The values
| n | Continuous optimum | Grid maximum | Proof |
|---|---|---|---|
| 1 | 0.288675134 | §1 | |
| 2 | 0.384900179 | §1 | |
| 3 | 0.549038104 | §1 | |
| 4 | 0.617524884780866… | 0.617524883 | §2–4, §6–7 |
| 5 | 0.693904526840437… | 0.693904525 | §2–4, §6–7 |
| 6 | 0.803847575 | §2–4, §6–7 | |
| 7 | 0.855687593470088… | 0.855687591 | §2–4, §6–7 |
| 8 | 0.912454563520543… | 0.912454561 | §2–4, §6–7 |
| 9 | 0.976596443373822… | 0.976596440 | §2–4, §6–7 |
| 10 | 1.056624324 | §2–4, §6–7 | |
| 11 | 1.098475709724551… | 1.098475706 | §2–4, §6–7 |
| 15 | 1.308432222 | §2, §5–7 |
Values without a closed form are the radius sum of the unique positive root of one contact system; the table shows 15 digits of our 60-digit solution, accurate to 10⁻³⁰. No continuous optimum lies on the grid, and each grid maximum is strictly below it.
0. Setting
The container has vertices (0,0), (1,0) and (1/2, √3/2). Circle i has centre (xᵢ, yᵢ) and radius rᵢ ≥ 0. The constraints are the three walls yᵢ ≥ rᵢ, √3xᵢ − yᵢ ≥ 2rᵢ and √3(1 − xᵢ) − yᵢ ≥ 2rᵢ, and |cᵢ − cⱼ| ≥ rᵢ + rⱼ for every pair (tangency allowed). Write z = (x₁, y₁, r₁, …, xₙ, yₙ, rₙ) ∈ ℝ³ⁿ; the objective is c·z = Σrᵢ. A site answer has at most nine decimals per number, and its score is the integer 10⁹·Σrᵢ.
Let Mₙ be the continuous optimum for n circles. The symmetry group D3 of the triangle (three rotations, three reflections) and any relabelling of the circles preserve feasibility and the objective.
1. n = 1, 2, 3
n = 1
The distances from the centre to the three sides always sum to the height √3/2 (add the areas of three small triangles), and each is at least r, so 3r ≤ √3/2 with equality for the incircle: M₁ = √3/6. On the grid an integer m needs (6m)² ≤ 3·10¹⁸, and (6·288675134)² < 3·10¹⁸ < (6·288675135)², so the maximum is 288675134, attained by the incircle at r = 0.288675134.
n = 2
Let ρ = √3/6. A circle of radius r has its centre in the triangle T_r, each side moved inward by r. For fixed r and s the distance between the centres is convex, so its maximum is at a pair of vertices of T_r and T_s. Put t = r + s and d = |r − s| ≤ 2ρ − t. Vertices at the same corner are 2d apart; vertices at different corners are at squared distance (1 − √3t)² + d².
If t > 4ρ/3 then d ≤ 2ρ − t < 2ρ/3 < t/2, so the same-corner distance 2d < t is too short, and
(1 − √3t)² + (2ρ − t)² − t² = 3(t − 4ρ/3)(t − 4ρ) < 0 (4ρ/3 < t ≤ 2ρ)
so different corners fail too. Hence r + s ≤ 4ρ/3 = 2√3/9, attained by the incircle and a tangent circle of radius ρ/3 in one corner gap. On the grid (9·384900179)² ≤ 12·10¹⁸ < (9·384900180)², so the maximum is 384900179, attained by (0.5, 0.288675134, 0.288675134) and (0.5, 0.673575313, 0.096225045).
n = 3
With the same vertex argument, the squared different-corner distance minus the same-corner one is (1 − 2√3a)(1 − 2√3b) ≥ 0, so every pair must satisfy
F(a, b) = 1 − 2√3(a + b) + 3a² + 2ab + 3b² ≥ 0.
F is strictly decreasing in b ∈ [0, ρ], so b ≤ g(a) = [√3 − a − 2√(a(√3 − 2a))]/3. Let R = (√3 − 1)/4 and order the radii a ≥ b ≥ c. If a ≤ R the sum is at most 3R. If a ≥ R it is at most H(a) = a + 2g(a); with h = √(a(√3 − 2a)), g″(a) = 1/(2h³) > 0, so H is convex and peaks at an endpoint: H(R) = 3R and H(ρ) = 5ρ/3 < 3R. So M₃ = 3R = 3(√3 − 1)/4, attained by three circles of radius R in the corners, pairwise tangent.
Grid: ⌊10⁹M₃⌋ = 549038105, but it cannot be reached. If the largest radius is at least 0.183012703, both endpoint values of H are below 0.549038105 (checked with rational intervals); otherwise all radii are at most 0.183012702, and a sum of 0.549038105 needs two of them equal to 0.183012702, where F(b, b) < 0. So the maximum is 549038104, which the current record attains.
2. Reduction to full contact graphs
Lemma 2.1 (positive radii). The feasible set is compact, so a maximum exists. Suppose a maximiser had a circle of radius 0. A small neighbourhood of a triangle vertex is always uncovered (a disc inside the triangle cannot reach a vertex), and a small positive circle placed there would increase the objective.
Lemma 2.2 (full contact). Some maximiser has active constraints (tangent pairs and wall contacts) whose gradients span ℝ³ⁿ. Among all maximisers take one maximising |z|². If the active gradients do not span, take a common null vector v ≠ 0. The pair function |cᵢ − cⱼ| − rᵢ − rⱼ is convex, so it cannot drop below its linearisation, which is zero along v; wall constraints are affine; inactive constraints survive small steps. So z + tv is feasible for small t of either sign. If c·v ≠ 0 one side increases the objective; if c·v = 0 both sides are maximisers and |z + tv|² strictly increases on one of them. Either way, a contradiction.
Lemma 2.3 (triangulation). With the three walls as three vertices, the contact graph plus the outer triangle is a simple plane graph, so it has at most 3(n + 3) − 6 = 3n + 3 edges; without the three wall–wall edges, at most 3n contacts. Lemma 2.2 gives at least 3n, so exactly 3n, and the plane graph is maximal: a triangulation. (The contact-free arcs near the corners are kept and then contracted; the package’s contact lemma treats each such boundary case.)
So the continuous optimum is the largest radius sum realised by one of finitely many triangulations whose outer face is the three walls and whose n interior vertices are the circles.
3. Each contact graph has at most one realisation
Around circle i, the angle at i of each triangular face is 2π/3 for two walls, 2 arctan √(rⱼ/rᵢ) for a wall and circle j, and 2 arctan √(uv/(1 + u + v)) for circles j and k, where u = rⱼ/rᵢ and v = rₖ/rᵢ. The last two are strictly increasing in each neighbour ratio, and the angles around every circle sum to 2π.
Lemma 3.1. Two positive realisations of the same contact graph have the same radii. Take the circle whose new/old ratio is largest; none of its neighbour ratios can increase, and if one strictly decreased its angle sum would fall below 2π. So all its circle neighbours share the ratio, and since the circle-to-circle graph is connected, every circle does. Finally the chain of tangent points along a wall is homogeneous of degree 1 in the radii (a corner circle contributes √3r, consecutive tangent circles 2√(rᵢrⱼ)) and equals 1 in both realisations, so the common ratio is 1.
Once the radii are fixed the centres are too: start from the circle in a wall–wall corner face and walk the dual graph; each new centre is one of two points at known distances from known points, and the orientation of the triangulation picks one.
4. n = 4–11: enumerating every contact graph
Triangulations with a fixed outer face are connected under edge flips (Bose and Verdonschot, arXiv:1206.0303, §2, Lemmas 1 and 3), so closing one triangulation under all legal flips produces every one. The package gives, for every legal flip of every graph, an explicit isomorphism to a listed graph; we regenerated the lists from scratch with our own code.
| n | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 |
|---|---|---|---|---|---|---|---|---|
| graphs | 16 | 78 | 457 | 2 938 | 20 118 | 144 113 | 1 065 328 | 8 068 332 |
For every graph with n = 4–9 the package stores a rational box and an approximate inverse R and verifies, in exact rational arithmetic, the Krawczyk conditions ‖I − RJ(X)‖∞ < 1 and self-mapping, so the box holds exactly one root; every non-contact pair and wall has a strictly positive gap, so the root is a valid packing. The best graph’s lower bound exceeds every other graph’s upper bound. For n = 10 and 11 the package takes shorter routes (type bounds as in §5; for n = 11, vertex-split coverage with comparison bounds). Independently, for every n = 4–11 we solved the angle-sum equations of every graph and checked every contact and non-contact, and found the same best graph, agreeing with the package to every printed digit.
For n = 6 and 10 the best graph is the equal-circle triangular array (3 and 4 rows): with m rows, r = 1/(2√3 + 2(m − 1)), giving 6 − 3√3 and 5(3 − √3)/6.
5. n = 15: power cells
For n = 15 there are too many graphs to enumerate. Take the full-contact maximiser of §2. If a circle touches all three walls it is the incircle, and area with Cauchy–Schwarz gives Σr ≤ √3/6 + √(14(√3/(4π) − 1/12)) < 1.308432221, below the target. From here on every circle touches at most two walls.
Power cells. Cᵢ = triangle ∩ {x : |x − cᵢ|² − rᵢ² ≤ |x − cⱼ|² − rⱼ² for all j}. These convex polygons partition the triangle, and circle i lies in Cᵢ. Each physical contact gives Cᵢ a side of positive length; the plane dual already has the 3n contact edges plus 3 outer edges, the maximum possible, so there are no other adjacencies: the number of sides kᵢ is exactly the degree of circle i, and each corner lies in exactly one cell. With wᵢ the wall contacts and tᵢ ∈ {0, 1} the corners in the cell, define the charge qᵢ = kᵢ + wᵢ − 6. Counting gives Σqᵢ = 0 and Σtᵢ = 3.
Three relations. (i) Area: every side of Cᵢ is tangent to circle i; with the outward-normal gaps β₁…β_k summing to 2π, area = rᵢ² Σ tan(βⱼ/2) ≥ Cᵢ·rᵢ² by convexity of tan, where Cᵢ depends only on the type (k, w, t) and the allocation of angles; the areas sum to √3/4. (ii) Wall length: each wall’s chain of tangent points has length 1, and 2√(rᵢrⱼ) ≤ rᵢ + rⱼ gives 3 ≤ Σ bᵢrᵢ with bᵢ = 2wᵢ + 2(√3 − 1)tᵢ. (iii) Wall junctions: exactly two cells meet at each junction on a wall, with angles summing to π, so Σ(Sᵢ − πwᵢ/2) = −3π/2, where Sᵢ is cell i’s sum of normal half-gaps at its wall junctions.
Weighting the three by λ > 0, μ ≥ 0 and ρ ≥ 0 and completing the square in each rᵢ gives
Σr ≤ λ√3/4 − 3μ + [Σᵢ F(typeᵢ) + 3πρ/2] / (4λ), F = maxangles [(1 + μb)²/C + ρ(S − πw/2)].
Bounding F for each cell type (k, w, t) is a finite rational computation: bisect the normalised angle domain and, on every leaf box, bound the area coefficient from below with truncated rational sine and cosine series. For n = 15, k ≤ 16: 135 types, 22,937 nodes.
Cases by charge. If not every charge is 0, then there are two positive cells, or a cell of charge at least 2, or a defect at a corner, a non-corner two-wall cell, or both defects on a wall; seven structural dual certificates (55 exact inequalities each) push all of these below the target. What is left is exactly one +1 and one −1: the three corner cells are of type (4, 2, 1), and by the number B of boundary circles and the defects’ positions there are 33 type histograms, each excluded by its own angle-domain certificate (3,101 nodes in all) strictly below 1.308432221. So every charge of a maximiser is 0.
Zero-charge lemma. At a circle of largest radius, three-circle angles are at most 60°, one-wall angles at most 90° and two-wall angles exactly 120°, so the upper bounds sum to (k + w)π/3 = 2π and must all be attained: every neighbour has the same radius, and connectivity spreads this to every circle. Equal circles in full contact have regular wall chains and form a complete triangular array, so n = m(m + 1)/2. For n = 15 that is 5 rows, r = (4 − √3)/26 and Σr = (60 − 15√3)/26, with centres x_ij = r(√3 + j + 2i), y_ij = r(1 + j√3), j = 0…4, i = 0…4 − j.
6. The grid: locking every high-scoring packing into small boxes
Let T be the score to exclude, one unit above the grid record. T exceeds Mₙ₋₁, so every packing scoring at least T has positive radii (removing a circle leaves n − 1); T also exceeds the bound of every contact graph except the best one.
Lemma 6.1. Every connected component of the set H of packings scoring at least T contains a D3 image of the optimum z* (after relabelling). Within the component maximise the radius sum and then |z|²; the argument of Lemma 2.2 applies verbatim (the perturbation stays in the component) and gives a full-contact point, whose graph can only be the best one, so by §3 it is an image of z*.
Lemma 6.2 (quadratic separation). At z* take the 3n active constraints (affine wall distances, and |cᵢ − cⱼ|² − (rᵢ + rⱼ)² for pairs); the Jacobian J is invertible and the multipliers λ = −J⁻ᵀc are all strictly positive. With d = z − z* and δ = |d|, F(z) = Jd + Q(d), where wall remainders vanish and pair remainders are at most 2δ². With p the indicator of pair rows and ε = M − T, take
A ≥ maxᵢ |(J⁻¹)·i| / λᵢ, B ≥ 2A Σpairs λᵢ + 2|J⁻¹p|.
Then every feasible point scoring at least T satisfies δ ≤ Aε + Bδ². If 4ABε < 1, no such point lies on the sphere δ = 1/(2B); a component contains an image of z* and cannot cross the sphere, so it lies inside, where the same inequality gives δ < 2Aε. Hence every high-scoring packing, after relabelling, lies in one of six coordinate boxes of half-width 2Aε around the D3 images of z*. n = 11 and 15 use the same argument in the ∞-norm.
The constants are computed with rational intervals: ‖I − RJ‖∞ ≤ θ < 1 bounds every entry of J⁻¹ − R by θ‖R‖∞/(1 − θ); A and B are rounded up, √3 is enclosed by rational squares, and box bounds are rounded inward to 10⁻⁹ (lower up, upper down). Each of the six D3 images gives its own integer box; identical boxes merge only with an explicit circle permutation (n = 11 also pairs the six images through x → 1 − x).
7. The grid: excluding every point in the boxes
In each integer box, expand every pair constraint over the box with its worst quadratic remainder moved to the right, and write the walls with a rational upper bound on √3, giving necessary linear inequalities A w ≤ b in integer offsets w from the box base, plus the target row −Σwᵣ ≤ ΣR − T. This is an outer model: every truly feasible point satisfies it.
Then a branch tree: at each node, nonnegative rational dual multipliers project bounds on each coordinate, rounded to integers, and the node splits one integer variable into ≤ k and ≥ k + 1. Every leaf either bounds the objective below T or has a strictly positive phase-I bound (infeasible), checked exactly from stored nonnegative rational multipliers with residuals compensated by the full variable bounds. No floating-point solver status is used as proof. For n = 15 each node also adds integer-hull cuts (walls from ⌊√3x⌋ tables, pairs from ⌈√(R² − dy²)⌉ tables), each verified against every relevant integer.
| n | Excluded scores | Certificate |
|---|---|---|
| 4 | ≥ 0.617524884 | 3 boxes, 85 nodes |
| 5 | ≥ 0.693904526 | 6 boxes, 474 nodes |
| 6 | ≥ 0.803847576 | 1 box, 287 nodes |
| 7 | ≥ 0.855687592 | 2,844 nodes |
| 8 | ≥ 0.912454562 | 4,400 nodes |
| 9 | ≥ 0.976596441 | 5,276 nodes |
| 10 | ≥ 1.056624325 | 1 box, 3,997 nodes |
| 11 | ≥ 1.098475707 | 3 trees, 368,119 nodes |
| 15 | ≥ 1.308432223 | 1 box, 20,191 nodes |
Each tree covers every integer point of its box and every leaf is closed, so no nine-decimal packing scores T or more; the current record reaches T − 1, which is therefore the grid maximum.
8. Our independent verification
zzzcy #308 emailed the proof package on 4 October 2026 (about 1.1 GB, written with AI assistance, as stated). Under the site’s rules we ran none of its scripts; we read only its data and decided every step with our own code:
- n = 1–3: recomputed from the derivations above at 60 digits and in exact integers, with the three answers checked by the site verifier.
- n = 4–11, continuous: our own enumeration of every triangulation and its radii (tools/p93-contact-graphs.py, tools/p93-contact-enum.cpp).
- n = 4–11, grid: A, B, ε and the boxes re-derived at 60 digits, the outer model rebuilt, and every leaf multiplier checked (tools/p93-grid-certificates.py, tools/p93-n11-grid-certificate.py); n = 4–8 were also reproduced by an independent search.
- n = 15: every type bound, structural certificate and all 33 histograms (tools/p93-power-cell-certificates.py), and the localisation with the 20,191-node integer-hull tree (tools/p93-n15-grid-certificate.py).
Each n takes seconds to minutes to check (n = 11’s 368,119 nodes take about 82 seconds). Every value agrees with the package, and we found no error in the reasoning.
Credit and scope
The constructions were known: David W. Cantrell (2011) for n = 4–11, and the equal-circle arrays for n = 6, 10 and 15 in Erich Friedman’s compilation. This proof shows they are optimal and settles the highest grid score. The adopted contribution earns zzzcy #308 one permanent +2 proof award.
The same package also proves the continuous optima for n = 21, 28, 36 and 45, which we have verified; their grid scores can still rise, and they will be handled after week 6 ends on 11 October.