Applications of Geometric

Algebra I
Chris Doran
Cavendish Laboratory
Cambridge University

[email protected]
www.mrao.cam.ac.uk/~clifford
 3D Algebra
• 3D basis consists of 8 elements
• Represent lines, planes and volumes, from a
common origin

Grade 0 Grade 1 Grade 2 Grade 3

Scalar Vector Bivector Trivector
1 e 1 , e2 , e 3 e1 e 2 , e2 e 3 , e 3 e 1 I
 Algebraic Relations
• Generators anticommute e 1 e2  e 2 e1
• Geometric product ab  a  b  a  b
1
• Inner product a  b  2
ab  ba
• Outer product a  b  12 ab  ba
• Bivector norm e 1  e2  2  1
• Trivector I  e 1 e2 e 3
• Trivector norm I 2  1
• Trivectors commute with all other elements
 Lines and Planes
• Pseudoscalar gives a map between lines and
planes a
B  Ia
a  IB B
• Allows us to recover the vector (cross)
product
a  b  I a  b
• But lines and planes are different
• Far better to keep them as distinct entities
 Quaternions
• For the bivectors set
i  e 2 e3 , j  e 3 e1 , k  e 1 e2
• These satisfy the quaternion relations
i 2  j2  k 2  ijk  1
• So quaternions embedded in 3D GA
• Do not lose anything, but
– Vectors and planes now separated
– Note the minus sign!
– GA generalises
 Reflections
• Build rotations from reflections
• Good example of geometric product – arises
in operations
b
a   a  nn a
a   a  a  nn
n
• Image of reflection is
b  a   a   a  2a  nn
 a  an  nan  nan
 Rotations
• 2 successive reflections give a rotation

Initial vector in red

Reflection in green
Rotated in blue
 Rotations
• Direction perpendicular to the two reflection
vectors is unchanged
• So far, will only talk about rotations in a plane
with a fixed origin (more general treatment
later)
 Algebraic Formulation
• Now look at the algebraic expression for a
pair of reflections
a  mnanm  mnanm
• Define the rotor R  mn
• Rotation encoded algebraically by
a  RaR  R   nm
• Dagger symbol used for the reverse
 Rotors
• Rotor is a geometric product of 2 unit vectors
R  mn  cos  m  n
• Bivector has square
m  n  2  mn  cosnm  cos   sin2 
• Used to the negative square by now!
• Introduce unit bivector B  m  n
sin
• Rotor now written
R  cos  sinB
 Exponential Form
• Can now write R  expB 
• But:
– rotation was through twice the
angle between the vectors
– Rotation went with orientation n  m
• Correct these, get double-sided, half-
angle formula
a  RaR  R  expB /2
• Completely general!
 Rotors in 3D
• Can rewrite in terms of an axis via
R  expIn/2 
• Rotors even grade (scalar + bivector in 3D)
• Normalised: RR   mnnm  1
• Reduces d.o.f. from 4 to 3 – enough for a
rotation
• In 3D a rotor is a normalised, even element
• The same as a unit quaternion
 Group Manifold
• Rotors are elements of a 4D space,
normalised to 1
• They lie on a 3-sphere
• This is the group manifold
• Tangent space is 3D
• Natural linear structure for rotors
• Rotors R and –R define the same rotation
• Rotation group manifold is more complicated
 Comparison
• Euler angles give a standard parameterisation
of rotations
cos cos  cos sin sin  sin cos  cos sin cos sin sin
cos sin  cos cos sin  sin sin  cos cos cos  sin cos
sin sin sin cos cos

• Rotor form far easier

R  expe 1 e2 /2  expe 2e 3 /2  expe1 e 2 /2 
• But can do better than this anyway – work
directly with the rotor element
 Composition
• Form the compound rotation from a pair of
successive rotations
a  R 2 R 1 aR 1 R 2
• Compound rotor given by group combination
law R  R 2 R 1
• Far more efficient than multiplying matrices
• More robust to numerical error
• In many applications can safely ignore the
normalisation until the final step
 Oriented Rotations
• Rotate through 2 different orientations
• Positive Orientation e2
R  expe 1 e 2 /2
 expe 1 e 2 /4 e1
• Negative Orientation
S  expe 1 e 2 /2
 expe 1 e 2 3/4  R
• So R and –R encode the same absolute
rotation, but with different orientations
 Lie Groups
• Every rotor can be written as expB/2 
• Rotors form a continuous (Lie) group
• Bivectors form a Lie algebra under the
commutator product
• All finite Lie groups are rotor groups
• All finite Lie algebras are bivector algebras
• (Infinite case not fully clear, yet)
• In conformal case (later) starting point of
screw theory (Clifford, 1870s)!
 Interpolation
• How do we interpolate between 2 rotations?
• Form path between rotors
R0  R 0
R  R 0 expB 
R1  R 1

• Find B from expB   R 0R 1
• This path is invariant. If points transformed,
path transforms the same way
• Midpoint simply R1/2   R 0 expB/2 
• Works for all Lie groups
 Interpolation - SLERP
• For rotors in 3D can do even better! R1
• View rotors as unit vectors in 4D
• Path is a circle in a plane 
• Use simple trig’ to get SLERP
R0
R  1 sin1    R 0  sin R 1 
sin
• For midpoint add the rotors and normalise!
sin/2 
R1/2  R 0  R 1 
sin
 Applications
• Use SLERP with spline constructions for
general interpolation
• Interpolate between series of rigid-body
orientations
• Elasticity
• Framing a curve
• Extend to general
transformations
 Linearisation
• Common theme is that rotors can linearise
the rotation group, without approximating!
• Relax the norm constraint on the rotor and
write RAR   A 1
• ψ belongs to a linear space. Has a natural
calculus.
• Very powerful in optimisation problems
involving rotations
• Employed in computer vision algorithms
 Recovering a Rotor
• Given two sets of vectors related by a
rotation, how do we recover the rotor?
• Suppose b i  Ra i R 
• In general, assume not orthogonal.
1
• Need reciprocal frame a
a 2  a3I
a1 
a1  a 2  a3 I  a3
i a2
• Satisfies a  a j   ij
 Recovering a Rotor II
• Now form even-grade object
b i a i  Ra i   Ba i  R3  B  1  4R
• Define un-normalised rotor
  b i ai  1
• Recover the rotor immediately now as

R
||
• Very efficient, but
– May have to check the sign
– Careful with 180o rotations
 Rotor Equations
• Suppose we take a path in rotor space R
• Differentiating the constraint tells us that
d RR    R  R   RR   0
d
• Re-arranging, see that
R  R   R  R     Bivector
• Arrive at rotor equation
R    12 B R
• This is totally general. Underlies the theory of
Lie groups
 Example
• As an example, return to framing a curve.
• Define Frenet frame n
b
• Relate to fixed frame
 t
t, n, b  R e i R
• Rotor equation
R    12 R    1 e2 e 1   2 e 3 e2
• Rotor equation in terms of curvature and
torsion
 Linearisation II
• Rotor equations can be awkward (due to
manifold structure)
• Linearisation idea works again
• Replace rotor with general element and write
    12 B
• Standard ODE tools can now be applied
(Runge-Kutta, etc.)
• Normalisation of ψ gives useful check on
errors
 Elasticity
• Some basics of elasticity (solid mechanics):
– When an object is placed under a stress
(by stretching or through pressure) it
responds by changing its shape.
– This creates strains in the body.
– In the linear theory stress and strain are
related by the elastic constants.
– An example is Hooke’s law F=-kx, where k
is the spring constant.
– Just the beginning!
 Bulk Modulus
• Place an object under uniform
pressure P
• Volume changes by

 P  B V
V

• B is the bulk modulus

• Definition applies for small
pressures (linear regime)
 Shear Modulus

• Sheers produced by
combination of tension  
and compression
• Sheer modulus G is
Shear stress / angle 

G  θ
2
 LIH Media
• The simplest elastic systems to consider are
linear, isotropic and homogeneous media.
• For these, B and G contain all the relevant
information.
• There are many ways to extend this:
– Go beyond the linearised theory and treat
large deflections
– Find simplified models for rods and shells
 Foundations
• Key idea is to relate the spatial configuration
to a ‘reference’ copy.

x y=f(x)

• y=f(x) is the displacement field. In general,

this will be time-dependent as well.
 Paths
• From f(x) we want to extract information about
the strains. Consider a path

• Tangent vectors map to

fx  a  fx  a  fx  Fa
• F(a)=F(a;x) is a linear function of a. Tells us
about local distortions.
 Path Lengths
• Path length in the reference body is
1/2
 d d
dx  dx d
• This transforms to
F x 
  F x  1/2
  d
• Define the function G(a), acting entirely in the
reference body, by
Ga  F Fa
 The Strain Tensor
• For elasticity, usually best to ‘pull’ everything
back to the reference copy
• Use same idea for rigid body mechanics
• Define the strain tensor from G(a)
– Most natural is
1
Ea  2
Ga  a
– An alternative (rarely seen) is
1
Ea  2
lnGa
 The Stress Tensor
• Contact force between 2 surfaces is a linear
function of the normal (Cauchy)
n τ(n)

• τ(n)=τ(n;x) returns a vector in the material

body. ‘Pull back’ to reference copy to define
Tn  F 1 n
 Constitutive Relations
• Relate the stress and the strain tensors in the
reference configuration
• Considerable freedom in the choice here
• The simplest, LIH media have
Ta  2GEa  B  23 GtrE a
• Can build up into large deflections
• Combined with balance equations, get full set
of dynamical equations
• Can get equations from an action principle
 Problems
• Complicated, and difficult numerically
• In need of some powerful advanced
mathematics for the full nonlinear theory
(FEM…)
• Geometric algebra helps because it
– is coordinate free
– integrates linear algebra and calculus
smoothly
• But need simpler models
• Look at models for rods and beams
 Deformable Rod
• Reference configuration is a cylinder

λ Line of
centre of
y mass
σ
x

Configuration encoded
in a rotor y  x, t  RR 
 Technical Part
• Spare details, but:
• Write down an action integral
• Integrate out the coordinates over each disk
• Get (variable) bending moments along the
centre line
• Carry out variational principle
• Get set of equations for the rotor field
• Can apply to static or dynamic configurations
 Simplest Equations
• Static configuration, and ignore stretching
• Have rotor equation
dR   1 R
B
d 2
• Find bivector from applied couple and elastic
constants. I(B) is a known linear function of
these mapping bivectors to bivectors
 B  I 1 R  CR
• Integrate to recover curve
x   Re1 R 
 Example
• Even this simple set of
equations can give highly
complex configurations!

Small, linear
deflections build
up to give large
deformations
 Summary
• Rotors are a general purpose tool for
handling rotations in arbitrary dimensions
• Computationally more efficient than matrices
• Can be associated with a linear space
• Easy to interpolate
• Have a natural associated calculus
• Form basis for algorithms in elasticity and
computer vision
• All this extends to general groups!
 Further Information
• All papers on Cambridge GA group website:
www.mrao.cam.ac.uk/~clifford
• Applications of GA to computer science and
engineering are discussed in the proceedings
of the AGACSE 2001 conference.
www.mrao.cam.ac.uk/agacse2001
• IMA Conference in Cambridge, 9th Sept 2002
• ‘Geometric Algebra for Physicists’ (Doran +
Lasenby). Published by CUP, soon.
 Revised Timetable
• 8.30 – 9.15 Rockwood • 1.30 – 2.00 Doran
Introduction and outline Beyond Euclidean
of geometric algebra Geometry
• 9.15 – 10.00 Mann • 2.00 – 3.00 Hestenes
Illustrating the algebra I Computational Geometry
• 10.00 -10.15 Break • 3.00 – 3.15 Break
• 10.15 – 11.15 Doran • 3.15 – 4.00 Dorst
Applications I Illustrating the algebra II
• 11.15 – 12.00 Lasenby • 4.00 – 4.30 Lasenby
Applications II Applications III
• 4.30 Panel