Valiant's Depth Reduction Lemma #
The length of a directed path is the number of nodes in it. The depth of a directed graph is the length of a longest directed walk.
Lemma (Valiant, 1977). In any directed graph with S edges and
depth d = 2 ^ k, for any 1 ≤ r ≤ k, it is possible to remove
r * S / k edges so that the depth of the resulting graph does not
exceed d / 2 ^ r.
Reference: L. G. Valiant, Graph-theoretic arguments in low-level complexity, MFCS 1977. Stated as Lemma 1.4 in Jukna, Boolean Function Complexity.
Note on cycles: depth is defined here as the supremum (in ℕ) of
walk lengths, with no injectivity requirement on the walk. In the
presence of a directed cycle, walks can be made arbitrarily long, so
the supremum's underlying set is unbounded; by ℕ's convention this
gives depth = 0. Consequently, G.depth ≤ 2 ^ k is satisfied
trivially (with F = ∅) for cyclic graphs, and is the meaningful
bound only on acyclic graphs — the case where Jukna's labeling
argument proceeds.
The proof machinery — canonical labelings, the edge partition by
first-differing bit, averaging, and the relabeling-after-removal
bound — lives in Circ.Internal.Valiant.
Valiant's Depth Reduction Lemma (Valiant, 1977).
In any finite directed graph G with S edges and depth at most
2 ^ k, for any r ≤ k, there exists a set F of edges such
that:
Fis a subset of the edge set,k * F.card ≤ r * S(equivalent to|F| ≤ r * S / kwithout integer division), and- after removing
F, the resulting digraph has depth at most2 ^ k / 2 ^ r.
For cyclic G the hypothesis is satisfied trivially (with depth = 0)
and the conclusion holds with F = ∅. The substantive content is the
acyclic case, handled via Jukna's canonical labeling argument.