J. Austral. Math. Soc. Ser. B 35(1993), 103-113
NUMERICAL COMPUTATION OF SYMMETRY-BREAKING
BIFURCATION POINTS
B. S. ATTILI1
(Received 14 February 1989; revised 26 February 1991)
Abstract
We consider symmetry-breaking bifurcation points which arise in parameterdependent nonlinear equations of the form /(JC, X) = 0. These types of bifurcation points are connected to pitchfork bifurcation points. A direct method is used to
compute such points. Multiple shooting is used to discretise the two-point boundaryvalue problems to obtain a finite-dimensional problem.
1. Introduction
Many physical problems can be formulated as a parameter-dependent nonlinear
equation of the form
/Oc,A.) = 0,
(1.1)
where / is a C2-function which maps Rn x R into R". The point (x0, Xo) is
called a regular point of f(x, X) = 0 if fx(x, X) is nonsingular. It is called a
simple singular point of f{x, X) = Oif fx(x, X) is singular; / j 0 = fx(x0, Xo) and
f°J each has one-dimensional null space spanned by 0o and Vo respectively.
Whether (or not) bifurcation occurs at a singular point depends on the type
of singularity, which is described by (in)equalities involving the derivatives of
f(x,X) at (x0, Xo). Typically, turning points, and transcritical and pitchfork
bifurcation occur.
'K.F.U.P.M., P.O. Box 1927, Dhahran 31261, Saudi Arabia
© Australian Mathematical Society, 1993, Serial-fee code 0334-2700/93
103
Downloaded from https://www.cambridge.org/core. IP address: 3.235.21.12, on 27 May 2020 at 15:54:44, subject to the Cambridge Core terms of
use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0334270000007293
104
B. S. Attili
[2]
To compute such points, we use the extended system proposed by Griewank
and Reddien [4, 5]; that is,
where g(x, k) = uJ fx(x, k) • v and u and v are the left and right null vectors of
/JJ0 respectively. More on this system will be given in Section 3 (see also Attili
[1,2]).
The outline of this paper will be as follows. Some assumptions and definitions
are given in Section 2. In Section 3, we present the relation between symmetrybreaking bifurcation points and pitchfork bifurcation, and details on the extended
system will also be presented. In this section we also present the main theorem
of this paper, which shows that a symmetry-breaking bifurcation point is an
isolated solution of the extended system. Numerical details are given in the final
section.
It is important to note that Werner and Spence [9] have considered computation of such points using a different extended system; also some other
investigators considered systems which do not utilise the symmetry, see Moore
[6] and Werner [10]. But the system considered here is more direct, and reduces
the number of equations and unknowns considerably, in comparison with the
system considered by Werner and Spence [9].
2. Assumptions and definitions
Before defining what is meant by a symmetry-breaking bifurcation point,
we shall recall some standard theory on bifurcation. The point (*0> k0) is a
simple singular point of f(x, X) — 0 if f(x0, k0) = 0, N(f°) =span{</>0} and
*(/?) = iy € R" : fh = 0}.
DEFINITION 2.1. A simple bifurcation point (xo, k0) is called a pitchfork bifurcation point of f(x, k) = Oif V0TA° = 0, ^o T />o0o = Oand ^ ( / A + Z i ^ o V o ) #
0 where f°v0 = — / x °. If Vro"/tox0o0o # 0, the bifurcation point is called transcritical.
We shall assume that (Al): There exists 5 e L(R"), the bounded linear
operators mapping R" into R", with 5 # / and S2 = I and f(Sx, A.) = Sf(x, k)
for all JC e Rnandk e Rn.
Downloaded from https://www.cambridge.org/core. IP address: 3.235.21.12, on 27 May 2020 at 15:54:44, subject to the Cambridge Core terms of
use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0334270000007293
[3]
Numerical computation of symmetry-breaking bifurcation points
105
Here (Al) will give rise to the natural decomposition
R" = R"s 0 R"a,
(2.1)
/?? = [x e R" : Sx = x]
(2.2)
where
is the set of symmetric elements, and
Rna = {x e R" : Sx = -x}
(2.3)
is the set of antisymmetric elements, see Sattinger [7]. The decomposition arises
because x e R" can be written as 2x = (/ + S)x + (/ - S);t,where Sx e R"
and (/ - S)x e R"a. To see this relation, note that S(I + S)x = Sx + S2x =
Sx + Ix = (/ + 5)JC which implies that (/ + S)x e Rns and also (/ - S)x e Rna
since S(/ - S)* = (5 - S2)x = (5 - 7)jc = - ( / - 5)JC.
DEFINITION 2.2. If (x0, Xo) is a singular point of f(x, X) = 0, then (x0, Xo) is
called a symmetry- breaking bifurcation point of f(x, k) ifx0 e ^?" and0 o € /?^
where span{0o} =
3. Symmetry breaking and pitchfork bifurcation points
The singular points in definition (2.2) will be shown to be pitchfork bifurcation
points of f{x, X) = 0 under certain generic conditions. Before doing that, we
state some results which can be obtained from assumption (Al).
LEMMA
3.1. Assume that (Al) holds. Then for all X e R andx, w,v 6 Rn,we
have
MSx,X)
fASx,X)Sv
fkxiSx,X)Sv
fxxiSx,X)SvSw
=
SfdSx.k),
=
SfASx,X)v,
= SfkxiSx,X)v,
=
P
-
;
SfxxiSx,X)vw.
All results in (3.1) can be obtained by differentiating fiSx, X) = Sfix, X)
with respect to X or x. Using (3.1), one can show that for X e R and x e /?"
fix,X),Mx,X)eR"s.
(3.2)
This is true if x e /?; then Sx = x and so fix, X) = fiSx, X) = Sfix, X),
which implies that fix,X) e R". The same can be said about /x(x, X). Note
Downloaded from https://www.cambridge.org/core. IP address: 3.235.21.12, on 27 May 2020 at 15:54:44, subject to the Cambridge Core terms of
use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0334270000007293
106
B. S. Attili
that R" and /?£ are invariant with regard to fx(x, k) and fkx(x,
since if v € R" then Sv = v and so
fAx,k)v = Sfx(x,k)v.
[4]
k). This follows
(3.3)
Also, for v, w 6 R" or v, w e fl", fxx(x, k)vw e /?". The reason is that
if v, w are in R", then Sv = v and 5io = w, which yields / „ ( * , A)vu> =
fxx(x, X)Svw = Sfxx(x, X)vw, and thus fxx(x, k)vw e /?;. If v e R", w e R"
thenfxx(x,k)vw
e Rna.
LEMMA 3.2. For every x € /?", we have ^x = 0.
PROOF. Since R" is invariant with respect to f° as in (3.3), then R" is the range
of ./j 0 . The conclusion now follows from the Fredholm Alternative.
Standard theory of bifurcation shows that at a symmetry-breaking bifurcation point, the solution set of (1.1) will consist of two smooth transversally
intersecting branches (see Brezzi, Rappaz and Raviart [3]); that is,
C, = {(x, k):k = ko
and
Ca = {(JC, k) : k =
k0+oU2),
where Cs and Ca refer to the symmetric and antisymmetric branches respectively.
The Implicit Function Theorem will guarantee that Cs C R" since f£ is restricted
to R" is an isomorphism of /?". Hence Ca bifurcates from a symmetric branch of
solution Cs at a symmetry-breaking bifurcation point. Such branching is given
by 0o € /?a- To compute symmetry-breaking bifurcation points, let us consider
the extended system
-ulfx(x,k)v
where u = \jr0 and v = <pQ, the left and right null vectors of f° respectively,
g{x, k) is obtained from the following lemma.
LEMMA 3.3. The systems
Downloaded from https://www.cambridge.org/core. IP address: 3.235.21.12, on 27 May 2020 at 15:54:44, subject to the Cambridge Core terms of
use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0334270000007293
[5]
Numerical computation of symmetry-breaking bifurcation points
107
and
where v and u are n x 1 vectors,
vectors, are both uniquely solvable
g is a scalar function,
and
T and R are n x l
8 = ~uTfxV,
3 ?)
f/ie prime denotes differentiation with respect to x or A..
Here 7"T and /? are normalisations for v and u respectively, which forces fx
to be singular, see Attili [1,2] and Griewank and Reddien [4, 5]. Also v and R
are chosen to be in Rna. It follows from (3.2) and (3.3) that F(x, A.) maps Rf x R
into R" x R. Hence a symmetry-breaking singular point (*o, A-o) corresponds to
a solution (JC0, A.o) of (3.4). Moreover, the next theorem shows the relevance of
(Al) for pitchfork bifurcation. See Brezzi, Rappaz and Raviart [3].
THEOREM 3.1. Assume that (Al) it holds and let (x0, A.o) be a symmetry breaking
simple singular point. Let F(x, A.), as in (3.4), be a mapping from R" x R into
R" x R then (x0, A.o) is an isolated solution of (3.4) if and only if(xo, A.o) is a
pitchfork bifurcation point.
PROOF.
The Jacobian of (3.4) is
At (*<), A.o), consider the system
r /?
A°
l r^i-o
L ^ o / ^ tfUfo J ' L <* \
which when expanded gives
(i)
(») ^
Since -/ A ° = / ? • u0 and /X°0O = 0, then from (3.9-i), we have
(3.10)
Downloaded from https://www.cambridge.org/core. IP address: 3.235.21.12, on 27 May 2020 at 15:54:44, subject to the Cambridge Core terms of
use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0334270000007293
108
B. S. Attili
[6]
Substituting (3.10) in (3.9-ii) we get
&
T
lf
>
= 0.
(3.11)
Moreover t/^J fxx(f>o<po = 0 since (x0, k0) is a pitchfork bifurcation point, and
hence (3.11) becomes
«[^ 0 T (/A + f?x4>ovo)] = 0.
But ^ofxX(f>o + fxxfovo # 0 since (x0, Ao) is a pitchfork bifurcation point,
this yields a = 0. Thus / = £0 O which implies t e /?", but ? e /?", then
^ e Rns PI /?^ = {0} and hence (3.10) has only the trivial solution. The solution
{t, a)J is nontrivial if and only if there exist ana e R and t e R" such that
A° + ft = o,
and
or equivalently, the null space of (3.8) is trivial if and only if
and t satisfies
A° + ft = o.
Thus (xo, ^-o) is a pitchfork bifurcation point.
The above theorem shows that the symmetry-breaking bifurcation point is an
isolated solution of the extended system (3.4), hence it will be used to compute
such points. Moreover, the same system can be used to compute certain double
singular points (x0, kQ) of f{x, k) = 0, as we shall show here that (x0, ^-o) is
again an isolated solution to (3.4), where F(x, k) = 0 (the extended system)
maps R" x R into R" x R. Now assume that
N(f?) = span{0o0i}, <£o e K - {0} and </>, e Rns - {0},
fl(/°) = {yeRn:
fry
= 0, i = 0, 1},
(3.12)
(3.13)
and
tfx
=0
forx e R",
xfrjx = 0 for* e Rna.
(3.14)
With these assumptions, we will have the following theorem.
Downloaded from https://www.cambridge.org/core. IP address: 3.235.21.12, on 27 May 2020 at 15:54:44, subject to the Cambridge Core terms of
use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0334270000007293
[7]
Numerical computation of symmetry-breaking bifurcation points
109
THEOREM 3.2. Assume that (Al) holds and let (x0, k0) be a double singular
point with x0 € R". Assume also that (3.12), (3.13) and (3.14) hold. With
F(x,k) given by (3.4) considered as a mapping on R" x R, then (x0, XQ) is an
isolated solution of (3.4) if and only if f® g range(/J°) and t/fo/t°J</>o0i # 0.
PROOF.
Consider
^!</>oJ• U _ r 0 >
(3l5)
where the matrix in (3.15) above is the Jacobian of (3.4). Equation (3.15) implies
(i)
f!'t+af°
= 0,
(ii) ^o T />or+a^ 0 T /c>o = 0.
^O)
Now since / ° £ range(/;c0), then rfrjf° ^ 0 and so (3.16-i) implies that
f] f?t+ax//] f° = 0. Further since ifrjf? = 0, t h e n a ^ A 0 = O . b u t ^ / " # 0,
this shows that a = 0. Now (3.16-i) becomes f°t — 0, and thus t e N(f°), or
equivalently t = a</>0 + b<f>\. Substituting this in (3.16-ii) we obtain
«^oT/,>o0o + bflf^fa
= 0.
(3.17)
From (3.12), </>0 € /?„ and so f°x(pO(f>o 6 /?" as shown earlier in this section, also
irjf°x4>o4>o = 0 since \frjx = 0 for every x e R" (Lemma 3.2), (3.17) becomes
b\//Q fxx(po<f)i = 0, which means b = 0 since irj f°x4>o<l>i # 0, and so / = a</>0Here f = a</>0 implies t e Rna since 0 O € ^ J . but t € R" since F ( x , A.) maps
/?; x R into /?; x /?, which shows t e R? D Rna - {0} and so a = 0. Thus (3.15)
has only the trivial solution. The other side of the "if and only i f statement
follows directly.
A result similar to the above is stated without proof in Werner and Spence
[9], where they considered a different extended system than the one presented
here.
4. Numerical examples
4.1. Coupled cell reaction. Consider the nonlinear operator
Downloaded from https://www.cambridge.org/core. IP address: 3.235.21.12, on 27 May 2020 at 15:54:44, subject to the Cambridge Core terms of
use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0334270000007293
110
B. S. Attili
[8]
where x = (x\, x2) € ./? x R, which represent a two-cell exothermic reaction
(see Werner and Spence [9]), where
h(xi,x2, A.) = xi +e(x, -x2)-Xexp[xi/(l
+Sx,)}.
Here e is some coupling coefficient and the exponential term is the Arrhenius reaction rate term. Solving the extended system (3.4) with e = .01
and £ = .23 we obtained (Xi,x2,k) = (7.7773552,7.7773261,-47827289)
= (2.430591,2.4306056,-51124817), which are two simple
and (xux2,k)
symmetry-breaking bifurcation points. With e = 0, the simple symmetrybreaking bifurcation coalesces with the simple turning point. This means that
assumption (Al) will split the double turning point referred to in Theorem 3.2
into a simple turning point and a simple symmetry-breaking bifurcation point.
4.2. The Duffing equation. The system (3.4) in the previous section will be
a basis for the computation of symmetry breaking; that is, pitchfork bifurcation points. Such points will be isolated solutions of (3.4). For more on the
experimental aspects, consider the Duffing equation which exhibits secondary
bifurcation, see Seydel [8]; that is,
Jc + 2;c3=A.cosf,
x(0)=x(2n).
(4.1)
We are seeking 2n-periodic solutions. The assumption (Al) of Section 2 is
satisfied directly for (4.1) because of the symmetry properties of the cosine for
S € (Sa, Sfi, Sy), where
(Sax)(t) = x(-t),
(Spx)(t) = -x{n-t),
(4.2)
(Syx)(t) = -x(t - 7t).
Equation (4.1) has a branch of 2n-periodic functions which are symmetric
with respect to all symmetries. On this branch there are several symmetrybreaking bifurcation points at which two symmetries are broken and one symmetry is preserved. See Werner and Spence [9].
We shall discretise (4.1) using the shooting method to obtain a finite-dimensional problem. We start by dividing the interval [0, 2n] into four equal subintervals. Let or (0) = S\, x(0) = s2, x(n) = s3 and JC(TT) = *4 be initial guesses at
0 and n. With s\ and s2 as initial guesses, solve the two initial-value problems on
[0, n/2] and backward on [37r/2, 2n] to obtain xx, x\ at n/2 and x4, x4 at 37r/2
respectively. Also with s^ and s4 as initial guesses, solve the two initial-value
Downloaded from https://www.cambridge.org/core. IP address: 3.235.21.12, on 27 May 2020 at 15:54:44, subject to the Cambridge Core terms of
use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0334270000007293
[9]
Numerical computation of symmetry-breaking bifurcation points
111
problems backward on [n/2, n] and forward on [n, 3n/2] to obtain x2, x2 at
n/2 and JC3, i 3 at 3n/2 respectively. Now
" xx -x2
F(s, k) =
X\
-x2
(4.3)
X3 -Xi,
-x4
_
For the finite-dimensional problem, (Sax)(t) = —x(t) will be equivalent to
saying s2 = 0 and s4 = 0, and so an analogous symmetry to Sa is
5,=
1 0
0 0
0 - 1 0 0
0 0
1 0
0 0
0 - 1
(4.4)
The second symmetry (Spx)(t) = —x(n — t) will be analogous to
0
0 - 1 0
0 0 0 1
- 1 0
0 0
0 1 0
0
(4.5)
Also, (SYx)(t) = —x(t — n) is analogous to
0
0 - 1 0
0
0
0 - 1
- 1 0
0
0
0 - 1 0 0
(4.6)
To show that assumption (Al) holds for the finite-dimensional problem, we
shall deal with F(s, k), where
F(s, k) = QF(s, k)
and
Q=
l l l
- l
1 1
l l i
i - l
l l l
- 1 1
(4.7)
Downloaded from https://www.cambridge.org/core. IP address: 3.235.21.12, on 27 May 2020 at 15:54:44, subject to the Cambridge Core terms of
use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0334270000007293
112
B. S. Attili
[10]
Note that solving F(s, k) — 0 is equivalent to solving F(s,k) = 0 , since Q
is invertible. The function F is given by
F(s, k) =
X\ — Xi + *3 — X4 — X\ + X2 + XT, — X4
X\ — X2 + X3 — X4 + X\ — X2 — JC3 + X4
—X\ + X2 + X-} — Xj + X\ — X2 + X3 — X4
Xl — JC2 — *3 + X4 + X\ — X2
(4.8)
Now, under the 5i symmetry, we have
si
(4.9)
S\s = Si
\ s4 J
\ —s4
Solving the initial-value problems with S\, —s2 and s3, —s4, we obtain
x4, —x4 on [0, n/2], JC3, — x3 on [n, n/2], x2, —x2 on [n, 3n/2] and xu —xt
on [In, 3n/2]. Using these values we find F(Sys, k) is the same as F(s, k)
with the second component and the fourth component reversed in sign, which is
SiF(s, k), and so
Similarly, one can prove the same result for the other symmetries, which shows
that assumption (Al) is satisfied.
For the computations, we can find the symmetric solution by solving
jci +2x\ = kcost,
(4.10)
on [0, n/2]. In order to find the antisymmetric solution, we solve
jc, +2x\ = kcost,
5i,
(4.11)
on [0, n/2]. There will be no need to solve the other initial-value problems on
the other subintervals. The solutions may be extended based on the symmetries.
We transform (4.10) and (4.11) into a first-order differential equation to obtain
X\
=
X2 ,
x2
=
— 2x\ + kcost,
(4.12)
together with the initial conditions Xi(0) = Si and ^2(0) = 0 in the symmetric
case and ^ ( 0 ) = 0 and x2(0) = st in the antisymmetric case. In the symmetric
case, we need to satisfy x, (n/2) = 0 and so
Downloaded from https://www.cambridge.org/core. IP address: 3.235.21.12, on 27 May 2020 at 15:54:44, subject to the Cambridge Core terms of
use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0334270000007293
[11]
Numerical computation of symmetry-breaking bifurcation points
To find g we used R = 1 and 7 T = 1. With this choice, g = BF/dsx.
function dF/dsi can be obtained by solving the initial-value problem
u, = v2,
v2 = — 6u]vi,
V](0) = 0,
113
The
v2(0) = 1
and 3 F / 3 A. is obtained by solving
u>i = w2,
w2 = cost — 6u\w\,
tUi(0) = 0,
u;2(0) = 0.
Using Newton's method, (£) = 0 was solved with ^ = 2.88 and A, = 11
as initial guesses. We find s{ = 2.97410 and A. = 12.0776. The value of
g = —.111768 - 10.
The same system was solved using other initial guesses, s\ — 1.2 and X =
1.35. We find 5, = 1.33150 and A. = 1.57408. Note that this problem was
solved by Werner and Spence [9] using finite differences, and similar results
were obtained.
Acknowledgement
The author would like to thank the referees for their valuable comments.
References
[1] B. Attili, "Multiple shooting and the calculation of some types of singularities in B.V.P's",
Int. J. Comp. Math. 32 (1990) 97-111.
[2] B. Attili, "A direct method for the characterization and computation of bifurcation points
with corank 2", Computing 48 (1992) 149-159.
[3] F. Brezzi, J. Rappaz and P. Raviart, "Finite dimensional approximation of nonlinear problems, Part III, Simple bifurcation points", Numer. Math. 38 (1981) 1-30.
[4] A. Griewank and G. Reddien, 'The approximation of generalized turning points by projection methods with super convergence to the critical parameter", Num. Math. 14 (1975)
354-366.
[5] A. Griewank and G. Reddien, "Characterization and computation of generalized turning
points", SIAMJ. Numer. Anal. 21 (1984) 186-196.
[6] G. Moore, "The numerical treatment of nontrivial bifurcation points", Num. Fund. Anal.
Optim. 2(1980)441^72.
[7] D. Sattinger, "Group theoretic methods in bifurcation theory", in Lecture Notes in Math.
762,(1979).
[8] R. Seydel, "Branch switching in bifurcation problems for ordinary differential equation",
Num. Math. 41 (1983) 93-116.
[9] B. Werner and A. Spence, "The computation of symmetry breaking bifurcation points",
SIAMJ. Num. Anal. 21 (1984) 388-399.
[10] H. Werner, "On the numerical approximation of secondary bifurcation problems", in Lecture
Notes in Math. 878, (1981), 407^425.
Downloaded from https://www.cambridge.org/core. IP address: 3.235.21.12, on 27 May 2020 at 15:54:44, subject to the Cambridge Core terms of
use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0334270000007293