Wiki
Wiki

Research notes on every problem, and a library of the papers behind them. Built from the open erdos repository.

Updated


Source. Item (29), "Correctness and bound," pp. 59--60, with the algorithm (19)--(28) on pp. 58--59 and Item (30) on p. 60, of W. H. Cunningham and A. B. Marsh III, "A primal algorithm for optimum matching," Mathematical Programming Study 8 (1978), 50--72, https://doi.org/10.1007/BFb0121194. The edition read is identified on the source card.

Statement

Setting (pp. 52--57). The problem is to find a perfect matching MM of GG maximizing ∑(cj:j∈M)\sum(c_j:j\in M), with the odd-set dual program (3) of p. 52 (see Theorem (46)). For a pair (y,Y)(y,Y) write dj=∑(yv:j∈δ(v))+∑(YS:j∈γ(S))−cjd_j=\sum(y_v:j\in\delta(v))+\sum(Y_S:j\in\gamma(S))-c_j. The algorithm keeps a real vector yy, a nonnegative vector YY, a shrinking family S\mathcal S of GG and a perfect matching MM of the graph obtained from GG by shrinking the maximal members of S\mathcal S, subject to (17): dj=0d_j=0 for j∈Mj\in M and for every edge jj of the odd polygon P(S)P(S) kept with each S∈SS\in\mathcal S; and (18): YS=0Y_S=0 for S∉SS\notin\mathcal S. It does not require dj≥0d_j\ge0 until the end (p. 53).

Item (29) (pp. 59--60). Each step of the primal algorithm preserves these properties; if the algorithm terminates, its final step (28) (p. 59) extends MM to a perfect matching M1M_1 of GG that is optimal, with (y,Y)(y,Y) optimal for (3). The algorithm terminates: the choice of vertex uu in step (19) happens at most ∣V(G)∣|V(G)| times; there are at most 32∣V(G)∣2\tfrac32|V(G)|^2 occurrences of steps (20)--(24) and at most ∣V(G)∣|V(G)| occurrences of (25)--(27); and the paper establishes a computation bound of O(∣V(G)∣2⋅∣E(G)∣)O(|V(G)|^2\cdot|E(G)|).

Item (30) (p. 60) adds that a bound of O(∣V(G)∣3)O(|V(G)|^3) can be achieved with considerable care, without giving the details. Item (31) (pp. 60--61) explains how to start when no perfect matching is known, by adding artificial edges of sufficiently small weight.

Read depth. Claims checked: the algorithm and the argument of (29) were read on the print.

Proof pointer

pp. 59--60. Once dj≥0d_j\ge0 holds for an edge it is never lost, and uu is changed only when every edge at uu satisfies it, which bounds the number of stages by ∣V(G)∣|V(G)|. Within a stage each tree-growing, shrinking or expanding step raises ∣O(T)∣−∣I∣|O(T)|-|\mathcal I| by at least one, where O(T)O(T) is the set of vertices of GG in odd vertices of the tree and I\mathcal I the members of S\mathcal S not inside an odd pseudo vertex; the nested-family bound (9) (p. 54) confines this quantity to a range of length about 32∣V(G)∣\tfrac32|V(G)|, which bounds the steps in a stage.

Dependencies

Theorem (7) (p. 54): if S\mathcal S is a shrinking family of GG, every perfect matching of the graph obtained by shrinking its maximal members is contained in a perfect matching of GG; and the bound (9) (p. 54) on nested families.

Bears on

No Erdős problem is recorded for this result. It underlies Theorem (46).