![Page 1: On Strong Structural Controllability of Networked Systems ...depts.washington.edu/uwrainlab/wordpress/wp...Distributed Space Systems Lab (DSSL) University of Washington Airlie Chapman](https://reader034.vdocuments.us/reader034/viewer/2022042112/5e8e667afecd1578d7621806/html5/thumbnails/1.jpg)
On Strong Structural Controllability of NetworkedSystems: A Constrained Matching Approach
Airlie Chapman and Mehran Mesbahi
Robotics, Aerospace and Information Networks Lab (RAIN Lab)Distributed Space Systems Lab (DSSL)
University of Washington
Airlie Chapman and Mehran Mesbahi (Robotics, Aerospace and Information Networks Lab (RAIN Lab) Distributed Space Systems Lab (DSSL))On Strong Structural Controllability University of Washington 1 / 16
![Page 2: On Strong Structural Controllability of Networked Systems ...depts.washington.edu/uwrainlab/wordpress/wp...Distributed Space Systems Lab (DSSL) University of Washington Airlie Chapman](https://reader034.vdocuments.us/reader034/viewer/2022042112/5e8e667afecd1578d7621806/html5/thumbnails/2.jpg)
The Network in the Dynamics
General Dynamics
x(t) = f (G,x(t),u(t))
y(t) = g(G,x(t),u(t))
Effective interfaces:
Network System Dynamics
Effective resistance H2 normAutomorphisms Homogeneity
Graph products/factorization Controllability composition/decompositionBipartite Matching Structural controllability
Airlie Chapman and Mehran Mesbahi (Robotics, Aerospace and Information Networks Lab (RAIN Lab) Distributed Space Systems Lab (DSSL))On Strong Structural Controllability University of Washington 2 / 16
![Page 3: On Strong Structural Controllability of Networked Systems ...depts.washington.edu/uwrainlab/wordpress/wp...Distributed Space Systems Lab (DSSL) University of Washington Airlie Chapman](https://reader034.vdocuments.us/reader034/viewer/2022042112/5e8e667afecd1578d7621806/html5/thumbnails/3.jpg)
The Network in the Dynamics
First Order, Linear Time Invariant model
xi (t) = ∑i∼j±wijxj (t)+ui (t)
yi (t) = xi (t)
Dynamics
x(t) = A(G)x(t) +B(S)u(t)
y(t) = C (R)x(t)
A(G): e.g. Z-matrix, Laplacian (wii =−∑wij) and Advection matrices(wii =−∑wji )
Input node set S = {vi ,vj , . . .}, B(S) = [ei ,ej , . . . ]
Output node set R= {vp,vq, . . .}, C (R) = [ep,eq, . . . ]T
Airlie Chapman and Mehran Mesbahi (Robotics, Aerospace and Information Networks Lab (RAIN Lab) Distributed Space Systems Lab (DSSL))On Strong Structural Controllability University of Washington 3 / 16
![Page 4: On Strong Structural Controllability of Networked Systems ...depts.washington.edu/uwrainlab/wordpress/wp...Distributed Space Systems Lab (DSSL) University of Washington Airlie Chapman](https://reader034.vdocuments.us/reader034/viewer/2022042112/5e8e667afecd1578d7621806/html5/thumbnails/4.jpg)
Flavors of Network Controllability
(1) General Controllability: Based on A(·) and G
(2) Structural Controllability: Based on G alone
Airlie Chapman and Mehran Mesbahi (Robotics, Aerospace and Information Networks Lab (RAIN Lab) Distributed Space Systems Lab (DSSL))On Strong Structural Controllability University of Washington 4 / 16
![Page 5: On Strong Structural Controllability of Networked Systems ...depts.washington.edu/uwrainlab/wordpress/wp...Distributed Space Systems Lab (DSSL) University of Washington Airlie Chapman](https://reader034.vdocuments.us/reader034/viewer/2022042112/5e8e667afecd1578d7621806/html5/thumbnails/5.jpg)
Structural Controllability
A pair (A(G),B(S)) is weak/strong structurally controllable (s-controllable),with weak/strong inputs S , if over every possible weighting of graph G it hasone/all controllable realization(s)
Conceived: Lin ’74, Mayeda and Yamada ’79Recently: Liu et al. ’11, Reinschke et al. ’92, Bowden et al. ’12Airlie Chapman and Mehran Mesbahi (Robotics, Aerospace and Information Networks Lab (RAIN Lab) Distributed Space Systems Lab (DSSL))On Strong Structural Controllability University of Washington 5 / 16
![Page 6: On Strong Structural Controllability of Networked Systems ...depts.washington.edu/uwrainlab/wordpress/wp...Distributed Space Systems Lab (DSSL) University of Washington Airlie Chapman](https://reader034.vdocuments.us/reader034/viewer/2022042112/5e8e667afecd1578d7621806/html5/thumbnails/6.jpg)
Pattern Representation
A pattern matrix A is a matrix composed ofzeros and crosses. A realization A of A maintainsthe zero structure
A(G) defined s.t. one realization is the adjacencymatrix A(G) (Similarly for B(S) and C(R))
Patterned Dynamics
x(t) = Ax(t) +Bu(t)
y(t) = Cx(t)
Here, A ∈A(G), B ∈B(S),C ∈ C(R).
Example:
A(G) =
× 0 ×× 0 ×0 × 0
, B(S) =
0×0
, C(R) =
00×
A pair (A(G),B(S)) is weak/strong s-controllable, with weak/strong inputsS , if it has one/all controllable realization(s) (A,B)
Airlie Chapman and Mehran Mesbahi (Robotics, Aerospace and Information Networks Lab (RAIN Lab) Distributed Space Systems Lab (DSSL))On Strong Structural Controllability University of Washington 6 / 16
![Page 7: On Strong Structural Controllability of Networked Systems ...depts.washington.edu/uwrainlab/wordpress/wp...Distributed Space Systems Lab (DSSL) University of Washington Airlie Chapman](https://reader034.vdocuments.us/reader034/viewer/2022042112/5e8e667afecd1578d7621806/html5/thumbnails/7.jpg)
Bipartite Representation
Bipartite representation H= (V+,V−,E ) of A(G) ∈ Rp×q
V+
A(G) =
× 0 ×× 0 ×0 × 0
V−
Airlie Chapman and Mehran Mesbahi (Robotics, Aerospace and Information Networks Lab (RAIN Lab) Distributed Space Systems Lab (DSSL))On Strong Structural Controllability University of Washington 7 / 16
![Page 8: On Strong Structural Controllability of Networked Systems ...depts.washington.edu/uwrainlab/wordpress/wp...Distributed Space Systems Lab (DSSL) University of Washington Airlie Chapman](https://reader034.vdocuments.us/reader034/viewer/2022042112/5e8e667afecd1578d7621806/html5/thumbnails/8.jpg)
Rank and Matchings
(A,B) is controllable iff rank[A−λ I ,B] = n for all eigenvalues λ of A
Combinatorial Criteria: =⇒ Rank Criteria:A has a t-matching ∃A ∈ A with rank(A)≥ t
A has a t-matching if there are t edges inH= (V+,V−,E ) between I+ ⊆ V+ andI− ⊆ V− (|I+|= |I−|= t) where no twoedges share a node
Nodes in V− \ I− are called unmatched
Airlie Chapman and Mehran Mesbahi (Robotics, Aerospace and Information Networks Lab (RAIN Lab) Distributed Space Systems Lab (DSSL))On Strong Structural Controllability University of Washington 8 / 16
![Page 9: On Strong Structural Controllability of Networked Systems ...depts.washington.edu/uwrainlab/wordpress/wp...Distributed Space Systems Lab (DSSL) University of Washington Airlie Chapman](https://reader034.vdocuments.us/reader034/viewer/2022042112/5e8e667afecd1578d7621806/html5/thumbnails/9.jpg)
Weak S-Controllability (Liu et al.)
Weak Inputs
S is weak iff A has an (n−|S |)-matching with S unmatched and input accessible.
S = {2}, R = {3}are weak
Airlie Chapman and Mehran Mesbahi (Robotics, Aerospace and Information Networks Lab (RAIN Lab) Distributed Space Systems Lab (DSSL))On Strong Structural Controllability University of Washington 9 / 16
![Page 10: On Strong Structural Controllability of Networked Systems ...depts.washington.edu/uwrainlab/wordpress/wp...Distributed Space Systems Lab (DSSL) University of Washington Airlie Chapman](https://reader034.vdocuments.us/reader034/viewer/2022042112/5e8e667afecd1578d7621806/html5/thumbnails/10.jpg)
Weak Features
Weak Inputs
S is weak iff A has an (n−|S |)-matching with S unmatched and input accessible.
Efficient algorithms for maximum bipartite matching
Deterministic O(√|V | |E |
), Probabilistic O
(|V |2.376
)
S is generically controllable but real-world systems canbe atypical, e.g., undirected unweighted consensus
Adding edges tends to improve weak s-controllability
A self-dampednetwork
Airlie Chapman and Mehran Mesbahi (Robotics, Aerospace and Information Networks Lab (RAIN Lab) Distributed Space Systems Lab (DSSL))On Strong Structural Controllability University of Washington 10 / 16
![Page 11: On Strong Structural Controllability of Networked Systems ...depts.washington.edu/uwrainlab/wordpress/wp...Distributed Space Systems Lab (DSSL) University of Washington Airlie Chapman](https://reader034.vdocuments.us/reader034/viewer/2022042112/5e8e667afecd1578d7621806/html5/thumbnails/11.jpg)
Rank and Matchings Revisited
(A,B) is controllable iff rank[A−λ I ,B] = n for all eigenvalues λ of A
Combinatorial Criteria:=⇒
Rank Criteria:A has a t-matching ∃A ∈ A with rank(A)≥ t
A has a constrained t-matching ∀A ∈ A, rank(A)≥ t
A t-matching is constrained if it is theonly t-matching between I+ and I−
A matching is Vs -less if it contains noedges correponding to self loops, i.e.,{v+i ,v−i
}.
Airlie Chapman and Mehran Mesbahi (Robotics, Aerospace and Information Networks Lab (RAIN Lab) Distributed Space Systems Lab (DSSL))On Strong Structural Controllability University of Washington 11 / 16
![Page 12: On Strong Structural Controllability of Networked Systems ...depts.washington.edu/uwrainlab/wordpress/wp...Distributed Space Systems Lab (DSSL) University of Washington Airlie Chapman](https://reader034.vdocuments.us/reader034/viewer/2022042112/5e8e667afecd1578d7621806/html5/thumbnails/12.jpg)
Rank and Matchings Revisited
(A,B) is controllable iff rank[A−λ I ,B] = n for all eigenvalues λ of A
Combinatorial Criteria:=⇒
Rank Criteria:A has a t-matching ∃A ∈ A with rank(A)≥ t
A has a constrained t-matching ∀A ∈ A, rank(A)≥ t
A t-matching is constrained if it is theonly t-matching between I+ and I−
A matching is Vs -less if it contains noedges correponding to self loops, i.e.,{v+i ,v−i
}.
Airlie Chapman and Mehran Mesbahi (Robotics, Aerospace and Information Networks Lab (RAIN Lab) Distributed Space Systems Lab (DSSL))On Strong Structural Controllability University of Washington 11 / 16
![Page 13: On Strong Structural Controllability of Networked Systems ...depts.washington.edu/uwrainlab/wordpress/wp...Distributed Space Systems Lab (DSSL) University of Washington Airlie Chapman](https://reader034.vdocuments.us/reader034/viewer/2022042112/5e8e667afecd1578d7621806/html5/thumbnails/13.jpg)
Strong S-Controllability
Strong inputs
S is strong iff A has a constrained (n−|S |)-matching with S unmatched andA× has a constrained Vs -less (n−|S |)-matching with S unmatched.
Pattern matrix A× is formed by placing crosses along the diagonal of A
S = {2} is strongR = {3} is not
Airlie Chapman and Mehran Mesbahi (Robotics, Aerospace and Information Networks Lab (RAIN Lab) Distributed Space Systems Lab (DSSL))On Strong Structural Controllability University of Washington 12 / 16
![Page 14: On Strong Structural Controllability of Networked Systems ...depts.washington.edu/uwrainlab/wordpress/wp...Distributed Space Systems Lab (DSSL) University of Washington Airlie Chapman](https://reader034.vdocuments.us/reader034/viewer/2022042112/5e8e667afecd1578d7621806/html5/thumbnails/14.jpg)
Strong Features
A strongly controllable input set S can be considered a type of controllabilityrobustness
For connected networks, adding edges tends to worsen strong s-controllability
Algorithms:
Golumbic (2001) - O (|V |+ |E |) to check amatching is constrained
Golumbic (2001) - NP-complete to find amaximum constrained matching
Misha (2011) - Polynomial time algorithm toapproximate a maximum constrained matching
can do no better than 1
2 3√9|V |
13−ε for any ε > 0
We have an O(|V |2) algorithm to check if S isstrong and to find a (not-necessarily minimal)strong input set
Student Version of MATLAB
Airlie Chapman and Mehran Mesbahi (Robotics, Aerospace and Information Networks Lab (RAIN Lab) Distributed Space Systems Lab (DSSL))On Strong Structural Controllability University of Washington 13 / 16
![Page 15: On Strong Structural Controllability of Networked Systems ...depts.washington.edu/uwrainlab/wordpress/wp...Distributed Space Systems Lab (DSSL) University of Washington Airlie Chapman](https://reader034.vdocuments.us/reader034/viewer/2022042112/5e8e667afecd1578d7621806/html5/thumbnails/15.jpg)
Self-damped Undirected Networks
Tested all undirected connected graphs for n ≤ 10
For S a minimum cardinality weak/strong set, nD := |S ||V |
2 3 4 5 6 7 8 9 100.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
n
n D
StrongWeak
Smallest weak/strong S is a lower/upper bound on the smallest controllableinput set for an arbitrary realization
The only strong single input is the end nodes of a path graph
The only realization requiring a strong n−1 input set is the complete graph
Airlie Chapman and Mehran Mesbahi (Robotics, Aerospace and Information Networks Lab (RAIN Lab) Distributed Space Systems Lab (DSSL))On Strong Structural Controllability University of Washington 14 / 16
![Page 16: On Strong Structural Controllability of Networked Systems ...depts.washington.edu/uwrainlab/wordpress/wp...Distributed Space Systems Lab (DSSL) University of Washington Airlie Chapman](https://reader034.vdocuments.us/reader034/viewer/2022042112/5e8e667afecd1578d7621806/html5/thumbnails/16.jpg)
Directed Erdos-Renyi random networks
Randomly generated on n nodes with a directed edge (i , j) ∈ E existing withprobability p and mean degree 〈k〉= 2np
Tested 1200 graphs on 20 nodes for each 〈k〉= 2,4, . . . ,20
0 2 4 6 8 10 12 14 16 18 20
0
0.2
0.4
0.6
0.8
1
<k>
n D
StrongWeak
kc = 2logn ≈ 6 is a sharp threshold for the disoriented connectedness ofErdos-Renyi random networks
k = kc presents on average the minimum strong input set
Airlie Chapman and Mehran Mesbahi (Robotics, Aerospace and Information Networks Lab (RAIN Lab) Distributed Space Systems Lab (DSSL))On Strong Structural Controllability University of Washington 15 / 16
![Page 17: On Strong Structural Controllability of Networked Systems ...depts.washington.edu/uwrainlab/wordpress/wp...Distributed Space Systems Lab (DSSL) University of Washington Airlie Chapman](https://reader034.vdocuments.us/reader034/viewer/2022042112/5e8e667afecd1578d7621806/html5/thumbnails/17.jpg)
Directed Erdos-Renyi random networks
Randomly generated on n nodes with a directed edge (i , j) ∈ E existing withprobability p and mean degree 〈k〉= 2np
Tested 1200 graphs on 20 nodes for each 〈k〉= 2,4, . . . ,20
0 2 4 6 8 10 12 14 16 18 20
0
0.2
0.4
0.6
0.8
1
<k>
n D
StrongWeak
Almost surely Almost surelydisconnected connected
←− −→Matchings Constrained
hard to matchings hard toform maintain
kc = 2logn ≈ 6 is a sharp threshold for the disoriented connectedness ofErdos-Renyi random networks
k = kc presents on average the minimum strong input set
Airlie Chapman and Mehran Mesbahi (Robotics, Aerospace and Information Networks Lab (RAIN Lab) Distributed Space Systems Lab (DSSL))On Strong Structural Controllability University of Washington 15 / 16
![Page 18: On Strong Structural Controllability of Networked Systems ...depts.washington.edu/uwrainlab/wordpress/wp...Distributed Space Systems Lab (DSSL) University of Washington Airlie Chapman](https://reader034.vdocuments.us/reader034/viewer/2022042112/5e8e667afecd1578d7621806/html5/thumbnails/18.jpg)
Conclusion
Linked strong s-controllability to a bipartite matching property and as suchillustrated the computational challenges of the problem
Compared features of weak and strong inputs
Provided an efficient algorithm to generate strong inputs
Brief examination of self-damped undirected and Erdos-Renyi networks
Future Direction: Output weak and strong s-controllability, degree ofcontrollability, edge s-controllability
Airlie Chapman and Mehran Mesbahi (Robotics, Aerospace and Information Networks Lab (RAIN Lab) Distributed Space Systems Lab (DSSL))On Strong Structural Controllability University of Washington 16 / 16
![Page 19: On Strong Structural Controllability of Networked Systems ...depts.washington.edu/uwrainlab/wordpress/wp...Distributed Space Systems Lab (DSSL) University of Washington Airlie Chapman](https://reader034.vdocuments.us/reader034/viewer/2022042112/5e8e667afecd1578d7621806/html5/thumbnails/19.jpg)
Strong S-Controllability: Rough Proof
Rank test
(A,B) is controllable iff [A−λ I ,B] has full column rank for every eigenvalue λ ofA.
[A,B] is full rank iff ∃ permutation matrices P1 and P2, s.t.,
P1 [A,B]P2 =
⊗ ·· · ⊗ × 0 · · · 0...
.... . . ×
. . ....
. . .. . . 0
⊗ ·· · ⊗ · · · · · · ⊗ ×
,
where ⊗-elements can be either zero or crosses
For λ = 0
[A,B(S)] has this form iff [A,B(S)] has a constrained n-matching
Removing S rows of [A,B(S)] leaves a constrained (n−|S |)-matching
Implies A has a constrained (n−|S |)-matching with S unmatched
Similarly for λ =×, and A×Airlie Chapman and Mehran Mesbahi (Robotics, Aerospace and Information Networks Lab (RAIN Lab) Distributed Space Systems Lab (DSSL))On Strong Structural Controllability University of Washington 17 / 16
![Page 20: On Strong Structural Controllability of Networked Systems ...depts.washington.edu/uwrainlab/wordpress/wp...Distributed Space Systems Lab (DSSL) University of Washington Airlie Chapman](https://reader034.vdocuments.us/reader034/viewer/2022042112/5e8e667afecd1578d7621806/html5/thumbnails/20.jpg)
Self-damped Undirected Networks
Self-damped
Given a maximum constrained self-less matching of A× with unmatched nodes S .Then, S is strong with minimum cardinality.
Tested all undirected connected graphs for n ≤ 10
For S a minimum cardinality strong set, nD := |S |/ |V |
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
nD
Fra
ctio
n of
Gra
phs
45678910
2 3 4 5 6 7 8 9 100.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
n
n D
StrongWeak
The only strong single input for a connected self-damped undirected networkis the end nodes of a path graph
The only connected self-damped undirected network requiring a strong n−1input set is the complete graph
Airlie Chapman and Mehran Mesbahi (Robotics, Aerospace and Information Networks Lab (RAIN Lab) Distributed Space Systems Lab (DSSL))On Strong Structural Controllability University of Washington 18 / 16
![Page 21: On Strong Structural Controllability of Networked Systems ...depts.washington.edu/uwrainlab/wordpress/wp...Distributed Space Systems Lab (DSSL) University of Washington Airlie Chapman](https://reader034.vdocuments.us/reader034/viewer/2022042112/5e8e667afecd1578d7621806/html5/thumbnails/21.jpg)
Directed Erdos-Renyi random networks
Randomly generated on n nodes with an edge (i , j) ∈ E existing withprobability p and mean degree 〈k〉= 2np
Tested 1200 graphs on 20 nodes for each 〈k〉= 2,4, . . . ,20
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.80
0.1
0.2
0.3
0.4
0.5
nD
Fra
ctio
n of
gra
phs
2468101214161820
0 2 4 6 8 10 12 14 16 18 20
0
0.2
0.4
0.6
0.8
1
<k>n D
StrongWeak
kc = 2logn ≈ 6 is a sharp threshold for the disoriented connectedness ofErdos-Renyi random networks
k = kc presents on average the minimum strong input set
Airlie Chapman and Mehran Mesbahi (Robotics, Aerospace and Information Networks Lab (RAIN Lab) Distributed Space Systems Lab (DSSL))On Strong Structural Controllability University of Washington 19 / 16