Research log for issue #62. Use the centers of disks of diameter one. A packing of these disks in an equilateral triangle of side is a set of centers at pairwise distance at least one in the inward equilateral triangle of side .
For an integer put . The classical triangular arrangement packs centers in side . The conjecture says that side is still required for centers.
Literature
- Oler (1961) proves optimality for the complete triangular numbers. Folkman and Graham (1969) give a triangulation proof of Oler's Euclidean inequality.
- Melissen (1993) studies the exact small packings and the equality cases of the area-and-perimeter bound. Payan (1997) proves the 14-disk case and states that his argument can be extended to 20 disks. Graham and Lubachevsky report further packings and conjectures.
- The problem is listed in Research Problems in Discrete Geometry and in the Wikipedia article on equilateral-triangle packing.
- Sahir's 2025 proof claim uses the planar hexagonal density to argue that a smaller triangle cannot contain disks. Even granting that density as an upper bound for the finite container, for disks of radius one it gives only for the outer side . The desired side satisfies . Thus the required finite lower bound does not follow. A separate posted claim assumes the necessary correspondence between numbers of disks touching the sides and triangular arrangements, without establishing it for an arbitrary packing. Neither argument establishes the unrestricted conjecture.
A directional bound for the hull boundary
Lemma. In an equilateral triangle of side , a convex polygon with successive vertices at distance at least one has at most boundary vertices. Points on a straight side of the polygon may be counted as successive vertices.
Orient the polygon and the containing triangle counterclockwise. Assign each directed edge of the polygon to the closest of the three directed side directions of the triangle. Fix a half-open convention at the ties. The assigned edges for any one side are a single consecutive chain, because the edge directions of a convex polygon are cyclically nondecreasing.
Rotate coordinates so the assigned side is the bottom side, from to . The edge directions of this chain lie between and . In particular and for each edge. Squaring the two sides gives
Along the chain the height first decreases and then increases, since the directions are nondecreasing. If the chain starts at and ends at , with minimum height , its total variation in height is . Its length is at most
The middle inequality uses the two sloping sides of the containing triangle at the endpoints. Thus there are at most edges in the chain. The three chains partition the boundary edges, proving the lemma.
All lattice packings and a condition on unrestricted counterexamples
Proposition. The Erdős–Oler lower bound holds, for every , for centers lying in a translate of any lattice whose minimum nonzero vector length is at least one. The lattice orientation and shape are arbitrary. Equality is attained in the class of such packings by the aligned triangular lattice.
In fact, let be any noncollinear subset of such a lattice in an equilateral triangle of side , let be its convex hull, and triangulate with vertex set . If points of are on the hull boundary, Euler's formula gives triangular faces. Each nondegenerate lattice triangle has area at least half the lattice covolume, which is at least . To see the latter bound, take a shortest vector and extend it to a basis . The shortest vector is primitive in the lattice. Reduce by multiples of so its projection onto has magnitude at most . Since , the covolume is at least .
Consequently, by the directional bound,
For the right side is strictly below , so the integer count is at most . A collinear packing has at most centers and also meets the desired bound when . The aligned triangular lattice attains side after deleting one site.
The same proof applies to any packing with a triangulation whose faces all have area at least . Conversely, suppose an unrestricted counterexample has centers in a triangle of side . It has at least triangular faces in any triangulation, since . Its hull has area strictly less than . Therefore every triangulation has at least two faces of area less than : a single deficient face could contribute a deficit no greater than .
In a Delaunay triangulation, each of these faces has circumradius greater than one. Indeed, if a triangle has all sides at least one and circumradius at most one, each angle is at least by the sine rule. Its largest angle lies in . The two sides incident to this angle then give area at least . Each of the deficient Delaunay faces also has circumcenter outside the containing triangle. Otherwise the empty circumdisk would allow adding that circumcenter to the packing, contrary to Oler's bound for centers.
The boundary-face analysis shows that a counterexample can be assumed to occupy all three corners. Its exterior Delaunay faces lead to individual boundary edges, and the total normalised area of the deficient faces is less than . The possibility of several very thin deficient faces remains.
Local rigidity after deleting any site
Let be the aligned triangular lattice patch of side , and remove any one site . The packing is infinitesimally rigid inside the fixed containing triangle. In fact, if another packing in the same triangle can be put in correspondence with these sites so that each displacement has norm less than , all displacements are zero. This concerns perturbations of a single-deletion configuration. A global comparison with all possible packings is still needed.
Balancing contact forces
Start with the full lattice patch. Give weight one to every unit contact. At a site let the force be the sum of the weighted unit vectors from to the sites that touch it. The force is zero at interior sites. At a side site it is times the inward unit normal. At a corner it is times the sum of the two inward unit normals.
After deleting , double the weight of each contact joining consecutive surviving neighbors in the unit ring about . If a neighbor has its two adjacent ring neighbors , the added force there is . This replaces the original force of the contact with exactly.
If is on a side or at a corner, the surviving part of the ring is a path. At either endpoint , the net change in force after the deletion and doubling is the one adjacent unit vector in the ring. Write that vector as , with the inward unit normal to the wall along . From follow that wall in direction to the next corner. Reduce the weights of all contacts along this segment from one to one half. If the segment has positive length, this removes the tangential component at its starting point, leaves the net force unchanged at intervening points, and adds to the force at its far corner. If the segment has length zero, the force change is already at that corner.
The two endpoints lead to different corners, so each remaining corner receives at most one tangential adjustment. In terms of the two inward unit normals at the far corner, adding reduces their coefficients by and , in some order. Thus at every surviving boundary site the modified force is still a positive sum of all its active inward normals, with each coefficient at least . Every contact weight is between one half and two.
Infinitesimal motions
Give each surviving site a velocity that is feasible to first order. An edge requires . At every active wall with inward normal , we require . Sum the edge derivatives with the modified weights. By the force balance, the sum is the negative of the sum of inward wall derivatives with their positive coefficients. Hence all edge derivatives and all wall derivatives vanish.
At a remaining corner both wall projections vanish, so the velocity is zero. Each connected part of a side after one deletion contains a remaining corner. Along that part, the side's unit contacts carry the zero tangential projection from the corner to every site, and the zero wall derivative gives the other projection. All boundary velocities are zero.
At an interior site, two of the three lattice-direction rows through the site avoid . Indeed, express the sites by three integer barycentric coordinates summing to . A row in one direction fixes one coordinate, and two distinct sites cannot agree in two coordinates. The velocities along each intact row have equal projections into its direction, by the contact derivatives, and the endpoints are on the boundary. Thus the two corresponding projections at the interior site are zero. The directions are independent, so this velocity is zero as well.
A displacement bound
For an actual packing inside the same triangle, let , now with the displacement of the point matched to site . Put for each unit contact. The condition on the new length gives , and all inward wall projections of displaced sites are nonnegative. There are at most contacts. The sum of the negative parts of the weighted edge derivatives is at most . The force identity bounds the sum of the positive parts of the weighted derivatives and the total weighted wall projection by this same quantity. In particular,
At a corner the two projections imply . Along every side component, the first inequality bounds the variation of the tangential projection from its remaining corner. Together with the wall projection this gives at each boundary site. At any interior site, the two intact rows give two directional projections each at most in absolute value. For unit directions meeting at or , a vector's norm is at most twice the sum of its two absolute projections. Consequently . If then , as asserted.
A separation radius independent of the side length
The local rigidity radius can be chosen uniformly in . There is an absolute such that a packing in the same triangle matched to a single-deletion triangular arrangement with maximum displacement less than must coincide with that arrangement. The proof combines the contact stress with a discrete strain inequality.
Extending through the deleted site
Use the standard small equilateral faces of side one in the full lattice patch. For a vector-valued displacement write for the sum of the squares of its projected contact differences and of all its normal projections on active walls. In this definition use only the surviving sites and their contacts.
We first extend to the missing site so the same strain energy on the full patch is at most with an absolute . It suffices to take , since the explicit displacement bound above treats smaller . In integer barycentric coordinates, write the deleted site as with . Choose integers with . This is possible since the sum of the lower bounds is at most six and the sum of the upper bounds is . The sites with coordinates at least in each position form an aligned triangle of side six. The hole has local coordinates . The small triangle contains every contact incident to the hole in the full triangle. Each local side for which the corresponding is a wall of the full triangle, since in that case .
On the small patch, use the energy of the surviving contacts and the normals on those physical walls. Consider its kernel. Zero projected difference on each edge of a unit triangle makes the displacement on it infinitesimally rigid. A common edge forces the same rigid motion on two neighboring triangles. This follows because an infinitesimal rigid motion in the plane is a translation plus a scalar multiple of rotation by a right angle, and two point values determine it.
If at most one local coordinate is at most one, the union of the removed unit triangles meets at most one side. The other unit triangles are connected by common edges and contain every surviving vertex. For the connectivity, the removed triangles form the hexagon of unit neighbors of the hole, clipped by the containing triangle in the boundary case. Its complement has connected interior when it meets at most one outer side. Each surviving interior vertex belongs to six original unit triangles, and each noncorner boundary vertex belongs to three. At most two incident triangles can be removed at any one surviving vertex. No corner is a neighbor of the hole in this case. Thus the kernel displacement is one rigid motion on the patch, and its value at the hole gives zero projected difference on the restored contacts. If the hole is on a physical wall, this rigid motion already has zero normal projection at two surviving sites on that wall, so it also has zero normal projection at the hole.
The cases with two local coordinates at most one are, up to permuting the coordinates, , and . For the deleted site is the corner. The remaining unit triangles form the truncated triangular patch and are connected by edges. The two physical walls through the corner constrain the resulting rigid motion to zero.
For let be the nearby corner and let be the two unit directions along its sides, with the hole at . The corner is the only surviving site in no remaining unit triangle. The bulk triangles form a connected patch meeting both physical walls in runs of at least two sites. These wall projections force its rigid motion to vanish. The corner has two wall projections and is fixed. For the hole is at . The closed hexagon of removed triangles meets the two sides of between distances one and two on each side. It leaves the separate unit triangle with vertices and the connected bulk patch. Each component has contacts with both physical walls on intervals, giving zero motion on each. In these three cases the zero extension at the hole works for the kernel.
There are finitely many choices of the hole in the side-six patch. In each pattern choose a linear extension at the hole that agrees with the indicated one on the kernel of the local strain energy. For example, extend its orthogonal projection to that kernel and set the complementary part to zero at the hole. The resulting projected differences on the new contacts and wall projections at the hole form a linear map vanishing on the kernel. Its squared norm is therefore at most a constant times the local strain energy by equivalence of norms on the finite-dimensional quotient. Taking the maximum constant gives the claimed extension estimate uniformly.
From strain to full displacement differences
Interpolate the extended displacements affinely on each small equilateral face, obtaining a continuous function on the triangle . Korn's inequality on the unit triangle, followed by dilation, gives
Here is the boundary step in this form of the inequality. On the unit triangle, the usual Korn inequality modulo infinitesimal rigid motions bounds in for an appropriate rigid motion . The trace estimate controls the boundary normal projection of . A rigid motion with zero normal projection along the triangle boundary must be zero, since it would be tangent along two nonparallel line segments. Its norm is therefore controlled by its normal trace by finite dimensionality. The factor comes from scaling boundary length. See Ayuso de Dios, Brezzi, Marini, Xu and Zikatanov, §2 for the normal-trace argument in Korn inequalities.
On each unit face the three directional projections along its edges determine the symmetric part of the gradient. Thus the integral of is bounded by a constant times the sum of the squared contact projections. The boundary integral is bounded by the sum of the squared normal projections at the lattice sites by linear interpolation. The integral of controls the sum of squares of the full vector differences across the surviving contacts. By the extension estimate,
with a constant independent of and of the deleted site.
For an actual packing in the same triangle we have , and the wall projections are nonnegative. The contact stress implies that the sum of absolute contact projections and the sum of nonnegative wall projections are both . If , each projection has absolute value at most . Hence . The two inequalities force when is below an absolute constant. Then all the contact and wall projections vanish, and the earlier rigidity argument gives .