THE SHAPE THAT PACKS WORST: A ROUNDED OCTAGON?
Take a flat shape — convex, and symmetric about its centre, like a coin or a rounded tile. Lay identical copies on a regular grid, every copy shifted from the next by the same steps, without any overlap, and push them together as tightly as possible. Part of the plane always stays uncovered. The best density — the share of the plane covered — a shape can reach this way depends on the shape; stretching or shearing it does not change it.
The Reinhardt problem asks which shape has the lowest best density: the shape that packs worst, even when packed as well as it can be.
An odd candidate
In 1934, Reinhardt proposed an answer: the smoothed regular octagon, a regular octagon whose corners are rounded off by arcs of hyperbolas. As the new preprint notes, the problem is unusual among packing questions: the shape that loses is neither a polygon nor an ellipse.
Progress came in steps. In 1947, Mahler split the problem in two: is the worst shape always a “smoothed polygon”, made of a finite number of straight and hyperbolic pieces? And if so, must it be Reinhardt’s octagon? In 1986, Nazarov showed that the octagon is at least a local minimum: nudging its shape slightly cannot make it pack worse. Later, Hales recast the question as a problem of drawing the boundary arc by arc, each arc traced by holding one setting fixed (an “optimal-control” problem), and Hales and Vajjha proved Mahler’s first question — but without a number for how many pieces the worst shape could have.
The claimed proof
In a 472-page preprint dated 17 November 2026, Fei Xue claims to settle the second question. Its main theorem states that every such shape has a best packing density at least
(8 − 4√2 − log 2) / (2√2 − 1),
the value reached by the smoothed octagon, with equality only for the octagon and its stretched or sheared copies.
The introduction lays out the plan. First, an explicit, astronomically large limit on the number of arcs: fewer than 2¹⁶⁵. Then sharper estimates — lower bounds on a local “cost” attached to short runs of arcs — bring it down to at most 21. That leaves 1,355 possible arrangements of arcs (“words”); exact calculations rule out 1,334 of them, leaving 21 (words this time, not arcs). Further geometry narrows these to six types, named R01 to R06. The preprint then excludes each one with its own equations, geometric reductions and computer-checked calculations, and concludes that only the octagon remains. An appendix describes the verification algorithms, and the data come with the paper.
Who did what
The acknowledgements are unusually frank. The author proposed the argument giving the first explicit limit on the number of arcs. “ChatGPT dot”, the text says, carried out the sustained exploration of these local cost bounds, and “ChatGPT 6.1 Sol carried out the proofs of the six remaining types R01–R06 and the organization of the manuscript”. The author retains responsibility for the mathematical claims.
Fei Xue writes of meeting the problem in seminar discussions with Chuanming Zong and Thomas Hales around 2015, and of being left restless by it for more than a decade.
What comes next
For now, this is a proof announced, by a single author, at a length few readers will go through quickly: the preprint has not been checked by other mathematicians. The author does not hide it either, closing the introduction with the hope that “further work will lead to a shorter and more beautiful proof”.
