I'm looking for the name of the combinatorial satisfying the following requirements:
- There are $n$ different tasks.
- There are $n$ different participants.
- There are $n$ phases of the experiment
- Exactly one participant occupies each task at each phase.
- At the end of the last phase, every participant has performed each of the tasks once.
So far, just a regular latin square. Here's the complication:
- At the start of each phase (except the first), the person who performed the task in the last phase explains the task to the person who is about to perform it.
- At the end of the last phase, every participant has explained one task each to every one of the other $n-1$ participants. (equivalently, has had one task explained to them by each of the other $n-1$ participants.
Examples: n=4, persons {A,B,C,D} and tasks {0,1,2,3}
Notation: Let the notation $2/AD/1$ stand for "At the start of phase 2, A teaches D how to perform task 1"
BAD DESIGN
phase A B C D
1st 1 2 3 4
2nd 2 3 4 1
3rd 3 4 1 2
4th 4 1 2 3
Why it fails: We have 2/AD/1, 3/AD/2, 4/AD/3, i.e. at the start of each phase, A always teaches D. In fact, this is maximally bad design, because all the pairs are fixed throughout the experiment.
GOOD DESIGN
phase A B C D
1st 1 4 3 2
2nd 2 3 1 4
3rd 3 2 4 1
4th 4 1 2 3
2/AC/1,2/BD/4,2/CB/3,2/DA/2
3/AB/2,3/BA/3,3/CD/1,3/DC/4
4/AD/3,4/BC/2,4/CA/4,4/DB/1
Every pairing appears exactly once, so this meets all the criteria.
What I know so far
Not much. I've found such designs for $n=2$ (trivial), $n=4$ and $n=6$, but (assuming my program is bug-free) no such design exists for $n=3$, $n=5$. My simple backtracking code bogs down by $n=8$, so I can't even get enough data to search the OEIS.
Is this a known design? Can such designs be found for all even n?
P.S. For the curious, here's an $n=6$ case:
0 1 2 3 4 5
1 2 0 4 5 3
3 4 5 1 2 0
2 0 1 5 3 4
5 3 4 0 1 2
4 5 3 2 0 1
No luck finding prior work on these so, provisionally, I'll call these "Maximally Social" Latin squares, with an eye on the motivating scenario in OP.
The first few cases already showed a lot of symmetry:
The second property means a dynamic programming algorithm can be used which makes finding such squares very fast (n=100 in well under a minute).
I've verified that:
The symmetry properties suggest that odd n solutions are not possible.
In Reduced form (first row and column are sorted), exhaustive search finds exactly one solution for $n=2$ (obviously) and $n=4$, and exactly two solutions for $n=6$ and $n=8$. The members of these solution pairs are related to one another by the simple transformation of flipping the inner nodes of the two center columns upside-down.
$n=6$:
$n=8$:
Here's the largest square that fits into SE's input box, n=30: