Kelvin problem

Imagine filling space with cells that all enclose the same amount of volume. The cells share walls. Which repeating arrangement needs the least wall area per cell? In two dimensions the familiar hexagonal honeycomb wins the analogous boundary problem. Kelvin’s problem asks for the three-dimensional answer, where surfaces can curve and neighboring cells must fit together.[1] [2]

Equal-volume cells fill three-dimensional space and share walls. Which periodic arrangement minimizes the wall area per cell when any finite number of cell types is allowed? Take a periodic partition with m unit-volume cells in a translation period. Its objective is the area of the interface set in the corresponding quotient, divided by m; shared interfaces contribute once, including those crossing the period boundary. Allow the lattice and m to vary. This formulation avoids a choice of finite observation window and retains multi-cell candidates. Scaling all lengths by s multiplies cell volume by s³ and area by s², so fixing volume removes a trivial scaling freedom.[1] [2]

Kelvin proposed one arrangement, but a later construction uses less surface area. That defeats his proposed answer without supplying a proof of the best possible answer. The leading candidate still must be compared with every allowed periodic arrangement.[1] [2]

The 2025 Cesaroni–Novaga paper states that the volume-constrained Kelvin problem remains open and then proves results for a distinct fixed-inradius model. Local equilibrium and regularity conditions on soap-film interfaces constrain potential minimizers but do not compare their global areas. A proof must defeat every admissible period and arrangement, not merely variations of a chosen candidate. The periodic restriction here is itself unresolved and does not assert a solution for all nonperiodic partitions.[1] [2]

A shared wall serves two cells, so efficient cells cannot be designed one at a time. The problem reveals how surface minimization and the need to fill space constrain each other.[1] [2]

Unlike nearest-point quantization, the objective depends on interfaces rather than distances to selected centers. The best way to reduce squared approximation error need not give the least wall area, even when both problems suggest related cell shapes. Kelvin’s problem isolates how a global tiling constraint changes surface optimization beyond the isoperimetric problem for one isolated region.[1] [2]