Wiki
Wiki

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

Updated


Source. Sections 3 and 5, equations (5)–(16), printed pp. 126–128 (published original). The nested notation expands the source's successive contractions.

Statement

Let a feasible weighted hierarchy have a tight current matching Mˉ\bar M. It has a compatible lift MM to original edges such that every exposed current block chooses a minimum-weight child at each successive expansion. Define

yv=w(v)−∑B∋vdB,zB=2dB,(1)y_v=w(v)-\sum_{B\ni v}d_B,\qquad z_B=2d_B, \tag{1}

and set zS=0z_S=0 for odd sets not in the hierarchy. These variables are dual feasible, every edge of MM is tight, and each blossom BB contains exactly (∣B∣−1)/2(|B|-1)/2 edges of MM. Moreover,

U(y,z)−Wc(M)=∑A exposed in the current quotientw(A).(2)U(y,z)-W_c(M) =\sum_{A\text{ exposed in the current quotient}}w(A). \tag{2}

In particular, if every exposed current node has weight zero, MM is maximum weight among all original matchings.

Proof

First lift the matching. An odd circuit with one specified child unmatched has a near-perfect circuit matching omitting that child. If a current block is met by a matching edge, its remembered original attachment determines the child to omit. If it is exposed, choose a child of weight mBm_B. Apply the same rule recursively. The odd-circuit lifting lemma shows that this gives a matching with the prescribed external edges. It also proves inductively that a blossom of bb original vertices has (b−1)/2(b-1)/2 internal matching edges. All child sizes are odd, so every blossom size is odd. No block is met by more than one external matching edge.

For an original vertex vv in current block AA, the telescoping identity gives

yv=w(A)+aA(v)≥0.(3)y_v=w(A)+a_A(v)\ge0. \tag{3}

Also zB≥0z_B\ge0. If e=uve=uv joins two current blocks A,DA,D, no blossom contains both endpoints. Therefore its dual slack is

yu+yv−ce=w(A)+w(D)−cˉe≥0.(4)y_u+y_v-c_e =w(A)+w(D)-\bar c_e\ge0. \tag{4}

If instead BB is the smallest blossom containing both endpoints, let A,DA,D be its two distinct children containing them. All common ancestors contribute equally to the two subtractions in (1) and to the terms 2d2d in the dual inequality; those terms cancel. Applying the same telescope inside AA and DD leaves

yu+yv+∑S⊇{u,v}zS−ce=w(A)+w(D)−cˉe B,(5)y_u+y_v+\sum_{S\supseteq\{u,v\}}z_S-c_e =w(A)+w(D)-\bar c_e^{\,B}, \tag{5}

where cˉe B\bar c_e^{\,B} is the stored reduced weight before BB was contracted. This is nonnegative by the stored inequality. Thus every edge inequality is feasible.

Every edge selected in the lift is either a tight current matching edge or a remembered circuit edge at the first blossom containing its endpoints. Equations (4)–(5) therefore give equality on MM. For an exposed original vertex, the path through its containing blossoms follows a minimum child every time. Each summand in its offset is zero, so (3) says that its yy equals the weight of its exposed current block. There is exactly one such vertex in each exposed current block.

Sum the tight edge inequalities over MM. Since MM is a matching and every blossom is internally saturated, this gives

Wc(M)=∑v matched by Myv+∑B∣B∣−12zB.W_c(M)=\sum_{v\text{ matched by }M}y_v+ \sum_B\frac{|B|-1}{2}z_B.

Subtracting from UU proves (2). For any feasible primal vector xx, nonnegativity and the vertex and odd-set inequalities give

∑ecexe≤∑vyv∑e∋vxe+∑BzB∑e∈E(G[B])xe≤U(y,z).(6)\sum_e c_ex_e \le\sum_v y_v\sum_{e\ni v}x_e+ \sum_B z_B\sum_{e\in E(G[B])}x_e \le U(y,z). \tag{6}

This is weak duality proved directly. If all exposed current weights are zero, (2) and (6) certify equality for the matching indicator. No general strong-duality theorem is needed. □\square

The minimum-base rule also checks condition (e) of Theorem M. Whenever a circuit vertex is exposed in an intermediate graph, it lies on the recursively selected exposed chain and is a minimum-weight child. A block met by an external matching edge has no exposed vertex in that intermediate graph.

For later bookkeeping, (1) also yields

U=∑v∈Vw(v)−∑BdB,(7)U=\sum_{v\in V}w(v)-\sum_B d_B, \tag{7}

because zB=2dBz_B=2d_B has objective coefficient (∣B∣−1)/2(|B|-1)/2 and dBd_B is subtracted from ∣B∣|B| vertex variables. The original node weights in (7) may change while their vertices are current; this is not a claim that arbitrary intermediate weights are automatically bounded.