Optimal Networks Mirka Miller School of Information Technology and Mathematical Sciences University of Ballarat m.miller@ballarat.edu.au CIAO Centre for Informatics and Applied Optimization 1 Outline of the talk Interconnection networks Degree/diameter problem – directed and undirected Recent new results concerning graphs close to Moore bound Three extremal problems Open problems 2 Interconnection networks Examples: Communications Transportation Computer Social Networks can be modeled as graphs. The structure (topology) of the graph is useful when designing algorithms (for communication, broadcasting etc.) and for analyzing the performance of a network. 3 Small-world networks 4 WWW (2000) 5 Network parameters Topology Nodes and edges and their arrangement. Diameter Maximum distance between any pair of nodes. Connectivity Number of neighbours of a given node: d degree Clustering Are neighbours of a node also neighbours among themselves? 6 Most “real life” networks small-world small diameter Milgram 1967 clustered Watts & Strogatz 1998 scale-free degrees follow a power-law Barabási & Albert 1999 7 Network topology Structured high clustering large diameter regular |V| = 1000 d = 10 D = 100 C = 0.67 Small-world Random high clustering small diameter almost regular low clustering small diameter |V|=1000 d = 8-13 D = 14 C = 0.63 |V|=1000 d = 5-18 D=5 C = 0.01 8 Definitions We study graphs with respect to 3 parameters: diameter, degree and order. Diameter The longest distance between any two vertices in the graph. Degree The number of edges attached to a vertex. Order The number of vertices in the graph. 9 Degree/diameter problem Determine the largest number of vertices n of a graph G for given maximum degree d and diameter at most k. Survey by Miller and Siran “Moore bound and beyond: A survey of the degree/diameter problem” Research supported by a grant from Australian Research Council (ARC) CI’s: McKay and Miller 10 Degree/diameter problem Directed version: Determine the largest number of vertices n of a digraph G for given maximum out-degree d and diameter at most k. 11 Degree/diameter problem Approaches to attack the problem Increase the lower bound: by construction o o o o Voltage assignments Line digraphs Computer search Etc. Decrease the upper bound: by proving a graph with given parameters cannot exist 12 Upper bound A natural upper bound on the number of vertices n of digraph G with given maximum out-degree d and diameter at most k is: n Md,k = 1+d +d 2 + … + d k. v This bound is called the Moore bound. A digraph attaining this bound is called a Moore digraph. 13 Moore digraphs Plesnik & Znam ’74, Bridges & Toueg ’80 : Moore digraphs exist only for trivial cases, namely for d =1 (cycles of k+1 vertices) or k =1 (complete digraphs on d+1 vertices). 14 Moore digraphs Outline of the PROOF that (d,k)-digraphs do not exist when d > 1 and k > 1 : I+A+…+Ak=J where A is adjacency matrix of G, J is unit matrix, I is identity matrix J has eigenvalues n (once) and 0 (n-1 times) A has eigenvalues d (once) and n-1 roots of the characteristic equation 1+ x+x2+…xk = 0 x = 0 and roots of xk+1 -1 =0 Since tr(Aj)=0 for 1 jk, we obtain –d=-dk and so d=1 or k=1 are the only solutions. 15 Diregularity of AMDs Is Md,k–1 attainable for all d>1 and k>1? Notation. (d,k)-digraph is a digraph of maximum out-degree d, diameter k and order n = Md,k – 1. (We also call such a digraph almost Moore digraph). Miller, Gimbert, Siran & Slamin, ‘00: The (d,k)-digraphs are diregular of degree d. Note that: to show the regularity of out-degree is easy (by a counting argument). However, to show the regularity of in-degree is not easy. 16 Repeat of a vertex Let G be a (d,k)-digraph. For every vertex x of G there exists a vertex y, called the repeat of x, such that there are two walks of lengths k from x to y. If r(x) = x then vertex x is called a selfrepeat. The function r : V(G) V(G) is an automorphism on V(G); namely, (x,y) E(G) iff (r (x), r (y)) E(G). Let the order of vertex v be the smallest positive integer (v) so that r (v)(v)= v. 17 Diameter 2 Fiol, Allegre, and Yebra ’83: (d,2)-digraphs exist for any d 2. Example: The line digraph of L(Kd+1) of the complete digraph Kd+1. But Gimbert ’01 showed that this line digraph is the only (d,2)-digraph if d 3. 18 Diameter 2, degree 2 There are exactly three (2,2)-digraphs. 1 2 1 3 4 4 5 3 2 1 4 5 2 5 6 3 6 (1)(2)(3)(4)(5)(6) (123)(456) All selfrepeats All order 3 6 (12)(3456) Two order 2; four order 4 19 Degree 2 and degree 3 Miller and Fris ’92: There are no (2,k)-digraphs for any k 3. Baskoro, Miller, Siran & Sutton (in press): There are no (3,k)-digraphs for any k 3. The remaining cases are still open: Do there exist any (d,k)-digraphs, d 4, k 3? 20 Existence of (d,k)-digraphs A (d,k)-digraph, d 4, k 3 (if it exists) may contain a selfrepeat or no selfrepeat. Further study may focus on the existence of: (d,k)-digraphs with selfrepeats (d,k)-digraphs with no selfrepeats 21 Structure of the orders of vertices A (d,k)-digraph contains either k selfrepeats or none. [Baskoro, Miller, Plesnik, ’98] 22 The orders of vertices We can determine the structure of the orders of vertices in a (d,k)-digraph with selfrepeats, d 2, k 2. [Baskoro, Cholily, Miller, ’04] 23 The orders of vertices Example. k=2, d=6. Label vertices 0,1,2,…,41. Suppose the digraph G contains a selfrepeat and that the outneighbourhood of a selfrepeat consists of vertices of orders 2 and 3 (as well as a selfrepeat). Then (up to isomorphism) the permutation cycles of repeats of G are (0) (1) (2,3)(4,5,6) (7,8)(9,10,11)(12,18)(13,19)(14,20) (15,21,16,22,17,23)(24,30,36)(25,31,37)(26,32,38)(27,33,39) (28,34,40)(29,35,41) Two cycles of length 1, five cycles of length 2, eight cycles of length 3, one cycle of length 6. 24 Open problems Are there any (d,k)-digraphs, d 4, k 3, with selfrepeats? Are there any (d,k)-digraphs, d 4, k 3, without selfrepeats? For d = 3, k 3, are there any digraphs of order M3,k – 2? For d = 2, k 3, are there any digraphs of order Md,k – 3? Are almost almost Moore digraphs diregular? 25 Degree/diameter problem Undirected version: Determine the largest number of vertices n of a graph G for given maximum degree d and diameter at most k. 26 Upper bound A natural upper bound on the number of vertices n of a graph G of given maximum degree d and diameter at most k is: n Md,k=1+d+d(d-1)+…+d(d-1)k-1 v This bound is called the Moore bound. A graph attaining this bound is called a Moore graph. 27 Moore graphs k =1: Moore graphs are complete graphs on d+1 vertices. Hoffman and Singleton, ’60: k =2: Moore graphs exist for d =2 (pentagon) or d =3 (Petersen graph) or d =7 (Hoffman-Singleton graph) or d =57? k =3: Moore graph exists for d =2 (7-gon). Damerell; Bannai and Ito, ’73: k >3: Moore graph are (2k+1)-gons. 28 Almost Moore graphs Is Md,k–1 attainable for all d>2 and k>1? Erdos, Fajtlowicz and Hoffman, ’80: k =2: almost Moore graph exists only for d =2 (4-cycle). Bannai and Ito; Kurosawa and Tsujii, ’81: k >3: almost Moore graphs exist only for d =2 (2k-gons). 29 Graphs with defect > 1 Defect 2: d =2: (2k-1)-gons. d >2: only 5 such graphs are known so far: (d,k) = (3,2) (two); (4,2); (5,2); (3,3) (unique). 30 Almost almost Moore graphs There are no almost almost Moore graphs of degree 3 and diameter k>3. [Jorgensen, ’92] Theorem 6. For k>2, there are no almost almost Moore graphs of degree 4. [Miller and Simanjuntak, ’04] 31 Repeat of a vertex Define (d,k,d)-graph to be a graph of degree d, diameter k and defect d. Vertex y is a maximal repeat of x if y appears in R(x) d times (x has no other repeats). Theorem 7. For d >1, the number of maximal repeats in a (d,2,d)-graph is 0 or 2 or 6. [Nguyen and Miller, ’04] 32 Structure of a (d,2,2)-graph Possible repeat configurations in a (d,2,2)-graph: u r1(u) u u r2(u) r(u) u u r1(u) r2(u) r1(u) r1(u) r2(u) r2(u) Define n0,n1,n2a,n2b,n2c. 33 Structure of a (d,2,2)-graph Theorem 7. A (d,2,2)-graph contains if d is even then n0 = 3 and n2b = d2 – 4 if d = 3 then (n0,n1,n2c) = (3,2,3) if d is odd then (n0,n1,n2c,n2a,n2b) = (9,6,9,4a,d2-25-4a) or n2b = d2 – 1. [Nguyen and Miller, ’04] 34 Open problems Are there any (d,2,2)-graphs for d 6? Are there any (d,k,2)-graphs for d 5 and k 3? Are there any (3,k,3)-graphs for k 4? Are there any (4,k,3)-graphs for k 3? . . . Is there a Moore graph with diameter 2 and degree 57? 35 Open problems We know that for directed graphs N(d,k) is monotonic in both d and k. Let K(n,d) be the smallest possible diameter of a digraph on n vertices and maximum out-degree d. Is K(n,d) monotonic in n? For undirected graphs we know that the corresponding K(n,d) is not monotonic in n. 36 Example. K(10,3) = K(8,3) = 2 n=8 n = 10 but K(9,3) = 3! n cannot be more than 10 37 Three optimisation problems = max{n: G(n,d,k)} D(n,k) = min{d: G(n,d,k)} K(n,d) = min{k: G(n,d,k)} N(d,k) G(n,d,k) denotes the set of all directed graphs of order n, degree d, and diameter k. Question: are these three problems equivalent to each other? 38 Known monotonic relationships d1<d2 implies N(d1,k)< N(d2,k) k1<k2 implies N(d,k1)< N(d,k2) d 1 <d 2 implies K(n,d 1 ) K(n,d 2 ) ? n1<n2 implies K(n1,d) K(n2,d) ? k1<k2 implies D(n,k1) D(n,k2) ? n1<n2 implies D(n1,k) D(n2,k) 39 K(n,d) problem digraph G of degree d, diameter k and order n line digraph construction digraph L(G) of degree d, order dn and diameter k+1 Vertex deletion scheme 40 K(n,d) problem Theorem [Slamin & Miller,’00] If L(G) G(dn,d,k) is a line digraph of a diregular digraph GG(n,d,k-1) then there exists digraph L(G) G(dn-r,d,k’), k’k, for every 1r (d-1)n-1 n nd 41 K(n,d) problem The largest n such that G(n,d,k-1) is not empty for given d. G(nd,d,k) The largest n’ such that G(n’,d,k) is not empty for given d. 42 K(n,d) problem K(n,d) is monotonic in n in some intervals of d values K(n,d) problem is equivalent to N(d,k) in those intervals N(d,k) K(n,d) D(n,k) For directed graphs, are these three problems in fact all equivalent? 43