Академический Документы
Профессиональный Документы
Культура Документы
9 Springer-Verlag 1993
V. JURDJEVlC
C o m m u n i c a t e d by P. HOLMES
Introduction
R ( t ) is the rotation matrix in SOs(JR), and Ul and u2 are the control func-
tions. The problem is to minimize 89~0z (Ul2 + u~) dt over all solution curves
g ( t ) = (x~(t), xz(t), R ( t ) ) of (1) which satisfy the prescribed boundary con-
ditions g(0) = (x ~ x2~ R0) and g ( T ) = (xl, xz~, R~) in G. For convenience
we have chosen the components of velocity of the center of the ball as the
control functions, rather than the components of the velocity of the moving
plate, and we have also assumed that the radius of the ball is 1. Then the
velocity of the moving plate is twice the velocity of the center of the ball.
Finally, we have normalized the problem by choosing c = ~.
Aside from a brief discussion of the plate-ball problem, BROC~ZETT& DAI
[2] do not attempt to find its solutions. Rather, they consider the problem of
. . . .
mlmmlzmg 89~0T (u 2 + v2) dt over the solution curves of the following poly-
306 V. JVRDJEWC
1. Kinematic equations
We begin by deriving the kinematic equations of motion for the ball. Let
q = (ql, q2, q3) denote the coordinates of a point in fR3 relative to a fixed
right-handed orthonormal frame el, e2, e3 centered at a point O in the sta-
tionary plane, with e3 perpendicular to this plane and pointing upward.
al, a2, and a3 denote a right-handed orthonormal frame fixed to the ball at
its center. This setting is shown in Fig. 1.
,t obO4
i/
Fig. 1. Kinematic configuration
The frame al, a2, a3 is called the moving frame. The coordinates of any
point of N3 relative to this moving frame are denoted by Q = (Q1, Q2, Q3),
R denotes the rotation matrix which transforms the coordinates relative to the
moving frame onto the coordinates relative to the stationary frame. That is,
~ x S 0 3 ( ~ ) at each point a ( t ) .
We identify the tangent space of SO 3 (JR) at any point R in S 0 3 ( ~ ) as the
set of all matrices R ~ with 0 an antisymmetric matrix. Then,
dR [ o - o3(t) o2(t)
- - = R ( t ) ~ co3(t) 0 -co,(t)]
dt k,-co2(t) col(t) o/
for some functions oil(t), ~2(t) and ~o3(t). The vector (-0 = ~1el + co2e2 +
r 3 is the angular velocity generated by R ( t ) .
The kinematic equations which follow will be based on the assumption that
there is no slipping between the ball and the parallel planes which are in con-
tact with the ball. That means that --dq = 0 when A is the point of contact
dt
of the ball with the fixed plane, and that the velocity dq of the point of
dt
contact with the moving plane is equal to the velocity of the moving plane.
If we denote by uel + re2 the velocity of the upper plane, then the non-slip-
ping assumptions imply that
- - (.0 2 (-01
dq(t) _ dqo(t)
with a~(t) = (a~l, (-02, o)3). Therefore, + r ( t ) xco(t). The non-
dt dt
slipping assumptions mean that
' dt'
0) + (0, 0, -1) xco,
dx dy & dy
Consequently, dt -0)2' dt o01 , u dt o)2, and v dt + o ) t . It then
follows that u = - 2 0 ) 2 and v = 20)1. If we now set us = 89u, u2 = 89v, we
get that
--
dt u~,
--
dt
u2, -- = R
dt
(0 o) 3
IA1
0
U2
o:)
2 9
the vector ~ = (al, a2, a3) in ~3. This choice of identification is motivated
310 V. JURDJSVIC
ad B(A) = [B,A] = A x / ~ = A B ,
(00 )
It is convenient to work with the following basis for 2 ( G ) : el and e2 are
the standard unit vectors in N2, and
A1 = 0 0 1 , A2 = 0 0 0 , A3 = 1 0 .
0 1 -1 0 0 0 0
Then A1, A2, and As are the standard unit vectors in ~3. The following com-
mutator relations are easy to verify:
(Recall that the Lie bracket of left-invariant vector fields corresponds to the
negative of the matrix commutator rule.)
It is convenient for future reference to assemble the structure of 2 ( G ) as
in Table t. With this notation, equation (1) may be written as
de
(2) --o = Ul(t) Xl(g) + uzXz(g)
dt
The Geometry of the Plate-Ball Problem 311
Table 1
[,] A1 A2 A3 E1 E2
A1 0 -A 3 A2 0 0
A2 As 0 -A~ 0 0
A3 -A 2 A1 0 0 0
E1 0 0 0 0 0
~2 0 0 0 0 0
where g(t) = (x(t), y(t), R ( t ) ) , and X1 and X2 are respectively the left-in-
variant vector fields E1 - A 2 and E z + A1. It can be easily verified that the
Lie algebra generated by E 1 - L 2 and Ez +L1 is equal to 2 ( G ) , and
therefore it follows that for any go and g~ in G, and T > 0 there exist control
functions ul and u2 defined on [0, T] such that the corresponding trajectory
g(t) which emanates from go at t = 0, passes through gl at time T. That is,
the system is controllable, and therefore our optimal problem is well-posed.
Proposition 1. Suppose that F and H are any smooth functions on the cotangent
bundle T*M of a smooth manifold M. Denote by {exp tH : t ~ R) the one-parameter
group of diffeomorphisms generated by the Hamiltonian vector field H of H. Then,
d
(i) {F,, H) (x) = -- F oexp tH(x)lt= 0 for each x in T*M.
dt
(ii) The Lie bracket [F, H] of two Hamiltonian vector fields F and H is a Hamil-
tonian vector field, and [F,,H] = E where E = {F, H].
(iii) If Hx and Hy are the Hamiltonians which correspond to smooth vector fields'
X and Y on M, then {Hx, Hy] = H[H,y].
tively, and by H1, H2, and H s the Hamiltonians associated with A~, Az and
As. Their Poisson table is isomorphic with Table 1 and is given by Table 2:
Table 2
{, } //1 H2 H3 h1 ha
~, 0 -H3 h'2 0 0
& H3 o -H~ 0 0
H3 -& & 0 0 0
h~ 0 0 0 0 0
h2 0 0 0 0 0
(3) H 2 = _ 89~. (/A2 .{_/,/2) _~_ //1 (hi - H2) -I- //2 (h2 + H1), 2 = 0, 1.
for almost all t in the interval [0, T]. In this notation u(t) = ( u l ( t ) , u2(t)),
and v = (vl, v2).
Suppose that ~(t) is an integral curve of H 1. Since H 1 is a concave func-
tion of //i, and u2, and since there are no bounds on the size of controls,
its maximum occurs for //1 = h i - / 4 2 and //2 = h2 + HI. Hence, the max-
imality condition (4) means that such an optimal trajectory is the projection
of an integral curve of a single Hamiltonian vector field H associated with
H = 89(hi - H2) 2 + 89(h2 + HI) 2. We refer to the integral curves of H as the
regular extrema associated with our optimal problem.
/_/0 is a linear function of ul and //2. Along an integral curve ~ of H ~
generated by the controls ( u l ( t ) , /,/2(/')) the maximality condition ( 4 ) m e a n s
that (h 1 -/-/2) (~(t)) = 0 and (h 2 --k H1) ( { ( t ) ) = 0. Thus, if g(t) is the pro-
jection of an integral curve g of H ~ then the maximality condition (4)
results in further constraints through which the optimal control may yet be
determined. We refer to the integral curves of/_/0 which satisfy the maximali-
ty condition (4) as the abnormal extremals. It is known that an optimal trajec-
tory may be a projection of an abnormal extremal, but not a projection of
a regular extremal (Morrax~om~.Y [7].) However, in our situation, as we shall
The Geometry of the Plate-Ball Problem 313
d
If Ul = u2 = 0, then dt H3 (~(t)) = 0, and therefore H3 (~(t)) = constant. But if
We shall presently see that the regular extremals include the abnormal ex-
tremals with u s = constant, u2 = constant, and H ( ( ( t ) ) = constant.
The regular extremals are the integral curves of the Hamiltonian vector
field H which corresponds to the system Hamiltonian H = 89(h 1 - H 2 ) 2 +
!2 (h 2 + H 1 ) 2 . Along any such curve ~(t) the controls are of the feedback
type us(t) = (h 1 - / / 2 ) (~(e)), u2(t) = (h2 + HI) ( ~ ( t ) ) . We have the follow-
ing proposition.
Corollary 1. There are four integrals of motion, along each regular extremal. They
are h~ = constant, ha = constant, H = 89(hi -- H2) 2 + 89 (h2 + Ha) 2, and M =
Again the proof is evident and is omitted. The solutions singled out in
Corollary 2 justify the claim made earlier that the abnormal extremals which
correspond to optimal solutions are also regular: The solution in (i) corres-
pond to ul = u2 = 0. The solution in (ii) corresponds to Ul = h i - / / 2 =
- ( 2 + 1//2, and ~ = h 2 + H l = ( 2 + 1) H1. Hence u l = f H 2 and u z = f H 1
with f = - ( 2 + 1).
We are now ready to draw an analogy with the equations for the rigid-body
motion. The system Hamiltonian H is analogous to the total energy of the
body, while M is analogous to its angular momentum. However, in this case
H is a cylinder, rather than the energy ellipsoid for a rigid body. Therefore,
the extremals are the intersections of the energy cylinder with the momentum sphere
(Fig. 2).
o -H3(t) Hz(t)
P(t)= H3(t) 0 -H l(t)),
\-H2(t) Hi(t)
0 0 -- (h I -- H2) )
t2(t) = o o -(he +H~) ,
(hi -/-/2) (h2 + Hi) 0
In the notations introduced in Section 2, the extremal equations for
d
H 1, H: and /-/3 can be rephrased as - - ( H I , / / 2 , / / 3 ) = (H1, H2,//3) x
dt
316 V. JURDJEVIC
d~
(h2 + H1, - (hi - H2), 0). Hence, - - = / S x t ? , and using formula (i) o f
dt
Section 2, we get that
(6) dP [s P]
dt
5. Integration
Each extremal solution (x(t), y(t), R(t), hl, h2, P(t)), with P(t) as
defined by equation (6), is contained in a four-dimensional integral sub-
manifold of T*G defined by the following equations:
Then
Hence,
dO(t)
- H3 ( t ) .
dt
The Geometry of the Plate-Ball Problem 317
Since - h I sin(0 + c~) + h 2 COS(0 -t- 0/) = - N/h -2 -t- h 2 sin 0, it follows that
d20
(7) -- + A sin 0 = 0, with A = ~/2H(h 2 + h~).
dt
Equation (7) is the equation for the mathematical pendulum. Its total
energy E is determined by our integrals of motion M = H~ + H22 + H~, the
system Hamiltonian H, and the constant h~ + h 2 . We have
Therefore
1 (M-2H-h21-h22) = -Acos0+ 1 (dO'~ 2
7 2 ~dt] "
Thus, E = 89( M - 2 H - (h~ + h22)) is the total energy of the pendulum, and
dO
(8) --= 4- X/2 (E + A cos 0) .
dt
Recall that ul -- ha - H2 -- x / ~ cos(0 + o~) and that u2 = h2 + H1 =
x / ~ sin(0 + od.
The extremal equations simplify considerably if we transform all the coor-
dinates according to the following transformation:
(2, y) = (x, y) ( cos c~ - s i n c~],
\ sin o~ cos a /
(cosc~ -sino~
(9) /~ = RRo with Ro = siO
n ~ cos
0
COS o~
i)
- s i n o?~
P = RolPRo, (hlh2) = (hi, h2) sin / o
COS ~ ]
318 V. JURDJEVIC
(00 o:)
It is convenient to introduce the following notations:
= 0 0 2 , ua=X/~cosO, u2=x/~sinO.
~2 ~2
In terms of these notations,
dR dR
- Ro = R~Ro = RRo(R~-2~Ro) = R ~ ,
dt dt
dff dP
(0
dtt = R~ dt R~ = RO-2[g2' P] R~ = [~' P]"
where the matrices A2, A2, and A 3 have the same meaning as before (Ai = ei,
i = 1, 2, 3). This representation agrees with the usual representation in terms
of Euler angles.
Since dfi/dt = [s P], it follows by the same argument used previously, that
R(t) f i ( t ) / ~ - l ( t ) = J l, for some constant matrix A. We assume that A is of
the form A = X/MA 3 . This condition can always be arranged through a
suitable rotation from the left. (That is, the original system is left-invariant,
and therefore left multiplications permute solutions.) This particular choice of
The Geometry of the Plate-Ball Problem 319
_/[ adapts naturally to the choice of coordinates in (11) for the following
reasons:
/:3(t) = %/rMt~ -1 (t) A3/~ (t) = x/M e - (h3A3 e - q52Aze - $1A3A3e 4'1A3e*2A2e$3A3
(13)
( •
- s i n 42 0 cos 42
0 •
1 )
e 4)2A2 = 0 1 0 .
+ d~b3 e61A3e~aA2e~3A313.
dt dt dt dt
R( t) (( e-A34)3e-A24~2e-A34hA3eA301eAz(~2eA3(~3)T +( e-A303A2eA303)~-
dO2+A3 9 )
d R --m
Since - - = RO, it follows that
dt
~ m 1 --
Using the symmetry relation /~/5/~-1 _~. ~/M A3 we get that ~ P = R - A3R.
Upon substituting into the equation above, we get that x/M
d03
x/M dt
d()l + (e-A3Cb3A2eA363)
ddOt2
+ A3 dt = u2A1 - ulA2"
Then,
e A30
/cos03OS030
sin03 i)
and therefore e-A3~3e2= sin 03el + cos 03e2. Upon substituting this matrix
into the preceding equation we get
d03
1 dt (/qlel+/~2e2+/-t3e3) + d02
d01 dt (sin 03el + COS 03e2) + dtt e3
= u2el -- ~le2 .
Proceeding further we get
1 dq52 1 d01/~2 d02
d01/q 1 + sin 03 = ~2, + - - cos 03 = - ~ ,
dt dt ~ dt dt
1 drY1/q3 d03 = 0.
1 ~ '
~--/~2 COS03 62 --Ul
The Geometry of the Plate-Ball Problem 321
A = - s i n q 5 z = -y- -2(E+Acos0).
H2 --tql
Since sin 03 - x/M sin 4z , and cos~b 3 - x / M sinqSz , it follows that
-/71~ 2 +/72~ 1
g2 cos ~3 + ~1 sin ~3 = . Then,
sin ~2
d~b3
m
tq3 dO1 (AcosO-2H)x/2(E+AcosO)
dt x/M dt M-2(E+Acos 0)
Equations (8), (10), and (14) contain the necessary integrability information
required for the solutions, and justify the claim made on page 16.
Let us concentrate first on the dynamics of the center of the ball. The ap-
propriate equations are (10), along with the equation for 0. When
reparametrized by 0, these equations become
d2 ~ 4 ~ cos 0 dy X / ~ sin 0
(15) - - = , - - =
As we remarked earlier these equations are the same as those for the planar
elastica (see, for instance, JURDJEVIC [4]). The equations of the planar elastica
go back to EuL~R who made their initial study in 1744, and who also is the
originator of their name. For convenience to the reader we quickly summarize
the basic details concerning the elastic problems of Euler.
Let y(t) = (x(t), y(t)) be any curve in the plane, parametrized by its arc
length relative to some point on the curve. The geodesic curvature k along the
curve is given by k 2 = d2x d2Y
dt-~ +~t 2 . The elastica problem consists in finding
a curve, called the elastica, which minimizes 89~0r k2ds among all curves y
which satisfy the prescribed two-point boundary conditions. These boundary
conditions consist of an initial point in the plane, an initial tangent vector,
a terminal point, and a terminal tangent vector. If 0 is the angle that the unit
tangent makes with the horizontal line, then dx/dt = Cos 0, and dy/dt = sin 0.
Thus, k(t) = •
It turns out that along an elastica, 0 satisfies the equation for a mathe-
matical pendulum. The mathematical pendulum associated with angle 0 is
what KIRCHHOFF called the kinetic analogue of the elastic problem. Love [6]
distinguishes between two geometric types which he calls inflectional and non-
inflectional. The non-inflectional type occurs when E > A . Then, k(t)=
x/2(E + A cos0) is always positive (or, the pendulum has sufficient energy
to clear the top, and therefore O(t) is a monotone increasing function of
time.)
The inflectional type occurs when E < A. Then, there is a cut-off angle
00. The function O(t) is a periodic function of t. The function k(t) alternates
in sign as 0 oscillates back and forth between - 0 0 and 00. The critical case,
when the pendulum has just enough energy to reach the top, occurs when
E=A.
In all cases, the equation for y is easily integrated to yield y(O)=
1 a/2E + 2/1 cos 0 + constant, while the integral x(O) = I cos 0 dO
A t~X/f(E + A cos 0)
is in general an elliptic integral (of the third kind). The elliptic integral reduces
to a trigonometric integral in the critical case when E = A.
The sketch below (Fig. 3) describes these various geometric possibilities (the
reproductions of EULER'S original sketches can be found in the survey paper
of TRUESDELL [I0]),
Since EUL~R, the elastic problem has drawn a considerable amount of
mathematical attention continuing up to the present (BRYANT& GRIFFITHS [3],
LAN~ZR & SINCER [5], JURagJ~VlC [4], and MUMFORD [8]). In the present situa-
tion, /-/3 plays the role of geodesic curvature with E = 8 9 ( M - 2 H - ( h 2 + h ~ ) )
corresponding to the total energy of the Hamiltonian function associated with
its kinetic analogue.
This curious connection between the elastica problem, and the plate-ball
problem begs for further mathematical investigations, which unfortunately I
am not able to provide now. Instead, I just state the result explicitly for future
references in the form of a proposition.
The Geometry of the Plate-Ball Problem 323
J
I X
(o)
(b)
(c)
Fig. 3
Proposition 4. Each optimal path of the center of the ball which corresponds to
the energy level H = 1 also minimizes 89 ~ k2 ds with k equal to the geodesic
curvature of the path.
Let us now consider the rotational kinematics of the problem, with a par-
ticular focus on the geometric significance of the kinetic analogue and its
geometric type. The simplest motions occur when h 2 = h 2 + h 2 = 0.
Then A = 0, and therefore dO/dt = • (cf. (8)). Thus, 0 is constant
when E = 0. In such a case, it follows from (10) that the center of the ball
moves along a straight line. The controls (u~, u2) which produce the straight
line motion are constant, and therefore the ball rotates with constant angular
velocity 09 = u z e t - ule2. As expected, this angular velocity is perpendicular
to the line traced by the center. The magnitude of co is equal t o [[co I1 =
+ = 2H.
Suppose now that E > 0, and maintain the assumption that h 2 + h 2 = 0.
Then 0 ( t ) = • 00. But then, using equations ( 1 5 ) w e get that
y(O) = -T- cos 0 +Y0, and the center of the ball moves along a circle
/_.___..
These rotations are described by two angular velocities cot = x/M e 3 and
o)2 = - x / ~ e 3 . The entire motion, including the circular motion of the cen-
ter of the ball may be viewed as a motion on T 3 = S 2 • 2. This motion
is closed if and only if [[coi IJ/llco2]l is rational. Recall that when h 1 = h 2 =0,
then M = H 2 + H22 + H3, 2
H = g1 ( H 21 + H2) 2
, and E = 89(M - 2 H ) . Thus,
2E = H~, and M = 2H + H~. Hence,
flcolrl M z-/
+1,
l/co2lp 2 - 2 E - E
and therefore the motion is closed if and only if the radius traced by the center
of the ball is a rational number. The sketch of these solutions is contained
in Fig. 4.
Consider now a general solution. We distinguish between the following
geometric types: non-inflectional if E > A, critical if E = A, and inflectional
if E < A . Recall that E = ~ ( M - 2 H - ( h I 2+h22)), and A = x / 2 H ( h 2+h~).
These types can be easily seen in terms of Fig. 2, which represents the intersec-
tions o f the energy cylinder H = 89((/-/2 - hi) 2 + (H1 + h2) 2) with the mo-
mentum sphere M = H 2 + H22 + H32.
The effect of the transformation employed earlier which transformed
the variables (H~,/-/2, H3) into H~, H2 and 1-73 is to rotate (hi, h2) into
(~/~12 + h 2 , 0), and thus to place the center of the energy cylinder on the
/72-axis at the point ~ - 2 + h ~ . Since 2 ( E - A ) = M - 2 H - ( h ~ + h 2 2 ) -
2x/2H(h 2+hE) = M - ( ' V ~ + x[h 2 + h~)2, the geometric types are
determined according to the sign of M - ( x / ~ + x/(h~ + h22)) 2. But,
( x / ~ + x/h~ + h22 )2 is the value of the extreme point on the/72-axis which
is on the energy cylinder. We have therefore obtained the following description
for the geometric types:
Let $1 and $2 be the circles on the momentum sphere and the energy
cylinder respectively, contained in the plane /73 = 0. Then, the non-inflec-
tional type occurs precisely when $2 is contained in $1. The critical type oc-
curs when these circles are tangent to each other along the/72-axis, and the
inflectional type occurs when these circles intersect transversally. The straight-
line movement occurs in the critical case ($1 = $2), while the circular move-
ment is of non-inflectional type because $2 is contained inside $1 (see Fig. 2).
We now assume that H and M are fixed, and w e study the kinematic
behaviour as the center of the energy cylinder changes. For simplicity we assume
that H = 89 Then A = x/2H(h~ + h e) = ~ + h22. So we study the depen-
dence of the solution on the parameter A. In particular we are interested in
the influence on the kinematics of the ball of the bifurcation occurring at the
critical case M = (1 + A ) 2. Our observations will be made in terms of the
Euler angles ~b1 , q~2, and ~b3 defined by (1!).
Recall that equations (14) are derived under the assumption that A =
that
(1-A)2=M-2(E+A) <=M-2(E+AcosO)<=M-2(E-A) = (I+A) 2
In order to ensure that all solutions remain in this coordinate chart we assume
that 0 < A < 1. The critical case M = (1 + A ) 2 also remains in this coor-
dinate chart for M < 4.
We begin with the discussion of the non-inflectional type. Since E > A,
there is no value of 0 such that x / 2 ( E + A cos 0) is equal to zero. We
assume that dO~dr = x/2 (E + A cos 0) > 0. Since H3 = dO/dt, this assumption
implies that we are choosing the solution branch in the upper hemisphere of the
momentum sphere (Fig. 2a).
We recall equations (14):
It is clear that dq~l/dt > 0, and d~3/dt < 0 for all t. Hence, both 4q(t) and
-(h3 (t) are monotone increasing with time. Moreover, it follows from (12) that
1
/-/3 = ~ cos 4~2. Thus cos q52 > 0, and therefore cos Oz = ~ x/2 (E + A cos 0).
Let us now consider the inflectional type (Fig. 2c). Since E < A, the angle
0 is confined to the interval - 0 e _-< 0 _< 0e with 0~ = cos-1E/A. As the pen-
dulum swings from -Oc to 0c, H3(t) is positive because H3(t) = dO/dt, and
//~ / "\\ /
H..!%
(Q)
e2
i
H1 ;-;" el
Ib)
(c)
Fig. 4. h = 0 .
The Geometry of the Plate-Ball Problem 327
is equal to zero at the extreme points 4-0c. Since dH3/dt = A sin 0, it follows
that /-/3 reaches its maximum at 0 = 0. //3 is negative while the pendulum
swings from 0c back to -0~, reaches its minimum at 0 = 0, and then in-
creases to zero at 0 = - 0 ~ . The same pattern repeats with each swing of the
pendulum.
It is easy to see that d4~3/dt = 0 at the extreme values 0 = 4-0c. Thus,
oh3(t) oscillates between its extreme values cpa• and 4'3rain. These values are com-
~ e2
(a)
f
el
"/ ~ \
(b)
e2
e~
,. I . / - J*~ "S-~ n
(c)
Fig. 5. h # 0 .
328 V. JURDJEVIC
We then have
01(Oc) - 0 1 ( - 0 c ) = ~ (1 - cos 4) dO
( A S + 1 - 2 A cos 0 ) ~ / 2 ( M - A 2 - 1 + A cos 4)
-0c
is a rational multiple of 20c.
These last two propositions reflect the qualitative changes in the rotational
kinematics of the ball near the critical case, and clarify the remarks made in
the introduction: As the parameter A increases through the critical value
M = (1 + A) 2, the ball changes its movement from rolling in the e3 direction
to wobbling in the e3 direction. Figures 4 and 5 contain a pictorial s u m m a r y
of our results.
References
Department of Mathematics
University of Toronto
Toronto, M5S IA1