Kronecker Product of Networked Systems and their Approximates

Transcription

1 Kronecker Product of Networked Systems and their Approximates Robotics, Aerospace and Information Networks (RAIN) University of Washington (Robotics, Aerospace Kronecker and Information Products Networks (RAIN)) University of Washington 1 / 19

2 Graph Products: Networks within Networks Many ways to compose graphs G and H Cartesian product G H Tensor/Direct/Kronecker product G H Strong product G H Lexicographic product G H Rooted product G H Corona product G H Star product G H How does modularity of the network manifest itself as modularity within the state dynamics? Kronecker Product: (Graphs, ) (Dynamics, ) (Robotics, Aerospace Kronecker and Information Products Networks (RAIN)) University of Washington 2 / 19

3 Graph Product Examples Periodic Structures: e.g., hypercube multiprocessors, building trusses Compartmental Networks: e.g., air traffic networks, chemical reactions Constant degree expander graphs: e.g., computer networks, sorting networks, cryptography Australian Academic Research Network (AARNET) Cite: Parsonage et al. Generalized Graph Products for Network Design and Analysis, (Robotics, Aerospace Kronecker and Information Products Networks (RAIN)) University of Washington 3 / 19

4 Graph Kronecker Product Kronecker product G H Vertex set: V (G H) = V (G) V (H) Edge set: (x 1,x 2 ) (y 1,y 2 ) is in G H if x 1 y 1 and x 2 y 2 Algebraic Representation A(G H) = A(G) A(H) (Robotics, Aerospace Kronecker and Information Products Networks (RAIN)) University of Washington 4 / 19

5 Graph Kronecker Product Weighted: Directed: Multiple Products: (Robotics, Aerospace Kronecker and Information Products Networks (RAIN)) University of Washington 5 / 19

6 Relevance and Intuition: Fractal Nature Recursive growth of graph communities: Nodes get expanded to micro communities A(K) A(K K) A(K K K) A(K K K K) Obey common network features: Degree distribution, density power law, diameters, spectra [Leskovec et al. 10] (Robotics, Aerospace Kronecker and Information Products Networks (RAIN)) University of Washington 6 / 19

7 Relevance and Intuition: Null-Models Attribute representations, e.g., High/Low GPA, Year in school Nodes describe attributes, e.g., u = (High GPA,Yr 12), v = (Low GPA,Yr 11) Edges describe interaction probabilities, e.g., p(u v) = = 0.02 GPA High Low High Low Class Yr12 Yr11 Yr Yr A(G 1 ) A(G 2 ) (GPA,Class) (High,Yr12) (High,Yr11) (Low,Yr12) (Low,Yr11) (High,Yr12) (High,Yr11) (Low,Yr12) (Low,Yr11) A(G 1 G 2 ) (Robotics, Aerospace Kronecker and Information Products Networks (RAIN)) University of Washington 7 / 19

8 Dynamics over Kronecker Product The factor dynamics for i = 1,2,... Consider the discrete dynamics x i (k + 1) = A(G i )x i (k) y i (k) = C i x i (k) x(k + 1) = A(G 1 G 2...)x(k) = A( G i )x(k) y(k) = (C 1 C 2...)x(k) = C i x(k) For output node sets S 1,S 2,... then C(S 1 ) C(S 2 ) = C(S 1 S 2...) Here A( ) preserved the Kronecker product, e.g., Adjacency A(G 1 ), Row stochastic adjacency [A s (G)] ij = [A(G)] ij j [A(G)] ij How does the features of the factors compare to the composite system? (Robotics, Aerospace Kronecker and Information Products Networks (RAIN)) University of Washington 8 / 19

9 Trajectory and Stability When initialized from x(0) = x i (0) the composite trajectory is x(k + 1) = x i (k) Consequence: If the factors are stable then the composite is stable (Robotics, Aerospace Kronecker and Information Products Networks (RAIN)) University of Washington 9 / 19

10 Observability Dynamics are observable if for any unknown x(0), t f there exists a t f such that knowledges of u(t) and y(t) over [0,t 1 ] uniquely determine x(0). Significant in networked robotic systems, human-swarm interaction, network security, quantum networks. Challenging to establish for large networks Known families of observable graphs for selected outputs Paths (Rahmani and Mesbahi 07) Circulants (Nabi-Abdolyousefi and Mesbahi 12) Grids (Parlengeli and Notarsefano 11) Distance regular graphs (Zhang et al. 11) Cartesian products (Chapman et al. 14) (Robotics, Aerospace Kronecker and Information Products Networks (RAIN)) University of Washington 10 / 19

11 Observability Consider diagonalizable A(G 1 ) and A(G 2 ) with distinct eigenvalues of λ 1,..., λ p and µ 1,..., µ q Theorem The pair (A(G 1 G 2 ),C 1 C 2 ) is observable if and only if (1) the pairs (A(G 1 ),C 1 ) and (A(G 2 ),C 2 ) are observable and (2) for λ 1 µ 1 = λ 2 µ 2 = = λ p µ p, λ i λ j i j, p > 1, C T 1 [U 1,U 2,...,U p ] and/or C T 2 [V 1,V 2,...,V p ], where the columns of U i are the orthogonal right eigenvectors of eigenvalues λ i of A(G 1 ) (sim. for pairs ( µ i,v i ) of A(G 2 )). If the factors are observable then all modes such that λ i µ s λ j µ t are observable If (1) and (2) satisfied, and C 1 and C 2 are minimal rank observable on the factors then C 1 C 2 is a minimal rank observable on the composite (Robotics, Aerospace Kronecker and Information Products Networks (RAIN)) University of Washington 11 / 19

12 Observability Example (A(G 1 ),C(S 1 )) is observable with S 1 = {blue, green} (A(G 2 ),C(S 2 )) is observable with S 2 = { } No new multiplicities are introduced in A(G 1 G 2 ) = (A(G 1 G 2 ),C(S 1 S 2 )) is observable with S 1 S 2 = {blue, green } (Robotics, Aerospace Kronecker and Information Products Networks (RAIN)) University of Washington 12 / 19

13 Observability Factorization - Idea of the Proof Popov-Belevitch-Hautus (PBH) test (A, C) is unobservable if and only if there exists a right eigenvalue-eigenvector pair (λ,v) of A such that Cv = 0. Eigenvalue and eigenvector relationship: A(G 1 ) A(G 2 ) A(G 1 G 2 ) Eigenvalue λ i µ j λ i µ j Eigenvector v i u j v i u j Also (C 1 C 2 )(v i u i ) = C 1 v i C 2 u i The proof for simple eigenvalues follows from these observations. (Robotics, Aerospace Kronecker and Information Products Networks (RAIN)) University of Washington 13 / 19

14 Graph Factorization A graph can be factored as well as composed... Theorem (Sabidussi 1960) Every connected graph can be factored as a Kronecker product of prime graphs. This is NOT unique up to reordering of the factors. Primes: G = G 1 G 2 implies that either G 1 or G 2 is K 1 Number of prime factors is at most log G Algorithms Van Loan and Pitsianis (1993) - Exact and an approximation O(n 3 ) (more later) Leskovec et al. (2010) - KronFit approximation of the form G 1 G 1 G 1 (Robotics, Aerospace Kronecker and Information Products Networks (RAIN)) University of Washington 14 / 19

15 Kronecker Product Approximations Van Loan and Pitsianis developed an efficient method to solve min A A 1 A 2 2,F, i.e., the closest Kronecker product G G 1 G 2 G 1 G 2 G 1 G 2 G How does the features of these approximate factors compare to the composite system? (Robotics, Aerospace Kronecker and Information Products Networks (RAIN)) University of Washington 15 / 19

16 Trajectory Approximation Approximate Kronecker Dynamics x(k + 1) = (A(G 1 G 2 ) + )x(k) y(k) = (C 1 C 2 )x(k) For unforced dynamics if x(0) = x 1 (0) x 2 (0) then the trajectory can be approximated by x a (k) = x 1 (k) x 2 (k) where A = A(G 1 G 2 ) A x(k) x a (k) x(0) I k A I k A I A I (Robotics, Aerospace Kronecker and Information Products Networks (RAIN)) University of Washington 16 / 19

17 Distance to Instability Distance to instability is d A = inf( : A + is unstable) = 1 ρ(d A ). If the distance to instability of the factors is d G1 and d G2 then the distance to instability of the composite is d G1 G 2 = d G1 + d G2 d G1 d G2 Consequence: A stable composite dynamics is always more stable than its factors dynamics, i.e., d G1 G 2 max(d G1,d G2 ) (Robotics, Aerospace Kronecker and Information Products Networks (RAIN)) University of Washington 17 / 19

18 Distance to Unobservability Distance to unobservability is d A,C = inf { : (A +,C) is unobservable } Let λ 1 and µ 1 are the smallest magnitude eigenvalues of A(G 1 ) and A(G 2 ), respectively If the distance to unobservability of the factors is d G1,C 1 and d G2,C 2 then the distance to unobservability of the composite is bounded as d G1 G 2,C min( λ 1 d G2,C 2, µ 1 d G1,C 1 ) Consequence: For factor dynamics with a stable mode, the composite dynamics is always closer to unobservability than the factors, i.e., d G1 G 2,C min(d G2,C 2,d G1,C 1 ) (Robotics, Aerospace Kronecker and Information Products Networks (RAIN)) University of Washington 18 / 19

19 Conclusion Kronecker product dynamics related to its factors examining Trajectory Stability Observability Approximate Kronecker dynamics related to its approximate factors examining Bounded Trajectory Distance to instability Distance to unobservability Future work: Lower bounds for distance to unobservability Input/Output approximations Powers of Kronecker products (Robotics, Aerospace Kronecker and Information Products Networks (RAIN)) University of Washington 19 / 19