1212G5346785349111078134G 1 6951778 102 4G11210edge appearing in both <strong>of</strong> G 1 and G2edge between the nodes that appearin <strong>on</strong>ly <strong>on</strong>e <strong>of</strong> G 1 and G25917382 4911106(a) Uni<strong>on</strong>11611(b) Check <strong>of</strong> separatenessFigure 6: Markovian combinati<strong>on</strong> <strong>of</strong> graphs. In panel (a), there are 12 green edges for the pairs <strong>of</strong>nodes that do not appear in the same graph <strong>of</strong> G 1 or G 2 . Three <strong>of</strong> the green edges remain in panel(b).node-separateness in both <strong>of</strong> the graphs.The combined graphs which are obtained through the two operati<strong>on</strong>s are maximal CMS’s <strong>of</strong> agiven set <strong>of</strong> Markovian subgraphs as shown in the theorem below.THEOREM 7. The combinati<strong>on</strong> process by the two operati<strong>on</strong>s <strong>of</strong> Uni<strong>on</strong> and Check <strong>of</strong> separatenessproduces a maximal CMS.Pro<strong>of</strong>. Let M be a set <strong>of</strong> Markovian subgraphs <strong>of</strong> a graph. The “Uni<strong>on</strong>” operati<strong>on</strong> puts an edgebetween a pair <strong>of</strong> nodes, u and v say, unless u and v are both in a graph in M and separated therein.Denote the graph from this operati<strong>on</strong> by G ∗ . It is obvious that G ′ ⊆ e G ∗ V (G ′ ) for every G′ ∈ M.The “Check <strong>of</strong> separateness” operati<strong>on</strong> removes edges from G ∗ in such a way that the followingc<strong>on</strong>diti<strong>on</strong> is satisfied for every G ′ in M:For any pair <strong>of</strong> n<strong>on</strong>-adjacent nodes u and v in G ′ and a set C in G ′ which is disjoint with {u, v},〈u|C|v〉 G ′ if and <strong>on</strong>ly if 〈u|C|v〉 G ∗. (6)Denote a graph obtained from this check-<strong>of</strong>-separateness by G ∗∗ . Then any pair <strong>of</strong> n<strong>on</strong>-adjacentnodes, u and v say, in G ∗∗ mean either (i) that they are n<strong>on</strong>-adjacent in at least <strong>on</strong>e <strong>of</strong> the graphsin M or (ii) that they bel<strong>on</strong>g to different graphs each and putting an edge between the nodes incursc<strong>on</strong>flicts with the node-separateness in some <strong>of</strong> the graphs in M.Therefore, adding any edge to G ∗∗ into another graph G ′′ disqualifies G ′′ as a CMS <strong>of</strong> the graphsin M. This means that G ′′ is a maximal CMS <strong>of</strong> the graphs in M.13
2213 1⇐⇒3 1 3(1a)(1b)Interacti<strong>on</strong> graphs (G 1 ) <strong>of</strong> X 1 , X 2 , X 3 G ′222134134 134⇐⇒(2a) (2a’) (2b)Interacti<strong>on</strong> graphs (G 2 ) <strong>of</strong> X 1 , · · · , X 4 G ′′Figure 7: Some simple examples where each <strong>of</strong> the graphs in the right column are not a Markoviansubgraph <strong>of</strong> any <strong>of</strong> the graphs <strong>on</strong> the left-hand side <strong>of</strong> ⇐⇒.1347 FURTHER DISCUSSIONIn Theorem 2, we are given a set <strong>of</strong> Markovian subgraphs <strong>of</strong> G. But in reality, we are <strong>of</strong>tengiven a set <strong>of</strong> <strong>marginal</strong> <str<strong>on</strong>g>model</str<strong>on</strong>g> <str<strong>on</strong>g>structures</str<strong>on</strong>g> that are assumed to be interacti<strong>on</strong> graphs <strong>of</strong> the <strong>marginal</strong><str<strong>on</strong>g>model</str<strong>on</strong>g>s. The interacti<strong>on</strong> graphs may not be Markovian subgraphs <strong>of</strong> the unknown G. In this case,maximal CMS’s may not c<strong>on</strong>tain G as an edge-subgraph. Simple examples <strong>of</strong> this situati<strong>on</strong> aredisplayed in Figure 7. In the first row <strong>of</strong> the figure are two interacti<strong>on</strong> graphs (G 1 ) for X 1 , X 2 , X 3and a subgraph G ′ which is not Markovian with respect to G 1 , and similarly in the sec<strong>on</strong>d row forX 1 , · · · , X 4 . Under the hierarchy assumpti<strong>on</strong> for c<strong>on</strong>tingency tables, n<strong>on</strong>e <strong>of</strong> the graphical loglinear<str<strong>on</strong>g>model</str<strong>on</strong>g>s (1a), (2a), and (2a’) is compatible with the graphical sub<str<strong>on</strong>g>model</str<strong>on</strong>g>s at the right ends <strong>of</strong>the corresp<strong>on</strong>ding rows by Theorem 2.3 <strong>of</strong> Asmussen and Edwards (1983). The <str<strong>on</strong>g>model</str<strong>on</strong>g> G ′ in Figure7 is possible with the graphical log-linear <str<strong>on</strong>g>model</str<strong>on</strong>g> (1b) in the figure whenE[(P (X {1,3} = x {1,3} |X 2 )] = P (X 1 = x 1 )P (X 3 = x 3 ) for all x {1,3} ∈ X {1,3} , (7)where X i is the support <strong>of</strong> X i and X a = ∏ i∈a X i. The graphical log-linear <str<strong>on</strong>g>model</str<strong>on</strong>g> G ′′ in Figure 7 isalso possible from the graphical <str<strong>on</strong>g>model</str<strong>on</strong>g> (2b) in the figure. Instances <strong>of</strong> this phenomen<strong>on</strong> follow.Example 1. Probability distributi<strong>on</strong>s corresp<strong>on</strong>ding to some <strong>of</strong> the graphs in Figure 7. We willpresent c<strong>on</strong>tingency tables for which the pair <strong>of</strong> <str<strong>on</strong>g>model</str<strong>on</strong>g>s, (1b) and G ′ in Figure 7, are possible and soare the pair <strong>of</strong> <str<strong>on</strong>g>model</str<strong>on</strong>g>s, (2b) and G ′′ . When X i and X j are c<strong>on</strong>diti<strong>on</strong>ally independent given X k , wewill simply write i ⊥j|k.(a) C<strong>on</strong>cerning <str<strong>on</strong>g>model</str<strong>on</strong>g>s (1b) and G ′ :x 2 x 1 x 3 P (X = x) x 2 x 1 x 3 P (X = x)0 0 0 1/24 1 0 0 2/241 3/24 1 6/241 0 2/24 1 0 1/241 6/24 1 3/24This distributi<strong>on</strong> satisfies that 1 ⊥3|2 and 1 ⊥3.(b) C<strong>on</strong>cerning <str<strong>on</strong>g>model</str<strong>on</strong>g>s (2b) and G ′′ :14