2

GYRODYNAMICS
Introduction to the dynamics of rigid bodies

Introduction. Though Newton wrote on many topics—and may well have given
thought to the odd behavior of tops—I am not aware that he committed any
of that thought to writing. But by  Euler was active in the field, and it
has continued to bedevil the thought of mathematical physicists. “Extended
rigid bodies” are classical abstractions—alien both to relativity and to quantum
mechanics—which are revealed to our dynamical imaginations not so much by
commonplace Nature as by, in Maxwell’s phrase, the “toys of Youth.” That such
toys behave “strangely” is evident to the most casual observer, but the detailed
theory of their behavior has become notorious for being elusive, surprising
and difficult at every turn. Its formulation has required and inspired work
of wonderful genius: it has taught us much of great worth, and clearly has
much to teach us still.
Early in my own education as a physicist I discovered that I could not
understand—or, when I could understand, remained unpersuaded by—the
“elementary explanations” of the behavior of tops & gyros which are abundant
in the literature. So I fell into the habit of avoiding the field, waiting for
the day when I could give to it the time and attention it obviously required
and deserved. I became aware that my experience was far from atypical:
according to Goldstein it was in fact a similar experience that motivated Klein
& Sommerfeld to write their 4-volume classic, Theorie des Kreisels (–).
In November  I had occasion to consult my Classical Mechanics II
students concerning what topic we should take up to finish out the term. It
was with feelings of mixed excitement and dread that I heard their suggestion
that we turn out attention to the theory of tops. The following material takes
as its point of departure the class notes that were written on that occasion.
2 Gyrodynamics

1. Kinematic preliminaries. Let xi be the inertial Cartesian coordinates of an

arbitrary—and, for the moment, not necessarily rigid —assemblage of point
masses mi . Writing
xi = X + r i (1)
the total kinetic energy of the system becomes

T = 12 mi ẋ2i
i
 
= 1
2 M Ẋ
2
X·
+ Ẋ miṙri + 1
2 mi ṙi2 (2)
i i

Noting that the variables {XX , r 1 , r 2 , . . . , rN } cannot, on numerological grounds,

be independent, but must be subject to a single vectorial constraint, it is “to
kill the cross term” in (2) that we require

mi r i = 0 (3)
i

It follows then from (1) that

 
mixi = MX X with M ≡ mi (4)
i i
X ≡ coordinates of the center of mass

Time-differentiation of (4) gives


P = pi
i

where pi = miẋ xi . P = M Ẋ
X is the total linear momentum of the system and is
for isolated systems conserved, whatever may be the nature of the intra-system
interactions. In this notation (2) has become

T =T0 + T (6)

T= 1
2 miṙr2i = kinetic energy relative to the center of mass
i

T0 = 12 M Ẋ 2 = 1
2M P
2
= kinetic energy of the center of mass

To impose rigidity upon the N -particle system amounts formally to

imposing the stipulation that the numbers

aij ≡ |x
xi − xj | = |rri − r j |

shall be time-independent constants:

ȧij = 0 (7)

Equivalently, we might stipulate that the numbers r i· r j be constant. It is,

however, intuitively obvious that the conditions (7) are highly redundant:
Kinematic preliminaries 3

number N of particles number n of degrees of freedom

1 3
2 5
3 6
4 6
5 6
.. ..
. .
We on this basis expect a rigid blob of material to have six degrees of freedom.
The point is sharpened by Chasle’s theorem,1 according to which the constituent
parts mi of a rigid assemblage have instantaneous positions that can be
described
xi (t) = X (t) + R(t)rr0i (8)
where X (t) locates the moving center of mass, where R(t) is a time-dependent
(proper) rotation matrix

R(0) = I and R(t) RT (t) = I : all t
0 (9)

and where the constant vectors r 0i record the initial and enduring design of
the rigid assemblage: r 0i = xi (0) − X (0). Three degrees of freedom enter
into the specification of X , and (in 3 -dimensional space) three more into the
specification of R.
Chasle’s equation (8) amounts to the assertion that in a rigid assemblage

r i = R r 0i (10)

The basic decomposition (1) assumes therefore the sharpened form

xi = X + R r0i

which after t-differentiation becomes

X + Ṙ r 0i
xi = Ẋ
ẋ
X + Ṙ R–1 r i
= Ẋ

But it is a familiar implication of (9) that A ≡ Ṙ R–1 is invariably/necessarily

antisymmetric. Writing
 
0 −ω3 ω2
A =  ω3 0 −ω1  (11)
−ω2 ω1 0

we obtain finally
ẋ X + Arri = Ẋ
xi = Ẋ ω × ri)
X + (ω (12)

1
See E. T. Whittaker, Analytical Dynamics (4th edition ), page 4.
4 Gyrodynamics

where A and (equivalently) ω will, in general, be t-dependent.

Returning with this information to (6) we find that for rigid bodies the
kinetic energy relative to the center of mass—the “intrinsic” kinetic energy—
can be described
 
T = 12 mir Ti AT Arri = 12 ω × r i )·· (ω
mi (ω ω × ri)
i i

= 1
2 mi (rri × ω )·· (rri × ω )
i
 
 0 −r3 r2
= 1
2 miω T BTi Bi ω with Bi ≡  r3 0 −r1 
i −r2 r1 0
= 1
2 ωTI ω (13)

where

I≡ mi BTi Bi
i
 
 r 2 − r1 r 1 − r 1 r2 − r 1 r3
= mi  − r 2 r1 r 2 − r2 r2 − r 2 r3  (14)
i − r 3 r1 − r 3 r2 r − r3 r3 i
2

serves to define the “moment of inertia matrix”—also called the “moment of

inertia tensor” or simply the “inertia matrix.” Here ri2 ≡ r i · r i , and in the
continuous limit we expect to have
 2 
 r − r1 r 1 − r 1 r2 − r 1 r3
=  − r2 r1 r 2 − r2 r 2 − r2 r3  ρ(rr) dr1 dr2 dr3
− r3 r1 − r 3 r2 r 2 − r3 r3

REMARK: In the little argument that led to (13)—whence to the

invention of the moment of inertia matrix I—we made essential
use of a property of the “cross product:” r × ω = − ω × r . But
the cross product is a peculiarly 3-dimensional construct, and so
also, therefore, are the results reported above. In a one-dimensional
world smoothly graded rotation is impossible, and the theory of
rigid bodies trivializes. It is established in introductory physics
courses that for rigid bodies that are constrained to move in what
is, in effect, two dimensions one has

T = 12 M Ẋ 2 + 12 Iω 2 with I ≡ r2 ρ(rr)dr1 dr2

It is in the cases of spatial dimension N
4 that things become

mathematically interesting, but in those cases we have no physical
interest.
Kinematic preliminaries 5

Look finally to the total angular momentum of a rigid body. For any
multi-particle system one has
 
J= xi × pi = mi xi × ẋ
xi
i i

which by (1) becomes


= X + ri ) × (Ẋ
mi (X X + ṙri )
i

The expression on the right develops into a sum of four terms, of which two—
the “cross terms”—vanish in consequence of the fundamental constraint (3).
We are left with
= Lorbital + Lintrinsic (15)
with

 angular momentum of the
Lorbital = X ×M Ẋ
X = X ×P P = center of mass, relative to

the coordinate origin
 
angular momentum relative to
Lintrinsic = mi r i × ṙri =
the center of mass
i

The intrinsic angular momentum of rigid multi -particle systems is called spin:
one has

Lintrinsic −→ S = mi r i × Arri
i

= mi r i × (ω
ω × ri)
i

=− mi r i × (rri × ω )
i

= mi BTi Bi ω
i
= Iω (16)

and agrees, in place of (15), to write

J =L+S (17)

Several concluding remarks are now in order:

• Equations (6) and (15) share a structure

total = orbital + intrinsic

which they owe to the disappearance of certain cross terms. That

disappearance is a consequence of the way  the center of mass was
defined: a consequence, that is to say, of (3): mi r i = 0
6 Gyrodynamics

• One does not write P total = P orbital +P

P intrinsic because, by (3), P intrinsic
vanishes identically.
• Familiar analogs of the conditions that yield the conservation law Ṗ P =0
J = 0, which may but need not arise from conservation
also yield J̇
separately of L and S .
• We write p = mvv , and from the fact that m is a scalar conclude that
the momentum p of a particle is always parallel to (in fact: a fixed
multiple of) its velocity v . But from S = I ω and the fact that I is a
3 × 3 matrix we see that the spin angular momentum S and angular
velocity ω need not—and typically will not—be parallel.
• Returning with ω = I –1S to (13), we find

T = 12 S T I –1 S provided det I
= 0 (18.1)

which is of interest as the formal analog of T = 12 p T m–1 p. Equivalently

= 12 ω · S (18.2)

2. Nature & properties of the moment of inertia matrix. Let the distribution
function ρ(x) describe how some material of interest (mass, let us say) is
distributed along the real line. One writes

M = m = x0 ρ(x) dx
(0)
: 0th moment

M X = m(1) = x1 ρ(x) dx : 1st moment

m = x2 ρ(x) dx
(2)
: 2nd moment
..
.

to define the “moments” of the distribution.2 From the set of all moments one
can construct the “moment generating function”
∞
 
ϕ(k) ≡ 1 n (n)
n! (ik) m = eikx ρ(x) dx
n=0

and by Fourier transformation recover the distribution itself:


ρ(x) = 2π e−ikx ϕ(k) dk
1

If we translate the coordinate origin to the center of mass—which is to say: if

we introduce new coordinates r = x − X and proceed as before we obtain the
2
In conventional terminology and notation, M is the “total mass” and X
defines the position of the “center of mass.”
Nature & properties of the moment of inertia matrix 7

so -called “centered moments”3


(n)
µ = rn ρ(r) dr

Evidently

µ(0) = M

the center of mass relative to the
µ(1) = m(1) − Xm(0) = 0 :
center of mass resides at the origin
All of which carries over straightforwardly to higher-dimensional situations. In
three dimensions we have

M= x) d3 x
ρ(x : solitary 0th moment
 
 x1
MXX=  x2  ρ(x x) d3 x : vector of 1st moments
x3
 
 r1 r 1 r 1 r 2 r 1 r 3
M=  r2 r1 r2 r2 r2 r3  ρ(rr) d3 r : matrix of centered 2nd moments
r3 r 1 r 3 r2 r3 r 3
and are placed now in position to recognize that the moment of inertia matrix
is an object assembled from centered second moment data:4
I = (trace M)·U − M (19)
Remarkably, the low- order moment data built into the designs of M , X and I
is central to the dynamical theory of rigid bodies, but the moments of higher
order are (in most contexts) utterly irrelevant: distinct rigid bodies can be
expected to move identically if they have identical 0th , 1st and 2nd order
moments.
The moment of inertia matrix I is manifestly real and symmetric : IT  = I.
We
are assured,
 therefore, that the eigenvalues of I (call them I1 , I 2 , I3  or
A, B, C
 ) are real , and the associated eigenvectors (call them e 1 2 3 or
, e , e
a, b, c ) are—or can always be taken to be—orthogonal : ei· ej = δij . We are
assured, moreover, that I can in every case be diagonalized by a suitably chosen
rotation matrix :
   
I1 0 0 1 0 0
R T I R =  0 I2 0  with R T R =  0 1 0 
0 0 I3 0 0 1

3
I have here allowed myself to write ρ(r) where I should more properly have
written something like ρ̃(r) ≡ ρ(X + r). Similarly, I will later write ρ(rr) when
x) that was given and ρ(X
actually it was ρ(x X + r ) that is intended.
4
Since the symbol I is busy, I have here had to use U to represent the 3 × 3
identity matrix.
8 Gyrodynamics

The orthonormal triple e1 , e2 , e3 serves to define the “principal axes” of the
rigid body. With respect to the “principal coordinate frame” that has its origin
at the center of mass and coordinate axes parallel to the principal axes one has
    
I1 0 0 1 1
0 I2 0   0  = I1  0 
0 0 I3 0 0
    
I1 0 0 0 0
0 I2 0   1  = I2  1 
0 0 I3 0 0
    
I1 0 0 0 0
0 I2 0   0  = I3  0 
0 0 I3 1 1

If r1 , r2 , r3 refer to the principal frame, then the continuous version of (14)
supplies
   2 
I1 0 0  r2 + r32 0 0
0 I2 0  =  0 r12 + r32 0  ρ(rr)d 3 r (20)
0 0 I3 0 0 r1 + r22
2

from which it follows trivially that

All the eigenvalues of I are positive. (21.1)

More interestingly,
 
 2   2 
I1 + I2 = r1 + r2 + 2r3 ρ(rr)d r

2 2 3
r1 + r22 ρ(rr)d 3 r = I3
..
.
etc.

—the implication being that

No eigenvalue can exceed the sum of the other two, (21.2)

which is to say:

The eigenvalues of I satisfy the triangle inequality: (21.3)

I3 I2

I1
Moment of inertia with respect to an axis 9

This occurance of the triangle inequality is, in view of (20), not at all surprising,
for if α, β and γ are any positive numbers then

i=α+β
j =β+γ
k =γ+α

invariably and automatically satisfy the  inequality.5

We have been brought to the conclusion that to every rigid blob can be
associated
• a naturally preferred point (the center of mass);
• a naturally preferred “principal axis frame,” with origin at that point
(defined by the eigenvectors of the moment of inertia matrix I);

• non-negative numbers A, B, C associated with the respective legs of
the principal axis frame. Those numbers (eigenvalues of I) can in all
cases be identified with the sides of a triangle, or alternatively: with
the semi -axes of an ellipsoid
r12 r22 r32
+ + =1
A2 B2 C2

Figure 1: Two-dimensional representation of a rigid body with

preferred point and eigenvalue-weighted principal axes. Note that
those attributes attach instantaneously even to non-rigid blobs, but
it is only in the presence of rigidity that they acquire importance.

3. Moment of inertia with respect to an axis. In introductory physics one learns

to write 
I= r2 dm

to describe the moment of inertia of a rigid body with respect to a prescribed

axis: here r is understood to denote the normal distance from the mass element

5
I am indebted to Tom Wieting for this observation.
10 Gyrodynamics

n
λn x

Figure 2: In the text we compute the shortest distance from a point

to a line. The figure explains the notation used in that argument.

dm to the axial line. The question now before us: how does that scalar moment
of inertia relate to the matrix-valued construction I ?
We confront first a simple geometrical problem: how to describe the length
of the normal dropped from a point to a line? To describe a line through the
n (λ variable, n a fixed unit vector). The condition that the
origin we write λn
vector p(λ) ≡ x − λn n · p(λ) = 0) enforces λ = n · x. The
n be normal to the line (n
length of the normal dropped from x to the line can therefore be described

r2 (x x − (n
x) = [x n · x)nn]·· [x n · x)n
x − (n n]
= x · x − (x
x · n)(nn · x)
= x · (U − N ) x (22.1)
= n · [(trace X)·U − X ] n (22.2)

where, as before, U is the identity matrix 4 , where

 
n 1 n1 n 1 n2 n 1 n3
N ≡  n 2 n1 n 2 n2 n 2 n3  projects onto n
n 3 n1 n 3 n2 n 3 n3

and where X is constructed similarly:

 
x1 x1 x1 x2 x1 x3
X ≡  x2 x1 x2 x2 x2 x3 
x3 x1 x3 x2 x3 x3
Moment of inertia with respect to an axis 11

It now follows that


 
 moment of inertial about an axis through
2 3
I0 = x)ρ(x
r (x x) d x = the center of mass, in the direction defined


by the unit vector n
can, by (19), be written
I0 = n · I n (23)
which could hardly be simpler or prettier.
What can one say about the moment of inertia about an axis that does not
pass through the center of mass? To describe points on such an axis we write
n and assume without loss of generality that a ⊥ n. Proceeding as before,
a + λn
we introduce p(λ) ≡ x − (a n), insist that p(λ) ⊥ (a
a + λn a + λn
n) and are led to
the conclusion that the line dropped normally from x to the displaced axis has
squared length
r2 (x x − a)··(U − N )(x
x) = (x x − a)
which, we note, does give back (22.1) at a = 0. More particularly, we have
= x · (U − N )x
x − 2x
x · (U − N )a
a + a · (U − N )a
a

which we introduce into


I= r2 (x x) d 3 x
x)ρ(x

and by quick easy steps recover the “parallel axis theorem:”

I = I0 + M a2 (24)

Equation (23) indicates how moments with respect to axes (numbers of the
type I0 ) can be extracted from the data written into I. One can also proceed
in the reverse direction. Suppose, for example, we were to set
 
1
n = 0
0
We would then have
measured value of I0 = I11
More generally, we have, for each selected n,
n) = n1 n1 I11 + 2n1 n2 I12 + 2n1 n3 I13
measured value of I0 (n
+ n2 n2 I22 + 2n2 n3 I23
+ n2 n2 I22
Given six well-chosen instances of that equation, we would have enough
information to compute all the Iij by straightforward linear algebra.6
6
Is there an computationally optimal way to select n1 , n2 , n3 , n4 , n5 , n6 ?
12 Gyrodynamics

4. How the moment of inertia matrix responds to rotations. Shown below is

a representation of a rigid body that has rotated about its fixed center of
mass. The component parts of the body have retained their relative positions,
but their positions relative to fixed exterior reference frames have, in general,

Figure 3: Schematic representation of a rigid body has been rotated

about its immobile center of mass.

changed. The issue before us: How is I final related to I initial ? Working from
(14), i.e., from (compare (19))
 
   r1 r 1 r 1 r2 r1 r3
I= mi r2 · U − W i with W ≡  r2 r1 r 2 r2 r2 r3 
i r3 r 1 r 3 r2 r3 r3

The effect of body rotation—as was remarked already at (10)— can be described

r 0i −→ r i = R r 0i : all i

which induces

0
Wαβ ≡ rα0 rβ0 −→ Wαβ ≡ rα rβ = Rαµ rµ0 rν0 Rβν = Rαµ Wµν
0
Rβν

giving W = R W0 RT = R W0 R –1 . Immediately r 2 = traceW = traceW0 = (r0 )2

so we have
I 0 −→ I = R I 0 R –1 (25)
Response to rotations 13

The importance of this result emerges when one supposes R to be a function

of time, writing
I(t) = R(t) I(0)R –1 (t)
Differentiation of R RT = U leads quickly to the important conclusion (remarked
already on page 3) that

Ṙ = A R with AT = −A

so we have
İ = A I − I A (26)

Look in this light back to the equation

S = Iω

that at (16) was seen to relate spin to angular velocity. Time-differentiation

gives
S = I ω̇
Ṡ ω + (A I − I A)ω
ω
But A was seen at (11) to be just another name for the operation ω ×, so we
have

S = I ω̇
Ṡ ω + ω× Iω − Iω× ω
= I ω̇
ω + ω ×S
S (27)

5. Equations of rigid body motion: Newtonian approach. Look now to the

Newtonian dynamics of the N -body system contemplated at the beginning of
§1, again suspending for the moment any assumption concerning the rigidity of
the system. Immediately
 interactive
F i ≡ F iimpressed + F ij xi
= miẍ
j

X + mir̈ri
= miẌ (28)

where evidently F iiinteractive = 0 while by the 3rd law F ij

interactive
= −F interactive
F ji .
Summation on i supplies
 impressed
F total impressed ≡ Fi = M ẌX
i

P
= Ṗ (29)

where the 3rd law has killed the interactive force terms, and the constraint (3)
has served to kill the relative acceleration terms. Application of
 
xi × = X + r i )×
(X (30)
i i
14 Gyrodynamics

to the left side of (28) gives

 
X × F total impressed + r i × F iimpressed + interactive
r i × F ij
i ij

where the final term actually vanishes if—as we will assume—the interactive
interactive
forces are central: F ij  (rri − r j ). On the other hand, application of
(30) to the right side of (28) was already seen at (15) to give Lorbital + Lintrinsic .
So we have, for any centrally interactive N -body system,

N orbital + N intrinsic = L̇
Lorbital + L̇
Lintrinsic

where
N orbital ≡ X × F total impressed

N intrinsic ≡ r i × F iimpressed
i

But it follows already from (29) that N orbital = L̇

Lorbital so, collecting the results
now in hand, we have

F total impressed = Ṗ
P
: refer to motion of the center of mass
Lorbital
N orbital = L̇

N intrinsic = L̇
Lintrinsic : refers to motion relative to the center of mass

Now impose the assumption of rigidity upon our N -body system, and
emphasize that we have done so by notational adjustment: Lintrinsic → S = I ω .
Drawing upon (27) we then have

S
N intrinsic = Ṡ
= I ω̇
ω + ω ×Iω (31)

Here N intrinsic , ω and the integrals that assign instantaneous meaning to I all
refer to the space frame, a (generally non-inertial) translated copy of the inertial
lab frame. Several circumstances limit the utility of this result:
• The value of N intrinsic will, in the general case, depend upon both the
location and the orientation of the rigid body, and may even depend upon
its instantaneous state of motion (as would happen if the body had been
assembled from charged particles and were placed in a magnetic field).
This circumstance introduces an element of circularity into the discussion:
one must know the net effect of all past motion to understand what (31)
has to say about present motion.
• Even in the simplest case N intrinsic = 0 an awkward time-dependence lurks
in the design of I, which changes moment to moment as the body rotates.

The latter difficulty can be circumvented by a strategy introduced by Euler

(), the essential idea being to kill the time -dependence of I by passing to
Equations of rigid body motion: Newtonian approach 15

a frame that is fixed in the body. If I 0 ≡ I(0) refers to the initial orientation
of the body, and I ≡ I(t) to its evolved orientation, then by (25) we have

I = R I 0 R –1 (32)

where R ≡ R (t) refers to the rotational that has been accomplished during the
interval in question. In this notation (31)—after multiplication on the left by
R –1 —becomes

N 0 = I 0 R –1 ω̇
ω + R –1ω × R I 0 R –1 ω with N 0 ≡ R –1N
= I 0 · R –1 ω̇
ω + R –1ω × R · I 0ω 0 with ω 0 ≡ R –1 ω

Our further progress hinges on the following

LEMMA: Notice that if A is a 3 × 3 antisymmetric matrix

   
0 −a3 a2 a1
A =  a3 0 −a1  and a =  a2 
−a2 a1 0 a3

then Axx = a × x establishes the sense in which “A = a×”. Let R

be a proper 3 × 3 rotation matrix: R –1 = R T and det R = 1. Then

R A R –1 = (R a)× (33)

We can on this basis write R –1ω × R = ω 0 ×. Differentiation of ω 0 = R –1 ω

ω 0 = R –1 ω̇
leads moreover to the conclusion that ω̇ ω + Ṙ Tω . But transposition of 7
Ṙ = A R gives Ṙ =−R A =−R (ω
T T T
ω ×) so we have ω̇ ω 0 =R –1 ω̇
ω − R Tω × ω =R –1 ω̇
ω,
so we have
N 0 = I 0 ω̇
ω 0 + ω 0 ×I0 ω 0 (34)

The preceding equation describes the motion of ω as viewed by a

non-inertial observer who is at rest with respect to the gyrating rigid body
—difficult to imagine when you contemplate a spinning/precessing top, yet an
entirely commonplace notion: you sit comfortably on the non-inertial earth,
look up into the night sky and see ω 0 as the vector about which the stars
appear to revolve. And if you wait long enough (thousands of years) will notice
that ω 0 traces a curve in the patterned fixed stars. What is remarkable is that
(34) is structurally identical to (31).8

7
See again page 3.
8
On pages 27–35 of gyrodynamics (/) and again on page 92 below
I discuss in detail how Coriolis and centrifugal forces—universal symptoms of
non-inertiality—conspire to achieve this remarkable result.
16 Gyrodynamics

Computation9 shows the vector ω 0×I 0 ω 0 to be, in the general case, a fairly
intricate object. It is, however, entirely natural to identify the body frame with
the principal axis frame
   
I11 I12 I13 I1 0 0
 • I22 I23  −−−−−−−−−−−−−−−−−−−−−−→  0 I2 0 
passage to principal axis frame
• • I33 0 0 I3

Major simplifications are then achieved: one is led from (34) by quick
calculation to the so-called Euler equations

N10 = I10 ω̇10 + (I30 − I20 ) ω30 ω20 

N20 = I20 ω̇20 + (I10 − I30 ) ω10 ω30 (35)


N30 = I30 ω̇30 + (I20 − I10 ) ω20 ω10
We have here a coupled system of three non-linear first-order differential
conditions on three unknown functions ω 0 (t).
Suppose for the moment that equations (35) have been solved . How does
knowledge of ω 0 (t) determine the rotation matrix R(t) by means of which we—
if not riding on the body but watching it from a position at rest with respect
to the space frame—propose to understand the perceived motion of the rigid
body? We have

Ṙ R –1 = A = ω ×
= (R ω 0 )×
ω 0 ×)R –1
= R (ω by lemma (33)
giving
Ṙ = R A0 with (we may assume) R (0) = U (36.1)
Equivalently
 t
R (t) = U + R (τ )A0 (τ ) dτ (36.2)
0
which can, in principle, be solved by iteration. But except in the simplest of
cases we can expect the solution of (35) to be very difficult, and the solution of
(36.1) to be also very difficult.

6. Equations of rigid body motion: Lagrangian approach. The idea here is to

look upon the elements Rij of R as “generalized coordinates,” to construct a
Lagrangian of the form L( Ṙ, R), and then to write
 
d ∂ − ∂ L( Ṙ, R) = 0
dt ∂ Ṙ ∂ R
We confront, however, the fundamental difficulty that the nine variables Rij
9
Which might be entrusted to Mathematica. Or see pages 24–25 in the
notes just cited.
Equations of rigid body motion: Lagrangian approach 17

are not independent, but are subject by R R T=U to six constraints. To deal with
that circumstance we will bring into play the “method of Lagrange multipliers,”
following the lead of Saletan & Cromer.10
Backing up to the equation that led to (13), we have

T= 1
2 miṙri· ṙri
i

= 1
2 mi (Ṙ r i0 )··(Ṙ r i0 )
i

1
= 2 Ṙµα Mαβ Ṙµα
µ,α,β

0
Mαβ ≡ 0 0
mi riα riβ
i

= 1
2 tr Ṙ M 0 Ṙ T

If the Rαβ were unconstrained independent variables we might at this point be

tempted to introduce a Lagrangian of the form L( Ṙ, R) = 12 tr Ṙ M 0 Ṙ T − U(R),
where the U(R) -term is intended to model such torques as may be impressed
upon the gyrating body. We have, however, a symmetric array of (in effect)
six constraints (R T R − U)αβ to honor. To that end we introduce a symmetric
array of Lagrange multipliers λβα , and with their aid construct

λβα (R T R − U)αβ = tr A
• (R R − U)
T

α,β

whence
L( Ṙ, R, Ȧ
•,A• ) =
1
2 tr Ṙ M 0 Ṙ T − U(R) + tr A
• (R R − U)
T

Drawing upon the symmetry of M 0 and A

• , one readily computes

 
d ∂ − ∂ L = (R̈ M 0 )αβ + ∂U/∂Rαβ − 2(R A
• )αβ = 0
dt ∂ Ṙαβ ∂Rαβ
 
d ∂ − ∂ L = (R T R − U)αβ = 0
dt ∂ λ̇αβ ∂λαβ

which in matrix notation (after multiplying the first equation by R –1 on the left)
become
R –1 R̈ M 0 + R –1 ∂U/∂R = 2A
•

R TR − U = O
Returning with R –1 = R T (information supplied by the second equation) to the
10
E. J. Saletan & A. H. Cromer, Theoretical Mechanics (), pages 144–146.
See also J. V. José & E. J. Saletan, Classical Dynamics: A Contemporary
Approach (1998), pages 514–519.
18 Gyrodynamics

first equation, and noting that A

• − A
•
T
= O, we obtain

(∂U/∂R)T R − R T (∂U/∂R) = R T R̈ M 0 − M 0 R̈ T R (37)

from which all reference to the Lagrange multipliers has been eliminated. This
second-order matrix equation is (by antisymmetry) equivalent to a system of
three scalar equations, and will now be shown to comprise no more nor less
than a matrix formulation of the Euler equations (35).
Let us first of all agree to write

(∂U/∂R)T R − R T (∂U/∂R) ≡ N 0

and to note that N 0 is antisymmetric. Turning our attention to the expression

on the right side of (37), we notice that differentiation of Ṙ = AR supplies
R̈ = (Ȧ + A2 )R whence
R T R̈ = R –1 (Ȧ + A2 )R
But it was the upshot of our lemma (33) that if A = ω × then

If A = ω × then R –1A R = A0 with A0 = ω 0 × and ω 0 = R –1 ω

Moreover, we by differentiation of A0 have

Ȧ0 = Ṙ TA R + R TȦ R + R TA Ṙ
= R T (−A2 + Ȧ + A2 )R
= R –1 Ȧ R

according to which the derivative of the transform is the transform of the

derivative of A. What we have established is that

R T R̈ = Ȧ0 + A0A0

from which information it follows that (37) can be written

N 0 = (Ȧ0 + A0A0 )M 0 − M 0 (−Ȧ0 + A0A0 )

= (Ȧ0 M 0 + M 0Ȧ0 ) + (A0 A0 M 0 − M 0A0 A0 ) (38)

But11 I = (trM)·U − M entails trI = 2 trM whence

M 0 = 12 (trI 0 )·U − I 0

giving finally
 
N 0 = (trI 0 )·Ȧ0 − (Ȧ0 I 0 + I 0Ȧ0 ) − (A0 A0 I 0 − I 0A0 A0 ) (39)

The claim now is that (39) stands to Euler’s equations (35) in precisely the
11
See again (19) on page 7.
Euler angles 19

relation that A stands to ω . The point is most easily established by direct

(Mathematica -assisted) calculation: set

   
0 −ω30 ω20 I10 0 0
A0 =  ω30 0 −ω10  and I =
0  0 I20 0
−ω20 ω10 0 0 0 I30

and discover that the expression on the right side of (39) can be written

 
0 −[I30 ω̇30 +(I20 −I10 ) ω20 ω10 ] [I20 ω̇20 +(I10 −I30 ) ω10 ω30 ]
 [I30 ω̇30 +(I20 −I10 ) ω20 ω10 ] 0 −[I10 ω̇10 +(I30 −I20 ) ω30 ω20 ] 
−[I20 ω̇20 +(I10 −I30 ) ω10 ω30 ] [I10 ω̇10 +(I30 −I20 ) ω30 ω20 ] 0

Taking the antisymmetric matrix on the left side of (39) to mean

 
0 −N30 N20
 N30 0 −N10 
−N20 N10 0

we see that the Lagrangian formalism has led us to what is in effect the dual
of Euler’s system of equations.

7. Euler angles. We could sidestep the constraint problem altogether if we could

produce a parameterized description of the elements of O(3), analogous to the
description which
 
cos θ sin θ
R (θ) ≡
− sin θ cos θ

provides of the elements of O(2). This was first accomplished by Euler,12 who
observed (see the following figure) that if one
• rotates through an appropriate angle φ about the 3-axis, then
• rotates through an appropriate angle θ about the new 1-axis, then
• rotates through an appropriate angle ψ about the newest 3-axis
one can bring any frame into coincidence with any other frame.13 Reading from
12
L. Euler (–) retained an interest in dynamics—particularly the
dynamics of rigid bodies—throughout his professional career, but his papers on
the subject were written mainly between  and , when he was attached
to the court of Frederick the Great, in Berlin. I suspect it was his physical
work which stimulated the invention of the Euler angles, though they are of
independent mathematical interest and importance.
13
We assume, of course, that the two frames share the same origin and are
similarly handed.
20 Gyrodynamics

2
φ
1

1
ψ

Figure 5: The sequence of axial rotations that give rise to Euler’s

description of
the elements
 of the group O(3) of rotations in 3-space.
The angles φ, θ, ψ are called “Euler angles” and the two planes
intersect in what is called the “line of nodes.”

the figure we have

   
cos ψ sin ψ 0 1 0 0 cos φ sin φ 0
r =  − sin ψ cos ψ 00 cos θ sin θ   − sin φ cos φ 0 r (40.1)
0 0 1 0 − sin θ cos θ 0 0 1

≡ E(φ, θ, ψ) r
Euler angles 21

serves to describe how the coordinates r relative to the rotated red frame of a
fixed point P are related to its coordinates r relative to the black frame:

x
x

Figure 6: The passive (or “alias”) interpretation that at (40.1)

has been used to introduce the Euler angles.

By Mathematica-assisted calculation we find

 
cos ψ cos φ−cos θ sin ψ sin φ cos ψ sin φ+cos θ sin ψ cos φ sin θ sin ψ
E(φ, θ, ψ) =  − sin ψ cos φ−cos θ cos ψ sin φ − sin ψ sin φ+cos θ cos ψ cos φ sin θ cos ψ 
sin θ sin φ − sin θ cos φ cos θ

= E(φ + π, − θ, ψ + π) (40.2)

So archetypically symmetric would O(3) appear to be that it seems distinctly

odd that Euler would have adopted such an asymmetric procedure to assign
names to the elements of O(3). It becomes in this light natural to ask: “Can
a more symmetrical variant of Euler’s procedure be devised—a procedure that
extends straightforwardly to the general case O(n)?” Whatever may be the
answers to those questions, it will be noticed (see Figure 7) that Euler’s defining
procedure does relate very naturally/directly to the casually observed behavior
of tops.
From our interest in the motion of rigid bodies we acquire interest in the
rotation-induced adjustments x 0 −→ x of the coordinates relative to the space
frame of points that are fixed in the body (i.e., with respect to the body frame).
The transformation of interest to us is therefore not passive but active (not an
“alias” but an “alibi transformation” in Wigner’s language): compare Figure 8
with Figure 6. We write, as has been our practice since page 2,

x = Rx
x0

where now
R = E –1 (φ, θ, ψ)
22 Gyrodynamics

Figure 7: It is, I suspect, not by accident that Euler’s angles equip

one to describe very simply and directly the motion executed by tops:
φ measures precession,
θ measures tilt (nutation),
ψ measures spin.
The line of nodes is shown here as a dotted line.

y
y0

x x0

Figure 8: The active (or “alibi”) interpretation that at (41) is used

to introduce Euler angles into the theory of tops.
Euler angles 23

Explicitly
   
cos φ − sin φ 0 1 0 0 cos ψ − sin ψ 0
R =  sin φ cos φ 0 0 cos θ − sin θ   sin ψ cos ψ 0 
0 0 1 0 sin θ cos θ 0 0 1

= R 3 (φ) · R 1 (θ) · R 3 (ψ) (41)

To compute ω we proceed from

 
0 −ω3 ω2
 ω3 0 −ω1  = Ṙ R –1
−ω2 ω1 0


= Ṙ 3 (φ)·R 1 (θ)·R 3 (ψ) + R 3 (φ)· Ṙ 1 (θ)·R 3 (ψ) + R 3 (φ)·R 1 (θ)· Ṙ 3 (ψ)
· R 3T (ψ)·R 1T (θ)·R 3T (φ)
= Ṙ 3 (φ)·R 3T (φ) + R 3 (φ)· Ṙ 1 (θ)·R 1T (θ)·R 3T (φ)
+ R 3 (φ)·R 1 (θ)· Ṙ 3 (ψ)·R 3T (ψ)·R 1T (θ)·R 3T (φ)

But by quick calculation

   
0 −1 0 0
Ṙ 3 (φ)·R 3T (φ) = φ̇  1 0 0  = φ̇  0 ×
0 0 0 1

and similarly
 
1
Ṙ 1 (θ)·R 1T (θ) = θ̇  0 ×
0
 
0
Ṙ 3 (ψ)·R 3T (ψ) = ψ̇  0 ×
1

Drawing upon lemma (33) we therefore have the deceptively suggestive formal
statement
φ + θ̇θ + ψ̇
ω = φ̇ ψ (41.1)

where
     
0 1 0
φ ≡ φ̇  0  ,
φ̇ θ̇θ ≡ θ̇ R 3 (φ) 0  , ψ ≡ ψ̇ R 3 (φ) R 1 (θ)  0 
ψ̇
1 0 1
24 Gyrodynamics

give      
0 cos φ sin θ sin φ
φ = φ̇  0  ,
φ̇ θ̇θ = θ̇  sin φ  , ψ = ψ̇  − sin θ cos φ 
ψ̇ (41.2)
1 0 cos θ
whence finally  
θ̇ cos φ + ψ̇ sin θ sin φ
ω= θ̇ sin φ − ψ̇ sin θ cos φ  (42)
φ̇ + ψ̇ cos θ

The symbols φ̇φ, θ̇θ and ψ̇

ψ are “deceptively suggestive” in that they are
intended to be read wholistically: they are not intended to be read as references
to the time derivatives of vectors φ, θ and ψ . Suppose it were possible to write
 
α1 (φ, θ, ψ)
ω = α̇α with α =  α2 (φ, θ, ψ) 
α3 (φ, θ, ψ)

We would then have

 
α1,φ (φ, θ, ψ)dφ + α1,θ (φ, θ, ψ)dθ + α1,ψ (φ, θ, ψ)dψ
ω dt =  α2,φ (φ, θ, ψ)dφ + α2,θ (φ, θ, ψ)dθ + α2,ψ (φ, θ, ψ)dψ 
α3,φ (φ, θ, ψ)dφ + α3,θ (φ, θ, ψ)dθ + α3,ψ (φ, θ, ψ)dψ
 
and the functions αk,λ (φ, θ, ψ) : k ∈ 1, 2, 3 , λ ∈ φ, θ, ψ would assuredly
satisfy the integrability conditions

αk,φθ = αk,θφ , αk,φψ = αk,ψφ , αk,θψ = αk,ψθ

Which the αkλ -functions latent in (42) obviously do not satisfy: it is not possible
α. Right at the heart of 3-dimensional rotational kinematics lives
to write ω = α̇
a non-integrability condition. Contrast this with the 2-dimensional situation,
where if  
cos θ sin θ
R=
− sin θ cos θ
then
 
0 θ
Ṙ R T = d
, which by dualization becomes simply ω = d
dt θ
dt −θ 0

We found at (13) that the intrinsic (or rotationial) kinetic energy can be
described
T = 12 ω T I ω
which by (25) becomes

T = 12 ω T R I 0 R –1 ω
= 12 ω 0 T I 0 ω 0 with ω 0 ≡ R –1 ω (43)

One could evaluate ω 0 by Mathematica -assisted brute force. Or one could work
Lagrangian formalism using Euler angles 25

from the dual of (43); i.e., from

A0 = R –1A R
= R –1 Ṙ R –1 R = R –1 Ṙ

by the methods that led to (42). By either procedure one is led to

 
φ̇ sin θ sin ψ + θ̇ cos ψ
ω 0 =  φ̇ sin θ cos ψ − θ̇ sin ψ  (44)
φ̇ cos θ + ψ̇

8. Lagrangian formalism using Euler angles. It follows clearly from results now
in hand that the rotational dynamics of a rigid body can be considered to
devolve from a Lagrangian of—if we exercise our option to identify the body
frame with the principal axis frame—the form

L = L(φ̇, θ̇, ψ̇, φ, θ, ψ)

= 12 I10 (φ̇ sin θ sin ψ + θ̇ cos ψ)2 + 12 I20 (φ̇ sin θ cos ψ − θ̇ sin ψ)2
+ 12 I30 (φ̇ cos θ + ψ̇)2 − U(φ, θ, ψ) (45)

The resulting Lagrange equations are coupled differential equations of second

order, and are of much more complicated appearance than the Euler equations,
which were seen at (35) to be a symmetric set of first order equations. Note,
however, that  
d ∂ − ∂ L=0 (46.1)
dt ∂ ψ̇ ∂ψ
becomes
d
−∂ U/∂ψ =I30 dt (φ̇ cos θ+ψ̇) +(I 0 −I 0 ) (φ̇ sin θ cos ψ−θ̇ sin ψ) (φ̇ sin θ sin ψ+θ̇ cos ψ)
      2 1     
N30 ω30 ω20 ω10

which precisely reproduces the third of the Euler equations (35). Certain linear
combinations of the remaining Lagrange equations
 
d ∂ − ∂ L=0 (46.2)
dt ∂ θ̇ ∂ θ
 
d ∂ − ∂ L=0 (46.3)
dt ∂ φ̇ ∂φ
serve to reproduce first and second of the Euler equations.
One can understand the relative complexity the Lagrange equations (46)
on grounds that they undertake
 to accomplish more than Euler equations.
Solutions φ(t), θ(t), ψ(t) of (46) serve in themselves to describe how the body
gyrates, while solutions ω (t) of (35) leave us—as we saw at (36)—one awkward
integration away from such explicit information.
26 Gyrodynamics

If I10 = I20 then the rigid body (or “top”) is said to be symmetrical,14
and the Lagrangian (45) assumes the form

L = 12 I10 (φ̇2 sin2 θ + θ̇2 ) + 12 I30 (φ̇ cos θ + ψ̇)2 − U(φ, θ, ψ) (47)

while the Euler equations (35) become


N10 = I10 ω̇10 − (I10 − I30 ) ω30 ω20 

N20 = I20 ω̇20 + (I10 − I30 ) ω10 ω30 (48)


N30 = I30 ω̇30

Had we instead set (not I10 = I20 but, say) I10 = I30 then the simplification of
(45) would have been masked or disguised, while the simplified Euler equations
would be precisely similar to (48). It is, within the Lagrangian formalism, as
a mere convenience, and without real loss of generality, that one identifies the
3-axis of the principal axis frame with the symmetry axis of a symmetrical top.
If I10 = I20 = I30 then the Lagrangian simplifies still further, to become

L = 12 I10 (φ̇2 + 2φ̇ψ̇ cos θ + ψ̇ 2 ) − U(φ, θ, ψ) (49)

while the Euler equations assume the trivial form

N 0 = I10 ω̇
ω0 (50)

We will return to discussion of some of the remarkably rich physics that arises
in these important special cases; i.e., to a discussion of the solutions of the
associated equations of motion.

9. Free motion of a rigid body. Working first in the space frame (since it is from
a position at rest with respect to the almost-inertial laboratory that we expect
to view our gyrating rigid objects), we return to (31) and, setting N = 0, obtain

S = I ω̇
Ṡ ω + ω×I ω = 0 (51)

according to which

Spin S is a constant of the free motion of a rigid body (52)

Trivially,
So also is S 2 ≡ S·S a constant of the free motion (53)

14
The term conveys no information about the actual shape of the top.
Free motion of a rigid body 27

But the general constancy of S does, in general, not imply “uniformity” of

the rotation: rotation typically causes I to become time -dependent, which by
ω = I –1 (t)S
S forces ω to be time -dependent. Rotational uniformity ω̇ ω = 0 is
seen by (51) to entail that ω × I ω = 0; i.e., that ω and I ω be parallel :
Rotational uniformity ω̇ ω = 0 requires that
ω be  (54)
an eignevector of I : I ω = λω
ω with λ ∈ I1 , I2 , I3

For a free rigid body the rotational energy is all kinetic, and as we saw at
(18.2) can be described T = 12 ω · S . To prepare for a proof that Ṫ = 0 I digress
to establish the following

LEMMA: The objects

   
0 −a3 +a2 a1
A = Aij =  +a3 0 −a1  and a =  a2 
−a2 +a1 0 a3

encountered on page 15 are “dual” in the sense that

Aij = &ikj ak (55)

The Levi-Civita symbol is known15 to assume the same numerical

values in every coordinate system if transformed as a tensor density
of weight w = −1:

&˜ikj ≡ W –1 · W a i W c k W b j &acb where W ≡ detW

= &ikj

Let it be assumed that the ak transform as components of a

weightless contravariant vector, and that the Aij transform as
components of a covariant tensor density of second rank and
negative unit weight. The assertion (55) then preserves its design
in all coordinate systems, and in

Ãij = &ikj ãk (56.1)

we have simply the statement that

transform of dual = dual of transform (56.2)

Explicitly
W –1 · W m i W n j Amn = &ikj M k p ap
which in index-free notation becomes

(det W)–1 · W TAW = (Ma

a)× (56.3)

where W ≡ M –1 . Notice that (56) gives back (33) as a special case.

15
See classical electrodynamics (/), pages 172/3.
28 Gyrodynamics

Returning now to the demonstration that T is conserved, we have

Ṫ = 12 ( ω̇
ω · S + ω · Ṡ
S)
ω·S
= 12 ω̇ S =0
by (51): Ṡ
But—again by (51)—
ω = −I –1 (ω
ω̇ ω × I ω)
= −I –1 (ω
ω ×)I –1 · I 2ω
But the symmetry of I implies that of I –1 , so we can use lemma (56.3) to obtain
= −(det I)–1 (I ω ) × I 2ω (57)
giving
Ṫ = 12 (det I)–1 (I 2 ω × I ω )·· I ω
But the triple scalar product
 
 a1 b1 c1 

a × b)·· c =  a2
(a b2 c2  = 0 unless a, b and c are linearly independent
 a3 b3 c3 
Which in the preceding equation is clearly not the case. So we have
Ṫ = 0 : T is a constant of the free motion (58)

It is interesting to notice that while

ω · I 2 ω = S2
and
ω · I ω = 2T
are constants of free rigid body motion, ω · ω = ω 2 is typically not constant, for
2 dt ω = ω̇ω = ω̇ · ω which by (57) supplies
1 d 2
ω
1 d 2
2 dt ω = (det I)–1 (I 2ω × I ω )·· ω
= (det I)–1 [I 2ω , I ω , ω ] in a standard triple scalar product notation
= (det I)–1 (I1 − I2 )(I1 − I3 )(I2 − I3 )ω1 ω2 ω3

Drawing upon ω = R ω 0 , S = RS S 0 , I = R I 0 R–1 we see by easy arguments

• That S 0 is not constant unless R happens to describe spin about the
(invariable) S -axis;
• That therefore ω 0 = (I 0 )–1S 0 is generally not constant either;
• That S 2 = S · S = S 0· S 0 provide alternative descriptions of S 2 (conserved);
• That T = 12 S T I –1S = 12 S 0 T (I0 ) –1S 0 provide alternative descriptions of T
(conserved);
• That ω 2 = ω · ω = ω 0· ω 0 provide alternative descriptions of ω 2 (not
conserved).
Free motion of a rigid body 29

We look now to the time -dependence of ω 2 . Drawing upon (57) we have

1 d
2 dt
ω 0· ω 0
ω 2 = ω̇
= (det I 0 )–1 (I 02 ω 0 × I 0 ω 0 )·· ω 0

Electing to work in the principal axis frame, where I 0 is diagonal, and agreeing
to omit all 0 superscripts for the duration of this argument, we therefore have
 2 
 I1 ω1 I1 ω1 ω1 

= 1  I22 ω2 I2 ω2 ω2 
I1 I2 I3  2
I3 ω3 I3 ω3 ω3 
 
I12 I22 I32
det J
= ω1 ω2 ω3 with J ≡  I1 I2 I3  (59)
I1 I2 I3
1 1 1

To bring our conservation laws into play we write

   2  
ω12 S α1
J  ω22  =  2T  ≡  α2 
ω32 ω2 α3

which gives
   
ω12 α1
 ω22  = J –1  α2 
ω32 α3
 
(I2 − I3 ) −(I2 + I3 )(I2 − I3 ) I2 I3 (I2 − I3 )
J –1 = (det J )–1 ·  (I3 − I1 ) −(I3 + I1 )(I3 − I1 ) I3 I1 (I3 − I1 ) 
(I1 − I2 ) −(I1 + I2 )(I1 − I2 ) I1 I2 (I1 − I2 )

From this equation it follows in particular that

(I2 − I3 )(I3 − I1 )(I1 − I2 ) 

(ω1 ω2 ω3 )2 = etc.
(det J )3

with
    
etc. ≡ α1 −(I2 +I3 )α2 +I2 I3 α3 α1 −(I2 +I3 )α2 +I2 I3 α3 α1 −(I2 +I3 )α2 +I2 I3 α3

But det J = −(I2 − I3 )(I3 − I1 )(I1 − I2 ), so we have

 2 
det J 1
ω1 ω2 ω3 =− etc.
I1 I2 I3 (I1 I2 I3 )2
= (λ1 − ω 2 )(λ2 − ω 2 )(λ3 − ω 2 )
30 Gyrodynamics

with
2(I2 + I3 )T − S 2
λ1 =
I2 I3
2(I3 + I1 )T − S 2
λ2 =
I3 I1
2(I1 + I2 )T − S 2
λ3 =
I1 I2
which, it will be noticed, are assembled from frame-independent
• system parameters and
• constants of the free motion
Returning with this information to (59) we come at last to the statement

ω dω = (λ1 − ω 2 )(λ2 − ω 2 )(λ3 − ω 2 )
dt
which can be used to compute

τ ≡ transit time: ωinitial → ωfinal

 ωfinal
=  ω dω (60)
ωinitial (λ1 − ω 2 )(λ2 − ω 2 )(λ3 − ω 2 )

The integral leads to elliptic functions with complicated arguments, but is an

integral with which Mathematica appears to be quite comfortable.16
Looking again to points ennumerated at the bottom of page 28, we see that
S 0 ranges simultaneously on a sphere of radius S in spin space

(S10 )2 + (S20 )2 + (S30 )2 = S 2

and on an energy ellipsoid

 S 0 2  S 0 2  S 0 2
√1 + √2 + √3 =T
2I1 2I2 2I3

with semi-axes that we may, as a handy convention, assume to have been

indexed in ordered sequence
  
2TI1
2TI2
2TI3

It is clear therefore on simple geometrical grounds (see Figure 9) that if S 2 is

given/fixed then
Tleast  T  Tmost

16
My derivation of (60) has been freely adapted from the discussion that
can be found in §137 of E. J. Routh, Advanced Dynamics of a System of Rigid
Bodies (6th edition ).
Free motion of a rigid body 31

Figure 9: The black circle represents the sphere of radius S 2 in

3-dimensional S 0-space. From the relationship of the small blue
energy ellipse to the sphere one deduces on purely geometrical
grounds that
2
Tleast = S (61.1)
2Ilargest
while from the red ellipse one obtains

Tmost = S2 (61.2)
2Ismallest
The intermediate ellipses were produced by incrementing the energy
in equal steps.

Figures 10 are taken from a filmstrip,17 and reveal a physically important

new aspect of the situation that becomes evident only when advances from two
to three dimensions. The free gyration of a rigid body causes S 0 to wander
(precisely how?) along the intersection of the S-sphere and T-ellipsoid. We
learn from the figures to expect
• stable rotation about the major axis if T = Tleast ;
• stable rotation about the minor axis if T = Tmost ; but
• rotation about the intermediate axis (energy T = Tcritical ) to be unstable.
17
I have written
x2 + y 2 + z 2 = 22
to describe the angular momentum sphere, and

(x/4)2 + (3y/8)2 + (z/2)2 = ( 14 + n 16

1
) : n = 0, 1, 2, . . . , 12

to describe a sequence of progressively more energetic ellipsoids. We then have

x-puncture at n = 0, y -puncture at n = 5, z -puncture at n = 12. For relevant
discussion see P. L. Lamy & J. A. Burns, “Geometrical approach to torque free
motion of a rigid body having internal energy dissipation,” AJP 40, 441 (1972)
and W. G. Harter & C. C. Kim, “Singular motions of asymmetric rotators,”
AJP 44, 1080 (1976).
32 Gyrodynamics

Figure 10 a: The ellipsoid of least energy is entirely interior to the

S-sphere, to which it is tangent at only two points—the puncture
points of the major principal axis.

Figure 10 b: At an energy T somewhat greater than Tleast the

energy ellipsoid has become visible in the neighborhood of the major
axis. Note that sphere and ellipsoid intersect on a roughly elliptical
curve that envelops the major axis.
Free motion of a rigid body 33

Figure 10 c: At a specific still greater energy

Tcritical = S2 (61.2)
2Iintermediate

the curves that enveloped the major axis have fused, to produce an ×
at the puncture-point of the intermediate principal axis.

Figure 10 d: At energies T > Tcritical the energy and angular

momentum surfaces intersect on curves that envelop the minor
principal axis.
34 Gyrodynamics

Figure 10 e: The ellipsoid of greatest energy is entirely exterior to

the S-sphere, to which it is tangent at only two points—the puncture
points of the minor axis.

0
10. Poinsot’s construction. The preceding constructions live in S -space: they
tell us—on the assumption that the values of I1 , I2 , I3 are known, and that the
values of S and T have been prescribed—where S 0 is free to roam, but not how
in time it elects to do so. We have

ω10 = S10 /I1 , ω20 = S20 /I2 , ω30 = S30 /I3

but have, as yet, no diagramatic interpretation of the motion of ω 0 . And even

if we did possess ω 0 (t), it would be a long and arduous journey back to R(t),
to understanding of how the rigid body itself moved.

These limitations were neatly circumvented by Louis Poinsot (–),

who devised a construction that owes its striking success mainly to the fact that
it employs variables that refer not to the body frame but to the space frame.
The equation
T(ωω ) ≡ 12 ω T I(t)ω
ω=T

defines what I will call the T-ellipse: its center is pinned to the origin of ω -space,
and its axes—of lengths
  
2T/I1
2T/I2
2T/I3

—wobble about, reflecting the t-dependence of I and remaining always in

coincidence with the principal axes of the body (though the latter live not
in ω -space but in r -space).
Poinsot’s construction 35

d ω

Figure 11: Physical variables S and ω⊥ ≡ 2T/S determine the

placement and orientation of Poinsot’s invariant plane , to which
the T-ellipse remains ever tangent

Poinsot’s construction proceeds now from two key observations, the first
of which is that
∇ω T(ω
ω) = I ω = S

In words: S stands normal to the plane which tangentially kisses the T-ellipsoid
at ω . The planes thus constructed at various times t will, by the invariance of
S , be parallel . Poinsot observed further that those planes are in fact identical,
for the normal distance from origin to plane

ω⊥ = ω · Ŝ
S = 2T/S = constant of free gyro motion

S ≡ S /S and ω⊥ = 2T/S serve conjointly to identify a unique

In short: Ŝ
S , T)—and since they are constants of the motion the plane is
plane—call it Π(S
invariable. Evidently the T-ellipsoid is forced by the Euler equations to move
in such a way that
• while the center of the ellipsoid remains pinned at the origin
• the ellipsoid remains at every instant tangent to the invariable plane;
• the point ω of tangency announces the instantaneous angular velocity of
the ellipsoid (and evidently traces on the surface of the ellipsoid a closed
curve).
36 Gyrodynamics

To complete Poinsot’s construction we return to page 11, where it was

reported that the moment of inertia about the axis (through the center of mass)
defined by the unit vector n can be described
n · I n = M R2 (n
n) (62)
n) is the so-called
where M refers to the total mass of the rigid body, and R(n
radius of gyration. Introducing

ρ ≡ n/ M R2 (n n) (63)
we find that (62) can be written
ρ·Iρ = 1 (64)
which serves to define the “inertia ellipsoid” in ρ -space. The point to notice is
that (64) follows also from ω T I(t)ω
ω = 2T upon setting
√
ρ = ω / 2T (65.1)
Evidently the inertia ellipsoid is—though it lives not in ω -space but in ρ -space
—a similarly oriented but rescaled copy of the T-ellipsoid. As such, it moves in
such a way as to be ever tangent to an invariable plane, from which its center
maintains a distance √
ρ⊥ = ω⊥ / 2T (65.2)

While it would border on deceptive absurdity to attempt to “unpin the

center of the T-ellipsoid and to transport it from point to point in ω -space,” it
is meaningful to do such a thing with the inertia ellipsoid in ρ -space.18 With
Poinsot, we observe that the vector
ρ : directed center −→ contact point
is, by (65.1), momentarily axial: points on that line are therefore momentarily
at rest. Which is to say: the inertial ellipsoid rolls without slipping on the
invariant plane, and at an instantaneous rate proportional to the length of the
ρ vector (see Figure 12.) Since the principal axes of the inertial ellipsoid
remain ever parallel to those of the rigid body itself, Poinsot has given us what
is, in effect, an ingenious special-purpose analog computer—its only limitation
being that it lives in a fairly abstract space.
It should, perhaps, be noted that the figure of Poinsot’s inertia ellipsoid
—with its semi-axes of lengths
  
1/I1
1/I2
1/I3 (66.1)
—is in an obvious sense “reciprocal” to that of the energy ellipsoid encountered
on page 30 and in Figures 10: the former lives in ρ -space, the latter in S -space,
and has semi-axes of lengths
  
2TI1  2TI2  2TI3 (66.2)

18
That is why we took the trouble to introduce ρ -space !
Poinsot’s construction 37

ρ⊥
ρ

invariant plane

Figure 12: Poinsot’s construction: The inertial ellipsoid, which

lives in ρ -space, rolls without slipping on the invariant plane, with
ρ⊥ held constant.. The moving contact point traces a closed curve,
called the “polhode” (from πóλoς = axis + óδóς = path), on the
surface of the ellipsoid, and a typically more complicated curve
called the “herpolhode” on the plane. In the figure the invariant
plane has—for clarity—been laid flat, and only the herpolhode is
shown.

It will prove useful to observe in this connection that if the physical ellipsoid

(x1 /a1 )2 + (x2 /a2 )2 + (x3 /a3 )2 = 1

is filled with material of uniform density d then, by quick calculation,19 the

total mass of the object is given by

M = d · 43 πa1 a2 a3

and the moment of inertia matrix becomes

   1 2 2
I1 0 0 
 I1 = 5 M (a2 + a3 )
I =  0 I2 0  with I2 = 15 M (a21 + a23 )


0 0 I3 I3 = 15 M (a21 + a22 )

(which gives back a very familiar result when a1 = a2 = a3 = r). Let it be

assumed that a1
a2
a3 . We then have
  
I1  I2  I3 whence 1/I1
1/I2
1/I3

19
See Problem 2-1.
38 Gyrodynamics

—the implication being that the physical ellipsoid and the associated Poinsot
ellipsoid have distinct but qualitatively similar figures. During the first half of
the 19th Century close study of the geometry of polhode and herpolhode appears
to have been a flourishing industry. Many wonderful facts were discovered,
most of which are now forgotten, but some of which can be recovered from
the old textbooks. Webster reports, for example,20 that while polhodes are
invariably reentrant, herpolhodes are usually not, and never possess inflection
points: their name (from ´ρπιν = to creep like a snake) is therefore somewhat
misleading. The old literature provides elaborate figures produced by laborious
hand calculation. It would be amusing—possibly instructive—to use modern
computer resources to recreate some of that material, to produce animated
images of rolling Poinsot ellipsoids, etc.
I must emphasize that Poinsot’s construction pertains to the gyrodynamics
of free bodies. The application of torques would, in general, cause both S and
T to become time-dependent. The formerly “invariant plane” would begin to
move, to wobble, and the center of the inertial ellipsoid to rise and fall with
respect to that plane: the whole construction would become “seasick,” and
rapidly lose its utility.

11. First look at the free gyration of a symmetric top. By “symmetric” we refer
here not to the shape of the body itself, but to the shape of its only dynamically
relevant feature—the inertia ellipsoid (or—reciprocally—the energy ellipsoid in
spin space) . . . though in practice most rigid bodies that are symmetric in the
above sense are axially symmetric also in their spatial form. I will occasionally
allow myself to call such bodies “tops.”
We are obliged at the outset to distinguish (see Figure 13) two principal
classes of axially symmetric tops:
oblate prolate
A ≡ I1 > I2 = I3 ≡ B B ≡ I1 = I2 > I3 ≡ A

Standing at the interface between those two classes is the essentially trivial class
spherical
A ≡ I1 = I2 = I3
of fully symmetric tops. In the presence of symmetry the Euler equations (35)
simplify: we have
 0
I1 ω̇1 =0
 
 I3 ω̇2 + (I1 − I3 )ω1 ω3 = 0  : oblate case
I3 ω̇3 − (I1 − I3 )ω1 ω2 = 0

20
A. G. Webster, The Dynamics of Particles and of Rigid Bodies (2nd edition
; Dover reprint ), page 264; see also J. B. Hart, “Incorrect herpolhodes
in textbooks,” AJP 37, 1064 (1969).
First look at the free gyration of a symmetric top 39

B B B B

A A

Figure 13: Symmetric tops come in two flavors. At the top are
triangle diagrams of the sort first encountered on page 8, and below
are oblate/prolate Poinsot ellipsoids (figures of revolution) typical
of that symmetry class.

and  0
I1 ω̇1 − (I1 − I3 )ω3 ω2 = 0
 
 I1 ω̇2 + (I1 − I3 )ω3 ω1 = 0  : prolate case
I3 ω̇3 =0
which share the same abstract structure, together with
 0
I1 ω̇1 =0
 
 I1 ω̇2 = 0 : spherical case
I1 ω̇3 =0

which is trivial. Looking first to the former, we have

ω10 (t) = λ : constant

    0 
ω̇20 0 −Ω ω2 I1 − I3
= with Ω≡ λ
ω̇30 Ω 0 ω30 I3
so  
Ω
ω0 = Ω × ω
ω̇ with Ω ≡  0 
0
40 Gyrodynamics

of which the immediate solution is

 
0 0 0
ω 0 (t) = etWω 0 (0) with W =  0 0 −Ω  = Ω× (67)
0 Ω 0

We conclude (in a phrase that describes not the motion of the top itself, but
the motion of the vector ω 0 (t) in ω 0 -space) that

ω 0 (t) precesses about the Ω -vector with angular velocity Ω

For a prolate top the same line of argument gives


0
0
ω = −Ω
ω̇ Ω× ω with Ω ≡  0 
Ω
I1 − I3
Ω≡ λ
I1
λ ≡ constant value of ω30 (t)

—the conclusion being similar to within a sign. Reverting to the notation

introduced at Figure 13, we have

 A − B λ in the oblate case (A > B): precession prograde


 B
Ω= (68)



 B − A λ in the prolate case (B > A): precession retrograde
B

The situation is illustrated in Figure 14. For discussion of a geophysical

instance of this kind of free rotor precession, and its relation to “Chandler
wobble,” see Goldstein.21
Looking back to (59) on page 29, we see that axial symmetry of any type
—be it oblate, spherical or prolate—immediately entails

det J = (I1 − I2 )(I1 − I3 )(I2 − I3 ) = (A − B)(A − C)(B − C) = 0

giving
det J
d
dt ω2 = ω1 ω2 ω3 = 0 : ω 2 is conserved (69)
I1 I2 I3
For tops-in-general ω and ω 2 are both time-dependent (unless ω happens to be
an eigenvector of I, in which case both are constant: see again (54)), but for all
symmetric tops ω 2 becomes constant, though ω typically continues to wander.

21
H. Goldstein, Classical Mechanics (2nd edition ), page 212.
Detailed account of the free gyration of a symmetric top 41

Figure 14: ω 0 is seen in the body frame to precess prograde around

the symmetry axis in the oblate case, retrograde in the prolate case.
The component ω 0
of ω 0 (shown here in red) that parallels the
symmetry axis is stationary, while ω 0⊥ revolves with constant
angular velocity Ω. That the motion of ω 0 is—for symmetric tops
—length-preserving is seen in this light to be not at all surprising.

12. Detailed account of the free gyration of a symmetric top. Knowledge of ω 0 (t),
or even of ω (t), leaves one still an integration away from a description of the
motion R(t) of the physical top itself. To gain the latter kind of understanding
we look to the motion of the Euler angles that serve to describe the relation of
the body frame to the space frame.
We know that S is conserved, and will (without real loss of generality)
look to those motions with the property that S is aligned with the 3 -axis of the
space frame:  
0
S = 0  (70)
S
It is also without loss of generality that we will identify the symmetry axis of
the top with the 30 -axis of the body frame . . . which is to say: we will stipulate
42 Gyrodynamics

the symmetry of the top by setting I1 = I2 . The Lagrangian then becomes (see
again (47) on page 26)

L = 12 B(φ̇2 sin2 θ + θ̇2 ) + 12 A(φ̇ cos θ + ψ̇)2 (71)

in which, it will be noticed, neither φ nor ψ makes an appearance, so the

conjugate momenta pφ and pψ are known already to be constant. Writing out
the Lagrange equations, we have

∂L = B φ̇ sin2 θ + A(φ̇ cos θ + ψ̇) cos θ = constant 



∂ φ̇ 


∂L = A(φ̇ cos θ + ψ̇) = constant
(72)
∂ ψ̇ 



d ∂L − ∂L = B θ̈ − B φ̇2 sin θ cos θ + A(φ̇ cos θ + ψ̇)φ̇ sin θ = 0 


dt ∂ θ̇ ∂θ

As a preliminary to discussion of the implications of (72) I digress to

translate (70) into a statement relating Euler angles and their derivatives. We
introduce (44) into
 
B 0 0
S =  0 B 0  ω0
0

0 0 A

and obtain  
B(θ̇ sin φ sin ψ + θ̇ cos ψ)
=  B(θ̇ sin φ cos ψ − θ̇ sin ψ) 
A(φ̇ cos θ + ψ̇)

S 0 and taking R from (41), we compute

Recalling now that S = RS
 
S1
S =  S2 
S3
 
B(−φ̇ sin θ cos θ sin φ + θ̇ cos φ) + A(φ̇ cos θ + ψ̇) sin θ sin φ
=  B(+φ̇ sin θ cos θ cos φ + θ̇ sin φ) − A(φ̇ cos θ + ψ̇) sin θ cos φ  (73)
B φ̇ sin2 θ + A(φ̇ cos θ + ψ̇) cos θ

Evidently
S1 cos φ + S2 sin φ = B θ̇

But at (70) we set S1 = S2 = 0, so we have θ̇ = 0 giving

θ(t) = θ0 : constant (74)

It proves handy to note also that (taking R –1 from (40.2))

Detailed account of the free gyration of a symmetric top 43

   
0 S sin θ0 sin ψ
S 0 = R –1 0  =  S sin θ0 cos ψ 
S S cos θ0

From the last of the equations (73) we see that the first of the Lagrange
equations (72) can be formulated

B φ̇ sin2 θ + A(φ̇ cos θ + ψ̇) cos θ = S

and when multiplied by cos θ, becomes

 
B φ̇ cos θ − A(φ̇ cos θ + ψ̇) sin2 θ + A(φ̇ cos θ + ψ̇) = S cos θ

The third Lagrange equation, by θ̈ = 0, has become

 
B φ̇ cos θ − A(φ̇ cos θ + ψ̇) φ̇ sin θ = 0

which when brought to the preceding equation tells us that the second Lagrange
equation can be written

A(φ̇ cos θ + ψ̇) = S cos θ (75)

The equation
 
B φ̇ cos θ − A(φ̇ cos θ + ψ̇) = 0
 
has therefore become B φ̇ − S cos θ = 0, which supplies

φ̇ = S : constant (76)
B
Returning with this information to (75) we obtain

ψ̇ = B − A S cos θ0 : constant (77)

AB
The functions φ(t) and ψ(t) are therefore linear in t, and the exact solutions of
the Lagrange equations (72) can be presented

θ(t) = θ0 




φ(t) = Ωφ t + φ0 with Ωφ ≡ S (78)
B 



ψ(t) = Ωψ t + ψ0 with Ωψ ≡ B − A S cos θ0 
AB

The fixed “angle of tilt” θ0 —whence also Ωφ and Ωψ —is determined by

joint specification of S and T (together with A and B, which describe the
44 Gyrodynamics

effective figure of the symmetric top), as I now show: Returning to (71) with
descriptions of B φ̇ and A(φ̇ cos θ + ψ̇) that were developed just above, we obtain
 
T = L = 1 1 sin2 θ0 + 1 cos2 θ0 S 2 (79.1)
2 B  A 
= 1 S −
2 1 1 − 1 S 2 sin2 θ
0
2A 2 A B 
= 1 S 2 + 1 1 − 1 S 2 cos2 θ0 (79.2)
2B 2 A B

giving    
tan θ0 = 1 − 2T / 2T − 1 (80)
A S2 S2 B
It is gratifying to note in this connection that (61) supplies

1
2T
1 : oblate case
B S2 A
1
2T
1 : prolate case
A S2 B

so the parenthetic expressions under the radical have in all cases the same sign:
we are at (80) never asked
√ to take the square root of a negative number. Making
use now of cos2 = 1/ 1 + tan2 we obtain finally
   
Ωψ = B − A S 2T − 1 / 1 − 1
AB S2 B A B
  
=S 1 − 1 2T − 1 (81)
A B S2 B

where again—for reasons just stated—the expression under the radical is in all
cases non-negative.
I enter now upon a series of elementary remarks that culminate in a
celebrated geometrical interpretation of the R(t) implicit in (78):
• From T = 12 ω ·S = 12 ωS cos α and the established facts that for a symmetric
free top not only T and S but also ω are constants of the motion, we see
that for such a top the angle S ∠ ω (we have named it α) is invariant. We
have
2T/S
cos α = (82)
ω
which in the notation of Figure 11 becomes simply cos α = d/ω
• Bringing θ̇ = 0 to the description (42) of ω , we find

ω 2 = ψ̇ 2 + 2ψ̇ φ̇ cos θ + φ̇ 2 (83)

which invites the diagramatic representation shown in the following figure:

Detailed account of the free gyration of a symmetric top 45

ω
φ̇

θ
ψ̇
It follows in any event from (78) that

ω 2 = Ω2ψ + 2Ωψ Ωφ cos θ0 + Ω2φ

      
= 1 − 1 2 + 2 1 − 1 1 cos2 θ + 1 2 S 2
0
 A B A B B B
= 1 2 1 2
cos θ0 + 2 sin θ0 S 2
(84)
A2 B

• Returning with this information and (79) to (82) we have

1
cos2 θ0 + 1
sin2 θ0
cos α =  A B
(85)
1
A2 cos2 θ0 + 1
B2 sin2 θ0

Evidently α → 0 as A → B
• The angle ω ∠(symmetry axis)—call it β—can be obtained from

ω30
cos β =
ω
φ̇ cos θ0 + ψ̇
= by (44)
ω
S
A cos θ0
=  by (75) and (84)
S A12 cos2 θ0 + B12 sin2 θ0

= 1 (86)
!
A 2
1+ B tan2 θ0

√
Drawing again upon cos = 1/ 1 + tan2 , we have

tan β = ± B
A
tan θ0 (87)

We have learned that β is (like α) a dynamical invariant. Evidently β → θ0

as A ↑ B.
46 Gyrodynamics

β
θ0 α

invariant plane

space frame

α θ0

β
invariant plane

space frame

Figure 15: Two-dimensional sections of figures drawn in ω -space,

where the colored triangles become cones. Conserved S sets the
orientation of the invariant plane, T sets its distance above the
origin. The dashed line identifies the symmetry axis, so the top
figure refers to the oblate case, the lower figure to the prolate
case. The blue “space cone” is dynamically invariant, the red “body
cone”—interior to the space cone in the oblate case, exterior in
the prolate case—is fixed with respect to the top, and its motion
(“rolling-without-slipping,” as described in the text) provides a
representation of the motion of the physical top.

• From the figure, which was designed to make plain the meanings of the
invariant angles θ0 , α and β, we read
"
β−α : oblate cases
θ0 = (88)
β+α : prolate cases
Detailed account of the free gyration of a symmetric top 47

In either case we have

cos α = cos(θ0 − β) = cos β cos θ0 (1 + tan β tan θ0 )

= cos β cos θ0 (1 + A
B tan2 θ0 ) by (87)

where the minus sign in (87) has been dismissed as an artifact. It is

gratifying to notice that if we draw upon (86) we obtain
A
(1 + tan2 θ0 )
= cos θ0  !
B
A 2
(1 + B tan2 θ0 )

from which, after simplifications, we are led back again to precisely (85).
• At (79.2) we had an equation which by (76) and (77) can be written

2T/S = φ̇ + ψ̇ cos θ0

so the equation (82) at which we introduced α becomes

φ̇ + ψ̇ cos θ0
cos α = (89.1)
ω
This equation is structurally reminiscent of an equation

ψ̇ + φ̇ cos θ0
cos β = (89.2)
ω

encountered in the derivation of (86). Looking to the square of (89.1) and

drawing upon the description (83) of ω 2 , we have

φ̇2 + 2φ̇ψ̇ cos θ0 + ψ̇ 2 [1 − sin2 θ0 ]

cos2 α =
ψ̇ 2 + 2ψ̇ φ̇ cos θ0 + φ̇2
ψ̇ 2 sin2 θ0 2
=1− =⇒ ψ̇ 2 = ω 2 sin2 α (90.1)
ω2 sin θ0

while by a similar argument

sin2 β
φ̇2 = ω 2 (90.2)
sin2θ0

• Evidently

φ̇ sin α = ψ̇ sin β i.e., Ωφ sin α = Ωψ sin β (91)

which provides the basis for the claim—developed in the following figures—
that the body cone (with vertex angle β) rolls without slipping on the
stationary space cone (vertex angle α).
48 Gyrodynamics

r sin α
r sin β
r
α β

Figure 16: Attempts to represent the body cone rolling-without-

slipping on the space cone in the prolate case. Rotation of the
space cone through angle dφ about the verticle S axis will mesh with
rotation of the body cone through angle dψ only if

dφ · r sin α = dψ · r sin β

But that is precisely the upshot of (91). The lower figure provides
another representation of the same principle, and makes clear the
fact that  rotation of the body cone causes its center to advance
around the space cone in that same  sense. The red axis in the
upper figure represents the symmetry axis of the top.
Detailed account of the free gyration of a symmetric top 49

Figure 17: Preceding constructions, here adapted to the oblate

case. Note that, while the body cone lies exterior to the space cone
the prolate case, it envelops the space cone in the oblate case. Note
also that—here as before— rotation of the body cone produces a
same-sense -advance of its center.
50 Gyrodynamics

S
ω
30

10

20
ω S 30

10

20
Figure 18: Alternative representations of the body cone rolling
without slipping on the space cone (prolate case above, oblate case
below). The rolling body cone controls the motion of the symmetry
axis of the body—the 30 -axis. The uniform rotation (about that
axis) is controlled by the Ω-vector that was introduced on page 40:
it is retrograde in the prolate case, prograde in the oblate case.
Instability of spin about the intermediate axis 51

The general results developed above can (in leading approximation) be

simplified in limiting special cases. One can, for example, readily imagine having
geophysical/astrophysical interest in slightly oblate symmetrical bodies:

B B
A 0< A−B 1
B

Flipped coins inspire interest in the gyrodynamics of highly oblate tops

B B B − 12 A
0< 1
A B

while tumbling needles are, in effect, highly prolate tops

B B
0< A 1
A B

I invite the reader to construct variants of figures 16–18 appropriate to those

cases.

13. Instability of spin about the intermediate axis.22 Let the Euler equations (35)
of a free asymmetric top be written
 0
I1 ω̇1 − (I2 − I3 )ω2 ω3 = 0
 
 I2 ω̇2 − (I3 − I1 )ω3 ω1 = 0 
I1 ω̇3 − (I1 − I2 )ω1 ω2 = 0

22
My primary source here has been the discussion presented by S. T. Thornton
& J. B. Marion in §11.12 of their Classical Dynamics of Particles and Systems,
(5th edition ).
52 Gyrodynamics

Assume I1 > I2 > I3 and agree for the purposes of this discussion to omit the 0 s.
Write  
ω1
ω =  2  : 2 and 3 infinitesimal
3
to express our presumption that ω lies in the immediate neighborhood of the
principal axis of greatest moment. In leading order (i.e., after abandoning terms
of 2nd order) we have

I1 ω̇1 = 0 ⇒ ω1 is constant
I2 ˙2 − (I3 − I1 )ω1 3 = 0
I1 ˙3 − (I1 − I2 )ω1 2 = 0

The last pair of equations can be “separated by differentiation.” One is led to

the conclusion that 1 and 2 are both solutions of an equation
(I1 − I2 )(I1 − I3 )
χ̈ = −ω12 χ
I2 I3
of which the general solution is

−iΩ1 t (I1 − I2 )(I1 − I3 )
χ(t) = P e +iΩ1 t
+ Qe with Ω1 ≡ ω1
I2 I3

Had we proceeded from

 
1
ω =  ω2  : 1 and 3 infinitesimal
3

we would have been led by a similar argument to the conclusion that 1 and 3
both move like

−iΩ2 t (I2 − I3 )(I2 − I1 )
χ(t) = P e+iΩ2 t
+ Qe with Ω2 ≡ ω2
I3 I1

while  
1
ω =  2  : 1 and 2 infinitesimal
ω3
leads to

−iΩ3 t (I3 − I1 )(I3 − I2 )
χ(t) = P e +iΩ3 t
+ Qe with Ω3 ≡ ω3
I1 I2

The point to notice is that

I1 > I2 > I3 ⇒ Ω1 and Ω3 are real, but Ω2 is imaginary

Instability of spin about the intermediate axis 53

The implication is that

• ω 0 , if initially nearly parallel to either the principal axis of greatest moment
or the principal axis of least moment, moves like a 2-dimensional oscillator,
tracing and retracing an ellipse: spin about either of those axes is stable;
• ω 0 , if initially nearly parallel to the principal axis of intermediate moment,
moves away from that neighborhood with exponential haste: spin about
the intermediate axis is unstable.
This analytical result conforms very nicely to the lesson latent in figures 10b,
10c & 10d (though those relate to how S 0 wanders in S -space, while we have
been working here in ω 0 -space).
If we write τi to denote the time it takes ω 0 to complete a circuit in the
immediate neighborhood of the ith principal axis, then we have

τ1 = 2π/Ω1 

τ2 = undefined (92)


τ3 = 2π/Ω3
where τ2 is “undefined” because ω 0 does not remain confined to neighbor-
hoods that contain the intermediate axis: it cyclically departs from the neigh-
borhood, visits its antipode, returns. Such excursions are actually remote tours
around either the axis of greatest moment or the axis of least moment. It
becomes of interest, therefore, to discover how to describe the the periods of such
“remote excursions.” The problem is addressed in §150a of Routh,16 who builds
upon Kirchhoff’s account of Jacobi’s discovery—anticipated by Euler—that the
general solution of the asymmetric top problem can be developed in terms of
elliptic functions. Here I must be content merely to discuss some properties
and implications of the results reported by Routh. Writing A > B > C in place
of I1 > I2 > I3 , we will consider T to be given/fixed, and S 2 to range on what
we discovered at (61) to be the physically allowed interval 2TC
S 2
2TA:

2TC S2 2TB 2TA

    
S 2 − 2TC 2TA − S 2
Circulation in the immediate neighborhood of the axis of greatest moment
requires that S 2 lie very near the right end of the blue region in the preceding
figure. At more remote blue points the circulation about that axis the period
is reported by Routh to be given by
  12 π
τ1 = 4 ABC  1 dφ (93)
(A − B)(S 2 − 2TC) 0 1 − k 2 sin2 φ
with
(B − C)(2TA − S 2 )
k2 =
(A − B)(S 2 − 2TC)

Mathematica recognizes the integral to be just the complete elliptic integral
EllipticK[k2 ]. If we construe τ1 to be a function of S 2 and, in order to learn
54 Gyrodynamics

the value assumed by τ1 (S 2 ) as S 2 ↑ 2TA, ask Mathematica to develop

τ1 (2TA − x2 ) as a power series in x, we obtain

τ1 (2TA − σ 2 ) = 2π ABC + (complicated term)σ 2 + · · ·
(A − B)(A − C)2T

which (since in leading order 2T = A ω12 ) can be written

= 2π/Ω1 + · · ·

with 
(A − B)(A − C)
Ω1 ≡ ω 1
BC
—in precise agreement with the result obtained by simpler means on page 52.
Formulæ appropriate to cases in which S 2 falls on the red interval in the
preceding figure (cases, that is to say, in which 2TC
S 2 < 2TB and ω 0
circulates about the principal axis of least moment) can be obtained from the
preceding formulae by A  C interchange.
Return now to (93) and set S 2 = 2TB to obtain
  1
2π
τ1 → τ 2 = 4 ABC  1 dφ (94)
(A − B)(B − C)2T 0 1 − p2 sin2 φ

with
(B − C)(A − B)
p2 = =1
(A − B)(B − C)

This result is invariant underA  C (we have therefore τ1 → τ2 ← τ3 , which

is gratifying), and—since the diverges—leads to the conclusion that τ2 = ∞ :
a top set spinning about the intermediate axis does not wander. But that
situation is, as we have recently established, unstable: if the alignment is not
absolutely precise the top does wander, and we have interest in computing the
period in such more realistic cases.
2
To signal our intent to approach Scritical = 2TB from above, we return to
(93)—which we agree now to notate

τ1 (S 2 ) = F1 (S 2 ) · EllipticK[ p21 (S 2 )]

—and set S 2 = 2TB + σ 2 . Mathematica supplies

  
ABC 1 1
2
F1 (2TB + σ ) = 4 1− σ2 + 3
8 [ T 2 (B−C)]2 σ − ···
4
2 T(A−B)(B−C) T(B−C)
A−C A−C
p21 (2TB + σ 2 ) = 1 − σ2 + σ4 − · · ·
(A−B)[2 T(B−C)] (A−B)[2 T(B−C)]2

≡ 1 − q (σ )
2 2
Instability of spin about the intermediate axis 55

Mathematical handbooks23 supply moreover the information that

EllipticK[ p2 ] ≡ EllipticK[ 1 − q 2 ]

can, for small values of q 2 , be developed

= Q + 14 (Q − 1)q 2 + 64
9
(Q − 76 )q 4 + 256
25
(Q − 37
30 )q + · · ·
6

where
Q ≡ log(4/q) : blows up logrithmically as q ↓ 0
2
The results now in hand could be used to compute τ1 (Scritical + σ 2 ), and by
A  C one could without labor obtain a description of τ3 (Scritical − σ 2 ). For
2

more detailed discussion see gyrodynamics (/), pages 139–144 or the

previously cited paper by W. G. Harter & C. C. Kim.17 But to obtain a good
qualitative understanding of the situation it is, I think, most instructive to
proceed not analytically but graphically: letting τ1 (S 2 ; A, B, C, T) denote the
expression that appears on the right side of (93), we define

τ1 (S 2 ; A, B, C, T) : 2TB < S 2
2TA
τ (S 2 ; A, B, C, T) ≡
τ1 (S 2 ; C, B, A, T) : 2TC
S 2 < 2TB

Suppose, for example, we set T = 1 and assign to the principal moments the
values A = 4, B = 83 and C = 2 that were used to construct the figures on
pages 32–34: then 2TA = 8, 2TB = 16 3 = 5.333, 2TC = 4 and Mathematica
constructs the graph of τ (S 2 ; 4, 16
3 , 2, 1) presented here as Figure 19.

80

60

40

20

5 6 7 8

Figure 19: Graph of τ (S 2 ) in a typical case. ω 0 circulates around

the axis of least moment on the left side of the spike, the axis of
2
greatest moment on the right. The spike is situated at Scritical .

23
See E. Jahnke & F. Emde, Tables of Functions (), page 73; J. Spanier
& K. B. Oldham, An Atlas of Functions (), page 612.
56 Gyrodynamics

r N

Figure 20: Diagram of the sort standard to elementary discussions

of the precession exhibited by spun toy tops. Not shown is the
horizontal component of the support force. We will take θ to be the
“angle of tilt,” and will use φ to describe “twirl about the vertical
space-axis.”

14. Symmetric tops with a fixed point. Astrophysical reality presents many
examples of semi-rigid bodies rotating semi-freely in intertial space, but we are
seldom inclined to call such objects “tops,” as has been my practice. The tops
of playroom experience derive much of their fascination from the fact that—
inevitably—they are spun in uniform gravitational fields, and are supported
. . . which is to say: they are not free.
Discussions of the precession of tops such as are found in introductory
textbooks24 standardly proceed from diagrams resembling Figure 20. Writing

F = F vertical + F horizontal

one claims (though it is generally untrue!) that F vertical + W = 0, and cleverly

circumvents the awkward fact that F horizontal is generally unknown by taking
the contact point to be the reference point with respect to which all torques
and angular momenta will be defined: with respect to that point F gives rise
24
See, for example, Paul A. Tipler, Physics for Scientists and Engineers
(3rd edition ), §8 -8; Richard Wolfson & Jay M. Pasachoff, Physics (),
pages 297–299.
Symmetric tops with a fixed point 57

x3
x03

x02

x2

ψ
φ x01
x1
line of nodes

Figure 21: Indication of the meanings of the variables employed

in a more careful treatment of the problem.

to no torque, and becomes effectively irrelevant. From N = r × W we obtain

 
− sin φ
N vertical = 0 and N horizontal = mgr sin θ
cos φ

L we learn that L vertical is conserved, while from

From N = L̇
 
cos φ
L horizontal = L sin θ
sin φ
we get  
− sin φ
L horizontal = Lφ̇ sin θ
L̇
cos φ
The implication appears to be that the top will precess with angular frequency
mg r
φ̇ =
L
The physical fact of the matter is, however, that tops exhibit much more
complicated kinds of motion that this simple theory leads one to anticipate!
A full-blown dynamical theory of tops must account both for the motion of
the center of mass and rotation about the center of mass, and might therefore
appear (Chasle’s theorem: page 3) to entail that we keep track of six variables.
58 Gyrodynamics

We require, however, that the inertial coordinates of one point—the point of

support—be fixed. That requirement imposes upon our six variables three
holonomic constraints. Our system has, therefore, only three degrees of freedom,
all of which refer to rotations about the support point. It becomes in this light
natural to take as generalized coordinates the Euler angles that (figure 21)
relate
• an inertial Cartesian space frame erected at the support point to
• a translated copy of the principal axis frame of the top.
This done, a slight modification of the argument that gave (71) gives

L = 12 B(φ̇2 sin2 θ + θ̇2 ) + 12 A(φ̇ cos θ + ψ̇)2 − mgh cos θ (95)

where m refers to the total mass of the top, where the final term on the right
is a potential enegy term that was absent from the theory of free rigid rotators,
and where it is to be understood that in the present instance25
A = (center of mass value)
B = (center of mass value) + mh2

In place of the equations of motion (72) we now have


pφ = B φ̇ sin2 θ + A(φ̇ cos θ + ψ̇) cos θ = constant 


pψ = A(φ̇ cos θ + ψ̇) = constant (96)


B θ̈ − B φ̇2 sin θ cos θ + A(φ̇ cos θ + ψ̇)φ̇ sin θ −mgh sin θ = 0 
  
ṗθ
where pφ , pψ and pθ are (angular) momenta conjugate to the angles φ, ψ and θ.
From the second of the preceding equations we get

pψ − Aφ̇ cos θ
ψ̇ =
A
which when brought to the first equation gives
pφ − pψ cos θ
φ̇ = (97.1)
B sin2 θ
whence
pψ pφ − pψ cos θ
ψ̇ = − cos θ (97.2)
A B sin2 θ

25
This follows directly from (14) if one makes the replacements

r1 → r1
r2 → r2
r3 → r3 + h

and uses (3) to eliminate terms of the form r hρ(ρρ) dr1 dr2 dr3 .
Symmetric tops with a fixed point 59

Note that if θ(t) were known then we could in principle use (97) to figure out
φ(t) and ψ(t).
From the design (95) of L it follows that total energy of the spinning top
is conserved:

E = 12 B(φ̇2 sin2 θ + θ̇2 ) + 12 A(φ̇ cos θ + ψ̇)2 + mgh cos θ = constant (98)

But 12 A(φ̇ cos θ + ψ̇)2 = 12 p2ψ /A was seen at (96) to be conserved all by itself, so
we have conservation of

E ≡ 12 B(φ̇2 sin2 θ + θ̇2 ) + mgh cos θ

which upon elimination of φ̇ becomes

(pφ − pψ cos θ)2

= 12 B θ̇2 + 2 + mgh cos θ (99)
 2B sin θ  
|—“effective potential,” call it V(θ)

At this point it becomes natural to mimic methods borrowed from the mechanics
of one-dimensional conservative systems, writing (for example)
 
dθ =
B E − V(θ)
2
dt
⇓
 θ


transit time θ  → θ  =   1 dϑ
θ
2
B E − V(ϑ)

5000

4000

3000

2000

1000

Π Π

2

Figure 22: Graph of the effective potential V(θ) in the case pφ = 5,

pψ = 3, 2B = 0.05 and mgh = 500. It is claimed not that the
numbers are (or are not) physically reasonable, only that the figure is
qualitatively typical.
60 Gyrodynamics

For many purposes it is, however, very useful to notice that (99) can, by a
change of variables θ → u ≡ cos θ, be brought to the form
u̇2 = (α − βu)(1 − u2 ) − (p − qu)2
= f (u) : cubic in u (100)
and to proceed under the presumption that physically self-consistent values
have been assigned to α ≡ 2E/B, β ≡ 2mgh/B, p ≡ pφ /B and q ≡ pψ /B.
From
f (u) ∼ βu3 for u large
from
f (±1) = −(p ∓ q)2 < 0
(we agree to exclude temporarily the exceptional cases p ∓ q = 0) and from the
fact that for our results to admit of physical interpretation it must be the case
that −π < θ < π (−1 < u < +1), we conclude that in physically realistic cases
f (u) must be of the form graphed in the following figure:

-1 1
-1

-2

-3

Figure 23: Graphical demonstration that the largest root of f (u)—

call it u3 —must necessarily be unphysical : u3 ≡ cos θ3 > 1 (unless,
exceptionally, u2 = u3 = 1). In all physically possible circumstances
the other roots (u1 and u2  u1 ) must lie on the interval [−1, +1 ].
From (100), i.e., from f (u) = u̇2  0, we learn that only u-values
on the interval u1
u
u2 refer to physical reality.

Mathematica is happy to provide explicit descriptions of u1 , u2 and u3 , but they

are, generally speaking, so complicated as to be worthless (except in concrete
cases where they can be presented as numbers). Mathematica is happy also to
supply
 u

transit time u → u = 1 dw
u
f (w)
= sum of incomplete elliptic functions
with complicated arcsine arguments
≡ t(u; u  )
Symmetric tops with a fixed point 61

from which it would not be feasible to extract u(t; u  ). We proceed therefore

qualitatively:
• the top spins under the control of ψ(t), the detailed motion of which is,
by (97.2), under the control of u(t):

p − qu
ψ̇ = (B/A)q − u
1 − u2

• as it spins it precesses under the control of φ(t), the detailed motion of

which is, by (97.1), again under the control of u(t):

p − qu
φ̇ = (101)
1 − u2

• as it spins and precesses it nutates under the control of θ(t), the detailed
motion of which is, by
θ = arccos u
again under the conrol of u(t), which oscillates back and forth—periodically
but non-sinusoidally—between the turning points u1 and u2 .
Looking to (101) we see that φ̇ vanishes at u = u
≡ p/q. If u
lies between
u1 and u2 then φ̇ reverses sign as u proceeds u1 → u2 and the symmetry axis
of the top traces a looping curve such as appears at the top of the following
figure. If—exceptionally—u
= u2 we get cusps (middle of the figure),26 while
if u
> u2 then φ̇ retains the same sign as u proceeds u1 → u2 and we get the
undulating curve shown at the bottom of the figure.
The elementary theory sketched on pages 56–57 provided no indication of
the nutation exhibited by real tops. We are in position now to recognize that
a top will display nutation-free precession if and only if the turning points u1
and u2 are coincident: u1 = u2 = u0 , where u0 marks the point at which the
effective potential

(p − qu)2
V (u) ≡ V(arccos u) = + βu
1 − u2

assumes its minimal value. To discover the value of u0 we construct

dV (u) 2u(p − qu)2 2q(p − qu)

= − +β
du (1 − u2 )2 1 − u2
2u(p − qu)2 − 2q(1 − u2 )(p − qu) + β(1 − u2 )2
=
(1 − u2 )2

and look for the root of the quartic numerator that lies on the physical interval
[−1, +1 ]. We note in this connection that the numerator is quadratic in (p−qu),
26
Though mathematically exceptional, such cusps are in fact observed if one
spins up a top and then “drops” it with (initially) φ̇ = θ̇ = 0.
62 Gyrodynamics

Figure 24: Typical patterns traced by the symmetry axis of a

nutating top as it precesses.
Symmetric tops with a fixed point 63

and that solving the quadratic gives

(1 − u20 ) !  "
p − qu0 = q ± q 2 − 2βu0
2u0

which by (101) becomes

!  "
φ̇0± = 1 q ± q 2 − 2βu0
2u0

We conclude that nutationless (or “steady”) precession can occur only if

q 2 − 2βu0  0, and that if q 2 − 2βu0 > 0 then such precession can be either
fast or slow . It follows, moreover, that if q 2 − 2βu0  0 then

q 2− 12 k − 18 k 2 − · · · 2βu 4Bmgh cos θ0
φ̇0± = · with k ≡ 2 0 = <1
2u0 2k + 8k + · · ·
1 1 2 q p2ψ

so in leading order
q pψ A
φ̇0+ = = = ω
u0 B cos θ0 B cos θ0 spin
β mgh mgh
φ̇0− = = =
2q pψ A ωspin

It is interesting that φ̇0− is (in leading order) θ0 -independent; i.e., that steady
precession can occur at any tilt. This may account for why it is that—according
to Thornton & Marion 22 that “it is the slower of the two [steady] precessional
velocities that is usually observed.” In any event (to rephrase in more physical
terms a point established just above), steady precession of neither sort can
occur unless 
4Bmgh cos θ0
ωspin >
A2

Much more could be said about the physics of toy tops, a subject which has
first charmed, then challenged, many of the greatest classical theorists. Here I
quote the 25 -year- old Maxwell, writing in :27
“To those who study the progress of exact science, the common
spinning-top is a symbol of the labours and the perplexities of men
who had successfully threaded the mazes of the planetary motions.
The mathematicians of the last age, searching through nature for
problems worthy of their analysis, found in this toy of their youth,
ample occupation for their highest mathematical powers. . . We find
Euler and D’Alembert devoting their talent and their patience to the
establishment of the laws of the rotation of solid bodies. Lagrange
has incorporated his own analysis of the problem with his general

27
See pages 246–262 in Volume I of W. D. Niven (editor), The Scieintific
Papers of James Clerk Maxwell ().
64 Gyrodynamics

treatment of mechanics, and since his time Poinsôt has brought the
subject under the power of a more searching analysis than that of the
calculus, in which ideas take the place of symbols, and
intelligible propositions supersede equations.” Maxwell continues
with a reference to “. . . the top which I have the honour to spin
before the Society. . . ”
But for further particulars and finer details I must refer my reader to §5 -7 in the
1st and 2nd editions ( and ) of Goldstein, and to additional references
cited there.

15. Nonparallelism of angular velocity and spin. It was remarked in passing

already on page 6 that while
• the linear momentum p and linear velocity v of a point particle stand in
the relation p = mvv , where m is a scalar,
• the intrinsic angular momentum (or spin) S and intrinsic angular velocity ω
of a rigid body stand in the relation S = I ω , where I is a symmetric matrix.
So while p and v are invariably parallel, S and ω are typically not parallel
but stand in an ever-shifting angular relationship, even in the total absence of
impressed torques. We have learned to attribute largely to this circumstance
the fact that the motion of free rigid bodies is so mucy more intricate than the
motion of free point particles. I propose to address this question: How great
can the angle S ∠ ω be?
We look first, by way of preparation, to a 2-dimensional model of the
3-dimensional issue. Let  
1
ω=
0
Let I range over the set I of all real symmetric 2 × 2 matrices with prescribed
eigenvalues A and B; i.e., let I be rotationally equivalent to the diagonal matrix
 
A 0
I0 =
0 B

and look to the set of vectors S = I ω that is generated as I ranges over I.

Typical elements of I can be described
 
cos θ sin θ
I(θ) = RT (θ)I 0 R(θ) with R(θ) ≡
− sin θ cos θ
 
A cos θ + B sin θ (A − B) cos θ sin θ
2 2
=
(A − B) cos θ sin θ B cos2 θ + A sin2 θ

Multiplication into ω gives

 
A cos2 θ + B sin2 θ
S (θ) =
(A − B) cos θ sin θ
Nonparallelism of angular velocity and spin 65

S (θ)

δ 2θ
ω B A

Figure 25: The vector S (θ) = I(θ) ω ranges—twice—around the

red circle as θ ranges on [0, 2π]. The circle intersects the ω -axis at
the eigenvalues of I(θ), so is centered at 12 (A + B) and has radius
2 (A − B). The double-valuedness of the construction reflects the
1

fact that the map I 0 → RT I 0 R is insensitive R → −R .

Writing

S1 = 12 A(cos2 θ + 1 − sin2 θ) + 12 B(sin2 θ + 1 − cos2 θ)

= 12 (A + B) + 12 (A − B) cos 2θ
2 (A − B) sin 2θ
1
S2 =

we have # $ # $2 # $2
A+B 2
S1 − 2 + S2 = A−B
2

The implication is that the vectors S (θ) all lie on the circle shown in the
preceding figure. On two occasions S (θ) and ω are parallel:

S (0) = A ω and S (π) = B ω

The greatest angular deviation is

! "
δ = arcsin A − B (102)
A+B
and occurs at the solution θmax of

2θ = 12 π + δ

Turning now to the 3-dimensional case, we set

   
0 A 0 0
ω = 0, I0 =  0 B 0
1 0 0 C
66 Gyrodynamics

and—assuming our rotation matrix R to have been presented in the Euler

representation R(φ, θ, ψ) that was spelled out at (42), and entrusting all
calculation to Mathematica—obtain a description of

I(φ, θ, ψ) = RT (φ, θ, ψ) I 0 R(φ, θ, ψ)

that when multiplied into ω gives

 
A sin θ sin ψ
S (φ, θ, ψ) = RT (φ, θ, ψ)  B sin θ cos ψ 
C cos θ
  
cos φ − sin φ 0 Σ1 (θ, ψ)
=  sin φ cos φ 0   Σ2 (θ, ψ) 
0 0 1 Σ3 (θ, ψ)
≡ R(φ)Σ
Σ (ψ, θ)

with
Σ1 ≡ 12 (A − B) sin 2ψ sin θ
Σ2 ≡ 12 [A sin2 ψ + B cos2 ψ − C ] sin 2θ
Σ3 ≡ (A sin2 ψ + B cos2 ψ) sin2 θ + C cos2 θ
Our assignment is to describe the Σ (ψ, θ)-vector, which R(φ) serves simply to
twirl about the 3-axis (the ω -axis), with these consequences:

S12 + S22 = Σ21 + Σ22

S 3 = Σ3

We begin with the observations that Σ3 is manifestly non-negative, and

can be written

Σ3 = [A sin2 ψ + B cos2 ψ − C ] sin2 θ + C

We have had occasion to note the identity

A sin2 ψ + B cos2 ψ = 12 (A + B) − 12 (A − B) cos 2ψ

so with the introduction of a ≡ A − C and b ≡ B − C we have

A sin2 ψ + B cos2 ψ − C = 12 (a + b) − 12 (a − b) cos 2ψ

giving
2 (a − b) sin 2ψ] sin θ
1
Σ1 = [
Σ2 = [ 2 (a + b) − 2 (a − b) cos 2ψ] sin θ cos θ
1 1

S3 = Σ3 = [ 12 (a + b) − 12 (a − b) cos 2ψ] sin θ sin θ + C

Nonparallelism of angular velocity and spin 67

Proceeding on the assumption that

A>B>C

we note it to be an easy implication of results now in hand that

A
S3
C

Next we notice that

S12 + S22 + (S3 − C)2 = 12 [a2 + b2 − (a2 − b2 ) cos 2ψ] sin2 θ

(a2 + b2 ) − (a2 − b2 ) cos 2ψ
= · (S3 − C)
(a + b) − (a − b) cos 2ψ
≡ 2k · (S3 − C)
(a2 + b2 ) − (a2 − b2 ) cos 2ψ
k = k(ψ) ≡ 12
(a + b) − (a − b) cos 2ψ
1 a2 sin2 ψ + b2 cos2 ψ
= >0 (103)
2
a sin2 ψ + b cos2 ψ

can be written
S12 + S22 + (S3 − C − k)2 = k 2 (104)
From
dk(ψ) (A − B)(A − C)(B − C) sin 2ψ
=
dψ (a sin2 ψ + b cos2 ψ)2
we learn that

kmin = k(0) = 12 (A − C) and kmax = k( 12 π) = 12 (B − C)

It is on the basis of (104) that I have constructed Figure 26, and from the
figure we discover that half the story remains untold , for the figure appears to
indicate that I possesses (in addition to the isolated eigenvalue C ) a continuum
of eigenvalues lying between B and A. The point to notice is that we obtained
(104) by eliminating θ between the following two equations:
 2  a+b a−b 2 
S12 + S22 = a−b
4 sin 2ψ + 4 − 4 cos 2ψ cos 2
θ sin2 θ
 a+b a−b 2 2
S3 = 4 − 4 cos 2ψ sin θ + C

If, on the other hand, we had undertaken to eliminate ψ we would have obtained

S12 + S22 + (S3 − A+B 2

2 ) = 2 (105)

with  
a+b 2
 = (θ) ≡ 2 − ab sin2 θ
 2
= A+B
2 − C − (A − C)(B − C) sin2 θ
68 Gyrodynamics

Figure 26: Cross section of a family of nested spheres obtained

from (104), with A = 4, B = 3, C = 1. The angular parameter ψ
π
has been stepped through the values n 20 (n = 0, 1, 2, . . . , 10) and is
constant on each sphere.

and where it is evident that

min = ( 12 π) = 12 (A − B) and max = (0) = 12 (A + B) − C

= 12 (A − B) + (B − C)

From (105) we are led to Figure 27, which again tells only half—the
other half—of the story. It is only by conflating those figures—by taking (104)
and (105) in combination—that we obtain a description of the set of points to
which parameters φ, θ and ψ can be simultaneously assigned, a description of
the curious region to which S (φ, θ, ψ) is necessarily confined.
The ω -axis punctures that “crescent of revolution” at only three points,
and those mark the eigenvalues of I.
It is evident from Figure 28 that
 
1
[(greatest eigenvalue) − (least eigenvalue) ]
δ = arcsin 12 (106)
2 [(greatest eigenvalue) + (least eigenvalue) ]

which serves very nicely as a generalization of (102). I invite my reader to

consider the limiting cases B ↑ A, B ↓ C and A = B = C.
Nonparallelism of angular velocity and spin 69

Figure 27: Cross section of a family of nested spheres obtained

from (105), with A = 4, B = 3, C = 1. The angular parameter θ
π
has been stepped through the values n 20 (n = 0, 1, 2, . . . , 10) and is
constant on each sphere.

Figure 28: Superimposed figures. S (φ, θ, ψ) lies necessarily in the

crosshatched region, which has the form of a sphere with two interior
spherical exclusions
70 Gyrodynamics

I originally developed this material to resolve some procedural problems

that arose in connection with some experiments I was performing () with
air-supported gyros of various designs. Only later did I realize that the
mathematics has nothing specific to do with the relationship of spin to angular
velocity: it pertains simultaneously to all statements of a form

y = Mx
x : M a real 3 × 3 symmetric matrix

that is encountered very commonly in physics, especially in linearized

phenomenological theories of all sorts.28 It is therefore not surprising that
the essentials—at least the 2-dimensional essentials—of the material have been
reinvented many times by many people. The first occurance of my Figure 25
appears to have been in a publication of Christian Otto Mohr (), who
had himself built upon a suggestion of Karl Culmann (). Culmann and
Mohr were concerned not with the dynamics of tops but with stress analysis
and the fracture of brittle materials.29 Some variants and generalizations
of “Mohr’s construction” are discussed in my “Non-standard applications of
Mohr’s construction” ().

16. Theory of celts. Footballs, hardboiled eggs, tippy tops . . . all behave in
counterintuitive ways when spun, and each has generated a literature.30 Here I
propose to discuss only one of those curiosities. The story begins in the British
Museum, where one day in the s the physicist G. T. Walker had reason to
examine that museum’s collection of “celts”—smooth axhead-like stones found
in abundance at paleolithic sites all over Europe and the British Isles—and
chanced to notice that many of them, while they spun easily in one direction,
first wobbled and then reversed course when spun in the opposite direction.

28
Such an equation relates stress to strain in elastic media, polarization to
electric field strength in dielectric media, etc.: the list could be very greatly
extended.
29
Culmann (–) was a German professor of civil engineering who
is remembered today mainly for his contributions—some of which had been
anticipated by Maxwell—to “graphical statics.” Mohr (–) taught civil
engineering first in Stuttgart and then (from  until his retirement in ) in
Dresden. He was said by his student A. Föppl (who himself figures importantly
in the history of electrodynamics, and whose texts influenced the development
of the young Einstein) to have been an outstanding teacher: a tall, proud and
taciturn man who spoke and wrote with simplicity, clarity and conciseness.
“Mohr’s stress circle”—the contribution for which he is today remembered—
provided the basis for his theory of stress failure; for an account of something
called the “Coulomb-Mohr fracture criterion” see (for example) C. C. Mei,
Mathematical Analysis in Engineering (), p. 150. For an electrodynamical
application of Mohr’s idea—having nothing at all to do either with tops or with
fracture—see my classical electrodynamnics (), p. 127.
30
A fairly extensive bibliography—which is, however, by no means complete
—begins at page 146 in gyrodynamics (/).
Theory of celts 71

Walker soon demonstrated31 that this odd behavior could be attributed to

the circumstance that the geometrical axes defined by the (approximately)
ellipsoidal base of the object and its principal axes are misaligned. Walker’s
lectures and demonstrations were witnessed by Arnold Sommerfeld (Trinity
College, Cambridge, ), who later recalled the powerful impression they
made upon him, and who gave brief attention to the subject on pages 149–150
of his Lectures on Theoretical Physics : Volume 1. Mechanics (). I myself
learned of the “celt phenomenon” from §2.72 of J. Walker’s Flying Circus of
Physics: I wrote up a modernized version of G. T. Walker’s original theory,
and—taking as my model the jade celt that Frank Oppenheimer one day pulled
from his desk drawer and showed me—fashionedd from Brazilian rosewood
what has become known locally as “Wheeler’s banana top.” My work came
to the attention of J. Walker, and is mentioned in the “Amateur Scientist”
column (of which Walker was then the editor) in the October  issue of
Scientific American. It was J. Walker who on that occasion attached the name
“rattleback” to these objects, and it is under that head that one should approach
Google for recent references.
The Scientific American article generated a flood of correspondence—much
of it goofy, some of it not—that continues to this day, 25 years later. It was
one of my correspondents who directed my attention to a then-recent article
by Sir Herman Bondi.32 Bondi’s objective was to write an improved and more
complete revision of G. T. Walker’s paper. Close study of Bondi’s paper leaves
me unconvinced, however, that he achieved his objective. Here I have taken as
my source the account of Walker’s paper that appears in Chapter 17, §§1–3 of
A.Gray’s A Treatise on Gyrostatics and Rotational Motion ().

design considerations The convex surface of our resting celt, in the

immediate vicinity of its support point, can in leading approximation be
considered to be ellipsoidal. To deal most simply with that fact we will consider
the celt to be ellipsoidal not just locally but globally; i.e., to have a surface of
which
2
2
2
x + y + z =1 (107)
a b c
provides an implicit description. We will assume a > b > c : the point of static
support resides then (see Figure 29) at (0, 0, −c). Stability requires that

least radius of curvature at support point > c

We are led to look at the bottom of the x = 0 cross-section of the celt, where

z(y) = −c 1 − (y/b)2

31
Quarterly Journal of Pure & Applied Mathematics 28 (1896), pages 175–
184.
32
“The rigid body dynamics of unidirectional spin,” Proceedings of the Royal
Society (London) 405A, 265 (1986).
72 Gyrodynamics

Figure 29: Assumed shape of the celt. The coordinate system—

with origin at the geometric center of the ellipsoid, and aligned in
coincidence with its symmetry axes—will eventually (after the top
has been properly “loaded”) be abandoned in favor of a coordinate
system that diagonalizes the moment of inertia matrix.

The curvature of such a plain curve can be described33

K ≡ d (slope) =  1 d arctan dz = z 
3
ds 
1 + (z ) dy
2 dy [1 + (z  )2 ] 2
which in the instance at hand supplies

K(0, y) = complicated expression (ask Mathematica)

↓
K(0, 0) = c/b2
= 1
least radius of curvature
so to achieve stability-at-rest we have only to require that b > c, and this we
have in fact already done.
It is our intention to load the ellipsoid in such a way as to cause the
horizontal principal axes to be slightly misaligned with respect to their
geometrical counterparts (see Figure 30). We confront therefore this small
mathematical problem: How to describe an ellipsoid that has been thus slewed
with respect to its principal axes? The answer, as will emerge, lies already at

33
Here s denotes arc length: ds = 1 + (z  )2 dy.
Theory of celts 73

Figure 30: The formerly homogeneous celt has now been “loaded”
in such a way as to preserve the location of the center of mass, and
to preserve also the z-axis as a principal axis, but to slew the other
principal axes with respect to the associated geometrical axes.

hand. Let (107) be notated

px2 + qy 2 = k 2

with p ≡ 1/a2 , q ≡ 1/b2 , k 2 ≡ 1 − (z/c)2 .34 Or again

 T  
x x p 0
M0 = k 2 with M 0 ≡
y y 0 q

To rotate that ellipse through an angle ψ we have—as was established already

on page 64—only to make the replacement

p cos2 ψ + q sin2 ψ (p − q) cos ψ sin ψ
M 0 −→ M(ψ) ≡ (108)
(p − q) cos ψ sin ψ q cos2 ψ + p sin2 ψ

—the effect of which is illustrated in Figure 31.35 Assuming ψ to have been

prescribed/fixed, we will write

P R
=
R Q

34
Note that a > b entails p < q.
35
Mathematica confirms that the eigenvalues of M(ψ) are {p, q} and that
det M(ψ) = pq for all values of ψ.
74 Gyrodynamics

0.5

-1 -0.5 0.5 1

-0.5

-1

Figure 31: Graph of the ellipse x2 + 4y 2 = 1 and (in red ) the

result of rotation through angle ψ = π/6, constructed with the aid
of (108).

in which notation (107) becomes

2
P x2 + 2Rxy + Qy 2 + z = 1
c

To say the same thing another way, we have


z=c 1 − (P x2 + 2Rxy + Qy 2 )

which in the near neighborhood of the point of static support (i.e., for small x
and y) becomes
 
= − c 1 − 12 (P x2 + 2Rxy + Qy 2 ) − · · · (109)

This equation will serve to describe—relative to the principal axes—all relevant

aspects of the shape of the celt. It is the presence of the R-term, which is under
the control of the angular parameter ψ, that accounts for the chirality of celts.

Importance will attach in the dynamical theory to the unit normal vector
at the point to contact (see Figure 32). To obtain a description of that vector,
Theory of celts 75

Figure 32: Celt in a typical non-equilibrium position. Shown in

blue is the unit vector that is normal to the celtic surface (and
normal also to the support surface) at the point of contact. Shown
in red is the space frame, the origin of which rides irrotationally
with the center of mass.

introduce the “scalar field”

ϕ(x, y, z) ≡ 12 c (P x2 + 2Rxy + Qy 2 ) − z

and construe (109) to describe a “surface of constant ϕ ”:

ϕ(x, y, z) = c

We then have  
c(P x + Ry)
∇ϕ =  c(Rx + Qy) 
−1
giving

∇ϕ
n=  
∇ϕ|
|∇
c(P x + Ry)
= 1  c(Rx + Qy) 
1 + c (P x + Ry)2 + c 2 (Rx + Qy)2
2
−1

This equation describes—relative to the body frame—the unit normal at the

point of contact, but only if x and y are assigned the values that describe the
instantaneous point of contact. In the near neighborhood of the origin (point
of resting contact) we have Hoyle’s
 
c(P x + Ry)
n= c(Rx + Qy)  + ···
−1 + 12 [c 2 (P x + Ry)2 + c 2 (Rx + Qy)2 ]

where the abandoned terms are of higher than second order in x and y.
76 Gyrodynamics

If, however, we elect with Walker/Gray to work in first order 36 we have

 
c(P x + Ry)
n0 =  c(Rx + Qy)  + · · · (110)
−1

In that same leading approximation the vector that extends from the center of
mass to the instantaneous contact point becomes
 
x
r0 =  y  (111)
−c

from which it follows that n0 can be described

 
cP cR 0
n0 =  cR cQ 0  r 0 (112.1)
0 0 c –1

In the preceding equations I have installed 0 s to emphasize that they refer to

the body frame.

equations of motion Looking first to the motion of the center of mass,

Newton’s 2nd law supplies
m v̇v = mg n + f
where f refers to the net force (support and friction) exerted on the celt by
the table. Here all vectors are relative to the inertial “table frame,” of which
the “space frame” (origin riding on the center of mass) is a non-inertial trans-
lated copy. To express inertial vectors in terms of body-frame vectors we write
v = R v 0 , f = R f 0 , n = R n0 , etc. The time -dependence of R entails

v̇v = R v̇v 0 + Ṙ R–1 v

= R v̇v 0 + ω × v
= R (v̇v 0 + ω 0 × v 0 )

so we have
m(v̇v 0 + ω 0 × v 0 ) = mg n0 + f 0
The motion of ω 0 is described by Euler’s equation

ω 0 + ω 0 × I 0 ω 0 = r 0 ×ff 0
I 0ω̇

36
It is principally because Hoyle works in 2nd order that his equations are
relatively so complicated, and the significance of his results so hard to grasp
intuitively.
Theory of celts 77

which, if we use Newton’s law to eliminate reference to the presently unknown

force f 0 , becomes
 
ω 0 + ω 0 × I 0ω 0 − m r 0 × v̇v 0 + ω 0 × v 0 − g n0 = 0
I 0ω̇ (112.2)

Here  
A 0 0
I0 =  0 B 0 (112.3)
0 0 C
n = 0 because the table is (by assumption) flat, giving
Additionally we have ṅ

n0 + ω 0 × n 0 = 0
ṅ (112.4)

And finally, because we assume the celt rolls without slipping, we have

v0 + ω 0 × r0 = 0 (112.5)

Equations (112) provide the physical basis of Walker’s theory (also of Hoyle’s).
In working out the consequences of these equations let us now agree, as a matter
of typographic convenience, to drop the 0 s.
a × b)× c = (cc · a)bb − (cc · b)a
Cross (112.4) into the unit vector n and use (a a
n ×n
to obtain ṅ n· ω )n
n + (n n − (n
n· n)ω
ω = 0 or

n ×n
ω = ṅ n + ω n with ω ≡ (n
n· ω ) (113.1)
= magnitude of ω

Returning with this result to (112.5) we have

v = r × (ṅ
n ×n
n + ω n) (113.2)

Equations (113)—taken together with their time-partials37

ω = n̈
ω̇ n ×n n + ω ṅ
n
v̇v = ṙr × (ṅ
n ×n
n + ω n) + r × (n̈
n ×n
n + ω ṅ
n)

—can be used to turn (112.2) into an equation involving only n, r and their
derivatives. I postpone that substitutional exercise.
The unit normal n, though fixed with respect to the table frame, moves
relative to the body frame (in which we are now working), and Walker takes
that apparent motion to be the indicator of what the celt is doing. Hoyle, on
the other hand, elects to watch the motion of r (i.e., of x and y, which refer
37
Here we accept Walker’s intuition-based assertion that ω will be constant
in leading order, that its temporal variation will be a higher-order effect. It
was in an effort to avoid such ad noc assertions that Hoyle worked in second
order.
78 Gyrodynamics

to the instantaneous location of the contact point). Methodologically it is six

one way, half a dozen the other: I will follow Walker’s lead, which calls for
elimination of all r and ṙr terms. To that end we return to (112.1), which
supplies
 
cQ −cR 0
r= 2 1  −cR cP 0 n
c (P Q − R2 )
0 0 c3 (P Q − R2 )
 
α ρ 0
≡  ρ β 0n
0 0 c
≡ Jn (113.3)

ṙr = J ṅ
n

With Walker, we proceed now in the assumption that in first order38

     
4 n1 4 ṅ1 4 n̈1
n =  4 n2  whence ṅ n =  4 ṅ2  , n̈
n =  4 n̈2 
−1 0 0

It then follows by (113.1) that

   
4( − ṅ2 + ω n1 ) 4( − n̈2 + ω ṅ1 )
ω =  4( + ṅ1 + ω n2 )  , ω̇
ω =  4( + n̈1 + ω ṅ2 ) 
−ω 0

and from (113.3) that

   
4(α n1 + ρ n2 ) 4(α ṅ1 + ρ ṅ2 )
r =  4( ρ n1 + β n2 )  , ṙr =  4( ρ ṅ1 + β ṅ2 ) 
−c 0

Bringing this informtion to (113.2) we find

 
4(c ṅ1 + ω[c n2 − β n2 − ρ n1 ])
v =  4(c ṅ2 − ω[c n1 − α n1 − ρ n2 ]) 
0
 
4(c n̈1 + ω[c ṅ2 − β ṅ2 − ρ ṅ1 ])
v̇v =  4(c n̈2 − ω[c ṅ1 − α ṅ1 − ρ ṅ2 ]) 
0

38
The 4 factors have been introduced to identify terms that we imagine to
be “small,” and to provide Mathematica with means to identify and discard
second order terms as they arise: at the end of the day we will set 4 = 1.
Theory of celts 79

Returning with this information to (112.2) we are led—after abandoning

terms of O[42 ]—to three equations, the third of which reads 0 = 0 and the first
pair of which (after we reverse their order and change a sign) read

0 = + n̈1 [B + mc 2 ]
+ ṅ1 [−mcρ ω]
+ ṅ2 [+(A + B − C)ω + 2mc 2 ω − mc β ω]
+ n1 [(−A + C − mc 2 + mc α)ω 2 − mg(c − α)]
+ n2 [mρ(g + cω 2 )]

0 = + n̈2 [A + mc 2 ]
+ ṅ1 [−(A + B − C)ω − 2mc 2 ω + mc α ω]
+ ṅ2 [+mcρ ω]
+ n1 [mρ(g + cω 2 )]
+ n2 [(−B + C − mc 2 + mcβ)ω 2 − mg(c − β)]

We have here a pair of coupled linear equations that can be written

M n̈
n + (S + A) ṅ
n + Kn = O (114)

where
 
B+mc 2 0 

M= 

0 A+mc 2 

 



− mc ρ ω 
1
2 mc (α−β) ω
S= 1 

2 mc (α−β) ω + mc ρ ω
 (115)
0 +[A+B−C+2mc2 − 12 mc(α+β)] ω 

A= 

−[A+B−C+2mc2 − 12 mc(α+β)] ω 0 

 


(−A+C−mc 2 +mc α) ω 2 −mg(c−α) mρ(g+c ω 2 ) 

K= 

mρ(g+c ω )2
(−B+C−mc 2 +mcβ) ω 2 −mg(c−β)

and where n is understood now to be the 2-vector


n1
n≡
n2

The matrices (115) are assembled from

• physical parameters m and g;
• principal moments A, B and C;
• parameters a, b and c that set the shape of the celt (particularly of its
foot); and
• an angle θ that describes the misalignment of the symmetry and principal
axis systems.
80 Gyrodynamics

The later four parameters are hidden in the designs of α, β and ρ. In the absence
of misalignment we have ρ = 0. The matrices M, S and K are symmetric; A is
antisymmetric.
Notice that reversing the sign of what Hoyle calls the “spin” (ω → −ω)
sends
M n̈
n + (S + A) ṅ
n + Kn = O
↓
M n̈
n − (S + A) ṅ
n + Kn = O
Which is to say: Celts spun  or  are described by distinct equations of motion,
already in first- order theory. The chirality of celt dynamics is thus made
immediately apparent.
The Walker/Gray argument becomes at this point a cleverly executed
exercise in stability theory. To expose the elegance of their idea without the
distraction of notational clutter, let us write
 
m1 0 

M= 

0 m2 
 



−ρσ s 

S = 

s ρσ
 (116)
0 a 

A = 

−a 0 

 



k1 ρκ 

K= 
ρκ k2

to abbreviate the structural essentials of (115). We look for solutions of the

form
n (t) = ν ei Ωt (117)
From (114) we obtain

[−Ω2 M + iΩ( S + A) + K ] ν = 0 (118)

which entails

0 = det[−Ω2 M + iΩ( S + A) + K ]

= (k1 k2 − ρ2 κ2 ) + iρ[σ(k1 − k2 ) − 2sκ ]Ω

− (m1 k2 + m2 k1 + a2 − s2 − ρ2 σ 2 )Ω2
− iρ σ(m1 − m2 )Ω3
+ m 1 m2 Ω 4 according to Mathematica

≡ (K − ρ2 κ2 ) − (µ − ρ2 s2 )Ω2 + M Ω2
 
+ ρ i(σ · ∆k − 2sκ)Ω − iσ · ∆m Ω3

≡ f (Ω) + ρ · g(Ω) + ρ2 · h(Ω) (119)

Theory of celts 81

Quartics are awkward. Note, however, that at ρ = 0; i.e., in the absence of

misalignment, the preceding equation becomes quadratic in Ω2 : it becomes
f (Ω) ≡ K − µΩ2 + M Ω4 = 0
and supplies 
µ2 − 4M K
µ±
Ω20 =
2M
We conclude—since M = (A + mc 2 )(B + mc 2 ) > 0 in all cases—that Ω0 will
be real only in those parts of parameter space where
µ2 − 4M K
0 and µ>0 (120)
When those conditions are satisfied we have
n(t) = ν fast cos(Ωfast t) + ν slow cos(Ωslow t + δ) (121)
where    12
µ+ µ2 − 4M K
Ωfast ≡
2M
   12
µ− µ2 − 4M K
Ωslow ≡
2M
and where ν fast , ν fast are associated solutions of the homogeneous system (118),
our present assumption being that ρ has been set to xero.39 Equation (121)
shows the small-amplitude motion of n to trace a (generally aperiodic) “skew-
Lissajous figure”:
1.5

-2 2

-1.5

Figure 33: “Skew-Lissajous figure” generated by (121) in the case

 
1.0 −0.6
Ωfast = 1.00, Ωslow = 0.68, ν fast = , ν slow =
0.8 0.5

39
It is tempting but would be incorrect to call the Ω’s “eigenvalues”—though
they are, like eigenvalues, roots of a polynomial—and it would for that same
reason be incorrect to call the ν ’s—which are in general not orthogonal—
“eigenvectors.”
82 Gyrodynamics

But setting ρ = 0 destroys the phenomenon  of interest,

 for on looking
back again to (115) we see that the parameters µ, M, K that enter into the
construction of Ω0 are all invariant under ω → −ω.
At this point it is Walker/Gray’s resourceful idea to assume ρ to be
non-zero but small , and to pass from quadratic to quintic by the methods of
perturbation theory.
MATHEMATICAL DIGRESSION: A toy perturbation theory. Let x0 be
a zero of f (x). What values should be assigned to {x1 , x2 , . . .} to
make x0 + 4x1 + 42 x2 + · · · a zero of the perturbed function
F (x) ≡ f (x) + 4g(x) + 42 h(x) + · · ·
Asking Mathematica to expand F (x0 + 4x1 + 42 x2 + · · ·), we are
led to
x1 f  (x0 ) + g(x0 ) = 0
x2 f  (x0 ) + x1 g  (x0 ) + 12 x21 f  (x0 ) + h(x0 ) = 0
..
.
which can be solved recursively. In leading order we have
g(x0 )
x1 = − (122)
f  (x0 )

Taking our definitions of f (Ω) and g(Ω) from (119)

f (Ω) = K − µΩ2 + M Ω4
g(Ω) = i(σ · ∆k − 2sκ)Ω − iσ · ∆m Ω3
we have first-order interest in the roots Ω = Ω0 + ρ·Ω1 of f (Ω) + ρ·g(Ω), which
according to (122) are given by
± Ωfast + ρ · iΓfast

σ[∆m(µ + µ2 − 4KM ) − 2M ∆k] + s[4κM ]
Γfast = + 
4M µ2 − 4KM
and
± Ωslow + ρ · iΓslow

σ[∆m(µ − µ2 − 4KM ) − 2M ∆k] + s[4κM ]
Γslow = − 
4M µ2 − 4KM
where
∆m ≡ m1 − m2
=B−A
and
∆ k = k1 − k2
=B−A
are, by our assumptions, positive.
Theory of celts 83

In first order, (121) has become40

n(t) = ν fast e−Γ fast t cos(Ω fast t)

(123)
+ν slowe−Γslow t cos(Ω slow t + δ)

Go to a point in parameter space where Γfast and Γslow are both positive. Both
of the exponentials in (123) then diminish as t increases. Such motion is stable.
Notice now that
s ≡ ω · 12 mc(α − β)
σ = ω · ρ mc
—which enter as factors into the construction of both Γ ’s—reverse sign when
the celt is spun in the opposite direction: ω → − ω. That sign reversal is
passed on to the Γ ’s, with the consequence that the exponentials blow up: the
purported motion (123) has been revealed to be unstable: the small-amplitude
theory has declared itself irrelevant to the subsequent physics.
To pursue that physics one would have to wrestle with the equations of
motion in their full non-linear complexity (which, so far as I am aware, has
never been attempted). The most variable to watch (because most informative)
would be not n but ω , for which one expects to obtain motion ω (t) of the form
     
small large oscillations small
 small  −→  large oscillations  −→  small 
large small large, but of opposite sign

Hoyle has remarked in this connection that so far as concerns the spin of the
celt (angular momentum with respect to its center of mass) we have (in the
space frame)
d
dt S = r ×f
f
d
= mrr × [ dt v − gn]
from which it follows in particular that
d
d
dt (normal component of S ) = n · dt S
n·(rr ×
= mn d
dt v )
= −m dt
d
v ·(n
n × r)

Evidently the normal component of S is constant for spinning objects that

sit always on their bottoms (rr  n). Hoyle remarks that even in cases where
r and n are not parallel, the normal component of spin typically “fluctuates
imperceptibly” unless the dtd
n × r) terms act in concert. . . which celts
v and (n
are designed specifically to achieve.
40
We have no present interest in the first-order adjustments experienced
by the vectors ν fast and ν slow (which, by the way, are defined only to within
multiplicative constants).
84 Gyrodynamics

The stable solutions of (123) die exponentially, which is to say: they lose
energy—an effect we normally attribute to dissipation. Odd in the present
instance, since no dissipation mechanism was built into the model . The
resolution of this little paradox must lie in the circumstance that energy leaks
into higher-order aspects of the motion which the first- order theory is powerless
to take into account. . . just as the first- order theory is powerless to temper the
seeming “explosions” in the unstable case.41

I draw attention finally to the fact that it is the S -term in (115) that
accounts for celtic chirality. One has
   
0 s 0 a
lim S = : resembles A=
ρ↓0 s 0 −a 0

where both s and a depend linearly upon ω. But in the 0th -order theory one
encounters only s2 and a2 , both of which are insensitive to the sign of ω. In
first order one encounteres s (also σ) and a2 : A is again blameless.

17. Deformable bodies. Reasonably good approximations to rigid bodies can

be found in toy rooms, on ball fields, on battle fields (among the small hunks
of matter hurled back and forth), in the astroid belt. But the earth is “rigid”
only in zeroth approximation: the circulation of atmosphere, oceans—even, on
a longer time scale, of continents—has an observable effect on its rotational
dynamics. Similar effects afflict almost all astrophysical bodies, and pertain
with especial importance to systems of interacting deformable bodies: it is
because the moon is deformable that it has come to present always the same
face to earth. Even an isolated body, made of material of finite strength, will
deform (expand at the equator) in response to its own rotation. Acrobats and
divers are deformation virtuosi: they tour many points in “shape space” with
rapid precision, and can never be accused of suffering from rigor mortis. We
evidently stand in need of a rotational dynamics of deformable bodies. But
confront at the outset several perplexing questions:

Once we abandon the concept of “rigidity,” what is left? Cannot every

isolated many-body system be considered to be a“deformable body”? How are
we to give physical meaning to our intuitive sense that some bodies are “almost
rigitd”? Can it be of dynamical relevance that the deformable bodies that
come most naturally to mind possess (in leading approximation) a well-defined
and shape -independent fixed volume? How many degrees of freedom has a
deformable body? The number appears to be indefinite: a swarm of N bees
has 3N degrees of freedom, but the system comprised of a bug walking around
on a rigid sphere has only eight (six for the sphere, two for the bug). The great
simplification brought to rigid body dynamics by Chasle’s theorem is clearly no
41
Though Hoyle works in second order, he does not escape the paradox:
escape would appear to require that one work not in truncated order, but
exactly.
Deformable bodies 85

longer operative. Finally, how—if at all—is one to attach a “body frame” to a

deformable body?42
We begin by looking to one respect in which the rotational physics of
deformable bodies differs profoundly from that of rigid bodies. Consider the
mechanism shown in Figures 35. No external forces/torques are impressed
upon the device, but an internal energy supply enables it to
• flex its elbows
• twist at the waist
We have
ϕ1 − ϕ2 = α whence ϕ̇1 − ϕ̇2 = α̇ (124.1)
while by angular momentum conservation43

I1 ϕ̇1 + I2 ϕ̇2 = 0 (124.2)

It follows that
 
I2   I2
ϕ̇1 = + α̇   dϕ1 = + dα
I1 + I2 I1 + I2
⇐⇒ (125)
I1 
 
 I1
ϕ̇2 = − α̇ dϕ2 = − dα
I1 + I2 I1 + I2

from which (124) are readily recovered as corollaries. Suppose now that the
device has been programed so as to cause α to increase/decrease periodically,
and to synchronously flex its elbows in such a way as to make

large when α̇ > 0 : “ inhaling”
J1 ≡ I2 /(I1 + I2 )
small when α̇ < 0 : “exhaling”

J1

α
42
This last problem has been addressed in a profound way by A. Shapere &
F. Wilczek in “Gauge kinematics of deformable bodies,” AJP 57, 514 (1989).
The paper appears also as §8.3 in Geometric Phases in Physics (), which
they edited, and provides the basis of the discussion which begins on page 89
below.
43
Actually, spin conservation. Without essential loss of generality we will
assume that initially—and therefore for all time, in the continued absence of
impressed torques—S S = 0.
86 Gyrodynamics

Figure 35A: Side views of the automaton discussed in the text,

showing (above) the elbows as they might be flexed when the device
is “inhaling” (α̇ > 0) and (below) as they might be flexed when the
device is “exhaling” (α̇ < 0).
Deformable bodies 87

ϕ1

ϕ2

Figure 35B: View from above of the automaton when configured

as shown at the top of the preceding figure. All angles are to be
assigned the same counterclockwise sense. By repeatedly performing
its carefully choreographed routine the device manages to rotate

without standing ever in violation of the condition S = 0 .
It becomes immediately evident from the diagram at the bottom of page 85 that
the device achieves a net ϕ1 -advance per α -cycle (and an identical ϕ2 -advance)
that can be described 
∆ϕ1 = J1 (α) dα (126)
It is clear also that ∆ϕ1 is independent of all temporal aspects of the cycle: in
that respect the striking phenomenon here at issue is (as Shapere & Wilczek
have emphasized) not so much “physical” as “geometrical .”
Suppose, for example, that the elbows of our automaton flex in such a way
as to achieve
I1 (t) = a − b cos ωt
(127.1)
I2 (t) = a + b cos ωt
and that the breathing of α can be described (see Figure 36)
α(t) = 12 αmax (1 + sin ωt) (127.2)
We then—by (125)—have
 t
ϕ(t) = a + b cos ωt α
max ω cos ωt dt
4a
0
 
α
= max 12 bω t + a sin ωt + b sin 2ωt
4a
b
= 8a αmax · ωt + oscillatory term
88 Gyrodynamics

2Π 4Π

Figure 36: The green curve (short dashes) describes the motion of
I1 , the solid red curve the motion of I2 , the blue curve (long dashes)
the motion of α. Notice that I2 is dominant when α is increasing,
I1 is dominant when α is decreasing. In constructing the figure I
have set a = 2, b = ω = αmax = 1.

2Π 4Π 6Π

Figure 37: Graph of the resulting motion of ϕ1 . The horizontal

lines indicate the ϕ1 -advance/period, which in the present instance
is ∆ϕ1 = π/8. The finer blue curve traces the advance of ϕ2 , which
for obvious reasons must achieve the same advance per period.

which yields a ϕ1 -advance per period (equal necessarily to the ϕ2 -advance per
period) given by
π
∆ϕ = 4 (b/a)αmax

Notice that t enters into the preceding integral only via the dimensionless
product ωt: it is for this reason that ∆ϕ is independent of the time τ = 2π/ω
that it takes for the device to complete a stroke, a deformation cycle.
Deformable bodies 89

Though striking, the phenomenon of rotation without angular momentum

accounts for how dropped cats manage to land on their feet,44 and is in evidence
whenever a diver departs the diving platform with zero angular momentum
and yet manages to perform complicated somersaults.45 I myself first became
interested in the phenomenon when, in , I attended a lecture by Thomas
Gold in which he advanced the thesis that in view of the plasticity of the earth
one should expect to find paleomagnetic evidence of large -scale polar wander
(mantle shifting with respect to the spin axis) over geologic time.46,47
I propose now to discuss the operation of our automaton from the more
readily generalizable point of view advocated by Shapere & Wilczek.42 The
device has five adjustable features: α and the four elbow angles β1 , β2 , β3 and
β4 . But if we insist that deformations preserve (relative to the device itself)
the location of the center of mass then three of the elbow angles become slaves
of the fourth (see Figure 38). To describe the shape of the device it would
suffice then to specify the values of α and β : though the “shape space” S of
deformable bodies can, in general, be expected to be ∞-dimensional, it is in
the present instance only 2-dimensional. We would write {α(t), β(t)}—which
is to say: we would inscribe a t -parameterized curve C on shape space—to
describe a temporal sequence of deformations. If the deformations are cyclic
(as deformations with locomotive intent tend to be) then C would have the
form of a closed loop. Note, however, that cyclicity does not, of itself, imply
temporal periodicity.
Imagine now that onto each shaped object we have—for future reference—
stamped a Cartesian frame, with origin at the center of mass. How this is
accomplished is a matter of fundamental indifference (see Figure 39), though
some frame-assignment procedures (adoption of the principal axis frame?) may
prove more useful—or at least feel more natural—in specific contexts. Clearly,
a rotation-matrix-valued function Q(α, β ) would serve to relate any such frame
assignment to any alternative assignment.
Consider now a blob—our carefully crafted device has, for the purposes
44
T. R. Kane & M. P. Scher, “A dynamical explanation of the falling cat
phenomenon,” J. Solids Struct. 5, 663 (1969).
45
C. Frohlich, “The physics of somersaulting & twisting,” Scientific American
263, 155 (March 1980); “Do springboard divers violate angular momentum
conservation?” AJP 47, 583 (1979).
46
T. Gold, “Instability of the earth’s axis of rotation,” Nature 175, 526
(1955).
47
It is important to notice that, while one can, by contortion, rotate about
one’s center of mass/change the way one faces in inertial space/“translate in
an angular sense” without external assistance, one cannot, by any amount of
contortion, translate one’s center of mass. And it is in this light interesting
that, according to Jack Wisdom (“Swimming in spacetime: motion by cyclic
changes in body shape,” Science 299, 1865 (March 2003)), one can do so in
curved spacetime: the effect is relativistic (disappears in the limit c ↑ ∞), and
such swimming is becomes impossible in flat spacetime.
90 Gyrodynamics

β β

β β

Figure 38: If the automaton flexes its elbows in such a way as to

preserve the location of its center of mass, then all four elbows must
be under the control of a single parameter, β.

of this discussion, become a generic “blob”—which moves through some

t-parameterized continuous sequence of shapes:

{α, β}initial −−−−−−−−−−−−−−−−−−→ {α, β}final

{α(t),β(t)}

Identifying the frame attached to {α, β}initial with a fixed reference frame in
physical space (the space frame), we introduce R (t) to describe the relationship
of that frame to the frame carried by the deformed blob after it has been
transported along C to the shape {α(t), β(t)} (see Figure 40). The oriention of
the deformed blob is determined by physical principle (conservation of angular
momentum), but the way it wears its frame is arbitrary. Our problem is to find
some way to distinguish what’s physical from what’s merely conventional.
From R (t)T R(t) = I it follows familiarly that it is always possible to write

Ṙ = A R with AT = −A

or—which is to say the same thing another way—

 t
R(t) = R0 + A(t ) R(t ) dt
0

t
Now, the solution of ṙ(t) = a(t)r(t) (i.e., of r(t) = r0 + 0
a(t )r(t ) dt ) is easily
seen to be  
t
r(t) = exp a(t )dt · r0 (128)
0
Deformable bodies 91

Figure 39: To assign frames to the various shapes of a 2-blob we

have here used center of mass ◦ and a pimple • as our guide—
a procedure that does not work for 3-blobs. We might alternatively
have adopted (say) the principal axis frame, though such a procedure
would become ambiguous when the principal moments are identical.

Figure 40: A rotation matrix R (t) relates the frame of a deformed

blob to the frame of the original blob.

in which connection we note that a temporal rescaling t → τ = τ (t) sends

d d
dt r(t) = a(t)r(t) =⇒ dτ R(τ ) = A(τ )R(τ )

with R(τ ) ≡ r(t) and A(τ ) = a(t)/τ̇ (t)

and this, by dτ = τ̇ (t)dt, means that the solution of ṙ = ar is scale -invariant:

 τ  t
A(τ ) dτ = a(t) dt
0 0
92 Gyrodynamics

A similar remark pertains to the iterative solution48

  t  t t

     
R(t) = I + A(t )dt + A(t )A(t ) dt dt + · · · · R0
0 0 0
 t 
= P exp A(t) dt · R0 (129)
0

of Ṙ = A R . Shapere & Wilczek interpret this to mean that the R(t) that
results from transport along a “curve in shape space” depends not at all upon
temporal specifics of the process, but only upon the geometry of the curve C.
Equation (129) makes clear also that specification of A(t) is sufficient in
principle to determine R(t).
Our further progress will be facilitated by some notational adjustment. Let
us write α1 = α, α2 = β, and let us recognize that R(t) means R(α1 (t), α2 (t))
which we will abbreviate R(α(t)). Then

d R = AR dt becomes d R = R ,i α̇i dt = R ,i dαi (130)

where R ,i ≡ ∂i R ≡ ∂R(α)/∂αi and where the Einstein summation convention

is understood to be in force. The differential dαi is tangent to C at the point α.
Immediately we confront a major problem: if we exercise our local frame
reassignment option
R(α) −→ R̂(α) = Q(α)R(α)
then
R̂ ,i = Q R ,i + Q ,i R
and the added term destroys the “reassignment covariance” of (130). To remedy
this defect we resort to a standard device: we introduce “compensating terms”
or “gauge fields” Ai —one for each degree of freedom in shape space—writing

R ;i ≡ R ,i − Ai R (131)

We then have
QR ;i = Q R ,i − QAi R
= (R̂ ,i − Q ,i R) − Q Ai R
= (R̂ ,i − Q ,i Q –1 R̂) − Q Ai Q –1 R̂
and insist upon
= R̂ ,i − Âi R̂

t
48
This is obtained by iteration of R(t) = R0 + 0 A(t )R(t ) dt , and gives
back (128) when all A(t) -matrices commute with one another. Here P is the
“chronological ordering” operator, the characteristic action of which becomes
evident from 
A(t1 )A(t2 ) : t1  t2
P[A(t1 )A(t2 )] =
A(t2 )A(t1 ) : t2  t1
Deformable bodies 93

which entails setting

Âi = QAi Q –1 + Q ,i Q –1 (132)
In short: if the “covariant derivative” R ;i is understood to be defined by (131),
and if the gauge fields are understood to transform Ai → Âi by the rule (132),
then
R → R̂ = Q R induces R ;i → R̂ ;i = Q R ;i (133)
Notice also that, since Q is a rotation matrix (Q –1 = Q T ), it is an implication
of (132) that if the matrices Ai are antisymmetric, then so are the matrices Âi .
Associated with every infinitesimal displacement in shape space α → dα
are not one but two distinct types of infinitesimal rotation. On the one hand
we have
R(α) −→ R(α + dα) = R(α) + R ,i (α)dαi (134.1)
that refers straightforwardly to the gradient structure of the frame-field that
we have (arbitrarily) deposited on shape space. On the other hand, we have49

R(α) −→ R(α + dα) = R(α + dα) − R ;i (α)dαi (134.2)

= R(α) + Ai (α)R(α)dαi

the precise meaning of which depends upon the structure assigned to the gauge
fields Ai (α), which are constrained only by (132): it is here—by contrived
specification of Ai (α)—that we will have an opportunity to slip some physics
into this formal scheme. The matrices R(α + dα) and R(α + dα) will, in general,
be distinct. If, however, they are identical then we say that R(α) −→ R(α +dα)
has proceeded by parallel transport , and can write

R(α + dα) = R(α) + Ai (α)R(α)dαi

If the parallel transport is along a curve α(t) in shape space—here t might but
need not signify time—then we have R(t + dt) = R(t) + Ai (t)R(t)α̇i dt or

Ṙ = A R with A ≡ Ai α̇i (135)

If we assume without real loss of generality that R(0) = I then, by (129), we

have   t  t t 
R(t) = I + A(t )dt + 12 P[A(t )A(t )] dt dt + · · · (136)
0 0 0

49
Said another way, we have

R(α + dα) − R(α) = R ,i (α)dαi

R(α + dα) − R(α + dα) = R ;i (α)dαi

It is from the circumstance that both matrices on the left side of the second
equation attach to the same point in shape space that R ;i (α) acquires its
superior transformation properties.
94 Gyrodynamics

It is important to appreciate that R(t) is a path-dependent object: it is

the result of parallel-transporting I from α(0) to α(t) along path C. Transport
along a different path linking the same endpoints can in general be expected to
yield a different result. To get a handle on the situation we ask: What is the
difference δ R = R2 −R1 that results when C2 and C1 differ only infinitesimally?

α(t)

α(s)

α(0)

Figure 41: Two curves inscribed on shape space that are coincident
except in the localized neighborhood of a pimple, where they differ
only infinitesimally.

Assuming the curves to differ only in the neighborhood of α(s), we have

 t   s 
δs R = P exp A(τ ) dτ · δA(s) · P exp A(τ ) dτ
s 0

and to distribute such pimples along the length of the curve we have only to
write
 t  t   s 
δR = P exp A(τ ) dτ · δA(s) · P exp A(τ ) dτ ds
0 s 0

But
δA(s) = Aj (s) · δ α̇j (s) + α̇i (s) · δAi (s)
∂Ai j
= Aj · ds
d
δαj (s) + α̇i (s) · δα (s)
∂αj
so  t  t   s 
δR = P exp etc. · Aj ds d
δαj · P exp etc. ds
0 s 0
 t  t   s 
∂Ai i
+ P exp etc. · j
α̇ · P exp etc. δαj (s)ds
0 s ∂α 0
where it is understood that the factors between dots are to be evaluated at s.
The first term we integrate by parts to obtain (after noting that by assumption
δαi (s) vanishes at the endpoints: δαi (0) = δαi (t) = 0 )
 t  t    s 
∂Aj i
− P exp etc. − Ai ȧ Aj +
i
α̇ + Aj Ai ȧ P exp
i
etc. δαj (s)ds
0 s ∂αi 0
Deformable bodies 95

giving finally
 t   t   s 
δR = δα (s) P exp
j
etc. · Fji · P exp etc. α̇i (s)ds (137)
0 s 0

where Fij refers to the antisymmetric array of 3 × 3 antisymmetric matrices

defined

Fij ≡ ∂i Aj − ∂j Ai − [Ai , Aj ] (138)

= −Fji

We have come here—by an argument adapted from a paper published

by Peter G. Bergman50 —upon a particular manifestation of object known to
differential geometers and general relativitists as the “Riemann-Christoffel
curvature tensor” and to field theorists as the “gauge field tensor.”51 Of the
many remarkable properties with which Fij is endowed, I will mention only one:
working from (132) we compute

[ Âi , Âj ] = Q[Ai , Aj ]Q –1 + stuff

and52
∂i Âj − ∂j Âi = Q(∂i Aj − ∂j Ai )Q –1 + same stuff!

—the implication being that Fij responds tensorially to gauge transformations:

F̂ij = Q Fij Q –1 (139)

50
“On Einstein’s λ transformations,” Phys. Rev. 103, 780 (1956). For my
immediate source, see pages 134–137 in Chapter 2 of classical dynamics
(/).
51
Note in this connection that if the matrices Ai and Q were number-valued
instead of matrix-valued, then (132) would read

Âi = Ai + ∂i Q

and (138) would become

Fij = ∂i Aj − ∂j Ai = F̂ij

These are equations that we recognize to be fundamental to electrodyamics.

“Non-abelian gauge field theory” is a generalization of electrodynamics in which
importance is assigned to the non-commutivity the gauge fields Ai .
52
Here one has need of (Q –1 ),i = −Q –1 Q ,i Q –1 , which follows directly from
(Q Q),i = O and is the non-commutative analog of dt
–1
q = −q −2 q̇.
d –1
96 Gyrodynamics

Figure 42: A closed curve C = ∂R bounds a region R in shape

space, which has been resolved into differential patches. Parallel
transport around C can be achieved by superimposing the results of
transport around each of the patches.

At (137) we managed to establish in effect that the result of parallel

transport around a differential patch (or “pimple”) at shape α can be described

δR = Fij (α) dαi ∧ dαj

We conclude (see the preceding figure) that transport of I around a finite closed
curve C—a cycle of shapes—has a rotational consequence that can be described
  
RC ≡ P exp A(τ ) dτ I = Fij (α) dαi ∧ dαj
C R

and that (since the initial and final reference frames are identical ) RC is gauge
invariant.53
We have now to pour some physics into the mathematical vessel that we
have been at such pains to construct. Erect an inertial frame at (let us say) the
center of mass of a (let us say) spinless system of particles:

S= mi r i × ṙri = 0
i

53
Recall from electrodynamics that
 
A · ds = (Ai,j − Aj,i ) dxi ∧ dxj
∂R

is invariant under A → A + ∇Q.

Deformable bodies 97

Imagine now that we have—in some continuous but otherwise arbitrary way—
associated a frame (origin coincident with that of the space frame) with every
possible configuration of the system. At time t the system finds itself in some
specific configuration, to which we have associated a frame. We agree to write
r i to describe the position of mi relative to that momentary frame, and

r i = R(t)rri

to describe the relationship of that frame to the space frame. In (non-inertial)

red variables the statement S = 0 becomes

mi [Rrri ] × [R ṙri + Ṙrri ] = 0
i
  
=R mi ri × ṙri + mi ri × R –1 Ṙrri
i i

Because R is a rotation matrix we have Ṙ = A R = R B where A and B are both

antisymmetric, but generally distinct. Writing B = Ω ×, we have (see again
page 5)
 
mir i × ṙri + mir i × (Ω
Ω × ri) = S + I Ω = 0
i i

giving
Ω = −I –1 S

whence54
B = Bij  with Bij = 2ijk (I –1S )k

At this point we
• can (but are under no obligation to) identify the particles mi with the
component parts of our deformable blob;
• can (but are under no obligation to) take our configuration-associated
frames to be principal axis frames.
Whatever our position with respect to the exercise of those options, we
• extract from the physics of the system a description of (compare (135))
B = Bi α̇i whence of Bi (which will be defined not everywhere in shape
space, but only where it is needed: on the curve pursued by the system),
with the aid of which we play the parallel transport game. Should the
system ever revisit a point in shape space we will be able to announce
whether it has experienced a net rotation as a result of its dynamical
zero -spin adventures (contortions).

54
It becomes clear only at this point why Shapere & Wilczek look to the
iterative solution of Ṙ = R B rather than (which is more common) of Ṙ = A R.
The resulting formalism is literally the transpose of that described on pages
90–96. Transposition entails reversal of the chronological ordering.
98 Gyrodynamics

It is an ambitious program, which I can expect to be computationally

feasible only in the simplest cases.55 It would be interesting to see whether it
can be brought to bear on the motion of our simple automaton (Figures 35)
which, as we have seen, admits of detailed analysis by elementary means. But
that is an exercise I must leave for another day.

18. Transformational aspects of rigid body mechanics. At (31) we obtained rigid

body equations of motion that read

S
N intrinsic = Ṡ
= I ω̇
ω + ω ×Iω

when referred to the space frame (generally non-inertial translated copy of the
inertial lab frame), on page 15 we drew attention to the fact that those equations
read
N 0 = I 0 ω̇
ω 0 + ω 0 ×I0 ω 0
—which is to say: they preserve their structure—when referred to the wobbly
body frame. I want now to discuss how this remarkable fact comes about.
Let a wobbly red frame which shares the origin of—but be in a state of
arbitrary rotation with respect to—the space frame, and write

r = Wrr : W is an arbitrarily t-dependent rotation matrix (140)

to describe the relationship between the red and black coordinates of any given
point. Immediately

ṙr = W ṙr + Ẇrr

= W ṙr + Urr with U ≡ Ẇ W –1 = −U T

which can be written

W ṙr = ṙr + Urr with U ≡ −U (141.1)

d 
= dt +U r

Extensions of the same basic line of argument give

d 2
Wr̈r = dt +U r
= r̈r + 2U ṙr + ( U̇ + U 2 )rr (141.2)
..
.
d n
Wrr(n) = dt +U r
55
Some specific examples are discussed by Shapere & Wilczek, both in the
paper cited previously42 and in a companion paper of slightly earlier date:
“Geometry of self-propulsion at low Reynolds number,” J. Fluid. Mech. 198,
557 (1989). This paper also is reprinted (as §8.4) in Geometric Phases in
Physics.
Transformational aspects of rigid body mechanics 99

If the space frame were inertial (which generally it is not) then to describe
the dynamics of a single particle we would write F = mr̈r, which when referred
to the wobbly frame becomes

F = m r̈r + 2U ṙr + ( U̇ + U 2 )rr

with F ≡ WF
F and m ≡ m. The preceding equation is often written

F + F Coriolis + F centrifugal = mr̈r (142)

with

F Coriolis ≡ −2mU ṙr

= −2mΩ
Ω × ṙr (143.1)
F centrifugal ≡ −m( U̇ + U 2 )rr
= −m Ω̇
Ω ×rr − mΩ
Ω × (Ω
Ω ×rr) (143.2)

Look now to the intrinsic

! angular momentum (“spin”) of a loose system of
particles. Hitting S = i mixi ×x xi with W gives

WSS= mi W(rri × W –1 W ṙri )
i

= mir i × W ṙri by the lemma of page 15
i

= mir i × (ṙri + Urri )
i

= mir i × (ṙri + Ω × r i )
i
 
= mir i × ṙri − mi (rri ×)2 Ω
i i
≡ S + IΩ (144)

which is not the result ( WS S = S ) that one might have anticipated. We note
also in this connection that, by appeal once again to the lemma,
 
W I W –1 = − mi W (rri ×)W –1 W (rri ×)W –1 = − mi (rri ×)2 ≡ I (145)
i i

Look now to the motion of S . By differentiation of (144) we have

N ≡ WN N
= WṠS
= Ṡ + İ Ω + I Ω̇
S Ω − ẆS
S
S + İ Ω + I Ω̇
= Ṡ Ω − Ẇ W –1 · WS
S
S + İ Ω + I Ω̇
= Ṡ Ω + Ω × (S
S + I Ω)
S + Ω ×S
= (Ṡ S ) + ( [ İ Ω + I Ω̇
Ω ] + Ω×IΩ)
100 Gyrodynamics

which—by seemingly trivial rearrangement—becomes

N − (Ω
Ω ×S
S + İ Ω ) − (I Ω̇
Ω + Ω × I Ω ) = Ṡ
S (146)

I turn now to manipulations intended to clarify the meaning of (146). We

have

Ω ×S
(Ω S + İ Ω ) = Ω × mi (rri × ṙri )
i
  
− mi (ṙri ×)(rri ×) + mi (rri ×)(ṙri ×) Ω
i i
  
= mi Ω × (rri × ṙri ) + ṙri × (Ω
Ω ×rri )
i

+ mir i × (Ω
Ω × ṙri )
i

which by a × (bb ×cc) + c × (a

a ×bb) + b × (cc ×a
a) = 0 becomes
  
= r i × 2mi (ΩΩ × ṙri ) (147.1)
i

A similar argument (I omit the details) supplies

  
Ω + Ω×I Ω) =
(I Ω̇ ri × mi Ω̇Ω ×rri + mi Ω × (Ω
Ω ×rri ) (147.2)
i

But the expressions in braces are familiar already from (143). The striking
implication is that (146) can be written

N + N Coriolis + N centrifugal = Ṡ
S (148)

where

N Coriolis ≡ r i ×F
F iCoriolis (149.1)
i

N centrifugal ≡ F icentrifugal
r i ×F (149.2)
i

serve to define what might be called—though I have never encountered such

terminology in the literature —the “net Coriolis and centrifugal torques.” These
we recognize to be “fictitious torques,” artifacts of the non-inertiality of the
wobbly red frame.
If, as a special circumstance, our many-particle system is rigidly assembled
then it becomes natural—not mandatory, but natural—to

identify the wobbly red frame with the body frame,

Transformational aspects of rigid body mechanics 101

with respect to which all particles are at rest: ṙri = 0 (all i). From (147.1) and
!
S ≡ mir i × ṙri it then follows that (Ω
Ω × S + İ Ω ) and Ṡ
S both vanish: (146)
therefore becomes
N − (I Ω̇
Ω + Ω×IΩ) = 0 (150.1)
which are usually written
N = I Ω̇
Ω + Ω×IΩ (150.2)
and called the Euler equations. Equations (150.1) serve at once to
illuminate and to deepen the mystery that motivaated this discussion, for they
establish that Euler’s equations might most properly be expressed

N + N centrifugal = 0 in the non-inertial body frame of a rigid system

Notice that if the system were deformable (which is to say: not rigid) then
it would be impossible to select a frame with respect to which all ṙri terms vanish:
one would be forced to work with some instance of (148). The implication is
that it should be possible to get from (148) to the Liouville equations,56 which
are used by astrophysicists to study the rotational dynamics of stars, planets
and astroids. I must admit, however, that I do not at present know how to do
so.

56
See Problem 10 in Chapter 5 in H. Goldstein’s Classical Mechanics (2nd or
rd
3 editions).