MathematicsPreprintTheory3 min read

THE ONLY MAHJONG HAND THAT WAITS ON EVERY TILE

Mahjong is played with tiles ranked 1 to 9 in three suits, with four copies of each tile. To win, a player must assemble one pair and four melds, where a meld is either three identical tiles or three tiles of consecutive ranks. Players draw and discard in turn, hoping to complete such a hand.

Among all possible hands, one has a legendary status. Nine Gates uses thirteen tiles of a single suit: 1-1-1, 2, 3, 4, 5, 6, 7, 8, 9-9-9. Whatever tile of that suit comes next, the hand is complete. Draw a 5, for instance: the pair is 5-5, and the melds are 1-1-1, 2-3-4, 6-7-8 and 9-9-9. This “nine-sided wait” earns the highest category of award in many variants of the game.

A folk theorem, checked by machine

Players have long held that Nine Gates is the only single-suit hand with this property, as long as the rule of four copies per tile is respected. Until now, that belief rested on computers. A 2019 study enumerated all 93,600 possible thirteen-tile hands of one suit and confirmed it; another classified every waiting pattern into 828 categories.

Shubin Xue, of the Beijing Institute of Technology, and Xingyu Zhong, of the National University of Singapore, wanted a proof that explains rather than counts. They also asked a broader question. What if a suit had 3n ranks instead of nine, and a winning hand needed one pair and n + 1 melds? Ordinary mahjong is the case n = 3.

Eight hands, one survivor

Their answer is a complete classification. For any suit of six, nine, twelve or more ranks (any multiple of three from six up), there are exactly eight hands that win whatever rank is drawn, if you set aside the four-copy limit. One is the generalised Nine Gates: three copies of the lowest rank, one of each middle rank, three of the highest. The seven others, three of which are mirror images of three more (the same hand with the ranks read from the top down), all share a flaw: each already holds four copies of some rank, so drawing that rank would require a fifth tile that does not exist. Nine Gates is therefore the only one that can be played with a real set. For the smallest suit, with three ranks, there are six such hands, all playable.

In ordinary mahjong, one of the impostors reads 1-1-1, 2, 3, 4, 5, 6-6-6-6, 7, 8. Another runs 2, 3, 4-4-4-4, 5, 6-6-6-6, 7, 8, symmetric around its middle.

The proof involves no exhaustive search and does not assume the answer in advance. It works with hands as lists of counts per rank, and tracks which pair each draw can use to win. Two facts emerge: a hand that waits on every rank always contains every rank other than the lowest and the highest at least once, and it can always be split into two pairs plus a set of melds. With fewer melds — say a pair and three melds on nine ranks — the authors show that no hand waits on every tile.

When the hand grows

Taiwanese mahjong uses sixteen-tile hands: one pair and five melds. There, the property becomes common, because adding melds to a waiting hand keeps it waiting. A computer search by the authors finds 149 hands that wait on every tile, of which 11 can actually be played, confirming an earlier count. To tame this abundance, the authors propose focusing on “minimal” hands, which cannot be reduced by removing melds, and conjecture a formula for the largest number of melds such a hand can use.

A proof being checked by machine

The authors are building a formal version of their proof in Lean, a programming language for verified mathematics, and describe it as ongoing and curated by hand. They state that GPT and Gemini models helped draft and edit the manuscript, prove auxiliary lemmas, draft candidate Lean code and explore research directions, and that they independently verified every argument. An online tool lets anyone explore the hands for themselves.