Jump to ContentJump to Main Navigation
Foundation of Statistical Energy Analysis in Vibroacoustics$

A. Le Bot

Print publication date: 2015

Print ISBN-13: 9780198729235

Published to Oxford Scholarship Online: August 2015

DOI: 10.1093/acprof:oso/9780198729235.001.0001

Show Summary Details
Page of

PRINTED FROM OXFORD SCHOLARSHIP ONLINE (www.oxfordscholarship.com). (c) Copyright Oxford University Press, 2017. All Rights Reserved. Under the terms of the licence agreement, an individual user may print out a PDF of a single chapter of a monograph in OSO for personal use (for details see http://www.oxfordscholarship.com/page/privacy-policy). Subscriber: null; date: 25 February 2017

(p.281) Appendix A Stability

(p.281) Appendix A Stability

Source:
Foundation of Statistical Energy Analysis in Vibroacoustics
Publisher:
Oxford University Press

The study of system stability may be found in textbooks on differential equations (Gourmelin and Wadi, 2009) or dynamical systems (Strogatz 1994, Jean, 2011). For general aspects on differential calculus see (Cartan 1985).

Differential equation

Let us consider a differential equation,

(A.1)
Y˙=f(Y)

where f:URNRN is a continuously differentiable function defined on an open set U. A solution Y:IU is a function differentiable on an interval IR which verifies eqn (A.1) on I. A maximal solution is a solution defined on the largest possible interval. The Cauchy–Lipschitz theorem states the existence and uniqueness of a maximal solution for any initial condition Y(0)=Y0U. A point YU is an equilibrium if f(Y)=0.

The differential equation is said to be linear if f is a matrix A,

(A.2)
Y˙=AY

For all linear systems, the origin is an equilibrium (but not necessarily the only one). For any Y0RN, the unique solution of eqn (A.2) verifying Y(0)=Y0 is

(A.3)
Y(t)=exptAY0

and is defined on I=R. Let us recall that the exponential of matrix A is defined by the series n=0An/n!.

Example A.1

In Chapter 2, we introduced the following N-degrees-of-freedom system,

(A.4)
MX¨+(C+G)X˙+KX=0

where M, K, C, and G are respectively the mass, stiffness, damping, and gyroscopic matrices. They are all real-valued and M, K are symmetric and positive-definite, i.e. MT=M, KT=K, M>0, and K>0. The gyroscopic matrix is antisymmetric GT=G and C=diag(c1,,cN) is diagonal with non-negative entries ci0 hence CT=C. If all ci>0 then C is positive-definite but only semipositive when at least one ci=0.

The matrices M and K are invertible since they are diagonalizable (real-valued and symmetric) and all their eigenvalues are non-zero.

By introducing the state vector,

(A.5)
Y=XX˙

which has values in R2N, the governing equation (A.4) takes the form (A.2) with

(A.6)
A=(0IM1KM1(C+G))

All properties of system (A.4) are embodied in the matrix A.

(p.282) Stability

The stability of an equilibrium is defined in the sense of Lyapunov by the following.

Definition A.1

The equilibrium Y is stable if for any ϵ>0 there exists δ>0 such that

(A.7)
ifY(0)Y<δthent0Y(t)Y<ϵ

Stability means that any solution starting in a neighbourhood of equilibrium remains indefinitely in a neighbourhood of equilibrium (Fig. A.1). A stronger definition is the asymptotic stability.

Definition A.2

The equilibrium Y is said to be attracting if there exists δ>0 such that

(A.8)
ifY(0)Y<δthenlimtY(t)=Y

An equilibrium Y is asymptotically stable if it is both stable and attracting.

Appendix A Stability

Figure A.1 Example of (a) stable equilibrium; (b) asymptotically stable equilibrium.

Asymptotic stability means that any solution starting in a neighbourhood of equilibrium tends to equilibrium by remaining arbitrarily close to it for all time (Fig. A.1). Asymptotic stability trivially implies stability but the converse is not true. An equilibrium may be attracting but not stable. An example of such a pathological case is given by the differential equation θ˙=1cosθ and the equilibrium θ=0. For all initial conditions θ00, the solution is θ(t)=2arctan1/1/tan(θ0/2)t. The point 0 is attracting since limθ(t)=0 when t for any θ0. But 0 is not stable since limθ(t)=π when t1/arctan(θ0/2). The solution cannot be confined to the neighbourhood of 0.

(p.283) In the linear case, stability depends on the position of eigenvalues of A in the complex plane. The following theorem is established by discussing the solution (A.3).

Theorem A.1

Let us consider a matrix A and the linear differential equation (A.2). The origin is an equilibrium. We note λi, i=1,,N the complex eigenvalues of A and Ei the related eigenspaces. Then,

  • the origin is stable if and only if Re(λi)0 for all i and when Re(λi)=0 the dimension of Ei is equal to the multiplicity of λi in the characteristic polynomial; and

  • the origin is asymptotically stable if and only if Re(λi)<0 for all i.

Although this criterion only applies to linear systems, it may be also useful for non-linear systems by linearizing the differential equation about equilibrium. But the result is weakened. The following theorem is known as the first Lyapunov method.

Theorem A.2

Let Df be the Jacobian matrix of a continuously differentiable function f:URN at equilibrium Y and let λi, i=1,,N be its complex eigenvalues. Then

  • if Re(λi)<0 for all i then Y is asymptotically stable; and

  • if Re(λi)>0 for at least one i then Y is not stable.

When Re(λi)=0 for one or more eigenvalues, one cannot conclude. For instance, the differential equation y˙=y3 has Df=0 at y=0. Theorem 2 does not apply. However, the origin is asymptotically stable since the solution is y(t)=sgn(y0)/2t+1/y02 for t0 and goes to zero when t. Conversely, the differential equation θ˙=1cosθ has also Df=0 at θ=0 but we have previously seen that 0 is not stable.

Example A.2

Let us discuss the stability of eqn (A.4) with respect to Y=0. We have seen that A is given by eqn (A.6). Stability is therefore related to the eigenvalues of A, that is the roots of the characteristic polynomial,

detAλI=detλIdetM1(C+G)λIdetM1KdetI=(detM)1detλ2M+λ(C+G)+K

Thus, the system stability depends on the position of zeros of the polynomial,

(A.9)
Δ(λ)=detλ2M+λ(C+G)+K

with respect to the axis Re(z)=0 in the complex plane.

(p.284) Lyapunov functions

To study stability in the general case, we introduce the concept of Lyapunov functions.

Definition A.3

A continuously differentiable function V defined in an open set U such that YURn is said to be a Lyapunov function if

  1. 1. V(Y)=0

  2. 2. V(Y)>0 for all YY

  3. 3. V˙=f.V(Y)0

If, furthermore, V˙(Y)<0 for all YY then V is a strict Lyapunov function.

Appendix A Stability

Figure A.2 Lyapunov function. Velocity along trajectories is oriented towards decreasing values of V.

In the above definition, V is the gradient of V defined in U. This is a vector oriented in the direction of increasing values of V. The velocity Y˙=f(Y) is a vector tangential to the trajectory at point Y. The scalar product f.V is therefore the time derivative of V along trajectories. The third condition imposes that trajectories are oriented toward lower levels of V (Fig. A.2).

We may now enunciate Lyapunov’s second theorem on stability.

Theorem A.3

Let us consider the differential equation (A.1) and an equilibrium point Y. If there exists a Lyapunov function in a neighbourhood of Y then Y is stable. If there exists a strict Lyapunov function in a neighbourhood of Y then Y is asymptotically stable.

A common application of Lyapunov functions is concerning the energy. The theorem then states that if the energy is non-increasing (no energy is injected), the equilibrium is stable. Furthermore, if the system dissipates at all times, the energy is a strict Lyapunov function and the equilibrium is asymptotically stable.

Example A.3

To establish the stability of system (A.4), we consider the Lyapunov function V(Y)=12×YTPY where

(A.10)
P=K00M

Since M and K are symmetric and positive-definite, P is also symmetric and positive-definite. The first two conditions of a Lyapunov function are therefore fulfilled. By developing the product YTPY, we obtain

(A.11)
V(Y)=12XTKX+12X˙TMX˙

so that V(Y) is simply the energy of system.

To check the third condition, we calculate V˙(Y):

(A.12)
V˙(Y)=12Y˙TPY+12YTPY˙=12YTATP+PAY

By denoting Q=12×(ATP+PA), a mere calculation gives

(A.13)
Q=000C

and V˙(Y)=X˙TCX˙. This equality is nothing other than the energy balance. Since C is semipositive, we obtain V˙(Y)0 and the third condition is also fulfilled. By applying the Lyapunov theorem, we conclude that system (A.4) is stable about the origin, that is all zeros of the polynomial Δ(λ) are in the half-plane Re(z)0. But, unfortunately, Q is not positive definite and V is not a strict Lyapunov function. Therefore, we cannot conclude asymptotic stability. To do that, we need a stronger result.

(p.285) Invariance principle

We now introduce an important extension of Lyapunov’s second method. Standard references are (Krasovskii 1963), and (LaSalle 1960).

Definition A.4

Let us consider a differential equation Y˙=f(Y) with an equilibrium YU. A set M is termed invariant if any solution starting from M is entirely in M, i.e. Y(0)MY(t)M for all t, and termed positively invariant if Y(0)MY(t)M for all t0.

Theorem A.4

Let V be a continuously differentiable function on U such that V˙(Y)0. Let S be the set of all points in U such that V˙(Y)=0. If M is the largest invariant set in S then every solution Y(t) bounded for t>0 approaches M as t, i.e.\ limtd(Y(t),M)=0.

(p.286) Let us choose V as the energy of a dissipating system. If no energy is provided to the system, the function V is non-increasing and any trajectory tends to a state which does not dissipate.

Theorem A.5

Let V be a Lyapunov function on U and Ω‎ a compact neighbourhood of Y. If Ω‎ is positively invariant and if the only solution Y(t) contained in the set S=YU,V˙(Y)=0 is the trivial solution Y(t)=Y, then Y is asymptotically stable. Furthermore all points in Ω‎ tend to Y.

What is interesting about this theorem is that asymptotic stability is obtained with a non-strict Lyapunov function.

Finally, let us mention this last result.

Theorem A.6

Let V be a Lyapunov function on RN such that limYV(Y)=. If the only solution Y(t) contained in the set S=YRN,V˙(Y)=0 is the trivial solution Y(t)=Y, then Y is globally asymptotically stable.

Globally asymptotic stability means that all solutions tend to Y, not only those starting in a neighbourhood of Y.

Example A.4

Let us consider the special case of system (A.4) when C is positive-definite (all resonators are damped). The condition V˙(Y)=X˙TCX˙=0 implies that X˙=0. Then, the set S reduces to S=Y=(X,0). A trajectory contained in S must verify X˙0 at any time. Therefore X¨0 and by substitution into eqn (A.4), we get X0 and finally Y0. Applying the invariance principle, we conclude that system (A.4) is asymptotically stable. In particular, the zeros of Δ‎ are all in the half-plane Re(z)<0.

Example A.5

Let us examine an example with a semipositive damping matrix C. The system is shown in Fig. A.3. The set of governing equations is

(A.14)
{m1X¨1+c1X˙1+(k1+K)X1KX2=0m2X¨2+(k2+K)X2KX1=0

The mass, stiffness, and damping matrices are

(A.15)
M=(m100m2) K=(k1+KKKk2+K) C=(c1000)

Matrix C is only semipositive because resonator 2 is undamped. The system energy is

(A.16)
V=12m1X˙12+12m2X˙22+12(k1+K)X12KX1X2+12(k2+K)X22

is positive-definite. The time derivative of the energy is

(A.17)
V˙=c1X˙12
Appendix A Stability

Figure A.3 Two coupled resonators, one of which is undamped.

and is seminegative. Hence S=Y=(X1,X2,X˙1,X˙2)R4/X˙1=0. If a solution X1(t),X2(t) stays in S, then X˙10 and therefore X¨10. The first governing equation gives (k1+K)X1KX2=0. Furthermore, X1 being constant in time, X2 is also constant and therefore X¨2=0. The second governing equation gives KX1+(k2+K)X2=0. In a matrix form this reads KX=0 and since K is invertible, X=0. Applying the invariance principle, we conclude that the system is asymptotically stable and therefore all zeros of Δ(λ) are in the half-plane Re(z)<0.

A direct method to check this last result would be to expand Δ‎,

(A.18)
Δ(λ)=m1λ2+c1λ+(k1+K)m2λ2+(k2+K)K2

and to calculate the four complex roots. But this way is evidently more tedious.

(p.287) In most cases, the stability of a system reduces to localizing the roots of the characteristic polynomial Δ(λ) of A. A real polynomial Δ(λ)R[X] is said to be Hurwitz stable if all its roots are located in the left half-plane Re(z)<0. A first simple property of Hurwitz stable polynomials is the following.

Theorem A.7

Let Δ(λ)R[X]. If Δ(λ) is Hurwitz stable then all its coefficients have the same sign.

This property is easy to prove by remarking that since Δ(λ)R[X] its complex roots appear in pairs λj, λ¯j. The factorized form of Δ(λ),

αi(λλi)j(λ22Re(λj)+λjλ¯j)

(p.288) where α‎ is the leading coefficient of Δ(λ), shows a product of polynomials whose coefficients are all positive. Developing can only give non-negative coefficients.

The converse is generally false. For instance, the polynomial λ3+λ2+λ+6=(λ+2)(λ2λ+3) has positive coefficients but two complex roots in the half-plane Re(z)>0.

The Routh–Hurwitz algorithm provides a criterion to recognize a Hurwitz stable polynomial.

Theorem A.8

Given anλn+an1λn1++a0 where ai are real coefficients, we construct the Routh array,

anan2an4an1an3an5ln2,1ln2,2ln2,3l0,1

completed by zeros on the right. The third and subsequent rows are calculated by the recursive rule,

lk,i=lk+1,ilk+2,i+1lk+2,1lk+1,i+1/lk+1,1

If all coefficients in the left column are positive (respectively negative), then Δ(λ) is Hurwitz stable.

We can now revisit the last example.

Example A.6

The characteristic polynomial (A.18), Δ(λ)=m1m2λ4+m2c1λ3+m1(k2+K)+m2(k1+K)λ2+c1(k2+K)λ+k1k2+K(k1+k2) leads to the Routh array,

m1m2m1(k2+K)+m2(k1+K)k1k2+K(k1+k2)0m2c1c1(k2+K)00m2(k1+K)k1k2+K(k1+k2)00c1K2/(k1+K)000k1k2+K(k1+k2)000 

All coefficients of the left column are positive. The polynomial is Routh stable and the system is asymptotically stable.