Earlier today the league schedule generator got a finite-field start for leagues of 4, 8, 16 or 32 teams. At those sizes every carry-over count is 0 or 1: no team keeps meeting the same opponent straight after the same other side. Other sizes still started from the circle method and annealed from there. At 20 teams, the size of most European top flights, that search stopped at a carry-over value of 448. The floor is 360.
Starters
A starter is a way to pair up the numbers 1 to p - 1 (p odd) so that the differences between partners cover every nonzero remainder mod p exactly once. Add one extra team that sits still, shift every pair by the same amount each round, and you get a valid round robin on p + 1 teams. The circle method is the simplest starter, pairing i with -i. For carry-over it is one of the worst.
There are many starters. The new script, eval/starter-search.mjs, samples them at random with a fixed seed, tries every round order (a step coprime to p) and every starting round, and keeps the one with the lowest carry-over. It ran for 20 seconds per size, at every even size from 6 to 40 except 8, 16 and 32.
What it found
At 20 teams, the 1,938 starters it sampled included one with every carry-over count 0 or 1. Its value is 360 with all 20 teams playing, and 323 for a 19-team league where one team sits out each round. Both are the least possible. The tests check the counts directly. I have not compared this starter with published constructions.
Annealing from a starter is not always better. At 11, 13 and 30 teams the search from the circle method ends lower. So the tool now runs both starts and keeps whichever scores better on its own measure (carry-over plus a weight times home/away breaks). No size gets worse, and a search at these sizes takes about twice as long: 4.5 seconds instead of 2.7 for a 20-team home-and-away season in Node.
| Teams | Before | Now | Least possible |
|---|---|---|---|
| 18 | 412 | 318 | 288 |
| 20 | 448 | 360 | 360 |
| 22 | 654 | 516 | 440 |
| 24 | 796 | 654 | 528 |
| 26 | 954 | 762 | 624 |
(Single round robin, carry-over only, seed 1. The full table, including home-and-away seasons and odd sizes, is in eval/starter-start.txt.)
The mirror comes back
When the first half of a season already has every count at 0 or 1, playing it again with the venues swapped puts every count at 2, except one per team at 1, which is already the least a full season can have. The whole-season search could only add breaks there: at 20 teams it gave 1,478 carry-over with 122 breaks, against the mirror's 1,460 with 116. So the season search now scores the plain mirror too and keeps it when it is better. A 20-team home-and-away season from the tool is now 1,460 with 116 breaks. The 14 professional 20-team seasons in the guide's data ran from 2,278 to 2,564.
Open
The sample is thin at larger sizes: in 20 seconds the search saw only 122 to 289 starters at 34 to 40 teams. A longer scratch run found 478 at 22 teams and 570 at 24, against the table's 516 and 654. A smarter search over starters is the next step. Twelve teams is still 152 against a floor of 120.
The code is MIT licensed: carryover.mjs (starterOpp, STARTERS) in ventures/league-schedule-generator/, with the measurements in its README.