Upload
others
View
2
Download
0
Embed Size (px)
Citation preview
Controllability and Identification of Complex Networks
Jorge CortesUNIVERSITY OF CALIFORNIAUNOFFICIAL SEAL
Attachment B - “Unofficial” SealFor Use on Letterhead
Mechanical and Aerospace EngineeringUniversity of California, San Diegohttp://carmenere.ucsd.edu/jorge
IMA Annual Program Year WorkshopAnalysis and Control of Network Dynamics
University of MinnesotaOctober 23, 2015
Joint work with Yingbo Zhao
Complex interconnected systems
Networked systems play increasingly essential role in modern society
power networks
wireless communication networks
social networks
biological networks
transportation networks
sensor&robotic networks
US power grid Google Loon Ocean science
Seek to understand how complex networks work, how individual(s) affect networkbehavior, how robust network is against external disruption and shocks,...
J. Cortes (UC San Diego) Controllability and Identification of Complex Networks October 23, 2015 2 / 40
Network science and distributed control
Challenging goals, different flavors
Modeling behavior and understanding interactions in large-scale networks
Design provably correct distributed coordination algorithms to make networkachieve desired task
Analyze properties of complex networks, mutual effect between dynamics andtopology
Focus today on analysis of network reachability and identification
Understanding how to drive the network state efficiently by controlling only afew selected components
Identifying network structure from input/output data
Plethora of other relevant issues: understanding noise propagation, fragility,optimizing placement of limited resources for enhancing network properties
J. Cortes (UC San Diego) Controllability and Identification of Complex Networks October 23, 2015 3 / 40
Network science and distributed control
Challenging goals, different flavors
Modeling behavior and understanding interactions in large-scale networks
Design provably correct distributed coordination algorithms to make networkachieve desired task
Analyze properties of complex networks, mutual effect between dynamics andtopology
Focus today on analysis of network reachability and identification
Understanding how to drive the network state efficiently by controlling only afew selected components
Identifying network structure from input/output data
Plethora of other relevant issues: understanding noise propagation, fragility,optimizing placement of limited resources for enhancing network properties
J. Cortes (UC San Diego) Controllability and Identification of Complex Networks October 23, 2015 3 / 40
Complex network as linear control system
Linear control systems provide simplest (but not simple!) model
x(k + 1) = Ax(k) + Bu(k)
each component of x corresponds to the state of a node
network adjacency matrix A
each column of B only affects a specific node
J. Cortes (UC San Diego) Controllability and Identification of Complex Networks October 23, 2015 4 / 40
Complex network as linear control system
Linear control systems provide simplest (but not simple!) model
x(k + 1) = Ax(k) + Bu(k)
each component of x corresponds to the state of a node
network adjacency matrix A
each column of B only affects a specific node
Deciding controllability: how many nodes, how robust, influence of topology
Y. Liu, J. Slotine, and A. Barabasi (Nature 2011): structuralcontrollability and degree distribution
A. Olshevsky (TAC 14): minimal controllability problems
A. Rahmani, M. Ji, M. Mesbahi, and M. Egerstedt (SIAM JCO, 09):controllability of consensus-type linear networks
C. Aguilar and B. Gharesifard (ACC 14): controllability ofconsensus-type nonlinear networks
J. Cortes (UC San Diego) Controllability and Identification of Complex Networks October 23, 2015 4 / 40
Complex network as linear control system
Linear control systems provide simplest (but not simple!) model
x(k + 1) = Ax(k) + Bu(k)
each component of x corresponds to the state of a node
network adjacency matrix A
each column of B only affects a specific node
Controllability metrics: how much energy to steer the network, strategies
G. Yan et al (PRL 12): how much energy is needed to control complexnetworks
F. Pasqualetti, S. Zampieri, and F. Bullo (IEEE TCNS 14):Controllability metrics, asymptotic regimes
J. Cortes (UC San Diego) Controllability and Identification of Complex Networks October 23, 2015 4 / 40
Complex network as linear control system
Linear control systems provide simplest (but not simple!) model
x(k + 1) = Ax(k) + Bu(k)
each component of x corresponds to the state of a node
network adjacency matrix A
each column of B only affects a specific node
Actuator and sensor placement: node location
T. Summers and J. Lygeros (IFAC World Congress 14): optimal sensorand actuator placement
V. Tzoumas, M. A. Rahimian, G. J. Pappas, A. Jadbabaie (ACC15): minimal actuator placement with optimal control constraints
...
J. Cortes (UC San Diego) Controllability and Identification of Complex Networks October 23, 2015 4 / 40
Effective connectivity in the brain
Brain areas with different functions represented as nodes, brain modeled asnetwork of nodes
Effective connectivity: external inputs not only have direct effect on the state ofthe brain in a particular area, but can also activate the connections amongdifferent brain areas
J. Cortes (UC San Diego) Controllability and Identification of Complex Networks October 23, 2015 5 / 40
Complex network as bilinear control system
Bilinear control systems provide next (in the complexity scale) simplest model
x(k + 1) = Ax(k) + Bu(k) + x(k)TFu(k)
Inputs not only affect the state of individual nodes, but can also modulateagent-to-agent interconnection
1 11 1
13 2 3 11 3
( 1) ( )
( ) ( )) )( (u k f
x k a x k
a x k u k
+ = +
+ +
21a
2 22 2
21 1 2
( 1) ( )
( ) ( )
x k a x k
a x k u k
+ = +
+
3 32 2
33 3 3
( 1) ( )
( ) ( )
x k a x k
a x k u k
+ = +
+
32a
13a2( )u k
13f
Akin to “dynamic causal modeling” in neuroscience: K. Friston, L.
Harrison, W. Penny (Neuroimage, 03)
J. Cortes (UC San Diego) Controllability and Identification of Complex Networks October 23, 2015 6 / 40
Outline
1 Reachability metrics for bilinear control systemsReachability GramianGramian-based reachability metrics and actuator selectionAdding bilinear inputs to linear networks
2 Network identification with latent nodesManifest and latent nodesAsymptotically exact identification via auto-regressive modelsLeast-squares auto-regressive estimation
J. Cortes (UC San Diego) Controllability and Identification of Complex Networks October 23, 2015 7 / 40
Reachability metrics for bilinear control systems
How easy/difficult to control is a given bilinear control system?How much energy is needed to steer the system to desirable state?Do bilinear systems offer any advantage compared to linear ones?
6 7
67 7( ) or ( )?u k u k
1
5 3
4
21( )u k4( )u k
Given desired performance and constraints (sparsity, limited inputenergy), is it better to control a node state directly or modulate(strengthen, weaken, or create) an interconnection?
J. Cortes (UC San Diego) Controllability and Identification of Complex Networks October 23, 2015 8 / 40
Controllability of bilinear systemsIncomplete literature review
Most available results study controllability as a binary property
W. Boothby and E. Wilson (SIAM JCO, 79): determination of transitivity
D. Koditschek and K. Narendra (IEEE TAC 85): controllability of planarcontinuous-time bilinear systems
U. Piechottka and P. Frank (Automatica 92): uncontrollability test for3-d continuous-time systems
M. Evans and D. Murthy (IEEE TAC 77), T. Goka, T. Tarn, and J.
Zaborszky (Automatica 73): controllability of discrete-timehomogeneous-in-the-state bilinear system using rank-1 controllers
L. Tie and K. Cai (IEEE TAC 10): near controllability of discrete-timesystems
Missing quantitative measures to ascertain energy amount required to steer state
J. Cortes (UC San Diego) Controllability and Identification of Complex Networks October 23, 2015 9 / 40
Minimum-energy optimal control problem
Minimum energy required to steer state from 0n to desired xf in K time steps
min{u}K−1
∑K−1k=0 uT (k)u(k)
s.t. ∀k = 0, . . . ,K − 1,x(k + 1) = Ax(k) +
∑mi=1(Fix(k) + Bi )ui (k)
x(0) = 0n, x(K ) = xf .
Necessary optimality conditions for solution lead to nonlinear two-pointboundary-value problem without known analytical solution
For linear systems (F = 0), optimal solution and value
u∗(k) = BT
(AT
)K−k−1W−11,Kxf
K−1∑k=0
u∗(k)Tu∗(k) = xTf W−11,Kxf
(W1,K ,∑K−1
k=0 AkBBT (AT )k is K -step controllability Gramian)
J. Cortes (UC San Diego) Controllability and Identification of Complex Networks October 23, 2015 10 / 40
Gramian-based reachability metrics for linear networks
Define controllability Gramian
W1 = limK→∞
W1,K =∞∑k=0
AkBBT (AT )k
For arbitrary time horizon K , lower bound on minimum input energy
K−1∑k=0
(u∗(k))Tu∗(k) = xTf W−11,Kxf > xTf W−1
1 xf
Gramian-based quantitative reachability metrics
λmin(W1): worst-case minimum input energytr(W1): average minimum control energy over {x ∈ Rn|‖x‖= 1}det(W1): volume of ellipsoid reachable using unit-energy control inputs
J. Cortes (UC San Diego) Controllability and Identification of Complex Networks October 23, 2015 11 / 40
Reachability Gramian for bilinear control systems
Even though bilinear systems are not linear, analogous concept exist
For stable bilinear system (A,F ,B), define recursively
W1 =∞∑
k1=0
Ak1BBT (Ak1 )T
Wi =∞∑
ki=0
Aki( m∑j=1
FjWi−1FTj
)(Aki )T , i ≥ 2
Reachability Gramian is
W =∞∑i=1
Wi
J. Cortes (UC San Diego) Controllability and Identification of Complex Networks October 23, 2015 12 / 40
Properties of the reachability Gramian
Reachability Gramian W is positive semi-definite
Reachability Gramian W satisfies generalized Lyapunov equation
AWAT −W +m∑j=1
FjWFTj + BBT = 0n×n
Unique positive semi-definite solution exists to Lyapunov equation iff
ρ(A⊗ A +m∑j=1
Fj ⊗ Fj) < 1
Im(W) is invariant under bilinear control dynamics (A,F ,B)—hence any state xf reachable from the origin belongs to Im(W)
J. Cortes (UC San Diego) Controllability and Identification of Complex Networks October 23, 2015 13 / 40
Lower bound on minimum input energy∗
Gramian-based lower bound for bounded inputs
For arbitrary time horizon K , if
‖u(k)‖∞ ≤1
2
( m∑i,j=1
‖FTj ΨFi‖
)−1β, k = 0, 1, . . . ,K − 1
then input energy is lower bounded as
K−1∑k=0
uT (k)u(k) ≥ xT (K )W−1x(K )
input u
state x
Ψ ,W−1 −W−1B(BTW−1B − Im)−1BTW−1
β , −m∑j=1
‖AT ΨFj + FTj ΨA‖+(( m∑
j=1
‖AT ΨFj + FTj ΨA‖)2 − 4
m∑i,j=1
‖FTj ΨFi‖·λmax(AT ΨA −W−1))1/2
J. Cortes (UC San Diego) Controllability and Identification of Complex Networks October 23, 2015 14 / 40
Lower bound on minimum input energy – cont’d
input u state x
Similar bound for continuous-time bilinear systems
W. Gray and J. Mesko (IFAC NCSDS 98) underintegrability condition
P. Benner and T. Damm (IJC 11): underdiagonal reachability Gramian
Bound on input energy does not hold if input norm is unconstrained
input u state x
inf
∑K−1k=0 uT (k)u(k)
‖x(K )‖2= 0
J. Cortes (UC San Diego) Controllability and Identification of Complex Networks October 23, 2015 15 / 40
Gramian-based reachability metrics
Lower bound on minimum input energy justifies following reachability metrics
λmin(W): minimum input energy in the worst casetr(W): average minimum input energy over unit hypersphere in state spacedet(W): volume of the ellipsoid containing the reachable states using
inputs with no more than unit energy
Plenty of (open) questions
how does Gramian and metrics depend on network?
what networks are better?
how robust are these properties against malfunctions/failures?
how to select inputs to make network better?
. . .
J. Cortes (UC San Diego) Controllability and Identification of Complex Networks October 23, 2015 16 / 40
Actuator selection
Consider M candidate actuators, labeled V = {1, . . . ,M}Actuator i associated to pair (bi ,Fi ) ∈ Rn × Rn×n
Actuator selection: select m ≤ M actuators S = {s1, . . . , sm} for
maxS⊆V
f (W(S)) (f is either tr, λmin or det)
Increasing returns property of map from S to W(S)
For any S1 ⊆ S2 ⊆ V and s ∈ V \S2,
W(S2 ∪ {s})−W(S2) ≥ W(S1 ∪ {s})−W(S1)
Maximization of supermodular functions under cardinality constraints is NP hard
J. Cortes (UC San Diego) Controllability and Identification of Complex Networks October 23, 2015 17 / 40
Actuator selection
Consider M candidate actuators, labeled V = {1, . . . ,M}Actuator i associated to pair (bi ,Fi ) ∈ Rn × Rn×n
Actuator selection: select m ≤ M actuators S = {s1, . . . , sm} for
maxS⊆V
f (W(S)) (f is either tr, λmin or det)
Increasing returns property of map from S to W(S)
For any S1 ⊆ S2 ⊆ V and s ∈ V \S2,
W(S2 ∪ {s})−W(S2) ≥ W(S1 ∪ {s})−W(S1)
Maximization of supermodular functions under cardinality constraints is NP hard
J. Cortes (UC San Diego) Controllability and Identification of Complex Networks October 23, 2015 17 / 40
Bounds on optimal value
Upper bound on optimal value: Lagrangian dual and continuous relaxationsolvable in polynomial time [G. Gallo and B. Simeone (Math. Prog. 89)]
Lower bound on reachability metrics
Let f be either tr, λmin or det. For any m actuators S ,
f (W(S)) ≥∑s∈S
f (W(s))
Greedy algorithm: maximize lower bound by
1 compute f (W(s)) for every s ∈ V
2 order results in decreasing order
3 select actuators sequentially starting with the one with largest value
J. Cortes (UC San Diego) Controllability and Identification of Complex Networks October 23, 2015 18 / 40
Example: controller selection via greedy algorithm
Baseline system on R5: (A,F0, b0)
b = (0.8, 0.6, 0.4, 0.2, 0.5), F = diag(0.1, 0.2, 0.3, 0.4, 0.5)
Additional actuator candidates: {(Bi ,Fi )}3i=1, Bi = ei ,
F1 =
0 0.02 0 0 00 0 0.02 0 00 0 0 0 00 0 0 0 00 0 0 0 0
, F2 =
0 0 0 0 00 0 0 0 0.010 0 0 0 00 0.05 0 0 00 0 0 0 0
, F3 =
0.05 0 0 0 0
0 0 0 0 00 0 0 0 00 0 0 0 0.020 0 0 0 0
S tr(W (S)) λmin(W (S)) det(W (S)) S tr(W (S)) λmin(W (S)) det(W (S))
{0} 14.42 0.027 0.242 {0, 2} 19.91 0.07 3.32
{1} 5.03 0.023 0.025 {0, 3} 18.69 0.05 1.13
{2} 4.04 3 × 10−5 9 × 10−7 {0, 1, 2} 26.50 0.137 46.15
{3} 3.03 1.6 × 10−6 4 × 10−11 {0, 1, 3} 25.28 0.125 28.68
{0, 1} 20.98 0.09 11.704 {0, 2, 3} 24.19 0.103 8.34
Lower bound is tight for tr, but not for λmin and det
Actuators with a large individual contribution are involved in combinatorialchoices with large values
J. Cortes (UC San Diego) Controllability and Identification of Complex Networks October 23, 2015 19 / 40
Outline
1 Reachability metrics for bilinear control systemsReachability GramianGramian-based reachability metrics and actuator selectionAdding bilinear inputs to linear networks
2 Network identification with latent nodesManifest and latent nodesAsymptotically exact identification via auto-regressive modelsLeast-squares auto-regressive estimation
J. Cortes (UC San Diego) Controllability and Identification of Complex Networks October 23, 2015 20 / 40
Asymptotic behavior of worst-case minimum input energy
Difficult-to-control (DTC) networks
A class of networks is difficult to control if, for fixed number of inputs,
limn→∞
supxf∈Rn
inf{u}∞:u(k)∈Rm
‖{u}∞‖2
‖xf ‖2→∞
i.e., worst-case minimum input energy grows unbounded with network size
For linear networks (A(n), 0n×nm,B(n)),
supxf∈Rn:‖xf ‖2=1
inf{u}∞:u(k)∈Rm
‖{u}∞‖2 = λ−1min(W1(n))
Stable symmetric linear networks are DTC [Pasqualetti et al (TCNS 14)]
λ−1min(W1(n)) increases exponentially with rate n
m for any choice ofB(n) ∈ Rn×m whose columns are canonical vectors in Rn
J. Cortes (UC San Diego) Controllability and Identification of Complex Networks October 23, 2015 21 / 40
Asymptotic behavior of worst-case minimum input energy
Difficult-to-control (DTC) networks
A class of networks is difficult to control if, for fixed number of inputs,
limn→∞
supxf∈Rn
inf{u}∞:u(k)∈Rm
‖{u}∞‖2
‖xf ‖2→∞
i.e., worst-case minimum input energy grows unbounded with network size
For linear networks (A(n), 0n×nm,B(n)),
supxf∈Rn:‖xf ‖2=1
inf{u}∞:u(k)∈Rm
‖{u}∞‖2 = λ−1min(W1(n))
Stable symmetric linear networks are DTC [Pasqualetti et al (TCNS 14)]
λ−1min(W1(n)) increases exponentially with rate n
m for any choice ofB(n) ∈ Rn×m whose columns are canonical vectors in Rn
J. Cortes (UC San Diego) Controllability and Identification of Complex Networks October 23, 2015 21 / 40
Does interconnection modulation help with DTC networks?
1mu− m
u1u
2u
Symmetric linear line networks with finitely many controlled nodes are DTC
1v
pv
1u
2u
m pu−
Symmetric linear line networks with finitely many controlled nodesAND finitely many controlled interconnections
J. Cortes (UC San Diego) Controllability and Identification of Complex Networks October 23, 2015 22 / 40
Does interconnection modulation help? Negative results
Symmetric linear networks remain difficult to control after
addition of finitely many bilinear inputs
with bounded homogeneous self-loop modulation
J. Cortes (UC San Diego) Controllability and Identification of Complex Networks October 23, 2015 23 / 40
Does interconnection modulation help? Negative results
Symmetric linear networks remain difficult to control after
addition of finitely many bilinear inputs
with bounded homogeneous self-loop modulation
DTC symmetric linear networks w/ finitely many bilinear inputs
Consider a class of DTC symmetric linear networks (A(n), 0n×nm,B(n))and let the number of nonzero entries in F (n) ∈ Rn×nm and‖F (n)‖max = maxi,j |Fij(n)| be uniformly bounded (with respect to n)
Then the class of bilinear networks (A(n),F (n),B(n)) is DTC
J. Cortes (UC San Diego) Controllability and Identification of Complex Networks October 23, 2015 23 / 40
Does interconnection modulation help? Negative results
Symmetric linear networks remain difficult to control after
addition of finitely many bilinear inputs
with bounded homogeneous self-loop modulation
DTC symmetric linear networks w/ self-loop modulation
Consider a class of DTC symmetric linear networks (A(n), 0n×nm,B(n))
with ρ(A(n)) <√
1− T−1m (Tm ,
⌈nm
⌉− 1)
Then the reachability Gramian of the class of bilinear networks(A(n), αIn,B(n)) with |tr(αIn)| ≤ µ satisfies, for any n > m−1µ2,
λmin(W) ≤ (1− Tmα2)−1
1− ρ2(A)− T−1m
ρ2Tm(A)
The term(1−Tmα
2)−1
1−ρ2(A)−T−1m
decreases in n and limn→∞(1−Tmα
2)−1
1−ρ2(A)−T−1m
= (1 − ρ2(A))−1
J. Cortes (UC San Diego) Controllability and Identification of Complex Networks October 23, 2015 23 / 40
Line network with self-loop modulation
Line networks with adjacency matricesA(n) with
aij =
{0.25 if |i − j | ≤ 1
0 otherwise
m = 1 and | tr(αIn)| = µ = 0.9
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15−18
−16
−14
−12
−10
−8
−6
−4
−2
0
2
Number of nodes (n)
log 1
0(λ
min(W
))
(A,0n×n, B) : log10(λmin (W1))
(A,αIn,B) : log10(λmin(W ))
J. Cortes (UC San Diego) Controllability and Identification of Complex Networks October 23, 2015 24 / 40
Does interconnection modulation help? Positive example
Line networks (A(n),B(n)) with
aij =
{0.05 if |i − j | ≤ 1
0 otherwiseB(n) = e1
plus single bilinear input F (n) with
fij =
{1 if i = j + 1
0 otherwise1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
−40
−35
−30
−25
−20
−15
−10
−5
0
5
Number of nodes
log 1
0(λ
min(W
))
(A,0n×n, B) : log10(λmin (W1))
(A,F,B) : log10(λmin(W))
DTC linear network can be made easy to control by adding single bilinear inputaffecting infinite number of interconnections (with constant strength)
J. Cortes (UC San Diego) Controllability and Identification of Complex Networks October 23, 2015 25 / 40
Outline
1 Reachability metrics for bilinear control systemsReachability GramianGramian-based reachability metrics and actuator selectionAdding bilinear inputs to linear networks
2 Network identification with latent nodesManifest and latent nodesAsymptotically exact identification via auto-regressive modelsLeast-squares auto-regressive estimation
J. Cortes (UC San Diego) Controllability and Identification of Complex Networks October 23, 2015 26 / 40
Identification of network topology from measured data
In many applications,network itself is unknown
Aim is to identify underlying dynamics or topology
1 2
1u 2u
3 4
3u 4u
1
2
3
4
( )
( )
( )
( )
u t
u t
u t
u t
1
2
3
4
( )
( )
( )
( )
x t
x t
x t
x t
Identify network using time series of stimulated node states
user may choose some of the inputs (control nodes) and measure some of thestates (observed nodes)
determine causal relationships among nodes: identify network transferfunction or adjacency matrix
J. Cortes (UC San Diego) Controllability and Identification of Complex Networks October 23, 2015 27 / 40
Network topology identificationIncomplete literature review
If every node manipulated&measured, large number of experiments allowed
M. Timme (PRL 07): identify networks of coupled phase oscillators
M. Nabi-Abdolyousefi, M. Mesbahi (IEEE TAC 12): identifyconsensus-type networks using node knockout
S. Shahrampour, V. Preciado (IEEE TAC 15): identify directed LTInetworks using node knockout
W/o ability to manipulate every node, perform large number of experiments, it isdifficult to accurately identify network topology
J. Goncalves, S. Warnick (IEEE TAC 08): not possible to identify evenBoolean structure of a general network using its transfer function
J. Cortes (UC San Diego) Controllability and Identification of Complex Networks October 23, 2015 28 / 40
Network topology identification-cont’dIncomplete literature review
Focus on particular network realizations (e.g., sparsest) that explain data
A. Julius, M. Zavlanos, S. Boyd, G. J. Pappas (Systems Biology09): genetic network identification using convex programming
D. Materassi, G. Innocenti, L. Giarre, M. Salapaka (SCL 13):trade-off between accuracy and sparsity
D. Materassi, G. Innocenti (IEEE TAC 10): identify tree networks
Our work inspired by wide use of auto-regressive (AR) models and time series toanalyze brain data via Granger causality
E. Bullmore, O. Sporns (Nature Reviews Neuroscience, 09): graphtheoretical analysis of structural and functional brain systems
S. L. Bressler, A. K. Seth (Neuroimage, 11): Wiener–Grangermethodology for determining causality
J. R. Iversen et al: (IEEE Engineering in Medicine and Biology, 14)Causal analysis of cortical networks involved in reaching to spatial targets
J. Cortes (UC San Diego) Controllability and Identification of Complex Networks October 23, 2015 29 / 40
Network model: linear system
Discrete-time, linear time-invariant state-space representation
x(k + 1) = Ax(k) + Bu(k)
y(k) = Cx(k)
manifest nodes: those whose state can be observedlatent nodes: those whose state cannot be observed
Assume manifest nodes can be controlled individually and latent nodes are passive(no control input)
[xm(k + 1)xl(k + 1)
]=
[A11 A12
A21 A22
] [xm(k)xl(k)
]+
[um(k)
0
]y(k) = xm(k)
J. Cortes (UC San Diego) Controllability and Identification of Complex Networks October 23, 2015 30 / 40
Network identification goal
Objective: reconstruct transfer function Txmum(ω) (from um to xm) formanifest subnetwork, absent knowledge of latent nodes
Characterize network behavior via auto-regressive (AR) model
Txmum(ω) : xm(k + 1) =τ−1∑i=0
Ai xm(k − i) + um(k)
How well can Txmum(ω) approximate original Txmum(ω)?
J. Cortes (UC San Diego) Controllability and Identification of Complex Networks October 23, 2015 31 / 40
Asymptotically exact identification via AR model
AR transfer function converges to manifest transfer function
If (1− ‖A11‖)(1− ‖A22‖) > ‖A12‖ · ‖A21‖, then there exists γ ∈ R>0 such that,for any τ ∈ Z≥0, there is a matrix sequence A∗0 , · · · , A∗τ−1 ∈ Rnm×nm satisfying
‖Txmum(ω, τ)− Txmum(ω)‖∞ ≤ γ‖A22‖τ
In fact, one such sequence is given by
A∗0 = A11, A∗i = −A12Ai−122 A21, i ≥ 1
J. Cortes (UC San Diego) Controllability and Identification of Complex Networks October 23, 2015 32 / 40
Asymptotically exact identification via AR model
AR transfer function converges to manifest transfer function
If (1− ‖A11‖)(1− ‖A22‖) > ‖A12‖ · ‖A21‖, then there exists γ ∈ R>0 such that,for any τ ∈ Z≥0, there is a matrix sequence A∗0 , · · · , A∗τ−1 ∈ Rnm×nm satisfying
‖Txmum(ω, τ)− Txmum(ω)‖∞ ≤ γ‖A22‖τ
In fact, one such sequence is given by
A∗0 = A11, A∗i = −A12Ai−122 A21, i ≥ 1
Stability margin vs interaction strength between manifest&latent subnetworks
as stability margin increasingly small, (1− ‖A11‖)(1− ‖A22‖) gets small
interaction strength between subnetworks (‖A12‖ · ‖A21‖) should beincreasingly small as well
J. Cortes (UC San Diego) Controllability and Identification of Complex Networks October 23, 2015 32 / 40
Asymptotically exact identification via AR model
AR transfer function converges to manifest transfer function
If (1− ‖A11‖)(1− ‖A22‖) > ‖A12‖ · ‖A21‖, then there exists γ ∈ R>0 such that,for any τ ∈ Z≥0, there is a matrix sequence A∗0 , · · · , A∗τ−1 ∈ Rnm×nm satisfying
‖Txmum(ω, τ)− Txmum(ω)‖∞ ≤ γ‖A22‖τ
In fact, one such sequence is given by
A∗0 = A11, A∗i = −A12Ai−122 A21, i ≥ 1
Exact identification: for acyclic latent subnetworks, matrix sequence ensures
Txmum = Txmum
J. Cortes (UC San Diego) Controllability and Identification of Complex Networks October 23, 2015 32 / 40
Direct versus latent interactions between manifest nodes
From linear state-space representation[xm(k + 1)xl(k + 1)
]=
[A11 A12
A21 A22
] [xm(k)xl(k)
]+
[um(k)
0
]
Given manifest nodes p and q, interaction is
direct iff A∗0(q, p) = A11(q, p) 6= 0
indirect (through latent nodes) iff A∗i (q, p) 6= 0 for some i ≥ 1
From A∗i = −A12Ai−122 A21, p first affects latent nodes (through
A21), then propagates through latent subnetwork (reflected byAi−1
22 ), and finally affects q (through A12)
If latent subnetwork is acyclic, A∗i (q, p) 6= 0 implies there are exactly i latentnodes in a path connecting p to q
J. Cortes (UC San Diego) Controllability and Identification of Complex Networks October 23, 2015 33 / 40
Direct versus latent interactions between manifest nodes
From linear state-space representation
xm(k + 1) =k∑
i=0
A∗i xm(k − i) + A12Ak22xl(0) + um(k)
Given manifest nodes p and q, interaction is
direct iff A∗0(q, p) = A11(q, p) 6= 0
indirect (through latent nodes) iff A∗i (q, p) 6= 0 for some i ≥ 1
From A∗i = −A12Ai−122 A21, p first affects latent nodes (through
A21), then propagates through latent subnetwork (reflected byAi−1
22 ), and finally affects q (through A12)
If latent subnetwork is acyclic, A∗i (q, p) 6= 0 implies there are exactly i latentnodes in a path connecting p to q
J. Cortes (UC San Diego) Controllability and Identification of Complex Networks October 23, 2015 33 / 40
Direct versus latent interactions between manifest nodes
From linear state-space representation
xm(k + 1) =k∑
i=0
A∗i xm(k − i) + A12Ak22xl(0) + um(k)
Given manifest nodes p and q, interaction is
direct iff A∗0(q, p) = A11(q, p) 6= 0
indirect (through latent nodes) iff A∗i (q, p) 6= 0 for some i ≥ 1
From A∗i = −A12Ai−122 A21, p first affects latent nodes (through
A21), then propagates through latent subnetwork (reflected byAi−1
22 ), and finally affects q (through A12)
If latent subnetwork is acyclic, A∗i (q, p) 6= 0 implies there are exactly i latentnodes in a path connecting p to q
J. Cortes (UC San Diego) Controllability and Identification of Complex Networks October 23, 2015 33 / 40
Outline
1 Reachability metrics for bilinear control systemsReachability GramianGramian-based reachability metrics and actuator selectionAdding bilinear inputs to linear networks
2 Network identification with latent nodesManifest and latent nodesAsymptotically exact identification via auto-regressive modelsLeast-squares auto-regressive estimation
J. Cortes (UC San Diego) Controllability and Identification of Complex Networks October 23, 2015 34 / 40
Least-squares AR estimation
Manifest subnetwork can be reconstructed
without knowledge of passive latent nodes, but
computation of matrix sequence requires network adjacency matrix
Given input/output data
persistently exciting input: {um}N1 zero-mean stochastic process withi.i.d. random vectors
measure response data sequence {y}N1
Least-squares AR estimation: given measurement data {y = xm}N1 ⊂ Rnm ,
determine {A}τ−10 minimizing 2-norm of residual sequence {e}N−1
τ
e(k) = y(k + 1)−τ−1∑i=0
Aiy(k − i)
J. Cortes (UC San Diego) Controllability and Identification of Complex Networks October 23, 2015 35 / 40
Least-squares AR estimation
Manifest subnetwork can be reconstructed
without knowledge of passive latent nodes, but
computation of matrix sequence requires network adjacency matrix
Given input/output data
persistently exciting input: {um}N1 zero-mean stochastic process withi.i.d. random vectors
measure response data sequence {y}N1
Least-squares AR estimation: given measurement data {y = xm}N1 ⊂ Rnm ,
determine {A}τ−10 minimizing 2-norm of residual sequence {e}N−1
τ
e(k) = y(k + 1)−τ−1∑i=0
Aiy(k − i)
J. Cortes (UC San Diego) Controllability and Identification of Complex Networks October 23, 2015 35 / 40
Convergence of least-squares estimate
Least-squares estimate converges in probability
Least-squares estimate satisfies
‖plimN→∞ Aτ ({y}N1 )− A∗τ‖max ≤ βτ‖A22‖τ
β only depends on network adjacency matrix and input covariance σ2um
Direct/indirect interactions between manifest nodes can be ascertained from
A0 → A∗0 = A11 and Ai → A∗i = −A12Ai−122 A21, i ≥ 1
Least-squares estimate converges in mean square
If latent subnetwork is acyclic (Aτ2222 = 0nl×nl ), then for τ ≥ τ22 + 1,
least-squares estimate satisfies
limN→∞
E[(Aτ ({y}N1 )− A∗τ )T (Aτ ({y}N1 )− A∗τ )] = 0nmτ×nmτ
J. Cortes (UC San Diego) Controllability and Identification of Complex Networks October 23, 2015 36 / 40
Convergence of least-squares estimate
Least-squares estimate converges in probability
Least-squares estimate satisfies
‖plimN→∞ Aτ ({y}N1 )− A∗τ‖max ≤ βτ‖A22‖τ
β only depends on network adjacency matrix and input covariance σ2um
Direct/indirect interactions between manifest nodes can be ascertained from
A0 → A∗0 = A11 and Ai → A∗i = −A12Ai−122 A21, i ≥ 1
Least-squares estimate converges in mean square
If latent subnetwork is acyclic (Aτ2222 = 0nl×nl ), then for τ ≥ τ22 + 1,
least-squares estimate satisfies
limN→∞
E[(Aτ ({y}N1 )− A∗τ )T (Aτ ({y}N1 )− A∗τ )] = 0nmτ×nmτ
J. Cortes (UC San Diego) Controllability and Identification of Complex Networks October 23, 2015 36 / 40
Least-squares recovers manifest transfer function
Least-squares consistently estimates manifest transfer function
There exist positive scalars β, γ and τ0 such that, for τ ≥ τ0,
‖plimN→∞ Tye({y}N1 , τ)− Txmum‖∞ ≤ (βτ 2 + γ)‖A22‖τ
As a consequence, plimN→∞,τ→∞ Tye({y}N1 , τ) = Txmum
Critical model order τ0 increases with ‖A22‖:less stable latent subnetworks require higher order of the AR model
Exact identification for acyclic latent subnetworks
If latent subnetwork is acyclic (Aτ2222 = 0nl×nl ), then for τ ≥ τ22 + 1,
plimN→∞ Tye({y}N1 , τ) = Txmum
J. Cortes (UC San Diego) Controllability and Identification of Complex Networks October 23, 2015 37 / 40
Least-squares recovers manifest transfer function
Least-squares consistently estimates manifest transfer function
There exist positive scalars β, γ and τ0 such that, for τ ≥ τ0,
‖plimN→∞ Tye({y}N1 , τ)− Txmum‖∞ ≤ (βτ 2 + γ)‖A22‖τ
As a consequence, plimN→∞,τ→∞ Tye({y}N1 , τ) = Txmum
Critical model order τ0 increases with ‖A22‖:less stable latent subnetworks require higher order of the AR model
Exact identification for acyclic latent subnetworks
If latent subnetwork is acyclic (Aτ2222 = 0nl×nl ), then for τ ≥ τ22 + 1,
plimN→∞ Tye({y}N1 , τ) = Txmum
J. Cortes (UC San Diego) Controllability and Identification of Complex Networks October 23, 2015 37 / 40
Example: Erdos–Renyi random network
Group of G (10, 0.25) Erdos–Renyi random networks
3 randomly chosen manifest nodes each
edge weight with uniform distribution in (0, 0.35)
N = 105 data points
24
68
24
68
10
0.1
0.2
0.3
0.4
0.5
Model orderNetwork index
H∞ n
orm
err
or
1 2 3 4 5 6
0.05
0.1
0.15
0.2
0.25
Model candidate order
H∞ n
orm
err
or
LSAR estimationAR model in Theorem 1
‖Txmum − Txmum‖∞ Network w/ index 1 Comparison w/ opt matrix seq
J. Cortes (UC San Diego) Controllability and Identification of Complex Networks October 23, 2015 38 / 40
Example: Erdos–Renyi random network
Group of G (10, 0.25) Erdos–Renyi random networks
3 randomly chosen manifest nodes each
edge weight with uniform distribution in (0, 0.35)
N = 105 data points
24
68
24
68
10
0.1
0.2
0.3
0.4
0.5
Model orderNetwork index
H∞ n
orm
err
or
1 2 3 4 5 6
0.05
0.1
0.15
0.2
0.25
Model candidate order
H∞ n
orm
err
or
LSAR estimationAR model in Theorem 1
‖Txmum − Txmum‖∞ Manifest network Comparison w/ opt matrix seq
J. Cortes (UC San Diego) Controllability and Identification of Complex Networks October 23, 2015 38 / 40
Summary
Reachability for bilinear control systems
inputs might not only have effect on state of nodes but also on theinteractions among nodes
Gramian-based reachability metrics capture ease of control
actuator selection
linear networks that remain difficult to control after addition of bilinear inputs
Network identification with latent nodes
linear model with manifest and latent nodes
existence of AR models that asymptotically identify manifest transfer function
reconstruction of direct/indirect interactions among manifest nodes
least-squares auto-regressive estimate: performance in probability
Y. Zhao and J. Cortes. Gramian-based reachability metrics for bilinear networks.IEEE Transactions on Control of Network Systems, 2015.Submitted
Y. Zhao and J. Cortes. Identification of linear networks with latent nodes.In American Control Conference, Boston, MA, July 2016.Submitted
J. Cortes (UC San Diego) Controllability and Identification of Complex Networks October 23, 2015 39 / 40
Plethora of open questions
Reachability for bilinear control systems
dependence of Gramian and metrics on network. Asymptotic results withnetwork scale
identify networks for which bilinear control does quantitatively better
robustness of reachability metrics against malfunctions/failures
optimal selection of actuators
. . .
Network identification with latent nodes
active latent nodes with unknown control inputs
linear network models of higher (unknown, time-varying) order
identification of properties that make networks easier to identify
applications to network science: brain and social networks
. . .
J. Cortes (UC San Diego) Controllability and Identification of Complex Networks October 23, 2015 40 / 40