Homework 3
Transcription
Homework 3
Advanced Topics in Control: Distributed Systems and Control Homework set #3 Note: The homework is due Friday, April 17 before 10:15 am. Exercise 1: The Green matrix of a Laplacian matrix (i.e., alternative definition and property of the pseudoinverse Laplacian matrix) Assume L is the Laplacian matrix of a weighted connected undirected graph with n nodes. Show that [20%] 1. the matrix L + n1 1n 1Tn is positive definite, 2. the so-called Green matrix −1 1 1 − 1n 1Tn X = L + 1n 1Tn n n is the unique solution to the system of equations: ( LX = In − n1 1n 1Tn , 1Tn X = 0Tn , 3. X = L† , where L† is defined in Exercise 6.6. Exercise 2: Properties of saddle points Prove Lemma 7.3: [20%] Let L = LT ∈ Rn×n be a symmetric Laplacian associated to an undirected, connected, and weighted graph, and consider the Lagrangian L, where each fi is strictly convex and twice continuously differentiable for all i ∈ {1, . . . , n}. Then 1. if (y ∗ , z ∗ ) ∈ Rn × Rn is a saddle point of L, then so is (y ∗ , z ∗ + α1n ) for any α ∈ R; 2. if (y ∗ , z ∗ ) ∈ Rn × Rn is a saddle point of L, then y ∗ = x∗ 1n where x∗ ∈ R is a solution of the original optimization problem (7.3); and 3. if x∗ ∈ R is a solution of the original optimization problem (7.3), then there are z ∗ ∈ Rn and ∂ ˜ ∗ y ∗ = x∗ 1n satisfying Lz ∗ + ∂y f (y ) = 0n so that (y ∗ , z ∗ ) is a saddle point of L. Exercise 3: The edge Laplacian matrix For an unweighted undirected graph, analogously to the Laplacian matrix of L = BB T ∈ Rn×n , we can define the edge Laplacian matrix Ledge = B T B ∈ Rm×m . Show that 1. ker(Ledge ) = ker(B); 2. for an acyclic graph Ledge is nonsingular; and 3. neglecting distinct eigenvalue multiplicities, L and Ledge have the same eigenvalues and, therefore, the same rank. 1 [20%] Advanced Topics in Control: Distributed Systems and Control Homework set #3 Exercise 4: Continuous distributed estimation from relative measurements Consider the continuous distributed estimation algorithm given by the affine Laplacian flow (8.4). Show that for an undirected and connected graph G and appropriately initial conditions x ˆ(0) = 0n , the affine Laplacian flow (8.4) converges to the unique solution x ˆ∗ of the estimation problem given in Lemma 8.5. [10%] Exercise 5: Averaging with distributed integral control Consider a Laplacian flow implemented as a relative sensing network over a connected and undirected graph with incidence matrix B ∈ Rn×|E| and weights aij > 0 for i, j ∈ E, and subject to a constant disturbance term η ∈ R|E| , as shown in Figure 1. .. u _ . x x˙ i = ui .. BT B ⌘ + . + .. z . aij .. . y Figure 1: A relative sensing network with a constant disturbance input η ∈ R|E| . 1. Derive the dynamic closed-loop equations describing the model in Figure 1. 2. Show that asymptotically all states x(t) converge to some constant vector x∗ ∈ Rn depending on the value of the disturbance η, i.e., x∗ is not necessarily a consensus state. Consider the system from Figure 1 with a distributed integral controller forcing convergence to consensus, as shown in Figure 2. Recall that 1s is the the Laplace symbol for the integrator. 3. Derive the dynamic closed-loop equations describing the model in Figure 2. 4. Show that the distributed integral controller in Figure 1 asymptotically stabilizes the set of steady states (x∗ , p∗ ), where x∗ ∈ span(1n ) corresponds to consensus. Hint: To show stability, use Lemma 7.4. 2 [30%] Advanced Topics in Control: Distributed Systems and Control .. u _ . Homework set #3 x x˙ i = ui .. . BT B ⌘ + + + p z .. .. . aij .. . .. . y .1 s Figure 2: Relative sensing network with a disturbance η ∈ R|E| and distributed integral action. 3