Dynamical Systems,
and Linear Algebra
Differential Equations,
Dynamical Systems,
and Linear Algebra
This I. d volumc in
PURE AND APPLIED MATHEMATICS ACADEMIC PRESS. INC.
H a r u r t B- Jovamvich, Publidvrs
A Scrlcc of Monographs and Textbooks San Diego New York Boston
London Sydney Tokyo Toronto
Editors: SAMUEL
EILENBERG BASS
AND HYMAN
A cotnplcrc Ir\l of titles in this series is ava~lablefrom the Publishers upon request.
Contents
340
Preface
REFERENCES
Subject Index
This book is about dynamical aspects of ordinary differential equations and the
relations between dynamical systems and certain fields outside pure mathematics.
A prominent role is played by the structure theory of linear operators on finite-
d;mensional vector spaces; we have included a seli+ontained t r m t m m t of that
subject.
The background material needed to understand this book is differential c~lculus
of several variables. For example, Serge Lang's Coldus o j SNcml VaricrMea, up to
the chapter on integration, contains more than is needed to understand much of our
text. On the other hand, after Chapter 7 we do we several resulta from elementary
analysis such as theorem on uniform convergence; these are stated but not proved.
This mathematics is contained in Lang's Anolysid I, for instance. Our treatment of
linear algebra is systematic and self<ontainrd, although the most elementary parta
have the character of a review; in any case, h g ' a Coldua oj Sncrd Variables
develops this elementary linear algebra a t a leisurely pace.
While this book can be wed as early ap the sophomore year by ntudents with a
strong first year of calculus, i t is oriented mainly toward upper diviaion mathematiea
and science students. I t can also be ueed for a graduate course, especially if t h e l a t e
chapters are emphasized.
I t has been said that the subject of ordinary differential equations is a collection
of tricks and hints for finding solutions, and that i t is important because i t cao
solve problem in physics, engineering, ete. O w view is that the subject can be
developed with considerable unity and coherence; we have attempted such a de-
velopment with this book. The importance of ordinary differentid equations
vtd d uid other areas of science lim in its power to motivate, unify, .ndgive l o r n to
those arena. Our four chapters on "applications" have been writtea to do ewretly
this, and not merely to provide examples. Moreover, an underatanding of the ways
that differential equations relates to other subjects is a prinlary m m of ineight
and inspiration for the student and working mathematician dike.
Our goal in this book is to develop nonlinear ordinary differentid equations in
open subsets of real Cartesian space, R', in such a a a y that the extension to
manifolds is simple and natural. We treat chiefly autonomous emphasiring
qualitative behavior of solution curves. The related themes of d i t y and physieal
significance pervade much of the material. Many topics have been omitted, such as
Lsplsee transforms, series solutions, Sturm theory, and special iunctions.
The level of rigor is high, and almost everything is proved. More important.
however, is that ad hoc methcds have been rejected. We have tried to develop
S PREFACE PREFACE xi
I , ~ O I I ~ tSI I I L ~;idd
, insight to the theorem and t,hat are important methods in their l'lrc. followi~lgrcmarks s h ~ ~ u hvlp
l d lht, rl,nd<.rd~,cid<. U I I \\-hich chapt,,rs 10 rttad
I,, arbitrary constant coefficient linear differential equations. Examples arc included.
In particular, the S + N decomposition is used to compute the exponential of a n
arbitrary square matrix.
Chapter 2 is independent from the others and jncludea an elementary account
of the Keplerian planetary orbits.
The fundamental theorem on existence, uniquenm, and continuity of solutions
of ordinary differential equations are developed in Chapters 8 and 16. Chapter 8 is
rrst rirtvd t r ~t hc autonomous CMP, in line with our basic orientation toward dynami-
,.,!
(,nl - t r,ll\h
( 'll:ll~tc.r\10, 12, and 14 are devoted tosystematic introductions to mathematical
m~rlt1. c t f ~ ~ l ~ ~ ~ circuits,
. l r i c a l population theory, and classical mechanics, respectively.
'1'111,I\r:~ytor~ ?LIosrrcircuit theory is presented a s a special casr of the more general
t I~wjrl-rr.curt ly developed on manifolds. The Volterra-Lotkaequationsofcompeting
spr~c.i~.,q tirr: :~nalyzcd, along with some generalizations. In mechanics we develop The book owerr much to many people. Wc only mention four of them her?. lkuko
thc 1ltrrnilli)nian f o r d i s m for conseniative systems whose configuration space is Workman and Ruth Suzuki did an cxccllcnt job of typing the manuscript. Dick
arl I I ( I I . I I s11b*rtOf a vector space. Palais made a number of uscful comments. Special thanks arc due to Jacob Palis,
T11,. rcw~:rinin~five chapters contain a subatantid introductioh to the phase who rc*ad tllu manuscript thoruughly, found nrany minor c m m , and suggcatrd
portr:~itannlysis of nonlinear autonomous system. They include a dkcuasion of wvcralsubstantial improvcmcnts. I'r~~f(~ssor llirsclr is grutrful to the hlillcr lnatitute
"gc.nvric" ~,n~pertics of linear flows, 1.iapunov and structural stability, PoincarC for its support during part of thc wriling of tlrc b,w,k.
U~ndixsu~l tllenry, periodic attractors, and perturbations. Wa conclude with an
Aftc.r\~ord which points thc way toward manifolds.
Chapter 1
First Examples
The purpose of this ahort chapter is to develop some simple examplea of diUerea-
tial equations. Thk development motivate. the linear algebra treated sutmequently
and moreover gives in an elementary context nome of the bssic idem of ordinsry
differential equations. Later these ideas will be put into a more ay-tic expixi-
tion. In particular, the -plea themselves are special eases of the d&Urn-
tial equations eonsidered in Chapter 3. We regard this chapter ma important &ea
some of the most basic ideas of differential equations are 8een in simple form
is the simplest differential equation. I t is also one of the most important. First,
what does it mean? Here z = + ( t ) is an unknown real-valued function of a red
variable t and &Id( is its derivative (we will also use z' or x'(t) for thisderivative).
The equation tells us that for every value of t the equality
0 .o a =0 a a(0
FIG. A This is a very h p l e system; however, many more-comphcated systems of two
equations ean be reduced to this form as we shall see a little later.
hIoreover, there are no olher solulions. To see this, let u ( t ) be any mlution and Since there is no relation specified between the two unknown functions z , ( t ) ,
compute the derivative of u(t)eW': x 1 ( t ) ,they are "uncoupled"; we can immediately write down all aolutiona (aa for
(1)):
) K , e x p ( a l f ) , KI = constant,
z ~ ( t=
z i ( t ) = K1 exp(&t), Kt = constant.
Here K , and K S are determined if initial conditions z , ( b ) = u l , - ( 4 ) = u, are
Thrrcfur,, rc(1 lc.*' is a constant K, so u(1) = KC'. This proves our w r t i o n
' r h rc,rlst;rnt K appeurirlg in the solution is completely determined if the vhkttttel i specified. (We sometimes write exp a for e.)
IPt US consider equation ( 2 ) from a more geometric point of view. We ewsider
I I O of t l i ~
holul ion at a singlr, point t, is specified. Suppose that a function z ( t ) satisfy-
two functions z,( f ) ,z i ( t ) as specifying an unknown cum z ( t ) = ( x , ( t ) ,a ( 1 ) ) in
i r l ~I 1 I i- rr~~uired such that x(t,) = uo, then K muat satisfy KC'@= %. Thus
the ( z , ,z i ) plane Rz.That ia to say, z is a m p from the real numbers R into R1,z :
~ I I U ~ L ~ I II I1I I L IIBS a u n i q u ~
solution satisfying a specified initial condilion z ( 4 ) = %.
R -+ R2. The right-hand side of (3) expresses the tangent uecfurzl(t) = ( z ; ( t ) ,z ; ( t ) )
1:or >i~~il~lir,ity, we often take L = 0; then K = %. There is no loss of generality
to the curve. Using vector notation,
in tt~killglo = 0, for if u ( t ) is a solution with u ( 0 ) = m, then the function v ( t ) =
u ( l - 1,) is n solution with u ( 4 ) = %. (3') z' = Az,
I t i:. cotnrlrun to restate (1) in the form of an initial v a l u e problem: where A z denotes the vector ( a l z l ,a b ) , which one ahould t h i d of as being b d
a t 2.
A solution z ( 1 ) to ( 2 ) must not only satisfy the first condition ( I ) , but must slso
takr on the prescribed initial value K a t t = 0.We have proved that the initial
valuc prublvm ( 2 ) has a unique solution.
The constant a in the equation z' = az can be considered as a parameter. If a
ch:r~~pq,c. t11v (quation changrs and so do the solutions. Can we describe qualita-
livi,l) 111{.!r :i? tlrc- solut ion3 cllangc?
'1'111.srq~tof a is crueirrl here:
if a > 0 , lim,,,K Z Lequals m when K > 0, and equals - m when K < 0;
if a = 0, K e l = constant;
if a < 0,lim,,, KC' = 0.
lr~itlulcv~lditionsare of the form z(b) = u where u = (y, y ) is 6 p i h q ~ & ~ t
of R2 (;~ornetrically, thii means t h t when t = b the curve is required to p W
through t hr gi Ten point u.
FIG. D. Some aolution cum- tor' - Az, A -C -:I
Tbt. lrurp (that is, function) A : W -P (or r -r Az) Can be considered a vcclor
&ld on RZ.Thii m e w that to eech point z in the plane we aasign the vector Az. solutions to (3) require a thedimensional picture wbieh the nsdw is invited to
For purposes of viaualirstion, we picture A r as a vector "based at I"; that is, we sketch!
+
w i g n to z the directed line w e n t from r to z Ar. For example, if q = 2, Let us consider equation (3) as a dynamieol aydem. This meana that the i m b
s = -3, and r = (1, l ) , then a t (1, 1) we picture an arrow pointing from (1, 1) pendent variable t ia interpreted as time and the solution curve r ( f ) could be thought
+
to (1. 1) (2, -$) = (3, f ) (Fig.B). Thus if Ar = (221, - f n ) , we attach to of, for example, as the path of a particle moving in the plane R '
. We ean hmg&
+
each point z in the plane an arrow with tail a t z and head at r Ax and obtain a particle placed at any point u = (u,, w) in R' at time t = 0. As time p d
the picture ill Fig. C. the particle moves along the solution curve z(t) that satisfies the initial condition
Sulvirlg thr differential equation (3) or (3') with initial conditions ( u ~y , ) at z(0) = u. At any later time t > 0 the particle will be in another +tion z(t). And
1 = 0 niralls finding in the plane a curve z(t) that satisfies (3') and passes through a t an earlier time t < 0, the particle was a t a position z(t). T o i n d i ~ t ethe d*
the poi I I I~I = (UI, w) when t = 0. A few solution curves are aketched in Fig. D. pendence of the position on t and u we denote it by d,(u). Thua
The trivial solution ( z ~ ( t )zi(t))
, = (0, 0) is also considered a "curve."
The family of all solution curves as subeels OF R1is called the "phase portrait"
of equation (3) (or (3')). We can imagine particles plseed at each point of the plane and all movir& simul-
The oncdimensional equation z' = at can also be interpreted geometrically: the taneously (for example, dust particles under a steady wind). The aolution curvm
phase portrait is as in Fi. E, which should be compared with Fig. A. It is clearer are spoken of as trajectories or orbits in this context. For each 6xed t in q we have
to picture the graph8 of (1) and the solution curves for (3) since two-djmcnsional a transformation assigning to each point u in the plane another point +,(u). This
pictures are better than either one- or threedimensional pictures. The g~aphsof transformation denoted by 6,:R' -+ R' is clearly a linear transFormation, that m,
81. THE BIMPLEST EXAMPLE8
a ,O o C 0
FIG. E
+
+ , ! u + i.) = + , ( , I ) +,(I,) and @,(Xu) = A+,(u), for all vectors u, Y, and all
real nurnbprs h.
As time proceeds, every point of the plane moves simdtmeously along the tra-
jectory passing through it. In this way the collection of maps 4,: R' + R', 1 E R, i
a one-parameter family of transformations. This family is called the pow or dynami-
cal8yskm or W determined by the vector field t + A+, which in turn is equivalent
to the systenr (3).
Th? d).namical system on the real line R companding to equation (1) is par-
ticul;rrl,v ~ ~ i \ st \ ~- dvscribr:
a > 0.:!I1 ~ l , ~ i l vxcrpt
)
rt\
-
if a < 0, all points move toward 0 as t,ime goes to ; if
0 n~ovcaway from 0 toward f m ;if a = 0, all points stand
still. i.
I\'(, I,;i\.r sta~rtrdfrom a diRerential equation and have obtained the dynamical
s.v>tprn + , >'Iris procrss is established through the fundamentat theorem of ordinary
difT(-rr,~ltial c.rluations as we shall see in Chapter 8.
Lntcr \vc shall also reverse this process: starting from a dyrprnical system +,, a
differential equation nil1 be obtained (simply by differentiatink +,(u)with respect
to I).
I t is seldoni that differential equations are given in the simple uncoupled form T o find y;, y; differentiate the equations defining yl, yz to obtain
( 3 ) . Consider, for example, the system: I
y; = 2r;+ r;,
y; = r; + r;.
By substitution
or in vrctor notation Y; = + + (-621 - 42.) = 411 + !h,
2 ( 5 ~ 1 3%)
y: = (5x1 + 3zr) + (42,- 43%) = -a. -I,
At this point we are not trying to solve (1) ; rather, we want to place it in a geo- differentiable, then the map z in d e d differentiable; its derivative is dehed to be
metrical and algebraic setting in order to understand better what a solution means.
At the most primitive level, a solution of (1) is a set of n differentiable real- -
dl = zf(l) = (z:(t), . . , , z:(t)).
dz
valued f~lnrtionsz.(l) that make (1) true.
I n or(lrr t n , r~.ncha more conceptual understanding of (1) we introduce reol ndi- Thus the derivative, as a function of 1, is again a map from R to R-.
ntr~tslo,,rri('nrlrsian space R".This is simply the set of all n-tuples of real numbers. The derivative can also be expressed in the form
An CIOIII(~III"1 R" is a "point" z = (11, ...
, I") ;the number z. is the ith cwrdin.de
of the , ~ ~ i L. ~ i I'oints
t x, y in Rmare added coordinatewise:
1
+
zr(t) = lim- ( r ( t h) - ~ ( 0 ) .
, h
.t + !, = (Z,, . . . , I") + (y,, , . , y.,)
, = (21 + y,, . . . , + ym).
Z.
I t ha5 a natural geometric interpretation a5 the vector o(f) based a t z(t), which is
-
Also, if X is n rcnl numbcr wc define t h prducl
~ of A and z to he a trsnalate of st(;). Thia vector is called the tongent usdor to the m e U t (or a t
~(1)).
If we imagine 1 aa denoting time, then the leagth I d ( f ) I of the t a r y p t vwbr m
interpreted physically as the Bpeed of a particle deacrihing I(#).
Thr distance between points z, y in R is defined to he To write (1) in an abbreviated form we d the doubly indexed mt of numbaa
lz -yl = [(rl - y1)' + ... + (2" - y.)']"'. a,, an n X n m a t e A, denoted thw:
The length of z is
IzI= (21' + . .. + z.l)l'l.
A vc.ctor based al z t R'. is an ordered pair of points z, y in Rn,denoted by 3.
We thi~lkof this as an arrow or line segment directed from r to y. We 8ay ?$ ia
baser1 at I Next, for each z € R. we define a vector Az € R' whoee ith coordinate is
A vector based a t the origin
note that thin ia the ith row in the righbhand uide of (1). In this way the matrix A
in identified with the point z € R". in interpreted as a map
To a vector Z b a s e d a t z is associated the vector y - z based a t the origin 0. A:Rm+R=
We call the vectors r t a n d y - z lramlal+s of each other. which to r auaigns Az.
From now on a vector based a t 0 in called simply a vector. Thus an element of With thie notation (1) is rewritten
R' can be considered either as an n-tuple of real numbers or as an arrow issuing (2) z' = Az.
from t k . origin.
I t is only for purposes of viaualisation that we consider vectors based a t points Thus the ayetern (1) can be conaidered M a uingle "vector differential e q \ u h "
other than 0. 1:or computations, all vectors are based a t 0 since such vectors can (2). (The word cquution is classically reserved for the case of just one rvkbb;re
be addrd and multiplied by real numbers. ahd call (2) both a syetem m d an equation)
We r ~ t u r nto the system of differential equations (1). A candidate for a solution Wethinkof t h e m s p A : R ' - t R a s a vcGlor)feidonR=:to each points € Rg
ia a cum in R": ~t a.ss~gns the vector based at r which is a translate of Az. Then a solution of (2)
is a curve z: R -t R- w h m tangent vector a t any given t is the veetorAs(t) (tmzm
. Fig.D of Seetion 1.
lated to ~ ( 1 ) )See
In Chapters 3 m d 4 we nhdl give methods of explicitly solviag (2), or equk-
By this we mean a map lently (1). In nubaequent chspters it will be ahown that in fact (2) IIM a uniqua
solution z(t) satisfying any given initial eonditioo z(0) = UI E II.. This is tbs
Such a map is dexribed in terms of coordinates by (*). If each function z.(t) ia S d o n 1 this was proved for the apechl case n -
fundamental theorem 01linear differentid equations with mnstmt m&chU; in
1.
-
(b) Let A [' -11. Find solutions u(l), v(1) to z' -
A t such that every
+
solution can be expressed in the form m ( l ) Bo(1) for nuitable con-
stanta a,8.
1. For each of the following matrices A sketeh the vector field z + Az in R'.
(\Iising matrix entries are 0.)
The background needed for a reader of Chapter 1 is a good first year of college
calculus. One good source is 8. Lang'e Stermd Coursc in C a h l w [12, Chapters I,
11, and 1x1.In thin refemme the material on derivative, mwee, .nd vecbm in
R. and matrice ia dieeuesed much more thoroughly than in our Section 2.
of z' -
(b) What is the geometric relation between solution curves of z' a Az and
-A27
- -
8. (a) Let u(i), v(i) be solutions to z' Az. Show that the curve ~ ( t )
rru(1) f ~ ( 1 is) a solution for all real numbers e,8.
$1. HARMONIC OSCILLATORS 15
We shall go into details of this field in Section 6. Other important examples of form
fields are derived from electrical forces, magnetic forces, and so on.
Ch,apter 2 The connection between the physical concept of force field and the mathematid
concept of differential equation is Neufun's second lalt.: F = mu. This h w e t e
that a particle in a force field moves in such a nay that the force vector a t the loen-
tion of the particle, a t any instant, equals the acceleration vector of tbe particle
Newton's Equation and Kepler's Luw times the maps m. If z(1) denotes the position vector of the particle a t time 1, where
z : R + Rais a sufficiently differentiable curve, then the acceleration v w r is t h
second derivative of z ( t ) uith respect to timr
(We follow tradition and use dots for time derivatives in t h i chapter.) Newton's
second law states
F(z(1)) = ?nZ(l).
Thus we obtain a eecond order differentialequation:
We develop in this chapter the earliest important examples of differential equa-
tions, which in fact are connected uith the origins of calculus. Theseequations were
used bv Newton lo derive and unify the three laws of Kepler. These laws were
found from the earlier astronomical observations of Tycho Brahe. Here we give a In Newtonian physics i l is assumed that nr is a positive constant. Newton's law of
brief derivation of two of Kepler's laws, while a t the aame time setting forth some gravitation is used to derive the exact form of thr function F ( z ) . While these q u a -
general ideas about differential equations. tions are the main goal of thin chapter. we first d i i u s s simple harmonic motion
The equations of Newton, our starting p i n t , have retained importance through- and then basic background material.
out the history of modem physics and lie a t the root of that part of physics called
classical mechanics. ,, 1'
The first chapter of this book dealt with linear equations, but Newton's equa- $1. Harmonic Oecillators
tions are nonlinear in general. I n later chapters we shall pursue the subject of non-
linear differrntial equations somewhat systematically. The-examples here ~ r o v i d e
us tr-ilh concrete examples of historical and scientific importance. Furthermore, the We consider a particle of mass at moving in one dimension, its position a t time
case we consider most thoroughly here, that of a particle moving in a central force 1 given by a function 1 + z(t), z: R + R. Suppose the force on the particle a t a
grnv~tationalfield, is simple enough so that the differential equations can be solved point I: E R is given by -mpyz, where p is some real constant. Then according
explicitly using exact, classical methods (just calculus!). This is due to the existence to the laws of physics (compare Section 3) the motion of the particle aatiafies
of certain invariant functions called inkgrala (aometirnes called "first integrals";
rr-e do not mean the integrals of elementary calculus). Physically, a n integral is a
(1) i' + p'z = 0.
cons<,rvatiol~law; in the case of Newtonian mechanics the two inkgrals we find oscillator and ( 1) is the equation of the harmonic
corrcalrol~d tc, consrmation of energy and angular momentum. htathematically
a n int~.gralrrduces t,he number of diienaions. An example of the harmonic ornillator is the simple pendulum moving in a plane,
\Vr shall br working with a particle moving in afield ojforce F . Mathematically when one makes an approximation of sin z by 2 (compare Chapter 9 ) . Another
F is a ~,cclorJFeM on the (configurntion) space of the particle, which ir. our csse we example is the caw where the force on the particle is c a d by a spring.
rmppose to bc Cartesian three space R8.Thus F is a map F: R' + Rathat apsigns I t ie easy to check that for any constants A , B, the function
to a point. I in R' anolher point F ( z ) in R:. From the mathematical point of view,
F ( z ) is tl~oughtof (YJ a vector based a t z. From the physical point of view, F ( x )
(2) ~(t=
) A cos pl + B sin pl
is the iorce exerted on a particle located a t z. is a solution of ( I ) , with initial eonditionsx(0) = A , %(O) = pB. Infact, asis proved
The example of a force field we shall be most concerned with is the gravitational
ficld oi the sun: F ( z ) is the force on a particle located a t z allracling i l to the sun.
often in calculus courses, (2) is the only solution of (1) aatiafying these initial condi-
Thue (z, z ) -
I z .1' If z, I/:I +Rn are C" functioaa, then a &of the Lcibai.
product rule for derivatives in
tions. I s t e r we will show in a systematic way that these facts are true.
Using hasic trigonometric identities, (2) may be rewritten in t,M form
~ ( 6 )= a coa (p6 + fa), M esn be d y checked using coordinate functions.
(3)
where a - (A' + F ) " ' ia called the amplitude, and coa lo = A(A' +
8)-"'.
In Section 6 we will consider equation (1) where a wnstant term ia added (repre-
We will have o&on to mnr6der functions f: Rm- + R (which, for mmple,
muld be given by temperature or density). Such a n u p f i e d e d O if the nup
R- -r R given by e3eb paniPl derivative z -+ df/h;(r) is dedned and mtinuous
senting a cutlstant diatwhing force) :
(in Cbspter 5 we dieeuss continuity in more detail). In this cmc the gdirnt of
(4) t +p*z = K , ..
f, d e d prad f, is the map Rm* R- tbat sen& z into ( a f / h ( z ) , . , a j / k ( r ) ).
Gradf is ul ewnple of a vector field on Rm. (In Chapter 1 we mnnided d y
Then, sinrilarly to ( I ) , every solution of (4) has the form
li- vector fielde, but gnd f may be more general.)
z(1) acos (pl + lo) + P'-K Next, mnsider the composition of two C maps as follom:
(5) ,
i =
' /
I*R-*R
r
The two-dimensi&l version of the harmonic oscillator concerns a map x:: R -+ R'
and a force F ( z ) = Lmkz (where now, of course, z = (z,, Q) E R'). Equation The chain rule can be expressed in this eontext .s
(1) now has the aame form
with solutions given by wing the definitions of g m d b t and inner product, the reader an prooe tbst thi.
is e q u i d e n t to
) Ccoakl+Dinkl.
~ ( 1 =
See Prohlem 1.
Planar motion will be considered more generally and in more detail in later sec-
tions. But first we go over some mathemstical preliminaries.
The planar harmonic oscillation of Section 1 corresponds to the force field These facts reduce the proof to showing that
m(i,1 ) + (prnd V, x) = 0
'l'lti,. 1ie.lrl i. cwnsrrvative, with potential energy or (mz:+ Y , f ) = 0. But this is M since Newton's m n d law is mZ +
grad V(z) = 0 in this instance.
{'(I) = bmk 1 z 1'
:ts i s I :t.il? v<,rifird.
I'lrr :tny nroving particle z ( t ) of mans m, the kinetic energy is defined to be $4. Central Force Fielda
her^ x ( 0 is interpreted as the velocily vector a t time I; i b length I P(1). I is the speed A force field F is called central if F ( z ) points in the direction of the line through
a t timr 1. If \ye consider the function z : R-+ R' as describing a curve in R', then z , for every z. In other words, the vector F ( z ) is always a scalar multiple of z, the
f ( 1 ) is the tangent vector to the curve a t ~ ( 1 ) . coefficient depending on z:
I.'or a particle moving in a conservative force field F = -grad V, the potential F ( z ) = X(z)z.
e n e r p at I is defined to be V(z). Note that whereas the kinetic energy depends on We often tacitly exclude from consideration a particle at the origin; many central
the velocity, the potential energy is a function of position. force fields are not defined (or are "infinite") at the origin.
The lold energy (or sometimes simply m g y ) is
L e m m a Let F be a eonscrvdite force field. Then the following stahmda a n
This ha_. thr folloning meaning. If z(l) is the trajectory of a particle moving in equivalat:
the conservative force field, then E is a real-valued function of time: (8) F d r d ,
(b) F(z) = f ( l z O z ,
( c ) F ( z ) = -grad V(z) and Ir(z) = g(l z 1).
Theorem (Conservation of Energy) Lei z(t) be he tr'&ry of a parttde moving Proof. Suppose (c) is true. To prove (b) we find, from the chain rule:
force field F = -grad V. Then the low mergy E is independent of
i r ~n ror~xer~aliue
tIff1,
this proves (b) with /(I z 1) = g'(( z ()/I z 1. I t is clear t b t (b) i m p w (a). To
ahow that (a) implies (c) we must prove that V is wnstant on each sphere.
d 1
(3 m I + ~ ( z ( t ) ) )= 0.
Since any two points in S. can be connected by a curve in S., it Suaim to show that
It fol1ou.s from calculus that V is constant on any curve in S . Hence if J C R is an interval and u: J -t S. is
a C' map, we must show that the derivative of the composition 1' u
v
J-S.C Ra-R
( a vprslon u f the Leibniz product formula) ; and also that
is identically 0.This derivative is
(thr chnir~rulr)
@. CENTRAL PORCE FIELDS 21
ns in Prrtion 2. Now grad V(z) = -F(z) = -A(z)z since F is central: origin and the particle, as ia the force on the particle. This d e a it plwsible th.t
the particle always moves along the same line through the origin. T o prove thi8 let
( s ( l ) , z ~ ( t )~, ' ( 1 ) )be the coordinstes of ~ ( 1 ) Then
. we have three diKerenti.l
equations
= 0 By intwation we find
because I u(l) I a.
z.(t) = @ l l l ~ . ( k ) , h(t) = \ ' q ( 8 ) &,
R
I n Section 5 we shall consider a special conservative central force field obtained
from Newton's law of gravitation. Therefore z ( l ) ia always a scalar multiple of z ( k ) and so z(1) moves in a hxed line,
Consider now a central force field, not necessarily conaewative. and hence in a fixed plane, as asserted.
SGppns? a t some time b, that P C Ra denotes the plane containing the particle, We restrict attention to a conservative central force field in a plane, which we
tl~t-vrh)city vector of the particle and the origin. The force vector F(z) for any take to be the Carlesian plane R '. T h u z now denotes a point of R', the p0kati.l
point z in P also lies in P.This makes it plaueible that the particle etaye in the plane energy V ia defined on R' and
P for all time. In fact, this is W e : a particle moving in a central f o m field moves
in a fixed plane.
The proof depende on the cross product (or vector product) u X v of vectors U,
v in R1.We recall the definition Introduce polar coordinates (r, B ) , with r = 12 .1
Define the angular m a e n l u m of the particle to be
h = mf8,
andthut u X v = - u X u = IuI IvINmn8,whereNisaunitvectorperpendi~v where 8 is the time derivative of the angular coordinate of the particle.
lar to u and v , ((I,cl, N ) oriented as the axes ("right-hand rule"), and 8 is the angle
betwern u and u.
Then the vector u X u = 0 if and only if one vector is a scalar multiple of the Theorem (Consmation of Angular Momentum) For a particle rnODImODin
Ing a
central jorce J&M:
other; if u X v # 0, then u X v ia orthogonal to the plane containing u and u. If
u and L- are functions of 1 in R, then a version of the Leibnic product rule asserts
(as one can check u i n g Cartesian coordinates) :
Proof. Let i = i(t) be the unit vector in the h t i o n z(1) so z = n . Let j
j(1) be the unit vector with e 90" angle from i to j. A computation ahom that di/&
--
Now let z ( t ) be the path of a particle moving under the influence of a central Bj, dj/dl = -t% and hence
Iorvv fivld. We have t = ii + dJ,
Differentiating again yielb
We must now face the fact that b o h bodie4 will move. However, if mi is much Lemma
gwater than mr, its motion will he much less since acceleration is inversely propor-
tional to mass. We therefore make the simplifying assumption that one of the
hodies d o e not move; in the case of planetary motion, of course it is the sun that
is awumrd a t rest. (One might also proceed by taking the center of maas a t the Proof. k'rom the formula for 2 in Section 4 and the d e f i ~ t i o nof T we have
origi11,without making this simplifying assumption.)
$VI% place the sun a t the origin of R
' and consider the force field corresponding
to a planet of given mkw m. This field is then
-1 du - h- du
f = -- 8 = -
u* lie in dB
I)y the chain rule and the definitions of u and h ; and also
\r-)jc.re C I& a tonstant. We then change the units in which force is measured to obtain
tlrr nirn)>lr,rformula
where ~ ' ( 8 )is t h e (unit) w e n t vector to the path. Prove that the force lield
i s conservative if and only 8 the work is independent of the path. I n fact if
F = -grad V, then the work done ia V(z,) V ( a ) .-
C l ~ a ~ . lcv I
, 1 if and only if E 1 0. Therefore the orbit is a hyperbola, parabola, or 5. How can we determine whether the orbit of (a) Earth .nd (b) Pluto is an
(4lipst. arrording to wnether E > 0,E = 0,or E < 0. ellipse, parabola, or hyperbola?
-1'lis,rli1antit.v 11 = l / r ia always positive, From (4) i t follows that
6. Fill in the detaila of the proof of the theorem in Section 4.
7. Prove the angular momentum h, energy E, and maas m of a planet rue related
by the inequality
Note
.I ,hi. I I<
~ h'
I. I ' , ~ I ~ ~ \ t. o * I0I.~For some of the planets, including the earth, complete
rl i r I ~ l t ~ , h~jlr~r .1~.<,11
r~b,wrvrd;for these planets cos8 = -1 nt least once a year. Lang'a Scebnd Courr in Calculus [I23 is a good background reference for the
.l'i~,.~<,fc'r,, 11i~ir orl,its arr rllipses. In fact from a few o b s e r ~ ~ t i o of
n s any planet it mathematics in this chapter, especially hi^ Chapters 3 and 4. The physics materid
,.:)I, I t < . 41tnjri t h a t thr orbit is in fact an elliose.
is covered extensively in a fairly elementary (and perhaps old-fashioned) way in
2s 2. N E W ~ N ' S EQUATIOX AND KEPLER'S LAW
Prilzciples of JIechanicg by Synge and Griffith 1231.One can also find the mechanics
d i s c u s s in the book on advanced cdculus by
I:<].
trf
Loomis and Sternberg [15, Chapter
Thr r~nsysternaticad hoc methods used in Section 6 are successful here because
OIVn-lativc simplicity of the equations. These methods do not extend very far
Chapter 3
Intct ~~it-r~li:r~i~rs.
In general, there are not e n o u ~ h"integrals."
' l ' l ~ t b 111g~lc1
of planetary motion in this chapter i quite idealized; it ignores the
Linear Systems with Constant
P ~ : I V I ~ H I I ~ I Ieffect of t.hc other planets.
IR~
Coefiients and Real Eigenvalues
The purpose of this chapter is to begin the study of the theory of linesroperatom,
which are basic to difierential equations. Section 1 is an outline of the necessary
facb about vector spaces. Since it is long it is divided into Paria A through F. A
reader familiar with some linear algebra should use Section 1 mainly as a reference.
In Section 2 we show how to diagonalizc an operator having real, distinct eigen-
values. This technique is used in Section 3 t o aolve the linear, constant aoeffeient
system I' = A z , where A is an operator having real distinct eigenvalues. The laat
seetion is an introduction t o complex eigenvdues. Thii subject will be Btudied
further in Chapter 4.
We emphasize that for many readers this wetion should be used only ss a refer-
ence or a review.
The setting for moat of the differential equations in this book is Csrteaian space
R";thie space was defined in Chapter I, Section 2, as were the operators of addition
and eealar multiplication of vectors. The following f d i properties of t h e
:a 3. L I N E A R 8 Y S T E M 8 : CONSTART COEFFICIENTS, REAL EIGENVALUES $1. U 4 S l C LINEAR ALGEBRA
olwr:tfio~~-
ltrr imnrtdiatr consequences of the definitions: with a 1 ill the kth place, zeros elsewhere, then
c,, = x b,ras.
k-1
:it, 1~:111~~i1
' I ' i t , .I, lrr~rattlyproprrliea. Any map A: Rm-R* ~atisfying1.1 and L2 To see thin we compute the image of e, under TS:
ih I.;III,.~ .I I~rrrar.rrrnp. Kvrn murc gcncrally, a map A: R" -+ R" (pcrhaps different
c i , ~ r l ~;it111
a i ~ rang(,)
~ that eatisfirs 1,l and L2 is called linear. In the csse where the (TS)e, = B(Ae,) = B ( X ar,ck)
k
[lomain and range arc thr same, A is also called an operalor. The set of all operators
mi R nii ~lrnotedby L ( R " ) . = C a,(Bet)
k
S o t c that if er C Rais the vector
-Z k
ai(Zbt~t).
I
91. B.<SIC LlSEAll ALGEBRA 33
I t is easy to see that if A and B are the respective matrices of S and T, then the
nrxtrix of S + +
T is A B = [ a , , b.,J. +
I lltt,r:~t<,rs
and tnatrices obey the two distributive laws
provided the d e t m i m n l D # 0. If D = 0, A is not invertible. (Determinants are
T i r I, sp<*eialoperators are 0: t
the c o ~ r r s ~ o n d i nmatrices.
g
-
0 and I: z -
x. We also use 0 and I to denote
All entries of 0 are 0 E R while I = [6,,] where 6;i
considered in Part E.)
Tlrr. ti~tn!leof rl is the sct It is cssy to check that Ker A and Im A are subspaces of E, and Et, respectively.
A sirnplr but important property of I<er A is this: A is onc-lwne if and a l y if
Irn A = [ y € E, / AI = y for some z E El) Iier A = 0. For suppose A is one-to-one, nnd z E Ker A. Then Az = 0 = AO.
= A(E1). Hence z = 0; therefore 0 is tbe only element of Ker A. Conversely, suppose Ker A =
0, and Az = Ay. Then A(z - y ) = 0, so x -
y E Ker A. Then A(+ - y) = 0,
Let F be a vector space. A set $ = [ a l ,. . . , a*\ of vecton, in F is enid to span F s o z - y E 1CerA.Thusz - y = O s o z = y.
.
if every vrctor in F is a linear combination of a,, . . , a,; that is, for every z E F The kernel of a linear map R' -+ R" is connected with linear equstione (algebraic,
there an- scalars 11, . . . , l r sucb that not differential) as follows. Let A = [a,,] be the m x n mntrix of the map. Then
s = llal + .. . + 1 ~ . x = (z,,. . . , I.) is in Ker A if and only if
(III.,, 1, = . = l1* 0.
.\ hilsis I I I F is a n ordrrcd set of vectors in F that is independent and which s p m In other uords, (z,,. . . , z.) is n solution to the above system of m linear homo-
F. geneous equations in n unknouna. In this case Ker A is called the dudion spou of
T h e follotving basic fact is proved in Appendix I. the system. "Solvmng" the syslem means finding a basis jotlKer A.
If a linear map T : E -4 F is both one-bone and onto, tben there is a unique
map S : F-+ E such that S T ( z ) = z and TS(y) = y for all z E E, y E F. The
Proposition 1 E m veclor space F ha8 a basis, and every boaid of F hw Ute same map S is also linear. In this case we call T an isomorphism, and say that E and F
rlunrber of elentenla. If (el, . . . ,ell C F id an indspenderU subsel l h d is no1 a basis,
are isomorphic vector spaces.
by adjoininy lo it milable uectors e*+~,.. . , e, one can form a basis [ e l , . . . , &,I.
'I.Ir(.nrlmber of clern~ntsin a basis of F is called the dimension of F, denoted by
Proposition 2 T u v vector spaces are isontorphic if and only if 1h.q hoa Ilu m e
dim b' I f I el. . . . , e". 1 is a basis of F, then every vector I: E F can be expressed
dimension. In p a r l k l a r , every n 4 i n r m s i o d m t o r spcrcc is isomorphic to R
'
.
rn
x = X Jie., 1. € R, Proof, Suppose E and F are isomorphic. If ( e l , . . , , a)-iss back for E, it b
r- I
easy to verify tbat Te,, . . . , Te. span F (since T is onto) and are independent
C ~P,. span F. ]\loreover, the numbers 11, . . . , t, are unique. To see this,
P I I ~thc (since T is one-toone). Therefore E and F have the Bame dimension, n.Conversely,
. .
suppnsc also that
z =
,"
s,e,.
suppose ]el, . . , e.1 and ( f ~., . , f.1 are bases for E and F, respectivtIy. D&ne
T : E + F to be the unique linenr map such tbat Te. = f,, i = I , . . . , n: if +
X z.e. E h', then T z = C zJ.. Then T is onto since the f , span F, and Ker T = 0
-
I-,
since the f , are independent.
The following important proposition is proved in Appendix I.
.
To every basis {el, . . , e. 1 of a vector space E we have associated a system of
coordinates aa follows: to each vector z E E we assign the unique n-tuple of real
numbers ( t , , . . . , I.) such that z = C z.c,. If we consider z, as a function of z,
we may define a map
In fact, Q is the matrix of the linear operator J.rp-': R" R'.
How are the matrices P and Q related? To answer this we fvst relnte the bsses
-
with their corresponding coordinates:
C ( C 9klPiP1,)
-
=
111 1111sn I nrnvr- nt thr follouing definition: A coordinate ayakm on a vector
I wr
~ 1 i
i ~ u m o r ~ ~ (P. tis m Rm.(Of course, n = dim E.) The coordinates
l~~
-
h l l w l - II ' L att by (5). Each term of the intarnal eum on the right is 0 unlee 1 = j, in which caa
. .
:1 F.' :nrc ( 2 1 , . . . , z . ) , where (p(z) = (21, . , z*). Each coordinate z, is a it is g r , ~ . Thus
~.
Trrnrt~onz , ; E
I~III.:!~ R.
I\
and 811 ~aurnorphisrnE -
r thus have three equivalent concepts: a basin of E , a coordinate system on E ,
RR.
Iieaders familiar with the theory of dual vector s p a c e (see Chapter 9) will
Qr; = C qk;pi,.
i
- -
. .
rrcoguizv the coordinate functions z, as forming the basis of E* dual to (e,, . ,c J ;
P,by
R = [r,,], r,, = p,,.
hrrc A'' is the "dual space" of E , that is, the vector space of linear maps E R.
The coordinate functions z , are the unique Linear functions E R such that Each row of R is the corresponding column of P. Then,
6.. = E grj rj.
I
\\-it11coordi~~ates +: -
1.t.t u.. I..' 4 R" br thv corresponding isomorphism. Let ( f ~.,. . ,f.! be a new basis,
(y,, . . . , y.). Let E R' be the corresponding isomorphiem.
Each victor j, is a linear combination of the e,; hence we define an n X n matrix:
Thus
I =
Q = ( P i ) - : = (P')',
QP'.
The last equalitv follows from the identities I' = I, and (AB)' - FA' for MY
50 (P)-' = (PI)',
I = (PP-')' = P8(P')',
38 3. LINEAR BYBTEW8: CONSTANT COEmCIENlU, REAL EIOENVALUEE $1. BASIC' LINEAR AWEBRA 39
We have proved: Now suppase a new system of coordinatrs ( y , , . . . , y.) is introduced in E, cor-
responding to a new bask (h,. . . ,f.1. Let B be the matrix of T in the new coordi-
Proposition 4 The matrix aprwsing nno mordina!a i n leima oj the dd u ~JICin- nates. How is B related to A?
o v s r Iraiispose oj Uc matrix czprwsing Ue nnu basis in lcrms oj the old. The new coordinates are related t o the old ones by an invertible matrix Q = [q,,],
as explained in Part C. If z E E is anv point, ite two seta of coordinate.
z = ( X I , . . . , zl) and y = ( y l , . . . , yl) are related by
D. Operator, boaea, and mdtricea
I/ = Qz; z = Q1y.
In Part A we -iated to an operator T on R' a matrix [at,] by the rule (Here we think of z and y as point8 in R'.) The imsge TIalso bss two sets of eo-
(7)
'
..
Tef = C a c ~ i ; i = I, . ,n, ordinates, Az and By, where B is the matrix of T in the new coordinstes. Therefore
.
where ( e l , . . , s )is the standmi bnain of R
'. Equivalently, the ith coordinate
Hence
of T z , z = ( X I ., . . , z . ) , is
(S) C asp,.
1 for all y E- R". I t follows that
It I- 11ic.fitltorepresent (8)as the product of a n n x n matrix and a n n x i matrix :
This is a basic fact. I t is worth restating in terms of the matrix P expreaing the
.
new basis vectors f; in terms of the old basis ( e , ,. . , emf:
\\'e carry out exactly the same procedure for an operator T: E -+ E, where E In Part C we saw that Q is the inverse transpaw of P. Therefore
.
is any vector space and ( e l ,. . , e,,l is a given basis of E. Namely, ( 7 ) dehnea a
matrix [a,;]. The coordinates of T z for the h& {el, . . . ,e. 1 are computed by ( 8 ) .
I t is helpful to use the following mlea in constmeting the matrix of an operator The matrix P' can be dmribed as followa: the tth cnlumn of P consista of the m-
in a given bssis: ordinam of the new basis vector j, in the old basis [el, . . . , e,,). Observe t M in
(10) and ( 1 1 ) the inverse s i p - i appear in different places.
The j t h c d u m n of the matrix give8 the coordinates of the image of the jth bssis Two n X n matrices B and A related as in (10)by some invertible matrix Q are
vector, as In ( 7 ) . called s i m i h r . This is a basic equivalence relation on matrices. Two matried are
.I'lrr 1111 row of the matrix e x p r m the ith coodin& of the imsge of z s d a linear
-
similar if and only if they represent the same operatar in ditTerent baaea. Any matrix
function of the coordinates of z, as in ( 8 ) . property that is preaemed under similarity is a property of the underlying linear
I f we think of the coordinate. as linear function8 xi: E R, then ( 7 ) is expreaaed transformation. One of the main goals of linear algebra is to h o v e r criteria for
succinctly by the similarity of matrices.
(9) z,T = C a s f ; i = i ,..., n. We also call two operators S , T E L ( E ) similar if T = QSQl for some invertible
operator Q t L ( E ) . This is equivalent to similarity of their matrioes. Similar
This looks very pretty when placed next to ( 7 ) ! The left side of ( 9 ) is the corn&- operators define differential equations that have the same d y n a r n i d proper&.
tion
r ii
E-E-R. E. Determinant, trace, a n d rank
The right-hand side of ( 9 ) is a linear combination of the linear functiom 21,. . ,Z.. .
The meaning of ( 9 ) in that the two linear functions on E, expressed by the left and
right sides of (9),are equah Det: Af. -
We recall briefly the main properties of the determinant function
R,
~n 3. L I N E A R SYSTEMS: CONBTANT COEFFICIENTS, REAL EIOENVALUEB $1. BABIC LINEAR ~mEnR.4 41
-
( b ) 1G.r A = 0,
(c) A is one-to-one,
(d) A is &, F. Direct sum decomposition
(e) A is rnucrldle.
I n particular, Det A = 0 ij and only ij A x = 0 for some wdor z # 0. ..
Let El, . , E, be aubspaces of E. We any E is the dircd srm of them if
vector x in E can be expressed uniquely:
Another important similarity invariknt is the irace of a matrix A = [a,,]:
This is denoted
the union of the basis elements of the Ej to obtain a basis for E, T has the matrix (See Part E of the previous section) To do thw let A he a rep-tstive of T. Then
(1) ia equivalent to
By (2) there must exist a nonzero soIution vector I . The solution space of (3) is
$2. Real Eigenvaluea exactly the a-eigenspace.
Emmple. Conaider the operator A = [_:
_:I on R', used to describe a differen-
tial equation (4) in Chapter 1. The characteristic polynomial ia
k t T bc an operator on a vector space E. A nonsero vectur z E E is called a
(real) eigenueclor if Tz = az for some r e d number a. This u is called a red eigigen-
d u e ; we say z belongs lo a.
E;gcnvaluea and eigenvectom are very impartant: Many problem in phymw
and other sciences, as w-11 as in mathematics, are equivalent to the problem of = (A - 2) (A + 1).
finding eigenvectors of an operator. Moreover, eigenvectors can often he used to The eigenvaluea are therefore 2 and - 1. The eigenvectom belonging to 2 are solu-
find an especially simple matrix for an operator. tions of the equations ( T - 21)x = 0, or
Tlw condition t hat a is a real eigenvalue of T menna that the kernel of the operator
The solutions am If a real eigenvalue a is known, the general procedure for finding +vectors
belonging to a are found as follows. Let A be the matrix of T in a lmsia@. The matrix
equation ( A - a1)z = 0 is equivalent to the system of linear equation8
p.. 1
I t is rlear that ( - 1, 2) i a basis for the solution space. Therefore the v e c h
j, = ( - 1, 2) R2 is a' basis for the ( - 1) -eigenapace of T.
The two vectors all. other entries being 0. We say T is diagonaluable.
jl = ( 1 , -I), j: = ( - 1 . 2) I t is very easy to compute with D. For example, if z E E haa componente
(21, .. . , z.), thBt is, z = C zJi. then T z = ( a m , . . . , a.z;). The kth power
form a new bask ( j i ,f*)for R'. In this basis T bss the diagonal matrix D'= D . - . D (kfactors) isjustdiag(a:, . . . , a : l .
This is how the diagonalizing change of coordinates was found in Section I of Chap-
ter 1.
Example. Let T have the matrix [:-:I. The characteristic polynomial is
S111~1.r , . . . , em arc indcpendent, 83. Differential IEquatione with Real, Distinct Eigenvalua
S I I I ~aV, # n, by aqumption, each 1, = 0. Therefore, e. = 0, contradicting e. We use the resulb of Section 2 to prove an important result.
..
beir~gan ~igenvector.Hence le,, . , en) is a bask, so T is diagonalizable.
Theorem 1 Lel A be a n operafor on R^ having n dis(ind, r e d e+uduu. Thm
The lolb\\ing theorem interprete Theorem 1 in the language of matrices. R
'
, the linurr diflerential cpualia
jor all xo E
Thaorem 2 Lel A be an n X n matrix having n disltncl real kgewatues A,, . . . ,A. (1) x' = Ax; z(0) = G,
Then there exisla an invertrMe n x n mat& Q such thal has a unique solution.
Proof. Theorem 2 of Section 2 implies the existence of an invertible m a t h Q
Proof. Let (el, . . . , e.1 be the standard besi in Rn with corresponding co- such that the matrix Q A Q Lis diagonal:
.
ordinntrs (I,, . . , x.). Let T be the operator on R' where the matrix in the stand- QAQ' = diaglA1, . . . , 1.A = B,
ard basis i s A. Suppose / jl, . . . ,j.1 ia a basis of eigenvectors of T, so that Afi =
h i , , j = 1, . . . , n. Put j,- ..
(j,,,. , ji.\. If Q is the matrix whose jth column is
J,, then QAQ-' is the matrix of T in the bask / jl, . . . ,I.), as shown in Part D of
.
where A,, . . , A. are the eigenvaluea of A. Introducing the new coordinates y = Qz
in R", with x = Q'y, we find
Section 1 . But this matrix is di&(h,, . , A.1. . . y' = Qz' = QAz = QA (QLy)
80
WT \ d l often use the expression " A has real dislinel eigavalues" for the hypothesis
of Theorem 1 and 2. (2) y' = By.
Another useful condition implying diagonalizability ia that an operator have a Since B is diagonal, this means
syit~ntefn'cmatrix (aii = aii) in some bash; eee Chapter 9.
Ict us examine a general operator T on R' for diagonalizability. Let the matrix (2') -
y.' = A@,; i 1, . . . ,n.
[:
be :I; the characteristic polynomial pr(A) is Thua (2) ia an uncoupled form of ( I ) . We know that (2') has unique sdutiona for
every initial condition yi(0) :
= A% - (a + d)A + (bd - bc). TO solve ( I ) , put y(0) = QQ. If y(t) ia the eorreaponding solution of (2), then
+
Wotice that a d is the trace Tr and ad - bc is the determinant Det. The mote
the solution of ( I ) ia
x (t) = Q-'y (1).
) , hence the eigenvaluea of T, are therefore
of ~ T ( Aand
hIore explicitly,
+ (T? - 4 Det)"'].
)mr
z(l) = Q1(yr(0) exp(A11). . . . , y.(O) exp(A,L)).
Tllc roots are real and distinct if Tr' - 4 Det > 0; they are nonreal complex con-
jugr~tcsii Tr - 4 Det < 0; and there is only one mot, necesssrily real, if T r - DiRerentiation shows that
J 1)rt - 0 Therefore T is diagonalizable if TI'- 4 Det > 0. The remaining case,
'l'r' - 4 Uet = 0 is ambiguous. If T is diagonalieable, the diagonal elements are
eigcnvcctora. If pr has only one root a,then T has a matrix [; :I. Hence T = al.
Rut this means any matrix for T ia diagonal (not just diagonalieable) ! Therefore
x h m Tr' - 4 Det = .O either every matrix for T,or no matrix for T, is diagonal.
Tlie oprrator represented by [I :j cannot be disgonalized, for example..
4X 3. L I N E A R SYSTEDIS: CONSTANT COEFFICIENTB, H E A L EIOENVALUES 43. DIFFERENTIAL EQUATIONS WITH REAL, DISTIN(T EIOENVALUEB 49
Thus r ( l ) really does solve (1). writing z ( 1 ) and y (1) as column vettors,
To prove that there are no other solutions t o ( I ) , we note that z(1) is a mlution
to (1) if and only if Qz(f) is a mlution to
Henre two different solutions t o (1) would lead to two different solutions to (3), T o find a mlution z(f) with a specified initial value
wllirh is ir~iposisiblesince B is diagonal. This prover, the theorem.
z(0) = u = (U,, . .. , u.),
It is inrportant to obsrrve that the proof in constructive; i t actually shows how one substitutes 1 = 0 in (6),equates the right-hand side to u, and mlveu the result-
to tir~dscjlutions in any specific case. For the proof of Theorem 1 of Section 2 .
ing syetem of linear algebraic equations for the unknowna (a,, , . ,&) :
shows how to find the diagondizing coordinate change Q (or Q-I). We review t h
procedure.
F i t , find the eigenvaluea of A by finding the roots of t,he charaeteriatic poly-
nomial of A. (This, of course, may be very difficult.) For each eigenvalue A; find a This im equivalent to the matrix equation
corresponding eigenvector f i by solving the a p t e m of linear equations corresponding
to the vector equation
(A - AJ)f; = 0. Thus a = ( P L ) - L ~Another
. way of saying this is that the initial valueu z(0) = u
corresponds to the initial value y (0) = (Pt)-'u of (5). If one ie interestedonly in a
(This is purely mechanical but may take a long time if n is lsrge.) Write out each specific vector u, i t is easier to solve (7) directly then to invert the matrix P.
eigenvector j. in coordinates: Here is a simple example. Find the general solution to the system
(8) 2: = z,
obtaining a matrix P = [M.Then the yi are defined by the equation
2; = 2, - q.
The corresponding matrix is
Sotr t hr order of the aubscripta in (4) ! The ith wlumn of P' consiets of the coordi-
11ntc*soff.. The matrix Q in the proof is t h e i n v e m of P.However, for mme pur-
poses, it is not necessary to compute Q.
Ilr thc new coordinates the original differential equation becomes
Since A is triangular,
Det(A - XI) = (1 - A) (2 - A ) ( - 1 - A)
Hence the eigenvalues are 1, 2, - 1. They are real and distinct, so the theorem
so tlle general solution is applies.
The matrix B is
.
where a,, . . , a. are arbitrary constants, a; = y,(O). The general solution to the
original rvluation is found from (4) :
$3. DIFFEHEXTIAL EQUATIONS WITH HEAL, DIUTINCI. EIOENVALUEB
,d) 3. LINEAR SYSTEMS: CONSTANT COEPFICIEN'TE, H E A L ErOENvALUEB
From z = P5, we have
In the new coordinates the equivalent differential equation is
yI = Ill1
y: = 2 ~ ,
y; = -lh,
this leads to the numerical equation for the unknowns a, b, c. Thii amounts to inverting the matrix of coeficienta d the
lefthand side of (lo), which ie exactly the matrix PI. For particdar value8 of u,, w,
us, i t is w i e r to solve (10) directly.
This procedure can, of course, be used for more general initial valuen, z(k) = u
The following observation h an immediate consequence of the proof of Theorem 1.
nonzero solution will do; we take w = 2, U, = -2, IJ=
I Thus
Theorem 2 Lel the n X n mdriz A have n disfincf real &&uu XI, ... , X..
Then every solution lo the differenlial equalwn
The matrix PL has for ita columna the triplm fh, ft, fa:
2' = Az, z!O) = u,
18 of the fonn
By usirrg this theorem we get much information about the general character of Therefore
tllr solutions dircctly from the knowledge of the eigenvaluea, without explicitly
snlvit~gthe differential equation. For example, if d l the eigenvduea are negstive,
cviclrr~tly
lim z(l) = 0
c- which reduces to
for every solution x(I), and conversely. This aspect of linear equations will be
investigated in later chapters. cn = 0.
Theorrni 2 leads to another method of solution of (1). Regard the coefficients ci, From z; = 21 - zl we obtain
as unknowns; eet
+ 2cd' - tar' = - csl)el + (cu - ca)2' + (qI
z,(l) -C I
ci, exp(1.4,) ; i= 1, . ..,n, cnel
Therefore
(GI - co)cl.
Cn = CII - Cn,
and substitute i t into
2czr = c* - c",
z' = Az, z(0) = u.
-c., = C" - a,
Then rquate coefficients of exp(6Ai) and solve for the c.,. There r d t s a ~ y s t e m which bd1 down to
of linear algebraic equations for the c., which can always be satisfied provided Cn = ~CIL,
.
x,l . . , A. are r e d and distinct. This ie the method of "undetermined coefficients!'
La, = 0.
As an example we consider the same eystem aa before,
Without uaing the initial condition yet, we have found
z ~ ( l )= cue',
with the initial condition which L equivalent to (9). From (z1(0), a(O), ~ ~ () 0=) (1,0,O) we find
z(0) = (1,0,0).
c,1=1, c"=l, cu=-+.
,,
I IIC- rige~ivalueaare A1 = 1, A t = 2, AI = -1. Our solution m w t be of the form The solution is therefore
z(t) = (el, -e8 + 11, tet - 48').
We remark that the conclusion of Theorem 2 is definitely falae for nome 0perat.m~
with real, repeated eigenvaluee. Consider the operator A = :I, whose only eigen- [:
value ir, 1 and the sy&m z' = Ax:
TIlrn from z:(I) = XI we obtain (11) 2: = 21,
~ ( 0= ) 3.
) 0, ~ ( 0 = xl(1) = I, - ( I ) = 1
(c) Z' = Az; (d) z' = Az, show that if F - 4c > 0, then (*) has a unique solution r ( t ) for every
z(0) = (3,O) ; z(0) = (0, -b, b), initial condition of the form
( H ~ n tTheorem
: 2.)
56 3. LINEAH BYSTEMB: CONSTANT COEmCIENT8, REAL EIQENVALUEB W. COMPLEX EIGENVALUES 57
Then: P r o v i d i i b > 0, T., id a mnlerclockwiue roeolion through B rodiaw f o h d T does not correspond to mnltiplication by a complex number since its matrix
by a strefching (or shrinking) of Ilu Imgth of each Yedm by a fador of r. is not of the form Am,&.But i t is possible to introduce new coordinates in R-that
That is, if IL denotes rotation through 0 radians, thcn is, to find a new hasia-giving T a matrix A*,.
Let (u, a)be the dandard coordinates in R2.Make the substitution
T.,b(+) = r&(z) = &(rz).
To see this first observe that 1
' =A + YT,
a = r eos 8, b = r sin 8. zr = -y1,
In the standard basis, the matrix of IL ia so that the new coordinates are given by
Yl = -a,
-yl=z1+zr.
the matrix of scalar multiplication by r is r l = C I].
The equality The matrix of T in the y-coodimtea i s [: = A].,. For this matrix r = 32,
0 = r/4. Therefore in the ( y ~ ,y*)-plane T is rotation through r / 4 fdlowed
with stretching by @. I n the original coordinates (G, z,), T is a kind of "elliptical
rotation" followed by the fi-atretch. If vectors in R' are identified with complex
numbers via the y-eoordinates-the vector whose yeoodinates are (B, B) becomes
viclds our assertion.
yl + i y t h e n T corresponds t o multiplication by 1 + i.
There is another al~ebraicinterpretatim of Tar.Identify the plane R' with the
This shows that although T is not diagonalizable, coordinates can be introduced
field of complex numbers under the identification
in which T has asimple geometrical interpretation: a rotation Iollowed by aunifonn
(z, I) - z + iy. stretch. Moreover, the amount of the rotation and stretch can be deduced from
Then with tbia identification, t L operator T.& cmrcspoRdB to mdliplicolin by
+
the roots of the characteristic polynomial, since r / 4 = arg(l i ) ,32 = 1 1 + i 1.
We shall explain in Chapter 4, Section 3 how the new coordinates were found
a +ib:
(z, Y) -2 ;Y + We show now how the complex structure on R' (that is, the identification of
R' with C) may be used to solve a corresponding class of differential equations.
operate by T.4
1
(az- by, bz
I -I
+ ay) (QZ- by)
mnltiply by a + ib
+ i(br + QY)
Consider the system
Notice also that r is the nonn (absoluta value) of a + bi and B is its argument.
Readers familinr with complex functions will recall the formula a + ib re' (aee
Appendix I).
-
The geometric interpretation of T.., makes it eksy to compute with. For example,
t o compute the pth power of T..,: We use complex variables to formally find a solution, check that whst we hsve
found solves (1 1, and postpone the uniqueness proof (but see Problem 5).
Thus replace (2, y) by z + +
iy = z, and [; -:] by a bi = r. Then (1) becornea
'cos@
sin @
-r*mn@
9 cm @ 1.
-
=
Following the lead from the beginning of Chapter 1, we write a solution for (2),
Next, we consider the operator T on R' where the matrix is C -:I .T h e c h a r t(t) KC'^. Let us interpret this in terms of complex and r e d nurnbm. Write
+
the complex number K aa u iv and set z(t) = zit) + iy(f),6. = emem. A stand-
acterlstlc polynomial isA1 - 2A + 2, where rootsare
ard formula from complex numbers (see Appendix I ) ssp that eW = u m tb +
i sin lb. Putting this information together and taking real and h q i m r y parts we
.is 3. LINEAR SYSTEMS: C O N S T A N T COEIYtCIENT8, REAL ElGENVALUELi 64. COMPLEX EIOENVALUEB
The reader who is uneasy about the derivation of (3) can regard the preceding z(t) = 5b sin 16 + zocos tb,
par:bgrapll simply aa motivation for the formulas (3); it is easy to verify directly
by diRcrentiation t h t (3) indeed providerr a solution to (1). On the other hnnd,
all the steps in the derivation of (3) are justifiable.
Y (6) = yo cos Ib - basin tb,
We have just seen how introduction of complex variables can be an aid in mlving a s can be vprified by differentiation.
differential equations. Admittedly, this use waa in a very special case. However, We can put this solution in a marc perspicuous form as follows. Let C =
many s.vstema not in the form (1) can be brought to that form through a change +
[ ( V O / ~ ) ' ~'1'''and write, d n g C f 0,
of roordinatea (see Problem 5). In Chapter 4 we shall pursue this idea systemati-
cally. At present we merely give an example which waa treated before in the Kepler
problc.111 511Chapter 2.
('onsi[lrr the system
Then u' + v' = 1, and
z(1) = CCvcosIb - uein tb].
Let 1, = bkLarc con v, m that
The rorrespanding matrix is
cos bio = v, sin b4 = u.
Then r ( I ) - C (coa bl cos bb - sin bt sin bk), or
\\-hose eigenvnluea are fbi. I t is natural to ask whether A can be put in the form
(5) z(t) = C cos b(t - Lo) ;
and
through a coordinate change. The answer ia yes; without explaining how we dip (6) y (1) = bC sin b(l - Lo)
covered them (this will be done in Chapter 4), we introduce new coordinates (u, U) as the reader can verify; C and 10 are arbitrary constants.
by setting z = v, y = bu. Then From (5) and (6) we see that
\\-r hnr-e already solved the system Thus the solution curve (z(l), y ( l ) ) goes round and round an &pee.
u' = - bv, Returning to the system (4), the reader hw probably recognized t h s t i t ia equiva-
lent to the second order equation on R
v' = bu;
tbc solution with (u(O), ~ ( 0 ) =
) (uo, vo) ie
obtained by differentiating the Grst equation of (4) and then substikiting the
) uocw 16 -
~(f= sin rb, second. This is the famous equation of "simple harmonic motion," whose p r d
v(1) = u o s i n 1 6 + g c w I b solution is (5).
60 3. LINEAR SYSTEMS: CON8TANT COEPFICIENW, HEAL EIOENVALUU w. COMPLEX EIOENVALUEB
61
PROBLEMS 6. Let A = -3. me sokutiona of z' = A r depend continuously on initid values.
(See Problem 5, Seotion 3.)
7. Solve the initial value problem
I . Solve the following initial value problem.
(a) z ' = -ye
y1 = 2;
z(0) = 1, y(0) = 1.
2: -
(b) z: = - 2%
22,;
Z , ( C ) = 0, a ( 0 ) = 2.
2' = -4y,
y' = z;
(c) z' = Y, (d) 2' = Az, ~(0=
) 0, ) -7.
~ ( 0=
y1 ' -2; x(0) = (3, -9) ;
3. Let A - C; -f] and let z(l) be a solution t o z' = Az, not identically 0. The
(a) a circle if a -
curve z(f) is of the following form:
0;
(b) a spiral inward toward (0,O) if a < 0, b f 0;
(c) a spiral outward away from (0, 0 ) if a > 0, b f 0.
What effect hes the sign of b on the spirals in (b) and (c)? What in the phase
portrait if b = O?
where I = [i :].
( b ) Show that there exists a 2 X 2 matrix Q such that AQ = QB.
(Hint: Write out the above equation in the four entrim of Q = [qc!]
Show> the resulting syatern of four-linear hornogensoua equations rn
the ur unknowna q;, has the coefficient matrivof part (a).)
(c) Show that Q can be chosen invertible.
Thvn~forcthe system z' = Ax has unique solutions for given initial conditions.
$1. COMPLEX VECFOH SPACES 63
.
', then X1 is the vector (As, . , , h.) ; this b d a r multi-
z = (21, . . . , 2.) i~ in C
plication. Note that R ' is contained naturally in C. as the set of all (s,, . . , z.),
<i~rt.rI I\ I , > C" tlnd c ~ m l ~ l evxe c t ~ rspaces hy ~ystcmaticallyreplacing the real num-
I>wr-.R iiith colafi;.x nrtmhers C. We makr thi3 more p r c r q now. Observe that the above theorem is stronger than the c o m a p o d h g theorem
f '<,!!ipfrs(,'a,-!es~a,r space C- is the set all n-tuplrs z = ( z a . . . , 2.) of complex in the real case. The latter demanded the lurther substantial condition that the
I ~ I I ~ I K( STC S~ Ap~rv~idix I fur the definition of colnplex nrllllhcrs). We call 2 in C' roots of the characteristic polynomial bs real.
a ,.<lrr~lilvx vvctur or son~ctinlrsa point in C'.Cornplcx vrctors are added exactly Say that an operator T on a complex vector apace is miaimpie if it b dingod-
likt. v,.rtors ill R" (SCP C8aptrr I , Section 2 ) . Also, if is a cr~rnplexnumhcr and
4. LINEAH 8YBTEM8: CONSTANT COEFFICIENTB, COMPLEX EIOWVALUEB $1. COMPLEX VECTOH SPACEB 65
izohle. Thus by the theorem above T is semisimple if its characteristic polynomial Similarly, y E F. I t follows easily that F = FHC,that is, F is the complexification
hns 111stinrtroots (but not conversely as we shall see in Chapter 6). of the space of real vectors in F. The conveiw is trivial.
. \ % n,. Ilnvc la,tc.cl, R" C C". WP consider now more generally the relations be- Just as every subupace E C R' has a complcxification Ec C C",every operator
t a a . , . ~v(.(.tor
~ spacrs in R" and complex vector space8 in C". Let F be a complex T: B -+ E han an extsnsion to a complcx linear operator
sul)r;,,npr. C.. Then FR = F fl R" is the set of all n-tuplea (21. . . . , 2.) that are Tc: Ec +&,
in F' and arc real. Clearly, FH is closed under the operations of addition as well
as srahar multiplication by real numbers. Thus FH is a real vector space (subspace eallc-d the compleziliea(wnof T.T o define Tc,r E Ec, let
of R'j. (1) z= A9j; x, E C, z, E E.
Considrr now the converse process. Let B C R" be a subspaee and let E c be the Then
subset (lf C" obtained by taking all linear combinations of vectors in E, with complex Tcz = X,Tz,.
coefficients. Thus
I t is easy to see that this definition doea not depend on the ehoiee of the representa-
tion (1).
If (el, ..., erl = 6~is a basis for E , it is also a basis for the complex vector space
Ec; and the @matrix for Tc ia the same as the &matrix for T.
In particular, if T E L(Rn) is represented by an ti X n matrix A (in the mud
and Ec is a complex subspaee of C". Note that (Ec)n = E. We call Ec the corn-
plexification of E and FH the qace of real wdors in F .
In drfining Ec, FH we used the fact t h a t all the spaces considered were subseta
way), then Tc E L(C') is also represented by A.
- -
The queation arises ss to when an operator Q: EC EC is the compledfication
of an operator T:E E.
of C". The essential element of structure here, befidea the algebraic structure, is
thc operation of complex wnjugalwn. Proposition Lel E C R' be a r e d veclor spuce and Ec C C. its wmplen'$mtion.
+
Itccall if r = z iy is a complex number, then i = z - iy. We often write If Q E L(Ec) fhm Q = Tc for some T t L(&i if and only if
i = a(2) so that a : C + C as a map with the property a' = a . a = identity. The
s1.1q t f fix(.rl pointn of a, that is, the set of a such that a(z) = r, b precisely the set
8)f ~-m.:il1n11l11,r-rs in C.
. L ' t ~ i y ii~~.riition can bo extended immediately to C' by defining
a, or rc~nju~ation, Prwj. If Q = Tc,we leave it to the rradrr 1*1 pmvc that Qa = sQ.Conversely,
0 : C- -+ C by conjugating each coordinate. That is, assume Q commntea with a. Then Q(E) C 15; for if z E E, then uz = x, henee
aQz = Qaz = QZ
so
For this extension, the set of l i d poinb w R". Q z E ly E E c I q = yl = ECR= E.
Note also that if F is a eomplex subspace of C*, auch that a F = F, then the set Let Q I E = T E L ( E ) ; it is clear from the definition of Tc that Tc = Q.
of fixrd points of a on F is precisely FH. This map a plays a crucial role in the relk
tiun bet\\-cen real and complex vector spaces. We close this section with a property that will be very important in later chapters.
r.t.t F C C" be a C-invariant linear subspace of C". Then it follows that for v E F
An operator T on a real vector space E is s e r n ~ i m p l eif its complexXcation Tc ia a
X C, s l X v ) = r ( A ) u ( ! ' )or if we write u(W) = W for W € F, hV = XI. Thus a is dingonalbable operator on Ec. Then the t h r e m proved earlier irnpliea that a
t t r d VDIII~IVX Linrar. However, o ( v + W ) = o(v) + a(W). sufficient (but not necessary) condition for semisimplicity is that the chanrctmktic
I I ft~llc~i\s tliat for any subspacf: F C C", polynomial should have distinct roots.
PROBLEMS
',
-. IJct I: C R. and F C C" bc subspacrs. What relations, if an),, exist bet\veen T h e o r e m 1 Let T: E + E be a red operdor uilh dislrnd G e n n d u u i 1% as i n
din1 E 2nd dim Kc? Urtween dim F and dim Fn? the previot~sproposilwn. Then E and T have a direcl sun? deco~nposilwn (scc Seelion
3. 1II.. C C" is any subspace, \\.hat relation is there betwecn F and Fnc? 1F of Chapter 3),
4. f.<.t P: 1 , a~ rrnt vrrtor space and T E L ( 6 ) . Show that (Ker T ) c = l i e r ( T c ) , E = E.reEb, T = T.etTb, T.:E.+E., Tb:E&+Eb,
(lI11 = I m ( T c ) , and (T-I)C = (TC)-I if T is invertible.
where T. has red etgenvalued and Tb nonreol eyenuaim.
For the proof a.e pass to the ~om~lexification TC and apply the theorem of the
$2. lleul Operators with Complex Eigenvalues preceding section together with the above proposition. This yields a basis for Ec
(e,, . . . , e,, f,, j,, . . . ,,I,,j,) of eigrnvectom of Tc corresponding to the eigenvaluw
(A,, . . . , A,, M I , PI,. . . , Pa, n.).
I\-r incwr toward understanding the lincar differential cquation with constant
No\\.let F. h~ thr romplex subspacr of F,c' spanncd by (el, . . . , and F, be
c~~rliici~~nts the subspace spanned by If,, fe, . . . ,f., .?.I.Tlrus F. and Fs are invariant subspaern
lor TC on Kc and form a direct sum decomposition for Ec,
mentioned a t the beginning of the section. \I-c nray rewrite this equation a s s pai
of equations
1 . I'irst. rrl,svrvt: tlr:rt the. rigrnvalum of T coincid,! with the ei~envalues
(,I' itc ~~~,~~~lrl,-xilirati,~~i TC I)I.C~USC both T and TC l ~ a v ct l ~ same : charactrristic
l i l 1 I ? l l l ~ ~I,vt
~ ~ iA, ~I)(%
l ~ L I LC~KCIIVLLIUC of Tc and q a corrrsl)ondinp, rigrnvcctor in
I., , 'i~+- = i q . . \ I , I I ~ ~ ~tIhI r~ cir~~jup;ation
owration e t c ~both sidc.8, we find
For the proof of Theorem 2, aimply let Fi be the compler s u m of Ec spanned PROBLEM
by tbe eigenveetors, /, ji co- ta the ekenvduee pr, pi. Then let Ei bs
Pi n E. The rest follow. For each of the following operators T on Rt find an invariant twodimenaional
E C R' and a bsais for E giving T I E a metrix of the form 3:
Theorems 1 and 2 reduce in principle the study of .nopentor witb dintinct eigen-
values to the case of rn opentor on s d teed wetor qmw with n o d (8)
eigenvalues.
-
We now give the proof of Theorem 3.
Let T c : Ec Ec be the complexificationof T . Since Tc bas the same dues
nu T , there are eigenvectors 0 , 3 in Ec belonging to p, p, respectively.
-
Let* u + ivwithu,vE R-.Then+ = u - w.NotathaturndvsreinE~,for wbere T ia an operator on Rn (or equivalently, an n X n matrix). Sup- that
T has tl distinct eigenvalua Then Theorems I , 2, and 3 of tbe prerious section
apply to uncouple the equation and, after finding the new bssis, one can obtaiu
Hence u and u are in Ec n R m = E. Moreover, it ia easy to see that u and v sre
the solution. Letting E = R', we first apply Theorem 1 to obtain the following sys
independent (we the independence of 7, a). Therefore Iv, u] is a W for E.
To compute the matrix of T i n this baaia we Btnrt from tern, equivalent to ( 1) :
Tc(u + iv) = ( a + bi) ( u + W)
= (-bv + au) + i ( w f bu).
+ w) - Tu + iTv.
Also,
T ~ ( U db
(2b) -
dt = TS~.
Tu - au + bu,
Therefore
Here
Tu - -bv f au. T = T. 0 TI, t = (z., a) E E. e E, = E,
This me- t b t the matrix of T in the basis (v, u ) is F <I, cornpletiq the proof. T. haa real eigenvdues, and T,nonreal eigenvaluea
Note that (2s) and (2b) are equations defined not on R-, but on nubopaced E.
In the c o w of the proof we have found the following interprdation of 4 complsx and EA. But our defmitiona and &on of d i f f e r e n t i equstions rpply just m
eigenvalue of a real operator T C L(E), E C Rm: well to subapacea of R '. To find explicit solutions to the original equatim, baaea
for tbeae mbapaees must be found. This in done by finding +vectom of the
complexificatim of T , as will be explained below.
I f we obtain solutions .ad properttea of (2s)and (2b) separately, corrpsponding
information M gained for (2) and (1). Furthermore, ( 2 4 received 4 complete
Note that u and v can be obtained directly from (p and a (without reference to discwnion in Chapter 3, Section 3. Thua in principle it is euflicient togive an adyak
C*)by the formulae in the proof of Theorem 3, of (2b). To this end, Theorem 2 of Section 2 appliea to give the loll* system,
io 4. LlxE.4~SYBTEM8: CONSTANT COEFFTClENTB, COMPLEX ElGENvALUES 93. APPLICATION O P COMPLEX LINEAH A I G E B H A TO DlPPZRENTlAL EQUATIONS 71
+ L,
where T = TIm . . . m T., y = (y,, . . . , y,) € Eb = El m . . . m E. and each Ei
hm two dimensions.
21 = YI
h = -us;
The new coordinates are given by
l or x = Py,
Thua (L'b) and hence (2), (1) are reduced to the atudy of the equation
14) -
dy 4 =
T,yi on two-dimensional E,,
dl
nIu,rr, r:wI, T, has nonrenl ~igciivalues. Finally, Thmrem 3 of Section 2 applies The matrix of A in the y-coordinates is
t o put ( 4 I i l l t h r lorn1 of the equation analyzed in Section 4 of Chapter 3.
Exontple I Consider the equation
z: = -221,
2; = XI + 2% or B = AI,Iin the notation of Section 4, Chapter 3.
Thus,as we saw in that section, our diflerential equation
or
[ - 1-'i][3 = 0;
y ~ ( t )= w' cos 1 - vet sin t,
-
~ ( t ) w'sin I + vet cos 1.
The ori&l equation has as its general solution
r~(1) = (U + v)el eos 1 + ( u - v ) e' sin 1,
The first equation is equivalent to the wand, as is aeen by multiplying the second
by { - 1 - i). From the second equation we me that the solutions are all (eomplex)
~ r ( 1 )= -uel cos t + uel 8in 1.
+
nmltiplce of any nonzero complex vector w such that w, = ( - 1 i)wr; for exam- Example 2 Consider on R' the differential equation
+
plr, ~r., = -1, wl .- l a. Thua
is a con~plcxeigenvector belonging to 1 f i.
We choose the new basin ( 0 , ul for Rr CC', with u -
(1, - I ) , u = (1,O).
To find new coordinates yl, y, corresponding to thia new bask, note that m y t The characteristic equation Det(A - 11) = 0 is ( 1 - 1) ( ( 2 - f)* + 9) = 0. Ita
can be written +
solutions, the eigenvalues for A, are X = 1, r = 2 3i, p = 2 - 3;. Eigenvedom
in C1 for the complexified operator are found by solving the hornqpneous aystema
of some texts. In geometry, physics, and other kinds of applied msthematim, one
-
I:vr rvvry operator A , another operator eA, called the e z p o d i a l of A , in defined
in Sr-rtirln4. The function A e* has formal properties similar to those of ordinary or
+
B ar, x kg) 2 o +
-
rxpc~nrntialsof real numbers; indeed, the latter in a special case of the former.
Likr\rise the function t erA (1 E R) resembles the familiar elb, where a E R. I n
particular, it is shorn that the solutions of ( I ) are exactly the maps z: R +RR
+
(x, z ) A'(Y, YJ +
2A(2, y) 2 0
Writing -(z, Y)/(Y, Y) for A yield6 the inequality.
given by The bssic properties of the norm me:
z(t) = el*K (K E R"). (I) I z I 2 O a n d I z I =Oifandonlyifz=O;
Thus lye establish existence and uniqueness of solution of ( I ) ; "uniqueness" means (2) I z + y l 5 l z l + l ~ l ;
that there is only one solution z(t) aatisfyiying a given initial condition of the form (3) I m l = l a l l z l ;
~ ( 1 0 )= KO. where I a I is the ordinary abaolute value of the scalsr a . To prove the triangle
E:xponcntials of operators are defined in Section 3 by means of an infinite series inequality (21, it suffices to prove
in the operator space L(R") ; the series is formally the same as the usual series for
es. Convergence is established by means of a epecial norm on L(R'), the unijom
It+y115 lzl=+lYl'+2lzlIyl.
Since
rrnrnr. Korrns in general are d k w d in Section 2, while Section 1 briefly reviews
sonlr bkqic topology in R". Iz+yIt= (Z+Y,Z+Y)
Srrtions 5 and 6 are devoted to two leas-central types of differential equations. = IXI1+ I Y 11+2(x, Y),
Onc is a simple inhomogeneoua system and the other a highcr order equetion of one thia follows from Cauchy's inequality.
variable. We do not, however, follow the heavy e m p h i a on higher order equations
i(i 5. LINEAR BYSTEMS AND EXPONENTIAL% OF OPERATORB
it is true that
Geometrically, / t [ is the length of the vector t and
T l n ~ ai ~ )t.hr
r max norm we can take A = I/*, B - 1, or, equivdently,
E , define a norm N' on R* by
N'b) - N(v) f 1 +I,
where
z = y + z , ~ E E , z EF,
Now let N : R" -r R be any norm. We show that N is w d i n w u s . We have and I z I in the Euclidean norm of z. I t is easy to verify that N' is a norm on Rg r a d
N(x) = N ( C < C I z, I N ( e j ) ,
t i ~ j )
N' I E = N .
From thb the equivalence of norms on E follows. For let N be s norm on E. Then
where el, . . . , c. in the standard bsein. If we m y sssurne N ia restriction to E of a norm on R', alao denoted by N. There
m s ~ r l N ( a ).,. ., N ( L ) 1 = M , exist A , B E R such that ( 4 ) holds for all z in R', so it holds a fortiori for dl t
then in E.
N(z) l M Z= I zi ( _< Mn l t .= , I We now define a normed veetor space ( E . N ) to be s vector apaee E (that is, a
subspace of some R") together with a particular norm N on E.
We shall frequently use the following corollary of the equivalence of norm:
By the triangle inequality,
INlz) - N ( Y ) I _i N ( t - Y ) Proposition 2 Lel ( E , N ) be any nonned w l a r space. A aqumx ( t r l in E con-
ucrgea Lb y ifand only if
<MnIz-y\.
(5) lim N ( t r - y) = 0.
This shows that N in continuous; for suppose lim z k = y in R*: .-*
I N ( z ~-
) N ( v ) I ':Mn I - Y I, Proof. Let A > 0, B > 0 be se in ( 4 ) . Suppose (5) holds. Then the inequdity
so lim N ( a ) = N ( y ) inR. O<(a-y[<A-'N(*,-y)
S;nee N is continuous, i t a t t a b a maximum value B and a minimum v d u e A
on tlre closed bounded set
shows that limb,, I n- y I - 0, hence n + y. The converee in proved *ly.
.
( d ) Let 1% . . . G.1 be a bask for E. Show that there is a unique inner
product on E mch that
In other u-or&, 11 T 11 is the maximum value of I Tz on the unil baU
(e,,e,)=6., f o r d i,~.
6. Which of the following formulas define n o m on R'? (Let ( t , y) be the co-
ordinates in R'.)
-
The e d e n c e of thin maximum value followa from the compactness of D (Section
1, Propaition 3) and the continuity of T: Rn Rn.(This continuity follow im-
mediately from a matrix representation of T.)
+ +
(a) ( 9 t y yt)l/'; (b) (i.'- 3ty $)lf" + The uniform norm on L(RS) depends on the norm chosen for R-. If no norm on
R" is Bpecified, the standard Euclidean norm is intended.
c + y ; (4 t ( l ~ I + l ~ I ) + ~ ( ~ + f ) ~ ~ ~ .
5. 1st U C R" be a bounded open set wntnioing 0. Suppose U is wnvez: if t E U
+
and y E U , then the line segment [Lt (1 - t)y 10 5 15 11 k in U . For
Lemma 1 Le4 R' be gim a norm I t I. The wrrmponding uniform norm an L(Rm)
has the following propertics:
each z C Rq deiine
a(z) = lead upper bound of ( A 2 0 1 kz E U I . (a) I f ] [ TI1 = k,ihen I T t I Ik l t I joralltinRm
(b) II S T I I 5 II 11.11 T II.
Then the function (c) I I T - ~ [ < l I T 1 [ - j o r o l l r n = 0 , 1 . 2, . . . .
Proof.
y = 1 t I-'%,
(a) If
then
t - 0, then I TI I = 0 = k I t I. If z Z 0, then Iz[Z 0. Let
is a norm on R'. I
l y l = - - l t l = 1.
8. Let nl. be the vector space of n X n matrices. Denote the trampme of A E M. It I
by A'. Show that an inner product (see Problem 5) on M, is defined by the Hence
formula
( A , B) = Tr(A'B)
Express thin inner product in term of the entries in the matrices A and B.
from which (a) follow.
9. IXnd the orthogonal complement in ni. (exProblem 8) of the eubepaee of (b) Let ( t ( 5 1. Then from (a) we have
diagonal matrices.
IS(Tz)I 5 IISII.ITtl
10. r i d a basis tor the subspace of M. of matrices of trace 0. What is the ortho-
g o d complement of this subspace? I llSII.lI TII.It.1
5 II SII.11 TII.
S i e 11 S T 11 is the maximum value of I STz 1, (b) follows.
53. Exponentinls of Operators Finally, (c) is an immediate consequence of (b).
We now define an important aeries generalizing the usual exponential eeries. For
The sct L(R") of operators on R" is identified with the set nl. of n X n matrices. any operator T: R" + R' define
This in turn k the ssme as R"' since a matrix is nothing but a L i t of n' numbers.
(011~ cl~noscean ordering for these numbers.) Therefore L(Rw)is a vector apace
under the usual addition and scalar multiplication of operators (or metricen). We
may thus speak of n o m on L(Rm),convergence of series of operators, and so on. (Here k! is k f a c l o d , the product of the 'rat k positive integer8 if k > 0, and
A frequently used norm on L(Rm)is the uniform wrm. Thin norm in defined in O! = 1 by definition.) This is a aeries in the vector space L(R").
t e r m of a given norm on R' = E, which we shall write as I t 1. If T: E 4 E is an
operator, the uniform norm of T is defined to be
Theorem The ezponerJd a r i c a Tk/k! i a ab8alukIy cotwc~genifar s ~ r y
IITII = ~ l l T z l I I z Il l l . operafar T.
84 5. LINEAR B T a n Y L I AND EXPONENTIAIB OF OPZRATORB
Proof. Let It T (1 = a 2 0 be the unifonn norm (for some norm on R'). ( b ) if S T = T S , Utm cd+" = cde';
Then <
11 T'/klll d / k ! , by L m u ~1 , proved earlier. Now the real series ( c ) e-* = (@)-I;
(d) i f n = Z a n d T = C ; 3 , U t m
d / k ! coovergea to G (where e is the base d natural logarithms). Therefore
the exponential aeries for T converges mbaolutely by the m p s r i s o n t . ( k t i o o
2).
We have also proved that
11 & 11 5 e''&IL. The proof of (a) follows from the idsntities P ( A + B ) P ' = P A P ' + PBP-'
We ahall need the following d t . and ( P T P 1 ) ' = P P P . Therefore
while Therefore
where cJ ia the complex number ( d ) ' / k ! . Utb,g P = -1, we &d the R.1
I( A, ( 1 < m. Similarly, C" [I A, ll.I\ Bk 11 part of e'to be thesum of the Taylor series ( a t 0 ) for coa b; simil.rly, the-
This tends to 0 as n -+ m &ce
-
+
part is sin b. This proves ( d ) .
0 as n -+ m. Therefore lim,,(m - ad.) 0, pro* the lemma Obsave that ( c ) implien that 8 b invertible for every operator 5. Tha is d
wua to tbe fact that Z # 0 for every real nlsmber a.
The next result is uaeful m cornpuling with enponmt*. As an example we compute tbe exponential of T = C; 3. We rrite
(a) ij Q -
Propodtion Lei P, S, T dmdc o p a d m 8 on R'. Then:
P T P I , Lhtn 8 - PerP1;
$3. EXPONENTIALS OF OPERATORB
$8 5. LINEAR 8YSTEYB AND EXPONENTIAL8 OP OPERATOR8
For, from T t we
= a, obtain
\'<,tr that n l commutes uith B. Hence
er = er@ = &.
Thus
We conclude this section with the observation that d l that hea been aaid lor
exponentiab of operators on R' also holds for operatom on the m p l a veetor spsee
C*. This is becauee C' can be conaidered aa the real vector spree R" by simply
ignoring nonreal scalars; every complex operator is a forfiwi a red operator. In
addition, the preeedii statement about eigenvectors in equdiy d i d when eomplu
\\-e enn now compute e" for any 2 x 2 matrix A. We will see in Chapter 6 that eigenvaluea of an operator on Cn are considered; tbe proof is the same.
car1 find an invertible matrix P such that the matrix
B = PAP1
I I T Ill0 < t + f, 14. Show that if 1 1 T - 111 is suficiently smell, then there i an operator S such
+
that es = T. (Hinf: Expand log(1 t ) in a Taylor series.) To what extent
u-here 11 T 11s is the uniform norm of T corresponding to the Euclidean is S unique?
(c)
&norm on R'.
For any basis 63 of R', 15. Show that there is no real 2 X 2 matrix S such that h = [< 3.
I/ T Ila > t.
6. ( a ) Showthat
I1 T 11.11
T-lIl l 1 04. Homogeneous Linear Systems
for wery invertible operator T .
( b ) If T haa two distinct real eigenvalues, then Let A be an operator on R'. In this section we shall express mlutiws to the
equation :
"
H. Let A E L(R") be invertible. Find r > 0 such that if / I B A 1 1 < e, then Proposition
B is invertible. (Hinl: F i t show A-'B is invertible by applying Problem 7
to T = A-'B.)
9. Compute the exponentiala of the following matricea ( i = 0) :
In other words, the derivative of the operator-valued function e" b another
1' 1::]
operator-valued function Ae'". This means the composition of elAwith A ; the order
of composition doea not matter. One can think of A and ccAas matrices, in which
erA@A - c~~
= lirn ---
A 4 h
\Ye car) now solve equation ( I ) . We recall from Chapter 1 that the general solu- whcre a, b are constants. In matrix n o t a t i ~ nthis is
tion of thc scalar equation
' f 1 1 ~ . *:ilur is true where z,a, and k are allowed to be complex numbers (Chapter 3 ) . el*K.
'l'11c.s-c rcsults are special cases of the following, which can be considered aa the In Section 3 we rwrw that
fundanlentgl theorem of Enear differential equations with constant coefficients.
Throrenl I,et A be an operdor on Rm.Then Uu solulwn of the initial value problem Thus
(1'1 t' = Az, z(0) = K € R", el" = (ehK1, e"(tbK1 + K*))
Thus the solution to (3) eatisfying
SIIIV~, P'h' = K, it follows that ( 2 ) is a solution of (1'). To see that there are no
ottrt.r ic,lrttions, let z(t) be any solution of (1') and put
Since at know how to compute the exponential of any 2 X 3 matrix (Section 3),
wv ran explicitly solve any two-dimensional system of the form z' = A z , A E L(R1). I"'
IVithout finding explicit solutions, we can also obtain important qualitative in-
formation about the solutions from the eigenvaluea of A. We consider the most
important special cases.
Case 1. A h a r d &endues of opposik signs. In this ease the origin (or some-
times the differential equation) is called a soddk. A s we saw in Chapter 3,after a
suitahle change of coord~natesr = Py,the equation becomes
In the ( y,, y?) plane the phase portrait looks like Fig. A on p. 91
Case 11. ~ ( @ive
.4U q p w d l ~ havc real pa~t8.This important case is called
a smk. 1t hau the characteristic property that
are of the form with y(l) as above and P t L(R2); clearly, x(l) -+O as I-- m .
The p h portrait for these subcases looks like Fig. B if the eigenvalues are
equal ( a j m ) and like Fig.C if they are unequal ( a node).
If the eigenvalues are negative but A is not diagonaliible, there is a change
of coordinates z = Py (see Chapter 6) giving the equivalent equation
Y' = BY,
where
Cane I V . The eigenvalues ore pure imaginary. This is called a center. I t is charac-
to obtain the equivalent system
terized by the property that all solutions are periodic with the same period. To see
= By,
B 2 [; -:I. this, change coordinates to obtain the equivalent equation
given by
+,(z) = eIAz.
The collection of maps 1 + , 1 t a is called the Pow corresponding to the differential
equation (1). Thia flow has the basic property
this is proved in the proposition in Section 2. The flow is called linear bemuse each
map @, : R" -+ R" is a linear map. I n Chapter 8 we s h d define more general nonlinmr
flows.
The phsse portraits h u s e e d above give a good visualization of the wrrespond-
ing flows. Imagine points of the plane all moving at once along the curves in the
direction of the m w s . (The origin stays put.)
PIG. G
The portrait in the y-crmrdinatea consists of concentric circles. In the original PROBLEMS
z-coordinates the orbits may be ellipses aa in Fig. F. (If b < 0, the m w a point
rln(,k\\-isr.)
1;iprr G summarizes the geometric information about the phase portrait of 1. Find the general solution to eaeh of the following systems:
r' : A r that can be deduced from the characteristic polynomial of A. We write
this polynomial as
+
hy - (Tr A)X Det A.
The d i s m ' m i ~ dA is defined to be
A = (TrA)' - 4 Det A.
The eigenvalues are
+ (Tr A * <A).
T~IIIS r1,:11eigenvalue correspond to the case A 2 0; the eigenvalues have negstive
rtnl 11:irt when Tr A < 0; and so on.
'rhv cr~nnctricintrrpretation of z' = Az is aa foUows (compare Chapter 1). The
n1:1p R,, -t R" which sends z into Ax is a vector field on R-.Given s point K of
R". t1cr.r~.is s uniqw curve 1 --t el*K which atsrte at K a t time %em,and is a aolution
. In ( a ) , ( b ) , and (c) of Problem 1, find the solutions 8atisIyinS each of tbe
uI ( 1) I K r interpret 1 as time.) The tangent vector to this curve a t a time b is the
following initial conditions:
vcctor Ar(4) of the vector field a t the point of the curve z(b).
We may think of points of R" flowing simultaneously along them solution curves.
(a) z(0) = 1, y(0) = -2; (b) z(0) = 0, y(0) = -2;
The [mitiun of a point z E R" a t time 1is denoted by (C) z(0) = 0, y(0) = 0.
3 Let A r R' -+ R' be an operator that leaves a subpace E C R. invariant.
Let 2: R --r R" be a solution of z' = A I . If z(b) E E for some 4 E R, show
Thus for each 1 E R we have a map that r(1) € A' for all 1 € R.
+,:R"-+R" (1ER) 1. Suppose A E L(Rm)has a real eigenvalue X < 0. Then the equation d = A2
5. LIREAR SYSTEMS AND EXPONENTIAL6 OF OPERATORB 8.5. A NONHOMOGENEOUS E Q U A T I O N 99
if Tr .4 = 0, and that in this case the origin is not a sink or a ,*wee. (Hint:
hrrs a t least one nontrivial solution z(r) such that
An operator is area-preserving if a ~ r donly if the detrrmi~rantis *I.)
lim ~ ( 1 =
) 0.
,-- 12. Describe in words the phase portraits of z' = Az for
10 \\'l~irhvalues (if any) of the parameter k in the following matrices makes the (1) z' = Ax + B(1).
(#ripina sink for the corresponding differential equation z' = Az? Here A is an operator on R. and B: R --r R" is a continuous map. This equation in
called nonhomogeneous because of the tcrm B(1) which prevents ( 1 ) from being
strictly linear; the fact that the right side of (1) depends explicitly on f makes i t
nonuulonommra. I t is difficult to interpret solutions geometrically.
We look for a solution having the form
Since z ia assunled to bs a mhrtion of (2), Note that u ( l ) ia also a solution to ( I ) , while ed'K isa solution to &e h v
equation
obtained from (1) by repking B(1) with 0. In fact, if v(f) ia any dution to ( I )
By integration +
and y(1) any solution to (4), then clesrly z = v y in another mlution to (1).
f(1) - fCbB(8) * + K. Hence the general solution to (1) ia obtained from a solution by adding
to it the general solution of the cmmsponding homogenema equation. In summary
so nu a candidate for a solution of (1) we hrw Theorem Lcf u ( f ) be a parfinJ4r sdulwn o j Lhc nonlbmopmamu linmt diflcrenfid
mdwn
Let us examine (3) to see that it indeed m&es sem. The i n w d in (3) and
the prrvioua equation ia the vector-dued hrnction 8 -+c*'B(r) mapping R into
R..In fact, for any continuowmap g of the reds into a vector apace Re,the integral (4') a? = Az.
can bs defined .a an element of R.. Given a baain of R", this integd ia a vector
whom c o o d b t e s M the integrds of the eoordiarte. function8 of g. Conueraely, flu tun1 of a dclulion of (1) and a .d- of (4') u a dulion of (1).
The in- .a 4 hrnction of its upper limit t in a map from R into R". For each
1 the operator acb on tbe in- to give an dement of R". So I -+ z(t) in a well- If the function B(1) in at all eomplicbted it w i l l probably bs hpcmible to replace
defmed map from R into E. the integral in (3) by a h p l e formula; sometimes, howwer, this esn bs done.
To check that (3) is a d u t i o n of ( I ) , we differentiate z(1) in (3) : Erampk. Find the gPnerd solution to
(5) x;= -a,
Here
t; = 4 + f.
Thus (3) ia indeed a solution of (1).
That every solution d (1) must bs of the form (3) can bs eeea M fo~owa.Let
y: Rm-+E bs a eecond solution of (1). Then
I
Performing the matrix multiplication and aimpl'if+ yields Thue if z(t) = ( z ~ ( t )a, ( t ) ) is a solution of (2), then s(t) = =I(() ia a Bdution
z,(t) = -1 + KIOOS1 + (1 - KI) 8in 1, d (1) ; if s(1) is a solution of ( I ) , then z(1) = (s(l), sl(l)) ia a solution of (2).
This procedure of introducing new variables works very generally to redue
zt(t) = 1 - (1 - Kg)cm t + KI sin 1. &her order equations to first ordm ones. Thus consider
86. Higher Otder Systems Proof. One uses induction on n. For n = 2, this is easily checked. Aawne the
truth of the propwition for n - 1, and let A, be the (n - 1) X (n - 1) sub
matrir of A ooneiating of the laat (n - 1) rows snd laat (n - 1) columns. Then
Consider a linear differential equation with constant coefficiente which involves Det(XI - A) is easily computed to be X Det (XI - A-I) +
a.by expndbq along
a derivative higher than the first; for example, the first column. The induction hypothesis yields the desired chrracteristie
polynanial.
By introducing new variables we are able to reduce (1) to a f i t order system The point of the propmition is that it giver the c h a r a c ~ pcdymmd
c dinxtly
of two equations. Let q = s and = 2; = a'. Then (1) becomes equivalent to the from the equation for the higher order differential equation (3).
104 5. LINEAR SYSTEM8 AND EXPONENTIAL0 OF OPERATOR8
06. HIGHER ORDER 8Y8TEMS
105
nt lust t l ~ ~~ ~t C Sroots
P are real and distinct. Then (1) reduces t o the equation of CI = 1,
Irr~tordt,r ( 2 ) ; onr can find a diagonaliing system of coordinates (yl, yz). Every
-C1+C* = 2.
x,lrltio~l<,I ( 2 ) for thrsr coordinatesisthen yl(t) = K, exp(All), y,(l) = Ktexp(Atl),
\ \ i t 1 1 nrlnt r a r ? runstants K1, K,. Thua r l ( t ) or a(!) is a certain linear eombimation Hence CZ = 3 and the solution to ( 5 ) is
a ( ( ) = p , , K , exp(A11) + p12Ktexp(Aat). We conclude that if At, At are real and
distinct then every solution of ( I ) is of t h e form
a(!) = C' + 3le-I,
The reader may verify that this actually is a solution to (5) !
The final case to consider is that when AI, A, are nonreal comple* coajugk n--
for some (real) constants C,, C,. These constants can be found if initial values bers. Suppose A, = u +
iu, A: = u - iu. Then we get a solution (ssin Ghspter 3) :
s(/,), a l ( t O ) are given.
S r x t , suppose that AI = Ar = A and that these eigenvalues are real. In this case
yl(f) = eg'(KI cos vl - Kz sin vl),
tlrv 2 X L' nlatrix in ( 2 ) is similar t o a matrix of the form ~ ( t=
) e"(K18invl + Kteosvt).
Thus we obtain s(1) aa a linear combination of yl(l) and y,(l), so that findly,
s(t) = egl(Clcos 01 + C1 sin A)
11s uill shown in Chapter 6. In the new coordinates the equivalent fmt-order for some constants CI, Cl.
systcnr i? A special caap of the laat equation is the "harmonic oaeillator":
Y; = Ayr,
a"+ bts = 0;
the eigenvalues are *ib, and the general solution is
thv nlrthods of Section 4 we find that the general solution t o such a system is
CI cos M + CI sin M.
We 8ummarir.e what we have found.
Kl ilnd K 2 being arbitrary constants. In the original coordinates the stlutions to Tbeorcm Lei Ah, A. be the rooh oj Ihc polylmniol A' + aA + b. Then nmy didion
1111. ~.,~rlivalmt first order s y s t ~ mare linear combinations of these. Thus \ye con- oj the d i h e n l i o l egudion
cll~rl~.t l ~ n t~f the chrrracteristic polynomial of ( I ) has only one root A € R, the
solrlti~llrshave the form (1) 8" + as' + ba = o
is oj lhe following type:
(6)
,(-I+ alsi--"+ . . . + Iha = b(1)
PROBLEMS
r a n Irc solved (in principle) by reducing it to a h t order inhomogeneous linear
1. Which of the fouowing functions nntiafy an equation of the fonn a" + +ad
and applying variation of constants (Section 5). Note that bs = O?
( a ) let (b) P - 1 +
(c) cos 21 3 9in 21
r 0 i +
(d) cos 21 2 sin 31 (e) r1 ws 21 (I) e1 4+
(g) 31 - 9
2. Find solutions to the following equatiom ha* the w e d initial values.
( a ) a" +
4s = 0 ; a(0) = 1, ~ ' ( 0 )= 0.
(b) 8" - 3a'+ 21 = 0 ; a(1) = 0, ~ ' ( 1 )= -1.
h s in the case of first order systems, the gene& mlution to (6) can be expressed
3. For each of the following equations find a basis for the solutions; that ia, find
~ lthe
i general solution to the corresponding homogeneous equation
+
two solutions al(l), e(1) such that every solution has the form a ( l ) Be(1)
for suitable constants a,8 :
(a) 8" +
36 = 0 (b) a" - 3s = 0
plus s particular solution of (2). Consider, for example,
( c ) (I" - 8' - 6a = 0 (d) a" 8'+ +
8 = 0
A particular solution to (7) is 5. State and prove a generalization of Problem 4 for for nth order differen-
E(1) = 1 - 1 tial equations
Hrnrc the general solution to (7) ia
+
+ a,a("-ll ... +a. = 0,
where the polynomial
A c o s I + B s i n l + l - 1. A' + a,A'-1 + . . . + a.
Finally, we point out that higher order rydcms can be reduced to fvst order hea n distinct roots with negative red parta.
10% 5. LINEAR BYSTEM8 AND EXPONENTIAIG OF OPERATOH8
l~as
+ +
as b = 0 B U C ~that the equation
a(1) = 0 Chapter 6
(a I nr* sohition; Linear Systems and Canonical
(11) :I pnsitivc finitp number of solutions;
( r ) i r ~ f i n i t rmany
l~ mlutions? Forms of Operators
; I : , , ~,.:,,-ho f t h c following equations sketchthe phase portrait of the cnrrpspond-
ilra first ordrr gystcm. Then sketch the graphs of several solutions a(!) for
diflprcnt initial conditions:
$8, + =0 (b) 8'' - a = 0 (c) 8" + +
8' 8 =0
(a) n- f r . 2s' = 0 (e) a" - a' 8 .
\ -
+
8. Which quatiom 8" + as' + bs = 0 have a nontrivial periodic solution? What
is the period?
The aim of this chapter is to achieve deeper insight. into the solutions of the
10. Find a real-valued function 8(l) such that differential equation
8" + 4s = ca9 21, (1) zl=Az, AEL(E), E=Rm,
s(O)=O, s'(O)=l. by decompoeing the operator A into operators of particularly Bimple kin&. I n
Section. 1 and 2 we d e c o m p the vector space E into a direct 8um
11. I'ind all of functjons z(l), y(t) that satisfy the system of differential
equations E = E , s . . . mE,
2' = -y, and A into a dire& ~ u m
-
Srrlion 6 applies the results of Section 5 to the higher order one-dimensioaal Let us s w what this decomposition means. Suppose first that there is only one
linc:tr t~o~rrogeneousequation with constant coefficient8 eigenvalue X, of multiplicity n = dim E. The theorem implies E = E(T, A). Put
('2) +
a(-' als{-l + ..
. + L a = 0. N=T-XI, S=M.
are easily found if the roota of the charncteristic polynomial
Sulu~ior~s n e n , clearly, T = N +
S and SN = NS.Moreover, S is diagonal (j,, b)
A" + alAW' + .. . + a. and N is nilpotent, for E = E ( T , A) = Ker N". can t~~ we
-&tely
compute
are known. A different approach to (2). via operstom on function spaces, is very
briefly d i i c u d in the last section. "-1 Nb
er=EBeH=&x---;
Ttlc first four sections deal not with differential equations, only linear algebra. r k!
This linr~aralgebra, the eigenvector eheory of a real operator, is, on one hand, there is no difficulty in finding it.
randy t r r a t d in tpxta, and, on the other hnnd, important for the study of linear
differentid equations. Emmpk 1 Let T = J]. The characterhtic polynomial i.
p(l) = b - 41 + 4 = (1 - 2)'.
There is only one eigenvalue, 2, of multiplicity 2. Hence
8 1. T h e Primary Decomposition
e r = eS@ = 8 ( l + N)
.
Here A,, . . , L are the distinct roota of p(t), and the integer n, 2 I is the multi-
+ .. +
plicity of XI; note thnt nt . n. = dim E. More generally,
We recall that the eigenspsce of T belonging to Xr is the subpsce
err = e"eM = e"(I f tN)
Ker(T - XI) C E
(w-e write Xt for the operntor L I ) . Note tbst T is diagonalicable if and only if E
is the direct sum of the eigenapaces (for this meam E has a basis of eigenvectors).
We define the generdired eigmspou of T w n g lo Xt to be the subspace Thus the method applies directly to mlving the differential equation z' = Tz
E( T, L) = Ker (T - At)" C E. (aee the previoua chnpter).
For comparieon, try to compute directly the limit of
Note that thie subspnce is invariant under T.
Thr following primary decapouitim theorein is proved in Appendix. 111. "'[I
Mk! 1
-73
Theorem 1 Lel T be an operator on E, w h E is a cmplct vector F . OT e b e h' I n the general ease put
G real and T has real e+nt.dues.Then E is lhe direct arm of the gmrdized eigen-
spaces of, T. The d i e of each gmrdized cigmspoce e#d, Me mul*ipIin'ty o f & ~ = T I E ( . L TI.
,
corresponding eigenualue. Then = Ti e .. . e T,. Since each TI hss only the one & e n d u e ht, we can
1 12 6. LINEAR BYSTEMS A N D CANONICAL YORM8 OF OPERATORE $1. THE PRIMARK DECOMFmSlnON
113
apply thv previous result. Thus The one-dimensional generalied eigenapace of 4 1 is the mlutian sp.ee of the
eystem of q u a t i o m
Tk=Sk+Nk; Sk,Nk€L(E(Xk,T)),
( T o - 1)r = 0,
xvhere St= 1.1 on E ( h , T ) , and Nt = Tk- Sk is nilpotent of order nk. Then
or
T=S+N,
where
S =&e ... es,,
N = N l e . . . eN,.
Clearlv, S N = N S . Moreover, N ia nilpofd and S ia diagonaliE&ble. For if m =
..
max (n,,. , n.) , then one can verify that the vector
N" = ( N , ) " e ...
e (N.), = 0; Qa = (0,2,1)
and S is diagonalizcd by a basia for E which is made up of bases for the generalized is a basis.
rigenspares. , ,a , ) of R'. Let T = S
Let 03 be the basis ( a ~I% + N bs as in Theorem 2. I n
We have proved: d3-coordiitea, S hss the mntkx
[: :3
nates is
(I-')' = 0 1 0 ,
11at r i ~r~ultiplication
~ give8 Finally, we obtain
which gives
We have now computed the matrices of S and N. The reader might verify that
N9=OandSN=NS.
We compute the matrix in standard coordinates of @ not by computing the matrix
eSQdirectly from the definition, which involves an infinite series, but as follows:
-
s m m h g E is a complex vector apace.
I t suffices to ahow that P ( A ) z 0 for all z in an tubitrsry g e m d i d e&nquca
42. THE 8 +N DGCOMWBITION 117
10. I f A and B are wmmuting operators, find a formula for the semisimple and
nilpotent parts of A B and A +
B in terms of the corresponding parts of A
and B. Show by example that the formula is not always valid if A and B do not
1. For each of the following operators T find bases for the generalized eigenspacea; commute.
pivr the rnatriccs (for the standard basis) of the ermieimple and nilpoknt
11. Identify R'+' with the eet P. of polynomiale of degree _< n, via the c o r n
parts of T.
spondence
13. A rotation around a l i e in R' and reflection in a plane in R' are semisimple
operators.
14. Let Sbesemisimpleand Nnilpotent.If S N = N S = 0 , then S = Oor N = 0.
2. A matrix [axi] such that aii = 0 for 1 5 j is nilpotent.
(Hint: Consider generalised eigenspacea of S 3- N.)
3. What are the eigenvdues of a nilpotent matrix?
15. If II = T, then T is diagondizable. (Hint: Do not use any d t s in this
4. For a h of the following matricea A, compute elA, 1 E R: chapter!)
16. Find necessary and sufficient conditions on a, b, c, d in order that the operator
C: :I be
(a) diagondkable; (b) semisimple; (c) nilpotent.
17. Let F C E be invariant under T E L ( E ) . If T is nilpotent, or seminimple, or
diagonalisable, m is T I F.
with S semisimple, N nilpotent, and SN = N S . We dso found a c a n o n i d fm We call the matrix in Theorem 1 the canonical f m of N .
for S , that is, a type of matrix representing S which is uniquely determined by Let A be an elementary nilpotent matrix. I t is evident that the rank of A is
T , except for the ordering of diagonnl blocks. In the complex case, for example, n - I ; hence
d i i K e r A = 1.
wl~rreA,. . . . , X. we the r m t s of the charncteristic polynomial of T , listed with Thii implies the following corollary of Theorem 1.
their proper multiplicities.
Although we ahowed how to find r o m matrix repreaentstion of N , we did not Theorem 2 In Themem 1 the number r oj blocks ia quai Lo dim Ker A.
give any special one. In this section we shall find for any nilpotent operator a matrix
that is uniquely determined by the operator (except for order of diagonal blocks). We define the canonical form of a nilpotent matrix to be the canonical form
From this we shall obtain a special matrix for any operator, called the Jordan of the corresponding operator; this is a matrix similar to the original one. Since
1;1
esnonical form. aimiiai matricea correspond to the same operator, it followe that they have the
An elemmlary ni2pdcnl Mock is a mat& of the form same canonical form. Fmm this we conclude:
(1)
Theorem 3 Two nilpotml n X n malriccs, or Lwo nilpot& operatme a ULc
vector spcce, are sirnilor if and mdy Chey hove Ihe sam mnonical jmm.
1 0
The queation arisee: given a nilpotent operator, how is it8 canonical form found?
To answer this let us examie a nilpotent matrix which is sfready in canonical
form, say, the 10 X 10 matrix
with 1's just below the d i o d nnd 0's elsewhere. We include the one-by-one
mntrix [O].
If .V: E --r E is m operator represented by euch a matrix in a basis el, . . ., e.,
then A.' behaves ss folloanr on the bdis elemente:
N o t e that .,
= 0 if k > 3. The numbers 4 depend on the operator, and can be
computed from any matrix for T. On the other hand, if we know the vr we can Subtracting the second of thew, equtiona from the h t gkcs
immediately write down the matrix N. The problem, then, is to compute the v,
in terms of the 6r.
Considrr
.I - 26, - a,.
61 = dim Ker T. Subtrscthg the (k + l ) t h from the kth giwa
From Theorem 2 we find
il = total number of blocks - rl + r1 + rs. and the Lst equation gives P. Thus we have proved the fob* tbsnm,in which
put (b) dawa IMto compute the nxamid form of any nilpotent -1~:
Next, consider 6, = dim Ker II. Each 1 X 1 block (thnt in, the blocks LO])
contributes one dimension to Ker P. Each 2 X 2 block contributes 2, while the
3 x 3 block also contribute0 2. Thus Theopcpl4 L d T b e a n i l p d m t o p c r d m a a n k d i m n u i o n o l a t d o r s p a c e . If.,
w & ~ o f k ~ k ~ i n & a o ~ n i o o t f a n o j T -, dd i & mKaP,h
k following squdwnr a n did:
l
For 6, = dim Ker P , we see that the 1 X 1 blocks each contribute 1; the 2 X 2
blocks each contribute 2; and the 3 X 3 block contributes 3. Hence
(8)
(b)
- CIS-
.,- 2.4- a,,
h krr + m Lsrj;
. m - I, . . ., n;
+
(Some of the d i i n a l blocks may be 1 x 1 matrices [XI.) Thnt is, AZ A has
X's along the diionrd; below the diagonal are 1's and 0's; all other entries are 0.
The blocks making up XI + A are c d l d eknienlary Jordan mafrius, or clmunkry
X4IwIC8. A matrix of the form (1) is cdled a Jmdun matrir belongitlg to X, or briefly,
a Jardon Adlock.
..
Qnaider next an operator T: E + E whose distinct eigenvduea are XI, . , L;
na usual E is complex if some cigenvalue is nonreal. Then E = El e ... e E,,
-1 -1 -1 0 where E, is the generalized A,eigenspacc, I; = I , . . . , m. We know thst T I EI =
Akl + Nk with Nk nilpotcnt. We give El a basis a,,,,which givea TI Ek a Jordan
.
matrix belonging to A,. The basis 63 = @I U . . U @t of E gives T a matrix of the
Q4. Jordan and R e a l Cenoniul Forms form
tion of T to the generalized h-eigenspace. Thua M must be the matrix we con- of diagonal blocks of the fnm
structed, perhaps with the X-blocks rearranged.
I t is easy to prove that similar operators have the Bame Jordan forma (perhaps
with rearranged A-blocks). For if P T P 1 = TI, then P maps each generaliaed
X-eigenspace of T. isomorphicdly onto the generalized beigenapace of TI; hence
the Jordan A-blocksare the same for TOand TI.
In summary :
Except for the order of these blocks, tbe matrix is uniquely determined by T.
D =
br a bmis for E(p, Tc), giving Tc IE(#, Tc) a Jordan matrix belonging to p.
In Section 2 we BBW that
Theorem2 I ~ L T : E - - r E b r o n o p c r d m m a r c o l u c e l mT~h. m E k r s a b o a i s
i s E . It is easy to see that in tbis basis, T 1 E.
nrc n l ~ t ~ s for has a matrix composed qiving T a mafrir compated of diagonal blocks of 1/18 forms ( 1 ) and ( 2 ) . The dioparcll
130 6. LINEAR SYBTEWB AND CANONICAL mRm OF OPERATORB
I)rlinitir,nn The matrix drscribcd in the thcorem ia called the real canonical form
of 1. 11 7' lu~sonly r ~ n rigcnvalue8,
l it is thc samp as the Jordan form. If T is nil-
pr~t<.~lt.it I S the same HS the canonical form discueeed earlier for nilpotent operatom.
is dim Kc.r(T - A ) . The number of blocks of fhefonn (2) ia dim Ker (Tc ( a a)). - +
Thr real canonical form of an operator T exhibits tbe eigenvalues as p u t of a in the real Jordan form of T; and
matrix for T. This ties them to T much more directly than h e i r definition as roots
of t111, rh~~ractcristic
polynomial. For example, it is easy to provc: 6t (A) = dim Ker(Tc - X)
as 8 compkx vector space. One obtains:
3
'Ihcv,rc*~n Let A,. . . . , A. be the e i g e n d u e a (with multiplicities) o j an opalor T.
l'l,, I, Thmrem 4 Let T be an operator on a real trdtmensirmal ucdm spau. Then he
(H) Tr(T) = XI + . . . + A,, real Jordan form o j T ia w i n e d by the j W n g equalions:
The rharacteristic polynomial ia 10. Show that the number of real logarithm of an operator on Rmis eitber 0, 1,
(t - (1 + t)(l - (1 - i))(l - 2)'. or countably idbite.
-
Thc ~igrnvaluesare thus 1 + i, 1 i, 2, 2. Biice 1 + i has multiplicity 1, there
can only be one block [: -:I.
A computation ahom W. O n o n i d Form. and Dieennttl -do-
&(2) = 1.
After a long algebraic digression we return t~ the differentia equation
This is p r o v d most easily by showing that rank ( T - 2)
[: i :I.
= 3. Hence there is only
one elementary 2-block. The real canonical form is thus: (1) ?=Az, AEL(R-).
Suppoee A is Jordaa A-block, E R:
Thr-rr rrmaina the problem of finding a basis that p u b an operator in real canon-
ical form. An algorithm can be derived from the procedure in Appendix 111 for
putting nilpotent operators in canonical form. We shall have no need for it, however.
h m the decornuonition
PROBLEMS
I.
"1 [
0 1'1
-1 0
'b) ' -;I
Firrd the Jordan f o m of the following operators on C":
"' ['ii I
1+i
2. Find the real canonical forms of the operators in Problem 1, Section 2.
3. Find the real canonical f o m of operntora in Problem 4, Section 2.
L 01
fiad by the exponentid m ~ t h a d(Chapter 5) that the solution t o (1)+th hitkl
d u e z(0) = C € R* b
+ +
(7) If X = a bi, b 0, and n = 2m, then each coordinate zr(f) of m y aolution
to ( I ) 18 a linear c o m b i t i o n of the functions
elfk cas bf, eLrsin bf ; 0 <k 5 m.
Let m be the number of blocks D so that n = 2m. The aolution to (1) can be com- Consider now Fq.(1) where A is any real n X n matrix. By a suitable change
puted using exponential*. I t is easiest to consider the equation or coordinates x = P y we tramform A into real canonical form B = P A P ' . The
(3)
mhcrr z : R - z' = Bz,
C" is an unknown map and B is the complex m X m matrix
1
equation
(8) y' = By
is equivalent to (1) : every solution x(1) to (1) has the form
I([) = Pv(t),
where y (f) aolvea (8).
Equation ( 8 ) breaks up into a net of uncoupled equations, each d the form
U' = B.u,
where 8. is one of the blocks in the real canonical form B of A. Therelore the m-
Wr identify C- with Rh by the correspondence ordinates of solutionu to ( 8 ) are linear coordinates of the function deamibed in (6)
(71+ i y ~ ., . . , z- + iw-) = (21,UI,. . . , %, y-1. +
and (7), where X or a bi ia an eigenvalue of B (hence of A). The m e therefore
is true of the original equation ( 1 ) .
'I'lu, solution to (3) is formally the same as (2) with a change of notation:
Theorem 1 Lel A E L(Rm) and lei I(!) be a sdufim of z' = A2. Then d c e
d i n a t e x, (t) is a tinear cornbindion oflhefiLm'm
reu bf, tle* sin bf,
Put Cb = L, + iMk,k = 1, . . . , m, and take real and imaginary pa& of (4);
+
COB
where a bi rune through all the eigenvalues of A with b 2 0, and k and 1 run lliraugh
us in^ the identity
el"-0 = e'(cos bf + i sin bl) aU the iniqeru 0 , . . . , n - 1. Moream, for each = a + bi, k and ia n lua Lhon
one obtains the mze of the brgwl &-block in the rcal comiml form of A.
j-1 p
rj(f) c' - [Lj4cas bt - M I A sin bf], Notice that i( A has real eigenvaluea, then the functions displayed in Theorem 1
=
,
k! include theae of the form kb.
j-I p This result does not tell what the aolutions of (1) are, but i t telb us what form
y i ( f ) = 1' C - [MI+ cas bf
,kt
+ LId sin bl]; the aolutions take. The following is a typieal and very importurt application of
Thwrean 1.
w. C A N O N I C A L FORMS A N D DIFFERENTIAL EQUATIONS
T h w r e m 2 Suppose amy eigenualue of A E L(Rm)had negafk red part. Then solution to z' = At, then
l i m ~ ( t =) O
I-.
for ewry solution to I' = Ax. for all t 2 0. Find such a k and a for each of the following operaton, A :
Proof. This is an immediate consequence of Theorem 1, the inequalities
IcwbtI_<l, lsinMlI1,
and the fact that
lirn Pet
I-,
- 0 for all k if a < 0.
The converse to Theorem 2 ia easy:
2. Let A E L ( R m ) Suppose
. all mlutiona of z' = A z are periodic with the lume
T h e o r e m 3 I f m r y a d u t i o n o f z ' = AztcndstoOast-+w,Lhenamyeigmvalue period. Then A is semisimple and the charactektic p o l y n d is a power d
u j :I has rwgalive red parl. 1'+ a a C R.
,
'
Proof. +
Suppose r = a 1% ia an eigenvalue with a > 0. From ( 5 ) we obtain 3. Suppose at least one eigenvalue of A E L ( R m )has pwitive red part. Prove
a solution (in suitable coordinates) that for any a E Rm,t > 0 there ia a solution z(1) to I' = A t mch tlut
zl ( 1 ) = 1'eos M , It(0)-al<r and liIt(t)l=m.
I--
y ( t ) = 1'sin bt,
4. Ld A € L ( R m )and , mppose all eigenvaluea of A have nonpoeitive red p-.
zfU) = Y I V ) = 0, j 2 1,
( a ) If A is semisimple, ahow that every solution of z' = Ax ia bounded (that
which doea not tend to zero as t + m . is, there is a constant M, depending on t ( O ) , such h t It(t)I S for
all 1 E R ) .
A n arpument similar to the proof of Theorem 2 shows: ( b ) Show hy example that if A ia not eemisiiple, there may ex& a solution
such that
'Um 1 r(1)I = w .
T h w r r m 4 If every eigenualue of A E L ( R n )haa posilive real parl, then a*.
Q6. Nipher Order Linear Equations A matrix of thin form in ealled the companion mdrir of the polynomid
(4) p(A) = A' +
a1X-I + ... +
%-,A Q,.+
Consider the onedimenaional, nth order homogeneous linear differential equation I n Chapter 5 it was shown that this in the characteristic polynomial of A.
with constant coefficienta Companion matrices have ~pecialproperties as operators. The key to mlvkg (1)
in the following fact.
-
Here s: R R in an unknown function, al, . . . , Q, are comlsntd, and )'(x
kth derivative of 8.
means the
Pmpositlon 2 Let A E C be a red or cmplez e i g d u e of a companion mabiz
A, Then Ule red alnunicolform of A ha8 only one A-block.
Proof. We eoneider A as an operator on C'. The number of X b l o e b M
Proposition 1 (a) A linear combinufion of sdutions of (1) is again a sdution.
I 1)) The derivative of a solution of ( I ) w again a sdution. dim Ker(A - A),
.
Frooj. By a linear cornbindion of functions fl, .. , f,, having a common do- considering Ker(A - A) as a complex vector apace.
main, and whose values are in a common vector apace, we mem a function of the Thefirstncolumnnof A - Aformthe (n -
1) x n m t r i x
fonn
f ( r ) = &(z) + +
.* . c ~ - ( z ) ,
. .
where c,, . . , G are conslank.Thus (a) means that if rl(t), . . , &(I) are solutions
of (1) and cl, . . . , G are constanta, then ca,(t) .- +. +
c,s, (t) is also a aolution;
thin follows from kinrarity of derivatives.
I'art ( h ) is immediate by differentiating both idea of (1)-provided we know
that 8 soll~tionis n f 1 times differentiable! This is in fact true. To prove it, con-
siclr.r thv ~quivalentlinear system
w h i c h h a s r a n k n - 1.HenceA - Ahasranknorn- 1,butrankninnrledout
aince A is an eigenvalue. Hence A - A has rank n - 1, no Ker(A - A) hrs E' .
I. Thia provea Propoaition 2.
..
Definition A bauia of mlutions to (1) in a net of solutions a,, . , r , auch that
If s is a solution to ( I ) , then every solution is "xprexible ss a linear combination of 81, . . . , b in one .adonly
one way.
z = (a, a', . . . , d-")
in a solution to (1). From Theorem 4, Sectisn 1 we know that every solution to The following theorem is the basic reault of this section.
(2) has derivatives of all orders.
The matrix of coefficienta of the linear system (2) in the n X n matrix Theorem The following n funclias form a bauia for the sdutimrs of (1) :
(a) the funclion reU, where A ncns Ihrarph the dislincl rml roo& of the k
feridfic pdpomial (4), and k i s anonnegatiw i n k w r i n the range0 5 k <
mdtiplicily of A; logcUIcr toilh
(h) thefuncrimrs
re1COE bf and Petsin bl,
whma+binrnsU1r~hhcomplezr&of(4) lraMnqb>Oandkua
nonnegative in@ i n the range 0 I k < multiplici& of a bi. +
140 6. L ~ N E A R SYSTEMS AND CANONICAL FORMS 01OPERA'FORB
06. HIGHER ORDER LINEAR EQUATIONS
Proof. We call the functions listed in the proposition basic functions. I t follows
To find a solution with given initial conditions, say,
from Theorem 1 of the previous section that every solution is a linear combination
of basic functions. (8) a(0) = 0,
Tbe proof that each basic function is in fact a solution in given in the next section. s'(0) = -1,
By Proposition 1 it followsthat the solutions to (1) are exactly the linear comhina-
tinns of hasic functions. ~ " ' ( 0 ) = -4,
It n.n~ninsto prove tbat each solution is a unipuc linear combination of basic a"' (0) = 14,
f~~n~-ti<llrr.lor thia wr first note that tbrre are precisely n function8 listed in (a)
nntl (1,) . t l ~ nurnhvr
r nt functions liated q u a l a the aum of the multiplicities of the we compute the left-hand side of (8) from ('I),
to get:
rr.111roots of p(A), plus tuice the sum of the multiplicities or the complex roota with (9) 40) = A +C -0,
positive imaginary parts. Since nonreal m t a come in conjugate pairs, this tatal
sf(0)=-2A+ B +D=-1,
is the sum of the multiplicities of all the roota, which is n.
.
Define a map q : R" -+Rna# follows. Let fi, . . , f. be an ordering of the basic a(*'(O)= 4A- 4B-C =-4,
functions. For each a = (a,, . .. , a.) E Rnlet s.(l) be the mlution ~ ( ~ ' ( 0=) -8A + 128 - D = 14.
are
4. What is the smallest integer n >0 for which there is a different. equation
Chapter 7 -
does not change with time. Mathematically, this means that the eonstrnt map
t 2 is a solution to (1) ; equivalently, f(2) = 0. Hence we define an aguilb
rium of (1) to be a point 5 E U such that f ( f ) = 0.
From a physical point of view only equilibria that are "stable" sre of interest.
Contractions and Generic Properties A pendulum balanced upright is in equilibrium, but this is very unlikely to occur;
moreover, tbe slightest diiturbance will completely dte tbe pendulum's behavior.
of Operators Such an equilibrium is urntable. On the other hand, the downward rest poeition ia
stable; if slightly perturbed from it, the pendulum will swing around i t rod (bemuse
of friction) gradually approach it again.
Stability is studied in detail in Chapter 9. Here we restrict a t t e n t i e to linear
systems and concentrate on the simplest and most important type d stable
equilibrium.
Consider a linear equation
(2) z' = Az, A E L (R") .
The origin 0 E R' in e d e d a sink if all the eigenvaluea of A have n-tive red
parts. We also nay the linear flow c'" is a cxmiractia.
In Chapter 6, Theorem 2 and 3, Seetion 5, it was shown that 0 in a sink if and
In this chapter we study some important kinds of linear flown elA, particularly
only if every trnjectory tends to 0 a8 t 4 m . (This is ealled aqmpblie stability.)
contractions. A (Linear) contraction in cbaracterbed by the property that every
From Problem 1, Section 5 of that chapter, it follown that trajectorieg approaeh
.
trajectory tends to 0 a8 t --r m Equivalently, the eigenvalues of A have negative
a sink ezpmtiolly. The fdlowing r d t makes this more preck.
real parts. Sucb f i m form the banh for the study of asymptotic stability in Chapter
9. Contractions and their extreme oppcaitee, expamiow, are studied in Seetion 1.
Section 2 is devoted to hyperbolic ftows el*, cbaractericed by the condition that
the eigenvalues of A have noluero real psrts. Such a flow in the direct aum of a
contraction and an expnaion. Thus their q d t a t i v e behavior is very simple.
In Section 3 we introduce the notion of a generic property of operators on Rm; (a) The migin is a sink for h& d y m m i c d syalem 2' = A+.
(b) ForanynmminELhcreareumslan&k>O,b>Oareh1M
t h i ~means that the get of operators which have that property contairna dense
ope11 subset of L(Rq). I t is shown that "semisimple" is a generic property, and
also, "generating hyperbolic flows" is a generic property for operators.
Tbe concept of a generic property of operators in a mathematical way of making f O T d t > O , l ! E E.
precise the idea of "ahnost all" operutore, or of a "typical" operator. This point is (c) There e n i k b > 0 and a bomb (B of E w k unseaponding twrm did*
discussed in Seetion 4.
01. S i n b and So- Proof. Clenrly, (c) implies (b) by equivalence of norms; and (b) implies (a)
by Theorem 3 of Cbapter 6, Section 5. That (a) implies (b) f o U m easily from
Theorem 1 of that section; the detaile are left to the reader.
Suppose that n state of aome "physieal" (or mechanical, biological, economic, I t remains to prove that (a) implies (c). For this we use the following purely
etc.) system in determined by the d u e s of n psrametem; the apace of all states
algebraic fact, wboee pmof is postponed.
is taken to be an open set U C Itg.We m p p that the dynamic behavior of the
Recall thbt R A is the real psrt of A.
system is modeled mathematically by the solution curves of a differential equation
(or d y n m i c d sya&m)
(1) a! -
f(z), f: U + R n .
We are interested in the long-run behavior of trajectories (that is, solution curves)
Lemma Let AbcanoperdmrnarcalucclmspoccEandarpporc
14(j 7. CONTRACTIONB A N D GENERIC PROPERTIES OF OPERATURB #I. EINM AND W W W ~ 147
for every etgenvalue X of A. Thcn E has a bimti arch lhuf i n he mespading i n w Let c E R be euch thst
product and n m , Rx<c<@
for every eigenvalue A of A.
(4) a l z l ' 5 (Azlz) 5 B I z 1 1 Suppoee first that A ia semislnple. Then R' has a direct sum demmpoitian
lor all r C E.
Assr~rni~~rthe truth of the lemma, we derive an estimate for solutions of I' = Az. where each Bi in a onedirnenaiond aubepsce spanned by an eigenv&tor e, of A
I t , . . , r.) be coordinates on E corresponding to a basin E? eucb that (4) comeqanding to a real eigenvalue X i ; and each FA is a twodimendonrl eubqmce
holds, and Irt invariant under A, having a bsais [fj, oil giving A ( F Athe matrix
be a solution to z' = Az. Then for tbe norm and inner product defined by E? we bave where a* + rb, is an eigenvslue of A. By .asumption
Given Rmthe inner product defined by
--CCC ziz;
(zi)'Pn' and all other inner products among the q,f,, and gb being 0. Then r m p u t a t i o n
Hence rhows
d
,I .I =
(2 2') - (2, A+)
IzI Izl '
it follows easily that
CI2,Z) 5 c I 2 It
Thcrcfore, from (4), we have for all z E Ra,IW required.
Now let A be any opentor. We fvet give R' a bssiseo that A has a matrix in red
c a n o n i d form
...
A = diagfA~, , A,I,
wbere each A, has the form
5 1% l t ( l ) l - 1% 1 z(O)l 5 Bl;
hence
If we give a rmbapace El of E, conmpondiag to a block A, a basis aatialying the Therefore if t is sufficiently m a l l , the basis a. satisfies the lemma for a block (5).
lemma for Aj, then all these bases together fulfill the lemma for A. Therefore we The ease of a block (6) is similar and is left to the reader. This completee the proof
may m u m e A is a single block.
1"or tbe fvst kind of block ( 5 ) , we cm writa A
a,l and N hse the matrix
-
S N where S hrs the matrix +
of the lemma.
The qualitative behavior of a flow near a sink haa a simple geometrid inter-
pretation. Suppose 0 E R' ia a sink for the linear diflerential equation i = j(r).
The proof is like that of Theorem 1, using the lemma and the first inequality of After contraetions and expansions, hyprrbolie linear flows have the simplest
(4). types of phase portraits. Their importance stems from the fact that almost every
linear flow is hyperbolic. This will be made precise, and proved, in the next section.
The following theorem my8 that a hyperbolic flow is the direct sum of a contrac-
tion and an expansion.
(b) Show that the corresponding first order system has a sink a t (0, 0).
(c) What do you conclude about the long-run behavior of this physical for all 1 2 0. Hence I z I = 0. This shows that F* C B. The same argument shows
system?
that E' C Fe; hence E' = F*. Similar reasoning about e l A shows that En = Fu.
i.I f cfA is a contraction (expansion), &.ow that el+*' is an expansion (respec- This comptetes the proof.
tively, contraction). Therefore a contraction is characterized by every trajeo
tory going to a as L -+ - ; and an expawion, by every trajectory going to A hyperbolic flow may h a contraction (E" = 0) or an expansion (E* = 0).
Oast-+-00. If neither E' nor Emis 0, the phase portrait may look like Fig. A in t h e twodimen-
sional case or like Fig. B in a three-dimensional case.
I
If, in addition, the eigenvalues of A E. have nonzero imaginary part, all tra-
v2. I ~ l l w r b o l iFlows
~ jectories will spiral toward Eu (Fig. C).
Other three-dimensional phase portraits are obtained by reversing the arrows in
Figs. B and C.
.\ typv of linrar flov elA that is more general than contractions and expansions is
The letter8 s and u stand for alable and unslable. E' and E' are sometimes called
thr hypcrbolzcjour: all eigenvalues of A have nonzero real part. the stable and unslaMe subspaces of the hyperbolic flow.
7. COHTRACI'IONB AND GENERIC PROPERTIES OP OPERATORS $3, OENrrUC P R O P M R m OF O-TDm
PROBLEMS
Let F be a normed vector apace (Chapter 5). Recall that a set X F ia opcn
if whenever z E X there ia mn open bdl about z contained in X; that is, for some
o > 0 (depending on r) the open bdl about t of radiun o ,
IY E ~ ~ ( u - <za ll ,
PROBLEMS
This i)rovt*nthat 81 is dense.
To provp that S1 is open we argue by contradiction. If it is not open, then there
is a sequence A,, A t , . . . of operators on Rg that are not in 81 but which converges 1. Each of the following properties defines a set of real n X n matriees. Ftod out
to an uprrator A in 8,. There is a n upper bound for the norms of the AI and hence which arts are dense, and which are open in the apace L(Rm)of d linear opera-
for their eigenvalues. By assumption eaeh A, has an eigenvalue Ak of multiplicity tors on R":
a t least two. (a) detetminant # 0 ;
At -
Suppose a t first that d l XI, are real. Passing to a subeequence wemay assume that
A E 8,. For each k, there are two independent eigenvectors zk, y, for A t be-
longing to the eigenvalue Ah. We may clearly suppose I s I = I yr I = 1. Moreover
(b) trace is rational;
(c) entries are not integers;
(d) 3 5 d e t e r m i m t < 4;
(e) - l < ( A I < l f o r e v e r y e i g e n v a l u e ~ ;
we may assume zr and y h orthogonal, otherwise replacing by
(f) no real eigenvalues;
(g) each real eigenvalue has multiplicity one.
Passing again to uubsequencen we may assume -t z and y, -,y. Then z and y 2. Which of the following propertien of operators on R- are generic?
are independent vectors. Fmm the relations A b z k = h z t and Atyt = XI# we find (a) I A I # 1 for every eigmvalue A;
in the limit that A z = hz and Ay = Ay. But thia contradicts A E 81. (b) n = 2 ; some eigenvalue is not real ;
If somp of the h are nonreal, the eame contradiction is reached by considering (c) n = 3; some eigenvalue is not real;
the complrxifications of the A,; now zk and yt are vectora in C". In place of the ( d ) no solution of z' = A z is periodic (except the zero solution) ;
Euclidean inner product on R" we use the Hemition inner product on Cmdefined (e) there an! n distinct eigenvalues. with distinct imaginary parts;
by (1, w ) = x7-I Z,I@~, and the corresponding norm I z I = (2, r)"'. The rest of the ( f ) A z f z a n d A z # - z f o r a l I z f 0.
argument is formally the same aa before.
3. The set of operators on Rmthat generate contractions is open, but not dense, in
Note that the operatom in SI are all seminimple, by Chapter 4. Therefore an L (R"). Likewise for expansions.
irnnrdinte consequence of Theorem 1 is
155 7. CONTRACTIONS A N D GENERIC PROPERTIES O F OPERATUR8
(.'
S o t ? that th? definition implies that the map 6 , :O * O is C' for each t and has a
invrrsr .$-, ( t a k e s = - 1 in ( b ) ) .
.4n example of a dynamical system is impticitly and approximately defined by -
Throughout the rest of this chapter, E will denote a vector space with a norm;
W C E, an open set in E ; and j: W E a continuou map. By a d u l h d the
-
trp differential equations in the Newton-Kepler chapter. However, we give a pre- differential equation
cise example as foUows.
Lct A be an operator on a vector apace E ; let E = d and 6 : R X d s b e de-
, = elAz.Thus 6 1 :S d s can be represented by 6 , = e l A . Clearly,
f i n d by ~ ( 1 z) we mean a differentiable function
+O = 8 = the identity operator and eince ecl+*" = c'*e'*, we have defined a dy-
namical system on E (see Chapter 5 ) .
This example of a dynamical system is related ta the differential equation dz/dt =
defined on some interval J C R such that for dl t E J
A r on E. A dynamical ~ y s t e m6 , on s i n general gives riee to a differential equation u'(9 = f ( u ( O ) .
on S, that is, a vector field o n 8, f : s -.E . Here S is suppoaed to be an open net in Here J could be an interval of real numbera which in open, e l d , or brlf open, Iulf
the v ~ c t o space
r E. Given +,, define f by closed. That ia,
(a,b) = I t € R I a c t c b l .
-
means that the function f doen not depend on time. One can alao consider a C
map f : I x W E where I is an interval and W is an open set in the vector space.
The equation in that ease is
rind is r111lr.d nonautonomous. The existence and uniqueness theory for (I1) will
8. F U N D A M E N T A L THEORY
u. EXISTENCE A N D UNIQUENE-
solution
z' = f ( 4
satisfying the initial condition
z(0) = 4.
53. Existence a n d V ~ q u e n e s s
-
FIG. B
A function f : W +E, W an open w t of the norrned vector apace E , is aaid to be
An initial condition for a solution u :J W is a condition of the form u ( b ) = Q Lipschitz on W if there exists a constant K such that
where b E J, € W. For simplicity, we usually take = 0 .
A differential equation might have several solutions with a given initial condition.
For example, consider the equation in R, for all z, y in W . We call K a Lipschifz constant for f .
We have aasumed a norm for E. In a different norm f will stffl be Lipsehitz be
2' = 391a. causeof the equivalence of norms (Chapter 5) ;the constant K m y change, however.
Hrre 1V = R = E , f : R +R is given by f ( + ) = 39'. More generally, we call f locoUy Lipsdrila if each point of W (the domain of f )
The identically zero function uo: R 4 R given by % ( t ) = 0 for all t is evidently has a neighborhood W. in W such that the restriction f I W Ois Lipschita. The Lip
a solution with initial condition u ( 0 ) = 0.But so is the function defined by z(L) = schitz constant off I Womay vary with WO.
1" The graphs of some solution curves are shown in Fig. B.
Thus it is clear that to ensure unique solutions, extra conditions must be imposed
on the function f . That f be continuously differentiable, turns out to be sufficient,
Lemma Let Ihe function f : W - E be C. Then j is i d l y Lip&&.
as we shall see. Thus the phenomenon of nonuniqueness of sohtions with given Before giving the proof we recall the meaning of the derivative D f ( r ) for x € W.
initial conditions is quite exceptional and rarely arises in practice. This is a linear operator on E ; i t aasigna to a vector u € E, the vector
In addition ta uniqueness of solutions there is the question of existence. Up to
1
this point, we have been able to compute solutions explicitly. Often, however, this
is not possible, and in fact it is not a priori obvious that an arbitrary differential
D f ( z ) u = lim - ( f ( z
.4 8
+ ar) - / ( z ) ) , a € R,
equation has any solutions a t all. which will exist if D f ( z ) is defined.
We do not give an example of a differential equation without a solution because .
In coordinates ( I , , . .,r.)on E, let f ( z ) = (f,(11, . . ., c.), . .., f . ( z ~ , ..., I . ) ) ;
in fact ( I ) has a solution for all initial conditions provided f is continuous. We then D f ( z ) is repres~ntedby t h e n X n matrix of partial derivatives
shall not prove this; instead we give an easier proof under hypotheses that also
gunrantrc uniqueness.
Thr following is the fundamental loesl theorem of ordinary differential equationi. Convrrscly, if all the partial derivatives rxist and are continuous, then f is CL.For
I t is called a "local" theorem bemuse it de& with the nature of the v e c h field each r: E W, therc is defined the oprrator norm I( D f ( r ) 11 of the linw operator
f : W 4 E near some point Q of W. D f ( z ) E L ( E ) (ace Chapter 5 ) . If u E E, then
Xext, we prove that there is a constant L 2 c such that for all k 2 0: (by uniform convergence)
I w + I ( ~ - u ( l ) l 5 (Ka1.L.
Put L = max(1 u ~ ( t-
) %(f)l:I I I 5 a ) . We have
= *+ l f ( Z ( 8 ) ) &
5 KLI -w-I(~)I~.
< (aK) ( a K ) h 1 L= (aK)'L.
Therefore we aee that, putting aK =a < 1 , for 8nY r >8>N
5 aKQ.
Since aK < I , this is impossible unless Q = 0. Thus
for any prescribed e > 0 provided N is large enough. Another proof of uniqueness followa from the le- of the next seetion.
By the lemma from analysis, this & o w that the eequence of functions uo, ul, ... We have proved Theorem 1 of Section 2. Note that in the c o w of the pmof
converges uniformly to a continuous function z : J 4 B. From the identity the following was shown: Given m y ball We C W of radius b &out with
max.cw, I f ( z ) J5 M , where f on Wo has Lipschit. constant K and 0 < a <
mini b l M , l / K 1, then there is a unique solution z: (-a, a ) 4 W of ( 3 ) such that
%.(I) = + je f(l(t(8)) da,
z ( 0 ) = I+.
we find by taking limits of both Bides that Some r e m a r b are in order.
Conaider the situation in Theorem 1 with a C' map f : W -+E, W open in E.
Two sduliun c u m of z' = f ( z ) eonnd cross. This is an immediate eonseguenee of
uniqueness but is worth emphasizing geometrically. Suppose p: J -,W , $: J I -r W
are two solutions of z' = f ( z ) such that p(t,) = S(h). Then p(h) in not a c w
+
because if we let J.l(t)= J.(h - II t ) , then $1 is also a solution. Since +,(t~) =
$ ( h ) = ~ ( L L ) ,it follows that $I nnd p agree near tg by the uniquenem atstement of
Theorem 1. Thus the aituation of Fig. A is prevented. Similarly, a eolution curve and so
cannot cross itself ss in Fig. B.
FIG. A FIG. R
M. Continuity of solution^ i n Initid Gnditions
If, in fact, a solution curve cp: J -+ W of z' = f(z) satisfies ~(11)= v(C + w)
for some I, and w > 0, then that solution curve muat close up as in Fig. C.
For Theorem 1 of Section 2 to be a t all intereating in any physical sense (or even
mathematically) it needs to he complemented hy the property that the solution
r(l) depends continuously on the initial condition z(0). The next theom gives s
precise statement of thin property.
u(1) 5C + ltKU(S)&
0
1f C = 0, then apply the above argument for a sequence of positive c, that tend
to 0 ss i --+ m . This prove4 the lemma.
*
fflnce x(f) --t m as f -+c f r / 2 .
We turn to the proof of the theorem. Now consider a general equation (1) z' = f ( z ) where the C function f ia defined
Define on an open w t W C E. For each 20 E W &re is a mazimum upen intend (a. 8 )
M i n i n g 0 on which h e w a 8dU1Mn. ~ ( f with ) x(0) = 4 There ia some such
~ ( 1 )= I ~ ( 1 )- 0 ) I. interval hy Theorem 1 of Section 2; let (a, 8) be the union of aU open intervnb
Since
containing 0 on which there is a solution with z(0) = 4.(Pomibly, a = - m or
Y(O - ~ ( =0~ -
( u zio + /' [firw ) - f m )) I & . 8 = +a,, or both.) By the lemma the solutions on any two intervds in the union
agree on the i n t e m t i o n s of the two intervals. Hence there ia a solution on all of
we have (a,8).
Next, we investigate what happens to a solution as the limits of its domain are
approached. We state the result only for the right-hand limit; the other caae ia
similar.
Now apply the l m a to the function u(f) = v(b + f) to get
Theorem Let W C E b e o p e n , letf: W - + E b e a C m a p .W y(f) beard&
"(9 I V(b) exp(K(t - 4 ) ) , on a m a d open inlnvcll J = (a,8) C R wiUl 8 < m . Then piDm any wmpoct
which ia just the conclueion of the theorem. aef K C W, there w sonw f E (a, 8) mMIhy(L) 4 K.
then it leaves any compact &.
boundary of W or I y(f)\ Lnds to - {or both).
-
This theorem rays that if a solution y ( t ) cannot be extended to a brger interval,
This i m p b that aa f @ e i t k ~ ( f Unds
) Lo the Prooj. Let [0, 8) be the msldmal half-open intend on which t h e n is rr mlution
y as above. Then y([O, 8 ) ) C A, and so 8 Cannot be h i t . by the theorem.
Proof of the theorem. Suppose y (f) E K for dl f E (a,8). Sincef M continu-
ous, there &ta M > 0 such that 1 f ( z ) J _< M if z E K.
Let 7 E (a, 0). Now we prove that y extends to a continuous map [r, a ] -r E. 66. Global Solutions
By a lemma from analyeis it aufficestto prove y: J -* E uniformiy continuous. For
1, < 11 in J we have
We five here a stronger theorem on the continuity of mlutiona in terms of initid
conditions.
I n the theorem of Section 4 we assumed that both m l u h were defined on the
same interval. I n the next theorem it is not necesssry to sssume thia. The theorem
showa that enlutiona atsrting at nearby points will be d e b e d on the m e closed
interval and remain nesr to each other in thia interd.
Now the extended curve y: [a, 814 E M differentiahle a t 8. For
Theorem led f(z) be C . Let y(f) be a adulh to i = f(z) d c f i d OI Lhs M
i& [b, f,], loilh y(b) = YO. Then is a nkghborhood U C E of y, a cudad
K w h Ihd if z, E U , fhm h e is a unique udulh z(l) nbo defined on 14, f,] with
z(L) =s;andxrolyfirs
1 y(f) - r(f)I < K I yo - Y I exp{K{f - 4 ) )
for ad 1 E CL,a.
For the proof we will uee the foUowing lemma.
hence ord
Lemma I f f : W -+ E is locally Lipwhdz and A C W is a m@
bounded) a d , llrm f I A is Lipbchh.
P-f. Suppose not. Then for ewry K > 0, no matter how Isrge,we aro find
for all f between 7 and 8. Henee y M differentiahle at 8, and in fact y'{B) = f(y{p)). z and y in A with
Therefore y ia a solution on [r, 81. S i c e there ia a dution on an interval D, a), If@) - f(y)I > K I z - Y I.
b > 8, we ean extend y to the interval {a,4). Renee (a,8) could not be a maximd
domain of a solution. This completes the p m f of the theorem. In particular, we cam find z., y. 811th that
(1) If(&) - f b . ) l L n I ~ . -y.I for n = 1, 2 , . . . .
The following important fact follow8 immediately from the theorem.
Since A is compact, we ean choose convergent subsequmw of the Z. and fi
wabeliag, we may aesume I.-+ 2 and y. -r y' with 2 and y' in A. We 0km-a
Prq-~tion L c l A b e a ~ ~ o f L h s o p o l s c l W C E a dW
~ -f+:E that z* = y*, eince we have, for dl n,
beC. Lely.
Y: CO, el -
E A andsu~ilisknown~mrl(adulirmnrrucafLhsfm
w,
lied mlirdy i n A. Then thore u a dulia
-
~ ( 0 ) %,
.--
I I* - ) = lim.1 z. - y. 1 5 n-I I f(z.j - f(y.)( 5 n-'2M,
where M ia the maximum value off on A. There ia a neighborhood W. of 2 for
which f I W,has a Lipschit5 constant K in t. There ia an n.pleh that I . € We if
v:[o,-)-W, y(O)=yh and y(t)EA n 2 np. Therefore, for n 2 n.:
f w d f 20. If(z.) -f(y.)l 5 Kt?. -u-In
which contradicts (1) for n > K. Thb proves the lemma.
$7. THE F L O W O F A DIFFERENTIAL EQUAfON
The proof of the theorem now goes as follom Let n C R X W be the following set:
By campaetnees of [b h], there exieta e > 0 such that z E W if I z - y ( f ) 1 5 t.
The set of dl such points is a cotnpnet subset A of W. The C map f is
locally Lipacbits (Section 3). By the lemma, it follona that f I A has a Lipschita The map ( t , y) -+ 0(1, y ) is then a function
constant k.
I r t a > 0 be so amall that 6 $ c and b exp(k I h - b 1) 5 t . We aeaert that if +:n+ w.
1 % - y, I < 6, then there is a unique solution through 4 defined on all of [4, 111. We call + the Pow of equation ( 1).
Fit of all, s E W since 14 - y(b)l < r , so there is a solution s(1) through zo on We shall often write
a tneximal interval [b, 8 ) . We prove 6 > 4. For mppcee 8 5 4. Then by the ex- o(1, 2 ) = Ol(2).
ponrntial rtstimate in Section 4, for d l 1 € [ b , @),we have
Example. Let f ( z ) = Az, A E L(E).Then +t(z) = cUz.
Thus z(t) lies in the compact set A ; by the theorwl of Section 5, [b, 8 ) could not
be a solution domain. Therefore ~ ( 1 is) defined on [to, h ] . The exponential Proof.Fit,suppoee s and 1 are positive and +.(+,(z)) is defined. Thia m e w
estimate follons from Section 4, and the uniquenew from the lemma of Section 5. and a E J(+,(z)). Suppoee J ( z ) = (a, 8). Then a < 1 < 8 ; we 8hall
f € J(z)
We interpret the theorem in another way. Given f ( z ) as in the theorem and a show @ >s + f. Define
solution y (t) defined on [ b , t,], we 8ee that for all 4 d c i e n t l y cloee to yo = y(b), y: (u,s+1]-+W
there is a unique solution on [f., t1] starting a t a at time sero. Let us note thin
solution by t -+ u ( l , 4 ) ; thus u ( 0 , ~ =
) a,m d ~ ( fy.)
, = y(1).
Then the theorem implies:
Then PROBLEMS
1. Write out the first few terms of the Picard iteration scheme (Setion 3) for
The second term on the right goes to 0 with 6 b e c s w the solution through s is each of the following initial value problrm. Where possible, use any method
co~itinuous(even differentiable) in t. The first term on the right, by the estimate to find explicit solutions. Discuss the domain of the solution.
in Section 6, is bounded hy I S 8 which alao goen to 0 with I. This proves Theorem 2.
+
( a ) r' = t 2; t ( 0 ) = 2.
(h) 2' = 9,';t ( 0 ) = 0.
In Chaptcr 16 we shall show that in fact 4 is C.
(c) I' = x'"; r ( 0 ) = 1.
S o w suppose ( 1 , 10) E O; then 4 haa a neighborhood U C W with t x U C 0,
(d) r' = sin z ; z ( 0 ) = 0.
sinrc wr know O is open in R X W. The function z -r 4t(1) defines a map
( e ) t' = 1 / 2 7 ; z ( l ) = 1.
2. Let A be an n X n matrix. Show that the Picard method for solving z' = At,
z ( 0 ) = u gives the solution etAu.
3. Derive the Taylor aeriea for sin t hy applying the P i e d method to the f i a t
Theorem 3 The map 4 , sends U MUO an open #el V and 6 ,ti defined on V and order system corresponding to the second order initial value problem
sends V malo U.The oompodilimr &,#, ti the identity map of U ; the wmposition +,&I
is the idmlily map of V .
P r m f . If y = + , ( z ) ,then t E J ( z ) . I t M easy to see that then - 1 E J(y), 4. For each of the following functions, find a Lipschits constant on the region
for the function indicated, or prove there is none:
8-+ A - ~ ( Y ) (a) f ( r ) = 1 1 1 , - - < z < -.
(h) f ( r ) = rl/', - 1 <_ z _< 1.
is a solution on [- t, 0 ] sending 0 toy. Thus 4-1 in defined on 4 , ( U ) = V; the state- (c) f ( t )= l / z , 1 5 z 5 -.
mrnt ab~nltcompositions ia obvious. I t remaim to prove V is open. Let V* V > (d) f ( z , y ) = ( Z + ~ Y , --I), ( Z , I ) E R'.
hc. tlrr rnaximal suhset of W on which 4-, is defined. V * is open because O is open,
and 0-,: V b-r W is continuous because 4 is continuous. Therefore the inverse
image of the open set U under 4-, is open. But this inverse image is exactly V.
-
letting 1 -+ + ( ( I ) = +(t, z ) be the maximal solution curve taking 0 to a. Tbere
is an open set U ,C W on which the map 6,:U , W is defined. The maps +, satisfy
O,+,(z) = &+,(I) as in Theorem 1. Each map 6 , is a homeomorphism; that b,6,
ternal [O, 81.
(b) For what values of a are there infinitely many solutions on [O, a ] satisfy-
ing z ( 0 ) = -1?
is one-bone and has a continuous inveree; the i n v e m is 4-1.
6. Let f : E + E be continuous; suppose f ( x ) _< M . For each n = 1, 2, . . . , let
If
z,: [0, I ] -+ E be a solution to r' = f ( z ) . If ~ " ( 0converges,
) &ow that a
f(r) = At, A E L(E), subsequence of Ix.1 converges uniformly to a solution. (Hint:Look up kseoli's
then theorem in a hook on analysis.)
* , ( z ) = c"z.
7. Use Prohlem 6 to show that continuity of solutions in initial conditions fol!ows
In this case O = R x E and each 4 , is defined in all of E. from uniqueness and existence of solutions.
8. Prove the followinggeneral fact (seealso Section4) :if C 10 and u, v: [O, 81 -r apectively. Consider C maps f, g,
R are continuous and nonnegative, and
The chah rule of calculus can be stated as: the derivative of the componition ia the
composition of the derivatives. In other words, if z E U, then
then
Out treatment of calculus tends to be from the modem point of view. The deriva-
tive is viewed as a linear translomrstion.
Suppose that U is an open set of a vector Vace E and that g: U -+ F ie m e map,
F a second vector space. What is +he derivative of g a t a E UO?We my that this
derivative exista and is denoted by & ( a ) E L(E, F) if
Then, if, for each z E U, the derivative Dg(z) exists, this derivative defines a
map
some physical (or biological, economic, or the like) system described by ( I ) , then
2 is an "equilihrium state": if the 8ystem is a t 2 it alwaye will be (and alwap
was) a t 2.
Let +: n -+ W be the flow associated with (1) ; D C R X W is sn open aet, and
Stability of Equilibria for each z E W the map t -+ +(L, z) = +,(z) is the solution p w i n g through z when
I = 0;it ti defined for I in some open interval. If i is an equilihriurn, then +,(t) = i.
for all t E R. For this reaeon, t is also called a alalianary poinf, or frud point, of
Sinrr, -4 = D/(O) and f(0) = 0, by the deftnition of derivative, Remernher that the spheres are not necessarily "mud" spherea; they are apheres
in a special norm. In standard coordiites they may be ellipsoids.
A simple physical example of a nonlinear sink is given hy a pendulum moving in
a vertical plane (Fig. B). We assume a constant downward gravitational force
Therefore hy Cauchy's m e q u d t y , equal to the mass m of the boh; we neglect the mass of the rod supporting the
lii
U(z) - Az,z) = 0,
hoh. We =ume there is a frictional (or viscous) force resisting the motion, pro-
portional to the speed of the boh.
rO IZI' Let 1 he the (constant) length of the rod. The boh of the pendulum moves along
I t f o l l o ~ sthat there exists I > 0 so m s l l that if I z I 5 6, then z E W and a cirrlr of radiw I. If B(f) is the countrrclockwiw angle from the vertical to the
rod a t timr f , then the angular vclocity of thr bob is @/dl and the veloeity is
U(z), z ) 5 - C I z 1.'
1 &/dl. Therefore the frictional force is -kl &/dl, k a nonnegative constant; this
1'utrr= IrE~"~/r(I6).Letz(f),O<f<b,bea~olutioncuweinU, force is tangent to the circle.
~ ( #0 0. Then The downward gravitational force m has component -m sin B(f) tangent to the
circle; this is the force on the boh that produces motion. Therefore the total force
tangent to the circle at time f is
This shows, first, that I z(f) I is d e e m i n g ; hence I r ( f ) I E U for all f E [O, 41.
S i c r U is compact, it follows from Section 5, Chapter 8 that the trajectory z(f)
is defined and in U for all f 1 0. Secondly, (2) implies that
hence, from Newton's law a = F/m, we have
I W ) I 5 e-'* I Z(0) I
for all 1 2 0. Thus (a) and (h) are proved and (c) follows from equivalence of
norms.
The phase portrait. a t a nonlinear sink Z looks like that of the linear part of the
vector field: in a suitable norm the trajectories point inside all uufficiently s d
epheres about f (Fig. A). P'= k 1
sin e.
m
Introducing a new variable
w = P
{interpretd as angular velocity), we o h t i n the equivalent first order system hut its motion is too m a l l to observe. A better explanation is that the msthematical
'6 = 0,
model (3) of its motion is only an apgroximation to reality.
(3)
PROBLEMS
This aonlinrar, autonomous equation in R' has equilihria a t the points
1. (a) State and prove a converse to the theorem of Section 1.
(h) Define "sources" for nonlinear vector fields and prove an interesting
theorem about them.
We concentrate on the equilihrium (0,O).
The vector field defining (3) is 2. Show hy example that if f is a nonlinear C vector field and f(0) = 0, i t in
possible that lim,-, z ( t ) = 0 for all solutions to z' = f (z), without the eigen-
values of Df(O) having negative red park-.
3. h u m e f is a C vector field on R ' and f(0) = 0. Suppme m e eigenvalue of
Its derivative a t (8, w) is Df(0) has positive real part. Show that in every neighborhood of 0 there is a
solution z(f) for which ( z(f) I ie increasing on some interval [O, 43, 4 > 0.
4. If Z L a sink of a dynamical aystem, i t has a neighborhood containing no other
equilibrium.
Hence
$2. Stability
equation
(2)
where A has pure imaginary eigenvalues. The orbits are all ellipses (Fx.D).
L
A. FIG. Stability.
Sin~ilnrly,for any b > 0 there exists a Euclidean norm on El such that Furthermore, if (2,u) = z E U,thenDg(z)(f(z)) = D d z , u)U~(z,u),ft(z,u ) ) =
(z,f,(z, I) ) - (Ill, fa(*, Y))which will be positive if r E ~ ~ (by0 (a).
) Thia implies
(4) (By, t ~ )< b I u Is, all u E El. that on a solution z(L) in U psasing t,hrough the boundmy of C, g is i w a ' n q
since by the chain rule, (d/dt)(g(z(L)) = Dg(z(l))f(z(L)). Therefore M dulion
W r choose b so that
which atarb in C can lmue C bejore ii leaves U. F i e E gives the idea.
0 < b < a. Geometrically (b) implies tbst each vector j(r) a t z E C points outward fmm
We take the inner product on E = El e E; to be the direct sum of these inner the ephere about 0 passing through z. See Fig.F.
products on E l and E;; we also use the n o m associated to these inner products
on El, El, E. If z = (I, y) € E Le El, then I z I = (1 z I' I u Is)'". +
We shall use the Taylor expamion off around 0:
Thus, given any r > 0, there ex* i > 0 such that if U = B1(O) (the hall of
radius 6 about 0),
(5) I Q(z) I I r l z I for 2 E U.
FIG. F
BO (b) implies
Thus each nontrivial solution z(1) starting in C fl U moves away from 0 a t an types of dynamical systems (the gradient systenrs of Section 4), almast every
exponential rate as long as it is defined and in C n U. state b in the basin of some sink; other states are "improhable" (they constitute
If y (1) is not defined for all 1 2 0, then, by Chapter 8, Section 5, i t must leave a set of measure 0). For such a system, thr s i n h represent the different types of
the compact set C n U ;as we have seen above, it must therefore Leave U.On the long term behavior.
other hand, if y(i) is defined for all 1, i t muat also leave U since U is the ball of I t is often a matter of practical importnncr to determine the basin of a sink f.
radius 6 and e1I z(0) I > 6 for large t. Therefore there are solutions starting arbi- For cxamplr, suppoar L repre8rnts sonlr dcwirtd equilibrium state of a physical
trarily close to 0 and leaving U.Thus (easuming the truth of the lemma), the system. The extent of the basin tulla us how large a perturhation fmm equilibrium
vector field f does not have 0 as a point of stable equilibrium. we can allow and still be sure that the eystpm will return to equilibriun~.
We now give the proof of the lemma. First, part (b) : if (I, y) = z E C n U , We conclude thin section by remarking that James Clerk .\laxwell applied
stability theory to the study of the rings of the planet Saturn. He decided that
they must be composed of many small separate bodies, rather than being solid or
so, by (3), (4), ( 5 ) : fluid, for only in the former case arr there sMle solutions of the equations of mo-
u(z), z) 2 a I 1P - b I u I' - e I 2 1.' tion. He discovered that while solid or fluid ringu were mathematically p d b l e ,
the slightest perturhation would destroy their configuration.
I n C , I z \ Z I y I a n d I x l ' L t ( l x I S + I y I t ) 2 f I ~ ( ~ . T h(f(a),z)
w 2 (a/2 -
b/2 - e ) 1 z 1.' We choose t > 0 and then 6 > 0 so that a = a/2 - b/2 - t > 0.
This proves (b).
To chrck (a), note that the lefbhand side of (a) is PROBLEMS
(Ax, I) - (By, u) + (2, R(x, y)) - S(2, u)),
but
1. (a) k t e be a stable equilibrium of a dgnamieal system corresponding to a
(2, R ( r , u)) - (u, S(+, u)) 1
I <
2 1 (2, Q(z)) I. C' vector field on an open set Nr C E. Show that for every neighborhood
We may proceed just as in the previous part; finally, 6 > Ois chosen sothat a / 2 - U of Z in W, there is a neighborhood U' of i in U such that every solution
b/2 - 2t > 0. This yields the propmition. curve r(1)with z(0) f U' is defined and in U' for all 1 > 0.
(b) If Z is asymptotically stable, the neighborhood U' in (a) can he ehosen
In Chapter 7 we introduced hyperbolic linear flows. The nonlinear analogue is to have the additional property that Lim,,, r(1) = i if z(0) t U'.
a hyperbolic equilibrium point 3 of a dynamical system t' = f ( r ) ; and to repeat, ( H z I I ~ :Consider the set of all points of U whose trajectories for 1 2 0 enter
thi nwans that the eigenvalues of Df(i) have nonrero real parts. If these real psrte the set U,in Definition 1 or 2.)
art. all nrgative, i is, of course, a sink; if they are all positive, 3 is called a smtrce. If
both signs occur, t ia a saddle point. From the theorem we see that a 2. For which of the following linear operators A on R mis 0 E R. a stable equi-
saddle poinl is undabb. librium of z' = Ax?
If f is an asymptotic equilibrium of a dymmical aystem, by definition there is
a neighborhood N of i such that any solution curve starting in N tends toward 2.
The union of all eolu tion curves that tend toward z (as I + m ) b d e d the badin
of i,denoted by B(3).
I t in clear that any solution curve which meets N is in B(2); and, eonveraely,
any 8Olution curve in B ( i ) must meet N . I t follors that B ( Z ) is an opm a l ; for,
by continuity of the flow, if the trajectory of I meeta N, the trajectory of any
nearby point alm meets N.
Sotir.1. that B ( f ) and B(g) are disjoint if 2 and y are different asymptotically
stsblr eq~~iEbria. For if a trajectory tends toward 2, it cannot also tend toward Y.
If a dynamics1 system represents a system, one can practically identify 3. Let A be a linear operator on Rn all of whose eigenvalues have real part 0.
the states in B ( i ) with 2. For every state in B(2) will, after a period of transition, Then 0 t R- is a stable equilibrium of z' = Ax if and only if A is semisimple;
stay 80 close to 2 as to be indistinguishable fmm it. For some frequently occurrins and 0 is never asymptotically stable.
192 9. BTABILITY OF EQUILIBRIA
4. Show that the dynamical aystem in R*,where equations in polar awrdin8t.m by the chain rule. Consequently, if P(z) is negative, then V dec- along th
are aolution of (1) through z.
9b 1 t > 0, We can now state Lbpunov's stability theorem:
#=I, .rf-1
0, r = 0, Theorem 1 k t Z E W & an quiltbiurn jor (1). Lct V: U -+ R & a calinuow
has a stable equilibrium a t the origin. (Hint: Every neighborhood of the fundion dejned on a neighborhood U C W of x, differed& an U - d, meh lhd
origin contains a solution curve encircling the origin.)
Let f : R" + R- be C and suppoee f(0) = 0. If mme e i g e n d u e of Dj(0) hss
(a) V(Z) - Oand V(z)
(b) V < 0 i n U - i.
> Oifz + i;
5.
positive real part, there is a nonzero solution z(t), - m < 1 5 0, to z' = f ( d , Thm d it sloblc. Furlhermare, if &o
such :hat lim,,, z(t) = 0. (Hint: Use the instability theorem of Section 3 to
find a sequence of solutions GO), L I t 5 0, in BdO) with I z.(O) I = 6 and (c) v<OinU-Z,
lirn..-- rm(k)= 0 . ) h 3 it ~ p I d d .&ilk.
y
6. Let g : R m 4 Rn be C' and auppoee f (0) = 0. If some eigenvalue of Dg(0) hss
negative real part, there ia a solution g(t), 0 5 t < m , tO Z' = g ( ~ )E, U C that
~
A function V antisfying (a) and (b) is &led a L i o p u w f i l i d h for 2. If (c)
also holtls. WP rwll V it stnct L I ~ ~ I J Uftilbrt~ot~.
I ~ O I . TIIC01i1y cq(~illl)riil~n
I?; tlie origin
lime-, g(t) = 0. (Hint: Compare previous pmblem.)
r = y =O.
We emphmiie that Liapunov's theorem can be applied without solving th
differential equation. On the other hand, there is no cut-and-dried method of
$3. Liapunov Function. finding Liapunov functions; i t is a matter of ingenuity and trid and e m r in each
ease. Sometimes there are natural functions to try. I n the case of mechanical or
electrical systems, energy is often a Lkpunov function.
In Section 2 we defined stability and asymptotic atability of a n equilibrium d
Exampfe 1 Conaider the dynamic$ sptem on R' described by the system of
of a dynamic81 system
differential equations
2' = 2y(z - I ) ,
where f : W + R" is a C map on an open set W C R". If 2 is a sink, atab'ility can
be detected by examining the eigendues of the linear part Df(2). Other than that,
however, as yet we have no way of determining &ability except by a c t d l y hnding
all solutions to ( I ) , which may be diffrcnlt if not impossible. The z-axis ( - { (z, y, r ) [ z = y = 01) consists entirely of equilibrium p o i n b
The Russian mathematician and engineer A. M. Liapunov, in his 1892 doctoral Let us investigate the origin for &ability.
thpsis, found a very uaefnl criterion for stability. I t is a generahation of the ides The linear part of the system at (O,O, 0) is the matrix
that fnr a sink there is a norm on R*such thqt I z(l) - i! I decreases for mlutiom
I ( 0 111.arr . Liapunov showed that certain other functions could be used inntesd
r g l I lnr rlrtrrtt to g u n r a n k stability.
14.1 I' 1' -+ R bc a diflcrentiable function defined in a neighborhood U C W
of I . WI. ~lc.nclteby V : U -t R the function defined by
Them are two imaginary eigenvalues and one zero eigendue. AU we an conclude
fmm tbia is thabthe origin is not a mink.
Here the right-hand Bide is aimply the operator DV(z) applied to the vector Let us look for a Liapunov funetion for (0, 0, 0) of the form V(x, y, r ) = ai f
f ( 7 ) . Then if +,(z) is the solution to (1) though z when t = 0,
+ d,with a, b, c > 0. For mch a V,
194 9. BTABILITY Or EQUILIBRIA $3. WAPUNOV PUNCl'IONS
1, -
s u l ~ ~ t ~shrting
on
, ri~(.h
ill t', - i and supposr t(t.) -+ 20 6 Ba(f) for some sequence
a scqurncr exists by compactnem of B,(f).We assert s 2. To see-
I'!i(f.) 1-
tl115.o l ~ > ~ ~thatr r r I ' ( r ( 1 ) ) > V ( y ) for all t 2 0 since I;(r(t)) decreases and
l'(z,) by continuity of V. If 4 # t , let z(t) be the solution atarting
a! O: I'm any s > 0, w e have V (~(s)) < V(4). Hence for any solution y (a) starting
@. LIAPUNOV PUNCTIONB
PROBLEMS
$4. Gradient Systeme
where
Find 6 > 0 as large as you can such that the open disk of radiua I and center V: U + R
(0,O) ia contained in the basin of (0,O). is a C function, and
.?. Discuss the stability and basins of the equilibria of Example 1in the text.
3. A particle moves on the straight line R under the inlluence of a Ncntonim
force depending only upon the position of the particle. If the force ia d m y 8 ia the gradient vector field
directed toward 0 E R, and Mniahm a t 0, then 0 is a stable equilibrium. (Hid: grad V: U + R.
of V. (The negative sign in (1) is traditional. Note that -grad V(x) = this kemrl is the ( n - 1)dimemional subspace of vectors perpendicular to grad
gsd(-.V(d).) V(u) (translated parallelly to u). Therefore we have shown:
Gradlent syatmnu have apecia1 propertien that make their flows rather aimple.
The following equality is fundamental:
Theorem 2 At r&r poinla, Lhe uecfor W - g a d V(E) b perpendieday to Lhe
(2) DV(z)y = (grad V(L.),u). hvel mrjaces of V.
Tlris Rsye that the derivative of V a t r (which is a l i n k map Rn --r R), evaluated
on y € Rm,gives the inner product of the veetars grad V(z) and y. To prove (2), Note by (2) that the nonregular or criticd points of V sre preckly the equi-
we observe that librium points of the syatem (1).
Since the trajectories of the gradient system (1) are tangent to -grad V(r),
we have the following geometric description of the flow of a gradient syetem:
which %exactly the inner product of grad V(z) and y = (yl, . . . , y,).
Let V: U -r Rmba the derivative of V along trajectariea of ( I ) ; that is, Theorem 3 lkt
E' = -grad V(z)
be a gradient system. At regular poinla Lhe trajedmied cross level aujacu o r h p d i y .
~Vonrepularpanla are equilzkiu of lhc system. Isdolcd minima art aupmptoticdly
Theorem I V ( Z ) < 0 for all z E U ;and V(Z) = 0 if ond bnly if z is an cqui- 8hfJh.
librzum qf (1).
Prooj. By the chain rule
Ezarn,.de. Let V: R' -+ R be the function V(r, y) = z'(z - I)' + 9.Then we
have, putting z = (2, y):
V(E) = DV(z)zl
by (2) ; hence
- (grad V(+), -grad V(z))
V(r) = - I g a d V ( t ) 1.'
This proves the theorem.
Prwj. I t is easy to verify that the function z 4 V(Z) - V(f) is a strict The study of thin differential equation starb with the equilihrk These are
Liapunov function for 2, in some neighborhood of f . found by setting the righbhand aiden eqnal to 0, or - 2z(z - 1)(% - 1) = 0,
To understand a g d i e n t flow geometrically one looks a t the level svgaces of -29 = 0.
the function V: U -+ R. These sre the subsets V-'(4, c E R. If u E V"(c) is a We obtain precisely three equilibria: rl = (0, O), 21.11= (t,0). +m = (I, 0 ) . T o
regular puint, that is, grad V(z) # 0, then V-'(c) looks like a "surface" of dimen- check their stability properties, we compute the derivative Dj(r) which in cc-
sion n - 1 near E. TO 8ee this, assume (by renumbering the coordinates) that ordinates is
;1
dV/dr.(u) +! 0. Using the implicit function theorem, we find a function g:
R--1 -r R such that for z near u we have identically
(-2z(z - I)(& - I))
.:
(c) y sin z 5 9 42 4y +.4
(F) P + f - Z (f) f ( 2 - 1) +
Y'(u - 2 ) 'z + a*is called the basis dual to a.
Now euppose E ia given an arbitrary inner product ( , ). We define an lursoeiated
rmrrrenl if ~ ( 1 . ) -
2 . Suppose a dynarnical ayntem ia given. A trajectory z ( t ) , 0 5 t < -, is called
z ( 0 ) for some EWJence f + m. Prove that a grdient
dynanlical ayatern hsa no nonconstrrnt recurrent trsjectoriea.
map E -+ E* (as in Theorem 1) by B ( z ) ( y ) = (z,y ) . Clearly, cp is an b
morphism by Theorem 1, mnce its kernel is 0 .
Next, let V : W -r R be a continuously differentiable map defined on an open
3. Let V: E + R be C and suppose V-I(- m , c ] is compact for every c E R. set W C E. The derivative of V ia a continuous map
Suppose alea D V ( z ) # 0 except for a finite number of points pl, . . . ,p,. Prove: D V : W -+ L ( E , R ) = E*.
(a) Every solution z ( t ) of d = -grad V ( z ) ia defined for all 1 2 0;
(b) li, z ( t ) exists and equals one of the equilibrium points P I , . . . , P., A map W --r E.is called a I-form on W. An ordinary differentid equation ia the
for wery solution z ( t ) . same as s v&r field on W,that ia, a map W -r E. We we + I : F + E to con-
vert the 1-form D V : W + C into s vector field grad V : W 4 E:
of all linear maD8 E + R . We now prove some results of the preceding section concerning the differential
equation
Theorem 1 E* is isomorphic lo E and Ihw ha8 fhe hem dimnsirm. (2) z' = -grad V ( z ) ,
Proof. .
Let let, . . , e.1 be a basis for E and ( , ) the induced inner ~ r o d u c t . wing our new definition of grad V .
Thrn define u : E + P by z --r u.where % ( y ) = (2, y). Clearly, u is a linear map.
,
d
V ( Z ( ~ )=) - I grad V ( x ( t ) ) Is.
If z(1) ts no1 consbnl, h V(z(L)) is a decrea~t'ngjunction of 1. 0.. - 0'.
(Tz,y) = (I, Ty), for all
In an orthonormal basis thia means the matrix [a,j] of T is syn&,
I, y E E.
that is,
jr.) I, - I..
Theorem 5 la&E be a red veclor spow with an i n w produe(. Then any aelf- 4. If A i s 8 njmmetric operator, show that the vector field r + Az is the gsdient
-
adjvittt oprralm on E can be diugonal~ed. of some function.
P r m f . Let T : E E be self-adjoint. Sinoe the eigenvaluraa of T are real, there
is a nonzero vector el E E nuch that TeI = Xlel, XI E R.Let
Notes
t h orthogonal
~ complement of e,. If z E El,then Tz 6 El, for A etstement and proof of the implicit function theorem uaed in Section 4, is
(Tx, el) = ( x , TCI)= (z,&I) = h(z. el) = 0. given in Appendix 4. See P. Ualmcm' Finite Dimmsiao( Vcdor Spacu [8] for a
more extended treatment of self-adjoint boperators. One can find more on
Hence T lcaves El invariant. Give El the same inner product sa E ; then the operator Liapunov theory in I a S d e and Lefacheh'a S W d y by L i o p u d a Dirdd M e M
WiUL ApplicaCiOns [14]. Pontryngm's text [lo] on ordinary d~fferent~al equations
is recommended; in particular, he hsa m interesting application of tispunov
is self-adjoint. In the aame way we find a nonsero vector ,e E El such that theory to the study of the governor of a steam engine.
Te,= A*; A, E R.
Note that el and e, m independent, since (el, e,) = 0. Continuing in thin way, we
fmd a maximal independent set (8 = Iel, .. .
,em]of eigenvectas of T.These must
span E, otherwise we could enlarge the net by looking a t the restriction of T to
..
the subspace orthogonal to el, . , c.. In this bspis 63, T in diagonal.
flowing into a node is equal to the total current Bowing out of that node. (Think of fying Kirchholf'slswsforma three-dimenaional subpace K of the form K = K , x
tho \vntc,r analogy to make this plausible.) For our circuit t h b ia equivalent to Kt C R' x R'.
Next, we give a mathematid definition of the three kinds of electrid devices
KCL: in = iL = -ic.
of the circuit.
T h i d,afines a onedimensional subspace K , of R' of physiwl currenf st&. Our First consider the resiator element. A resistor in the R branch impoees a "fun*
cholcr of orientation of the capacitor branch may aeem unnatural. In fact the tional relationship" on in, OR. We take in our example this relationship to be de-
oricntntions are arbitrary; in the example they were chosen so that the equations fined by a C' real function f of a real variable, so that vn = !(is). If R denotes a
eventually obtained relate most directly to the history of the subject. conventional linear resiator, then f is linear and UR = f ( i ~ is ) a statement of Ohm's
The state of the circuit is characteri~edby the current itogether with the voltage law. The graph o f f in the (in, us) plane ia called the characlmslu of the resistor.
(or Iwttrr, voltage drop) acrow each branch. These voltages are denoted by vn, V L , uc A couple of examples of characteristics are given in Figs. B and C. (A characteristic
for the resistor branch, inductor branch, and capacitor branch, respectively. In the like that in Fig. C occurs in the,"tunnel diode.")
water analogy one thinks of the voltage drop as the difierence in preseures a t the A physical st& (i, u) E R' X R' = S will be one which sstiifiea KCL and KVL
two ends of a pipe. To measure voltage one placesa voltmeter (imagine8 water pres- or (i, u) E K and also f (i,) = u,. These conditions define a aubeet 2 C K C s
sure meter) at each of the nodes a,p, v which reads V(a) a t a , and so on. Then UB Thus the eel of physical slates Z is that set of points ( t ~i , ~ic,, on, VL, vc) in R' X Ra
is thc differellee in the reading a t a and p mtisfying:
V(p) - V(a) = VR.
T h r oricmtation or arrow tells us that V R = V(p) - V[a) rather than V(a) V(p).- vm +VL - vc = 0 (KVL),
An ur~reslricledvolloge st& of the circuit is then a point v = (UR,UL,uc) in 2'. f(in) = VR (generalized Ohrn'a law).
Again R Iiirchholf law puts a p h y d d restriction on v:
Next we concern owselves with the pasmge in time of a state; this defines a
KVL: ve + v~ - vc = 0. curve in the atate space S:
This defines a twodimensional linear aubspace K, of R'. From our explanation of
tbe vn, uL,vc in t e r m of voltmeters, KVL is clear; that is, The inductor (which one may think of as a coil; it is hard to find a water analogy)
UR + UL - vc = (V(p) - V(a)) + (V(a) - V(v)) - (V(8) - V(v)) = 0. specifies that
In a general circuit, one veraion of KVL ssserta that the voltages can be derived L- d i ~ ( t )= VL(~) (Fareday's law),
from a "voltage potential" function V on the nodes as above.
We summarise that in the product space, R' X R' = S, t h m ntatea (i, V) =tie- where L is a positive conatant called the inductance
I "#
I IVR
214 10. DIFFERENTIAL EQUATION6 FQR ELErnRLCAL ClRCUITB 02. LVALYSIS O F THE CIRCUIT EQUATIONS 215
011thr. other hand, the capacitor (which may be thought of as two metal p L a b PROBLEMS
sf.parat~cjhy some insulator; in the water model it is a tank) imposes the condition
1. Find the differential equations for the nctwork in Fig. D, where the rcskbr ia
voltage controlled, that is, the resistor characteristic is the graph of a CJfunc-
tion g : R -+ R, #(vn) = in.
n h m . C is a positive constant called the capacitance.
We summarize our development so far: a state of our circuit is given by the nix
numbers (in, i r , ic, V R , UL, V C ) that
, is, an element of R1 X R'. These numbera are
subject to three restrictions: Kirchhoff's current Law, Kirchhoff's voltage Law, and
the resistor characteristic or "generalized Ohm's law." Therefore the space of
physical statps is a certain subset Z C R' X R1. The way a state changes in time
is dctcrmined by two differential equations.
Ycut. \vr simplify the state space Z by ohserving that i~ and uc determine tb
otl~vrforlr coordinates, since i ~ = i Land ic = - i ~by KCL, V E = j ( i ~ )= j ( i ~ by
)
tht' pcen~~mlized Ohm's law, and U L = uc - un = uc - f(ir,) by KVL. Therefore
rvr can use Rt as thc state space, interpreting the coordinates as ( i ~vc). , Formally,
we define a map r : R' X R' 4 If', sending (I, u ) E R' x R' to ( i ~V ,C ) . Then m FIG. D
sct ra = x I Z,the restriction of r to 2 ;this map ro:Z + R ' b one-to-one and onto;
its inverse is given by the map u : RI -+ Z, 2. Show that the LX circuit consisting of one inductor and one capacitor wired
in a closed loop oscillates.
~ ( i ,V,C ) = ( i ~i ,~- ,i ~ ,W L ) ,vc - j ( i ~ )V, C ) .
I t is easy to check that q(ir, vc) satisfies KCL, KVL, and the generalised Ohm's
law, so p does map R' into E ;it ia also easy to see that ro and sre inverae to each $2. Analyais of the Circuit Equations
other.
We therefore adopt R ae our stab space. The differential equations goy-
the change of state must be rewritten in t e r n of our new coordinates ( i ~V,C :) Here we begin a study of the phase portrait of the planar differential equation
derived from the circuit of the previous section, namely:
di
L--L =
dt
UL = vc - j ( i ~ ) ,
"
dl
= ,-
ill f;wt a sink. Every state tends to zero; physically this is the dissipative effect of uuppresging t, this is equal to
thr rrsistor. Furthermore, one can see that ( 0 , 0 ) will be a spiral pink precisely
\\IIVII K < 2.
S ~ xwc
t consider the equilibria of ( 1 ) for a general C1 function j. by ( 1 ) . Here J could be any interval of real numbers in the domain of z.
Thrre is in fact a unique equilibrium i of ( 1 ) obtained by setting The statement of the proposition has an interpretation for the electric circuit
that gave rise to ( 1 ) and which we will pursue later: energy d e c along ~ the
solution curves according to the power dissipated in the resistor.
In circuit theory, a resistor whose characteristic is the graph of j: R + R, is
called parmve if its chsracteristic is contained in the set c o n $ s t i of ( 0 , 0 ) and
the interior of the first and third quadrant (Fig.A for example). Thus in the case
of a passive resistor - r j ( z ) i negative except when z = 0 .
The matrix of first partial derivatives of ( 1 ) at i is
The goal here is to continue the study of Lienard's equation for a certain func-
Prwj. Apply the chain rule to the c o m p i t i o n tion j.
J:R'~R
to obtain
10. DlFFEHENTlAL EQUATIONS FUR ELECTRICAL CIRCUIT[I b3. VAN DER POL'S EQUATION 219
These curves are disjoint; together with the origin they form the boundaries of
the four regions.
Next we see how the vector field ( z ' , d ) of (1) behaves on the boundary curves.
I t is clear that y' = 0 a t ( 0 , O ) and on v+ U u ,and nowhere elm; and z' = 0 exnctly
on g+ U r U ( 0 , 0 ) . Furthermore the vector (x', y') i~ horizontal on v+ U tr and
points right on v+, and left on rr ( f i g . B ) . And ( r ' , y') is vertical on g+ U g-, p o i n t
ing downward on g+ and upward on g-. I n each region A , B, C. D the signs of
x' and y' are constant. Thus in A , for example, we have z' > 0 , y' < 0 , and eo the
vector field always points into the fourth quadrant.
The next part of our analysis concerns the nature of the flow in the interior of
the regions. Figure B suggests that trajectories spiral around the origin clockwise.
C
I v-
The nrxt two propositions make this precise.
FIG. A
"
dt
= -=,
la this case we can give a fairly complete phase portrait analysis
11.17knr,\v from thr previous section that (2) has a unique equilibrium a t ( 0 , O), FIG. B
ntrrl it is ;I source. Thv next step is to show that every nonequilihrium ~olution
"rotatt>sVin s certain sense around the equilibrium in a clockwise direction. TO Proposition 1 A n y trajectory sfarfing a v+ enters A . A n y trajeclmy starting in A
this cnd \vc. divide the ( I , y ) plane into four disjoint regions (open sets) A , B, meek g + ; furlhennme it meek g+ befme it meek r,g- or vt.
C, L) in Iig. A. These rrgions m a k ~up the complement of the curves
Prooj. See Fig. B . Let ( z ( t ) ,y ( t ) ) be a solution CWE to ( 1 ) . If ( r ( O ) , y ( 0 ) ) E
v+, then x ( 0 ) = 0 and y ( 0 ) > 0 . Since t ' ( 0 ) > 0 , z ( t ) increaaea for anall t and
so z ( t ) > 0 which implies that y ( t ) decreases for small t. Hence the curve enters A .
Before the curve leaves A (if it does), z' must become 0 again, so the c w e must
TI~esrcurves ( 3 ) thus form thc boundaries of the four regions. Let us make this
cross g+ before it meets rr, r or u+. Thus the first and last statements of the pmpo-
nlorc. prl,r.isr. Definp four curves
sition are proved.
I t remains to show that if ( 2 ( 0 ) , y ( 0 ) ) 6 A then ( x ( t ) , y ( t ) ) E g+ for some
t > 0 . Sup- not.
Let P C Ra be the compact w t bounded by ( 0 , O ) and vt, g+ and the line y = y ( 0 )
as in Fig, C. The solution curve ( ~ ( t )y (, l ) ) . 0 < t < @ is in P.From Chapter 8,
i t follows since ( ~ ( t )y ,( t ) ) does not meet g+, i t is defined for all t > 0 .
Since z' > 0 in A , r ( t ) 2 a for t > 0 . Hcncc from ( I ) , y'(t) 5 -a for t > 0 .
-
10. DIFFERENTIAL EQUATIONS FOR ELECTIlICAL CIRCUITS ~ 3 . VAN DER m ~ l aEQuAmoN 2-21
-
P r w j . If r ( p ) = p, then &t,(p) = p, where 1, = t,(p) is as in the delinition
For these values oft, then of u. Suppoee on the other hand that u ( p ) # p. Let v* = vf u ( 0 , 0 ) . We observe
'rat that r extenda to a map u* v* which is again continuous and one to one,
u(t) = I' ~ ' ( 8d). 6 u ( 0 ) - .L sending ( 0 , 0 ) to itself. Next we identify u* with ( y t R I y 2 0 ) b y amignhg to
each point its y-~oordinate.Hence there is a natural order on v': ( 0 , y) < ( 0 , z) if
y < z. It follow8 from the intermediate value theorem that u : tf -r tf is order
This is impossible, unless our trajectory meets g+, proving Proposition 1.
preserving. If r ( p ) > p, then d ( p ) > r ( p ) > p and by induction e ( p ) > p,
Similar arguments prove (aee fig. D) :
..
n = 1, 2, . . This means that the trajectory of p never c- u+ agdn a t p.
Hence & ( p ) # p for dl t # 0. A simiiar argument applien if o ( p ) < p. Themfore
if r ( p ) Z p, p is not on a periodic trajectory. The lsllt statement of Proposition 3
follows from Proposition 2 which implies that every trajectory (except ( 0 , 0 ) )
meets vf.
For every point p E vf let k ( p ) = k be the smallest t > 0 auch thst +,(p) E w.
Define a continuow map
To analyze further the Bow of the Van der Pol oscillator we define a map I"-
e: v+-+ vf
FIG. E. The map *: o* - .*.
10. DIFFERENTIAL EQUATIONS FOR mECTRICAL ClRCUlTB w. VAN DER POL'S WUATION
Since r has only one fixed point q, = q This ahows that h e h i- of p *rob
totwrdrasf+w.Thesamething~tmeifp<~;thedetailsmlefttothe
reader. Since every trajectory except (0,O) meeta v+, the proof of the main theorem
in complete.
-
I t remains to prove Proposition 4.
We adopt the following notation, Let 7 ; [a, b] + R' be a C1 curve in the p h e ,
written y(f) = (z(t), y(t)). If F: R' R in C , define
a(p) = [ -
- = ( t ) ( ~ ( t ) ~ 2 ~ ) dl;
)
where f(z) = i - t. As p moves up the -is,
where
\\.F her? a tiny of finding the ordinary differential equations for a claas of Proof. It is d c i e n t to check the condition for each node a E A. Thus (dt?. =
elevtricnl !rr,tworksor circuits. We consider networks made up of resistors, capaci- 0 if and only if
tors, and inductom. Later we diacuaq briefly the nature of these objeets, called the Cc& = 0,
branches of the circuit; at premnt it suffices to consider them as devices with twb DfB
23 10. DIFFERENT~A~
EQUATIONS rOR ELECTRICAL CIRCUITS 05. MORE GENERAL CIRCUIT EQUATION6
u*: F* -
I f u . E -,F is a linear trnnsforniation, then its adjoin! or dual is a lirlrar map
E* defined by u8(+) (y) = +(u(y)), where I E F*, y E E. (Here u*(r)
t E* and maps E -,R.)
is an e l r m ~ n of
W ( f a ) , where f. C I, has kth coordina~e1 and all other coordinates 0.Then V is
a voltage potential for v since the voltage which v assigns to the branch 6 is
Now let + be the natural bilinear map defined on the Cartesian product vector
space E X E* with values in R : if (e, e*) € E X E*, then +(e, e*) = eL(e).
4(e, e*)
E K so that u(e)
= 4(e, u*y) =
= 0 and e*
(u*y) (e)
- ufy for some y € F'.
= y (u(e)) = 0.
Then Thr space oJ utlreslricled stales of the circuit is the Cartesian space # X d*. Thoee
ststm which satisfy KCL and KVL constitute a linear subspace K C d X s*. By
Theorems I and 2,
This proves the proposition.
K = Ker d x Im d* C s x s*
Remark. A further argument shows that dim K = dim E. An actual or physical state of the network must lie in K.
We return to the analysis of the voltage and current states of a network. It Thc p o w 6 in a network is a real function d~finedon the big state spaee s X S *
turns out to be useful, as we shall see presrntly, to identify the space with the and in fact is just the natural pairing diiussrd earlier. Thus if ( r , v) E d X do,
dual spare 8' of 9. \Iathematically this is no problem sinrr both 'U and 9 ' are
the power +(i, 11) = u(i) or in terms of Cartesian coordinates
nnt,lr:~llrisonlorphic to Rb. With this idrntification, the voltage which a voltage
.I:LIV 1' . g * axsigns to the kth branch 6 is just u(id), whrre 1 8 E # is thv vector
u I ~ ~ , rt itl .t . kt11 cnordinate is 1 and where other coordinatvs art. 0.
\Vt- rnn now rxprpss I<VL more elcgantl3:
Thv prrvious proposition givcs us
Yo\\- iar describe in mathematical terms the three different types of devices in Let C be the space of all currents in the inductor branchea, 80 that E is n a t u d y
tllr ~ ~ r t n - t ~the
r k :resistor, inductor, and capacitor. These devices impom conditionn isomorphic to R',where 1 is the number of inductors. A point i of 2 will be d e n o w
on thr state, or on how the state changes in time, in the corresponding branch. by i = (11, . . . , i d ) where ir is the current in the Ath branch. There is a naturd
Kwh resistor imposes a relation on the current and voltage in its branch. This map ( a projection) ir: # -+ which just eenda a current state into its components
-
An inductor or capacitor does not impose conditione directly on the state, but
only on how the state in that b m c h chsngee in time. In particular let A be an
inductor branch with current. voltage in that branch denoted by Q, vh. Then the
-
Hypotheaim The map I: Z -
;e
E X e* haa an inverrre whkh is a C map
x e* -, e c s x a*.
xth inductor imposes the condition:-
d i ~ Under thii hypothesis, we may identify the space of physical states of the n e t
(18) L*(ir) - dl = ur. work with the space E X e*. This is convenient because, as we shall see, the dif-
Hrrc I,r is determined by the inductor and is called the inductance. I t is assumed ferential equations of the circuit have a simple formulation on E X e*. In words
t u Irr. n ('I positivc function of ir.
S1111ilnr1y 4 capacitor in the 7th branch defines a C positive function v,
t.rillvcl tlw capacitancr; and the current, voltage in the Tth branch satisfy.
-
C,(u,)
the hypothesis may be stated: the current in the inductors and the voltages in
the capacitors, via Kirchhoff's taws and the laws of the &tor
determine the currents and voltages in all the branches.
chanrctaistiea,
Although thii hypothesis is strong, it makea some sense when one reslim that
du the "dimension" of Z should be expected to be the same as the dimemion of
c,(v,) -2 = i,.
dl E X 12'. This follows from the remark after the proposition on dim K, and the fact
that Z is defined by r additional equations.
We now examine the resistor conditione more carefully. These are conditionn on
the state3 themselves and have an effect similsr to Iiirchhoff's laws in that they T o state the equationn in this case we define a function P : 9 X #* + R called
place physical restrictions on the space of all states, s X S*. We define Z to be the the mired pdenlicrl. We will follow the convention that indices p refer to resistor
branchea and auma over such p means summation over the resistor bnmehes.
-
sobst-t of g X 8. consisting of states that satisfy the two Kirchhoff laws and the
rrriitr~rronditions. This space Z is called the space of p h y s i d a t a h and is de- Similarly A is used for inductor branches and 7 for capacitor branches. Then
zcribcd lry P: # X 8. R b defined by
3 = l(t,s)CeXebj(i,v)EK,f,(i,) =v,,p=l, . . . , rl.
HI^ ( I , , ~ b )denotes the components of i, u in the pth branch and p varies over
tllr rt.sistor branches, r in number.
1indt.r rather generic conditions, E will be a manifold, that is, the higher dimen-
sional nnalog of a surface. Differential equations can be defined on msnifolda; the
arbitrary conntant. Now P by restriction may be considered as s map P : Z
and fiually by our hypothesis may even be considered as a map
-
Here the integral tefera to the indefinite integral so that P is defined only up to an
R
-
rapncitors and inductors in our circuit will determine diRerential equations on Z
\\.lrosr corresponding flow *,:E Z describes how a atats changed with time.
Bccauae we do not have a t our d i i the notionn of dilTerentiable manifolds, (By an "abuse of language" we use the m e letter P for all three map.)
we will make a simplifying assumption before pmeeeding to the differential eqw Now assume we have a particular circuit of the type we have been considering.
tions of thc circuit. Thia is the assumption that the space of currents in the in- At a given instant 4 the circuit is in a particular current-voltage atate. The states
durtors and voltsges in the capacitors may be wed to give coordinates to Z. We will ehange as time goes on. In this way a c w e in # X #* is obtained, depending
makt*this more precise. on the initial state of the circuit.
Id4 10. DIFFERENTIAL EQUATIONS FOR ELECTIIICAL CIRCUIT8 $5. MORE CENEML CIRCUIT EQUATIONS 235
Tlle cunlponents ip(t), vp(t), 0 E B of this eurve must satisfy the conditions from the drfinition of P and the generalized Ohm's laws. By the chain rule
inrposd by Kirchhoff's laws and the resistor characteristics; that is, they must
be in P. In addition a t each instant of time the components dir/dt and du,/dt of
the tangent vectors of the curve must satisfy the relations imposed by ( l a ) and
( 111).A curve satisfying these conditions we call a phygicol trajectory.
From the last two equations we find
If tlit. circuit satisfirs our special hypothesis, each physical trajectory is identi-
n rurve in & X e*. The following theorem says that the curves so obtained
f l ~ d\\-~tli
arc- cxartly the solution curves of a certain system of differential equations in c (2+ "r) ibl + c ('- i,) v; = 0.
a: x e*.
Since 4' and vA' can take any values,
Some remarks on this theorem are in order. First, one can follow this develop
ment for the example of Section 1 to bring the generality of the above down to
earth. Secondly, note that if there are either no inductors or no capacitors, the
Brayton-Moser equations have many features of gradient equations and much of
zr.li,.~<h rlrirl 7 rut8 throagh all in(luctors and capoc'fors o j the circuit respectively. the material of Chapter 9 can be applied; see Problcm 9. In the more general cnse
('utzr-rrsrly, every solutior~curve fo t h e equations is a physual trajectory.
-
the equations have the character of a gradient with respect to an indefinitemetric.
We add some final remarks on an energy theorem. Suppose for aimplieity that
Here P is the map E X e* R defined above. The right-hand sides of the all the LAand C, are corwtant and let
d;fferential equations are thus functionn of all the i,, u,.
Proor. Consider an arbitrary C1 curve in E X e*. Because of our hypothesis be the function W(i, u) = f Lrir' + + x,
Cp,'. Thus W haa the form of a
a . irlrntii>-
~ a: X e* with Z C 9 X 9'; hence we write the curve norm square and its level surfaces are generalized ellipsoids; W may be interpreted
aa the energy in the inductor and capacitor branches. Define P.: t X e* + R
(power in the resistors) to be the composition
I \ \ liirc~lll~olf's Ian- (l'hmrcm 1) i(1) f I<er d. Hence i'(t) F Ker d. By Theorem 2
1.1 1r I I I I (1'. 1%-l'~~llc,~vn's
tht.orrm, for all 1
where P . ( t , u) = x i,v, (summed over resistor branches). We state without proof:
U Prr\vnte this as T h e o r e m 5 Let +: I + E X e* be any solution o j the equatirms of Ule prenkxa
C v,i, + x u l i r + C u,i, = 0.
theorem. Then
3. Prove that dim K = dim E (see the proposition in the text and the remark
at ter i t ) .
do,.
C' -
dl
= a& - J ( i J , . ) .
4. I'ruvr Throrcm 5.
Here i = / ( v ) gives the mistor characteristic.
5. G1ns1dr.r resistors whose characteristic in of the form F(a,, v,) = 0, where F is
:Irr,~rl-valuedCLfunction. Show that an RLC circuit (Fig. A) with this kind of 7. Suppoae given a circuit satisfying the basic hypothesis of this & 44d d
rcmsistr~rantisfie the special hjpoth* if and only if the resistor is current the other ammptions except that the characteristic of one resistor is given
rontn,llcd, that is, F has the form ,
by a voltagecontrolled char~cteristici = f ( v ) , not neeeesarily current con-
trolled. Show that if the corresponding term of the mixed poteatia Pis rep-
by Ivf'(v) du, then Theorem 4 is still true.
8. F i d the differential equations for this circuit (Brayton) (Fig. C). Here1111
denotes a hattery (reaistm with characteristic : v = conat.), q&&. denotes s
FIG. C
238 10. D~FPERENTIAL EQUATIONS FUR ELECTRICAL CIRCUITU
linear resistor, and the box is a reaintor with characteristic given by i = f(v).
9.
Find the mixed potential and the phase portrait for some choice of f. See
Problem 7.
We refer to the Brayton-Moser equations. Suppose there are no capacitonr.
Chapter 11
( a ) Show that the function P: C --. R decreases along nonequilibrium tra- The Poincarb-Bendixson Theorem
jt-ctoriea of the Brayton-Jtoaer equations.
(1)) I r t n be the number of inductors. If each function I., is a constant,
find an inner product on R' = C which makea thr*vector
Notes
We have already Been how periodic solutions in plnnar d @ d systems play
an important role in electrical circuit theory. In fact the periodic wlution in Van
der Pol's equation, coming from the simple circuit equation in the preview chapter,
This chapter follows to a large extent "%tathemtical foundations nf clcctriral haa feature that go well beyond circuit theory. This periodic wlutioa is a 'limit
c ~ r r t ~ ~ 1~ .3.Snlalr
"~ In tl~rJor(nmlofD~~e~'e,ztinI
Geot~tetry1221.1'111.I I I I ~ ~ I ~ I . E I . : I , ~ \ I . I I ~ * cycle," a concept we make precise in this chapter.
trxt nn rlrctrical circuit theory by Desoer and Icuh [53 is excellent for a trratrnnlt
The Poinca&Bend'uson theorem givea a criterion for the detection of limit
of nluny rrlated subjects. Hartman's h w k [9], ~nentionrdu l ~ oIn Cl~nptcr 11, cycler ~nt l ~ e ~ l a o(I~isrriterloncould
e. have been used to findthevan der Potoscilla-
goes extensively into the material of our Sections 2 and 3 with many historical
tor. On the other hand, this approach would have miaaed the uniqu-
refercnces. Lefwhetz's book Differential Epuolions, Geometrical Theory [I41 also Poinc&Bendixson is a basic tool for understanding p hd v i d~ I M
discusses these nonlinear planar equations. Van der Pol himsrlf related his quation but for diflerential equations in higher dimensions i t has no -tion or
to heartbeat and recently E. C. Zeeman haa done very intrrrsting work on this counterpd. Thus nfter the first two rather basic sections, we restrict o d v e s to
. some phyeical background of circuit theory, onr can see The Feynman
s u b j ~ t For
planar dynamical system. The first section gives wme properties of the Ilnitirq
Lccturca on Phy& [6].
behavior of orbits on the level of abatraet topological dynamics while in the next
section we analyze the flow near nonequilibrium points of a d y n a m i d aystao.
Throughout this chapter we consider a dynamical ayatem on an open set W in a
vector spaee E, that is, the flow 9,defined by a C1 vector field f: W 4 E.
if there is a Pequence I . --
We recall from Chapter 9, Section 3 that y C W is an o-limd poi& of z C W
such that lim.,, d , . ( t ) = y The set of all *limit
-
points of y is the o-limit eel L.(u). We define a-limit poinla and the d i m i r ad L. ( y )
by replacing I. -+ a, with 1. - m In the above definition. By a limit at we
mean a set of thr. form L..(y) or L.(y).
Here are some examplea of limit e t s . If r is an asymptotically strbte equilib
rium, it is the o-limit set of every point in ita basin (seeChapter 9, Seetion 2). Any
P P ~ (a) . Suppoee y E L ( z ) , a n d +(z) = 2. If +L(z) -+YI ~JWI +-(z) v.
-
+
Renee y E L.(z).
(b) If f. rn and +,.(z) 2 y E 4 ( z ) , then f. 2 0 for *~JY b n
m that +,.(z) E 0. Renee y E D = D.
(c) Follow from (b).
PROBLEMS
- FIG. A
1. Show that acompact limitset isconnected (that IS.not the union oftwodisjoint
nonempty closed eets.
2. Identify R' with C' having two complex coordinates (w, r), and w d d e x tbe
linear s y d m
(*) w' = 2riw,
orbit 1 ;it is the -limit of every point except the origin (Fig. A ) . The origin is t b (b) Let +, be the flowof (*). Show that for n an integer,
a-limit set of every point inside 7. If y is outside y, then L ( y ) is empty.
There are examples of limit sets that are neither closed orbits nor equilibria, for
example the f~gure8 in the flow suggested by Fig. B. There %rethree equilibria, two (c) a) belong to the toma C X C C C'. Use (a), (b) to sbow thrt
Let (a,
aaurees, and one saddle. The figure 8 is the ~ l i r n iset
t of all points outside it. The
right half of the 8 the A m i t set of all points inside i t except the equilibrium, and
similarly for the left half. (d) Find 4 and L. of an srbitrary point of C'
In three dimensions there are extremely complicated examples of limit sets,
although they are not easy to describe. In the plane, howwer, limit sets are f d y 3. Find a linear aystan on
CX ...
XCCCk,then
R* - Cb mch that if a belonga to the &torum
simple. In fact Fig. B is typical, in b t one can ahow that a limit set other than a
closed orbit or equilibrium is made up of equilibria and trajectories joining them.
The Poincan%Beodixson theorem says that if a compact limit set in the plane
contains no equilibria it is a closed orbit. 4. I n Problem 2, mppoee instead that 0 is raliond. Identify L and L d nery
We recall fmm Chapter 9 that a limit set is c l a d in W, and in invariant under point.
the flow. We ahall alao need the following result: 5. Let X be a nonempty compact invsriant set for a (? dynamidsyatcm. Buppwe
that X is minimal, h t is, X contains no compact in&t nonempty proper
subset. Prove the following:
Propodtion (a) If r and r are on Ulc aume trajcdory, then 4 (2) = L.(z) ;8imC
hrly for a-limits.
( h ) If D w a dodd po&bdy inadant ad and o € D, lhm L ( z ) C D ;s i m h i y
-
(a) Every hjeetory in X in d e w in X ;
-
(b) La(?) 4 ( z ) X f o r e a c h t E X;
(c) For any (relatively) open set U C X, there is a number P > 0 such th.(
for nqalivcly inadant & and d i m * . f o r a n y r E X , b E R , C h e r e e ~ t m c h b t + , ( zE) U a n d I t - 4 1 <
(c) A c l o d iwarianl ad, in particuhr a limd set, conhind the d i m i l and dimit
add of aey point i n il.
p;
141 11. THE WIN CAR^-BENDIXSON THEOREM $2. IACAL SECTIONS AND FLOW BOX-
6. Lrt X be a closed invariant set for a CLdynamics1 Bystem on R', such that
1 N - LQ(NI
g,(z) is defined for all t E R, z € X. Suppose that L.(r) = L.(z) = X for FIG. A. The flow box.
all r : X. I'rove that X is compact.
S X (-a, a),where S C H is a section at 0 and a > 0. I n this ease u-e sometimes
.
write V. = .(N) and call V. a pow bor a t (or about) 0 in E. See Fig. A. An
52. l ~ x - a lScctiona a n d Flow Boxen important proprrty of a flow box is that if z E t'.. then +,(I) E S for a unique
1 € (-u, a ) .
From the definition of it follows that if e L ( p )= (s,y ) , then W1(+,(p)) =
\Vr consitlrr again the flow 9, of the C' vector field J: W -- E. Suppose the origin (8 + 1, y) for sufficiently small I a ] , I f I.
0 F E belongs to W . We remark that a flow box can be defined about any nonequilibrium point 4
A local srcfion at 0 of Jis an open set S containing 0 in a hyperplane H C E which The assumption that a = 0 i~ no real restriction since if . z is any point, one can
is transvrrw to J By a hyperplane we mean a linear subspace whose dimension replace J(z) by f ( z - a ) to convert the point to 0.
is oncbIps8 than dim B. T o say that S C H is m w s e to j means that j(z). 6 H If S is a local eection, the trajectory through a point r, (perhaps far from S) may
for all r E S. In particular j ( z ) # 0 for z E S. reach 0 E S in a certain time 4; 8ee Fi.B. We ahow that in a c& I d sense,4
Our first use of a local section a t 0 will be to aonstmct a "flow box" in a neighbor- is a continuous function of zp More precisely:
hood of 0. A flow box gives a complete description of a flow in a neighborhood of
any ~ronr~uilibrium point of any flow, by m w of special (nonlinear) coordinates
\\.v I I I ~ ~ this -
Thp drscription is simple: points move in parallel straight lines a t constant s p e d .
L ~ , I)r~ciaca s follows. A diffmorphian t:U V is a differentiable
map f t , > n ~t911,. opcn set of a vcctor space to-another with a diflerentiablc invew.
.\ ,flair. bor is a diffeomorphism
.
of j is thereby converted to a simple flow on R X H : FIG. B
.
for (I, y ) in a sufficiently small neighborhood of (0,0) in R X H . One appeals b
I, and
+.I., (1) E S
.
('hnptrr 15 tosee that I is a CLIMP. The derivative of a t (0,O) is easily computed for all z f Z'.
to bv the linear map ~ h i c his the identity on 0 X H, and on R = R X 0 it sends
Proof. Let h : E +R be a linear map whose kernel H is the hyperplnae eon-
1 to j ( 0 ) . Since j ( 0 ) is tramverse to H, it follows that DP(0, 0) is an isomorphism. taining S. Then h ( j ( 0 )) # 0. The function
Hence by the inverse function theorem maps an open neighborhood N of (0,O)
diffc~omt~rpl~irally onto a nrighborl~nod V of 0 ill E . Wt. taka- N of the forpl.
in 0, and
(3
FIG. C
$3. MONO'IQNE SEQUENCES I N PLANAR D Y N A M I C h L S Y S T E M S 247
It follo\vs that +(y) = +,(y) ; hence 6-,(y) = y, r - s > 0.Since L.(z) contains
no equilibrium, y belongs to closed orbit. 1. Consider a CLdynarnical sy&m in R1having only a finite number of equilibrk
It remains to prove that if y is a closed orbit in L ( z ) then 7 = L.(z). I t is (8) Show that every limit set is either s closed orbit or the mion of equilibria
enough to show that r o d trajectories rp,(z) such that Lim,,+,(z) and Lim,,+,(z) are
equilibria.
(b) Show by example (draw a picture) that the number of diatinet trsjeetories
in L ( z ) may be infinite.
where d ( + , ( z ) , v) is the distance from x to the compact set y (that is, the distance 2. Let 7 be a c l o d orbit oi a C dynsmical system on an open net m R '
. Let
from +,(z) to the neareat point of 7). be the period of y. Let (7.) be a eequeme of closed orbits; glppome the period
?,-PO li. THE PO IN CAR^-BENDIX~N THEOHEM
(This result can be false for higher dimensional syatems. I t is true, however, that
if A. -+ p, then p is an integer multiple of A.)
In the proof of the Poinc&Bendixson theorem it waa shown that limit cyclea is positiveiy invariant, as is the net B = A - y . I t is easy t o m that +,(y) s p k l s
enjoy a certain property not shared by other closed orbits: if y is an w-limit cycle, toward y for ail y B. A useful consequence of this is
there exists x f y such that
A = l y l - f = L.(y)I - y
We sd
Proof. i.'or suffici~ntly large t > 0,+, (z)is in the interior of the net A described
above. Hence + , ( y ) E A for y sufficiently ciose to x. This implies the pmpoeition.
Limit cycies possess a kind of one-sided stability. Suppose r is an w-limit cycle The next result exploits the spiraling property of limit cyclea.
and iet +,(z) spirai toward y as t -+ m. Let S h e local section a t z € y . Then there
will br an interval T C S disjoint from y bounded by + , , ( x ) , +,,(z), with lo < 11
a ~ r dnot meeting thc trajectory of x for < t < 1, (Fig. B). The region A bounded Proposition 2 Let y be a closed orb11arid suppose thaf the domain W of& d y o m i c d
I,?. r , T and the curve Lr enclosed by y. Then U amtaitu d h m an
systern inclruies the tuhole open region
equilibrium or a lintit cycle.
l+dz)I lo 5 t 5 tll
2.19 11. THE P ~ ~ H C A R ~ - B E N D I XTHEOREM
$ON
6.APPLICATIONB OF THE POINCAR~-BENDIXSON THEOREM 253
Proof. Let D be the compact set U U 7 . Then D is invariant since no trajectory PROBLEMS
from O can cross 7. If U contains no limit cycle and no equilibrium, then, for any
x t i',
L.(z) = L.(z) = 7 1. The celebrated Brouwcr fcrd poinf Ihwrem statea that any cvntinuous map f
of the closed unit ball
- If S is a local swtion a t a point z
b! l'r,~~~~,:rrf-Rclrdixson.
I, a . s, -t - rn such that
t 7 , there are sequences
+,.(a) € S, +d.(z) -+z, into itaelf has a fued point (that is, j ( z ) = z for some x).
and (a) Prove this for n = 2, ~nsumingthat J is C', by finding an equilibrium for
. t s, +..(X) -+z. the vector field g(z) = /(z) - z.
But this leads to a contradiction of the ~ropmitionin Section 3 on monotone (b) Prove Brouwer'rr theorem for n = 2 using the fact that any continuous
map is the uniform l i t of C' maps.
sequerrces.
2. Ifit f be a CLvector field on a neighborhood of the annulus
Actunlly this last result can be considerably sharpened:
Suppose that, f haa no zeros and that j is transverse to the boundary, pointing
Theorem 2 Let 7 be a closed orbit enclosing an open ael L' contained in the domain inward.
1V of the dytramieal system. Then U contains an equilibrium.
( a ) Prove there is a closed orbit. (Notice that the hypothesis is weaker than
Proof. Suppose U contains no equilibrium. If z. -z in U and each x. lies in Problem 1, Section 3.)
on a clost.d orbit, then x must lie on a closed orbit. E'or otherwise the trajectory of (b) If there are exactly seven closed orbits, show that one of them has orbits
z I\-ould spiral toward a limit cycle, and by Proposition 1 so would the trajectory spiraling toward i t from both side.
of some z..
3. Let J; R1 -+ R' be a C1vector field w i t h no zeros. Suppose the flow +I generated
Let A 2 0 be the greatest lower bound of the areas of regions enclosed by closed
by j preserve area (that is, if S is any open set, the area of +,(S) ie independent
orbits in U . Let (7.1 be a sequence of closed orbits enclosing regions of a m A.
z, -
such that lim,, A. = A. Let z. E 7.. Since 7 U U ie compact we may assume
x c c', Then if U contains no equilibrium, x lies on a closed orbit B of area
.4 (01.Tlrc usual section argument shorn thnt 89 n + Q , 7, gets arbitrarily close
4.
of 1). Show that every trajectory is a closed act.
Let J be a CLvector field on a neighborhood of the annulus A of Pmblem 2.
Suppose that for every boundary point x, J(x) is a nonzero vector tangent to
~ area A. - A (B), of the region between 7. and 8, goes to 0.Thus
to p I I I I ~I I ( , I I C the the boundary.
r. (0) = ..I. ( a ) Sketch the pmible phase portraits in A under the further assumption
\Ye have shown that if U contains no equilibrium, i t contains a closed orbit ,3 that there are no equilibria and no closed orbits besides the boundary
enclosing a region of minimal area. Then the region enclosed by B contains neither circles. Include the cam where the boundary trajectories have opposite
an rquilibrium nor a closed orbit, contradicting Proposition 2. orientations.
(b) Suppose the boundary trajectories are oppositely oriented and that the
The following result uses the spiraling properties of limit cycles in a subtle way. flow preserves area. Show that A contains an equilibrium.
5. Let J and g be C' vector fields on R2 such that Cf(x), g(x)) = 0 for all z. If j
Theorem 3 Lef H be a $ref inlcgral of a planar CLdynamical a y a h (thal is, H has a closed orbit, prove that g has a zero.
LS a r e a l - r a l d Junction lhal i s conslant on lrajcclorics). IJ H is not cmlanl on any
opetl sel, /hen here are no limd cycles. 6. LetJbeaC'vectorfieldonanopensetWCR'andH:W+RaChurction
such that
P~.ooj. Suppose there is a limit cycle 7; let c E R be the constant value of H
on r. If z(t) is a trajectory that spirale toward 7 , then H(z(1)) E c by continuity DH(x)/(z) = 0
of H. In Proposition 1 we found an open set whose trajectories spiial toward 7 ; thua for all x. Prove that:
H is constant on an open set. (a) H is constant on solution curves of z' = /(x) ;
I 11 I
I rI
1)H (r) = 0 if r belongs to a limit cycle;
If x belongs to a compact invariant set on H-hichDH is never 0,then x
1ic.s on a eloped orbit. Chapter 12
Ecology
Notes
1'. Hartman's Ordinary Diferenlial Epolions [9], a good but advanced book,
coven cxtcnsively the material in this chapter.
I t should be noted that our diiussion implicitly used the fact that a closed curve
in R?n-hich does not intersect itself must separate Rzinto two connectvd rrgions, a
bounded one and an unbounded one. This theorem, the Jordan curve theorem, while
naively obvious, needs mathematical proof. One can be found in Newman's Topology
of I ' l a ~ ~ Sols
c [i7].
In this chapter we examine some nonlinrar two dimrnsional systems that have
been usrd as mathematical models of the g r ~ ~ wof t htw-o sppcies sharing a common
environment. In the first section, which treats only a single species, various math*
metical assumptions on the growth rate are discussed. These are intended to capture
mathematically, in the simplest way, the dependence of the growth rate on food
supply and the negative effects of ovemrowding.
In Section 2, the simplest types of equations that model a predator-prey ecology
are invcstigated: the object is to find out the long-run qualitative behavior of tra-
jectories. A more sophisticated approach is used in Section 3 to study two competing
species. Instead of explicit formulas for thr equatiuns, certain qualitative assump
tions are made ahout the form of the equations. (A similar approaeh to predator
and prey is outlind in one of the problems.) Such assumptions are more plausible
than any set of particular equations can he; one has correspondingly more confidence
in the conclusions reached.
An interesting phenomenon ohserved in Section 3 is bifurcation of behavior.
RIathernatically this means that a slight quantitative change in initial conditions
leads to a large qualitative difference in long-term behavior (because of a change of
w-iimit sets). Such bifurcations, also called "catastrophes," occur in many applica-
tions of nonlinear systems; several recent thwries in mathematical biology have
been based on bifurcation theory.
The birth rate of a human population is unually given in terms of the number
of births per thousand in one year. Thr ~rumbrrone thousand is used merely t~
avoid decimal places; instead of a hirth rate of 17 per thousand one could just ne
2.)6 12. ECOLOGY $1. ONE SPECIES 257
uell spenl, of 0.017 pcr individual (although this is harder to visualize). Similarly, dependent on the partieular environment but constant for a given ecology. ( I n
thr prriod of one year is also only a convention; the hirth rate could just es well the next section o will he another species satisfying a second differential equation.)
br plivc~nill terms of a week, a sccond, or any other unit of time. Similar remarks The preceding quation is readily solved:
al,]>lyto the death ratc and to thr growlh mk, or hirth rate minus death rate. The
~rowtlirate is thus the nct change in population per unit of time divided hy the
total population at the hrginning of thc timc period. Thus the population muat increase without limit, remain constant, or approach
+
Suppose the population y ( 1 ) at time t changes to y Ay in the time interval 0 es a limit, depending on whether o > oa, = -0, or o < so. If we recall that actu-
[f, f +At]. Then the (average) growth rate is ally fractional values of y (f) are meaningless, we see that for all practical purposer,
"y (1) -+ 0" really means that the population dies out in a finite time.
In reality, a population cannot increase without limit; a t least, this has never
been observed! I t is more realistic to assume that when the population level exceeds
I I I praeticr y ( t ) is found only at such times b, h, . . . when population is counted; a certain value q , the growth rate is negative. We call this value v, the limiting
and ~ t valur
s IS a nonnegative integer. We assume that y is extended (hy interpola- population. Note that ? is not necessarily an upper bound for the population. Rea-
tion or some other method) to a nonnegative real-valued function of a real variahle. Eons for the negative growth rate might be insanity, decreased food supply, over-
Wr assume that y has a continuous derivative. crowding, smog, and so on. We refer to these varioua unspecified eauses es social
Giving in to an irresistible mathematical urge, we form the limit phenomena. (There may be p i t i v e soeial phenomena; for example, a medium size
population may he better organized to resist predators and ohtain food than a
small one. But we ignore this for the moment.)
Again making the simplest mathematical assumptions, we suppose the growth
rate is proportional to 1 - y :
'11111s~ U I I I . ~ I Cof> I ~1 is the growth rate of the population a t time 1.
'1'111. sllnplrst assumption is that of a constant growth rate a . This is the case a = C ( T - y), c > 0 a constant.
if tlri! number of births and deaths in a small time period A1 have a fixed ratio to Thus we ohtain the eqwtion of limiled gro]uIh:
the total population. These ratios will he linear functions of At, hut independent
111tllc size. of the population. Thus the net change will be ay At where a is a constant;
lit,llrv
;
a = -y' = - logy;
Notr that this suggests
i ~ ~ t ~ ' ~ r nwe
t i nohtain
g the familiar formula for unlimiled growth:
This means that during the period At the population changc is cy2Af less than it
would be without social phenomena. We can interpret cy' as a number propor-
The growth rate can depend on many things. Let w assume for the moment that tional t o the averagc numher of encounters hetween y individuals. Hence cy' is a
it dcpends only on the per capita food supply o, and that a 2 0 is constant, There kind of sorial friction.
will be a m i ~ m u muo necegsary to sustain the population. For o > oo, the growth
rate is positive; for o < no,it is negative; while for s = ro,the growth rate is 0. The
-
The guilihria of (2) oecur a t y 0 and y = q . The equilibrium a t q isasyrnptot
ically stahle (if c > 0 ) since the derivative of c ( q - y)y at t j is -cq, which is
simplrst way to ensure this is to make the growth rate a linear function of o - oo: negative. The hasin of q is 1 y I y > 01 since y (1) will increase to q aa a limit if 0 <
y (0)< II, and decrease to 1 as a limit if < y (0). (This can be seen by eonsidering
the sign of dyldt.)
A more realistic model of a single species is
Here the variahle gmwth rate M is assumed to depend only on the tobl population
Elc,r~.11 : I I I ~a,,nrv co~istants,dependent only on the species, and r is a parameter, w.
Q'2. PREDATOR AND PREY 259
I t is plausiblr t o assume as before that there is a limiting population II such that lations, and ia proportional to At; we write it asf(r, y ) At. What should we p t u l a t e
I f (,) = 0 and M ( y ) < 0 for y > v. If very small populations behave like the about !(I, g ) ?
unlimited growth model, we assum? hf (0) > 0. I t is reasonable that I(+, y) be proportional to y: twice as many cats will eat
twice as many mice in a small time period. We also assumef(z, y) ia proportiand
to r : if the mouse population is doublrd, a cat will come across a mouse tuice as
often. Thus we put /(z, y) = Bry, 6 a positive constant. (This assumption is 1-
plausibl~if thr ratio of prey to prrdators is very large. If a cat is placed among a
xufficientl? large mouse population, after a while it will ignore t h e mice.)
1. .\ population y (t) is governed by an equation The prey spwies is assumed to have a constant per capita food supply available,
sufficirnt to increase its population in the absence of predators. Therefore the prey
Y' = M(v)v. is subjrct to a differential equation of the form
Provr that:
-
(ti) equilibria occur a t y 0 and whrnrvrr ill (y) = 0;
( b ) tlie r~quilibriuma t y = 0 is unstable;
I' = A t - Rxy
In this \ca.v wr arrivr s t thr predator-prey equations of Voltem and Lotka:
I!,) un ccluilibrium t > 0 is asymptoticnlly stabbb if ant1 only if tllcsrr exists
6 > 0 Y U C ~ I that Bf > 0 on the illtrwal [E - t, f ) and hf < 0 on
+
cc, E *I.
2 Suppose the population of the United States obeys limited growth. Compute
the limiting population end the population in the year 2000, using the follou.ing This system has equilibria a t (0, 0 ) and z = (DJC, AjR). It is pas?, to see that
data: (0, 0 ) is a saddle, hence unstable. The eigenvalues a t (D/C, .4/B) are pun? i@-
nary, however, which gives n o information about stability.
Year Population We investigate the phase portrait of ( I ) by drawing the two Lines
19fd 150,697,361
1960 179,323,176
1970 203,184,772
r l r O !y" 0 r ' c 0 Therefore (Chapbr 8) ( z ( r ) ,v ( r ) ) is defined and in the boundary of that region;
y' < 0 y', 0 since z(1) is decreasing, ~ ( r = ) DIC. Thus the trajectory enters quadrant 11.
Similarly for other quadrants.
We cannot yet tell whether trajectories spiral in toward z, spiral toward a limit
,
/8 ---- - - - - - -- - 1 1 1
8'. 0
cycle, or spiral out toward "infinity" and the coordinate axes. Let us try to finda
tiapunov function H.
Borrowing the trick of separaliun of variables from partial differential equations,
*'>0 Y x;. 0 we look for a function of the form
y'c 0 Y '0
4
!
I
FIG.A
A s lcrllg as 1
Cu-D=s>O.
E J , z(1) is decreasing and y(1) is incren6ing. Hence
We obtain H - 0 provided
d 2'
;log x(1) = -Z = A - B y 5 -r,
Since r and y are independent variables, this is possible if and only if
d_logy(t) = = CZ- D 1 s. z dF/dz y dC/dy
dl Y -= -= Constant.
Cx-D By-A
Therefore
Putting the constant equal to 1 tie get
(2) < z(1) 5 ue-I,
C -
for 0 _< 1 < 7. From the sreond inequality of (2) we see that r is finite. From (2)
and (3) we see that for 1 E J , ( ~ ( f )y, (1)) is confined to the compact r e o n integrating we find
F (2) = Cz - D log 2,
O (y) = By - A logy.
Thus the function
By considering the signs of J H / J r and JH/Jy i t is easy to srr that the rquilibrium No matter what the numbers of prey and predator are, neither species will die
r = (DiC, A / B ) is an absolute minimum for H.I t follows that H (more precisely, out, nor will it grow indefinitely. On the other hand, vxcept for the state z, which
H - H ( z ) ) is a Iiapunov function (Chapter 9). Therefor? z is a shble equdi6riua1. is improbable, the populations will not remain constant
We note next that there are no limil cycles; this follows from Chapter I1 because Let us introduce social phenomena of Section 1 into the equations (1). We ohtain -.
H is I N I ~rnnstant on any open set. the follouing predalor-prey equatiuna o j s p i e s with l i m W growth:
\VV I l l l l l {)rOvr
T l l c o r c ~ n1 Er-rry trojrctory o j the I'olterra-Lotka equations ( 1 ) is o clnaed orbit y ' = (Cz- D -py)y
(crc,,pl lhc, r.qrrilibriurn z n n d the coordinate ares). The constants . I , R, C , D, A, r are all positive.
Proof. C ~ ~ ~ n i dar -point
r w = (u, u ) , u > 0, u > 0; Icr # z , Then thrrr is a Wc dividc. thc. upper-right quadrant Q ( I > 0, y > 0) into sectors by the two
scquencr . . . < 1-1 < 4 < II
,Ir~t11jIyi~~linit,, < . . . nurl~that +,.(w) is on tllr line liow
r = 1) 1 ', I ~ I I ~
L : A - Ry - Ar = 0 ;
t. --+ m as n - m ,
Along these lines t' = 0 and y' = 0, respectively. There are two pansibiiities, ac-
If i< llltt ill a closed orbit, the points +,.(w) are monotone along the line 1: = D/C
11.
cording to whether th& lines intersect in Q or not. If not (Fig. C ) , the predators
I ('t~:~~,t,.r 1 I ). Sinre tllerr are no limit cyclrs, either
die out and the prey population approaches its limiting value A/A (where L meets
the I-axis).
Siurc. If is constant on the trajectory of w, this implies that H (w) = H(z). But this
rol~tr~rlicts minimalit? of H ( 2 ) .
A/ x
FIG. C. Predalom - -
0 ; prey A/&.
This is because it is impossible for both prey and predators to increase a t the
same time. If the prey is above its limiting populatio~~ it must derrease and after
a while the predator population also starts to decrease (when the trajectory crosses
M).After that point the prey can never increase past A/A, and so the predntora
~l<,ii
\\I, II:IVP t l ~ fi,lln\~ing
. (schcmatic) phase portrait (I'ig. B). Thrrc.forr, for continue to decrease. If the trajectory crosspj L. the prey increases again (hut not
;III! r i ~ t initial
. ~ ~ ~)upulations(r(O), ~ ( 0 ) with ) r ( 0 ) # 0 , and y(0) # 0, other past :l/A), while the predators continue to die off. In the limit the predators dis-
111:<11 :. popuI:~tionsof predatcx and prey will rlscillatr cyclically. appear and the prey population stabilizes a t 11 /A.
FIG. D
Suppose now that L and M crow at a point z = (2, y ) in the quadrant Q (Fig. FIG. E
D ) ; of courser is an equilibrium. The tinear part of the vector field (3) a t r is
r-is - 3 q We alsa see that in the long run, a trnjectory either appronches r or ebe spirals
down to s limit cycle.
From a prnctic~lstandpoint a trajectory that tends toward z is indistinguishable
The characteristic polynomial has paaitive coefficients. Both mta of such a poly-
from z after a certain time. Likewise a trajectory that appronchea a Limit cycle 7
nomial have negative real pa&. Therefore z w asyrnplotually amble.
can be identified with 7 nfter it is sufficiently close.
Nntr that in addition tc the equilibria a t r and (0,O), there is also an quilibrium,
The conclusion ia that ony 4 of prddalora and p r q which obeys s p u a l h ( 2 )
n caddlr, a t the intenection of the line L with the z-axis.
d u a l l y selflea down lo ci(hcr a cunatunl m p e d u pqpuhlian. There we obadule
I t 1s not easy to determine the basin of 2 ; nor do we know whether there are any
upper Sounds fhal no populotia can e& in the long run, M m& whal the indid
limit cyclrs Nevertheless we can obtain some information.
popJolimu are.
Let L meet thc axis a t (p, 0) and the y-axis a t (0, q ) . Let r be a rectangle
whose corners are
(0, 01, (P,01, (0, Q), ( P , Q) PROBLEM
with p > p, > q, and ( p , @)E M (Fig. E). Every trajectory a t a boundary point
of r either enters r or is part of the bound*. Therefore r w porilwely iwarionl.
Every point in Q is contained in such a rectangle. Show by examples that the equilibrium in Fig. D can be either a spiral sink or a
~ Poincark-Bcndixson theorem the w-limit set of any point ( z , y ) in r, with
1 3 tllc node. Draw dragrams.
I > 0, > 0, must be a limit cycle or one of the three equilibria (0, O ) , z or ( p ,0).
W'r r~iluout ( 0 , 0) and ( p , 0 ) by noting that z' is increasing near (0, 0); and y' is
Inc.rl,:mlrll: near ( p , 0 ) . Therefore L.(r) is either z or a limit cycle in l". By a con-
sr.clurnl7rof the Poincd-Bendixson theorem any limit cycle must surround z. $3. Competing Spccia
WP ~ h ~ r r vfurther
e that any such rectangle r contains all limit cycles. For a
l i ~ l ~P>.cII,
it (Like any trajectory) must enter r, and r is positively invariant.
E'ixrrig (P,0 ) as above, i t follows that for any inild valuer (z(O), y ( O ) ) , here We consider now two species z , y which cotnlwte for a cornmon food supply.
errs18 I, > 0 m h U~at Instead of analyzingspec~ficeguations we follow ndrfferent procedure: weconalder
a large class of equations about which we assume only a few qualitative features. In
this way constderable generahty is gamed, and l ~ t t l eIS last b e c a w specrfic
One can also find eventual lower bounds for Z ( L ) and y ( L ) equations can be very diflicult to analyze
$3. COMPETING SPECIES 267
Thr equations of growth of the two species are written in the form
whrrr: g : [O, b] -r R is a nonncgativc C' map with g-'(0) = b. The function S is
(1) r' = M (E, y)r, positive to the lrft of v and negative to the right
Supposr fi and v do not intersect and that fi is brIoa v . Thrn a haw portrait can
v' = N(x, Y)V, bv founrI in a xtraightfrJrward way following mrthrds of the previous section. The
. c h r r ~the growth ratm M and N are C1 functions of nonnegative variables z, y. rquilibria are (0, O), (a, 0 ) and (0, b). All carhits tend to one of the three quilibria
Thr fi>llowingassumptions are made: but most tu thr asymptotically stabh. (aquilibrium (0, b ) . SIT Fig. B.
(11) If either species increases, the growth rate of the other goes down. Hence
dM t)N
-<O and -<0.
aY ar
( b ) If either popuIation iy very large, neither species can multiply. Hence
there exists K > 0 such that
Ji(1,y)50 and N(z,y)lO if t > K or v Z K .
(c) In the absence of either species, the other has a positive gmwth rate up to
a r ~ r t a i npopuIation and a negative growth rate beyond it. Therefore there are
a.c~n>t:llrIs a > 0, b > 0 such that
.If(1.0 ) >0 for t <a and M(z, 0 ) < 0 for x > a,
.l(O,y)>O for y < b and N(O,y)<O for y > b
I I T I I I . ~ II ~I I I I ~r x R I I I C T . ~ ~1t)e s e t I( = M ' (0) P X H C L ~ Yonce
I{? ( : I ) :tnd (I.) i.i~~.Ia
if 0 5 r 5 a and not at all if E > a. By (a) and the implicit function theorem fi
ir the ~(raphof a nonnegative C1 map j : [0, a] -+ R such that j-'(0) = a . Below
t I ~ rurvr
r p. M > 0 and a h v p it JI < 0 (Fig. A).
Let rul C p and vo C v be the open arcs of ordinary boundary points having p
s a common end point. If w U a consists mtirr.lj of inward nr entirely of outward
points of dB, we call p goodjor B ; otherwise p is bad for B. I t is essy to see that if
p is good for B, it is good for theother three bmic regions adjacent top, and similarly
for bad (Fig. D). This is because (z', y') rcvcnrs direction as one proceeds along
p or u p a t a prosing p i n t . Hence it makes sense to call a vertex simply gwd or bad.
Bad Good
FIG. U
Consider first of all the region Bo whose boundary contains (0,O). This is of type
I (I' > 0,y' > 0). If q is an ordinary point of n aBo, we can connect q to a point
inside 8 0 by a path which avoids v. Along such a path y' > 0. Hence (z', y') pointa
l o . 01
upward out of Bo a t q since p is the graph of a function. Similarly a t an ordinary
FIG. C
point r of v n dBo, (z',y') points to the right, out ul BOa t r. Hence BOis good, and
Thr boundary dB of a basic region B is made up of points of the following typen: so every vertex of Bo is good.
points of p 0 u, d e d vertices; pointa on p or v but not on both nor on the coordinats Next we show that if B is a basic region and aB contains one good vertex p of
axes, called o r d i ~ r yboundary points; and points on the axes. p il u , then B is good. We assume that near p, the vector field along dB points into
A vertex in an equilibrium; the other equilibria are a t (0, 0). (a, O), and (0, b ) . B ; we also assume that in B, z' < 0 and y' > 0. (The other caves are similar.) Let
At an ordinary boundary point w E dB, the vector (z', y') is either vertical (if rul C p, vo C v b arcs of ordinary boundary points of B adjacent to p (Fig. E). For
w E p ) or horizontal (if w E v ) . I t points either into or out of B since r has no example let t he any ordinary point of dB n p and q any ordinary point of m. Then
vertical tangents and v hss no horizontal tangents. We call w an inward or outward y' > 0 s t q. As we move along p from q to r thv sign of y' changes each time we cross
point of dB, accordingly. v. Thc. number of uuch crossing8 is men becaus~.r and q are on the same side of r.
The following technical result in the key to analysing equation ( I ) : Hence y' > 0 a t r . This means that (z', y ' ) points up a t r. Similndy, I' < 0 a t
every ordinary point of v n dB. Therefore along p the vector (i, y') points up;
L e m m a Lcl B be a basic region. Then Ihc ordinary boundary goink of B are ciUlcr along v ic p i n t s left. Then B lies above p and lejt of v. Thus B is good.
nU inward or aU oufword. This proves the lemma, for we can pass from any vertex to any other along cry
Proof. If the lemma holds for B, we call B good. starting from a good yertex. Since successive vrrtices belong to the boundary of a
Let p be a vertex of B where p and r cross. Then p is on the boundary of four common basic region, each vertex in turn is proved good. Hence all are good.
b w i c regions, one of each type. Types I1 and IV,and types I and 111, are d ' i o n d i y As a consequencc of the lemma, each baa& region, and iis closure, L eiUlet PEG
opposite paim. lively or negafiuely invariant.
FIG. F
hamining the equilibria for stability, one finds the following d t a . A vertex
where p and r each hsve negstive dope, but p is steeper, is s s y m p b t i d y stable
(Fig. G ) . One aeea this by drawing a a d rectangle with aides parallel ta the axes
around the equilibrium, putting one corner in each of the bur adjacent regions.
Such a rectangle is positively invariant; nince it can be arbitrarily small, the equilib
rium is asymptotically stable. Analytically this ia expressed by
slope of %, - --
M=<slopeofv =
M.
- N*
-
N, < O1
where M, = dM/dz, M, -
dM/dy, and ru, on, a t the equilibrium, from which rr
computation yielda eigenvalues with negative real partu. Hence we have a sink.
FIG. E
\Vhat are the possible e l i n i t pointa of the flow ( I ) ? There are no closed orbite.
For n closed orbit must be contained in a b ~ i region, c but this is impossible since
z ( 1 ) and y (1) are monotone slang any mlution curve in a basic region. Therefore
aU -limit pointa are equilibria.
We note also that each trajectory is defined for aU 1 2 0,because any point
lies in a large rectangle r spanned by (0, O), (zo,O), (0,yo), (G,yo) with q > a,
yo > b ; such a rectangle is compact and positively invariant (Fig. F). Thus we
havc shown :
exish and is one of a fin& number of eguilibria. A case by CW study of the different ways p and * can cross ahows that the only
other ssymptotidly atable equilibrium is (b, 0) when (b,0) is above p , or (a,0 )
Wc conclude that lhe populaliona of two competing species always lnul lo one of a when (a,0) is to the right of v. All other equilibria are unstable. For enample, q
.fin& nurtiber of limaling populaliona. in Fig. H is unstable because arbitrarily near it, to the kft, is a trajectory with z
272 12. ECOWCY $3. CO\IPETISG SPECIES L ?3
drcrcilsing; such a trajectory tends toward (0, b ) . Thus in Fig. H, ( 0 , b ) and p Xotr that both populations are positive at p. Suppose that some unusual event
arr asymptotically stable, while q , r, s and (a, 0 ) are unstable. Note that r is a occurs, I I I I ~ilccountrd for hy our model, and the state of the ecology c h a n p sud-
source. d m l y from rqo to 'I. Such an w e n t m l ~ hb~
t introduction of a new psticide, importa-
There musf be a1 leas1 one mymplotuatly alable equilibrium. If (0, b ) is not one, tion of additional mcmhcrs of onc of thc spcrirs, a forest fire, or the like. .\lathe-
then it lies under p ; and if ( a , 0 ) is not one, it lies over r . In that case p and r cross, matically the rvcnt is a jump from the hasin of p to that of (0, b).
and thr first crossing to the left of (a, 0 ) is asymptotically stable. Sucll a change, eve11 though quite small, is an ecological catastrophe. For the
Ilvc*ry trajectory tends to an equilihrium; it is instructive to see how these trajectory of ur has quite a di Rerent fate: i t goes to (0, b ) and the E speeies is wiped
a-limits change as the initial state changes. Lpt us suppose that q is a saddle. Then out!
it can br shown that exactly two trajectories a,a' approach q, the so-called slabk Of course in practical ecology one rarely has TiK.H to work with. Without it, the
nlangolds of q, or sometirnea separalrices of q. We concentrate on the one in the
change from q to ul does not seem very different from the insignificant change from
unbounded basic region B,, labeled a in Fig. H.
uo to a near state u2, which also goes to p. The moral 4 c l ~ a rin. the absence of com-
prehensive knowledge, a deliberate change in the ecology, even an apparently minor
m e , is a very risky proposition.
1. The equations
2' = x(2- I- y),
Note*
Periodic Attractors Theorem 1 Let 7 be an aaymploLakdly sloble closed orbil of penid A. Thtn 7 has a
neighborhood U C W utuh lhal awry pard o j C haa aympldic perid A.
Proof. Let U be the open set Cr, in the definition of asymptotically stable with
W o = U1. Let z C U and fiT e > 0. There exists 6, 0 < 6 5 r, such that if z € 7
and 1 y - z I < 6, then I + ~ ( y )- + A ( z ) ~< e (by continuity of the Row). Of eourae
+r(z) = z. Since d ( + , ( z ) ,y ) + 0 as 1 -r m, there exista 1, 2 0 such that if L 2 ,,f
there is a point 2 , E 7 such that I +,(z) - r , I < I. Keeping in mind #A(z,) = r ,
we have for L 2 b:
-
a b u t the vector field Nevertheles ~tE ol'grrat ~ n i p o n ; i ~ ~ r r .
Let f: IV Rmbe a C' vector field on an open set W C R";the flow of the dif- Theorem 2 Let 7 be achedorbilojpcrimi A ojllredyomicolsydrm ( 1 ) . L c l p E 7 .
frrrntial equation Suppose lhal n - 1 o j fie cige?umlw of lht linear map D k ( p ) : E + E are lcds than
0) x' = j(t) 1 in abmLulc d u e . Thm t is oayllploficdly alable.
i~ 1I8.11otrfI11y +I.
%me remarks on this theorem are in order. First, it aasumes that $1 is diffemti-
1 . ~ 1 r C If- be a closed orbit of the flow, that is, a nontrivial periodic solution
able. In fact, +,(z) is a C function of (L, z ) ; this is proved in Chapter 16. Second,
curvv JVc call r aay~~tp/o/ically dfuble if for every open set Ul C U', with Y C UI
the condition on D+l(p) is independent of p E 7 . For if q E 7 is a different point,
tllr,rr Ir r i l l ope11 set i . ~ Y, C L'I C U, B U C ~that +,(U1) C 9 for all 1 > 0 and
let r € R be such that 4.(p) = q. Then
The elpenvalue condition in Theorem 2 is stronger than asymptot~cstability. space" of somr sort, then g(z) is the state of thr system 1 unit of time after it is in
If it holds, we cdl 7 a periodic @ r a c k . Not only do trajectoriee neer a state x. After 2 units of time it will be in state g2(x) = g ( g ( z ) ); after n units, in
attractor 7 have the enme asymptotic period M 7, but they are ~ y r n p t o t i c d y state g m ( r ) .Thus instead of a continuou~family of states j&(z) I t f R) we have
"in ~'hasr"u-ith 7. This is stated preeieely in the following thwrem. the discrete family [ga(z)I n E Z I , where Z is the set of integers.
The diReomorphism might be a linrar owrator T: E + E. Sueh systems are
s t u d i ~ din linear al~r.hra.We get rather eomplrte information a b u t their s t ~ c t u r e
Theorem 3 Let -, be a periodic &actor. I f lim-. d(+,(z), 7 ) = 0, Ihen Ullre
1 +,(I) - ~,(z)l 0.
ia o unlpue point z E 7 arch thnf h+, - from the ranonical form throrrms of Chaptrr 6.
Suppo~eT = eA, A F L ( E ) . Thrn T is thr "time one map" of the linear flow
cU. If this continuous flow elA reprewnts some natural dynamical process, the
This means that any point &ciently near to 7 has the same fate as a defurite
point of 1 . discrete flow T' = em*is like a series of photographs of the process taken a t regular
I t can be proved (not cadsily)that the closed orbit in thk Vm der Pol oscillatar time intervals. If these intervaki arc very small, the discrete flow is a good appmxi-
mation to the continuous one. A motion picture, for example, is a discrete flow
is a periodic attnctor (see the Problems).
The proofs of Theorems 2 and 3 occupy the rest of this chapter. The proof of
Thmrrm 2 depends on a locd section S to the flow at p, analogous to thoee in C h a p
ter 10 for planar flow-s: S is an open subset of an ( n - 1)-dimensional subpaee
that is hard to distinguish from a continuous one.
-
The analogue of an equilibrium for a discrete system g: E E is a fir& poid
Z = g ( i ) . For a linear ojxrator T, the origin is a fixed point. If there are other
transvcrsr to the veetor field a t p. FoUowhg trajectories from one point of 3 to fixed points, they are eigenvectors belonging to the eigenvalue 1.
another, defines a C1 map h: So-+ S, where Sois open in S and c o n k h a p. We call We 8hall be interested in stability propertie8 of fixed points. The key e u u n p l ~is a
h t h i'nrnrorL
~ map. The following section studies the " d i r e t e dynamical aystem" linear coir!raction: an operator T F L(E) such that
h : So-+ S.In particular p E Sois shown to be an ~ y m p t o t i c d ystable hxed point
of h , and this w i l y impties Theorem 2.
for all z i E. The time one map of a contracting flow is a linear contraction.
PROBLEM
Proposition The jollotcqings€olements are equivakrrt:
Let y be a closed orbit of period h 7 0 in a plsnar dynamical system z' = .f(z). (a) T is a linear confractiot~;
L e t p E 7. (b) the etgenvalues of T have absolure d u e s less Iha7r 1;
( a ) If
I Det D+h(p)l < 1,
,,
(c) them is a norm orr E , and < 1, such /ho(
is a C map g: W -
An important example of a discrete dynamieal aystem (preciee definition Ister)
W on an open wt W of vector space which haa a C' inverae
g-I: W + W. Such a map is cdled a difleamarphiam of W. If W represents a "atate
zr f VA,then define I z 1 = maxi/ 1 ) . Thus we may replace Ec h\. VA, or
what is the same thing, sssume that T has only one eigenvalue A.
A aimilar argument reduces us t o the cam where the Jordan form of Tc hay only
13. PERIODIC AlTRAC1URB $3. S T A B I L I T Y A X D CLOSED ORBITS 281
orrr +*l(.m~,nt;iry
Jordan block for all I i E. Let 0 < r < 1 - r. RJ- Taylor's thmrem tbrre is a neighborbood
1. C M' of 0 so small that if z f t., then
Tbc preceding argument can be slightly modified t o show tbat in the specified
norm,
I ~ ) - ~ Y I I - Y Ir < l ,
7'11is \\.as proved in Chaptpr 7. Give F.'c the max norm for this basis.
for all I,y in some neighborbood of 0 in H'.
\vht,n. at. . . , o. are arbitrary complrx numbers. Then if I A I < 1 and t is suffi-
cir11tl.vsniall, ( c ) is satisfied This completes the proof of Proposition 1. 53. Stability and Cloned Orbile
-
Let g"(z) =
Since . --, 0. A. -, A
I (the period of 7 ) . Thus there is an upper bound r for Define
II A. II 11 Z 01. By continuity of the flow, as n m , Ul = [ + , ( t ) E B", 12 0 ) .
See Fig. C. Then U, is a neighborhood of y which is positively invariant. Moreover
Ul C U. For let y C UI. Then y = +,(I)for m e z E B,, 1 1 0. We scnrert that
unijomrly in 8 [O, r]. For any 1 > 0, there exist 8(1) t [O, r], and an integer (1, z) can be chosen so that 0 < I 5 r ( z ) . For put pm(z)= +. Then z. E V for
all n 2 0. There exista n such that y is between 1. and z.+~on the trajectory of
z ; since z. E V , +(z.) < 2h; and y = &(z) = Q,(z.) for 0 5 I < 2A. Then y E U
because z. E N.
-
Finally, d(+,(y), y ) -+ 0 as 1 m for all y 6 U. For we cnn wirite, for given y,
The following reault links the derivative of the Poincad map to that of the fh.
-
We keep the same notation.
D ~ ( 0 1= D+r(O)l H .
$3. STABILITY AND C W S E D ORBITS
If ~ . ~ ( isz defined
) and sufficiently near 0 for n = I, . . . , !ithen
,
where
Hence if t is sufficiently small, the sequence +.~(z) stays near 0 and can be mn-
-
P r w j . 1,rt r : So R be the CLmap such that r(0) = k and g(z)
By the remark at the end of Section 2, Chapter 11, we have
= O(+(Z),2)
tinued for all positive integers n, and the above inequalities are valid for all n. It
follows that the sequence It.) is Cauchy and converges to some a t R. Thus +~.z()
converges to &(O) = z f 7. This implies Theorem 3 of Section 1.
PROBLEMS
Since I)o~(O)( H ) = H = Ker h , Dr(0) -i 0. Hence by the chain m l e
Thus Euclidean three space, tbe configuration spacc of onc. body, is a thrce- We deal thrn with t h r space of noncollision stat- which is ( M A) X Y. -
diinrnsionsl vector space together with an inner product. Newton's equations are second order equations an .ti - A which may be written
Tlu, rolifiguration space d l for the it-body problem is the Cartesian product of m.?,=-grad.V(r) for i = l , . . . , n.
.
E \ritll it-~,lf11 tinlcs; thus 111 = ( E ) ' and r = ( r ~ ., . , x.), where I, E E is thc
~>osirl<,~l o f the ith body. S o t r that I, denotes a point in E, not a number. Here the partial derivative D.V of V with respect to I, is a map from JI - A to
0 1 1 , . nlil? drducc thr space of states from the configuration space as the space L ( E , R ) ; then t h e innerproduct on E conrerL~D,V(x)t o a rertorwhtch weeall grad,
7',, of all t:ingcSntvectors to all possible CUNPB in h l . Onr may think of T M as the V ( r ) .T h e process E similar to the defin~tionof gradierlt i r ~Chapter 9. Thus tlre
l r t X .If and represent a state a s ( I ,v) t di X hf, whrrc I is a configura-
~ ) r o < l ~.lI rquatlons make sense as written.
tior! as bc.fa~rv and r = (o,, . . . , v . ) , u, t E being the velocity of t h r ith body. A One may rewrite Newton's equations in such a way that they k o m e a first
stat,, of thv s-stem gives complete information about the systrm a t a given moment -
order system on the space of states (,If A) x I f .
and ( a t Irast in classical mechanics) determines the complc*te lift history of the
stat(,.
'1'111. dvtc.rinination of this life history gors via the ordinary diRcrcntial equations m,v, = -grad, V ( s ) . for I = . . .~ t .
of motion, Xewton's equations in this instance. Good insights into these equations
can b~ obtained by introducing kinetic and potential energy. The flow obtained from this diflcrential equation then determines how a state
T h e kinetic energy is a function K : hi X hi 4R on the spacc of states which moves in time, or the life history of the n b o d i e o n n thrir positions and velocities
is givrn by are given. Although there is a vast literaturr of sevrral centuries on these equa-
I " tions, no clear picture has emerged. I n fact it is still not cvrn clrnr what the basic
K ( I , u) = - C m. I u. 1.' qurstiona are for thin "problem."
2 i-1
%imc*of thr clurstions that h a v ~bern s t u d i d irlrludr,: I s it true that almnat all
HI^, 1111 110)rnrof u , in thr Euclidran norm on E. Onr mav also consider K to be stntca do not h a d to collisions? To what rxtrnt are periodic ~olutionsstable? How
givvrl cIirw.tl>.by a n inner product B on hf by to show the existenec of periodic solutions?How to r r l a t ~t h r thcwry of the n-body
problem to the orbits in the solar systcm?
Our present goal is simply t o puc Newton's equations into the fmnrrwork of
this book and to see how they fit into the more abstract fran~eworkof Hamiltonian
mechanics.
K ( x , u) = B(v, v). We put the n-body problem into a little more gtsneral setting. The key ingrcdi~nts
I t is clear that B defines a n inner product on hf where (u., w.)means the original are:
inner product on E. Configuration space Q, a n open set in a vrrtor spacr E (in t h r above ease
(1)
T ~ pofolfial
P energy V is a function on M dcfined by Q = hi - A a n d E = M).
(2) A C function K . Q X E 4 R , kinetic enrrgy, such that K ( r , v) has the
form K ( r , a) = K.(v, a ) , where K. is an inner product on E (in the above
case K, was independent of 3, but in problems with constraints, K. de-
\VI, silpposr that the gravitational constant is I for simplicity. Note that this pends on I).
functioi~is not defined a t any "collision" (where I, = I,). Lct A,, be the subspaee of (3) A (? function V: Q + R , potential cnergy.
collisions 111 the ith and jth bodies so t h a t
T h e triple (Q, K, V) is called a simple mechanical syalem, and Q X E the stale
A,, = { rE Ai I r. = z j l , i < j. apace of the system. Givcn a simple mechanical system ( Q , K, V) the energy or
Thus A,, is a linear subspacc of the vector spacc Al. Drnotc the spacct of all collisions total cnrrg). is thr function e: Q X E -. R d r f i n d by e(r, v ) = K(I, v ) I.'(r). +
I)? A C .I1 so tlwt A = U Thctn proprrly nprakinp;, the ,lt,tnnin 111 thr pntcntial For a simplr ~nochanicalsystem, one can canonically define a vector field on
,.nr.rK! i.: 11 - A : Q X B wl~ichgive-s the rquations of motion for thc statcq (points of Q X E ) . W e
1': A1 - A + R.
will sce how this can be done in the ncxt section.
I < X ~ I I I I I ~ofI ~ simple
: mechanical systenls beside the n-body problem include a It can be shown that every avmplectic form is of thiri type for some repwen&
p n r t i r 4 ~rlloving in a conservative central force field, a harmnnic oscillator, and a tion of F ss E X EL.
fr~rtlor~l~~s:. pendulum. If one r x t ~ n d sthe definition of sitnplc mechanical systems Now let U be an open subset of a vector space F provided with a symplectie
to p<.rniit Q to be a manifold, then a large part of classical mechanics may be form n. There is a prescription for assigning to any (? function H : U --9 R,a CI
analyzed in this framework. vector field X H on U called the Hamillonian u c l m JWof H . In thie eontext H m
called a H a m i l h i a n or a Hmiltonian function. To obtain X H let DH: V -r P
be the derivative of H and simply write
92. I l u t ~ ~ i l t o n l aMechanim
n
setting, 11 plays the role of energy, and the solution curves represent the
i l l tl~i-.
Then set
nlotlorrs of states of the s ~ s t e m . 0 ) = (9,A,(u)).
/I. I -
T l l r o r r m (Conservation of Energy) Let U be on open set o j a vector s w e F,
sul~rli~tt
: ~ II a symplectic jorm on F. Then H is conslant on Ihe
R arty Cv j u ~ ~ c l t oattd
crrrres defitrerl by /be wclor field X u .
K ( q , r') = inr / u 1.' Thrn A: E X E -
tial e n r r p 1'. I n this rasp s t a t r space is E X E and K: E x E -
moving in Euclidean t h r w space E under a consrrvativc. force firld given by poten-
R is given by
E X E* is given h,. A,(u) = p f E*, \vhere
p(11') = 2 K o ( i t ,10); o r
Proof. If + , ( r )is a solution curve of the vector f i ~ J dX,r, then it has to h shown p ( t u ) = m (e, s )
that and ( , ) is thr inner product on E. In a Cartesian coordinate system on E, p =
mr, so t h a t the image p of u under A is indwd the classical momentum, "conjugate"
to v.
Rrturning to our simple mechanical s>.stem in grneral, note that t h e Legendre
l'lrib t.sl~rl.fisionby the chain rule is transformation has a n inverse, so that A is a diffeomorphism from t h e state space
t o the phase space. This permits one to transfer the e n e r e - function o on s t a t e
space to a function H on phase space called t h e H u ~ l t i ~ l o l of
i a a~ simple nrechanicd
sysfem. Thus 13.e havr
B u t U H ( X H ) is simply, by the definition of X H ,~ ( X H , which is 0 since n is
XH)
antis) rn~rrvtrir.This rnds thr proof.
lilwar i % o n ~ u r ~ h A,: -
b,v tlrv I.r!/?rldre lransjormation A: Q X E + Q X E*. To d c f i n ~A, fitst drtine a
i s nE~ E*, for each q E Q, by
A,(v)u* = 2K,(u, w ) ; u E E, IV E B
Ia\3-sdo not dirrctl~- apply (such asaspinning top), b u t which fit into the framework
of "sin~plr.mctllhniral systems," especiall!. if the conliguratic~ns p c c is allou,ed
to bra a surfact*o r higher dimemionel manifold. For many such s!-stems, Hamilton'a
wuations havr been verified experimentallj.
It is meaningleas, however, to tu to deduce Hamilton's q u a t i o m from Newton's and show that H is constant on orbits. The critical points of H are s t a = 0,
on the abstract level of simple mechnnical eystem (Q, K, V). For there is no f(z) = 0 ; use H
. = j ' ( z ) , H.. = 1.)
identification of the elements of the "configuration npace" Q with m y particular
physical or geometrical parametera. 4. Consider the equation
Cuns~drra s an example the apecial caw above where K ( q , v ) = t C mu,' in
+
Curtral~lrlcoordinates. Then m,q = pc and H@, q ) = C ( R 2 / ~ ) V ( q ) ;Hamil-
2 + g(+ + j(l) = 0,
ton's ~qrir~tlonsbecome where g ( r ) > 0, and j is as in Problem 3. Describe the phase portrait (the
function y may be interpreted as coming from friction in a mechanical problem).
Pi1 "
av .
--
3%
Diflerentiating the first and combining them equatiorur yield One modern approach to mechanics is Abraham's book, Faundolions o j M&nica
[I]. Wintner's A d y t i c d Fmndotiona oj Celestial Mechnlur 1251 has a very ex-
tenaive treatment of the n-body problem.
These are the familiar Newton's equations, again. Convenrely, Newton's equations
imply Hamilton'e in this csse.
PROBLEMS
1. Show that if the vector spsee F has a aymplectic form U on it, then F hrs even
dienaion. H i d : Give F m inner product ( , ) m d let A : F -,F be the operator
defined by (Az, y) = fl(r,y). Consider the eigenvectors of A.
2. (Lagrange) Let (Q, K , V) be a simple mechnnical system and X N the ~ a a e i -
ated Hamiltoninn vector 6eld on phase npace. Show that (q. 0 ) is an equi-
librium for X x if and only if DV(q) = 0 ; and (q, 0) is a atable equilibrium if
9 ia an isolated minimum of V. ( H i d : Uee consewation of energy.)
3. Consider the second order differential equation in one variable
r+f(z) =o,
where j; R 4 R is C and if j ( z ) = 0, then j l ( r ) # 0. Deacribe the orbit stmc-
ture of the associated system in the plane
* = v
8 = -f(+)
whm j ( x ) = z - 9.Discm thin p h a ~ p o r t r a iint general. ( H i d : Wder
Wr call the function f ( t , r ) Lipschilz i n r if tlicrc. is a constant K 2 0 such thet
I f ( 4 1 1 ) - f(1, dl 5K 111 - 19 I
for all ( I , 11) and ( 1 , I S ) in IV.
The fundanil*ntallocal thwrcm for nnnautonomous equations is:
Nonautonomous Equations
Theorem 1 ].el W C R X E be open arul f : IV 4 E a continuow m a p thaf ia
and Dierentiability of Flozus 1,ips~hitzill r. If (6,rb) E W , t b r e is art open inlerual J conloining 1 and a unique
sdulion to ( I ) dejned on J .
The proof is the samc as that of thr fundamental theorem for autonomous equa-
tions (Chaptrr 8),the extra variable L be in^ insrrtrd a h r r e appropriate.
Thc theurrm applirs in particular to functions f(1, r ) that are CL,or even con-
tinuously difirrrntiablc only in r ; for such an f iy locally Lipsehitz in r (in the
obvirms scnsi.). In particular we can prove:
V' = g(1. V ) ,
298 15. NONAUTONOMOU8 EQUATION8 AND DIWERENTIABILITY OF PLOWS $2. DIFFERENTIABILITY OF THE ww OF A U T O N O H O U ~ EQUATIONB . 299
x-.
L*
respeclive/y, on rome infemd J, and ~ ( 4 =
) ~ ( 4 ) ~ where f is assumed C'. Our aim is to show that the flow
1 z(t) - v ( 9 l l$ (exp(K I t - 4 0 - 1)
defined by ( I ) is a C' function of two variables, and to identify a+/&.
for all 1 t J. To this end let y(1) be a particular solution of ( I ) for t in some open i n t e r 4
J . Fix 4 E J and put y(4) = yo. For each 1 f J put
Proof. For t E J we have
I[:KI~(~) - Y(~)ILC/'.&. be the solution to (2) which sends 4 to €.If E and yo + E E W, let the map
u(l) + ' < ' e x p ( K l t - b l ) , This meana that for every r > 0, there exists 6 > 0 such that if I E I Id, then
K-K
which yields the theorem.
f o ~all t t Jo. Thus as E -+ 0, the curve t + y ( 1 ) + u(L, E) is a better snd better
+
approximation to y(f, 5). In many applications y(t) u(t, E) is used in pbee of
(2. DilTercntiebility of the Flow of Autonomous Equations y (1, E) ; this is convenient because u(t, [) is linear in 5.
We will prove the presently. First we use (3) to prove:
Proof. Of course a+(!,z ) / a l is just f ( + , ( z ) ) ,which is continuous. T o compute The Taylor estimate for f says
d+/dr \\-r2 have, for small
j(y) -f (2) = Df ( I ) ( y - r ) + R(z, Y - 4,
d ( l , yo + f ) - +(L, Yo) = V ( l ,f ) - ~ ( 1 )
limR(z, y - , ? ) / I y - z ( = O
Thc proposition now implies that a4(1, a)/ax E I,(@ is the linear map -+ u(L, f ) . 9-Y
Thc continuity of d+/dr is a consequence of the continuity in initial conditions and
data of solutions to the variational equation (2). uniformly in y for y in a given compact set. We apply this t o y = y(8, E), z = y(8) ;
Drnoting the flow again by + , ( I ) ,we note that for each 1 the derivative D+t(r) from linearity of D f ( y ( 8 ) ) and (4) we get
of t l ~ rrl1;11) +, a t I E W is the same as a + ( / ,r)/d+. W e call this the space der~ualiw
of t h r f l o ~ v ,as opposed to the lime derivdive z)/dl.
TII(. proof of thc preceding thmrem actually cornputm & , ( a ) as the solution
to a n initial value problem in the vector space L ( E ) : for each r0 t W fhe apace
derit'alive of /he pow satisfies
N 1 € Jal.
-
= maxlJIDf(y,s)ll 8
Il,.rr. L\-P rcgzrrd roas a parameter. An important special case is that of an equilibrium
i so that +,(q i . Putting D f ( i ) = A E L ( E ) , we get
Then from ( 5 ) we get
if 1f 1 I a, and
E Ja. (I
Assume now that ( 6 ( 5 61. From ( 6 ) , (7), and (8) r e find, for i C do,
for some constant C depending only on K and the lrngth of JO.Applying Gronu-all's
$2. DlFPEHENTlABlLITY OP THE PLOW OP AUTONOMOUS EQUATIONB 303
or equivalently,
(9) 2' = f ( x ) , up = Df(2)u.
S j c e F k C-I, the Bow 9 of (9) is O n .But this ia just
aince the aecond equation in (9) M the vsrktioarl equation d the first equation.
Therefore a+/ax is a Cr' function of (1, z), &w &/ax = D+,(z). MOreover
@It31 1s Cr-' (In fact, C
' in t) since
I t follows thst 4 is C' ainee its firat partid derivstivm are O'.
51. PI:IlSlSTESCE O F EQUILIBRIA 3G
Ij(r)--g(z)I and IlDf(z)-Dg(~)Il The proof of Throrem 1 has nothing to do with differential equations; rather, it
depends on thr following result about C' maps:
arr small for all r F W .
T,, 1n:lht. this nlrirr P f ~ ~Irti V(1V) ~ ~ , be the set of all C1 vrctor fields on W. If
F,' h:!. :I 11, ~ I I I I . 11-\-r-drfinr thv C1-norm 11 h 111 of a vector field h E v ( W ) to be the Proposition L e t / : W 4 E k C 1 and suppose r, 6 IV is such t h d the l i t t a r operalor
1 ~ : t . t 11111)c.rI N I U I ) ~of all thr nunihrrs DJ(so): R 4 E rs invertible. Then there is a r~eighborhuod 3L C V(W) of J and an
opor sel 1' C W conlaintng r o such lhal rf y t X , Lher~
(a) [j I 1 is one-10-OILP,
and
\ .a, :~il,n\111cpossibility ( 1 h 11, = rn if thrse numbem are unbounded. (h) f(10) c g(U).
gk. PERslSTENCE OP EQUILIBBIA
-
Theorem 1 follows by taking z. Z and/(*) = 0, for then g(y) = 0 for a unique
0 t C'. To make I p - z I < e (assuming E is nonned now) we eimply replace W
Therefore
u = Df(y)-'(Dj(y)u); ) w Is
H ( Y ) = i I ~ ( Y-
hence takes a minimal value. Note that yo cannot be in dB, for if y E aB, then
l u l 5 l l D ~ ( Y ) - ' II I D f ( ~ ) uI.
so, from ( I ) , Ij(Y) - w I ? Ij(Y) - j ( ~ a ) l- lj(-%) -wI
Iu I > 2a - 6.
(3) I DJ(y)(u)l > T- Hence
) w l > a > I/(%) - w l ,
I ~ ( Y-
+
Now let y, 2 be dietinct points of V with r = y u. Note that since V is a ball, showing that I j ( y ) - w I is not minimal if y E aB.
y + t u E V f o r a l l t E [O, 1). DefineaCmapq:[O, 1 1 - + E b y Since the norm on E comes from an inner product, I z 1' is differentiable; its
q(L) = j ( y e tu). derivative a t I is the linear map z -+ (x, 2). By the chain rule, H is differentiable
and its derivative a t yo is the linear map
Then
4 ) =j(y), *(l)=j(z).
Since ye is a critical point of H and yo in an interior point of B, DH(y,) = 0.
By the rhain ~ l e , Since Dj(y0) is invrrtible, there exists u t E with
v'(L) = Dj(y + tu)u. D ~ ( Y o )YO)
~ - w.
Hencr Then
0 = DH(yo)u
= V ( Y-
~ w, YO) - w)
= I j(@) - w 1'.
Therefore j(y0) = w, proving Lemma 2.
$2. PERSISTENCE OF CLOSED ORBIT8
308 16. PERTURBATION THEORY AND STRUCTURAL BTABILI~Y
We can find still smaller neighborhoods, XIC X* of f and N I C I V ~of 10, such that
if g 6 XIand y, r C N I , then In this section we consider a dynamics1 system 4, defined by a CLvector field
f : W + E where W C E is an open set. We suppose that there is a c l d orbit
r C W of period A > 0. For convenience we assume the origin 0 t E in in 7. The
main rwult is:
If I<,iiorrafrarnLnrlfna 1 lhal for arry ball V C N a n d any g t 318, g I V is one-lo-
" <. Theorem 1 Let u : SO+ S be a Pmneard m a p for a loud section Sat 0. k t U C W
1.'1x II 1>1111 1' C S1around
Let B C V be a closed ball around r o and choose
20. be a neighborhad of 7. Suppoae lhat 1 is not a n eigenvalue of Du(0).T b m b e eriab
6 > 0 as I I I 2.aThere is a neighborhood n C aroff such that if g t 32, thcn
1~~mm a & l t b m h w d n C ~ ( W ) o f f such that meevery u e c f m f i e l d g ~n h a d o s c d a b i c
BC U .
mini I g ( y ) -~(Io) 1 Iy t dB1 > 26 > 0 .
It f,,ll,~rvsthat if I w - g(ro)l < 6 and g E n , then w t g ( B ) . The proposition is The condition on the PoincarC map in Theorem 1 is equivalent ta the condition
noa p r ~ , r f ~using
d this Iand taking U = V. that the e~ut=n\-alue10fD4~ ( 0 )ha< nlulti])l~c~ty
1. Cnfortunately, no equ~valent
condition on the rector fieldf IS known.
We have not d i s c u d the important topic of nonautonomous perturbations. Proof of the theorem. Let T: + R be the C map such that r ( 0 ) = k and
Problem 2' shows that in a certain sense the basin of attraction of a sink persisrs
under small nonautonomous ~erturbations. 44 = *.<S)(X).
3 10 16. PERTURBATION THEORY AND STRUCTURAL STABILITY
We n i a j assume that the closure of So is a compact subset of S. Let a > 0. There Hc,ncr 1has a unique zrro y E So;and y lies on a closrd orbit @ of g. Xloreover, we
exists bo > 0 such that if g t X(W) and I g(x) - f ( t ) l < L for all z t 8, then, can nlalir y so close to 0 that 0 C U.This proves Theorem I.
fiwt, S will be a local section a t 0 for g, and second, there is a C' map 0 : & + R
such that The question of the uniqurness of the closed orbit of the perturbation is interest-
ing. I t is not necessarily unique; in fact. it is poasible that all points of U lie on
l-(x) - ~ ( 4 < 1 a,
closd orbits off! But it is true that closrd orbits other than y will have periods
much b i g ~ r rthan y. In fact, given t > 0, there exists 6 > 0 so small that if 0 <
and d(r, 7 ) < 6 and + , ( I ) = I, 1 > 0, then 1 > 2A - r. The same will hold true for
sufficiently small perturbations of 7 : the fix& point y of v that we found above
where $, is the flow of g. lies on a closcd orbit 0 of g whose p ~ r i o dis within t of A ; while any other closed orbit
of g that nltLctsSOwill have to circlc around 4 seaeral limrs before it closcs up. This
I'u t
sections; see Fig, A.
fnllo\r.s fro111t l ~ rrc.lntion of ckjstd orbits to 1111%
$.(x>(x) = u ( 4 .
Thrn
v:&-+S
Is n ( " Innlr l r hich is a kind of Poincar6 map for the flow $,.
C;ivvri a n $ 1, > 0 and any compact set K C W, and any r >0 we can ensure
thnt
I I Db(1.1 - m , ( ~ ) l <l v
for all t E [- 4 41, I C K, provided we make (1 Q - f 11, small enough. This follows
from continuity of solutions of differential equations a s functions of the original
data and initial conditions and the expression of dJ.,(z)/dx as solutions of the
nonautonomous quation in L ( E ) ,
-
so that ~ ( 0 ) 0. Now
[(I)
DE(0)
= U(Z)
= Du(0)
-I
- I.
ia entirely contained in V . I t is possible, however, for every neighborhood of a hyper-
bolic closed orbit to i n t e m t other closed orbits, although t b is hard to picture.
We noup state without proof an important approximation result. Let B C R.
whcrr I . M -+ H is the identity. Since 1 is not an eigenvalue of h ( 0 ) we know br a closed ball and dB its boundary sphere.
that 0 i.; nr~tan eigrnvalue of D[(O), that is, D((0) ia invertible. From the proposi-
tion In t11r j~rc.cedingsection we can find a neighborhood C V(So) o f t such that Theorem 2 I,tl W C R" be a n open set containing B and f : W 4 R. a C vedm
IIUlI;I ulll(illr5zrro y S,,.If I / g - f 1 1 , lc; sufficiently small, 1, { 3n.
;In)- I I I : ~ ~1 1 1) ,? jieM which is lransuerse lo dB al every pmnl of dB. Lel 32 C V ( W ) be any n&hborhood
$11' 16. PERTURBATION THEORY A N D STNULTVHAL STABILITY
and
(a) if
(b) if
zE B id an equilabrium of g, then r is hyperbolic;
y C B is a closed orbil of g, fhen y is hyperbolic.
aD.
-
= (IER"[IEI =]I.
Cotwider C' vector fields f : W R" defined on some open set W containing D-
such that (1(z). z ) < O for each E in 8D".Such an f is called atmdurally alobk on
The condition that f be transverse to r3B is not actually necessary, and in fact, B D" if there exists a neighborhood 32 C V(W) such that if g: W -+R"is in 32. then
can be replaced by any compact subset of W. flows of fand gare topologically equivalent on D".This means there exists a borne+
morphism h. D' 4 D" such that for rach r 6 D m ,
h ( l + , ( ~ )1I 2 01) = I h ( h ( r ) ) l 1 2 01,
where $, is the Row of g; and if z is not an equilibrium. h preserves the orientation
of the trajectory. (The orientation of the trajeetor). is simply the direction that
SIIO~ t l ~ tthc rigrnvalur condition in the main theorem of this section points move along the curve as 1 increases.)
1.
is nccrssar>-.
This is a very strong condition on a vector field. It means that the flow g, can-
not have any "exceptional" dynamics1 features in D'. For example, it can he shown
2. Let y br a periodic attractor of r' = f(x). Show there is a C' real-valued that if t E int D' is an equilibrium, then i t must be hyperbolic; the basic tenson
function V(r) on a neighborhood of y such that V 2 0, V-'(0) = y, and is that linear flows with such equilibria are generic.
(did()( l ' ( r ( t ) ) < 0 if x(1) is a solution curve not in 7. (Hinl: Let z(1) be The harmonic oscillator illustrates the necessity of this condition as follows.
tht. solution rurvr in 7 .such that r ( l ) - t ( t ) - 0 aa 1 4 w ; s r r Chapter 13. Suppow that f; W 4 R', with W 3 0, is a vector firld which in some neighborhood
-
Svrtion I , 'l'hrorrm 3. Consider ~d 1 r(1) r(l)ladl for somr large constant T.) of 0 is given by
3, 1.t.t i I (1 R" br, rlr,t.n a ~ ~ r l y 111n prriodic attractor for a (." vc-ctor field f : W -+
-
R * SII,,\Vthat 7 ham n nrighborhood ( I with the following property. For any
+ > 0 t11tmrc. rxists 6 > 0 such that if g: R X W R" is C' and I g(1, r ) -
zr=Ax, A=[-: :I.
j 1 1< 6 . then every solution ~ ( t to
m ) I' = g(1, I) with r ( h ) E II satisfie8 By arbitrary slight perturbation, the matrix A can be changed to make the origin
r i I I .. 1 . for all 1 2 I, and d ( r ( t ) , 7) < t for all 1 g r e a t ~ than
r some !I. (Hint: either a sink, saddle, or source. Since thme havr different dynamic behavior, the
l'rolil,,~tt2 , and Problem 2 of Srction 1.) flou~yare not t o ~ o l o g ~ r a l the
l y sanw Hrncr / ' I > nor t r ~ ~ t - t ~ l r stahlr.
> ~ I I y In cntltm.t.
11LS known tf~atthe Van der Pol ocr~lli~tor I. strurtur;~llyrt;llllr.
The folloning is the main result of this section. I t givw an example of a class of
$3. S t r u c t u r a l Stability s t m c t u r d y stable systems. (See Fig. A.)
In the previous sections we saw that certain features of a flow may be preserved
under small perturbations. Thus if a flow has a sink or attractor, any nearby flow
will have a nearby sink; similarlv, for periodic attmtors.
I t sometimes happens that an!, nearby flow is topologically the same as a given
flow, that is. fur any sufficiently small perturbation of the flow, a homeomorphism
rxists rhnt rnrrirs rach trajpctory of the original flow onto a trajectory of the per- FIG A. A structurally stable vector field
turb:ht~,~n. ( A /tomsomurplr~srnis simply a continuous map, having a continuou
invvr..(-.i Surli a homromr~rphism sets up a onc-toone correspondence between
rquilit>rl:~ uf thr two flows, closed orbits, and so on. In tllis case the original flow
(or i t b vect~xfield) is callrd slrucfurally stable.
H c ~ cis thr prrrisr definition of structural stability, a t least in the restricted
setting of vtrtor fields which point in on the unit disk (or ball) in R '. Let
314 16. PERTURBATION THEORY A N D STRUCTURAL RT4BILlTY 63. STRVCTURAL STABILITY
Theorem 1 Lel f : W + R" be a C1wclor field on a n open sel W 3 D" luilh lhe
following properlies:
Before proving this we mention three other results on structural stability. These
concern a C' vector field f : W + R' whrrr W C R' is a neighborhood of P.The
fist is from thr original paper on structural stability by Pontr~sginand Andronov.
S = {f t Vo(W ) I f is structurally stable on P J Theorem 4 The sel of slruclurally slable syslems cortlaatred i n grad (D') id open
and dense irr grad (D").
i s dense awI open. That is, every elemenl of S has a neighborhood in Vo(W) contained
171 S, and every open sel i n Vo(W) contains a veclm field which i s slnrclurally stable We turn to the proof of Theorem 1. In outline it proceeds as follows. A vector
on Dl. firld g sufficiently close to f is shown to have a unique equilibrium a E D' near
Proporitionn LelO C E be a sinkfma CLu e c h j i e l d f: W 4 E where W is an open We now prove Theorem I. Since Dmis compact and f ( r ) points inward along the
s e t , , I I , ~ ~ I0. I I I ~erisls an inner product on E, a number r 0, and a neighbor-
HThere > boundary, no solution curve can leave D". Hence Dmis ~ositivelyinvariant. Choose
hoot1 X C P ( U ' ) off such lhal the following holds: for each g € % lh~reis a sink a = r > 0 and 32 C: V ( W ) aa in the proposition. let OJLd C 32 be a neighborhood of j
a ( y ) f o r y such lhal the sel so small that if g E &, then g ( r ) points inward along JD'. Let $I be the flow of
B7= lx€ E l l r l S r l g E 32p. Note that Dmis also positively invariant for ),.
For every z t Dm- int B,, th?rr is a neighborhood U . C W of r and C, > 0
contains a, is an the basin of a, and is poailiwly invariant under the flow of g. much that if y t U. and 1 1 l,, then
Proof. From Chapter 9 we give E an inner product with the followingproperty.
For snmr u < 0 and 2r > 0 it is true that
By compactness of dD" a finite number K,, . . . , U,. of the sets U . cover aD". P u t
For let x D" : then y = $,,(z) B,. andB, C bnsln ofa i111alcrJ,.
if r F: B.(a). To see this, write
(y(r), r - a ) = U ( r - a),x - a ) + (g(z) - f ( z - a ) , z - a )
It also implies that every y D" - a 1s of the fonn $ , ( x ) for sonre r dD" and
1 2 0. For otherw~se
hence y = o .
L.(y) is not empty : hut z L.(y). t l w t ~+,(z)+ a a-t x.-
Fix g E n l . We h a w proved RO far that the map
Afterword
1. Slii~ii. I l l a tif j: R2 + R2 is structurally stable on p' and f ( 0 ) = 0, thpn 0 is a
11) IWI lx~livc~~~t~ilil~rium.
?. l . ~ t7 C R",tt > 1 br the circle
7 = ( I E R . I Z : + ~ : - ~ , . Z ~ = Ofor k > 2 1 .
1,vt
N = ( . z € ReId!s,.r) 5 1 1 .
Let W C R" be a neighborhood of N a n d f : W -r R" a CLvector field. Suppose
f ( r ) points into N for d l r in aN = ( x € RmI d(r, 7 ) = I J.If y is a periodic
attrnrtr~rand y = L.(I) for all E E N, prove that f is structurally stable on This book is only an introduction to the subject of dynamical systems. To pro-
.S. (Ser Fig. C for n = 3.) ceed furthrr requires the treatment of diRrrrntial equations on manifolds; and
the formidable complications arising from infinitely many closed orbits must be
faced.
This is not the place to develop the theory of manifolds, but we can try to indi-
cate thrir use in dynamiral systems. The surface S of the unit ball in R' is an exam-
ple of the twodimensional manifold. A vector field on R' might be tangent to S
FIG. C a t all pointsof S ; if it is, thrv S is invarinr~turtdcr tltv f l ~ ~ In
e - this s a y n-eget an
rxnrnplr nf a dynamical sysirm on thr ma~~ifold $ (xr.v Fig. A).
Thc eompaetnPss of S implies that solution curves of such a system are defined
for all 1 i R. This is in fact true for all flows on compact manifolds, and is one
reason for the introduction of manifolds.
hlanifolds arise quite naturally in mechanics. C'onsidpr for example a simple
mechanical system as in Chapter 14. There is the Hamiltonian function H: L' + R ,
3 I f f E V(1V) is structurally stable on DmC R", show that f has a neighborhood where U is an open s u k t of a vector space. The "conservation of energy" theorem
3t such that every g t 3t is structurally stable. a t a h that H is constant on trajectories. Another way of saying the same thing
is that if H ( z ) = c, then the whole trajectory of 3. lies in the subset H-'(c). For
4. Shov that Theorem 1 can be shalpened a s follows. For every > 0 there is a
"most" values of c this subset is a submanifold of 11, just as thr sphere S i n R' c a n
~~eigtiborhood n of fsuch that if g 5 3t the homeomorphism h (in the definition
of structural stability) can be chosen so that I h(x) - t I < c for all x E Dm.
be viewed as H-I( 1) where H ( z , y, z) = z ' + y2 + z2. The dimension of H-'(c)
is one less than that of U.Other first integrals cut doun the dimension even further.
5 . Find necessary and sufficient conditions that a vector field f : R + R be struc- In the planar Kepler problem, for example, thr state space is originally an open
turally stable on a compact interval. subset U of R'. The flow conserves both total energy H and angular momentum
h. For all values of c, d the subset (r. E 1' 1 H ( L ) = C, h ( r ) = dl is a manifold
6. l e t A be an operator on R' such that the linear Row e l A is hyperbolic. Find
that is invariant undrr the flow.
t > 0 such that if B is an operator on R" satisfying I( B - A 1 1 < t, then there hlanifolds also arise in mechanical problems with constraints. A pendulum in
IS n 11(1111(nrnorphisnlof R* ontc itsclf that tskev each trajectory of thr dif-
three dimensions has a configuration space rondisting of the Bsphere S, and it8
fr.rlmtirll vquation I' = A+ onto a trajectory of y' = By.
state space is thc manifold of tangent vectors to S. The configuration space of a
must confront lin~itsets which can be rxtrenlrly ronlplicated, even for structurally
stable systems. I t can happen that a compact rrgion contains infinitely many
prriodic solutions with periods approaching infinity. I'oincare was dismayed by
11i.q discovery that this could happrn w e n in t l ~ vSr\~-tonianthree-body problem,
and expressed despair of comprehending such a plrenon~tmon.
In spite of t h r ~ ~ r c v a l r n of
c r such systems it is not easy t o prove their existence,
and UP cannot go into details here. But t o give some idea of how they arise in ap-
parently simple situations, 1r.e indicate in I:ig. B a discrvtr dxnarnical system in
the planr. Hrre the rrctanglr AHCD is srnt t o its i n ~ a ~A'BC'D' c in the most
obvious way by a diffeomorphism f of RZ,thus j1.4) = . I f , and so on. I t can be
s1hotr.n that j \\-ill have infinitely many I,c.r~od~cj)oi~~ts, u ~ l dthat this property is
prrsrwrd by perturbations. ( A point p is pcridic if f'(),\ = 11 for some 11 > 0.)
Considering R2a s embedded in RJ,one can construct a flow in R' tsansverse t o R'
whose time one map Icaves R' invariant and is just the diffromorphism f in Rz.Such
a flow has closed orbits through the periodic points of f .
rigid body n i t h one point fixed is a compact three-dimensional manifold, the set
of rotations of Euclidean three space.
The topology (global structure) of a manifold plays an important role in the
analyais of dynamical systems on the manifold. For examplr, a dynamical s y s t m
on the two sphvre S must have an equilibrium; this can be proved using the
I'oincarCFkndixson throrem.
The nrathemstical treatment of electrical circuit theory tan h e extended if mani- I n spite of PoincarB's discouragement therr has been much progress in recent
folds are usrd. The very restrictive special hypothesis in Chaptrr 10 was made in years in underfitanding the global behavior of fairly general types of dynamical
orrllsr t o avoid manifolds. T h a t hypothesis is that the physical stater of a circuit aystcms, including those exhibiting arbitrarily long closed orbits. On the other
(1111r.yi11gIiirrhhnR's and ~ r n r r a l i z r dOhm's laws) c a n br {raramrtrizrd hy the hand, we a r r far from a clear picture of the suk~jrctand many intrrmting problcms
inlh~ctorcnrrrnts and capacitor voltag~r.This convcxrts thc flow on the space of are unsolved.
physical states into a flow on a vector space. Unfortunatrly this w u m p t i o n ex- The lollouing books are recommendrd t o the reader 1%-houqshes t o see how the
cludes many circuits. The more general theory simply deals with the flo\r. directly subject of dynamied s y s t m s has developed in recent years. They represent a good
on the space of physical states, which is a manifold under "generic" hypotheses cross section of current research. Proceedings of Synzposta in Pure dlathnnolica
on the circuit. Volunre X I V , Global Analysis [33 and Dynar~~ical S y s l n ~ ~[19].
s See also Sitecki's --
f i l n n i f < ~ lenter d ~ into difierential equations in anothrr way. The set of points Diflerenliable Dyncmrica [AS].
~ < ) I I , S ~t ,r : r j ~ ~ r t t ~ r i r s to a given hyperbolic equilibrium form a submanifold called
trnd
t l ~ ht t ; ~ l r i ~~nilnifolcl
. <,f t11l.rquilibrium. T l ~ c s esubmanifolds are a key t o any deep
g1oll;ll ti~~~lr>rstanding of dynamical systcms.
Our an:ll?sis of thl. longterrn behavior of trajectories has been lirnitrd Lo the
sinrplest Lillds of limit sets, equilibria and closed orbits. 1;or some types of systems
thrsr arr essentially all that can occur, for example gradient flows and planar sys-
tt3rnP But t o achieve any kind of grneral picture in dimensions higher than two, one
ELEMENTAHY FACTS
nhpre the I , arc elements of a vector space. If there is not much ambiguity, the
Elementary Facts limits arc omitted:
C z, = r , - . + + I..
2. Complex Numbers
t:S -
f ( r ) = y of Y. In this case ~ v often
r write r -+ y or I --t ,'(I). T h e identity map
S i* defined by i ( z ) = r and if Q is a subset of X , Q C X, the inclusion
(c) 1z = z (hrrc I = I
( d ) If z = z +
+i-0)
iy is not 0 , thcn
-
nl:qr n . Q 4 .Y is defined h y a ( q ) = q. I f f : X --t Y, and g: 'I --t Z are two maps,
th,. r*~mlr<i>~tiong * f (or somrtimes w r i t t ~ ngf) is drfined by g f ( r ) = g U ( z ) ) .
TIIVtn:ql / S --t Y is said to bc one-to-one if whenr!vcr r, r' C X , r + z', thcn
*
f ( s 1 f(1'1. T h r Image o f f is t h r set dt.scrihvd as (P) If t is real (that is, t = I + i . O ) , tIlt.11multiplication b,. t cdncidrs with
scalar multiplication ill R2.If z and 2' are both real, complex multiplication
- -
1' and f is O I I P to ulie, thcn J has an inverse and conversely.
If /: S Y is a [nap and Q C X, then f I Q: Q Y denotes t b e restrictiun of
i = s - iy. Thus cr~njugationis a map e : C -
Thc cor)~plcrconjugate of a complex ~ ~ u n ~ btc .=r x +iy IS the complcx ~lurnber
C, s ( z ) = i. \vhiclr has 3s its set
of fixed poi~ltsthe real numbers; t h a t is to say i = z if and only if z is rcal Simple
J t(2 O PO J I Q (9) = f (9).
ELEMENTARY FACT8 3%
prvpt.rtil.s t r f ronjugation arc: Thus the expression on the right does not depend on i and furthermore g i v e a
i = z, way of finding (or defining) Det A inductively. The determinant of a 2 X 2 matrix
:]
C: is od - be. For a 3 X 3 matrix
(z + 2') = f + i',
E' = 22'.
' f ~ , ,:lll+llllll(.
. VIIIUI.
of 11 C O I I I ~ Inumber
PX z = z + iy is
/z/ = ( ~ i )=~ (zZ
f' + y7)lf2. one obtains
Thrn
Iz / = 0 if and only if z = 0, Det (.4) = all DetI,"[ -m Det [zl 21+ [z 21 ai. Det
1z+z'1112l+ 12'1,
Recall that if Det A # 0 , then A has an inverse. One way of finding t h b inverse
lzz'l = I z I l z ' l I is to solre r x p l ~ c ~ thp
B i% en inverse A ' forA
t l ~ system
. of ~ C ~ U ~ I ~ IAx- 1.g= By : then
O I I < y for x o b t a ~ n ~ n
l'ltis usr of the expunrntial symbol can be justified by showing that it is eon-
sistrnt with a ronvrrgcnt power scrirs representation of e'. Here one takes the
po\\-c.r svrics ~f ca+" or rmr dom for ordinary real rxponcntinlu; thus 4. Two Proponitions on Llncar Algebra
The purpose of this section is to prove Propositions 1 and 3 of Section lB, Chap
ter 3.
011i. c:itt olwratr uith complex exponentials by the same rules as for real ex-
~)I)II~~III~;I~S
Proposition I Every vecfor space F has a b a h , and every basis of F h a the m e
number of e h m k . If (el, . . , 4.
1 C F id an independen1 subsel tbaf is not a baaid,
3. Determinants
by adjoining lo if suitable veUma e+,, . . ., e,. one can form a basis el, ., CI. ..
The proof goes by some easy lemmas.
Oar. me? fiud a good account of d c t ~ r m i m n tin
s Lang's Second Course in Colcdur.
[12]. H u t \<I- just \\.rite do~vna couple of facts that arc usrful. Lemma 1 A s y
sm of n linear homogmears eqtdwm i n n + 1u k altwus
I,.lrrt give a general expression for a determinant. k t A = [a,,] be the
\rr. has a d r i v i a l solution.
I fr X 1 1 ) ~llrltrixwhose entr). in the it11 row and jth colu~nnis a,,. Denote by .4,>
thi. I 11 l i X ( n - 1) matrix obtained by deleting the ith row and jth column.
- The proof of Lemma 1 in done by the process of elimination of one unknown to
<
'I'hc~i~f I is a fixed intcger, 1 5 i n, the determinant mtisfies obtain a system of n - I equations in n unknowns. Then one is finished by indue-
tion (the first case, n = 2, being obvious). The elimination is done by udng the
nrt .4 = (-l).+lo,, Det A,, + --- + (-I)'+"a,. Dct A,.. first equation to adve f m one variable as a linear eombiition of the rest. The
327
-
ELEMENTARY PACW
exprcs-ion obtained is substituted in the remaining equations to make the re- Proposition 3 Lel T : E F be a linear map. Then
durtion
dim(Im T ) + dim(Ker T) = dim E.
Lemma 2 Lel l el, . . . , e.) be a h i s for a veclor space F. If vl, . . . , U . are linearly
i t l d e p r ~ ~ d eelements
Proof.
~it of F , lhen m n.
I t is sufficient to show that m
c, is a l i n ~ a rcombination of the ei,
<
+ n + 1. Suppose othe&. Then each
(a) Ker T
(b)
(c)
Im T -
In parlicular, suppose dim E = dim F. Then [he folh<rrg are epuiunlenl dolemenfa:
=
F;
0;
T if a n h o r p h i s m .
P r o o f . The second Dart follows from the first part (and things said in Section 1
of chapter 3 ) .
T o prove the first part of the proposition let fi, . . . ,f. be a basis for I m T . Choose
B \ Lemma 1, the system of equations .
el, . . . , eb euch that Te. = f,. Let gl, . . , 81 be a basis for Ker T . I t is eufficient
"41
to show that
C+,a.r= 0, k = I , . . . , Q, [ e l , . . . , eh, g ~ ., . . , g11
i-l
is a basis for E since k = dim Im T and 1 = dim Ker T .
hns a nontrivial solution t = (21,. . . , I"+,).Then F i ,these elements are independent: for if C A.e. + 1 Mag, = 0, application
of T yields 1 A,Te, = C h f , = 0. Then the A. = 0 since the j; sre independent.
Thus C Msi = 0 and the Mi = 0 since the gi are independent.
Second, E is spanned by the e. and g,, that is, every element of E can be written
so that the v , are linearly dependent. This contradiction proves Lemma 2.
as a linear combination of the e, and the g,. Let e be any element of E. Define
From Lemma 2 we obtain the part of Proposition I which say8 that two bases
v = A*., where Te = x
A& defines the A,. Then e = (e - v ) +
u. Now
T ( e - u ) =Osee- v i KerTandthus(e-u)canhewrittenasaLinesrcombh-
. .
havr the same number of elements. If [el, . , 1.e and (a,, . ..
, v.1 are the two ation of the gj.
bases, thvn the lemma says m <
n. An interchange yields n 5 m.
.
say that a set S = ( U I , . . , u,J of linearly independent elements of F is mmimal
if for e v p r 11 in F, v 4 S , the set { v , V I., . . , v.1 is dependent.
We may take k ? I . Then vk is the minimum value of I p(z) 1, for if r E D., then
>
I p(r ) I I ur, while if z 4 D,, / p (z) I ul by (2); and U I > vr since DI C DI.
Proof of theorem. Let I p ( s ) I be minimal. The function
Proof. For 2 # 0 we can write In other words, w ib a kth m t of Such n root exists, for if
-
-aa = P(cos 8 + i sin U),
ah
then we can take
I P ( ~ w ) <I l c y l .
This contradicts minirnality of I p ( 0 ) I = I cy I. Hence I p(0) I = 0.
Appendix I I k
On Canonical Forms
Corollary ..I polynomial p of &gree n can be foelored:
Every term on the right with j > 0 has a factor of z - A; hence 1. A Deeornpanition Theorem
C a,tNkz,,
t*
nJ = nil (z,), Appendix Iv
Hrnrr rr, = p,(~V)z,for the polynomial p,(l) = C z l a,,lk. Therefore .Vu, = The Inverse Function Theorem
p,(.\-)y, = 0. By Lemma 2, p,(L) is divisible by L"if nt _< nil(y,). Since I 5 nil(y,),
!v(* can \x-rlt(-
P,(t) = S,(l)t
for some polynomial s,(l).
But nnn-. substituting I\' for I, we have
uj = s,(N)Xxj
= s,(.V)y, E Z(y,).
Tlr(*rrf~ar
u , = 0 since the Z(y,) are indcpend~nt.
H'r now show that
In this appendix we prove the inverse function theorem and the implicit function
(1 1 V=Z(q)a...eZ(z,).L theorem.
with L C lirr A'. h t Krr ,V = K and Irt I, br a subspace of K such that
K = ( K f l N ( V ) )e L . Inverse function theorem Let W be an open aei in a uedm tpace E and ld /:
W -4E be a C m p . Suppaae zr E W id arch fhaf D j ( s ) ra an iMcrliblc linear
'l'11' I i. i~~~lt*pi.nd~nt
II from the Z ( r , ) . T o see this, let v E ( @ Z ( z , ) )n L. Then
operaloronE. Then~~hasanopenneighborhocdV C Wsuehfhaffalf V w a d i f f w
I I @%(I,)1 n li, and by an argument similar to the one above, this implies
murphimn nlo an open 8cl.
I \ I I ' ) 1h1t .\' (I') n I, = 0,hence u = 0.
It I r rl<.:~r tliat cvery cyclic subspace in K, and hence in L, is one dimensional. Prwj. By continuity of Df: W -t L ( E ) there is an open MI V C W about zr
Thcr~.f~or(% I, = Z (wl) e . . . m Z ( w . ) , where (w,, . . . , w,l is a basis for L. Finally, and a number r > 0 such that if y, z € V, then Dj(y) is invertible,
'I = Z ( Z I ) e . . . e Z ( z , ) eZ(w1) e . . . e Z ( w8 ).
and
This proposition implies the theorem, except for the question of uniqueness of the
rnittricrs .4,. . . . , A,. This uniqueness is equivalcnt to the assertion that thc oper-
:itor \' dr,tr.rmincs t l ~ csizrs of the blocks .A, (or the dimension$ of the cycllc sub- I t follows from Lemma I of Chapter 16, Seetion I , that j I V is one-to-one. hlore-
~ [ I : I ~1 , V - : is don? I ) \ induction on dim V.
'1'111. over, Lemma 2 of that seetion implies tbat j ( V ) is an open set.
IIII \ . to its image N ( V ) = F:
( ' o i , ~ ~ ,fIl *~ r, rt,rtrietiol~
, The map f ': j ( V ) + V is continuous. This foliows from I d compactness of
.V / F : F - F .
j ( V ) . Alternatively, in the proof of Lemma 1 it is shown that if y and r are in
V, then
It I. t,;~-i wr, tl~;ltII1' is thr tlirttct aum of cyclic subsparc.~ZI C I . . . e %. m !I.;.
\\t1,.1# 11 , ~ dim Z , > I , t111.n , V ( I ' ) is thr
C l i ~ l r.\'. : I I I CZI~ is gr1111rntrdI I zlr
hence, putting j(y) = a and j(z) = b, we have
r11r.,vt ~ I I I ~ I
N(Z1) e , . . m A'(Z,),
o h , r r \ rZk) I.<cyclic, Rrul.ratcd by ,\'( r k ) ,and dim N ( Z h ) = dim 2, - I . Since which provesf continuous.
111111 \ (P') < dim l', t l ~ nurnbvrs
r {dimZt -. I I are drtermincd by A' 1 F , hcnce I t remains to prove that j-' 4 CL.The derivative of /-I at a = j ( r ) E j ( V ) is
b! .\ 11 Llh,ss that ( d ~ m ZbJ arc also drtcrmined by N . Dj(z)-'. To see this, we write, for b = j(y) E / ( V ) :
Tl1t5liil~*lr~..*
t l ~ rpm)I I J thr
~ tlicorem.
APPENDIX IV THE INVERSE FUNCTION THEOREM 339
Nor Bt.forc brginning the proof we remark that the conelusion can be rephrased thus:
f(v) - f ( z ) = Dj(z) (V - 2) + R(Y. 2). ik graph oj g id the set
where
Thus F-l(c) is a "hypersurface" in a neighborhood of ( r ~yo).
,
To prove the implicit function theorem we apply the inverse function theorem
to the map
Hence
j: W + E, x E2,
19 - 2 - Dj(z)-'U(v) - f(z)) I = I Y - r - Df(z)-I(Df(y - z) + R(V,z ) ) 1 f k . v) = (I, F(I. Y ) ) .
= I D/(Z)-~(R(V,
2) 1. The derivative off at (z, y) C U' is the linear map
Hence
of(=,v ) : ELX E; 4 EnX G,
is C1.
h(z, w) = (2, 4 2 ,
p : u x Y-I.
u')),
Define a C1 map
g: C' - V,
a.I . Synge and B. Criffiths, Principlrs of hfrchnrra (New York: hlcGraw-Hill, 1949).
24 R. Thorn, SLabifrlt Strudurrfk d Y o r p h o g h h e : Esaai d'unc thloric gh(ro1r dcamodhlrs (Read-
ing, 3fa~sachusetfs:Addlson-Wedey, 1973).
25. A . Wlntaer, The A n u l y f ~ c o fFrmndatim of C e M l a f Mechan~ca (Princeton, New Jersey:
Princeton Univ. Preus, 1941).
26. E. Zeeman, Diflcrentisl cqus(ians for heartbeat and nerve irnpulaes, in D y n a m ~ c dSyaLmu
( 3 1 . h l . Peixoto, ed.), p. 683 (New Yark: Academic Press, 1973).
Answers to Selected Problems
Chapter I
Chapter 2
Page 27
m e a n s thr s e t
".\lostv initirrl rrlnditi<~ns is, v ) . R2 X R1such that r i3 not
colltnr,ar \\-ith r.
2. ( a ) n-ith V ( s ,y) = -2
- -
2d and ( c )
r2
\r-~thI,(+, y) = -
3 -3 2
7. H i l l / , Usr ( 4 ) Srction 6.
ANSWERS TO BELECl'ED PROBLEMS Y S W E R S TO SELECTED PROBLEMS
Section 2, page 81
Section 1 , page 65 3. A = l , B = f i
4. ( R ) fi (h) 4 (c) I td) %
2, dim K = dim Kc and dim F 2 dim Fa
0. ( a ) and (dl
3 F 3 RCR
Section 3, page 87
Sect ion 2. pape 69
Section 3, page 73
4. (a) T h r norm L=, 1.
Introduce thr new basis ( 1 . 0 , O), (0, -*,a), (1, -2, - I ) , and new coordinates 7. Hinf: Use geometric series
(y,, yz. Y ) related t o the old by
*-,
I 1 = Yl + uz, x z ' = L for O < z < 1, w i t h r = ( l I - T I [
2
, = /Z - 2y,, ,, I - =
I, = y2 - y,. 13. H i ~ l l : Show that all the terms in t h r pPa1.r srries for e* leave E invariant.
ANSWERS TO BELECTED PROBLEM6
"
1;. .4 prr.ir.rvrs cach grneralizd v,gc.nspacc G:hence it suffices to cr~nsidr-rthc
10 0111y if a < -2 a r r thrrr any values of such k and in this case for
Ia)
rc?;trirtionsof A and T t c ~EA.I f T = S + .Y, thrn S 1 E, = $.I whichcommutes
k >6 2 ; . with A . Thus S and T hoth commute with 4 , so thrrefore does .\' = T - S.
I lii !it, values (,f A-.
S. Ksr thv Cajlcy-Harn~lton thwrrm.
I4 Ilrrtl: There is a rrsl cigrnvalur.. Study 2' on its (~igrnspace.
1.5. O j n s i d ~ bascs
r of the* kernrl and thr image.
I. Canonica~forms:
ill1 ,111 = -fs[4t + I ] + e"-
18
+ e41k ( a ) 0 0I 0!]
trr r 1 0 = Acns1 + B s i n / .
y(1) = - A sin 1 + B ens 1 + 21 0 0 0
0 0
S e c t i o n 6 , page 107 4. Assunll that S is in nilpotent canonical form. k t b denote t h e number of hlocks
2. (:I) s(l) eos21. (h) s(1) = -e" + elt-'.
nntl s t h r ~naximalnumber nf rows in a block. Then bs 5 n ; a l w b = n r and -
=
s 5 k.
. I r 1, I i (h) cxpV2 1 , cxp - f i t 4. Sirnilar pairs a r r ( a ) , ( d ) and ( I ) ) , (c).
4 ffz,at. ('llc.ck caw? (a), ( b ) , ( c ) of t h r thcorum.
X a = 0, b > 0 ; prrind is <b/2r.
S e r f ion 4, page 132
Chapter 6
S e r t i o ~2.~page 120
By induction
P
I. h({) =?-I.
Chapter 7
ul(l) = Ods = 0,
Section 1, page 150 0
(c) 1
5. ( a ) For 0 < c < E let
Chapter 9 Chapter 10
":i1. (1,) (,.) 1. Every solut~onis prriodic! Hifrl: If ( r ( l ) ,y(1) ) is a solution, so is ( - I ( -I),
13. I l i r i ~ : I,ook a t the Jnrdan f o r n ~of A . I t suffices t o considr.r a n r l ~ m e n t a r y y(-Oj,
.Tordnn hlork.
Section 4, p a g e 228
S. 1'-'LO, r ] I < positively invariant. The a-limit set of any point of 1'-1[O, c ] con-
siqts entirrly of rquilibria in 1'-'[O, c ] , hence it is just 2'. Section 5, p a g e 237
Section 3, p a g e 447
2. H i n l : Apply Proposition 2
1. I f ir~ts: ( a )If z is not an equilibrium, take a local section a t z. (b) See Section 1 , p a g e 309
Prnhlrm 2 of Section 1.
I. Hitrt: If B is ckrtre to A , each eigenvalue of B having negative real part will
be closr to a similar elgrnvaluc A of A . Arguing as in the proof that S , is open
in Theorem 1 of Chapter 7, Spction 3, show that the sum of the multiplicities
of thesr rigenvalues p , of B nrar A equals the multiplicity of A. Then show that
2. H i n t : Lct y C r. Take a local section a t y and apply Proposition 1 of the pre- bases for the generalized eigenspaces of the ,I can be chosen near corresponding
villus section. bases for A .
2. Hints: (a) Use Poincad-F3endiison. (b) D o the problem for 2 n l + 1. Suppose Df(0) hss 0 as an eigenvaluc, let g,(r) = f ( r ) 4- r r , t # 0. For.1 t I
sufficiently small, one of g-., g. ~111be a saddle and the other a source o r s ~ n k ;
closed orbits; use induction on n .
hence f cannot have the same phase portrait as both g, and g.. If Dj(0) h
5 . H i n l : Let U be the region bounded by a closed orbit r o f f . Then g is trans-
3 3 1 , A > 0,as an eigenvalue, then g-. is a sink and g,, is a source.
verse to the boundary r of U . Apply Poincar&Bendixson.
ti. H i n t : First consider the case where e l A is a contraction or expansion. Then use
Problem 1 of Section 1.
Section 1 , p o g e Z78
S e r t i o n 3, prrge 285
Chapter 15
4 (hxllge