Suficient Conditions For Stability and Instability of Autonomous Functional-Differential Equations
Suficient Conditions For Stability and Instability of Autonomous Functional-Differential Equations
Suficient Conditions For Stability and Instability of Autonomous Functional-Differential Equations
JACK K. HALE
where the 7j are positive constants. In this paper, we will be concerned with
the application of the concept of Lyapunov functionals to the determination
of sufficient conditions for the stability of such systems. Lyapunov functionals
have been applied to these equations by many authors and the reader should
consult the book of Krasovskii [9] for a detailed bibliography. Lyapunov
functionals can also be employed to discuss the stability of nonautonomous
systems; that is, systems in which F contains t explicitly, but we restrict
ourselves to the autonomous case since more general results are obtainable.
In particular, we wish to discuss the implications of a theorem of LaSalle [7j
for ordinary differential equations properly extended to functional-differen-
tial equations. Some of the results of this paper have been announced in [5j,
and the present paper contains the complete proofs of those results as well
as some new results on instability, and many applications. The extension
of the results to infinite time lag is contained in Section 4.
To discuss functional-differential equations in the proper setting, it is
necessary to introduce some notation, En will denote the real Euclidean space
of n-vectors and ] x ] will denote the norm of the vector x in En. If r > 0 is
given, C = C([ - r, 01, En) will denote the space of continuous functions
with domain [ - r, 0] and range in E n. The norm in this space will be the
uniform one: \j y 11= max-,rGeGO1(p(e)( for ‘p in C. Suppose x is any given
function with domain [- t, co) and range in Es. For any t > 0, we will let xt
denote a translation of the restriction of x to the interval [t - Y, t]; more
specifically, xt is an element of C defined by xt(e) = x(t + f?), - I < 8 < 0.
In other words, the graph of xt is the graph of x on [t - Y, t] shifted to the
interval [ - r, 01. The reader is urged at this point to use extreme caution
in distinguishing between the following symbols: x is a function taking
[- r, a~) into En, xt for t > 0 is a function taking [ - Y,0] into En, x(t) is the
value of x at t and x$(e) is the value of xt at 0, x(v) also has a special meaning
to be explained in a moment.
If H is a given positive constant, we use the notation C, for the
set (9 in C : 11y [I < H); that is, C, is the open ball in C of radius H.
Iff(v) is a function defined for every upin C, and 3(t) is the right hand
derivative ofx(t), we consider the following autonomous functionaldifIerentia1
equation :
*(t> =f(xt), t> 0. (1)
LEMMA 2. If 9)in C, is such that the solution x(v) of qtem (1) with initial
function q~at 0 is defined on [ - Y, 00) and I/ xt(v) 11< H1 < H for t in [0, oo),
then Q(v) is a nonempty, compact, connected, invariant set and
&St (X&P,>,-Q(d) - 0 a.~ t--+03.
Proof. From Lemma 1, we know that the family of functions xi(v),
t > Y, belongs to a compact subset S of C and, furthermore, S could be
STABILITY OF FUNCTIONAL EQUATIONS 455
chosen to be the set of 4 in C such that /I$ I/ < Hi , I/ 4 11< K for some
constant K. This shows that Q(p) is nonempty and bounded.
If IJ belongs to 9(v), then there exists a sequence tn , tn -+ co as n + co,
such that 11xt,(v) - I+% (I + 0 as n + co. For any integer N, there exists a
subsequence of the t,, , which we keep with the same designation, and a
function g7(p)) defined for - N < T < N such that II x~,+.,(P))- g,(v) II--+ 0
as n + oc),uniformly for T in [ - N, N]. By the diagonalization process, one
can choose the t, so that 11xtn+,(v) - g,(v) Ij + 0 as n + co uniformly on all
compact subsets of (- co, co). In particular, the sequence x~~+,(P)) defines
a function g,(v) for - co < 7 < co. It is easy to see that g,(p)) satisfies (1).
Since g&p)) = #, it follows that the solution x~(#) of (1) with initial value y%
at 0 is defined for all values of t in (- co, co) and, furthermore, is in Q(v),
since II~t,+dv) - 44 II- 0 as n + co for any fixed t. This shows that
Q(v) is invariant. It is clear that Q(v) is connected.
To show Q(v) is closed, suppose #n in Q(p) approaches I/ as n --f 00. There
exists an increasing sequence of t, = t&,,) + CO as n + CO such that
IIxt,(v) - A II- 0 = n - ~0. Given any z > 0, choose n so large that
IIA - # II < 42 and IIxt,(~) - A II < 42. Then IIX#P) - 4 II < E for n
large enough which shows that 4 is in Q(p) and Q(v) 1s closed. But, clearly
Q(v) C S and since S is compact, it follows that 52(v) is compact.
To show the last assertion of the lemma, suppose that there is an increasing
sequence of t,, -+ co as n + co and an a: > 0 such that 11xt,(y) - 1,4II > OL
for all 9 in Q(v). Since x,,(cp) belongs to a compact subset of C there exists a
subsequence which converges to an element # in C and thus # is in Q(v).
This is a contradiction of the above inequality and completes the proof of
the lemma.
Remark. In the proof of Lemma 2, we only used the fact that xt(v) was
continuous in t, y and that x~(F,)belonged to a compact subset of C. Therefore,
the Lipschitz condition off could have been replaced by 11v I( < HI < H
implies If(q) I < L for some L.
for all t 2 0. Therefore I x(v) (t) ] < E for all t > 0 and, thus, I/ xt(v) /) < E
for t 3 0. This shows the solution x = 0 is stable. The second part of the
theorem is proved as follows. Since u is increasing the set Ut of q~for which
V(v) < I satisfies the conditions of Theorem 1 if I < I,, and, thus, by Corol-
lary 1, every solution in U, approaches zero as t + co. This completes the
proof.
We can now deduce the following interesting corollary of Theorem 2. This
corollary is also true for nonautonomous equations and may be found in
Krasovskii [9]. Of course, the proof for the nonautonomous case is more
difficult.
STABILITY OF FUNCTIONAL EQUATIONS 457
Then the solution x = 0 of (I) is stable and zf w(s) < 0 for s # 0, then it is
asymptotically stable.
505-S
458 HALE
(iii) V(p) < ~(1~(0) 1) on U n Cy; u(s), continuous, nonnegative and non-
decreasingon [0, H), u(O) = 0;
(iv) v&(v) 2 0 on Ct( 0’ n C,) and the set R of v in CZ(U n C,) such
that p&(v) = 0 contains no invariant set of (1) except qo= 0.
Under theseconditions, the solution x = 0 of (1) is unstable and the trajectory
of each solution of (1) with initial value in U n C, must intersect cY in some
finite time.
Proof. SUPPOSE q. E U n C,. By hypothesis (iii), j v,,(O) / 2 u-‘(V(~~))
and (iii) and (iv) imply that xt = x,(& satisfies
3. APPLICATIONS
where g(x) is a real function defined for all real x, locally Lipschitzian in x,
and a(t) is continuous together with its first and second derivatives on [0, Y].
We also suppose that
n(t)
=- j"u(
- e)
g(x(l
+e))
de=- stt-r
--I
a(t - u)g(x(u)) du. (8)
We wish to investigate along the same lines as Levin and Nohel [S] the
relations between the solutions of (8) and the second order ordinary differen-
tial equation
f + u(O)g(jc) = 0. (9)
First of all, we derive the second order functional-differential equation
which all solutions of (8) must satisfy. If x is a solution of (8), then it has
continuous second derivatives. Differentiation of (8) and making use of (7)
yields
w + 4Mw
=- t d(t - u)g(x(u)) du
I t-r
w + 4omw = - 4~) $ gw + 4) de
-r
the neutron density in the reactor. Also, the same equation arises in some
one dimensional problems in viscoelasticity where x then represents the
strain and a the relaxation function.
Suppose y(t) is any continuous function defined on (- co, co). In the
following, we shall be interested in such functions y(t) which satisfy the
additional conditions:
if d(r) f 0, then
We shall say that Q(y), the w-limit set of an orbit of (8) through 9, is the union
of orbits of solutions of (9) which satisfy (II) if Q(q) is the union of sets of
the form u-, <t<m 4~ 81, where 401, B) (t) is a solution of (9) defined for
- co < t < co, U(OL,j3) (0) = 01,ti(~~,/3) (0) = /3, and u(or, j3) (t) satisfies (11).
Notice that any solution of (9) which satisfies (11) also satisfies (10).
Not all solutions of (10) are solutions of (8). In fact, integration of (10)
from 0 to t yields
L+(t)
- k(0-t)= - J:-,u(t- u)&(u))au+ 1’--F
a(--@g@(e))
du, (12)
where C is given by
THEOREM 5. If system (8) satisfies conditions (6) and (7), then every
solution of (8) is bounded and the w-limit set of any solution of (8) is the union,
as explained above, of orbits of solutions of (9) which satisfy (II).
Proof. We use the same Lyapunov functional introduced by Levin and
Nohel [S]. If M is the minimum of G(x) defined in (6) and
0
-sgh49) de = 0 for any s in [0, r] for which ii(s) # 0. (14)
But all solutions of (8) satisfy (10) and any solution in an w-limit set must be
defined on (- co, co) and satisfy (14) for all t in (- co, co). Finally this
implies these solutions must satisfy (9) and (1 I), completing the proof of the
theorem.
Two questions now present themselves in a natural manner: (i) What are
the possible solutions of (9) which satisfy (11) ? What additional conditions
on u(t) and g(x) will ensure that Q(F) is generated by only one solution of (9) ?
We now give some partial results in an attempt to answer these questions.
462 HALE
COROLLARY 4. If system (8) satisfies conditions (6) and (7) and there is an
s in [0, T] such that tits) > 0, then (8) has no nonco~tant periodic solutions.
Proof. There exist s0 < si such that ii(s) > 0 for s,, < s < si . If (8) has a
periodic solution xi(v) then Q(v) = utx,(cp), and Theorem 5 states that it
must be generated by a nonconstant periodic solution u(t) of least period p
of (9) satisfying (1 lb) for s0 < s < si . But integration of (9) yields
which implies G(t) is periodic of period s for so < s < s1 . But this is impossi-
ble since zi(t) is periodic of period p. This proves the corollary.
COROLLARY 5. If system (8) satisfies conditions (6), (7), xg(x) > 0 for
x # 0 and there tS an s in [0, Y] such that ii(s) > 0, then the solution x = 0 of (8)
is globally asymptotically stable.
Proof. The condition ii(s) > 0 for some s implies a(O) > 0 and this
together with the conditions on g(x) imply that every solution of (9) is periodic
and the only constant solution is zero. Theorem 5 and Corollary 2 imply the
conclusions of the theorem.
Corollary 5 is essentially the same as the one of Levin and Nohel [S].
Corollary 5 is a statement of global asymptotic stability of the solution
x = 0 of (8). Can we use this result to draw any conclusions about the solu-
tions of (IO) ? Any constant function satisfies (10) and one would suspect
that any solution of (10) approaches a constant as t -+ co. This seems rather
difficult to prove in the general case, but we can prove the following simple
result from Corollary 5. If for any real constant b, there exists a real 6 such that
where g(y) satisjies the conditions of corollary 4, Gzrz fo-r a given solution x of
(lo), there is a constant a such that x ---f a as t -+ co. In fact, if the solutions
of (10) are required to have continuous first derivatives as stated in the remarks
preceding Theorem 5, then the system (10) is equivalent to the system
c =do-)+ so
-F
4- 4&(W, yeso-? a(- 0) d0
STABILITY OF FUNCTIONAL EQUATIONS 463
g(x) = - x + 2. (15)
Suppose there are s,, < sr such that ii(s) > 0, s,, < s < s, . The conditions
of Theorem 5 and Corollary 4 are satisfied and the w-limit set of any solution
of (8) with g(x) given by (15) must satisfy
R-x+3CS=O
FIG. 1.
By Corollary 4, the only possible candidates for an Q(v) of (8) are shown in
Fig. 2. The curves around 1, - 1 are determined by solutions z+(t) > 0,
%(t) -=c0, - co < t < co, respectively. But there is a T such that
0 < z+(t) < 1, - 1 < t+(t) < 0, for t > T and, thus, for any s # 0
for t > T.
464 HALE
This shows that ui(t), us(t) cannot have orbits which belong to an Q(q).
Consequently, the w-limit set of any solution of (8) with g satisfying (15)
must be one of the three constant functions - I, 0, 1.
For this same particular example let us analyze the stability properties of
each of the constant solutions - 1, 0, 1. If 01is either of these solutions,
the linear variational equations relative to LYare
FIG. 2.
It should be clear the particular form g(x) given in (15) is not at all essential
for the above type of results and one could have a function g(x) with many
zeros on the axis.
If ii(s) = 0 for all s then the behavior of solutions of (8) may be more
complicated. The following result indicating this fact is due to Levin and
Nohel [S]. The proof of the theorem is essentially the same as in [S].
w-limit set of any solution of (8) consistsof the orbit of somesolution u(t) of (9)
which satisfies the condition
0
-rg(4t + '3) de = 0, -co<t<~m. (18)
s
In this case, u(t) must be periodic of some period p and there must be an
integer m such that mp = Y. If ii + g(u) = 0 and mp = Y then u satisfies (8).
Remark. If the only solution of (9) which satisfies (18) is the solution
x(t) = 0 for all t, the zero solution of (8) is globally asymptotically stable. In
particular, this is true if g(x) = a%, G > 0, and Y f 0 (mod 27r/u). If Y s 0
(mod 24~) and g(x) = a2x, th en every periodic solution of (9) is a periodic
solution of (8). Also, it will be clear from the proof that g(0) = 0 and (6)
implies the zero solution of (8) is uniformly stable.
v(ut(a))
=G(-a)++jr 9
[G(- a) - G(u(T, a))] dT
where /I(a) > 0 is defined as the unique positive number such that
G(- a) = G@?(a)). F rom the properties of g, it follows that V(u,(a)) is a
strictly increasing function of a and this is a contradiction unless aI = aa .
To prove the last assertion of the theorem, note that
ji + h(y) = 0
STABILITY OF FUNCTIONAL EQUATIONS 467
cw
System (21) is more general than (19) and for (21) to be equivalent to (19)
the initial conditions for (21) must be chosen in a special manner. However,
we are going to discuss the more general system (21). Following Volterra [IO],
we define V(v, +) as
Then
Consequently,
all t. Thus, the largest invariant set of (21) in the set whete r(zl)(v, 4) = 0
is (0,O). Since V satisfies the conditions of Theorem 2, the origin si asympto-
tically stable. This completes the proof of the theorem.
Proof. Using relation (23) for rca,, we see that the set where r = 0
consist precisely of those initial functions for which +-T) = p(O). Con-
sequently for a solution to belong to the set where P = 0 it must be periodic
of period r. But any nontrivial periodic solution of a linear equation with
constant coefficients has zero average over the period and it follows that the
largest invariant set in the set where p = 0 must be periodic solutions of
period r which satisfy (24). The conditions of Theorem 2 are satisfied and-
the theorem is proved.
In the case where x is a scalar, it is easy to show that the w-limit set of any
solution of (21) satisfying the conditions of Theorem 8 is generated by at
most one solution of (24).
We now prove the following instability theorem for the case where M < 0.
has an unlimited food supply and species B relies upon A for his development.
For positive constants r, ci, yi , j = 1, 2, and nonnegative functions Fr(8),
F,(B), 0 < 0 ,< r, the species are assumed to evolve according to the relations
where
4 q= +s‘F(e)
P= +f W’)
7iK2 de,
0
Y,&
0
G=K&>O,
F = K,F2 > 0,
and we make the assumption that p > 0, Q > 0. If we can show that system
(26) is asymptotically stable, then the equilibrium point (Kr , Ka) of (25) is
asymptotically stable. Notice that (26) is a generalization of the scalar version
of (21). In fact, we obtain (21) by putting G = 0 and replacing y by - y.
THEOREM10. 1fp > 0, q > 0, F(B) >, 0, G(8) 2 0, I’<@ < 0, e(e) d 0,
0 < ~9< Y, and there is a 0, in [0, Y] such that either P(L),) < 0 or G((e,) < 0,
then every solution of (26) approacheszero as t -+ Q).
STABILITY OF FUNCTIONAL EQUATIONS, 471
Proof. Define
It follows that
Therefore all solutions of (26) are bounded. If there is a 0, in [0, r] such that
$‘((e,) < 0, then xt in the largest invariant set in the set where P = 0 must
satisfy x(t - e) - x(t) = 0 for e in an interval containing e,. Consequently,
x = c a constant and the differential equation implies x = 0, y = 0. There-
fore, Theorem 2 implies every solution approaches zero as t -+ co. The
other case is treated in exactly the same manner and the theorem is proved.
One could obviously generalize this example to a system of equations of
a special type.
w = r(t)
90) = - ww + j’wkwt - 4 - mi de (27)
0
where h(x), g(x) are continuous functions of x such that x/z(x) > 0, xg(x) > 0
ifx#Oand
If
G(x) = j’g(s) ds
0
and
tic2,)(xt , yt) = -F(r) G[x(t - I) - x(t)] + j-i 2$3) G[x(t - 0) - x(t)] de.
472 HALE
If we suppose F(B) 3 0 for all 0 and there is a 0, such that P(0,,) < 0, then,
as in the previous example 2, one shows that every solution of (27) approaches
zero as t + a3. If P(0) = 0, 0 < 0 < Y and F(r) > 0, then every solution of
(27) is bounded and the w-limit set of any solution of (27) must consist of
periodic solutions of
k”--y j = - h(x) (28)
w = YW
for all x, If h(x) = dg(x)/dx, system (29) is related to a second order differen-
tial-difference equation. For any scalar functions v, # defined and continuous
on [- r,O], define
where
If T < b/L, then one can always find a positive v for which the integrand is
positive definite in #(O), #(e) w h ic h im pl ies ~~2s)(~,I#) = 0 if and only if
4 = 0. But, the only invariant set of (29) which lies in the set of (q,, 4) for
which $ = 0 is (0,O). Consequently, the above theorems may be applied
to obtain the results that the solution x = 0, y = 0 of (29) is asymptotically
stable and if, in addition G(x) -+ co as j x I-+ co, this solution is globally
asymptotically stable.
STABILITY OF FUNCTIONAL EQUATIONS 473
and let
for some convenient v > 0 (and in fact, any v < 2b/r) if Y < b/L and
p&,(y, #) = 0 if and only if 4 = 0. But, the only invariant set of (29) for
which # = 0 is v = 0. Since V satisfies the conditions of Theorem 4, the
origin is unstable.
As a special case of this example, let us consider the Rayliegh equation
with retardation,
which, for Y = 0, can be transformed into the well known van der Pol equa-
tion. If this equation is considered as a system,
w = r(t)
j(t) = < [l - fp] y(t) - x(t) + f,Y(t + 4 de, (30)
-fcY)=+ -G)Y
g(x) = x, h(x) = 1.
For any y, 0 < y < 1, if y2 < r2 < 3, then -f(y)y > ~(1 - r2) > 0.
Consequently, from Theorem 4 and the computation in the above example, it
follows that the solution x = 0 of (30) is unstable if r < ~(1 - p). The region
U where V > 0 consists of all 9, I/ for which
505-6
474 HALE
where v > 0 can be chosen as small as we like with the only restriction on v
being that v < ~(1 - y2)/r. Theorem 4 implies that any solution of (30) with
initial function q~E U n C+ must reach the boundary of C,+ in a finite
time. If r = 0 this is the same result as one obtains for ordinary differential
equations; namely, if Eq. (30) has a cycle it must be outside the set U n Cd,
if r < ~(1 - y2). This type of result states that the properties of the type men-
tioned above are continuous in r.
Example 7 (Krasovskii [9], p. 10). Consider the scalar equation
if
\
\
\
\
/
/
/
/
/
/
/
/
/
/
/
476 HALE
Consequently, if for given a, b and Y we can choose 01and ,8(e) so that relations
(33) and (35) are satisfied, then the point (a, b, Y) will correspond to a set of
the parameters for which (31) is asymptotically stable.
Let us check first of all to see whether we can obtain the same results as
before for the region of stability which is independent of Y. For 01= 0,
inequalities (33) and (35) do not depend on Y explicitly and reduce to /3(B) > 0,
B(e) > 0, - Y G e G 0, a > p(O), b2 < 4(u - /3(O))fl(- Y). Suppose there
is a constant q < 1 such that b2 < qu2, a > 0. Then we can satisfy the inequa-
lities if we choose /3(e) close enough to u/2, which yields the same result as
before.
We now make some rather crude estimates of 01,/3 to obtain some informa-
tion about that part of the stability region which depends on Y and is con-
tained in the region a + b > 0. In particular, we want to see if it is possible
to obtain the region of stability as the region a + b > 0 if Y -+ 0. If we let
01= - b - 2/3- Y) [this maximizes A,], then our inequalities (33), (35)
will be satisfied if
a+b<O
F’*(d V(v)
<vgp,
3PP”(O),
for all 9, in C. If
then U satisfies (i) and (ii) of Theorem 4 and the remark after Theorem 4
implies the solution x = 0 is unstable.
we have
and if
vw = $Q$ + Q1’-Tv6(e)
de
and use Corollary 2 to prove the zero solution is stable.
Notice that the same V functionals may be used to show that the zero
solution of n(t) = d(t) + W(t - r) is stable or unstable according as
a < 0 or > 0, regardless of the size of b. One simply must operate in a
sufficiently small neighborhood of the origin.
STABILITy OF FUNCTIONAL EQUATIONS 479
SUP I de) I *
IIY II Ef IIP Ilr-Ml]= -r<e<o
fh 4)= f$ N=O
mN 3 mN = min PvN, IIv - # II[-(N+I).--N-J). (37)
One can show that a sequence vn --+ v [that is, p(~~ , p’) + 0 as n -+ co] if
and only if for every nonnegative integer N, /j yn - y$](r-N,O,+ 0 as n-+ CO.
For more details on this space of functions, see Arens [I], Bourbaki [3], and
Kelley [6].
If H is a positive constant, we use the notation C, to denote the set
GP in C : sups in r--h,O~ I ~(4 I < HI.
If u is any real number and x is any continuous function with domain
[U - Y, co) and range in En, we let xt , t 3 u, denote a translation of the
restriction of x to the interval [t - Y, t] defined by ~~(0) = x(t + 8),
- co < 0 < 0; that is, xt belongs to C.
If f(y) is a function defined for every v in C, and i(t) is the right hand
derivative of x(t), we consider the following autonomous functional differen-
tial equation:
*.(t> =f(xt). (38)
A solution is defined the same way as for the case of finite lag. Relation (37)
allows one to prove the following existence and uniqueness theorem: Iff(v)
480 HALE
function ~4in C(6, , En) such that 11xtn - $ IIaN+ 0 as n -+ cc. Application
of the diagonalization procedure yields a sequence s,, , independent of N,
and a function $ in C([- Y, 01, En) such that