The top Lyapunov exponent of switched linear systems...

69
Contents Linear time-varying systems and growth rates Constrained switching The top Lyapunov exponent of switched linear systems with dwell times Fabian Wirth Institute of Mathematics University of W¨ urzburg The Dynamics of Control Ocotber 1–3, 2010.

Transcript of The top Lyapunov exponent of switched linear systems...

Page 1: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

The top Lyapunov exponentof switched linear systems with dwell times

Fabian Wirth

Institute of MathematicsUniversity of Wurzburg

The Dynamics of ControlOcotber 1–3, 2010.

Page 2: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

Linear time-varying systems and growth ratesGrowth ratesSwitched linear systemsIrreducibilityA converse Lyapunov theorem

Constrained switchingDwell timesA converse Lyapunov theoremAverage Dwell Time

Page 3: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

Growth rates

Families of Linear Time-Varying Systems

Consider a family of time-varying linear systems

x = A(u(t))x , t ≥ 0

x(0) = x0 ∈ Rn,

u :R+ → U measurable

The evolution operator is denoted by Φu(t, s), t ≥ s ≥ 0.

Page 4: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

Growth rates

Exponential growth rates

x = A(u(t))x , u : R+ → U measurable

The Lyapunov exponent of a solution is

λ(x0, u) = lim supt→∞

1

tlog ‖Φu(t, 0)x0‖,

Lyapunov exponents for periodic u are called Floquet exponents.

The upper Bohl exponent corresponding to u is

β(u) = lim supt,s→∞

1

tlog ‖Φu(t + s, s)‖

The collections of all Lyapunov, Floquet and Bohl exponents are

ΣLy(A,U), ΣFl(A,U), ΣBohl(A,U).

Page 5: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

Growth rates

Exponential growth rates

x = A(u(t))x , u : R+ → U measurable

The Lyapunov exponent of a solution is

λ(x0, u) = lim supt→∞

1

tlog ‖Φu(t, 0)x0‖,

Lyapunov exponents for periodic u are called Floquet exponents.

The upper Bohl exponent corresponding to u is

β(u) = lim supt,s→∞

1

tlog ‖Φu(t + s, s)‖

The collections of all Lyapunov, Floquet and Bohl exponents are

ΣLy(A,U), ΣFl(A,U), ΣBohl(A,U).

Page 6: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

Growth rates

Exponential growth rates

x = A(u(t))x , u : R+ → U measurable

The Lyapunov exponent of a solution is

λ(x0, u) = lim supt→∞

1

tlog ‖Φu(t, 0)x0‖,

Lyapunov exponents for periodic u are called Floquet exponents.

The upper Bohl exponent corresponding to u is

β(u) = lim supt,s→∞

1

tlog ‖Φu(t + s, s)‖

The collections of all Lyapunov, Floquet and Bohl exponents are

ΣLy(A,U), ΣFl(A,U), ΣBohl(A,U).

Page 7: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

Growth rates

Read this

F. Colonius, W. Kliemann. Infinite-time optimal control andperiodicity. Appl. Math. Optim. 1989.F. Colonius, W. Kliemann. Stability radii and Lyapunov exponents.Proc. Workshop Bremen. 1990.F. Colonius, W. Kliemann. Linear control semigroups acting onprojecitve space. J. Dyn. Diff. Equations. 1993.F. Colonius, W. Kliemann. Minimal and maximal Lyapunovexponents of nonlinear control systems. J. Diff. Equations. 1993.F. Colonius, W. Kliemann. The Lyapunov spectrum of families oftime-varying matrices. Trans. Amer. Math. Society. 1996.F. Colonius, W. Kliemann. The Morse spectrum of linear flows onvector bundles. Trans. Amer. Math. Society. 1996.

Page 8: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

Growth rates

The Colonius-Kliemann intuition

Considerx = A1x andx = A2xin R2.

Page 9: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

Growth rates

The Colonius-Kliemann intuition

Considerx = A1x andx = A2xin R2.

Page 10: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

Growth rates

The Colonius-Kliemann intuition

Project thesystems onto theunit sphereor ratherprojective space.

Page 11: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

Growth rates

The Colonius-Kliemann intuition

Project thesystems onto theunit sphereor ratherprojective space.

Page 12: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

Growth rates

The Colonius-Kliemann intuition

Project thesystems onto theunit sphereor ratherprojective space.

Page 13: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

Growth rates

The Colonius-Kliemann intuition

Project thesystems onto theunit sphereor ratherprojective space.

Page 14: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

Growth rates

The Colonius-Kliemann intuition

Project thesystems onto theunit sphereor ratherprojective space.

DC

Page 15: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

Growth rates

The Colonius-Kliemann spectrum of families oftime-varying matrices

Generically, there are a finite number of control sets. In particular,one invariant control set.To each control set we can associate a set of FLoquet exponentsand Lyapunov exponents.In particular we have the Gelfand formula

sup ΣFL = max ΣLy = max ΣBohl .

Assumptions: cl intU = cl U, local accessibility of the projectedsystem, ...

Page 16: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

Switched linear systems

We consider a finite set of matrices M = {A1, . . . ,Am} ⊂ Rn×n

and the associated switched linear system

x(t) = Aσ(t)x(t) (1)

whereσ : R→M

is the switching signal.

Page 17: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

Switched linear systems

Exponential growth rates

M = {A1, . . . ,Am} ⊂ Rn×n x(t) = Aσ(t)x(t) (1)

Options for defining a uniform exponential growth rate of (1):

I trajectory-wise (Lyapunov exponents):

κ := maxσ∈S,x0

{lim supt→∞

1

tlog ‖ϕ(t, x0, σ)‖

},

I using norms of evolution operators (Bohl exponents):

ρ := limt→∞

1

tlog max

σ∈S‖Φσ(t, 0)‖ ,

It is known by Fenichel’s uniformity lemma thatκ = ρ .

Page 18: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

Switched linear systems

Exponential growth rates

M = {A1, . . . ,Am} ⊂ Rn×n x(t) = Aσ(t)x(t) (1)

Options for defining a uniform exponential growth rate of (1):I trajectory-wise (Lyapunov exponents):

κ := maxσ∈S,x0

{lim supt→∞

1

tlog ‖ϕ(t, x0, σ)‖

},

I using norms of evolution operators (Bohl exponents):

ρ := limt→∞

1

tlog max

σ∈S‖Φσ(t, 0)‖ ,

It is known by Fenichel’s uniformity lemma thatκ = ρ .

Page 19: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

Switched linear systems

Exponential growth rates

M = {A1, . . . ,Am} ⊂ Rn×n x(t) = Aσ(t)x(t) (1)

Options for defining a uniform exponential growth rate of (1):I trajectory-wise (Lyapunov exponents):

κ := maxσ∈S,x0

{lim supt→∞

1

tlog ‖ϕ(t, x0, σ)‖

},

I using norms of evolution operators (Bohl exponents):

ρ := limt→∞

1

tlog max

σ∈S‖Φσ(t, 0)‖ ,

It is known by Fenichel’s uniformity lemma thatκ = ρ .

Page 20: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

Switched linear systems

Exponential growth rates

M = {A1, . . . ,Am} ⊂ Rn×n x(t) = Aσ(t)x(t) (1)

Options for defining a uniform exponential growth rate of (1):I trajectory-wise (Lyapunov exponents):

κ := maxσ∈S,x0

{lim supt→∞

1

tlog ‖ϕ(t, x0, σ)‖

},

I using norms of evolution operators (Bohl exponents):

ρ := limt→∞

1

tlog max

σ∈S‖Φσ(t, 0)‖ ,

It is known by Fenichel’s uniformity lemma thatκ = ρ .

Page 21: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

Switched linear systems

Arbitrary switching

We are considering

M = {A1, . . . ,Am} ⊂ Rn×n x(t) = Aσ(t)x(t) (1)

where the switching signal is any measurable function.An equivalent formulation considers the linear inclusion

x(t) ∈ {Ax(t) | A ∈M} .

Denote the exponential growth rate of this inclusion by ρ.

Page 22: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

Switched linear systems

Arbitrary switching

We are considering

M = {A1, . . . ,Am} ⊂ Rn×n x(t) = Aσ(t)x(t) (1)

where the switching signal is any measurable function.An equivalent formulation considers the linear inclusion

x(t) ∈ {Ax(t) | A ∈M} .

Denote the exponential growth rate of this inclusion by ρ.

Page 23: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

Irreducibility

IrreducibilityM is called irreducible, if only the trivial subspaces {0} and Kn areinvariant under all A ∈M and otherwise reducible.

Modulo a similarity transformation for reducible M all A ∈M areof the form

A11 A12 . . . . . . A1d

0 A22 A23 . . . A2d

0 0 A33...

.... . .

. . ....

0 . . . 0 Add

,

whereMii := {Aii ; A ∈M}

is irreducible or Aii = 0 for all A ∈M.

Page 24: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

Irreducibility

IrreducibilityM is called irreducible, if only the trivial subspaces {0} and Kn areinvariant under all A ∈M and otherwise reducible.Modulo a similarity transformation for reducible M all A ∈M areof the form

A11 A12 . . . . . . A1d

0 A22 A23 . . . A2d

0 0 A33...

.... . .

. . ....

0 . . . 0 Add

,

whereMii := {Aii ; A ∈M}

is irreducible or Aii = 0 for all A ∈M.

Page 25: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

A converse Lyapunov theorem

A converse Lyapunov theorem

Theorem (Barabanov 1988)If M is irreducible there exists a norm v on Kn such that for allx ∈ Kn, t ≥ 0:

∀σ : v(Φσ(t, 0)x) ≤ eρtv(x)

∃σ : v(Φσ(t, 0)x) = eρtv(x)

The proof relies in many ways on the fact that the set of evolutionoperators

{Φσ(t, s) | σ measurable , t ≥ s ≥ 0 }

is a semigroup.

Page 26: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

A converse Lyapunov theorem

A converse Lyapunov theorem

Theorem (Barabanov 1988)If M is irreducible there exists a norm v on Kn such that for allx ∈ Kn, t ≥ 0:

∀σ : v(Φσ(t, 0)x) ≤ eρtv(x)

∃σ : v(Φσ(t, 0)x) = eρtv(x)

The proof relies in many ways on the fact that the set of evolutionoperators

{Φσ(t, s) | σ measurable , t ≥ s ≥ 0 }

is a semigroup.

Page 27: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

A converse Lyapunov theorem

Three proofs of the Gelfand formula

I Berger, Wang, Linear Algebra and its Applications, 1992 -algebraic

I Elsner, Linear Algebra and its Applications, 1995 - usingBarabanov norms

I Shi, Wu, Pang, Linear Algebra and its Applications, 1997 -using Barabanov norms

note: continua of matrices not needed, acessibility not needed.

Page 28: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

Constrained Switching

So far we have dealt with unrestricted switching.There are many suggestions in the literature for restrictedswitching with some type of dwell-time condition

Page 29: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

Dwell TimeWe consider a finite set of matrices M = {A1, . . . ,Am} ⊂ Rn×n

and the associated switched linear system

x(t) = Aσ(t)x(t) (2)

whereσ : R→M

is the switching signal.

Now switching signals satisfying a dwell-time condition areconsidered, i.e. for h > 0 the set of admissible switching signals is

Sdwell(h) := {σ | σ is piecewise constant and

its discontinuities are at least h apart.}

Page 30: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

Dwell TimeWe consider a finite set of matrices M = {A1, . . . ,Am} ⊂ Rn×n

and the associated switched linear system

x(t) = Aσ(t)x(t) (2)

whereσ : R→M

is the switching signal.Now switching signals satisfying a dwell-time condition areconsidered, i.e. for h > 0 the set of admissible switching signals is

Sdwell(h) := {σ | σ is piecewise constant and

its discontinuities are at least h apart.}

Page 31: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

M = {A1, . . . ,Am} ⊂ Rn×n

x(t) = Aσ(t)x(t)

Switching signals look like this

≥ h

Page 32: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

Average Dwell Time

For t < T let Nσ(T , t) denote the number of discontinuities of σin [t,T ].Switching signals are said to have an average dwell time h if for allt < T it holds that

Nσ(T , t) ≤ N0 +T − t

h.

Sav (h,N0) := {σ : R→M | σ is piecewise constant and

satisfies the average dwell time condition.}

Page 33: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

Average Dwell Time

For t < T let Nσ(T , t) denote the number of discontinuities of σin [t,T ].Switching signals are said to have an average dwell time h if for allt < T it holds that

Nσ(T , t) ≤ N0 +T − t

h.

Here h is called the average dwell time andN0 is the chatter bound.

Page 34: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

Average Dwell Time - Switching Signals

M = {A1, . . . ,Am} ⊂ Rn×n, S = Sav (h,N0)

Page 35: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

A remark on topologyPropositionThe sets Sdwell(h) and Sav (h,N0) are compact in the weak∗

topology of L∞, resp. `∞.In particular, the systems

x(t) = Aσ(t)x(t)

together with the shift on Sdwell(h) resp. Sav (h,N0) are linearflows with compact base space.We therefore have ρ = eκ for these flows by Fenichel’s uniformitylemma.Conley’s theorem says that if 0 is uniformly exponentially stable,there exist a Lyapunov function for the set {0} × S.

Page 36: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

Construction of Lyapunov functions

x(t) = Aσ(t)x(t) , σ ∈ Sdwell(h),Sav (N0, h).

Remark 1: It is not reasonable to look for a single Lyapunovfunction V for this type of system:If one Lyapunov function exists such that

∇V (x)Ax < −α(‖x‖) , ∀A ∈M

and a positive definite function α, then the system is exponentiallystable with unrestricted switching. But we are interested in thecase that the switching is restricted.Remark 2: By imposing dwell-time conditions the stabilityproperties of a switched system can change.

Page 37: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

Long Dwell Times

PropositionLet M be a compact set of Hurwitz stable matrices, then thereexists a T > 0 such that for all h ≥ T the switched system

x(t) = Aσ(t)x(t) , σ ∈ Sdwell(h)

is uniformly exponentially stable.(easy consequence of results on slowly varying systems: Cesari1967, Desoer 1969)

Page 38: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

Converse Lyapunov Theorems with Dwell-Time

There is one fundamental problem in the construction of Lyapunovfunctions for systems with dwell time: The concatenation ofswitching signals.

Page 39: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

Concatenation

Page 40: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

Concatenation

Page 41: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

Concatenation

Page 42: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

Dwell times

Consider a finite set of matrices M = {A1, . . . ,Am} ⊂ Rn×n andthe associated switched linear system

x(t) = Aσ(t)x(t)

whereσ : R→M

is the switching signal.

We consider switching signals satisfying a dwell-time condition, i.e.for h > 0 the set of admissible switching signals is

S(h) := {σ | σ is piecewise constant and

its discontinuities are at least h apart.}

Page 43: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

Dwell times

Consider a finite set of matrices M = {A1, . . . ,Am} ⊂ Rn×n andthe associated switched linear system

x(t) = Aσ(t)x(t)

whereσ : R→M

is the switching signal.We consider switching signals satisfying a dwell-time condition, i.e.for h > 0 the set of admissible switching signals is

S(h) := {σ | σ is piecewise constant and

its discontinuities are at least h apart.}

Page 44: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

Dwell times

Reprise: The converse Lyapunov theorem for unrestrictedswitching

TheoremIf M is irreducible there exists a norm v on Rn such that for allx ∈ Rn, t ≥ 0:

∀σ : v(Φσ(t, 0)x) ≤ eρtv(x)

∃σ : v(Φσ(t, 0)x) = eρtv(x)

The proof relies in many ways on the fact that the set of evolutionoperators

{Φσ(t, s) | σ measurable , t ≥ s ≥ 0 }

is a semigroup.

This is not the case if a dwell time condition holds.

Page 45: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

Dwell times

Reprise: The converse Lyapunov theorem for unrestrictedswitching

TheoremIf M is irreducible there exists a norm v on Rn such that for allx ∈ Rn, t ≥ 0:

∀σ : v(Φσ(t, 0)x) ≤ eρtv(x)

∃σ : v(Φσ(t, 0)x) = eρtv(x)

The proof relies in many ways on the fact that the set of evolutionoperators

{Φσ(t, s) | σ measurable , t ≥ s ≥ 0 }

is a semigroup. This is not the case if a dwell time condition holds.

Page 46: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

A converse Lyapunov theorem

A Converse Lyapunov Theorem - Dwell TimesTheoremLet M = {A1, . . . ,Am} ⊂ Rn×n be irreducible and consider a dwelltime h > 0. The following two statements are equivalent

(i) ρ(M, h) = ρ,

(ii) there are norms v1, . . . , vm on Rn with the followingproperties:

vi (eAi tx) ≤ eρtvi (x) for all t ≥ 0, x ∈ Rn, i = 1, . . . ,m,

vj(eAj tx) ≤ eρtvi (x) for all t ≥ h, x ∈ Rn, i , j = 1, . . . ,m.

and for all x0 ∈ Rn there exists a σ such that

vσ(ti )(φ(t, x0, σ)) = eρtvσ(0)(x0) , ∀t ∈ [ti + h, ti+1], i .

Page 47: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

A converse Lyapunov theorem

A Converse Lyapunov Theorem - Dwell TimesTheoremLet M = {A1, . . . ,Am} ⊂ Rn×n be irreducible and consider a dwelltime h > 0. The following two statements are equivalent

(i) ρ(M, h) = ρ,

(ii) there are norms v1, . . . , vm on Rn with the followingproperties:

vi (eAi tx) ≤ eρtvi (x) for all t ≥ 0, x ∈ Rn, i = 1, . . . ,m,

vj(eAj tx) ≤ eρtvi (x) for all t ≥ h, x ∈ Rn, i , j = 1, . . . ,m.

and for all x0 ∈ Rn there exists a σ such that

vσ(ti )(φ(t, x0, σ)) = eρtvσ(0)(x0) , ∀t ∈ [ti + h, ti+1], i .

Page 48: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

A converse Lyapunov theorem

A Converse Lyapunov Theorem - Dwell TimesTheoremLet M = {A1, . . . ,Am} ⊂ Rn×n be irreducible and consider a dwelltime h > 0. The following two statements are equivalent

(i) ρ(M, h) = ρ,

(ii) there are norms v1, . . . , vm on Rn with the followingproperties:

vi (eAi tx) ≤ eρtvi (x) for all t ≥ 0, x ∈ Rn, i = 1, . . . ,m,

vj(eAj tx) ≤ eρtvi (x) for all t ≥ h, x ∈ Rn, i , j = 1, . . . ,m.

and for all x0 ∈ Rn there exists a σ such that

vσ(ti )(φ(t, x0, σ)) = eρtvσ(0)(x0) , ∀t ∈ [ti + h, ti+1], i .

Page 49: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

A converse Lyapunov theorem

A Converse Lyapunov Theorem - Dwell TimesTheoremLet M = {A1, . . . ,Am} ⊂ Rn×n be irreducible and consider a dwelltime h > 0. The following two statements are equivalent

(i) ρ(M, h) = ρ,

(ii) there are norms v1, . . . , vm on Rn with the followingproperties:

vi (eAi tx) ≤ eρtvi (x) for all t ≥ 0, x ∈ Rn, i = 1, . . . ,m,

vj(eAj tx) ≤ eρtvi (x) for all t ≥ h, x ∈ Rn, i , j = 1, . . . ,m.

and for all x0 ∈ Rn there exists a σ such that

vσ(ti )(φ(t, x0, σ)) = eρtvσ(0)(x0) , ∀t ∈ [ti + h, ti+1], i .

Page 50: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

A converse Lyapunov theorem

Outline of proof

Concatenation:

(u �t w)(s) :=

{u(s) , s < tw(s − t) , t ≤ s

.

For each i define the set of switching signals that can beconcatenated at time t to a switching signal that has been equalto i on (t − h, t).

S(i) := {σ ∈ Sh | σ(0) = i or t0(σ) ≥ h} .

Tt(i) := {Φσ(t, 0) | σ ∈ S (i)}.

Note Ti := ∪t≥0Tt(i) is not a semigroup.

Page 51: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

A converse Lyapunov theorem

Outline of proof

Concatenation:

(u �t w)(s) :=

{u(s) , s < tw(s − t) , t ≤ s

.

For each i define the set of switching signals that can beconcatenated at time t to a switching signal that has been equalto i on (t − h, t).

S(i) := {σ ∈ Sh | σ(0) = i or t0(σ) ≥ h} .

Tt(i) := {Φσ(t, 0) | σ ∈ S (i)}.

Note Ti := ∪t≥0Tt(i) is not a semigroup.

Page 52: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

A converse Lyapunov theorem

Outline of proof

Concatenation:

(u �t w)(s) :=

{u(s) , s < tw(s − t) , t ≤ s

.

For each i define the set of switching signals that can beconcatenated at time t to a switching signal that has been equalto i on (t − h, t).

S(i) := {σ ∈ Sh | σ(0) = i or t0(σ) ≥ h} .

Tt(i) := {Φσ(t, 0) | σ ∈ S (i)}.

Note Ti := ∪t≥0Tt(i) is not a semigroup.

Page 53: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

A converse Lyapunov theorem

Outline of proof II

Tt(i) := {Φσ(t, 0) | σ ∈ S (i)}.If M is irreducible, then

Vi :=⋃t≥0

ρ−tTt(i)

is bounded and irreducible.

Thenwi (x) = sup{‖Φx‖ | Φ ∈ Vi}

defines Lyapunov functions with the property

wi (eAi tx) ≤ eρtwi (x) for all t ≥ 0, x ∈ Rn, i = 1, . . . ,m,

wj(eAj tx) ≤ eρtwi (x) for all t ≥ h, x ∈ Rn, i , j = 1, . . . ,m.

Page 54: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

A converse Lyapunov theorem

Outline of proof II

Tt(i) := {Φσ(t, 0) | σ ∈ S (i)}.If M is irreducible, then

Vi :=⋃t≥0

ρ−tTt(i)

is bounded and irreducible.Then

wi (x) = sup{‖Φx‖ | Φ ∈ Vi}defines Lyapunov functions with the property

wi (eAi tx) ≤ eρtwi (x) for all t ≥ 0, x ∈ Rn, i = 1, . . . ,m,

wj(eAj tx) ≤ eρtwi (x) for all t ≥ h, x ∈ Rn, i , j = 1, . . . ,m.

Page 55: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

A converse Lyapunov theorem

Outline of proof III

This is not enough!!

Tt(i) := {Φσ(t, 0) | σ ∈ S (i)}.

If M is irreducible, then

T∞(i) := lim supt→∞

e−ρtTt(i)

is compact and irreducible.The norms

vi (x) = max{‖Φx‖ | Φ ∈ T∞(i)}

define the norms for which all assertions hold.

Page 56: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

A converse Lyapunov theorem

Outline of proof III

This is not enough!!

Tt(i) := {Φσ(t, 0) | σ ∈ S (i)}.

If M is irreducible, then

T∞(i) := lim supt→∞

e−ρtTt(i)

is compact and irreducible.The norms

vi (x) = max{‖Φx‖ | Φ ∈ T∞(i)}

define the norms for which all assertions hold.

Page 57: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

A converse Lyapunov theorem

A Common Quadratic Lyapunov VersionConsider a finite set of matrices M = {A1, . . . ,Am} ⊂ Rn×n andthe associated switched linear system

x(t) = Aσ(t)x(t) , σ ∈ Sdwell(h) .

Theorem(Colaneri, Geromel, 2005)If there are positive definite matrices Pi > 0, i = 1, . . . ,m suchthat

(i) ATi Pi + PiAi < 0, i = 1, . . . ,m,

(ii)

eATj hPje

Ajh < Pi , i , j = 1, . . . ,m .

then the switched system is exponentially stable.Remark For Hurwitz matrices, the condition of the theorem canalways be satisfied if h is large. So again large enough dwell timesensure stability.

Page 58: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

A converse Lyapunov theorem

Consequences of the Converse Lyapunov Theorem

With the existence of the Lyapunov functions for systems withdwell time, the following results may be proved.

(i) ρ(M) = ρ(cl convM) is no longer true.

(ii) The maximal exponential growth rate may be approximatedby periodic switching signals – Gelfand formula

(iii) ρ is jointly continuous in M and h.

(iv) ρ is locally Lipschitz continuous in M and h, for Mirreducible.

Page 59: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

Average dwell time

Average Dwell Time

For t < T let Nσ(T , t) denote the number of discontinuities of σin [t,T ].Switching signals are said to have an average dwell time h if for allt < T it holds that

Nσ(T , t) ≤ N0 +T − t

h.

Sav (h,N0) := {σ : R→M | σ is piecewise constant and

satisfies the average dwell time condition.}

Page 60: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

Average dwell time

Average Dwell Time

For t < T let Nσ(T , t) denote the number of discontinuities of σin [t,T ].Switching signals are said to have an average dwell time h if for allt < T it holds that

Nσ(T , t) ≤ N0 +T − t

h.

Here h is called the average dwell time andN0 is the chatter bound.

Page 61: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

Average dwell time

Average Dwell Time - Switching Signals

M = {A1, . . . ,Am} ⊂ Rn×n, S = Sav (h,N0)

Page 62: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

Average dwell time

Is this really more general ?

A system stable with dwell time h may be destabilized byincreasing the chatter bound:Consider switching between two Hurwitz matrices. There areexamples where such systems are just unstable ρ = 0 and thedestabilizing switching is periodic with two switches.

hh

Page 63: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

Average dwell time

Is this really more general ?

In this situation the system is stable with respect to the dwell timeh, but not stable with respect to the set Sav(h, 2).

hh

Page 64: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

Average dwell time

Gelfand formula

A converse Lyapunov theorem for the class Sav(h,N0) is notknown.

But:The Gelfand formula

sup ΣFl = ρ

may still be shown, using new techniques from ergodic theory.I’s still to be dotted and t’s to be crossed.

Page 65: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

Average dwell time

Gelfand formula

A converse Lyapunov theorem for the class Sav(h,N0) is notknown.But:

The Gelfand formulasup ΣFl = ρ

may still be shown, using new techniques from ergodic theory.I’s still to be dotted and t’s to be crossed.

Page 66: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

Average dwell time

Gelfand formula

A converse Lyapunov theorem for the class Sav(h,N0) is notknown.But:The Gelfand formula

sup ΣFl = ρ

may still be shown, using new techniques from ergodic theory.

I’s still to be dotted and t’s to be crossed.

Page 67: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

Average dwell time

Gelfand formula

A converse Lyapunov theorem for the class Sav(h,N0) is notknown.But:The Gelfand formula

sup ΣFl = ρ

may still be shown, using new techniques from ergodic theory.I’s still to be dotted and t’s to be crossed.

Page 68: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

Average dwell time

Conclusions

I In stability theory of linear time-varying systems puzzles stillabound – and new ones are invented all the time.

I Fritz and Didi have put me on the tracks of a subject thatcontinues to fascinate me.

I At a certain point stop followong your teachers as far astechniques are concerned.

Page 69: The top Lyapunov exponent of switched linear systems …num.math.uni-bayreuth.de/en/conferences/irsee_2010/home/download/... · Contents Linear time-varying systems and growth rates

Contents Linear time-varying systems and growth rates Constrained switching

Average dwell time