Professional Documents
Culture Documents
Clifford Algebras, Dynamical Systems, and Periodic Orbits: IMA Journal of Mathematical Control Information
Clifford Algebras, Dynamical Systems, and Periodic Orbits: IMA Journal of Mathematical Control Information
Certain differential systems can be lifted to algebras (of matrices), which greatly
1. Introduction
The Carleman linearization technique has been developed in a number of papers [1, 2,
3]. For a nonlinear one-dimensional equation of the form
.X =f(x), x(O) = xo, x(t) E ~' (1.1)
the method is very simple. We define the new variables
c/J;=Xi.
Then
}=0
for some values ail depending on the derivatives off. Then we can write the equation
as an infinite-dimensional linear one:
cP = A<P (1.2)
where <P = (¢o,¢J, .. y and A= (a;1)0 ,;;; i,J<oo· The solution is
<P(t) = eA 1<P0 (0 ~ t ~ T),
where 4> 0 = (l,x0 ,x5,x6, .. y,
provided the Taylor series of the solution of the
original system (1.1) converges in [0, T].
The main drawback with the method is that, for vector systems, we must consider
Taylor monomials of the form
and proceed as before. However, we then obtain a tensorial equation of the form
cP = A<P,
279
©Oxford University Press 1996
280 D. McCAFFREY, S. P. BANKS AND A. MOSER
where 1> = (¢;1 ... ;J(i1 ;;;. o) is a rank-n tensor and A is a tensor operator. This approach
has been used in Ref. 4. One can, of course, string 1> out into a long vector as in Ref. 6,
but this destroys the essential structure of A.
In this paper we shall consider systems which can be 'lifted' to associative algebras,
so that we have
x =f(x), x(O) = x0 , xEA, (1.3)
where x 0 E A for some associative algebra A. Assuming that/ has a convergent Tay-
lor series in this algebra, then we can use the expansion functions xi which are well
defined in the algebra. This will lead to a matrix linearization of (1.3), since these
functions can be treated a essentially 'scalar' objects.
The most useful algebras for our purposes are the standard matrix algebras and the
2. Clifford algebras
In this section we shall give a brief introduction to Clifford algebras and their
properties which are needed for global linearization of systems. For more details
see Ref. 7. A quadratic space (V, Q) is a vector space over a field F( = JR., C, or lHI)
together with a nondegenerate quadratic form Q, i.e. a mapping Q : V--+ :F, such
that
(i) Q(-\v) = ,\2 Q(v) (,\ E F, v E V),
and such that the map B(. , .), defined by
(ii) B(v, w)~ HQ(v) + Q(w)- Q(v- w)] (v, wE V),
is bilinear. The associated form B(v, w) can be used to define an inner product on V.
Also, if {e1 } is a basis for V, then
Note that (lRn,o, Qn,o) is just JR.n with the quadratic form -llxll 2 while (JR 0·n, Qo.n) is JR.n
with the quadratic form llxll 2 (i.e. usual Euclidean space). 0
ExAMPLE 2.2 (C, Qn) is a complex quadratic space with
Qn(z) = ZT + ... + z~.
The associated bilinear form is Bn(z, w) = 2:7=I ziwi (rather than the usual form
2::7= 1 Z;W;). 0
1 = ao = [~ ~]. a1 = [~ ~1 ], az = [ 0i -i]
0 ,
Then
aT=1 (O~i~3), E~ = 1, Er = -1, ( 1 ~ i ~ 3)
282 D. McCAFFREY, S. P. BANKS AND A. MOSER
Since JR 0•0 has dimension 0, its universal Clifford algebra has dimension 2° = 1, and so
Ao,o = {..\1 :.A E JR} = JR,
together with the embedding v: JR0•0 __, Ao,o given by v(O) = 0. Since a~ = 1, the
embedding v: JR 1•0 __,A1,o given by v(y) = ycr3 gives rise to the realization
A~,o={[; :J:x,yElR}=lREBlR
with basis e0 = 1 = cr0 and e1 = cr3 • Similarly, A 0,1 has the realization
Similarly,
A 0 ,2 = { [ xo + ix1
. Xz + ix3]
. : x 0 , ... , x 3 E lR } = lHI.
-X2 + lXJ Xo- 1X 1
D
Higher-order Clifford algebras can be generated from the above special cases, since it
is easy to see that, if A(V, Q) denotes the universal Clifford algebra for (V, Q), then
is a universal Clifford algebra of dimension 4 x dim (A(V, Q)). For example, consider
the quadratic space
0
v: (v,x)f---+ [ v-x V +X]
0
.
In fact, it can be shown that, if A(V, Q) is the universal Clifford algebra for a non-
degenerate real quadratic space (V, Q), then it is isomorphic to a real subalgebra of one
of the matrix algebras C 2m x 2m or C 2m x 2m EB C 2m x 2m if dim V is even or odd, respectively.
All these operators are extended by linearity to the whole of the algebra. Using these
operators we can define the norm function
L1: A---+A
by
L1(x) =XX.
Care should be taken, however, with this definition, since it does not generally resemble
the usual idea of norm in a vector space: it is more like the determinant of a matrix,
and indeed replaces the det function in the general theory of spin groups. This can be
seen clearly in the cases Ao, 1 and A 1,0 • In the former case we have the expansion of any
element z = xe0 + ye 2 E Ao, 1 in terms of the basis e0, e2. Then
L1(z) = (xeo- ye2)(xeo + ye2) = (x 2 + i)t,
and this is indeed isomorphic to the standard Euclidean norm. In the latter case,
however, we have z = xa0 + ya3 for any element in A 1,0, and then
X= L
aE2{1, .. , •l
eaXa, (2.3)
284 D. McCAFFREY, S. P. BANKS AND A. MOSER
then
2
X a. (2.4)
aE2{1, .. ,n}
Moreover, this norm is independent of the basis. We shall use the norm given by
(2.3)-(2.4) even in the non-Euclidean case, but it should be noted that (2.2) no longer
holds, and the norm is not independent of the basis {e"}, although all such norms are
equivalent.
where
and {e 13 )JE:l U {ea,h,;;,;n is a complete basis of Ap,q such that {e 13 )JE:l and
{ea,h,;;,;n are disjoint. We then consider the 'extended' differential equation
LEMMA 3.1 If the equation (3.2) is a lifting of equation (3.1) to Ap,q then any solution
X(t) of (3.2) with ~i = 0, for j E ..J, gives rise to a solution x(t) of (3.1):
Proof We have
d d
(X, ea ) = -d (X, ea ) = -d X;
0
' t ' t
and
[~I ~2
x2 XI
l [ x 21 +x22 2x 1x 2 ]
x,r
2x 1x 2 xi +x~
[XI
x2 XI
on Ao,2> where
x3 + ~x4].
X1- IX2
Now let
F(X) = L
00
ix;,
i=O l.
(3.5)
where a; E F (i ~ 0). The solution of (3.4) is given directly by the Lie series
Xo = [~ ~l
The solution is given by (3.6), i.e.
X(t) = [exp(tx
3
d~)x]X=Xo
=L ;
00 i
(2i- 1)!!Xii+l
i=O l.
we have
00 00 00
= '""'. ak xk+i-1
~~ k' = '""'. ak "'
~~ k' 'f'k+i-1 = '""' "'
~CX.jf'f'f
k=O . k=O . f=O
where
. ae-i+1
if£+ 1 ~ i,
a;e= { ~(£-i+1)!
otherwise.
Hence, writing ci> = (¢0 , ¢1 , ¢2 , •. y, we have
<i> = Aci>, (4.2)
where A= (a;j) 0 ~ i,J<oo· Note that ci> E A~q~ EBk'=oAp,q• and the latter can be made
into a Banach space by defining
Then (as has been shown in Ref 3) eAr exists as a bounded operator for t E [0, r) on
(A~J, II·IIA;), and the solution of (4.2) is given by
ci>(t) = eA
1
ci>(O),
where
r is the maximal time of existence of the solution, and A~J is the subset of A~q
consisting of all elements of the form ( 1, X' X 2 ' X 3 , ... ) T.
REMARK 4.1 If F(O) = 0, i.e. a 0 = 0, we can consider the vector ci> = (¢ 1 , ¢ 2 , ... )T,
i.e. we only need to take into account Taylor monomials for i > 0. 0
288 D. McCAFFREY, S. P. BANKS AND A. MOSER
= iXi+2
A= [~ ~
0 0 0 0
0 2
3
.....'1
This is easily checked to lead to the same solution as in (3.7). 0
THEOREM 4.1 If F(O) = 0, the system (4.1) is stable if and only if, given any positive
definite (infinite) matrix Q, there exists a positive definite matrix P such that
ATP+PA=-Q.
where by positive definite we mean that
cpT PeP> 0
for all cP E Ap,q of the form cP =(X, X 2 ,X,
oo· 3
... )T . 0
- - -2 -3 T - .
(Here, cP = (X, X , X , ... ) , where X denotes the conjugate of X.) A Lyapunov
function is then given directly by
CLIFFORD ALGEBRAS, DYNAMICAL SYSTEMS, AND PERIODIC ORBITS 289
Let
l l
0 0 0
P~ r~ 0 0
Q~ r: 0 0
Then,
l
0 1
A~-[: 0
0
0
0
2
0 3
and so
Also,
V = <PT PC/J =XX= L1(X) = llxll 2 1
and
D
Remark 'Positive definite' now means something rather different than in the usual
sense.
5. Periodic orbits
The proof of the existence of periodic orbits of nonlinear dynamical systems has a
long history, including justifications of the harmonic linearization techniques [8],
Lyapunov theory [5], and the use of index theory [9]. The method developed here
will allow us to prove directly the existence of periodic orbits for systems which
can be lifted to a Clifford algebra. Thus, consider again the system (4.1) on Ap,q
and its linearization (4.2) on A;.'q· Suppose that X(O) lies on a periodic orbit of period
T. Then
where C/J(O) = (X(O), X 2 (0), X 3 (0), .. y. Hence, eAr has an eigenvalue 1 with
290 D. McCAFFREY, S. P. BANKS AND A. MOSER
THEOREM 5.1 A necessary condition for (4.1) to have a periodic orbit of period T is
that the associated operator A E .C(A~q) has an eigenvalue of 2mjT. 0
A sufficient condition is that A satisfies the 'necessary' condition of Theorem 5.1 and
that eAT has an eigenvector corresponding to the eigenvalue 2m/T of the form <P(O)
above. Moreover, each element of the vector eAt<P(O) is a formal power series which
We must solve these equations with initial condition (0,0, ... , 1,0, ... ) (1 in the
(2p + 1)th position). Thus
Y2p+l =0 ===} Y2p+l = 1,
j; 2p-l = (2p- l)e2it · 1
CLIFFORD ALGEBRAS, DYNAMICAL SYSTEMS, AND PERIODIC ORBITS 291
so
Y2p- 1 = 1 (e2it - 1) .
(2p - 1) 2i
Next,
Y2p-3 = (2p- 3)e
21
i Y2p-l
and so
Thus,
Hence, if
ll(e2it-
2
l)X?I ~!
0 "' 2'
then the solution is given by the above power series. Since A has i as an eigenvalue, all
solutions with
2 1 I
IXol ::::;
SUPo ,;; 21r le
2it
- 1
=2
1 ,;; J
lie on periodic orbits of period 21r. In particular, if A = Ao.1 ~ C, then the equation
292 D. McCAFFREY, S. P. BANKS AND A. MOSER
1.6
1.2
0.8
--{).4
-1.2
-1.6
-2.0 L__ _.::::::::~====::::::::::=.._L~_!__ _ _ __::::::::=-=====:::::::__ _j
becomes
Le.
[ ~I]Pi
[-1 01] [~:]
0
aH
ap,
CLIFFORD ALGEBRAS, DYNAMICAL SYSTEMS, AND PERIODIC ORBITS 293
and
. . .. 8H .8H
z=q 1 +rp 1 = - - I -
8p1 oq,
= _ 2i(~ 8H _ !._ 8H) (6.2)
2 aq, 2i ap,
-2i 8H
=
oz.
Similarly, for n = 2, we can 'lift' the equation (6.1) to the quaternions IHI in a similar
manner. Thus, define
where ij = k. Then,
. . 8H
ZJ = - 21 -
0ZJ '
whence
. . 8H .. 8H
w=- 2r - + 2J I -
8z1 8z 2
= -2i (Bz 1 + j8z2 )H(w),
where
Hence
w= -2iDH(w) (6.3)
where
Next we seek to lift (6.1) for a general n to A 2m the universal Clifford algebra over
(~ 2n, 11·11 2 ). We need the following characterization of An+ 2 in terms of An-
LEMMA 6.1 An+ 2 can be realized as M(2, An) (the 2 x 2 matrix algebra with ele-
ments in An·
Proof Let e 1 , ... , e" be an orthonormal basis of ~n and define the elements
1'i E M(2, An) (i ~ j ~ n + 2) by
i 0]
l'n+i = [0 -i '
294 D. McCAFFREY, S. P. BANKS AND A. MOSER
Then
'Y] = [ eo] 2
eJ
0]=[-10 -1OJ (1 ~ j ~ n), 2
'Yn+1 =
[ -1
O
'Y~+2 = [ ~1 ~1].
Moreover, the Clifford condition
'Y}'Yk + 'Yk'Yj = -2c5f
also holds. A simple dimension argument finishes the proof. D
Returning to the problem oflifting a general Hamiltonian system of the form (6.1) to
Note that v: JR:.2 ---+ A 2 is given by v(x 1, x 2 ) = (x 1e1 + x 2 e2 ) 2 = -(xi+ x~)I, so that
e 1 and e2 generate the algebra A 2 . Now,
z = q[ 0i 0] + p[-10 01]
-i
an [ i o J an [ o
= ap o -i - aq -1 ~]
o -i J [ o 1 J an [o o. Jan
= [ -i o -1 o ap + -i -l
- = -J2 DH,
aq
i]
where
0
12 = [ i 0
and
D= [oi -io Jaqa + [-1o
Note that
In order to lift the system from An_ 2 to An we introduce the following definition
DEFINITION 6.3 Let n be even; say n = 2k, where k ~ 2. Define
2
z<n- ) = q1e1 + P1e2 + q2e3 + P2e4 + · · · + qk_Jen-3 + Pk-len-2•
and suppose that
.z(n-2) = -Jn-2DH
D
a
= e1-+e2-+ ··· +en-3--+en-2--
a a a
aql aP1 aqk-1 aPk-1
Is a Hamiltonian system on An_ 2. We define the double of this system on An to be the
where
w<n-
2) = qn1 +P!''/2 + ··· +qk-i'Yn-3 +Pk-i'Yn-2•
- a a a a
D = '1'1-a +'1'2-a + · · · +'Yn-3-a--+'l'n-2-a--,
'Pk-1
q1 'Pi . qk-1
')';= [O
e;
e;l0 (i=l, ... ,n-2),
and
Jn-2 0 ]
Jn = [ .
0 Jn-2
0
Remark 6.4 Clearly, the double of a system on An_ 2 is of the form
2
0 .z(n- )] _ [ 0 -Jn-2DH]
[ .z(n-2) 0 - -Jn-2DH 0 .
0
EXAMPLE 6.5 We now lift the system in example 6.2 to A 4 . First, the double is the
system
296 D. McCAFFREY, S. P. BANKS AND A. MOSER
or
. . J (EJH EJH )
q111 + P112 = - 4 aq 11 + ap 12
1 1
To extend the system to~ consider the system
0 0
0 0
0 0
42 +h
-i 0 -1 0
0 0
0 -i 0 -1
x.~[~ 0 ~o~]
0
. EJH
P2 = --. ·
8q2
Moreover, the matrices /~o 1 2, 1 3 , 1 4 generate A 4 , by Lemma 6.1. Since the systems
(6.4) and (6.5) are independent, we can combine them into a single equation
w= -J4DH(w)
where
and
a a a a
D=11-a +12-a +K4r3-a +K4r4-a .
ql 'PI q2 'P2
The general technique is now clear and we state it as follows.
THEOREM 6.6 Suppose that the (n- 2)-dimensional Hamiltonian system can be
written in the form
CLIFFORD ALGEBRAS, DYNAMICAL SYSTEMS, AND PERIODIC ORBITS 297
wn -- [ 0
wn-2
wn
0
-2] +[~ ~Jqn +[~~ ~ ]Pn,
and
Dn [ 0
Dn-2
Dn-2] +K2n [i
0 0 ~J +K2·[~ ~],
0
0
0
7. Conclusions
In this paper we have studied nonlinear dynamical systems which can be lifted to
Clifford algebras. Such systems have a considerable degree of symmetry and exhibit
interesting behaviour. In particular, we have shown that Lyapunov's equation for
linear system stability directly generalizes to Clifford systems and that periodic orbits
may often be computed explicitly for these types of system. We have presented only
some very simple examples to illustrate the theory. Clearly, many other kinds of
systems are expressible in this form, and a thorough study of some higher-
dimensional examples should provide some very interesting behaviour. For
example, are any of these systems chaotic? If so, the global linearization technique
will provide a method for obtaining explicit characterization of their properties, as
in the case of periodic orbits. These questions will be examined in a future paper.
Acknowledgement
This work was partially supported by the EPSRC under grant numbers GR/H 73066
and GR/J 75241.
REFERENCES
1. BANKS, S. P., 1985. On the generation of infinite-dimensional bilinear systems and Volterra
series. Int. J. Systems Sci. 16, 145-60.
2. BANKS, S. P., 1987. Global bilinearization of nonlinear systems and the existence of Volterra
series. Int. J. Contro/46, 1331-42.
298 D. McCAFFREY, S. P. BANKS AND A. MOSER
3. BANKS, S. P., 1987. Global Volterra series for nonlinear meromorphic systems on complex
manifolds. IMA J. Math. Control & lnf. 4, 131-41.
4. BANKS, S. P., & CHANANE, B., 1988. A generalized frequency response for nonlinea.r systems.
IMA J. Math. Control & Inf 5, 147-66.
5. BANKS, S. P., 1986. Control systems engineering. Prentice-Hall, London.
6. BROCKETT, R., 1976. Volterra series and geometric control theory. Automatica 12, 167-76.
7. GREUB, W., 1978. Multilinear algebra (2nd edn). Springer, Berlin.
8. ScHMIDT, G., & ToNDL, A., 1986. Nonlinear vibrations. Cambridge University Press,
London.
9. MAWHIN, J., & WILLEM, M., 1985. Critical point theory and Hamiltonian systems. Springer,
Berlin.