On Systems of Linear Quaternion Functions

arXiv:math/0702084v1 [math.RA] 4 Feb 2007

Todd A. Ell
February 2, 2008

Abstract
A method of reducing general quaternion functions of first degree, i.e.,
linear quaternion functions, to quaternary canonical form is given. Linear
quaternion functions, once reduced to canonical form, can be maintained
in this form under functional composition. Furthermore, the composi-
tion operation is symbolically identical to quaternion multiplication, mak-
ing manipulation and reduction of systems of linear quaternion functions
straight forward.

1 Introduction
Real linear functions take the monomial form: f (x) = mx, where x, m ∈ R.
By contrast the general quaternion linear function has a multi-nomial form:
P
X
f (q) = mp q n p , (1)
p=1

where all factors are quaternion valued, i.e., q, mp , np ∈ H. The number of terms
in the summation can be arbitrarily large due to quaternion multiplication being
non-commutative.
Unlike Hardy [1, p.199-200] who studied the inversion of these functions
or Hathaway [2, chp. 4] who was interested in their properties, this work ad-
dresses the reduction of systems or networks of such functions, formed by scalar-
weighted sums and functional composition. The number of monomial terms in
such systems quickly become intractable.
Herein is a method for reducing such systems in an efficient and transparent
manner. Two reduction methods are key to achieving this goal. First is conver-
sion of general linear quaternion functions, as given in equation (1), to at most
a four-tuple form. It will be shown that equation (1) can always be reduced to
the form:
f (q) = A q + B q i + C q j + D q k, (2)
where A, B, C, D ∈ H and i, j, k are the standard hypercomplex operators. Lin-
ear quaternion functions of this type shall be said to be in quaternary canon-
ical form. Just as a quaternion can be associated with a 4-tuple of reals as

1
a1 + bi + cj + dk ⇔ (a, b, c, d), likewise the canonical linear function can be
associated with a 4-tuple of quaternions as
A q + B q i + C q j + D q k ⇔ {A, B, C, D} .
The four-tuple of quaternions will be used as a shorthand notation for the
canonical form.
The linear sum of two such functions, say f1 and f2 , is immediately reducible
to canonical form, i.e.,
f3 (q) = f1 (q) + f2 (q) = A3 q + B3 q i + C3 q j + D3 q k
where
A3 = A1 + A2 , B3 = B1 + B2 , C3 = C1 + C2 , D3 = D1 + D2 .
However, functional composition of two such functions naturally expand into
sixteen terms, which require reduction back to four terms. The number of terms
in multiple compositions expand exponentially. The second reduction method
provides a simple composition rule which automatically yields a quaternary
form, provided the two functions being composed are already in this form.
The next section provides preliminary material to introduce notation and
nomenclature used in this work, as well as, matrix and involution operations
useful in deriving the main results. Section 3 presents the method of reduction to
quaternary form using either equivalent matrix operations or simpler involution
operators. The matrix-based version provides a numerically efficient algorithm,
while the involution-based version is better suited to symbolic analysis. Section
4 presents the composition algorithm.
Motivation behind this work is in the use of arrays of linear quaternion
functions as filtering operations on quaternion valued data arrays. This allows
for complex geometric operations on three- and four-dimensional data sets. Only
after the basic manipulation of systems of such functions are detailed can the
more difficult problem of linear quaternion filters be addressed.

2 Preliminaries
2.1 Quaternions
The quaternions, H, are a four dimensional, non-commutative division algebra
over the reals, R. Multiplication is defined by the hypercomplex operator rules
i2 = j 2 = k 2 = ijk = −1.
The typical quaternion is written in Cartesian form as
p = p0 + p1 i + p2 j + p3 k
where p0,1,2,3 ∈ R.
The quaternion conjugate, p̄, is defined as negating the non-real part, i.e.,
p̄ = p0 − p1 i − p2 j − p3 k

2
2.2 Quaternion-Matrix Homomorphism
The equivalence between quaternion and 4×4 real matrix multiplication was
first established by Peirce [3], later verified by Spottiswoode [4], made explicit
by Cayley [6, 7] and later in Tait’s book [5]. Walker [8] provides a concise
introduction and Ickes [9] uses this equivalence to provide fast numerical com-
putations on linear quaternion equations.
These quaternion matrix representations are not unique. Farebrother et al.
[10] identified 48 different representations, not considering all automorphisms
of the quaternion algebra. Later, Ducati [11] divided these representations into
two groups based on the left and right action of the hypercomplex operators.
Mackey [12] used this concept to solve the eigenvalue problem. Fassbender et
al. [13] used them to solve the Hamiltonian eigenproblem. DeLeo and Rotelli
[14] used the same concept applied to complex 2×2 matrices.
Using the hypercomplex operator product rules, the quaternion product of
two arbitrary quaternions, q, p ∈ H, written in Cartesian form as
p = p0 + p1 i + p2 j + p3 k
q = q0 + q1 i + q2 j + q3 k
is
r0 = p0 q0 − p1 q1 − p2 q2 − p3 q3
r1 = p1 q0 + p0 q1 − p3 q2 + p2 q3
r2 = p2 q0 + p0 q2 + p3 q1 − p1 q3
r3 = p3 q0 + p0 q3 − p2 q1 + p1 q2
where r = pq = r0 + r1 i + r2 j + r3 k. Gathering terms into matrix-vector
notation we obtain two distinct matrix-vector products: the standard form
    
r0 p0 −p1 −p2 −p3 q0
 r1   p1 p0 −p3 p2   q1 
 r2  =  p2 p3 (3)
    
p0 −p1   q2 
r3 p3 −p2 p1 p0 q3
which preserves the order of the product, and the transmuted form
    
r0 q0 −q1 −q2 −q3 p0
 r1   q1 q0 q3 −q2    p1 
 
 r2  =  q2 −q3 q0 (4)
  
q1   p2 
r3 q3 q2 −q1 q0 p3
which reverses the order.
Following Ickes [9], let the standard matrix form be denoted [p] and the
transmuted matrix form as [q]† , where the ‘†’ signifies transposing the lower-
right 3×3 sub-matrix.
In the quaternion product r = pq, if we regard p as a left-handed operator
on the state variable q, then the p-operator is encoded as a 4×4 real matrix and
the q-state as a real 4×1 column vector. Conversely, if q is regarded as a right-
handed operator on the state variable p, then the operator is again encoded as
a real matrix (transmuted to the left) and the state as a real vector.

3
2.3 Quaternion Operators
Since left and right quaternion multiplications are unique, we can view various
combinations as independent operators. Following De Leo and Rotelli [14], one
can define a bar -operator notation to provide a shorthand representation for
these combinations. The bar -operator, denoted e1 |e2 , is defined as:

e1 |e2 (q) → e1 (q) e2 , where en ∈ {1, i, j, k}

so there are sixteen such operators as listed in tabular form below:

1|1, 1|i, 1|j, 1|k,
i|1, i|i, i|j, i|k,
j|1, j|i, j|j, j|k,
k|1, k|i, k|j, k|k.

These quaternion operators are consistent with the standard hypercomplex

operators and have similar operator formulae
2 2 2
(i|1) = (j|1) = (k|1) = (i|1) (j|1) (k|1) = −1|1
(1|i)2 = (1|j)2 = (1|k)2 = (1|i) (1|j) (1|k) = −1|1

so that various operator equations can be manipulated directly, e.g., (i|j) (j|k) =
(ij|kj) = − (k|i).
Using the standard and transmuted matrix homomorphisms given in the
†
previous section, i.e., e1 |e2 → [e1 ] [e2 ] , each can be written as a matrix. For
example,  
0 0 +1 0
†  0 0 0 −1 
(k|i) → [k] [i] =   +1 0
.
0 0 
0 −1 0 0
The basic quaternion operator formulae can be extended by forming weighted
linear combinations. Linearity of the quaternion operations implies that for four
arbitrary quaternions, A, B, C, D ∈ H, written in component form as
A = α0 + α1 i + α2 j + α3 k,
B = β0 + β1 i + β2 j + β3 k,
C = γ0 + γ1 i + γ2 j + γ3 k,
D = δ0 + δ1 i + δ2 j + δ3 k,

when combined with four operators 1|1, 1|i, 1|j, and 1|k, respectively, form the
four extended operators
(A|1) = α0 (1|1) + α1 (i|1) + α2 (j|1) + α3 (k|1) ,
(B|i) = β0 (1|i) + β1 (i|i) + β2 (j|i) + β3 (k|i) ,
(C|j) = γ0 (1|j) + γ1 (i|j) + γ2 (j|j) + γ3 (k|j) ,
(D|k) = δ0 (1|k) + δ1 (i|k) + δ2 (j|k) + δ3 (k|k) .

4
 In the standard case the quaternions ±1, ±i, ±j, and ±k form a non-
abelian group of order eight, with multiplication as the group operator. In
the quaternion-operator case, these become ±1|1, ±1|i, ±1|j, ±1|k, ±i|1, ±j|1,
and ±k|1 giving a 14th order group with composition being the group operator.

2.4 Quaternion Involutions

A function, f , is an involution 1 if it has the following properties [16]:
1. is it’s own inverse: f (f (q)) = q,
2. is linear: f (q1 + q2 ) = f (q1 ) + f (q2 ), and
3. is anti-homomorphic: f (q1 q2 ) = f (q2 )f (q1 ).
A function is an anti-involution 2 if the third property, above, is homomor-
phic, i.e., f (q1 q2 ) = f (q1 )f (q2 ).
Using the quaternion bar-operators in conjunction with conjugation provides
a generalized conjugate,
q̄ ε → −εq̄ε, ε ∈ {i, j, k}
for any q ∈ H, which are involutions under these definitions [15]. The conjugate
of these functions,
q̄ ε → −εqε, ε ∈ {i, j, k}
are anti-involutions [17].
For an arbitrary quaternion q = q0 + iq1 + jq2 + kq3 the following identities
hold in terms of involutions
q1 = 21 q̄ i + q̄ i

q2 = 12 q̄ j + q̄ j


(5)
q3 = 21 q̄ k + q̄ k

which are analogous to the formula for the scalar part of a quaternion using the
standard conjugate
q0 = 21 (q + q̄) (6)
except they extract the remaining real-valued component-parts of any quater-
nion.
Finally, it should be pointed out that all of these functions (generalized
conjugates, and component-part functions) are readily placed in quaternary
form. This is achieved by first expressing the standard quaternion conjugate in
involutory form, i.e.,
1
(−q + iqi + jqj + kqk) .
q̄ = 2

Hence its canonical four-tuple is q̄ = − 21 , 2i , 2j , k2 . Proof of this involutory



form is given in [17].

1 This is also known as an involutory anti-automorphism.
2 Likewise, this is also known as an involutory automorphism.

5
3 Reduction to Canonical Form
3.1 Matrix Method
At its core, the matrix method is the process for taking an arbitrary operator
matrix and converting it to the sum of four operator matrices, each matrix
equivalent to an operator quaternion.
The matrix-quaternion homomorphism of section 2.2 suggest that the indi-
vidual terms in equation (1) can be converted into matrix form as
 
† †
mp (qnp ) → [mp ] [np ] ~q = [mp ] [np ] ~q

where ~q denotes the state-vector form of quaternion q.

If a numerical reduction to the problem is desired, then the following matrix
equation could be used
( P )
†
X
f (q) → [mp ] [np ] ~q.
p=1

Even though the resulting 4-vector represents a quaternion, the intermediate

†
matrices [mp ] [np ] cannot be decoded into a equivalent single quaternion since
in general it does not have the correct pattern of elements. A general 4×4
real matrix has sixteen degrees of freedom, whereas a quaternion has only four
(whether or not it is in matrix form). So arbitrary real matrices cannot be
written as a single quaternion. However, it can be written as the direct sum of
four matrices, each matrix equivalent to a quaternion.
Let A, B, C, D ∈ H, be defined as
A = α0 + α1 i + α2 j + α3 k,
B = β0 + β1 i + β2 j + β3 k,
C = γ0 + γ1 i + γ2 j + γ3 k,
D = δ0 + δ1 i + δ2 j + δ3 k,
then each of the above quaternions contains four degrees of freedom. Hence the
combined equation
(A|1) + (B|i) + (C|j) + (D|k)
has a total of sixteen; the number necessary to encode an arbitrary 4×4 matrix.
Writing each quaternion operator in matrix form as
 
α0 −α1 −α2 −α3
 α1 α0 −α3 α2 
(A|1) → [A] [1]† =  ,
 α2 α3 α0 −α1 
α3 −α2 α1 α0
 
−β1 −β0 β3 −β2
†  β0 −β1 −β2 −β3 
(B|i) → [B] [i] =  β3 −β2 β1
,
β0 
−β2 −β3 −β0 β1

6
  
−γ2 −γ3 −γ0 γ1
†  −γ3 γ2 −γ1 −γ0 
(C|j) → [C] [j] =  ,
 γ0 −γ1 −γ2 −γ3 
γ1 γ0 −γ3 γ2
and  
−δ3 δ2 −δ1 −δ0
δ2 δ3 δ0 −δ1 
(D|k) → [D] [k]† = 

,
 −δ1 −δ0 δ3 −δ2 
δ0 −δ1 −δ2 −δ3
then the combined operator in matrix form is

(A|1) + (B|i) + (C|j) + (D|k) →


α0 − β1 − γ2 − δ3 −α1 − β0 − γ3 + δ2
 α1 + β0 − γ3 + δ2 α0 − β1 + γ2 + δ3

 α2 + β3 + γ0 − δ1 ···
α3 − β2 − γ1 − δ0
α3 − β2 + γ1 + δ0 −α2 − β3 + γ0 − δ1

−α2 + β3 − γ0 − δ1 −α3 − β2 + γ1 − δ0
−α3 − β2 − γ1 + δ0 α2 − β3 − γ0 − δ1 
.
α0 + β1 − γ2 + δ3 −α1 + β0 − γ3 − δ2 
α1 − β0 − γ3 − δ2 α0 + β1 + γ2 − δ3

Setting this equal to an arbitrary matrix, R, with elements {rij }, and solving
for the components of the quaternions one finds
   
α0 +r11 + r22 + r33 + r44
 α1  1  +r21 − r12 + r43 − r34 
A=  α2  = 4  +r31 − r13 − r42 + r24  ,
   (7)
α3 +r41 − r14 + r32 − r23
   
β0 +r21 − r12 − r43 − r34
 β1  1  −r11 − r22 + r33 + r44 
 β2  = 4 
B=   , (8)
−r41 − r14 − r32 − r23 
β3 +r31 + r13 − r42 − r24
   
γ0 +r31 − r13 + r42 − r24
 γ1  1  +r41 + r14 − r32 − r23 
C=
 γ2  = 4 
  , (9)
−r11 + r22 − r33 + r44 
γ3 −r21 − r12 − r43 − r34
and    
δ0 +r41 − r14 − r32 + r23
 δ1  1  −r31 − r13 − r42 − r24 
D=
 δ2  = 4 
  . (10)
+r21 + r12 − r43 − r34 
δ3 −r11 + r22 + r33 − r44
This is not as difficult as it appears since the problem decomposes into
four sets of four unknowns with each set independent of the others, e.g., the

7
coefficient set {α0 , β1 , γ2 , δ3 } is solved using only the diagonal terms in both
matrices.
The steps in reducing equation (1) to quaternary form are:

1. Map coefficients mp and np to matrix-space using equations (3-4)

†
mp → [mp ] , np → [np ] .

2. Construct single matrix R as

P
†
X
R= [mp ] [np ] .
p=1

3. Construct equivalent quaternion factors using equations (7-10)

R → (A|1) + (B|i) + (C|j) + (D|k) .

3.2 Involution Method

The matrix equivalent method outlined above is useful for numerical computa-
tions, however, it requires manipulation at the component level of each quater-
nion, mp and np . The involution method, outlined in this section, provides
an alternative solution in terms of mp and np directly, without these explicit
references, making it more suitable for symbolic analysis.
Starting with equation (1), let each np be expanded into component form as

np = wp + ixp + jyp + kzp

then the full equation can be expanded as

P
X
f (q) = mp q (wp + ixp + jyp + kzp )
p=1
P
X P
X
= mp wp q + mp xp q i
p=1 p=1
P
X P
X
+ mp y p q j + mp z p q k
p=1 p=1
" P # " P #
X X
= mp wp q + mp xp q i
p=1 p=1
" P # " P #
X X
+ mp y p q j + mp zp q k.
p=1 p=1

8
Now let
P
X P
X
A= mp wp , B= mp xp ,
p=1 p=1
P
X P
X
C= mp y p , D= mp z p ,
p=1 p=1

so that
f (q) = Aq + Bq i + Cq j + Dq k.
The problem with this formula is that we are required to decompose quaternions
np into component form, whereas we are looking for a formula that does not
require this step. The answer to this conundrum is to apply the involution
formulas (5-6). Namely, let
wp = 12 (np + n̄p ) , xp = 21 n̄ip + n̄p i,

yp = 12 n̄jp + n̄p j, zp = 12 n̄kp + n̄p k,

so that
P
X
1
A= 2 mp (np + n̄p ),
p=1
P
X
1
mp n̄ip + n̄p i,


B= 2
p=1
P
X
1
mp n̄jp + n̄p j,


C= 2
p=1
P
X
1
mp n̄kp + n̄p k.


D= 2
p=1

4 System Reduction
One could use the methods of the previous section to reduce compositions of two
linear functions, i.e., where f2 (q) ◦ f1 (q) = f2 (f1 (q)). What is not apparent
is that functional composition can be reduced in a more elegant way. This
alternative lends itself to better insight into what happens in these composition
operations.
Using bar-operator notation, composition is reduced to ordered multiplica-
tion, hence one can build a composition table showing how components of the
composed functions are reduced. The composition table for f2 ◦f1 is given below
f2 ◦ f1 A1 |1 B1 |i C1 |j D1 |k
A2 |1 +A2 A1 |1 +A2 B1 |i +A2 C1 |j +A2 D1 |k
B2 |i +B2 A1 |i −B2 B1 |1 −B2 C1 |k +B2 D1 |j
C2 |j +C2 A1 |j +C2 B1 |k −C2 C1 |1 −C2 D1 |i
D2 |k +D2 A1 |k −D2 B1 |j +D2 C1 |i −D2 D1 |1

9
 Collecting like terms, we find that
f2 ◦ f1 (q) = A3 q + B3 q i + C3 q j + D3 q k
where
A3 = A2 A1 − B2 B1 − C2 C1 − D2 D1 ,
B3 = A2 B1 + B2 A1 − C2 D1 + D2 C1 ,
C3 = A2 C1 + B2 D1 + C2 A1 − D2 B1 ,
D3 = A2 D1 − B2 C1 + C2 B1 + D2 A1 .
Careful examination of the 4-tuple resulting from a product of two quater-
nions with the 4-tuple resulting from the functional composition of two canoni-
cal equations reveals that they are transmuted versions of the same operational
formula. This is a consequence of the bar-operators having a similar operator
formula as the standard hypercomplex operator formula. If the canonical form
had been defined with the full quaternions on the right, e.g.,
f (q) = q A′ + i q B ′ + j q C ′ + k q D′ ,
then the composition formula would have been identical to the quaternion mul-
tiplication formula.

5 Conclusions
The goal of this work is the reduction and manipulation of systems of linear
quaternion functions. To this end two results have been provided. First, even
though a general linear quaternion function may contain an arbitrary number of
terms, it can always be reduced to at most four terms. This reduced function,
called the quaternary canonical form of the function, is completely specified with
a four-tuple of quaternions. Second, series (cascaded composition) and parallel
(weighted-sum) combinations of such canonical functions can be written directly
in canonical form without recourse to the previous reduction method.

References
[1] Arthur Sherburne Hardy, Elements of Quaternions, Ginn, Heath and Co.,
Boston, 1881.
[2] Arthur S. Hathaway, A Primer of Quaternions, 1896.
[3] Peirce, Linear Associative Algebra, American Journal of Mathematics, Vol.
4, p. 132, 1881.
[4] William Spottiswoode, Proceedings of London Mathematical Society, Vol.
4, pp. 156-159, 1882.
[5] P.G. Tait, An Elementary Treatise on Quaternions, 3rd Ed. Cambridge
University Press, 1890.

10
 [6] A. Cayley, A Memoir on the Theory of Matrices, Philosophical Transactions
of the Royal Society of London, 148, pp. 1737, 1858.
[7] A. Cayley, Mathematical Papers, Proceedings of London Mathematical So-
ciety, Vol. 12, pp. 311-313, 1885.
[8] M. J. Walker, Quaternions as 4-Vectors, American Journal of Physics, Vol.
24, No 7, pp. 515–522, October 1956.
[9] B. P. Ickes, A new method for performing digital control system attitude
computations using quaternions, AIAA Journal, Vol. 8, No. 1, pp. 13–17,
1970.
[10] R. W. Farebrother, J. Gro and S. Troschke, Matrix representation of quater-
nions, Lin. Alg. Appl. 362, 251-255 (2003).
[11] G. Ducati, Comments on the Matrix Representations of Quaternions,
oai:arXiv.org:math/0310319.
[12] N. Mackey, Hamilton and Jacobi meet again - quaternions and the eigen-
value problem, SIAM Journal on Matrix Analysis and Applications 16 (2)
(1995) 421-435.
[13] H. Fassbender, D. S. Mackey, N. Mackey, Hamilton and Jacobi come full
circle: Jacobi algorithms for structured hamiltonian eigenproblems, Linear
Algebra and its Applications 332 (2001) 37-80.
[14] S. DeLeo and P. Rotelli, Translations between quaternion and complex
quantum mechanics, Prog. Th. Phys. 92 pp. 917-926, 1994.
[15] I. M. Benn and R. W. Tucker, An Introduction to Spinors and Geometry
with Applications is Physics, Adam Hilger/IOP Publishing, Bristol and
New York, 1987.
[16] M. Hazewinkel, Ed., Encyclopaedia of mathematics: an updated and anno-
tated translation of the Soviet Mathematical Encyclopaedia, Kluwer, Dor-
drecht, c198894.
[17] T. A. Ell and S. J. Sangwine, Quaternion Involutions and Anti-Involutions,
Computers & Mathematics with Applications, (Accepted) October 2006.

11