Geometry, Analysis and Computation

for Network Systems

Francesco Bullo
Department of Mechanical Engineering

Center for Control, Dynamical Systems & Computation

University of California at Santa Barbara

http://motion.me.ucsb.edu

Workshop on Resilient Control of Infrastructure Networks

September 24-27, 2019 - DISMA, Politecnico di Torino
Acknowledgments

Saber Jafarpour Xiaoming Duan Kevin D. Smith

UCSB UCSB UCSB

NSF AFOSR ARO ONR DOE

 Lectures on Network Systems

Lectures on Network Systems, Francesco Bullo,

Lectures on Createspace, 1 edition, 2018, ISBN 978-1-986425-64-3
Lectures on Network Systems

Network Systems
1. Self-Published and Print-on-Demand at:
https://www.amazon.com/dp/1986425649

2. PDF Freely available at

http://motion.me.ucsb.edu/book-lns:
For students: free PDF for download
For instructors: slides, classnotes, and answer keys

3. incorporates lessons from 2 decades of research:

robotic multi-agent, social networks, power grids

4. now v1.3
v2.0 will expand nonlinear coverage
Francesco Bullo
Francesco Bullo

316 pages
With contributions by 205 pages solution manual
Jorge Cortés
Florian Dörfler 4.4K downloads Jun 2016-Aug 2019
Sonia Martínez 164 exercises with solutions
33 instructors in 15 countries
Outline

Linear Network Systems and Metzler Matrices

X. Duan, S. Jafarpour, and F. Bullo. Graph-theoretic small gain theorems for
1 Metzler matrices and monotone systems.
IEEE Transactions on Automatic Control, June 2019.
Submitted.
URL: https://arxiv.org/pdf/1905.05868.pdf

2
An emerging theory for Nonlinear Network Systems

3
Kuramoto Synchronization (existence and lack of uniqueness)
Linear network systems

ẋ(t) = Ax(t)

network structure ⇐⇒ function = asymptotic behavior

Model Dynamics Asy Behavior Graph property

averaging flow ẋ = −Lx consensus ∃ globally reach node
(Abelson ’64) Laplacian matrix
network flow ẋ = −L> x stationary dis- ∃ globally reach node
(Noy Meir ’73) transpose Laplacian tribution
network flow with ẋ = Cx stability outflow-connected
decay (outflows) C = −L> − diag(d)
compartmental matrix
network flow with ẋ = Mx stability unknown
decay/growth M = −L> + diag(g − d)
Metzler matrix
Network flow systems
X
precipitation soil evaporation, drainage, runo↵ q̇i = (Fj→i − Fi→j ) − Fi→0 + ui
j
uptake plants transpiration
Fi→j = fij qi , F = [fij ]
drinking herbivory

animals evaporation
q̇ = F T − diag(F 1n + f0 ) q + u


Water flow model for a desert ecosystem (Noy-Meir ’73) | {z }
=: C

C compartmental matrix:
quasi-positive (off-diag ≥ 0) and non-positive column sums (f0 ≥ 0)
analysis tools: PF for quasi-positive, inverse positivity, algebraic graph

system (= each condensed sink) C is Hurwitz

is outflow-connected

limt→∞ q(t) = −C −1 u ≥ 0
(−C −1 u)i > 0 ⇐⇒ ith compartment is inflow-connected
Stability of network flow systems

A Metzler M is Hurwitz iff any following equivalent condition hold:

1 there exists ξ ∈ Rn such that ξ > 0n and Mξ < 0n ;
2 there exists η ∈ Rn such that η > 0n and η > M < 0>
n ; or
3 there exists a diagonal matrix P  0 such that M > P + PM ≺ 0.

(a) maxi∈{1,...,n} xi /ξi (b) η > x (c) x > Px

Goal: graph-theoretic conditions for stability

Reducible and acyclic graphs
Reducible graphs

M ∈ Rn×n is Hurwitz
m
Strongly connected components
are Hurwitz

Implication: large-scale system may be decomposed into smaller systems

Directed acyclic graphs

M ∈ Rn×n is Hurwitz
m
diagonal entries are negative

Implication: study cycles!

Basic ideas: a simple cycle

m11 1
  m12
m11 m12 0 ··· 0
 0 m22 m23 ··· 0 mn1
 mnn m22

 n 2
M =  ... .. .. .. ..
 
 . . . .

 0 0 ··· mn−1,n−1 mn−1,n  ···
mn1 0 ··· 0 mnn
mii

    
m12 m23 mn1
M Hurwitz ⇐⇒ ... <1
−m11 −m22 −mnn

where
mij
−mii represents a “gain” for subsystem i with respect to j
test: composition of “gains” along the cycle is less than 1
Basic ideas: Small-gain network stability
Cyclic Small-Gain Theorem
a network of systems with input is ISS if

cycle gain < 1

about each simple cycle,

for appropriate interconnection gains

1 V. Lakshmikantham, V. M. Matrosov, and S. Sivasundaram. Vector Lyapunov

Functions and Stability Analysis of Nonlinear Systems.
Kluwer Academic Publishers, 1991
2 S. N. Dashkovskiy, B. S. Rüffer, and F. R. Wirth. Small gain theorems for large
scale systems and construction of ISS Lyapunov functions.
SIAM Journal on Control and Optimization, 48(6):4089–4118, 2010.
doi:10.1137/090746483
3 T. Liu, D. J. Hill, and Z.-P. Jiang. Lyapunov formulation of ISS cyclic-small-gain in
continuous-time dynamical networks.
Automatica, 47(9):2088–2093, 2011.
Summary of results

Thm 1: Input-to-state interconnection gains for Metzler systems

Thm 2: Max-interconnection gains and graph-theoretic conditions
Thm 3: Sum-interconnection gains and graph-theoretic conditions

X. Duan, S. Jafarpour, and F. Bullo. Graph-theoretic small gain theorems for Metzler
matrices and monotone systems.
IEEE Transactions on Automatic Control, June 2019.
Submitted.
URL: https://arxiv.org/pdf/1905.05868.pdf
Possible notions of ISS gains
An interconnected nonlinear system with subsystem dynamics

ẋi = fi (xi , xNi , ui ), ∀i ∈ {1, . . . , n}.

system has sum-interconnection gains {γij } if

X
|xi (t)| ≤ βi (|xi (0)|, t) + γij (kxj k[0,t] ) + γi (kui k∞ ).
j∈Ni

where βi ∈ KL, γij ∈ K, and γi ∈ K.

system has max-interconnection gains {ψij } if

|xi (t)| ≤ max βi0 (|xi (0)|, t), ψij (kxj k[0,t] ), ψi (kui k∞ ) .

j∈Ni

where βi ∈ KL, ψij ∈ K, and ψi ∈ K.

Thm 1: ISS gains for Metzler systems

Thm 1: ISS gains for Metzler systems

For Metzler system ẋ = Mx + u, M with negative diagonals,
1 sum-interconnection gains {γ } satisfy
ij

mij
≤ γij , ∀i ∈ {1, . . . , n}, j ∈ Ni
−mii
2 max-interconnection gains {ψij } satisfy
X  mij 
ψij−1 < 1, ∀i ∈ {1, . . . , n}
−mii
j∈Ni

For c = (i1 , i2 , . . . , ik , i1 ) be a simple cycle

1 the sum-cycle gain of c is γc = (γi2 i1 ) (γi3 i2 ) . . . (γi1 ik )
2 a max-cycle gain of c is ψc = (ψi2 i1 ) (ψi3 i2 ) . . . (ψi1 ik )
Thm 2: Max-cycle gains and graph conditions

Thm 2: Conditions based on max-cycle gains

Given an irreducible Metzler matrix M ∈ Rn×n with negative diagonal
elements and the set of simple cycles Φ, the followings are equivalent:
1 M is Hurwitz;
2 for every i ∈ V and j ∈ Ni , there exists ψij > 0 such that
X  mij 
ψij−1 < 1, ∀i ∈ {1, . . . , n},
−mii
j∈Ni

ψc < 1, ∀c ∈ Φ.

“cycle gain < 1 about each simple cycle” is now IFF

convex problem
Thm 3: Sum-cycle gains and graph conditions

Thm 3: Conditions based on sum-cycle gains

Given an irreducible Metzler matrix M ∈ Rn×n with negative diagonal
elements, the followings are equivalent:
1 M is Hurwitz;
2 for each i, let Φi be simple cycles over {1, . . . , i} (or renumbered)
X X X
γc1 − γc1 γc2 + · · · + (−1)ri −1 γc1 . . . γcri < 1
c1 ∈Φi {c1 ,c2 }⊂Φi {c1 ,...,cri }⊂Φi
c1 ∩c2 =∅ ci ∩cj =∅

condition 2 ⇐⇒ certain sums of products of gains < 1

computation of sum-cycle gains and “sums of products” is
straightforward (not iterative)
Thm 3: Example

1 V1 = {1} =⇒ ∅
c1 V2 = {1, 4} =⇒ {γc1 < 1}
c4 V3 = {1, 4, 2} =⇒ {γc1 + γc4 < 1}
4 2
V4 = {1, 4, 2, 3} =⇒ {γc1 + γc4 < 1,
c2
c3 γc1 + γc2 + γc3 + γc4 − γc1 γc3 < 1}
3

Hence, stability certificate

γc 1 + γc 4 < 1
γc1 + γc2 + γc3 + γc4 − γc1 γc3 < 1
Outline

1
Linear Network Systems and Metzler Matrices

An emerging theory for Nonlinear Network Systems

2 F. Bullo. Lectures on Network Systems.
Kindle Direct Publishing, 1.3 edition, July 2019.
With contributions by J. Cortés, F. Dörfler, and S. Martı́nez.
URL: http://motion.me.ucsb.edu/book-lns

3
Kuramoto Synchronization (existence and lack of uniqueness)
Nonlinear network systems

Rich variety of emerging behaviors

1 equilibria / limit cycles / extinction in populations dynamics
2 epidemic outbreaks in spreading processes
3 synchrony and multi-stability in coupled oscillators

Rich variety of analysis tools

1 nonlinear stability theory
2 passivity, small gain theorems, and dissipativity
3 contractivity and monotonicity

(infection rate)
Susceptible Infected
(recovery rate)
Example: Population systems in ecology
(Vito Volterra, Universita’ di Torino, 1860-1940)

Lotka-Volterra: xi = quantity/density

ẋi X
= bi + aij xj
xi j

Mutualism clownfish / anemones (Takeuchi et al ’78) ẋ = diag(x) Ax + b)

interaction matrix A:
(+, +) mutualism, (+, −) predation, (−, −) competition
rich behavior: persistence, extinction, equilibria, periodic orbits, . . .

1 mutualism: aij ≥ 0
2 either unbounded evolution or
exists unique steady state −A−1 b > 0
limt→∞ x(t) = −A−1 b from all x(0) > 0
Dichotomy in mutualistic Lotka-Volterra system

x2 -null-line x1 -null-line

x2 -null-line

r2 /a22
r2 /a22

x⇤2 =
x⇤2 =

x1 -null-line

x⇤1 = r1 /a11 x⇤1 = r1 /a11

Case I: a12 > 0, a21 > 0, Case II: a12 > 0, a21 > 0,
a12 a21 > a11 a22 . There exists no a12 a21 < a11 a22 . There exists a
positive equilibrium point. All unique positive equilibrium point.
trajectories starting in R2>0 diverge. All trajectories starting in R2>0
converge to the equilibrium point.
Research questions in Nonlinear Network Systems

1 what are key example systems?

2 what is a useful underlying structure?

3 what is a practical, simple, rich technical approach?

4 how do we treat dichotomy and richer behaviors?
5 how do we automatically generate Lyapunov functions?
Example systems
Kuramoto oscillators (’75) Yorke network propagation (’76)
Xn
θ̇i = ωi − aij sin(θi − θj )
j=1

ẋ = β In − diag(x) Ax − γx
Metzler Jac: phase cohesive region
Ex: active power flow, motion patterns Metzler Jac and positive
Ex: network SIR, patchy SIS
Lotka-Volterra population (’20) Daganzo cell transmission (’94)
ρ̇e = fein (ρ) − feout (ρ)


ẋ = diag(x) Ax + r

Metzler Jac: mutualistic interactions Metzler Jac: free flow region

Ex: biochemical networks, repressilator Ex: monotone distributed routing (Como,
with 2 genes Savla, et al), Maeda ’78, Sandberg ’78
Matrosov interconnection of ISS systems (’71)
ẋi = fi (x1 , . . . , xn , ui ) =⇒ v̇ ≤ −A(v ) + Γ(v ) + G (w )

Metzler Jac and positive

A review of Contraction Theory
given norm, the matrix measure of A is
kIn + hAk − 1
µ(A) := lim+
h→0 h
assume: vector field f is infinitesimally contracting over C , that is,


µ Df (x) ≤ c < 0, for all x ∈ C

assume: set C is f -invariant, closed and convex

Desirable consequences
1 flow of f is a contraction, i.e.,
distance between solutions exponentially decreases with rate c
2 there exists an equilibrium x ∗ , unique, globally exponentially stable
with global Lyapunov functions

x 7→ kx − x ∗ k2 and x 7→ kf (x)k2
 <latexit sha1_base64="QZhN6tfOg1v6HnMBc8DmxzG6IAE=">AAAB/HicbZDLSgMxFIYz9VbrrerSTbAIFUqZqYLiquDGZQV7gelQMmmmDc0kQ5IpDkN9C7e6cSdufRfBhzGdzkJbDwQ+/v8czsnvR4wqbdtfVmFtfWNzq7hd2tnd2z8oHx51lIglJm0smJA9HynCKCdtTTUjvUgSFPqMdP3J7dzvTolUVPAHnUTEC9GI04BipI3k9qMxrepaMrDPB+WKXbezgqvg5FABebUG5e/+UOA4JFxjhpRyHTvSXoqkppiRWakfKxIhPEEj4hrkKCSqpqajDLz0MTt+Bs+MN4SBkOZxDTP192yKQqWS0DedIdJjtezNxf88N9bBtZdSHsWacLxYFMQMagHnScAhlQRrlhhAWFJzNcRjJBHWJq+SicNZ/vwqdBp156LeuL+sNG/yYIrgBJyCKnDAFWiCO9ACbYCBAM/gBbxaT9ab9W59LFoLVj5zDP6U9fkDZCeUvg==</latexit>
(t, y0 )

y0
<latexit sha1_base64="Q+kR5PobiieHyykm1KTcp0E+PlU=">AAAB6nicbVBNS8NAEJ3Ur1q/qh69LBbBU0mqoHgqePFY0X5AG8pmu2mXbjZhdyKE0p/gxYMiXv1F3vw3btsctPXBwOO9GWbmBYkUBl332ymsrW9sbhW3Szu7e/sH5cOjlolTzXiTxTLWnYAaLoXiTRQoeSfRnEaB5O1gfDvz209cGxGrR8wS7kd0qEQoGEUrPWR9t1+uuFV3DrJKvJxUIEejX/7qDWKWRlwhk9SYrucm6E+oRsEkn5Z6qeEJZWM65F1LFY248SfzU6fkzCoDEsbalkIyV39PTGhkTBYFtjOiODLL3kz8z+umGF77E6GSFLlii0VhKgnGZPY3GQjNGcrMEsq0sLcSNqKaMrTplGwI3vLLq6RVq3oX1dr9ZaV+k8dRhBM4hXPw4ArqcAcNaAKDITzDK7w50nlx3p2PRWvByWeO4Q+czx8KmI2c</latexit>

(t, x0 )
<latexit sha1_base64="61UE9xPCc52PjErOqMKwQccv1Cc=">AAAB8nicbVC7SgNBFJ2NrxhfUTttFoMQQcJuLBSrgI2NEME8YLOE2clsMmR2Zpm5K4YlWPkJNhaK2PoNfoSdP2Lt5FFo4oELh3Pu5d57gpgzDY7zZWUWFpeWV7KrubX1jc2t/PZOXctEEVojkkvVDLCmnAlaAwacNmNFcRRw2gj6FyO/cUuVZlLcwCCmfoS7goWMYDCS14p7rAjHd23nqJ0vOCVnDHueuFNSqOw9FL7vrz6q7fxnqyNJElEBhGOtPdeJwU+xAkY4HeZaiaYxJn3cpZ6hAkdU++n45KF9aJSOHUplSoA9Vn9PpDjSehAFpjPC0NOz3kj8z/MSCM/8lIk4ASrIZFGYcBukPfrf7jBFCfCBIZgoZm61SQ8rTMCklDMhuLMvz5N6ueSelMrXJo1zNEEW7aMDVEQuOkUVdImqqIYIkugRPaMXC6wn69V6m7RmrOnMLvoD6/0H13OT9w==</latexit>

x0
<latexit sha1_base64="bBxD3sM5gabNhoCQ5fwtTlnKoN0=">AAAB6nicbVBNS8NAEJ34WetX1aOXxSJ4KkkVFE8FLx4r2g9oQ9lsN+3SzSbsTsQS+hO8eFDEq7/Im//GbZuDtj4YeLw3w8y8IJHCoOt+Oyura+sbm4Wt4vbO7t5+6eCwaeJUM95gsYx1O6CGS6F4AwVK3k40p1EgeSsY3Uz91iPXRsTqAccJ9yM6UCIUjKKV7p96bq9UdivuDGSZeDkpQ456r/TV7ccsjbhCJqkxHc9N0M+oRsEknxS7qeEJZSM64B1LFY248bPZqRNyapU+CWNtSyGZqb8nMhoZM44C2xlRHJpFbyr+53VSDK/8TKgkRa7YfFGYSoIxmf5N+kJzhnJsCWVa2FsJG1JNGdp0ijYEb/HlZdKsVrzzSvXuoly7zuMowDGcwBl4cAk1uIU6NIDBAJ7hFd4c6bw4787HvHXFyWeO4A+czx8JEo2b</latexit>

Figure: Any two trajectories of an infinitesimally contracting system converge.

Common matrix measures
Vector norm Matrix measure
Xn  Xn 
kxk1 = |xi | µ1 (A) = max ajj + |aij |
i=1 j∈{1,...,n} i=1,i6=j
= max column “absolute sum” of A
r
Xn  A + A> 
kxk2 = x2 µ2 (A) = λmax
i=1 i 2
 Xn 
kxk∞ = max |xi | µ∞ (A) = max aii + |aij |
i∈{1,...,n} i∈{1,...,n} j=1,j6=i
= max row “absolute sum” of A

Simplifications for a Metzler matrix M

Xn
µ1 (M) = max mij = max(M > 1n ) = max column sum of M
j∈{1,...,n} i=1
Xn
µ∞ (M) = max mij = max(M1n ) = max row sum of M
i∈{1,...,n} j=1
The Euclidean case: works by Krasovskiı̆ & Vidyasagar

Vidyasagar ’78: Lyapunov functions and matrix measures

Given P  0 and c ∈ R,

µ2,P (A) < c ⇐⇒ A> P + PA ≺ 2cP

1 A Hurwitz ⇐⇒ A has negative weighted 2-norm (w.r.t. some P)

2 inf µ2,P (A) = spectral abscissa of A
P0

Krasovskiı̆ ’60: method to design Lyapunov function

f is weighted 2-norm contracting if ∃P  0 and c < 0

P Df (x) + Df (x)> P  2cP, for all x ∈ Rn

Constant Lyapunov weight P at each x implies desirable consequences

The non-Euclidean case for Metzler Jacobians
Coogan ’16: matrix measures of a Metzler matrix M
Given vectors η, ξ > 0m and c ∈ R,

µ1,diag(η) (M) <c ⇐⇒ η > M < cη > , and

µ∞,diag(ξ)−1 (M) < c ⇐⇒ Mξ < cξ,

1 M Hurwitz ⇐⇒ M has negative weighted 1- or ∞-measure

2 inf µ1,diag(η) (M) = inf µ∞,diag(ξ)−1 (M) = spectral abscissa of M
η>0m ξ0m

Sum-separable and max-separable Lyapunov functions

f with Metzler Jac is weighted 1-norm contracting if ∃η > 0n and c < 0

η > Df (x) ≤ cη > , for all x ∈ Rn

Constant column weights η at each x implies desirable consequences

Krasovskiı̆ Lyapunov functions
for systems with Metzler Jacobians and constant weights
Weighted diagonal 2-norm:
n
X n
X
kx − x ∗ k2P = pi (xi − xi ) 2
and kf (x)k2P = pi fi (x)2
i=1 i=1

Weighted 1-norm
n
X n
X
kx − x ∗ k1,η = ηi |xi − xi∗ | and kf (x)k1,η = ηi |fi (x)|
i=1 i=1

Weighted ∞-norm

|xi − xi∗ | |fi (x)|

kx − x ∗ k∞,ξ−1 = max and kf (x)k∞,ξ−1 = max
i∈{1,...,n} ξi i∈{1,...,n} ξi

Recall: sublevel sets of Lyapunov functions are f -invariant

Example application to Lotka-Volterra

1 change of variable y = ln x, so that x ∈ Rn>0 maps into y ∈ Rn and

ẏ = A exp(y ) + r := fLVe (y )

2 pick v > 0n such that v > A < 0n and show

v > DfLVe (y ) = v > A diag(exp(y )) < −cv > diag(exp(y )) ≤ 0.

3 fLVe , and so fLV , has a unique globally exponentially stable equilibrium

with sum-separable global Lyapunov functions

ky − y ∗ k1,diag(v ) and kfLVe (y )k1,diag(v )

that is,
n
X n
X
x 7→ vi | ln(xi /xi∗ )|, x 7→ vi |(Ax + r )i |
i=1 i=1
Why is this relevant for infrastructure networks?

Consider a network flow system ẋ = f (x) preserving a commodity

constant = 1>
n x(t)
=⇒ 0 = 1> >
n ẋ(t) = 1n f (x(t))
=⇒ 0n = 1>
n Df x(t)

If additionally f has Meztler Jacobian, then f is automatically weakly

contracting (non-expansive) with respect to the `1 norm.
Outline
1
Linear Network Systems and Metzler Matrices

2
An emerging theory for Nonlinear Network Systems

Kuramoto Synchronization (existence)

3
S. Jafarpour and F. Bullo. Synchronization of Kuramoto oscillators via cutset projections.
IEEE Transactions on Automatic Control, 64(7):2830–2844, 2019.
doi:10.1109/TAC.2018.2876786

1 problem statement
2 solution

Kuramoto Multi-Stability (lack of uniqueness)

S. Jafarpour, E. Y. Huang, K. D. Smith, and F. Bullo. Multistable synchronous power
4 flows: From geometry to analysis and computation.
SIAM Review, January 2019.
Submitted.
URL: https://arxiv.org/pdf/1901.11189.pdf
 Today: Sync & Multi-Stability in Coupled Oscillators

Kuramoto model
n oscillators with angle θi ∈ S1
non-identical natural frequencies ωi ∈ R1
coupling with strength aij = aji

n
X
θ̇i = ωi − aij sin(θi − θj )
j=1
Model #1: Spring network analog and applications

⌧3
Coupled swing equations
k34
⌧4 Euler-Lagrange eq for spring network on ring:
k23
X
mi θ̈i + di θ̇i = τi − kij sin(θi − θj )
⌧2 k24 j
k12

⌧1

Kuramoto coupled oscillators

X
θ̇i = ωi − aij sin(θi − θj )
j

Kuramoto equilibrium equation

X
0 = ωi − aij sin(θi − θj )
j
Model #2: Active Power Flow Problem

AC, Kirckhoff and Ohm, quasi-sync, lossless lines, constant voltages.

supply/demand pi , max power coeff aij , voltage phase θi
Xn
pi = fij , fij = aij sin(θi − θj )
j=1

aij
power flow
<latexit sha1_base64="SLy9HAEioHHCM36ABtCERYc8mnE=">AAAB43icbVDLSgNBEOyNrxhfUY9eBoPgKeyKYI4BLx4jmAckIcxOepNJZh/M9AphyRd4EfGi4P/4C/6Nk2QviRYMFFU1dFf7iZKGXPfHKWxt7+zuFfdLB4dHxyfl07OWiVMtsCliFeuOzw0qGWGTJCnsJBp56Cts+9P7hd9+Rm1kHD3RLMF+yEeRDKTgZKUWH2RyMh+UK27VXYL9JV5OKpCjMSh/94axSEOMSChuTNdzE+pnXJMUCuelXmow4WLKR5gtd5yzKysNWRBr+yJiS3Utx0NjZqFvkyGnsdn0FuJ/XjeloNbPZJSkhJFYDQpSxShmi8JsKDUKUjNLuNDSbsjEmGsuyJ6lZKt7m0X/ktZN1XOr3uNtpV7Lj1CEC7iEa/DgDurwAA1ogoAJvMIHfDrovDhvzvsqWnDyP+ewBufrF6bwiuI=</latexit>

fij
<latexit sha1_base64="g8st85giIEjIQxoaP6DXlhF7+Vc=">AAACE3icbZC7TsMwFIYdriXcCowsFi0SA6oSFjpWYmEsEr1ITVU57klr6tiR7VCqqCtvwFuwwsKGWHkAJB4G9zJAyz99Ov859jl/mHCmjed9OSura+sbm7ktd3tnd28/f3BY1zJVFGpUcqmaIdHAmYCaYYZDM1FA4pBDIxxcTfzGPSjNpLg1owTaMekJFjFKjC118jigIAwoJnpuIoegcMTlMAjcYtTJ2N242MkXvJI3FV4Gfw4FNFe1k/8OupKmsX2WcqJ1y/cS086IMoxyGLtBqiEhdEB60LIoSAz6XN/3ptDOHqY3jfGp9bo4knYhKQyeVn/PZiTWehSHtjMmpq8XvUnxP6+VmqjczphIUgOCzj6KUo6NxJOAcJcpoIaPLBCqmN0a0z5RhNqUtGvj8BePX4b6Rcn3Sv7NRaFSngeTQ8foBJ0hH12iCrpGVVRDFD2iZ/SCXp0n5815dz5mrSvOfOYI/ZHz+QNIi53n</latexit>

fij
<latexit sha1_base64="7IyjuMywHDGDQ4Z9D8xaM6McRGw=">AAAB9HicbZDLSgMxFIbP1Futt6pLN8EiuJCSEUGXBTcuK9oLtEPJpJlpaCYzJJliGfoIbnXjTtz6PoIPYzqdhbYeCHz8/zmck99PBNcG4y+ntLa+sblV3q7s7O7tH1QPj9o6ThVlLRqLWHV9opngkrUMN4J1E8VI5AvW8ce3c78zYUrzWD6aacK8iISSB5wSY6WHZMAH1Rqu47zQKrgF1KCo5qD63R/GNI2YNFQQrXsuToyXEWU4FWxW6aeaJYSOSch6FiWJmL7QkzAHL3vKj56hM+sNURAr+6RBufp7NiOR1tPIt50RMSO97M3F/7xeaoIbL+MySQ2TdLEoSAUyMZongIZcMWrE1AKhiturER0RRaixOVVsHO7y51ehfVl3cd29v6o1cBFMGU7gFM7BhWtowB00oQUUQniGF3h1Js6b8+58LFpLTjFzDH/K+fwBdvuSAg==</latexit>
pi <latexit sha1_base64="WTNzHc9Xoqra4AzTSjaHSwxqr0w=">AAAB93icbZDNSgMxFIXv1L9a/6ou3QSL4EJKRgRdFty4rGBboR1KJs20aTOZIckUh6Hv4FY37sStjyP4MKbTWWjrhcDHOfdyb44fC64Nxl9OaW19Y3OrvF3Z2d3bP6geHrV1lCjKWjQSkXr0iWaCS9Yy3Aj2GCtGQl+wjj+5nfudKVOaR/LBpDHzQjKUPOCUGCu1g37Gx7N+tYbrOC+0Cm4BNSiq2a9+9wYRTUImDRVE666LY+NlRBlOBZtVeolmMaETMmRdi5KETF/o6TAHL3vK756hM+sNUBAp+6RBufp7NiOh1mno286QmJFe9ubif143McGNl3EZJ4ZJulgUJAKZCM1DQAOuGDUitUCo4vZqREdEEWpsVBUbh7v8+VVoX9ZdXHfvr2oNXARThhM4hXNw4RoacAdNaAGFMTzDC7w6qfPmvDsfi9aSU8wcw59yPn8AAbyTeA==</latexit>

power angle ✓i
<latexit sha1_base64="7RFTMqMbRTAIXDFIkTcXgxfnB2c=">AAACEnicbZC/TsMwEMad8q+UfwFGBixaJAaoEhYYkVgYQaK0UltFjnttDY4T2RegqjryBrwFKyxsiJUXQOJhcNMMQPkkSz/dd6fzfWEihUHP+3QKM7Nz8wvFxdLS8srqmru+cWXiVHOo8VjGuhEyA1IoqKFACY1EA4tCCfXw5nTs129BGxGrSxwk0I5YT4mu4AxtKXC3k/gONGWqJ4FWWtgHZIE4yOG6Erhlr+plotPg51Amuc4D96vViXkagUIumTFN30uwPWQaBZcwKrVSAwnjN6wHTYuKRWD2zW0vg/bwPjtpRHet16HdWNunkGbVn7NDFhkziELbGTHsm7/euPif10yxe9weCpWkCIpPFnVTSTGm43xoR2jgKAcWGNfC/pryPtOMo02xZOPw/x4/DVeHVd+r+heH5ZPjPJgi2SI7ZI/45IickDNyTmqEkwfyRJ7Ji/PovDpvzvukteDkM5vkl5yPb3VmnXc=</latexit>
✓j

Given: network parameters & topology, load & generation profile,

Phenomenon #1: Transition from incoherence to sync

Function = synchronization Xn
θ̇i = ωi − aij sin(θi − θj )
j=1

✓i (t) ✓i (t)

large |ωi − ωj | & small coupling small |ωi − ωj | & large coupling
⇒ incoherence = no sync ⇒ coherence = frequency sync
 Erie. These transmission lines run through the states of New York,

Phenomenon #2: Multiple power flows

chigan and the Canadian province of Ontario, and traverse the jurisdictions of
ncluding the NYISO, PJM, IESO, and Midwest Independent Transmission System

Theoretical observation: multiple solutions exist

power system are complex and ordinarily occur as the result of a combination of

nergy transactions between the areas controlled by grid operators

tricity supply within each grid operator’s system

Practical observations:
ctricity within each grid operator’s system

utages sometimes undesirable power flows around loops

ages
sometimes sizable difference between predicted and actual power flows
THEMA-Report 2013-36 Loop flows – Final advice
nterconnected transmission systems, as the flow of electricity follows physical
Lake Erie Figure 8: Average unscheduled flows for the years 2011 and 2012, MWh/h8

xpected
loop flows
onal
he
ow and
s affecting
ent to the

ears, Lake
Avg. physical flow
average, [MWh/h]
534

Avg. scheduled flow

ockwise [MWh/h]

Avg. unscheduled flow

289
general [MWh/h]

Source: THEMA Consulting Group, based on data from 16 TSOs

New York Independent System Operator, Lake Erie THEMA Consulting Group, Loop-flows - Final ad-
In this report, we focus on three different cases of loop and transit flows, indicated in Figure 8:
Lake Erie Loop Flow Mitigation | November 2008 | 6
Loop Flow Mitigation, Technical Report, 2008 vice, Technical Report prepared for the European
loop flows in Central Eastern Europe, involving Germany, Poland, Czech Republic and Austria,
transit flows in Central Western Europe, involving Germany, the Netherlands, Belgium, and
Commission, 2013
France, and a loop flow in Central South Europe, where we identify a loop flow from France
through Germany and Switzerland, back to France. These flows are analysed in more detail in the
following section.
The loop and transit flows are identified visually (cf. Figure 8) and by analysing how well the
Outline
1
Linear Network Systems and Metzler Matrices

2
An emerging theory for Nonlinear Network Systems

Kuramoto Synchronization (existence)

3
S. Jafarpour and F. Bullo. Synchronization of Kuramoto oscillators via cutset projections.
IEEE Transactions on Automatic Control, 64(7):2830–2844, 2019.
doi:10.1109/TAC.2018.2876786

1 problem statement
2 solution

Kuramoto Multi-Stability (lack of uniqueness)

S. Jafarpour, E. Y. Huang, K. D. Smith, and F. Bullo. Multistable synchronous power
4 flows: From geometry to analysis and computation.
SIAM Review, January 2019.
Submitted.
URL: https://arxiv.org/pdf/1901.11189.pdf
Primer on algebraic graph theory (slide 1/2)

Weighted undirected graph with n nodes and m edges:

Incidence matrix: n × m matrix B s.t. (B > pactv )(ij) = pi − pj
Weight matrix: m × m diagonal matrix A
Laplacian stiffness: L = BAB > ≥ 0

Linearization of Kuramoto equilibrium equation:

pactv = BA sin(B > θ) =⇒ pactv ≈ BA(B > θ) = L θ

Algebraic connectivity:

λ2 (L) = second smallest eig of L

= notion of connectivity and coupling
Primer on algebraic graph theory (slide 2/2)
Laplacian linear balance equation

+1
<latexit sha1_base64="F0ZijgdLySn1qco4YVCzzLpOlVY=">AAAB8nicdVDLSsNAFJ3UV62vqks3g0UQlDCJoY9dwY3LKsYW2lAm00k7dPJgZlIsoX/gVjeuxK0/JPgxTtoKKnrgwuGce7n3Hj/hTCqE3o3Cyura+kZxs7S1vbO7V94/uJNxKgh1Scxj0fGxpJxF1FVMcdpJBMWhz2nbH1/mfntChWRxdKumCfVCPIxYwAhWWro5s/rlCjIb9artVCEyEapZtpUTu+ZcONDSSo4KWKLVL3/0BjFJQxopwrGUXQslysuwUIxwOiv1UkkTTMZ4SLuaRjik8lxOhnPiZffzm2fwRHsDGMRCV6TgXP0+m+FQymno684Qq5H87eXiX143VUHdy1iUpIpGZLEoSDlUMcwDgAMmKFF8qgkmgumrIRlhgYnSMZV0HF8/w/+Ja5sNE107lSZa5lIER+AYnAIL1EATXIEWcAEBAXgAj+DJSI1n48V4XbQWjOXMIfgB4+0Tx0+RMQ==</latexit> <latexit sha1_base64="xw3lWDRfInNhNtt9rrU1Uk4Gwck=">AAAB8nicdVDLSsNAFJ3UV62vqks3g0VwoWESQx+7ghuXVYwttKFMppN26OTBzKRYQv/ArW5ciVt/SPBjnLQVVPTAhcM593LvPX7CmVQIvRuFldW19Y3iZmlre2d3r7x/cCfjVBDqkpjHouNjSTmLqKuY4rSTCIpDn9O2P77M/faECsni6FZNE+qFeBixgBGstHRzbvXLFWQ26lXbqUJkIlSzbCsnds25cKCllRwVsESrX/7oDWKShjRShGMpuxZKlJdhoRjhdFbqpZImmIzxkHY1jXBI5ZmcDOfEy+7nN8/gifYGMIiFrkjBufp9NsOhlNPQ150hViP528vFv7xuqoK6l7EoSRWNyGJRkHKoYpgHAAdMUKL4VBNMBNNXQzLCAhOlYyrpOL5+hv8T1zYbJrp2Kk20zKUIjsAxOAUWqIEmuAIt4AICAvAAHsGTkRrPxovxumgtGMuZQ/ADxtsnymuRMw==</latexit>
1

<latexit sha1_base64="POg+i3S4EYtqfKmvrgZnjs4JV0w=">AAAB8nicdVDLSsNAFJ34rPVVdelmsAguNEzS0Meu4MZlFWMLbSiT6aQdOnkwMymW0D9wqxtX4tYfEvwYJ20FFT1w4XDOvdx7j59wJhVC78bK6tr6xmZhq7i9s7u3Xzo4vJNxKgh1Scxj0fGxpJxF1FVMcdpJBMWhz2nbH1/mfntChWRxdKumCfVCPIxYwAhWWrq5qPRLZWQ26lXbqUJkIlSzbCsnds2pONDSSo4yWKLVL330BjFJQxopwrGUXQslysuwUIxwOiv2UkkTTMZ4SLuaRjik8lxOhnPiZffzm2fwVHsDGMRCV6TgXP0+m+FQymno684Qq5H87eXiX143VUHdy1iUpIpGZLEoSDlUMcwDgAMmKFF8qgkmgumrIRlhgYnSMRV1HF8/w/+Ja5sNE1075SZa5lIAx+AEnAEL1EATXIEWcAEBAXgAj+DJSI1n48V4XbSuGMuZI/ADxtsnzYWRNQ==</latexit>
3 +1
<latexit sha1_base64="F0ZijgdLySn1qco4YVCzzLpOlVY=">AAAB8nicdVDLSsNAFJ3UV62vqks3g0UQlDCJoY9dwY3LKsYW2lAm00k7dPJgZlIsoX/gVjeuxK0/JPgxTtoKKnrgwuGce7n3Hj/hTCqE3o3Cyura+kZxs7S1vbO7V94/uJNxKgh1Scxj0fGxpJxF1FVMcdpJBMWhz2nbH1/mfntChWRxdKumCfVCPIxYwAhWWro5s/rlCjIb9artVCEyEapZtpUTu+ZcONDSSo4KWKLVL3/0BjFJQxopwrGUXQslysuwUIxwOiv1UkkTTMZ4SLuaRjik8lxOhnPiZffzm2fwRHsDGMRCV6TgXP0+m+FQymno684Qq5H87eXiX143VUHdy1iUpIpGZLEoSDlUMcwDgAMmKFF8qgkmgumrIRlhgYnSMZV0HF8/w/+Ja5sNE107lSZa5lIER+AYnAIL1EATXIEWcAEBAXgAj+DJSI1n48V4XbQWjOXMIfgB4+0Tx0+RMQ==</latexit>
+2
<latexit sha1_base64="sUWlV3UIi9iPazXzOmLB2PUa61w=">AAAB8nicdVDLSsNAFJ3UV62vqks3g0UQlDCJoY9dwY3LKsYW2lAm00k7dPJgZlIsoX/gVjeuxK0/JPgxTtoKKnrgwuGce7n3Hj/hTCqE3o3Cyura+kZxs7S1vbO7V94/uJNxKgh1Scxj0fGxpJxF1FVMcdpJBMWhz2nbH1/mfntChWRxdKumCfVCPIxYwAhWWro5s/vlCjIb9artVCEyEapZtpUTu+ZcONDSSo4KWKLVL3/0BjFJQxopwrGUXQslysuwUIxwOiv1UkkTTMZ4SLuaRjik8lxOhnPiZffzm2fwRHsDGMRCV6TgXP0+m+FQymno684Qq5H87eXiX143VUHdy1iUpIpGZLEoSDlUMcwDgAMmKFF8qgkmgumrIRlhgYnSMZV0HF8/w/+Ja5sNE107lSZa5lIER+AYnAIL1EATXIEWcAEBAXgAj+DJSI1n48V4XbQWjOXMIfgB4+0TyNyRMg==</latexit>

current current
source source

(a) spring network (b) resistive circuit

Lstiffness x = fload and Lconductance v = cinjected

Laplacian linear balance equation: pactv = L θ

X
if pi = 0 in pactv = L θ, then equilibrium exists : θ = L† pactv
i
pairwise displacements : B > θ = B > L† pactv
From Old to New Tests
Question: Given balanced pactv , do angles exist satisfying

pactv = BA sin(B > θ)

Old Tests: Equilibrium angles (neighbors within π/2 arc) exist if

kB > pactv k2 < λ2 (L) for unweighted graphs (Old 2-norm T)

> †
kB L pactv k∞ < 1 for trees, complete (Old ∞-norm T)

New Tests: Equilibrium angles (neighbors within π/2 arc) exist if

kB > L† pactv k2 < 1 for unweighted graphs (New 2-norm T)

kB > L† pactv k∞ < g (kPk∞ ) for all graphs (New ∞-norm T)
where g is monotonically decreasing

g : [1, ∞) → [0, 1]
y (x) + sin(y (x)) y (x) − sin(y (x)) 
x 7→ −x x − 1

2 2

y (x) = arccos
x +1
!'"

"')

"'(
= g(x)
"'%

"'#

"'"
!" !" #" $" %" &"
x=
and where P is a projection matrix

P = B > L† BA = oblique projection onto Im(B > ) parallel to Ker(BA)

1 1
1 2 1 2

2 3 3 2 3 3
m >
R
|{z} = Im(B ) ⊕ Ker(BA)
| {z } | {z }
edge space cutset space weighted cycle space
flow vectors cycle vectors

1 if G unweighted, then P is orthogonal and kPk2 = 1

2 if G acyclic, then P = Im and kPkp = 1
3 if G uniform complete or ring, then kPk∞ = 2(n − 1)/n ≤ 2
New Tests: Equilibrium angles (neighbors within π/2 arc) exist if

kB > L† pactv k2 < 1 for unweighted graphs (New 2-norm T)

> †
kB L pactv k∞ < g (kPk∞ ) for all graphs (New ∞-norm T)

Unifying theorem with a family of tests

Equilibrium angles (neighbors within γ arc) exist if, in some p-norm,

kB > L† pactv kp ≤ γαp (γ) for all graphs (New αp T)

where nonconvex optimization problem:

αp (γ) := min amplification factor of P diag[sinc(x)]

Proof sketch 1/2: Rewriting the equilibrium equation

For what B, A, pactv does there exist θ solution to:

pactv = BA sin(B > θ)

STEP 1: For what flow z and projection P onto cutset/flow space,

does there exist a flow x that solves

P sin(x) = z

⇐⇒ P diag[sinc(x)]x = z
⇐⇒ x = (P diag[sinc(x)])−1 z =: h(x)
Proof sketch 2/2: Amplification factor & Brouwer
STEP 1: look for x solving

x = h(x) = (P diag[sinc(x)])−1 z

IDEA: assume kxkp ≤ γ and ensure kh(x)kp ≤ γ

STEP 2: if one defines min amplification factor

αp (γ) := min min kP diag[sinc(x)]y kp

kxkp ≤γ ky kp =1

then kh(x)kp ≤ max max k(P diag[sinc(x)])−1 y kp · kzkp

x y

−1 kzkp
= min min kP diag[sinc(x)]y kp kzkp ≤
x y αp (γ)

STEP 3: kzkp ≤ γαp (γ), then kh(x)kp ≤ γ so that h satisfies Brouwer

Comparison of sufficient and approximate sync tests

Any test predicts max transmittable power (before bifurcation).

Compare with numerically computed.

ratio of test prediction to numerical computation

Test Case old 2-norm new ∞-norm g (kPk∞ ) ≈ 1 α∞ test
approximate fmincon
IEEE 9 16.54 % 73.74 % 92.13 % 85.06 %†
IEEE 14 8.33 % 59.42 % 83.09 % 81.32 %†
IEEE RTS 24 3.86 % 53.44 % 89.48 % 89.48 %†
IEEE 30 2.70 % 55.70 % 85.54 % 85.54 %†
IEEE 118 0.29 % 43.70 % 85.95 % —*
IEEE 300 0.20 % 40.33 % 99.80 % —*
Polish 2383 0.11 % 29.08 % 82.85 % —*
†
fmincon with 100 randomized initial conditions
*
fmincon does not converge
Summary: Kuramoto equilibrium and active power flow

Given topology (incidence B), admittances (Laplacian L), injections pactv ,

Xn
pi = aij sin(θi − θj )
j=1

Equilibrium angles exist if, in some p-norm,

kB > L† pactv kp ≤ γαp (γ) for all graphs (New αp T)

For p = ∞, after bounding,

kB > L† pactv k∞ ≤ g (kPk∞ ) (New ∞-norm T)

Q1: ∃ a stable operating point (with pairwise angles ≤ γ)?

Q2: what is the network capacity to transmit active power?
Q3: how to quantify robustness as distance from loss of feasibility?
Outline
Introduction to Network Systems
1 F. Bullo. Lectures on Network Systems.
Kindle Direct Publishing, 1.3 edition, July 2019.
With contributions by J. Cortés, F. Dörfler, and S. Martı́nez.
URL: http://motion.me.ucsb.edu/book-lns

Synchronization (existence)
2
S. Jafarpour and F. Bullo. Synchronization of Kuramoto oscillators via cutset projections.
IEEE Transactions on Automatic Control, 64(7):2830–2844, 2019.
doi:10.1109/TAC.2018.2876786

Multi-Stability (lack of uniqueness)

S. Jafarpour, E. Y. Huang, K. D. Smith, and F. Bullo. Multistable synchronous power
3 flows: From geometry to analysis and computation.
SIAM Review, January 2019.
Submitted.
URL: https://arxiv.org/pdf/1901.11189.pdf
 Erie. These transmission lines run through the states of New York,

Phenomenon #2: Multiple power flows

chigan and the Canadian province of Ontario, and traverse the jurisdictions of
ncluding the NYISO, PJM, IESO, and Midwest Independent Transmission System

Theoretical observation: multiple solutions exist

power system are complex and ordinarily occur as the result of a combination of

nergy transactions between the areas controlled by grid operators

tricity supply within each grid operator’s system

Practical observations:
ctricity within each grid operator’s system

utages sometimes undesirable power flows around loops

ages
sometimes sizable difference between predicted and actual power flows
THEMA-Report 2013-36 Loop flows – Final advice
nterconnected transmission systems, as the flow of electricity follows physical
Lake Erie Figure 8: Average unscheduled flows for the years 2011 and 2012, MWh/h8

xpected
loop flows
onal
he
ow and
s affecting
ent to the

ears, Lake
Avg. physical flow
average, [MWh/h]
534

Avg. scheduled flow

ockwise [MWh/h]

Avg. unscheduled flow

289
general [MWh/h]

Source: THEMA Consulting Group, based on data from 16 TSOs

New York Independent System Operator, Lake Erie THEMA Consulting Group, Loop-flows - Final ad-
In this report, we focus on three different cases of loop and transit flows, indicated in Figure 8:
Lake Erie Loop Flow Mitigation | November 2008 | 6
Loop Flow Mitigation, Technical Report, 2008 vice, Technical Report prepared for the European
loop flows in Central Eastern Europe, involving Germany, Poland, Czech Republic and Austria,
transit flows in Central Western Europe, involving Germany, the Netherlands, Belgium, and
Commission, 2013
France, and a loop flow in Central South Europe, where we identify a loop flow from France
through Germany and Switzerland, back to France. These flows are analysed in more detail in the
following section.
The loop and transit flows are identified visually (cf. Figure 8) and by analysing how well the
Lack of uniqueness and winding solutions

Given topology (incidence B), admittances (Laplacian L), injections pactv ,

Xn
pi = aij sin(θi − θj )
j=1

1 is solution unique?
2 how to localize/classify solutions?

triangle graph, homogeneous weights (aij = 1), pactv = 0

phase sync
<latexit sha1_base64="1t54r6HgAZTnmwKVcD14MkZR2YQ=">AAAB+3icbVDLSgMxFL3js9ZX1aWbYBFcSJmpC10W3LisYB/YDiWTZtrQJDMkmeIw9C/c6saduPVjBD/GdDoLbT0QOJxzD/fmBDFn2rjul7O2vrG5tV3aKe/u7R8cVo6O2zpKFKEtEvFIdQOsKWeStgwznHZjRbEIOO0Ek9u535lSpVkkH0waU1/gkWQhI9hY6TEe2yzSqSSDStWtuTnQKvEKUoUCzUHluz+MSCKoNIRjrXueGxs/w8owwums3E80jTGZ4BHtWSqxoPpST0c58bOn/PYZOrfeEIWRsk8alKu/sxkWWqcisJMCm7Fe9ubif14vMeGNnzEZJ4ZKslgUJhyZCM2LQEOmKDE8tQQTxezViIyxwsTYusq2Dm/586ukXa95VzX3vl5tNIpiSnAKZ3ABHlxDA+6gCS0gIOEZXuDVmTlvzrvzsRhdc4rMCfyB8/kDxliVDg==</latexit>
splay state
<latexit sha1_base64="62GmABInUXputbPgpJG06ogo0qY=">AAAB/HicbZDLSgMxFIYzXmu9VV26CRbBhZSZutBlwY3LCvYC06Fk0kwbmkmG5ExxKPUt3OrGnbj1XQQfxnQ6C209EPj4/3M4J3+YCG7Adb+ctfWNza3t0k55d2//4LBydNw2KtWUtagSSndDYpjgkrWAg2DdRDMSh4J1wvHt3O9MmDZcyQfIEhbEZCh5xCkBK/kmESTDBgiwfqXq1ty88Cp4BVRRUc1+5bs3UDSNmQQqiDG+5yYQTIkGTgWblXupYQmhYzJkvkVJYmYuzWSYQzB9zI+f4XPrDXCktH0ScK7+np2S2JgsDm1nTGBklr25+J/npxDdBFMukxSYpItFUSowKDxPAg+4ZhREZoFQze3VmI6IJhRsXmUbh7f8+VVo12veVc29r1cbjSKYEjpFZ+gCeegaNdAdaqIWokihZ/SCXp0n5815dz4WrWtOMXOC/pTz+QOwvZWU</latexit>
Winding number of n angles

Given undirected graph with a cycle σ = (1, . . . , nσ ) and orientation

1 winding number of θ ∈ Tn along σ is:
nσ
1 X
wσ (θ) = dcc (θi , θi+1 )
2π
i=1

w(✓) = 0
<latexit sha1_base64="3pRNU2xuC/og4Qoc0GKvRAlAPdM=">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</latexit>
w(✓) = ±1
<latexit sha1_base64="d7rxExj71ve8ubsN3Y2rqOvEwQI=">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</latexit>

2 given basis σ1 , . . . , σr for cycles, winding vector of θ is

w (θ) = (wσ1 (θ), . . . , wσr (θ))

“Kirckhoff Angle Law” and partition of the n-torus
Theorem: Kirchhoff angle law on Tn

wσ (θ) = 0, ±1, . . . , ±bnσ /2c

=⇒ w (θ) is piecewise constant
=⇒ w (θ) takes value in a finite set

Theorem: Winding partition

For each possible winding vector u, define

WindingCell(u) := {θ ∈ Tn | w (θ) = u}

Then
Tn = ∪u WindingCell(u)
Winding partition of triangle graph

w = −1 w =0 w = +1

each winding cell is connected 1

⇡
⇡/2 ⇡

each winding cell is invariant under rotation

dcc (✓2 , ✓3 )

✓ )
0 0
bijection:
reduced winding cell ←→ open convex polytope ⇡/2

⇡
⇡ ⇡/2 0 ⇡/2 ⇡

dcc (✓1 , ✓2 )
The Kuramoto model and the winding partition

Given topology (incidence B), admittances

(Laplacian L), injections pactv ,
X
θ̇i = pi − aij sin(θi − θj )
j

Theorem: At-most-uniqueness and extensions

1 each WindingCell has at-most-unique equilibrium with ∆θ < π/2

2 equilibrium loop flow increases monotonically wrt winding number
3 existence + uniqueness in WindingCell(u) with ∆θ < π/2 if

kB > L† pactv + Cuk∞ ≤ g (kPk∞ ), or (Static T)

∃ a trajectory inside WindingCell(u) with ∆θ < π/2 (Dynamic T)
Summary and Future Work

averaging compartmental flows mutualism virus spread coupled oscillators social systems

Review
1 a rather comprehensive theory of linear network systems

2 an emergent theory of nonlinear network systems based on

contractivity and monotonicity

3 existence and multistability for Kuramoto

Future research
1 a little bit more on Metzler matrices

2 much work on monotonicity and contractivity

3 applications to other dynamic flow networks

4 outreach/collaboration opportunities for our community with

sociologists, biologists, economists, physicists ...