Escolar Documentos
Profissional Documentos
Cultura Documentos
SUMMARY
We present an alternative derivation of Simo and Vu Quoc’s numerical algorithm1 for modelling the non-linear
dynamic behaviour of rods. The original derivation uses dierential topology, describing large rotations using
the Lie group SO(3) and Lie algebra so(3), but resorting to quaternions for the numerical implementation.
The new derivation uses Cliord or geometric algebra as developed by Hestenes2; 3 for both formulation and
implementation. We contend that the new approach is considerably simpler to follow, and thereby allows
alternative modelling strategies to be more readily investigated. The new description is also novel in that all
formulae for rotational kinematics are applicable in a Euclidean space of any dimension. Copyright ? 1999
John Wiley & Sons, Ltd.
KEY WORDS: Cliord algebra; geometric algebra; rods; large rotations
1. BACKGROUND
Over the past few decades the foundations of classical mechanics have been rigorously reformu-
lated in terms of dierential topology. The major contributors to this process have been Abraham
and Marsden in America, with their deÿnitive tome ‘Foundations of Classical Mechanics’4 and
Arnol’d in Russia, whose work in this area is typiÿed by his book ‘Mathematical Methods of
Classical Mechanics’.5 The early contributions focused on classical particle mechanics and rigid
body mechanics. A subsequent advance was the publication of Marsden and Hughes ‘Mathematical
Foundations of Elasticity’6 where the dierential topology reformulation was extended to elastic
continuum mechanics, providing a fundamental alternative to the earlier description based upon
classical tensor analysis as exempliÿed by the works of Truesdell, Noll, Gurtin, Naghdi and others
in Vols. III, IV and VI of ‘Handbuch Der Physik’.7 This reformulation has continued apace into
the realms of computational continuum mechanics most notably through the work of Simo and
coworkers, in particular, on the ÿnite element analysis of the large displacement behaviour of rods
and shells (e.g. References 1, 8–11). This presents a particular problem in that dierential topology
rarely forms part of an engineering education, and thus the foundations of applied mechanics are
now written in a language that is impenetrable to most engineers.
∗Correspondence to: F. A. McRobie, Department of Engineering, Cambridge University, Trumpington Street, Cambridge
CB2 lPZ, U.K. E-mail: fam@eng.cam.ac.uk
Over the last ten years an alternative reformulation has been pioneered by Hestenes and
co-workers who, building upon Cliord algebra and Grassman calculus, have developed a math-
ematical framework they refer to as Geometric Algebra. In the introductory ‘New Foundations
of Classical Mechanics’2 and in the more formal ‘Cliord Algebra to Geometric Calculus’3 the
foundations of this alternative approach are presented. Again any physical emphasis is on particle
mechanics. The initial extension of the approach to continuum mechanics has been commenced by
Gull et al.12 In this paper we provide the initial extensions to computational continuum mechanics
by undertaking the reformulation of the Simo–Vu Quoc algorithm for the non-linear dynamics
of rods. Although this algorithm has many shortcomings, both in terms of its numerical perfor-
mance and in the physical description of rod mechanics, it provides perhaps the simplest entry
point to the literature on computational models based on dierential topology. This extensive liter-
ature, as well as the numerous contributions of Simo and co-workers, includes such as References
13–16.
Geometric algebra has been proposed as the natural mathematical framework for physics and
mechanics.2 In a number of non-classical areas of theoretical mechanics such as quantum physics
and general relativity, the use of geometric algebra is somewhat controversial: reformulating the
existing theories in the new language has led to new and dierent predictions of behaviour rather
than mere transcription.17 – 19 However, in engineering, it is envisaged that there will be no such
controversy. The principal argument for the adoption of geometric algebra is that it provides
a single, simple mathematical framework which eliminates the need for the plethora of diverse
mathematical descriptions and techniques that it would otherwise be necessary to learn.
A particular area where geometric algebra provides a unifying language is in the description
of large rotations. The most fundamental modern treatments such as those of Abraham, Marsden,
Arnol’d and Simo use dierential topology and describe rotations in terms of the Lie group SO(3).
A rotation is thus an element of a dierentiable manifold, and combinations of rotations are
described using the group action. Innitesimal rotations and rotational velocities live in the tangent
bundle TSO(3), a dierentiable manifold with distinctly non-trivial topology, from where they can
be transported to the tangent space at the identity, identiable with the Lie algebra so(3). As
throughout all of the dierential topology formulation of mechanics, a proliferation of manifolds
occurs. Whilst the proper placement of each object in its appropriate topological space enables
proper rigour, it can lead to even further complications when isomorphisms between these spaces
endow a single geometric object with a number of mathematical manifestations. As a simple
example, an angular velocity may be described both by a vector and a skew-symmetric matrix.
In geometric algebra there is no such proliferation of manifolds: the mathematical arena consists
only of elements of the algebra and nothing more. In most classical applications this will be the
algebra of three-dimensional physical space. (In non-classical applications it will probably be the
so-called Space–Time Algebra corresponding to four-dimensional space).
We feel that the sheer mathematical sophistication of the current description of mechanics puts
modern theoretical developments beyond the reach of most engineers. Few understand rigid body
mechanics in terms of the Lie–Poisson structure on a Poisson manifold and concepts such as
Lie–Poisson reduction and coadjoint orbit actions that pervade the works of the dierential topology
school may be powerful, but are virtually inaccessible to practising engineers. However, such
austere formulations have already found their way into the computational packages routinely used in
design oces. Consider for example the numerous references to the works of Simo in the ABAQUS
Theory Manual.20 One wonders what fraction of engineers who regularly use this package have
read and understood these papers.
Copyright ? 1999 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 45, 377–398 (1999)
SIMO–VU QUOC RODS USING CLIFFORD ALGEBRA 379
Given the diculty of the mathematical formulations described above, it is no surprise that
the description of large rotations elsewhere in the engineering literature has seen a proliferation
of alternative descriptions and techniques that avoid such arcane methods. For example, a typ-
ical textbook on spacecraft attitude dynamics21 uses a number of descriptions for the rotational
kinematics which include rotation matrices, direction cosines, Euler angle parameterization, Euler
symmetric parameters (or quaternions), Cayley–Klein parameters and Gibbs vectors. In compu-
tational mechanics, Argyris’s paper 22 on semi-tangential rotations is much quoted. Texts on the
kinematics of mechanisms employ quaternions and bi- or (dual) quaternions, along with screw
theory, Study co-ordinates and Plucker coordinates; three-, ve- and seven-dimensional projective
spaces also gure prominently. In computer vision, four-dimensional homogeneous co-ordinates
are frequently employed. If we extend the realm of applied mathematics to include physics we
observe that many physics texts elect to follow the spinor description of rotations with Pauli, Dirac
(and even Weyl) matrices for descriptions of quantum mechanical spin. In higher dimensions we
could even include twistor theory.
We contend that the description of large rotations contained naturally within geometric algebra
provides a simple cohesive alternative to both the pure mathematical description with its prolif-
eration of manifolds, and the applied literature with its proliferation of methods, techniques and
descriptions. In one sense though, the geometric algebra description is not new, in that in any
particular engineering application, it is equivalent to any one (or indeed all) of the other methods.
For classical problems the end results should be identical: it will only be the description of the
various objects and the way in which they are manipulated that will dier.
Although there are many more mathematically rigorous (e.g. Reference 3) and more complete
introductory presentations available (e.g. Reference 2) than the one we provide here, given the
general unfamiliarity of most engineers with geometric algebra we feel that the inclusion of some
such introduction is appropriate.
In essence, geometric algebra is merely ordinary undergraduate vector analysis extended by
a handful of simple concepts: in particular, the denition of the wedge and the geometric (or
Cliord) products, and the admission that it is permissible to add objects of dierent geometric
type (a scalar plus a vector, for example) to obtain new geometric objects.
Assume, for the purposes of this introduction, that one has a vector space spanned by an
orthogonal set of unit basis vectors {ei }. Dene then the ‘wedge’ product u ∧ v of any two vectors
u and v as the oriented area swept out when u is swept along v, see Figure 1.
Such an oriented area is a new geometric object called a ‘bivector’. The orientation of u ∧ v is
dened to be positive in the sense where one circumscribes the perimeter of the area by setting
o rst along the vector u. The wedge product is therefore clearly skew-symmetric:
u ∧ v = −v ∧ u (1)
The Cliord (or geometric) product uv of two vectors u and v is dened as
uv = u·v + u ∧ v (2)
The rst term is a scalar, equal to the dot product of u and v, and the second term is the
bivector discussed earlier. As in Grassman calculus, geometric algebra allows complete freedom
Copyright ? 1999 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 45, 377–398 (1999)
380 F. A. MCROBIE AND J. LASENBY
to add elements of dierent geometric types (scalars plus vectors plus bivectors, etc.) to obtain
new geometric entities, called multivectors. Geometric objects of higher order but single type
(trivectors, quadrivectors, etc.) are simply oriented volumes of correspondingly higher dimension.
Obviously trivectors cannot exist in a two-dimensional vector space, and so forth.
In a more rigorous approach, the Cliord product is taken as the fundamental object, and the dot
and wedge products of vectors are the symmetric and anti-symmetric parts of this. One consequence
is that every vector squares to give a scalar (i.e. the geometric product of a vector with itself is
a scalar). This is not true of the squares of general multivectors.
The geometric algebra of a two-dimensional vector space has four elements: a scalar, two vectors
and a bivector. The algebra of a three-dimensional vector space has eight elements: one scalar,
three vectors, three bivectors and one trivector. And so on in higher dimensions.
The highest dimensional element of the algebra (corresponding to an oriented volume of the
same dimension as the vector space) is called the pseudoscalar. In the geometric algebra literature
(but not in this paper) it is usually denoted evocatively by the letter i in two and three dimensions,
and by I in higher dimensions.
Note the following properties of the geometric product of basis vectors. Since
ei ej = ei ·ej + ei ∧ ej
then: if i = j; ei2 = 1 (scalar part only)
if i 6= j; ei ej = ei ∧ ej (bivector part only)
A basis for bivectors is thus given by geometric products of unit vectors (i.e. ei ej is a unit
bivector).
Starting from the orthonormal basis {ei } for the vector space, a basis for the entire algebra is
readily generated from geometric products of basis vectors.
In 2D, the basis is
1; {e1 ; e2 }; e1 e2
In 3D, the basis is
1; {e1 ; e2 ; e3 }; {e2 e3 ; e3 e1 ; e1 e2 }; e1 e2 e3
and so on for higher dimensional spaces.
Although many interesting descriptions can be obtained by developing geometric algebra with
non-Euclidean metrics, throughout our exposition all spaces are assumed to be Euclidean.
In a (Euclidean) space of any dimension, note the following property of the square of a unit
bivector:
(ei ej )2 = (ei ej )(ei ej ) (no summation) (3)
Copyright ? 1999 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 45, 377–398 (1999)
SIMO–VU QUOC RODS USING CLIFFORD ALGEBRA 381
In particular, in two dimensions if we temporarily represent the unit bivector e1 e2 (the pseu-
doscalar) by the symbol i we obtain
i2 = −1 (7)
The geometric algebra interpretation of the unit imaginary, i, is thus that it is not imaginary at all,
but that it is simply the unit bivector in a two-dimensional vector space.23
In the same way that multiplying a complex number z = x + iy by the unit imaginary i rotates
z by 90◦ anti-clockwise on the Argand diagram, so in the 2D geometric algebra, post-multiplying
a vector v by the unit bivector e1 e2 rotates v through 90◦ anti-clockwise.
In higher dimensions, rotations of a vector through 90◦ are equally easy to accomplish: if the
vector v lies in the plane spanned by e1 and e2 then the post-multiplication by the bivector e1 e2
rotates v by 90◦ anti-clockwise in that plane. Explicitly
if v = v1 e1 + v2 e2 (8)
then ve1 e2 = v1 e1 e1 e2 + v2 e2 e1 e2 (9)
= −v2 e1 + v1 e2 (10)
Obviously for more general rotations in higher dimensions the representation of a rotation in
terms of a rotation vector is an incomplete and inappropriate description. Instead a description of
a plane of the rotation is more apposite, and since in geometric algebra a unit bivector denes a
plane, a rotation of a given magnitude is represented by a bivector of that magnitude.
This is the rst real departure from conventional descriptions of physical quantities: in this
description we have rotation bivectors.
Remark 1. One of the advantages of geometric algebra is the ease with which results can be
generalized to any dimension, and for the remainder of this section, and throughout Section 3,
since it involves almost no extra eort, we shall write all results in a form valid in n-dimensional
Euclidean space unless stated otherwise. There is a technicality, however. All formulae that we shall
present will refer to planar rotations, which are dened by elementary (or planar) bivectors, i.e.
bivectors associated with a single plane. In higher dimensions it is possible to dene more general
rotations, which may be built up from compositions of planar rotations. We do not present any
analysis for such general rotations here, but suggest that it is straightforward using the language,
methods and results presented below. In dimensions greater than three, not all bivectors can be
associated with a single plane (B = Ba e1 e2 + Bb e3 e4 , for example), and one consequence of our
restriction to planar rotation bivectors is that the square of any such bivector is a scalar. This is
not true for general bivectors, which may square to a scalar plus a quadrivector.
√ For example, for
a bivector we shall often write a2 = −1 where a ≡ = and ≡ i i but this is only true
if is planar. However, in three dimensions, all rotations and bivectors are planar and such
technicalities can be ignored in this case of most concern to computational mechanics.
In three-dimensions, to obtain the rotation bivector from a rotation vector one simply multiplies
the rotation vector by the pseudoscalar (unit trivector) e1 e2 e3 (or i). Consider a rotation in a
three-dimensional vector space about the direction X̂ by an angle , represented by the rotation
Copyright ? 1999 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 45, 377–398 (1999)
382 F. A. MCROBIE AND J. LASENBY
Copyright ? 1999 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 45, 377–398 (1999)
SIMO–VU QUOC RODS USING CLIFFORD ALGEBRA 383
As a nal example, note the way that Hamilton’s relations drop out of this approach: setting
i ≡ e2 e3 , j ≡ −e3 e1 , and k ≡ e1 e2 , one readily obtains
i2 = j2 = k2 = ijk = − 1 (18)
Comparing this with our earlier statement that the unit bivector in a two-dimensional vector
space squares to −1, it is a trivial matter to generalize Hamilton’s relations to a space of any
dimension by stating the relationship between products of unit bivectors.
In this section we introduce the commutator and anti-commutator notation common throughout
the physics literature. This notation is rarely used by proponents of geometric algebra who adopt a
more succint notation for algebraic manipulation. However, in its most succint form, this requires
familiarity and dexterity with the precedence relations of the various multiplication operators.
Accepting some slight loss of brevity, we adopt instead the commutator and anti-commutator
notation for its unambiguous simplicity.
As usual the commutator and anti-commutator of two multivectors U and V are dened as
[U; V] ≡ 12 (UV − VU) commutator (19)
(U; V) ≡ 12 (UV + VU) anti-commutator (20)
For example, if U is a scalar u then [u; V] = 0 and (u; V) = uV = Vu (scalars commute with all
multivectors).
Similarly, if both are vectors then (u; v) = u·v = v·u = ui vi . This is the symmetric part of the
geometric product. Since uv = (u; v) + [u; v] it follows that [u; v] is the antisymmetric part, and in
the particular case of a three-dimensional vector space we have [u; v] = u ∧ v = iu × v. The vector
cross-product of two vectors in 3D can thus be obtained from their commutator. Note that both
the commutator and wedge product make sense in any dimension, whereas vector cross-products
can only be dened in 3D.
The purpose of using this notation in this paper is that much of the novelty lies in the use
of bivectors to describe rotations, angular velocities, bending moments, etc., and a considerable
amount of bivector manipulation is involved. The commutator and anti-commutator notation allows
a very convenient yet rigorous method for handling such manipulations.
In this section we develop the description of rotations and their derivatives which will be necessary
for any subsequent computational mechanics applications. We note that much of the following is
applicable far more widely than to the mere mechanics of rods. In particular, since it entails almost
no extra eort, we write all statements in this section on rotational kinematics in a form valid in
a Euclidean space of any dimension. (Refer to Remark 1 earlier, however, for some background
technicalities.) Perhaps such novel descriptions may nd application in elds such as dynamical
systems theory, signal processing and linear systems theory, where higher dimensional phase spaces
are commonplace. We cannot however foresee any need for a general n-dimensional description of
rotational mechanics, and so subsequent sections which involve concepts such as momenta will be
Copyright ? 1999 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 45, 377–398 (1999)
384 F. A. MCROBIE AND J. LASENBY
restricted to the three-dimensional setting. However, even there, many of the results are applicable
to rigid and
exible structures far more general than rods alone.
It is common in the geometric algebra literature to reserve lower-case bold letters for vectors.
We endeavour to adhere to this, but have found it necessary to make some exceptions. In particular,
the quantities a; q; g and ki that we use in this and subsequent sections are bivectors.
Let
V = RUR̃ (21)
where U = U(t) is a time-varying multivector and R = R() is the rotor corresponding to a time-
varying bivector (t).
Straightforward time-dierentiation gives
We have dened
S ≡ −2ṘR̃ (27)
˙ This is
but in order to be able to use this we shall need to express it as a function of and .
achieved as follows. We have
R() = cos − a sin (28)
2 2
Copyright ? 1999 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 45, 377–398 (1999)
SIMO–VU QUOC RODS USING CLIFFORD ALGEBRA 385
Now,
2 = − (34)
so
2˙ = −
˙ −
˙ (35)
thus,
1 ˙
˙ = − 1 ( ;
˙ = − + ˙ ) = − ( ;
˙ a) (36)
2
Also, since
˙
˙
a= then ȧ = − 2 (37)
and using (36),
˙
˙ a) 1 n
( ; o
ȧ = + = ˙ + ( ;
˙ a)a (38)
2
Dening S ≡ sin(=2)==2 and collecting terms in the above we obtain the nal result:
˙ a)a + S{
S ≡ −2ṘR̃ = −( ; ˙ + ( ;
˙ a)a}R̃ (39)
which expresses the spatial angular velocity bivector in terms of the rotation bivector and its
˙
derivative .
Although derivation of this simple formula (39) has taken a little eort, that eort should be
compared with that required to derive the corresponding equation (equation A.9, Reference 8) of
the dierential topology approach. Derivation of that equation by the dierential topology methods
Copyright ? 1999 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 45, 377–398 (1999)
386 F. A. MCROBIE AND J. LASENBY
we leave as an exercise for the reader. Note that it is not merely a question of algebraic simplicity
but also of considerably increased mathematical power. Equation (39) above gives the formula
applicable to the derivative of any rotated multivector quantity in a space of any dimension,
whereas the dierential topology equation, for all the severity of its derivation, is only applicable
to the rotations of vectors in three dimensions.
Remark 2. (This may be skipped by readers interested only in the three-dimensional
formulation.) An interesting technical point concerning the n-dimensional result arises from equa-
tion (36). The left-hand side, ˙ is clearly a scalar. The right-hand side −( ;
˙ a) is therefore also
a scalar. However, if ˙ were a general bivector, then −( ; ˙ a) may contain quadrivector parts.
Since this would be a contradiction, we conclude that there are constraints on . ˙ These con-
straints, which are dened by equation (36), amount to the requirement that as changes, it must
remain planar. In the same way that not all bivectors can dene rotations, not all derivatives are
allowed either. It follows from equation (39) that there are also constraints on the possible angular
velocities
S .)
The derivation of the spatial angular velocity bivector above is completely coordinate-free. How-
ever, in applications, dierent frames of reference are involved. If we take the above derivation
to be with respect to an underlying set of basis vectors xed in space (hence the name spatial
angular velocity), then we can dene, as usual, another (bivector) quantity called the material
angular velocity bivector
B dened by
B ≡ R̃
S R = − 2R̃Ṙ (40)
Equation (39) is our primary description of the spatial angular velocity bivector
S . Note, in
particular, that it is a non-linear function of the rotation bivector but a linear function of the
derivative .˙ It is useful to write
S = ˙ ≡ −( ;
( )
˙ a)a + S{
˙ + ( ;
˙ a)a}R̃ (41)
Thus, for given , is a linear function mapping the bivector ˙ to the bivector
S .
There are occasions when we require the inverse of this function, and this is readily obtained.
Substituting ˙ + ( ;
˙ a)a = − [ ;
˙ a]a and taking the anticommutator of (41) with a, we obtain
2
(using aR̃ = R̃a and a = − 1) the result
˙ a)
(
S ; a) = ( ; (42)
˙ + (
S ; a)a}R̃
S = − (
S ; a)a + S{ (43)
Copyright ? 1999 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 45, 377–398 (1999)
SIMO–VU QUOC RODS USING CLIFFORD ALGEBRA 387
which rearranges to
˙ = −(
S ; a)a + 1 {
S + (
S ; a)a} R
(44)
S
≡ −1 (
S ) (45)
Again, we stress that these formulae are applicable in any dimension.
˙ =
B and −1 (
B ) =
(Similar linear functions ( ) ˙ can be readily determined in the ma-
terial setting.)
Almost all of the preceding analysis has referred to time-derivatives of multivectors, and yet
computational mechanics formulations usually also require the evaluation of innitesimal variations
of objects. Here, we derive the appropriate variational expressions. Obviously, most of the analysis
parallels the earlier work on derivatives, and thus we may be more succint here. Suppose
Now dene
where denotes an innitesimal variation, such that terms of higher than rst order in may be
neglected.
Following through the analysis parallel to Sections 3.1 to 3.4, we obtain
V = R U R̃ + [ V; ()] (46)
where
Given a rotor R() and its variation R, one can dene other innitesimal bivectors associated
with . We have already encountered the additive form, denoted , related to R by
Another is the left multiplicative form we denote by , dened such that
Copyright ? 1999 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 45, 377–398 (1999)
388 F. A. MCROBIE AND J. LASENBY
We have
−
R() = exp =1 − + higher-order terms (50)
2 2
Also, from (47),
1
R = − 2 ()R (51)
so, since
R( + ) = R()R() (52)
then
1
R() + R = R + − ()R = 1− R (53)
2 2
thus
1
1− () = 1 − (54)
2 2
so, nally,
−1
() = and = () (55)
These nal relationships clearly illustrate the meaning of the function dened earlier in
the context of the angular velocity. It is that relates innitesimal additive and multiplicative
variations.
A common problem in the kinematics of large rotations is that many functions are multi-
valued. A rotation by 4 is the same as no rotation at all, for example. It is therefore useful
in computational mechanics to employ compound rotations from a given rotation. If compound
rotations are suciently moderate then all ambiguities in multi-valuedness can be avoided. By
moderate, we mean of magnitude less than , say. There is no intention that use may be made
of this moderation for approximation purposes, neglecting terms of higher order in the compound
rotations, for example.
During the evolution of the equations of motion, assume that the rotation bivector at time tn
is n . At time t let the rotation bivector be . For moderate time dierences, we can unambigu-
ously dene a compound rotation bivector q via
R() = R(q)R(n ) (56)
There is more to the use of compound rotations than the desire to avoid ambiguities in multi-
valuedness. As pointed out in Criseld and Jelenic,26 interpolation between rotations involves a
number of counter-intuitive subtleties. Perhaps the most obvious interpolation between two rotations
which are dened by bivectors 0 and 1 would be linear interpolation between the bivectors,
the formula
s = (1 − s)0 + s1 (57)
Copyright ? 1999 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 45, 377–398 (1999)
SIMO–VU QUOC RODS USING CLIFFORD ALGEBRA 389
Since the numerical algorithms that we consider in subsequent sections involve much interpolation
and extrapolation of rotations, the use of compound rotations will gure prominently.
Consider now a conguration dened by the rotation bivector , and say we require the exact
linearisation of our equations around this conguration. We thus let → + and q → q + q,
and dene the associated innitesimal multiplicative rotation variations and q via R( +
) = R()R() and R(q + q) = R(q)R(q) for the total and compound rotation bivectors
and q, respectively (as described in Section 3.5).
Taking innitesimal variations of equation (56) above, we obtain
Therefore,
thus
giving
= q (63)
Since R(q + q) = R(q)R(q), note that an analysis parallel to equations (48) – (55) (replacing
by q) gives
q (q) = q (64)
is the geometric algebra statement of the Simo–Vu Quoc expression (equation (3.11), Reference 1)
for the linear part of the compound rotations
(i)
n(i) = T (n(i) )n+1 (66)
Again, readers are invited to derive the formula for T () (equation (3.12), Reference 1) using the
methods of dierential topology.
Copyright ? 1999 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 45, 377–398 (1999)
390 F. A. MCROBIE AND J. LASENBY
which are the relations between the various descriptions of the innitesimal variations of the
rotations dened so far.
Another compound rotation variable Q can be dened by
(cf. the denitions of n(i) and n(i+1) in the rubric of Table 4, Reference 1).
Since n is xed, taking variations leads to
−1
Q = R̃(n )[ q (q)]R(n ) (70)
(The corresponding term in Simo–Vu Quoc notation reads tn T (n(i) )n+1
(i)
appearing at the end
of their equation (3.15) Reference 1, for example.)
We shall now derive the Simo–Vu Quoc algorithm using geometric algebra. As previously stated,
this algorithm has a number of limitations, and thus rather than reproducing the whole model in
this paper, we derive only the rotational components of the tangent inertia matrix (equation (4.3c)
Reference 1).
First however, we must derive the appropriate terms in the governing equations of motion.
4.1. Preliminaries
Simo and Vu Quoc’s rod model is a Cosserat model, in that the deformation of the rod is
completely specied by independent displacement and rotation elds. We consider the simplest
case where, in the reference conguration, the rod is unstressed and straight, pointing along one of
the unit basis vectors (e1 , say) xed in space. Cross-sections of the rod are dened in the reference
conguration as being those surfaces perpendicular to the e1 direction. We assume for simplicity
that the centre of mass of each section lies on the e1 axis, and that the line of centres of mass
(the centreline) may be parameterized by a length parameter s in the reference conguration. The
centre of mass of a reference section thus lies at the point se1 . Material axes for the cross-section
are then dened in the reference conguration by the directions e2 and e3 .
In a deformed conguration at time t, the new centreline of the rod is specied by a displacement
(vector) eld r(s; t).
In the deformed conguration, a material reference cross-section spanned by the vectors e2 and
e3 is mapped to a new cross-section spanned by vectors t2 and t3 .
Copyright ? 1999 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 45, 377–398 (1999)
SIMO–VU QUOC RODS USING CLIFFORD ALGEBRA 391
4.2. The angular momentum bivector of a rigid body and the inertia tensor
The description we present here is essentially that given in Gull et al.,12 although the notation
diers slightly.
Consider a rigid body whose centre of mass is at r, with an orthonormal frame {ti } xed in
the body with origin at the centre of mass. Let there be a frame of orthonormal basis vectors {ei }
xed in space at some global origin.
A material point in the body has location u = r + xi ti , and there is a rotation bivector such
that ti = Rei R̃ with R = exp(−=2).
The velocity of this point is u̇ = ṙ+xi ṫi , but using equation (26), ṫi = [ti ;
S ], and since ti = Rei R̃
and
B ≡ R̃
S R, we obtain ṫi = R[ei ;
B ]R̃ thus
u̇ = ṙ + xi R[ei ;
B ]R̃ (71)
The angular momentum about the origin of a rigid innitesimal element of volume dV and
density at this point is the bivector dL dened by dL ≡ u ∧ u̇ dV and thus the total angular
Copyright ? 1999 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 45, 377–398 (1999)
392 F. A. MCROBIE AND J. LASENBY
For the lamina of innitesimal thickness s in the undeformed conguration, the inertia tensor
for the general rigid body derived above simplies. Since the origin is at the centre of mass, all
integrals involving a single x1 (through the thickness) vanish, and the term involving x12 can be
neglected, leaving
I(
B ) = (I22 + I33 )
B1 k1 + (I33
B2 − I32
B3 )k2 + (I22
B3 − I23
B2 )k3 (78)
If C = I(
B ) the components of the bivectors are related by the matrix expression
C1 I22 + I33 0 0
B1
C2 = 0 I33 −I32
B2 (79)
C3 0 −I23 I22
B3
One must remember, however, that the rows and columns of the matrix refer to the underlying
unit bivector basis {ki } and not a vector basis as usual.
Copyright ? 1999 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 45, 377–398 (1999)
SIMO–VU QUOC RODS USING CLIFFORD ALGEBRA 393
The angular momentum bivector L of the lamina (equation (76)) may be separated into terms
Lg = M [r; ṙ] and Lc = RI(
B )R̃ where Lg gives the angular momentum due to the motion of the
centre of mass of the lamina about the global origin, and Lc gives the angular momentum of the
lamina about its own centre of mass.
In the Simo–Vu Quoc formulation, the rotational components of the local momentum balance
equations are given about the centre of mass of the lamina (Table 1, Reference 1), and the rotational
inertia component is simply the rate of change of (local) angular momentum. Our corresponding
local inertia term is
d
g ≡ L̇c = RI(
B )R̃ (80)
dt
dI(
B )
=R R̃ + [RI(
B )R̃;
S ] using equation (26) (81)
dt
= RI(
˙ B )R̃ + [RI(
B )R̃; R
B R̃] (82)
= RHR̃ (83)
where
˙ B ) − [
B ; I(
B )]
H ≡ I(
(84)
Above, we have the rotational inertia term g = RHR̃. We now linearise this about the rotation
bivector at time t by taking innitesimal variations. This gives
g = R H R̃ + [g; ()] using equation (46) (85)
= R H R̃ + [g; q] using (67) (86)
Since
˙ B ) − [
B ; I(
B )]
H = I(
(87)
then
H = I(
˙ B ) − [
B ; I(
B )] − [
B ; I(
B )] (88)
In this section we recast Simo and Vu Quoc’s extension of the Newmark beta algorithm24 to
deal with rotational accelerations and velocities. In light of the above analysis, we nd their claim
that it is the ‘canonical’ extension to be questionable. However, despite the superiority of other
algorithms11; 25 for the time-integration of the sti, nonlinear equations of structural dynamics, we
do not pursue alternative formulations of the Newmark beta algorithm in any detail here.
The Newmark beta algorithm applied to translational velocities and accelerations makes use of
the familiar formula
s = ut + 12 at 2 (89)
Copyright ? 1999 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 45, 377–398 (1999)
394 F. A. MCROBIE AND J. LASENBY
which is strictly true if the acceleration is constant. Integration parameters and
are introduced,
weighting values at the beginning and end of a time-step to deal with the case of nonconstant
acceleration.
The displacement eld for the rod model in our geometric algebra formulation is given by the
multivector r + . Standard Newmark beta for translational motion applies equation (89) to r and
its derivatives. The natural extension to cope with rotational components would thus appear to be
to apply equation (89) to the rotation bivector and its derivatives. This, however, would not be
appropriate. As discussed in Section 3.7, interpolation between rotations is better achieved using
compound rotations in order to obtain a rotationally invariant interpolation. Simo and Vu Quoc
use Q and apply equation (89) to the three parameters (in our notation) Q;
B , and
˙ B . (In their
notation these are ; W, and A = Ẇ (Table 2, Reference 1).) Note in particular that Q̇ 6=
B . (In
their notation, this reads ˙ 6= W.) An alternative Newmark beta scheme is proposed in Cardona
and Geradin.15 However, for the reasons stated above, we do not wish to deliberate on the relative
merits of alternative formulations here but merely proceed with our translation of the Simo–Vu
Quoc algorithm into geometric algebra. Further discussion of subtleties regarding the updating of
rotational velocities and accelerations may be found in Criseld,27 Section 24.16.
At the end of the preceding time-step the rotation eld is specied by the rotation bivector
n . At some iteration (the ith, say) in the course of the Newton procedure to obtain the rotation
eld at the end of the time-step under consideration, the latest estimate of the rotation bivector is
(i) . We shall drop the superscript (i) for clarity. The aim is to obtain the next estimate (i+1) ,
which we shall denote by + d. The incremental quantity d is a small, but non-innitesimal,
bivector. (The need to use the symbol d for the small update quantities is unfortunate. However,
and have been used throughout for innitesimal quantities.)
A straight translation of the Simo–Vu Quoc update formulae (equations (3.4)–(3.7), Reference 1)
into our notation reads
˙ B (n ) +
Q + dQ = h
B (n ) + h2 ( 12 − )
˙ B ( + d) (90)
˙ B (n ) +
Q = h
B (n ) + h2 ( 12 − )
˙ B () (91)
Subtraction gives
˙ B () + 1 dQ
˙ B ( + d) =
(92)
h2
whence
˙ B) = 1
d(
dQ (93)
h2
B ( + d) =
B (n ) + h[(1 −
)
˙ B (n ) +
˙ B ( + d)] (94)
˙ B (n ) +
B () =
B (n ) + h[(1 −
)
˙ B ()] (95)
Copyright ? 1999 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 45, 377–398 (1999)
SIMO–VU QUOC RODS USING CLIFFORD ALGEBRA 395
For the weak form combined with a Newton iterative procedure, the exact statement of the
local momentum balance (containing the inertia term g) is rst represented by an approximation
on some approximating subspace. This is then projected onto some space of trial functions as is
usual in any weighted residual approach. The local linearization of this is then required for the
Newton procedure.
There is considerable freedom in the choice of trial functions: Simo and Vu Quoc select com-
pound rotations. In our notation, let Á(s) be an innitesimal compound rotation variation around
the conguration (s), such that rotors R() are varied to R(Á)R().
The rotational inertia term of the weak form of the momentum balance equations is thus
Z
G= (Á; g()) ds (104)
[0; L]
Copyright ? 1999 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 45, 377–398 (1999)
396 F. A. MCROBIE AND J. LASENBY
where
with
H = I(
˙ B ) + [I(
B );
B ] − [
B ; I(
B )] from (88) (107)
This exact linearization is now used in the Newton procedure to extrapolate the values at the
next time increment. Incremental quantities are nite. However, in our rst analysis, we compare
the results of the two derivations assuming incremental quantities are suciently small that we
may use any of the derived linear relationships to transform between them. Essentially this means
that incremental quantities are assumed so small that additive quantities prexed by d can be
written with a prex . (We do not, however, assume that compound rotations q or Q are small.)
Certainly, both derivations should agree in this limit.
First, we substitute the quantities
B = (
=h)Q and
˙ B = (1=h2 )Q using (96) and (93),
−1
and using Q = R̃(n ) q (q) (70) we obtain from (106)–(107) an expression for g as an
operator acting upon q.
This may be compared term by term with equation (3.15) (Reference 1) where it will be seen
that the nal terms (involving H) are in perfect agreement, but the rst term
Z Z
(Á; [g; q]) ds = (Á; R()[I(
˙ B ) − [
B ; I(
B )]; q]) ds (108)
[0; L] [0; L]
disagrees. In particular, this term operates on q whereas the corresponding term of Simo and
Vu Quoc operates on nt T (n(i) )n+1
(i)
, which corresponds to our bivector Q.
Correcting also a second, typographical error in Simo and Vu Quoc’s expression concerning
the location of the rst superscript V , we conclude that the nal integral in equation (3.15)
(Reference 1) should read (in their notation)
Z n oV
(i) (i) (i) (i)
· −n+1 {J An+1 + Wn+1 × J W n+1 }
[0; L]
n oV
1 (i) (i) (i) t (i) (i)
+ J − h
J W
n+1 + h
W J
n+1 n T (n ) n+1 dS (109)
h2 n+1
The primary dierence is in the location of the nal closing bracket, such that the rst expression
(i)
operates only on n+1 (our q) and not on tn T (n(i) )n+1
(i)
(our Q). The error is carried
through to their equation (4.3c), and was also incorporated into a literal implementation of the
Simo–Vu Quoc algorithm written by one of the authors of this paper. When corrected, improved
convergence at each time-step was obtained. (During the preparation of this manuscript, the authors
were kindly sent a preprint of Reference 26 by Jelenic and Criseld, where this error in Simo
and Vu Quoc’s tangent inertia matrix was also pointed out (Note A.2, Reference 26.) We are also
grateful to Dr. G. Jelenic for pointing out the incorrect placement of the rst superscript V , an
error not present in the geometric algebra derivation.)
Copyright ? 1999 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 45, 377–398 (1999)
SIMO–VU QUOC RODS USING CLIFFORD ALGEBRA 397
Remark 4. We comment nally that if incremental quantities are considered nite there is
some choice of which increments to relate by linear relations and which to relate by more exact
expressions. We do not wish to discuss this here. The primary purpose of this paper is to show
how the algorithm may be more easily understood, and a detailed analysis of the possible update
procedures would presuppose acceptance of the underlying Q;
B ;
˙ B Newmark scheme.
5. SUMMARY
In this paper we have laid the foundations for a complete reformulation of computational mechanics
in the language of geometric algebra. The motivation for this is our claim that geometric algebra
is considerably easier to learn and to manipulate than the dierential topology used for current
advanced formulations of computational mechanics, whilst at the same time possessing the full
mathematical power. Our claim that geometric algebra oers a further possible advantage in that
it readily generalizes to higher dimensions is something that we cannot yet see how to exploit in
the eld of computational mechanics. However, since no extra work was required, in Section 3
we presented some of the fundamental formulae for rotational kinematics in the full n-dimensional
setting. This work may nd applications in the higher-dimensional phase spaces common in other
disciplines.
In Section 4, we presented a reformulation of one of the terms in the Simo–Vu Quoc rod
algorithm. Although this algorithm has many evident short-comings in terms of the numerical
integration scheme and certain physical limitations as a mathematical model of a rod, it provides
the simplest clear starting point for anyone wishing to enter the eld of computational mechanics
using the dierential topology approach. We considered only the rotational inertia term, and thus
much of the work is more widely applicable than to rods alone. Using the methods presented
here the rest of the algorithm follows comparatively easily. Curvatures of the rods, for example,
are bivectors dened in much the same way as the angular velocities, but using the rotation
eld and its derivatives with respect to undeformed arc-length s, rather than t. (For example,
ZS ≡ − 2R0 R̃ = (0 )) To see just how far this reformulation reaches into fundamental structural
mechanics, note that in this description, a bending moment diagram is the graph of a bivector-
valued function.
The intention in this paper was to demonstrate that the simple language of geometric algebra is
capable of replicating the full power of the dierential topology formulation. In the course of that
demonstration, we concluded that the dierential topology statement in the original paper, despite
the daunting austerity of its derivation, in fact contains an error. It is doubtful that we would have
detected this error if we had not had the simpler tools of geometric algebra available to us.
Having presented the foundations for this approach it is hoped that we have cleared the way
for more incisive incursions into advanced computational mechanics in the future, not only by the
authors, but by means of this simpler language, by a wider spectrum of engineers than currently
possible.
ACKNOWLEDGEMENTS
Allan McRobie is grateful for support from the EPSRC Applied Nonlinear Mathematics Pro-
gramme, and Joan Lasenby is grateful to the Royal Society of London for a University Research
Fellowship.
Copyright ? 1999 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 45, 377–398 (1999)
398 F. A. MCROBIE AND J. LASENBY
REFERENCES
1. J. C. Simo and L. Vu-Quoc, ‘On the dynamics in space of rods undergoing large motions—a geometrically exact
approach’, Comput. Meth. Appl. Mech. Engng., 66, 125 (1988).
2. D. Hestenes, New Foundations for Classical Mechanics, Reidel, Dordrecht, 1986.
3. D. Hestenes and G. Sobczyk, Cliord Algebra to Geometric Calculus, Reidel, Dordrecht, 1984.
4. R. Abraham and J. E. Marsden, Foundations of Mechanics, 2nd edn., Benjamin=Cummings, Reading, MA, 1978.
5. V. I. Arnol’d, Mathematical Methods of Classical Mechanics, 2nd edn., Springer, Berlin, 1989.
6. J. E. Marsden and T. J. R. Hughes, Mathematical Foundations of Elasticity, Prentice-Hall, Englewood Clis, N.J.,
1983.
7. S. Flugge (ed.), Handbuch der Physik, Springer, Berlin, 1960.
8. J. C. Simo and L. Vu-Quoc, ‘A three-dimensional nite-strain rod model. Part II. Computational aspects’, Comput.
Meth. Appl. Mech. Engng., 58, 79 (1986).
9. J. C. Simo, ‘On a stress resultant geometrically exact shell model. Part VII: Shell intersections with 5=6 DOF nite
element formulations’, Comput. Meth. Appl. Mech. Engng., 108, 319–339 (1993).
10. J. C. Simo, ‘The (symmetric) Hessian for geometrically nonlinear models in solid mechanics: intrinsic denition and
geometric interpretation’, Comput. Meth. Appl. Mech. Engng., 96, 189 –200 (1992).
11. J. C. Simo, N. Tarnow and M. Doblare, ‘Nonlinear dynamics of three-dimensional rods: exact energy and momentum
conserving algorithms’, Int. J. Numer. Meth. Engng., 38, 1431–1473 (1995).
12. S. F. Gull, C. J. L. Doran and A. N. Lasenby, Dynamics, in W. E. Baylis (ed.), Geometric (Cliord) Algebras in
Physics, Birkhauser, Boston, 1996.
13. C. Sansour and H. Bednarczyk, ‘The Cosserat surface as a shell model, theory and nite-element formulation’, Comput.
Meth. Appl. Mech. Engng., 120, 1 (1995).
14. G. Jelenic and M. A. Criseld, ‘Objectivity of strain measures in geometrically exact 3D beam theory and its nite
element implementation’, Proc. R. Soc. Lond., (1998), submitted.
15. A. Cardona and M. Geradin, ‘A beam nite element non-linear theory with nite rotations’, Int. J. Numer. Meth.
Engng., 26, 2403–2438 (1988).
16. A. Cardona and M. Geradin, ‘Time integration of the equations of motion in mechanism analysis’, Comput. Struct.,
33(3), 801–820 (1989).
17. A. D. Challinor, A. N. Lasenby, S. S. Somaroo, C. J. L. Doran and S. F. Gull, ‘Tunnelling times of electrons’, Phys.
Lett. A227, 143 (1997).
18. A. N. Lasenby, C. J. L. Doran and S. F. Gull, ‘Gravity, gauge theories and geometric algebra’, Phil. Trans. R. Soc.
Lond., A356, 487–582 (1998).
19. C. J. L. Doran, ‘Integral equations and Kerr-Schild elds. I. Spherically-symmetric elds’, Class. Quantum. Grav.
(1997), submitted.
20. ABAQUS Theory Manual, (Version 5.4). Hibbitt, Karlsson & Sorensen, Inc., Pawtucket, RI, 1994.
21. J. R. Wertz, (ed)., Spacecraft Attitude Determination and Control, Kluwer, Dordrecht, 1978.
22. J. Argyris, ‘An excursion into large rotations’, Comput. Meth. Appl. Mech. Engng., 32, 85 –155 (1982).
23. S. F. Gull, A. N. Lasenby and C. J. L. Doran, ‘Imaginary numbers are not real—the geometric algebra of spacetime’,
Found. Phys., 23(9), 1175 (1993).
24. N. M. Newmark, ‘A method of computation for structural dynamics’, J. Eng. Mech. ASCE, EM3, 85, 67–94 (1959).
25. H. M. Hilber, T. J. R. Hughes and R. L. Taylor, ‘Improved numerical dissipation for time-integration algorithms in
structural dynamics’, Earth. Engng. Struct. Dyn., 5, 283–292 (1977).
26. G. Jelenic and M. A. Criseld, ‘Interpolation of rotational variables in nonlinear dynamics of 3D beams’, Int. J.
Numer. Meth. Engng., (1997), submitted.
27. M. A. Criseld, Nonlinear Finite Element Analysis of Solids and Structures, Vol. 2, Advanced Topics, Wiley,
Chichester, 1997.
Copyright ? 1999 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng. 45, 377–398 (1999)