{
  "id": "kepler-conjecture",
  "title": "Kepler Conjecture",
  "lang": "en",
  "source_lang": "en",
  "summary": "The Kepler conjecture, now a theorem, states that no packing of congruent spheres in three-dimensional Euclidean space exceeds the density of close packing.",
  "category": "mathematics",
  "subcategory": "geometry",
  "type": "concept",
  "aliases": [
    "Kepler's conjecture",
    "Kepler sphere-packing conjecture"
  ],
  "generated_at": "2026-11-05",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/kepler-conjecture/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/kepler-conjecture/",
      "json": "https://aiwiki.page/zh/kepler-conjecture.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "The Kepler conjecture states that the greatest possible density of a packing of congruent spheres in three-dimensional Euclidean space is\n\n\\[\\frac{\\pi}{\\sqrt{18}}=\\frac{\\pi}{3\\sqrt{2}}\\approx 0.7404804897.\\]\n\nThis density is achieved by familiar arrangements of stacked balls, including face-centered cubic and hexagonal close packing. Proposed by Johannes Kepler in 1611, the conjecture remained unresolved for nearly four centuries because proving that a particular arrangement is efficient is much easier than excluding every possible irregular arrangement.\n\nThomas Hales, working with Samuel Ferguson, announced a computer-assisted proof in 1998. Major accounts of the proof appeared in 2005 and 2006. The subsequent Flyspeck project produced a formally verified proof, completed in 2014 and published in 2017. Although its historical name retains the word “conjecture,” the statement is now an established theorem. Its resolution is important both in discrete geometry and in the development of computer-assisted and machine-checked mathematical proof."
    },
    {
      "heading": "Mathematical statement and the meaning of density",
      "text": "A sphere packing is an arrangement of solid balls whose interiors do not overlap. For congruent balls of radius \\(r\\), this condition is equivalent to requiring that the distance between any two distinct centers be at least \\(2r\\). Tangency is permitted. Because multiplying every length by the same positive factor changes neither overlap nor the proportion of occupied space, the problem can be normalized to unit balls.\n\nFor an infinite packing, density requires a limiting definition: both the total volume of the balls and the volume of space are infinite. Let \\(P\\) denote the union of the packed unit balls, and let \\(B_R\\) be a ball of radius \\(R\\) centered at the origin. One standard definition of upper density is\n\n\\[\\overline{\\delta}(P)= \\limsup_{R\\to\\infty} \\frac{\\operatorname{vol}(P\\cap B_R)} {\\operatorname{vol}(B_R)}.\\]\n\nThe upper limit accommodates arrangements whose occupied fraction does not converge as the observation window grows. Volume here has its usual meaning in three-dimensional measure theory. Equivalent formulations can count centers in large regions, with suitable attention to balls crossing their boundaries.\n\nThe theorem asserts that every such packing satisfies\n\n\\[\\overline{\\delta}(P)\\leq \\frac{\\pi}{\\sqrt{18}},\\]\n\nand that equality is attained. The upper bound is the difficult part. Constructing an arrangement of this density gives only a lower bound for the optimum; it does not establish that a more complicated arrangement cannot do better. In particular, a proof restricted to periodic patterns would not automatically settle the unrestricted problem.\n\nThe distinction between finite and infinite packing is essential. A small container may favor an arrangement that would be inefficient in bulk, because its walls reward certain contacts or force empty spaces. Conversely, cutting a finite cluster from an infinite close packing introduces a surface layer whose occupied fraction differs from the bulk value. For regular expanding windows, such boundary contributions are of lower order than volume: surface-scale terms grow like \\(R^2\\), whereas volume grows like \\(R^3\\).\n\nThe theorem does not assert that every maximizing packing has one particular global pattern. It identifies the optimal density, not a unique arrangement. It also concerns equal spheres in flat, unbounded, three-dimensional space. Unequal radii, curved spaces, finite containers, and nonspherical particles lead to different optimization problems, even when their physical appearance resembles a stack of balls."
    },
    {
      "heading": "Close-packed layers and the optimal constructions",
      "text": "The optimal density can be calculated directly from face-centered cubic packing, usually abbreviated FCC. A conventional cubic cell has a sphere center at each corner and at the center of each face. Counting the fractions belonging to one cell gives four spheres: the eight corner spheres contribute one in total, and the six face-centered spheres contribute three. If the sphere radius is \\(r\\), contact along a face diagonal requires\n\n\\[a\\sqrt{2}=4r,\\]\n\nwhere \\(a\\) is the cube’s edge length. Consequently,\n\n\\[\\delta_{\\mathrm{FCC}} =\\frac{4(4\\pi r^3/3)}{(2\\sqrt{2}\\,r)^3} =\\frac{\\pi}{3\\sqrt{2}}.\\]\n\nA second description starts with a triangular layer of touching spheres. Each center has six nearest neighbors in that layer. A second layer can be placed in one of two families of triangular hollows above it. For unit spheres, the horizontal distance from the center of a hollow to each of its three supporting centers is \\(2/\\sqrt{3}\\). The separation between consecutive layers is therefore\n\n\\[h=\\sqrt{2^2-\\left(\\frac{2}{\\sqrt{3}}\\right)^2} =\\sqrt{\\frac{8}{3}}.\\]\n\nThe triangular lattice allocates area \\(2\\sqrt{3}\\) to each center, so the corresponding volume per sphere is \\(2\\sqrt{3}\\,h=4\\sqrt{2}\\), again giving the optimal density.\n\nThe possible layer positions are conventionally labeled \\(A\\), \\(B\\), and \\(C\\). Adjacent layers cannot have the same label, but after layers \\(A\\) and \\(B\\), the third layer may occupy either an \\(A\\) position or a \\(C\\) position. Repeating \\(ABCABC\\ldots\\) produces FCC packing. Repeating \\(ABAB\\ldots\\) produces hexagonal close packing, or HCP. These arrangements have different symmetries and different longer-range neighborhoods, but the same volume per sphere.\n\nMore generally, arbitrary stacking sequences with unequal adjacent labels give Barlow packings. Some are periodic and others are not. Their common density follows from the identical spacing and population of every layer, not from a shared three-dimensional period. Thus nonperiodicity alone is no obstacle to attaining the optimum.\n\nIn both FCC and HCP, each sphere touches twelve others: six within its layer, three above, and three below. The twelve contacts help explain their mechanical and crystallographic importance, but contact number is not itself a density proof. Different arrangements may have the same number of contacts while distributing empty space differently. Nor does optimality force exact order everywhere: removing finitely many spheres changes local structure without changing asymptotic density. The density statement is therefore less restrictive than a complete classification of all arrangements attaining it."
    },
    {
      "heading": "From Kepler to the twentieth-century problem",
      "text": "Kepler discussed the packing problem in his 1611 work Strena seu de nive sexangula, commonly translated as The Six-Cornered Snowflake. His consideration of closely packed spheres formed part of a broader investigation of shape and order in nature. The triangular arrangement of spheres in a layer, and the way such layers stack, provided a geometrical model of dense organization. The setting also belonged to an early modern tradition of questions about stacked cannonballs, including problems associated with Thomas Harriot.\n\nA decisive restricted result came from Carl Friedrich Gauss in 1831. Gauss proved that FCC gives the greatest density among lattice packings in three dimensions. In a lattice packing, every center is an integer linear combination of three independent vectors. This algebraic regularity sharply reduces the range of possibilities. An arbitrary packing need not have any translational symmetry, so the lattice theorem left the central difficulty untouched.\n\nThe two-dimensional analogue was more accessible. Equal circles attain maximum planar density\n\n\\[\\frac{\\pi}{\\sqrt{12}}\\]\n\nin the triangular arrangement. Axel Thue developed proofs around the turn of the twentieth century, and László Fejes Tóth later supplied a systematic geometric treatment. The planar case encouraged approaches based on partitioning space into cells, but the corresponding three-dimensional cells have much greater combinatorial variety. Methods that control polygons in the plane do not transfer directly to polyhedra.\n\nIn 1900, David Hilbert included sphere-packing questions within his eighteenth problem. During the twentieth century, researchers obtained progressively stronger upper bounds without reaching the conjectured value. C. A. Rogers established an influential bound of approximately \\(0.7797\\), using local simplex geometry. Such results narrowed the numerical gap but also exposed why a sharp proof could not rely on a simple estimate applied independently to every small piece of space.\n\nFejes Tóth advocated reducing the problem to a finite examination of local configurations and anticipated a role for computation. Wu-Yi Hsiang proposed a proof in the early 1990s, but it was not accepted as a resolution because its argument did not establish the required bounds to the satisfaction of specialists. Hales developed a different program based on local geometric inequalities, combinatorial classification, and rigorous computation. His collaboration with Ferguson culminated in the 1998 announcement.\n\nPublication followed an unusually extensive review process. Hales’s main account appeared in the Annals of Mathematics in 2005, while detailed supporting papers appeared in Discrete & Computational Geometry in 2006. The subsequent formal verification addressed the difficulty of checking the large computational component through ordinary mathematical refereeing alone."
    },
    {
      "heading": "Local geometry and the obstruction to simple proofs",
      "text": "A natural way to study density is to assign a region of space to each sphere center. The Voronoi cell of a center consists of all points at least as close to that center as to any other. These cells partition space, apart from shared boundaries. For unit spheres, each cell contains its associated sphere, and in a periodic packing the density is the sphere volume divided by the average cell volume.\n\nIn FCC packing, each Voronoi cell is a rhombic dodecahedron of volume \\(4\\sqrt{2}\\). This suggests trying to prove that every unit-sphere Voronoi cell has at least that volume. Such an argument cannot work: some admissible local arrangements have smaller cells. A prominent example is a regular dodecahedron circumscribed about a unit sphere, which has smaller volume than the FCC cell. The corresponding surrounding sphere centers can satisfy the nonoverlap constraints.\n\nThe obstruction is global compatibility. A locally economical cell need not fit together with copies of itself to partition all space. Regular dodecahedra do not tile Euclidean three-space. Their favorable local volume therefore cannot simply be repeated everywhere. The separate dodecahedral conjecture, proved by Hales and Sean McLaughlin, concerns the minimum possible volume of an individual Voronoi cell. Its distinction from Kepler’s conjecture illustrates why minimizing one cell and minimizing an average are different tasks.\n\nAnother approach uses Delaunay decompositions, which connect neighboring centers and divide space into tetrahedra and, when degeneracies occur, more general cells. Four mutually touching unit spheres form a regular tetrahedron of edge length \\(2\\). The proportion of that tetrahedron occupied by the spherical sectors at its vertices is approximately \\(0.7797\\), exceeding the global optimum. This does not contradict Kepler’s theorem because regular tetrahedra cannot fill space without gaps.\n\nIndeed, the internal dihedral angle of a regular tetrahedron is\n\n\\[\\arccos(1/3)\\approx70.53^\\circ.\\]\n\nFive such angles add to less than \\(360^\\circ\\), while six add to more. An arrangement cannot surround every edge with regular tetrahedra and remain a gap-free Euclidean tiling. Local density estimates that ignore this incompatibility are therefore too generous.\n\nA successful proof must account for how neighboring pieces constrain one another. It can allow some regions to be unusually dense, provided the resulting inequalities force compensation elsewhere. This is the conceptual reason for introducing carefully designed local scores rather than attempting to bound every raw cell density by the final packing constant."
    },
    {
      "heading": "Reduction to finitely many configurations",
      "text": "Hales’s approach converts the global density problem into local inequalities around sphere centers. One first restricts attention to saturated packings: arrangements in which no additional unit sphere can be inserted. Any packing can be enlarged to a saturated one without removing existing spheres. In such a packing, every point lies within distance less than \\(2\\) of some center; otherwise another unit sphere could be added there. This gives control over the extent of relevant local regions.\n\nThe original proof assigned a score to each local configuration, called a decomposition star. Its construction combined information from Voronoi regions and carefully selected simplices. The score was not merely the density of a small cluster. It was an accounting device designed so that, after contributions from many centers were combined, a universal score bound implied the required global density bound.\n\nThe crucial estimate said that no allowable local configuration could score better than the close-packed reference configurations. To prove it, the local geometry was encoded by a graph. Nearby centers determine directions from a distinguished center; suitable connections between these directions produce a planar combinatorial structure. Restrictions on distances, angles, and nonoverlap then constrain the possible vertices, edges, and faces.\n\nThe use of graph theory separates two kinds of difficulty. A graph records which neighbors are related, while continuous variables record their actual distances and angles. A finite list of combinatorial types can therefore represent infinitely many geometric configurations. Exhaustiveness is essential: finding many plausible types is not enough, because an omitted type might contain a counterexample.\n\nHales’s program identified a finite class of relevant “tame” plane graphs. Each surviving graph then had to be shown incompatible with a score exceeding the target. Many cases were eliminated by relatively coarse inequalities. More difficult cases required stronger estimates, subdivision of parameter ranges, or separate geometric arguments. Ferguson’s work dealt with a particularly difficult configuration known as the pentahedral prism.\n\nThe later Flyspeck proof reorganized substantial parts of this machinery, including the local inequalities and classification. Consequently, the formally verified argument should not be understood as a literal transcription of every step of the 1998 proof. The shared strategy is the passage from an unrestricted infinite arrangement to bounded local data, then to finitely many combinatorial cases, and finally to rigorous inequalities. The finite reduction makes computation possible, but its correctness is itself a central mathematical achievement."
    },
    {
      "heading": "Linear programming, interval bounds, and rigorous computation",
      "text": "The computer-assisted proof used computation to exclude entire families of configurations, not merely to search for dense examples. This distinction is fundamental. A numerical search may discover FCC or HCP repeatedly without ruling out a better arrangement elsewhere. To prove optimality, every permissible competitor must be covered by a logically valid exclusion or bound.\n\nOne major tool was linear programming. Geometric quantities such as edge lengths, angles, solid angles, and local scores satisfy numerous inequalities. Replacing nonlinear relations with valid linear bounds produces a relaxed feasible region. If the maximum score over this larger region is already below the target, then the original geometric family is also excluded. The relaxation can be deliberately generous, provided it remains rigorous.\n\nIf a linear relaxation is inconclusive, the variables can be split into smaller ranges and the bounds strengthened. This resembles branch-and-bound in mathematical optimization: partition a problem, bound each part, and discard parts that cannot violate the desired conclusion. The success of the method depends on finding estimates that are computationally manageable without allowing too much geometric freedom.\n\nNonlinear inequalities required another kind of control. In interval arithmetic, a quantity is represented by an interval guaranteed to contain its exact value. Arithmetic operations produce enclosing intervals rather than single approximate numbers. For example, if an entire admissible box of parameters yields a strictly negative upper bound for the difference between a score and its permitted maximum, every configuration in that box is certified at once.\n\nDerivative estimates and Taylor expansions help keep these bounds tight. For a sufficiently smooth function, values and derivative bounds near the center of a box constrain the function throughout the box. When an estimate is too weak, subdivision reduces the uncertainty. These techniques must also handle domain restrictions, denominators, square roots, and other operations whose validity depends on signs or parameter ranges.\n\nOrdinary floating-point arithmetic cannot by itself supply this guarantee. Rounding can shift a value across an inequality threshold. Rigorous numerical work therefore needs controlled rounding, exact arithmetic, formally justified error bounds, or equivalent certification. A displayed decimal approximation is evidence of a computation, not automatically evidence of the theorem it is intended to support.\n\nThe Kepler proof combined human-designed geometry with extensive machine checking. The theoretical argument specified what had to be verified; the computation handled the volume of finite cases and numerical inequalities. Its reliability depended on both sides: correct programs cannot repair an incomplete geometric reduction, and a correct reduction cannot justify unchecked numerical output."
    },
    {
      "heading": "Flyspeck and formal verification",
      "text": "The scale of the original computation made traditional review difficult. Referees could assess the architecture of the proof and inspect many arguments, but independently checking every computational detail was a different undertaking from reviewing a conventional journal article. Hales initiated Flyspeck in 2003 to reconstruct the proof within systems capable of checking individual logical inferences. The name abbreviates “Formal Proof of Kepler.”\n\nA formal proof represents statements and deductions in a precisely specified logical framework. A proof assistant checks that each derived statement follows from previously established results through permitted inference rules. The goal is not to have a machine judge whether prose is persuasive, but to supply an explicit derivation of the mathematical claim. This required formalizing much of the surrounding geometry, analysis, and combinatorics.\n\nFlyspeck used HOL Light for the main mathematical development and Isabelle for the enumeration of tame graphs. These systems played complementary roles. The geometric argument needed a verified classification of combinatorial possibilities, while the classification needed to match the definitions and hypotheses used by the geometric proof. The project also developed formally checked methods for the nonlinear inequalities.\n\nCompletion was announced in August 2014. A multi-author account, A Formal Proof of the Kepler Conjecture, appeared in 2017. The result established the density theorem through machine-checked mathematics, including the computationally intensive components that had made the earlier proof difficult to audit. It also required revisions and simplifications of the mathematical argument to make formal verification practical.\n\nFormal verification narrows, rather than abolishes, the basis of trust. A small logical kernel checks proof steps, reducing dependence on large, specialized search programs. A complicated program may propose a result, but its output must be converted into a derivation or certificate that the trusted checking machinery accepts. As with other computer-based verification, assumptions remain about the logic, the implementation, and the execution environment.\n\nThe human role also remains substantial. Definitions must express the intended packing problem; formal lemmas must connect to the density statement; and the mathematical strategy must be invented. Flyspeck did not replace those activities with automatic discovery. Its significance was to show that a lengthy proof combining continuous geometry, discrete enumeration, and numerical bounds could be brought under systematic logical checking, rather than leaving its computational parts as externally trusted calculations."
    },
    {
      "heading": "Related packing theorems and dimensional contrasts",
      "text": "The Kepler conjecture belongs to the broader sphere-packing problem in \\(n\\)-dimensional Euclidean space. If the packed balls have radius \\(r\\), their volume is\n\n\\[V_n(r)=\\frac{\\pi^{n/2}}{\\Gamma(n/2+1)}r^n,\\]\n\nwhere \\(\\Gamma\\) is the gamma function. Scale invariance persists in every dimension, but the geometry of optimal arrangements changes markedly. Three-dimensional intuition is often a poor guide to higher-dimensional behavior.\n\nIn one dimension, congruent intervals can tile the line, giving density \\(1\\). In two dimensions, the triangular circle packing has optimal density \\(\\pi/\\sqrt{12}\\). Three dimensions are more difficult partly because locally favorable tetrahedral structures cannot fit together uniformly. The particular local obstructions and compensating inequalities used in the Kepler proof are therefore closely tied to dimension three.\n\nAnother neighboring problem is the kissing number problem: how many nonoverlapping congruent spheres can touch one central sphere? In three dimensions, the answer is twelve, established by Kurt Schütte and Bartel Leendert van der Waerden in 1953. This is compatible with the twelve contacts in FCC and HCP, but it does not prove their density is optimal. A contact bound controls one neighborhood, whereas density depends on how all neighborhoods fit together.\n\nExceptional higher-dimensional cases have yielded exact solutions by different methods. In 2016, Maryna Viazovska proved the optimality of the \\(E_8\\) lattice packing in dimension eight. Soon afterward, Henry Cohn, Abhinav Kumar, Stephen D. Miller, Danylo Radchenko, and Viazovska established the optimality of the Leech lattice packing in dimension twenty-four. These packings have densities \\(\\pi^4/384\\) and \\(\\pi^{12}/12!\\), respectively.\n\nThose proofs used sharp auxiliary functions within the Cohn–Elkies linear-programming framework. Conditions on a function and its Fourier transform imply upper bounds on packing density. Despite the shared phrase “linear programming,” this analytic method is different from the finite-variable linear programs used to eliminate local configurations in the Kepler proof.\n\nThe higher-dimensional results also reveal links with information theory. A collection of well-separated points can model distinguishable signals: sufficient separation makes them less vulnerable to confusion under bounded noise. Lattices and sphere packings consequently connect geometric optimization with coding theory. The exact three-dimensional theorem is one particularly famous instance of a much wider family of questions, most of which do not have known exact answers."
    },
    {
      "heading": "Physical relevance and the limits of the theorem",
      "text": "Close packing is important in crystallography because FCC and HCP describe common arrangements of atomic positions in idealized crystal structures. Aluminum and copper have FCC structures under ordinary conditions, while magnesium has an HCP structure. The hard-sphere picture explains their coordination geometry and occupied-volume fraction, but it is only a model: atoms are not rigid balls with sharp boundaries.\n\nThe theorem also provides a reference point for granular materials. Experiments and simulations often produce disordered sphere assemblies with densities near \\(0.64\\), lower than the crystalline maximum of approximately \\(0.74048\\). The exact density depends on preparation, including compression, agitation, friction, and permitted ordering. The expression “random close packing” does not specify a unique mathematical optimization problem unless randomness and mechanical constraints are defined.\n\nDensity and jamming are distinct properties. A packing can resist particular particle motions without achieving the largest possible occupied fraction. Conversely, the density theorem alone does not determine mechanical stability under forces, friction, or boundary constraints. Different notions of jamming specify whether individual particles, finite groups, or the entire packing and its enclosing boundaries may move.\n\nThermal hard-sphere systems introduce another distinction. At nonzero temperature, equilibrium behavior depends on the number of accessible configurations, not simply on the density of a single static arrangement. Hard-sphere freezing and the relative free energies of FCC and HCP are therefore questions in statistical mechanics, not direct consequences of Kepler’s theorem. Equal maximum density does not imply identical behavior at lower densities.\n\nThe equal-radius assumption is especially restrictive. In a close-packed array of spheres of radius \\(R\\), the largest spheres fitting into ideal octahedral holes have radius\n\n\\[r_{\\mathrm{oct}}=(\\sqrt{2}-1)R,\\]\n\nwhile those fitting into tetrahedral holes have radius\n\n\\[r_{\\mathrm{tet}}=\\left(\\sqrt{\\frac{3}{2}}-1\\right)R.\\]\n\nAdding smaller spheres to such gaps can increase the total occupied fraction without moving the large ones. This does not violate the theorem because the resulting balls are no longer congruent.\n\nChanging the particle shape likewise changes the problem. Congruent ellipsoids, tetrahedra, or other bodies can have packing densities exceeding the sphere value. Finite-container problems add further dependence on the number of objects and the container’s dimensions. The Kepler theorem gives an exact bulk limit for one sharply defined geometric setting; quantitative predictions for crystals, powders, or engineered particle mixtures require additional assumptions about shape, interactions, disorder, and boundaries."
    }
  ],
  "links": [
    "euclidean-space",
    "theorem",
    "mathematical-proof",
    "euclidean-distance",
    "measure-theory",
    "carl-friedrich-gauss",
    "david-hilbert",
    "voronoi-diagram",
    "delaunay-triangulation",
    "graph-theory",
    "linear-programming",
    "branch-and-bound",
    "mathematical-optimization",
    "interval-arithmetic",
    "floating-point-arithmetic",
    "formal-proof",
    "gamma-function",
    "fourier-transform",
    "information-theory",
    "statistical-mechanics"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}