275-Consistent Quaternion Interpolation For Objective Finite Element Approximation of Geometrically Exact Beam

Comput. Methods Appl. Mech. Engrg.
198 (2008) 555–571
Contents lists available at ScienceDirect
Comput. Methods Appl. Mech. Engrg.

journal homepage: www.elsevier.com/locate/cma
Consistent quaternion interpolation for objective finite element

approximation of geometrically exact beam
S. Ghosh, D. Roy *
Structures Lab, Department of Civil Engineering, Indian Institute of Science, Bangalore, Karnataka 560 012, India
a r t i c l e i n f o a b s t r a c t
Article history: We explore an isoparametric interpolation of total quaternion for geometrically consistent, strain-objec-
Received 4 April 2008 tive and path-independent finite element solutions of the geometrically exact beam. This interpolation is
Received in revised form 16 June 2008 a variant of the broader class known as slerp. The equivalence between the proposed interpolation and
Accepted 4 September 2008
that of relative rotation is shown without any recourse to local bijection between quaternions and rota-
Available online 18 September 2008
tions. We show that, for a two-noded beam element, the use of relative rotation is not mandatory for
attaining consistency cum objectivity and an appropriate interpolation of total rotation variables is suf-
Keywords:
ficient. The interpolation of total quaternion, which is computationally more efficient than the one based
Geometrically exact beam
Finite rotation
on local rotations, converts nodal rotation vectors to quaternions and interpolates them in a manner con-
Rotation manifold sistent with the character of the rotation manifold. This interpolation, unlike the additive interpolation of
Tangent space total rotation, corresponds to a geodesic on the rotation manifold. For beam elements with more than two
Relative rotation nodes, however, a consistent extension of the proposed quaternion interpolation is difficult. Alterna-
Objectivity tively, a quaternion-based procedure involving interpolation of relative rotations is proposed for such
Path-independence higher order elements. We also briefly discuss a strategy for the removal of possible singularity in the
interpolation of quaternions, proposed in [I. Romero, The interpolation of rotations and its application
to finite element models of geometrically exact rods, Comput. Mech. 34 (2004) 121–133]. The strain-
objectivity and path-independence of solutions are justified theoretically and then demonstrated through
numerical experiments. This study, being focused only on the interpolation of rotations, uses a standard
finite element discretization, as adopted by Simo and Vu-Quoc [J.C. Simo, L. Vu-Quoc, A three-dimen-
sional finite rod model part II: computational aspects, Comput. Methods Appl. Mech. Engrg. 58 (1986)
79–116]. The rotation update is achieved via quaternion multiplication followed by the extraction of
the rotation vector. Nodal rotations are stored in terms of rotation vectors and no secondary storages
are required.
Ó 2008 Elsevier B.V. All rights reserved.
1. Introduction ters adopted. Over the years, several such parametrizations have
been adopted, viz., direction cosine matrices or orthogonal matri-
Geometrically exact 3D beam theories and their finite element ces, Euler angles, quaternions or Euler parameters, Rodrigues’
(FE) implementations have been extensively studied over the last parameters or Gibbs vectors, conformal rotation vectors and rota-
few decades. Simo and and Vu-Quoc [2–4] have provided a beam tion vectors. Detailed accounts of these parametrizations and their
theory and FE formulation that does not affect the geometric exact- mutual relations may be found in [13–16]. Different parametriza-
ness even under finite rotational deformations in 3D. This ap- tions have their respective advantages and disadvantages depend-
proach has been followed up by several authors [5–12], who ing on the application. A rotation vector parametrization has the
have contributed in setting up a theoretical and computational ba- following advantages. First, it is easier to implement in the FE con-
sis for the geometrically exact 3D beam theory. text. Then, it requires only three parameters and has a simple geo-
Parametrization of finite rotation is an important issue, espe- metric interpretation. Moreover, every rotation corresponds to a
cially as practical implementations of computational schemes rotation vector, notwithstanding the non-uniqueness associated
involving large rotations are crucially dependent on the parame- with it. Quaternion-based rotations constitute one of the most con-
venient representations of the double cover of SO(3) by S3 .1 Besides
* Corresponding author. Tel.: +91 80 2293 3129; fax: +91 80 2360 0404.
E-mail addresses: susanta@civil.iisc.ernet.in (S. Ghosh), royd@civil.iisc.ernet.in 1
SO(3) is the Special Orthogonal Group. S3 is a unit sphere, centered at the origin,
(D. Roy). in R4 and is also known as unit 3-sphere.
0045-7825/$ - see front matter Ó 2008 Elsevier B.V. All rights reserved.
doi:10.1016/j.cma.2008.09.004
556 S. Ghosh, D. Roy / Comput. Methods Appl. Mech. Engrg. 198 (2008) 555–571
less number of parameters, quaternions have several other advanta- Steinmann [22] have eliminated the explicit reference to finite
ges over rotation matrices. Details about quaternions can be found in rotation tensors and proposed new formulations of a geometrically
[14,15]. Since the finite rotation evolves over a nonlinear manifold, exact beam with two director vectors. These formulations are
the rotation vector is not generally amenable to vector-space opera- strain-objective.
tions. A non-adherence to such characteristics may lead to certain In view of the limitations of the updated Lagrangian formula-
anomalies, e.g. a compromise on geometric exactness, path-depen- tion, our present aim is to look for efficient means of retaining geo-
dence of solutions, non-objective strains etc. metric consistency along with objectivity while explicitly using the
In Simo and Vu-Quoc [2], rotations are parametrized by the rotation variables. In particular, we use the Eulerian formulation
nine-parameter rotation matrix and FE nodal rotation updates (similar to [2,5]) with an interpolation of nodal total-quaternions.
are through matrix multiplications. In a major departure from The proposed total quaternion interpolation for the two-noded ele-
[2], Cardona and Géradin [5] have parametrized rotations as rota- ment is computationally more efficient with respect to the interpo-
tion vectors. Here FE formulations are categorized as (i) Eulerian, lation of local rotations. Moreover, in view of its computational
(ii) total Lagrangian or (iii) updated Lagrangian, depending upon efficiency over the existing interpolation of local rotations [19],
whether the incremental rotation vector belongs to the tangent we also consider quaternion-based interpolation of local rotations
space at current rotation, at identity rotation or at the previously for elements with more than two nodes. Steps to obtain the curva-
converged rotation, respectively. During FE implementations via ture using the interpolated rotations are presented in a way consis-
Eulerian [2], total Lagrangian [5] and updated Lagrangian tent with the nonlinear characteristics of the rotation manifold.
[5,17,18] formulations, the rotation variables interpolated along The study emphasizes on maintaining the geometric consistency,
the length of the beam are the iterative rotation vectors (measured objectivity and path-independence via rotation vector parametri-
with respect to the configuration of the last iteration), total rota- zation at no extra effort. We use the quaternion product to update
tion vectors (measured with respect to the initial configuration) nodal rotations and thus avoid matrix multiplications. By working
and incremental rotation vectors (measured with respect to the with the Eulerian framework based on (total) rotation vectors,
last converged configuration), respectively. In the total Lagrangian requirements of secondary storages are also eliminated.
formulation, the incremental rotation vectors in different tangent
spaces are transferred to the tangent space at identity via a linear
transformation, presently called the tangential transformation. 2. Geometrically exact beam theory
Due to the well known singularity of the tangential transformation,
the total Lagrangian formulation is restricted to rotations less than Strain measures and governing equations of the 3D geometri-
the full angle. This restriction is removable via the updated cally exact beam theory are summarized in this section. Details of
Lagrangian formulation, wherein we need to store the last con- the kinematic variables and configuration space can be found in
verged configuration. The formulation via incremental rotation [2,5]. Consider an initially straight beam of cross-section X and
vectors [17] uses a similar concept. Since rotations at different length L, with reference configuration at B ¼ X ½0; l R3 . The
nodes (or Gauss points) of an element are likely to be different in beam kinematics are assumed to be governed by the beam cen-
the deformed configuration, rotation increments at those points tre-line and orientations of undeformable cross-sections of the
would belong to different tangent spaces. Even as these inconsis- beam (Fig. 1). Let fe1 ; e2 ; e3 g and fE1 ; E2 ; E3 g be the inertially fixed
tencies remain in FE formulations involving both Eulerian and up- (world) and material Cartesian frames, respectively. The material
dated Lagrangian formalisms, converged estimates of rotational frame may be attached to the beam cross-section and its axes ori-
deformations beyond the full angle are still obtainable whilst ented along the principal axes of inertia of the body at its initial
allowing additive interpolations of rotation variables with different configuration. Let the beam be initially so oriented that fe1 ; e2 ; e3 g
base rotations. However, as shown in [19], the additive interpola- and fE1 ; E2 ; E3 g remain parallel. The orientation of the moving
tion results in non-objective strain measures under superposed ri- frame attached to the cross-section at position s is specified
gid-body rotations and path-dependent solutions for iterative and through an orthogonal matrix R(s) such that ti ¼ REi ; i ¼ 1; 2; 3; s
incremental formulations. Indeed, loss of frame-invariance and denotes the length parameter. ft 1 ; t 2 ; t 3 g constitute the body-fixed
path-independence is attributable to the inconsistent interpolation Cartesian frame. At every deformed configuration, t 2 and t3 are
of rotation variables. As a remedy, Crisfield and Jelenić [19,20] have aligned with the principal axes of inertia and the vector t1 remains
suggested interpolations of local rotations (or relative rotations). normal to the cross-section. Thus the kinematic assumption gives
Here the total rotation is decomposed into a reference rotation the deformation map as x ¼ x0 ðsÞ þ X i t i ðsÞ; i ¼ 2; 3. Here X i ði ¼
and a local rotation. While the reference rotation moves with the 2; 3Þ are coordinates in the material frame and x is the position vec-
beam element (like convected coordinates), local rotations are tor of any point in the beam. The current configuration UðsÞ of the
interpolated to obtain the rotation field within an element. This beam is completely characterized through the position of the cent-
formulation is referred to as the invariant-formulation. Due to the roidal line and the orientation of the moving frame:
complexity in interpolating relative rotations and the associated
UðsÞ ¼ ðx0 ðsÞ; RðsÞÞ;
difficulties in linearization, an approximated geometric tangent
stiffness is used in [20,21]. In the context of frame-invariant for- x0 : ½0; L Rþ ! R3 ;
mulations, a review of interpolation of rotation variables is pro- R : ½0; L Rþ ! SOð3Þ:
vided in [1]. All of these articles, addressing objectivity, use
orthogonal matrix parametrization. Only the work of Ibrahimbeg- Even though the rotational component of U is written as a matrix
ović and Taylor [18] addresses the issue of objectivity while using (tensor) R, it is presently parametrized by the total rotation vector
the rotation vector parametrization. They use an updated Lagrang- (denoted as W). A rotation tensor R represents a rotation with re-
ian formulation, which is generally neither path-independent nor spect to a vector WI through an angle kWI k. This rotation vector is
frame-invariant. They show that frame-invariance is achievable related to the rotation tensor via the exponential mapping:
provided that the applied load is treated as a follower load; how- 2
ever this is tantamount to changing the problem itself. Moreover e I Þ ¼ I þ sin kWI k W
R ¼ expð W e I þ 2 sinðkWI k=2Þ W e 2: ð1Þ
I
kWI k kWI k
it takes some extra effort to handle the follower loading and needs
secondary storage. In a different approach from Simo’s geometri- Throughout this text, ð~
Þ will denote the skew-symmetric tensor
cally exact theory, Romero and Armero [23] and Betsch and corresponding to the axial vector ðÞ. The subscript I indicates that
S. Ghosh, D. Roy / Comput. Methods Appl. Mech. Engrg. 198 (2008) 555–571 557
t3
t1
Φ
x0
e3 E3
O X
e1 E1
Fig. 1. Initial and deformed configurations of the beam. Only two dimensions are shown for clarity, the third dimension is perpendicular to the XZ plane.

W is represented in the non-rotated material frame fEi g, i.e. when sin h sin h 1 cos h
TðWÞ ¼ Iþ 1 ee ~e ; h ¼ kWk
base rotation is identity. Thus WI belongs to the tangent space at h h h
identity. However, for notational convenience, the subscript I is W
henceforth dropped. We do not present details of the rotation man- and e ¼ : ð4Þ
h
ifold (SO(3)) and its parametrizations and closely follow the nota-
The tangential transformation T [5,7,17] has singularities for
tions used in [5].
h ¼ 2kp; k ¼ 1; 2; . . .. Indeed, TðWÞ is only an approximation to the
The strain measures are derived from expressions of internal
exact transformation. This is apparent form the derivation provided
stress power equation and kinematics of the beam [2,3,5]. The
in [5].
same however is also obtainable through the differential equation
The variational material strain measures are obtained by taking
of equilibrium and the principle of virtual work ([19]). The strain
directional derivatives along the admissible variations as
conjugates to the resultant force and resultant moment are de-
noted, respectively, as C and K in the material representation. They " t dðdx0 Þ #
dC R ds þ Rt dðdx 0Þ
dH
are given by ½de ¼ ¼ dðdHÞ
ds
dK þ K dH
ds
dx0 d " #
e ¼ ½Ct ; K t t where C ¼ Rt E1 and K ¼ Rt R: ð2Þ t d
R ds t dx0
R ds dx0
ds ds ¼ ¼ BðUÞdU: ð5Þ
0 IdþK e dH
For the beam centre-line, there exists an axis of rotation at each ds
point on the curve. This axis is given in spatial coordinates by k Here BðUÞ denotes the strain operator. The choice of material
and in body coordinates by K ¼ Rt k. Within the variational description has the benefit of circumventing the Lie-derivative dur-
framework, let dUðsÞ ¼ ½dx0 ; dHt denote an admissible variation ing linearization. The weak form of the beam equilibrium equation
of the deformation. Thus, following the kinematics of the beam, may be written as
the perturbed configuration corresponding to U ¼ ðx0 ; RÞ is given Z
by GðdU; UÞ ¼ ðde r dU F ext Þds ð6Þ
½0;L
x0 ¼ x0 þ dx0 e :
and R ¼ R exp½d H ð3Þ ¼ Gint ðdU; UÞ Gext ðdUÞ ¼ 0: ð7Þ
Left translation is used to obtain the perturbed rotation. Variation of The external forces are assumed to be conservative so Gext is not
the rotation tensor is obtained via its directional derivative along dependent on the current configuration. In the linear elastic case,
the superposed infinitesimal rotation dH as the stress resultants rðsÞ can be obtained from e by simply multi-
plying with the elasticity tensor.
d e:
dR ¼ R ¼ Rd H
d¼0 2.1. Linearization of the weak form
This is also obtainable by perturbing R via the variation of W as
The weak form (Eq. (6)) is linearized to obtain the tangent stiff-
d d e Þ ¼ R TðWg
e þ dW ness matrix. The linearized weak form at the current configuration,
dR ¼ R ¼ expð W ÞdW ;
d ¼0 d ¼0 UðsÞ ¼ ðx0 ; RðsÞÞ, is written as
TðWÞ was given by [5] to relate dH and dW as L½GðdU; UÞ ¼ GðdU; UÞ þ D½GðdU; UÞ DU ¼ 0 ð8Þ
dH ¼ TðWÞdW DUðsÞ ¼ ðDx0 ; DUR Þ denotes the incremental material generalized
where displacements. Depending upon the type of material rotational
increment, the linearized operators are different. Based on the tan- DHh ðnÞjLe N i ðnÞDHi : ð13Þ
gent space to which the material rotational increment DUR belongs,
formulations are divided into three categories [5]: The inaccuracy due to this approximation may apparently be made
smaller through mesh refinement (thereby forcing the function
(1) Eulerian formulation, where the incremental rotations ðTðWh ÞÞTðWi Þ1 of Eq. (13) closer to identity). A more precise
DH 2 left T R SOð3Þ description of the error involved with the above interpolation is
(2) total Lagrangian formulation, where the incremental rota- provided in Section 3.2.2. We must however emphasize that all for-
tions DW 2 T I SOð3Þ mulations that allow rotations beyond 2p have to use some approx-
(3) updated Lagrangian formulation, where the incremental imation of this kind at some stage of implementation. This is
rotations DWinc 2 left T Rref SOð3Þ. evident as the rotation manifold is compact and a smooth global
1-1 representation via rotation vectors (which are elements of an
These are linearizations in left T R SOð3Þ, T I SOð3Þ and left T Rref SOð3Þ, open set) is impossible. For instance, in the updated Lagrangian set-
respectively.2 R and Rref correspond, respectively, to the current ting, the increment is measured with respect to the last converged
and last converged/known configurations. Since we use the Euleri- configuration. But, following deformation, one has two different
an formulation, this is briefly described in Appendix B. Lineariza- rotations at two different nodes of an element. Rotation increments
tion of Gint is decomposed into a sum of material and geometric of these two nodes belong to two different tangent spaces with the
tangent stiffnesses, derivable, respectively, from linearizations of two nodal rotations as base points. Hence the problem of handling
r and de of Eq. (6). The expressions are available in [2,5]. two vectors in two different tangent spaces persists in this setting. A
similar approximation in the strain-configuration matrix also arises
2.2. Interpolations of displacements and rotations in the invariant-formulation given in [20]. Thus, for effectively deal-
ing with rotations beyond 2p, the approximation as in Eq. (13) is
For finite element implementations, both displacements and widely used. Use of this approach for computing the strain-config-
rotations need to be interpolated. While interpolation of vectors uration matrix enables an easy derivation of the tangent stiffness.
is additive and straightforward, interpolation of rotation remains
a difficult yet crucial part. A review of interpolations of rotations 2.2.2. Interpolations of local rotations
is available in [1]. In order to bring into focus the present contribu- In the interpolation methods enumerated above, rotations of
tion, we briefly discuss some existing rotation interpolations and the current frame, measured with respect to some chosen refer-
then try to critically assess them for motivating the propositions ence frame, are interpolated. In [19], Crisfield and Jelenić have pro-
of this study. vided an approach that marks a paradigm shift in interpolations of
rotations and showed that the previous works are not frame-indif-
2.2.1. Interpolations of total, iterative and incremental rotation ferent. In particular, they have proposed interpolations of relative
vectors (local) rotations, which are measured with respect to some refer-
For linearization on T I SOð3Þ, rotations may be interpolated like ence rotation in a beam element. We describe the procedure in
vectors the context of a two-noded element with nodal rotations R1 and
R2 . Taking the first node as the reference point, the interpolated
xh0 ðnÞ ¼ N i ðnÞ x0i ; where i ¼ 1; 2; . . . ; ne ð10Þ rotation is given by
h s
W ðnÞ ¼ N i ðnÞ Wi ð11Þ
Rh ðsÞ ¼ R1 exp ~ 12
/ ~ 12 Þ ¼ Rt R2 :
where expð/ 1 ð14Þ
N i denotes the linear isoparametric interpolation function based at L
node i (e.g. ½N 1 ðnÞ; N2 ðnÞ ¼ ½ð1 nÞ=2; ð1 þ nÞ=2) and Einstein’s sum- Spurrier algorithm is used to extract /12 from Rt1 R2 ; owing to small-
mation convention is used (whenever the summation-sign is not ness of relative nodal rotations of elements, the extraction does not
explicitly employed) extending over nodes 1 to ne . n denotes the cause any singularity. The main advantage of this interpolation is
independent variable in natural coordinate corresponding to the that it leads to an objective formulation under superposed rigid
beam arc-length parameter s. Similar interpolation is valid for vir- body rotations. The linearization of Eq. (14) is however quite in-
tual generalized displacements ðdx0 ; dWÞ and increments of general- volved [20]. Instead of the first node, we may use relative rotations
ized displacements ðDx0 ; DWÞ. In [5], it is indicated that this form of with respect to any other reference node or a non-nodal reference
interpolation is valid since Wi ; dWi and DWi belong to the same vec- point.
tor space at T I SOð3Þ. The same interpolation is also used for incre-
mental rotations within an updated Lagrangian setting (see 2.2.3. Non-orthogonal direct interpolation of rotation tensors
[5,17]), where the incremental rotations are measured with respect In a different approach [22,23] that maintains objectivity in the
to the last converged rotation. FE discretization but does not retain orthogonality, the following
This form of interpolation is however not applicable to rotational interpolation is used:
variables H; dH and DH lying on T R SOð3Þ. For different i, dHi belongs
to different tangent spaces denoted by T Ri SOð3Þ. Hence it is not ame- Rh ðsÞ ¼ Ni ðsÞRi
nable to interpolations valid in Rn . However, we may write Ni ðsÞ denotes linear interpolation functions e.g. ½N1 ðsÞ; N2 ðsÞ ¼
X
ne ½ð1 sÞ=L; s=L. It is readily verifiable that, under superposed rigid
DHh ðnÞjLe ¼ TðWh Þ Ni ðnÞ TðWi Þ1 DHi : ð12Þ body motion, it satisfies the objectivity (see Eq. (36)). The key idea
i¼1 is to impose the orthogonality constraint of the rotation field only at
Note that the iterative rotations have been interpolated additively nodes and relax it at other points. The theory has also been modified
in [2], i.e. to cater to the lack of orthogonality of the rotation field.
2.2.4. Non-orthogonal direct interpolation followed by

2
Left-invariant (material) vector field on SOð3Þ at R is defined as orthogonalization
left T R SOð3Þ
e R ¼ ðR; H
¼ fH e Þj with left translation R H
e ; R 2 SOð3Þ; H
e 2 SOð3Þg ð9Þ To set right the non-orthogonality almost everywhere (except-
e R is an element of material vector field left T R SOð3Þ. For notational convenience, the ing the nodes), a new interpolation of quaternions is proposed in
H
subscript (R) is omitted. [1]:
1
Q h ðsÞ ¼ Ni ðsÞQ i ð15Þ sin nÞX 2
ð1
c1 ðnÞ ¼ ; ð19Þ
sin X
Unfortunately, such additive interpolation is not valid for quater-
sin 12 ð1 þ nÞX
nions. Upon conversion by formula (A.3), the resulting quaternion c2 ðnÞ ¼ : ð20Þ
will produce an inaccurate orthogonal tensor. Further, it is not free sin X
from singularity [1]. Even though this interpolation provides objec- Thus the interpolated quaternion is given by
tivity, very significant errors occur in the computation of stress and
sinðNi ðnÞXÞ
reaction forces, particularly for coarse meshes. The review paper [1] Q h ðnÞ ¼ Qi i ¼ 1; 2; ð21Þ
on interpolation of rotation concludes that, amongst the existing sin X
methods, the method of interpolation of local rotations by [19] where cos X ¼ Q 1 Q 2 . Henceforth, we will denote this interpola-
has by far the best performance. tion as slerpðn; Q 1 ; Q 2 Þ (slerp is a short form of Spherical Linear
Interpolation) following the computer graphics literature [24,25],
where similar interpolations are often used. In order to derive the
3. Present implementation above interpolation formula, at first the condition of unit length is
enforced on interpolated quaternion, then the coefficients of given
In this section, we will discuss different interpolation proce- unit-quaternions are obtained from considerations of either ana-
dures and rotation update through quaternion. In [20], the author lytic geometry or trigonometry. The above construction ensures
appears to suggest that the total rotation has to be decomposed in that slerp gives the shortest path between its quaternion end points.
a reference and a local rotation to achieve objectivity along with Once the interpolated quaternion field is available, rotation ten-
geometric consistency. We note that it is not essential. Toward this sors at Gauss points may be obtained by using Eq. (A.2). Note that
we propose isoparametric interpolation of two nodal quaternions, this interpolation yields unit quaternions as long as nodal quater-
also known as slerp. slerp was briefly mentioned in [1], but, to the nions are unit and hence normalization of quaternions is not re-
best of our knowledge, the objectivity cum consistency through quired. Even though Q 1 and Q 1 represent the same rotation,
slerp and equivalence between slerp and interpolation of relative slerpðn; Q 1 ; Q 2 Þ and slerpðn; Q 1 ; Q 2 Þ do not represent the same
rotation is not available in the literature. Moreover, the nine- rotation field. Thus we choose the sign ‘‘1” on Q 2 so that
parameter orthogonal matrix representation and the use of Spurri- Q 1 1 Q 2 P 0. This helps maintain an acute angle between these
er algorithm to extract the local rotations are computationally two quaternions (i.e. ensures that p=2 6 X 6 p=2) and avoids ex-
expensive. Therefore a cheaper alternative through quaternion tra spinning caused by interpolated rotations. It is evident that this
interpolation is worth investigating. condition remains valid provided the angle of relative rotation be-
The rotation update for orthogonal tensor parametrization is tween the two nodal rotations is less than the full angle. We note
straightforward as there is no non-uniqueness associated with that the formula (Eq. (21)) reduces to the corresponding symmetric
the nine-parameter representation. However, this is not the case formula for linear interpolation as X ! 0.
with the three-parameter rotation vector representation. The up- This interpolation may be extended to beam elements with
dated rotation vector need not be and indeed not the only rotation more than two nodes using the same formula (21) between two
vector to represent the sought rotation. Therefore the update pro- consecutive nodes. However the resulting interpolation will nei-
cedure for vectors must be handled more carefully than that for ther be isoparametric. However, such interpolations will remain
matrices. In Section 3.6, we propose an updating strategy that ob- rotation manifold consistent for all values of n. Further discussion
tains the (updated) rotation vector of magnitude bounded by p. on this issue will be presented in Section 3.2.
For obtaining the weak form and its linearization, the derivative
3.1. Interpolation based on quaternion of the rotation field is needed. Using the quaternion representation,
one may thus invoke quaternion calculus. On the other hand, we
As an alternative to the interpolation of local rotations, we pro- may also take the derivative of Eq. (21) to obtain derivative of qua-
pose an isoparametric interpolation based on quaternions for com- ternion, i.e.:
puting the rotation field within a two-noded element. Let the nodal
0 X2
cosðNi ðnÞXÞ 1 d
rotations be depicted by rotation vectors W1 and W2 . First, using Eq. Q h ðnÞ ¼ X Ni ðnÞ Q i ; i ¼ 1; 2 ð22Þ
(A.4), we convert the rotation vectors to corresponding quater- i¼1
sin X J dn
nions Q 1 and Q 2 . The angle X between two quaternions can be ob-
ðÞ0 denotes derivative with respect to s and J denotes the jacobian of
tained as cos X ¼ Q 1 Q 2 . The interpolated quaternion may be
the isoparametric transformation. Since we have not used the unit
written as
quaternion condition in finding the derivative, we will use Eq.
Q h ðnÞ ¼ ci ðnÞ Q i i ¼ 1; 2; ð16Þ (A.3) for obtaining the derivative of the rotation tensor. Thus we get
0
where ci ðnÞ; i ¼ 1; 2 are real-valued functions with domain [1, 1], 0 Qh Qh 2 0 0
such that c1 ð1Þ ¼ 1; c1 ð1Þ ¼ 0; c2 ð1Þ ¼ 0; c2 ð1Þ ¼ 1. We want Rh ¼ 2 ðI þ Rh Þ þ ð2qh0 qh0 I þ qh qh

kQ h k kqh k2
the interpolated quaternion Q h ðnÞ to lie on S3 (unit three-sphere).
þ qh qh þ qh0 q~h þ qh0 q~h Þ
0 0 0
The idea is that the angle X between these two unit-quaternions ð23Þ
should be traversed at uniform speed. Therefore, the quaternion h0 0 0
Q is given by ðqh0 ; qh Þ
in the above relation. The curvature
dot product between Q h ðnÞ and Q 1 should be cos 12 ð1 þ nÞX and e h ¼ Rh t Rh0 may now be computed using the above formula obtain-
K
Q h ðnÞ and Q 2 should be cos 12 ð1 nÞX . From this we get
ing the derivative of the rotation tensor. Note that entire process
avoids computations of relative rotations and inverse trigonometric
1
cos ð1 þ nÞX ¼ Q h ðnÞ Q 1 ¼ c1 ðnÞ þ c2 ðnÞ cos X; ð17Þ functions.
2

1
cos ð1 nÞX ¼ Q h ðnÞ Q 2 ¼ c2 ðnÞ þ c1 ðnÞ cos X: ð18Þ 3.2. Comparisons of different interpolations
2
Solving these two equations for the two unknowns, ci ’s are given by We will consider the following three approaches:
(a) interpolation of relative rotations, tangent space onto elements in the manifold along geodesic curves
(b) interpolation of quaternions and (see Section 7.3, ‘‘The Exponential Map”, of [26]). Due to the equiv-
(c) additive interpolation of rotation vectors. alence between slerp and interpolation of relative rotation based
on two nodal rotations, slerp also corresponds to the geodesic on
To start with, we show an equivalence between the proposed SO(3) between two end rotations. This assumes importance since
interpolation and that of relative rotations; i.e. they provide identi- it inherits the important shortest-path property of the linear inter-
cal rotation fields along the beam element. It will also be shown polation in Euclidean spaces.
that (a) and (b) correspond to a geodesic on SO(3) between two
end rotations, which is not the case for (c). Subsequently, differ- 3.2.2. Error in additive interpolation
ences between these two interpolations and the additive interpola- We now investigate the differences between interpolation of
tion are described. relative rotation (i.e. type (a) and additive interpolation of rotation
vector (i.e. type (c). First, we recapitulate the Baker–Campbell–
Hausdorff (BCH) formula. Let expðwÞ ~ ¼ expðu ~ Þ expðv
~Þ, where
3.2.1. Equivalence between interpolation of relative rotations and
~; v
u ~; w
~ 2 soð3Þ. The first few terms of the BCH expansion are given
interpolation of quaternions
below
We will show that the curve segment slerpðQ 1 ; Q 2 ; nÞ is equiva-
g
lent to R1 expðN 2 ðnÞ DW Þ, where DW is the rotation vector for Rt1 R2 . 1 1 1
~ ¼u
w ~þv ~; v
~ þ ½u ~ þ ~; ½u
½u ~; v
~ þ ~; ½v
½v ~; u
~ þ HOT;
Toward this we use Q 1 1 Q 2 ¼ ðq01 q02 þ ðq1 ; q2 Þ; q01 q2 q02 q1 2 12 12
q1 q2 Þ and cos X ¼ Q 11 Q 2 ¼ q01 q02 þ ðq1 ; q2 Þ; X is the ‘angle’ where ‘HOT’ denotes ‘higher order terms’ and ½; denotes Lie brack-
between Q 11 and Q 2 . Now we have et. The interpolated relative rotation field (see Eq. (14)) is given by
ðq01 q2 q02 q1 q1 q2 ; q01 q2 q02 q1 q1 q2 Þ e 12 Þ;
Rh ðnÞ ¼ R1 expðN2 ðnÞ logðRt1 R2 ÞÞ ¼ expðN2 ðnÞD W ð27Þ
¼ q201 ðq2 ; q2 Þ q202 ðq1 ; q1 Þ 2q01 q02 ðq1 ; q2 Þ þ ðq1 q2 ; q1 q2 Þ
where R1 ¼ expð W e 1 Þ and R2 ¼ expð W
e 2 Þ. Using the BCH formula, we
¼ 1 q201 q202 ðq1 ; q2 Þ2 2q01 q02 ðq1 ; q2 Þ e 12 as
can obtain D W
2
¼ sin X:
e1 þW
~ 12 ¼ W
DW e 2 þ 1 ½ W
e 1; We 2 þ 1 ½ W e 1; W
e 1 ; ½ W e 2
2 12
Therefore the vector part of Q 1
1 Q 2 ¼ sin X e, where e is the unit 1 e
vector along q01 q2 q02 q1 q1 q2 . From Eq. (21), we have þ e 2; W
½W2; ½W e 1 þ HOT: ð28Þ
12

sinðN 1 ðnÞXÞ 1 sinðN2 ðnÞXÞ 1 e h Þ; then using Eq. (28) and the BCH formula, we
Let Rh ðnÞ ¼ expðr W
Q h ðnÞ ¼ Q 1 Q1 Q1 þ Q1 Q2
sin X sin X get:

sinðN 1 ðnÞXÞ 1 sinðN2 ðnÞXÞ cos X 1
¼ Q1 þ eh ¼ W
W e 12 þ ½ W
e 1 þ N2 ðnÞD W e 12
e 1 ; N2 ðnÞD W
sin X 0 sin X sin X e r
2
cos N 2 X h 1 e e 12
e 1 ; N2 ðnÞD W
¼ Q1 ¼ Q 1 DQ ðnÞ; ð24Þ þ ð½ W 1 ; ½ W
sin N2 X e 12
þ ½N 2 ðnÞD W e 12 ; ½N2 ðnÞD We 12 ; We 1 Þ þ HOT: ð29Þ
where DQ h ðnÞ ¼ ðcos N 2 X; sin N 2 X eÞ.
Substituting Q 1 Q 2 ¼ ðq01 q02 ðq1 ; q2 Þ; q01 q2 þ q02 q1 þ q1 q2 Þ The left-subscript r is used to indicate interpolation of relative rota-
in Eq. (A.2) and simplifying we get HðQ 1 Q 2 Þ ¼ HðQ 1 ÞHðQ 2 Þ ¼ tion. Substituting the series expression of D We 12 from Eq. (28) in the
R1 R2 . Therefore, HðQ h ðnÞÞ ¼ R1 HðDQ h ðnÞÞ. Using Eq. (A.2), we get last equation and retaining terms up to two Lie brackets, we obtain
N1 N2 e
HðQ h ðnÞÞ ¼ R1 ðð2 cos2 N 2 X 1ÞI þ sin 2N 2 X e
~ r
e h ðnÞ ¼ N1 W
W e 1 þ N2 W
e2 e 1; W
½W1; ½W e 2
2
12
þ 2 sin 2N2 X e eÞ: ð25Þ N1 N2 e
e 2; W
½W2; ½W e 1 þ HOT: ð30Þ
12
Using Eqs. (1) and (A.4), we finally have
Thus, the additive interpolation (Eq. (11)) gives the rotation field as
h g e 1 þ N2 W
Rh ðnÞ ¼ expðN 1 W e 2 Þ ¼ expð W
e h ðnÞÞ. Hence, it produces the
~Þ ¼ R1 expðN2 ðnÞ D
HðQ ðnÞÞ ¼ R1 expð2N2 Xe W Þ: ð26Þ
following error terms:
Therefore, slerp provides equivalent quaternion representation for N1 N2 e
interpolation of relative rotation based on two nodal rotations. r
e h ðnÞ W
W e h ðnÞ ¼ ½W1; ½W e 2 N 1 N2 ½ W
e 1; W e 2; W
e 2; ½W e 1
12 12
We note that the locally injective property between rotation matri-
þ HOT: ð31Þ
ces and quaternions is not required in the above proof.
Since the exponential map from soð3Þ to SO(3) is surjective, we Therefore, terms with two or more nested Lie brackets will contrib-
always have some Ri , given Wi ; 2 soð3Þi ¼ 1; 2, such that ute to the error.
expð W fi Þ ¼ Ri . The use of relative rotation offers a valid interpola- Accurate interpolation of iterative rotation would be through
tion as described by Eq. (14) i.e. Rh ðnÞ ¼ R1 expðð1 þ nÞ=2 log interpolation of relative rotation. In this context, it is useful to take
ðRt1 R2 ÞÞ. The logarithm is a multi-valued function; but it is known a look at the additive interpolation of iterative rotation of Eq. (13).
(see Section 3.6) that as long as R1 ¼ expð W e 1 Þ and R2 ¼ expð W
e 2Þ In particular assuming h to be the characteristic nodal spacing, the
are close enough (such that their relative rotations are bounded above analysis reveals that the error oder in additive interpolation
2
by p), the logarithm yields a unique value. In that case, the curve is of Oðh Þ. This observation, further substantiated through our
segment R1 expðN 2 ðnÞ logðRt1 R2 Þ for n 2 ½1; 1 produces the short- numerical experiments, suggests that it is reasonable to use (13)
est path between R1 and R2 . This is deducible from the fact that in deriving the weak form.
geodesic curves are invariant under left translations of the group. Now we provide further illustrations of the above observation
For example, if R1 ¼ I, then DW12 ¼ W2 , hence Rh ðnÞ ¼ expðN 2 ðnÞ through numerical results. We recall that the unit sphere S2 is a
We 2 Þ is a geodesic curve on the SO(3) manifold. Note that the expo- homogeneous space with respect to an SO(3)-action. In Figs. 2 and
nential mapping on a (Riemannian) manifold maps elements in the 3a, we have illustrated the interpolation in SO(3) by its action on
a b
0
1 10
0.5
–5
10
0
class–I
–10
10 class–II
class–III
–0.5
–0.5
–15
10
–1 0
1
0.5 0.5
0
–0.5 –20
10
–1 0 0.2 0.4 0.6 0.8 1
2 2
Fig. 2. The interpolated curves on S is RðnÞe3 2 S , where RðnÞ is interpolated curve in SO(3). The two end points corresponds to rotation vector
W1 ¼ ½0:6108; 0:5895; 0:0429; and W2 ¼ ½0:3920; 0:8072; 2:4424. (a) firm-line and dotted line via interpolation type (c) and (a), respectively (mentioned in Section
3.2); (b) norm of difference in RðnÞ e1 , where RðnÞ is obtained by different pair of interpolation methods as, for class-I: (c) and (a), for class-II: (a) and (b), for class-III: (b) and
(c).
a 1 b
100
0.5
0 10–5
–0.5
class–I
10–10 class–II
class–III
–1
0.5 10–15
1
0
0
–0.5
10–20
–1 0 0.2 0.4 0.6 0.8 1
Fig. 3. Interpolation results same as (Fig. 2), with end rotation vectors W1 ¼ ½0:9659; 0:6512; 0:7885; and W2 ¼ ½0:1056; 1:1294; 0:6813.
S2 . Given two rotation vectors, W1 and W2 , the interpolated rotation uses the additive interpolation of material rotational incre-
tion acts on e3 ¼ ½0; 0; 1. The vector RðnÞe3 is plotted on S2 via dif- ments to obtain the test functions. We however note that, with
ferent interpolation methods ((a), (b) and (c)). It is observed that the rotation vector parametrization, no finite element discretiza-
the interpolation of relative rotation (i.e. (a)) provides identical re- tion can be entirely consistent for rotations beyond the full angle.
sults as those via interpolation of quaternions (i.e. (b)). To show the This is because the SO(3) is a compact manifold as contrasted with
error numerically, the differences in RðnÞe3 obtained via the three the set of all rotation vectors forming an open set (already pointed
different methods are plotted in Figs. 2 and 3b. The claimed equiv- out in Section 2.2.1). While the Eulerian formulation could reduce
alence of the proposed interpolation with the interpolation of the rate of convergence, FLOPS per iteration are the least due to the
relative rotation is readily verifiable. The neglected terms (refer relative simplicity of the formulation [2]. For brevity, we skip the
Eq. (30)) in the additive interpolation of rotation vectors (i.e. (c)) details of the finite element discretization.
lead to errors, which may possibly be small for refined meshes.
Nevertheless, it does produce non-objective strains and path- 3.3. Slerp for a three-noded element
dependent solutions.
The isoparametric interpolation of Eq. (21) is valid for only two-
3.2.3. Finite element discretization and strain-configuration matrix noded elements. For elements with three or still higher number of
The finite element discretization in this study follows Simo and nodes, derivation of a similar interpolation is rather difficult. In the
Vu-Quoc [2]; therefore the discretized strain-configuration matrix following, we propose a few such interpolations for a three-noded
(see Eq. (C.1)) remains similar to that of the Eulerian formulation of element.
[2]. Though different discretizations should have implications in fi- Given three nodal quaternions Q 1 , Q 2 , Q 3 , we choose interme-
nite element solutions, we have kept possible innovations in dis- diate quaternions Q 1 and Q 2 to allow control of the interpolated
cretization out of the scope of the present study (a point of view curve. Interpolate Q U along the geodesic from Q 1 to Q 3 using slerp.
that has been adopted earlier by others [1]). The Eulerian formula- Simultaneously interpolate Q V along the geodesic from Q 1 to Q 2 .
This is followed by interpolation of Q U and Q V to get the final re- implement this in our FE code. Instead, in the next sub-section, we
sult as will explore the possibility of efficient use of quaternions for the
interpolation of relative rotations for higher order elements.
Q h ðnÞ ¼ slerpð2N 1 ðnÞN 2 ðnÞ; slerpðN2 ðnÞ; Q 1 ; Q 2 Þ;
slerpðN 2 ðnÞ; Q 1 ; Q 2 ÞÞ; ð32Þ 3.4. Efficient interpolation of relative rotations via quaternions
where the intermediate points are computed as Q 1 ¼ slerpð2; Q 2 ; The formula (Eq. (21)) for interpolating total quaternion is valid
Q 3 Þ; Q 2 ¼ slerpð2; Q 2 ; Q 1 Þ. The functions N1 ðnÞ and N 2 ðnÞ are as in only for a two-noded element. For more than two nodes, a similar
Eq. (21). The idea is to interpolate between two interpolated quater- derivation is not possible due to the presence of more than one axis
nions. The intermediate points are not unique and can be chosen of 3D rotation between different nodes. In such cases, interpolation
according to requirements. Here we have chosen intermediate of relative rotation appears to be the best available solution. In
points such that the interpolated quaternion comes close to Q 1 at [20], interpolation of relative rotation (Eq. (14)) is performed by
mid point of the beam. obtaining the relative quaternion from the relative rotation tensor,
It is readily seen that the above interpolation, being based on followed by the extraction of the relative rotation vector. Their ap-
slerp, is invariant under rigid body rotation (the invariance of slerp proach requires both tensors and quaternions, and quaternions are
will be shown in Section 4). But it does not reproduce Q 2 at the extracted through Spurrier algorithm. In the following, we will
middle node. This is however possible in another form of cubic propose a computationally efficient alternative implementation
interpolation given below for interpolating relative rotations. In the process, we will directly
compute relative quaternions from total rotation vectors and then
Q h ðnÞ ¼ slerpðN 2 ðnÞ; slerpð4N1 ðnÞN2 ðnÞ; Q 1 ; Q 2 Þ; slerpðN2 ðnÞðN2 ðnÞ
obtain the nodal relative rotation vectors. The relevant steps for
N1 ðnÞÞ; Q 2 ; Q 3 ÞÞ: ð33Þ such an isoparametric interpolation are given below
While such composite slerp approximations are invariant, they are
not necessarily of higher order consistency than the two-point slerp (1) Find quaternions for nodal and reference rotations Q i ¼
approximation. To obtain derivatives of this kind of interpolation, F1 ðWi Þ.
we rewrite the above equation as (2) Find relative quaternions with respect to the reference qua-
ternion Q r : DQ ri ¼ ðDq0 ri ; Dqri Þ ¼ Q 1
r Q i.
Q h ðnÞ ¼ slerpðN 2 ðnÞ; UðnÞ; VðnÞÞ: ð34Þ (3) Referring to Eqs. (A.4)–(A.6), obtain the terms a and e for
Dqri
The derivative d
Q h
ðnÞ may be obtained as each node using ari ¼ cos1 ðQ r Q i Þ, eri ¼ sinð ari Þ.
ds
(4) Define vri ¼ ari eri . Interpolate vri as vðnÞ ¼ N i ðnÞvri and v0 ðnÞ ¼
1 N 0i ðnÞvri
Q 0h ¼ ðcos N1 XUV ðN01 XUV þ N1 X0UV Þ
sin XUV (5) Use Rodrigues’ formula (Eq. (1)) on 2v to obtain the rotation
sin N1 XUV 0 field RðnÞ ¼ Rr expð2v ~Þ. The curvature is obtained as K ¼ 2T
sin N1 XUV cot XUV X0UV ÞU þ U ð2vÞv0 .
sin XUV
þðcos N2 XUV ðN02 XUV þ N2 XUV Þ sin N2 XUV cot XUV X0UV ÞV
0
Note that v is half of the relative rotation vector. The function
sin N2 XUV 0
þ V : F1 is defined in Eq. (A.4).
sin XUV
Remark 1. Relative Rotation Vector through slerp
Here, cos XUV ¼ U V: U 0 , V 0 and X0UV can be obtained from their For higher order elements, consistent isoparametric interpola-
respective definitions. The above two interpolations are elaborated tion of total rotation is not viable and relative rotation is advanta-
through Fig. 4. It appears that the above formula is rather cumber- geous. Nevertheless, it can be shown that the relative rotation
some and the very purpose of it, efficiency, gets defeated. Moreover, vector is obtainable from total rotation via slerp. Consider ðslerpðn;
a similar formula for elements with four or mode nodes will be very 0
Q r ; Q i ÞÞ1 ðslerp ðn; Q r ; Q i ÞÞ; it can be shown to be equal to
difficult to derive and handle. To summarize, the concept of slerp is
1 1
not very effective beyond two-noded elements. Therefore we do not ðsin Nr Xri Q 1 1
r þ sin N i Xri Q i Þ ðN0 Xri cos Nr Xri Q r
sin Xri sin Xri r
Xri 0 0
þ N 0i Xri cos Ni Xri Q i Þ ¼ ¼ :
sin Xri q0r qi q0i qr qr qi vri
0
Therefore, vri is the vector part of ðslerpðn; Q r ; Q i ÞÞ1 ðslerp
ðn; Q r ; Q i ÞÞ (this, however, is also obtainable from the definition of
logarithm of quaternion). This route of extracting vri is less efficient
than the algorithm above.
3.5. Singularity free additive interpolation of total quaternions

followed by normalization
The interpolation described in Section 2.2.4 [1] is susceptible to

singularity even for very small relative rotation. Even when it is not
singular, the method may yield grossly erroneous interpolated
quaternions. This is due to 2-1 correspondence between quater-
nions and rotations. The additive interpolation is allowed to inter-
polate quaternions at two different hyper-hemispheres of S3 . Here,
Fig. 4. Different quaternion-based interpolations between three nodes (plotted on one should ensure that the nodal quaternions are in the same hy-
unit sphere as in (Fig. 2)), with end rotation vectors W1 ¼ ½0:6108; 0:5895;
0:0429, W2 ¼ ½0:3920; 0:8072; 2:4424; and W3 ¼ ½0:920; 0:8072; 1:4424. Note
per-hemisphere of S3 . One may do so by choosing a suitable qua-
that all the curve lie on the unit sphere. Legend: dashed line is by Eq. (32); firm line ternion as described in Section 3. It is checked whether Q i Q r is
is by Eq. (33); –+– is by Eq. (14); dotted line is by slerpðN 2 ðnÞ; Q 1 ; Q 1 Þ of Eq. (32). positive; if not then one must use the negative of Q i . Note that
the procedure restricts the angle of relative rotation between two matrices and then performing the update via matrix multiplica-
end nodes within p. tions; but this would require substantially more computational
We will further explain this through an example. Let effort.
Q 1 ¼ ð0:9110; 0:2963 0:2860 0:0208Þ and Q 2 ¼ ð0:2665;
0:1452 0:2990 0:9047Þ. The interpolated quaternion at the middle 4. Invariance of strain measures under superposed rigid body
(via Eq. (15) is Q h ¼ ð0:5158; 0:1209 0:4681 0:7073Þ, which motion
is not the desired interpolation. This is because Q 1 Q 2 < 0 (so the
additive formula interpolates two quaternions at two different Material strain measures at a particular configuration are objec-
hyper-hemispheres of S3 ). This gives a crude approximation since tive if they are invariant under superposed rigid-body motion. It is
this formula does not respects the geometric structure and, in known that the strain measures defined by Eq. (2) are frame invari-
such cases, the normalization of quaternion cannot bring it ant. In this section we will prove the objectivity of strain measures
closer to the desired curve on S3 . By choosing the appropriate approximated by the proposed isoparametric interpolation. We
quaternion for Q 2 , we get the interpolated quaternion as present the proof for slerp as applied to a two-noded element.
Q h ¼ ð0:7541; 0:2827 0:0084 0:5927Þ, which is closer to the geo- For a beam element with an arbitrary number of nodes, relative
desic through R1 and R2 . rotation can be used which is known to be objective. Now consider
a two-noded element. Let the current configurations at node 1
3.6. Rotation vector update based on quaternions þ
ðx0 1 ; Q 1 Þ and node 2 ðx0 2 ; Q 2 Þ be modified to ðxþ 0 1 ; Q 1 Þ and
þ þ
ðx0 2 ; Q 2 Þ, respectively, via the superposition of an arbitrary rigid
After solving for the increments from the discretized version of body motion ðx0r ; Q r Þ; i.e.
the linearized weak form, they are used to update both the position
and rotation vectors at each node. While the displacement update xþ0 1 ¼ FQ r ðx0 1 Þ þ x0 r Q þ1 ¼ Q r Q 1; ð36Þ
follows the standard vector addition, the rotation update must re- xþ0 2 ¼ FQ r ðx0 2 Þ þ x0 r Q þ2 ¼ Qr Q 2: ð37Þ
tain the nonlinear character of the rotation manifold. The standard 3
The rotation (R) and its action on a vector ðRv 8v 2 R Þ are, respec-
approach to the rotation update is either via matrix multiplication
tively, represented by a quaternion Q and the rotation of the same
or quaternion product; see [2,18]. We presently propose a simple
vector by the quaternion FQ ðvÞ. A quaternion Q ¼ ðcos h=2;
procedure, based on quaternions, to yield a precise update that by-
sin h=2 eÞ represents a rotation of a vector v by an angle h about
passes the computation of the rotation matrix. This approach is
the axis e. The function FQ ðvÞ maps the vector v to the rotated vec-
motivated by the well known advantages of quaternion products
tor (Rv). This is done in two steps. First the quaternion representa-
over matrix multiplications.3 We achieve rotation updates by con-
tion of the rotated vector is obtained as Q ð0; vÞ Q . This is
verting both nodal rotation vectors and their increments to quater-
known as conjugation by Q . Then the rotated vector is obtained by
nions (via Eq. (A.4) followed by quaternion multiplications.
just selecting the vector part of the resulting quaternion. Note that
Following this, we extract rotation vectors from quaternions. The
the scalar part of it is zero. Thus the function FQ may be defined as
steps for the ith node are enumerated below:
conjugation by Q followed by the selection of the vector part.
Since the proposed interpolation chooses the quaternion such
(1) Q ki ¼ F1 ðWi Þ DQ ki ¼ F1 ðDHki Þ
that the angle between the quaternions of two consecutive nodes
(2) Q kþ1
i ¼ Q ki DQ ki
is acute, a 1-1 relation between R and Q is ensured. Thus, the
(3) Let Q kþ1
i ¼ ðqkþ1
0i ; qi
kþ1
Þ, if qkþ1
0i < 0 then, Q kþ1
i ¼ Q kþ1
i
kþ1 kþ1 non-uniqueness of quaternions does not affect the following argu-
(4) Wi ¼ F2 ðQ i Þ
ments. To establish that the frame-invariance of strain measures is
inherited by the proposed interpolation, we need to prove the
The function F1 is given in Eq. (A.4) and F2 , the inverse of F1 ,
following:
is defined below:
" # " #
2 sin kqk
1
Cþh ðnÞ Ch ðnÞ
W ¼ F2 ðQ Þ ¼ q: ð35Þ ¼ : ð38Þ
kqk K þh ðnÞ K h ðnÞ
Recall that the map from the set of unit quaternions to SO(3) is 2-1.
At first, we will show the following properties of the interpolated
Step (3) above ensures that we choose the updated quaternion such
variables, x0 ðnÞ and Q ðnÞ:
that it always remains in that hyper-hemishpere of S3 that corre-
sponds to q0 P 0. This step is essential as the function F2 cannot xþh h
0 ðnÞ ¼ FQ r ðx0 ðnÞÞ þ x0 r Q þh ðnÞ ¼ Q r Q h ðnÞ: ð39Þ
extract a rotation vector of magnitude greater than p. This is be-
1 Using the standard isoparametric interpolation for the centroidal
cause sin works on the norm of a vector q, which is always posi-
curve and the proposed quaternion interpolation of rotation of the
tive. While the extraction of rotation vector from quaternion is not
cross-sectional plane, the configuration before rigid body motion
unique, given a quaternion Q ¼ ðq0 ; qÞ with q0 P 0, there is a un-
is given by:
ique rotation vector with magnitude less than or equal to p and this
is extracted by F2 . However the choice of a quaternion or its neg-
xh0 ðnÞ ¼ Ni ðnÞx0 i ;
ative does not affect our formulation since, during interpolation
(Eq. (21), we choose signs of quaternions of consecutive nodes such sinðNi ðnÞXÞ
Q h ðnÞ ¼ Qi i ¼ 1; 2;
that the ‘angle’ between them is acute. Similar rotation vector up- sin X
date may be performed by converting rotation vectors to orthogonal where cos X ¼ Q 1 Q 2 . The configuration after the superposed rigid
body motion is given by
h
3
The standard way of multiplying two 3 3 rotation matrices takes 27 multipli- xþh
0 ðnÞ ¼ N i ðnÞðFQ r ðx0 i Þ þ x0 r Þ
cations and it is possible to show that two 3 3 matrices cannot be multiplied with
X ne
fewer than 17 multiplications. In contrast, two quaternions may be multiplied with
¼ FQ r ðNi ðnÞxh0 i Þ þ x0 r Ni ðnÞ
just twelve multiplications and a few additions. Also, round-off errors in matrices are
1
more than in quaternions. Round-off errors lead to loss of orthogonality and the
requirement of re-orthogonalization. While quaternion renormalization is as easy as ¼ FQ r ðxh0 ðnÞÞ þ x0 r ð40Þ
normalizing a vector, rotation matrix orthonormalization is far more costly.
sinðNi ðnÞXþ Þ The function H maps a quaternion to rotation tensor and defined by
Q þ h ðnÞ ¼ Q r Q i ; i ¼ 1; 2 0 0
sin Xþ Eq. (A.2). Noting that Q þh ¼ Q r Q h ¼ Q r ðQ h logðQ ÞÞ, we get

sinðNi ðnÞXþ Þ
¼ Qr Q i : ð41Þ e þh ¼ HððQ h
K Q r Þ ðQ r ðQ h logðQ ÞÞÞÞ
sin Xþ
h h
¼ HðQ ðQ logðQ ÞÞÞÞ
Now, if we show that cos Xþ ¼ cos X, then from finiteness of an ele-
h 0
ment and continuity of rotation, we will get Xþ ¼ X. We provide ¼ HðQ QÞ
this proof below and presently proceed assuming that it is true. ht
e h:h0
¼R R ¼K
Thus the last equation becomes
Q þh ðnÞ ¼ Q r Q h ðnÞ: In the above, we have used the associativity of quaternion multipli-
cation. Thus the proof of Eq. (38) is complete. The specific form of
Hence, the properties of interpolated variables given by Eq. (39) quaternion interpolation, presently employed, ensures that the
hold. interpolated quaternion is always on the unit hyper-sphere S3 and
We will now show that cos Xþ ¼ cos X. Using the following this is the key to the above proof.
relation between quaternion dot product and quaternion product Path-independence means that the final solution is independent
(Grassmann product): of the path of deformation. It is evident that, due to the proposed
Q þ1 Q þ2 þ Q þ2 Q þ1 interpolation (Eq. (21), the interpolated strain measures Ch and
Q þ1 Q þ2 ¼ ð42Þ K h depend only on the current configurations of nodes ðx0 i ; Q i Þ.
2
Hence the Ch and K h are independent of the history of deformation.
and the formula for conjugate quaternion, we get
Q þ1 Q þ2 ¼ qþ1 0 qþ2 0 þ ðqþ1 ; qþ2 Þ: ð43Þ 5. Numerical simulations
We have used the notations ¼ Qþ

1 ðqþ þ
and Q þ
1 0 ; q1 Þ
þ þ
2 ¼ ðq2 0 ; q2 Þ. Via In the first two examples, we show the performance of the pro-
þ þ
the definition of quaternion product, we write Q 1 and Q 2 as posed interpolation for any rotation. We present numerical evi-
Q þ1 ¼ ðqr 0 q1 0 ðqr ; q1 Þ; qr 0 q1 þ q1 0 qr þ qr q1 Þ; dence of invariance under superposed support rotations and
path-independence of final displacements in Examples 3–5. All
Q þ2 ¼ ðqr 0 q2 0 ðqr ; q2 Þ; qr 0 q2 þ q2 0 qr þ qr q2 Þ:
along this section, we compare the results obtained through the
Identifying the scalar and vector parts in the last equation and present method with those available in the literature. In all the
substituting back in Eq. (44), we get numerical simulations, one-point Gauss quadrature is used for
numerical integration and the equilibrium is considered to be
Q þ1 Q þ2 ¼ ½qr 0 q1 0 ðqr ; q1 Þ½qr 0 q2 0 ðqr ; q2 Þ established when the residual norm reaches below 106 .
þ ð½qr 0 q1 þ q1 0 qr þ qr q1 ; ½qr 0 q2 þ q2 0 qr þ qr q2 Þ:
5.1. Example-1. Application to large rotation: beam bent to a helical
Following some algebraic manipulations, the above equation sim-
form
plifies to
Q þ1 Q þ2 ¼ ½q2r 0 þ ðqr ; qr Þ½q1 0 q2 0 þ ðq1 ; q2 Þ ¼ ½Q r Q r ½Q 1 Q 2 : This example has been used in [17] and [12] in the contexts of
the incremental rotation vector formulation and the switching
The fact that Q r is an unit quaternion leads to the desired result, i.e.
beam element technique. Here, we examine a cantilever beam of
cos Xþ ¼ cos X. This completes the proof of the properties given by
length 10, axial stiffness EA ¼ 104 , shear stiffness GA2;3 ¼ 104 ,
Eq. (39).
bending stiffness EI2;3 ¼ 102 and torsional stiffness GJ ¼ 102 . We
What remains is to prove Eq. (38). Using Eq. (39) and noting that
apply a tip load F ¼ 50 in the Y-direction and a couple M ¼ 200p
the proposed interpolation ensures locally invertible correspon-
about the Y-axis. This example assumes importance as it shows
dence between rotation RðnÞ and quaternion Q ðnÞ, we may write
the performance of the proposed method for rotation angles
t 0 exceeding 2p. In order to ensure accuracy in the solution despite
Cþh ¼ Rþh xþh
0 E1 ;
0
large deformations, we use 100 beam elements and 100 equal load
¼ FQ þh ðxþh
0 Þ E1 : increments as done in [17]. We have obtained the same results as
reported in [17]. The deformed shape is given in Fig. 5.
Using the identities
Q þh ¼ ðQ r Q hÞ 5.2. Example-2. Application to large rotation: deployment of a ring

0 0
x0þh0 ¼ Rr xh0 ¼ FQ r ðxh0 Þ
This example studies very large deformations of a circular ring,
and which considerably reduces its original size; see (Fig. 6a). This
0 0
problem has been considered in [27,10,1]. Rotations (of magnitude
FQ þh ðFQ r ðxh0 ÞÞ ¼ FQ h ðxh0 Þ p) at two diametrically opposite points A and C are imposed,
respectively, along the positive and negative directions of the Y-
we have
axis. The deformation stages for every p=8 rotation increment at
0
Cþh ¼ FQ þh ðFQ r ðxh0 ÞÞ E1 each of the control nodes A and C are shown in (Fig. 6). The follow-
0 ing properties are considered for the ring; radius = 20, rectangular
¼ FQ h ðxh0 Þ E1 cross-section with height = 1 and width = 1/3, polar moment of
t
¼ Rh x00 E1 inertia = 9.753 103, Young’s modulus = 21 106 and Poisson’s
ratio = 0.3. Like [10], we have used 48 linear isoparametric beam
¼ Ch : elements. Note that the total rotation is imposed in increments.
Similarly, for the bending curvature, we have With the present formulation, it takes 42 increments to impose
the maximum rotation. The deformed shape matches well with
e þh ¼ Rþht Rþh0 ¼ HðQ þh
0
K Q þh Þ: that reported in the existing literature.
a 4 b
3
0.3
2
0.25
Z
1 0.2
Z
0.15
0
0.1
—1 0.05
0.2
0
—2
0 0 —0.1 0.1
2 0 0.1
4 Y 0.2 0
6 8 0.3 0.4 —0.1
10 —0.2 0.5 Y
X X
Fig. 5. Cantilever under tip load and tip moment: (a) initial configuration and (b) deformed shape.
a A b c
D
5 20 5 20 5 20
0 B 10 0 10 0 10
Z
Z
—5 0 0 0
C —5 —5
—20 —10 Y —20
—10 Y —20
—10 Y
—10 —10 —10
0 —20 0 —20 0 —20
10 10 10
20 X 20 20
X X
d e f
5 20 5 20 5 20
10 10 10
Z
0 0 0
Z
Z
0 0 0
—5 —5 —5
—20 —10 Y —20 —10 Y —20 —10 Y
—10 —10 —10
0 —20 0 —20 0 —20
10 10 10
X 20 X 20 X 20
g h i
5 20 5 20 5 20
10 0 10
Z
0 10 0
Z
Z
0 0 0
—5 —5 —5
—20 —10 Y —20 —10 Y —20 —10 Y
—10 —10 —10
0 —20 0 —20 0 —20
10 10 10
X 20 20 X 20
X
Fig. 6. Different deformation stages of the deplyable ring. Sub-figures (a)–(i) correspond to imposed rotation at point A and C of magnitude 0 to p at an interval of p=8.
5.3. Example-3. Objectivity test: L-shaped cantilever subjected to (2) the load and support rotation applied together in a single
support rotation and tip load stage.
The first part of this example shows path-independence. This is Using the present method, the X-component of the tip displace-
followed by a series of objectivity tests for different cases of load- ment is obtained exactly as 10 (up to 14 decimal places) in both
ing and prescribed rotation. Different parts of this example have the loading sequences, verifying that the method is frame-indiffer-
earlier been considered in [20,18] We examine an L-shaped canti- ent. Recall that we use the interpolation of iterative rotations,
lever clamped at one end, subjected to a concentrated vertical force However, as reported in Table 4 of [20], the existing formulation
or moment at its free end and different support rotations. The based on interpolation of iterative rotation vectors is not objective,
structure is composed of two equally long mutually perpendicular producing different solutions for two different sequences of load-
legs of length 10 units, which are parallel to X and Y axes, respec- ing. The deformed shapes via the present method are shown in
tively. It is modeled with five elements in each leg. Different load (Fig. 7). The robustness of the present scheme is highlighted by
and support rotations are considered leading to variety of de- the fact that the support rotation, currently imposed in a single
formed shapes for the same structure. They are as follows: step, is indeed very large.
For yet another check on the accuracy of the present method,
5.3.1. Case-I solutions for the tip-coordinates under only the vertical tip load
We use the example in [20] to check path-independence of the are given in Table 1. Here we compare our result with those of
present formulation. Accordingly, two different loadings are ap- [18]. The results show a good match. (Fig. 8a) shows the initial
plied as follows: and deformed shapes.
(1) a tip load of magnitude five in the negative Z-direction (see 5.3.2. Case-II
(Fig. 7) in the first stage followed by a support rotation p=2 A test of objectivity is obtained by allowing further rotation of
(in a single increment) with respect to the negative Y-axis in the support in the same direction up to 100 full revolutions. The
the second stage; displacement component of the tip is monitored after each com-
10
Z
4
0
—5
0
15
X 5 10
5
10 0 Y
—5
Fig. 7. L-shaped cantilever under tip load (of magnitude 5) and support rotation (of magnitude p=2) around negative Y-axis.
Table 1 example has been considered via an updated Lagrangian formula-

L-shaped cantilever: tip displacement components in (x, y, z)-direction for tip load tion in [18]. The loading is presently provided in two phases. In the
Methods Displacement first phase, either (i) a tip load (of magnitude five along the nega-
tive Z-axis) or (ii) a tip moment (of magnitude 10p about the Y-
Present (five elements per leg) (0.4126, 1.7511, 6.7468)
Present (20 elements per leg) (0.4259, 1.7512, 6.7671)
axis) is applied. (Fig. 8) shows the deformed shapes for the two
Orthogonal tensora [18] (0.43081, 1.7368, 6.7329) types of loading that are applied in a single increment. In the sec-
Incremental rotation vectora [18] (0.43080, 1.7367, 6.7330) ond phase, subsequent to the loading as above, a support rotation
a
Orthogonal tensor and incremental rotation vector represent two different
about either (i) the Z-axis or (ii) the X-axis is imposed.
choices of rotation parameters. Let the tip force/moment be followed by support rotation
about the Z-axis, with the increment of rotation being p=2. This
plete revolution. However, for tackling such huge rotations, the results in 100 revolutions in only 400 increments. Fig. 10a and
size of each rotation increment is reduced and taken as p=3 (which b is consistent with our earlier observation on objectivity; in both
is also quite large). The rigid body rotation (i.e. the support rota- the cases we find that the deformed shapes remain identical fol-
tion) should not perturb the existing state of stress established at lowing every complete revolution. In this case too, we note that
the end of the first loading phase. Since the second loading phase the support rotation is as high as p=2 over a single increment
results in the whole structure being rotated as a rigid body, there and that this is far higher than p=20, earlier used with the up-
should not be any differences in deformed shapes corresponding dated Lagrangian formulation in [18]. Of related relevance is also
to support rotations of magnitudes, say, 0 and 2kp; k ¼ 1; 2; . . .. the fact that the first phase of loading (i.e. tip force/moment) is
The deformed shapes for each increment over 100 revolutions presently applied in a single increment. This may be contrasted
are plotted in Fig. 9a. We observe that, after each complete turn, with five increments needed with the updated Lagrangian
the deformed shapes remain identical. By way of another instance formulation.
of the numerical proof of objectivity, we plot in Fig. 9b changes in
the strain energy following every completer turn. As anticipated, 5.3.4. Case-IV
the changes are negligibly small. Let the same tip force/moment be followed by support rotation
about the X-axis. The tip load and the tip moment cases need 3100
5.3.3. Case-III and 1700 support rotation increments, respectively, to reach 100
The above example is now extended by introducing yet another complete revolutions. The formulation given in [18] takes 4000
loading and changing the direction of the support rotation. This support rotation increments in both the cases. In Fig. 11, deformed
a b
2
2
0
0
—2
Z
—2
Z
—4 0
—4 0
—6 5
—6 5
X
X 0
0 2 2 10
4 10 4
6 6
8 8
10 Y 10
Y
Fig. 8. L-shaped cantilever under (a) tip load and (b) tip moment.
a b
5
Change in strain energy

10–11.685
Z
–5
10–11.687
–10
10–11.689
–5
0
10 10–11.691
X 5
5 0 20 40 60 80 100
10 0 Y Number of turns
Fig. 9. L-shaped cantilever under tip load and support rotation around negative Y-axis: (a) undeformed and deformed shapes, (b) change in strain energy.
a 5 b
5
0
0
Z
–5 –5
–10 –10
–10 –10
–5 –5
10
0 10 0 5
X 5 5 X 5 0
0
–5 –5
10
–10 Y 10 –10
Y
Fig. 10. L-shaped cantilever: deformed shapes under (a) tip load, or (b) tip moment, and subsequent support rotation about Z-axis.
a 5 b 10
5
0
Z
0
Z
—5
—5
—10
0 —10
0
5
10 5 10 5 10
0 0 5
15 5 X 10 10 5
Y Y
X
Fig. 11. L-shaped cantilever: deformed shapes under (a) tip load (two-noded, five elements per leg), or (b) tip moment (six-noded, two elements per leg), and subsequent
support rotation about X-axis.
shapes for all 100 revolutions are plotted. These results are again Indeed, in this case, using a Gauss quadrature rule involving less
consistent with the stated objectivity of the present method. The than five quadrature points leads to singularity of the stiffness
Fig. 11b is obtained by using two six-noded beam elements per matrix.
leg. The interpolation described in Section 3.4 is used for Fig. It was found that the changes in the free tip-coordinates and in
11b. We have used five point Gauss quadrature rule for the six- the strain energy after each complete support revolution remain
noded beam element (against the three point quadrature rule, within 1011 .
which would normally have been sufficient). We note that the In all the above cases (Case-II through Case-IV), norms of
requirement of higher order Gauss quadrature probably arises changes in tip-coordinate after 100 revolutions are restricted with-
out of the geometric inconsistencies in the choice of test functions. in 1014 .
5.4. Example-4. Objectivity test: rigid motion of a deformed beam
0.15032767877715
0.03524548915147
0.03524548915147
1.76165746007886
0.15032767877710
0.00000000000000
As yet another test of objectivity, we use the example on rigid
rotation of the deformed rod, earlier studied in [23,1]. A straight
rod with two ends at A(0, 0, 2) and B(0, 0, 5) is discretized with nine
elements. The following properties are used: rod cross-sectional
area = 0.1, polar moment of area = 1.6 104, other two moments
60 of area = 8.3 105, Young’s modulus = 1.2 108 and Poisson pffiffiffi ra-
tio
pffiffiffi = 0.3. To deform the rod, node B is displaced by 0:5½1= 2; 1=
0.15032767877713 2; 0 along X, Y, Z directions, respectively (without any rotation).
0.15032767877711
0.03524548915147
0.03524548915147
1.76165746007870
0.00000000000000
During the imposed displacement, node A is kept fixed. Due to the
imposed displacements, the rod is in a state of axial, shear and
bending deformations. This is followed by the superposition of
six rotations of magnitudes p=18; 2p=18; . . . ; 6p=18 about the axis
p1ffiffi ½1; 1; 1 at the two ends. All of the convected stresses, i.e. axial,
3
50
shear, torsional and bending, are computed for the element adja-
cent to node A in each configuration and provided in Table 2. This
shows the objectivity of strain measures as none of the stresses
0.15032767877712
0.03524548915147
0.03524548915147
0.15032767877711
1.76165746007880
0.00000000000000
changes when the beam undergoes rigid body rotations. Stresses

shown in this table can be compared with results reported with
other formulations in [23,1]. The objectivity is also evident form
a negligible change in the strain energy during the rigid rotation
(as depicted in (Fig. 12)).
Via Examples 3 and 4, we have sufficient numerical evidence for
40
the computed response remaining frame-invariant. Based on a

rotation vector parametrization, the only other work in [18], which
also achieves objectivity, needs the loading in the first phase to fol-
0.15032767877714
0.03524548915147
0.03524548915147
0.15032767877712
1.76165746007875
0.00000000000000
low the support rotation to preserve the stress state corresponding

to the first phase. In other words, the loading itself needs a (spuri-
ous) modification. The present approach however does not impose
such a restriction while achieving objectivity.
30
5.5. Example-5. Path-independence
We take up the well known example of a 45°-bent cantilever.

0.15032767877713
0.03524548915147
0.03524548915147
0.15032767877713
1.76165746007870
0.00000000000000
This problem has been considered in [20,18] to check path-inde-

pendence. The structure is loaded to a peak load in increments
and then the load is also removed in decrements. The objective is
to investigate whether, upon unloading, the finally deformed shape
corresponds to zero deformations (modulo the round-off errors) at
Example 4: stresses 105 in the element adjacent to node A for superposed rigid rotations
all nodes.
20
The bend has a radius of 100 units and the cantilever has unit
square cross-section uniformly. The material properties are:
Young’s modulus = 107, Poisson’s ratio = 0. We have used eight
0.03524548915147
0.03524548915147
0.15032767877712
0.15032767877712
1.76165746007870
0.00000000000000
beam elements. The beam is loaded with a vertical tip load of

600 units, which is applied in 6 equal load increments and then un-
loaded using the same number of decrements. Fig. 13 reports the
10
10—9
0.15032767877712
0.03524548915147
0.03524548915147
0.15032767877712
1.76165746007878
0.00000000000000
Change in strain energy

Rotation angle (°)
10—10
0
Bending moment (1)

Bending moment (2)
Torsional moment
Shear force (1)
Shear force (2)
10—11
Axial force
0 10 20 30 40 50 60
Rotation angle (Degrees)
Table 2
Fig. 12. Change in strain energy under rigid rotation.

50
tions. We have noted that, unlike an additive interpolation of total
rotation vectors, the two-noded interpolation of total quaternions
corresponds to a geodesic on the rotation manifold. We have also
40
obtained the error terms in the additive interpolation. However,
constructing a geodesic through more than two-points is generally
30 not feasible; thus slerp cannot be systematically extended in the
Z
case of more than two nodes for achieving higher order consistency.
20 For approximation over an arbitrary number of nodes, we have pro-
posed an efficient strategy for interpolation of relative rotations via
Z quaternions. In the sequel, we have also traced the reason of possi-
10
ble singularity and inaccuracy in an additive interpolation of quater-
nions [1] and worked out a strategy to bypass these problems.
0 X
0 The quaternion-based multiplicative rotation update, used here,
20
0 perhaps takes the least computational effort amongst the existing
40 10
60 20 schemes. Several numerical experiments are reported to verify that
Y X
the slerp over two nodes works effectively for very large deforma-
Fig. 13. Initial and deformed shape of the 45° bent cantilever under six load tions whilst maintaining strain-objectivity and path-indepen-
increments. dence. This is contrasted with the updated Lagrangian
formulation [18], which needs to spuriously affect the loading
characteristics in order to attain objectivity. We also do not need
Table 3 any secondary storages at Gauss points. In addition, the results
45° bent cantilever: displacements 1015 for all nine nodes in (x, y, z)-direction, after suggest that the present approaches (i.e., slerp over two nodes as
unloading, via present method
well as interpolation through relative quaternions over several
Node number ? 1 2 3 4 5 6 7 8 9 nodes) work with larger rotation increments than most existing
0 0.056 0.222 0.888 0.888 0 0 0 0 methods based on the rotation vector parametrization. If compared
0 1.776 0 0 0 0 0 0 0 with the invariant formulation in [20], besides a three-fold reduc-
0 0 0 0 0 0 0 0 0 tion in memory requirements due to rotation vector parametriza-
tion, the present methods also reduce the computational cost of
interpolation and update of rotation vectors.
deformed shapes for each of the six load increments leading to the
maximum load. The tip displacement vector for maximum load is
Appendix A. Parametrization of rotation by quaternion
(13.6187, 23.4184, 53.3743). Table 3 reports the final nodal dis-
placements upon complete unloading. The node numbering is
Quaternions, presently denoted by Q ¼ ðq0 ; qÞ,4 have four
sequential and starts from the fixed end. It is found that rotation
parameters, with a scalar part q0 and the vector part q. q can be
components are all zeros (i.e., of a typical order of 1021 ) and hence
thought of as an element of R3 . The norm of a quaternion is defined
we do not provide the numbers explicitly. However, not all the no- qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
dal displacements reach zeros with such a high order; the maxi- as kQ k ¼ q20 þ kqk2 . Additions and subtractions of quaternions are
mum departure from zero is at the second node (of order 1015 ). defined as:
For this 45°-bent cantilever, with the same maximum load with
Q1 Q 2 ¼ ðq1 0 q2 0 ; q1 q2 Þ:
10 load increments, Ibrahimbegović and Taylor [18] have reported
an error of order 105 or 106 at the free end. It is thus numerically Quaternion dot product is defines as
evidenced that the present method has remarkably improved Q 1 Q 2 ¼ q1 0 q2 0 þ ðq1 ; q2 Þ:
path-independence characteristics than the updated Lagrangian
The conjugate of a quaternion Q is defined as Q ¼ ðq0 ; qÞ. The
formulation [18]. We also note that similar path-independence is
multiplicative inverse of a quaternion is denotes as Q 1 , and has
observed as the deformed shape of Example-1 is attained through
the property QQ 1 ¼ Q 1 Q ¼ 1. It is constructed as Q 1 ¼ Q =
different loading paths.
kQ k. The quaternion product, also known as Grassman product, is
a non-commutative product of two quaternions. Given two quater-
6. Discussion and conclusions nions ðp0 ; pÞ and ðq0 qÞ, their quaternion product ðr 0 ; rÞ is obtained by
We have used the rotation vector parametrization in the geomet- ðr 0 ; rÞ ¼ ðp0 ; pÞ ðq0 ; qÞ ðA:1Þ
rically exact beam theory with emphasis on the interpolation of to-
via the formula r 0 ¼ p0 q0 ðp; qÞ and r ¼ p0 q þ q0 p þ p q. The
tal rotation variables leading to geometric consistency, objectivity
above product is non-commutative, associative and distributive.
of strain measures and path-independence of solutions. Among
One of the main advantages of quaternions over orthogonal tensors
existing interpolation strategies, the interpolation of relative rota-
is the reduced cost of multiplication. Moreover, quaternion multi-
tion [20] is considered the best; but it is computationally expensive.
plications are less prone to round-off errors.
Presently, we have proposed and extensively explored a two-noded
Generally unit quaternions are used for rotations. Unit quater-
isoparametric quaternion interpolation (slerp), which does not need
nions produce a unit sphere S3 embedded in R4 ; they are subject
to compute any local rotations, as an computationally efficient
to the constraint q20 þ kqk2 ¼ 1. Hence, coordinates of unit quater-
alternative. slerp does not require an extraction scheme (such as
nions are not independent. They may be used to construct an
the Spurrier algorithm) for relative rotation vectors, typically
orthogonal tensor according to
needed with the interpolation of local rotations. This scheme is
somewhat at variance with the idea that relative rotation is essen- R ¼ ð2q20 1ÞI þ 2q0 q
~ þ 2q q or R ¼ HðQ Þ: ðA:2Þ
tial to attain geometric consistency. It is proved that slerp retains
the objectivity of strain measures in the discretized form. A promi-
nent aspect of this work is the proof that the slerp interpolation of 4
The symbol ð; Þ also denotes the dot product of vectors; the distinction should be
total quaternions is equivalent to the interpolation of relative rota- clear from the context.
Z
The symbol ð ~Þ denotes the skew-symmetric matrix corresponding
Dg Gint ðdU; UÞ DU ¼ ð dUÞ DðUÞð DUÞds; ðB:2Þ
to ðÞ. For every nonzero quaternion, a unique rotation can be ob- ½0;L
tained. For a non-unit quaternion, the rotation matrix is given by
where
1 2d 3 2 3
R¼ ð2q20 I þ 2q0 q
~ þ 2q qÞ 1: ðA:3Þ I 0 0 0 e
R N
kQ k ds
6 d 7 6 7
¼4 0 ds
I 5; DðUÞ ¼ 4 0 0 0 5;
This formula for converting from a unit quaternion to an orthogonal ðB:3Þ
0 I e t
NR e
M A
matrix involves no trigonometric functions. From Eq. (A.3) it is clear
that the map from the set of unit quaternions to SO(3) is 2-1, since A ¼ ½ðRt x00 Þ N ððRt x00 Þ NÞI þ K M ðK MÞI:
ðq0 ; qÞ and ðq0 ; qÞ yield the same R. There are no approximations in this linearization, i.e. the explicit
Let the unit quaternions ðp0 ; pÞ; ðq0 ; qÞ; ðr 0 ; rÞ be associated with expressions are exact.
orthogonal matrices P; Q ; R 2 SOð3Þ. Matrix and quaternion multi-
plications are 1-1, i.e. the product R ¼ PQ is 1-1 with
Appendix C. Expression for residual in T R SOð3Þ
ðr 0 ; rÞ ¼ ðp0 ; pÞ ðq0 ; qÞ. In absence of roundoff errors during qua-
ternion multiplication (Eq. (A.1), orthogonality of the tensor ob-
The expressions for strain–displacement matrix is given by
tained by Eq. (A.2) is preserved. Even when roundoff errors are
" #
present, the orthogonality property may be precisely satisfied via N0i Rt Ni Rg
t 0
x0
quaternion multiplication along with the renormalization Bhi ðUÞ ¼ : ðC:1Þ
0 e
N0i I þ N i K
procedure:
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi The nodal internal and external force vectors are given by
ðr0 ; rÞ#ðr 0 =l1 ; r=l1 Þ; where l1 ¼ r 20 þ ðr; rÞ: Z 1
Ginti ¼ Bht
i ðnÞrðnÞJ dn; ðC:2Þ
If there are no roundoff errors then l1 ¼ 1. Note that the inverse of a 1
Z 1
unit quaternion and the product of two unit quaternions are also
unit quaternions.
Gexti ¼ Bht
i ðnÞF ext ðnÞJ dn: ðC:3Þ
1
The rotation vector W may be transformed to a unit quaternion
via the following formula: Nodal residual force is obtained by the sum
Gi ¼ Ginti Gexti : ðC:4Þ
ðq0 ; qÞ ¼ ðcos h=2; sin h=2 eÞ where e ¼ W=h; h ¼ kWk: ðA:4Þ
The extraction of W from ðq0 ; qÞ is not unique. References

It may be observed that quaternion product of pure quaternions
made of vector e, is zero quaternion, i.e. ð0; eÞ ð0; eÞ ¼ [1] I. Romero, The interpolation of rotations and its application to finite element
ð1; f0; 0; 0gÞ. The similarity with unit length complex numbers models of geometrically exact rods, Comput. Mech. 34 (2004) 121–133.
[2] J.C. Simo, L. Vu-Quoc, A three-dimensional finite rod model part II:
ðcos a; i sin aÞ is clear; here h=2 is written as a. Similarly the Euler’s computational aspects, Comput. Methods Appl. Mech. Engrg. 58 (1986) 79–116.
identity for complex numbers can be extended for quaternions [3] J.C. Simo, A finite strain beam formulation. The three-dimensional dynamic
as problem: part I, Comput. Methods Appl. Mech. Engrg. 49 (1985) 55–70.
[4] L. Vu-Quoc, Dynamics of flexible structures performing large overall motions:
expðð0; aeÞÞ ¼ ðcos a; sin a eÞ: ðA:5Þ a geometrically-nonlinear approach, PhD thesis, UC Berkeley, Dissertation, ERL
Memorandum UCB/ERL M86/36, 1986.
This exponential can be evaluated symbolically by substituting [5] A. Cardona, M. Géradin, A beam of the finite element nonlinear theory with
finite rotations, Int. J. Numer. Methods Engrg. 26 (1988) 2403–2438.
ð0; aeÞ in the standard power series of exponential function of real [6] J.C. Simo, L. Vu-Quoc, A geometrically exact rod model incorporating shear and
variables, and then using the identity ð0; eÞ ð0; eÞ ¼ ð1; f0; 0; 0gÞ. torsion warping deformation, Int. J. Numer. Methods Engrg. 27 (1991) 371–
Using this identity, one may define the power of a quaternion Q as 393.
[7] A. Ibrahimbegović, F. Frey, I. Kozar, Computational aspects of vector-like
ðQ Þt ¼ ððcos a; sin a eÞÞt ¼ expðð0; t aeÞÞ ¼ ðcos ta; sin t a eÞ: ðA:6Þ parametrization of three-dimensional finite rotations, Int. J. Numer. Methods
Engrg. 38 (1995) 3653–3673.
The logarithm of a unit quaternion may be defined as
[8] A. Ibrahimbegović, Finite element implementation of Reissner’s geometrically
logðQ Þ ¼ logðexpðð0; aeÞÞÞ ¼ ð0; aeÞ: ðA:7Þ nonlinear beam theory: three dimensional curved beam finite element,
Comput. Methods Appl. Mech. Engrg. 122 (1995) 10–26.
t [9] G. Jelenić, M. Saje, A kinematically exact space finite strain beam model finite
The derivative of the function ðQ Þ , where Q is a constant unit qua-
element formulation by generalized virtual work principle, Comput. Methods
ternion is Appl. Mech. Engrg. 120 (1995) 131–161.
[10] W.M. Smolenski, Statically and kinematically exact nonlinear theory of rods
d
ðQ Þt ¼ ðQ Þt logðQ Þ: ðA:8Þ and its numerical verification, Comput. Methods Appl. Mech. Engrg. 178
dt (1999) 89–113.
[11] F. McRobie, J. Lasenby, Simo–Vu-Quoc rods using clifford algebra, Int. J. Numer.
Methods Engrg. 45 (4) (1999) 377–398.
Appendix B. Linearization of the weak form on T R SOð3Þ [12] J. Mäkinen, Total Lagrangian Reissner’s geometrically exact beam element
without singularities, Int. J. Numer. Methods Engrg. 70 (2007) 1009–1048.
[13] P.E. Nikravesh, Computer Aided Analysis of Mechanical Systems, Prentice Hall,
Linearized Gint is decomposed, as usual, into a sum of material Englewood Cliffs, New Jersey, 1988.
and geometric tangent stiffnesses, derivable, respectively, from lin- [14] M. Géradin, A. Cardona, Flexible Multibody Dynamics: A Finite Element
Approach, John Wiley and Sons, New York, 2001. ISBN 0-471-48990-5.
earizations of r and de of Eq. (6). The relevant expressions are pro-
[15] K. Spring, Euler parameters and the use of quaternion algebra in the
vided below: manipulation of finite rotations: a review, Mech. Machine Theory 21 (1986)
2 3t 365–373.
Z
Rt dðdx 0Þ
þ Rt dx 0
dH [16] M. Géradin, D. Rixen, Parametrization of finite rotations in computational
Dm Gint ðdU; UÞ DU ¼ 4 ds ds 5 dynamics: a review, Revue européenne des éléments finis 4 (1995) 497–553.
ddH
½0;L þ K d H [17] A. Ibrahimbegović, On the choice of finite rotation parameters, Comput.
2 ds 3 Methods Appl. Mech. Engrg. 149 (1997) 49–71.
Rt dðDdsx0 Þ þ Rt dxds0 DH [18] A. Ibrahimbegović, R.L. Taylor, On the role of frame invariance in structural
C4 5ds; ðB:1Þ mechanics models at finite rotations, Comput. Methods Appl. Mech. Engrg. 191
dDH
ds
þ K DH (2002) 5159–5176.
[19] M.A. Crisfield, G. Jelenić, Objectivity of strain measures in the geometrically [23] I. Romero, F. Armero, An objective finite element approximation of the
exact three-dimensional beam theory and its finite element kinematics of geometrically exact rods and its use in the formulation of an
implementation, Proc. Royal Soc. Lond., Ser. A – Math. Phys. Engrg. Sci. energy-momentum conserving scheme in dynamics, Int. J. Numer. Methods
455 (1999) 1125–1147. Engrg. 54 (12) (2002) 1683–1716.
[20] G. Jelenić, M. Crisfield, Geometrically exact 3D beam theory: implementation [24] K. Shoemake, Animating rotation with quaternion curves, in: SIGGRAPH’85:
of a strain-invariant finite element for statics and dynamics, Comput. Methods Proceedings of the 12th Annual Conference on Computer Graphics and
Appl. Mech. Engrg. 171 (1999) 141–171. Interactive Techniques, ACM, New York, NY, USA, 1985, pp. 245–254.
[21] S. Ghosh, D. Roy, A frame-invariant scheme for the geometrically exact beam [25] J.B. Kuipers, Quaternions and Rotation Sequences, Princeton University Press,
using rotation vector parametrization, Comput. Mech., submitted for Princeton, New Jersey, 1999.
publication. [26] J. Huang, Lectures on Representation Theory, World Scientific, Singapore,
[22] P. Betsch, P. Steinmann, Frame-indifferent beam element based upon the 1999.
geometrically exact beam theory, Int. J. Numer. Methods Engrg. 54 (2002) [27] Y. Goto, Y. Watanabe, T. Kasugai, M. Obata, Elastic buckling phenomenon
1775–1788. applicable to deployable rings, Int. J. Solids Struct. 29 (1992) 893–909.

275-Consistent Quaternion Interpolation For Objective Finite Element Approximation of Geometrically Exact Beam

Uploaded by

Copyright:

Available Formats

275-Consistent Quaternion Interpolation For Objective Finite Element Approximation of Geometrically Exact Beam

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

275-Consistent Quaternion Interpolation For Objective Finite Element Approximation of Geometrically Exact Beam

Uploaded by

Copyright:

Available Formats

Comput. Methods Appl. Mech. Engrg.

198 (2008) 555–571

Contents lists available at ScienceDirect

Comput. Methods Appl. Mech. Engrg.

Consistent quaternion interpolation for objective ﬁnite element

2.2.4. Non-orthogonal direct interpolation followed by

such that c1 ð1Þ ¼ 1; c1 ð1Þ ¼ 0; c2 ð1Þ ¼ 0; c2 ð1Þ ¼ 1. We want Rh ¼ 2 ðI þ Rh Þ þ ð2qh0 qh0 I þ qh  qh

3.5. Singularity free additive interpolation of total quaternions

The interpolation described in Section 2.2.4 [1] is susceptible to

sin Xþ Eq. (A.2). Noting that Q þh ¼ Q r Q h ¼ Q r ðQ h logðQ ÞÞ, we get

We have used the notations ¼ Qþ

Q þh ¼ ðQ r Q hÞ 5.2. Example-2. Application to large rotation: deployment of a ring

Table 1 example has been considered via an updated Lagrangian formula-

Change in strain energy

5.4. Example-4. Objectivity test: rigid motion of a deformed beam

changes when the beam undergoes rigid body rotations. Stresses

the computed response remaining frame-invariant. Based on a

low the support rotation to preserve the stress state corresponding

5.5. Example-5. Path-independence

We take up the well known example of a 45°-bent cantilever.

This problem has been considered in [20,18] to check path-inde-

beam elements. The beam is loaded with a vertical tip load of

Change in strain energy

Bending moment (1)

Fig. 12. Change in strain energy under rigid rotation.

The extraction of W from ðq0 ; qÞ is not unique. References

You might also like

such that c1 ð1Þ ¼ 1; c1 ð1Þ ¼ 0; c2 ð1Þ ¼ 0; c2 ð1Þ ¼ 1. We want Rh ¼ 2 ðI þ Rh Þ þ ð2qh0 qh0 I þ qh qh