WorksMathematics
Improved lower bounds for the kissing number in dimensions 49 to 61
Abstract
We give explicit kissing configurations in R^d for every d = 49, ..., 61. The lower bound is improved over the best previously published value in all thirteen dimensions, by margins from 232 to 159,220 points and by 506,461 points in total. The configurations are obtained from the antipodal block lift of Yang et al. applied to the 52,416,000-point minimal shell of the extremal even unimodular lattice P_48p, using a class of 7,193 minimal lines and, in each dimension, a kissing configuration of R^k (k <= 15) with the largest available number of antipodal pairs. Two ingredients are new. The published class of 7,077 lines is grown to 7,193 by local search in three stages, as a sequence of exchanges between candidate lines and class lines; and the point count, which equals 52,416,000 + 2S + 2t with t the number of antipodal pairs and S the size of the union of t isometric images of the class, is maximised as a maximum-coverage problem by greedy search with iterated local search over pools of 4,000 automorphic images, reaching 92-100% of the trivial bound 7,193t. All results are verified in exact integer arithmetic, and the certificates are distributed with the code. Code and data: https://github.com/dimkadimon/KissingNumbers
Introduction
How many equal spheres can touch a given sphere, all of the same size and without overlapping? This question was put to Isaac Newton by David Gregory in 1694. Newton held that the answer is 12; Gregory suspected 13. Newton was right, but a complete proof took more than three centuries [9]. The quantity is now called the kissing number : the largest number of non-overlapping unit spheres in that can all touch one central unit sphere, an arrangement of points on the unit sphere with pairwise distances at least one.
In the plane the answer is easy to see: six equal circles fit around a seventh, , and the centres form a hexagonal pattern. In 3D space , the centres sitting at the vertices of a cuboctahedron, but here already the proof is delicate, and the visual intuition is misleading — the twelve neighbours leave visible gaps, which for a long time made it unclear whether a thirteenth sphere could be squeezed in. Beyond three dimensions the problem becomes a benchmark for the limits of both geometry and computation: despite much effort, is known exactly only in dimensions 1, 2, 3, 4, 8, 24 [6, 10, 1]. In every other dimension one has to settle for lower bounds, supplied by explicit constructions, and upper bounds. This paper improves the lower bounds in dimensions 49 to 61.
Equivalently, one asks for the largest set of points on the unit sphere with pairwise distances at least one, that is, a code with a fixed minimum separation. In the notation used below, is the largest cardinality of a set with for distinct . The published tables of lower bounds traditionally stop at dimension 48, where the extremal even unimodular lattice of Nebe [7, 8] has 52,416,000 minimal vectors [2], and resume at dimension 64, where the code construction of Edel, Rains and Sloane [4] gives 331,737,984. The intermediate range 49–63 was therefore, until recently, occupied only by the trivial observation that a direct sum with a good low-dimensional configuration gives 52,416,000 + .
That changed with two independent works. Yang et al. [11] introduced an antipodal block lift (their Theorem 2.5) which, applied to the minimal shell of together with a class of 7,077 minimal lines, produces configurations in dimensions 49–55; and the repository [5] evaluated the same family in dimensions 49–61 with a class of 7,069 lines and better tails. The two overlap in 49–55 and disagree in 56–61, where [5] is the only published source. Table 3 lists the best previously published value in each dimension, with its source.
| previous best | source | this work | margin | |||
| 49 | 1 | 1 | 52,430,156 | [11] | 52,430,388 | 232 |
| 50 | 2 | 3 | 52,458,468 | [11] | 52,459,164 | 696 |
| 51 | 3 | 6 | 52,500,930 | [11] | 52,502,324 | 1,394 |
| 52 | 4 | 12 | 52,585,772 | [11] | 52,588,576 | 2,804 |
| 53 | 5 | 20 | 52,698,696 | [11] | 52,703,384 | 4,688 |
| 54 | 6 | 36 | 52,923,774 | [11] | 52,932,310 | 8,536 |
| 55 | 7 | 63 | 53,301,140 | [11] | 53,316,308 | 15,168 |
| 56 | 8 | 120 | 54,085,392 | [5] | 54,117,390 | 31,998 |
| 57 | 9 | 153 | 54,534,176 | [5] | 54,574,792 | 40,616 |
| 58 | 10 | 255 | 55,893,422 | [5] | 55,960,474 | 67,052 |
| 59 | 11 | 302 | 56,505,668 | [5] | 56,583,474 | 77,806 |
| 60 | 12 | 378 | 57,477,123 | [5] | 57,573,374 | 96,251 |
| 61 | 13 | 577 | 59,898,694 | [5] | 60,057,914 | 159,220 |
Table 3. Lower bounds for , . The margin is the difference from the best previously published value; total margin 506,461. In dimensions 49 and 50 the images happen to be pairwise disjoint (), so the ceiling of the construction is attained exactly.
This paper keeps the construction of [11] and changes one ingredient: the choice of isometric images. Because each selected image contributes to the count through the size of the union of the images and not through its own size, the images must be chosen jointly. We formalise this as a maximum-coverage problem, solve it with a greedy pass followed by exact 1-swap local search and iterated local search over pools of up to 4,000 images, and obtain an improvement in every dimension 49–61.
A second lever points in the same direction. Every line of the class is worth points in dimension , so enlarging the class by one line improves all thirteen bounds at once. § 3 describes the searches that grow the published class of 7,077 lines to 7,193 lines, in three stages; the images of § 4 are then re-optimised for the enlarged class.
The lift and the count
We begin by recalling the construction and fixing notation. Let be the published integral Gram matrix of , so that a lattice vector is an integer row and its squared norm is . The minimal shell has 52,416,000 vectors of norm 6 and is antipodal; a line is a pair , represented by the member whose first nonzero coordinate is positive.
A class is a set of minimal lines with
which after normalisation is . The class used here has lines: it is obtained in § 3 from the 7,077-line class of [11] (one of whose elements is a 7,069-line subclass used in [5]) by adding 125 lines and deleting 9.
Theorem 1 (Antipodal block lift; [11 Theorem 2.5]). Let be a kissing configuration and pairwise disjoint subsets with for distinct . Let be a kissing configuration containing distinct antipodal pairs . Then
is a kissing configuration in with .
Apply this with the minimal shell of and with blocks obtained from a class: given isometries of the Gram matrix , assign each line of to the first image in which it occurs. The assigned parts are disjoint as line sets, their union has
and every image is again a class by (1). Taking both signs of each assigned line gives blocks of size satisfying the hypothesis of Theorem 1, and
since a tail with antipodal pairs has points. This is equation (8.8) of [11].
Two consequences shape everything that follows. First, is determined only by the tail and is independent of the lattice: any kissing configuration of with antipodal pairs will do, so the same class family serves every dimension. Second, the images enter (3) only through the single number , which is at most and is smaller whenever two chosen images share a line. The class itself enters linearly: one more line is worth two more points per image, that is points in dimension .
Growing the class
Fix the class problem: find as many minimal lines of as possible with (1). Two lines conflict when ; a class is an independent set in the conflict graph on the minimal lines, and the published class of 7,077 lines is one such set. Enlarging it is the only lever that scales every dimension at once, one added line being worth points in dimension . We use three searches in sequence, each of which exchanges candidate lines for class lines; the sizes grow from 7,077 to 7,082, then to 7,108, and finally to 7,193.
Stage 1: conflict-filtered exchanges. A line that conflicts with few class lines can enter the class if those few conflicts leave. We draw a pool of 20 random isometric images of the class — about distinct lines — and count, for every pool line, its conflicts against the whole class. The distribution is sharply concentrated: over 141,140 pool lines, 57 lines conflict with exactly one class line, 105 with two, 431 with three, 1,478 with four, 3,619 with five, and the median is ten. Only lines with at most five conflicts are kept, which keeps the exchange problem small: if is the set of class lines hit by the kept candidates and an independent set among them, then is again a class, and it is larger exactly when . We solve this induced maximum-independent-set instance by iterated local search — delete random lines of the current solution, refill greedily — and rebuild the pool around each improved class. Six passes give 7,082 lines.
Stage 2: streamed tightness search. The second search works with the tightness of a candidate, its number of conflicts with the current class. Each batch draws 20 new automorphisms and forms all images of the current class lines; only candidates with tightness at most 2 are kept, which bounds memory while retaining the useful ones. On that restricted candidate set the search maintains the tightness of every candidate incrementally as lines enter and leave the solution.
Throughout, an -swap deletes class lines and inserts candidates in their place, so the class grows by lines; an -swap leaves the size unchanged. The search repeatedly applies
free adds, for candidates of tightness 0;
-swaps: two candidates of tightness 1, conflicting with the same single class line, are compatible with each other, and replacing that line by the pair gains one;
random -swaps on candidates of tightness 1: these leave the size of the class fixed and let the search drift across a plateau of equally large classes.
Stage 3: exchanges and restarts. The first two stages never consider a candidate blocked by more than one class line. Stage 3 adds them: candidates that share the same set of blocking class lines can be exchanged for that set whenever of them are mutually compatible, which is a -swap in the above sense; the blocker sets are recomputed from scratch for each attempt, so the move does not depend on incremental counters. The search is then restarted many times from the current best class with fresh automorphisms, which turns it into a randomised hill climb with restarts.
Result. Fifty batches of stage 2 give 7,108 lines. The final class, after stage 3, has 7,193 lines, obtained from the published 7,077 by adding 125 lines and deleting 9. It passes the same exact checks — norm , for all pairs, distinct lines — and the extra lines are worth up to points each in dimension : at alone the ceiling rises from 60,589,782 (with the published 7,077 lines) to 60,717,876.
None of the three searches is a certificate of maximality; all are hill climbs from a known solution. The conflict statistics also explain why the naive alternative — drop some lines at random and refill — fails: a random pool line conflicts with about ten class lines, so removing a handful frees almost no candidate, while removing enough to help destroys the structured part of the class faster than refilling restores it.
Choosing the images
For a fixed , (3) is maximised by maximising . This is a maximum-coverage instance: the ground set is the set of minimal lines of that occur in at least one image, the sets are the images , and exactly of them must be chosen.
Pool. Distinct isometries are drawn as words of length 170 in the three published generators of , retaining those that satisfy exactly. Pools of 400–4,000 images are used; when re-optimising a previous family we seed the pool with its images, which makes a pool of up to 4,577. Each line of each image is replaced by a sign-canonical 64-bit key so that identical lines coincide; for 4,577 images this is a 32,000,000-entry join, and the number of distinct lines in such a pool is about 17.4 million.
Greedy. Images are added one at a time, each time the one covering the most uncovered lines. This single pass produces the whole family: after additions we have a valid configuration for dimension , so one run of the optimiser yields every dimension at once.
Local search. We use exact best-improvement 1-swaps — replace a chosen image by an unchosen one and evaluate the true change in — restricted to the 26 most promising incoming and 8 lowest-value outgoing images, followed by iterated local search: remove a random sixth of the family, refill greedily, polish, and keep the result if the union grew.
What binds. The pool binds, not the local search. On the 7,193-line class the search reaches at , i.e. 92.05% of ; the small- entries sit at 99.95% or above, where 1-swaps alone close everything that is left. Starting the search from an existing family and adding 4,000 fresh images moves by thousands of lines, whereas local search on a fixed pool moves it by tens.
Ablation. Table 1 separates the two levers: the families are held fixed and only the class changes. Every stage of the growth from 7,077 to 7,193 lines increases at every , and at the last stage alone is worth 42,202 lines of union, out of the that its 85 new class lines could supply in the absence of overlap; the shortfall is the overlap the larger class forces, not a failure of the search.
| 49 | 1 | 7,077 | 7,082 | 7,108 | 7,193 | 100.000% | 100.000% |
| 50 | 3 | 21,231 | 21,246 | 21,324 | 21,579 | 100.000% | 100.000% |
| 51 | 6 | 42,460 | 42,490 | 42,646 | 43,156 | 99.995% | 99.995% |
| 52 | 12 | 84,886 | 84,946 | 85,257 | 86,276 | 99.955% | 99.954% |
| 53 | 20 | 141,360 | 141,459 | 141,977 | 143,672 | 99.873% | 99.869% |
| 54 | 36 | 253,981 | 254,159 | 255,082 | 258,119 | 99.690% | 99.680% |
| 55 | 63 | 442,905 | 443,214 | 444,819 | 450,091 | 99.339% | 99.323% |
| 56 | 120 | 837,045 | 837,625 | 840,646 | 850,575 | 98.564% | 98.542% |
| 57 | 153 | 1,062,166 | 1,062,901 | 1,066,697 | 1,079,243 | 98.096% | 98.066% |
| 58 | 255 | 1,744,279 | 1,745,473 | 1,751,636 | 1,771,982 | 96.656% | 96.607% |
| 59 | 302 | 2,051,070 | 2,052,463 | 2,059,649 | 2,083,435 | 95.968% | 95.910% |
| 60 | 378 | 2,538,709 | 2,540,409 | 2,549,175 | 2,578,309 | 94.901% | 94.827% |
| 61 | 577 | 3,763,184 | 3,765,682 | 3,778,178 | 3,820,380 | 92.157% | 92.049% |
Table 1. The ablation of Section 4. The families are those of Table 3 and only the class changes—the class is grown by +5, +26 and +85 lines in the three stages, so each column shows the effect of the class alone on the union of a fixed set of images. Every stage increases at every ; at the last stage adds 42,202 lines of union out of the 49,045 that its 85 new class lines could supply, the difference being the overlap that a larger class forces. The last two columns give the fraction of attained with the smallest and with the largest class: the small decrease as the class grows is that same overlap, which is why the marginal line pays slightly less than points.
Why the images cannot simply be taken disjoint. One might hope for a family whose images are pairwise disjoint, which would attain exactly. Write for an automorphism ; the loss is at first order. For 120 random automorphisms the mean of is 2.40 (histogram: 0 in 12 cases, 1 in 21, 2 in 37, 3 in 26, 4 in 12, 5 in 7, 6 in 3, 7 in 2), so a random family loses about 2.4 lines, which is the same order as the losses we observe. Cyclic families behave differently: if has odd order , the union of the powers reaches 99.6–99.9% of (for , 121 lines lost out of 77,847), but every even-order element we sampled had , so its cyclic family contains only distinct images and the union collapses to at most half the ceiling. Disjointness by construction therefore needs a family with no short relations, not a single generator; this is the same obstruction that the ceiling analysis of [5] describes, and we have not overcome it.
Tails
By Theorem 1 the tail may be any kissing configuration of containing antipodal pairs; it does not have to be a lattice shell. Table 2 lists the tails used, with the number of points and of antipodal pairs . All of them are antipodal, which is what the lift requires: it consumes antipodal pairs, not points.
| 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 | 12 | 13 | |
| tail | |||||||||||
| points | 12 | 24 | 40 | 72 | 126 | 240 | 306 | 510 | 604 | 756 | 1154 |
| pairs | 6 | 12 | 20 | 36 | 63 | 120 | 153 | 255 | 302 | 378 | 577 |
Table 2. The tails used, by dimension of the added space. , (not listed) give ; the root systems – give the classical values; , , , are the record kissing configurations of Cohn [3]; is the 756 minimal vectors of the Coxeter–Todd lattice. Every one of them is antipodally symmetric, as the lift requires, which fails for the 841-point record configuration of , so the lattice value 378 is used there instead. Each direction system was re-verified here in exact arithmetic: constant norm, antipodal, all pairwise inner products , numerical rank .
Results
Table 3 gives the results. In each dimension the value is with the and of Table 4.
| loss | |||||
| 49 | 1 | 7,193 | 7,193 | 0 | 100.000% |
| 50 | 3 | 21,579 | 21,579 | 0 | 100.000% |
| 51 | 6 | 43,156 | 43,158 | 2 | 99.995% |
| 52 | 12 | 86,276 | 86,316 | 40 | 99.954% |
| 53 | 20 | 143,672 | 143,860 | 188 | 99.869% |
| 54 | 36 | 258,119 | 258,948 | 829 | 99.680% |
| 55 | 63 | 450,091 | 453,159 | 3,068 | 99.323% |
| 56 | 120 | 850,575 | 863,160 | 12,585 | 98.542% |
| 57 | 153 | 1,079,243 | 1,100,529 | 21,286 | 98.066% |
| 58 | 255 | 1,771,982 | 1,834,215 | 62,233 | 96.607% |
| 59 | 302 | 2,083,435 | 2,172,286 | 88,851 | 95.910% |
| 60 | 378 | 2,578,309 | 2,718,954 | 140,645 | 94.827% |
| 61 | 577 | 3,820,380 | 4,150,361 | 329,981 | 92.049% |
Table 4. The union sizes behind Table 3. The loss is the number of lines counted twice: the images hold 7,193 lines in total, but a line occurring in two or more of them is one line of their union, so , and each lost line costs two points. The last column is the fraction of the trivial maximum 7,193 attained. The loss grows quadratically in , as expected from with for a random automorphism .
The margins are largest where the published families were furthest from their own ceiling. For the value in [5] corresponds to , i.e. 91.6% of , while the family given here reaches 92.05% of and gains 79,610 lines — worth 159,220 points, since every line is worth two points. At most of those can be attributed to the new class lines, so at least 14,970 come from the image selection alone. The split between the two levers depends on : at the family sits exactly at the ceiling, so the whole margin is the class growth (); for the measured margin already exceeds , the excess coming from the image selection.
In the low dimensions the gains are accordingly small in absolute terms: at the ceiling allows lines and only two of them are lost.
Two independent checks confirm that these are real gains rather than bookkeeping. First, the ceiling of the construction is exact and known: with every one of the classes at the maximum 7,193 lines the totals would be 52,416,000 , and no family can exceed them; every entry of Table 3 is at or below its ceiling, and the two entries at the ceiling are attained exactly. The ablation of Table 1 separates the two levers: its oldest-class column, with the published 7,077 lines and the same families, already reaches 3,763,184 lines at , against the 3,740,770 of [5]; crediting all eight extra class lines to the class still leaves at least 17,798 lines, that is 35,596 points, due to the choice of images. Second, a random control — the same number of images drawn at random rather than chosen by coverage — loses roughly lines, since the mean of over random automorphisms is 2.40; at that is of the order of lines, more than the 329,981 lost by the optimised family. The coverage optimisation, and the class growth behind it, are doing real work. Figure 1 shows the same margins in points, and relative to the previous best and to the ceilings of (4); § 7 quantifies the room left above them.

Figure 1. Left: the margin of Table 3 over the best previously published value, as a percentage of that value (the line at 0% is the previous best), together with the same quantity for the ceilings of (4) with and . The curves are far below their ceilings, and the gap widens with , which is where the remaining room lies. Right: the margin in points; total 506,461.
Where this method stops
The construction is capped by
which for and () is 60,717,876. Our value is 98.9% of that wall, so at most 659,962 further points are available in this dimension from any image selection. Note that the wall moves with the class: the ceiling quoted in [5] for is 60,574,780, which is (4) with their 7,069-line class. (With the 7,077-line class of [11] it would be 60,589,782.) Every line found by the searches of § 3 raises the ceiling by ; the 116 net lines we add are worth up to points there, and a class of 8,250 lines would put above 61.9 million. This is the “class problem”, open in the same sense for the Leech lattice and for [5].
Future work. Three directions seem most promising. First, the class: the searches of § 3 were still gaining lines when they were stopped, and nothing in them addresses the class problem itself, for which an upper bound would be as valuable as a larger class; comparison with a spectral or Caro–Wei-type bound (which guarantees only 712 lines for [5]) would show how far a local search can plausibly go. Second, the images: the loss grows quadratically in and eats 7.95% of the ceiling at ; a family with few short relations, which would make the images nearly disjoint, is the natural target, and we have only ruled out the cyclic families that came to mind. Third, the framework: the same lift applies to any lattice shell with a large class of lines at inner product , so the Leech lattice and [5] are the obvious next test beds, and the tail side is limited only by the record antipodal kissing configurations of small dimensions.
Above the method is not competitive and we make no claim there. Dimensions 62 and 63 are dominated by the Edel–Rains–Sloane code construction [4], which gives 71,310,732 and 138,419,844; (4) with the largest antipodal tails there, and [3], gives only 66,314,808 and 70,861,416, below even the level-0 sign code of that chain. In dimension 64 and beyond [4] gives 331,737,984 and more, which is five times what this family can reach.
Verification
Each configuration is specified by the integer matrices and is checked from scratch by code/verify.py, which
checks for every , in exact integer arithmetic without floating point;
applies
data/class_patch.jsonto the published class, which adds 125 lines and deletes 9, and checks the resulting class: for all 7,193 rows and for every pair of distinct rows, and that the 7,193 lines are distinct;
checks that no two selected images coincide as line sets;
recomputes without hashing: the sign-canonical integer rows of all images are stacked and de-duplicated with
np.unique(axis=0), so the union size is exact by construction and independent of the search code;
reports .
The script code/certificate.py runs the same checks over the whole table in one pass and writes results/certificate.txt, one line per dimension, ending with an assertion that every entry agrees with results/verified.json, and records the SHA-256 checksums of data/class_patch.json and results/verified.json, so that a revision can be cited unambiguously; pool size, word length and random seed behind each family are recorded in results/pool_stats.json. The premises taken from the literature are only two: that the minimal shell of has 52,416,000 vectors of norm 6 [2, 8], and Theorem 1 [11]. The 7,077-line class is the one supplied with [11], and the added lines are those of data/class_patch.json; the property (1) of the resulting class is re-verified here rather than assumed.
Contribution statement
The author proposed the problem, the overall approach and the specific experiments, and reviewed all results. The research code, the search and the numerical experiments reported in this paper were written and executed by an AI agent (Arena.ai Agent Mode), including the implementation of the maximum-coverage search, of both class searches, of the exact verification program, the comparison with the published bounds, and the preparation of this manuscript. The author checked the verification code and the distributed certificates.
References
- [1] [CKMRV17] H. Cohn, A. Kumar, S. D. Miller, D. Radchenko and M. Viazovska, The sphere packing problem in dimension 24, Ann. of Math. 185 (2017), 1017–1033.DOI
- [2][CS99] J. H. Conway and N. J. A. Sloane, Sphere Packings, Lattices and Groups, 3rd ed., Grundlehren der mathematischen Wissenschaften 290, Springer, New York, 1999.
- [3][Coh] H. Cohn, Table of kissing number bounds, https://cohn.mit.edu/kissing-numbers/, accessed October 2026.
- [4][ERS98] Y. Edel, E. M. Rains and N. J. A. Sloane, On kissing numbers in dimensions 32 to 128, Electron. J. Combin. 5 (1998), Research Paper 22; arXiv:math/0207291.
- [5][Kra26] A. Kravatsky, KissingNumbers, public repository, 2026, https://github.com/alexlegeartis/KissingNumbers, in particular verifications/improved/dim49–63-p48-caps.
- [6][Mus08] O. R. Musin, The kissing number in four dimensions, Ann. of Math. 168 (2008), 1–32.arxiv.org/abs/math/0309430
- [7][Neb98] G. Nebe, Some cyclo-quaternionic lattices, J. Algebra 199 (1998), 472–498.DOI
- [8][NS] G. Nebe and N. J. A. Sloane, A Catalogue of Lattices, https://math.rwth-aachen.de/~Gabriele.Nebe/LATTICES/, accessed October 2026.
- [9][Sch53] K. Schütte and B. L. van der Waerden, Das Problem der dreizehn Kugeln, Math. Ann. 125 (1953), 325–334.DOI
- [10][Via17] M. S. Viazovska, The sphere packing problem in dimension 8, Ann. of Math. 185 (2017), 991–1015.DOI
- [11][Yan26] S. Yang, R. Zhao, J. Wu, Y. Wang, F. Chen, K. Zhu, W. Li, Z. Li, Y. Li, S. Hong, Y. Pan, J. Yang, T. Deng, J. Mi, H. Chen and Y. Yang, Large-scale autonomous discovery of kissing number constructions, arXiv:2609.35051 (2026).arxiv.org/abs/2609.35051