non-euclidean contraction and its extensions with
TRANSCRIPT
Non-Euclidean Contraction and its Extensionswith Application to Network Systems
Saber Jafarpour
Center for Control, Dynamical Systems & ComputationUniversity of California at Santa Barbara
Georgia Institute of Technology
May 3, 2021
Acknowledgment
Francesco BulloUCSB
Alexander DavydovUCSB
Pedro Cisneros-VelardeUCSB
SJ and P. Cisneros-Velarde and F. Bullo. Weak and Semi-Contraction for Network Systems andDiffusively-Coupled Oscillators. IEEE Transactions on Automatic Control, accepted, Mar. 2021.
SJ and A. Davydov and F. Bullo. Non-Euclidean Contraction Theory for Monotone and PositiveSystems. Submitted, Apr. 2021.
P. Cisneros-Velarde and SJ and F. Bullo. Distributed and time-varying primal-dual dynamics viacontraction analysis. Submitted, May 2020.
SJ (UCSB) Non-Euclidean Contraction Theory May 3, 2021 2 / 39
Motivation: Large-scale Nonlinear Networks
Power grids Brain neural network Transportation network
Nonlinearity:
Multiple equilibria
Transient stability
Cluster synchronization
Large-scale:
Stochastic
Distributed
“... As [power] systems become more heavily loaded,nonlinearities play an increasingly important role in powersystem behavior ... ” [I. Hiskens,1995]
“... in Oahu, Hawaii, at least 800,000 micro-invertersinterconnect photovoltaic panels to the grid... ”[IEEESpectrum, 2015]
SJ (UCSB) Non-Euclidean Contraction Theory May 3, 2021 3 / 39
Presentation outline
non-Euclidean contraction theory
one-sided Lipschitz constant
characterization of contraction wrt non-Euclidean norms
contraction-based small gain theorem
weakly-contracting systems
definition and examples
dichotomy in asymptotic behavior
example: distributed primal-dual
semi-contracting systems
definition and examples
convergence to invariant subspaces
example: diffusively-coupled oscillators
SJ (UCSB) Non-Euclidean Contraction Theory May 3, 2021 4 / 39
Contraction theory: a brief reviewDefinition
Definition: Contracting systems
x = f (t, x) is contracting wrt to ‖ · ‖ with rate c > 0:
‖x(t)− y(t)‖ ≤ e−ct‖x(0)− y(0)‖.
Contracting system: flow is a contracting map.
x0
y0
y(t)
x(t)
circle with radius e�ct
SJ (UCSB) Non-Euclidean Contraction Theory May 3, 2021 5 / 39
Contraction theory: a brief reviewHistorical notes
B. P. Demidovich. Dissipativity of a nonlinear system of differential equations.Uspekhi Matematicheskikh Nauk, 16(3(99)):216, 1961
Application in control theory: W. Lohmiller and J.-J. E. Slotine. On contraction analysisfor non-linear systems.Automatica, 34(6):683–696, 1998
Differential framework: F. Forni and R. Sepulchre. A differential Lyapunov framework forcontraction analysis.IEEE Trans. Autom. Control, 59(3):614–628, 2014
Non-Euclidean contraction: S. Coogan. A contractive approach to separable Lyapunovfunctions for monotone systems.Automatica, 106:349–357, 2019
Review: M. Di Bernardo, D. Fiore, G. Russo, and F. Scafuti. Convergence, consensus andsynchronization of complex networks via contraction theory.In Complex Systems and Networks: Dynamics, Controls and Applications, pages 313–339.Springer, 2016
Review: Z. Aminzare and E. D. Sontag. Contraction methods for nonlinear systems: Abrief introduction and some open problems.In Proc CDC, pages 3835–3847, Dec. 2014
SJ (UCSB) Non-Euclidean Contraction Theory May 3, 2021 6 / 39
Contraction theory: a brief reviewProperties of contracting systems
Highly ordered asymptotic and transient behaviors:
1 initial conditions are forgotten
2 no overshoot in distance between trajectories
3 time-invariant f : unique globally stable equilibrium
4 periodic f : unique globally stable periodic solution
5 robustness properties: input-to-state stability even in the presence ofunmodeled dynamics.
SJ (UCSB) Non-Euclidean Contraction Theory May 3, 2021 7 / 39
Contraction theory: a brief reviewRole of non-Euclidean norms
Why non-Euclidean norms?
1 systematic and efficient stability analysis:
conservation law: 1>n x = c
geometric symmetry: f (x + 1n) = f (x)
2 `2-norm gives conservative estimates:
contraction: lack of symmetry in dynamics: ∂fi∂xj6= ∂fj
∂xi
nonlinearity: frequency instability in power grids, cluster synchronization inBrain neural networks
3 error analysis for large-scale networks:
x ∈ N (0, σIn) =⇒ E(‖x‖22) = nσ2,
E(‖x‖2∞) ∼ 2 ln(n)σ2
SJ (UCSB) Non-Euclidean Contraction Theory May 3, 2021 8 / 39
Characterization of contracting systemsContraction tests
Fundamental Question
How to check if a system is contracting?
wrt `2-norm
1 Differential condition: Linear matrix inequality2 Integral condition: one-sided Lipschitz constant
wrt non-Euclidean norms
1 Differential condition: matrix measure
Differential conditions computationally intensive
Differential conditions challenging for switching systems
SJ (UCSB) Non-Euclidean Contraction Theory May 3, 2021 9 / 39
One-sided Lipschitz constant: scalar vector fields
Let x = f (x), where f : R→ R:
Lipschitz constant ` ∈ R
|f (x)− f (y)| ≤ `|x − y | =⇒ −` ≤ f ′(x) ≤ `
One-sided Lipschitz constant b ∈ R
(x − y)(f (x)− f (y)) ≤ b(x − y)2 =⇒ f ′(x) ≤ b
〈f (x)− f (y), x − y〉 ≤ b|x − y |2
By the Growall-Bellman Lemma: |x(t)− y(t)| ≤ ebt |y(0)− x(0)|.Numerical analysis: E. Hairer, S. P. Nørsett, and G. Wanner. Solving OrdinaryDifferential Equations I. Nonstiff Problems.
Springer, 1993
SJ (UCSB) Non-Euclidean Contraction Theory May 3, 2021 10 / 39
Contraction for `2 norm
For x ∈ Rn and differentiable f :
x = f (t, x)
For P = P> � 0, define ‖x‖2P = x>Px
equivalent properties:
1 osL: (f (t, x)− f (t, y))>P(x − y) ≤ −c‖x − y‖2P , for all x , y , t
2 LMI: PDf (t, x) + Df (t, x)>P � −2cP for all x , t,
3 dIS: D+‖x(t)− y(t)‖P ≤ −c‖x(t)− y(t)‖P , for all soltns x(·), y(·)4 IS: ‖x(t)− y(t)‖P ≤ e−c(t−t0)‖x(t0)− y(t0)‖P , for all soltns x(·), y(·)
If f not differentiable, then osL ⇐⇒ dIS ⇐⇒ IS
Question: How to extend osL and LMI to non-Euclidean norms?
SJ (UCSB) Non-Euclidean Contraction Theory May 3, 2021 11 / 39
Matrix measuresDefinition and properties
The matrix measure of A ∈ Rn×n wrt to ‖ · ‖:
µ‖·‖(A) := limh→0+
‖In + hA‖ − 1
h.
Directional derivative of norm ‖ · ‖ in direction of A,
Logarithmic norm:T. Strom. On logarithmic norms.
SIAM Journal on Numerical Analysis, 12(5):741–753, 1975
spectral property: −‖A‖ = <(λ) ≤ µ‖·‖(A) ≤ ‖A‖, for every λ ∈ spec(A)
<latexit sha1_base64="vOrRKPBiVk8xg8hJIHP4AKYwBzE=">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</latexit>
� 2 spec(A)
<latexit sha1_base64="PVarAe+K0M9BqBqm3WkEpzwluV8=">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</latexit>
µ(A)<latexit sha1_base64="+2YI3P4S02yI693kM1YWgCI1pDs=">AAACmnicdZFLT+MwEMfdwC6QffA6wiHaConDqkpWi+DIYw8ghATaFpDqCtnOpLWwk8ieIKoQ7UfgCh+Nb4Mbinh1R7L084xn/uMZnitpMQwfGt7U9KfPM7Nz/pev377PLywundqsMAI6IlOZOefMgpIpdFCigvPcANNcwRm/3BvFz67AWJmlbRzm0NOsn8pECobO1aE3O/TmYqEZtsLago8QjaFJxnZ8sdj4R+NMFBpSFIpZ243CHHslMyiFgsqnhYWciUvWh67DlGmwP+1Vv4ZeeV33XQVrLhYHSWbcSTGova9zS6atHWruXmqGA/s+NnJOinULTLZ6pUzzAiEVT0JJoQLMgtEQglgaEKiGDpgw0nUdiAEzTKAb1RsVrivfp3/AfdPAkRPaUfmAccCS1rKJxKpsR1VJbRJDwgrl7lTHL2zVM/v0L+D/avBMxRMq8WRiJbey6P2CPsLpr1a00QpPfje3w/HyZskK+UHWSUQ2yTbZJ8ekQwSR5JbckXtv1dv1DrzDp6deY5yzTN6Y134Ezw7SWQ==</latexit>kAk<latexit sha1_base64="KsOMA6ML84GanpF2NvPqQL08GII=">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</latexit>�kAk <latexit sha1_base64="YxGalPK4r7wyX3o0p7Jq9dJoizQ=">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</latexit>�µ(�A)
<latexit sha1_base64="HSNx8IVlHW3cdYTPbQXGy0ebtu0=">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</latexit>
↵(A) = max<(�)
<latexit sha1_base64="V+M2/Tg9DKokEeJ5QAur254UeBk=">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</latexit>
min<(�)
<latexit sha1_base64="TfFGiSI/Owjpnlfprkm3TijV/IU=">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</latexit>
logarithmicine�ciency
SJ (UCSB) Non-Euclidean Contraction Theory May 3, 2021 12 / 39
Weak semi-inner productsDefinition and properties
A weak semi-inner product (WSIP) is J·, ·K : Rn × Rn → R satisfying
1 Jx1 + x2, yK ≤ Jx1, yK + Jx2, yK and x 7→ Jx , yK is continuous,
2 Jαx , yK = Jx , αyK = α Jx , yK for α ≥ 0 and J−x ,−yK = Jx , yK,
3 Jx , xK > 0, for all x 6= 0,
4 | Jx , yK | ≤ Jx , xK1/2 Jy , yK1/2,
5 (compatibility) Jx , xK = ‖x‖2 for all x
Extension of semi-inner product: G. Lumer. Semi-inner-product spaces.
Transactions of the American Mathematical Society, 100:29–43, 1961
Properties:
Matrix measure: µ(A) = sup‖x‖=1
JAx , xK ,
Norm derivative formula: ‖x(t)‖D+‖x(t)‖ = Jx(t), x(t)K .
SJ (UCSB) Non-Euclidean Contraction Theory May 3, 2021 13 / 39
Norm WSIP Matrix measure
‖x‖P =√x>Px Jx , yK2,P1/2 = x>Py
µ2,P1/2 (A) = min{b ∈ R | A>P + PA � 2bP}
= max‖x‖P=1
x>PAx
‖x‖p =(∑
i
|xi |p)1/p
p ∈ ]1,∞[
Jx , yKp = ‖y‖2−pp (y ◦ |y |p−2)>x µp(A) = max
‖x‖p=1(x ◦ |x |p−2)>Ax
‖x‖1 =∑i
|xi | Jx , yK1 = ‖y‖1sign(y)>x
µ1(A) = maxj∈{1,...,n}
(ajj +
∑i 6=j
|aij |)
= sup‖x‖1=1
sign(x)>Ax
‖x‖∞ = maxi|xi | Jx , yK∞ = max
i∈I∞(y)yixi
µ∞(A) = maxi∈{1,...,n}
(aii +
∑j 6=i
|aij |)
= max‖x‖∞=1
maxi∈I∞(x)
xi (Ax)i
Table of norms, WSIPs, and matrix measures for weighted `2, `p for p ∈ (1,∞), `1, and`∞ norms. Note: I∞(x) = {i ∈ {1, . . . , n} | |xi | = ‖x‖∞}.
SJ (UCSB) Non-Euclidean Contraction Theory May 3, 2021 14 / 39
Contraction for arbitrary norm
For x ∈ Rn and differentiable f :
x = f (t, x)
For norm ‖ · ‖ with matrix measure µ(·) and compatible WSIP J·, ·K,equivalent properties:
1 osL: Jf (t, x)− f (t, y), x − yK ≤ −c‖x − y‖2 for all x , y , t ≥ 0,
2 MM: µ(Df (t, x)) ≤ −c , for all x , t ≥ 0,
3 dIS: D+‖x(t)− y(t)‖ ≤ −c‖x(t)− y(t)‖, for soltns x(·), y(·),
4 IS: ‖x(t)− y(t)‖ ≤ e−c(t−t0)‖x(t0)− y(t0)‖, for all soltns x(·), y(·)
SJ (UCSB) Non-Euclidean Contraction Theory May 3, 2021 15 / 39
Measure Demidovich One-sided Lipschitzbound condition condition
µ2,P(Df (x)) ≤ −c PDf (x) + Df (x)>P � −2cP (x − y)>P(f (x)− f (y)
)≤ −c ‖x − y‖2
P
µp(Df (x)) ≤ −c (v ◦ |v |p−2)>Df (x)v ≤ −c‖v‖pp ((x − y) ◦ |x − y |p−2)>(f (x)− f (y)) ≤ −c‖x − y‖pp
µ1(Df (x)) ≤ −c sign(v)>Df (x)v ≤ −c ‖v‖1 sign(x − y)>(f (x)− f (y)) ≤ −c ‖x − y‖1
µ∞(Df (x)) ≤ −c maxi∈I∞(v)
vi (Df (x)v)i ≤ −c ‖v‖2∞ max
i∈I∞(x−y)(xi − yi )(fi (x)− fi (y)) ≤ −c ‖x − y‖2
∞
Table of equivalences between measure bounded Jacobians, differential Demidovich andone-sided Lipschitz conditions. Note: I∞(v) = {i ∈ {1, . . . , n} | |vi | = ‖v‖∞}.
SJ (UCSB) Non-Euclidean Contraction Theory May 3, 2021 16 / 39
Contraction-based small-gain theoremSetting
n interconnected systems xi = fi (xi , x−i ) where xi ∈ RNi
Lyapunov functions Vi : RNi → R≥0:
LfiVi (xi ) =∂Vi
∂xifi (xi , x−i ) ≤ gi (Vi (xi ),V−i (x−i ))
Classical small-gain:
gi (v) = −αi (vi ) +∑j 6=i
γij(vj)
for class K∞ functions αi , γij .
Comparison system
Study properties of
v = g(v), g = (g1, . . . , gn)>
for v ∈ Rn≥0.
SJ (UCSB) Non-Euclidean Contraction Theory May 3, 2021 17 / 39
Contraction-based small-gain theoremMain result
V (x) = (V1(x1), . . . ,Vn(xn))>
J·, ·Kp,R is associated with ‖R(·)‖p, for p ∈ [1,∞], R ∈ Rn×n non-negative.
Theorem: Contraction-based small-gain
If there exists c > 0 such that for all v ≥ w ≥ 0n,
Jg(v)− g(w), v −wKp,R ≤ −c‖v −w‖2p,R ,
Then
‖V (x(t))‖p,R ≤ e−ct‖V (x(0))‖p,R
For R = diag(η) where η ∈ Rn>0
if p ∈ [1,∞), sum-separable Lyapunov function:∑n
i=1 ηpi V
pi (xi )
if p =∞, max-separable Lyapunov function: maxi{ηiVi (xi )}
SJ (UCSB) Non-Euclidean Contraction Theory May 3, 2021 18 / 39
Contraction-based small-gain theoremProof of the main result
Key Lemma
Jx , yKp,R ≤ Jz , yKp,R , for every x ≤ z and y ≥ 0n.
‖V (x(t))‖p,RD+‖V (x(t))‖p,R =rV (x(t)),V (x(t))
z
p,RNorm derivative formula
≤ Jg(V (x(t))),V (x(t))Kp,R Key Lemma
≤ −c‖V (x(t))‖2p,R osL
Unlike classical small-gain theorems we do not need g to be monotone.
SJ (UCSB) Non-Euclidean Contraction Theory May 3, 2021 19 / 39
Presentation outline
non-Euclidean contraction theory
one-sided Lipschitz constant
characterization of contraction wrt non-Euclidean norms
contraction-based small gain theorem
weakly-contracting systems
definition and examples
dichotomy in asymptotic behavior
example: distributed primal-dual
semi-contracting systems
definition and examples
convergence to invariant subspaces
example: diffusively-coupled oscillators
SJ (UCSB) Non-Euclidean Contraction Theory May 3, 2021 20 / 39
Contraction theory for networks
Challenge: many real-world networks are not contracting.
Network flow system x = f (x) preserving commodity x :
constant = 1>n x(t)
=⇒ 0 = 1>n x(t) = 1>n f (x(t))
=⇒ 0n = 1>n Df (x(t))
If additionally f has Metzler Jacobian, then µ1(Df (x)) = 0.
SJ (UCSB) Non-Euclidean Contraction Theory May 3, 2021 21 / 39
Weakly-contracting systemsDefinition and examples
Definition: Weakly-contracting systems
x = f (t, x) with f continuously differentiable in x is weakly-contracting wrt ‖ · ‖:
µ‖·‖(Df (t, x)) ≤ 0
1 Lotka-Volterra population dynamics (Lotka, 1920; Volterra, 1928) (`1-norm)
2 Kuramoto oscillators (Kuramoto, 1975) and coupled swing equations (Bergen and Hill,1981) (`1-norm and `∞-norm)
3 Daganzo’s cell transmission model for traffic networks (Daganzo, 1994), (`1-norm)
4 compartmental systems in biology, medicine, and ecology (Sandberg, 1978; Maeda et al.,1978). (`1-norm)
5 saddle-point dynamics for optimization of weakly-convex functions (Arrow et al., 1958).(`2-norm)
SJ (UCSB) Non-Euclidean Contraction Theory May 3, 2021 22 / 39
Weakly-contracting systemsPart I: Dichotomy in asymptotic behavior
Theorem: Dichotomy for weakly-contracting systems
For a weakly-contracting system x = f (x), either
1 f has no equilibrium and every trajectory is unbounded, or
2 f has at least one equilibrium x∗ and every trajectory is bounded.
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5-0.6
-0.4
-0.2
0
0.2
0.4
0.6
-0.6 -0.4 -0.2 0 0.2 0.4 0.6-0.6
-0.4
-0.2
0
0.2
0.4
0.6
-0.6 -0.4 -0.2 0 0.2 0.4 0.6-0.6
-0.4
-0.2
0
0.2
0.4
0.6
x1 = �x2
x2 = x1
x1 = �2x1 + x2
x2 = �x1
x1 = �x1 + sin(x2) + 3
x2 = x1
SJ (UCSB) Non-Euclidean Contraction Theory May 3, 2021 23 / 39
Weakly-contracting systemsPart II: bounded trajectory case
Theorem
If x = f (x) is weakly-contracting and f has at least one equilibrium x∗ then:
(i) each equilibrium x∗∗ is stable with weak Lyapunov function x 7→ ‖x − x∗∗‖,
(ii) if the norm ‖ · ‖ is a (p,R)-norm, p ∈ {1,∞} and f is piecewise real analytic,then every trajectory converges to the set of equilibria,
(iii) x∗ is locally asy stable =⇒ x∗ is globally asy stable.
Idea of the proof
µ(Df(x)) 0
x⇤ x0xtmp
B(x⇤, ✏)x⇤
is locally asym stable
x(t)
SJ (UCSB) Non-Euclidean Contraction Theory May 3, 2021 24 / 39
Example: Primal-dual algorithmDistributed implementation over networks
Optimization problem
minx∈Rk
f (x) = minx∈Rk
n∑i=1
fi (x)
Distributed implementation
n agents communicate over a undirected weighted graph G ,
agent i have access to function fi and can exchange xi with its neighbors.
minx∈Rk
n∑i=1
fi (xi )
x1 = x2 = . . . = xn
In matrix form by assuming x = (x>1 , . . . , x>n )> ∈ Rnk :
minx∈Rk
n∑i=1
fi (xi )
(L⊗ Ik)x = 0nN
SJ (UCSB) Non-Euclidean Contraction Theory May 3, 2021 25 / 39
Example: Primal-dual algorithmDistributed implementation over networks
If each fi is continuously differentiable in xi :
Lagrangian
L(x , ν) =∑n
i=1 fi (xi ) + ν>(L⊗ Ik)x
Distributed primal-dual algorithm (component form):
xi = −∂L∂xi
= −∇fi (xi )−n∑
j=1
aij(νi − νj),
νi =∂L∂νi
=n∑
j=1
aij(xi − xj)
Distributed primal-dual algorithm (vector form):
x = −∇f (x)− (L⊗ Ik)ν,
ν = (L⊗ Ik)x
SJ (UCSB) Non-Euclidean Contraction Theory May 3, 2021 26 / 39
Example: Primal-dual algorithmStability and rate of convergence
Assume
1 f has a minimum x∗ ∈ Rk ,
2 for each i ∈ {1, . . . , n}, fi is twice differentiable, ∇2fi (x) � 0 for all x , and∇2fi (x
∗) � 0, and
3 the undirected weighted graph G is connected with Laplacian L.
Theorem: Distributed primal-dual dynamics
The distributed primal-dual algorithm
1 is weakly-contracting wrt `2-norm,
2 (x(t), ν(t))→ (1n ⊗ x∗,1n ⊗ ν∗), with ν∗ =∑n
i=1 νi (0),
3 exponential convergence rate is −αess
([−∇2f (x∗) −L⊗ IkL⊗ Ik 0
])where
αess(A) := max{<(λ) | λ ∈ spec(A) \ {0}}.
SJ (UCSB) Non-Euclidean Contraction Theory May 3, 2021 27 / 39
Semi-contracting systemsSemi-norms
Idea: flows converges to each other only in certain directions.
Definition: Semi-norms
|||·||| is a semi-norm if
1 |||cv ||| = |c ||||v |||, for every v ∈ Rn and c ∈ R;
2 |||v + w ||| ≤ |||v |||+ |||w |||, for every v ,w ∈ Rn.
Define the subspace Ker |||·||| = {v ∈ Rn | |||v ||| = 0}.Example: for k < n, R ∈ Rk×n, and norm ‖ · ‖, we get |||x |||R = ‖Rx‖.Example: for a network G with edge set E and incidence matrix B:
|||x |||E := max(i,j)∈E
|xi − xj | = ‖B>x‖∞
For strongly connected graphs Ker |||·|||E = span{1n}
SJ (UCSB) Non-Euclidean Contraction Theory May 3, 2021 28 / 39
Semi-contracting systemsMatrix semi-measures
Definition: Matrix semi-measures
The matrix semi-measure of A ∈ Rn×n wrt |||·|||:
µ|||·|||(A) = limh→0+
|||In + hA||| − 1
h.
Directional derivative of |||·||| in direction of A.
if Ker |||·||| is invariant under A then <(λ) ≤ µ|||·|||(A), for every
λ ∈ specKer |||·|||⊥(A>).
SJ (UCSB) Non-Euclidean Contraction Theory May 3, 2021 29 / 39
Semi-contracting systemsDefinition and examples
Definition: Semi-contracting systems
x = f (t, x) with f continuously differentiable in x is semi-contracting wrt thesemi-norm |||·||| with rate c > 0:
µ|||·|||(Df (t, x)) ≤ −c
1 Kuramoto oscillators (Kuramoto, 1975) and coupled swing equations (Bergen and Hill,1981), (`1-norm)
2 Chua’s diffusively-coupled circuits (Wu and Chua, 1995), (`2-norm)
3 morphogenesis in developmental biology (Turing, 1952), (`1-norm)
4 Goodwin model for oscillating auto-regulated gene (Goodwin, 1965). (`1-norm)
SJ (UCSB) Non-Euclidean Contraction Theory May 3, 2021 30 / 39
Semi-contracting systemsSemi-contraction and asymptotic behavior
Theorem: Semi-contracting systems
Consider x = f (t, x) with f continuously differentiable in x and assume
f is semi-contracting wrt the semi-norm |||·||| with rate c > 0, and
(Affine invariance): f (t, x∗ + Ker |||·|||) ⊆ Ker |||·||| for every t
Then,
1 for every trajectory x(t),
|||x(t)− x∗||| ≤ e−ct |||x(0)− x∗|||, for every t ≥ 0.
2 every trajectory converges to x∗ + Ker |||·|||.
partial contraction and horizontal contraction
SJ (UCSB) Non-Euclidean Contraction Theory May 3, 2021 31 / 39
Example: Diffusively-coupled oscillators
n agents connected by a weighted undirected graph G ,
identical internal dynamics f : R≥0 × Rk → Rk
xi = f (t, xi )−∑n
j=1 aij(xi − xj), i ∈ {1, . . . , n}
synchronization:
limt→∞ ‖xi − xj‖ = 0 for every i , j
synchronization of diffusively-coupled oscillators:
1 contractivity of the internal dynamics2 strength of the diffusive coupling
SJ (UCSB) Non-Euclidean Contraction Theory May 3, 2021 32 / 39
Example: Diffusively-coupled oscillators
Introduce a local-global mixed norm: (2, p)-tensor norm on Rnk = Rn ⊗ Rk
‖u‖(2,p) = inf{( r∑
i=1
‖v i‖22‖w i‖2
p
) 12 ∣∣ u =
r∑i=1
v i ⊗ w i}.
closely related to, but different from, the projective tensor product norm
R. A. Ryan. Introduction to Tensor Products of Banach Spaces.
Springer, 2002
different from the mixed global norm
G. Russo, M. Di Bernardo, and E. D. Sontag. A contraction approach to the hierarchicalanalysis and design of networked systems.
IEEE Trans. Autom. Control, 58(5):1328–1331, 2013
Global norm: `2-norm for the interactions between agents
Local norm: `p-norm for internal dynamics of each agent
SJ (UCSB) Non-Euclidean Contraction Theory May 3, 2021 33 / 39
Example: Diffusively-coupled oscillators
The orthogonal projection P ∈ Rn×n
P = In − 1n1n1>n =
n−1n − 1
n . . . − 1n
− 1n
n−1n . . . − 1
n...
.... . .
...− 1
n − 1n . . . n−1
n
(P ⊗ Ik)x measures dissimilarity of the states xi
x =1n ⊗ x∗ =⇒
(P ⊗ Ik)x = (P ⊗ Ik)(1n ⊗ x∗) = P1n ⊗ x∗ = 0(n−1)×k .
SJ (UCSB) Non-Euclidean Contraction Theory May 3, 2021 34 / 39
Example: Diffusively-coupled oscillators
G is an undirected weighted graph with Laplacian L,
p ∈ [1,∞], Q ∈ Rk×k
xi = f (t, xi )−∑n
j=1 aij(xi − xj), i ∈ {1, . . . , n}
Theorem: diffusively-coupled oscillators are semi-contracting
Suppose that
µp,Q(Df (t, x)) ≤ λ2(L)− c , for every t, x
then
1 the dynamics is semi-contracting wrt ‖ · ‖(2,p),(P⊗Q);
2 for every trajectory x(t),
‖x(t)− 1n ⊗ xave(t)‖(2,p),(P⊗Q) ≤ e−ct‖x(0)− 1n ⊗ xave(0)‖(2,p),(P⊗Q).
3 the system achieves synchronization: limt→∞ x(t) = 1n ⊗ xave(t)
where xave(t) = 1n
∑ni=1 xi (t)
SJ (UCSB) Non-Euclidean Contraction Theory May 3, 2021 35 / 39
Example: Diffusively-coupled oscillators
µp,Q(Df (t, x)) ≤ λ2(L)− c , for every t, x
trade off between internal dynamics and coupling strength
f time-invariant: every trajectory converges to the unique equilibrium point.
f periodic: every trajectory converges to the unique periodic orbit.
Unstable dynamics f , sufficiently strong coupling =⇒ λ2(L) large =⇒ thenetwork synchronizes.
SJ (UCSB) Non-Euclidean Contraction Theory May 3, 2021 36 / 39
Summary
notion of one-sided Lipschitz constant and weak semi-inner product
characterization of contraction wrt non-Euclidean norms
contraction-based small-gain theorem
two extensions of classical contraction:
weak contractionsemi-contraction
properties of weakly-contracting and semi-contracting systems
SJ (UCSB) Non-Euclidean Contraction Theory May 3, 2021 37 / 39
Future research
contraction-based compositional analysis of interconnected systems
scalable stability certificates using non-Euclidean contraction.
computing equilibra of contracting and weakly-contracting systems
explicit and implicit integration algorithms
accelerated convergence.
optimization algorithms using contraction theory
extension to gradient descent algorithms and time-varying algorithms.
connection with discrete-time algorithms for optimization.
implicit deep learning using contraction theory
use contraction condition for well-posedness in the optimization problem.
SJ (UCSB) Non-Euclidean Contraction Theory May 3, 2021 38 / 39
Contraction-based small gain theoremA simple example
Consider the following system on R2:
x1 = −x1 + 3x42 x1 − x3
1
x2 = −x2 − 3x41 x2 − x5
2
and pick Vi (xi ) = x2i for i ∈ {1, 2}. Then
V1 = −2V1 + 6V1V22 − 2V 2
1 ≤ −2V1 + 6V1V22 := g1(V1,V2)
V2 = −2V2 − 6V 21 V2 − 2V 3
2 ≤ −2V2 − 6V 21 V2 := g2(V1,V2)
but g = (g1, g2)> is not monotone. However, for every v ≥ w ≥ 0n,
Jg(v)− g(w), v −wK2
=[v1 − w1 v2 − w2
] [−2v1 + 2w1 + 6v1v22 − 6w1w
22
−2v2 + 2w2 − 6v21 v2 + 6w2
1w2
]≤ −2|v1 − w1|2 − 2|v2 − w2|2 = −2‖v −w‖2
2
SJ (UCSB) Non-Euclidean Contraction Theory May 3, 2021 39 / 39