Optimal Packings of Equal Circles in a Square

Kari J. Nurmela Department of Computer Science and Engineering Helsinki University of Technology P.O. Box 1100, FIN 02015 HUT, Finland Kari.Nurmela@hut. Patric R. J. sterg rd Department of Mathematics and Computing Science Eindhoven University of Technology P.O. Box 513, NL 5600 MB Eindhoven, The Netherlands Patric.Ostergard@hut.
The problem of nding the maximum radius of n non-overlapping equal circles in a unit square is considered. A computer-aided method for proving global optimality of such packings is presented. This method is based on recent results by De Groot, Monagan, Peikert, and W rtz. As an example, it is shown how the method can be used to get an optimality proof for the case n = 7, which has not earlier been published.

Abstract

1 Introduction
We consider the problem of nding values of rn , the maximum radius of n non-overlapping equal circles in a unit square. An equivalent problem is that of spreading n points in a unit square so as to maximize the least distance between any two points; the maximum is denoted by dn . Circle packings in other geometrical objects such as, for example, equilateral triangles 4] and circles 6] have also been considered recently. For a thorough discussion of solved and unsolved packing problems, see 1, 3]. Although packing problems have a long history, packing circles in a square has been considered in the mathematical literature only since the early 1960s, when the problem was raised by L. Moser 9]. A packing of circles with maximum radius (and the corresponding spread of n points) is called optimal. During the last three decades, optimality proofs for n 20, n = 25, and n = 36 and good packings
Graph Theory, Combinatorics, Algorithms, and Applications, 1996

Kari J. Nurmela and Patric R. J. sterg rd

of other values of n have been published; see 2, 5, 10, 12] and their references. Proofs for most of the optimal packings for 10 n 20 were recently obtained in 2, 12]. These proofs involve extensive use of computers and cannot be checked by hand. In this paper we intend to further develop this technique. Our method has enabled us to extend the number of circles up to 26. Due to lack of space, these results will be published in a forthcoming paper 11]. In Section 2 the computer-aided method used to prove optimality of packings is presented. In Section 3 we show how this method can be used to obtain a proof for the case n = 7; the fact that this case has been proven was reported already in 13] but the proof has never been published. Some other results are also brie y mentioned.

2 A Computer-Aided Method for Optimality Proofs

2.1 Overview of the Method
We consider the problem of spreading n points in a unit square. The method used is based on that in 2], but contains some modi cations. The last step is the least convincing in 2]; we hope that our method will convince the reader that the results are correct, assuming that the computer programs are correct. In the rst step, the square is tiled in such a way that two points of an optimal packing cannot be in the same tile (using a lower bound on dn obtained from the best packing known). The shape of the tiles is irrelevant, we use square or rectangular tiles. We now consider all possible ways of choosing n of the tiles; the number of combinations is reduced by taking symmetries into account. In the second step, the combinations obtained in the rst step are considered, one by one. Areas of the tiles that cannot contain a point are eliminated, nally leading to elimination of most of the combinations. Approaches to eliminate hard cases are also considered. Before the nal step, we know small regions in which points of an optimal packing must reside. In the nal step, we guess an optimal packing and use an elimination procedure to prove that it is optimal.

Optimal Packings of Equal Circles in a Square

2.2 Finding Initial Combinations

First of all, we need a lower bound on dn . This lower bound, dlow , is obtained from the densest known packing of n circles. Moreover, care must be taken that dlow is less than the value obtained in this way, so that the nite precision of the oating point arithmetic will not a ect the results. Now the square is tiled in such a way that the maximum distance between two points in a tile (the diameter of the tile) is smaller than the aforementioned bound. Consequently, no two points in an optimum packing can be in the same tile. We denote the total number of tiles by t. The next step is to consider all possible ways of choosing n out of t tiles. The chosen tiles contain a point and they are called full. Other tiles are called empty. By taking symmetries into account, we need not consider all possible combinations. As the symmetry group of a square, D4 , has order 8 and the symmetry group of a rectangle, D2 , has order 4, these orders give the maximum factors by which the number of cases can be reduced. The exact number of remaining cases can be calculated using Burnside's lemma.

Theorem 1 (Burnside) Let G be a group acting on a set X . For g 2 G let (g) denote the number of elements of X xed by g. Then P 1 the number of orbits of G is jGj g2G (g).
In our case, X is the set of all possible choices of n full tiles, and G is one of the two aforementioned symmetry groups. Now, we try to eliminate combinations that cannot contain a packing with minimum distance at least dlow . We do this by eliminating parts of full tiles until, hopefully, one of them is completely eliminated, in which case we can discard the combination. The area of a full tile that has not yet been eliminated is called active. Only the active area of a tile can contain a point. Initially the whole area of full tiles of a combination is active. This area is reduced by considering the active area of other full cells. Where each tile in 2] is divided into smaller squares (division in two directions), we divide the tiles into rectangles (division in only one direction). The elimination goes as in 2], with the di erence that

2.3 Eliminating Combinations

Kari J. Nurmela and Patric R. J. sterg rd

Figure 1: Eliminating active area. here the rectangles are made smaller and smaller (and completely eliminated), whereas in 2] squares are eliminated. Sometimes a part in the middle of a rectangle can be eliminated; to keep the algorithms and data structures simple, no elimination is done in that case. Figure 1 illustrates this process where the left full tile a is used to eliminate active areas in the other tiles. The shaded parts of tiles b and c can be eliminated, whereas tile d is not a ected. Initially, each tile consists of only one rectangle. When no area can be eliminated any more, rectangles with active area are split into two rectangles. This is repeated until either a full tile has no active area or a prescribed number of splits has been done; in the latter case, the combination is a candidate for containing an optimal packing. Figure 2 illustrates how active area of the optimal combination for n = 7 is eliminated.

2.4 Eliminating Hard Cases

The previous step is sometimes not able to eliminate all combinations that cannot contain a packing with minimum distance at least dlow .

Optimal Packings of Equal Circles in a Square

Figure 2: Obtaining optimal packing for n = 7. This happens especially when there are several good packings close to each other. It is possible to cope with this problem in several ways. One possibility is to continue the procedure in the previous subsection and consider even narrower rectangles. Unfortunately this is a time-consuming approach and we are still not able to eliminate many cases that cannot contain an optimal packing. A better approach following the ideas in 2] is to split the problem into subproblems at a certain stage. This is here done by choosing a full tile and by separately treating all rectangles with active area, one by one (assuming that the rectangle in consideration is the only active area in its full tile). In some cases even more elaborate methods for dividing the problem into subproblems are necessary.

2.5 Proof of Optimality

The approach in the proof of optimality was rst used by Schaer to (manually) prove optimality of the densest packing of 9 circles 13]; see also the papers by Kirchner and Wengerodt 8, 14, 15, 16]. When all but the (possibly more than one) combinations leading to optimal packings have been eliminated, the process of splitting rectangles can be continued to get more and more accurate solutions. However, in this way one will never nd the exact solution but only better and better approximations. After the previous steps of the algorithm we know the (small) regions in which points of an optimal packing can reside. We now guess an optimal packing (that is, which points are at minimum distance from each other) and construct the corresponding system of equations. In many cases (especially when there are strong symmetries in the solution), the system can be solved symbolically. The existence of these packings can be veri ed easily, because the coordinates of

Kari J. Nurmela and Patric R. J. sterg rd

the points can be computed symbolically. In the cases where the exact values cannot be found, we solve the system numerically to get approximate positions of the points of a solution, p0i . In these cases, using the interval Newton/bisection program INTBIS 7], we prove that such a packing really exists with points pi for which jp0i ? pi j < d0 holds for some given small d0 . Next, we draw congruent error squares (circles are used in 2], but in 8, 13, 14, 15, 16] Schaer and others indeed used squares) of side r around the points p0i (which are in the center), such that the active area remaining from the previous steps of the algorithm is within these squares (this requirement must also hold if the squares are drawn around the points pi ). For a point on a side, only a half square is used, and for a point in a corner, a quarter of a square is used. A square that is around a point not on a side must t within the unit square (both with p0i and pi as a center). Finally, an elimination procedure similar to that in the previous step is started. This elimination procedure can be used to prove optimality whenever the following theorem can be applied.
d0 < dn . If the active area around the points p0i can be reduced by the elimination process (with lower bound d ) to t within squares of side qr, with 0 < q < 1 (with the points p0i in the centers), then the guessed packing with the points pi is the best packing within the original error squares. Furthermore, there are no other equally good such packings.
low

Theorem 2 Assume that d

low

+2

Proof. The premises of the theorem state that within a nite number of steps in the computed elimination process, each of the active areas around the points p0i are reduced. Let us denote the process consisting of these rst k eliminations by P10 . The proof of the theorem is structured as follows. In the rst part of the proof we will show that there exists another elimination process, P1 , with speci c properties. The process P1 consists of k eliminations (like P10 ) and it will start with active areas congruent to those of P10 . In P1 the areas are centered around the points pi . We will prove that whenever an active area is reduced in P10 , a congruent area reduction can be performed in P1 . Note that since we will not have to actually compute the elimination process P1 , we can assume exact arithmetic in P1 and use the reductions in P10 as lower bounds for reductions in P1 .

Optimal Packings of Equal Circles in a Square

x
.

y y

Figure 3: First part of proof. In the second part of the proof we will show that the elimination process P1 can be extended to an in nitely long elimination process P where the sides of the error squares tend to zero; the guessed packing is thus the unique best packing within the original error squares. duction. Initially, all corresponding active areas in P10 and P1 are congruent. Assume that this holds at some stage. In the next step, active area around p0a is reduced. We shall now show that active area around pa can be reduced in P1 to get a congruent new active area. The idea is to use a slightly larger value of dlow in P1 , namely dlow2 = dlow + 2d0 < dn . Take any point x0 that is in the area to be eliminated (around 0a ). The point x0 is then at distance less than dlow from all points p in another active area around p0b . Now, a point x that is in the active area around pa and congruent to x0 is at distance less than dlow + 2d0 = dlow2 from all points in the active area around pb (as jp0a ? paj < d0 and jp0b ? pbj < d0) and can be eliminated in procedure P1 . See Figure 3. This concludes the induction. nal part of the proof we construct an elimination process P where the active areas will converge to the packing with points pi . The active area around pi is reduced from active squares of side r to t within squares of side qr by P1 . We show that the existence of process Pi implies that there is another elimination process, Pi+1 ,

First part of the proof: existence of P . The proof is by in1

Second part of the proof: extension of P . In the second and

1

Kari J. Nurmela and Patric R. J. sterg rd

which starts with error squares resulting from Pi (sides equal to qi r) and reduces the active areas to t within squares with sides equal to q(i+1) r by using the same number of steps as in Pi . Clearly, by concatenating all the elimination processes P1 , P2 , : : : we obtain an elimination procedure where the sides of the error squares tend to zero and thus the points pi form the best possible packing within the original error squares. Furthermore, the best packing is unique; should it not be the case, there would not be an elimination process where the square sides converge to zero. Now we can use for dlow the exact value of d for the packing with points pi , because the existence of such a packing has been proven separately. We have already proven that P1 performs a required reduction in the active areas. Now assume that Pi reduces the error squares from q(i?1) r to qi r. We show that at each stage all parts of the new elimination procedure Pi+1 (starting from active squares of side qi r) are homothetic with scaling factor q with those of Pi . Initially, this is clearly the case. For induction, assume that this holds at some stage. In the next step, active area around pa is reduced in Pi . We shall now show that active area around pa can be reduced in Pi+1 to get a homothetic new active area. Take any point x0 that is in the area to be eliminated. The point x0 is then at distance less than d from all points in another active area around pb . Take one such point, y0 . Let x (y) be on the line going through pa and x0 (pb and y0 ), such that x = pa + q(x0 ? pa ) (y = pb + q(y0 ? pb )); see Figure 4. Now the distance between x and y is less than d (this is easily calculated; see also 2, 12]), and x is in the area that can be eliminated in procedure Pi+1 . This concludes the proof. 2 To be precise, the maximal error in a calculation due to nite precision of the computer used (e) must also be taken into account (dlow + 2d0 + e < dn ).

3 The Results
We shall here show how the method presented in the previous section can be used to get an optimality proof for the case n = 7. Schaer wrote a manuscript containing an outline of a proof, but it remained unpublished; we ll here that gap in the circle packing literature.

Optimal Packings of Equal Circles in a Square

x y x pa p
b

Figure 4: Second part of proof.

Figure 5: The optimal packing of 7 circles. For the best known packing with n = 7 (depicted in Figure 5 and p which will turn out to be optimal), d = 4 ? 2 3 > 0:5358983. We can then use a 3 3 tiling of a square, as the diameter of the tiles then p ? is 2=3 0:4714045. The total number of combinations is 9 = 36, 7 but by taking symmetries into account this number is reduced to 8. By using the elimination procedure (Section 2.3) and stopping it after 5 rounds, 2 combinations remain: 111110101 and 101111101, where a 1 represents a full tile and a 0 represents an empty tile, starting from the leftmost, lowermost tile and proceeding upwards row after row. The second of these cannot be eliminated by using more rounds instead the ideas described in Section 2.4 can be used. Finally, it is not di cult to show that the rigid structure of 6 circles indeed exists, and the prerequisites of Theorem 2 are ful lled as convergence occurs with error squares of, for example, size 0.01 (and the original procedure is able to eliminate area so that active area ts into these error squares).

10
21

Kari J. Nurmela and Patric R. J. sterg rd

The method has further been used to nd optimal packings for n 26. It is feasible to consider even larger values of n by using an initial phase in which packings in parts of a square are found and combined to reduce the number of combinations to be considered in the main elimination procedure. Due to lack of space these results will be published elsewhere 11]. The approach can with minor modi cations be used for packings in others geometrical objects (such as equilateral triangles and circles).

Acknowledgments
The rst author would like to thank the Emil Aaltonen Foundation and the Vilho, Yrj and Kalle V is l Foundation, and the second author would like to thank the Academy of Finland and the Alfred Kordelin Foundation for nancial support.

References
1] H. T. Croft, K. J. Falconer, and R. K. Guy, Unsolved Problems in Geometry, Springer-Verlag, New York, 1991. 2] C. de Groot, M. Monagan, R. Peikert, and D. W rtz, Packing circles in a square: a review and new results, in P. Kall (ed.), System Modelling and Optimization (Proc. 15th IFIP Conf. Z rich 1991), 45 54, Lecture Notes in Control and Information Sciences, Vol. 180, Springer-Verlag, Berlin, 1992. 3] L. Fejes T th, Lagerungen in der Ebene, auf der Kugel und im Raum, 2nd ed., Springer-Verlag, Berlin, 1972. 4] R. L. Graham and B. D. Lubachevsky, Dense packings of equal disks in an equilateral triangle: from 22 to 34 and beyond, Electron. J. Combin. 2 (1995) A1, 39 pp. (electronic). 5] R. L. Graham and B. D. Lubachevsky, Repeated patterns of dense packings of equal disks in a square, Electron. J. Combin. 3 (1996) R16, 17 pp. (electronic). 6] R. L. Graham, B. D. Lubachevsky, K. J. Nurmela, and P. R. J. sterg rd, Dense packings of congruent circles in a circle, Discrete Math., to appear.

Optimal Packings of Equal Circles in a Square

11

7] R. B. Kearfott and M. Novoa III, Algorithm 681: INTBIS, a portable interval Newton/bisection package, ACM Trans. Math. Software 16 (1990), 152 157. 8] K. Kirchner and G. Wengerodt, Die dichteste Packung von 36 Kreisen in einem Quadrat, Beitr ge Algebra Geom. 25 (1987), 147 159. 9] L. Moser, Problem 24 (corrected), Canad. Math. Bull. 3 (1960), 78. 10] K. J. Nurmela and P. R. J. sterg rd, Packing up to 50 equal circles in a square, Discrete Comput. Geom., to appear. 11] K. J. Nurmela and P. R. J. sterg rd, More optimal packings of equal circles in a square, in preparation. 12] R. Peikert, Dichteste Packungen von gleichen Kreisen in einem Quadrat, Elem. Math. 49 (1994), 16 26. 13] J. Schaer, The densest packing of nine circles in a square, Canad. Math. Bull. 8 (1965), 273 277. 14] G. Wengerodt, Die dichteste Packung von 16 Kreisen in einem Quadrat, Beitr ge Algebra Geom. 16 (1983), 173 190. 15] G. Wengerodt, Die dichteste Packung von 25 Kreisen in einem Quadrat, Ann. Univ. Sci. Budapest. E tv s Sect. Math. 30 (1987), 3 15. 16] G. Wengerodt, Die dichteste Packung von 14 Kreisen in einem Quadrat, Beitr ge Algebra Geom. 25 (1987), 25 46.