HOPF BIFURCATION OF A DELAY DIFFERENTIAL

EQUATION WITH TWO DELAYS

JAUME LLIBRE
1
AND ALEXANDRINAALINA TART A
2,3
Dedicated to Professor Dr. Ioan A. Rus on his 70th birthday.
Abstract. We consider a delay dierential equation with two delays.
The Hopf bifurcation of this equation is investigated together with
the stability of the bifurcated periodic solution, its period and the
bifurcation direction. Finally, three applications are given.
1. Introduction
Hopf [9] was the rst who state a theorem concerning the bifurcation of
periodic solutions from a singular point of an ordinary dierential equation.
Many generalizations to innitedimensional systems have been given (see
[12] for references). As far as we know, the rst statement similar to this
theorem for retarded functionaldierential equations was given by Chow
and MalletParet in a course at Brown University in 1974, see [7]. Ecient
procedures for determining the stability and the amplitude of the bifurcating
periodic orbit using a method of averaging have been given by Chow and
MalletParet [4]. The global existence of a Hopf bifurcation as a function
of initial data and the period has been discussed by Chow and Mallet
Paret [5] and Nussbaum [14]. The interest on the periodic orbits of a delay
dierential equation has increased strongly these last years, see for instance
[2], [3], [13], [15][17].
In this paper we study the delay dierential equation of the form
x(t) = (a +)[x(t 1) +x(t 2) +G
2
(x(t), x(t 1), x(t 2))] (1)
[1 +G
1
(x(t), x(t 1), x(t 2))],
where 9/100 a

3/9, and G
1
(x, y, z) and G
2
(x, y, z) are analytic func-
tions in a neighborhood of 0 R
3
, starting with terms of degree at least
1 and 2 respectively. We prove that equation (1) exhibits a Hopf bifurca-
tion and we discuss for distinct functions G
1
and G
2
about the period, the
1991 Mathematics Subject Classication. Primary 34K18.
Key words and phrases. Delay dierential equation, Hopf bifurcation.
1
2 J. LLIBRE AND A. TART A
stability of the bifurcated periodic orbit and its direction in the parameter
space.
The periodic orbits of the delay dierential equation (1) for the particular
case a =

3/9, G
1
(x, y, z) = x
2
and G
2
(x, y, z) = 0 were studied by Jones
[10], and Kaplan and Yorke [11]. The Hopf bifurcation for this particular
delay dierential equation has been analyzed in Hassard, Kazarino and
Wan [8].
2. Analysis of the equation
Consider the delay dierential equation (1). We dene the function
F(x, y, z, ) = (a +)[y +z +G
2
(x, y, z)][1 +G
1
(x, y, z)].
Clearly it satises
(i) F is analytic in (x, y, z) and in a neighborhood of (0, 0) R
3
R,
(ii) F(0, ) = 0 for in an open interval containing 0.
Additionally we assume that
(iii) 0 R
3
is an isolated stationary zero of F for in an open interval
containing the zero.
The linearization of (1) about x = 0 is
(2) x(t) = (a +)(x(t 1) +x(t 2)),
and its characteristic equation is given by
(3) (a +)(e

+e
2
) = 0,
for more details see [8, 6] or Appendix 2.
Proposition 1. At = 0 equation (3) has exactly two pure imaginary roots
and no roots with positive real part.
Proof: The straight line u = /(a) does not intersect the curve u =
e

+ e
2
for 0. Thus, at = 0, equation (3) has no positive real
roots.
In order that equation (3) has a pure imaginary root = i with > 0 at
= 0 it must be that its real and imaginary parts are zero, or equivalently
that
(4) cos + cos 2 = 2 cos

2
cos
3
2
= 0,
and
(5) = a(sin + sin2).
HOPF BIFURCATION OF A DELAY DIFFERENTIAL EQ. WITH TWO DELAYS 3
Equation (5) can be written in the following form
(6) sin
3
2
cos

2
=

2a
If cos(/2) = 0, equation (6) is not satised. Thus cos(3/2) = 0. That
means 3/2 /2 +k[k Z. In this case, sin(3/2) = 1. Using the
graphics of the functions cos(/2) and /(2a), it is easy to see that for
a [9/100,

3/9], the equation cos(/2) = /(2a) has exactly one solu-

tion,
0
[17/100, 33/100], and that the equation cos(/2) = /(2a)
has no solution for > 0.
It remains to prove that equation (3) has no complex roots with positive
real part for = 0. Suppose = + i, > 0 is a solution of (3). Then
i is also a solution. So we can consider > 0. We note that is a
root of (3) at = 0 if and only if
(7) = a(e

cos +e
2
cos 2)
and
(8) = a(e

sin +e
2
sin2).
If > 0, (8) implies that [ [ 2a. Therefore we only need to look for
solutions of (7)(8) for (0, 2a]. Now (7) implies
(9)

a cos
= e

+
cos 2
cos
e
2
.
Since cos 2t/ cos t 1 if t (0, /3], the righthand side of (9) is positive.
When 1/6 < a

3/9, we need only to look for solutions of (7)(8) with

(/3, 2a). But
K() = ae

sin ae
2
sin2
has a positive derivative for such and K(/3) > 0. Hence (3) has no
complex roots with positive real parts for = 0.
Theorem 2. The delay dierential equation (1) satisfying (iii) has a family
of Hopf periodic solutions bifurcating from the origin at = 0.
Proof: Clearly from (i), (ii) and (iii) the delay dierential equation (1)
satises the conditions (a) and (b) of Theorem 5 from Appendix 1.
Derivating the characteristic equation (3) with respect to we get

() = (a +)(e
()
+ 2e
2()
)

() (e
()
+e
2()
).
Evaluating the above equation at = 0 and noticing that (0) = i
0
we
obtain

(0) =
e
i
0
+e
2i
0
1 a(e
i
0
+ 2e
2i
0
)
,
4 J. LLIBRE AND A. TART A
or equivalently

(0) =
2 cos(

0
2
)
2
(1 + 3a 2 cos
0
) +i sin
0
(1 +a + 2 cos
0
)
1 + 5a
2

2
2a[(1 2a) cos
0
+ 2 cos 2
0
]
=

(0) +i

(0).
Hence, we obtain that

(0) =
2 cos(

0
2
)
2
(1 + 3a 2 cos
0
)
1 + 5a
2

2
2a[(1 2a) cos
0
+ 2 cos 2
0
]
.
Substituting
0
with 2v and noticing that cos v = v/a we get

(0) =
2(3a
2

2
(1 +a) 4v
2
)v
2
a
4

4
(1 +a)
2
+ 4a
3

3
(7 + 2a)v
2
32v
4
.
Replacing 2v with
0
, we have

(0) =
(3a
2

2
(1 +a)
2
0
)
2
0
2a
4

4
(1 +a)
2
+ 2a
3

3
(7 + 2a)
2
0
4
4
0
.
In the following we prove that

(0) is positive. Since the denominator of

(0) is positive (because is the the square module of a complex number) it is

enough to show that the numerator is positive, i.e. 3a
2

2
(1 +a)
2
0
> 0.
This is obvious because the derivative of h(a) = 3a
2

2
(1 + a)
2
0
is
positive for a [9/100,

3/9] and since

0
= 0.544648 for a = 9/100, we
have that h(9/100) = 0.0110011 > 0.
In particular the transversality condition

(0) > 0 of statement (c) of

Theorem 5 is satised. Consequently, by Proposition 1, assumptions (c)
and (d) of Theorem 5 of Appendix 1 hold. In short, we can apply the Hopf
bifurcation theory to equation (1) at = 0. Therefore, a family of Hopf
periodic solutions bifurcates from the origin at = 0.
If we are in the assumptions of Theorem 2 due to Theorem 5 of Appendix
1, the following result is satised.
Theorem 3. For the delay dierential equation (1) satisfying (iii) the fol-
lowing statements hold.
(a) There is an
0
> 0 and an analytic function () =
2

2
+ O(
3
)
for 0 < <
0
such that for each (0,
0
) there exists a periodic
solution p

(t) occuring for = ().

(b) The period T() = 2[1 +
2

2
+ O(
3
)]/
0
of p

(t) is an analytic
function.
(c) The periodic solution p

(t) is orbitally asymptotically stable if

2
<
0, and unstable if
2
> 0.
HOPF BIFURCATION OF A DELAY DIFFERENTIAL EQ. WITH TWO DELAYS 5
Now we shall compute
2
,
2
and
2
for some function G
1
and G
2
.
For the delay dierential equation (1) we have that r = 2. Now we
follow Appendix 2. The function from the Riesz representation theorem is
d(, ) = (a + )[( + 2) + ( + 1)]. For equation (1) it follows that
A and R are
(10) A()() =
_

_
d
d
if 2 < 0,
(a +)[(1) +(2)] if = 0,
and
R() =
_
_
_
0 if 2 < 0,
f(, ) if = 0,
where
(11)
f(, ) = (a +)[G
2
((0), (1), (2))+
((1) +(2))G
1
((0), (1), (2))+
G
2
((0), (1), (2))G
1
((0), (1), (2))].
We dene q() = e
i
0

and q

() = De
i
0

.
Lemma 4. Normalizing q and q

by the condition q

, q) = 1 we obtain
that D = ReD + ImD where
ReD =
2a
3

3
2a
4

4
+ 7a
2

2
0
2
4
0
2a
3

3
(a 1)
2
+ 2a
2

2
(7 + 2a)
2
0
4
4
0
,
ImD =
3a
2

0
_
4a
2

2
0
+ 2
3
_
4a
2

2
0
2a
3

3
(a 1)
2
+ 2a
2

2
(7 + 2a)
2
0
4
4
0
.
Proof: Since
q

, q) = q

(0)q(0)
_
0
=2
_
_

=0
q

( )q()d
_
d()
=

D
_
0
=2
_
_

=0

De
i
0
()
e
i
0

d
_
d()
=

D +a

D
_
0
=2
_
_

=0
e
i
0

d
_
(( + 1) +( + 2))d
=

D +a

D
_
0
=2
e
i
0

(( + 1) +( + 2))d
=

D
_
1 a(e
i
0
+ 2e
2i
0
)
_
,
6 J. LLIBRE AND A. TART A
we obtain

D =
_
1 a(e
i
0
+ 2e
2i
0
)
_
1
= (1 a(cos
0
+ 2 cos 2
0
) +ia(sin
0
+ 2 sin2
0
))
1
.
Rationalizing the above equation, we get

D =
1 a(cos
0
+ 2 cos 2
0
) ia(sin
0
+ 2 sin2
0
)
1 + 5a
2

2
2a[(1 2a) cos
0
+ 2 cos 2
0
]
.
Taking into account that cos(
0
/2) =
0
/(2a), we obtain the value of
D.
From the denition of z and w in Appendix 2, and (24) we have
z(t) = i
0
z(t) +

Df(w(z(t), z(t), 0) + 2Re(z(t)q(0)), 0)
= i
0
z(t) +

Df
0
(z(t), z(t)) (12)
= i
0
z +g(z, z).
Next step is to solve system (29) and nd the coecients w
ij
for i +j = 2.
From the equations (26) and (27) we have for [2, 0) that
H(z, z, ) =

Df
0
q() D

f
0
q() = gq() g q().
Using the expression for H obtained in (28) and for g given in (30), we get
(13) H
20
() = g
20
q() g
02
q(),
(14) H
11
() = g
11
q() g
11
q().
Using the denition of A, the rst equation of (29) and (13) we have
w
20
() = A()(w
20
())
= 2i
0
w
20
() H
20
()
= 2i
0
w
20
() +g
20
q() + g
02
q().
Noticing that q() = e
i
0

, we solve the linear dierential equation in the

variable w
20
(), and we obtain
(15) w
20
() =
i

0
g
20
e
i
0

+
i
3
0
g
02
e
i
0

+ce
2i
0

,
where c is a complex constant.
Similarly, using the denition of A, the second equation of (29) and (14) we
obtain
w
11
() = g
11
q() + g
11
q(),
which leads to
(16) w
11
() =
i

0
g
11
e
i
0

+
i

0
g
11
e
i
0

+d,
HOPF BIFURCATION OF A DELAY DIFFERENTIAL EQ. WITH TWO DELAYS 7
where d is also a complex constant.
In the following we have to nd the constants c and d. From equations (26),
(27) and (12), for = 0 we have that
H(z, z, 0) =

Df
0
q(0) D

f
0
q(0) +f
0
= gq(0) g q(0) +
q

(0)f
0
q

(0)
= g g +
g

D
.
Using the expression for H obtained in (28) and for g obtained in (30), we
get
(17) H
20
(0) = g
20
g
02
+
g
20

D
,
(18) H
11
(0) = g
11
g
11
+
g
11

D
.
From the denition of A (see (10)), the rst two equations of (29), (17) and
(18), for = 0 we have that
a(w
20
(1) +w
20
(2)) = 2i
0
w
20
(0) H
20
(0)
= 2i
0
w
20
(0) g
20
g
02
+

D
1
g
20
,
and
a(w
11
(1) +w
11
(2)) = H
11
(0)
= g
11
g
11
+

D
1
g
11
.
Substituting w
20
and w
11
in the above equations we obtain
a(
i

0
g
20
e
i
0
+
i
3
0
g
02
e
i
0
+ce
2
0
+
i

0
g
20
e
2i
0
+
i
3
0
g
02
e
2i
0
)+ce
4
0
)
= 2(g
20
+
1
3
g
02
ic
0
) g
20
g
02
+

D
1
g
20
,
and
a (
i

0
g
11
e
i
0
+
i

0
g
11
e
i
0
+d
i

0
g
11
e
2i
0
+
i

0
g
11
e
2i
0
+d)
= g
11
g
11
+

D
1
g
11
.
Solving these two equations for c and d, we get
c =
ia

0
(g
20
e
i
0
+
1
3
g
02
e
i
0
+g
20
e
2i
0
+
1
3
g
02
e
2i
0
)3g
20

5
3
g
02
+

D
1
g
20
ae
2i
0
+ae
4i
0
+ 2i
0
,
and
d =
ia

0
(g
11
e
i
0
g
11
e
i
0
+g
11
e
2i
0
g
11
e
2i
0
) +g
11
+ g
11

D
1
g
11
2a
.
8 J. LLIBRE AND A. TART A
From the denition of w we have that x
t
() = w(t, ) + zq() + z q().
Replacing w with the expression obtained in (23), we nd that
(19)
x
t
() = ze
i
0

+ ze
i
0

+w
20
()
z
2
2
+w
11
()z z +w
02
()
z
2
2
+O([ z, z [
3
).
Taking into account (11), (12) and (19), we can nd g
20
, g
11
, g
02
, g
21
and

2
=
1
2
0

(0)
(
0
Re(g
21
) + Im(g
20
)Re(g
11
) + Im(g
11
)Re(g
20
)),

2
=
1
6
2
0
[Im(g
02
)
2
6Im(g
11
)
2
3Im(g
11
)Im(g
20
)Re(g
02
)
2
6Re(g
11
)
2
+ 3Re(g
11
)Re(g
20
) + 3Im(g
21
)
0
]
1

(0)
and

2
=
1

0
(
0
Re(g
21
) Im(g
20
)Re(g
11
) Im(g
11
)Re(g
20
)).
3. Applications
In this section we give some applications of the theory developed in Sec-
tion 2 .
3.1. Example 1. We consider the delay dierential equation (1) with
9/100 a

3/9, G
1
= mx and G
2
= mxy + mxz + nxz
2
where
m, n are arbitrary real numbers. Thus equation (1) has a Hopf periodic
orbit for the value of the parameter =
2

2
+ O(
3
) having period
T =
2

0
(1 +
2

2
) + O(
3
), where
0
is the imaginary part of the pure
imaginary root of the characteristic equation of delay dierential equation
(1). This periodic orbit is asymptotically stable if
2

2
+ O(
3
) < 0 and
unstable if
2

2
+O(
3
) > 0. In this case
2
,
2
and
2
are given by

2
=
1
6a 4 cos(w
0
) + 2
(an
_
5
2
a
2
+ 2(2a 1) cos(w
0
)a
4 cos(2w
0
)a+1) sec
2
_
w
0
2
_
(ReD(cos(4w
0
)+2)ImDsin(4w
0
))),
HOPF BIFURCATION OF A DELAY DIFFERENTIAL EQ. WITH TWO DELAYS 9

2
=
1
w
0
(3a + 2 cos(w
0
) 1)
(an sec
_
w
0
2
_
(6aImD cos
_
w
0
2
_
2ImDcos
_
3w
0
2
_
ImDcos
_
5w
0
2
_
+aImD cos
_
7w
0
2
_
+ 2aImD cos
_
9w
0
2
_
+ 2aReD sin
_
w
0
2
_
+ 2ReDsin
_
3w
0
2
_
ReDsin
_
5w
0
2
_
+aReD sin
_
7w
0
2
_
+ 2aReD sin
_
9w
0
2
_
)),
and

2
= 2an(ReD(cos(4w
0
) + 2) ImDsin(4w
0
)),
where ReD and ImD are given in Lemma 4.
3.2. Example 2. We consider the delay dierential equation (1) for a =

3/9. It follows that

0
= /3. We take G
1
= 0 and G
2
= mx
2
+ny
2
+pz
2
with m, n, p arbitrary real numbers. Here
2
,
2
and
2
are

2
=
=
1

21
(2

3(59049
7
(34234m
2
+ 1200nm15917pm15917n
2
133336p
2
132136np)3486784401

3(8m
2
+9nm+5pm+5n
2
5p
2
+ 4np) 117649(2m
2
+12nm+11pm+11n
2
+28p
2
+40np)
13
+ 151263

3(16m
2
+ 5nm3pm3n
2
49p
2
44np)
12
+ 194481(62m
2
+ 8nm23pm23n
2
224p
2
216np)
11
+ 3695139

3(8m
2
+ 9nm+ 5pm+ 5n
2
5p
2
+ 4np)
10
+ 13931190(2m
2
+ 12nm+ 11pm+ 11n
2
+ 28p
2
+ 40np)
9
275562

3(964m
2
131nm613pm613n
2
4249p
2
4380np)
8
531441

3(9208m
2
+5497nm+893pm+893n
2
20341p
2
14844np)
6
9565938(1778m
2
+ 944nm+ 55pm+ 55n
2
4280p
2
3336np)
5
15943230

3(1468m
2
+1203nm+469pm+469n
2
2263p
2
1060np)
4
43046721(1138m
2
+ 848nm+ 279pm+ 279n
2
2008p
2
1160np)
3
129140163

3(344m
2
+ 335nm+ 163pm+ 163n
2
371p
2
36np)
2
387420489(122m
2
+ 108nm+ 47pm+ 47n
2
164p
2
56np))),
where

21
= 117(27 + 3

3 + 7
2
)(27

3 +(9 + 7

3))
2

(19683 +(6561

3 +(17496
+ (3483

3 + 7(648 + 7(9

3 + 7))))));
10 J. LLIBRE AND A. TART A

2
=
=
1

21
(2(8m
2
+ 9nm+ 5pm+ 5n
2
5p
2
+ 4np)(31381059609

3
+ (66248903619 +(65861483130

3 +(91043814915
+ (47939698287

3 +(46193914602 +(15740750979

3
+ (10381699935 + 7(323883765

3 + 7(17281674
+ 7(187353

3+7(11421+7(810

3+7(45+7

3))))))))))))))),
where

21
=
= 13(729 +(162

3 +(405 + 7(6

3 + 7))))
2
(19683 +(6561

3
+ (17496 +(3483

3 + 7(648 + 7(9

3 + 7))))));
and

2
=
=
1

21
(6(59049
7
(34234m
2
+ 1200nm15917pm15917n
2
133336p
2
132136np) + 3486784401

3(8m
2
+ 9nm+ 5pm+ 5n
2
5p
2
+ 4np) + 117649(2m
2
+12nm+11pm+11n
2
+28p
2
+40np)
13
151263

3(16m
2
+ 5nm3pm3n
2
49p
2
44np)
12
194481(62m
2
+ 8nm23pm23n
2
224p
2
216np)
11
3695139

3(8m
2
+ 9nm+ 5pm+ 5n
2
5p
2
+ 4np)
10
13931190(2m
2
+ 12nm+ 11pm+ 11n
2
+ 28p
2
+ 40np)
9
+ 275562

3(964m
2
131nm613pm613n
2
4249p
2
4380np)
8
+ 531441

3(9208m
2
+5497nm+893pm+893n
2
20341p
2
14844np)
6
+ 9565938(1778m
2
+944nm+55pm+55n
2
4280p
2
3336np)
5
+ 15943230

3(1468m
2
+1203nm+469pm+469n
2
2263p
2
1060np)
4
+ 43046721(1138m
2
+ 848nm+ 279pm+ 279n
2
2008p
2
1160np)
3
+ 129140163

3(344m
2
+ 335nm+ 163pm+ 163n
2
371p
2
36np)
2
+ 387420489(122m
2
+ 108nm+ 47pm+ 47n
2
164p
2
56np))),
where

21
= 13(729 +(162

3 +(405 + 7(6

3 + 7))))
2
(19683
+ (6561

3+(17496+(3483

3+7(648+7(9

3+7)))))).
HOPF BIFURCATION OF A DELAY DIFFERENTIAL EQ. WITH TWO DELAYS 11
3.3. Example 3. We consider the delay dierential equation (1) for
9/100 a

3/9, G
1
= x and G
2
= xy + xz + x
3
. Thus,
2
,
2
and
2
are given by

2
=
6aReD
_
5
2
a
2
+ 2(2a 1) cos(w
0
)a 4 cos(2w
0
)a + 1
_
(6a 4 cos(w
0
) + 2)(cos(w
0
) + 1)
,

2
=
1
w
0
(3a 2 cos(w
0
) + 1)
(3a sec
_
w
0
2
_
(3aImD cos
_
w
0
2
_
ImDcos
_
3w
0
2
_
+ ReD(a + 2 cos(w
0
) + 1) sin
_
w
0
2
_
)),
and

2
= 6aReD,
where ReD and ImD are given in Lemma 4.
4. Appendix 1: The Hopf bifurcation Theorem
The following result can be found in [8].
Theorem 5. We consider the delay dierential equation
(20) x(t) =
dx(t)
dt
= F(x(t), x(t r
1
), . . . , x(t r
n1
), )
If
(a) F is analytic in x and in a neighborhood of (0, 0) in R
n
R,
(b) F(0, ) = 0 for in an open interval containing 0, and x(t) = 0 is
an isolated stationary solution of (20),
(c) the characteristic equation of (20) has a pair of complex conjugate
eigenvalues and

such that () = () + i() where (0) =

0
> 0, (0) = 0,

(0) ,= 0,
(d) the remaining eigenvalues of the characteristic equation have strictly
negative real parts,
then the delay dierential equation (20) has a family of Hopf periodic solu-
tions. More precisely, there is an
0
> 0 and an analytic function () =

i=2

i
for 0 < <
0
such that for each (0,
0
) there exists a
periodic solution p

(t) occurring for = (). If () is not identically

zero, the rst nonvanishing coecient
i
has an even subscript, and there
is an
1
(0,
0
] such that () is either strictly positive or strictly neg-
ative for (0,
1
). For each L > 2/
0
there is a neighborhood V of
x = 0 and an open interval I containing 0 such that for any I the
only nonconstant periodic solutions of the delay dierential equation (20)
with period less then L which lie in V are members of the family p

(t)
12 J. LLIBRE AND A. TART A
for values of a satisfying () = , (0,
0
). For 0 < <
0
the pe-
riod T() = 2
_
1 +

i=2

/
0
of p
(t)
is an analytic function. Exactly
two of the Floquet exponents of p

(t) approach 0 as 0. One is 0 for

(0,
0
), and the other is an analytic function () =

i=2

i
for
0 < <
0
. The periodic solution p

(t) is orbitally asymptotically stable if

() < 0, and unstable if () > 0.
5. Appendix 2: The algorithm for computing the period and
the stability of the Hopf periodic orbit
The following algorithm also follows from [8]. Since the algorithm for
computing the period and the stability of the Hopf periodic orbit does not
depend from the fact that there is one or more delays, we describe it for
one delay dierential equation with a unique delay.
Consider the autonomous equation
(21)
dx(t)
dt
= L

x
t
+f(x
t
(), ), t > 0, R,
= 0,
where for some r > 0
x
t
() = x(t +), x : [r, 0] R, [r, 0].
We denote by C[r, 0] the set of all continuous functions from [r, 0] to R.
In C[r, 0] we put the topology of the supremum. Then we dene C as the
set of all continuous functions from C[r, 0] to R. An orbit corresponding to
a solution x(t) of (21) is a curve in C traced out by the family of functions
x(), (x
t
() = x(t + )) as t rangers over (0, ); the orbit of a periodic
solution is a closed curve in C. The individual periodic orbits will belong
to slices C

, (constant) of C.
Here L

: C[r, 0] R and the operator f(, ) : C[r, 0] R contains

the nonlinear terms, beginning with at least quadratic terms, i.e.
f(0, ) = 0, f

x
(0, ) = 0.
For simplicity we consider that f(, ) is analytic and L

depends analyti-
cally on the bifurcation parameter for [ [ small.
In the following we transform the linear problem x = L

x
t
. By the Riesz
representation theorem, there exists a function (, ) of bounded variation
for [r, 0] such that for all C[r, 0],
L

=
_
0
r
()d(, ).
HOPF BIFURCATION OF A DELAY DIFFERENTIAL EQ. WITH TWO DELAYS 13
In particular
L

x
t
=
_
0
r
x(t +)d(, ).
For example, if L

x
t
= x(t 1), then d(, ) = ( + 1), where () is
the Dirac delta function. The choice d(, ) = () corresponds to the
ordinary dierential equation x = x.
We make the usual Hopf assumption on the spectrum () =
=
_
[ L

= 0
_
of L

; namely
(1) there exists a pair of complex, simple eigenvalues () and

()
such that () = () +i(), where and are real and (0) =
0, (0) =
0
> 0, and

(0) ,= 0 (the transversality hypothesis);

(2) all the other elements of (0) have negative real parts.
We identify () by choosing (0) = i
0
. Of course, L

= 0 is the
characteristic equation of (21).
Next we dene for C
1
[r, 0]
A()() =
_

_
d
d
if r < 0,
_
0
r
(s)d(s, ) = L

if = 0,
and
R() =
_
_
_
0 if r < 0,
f(, ) if = 0.
Then since dx
t
/d = dx
t
/dt, (21) can be written as
(22) x
t
= A()(x
t
) +R(x
t
),
which is a more mathematically pleasing one because this equation involves
a single unknown variable x
t
rather than both x and x
t
.
We shall obtain explicit expressions only for
2
,
2
and
2
(see Appendix
1). We dene q() to be the eigenvector for A(0) corresponding to (0);
namely A(0)(q()) = i
0
q(). The adjoint operator A

(0) is dened by
A

(0)((s)) =
_

_
d
ds
if 0 < s r,
_
0
r
(t)d(t, 0) if s = 0.
We shall henceforth simply right A for A(0), A

for A

(0), (s) for (s, 0),

etc. Since A(q()) = (0)q() = i
0
q(),

(0) is an eigenvalue for A

, and
A

(q

) = i
0
q

for some nonzero vector q

.
14 J. LLIBRE AND A. TART A
To construct coordinates to describe C
0
near 0 R
3
, we need an inner
product. For C[0, r] and C[r, 0] that is dened by
, ) =

(0)(0)
_
0
=r
_
_

=0

( )()d
_
d(),
Then, as usual, , A) = A

, ) for (, ) D(A) D(A

), where D(A)
denotes the denition domain of A. We normalize q and q

by the condition
q

) = 1. Of course q

, q) = 0, since i
0
is a simple eigenvalue for A. For
each x D(A), we may than associate the pair (z, w) where z = q

, x) and
w = x zq z q = x 2Re(zq).
For x
t
a solution of (22) at = 0, we dene z(t) = q

, x
t
) and then
dene w(t, ) = x
t
() 2Re(z(t)q(t)). On the manifold C
0
, w(t, ) =
w(z(t), z(t), ) where
(23) w(z, z, ) = w
20
()
z
2
2
+w
11
()z z +w
02
()
z
2
2
+w
30
()
z
3
6
+. . . .
In eect, z and z are local coordinates for C
0
in C in the directions of q

and q

. Note that w is real if x

t
is; we shall deal with real solutions only.
It is easy to see that q

, w) = 0.
Now, for solutions x
t
C
0
of (22), q

, x
t
) = q

, A(x
t
) +R(x
t
)) or, since
= 0,
z(t) = q

, x
t
) + q

, A(x
t
) +R(x
t
))
= i
0
z(t) + q

, A(x
t
)) + q

, R(x
t
))
= i
0
z(t) + q

(0)f(x
t
, 0) (24)
= i
0
z(t) + q

(0)f(w(z(t), z(t), 0) + 2Re(z(t)q(0)), 0)

= i
0
z(t) + q

(0)f
0
(z(t), z(t)),
which we write in abbreviated form as
(25) z = i
0
z +g(z, z).
Our next object is to expand g in powers of z and z; and then to obtain
from this expansion the values of
2
,
2
and
2
. First, it is required to derive
equations for the coecient w
ij
(). We write w = x
t
zq

z q and use (25)
and (22) to obtain
(26) w =
_
_
_
A(w) 2Re( q

(0) f
0
q()) if r < 0,
A(w) 2Re( q

(0) f
0
q(0)) +f
0
if = 0,
HOPF BIFURCATION OF A DELAY DIFFERENTIAL EQ. WITH TWO DELAYS 15
which we rewrite as
(27) w = A(w) +H(z, z, ),
where
(28) H(z, z, ) = H
20
()
z
2
2
+H
11
()z z +H
02
()
z
2
2
+. . . .
On the other hand, on C
0
near the origin w = w
z
z+w
z

z. Using (23) and (25)
to replace w
z
and z and their conjugates by their power series expansions
(which involve the w
ij
), we get a second expression for w. We equate this
to the righthand side of (27). The result is an equation from which we can
derive equations for the nvectors w
ij
(), (i +j = 2, 3, . . . ). These are
(29)
2i
0
w
20
() A(w
20
()) = H
20
(),
A(w
11
()) = H
11
(),
2i
0
w
02
() A(w
02
()) = H
02
(),
. . .
Now the H
ij
with i + j = 2 do not involve any of the w
ij
with i + j > 2.
Further, by hypothesis 2i
0
and 0 are not eigenvalues of A. Thus, the rst
three equations (29) can be solved for w
20
, w
11
and w
02
= w
20
. At each stage
the equations for w
ij
(i +j k +1) only involve via the H
ij
coecients w
ij
with i +j k. Hence the equations (29) can be solved successively for the
w
ij
. Only the values of w
ij
(i +j = 2) are needed to compute
2
,
2
and
2
.
If
2k
,
2k
and
2k
are desired for some k > 1, then must not be set equal
to 0 in the previous analysis.
Once the w
ij
are determined, the dierential equation (25) for z can be
written as
(30) z = i
0
z +g
20
z
2
2
+g
11
z z +g
02
z
2
2
+g
21
z
2
z
2
+. . . ,
where the coecients g
ij
for i +j 3 may be computed by expanding the
expression
(31) q

(0) f
_
zq() + z q() +w
20
()
z
2
2
+w
11
()z z +w
02
()
z
2
2
_
.
The coecient c
1
(0) of the Poincare normal form (see [8], pp. 2536, 4551
and [1], chapters 5 and 6) is given in terms of these g
ij
by formula
c
1
(0) =
i
2
0
_
g
20
g
11
2 [ g
11
[
2

1
3
[ g
02
[
2
_
+
g
21
2
.
The following formulas give us the values of
2
,
2
and
2
(32)
2
=
Re(c
1
(0))

(0)
,
16 J. LLIBRE AND A. TART A
(33)
2
=
1

0
[Im(c
1
(0)) +
2

(0)] ,
(34)
2
= 2Re(c
1
(0)).
For more details see [8] pages 29, 31 and 44, respectively. We recall that
=
2

2
+ O(
3
), T() = 2(1 +
2

2
+ O(
3
))/
0
, and that the Hopf
periodic orbit is orbitally asymptotically stable if
2
< 0, and unstable if

2
> 0.
Acknowledgements
The rst author was partially supported by a DGICYT grant number
MTM200506098C0201 and by a CICYT grant number 2005SGR 00550;
the second author thanks to Centre de Recerca Matem`atica for its hospi-
tality and facilities for doing this work.
References
[1] V. Arnold, Chapitres supplementaires de la theorie des equations dierentielles
ordinaires, Mir, Moscow, 1984.
[2] P. Bi and M. Han, Bifurcation of periodic solutions of delay dierential equation
with two delays, J. Math. Anal. Appl. 284 (2003), 548563.
[3] P. Bi, M. Han and Y. Wu, Existence and bifurcation of periodic solutions of three
dimensional delay dierential equations, J. Bifurcation and Chaos 14 (2004), 3921
3929.
[4] S.-N. Chow and J. MalletParet, Integral averaging and bifurcation, J. Dierential
Eqns. 26 (1977), 112159.
[5] S.-N. Chow and J. MalletParet, The Fuller index and global Hopf bifurcation,
J. Dierential Eqns. 29 (1978), 6685.
[6] K. Gopalsamy, Stability and oscilations in delay dierential equations of popula-
tion dynamics, Mathematics and Its Applications 74, Kluwer Academic Publisher,
London, 1992.
[7] J.K. Hale, S.M.V. Lunel, Introduction to functional dierential eqations, Appl.
Math. Sciences 99, SpringerVerlag, New York, 1993.
[8] B.D. Hassard, N.D. Kazarinoff and YH. Wan, Theory and applications of Hopf
bifurcation, London Math. Soc. Lecture Note Series 41, Cambridge University Press,
Cambridge, 1981.
[9] E. Hopf, Abzweigung einer periodischen losung eines dierential systems, Berichen
Math. Phys. Kl. Sach. Akad. Wiss. Leipzig 94 (1942), 122.
[10] G.S. Jones, Periodic motions in Banach space and applications to functional dif-
ferential equations, Contrib. Di. Eqns. 3 (1964), 75106.
[11] J.L. Kaplan and J.A. Yorke, Ordinary dierential equations which yield periodic
solutions of dierential delay equations, J. Math. Anal. Appl. 48 (1974), 317324.
[12] J.E. Marsden and M.F. MacCracken, The Hopf bifurcation and its applications,
Appl. Math. Sci. 19, SpringerVerlag, 1976.
[13] X. Meng and J. Wei, Stability and bifurcation of mutual system with time delay,
Chaos, Solitons and Fractals 21 (2004), 729740.
HOPF BIFURCATION OF A DELAY DIFFERENTIAL EQ. WITH TWO DELAYS 17
[14] R.G. Nussbaum A Hopf bifurcation theorem for retarded functional dierential equa-
tions, Trans. Amer. Math. Soc. 238 (1978), 139164.
[15] L.P. Shayer and S.A. Campbell, Stability, bifurcation, and multistability in a
system of two coupled neurons with multiple time delays, SIAM Appl. Math. 61
(2000), 673700.
[16] Y. Song, M. Han and J. Wei, Stability and Hopf bifurcation analysis on a simplied
BAM neural network with delays, Physica D 200 (20059), 185204.
[17] L. Wang and X. Zou, Hopf bifurcation in bidirectional associative memory neural
networks with delays: analysis and computation, J. Comput. Appl. Math. 167 (2004),
7390.
1
Departament de Matem` atiques, Universitat Aut` onoma de Barcelona, 08193
Bellaterra, Barcelona, Spain
E-mail address: jllibre@mat.uab.es
2
Department of Applied Mathematics, Babes Bolyai University, 1 Kogal-
niceanu str., 400084 ClujNapoca, Cluj, Romania
3
Centre de Recerca Matem` atica, Apartat 50, 08193, Bellaterra, Spain
E-mail address: atarta@crm.es