Sharp Bounds for Sums Associated to a Graph of Matrices James Mingo (joint work with Roland Speicher) Queen’s University at Kingston Brazilian Operator Algebras Symposium FlorianoĢpolis, 2011 [arXiv:0909.4277] 1 / 15 Graph Sums I G = (V, E) is a directed graph, e ∈ E runs from s(e) to t(e) I T : E −→ MN (C), e 7→ Te = (tij ) is an oriented assignment of matrices to edges (flip edge = transpose matrix) tij I (e) S(G, T) = X Y i (e) tit(e) is(e) j where the sum runs over all i:V→[N] e∈E functions i : V → [N] = {1, 2, 3, . . . , N}; ik = i(k) Examples i T T S(G, T) = Tr(T) = i X i tii S(G, T) = j X tij i,j 2 / 15 More Examples j T1 i T1 T3 S(G, T) = j k T2 X i l (1) (2) (3) tij tjk tjl S(G, T) = i,j,k,l T3 X k (1) (2) (3) tij tjk tki i,j,k T4 i4 i3 T3 S( G, T) = N X (1) (2) (3) (4) (5) (6) ti1 i2 ti3 i2 ti3 i4 ti4 i4 ti5 i3 ti2 i5 T2 i1 T2 T 1 i2 T5 i1 ,i2 ,...,i8 =1 T6 i5 T11 i8 T12 i7 T10 T8 T7 (7) (8) (9) (10) (11) (12) × ti6 i5 ti6 i5 ti6 i6 ti7 i5 ti8 i7 ti8 i7 i6 T9 3 / 15 the question (inspired by Bai & Silverstein (2006) (1) ) S(G, T) = X Y (e) tit(e) is(e) i:V→[N] e∈E I given G = (V, E) find r(G) such that for all N and all T : E → MN (C) we have |S(G, T)| 6 Nr(G) Y kTe k e∈E where kTk is the operator norm of T ∈ MN (C) I note that r(G) > 1 I answer: r(G) depends on the number of leaves of a tree (or union of trees) associated with G (1) but with a different answer, also Yin & Krishnaiah (1983), and Bai (1999) 4 / 15 motivation I Let {AN }N be a sequence of symmetric random matrices with i. i. d. centred entries and TN,1 , . . . TN,k a sequence of deterministic matrices (i.e. “just a matrix”) I we want to find the leading order terms of E Tr(T1 AT2 A · · · ATk−1 ATk ) I this leads to finding which for partitions π of [2k] the sum N X (1) (k) ti1 i2 · · · ti2k−1 i2k i1 ,...,ik =1 (subject to the constraint that ir = is whenever r and s are in the same block of π) will contribute to the leading order I this can easily be written as a graph sum of a graph Gπ 5 / 15 some notation, so we can say what r(G) is I let G = (V, E) be a directed graph I a cutting edge is one whose removal disconnects the connected component which contains it I a subgraph is two-edge connected if it cannot be disconnected by removing a single edge I a two-edge connected component is a two-edge connected subgraph not properly contained in another two-edge connected subgraph I F(G) is the graph whose vertices are the two-edge connected components of G and whose edges are the cutting edges of G I F(G) is a union of trees (= a forest) I a leaf of a tree (or forest) is a vertex with only one incident edge 6 / 15 The Theorem I let (G = (V, E) be a directed graph with F(G) the forest of two-edge connected components I if G is connected r(G) = max{1, `/2} where ` is the number of leaves of F(G) I in general, r(G) is the sum of the contributions of each connected component T i4 i3 T2 T1 j i T3 r = 3/2 i1 T1 T 1 i2 T2 T5 T6 i5 T11 l i T3 T2 j k 4 T3 r=1 k i8 T12 i7 T10 T8 T7 i6 T9 r = 3/2 7 / 15 Input-Output Graphs I I I I I I I I G = (V, E), a directed graph is a input-output graph if we can write V as a disjoint union: Vin ∪ Vout ∪ Vint , with Vin = input vertices; Vout = output vertices; and Vint = internal vertices = everything else, and G contains no directed cycles; each v ∈ Vint is on a directed path from a vertex in Vin to a vertex in Vout ; each input vertex has only outgoing edges; and each output vertex has only incoming edges. 8 / 15 The Operator TG when G = (V, E) is input-output I I I for each v ∈ V, let Hv = CN , ξi = [0, . . . , 1, . . . , 0] ∈ Hv , √ and ξ(v) = ξ1 + · · · + ξN = [1, . . . , 1, . . . , 1] ∈ Hv , kξk2 = N O a basis element of Hv is specified by a function v∈Vin j : Vin → [N] and the corresponding vector is O ξjv v∈Vin I define TG : O Hv −→ v∈Vin D O w∈Vout ξjw , TG O Hv by v∈Vout O ξiv E v∈Vint = X Y k:V→[N] e∈E hξkt(e) , Te ξks(e) i (e) (=tk t(e) ks(e) ) where the sum runs over all k : V → [N] such that k|V = i and in k|Vout = j O O (w) ξ(v) S(G, T) = ξ , TG w∈Vout v∈Vin 9 / 15 S(G, T) = O w∈Vout O ξ(w) , TG ξ(v) v∈Vin thus |S(G, T)| 6 N(#(Vin )+#(Vout ))/2 kTG k I moreover we can write TG = Lr Tr Lr−1 · · · T1 L0 where the Li ’s are partial isometries and each Ti is a tensor product of some Te ’s L0 T1 L1 T2 L2 T3 L3 T4 L4 10 / 15 conclusion of method I suppose our graph is where L∗ L" L : H → H ⊗ H is given by L(ξi ) = ξi ⊗ ξi and L 0 : H → H ⊗ H ⊗ H is given by L 0 (ξi ) = ξi ⊗ ξi ⊗ ξi Theorem Let G be a graph. By splitting vertices and adding I, flipping edges we can modify a graph to obtain G 0 so that I the graph sums are the same; I the forests of two edge connected components are the same I the products of the norms kTe k are the same such that I G 0 is a input-output graph 11 / 15 I think of our graph as an algorithm, where we are feeding input vectors into the input vertices and then operate them through the graph, each edge doing some calculation, and each vertex acting like a logic gate, doing some compatibility checks. I the main problem is the timing of the various operations, in particular, how long one has to wait at a vertex, before applying an operator on an outgoing edge. I in algorithmic terms, it is clear that one has to wait until all the input information is processed; i.e. one has to wait for information to arrive along the longest path from an input vertex to the given vertex. 12 / 15 converting a two-edge connected graph to an input-output graph I let v and w be distinct vertices, declare v to be input and w to be output; I choose a simple path from v to w (no loops or backtracks), this makes a subgraph G1 which is an input-output subgraph; I we form an increasing sequence of subgraphs G1 ⊂ G2 ⊂ · · · ⊂ Gk by adding paths: if e is an edge with a vertex x in Gk and the other vertex z not in Gk , then there is a path from z to Gk without passing through x, say returning to Gk at y; orient this path connecting x to y and add to Gk to get Gk+1 so that Gk+1 has no oriented cycles; I this works unless x = y, in which case we split a vertex. 13 / 15 vertex splitting I sometimes to make a nice input-out graph we need to split a vertex (3) T3 i3 T2 (2) (4) ti2 i3 ti4 i2 T5 i2 I i4 i4 T4 T1 i1 i3 ti3 i4 i5 i1 (1) ti1 i2 i2 ! δi2 i!2 i2 (5) ti2 i5 i5 so we split the vertex and add an edge; on the new edge we put the identity matrix; this forces the indices at either end to be equal 14 / 15 optimality and edge collapsing How do we know that we can always find T : E → MN (C) such Y r(G) that |S(G, T)| = N kTe k? e I on each edge which is not a cutting edge place I; this collapses each two-edge connected component to a point; resulting graph os a tree (or forest) I on each edge which is not incident to a leaf place the identity matrix; flip vertices so that graph now looks like: V V V t V I √ let V(ξ1 ) = ( N)−1 ξ, V(ξi ) = 0 for i > 1 15 / 15