Survey
* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project
Random and Quantum Walks on Graphs: an Introduction Simone Severini ([email protected]) AIMS June 12, 2012 References I A. Ambainis, Quantum walks and their algorithmic applications, arXiv:quant-ph/0403120v3 I R. Diestel, Graph Theory, Graduate Texts in Mathematics, Springer-Verlag, Heidelberg, 2010. http://diestel-graph-theory.com/ I N. Konno, Quantum Walks: Lecture at the School "Quantum Potential Theory: Structure and Applications to Physics" Held at the Alfried-Krupp-Wissenskolleg, Greifswald, 26 February - 9 March 2007. I L. Lovász, Random walks on graphs: a survey. http://www.cs.elte.hu/~lovasz/erdos.pdf M. Santha, Quantum walk based search algorithms. arXiv:0808.0059v1 [quant-ph] I I R. de Wolf, Quantum Computing: Lecture Notes, CWI Amsterdam. http://homepages.cwi.nl/~rdewolf/qwalksintro.pdf (*) Plan I Graphs I Random walks I Quantum walks Syllabus I Natural networks, Technological networks, Social networks, Graphs, Subgraphs, Isomorphisms, Degrees, Degree distributions, Regular graphs, Complete graphs, Clique number, Independence number, Chromatic number, Paths and cycles, Connectedness, Regularity Lemma, Ramsey numbers, Max-Flow Min-Cut, Visibility graphs, Graph widths, Adjacency matrix, Adjacency spectrum, Expansion, Random walks, Mixing/hitting, PageRank, Grover matrix, A special case: regular graphs, Coined quantum walks, Mixing, Continuous-time quantum walks, Mixing matrices, Applications: keywords, Find the missing link, A general framework (1/2), A general framework (2/2), A mathematical open problem Natural networks Technological networks Social networks Graphs Graphs De…nition A graph G = (V , E ) is a pair of two sets: the elements of V are called vertices; E V V ffi, i g : i 2 V gg and its elements are called edges. I Generalizations: digraphs (directed graphs); weighted graphs. Subgraphs De…nition A graph H = (W , F ) is a subgraph of a graph G = (V , E ) if W and F E. V Isomorphism De…nition Two graphs G = (V , E ) and H = (W , F ) are isomorphic (G = H) if there is a bijection f : V ! W such that fi, j g 2 E i¤ ff (i ), f (j )g 2 F . The number of isomorphism classes on n vertices is at most (1 + o (1)) 2n (n 1 )/2 /n!. De…nition A map on graph is called a (complete) graph invariant (i¤) if it assigns equal values to isomorphic graphs. Degrees De…nition The neighbourhood of a vertex i is N (i ) := fj : fi, j g 2 E g. The nonnegative integer dG (i ) = jN (i )j is the degree of a vertex i. The maximum degree is denoted by ∆ (G ); the minimum degree by δ (G ). Lemma (Degree sum formula) In a graph G , ∑i 2V dG (i ) = 2 jE j. Lemma In any graph, there are two vertices of equal degree. Lemma (Handshaking lemma) The number of vertices of odd degree in a graph is even. Degree distributions De…nition The degree sequence of a graph G is the list of its degrees f dG ( i ) : i 2 V g . De…nition The degree distribution of a graph G on n vertices is the list fPG (k ) : k = 0, 1, ..., n 1g, where PG (k ) = nk /n and nk = jfi : dG (i ) = k gj. De…nition A unigraph is a graph characterized by its degree sequence. Degree distributions: example Regular graphs De…nition A graph G is k-regular if dG (i ) = k for every vertex i. Complete graphs De…nition A complete graph Kn is an (n 1)-regular graph on n vertices. Clique number De…nition A clique is a complete induced subgraph. De…nition The clique number ω (G ) of a graph G is the order of its largest clique. Independence number De…nition A set S V is an independent set in a graph G = (V , E ) if fi, j g 2 /E for every i, j 2 S. De…nition The independence number, α (G ), is the size of the largest independent set. Chromatic number De…nition The chromatic number χ (G ) of a graph G is the minimum number of colours for which every two adjacent vertices have a di¤erent colour. Lemma ω (G ) χ (G ) ∆(G ) + 1. Paths and cycles De…nition A path Pn = (V , E ) of length n is a graph of the form V = f1, 2, ..., n g and E = ff1, 2g, f2, 3g, ..., fn 1, n gg. De…nition A cycle Cn = (V , E ) of length n is a graph of the form V = f1, 2, ..., n g and E = ff1, 2g, f2, 3g, ..., fn 1, n g, f1, n gg. Connectedness De…nition A graph is connected if there is a path between any two of its vertices. De…nition A maximal connected subgraph of a graph G is said to be a connected component. Regularity Lemma (1) De…nition Given a graph G = (V , E ), let X , Y V with X \ Y = ∅. Let kX , Y k = jffi, j g 2 E : i 2 X , j 2 Y gj. The density of the pair (X , Y ) kX ,Y k is de…ned as d (X , Y ) = jX jjY j . De…nition Given a graph G = (V , E ), let A, B V with A \ B = ∅. Then (A, B ) is an e-regular pair if for all X A and Y B with jX j e jAj and jY j e jB j we have jd (X , Y ) d (A, B )j e. De…nition Let fV0 , V1 , ..., Vk g be a partition of V . The set V0 (possibly empty) is called an exceptional set. The partition is called an e-regular partition if: (i) jV0 j e jV j; (ii) jV1 j = ... = jVk j; (iii) all but at most ek 2 of the pairs Vi , Vj with 1 i < j k are e-regular. Regularity Lemma (2) Theorem (Szemerédi, 1978) For every e > 0 and every integer n 1 there exists an integer N such that every graph on at least m vertices has an e-regular partition fV0 , V1 , ..., Vk g with n k N. I The intuition behind Szemerédi regularity lemma is that every large enough graph can be divided into subsets of about the same size so that the edges between di¤erent subsets behave almost randomly. I see J. Komlós, A. Shokoufandeh, M. Simonovits, E. Szemerédi, LNCS, 2292 (2002), 84–112. Ramsey numbers I Ramsey Theory studies conditions when a combinatorial object contains necessarily some smaller given objects. De…nition The Ramsey number R (k; l ), where k and l are arbitrary integers, is de…ned as the smallest n such that for a graph G with n or more vertices ω (G ) = k or α (G ) = l. (Ramsey’s theorem guarantees this.) Example In any group of six people, it is possible to …nd three persons who all know each other or three persons who do not know each, i.e., 2 4k . p R (3, 3) = 6; R (k, k ) (2k k 1) k I F. P. Ramsey, Proceedings of the London Mathematical Society, 30 (1930), pp. 264–286. Max-Flow Min-Cut De…nition In a digraph D = (V , E ), a vertex of in-degree 0 (resp. out-degree 0) is called a source (resp. sink). Each arc (i, j ) has a capacity ci ,j 0. A 0 ‡ow is a map f : E ! R such that: (i) f (i, j ) ci ,j ; (ii) ∑j f (i, j ) = ∑j f (j, i ), for every i 2 V . Problem Maximize ftot = ∑j f (1, j ) = ∑j f (j, n ), where 1 is the source and n is the sink. De…nition A cut in a digraph is a set of arcs such that if they are deleted, there is not path from the source to the sink (see edge cut). The capacity of a cut is de…ned to be the sum of the capacities of every arc in the cut. Theorem (Max-Flow Min-Cut) In every digraph, the maximum ‡ow equals the minimum capacity of a cut. I Ford–Fulkerson, Elias–Feinstein–Shannon, 1956. Visibility graphs De…nition Let X = (x1 , x2 , ..., xn ) be a sequence with xi 2 R 0 for i = 1, 2, ..., n. Each element xi 2 X is associated to a line segment in R2 with end points (i, 0) and (i, xi ). The visibility graph of the sequence X , denoted by G (X ), is the graph with set of vertices f1, 2, ..., n g and an edge i j if, for any i < k < j, we have xk < xj + (xi xj ) jj ki . Graph widths I Most NP-hard problems are e¢ ciently solvable for graphs with bounded widths. Treewidth quanti…es how “close” a graph is to being a tree. De…nition A tree decomposition (T , X ) of a graph G = (V , E ) S consists of a tree T = (I , F ) and a function X : I ! 2V such that: (i) i 2I X (i ) = V ; (ii) for every fi, j g 2 E there is u 2 I such that fi, j g 2 X (u ); (iii) X (i ) \ X (k ) X (j ), for every j on the path between i and k. The width of a tree decomposition is maxi jX (i )j 1. The treewidth, tw (G ), is the minimum width over all tree decompositions. Random walks Adjacency matrix De…nition The adjacency matrix of a graph G = (V , E ) on n vertices is de…ned as follows: 1, if fi, j g 2 E ; [A (G )]i ,j = 0, otherwise. Lemma The number of walks of length l from vertex i to vertex j in a graph G is given by [A (G )l ]i ,j . I Canonical labeling (lexicographic smallest adjacency matrix). I Adjacency list. (Adjacency) spectrum De…nition The spectrum of a graph G is the multiset of the eigenvalues of A (G ): λ1 λ2 λn . Theorem (i) If diam (G ) = d then G has at least d + 1 distinct eigenvalues. (ii) A (connected) graph G is bipartite i¤ λ1 = λn . (iii) χ (G ) 1 + λn /λ1 . (iv) For any graph with m edges, ∑i λ2i = 2m. (v) (Ho¤man’s bound) For a d-regular graph G , α (G ) n/(1 d /λn ). (vi) If G is d-regular then (1, 1, ..., 1) is an eigenvector and d is an eigenvalue. Theorem (Babai-Grigoryev-Mount, 1982) If each eigenvalue has multiplicity O (1), isomorphism can be tested in polynomial time. Fact If two graphs are isomorphism then they are cospectral. The converse does not always hold. [Example] Expansion De…nition An ε-expander is a graph G = (V , E ) in which, for every S V with jS j jV j /2, the number of vertices in V S adjacent to some vertex in S is at least ε jS j. Theorem (Alon, 2001) Le G be a d-regular graph. (i) If d λ2 2εd then G is an ε-expander. (ii) If G is an ε-expander then d λ2 ε2 /5. Example V = Zp , p prime; fi, j g 2 E i¤ j = i + 1 and j = i 1 , for each i 2 Zp . De…nition A d-regular graph is said to be Ramanujan if λ2 Theorem (Friedman, 2003) For almost every d-regular graph λ2 tends to 0 as n ! ∞). I p 2 d p 2 d 1. 1 + o (1) (the o (1)-term S. Hoory, N. Linial, and A. Wigderson, Bull. Amer. Math. Soc. 43 (2006), no. 4, 439–561. Random walks De…nition The transition matrix of a graph G = (V , E ) on n vertices is de…ned as follows: 1/dG (i ) , if fi, j g 2 E ; [T (G )]i ,j = 0, otherwise. The matrix T (G ) is stochastic (i.e., nonnegative with row sums 1). Then, 1 λi (T ) 1, for every i. De…nition A (simple) random walk on a graph G = (V , E ) on n vertices is a sequence of vectors fv(1 ) , v(2 ) , ...v(k ) g de…ned by the rule T k (G ) ei = v(k ) , where ei is the i-th element of the standard basis of Rn . Theorem If G is connected and χ (G ) 6= 2 then limk !∞ v(k ) π G = 0, TV where π G is the (unique) stationary distribution given by [π G ]i = dG (i ) /2 jE j. Mixing/hitting Theorem The mixing rate is the rate of convergence of v(k ) towards π G and it is determined by 1 λ2 (T ), the eigenvalues gap of the graph. De…nition In a random walk starting from vertex i, the hitting time Hi ,j is the expected number of steps before vertex j. Theorem Let G be a (connected) graph with m edges. Then, Hi ,j = 2m ∑ k =2 1 1 λk (T ) [λk ]2i d (i ) [ λk ]i [ λk ]j p d (i )d (j ) ! , where λk (T ) is the k-th eigenvalue of the transition matrix, [λk ]i is the i-th component of its normalized eigenvector. PageRank I A digraph G = (E , V ); I Rank: p (j ) = ∑i :(i ,j )2E p (i )/dout (i ), p = (p (1), ..., p (n )). I Wout = Aout Dout1 , where A := AT (G ) and Dout is out-degree matrix. I PageRank vector: pWout = p. I Random jump: p = α n1 1 + (1 α)Wout p, with 0 < α < 1, so that the probability of reaching every vertex is at least α/n. I t 1. Quick convergence: p = nα ∑t 0 (1 α)t Wout P. Bonacich, American Journal of Sociology, 92:1170–1182, 1987; L. Page, S. Brin, R. Motwani, and T. Winograd, TR, CS Dept., Stanford University, 1998; J. Kleinberg, Journal of the ACM, 48:604–632, 1999. I Quantum walks Grover matrix De…nition ! The line digraph of G = (V , E ) is a digraph D G = (V 0 , E 0 ), (i, j ), (j, i ) 2 V 0 for each fi, j g 2 E and ((i, j ), (k, l )) 2 E 0 if j = k. De…nition Given a graph G = (V , E ), we denote by U (G ) the matrix de…ned as ! follows: for al (i, j ), (k, l ) 2 V ( D G ), [U (G )](i ,j ),(k ,l ) = 2/d (j ) 0, δi ,l , if j = k; otherwise. Fact (i) We have U (G ) U T (G ) = I . (ii) Let [U p (G )]i ,j := 1 if [U p (G )]i ,j > 0 and [U p (G )]i ,j := 0, otherwise. Then the spectrum of U 3 (G ) is a complete graph invariant for regular graphs on n 13 vertices. I I Remark: "The secret of quantum algorithms is not entanglement but minus signs." (Eddie Farhi, 11/06/2012) D. Emms, et al., Pattern Recognition 42(9): 1988-2002 (2009). A special case: regular graphs I Let G be a k-regular graph on n vertices. The adjacency matrix A(G ) can be decomposed into permutation matrices: k A(G ) = ∑ Pi . i =1 I Let C be a k k unitary matrix. Then the matrix UC ( G ) = k M Pi ( C In ) i =1 is also unitary. In the previous case, [C ]i ,j = 2/k C is often called the coin. δi ,j . The matrix Coined quantum walks I We …x a correspondence between the vertices of a k-regular graph G on n vertices and the elements of the standard basis fj1i, j2i, ..., jnig for a space called vertex-space. Also, each permutation corresponds to an element of fj1i, j2i, ..., jk ig for a space called coin-space. De…nition A coined quantum walk on G starting from jk i with coin-state jl i is de…ned by the rule UC (G )t (jk i jl i) = UC (G )UC (G )t 1 (jk i jl i) = jψt i, for any t. De…nition The probability of observing the walker at vertex j after t steps is given by k Pr[j jt] = ∑ j(hj j i =1 hi j)jψt ij2 . Mixing Fact A coined quantum walk does not mix. (In general, while random walks are distribution-preserving, quantum walks are norm-preserving. De…nition Two notions of mixing: (i) Instantaneous mixing time (choose a time t and measure); Average mixing (take the Cesaro mean) M (1, t ) = 1s ∑si=1 Pr[j jti ]. Theorem (Moore-Russell 2001) The mixing time of a random walk in the hypercube is Θ(n log n ). The mixing time of a coined (with Grover coin) quantum walks in the hypercube is (π/4)n. Continuous-time quantum walks I We …x a correspondence between the vertices of a graph G on n vertices and the elements of the standard basis fj1i, j2i, ..., jn ig for a space of dimension n. De…nition A continous-time quantum walk on G starting from jj i with is de…ned by the rule e iA (G )t jj i = jψct i, for any t. De…nition The probability of observing the walker at vertex j after a walk of time t is given by Pr[j jt] = jhj jψct ij2 . Mixing matrices I I De…ne a uni-stochastic matrix M (t ) := U (t ) U ( t ), where U (t ) := e iA (G )t . The average mixing matrix is R b (G ) := limT !∞ 1 T M (t )dt = ∑r Er 2 , where each Er is an M T 0 idempotent in the spectral decomposition of the adjacency matrix. b (Cn ) = n 21 J + 1 I , where J is E.g., for the n-cycle with n odd M n n the all-ones matrix. Theorem The average mixing matrix of a graph is rational. I We have uniform mixing at time t if U (t ) is ‡at, i.e., all entries of U (t ) have the same absolute value. Theorem If uniform mixing occurs on Cn then 4jn and n is the sum of two squares. Problem Determine which graphs have uniform average mixing. Applications: keywords I quantum algorithms I I universal quantum computation models of quantum information transport in spin systems I models of energy transfer in large molecules Find the missing link De…nition The (combinatorial) Laplacian matrix vertices is de…ned as follows: 8 < dG ( i ) , 1, [L (G )]i ,j = : 0, of a graph G = (V , E ) on n if i = j; if fi, j g 2 E ; otherwise. Theorem Let Kn be a complete graph on n = 4m vertices. A deleted edge from Kn can be found with certainty by the use of O (n 1) continuous-time quantum walks induced by the Laplacian, each one of time π/2. A general framework (1/2) I Let T (G ) be the transition matrix of a regular graph G on n vertices. Then λ1 = 1 and 1 λi < 1, for i 6= 1. Let u := π G = ( n1 , ..., n1 )T be the eigenvector of λ1 . Let jλi j 1 The number δ is called the spectral gap. δ. 2 I T k (G ) v u kv k2 (1 δ)2k . If δ is not too small then we get quick convergence independently of v . I Let us mark an ε-fraction of the vertices. Once we are close to the uniform distribution, we have probability roughly ε of hitting a marked vertex. Suppose it costs S to set up an initial distribution v , it costs U to perform one step of the random walk, and it costs C to check whether a given vertex is marked. I I Algorithm: Check if the current vertex is marked; if not run a random walk for roughly 1/δ steps. I The total cost of the algorithm is S + 1ε (C + 1δ U ). A general framework (2/2) I A procedure based on quantum walks gives S + p1ε C + p1 U . I The choice of the graph is important to encode the problem. I Triangle …nding, element distinctness, etc. I Lots of potential: e.g., graph isomorphism problem? I [See Lecture 7 in Ronald de Wolf’s Course] δ Mathematical open problem Problem Given a graph G . Find necessary and su¢ cient conditions such that there is a unitary matrix U for which [U ]i ,j 6= 0 if and only if [A(G )]i ,j = 1. [I.e., obtain a matrix with the same structure of A(G ) and with max. rank] [Open since the 1970s]