You are on page 1of 1

108 Problem reduction

procedure deals with j-constraints rather than just binary constraints. It may be
worth noting that the primal graph need not be represented and modified in the pro-
cedure. It was only mentioned above to help explain the algorithm.

Theorem 4.1
If adaptive-consistency is achieved in a CSP under an ordering, then a search
under this ordering is backtrack-free:

∀ csp((Z, D, C)): (∀ <: total_ordering(Z, <):


(adaptive-consistent((Z, D, C),<) ⇒ backtrack-free((Z, D, C), <)))

Proof (see [DecPea88a])


Assume that the variables are given the ordering x1, x2, ..., xn, and adaptive-
consistency has been achieved under this ordering. At any stage of a search,
a (possibly empty) sequence of variables x1, x2, ..., xk have been consistently
labelled. Let S be the set of parents of xk+1. When xk+1 is being labelled,
there are only two possibilities:

(1) The domain of xk+1 is an empty set, in which case the search may ter-
minate with failure being reported. Note that this can only be the case if
xk+1 has no parents (i.e. S is an empty set). This is because if S is non-
empty, then there must exist a constraint CS which is an empty set
(because no compound label for S is compatible with any value for
xk+1), and therefore S could not have been consistently labelled (which
contradicts the assumption).
(2) If the domain of xk+1 is nonempty, then since the parents of xk+1
(which could be an empty set) have been consistently labelled (by
assumption), there must exist a value for xk+1 which is compatible with
all its parents (because every compound label in CS has a compatible
value in xk+1).
In both cases, no backtracking is required.
(Q.E.D.)

Definition 4-4:

The primal graph of a CSP P after adaptive-consistency is achieved under


some ordering of the variables (i.e. possibly with new edges added) is called
the induced-graph of P under that ordering, denoted by induced-graph(P,
Ordering):

You might also like