You are on page 1of 44

# Quaternionic Analysis

A. Sudbery [as2@mailer.york.ac.uk]
Department of Mathematics
University of York
Heslington, York YO1 5DD
August 1977
Contents
1 Introduction 2
2 Preliminaries 5
2.1 The Algebra of Quaternions
1
. . . . . . . . . . . . . . . . . . . . . . 5
2.2 Quaternionic Differential Forms . . . . . . . . . . . . . . . . . . . . 9
3 Regular Functions 11
4 Cauchy’s Theorem and the Integral Formula 17
5 Some Immediate Consequences 25
6 Construction of Regular Functions 26
7 Regular Functions and Conformal Mappings 30
8 Homogeneous Regular Functions 32
9 Regular Power Series 39
1
1 Introduction
The richness of the theory of functions over the complex ﬁeld makes it natural to look
for a similar theory for the only other non-trivial real associative division algebra,
namely the quaternions. Such a theory exists and is quite far-reaching, yet it seems
to be little known. It was not developed until nearly a century after Hamilton’s discov-
ery of quaternions. Hamilton himself [1] and his principal followers and expositors,
Tait [2] and Joly [3], only developed the theory of functions of a quaternion variable
as far as it could be taken by the general methods of the theory of functions of several
real variables (the basic ideas of which appeared in their modern form for the ﬁrst time
in Hamilton’s work on quaternions). They did not delimit a special class of regular
functions among quaternion-valued functions of a quaternion variable, analogous to
the regular functions of a complex variable.
This may have been because neither of the two fundamental deﬁnitions of a regular
function of a complex variable has interesting consequences when adapted to quater-
nions; one is too restrictive, the other not restrictive enough. The functions of a quater-
nion variable which have quaternionic derivatives, in the obvious sense, are just the
constant and linear functions (and not all of them); the functions which can be repre-
sented by quaternionic power series are just those which can be represented by power
series in four real variables.
In 1935 R Fueter [4] proposed a deﬁnition of “regular” for quaternionic functions
by means of an analogue of the Cauchy-Riemann equations. He showed that this deﬁ-
nition led to close analogues of Cauchy’s theorem, Cauchy’s integral formula, and the
Laurent expansion [5]. In the next twelve years Fueter and his collaborators developed
the theory of quaternionic analysis. A complete bibliography of this work is contained
in ref. [6], and a simple account (in English) of the elementary parts of the theory has
been given by Deavours [7].
The theory developed by Fueter and his school is incomplete in some ways, and
many of their theorems are neither so general nor so rigorously proved as present-day
standards of exposition in complex analysis would require. The purpose of this paper
is to give a self-contained account of the main line of quaternionic analysis which
remedies these deﬁciencies, as well as adding a certain number of new results. By
using the exterior differential calculus we are able to give new and simple proofs of
most of the main theorems and to clarify the relationship between quaternionic analysis
and complex analysis.
Let Hdenote the algebra of quaternions, and let ¦1, i, j, k¦ be an orthonormal basis,
with the product on H given by the usual multiplication table (see section 2). The
typical quaternion can be written as
q = t + ix + jy + kz (1.1)
where t, x, y, z are real coordinates. Fueter deﬁned a function f : H →Hto be regular
if it satisﬁed the equation
∂f
∂t
+ i
∂f
∂x
+ j
∂f
∂y
+ k
∂f
∂z
= 0. (1.2)
2
This is an analogue of the Cauchy-Riemann equations for a complex function f : C →
C, which can be written as
∂f
∂x
+ i
∂f
∂y
= 0, (1.3)
where the variable is z = x + iy.
Fueter showed that any quaternionic function which is regular and also continu-
ously differentiable must satisfy an analogue of Cauchy’s theoremwhich can be written
as

C
Dq f = 0, (1.4)
where C is any smooth closed 3-manifold in H and Dq is a certain natural quaternion-
valued 3-form. Dq is deﬁned in section 2; it can be thought of as the quaternion repre-
senting an element δC of the 3-manifold, its magnitude being equal to the volume of
δC and its direction being normal to δC. Fueter also obtained an analogue of Cauchy’s
integral formula for such functions, namely
f(q
0
) =
1

2

∂D
(q −q
0
)
−1
[q −q
0
[
2
Dq f(q), (1.5)
where D is a domain in H in which f is regular and q
0
is a point inside D.
The complex Cauchy-Riemann equation (1.3) is equivalent to the statement that f
has a complex derivative, i.e. that there exists a complex number f

(z) such that
df = f

(z) dz. (1.6)
Fueter gave no corresponding characterisation of a regular function of a quaternion
variable, leaving the theorems (1.4) and (1.5) and the analogy with the Cauchy-Riemann
equations as sufﬁcient justiﬁcation of the deﬁnition (1.2). In this paper we will show
that a regular function can be deﬁned as one which has a certain kind of quaternionic
derivative; speciﬁcally, (1.2) is equivalent to the existence of a quaternion f

(q) such
that
d(dq ∧ dq f) = Dq f

(q). (1.7)
(the 2-form dq ∧ dq is described in section 2).
Cauchy’s theorem (1.4) and the integral formula (1.5) can be simply proved by
showing that
d(Dqf) = 0 (1.8)
and
d
¸
(q −q
0
)
−1
[q −q
0
[
2

Dq f(q) = ∆

1
[q −q
0
[
2

f(q)v (1.9)
where ∆ is the Laplacian on R
4
and v = dt ∧ dx ∧ dy ∧ dz is the standard volume
4-form. Since [q −q
0
[
−2
is the Green’s function for the Laplacian in R
4
, (1.5) follows
from (1.9). This is essentially how Fueter proved these theorems. Since both proofs
use Stokes’s theorem, they need the condition that the partial derivatives of the func-
tion should be continuous. Schuler [8] showed that this condition could be dropped
3
by adapting Goursat’s proof of Cauchy’s theorem, but he did not draw the full con-
sequences of this argument. In fact Cauchy’s theorem (1.4) can be proved for any
rectiﬁable contour C and any function f which is differentiable at each point inside C
and whose partial derivatives satisfy (1.2). The integral formula (1.5) has a similarly
wide range of validity. From this it follows, as in complex analysis, that if f is regular
in an open set U then it has a power series expansion about each point of U. Thus
pointwise differentiability, together with the four real conditions (1.2) on the sixteen
partial derivatives of f, is sufﬁcient to ensure analyticity.
The homogeneous components in the power series representing a regular function
are themselves regular; thus it is important to study regular homogeneous polynomials,
the basic regular functions from which all regular functions are constructed. The cor-
responding functions of a complex variable are just the powers of the variable, but the
situation with quaternions is more complicated. The set of homogeneous regular func-
tions of degree n forms a quaternionic vector space of dimension
1
2
(n+1)(n+2); this
is true for any integer n if for negative n it is understood that the functions are deﬁned
and regular everywhere except at 0. The functions with negative degree of homogene-
ity correspond to negative powers of a complex variable, and occur in the quaternionic
Laurent series which exists for any regular function which is regular in an open set
except at one point. Fueter found two natural bases for the set of homogeneous func-
tions, which play dual roles in the calculus of residues. (He actually only proved that
these bases formed spanning sets). In this paper we will study homogeneous regular
functions by means of harmonic analysis on the unit sphere in H, which forms a group
isomorphic to SU(2); this bears the same relation to quaternionic analysis as the theory
of Fourier series does to complex analysis.
Many of the algebraic and geometric properties of complex analytic functions are
not present in quaternionic analysis. Because quaternions do not commute, regular
functions of a quaternion variable cannot be multiplied or composed to give further reg-
ular functions. Because the quaternions are four-dimensional, there is no counterpart
to the geometrical description of complex analytic functions as conformal mappings.
The zeros of a quaternionic regular function are not necessarily isolated, and its range
is not necessarily open; neither of these sets need even be a submanifold of H. There is
a corresponding complexity in the structure of the singularities of a quaternionic regu-
lar function; this was described by Fueter [9], but without giving precise statements or
proofs. This topic is not investigated here.
The organisation of this paper is as follows.
In section 2 the basic algebraic facts about quaternions are surveyed and notation
is established; some special algebraic concepts are introduced, and quaternionic differ-
ential forms are described.
Section 3 is concerned with the deﬁnition of a regular function. The remarks in the
second paragraph of this introduction, about possible analogues of complex deﬁnitions
of analyticity, are ampliﬁed (this material seems to be widely known, but is not easily
accessible in the literature), and the deﬁnition (1.7) of regular functions by means of the
quaternionic derivative is shown to be equivalent to Fueter’s deﬁnition (1.2) by means
of a Cauchy-Riemann-type equation.
Section 4 is devoted to the analogues of the Cauchy-Goursat theorem and Cauchy’s
integral formula.
4
Section 5 contains analogues of Liouvilles’s theorem, the maximum-modulus prin-
ciple, and Morera’s theorem. After the work of section 4, only the last of these requires
proof.
In section 6 we show how regular functions can be constructed from functions
of more familiar type, namely harmonic functions of four real variables and analytic
functions of a complex variable.
Section 7 is concerned with the effect on regular functions of conformal transfor-
mation of the variable; these results are needed in the last two sections.
Section 8 is an investigation of homogeneous regular functions by means of har-
monic analysis on S
3
.
In section 9 we examine the power series representing a regular function and prove
analogues of Laurent’s Theorem and the residue theorem.
2 Preliminaries
2.1 The Algebra of Quaternions
1
The quaternions H form a four-dimensional algebra over the real ﬁeld R, with an iden-
tity element denoted by 1. We regard R as being embedded in H by identifying t ∈ R
with t1 ∈ H. Then we can write H = R⊕P, where P is an oriented three-dimensional
Euclidean vector space, and the product of two quaternions is deﬁned by
(a
0
, a)(b
0
, b) = (a
0
b
0
−a.b, a
0
b + b
0
a +a b) (2.1)
where a
0
, b
0
∈ R, a, b ∈ P, a.b denotes the inner product of a and b, and a b de-
notes the vector product determined by the orientation on P. (Conversely, the subspace
P, its inner product and its orientation can be deﬁned in terms of the multiplication on
H.)
Thus we can choose a basis ¦1, i, j, k¦ for H so that the multiplication is given by
i
2
= j
2
= k
2
= −1,
ij = −ji = k, jk = −kj = i, ki = −ik = j
¸
(2.2)
The typical quaternion will be denoted by
q = t + xi + yj + zk (t, x, y, z ∈ R) (2.3)
In performing calculations it is sometimes useful to denote the basic quaternions i, j,
k by e
i
(i = 1, 2, 3) and the coordinates x, y, z by x
i
(i = 1, 2, 3), and to use the
summation convention for repeated indices. In this notation eq. (2.3) becomes
q = t + x
i
e
i
(2.4)
and the multiplication rules (2.2) become
e
i
e
j
= −δ
ij
+
ijk
e
k
(2.5)
1
Proofs of the assertions in this section can be found in [10, chap. 10].
5
where
ijk
is the usual alternating symbol.
The centre of the algebra H is the real subﬁeld R. If q is any quaternion, the vector
subspace spanned by 1 and q is a subﬁeld of H, which is isomorphic to the complex
numbers Cif 1 and q are linearly independent. It is sometimes convenient to distinguish
a particular embedding of C in H; whenever we want to do this, we will take C to be
the subﬁeld spanned by 1 and i. Then any quaternion can be written as
q = v + jw (2.6)
with v, w ∈ C; if q is given by (2.3), then
v = t + ix and w = y −iz (2.7)
From an equation of the form v
1
+ jw
1
= v
2
+ jw
2
with v
1
, v
2
, w
1
, w
2
∈ C, we can
deduce that v
1
= v
2
and w
1
= w
2
. When quaternions are written in this form, the
basic law of multiplication is
vj = j¯ v. (2.8)
The conjugate of the quaternion q is
¯ q = t −ix −jy −kz. (2.9)
Conjugation is an involutive anti-automorphism of H, i.e. it is R-linear, q = q, and
q
1
q
2
= ¯ q
2
¯ q
1
. (2.10)
For every embedding of C in H, quaternion conjugation coincides with complex con-
jugation. q commutes with ¯ q, and their product is
q¯ q = t
2
+ x
2
+ y
2
+ z
2
. (2.11)
The modulus of q is the non-negative real number
[q[ =

q¯ q (2.12)
It follows from (2.11) that every non-zero quaternion has a multiplicative inverse
q
−1
=
¯ q
[q[
2
. (2.13)
The real part of q is
Re q = t =
1
2
(q + ¯ q) ∈ R (2.14)
and its pure quaternion part (or vector part) is
Puq = xi + yj + zk =
1
2
(q − ¯ q) ∈ P. (2.15)
It follows from (2.1) that
Re(q
1
q
2
) = Re(q
2
q
1
) (2.16)
and from (2.9) that
Re ¯ q = Re q. (2.17)
6
q is a unit quaternion if [q[ = 1. The set of unit quaternions, constituting the sphere
S
3
, forms a multiplicative Lie group isomorphic to SU(2); we will denote it by S. The
versor of a quaternion q is the unit quaternion
Un q =
q
[q[
(2.18)
Any quaternion has a polar decomposition q = ru where r = [q[ ∈ R and u = Un q ∈
S.
The classical notation for these functions of q is S q = Re q, Vq = Puq, Kq = ¯ q,
Tq = [q[, Uq = Un q.
The positive-deﬁnite quadratic form (2.11) gives rise to an inner product
'q
1
, q
2
` = Re(q
1
¯ q
2
)
= Re(¯ q
1
q
2
) by (2.16) and (2.17)
(2.19)
Note that
'aq
1
, q
2
` = 'q
1
, ¯ aq
2
` (2.20)
and
'q
1
a, q
2
` = 'q
1
, q
2
¯ a`, (2.21)
i.e. the adjoint of left (right) multiplication by a is left (right) multiplication by ¯ a, for
any a ∈ H. If u
1
and u
2
are unit quaternions, the map q → u
1
qu
2
is orthogonal with
respect to this inner product and has determinant 1; conversely, any rotation of H is
of the form q → u
1
qu
2
for some u
1
, u
2
∈ S. This is the well-known double cover
0 →Z
2
→SU(2) SU(2) →SO(4) →0.
This inner product induces an R-linear map Γ : H

→H, where H

= Hom
R
(H, R)
is the dual of H, given by
'Γ(α), q` = α(q) (2.22)
for a ∈ H

, q ∈ H. Since ¦1, i, j, k¦ is an orthonormal basis for H, we have
Γ(α) = α(1) + i α(i) + j α(j) + k α(k). (2.23)
The set of R-linear maps of H into Itself forms a two-sided vector space over H of
dimension 4, which we will denote by F
1
. It is spanned (over H) by H

, so the map
Γ can be extended by linearity to a right H-linear map Γ
r
: F
1
→ H and a left-linear
map Γ

: F
1
→H. They are given by
Γ
r
(α) = α(1) + i α(i) + j α(j) + k α(k) (2.24)
and
Γ

(α) = α(1) + α(i) i + α(j) j + α(k) k (2.25)
for any α ∈ F
1
.
The inner product (2.19) and the maps Γ
r
, Γ

: F
1
= Hom
R
(H, R) → H have
alternative characterisations in terms of quaternion multiplication and conjugation; in
fact they are obtained by slight modiﬁcations of standard procedures from the tensor of
type (1,2) which deﬁnes the multiplication on H. Let L
q
: H → H and R
q
: H → H
be the operations of multiplication by q on the left and on the right, respectively, and
let K : H →H be the operation of conjugation; then
7
Theorem 1
(i) 'q, p` = −
1
2
Tr(L
q
R
p
K)
(ii) 'Γ
r
(α), q` = Tr(R
¯ q
αK)
(iii) 'Γ

(α), q` = Tr(L
¯ q
αK) for all p, q ∈ H and α ∈ F
1
.
Proof The trace of an R-linear map α : H →H is
Tr α = Re [α(1)] −
¸
i
Re [e
i
α(e
i
)] (2.26)
Hence (i)
Tr(L
q
R
p
K) = Re
¸
qp +
¸
i
e
i
qe
i
p
¸
But
q +
¸
i
e
i
qe
i
= −2¯ q (2.27)
as can easily be veriﬁed, so
Tr(L
q
R
p
K) = −2 Re(¯ qp) = −2'q, p`. 2
(ii)
Tr(R
¯ q
αK) = Re
¸
α(1)¯ q +
¸
i
e
i
α(e
i
)¯ q
¸
= Re [Γ
r
(α)¯ q] by (2.24)
= 'Γ
r
(α), q`
2
(iii)
Tr(L
¯ q
αK) = Re
¸
¯ qα(1) +
¸
i
e
i
¯ qα(e
i
)
¸
= Re
¸
¯ qα(1) +
¸
i
¯ qα(e
i
)e
i
¸
by (2.16)
= Re [¯ qΓ

(α)] by (2.25)
= 'Γ

(α), q` 2
We will also need two other maps
¯
Γ
r
,
¯
Γ

: F
1
→H, deﬁned as follows:
'
¯
Γ
r
(α), q` = Tr(R
¯ q
α);
¯
Γ
r
(α) = α(1) −
¸
i
e
i
α(e
i
) (2.28)
'
¯
Γ

(α), q` = Tr(L
q
α);
¯
Γ

(α) = α(1) −
¸
i
α(e
i
)e
i
(2.29)
8
2.2 Quaternionic Differential Forms
When it is necessary to avoid confusion with other senses of differentiability which
we will consider, we will say that a function f : H → H is real-differentiable if it is
differentiable in the usual sense. Its differential at a point q ∈ H is then an R-linear
map df
q
: H → H. By identifying the tangent space at each point of H with H itself,
we can regard the differential as a quaternion-valued 1-form
df =
∂f
∂t
dt +
∂f
∂x
dx +
∂f
∂y
dy +
∂f
∂z
dz. (2.30)
Conversely, any quaternion-valued 1-form θ = a
0
dt + a
i
dx
i
(a
0
, a
i
∈ H) can be
regarded as the R-linear map θ : H →H given by
θ(t + x
i
e
i
) = a
0
t + a
i
x
i
(2.31)
Similarly, a quaternion-valued r-form can be regarded as a mapping from H to the
space of alternating R-multilinear maps fromH... H(r times) to H. We deﬁne the
exterior product of such forms in the usual way: if θ is an r-form and φ is an s-form,
θ ∧ φ(h
1
, ..., h
r+s
) =
1
r!s!
¸
ρ
(ρ)θ(h
ρ(1)
, ..., h
ρ(r)
)φ(h
ρ(r+1)
, ..., h
ρ(r+s)
), (2.32)
where the sum is over all permutations ρ of r + s objects, and (ρ) is the sign of ρ.
Then the set of all r-forms is a two-sided quaternionic vector space, and we have
a(θ ∧ φ) = (aθ) ∧ φ,
(θ ∧ φ)a = θ ∧ (φa),
(θa) ∧ φ = θ ∧ (aφ)

(2.33)
for all quaternions a, r-forms θ and s-forms φ. The space of quaternionic r-forms has a
basis of real r-forms, consisting of exterior products of the real 1-forms dt, dx, dy, dz;
for such forms left and right multiplication by quaternions coincide. Note that because
the exterior product is deﬁned in terms of quaternion multiplication, which is not com-
mutative, it is not in general true that θ ∧ φ = −φ ∧ θ for quaternionic 1-forms θ and
φ.
The exterior derivative of a quaternionic differential form is deﬁned by the usual
recursive formulae, and Stokes’s theorem holds in the usual form for quaternionic in-
tegrals.
The following special differential forms will be much used in the rest of the paper.
The differential of the identity function is
dq = dt + i dx + j dy + k dz; (2.34)
regarded as R-linear transformation of H, dq is the identity mapping. Its exterior prod-
uct with itself is
dq ∧ dq =
1
2

ijk
e
i
dx
j
∧ dx
k
= i dy ∧ dz + j dz ∧ dx + k dx ∧ dy (2.35)
9
which, as antisymmetric function on H H, gives the commutator of its arguments.
For the (essentially unique) constant real 4-form we use the abbreviation
v = dt ∧ dx ∧ dy ∧ dz, (2.36)
so that v(1, i, j, k) = 1. Finally, the 3-form Dq is deﬁned as an alternating R-trilinear
function by
'h
1
, Dq(h
2
, h
3
, h
4
)` = v(h
1
, h
2
, h
3
, h
4
) (2.37)
for all h
1
, ..., h
4
∈ H. Thus Dq(i, j, k) = 1 and Dq(1, e
i
, e
j
) = −
ijk
e
k
. The
coordinate expression for Dq is
Dq = dx ∧ dy ∧ dz −
1
2

ijk
e
i
dt ∧ dx
j
∧ dx
k
= dx ∧ dy ∧ dz −i dt ∧ dy ∧ dz −j dt ∧ dz ∧ dx −k dt ∧ dx ∧ dy.
(2.38)
Geometrically, Dq(a, b, c) is a quaternion which is perpendicular to a, b and c and has
magnitude equal to the volume of the 3-dimensional parallelepiped whose edges are a,
b and c. It also has the following algebraic expression:
Theorem 2 Dq(a, b, c) =
1
2
(c¯ ab −b¯ ac)
Proof For any unit quaternion u, the map q → uq is an orthogonal transformation of
H with determinant 1; hence
Dq(ua, ub, uc) = u Dq(a, b, c).
Taking u = [a[
−1
, and using the R-trilinearity of Dq, we obtain
Dq(a, b, c) = [a[
2
a Dq(1, a
−1
b, a
−1
c). (2.39)
Now since Dq(1, e
i
, e
j
) = −
ijk
e
k
=
1
2
(e
j
e
i
−e
i
e
j
), we have by linearity
Dq(1, h
1
, h
2
) =
1
2
(h
2
h
1
−h
1
h
2
) (2.40)
for all h
1
, h
2
∈ H. Hence
Dq(a, b, c) =
1
2
[a[
2
a(a
−1
ca
−1
b −a
−1
ba
−1
c)
=
1
2
(c¯ ab −b¯ ac) by (2.13). 2
Two useful formulae were obtained in the course of this proof. The argument lead-
ing to (2.39) can be generalised, using the fact that the map q → uqv is a rotation for
any pair of unit quaternions u, v, to
Dq(ah
1
b, ah
2
b, ah
3
b) = [a[
2
[b[
2
a Dq(h
1
, h
2
, h
3
)b; (2.41)
and the formula (2.40) can be written as
1|Dq = −
1
2
dq ∧ dq, (2.42)
10
where | denotes the usual inner product between differential forms and vector ﬁelds
and 1 denotes the constant vector ﬁeld whose value is 1.
Since the differential of a quaternion-valued function on H is an element of F
1
, the
map Γ
r
can be applied to it. The result is
Γ
r
(df) =
∂f
∂t
+ i
∂f
∂x
+ j
∂f
∂y
+ k
∂f
∂z
. (2.43)
We introduce the following notation for the differential operator occurring in (2.40),
and for other related differential operators:
¯

f =
1
2
Γ
r
(df) =
1
2

∂f
∂t
+ e
i
∂f
∂x
i

,

f =
1
2
¯
Γ
r
(df) =
1
2

∂f
∂t
−e
i
∂f
∂x
i

,
¯

r
f =
1
2
Γ

(df) =
1
2

∂f
∂t
+
∂f
∂x
i
e
i

,

r
f =
1
2
¯
Γ

(df) =
1
2

∂f
∂t

∂f
∂x
i
e
i

,
∆f =

2
f
∂t
2
+

2
f
∂x
2
+

2
f
∂y
2
+

2
f
∂z
2
.

(2.44)
Note that ∂

,
¯

, ∂
r
and
¯

r
all commute, and that
∆ = 4∂
r
¯

r
= 4∂

¯

(2.45)
3 Regular Functions
We start by showing that the concept of an analytic function of a quaternion variable as
one which is constructed from the variable by quaternion addition and multiplication
(possibly involving inﬁnite series) is the same as the concept of a real-analytic function
in the four real variables t, x, y, z.
Deﬁnition 1 A quaternionic monomial is a function f : H →H of the form
f(q) = a
0
qa
1
q...q
r−1
qa
r
(3.1)
for some non-negative integer r (the degree of the monomial) and constant quaternions
a
0
, ..., a
r
.
A quaternionic polynomial is a ﬁnite sum of quaternionic monomials.
A homogeneous polynomial function of degree r on H is a function f : H → H
of the form
f(q) = F(q, ..., q)
where F : H... H (r times) →H is R-multilinear.
A polynomial function on H is a ﬁnite sum of homogeneous polynomial functions
of varying degrees.
11
Theorem 3 Every polynomial function on H is a quaternionic polynomial.
Proof Any polynomial function f can be written as
f(q) = f
0
(t, x, y, z) +
¸
i
f
i
(t, x, y, z)e
i
where f
0
and f
i
are four real-valued polynomials in the four real variables t, x, y, z.
But
t =
1
4
(q −iqi −jqj −kqk),
x =
1
4i
(q −iqi + jqj + kqk),
y =
1
4j
(q + iqi −jqj + kqk),
z =
1
4k
(q + iqi + jqj −kqk).

(3.2)
Putting these expressions for t, x, y, z into the polynomials f
0
, f
i
, we obtain f(q) as a
sum of expressions in q of the form (3.1), so f is a quaternionic polynomial. 2
It is clear that conversely, every quaternionic polynomial is a polynomial function,
so we have
Corollary The class of functions which are deﬁned in a neighbourhood of the origin
in H and can be represented there by a quaternionic power series, i.e. a series of
quaternionic monomials, is precisely the same as the class of functions which are real-
analytic in a neighbourhood of the origin.
We now turn to an alternative attempt to parallel complex analysis, in which we
concentrate on the existence of a quaternionic derivative deﬁned as the limit of a differ-
ence quotient. We will show that only quaternionic polynomials of degree 1 (and not
all of them) possess such a derivative.
Deﬁnition 2 A function f : H → H is quaternion-differentiable on the left at q if
the limit
df
dq
= lim
h→0

h
−1
¦f(q + h) −f(q)¦

exists.
Theorem 4 Suppose the function f is deﬁned and quaternion-differentiable on the left
throughout a connected open set U. Then on U, f has the form
f(q) = a + qb
for some a, b ∈ H.
Proof From the deﬁnition it follows that if f is quaternion-differentiable on the left at
q, it is real-differentiable at q and its differential is the linear map of multiplication on
the right by
df
dq
:
df
q
(h) = h
df
dq
,
i.e.
df
q
= dq
df
dq
.
12
Equating coefﬁcients of dt, dx, dy and dz gives
df
dq
=
∂f
∂t
= −i
∂f
∂x
= −j
∂f
∂y
= −k
∂f
∂z
. (3.3)
Put q = v +jw, where v = t +ix and w = y −iz, and let f(q) = g(v, w) +jh(v, w),
where g and h are complex-valued functions of the two complex variables v and w;
then (3.3) can be separated into the two sets of complex equations
∂g
∂t
= −i
∂g
∂x
=
∂h
∂y
= i
∂h
∂z
,
∂h
∂t
= i
∂h
∂x
= −
∂g
∂y
= i
∂g
∂z
.
In terms of complex derivatives, these can be written as
∂g
∂¯ v
=
∂h
∂ ¯ w
=
∂h
∂v
=
∂g
∂w
= 0, (3.4)
∂g
∂v
=
∂h
∂w
, (3.5)
and
∂h
∂¯ v
= −
∂g
∂ ¯ w
. (3.6)
Eq. (3.4) shows that g is a complex-analytic function of v and ¯ w, and h is a complex-
analytic function of ¯ v and w. Hence by Hartogs’s theorem [11, p. 133] g and h have
continuous partial derivatives of all orders and so from (3.5)

2
g
∂v
2
=

∂v

∂h
∂w

=

∂w

∂h
∂v

= 0.
Suppose for the moment that U is convex. Then we can deduce that g is linear in ¯ w, h
is linear in w and h is linear in ¯ v. Thus
g(v, w) = α + βv + γ ¯ w + δv ¯ w,
h(v, w) = + ζ¯ v + ηw + θ¯ vw,
where the Greek letters represent complex constants. Now (3.5) and (3.6) give the
following relations among these constants:
β = η, ζ = −γ, δ = θ = 0.
Thus
f = g + jh = α + j + (v + jw)(β −jγ)
= a + qb,
where a = α+j and b = β−jγ; so f is of the stated formif U is convex. The general
connected open set can be covered by convex sets, any two of which can be connected
by a chain of convex sets which overlap in pairs; comparing the forms of the function
13
f on the overlaps, we see that f(q) = a + qb with the same constants a, b throughout
U. 2
Even if f is quaternion-differentiable, it will not in general satisfy Cauchy’s theo-
rem in the form

dqf = 0 (3.7)
where the integral is round a closed curve; in fact the only functions satisfying this
equation for all closed curves are the constant functions. We will prove this for the
inﬁnitesimal form of (3.7), namely
d(dqf) = 0. (3.8)
Theorem 5 If the function f : U → H is real-differentiable in the connected open set
U and satisﬁes d(dqf) = 0 in U, then f is constant on U.
Proof
d(dqf) = dq ∧ df = (dt + e
i
dx
i
) ∧

∂f
∂t
dt +
∂f
∂x
j
dx

=

∂f
∂x
i
−e
i
∂f
∂t

dt ∧ dx
i
+ e
i
∂f
∂x
j
dx
i
∧ dx
j
Thus
d(dqf) = 0 ⇒
∂f
∂x
i
= e
i
∂f
∂t
(3.9)
and
e
i
∂f
∂x
j
= e
j
∂f
∂x
i
(3.10)
From (3.9) we have, for example,
∂f
∂x
i =
∂f
∂y
j
while from (3.10),
∂f
∂x
j =
∂f
∂y
i.
Hence
∂f
∂x
= −
∂f
∂y
ji =
∂f
∂y
ij = −
∂f
∂x
.
So
∂f
∂x
= 0 and therefore
∂f
∂t
=
∂f
∂y
=
∂f
∂z
= 0 in U. It follows that f is constant in
U. 2
We will now give a deﬁnition of “regular” for a quaternionic function which is
satisﬁed by a large class of functions and opens the door to a development similar to
the theory of regular functions of a complex variable.
14
Deﬁnition 3 A function f : H → H is left-regular at q ∈ H if it is real-differentiable
at q and there exists a quaternion f

(q) such that
d(dq ∧ dq f) = −2Dq f

(q). (3.11)
It is right-regular if there exists a quaternion f

r
(q) such that
d(f dq ∧ dq) = −2f

r
(q) Dq.
Clearly, the theory of left-regular functions will be entirely equivalent to the theory
of right-regular functions. For deﬁniteness, we will only consider left-regular func-
tions, which we will call simply regular. We will write f

(q) = f

(q) and call it the
derivative of f at q.
Theorem 6 (the Cauchy-Riemann-Fueter equations)
A real-differentiable function f is regular at q if and only if
Γ
r
(df
q
) = 0. (3.12)
Proof Suppose f is regular at q. Then from (3.11),
dq ∧ dq ∧ df
q
= −2Dqf

(q)
Evaluating these trilinear functions with i, j, k as arguments gives
(ij −ji) df
q
(k) + (jk −kj) df
q
(i) + (ki −ik) df
q
(j) = −2f

(q)
while using 1, i, j as arguments gives
(ij −ji) df
q
(1) = 2kf

(q).
Hence
f

(q) = df
q
(1) = −¦i df
q
(i) + j df
q
(j) + k df
q
(k)¦ (3.13)
Comparing with (2.24), we see that Γ
r
(df
q
) = 0.
Conversely, if Γ
r
(df
q
) = 0 we can deﬁne f

(q) = df
q
(1) and then evaluating as
above shows that dq ∧ dq ∧ df
q
= 2Dqf

(q), so that f is regular at q. 2
Note that (3.12) and (3.13) can be written as
2
¯

f =
∂f
∂t
+ i
∂f
∂x
+ j
∂f
∂y
+ k
∂f
∂z
= 0 (3.14)
and
f

(q) =
∂f
∂t
. (3.15)
Hence f

= ∂

f. If we write q = v + jw, f(q) = g(v, w) +jh(v, w) as in theorem 4,
(3.14) becomes the pair of complex equations
∂g
∂¯ v
=
∂h
∂ ¯ w
,
∂g
∂w
= −
∂h
∂v
(3.16)
15
which show some similarity to the Cauchy-Riemann equations for a function of a com-
plex variable.
From (3.14) and (2.45) it follows that if f is regular and twice differentiable, then
∆f = 0,
i.e. f is harmonic. We will see in the next section that a regular function is necessarily
inﬁnitely differentiable, so all regular functions are harmonic.
The derivative of a regular function can be characterised as the limit of a difference
quotient which is analogous to that used to deﬁne the derivative of a complex-analytic
function.
Deﬁnition 4 An oriented k-parallelepiped in H is a map C : I
k
→ H, where I
k

R
k
is the closed unit k-cube, of the form
C(t
1
, ..., t
k
) = q
0
+ t
1
h
1
+ ... + t
k
h
k
.
q
0
∈ H is called the original vertex of the parallelepiped, and h
1
, ..., h
k
∈ H are
called its edge-vectors. A k-parallelepiped is non-degenerate if its edge-vectors are
linearly independent (over R). A non-degenerate 4-parallelepiped is positively ori-
ented if v(h
1
, ..., h
4
) > 0, negatively oriented if v(h
1
, ..., h
4
) < 0.
Theorem 7 Suppose that f is regular at q
0
and continuously differentiable in a neigh-
bourhood of q
0
. Then given > 0, there exists δ > 0 such that if C is a non-degenerate
oriented 3-parallelepiped with q
0
∈ C(I
3
) and q ∈ C(I
3
) ⇒[q −q
0
[ < δ, then

C
Dq

−1

∂C
dq ∧ dqf

+ 2f

(q
0
)

< .
Proof First note that [[h
1
∧ h
2
∧ h
3
[[ = [Dq(h
1
, h
2
, h
3
)[ is a norm on the real vector
space H∧H∧H. Now since df
q
is a continuous function of q at q
0
, so is dq ∧dq ∧df
q
;
hence we can choose δ so that
[q −q
0
[ < δ ⇒[dq ∧ dq ∧ df
q
(h
1
, h
2
, h
3
) −dq ∧ dq ∧ df
q0
(h
1
, h
2
, h
3
)[
< [Dq(h
1
, h
2
, h
3
)[, ∀h
1
, h
2
, h
3
∈ H.
Let C be as in the statement of the theorem, with edge-vectors h
1
, h
2
, h
3
. Then by
Stokes’s theorem,

∂C
dq ∧ dqf =

C
d(dq ∧ dqf) =

1
0
dq ∧ dq ∧ df
C(t)
(h
1
, h
2
, h
3
) dt
1
dt
2
dt
3
Since f is regular at q
0
,
2

C
Dqf

(q
0
) = 2

1
0
Dq(h
1
, h
2
, h
3
) f

(q
0
) dt
1
dt
2
dt
3
= −

1
0
dq ∧ dq ∧ df
q0
(h
1
, h
2
, h
3
) dt
1
dt
2
dt
3
16
Thus

∂C
dq ∧ dqf + 2

C
Dqf

(q
0
)

1
0
[dq ∧ dq ∧ df
C(t)
(h
1
, h
2
, h
3
)
−dq ∧ dq ∧ df
q0
(h
1
, h
2
, h
3
)[ dt
1
dt
2
dt
3
< [Dq(h
1
, h
2
, h
3
)[ =

C
Dq

. 2
The corresponding characterisation of the derivative in terms of the values of the
function at a ﬁnite number of points is
f

(q
0
) = − lim
h1,h2,h3→0
[Dq(h
1
, h
2
, h
3
)
−1
¦(h
1
h
2
−h
2
h
1
)(f(q
0
+ h
3
) −f(q
0
))
+(h
2
h
3
−h
3
h
2
)(f(q
0
+ h
1
) −f(q
0
))
+(h
3
h
1
−h
1
h
3
)(f(q
0
+ h
2
) −f(q
0
))¦]
(3.17)
This is valid if it is understood that h
1
, h
2
, h
3
are multiples of three ﬁxed linearly
independent quaternions, h
i
= t
i
H
i
, and the limit is taken as t
1
, t
2
, t
3
→ 0. The limit
(3.17) is similar to that used by Joly [3, art. 54] to deﬁne ∇f for a function satisfying
∂f
∂t
= 0, which would be obtained as the Cauchy-Riemann equations if dx ∧ dy ∧ dz
were substituted for Dq in the deﬁnition of regularity.
To make explicit the analogy between the deﬁnition of regular quaternionic func-
tions and the deﬁnition of regular complex functions by means of the Cauchy-Riemann
equations, note that the analogues of the 2-form
1
2
dq ∧ dq, the 3-form −Dq and the
equation d(dq ∧ dqf) = −2Dqf

(q) are the 0-form 1, the 1-form dz and the equation
df = dzf

(z). The analogy between theorem 7 and the difference-quotient deﬁnition
of regular complex functions can be made explicit by stating the latter as follows: If
f is regular at z
0
, then given > 0 there exists δ > 0 such that if L is a directed
line segment (i.e. an oriented parallelepiped of codimension 1 in C) with z
0
∈ L and
z ∈ L ⇒[z −z
0
[ < δ, then

L
dz

−1

∂L
f

−f

(z
0
)

< .
4 Cauchy’s Theorem and the Integral Formula
The integral theorems for regular quaternionic functions have as wide a range of valid-
ity as those for regular complex functions, which is considerably wider than that of the
integral theorems for harmonic functions. Cauchy’s theorem holds for any rectiﬁable
contour of integration; the integral formula, which is similar to Poisson’s formula in
that it gives the values of a function in the interior of a region in terms of its values on
the boundary, holds for a general rectiﬁable boundary, and thus constitutes an explicit
solution to the general Dirichlet problem.
Our route to the general form of Cauchy’s theorem will be to use Goursat’s method
to prove the theorem for a parallelepiped, and immediately obtain the integral for-
mula for a parallelepiped; then we can deduce that an analytic function is continuously
17
differentiable, and use Stokes’s theorem to extend Cauchy’s theorem to differentiable
contours. The extension to rectiﬁable contours also follows from an appropriate form
of Stokes’s theorem.
The heart of the quaternionic Cauchy’s theorem is the following fact.
Theorem 8 A differentiable function f is regular at q if and only if
Dq ∧ df
q
= 0.
Proof
Dq ∧ df
q
(i, j, k, l) = Dq(i, j, k)df
q
(l) −Dq(j, k, l)df
q
(i)
+ Dq(k, l, i)df
q
(j) −Dq(l, i, j)df
q
(k)
= df
q
(l) + idf
q
(i) + jdf
q
(j) + kdf
q
(k)
= Γ
r
(df
q
)
which vanishes if and only if f is regular at q, by theorem 6. 2
Theorem 9 (Cauchy’s theorem for a parallelepiped)
If f is regular at every point of the 4-parallelepiped C,

∂C
Dq f = 0.
Proof [8] Let q
0
and h
1
, ..., h
4
be the original vertex and edge-vectors of C. For each
subset S of ¦1, 2, 3, 4¦ let C
S
be the 4-parallelepiped with edge-vectors
1
2
h
1
, ...,
1
2
h
4
and original vertex q
0
+
¸
i∈S
1
2
h
i
; then

∂C
Dqf =
¸
S

∂CS
Dqf
Hence there is a C
S
—call it C
1
—such that

∂C1
Dqf

1
16

∂C
Dqf

.
Now perform a similar dissection of C
1
. Continuing in this way, we obtain a se-
quence of 4-parallelepipeds C
n
, with original vertices q
n
, such that C
n
has edge-
vectors 2
−n
h
1
, ..., 2
−n
h
4
, C ⊃ C
1
⊃ C
2
⊃ ..., and

∂Cn
Dqf

1
16
n

∂C
Dqf

, (4.1)
Clearly there is a point q

∈ ∩C
n
, and q
n
→q

as n →∞.
Since f is real-differentiable at q

, we can write
f(q) = f(q

) + α(q −q

) + (q −q

)r(q)
where α = df
q∞
∈ F
1
, and r(q) → 0 as q → q

. Then if we deﬁne r(q

) = 0, r
is a continuous function and so [r(q)[ has a maximum value ρ
n
on ∂C
n
. Since the C
n
converge on q

, ρ
n
→0 as n →∞.
18
Now

∂Cn
Dqf(q) =

Cn
d(Dq)f(q

) = 0
and

∂Cn
Dqα(q −q

) =

Cn
d(Dqα) = 16
−n
(Dq ∧ α)(h
1
, ..., h
4
) = 0
by theorem 8, since f is regular at q

. Thus

∂Cn
Dqf(q) =

∂Cn
Dq(q −q

)r(q).
Let F : I
3
→ H be one of the 3-parallelepipeds forming the faces of C
n
. Then
F ⊂ ∂C
n
, and the edge-vectors of F are three of the four edge-vectors of C
n
, say
2
−n
h
a
, 2
−n
h
b
, and 2
−n
h
c
. For q ∈ F(I
3
) we have [r(q)[ < ρ
n
and [q − q

[ ≤
2
−n
([h
1
[ + ... +[h
4
[); hence

F
Dq(q −q

)r(q)

≤ 8
−n
[Dq(h
a
, h
b
, h
c
)[ 2
−n
([h
1
[ + ... +[h
4
[) ρ
n
.
Let V be the largest [Dq(h
a
, h
b
, h
c
)[ for all choices of a, b, c; then since the integral
over ∂C
n
is the sum of 8 integrals over faces F,

∂Cn
Dqf(q)

≤ 8.16
−n
V ([h
1
[ + ... + [h
4
[) ρ
n
.
Combining this with (4.1), we ﬁnd

∂C
Dqf

≤ 8V ([h
1
[ + ... +[h
4
[)ρ
n
.
Since ρ
n
→0 as n →∞, it follows that

∂C
Dqf = 0. 2
Theorem 10 (the Cauchy-Fueter integral formula for a parallelepiped)
If f is regular at every point of the positively oriented 4-parallelepiped C, and q
0
is a
point in the interior of C,
f(q
0
) =
1

2

∂C
(q −q
0
)
−1
[q −q
0
[
2
Dq f(q).
Proof [7] The argument of theorem 8 shows that
Dq ∧ df
q
= −
¯

f(q)v (4.2)
where v = dt ∧ dx ∧ dy ∧ dz, for any differentiable function f. A similar calculation
shows that
df
q
∧ Dq =
¯

r
f(q)v. (4.3)
19
Hence if f and g are both differentiable,
d(gDqf) = d(gDq)f + gd(Dqf)
= dg ∧ Dqf −gDq ∧ df
=
¸
(
¯

r
g)f + g(
¯

f)
¸
v
(4.4)
Take g(q) =
(q−q0)
−1
|q−q0|
2
=
¯ q−¯ q0
|q−q0|
4
= ∂
r

1
|q−q0|
2

; then g is differentiable except at q
0
,
and
¯

r
g = ∆

1
[q −q
0
[
2

= 0.
If f is regular we have
¯

f = 0, and so
d
¸
(q −q
0
)
−1
[q −q
0
[
2
Dqf

= 0.
We can now follow the argument of theorem 9 to show that

∂C

(q −q
0
)
−1
[q −q
0
[
2
Dqf(q) = 0
where C

is any 4-parallelepipednot containing q
0
. By dissecting the given 4-parallelepiped
C into 81 4-parallelepipeds with edges parallel to those of C, we deduce that

∂C
(q −q
0
)
−1
[q −q
0
[
2
Dqf(q) =

∂C0
(q −q
0
)
−1
[q −q
0
[
2
Dqf(q),
where C
0
is any 4-parallelepiped containing q
0
which lies in the interior of C and has
edges parallel to those of C. Take C
0
to have edge-vectors δh
1
, ..., δh
4
, where δ is a
positive real number and h
1
, ..., h
4
are the edge-vectors of C, and suppose q
0
is at the
centre of C
0
(so that the original vertex of C
0
is q
0

1
2
δh
1
−... −
1
2
δh
4
); then
min
q∈∂C0
[q −q
0
[ = min
1≤a,b,c≤4

v(δh
1
, ..., δh
4
)
Dq(
1
2
δh
a
,
1
2
δh
b
,
1
2
δh
c
)

= Wδ
where W depends only on h
1
, ..., h
4
. Since f is continuous at q
0
, we can choose δ so
that q ∈ C
0
(I
4
) ⇒[f(q) −f(q
0
)[ < for any given > 0; hence

∂C0
(q −q
0
)
−1
[q −q
0
[
2
Dq¦f(q) −f(q
0

8V
W
3
(4.5)
where, as in theorem 9,
V = max
1≤a,b,c≤4
[Dq(h
a
, h
b
, h
c
)[.
Since the 3-form
(q−q0)
−1
Dq
|q−q0|
2
is closed and continuously differentiable in H`¦q
0
¦,
Stokes’s theorem gives

∂C0
(q −q
0
)
−1
Dq
[q −q
0
[
2
=

S
(q −q
0
)
−1
Dq
[q −q
0
[
2
20
where S is the 3-sphere [q − q
0
[ = 1, oriented so that Dq is in the direction of the
outward normal to S. Working in spherical coordinates (r, θ, φ, ψ), in which
q −q
0
= r

cos θ + i sinθ cos φ + j sin θ sin φe
−iψ

,
we ﬁnd that on S, i.e. when r = 1,
Dq = (q −q
0
) sin
2
θ sin φdθ ∧ dφ ∧ dψ
= (q −q
0
)dS
(4.6)
where dS is the usual Euclidean volume element on a 3-sphere. Hence

∂C0
(q −q
0
)
−1
Dq
[q −q
0
[
2
=

S
dS = 2π
2
and so (4.5) becomes

∂C0
(q −q
0
)
−1
[q −q
0
[
2
Dqf(q) − 2π
2
f(q
0
)

8V
W
3
.
Since was arbitrary, it follows that

∂C
(q −q
0
)
−1
[q −q
0
[
2
Dqf(q) =

∂C
(q −q
0
)
−1
[q −q
0
[
2
Dqf(q) = 2π
2
f(q
0
). 2
Because of the special role played by the function occurring in this integral formula,
we will use a special notation for it:
G(q) =
q
−1
[q[
2
. (4.6)
Note that
G(q) = −∂

1
[q[
2

= −∂
r

1
[q[
2

; (4.7)
it follows that
¯

G = 0, i.e. G is regular except at 0.
As an immediate corollary of the integral formula we have
Theorem 11 A function which is regular in an open set U is real-analytic in U.
Proof Suppose f is analytic in the open set U, and let q
0
be any point of U. Then we
can ﬁnd a 4-parallelepiped C
0
such that q
0
lies in the interior of C
0
and C
0
lies inside
U, and by theorem 10 we have
f(q
0
) =
1

2

∂C0
G(q − q
0
)Dqf(q).
Let C
1
be a 4-parallelepiped such that q
0
∈ C
1
(I
4
) ⊂ int C
0
(I
4
); then in this integral
the integrand is a continuous function of (q, q
0
) in C
0
(∂I
4
) C
1
(I
4
) and a C

func-
tion of q
0
in C
1
(I
4
). It follows that the integral deﬁnes a C

function of q
0
in C
1
(I
4
).
Thus f is C

throughout U. Since f is regular in U, so that
¯

f = 0, eq. (2.45) gives
∆f = 0, i.e. f is harmonic. It follows [12, p. 269] that f is real-analytic in U. 2
This fact enables us to give the most general formulation of Cauchy’s theorem and
the integral formula for a differentiable contour of integration.
21
Theorem 12 (Cauchy’s theorem for a differentiable contour)
Suppose f is regular in an open set U, and let C be a differentiable 3-chain in U which
is homologous to 0 in the differentiable singular homology of U, i.e. C = ∂C

for
some differentiable 4-chain C

in U. Then

C
Dq f = 0.
Proof By theorem 8, d(dqf) = 0. By theorem 11, the 3-form Dqf is inﬁnitely
differentiable in U; hence we can apply Stokes’s theorem to ﬁnd

C
Dqf =

∂C

Dqf =

C

d(Dqf) = 0. 2
In order to state the general form of the integral formula, we need an analogue of
the notion of the winding number of a curve round a point in the plane.
Deﬁnition 5 Let q be any quaternion, and let C be a closed 3-chain in H`¦q¦. Then
C is homologous to a 3-chain C

: ∂I
4
→S, where S is the unit sphere with centre q.
The wrapping number of C about q is the degree of the map C

.
Theorem 13 (the integral formula for a differentiable contour)
Suppose f is regular in an open set U. Let q
0
∈ U, and let C be a differentiable 3-chain
in U`¦q
0
¦ which is homologous, in the differentiable singular homology of U`¦q
0
¦, to
a 3-chain whose image is ∂B for some ball B ⊂ U. Then
1

2

C
(q −q
0
)
−1
[q −q
0
[
2
Dq f(q) = nf(q
0
)
where n is the wrapping number of C about q
0
.
Proof In the case n = 0, C is homologous to 0 in U`¦q
0
¦, so C = ∂C
0
where C
0
is
a differentiable 4-chain in U`¦q
0
¦. Since the 3-form G(q − q
0
)Dqf(q) is closed and
inﬁnitely differentiable in U`¦q
0
¦, Stokes’s theorem gives
1

2

C
G(q −q
0
)dqf(q) =
1

2

C0
d [G(q −q
0
)Dqf(q)] = 0.
In the case n = 1, C is homologous to a 3-chain C

: ∂I
4
→ ∂B where q
0

B ⊂ U and the map C

has degree 1, hence C is homologous to ∂C
0
, for some 4-
parallelepiped C
0
with q
0
∈ int C
0
(I
4
) and C
0
(I
4
) ⊂ U. Again using the fact that the
3-form G(q −q
0
)Dq is closed in U`¦q
0
¦, we have
1

2

C
G(q −q
0
)Dqf(q) =
1

2

∂C0
G(q −q
0
)Dqf(q) = f(q
0
)
by the integral formula for a parallelepiped.
For general positive n, C is homologous to a 3-chain C

of the form C

= ρ ◦ C

where C

(having image ∂B) is as in the previous paragraph and ρ : ∂B → ∂B is a
map of degree n, e.g.
ρ [q
0
+ r(v + jw)] = q
0
+ r(v
n
+ jw)
22
where r is the radius of B. Dissect C

as C

=
¸
n
=1
C

where the image of C

is the
sector

q = q
0
+ r(v + jw) ∈ ∂B :
2π( −1)
n
≤ arg v ≤

n

.
Then each ρ◦C

has image ∂B and wrapping number 1 about q
0
, and so by the previous
paragraph
1

2

C
G(q −q
0
)Dqf(q) =
1

2
n
¸
=1

ρ◦C

G(q −q
0
)Dqf(q) = nf(q
0
).
In the case n = −1, C is a homologous to a 3-chain C

of the form C

= C

◦ K,
where C

: ∂I
4
→ ∂B has degree 1 and K : ∂I
4
→ ∂I
4
has degree −1, for example
the reﬂection (t
1
, t
2
, t
3
, t
4
) →(1 −t
1
, t
2
, t
3
, t
4
). Then

C
=

C

= −

C

= −2π
2
f(q
0
),
the integrand G(q − q
0
)Dqf(q) being understood. For general negative n we dissect
the 3-chain C as for general positive n. This establishes the theorem for all n. 2
More generally, Cauchy’s theorem and the integral formula are valid for rectiﬁable
contours, which we deﬁne as follows.
Deﬁnition 6 Let C : I
3
→ H be a continuous map of the unit 3-cube into H, and
let P : 0 = s
0
< s
1
< ... < s
p
= 1, Q : 0 = t
0
< t
1
< ... < t
q
= 1 and
R : 0 = u
0
< u
1
< ... < u
r
= 1 be three partitions of the unit interval I. Deﬁne
σ(C; P, Q, R) =
p−1
¸
l=0
q−1
¸
m=0
r−1
¸
n=0
Dq(C(s
l+1
, t
m
, u
n
) −C(s

, t
m
, u
n
),
C(s

, t
m+1
, u
n
) −C(s

, t
m
, u
n
),
C(s

, t
m
, u
n+1
) −C(s

, t
m
, u
n
)).
C is a rectiﬁable 3-cell if there is a real number M such that σ(C; P, Q, R) < M
for all partitions P, Q, R. If this is the case the least upper bound of the numbers
(C; P, Q, R) is called the content of C and denoted by σ(C).
Let f and g be quaternion-valued functions deﬁned on C(I
3
). We say that fDqg
is integrable over C if the sum
p−1
¸
l=0
q−1
¸
m=0
r−1
¸
n=0
f(C(¯ s

,
¯
t
m
, ¯ u
n
))Dq(C(s
l+1
, t
m
, u
n
) −C(s

, t
m
, u
n
),
C(s

, t
m+1
, u
n
) −C(s

, t
m
, u
n
),
C(s

, t
m
, u
n+1
) −C(s

, t
m
, u
n
))g(C(¯ s,
¯
t
m
, ¯ u
n
)),
where s

≤ ¯ s

≤ s
+1
, t
m

¯
t
m
≤ t
m+1
and u
n
≤ ¯ u
n
≤ u
n+1
, has a limit in the
sense of Riemann-Stieltjes integration as [P[, [Q[, [R[ →0, where [P[ = max
0≤≤p−1
[s
+1

s

[ measures the coarseness of the partition P. If this limit exists, we denote it by

C
fDqg.
23
We extend these deﬁnitions to deﬁne rectiﬁable 3-chains and integrals over rectiﬁable
3-chains in the usual way.
Just as for rectiﬁable curves, we can show that fDqg is integrable over the 3-chain
C if f and g are continuous and C is rectiﬁable, and

C
f Dq g

≤ (max
C
[f[)(max
C
[g[)σ(C).
Furthermore, we have the following weak form of Stokes’s theorem:
Stokes’s theorem for a rectiﬁable contour. Let C be a rectiﬁable 3-chain in H with
∂C = 0, and suppose f and g are continuous functions deﬁned in a neighbourhood U
of the image of C, and that fDqg = dω where ω is a 2-form on U. Then

C
f Dq g = 0.
The proof proceeds by approximating C by a chain of 3-parallelepipeds with vertices at
the points C
a
(s

, t
m
, u
n
) where C
a
is a 3-cell in C and s

, t
m
, u
n
are partition points
in I. Stokes’s theorem holds for this chain of 3-parallelepipeds, and we can use the
same argument as for rectiﬁable curves (see e.g. [13, p. 103]).
We can now give the most general forms of Cauchy’s theorem and the integral
formula.
Theorem 14 (Cauchy’s theorem for a rectiﬁable contour)
Suppose f is regular in an open set U, and let C be a rectiﬁable 3-chain which is
homologous to 0 in the singular homology of U. Then

C
Dq f = 0.
Proof First we prove the theorem in the case when U is contractible. In this case, since
d(Dqf) = 0 and f is continuously differentiable (by theorem 11), Poincar´ e’s lemma
applies and we have Dqf = dω for some 2-form ω on D. But ∂C = 0, so by Stokes’s
theorem

C
Dqf = 0.
In the general case, suppose C = ∂C

where C

is a 4-chain in U. We can dissect
C

as C

=
¸
n
C

n
, where each C

n
is a 4-cell lying inside an open ball contained
in U and C

n
is rectiﬁable. Hence by the ﬁrst part of the theorem

∂C

n
Dqf = 0, and
therefore

C
Dq f =
¸
n

∂C

n
Dqf = 0. 2
A similar argument proves the following general form of the integral formula
Theorem 15 (the integral formula for a rectiﬁable contour)
Suppose f is regular in an open set U. Let q
0
∈ U, and let C be a rectiﬁable 3-chain in
U`¦q
0
¦ which is homologous, in the singular homology of U`¦q
0
¦, to a differentiable
3-chain whose image is ∂B for some ball B ⊂ U. Then
1

2

C
(q −q
0
)
−1
[q −q
0
[
2
Dq f(q) = nf(q
0
)
where n is the wrapping number of C about q
0
.
24
5 Some Immediate Consequences
Since regular functions are harmonic, they satisfy a maximum-modulus principle and
a Liouville theorem. As with functions of a complex variable, Liouville’s theorem
follows immediately from the Cauchy-Fueter integral formula, as in e.g. [14, p. 85
(second proof)].
Morera’s theorem also holds for quaternionic functions, but in this case the usual
proof cannot easily be adapted. The proof given here is based on an incomplete proof
by Schuler [8].
Theorem 16 (Morera’s theorem) Suppose that the function f is continuous in an
open set U and that

∂C
Dqf = 0 for every 4-parallelepiped C contained in U. Then
f is regular in U.
Proof The method is to show that f satisﬁes the integral formula and then argue as for
the analyticity of a regular function (theorem 11).
First we show that

∂C
G(q − q
0
)Dqf(q) = 0 if q
0
does not lie inside C, using
Goursat’s argument. As in theorem 9, we ﬁnd a sequence of 4-parallelepipeds C
n
,
converging on a point q

and satisfying

∂Cn
G(q −q
0
)dqf(q)

1
16
n

∂C
G(q −q
0
)dqf(q)

. (5.1)
Since q
0
lies outside C, G(q − q
0
) is a right-regular function of q inside C and so, by
the counterpart of theorem 8 for right-regular functions, dG
q−q0
∧Dq = 0. Now write
G(q −q
0
) = G(q

−q
0
) + α(q −q

) + (q −q

)r(q)
where α = dG
q∞−q0
∈ F
1
, so that α ∧ Dq = 0, and r(q) →0 as q →q

; and write
f(q) = f(q

) + s(q)
where s(q) →0 as q →q

. Then

∂Cn
G(q −q
0
)Dqf(q) = G(q

−q
0
)

∂Cn
Dqf(q) +

∂C
α(q −q

)Dqf(q
0
)
+

∂Cn
α(q −q

)Dqs(q) +

∂Cn
(q −q

)r(q)Dqf(q)
The ﬁrst term vanishes by assumption, the second because α ∧ Dq = 0. Now for
q ∈ ∂C
n
we have [q − q

[ ≤ 2
−n
L, where L is the sum of the lengths of the edges
of C; since α is linear, there is a number M such that [α(q − q

[ ≤ M[q −q

[. The
volume of each face of C
n
is at most 8
−n
V , where V is the volume of the largest face
of C. Hence

∂Cn
G(q −q
0
)Dqf(q)

≤ 2
−n
LM.8
−n
V σ
n
+ 2
−n

n
.8
−n
V ([f(q

)[ + σ
n
)
25
where ρ
n
and σ
n
are the maximum values of the continuous functions r(q) and s(q)
on ∂C
n
. Since ρ
n
→0 and σ
n
→0 as n →∞, it follows that
16
n

Cn
G(q −q
0
)Dqf(q)

→0 as n →∞
and so from (5.1) that

∂C
G(q −q
0
)Dqf(q) = 0.
Now consider the integral

∂C
G(q − q
0
)Dqf(q) where q
0
lies inside C. By what
we have just proved, the parallelepiped can be replaced by a small parallelepiped C
0
containing q
0
, as in theorem 10. After this point has been established, the proof of
theorem10 depends only on the continuity of f and is therefore valid under the present
conditions; hence
f(q) =
1

2

∂C
G(q

−q)Dq

f(q

)
for any 4-parallelepiped C with q ∈ int C(I
4
) ⊂ U. Since G(q

−q) is a continuously
differentiable function of q in the interior of C as long as q

lies on its boundary, it
follows, as in theorem 11, that f is differentiable and that
¯

f(q) =
1

2

∂C
¯

[G(q

−q)]Dq

f(q

) = 0
since G is regular. Thus f is regular. 2
6 Construction of Regular Functions
Regular functions can be constructed from harmonic functions in two ways. First, if f
is harmonic then (2.45) shows that ∂

f is regular. Second, any real-valued harmonic
function is, at least locally, the real part of a regular function:
Theorem 17 Let u be a real-valued function deﬁned on a star-shaped open set U ⊆ H.
If u is harmonic and has continuous second derivatives, there is a regular function f
deﬁned on U such that Re f = u.
Proof Without loss of generality we may assume that U contains the origin and is
star-shaped with respect to it. In this case we will show that the function
f(q) = u(q) + 2 Pu

1
0
s
2

u(sq)q ds (6.1)
is regular in U.
Since
Re

1
0
s
2

u(sq)q ds =
1
2

1
0
s
2

t
∂u
∂t
(sq) + x
i
∂u
∂x
i
(sq)

ds
=
1
2

1
0
s
2
d
ds
[u(sq)] ds
=
1
2
u(q) −

1
0
su(sq) ds,
26
we can write
f(q) = 2

1
0
s
2

u(sq)q ds + 2

1
0
su(sq) ds. (6.2)
Since u and ∂

u have continuous partial derivatives in U, we can differentiate under
the integral sign to obtain, for q ∈ U,
¯

f(q) = 2

1
0
s
2
¯

[∂

u(sq)] q ds+

1
0
s
2
¦∂

u(sq) + e
i

u(sq)e
i
¦ ds+2

1
0
s
2
¯

u(sq) ds.
But
¯

[∂

u(sq)] =
1
4
s∆u(sq) = 0 since u is harmonic in U, and

u(sq) + e
i

u(sq)e
i
= −2∂

u(sq) by (2.27)
= −2
¯

u(sq) since u is real.
Hence
¯

f = 0 in U and so f is regular. 2
If the region U is star-shaped with respect not to the origin but to some other point
a, formulae (6.1) and (6.2) must be adapted by changing origin, thus:
f(q) = u(q) + 2 Pu

1
0
s
2

u((1−s)a+sq)(q−a) ds (6.3)
= 2

1
0
s
2

u((1−s)a+sq)(q−a) ds + 2

1
0
su((1−s)a+sq) ds. (6.4)
An example which can be expected to be important is the case of the function u(q) =
[q[
−2
. This is the elementary potential function in four dimensions, as log [z[ is in the
complex plane, and so the regular function whose real part is [q[
−2
is an analogue of
the logarithm of a complex variable.
We take U to be the whole of H except for the origin and the negative real axis.
Then U is star-shaped with respect to 1, and [q[
−2
is harmonic in U. Put
u(q) =
1
[q[
2
, ∂

u(q) = −
q
−1
[q[
2
, a = 1;
then (6.3) gives
f(q) = −(q Pu q)
−1

1
[ Puq[
2
arg q if Pu q = 0
=
1
[q[
2
if q is real and positive,

(6.5)
where
arg q = log(Un q) =
Puq
[ Puq[
tan
−1

[ Puq[
Re q

, (6.6)
which is i times the usual argument in the complex plane generated by q. (In practice
the formulae (6.3) and (6.4) are not very convenient to use, and it is easier to obtain
(6.5) by solving the equations
∇.F = −
2t
(t
2
+ r
2
)
2
27
and
∂F
∂t
+∇F =
2r
(t
2
+ r
2
)
2
,
where t = Re q, r = Pu q and r = [r[—these express the fact that F : H → P is the
pure quaternion part of a regular function whose real part is [q[
−2
—and assuming that
F has the form F(r)r.)
We will denote the function (6.5) by −2L(q). The derivative of L(q) can most
easily be calculated by writing it in the form
L(q) = −
r
2
+ te
i
x
i
2r
2
(r
2
+ t
2
)
+
e
i
x
i
2r
3
tan
−1

r
t

; (6.7)
the result is

L(q) = G(q) =
q
−1
[q[
2
. (6.8)
Thus L(q) is a primitive for the function occurring in the Cauchy-Fueter integral for-
mula, just as the complex logarithm is a primitive for z
−1
, the function occurring in
Cauchy’s integral formula.
Theorem17 shows that there are as many regular functions of a quaternion variable
as there are harmonic functions of four real variables. However, these functions do not
include the simple algebraic functions, such as powers of the variable, which occur as
analytic functions of a complex variable. Fueter [4] also found a method for construct-
ing a regular function of a quaternion variable from an analytic function of a complex
variable.
For each q ∈ H, let η
q
: C → H be the embedding of the complex numbers in
the quaternions such that q is the image of a complex number ζ(q) lying in the upper
half-plane; i.e.
η
q
(x + iy) = x +
Puq
[ Puq[
y, (6.9)
ζ(q) = Re q + i[ Puq[. (6.10)
Then we have
Theorem 18 Suppose f : C → C is analytic in the open set U ⊆ C, and deﬁne
˜
f : H →H by
˜
f(q) = η
q
◦ f ◦ ζ(q). (6.11)
Then ∆
˜
f is regular in the open set ζ
−1
(U) ⊆ H, and its derivative is

(∆
˜
f) = ∆
˜
f

, (6.12)
where f

is the derivative of the complex function f.
Proof [6] Writing t = Re q, r = Pu q, r = [r[, and u(x, y) = Re f(x+iy), v(x, y) =
Imf(x + iy), the deﬁnition of
˜
f gives
˜
f(q) = u(t, r) +
r
r
v(t, r)
Re

¯

˜
f(q)

=
1
2

∂t
[u(t, r)] −∇.

r
r
v(t, r)

1
2
u
1
(t, r) −
v(t, r)
r

1
2
v
2
(t, r)
(6.13)
28
where the subscript 1 or 2 on u and v denotes partial differentiation with respect to the
ﬁrst or second variable; and
Pu

¯

˜
f(q)

=
1
2

∇[u(t, r)] +

∂t

r
r
v(t, r)

+∇

r
r
v(t, r)

=
1
2

r
r
u
2
(t, r) +
r
r
v
1
(t, r)
¸
.
Since f = u+iv is analytic, the Cauchy-Riemann equations give u
1
−v
2
= u
2
+v
1
= 0
and so
¯

˜
f(q) = −
v(t, r)
r
.
Hence
¯

˜
f(q) = ∆
¯

f(q) = −

2
∂t
2
+
1
r

2
∂r
2
r

v(t, r)
r
= −
v
11
(t, r) + v
22
(t, r)
r
= 0
since v is a harmonic function of two variables. Thus f is regular at q if f is analytic
at ζ(q).
A calculation similar to the above shows that

˜
f(q) =
1
2
(u
1
+ v
2
) +
v
r
+
1
2
r
r
(−u
2
+ v
1
)
= u
1
(t, r) +
r
r
v
1
(t, r) +
v(t, r)
r
by the Cauchy-Riemann equations for f. But f

= u
1
+ iv
1
, so ∂
˜
f =
˜
f

+ v/r; and
∆(v/r) = 0, so

(∆
˜
f) = ∆(∂

˜
f) = ∆
˜
f

. 2
Functions of the form f have been taken as the basis of an alternative theory of
functions of a quaternion variable by Cullen [15].
From (5.13) a straightforward calculation, using the fact that u and v are harmonic
functions, gives

˜
f(q) =
2u
2
(t, r)
r
+
2r
r

v
2
(t, r)
r

v(t, r)
r
2

(6.14)
The following examples are interesting: When
f(z) = z
−1
, ∆
˜
f(q) = −4G(q); (6.15)
when
f(z) = log z, ∆
˜
f(q) = −4L(q) (6.16)
29
7 Regular Functions and Conformal Mappings
Because the quaternions are four-dimensional, very little remains in quaternionic anal-
ysis of the relation between analytic functions and conformal mappings in complex
analysis. However, the conformal group of H acts on regular functions in a simple
way. The action of rotations and inversions will be needed in studying regular poly-
nomials, and so it seems appropriate to present here the action of the full conformal
group.
Theorem 19 Let H

= H ∪ ¦∞¦ be the one-point compactiﬁcation of H. If the
mapping f : H

→H

is conformal and orientation-preserving, f is of the form
f(q) = (aq + b)(cq + d)
−1
(7.1)
for some a, b, c, d ∈ H. Conversely, any such mapping is conformal and orientation-
preserving.
Proof Let C be the group of orientation-preserving conformal mappings of H

, and let
D be the set of mappings of the form (7.1). Then if f ∈ D, f has differential
df
q
= (ac
−1
d −b)(cq + d)
−1
c dq(cq + d)
−1
This is of the form αdqβ, which is a combination of a dilatation and a rotation, so f is
conformal and orientation-preserving. Thus D is a subset of C. Now C is generated
by rotations, dilatations, translations and the inversion in the unit sphere followed by a
reﬂection [16, p. 312], i.e. by the mappings q → αqβ, q → q + γ and q → q
−1
. If a
mapping in D is followed by any of these mappings, it remains in D; hence CD ⊆ D.
It follows that D = C. 2
The same argument can be used to obtain the alternative representation f(q) =
(qc + d)
−1
(qa + b).
We now show how a regular function gives rise to other regular functions by con-
formal transformation of the variable.
Theorem 20
(i) Given a function f : H →H, let If : H`¦0¦ →H be the function
If (q) =
q
−1
[q[
2
f(q
−1
).
If f is regular at q
−1
, If is regular at q.
(ii) Given a function f : H →H and quaternions a, b, let M(a, b)f be the function
[M(a, b)f](q) = bf(a
−1
qb).
If f is regular at aqb, M(a, b)f is regular at q.
30
(iii) Given a function f : H → H and a conformal mapping ν : q → (aq + b)(cq +
d)
−1
, let M(ν)f be the function
[M(ν)f](q) =
1
[b −ac
−1
d[
2
(cq + d)
−1
[cq + d[
2
f(ν(q)).
If f is regular at ν(q), M(ν)f is regular at q.
Proof (i) By theorem 8, it is sufﬁcient to show that
Dq ∧ d(If )
q
= 0.
Now If = G(f ◦ ι), where G(q) =
q
−1
|q|
2
and ι : H`¦0¦ →H is the inversion q →q
−1
.
Hence
Dq ∧ d(If )
q
= Dq ∧ dG
q
f(q
−1
) + Dq ∧ G(q)d(f ◦ ı)
q
= Dq G(q) ∧ ı

q
df
q
−1
since G is regular at q = 0. But
ı

q
Dq(h
1
, h
2
, h
3
) = Dq(−q
−1
h
1
q
−1
, −q
−1
h
2
q
−1
, −q
−1
h
3
q
−1
)
= −
q
−1
[q[
4
Dq(h
1
, h
2
, h
3
)q
−1
by (2.41). Thus
Dq G(q) = −[q[
2

q
Dq
and so
Dq ∧ d(If )
q
= −[q[
2

q
(Dq ∧ df
q
−1 )
= 0
if f is regular at q
−1
. 2
(ii) Let µ : H →H be the map q →aqb. Then by (2.41)
µ

Dq = [a[
2
[b[
2
a Dq b
and so
Dq ∧ d[M(a, b)f]
q
= Dq ∧ bµ

q
df
µ(q)
= [a[
−2
[b[
−2
a
−1

q
Dq)b
−1
∧ bµ

q
df
µ(q)
= [a[
−2
[b[
−2
a
−1
µ

q
(Dq ∧ df
µ(q)
)
= 0
if f is regular at µ(q). It follows from theorem 8 that M(a, b)f is regular at q. 2
(iii) The map q → ν(q) = (aq + b)(cq + d)
−1
can be obtained by composing the
sequence of maps
q →q

= cq(b −ac
−1
d)
−1
(7.2)
q

→q

= q

+ d(b −ac
−1
d)
−1
(7.3)
q

→q

= q
−1
(7.4)
q

→ν(q) = q

+ ac
−1
(7.5)
Clearly translation preserves regularity, i.e. if f is regular at q +α, f(q +α) is regular
at q. Applying this to the maps (7.3), part (i) of the theorem to (7.4) and part (ii) to
(7.2), we ﬁnd that M(ν)f is regular at q if f is regular at ν(q). 2
31
8 Homogeneous Regular Functions
In this section we will study the relations between regular polynomials, harmonic poly-
nomials and harmonic analysis on the group S of unit quaternions, which is to quater-
nionic analysis what Fourier analysis is to complex analysis.
The basic Fourier functions e
inθ
and e
−inθ
, regarded as functions on the unit cir-
cle in the complex plane, each have two extensions to harmonic functions on C`¦0¦;
thus we have the four functions z
n
, ¯ z
n
, z
−n
and ¯ z
−n
. The requirement of analytic-
ity picks out half of these, namely z
n
and z
−n
. In the same way the basic harmonic
functions on S, namely the matrix elements of unitary irreducible representations of
S, each have two extensions to harmonic functions on H`¦0¦, one with a negative
degree of homogeneity and one with a positive degree. We will see that the space of
functions belonging to a particular unitary representation, corresponding to the space
of combinations of e
inθ
and e
−inθ
for a particular value of n, can be decomposed into
two complementary subspaces; one (like e
inθ
) gives a regular function on H`¦0¦ when
multiplied by a positive power of [q[, the other (like e
−inθ
) has to be multiplied by a
negative power of [q[.
Let U
n
be the set of functions f : H`¦0¦ →Hwhich are regular and homogeneous
of degree n over R, i.e.
f(αq) = α
n
f(q) for α ∈ R.
Removing the origin from the domain of f makes it possible to consider both positive
and negative n (the alternative procedure of adding a point at inﬁnity to H has disad-
vantages, since regular polynomials do not necessarily admit a continuous extension to
H∪ ¦∞¦

= S
4
). Let W
n
be the set of functions f : H`¦0¦ →H which are harmonic
and homogeneous of degree n over R. Then U
n
and W
n
are right vector spaces over
H (with pointwise addition and scalar multiplication) and since every regular function
is harmonic, we have U
n
⊆ W
n
.
Functions in U
n
and W
n
can be studied by means of their restriction to the unit
sphere S = ¦q : [q[ = 1¦. Let
˜
U
n
= ¦f[S : f ∈ U
n
¦,
˜
W
n
= ¦f[S : f ∈ W
n
¦;
then U
n
and
˜
U
n
are isomorphic (as quaternionic vector spaces) by virtue of the corre-
spondence
f ∈ U
n

˜
f ∈
˜
U
n
, where f(q) = r
n
˜
f(u), (8.1)
using the notation r = [q[ ∈ R, u =
q
|q|
∈ S.
Similarly W
n
and
˜
W
n
are isomorphic.
In order to express the Cauchy-Riemann-Fueter equations in a form adapted to the
polar decomposition q = ru, we introduce the following vector ﬁelds X
0
, ..., X
3
on
H`¦0¦:
X
0
f =
d

f(qe
θ
)

θ=0
, (8.2)
X
i
f =
d

f(qe
eiθ
)

θ=0
=
d

f[q(cos θ + e
i
sin θ)]
θ=0
(i = 1, 2, 3). (8.3)
32
These ﬁelds form a basis for the real vector space of left-invariant vector ﬁelds on the
multiplicative group of H, and they are related to the Cartesian vector ﬁelds

∂t
,

∂xi
by
X
0
= t

∂t
+ x
i

∂x
i
, (8.4)
X
i
= −x
i

∂t
+ t

∂x
i

ijk
x
j

∂x
k
, (8.5)

∂t
=
1
r
2
(tX
0
−x
i
X
i
), (8.6)

∂x
i
=
1
r
2
(
ijk
x
j
X
k
+ tX
i
+ x
i
X
0
). (8.7)
Their Lie brackets are
[X
0
, X
i
] = 0, (8.8)
[X
i
, X
j
] = 2
ijk
X
k
. (8.9)
Using (3.4) and (8.5) the differential operators
¯

and ∆ can be calculated in terms
X
0
and X
i
. The result is
¯

=
1
2
¯ q
−1
(X
0
+ e
i
X
i
), (8.10)
∆ =
1
r
2
¦X
i
X
i
+ X
0
(X
0
+ 2)¦ . (8.11)
From (8.11) we can deduce the following (well-known) facts about W
n
:
Theorem 21
(i)
˜
W
n
=
˜
W
−n−2
(ii) dimW
n
= (n + 1)
2
(iii) The elements of W
n
are polynomials in q.
Proof The elements of W
n
, being homogeneous of degree n, are eigenfunctions of
X
0
with eigenvalue n. Since they are also harmonic, e.g. (8.11) shows that they are
eigenfunctions of X
i
X
i
with eigenvalue −n(n + 2). Now the vector ﬁelds X
i
are
tangential to the sphere S, so their restrictions
˜
X
i
= X
i
[S are vector ﬁelds on S; they
are a basis for the real vector space of left-invariant vector ﬁelds on the Lie group S,
which is isomorphic to SU(2). Thus if
˜
f ∈
˜
W
n
,
˜
f = f[S for some f ∈ W
n
, so f is an
eigenfunction of X
i
X
i
with eigenvalue −n(n +2) and therefore
˜
f is an eigenfunction
of
˜
X
i
˜
X
i
with eigenvalue −n(n+2). Conversely, if
˜
f is an eigenfunction of
˜
X
i
˜
X
i
with
eigenvalue −n(n + 2), then

r
n
˜
f(u)

=
1
r
2

r
n
˜
X
i
˜
X
i
˜
f + [X
0
(X
0
+ 2)r
n
]
˜
f
¸
= 0
so
˜
f ∈
˜
W
n
. Thus
˜
W
n
is the space of eigenfunctions of
˜
X
i
˜
X
i
with eigenvalue −n(n +
2).
33
It follows immediately that
˜
W
n
=
˜
W
−n−2
. 2(i)
˜
W
n
is the quaternioniﬁcation of the complex vector space
˜
W
c
n
of complex-valued
functions on S which have eigenvalue −n(n + 2) for
˜
X
i
˜
X
i
. It is known [17, p. 71]
that this space has a basis consisting of the matrix elements of the (n+1)-dimensional
representation of the group S. Hence
dim
H
˜
W
n
= dim
C
W
c
n
= (n + 1)
2
. 2(ii)
In particular,
˜
W
0
consists only of constant functions, and therefore so does W
0
.
Nowif f belongs to W
n
,
∂f
∂t
and
∂f
∂xi
belong to W
n−1
, and all the nth partial derivatives
of f belong to W
0
; hence all the (n + 1)th partial derivatives of f vanish and so f is a
polynomial. 2(iii)
Theorem 22
(i)
˜
W
n
=
˜
U
n

˜
U
−n−2
(ii) U
n

= U
−n−3
(iii) dimU
n
=
1
2
(n + 1)(n + 2)
Proof (i) Eq. (8.10) shows that the elements of U
n
, which satisfy X
0
f = nf and
¯

i
f = 0, are eigenfunctions of Ω = e
i
X
i
with eigenvalue −n. As in theorem 21, it
follows that U
n
consists of the eigenfunctions of Ω = e
i
X
i
with eigenvalue −n. Now
using (8.9), it can be shown that Ω satisﬁes the equation

2
−2Ω + X
i
X
i
= 0
and therefore
˜

2
−2
˜
Ω + X
i
X
i
= 0.
Hence
˜
f ∈
˜
W
n

˜
X
i
˜
X
i
f = −n(n + 2)f
⇒(
˜

−n−2
)(
˜

+n
)f = 0
It follows that
˜
W
n
is the direct sum of the eigenspaces of
˜
Ω with eigenvalues −n and
n + 2 (these are vector subspaces of
˜
W
n
since the eigenvalues are real), i.e.
˜
W
n
=
˜
U
n

˜
U
−n−2
2(i)
(ii) It follows from theorem 20(i) that the mapping I is an isomorphism between
U
n
and U
−n−3
. 2(ii)
(iii) Let d
n
= dimU
n
. By (1) and theorem 21 (ii),
d
n
+ d
−n−2
= (n + 1)
2
and by (ii), d
−n−2
= d
n−1
. Thus d
n
+d
n−1
= (n+1)
2
. The solution of this recurrence
relation, with d
0
= 1, is
d
n
=
1
2
(n + 1)(n + 2). 2(iii)
34
There is a relation between theorem 20(ii) and the fact that homogeneous regular
functions are eigenfunctions of Ω. Theorem 20(ii) refers to a representation M of the
group H
×
H
×
deﬁned on the space of real-differentiable functions f : H`¦0¦ →H
by
[M(a, b)f](q) = b f(a
−1
qb).
Restricting to the subgroup ¦(a, b) : [a[ = [b[ = 1¦, which is isomorphic to SU(2)
SU(2), we obtain a representation of SU(2) SU(2). Since the map q → aqb is a
rotation when [a[ = [b[ = 1, the set W of harmonic functions is an invariant subspace
under this representation. NowW = H⊗
C
W
c
, where W
c
is the set of complex-valued
harmonic functions, and the representation of SU(2) SU(2) can be written as
M(a, b)(q ⊗f) = (bq) ⊗R(a, b)f
where R denotes the quasi-regular representation corresponding to the action q →
aqb
−1
of SU(2) SU(2) on H`¦0¦:
[R(a, b)f]q = f(a
−1
qb).
Thus M[W is the tensor product of the representations D
0
D
1
and R[W
c
of SU(2)
SU(2), where D
n
denotes the (n + l)-dimensional complex representation of SU(2).
The isotypic components of R[W
c
are the homogeneous subspaces W
c
n
, on which R
acts irreducibly as D
n
D
n
; thus W
n
is an invariant subspace under the representation
M, and M[W
n
is the tensor product (D
0
D
1
) ⊗(D
n
D
n
). W
n
therefore has two
invariant subspaces, on which M acts as the irreducible representations D
n
D
n+1
and
D
n
D
n−1
. These subspaces are the eigenspaces of Ω. To see this, restrict attention
to the second factor in SU(2) SU(2); we have the representation
M

(b)(q ⊗f) = M(1, b)(q ⊗f) = [D
1
(b)q] ⊗[R(1, b)f]
where D
1
(b)q = bq. The inﬁnitesimal generators of the representation R(1, b) are the
differential operators X
i
; the inﬁnitesimal generators of D
1
(b) are e
i
(by which we
mean left multiplication by e
i
). Hence the inﬁnitesimal operators of the tensor product
M

are e
i
+ X
i
. The isotypic components of W are the eigenspaces of the Casimir
operator
(e
i
+ X
i
)(e
i
+ X
i
) = e
i
e
i
+ X
i
X
i
+ 2Ω.
But e
i
e
i
= −3, and X
i
X
i
= −n(n + 2) on W
n
; hence
(e
i
+ X
i
)(e
i
+ X
i
) = 2Ω−n
2
−2n −3.
and so the isotypic components of W
n
for the representation M

are the eigenspaces
of Ω. U
n
, the space of homogeneous regular functions of degree n, has eigenvalue −n
for Ω, and so M

[U
n
is the representation D
n+1
of SU(2).
Similar considerations lead to the following fact:
Theorem 23 If f is regular, qf is harmonic.
35
Proof First we show that f ∈ U
n
⇒ qf ∈ W
n
. From the deﬁnition (8.3) of the
operators X
i
we have
X
i
(q) = qe
i
Hence
X
i
X
i
(qf) = X
i
(qe
i
f + qX
i
f)
= −3qf + 2qe
i
X
i
f + qX
i
X
i
f
since the X
i
are real differential operators. If f ∈ U
n
, it is an eigenfunction of e
i
X
i
with eigenvalue −n and of X
i
X
i
with eigenvalue −n(n + 2), and so
X
i
X
i
(qf) = −(n + 1)(n + 3)f.
Since qf is homogeneous of degree n+1, it follows from(8.11) that it is harmonic. But
any regular function can be represented locally as a series f =
¸
f
n
with f
n
∈ U
n
,
and so the result follows for any regular f. 2
The representation M of SU(2) SU(2) can also be used to ﬁnd a basis of regular
polynomials. It belongs to a class of induced representations which is studied in [18],
where a procedure is given for splitting the representation into irreducible components
and ﬁnding a basis for each component. Rather than give a rigorous heuristic derivation
by following this procedure, which is not very enlightening in this case, we will state
the result and then verify that it is a basis.
Since the functions to be considered involve a number of factorials, we introduce
the notation
z
[n]
=
z
n
n!
if n ≥ 0
= 0 if n < 0
for a complex variable z. This notation allows the convenient formulae
d
dz
z
[n]
= z
[n−1]
, (8.12)
(z
1
+ z
2
)
[n]
=
¸
r
z
[r]
1
z
[n−r]
2
, (8.13)
where the sum is over all integers r.
The representation D
n
of S

= SU(2) acts on the space of homogeneous polynomi-
als of degree n in two complex variables by
[D
n
(u)f](z
1
, z
2
) = f(z

1
, z

2
),
where
z

1
+ jz

2
= u
−1
(z
1
+ jz
2
).
Writing u = v + jw where v, w ∈ C and [v[
2
+[w[
2
= 1, we have
z

1
= ¯ vz
1
+ ¯ wz
2
, z

2
= −wz
1
+ vz
2
.
Hence the matrix elements of D
n
(u) relative to the basis f
k
(z
1
, z
2
) = z
[k]
1
z
[n−k]
2
are
D
n
k
(u) = (−)
n
k!(n −k)!P
n
k
(u)
36
where
P
n
k
(v + jw) =
¸
r
(−)
r
v
[n−k−+r]
¯ v
[r]
w
[k−r]
¯ w
[−r]
(8.14)
The functions P
n
k
(q) are deﬁned for all quaternions q = v + jw and for all integers
k, , n, but they are identically zero unless 0 ≤ k, ≤ n.
Theorem 24 As a right vector space over H, U
n
has the basis
Q
n
k
(q) = P
n
k
(q) −jP
n
k−1,
(q) (0 ≤ k ≤ ≤ n).
Proof Using (8.14), it is easy to verify that Q
n
k
satisﬁes the Cauchy-Riemann-Fueter
equations in the form
∂P
n
k
∂¯ v
= −
∂P
n
k−1,
∂ ¯ w
,
∂P
n
k
∂w
=
∂P
n
k−1,
∂v
.
(cf. 3.16). Since the functions D
n
k
are independent over C as functions on S for 0 ≤ k,
≤ n, the functions P
n
k
are independent over C as functions on H for 0 ≤ k, ≤ n.
It follows that the functions Q
n
k
(0 ≤ k ≤ n + 1, 0 ≤ ≤ n) are independent over C
and therefore span a right vector space over H of dimension at least
1
2
(n + 1)(n + 2).
Since this space is a subspace of U
n
, which has dimension
1
2
(n + 1)(n + 2), the Q
n
k
span U
n
.
Since zj = j¯ z for any z ∈ C, it can be seen from the deﬁnition (8.14) that
P
n
k
j = j P
n
n−k,n−
and therefore
Q
n
k
j = Q
n
n−k+1,n−
.
Thus U
n
is spanned by the Q
n
k
(0 ≤ k ≤ ≤ n), which therefore form a basis for
U
n
. 2
Another basis for U
n
will be given in the next section.
We conclude this section by studying the quaternionic derivative ∂

. Since ∂

is a
linear map from U
n
into U
n−1
and dimU
n
> dimU
n−1
, ∂

must have a large kernel
and so we cannot conclude from ∂

f = 0 that f is constant. However, although the
result is far from unique, it is possible to integrate regular polynomials:
Theorem 25 Every regular polynomial has a primitive, i.e. ∂

maps U
n
onto U
n−1
if
n > 0.
Proof Suppose f ∈ U
n
is such that ∂

f = 0. Then
∂f
∂t
= e
i
∂f
∂x
i
= 0.
Thus f can be regarded as a function on the space P of pure imaginary quaternions.
Using vector notation for elements of P and writing f = f
0
+ f with f
0
∈ R, f ∈ P,
the condition e
i
∂f
∂xi
= 0 becomes
∇f
0
+∇f = 0, ∇.f = 0.
37
If n ≥ 0, we can deﬁne f(0) so that these hold throughout P, and so there exists a
function F : P →P such that
f = ∇F, f
0
= −∇.F,
i.e.
f = e
i
∂F
∂x
i
.
Then F is harmonic, i.e. ∇
2
F = 0.
Let T
n
be the right quaternionic vector space of functions F : P → H which are
homogeneous of degree n and satisfy ∇
2
F = 0; then dimT
n
= 2n+1. Let K
n
be the
subspace of T
n
consisting of functions satisfying e
i
∂F
∂xi
= 0; then K
n
= ker ∂

⊂ U
n
.
The above shows that e
i

∂xi
: T
n+1
→ T
n
maps T
n+1
onto K
n
; its kernel is K
n+1
,
and so
dimK
n
+ dimK
n+1
= dimT
n+1
= 2n + 3.
The solution of this recurrence relation, with dimK
0
= 1, is dimK
n
= n + 1. But
dimU
n
−dimU
n−1
=
1
2
(n + 1)(n + 2) −
1
2
n(n + 1) = n + 1.
It follows that ∂

maps U
n
onto U
n−1
. 2
Theorem 26 If n < 0, the map ∂

: U
n
→U
n−1
is one-to-one.
Proof We introduce the following inner product between functions deﬁned on the unit
sphere S:
'f, g` =

S
f(u)g(u)du
where du denotes Haar measure on the group S, normalised so that

S
du =
1
2
π
2
. For
functions deﬁned on H, we can use (4.6) to write this as
'f, g` =

S
f(q)q
−1
Dq g(q)
As a map : U
n
U
n
→H, this is antilinear in the ﬁrst variable and linear in the second,
i.e.
'fa, gb` = ¯ a'f, g`b for all a, b ∈ H
and is non-degenerate since 'f, f` = 0 ⇐⇒ f = 0.
Now let f ∈ U
n
, g ∈ U
−n−2
and let I denote the map : U
n
→ U
−n−3
deﬁned in
theorem 20(i). Then
'g, I∂

f` =

S
g(q)q
−1
Dq q
−1

f(q
−1
)
= −

S
g(q)ı

(Dq ∂

f),
38
where ι denotes the map q →q
−1
and we have used the fact that ι

Dq = −q
−1
Dqq
−1
for q ∈ S. Since f is regular, Dq∂

f =
1
2
d(dq ∧ dqf) and so
'g, I∂

f` = −
1
2

S
g(q)d[ı

(dq ∧ dq f)]
= −
1
2

S
d¯ g ∧ ı

(dq ∧ dq f) since ∂S = 0.
On S, the inversion ι coincides with quaternion conjugation; hence ι

dq = d¯ q and
therefore
'g, I∂

f` = −
1
2

S
d¯ g ∧ d¯ q ∧ d¯ qf(q
−1
)
= −
1
2

S
dq ∧ dq ∧ dgf(q
−1
)
= −

S
Dq ∧ ∂

g(q)f(q
−1
)
since g is regular. Since conjugation is an orthogonal transformation with determinant
−1, Dq(
¯
h
1
,
¯
h
2
,
¯
h
3
) = −Dq(h
1
, h
2
, h
3
); hence, because conjugation is the same as
inversion on S,
Dq = −ı

Dq = q
−1
Dq q
−1
.
Thus
'g, I∂

f` = −

S

g(q)q
−1
Dq q
−1
f(q
−1
)
= −'∂

g, If `.
But I is an isomorphism, the inner product is non-degenerate on U
−n−2
, and ∂

maps
U
−n−2
onto U
−n−3
if n ≤ −3; it follows that ∂

: U
n
→U
n−1
is one-to-one. 2
In the missing cases n = −1 and n = −2, theorems 25 and 26 are both true
trivially, since U
−1
= U
−2
= ¦0¦.
9 Regular Power Series
The power series representing a regular function, and the Laurent series representing a
function with an isolated singularity, are most naturally expressed in terms of certain
special homogeneous functions.
Let ν be an unordered set of n integers ¦i
1
, ..., i
n
¦ with 1 ≤ i
r
≤ 3; ν can also
be speciﬁed by three integers n
1
, n
2
, n
3
with n
1
+ n
2
+ n
3
= n; where n
1
is the
number of 1’s in ν, n
2
the number of 2’s and n
3
the number of 3’s, and we will write
ν = [n
1
n
2
n
3
]. There are
1
2
(n + 1)(n + 2) such sets ν; we will denote the set of all
of them by σ
n
. They are to be used as labels; when n = 0, so that ν = ∅, we use the
sufﬁx 0 instead of ∅. We write ∂
ν
for the n’th order differential operator

ν
=

n
∂x
i1
...∂x
in
=

n
∂x
n1
∂y
n2
∂z
n3
.
39
The functions in question are
G
ν
(q) = ∂
ν
G(q) (9.1)
and
P
ν
(q) =
1
n!
¸
(te
i1
−x
i1
)...(te
in
−x
in
) (9.2)
where the sum is over all
n!
n1!n2!n3!
different orderings of n
1
1’s, n
2
2’s and n
3
3’s.
Then P
ν
is homogeneous of degree n and G
ν
is homogeneous of degree −n −3.
As in the previous section, U
n
will denote the right quaternionic vector space of
homogeneous regular functions of degree n.
Theorem 27 The polynomials P
ν
(ν ∈ σ
n
) are regular and form a basis for U
n
.
Proof [17] Let f be a regular homogeneous polynomial of degree n. Since f is regular
∂f
∂t
+
¸
i
e
i
∂f
∂x
i
= 0
and since it is homogeneous,
t
∂f
∂t
+
¸
i
x
i
∂f
∂x
i
= 0.
Hence
nf(q) =
¸
i
(x
i
−te
i
)
∂f
∂x
i
.
But
∂f
∂xi
is regular and homogeneous of degree n − 1, so we can repeat the argument;
after n steps we obtain
f(q) =
1
n!
¸
i1...in
(x
i1
− te
i1
)...(x
in
−te
in
)

n
f
∂x
i1
...∂x
in
=
¸
ν∈σn
(−1)
n
P
ν
(q)∂
ν
f(q).
Since f is a polynomial, ∂
ν
f is a constant; thus any regular homogeneous polynomial
is a linear combination of the P
ν
. Let V
n
be the right vector space spanned by the P
ν
.
By theorem 21(iii), the elements of U
n
are polynomials, so U
n
⊆ V
n
; but dimV
n

1
2
(n + 1)(n + 2) = dimU
n
by theorem 22(iii). Hence V
n
= U
n
. 2
The mirror image of this argument proves that the P
ν
are also right-regular.
Theorem 28 The expansions
G(p −q) =

¸
n=0
¸
ν∈σn
P
ν
(q)G
ν
(p) (9.3)
=

¸
n=0
¸
ν∈σn
G
ν
(p)P
ν
(q) (9.4)
are valid for [q[ < [p[; the series converge uniformly in any region ¦(p, q) : [q[ ≤ r[p[¦
with r < 1.
40
Proof Just as for a complex variable, we have
(1 −q)
−1
=

¸
n=0
q
n
for [q[ < 1; the series converges absolutely and uniformly in any ball [q[ ≤ r with
r < 1. Hence G(1 − q) = (1 − q)
−2
(1 − ¯ q) can be expanded as a power series in q
and ¯ q which converges uniformly in any ball with radius less than 1. Because G has
the multiplicative property
G(q
1
q
2
) = G(q
2
)G(q
1
), (9.5)
it follows that G(p−q) can be expanded as a power series in p
−1
q, multiplied by G(p);
the series converges uniformly in any region [p
−1
q[ ≤ r with r < 1.
Regarding this series as a power series in q and identifying it with the Taylor series
of G about p, we have
G(p −q) =

¸
r,s=0
¸
i1...is
(−1)
r+s
(r + s)!

r+s
G
∂t
r
∂x
i1
...∂x
is
(p) t
r
x
i1
...x
is
. (9.6)
Since G is regular, each derivative with respect to t can be replaced by the combi-
nation −
¸
i
e
i

∂xi
, giving
G(p −q) =

¸
n=0
1
n!
¸
r+s=n
¸
j
1
...j
r
i
1
...i
s
(te
j1
)...(te
jr
)(−x
i1
)...(−x
is
)

n
G
∂x
j1
...∂x
jr
∂x
i1
...∂x
is
(p)
=

¸
n=0
¸
ν∈σn
P
ν
(q)G
ν
(p).
This proves (9.5). But G is also right-regular, i.e.
∂G
∂t
= −
¸
i
∂G
∂x
i
e
i
,
so the derivatives with respect to t in (9.6) can alternatively be replaced by combina-
tions of derivatives with respect to the x
i
with coefﬁcients on the right, thus giving
G(p −q) =

¸
n=0
¸
ν∈σn
G
ν
(p)P
ν
(q). 2
Theorem 29 Suppose f is regular in a neighbourhood of 0. Then there is a ball B
with centre 0 in which f(q) is represented by a uniformly convergent series
f(q) =
¸
n=0
¸
ν∈σn
P
ν
(q)a
ν
, (9.7)
41
where the coefﬁcients a
ν
are given by
a
ν
=
1

2

∂B
G
ν
(q) Dq f(q) (9.8)
= (−1)
n

ν
f(0). (9.9)
Proof Let S be a sphere with centre 0 lying inside the domain of regularity of f, B a
closed ball with centre 0 lying inside S. Then for q ∈ B the integral formula gives
f(q) =
1

2

S
G(q

−q)Dq

f(q

)
=
1

2

S

¸
n=0
¸
ν∈σn
P
ν
(q)G
ν
(q

)Dq

f(q

)
since [q[ < [q

[ if q ∈ B and q

∈ S. The series converges uniformly on B S, so it
can be integrated term by term to give a uniformly convergent series of the form (9.7)
with
a
ν
=
1

2

S
G
ν
(q)Dqf(q).
But the functions G
ν
are right-regular except at 0, so (4.4) gives d(G
ν
Dqf) = 0
outside B, and therefore Stokes’s theorem can be used to replace the contour S by ∂B.
Differentiating the integral formula gives

ν
f(q) =
(−1)
n

2

S
G
ν
(q

−q)Dq

f(q

)
Hence ∂
ν
f(0) = (−1)
n
a
ν
. 2
Corollary
1

2

S
G
µ
(q) Dq P
ν
(q) = δ
µν
,
where S is any sphere containing the origin.
This follows by putting f = P
µ
in (9.7).
Theorem 30 (the Laurent series) Suppose f is regular in an open set U except pos-
sibly at q
0
∈ U. Then there is a neighbourhood N of q
0
such that if q ∈ N and q = q
0
,
f(q) can be represented by a series
f(q) =

¸
n=0
¸
ν∈σn
¦P
ν
(q −q
0
)a
ν
+ G
ν
(q −q
0
)b
ν
¦
which converges uniformly in any hollow ball
¦q : r ≤ [q −q
0
[ ≤ R¦ , with r > 0, which lies inside N.
The coefﬁcients a
ν
and b
ν
are given by
a
ν
=
1

2

C
G
ν
(q −q
0
) Dq f(q), (9.11)
b
ν
=
1

2

C
P
ν
(q −q
0
) Dq f(q), (9.12)
42
where C is any closed 3-chain in U`¦q
0
¦ which is homologous to ∂B for some ball B
with q
0
∈ B ⊂ U (so that C has wrapping number 1 about q
0
).
Proof Choose R
1
so that the closed ball B
1
= ¦q : [q − q
0
[ ≤ R
1
¦ lies inside U,
and let N = int B
1
, S
1
= ∂B
1
. Given q ∈ N`¦q
0
¦, choose R
2
so that 0 < R
2
<
[q − q
0
[ < R
1
, and let S
2
be the sphere ¦q : [q − q
0
[ = R
2
¦. Then by the integral
formula,
f(q) =
1

2

S1
G(q

−q)Dq

f(q

) −
1

2

S2
G(q

−q)Dq

f(q

). (9.13)
For q

∈ S
1
we have [q

−q
0
[ < [q −q
0
[, and so as in theorem 29
1

2

S1
G(q

−q)Dq

f(q

) =

¸
n=0
¸
ν∈σn
P
ν
(q −q
0
)a
ν
(9.14)
where
a
ν
=
1

2

S1
G
ν
(q

−q
0
)Dq

f(q

)
and the series is uniformly convergent in any ball [q−q
0
[ ≤ Rwith R < R
1
. If C is a 3-
chain as in the statement of the theorem, it is homologous to S
1
; since d(G
ν
Dqf) = 0
in U`¦q
0
¦, Stokes’s theorem gives (9.11).
For q

∈ S
2
we have [q

− q
0
[ < [q − q
0
[ and so we can expand G(q

− q) as in
(9.4) to obtain

1

2

S2
G(q

−q)Dq

f(q

) =
1

2

S2
¸
n=0
¸
ν∈σn
G
ν
(q −q
0
)P
ν
(q

−q
0
)Dq

f(q

)
=

¸
n=0
¸
ν∈σn
G
ν
(q −q
0
)b
ν
(9.15)
where
b
ν
=
1

2

S2
P
ν
(q

−q
0
)Dq

f(q

).
The series is uniformly convergent in any region [q − q
0
[ ≥ r with r > R
2
. Since the
P
ν
are right-regular, d(P
ν
Dqf) = 0 in U`¦q
0
¦ and so the contour S
2
in the formula
for b
ν
can be replaced by any 3-chain C as in the statement of the theorem. In par-
ticular, b
ν
is independent of the choice of R
2
and so the same representation is valid
for all q ∈ N`¦q
0
¦. Now putting (9.14) and (9.15) into (9.13) gives (9.10) for all
q ∈ N`¦q
0
¦. 2
ACKNOWLEDGEMENTS
I am grateful for the hospitality of the Department of Applied Mathematics and Theo-
retical Physics in the University of Liverpool, where part of this work was done, and to
Drs P J McCarthy and C J S Clarke for many stimulating discussions.
43
References
[1] Hamilton, W R: Elements of Quaternions. London: Longmans Green 1866.
[2] Tait, P G: An Elementary Treatise on Quaternions. Cambridge: C.U.P. 1867.
[3] Joly, C J: A Manual of Quaternions. London: Macmillan 1905.
[4] Fueter, R: “Die Funktionentheorie der Differentialgleichungen ∆u = 0 und
∆∆u = 0 mit vier reellen Variablen.” Comment. math. Helv. 7, 307-330 (1935).
[5] Fueter, R: “
¨
Uber die analytische Darstellung der regul¨ aren Funktionen einer
Quaternionenvariablen.” Comment. math. Helv. 8, 371-378 (1936).
[6] Haefeli, H: “Hyperkomplexe Differentiale.” Comment. math. Helv. 20, 382-420
(1947).
[7] Deavours, C A: “The Quaternion Calculus.” Amer. Math. Monthly 80, 995-1008
(1973).
[8] Schuler, B: “Zur Theorie der regul¨ aren Funktionen einer Quaternionen-Variablen.”
Comment. math. Helv. 10, 327-342 (1937).
[9] Fueter, R: “Die Singularit¨ aten der eindeutigen regul¨ aren Funktionen einer Quater-
nionenvariablen.” Comment. math. Helv. 9, 320-335 (1937).
[10] Porteous, I R: Topological Geometry. London: Van Nostrand Reinhold 1969.
[11] Cartan, H: Elementary Theory of Analytic Functions of One or Several Complex
Variables. London: Addison-Wesley 1963.
[12] Courant, R and Hilbert, D: Methods of Mathematical Physics Vol II. London:
Interscience 1962.
[13] Heins, M: Complex Function Theory. London: Academic Press 1968.
[14] Titchmarsh, E C: The Theory of Functions, 2nd ed. Oxford: O.U.P. 1939.
[15] Cullen, CG: “An integral theoremfor analytic intrinsic functions on quaternions.”
Duke Math. J. 32, 139-148 (1965).
[16] Schouten, J A: Ricci-Calculus, 2nd ed. Berlin: Springer-Verlag 1954.
[17] Sugiura, M: Unitary Representations and Harmonic Analysis. London: Wiley
1975.
[18] McCarthy, P J and Sudbery, A: “Harmonic analysis of generalised vector func-
tions, generalised spin-weighted functions and induced representations.” J. Phys.
A 10, 331-338 (1977).
44

1 Introduction
The richness of the theory of functions over the complex ﬁeld makes it natural to look for a similar theory for the only other non-trivial real associative division algebra, namely the quaternions. Such a theory exists and is quite far-reaching, yet it seems to be little known. It was not developed until nearly a century after Hamilton’s discovery of quaternions. Hamilton himself [1] and his principal followers and expositors, Tait [2] and Joly [3], only developed the theory of functions of a quaternion variable as far as it could be taken by the general methods of the theory of functions of several real variables (the basic ideas of which appeared in their modern form for the ﬁrst time in Hamilton’s work on quaternions). They did not delimit a special class of regular functions among quaternion-valued functions of a quaternion variable, analogous to the regular functions of a complex variable. This may have been because neither of the two fundamental deﬁnitions of a regular function of a complex variable has interesting consequences when adapted to quaternions; one is too restrictive, the other not restrictive enough. The functions of a quaternion variable which have quaternionic derivatives, in the obvious sense, are just the constant and linear functions (and not all of them); the functions which can be represented by quaternionic power series are just those which can be represented by power series in four real variables. In 1935 R Fueter [4] proposed a deﬁnition of “regular” for quaternionic functions by means of an analogue of the Cauchy-Riemann equations. He showed that this deﬁnition led to close analogues of Cauchy’s theorem, Cauchy’s integral formula, and the Laurent expansion [5]. In the next twelve years Fueter and his collaborators developed the theory of quaternionic analysis. A complete bibliography of this work is contained in ref. [6], and a simple account (in English) of the elementary parts of the theory has been given by Deavours [7]. The theory developed by Fueter and his school is incomplete in some ways, and many of their theorems are neither so general nor so rigorously proved as present-day standards of exposition in complex analysis would require. The purpose of this paper is to give a self-contained account of the main line of quaternionic analysis which remedies these deﬁciencies, as well as adding a certain number of new results. By using the exterior differential calculus we are able to give new and simple proofs of most of the main theorems and to clarify the relationship between quaternionic analysis and complex analysis. Let H denote the algebra of quaternions, and let {1, i, j, k} be an orthonormal basis, with the product on H given by the usual multiplication table (see section 2). The typical quaternion can be written as q = t + ix + jy + kz (1.1)

where t, x, y, z are real coordinates. Fueter deﬁned a function f : H → H to be regular if it satisﬁed the equation ∂f ∂f ∂f ∂f +i +j +k = 0. ∂t ∂x ∂y ∂z (1.2)

2

This is an analogue of the Cauchy-Riemann equations for a complex function f : C → C, which can be written as ∂f ∂f +i = 0, (1.3) ∂x ∂y where the variable is z = x + iy. Fueter showed that any quaternionic function which is regular and also continuously differentiable must satisfy an analogue of Cauchy’s theorem which can be written as Dq f = 0, (1.4)
C

where C is any smooth closed 3-manifold in H and Dq is a certain natural quaternionvalued 3-form. Dq is deﬁned in section 2; it can be thought of as the quaternion representing an element δC of the 3-manifold, its magnitude being equal to the volume of δC and its direction being normal to δC. Fueter also obtained an analogue of Cauchy’s integral formula for such functions, namely f (q0 ) = 1 2π 2 (q − q0 )−1 Dq f (q), |q − q0 |2 (1.5)

∂D

where D is a domain in H in which f is regular and q0 is a point inside D. The complex Cauchy-Riemann equation (1.3) is equivalent to the statement that f has a complex derivative, i.e. that there exists a complex number f (z) such that df = f (z) dz. (1.6)

Fueter gave no corresponding characterisation of a regular function of a quaternion variable, leaving the theorems (1.4) and (1.5) and the analogy with the Cauchy-Riemann equations as sufﬁcient justiﬁcation of the deﬁnition (1.2). In this paper we will show that a regular function can be deﬁned as one which has a certain kind of quaternionic derivative; speciﬁcally, (1.2) is equivalent to the existence of a quaternion f (q) such that d(dq ∧ dq f ) = Dq f (q). (1.7) (the 2-form dq ∧ dq is described in section 2). Cauchy’s theorem (1.4) and the integral formula (1.5) can be simply proved by showing that d(Dqf ) = 0 (1.8) and d (q − q0 )−1 Dq f (q) = ∆ |q − q0 |2 1 |q − q0 |2 f (q)v (1.9)

where ∆ is the Laplacian on R4 and v = dt ∧ dx ∧ dy ∧ dz is the standard volume 4-form. Since |q − q0 |−2 is the Green’s function for the Laplacian in R4 , (1.5) follows from (1.9). This is essentially how Fueter proved these theorems. Since both proofs use Stokes’s theorem, they need the condition that the partial derivatives of the function should be continuous. Schuler [8] showed that this condition could be dropped

3

by adapting Goursat’s proof of Cauchy’s theorem, but he did not draw the full consequences of this argument. In fact Cauchy’s theorem (1.4) can be proved for any rectiﬁable contour C and any function f which is differentiable at each point inside C and whose partial derivatives satisfy (1.2). The integral formula (1.5) has a similarly wide range of validity. From this it follows, as in complex analysis, that if f is regular in an open set U then it has a power series expansion about each point of U . Thus pointwise differentiability, together with the four real conditions (1.2) on the sixteen partial derivatives of f , is sufﬁcient to ensure analyticity. The homogeneous components in the power series representing a regular function are themselves regular; thus it is important to study regular homogeneous polynomials, the basic regular functions from which all regular functions are constructed. The corresponding functions of a complex variable are just the powers of the variable, but the situation with quaternions is more complicated. The set of homogeneous regular func1 tions of degree n forms a quaternionic vector space of dimension 2 (n + 1)(n + 2); this is true for any integer n if for negative n it is understood that the functions are deﬁned and regular everywhere except at 0. The functions with negative degree of homogeneity correspond to negative powers of a complex variable, and occur in the quaternionic Laurent series which exists for any regular function which is regular in an open set except at one point. Fueter found two natural bases for the set of homogeneous functions, which play dual roles in the calculus of residues. (He actually only proved that these bases formed spanning sets). In this paper we will study homogeneous regular functions by means of harmonic analysis on the unit sphere in H, which forms a group isomorphic to SU(2); this bears the same relation to quaternionic analysis as the theory of Fourier series does to complex analysis. Many of the algebraic and geometric properties of complex analytic functions are not present in quaternionic analysis. Because quaternions do not commute, regular functions of a quaternion variable cannot be multiplied or composed to give further regular functions. Because the quaternions are four-dimensional, there is no counterpart to the geometrical description of complex analytic functions as conformal mappings. The zeros of a quaternionic regular function are not necessarily isolated, and its range is not necessarily open; neither of these sets need even be a submanifold of H. There is a corresponding complexity in the structure of the singularities of a quaternionic regular function; this was described by Fueter [9], but without giving precise statements or proofs. This topic is not investigated here. The organisation of this paper is as follows. In section 2 the basic algebraic facts about quaternions are surveyed and notation is established; some special algebraic concepts are introduced, and quaternionic differential forms are described. Section 3 is concerned with the deﬁnition of a regular function. The remarks in the second paragraph of this introduction, about possible analogues of complex deﬁnitions of analyticity, are ampliﬁed (this material seems to be widely known, but is not easily accessible in the literature), and the deﬁnition (1.7) of regular functions by means of the quaternionic derivative is shown to be equivalent to Fueter’s deﬁnition (1.2) by means of a Cauchy-Riemann-type equation. Section 4 is devoted to the analogues of the Cauchy-Goursat theorem and Cauchy’s integral formula. 4

(Conversely. ki = −ik = j (2. In section 9 we examine the power series representing a regular function and prove analogues of Laurent’s Theorem and the residue theorem. k} for H so that the multiplication is given by ij = −ji = k. jk = −kj = i. i. the maximum-modulus principle. b0 ∈ R. 2 Preliminaries 2.3) In performing calculations it is sometimes useful to denote the basic quaternions i. 3) and the coordinates x. a0 b + b0 a + a × b) (2. the subspace P . a. 5 . j. where P is an oriented three-dimensional Euclidean vector space. In section 6 we show how regular functions can be constructed from functions of more familiar type. In this notation eq. i2 = j 2 = k 2 = −1. b) = (a0 b0 − a. (2. Then we can write H = R ⊕ P .) Thus we can choose a basis {1. a. 3). with an identity element denoted by 1. only the last of these requires proof. its inner product and its orientation can be deﬁned in terms of the multiplication on H. Section 8 is an investigation of homogeneous regular functions by means of harmonic analysis on S 3 . a)(b0 . chap. j.5) of the assertions in this section can be found in [10.b. Section 7 is concerned with the effect on regular functions of conformal transformation of the variable.b denotes the inner product of a and b.3) becomes q = t + x i ei and the multiplication rules (2. 10]. x. We regard R as being embedded in H by identifying t ∈ R with t1 ∈ H. and Morera’s theorem. and to use the summation convention for repeated indices.2) The typical quaternion will be denoted by q = t + xi + yj + zk (t. 2.1 The Algebra of Quaternions1 The quaternions H form a four-dimensional algebra over the real ﬁeld R. z by xi (i = 1.1) where a0 . and the product of two quaternions is deﬁned by (a0 . b ∈ P .4) ijk ek (2. After the work of section 4. k by ei (i = 1.Section 5 contains analogues of Liouvilles’s theorem. y. z ∈ R) (2. 2.2) become ei ej = −δij + 1 Proofs (2. namely harmonic functions of four real variables and analytic functions of a complex variable. these results are needed in the last two sections. and a × b denotes the vector product determined by the orientation on P . y.

we will take C to be the subﬁeld spanned by 1 and i. then v = t + ix and w = y − iz (2. which is isomorphic to the complex numbers C if 1 and q are linearly independent.14) and its pure quaternion part (or vector part) is Pu q = xi + yj + zk = It follows from (2.9) (2. the vector subspace spanned by 1 and q is a subﬁeld of H. ¯ ¯ (2.13) (2. ¯ q1 q2 = q 2 q 1 . Then any quaternion can be written as q = v + jw with v. i.6) Conjugation is an involutive anti-automorphism of H.15) . v2 . q commutes with q . we can deduce that v1 = v2 and w1 = w2 . w2 ∈ C. and their product is ¯ q q = t 2 + x2 + y 2 + z 2 . ¯ 6 (2. whenever we want to do this. It is sometimes convenient to distinguish a particular embedding of C in H. if q is given by (2.16) 1 2 (q − q ) ∈ P. ¯ The modulus of q is the non-negative real number √ |q| = q q ¯ q ¯ .3).12) It follows from (2. v (2. and (2.9) that Re q = Re q.11) (2. The centre of the algebra H is the real subﬁeld R.1) that Re(q1 q2 ) = Re(q2 q1 ) and from (2. |q|2 (2. the basic law of multiplication is vj = j¯.where ijk is the usual alternating symbol.e. w ∈ C. q = q.8) The conjugate of the quaternion q is q = t − ix − jy − kz. w1 .11) that every non-zero quaternion has a multiplicative inverse q −1 = The real part of q is 1 Re q = t = 2 (q + q ) ∈ R ¯ (2. ¯ (2. If q is any quaternion. When quaternions are written in this form.7) From an equation of the form v1 + jw1 = v2 + jw2 with v1 . it is R-linear.17) (2.10) For every embedding of C in H. quaternion conjugation coincides with complex conjugation.

19) . q2 = q1 .25) for any α ∈ F1 . the adjoint of left (right) multiplication by a is left (right) multiplication by a. respectively. constituting the sphere S 3 . in fact they are obtained by slight modiﬁcations of standard procedures from the tensor of type (1. q2 = q1 . we will denote it by S. ¯ T q = |q|. This inner product induces an R-linear map Γ : H∗ → H.17) q Note that aq1 .16) and (2. q ∈ H.18) |q| Any quaternion has a polar decomposition q = ru where r = |q| ∈ R and u = Un q ∈ S. Since {1. for ¯ any a ∈ H. q = α(q) (2. then 7 (2. which we will denote by F1 .21) i. The inner product (2. Γ : F1 = HomR (H. we have Γ(α) = α(1) + i α(i) + j α(j) + k α(k). The positive-deﬁnite quadratic form (2. so the map Γ can be extended by linearity to a right H-linear map Γr : F1 → H and a left-linear map Γ : F1 → H. q2 a . Let Lq : H → H and Rq : H → H be the operations of multiplication by q on the left and on the right.q is a unit quaternion if |q| = 1. conversely.19) and the maps Γr . ¯ (2.24) (2.20) (2.11) gives rise to an inner product q1 . given by Γ(α). The set of unit quaternions. If u1 and u2 are unit quaternions. (2. The classical notation for these functions of q is S q = Re q. the map q → u1 qu2 is orthogonal with respect to this inner product and has determinant 1.22) for a ∈ H∗ . It is spanned (over H) by H∗ . and let K : H → H be the operation of conjugation. aq2 ¯ and q1 a. K q = q . R) is the dual of H. U q = Un q. any rotation of H is of the form q → u1 qu2 for some u1 . forms a multiplicative Lie group isomorphic to SU(2).e. i. V q = Pu q. The versor of a quaternion q is the unit quaternion q Un q = (2. R) → H have alternative characterisations in terms of quaternion multiplication and conjugation. k} is an orthonormal basis for H. j. q2 = Re(q1 q2 ) ¯ = Re(¯1 q2 ) by (2.2) which deﬁnes the multiplication on H.23) The set of R-linear maps of H into Itself forms a two-sided vector space over H of dimension 4. u2 ∈ S. where H∗ = HomR (H. This is the well-known double cover 0 → Z2 → SU(2) × SU(2) → SO(4) → 0. They are given by Γr (α) = α(1) + i α(i) + j α(j) + k α(k) and Γ (α) = α(1) + α(i) i + α(j) j + α(k) k (2.

q (iii) Tr(Lq αK) = Re q α(1) + ¯ ¯ i ei q α(ei ) ¯ q α(ei )ei ¯ i = Re q α(1) + ¯ = Re [¯Γ (α)] q = Γ (α). ¯ Proof The trace of an R-linear map α : H → H is Tr α = Re [α(1)] − Hence (i) Tr(Lq Rp K) = Re qp + i Re [ei α(ei )] i (2. q = Tr(Rq αK) ¯ (iii) Γ (α). q = Tr(Lq α). q = Tr(Rq α). q ∈ H and α ∈ F1 . q (ii) Tr(Rq αK) = Re α(1)¯ + q ¯ i 2 ei α(ei )¯ q by (2.16) by (2.29) α(ei )ei i 8 .27) as can easily be veriﬁed.Theorem 1 1 (i) q.24) 2 = Re [Γr (α)¯] q = Γr (α). q by (2. p .26) ei qei p But q+ i ei qei = −2¯ q (2. Γ : F1 → H. so Tr(Lq Rp K) = −2 Re(¯p) = −2 q. ¯ Γr (α) = α(1) − ¯ Γ (α) = α(1) − ei α(ei ) i (2.25) 2 ¯ ¯ We will also need two other maps Γr . p = − 2 Tr(Lq Rp K) (ii) Γr (α). q = Tr(Lq αK) for all p.28) (2. deﬁned as follows: ¯ Γr (α). ¯ ¯ Γ (α).

The differential of the identity function is dq = dt + i dx + j dy + k dz. . (2. we will say that a function f : H → H is real-differentiable if it is differentiable in the usual sense.. The following special differential forms will be much used in the rest of the paper. and Stokes’s theorem holds in the usual form for quaternionic integrals. we can regard the differential as a quaternion-valued 1-form df = ∂f ∂f ∂f ∂f dt + dx + dy + dz. . Its exterior product with itself is dq ∧ dq = 1 2 ijk ei dxj ∧ dxk = i dy ∧ dz + j dz ∧ dx + k dx ∧ dy 9 (2.2 Quaternionic Differential Forms When it is necessary to avoid confusion with other senses of differentiability which we will consider. dx. Then the set of all r-forms is a two-sided quaternionic vector space. ai ∈ H) can be regarded as the R-linear map θ : H → H given by θ(t + xi ei ) = a0 t + ai xi (2..30) Conversely. a quaternion-valued r-form can be regarded as a mapping from H to the space of alternating R-multilinear maps from H × .. hr+s ) = 1 r!s! ρ (ρ)θ(hρ(1) . which is not commutative. Its differential at a point q ∈ H is then an R-linear map dfq : H → H. ∂t ∂x ∂y ∂z (2. and (ρ) is the sign of ρ. We deﬁne the exterior product of such forms in the usual way: if θ is an r-form and φ is an s-form... The space of quaternionic r-forms has a basis of real r-forms. (2. The exterior derivative of a quaternionic differential form is deﬁned by the usual recursive formulae.33)   (θa) ∧ φ = θ ∧ (aφ) (2..35) . for such forms left and right multiplication by quaternions coincide. .. it is not in general true that θ ∧ φ = −φ ∧ θ for quaternionic 1-forms θ and φ.31) Similarly. hρ(r+s) ). where the sum is over all permutations ρ of r + s objects. and we have  a(θ ∧ φ) = (aθ) ∧ φ. consisting of exterior products of the real 1-forms dt... hρ(r) )φ(hρ(r+1) . dz.32) for all quaternions a.2... dy. × H (r times) to H. Note that because the exterior product is deﬁned in terms of quaternion multiplication. By identifying the tangent space at each point of H with H itself. any quaternion-valued 1-form θ = a0 dt + ai dxi (a0 . r-forms θ and s-forms φ. θ ∧ φ(h1 .34) regarded as R-linear transformation of H.   (θ ∧ φ)a = θ ∧ (φa). dq is the identity mapping.

(2. ej ) = − ijk ek 1 = 2 (ej ei − ei ej ).39) Dq(1.41) (2. gives the commutator of its arguments. c) = 1 |a|2 a(a−1 ca−1 b − a−1 ba−1 c) 2 = 1 (c¯b − b¯c) a 2 a by (2. as antisymmetric function on H × H. h3 . the 3-form Dq is deﬁned as an alternating R-trilinear function by h1 . h2 ) = 1 (h2 h1 − h1 h2 ) 2 for all h1 . h2 . hence Dq(ua. c) is a quaternion which is perpendicular to a. k) = 1. i. Dq(h2 . the map q → uq is an orthogonal transformation of H with determinant 1. h1 . h4 ) = v(h1 . ei . to Dq(ah1 b.which. ei . we obtain Dq(a. Taking u = |a|−1 .. a−1 c). Thus Dq(i. b. j. h4 ) (2. Now since Dq(1. we have by linearity (2. (2. ej ) = − coordinate expression for Dq is ijk ek . Finally. b and c.40) Two useful formulae were obtained in the course of this proof.37) for all h1 .42) 10 . uc) = u Dq(a. h2 . . ah2 b. It also has the following algebraic expression: 1 Theorem 2 Dq(a. using the fact that the map q → uqv is a rotation for any pair of unit quaternions u.36) so that v(1. h3 )b..39) can be generalised.. (2. c) = |a|2 a Dq(1. c). ah3 b) = |a|2 |b|2 a Dq(h1 . Hence Dq(a. b. b. For the (essentially unique) constant real 4-form we use the abbreviation v = dt ∧ dx ∧ dy ∧ dz. h2 ∈ H. b. v.40) can be written as 1 1 Dq = − 2 dq ∧ dq. 2 (2. c) = 2 (c¯b − b¯c) a a Proof For any unit quaternion u. j. ub. a−1 b. and using the R-trilinearity of Dq.38) Geometrically. h3 . k) = 1 and Dq(1. b. b and c and has magnitude equal to the volume of the 3-dimensional parallelepiped whose edges are a. and the formula (2. The Dq = dx ∧ dy ∧ dz − 1 ijk ei dt ∧ dxj ∧ dxk 2 = dx ∧ dy ∧ dz − i dt ∧ dy ∧ dz − j dt ∧ dz ∧ dx − k dt ∧ dx ∧ dy. The argument leading to (2. h4 ∈ H.13). Dq(a.

.. the map Γr can be applied to it.44) ∂ + ei . 2  2 ∂t ∂xi     ∂f 1 ∂f  ¯  ∂r f = 1 Γ (df ) = − ei .qr−1 qar (3. The result is Γr (df ) = ∂f ∂f ∂f ∂f +i +j +k .. y.1) for some non-negative integer r (the degree of the monomial) and constant quaternions a0 . x. z.45) 3 Regular Functions We start by showing that the concept of an analytic function of a quaternion variable as one which is constructed from the variable by quaternion addition and multiplication (possibly involving inﬁnite series) is the same as the concept of a real-analytic function in the four real variables t.where denotes the usual inner product between differential forms and vector ﬁelds and 1 denotes the constant vector ﬁeld whose value is 1. q) where F : H × .43) We introduce the following notation for the differential operator occurring in (2. ar . ∂ . A homogeneous polynomial function of degree r on H is a function f : H → H of the form f (q) = F (q.  2 ∂t ∂x ∂y ∂z ¯ ¯ Note that ∂ . . . ∂t ∂x ∂y ∂z (2. ∂r and ∂r all commute. Deﬁnition 1 A quaternionic monomial is a function f : H → H of the form f (q) = a0 qa1 q. 11 . Since the differential of a quaternion-valued function on H is an element of F 1 . ∂ f = 2 Γr (df ) = + ei   2 ∂t ∂xi      ∂f 1 ∂f  1¯ ∂ f = 2 Γr (df ) = . and for other related differential operators:  ∂f 1 ∂f 1 ¯  . A quaternionic polynomial is a ﬁnite sum of quaternionic monomials... A polynomial function on H is a ﬁnite sum of homogeneous polynomial functions of varying degrees... and that ¯ ¯ ∆ = 4∂r ∂r = 4∂ ∂ (2.. − ei   2 ∂t ∂xi     1 ∂f ∂f ¯r f = 1 Γ (df ) = (2.   2  2 ∂t ∂xi     2 2 2 2  ∂ f ∂ f ∂ f ∂ f  ∆f = 2 + + 2 + 2.40).. × H (r times) → H is R-multilinear..

a series of quaternionic monomials. f has the form f (q) = a + qb for some a. in which we concentrate on the existence of a quaternionic derivative deﬁned as the limit of a difference quotient. 2 It is clear that conversely. fi . We will show that only quaternionic polynomials of degree 1 (and not all of them) possess such a derivative. We now turn to an alternative attempt to parallel complex analysis. i. y. it is real-differentiable at q and its differential is the linear map of multiplication on df the right by dq : df dfq (h) = h . z. x. dq where f0 and fi are four real-valued polynomials in the four real variables t. y. so f is a quaternionic polynomial.e.      1 z = 4k (q + iqi + jqj − kqk). Proof From the deﬁnition it follows that if f is quaternion-differentiable on the left at q. x. Deﬁnition 2 A function f : H → H is quaternion-differentiable on the left at q if the limit df = lim h−1 {f (q + h) − f (q)} dq h→0 exists.  (3. .Theorem 3 Every polynomial function on H is a quaternionic polynomial. x. Proof Any polynomial function f can be written as f (q) = f0 (t.e. dq i. z into the polynomials f0 . But  1 t = 4 (q − iqi − jqj − kqk). x.2) 1 y = 4j (q + iqi − jqj + kqk). dfq = dq 12 df . Then on U.     1 x = 4i (q − iqi + jqj + kqk). z) + i fi (t. we obtain f (q) as a sum of expressions in q of the form (3. is precisely the same as the class of functions which are realanalytic in a neighbourhood of the origin. every quaternionic polynomial is a polynomial function. y. so we have Corollary The class of functions which are deﬁned in a neighbourhood of the origin in H and can be represented there by a quaternionic power series.1). Theorem 4 Suppose the function f is deﬁned and quaternion-differentiable on the left throughout a connected open set U . z)ei Putting these expressions for t. b ∈ H. y.

f = g + jh = α + j + (v + jw)(β − jγ) = a + qb. then (3.Equating coefﬁcients of dt. ∂¯ v ∂w ¯ ∂v ∂w ∂g ∂h = . dq ∂t ∂x ∂y ∂z (3.3) Put q = v + jw. where v = t + ix and w = y − iz. 133] g and h have ¯ continuous partial derivatives of all orders and so from (3.5) ∂2g ∂ = ∂v 2 ∂v ∂h ∂w = ∂ ∂w ∂h ∂v = 0.5) and (3. where a = α+j and b = β −jγ. comparing the forms of the function 13 .6) give the following relations among these constants: β = η.5) ∂g ∂h =− . any two of which can be connected by a chain of convex sets which overlap in pairs. Thus ¯ g(v. w) = + ζ v + ηw + θ¯w. and h is a complex¯ analytic function of v and w. where g and h are complex-valued functions of the two complex variables v and w. The general connected open set can be covered by convex sets. (3. δ = θ = 0. w) + jh(v. Then we can deduce that g is linear in w.4) (3. dx. p. (3.3) can be separated into the two sets of complex equations ∂g ∂g ∂h ∂h = −i = =i .6) ∂¯ v ∂w ¯ Eq. ¯ v where the Greek letters represent complex constants. so f is of the stated form if U is convex. ∂t ∂x ∂y ∂z In terms of complex derivatives. ∂t ∂x ∂y ∂z ∂h ∂g ∂g ∂h =i =− =i . w) = α + βv + γ w + δv w. and let f (q) = g(v. Now (3. h ¯ is linear in w and h is linear in v . Thus ζ = −γ. dy and dz gives ∂f ∂f ∂f ∂f df = = −i = −j = −k . ¯ ¯ h(v. Hence by Hartogs’s theorem [11. these can be written as ∂g ∂h ∂h ∂g = = = = 0. w). ∂v ∂w and (3.4) shows that g is a complex-analytic function of v and w. Suppose for the moment that U is convex.

7). 2 Even if f is quaternion-differentiable.10) So U. ∂f ∂x 14 .7) where the integral is round a closed curve. ∂f ∂f j= i. b throughout U. we see that f (q) = a + qb with the same constants a. ∂x ∂y ∂y ∂x ∂f ∂f − ei ∂xi ∂t ∂f ∂f dt + dx ∂t ∂xj ∂f dt ∧ dxi + ei dxi ∧ dxj ∂xj ∂f ∂f = ei ∂xi ∂t (3. in fact the only functions satisfying this equation for all closed curves are the constant functions. We will prove this for the inﬁnitesimal form of (3. it will not in general satisfy Cauchy’s theorem in the form dqf = 0 (3.f on the overlaps. namely d(dqf ) = 0. (3. It follows that f is constant in ∂t ∂y ∂z 2 We will now give a deﬁnition of “regular” for a quaternionic function which is satisﬁed by a large class of functions and opens the door to a development similar to the theory of regular functions of a complex variable. = 0 and therefore ∂f = ∂f = ∂f = 0 in U .9) we have.8) Theorem 5 If the function f : U → H is real-differentiable in the connected open set U and satisﬁes d(dqf ) = 0 in U . ∂f ∂f i= j ∂x ∂y while from (3. ∂x ∂y Hence ∂f ∂f ∂f ∂f = − ji = ij = − .9) ∂f ∂f = ej ∂xj ∂xi (3. then f is constant on U . Proof d(dqf ) = dq ∧ df = (dt + ei dxi ) ∧ = Thus d(dqf ) = 0 ⇒ and ei From (3.10). for example.

For deﬁniteness.Deﬁnition 3 A function f : H → H is left-regular at q ∈ H if it is real-differentiable at q and there exists a quaternion f (q) such that d(dq ∧ dq f ) = −2Dq f (q). Clearly. we will only consider left-regular functions. we see that Γr (dfq ) = 0. Then from (3. It is right-regular if there exists a quaternion fr (q) such that d(f dq ∧ dq) = −2fr (q) Dq. which we will call simply regular. Theorem 6 (the Cauchy-Riemann-Fueter equations) A real-differentiable function f is regular at q if and only if Γr (dfq ) = 0. If we write q = v + jw. 2 Note that (3.24). the theory of left-regular functions will be entirely equivalent to the theory of right-regular functions. dq ∧ dq ∧ dfq = −2Dqf (q) Evaluating these trilinear functions with i. (3.12) (3. Conversely. i. w) + jh(v.11) ∂f .14) becomes the pair of complex equations ∂h ∂g = .16) .15) ∂t Hence f = ∂ f . k as arguments gives (ij − ji) dfq (k) + (jk − kj) dfq (i) + (ki − ik) dfq (j) = −2f (q) while using 1. (3. Hence f (q) = dfq (1) = − {i dfq (i) + j dfq (j) + k dfq (k)} (3.14) (3.12) and (3. ∂¯ v ∂w ¯ 15 ∂g ∂h =− ∂w ∂v (3. We will write f (q) = f (q) and call it the derivative of f at q. f (q) = g(v. w) as in theorem 4. Proof Suppose f is regular at q. if Γr (dfq ) = 0 we can deﬁne f (q) = dfq (1) and then evaluating as above shows that dq ∧ dq ∧ dfq = 2Dqf (q).13) Comparing with (2.11). j as arguments gives (ij − ji) dfq (1) = 2kf (q). j.13) can be written as ¯ 2∂ f = and ∂f ∂f ∂f ∂f +i +j +k =0 ∂t ∂x ∂y ∂z f (q) = (3. so that f is regular at q.

Now since dfq is a continuous function of q at q0 . .. The derivative of a regular function can be characterised as the limit of a difference quotient which is analogous to that used to deﬁne the derivative of a complex-analytic function. h3 ∈ H. h2 . then ∆f = 0.. hence we can choose δ so that |q − q0 | < δ ⇒|dq ∧ dq ∧ dfq (h1 . h2 .. f is harmonic. From (3. h2 . hk ∈ H are called its edge-vectors. + tk hk . Then by Stokes’s theorem. h2 . with edge-vectors h1 . so is dq ∧ dq ∧ dfq .. so all regular functions are harmonic. then −1 Dq C ∂C dq ∧ dqf + 2f (q0 ) < . Deﬁnition 4 An oriented k-parallelepiped in H is a map C : I k → H. A k-parallelepiped is non-degenerate if its edge-vectors are linearly independent (over R). h3 )| < |Dq(h1 . Then given > 0. h3 )|. Let C be as in the statement of the theorem.. Theorem 7 Suppose that f is regular at q0 and continuously differentiable in a neighbourhood of q0 . h4 ) < 0.14) and (2. q0 ∈ H is called the original vertex of the parallelepiped. tk ) = q0 + t1 h1 + .. h3 )| is a norm on the real vector space H ∧ H ∧ H.. . h3 ) dt1 dt2 dt3 Since f is regular at q0 . 1 ∂C dq ∧ dqf = C d(dq ∧ dqf ) = 0 dq ∧ dq ∧ dfC(t) (h1 . and h1 ... Proof First note that ||h1 ∧ h2 ∧ h3 || = |Dq(h1 . h3 . h2 . h3 ) dt1 dt2 dt3 16 . i. A non-degenerate 4-parallelepiped is positively oriented if v(h1 . of the form C(t1 . h3 ) − dq ∧ dq ∧ dfq0 (h1 .45) it follows that if f is regular and twice differentiable. 1 2 C Dqf (q0 ) = 2 0 Dq(h1 . h4 ) > 0. . We will see in the next section that a regular function is necessarily inﬁnitely differentiable. there exists δ > 0 such that if C is a non-degenerate oriented 3-parallelepiped with q0 ∈ C(I 3 ) and q ∈ C(I 3 ) ⇒ |q − q0 | < δ. negatively oriented if v(h1 .. ..e. h2 . where I k ⊂ Rk is the closed unit k-cube... h3 ) f (q0 ) dt1 dt2 dt3 1 =− 0 dq ∧ dq ∧ dfq0 (h1 .which show some similarity to the Cauchy-Riemann equations for a function of a complex variable. h2 . h2 .. ∀h1 . h2 .

holds for a general rectiﬁable boundary. t2 . and thus constitutes an explicit solution to the general Dirichlet problem. h3 ) − dq ∧ dq ∧ dfq0 (h1 .h3 →0 lim [Dq(h1 . which is considerably wider than that of the integral theorems for harmonic functions. The analogy between theorem 7 and the difference-quotient deﬁnition of regular complex functions can be made explicit by stating the latter as follows: If f is regular at z0 . 54] to deﬁne f for a function satisfying ∂f ∂t = 0. The limit (3. and immediately obtain the integral formula for a parallelepiped. art. h3 are multiples of three ﬁxed linearly independent quaternions.17) This is valid if it is understood that h1 . h2 . t3 → 0. To make explicit the analogy between the deﬁnition of regular quaternionic functions and the deﬁnition of regular complex functions by means of the Cauchy-Riemann equations. h3 )| dt1 dt2 dt3 < |Dq(h1 .h2 . h3 )| = Dq . h2 . the integral formula. and the limit is taken as t1 . h2 . which is similar to Poisson’s formula in that it gives the values of a function in the interior of a region in terms of its values on the boundary. C 2 The corresponding characterisation of the derivative in terms of the values of the function at a ﬁnite number of points is f (q0 ) = − h1 . an oriented parallelepiped of codimension 1 in C) with z 0 ∈ L and z ∈ L ⇒ |z − z0 | < δ. then given > 0 there exists δ > 0 such that if L is a directed line segment (i. note that the analogues of the 2-form 1 dq ∧ dq. the 3-form −Dq and the 2 equation d(dq ∧ dqf ) = −2Dqf (q) are the 0-form 1. Our route to the general form of Cauchy’s theorem will be to use Goursat’s method to prove the theorem for a parallelepiped.Thus 1 ∂C dq ∧ dqf + 2 C Dqf (q0 ) ≤ 0 |dq ∧ dq ∧ dfC(t) (h1 . then we can deduce that an analytic function is continuously 17 . h3 )−1 {(h1 h2 − h2 h1 )(f (q0 + h3 ) − f (q0 )) +(h2 h3 − h3 h2 )(f (q0 + h1 ) − f (q0 )) +(h3 h1 − h1 h3 )(f (q0 + h2 ) − f (q0 ))}] (3.17) is similar to that used by Joly [3. Cauchy’s theorem holds for any rectiﬁable contour of integration. then −1 dz L ∂L f − f (z0 ) < . the 1-form dz and the equation df = dzf (z). hi = ti Hi .e. which would be obtained as the Cauchy-Riemann equations if dx ∧ dy ∧ dz were substituted for Dq in the deﬁnition of regularity. 4 Cauchy’s Theorem and the Integral Formula The integral theorems for regular quaternionic functions have as wide a range of validity as those for regular complex functions. h2 . h2 .

j. ρn → 0 as n → ∞. and qn → q∞ as n → ∞. Continuing in this way. ∂C ∂C1 Now perform a similar dissection of C1 . l)dfq (i) + Dq(k. we can write f (q) = f (q∞ ) + α(q − q∞ ) + (q − q∞ )r(q) where α = dfq∞ ∈ F1 . Since the Cn converge on q∞ .. The extension to rectiﬁable contours also follows from an appropriate form of Stokes’s theorem. we obtain a sequence of 4-parallelepipeds Cn . C ⊃ C1 ⊃ C2 ⊃ . 2 h4 2 1 and original vertex q0 + i∈S 2 hi . i. and r(q) → 0 as q → q∞ . For each 1 subset S of {1...1) Clearly there is a point q∞ ∈ ∩Cn . ∂C (4.. k)dfq (l) − Dq(j. Since f is real-differentiable at q∞ . k. ∂C Proof [8] Let q0 and h1 .. .. l.differentiable. 2−n h4 . . i)dfq (j) − Dq(l. 18 ... Then if we deﬁne r(q∞ ) = 0. Proof Dq ∧ dfq (i. such that Cn has edgevectors 2−n h1 . and use Stokes’s theorem to extend Cauchy’s theorem to differentiable contours... then Dqf = ∂C S ∂CS Dqf Hence there is a CS —call it C1 —such that Dqf ≥ 1 16 Dqf .. Theorem 9 (Cauchy’s theorem for a parallelepiped) If f is regular at every point of the 4-parallelepiped C. h4 be the original vertex and edge-vectors of C. The heart of the quaternionic Cauchy’s theorem is the following fact. Theorem 8 A differentiable function f is regular at q if and only if Dq ∧ dfq = 0. l) = Dq(i. and ∂Cn Dqf ≥ 1 16n Dqf . by theorem 6. 3. with original vertices qn . j. r is a continuous function and so |r(q)| has a maximum value ρ n on ∂Cn . k. 2. . 4} let CS be the 4-parallelepiped with edge-vectors 1 h1 . Dq f = 0.. j)dfq (k) = dfq (l) + idfq (i) + jdfq (j) + kdfq (k) = Γr (dfq ) 2 which vanishes if and only if f is regular at q.

. say 2−n ha . hb . since f is regular at q∞ . hence F Dq(q − q∞ )r(q) ≤ 8−n |Dq(ha . hc )| 2−n (|h1 | + . |q − q0 |2 ∂C Proof [7] The argument of theorem 8 shows that ¯ Dq ∧ dfq = −∂ f (q)v (4. c. 2 Theorem 10 (the Cauchy-Fueter integral formula for a parallelepiped) If f is regular at every point of the positively oriented 4-parallelepiped C. b. For q ∈ F (I 3 ) we have |r(q)| < ρn and |q − q∞ | ≤ 2−n (|h1 | + .16−nV (|h1 | + . Dqf (q) ≤ 8. 2−n hb . + |h4 |). hb .2) where v = dt ∧ dx ∧ dy ∧ dz. and the edge-vectors of F are three of the four edge-vectors of C n . Thus Dqf (q) = ∂Cn ∂Cn Dq(q − q∞ )r(q). for any differentiable function f .1). Let V be the largest |Dq(ha .. h4 ) = 0 ∂Cn Cn by theorem 8. and q 0 is a point in the interior of C... ∂C ∂C Since ρn → 0 as n → ∞.. Then F ⊂ ∂Cn . Let F : I 3 → H be one of the 3-parallelepipeds forming the faces of Cn . + |h4 |) ρn . A similar calculation shows that ¯ dfq ∧ Dq = ∂r f (q)v. ∂Cn Combining this with (4.. we ﬁnd Dqf ≤ 8V (|h1 | + .3) 19 . (4. and 2−n hc .... then since the integral over ∂Cn is the sum of 8 integrals over faces F . it follows that Dqf = 0. + |h4 |) ρn . + |h4 |)ρn . hc )| for all choices of a.. . f (q0 ) = 1 2π 2 (q − q0 )−1 Dq f (q).Now Dqf (q) = ∂Cn Cn d(Dq)f (q∞ ) = 0 and Dqα(q − q∞ ) = d(Dqα) = 16−n (Dq ∧ α)(h1 ..

2 δhb .Hence if f and g are both differentiable. By dissecting the given 4-parallelepiped C into 81 4-parallelepipeds with edges parallel to those of C. Since the 3-form (q−q0 ) 0 |2Dq is closed and continuously differentiable in H\{q0 }.b. then 2 q∈∂C0 min |q − q0 | = 1≤a. 1≤a.. then g is differentiable except at q0 .... hence ∂C0 (q − q0 )−1 8V Dq{f (q) − f (q0 )} ≤ 3 |q − q0 |2 W V = −1 (4. |q−q Stokes’s theorem gives (q − q0 )−1 Dq = |q − q0 |2 20 (q − q0 )−1 Dq |q − q0 |2 ∂C0 S . δh4 ) = Wδ 1 Dq( 1 δha .c≤4 max |Dq(ha . − 1 δh4 ). where δ is a positive real number and h1 . we deduce that (q − q0 )−1 Dqf (q) = |q − q0 |2 (q − q0 )−1 Dqf (q). . 1 δhc ) 2 2 where W depends only on h1 . |q − q0 |2 (q − q0 )−1 Dqf (q) = 0 |q − q0 |2 We can now follow the argument of theorem 9 to show that ∂C where C is any 4-parallelepiped not containing q0 . and suppose q0 is at the 1 centre of C0 (so that the original vertex of C0 is q0 − 2 δh1 − . . ¯ ∂r g = ∆ ¯ If f is regular we have ∂ f = 0. |q − q0 |2 ∂C ∂C0 where C0 is any 4-parallelepiped containing q0 which lies in the interior of C and has edges parallel to those of C. hb .4) = q −¯0 ¯ q |q−q0 |4 = ∂r 1 |q−q0 |2 ..5) where... = 0.c≤4 min v(δh1 . h4 are the edge-vectors of C.b.. δh4 . h4 . hc )|. d(gDqf ) = d(gDq)f + gd(Dqf ) = dg ∧ Dqf − gDq ∧ df ¯ ¯ = (∂r g)f + g(∂ f ) v Take g(q) = and (q−q0 )−1 |q−q0 |2 (4.. .. and so d 1 |q − q0 |2 (q − q0 )−1 Dqf = 0. ... as in theorem 9.. we can choose δ so that q ∈ C0 (I 4 ) ⇒ |f (q) − f (q0 )| < for any given > 0. Since f is continuous at q0 . Take C0 to have edge-vectors δh1 ..

21 . we ﬁnd that on S.6) ∂C0 and so (4. in which q − q0 = r cos θ + i sin θ cos φ + j sin θ sin φe−iψ . (2. i. Dq = (q − q0 ) sin2 θ sin φdθ ∧ dφ ∧ dψ = (q − q0 )dS where dS is the usual Euclidean volume element on a 3-sphere. It follows [12.e. and let q0 be any point of U . i. f is harmonic. then in this integral the integrand is a continuous function of (q. p. Hence (q − q0 )−1 Dq = |q − q0 |2 dS = 2π 2 S (4.45) gives ∆f = 0.where S is the 3-sphere |q − q0 | = 1. q0 ) in C0 (∂I 4 ) × C1 (I 4 ) and a C ∞ function of q0 in C1 (I 4 ). 269] that f is real-analytic in U . G is regular except at 0. θ.e.7) ¯ it follows that ∂ G = 0.e. It follows that the integral deﬁnes a C ∞ function of q0 in C1 (I 4 ). ψ). |q − q0 |2 W (q − q0 )−1 Dqf (q) = 2π 2 f (q0 ). eq. oriented so that Dq is in the direction of the outward normal to S. Then we can ﬁnd a 4-parallelepiped C0 such that q0 lies in the interior of C0 and C0 lies inside U . we will use a special notation for it: G(q) = Note that G(q) = −∂ 1 |q|2 q −1 . φ. |q|2 = −∂r 1 |q|2 . Proof Suppose f is analytic in the open set U .6) (4. 2 This fact enables us to give the most general formulation of Cauchy’s theorem and the integral formula for a differentiable contour of integration. (4. and by theorem 10 we have f (q0 ) = 1 2π 2 G(q − q0 )Dqf (q). As an immediate corollary of the integral formula we have Theorem 11 A function which is regular in an open set U is real-analytic in U . when r = 1. so that ∂ f = 0. ∂C0 Let C1 be a 4-parallelepiped such that q0 ∈ C1 (I 4 ) ⊂ int C0 (I 4 ). it follows that (q − q0 )−1 Dqf (q) = |q − q0 |2 2 ∂C ∂C Because of the special role played by the function occurring in this integral formula. i. Working in spherical coordinates (r. ¯ Thus f is C ∞ throughout U .5) becomes (q − q0 )−1 8V Dqf (q) − 2π 2 f (q0 ) ≤ 3 . Since f is regular in U . |q − q0 |2 ∂C0 Since was arbitrary.

e. By theorem 11. C is homologous to a 3-chain C : ∂I 4 → ∂B where q0 ∈ B ⊂ U and the map C has degree 1. and let C be a closed 3-chain in H\{q}. 2 In order to state the general form of the integral formula.Theorem 12 (Cauchy’s theorem for a differentiable contour) Suppose f is regular in an open set U . so C = ∂C0 where C0 is a differentiable 4-chain in U \{q0 }. in the differentiable singular homology of U \{q 0 }. In the case n = 1. and let C be a differentiable 3-chain in U which is homologous to 0 in the differentiable singular homology of U . i. Deﬁnition 5 Let q be any quaternion. we have 1 2π 2 C G(q − q0 )Dqf (q) = 1 2π 2 ∂C0 G(q − q0 )Dqf (q) = f (q0 ) by the integral formula for a parallelepiped.g. C Proof By theorem 8. Since the 3-form G(q − q0 )Dqf (q) is closed and inﬁnitely differentiable in U \{q0 }. Stokes’s theorem gives 1 2π 2 C G(q − q0 )dqf (q) = 1 2π 2 C0 d [G(q − q0 )Dqf (q)] = 0. the 3-form Dqf is inﬁnitely differentiable in U . The wrapping number of C about q is the degree of the map C . where S is the unit sphere with centre q. C is homologous to a 3-chain C of the form C = ρ ◦ C where C (having image ∂B) is as in the previous paragraph and ρ : ∂B → ∂B is a map of degree n. Again using the fact that the 3-form G(q − q0 )Dq is closed in U \{q0 }. hence we can apply Stokes’s theorem to ﬁnd Dqf = C ∂C Dqf = C d(Dqf ) = 0. Then Dq f = 0. to a 3-chain whose image is ∂B for some ball B ⊂ U . C = ∂C for some differentiable 4-chain C in U . Let q0 ∈ U . we need an analogue of the notion of the winding number of a curve round a point in the plane. hence C is homologous to ∂C0 . d(dqf ) = 0. C is homologous to 0 in U \{q0 }. ρ [q0 + r(v + jw)] = q0 + r(v n + jw) 22 . Theorem 13 (the integral formula for a differentiable contour) Suppose f is regular in an open set U . and let C be a differentiable 3-chain in U \{q0 } which is homologous. For general positive n. Then C is homologous to a 3-chain C : ∂I 4 → S. e. for some 4parallelepiped C0 with q0 ∈ int C0 (I 4 ) and C0 (I 4 ) ⊂ U . Then 1 2π 2 C (q − q0 )−1 Dq f (q) = nf (q0 ) |q − q0 |2 where n is the wrapping number of C about q0 . Proof In the case n = 0.

C(s . t2 . tm . tm+1 . If this limit exists. tm . s ¯ ¯ C(s . R) is called the content of C and denoted by σ(C). tm+1 . un+1 ) − C(s .. and so by the previous paragraph 1 2π 2 G(q − q0 )Dqf (q) = 1 2π 2 n ρ◦C n C =1 G(q − q0 )Dqf (q) = nf (q0 ). C is a homologous to a 3-chain C of the form C = C ◦ K. Deﬁnition 6 Let C : I 3 → H be a continuous map of the unit 3-cube into H. tm . Q. Dissect C as C = =1 C where the image of C is the sector 2π 2π( − 1) . |Q|. un ))Dq(C(sl+1 . ≤ arg v ≤ q = q0 + r(v + jw) ∈ ∂B : n n Then each ρ◦C has image ∂B and wrapping number 1 about q0 . P. P. un+1 ) − C(s . un ))g(C(¯. un ). In the case n = −1. C(s .. Q. un ). t4 ). tm . un ). tm ≤ tm ≤ tm+1 and un ≤ un ≤ un+1 . R. the integrand G(q − q0 )Dqf (q) being understood. tm . where |P | = max0≤ ≤p−1 |s s | measures the coarseness of the partition P .. tm . Q. < ur = 1 be three partitions of the unit interval I. t2 . Then = C C =− C = −2π 2 f (q0 ). tm . un ) − C(s . un ). < sp = 1. C 23 +1 − .where r is the radius of B. Let f and g be quaternion-valued functions deﬁned on C(I 3 ). un ) − C(s . Deﬁne p−1 q−1 r−1 σ(C.. Q : 0 = t0 < t1 < . P. |R| → 0. for example the reﬂection (t1 . C is a rectiﬁable 3-cell if there is a real number M such that σ(C. un ) − C(s . un )). tm . which we deﬁne as follows. If this is the case the least upper bound of the numbers (C. and let P : 0 = s0 < s1 < . R) = l=0 m=0 n=0 Dq(C(sl+1 . we denote it by f Dqg. s ¯ ¯ C(s . 2 More generally. < tq = 1 and R : 0 = u0 < u1 < . Q. R) < M for all partitions P . un )). where C : ∂I 4 → ∂B has degree 1 and K : ∂I 4 → ∂I 4 has degree −1.. tm . un ) − C(s . Cauchy’s theorem and the integral formula are valid for rectiﬁable contours. t3 . tm . t4 ) → (1 − t1 . t3 . tm . has a limit in the ¯ ¯ sense of Riemann-Stieltjes integration as |P |. We say that f Dqg is integrable over C if the sum p−1 q−1 r−1 l=0 m=0 n=0 f (C(¯ . For general negative n we dissect the 3-chain C as for general positive n. ¯ where s ≤ s ≤ s +1 .. tm . This establishes the theorem for all n.

We can dissect ∗ ∗ ∗ C as C ∗ = n Cn . Then Dq f = 0. In the general case. so by Stokes’s theorem C Dqf = 0.We extend these deﬁnitions to deﬁne rectiﬁable 3-chains and integrals over rectiﬁable 3-chains in the usual way. where each Cn is a 4-cell lying inside an open ball contained ∗ in U and Cn is rectiﬁable. Let q0 ∈ U . un are partition points in I. C n ∗ ∂Cn A similar argument proves the following general form of the integral formula Theorem 15 (the integral formula for a rectiﬁable contour) Suppose f is regular in an open set U . suppose C = ∂C ∗ where C ∗ is a 4-chain in U . and we can use the same argument as for rectiﬁable curves (see e. Theorem 14 (Cauchy’s theorem for a rectiﬁable contour) Suppose f is regular in an open set U . tm . un ) where Ca is a 3-cell in C and s . In this case. and suppose f and g are continuous functions deﬁned in a neighbourhood U of the image of C. C Proof First we prove the theorem in the case when U is contractible. Just as for rectiﬁable curves. p. [13. we can show that f Dqg is integrable over the 3-chain C if f and g are continuous and C is rectiﬁable. since d(Dqf ) = 0 and f is continuously differentiable (by theorem 11). Poincar´ ’s lemma e applies and we have Dqf = dω for some 2-form ω on D. Then f Dq g = 0. Hence by the ﬁrst part of the theorem ∂C ∗ Dqf = 0.g. and that f Dqg = dω where ω is a 2-form on U . 103]). C C C Furthermore. and let C be a rectiﬁable 3-chain in U \{q0 } which is homologous. and let C be a rectiﬁable 3-chain which is homologous to 0 in the singular homology of U . in the singular homology of U \{q0 }. C The proof proceeds by approximating C by a chain of 3-parallelepipeds with vertices at the points Ca (s . tm . to a differentiable 3-chain whose image is ∂B for some ball B ⊂ U . we have the following weak form of Stokes’s theorem: Stokes’s theorem for a rectiﬁable contour. We can now give the most general forms of Cauchy’s theorem and the integral formula. Stokes’s theorem holds for this chain of 3-parallelepipeds. Then 1 2π 2 C where n is the wrapping number of C about q0 . Let C be a rectiﬁable 3-chain in H with ∂C = 0. and f Dq g ≤ (max |f |)(max |g|)σ(C). and n therefore Dq f = 2 Dqf = 0. 24 (q − q0 )−1 Dq f (q) = nf (q0 ) |q − q0 |2 . But ∂C = 0.

Now write G(q − q0 ) = G(q∞ − q0 ) + α(q − q∞ ) + (q − q∞ )r(q) where α = dGq∞ −q0 ∈ F1 . Liouville’s theorem follows immediately from the Cauchy-Fueter integral formula. Then f is regular in U . the second because α ∧ Dq = 0. As with functions of a complex variable. they satisfy a maximum-modulus principle and a Liouville theorem.g. and write f (q) = f (q∞ ) + s(q) where s(q) → 0 as q → q∞ . Theorem 16 (Morera’s theorem) Suppose that the function f is continuous in an open set U and that ∂C Dqf = 0 for every 4-parallelepiped C contained in U . Then ∂Cn G(q − q0 )Dqf (q) = G(q∞ − q0 ) + ∂Cn Dqf (q) + ∂Cn ∂C α(q − q∞ )Dqf (q0 ) α(q − q∞ )Dqs(q) + ∂Cn (q − q∞ )r(q)Dqf (q) The ﬁrst term vanishes by assumption. by the counterpart of theorem 8 for right-regular functions.1) ∂Cn ∂C Since q0 lies outside C. there is a number M such that |α(q − q∞ | ≤ M |q − q∞ |. The proof given here is based on an incomplete proof by Schuler [8]. dGq−q0 ∧ Dq = 0. so that α ∧ Dq = 0. Morera’s theorem also holds for quaternionic functions. where L is the sum of the lengths of the edges of C. [14. The volume of each face of Cn is at most 8−n V . p.8−n V σn + 2−n Lρn . 85 (second proof)]. and r(q) → 0 as q → q∞ . as in e. but in this case the usual proof cannot easily be adapted. since α is linear. using Goursat’s argument. Now for q ∈ ∂Cn we have |q − q∞ | ≤ 2−n L. (5. G(q − q0 ) is a right-regular function of q inside C and so. where V is the volume of the largest face of C. Proof The method is to show that f satisﬁes the integral formula and then argue as for the analyticity of a regular function (theorem 11). First we show that ∂C G(q − q0 )Dqf (q) = 0 if q0 does not lie inside C.5 Some Immediate Consequences Since regular functions are harmonic. we ﬁnd a sequence of 4-parallelepipeds C n . Hence G(q − q0 )Dqf (q) ≤ 2−n LM.8−n V (|f (q∞ )| + σn ) ∂Cn 25 . converging on a point q∞ and satisfying G(q − q0 )dqf (q) ≥ 1 16n G(q − q0 )dqf (q) . As in theorem 9.

6 Construction of Regular Functions Regular functions can be constructed from harmonic functions in two ways. as in theorem 10. Since 1 Re 0 s2 ∂ u(sq)q ds = = 1 2 1 2 1 0 1 0 s2 t s2 ∂u ∂u (sq) + xi (sq) ∂t ∂xi ds d [u(sq)] ds ds 1 1 = 2 u(q) − su(sq) ds. at least locally. the proof of theorem 10 depends only on the continuity of f and is therefore valid under the present conditions. the parallelepiped can be replaced by a small parallelepiped C 0 containing q0 . By what we have just proved. Proof Without loss of generality we may assume that U contains the origin and is star-shaped with respect to it. If u is harmonic and has continuous second derivatives.1) is regular in U . After this point has been established. Since G(q − q) is a continuously differentiable function of q in the interior of C as long as q lies on its boundary. Thus f is regular. 0 26 .where ρn and σn are the maximum values of the continuous functions r(q) and s(q) on ∂Cn . In this case we will show that the function 1 f (q) = u(q) + 2 Pu 0 s2 ∂ u(sq)q ds (6. Since ρn → 0 and σn → 0 as n → ∞.45) shows that ∂ f is regular. First. Second. it follows that 16n Cn G(q − q0 )Dqf (q) → 0 as n → ∞ and so from (5. Now consider the integral ∂C G(q − q0 )Dqf (q) where q0 lies inside C. as in theorem 11. that f is differentiable and that 1 ¯ ∂ f (q) = 2π 2 ¯ ∂ [G(q − q)]Dq f (q ) = 0 2 ∂C since G is regular. hence 1 G(q − q)Dq f (q ) f (q) = 2 2π ∂C for any 4-parallelepiped C with q ∈ int C(I 4 ) ⊂ U . any real-valued harmonic function is. there is a regular function f deﬁned on U such that Re f = u.1) that ∂C G(q − q0 )Dqf (q) = 0. the real part of a regular function: Theorem 17 Let u be a real-valued function deﬁned on a star-shaped open set U ⊆ H. it follows. if f is harmonic then (2.

6) 1 .5) by solving the equations .1) and (6.3) and (6. This is the elementary potential function in four dimensions.3) gives f (q) = −(q Pu q)−1 − 1 = 2 |q| where arg q = log(Un q) = 1 arg q | Pu q|2   if Pu q = 0   | Pu q| Re q    (6.4) are not very convenient to use. we can differentiate under the integral sign to obtain.5) if q is real and positive. We take U to be the whole of H except for the origin and the negative real axis. for q ∈ U .3) =2 0 s2 ∂ u((1−s)a+sq)(q−a) ds + 2 0 su((1−s)a+sq) ds. |q|2 ∂ u(q) = − q −1 . and |q|−2 is harmonic in U . Then U is star-shaped with respect to 1. Put u(q) = then (6. 1 ¯ But ∂ [∂ u(sq)] = 4 s∆u(sq) = 0 since u is harmonic in U . (In practice the formulae (6. and so the regular function whose real part is |q| −2 is an analogue of the logarithm of a complex variable.27) since u is real. which is i times the usual argument in the complex plane generated by q. |q|2 a = 1. and it is easier to obtain (6.we can write f (q) = 2 0 1 s2 ∂ u(sq)q ds + 2 0 1 su(sq) ds. (6. as log |z| is in the complex plane. (6.4) An example which can be expected to be important is the case of the function u(q) = |q|−2 . Pu q tan−1 | Pu q| . ¯ Hence ∂ f = 0 in U and so f is regular. thus: 1 f (q) = u(q) + 2 Pu 0 1 s2 ∂ u((1−s)a+sq)(q−a) ds 1 (6. formulae (6. and ∂ u(sq) + ei ∂ u(sq)ei = −2∂ u(sq) ¯ = −2∂ u(sq) by (2. 2 If the region U is star-shaped with respect not to the origin but to some other point a.F = − (t2 2t + r 2 )2 27 .2) Since u and ∂ u have continuous partial derivatives in U . (6. ¯ ∂ f (q) = 2 0 1 ¯ s2 ∂ [∂ u(sq)] q ds+ 0 1 s2 {∂ u(sq) + ei ∂ u(sq)ei } ds+2 1 0 ¯ s2 ∂ u(sq) ds.2) must be adapted by changing origin.

let ηq : C → H be the embedding of the complex numbers in the quaternions such that q is the image of a complex number ζ(q) lying in the upper half-plane. such as powers of the variable. Pu q y. r = Pu q.9) ηq (x + iy) = x + | Pu q| ζ(q) = Re q + i| Pu q|. y) = ˜ Im f (x + iy). and its derivative is ˜ ˜ ∂ (∆f ) = ∆f . The derivative of L(q) can most easily be calculated by writing it in the form L(q) = − the result is e i xi r r2 + tei xi . i. Proof [6] Writing t = Re q. v(x.8) |q|2 Thus L(q) is a primitive for the function occurring in the Cauchy-Fueter integral formula. these functions do not include the simple algebraic functions.11) −1 ˜ Then ∆f is regular in the open set ζ (U ) ⊆ H. (6. r) r r v(t. However. r = |r|. r) − − 2 v2 (t. r) − 2 u1 (t. + 3 tan−1 2r2 (r2 + t2 ) 2r t ∂ L(q) = G(q) = (6.10) Then we have Theorem 18 Suppose f : C → C is analytic in the open set U ⊆ C. (6.) We will denote the function (6. (6.12) where f is the derivative of the complex function f .13) 28 . just as the complex logarithm is a primitive for z −1 . r) 1 ∂ 1 1 ¯ ˜ Re ∂ f (q) = 2 [u(t. and u(x. y) = Re f (x + iy). (6. the function occurring in Cauchy’s integral formula. Fueter [4] also found a method for constructing a regular function of a quaternion variable from an analytic function of a complex variable. r)] − . r) + v(t. v(t.5) by −2L(q). the deﬁnition of f gives r ˜ f (q) = u(t. which occur as analytic functions of a complex variable.e. r) ∂t r r (6. Theorem 17 shows that there are as many regular functions of a quaternion variable as there are harmonic functions of four real variables. and deﬁne ˜ f : H → H by ˜ f (q) = ηq ◦ f ◦ ζ(q). (6. For each q ∈ H. ∂t (t + r2 )2 where t = Re q.7) and q −1 . r = Pu q and r = |r|—these express the fact that F : H → P is the pure quaternion part of a regular function whose real part is |q|−2 —and assuming that F has the form F (r)r.∂F 2r + ×F= 2 .

16) 29 . and ¯ ˜ Pu ∂ f (q) = = 1 2 1 2 [u(t. r) + v22 (t.14) The following examples are interesting: When f (z) = z −1 . r) + v1 (t. so ˜ ˜ ˜ ∂ (∆f ) = ∆(∂ f) = ∆f . so ∂ f = f + v/r. r) v(t. r) r r r u2 (t.where the subscript 1 or 2 on u and v denotes partial differentiation with respect to the ﬁrst or second variable. the Cauchy-Riemann equations give u 1 −v2 = u2 +v1 = 0 and so v(t. r) 2r + r r v2 (t. when f (z) = log z. r r Since f = u+iv is analytic. ˜ ∆f (q) = −4L(q) (6. Thus f is regular at q if f is analytic at ζ(q). r) ¯ ˜ ¯ ∂ ∆f (q) = ∆∂ f (q) = − + r ∂t2 r ∂r2 r v11 (t. using the fact that u and v are harmonic functions. r) . r) − r r2 (6. r) = u1 (t. r) + ∂t r × r v(t. From (5. ˜ ∆f (q) = −4G(q). But f = u1 + iv1 . r) =− r =0 since v is a harmonic function of two variables.15) (6. r) ¯ ˜ ∂ f (q) = − .13) a straightforward calculation. 2 Functions of the form f have been taken as the basis of an alternative theory of functions of a quaternion variable by Cullen [15]. A calculation similar to the above shows that v r ˜ ∂ f (q) = 1 (u1 + v2 ) + + 1 (−u2 + v1 ) 2 2r r r v(t. r) + v1 (t. gives ˜ ∆f(q) = 2u2 (t. r Hence ∂2 1 ∂2 v(t. r) + r r ˜ ˜ by the Cauchy-Riemann equations for f . r)] + ∂ r v(t. and ∆(v/r) = 0.

hence CD ⊆ D. Then if f ∈ D. b)f is regular at q. q → q + γ and q → q −1 . Theorem 19 Let H ∗ = H ∪ {∞} be the one-point compactiﬁcation of H. Thus D is a subset of C. it remains in D. any such mapping is conformal and orientationpreserving. The action of rotations and inversions will be needed in studying regular polynomials. If a mapping in D is followed by any of these mappings. b. by the mappings q → αqβ.1) for some a. f has differential dfq = (ac−1 d − b)(cq + d)−1 c dq(cq + d)−1 This is of the form αdqβ. translations and the inversion in the unit sphere followed by a reﬂection [16. 30 .7 Regular Functions and Conformal Mappings Because the quaternions are four-dimensional. M (a. We now show how a regular function gives rise to other regular functions by conformal transformation of the variable. If f is regular at aqb. (ii) Given a function f : H → H and quaternions a. If is regular at q. let M (a. It follows that D = C. If the mapping f : H ∗ → H∗ is conformal and orientation-preserving. p. b)f ](q) = bf (a−1 qb). d ∈ H. and let D be the set of mappings of the form (7. b)f be the function [M (a. which is a combination of a dilatation and a rotation. Theorem 20 (i) Given a function f : H → H. b. |q|2 If f is regular at q −1 . very little remains in quaternionic analysis of the relation between analytic functions and conformal mappings in complex analysis. dilatations. Now C is generated by rotations. and so it seems appropriate to present here the action of the full conformal group. However. 2 The same argument can be used to obtain the alternative representation f (q) = (qc + d)−1 (qa + b). f is of the form f (q) = (aq + b)(cq + d)−1 (7. i. let If : H\{0} → H be the function If (q) = q −1 f (q −1 ). so f is conformal and orientation-preserving. 312].1). the conformal group of H acts on regular functions in a simple way.e. Conversely. c. Proof Let C be the group of orientation-preserving conformal mappings of H∗ .

2) −1 d) (7. part (i) of the theorem to (7.3) (7.5) q Clearly translation preserves regularity.2). |b − ac−1 d|2 |cq + d|2 If f is regular at ν(q). But ∗ ıq Dq(h1 .e. h2 . Then by (2.3). M (ν)f is regular at q. Now If = G(f ◦ ι). 2 (iii) The map q → ν(q) = (aq + b)(cq + d)−1 can be obtained by composing the sequence of maps q →q q →q q → q = cq(b − ac−1 d)−1 = q + d(b − ac = q −1 + ac−1 −1 (7. b)f ]q = Dq ∧ bµ∗ dfµ(q) q = |a|−2 |b|−2 a−1 µ∗ (Dq ∧ dfµ(q) ) q =0 = |a|−2 |b|−2 a−1 (µ∗ Dq)b−1 ∧ bµ∗ dfµ(q) q q if f is regular at µ(q).41). Applying this to the maps (7. f (q + α) is regular at q. where G(q) = q 2 and ι : H\{0} → H is the inversion q → q −1 . |q| Hence Dq ∧ d(If )q = Dq ∧ dGq f (q −1 ) + Dq ∧ G(q)d(f ◦ ı)q = Dq G(q) ∧ ı∗ dfq−1 q since G is regular at q = 0. Proof (i) By theorem 8. −q −1 h3 q −1 ) −1 =− by (2.4) and part (ii) to (7.41) µ∗ Dq = |a|2 |b|2 a Dq b and so Dq ∧ d[M (a. i. let M (ν)f be the function [M (ν)f ](q) = 1 (cq + d)−1 f (ν(q)).4) (7.(iii) Given a function f : H → H and a conformal mapping ν : q → (aq + b)(cq + d)−1 . h2 . It follows from theorem 8 that M (a. b)f is regular at q. we ﬁnd that M (ν)f is regular at q if f is regular at ν(q). h3 ) = Dq(−q −1 h1 q −1 . −q −1 h2 q −1 . 2 (ii) Let µ : H → H be the map q → aqb. 2 31 → ν(q) = q . Thus and so q −1 Dq(h1 . h3 )q −1 |q|4 Dq G(q) = −|q|2 qı∗ Dq q Dq ∧ d(If )q = −|q|2 qı∗ (Dq ∧ dfq−1 ) q =0 −1 if f is regular at q . if f is regular at q + α. it is sufﬁcient to show that Dq ∧ d(If )q = 0.

one (like einθ ) gives a regular function on H\{0} when multiplied by a positive power of |q|. Let ˜ Un = {f |S : f ∈ Un }. . 3). In order to express the Cauchy-Riemann-Fueter equations in a form adapted to the polar decomposition q = ru. where f (q) = r n f(u). (8. Let Wn be the set of functions f : H\{0} → H which are harmonic and homogeneous of degree n over R. Functions in Un and Wn can be studied by means of their restriction to the unit sphere S = {q : |q| = 1}. each have two extensions to harmonic functions on C\{0}. ˜ then Un and Un are isomorphic (as quaternionic vector spaces) by virtue of the correspondence ˜ ˜ ˜ f ∈ U n ⇔ f ∈ Un . harmonic polynomials and harmonic analysis on the group S of unit quaternions. X3 on H\{0}: X0 f = d f (qeθ ) θ=0 . We will see that the space of functions belonging to a particular unitary representation. namely the matrix elements of unitary irreducible representations of S.2) (i = 1. (8. 2. ˜ Wn = {f |S : f ∈ Wn }. regarded as functions on the unit circle in the complex plane. ˜ Similarly Wn and Wn are isomorphic. Then Un and Wn are right vector spaces over H (with pointwise addition and scalar multiplication) and since every regular function is harmonic. Let Un be the set of functions f : H\{0} → H which are regular and homogeneous of degree n over R. namely z n and z −n .1) q using the notation r = |q| ∈ R. In the same way the basic harmonic functions on S. z n . which is to quaternionic analysis what Fourier analysis is to complex analysis. z −n and z −n .. we have Un ⊆ Wn . corresponding to the space of combinations of einθ and e−inθ for a particular value of n. i. Removing the origin from the domain of f makes it possible to consider both positive and negative n (the alternative procedure of adding a point at inﬁnity to H has disadvantages. since regular polynomials do not necessarily admit a continuous extension to ∼ H ∪ {∞} = S 4 ). thus we have the four functions z n . u = |q| ∈ S. The requirement of analytic¯ ¯ ity picks out half of these. dθ d d Xi f = f (qeei θ ) θ=0 = f [q(cos θ + ei sin θ)]θ=0 dθ dθ 32 (8. the other (like e−inθ ) has to be multiplied by a negative power of |q|. one with a negative degree of homogeneity and one with a positive degree. The basic Fourier functions einθ and e−inθ .8 Homogeneous Regular Functions In this section we will study the relations between regular polynomials.. f (αq) = αn f (q) for α ∈ R. each have two extensions to harmonic functions on H\{0}.e..3) . we introduce the following vector ﬁelds X 0 . can be decomposed into two complementary subspaces.

11) we can deduce the following (well-known) facts about W n : Theorem 21 ˜ ˜ (i) Wn = W−n−2 (ii) dim Wn = (n + 1)2 (iii) The elements of Wn are polynomials in q. 33 . being homogeneous of degree n. r From (8. 2¯ 1 ∆ = 2 {Xi Xi + X0 (X0 + 2)} . Thus Wn is the space of eigenfunctions of Xi Xi with eigenvalue −n(n + 2).4) and (8. ∂t ∂xi ∂ ∂ ∂ = −xi +t − ijk xj . Now the vector ﬁelds Xi are ˜ tangential to the sphere S. and they are related to the Cartesian vector ﬁelds ∂t .10) (8.5) (8. are eigenfunctions of X0 with eigenvalue n. [Xi .11) ˜ ˜ ˜ ˜ ˜ so f ∈ Wn . r (8. Proof The elements of Wn .These ﬁelds form a basis for the real vector space of left-invariant vector ﬁelds on the ∂ ∂ multiplicative group of H. Thus if f ∈ Wn . Xi ] = 0. so their restrictions Xi = Xi |S are vector ﬁelds on S.g. ˜ ˜ ˜ which is isomorphic to SU(2).8) (8.9) ¯ Using (3.7) (8. ∂xi by X0 = t Xi ∂ ∂t ∂ ∂xi Their Lie brackets are [X0 .5) the differential operators ∂ and ∆ can be calculated in terms X0 and Xi . e. Xj ] = 2 ijk Xk . Conversely. ∂ ∂ + xi . The result is ¯ ∂ = 1 q −1 (X0 + ei Xi ). f = f |S for some f ∈ Wn . then ˜ ∆ rn f (u) = 1 r2 ˜ ˜ ˜ ˜ rn Xi Xi f + [X0 (X0 + 2)rn ] f = 0 (8.6) (8. if f is an eigenfunction of Xi Xi with ˜ ˜ ˜ of X eigenvalue −n(n + 2). (8.11) shows that they are eigenfunctions of Xi Xi with eigenvalue −n(n + 2). r 1 = 2 ( ijk xj Xk + tXi + xi X0 ).4) (8. ∂t ∂xi ∂xk 1 = 2 (tX0 − xi Xi ). they are a basis for the real vector space of left-invariant vector ﬁelds on the Lie group S. so f is an ˜ eigenfunction of Xi Xi with eigenvalue −n(n + 2) and therefore f is an eigenfunction ˜ ˜ i Xi with eigenvalue −n(n+2). Since they are also harmonic.

∂f and ∂xi belong to Wn−1 . 2 34 2(iii) . ˜ ˜ ˜ ˜ f ∈ Wn ⇒ Xi Xi f = −n(n + 2)f ˜ ˜ ⇒ (Ω−n−2 )(Ω+n )f = 0 ˜ ˜ It follows that Wn is the direct sum of the eigenspaces of Ω with eigenvalues −n and ˜ n + 2 (these are vector subspaces of Wn since the eigenvalues are real). and therefore so does W0 . (8. dn + d−n−2 = (n + 1)2 and by (ii). W0 consists only of constant functions. 71] ˜ functions on S which have eigenvalue −n(n + 2) for X that this space has a basis consisting of the matrix elements of the (n + 1)-dimensional representation of the group S. 2(i) ˜ n is the quaternioniﬁcation of the complex vector space Wn of complex-valued ˜c W ˜ i Xi . 2(ii) (iii) Let dn = dim Un . ∂f Now if f belongs to Wn . Now using (8.e. ˜ ˜ ˜ Wn = Un ⊕ U−n−2 2(i) (ii) It follows from theorem 20(i) that the mapping I is an isomorphism between Un and U−n−3 . As in theorem 21. it follows that Un consists of the eigenfunctions of Ω = ei Xi with eigenvalue −n. i. is dn = 1 (n + 1)(n + 2).˜ ˜ It follows immediately that Wn = W−n−2 . Hence c ˜ dimH Wn = dimC Wn = (n + 1)2 . p. It is known [17. The solution of this recurrence relation. are eigenfunctions of Ω = ei Xi with eigenvalue −n. 2(ii) ˜ In particular. it can be shown that Ω satisﬁes the equation Ω2 − 2Ω + Xi Xi = 0 and therefore Hence ˜ ˜ Ω2 − 2Ω + Xi Xi = 0. Thus dn +dn−1 = (n+1)2 . 2(iii) Theorem 22 ˜ ˜ ˜ (i) Wn = Un ⊕ U−n−2 (ii) Un ∼ U−n−3 = 1 (iii) dim Un = 2 (n + 1)(n + 2) Proof (i) Eq. hence all the (n + 1)th partial derivatives of f vanish and so f is a polynomial. which satisfy X0 f = nf and ¯ ∂i f = 0. d−n−2 = dn−1 .10) shows that the elements of Un . By (1) and theorem 21 (ii). with d0 = 1.9). and all the nth partial derivatives ∂t of f belong to W0 .

Wn therefore has two invariant subspaces. which is isomorphic to SU(2) × SU(2). b)f where R denotes the quasi-regular representation corresponding to the action q → aqb−1 of SU(2) × SU(2) on H\{0}: [R(a. b)f ] where D1 (b)q = bq. and Xi Xi = −n(n + 2) on Wn . we obtain a representation of SU(2) × SU(2). Un . Hence the inﬁnitesimal operators of the tensor product M are ei + Xi . The isotypic components of W are the eigenspaces of the Casimir operator (ei + Xi )(ei + Xi ) = ei ei + Xi Xi + 2Ω. Restricting to the subgroup {(a. and so the isotypic components of Wn for the representation M are the eigenspaces of Ω. on which M acts as the irreducible representations D n ×Dn+1 and Dn × Dn−1 . where W c is the set of complex-valued harmonic functions. b)(q ⊗ f ) = [D 1 (b)q] ⊗ [R(1. To see this. hence (ei + Xi )(ei + Xi ) = 2Ω − n2 − 2n − 3. c The isotypic components of R|W c are the homogeneous subspaces Wn . and the representation of SU(2) × SU(2) can be written as M (a. Now W = H⊗C W c . b) are the differential operators Xi . thus Wn is an invariant subspace under the representation M . the inﬁnitesimal generators of D 1 (b) are ei (by which we mean left multiplication by ei ). The inﬁnitesimal generators of the representation R(1. But ei ei = −3. and so M |Un is the representation D n+1 of SU(2). Theorem 20(ii) refers to a representation M of the group H× × H× deﬁned on the space of real-differentiable functions f : H\{0} → H by [M (a. b)f ](q) = b f (a−1 qb). Similar considerations lead to the following fact: Theorem 23 If f is regular. b) : |a| = |b| = 1}. 35 . the space of homogeneous regular functions of degree n. These subspaces are the eigenspaces of Ω. qf is harmonic.There is a relation between theorem 20(ii) and the fact that homogeneous regular functions are eigenfunctions of Ω. Thus M |W is the tensor product of the representations D 0 × D1 and R|W c of SU(2)× SU(2). and M |Wn is the tensor product (D 0 × D1 ) ⊗ (Dn × Dn ). b)(q ⊗ f ) = (bq) ⊗ R(a. where D n denotes the (n + l)-dimensional complex representation of SU(2). Since the map q → aqb is a rotation when |a| = |b| = 1. has eigenvalue −n for Ω. the set W of harmonic functions is an invariant subspace under this representation. on which R n n acts irreducibly as D × D . restrict attention to the second factor in SU(2) × SU(2). we have the representation M (b)(q ⊗ f ) = M (1. b)f ]q = f (a−1 qb).

w ∈ C and |v|2 + |w|2 = 1. It belongs to a class of induced representations which is studied in [18]. we have z1 = v z1 + wz2 . z2 ) = f (z1 . r (8. Since qf is homogeneous of degree n+1. This notation allows the convenient formulae d [n] z = z [n−1] . Since the functions to be considered involve a number of factorials. Rather than give a rigorous heuristic derivation by following this procedure. z2 ) = z1 z2 n n Dk (u) = (−)n k!(n − k)!Pk (u) are 36 . dz [r] [n−r] (z1 + z2 )[n] = z1 z2 . z2 ).Proof First we show that f ∈ Un ⇒ qf ∈ Wn . Writing u = v + jw where v. If f ∈ Un . it is an eigenfunction of ei Xi with eigenvalue −n and of Xi Xi with eigenvalue −n(n + 2). where z1 + jz2 = u−1 (z1 + jz2 ).11) that it is harmonic. we will state the result and then verify that it is a basis. and so Xi Xi (qf ) = −(n + 1)(n + 3)f.13) where the sum is over all integers r. we introduce the notation zn z [n] = if n ≥ 0 n! = 0 if n < 0 for a complex variable z. 2 The representation M of SU(2) × SU(2) can also be used to ﬁnd a basis of regular polynomials. ¯ ¯ z2 = −wz1 + vz2 .3) of the operators Xi we have Xi (q) = qei Hence Xi Xi (qf ) = Xi (qei f + qXi f ) = −3qf + 2qei Xi f + qXi Xi f since the Xi are real differential operators. and so the result follows for any regular f .12) (8. [k] [n−k] Hence the matrix elements of D n (u) relative to the basis fk (z1 . But any regular function can be represented locally as a series f = fn with fn ∈ Un . From the deﬁnition (8. it follows from (8. where a procedure is given for splitting the representation into irreducible components and ﬁnding a basis for each component. which is not very enlightening in this case. The representation D n of S ∼ SU(2) acts on the space of homogeneous polynomi= als of degree n in two complex variables by [Dn (u)f ](z1 .

(−)r v [n−k− +r] [r] v w[k−r] w[ ¯ ¯ −r] (8. it is easy to verify that Qn satisﬁes the Cauchy-Riemann-Fueter k equations in the form n n ∂Pk−1. ∂Pk =− . Since the functions Dk are independent over C as functions on S for 0 ≤ k. ∂ must have a large kernel and so we cannot conclude from ∂ f = 0 that f is constant.f = 0.14). . it is possible to integrate regular polynomials: Theorem 25 Every regular polynomial has a primitive.14) that ¯ n n Pk j = j Pn−k. 37 . Un has the basis n n Qn (q) = Pk (q) − jPk−1.14) Theorem 24 As a right vector space over H. 2 Since this space is a subspace of Un . although the result is far from unique. It follows that the functions Qn (0 ≤ k ≤ n + 1. Proof Suppose f ∈ Un is such that ∂ f = 0. ∂t ∂xi Thus f can be regarded as a function on the space P of pure imaginary quaternions. Then ∂f ∂f = ei = 0. ∂ maps Un onto Un−1 if n > 0. Proof Using (8. n. ∂¯ v ∂w ¯ n n ∂Pk−1.e. ≤ n. 2 Another basis for Un will be given in the next section.16). which has dimension 1 (n + 1)(n + 2).where n Pk (v + jw) = r n The functions Pk (q) are deﬁned for all quaternions q = v + jw and for all integers k. Using vector notation for elements of P and writing f = f0 + f with f0 ∈ R. We conclude this section by studying the quaternionic derivative ∂ . 0 ≤ ≤ n) are independent over C k and therefore span a right vector space over H of dimension at least 1 (n + 1)(n + 2).n− . ∂f the condition ei ∂xi = 0 becomes f0 + × f = 0. ∂w ∂v n (cf. ≤ n. the functions Pk are independent over C as functions on H for 0 ≤ k. 3.n− and therefore Qn j = Q n k n−k+1. but they are identically zero unless 0 ≤ k. f ∈ P . However. (q) k (0 ≤ k ≤ ≤ n). Since ∂ is a linear map from Un into Un−1 and dim Un > dim Un−1 . ∂Pk = . . the Qn k 2 span Un . i. n ≤ n. which therefore form a basis for k Un . it can be seen from the deﬁnition (8. Thus Un is spanned by the Qn (0 ≤ k ≤ ≤ n). Since zj = j z for any z ∈ C.

b ∈ H ¯ and is non-degenerate since f. we can deﬁne f (0) so that these hold throughout P .e. f = ei × F. is dim Kn = n + 1. g ∈ U−n−2 and let I denote the map : Un → U−n−3 deﬁned in theorem 20(i).e. Then g. then Kn = ker ∂ ⊂ Un .F. I∂ f = S g(q)q −1 Dq q −1 ∂ f (q −1 ) g(q)ı∗ (Dq ∂ f ). this is antilinear in the ﬁrst variable and linear in the second. g = S f (u)g(u)du S where du denotes Haar measure on the group S. Now let f ∈ Un . f0 = − . g b for all a. Let Tn be the right quaternionic vector space of functions F : P → H which are homogeneous of degree n and satisfy 2 F = 0. ∂F . Let Kn be the ∂F subspace of Tn consisting of functions satisfying ei ∂xi = 0.e. 2 Theorem 26 If n < 0. 2 F = 0. and so there exists a function F : P → P such that f= i. i. But dim Un − dim Un−1 = 1 (n + 1)(n + 2) − 1 n(n + 1) = n + 1. 2 2 It follows that ∂ maps Un onto Un−1 . ∂ The above shows that ei ∂xi : Tn+1 → Tn maps Tn+1 onto Kn . i. its kernel is Kn+1 . g = S du = 1 π 2 . ∂xi Then F is harmonic. f = 0 ⇐⇒ f = 0. with dim K0 = 1. For 2 f (q)q −1 Dq g(q) As a map : Un ×Un → H.6) to write this as f. we can use (4. gb = a f. the map ∂ : Un → Un−1 is one-to-one. Proof We introduce the following inner product between functions deﬁned on the unit sphere S: f. normalised so that functions deﬁned on H. and so dim Kn + dim Kn+1 = dim Tn+1 = 2n + 3.If n ≥ 0. f a. The solution of this recurrence relation. S =− 38 . then dim Tn = 2n + 1.

If . and the Laurent series representing a function with an isolated singularity. where n1 is the number of 1’s in ν. hence ι ∗ dq = d¯ and q therefore 1 d¯ ∧ d¯ ∧ d¯f (q −1 ) g q q g. h2 . theorems 25 and 26 are both true trivially. in } with 1 ≤ ir ≤ 3. since U−1 = U−2 = {0}.where ι denotes the map q → q −1 and we have used the fact that ι∗ Dq = −q −1 Dqq −1 for q ∈ S. 9 Regular Power Series The power series representing a regular function. when n = 0. Since f is regular. ν can also be speciﬁed by three integers n1 . and ∂ maps U−n−2 onto U−n−3 if n ≤ −3. it follows that ∂ : Un → Un−1 is one-to-one. Since conjugation is an orthogonal transformation with determinant ¯ ¯ ¯ −1.. I∂ f = − ∂ g(q)q −1 Dq q −1 f (q −1 ) S = − ∂ g. n3 with n1 + n2 + n3 = n. g S On S. .. Dq∂ f = 1 d(dq ∧ dqf ) and so 2 g.. because conjugation is the same as inversion on S. h3 ). so that ν = ∅. we will denote the set of all 2 of them by σn . Dq = −ı∗ Dq = q −1 Dq q −1 . hence. are most naturally expressed in terms of certain special homogeneous functions. n2 . Dq(h1 . and we will write ν = [n1 n2 n3 ]. They are to be used as labels. Thus g. h2 .. the inner product is non-degenerate on U −n−2 . we use the sufﬁx 0 instead of ∅. I∂ f = − 2 S 1 =− dq ∧ dq ∧ dgf (q −1 ) 2 S =− S Dq ∧ ∂ g(q)f (q −1 ) since g is regular. n2 the number of 2’s and n3 the number of 3’s. h3 ) = −Dq(h1 . But I is an isomorphism. I∂ f = − 1 2 1 =− 2 S g(q)d[ı∗ (dq ∧ dq f )] d¯ ∧ ı∗ (dq ∧ dq f ) since ∂S = 0. We write ∂ν for the n’th order differential operator ∂ν = ∂n ∂n = . Let ν be an unordered set of n integers {i1 .. n1 ∂y n2 ∂z n3 ∂xi1 . the inversion ι coincides with quaternion conjugation. 2 In the missing cases n = −1 and n = −2. There are 1 (n + 1)(n + 2) such sets ν.∂xin ∂x 39 .

. As in the previous section.in (xi1 − tei1 ). ∂f + ∂t and since it is homogeneous. ν∈σn Since f is a polynomial.2) n! where the sum is over all n1 !n2 !n3 ! different orderings of n1 1’s. ∂xi ∂f .. Un will denote the right quaternionic vector space of homogeneous regular functions of degree n. ∂ν f is a constant. n2 2’s and n3 3’s.3) (9..∂xin (−1)n Pν (q)∂ν f (q).(tein − xin ) (9. so we can repeat the argument.(xin − tein ) ∂nf ∂xi1 . Since f is regular ei i ∂f =0 ∂xi ∂f + ∂t xi i ∂f = 0. The mirror image of this argument proves that the Pν are also right-regular. thus any regular homogeneous polynomial is a linear combination of the Pν .4) Gν (p)Pν (q) n=0 ν∈σn are valid for |q| < |p|. Theorem 27 The polynomials Pν (ν ∈ σn ) are regular and form a basis for Un ... the elements of Un are polynomials.1) (9.The functions in question are Gν (q) = ∂ν G(q) and Pν (q) = 1 n! (tei1 − xi1 )... Hence Vn = Un . By theorem 21(iii). after n steps we obtain f (q) = = 1 n! i 1 . Theorem 28 The expansions ∞ G(p − q) = = Pν (q)Gν (p) n=0 ν∈σn ∞ (9. the series converge uniformly in any region {(p. Let Vn be the right vector space spanned by the Pν . 40 . t Hence nf (q) = i Proof [17] Let f be a regular homogeneous polynomial of degree n. but dim Vn ≤ 1 2 2 (n + 1)(n + 2) = dim Un by theorem 22(iii). so Un ⊆ Vn . ∂xi (xi − tei ) ∂f But ∂xi is regular and homogeneous of degree n − 1. Then Pν is homogeneous of degree n and Gν is homogeneous of degree −n − 3. q) : |q| ≤ r|p|} with r < 1..

giving G(p − q) = 1 n! r+s=n n=0 ∞ (tej1 ).Proof Just as for a complex variable.. multiplied by G(p). But G is also right-regular.(tejr )(−xi1 ).∂xis ∞ = n=0 ν∈σn Pν (q)Gν (p). This proves (9. the series converges absolutely and uniformly in any ball |q| ≤ r with r < 1.7) 41 . each derivative with respect to t can be replaced by the combi∂ nation − i ei ∂xi .e.is ∂nG (p) ∂xj1 ....5) it follows that G(p−q) can be expanded as a power series in p−1 q. i. (9... r ∂x .∂xjr ∂xi1 . Regarding this series as a power series in q and identifying it with the Taylor series of G about p.... ∂xi i so the derivatives with respect to t in (9... Then there is a ball B with centre 0 in which f (q) is represented by a uniformly convergent series f (q) = n=0 ν∈σn Pν (q)aν .. the series converges uniformly in any region |p−1 q| ≤ r with r < 1.jr i1 .s=0 i1 .6) Since G is regular..6) can alternatively be replaced by combinations of derivatives with respect to the xi with coefﬁcients on the right..5).xis .∂x (r + s)! ∂t i1 is (9. Because G has ¯ the multiplicative property G(q1 q2 ) = G(q2 )G(q1 )... (9.is ∂ r+s G (−1)r+s (p) tr xi1 . Hence G(1 − q) = (1 − q)−2 (1 − q ) can be expanded as a power series in q ¯ and q which converges uniformly in any ball with radius less than 1. n=0 ν∈σn 2 Theorem 29 Suppose f is regular in a neighbourhood of 0. we have ∞ (1 − q)−1 = qn n=0 for |q| < 1.(−xis ) j1 .. we have ∞ G(p − q) = r. thus giving ∞ G(p − q) = Gν (p)Pν (q). ∂G =− ∂t ∂G ei ..

∂ν f (q) = Theorem 30 (the Laurent series) Suppose f is regular in an open set U except possibly at q0 ∈ U . f (q) can be represented by a series ∞ f (q) = n=0 ν∈σn {Pν (q − q0 )aν + Gν (q − q0 )bν } which converges uniformly in any hollow ball {q : r ≤ |q − q0 | ≤ R} . 42 (9. Gν (q − q0 ) Dq f (q). The coefﬁcients aν and bν are given by aν = bν 1 2π 2 1 = 2π 2 C with r > 0. 2π 2 S But the functions Gν are right-regular except at 0. This follows by putting f = Pµ in (9. Then for q ∈ B the integral formula gives f (q) = 1 2π 2 S G(q − q)Dq f (q ) ∞ 1 = 2 2π Pν (q)Gν (q )Dq f (q ) S n=0 ν∈σn since |q| < |q | if q ∈ B and q ∈ S. (9. The series converges uniformly on B × S. B a closed ball with centre 0 lying inside S.7). so (4. and therefore Stokes’s theorem can be used to replace the contour S by ∂B.9) Proof Let S be a sphere with centre 0 lying inside the domain of regularity of f . so it can be integrated term by term to give a uniformly convergent series of the form (9. 2π 2 S where S is any sphere containing the origin.8) (9. Then there is a neighbourhood N of q0 such that if q ∈ N and q = q0 . 2 Corollary 1 Gµ (q) Dq Pν (q) = δµν .12) C .4) gives d(Gν Dqf ) = 0 outside B. Differentiating the integral formula gives (−1)n Gν (q − q)Dq f (q ) 2π 2 S Hence ∂ν f (0) = (−1)n aν . Pν (q − q0 ) Dq f (q). which lies inside N .where the coefﬁcients aν are given by aν = 1 Gν (q) Dq f (q) 2π 2 ∂B = (−1)n ∂ν f (0).11) (9.7) with 1 aν = Gν (q)Dqf (q).

and let N = int B1 .13) S1 S2 For q ∈ S1 we have |q − q0 | < |q − q0 |. If C is a 3chain as in the statement of the theorem. Since the Pν are right-regular.13) gives (9.15) into (9. d(Pν Dqf ) = 0 in U \{q0 } and so the contour S2 in the formula for bν can be replaced by any 3-chain C as in the statement of the theorem.11). Stokes’s theorem gives (9.4) to obtain − 1 2π 2 G(q − q)Dq f (q ) = = n=0 ν∈σn S2 1 2π 2 ∞ S2 n=0 ν∈σn Gν (q − q0 )Pν (q − q0 )Dq f (q ) Gν (q − q0 )bν (9. 2 ACKNOWLEDGEMENTS I am grateful for the hospitality of the Department of Applied Mathematics and Theoretical Physics in the University of Liverpool. where part of this work was done.10) for all q ∈ N \{q0 }.14) S1 Gν (q − q0 )Dq f (q ) and the series is uniformly convergent in any ball |q−q0 | ≤ R with R < R1 . and so as in theorem 29 1 2π 2 where aν = ∞ S1 G(q − q)Dq f (q ) = 1 2π 2 n=0 ν∈σn Pν (q − q0 )aν (9. and to Drs P J McCarthy and C J S Clarke for many stimulating discussions. it is homologous to S1 .where C is any closed 3-chain in U \{q0 } which is homologous to ∂B for some ball B with q0 ∈ B ⊂ U (so that C has wrapping number 1 about q0 ).15) where bν = 1 2π 2 S2 Pν (q − q0 )Dq f (q ). bν is independent of the choice of R2 and so the same representation is valid for all q ∈ N \{q0 }. Proof Choose R1 so that the closed ball B1 = {q : |q − q0 | ≤ R1 } lies inside U . The series is uniformly convergent in any region |q − q0 | ≥ r with r > R2 . since d(Gν Dqf ) = 0 in U \{q0 }. (9. S1 = ∂B1 . Then by the integral formula. For q ∈ S2 we have |q − q0 | < |q − q0 | and so we can expand G(q − q) as in (9. In particular. 43 .14) and (9. Now putting (9. f (q) = 1 2π 2 G(q − q)Dq f (q ) − 1 2π 2 G(q − q)Dq f (q ). and let S2 be the sphere {q : |q − q0 | = R2 }. choose R2 so that 0 < R2 < |q − q0 | < R1 . Given q ∈ N \{q0 }.

Helv. C A: “The Quaternion Calculus. [10] Porteous. 320-335 (1937).” Comment. London: Macmillan 1905. 371-378 (1936).” Comment. [8] Schuler. Helv. R: “Die Funktionentheorie der Differentialgleichungen ∆u = 0 und ∆∆u = 0 mit vier reellen Variablen. [13] Heins. E C: The Theory of Functions. [15] Cullen. London: Longmans Green 1866. 44 . W R: Elements of Quaternions. 20. 7. J. 331-338 (1977). 1939. 8. [2] Tait. Berlin: Springer-Verlag 1954. [9] Fueter. B: “Zur Theorie der regul¨ ren Funktionen einer Quaternionen-Variablen. math. London: Van Nostrand Reinhold 1969.U. [17] Sugiura. 307-330 (1935). P G: An Elementary Treatise on Quaternions. 327-342 (1937). Helv. London: Interscience 1962. [4] Fueter. H: Elementary Theory of Analytic Functions of One or Several Complex Variables. M: Unitary Representations and Harmonic Analysis.P. [16] Schouten. Helv. 1867. 9. C J: A Manual of Quaternions.” Comment. math.” a Comment. [6] Haefeli. [14] Titchmarsh.U. J A: Ricci-Calculus. Monthly 80. Phys.” Comment. 139-148 (1965). [18] McCarthy. math. R and Hilbert. 382-420 (1947). London: Wiley 1975.” Duke Math. I R: Topological Geometry. Oxford: O. Helv. [11] Cartan. Cambridge: C. D: Methods of Mathematical Physics Vol II. 995-1008 (1973). ¨ [5] Fueter. H: “Hyperkomplexe Differentiale. P J and Sudbery. 2nd ed. math. R: “Die Singularit¨ ten der eindeutigen regul¨ ren Funktionen einer Quatera a nionenvariablen. R: “Uber die analytische Darstellung der regul¨ ren Funktionen einer a Quaternionenvariablen. London: Addison-Wesley 1963. math. [3] Joly.P. 2nd ed. C G: “An integral theorem for analytic intrinsic functions on quaternions. A: “Harmonic analysis of generalised vector functions. 10. 32. Math.” J. A 10.References [1] Hamilton.” Amer. London: Academic Press 1968. generalised spin-weighted functions and induced representations. [7] Deavours. M: Complex Function Theory. [12] Courant.