.2. One must cover by small sets to compute length accurately. Here thelength of the spiral is well estimated by the sum of the diameters of the tiny balls, butgrossly underestimated by the diameter of the huge ball.

defining Hm�A�, then B D \k [j S

�k�j gives the desired Borel set. Therefore,

Hm is Borel regular. Later it will be proved that H

m gives the “correct” areafor C1 submanifolds of Rn.

The definition of Hausdorff measure extends to any nonnegative realdimension. [The definition of ˛m is extended by the function: ˛m D�m/2/�m/2 C 1�]. Notice that H

0 is counting measure; H0�A� is the number

of elements of A.


The Hausdorff dimension of a nonempty set A is defined as

inffm ½ 0: Hm�A� < 1g D inffm: H

m�A� D 0gD supfm: H

m�A� > 0gD supfm: H

m�A� D 1g.

The equivalence of these conditions follows from the fact that if m < kand H

m�A� < 1, then Hk�A� D 0 (Exercise 2.4). The Cantor set of Exer-

cise 2.6 turns out to have Hausdorff dimension ln 2/ ln 3. Figure 2.3.3 picturesa Cantor-like set in R3, called the Sierpinski sponge, which has Hausdorffdimension of about 2.7.

These Cantor-like sets are self-similar in the sense that certain homotheticexpansions of such a set are locally identical to the original set. Self-similarityappears in the coastline of Great Britain and in the mass in the universe.B. Mandelbrot has modeled many natural phenomena by random fractionaldimensional sets and processes, called fractals. His books, Fractals and The

Figure 2.3.3. The Sierpinski sponge is an example of a fractional dimensional set. ItsHausdorff dimension is about 2.7. (From Studies in Geometry by Leonard M. Blumen-thal and Karl Menger. Copyright 1979 by W. H. Freeman and Company. Reprintedwith permission.)

Measures 11

Fractal Geometry of Nature, contain beautiful, computer-generated pictures ofhypothetical clouds, landscapes, and other phenomena.

2.4. Integralgeometric Measure. In 1932, J. Favard defined another m-dimensional measure on Rn�m D 0, 1, . . . , n�, now called integral-geometricmeasure, I

m. It turns out that Im agrees with H

m on all smooth m-dimensio-nal submanifolds and other nice sets, but disagrees and often is zero onCantor-like sets.

Roughly, to define Im�A�, project A onto an m-dimensional subspace of

Rn, take the Lebesgue measure (counting multiplicities), and average over allsuch projections.

More precisely, let OŁ�n,m� denote the set of orthogonal projections pof Rn onto m-dimensional subspaces. For general reasons there is a uniquemeasure on OŁ�n, m� invariant under Euclidean motions on Rn, normalized tohave total measure 1. For example, the set OŁ�2, 1� of orthogonal projectionsonto lines through 0 in the plane may be parameterized by 0 � $ < �, andthe unique measure is �1/�� d$. For y 2 image p ¾D Rm, let the “multiplicityfunction,” N�pjA, y�, denote the number of points in A \ p�1�y�. Define anormalizing constant,

ˇ�n, m� D

(m C 1

2

)

(n� m C 1

2

)

(nC 1

2

)�1

��1/2.

Now define the integralgeometric measure of any Borel set B by

Im�B� D 1

ˇ�n,m�

∫p2OŁ�n,m�

∫y 2 imagep¾D Rm

N�pjB, y� d Lmy dp.

One checks that the function N�pjB, y� is indeed measurable and that Im

is countably subadditive. Finally extend Im to a Borel regular measure by

defining for any set A ² Rn,

Im�A� D inffI

m�B�:A ² B, B Borelg.

2.5. Densities [Federer, 2.9.12, 2.10.19]. Let A be a subset of Rn. For 1 �m � n, a 2 Rn, we define the m-dimensional density m�A, a� of A at a bythe formula

m�A, a� D limr!0

Hm�A \ Bn�a, r��

˛mrm,

where ˛m is the measure of the closed unit ball Bm�0, 1� in Rm. For example,the cone

C D fx2 C y2 D z2g


Figure 2.5.1. The cone fx2 C y2 D z2g has density 1 everywhere except at the vertex,where it has density

p2.

of Figure 2.5.1 has two-dimensional density

2�C, a� D

1 for a 2 C � f0g,0 for a /2 C,p

2 for a D 0.

Similarly, for � a measure on Rn, 1 � m � n, a 2 Rn, define the m-dimen-sional density m��, a� of � at a by

m��, a� D limr!0

��Bn�a, r��

˛mrm.

Note that for any subset A of Rn, m�A, a� D m�Hm A, a�, where H

m Ais the measure defined by

�Hm A��E� � H

m�A \ E�.

Hence density of measures actually generalizes the notion of density of sets.

2.6. Approximate Limits [Federer, 2.9.12]. Let A ² Rm. A function f:A ! Rn has approximate limit y at a if for every ε > 0, Rm � fx 2 A: jf�x��yj < εg has m-dimensional density 0 at a. We write y D ap limx!a f�x�. Notethat in particular A must have density 1 at a.

Measures 13

PROPOSITION. A function f:A ! Rn has an approximate limit y at a if andonly if there is a set B ² A such that BC has m-dimensional density 0 at a andfjB has the limit y at a.

Remark. In general, the word approximate means “except for a set ofdensity 0.”

Proof. The condition is clearly sufficient. To prove necessity, assume thatf has an approximate limit y at a. For convenience we assume y D 0. Thenfor any positive integer i,

Ai � Rm � fx 2 A: jf�x�j < 1/ig

has density 0 at a. Choose r1 > r2 > . . . such that

Hm�Ai \ Bn�a, r��

˛mrm� 2�i

whenever 0 < r � ri. Notice that A1 ² A2 ² . . .. Let BC D [�Ai \ B�a, ri��.Clearly fjB has the limit y at a. To show that BC had density 0 at a, let

ri > s > riC1. Then

Hm�BC \ B�a, s�� H

m�Ai \ B�a, s�� C Hm�AiC1 \ B�a, riC1��

C Hm�AiC2 \ B�a, riC2�� C Ð Ð Ð

� ˛m�sm Ð 2�i C rmiC1 Ð 2��iC1� C rmiC2 Ð 2��iC2� C Ð Ð Ð�

� ˛m Ð sm Ð 2��i�1�.

Therefore BC has density 0 at a, as desired.

DEFINITIONS. Let a 2 A ² Rm. A function f:A ! Rn is approximatelycontinuous at a if f�a� D ap limx!a f�x�. The point a is a Lebesgue point off if m�AC, a� D 0 and

1

˛mrm

∫A\B�a,r�

jf�x�� f�a�j d Lmx ��!

r ! 00.

The function f is approximately differentiable at a if there is a linear func-tion L: Rm ! Rn such that

ap limx!a

jf�x�� f�a� � L�x � a�jjx � aj D 0.

We write L D apDf�a�.


The following covering theorem of Besicovitch proves more powerful inpractice than more familiar ones, such as Vitali’s. It applies to any finite Borelmeasure ϕ.

2.7. Besicovitch Covering Theorem [Federer, 2.8.15; Besicovitch]. Sup-pose ϕ is a Borel measure on Rn, A ² Rn, ϕ�A� < 1, F is a collection ofnontrivial closed balls, and inffr: B�a, r� 2 Fg D 0 for all a 2 A. Then there isa �countable� disjoint subcollection of F that covers ϕ almost all of A.

Partial Proof. We may assume that all balls in F have radius at most 1.

PART 1. There is a constant .�n� such that, given a closed ball, B, ofradius r and a collection, C, of closed balls of a radius of at least r whichintersect B and which do not contain each other’s centers, then the cardinalityof C is at most .�n�. This statement is geometrically obvious, and we omitthe proof. E. R. Riefenberg [4] proved that for n D 2, the sharp bound is 18.(See Figure 2.7.1.)

PART 2. . C 1 subcollections of disjoint balls cover A. To prove this state-ment, we will arrange the balls of F in rows of disjoint balls, starting with

Figure 2.7.1. At most, 18 larger discs can intersect the unit disc in R2 withoutcontaining each other’s centers. Figure courtesy of J. M. Sullivan [2].

Measures 15

the largest and proceeding in order of size. (Of course, there may not alwaysbe a “largest ball,” and actually one chooses a nearly largest ball. This tech-nical point propagates minor corrections throughout the proof, which we willignore.)

Place the largest ball B1 in the first row. (See Figure 2.7.2.) Throw awayall balls whose centers are covered by B1.

Take the next largest ball, B2. If B2 is disjoint from B1, place B2 in the firstrow. If not, place B2 in the second row. Throw away all balls whose centersare covered by B2.

At the nth step, place Bn in the earliest row that keeps all balls in each rowdisjoint. Throw away all balls whose centers are covered by Bn.

Proceed by transfinite induction. The whole list certainly covers A, since wethrow away only balls whose centers are already covered. Each row consistsof disjoint balls, by construction. Hence it suffices to show that there are atmost . C 1 nonempty rows. Assume some ball, B, gets put in the . C 2 row.Then there are balls D1, . . . , D.C1, at least as large as B already in the first. C 1 rows and not disjoint from B. No Dj can contain another’s center, orthe smaller would have been thrown away when the larger was put in. Thiscontradiction of Part 1 completes the proof of Part 2.

PART 3. Completion of proof. By Part 2, some disjoint subcollection covers1/�. C 1� the ϕ measure of A. Hence some finite disjoint subcollection covers

1.

2.

3.

. . .

.

ζ+1

Figure 2.7.2. In the proof of the Besicovitch covering Theorem 2.7, the balls cover-ing A are arranged by size in rows or discarded. Intersecting balls must go in differentrows. For the case of R2, this requires at most .�2� C 1 � 19 rows. Then some rowmust provide a disjoint cover of at least 1/19 the total measure.


a closed subset A1 ² A with

ϕ�A1�

ϕ�A�½ 1

. C 2, i.e., 1 � ϕ�A1�

ϕ�A�� υ < 1.

Repeat the whole process on A � A1 with the balls contained in A � A1 toobtain a finite disjoint subcollection covering A2 ² A with

1 � ϕ�A2�

ϕ�A�� υ2.

Countably many such repetitions finally yield a countable disjoint subcollec-tion covering ϕ almost all of A.

We now give three corollaries as examples of the usefulness of Besicovitch’scovering theorem.

2.8. Corollary. Hn D L

n on Rn.

Proof. We will need the so-called isodiametric inequality or Bieberbachinequality, which says that among all sets of fixed diameter, the ball has thelargest volume. In other words, for any set S in Rn,

Ln�S� � ˛n

(diam S

2

)n

.

It follows immediately that Hn ½ L

n.There happens to be an easy proof of the isodiametric inequality. (See also

Burago and Zalgaller, Theorem 11.2.1.) We may assume that S is symmetricwith respect to each coordinate axis, since replacing each intersection of Swith a line parallel to the axis by a symmetric interval of the same one-dimensional measure does not change the Lebesgue measure and can onlydecrease the diameter. But now S is symmetric with respect to the origin andhence is contained in the ball B of the same diameter. Therefore

Ln�S� � L

n�B� D ˛n

(diam S

2

)n

,

as desired. Notice that the symmetrization step is necessary, because anequilateral triangle, for example, is not contained in a ball of the samediameter.

To prove that Hn � L

n, we will use the Besicovitch Covering Theorem.First, we note that it suffices to prove that H

n�A� � Ln�A� for A Borel and

bounded, or hence for A equal to the open R-ball Un�0, R� ² Rn, or hence

Measures 17

for A D Un�0,1�. An easy computation shows that Hn�A� < 1. Given ε > 0,

choose υ > 0 such that

�1� Hn�A� � inf

{˛n Ð

(diam Si

2

)n

:A ² [ Si, diam Si � υ

}C ε.

Apply the covering theorem with

F D fclosed balls contained in A with diameter � υg

to obtain a disjoint covering G of B ² A with Hn�A � B� D 0. Let G0 be a

covering by balls of diameter at most υ of A � B with

∑S2G0

˛n

(diam S

2

)n

� ε.

Then G [G0 covers A, and therefore

Hn�A� �

∑S2G[G0

˛n

(diam S

2

)n

C ε

�∑S2G

Ln�S�C

∑S2G0

˛n

(diam S

2

)n

C ε

� Ln�A� C ε C ε.

The corollary is proved. The fussing with A � B at the end was necessarybecause (1) does not apply to B.

2.9. Corollary. If A ² Rn is Lebesgue measurable, then the density n�A, x�equals the characteristic function 1A�x� almost everywhere.

Proof. It suffices to show that for every measurable set A, �A, x� D 1 atalmost all points x 2 A. (Considering AC then implies �A, x� D 0 at almostall x /2 A.) Assume not. We may assume 0 < L

n�A� < 1. We may furtherassume that for some υ < 1

�1� Ł�A, a� D limL

n�A \ B�a, r��˛nrn

< υ for all a 2 A,

by first choosing υ such that

Lnfa 2 A:Ł�A, a� < υg > 0


and then replacing A by fa 2 A:Ł�A, a� < υg. Choose an open set U ¦ Asuch that

�2� Ln�A� > υL

n�U�.

Let F be the collection of all closed balls B centered in A and contained inU such that

Ln�A \ B� < υL

n�B�.

By (1), F contains arbitrarily small balls centered at each point of A. Bythe covering theorem, there is a countable disjoint subcollection G coveringalmost all of A. Therefore,

Ln�A� < υ

∑S2G

˛n

(diam S

2

)n

� υLn�U�.

This contradiction of (2) proves the corollary.

2.10. Corollary. A measurable functionf: Rn ! R is approximately contin-uous almost everywhere.

Corollary 2.10 follows rather easily from Corollary 2.9. Exercise 2.9 givessome hints on the proof.

EXERCISES

2.1. Let I be the line segment in R2 from (0, 0) to (1, 0). Compute I1�I� directly.

(ˇ�2, 1� D 2/�.)2.2. Let I be the unit interval [0, 1] in R1. Prove that H

1�I� D 1.2.3. Prove that H

n�Bn�0,1�� < 1, just using the definition of Hausdorff measure.2.4. Let A be a nonempty subset of Rn. First prove that if 0 � m < k and H

m�A� <1, then H

k�A� D 0. Second, deduce that the four definitions of the Hausdorffdimension of A are equivalent.

2.5. Define a set A ² R2 as in the following figure by starting with an equilateraltriangle and removing triangles as follows. Let A0 be a closed equilateral triangularregion of side 1. Let A1 be the three equilateral triangular regions of side 1

3 inthe corners of A0. In general let AjC1 be the triangular regions, a third the size,in the corners of the triangles of Aj. Let A D \Aj. Prove that H

1�A� D 1.

Measures 19

2.6. To define the usual Cantor set C ² R1, let C1 D [0, 1]; construct CjC1 by remov-ing the open middle third of each interval of Cj and put

C D \ fCj: j 2 �Cg.

Let m D ln 2/ ln 3.

(a) Prove that Hm�C� � ˛m/2m and, hence, dimC � m.

(b) Try to prove that Hm�C� D ˛m/2m or at least that H

m�C� > 0 and hence thatthe Hausdorff dimension of C is m.

2.7. Give a function f: R2 ! R which is approximately continuous at 0, but for which0 is not a Lebesgue point.

2.8. Prove that if f: Rm ! R has 0 as a Lebesgue point, then f is approximatelycontinuous at 0.

2.9. Deduce Corollary 2.10 from Corollary 2.9.Hint : Let fqig be a countable dense subset of R, Ai D fx:f�x� > qig, and Ei Dfx:�Ai, x� D 1Ai g, and show that f is approximately continuous at each pointin \ Ei.

CHAPTER 3

Lipschitz Functions and Rectifiable Sets

This chapter introduces the m-dimensional surfaces of geometric measuretheory, called rectifiable sets. These sets have folds, corners, and more generalsingularities. The relevant functions are not smooth functions as in differentialgeometry, but Lipschitz functions. See also the survey, “What Is a Surface?”[Morgan 24].

3.1. Lipschitz Functions. A function f: Rm ! Rn is Lipschitz if there is aconstant C such that

jf�x�� f�y�j � Cjx � yj.

The least such constant is called the Lipschitz constant and is denoted byLipf. Figure 3.1.1 gives the graphs of two typical Lipschitz functions. Theo-rems 3.2 and 3.3 show that a Lipschitz function comes very close to beingdifferentiable.

3.2. Rademacher’s Theorem [Federer, 3.1.6]. A Lipschitz function f:Rm ! Rn is differentiable almost everywhere.

The Proof has five steps:

(1) A monotonic function f: R ! R is differentiable almost everywhere.

(2) Every function f: R ! R which is locally of bounded variation (andhence every Lipschitz function) is differentiable almost everywhere.

(3) A Lipschitz function f: Rm ! Rn has partial derivatives almost every-where.

21


f

xx

y

f

f (x) = |x| f (x,y) = √x2 + y2

Figure 3.1.1. Examples of Lipschitz functions.

(4) A Lipschitz function f: Rm ! Rn is approximately differentiable almosteverywhere.

(5) A Lipschitz function f: Rm ! Rn is differentiable almost everywhere.

Step (1) is a standard result of real analysis, proved by differentiation ofmeasures. Step (2) follows by decomposing a function of bounded variationas the difference of two monotonic functions. Step (3) follows immediatelyfrom Step (2) (modulo checking measurability). The deduction of (4) from(3) is technical, but not surprising, because the existence of continuous partialderivatives implies differentiability and a measurable function is approximatelycontinuous almost everywhere. If (3) holds everywhere, it does not follow that(4) holds everywhere.

The final conclusion (5) rests on the interesting fact that if a Lipschitz func-tion is approximately differentiable at a, it is differentiable at a. We concludethis discussion with a proof of that fact.

Suppose that the Lipschitz function f: Rm ! Rn is approximately differ-entiable at a but not differentiable at a. We may assume a D 0, f�0� D 0 andapDf�0� D 0.

For some 0 < ε < 1, there is a sequence of points ai ! 0 such that

jf�ai�j ½ εjaij.

Let C D maxfLipf, 1g. Then, for x in the closed ball B�ai, εjaij/3C�,

jf�x�j ½ εjaij � εjaij/3 ½ εjxj/2.

Thus, for x 2 E D [1iD1B�ai, εjaij/3C�,

jf�x�j ½ εjxj/2.

Lipschitz Functions and Rectifiable Sets 23

But E does not have density 0 at 0, because

LmB�ai, εjaij/3C�

˛m�jaij C εjaij/3C�m ½ �εjaij/3C�m�4jaij/3�m D εm

4mCm> 0.

This contradiction of the approximate differentiability of f at 0 completes theproof.

3.3. Approximation of a Lipschitz Function by a C 1 Function [Federer,3.1.15]. Suppose that A ² Rm and that f:A ! Rn is Lipschitz. Given ε > 0there is a C1 function g: Rm ! Rn such that L

mfx 2 A:f�x� 6D g�x�g � ε.

Note that the approximation is in the strongest sense: the functions coincideexcept on a set of measure ε. The proof of 3.3 depends on Whitney’s Exten-sion Theorem, which gives the coherence conditions on prescribed values fora desired C1 function.

3.4. Lemma (Whitney’s Extension Theorem) [Federer, 3.1.14]. Let A bea closed set of points a in Rm at which the values and derivatives of a desiredC1

function are prescribed by linear polynomials Pa: Rm ! R. For each compactsubset C of A and υ > 0, let ��C, υ� be the supremum of the numbers jPa�b��Pb�b�j/ja � bj, jjDPa�b�� DPb�b�jj, over all a, b 2 C with 0 < ja� bj � υ. Ifthe prescribed data satisfy the coherence condition that limυ!0 ��C, υ� D 0 foreach compact subset C of A, then there exists a C1 function g satisfying

g�a� D Pa�a�, Dg�a� D DPa�a�

for all a 2 A.

Remarks. A more general version of Whitney’s Extension Theorem givesthe analogous conditions to obtain a Ck function with values and derivativesprescribed by polynomials Pa of degree k. In the proof, the value g�x� assignedat a point x /2 A is a smoothly weighted average of the values prescribed atnearby points of A. The averaging uses a partition of unity subordinate to acover of AC which becomes finer and finer as one approaches A.

Sketch of Proof of 3.3. First extend f to a Lipschitz function on all ofRm (see [Federer [1, 2.10.43]). Second, by Rademacher’s Theorem 3.2, f isdifferentiable almost everywhere. Third, by Lusin’s Theorem [Federer, 2.3.5],there is a closed subset E of A such that Df is continuous on E and L

m�A�E� < ε. Fourth, for any a 2 E, υ > 0, define

�υ�a� D sup0<jx�aj<υ

x2E

jf�x�� f�a�� Df�a��x � a�jjx � aj


Since as υ ! 0, �υ ! 0 pointwise, then by Egoroff’s Theorem [Federer, 2.3.7]there is a closed subset F of E such that L

m�A� F� < ε and �υ ! 0 uni-formly on compact subsets of F. This condition implies the hypothesesof Whitney’s Extension Theorem (3.4), with Pa�x� D f�a�C Df�a��x � a�.Consequently there is a C1 function g: Rm ! Rn which coincides withf on F.

The following theorem implies for example that Lipschitz images of setsof Hausdorff measure 0 have measure 0.

3.5. Proposition [Federer, 2.10.11]. Suppose f: Rl ! Rn is Lipschitz andA is a Borel subset of Rl. Then

∫RnN�fjA, y� d H

my � �Lipf�mHm�A�.

Here N�fjA, y� � cardfx 2 A:f�x� D yg.

Proof. Any covering of A by sets Si of diameter di yields a covering off�A� by the sets f�Si�, of diameter at most �Lipf�di. Since the approximatingsum ˛m�diam/2�m for the Hausdorff measure contains �diam�m,

Hm�f�A�� Lipf�mH

m�A�.

Notice that this formula gives the proposition in the case that f is injective.In the general case, chop A up into little pieces Ai and add up the formulasfor each piece to obtain

∫f�A�

�the number of Ai intersecting f�1fyg� d Hmy � �Lipf�mH

m�A�.

As the pieces subdivide, the integrand increases monotonically to the multi-plicity function N�fjA, y�, and the proposition is proved.

The beginning of this proof illustrates the virtue of allowing coverings byarbitrary sets rather than just balls in the definition of Hausdorff measure. IffSig covers A, then ff�Si�g is an admissible covering of f�A�.

3.6. Jacobians. Jacobians are the corrective factors relating the elements ofareas of the domains and images of functions. If f: Rm ! Rn is differentiableat a, we define the k-dimensional Jacobian of f at a, Jkf�a�, as the maximumk-dimensional volume of the image under Df�a� of a unit k-dimensionalcube.

If rank Df�a� < k, Jkf�a� D 0. If rank Df�a� � k, as holds in mostapplications, then Jkf�a�2 equals the sum of the squares of the determinants


of the k ð k submatrices of Df�a�. If k D m or n, then Jkf�a�2 equals thedeterminant of the k ð k product of Df�a� with its transpose. If k D m D n,then Jkf�a� is just the absolute value of the determinant of Df�a�. In general,computations are sometimes simplified by viewing Df�a� as a map from theorthogonal complement of its kernel onto its image. If L: Rm ! Rm is linear,then L

m�L�A�� D JmL Ð Lm�A�.

3.7. The Area Formula [Federer, 3.2.3]. Consider a Lipschitz functionf: Rm ! Rn for m � n.

(1) If A is an Lm measurable set, then

∫AJmf�x� d L

mx D∫

RnN�fjA, y� d H

my.

(2) If u is an Lm integrable function, then∫

Rmu�x�Jmf�x� dLmx D

∫Rn

∑x2f�1fyg

u�x� d Hmy.

Remark. If f is a smooth embedding, then (1) equates the classical areaof the parameterized surface f�A� with the Hausdorff measure of f�A�.Therefore for all smooth surfaces, the Hausdorff measure coincides with theclassical area.

Sketch of the Proof of the Area Formula 3.7(1). We will split up A intotwo cases, according to the rank of Df. In either case, by Rademacher’sTheorem 3.2 and 3.5, we may assume that f is differentiable.

CASE 1. Df has rank m. Let fsig be a countable dense set of affine maps ofRm onto m-dimensional planes in Rn. Let Ei be a piece of A such that for eacha 2 Ei the affine functions f�a�C Df�a��x � a� and si�x� are approximatelyequal. It follows that

(1) det si ³ Jmf on Ei,

(2) f is injective on Ei, and

(3) the associated map from si�Ei� to f�Ei� and its inverse both have Lipschitzconstant ³ 1.

Because f is differentiable, the Ei cover A. Refine fEig into a countabledisjoint covering of A by tiny pieces. On each piece E, by (3) and 3.5,

Hm�f�E�� ³ H

m�si�E��

D Lm�si�E��


D∫E

det si d Lm

³∫EJmfd L

m.

Summing over all the sets E yields∫�number of sets E intersecting f�1fyg� d H

my ³∫AJmfd L

m.

Taking a limit yields∫N�fjA, y� d H

my D∫AJmfd L

m

and completes the proof of Case 1.We remark that it does not suffice in the proof just to cut A up into tiny pieces

without using the si. Without the requirement that for a, b 2 E, Df�a� ³Df�b�, f need not even be injective on E, no matter how small E is.

CASE 2. DF has rank < m. In this case the left-hand side∫A Jmf is zero.

Define a function

g: Rm ! RnCm

x ! �f�x�, εx�.

Then Jm�g� � ε�LipfC ε�m�1. Now by Case 1,

Hm�f�A�� H

m�g�A��

D∫AJmg

� ε�LipfC ε�m�1Lm�A�.

Therefore the right-hand side also must vanish. Finally we remark that 3.7(2)follows from 3.7(1) by approximating u by simple functions.

The following useful formula relates integrals of a function f over a set Ato the areas of the level sets A \ f�1fyg of the function.

3.8. The Coarea Formula [Federer, 3.2.11]. Consider a Lipschitz functionf: Rm ! Rn with m > n. If A is an L

m measurable set, then∫AJnf�x� d L

mx D∫

RnHm�n�A \ f�1fyg� d L

ny.


Proof. CASE 1. f is orthogonal projection. If f is orthogonal projection,then Jnf D 1, and the coarea formula is reduced to Fubini’s Theorem.

GENERAL CASE. We treat just the main case Jnf 6D 0. By subdividing Aas in the proof of the area formula, we may assume that f is linear. Thenf D L °P, where P denotes projection onto the n-dimensional orthogonalcomplement V of the kernel of f and where L is a a nonsingular linearmap from V to Rn. Now

∫AJnfd L

m D j det LjHm�A�

D j det Lj∫P�A�

Hm�n�P�1fyg� d L

ny

D∫L °P�A�

Hm�n��L °P�

�1fyg� d Lny

as desired.

3.9. Tangent Cones. Suppose that a 2 Rn, E ² Rn, and ϕ is a measure onRn. Define a measure ϕ E, “the restriction of ϕ to E,” by

�ϕ E��A� D ϕ�E \ A�.

As in 2.5, define m-dimensional densities [Federer, 2.10.19]

m�ϕ, a� D limr!0

ϕ�B�a, r��˛mrm

.

m�E, a� D m�H m E, a�

D limr!0

Hm�E \ B�a, r��˛mrm

.

Define the tangent cone of E at a consisting of the tangent vectors of E at a:

Tan�E, a� D fr 2 R: r ½ 0g[⋂ε>0

Clos{x � a

jx � aj : x 2 E, 0 < jx � aj < ε}]

[Federer, 3.1.21].Define the (smaller) cone of approximate tangent vectors of E at a:

Tanm�E, a� D \fTan�S, a�:m�E� S, a� D 0g

[Federer, 3.2.16]. See Figure 3.9.1.


Figure 3.9.1. A set, its tangent cone, and its approximate tangent cone at a. Theapproximate tangent cone ignores lower-dimensional pieces.

3.10. Rectifiable Sets [Federer, 3.2.14]. A set E ² Rn is called (Hm,m�

rectifiable if Hm�E� < 1 and H

m almost all of E is contained in the unionof the images of countably many Lipschitz functions from Rm to Rn. Thesesets are the generalized surfaces of geometric measure theory. They includecountable unions of immersed manifolds (as long as the total area stays finite)and arbitrary subsets of Rm.

Rectifiable sets can have countably many rectifiable pieces, perhaps connec-ted by countably many tubes and handles and perhaps with all points in Rn aslimit points (cf. Figure 3.10.1). Nevertheless, we will see that from the pointof view of measure theory, rectifiable sets behave like C1 submanifolds.

This book will call an �H m,m�-rectifiable and Hm-measurable set an m-

dimensional rectifiable set.

Figure 3.10.1. A two-dimensional rectifiable set in R3 consisting of the surfaces ofcountably many bicycles.


The following proposition shows that a measurable set E is rectifiable ifand only if H

m�E� < 1 and Hm almost all of E is contained in a countable

union of C1, embedded manifolds.

3.11. Proposition [cf. Federer 3.2.18]. In the definition of a rectifiable set E,one can take the Lipschitz functions to be C1 diffeomorphisms fj on compactdomains with disjoint images whose union coincides with E H

m almost every-where. Moreover, the Lipschitz constants of fj and f�1

j can be taken near 1.

Proof. It suffices to obtain 1% of the set; the rest can be exhausted by repe-tition. By subdividing them we may assume that the domains have diameter atmost 1. The first Lipschitz function f can be replaced by a C1 approximationg by Theorem 3.3. By the area formula 3.7, we may assume Dg is nonsin-gular. By subdividing the domain, we may assume it is reasonably small. Takejust a portion of the domain so that image g ² image f, Dg is approximatelyconstant, and hence g is injective. Altering domain g by a linear transforma-tion makes Dg ³ identity, and Lip g ³ Lip g�1 ³ 1. Finally the domain maybe replaced by a compact subset. Thus 1% of image f is obtained. Simi-larly replace the second Lipschitz function by a nice one with disjoint image.Continuing through all the original Lipschitz functions yields 1% of the set E.Countably many repetitions of the whole process prove the proposition.

The following proposition shows that in a certain sense a rectifiable set hasa tangent plane at almost every point. (Using different definitions, Hardt andSimon [2, 2.5, p. 22] or Simon [3, 11.6] shows that a modified “rectifiability”is equivalent to the existence of certain “approximate tangent planes” almosteverywhere.)

3.12. Proposition [Federer, 3.2.19]. If W is an m-dimensional rectifiablesubset of Rn, then for almost all points a in W, the density m�W, a� D 1 andTanm�W, a� is an m-dimensional plane. If f is a Lipschitz map from W to R.,then f is approximately differentiable H

m almost everywhere.

EXAMPLE. This example gives a modest indication of how bad rectifiablesets can be and hence how strong Proposition 3.12 is. Begin by constructing aCantor-like set of positive measure as follows. Start with the unit interval. First,remove the middle open interval of length 1

4 . (See Figure 3.12.1.) Second,from the two remaining intervals, remove middle open intervals of total length18 . At the nth step, from the 2n�1 remaining intervals, remove middle openintervals of total length 2��nC1�. Let C be the intersection. Clearly C containsno interval. However, since the total length removed was 2��nC1� D 1

2 , thelength remaining H

1�C� D 12 .


Figure 3.12.1. A Cantor-like set C with H1�C� D 1

2 .

Figure 3.12.2. The image of g intersects [0, 1] in the set C.

Now define g: [0, 1] ! R2 by

g�x� D �x, dist�x, C��.

See Figure 3.12.2.Then image g and hence E D [0, 1] [ �image g� are rectifiable, even though

E fails to be a submanifold at all points of C. Nevertheless, Proposition 3.12says that 1�E, x� D 1 and Tan1�E, x� is a line at almost all points x 2 C.

Remarks on Proof. The proof that m�W, a� D 1 almost everywhere usesa covering argument (see Corollary 2.9).

Proposition 3.11 implies that Tanm�W, a� contains an m-plane almost every-where. Since m�W, a� D 1, it can contain no more.

Similarly by Proposition 3.11, at almost every point, neglecting sets ofdensity 0,W is parameterized by a nonsingular C1 map g: Rm ! Rn. ByRademacher’s Theorem, 3.2, f ° g is differentiable almost everywhere, andhence f is approximately differentiable almost everywhere.

Here we state a general theorem which subsumes both the area and thecoarea formula.

3.13. General Area–Coarea Formula [Federer, 3.2.22]. Let W be an m-dimensional rectifiable subset of Rn, Z a 0-dimensional rectifiable subset ofR., m ½ 0 ½ 1, and f a Lipschitz function from W to Z. Then

∫WapJ0fd H

m D∫Z

Hm�0�f�1fzg� d H

0z.

More generally, for any Hm W integrable function g on W,

∫Wg Ð ap J0fd H

m D∫z

∫f�1fzg

g d Hm�0 d H

0z.


Note: If f has an extension f to Rn, ap J0f � J0f (where both are defined).

3.14. Product of measures [Federer, 3.2.23]. Let W be an m-dimensionalrectifiable Borel subset of Rn and let Z be a 0-dimensional rectifiable Borelsubset of R.. If W is contained in the image of a single Lipschitz function ona bounded subset of Rm, then Wð Z is rectifiable and

HmC0 �Wð Z� D �H m W�ð �H 0 Z�.

Remarks. In general, the additional hypothesis on W is necessary. If 0 D., it holds automatically. In particular, if W is an m-dimensional rectifi-able Borel subset of Rn, then Wð [0, 1]. is an �m C .�-dimensional recti-fiable subset of RnC.. If m D n and 0 D ., this proposition is just Fubini’sTheorem.

The proof, as that of Fubini’s Theorem, shows that the collection of sets onwhich the proposition holds is a 2-algebra.

3.15. Orientation. An orientation of an m-dimensional rectifiable subsetW of Rn is a (measurable) choice of orientation for each Tanm�W, a�. Atpresent no further coherence is required, but we will see in Section 4.2 that abad choice will make the boundary ∂W much worse. Every rectifiable setof positive measure has uncountably many different orientations (not justtwo).

3.16. Crofton’s Formula [Federer, 3.2.26]. If W is an m-dimensionalrectifiable set, then the integralgeometric measure of W equals its Hausdorffmeasure:

Im�W� D H

m�W�.

Remarks. Crofton’s Formula follows easily from the coarea formula. Theproof, although stated for one-dimensional measure in R2, applies virtuallyunchanged to m-dimensional measure in Rn.

Proof. For a one-dimensional measure in R2,

H1�W� D

∫W

(length of unit tangent) d H1

D∫W

1

ˇ�2, 1�

∫p2OŁ�2,1�

(length of projection of unit tangent) dp d H1

(because I1 (unit tangent) D 1)

D 1

ˇ�2, 1�

∫p2OŁ�2,1�

∫W

(length of projection of unit tangent) d H1 dp


D 1

ˇ�2, 1�

∫p2OŁ�2,1�

∫WJ1p d H

1 dp

D 1

ˇ�2, 1�

∫p2OŁ�2,1�

∫N�pjW,y� d H

1y dp

(by the Coarea Formula, 3.13, because W rectifiable)

D I1�W�.

The proof is virtually identical in general dimensions.

3.17. Structure Theorem [Federer, 3.3.13]. This striking theorem descri-bes the structure of arbitrary subsets of Rn. Proved for one-dimensional subsetsof R2 by Besicovitch in 1939, it was generalized to general dimensions byFederer in 1947.

Let E be an arbitrary subset of Rn with Hm�E� < 1. Then E can be decom-

posed as the union of two disjoint sets E D A [ B with A �H m,m� rectifiableand I

m�B� D 0.

Remarks. That Im D 0 means that almost all of its projections onto m-

planes have measure 0; we might say B is invisible from almost all directions.Such a set B is called purely unrectifiable.

The proof, a technical triumph, employs numerous ingenious coverings,notions of density, and amazing dichotomies. A nice presentation of Besicov-itch’s original proof of the structure theorem for one-dimensional subsets ofthe plane appears in [Falconer, Chapter 3].

Structure theory had been considered the most daunting component of theproof of the compactness theorem for integral currents, 5.5. In 1986, followingBruce Solomon, Brian White [3] found a direct argument that obviated thedependence on structure theory. In 1998, White [4] gave an easier proof byinduction of the structure theorem.

If E is Borel, so are A and B.

EXAMPLE. Purely unrectifiable sets result from Cantor-type constructions.For example, start with the unit square. Remove a central cross, leaving foursquares, each 1

4 as long as the first. (See Figure 3.17.1.) Similarly, removecentral crosses from each small square, leaving 16 smaller squares. Continue,and let the set E be the intersection.

The set E is purely unrectifiable. H1�E� D p

2, but I1�E� D 0. Almost all

projections onto lines have measure 0. For example, the projection onto thex-axis is itself a slim Cantor-like set of dimension 1

2 . A diagonal line (with


Figure 3.17.1. A purely unrectifiable one-dimensional set E. E is invisible fromalmost all directions.

slope 12 ) gives an exceptional case: the projection is a solid interval. If A is

any rectifiable set, then H1�A \ E� D 0.

EXERCISES

3.1. Give an example of a Lipschitz function f: [0, 1] ! R such that f is not differ-entiable at any rational point.

3.2. Use Theorem 3.3 to deduce that a Lipschitz function is approximately differen-tiable almost everywhere.

3.3. Give an example of a continuous function f: R ! R such that

(a) given ε > 0 there is aC1 function g: R1 ! R1 such that L1fx:f�x� 6D g�x�g <

ε, but(b) f is not Lipschitz.

3.4. Consider the map f: R2 � f0g ! R2 carrying Cartesian coordinates (x, y) to polarcoordinates (r, 6). What is J1f?

3.5. Consider a differentiable map f: Rn ! R. Show that J1f D jrfj.3.6. Compute H

2 of the unit two-sphere S2�0, 1� by considering the map

f: R2 ! R3

f: �ϕ, 6� ! �sinϕ cos 6, sinϕ sin 6, cosϕ�.


3.7. Verify the coarea formula for f: R3 ! R given by f�x, y, z� D x2 C y2 C z2, A DB3�0, R�.

3.8. Let E be an m-dimensional rectifiable Borel subset of the unit sphere in Rn withHm�E� D a0. Let C D ftx: x 2 E, 0 � t � 1g.

(a) Rigorously compute HmC1�C�.

(b) Compute mC1�C, 0�.(c) What is TanmC1�C, 0�?

3.9. Give an example of an (H2, 2)-rectifiable subset E of R3 which is dense in R3.

Can you also make fx 2 R3:2�E, x� D 1g dense in R3?

CHAPTER 4

Normal and Rectifiable Currents

In order to define boundary and establish compactness properties, it will beuseful to view our rectifiable sets as currents, i.e., linear functionals on smoothdifferential forms (named by analogy with electrical currents). The action ofan oriented rectifiable set S on a differential form ϕ is given by integratingthe form ϕ over the set:

S�ϕ� D∫Sϕ d H

m.

Currents thus associated with certain rectifiable sets, with integer multiplicities,will be called rectifiable currents. The larger class of normal currents will allowfor real multiplicities and smoothing.

The concept of currents is a generalization, by de Rham [2], of distri-butions. Normal and rectifiable currents are due to Federer and Fleming.Important earlier and contemporaneous work includes the generalized surfacesof L. C. Young [1, 2], the frontiers of E. De Giorgi [1, 3], and the surfaces ofE. R. Reifenberg [1–3]. For hypersurfaces, rectifiable currents are just bound-aries of the sets of finite perimeter of Caccioppoli and De Giorgi (see Giusti).The general reference for this chapter is [Federer, Chapter IV].

4.1. Vectors and Differential Forms [Federer, Chapter 1 and 4.1]. Con-sider Rn with basis e1, e2, . . . , en. There is a nice way of multiplying m vectorsin Rn to obtain a new object called an m-vector :

D v1 ^ Ð Ð Ð ^ vm.

35


This wedge product is characterized by two properties. First, it is multilinear:

cv1 ^ v2 D v1 ^ cv2 D c�v1 ^ v2�,

�u1 C v1� ^ �u2 C v2� D u1 ^ u2 C u1 ^ v2 C v1 ^ u2 C v1 ^ v2.

Second, it is alternating:

u ^ v D �v ^ u or u ^ u D 0.

For example,

�2e1 C 3e2 � 53� ^ �7e1 � 11e3�

D 14e1 ^ e1 � 22e1 ^ e3 C 21e2 ^ e1 � 33e2 ^ e3

� 35e3 ^ e1 C 55e3 ^ e3

D 0 � 22e1 ^ e3 � 21e1 ^ e2 � 33e2 ^ e3 C 35e1 ^ e3 C 0

D �21e12 C 13e13 � 33e23.

We have abbreviated e12 for e1 ^ e2.In general, computation of D v1 ^ Ð Ð Ð ^ vm yields an answer of the form

D∑

i1<ÐÐÐ<imai1ÐÐÐimei1ÐÐÐim .

The set of all linear combinations of fei1ÐÐÐim : i1 < Ð Ð Ð < img is the space mRn

of m-vectors, a vectorspace of dimension(nm

). It has the inner product for

which fei1...img is an orthonormal basis.The purpose of an m-vector D v1 ^ Ð Ð Ð ^ vm is to represent the oriented

m-plane P through 0 of which v1, Ð Ð Ð , vm give an oriented basis. Fortunately,the wedge product 0 D v0

1 ^ Ð Ð Ð ^ v0m of another oriented basis for P turns out

to be a positive multiple of . For example, replacing v1 with v01 D ∑

civiyields

v01 ^ v2 ^ Ð Ð Ð ^ vm D c1v1 ^ v2 ^ Ð Ð Ð ^ vm.

If v1, . . . , vm give an orthonormal basis, then D v1 ^ Ð Ð Ð ^ vm has length 1.A product v1 ^ Ð Ð Ð ^ vm is 0 if and only if the vectors are linearly dependent.For the case m D n,

v1 ^ Ð Ð Ð ^ vn D det[v1, . . . , vn] Ð e1...n.

An m-vector is called simple or decomposable if it can be written as asingle wedge product of vectors. For example, in 2R4, e12 C 2e13 � e23 D

Normal and Rectifiable Currents 37

�e1 C e3� ^ �e2 C 2e3� is simple, whereas e12 C e34 is not (see Exercise 4.5).The oriented m-planes through the origin in Rn are in one-to-one correspon-dence with the unit, simple m-vectors in mRn.

Incidentally, the geometric relationship between two m-planes in Rn is givenby m angles, as appeared at least as early as [Somerville, IV.12], with beautifullater applications to the geometry of Grassmannians (see Wong) and to areaminimization (see Morgan [1, §2.3]).

Let RnŁ denote the space of covectors dual to Rn, with dual orthonormalbasis eŁ

1, . . . , eŁn. We remark that dxj is a common alternate notation for eŁ

j. Thedual space to mRn is the space mRn m�RnŁ� of linear combinations ofwedge products of covectors, called m-covectors. The dual basis is feŁ

i1ÐÐÐim: i1 <

Ð Ð Ð < img.A differential m-form ϕ on Rn is an m-covectorfield, that is, a map

ϕ: Rn ! mRn.

For example, one 2-form on R4 is given by

ϕ D cos x1eŁ12 C sin x1eŁ

34

D cos x1 dx1 dx2 C sin x1 dx3 dx4.

The support, sptϕ, of a differential form ϕ is defined as the closure of fx 2Rn:ϕ�x� 6D 0g.

A differential m-form ϕ is a natural object to integrate over an oriented, m-dimensional rectifiable set S, because it is sensitive to both the location x 2 Sand the tangent plane to S at x. Let ES�x� denote the unit m-vector associatedwith the oriented tangent plane to S at x. Then

∫Sϕ

∫S

⟨ES�x�, ϕ�x�

⟩d H

mx.

In a classical setting, with no Hausdorff measure available, the definition ismore awkward. One uses local parameterizations and proves that the definitionis independent of the choice of parameterization. Even the appropriateness ofdealing with forms — functions on mRn — is obscured.

The exterior derivative dϕ of a differential m-form

ϕ D∑

fi1...im eŁi1...im

is the (mC 1)-form given by

dϕ D∑

dfi1...im ^ eŁi1...im ,


where df D �∂f/∂x1�eŁ1 C Ð Ð Ð C �∂f/∂xn�eŁ

n. For example, if

ϕ D fdy dz C g dz dx C h dx dy,

then

dϕ D(∂f

∂xC ∂g

∂yC ∂h

∂z

)dx dy dz

D div�f, g, h� dx dy dz.

If � is a differential l-form and ω is a differential m-form, then

d�� ^ ω� D �d�� ^ ω C ��1�l� ^ dω.

In addition to the dual Euclidean norms jj, jϕj on mRn and mRn, thereare the mass norm jjjj and comass norm jjϕjjŁ, also dual to each other, definedas follows:

jjϕjjŁ D supfjh, ϕij: is a unit, simple m-vectorg;jjjj D supfjh, ϕij: jjϕjjŁ D 1g.

It follows from convexity theory that

jjjj D inf{∑

jij: D∑

i, i simple}.

Consequently, jjϕjjŁ D supfjh, ϕij: jjjj D 1g, so that the mass and comassnorms are indeed dual to each other. Federer denotes both mass and comassnorms by jj jj.

4.2. Currents [Federer, 4.1.1, 4.1.7]. The ambient space is Rn. Let

Dm D fC1 differential m-forms with compact supportg.

For example, in R4, a typical ϕ 2 D2 takes the form

ϕ D f1 dx1 dx2 C f2 dx1 dx3 C f3 dx1 dx4 C f4 dx2 dx3

C f5 dx2 dx4 C f6 dx3 dx4

D f1eŁ12 C f2eŁ

13 C f3eŁ14 C f4eŁ

23 C f5eŁ24 C f6eŁ

34,

where the fj are C1 functions of compact support. The topology is generatedby locally finite sets of conditions on the fj and their derivatives of arbitraryorder.


The dual space is denoted Dm and called the space of m-dimensional cur-rents. This is a huge space. Under the weak topology on Dm, Tj ! T if andonly if Tj�ϕ� ! T�ϕ� for all forms ϕ 2 D

m.Any oriented m-dimensional rectifiable set may be viewed as a current as

follows. Let ES�x� denote the unit m-vector associated with the oriented tangentplane to S at x. Then for any differential m-form ϕ, define

S�ϕ� D∫S

⟨ES�x�, ϕ

⟩d H

m.

Furthermore, we will allow S to carry a positive integer multiplicity "�x�,with

∫S "�x� d H

m < 1, and define

S�ϕ� D∫S

⟨ES�x�, ϕ

⟩"�x� d H

m.

Finally, we will require that S have compact support. Such currents are calledrectifiable currents.

Definitions for currents are by duality with forms. The boundary of an m-dimensional current T 2 Dm is the �m � 1�-dimensional current ∂T 2 Dm�1

defined by∂T�ϕ� D T�dϕ�.

By Stokes’s Theorem, this agrees with the usual definition of boundary if Tis (integration over) a smooth oriented manifold with boundary. Notice thatgiving a piece of the manifold the opposite orientation would create additionalboundary (as in Figure 4.3.6). A boundary has no boundary; i.e., ∂ ° ∂ D 0, asfollows from the easy fact that d °d.

The boundary of a rectifiable current S is generally not a rectifiable current.If it happens to be, then the original current S is called an integral current.The support of a current is the smallest closed set C such that

�spt ϕ� \C D ∅ ) S�ϕ� D 0.

4.3. Important Spaces of Currents [Federer, 4.1.24, 4.1.22, 4.1.7, 4.1.5].Figure 4.3.1 gives increasingly general spaces of currents that play an impor-tant role in geometric measure theory. Figure 4.3.2 pictures some low dimen-sional examples. The first tier has a polygonal curve, an integral current offinite length and finite boundary, and a rectifiable region of finite area but infi-nite boundary length, bounded by an integral flat chain of infinite length. Thesecond tier allows real multiplicities and smoothing. The final rows illustratemore general currents without the same geometric significance.


Pm ² Im ² Rm ² Fm

integral integral rectifiable integralpolyhedral currents currents flat chainschains

\ \ \ \Pm ² Nm ² Rm ² Fm

real normal realpolyhedral currents flat chainschains \

Em ² Dm

Figure 4.3.1. The increasingly general spaces of currents of geometric measure theory.

Multiplicity 1/2 Multiplicity 1/2

I111

P1 N1 F1

R2

T 2 2 [0,1]2^i

2

1

T (fdx + gdy) = f ′(0)

T (fdx + gdy) = f ′(1,1) + f ′′(2,2) + f ′′′(3,3) + . . .

1

Figure 4.3.2. Examples of increasingly general types of currents, with finite or infi-nite mass or boundary mass. The second tier admits fractional multiplicities andsmoothing.

DEFINITIONS. Let

Dm D fm-dimensional currents in Rng,Em D fT 2 Dm: sptT is compactg,

Rm D frectifiable currentsgD fT 2 Em associated with oriented rectifiable sets, with integer

multiplicities, with finite total measure (counting multiplicities�g,


Pm D fintegral polyhedral chainsgD additive subgroup of Em generated by classically oriented simplices,

Im D fintegral currentsgD fT 2 Rm: ∂T 2 Rm�1g,

Fm D fintegral flat chainsgD fTC ∂S:T 2 Rm, S 2 RmC1g.

The definitions of the second tier of spaces will appear in Section 4.5.We also define two important seminorms on the space of currents Dm: the

mass M and the flat norm F.

M�T� D supfT�ϕ�: supx

jjϕ�x�jjŁ � 1g,

F �T� D inffM�A�C M�B�:T D AC ∂B, A 2 Rm, B 2 RmC1g.

The mass of a rectifiable current is just the Hausdorff measure of the associ-ated rectifiable set (counting multiplicities) as explained in Section 4.5. Notethat the norm supx jjϕ�x�jjŁ gives a weaker topology on Dm than the one towhich currents are dual, so that a general current may well have infinite mass.Similarly, F �T� < 1 if and only if T 2 Fm.

The flat norm gives a good indication of when surfaces are geometricallyclose together. For example, the two unit discs D1, D2 of Figure 4.3.3 areclose together under the flat norm F because their difference T D D2 � D1,together with a thin band A, is the boundary of a squat cylindrical region Bof small volume. On the other hand, in the mass norm, M�D2 � D1� D 2'.

The flat norm topology is clearly weaker than the mass norm topology butstronger than the weak topology. Actually it turns out that for integral currentsof bounded mass and boundary mass, the flat and weak topologies coincide asfollows from the Compactness Theorem 5.5, or really, just from Corollary 5.2(compare Simon [3, 31.2]).

D1B

A D2

Figure 4.3.3. The unit discs D1, D2 are close together in the flat form F becausetheir difference T D D2 � D1, together with a thin band A, is the boundary of a squatcylindrical region B of small volume. [Morgan 24, Figure 7].


To obtain a rectifiable current which is not an integral current, choose theunderlying rectifiable set E with infinite boundary. For example, let E be aconnected open subset of the unit disc bounded by a curve of infinite length,as in Figure 4.3.4.

E

Figure 4.3.4. Although a rectifiable set E must have finite area, its boundary canwiggle enough to have infinite length. Thus a rectifiable current need not be an integralcurrent. Here the width of each successive smaller square rapidly approaches one-thirdthe length of the larger square.

Alternatively, let E be a countable union of discs of radius 1/k:

E D⋃k 2 ZC

f�x, y, z�: x2 C y2 � k�2, z D k�1g.

See Figure 4.3.5.As a second alternative, decompose the unit disc into the infinitely many

concentric annuli

An D f1/�nC 1� < r � 1/ng

of Figure 4.3.6 with alternating orientations.In all three examples, the associated rectifiable current T is not an integral

current, and ∂T is an integral flat chain but not a rectifiable current.


Figure 4.3.5. This infinite collection of discs gives another example of a rectifiablecurrent which is not an integral current. There is finite total area, but infinite totalboundary length.

Figure 4.3.6. Giving alternating orientations to concentric annuli creates infiniteboundary so that the disc is no longer an integral current. [Morgan 24, Fig. 6].


Actually, only by having infinite boundary mass can a rectifiable current failto be an integral current. The difficult Closure Theorem, 5.4, will show that

�1�Im D fT 2 Rm: M�∂T� < 1g,

Rm D fT 2 Fm: M�T� < 1g.

(The equivalence of these two equalities follows immediately from the defi-nitions.)

Remarks on Supports and Notation. Let K be a compact C1 submanifoldof Rn, with or without boundary (or more generally, a “compact Lipschitzneighborhood retract”). Federer uses the subscript K to denote support in K.For example,

Rm,K D fT 2 Rm: sptT ² Kg.

(For arbitrary compactK, Rm,K has a more technical meaning [Federer, 4.1.29].)Similarly, a norm FK is defined by

FK�T� inffM�A�C M�B�:T D AC ∂B, A 2 Rm,K, B 2 RmC1,Kg.

If K is any large ball containing sptT, then FK�T� equals what we have calledF �T�, as can be seen by projecting the A and B from the definition of Fonto K. In the other main case of interest, when K is a compact C1 submani-fold of Rn, FK�T� ½ F �T�, with strict inequality sometimes. However, FK

and F yield the same topology on Fm,K, the integral flat chains supportedin K.

4.3A. Mapping Currents. Next we want to define the image of a compactlysupported current under a C1 map f: Rn ! R,. First, for any simple m-vector D v1 ^ Ð Ð Ð ^ vm 2 mRn, and point x in the domain of f, define thepush-forward of in mR,:

[m�Df�x��]�� D �Df�x��v1� ^ Ð Ð Ð ^ �Df�x��vm�.

The map m�Df�x�� extends to a linear map on all m-vectors.Second, for any differential m-form ϕ 2 D

m�R,�, define its pullback f-ϕon Rn by

h, f-ϕ�x�i D h[m�Df�x��]��, ϕ�f�x��i.


Finally, for any compactly supported current T 2 Dm�Rn�, define its push-forward f-T 2 Dm�R,� by

�f-T��ϕ� D T�f-ϕ�.

If T is the rectifiable current associated with some oriented rectifiable setE, then f-T is the rectifiable current associated with the oriented rectifi-able set f�E�, with the appropriate multiplicities (see Exercise 4.23). Theboundary ∂�f-T� D f-∂T. In many cases the smoothness hypothesis on fmay be relaxed.

4.3B. Currents Representable by Integration. A current T 2 Dm is calledrepresentable by integration if there is a Borel regular measure jjTjj on Rn,finite on compact sets, and a function ET: Rn ! mRn with jjET�x�jj D 1 forjjTjj almost all x such that

T�ϕ� D∫

hET�x�, ϕ�x�id jjTjjx.

The mass M�T� is just the total measure jjTjj�Rn�. We write T D jjTjj ^ ET.A current T 2 Dm of finite mass is automatically representable by integra-

tion, as follows from the representation theory of general measure theory. Onthe other hand, the current T 2 Dm�Rn� defined by

T�a1 dx1 ^ Ð Ð Ð ^ dxm C Ð Ð Ð� D ∂a1

∂x1�p�,

where p is a fixed point in Rn, has infinite mass, and is not representable byintegration.

Every rectifiable current S is representable by integration. Indeed, if E is theassociated set with multiplicity function l, then jjSjj is the measure l�H m E�and ES is the unit m-vectorfield orienting E. S D l�H m E� ^ ES D �H m E� ^/, where / D lES. The mass is

M�S� D jjSjj�Rn� D∫S0 d H

m.

For example, the rectifiable current associated with a unit disc D in R2 ² Rn

is H2 D ^ e12 and its mass is '.


4.4. Theorem [Federer, 4.1.28]. The following are equivalent definitions forT 2 Em to be a rectifiable current.

�1� Given ε > 0, there are an integral polyhedral chain P 2 Pm�R,� and aLipschitz function f: R, ! Rn such that

M�T� f-P� < ε.

�2� There are a rectifiable set B and an Hm B summable m-vectorfield /

such that / is simple, j/�x�j is an integer �“the multiplicity”�, Tanm�B, x�is associated with /�x�, and T�ϕ� D ∫

Bh/�x�, ϕ�x�i d Hm.

Remarks. In (1), if T is supported in a closed ball K, one may assumesptf-P ² K, by replacing f-P by its projection onto K. Actually, Federertakes (1) as the definition of Rm, whereas we have used (2).

A current �H m B� ^ / can fail to be rectifiable in several ways: the setB could fail to be rectifiable or to have compact closure; the total mass∫B jj/�x�jj d H

m could fail to be finite; the given m-vector /�x� could fail to betangent to B at x; or j/�x�j could fail to be an integer.

Proof Sketch. First suppose (1) holds. Since each side of the polyhedralchain is a subset of some Rm, its image under f is rectifiable and hence f-Pis a rectifiable current and satisfies (2). But now T, as a mass convergent sumof such, obtained by successive approximation, is a rectifiable current.

The opposite implication depends on the following lemma of measuretheory.

LEMMA. Let A be a bounded �L m-measurable� subset of Rm. Then givenε > 0, there is a finite disjoint set of m-simplices which coincide with A exceptfor a set of measure less than ε.

Proof of Lemma. We may assume that A is open, by replacing A by aslightly larger open set. Cover 1% of A by disjoint simplices (as in the proofof the Besicovitch Covering Theorem, 2.7). Repeat on what is left. After Nrepetitions, 1 � �.99�N of A is covered by disjoint simplices, as desired.

Completion of Proof of Theorem. Suppose T satisfies (2). The rectifiableset B is the union of Lipschitz images of subsets of Rm. Use the lemma toapproximate B by images of polyhedra.

4.5. Normal Currents [Federer, 4.1.7, 4.1.12]. In preparation for the defi-nitions of more general spaces of currents, we define a more general flat


norm, F. For any current T 2 Dm, define

F�T� D supfT�ϕ�: ϕ 2 Dm, jjϕ�x�jjŁ � 1 and jjdϕ�x�jjŁ � 1 for all xg

D minfM�A�C M�B�:T D AC ∂B, A 2 Em, B 2 EmC1g.The second equality shows the similarity of the norm F and the previouslydefined norm F. Inequality (�) is easy, since if T D AC ∂B as in the minimumand ϕ is as in the supremum, then

T�ϕ� D �AC ∂B��ϕ� D A�ϕ�C B�dϕ� � M�A�C M�B�.

Equality is proved using the Hahn–Banach Theorem (Federer [1, p. 367]).Now continuing the definitions of the spaces of currents in the diagram in

the beginning of Section 4.3, let

Nm D fT 2 Em: M�T�C M�∂T� < 1gD fT 2 Em:T and ∂T are representable by integrationg,

Fm D F-closure of Nm in Em,

Rm D fT 2 Fm: M�T� < 1g,Pm D f real linear combinations of elements of Pmg.

The important space Nm of normal currents allows real densities and smooth-ing. For example, if A is the unit square region

f�x, y�: 0 � x � 1, 0 � y � 1gin the plane, then S1 D p

2�H 2 A� ^ e12 is a two-dimensional normal currentwhich is not an integral current. (See Figure 4.5.1.) S1 is

p2 times the inte-

gral current �H 2 A� ^ e12. S2 D �H 2 A� ^ e1 is a one-dimensional normalcurrent which is not an integral current. To check that S2 is indeed a normalcurrent, compute ∂S2 from the definition:

∂S2�f�x, y�� D S2�df� D S2

(∂f

∂xeŁ

1 C ∂f

∂yeŁ

2

)

D∫A

⟨e1,∂f

∂xeŁ

1 C ∂f

∂yeŁ

2

⟩dH

2

D∫A

∂f

∂xdx dy

D∫ 1

0[f�1, y�� f�0, y�] dy

D∫ 1

0f�1, y� dy �

∫ 1

0f�0, y� dy.


S1Multiplicity

√2

S2

Figure 4.5.1. Currents with noninteger densities and one-dimensional currents spreadover two-dimensional sets give examples of normal currents which are not integralcurrents.

Therefore,

∂S2 D H1 f�1, y�: 0 � y � 1g � H

1 f�0, y�: 0 � y � 1g,

and M�∂S2� D 2 < 1. If B D f�x, 0�: 0 � x � 1g, T D H1 B ^ e1, and 4�x,y�

denotes translation by �x, y�, then

S2 D∫ 1

04�0,y�-T dy.

Thus S2 is an integral of integral currents.More generally, if T is any m-dimensional integral current in Rn and f

is a function of compact support with∫ jfjdL

n < 1, then the weightedsmoothing of T

S D∫x2Rn

f�x� Ð 4x-T d Lnx

is a normal current. Of course,

∂S D∫x2Rn

f�x� Ð 4x-∂T d Lnx.

Whether every normal current can be written as an integral of integral currentswithout cancellation has been a subject of research. A counterexample wasprovided by M. Zworski.

4.6. Proposition [Federer, 4.1.17]. The space Rm is the M-closure of Nm

in Em.

Proof. Clearly Rm is M-closed in Em. Suppose T 2 Rm. Given ε >0, choose S 2 Nm such that F�T� S� < ε. Hence there are currents A 2


Em and B 2 EmC1 such that T� S D AC ∂B and M�A�C M�B� < ε. SinceM�∂B� D M�T� S� A� < 1, ∂B 2 Nm. Therefore SC ∂B 2 Nm, and M�T��SC ∂B�� D M�A� < ε. Hence T is the M-closure of Nm, as desired.

We have seen examples of m-dimensional normal currents based on higher-dimensional sets. The following theorem shows that even real flat chainscannot be supported in lower-dimensional sets. The hypothesis that theintegral geometric measure I

m�sptT� D 0 holds if the Hausdorff measureHm�sptT� D 0, as follows easily from the definition of I

m�2.4�.

4.7. Theorem [Federer, 4.1.20]. If T 2 Fm�Rn� and Im�sptT� D 0, then

T D 0.

EXAMPLES. The current S H0 f�0, 0�g ^ e1 2 D1 is not flat because

I1�spt S� D I

1f�0, 0�g D 0. The current

T D H1 f�0, y�: 0 � y � 1g ^ e1

is not flat, because if it were, its projection on the x-axis, which is S, wouldbe flat. (See Figure 4.7.1.) This example illustrates the principle that fora flat current, the prescribed vectorfield must lie down “flat” (see Federer[1, 4.1.15]). The suggestiveness of the term flat is a happy accident.H. Whitney, also a student of music, coined the term for the smaller of hisflat and sharp norms, originally designated jj jj5, and jj jj-.

Figure 4.7.1. The current T is not flat; its prescribed vectorfield is not tangent to theunderlying set.


Outline of Proof.

I. Smoothing. A smooth normal current in Rn is one of the form Ln ^ ,

with a smooth m-vectorfield of compact support. Any normal current T canbe approximated in the flat norm by a smooth normal current Tε D L

n ^ asfollows. Let f be a smooth approximation to the delta function at 0, and putTε D ∫

x2Rn f�x� Ð 4x-T d Lnx.

II. If T 2 Fn�Rn�, then T is of the form Ln ^ for some vectorfield

[Federer, 4.1.18]. Notice the assumption of codimension 0, where the normsF and M coincide. Therefore T can be M-approximated by a normal currentand hence by smoothing by L

n ^ 1, with 1 a smooth n-vectorfield,M�T� L

n ^ 1� < 2�1, and hence

M�L n ^ 1� D∫

j1jdLn < M�T�C 2�1.

Likewise, T� Ln ^ 1 can be M-approximated by L

n ^ 2, with M�T�Ln ^ 1 � L

n ^ 2� < 2�2, and hence

M�L n ^ 2� D∫

j2j d Ln < 2�1 C 2�2.

Likewise, T� Ln ^ 1 � L

n ^ 2 can be M-approximated by Ln ^ 3,

with M�T� Ln ^ 1 � L

n ^ 2 � Ln ^ 3� < 2�3, and hence M�L n ^

3� < 2�2 C 2�3. Continue. Since∫ ∑1

jD1 jjj < M�T�C 2�1 C 2�1 C 2�2 C2�2 C Ð Ð Ð D M�T�C 2 < 1,

∑j converges in L1. Let D ∑

j. Then T DLn ^ as desired.

III. Completion of Proof. For the case m D n, the theorem follows imme-diately from part II. Let m < n. Since I

m�sptT� D 0, we may assume sptTprojects to sets of measure 0 in the m-dimensional coordinate axis planes. Fornotational convenience we take m D 1, so that T 2 F1�Rn�. We consider theaction of T on an arbitrary smooth 1-form

ϕ D f1eŁ1 C f2eŁ

2 C Ð Ð Ð C fneŁn.

Since T�ϕ� D ∑T�fjeŁ

j�, it suffices to show that T�fjeŁj� D 0. Let pj denote

projection onto the jth coordinate axis, and let T f denote the current definedby �T f��ϕ� D T�fϕ� (see 4.11). Then

T�fjeŁj� D �T fj��eŁ

j� D �T fj��p#je

Łj� D �pj#�T fj��eŁ

j�.


Since pj#�T fj� 2 Fm�Rm� is of the form Lm ^ by part II, and its support

has measure 0, it must be 0. Therefore T�fjeŁj� D 0, as desired.

4.8. Theorem [Federer, 4.1.23]. Given a real flat chain T 2 Fm and ε > 0.there is a real polyhedral approximation P 2 Pm satisfying F�T� P� � ε andM�P� � M�T�C ε.

Proof. Since the space Fm is defined as the F-closure of Nm, and ifM�T� < 1, T lies in the M-closure of Nm (Proposition 4.6), we may assumeT 2 Nm. By smoothing (cf. proof of 4.7, part I), we may assume T is ofthe form T D L

n ^ �x�, where �x� is a smooth m-vectorfield of compactsupport with

∫ j�x�jd Ln < 1. By approximating by step functions, we

may assume T is of the form T D Ln A ^ /, for some bounded set A and m-

vector /. We may assume / D e1ÐÐÐm and A is the unit cube f0 � xi � 1g ² Rn.Now we can approximate T D L

n A ^ / by layers. Take a large integer, M,let

B D fx 2 Rm: 0 � xi � 1g ð{

1

M,

2

M, . . . , 1

}n�m² Rn,

and letP D M��n�m��H m B� ^ e1ÐÐÐm.

Then M�P� D M�T� and for M large, F�T� P� < ε.

4.9. Constancy Theorem [Federer, 4.1.31]. Suppose B is anm-dimensionalconnected, C1 submanifold with boundary of Rn, classically oriented by 8. Ifa real flat chain T 2 Fm is supported in B and its boundary is supported in theboundary of B, then, for some real number r,

T D r�H m B� ^ 8.Of course if T is an integral flat chain, then r is an integer.

Proof. We must show locally that ∂T D 0 means T is constant. We mayassume locally that B D Rm ð f0g ² Rn. Then T is of the form L

m ^ for some m-vectorfield D f Ð e1ÐÐÐm (proof of 4.7, part II). For any smooth(m� 1)-form

ϕ D g1eŁ2ÐÐÐm � g2eŁ

13ÐÐÐm C Ð Ð Ð gmeŁ12ÐÐÐm�1

of compact support.

0 D ∂T�ϕ� D T�dϕ� D∫

h, dϕi d Lm


D∫f

(∂g1

∂x1C ∂g2

∂x2C Ð Ð Ð C ∂gm

∂xm

)d L

m

D∫f div g d L

m.

It follows that f is constant, as desired. (If f is smooth, integration by partsyields that

0 D �∫ (

∂f

∂x1g1 C Ð Ð Ð C ∂f

∂xmgm

)d L

m

for all gj, so that ∂f/∂xi D 0 and f is constant. For general f 2 L1,∫f div g D 0 for all g means the weak derivative vanishes and f is constant.)

4.10. Cartesian Products. Given S 2 Dm�Rn� and T 2 D"�R,�, one candefine their Cartesian product Sð T 2 DmC"�RnC,�. The details appear inFederer [1, 4.1.8. p. 360], but for now it is enough to know that it exists andhas the expected properties.

4.11. Slicing [Federer, 4.2.1]. The Coarea Formula 3.13 relates the area ofa rectifiable set W to areas of its slices. In this section we define (m � 1)-dimensional slices of m-dimensional normal currents by hyperplanes or byhypersurfaces fu�x� D rg. It will turn out that for almost all values of r, theslices themselves are normal currents, and that the boundary of the slice isjust the slice of the boundary. For rectifiable sets the two notions of slicingagree (4.13).

First, for any current T 2 Dm and C1 differential k-form ˛, define a currentT ˛ 2 Dm�k by

�T ˛��ϕ� D T�˛ ^ ϕ�.

In particular, if ˛ is a function (0-form) f, then �T f��ϕ� D T�fϕ�. Thesymbol for such “interior multiplication,” sometimes called “elbow,” pointsto the term of lower degree which gets pushed to the other side in the definition.

If T is representable by integration, T D jjTjj ^ ET, then it suffices to assumethat

∫ jfjd jjTjj < 1. Indeed, then T f D fjjTjj ^ ET: one just multiplies themultiplicity by f. Of course, even if T is rectifiable, T f will not be, unlessf is integer valued. For A ² Rn, define “T restricted to A,” T A D T ;A,where ;A is the characteristic function of A.

For a normal current T 2 NmRn, a Lipschitz function u: Rn ! R, and areal number r, define the slice

(1) hT, u, rCi �∂T� fx: u�x� > rg � ∂�T fx: u�x� > rg�D ∂�T fx: u�x� � rg�� ∂T� fx: u�x� � rg.


(See Figure 4.11.1 on page 54.) It follows that

�2� ∂hT, u, rCi D �h∂T, u, rCi.

PROPOSITION.

�3� MhT, u, rCi � �Lip u� limh!0C

jjTjjfr < u�x� < r C hg/h.

In particular, if f�r� D jjTjjB�x, r�, then for almost all r,

MhT, u, rCi � f0�r�.

Proof. If ; is the characteristic function of the set fx: u�x� > rg, then

hT, u, rCi D �∂T� ; � ∂�T ;�.

For small, positive h, approximate ; by a C1 function f satisfying

f�x� D{

0 if u�x� � r1 if u�x� ½ r C h

and Lipf� �Lip u�/h. Then

MhT, u, rCi ³ M��∂T� f� ∂�T f��

D M�T df�

� �Lipf�jjTjjfx: r < u�x� < r C hg� �Lip u�jjTjjfx: r < u�x� < r C hg/h.

Consequently,

MhT, u, rCi � �Lip u� limh!0

jjTjjfx: r < u�x� < r C hg/h.

PROPOSITION.

�4�∫ b

aMhT, u, rCi d L

1r � �Lip u�jjTjjfx: a < u�x� < bg.

Proof. Consider the function f�r� D jjTjjfu�x� < rg. Since f is monoton-ically increasing, its derivative, f0�r�, exists for almost all r. Hence

�Lip u�jjTjjfa < u�x� < bg D �Lip u�(f�b�� lim

x!aCf�x�)

½ �Lip u�∫ b

af0�r�dr ½

∫ b

aMhT, u, rCidr

by (3).


∂T

T

u = r u > r

u < r u = r u > r

Figure 4.11.1. The slice of the torus T by the pictured plane consists of 1 34 circles.

COROLLARY.

�5� hT, u, rCi 2 Nm�1

for almost all r.


The corollary follows directly from (4) and (2) (see Exercise 4.20). Of course,it follows that if T is rectifiable, so are almost all slices.

PROPOSITION.

�6�∫ b

aF[T fu�x� � rg]d L

1r � [b� aC Lip u]F�T�.

For a hint on the proof, see Exercise 4.21.

Remarks. If T is an integral current, so is almost every slice, as will followfrom the Closure Theorem, 5.4(2,3), and 4.11(5) (or as is shown directlyin Simon [3, §28]). Slicing can be generalized to a vector-valued functionu: Rn ! Rl [Federer, 4.3].

The following lemma considers slices of T by the function u�x� D jx � aj.If T has no boundary, then

hT, u, rCi ∂�T fx: u�x� � rg� D ∂�T B�a, r��.

The lemma says that if almost all such “slices by spheres” are rectifiable, thenT is rectifiable.

4.12. Lemma [Federer, 4.2.15]. If T is a normal current without boundaryand if, for each a 2 Rn, ∂�T B�a, r�� is rectifiable for almost all r 2 R, thenT is rectifiable.

Remarks. This lemma is crucial to proving the Closure and CompactnessTheorems of Chapter 5. The proof of this lemma in Federer uses structuretheory and a covering argument. In 1986, following the work of Solomon,White [3] discovered a way of circumvent the structure theory at almost nocost.

4.13. Proposition [Federer, 4.3.8, 3.2.22]. Let W be an m-dimensional rec-tifiable set in Rn, u a Lipschitz function from W to R". Then for almost all z 2R", W \ u�1fzg is rectifiable, and the associated current is the slice hT, u, rCiof the current T associated to W.

EXERCISES

4.1. Compute �e1 C 2e2 C 3e3� ^ �e1 C 2e2 � 3e3� ^ e4.


4.2. Consider the 2-plane P in R4 given by

P D f�x1, x2, x3, x4�: x1 C x2 C x3 D x3 C x4 D 0g.

Find a nonorthogonal basis u, v and an orthonormal basis w, z for P. Verify bydirect computation that u ^ v is a multiple of w ^ z and that jw ^ zj D 1.

4.3. Verify by direct computation that

�e1 C 2e2 C 3e3� ^ �e1 � e3� ^ �e2 C e3� D∣∣∣∣∣∣1 1 02 0 13 �1 1

∣∣∣∣∣∣ e123.

4.4. Prove that the 2-vector e12 C 2e13 C 2e23 is simple.4.5. Prove that e12 C e34 is not simple.4.6. Find the integral of the differential form

ϕ D x1�sin x1x2�eŁ12 C ex1Cx2Cx3 eŁ

13 C eŁ23

over f�x1, x2, x3� 2 R3: 0 � x1 � 1, 0 � x2 � 1, x3 D 0g (with the usual upwardorientation).

4.7. Find the integral of the differential form

ϕ D 2eŁ12 C 3eŁ

13 C 5eŁ23

over f�x1, x2, x3� 2 R3: x1 C x2 C x3 D 0 and x21 C x2

2 C x23 � 1g.

4.8. Prove that the boundary operator ∂ maps Im into Im�1 and Fm into Fm�1. Alsoprove that spt ∂T ² sptT.

4.9. For the rectifiable currents T 2 R1�R2� and for C1 functions f, g, and h,compute formulas for T�f dx C g dy� and ∂T�h�:

(a) T D H1 f�x, 0�: 0 � x � 1g ^ e1.

(b) T D �H 1 f�x, x�: 0 � x � 1g� ^ 3p

2�e1 C e2�.

4.10. Prove that Im is M dense in Rm and F dense in Fm.4.11. Prove that fT 2 Rm: sptT ² B�0, R�g is M complete and that fT 2 Fm: sptT ²

B�0, R�g is F complete.4.12. Prove that ∂ carries Nm into Nm�1 and Fm into Fm�1.4.13. Check this analogy to 4.3(1):

Nm D fT 2 Rm: M�∂T� < 1g,Rm D fT 2 Fm: M�T� < 1g.

4.14. Prove that, in analogy with the definitions of Im and Fm,

Nm D fT 2 Rm: ∂T 2 Rm�1g,Fm D fTC ∂S:T 2 Rm, S 2 RmC1g.


4.15. Prove that Im ² Nm,Rm ² Rm, and Fm ² Fm.4.16. For the currents T 2 D1�R2� representable by integration and for C1 functions

f, g, and h,

(i) write a formula for T�f dx C g dy�,(ii) write a formula for ∂T�h�, and

(iii) give the smallest space of currents from the table at the beginning ofSection 4.3 to which T belongs:

(a) T D ∑1kD1 H

1 f�k�1, y�: 0 � y � 2�kg ^ j.(b) T D H

2 f�x, y�: 0 � x � 1, 0 � y � 1g ^ i.(c) T D H

1 f�x, 0�: 0 � x � 1g ^ j.(d) T D H

0 fag ^ i.(e) T D H

2 f�x, y�: x2 C y2 � 1g ^ i.

4.17. Let E be the modification of the Cantor set obtained by starting with the unitinterval and removing 2n�1 middle intervals, each of length 4�n, at the nth step(n D 1, 2, 3 . . .).

(a) Show that H1�E� D 1

2 .(b) Show that H

1 E ^ i is a rectifiable current, but not an integral current.

4.18. Prove the second equality in 4.11(1) above.4.19. Prove 4.11(2).4.20. Deduce 4.11(5) from 4.11(4) and 4.11(2).4.21. Prove 4.11(6).

Hint : First show that, if T D AC ∂B with T,A, and B 2 N, then

T fu�x� � rg D A fu�x� � rg C ∂[B fu�x� � rg] � hB, u, rCi.

4.22. Prove that M is F lower semicontinuous on Dm, i.e., if Ti, T 2 Dm, and TiF!T,

then M�T� � lim inf M�Ti�.Hint : Work right from the definition of M in 4.3 and the first definition of Fin 4.5.

4.23. Suppose f is a C1 map from Rn to R, and S D l�H m E� ^ ES is a rectifiablecurrent represented by integration in terms of an underlying rectifiable set E andan integer-valued multiplicity function l.

(a) Assuming that f is injective, show that

f-S D l °f�1�H m f�E�� ^ �mDf��ES�

jmDf��ES�j.

(b) Without assuming that f is injective, show that

f-S D �H m f�E�� ^∑yDf�x�

l�x��mDf�x��ES�

j�mDf�x��ES�j.

Hint : Use the definitions and the general coarea formula, 3.13.

CHAPTER 5

The Compactness Theorem and theExistence of Area-Minimizing Surfaces

The Compactness Theorem, 5.5, deserves to be known as the fundamentaltheorem of geometric measure theory. It guarantees solutions to a wide classof variational problems in general dimensions. It says that a certain set T

of surfaces is compact in a natural topology. The two main lemmas are theDeformation Theorem, 5.1, which will imply in 5.2 that T is totally bounded,and the Closure Theorem, 5.4, which will imply that T is complete.

5.1. The Deformation Theorem [Federer, 4.2.9]. The DeformationTheorem approximates an integral current T by deforming it onto a gridof mesh 2ε > 0. (See Figure 5.1.1.) The resulting approximation, P, isautomatically a polyhedral chain. The main error term is ∂S, where S is thesurface through which T is deformed. There is a secondary error term, Q, dueto moving ∂T into the skeleton of the grid.

Whenever T 2 ImRn and ε > 0, there exist P 2 PmRn, Q 2 ImRn, and S 2ImC1Rn such that the following conditions hold with D 2n2mC2:

(1) T D PC QC ∂S.(2) M�P � [M�T C εM�∂T ],

M�∂P � M�∂T ,

M�Q � εM�∂T ,

M�S � εM�T .

Consequently, F �T� P � ε�M�T C M�∂T .

59


S1

Q1Q2

Q3

Q1

Q2 Q3

TT1

S2

P

S1

2ε

2ε

Figure 5.1.1. The Deformation Theorem describes a multistep process for deforminga given curve T onto a polygon P in the 2ε grid. During the process surfaces S1 andS2 are swept out. The endpoints of T trace out curves Q1, Q2, and Q3.

(3) spt P is contained in the m-dimensional 2ε grid; i.e., if x 2 spt P, thenat least n� m of its coordinates are even multiples of ε. Also, spt ∂P iscontained in the (m� 1)-dimensional 2ε grid.

�4 sptP [ sptQ [ spt S ² fx: dist�x, sptT � 2nεg.

Proof Sketch, Case m D 1, n D 3. Let Wk denote the k-dimensional ε grid:

Wk D f�x1, x2, x3 2 R3 : at least 3 � k of the xj are even multiples of εg.

Then W2 consists of the boundaries of 2εð 2εð 2ε cubes.First project the curve T radially outward from the centers of the cubes onto

W2. (For now, suppose T stays away from the centers.) Let S1 be the surfaceswept out by T during this projection, let Q1 be the curve swept out by ∂T,and let T1 be the image of T in W2. Then, with suitable orientations,

T D T1 C Q1 C ∂S1.

The Compactness Theorem and the Existence of Area-Minimizing Surfaces 61

The mass of T1 is of the same order as the mass of T, M�T1 ¾ M�T .Likewise M�∂T1 ¾ M�∂T , M�Q1 ¾ εM�∂T , and M�S1 ¾ εM�T .

Second, W1, the 1-skeleton of W2, consists of the boundaries of 2εð 2εsquare regions. Project the curve T1 radially outward from the centers of thesquares onto W1. Let S2 be the surface swept out by T1, let Q2 be the curveswept out by ∂T1, and let T2 be the image of T1 inW1. Then, T1 D T2 C Q2 C∂S2, and the masses are of order M�T2 ¾ M�T1 ¾ M�T , M�∂T2 ¾ M�∂T ,M�Q2 ¾ εM�∂T , and M�S2 ¾ εM�T .

Third, let Q3 consist of line segments from each point of ∂T2 to the nearestpoint in the 0-skeleton W0. Put P D T2 � Q3. Then not only does P lie inW1, but also ∂P lies in W0. In particular, P is an integral polyhedral chain.The masses satisfy

M�Q3 ¾ εM�∂T ,M�P D M�T2 C M�Q3 ¾ M�T C εM�∂T ,

M�∂P D M�∂T2 ¾ M�∂T .

Let Q D Q1 C Q2 C Q3 and S D S1 C S2. Then T D PC QC ∂S. Themasses satisfy

M�Q D M�Q1 C M�Q2 C M�Q3 ¾ εM�∂T ,M�S ¾ εM�T ,

completing the proof.There is one problem with the foregoing sketch. If the original curve winds

tightly about (or, worse, passes through) one of the centers of radial projection,the mass of its projection could be an order larger than its own mass. In thiscase, one moves the curve a bit before starting the whole process. If the originalcurve winds throughout space, it may be impossible to move it away from thecenters of projection, but the average distortion can still be controlled.

5.2. Corollary. The set

T D fT 2 Im : sptT ² Bn�0,c1 ,M�T � c2, and M�∂T � c3g

is totally bounded under F.

Proof. Each T 2 T can be well approximated by a polyhedral chain P inthe ε-grid with M�P � [c2 C εc3] and spt P ² Bn�0,c1 C 2nε by 5.1(4).Since there are only finitely many such P, T is totally bounded.


5.3. The Isoperimetric Inequality [Federer, 4.2.10]. If T 2 ImRn with∂T D 0, then there exists S 2 ImC1Rn with ∂S D T and

M�S m/�mC1 � M�T .

Here D 2n2mC2 as in the Deformation Theorem.

Remarks. That T bounds some rectifiable current S is shown by taking thecone over T. The value of the isoperimetric inequality lies in the numericalestimate on M(S). It was long conjectured that the worst case (exhibitingthe best constant) was the sphere, in all dimensions and codimensions. Thisconjecture was proven in 1986 by Almgren [2]. For merely stationary surfaces,an isoperimetric inequality still holds [Allard, § 7.1], but the sharp constantremains conjectural, even for minimal surfaces in R3. For more general(bounded) integrands than area, an isoperimetric inequality follows triviallyfor minimizers, but remains conjectural for stationary surfaces, even in R3.

The Isoperimetric Theorem 5.3 extends from Rn to a smooth compactRiemannian manifold M. One must assume that T is a boundary, and theconstant depends on M. A cycle of small mass is a boundary. These exten-sions are special cases of the isoperimetric inequality of 12.3, with B D ∅.

For more on these results, see Chapter 17.

The proof of 5.3 is a bizarre application of the Deformation Theorem,T D PC QC ∂S. One chooses ε large, a grid large enough to force the approx-imation P to T to be 0. See Figure 5.3.1.

Proof. Choose ε so that M�T D εm, and apply the Deformation Theoremto obtain T D PC QC ∂S. Since ∂T D 0, Q D 0. By 5.1(2), M�P � M�T .But, since P lies in the 2ε grid, M(P) must be a multiple of �2ε m, whichexceeds M�T by choice of ε. Therefore P D 0. Now T D ∂S, and, by 5.1(2),

M�S � εM�T D εmC1 D [M�T ]�mC1 /m.

5.4. The Closure Theorem [Federer, 4.2.16].

(1) Im is F closed in Nm,(2) ImC1 D fT 2 RmC1: M�∂T <1g,(3) Rm D fT 2 Fm: M�T <1g.

Consequently,(4) T D fT 2 Im: sptT ² Bn�0,R , M�T � c, and M�∂T � cg, is F complete.

Remarks. This result, specifically (1), is the hard part of the CompactnessTheorem. It depends on Lemma 4.12, the characterization of rectifiable sets


ε

S

TS

Figure 5.3.1. In the proof of the Isoperimetric Inequality, T is projected onto a largegrid.

by rectifiable slices. The original proof of Lemma 4.12 used structure theory.However, in 1986 White [3] found a simpler and more direct argument.

It follows directly from the definition of Fm that fT 2 Fm: sptT ² B�0,R gis F complete. Consequence (4) now follows by the lower semicontinuity ofM (Exercise 4.22) and the continuity of ∂.

Proof. We leave it as an exercise (5.2) to check that for each m, �1 )�2 ) �3 . Hence to prove (1), (2), and (3), it suffices to prove (1) for m,assuming all three conclusions for m � 1.

To prove (1), suppose that a sequence of integral currents Qi 2 Im convergesin the F norm to a normal current T 2 Nm. We must show that T 2 Im.

By induction, we may assume that ∂T 2 Im�1. By replacing T with T� T1,where T1 2 Im has the same boundary as T, we may assume that ∂T D 0.

By Lemma 4.12, it suffices to show that for all points p 2 Rn, for almostevery positive real number r, the slice ∂�T B�p, r is rectifiable. We mayassume that the Qi’s converge so rapidly that

F�Qi � T <1.


Thence by slicing theory 4.11(6), for 0 < a < b,

∫ b

aF[�Qi � T B�p, r ]dr <1.

Therefore Qi B�p, r F! T B�p, r and hence

∂�Qi B�p, r F! ∂�T B�p, r

for almost every r. Recall that by slicing theory 4.11(5), ∂�Qi B�p, r and∂�T B�p, r , and hence of course Qi B�p, r and T B�p, r , are normalcurrents. By induction on (2) above, Qi B�p, r and hence ∂�Qi B�p, r areintegral currents. By induction on (1), ∂�T B�p, r is an integral current (foralmost all r). Now by Lemma 4.12, T is a rectifiable current. Since ∂T D 0,T is an integral current, and (1) is proved. As mentioned, (2) and (3) follow.

To prove (4), let Tj be a Cauchy sequence in T. By the completeness offT 2 Fm: sptT ² B�0,R g, there is a limit T 2 Fm. By the lower semicontinuityof mass (Exercise 4.22), M�T � c and M�∂T � c. Finally by (3), T 2 Im.

5.5. The Compactness Theorem [Federer, 4.2.17]. Let K be a closed ballin Rn, 0 � c <1. Then

fT 2 ImRn: sptT ² K,M�T � c, and M�∂T � cg

is F compact.

Remark. More generally, K may be a compact C1 submanifold of Rn or acompact Lipschitz neighborhood retract (cf. Federer [1, 4.1.29]), yielding theCompactness Theorem in any C1 compact Riemannian manifold M. (Any C1

embedding of M in Rn, whether or not isometric, will do, since altering themetric only changes the flat norm by a bounded amount and does not changethe topology, as Brian White pointed out to me.)

Proof. The set is complete and totally bounded by the ClosureTheorem, 5.4, and the Deformation Theorem Corollary, 5.2.

As an example of the power of the Compactness Theorem, we prove thefollowing corollary.

5.6. The Existence of Area-Minimizing Surfaces. Let B be an (m � 1)-dimensional rectifiable current in Rn with ∂B D 0. Then there is an m-dimensional area-minimizing rectifiable current S with ∂S D B.


Remarks. S area minimizing means that, for any rectifiable current T with∂T D ∂S, M�S � M�T . That B bounds some rectifiable current is shown bytaking the cone over B. Even if B is a submanifold, S is not in general.

Proof. Let B�0,A be a large ball containing spt B. Let Sj be a sequenceof rectifiable currents with areas decreasing to inffM�S : ∂S D Bg. The firstproblem is that the spt Sj may send tentacles out to infinity. Let denote theLipschitz map which leaves the ball B�0,A fixed and radially projects pointsoutside the ball onto the surface of the ball (Figure 5.6.1). is distancenonincreasing and hence area nonincreasing. Therefore by replacing Sj by#Sj, we may assume that spt Sj ² B�0,A .

Π

ΠΠ

Π Π

B

B (0, A)

Figure 5.6.1. In the proof of the existence of an area-minimizing surface with givenboundary B, it is necessary to keep surfaces under consideration inside some largeball B�0,A . This is accomplished by projecting everything outside the ball onto itssurface.

Now using the Compactness Theorem we may extract a subsequence whichconverges to a rectifiable current S. By the continuity of ∂ and the lowersemicontinuity of mass (Exercise 4.22), ∂S D B and M�S D inffM�T : ∂T DBg. Therefore S is the desired area-minimizing surface.


5.7. The Existence of Absolutely and Homologically Minimizing Surfacesin Manifolds [Federer, 5.1.6]. Let M be a compact, C1 Riemannian mani-fold. Let T be a rectifiable current in M. Then among all rectifiable currents Sin M such that ∂S D ∂T (respectively, S� T D ∂X for some rectifiable currentX in M), there is one of least area.

S is called absolutely or homologically area minimizing. The methods alsotreat free boundary problems (cf. 12.3).

Proof. Given the Remark after the Compactness Theorem, 5.5, we justneed to check that a minimizing limit stays in the same homology class. IfF �Si � S is small, Si � S D AC ∂B, with M(A) and M(B) small. Let Y1 be thearea minimizer in Rn with ∂Y1 D A. Since M�Y1 is small, by monotonicity9.5, Y1 stays close to M and hence may be retracted onto Y in M. SinceSi � S D ∂YC ∂B, therefore S ¾ Si, as desired.

EXERCISES

5.1. Verify that the Isoperimetric Inequality, 5.3, is homothetically invariant. (A homo-thety &r of Rn maps x to rx.)

5.2. Prove that 5.4�1 ) �2 ) �3 .5.3. Try to find a counterexample to the Closure Theorem, 5.4. (Of course, there is

none.)5.4. Show that the Isoperimetric Inequality, 5.3, fails for normal currents.

CHAPTER 6

Examples of Area-Minimizing Surfaces

It can be quite hard to prove that any particular surface is area minimizing.After all, it must compare favorably with all other surfaces with the givenboundary. Fortunately, there are a number of beautiful examples.

6.1. The Minimal Surface Equation [Federer, 5.4.18]. Let f be a C2,real-valued function on a planar domain D, such that the graph of f is areaminimizing. Then F satisfies the minimal surface equation:

�1 C f2y �fxx � 2fxfyfxy C �1 C f2

x �fyy D 0.

Conversely, if f satisfies the minimal surface equation on a convex domain,then its graph is area minimizing.

The minimal surface equation just gives the necessary condition that undersmooth variations in the surface, the rate of change of the area is 0. Thiscondition turns out to be equivalent to the vanishing of the mean curvature.A smoothly immersed surface which is locally the graph of a solution tothe minimal surface equation (or, equivalently, which has mean curvature 0)is called a minimal surface. Some famous minimal surfaces are pictured inFigures 6.1.1–6.1.3b.

Theorem 6.1 guarantees that small pieces of minimal surfaces are area mini-mizing, but larger pieces may not be. For example, the portion of Enneper’ssurface pictured in Figure 6.1.2 is not area minimizing. There are two area-minimizing surfaces with the same boundary, pictured in Figure 6.1.4. Somesystems of curves in R3 bound infinitely many minimal surfaces. See Morgan[6] and references therein or the popular articles by Morgan [21, 18].

67


The plane

The catenoid√x2 + y2 = coshzEuler,1740

Scherk's surface,1835

ez cosy = cosxThe helicoidy tan z = xMeusnier,1776

Figure 6.1.1. Some famous minimal surfaces.

On the disc (or any other convex domain), there is a solution of the minimalsurface equation with any given continuous boundary values. We omit theproof. Because the minimal surface equation is not linear, this fact is not atall obvious, and it fails if the domain is not convex. Moreover, on nonconvex

Examples of Area-Minimizing Surfaces 69

x = Re(w − w3)y = Re(i (w + w3))z = Re(w 2) w ∈C

13

13

Figure 6.1.2. Enneper’s surface, 1864.

domains there can be solutions of the minimal surface equation whose graphsare not area minimizing (Figure 6.1.5).

The first part of the proof of 6.1 will afford an opportunity to illustratethe classical method and notation of the calculus of variations. In general,begin with a function f: D ! Rm, supposed to maximize or minimize somefunctional A�f� for the given boundary values fj∂D. Consider infinitesimalchanges υf in f, with υfj∂D D 0. Set to 0 the corresponding change in A,called the first variation υA. Use integration by parts to obtain an equation ofthe form ∫

DG�f� Ð υf D 0.

Since this equation holds for all admissible υf, it follows immediately thatG�f� D 0. This turns out to be a differential equation for f, called theEuler–Lagrange equation. When A�f� is the area of the graph of f, theassociated Euler–Lagrange equation is the minimal surface equation.

Proof of 6.1. First we will show that if the graph of f is area minimizing,then f satisfies the minimal surface equation. By hypothesis, f minimizes the


Figure 6.1.3a. The first modern complete, embedded minimal surface of Costa andHoffman and Meeks (also see Hoffman). Courtesy of David Hoffman, Jim Hoffman,and Michael Callahan.

area functional

A�f� D∫

D�1 C f2

x C f2y �1/2 dx dy

for given boundary values. Therefore, the variation υA in A due to an infinites-imal, smooth variation υf in f, with υfj∂D D 0, must vanish:

0 D υA D∫

D

1

2�1 C f2

x C f2y ��1/2�2fxυfx C 2fyυfy� dx dy

D∫

D�[�1 C f2

x C f2y ��1/2fx]υfx C [�1 C f2

x C f2y ��1/2fy]υfy� dx dy.


Figure 6.1.3b. One of the latest new complete, embedded minimal surfaces: thegenus one helicoid, discovered by David Hoffman, Hermann Karcher, and FushengWei (1993). Computer-generated image made by James T. Hoffman at the GANGLaboratory, University of Massachusetts, Amherst. Copyright GANG, 1993.

Figure 6.1.4. Area-minimizing surfaces with the same boundary as Enneper’s surface.


C C

Figure 6.1.5. A minimal graph over a nonconvex region C need not be area mini-mizing. The second surface has less area.

Integration by parts now yields an equation of the form∫

D G�f�υf D 0,where

G�f� D � ∂

∂x[�1 C f2

x C f2y ��1/2fx] � ∂

∂y[�1 C f2

x C f2y ��1/2fy]

D 1

2�1 C f2

x C f2y ��3/2�2fxfxx C 2fyfxy�fx

� �1 C f2x C f2

y ��1/2fxx

C 1

2�1 C f2

x C f2y ��3/2�2fxfxy C 2fyfyy�fy

� �1 C f2x C f2

y ��1/2fyy

D ��1 C f2x C f2

y ��3/2[�1 C f2y �fxx � 2fxfyfxy C �1 C f2

x �fyy].

Since∫

D G�f�υf D 0 for all smooth υf satisfying υfj∂D D 0, G�f� D 0, i.e.,

�1 C f2y �fxx � 2fxfyfxy C �1 C f2

x �fyy D 0,

the minimal surface equation.Second, we prove that the graph over a convex domain of a solution to the

minimal surface equation is area minimizing. This proof uses a calibration,i.e., a differential form ϕ which is closed (dϕ D 0) and has maximum valuesup ϕ�� D 1 as a function on the set of unit k-planes �. A surface is said to be


calibrated by ϕ if each oriented tangent plane � satisfies ϕ�� D 1. The methodshows that a calibrated surface is automatically area minimizing.

Given f: D ! R, define a 2-form ϕ on D ð R by

ϕ�x, y, z� D �fx dy dz � fy dz dx C dx dy√f2

x C f2y C 1

.

Then for any point �x, y, z� 2 D ð R and for any unit 2-vector �, ϕ�� 1,with equality if � is tangent to the graph of f at (x, y, z). Moreover, ϕ isclosed:

dϕ D � ∂

∂xfx�f

2x C f2

y C 1��1/2 � ∂

∂yfy�f2

x C f2y C 1��1/2 dx dy dz

D ��f2x C f2

y C 1��3/2��1 C f2y �fxx � 2fxfyfxy C �1 C f2

x �fyy� dx dy dz

D 0

by the minimal surface equation.Now let S denote the graph of f, and let T be any other rectifiable current

with the same boundary. Since D is convex, we may assume spt T ² D ð R Ddomain ϕ, by projecting T into D ð R if necessary without increasing area T.Now, since ϕ�� D 1 whenever � is tangent to S,

area S D∫

Sϕ.

Since S � T bounds and ϕ is closed,

∫S

ϕ D∫

Tϕ.

Since ϕ�� 1 for all 2-planes �,

∫T

ϕ area T.

Combining the two equations and the inequality yields

area S area T.

Therefore S is area minimizing, as desired.


Remark. The same argument with the same calibration shows that if thegraph of f has constant mean curvature, then it minimizes area for fixedvolume constraint.

6.2. Remarks on Higher Dimensions. For a function f: Rn�1 ! R, theminimal surface equation takes the form

(1) divrf√

1 C jrfj2 D 0.

The statement and proof of 6.1 apply virtually unchanged.In higher codimension there is a minimal surface system. For example, for

a function f: R2 ! Rn�2, the minimal surface system is

(2) �1 C jfyj2�fxx � 2�fx Ð fy�fxy C �1 C jfxj2�fyy D 0.

In general codimension, the graph of f need not be area minimizing, evenif the domain is convex. (In an attempted generalization of the Proof of6.1, ϕ generally would not be closed. For a counterexample, see Lawson andOsserman.)

6.3. Complex Analytic Varieties [Federer, 5.4.19]. Any compact portionof a complex analytic variety in Cn ¾D R2n is area minimizing.

With this initially astonishing fact, Federer provided some of the first examplesof area-minimizing surfaces with singularities. For example, fw2 D z3g ² C2 hasan isolated “branch-point” singularity at 0. Similarly, the union of the complexaxis planes fz D 0g and fw D 0g in C2 has an isolated singularity at 0. These areprobably the simplest examples. fw D šp

zg is regular at 0; it is the graph ofz D w2. The proof, as that of the second part of 6.1, uses a “calibration” ϕ.

Proof. On Cn D Rn ý Rn, let ω be the Kahler form

ω D dx1 ^ dy1 C Ð Ð Ð C dxn ^ dyn.

Wirtinger’s Inequality [Federer, 1.8.2, p. 40] says that the real 2p-form ϕ Dωp/p! satisfies

jϕ��j 1

for every 2p-plane �, with equality if and only if � is a complex p-plane.Now let S be a compact portion of a p-dimensional complex analytic

variety, and let T be any other 2p-dimensional rectifiable current with the


same boundary. Since S is complex analytic, ϕ�� is 1 on every plane tangentto S, and

area S D∫

Sϕ.

Since ∂S D ∂T, S � T is a boundary. Of course dϕ D 0, because ϕ has constantcoefficients. Therefore ∫

Sϕ D

∫T

ϕ.

Finally, since ϕ�� is always at most 1,

∫T

ϕ area T.

Combining the equalities and inequality yields

area S area T.

We conclude that S is area minimizing, as desired.The second part of the Proof of 6.1 and 6.3 proved area minimization by

means of a calibration or closed differential form of comass 1, as examplesof the following method.

6.4. Fundamental Theorem of Calibrations. Let ϕ be a closed differentialform of unit comass in Rn or in any smooth Riemannian manifold M. Let Sbe an integral current such that hES, ϕi D 1 at almost all points of S. In Rn,S is area minimizing for its boundary. In any M, S is area minimizing in itshomology class (with or without boundary).

Proof. Let T be any comparison surface. Then

area S D∫

Sϕ D

∫T

ϕ area T.

6.5. History of Calibrations (cf. Morgan [1, 2]). The original exampleof complex analytic varieties was implicit in Wirtinger (1936), explicit forcomplex analytic submanifolds in de Rham [1] (1957), and applied to singularcomplex varieties in the context of rectifiable currents by Federer [3, §4](1965). Berger [2, §6, last paragraph] (1970) was the first to extract the under-lying principle and apply it to other examples such as quaternionic varieties,followed by Dao (1977). The term calibration was coined in the landmarkpaper of Harvey and Lawson, which discovered rich new calibrated geometriesof “special Lagrangian,” “associative,” and “Cayley” varieties.


The method has grown in power and applications. Surveys appear in Morgan[1, 2]. Mackenzie and Lawlor use calibrations in the proof (Nance; Lawlor[1]) of the angle conjecture on when a pair of m-planes in Rn is area mini-mizing. The “vanishing calibrations” of Lawlor [3] actually provide sufficientdifferential-geometric conditions for area minimization, a classification ofall area-minimizing cones over products of m spheres, examples of nonori-entable area-minimizing cones, and singularities stable under perturbations.The “paired calibrations” of Lawlor and Morgan [2] and of Brakke [1, 2] andthe covering space calibrations of Brakke [3] prove new examples of soapfilms in R3, in R4, and above. Other developments include Murdoch’s “twistedcalibrations” of nonorientable surfaces, Le’s “relative calibrations” of stablesurfaces, and Pontryagin calibrations on Grassmannians (Gluck, Mackenzie,and Morgan).

Lawlor [2] has developed a related theory for proving minimization byslicing. Lawlor and Morgan [1] show, for example, that three minimal surfacesmeeting at 120 degrees minimize area locally.

EXERCISES

6.1. Verify that the helicoid is a minimal surface.6.2. Verify that Scherk’s surface is a minimal surface.6.3. Verify that Enneper’s surface is a minimal surface.

6.4. Prove that the catenoids√

x2 C y2 D 1

acosh az are the only smooth, minimal

surfaces of revolution in R3.6.5. Verify that for n D 3 the minimal surface equation, 6.2(1), reduces to that of 6.1.6.6. Use 6.2(2) to verify that the complex analytic variety

fw D z2g ² C2

is a minimal surface.6.7. Use 6.2(2) to prove that the graph of any complex analytic function g: C ! C is

a minimal surface.

CHAPTER 7

The Approximation Theorem

The Approximation Theorem says that an integral current, T, can be approx-imated by a slight diffeomorphism of a polyhedral chain, P, or, equivalently,that a slight diffeomorphism f#T of T can be approximated by P itself. (SeeFigure 7.1.1.) The approximation P actually coincides with f#T except for anerror term, E, of small mass.

7.1. The Approximation Theorem [Federer, 4.2.20]. Given an integralcurrent T 2 ImRn and ε > 0, there exist a polyhedral chain P 2 PmRn, sup-ported within a distance ε of the support of T, and a C1 diffeomorphism f ofRn such that

f#T D P C E

with ME� � ε, M∂E� � ε, Lipf� � 1 C ε, Lipf�1� � 1 C ε, jfx� � xj �ε, and fx� D x whenever dist x, spt T� � ε.

Proof.

CASE 1. ∂T polyhedral. Since T is rectifiable, T D Hm B� ^ �, with B

rectifiable and j�j integer valued. By Proposition 3.11, Hm almost all of B is

contained in a countable union [ Mi of disjoint C1 embedded manifolds. Atalmost every point x 2 B, the density of B and of [ Mi is 1 (Proposition 3.12),so that there is a single Mi such that B and Mi coincide at x except for a setof density 0. Now a covering argument produces a finite collection of disjointopen balls Ui ² Rn � spt ∂T and nearly flat C1 submanifolds Ni of Ui such

77


T

f#T

P

Figure 7.1.1. The Approximation Theorem yields a diffeomorphism f#T of T whichcoincides with a polyhedral P except for small measure.

that [ Ni coincides with B except for a set of small jjTjj measure. Gentlehammering inside each Ui, smoothed at the edges, is the desired diffeomor-phism f of Rn, which flattens most of B into m-dimensional planes, wheref#T can be M approximated by a polyhedral chain P1 (cf. Proof of 4.4).

Unfortunately the error f#T � P1, although small in mass, may have hugeboundary. Now we use the Deformation Theorem, 5.1, to decompose the error

f#T � P1 D P2 C Q C ∂S.

As usual, Q and S have small mass. But in this case, because f#T � P1 hassmall mass, so does P2 and hence so does the remaining term, ∂S. Becausef leaves spt ∂T fixed, ∂f#T � P1� D ∂T � ∂P1 is polyhedral, and hence theDeformation Theorem construction makes Q polyhedral. Now take

P D P1 C P2 C Q.

Then f#T � P D ∂S has small mass and no boundary.

The Approximation Theorem 79

GENERAL CASE. When ∂T is not polyhedral, first approximate ∂T by thefirst case,

f1#∂T D P1 C ∂S1,

with MS1� and M∂S1� small. Now f1#T � S1 has polyhedral boundary andcan be approximated by an integral polyhedral chain, P2,

f2#f1#T � S1� D P2 C ∂S2,

with M∂S2� small. Therefore

f2 ° f1�#T D P2 C f2#S1 C ∂S2�,

and the error term and its boundary have small mass, as desired.

CHAPTER 8

Survey of Regularity Results

In 1962 Wendell Fleming proved a regularity result that at first sounds toogood to be true.

8.1. Theorem [Fleming, 2]. A two-dimensional, area-minimizing rectifiablecurrent T in R3 is a smooth, embedded manifold on the interior.

More precisely, spt T � spt ∂T is a C1 embedded manifold.In the classical theory, such complete regularity fails. The disc of least

mapping area with given boundary is an immersed minimal surface (seeNitsche [2, §365, p. 318], Osserman [2, Appendix 3, §1, p. 143], or Lawson[§3, p. 76]), but not in general embedded. For the boundary pictured inFigure 8.1.1, a circle with a tail, the area-minimizing rectifiable current hashigher genus, has less area, and is embedded. Pictured in Figure 8.1.2, it flowsfrom the top, flows down the tail, pans out in back onto the disc, flows aroundfront, and flows down the tail to the bottom. There is a hole in the middle thatyou can stick your finger through. Incidentally, this surface exists as a soapfilm, whereas the least-area disc does not.

The regularity theorem 8.1 was generalized to three-dimensional surfaces inR4 by Almgren [6] in 1966 and up through six-dimensional surfaces in R7 bySimons in 1968. In 1969, Bombieri, De Giorgi, and Giusti gave an example ofa seven-dimensional, area-minimizing rectifiable current in R8 with an isolatedinterior singularity. Chapter 10 gives a short discussion of this counterexampleas well as an outline of the proof of the regularity results.

81


Figure 8.1.1. A least-area disc need not be embedded.

Figure 8.1.2. The area-minimizing rectifiable current is embedded.

Survey of Regularity Results 83

The complete interior regularity results for area-minimizing hypersurfacesare given by the following theorem of Federer.

8.2. Theorem [Federer 2]. An �n � 1�-dimensional, area-minimizing recti-fiable current T in Rn is a smooth, embedded manifold on the interior exceptfor a singular set of Hausdorff dimension at most n � 8.

Regularity in higher codimension, for an m-dimensional area-minimizingrectifiable current T in Rn, with m < n � 1, is much harder. Until 1983 it wasknown only that the set of regular points, where spt T is a smooth embeddedmanifold, was dense in spt T � spt ∂T [Federer, 5.3.16]. On the other hand,m-dimensional complex analytic varieties, which are automatically area mini-mizing (6.3), can have (m � 2)-dimensional singular sets. In a major advanceAlmgren proved the conclusive regularity theorem.

8.3. Theorem [Almgren 3, 1983]. An m-dimensional, area-minimizingrectifiable current in Rn is a smooth, embedded manifold on the interior exceptfor a singular set of Hausdorff dimension at most m � 2.

For example, a two-dimensional area-minimizing rectifiable current in Rn

has at worst a zero-dimensional interior singular set. In 1988 Sheldon Chang(Figure 8.3.1) proved that these singularities must be isolated, “classical branchpoints.”

Figure 8.3.1. Graduate student Sheldon Chang sharpened Fred Almgren’s regularityresults for two-dimensional area-minimizing surfaces. Pictured outside the Almgrenhouse in Princeton. Photo courtesy Jean Taylor.


The stronger regularity theory in codimension 1 comes from an elementaryreduction to the relatively easy case of surfaces of multiplicity 1. Indeed,a nesting lemma decomposes an area-minimizing hypersurface into nested,multiplicity-1 area-minimizing surfaces, for which strong regularity resultshold. If these surfaces touch, they must coincide by a maximum principle.(See Section 8.5 and Chapter 10.)

In general dimensions and codimensions, very little is known about thestructure of the set S of singularities. One might hope that S stratifies intoembedded manifolds of various dimensions. However, for all we know todate, S could even be fractional dimensional.

8.4. Boundary Regularity. In 1979 Hardt and Simon proved the conclusiveboundary regularity theorem for area-minimizing hypersurfaces.

THEOREM [Hardt and Simon, 1]. Let T be an �n � 1�-dimensional, area-minimizing rectifiable current in Rn, bounded by a C2, oriented submanifold�with multiplicity 1�. Then at every boundary point, spt T is a C1, embeddedmanifold-with-boundary.

Notice that general regularity is stronger at the boundary than on the interior,where there can be an (n � 8)-dimensional singular set. The conclusion ismeant to include the possibility that spt T is an embedded manifold withoutboundary at some boundary points, as occurs for example if the given boundaryis two concentric, similarly oriented planar circles. (See Figure 8.4.1.)

Multiplicity 2

Multiplicity 1

Figure 8.4.1. This example shows that part of the given boundary can end up on theinterior of the area-minimizing surface T.

The surface of Figure 8.1.2 at first glance may seem to have a boundarysingularity where the tail passes through the disc. Actually, as you move downthe tail, the inward conormal rotates smoothly and rapidly almost a full 360°.

B. White [5] has generalized regularity to smooth boundaries with multi-plicities.

Survey of Regularity Results 85

In higher codimension, boundary singularities can occur.For the classical least-area disc in R3, it remains an open question whether

there can be boundary branch points if the boundary is C1 but not real analytic(see Nitsche [2, §366, p. 320]).

8.5. General Ambients and Other Integrands. All of the above regu-larity results continue to hold in smooth Riemannian manifolds other than Rn.Federer [1, 5.4.15, 5.4.4] gives the proof of Theorem 8.2 for Rn and explainsthat the methods can be generalized to any smooth Riemannian manifold,except possibly a maximum principle. Because two regular minimal surfaces inlocal coordinates are graphs of functions satisfying a certain nonlinear, elliptic,partial differential system (see Morgan [7, §2.2] or Federer [1, 5.1.11]), sucha maximum principle is standard, due essentially to Hopf [Satz 10; see alsoSerrin, p. 184].

Incidentally, Schoen, Simon, and Yau, as well as Schoen and Simon, haveproved regularity results for hypersurfaces that are stable but not necessarilyarea minimizing in Riemannian manifolds of dimension less than eight.

Regularity for surfaces minimizing more general integrands than area isharder, and the results are weaker. For a smooth elliptic integrand (see 12.5)on an n-dimensional smooth Riemannian manifold, an m-dimensional -minimizing rectifiable current is a smooth embedded manifold on an open,dense set [Federer 1, 5.3.17]. If m D n � 1, the interior singular set has dimen-sion less than m � 2 (Schoen, Simon, and Almgren [II.7, II.9] plus an addi-tional unpublished argument of Almgren as in White [7, Theorem 5.2]). Theachievement of such general regularity results is one of the telling strengthsof geometric measure theory.

The regularity theory extends to hypersurfaces minimizing area subject tovolume constraints.

8.6. Theorem [Gonzalez, Massari, and Tamanini, Theorem 2]. Let T bean �n � 1�-dimensional rectifiable current of least area in the unit ball B ² Rn,with prescribed boundary in ∂B, bounding an oriented region of prescribedvolume.

Then T is a smooth hypersurface of constant mean curvature on the interior,except for a singular set of dimension at most n � 8.

Gonzalez, Massari, and Tamanini state the result in the equivalent termi-nology of sets of finite perimeter. Of course once smooth, T must have constantmean curvature by a simple variational argument.

By the methods of Almgren [1], the result also holds in a smooth Rieman-nian manifold, although there seems to be no clear statement in the literature.


EXERCISES

8.1. Try to come up with a counterexample to Theorems 8.1 and 8.4.8.2. Try to draw an area-minimizing rectifiable current bounded by the trefoil knot.

(Make sure your surface is orientable.)

8.3. Illustrate the sharpness of Theorem 8.3. by proving directly that the union of twoorthogonal unit discs about 0 in R4 is area minimizing.Hint : First show that the area of any surface S is at least the sum of the areas ofits projections into the x1-x2 and x3-x4 planes.

CHAPTER 9

Monotonicity and OrientedTangent Cones

This chapter introduces the two basic tools of the regularity theory of area-minimizing surfaces. Section 9.2 presents the monotonicity of the mass ratio, alower bound on area growth. Section 9.3 presents the existence of an orientedtangent cone at every interior point of an area-minimizing surface.

9.1. Locally Integral Flat Chains [Federer, 4.1.24, 4.3.16]. We need togeneralize our definitions to include noncompact surfaces such as orientedplanes. First, define the space F

locm of locally integral flat chains as currents

which locally coincide with integral flat chains:

Flocm D fT 2 Dm: for all x 2 Rn there exists S 2 Fm with x /2 spt�T � S�g.

For the local flat topology, a typical neighborhood Uυ of 0 takes the form

Uυ D fT 2 Flocm : spt�T � �A C ∂B�� \ U�0, R� D ∅,

A 2 Rm, B 2 RmC1,M�A� C M�B� < υg,where U(0, R) is the open ball fx 2 Rn: jxj < Rg.

The subspaces Ilocm ² R

locm ² F

locm of locally integral currents and locally

rectifiable currents are defined analogously:

Ilocm D fT 2 Dm: for all x 2 Rn there exists S 2 Im with x /2 spt�T � S�g.

Rlocm D fT 2 Dm: for all x 2 Rn there exists S 2 Rm with x /2 spt�T � S�g.

87


Alternative definitions of the locally rectifiable currents are given by

Rlocm D fT 2 Dm:T B�0, R� 2 Rm for all Rg

D fT 2 Dm:T B�a, R� 2 Rm for all a and all Rg.There are no similar alternatives for locally integral currents. Indeed,Figure 9.1.1 shows an integral current T such that T B�0, 1� is not an integralcurrent, because M�∂�T B�0, 1�� D C1.

Figure 9.1.1. T is a single curve with two endpoints. However, its restriction to theinside of the circle has infinitely many pieces and infinitely many endpoints. Hencethe restriction of an integral current need not be an integral current.

A locally rectifiable current T is called area minimizing if for all a andR,T B�a, R� is area minimizing.

The Compactness Theorem, 5.5, generalizes to unbounded locally integralcurrents in Rn or noncompact manifolds, as nicely explained in Simon [3,27.3, 31.2]. Without this perspective, Federer [1] must be forever judiciouslyrestricting currents, a troublesome complication.

9.2. Monotonicity of the Mass Ratio. For a locally rectifiable current T 2R

locm and a point a 2 Rn, define the mass ratio

m�T, a, r� D M�T B�a, r��/˛mrm,

where ˛m is the measure of the unit ball in Rm. Define the density of T at a,

m�T, a� D limr!0

m�T, a, r�.

The following theorem on the monotonicity of the mass ratio is one of themost useful tools in regularity theory.

9.3. Theorem [Federer, 5.4.3]. Let T be an area-minimizing locallyrectifiable current in R

locm . Let a lie in the support of T. Then for 0 <

r < dist�a, spt ∂T�, the mass ratio �T, a, r� is a monotonically increasingfunction of r.

Monotonicity and Oriented Tangent Cones 89

Remark. If the mass ratio is constant, then T is a cone (invariant underhomothetic expansions).

Monotonicity for Minimal Surfaces and Other Integrands. Monotonicityactually holds for stationary (minimal) surfaces or, in a weakened form, evenfor bounded mean curvature surfaces, but not for more general integrandsthan area [Allard, 5.1]. (Nevertheless 9.5 holds with a smaller constant forminimizers of general integrands.)

Before giving the Proof of 9.3, we state two immediate corollaries.

9.4. Corollary. Suppose T 2 Rlocm is area minimizing. Then m�T, a� exists

for every a 2 sptT � spt ∂T.

9.5. Corollary. Suppose T 2 Rlocm is area minimizing and a 2 sptT. Then

for 0 < r < dist�a, spt ∂T�,

M�T B�a, r�� ½ m�T, a� Ð ˛mrm.

For example, if furthermore T happens to be an embedded, two-dimensional,oriented manifold-with-boundary, then

M�T B�a, r�� ½ �r2.

See Figures 9.5.1 and 9.5.2.

r

Figure 9.5.1. The disc D is area minimizing and M�D B�0, r�� D �r2.

Proof of Theorem 9.3. For 0 < r < dist�a, spt ∂T�, let f�r� denoteM�T B�a, r��. Since f is monotonically increasing, for almost all r, f0�r�exists. Slicing by the function u�x� D jx � aj yields (4.11(3))

M�∂�T B�a, r�� f0�r�.

Since T is area minimizing, M�T B�a, r�� is less than or equal to the areaof the cone C over ∂�T B�a, r�� to a (Figure 9.5.3).


r

Figure 9.5.2. An area-minimizing surface with M�T B�0, r�� > �r2.

T C

Figure 9.5.3. If T B�a, r� is area minimizing, then it must of course have less areathan the cone C over its boundary.

By Exercise 3.8, M�C� D r

mM�∂�T B�a, r��. Assembling these inequalities

yields

f�r� � M�C� D r

mM�∂�T B�a, r�� r

mf0�r�.

Consequently,

d

dr˛m�T, a, r� D d

dr[r�mf�r�] D r�mf0�r� � mr�m�1f�r�

D m

rm�1

[ rmf0�r� � f�r�

]½ 0.

Hence the absolutely continuous part of �T, a, r� is increasing. Since anysingular part is due to increases in f,�T, a, r� is increasing as desired.

9.6. Corollary. Let T be an area-minimizing rectifiable current in RmRn.Then for all a 2 sptT � spt ∂T,m�T, a� ½ 1.

Proof. Since a rectifiable set has density 1 almost everywhere (3.12),there is a sequence of points aj ! a with m�T, aj� ½ 1. Let 0 < r <


dist�a, spt ∂T�, rj D dist�a, aj�. Obviously

M�T B�a, r�� ½ M�T B�aj, r � rj��.

But by monotonicity, M�T B�aj, r � rj�� ½ ˛m�r � rj�m. ConsequentlyM�T B�a, r�� ½ ˛mrm and �T, a� ½ 1.

9.7. Oriented Tangent Cones [Federer, 4.3.16]. We now develop a gener-alization to locally integral flat chains of the notion of the tangent plane to aC1 manifold at a point. (See Figure 9.7.1.)

0 0

Figure 9.7.1. The surface of the unit cube and the three quarter planes constitutingits oriented tangent cone at 0.

DEFINITIONS. A locally integral flat chain C is called a cone if every homo-thetic expansion or contraction �R#C D C. If T 2 F

locm , such a cone C is

called an oriented tangent cone to T at 0 if there is a decreasing sequencer1 > r2 > r3 > Ð Ð Ð tending to 0 such that �r�1

j #T converges to C in the local

flat topology. Note that an oriented tangent cone C is a current, whereas atangent cone Tan�E, 0� as defined in Section 3.9 is a set. In general, sptC ²Tan�sptT, 0�, but equality need not hold (cf. Exercise 9.6).

Remarks. Figure 9.7.2 illustrates that an oriented tangent cone is notnecessarily unique. As it approaches 0, this curve alternates between followingthe x-axis and following the y-axis for successive epochs.

In fact, one of the big open questions in geometric measure theory is whetheran area-minimizing rectifiable current T has a unique oriented tangent cone atevery point a 2 sptT � spt ∂T.

Figure 9.7.3 illustrates the need for specifying that C be a cone.

9.8. Theorem [Federer, 5.4.3(6)]. Let T be an area-minimizing rectifiablecurrent in Rm. Suppose 0 2 sptT � spt ∂T. Then T has an oriented tangentcone C at 0.


T

Figure 9.7.2. T alternates ad infinitum between the positive x-axis and the positivey-axis. Each axis is an oriented tangent cone.

T1

T2

Figure 9.7.3. T1 and T2 are both invariant under certain sequences of homotheticexpansions, but are not cones.

Because the proof is a bit technical, we will prove just this much: thereis a rectifiable current C supported in B(0, 1) and a sequence r1 > r2 >r3 > . . . tending to 0 such that the sequence �r�1

j #�T B�0, rj�� converges

to C. The hypotheses of the Compactness Theorem require bounds on bothM��r�1

j #�T B�0, rj�� and M��r�1j #∂�T B�0, rj��. Fix r0 < dist�0, spt ∂T�.

Then for r � r0, monotonicity of the mass ratio (9.3) says that

M��r�1#�T B�0, r�� D M�T B�0, r��r�m

� M�T B�0, r0��r�m0 � c,

the first desired bound. To get the second bound, slicing theory must beemployed to choose the sequence rj carefully. The theory (4.11(4)) says thatfor 0 < s < r0,

∫ s

s/2M�∂�T B�0, r��dr � M�T B�0, s�� csm.


Consequently, for some s/2 < r < s,

M�∂�T B�0, r�� csm

s/2� 2mcrm�1

andM��r�1#∂�T B�0, r�� 2mc.

Therefore a sequence of rj can be chosen satisfying both desired mass bounds.Now the Compactness Theorem guarantees that some subsequence convergesto the desired rectifiable current C, as we set out to prove.

A complete proof of the theorem involves considering not just B(0, 1),but a sequence of balls B�0, Rk� with Rk ! 1, successively applying theCompactness Theorem to extract subsequences convergent in each B�0, Rk�,and applying a diagonal argument. One difficulty comes in choosing the initialsequence r1 > r2 > r3 > . . . ! 0 such that the

limj!1

M��r�1j #�T B�0, rjRk�� < 1

for each k. The remark after 9.3 may be used to conclude that C is in fact acone.

9.9. Theorem. Let T be an area-minimizing rectifiable current in Rm.Suppose 0 2 sptT � spt ∂T. Let C be an oriented tangent cone to T at 0. Thenm�C, 0� D m�T, 0�.

Remark. Exercise 9.2 implies that C is itself area minimizing.

Proof. After replacing the sequence rj such that �r�1j #T ! C with a subse-

quence if necessary, for each j choose currents Aj, Bj such that

spt��r�1j #�T � �Aj C ∂Bj� \ U�0, 2� D ∅,

M�Aj� C M�Bj� � 1/j2.

Let u�x� D jx � aj and apply slicing theory 4.11(4) to choose 1 < sj < 1 C1/j such that

MhBj, u, sjCi � jM�Bj� � 1/j.

Note that

�r�1j #T B�0, sj� D C B�0, sj� C Aj B�0, sj�

C ∂�Bj B�0, sj�� hBj, u, sjCi.


Hence �r�1j #T B�0, sj� ! C B�0, 1� in the flat norm. By the lower semi-

continuity of mass (Exercise 4.22),

�C, 0� � �T, 0�.

Moreover, since �r�1j #TB�0, sj� is area minimizing and

C B�0, sj� C Aj B�0, sj� � hBj, u, sjCi

has the same boundary,

M��r�1j #T B�0, sj�� M�C B�0, sj�� C 2/j.

It follows that�T, 0� � �C, 0�.

EXERCISES

9.1. Give an example of an integral flat chain T 2 F 0R2 such that T B2�0, 1� is notan integral flat chain.

9.2. Let S1, S2, S3, . . . ! S be a convergent sequence of locally rectifiable currents.Suppose each Sj is area minimizing. Prove that S is area minimizing.

9.3. Let S be an area-minimizing rectifiable current in R 2R3 bounded by the circlesx2 � y2 D R2, z D š1 oppositely oriented. Prove that

sptS ²{√

x2 C y2 ½ R � 2pR}.

9.4. Prove or give a counterexample. If T 2 I2R3, then for all a 2 sptT �spt ∂T,2�T, a� ½ 1.

9.5. Let T be an m-dimensional area-minimizing rectifiable current in Rn and consider

f: Rn ! R,f�x� D m�T, x�.

(a) Mention an example for which f is not continuous, even on sptT � spt ∂T.(b) Prove that f is upper semicontinuous on Rn � spt ∂T.

9.6. Let T 2 Flocm , and let C be an oriented tangent cone to T at 0. Prove that sptC ²

Tan�sptT, 0� (cf. 3.9). Show by example that equality need not hold.9.7. Let T 2 F

locm , and consider oriented tangent cones to T at 0:C D lim�r�1

j #T and

D D lim�s�1j #T. Prove that, if 0 < lim sj/rj � lim sj/rj < 1, then C D D.


9.8. Let T be the polygonal curve which follows the x-axis from 1 D 2�02to 2�12

,then the y-axis from 2�12

to 2�22, and so on as in Figure 9.7.2, oriented outward.

(a) Show that the nonnegative x-axis and the nonnegative y-axis are each anoriented tangent cone at 0.

(b) Find a limit of homothetic expansions which is not a cone.(c) Let T0 D T\ x-axis. Show that the lower density Ł�T0, 0� D 0 while the

analogous upper density Ł�T0, 0� D 1/2.

CHAPTER 10

The Regularity of Area-MinimizingHypersurfaces

This chapter outlines some parts of the proof of the regularity theorem forarea-minimizing rectifiable currents in Rn�1Rn for n � 7. The purpose is togive an overview, illustrate basic arguments, and indicate why regularity failsfor n ½ 8. The deeper and more technical aspects of the theory are omitted.The first theorem proves a special case by methods that will be useful in thegeneral case.

10.1. Theorem. Let T be an area-minimizing rectifiable current in R 1R2.Then sptT� spt ∂T consists of disjoint line segments.

Proof. It will be shown that every point a 2 sptT� spt ∂T has a neigh-borhood U�a, r� such that sptT \ U�a, r� is a straight line segment.

CASE 1. If ∂T consists of two points (oppositely oriented), then T is theoriented line segment between them. Our assignment is to prove the mostfamous result in the calculus of variations: that a straight line is the shortestdistance between two points! We may assume ∂T D υ�1,0� � υ�0,0�. Let T0 bethe oriented segment from (0,0) to (1,0):

T0 D [�0, 0�, �1, 0�] D H1 f0 � x � 1, y D 0g ^ i.

97


We will actually show that T0 uniquely minimizes length among all normalcurrents N 2 N1R2 with the same boundary as T0. Indeed,

M�N� ½ N�dx� D ∂N�x� D 1 D M�T0�.

Therefore T0 is area minimizing. Furthermore, if M�N� D 1, then EN D i jjNjj-almost everywhere.

Next, supposing that M�N� D 1, we show that spt N ² fy D 0g. If not,for some ε > 0 there is a C1 function 0 � f�y� � 1 such that f�y� D 1 forjyj � ε and M�N f� < 1.

∂�N f� D �∂N� f�N df D ∂N� 0,

because EN D i jjNjj-almost everywhere and df�i� D 0. Since ∂�N f� D∂N D ∂T0 and T0 is mass-minimizing, therefore M�N f� ½ 1. Thiscontradiction proves that sptN ² fy D 0g.

Finally, note that ∂�N � T0� D 0. By the Constancy Theorem 4.9, N� T0

is a multiple of E1 � H1 ^ i. Since N� T0 has compact support, it must be

0. Therefore N D T0, uniqueness is proved, and Case 1 is complete.

CASE 2. If the density 1�T, a� equals 1, then spt T is a straight linesegment in some neighborhood U�a, r� of a. For almost all s, 0 < s <dist�a, spt ∂T�, the slice ∂�T B�a, s�� is a zero-dimensional rectifiable currentand a boundary, i.e., an even number of points (counting multiplicities). Therecannot be 0 points, because then T0 D 0 would have the same boundary andless mass than T B�a, s�. Therefore, M�∂�T B�a, s�� ½ 2. On the otherhand, by slicing theory 4.11(4),

s�1∫ s

0M�∂�T B�a, r�� dr � s�1M�T B�a, s��,

which converges to ˛11�T, a� D 2 as s ! 0. Therefore for some small r > 0,M�∂�T B�a, r�� D 2, and ∂�T B�a, r�� consists of two points. By Case 1,spt�T B�a, r�� is a line segment, as desired.

The general case will require the following lemma.

LEMMA [FEDERER, 4.5.17]. If R 2 Rn�1Rn with ∂R D 0, then there arenested, L

n measurable sets Mi�i 2 Z�, Mi ² Mi�1, such that

R D∑i2Z

∂�En Mi� and M�R� D∑i2Z

M�∂�En Mi��.

The Regularity of Area-Minimizing Hypersurfaces 99

Here En is the unit, constant-coefficient n-dimensional current in Rn, definedby

En D Ln ^ e1 ^ Ð Ð Ð ^ en.

Proof of Lemma. By the isoperimetric inequality (5.3), R D ∂T for someT 2 InRn. Such a T is of the form T D En f for some measurable, integer-valued function f, and M�T� D ∫ jfj. Just put Mi D fx:f�x� ½ ig. SeeFigure 10.1.1. All of the conclusions of the lemma except the last followimmediately. The last conclusion on M�R� means that there is no cancellationin the sum R D ∂�En Mi�. The idea of the proof is that, because the Mi arenested, their boundaries, if they happen to overlap, have similar orientations.Hence, in their sum, the masses add. We omit the details.

M2

M3

M1

Figure 10.1.1. Decomposing this curve as the boundaries of three nested regionsyields three pieces which never cross over each other.

CASE 3. GENERAL CASE. For every a 2 sptT� spt ∂T, spt T is a straightline segment in some neighborhood U�a, r� of a. Choose 0 < � <dist�a, spt ∂T� such that M�∂�T B�a, �� < 1. Let be a rectifiablecurrent supported in the sphere S�a, �� with ∂ D ∂�T B�a, ��(Figure 10.1.2). Apply the above lemma to T B�a, ��, and letTi D �∂�E2 Mi�� U�a, ��. Since T U�a, �� D Ti and M�T U�a, �� DM�Ti�, each Ti is mass minimizing. Since M�T U�a, �� < 1, it followsfrom monotonicity (9.5 and 9.6) that spt Ti intersects U�a, �/2� for onlyfinitely many i, and we will ignore the rest. Since Ti is of the form�∂�E2 Mi�� U�a, ��, it can be shown with some work that an oriented


S1

S2

T−

−

+

+

Ξ

Figure 10.1.2. Given the surface T B�a, �� in the ball with boundary in the sphere,there is a surface entirely in the sphere with the same boundary.

0

C

Figure 10.1.3. A pair of lines through 0 cannot be mass minimizing, as the dashedshortcut shows.

tangent cone C to Ti at any point in sptTi \ U�a, �� is of the formC D ∂�E2 N� (cf. Federer [1, 5.4.3]). Similarly if b 2 sptC� f0g, then anoriented tangent cone D to C at b is of the form D D ∂�E2 P�. The factthat C is a cone means that D is a cylinder in the sense of invarianceunder translations in the b direction. A relatively easy argument shows thata one-dimensional oriented cylinder is an oriented line with multiplicity(cf. Federer [1, 4.3.15]). Since D is of the form ∂�E2 P�, the multiplicitymust be 1. Consequently, 1�C, b� D 1�D, 0� D 1, for all b 2 sptC� f0g.By Case 2, C consists of rays emanating from 0. Since C D ∂�E2 N� hasno boundary, the same number must be oriented outward as inward. SinceC is area minimizing, oppositely oriented rays must be at a 180° angle;otherwise inside the unit circle they could be replaced by a straight lineof less mass. See Figure 10.1.3. Therefore C must be an oriented line with


multiplicity. Since C is of the form C D ∂�E2 N�, the multiplicity must be 1.Consequently 1�Ti, a� D 1�C, 0� D 1. By Case 2, for some 0 < ri � �/2,Ti U�0, ri� D �∂�E2 Mi�� U�0, ri� is an oriented line. Let r be the leastof the finitely many ri. Since Mi ² Mi�1, the various nonzero Ti U�0, r�coincide. Therefore T U�0, r� is an oriented line with multiplicity.

Now we state and sketch the proof of a complete interior regularity theoremfor area-minimizing rectifiable currents in Rn, for n � 7.

10.2. Regularity for Area-Minimizing Hypersurfaces Theorem (Simons;see Federer [1, 5.4.15]). Let T be an area-minimizing rectifiable current inRn�1Rn for 2 � n � 7. Then sptT� spt ∂T is a smooth, embedded manifold.

The proof depends on a deep lemma, which we will not prove. Our intentis to show how the pieces fit together. This lemma gives conditions on densityand tangent cones sufficient to establish regularity. A later regularity theoremof Allard [Section 8] shows that the hypothesis on an oriented tangent coneis superfluous.

10.3. Lemma [Federer, 5.4.6]. For 1 � m � n� 1, suppose T is anarea-minimizing locally rectifiable current in R

locm Rn, a 2 sptT� spt ∂T,

m�T, a� D 1, and some oriented tangent cone to T at a is an oriented, m-dimensional plane. Then spt T is a smooth, embedded manifold at a.

The second lemma gives a maximum principle for area-minimizinghypersurfaces. The result is contained in Federer [5.3.18], but it is easier toprove.

10.4. Maximum Principle. For n ½ 2, let S1 and S2 be (n� 1)-dimensional,area-minimizing rectifiable currents in Rn, and let Mi D spt Si. Suppose thatM1 and M2 intersect at a point a; that, in some neighborhood of a, M1 and M2

are smooth submanifolds; and that M2 lies on one side of M1. Then in someneighborhood of a, M1 and M2 coincide.

Remark. The maximum principle holds for singular area-minimizinghypersurfaces and more generally. Indeed, Ilmanen generalizes it to minimalhypersurfaces (stationary integral varifolds, not necessarily area-minimizing)whose singular sets have (n� 3)-dimensional Hausdorff measure 0. Solomonand White treat hypersurfaces stationary for smooth, even elliptic integrands,at least one of which is smooth.

Proof. At a, M1 and M2 may be viewed locally as graphs of functionsu1 and u2, satisfying the minimal surface equation (6.2(1)). By a standard


maximum principle, due essentially to Hopf [Satz 10], the functions u1 and u2

coincide.

The next lemma of Simons provided the final ingredient for the regu-larity theorem. In a remark following the proof of the theorem we discussits failure for n ½ 8. This lemma marks a solo appearance of differentialgeometry without measure theory.

10.5. Simons’s Lemma [Federer, 5.4.14]. For 3 � n � 7, let B be anoriented, compact, (n� 2)-dimensional smooth submanifold of the unit(n� 1)-dimensional sphere, such that the cone over B is area minimizing. ThenB is a great sphere.

It is customarily assumed that the submanifold B is connected. That weakerversion of the lemma, applied to the connected components of a general subman-ifold B, shows that all components are great spheres and hence intersect. Thiscontradiction shows thatB is connected and renders that assumption superfluous.

The final lemma has an easy proof by slicing.

10.6. Lemma [Federer, 5.4.8, 5.4.9]. For 1 � m � n� 1, let Q be a locallyrectifiable current in R

locm Rn. Then Q is area minimizing if and only if E1 ð Q

is area minimizing.

Proof of Theorem 10.2. The proof will be in two parts, by induction. Theinitial case, n D 2, was proved by Theorem 10.1.

PART I. Suppose S is an area-minimizing rectifiable current in Rn�1Rn andS is of the form S D �∂�En M�� V for some measurable set M and openset V. Then spt S \ V is a smooth embedded manifold. To prove Part I, leta 2 spt S \ V. An oriented tangent cone C to S at a can be shown to be ofthe form C D ∂�En N� (cf. Federer [1, 5.4.3]). Similarly, for b 2 sptC� f0g,an oriented tangent cone D to C at b is of the form D D ∂�En P�. The factthat C is an oriented cone means that D is an “oriented cylinder” of the formD D E1 ð Q for some Q 2 Rn�2Rn�1 (cf. Federer [1, 4.3.15]). Since S is areaminimizing, so are C and D D E1 ð Q. By Lemma 10.6 and induction, Q isan oriented, smooth, embedded manifold, possibly with multiplicity. HenceD is an oriented, smooth, embedded manifold, with multiplicity 1 becauseD is of the form D D ∂�En P�. Therefore any oriented tangent cone to Dat 0 is an oriented, (n� 1)-dimensional plane. In particular, since D is anoriented cone, D itself is an oriented, (n� 1)-dimensional plane. Therefore byLemma 10.3, sptC� f0g is a smooth, embedded manifold. By Lemma 10.5,spt C intersects the unit sphere in a great sphere. Hence C is an oriented,(n� 1)-dimensional plane, with multiplicity 1 because C is of the form C D


∂�En N�. A reapplication of Lemma 10.3 now shows that spt S is a smooth,embedded manifold at a, proving Part I.

PART II. Completion of proof. Let a 2 sptT� spt ∂T. Choose a small� > 0 such that ∂�T U�a, �� is rectifiable (cf. 4.11(5)). Let be a recti-fiable current supported in the sphere S�a, �� with the same boundary asT U�a, ��. By the Lemma of section 10.1, there are nested sets Mi ²Mi�1 such that T U�a, �� D Si, with Si D �∂�En Mi�� U�a, ��. More-over M�T U�a, �� D M�Si�, so that each Si is area minimizing. SinceM�T U�a, �� < 1, it follows from monotonicity (9.3 and 9.6) that spt Siintersects U�a, �/2� for only finitely many i. For such i, by Part I, spt Si is asmooth embedded manifold at a. The containments Mi ² Mi�1 imply that eachof these manifolds lies on one side of another. Therefore by the maximum prin-ciple, 10.4, spt T is a smooth, embedded manifold at a. The theorem is proved.

10.7. Remarks. The regularity theorem fails for n ½ 8. E. Bombieri, E. DeGiorgi, and E. Giusti (1969), gave an example of a seven-dimensional,area-minimizing rectifiable current T in R8 with an isolated singularity at0. This current T is the oriented truncated cone over B D S3�0, 1/

p2�ð

S3�0, 1/p

2� ² S7�0, 1� ² R8; ∂T D B. It also provides a counterexample toSimons’s Lemma 10.5, which is precisely the point at which the proof ofregularity breaks down.

Here we give a plausibility argument that such a counterexample shouldarise for some large n. First consider B D S0�0, 1/

p2�ð S0�0, 1/

p2� ² S1 ²

R2, the four points pictured in Figure 10.7.1. The associated cone does have asingularity at 0, but it is not mass minimizing. The mass-minimizing currentconsists of two vertical lines (or two horizontal lines).

R1

R1

− + +−

+ − + −

Figure 10.7.1. The X-shape, which is the cone over the four points S0 ð S0 in R2,is not mass minimizing. The mass-minimizing current consists of two vertical linesegments.

Second consider B D S1�0, 2/p

5�ð S0�0, 1/p

5� ² S2 ² R3, the twocircles pictured in Figure 10.7.2. Again the associated cone has a singularity at


R1

R2

Figure 10.7.2. The cone bounded by the two circles S1 ð S0 in R3 is not massminimizing. Neither is the cylinder on the right. The mass-minimizing surface is thecatenoid in the middle. It bows inward a bit toward the cone in order to shorten itswaist, but not too much to overstretch its sides.

Figure 10.7.3. In higher dimensions, the mass-minimizing surface bows furtherinward toward the cone. Finally in R8 it collapses onto the cone, which is massminimizing for the first time.

0, but it is not mass minimizing. The mass-minimizing current is the picturedcatenoid. The catenoid has less area than a cylinder. Although the curved cross-sections are longer than the straight lines of the cylinder, the circumference isless. The amount that the catenoid bows inward toward the cone is a balancingof these two effects.

Third consider B D S2�0, 1/p

2�ð S2�0, 1/p

2� ² S5 ² R6, pictured sche-matically in Figure 10.7.3. As the dimensions increase, the mass cost ofbeing far from the origin rises, and the mass-minimizing current bows farthertoward the cone. Finally in R8, it has collapsed onto the cone, which is massminimizing for the first time.

EXERCISES

10.1. Give two different area-minimizing rectifiable currents in R 2R3 with the sameboundary.

10.2. Prove or give a counterexample: If T 2 Rm�Rn ð Rl� is area minimizing and pdenotes orthogonal projection of Rn ð Rl onto Rn, then p#T is area minimizing.

10.3. Find a counterexample in I2R3 to Lemma 10.3 if the hypothesis that S be areaminimizing is removed.

CHAPTER 11

Flat Chains Modulo �, Varifolds, and(M, ε, υ)-Minimal Sets

A number of alternative spaces of surfaces have been developed in geometricmeasure theory, as required for theory and applications. This chapter givesbrief descriptions of flat chains modulo �, varifolds, and (M, ε, υ)-minimal sets.

11.1. Flat Chains Modulo n [Federer, 4.2.26]. One way to treat nonori-entable surfaces and more general surfaces is to work modulo 2, or moregenerally modulo � for any integer � ½ 2. Two rectifiable currents T1, T2

are congruent modulo � if T1 � T2 D �Q for some rectifiable current Q. Inparticular, T � �T (mod 2). The m-dimensional rectifiable currents modulo�, denoted R

�m , are defined as congruence classes of rectifiable currents.

For example, consider the Mobius strip of Figure 11.1.2 bounded by thecurve C. There is no way to orient it to turn it into a rectifiable current withboundary C. However, if it is cut along a suitable curve D, it can then beoriented as a rectifiable current T, and it works out that ∂T � C (mod 2). Ingeneral, rectifiable currents modulo 2 correspond to unoriented surfaces.

Two parallel circles, as in Figure 11.1.3, bound an interesting rectifiablecurrent modulo 3. The surfaces of both Figures 11.1.2 and 11.1.3 occur assoap films.

Most of the concepts and theorems on rectifiable currents have analogs forrectifiable currents modulo �: mass, flat norm, the Deformation Theorem, theCompactness Theorem, existence theory, and the Approximation Theorem.

One tricky point: to ensure completeness, the integral flat chains congruentto 0 modulo � must be defined as the flat-norm closure of those of the form �Q.

105


Figure 11.1.1. Bill Ziemer (right), who introduced flat chains modulo 2, with histhesis advisor, Wendell Fleming (left), and the author (center), at a celebration inZiemer’s honor at Indiana in 1994. Photo courtesy of Ziemer.

D

C

Figure 11.1.2. A rectifiable current modulo 2 can be nonorientable, like this Mobiusstrip.

Flat Chains Modulo �, Varifolds, and (M, ε, υ)-Minimal Sets 107

Figure 11.1.3. In a rectifiable current modulo 3, three sheets can meet along a curveinside the surface which does not count as boundary.

In codimension greater than 1, it is an open question whether they are all ofthe form �Q. (The counterexample in Federer [p. 426] is wrong.) However,every rectifiable current congruent to 0 modulo � is of the form �Q.

It is generally easier to prove regularity for area-minimizing rectifiablecurrents modulo 2 than for rectifiable currents themselves. Indeed, an m-dimensional area-minimizing rectifiable current modulo 2 in Rn is a smoothembedded manifold on the interior, except for a rectifiable singular set oflocally finite H

m�2 measure (Federer [2]; Simon [1]). The only standard regu-larity result I know which fails modulo 2 is boundary regularity for area-minimizing hypersurfaces (8.4).

For � > 2, even area-minimizing hypersurfaces modulo � can have codi-mension 1 singular sets, as Figure 11.1.3 suggests. Taylor [3] proved thata two-dimensional area-minimizing rectifiable current modulo 3 in R3 awayfrom boundary consists of C1 surfaces meeting in threes at 120° angles alongC1,˛ curves (cf. 13.9). White [8] proved that an �n � 1�-dimensional area-minimizing rectifiable current modulo 4 in Rn decomposes locally into a pairof area-minimizing rectifiable currents modulo 2. Moreover, White [7] provedthat for any odd �, an �n � 1�-dimensional area-minimizing rectifiable currentmodulo � in Rn away from boundary is a smooth embedded manifold, exceptfor a singular set of dimension at most n � 2. For even � > 2, almost every-where regularity remains an open question.

11.2. Varifolds [Allard]. Varifolds provide an alternative perspective to cur-rents for working with rectifiable sets. Varifolds carry no orientation, and hencethere is no cancellation in the limit and no obvious definition of boundary.


An m-dimensional varifold is a Radon measure on Rn ð GmRn, whereGmRn is the Grassmannian of unoriented unit m-planes through 0 in Rn.We have seen previously how to associate to a rectifiable set E the measureH

m E on Rn, but this perspective ignores the tangent planes to E. Instead,we now associate to a rectifiable set E, with unoriented tangent planes EE�x� DTanm�E, x� 2 GmRn, the varifold

v�E� � Hm f�x, EE�x��: x 2 Eg.

The varifolds which so arise are called integral varifolds. One allows positiveinteger multiplicities and noncompact support.

In general varifolds, the tangent planes need not be associated with theunderlying set. For example, if you cut an edge off a cubical crystal, as inFigure 11.2.1, because the crystal loves horizontal and vertical directions, theexposed surface forms very small horizontal and vertical steps. Such a corru-gation can be modeled by a varifold concentrated half and half on horizontaland vertical tangent planes in G2R3, whereas in space R3 it is concentratedon the diagonal surface.

Figure 11.2.1. If you cut an edge off a cubical crystal, the exposed surface formstiny steps, well modeled by a varifold.

The first variation υV of a varifold V is a function which assigns to anycompactly supported smooth vectorfield g on Rn the initial rate of changeof the area of V under a smooth deformation of Rn with initial velocityg. Roughly, the first variation is due to the mean curvature of V and theboundary of V. A varifold is called stationary if υV D 0. Geometrically,stationary integral varifolds include area-minimizing rectifiable currents, area-minimizing rectifiable currents modulo �, and many other physical surfacessuch as soap films. Some singularities of soap films unavoidably count as addi-tional boundary in the category of rectifiable currents or rectifiable currentsmodulo �, but fortunately do not add to their first variations as varifolds.


There is a compactness theorem for integral varifolds with bounds ontheir areas, first variations, and supports. There are also general isoperimetricand regularity theorems. However, it is an open question whether a two-dimensional stationary integral varifold in an open subset of R3 is a smoothembedded manifold almost everywhere.

11.3. (M, e, d)-Minimal Sets [Almgren 1]. Perhaps the best model of soapfilms is provided by the (M, 0, υ)-minimal sets of Almgren. A nonempty,bounded subset S ² Rn � B with H

m�S� < 1 and S D spt�Hm S� � B is

(M, 0, υ) minimal with respect to a closed set B (typically “the boundary”) if,for every Lipschitz deformation ϕ of Rn which differs from the identity maponly in a υ-ball disjoint from B,

Hm�S� � H

m�ϕ�S��.

Since ϕ need not be a diffeomorphism, it can pinch pieces of surface together,as in Figure 11.3.1. The M refers to area and may be replaced by a moregeneral integrand. For more general functions ε�r� D Cr˛, ˛ > 0, the inequalityimposed on deformations inside r-balls �r � υ� relaxes to

Hm�S� � �1 C ε�r��H

m�ϕ�S��.

S j (S )

Figure 11.3.1. A curve S is not (M, 0, υ) minimal if a deformation ��S� has lesslength.

Such (M, ε, υ)-minimal sets include soap bubbles (with volume constraints);see Section 13.8.

Three basic properties of (M, ε, υ)-minimal sets are �Hm, m� rectifiability

[Almgren 1, II.3(9)], monotonicity [Taylor 4, II.1], and the existence of an


Figure 11.3.2. It is an open question whether the cone over the tetrahedron is aleast-area soap film or whether there might be a smaller soap film of higher topologicaltype. Photo by F. Goro, computer graphics by J. Taylor, used by permission.


(M, 0, υ)-minimal tangent cone at every point [Taylor 4, II.2]. The conditionin the definition that S D spt�H

m S� � B may be replaced by the conditionthat S be rectifiable, with the understanding that S may be altered by a set ofH

m measure 0 [Morgan 17, §2.5].Almgren [1] has proved almost-everywhere regularity results for (M, ε,

υ)-minimal sets. In 1976, Taylor [4] proved that for two-dimensional (M, ε, υ)-minimal sets in R3, there are only two possible kinds of singularities: (1) threesheets of surface meeting at 120° angles along a curve and (2) four such curvesmeeting at approximately 109° angles at a point (see Section 13.9). Theseare precisely the two kinds of singularities that Plateau had observed in soapbubbles and soap films a hundred years earlier. (Warning: Almgren [1] andTaylor [4] are technical, and I think there are some (correctable) gaps.)

Boundary regularity remains conjectural; see Section 13.9.It remains an open question today whether a smooth curve in R3 bounds

a least-area soap film in the class of (M, 0, υ)-minimal sets, with variableυ > 0. The problem is that in the limit υ may go to 0. See Morgan [20]. Itis an open question whether the cone over the tetrahedron (the soap film ofFigure 11.3.2) is a least-area soap film or whether there might be a smallersoap film of higher topological type. Another open question asks whether theCartesian product of an (M, 0, υ)-minimal set with an interval is (M, 0, υ0)minimal (a property which holds trivially for most classes of minimal surfaces;cf. 10.6).

EXERCISES

11.1. Give an example of a boundary curve in R3 for which the area-minimizing flatchain modulo 4 has less area than the area-minimizing integral current.

11.2. Let S be the unit 2 disc, and let v�S� be the associated varifold. What isv�S��Rn ð G2Rn�?

11.3. Give an example of a two-dimensional set in R3 that is (M, 0, υ) minimal forsmall υ but not for large υ.

11.4. Give an example of a two-dimensional set in R3 that is (M, 0, υ) minimal forall υ > 0 but not area minimizing.

CHAPTER 12

Miscellaneous Useful Results

Federer’s treatise presents many basic methods of geometry and analysisin a generality that embraces manifold applications. This chapter describesFederer’s treatment of Sard’s Theorem, Green’s Theorem, relative homology,functions of bounded variation, and general parametric integrands.

12.1. Morse–Sard–Federer Theorem. The usual statement of Sard’s The-orem says that the set of critical values of a C1 function f: Rm ! Rn hasLebesgue measure 0. Federer’s refinement shows precisely how the Haus-dorff measure of the image depends on the rank of Df and the smoothnessclass of f.

THEOREM [FEDERER, 3.4.3]. For integers m > � ½ 0, k ½ 1, let f be a Ck

function from an open subset A of Rm into a normed vectorspace Y. Then

H�Cm��

k f�fx 2 A: rank Df�x � �g D 0.

Note that the usual statement may be recovered by taking Y D Rn, � D n � 1,k ½ m � n C 1. The latest improvement has been provided by Bates.

12.2. Gauss–Green–De Giorgi–Federer Theorem. The usual statementof Green’s Theorem says that a C1 vectorfield ��x on a compact region A inRn with C1 boundary B satisfies

∫B��x Ð n�A, x d� D

∫A

div ��x d Lnx,

113


where n�A, x is the exterior unit normal to A at x and d� is the element ofarea on B. Federer treats more general regions and vectorfields that are merelyLipschitz.

Federer allows measurable regions A for which the current boundary T D∂�En A is representable by integration: T�ϕ D ∫ hET, ϕi d jjTjj (cf. 4.3). IfA is compact, this condition just says that the current boundary has finitemeasure: M�T D jjTjj�Rn < 1. In any case, this condition is weaker thanrequiring that the topological boundary of A have finite H

n�1 measure.

DEFINITION. Let b 2 A ² Rn. We call n D n�A, b the exterior normal ofA at b if n is a unit vector,

n�fx: �x � b Ð n > 0g \ A, b D 0,

and

n�fx: �x � b Ð n < 0g � A, b D 0.

Clearly there is at most one such n. If b is a smooth boundary point of A,then n is the usual exterior normal. Even if ∂A is not smooth at b, n maybe defined, as Figure 12.2.1 suggests. The assertion in the theorem belowthat the measure jjTjj D H

n�1 domain n�A, x says roughly that the currentboundary of A coincides with the domain of n�A, x almost everywhere. Inthe final formula, div � exists almost everywhere because a Lipschitz functionis differentiable almost everywhere.

THEOREM [FEDERER, 4.5.6]. Let A be an Ln-measurable subset of Rn

such that T D ∂�En A is representable by integration. Then jjTjj D Hn�1

n

bA

Figure 12.2.1. The generalized normal n is defined at b because the arms from thesides have density 0 at b.

Miscellaneous Useful Results 115

domain n�A, x and, for any Lipschitz vectorfield ��x of compact support,∫

��x Ð n�A, x d Hn�1x D

∫A

div ��x d Lnx.

Generalizations to wilder sets have been provided by Harrison.

12.3. Relative Homology [Federer, 4.4]. Suppose B ² M are C1, compactsubmanifolds with boundary of Rn (or, more generally, compact Lipschitzneighborhood retracts; cf. Federer [1, 4.1.29, 4.4.1, 5.1.6]). Two rectifiablecurrents S, T in M are homologous with respect to B if there is a rectifiablecurrent X in M such that

spt�T � S � ∂X ² B.

We say that S and T are in the same relative homology class. Given a rectifiablecurrent S, there is a rectifiable current T of least area in its relative homologyclass.

EXAMPLE 1. Let M be a perturbed solid torus in R3, let B be its boundary,and let S be a cross-sectional disc. The area minimizer T relatively homologousto S provides a cross-sectional surface of least area. See Figure 12.3.1. Theboundary of T is called a free boundary.

EXAMPLE 2. Let M be a large, encompassing ball in R3, let B be the surfaceof a table (not necessarily flat), and let C be a curve which begins and ends in

S

T

M

Figure 12.3.1. The area-minimizer T relatively homologous to S provides the leastcross-sectional area.


B. Let S be a surface with spt�∂S � C ² B. The area minimizer T relativelyhomologous to S provides a surface of least area with the fixed boundary Cand additional free boundary in B. It can generally be realized as a soap film.See Figure 12.3.2.

T

B

C

Figure 12.3.2. A soap film minimizing area in its relative homology class.

Proof of Existence. As Brian White pointed out to me, there is a muchsimpler existence proof than that of Federer [1, 4.4.2, 5.1.6]. Consider a mini-mizing sequence Ti, viewed as locally integral currents in M � B (cf. 9.1). Bycompactness, we may assume that the Ti converge to a minimizer T. The hardpart is to show that T stays in the same relative homology class. Convergencemeans that, for some Y and Z1 in M with small mass,

E D Ti � T � Y � ∂Z1

is supported in a small neighborhood of B. Let Y1 be a minimizer in RN

with ∂Y1 D ∂Y. Since ∂Y1 lies in a small neighborhood of B and becauseM�Y1 � M�Y is small, by monotonicity, 9.5, Y1 lies in a small neighborhoodof B. Let Z2 be a minimizer in RN with ∂Z2 D Y � Y1. Then M�Z2 is small,and hence Z2 lies in a small neighborhood of M. Let E1 denote the projectionof the cycle E C Y1 onto B; then E C Y1 � E1 D ∂Z3, with Z3 in a smallneighborhood of B. We now have

Ti � T D E1 C ∂Z1 C ∂Z2 C ∂Z3,

with E1 in B and the Zi in a small neighborhood of M. Projecting onto Myields

Ti � T D E1 C ∂Z,

with Z in M, so that T lies in the same relative homology class as Ti, asdesired.


Isoperimetric Inequality. Very similar arguments provide a constant ,depending on M, such that, if X is an m-dimensional relative boundary inM, then there is a rectifiable current Y in M such that X D ∂Y (mod B) and

�1 M�Y m/�mC1 C M�∂Y � X � M�X .

Moreover, every relative cycle with small mass is a relative boundary.

Remarks on Regularity. Of course, away from B, a relatively homologi-cally area-minimizing rectifiable current enjoys the same regularity as anabsolutely area-minimizing rectifiable current. In addition, regularity resultsare known along the free boundary (see e.g. Taylor [1], Gruter [1, 2], andHildebrandt). In particular, if M is a smooth Riemannian three manifold withboundary B, then an unoriented homologically area-minimizing surface T rela-tive to B with T ² B is a C1 submanifold with boundary (Taylor [1, 1(7),and Theorem 5], preceded by decomposition into multiplicity-1 boundaries ofnested sets as in the Lemma of 10.1 and followed by a maximum principle).

12.4. Functions of Bounded Variation [Federer, 4.5.9; Giusti; Simon3, §6]. An important class of functions in analysis is the space BVloc offunctions of locally bounded variation. A real-valued function on R1 is inBVloc if it agrees almost everywhere with a function g of finite total variationon any interval [a, b],

sup

{k∑

iD1

jg�xi � g�xi�1 j: k 2 Z, a � x0 � Ð Ð Ð � xk � b

}< 1,

or, equivalently, if the distribution derivative Df is a locally finite measure.Similarly, a real-valued function on Rn is in BVloc if Df is a locally finite(vector-valued) measure. The associated space of currents fEn fg is preciselyNloc

n Rn, the locally normal currents of codimension 0. Here we give a samplingfrom Federer’s comprehensive theorem on BVloc.

THEOREM [FEDERER, 4.5.9]. Suppose f 2 BVloc.

�13 If $s is the characteristic function of fx:f�x ½ sg, then

Df D∫s2R

D$s ds

and

jDfj D∫s2R

jD$sj ds

almost everywhere.


�31 If n ½ 1, then there is a constant c such that

jjf � cjjLn/�n�1 � n�1˛�1/nn

∫jDfj,

where ˛n is the volume of the unit ball in Rn. If f has compact support, thenc D 0.

Remarks. The second statement of (13) just says that there is no cancel-lation in the first. In the notation of geometric measure theory, (13) becomes

∂�En f D∫

∂[En fx:f�x ½ sg] d L1s

and

jj∂�En f jj D∫

jj∂[En fx:f�x ½ sgjj d L1s.

An excellent comprehensive treatment of BV appears in Giusti.

12.5. General Parametric Integrands [Federer 1, 5.1]. In many mathe-matical and physical problems the cost of a surface depends not only on itsarea but also on its position or tangent plane direction. The surface energy of acrystal depends on its orientation with respect to the underlying crystal lattice(cf. Taylor [2]). Therefore one considers an integrand �x, � associating to arectifiable current T a cost or energy

�T D∫

�x, ET�x d jjTjj.

The most important case remains the area integrand A�x, � D j�j. One alwaysrequires to be continuous and homogeneous in �. Usually is positive (for �nonzero), even ��x,�� D �x, � , and convex ��x, �1 C �2 � �x, �1 C�x, �2 ; i.e., each

a�� a, �

is a norm. (If this norm is given by an inner product, then �x, � is a Rieman-nian metric.)

The adjective parametric just means that �T depends only on the geo-metric surface T and not on its parameterization (geometric measure theoryrarely uses parameterizations at all). Convexity of means each a is convex,

�1 a��1 C �2 � a��1 C a��2 ,


with the geometric interpretation that the unit ball fa�� 1g is convex. Itfollows easily that planes are a minimizing; i.e., if S is a portion of a planeand ∂R D ∂S, then

�2 a�S � a�R .

When (2) holds, one says F is semielliptic.Uniform convexity of gives an estimate on the strength of the convexity

inequality (1),

�3 a��1 C a��2 � a��1 C �2 ½ c�j�1j C j�2j � j�1 C �2j ,

with the geometric interpretation for smooth that the unit ball fa�� 1ghas positive inward curvature. It follows easily that, if S is a portion of a planeand ∂R D ∂S, then

�4 a�R � a�S ½ ,�M�R � M�S .

Unit Balls Φ Minimizers

{Φ1(x)≤1}

{Φ2(x)≤1}

{Φ3(x)≤1}

Convex, but not uniformly convex

Uniformly convex Diagonal is unique minimizer

Not convexDiagonal limit of sawtooth

minimizers is not minimizing

Minimizers include sawtoothcurves and smooth diagonal limit

Figure 12.5.1. Three integrands on R2 for which horizontal and vertical directionsare relatively cheap.


One says is elliptic. Ellipticity, introduced by Almgren, seems to be theright hypothesis for theorems, but much harder to verify directly than uniformconvexity. The two notions are equivalent in codimension 1.

All of these notions are invariant under diffeomorphisms of the ambient.In general, semiellipticity implies lower semicontinuity and the existence of

minimizers; ellipticity implies regularity of minimizers (see 8.5). Figure 12.5.1considers the cost of curves for three integrands �� on the plane for whichhorizontal and vertical directions are relatively cheap. 1 is not convex andnot lower semicontinuous, so limit arguments can fail to produce minimizers.2 is borderline convex and admits nonsmooth minimizers. 3 is uniformlyconvex, and the cheapest path between two points is uniquely a straight line.

CHAPTER 13

Soap Bubble Clusters

Soap bubble clusters illustrate simple mathematical principles. Yet despitenotable progress, they defy complete mathematical explanation.

A single soap bubble quickly finds the least-surface-area way to enclose thefixed volume of air trapped inside — the round sphere in Figure 13.0.1.

Similarly bubble clusters seek the least-area way to enclose and sepa-rate several regions of prescribed volumes. This principle of area minimiza-tion alone, implemented on Ken Brakke’s Surface Evolver (see 16.8), yieldscomputer simulations of bubble clusters, as in Figure 13.0.2, from the video“Computing Soap Films and Crystals” by the Minimal Surface Team at theGeometry Center.

Do soap bubble clusters always find the absolute least-area shape? Notalways. Figure 13.0.3 illustrates two clusters enclosing and separating the samefive volumes. In the first, the tiny fifth volume is comfortably nestled deep inthe crevice between the largest bubbles. In the second, the tiny fifth volumeless comfortably sits between the medium-size bubbles. The first cluster hasless surface area than the second. It might be still better to put the smallestbubble around in back.

As a matter of fact, until 2000 (see our new Chapter 14) it remained anopen question whether the standard double bubble of Figure 13.0.4 is theleast-area way to enclose two given volumes, as realized over the courseof an undergraduate thesis by Foisy, despite a proof due to White that thesolution must be a surface of revolution [Foisy, Theorem 3.4; cf. Morgan 3,Theorem 5.3]. It may seem hard to imagine any other possibilities, until yousee J. M. Sullivan’s computer-generated competitor in Figure 13.0.5, which

121


Figure 13.0.1. A spherical soap bubble has found in the least-area way to enclose agiven volume of air.

does, however, have more area and is apparently unstable. More generally,both regions might have several components, wrapped around each other.

In general, it is a difficult open question whether each separate region isconnected, or whether it might conceivably help to subdivide the regions ofprescribed volume, with perhaps half the volume nestled in one crevice here,and the other half in another crevice there.

Similarly, it is an open question whether an area-minimizing cluster mayincidentally trap inside “empty chambers,” which do not contribute to theprescribed volumes. Figure 13.0.6 shows a 12-bubble with a dodecahedralempty chamber on the inside, obtained by Tyler Jarvis of Mississippi StateUniversity using Brakke’s Surface Evolver. The computation postulated theempty chamber; without such a restriction, empty chambers probably neveroccur. Michael Hutchings proved that in a minimizing double bubble, thereare no such empty chambers.

This chapter began with my AMS-MAA address in San Francisco, 1991,available on video [Morgan 4] and written up in Morgan [13, 12].

Soap Bubble Clusters 123

Figure 13.0.2. Bubble clusters seek the least-area way to enclose several volumesof air. Enhanced from “Computing Soap Films and Crystals,” a video by the MinimalSurface Team, The Geometry Center. Computer graphics copyright John M. Sullivan,University of Illinois; color version at http://www.math.uiuc.edu/¾jms/Images/.

13.1. Planar Bubble Clusters. Many of the fundamental questions remainopen for planar bubble clusters — least-perimeter ways to enclose and separateregions of prescribed areas. A proof of the planar double bubble conjectureappeared in 1993, the work of a group of undergraduates: Joel Foisy, ManuelAlfaro, Jeffrey Brock, Nickelous Hodges, and Jason Zimba (see Figure 13.1.1).(Their work was featured in the 1994 AMS What’s Happening in the Mathe-matical Sciences.)


Figure 13.0.3. Soap bubble clusters are sometimes only relative minima for area.These two clusters enclose and separate the same five volumes, but the first has lesssurface area than the second.

Figure 13.0.4. A double bubble provides the least-area way to enclose two givenvolumes. Photo by Jeremy Ackerman, Washington University ’96.


Figure 13.0.5. This nonstandard, computer-generated double bubble has more areaand is apparently unstable. Computer graphics copyright John M. Sullivan, Universityof Illinois; color version at http://www.math.uiuc.edu/¾jms/Images/.

In the plane (but not in higher dimensions) the nontrivial general existencetheory [Morgan 19] admits requiring regions to be connected, although theythen must be allowed to bump up against each other. For this restrictedproblem, another group of undergraduates — Christopher Cox, Lisa Harrison,Michael Hutchings, Susan Kim, Janette Light, Andrew Mauer, and MegTilton — proved the standard triple bubble of Figure 13.1.2 minimizing.

13.2. Theory of Single Bubbles. Simplikios’s sixth century commentary onAristotle’s De Caelo (see Knorr [p. 273], recommended to me by D. Fowler),referring incidentally to work no longer extant, reports that

. . . it has been proved . . . by Archimedes [287–212 B.C.] and Zenodorus [¾200B.C.] that of isoperimetric figures the more spacious one . . . among the solids [is]the sphere.


Figure 13.0.6. It is an open question whether area-minimizing clusters may haveempty chambers, such as the dodecahedral chamber at the center of this 12-bubble.Graphics by Tyler Jarvis, Mississippi State University.

a

ce

d

bB2 B1 B2

B1

B2B1B2 B1B1

B2

B2

B2B1

B1

B1

Figure 13.1.1. The standard planar double bubble (a and b) and not some exoticalternative with disconnected regions or empty chambers (c, d, or e) provides theleast-perimeter way to enclose and separate two regions of prescribed area, as provedby a group of undergraduates [Foisy et al., 1993].


a b c

Figure 13.1.2. The standard planar triple bubble (a) provides the least-perimeter wayto enclose and separate three regions of prescribed area, at least in the category ofconnected regions, as proved by another group of undergraduates [Cox et al.].

But Archimedes and Zenodorus considered only a small class of solids, includ-ing of course the Platonic solids.

Over 2000 years later H. A. Schwarz (1884) apparently gave the firstcomplete proof, by a symmetrization argument. Schwarz symmetrization re-places slices by parallel hyperplanes in Rn with �n� 1�-discs centered onan orthogonal axis. De Giorgi [2] (1958) gave a simple completion ofan early argument of J. Steiner. Steiner symmetrization replaces slices byparallel lines with intervals centered on an orthogonal hyperplane. Perhapsthe simplest known proof is based on the divergence theorem [Gromov,§2.1; Berger 1, 12.11.4]. This result also follows from the Brunn-MinkowskiTheorem [Federer, 3.2.41]. Furthermore, given known existence and regularity,there is a very simple symmetry proof (see Section 14.3 and [Hutchings,Corollary 2.8]). For more history, results, and references, see the excellentreview by Burago and Zalgaller (especially §10.4). These isoperimetric resultsgeneralize to norms more general than area [Gromov, §2.1; Morgan 16, §10.6;Brothers and Morgan].

More General Ambients. Round balls are known to be minimizing also inSn and Hn [Schmidt]. Round balls about the origin are known to be mini-mizing in certain two-dimensional surfaces of revolution (see the survey byHowards et al.), in certain n-dimensional cones [Morgan and Ritore], andin Schwarzschild-like spaces by Bray and Morgan, with applications to thePenrose Inequality in general relativity. [Morgan and Johnson, Theorem 2.2]show that in any smooth compact Riemannian manifold, minimizers for smallvolume are nearly round spheres. There are results on R ð Hn by Hsiang andHsiang, on RP3, S1 ð R2, and T2 ð R by Ritore and Ros ([2]; [1], [Ritore]),on R ð Sn by Pedrosa, and on S1 ð Rn, S1 ð Sn, and S1 ð Hn by Pedrosaand Ritore. In RP3, the least-area way to enclose a given volume V is: for


small V, a round ball; for large V, its complement; and for middle-sized V, asolid torus centered on an equatorial RP1.

Although regularity theory (8.5) admits the possibility of singularities ofcodimension 8 in an area-minimizing single bubble, one might well not expectany. Nevertheless Hsiang (1993) announced an example of a singular bubblein the Cartesian product H7 ð S7 of hyperbolic space with the round sphere.Hsiang uses symmetry to reduce it to a question about curves in the plane.To understand his example, I like to think about the least-perimeter way toenclose a region of prescribed area A on the cylinder R1 ð S1. For small A,the solution is a disc, for large A, the solution is an annular band. Both typesoccur for a critical value of A, when the minimizer jumps from one type tothe other. If the minimizer were continuous in A, it would have to becomesingular to change type. This is what happens in H7 ð S7.

13.3. Cluster Theory. The existence of soap bubble clusters in Rn (seeFigure 13.3.1) is guaranteed by the following theorem, proved in a more

Figure 13.3.1. There exists a “soap bubble cluster” providing the least-area way toenclose and separate m regions Ri of prescribed volumes. Photo by Jeremy Ackerman.


general context by Almgren [1, Theorem V1.2] and specialized and simplifiedin Morgan [3, §4.4], where details can be found.

A cluster consists of disjoint regions R1, . . . , Rm (n-dimensional locallyintegral currents of multiplicity 1) with volume Ri D Vi, complement R0, andsurface area

A D 1

2

m∑iD0

M�∂Ri�.

(By including R0, the sum counts each surface twice, before multiplication by12 ). A region is not assumed to be connected.

13.4. Existence of Soap Bubble Clusters. In Rn, given volumesV1, . . . , Vm > 0, there is an area-minimizing cluster of bounded regions Riof volume Vi.

To outline a simple proof, we will need a few lemmas. Lemma 13.5, anextremely useful observation of Almgren’s (see Almgren [1, V1.2(3)] orMorgan [19, 2.2]), will let us virtually ignore the volume constraints in elim-inating wild behavior.

13.5. Lemma. Given any cluster, there exists C > 0, such that arbitrarysmall volume adjustments may be accomplished inside various small balls ata cost

jAj � CjVj

(see Figure 13.5.1).

Remark. There is much freedom in the placement of the balls. Selectingtwo disjoint sets of such balls now will allow later proofs to adjust volumesat such locations, away from the main argument.

Figure 13.5.1. Gently pushing R1 into R2 at a typical border point yields a smallvolume adjustment with jAj � CjVj.


Proof Sketch. By the Gauss–Green–De Giorgi–Federer Theorem, 12.2,at almost all points of ∂Ri, Ri has a measure-theoretic exterior normal andthe approximate tangent cone is a half-space. Hence at almost every pointof ∂R1, for example, this half-space fits up against that of some other Ri,say R2. Gently pushing R1 into R2 costs jAj � C12jVj. Combining overmany such neighboring pairs yields an arbitrary small volume adjustment withjAj � CjVj.

13.6. Lemma. An area-minimizing cluster is bounded in Rn.

Proof. Let V�r� denote the volume outside B�0, r�; let A�r� denote thearea outside B�0, r�. Truncation at almost any radius r saves A�r�, requirespatching by the slice hRi, u, rCi with u�x� D jxj and

MhRi, u, rCi � jV0�r�j

(4.11(3)), and requires replacing lost volume at a cost of CV�r� for large r byLemma 13.5. Therefore

�1� jV0�r�j C CV�r� ½ A�r�.

On the other hand, application of the isoperimetric inequality, 5.3, to the exte-rior of B�0, r� yields

�2� jV0�r�j C A�r� ½ �V�r��n�1�/n.

Adding inequalities (1) and (2) yields

2jV0�r�j ½ �CV�r�C �V�r��n�1�/n ½ 12�V�r�

�n�1�/n

for large r. If V�r� is never 0,

n�V1/n�0 D V��n�1�/nV0 � �c < 0

for almost all large r, which contradicts positive and nonincreasing V.

13.7. Sketch of Proof of Theorem 13.4. The main difficulty is that volumecan disappear to infinity in the limit. First we show that we can preserve somefraction of the volume. Let C˛ be a minimizing sequence of clusters with theprescribed volumes Vi. We claim there are constants S, υ > 0, such that if Rn

is partitioned into cubes Ki of edge length S, then for some Ki,

�1� vol�R1,˛ Ki� ½ υV1.


Indeed choose S large enough so that if vol�R Ki� � V1, then

area�∂R Ki� ½ � �vol�R Ki��n�1�/n,

for some isoperimetric constant � (12.3(1)). Let υ � �/A. Then for each Kj,

area�∂R1,˛ Kj� ½ �

maxi�R1,˛ Ki�

vol�R1,˛ Kj�.

Summing over j yields

A ½ �

maxi�R1,˛ Ki�

V1,

maxi�R1,˛ Ki� ½ �V1/A ½ υV1,

proving the claim. Therefore, by translating the C˛, we may assume that forsome fixed r > 0,

�2� vol�R1,˛ B�0, r�� ½ υV1.

By a compactness argument (see 9.1), we now may assume that the C˛

converge to a limit cluster C. By (2),

�3� vol�R1� ½ υV1.

The second step is to show that C is area minimizing for its volume. Other-wise, a compact improvement of C could be used to improve substantiallythe C˛, with only a small volume distortion. The volume distortion could becorrected by truncation, slight homothetic contraction until no volume is toobig, and the addition of tiny spheres. By Lemma 13.6, C is bounded.

Now if there was no volume loss to infinity, C solves our problem. Ifthere was a volume loss, repeat the whole process with translations of thediscarded material, obtained as the restriction of the C˛ to the exterior of anincreasing sequence of balls, judiciously chosen to make the required patchinginconsequential. Since each repetition recovers a fixed fraction of missingvolume, countably many repetitions can capture the total volume and yield asolution. (Since each is bounded, they do fit in Rn). Since the conglomeratesolution must be bounded by Lemma 13.6, we conclude that only finitelymany repetitions were actually needed.

General Ambient Manifolds. Theorem 13.4 holds in any smooth Rieman-nian manifold M with compact quotient M/ by the isometry group .


Regularity results for minimizing clusters begin with the following.

13.8. Proposition. In a minimizing cluster, the rectifiable set S D [�∂Ri� is�M, ε, υ� minimal:

Hn�1�S� � �1 C ε�r��H

n�1�ϕ�S��,

where ε�r� D 3Cr, r � υ.

Proof. The proof depends on Lemma 13.5, which provides volume adjust-ments at cost jAj � CjVj. υ must be chosen small enough so that a υ-ballhas small volume in the sense of Lemma 13.5 and so that any υ-ball is disjointfrom a set of balls used to readjust volumes.

Figure 13.8.1. Jean Taylor and Fred Almgren at their wedding, with Rob and AnnAlmgren. Photograph courtesy of the Almgren-Taylor family.


Consider a Lipschitz deformation inside an r-ball with r � υ. The totalvolume distortion in moving S to ϕ�S� is at most r�H

n�1�S�C Hn�1�ϕ�S��

and certainly less than the volume of the r-ball. By Lemma 13.5, the volumesmay be readjusted elsewhere at cost

jAj � CjVj � Cr�Hn�1�S�C H

n�1�ϕ�S��.

Since the original cluster is minimizing,

Hn�1�S� � H

n�1�ϕ�S��C Cr�Hn�1�S�C H

n�1�ϕ�S��,

Hn�1�S� � 1 CCr

1 �CrH

n�1�ϕ�S�� 1 C 3Cr�Hn�1�ϕ�S��

for r � υ small. Therefore S is �M, ε, υ� minimal for ε�r� D 3Cr.In 1976 Jean Taylor gave a definitive mathematical explanation of the struc-

ture of soap bubble clusters recorded more than a century before by J. Plateau,based on Fred Almgren’s theory of �M, ε, υ�-minimal sets. A beautiful descrip-tion of Taylor’s work appears in a Scientific American article by Taylor andAlmgren, who supervised her Ph.D. thesis. (See also Kanigel.) As anothernice result, Almgren and Taylor were married (see Figure 13.8.1).

13.9. Regularity of Soap Bubble Clusters in R3 [Taylor, 4]. A soap bubblecluster C in R3 ��M, ε, υ�-minimal set� consists of real analytic constant-mean-curvature surfaces meeting smoothly in threes at 120° angles along smoothcurves, in turn meeting in fours at angles of cos�1��1/3� ³ 109°.

Remark. The singular curves were proved C1,˛ by Taylor [4]; C1

by Nitsche [1]; and real-analytic by Kinderlehrer, Nirenberg, and Spruck[Theorem 5.1].

Comments on Proof. Consider a linear approximation or tangent cone Cat any singularity (which exists by monotonicity and further substantial argu-ments). By scaling, C is �M, ε, υ� minimal for ε D 0 and υ D 1. C mustintersect the unit sphere in a “net” of geodesic curves meeting in threes at120°, an extension of the more familiar fact that shortest networks meet onlyin threes at 120°. (This angle comes from a balancing condition for equilib-rium. A junction of four curves could be profitably deformed to two junctionsof three as in Figure 11.3.1). In 1964 Heppes [1], unaware of much earlierincomplete work of Lamarle (1864), found all ten such geodesic nets, picturedin Figure 13.9.1. Taylor, adding a final case to complete the work of Lamarle,showed all but the first three cones to be unstable by exhibiting area-decreasingdeformations as in Figure 13.9.2. In an ironic twist of fate, in the fourth case,


Figure 13.9.1. On the sphere there are exactly 10 nets of geodesics meeting in threesat 120°, providing 10 candidate cone models for soap bubble structures. Reprinted withpermission from Almgren and Taylor. Copyright 1976 by Scientific American, Inc.All rights reserved.

the comparison surface provided by Taylor has more area than the cone; thecorrect comparison surface, which reduces area by pinching out a flat triangularsurface in the center, had been given correctly by Lamarle.

Actually back in 1964, in work not published until 1995, Heppes [2, Lemma1] had already shown all but the first three cones to be unstable. When writing


Figure 13.9.2. All but the first three cones are unstable, as demonstrated by thepictured deformations of less area. The fourth one actually should have a horizontaltriangle instead of a vertical line segment in the center. Copyright 1976 by ScientificAmerican, Inc. All rights reserved.

the paper, he checked Math Reviews, discovered Taylor’s paper, and thuslearned of Lamarle’s work for the first time. Heppes finally met Taylor in1995 at a special session on Soap Bubble Geometry in Burlington, Vermont,organized by the author.


Incidentally, an octahedral frame, whose cone, of course, cannot be in equi-librium since surfaces meet in fours, bounds at least five interesting soap films[Isenberg, color plate 4.6]. It is an open question whether the smallest of thesefive is, in fact, the minimizer.

Thus, for the approximating cone, there are just three possibilities, corre-sponding to a smooth surface or the two asserted types of singularities.

The really hard part is to show that the cone is a good enough approximationto the original soap film. Taylor uses a deep method of Reifenberg [1–3],requiring the verification of a certain “epiperimetric inequality,” which saysroughly that cones near the special three are not too close to being minimizersthemselves.

Of course, where the surface is regular, a classical variational argumentyields constant mean curvature and hence real analyticity.

Boundary Singularities. Lawlor and Morgan [1] (see also Morgan [18,Lecture I, Section 11.3, p. 91] and [Sullivan and Morgan, Problem 12])describe ten conjectured types of smooth boundary singularities of soap films,as in Figure 13.9.3.

13.10. Cluster Regularity in Higher Dimensions. Almgren [1, TheoremIII.3(7)] proved that soap bubble clusters (�M, ε, υ�-minimal sets) in Rn

�n ½ 3� are C1,˛ almost everywhere. [White 9] has proved that they consistof smooth constant-mean-curvature hypersurfaces meeting in threes at 120degrees along smooth �n� 2�-dimensional surfaces, which in turn meet infours at equal angles along smooth �n� 3�-dimensional surfaces, which meetin an �n� 4�-dimensional set.

Brakke [2] has classified the polyhedral �M, 0,1�-minimal cones in R4,which include the cone over the hypercube. It is conjectured that there are nonon-polyhedral �M, 0,1�-minimal cones below R8.

For a simple treatment of l-dimensional �M, ε, υ�-minimal sets seeMorgan [7]. (Almgren’s regularity results technically do not apply [Almgren1, IV.3(1), p. 96].) Taylor’s [4] regularity results go through for soap bubbleclusters in R2. Only in R2 does existence theory provide the alternative optionof requiring that the regions be connected [Morgan 19]; such regions maybump up against each other.

13.11. Minimizing Surface and Curve Energies. In some materials, notonly interfacial surfaces but also singular curves may carry energetic costs.This modification alters behavior qualitatively as well as quantitatively, withfour surfaces meeting along a singular curve, for example, as in Figure 13.11.1.


(1) (2)

(3) (4)

(5) (6)

(7) (8)

(9) (10)

120°

120° 120°

120°

≥ 120°

Figure 13.9.3. Ten conjectured types of soap film boundary singularities. Lawlor andMorgan [1, Figure 5.3].

Some results on existence, regularity, and structure appear in Morgan andTaylor and in Morgan [22, 3].

13.12. Bubble Cluster Equilibrium. For an equilibrium bubble cluster inRn, each region has a pressure pi defined up to a constant (the externalpressure pext), such that the mean curvature of an interface between regionsRi and Rj is pi � pj. The rate of change of area with respect to volume


Figure 13.11.1. If singular curves carry energetic cost, the degree-four singularity ofthe soap film of Figure 11.3.1 can profitably decompose into two degree-three singu-larities, with four surfaces meeting along the singular curve between them. [Morganand Taylor, Figure 3]

dA/dVi D pi � pext. Equilibrium implies that

�1�n� 1

nA D

∑�pi � pext�Vi

[Cox et al. 1, Remark 4.4]. For an ideal gas piVi D NT and hence

�2�n� 1

nA D NT� pextV.

(Depending on units, there are usually some physical constants.) Condition(2) had an early if questionable derivation by S. Ross, with reference to anearlier suggestion by Tait.


Proof of (1). Under scaling by a factor �1 C t�,

A�t� D A Ð �1 C t�n�1, Vi�t� D Vi Ð �1 C t�n,

�n� 1�A D dA

dtD

∑�pi � pext�

dVidt

D n∑

�pi � pext�Vi.

(Equivalently the cluster is stationary for enthalpy A� ∑�pi � pext�Vi, as

observed by Graner, Jiang, et al.)

CHAPTER 14

Proof of Double Bubble Conjecture

The year 2000 brought an announcement of the proof the Double BubbleConjecture in R3 by Hutchings, Morgan, Ritore, and Ros. The Double BubbleConjecture says that the “standard” familiar double soap bubble of Figure 14.0.1,consisting of three spherical caps meeting at 120 degrees, provides the least-perimeter way to enclose two prescribed volumes in R3 (or similarly in Rn).Long believed (see Plateau [pp. 300–301] and Boys [p. 120]) but not published,the conjecture finally appeared in a 1991 undergraduate thesis at WilliamsCollege by Joel Foisy, who realized that no one knew how to prove it, despitethe fact that a minimizer was known to exist and be a surface of revolution.

The 1990 Williams College NSF “SMALL” undergraduate research Geo-metry Group, including Foisy, proved the Double Bubble Conjecture in R2,using original geometric arguments and some computation.

In 1995, Joel Hass, Michael Hutchings, and Roger Schlafly announced aproof of the Double Bubble Conjecture for equal volumes in R3, using newtheoretical tools and extensive computer computations. (See also Morgan [5],Hass and Schlafly, and Hutchings.)

In 2000, Hutchings, Morgan, Manuel Ritore, and Antonio Ros announceda proof of the Double Bubble Conjecture for arbitrary volumes in R3, usinga new instability argument. (See also Cipra.)

The Williams undergraduate research Geometry Group, consisting of BenReichardt, Cory Heilmann, Yuan Lai, and Anita Spielman, generalized theresult from R3 to R4, and, for the case when the larger volume is more thantwice the smaller, to Rn.

In R5, even the case of equal volumes remains open.

141


Figure 14.0.1. The standard double bubble provides the least-perimeter way to encloseand separate two prescribed volumes. Computer graphics copyright John M. Sullivan,University of Illinois; color version at http://www.math.uiuc.edu/¾jms/Images/.

This chapter outlines the latest proofs of the Double Bubble Conjecturein R3 and Rn. In R3, a major difficulty is ruling out a bubble with threecomponents, as in Figure 14.0.2. The main new idea is to prove such bubblesunstable by rotating pieces about well-chosen axes (see Proposition 14.15).

14.1. Proposition. For prescribed volumes v, w, there is a unique standarddouble bubble in Rn consisting of three spherical caps meeting at 120 degreesas in Figure 14.0.1.

The mean curvature H0 of the separating surface is the difference of themean curvatures H1,H2 of the outer caps.

Proof. Consider a unit sphere through the origin and a congruent or smallersphere intersecting it at the origin (and elsewhere) at 120 degrees, as inFigure 14.1.1. There is a unique completion to a standard double bubble.Varying the size of the smaller sphere yields all volume ratios precisely once.Scaling yields all pairs of volumes precisely once.

The condition on the curvatures follows by the law of sines (seeFigure 14.1.2) for R2 and hence for Rn. (Plateau presented this figure forconstructing standard double bubbles.) Since curvature is proportional to

Proof of Double Bubble Conjecture 143

Figure 14.0.2. In R3, a major difficulty is ruling out a bubble with three compo-nents. Here one region consists of a central bubble and a thin toroidal tube aroundthe outside, while the second region consists of a larger toroidal tube in between.Computer graphics copyright John Sullivan, University of Illinois; color version athttp://www.math.uiuc.edu/¾jms/Images/.

pressure difference, this condition implies that the pressure is well defined(up to a constant).

14.2. Remark. Montesinos has proved the existence of a unique standardk-bubble in Rn for k � nC 1. (See also Sullivan and Morgan.)

The following symmetry theorem is based on an idea of Brian White,written up by Foisy [Theorem 3.4] and Hutchings [Theorem 2.6]. Inciden-tally, the same argument shows that an area-minimizing single bubble is around sphere.

14.3. Theorem. An area-minimizing double bubble in Rn is a hypersurfaceof revolution about a line.

Proof sketch (for details see the beautiful proof of Hutchings). In R2, thestandard double bubble, which is symmetric about a line, is known to be theunique minimizer (Foisy et al.). First, we’ll show in R3 that a symmetrizedminimizer is symmetric about a line. Second, we’ll consider Rn. Third, we’llshow that every minimizer is symmetric about a line.


Figure 14.1.1. Varying the size of the smaller component yields all volume ratiosprecisely once.

α α − 60°

1/H0

60°

1/H2

180° − α

60°1/H1

120° − α

H2

H1

H0

Figure 14.1.2. By the law of sines, the mean curvature H0 of the separating surfaceis the difference of the mean curvatures H1,H2 of the outer caps.


First, to obtain a symmetrized minimizer, start with any minimizing doublebubble in R3. We claim that there are two orthogonal planes that split bothvolumes in half. Certainly, for every 0 � � � �, there is a vertical plane atangle � to the xz-plane, which splits the first region in half. These planes canbe chosen to vary continuously back to the original position, now with thelarger part of the second region on the other side. Hence for some intermediate�, the plane splits both volumes in half. Turning everything to make this planehorizontal and repeating the argument yield a second plane, as desired. (TheBorsuk-Ulam theorem, as applied to prove the ham sandwich theorem, gives adirect proof. See e.g. Munkres, Theorem 68.6, p. 405.) Hence we may assumethat the xz- and yz-planes bisect both volumes.

Reflecting the quadrant of least area now yields a symmetrized bubble Bof no more (hence equal) area, symmetric under reflection across both planesand hence under their composition, rotation by 180 degrees about the z-axis.Hence every plane containing the z-axis splits both volumes in half.

We claim that at every regular point, the bubble B is orthogonal to thevertical plane. Otherwise the smaller or equal half of B, together with itsreflection, would be a minimizer with an illegal singularity (which could besmoothed to reduce area while maintaining volume). Now it follows that B isa surface of revolution.

Second, in Rn, one analogously finds n� 1 orthogonal hyperplanes H1,. . . ,Hn�1 of reflective symmetry, symmetrizes, deduces rotational symmetryabout the intersection Hi \Hj of any two hyperplanes, and concludes rota-tional symmetry about their common intersection L D \Hi.

Third, suppose there is some minimizer B in Rn which is not symmetricabout the line L D \Hi. Then it is not symmetric about some Hi \Hj, sayH1 \ H2. By the following Lemma 14.4, the part of B on one side of H1 andits reflection, together called B0, is not symmetric about H1 \H2. Likewisethe part of B0 on one side of H2 and its reflection, together called B00, is notsymmetric about H1 \H2. But B00 is symmetric with respect to both H1 andH2, and hence is symmetric with respect to H1 \ H2 by the first part of thisproof, the desired contradiction.

14.4. Lemma. Suppose that both regions of a minimizing double bubble Bin Rn�n ½ 3� are split in half by a hyperplane H. Let B1, B2 be the halvestogether with their reflections. If each Bi is symmetric about an �n � 2�-planePi in H, then they are both symmetric about P1 (indeed, about P1 \ P2).

Proof sketch. It suffices to show that at every regular point of B2, thetangent plane contains the direction of rotation about P1 \ P2. By symmetry,almost every point p of B \H � P1 � P2 is a regular point of B1, B2, and B


(since a singular point yields a whole singular orbit). By analytic continuation,in a neighborhood of p, the tangent plane contains the directions of rotationabout P1 \ P2. By symmetry, this holds almost everywhere in B2, exceptpossibly at points where B2 is locally contained in the orbit of P1 � P2. Bythe symmetry of B2, such a piece of surface would continue down to H, wherethe symmetry of B1 shows that as a piece of B it does not separate distinctregions, a contradiction.

14.5. Concavity (Hutchings Theorem 3.2). The least area A�v, w� of adouble bubble of volumes v, w in Rn is a strictly concave function. In particular,A�v, w� is increasing in each variable.

Proof sketch. To illustrate the idea of the proof for the interesting casen D 3, we will just prove that for fixed w0, A�v, w0� is nondecreasing, whichsuffices to prove Corollary 14.6. If not, then there is a local minimum at somev0. For simplicity, we treat just the case of a strict local minimum. Considera minimizing double bubble B of volumes v0, w0. By Theorem 14.3 and itsproof, B is a surface of revolution about a line L D P1 \ P2, where P1 andP2 are planes that divide both regions in half. Choose a plane P3 near P2

which divides the second region in half but does not contain L. We claim thatit divides the first region in half. Otherwise, the half with smaller (or equal)area, reflected across the plane, would yield a bubble of no more area andslightly different volume, contradicting the assumption that v0 is a strict localminimum. Therefore P3 splits both volumes in half. Now as in the proof ofTheorem 14.3, B is symmetric about the line L0 D P1 \ P3 as well as about L.It follows that B consists of concentric spheres, which is impossible. ThereforeA�v, w0� must be nondecreasing as desired.

14.6. Corollary (Hutchings Theorem 3.4). An area-minimizing doublebubble in Rn has connected exterior.

Proof. If the exterior has a second, bounded, component, removing asurface to make it part of one of the two regions would reduce area andincrease volume, in contradiction to A�v, w� increasing.

14.7. Corollary (Hutchings Theorem 3.5). If the larger region of a mini-mizing double bubble has more than twice the volume of the smaller region, itis connected.

Proof. If not, create a new double bubble by calling the smallest compo-nent of the first region part of the second region. Of course total volume


v0 C w0 D v C w does not change and area does not increase. Each region nowhas volume greater than w, so that �v0, w0� is a convex combination of �v, w�and �w, v�. By Concavity 14.5, area must increase, a contradiction.

14.8. Decomposition Lemma (Cox et al. [2], Hutchings Lemma 4.1).Consider a minimizing double bubble of volumes v, w in Rn. If the first regionhas a component of volume x > 0, then

2A�v, w� ½ A�v � x, w�C A�x� C A�v � x, w C x�,

where A�x� D A�x, 0� denotes the area of a sphere of volume x in Rn.

Proof. The result follows from the decomposition of Figure 14.8.1. Notethat each piece of surface on the left appears twice on the right.

2 • x w v − x = x + w v − x x + w v − x+

Figure 14.8.1. A useful decomposition.

14.9. Hutchings Basic Estimate (Hutchings Theorem 4.2). Consider aminimizing double bubble of volumes v, w in Rn. If the first region has acomponent of volume x > 0, then

�1� 2A�v, w� ½ A�v C w�C A�w� C A�v�[v/x]1/n.

In particular, a region has only finitely many components.

Proof. By Decomposition 14.8,

2A�v, w� ½ A�v � x, w�C A�x� C A�v � x, w C x�.

Now by Concavity 14.5,

A�v � x, w� ½v � x

vA�v, w�C x

vA�w�,

A�v � x, w C x� ½v � x

vA�v, w�C x

vA�v C w�.

Substituting and rearranging terms yields (1). Of course this gives a bound onv/x and hence on the number of components of the first region.


14.10. Hutchings Structure Theorem (Hutchings Theorem 5.1). An area-minimizing double bubble in Rn is either the standard double bubble or anothersurface of revolution about some line, consisting of a topological sphere with atree of annular bands (smoothly) attached, as in Figure 14.10.1. The two capsare pieces of round spheres, and the root of the tree has just one branch. Thesurfaces are all constant-mean-curvature surfaces of revolution, “Delaunaysurfaces,” meeting in threes at 120 degrees.

Proof sketch. Regularity, including the 120-degree angles, comes fromapplying planar regularity theory [Morgan 19] to the generating curves;also the curves must intersect the axis perpendicularly. The bubble must

Figure 14.10.1. A nonstandard area-minimizing double bubble in Rn would have toconsist of a central bubble with layers of toroidal bands. Drawing by Yuan Lai.


S1 S2

R2R1

Figure 14.10.2. An intermediate piece of surface through the axis must branch intotwo spheres S1, S2.

be connected, or moving components could create illegal singularities (oralternatively an asymmetric minimizer). By comparison with spheres centeredon the axis and vertical hyperplanes, pieces of surface meeting the axis mustbe such spheres or hyperplanes. We claim that the number of such piecesis two (or three for the standard double bubble). If it were 0, an argumentgiven by [Foisy, Theorem 3.6] shows that the bubble could be improvedby a volume-preserving contraction toward the axis (r ! �rn�1 � ε�1/�n�1�).If it were 1, that piece of surface would not be separating any regions. Ifit were more than 2, some piece separates the two regions and eventuallybranches into two surfaces S1 and S2, as in Figure 14.10.2. We claim that S1

and S2 must be spherical. If, for example, S1 were not spherical, replacingit by a spherical piece enclosing the same volume (possibly extending adifferent distance horizontally) would decrease area, as follows from the area-minimizing property of the sphere. Since everything else can be rolled aroundS1 or S2 without creating any illegal singularities, they must be spheres andthe bubble must be the standard double bubble. The structure theorem nowfollows, since the only possible structures are bubbles of one region in theboundary of the other. The same rolling argument implies that the root of thetree has just one branch.

The Hutchings Basic Estimate 14.9 also has the following corollary.

14.11. Corollary. Let QA�v, w� be the area of the standard double bubble inRn of volumes v, w, or any other upper bound on the minimum double bubblearea. Consider a minimizing double bubble of volumes v, 1 � v. Then the firstregion has at most k components, where

A�v�k1/n D 2 QA�v, 1 � v�� A�1� � A�1 � v�.

Proof. Let x be the volume of the smallest component of the first region,so that the number of components is bounded by k D v/x. Since, of course, theminimumA�v, w� � QA�v, w�, Corollary 14.11 follows immediately from 14.9(1).

Remark. Mathematica graphs of this function k (for n D 2, 3, 4, 5) appearin Figure 14.11.1. Some results are summarized in Table 14.11.2. In particular,


6.625

6

5

44.06

3

2.5

2

1.57

10 0.5 1

n = 5

n = 4

n = 3

n = 2

Figure 14.11.1. The function k bounding the number of components of the firstregion in a minimizing double bubble of volumes v, 1 � v in R2 through R5. B.Reichardt [Heilmann et al. 1, Figure 2].

Table 14.11.2. Bounds on the number of components

R2 R3 R4 R5 Rn

Bounds on number of components inlarger or equal region

1 1 1 2 3

Bounds on number of components insmaller region

1 2 4 6 2n

in R2 both regions are connected, from which the Double Bubble Conjecturein R2 follows easily (as in Proposition 22 of Foisy et al.). The Rn bounds areelegantly deduced from the Hutchings Basic Estimate 14.9 in Heilmann et al.[1, Proposition 5.3].

14.12. Renormalization. If in Corollary 14.11 we consider instead a mini-mizing bubble of volumes 1, w, then the bound k on the number of componentsof the first region satisfies

(1) A�1�k1/n D 2 QA�1, w�� A�w C 1�� A�w�.

14.13. Remark on rigor. There are a number of ways to prove rigorouslythe computational bounds of Table 14.11.2 apparent from the Mathematica


plot of Figure 14.11.1. Unfortunately, obtaining useful bounds on the deriva-tive of the plotted functions seems difficult as well as ugly. For R3, Hutchingset al. use a simple, weaker, convex bound to prove the larger region connected;of course for a convex bound it is enough to check the endpoints. To provethat the smaller region has at most two components, they use a completelydifferent auxiliary instability argument. For R4 (and incidentally R3), Heil-mann et al. [1] develop a piecewise convex bound to cover all volumes.Ultimately Reichardt et al. do not need any such bounds, because they gener-alize the main instability argument of Hutchings et al. to any number ofcomponents for the smaller region (as long as the larger region is connected).

The following conjecture [Heilmann et al. 1, Conj. 4.10] would perhapsprovide the most elegant way to prove the bounds rigorously. I hereby renewthe $200 offer of my Math Chat column (at www.maa.org) of October 7, 1999for the first proof.

14.14. Conjecture. In Rn, let H0,H1,H2, respectively, denote the meancurvature of a sphere of volume w, a sphere of volume w C 1, and the exte-rior of the second region of the standard double bubble of volumes 1, w, assuggested by Figure 14.14.1. Then

2H2 > H0 C H1.

Remark. Since the derivatives of the terms of 14.12(1) are proportional to themean curvatures (see Morgan [16]), Conjecture 14.14 implies that the associatedfunction k of 14.11 and 14.12 is increasing in w and hence decreasing in v.

Proposition 14.15 contains the main idea of the proof of the Double BubbleConjecture in R3. The existence of points pi where certain rotational eigen-functions vanish implies instability (unless associated surfaces are all spheresand hyperplanes, which will be shown to be impossible). This idea goes back tothe Courant Nodal Domain Theorem, which relates the divisions of a domain

w w + 1 w

H0 H1 H2

120

120

1

Figure 14.14.1. It is conjectured that the pictured curvatures satisfy H2 > �H0

CH1�/2. This conjecture would provide an elegant way to prove rigorously the boundsof Table 14.11.2.


by nodal sets to the position of the eigenvalues [Courant and Hilbert, VI-6,p. 452].

Consider a minimizing double bubble of revolution about the x-axis L inRn�n ½ 3�, with cross-section consisting of circular arcs 0 meeting the axis,and other arcs i meeting in threes, with interiors i (see Figures 14.16.1,14.17.1). Consider the map f: � L ! L [ f1g which maps p 2 to thepoint L�p� \ L, where L�p� denotes the normal line to at p. If L�p� doesnot meet L, we define the image of p as f�p� D 1.

14.15. Proposition (Hutchings et al. Proposition 5.1). Consider a mini-mizing double bubble of revolution about the x-axis L in Rn�n ½ 3�. Supposethat there is a minimal set of points fp1, . . . , pkg in [i with x D f�p1� DÐ Ð Ð D f�pk� which separates . �See Figure 14.15.1.�

Then every component of the regular set which contains some pi is part ofa sphere centered at x (if x 2 L) or part of a hyperplane orthogonal to L �inthe case x D 1�.

Proof sketch for R3. Consider rotation about the line perpendicular to thecross-sectional plane at x. The normal component u of the rotation vectorfieldvanishes on the circular orbits of the pi about L as well as in the horizontalplane containing L. Let u1, u2, u3, u4 denote the restriction of u to the fourassociated components of the cluster minus this vanishing set. Choose scalarsai such that, for example, a1 D 1, the variation aiui preserves volumes

p3

p2

p1

u1

u4

u3

L x

u2

Figure 14.15.1. Rotation about an axis through x, which is tangential at the separatingset p1, p2, p3, leads to a proof of instability.


Γ2

Γ1Γ0 Γ0

Figure 14.16.1. There is no nonstandard minimizing double bubble of revolutionwith connected regions. [Hutchings et al.]

to first order, and some aj D 0. The variation �1 � ai�ui also preservesvolume to first order. Since the bubble is assumed to minimize area, bothvariations have nonnegative second variation of area. But their sum, a rota-tion vectorfield, has second variation 0. Therefore aiui has second vari-ation 0 and is an eigenfunction. On each regular piece S of the bubblecontaining a pi, the four components all meet. Since ajuj vanishes, by uniquecontinuation for eigenfunctions a1u1 D u1 vanishes. Repeating the whole argu-ment shows that each ui vanishes. It follows that S is spherical or planar asasserted.

14.16. Corollary. There is no nonstandard minimizing double bubble in Rn

in which both regions and the exterior are connected, as in Figure 14.16.1.

Proof. The line equidistant from the two vertices intersects the axis Lin a point p (unless the line is horizontal, which we will consider next).Hence 1 and 2 each has an interior point farthest from or closest to p, sothat p 2 f�1� \ f�2�. By Proposition 14.15, 1 and 2 are both spherical,which is impossible.

If the line is horizontal, 1 and 2 each has an interior point farthest left orright, so that 1 2 f�1� \ f�2�. By Proposition 14.15, 1 and 2 are bothvertical, which is impossible.

14.17. Corollary. Consider a minimizing double bubble in Rn of three com-ponents and connected exterior as in Figure 14.17.1. Then there is no x 2 Rsuch that f�1�x�� 0 contains points in the interiors of distinct j whichseparate .

Proof. If so, there must be points in 1, 2, or 3. By Proposition 14.15,one of them is spherical. By so-called “force balancing” (Korevaar, Kusner,


Γ5

υ345υ245

Γ4

Γ1

Γ3

Γ0Γ0

Γ2

R1

R1

R2

Figure 14.17.1. A candidate double bubble with three components. [Hutchings et al.]

and Solomon), when two spherical pieces met, the third piece is also spherical.Therefore 1, 2, and 3 are all spherical. But since 2 bounds a region ofpositive pressure by Concavity 14.5 and hence 2 curves to the right from itsvertex with 1, therefore 1 and 2 cannot both be spherical.

14.18. Proposition (Hutchings et al. Proposition 5.8). There is no mini-mizing double bubble in Rn in which the region of smaller or equal pressure isconnected, the other region has two components, and the exterior is connected,as in Figure 14.17.1.

Remarks on proof. Corollary 14.17 reduces this to plane geometry, aidedby a few simple facts about surfaces of Delaunay. Finding the separatingset of points requires consideration of a number of cases, as suggested byFigure 14.18.1.

14.19. Proposition. In a minimizing double bubble in Rn, the smaller regionhas larger pressure.

Proof. Consider the function A�v, 1 � v� giving the least area enclosingand separating regions of volume v, 1 � v. By Concavity 14.5, A is strictlyconcave and of course symmetric about v D 1/2. Since one way to obtainnearby �v, 1 � v� is to vary the separating surface of mean curvature say H,with dA/dv D 2H, the left and right derivatives of A must satisfy

A0R � 2H � A0

L.

Consequently H is positive for v < 1/2 and negative for v > 1/2. In otherwords, the smaller region has larger pressure.

14.20. Theorem (Hutchings et al. Theorem 7.1). The standard doublebubble in R3 is the unique area-minimizing double bubble for prescribedvolumes.


A A

A A A

A

Figure 14.18.1. The six principal cases to be eliminated.

Proof. Let B be an area-minimizing double bubble. By Corollary 14.11 andProposition 14.19, either both regions are connected, or one of larger volume andsmaller pressure is connected and the other of smaller volume and larger pres-sure has two components. By the Hutchings Structure Theorem 14.10,B is eitheras in Figure 14.16.1 or as in Figure 14.17.1. By 14.16 and 14.18, B must be thestandard double bubble.

Remark. Although the final competitors are proved unstable, earlier stepssuch as symmetry (14.3) assume area minimization. It remains conjecturalwhether the standard double bubble is the unique stable double bubble.

To obtain results in higher dimension, Reichardt et al. generalized Propo-sition 14.18 to Proposition 14.21. Since there is no bound on the number ofcomponents of the second region, the case-by-case analysis of Figure 14.18.1must be broken down into more general arguments about constitutive parts,with the help of new ideas.

14.21. Proposition (Reichardt et al.). A minimizing double bubble in whichthe larger or equal region is connected must be the standard double bubble.

14.22. Corollary (Reichardt et al.). In R4, the standard double bubble is theunique minimizer. In Rn, for prescribed volumes v > 2w, the standard doublebubble is the unique minimizer.


Proof. By 14.7 and 14.11, the larger (or equal) region is connected. ByProposition 14.21, a minimizer must be the standard double bubble.

14.23. Open Questions. It is conjectured by Hutchings et al. that the stan-dard double bubble in Rn is the unique stable double bubble. Sullivan [Sullivanand Morgan, Proposition 2] has conjectured that the standard k-bubble in Rn

(k � n C 1� is the unique minimizer enclosing k regions of prescribed volume.This remains open even for the triple bubble in R2, although Cox et al. haveproved it minimizing in a category of bubbles with connected regions (seeSection 13.1).

One could consider the Double Bubble Conjecture in hyperbolic space Hn

or in the round sphere Sn. The symmetry and concavity results still hold[Hutchings, 3.8–3.10]. The case of S2 was proved by J. Masters.

CHAPTER 15

The Hexagonal Honeycomb and KelvinConjectures

The Hexagonal Honeycomb Conjecture says that regular hexagons as inFigure 15.0.1 provide the most efficient (least-perimeter) way to divide theplane into unit areas. A proof was announced by Thomas Hales of the Univer-sity of Michigan in 1999, the same Hales who recently proved the 1611 Keplersphere-packing conjecture (see also Peterson [2], Klarreich, and Notices AMS47(2000), 440–449).

The Kelvin Conjecture describes a candidate for the least-perimeter wayto divide three-space into unit volumes. A counterexample was announcedby Denis Weaire and Robert Phelan of Trinity College, Dublin, in 1994 (seealso Peterson [1] and Klarreich). Whether their counterexample is optimalremains open.

Figure 15.0.1. The Hexagonal Honeycomb Conjecture, proved by Hales in 1999,says that regular hexagons provide the most efficient (least-perimeter) way to dividethe plane into unit areas [Hales].

157


15.1. Hexagonal Honeycomb History. Since antiquity philosophers andhoneybees have regarded hexagons as the ideal way to partition the plane intoequal areas, as in the honeycomb of Figures 15.1.1 and 15.1.2. Around 36 BC

Marcus Terentius Varro, in his book On Agriculture [III, xvi.5], wrote aboutthe bees’ honeycomb, “The geometricians prove that this hexagon . . . enclosesthe greatest amount of space.” Zenodorus (¾200 BC) had proved the regularhexagon superior to any other hexagon, triangle, or parallelogram [Heath,pp. 206–212].

The next major advance came in 1953, when L. Fejes Toth ([2, Chapter III,§9, p. 84] or [3, Corollary Section 26, p. 183] after [4]; see also [1] or [3,Section 29, pp. 206–208] on “wet films”) used a simple convexity argument

Figure 15.1.1. The bees’ honeycomb illustrates the efficiency of using hexagons toenclose equal spaces with the least partitioning. From T. Rayment, A cluster of bees(The Bulletin, Sydney), as it appeared in D’Arcy Thompson’s On Growth and Form,p. 109.

The Hexagonal Honeycomb and Kelvin Conjectures 159

Figure 15.1.2. Bees at work. Photo by I. Kitrosser (Realites 1 (1950), Paris), as itappeared in Herman Weyl’s classic book on Symmetry [p. 84].

to prove regular hexagons superior in the very restricted category of convex (orpolygonal) regions, modulo a truncation argument as in our Proposition 15.3.Many mathematicians came to have the erroneous impression that the problemwas solved (see e.g. [Weyl, p. 85]).

Finally in 1999, Hales announced a proof in the category of connectedregions.

When I arrived at St. John’s University to give a talk on this topic, I wasgreeted by the church of Figure 15.1.3.

A difficulty with the honeycomb and its symmetry was pointed out in 1994by Gary Larson (Figure 15.1.4).


Figure 15.1.3. St. John’s Abbey — University Church, St. John’s University, myfavorite place to speak on the hexagonal honeycomb.

Figure 15.1.4. Gary Larson pointed out another difficulty with the honeycomb andits symmetry. THE FAR SIDE 1994 FarWorks Inc. used by permission. All rightsreserved.


15.2. Definition of Clusters in R2. We will define a cluster C in R2 asa smooth, locally finite graph, with each face included in a unique region(nonempty union of faces) Ri or the exterior. We will usually assume that eachRi is connected (consists of a single face). For infinite clusters, we considerthe perimeter P�r� and area A�r� inside the ball B�0, r�. The truncated clusterC0�r� consists of the n�r� regions completely contained inside the ball B�0, r�,with perimeter P0�r� and area A0�r�.

15.3. The Truncation Lemma (Morgan [8, proof of Theorem. 2.1]). Let Cbe a cluster of connected regions of area at most 1 in R2. Then

lim infr!1

P0�r�

A0�r�� lim sup

r!1P�r�

A�r�D �.

Proof. For almost all r, P0�r� exists and bounds the number of points wherecluster boundaries meet the circle S�0, r�, by the co-area formula 3.13 appliedto the function f�w� D r. Given ε > 0, r0 > 0, we can choose r ½ r0 suchthat P0�r� < εA�r�. Otherwise, for almost all r ½ r1 ½ r0,

P0 ½ εA >ε

2�P.

Since P is nondecreasing, for large r,

P�r� ½ P�r1�e�ε/2��r�r1� > 2��r2,

which implies that A�r� > �r2, a contradiction.Therefore in forming the truncated cluster C0 from the restriction of C to

B�0, r�, at most εA�r� regions are discarded, and

P0�r�

A0�r�� P�r�

�1 � ε�A�r� � �

1 � ε ,

as desired.

The following inequality is the central idea of Hales’s proof. Its penaltyterms for bulging outward and for using more than six edges make the hexagonsuperior to the circle.

15.4. Hexagonal Isoperimetric Inequality (Hales Theorem 4). Considera curvilinear planar polygon of N edges, area A at least 3.6/N2, andperimeter P. Let P0 denote the perimeter of a regular hexagon of area 1. Foreach edge, let ai denote how much more area is enclosed than by a straightline, truncated so that –1/2 � ai � 1/2. Then

�1� P/P0 ½ minfA, 1g � .5ai � c�N� 6�,


with, for example, c D .0505/2 4p

12 ³ .013, with equality only for the regularhexagon of unit area.

Remark. The truncation in the definition of ai is necessary to preventcounterexamples as in Figure 15.4.1.

Figure 15.4.1. Without truncation in the definition of the ai, this would be a coun-terexample to (1) [Hales].

Proof sketch. It is convenient to work in the larger category of immersedcurvilinear polygons, counting area with multiplicity and sign. Then onemay assume that all the edges are circular arcs (or straight lines). Withoutthe truncation condition, one could further assume that all the edges havethe same curvature, since moving to that condition at constant net areadecreases perimeter. With the truncation condition, the reductions are a littlemore complicated. Let xi denote the excess area enclosed by an edge (incomparison with a straight line, before truncation, so that, for example,jaij D minfjxij, 1/2g). We may assume that all xi 2 ��1/2, 1/2� have the samecurvature, and that all xi 2 [�1/2, 1/2] have the same sign. We may assumethat each xi ½ �1/2, since increasing it to �1/2 decreases perimeter, leavestruncated area ai at �1/2, and increases total area. Hales now treats separatelytwo cases:

Case I: some xi > 1/2,

Case II: every jxij � 1/2.

The proofs use five lower bounds on perimeter:

�LN� The perimeter of the N-gon determined by the N vertices is at least asgreat as the perimeter LN of the regular N-gon of the same area.

�LC� “Standard isoperimetric inequality.” The perimeter is at least theperimeter LC of a circle of area A, namely 2

p�A.

�L�� The figure obtained by reflecting each edge across its chord has at leastas much perimeter as a circle of the same area.


�LD� “Dido’s inequality.” Each edge is at least as long as a semicircleenclosing the same area xi.

�L0N� If all the chords have length at most 1 and jxij � �/8, then the

perimeter is at least LN times the length of a circular arc of chord 1enclosing area jxij/LN. For example when xi D 0, L0

N D LN.

This last estimate is new, applying LN to the original figure. It follows imme-diately from the Chordal Isoperimetric Inequality 15.5 below.

Before treating Cases I and II, Hales treats the easy case of digons �N D 2�,which uses nothing more than the standard isoperimetric inequality LC.

Case I also is easy, using LC, LD, and a construction that reflects the edgesenclosing negative area.

Case II is more delicate, using LC for xi very positive, L� or LD for xivery negative, and LN in between. For N ½ 100, LC and LD alone suffice. Thesubcases N D 6 or N D 7, xi ³ 0, require special attention and L0

N or theChordal Isoperimetric Inequality.

15.5. Chordal Isoperimetric Inequality (Hales Proposition 6.1). Consideran immersed curvilinear polygon as in Figure 15.5.1 of length L, net excessarea X over the chordal polygon, jXj � �/8, and each chord length at most 1.Let L0 be the length of the chordal polygon. Then the length L is at least L0

times the length arc�jXj/L0� of a circular arc of chord 1 enclosing area jXj/L0.The function L0arc�jXj/L0� is increasing in L0.

Remark. The constant �/8 can probably be improved a bit. For L0 ½ 2the sharp bound is �/4. Of course if L0 � 1, no bound on X is necessary.

Proof (a simplification of Hales’s original proof). We may assume thateach edge is a circular arc. Since the derivative of arc�x� is just the curvature ofthe arc, arc�x� is convex up to a semicircle with x D �/8, as in Figure 15.5.2.Once we reduce to arcs at most semicircles, we’ll deduce the lemma byconvexity.

We may assume that each excess area xi has the same sign, by moving twowith opposite signs closer to 0, without changing X, reducing length, until oneis 0. We may assume that X ½ 0.

We now consider the less constrained problem of enclosing area X by arcsabove chords on the x-axis of length at most 1 and total chord length L0. Theminimizer consists of circular arcs. To be in equilibrium, they must have thesame curvature, because curvature is the rate of change of length with respectto area.

It pays to lengthen chords of (equal or) higher arcs, since moving a tinysubarc from the middle of a lower arc to the middle of a higher arc would


X

L

L0

Figure 15.5.1. The chordal isoperimetric inequality estimates the perimeter L of acurvilinear polygon in terms of the perimeter L0 of the chordal polygon and the excessshaded area X.

leave L0 fixed and increase X. Now reducing the area under the lower arcto restore X reduces length. (This simplifying argument was suggested byStewart Johnson.)

Hence if L0 � 1, a minimizer is a single arc on chord L0, and its length

L0arc�jXj/L02� ½ L0arc�jXj/L0�,

as desired.On the other hand if L0 ½ 1, a minimizer consists of m ½ 1 arcs over chords

of length Li D 1 and one arc over a chord of length 0 � LmC1 < 1. The lengthof each arc is Liarc�xi/Li2�.

Since by hypothesis X � �/8, none of the arcs over unit chords is greaterthan a semicircle, and hence a lower arc of the same curvature cannot begreater than a semicircle. Hence we are in the convex range of arc. Since eachLi � 1, the total length

Liarc�xi/Li2� ½ Liarc�xi/Li� ½ L0arc�X/L0�

by convexity, as desired.Finally we show that L0arc�jXj/L0� is increasing, or equivalently that

arc�x�/x is decreasing, or equivalently that arc0�x� < arc�x�/x. Since arc�x�is convex for 0 � x � �/8 and concave for �/8 � x as in Figure 15.5.2, it


suffices to check this at �/8 (the semicircle of radius 1/2), where indeed thecurvature arc0��/8� D 2 < ��/2�/��/8�.

π/8 1 x

arc (x)

2

π/2

1

Figure 15.5.2. The length arc�x� of a circular arc enclosing area x with a unit chordis convex up to a semicircle (x D �/8) and concave thereafter. It follows that arc�x�/xis decreasing.

15.6. Proposition (Hales Theorem 2). LetC be a cluster of n regions of areaat most 1 in R2. Then the ratio of perimeter to area is greater than the ratio ofperimeter to area for the unit regular hexagonal tiling.

Proof. More generally, without assuming that each area Ai � 1, we boundthe ratio of perimeter to minfAi, 1g. By considering connected components,we may assume that the regions are connected. To obtain a contradiction, takea counterexample to minimize the number of (connected) regions. We mayassume n ½ 2, since the perimeter to area ratio for a circle of unit area isgreater than the ratio for the hexagonal tiling, which is half the ratio for aregular hexagon.

By trivial modifications we may always assume that C is a smoothconnected graph with no loops, all vertices of degree at least three, and simplyconnected faces. We may assume that C is simply connected by incorporatingany bounded components of the exterior into the regions of C by removingedges. Each region with say N edges must have area at least 2�/N2

p3, or

removing its longest edge would yield a counterexample with fewer regions


(for details, see Hales Remark 2.6). If for the moment we think of C as lyingon a sphere, then the number of vertices, edges, and regions satisfy

2 D v � eC n �including the exterior�.

Each region with N edges contributes at most N/3 to v and exactly N/2 toe, because, for example, each edge bounds two regions (or possibly the sameregion twice). Therefore

2 � �N/3 �N/2 C 1� D �1 �N/6� �including the exterior�.

If we remove the contribution of the exterior region, we have

1 � �1 �N/6� �C, without the exterior�.

Hence if we sum the Hexagonal Isoperimetric Inequality 15.4(1) over allregions, the sum of the �c�N� 6� terms is positive. The sum of the ai termsfrom interior edges is 0, because the two regions sharing an edge make oppo-site contributions. The length Pi of each interior edge contributes twice to thesum of the perimeters. Therefore summing (1) yields the proposition exceptfor two discrepancies: the favorable undercounting of exterior perimeter Piand the possibly unfavorable contribution of any exterior .5ai > 0. But byDido’s inequality LD,

Pi ½√

2�ai D ai√

2�/ai ½ aip

4� > .5ai,

because ai � 1/2.

Proposition 15.6 now has the following corollary:

15.7. The Hexagonal Honeycomb Theorem. Let C be a planar cluster ofinfinitely many regions of unit area. Then the limiting perimeter to area ratio

� D lim supr!1

P�r�

A�r�

satisfies � ½ �0, where �0 is the ratio for the unit regular hexagonal tiling.

Proof. We actually prove the result for regions of area at most 1. Thenby considering the connected components, we may assume that the regionsare connected. By the Truncation Lemma 15.3, � is greater than or equal to alimit of ratios for finite truncations. Therefore by Proposition 15.6, � ½ �0.

15.8. The Bees’ Honeycomb. The bees actually have a more complicated,three-dimensional problem involving how the ends of the hexagonal cells areshaped to interlock with the ends of the cells on the other side. L. Fejes Toth[5], in his famous American Mathematical Monthly article on “What the bees


Figure 15.9.1. Candidate minimizers for two prescribed areas A1 > A2. The struc-tures should relax to circular arcs meeting at 120 degree angles. (From Morgan [8,Figure 2], based on figure from Branko Grunbaum and G. C. Shephard, Tilings andPatterns, Freeman and Co, New York, 1987.)

know and what they do not know,” showed that the bees’ three-dimensionalstructure can be improved slightly, at least for the mathematical model withinfinitely thin walls.

15.9. Unequal areas. One can generalize the planar partitioning problem tofinitely many prescribed areas A1, . . . , Ak and prescribed probabilitiesp1 C Ð Ð Ð C pk D 1 such that as r ! 1 the fraction of regions within B�0, r�of area Ai approaches pi. For approximately equal areas, the minimizer prob-ably consists essentially of regular hexagons, as G. Fejes Toth has shown canbe accomplished by partitions of the plane. Figure 15.9.1 suggests two other


candidate minimizers for two equally likely prescribed areas (k D 2, A1 >A2, p1 D p2 D 1/2), which do better than regular hexagons. Figure 15.9.1adoes better approximately for .117 < A2/A1 < .206, while Figure 15.9.1b doesbetter approximately for A2/A1 < .039. Figure 15.9.1c never does better.

15.10. Kelvin Conjecture Disproved by Weaire and Phelan. 1994 broughtstriking news of the disproof of Lord Kelvin’s 100-year-old conjecture byDenis Weaire and Robert Phelan of Trinity College, Dublin. Kelvin sought theleast-area way to partition all of space into regions of unit volumes. (Sincethe total area is infinite, least area is interpreted to mean that there is noarea-reducing alteration of compact support preserving the unit volumes.) Hisbasic building block was a truncated octahedron, with its six square facesof truncation and eight remaining hexagonal faces, which packs perfectly tofill space as suggested by Figure 15.10.1. (The regular dodecahedron, withits twelve pentagonal faces, has less area, but it does not pack.) The wholestructure relaxes slightly into a curvy equilibrium, which is Kelvin’s candidate.All regions are congruent.

Kelvin loved this shape, constructed models, and exhibited stereoscopicimages as in Figure 15.10.2.

Weaire and Phelan recruited a crystal structure from certain chemical“clathrate” compounds, which uses two different building blocks: theregular dodecahedron and a tetrakaidecahedron with 12 pentagonal faces and

Figure 15.10.1. Lord Kelvin conjectured that the least-area way to partition spaceinto unit volumes uses relaxed truncated octahedra. Graphics by Ken Brakke in hisSurface Evolver from Brakke’s early report [5].


Figure 15.10.2. Kelvin loved his truncated octahedron, constructed models, andexhibited the pictured stereoscopic images. (Crossing your eyes to superimpose thetwo images produces a three-dimensional view.) [Thomson, p. 15].

Figure 15.10.3. The relaxed stacked tetrakaidecahedra and occasional dodecahedraof Weaire and Phelan beat Kelvin’s conjecture by about 0.3%. Graphics by Ken Brakkein his Surface Evolver from Brakke’s early report [5].

2 hexagonal faces. The tetrakaidecahedra are arranged in three orthogonalstacks, stacked along the hexagonal faces, as in Figure 15.10.3. The remainingholes are filled by dodecahedra. Again, the structure is allowed to relax into astable equilibrium. Computation in the Brakke Evolver shows an improvementover Kelvin’s conjecture of about 0.3%. The rigorous proof, by Kusner andSullivan [1], with further details by Almgren et al., proves only about 0.01%.


In greater detail, the centers of the polyhedra are at the points of a latticewith the following coordinates modulo 2:

0 0 01 1 10.5 0 11.5 0 10 1 0.50 1 1.51 0.5 01 1.5 0

Given a center, the corresponding polyhedral region is just the “Voronoi cell”of all points closer to the given center than to any other center. The relaxationprocess also needs to slightly adjust the volumes to make them all 1.

Weaire and Phelan thus provided a new conjectured minimizer. Weaire’spopular account in New Scientist gives further pictures and details (see alsoKlarreich). Don’t miss the pictures of Kelvin, Weaire, and Phelan in DiscoverMagazine’s comic-book version by Larry Gonick.

15.11. Higher Dimensions. Kelvin’s truncated octahedron is actually ascaled “permutohedron,” the convex hull of the 24 permutations of (1, 2, 3, 4) inR3 D fx 2 R4:xi D 10g. Likewise the regular hexagon is the permutohedronin R2 D fx 2 R3:xi D 6g. Will permutohedra turn out to relax into optimalpartitions in higher dimensions? To the contrary, John M. Sullivan (personalcommunication, 1999) conjectures that the least-area partitioning of R4 intounit volumes is given by what Coxeter [Chapter VIII] calls the regularhoneycomb f3, 4, 3, 3g, with octaplex cells f3, 4, 3g, each with 24 octahedralfaces. Intriguingly enough, this polyhedral foam has no need of relaxation. Itssurfaces and lines all meet at the ideal angles. Likewise in R8, there is such acandidate based on the so-called E8 lattice. See Conway and Sloane.

15.12. How the Weaire-Phelan Counterexample to the Kelvin ConjectureCould Have Been Found Earlier. The clathrate compounds that inspiredWeaire and Phelan had just three years earlier inspired counterexamples byTibor Tarnai to related conjectures on the optimal way to cover a sphere withdiscs (see Stewart).

K. Brakke spent hours at his father’s old desk seeking counterexamples.Had he reached up and pulled down his father’s copy of Linus Pauling’sclassic, The Nature of the Chemical Bond, it would doubtless have fallen opento the illustration, in the clathrate compound section, of the chlorine hydratecrystal, essentially the Weaire-Phelan counterexample!


R. Williams, after spending years seeking a Kelvin counterexample, finallygave up and later published a well-illustrated The Geometrical Foundation ofNatural Structure: A Source Book of Design. In his Figure 5.22, he picturedthe Weaire-Phelan counterexample without realizing it. Of course, it wouldhave been hard to check without Brakke’s Surface Evolver.

15.13. Conjectures and Proofs. Optimal partitioning, as described in thischapter, is much harder than optimal packing or covering. Hales’s proof forthe hexagonal honeycomb in R2 did not come until 1999. Will the new Weaire-Phelan candidate for R3 take another century to prove?

EXERCISES

15.1. The perimeter of a finite cluster equals half the sum of the perimeters of theregions and the perimeter of their union. Moreover, the perimeter of a region isat least the perimeter of a round disc of the same area. Use these two facts togive a short proof of Proposition 15.6 for n � 398.

CHAPTER 16

Immiscible Fluids and Crystals

16.1. Immiscible Fluids. Clusters of immiscible fluids F1, . . . , Fm (withambient F0) such as oil, water, and mercury in air tend to minimize an energyproportional to surface area, where now the constant of proportionality aij > 0depends on which fluids the surface separates. We might as well assumetriangle inequalities aik � aij C ajk , since otherwise an interface between Fi

and Fk could profitably be replaced by a thin layer of Fj.

16.2. Existence of Minimizing Fluid Clusters. The existence of least-energy clusters of immiscible fluids in Rn follows as for soap bubbleclusters (13.4). Technically it is very convenient to use flat chains with “fluid”rather than integer coefficients, after Fleming [1], White [2], and Morgan [9]so that, for example, the superposition of an oil-water interface and a water-mercury interface is automatically an oil-mercury interface.

16.3. Regularity of Minimizing Fluid Clusters. Now suppose stricttriangle inequalities hold. White [6; 2, Section 11] has announced thatminimizing fluid clusters consist of smooth constant-mean-curvaturehypersurfaces meeting along a singular set of Hausdorff dimension at mostn � 2. A form of monotonicity holds [Morgan 9, Section 3.2]. G. Leonardihas proved that if two fluids occupy most of a ball in a minimizer, then theyoccupy all of a smaller ball.

The structure of the singularities is not as well understood as for soapbubble clusters. In planar singularities, many circular arcs can meet at an

173


isolated point. A conjecture on the classification of energy-minimizing conesin R2 was recently proved for up to five fluids and disproved for six by theWilliams College NSF “SMALL” undergraduate research Geometry Group[Futer et al.].

Examples of energy-minimizing cones in R3 include cones over the edgesof any tetrahedron, a regular polygonal prism, the regular octahedron, theregular dodecahedron, the regular icosahedron, and the cube; if, however, youmove just one vertex of the cube, the cone is not minimizing for any choiceof weights aij (Morgan, unpublished). It is an open question whether everycombinatorial type of polyhedron occurs as an energy-minimizing cone.

16.4. Crystals (see Taylor [2], Feynman, Morgan [16, Chapter 10]). Clus-ters of crystals C1, . . . , Cm tend to minimize an energy that depends ondirection as well as the pair of crystals separated, an energy given by normsij (cf. 12.5), sometimes assumed to be even so that ij �v� D ij v�.

Figure 16.4.1. Crystal shapes typically have finitely many flat facets corresponding tosurface orientations of low energy. (The first two photographs are from Steve Smale’sBeautiful Crystals Calendar. The third photograph is from E. Brieskorn. All threeappear in The Parsimonious Universe by S. Hildebrandt and A. Tromba [pp. 263–264]and in Morgan [16, Figure 10.1].)

Immiscible Fluids and Crystals 175

For a single crystal of prescribed volume, the unique energy minimizer is thewell-known Wulff shape (ball in dual norm; see Morgan [16, Chapter 10] andreferences therein). Typically the norm is not smooth, certain directions aremuch cheaper than others, and the Wulff shape is a polytope, like the crystalsof Figure 16.4.1. For salt, horizontal and vertical axis-plane directions arecheap and the Wulff shape is a cube.

If the energy functional is lower-semicontinuous, then the existence of least-energy clusters of crystals in Rn follows as for soap bubble clusters (13.4).Ambrosio and Braides [Example 2.6] (see Morgan [10]) show that the triangleinequalities ik � ij C jk do not suffice to imply lower-semicontinuity.

There are no general regularity or monotonicity results. For the case whenthe ij are multiples aij of a fixed norm, Leonardi [Remark 4.7] has provedthat almost every point in the surface has a neighborhood in which the surfaceis a minimizing interface between two regions (for fixed volumes).

16.5. Planar Double Crystals of Salt. For the case of double crystals ofsalt in R2, Morgan, French, and Greenleaf proved that three types of mini-mizers occur, as in Figure 16.5.1. (Here the energy norm is given by x, y� Djxj C jyj, so that horizontal and vertical directions are cheap.) Electron micro-scope photographs of table salt crystals in Figure 16.5.2 show similar shapes.Undergraduates Wecht, Barber, and Tice considered a refined model in whichthe interface cost is some fraction � of the external boundary cost. Theyfound the same three types (with different dimensions), except that when� � �0 ³ .56, the minimizer jumps directly from Type I to Type III (as shownin Figure 16.5.3).

A B B

A

B

A

Figure 16.5.1. The three types of minimizing double crystals of salt.

16.6. Willmore and Knot Energies. Physical rods or membranes may tendto minimize nonlinear elastic energies such as the Willmore energy

∫H2

(where H is the mean curvature). In 1965 [Willmore, Section 7.2] showedthat the round sphere minimizes this energy. The Willmore Conjecture says


Figure 16.5.2. Electronic microscope photographs of table salt crystals show shapessimilar to those of Figure 16.5.1. [Morgan, French, and Greenleaf, Figure 2.]


10

8

6

4

2

00 0.2 0.4 0.6

λ0.8 1

Type III

Type II

Type I

R = BA

Figure 16.5.3. If the interface carries just a fraction � of the exterior boundarycost, the transition to Type III occurs sooner, without passing through Type II when� � �0 ³ .56. [Wecht, Barber, and Tice, Figure 4.]

that among tori, a standard “Clifford” torus of revolution (with circular cross-section) minimizes energy. L. Simon [2] proved existence and regularity fora minimizing torus.

Kusner and Sullivan [2] studied a modified electrostatic energy on knotsand links and provided stereoscopic pictures of (nonrigorously) computedminimizers.

16.7. Flows and Crystal Growth. Energy minimization guides not onlyequilibrium shapes but also dynamical processes such as crystal growth. Lengthor area minimization gives rise to flows in the direction and magnitude ofthe curvature. Matt Grayson proved that a smooth Jordan curve in a closedsurface flows by curvature to a round point or a geodesic. Gerhard Huiskenproved that a smooth uniformly convex hypersurface in Rn flows by meancurvature to a round point. (Some nonconvex surfaces, such as a dumbbell,can develop singularities.) Other factors such as heat flow further complicatecrystal growth. There are many definitions of flow, which agree in the nicestcases (see e.g. the survey on “Geometric models of crystal growth” by Taylor,Cahn, and Handwerker). Some recent approaches require the full strength ofgeometric measure theory (see e.g. Almgren, Taylor, and Wang). Hubert Brayused a simultaneous flow of surfaces and metrics in his recent proof of the


Riemannian Penrose Conjecture in general relativity, which says roughly thatthe square of the mass of the universe is at least as great as the sum of thesquares of the masses of its black holes. This result uses and generalizes thePositive Mass Theorem of Schoen and Yau, which just says that the mass ofthe universe is positive.

Francis, Sullivan, Kusner, et al. used elastic energy gradient flow on Brak-ke’s Surface Evolver to turn a sphere inside out, starting as in Figure 16.7.1.The flow is downward in two directions from the halfway energy saddle pointpictured in Figure 16.7.2.

Figure 16.7.1. The Optiverse, a beautiful video by Sullivan, Francis, and Levy,begins its optimal way to turn a sphere inside out. http://new.math.uiuc.edu/optiverse/.

16.8. The Brakke Surface Evolver. Ken Brakke has developed andmaintained magnificent software for the computer evolution of surfaces, fromsoap films to crystal growth, with beautiful graphics. It was Brakke’s SurfaceEvolver that established the Weaire-Phelan counterexample to the Kelvin


Figure 16.7.2. The saddle configuration halfway through the Optiverse sphereeversion. http://new.math.uiuc.edu/optiverse/.

Conjecture (Chapter 15). It was used to help redesign the fuel tanks of the USspace shuttle.

A package containing the source code, the manual, and sample data filesis freely available at the Evolver Web site at http://www.susqu.edu/facstaff/b/brakke/evolver/.

CHAPTER 17

Isoperimetric Theorems in GeneralCodimension

The classical isoperimetric inequality (Section 13.2) for the volume of aregion in Rn in terms of its perimeter, maximized by the round ball, hasimportant generalizations to higher codimension and to other ambients. Itwas not until 1986 that the classical isoperimetric inequality was extendedto general codimension by Fred Almgren (Figure 17.0.1). While in codimen-sion 0 there is a unique region with given boundary, in higher codimensionthere are many surfaces (of unbounded area) with given boundary, and theisoperimetric inequality applies only to the one of least area.

17.1. Theorem (Almgren [2]). An m-dimensional area-minimizing integralcurrent in Rn�2 � m � n� has no more area than a round disc of the sameboundary area, with equality only for the round disc itself.

Remark. Previously for m < n this result was known just for m D 2. Inthat case the boundary, a curve in Rn, may by decomposition be assumed to bea closed curve through the origin. The cone over the curve can be developedin the plane and thus shown to have no more area than a round planar disc(see Federer 4.5.14).

Proof sketch. As in 13.7, among area-minimizing surfaces with the samearea as the round unit disc, there exists a Q of least boundary area.

The second step is to show that ∂Q has mean curvature at most 1 (the meancurvature of the boundary of the unit disc). Since ∂Q is not known to besmooth, this statement is formulated weakly, in terms of the first variation of∂Q (see Section 11.2).

181


(a)

(b)

Figure 17.0.1. (a) Fred Almgren always presented his favorite ideas with generousenthusiasm. (b) Fred joyfully waves his honorary flag at a celebratory dinner. Seatedon the near side of the table, starting on the right, are Aaron Yip, Jenny Kelley,David Caraballo, Karen Almgren, Jean Taylor, Elliott Lieb, and Gary Lawlor (withEd Nelson barely visible between and behind Lawlor and Lieb). At the head of thetable and then along the far side are Frank Morgan, Joe Fu, Mohamed Messaoudene,Melinda Duncan, Andy Roosen, Fred Almgren, Christiana Lieb, and Dana Mackenzie.Photos by Harold Parks, courtesy of Parks. These photos also appeared in the Almgrenmemorial issue of the Journal of Geometric Analysis 8 (1998).

Isoperimetric Theorems in General Codimension 183

To make the ensuing discussion specific we treat the case of a 2-dimensionalarea-minimizing surface Q of area in R3, chosen to minimize the length ofits boundary curve C. Also we will suppress some relatively routine concernsabout singularities.

Note that flowing the boundary of the unit disc (with unit curvature) byits curvature and then rescaling to the original enclosed area just restoresthe unit disc, with no change in enclosed area. Moreover, this flow, directlyinto the surface, realizes a general upper bound on the reduction of leastarea enclosed; any smaller new minimizer, together with the strip of surfaceswept out by the flowing boundary, would have less area than the originalminimizer, a contradiction. By comparison, if C somewhere had curvaturegreater than 1, a local variation followed by rescaling would reduce the lengthof C, a contradiction.

The third step is to show that the length of a curve with curvature boundedby 1 is at least 2, with equality only for a round circle. Consider theGauss map G mapping each point of the boundary B of the convex hull of Cto the unit sphere. Because B � C is generally flat in one direction as inFigure 17.1.1, the entire area 4 of the unit sphere is due to the singularcontribution along B \ C, where the surface has a dihedral angle 0 � ˛ � .The contribution of an element ds to the area 4 of the sphere is proportionalto �˛ ds. Consideration of the example of the unit circle with � D 1 and ˛ D shows the constant to be 2/. Therefore

4 D 2

∫C

�˛ ds � 2

1jCj

so that jCj ½ 2, as desired.If equality holds, then ˛ D , so that the convex hull has no interior and

must be planar; moreover � D 1, and C must be a round unit circle.

C

B

90°

180°

Figure 17.1.1. The boundary B of the convex hull of the curve C is generally flatin one direction, so that the entire area 4 of the image of the Gauss map is due tothe singular contribution along B \ C, depending on a dihedral angle ˛, here often90 degrees, while 180 degrees for a flat circle.


The following theorem of Almgren and Allard gives a general isoperi-metric inequality for m-dimensional minimal (stationary, not necessarily area-minimizing) surfaces (possibly with singularities) in Rn. It is conjectured thatthe round disc is the extreme case, but this remains open even for smoothminimal surfaces of several boundary components in R3; see Section 4 of thesurvey on “The isoperimetric inequality” by Osserman [1], and [Sullivan andMorgan, Problem 22].

17.2. Theorem (Allard Theorem 7.1 and Corollary 7.2). There is a constantc�m, n� such that an m-dimensional minimal submanifold with boundary ofRn has no more than c times the area of a round disc of the same boundaryarea. More generally, any submanifold M with boundary and mean curvaturevector H satisfies

�1� jMj�m�1�/m � c

(j∂Mj C

∫M

jH j)

.

Still more generally, any compact integral varifold V satisfies

�2� jVj�m�1�/m � cjυVj.Here j j is used for total area or measure.

Remark. The proof uses monotonicity (and a covering argument), andhence does not generalize to more general integrands. Of course, minimizersfor general integrands satisfy certain inequalities by comparison with mini-mizers for area.

17.3. General Ambient Manifolds. As described in Section 12.3, Federerprovides a very general isoperimetric inequality for area-minimizing surfacesin a general Riemannian manifold M with boundary, with little informationabout how the constant depends on M. Hoffman and Spruck [Theorem 22]give a delicate generalization of 17.2(1) to small submanifolds of a Rieman-nian manifold, depending on the mean curvature of the submanifold and thesectional curvature of the ambient.

In codimension 0, Yau [Section 4] (using divergence of the distance func-tion) and Croke (using integral geometry) give linear and standard isoperi-metric inequalities depending on diameter, volume, and Ricci curvature.

The Aubin Conjecture seeks sharp isoperimetric inequalities in a simplyconnected Riemannian manifold of sectional curvature K � K0 � 0. This hasbeen proved only in dimension two, and in dimension three by Bruce Kleinerby a generalization of Almgren’s proof of Theorem 17.1; see Morgan andJohnson.

Solutions to Exercises

Chapter 22.1.

I1�I� � 1

ˇ�2, 1�

∫p2 0Ł�2,1�

∫y 2 Imp

N�p/I, y�d L1dp

D

2

∫

0j cos �jd�

D 1.

2.2. Coverings by n intervals of length 1/n show that H1�I� � 1. Suppose H

1�I� < 1.Then there is a covering fSjg of I with

∑diam Sj < 1.

By slightly increasing each diam Sj if necessary, we may assume that the Sj areopen intervals �aj, bj�. Since I is compact, we may assume that there are onlyfinitely many. We may assume that none contains another. Finally we may assumethat a1 < a2 < Ð Ð Ð < an and hence bj > ajC1. Now

n∑jD1

diam Sj Dn∑

jD1

�bj � aj� ½n�1∑jD1

�ajC1 � aj�C �bn � an�

D bn � a1 > 1,

the desired contradiction.2.3. Covering [�1, 1]n by �2N�n cubes of side 1/N and radius

pn/2N yields

Hn�Bn�0, 1�� H

n�[�1, 1]n� � lim�2N�n˛n�p

n/2N�n D ˛nnn/2 <1.

185


2.4. For each υ > 0, there is a cover fSj�υ�g of A with diam�Sj�υ�� υ and

∑˛m

(diam Sj�υ�

2

)m

� Hm�A�C ε <1

Consequently,

lim∑

˛k

(diam Sj�υ�

2

)k

D ˛k

˛mlim

∑˛m

(diam Sj�υ�

2

)m (diam Sj�υ�

2

)k�m

� ˛k

˛m�H

m�A�C ε� lim(

υ

2

)k�m

D 0.

Therefore Hk�A� D 0.

It follows that for a fixed set A, there is a nonnegative number d such that

Hm�A� D

{1 if 0 � m < d,0 if d < m <1.

All four definitions of the Hausdorff dimension of A yield d . Incidentally, Hd�A�

could be anything: 0,1, or any positive real number, depending on what A is.2.5. The 3j triangular regions of side 3�j which make up Aj provide a covering of A

with ˛1�diam /2�1 D 1. Hence H1�A� � 1.

The opposite inequality is usually difficult, but here there is an easy way. Let denote projection onto the x-axis. Then H

1�A� ½ H1��A�� D 1.

2.6. (a) C can be covered by 2n intervals of length 3�n.(b) Suppose H

m�C� < ˛m/2m, so there is a covering by intervals Si with�diam Si�m < 1. By slight enlargement, we may assume that each Si is open.Since C is compact, we may assume that fSig is finite, of minimal cardinality.If no Si meet both halves of C, a scaling up of the covering of one of the twohalves yields a cheap covering of C of smaller cardinality, a contradiction. Ifexactly p ½ 1 Si’s meet both halves of C, replace each by

{x 2 Si: x � 1

3

}and

{x 2 Si: x ½ 2

3

}.

For 0 � s, t � 13 ,

f�s, t� � (sC 1

3 C t)m � sm � tm ½ 0,

because the partial derivatives are negative and f� 13 ,

13 � D 0; therefore, the new

covering still satisfies �diam Si�m < 1. p � 2 Si’s meet both halves, since anyin addition to the leftmost and rightmost such would be superfluous. If theoriginal covering had (minimal) cardinality n, the new one has cardinality atmost nC p. A scaling up of the covering of one of the two halves yields a cheapcovering of C of cardinality at most �nC p�/2. Since �nC p�/2 ½ n, n � p.

Solutions to Exercises 187

If n D 1, �diam Si�m ½ 1, a contradiction. Otherwise n D p D 2, each Si hasa diameter of at least 1

3 , and �diam Si�m ½ 1, a contradiction.2.7.

f =0

f Blowsup Fast

Set ofDensity 0at 0

2.8. Let ε > 0. Suppose 0 is a Lesbesgue point of f. Then

1

˛mrm

∫Bm�0,r�

jf�x�� f�0�jd Lm ! 0.

Consequently,

1

˛mrmL

mfx 2 Bm�0, r�: jf�x�� f�0�j ½ εg ! 0.

Therefore, f is approximately continuous at 0.2.9. Following the hint, let a 2 \Ei. It suffices to show that f is approximately

continuous at a, since by Corollary 2.9, almost every point lies in \Ei.Given ε > 0, choose

f�a�� ε < qi < f�a� < qj < f�a�C ε.

Then

�fjf�x�� f�a�j ½ εg, a�� ff�x� < qig, a�C�ff�x� > qjg, a� D 0C 0,

because a 2 Ei and a 2 Ej.

Chapter 33.1. Let fqig be an enumeration of the rationals and let

f�x� D1∑iD1

2�ijx � qij.


3.2. Given ε > 0, f is approximately differentiable at the points of density 1 offx 2 A: f�x� D g�x�g, i.e., everywhere except for a set of measure < ε.

3.3. On all of R, one can just take f�x� D x2. On [�1, 1], one can take f�x� D 3p

x.

3.4. J1f D{

1 for r ½ 11/r for r � 1

3.5.J1f � maxfDf�u�: u unit vectorg

D maxfrf Ð ug D jrfj.3.6.

H2�S2�0, 1�� D

∫ 2

0

∫

0J2fdϕ d�.

J2f D sin ϕ.The answer is 4.

3.7.

LHS D∫

AJ1f d L

3 D∫

B�0,R�2r d L

3 D∫ R

0�2r��4r2� dr

D 2R4.

RHS D∫ R2

04y dy D 2R4.

3.8. (a) Apply the area formula, 3.13, to f: Eð [0, 1] ! C given by f�x, t� D xt.

HmC1�C� D

∫Eð[0,1]

JmC1f D∫

Eð[0,1]tm D a0

m C 1.

Alternatively, apply the coarea formula, 3.13, to f: C! R given by f�x� Djxj. Then

HmC1�C� D

∫ 1

0a0y

m D a0

m C 1.

(b) mC1�C, 0� D limH

mC1�C \ B�0,r��˛mC1rmC1

D lima0rmC1/�m C 1�

˛mC1rmC1

D a0

˛mC1�m C 1�.

(c) The cone over the closure of fx 2 E: m�E, x� 6D 0g equals the cone overfx 2 E: H

m�B�x, r� \ E� > 0 for all r > 0g.3.9. Let fqig be an enumeration of the points in R3 with rational coordinates. Let

E D [1iD�1S�qi, 2�i�. By 3.12, 2�E, x� D 1 for almost all x 2 E. It follows thatfx 2 R3: 2�E, x� D 1g is dense in R3.


Chapter 44.1. �6e134 � 12e234.4.2. One possibility is

u D ��1, 0, 1,�1�,

v D �0,�1, 1,�1�,

w D(� 1p

3, 0,

1p3,� 1p

3

).

z D(

2p15

,� 3p15

,1p15

,� 1p15

).

u ^ v D e12 � e13 C e14 C e23 � e24

w ^ z D 1p5

e12 � 1p5

e13 C 1p5

e14 C 1p5

e23 � 1p5

e24

D 1p5u ^ v

jw ^ zj D 1.

4.4. e12 C 2e13 C 2e23 D �e1 C e2� ^ �e2 C 2e3�.4.5. Method 1: Assume e12 C e34 D �aiei� ^ �bjej�, and derive a contradiction.

Method 2: Clearly, if 0 is simple, 0 ^ 0 D 0. Since �e12 C e34� ^ �e12 C e34� D2e1234 6D 0, it is not simple. (Actually, for 0 2 2Rn, 0 simple , 0 ^ 0 D 0. For0 2 mRn, m > 2, 0 simple ) 0 ^ 0 D 0.)

4.6.∫ 1

0

∫ 10 he12, ϕi dx2 dx1, he12, ϕi D x1 sin x1x2.

Inside integral D � cos x1x2]1x2D0 D 1� cos x1.

Outside integral D x1 � sin x1]1x1D0 D 1� sin 1.

4.7. The surface is a unit disc with normal e1 C e2 C e3, unit tangent 0 D �e12 C e23 �e13�/

p3. ϕ�0� D 4/

p3. Integral D 4/

p3.

4.8. If I 2 Im, then ∂I 2 R m�1 by definition and ∂�∂I� D 0 2 R m�2. Therefore, ∂I 2Im�1. If F 2 F m, then F D TC ∂S, with T 2 R m, S 2 R mC1. Since ∂F D ∂T, ∂F 2F m�1.

To prove that spt ∂T ² spt T, consider a form ϕ 2 Dm�1 such that spt ϕ \

spt T D ∅. Then spt dϕ \ spt T D ∅, and consequently ∂T�ϕ� D T�dϕ� D 0. Weconclude that spt ∂T ² sptT.

4.9. (a) T�f dx C g dy� D∫ 1

0f�x, 0� dx.

∂T�h� D T

[∂h

∂xdx C ∂h

∂ydy

]

D∫ 1

0

∂h

∂x�x, 0� dx D h�1, 0�� h�0, 0�.


Hence ∂T D H0 f�1, 0�g � H

0 f�0, 0�g.

∂T

− +

(0,0) T (0,1)

(b) Let E D f�x, x�: 0 � x � 1g.

T�f dx C g dy� D∫

E3p

2�f�t, t�C g�t, t�� d H1

D∫ 1

03p

2�f�t, t�C g�t, t��p

2 d t

D 6∫ 1

0�f�t, t�C g�t, t�� d t.

∂T�h� D 6∫ 1

0

[∂h

∂x�t, t�C ∂h

∂y�t, t�

]d t

D 6�h�1, 1�� h�0, 0��.

T

(1, 1)

(0, 0)

∂T

Mult

iplici

ty 6

Notice that

T D 6H1 E ^ e1 C e2p

2.

4.10. It follows from Theorem 4.4(1) that Im is M dense in R m. To show that Im isF dense in Fm, let R 2 Fm, so that R D TC ∂S, with T 2 Rm and S 2 RmC1.Given ε > 0, choose T1 2 Im, S1 2 ImC1, such that M�T1 � T�CM�S1 � S� <ε. Then T1 C ∂S1 2 Im, and F��T1 C ∂S1�� R� D F��T1 � T�C ∂�S1 � S�� M�T1 � T�CM�S1 � S� < ε.


4.11. It follows from Theorem 4.4(1) that A D fT 2 Rm: spt T ² B�0, R�g is M com-plete. To show that B D fT 2 Fm: spt T ² B�0, R�g is F -complete, let Rj be aCauchy sequence in B. By taking a subsequence, we may assume that F �RjC1 �Rj� < 2�j. Write RjC1 � Rj D Tj C ∂Sj with M�Tj�CM�Sj� < 2�j. Since A isM complete, Tj converges to a rectifiable current, T, and Sj converges to arectifiable current S. It is easy to check that Rj ! R1 C TC ∂S 2 Fm.

4.12. That ∂ carries Nm into Nm�1 follows immediately from the definition of Nm andthe fact that ∂∂ D 0. Since ∂ is F continuous, it follows that ∂ carries Fm intoFm�1.

4.13. The first follows immediately from the definitions. The second is the definitionof Rm.

4.14. The first follows immediately from the definitions. By Exercise 4.12, Fm ¦fTC ∂S: T 2 Rm, S 2 RmC1g. Conversely, suppose R 2 Fm. Let Nj 2 Nm withF�Nj � R� < 2�j�1 and hence F�NjC1 �Nj� < 2�j. Hence �NjC1 �Nj� D Aj C∂Bj for currents Aj and Bj with M�Aj�CM�Bj� < 2�j. Since M�∂Bj� DM�NjC1 �Nj � Aj� <1, Bj 2 NmC1. Therefore Aj D NjC1 �Nj � ∂Bj 2 Nm.By Proposition 4.6, Aj 2 Rm and Bj 2 RmC1. Finally, R D �N1 CAj�C∂Bj is of the form TC ∂S, as desired.

4.15. That Im ² Nm follows immediately from the definitions. Next,

F m ² F -closure of Im ² F-closure of Nm D Fm;

the first inclusion follows from Exercise 4.10, whereas the second followsbecause Im ² Nm and F � F . Finally Rm ² Rm, because if T 2 Rm, then T 2Fm ² Fm and M�T� <1.

4.16. (a) T�f dx C g dy� D1∑kD1

∫ 2�k

0g�k�1, y� dy.

∂T�h� D1∑kD1

h�k�1, 2�k��1∑kD1

h�k�1, 0�.

T 2 R1.

(b) T�f dx C g dy� D∫ 1

0

∫ 1

0f�x, y� dx dy.

∂T�h� D∫ 1

0h�1, y� dy �

∫ 1

0h�0, y� dy.

T 2 N1.

(c) T�f dx C g dy� D∫ 1

0g�x, 0� dx.

∂T�h� D∫ 1

0

∂h

∂y�x, 0� dx.

T 2 E1.

(See example following Theorem 4.7.)


(d) T�f dx C g dy� D f�a�.

∂T�h� D ∂h

∂x�a�.

T 2 E1.

(See Theorem 4.7.)

(e) T�f dx C g dy� D∫

unit discf�x, y� dx dy.

∂T�h� D∫ 1

0h�√

1� y2, y� dy �∫ 1

0h��

√1� y2, y� dy.

T 2 N1.

4.17. H1�E� D 1�

1∑nD1

2n�1 Ð 4�n D 1/2.

Let T be the sum of all oriented intervals removed in defining E. Clearly T 2R1 � I1. Therefore H

1 E ^ i D [0, 1]� T 2 R 1 � I1.

4.18. ∂T D �∂T� fu > rg C �∂T� fu � rgD ∂[T fu > rg C T fu � rg].

4.19. Immediate from definition of hT, u, rCi.

4.20. MhT, u, rCi <1

for almost all r by (4).

M�∂hT, u, rCi� D Mh∂T, u, rCi(by (2))

<1

for almost all r by (4).4.21. Choose currents A and B such that T D AC ∂B and F�T� D M�A�CM�B�. Since

∂A D ∂T, A 2 N. Since ∂B D T� A, B 2 N. Now

T fu � rg D A fu � rg C ∂[B fu � rg]� hB, u, rCi

by 4.11(1). Therefore,

F�T fu � rg� � M�A�CM�B�CMhB, u, rCi.

Integration and 4.11(4) yield (6) as desired.


4.22. First consider the case M�T� <1. Given ε > 0, choose ϕ 2 Dm with jjϕ�x�jjŁ �

1 such that M�T� � T�ϕ�C ε. Then

M�T� � T�ϕ�C ε D limTi�ϕ�C ε

� lim inf M�Ti�C ε.

Second, if M�T� D 1, given ε > 0, choose ϕ 2 Dm with jjϕ�x�jjŁ � 1 such that

T�ϕ� > 1/ε. Then

lim inf M�Ti� ½ lim Ti�ϕ� > 1/ε.

Hence, lim inf M�Ti� D 1, as desired.4.23. We prove b, of which a is a special case.

f#S�ϕ� D S�f#ϕ� D∫

EhES,f#ϕildH

m

D∫

Eh^m�Df�x��ES�, ϕ�f�x��il�x�dH

mx

D∫

E

⟨^m�Df�x��ES�j ^m �Df�x��ES�j , ϕ�f�x��

⟩l�x�apJm�fjE�dH

mx

(where the contribution from points at which jm�Df�x��ES�jD ap Jm�fjE� D 0 is still interpreted to be 0)

D∫

f�E�

∑yDf�x�

⟨�^mDf�x��ES�j�^mDf�x��ES�j , ϕ

⟩l�x� d H

my

by the coarea formula, 3.13. Therefore,

f#S D �Hm f�E�� ^

∑yDf�x�

l�x��^mDf�x��ES�j�^mDf�x��ES�j .

Chapter 5

5.1. Applying a homothety 7r�x� D rx multiplies M�S� by rmC1, M�T� by rm, andhence both sides of the inequality by the same factor rm.

5.2. Immediately from the definitions, ImC1 ² fT 2 RmC1: M�∂T� <1g. The oppositeinclusion follows from 5.4(1) because ImC1 is F dense in FmC1 (Exercise 4.10).Also from the definitions, Rm ² fT 2 Fm: M�T� <1g. Conversely, suppose T 2Fm with M�T� <1. Then T D RC ∂S, with R 2 Rm, S 2 RmC1. Since M�∂S� DM�T� R� <1, it follows from (2) that S 2 ImC1. Therefore, T D RC ∂S 2 Rm,as desired.


5.3. One good candidate is the sequence

Tk D2k∑

jD1

[(j� 1

2

)2�k, j2�k

]2 I1R1,

which at first glance appears to converge to 12 [0, 1] 62 I1R1.

ž ! ! ! ! ! ! ! ! ž0 T3 1

5.4. Let T be the unit circle R2 with multiplicity 1/N. The only normal current S withS D ∂T is the unit disc with multiplicity 1/N. Since M�S� D /N and M�T� D2/N, the isoperimetric inequality does not hold for any constant 8 .

Chapter 66.1. Just plug f�y, z� D y tan z into the minimal surface equation, 6.1.6.2. Just plug f�x, y� D ln�cos x/ cos y� into the minimal surface equation, 6.1.6.3. Apply the minimal surface equation to z D f�x, y�. Let w D uC iv. Then

fx D ∂f/∂u

∂x/∂uC ∂f/∂v

∂x/∂v,

etc.6.4. A surface of revolution has an equation of the form r D g�z�, where r D√

x2 C y2. Differentiating g2 D x2 C y2 implicitly yields

gg0zx D x, gg0zy D y,

�g02 C gg00�z2x C gg0zxx D 1,

�g02 C gg00�z2y C gg0zyy D 1,

�g02 C gg00�zxzy C gg0zxy D 0.

Applying the minimal surface equation to z�x, y� yields

0 D [�1C z2y�zxx C �1C z2

x �zyy � 2zxzyzxy]gg0

D �1� z2y��1� �g02 C gg00�z2

x �

C �1C z2x ��1� �g02 C gg00�z2

y�

C 2zxzy�g02 C gg00�zxzy

D 2C �z2x C z2

y��1� g02 � gg00�

D �z2x C z2

y��1C g02 � gg00�.


Therefore gg00 D 1C g02. Substituting p D g0 yields gp�dp/dg� D 1C p2. Inte-gration yields p2 D ag2 � 1, i.e.,

dg√a2g2 � 1

D š dz.

Integration yields �1/a� cosh�1 ag D š z C c, i.e., r D g�z� D �1/a�ðcosh�š az C c� D �1/a� cosh�az Ý c�, which is congruent tor D �1/a� cosh az.

6.5. 0 D divrf

1C jrfj2

D ∂

∂x[fx�1C f2

x C f2y��1/2]C ∂

∂y[fy�1C f2

x C f2y��1/2]

D fxx�1C f2x C f2

y��1/2 � fx�1C f2

x C f2y��3/2�fxfxx C fyfxy�

C fyy�1C f2x C f2

y��1/2 � fy�1C f2

x C f2y��3/2�fxfxy C fyfyy�.

0 D �fxx C fyy��1C f2x C f2

y�� f2xfxx � 2fxfyfxy � f2

yfyy

D �1C f2y�fxx � 2fxfyfxy C �1C f2

x�fyy.

6.6. f�x, y� D �x2 � y2, 2xy�.

�1C jfyj2�fxx � 2�fx Ð fy�fxy C �1C jfxj2�fyy

D �2C 8jzj2, 0�� 0

C ��2� 8jzj2, 0�

D 0.

6.7. g�z� D �u�x, y�, v�x, y��, satisfying the Cauchy-Riemann equations ux D vy, uy D�vx , and hence uyy D �uxx and vyy D �vxx .

�1C jfy j2�fxx � 2�fx Ð fy�fxy C �1C jfxj2�fyy

D �1C u2y C v2

y��uxx, vxx�� 2�uxuy C vxvy��uxy, vxy�

C �1C u2x C v2

x��uyy, vyy�

D �1C u2x C u2

y��uxx, vxx�� 0C �1C u2x C u2

y��uxx,�vxx�

D 0.

Chapter 88.1. To contradict 8.1, one might try putting a half-twist in the tail of Figure 8.1(1), so

that the surface, like 8.1(2), would not be orientable. However, there is another


surface like 8.1(2), in which the little hole goes in underneath in front and comesout on top in back. To contradict 8.4, one might try a large horizontal disc centeredat the origin and a little vertical disc tangent to it at the origin:

However, the area-minimizing surface is

8.2.

Front view

Top view

(There are two sheets in the middle.)8.3. Let E be a unit element of surface area in R4 in a plane with orthonormal basis

u, v. Let P,Q denote projection onto the x1-x2 and x3-x4 planes.


Then

area PEC area QE D jPu ^ Pvj C jQu ^ Qvj� �jPujjPvj C jQujjQvj�� jPuj2 C jQuj2�1/2�jPvj2 C jQvj2�1/2

D jujjvj D 1 D area E.

Therefore the area of any surface is at least the sum of the areas of its projections.Now let S be a surface with the same boundary as the two discs D1 C D2. Since

∂�P#S� D ∂D1 and P#S and D1 both lie in the x1-x2 plane, P#S D D1. SimilarlyQ#S D D2. Therefore

area S ½ area P#SC area Q#S

D area D1 C area D2

D area �D1 C D2�.

Therefore D1 C D2 is area minimizing.

Chapter 99.1. ∂R, where R is the pictured rectifiable current of infinitely many components Cj

of length 2�j.

C1

C3

C2

9.2. Suppose S is not area minimizing. For some a, r > 0 there is a rectifiable currentT such that ∂T D ∂�S B�a, r�� and ε D M�S B�a, r��M�T� > 0. Choose j


such thatspt�Sj � S� �AC ∂B�� \ U�a, r C 1� D ∅,

A 2 Rm, B 2 RmC1, and M�A�CM�B� < ε. Let u�x� D jx � aj, and apply slicingtheory, 4.11(4), to choose r < s < r C 1 such that MhB, u, sCi � M�B�. NowSj B�a, s� has the same boundary as

TC S �B�a, s�� B�a, r��C A B�a, s�� hB, u, sCi

and more mass. Therefore Sj is not area minimizing.9.3. Suppose p 2 �spt S� \ f√x2 C y2 < R� 2

pRg. Then the distance from p to

spt ∂S exceeds 2p

R. By monotonicity,

M�S� > �2p

R�2 D 4R D area cylinder,

which contradicts S area minimizing.9.4. False:

a

9.5. (a) One example is two unit orthogonal (complex) discs in R4, where the densityjumps up to 2 at the origin (cf. 6.3 or Exercise 8.3).

(b) Suppose xi ! x but f�x� < limf�xi�. Choose 0 < r0 < dist �x, spt ∂T�such that m�T, x, r0� < limf�xi�.

Choose 0 < r1 < r0 such that

rm0 m�T, x, r0� < rm

i limf�xi�.

Choose i such that jxi � xj < r0 � r1 and

rm0 m�T, x, r0� < rm

1 m�T, xi�.

By monotonicity,rm

0 m�T, x, r0� < rm1 m�T, xi, r1�.


But since B�x, r0� ¦ B�xi, r1�,

rm0 m�T, x, r0� ½ rm

1 m�T, xi, r1�.

This contradiction proves that f is upper semicontinuous.9.6. Suppose x 2 S�0,1�� Tan �spt T, 0�. Then for some ε > 0, for all sufficiently

large r,spt 7r#T \ B�x, ε� D ∅.

Consequently x 62 spt C.Figure 9.7.2 pictures an example in which spt C 6D Tan�spt T, 0�.

9.7. C� D D lim7rj/sj#[C� 7r�1j #T] D 0.

9.8. (a) Let Tn be the homothetic expansion of T by 2n2�n (n odd). Tn consists ofthe interval [2�n, 2n�1] on the x-axis, stuff outside B�0,2n�2�, and stuff insideB�0,2�n� of total mass less than 2�nC2. As a limit of the Tn, the nonnegativex-axis is an oriented tangent cone. Similarly taking n even yields the y-axis.

(b) Let Sn be the homothetic expansion of T by 2n2(n odd). Sn consists of

the interval [1, 22n�1] on the x-axis, the segment from �0, 1� to �1, 0�, theinterval [2�2n�1, 1] on the y-axis, stuff outside B�0, 22n�2�, and stuff insideB�0,2�2n�1� of total mass less than 2�2nC1. Sn converges to the interval [1,1)on the x-axis, plus the segment from �0, 1� to �1, 0�, plus the interval [0, 1]on the y-axis. This limit is not a cone.

(c) Looking at balls of radius 2�n2(n odd) exhibits the lower density 0. Looking

at balls of radius 2�n2(n even) exhibits the upper density 1/2. (Of course, for

a subset of the nonnegative x-axis, the densities must lie between 0 and 1/2.)

Chapter 1010.1.

or see Figure 6.1.4.


10.2. False. [��1, 0, 0�, �1, 0, 0�]C [�0,�1, 100�, �0, 1, 100�] 2 R 1R2 ð R1 is areaminimizing, but its projection, [��1, 0�, �1, 0�]C [�0,�1�, �0, 1�], is not.

10.3.

Chapter 11

11.1. Four similarly oriented, parallel circles

bound a catenoid and a horizontal annulus with cross-sections.

Catonoid

Origin

Annulus

11.2. Because the tangent plane is constant, v�S��Rn ðG2Rn� D area S D .11.3. Two close parallel unit discs, which can be deformed to a surface like that in

Figure 11.1.3. Any unstable minimal surface, as in Figure 6.1.2, which can bedeformed to a surface in Figure 6.1.4.

11.4. The surface of Figure 11.1.3. The catenoid has less area.


Chapter 1515.1. By the two facts, the perimeter P satisfies

P ½ �n/2��2p

�C �1/2��2p

n�.

The perimeter to area ratio for the regular hexagonal tiling is easily computedas 121/4. The problem is finished by verifying the algebraic inequality that forn � 398,

P/n ½ �1C 1/p

n�p

½ 121/4.

Bibliography

Section numbers follow the entries, indicating the locations in this book where thepublications are cited.

Allard, William K. On the first variation of a varifold. Ann. Math. 95 (1972), 417–491.(§§5.3, 9.3, 10.2, 11.2, 17.2)

Almgren, F. J., Jr.[1] Existence and regularity almost everywhere of solutions to elliptic variational

problems with constraints. Mem. AMS No. 165 (1976). (§§8.6, 11.3, 13.3, 13.4,13.8, 13.10)

[2] Optimal isoperimetric inequalities. Indiana Univ. Math. J. 35 (1986), 451–547.(§§5.3, 17.1)

[3] Q-valued functions minimizing Dirichlet’s integral and the regularity of areaminimizing rectifiable currents up to codimension two. Bull. AMS 8 (1983),327–328. (§8.3)

[4] Questions and answers on area minimizing surfaces and geometric measuretheory. Proc. Symp. Pure Math. 54 (1993), Part 1, 29–53. (Preface)

[5] Review of “Geometric Measure Theory: A Beginner’s Guide.” Am. Math.Monthly 96 (1989), 753–756. (Preface)

[6] Some interior regularity theorems for minimal surfaces and an extension ofBernstein’s theorem. Ann. Math. 84 (1966), 277–292. (§8.1)

Almgren, Fred, Kusner, Robert B. and Sullivan, John M. A comparison of the Kelvinand Weaire-Phelan foams, in preparation. (§15.10)

Almgren, F. J., Jr., and Taylor, J. E. Geometry of soap films. Sci. Am. 235 (1976),82–93. (§§13.8, 13.9)

Almgren, Fred, Taylor, Jean E., and Wang, Lihe. Curvature-driven flows: a variationalapproach. Siam J. Control and Optimization 31 (1993), 387–437. (§16.7)

203


Ambrosio, Luigi and Braides, Andrea. Functionals defined on partitions in sets offinite perimeter II: semicontinuity, relaxation and homogenization. J. Math. PuresAppl. 69 (1990), 307–333. (§16.4)

Bates, S. M. Toward a precise smoothness hypothesis in Sard’s Theorem. Proc. AMS117 (1993), 279–283. (§12.1)

Berger, M.[1] “Geometry,” Vol. II. Springer-Verlag, New York, 1972. (§13.2)[2] Quelques problemes de geometrie Riemannienne ou Deux variations sur les

espaces symetriques compacts de rang un. Enseignement Math. 16 (1970),73–96. (§6.5)

Besicovitch, A. S. A general form of the covering principle and relative differentiationof additive functions I, II. Proc. Cambridge Phil. Soc. 41 (1945), 103–110; 42(1946), 1–10. (§2.7)

Bombieri, E., De Giorgi, E., and Giusti, E. Minimal cones and the Bernstein problem.Invent. Math. 7 (1969), 243–268. (§§8.1, 10.7)

Boys, C. V. “Soap-Bubbles: their Colors and the Forces which Mold them.” Dover,New York, 1959. (§14.0)

Brakke, Kenneth A.[1] Minimal cones on hypercubes. J. Geom. Anal. 1 (1991), 329–338. (§6.5)[2] Polyhedral minimal cones in R4. Preprint (1993). (§§6.5, 13.10)[3] Soap films and covering spaces. J. Geom. Anal. 5 (1995), 445–514. (§6.5)[4] The surface evolver. Exp. Math. 1 (1992), 141–165. (§§13.0, 13.12, 13.13)[5] Century-old soap bubble problem solved! Imagine That! 3 (Fall, 1993), The

Geometry Center, University of Minnesota, 1–3. (§13.13)Bray, Hubert. Proof of the Riemannian Penrose conjecture using the positive mass

theorem, preprint (1999). (Pref., §16.7)Bray, Hubert and Morgan, Frank. An isoperimetric comparison theorem for

Schwarzschild space and other manifolds, preprint (1999). (§13.2)Brothers, J. E. (Ed.). Some open problems in geometric measure theory and its

applications suggested by participants of the 1984 AMS Summer Institute. In“Geometric Measure Theory and the Calculus of Variations” (W. K. Allard andF. J. Almgren, Jr., Eds.). Proc. Symp. Pure Math. 44 (1986), 441–464. (Preface)

Brothers, John E., and Morgan, Frank. The isoperimetric theorem for general inte-grands. Michigan Math. J. 41 (1994), 419–431. (§13.2)

Burago, Yu. D., and Zalgaller, V. A. “Geometric Inequalities.” Springer-Verlag, NewYork, 1988. (§§2.8, 13.2)

Chang, Sheldon. Two dimensional area minimizing integral currents are classicalminimal surfaces. J. AMS 1 (1988), 699–778. (§8.3)

Cipra, Barry. Why double bubbles form the way they do. Science 287 (17 March2000), 1910–1911. (§14.0)

Conway, John H., and Sloane, Neil. The cell structures of certain lattices. In “Miscel-lanea Mathematica,” Peter Hilton, Friedrich Hirzebruch, Reinhold Remmert, ed.,Springer-Verlag, New York, 1993. (§15.11)

Bibliography 205

Costa, C. “Imersoes Minimas Completas em R3 de GOenero um e Curvatura TotalFinita.” Doctoral thesis. IMPA, Rio de Janeiro, Brasil, 1982; Example of acomplete minimal immersion in R3 of genus one and three embedded ends.Bull. Soc. Bras. Mat. 15 (1984), 47–54. (Figure 6.1.3a)

Courant, R. and Hilbert, D. “Methods of Mathematical Physics,” Vol. I, English ed.,Wiley, New York, 1953. (§14.14)

Cox, Christopher, Harrison, Lisa, Hutchings, Michael, Kim, Susan, Light, Janette,Mauer, Andrew, and Tilton, Meg.

[1] The shortest enclosure of three connected areas in R2. Real Anal. Exch. 20(1994/1995), 313–335. (§§13.1, 14.23)

[2] The standard triple bubble type is the least-perimeter way to enclose threeconnected areas. NSF “SMALL” undergraduate research Geometry Groupreport. Williams College, 1992. (§14.8)

Coxeter, H. S. M. “Regular polytopes.” Dover, New York, 1973. (§15.11)Croke, Christopher B. Some isoperimetric inequalities and eigenvalue estimates. Ann.

Scient. Ec. Norm. Sup. 13 (1980), 419–435. (§17.3)Dao Trong Thi. Minimal real currents on compact Riemannian manifolds. Izv. Akad.

Nauk. SSSR Ser. Mat. 41 (1977) [English translation in Math. USSR Izv. 11(1977), 807–820]. (§6.5)

De Giorgi, E.[1] “Frontiere Orientate di Misura Minima” (Sem. Mat. Scuola Norm. Sup. Pisa,

1960–1961). Editrice tecnico scientifica, Pisa, 1961. (§§1.0, 4.0)[2] Sulla proprieta isoperimetrica dell’ipersfera, nella classe degli insiemi aventi

frontiera orientata di misura finita. Mem. Acc. Naz. Lincei, Ser. 8, 5 (1958),33–44. (§13.2)

[3] Su una teoria generale della misura r � 1 dimensionale in un spazio ad r dimen-sioni. Ann. Mat. 4 (1955), 95–113. (§§1.0, 4.0)

de Rham, Georges.[1] “On the Area of Complex Manifolds. Notes for the Seminar on Several Complex

Variables.” Institute for Advanced Study, Princeton, NJ, 1957–1958. (§6.5)[2] Varietes differentiables, formes, courants, formes harmoniques. Act. Sci. Indust.

1222 (1955). (§4.0)Douglas, Jesse. Solution of the problem of plateau. Trans. AMS 33 (1931), 263–321.

(§1.2)Falconer, K. J. “The Geometry of Fractal Sets.” Cambridge University Press, 1985.

(§3.17)Federer, Herbert.

[1] “Geometric Measure Theory.” Springer-Verlag, New York, 1969. (citedthroughout)

[2] The singular sets of area minimizing rectifiable currents with codimension oneand of area minimizing flat chains modulo two with arbitrary codimension. Bull.AMS 76 (1970), 767–771. (§§8.2, 11.1)

[3] Some theorems on integral currents. Trans. AMS 117 (1965), 43–67. (§6.5)


Federer, Herbert, and Fleming, Wendell H. Normal and integral currents. Ann. Math.72 (1960), 458–520. (§§1.0, 4.0)

Fejes Toth, G. An isoperimetric problem for tessellations. Studia Sci. Math. Hungarica10 (1975), 171–173. (§15.9)

Fejes Toth, L.[1] An arrangement of two-dimensional cells. Ann. Univ. Sci. Budapest., Sect. Math.

2 (1959), 61–64. (§15.1)[2] “Lagerungen in der Ebene auf der Kugel und im Raum,” Die Grundlehren der

Math. Wiss., Vol. 65. Springer-Verlag, Berlin, 1953, 1972. (§15.1)[3] “Regular Figures,” International Series of Monographs on Pure and Applied

Mathematics Vol. 48. Macmillan Co., New York, 1964. (§15.1)[4] Uber das kurzeste Kurvennetz das eine Kugeloberflache in Flachengleiche

konvexe Teile zerlegt. Mat. Term. Ertesito 62 (1943), 349–354. (§15.1)[5] What the bees know and what they do not know. Bull. AMS 70 (1964), 468–481.

(§15.8)Feynman, Richard P. The internal geometry of crystals, “The Feynman Lectures on

Physics,” Vol. II. Addison-Wesley, Reading, MA, 1963–65, Chapter 30. (§16.4)Fleming, Wendell H.

[1] Flat chains over a finite coefficient group. Trans. AMS 121 (1966), 160–186. (§16.2)[2] On the oriented Plateau problem. Rend. Circ. Mat. Palermo (2)11, (1962), 1–22.

(§8.1)Foisy, Joel. “Soap Bubble Clusters in R2 and R3.” Undergraduate thesis. Williams

College, Williamstown, MA, 1991. (§§13.0, 14.0, 14.3, 14.9)Foisy, Joel, Alfaro, Manuel, Brock, Jeffrey, Hodges, Nickelous, and Zimba, Jason.

The standard double soap bubble in R2 uniquely minimizes perimeter. Pac. J.Math. 159 (1993), 47–59. Featured in the 1994 AMS “What’s Happening in theMathematical Sciences.” (Pref., §§13.1, 14.0, 14.3, 14.10)

Fomenko, A. T. “The Plateau Problem. I: Historical Survey. II: Present State of theTheory.” Gordon & Breach, New York, 1990. (Preface)

Francis, George, Sullivan, John M., Kusner, Rob B., Brakke, Ken A., Hartman,Chris, and Chappell, Glenn, The minimax sphere eversion, in “Visualizationand Mathematics,” Springer, 1997, 3–20. (§16.7)

Futer, David, Gnepp, Andrei, McMath, David, Munson, Brian A., Ng, Ting, Pahk,Sang-Hyoun, and Yoder, Cara. Cost-minimizing networks among immisciblefluids in R2. Pac. J. Math., to appear. (§16.3)

Giusti, Enrico. “Minimal Surfaces and Functions of Bounded Variation.” Birkhauser,Boston, 1984. (Pref., §§4.0, 12.4)

Gluck, Herman, Mackenzie, Dana, and Morgan, Frank. Volume-minimizing cycles inGrassmann manifolds. Mackenzie, Morgan. Duke Math. J. 79 (1995), 335–404.(§6.5)

Gonick, Larry. Fill’er up. Discover Magazine, August, 1994, 80–81. (§15.10)Gonzalez, E., Massari, U., and Tamanini, I. On the regularity of boundaries of sets

minimizing perimeter with a volume constraint. Indiana Univ. Math. J. 32 (1983),25–37. (§8.6)

Bibliography 207

Graner, F., Jiang, Y., Janiaud, E., and Flament, C. Equilibrium energies of 2D fluidfoams, preprint (2000). (§13.12)

Grayson, Matthew A. Shortening embedded curves. Ann. of Math. 129 (1989),71–111. (§16.7)

Gromov, M. Isoperimetric inequalities in Riemannian manifolds, Appendix I to“Asymptotic Theory of Finite Dimensional Normed Spaces” by Vitali D. Milmanand Gideon Schechtman, Lecture Notes in Mathematics, No. 1200. Springer-Verlag, New York, 1986. (§13.2)

Grunbaum, Branko and Shephard, G. C. “Tilings and Patterns.” Freeman and Co,New York, 1987. (Figure 15.9.1)

Gruter, Michael.[1] Free boundaries in geometric measure theory and applications. In “Variational

Methods for Free Surface Interfaces” (P. Concus and R. Finn, Eds.). Springer-Verlag, New York, 1986. (§12.3)

[2] Optimal regularity for codimension one minimal surfaces with a free boundary.Manuscript Math. 58 (1987), 295–343. (§12.3)

Hales, Thomas C. The honeycomb conjecture. Disc. Comp. Geom., to appear. (Pref.,Chapter 15)

Hardt, Robert, and Simon, Leon.[1] Boundary regularity and embedded solutions for the oriented Plateau problem.

Ann. Math. 110 (1979), 439–486. (§8.4)[2] “Seminar on Geometric Measure Theory.” Birkhauser, Boston, 1986. (Pref.,

§3.11)Harrison, Jenny. Stokes’ theorem for nonsmooth chains. Bull. AMS 29 (1993),

235–242. (§12.2)Harvey, Reese, and Lawson, H. Blaine, Jr. Calibrated geometries. Acta Math. 148

(1982), 47–157. (§6.5)Hass, Joel, Hutchings, Michael, and Schlafly, Roger. The double bubble conjecture,

Elec. Res. Ann. AMS 1 (1995), pp. 98–102. (Pref., §14.0)Hass, Joel and Schlafly, Roger.

[1] Bubbles and double bubbles. American Scientist, Sept-Oct, 1996, 462–467.(§14.0)

[2] Double bubbles minimize. Annals of Math. 151 (2000), 459–515. (§14.0)Heath, Sir Thomas. “A History of Greek Mathematics,” Vol. II. Oxford University

Press, London, 1921. (§15.1).Heilmann, Cory, Lai, Yuan Y., Reichardt, Ben W., and Spielman, Anita.

[1] Component bounds for area-minimizing double bubbles, NSF “SMALL” under-graduate research Geometry Group report, Williams College, 1999. (Chapter 14)

[2] Proof of the double bubble conjecture in R4 and certain higher dimensionalcases, NSF “SMALL” undergraduate research Geometry Group report, WilliamsCollege, 1999.

Heppes, A.[1] Isogonale spharischen Netze. Ann. Univ. Sci. Budapest Eotvos Sect. Math. 7

(1964), 41–48. (§13.9)


[2] On surface-minimizing polyhedral decompositions. Disc. Comp. Geom. 13(1995), 529–539. (§13.9)

Hildebrandt, Stefan. Free boundary problems for minimal surfaces and related ques-tions. Comm. Pure Appl. Math. 39 (1986), S111–S138. (§12.3)

Hildebrandt, Stefan, and Tromba, Anthony. “The Parsimonious Universe.” Coper-nicus, Springer-Verlag, New York, 1996. (Pref., Figure 16.4.1)

Hoffman, David. The computer-aided discovery of new embedded minimal surfaces.Math. Intelligencer 9 (1987), 8–21. (Figure 6.1.3a)

Hoffman, David, and Meeks, W. H., III. A complete embedded minimal surfacein R3 with genus one and three ends. J. Diff. Geom. 21 (1985), 109–127.(Figure 6.1.3a)

Hoffman, David and Spruck, Joel. Sobolev and isoperimetric inequalities for Rieman-nian submanifolds. Comm. Pure Appl. Math. 27 (1974), 715–727. Correction,28 (1975), 765–766. (§17.3)

Hoffman, David, Wei, Fusheng, and Karcher, Hermann. Adding handles to the heli-coid. Bull. AMS 29 (1993), 77–84. (Pref., Figure 6.1.3b)

Hopf, E. Elementare Bemerkungen uber die Losungen partieller Differential-gleichungen zweiter Ordnung vom elliptischen Typus, Sitzungberichte derPreussischen Akademie der Wissenshaften zu Berlin. Phys.-Math. Klasse 19(1927), 147–152. (§§8.5, 10.4)

Howards, Hugh, Hutchings, Michael, and Morgan, Frank. The isoperimetric problemon surfaces. Am. Math. Monthly 106 (1999), 430–439. (§13.2)

Hsiang, Wu-Yi. Talk at First MSJ International Research Institute, Sendai, 1993. (§13.2)Hsiang, W.-T. and Hsiang, W.-Y. On the uniqueness of isoperimetric solutions and

imbedded soap bubbles in noncompact symmetric spaces. Inv. Math. 85 (1989),39–58. (§13.2)

Huisken, Gerhard. Flow by mean curvature of convex surfaces into spheres. J. Diff.Geom. 20 (1984), 237–266. (§16.7)

Hutchings, Michael. The structure of area-minimizing double bubbles. J. Geom. Anal.7 (1997), 285–304. (§§13.0, 13.2, Chapter 14)

Hutchings, Michael, Morgan, Frank, Ritore, Manuel, and Ros, Antonio. Proof of theDouble Bubble Conjecture, preprint (2000). (Pref., Chapter 14)

Ilmanen, T. A strong maximum principle for singular minimal hypersurfaces. Calc.Var. Partial Diff. Eqns. 4 (1996), 443–467. (§10.4)

Isenberg, Cyril. “The Science of Soap Films and Soap Bubbles.” Dover, NY, 1992.(§13.9)

Kanigel, Robert. Bubble, bubble: Jean Taylor and the mathematics of minimal surfaces.The Sciences (May/June 1993), 32. (§13.8)

Kinderleharer, D., Nirenberg, L., and Spruck, J. Regularity in elliptic free boundaryproblems, I. J. Anal. Math. 34 (1978), 86–119. (§13.9)

Klarreich, Erica G. Foams and honeycombs. Amer. Scientist 88 (March–April 2000),152–161. (§§15.0, 15.10)

Kleiner, Bruce. An isoperimetric comparison theorem. Invent. Math. 108 (1992),37–47. (§17.3)

Bibliography 209

Knorr, Wilbur Richard. “The Ancient Tradition of Geometric Problems.” Birkhauser,Boston (§13.2)

Korevaar, Nicholas J., Kusner, Rob, and Solomon, Bruce. The structure of completeembedded surfaces with constant mean curvature. J. Differential Geom. 30(1989), no. 2, 465–503. (§14.7)

Kusner, Robert B. and Sullivan, John M.[1] Comparing the Weaire-Phelan equal-volume foam to Kelvin’s foam. Forma 11

(1996), 233–242. Reprinted in Weaire [2]. (§15.10)[2] Mobius energies for knots and links, surfaces and submanifolds. AMS/IP Studies

in Adv. Math. 2 (Part 1, 1997), 570–604. (§16.6)Lamarle, Ernest. Sur la stabilite des systemes liquides en lames minces. Mem. Acad.

R. Belg. 35 (1864), 3–104. (§13.9)Lawlor, Gary

[1] The angle criterion. Invent. Math. 95 (1989), 437–446. (§6.5)[2] Proving area minimization by directed slicing. Indiana U. Math. J. 47 (1998),

1547–1592. (§6.5)[3] A sufficient condition for a cone to be area-minimizing. Mem. AMS 91 (446)

(1991). (§6.5)Lawlor, Gary, and Morgan, Frank.

[1] Curvy slicing proves that triple junctions locally minimize area. J. Diff. Geom.44 (1996), 514–528. (§§6.5, 13.9)

[2] Paired calibrations applied to soap films, immiscible fluids, and surfaces ornetworks minimizing other norms. Pac. J Math. 166 (1994), 55–83. (§6.5)

Lawson, H. Blaine, Jr. “Lectures on Minimal Submanifolds,” Vol. 1. Publish or Perish,1980. (Pref., §§1.2, 8.1)

Lawson, H. Blaine, Jr., and Osserman, Robert. Non-existence, non-uniqueness andirregularity of solutions to the minimal surface system. Acta Math. 139 (1977),1–17. (§6.2)

Le Hong Van. Relative calibrations and the problem of stability of minimal surfaces.In “Lecture Notes in Mathematics,” No. 1453, pp. 245–262. Springer-Verlag,New York, 1990. (§6.5)

Leonardi, G. Infiltrations in immiscible fluids systems, Proc. Royal Soc. Edinburgh,to appear. (§16.3)

Mackenzie, D. See Nance.Mandelbrot, Benoit B.

[1] “The Fractal Geometry of Nature.” Freeman, New York, 1983. (§2.3)[2] “Fractals.” Freeman, San Francisco, 1977. (§2.3)

Masters, Joseph D. The perimeter-minimizing enclosure of two areas in S2. Real Anal.Exch. 22 (1996/7), 645–654. (§14.23)

Montesinos Amilibia, A. Existence and uniqueness of standard bubbles of givenvolumes in Rn. Asian J. Math., to appear. (§14.2)

Morgan, Frank[1] Area-minimizing surfaces, faces of Grassmannians, and calibrations. Am. Math.

Monthly 95 (1988), 813–822. (§§4.1, 6.5)


[2] Calibrations and new singularities in area-minimizing surfaces: a surveyIn “Variational Methods” (Proc. Conf. Paris, June 1988), (H. BerestyckiJ.-M. Coron, and I. Ekeland, Eds.). Prog. Nonlinear Diff. Eqns. Applns 4,329–342. Birkhauser, Boston, 1990. (§6.5)

[3] Clusters minimizing area plus length of singular curves. Math. Ann. 299 (1994),697–714. (§§13.0, 13.3, 13.11)

[4] “Compound Soap Bubbles, Shortest Networks, and Minimal Surfaces.” AMSvideo, 1993. (§13.0)

[5] The Double Bubble Conjecture, FOCUS, Math. Assn. Amer., December, 1995.(§14.0)

[6] On finiteness of the number of stable minimal hypersurfaces with a fixedboundary. Indiana Univ. Math. J. 35 (1986), 779–833. (§6.1)

[7] Generic uniqueness results for hypersurfaces minimizing the integral of anelliptic integrand with constant coefficients. Indiana Univ. Math. J. 30 (1981),29–45. (§8.5)

[8] The hexagonal honeycomb conjecture. Trans. AMS 351 (1999), 1753–1763.(§15.3)

[9] Immiscible fluid clusters in R2 and R3. Mich. Math. J. 45 (1998), 441–450.(§§16.2, 16.3)

[10] Lower-semicontinuity of energy of clusters. Proc. Royal Soc. Edinburgh 127A(1997), 819–822. (§16.4)

[11] �M, ε, υ-minimal curve regularity. Proc. AMS 120 (1994), 677–686. (§§11.3,13.10)

[12] Mathematicians, including undergraduates, look at soap bubbles. Am. Math.Monthly 101 (1994), 343–351. (§13.0)

[13] Minimal surfaces, crystals, shortest networks, and undergraduate researchMath. Intelligencer 14 (Summer, 1992), 37–44. (§13.0)

[14] Review of “Mathematics and Optimal Form” by S. Hildebrandt andA. Tromba. Am. Math. Monthly 95 (1988), 569–575. (Preface)

[15] Review of “The Parsimonious Universe” (Hildebrandt/Tromba). Amer. Math.Monthly 104 (April, 1997), 377–380. (Preface)

[16] “Riemannian Geometry: A Beginner’s Guide.” A. K. Peters, Natick, 1998.(§13.2)

[17] Size-minimizing rectifiable currents. Invent. Math. 96 (1989), 333–348. (§11.3)[18] Soap bubbles and soap films. In “Mathematical Vistas: New and Recent

Publications in Mathematics from the New York Academy of Sciences”(J. Malkevitch and D. McCarthy, Eds.), Vol. 607. New York Academy ofSciences, New York, 1990. (§6.1)

[19] Soap bubbles in R2 and in surfaces. Pac. J. Math. 165 (1994), 347–361.(§§13.1, 13.4, 13.10)

[20] Soap films and mathematics. Proc. Symp. Pure Math. 54 (1993), Part 1,375–380. (§11.3)

[21] Soap films and problems without unique solutions. Am. Sci. 74 (1986),232–236. (§6.1)

Bibliography 211

[22] Surfaces minimizing area plus length of singular curves. Proc. AMS 122(1994), 1153–1164. (§13.11)

[23] Survey lectures on geometric measure theory. In “Geometry and Global Ana-lysis: Report of the First MSJ International Research Institute, July 12–13,1993,” (Takeshi Kotake, Seiki Nishikawa, and Richard Schoen, Eds.), 87–110.Tohoku Univ., Math. Inst., Sendai, Japan, 1993. (§13.9)

[24] What is a surface? Amer. Math. Monthly 103 (May, 1996), 369–376. (Pref.,§1.2, Figure 4.3.3)

Morgan, Frank, French, Christopher, and Greenleaf, Scott. Wulff clusters in R2. J.Geom. Anal. 8 (1998), 97–115. (§16.5)

Morgan, Frank and Johnson, David L. Some sharp isoperimetric theorems forRiemannian manifolds, preprint (1999). (§§13.2, 17.3)

Morgan, Frank and Ritore, Manuel. Isoperimetric regions in cones, preprint (1999).(§13.2)

Morgan, Frank, and Taylor, Jean. The tetrahedral point junction is excluded iftriple junctions have edge energy. Scr. Metall. Mater. 15 (1991), 1907–1910.(§13.11)

Munkres, James. “Elements of Algebraic Topology.” Addison-Wesley Publishing Co.,Menlo Park, California, 1984. (§14.3)

Murdoch, Timothy A. Twisted calibrations. Trans. AMS 328 (1991), 239–257. (§6.5)Nance [Mackenzie], Dana. Sufficient conditions for a pair of n-planes to be area-

minimizing. Math. Ann. 279 (1987), 161–164. (§6.5)Nitsche, Johannes C. C.

[1] The higher regularity of liquid edges in aggregates of minimal surfaces. Nachr.Akad. Wiss. Gottingen Math.-Phys. Klasse 2 (1977), 75–95. (§13.9)

[2] “Vorlesungen uber Minimalflachen.” Springer-Verlag, New York, 1975 [Transla-tion: “Lectures on Minimal Surfaces.” Cambridge University Press, New York,1989]. (Pref., §§1.2, 8.1, 8.4)

Osserman, Robert.[1] The isoperimetric inequality. Bull. AMS 84 (1978), 1182–1238. (§17.1)[2] “A Survey of Minimal Surfaces.” New Dover edition, New York, 1986. (Pref.,

§§1.2, 8.1)Pauling, Linus. “The Nature of the Chemical Bond and the Structure of Molecules

and Crystals: an Introduction to Modern Structural Chemistry,” third edition.Cornell University Press, Ithaca, 1960. (§15.12)

Pedrosa, Renato H. L. Symmetrization for cylindrical spaces and applications toisoperimetric problems, preprint (2000). (§13.2)

Pedrosa, Renato H. L. and Ritore, Manuel. Isoperimetric domains in the Riemannianproduct of a circle with a simply connected space form. Indiana Univ. Math. J.48 (1999), 1357–1394. (§13.2)

Peterson, Ivars.[1] Constructing a stingy scaffolding for foam. Science News (March 5, 1994).

(§15.0)[2] The honeycomb conjecture. Science News 156 (July 24, 1999), 60–61. (§15.0)


Plateau, J. “Statique Experimentale et Theorique des Liquides Soumis aux SeulesForces Moleculaires.” Gauthier-Villars, Paris, 1873. (Pref., §13.8).

Rado, Tibor. “On the Problem of Plateau.” Springer-Verlag, New York, 1933 [reprinted1971]. (§1.2)

Reichardt, Ben W., Heilmann, Cory, Lai, Yuan Y., and Spielman, Anita. Proof ofthe double bubble conjecture in R4 and certain higher dimensional cases, preprint(2000). (Pref., Chapter 14)

Reifenberg, E. R.[1] Solution of the Plateau problem for m-dimensional surfaces of varying topolog-

ical type. Acta Math. 104 (1960), 1–92. (§§1.0, 4.0, 13.9)[2] An epiperimetric inequality related to the analyticity of minimal surfaces. Ann.

Math. 80 (1964), 1–14. (§§1.0, 4.0, 13.9)[3] On the analyticity of minimal surfaces. Ann. Math. 80 (1964), 15–21. (§§1.0,

4.0, 13.9)[4] A problem on circles. Math. Gazette 32 (1948), 290–292. (§2.7)

Ritore, Manuel. Applications of compactness results for harmonic maps to stableconstant mean curvature surfaces. Math. Z. 226 (1997), 465–481. (§13.2)

Ritore, Manuel, and Ros, Antonio.[1] The spaces of index one minimal surfaces and stable constant mean curvature

surfaces embedded in flat three manifolds. Trans. AMS 258 (1996), 391–410.(§13.2)

[2] Stable constant mean curvature tori and the isoperimetric problem in three spaceforms. Comm. Math. Helv. 67 (1992), 293–305. (§13.2)

Ross, Sydney. Bubbles and foam. Industrial and Engineering Chemistry 61 (1969),48–57. (§13.12)

Schmidt, Erhard, Beweis der isoperimetrischen Eigenschaft der Kugel im hyperbolis-chen und sphrarischen Raum jeder Dimensionenzahl. Math. Z. 49 (1943), 1–109.(§13.2)

Schoen, R., and Simon, L. Regularity of stable minimal hypersurfaces. Comm. PureAppl. Math. 34 (1981), 741–797. (§8.5)

Schoen, R., Simon, L., and Almgren, F. J. Regularity and singularity estimates onhypersurfaces minimizing parametric elliptic variational integrals. Acta Math.139 (1977), 217–265. (§8.5)

Schoen, R., Simon, L., and Yau, S.-T. Curvature estimates for minimal hypersurfaces.Acta Math. 134 (1975), 275–288. (§8.5)

Schoen, R., and Yau, S.-T. On the proof of the positive mass conjecture in GeneralRelativity. Comm. Math. Phys. 65 (1979), 45–76. (Pref., §16.7)

Schwarz, H. A. Beweis des Satzes, dass die Kugel kleinere Oberflache besitzt, alsjeder andere Korper gleichen Volumens. Nachrichten Koniglichen GesellschaftWissenschaften Gottingen (1884), 1–13. (§13.2)

Serrin, J. On the strong maximum principle for quasilinear second order differentialinequalities. J. Funct. Anal. 5 (1970), 184–193. (§8.5)

Bibliography 213

Simon, Leon[1] Cylindrical tangent cones and the singular set of minimal submanifolds. J. Diff.

Geom. 38 (1993), 585–652. (§11.1)[2] Existence of surfaces minimizing the Willmore functional. Comm. Anal. Geom.

1 (1993), 281–326. (§16.6)[3] “Lectures on Geometric Measure Theory.” Proc. Centre Math. Anal. Austral.

Nat. Univ. 3 (1983). (Pref., §§3.11, 4.3, 4.11, 9.1, 12.4)[4] Survey lectures on minimal submanifolds. In “Seminar on Minimal Submani-

folds” (E. Bombieri, Ed.), Ann. of Math. Studies 103, Princeton University Press,Princeton, NJ, 1983. (Preface)

Simons, James. Minimal varieties in Riemannian manifolds. Ann. Math. 88 (1968),62–105. (§§8.1, 10.5)

Solomon, Bruce and White, Brian. A strong maximum principle for varifolds that arestationary with respect to even parametric elliptic functionals. Ind. Univ. Math.J. 38 (1989), 683–691. (§10.4)

Sommerville, D. M. Y. “An Introduction to the Geometry of n Dimensions.” Dutton,New York, 1929. (§4.1)

Stewart, Ian. Circularly covering clathrin. Nature 351 (May 9, 1991), 103. (§15.12)Sullivan, John M.

[1] “The Optiverse” and other sphere eversions. Proc. ISAMA99, San Sabastian,Spain (June, 1999) and Proc. Bridges, Kansas (July, 1999). (§16.7)

[2] Sphere packings give an explicit bound for the Besicovitch covering theorem.J. Geom. Anal. 4 (1994), 219–231. (Figure 2.7.1)

Sullivan, John M., Francis, George, and Levy, Stuart. The Optiverse, in VideoMathFestival at ICM ‘98. Springer, NY, 1998, 7-minute video. (§16.7)

Sullivan, John M. and Morgan, Frank, ed. Open problems in soap bubble geometry.International J. of Math. 7 (1996), 833–842. (Pref., §§13.9, 17.1)

Tarnai, Tibor. The observed form of coated vesicles and a mathematical coveringproblem. J. Mol. Biol. 218 (1991), 485–488. (§15.12)

Taylor, Jean E.[1] Boundary regularity for solutions to various capillarity and free boundary prob-

lems. Com. PDE 2 (1977), 323–357. (§12.3)[2] Crystalline variational problems. Bull. AMS 84 (1987), 568–588. (§§12.5, 16.4)[3] Regularity of the singular sets of two-dimensional area-minimizing flat chains

modulo 3 in R3. Invent. Math. 22 (1973), 119–159. (§11.1)[4] The structure of singularities in soap-bubble-like and soap-film-like minimal

surfaces. Ann. Math. 103 (1976), 489–539. (§§11.3, 13.9)Taylor, J. E., Cahn, J. W., and Handwerker, C. A. Geometric models of crystal

growth. Acta Metall. Mater. 40 (1992), 1443–1474 (Overview No. 98–I).(§16.7)

Thompson, D’Arcy Wentworth. “On Growth and Form,” abridged edition. CambridgeUniversity Press, 1969. (Figure 15.1.1)


Thomson, William (Lord Kelvin). On the homogeneous division of space. Proc. RoyalSoc. London 55 (1894), 1–16. (§15.10)

Varro, Marcus Terentius. “On Agriculture.” The Loeb Classical Library. HarvardUniversity Press, 1934. (§15.1)

Weaire, Denis.[1] Froths, foams and heady geometry. New Scientist, (May 21, 1994). (§15.10)[2] ed., “The Kelvin problem. Foam structures of minimal surface area.” Taylor &

Francis, Ltd., London, 1996.Weaire, Denis and Phelan, Robert. A counter-example to Kelvin’s conjecture on

minimal surfaces. Phil. Mag. Lett. 69 (1994), 107–110. (§§15.0, 15.10)Wecht, Brian, Barber, Megan, and Tice, Jennifer. Double crystals. Acta Cryst., Sect.

A, 56, 92–95. (§16.5)Weyl, Hermann. “Symmetry.” Princeton University Press, 1952, Princeton Sci. Lib.

ed., 1989. (§15.1)White, Brian

[1] Existence of least-area mappings of N-dimensional domains. Ann. Math. 118(1983), 179–185. (§1.2)

[2] Existence of least-energy configurations of immiscible fluids. J. Geom. Anal. 6(1996), 151–161. (§§16.2, 16.3)

[3] A new proof of the compactness theorem for integral currents. Comm. Math.Helv. 64 (1989), 207–220. (§§3.17, 4.12, 5.4)

[4] A new proof of Federer’s structure theorem for k-dimensional subsets of RN. J.AMS 11 (1998), 693–701. (§3.17)

[5] Regularity of area-minimizing hypersurfaces at boundaries with multiplicity.In “Seminar on Minimal Submanifolds” (E. Bombieri, Ed.), pp. 293–301.Princeton University Press, Princeton, NJ, 1983. (§8.4)

[6] Regularity of the singular sets in immiscible fluid interfaces and solutions toother Plateau-type problems. Proc. Centre Math. Anal. Austral. Nat. Univ. 10(1985), 244–249. (§16.3)

[7] A regularity theorem for minimizing hypersurfaces modulo p. Proc. Symp. PureMath. 44 (1986), 413–427. (§§8.5, 11.1)

[8] The structure of minimizing hypersurfaces mod 4. Invent. Math. 53 (1979),45–58. (§11.1)

[9] In preparation. (§13.10)Whitney, Hassler. “Geometric Integration Theory.” Princeton University Press,

Princeton, NJ, 1957. (§4.7)Williams, Robert. “The Geometrical Foundation of Natural Structure: A Source Book

of Design.” Dover, 1979. (§15.12)Willmore, T. J. “Riemannian Geometry.” Clarendon Press, Oxford, 1993. (§16.6)Wirtinger, W. Eine Determinantenidentitat und ihre Anwendung auf analytische

Gebilde und Hermitesche Massbestimmung. Monatsh. Math. Phys. 44 (1936),343–365. (§6.5)

Wong, Yung-Chow. Differential geometry of Grassmann manifolds. Proc. Natl. Acad.Sci. USA 57 (1967), 589–594. (§4.1)

Bibliography 215

Yau, Shing-Tung. Isoperimetric constants and the first eigenvalue of a compactRiemannian manifold. Ann. scient. Ec. Norm. Sup. 8 (1975), 487–507. (§17.3)

Young, L. C.[1] On generalized surfaces of finite topological types. Mem. AMS No. 17, (1955),

1–63. (§§1.0, 4.0)[2] Surfaces parametriques generalisees. Bull. Soc. Math. France 79 (1951), 59–84.

(§§1.0, 4.0)Zworski, Maciej. Decomposition of normal currents. Proc. AMS 102 (1988), 831–839.

(§4.5)

Index of Symbols

restriction of measures, 27, or interior multiplication, 52ð Cartesian product, 31, 52^ wedge, 36Ł (superscript) dual, 37! ES denotes oriented tangent plane to S, 39, 45jj jj mass norm of m-vector, 38, or measure jjTjj associated

with current T, 45jj jjŁ comass norm, 38# (subscript) push-forward f#, 44# (superscript) pull-back f#, 44hÐ, Ði pairing of vectors and covectorshÐ, Ð, Ði slice, 52[a, b] current associated with oriented line from a to bapDfa� approximate derivative of f at a, 13apJfa� approximate Jacobian associated with apDfa�ap lim approximate limit, 13˛m volume of unit m-ball, 8, 9Bma, r� closed m-ball about a of radius rˇn,m� normalizing constant, 11BV functions of bounded variation, 117C (superscript) AC denotes complement of set ACk having continuous derivaties of order k�A characteristic function of set Ad exterior derivative, 37

217

218 Index of Symbols

D derivativeDm C1 differential forms with compact support, 38

D m currents, 38∂ boundary, 39υ variation, 70, 108da point mass current at adiam diameter of a set, 8div divergence, 38ei basis vector, 35eŁi dual basis vector, 37

eij basis 2-vector, 36En volume current in Rn, 99Em currents with compact support, 40F real flat norm of current, 46–47Fm real flat chains, 40, 47F integral flat norm of current, 41, 49, 87Fm integral flat chains, 41Hm Hausdorff measure, 8

Im integral currents, 39, 41Im integralgeometric measure, 11

i,j,k orthonormal basis for R3

Jkfa� Jacobian, 24K (subscript) support in K, 44Ln Lebesgue measure, 8

m push-forward of m-vector, 44mRn m-vectors, 35–37mRn m-covectors, 37Lip f Lipschitz constant, 21loc (superscript) localized spaces, 87, 117M mass of current, 41M, ε, υ� minimal, 109�r homothety, 66n exterior normal vector, 114Nm normal currents, 40, 46–47NfjA, y� multiplicity function, 24OŁn, m� projections of Rn onto m-dimensional subspaces, 11Pm real polyhedral chains, 40, 47Pm integral polyhedral chains, 41Rm real flat chains of finite mass, 47R�m rectifiable currents modulo �, 105

Index of Symbols 219

Sma, r� m-sphere about a of radius rspt support, 37, 39TanE, a� tangent cone, 27–28TanmE, a� approximate tangent vectors, 30mA, a� density of set A at point a, 11mŁ A, a� lower density of set A at point a, 17m�, a� density of measure � at point a, 12mT, a� density of current T at point a, 88mT, a, r� mass ratio of current T at point a, radius r, 88Uma, r� open m-ball about a of radius rvE� varifold associated to set E, 108!n� Besicovitch constant, 14

Name Index

(See also the bibliography, which includes cross-references to each citation within this

text.)

Ackerman, J., 124, 128Almgren, A., 132Almgren, F., viii, 2, 83, 129, 132, 182,

203, 212Almgren, K., 182Almgren, R., 132Archimedes, 125, 127Aristotle, 125

Barschdorf, M., viiiBesicovitch, A. S., 14–18, 32, 204Brakke, K., 170, 178, 204

see also Surface EvolverBray, H., vii, 177, 204Bredt, J., iii, viiiBrieskorn, E., 174

Caccioppoli, R., 35Callahan, M., 70Caraballo, D., 182Chang, S., 83, 204Cheng, B., viiiCourt, L., viii

De Giorgi, E., 1, 2, 35, 81, 103, 204,205

Douglas, J., 1, 205Duncan, M., 182

Enneper, 67, 69, 71Euler, L., 68, 69

Favard, J., 11Federer, H., viii, 1, 32, 35, 205Fleming, W., 1, 2, 35, 106, 206Foisy, J., 121, 123, 141, 206Fowler, D., 125Fu, J., 182

Giga, Y., viiiGoro, F., iv, 110

Hales, T., 157, 207Hausdorff, F., 8Hoffman, D., vii, 70, 71, 208Hoffman, J., 70, 71Hutchings, M., 147, 148, 205, 208

221

222 Name Index

Jarvis, T., 122, 126Johnson, S., 164

Kelley, J., 182Kelvin, Lord (W. Thomson), 168, 214Kevitt, L., viiiKitrosser, I., 159

Lai, Y., 148, 207, 212Larson, G., 160Lawlor, G., viii, 209Lieb, C., 182Lieb, E., 182

Mackenzie, D., 182, 211McIntosh, R., viiiMessaoudene, M., viii, 182Meusnier, 68Morgan, F., ii, 106, 135, 182, 227Morgan, F. E., iiMorgan, F.E., Sr., iv

Nelson, E., 182

Parks, H., 182Peters, K., viiiPhelan, R., 157, 168–171, 214

Rado, T., 1, 212Rayment, T., 158

Reifenberg, E. R., 1, 14, 35, 212Roosen, A., 182

Scherk, 68Schoen, R., vii, 178, 212Selemeyer, C., ii, 227Silva, C., ivSimons, J., 81, 213Simplikios, 125Smale, S., 174Solomon, B., 32, 55Steiner, J., 127Sullivan, J. M., iv, viii, 14, 121, 123, 125,

142, 143, 213

Taylor, J., vii, 110, 132, 133, 182, 203,211, 213

Thomson, W. (Lord Kelvin), 168, 214

Weaire, D., 157, 168–171, 214White, B., 64, 116, 121, 143, 213, 214Whitney, H., 49, 214

Yau, S-T., vii, 178, 212, 215Yip, A., 182Young, L. C., 1, 35, 215

Zenodorus, 125, 127, 158Ziemer, W., 106

Subject Index

Almgren’s lemma, 129Angles between planes, 37, 76Approximate

continuity, 13, 18, 19differentiability, 13limits, 12tangent vectors Tanm�E, a�, 27

Approximation theorem, 77Area formula, 25

general area-coarea formula, 30Area-minimizing, 64, 66, 67, 88

preserved in limit, 94Aubin Conjecture, 184Axis of instability, 152

Besicovitch covering theorem, 14Bicycles, 28Bieberbach inequality, 16Black holes, vii, 178Borel

regularity, 7sets, 7

Boundary, see also Free boundarycurrent, 39, 42–44regularity, 84, 111, 117

soap films, 136–137Bounded variation BV, 117

Branch point, 74, 83Burlington, 135

Calculus of variations, 1, 69Calibration, 72, 74

const. mean curv. surf., 74Fundamental Theorem, 75history, 75

Cantor set, 10, 19Cantor-like square, 33Cantor-like triangle, 19positive measure, 30, 57

Caratheodory’s criterion, 7Cartesian product

area-minimizing, 102currents, 52measures, 31rectifiable sets, 31

Catenoid, 68, 76, 104, 199Chordal isoperimetric inequality, 163Closure theorem, 65Coarea formula, 26

general area–coarea formula, 30Comass jjϕjjŁ, 38Compact Lipschitz neighborhood retract,

44, 64, 115Compactness theorem, 64, 92

223

224 Subject Index

Complex analytic varieties, 74, 75, 83Concavity, 146Cone, 91, see also Tangent cone

area, 34Constancy theorem, 51Constant mean curvature, 74, 85Courant Nodal Domain theorem, 151Covectors, 37

m-covectors, 37Convex integrand, 118Crofton’s formula, 31Crystals, vii, 108, 118, 174–177

growth, vii, 177Currents Dm, 4, 35, 39–51, see also Recti-

fiable currentscompact support Em, 39, 40

Cylinder, 100, 104

Decomposable m-vector, 36Decomposition lemma, 98, 147Deformation theorem, 59Delaunay surfaces, 148Density

currents m�T, a�, 88, 90, 93upper semicontinuity, 94

lower mŁ �A, a�, 17

measures m��, a�, 12sets m�A, a�, 11

equals �A a.e., 17Diameter of sets, 8Differential forms, 35Direct method, 3Double bubble, vii, viii, 121, 123–125,

141–156component bounds, 149–150concavity, 146connected exterior, 146equal volumes, vii, 141rotational symmetry, 143, 121structure, 148

Edge energy, 136–138Elliptic integrand, 119–120, see also

IntegrandEnneper’s surface, 67, 69, 71, 76

Euler–Lagrange equation, 69Existence

area-minimizing surfaces, 64, 66fluid clusters, 173homologically minimizing surfaces, 66,

115–116soap bubble clusters, 129

Exterior derivative d, 37

FAR SIDE, 160Finite perimeter sets, 35First variation of a varifold υV, 108Flat, see also Integral flat chains

chains modulo �, 105norm F, 46–47norm F, 41, 49, 87

Flatness theorem, 49Flows, 177, viiiForce balancing, 153Fractals, 10Free boundary, 66, 115–117Fubini’s theorem, 27, 31

Gauss–Green–De Giorgi–Federertheorem, 113, 130

Green’s theorem, 113

Hahn–Banach theorem, 47Hausdorff

dimension, 10, 18, 186measure H

m, 8, 16, 24Helicoid, 68, 71, 76Hexagonal honeycomb, iv, viii, 157–168

bees’, 158, 159, 160, 166Homology, 66, 115Homothety, 66, 193Honeycomb, see Hexagonal honeycomb

Immiscible fludis, 173, viiiIntegral

currents Im, 39, 41flat chains Fm, 41polyhedral chains Pm, 41varifolds, 108

Integralgeometric measure Im, 11, 31, 32

Integrands, 118, 85, 89Isodiametric inequality, 16

Subject Index 225

Isoperimetric inequality, 62, 66, 117,125–128, 184, viii

chordal, 163hexagonal, 161higher codimension, 181

Jacobians, 24

Kahler form, 74Kelvin conjecture, 157, 168Knot energy, 177

Least-area disc, 1–3, 81, 82, 85Lebesgue

measure Ln, 8

point, 13, 21Lipschitz

constant Lip f, 21functions, 21–24

approximation by C1 function, 23differentiable a.e., 21

Local flat topology, 87Locally defined currents, 87

Mapping currents, 44Mass

current M�T�, 41lower bound, 89lower semicontinuous, 57

ratio, 88m-vector jj�jj, 38

Maximum principle, 101, 85Mean curvature, 67, 108

constant 74, 85Measures, 7

associated with current jjTjj, 45�M,ε, υ�-minimal sets, 109, 133Minimal surface, 67

equation, 67examples, 68–71nonuniqueness, 67, 71, 104system, 74

Mobius strip, 105–106Modulo �, 105Monotonicity of mass ratio, 88, 89Morse–Sard–Federer theorem, 113

Multiplicity,current, 45function, 24

Nested sets, 98Nonuniqueness, 67, 71, 104, 121, 124Normal currents Nm, 40, 46–50Norms

on currents, 41, 46–47, 49, 87as integrands, 118, 127on vectors, 38

Octahedral frame, 136, 169, 174Orientation, 31, 39, 108Oriented tangent cones, 91, 93, 95

density, 93

Penrose conjecture, vii, 127, 178Parametric integrand, 118Polyhedral chains Pm, 44Polymers, viiPositive mass theorem, vii, 178Pressure, 137, 143, 154Pull-back f#, 44Purely unrectifiable sets, 32, see also

Cantor setPush-forward f#, 44

Rademacher’s theorem, 21Real

flat chains Fm, 40, 47polyhedral chains Pm, 40, 47

Rectifiablecurrents Rm, 3, 40–44, 46

modulo � R�

m , 105sets, 4, 28–30

Regularityarea-minimizing rectifiable currents, 6,

81–85, 97, 101, 117modulo � R

�m , 107

fluid clusters, 173�M,ε, υ�-minimal sets, 112, 133soap bubble clusters, 112, 128, 132–137

Relative homology, 115Representable by integration, 45

226 Subject Index

Riemannian manifolds, 44, 62, 64, 85,127, 131, 184

Riemannian metric, 118

Salt crystals, 175–177Sard’s theorem, 113Scherk’s surface, 68, 76Schwarzschild space, 127Self-similarity, 10Sierpinski sponge, 10Simons

cone over S3 ð S3, 81, 103–104lemma, 102

Simple m-vector, 36Slicing, 52–55, 76“SMALL” Geometry Group, vii, viii, 123,

125, 141, 174, 175Smoothing, 50, 48Soap bubbles, ii, vii, viii, 109–111,

121–156connected regions, 122, 126, 127, 129equilibrium, 137existence, 128–131planar, vii, 123–127pressure, 137, 143, 154regularity, 112, 128, 132–137single, 125standard, 143

Soap films, 109–111, 107, 108, 116,136–137

Solutions to exercises, 185–201Sphere eversion, iv, 178–179Stationary varifold, 108Stokes’s Theorem, 6, 39Structure theorem for sets, 32, 55, 63

Supportcurrent, 39differential form, 37

Surface Evolver, 178, 121, 122, 169, 171Surfaces, see also Rectifiable currents

mappings, 1–3Symmetrization, 16, 127, 143–146

Tangentcone Tan(E, a), 27–28

approximate Tanm�E, a�, 30oriented, 91, 93, 95, 89

vectors, 27approximate, 27

Tetrahedron, iv, 110–111, 134, 138Trefoil knot, 86, 196, 106Triangle inequalities, 173Triple bubble, 125, 156Truncation lemma, 161

Undergraduate research, vii, viii, 123,125, 141, 156, 174, 175

Uniform convexity, 119Unoriented surfaces, see Flat chains

modulo �

Varifolds, 107Vectors, 35–37

m-vectors, 35–37Volume constraint, 85

Weak topology on currents, 39Whitney extension theorem, 23Willmore conjecture, 175Wirtinger’s inequality, 74Wulff shape, 175

geometric measure theory

Documents