AN ELEMENTARY PROOF FOR THE DOUBLE BUBBLE PROBLEM IN 31 NORM PARKER DUNCAN parkeraduncan@tamu.edu 0 RORY O'DWYER 0 EVIATAR B. PROCACCIA 0 Here we are interested in the double bubble perimeter of two simply connected open sets A, B

We study the double bubble problem with perimeter taken with respect to the 31 norm on R2. We give an elementary proof for the existence of minimizing sets for any volume ratio parameter 0 < ³ ≤ 1 by direct comparison to a small family of parameterized sets. By simple analysis on this family we obtain the minimizing shapes found in [9]. In [5] and [6] the double bubble conjecture in R2 and R3 was established, stating that the unique perimeter-minimizing double bubble which encloses two fixed volumes consists of three spherical caps whose tangents meet at an angle of 120 degrees. Recently the Gaussian double bubble conjecture was established by Milman and Neeman [8]. These problems are an extension of the classical isoperimetric problem stating that the perimeter minimizing shape for a fixed volume is the sphere. The case of the isoperimetric problem in which the perimeter was taken with respect to any norm on Rn was solved as well. Namely, Taylor [10, 11] proved that the unique solution of the isoperimetric inequality with respect to any norm à is the renormalized ball in the dual norm, the so called Wulff construction [12]. For example the isoperimetric shape with respect to the 31 is the 3> ball [0, 1]n. Such non isotropic isoperimetric problems arose naturally in the field of probability, mostly in scaling limits of percolation clusters in a lattice [1, 2, 3, 4]. In this paper we study the double bubble problem with respect to the 31 norm. The result discussed in this paper was first proved in [9], and is based on previous geometric measure theory results. However our proof is self contained and considerably simpler. Moreover our simple approach, that uses no geometric measure theory, seems to be more amenable to generalizations to higher dimensions.

1. Introduction

1.1. Notations and results. For any Lebesgue-measurable set A ⊂ R2, let µ (A) be its Lebesgue measure. For a simple curve » : [a, b] → R2, not necessarily closed, where » (t) = (x(t), y(t)), define its 31 length by

N Ã (» ) = sup sup X

Ng 1 af t1f ...f tN f b i=1

x(ti+1) − x(ti) + y(ti+1) − y(ti) .

If we wish to measure only a portion of the curve » , it will be denoted à (» ([t, t2])), where [t, t2] ⊂ [a, b]. For simplicity we assume that [a, b] = [ 0, 1 ] unless otherwise stated.

We say that two curves », » 2 : [ 0, 1 ] → R2 intersect nontrivially if there are intervals [s, s2], [t, t2] ⊂ [ 0, 1 ] such that » ([s, s2]) = » 2([t, t2]), then their nontrivial intersection can be written as the union of curves » i such that » i([ 0, 1 ]) = » ([si, si+1]) → R2 for some intervals [si, si+1], and we define the length of the nontrivial intersection to be à (» ∩ » 2) := Pi à (» i). similarly for B, and where the intersection of the boundaries of A and B is a union of disjoint, rectifiable curves. The double bubble perimeter is defined as

à DB(A, B) = à (» ) + à (» 2) − à (» ∩ » 2), where "A = » ([ 0, 1 ]), and "B = » 2([ 0, 1 ]). We will also use the notation à (» ) = à ("A ).

For ³ ∈ (0, 1], define: ³ ³ = {(A, B) : A, B ⊂ R2, where A, B are disjoint, simply connected open sets, and "A ,"B ,"A ∩ "B are unions of closed, continuous, simple, rectifiable curves, with µ (A) = 1, µ (B) = ³ }.

Let

à DB(Γ³ ) := inf{à DB(A, B) : (A, B) ∈ ³ ³ }, be the infimum of the double bubble perimeter (bounded below by zero).

The main result in this paper is: Theorem 1. For 0 < ³ ≤ 1,

I. The set Γ³ := {(A, B) ∈ ³ ³ : Ã DB(A, B) = Ã DB(Γ³ )} is not empty.

II. The infimum, √ Ã DB(Γ³ ) = (4√1 + ³ + 2√³ )1(0, 688−44980√2 ] + (4 + 2 2³ )1 688−44980√2 , 1 + (2p6(1 + ³ ))1[ 12 ,1 ]. 2 III. For ³ = 6882 44980: 2 we have |Γ³ | ≥ 2, with Γ³ containing both sets in Figure 1 (a) and (b). Moreover for ³ ∈ [1/2, 1], Γ³ contains Figure 1 (c), for ³ ∈ Γ³ contains Figure 1 (b), and for ³ ∈ 0, 6882 44980: 2 , Γ³ contains Figure 1 (a).

6882 44980: 2 , 12 , √³

√³ ³ q Corollary 1.1. There are two critical ³ ’s at which à DB(Γ³ ) undergoes a phase transition. The first, at ³ = 6882 44980: 2 , is discontinuous in the first order derivative, while the second, at ³ = 1/2, is discontinuous in the second order derivative.

Before we explain the proof strategy we define in Figure 2 a finite family of set types abbreviated F³ ⊂ ³ ³ :

While the general case encapsulates the other cases, the kissing rectangles and embedded rectangle cases are important enough to annotate and include with names. We will be referring to these annotations later in the paper.

The strategy for proving Theorem 1 follows 3 steps: (1) Begin with any two sets (A, B) ∈ ³ ³ , and find sets (A˜, B˜) ∈ F³ with à DB(A˜, B˜) ≤ à DB(A, B). This part is done in Section 2. (2) Since the sets in F³ are very simple to analyze, and the family is finite, we can show the existence of

arg inf{Ã DB(A, B) : (A, B) ∈ F³ }.

This part is done in Section 3.

(3) Finally by the previous points, these sets achieve the infimum over all of ³ ³ proving the existence of an element in Γ³ . Moreover we get the phase transitions in ³ and show non uniqueness for the first phase transition. This is done in Section 4.

2. Finding the sets in F (³ )

Our goal in this section is to find elements of F³ with a smaller double bubble perimeter than given sets A with µ (A) = 1, and B with µ (B) = ³ .

Definition 2.1. (1)

A := [aleft, aright] × [abottom, atop], where aleft = inf{x : (x, y) ∈ A for some y ∈ R} aright = sup{x : (x, y) ∈ A for some y ∈ R} abottom = inf{y : (x, y) ∈ A for some x ∈ R} atop = sup{y : (x, y) ∈ A for some x ∈ R} Lemma 1. Ã (A ) ≤ Ã (A) and µ (A) ≤ µ (A ).

Proof. By definition, A ⊂ A . Therefore, by monotonicity of Lebesgue measure, µ (A) ≤ µ (A ).

Now we show that à (A ) ≤ à (A). Let H1 be the horizontal line passing through atop, and H2 be the horizontal line passing through abottom. Similarly define V1 and V2 to be the vertical lines passing through aleft and aright, respectively. Let DH1,H2be the distance from H1 to H2, and DV1,V2 be the distance between V1 and V2. "A must touch H1 in at least one point, say p1, and similarly must touch H2 in at least one point, say p2. See Figure 3 for an illustration of the notations.

Since A is open and "A is simple, there must be at least two disjoint paths in "A from p1 to p2, abbreviate them » i(t) = (xi(t), yi(t)) : [ 0, 1 ] → R2, for i ∈ {1, 2}. The vertical portion of these paths must be at least DH1,H2 i.e. for any 0 ≤ t1 ≤ t2 ≤ · · · ≤ tn ≤ 1, and i ∈ {1, 2} we have that

DH1,H2

N X j=1

V2 p1

H1

H2

DV1,V2 yi(tj+1) − yi(tj) ≥ DH1,H2 Similarly there must be at least two disjoint paths from a point on V1 to a point on V2, whose horizontal distances must be at least DV1,V2. We have so far found that the boundary of A must measure at least 2 · DV1,V2 + 2 · DH1,H2, which is the length of the boundary of A . That is, Ã (A ) ≤ Ã (A), as claimed.

Lemma 2. If (A, B) ∈ ³ ³ , and B ⊂ A , then there exists (A˜, B˜) ∈ F³ , such that à DB(A˜, B˜) ≤ à DB(A, B), and (A˜, B˜) are either kissing rectangles or embedded rectangles. ˜ B ˜ A ˜ A ˜ B Proof. If B is contained in A there are several options to consider, namely that one, two, three, or no edges of B could touch the same number of edges in A . Let’s take the instance when none of the edges of B touch any of the edges of A , that is, btop < atop, bright < aright, aleft < bleft, and abottom < bbottom. From this case we can easily derive the results for the other cases. Let H1 = R × {atop}, H2 = R × {btop}, H3 = R × {bbottom}, and H4 = R × {abottom}. Now we define the distance between H1 and H2 to be DH1,H2 = atop −btop, the distance between H2 and H3 to be DH2,H3 = btop −bbottom, and the distance between H3 and H4 to be DH3,H4 = bbottom − abottom. Similarly we call the vertical line through aleft V1, the vertical line through bleft V2, the vertical line through bright V3, and the vertical line through aright V4, naming the distances between these lines just as before. See Figure 4 for illustration. V1

V2

V3

V4 DH1,H2 DH2,H3 DH3,H4 DV1,V2

DV2,V3

Now, in "A ∪ "B we need to find three paths with vertical lengths of DH2,H3, and in "A two paths with vertical lengths of DH1,H2 and the same for DH3,H4. Further, we need these paths to be disjoint so we don’t count anything more than once. Then we would do the analogous process for horizontal distances. First, we find points p1 ∈ H1 ∩ "A , and p2 ∈ H4 ∩ "A . There must be two distinct paths in "A between these two points, which we call » i(t) = (xi(t), yi(t)) : [ 0, 1 ] → R2, for i ∈ {1, 2}, both of which must have vertical distance of at least DH1,H2 + DH2,H3 + DH3,H4. That is, for any 0 ≤ t1 ≤ t2 ≤ ... ≤ tN ≤ 1, we have for i ∈ {1, 2},

N X (|yi(tj+1) − yi(tj )|) ≥ DH1,H2 + DH2,H3 + DH3,H4.

j=1

To complete our search for enough vertical length, it remains to find one final path with vertical length DH2,H3 that we have yet to count among these crossings. For a path ³ : [ 0, 1 ] → R2 we say that a subpath ³ [a, b] is a crossing of S := R × (bbottom, btop) if ³ (a) ∈ H2, ³ (b) ∈ H3 and for all a < t < b, ³ (t) ∈/ {H2, H3} (or in the other direction). If "A contains more than two crossings, then each of these crossings must have vertical length at least DH2,H3, and we are done. So we may assume that "A only contains two crossings of S. We denote these two crossings ¿ 1, ¿ 2. Since "A contains exactly two crossings of S, S \ ¿ 1 ∪ ¿ 2 consists of three open sets, exactly two of which are unbounded and have joint boundary with both H2 and H3. We call these two open sets S1 and S2. Since B is open and connected and contained in S, B must be contained in one of these unbounded open sets, say w.l.o.g. S2 (note that if there are more than 2 crossings then B can be contained in a bounded set). Since B is open and "B is rectifiable, there are at least two distinct paths in "B from H2 to H3, which we call » 3 = (x3(t), y3(t)), and » 4 = (x4(t), y4(t)). Both of these paths must be in S2. By planarity only one of these paths might intersect ¿ 2 say w.l.o.g. » 3. This means that we have not counted the vertical part of » 4, and it must have vertical length at least DH2,H3. That is to say, that for any 0 ≤ t1 ≤ t2 ≤ ... ≤ tN ≤ 1, we have

N X (|y4(tj+1) − y4(tj)|) ≥ DH2,H3.

j=1

This is our third such length, and we are done finding vertical lengths. We need only find horizontal lengths now. But this is done in exactly the same manner as in our search for vertical lengths. We could even rotate our figures 90 degrees either left or right, so that vertical lines become horizontal and vice versa, and perform the exact same proof as above.

We now construct our figure. First, we have found a total length of

2 · (DH1,H2 + DH3,H4 + DV1,V2 + DV3,V4) + 3 · (Dh2,H3 + DV2,V3) .

This gives us enough length to construct A and still have left over a total length of DH2,H3 + DV2,V3. With these lengths, we construct a box in the corner of A with dimensions DH2,H3 × DV2,V3. This box, which we call B˜ in the corner has volume the same as B , and therefore volume at least ³ . We can shrink it easily so that it has volume exactly ³ , and by abuse of notation still call this possibly smaller rectangle B˜. On the other hand, A ∪ B ⊂ A , and therefore µ (A ) ≥ 1 + ³ . Therefore, µ (A \ B˜) ≥ 1 + ³ − ³ . So we can easily move the sides of A that don’t share joint boundary with B˜ inwards until the volume is exactly 1. This set we call A˜, and have completed our construction.

Now, suppose that bbottom = abottom, or bleft = aleft, etc. That is one of the sides of B is contiguous with one of the lines of A . The process would be as above, except we would have H3 = H4. This means we wouldn’t have to find two paths with vertical length DH3,H4. The rest of the proof would be the same. Similarly, if two sides of B are contiguous with two sides of A , say abottom = bbottom and aleft = bleft, then we wouldn’t have to find paths with vertical length DH3,H4 and we wouldn’t have to find paths with horizontal length DV1,V2. The rest of the proof would be the same.

The previous lemma took into account all of the cases when all four corners of B are contained in A . The only other two options are if two or one corner of B is contained in A .

Lemma 3. If (A, B) ∈ ³ ³ , and exactly one corner of B is contained in A , then there exists (A˜, B˜) ∈ F³ , such that à DB(A˜, B˜) ≤ à DB(A, B).

Proof. For the one corner case we argue that we can find sets, with a better double bubble perimeter and more joint volume than the original shapes, that looks like the general case of Figure 2:

Here, the rectangle can be either A or B , say A , and the other set is B \ A . Once we create these two sets, we are not necessarily done because the volumes may not be correct. This may cause somewhat more of a problem than in previous cases, but in any case the method remains similar.

In this proof we are assuming that µ (A) = 1, and µ (B) = ³ . Since A and B only intersect in one corner, either atop > btop, or btop > atop. We can suppose without loss of generality that atop > btop. Let H1 = R × {atop} be the horizontal line passing through atop, H2 = R × {btop} be the horizontal line passing through btop, H3 = R × {abottom} be the horizontal line passing through abottom, and H4 = R × {bbottom} be the horizontal line passing through bbottom. Note that the height of A ∩ B is the same as the distance between H2 and H3, which we will call DH2,H3. Furthermore, let the distance between H1 and H2 be DH1,H2, and the distance between H3 and H4 be DH3,H4. Now, let V1 = {bleft} × R be the vertical line passing through bleft, V2 = {aleft} × R be the vertical line passing through aleft, V3 = {bright} × R be the vertical line passing through bright, and V4 = {aright} × R be the vertical line passing through aright. We define the distance between V1 and V2 to be DV1,V2, the distance between V2 and V3 to be DV2,V3, and the distance between V3, and V4 to be DV3,V4. See Figure 5.

V1

V2

V3

V4 DH1,H2 DH2,H3 DH3,H4 H1 H2 H3

H4

Notice that to construct A and B \ A in this way we need two lengths of DH1,H2, DH3,H4, DV1,V2, and DV3,V4, as well as three lengths of DH2,H3, and DV2,V3. In other words, for à DB((A , B \ A )) to be at most à DB((A, B)), we need to find two vertical lengths of DH1,H2 in "A , two vertical lengths of DH3,H4 in "B , and three vertical lengths of DH2,H3 between "A and "B . Similarly for horizontal lengths.

First, there must be a point p1 ∈ H1 ∩ "A , and another point p2 ∈ H3 ∩ "A . Between these two points there must be at least two disjoint paths, which we call » i(t) = (xi(t), yi(t)) : [ 0, 1 ] → R2, i ∈ {1, 2}, in "A , both of which have vertical length at least DH1,H2 + DH2,H3. That is, as before, for any 0 ≤ t1 ≤ t2 ≤ ... ≤ tN ≤ 1, N X(|yi(tj+1) − yi(tj)|) ≥ DH1,H2 + DH2,H3 j=1

Similarly in "B , we can find two disjoint paths of vertical length at least DH2,H3 + DH3,H4. Since there can be no joint boundary below H3, the portions of these paths that measure at least DH3,H4 have not been counted yet. It remains to find a path in "B with vertical length at least DH2,H3 that we have yet to count.

To find this path we define the infinite strip of height DH2,H3, S := R × (abottom, btop). If "A has more than two crossings of S, then each of these crossings has vertical length of at least DH2,H3, as above, and we are done. So we may assume that "A contains exactly two crossings of S (it can’t be less than two as we noted above). Abbreviate these crossings ¿ 1, ¿ 2. In this case, consider S \ ¿ 1 ∪ ¿ 2. This consists of three open sets, but only two are unbounded sets, and in one of these open sets that we find B ∩ S. We call these sets S1 and S2 such that ¿ 1 ⊂ "S 1 and ¿ 2 ⊂ "S 2. Suppose without loss of generality that B ∩ S ⊂ S2. There is at least one point p3 ∈ "B ∩ H2, and at least one point p4 ∈ "B ∩ H3, and there must be two distinct paths in "B from p3 to p4. We call these paths » i(t) = (xi(t), yi(t)) : [ 0, 1 ] → R2, i = 3, 4. By planarity, only one of the paths, either » 3 or » 4 can have joint boundary with ¿ 2, say » 3. This means that we have yet to include the vertical length of » 4, and we have just found our third vertical length of DH2,H3. That is, for any 0 ≤ t1 ≤ t2 ≤ ... ≤ tN ≤ 1,

N X(|y4(tj+1) − y4(tj)|) ≥ DH2,H3.

j=1

Finding the three horizontal lengths of DV2,V3 follows the same argument. The reason we can be sure that we won’t double count anything is because, inside A ∩ B we have only counted vertical distance, and in the argument to find our three horizontal lengths of DV2,V3 we would only count horizontal lengths. So, we have proved that à DB(A , B ) ≤ à DB(A, B). It is clear that µ (A ∪ B) ≤ µ (A ∪ B ). Since A ⊂ A , 1 = µ (A) ≤ µ (A ). However, it is possible that µ (B \ A ) < µ (B) = ³ . This we must correct. For this purpose, let us refer to the notation we established in the introduction for the general case.

a c A b d e

B \ A

f

We look at the two intervals of lengths c, f and assume w.l.o.g that f > c (see Figure 6). Now we slide B \ A down towards the longer edge f . By doing so we do not enlarge the boundary and we can only increase the area. There are two options now: (1) If c + d + e ≤ a we get kissing rectangles where the rectangle on the left has greater area than B \ A . (see Figure 7).

Next we move area from A to the rectangle on the right until reaching the appropriate area ³ . Consider f 2 ≥ f such that µ (B˜) = f 2 · (c + d + e) = ³ , and let b2 = b − (f 2 − f ). Our final sets are of the general case class as can be seen in Figure 8. Since we have only enlarged the total area we are guaranteed that µ (A˜) ≥ 1. Reducing the area is easy and we have dealt with this before. b − (f2 − f) c + d + e ˜ B f2 ˜ B ˜ f d2 f a A˜ ˜ b

c

Now we need only deal with the case when two corners of B are in A . Lemma 4. If (A, B) ∈ ³ ³ , and two corners of B are contained in A , then there exists (A˜, B˜) ∈ F³ , such that à DB(A˜, B˜) ≤ à DB(A, B).

Proof. Here we must have one of the following: btop > atop, bright > aright, bbottom < abottom, or bleft < aleft. Let’s suppose, without loss of generality, that bright > aright. We construct our configuration in a similar way as before. First, since two corners of B are in A , it follows that aleft < bleft ≤ aright. However, if bleft = aright, we can just replace A with A and B with B . This will increase the volume of both A and B, and increase their joint boundary. Then we can reduce the volumes as necessary, which will only decrease the double bubble perimeter. So, we can assume that aleft < bleft < aright. Let H1 be the horizontal line passing through atop, H2 be the horizontal line passing through btop, H3 be the horizontal line passing through bbottom, and H4 the horizontal line passing through abottom. We define DHi,Hi+1 as before, i = 1, 2, 3. Similarly define V1, V2, V3, V4 as the vertical lines passing through aleft, bleft, aright, and bright, respectively. Let DV1,Vi+1 be defined as before, i = 1, 2, 3. Notice that we haven’t eliminated the possibility that H1 = H2, or H3 = H4, or both. See Figure 11.

We wish to find two disjoint paths with vertical lengths at least DH1,H2, two disjoint paths with vertical length as least as long as DH3,H4, three disjoint paths with vertical

V1

V2

V3

V4

DV2,V3 DV3,V4 lengths at least as long as DH2,H3, two disjoint paths with horizontal length at least as long as DV1,V2, three disjoint paths with horizontal lengths at least as long as DV2,V3, and two disjoint paths with horizontal lengths at least as long as DV3,V4 . We achieve this much the same as in the previous Lemma.

Finding the horizontal lengths is nearly identical. We can then proceed to construct our sets. If move B up until H1 = H2 we reduce the double bubble perimeter and we get a shape of the general case type. Now we can fix the volumes in the same way as in the previous Lemma.

3. KKT analysis

In the previous section, we constructed a finite list of set types F³ in which from any configuration (A, B) ∈ ³ ³ we obtain a configuration (A˜, B˜) ∈ F³ so that à DB((A˜, B˜)) ≤ à DB((A, B)). In the upcoming analysis, however, it is convenient to note that all the cases in F³ can be represented by the 6 parameter configuration we called “General Case” given in Figure 12.

ρDB = 2(a + b + c + f) + (d + e) ab = 1 or α b ≥ d, a ≥ e cd + cf + ef = α or 1 a, b, c, d, e, f > 0 a c

d b e

f

The other cases, kissing rectangles for instance, occur when some of the parameters defining the geometry of the configuration are set to zero. For kissing rectangles, this would consist of setting e and f to zero (or c = d = 0) in Figure 12. Using this configuration we have made a geometric problem into the minimization of a hyper-plane with algebraic inequality constraints. We want to minimize the à DB in Figure12, while remaining within the inequality contraints.

The Karush Kuhn Tucker method [ 7 ] is one method used for calculating minima in such problems; with some care the global minimum of each element of the above list can be found. In the previous lemmas, it was sometimes ambiguous which of the two shapes in a configuration had unit volume and which ³ . For our purposes, that just means we must alternate which shape has volume ³ and analyze both.

The original problem we would have to solve is 6 dimensional, but our two equality constraints reduce it to four. These are that ab = ³ or 1 and cd + cf + ef = 1 or ³ . Using these constraints, our problem becomes to minimize 2(a + (1 or ³ ) − cd c + e

The variable a cannot be zero, so we can exclude its inequality lagrange multiplier. The (³ or 1) ≥ ad constraint comes from b = ³ oar 1 ≥ d and (1 or ³ ) ≥ cd from (1 orc +³e)2 cd = f ≥ 0. Note that the conventional conditions required for KKT (equality constraints be affine, etc.) are not satisfied here. In this case, however, we can constrain our variable space in the upper octant and inside some hypercube. This is because if one of the variables becomes larger than some example double bubble perimeter, it is not optimal. A final precaution we address is that due to the exclusion of the constraint a ≥ 0 from our list of inequality constraints under analysis, our space of variables is not truly closed and not truly compact. In the upcoming analysis, other portions of the variable space will similarly be excluded (it will be explicitly mentioned when a portion of the variable space is excluded) because they are not valid configurations. However at some finite distance in variable space from these coordinates, the fact that a (as an example) becomes small forces another dimension to grow as a2 1. This growth eventually pushes à DB over the perimeter of that example configuration we used to build the hypercube. Therefore in a similar manner we bound our domain in these degenerate cases by curves a finite distance from the degenerate case; the boundary here doesn’t require explicit checking because by construction it is too large.

Thus our modified variable space is truly closed and bounded and therefore compact. The global minimum is either along the boundary or is the lowest local minimum inside the domain itself. The KKT method checks all of this by implementing the inequality constraints which define the boundary.

Our above à DB yeilds the lagrangian

L =2(a + ³ or 1 (1 or ³ ) − cd

+ c + a c + e + cµ 3 + dµ 4 + eµ 5 + ((1 or ³ ) − cd)µ 6.

) + (d + e) + ((³ or 1) − ad)µ 1 + (a − e)µ 2

Lets call ³ = ³ or 1, ³ = 1 or ³ , then the gradient of L becomes 2(1 − a2 ) − dµ 1 + µ 2, 2(1 − (c + e)2 ) + µ 3 − dµ 6, 2( c−+ce ) + 1 + µ 4 − cµ 6, ...

³ ³ + ed

For the KKT method, the inequality lagrangian multipliers are either positive and their conditions applied (i.e. if µ 1 > 0 then ³ = ad) or they are zero. This means without symmetry arguments that there are 26 systems of nonlinear equations to check for minima. If µ 6 > 0 then cd = ³ , then from our second equality constraint we have that (c + e)f + cd = 1 or ³ and (c + e)f = 0, or c = e = 0 or f = 0. The first is impossible and the second becomes kissing rectangles. As one may see, in parts of this calculation we will eliminate some of the 64 possible systems of equations by demonstrating geometrically what they resolve to, and then doing that case only once. In this case, this simple calculation got rid of 31 systems and left us only needing to calculate the general kissing rectangles case. Similarly if µ 3 > 0, µ 4 > 0, or µ 5 > 0 then c = 0, d = 0, or e = 0 and again we have kissing rectangles. This means we only have to calculate it with µ 1 and µ 2 possibly positive, leaving 4 systems of equations and the kissing rectangles to analyze. If we remember that µ 1 corresponds to b ≥ d, then by symmetry we realize that µ 1 and µ 2 being activated alone have the same effect by symmetry of the figure, so we need only check µ 1 > 0 and both µ 1 > 0 and µ 2 > 0. So we have only 3 systems and kissing rectangles to calculate. 3.1. Kissing rectangles. For kissing rectangles we let the sides of one rectangle be a,b, and the other c,d (notation is changed for convenience, see Figure 2). So ab = ³ while cd = ³ . And taking the b-d edge to be the kissing side, we then take b to be the larger size (b ≥ d or ³a ≥ ³c ). Our lagrangian becomes is

³ ³ L = 2(a + + c) + + µ 1(c³ − a³ )

a c where the inequality constraint µ 1 is b ≥ d or in terms of a and c c³ ≥ a³ . The gradient The unconstrained minimum has a = √³ = b, c = q ³2 , d = √2³ and

³ ³ 2(1 − a2 ) − ³µ 1, 2 − c2 + µ 1³

= 0. Ã DB = 2 2p³ +

+ p2³. r ³ 2

The perimeter becomes either 4√³ + 2√2 or 4 + 2√2³ . This first equation again has ³ = 1, ³ = ³ , a = √³ = b, and d = √2. Now we need b ≥ d, which would mean ³ ≥ 2. This is a contradiction, because ³ ≤ 1. The second equation, in the exact same way, forces us instead to have √³ ≤ p1/2 =⇒ ³ ≤ 1/2. For ³ > 1/2 we need to apply the constraint. We know now that ³ = ,³ ³ = 1, so c = a³ , and our gradient becomes

Multiplying the second equation in the gradient by ³ and adding it to the first we have 2³ − a12 + 2 − a22 = 0, so 2(³ + 1) = a32 , a = q 2(³ 3+1) , c = a³ . Plugging these into our equation for à DB we obtain the kissing rectangle solution Lemma 5. For any ³ ∈ (0, 1] the infimum among all (A, B) ∈ F³ of the kissing rectangles type is achieved and admits: 3.2. Embedded rectangles. Now we go on to the case of min(µ 1, µ 2) > 0. This means b = d and a = e which becomes an “Embedded rectangle” type. Again it becomes useful to breifly depart from our original notation. We can label this structure with a,b being the inner rectangle’s sides, c and d the outer (See Figure 2). This means ab = ³ the volume of the inner rectangle, cd = ³ + ³ the volume of the inner rectangle and the outer peice, and our lagrangian becomes

2(c + d) + a + b + » 1(ab − ³ ) + » 2(cd − (³ + ³ )) or in an unrestrained form , so a = b = √³, c = d = √³ + ³ , and altogether the perimeter is

à DB = 2p³ + 4p³ + ³, which is either 2√³ + 4√1 + ³ or 2 + 4√1 + ³ . This second one is too large and is proven by our example in the paper to be suboptimal. So for this case we have Lemma 6. For any ³ ∈ (0, 1] the infimum among all (A, B) ∈ F³ of the embedded rectangles type is achieved and admits:

inf{Ã DB(A, B) : (A, B) ∈ F³ are embedded rectangles} = 2√³ + 4√1 + .³ 3.3. General case. We now check the unconstrained case where all µ i = 0. For this purpose we go back to the original notation we established.

Assume all µ i = 0. Then we have the following gradient

2(1 − a³2 ), 2(1 − (³ c ++ ee)d2 ), 2( c−+ce ) + 1, 2(− (³c +− ec)d2 ) + 1 = 0

So a = √³, c = e, d = 3³c . The second constraint gives us (c + e)2 = ³ + ed. Plugging in our expression for e and d in terms of c, we have c = q ³ . 3

Now that we know a and c in terms of ³ and ³ , we can obtain the perimeter from the first expression in the KKT analysis (i.e. the perimeter for the general case in terms of our variables).

We obtain the following as perimeter: 2(2√³ + 2q ³3 ) + 2q ³3 = 4√³ + 6q ³3 , which is 4√³ + 6q 13 or 4 + 6p ³3 . For the perimeter to represent a valid shape, we need b ≥ d =⇒ √³ ≥ √3, which is never true. The second double bubble perimeter is never optimal as can be seen by comparing it to the double bubble perimeter of the sets in Theorem 1 part III .

Now to the last case where only µ 1 > 0. So ³ = ad and we have the following gradient 2(1 − a2 ) − dµ 1, 2(1 − (c + e)2 ), 2( c−+ce ) + 1, 2(− (c + e)2 ) + 1 − aµ 1 ³ ³ + ed ³ − cd So c = e and this gradient becomes (by reducing the dimension and removing e) ³ ³ + cd ³ − cd (2(1 − a2 ) − dµ 1, 2(1 − 4c2 ), 2(− 4c2 ) + 1 − aµ 1) = 0

Multiply the first equation by a, the third by d, and subtract the third by first to get 2(a − ³a + ³d 42cc2d2 ) − d. Next from the second equation we get 4c2 = ³ + cd, so d = ³d 4+cc2d2 . Applying this to the derived equation, we get d = 2( ³a − a). Now from our µ 1 constraint we know ad = ³ or 2(³ − a2) = ³ , so we get a = q ³2 and d = √2³ . Now plugging d back into 4c2 = ³ + cd we get 4c2 − c√2³ − ³ = 0. So from the quadratic equation we have c = e = : 2³ +: 82³ +16³ (the minus solution is negative so it can’t work). This obtains all the relevant variables. From this we can plug into the perimeter and obtain: Ã DB = 2 q ³2 + q³ β2 + : 2³ +: 82³ +16³ + ³ 2 2√√2β2+β+√8√28β2β++161γ6γ: 2³ ! + √2³ + : 2³ +: 82³ +16³ For this if let ³ = ³ , it is always more than double bubble perimeter of the sets in Theorem 1 part III. If ³ = ³ and ³ = 1, then we do get valid and smaller answers for ³ < 0.12. But we also need e ≤ a, or : 2³ +: 82³ +16 ≤ p ³2 which is only true outside of the range from 0 to 1. Therefore this is an invalid answer and we obtain that: Lemma 7. The minimum of the double bubble perimeter over every configuration in F³ is the minimum between the kissing rectangles and embedded rectangle given in Lemmas 5 and 6.

Remark 3.1. The two graphs for V ol(A) = 1, V ol(B) = ³ and V ol(A) = ,³ V ol (B) = 1 are different because in one case the encased rectangle has volume 1 and never gets small enough to be absorbed into the bigger rectangle like in the embedded rectangle case, so we only see the kissing rectangles case for it. The other case exhibits all portions of the minimum. See figure 13 for a comparison of the two cases.

³)( B DÃ 7 6 5 4

Vol(A)=³ ,Vol(B)=1

Vol(A)=1,Vol(B)=³ 0 In this section we collect the results of the previous sections to prove our main result.

In Lemmas 5, 6 and 7 we analyze each of the possible elements of F³ using the KKT method and by comparing them we can find a global minimizer in F³ which we call here Ç ³ ∈ F³ satisfying:

Lemma 8. I. For any 0 < ³ à DB(Ç ³ ) ≤ à DB((A, B)). ≤ 1 there is a Ç ³ ∈ F³ such that for any (A, B) ∈ F³ , √ √ √ à DB(Ç ³ ) = (4 1 + ³ + 2 ³ )1[0, 688−44980√2 ](³ ) + (4 + 2 2³ )1 688−44980√2 , 1 (³ ) 2 + (2p6(1 + ³ ))1[ 12 ,1 ](³ ) Ç ³ satisfies Figure 14 (c), for ³ ∈ ³ ∈ Proof of Theorem 1. First let 0 < ³ ≤ 1. Take some sequence Ç i ∈ ³ ³ such that lwi meio³>btaiÃnDBan(Ç eil)em=enÃtD Bǘ(iΓ∈³ )F.³ BsyucLhemthmatasà D2B, (3ǘia)n≤d à 4D, Bf(oÇ r ie).acBhyeLleemmemnta o8f ptahrits Isetqhueerencies a Ç ³ ∈ F³ satisfying for any i ∈ N,

à DB(Ç ³ ) ≤ à DB(Ç ˜i) ≤ à DB(Ç i).

We have that

à DB(Γ³ ) ≤ à DB(Ç ³ ) ≤ ilim à DB(Ç i) = à DB(Γ³ ),

³> and thus Ç ³ ∈ Γ³ and Γ³ is non empty, establishing part I of Theorem 1. Since we have that Ç ³ ∈ Γ³ , Lemma 8 parts II and III establishes Theorem 1 parts II and III.

[1] K. Alexander , J.T. Chayes , and L. Chayes. The wulff construction and asymptotics of the finite cluster distribution for two-dimensional bernoulli percolation . Communications in mathematical physics , 131 ( 1 ): 1 - 50 , 1990 . [2] M. Biskup , O. Louidor , E. B. Procaccia , and R. Rosenthal . Isoperimetry in two-dimensional percolation . Communications on Pure and Applied Mathematics , 68 ( 9 ): 1483 - 1531 , 2015 . [3] T. Bodineau , D. Ioffe , and Y. Velenik . Rigorous probabilistic analysis of equilibrium crystal shapes . Journal of Mathematical Physics , 41 ( 3 ): 1033 - 1098 , 2000 . [4] R. Cerf . The Wulff Crystal in Ising and Percolation Models: Ecole D'Et´e de Probabilit´es de Saint-Flour XXXIV - 2004 . Springer, 2006 . [5] J. Foisy , Manuel Alfaro G., J. Brock , N. Hodges , and J. Zimba. The standard double soap bubble in r2 uniquely minimizes perimeter . Pacific journal of mathematics , 159 ( 1 ): 47 - 59 , 1993 . [6] M. Hutchings , F. Morgan , M. Ritor´e, and A. Ros . Proof of the double bubble conjecture . Annals of Mathematics , 155 ( 2 ): 459 - 489 , 2002 . [7] W. Karush . Minima of functions of several variables with inequalities as side conditions . Master thesis , University of Chicago, 1939 . [8] E. Milman and J. Neeman. The gaussian double-bubble conjecture . arXiv preprint arXiv:1801.09296 , 2018 . [9] F. Morgan , C. French , and S. Greenleaf . Wulff clusters in r2 . The Journal of Geometric Analysis , 8 ( 1 ): 97 , 1998 . [10] J. Taylor . Existence and structure of solutions to a class of nonelliptic variational problems . In Symposia Mathematica, volume 14 , pages 499 - 508 , 1974 . [11] J. Taylor . Unique structure of solutions to a class of nonelliptic variational problems . In Proc. Symp. Pure Math. AMS , volume 27 , pages 419 - 427 , 1975 . [12] G. Wul . Zur frage der geschwindigkeit des wachstums und der auflosung der kristall achen . Z. Kristallogr , 34 : 449 - 530 , 1901 . ( Eviatar B. Procaccia2) Faculty of Industrial Engineering and Management, Technion - Israel Institute of Technology and Department of mathematics, Texas A &M University URL: http://www.math.tamu.edu/~procaccia E-mail address: eviatarp@technion.ac.il