- Code: Select all
`ab . . | . Bc .`

. . . | . . .

. . . | . . .

---------------

. . . | . . .

AD . . | . Cd .

. . . | . . .

Now in a constraint subset, there are the A constraints which have no overlapping values and exist entirely within the B constraints. In this example the A's are these columns and rows, each for a particular digit. In this case, advanced coloring tells us we can eliminate all other choices in each of those 4 cells, the B constraints. We could also look at this situation, where the cells are the A constraints and the columns and rows are the B's: Here, the same 4 digits are in those cells as the only candidates, but other candidates exist in the columns and rows in question.

- Code: Select all
`aA . . | . bB .`

. . . | . . .

. . . | . . .

---------------

. . . | . . .

cC . . | . dD .

. . . | . . .

If the digits involved are such that they line up into these two columns and these two rows, then other choices for each digit can be eliminated from the respective column or row they share.

This is sort of a forcing chain and the same logic can be extended to larger chains. This invalidates my initial assumption that a 9x9 grid can have constraint subsets no bigger than 9. In fact, I don't know the upper bound for such a subset, except of course for 2N^2. As I mentioned on the programmer's forum, there's also an interesting twist on this that covers situations advanced coloring doesn't, where a 3x3 set of cells might happen to converge with 9 digit-in-house constraints.