The s(110) landscape
We know — different ways to pack 110 unit squares into a square of side less than 11. Each one is a local minimum: no small motion of the squares lets the box shrink. They differ in side by at most —, they share one structure, and the deepest of them sit in two pockets. This page shows what that means for anyone hunting for a better packing.
Hundreds of floors, two deep pockets
Each dot is one certified local minimum, placed by its side length. Below 11 there is a single structure: a band of squares tilted about 28° along the left wall, a band tilted about 62° along the bottom, and 64 to 70 squares parallel to the walls filling the rest. The rows sort the minima by how many squares are tilted. That one number is the clearest dividing line in the census: 42 and 43 tilted squares hold the deepest minima, and every step away from them costs depth.
Hover or tap a dot to see its packing in the panel below; click to hold it.
Islands of similar layouts
Here the same minima are placed so that similar layouts sit close together. Similarity compares where the axis-parallel and the tilted squares sit, ignoring which square is which and the eight symmetries of the box. Positions within an island are meaningful; distances between islands only roughly so.
The minima form islands. An island usually keeps one count of axis-parallel squares, while the split between the two bands and the corner where they meet varies inside it: moving one square between the grid and a band is a small change. Hover a dot to ring every minimum with the same tilt counts; they are often spread over several islands. The two deep pockets are the record's island (43 tilted) and a second island, mostly 42 tilted, that also takes in the best 41-tilted minima.
To get from one minimum to the next, the box has to grow
Pick two minima. Match each square in one to a square in the other, so the total corner movement is as small as possible. Then look for the smallest box in which the squares can move from the first arrangement to the second without ever overlapping. That box size, minus the larger of the two sides, is the expansion the transition needs.
The animation runs a tour through — representative minima, from most of the larger islands and several from each deep pocket, chosen to keep the expansions small. In each step the packing first spreads out to fill the larger box, then rearranges, then settles into the next minimum. The chart below the packing plots the side along the way: each dip is a local minimum, and each hump is the expansion needed to reach the next one.
A small shake finds its way back; a slightly bigger one does not
Both runs start from the packing with side 10.996793 (the record before September 27). The box is loosened by 3 %, every square is nudged by a random amount, and a local optimizer squeezes the box down again while keeping the squares apart. On the left the nudges have standard deviation 0.01 in position and 0.2° in angle; on the right 0.03 and 0.6°.
In two of our four small-shake runs, including the one on the left, the squeeze lands in a better minimum than the start: Couzo's packing, side 10.996783. The other two jam just above 11. With three times the shake, all four runs end at 11 or above; two end at exactly 11, which is what a full row of 11 squares parallel to the walls forces.
Every contact pushes, and the pushes balance
To shrink the box, the walls must move in. A packing is stuck when the contacts can carry that inward push: there are positive contact forces, one for each touching pair or square-and-wall, that exactly balance a unit inward force on the walls. Then any small motion that keeps the squares apart costs side length, to first order or (where the first-order test is flat) to second order. The exact solver checks this at an 80-digit point and also writes an exact rational certificate for the side.
Stuck does not mean frozen. Some squares carry no force (rattlers), and in most minima some groups of squares can slide together without changing the side. Those are the flat motions; the record has — squares involved in them. They never lead anywhere lower.
- Census. Basin-exploration runs of our packer, started from the known record funnel and, in the latest round, from minima far from the best ones; every candidate below 11 is solved exactly. A minimum counts when it has a valid rational certificate and the solver's first- and second-order tests pass. A further 11 minima pass the first-order test, but each has a direction of negative curvature that contacts block only to first order, so the second-order test is inconclusive; they are left out until it can be settled. (Until October 15 a solver bug left 45 more minima in this group: contacts carrying very small forces, around 10−9, were taken as force-free. Fixed, they pass.) Snapshot: —.
- Distinct. Two minima are different when their exact sides differ. Sides that agree to nine digits still belong to visibly different packings here.
- Map. Each packing becomes two blurred 24 × 24 images of square centres (axis-parallel and tilted), compared by the summed absolute difference, minimised over the 8 symmetries of the box. This agrees closely with the corner-movement distance (rank correlation 0.93 on 11,175 pairs) at a fraction of the cost. The map is a UMAP layout of that distance (raised to the power 0.5; 250 neighbours, min_dist 1, spread 2). Islands: groups linked by distances below 0.02.
- Expansion. For each pair: spread the first packing to fill a box of side s, follow the straight-line morph to the second with small feasible steps (linear programs that keep every pair of squares apart), and bisect on s. Our walk is not mirror-symmetric, so tour steps were also tried with both packings reflected in the diagonal, keeping the smaller box. The result is an upper bound on the expansion needed, for this family of paths. Six candidate stops that no walk reached cheaply were left out of the tour.
- Not covered. The census only covers what our search reaches. Searches started far from the record found their way back to it within an hour, which points to a single funnel, but a different structure below 11, or a better minimum with a very small basin, is not ruled out.
- Code: search/packer/landscape/ and search/exact/exactsolve.py.