Remarks on Example 10
(with apologies)
This
example shows that, in the case of tournaments, the ranking by choosing
procedure induced by the Slater choice procedure is not very weakly monotonic.
This
example was obtained following a trial and error heuristic search process early
in 1995. At that time, I used the very nice piece of software by Alain
Guénoche, called A.B.C.D., that, for small instances, enumerates all the linear
orders at minimal distance from a tournament using a B&B algorithm
described in Barthélémy J.P., Guénoche A., Hudry O. (1989), Median linear
orders: Heuristics and a branch and bound algorithm, European Journal of
Operational Research 42, 313-325.
The
guiding idea in this heuristic search was to find a tournament having
"many" Slater winners and for which the linear orders at minimal
distance would be relatively far from the tournament. I expected that in such a
case, increasing the position of an alternative could lead to violations of
(weak) monotonicity of the ranking by choosing procedure.
I indeed
found such an example (with 9 alternatives, the linear orders at minimal
distance being at distance 10 from the tournament, whereas it is well know that
with 9 alternatives, the maximal possible distance is 12). This is example 10
reported in the paper. At that time, I used Alain Guénoche’s software to solve
the optimization problem. Unfortunately, this software only runs on
"ancient" Macintosh computers, which are not likely to be much
available nowadays.
When
preparing the paper, I checked the results obtained in 1995 using a standard
Integer Linear Programming formulation of the problem using a commercial ILP
code (CPLEX). I then looked for an analytical proof of reasonable length,
which, I confess, I have been unable to find. Since I think that the example is
interesting, I decided nevertheless to include it in the paper. I provide below
a "simple" but quite ugly "proof" that it works. I
would clearly be nice to find an elegant analytical proof. It would also be
interesting to investigate whether this example is minimal.
Consider
the tournament T on X defined by:
1 T 2, 1 T 5, 1 T 7, 1 T 8, 1 T 9,
2 T 3, 2 T 5, 2 T 6, 2 T 7, 2 T 9,
3 T 1, 3 T 4, 3 T 5, 3 T 6,
4 T 1, 4 T 2, 4 T 5, 4 T 9,
5 T 6, 5 T 8,
6 T 1, 6 T 4, 6 T 8, 6 T 9,
7 T 3, 7 T 4, 7 T 5, 7 T 6, 7 T 8,
8 T 2, 8 T 3, 8 T 4, 8 T 9,
9 T 3, 9 T
5, 9 T 7.
It is not
difficult to check that 5 is covered by 7 and that 9 is covered by 2. None of
them can be therefore at the top of a Slater order.
Considering
the scores of each alternative (it is well known that only alternatives having
a score greater or equal than the integer part of n/2 (here 4) can be Slater
winners) leaves again all alternatives except 5 and 9 as potential winners.
It is also
not difficult to check that this Tournament has 9 arc-disjoint circuits:
9 3 6 9
9 5 8 9
9 7 4 9
8 4 1 8
8 2 6 8
7 6 1 7
6 4 5 6
4 2 3 4
1 2 7 3 1
The orders
at minimal distance must therefore be at least at distance 9 from T.
We assert
in the paper that linear orders at minimal distance of T are at distance 10,
that there are exactly 40 such orders and that we have SL(X, T) = {1, 2, 4, 6,
7, 8}.
Lacking
any analytical proof of decent length, and since not all reader will have a
commercial LP package available, I wrote a few lines of code enumerating all
Hamiltonian paths of T (although this is definitely not nice, this, at least
gives a check that the example works while not requiring the use a commercial
LP software). It is well known that only such paths can be Slater orders. There
are exactly 2383 such paths (in order to keep the code as simple as possible, I
did not take into account the fact that 5 and 9 cannot be winners). For each of
them, I computed the distance wrt T. The list of all these paths (sorted by
increasing distance with T) is here (Excel file). This
enumeration only takes a few seconds on an ordinary PC.
It is
clear that the restriction of T to {3, 5, 9} is the linear order 9 > 3 >
5.
Hence we
have 9 >_{SL}(T) 3.
Consider
now the tournament V identical to T except that 9 V 1.
I assert
in the paper that there at 11 linear orders at minimum distance (this minimal
distance is 10) from V and that SL(X, V) = {2, 7, 8}. Again, I have no nice
proof of this fact and I resort to the same ugly enumeration trick as before. V
has exactly 2537 Hamiltonian paths. The complete list (sorted by increasing
distance with V) of these paths can be found here (Excel file).
I also
assert that SL(X\{2, 7, 8}, V) = {3}, so that we now have 3 >_{SL}(V) 9
although V improves the position of 9 wrt T, which shows that SL does not
induce a very weakly monotonic ranking by choosing procedure.
There is
only one linear order at minimum distance (2) from this reduced tournament. This
is easily proved. Indeed, we have the following reduced tournament:
1 V 5,
3 V 1, 3 V
4, 3 V 5, 3 V 6,
4 V 1, 4 V
5, 4 V 9,
5 V 6,
6 V 1, 6 V
4, 6 V 9,
9 V 1, 9 V
3, 9 V 5.
Considering
the scores of the alternatives, it is clear that 1 and 5 cannot be Slater
winners. Observe that any order having 3 on top is at least at distance 1 from
V. Considering now the remaining alternatives, there is no Condorcet winner
once 3 is removed.
Hence, an
order at minimal distance from V having 3 on top must be at least at distance 2
from V. There is one such order at distance 2 from V:
3 > 6 > 4 > 9 > 1 > 5.
Let us now
prove that 3 is the unique Slater winner, i.e. that any order having 4, 6 or 9
on top is at least at distance 3 from V. Let us take the case of alternative 4.
Alternative 4 is beaten by 3 and 6. hence the minimal distance must be at least
2. Now, there is no Condorcet winner once 4 is removed. Hence the minimal
distance is at least 3. The case of
alternatives 6 and 9 is deal with similarly.