Skew-symmetric endomorphisms in M1,n : A unified canonical form

with applications to conformal geometry

arXiv:2012.11999v1 [gr-qc] 22 Dec 2020

Marc Mars and Carlos Peón-Nieto

Instituto de Fı́sica Fundamental y Matemáticas, Universidad de Salamanca
Plaza de la Merced s/n 37008, Salamanca, Spain
May 13, 2021

Abstract
We show the existence of families of orthonormal,

future directed bases which allow to cast every skew-
symmetric endomorphism of M1,n (SkewEnd M1,n ) in a single canonical form depending on
a minimal num-
ber of parameters. This canonical form is shared by every pair of elements in SkewEnd M1,n differing by an
orthochronous Lorentz transformation, i.e. it defines the orbits of the orthochronous Lorentz group

under the
adjoint action on its algebra. Using this form, we
obtain the quotient topology of SkewEnd M1,n /O+ (1, n).
From known relations between SkewEnd M1,n and the conformal Killing vector fields (CKVFs) of the
sphere Sn , a canonical form for CKVFs follows immediately. This form is used to find adapted coordinates
to an arbitrary CKVF that covers all cases at the same time. We do the calculation for even n and obtain
the case of odd n as a consequence. Finally, we employ the adapted coordinates to obtain a wide class
of TT-tensors for n = 3, which provide Cauchy data at conformally flat null infinity I . Specifically, this
class of data is characterized for generating Λ > 0-vacuum spacetimes with two-symmetries, one of which
axial, admitting a conformally flat I . The class of data is infinite dimensional, depending on two arbitrary
functions of one variable as well as a number of constants. Moreover, it contains the data for the Kerr-de
Sitter spacetime, which we explicitly identify within.

Contents
1 Introduction 2

2 Classification of Skew-symmetric endomorphisms 4

3 Canonical form for skew-symmetric endomorphisms 11

4 Simple endomorphisms 15

5 O+ (1, d − 1)-classes 17
5.1 Structure of SkewEnd (V )/O+ (1, d − 1) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

6 Conformal vector fields 22

7 Adapted coordinates 28

1
8 TT-Tensors 38

1 Introduction
Having a Lie group G acting on a space X respresenting a set of physical quantities is always a desirable feature
in a physical problem, as Lie groups represent symmetries (either global or gauge) and their presence is often
translated into a simplification of the formal aspects of the problem. Roughly speaking, in this situation the
“relevant” part for the physics effectively happens in the quotient space X/G. For the study of these quotient
spaces, one may be interested in obtaining a unified form to give a representative for every orbit in X/G, i.e. a
canonical form (also known as normal form). A particularly relevant case is when X is a Lie algebra g and G its
Lie group acting by the adjoint action, in which case the orbits are also called conjugacy classes of G (see e.g.
[5]). In the first part of this paper, we study the conjugacy classes of the pseudo-orthogonal group O(1, n) (or
Lorentz group), for which we will obtain a canonical form. Our main interest on this, addressed in the second
part of the paper, lies in its relation with the Cauchy problem of general relativity (GR) (cf. [11] and e.g. [7],
[14]) and more precisely, its formulation at null infinity I for the case of positive cosmological constant Λ (cf.
[12], [13]). In the remainder of this introduction we summarize our ideas and results, and we will also briefly
review some results on conjugacy classes of Lie groups related to our case, as well as the Cauchy problem of
GR with positive Λ.
A typical example of a canonical form in the context described above, is the well-known Jordan form,
which represents the conjugacy classes of GL(n, K) (where K is usually R, C or the quaternions H). Besides
this example, the problem of finding a canonical representative for the conjugacy classes of a Lie group has
been adressed numerous times in the literature. The reader may find a list of canonical forms for algebras
whose groups leave invariant a non-degenerate bilinear form in [9] (this includes symmetric, skew-symmetric
and simplectic algebras over R, C and H) as well as the study of the affine orthogonal group (or Poincaré group)
in [8] or [19]. Notice that these works deal, either directly or indirectly, with our case of interest O(1, n), whose
algebra o(1, n) will be represented in this paper as skew-symmetric endomorphisms of Minkowski spacetime
M1,n . When giving a canonical form, it is usual to base it on criteria of irreducibility rather than uniformity
(e.g. [8], [9], [19]). This is similar to what is done when the Darboux decomposition is applied to two-forms (i.e.
elements of o(1, n)), for example in [23] or for the low dimensional case n = 3 (e.g. [18], [29]). As a consequence,
all canonical forms found for the case of o(1, n) require two different types of matrices to represent all orbits, one
and only one fitting a given element. Our first aim in this paper is to give a unique matrix form which represents
each element F ∈ o(1, n), depending on a minimal number of parameters that allows one to easily determine
its orbit under the adjoint action of O(1, n). This is obviously achieved by loosing explicit irreducibility in the
form. However, this canonical form will be proven to be fruitful by giving several applications. This same issue
has also been adressed in [24] for the case of low dimensions, i.e. O(1, 2) and O(1, 3), where in addition, several
applications are worked out. The present work constitutes a natural generalization of the results in [24] to
arbitrary dimensions.
The Lorentz group is well-known to be of particular interest in physics, as for example, it is the group of
isotropies of the special theory of relativity and the Lorentz-Maxwell electrodynamics (e.g. [22], [27], [29]).
Its study in arbitrary dimensions have received renewed interest with theories of high energy physics such as
conformal field theories [28] or string theories [19]. Related to the former and for our purposes here, a fact
of special relevance is that the orthochronous component O+ (1, n) ⊂ O(1, n) is homomorphic to the group of
conformal transformations of the n-sphere, Conf (Sn ). The conformal structure of I happens to be fundamental
for the Cauchy problem at null infinity of GR for spacetimes with positive Λ , as it is the gauge group for the
set of initial data. Such a set consists of a manifold Σ endowed with a (riemannian) conformal structure [γ],

2
representing the geometry of null infinity I := (Σ, [γ]), together with the conformal class [D] of a transverse
(i.e. zero divergence), traceless, symmetric tensor D (TT-tensor) of I . If the spacetime generated by the data
is to have a Killing vector field, the TT-tensor must satisfy a conformally covariant equation depending on a
conformal Killing vector field (CKVF) of I , the so-called Killing Initial Data (KID) equation [25].
The class of data in which [γ] contains a constant curvature metric (or alternatively the locally conformally
flat case) includes the family of Kerr-de Sitter black holes and its study could be a possible route towards a
characterization result for this family of spacetimes. Even in this particular (conformally flat) case, it is difficult
to give a complete list of TT-tensors. An example can be found in [3], where the author gives a class of solutions
with a direct and elegant method, but the solution is restricted in the sense that global topological conditions
are imposed on I . Namely, the solutions obtained by this method must be globally regular on S3 and hence
cannot contain the family of Kerr-de Sitter, which is known (see e.g. [23]) to have I with topology S3 minus two
points, which correspond with the loci where the Killing horizons “touch” I . The local problem for TT tensors
is much more difficult to solve with generality, so our idea is to simplify it by imposing two KID equations
to the data, so that the corresponding spacetimes have at least two symmetries. Using the homomorphism
between O+ (1, n + 1) and Conf (Sn ), we induce a canonical form for CKVFs from the canonical form obtained
for o(1, n + 1). Since this form covers all orbits of CKVFs under the adjoint action of Conf (Sn ), our adapted
coordinates fit every CKVF and in addition, since the KID equation is conformally covariant, we can choose
a conformal gauge where this CKVF is a Killing vector field, which makes the KID equation trivial. Hence, a
remarkable feature from our method is that by solving one simple equation, we are solving many cases at once.
This has already been done in the case of S2 in [24] and here we extend it to the more interesting and difficult
case of (open domains of) S3 .
Specifically, we obtain the most general class of TT-tensors on a conformally flat I such that the Λ > 0-
vacuum four-dimensional spacetime generated by these data admits two local isometries, one of them axial. It
is worth highlighting that this is a broad class (of infinite dimensions as it depends on functions) of TT-tensors
and it contains the Kerr-de Sitter Cauchy data at I . This provides a potentially interesting ”sandbox“ to try
the consistence of possible definitions of (global) mass and angular momentum (see [30] for a review on the
state of the art). Recall that symmetries are well-known to be related to conserved quantities, in particular,
axial symmetry is related to conservation of angular momentum and time symmetry to conservation of energy.
Moreover, for a spacetime to have constant mass, one may require no radiation escaping from or coming within
the spacetime, a condition which, following the criterion of [10], is guaranteed by conformal the flatness of I .
Finally, the presence of the Kerr-de Sitter data within the set of TT-tensors contributes to its physical relevance
and furnishes the possibility of looking for new characterization results for this family of spacetimes.
As an additional sidenote concerning our results, notice that both the canonical form of CKVFs as well as
the adapted coordinates are obtained in arbitrary dimensions, so similar applications may be worked out in
arbitrary dimensions which, needless to say, is a considerable harder problem. On the possible extension to
more dimensions of this type of TT-tensors, one should mention that the Cauchy problem at I for positive
cosmological constant is known to be well-posed in arbitrary even dimensions [2]. However the KID equations
are only known to be a necessary consequence of having symmetries, but sufficiency is an open problem in
spacetime dimensions higher than four.
The paper is organized as follows. In order to properly define the canonical form, in Section 2 we rederive
a classification result for skew-symmetric endomorphisms (cf. Theorem 2.6), employing only elementary linear
algebra methods. The results of this section are known (see e.g. [16], [17], [20]), but the method is original and
we believe more direct than other approaches in the literature. We also include the derivations in order to make
the paper self-contained. Section 2 leads to the definition of canonical form in Section 3. Section 4 deals with
a particular type of skew-symmetric endomorphisms (the so-called simple, i.e. of minimal matrix rank), which
will be useful for future sections. In Section 5 we work out some applications of our canonical form: identifying

3
invariants which characterize the conjugacy classes of the orthochronous Lorentz group (cf. Theorem 5.1) and
obtaining the topological structure of this quotient space (cf. Section 5.1).
In Section 6 we use the homomorphism between O+ (1, n + 1) and Conf (Sn ) and apply the canonical form
obtained for skew-symmetric endomorphisms to give a canonical form for CKVFs, together with a decomposed
form (cf. Proposition 6.1) which analogous to the one given for skew-symmetric endomorphisms. In Section
7, we adapt coordinates to CKVFs in canonical form, first in the even dimensional case, from which the odd
dimensional case is obtained as a consequence. Finally, in Section 8 we employ the adapted coordinates to find
the most general class of data at I corresponding to spacetime dimension four, such that I is conformally
flat and the (Λ > 0)-vacuum spacetime they generate admits at least two symmetries, one of which is axial. It
is remarkable how easily are these equations solved with all the tools developed so far. With this solution at
hand, we are able to identify the Kerr-de Sitter family within.

2 Classification of Skew-symmetric endomorphisms

In this section we derive a classification result for skew-symmetric endomorphisms of Lorentzian vector spaces.
Let V be a d-dimensional vector space endowed with a pseudo-Riemannian metric g. If g is of signature
(−, +, · · · , +), then (V, g) is said to be Lorentzian. Scalar product with g is denoted by h , i. An endomorphism
F : V −→ V is skew-symmetric when it satisfies
hx, F (y)i = −hF (x), yi ∀x, y ∈ V.
We denote this set by SkewEnd (V ) ⊂ End (V ). We take, by definition, that eigenvectors of an endomorphism are
always non-zero. We use the standard notation for spacelike and timelike vectors as well as for spacelike, timelike
and degenerate vector subspaces. In our convention all vectors with vanishing norm are null (in particular, the
zero vector is null). We denote ker F and Im F , respectively, to the kernel and image of F ∈ End (V ).
Lemma 2.1. [Basic facts about skew-symmetric endomorphisms] Let F be a skew-symmetric endomorphism in
a pseudo-riemannian vector space V . Then
a) ∀w ∈ V , F (w) is perpendicular to w, i.e. hF (w), wi = 0.
b) Im F ⊂ (ker F )⊥ and ker F ⊂ (Im F )⊥ .
c) If w ∈ ker F ∩ Im F then w is null.
d) If w ∈ V is a non-null eigenvector of F , then its eigenvalue is zero.
e) If w is an eigenvector of F with zero eigenvalue, then all vectors in Im F are orthogonal to w, i.e.
Im F ⊂ w⊥ .
f ) If F restricts to a subspace U ⊂ V (i.e. F (U ) ⊂ U ), then it also restricts to U ⊥ .
Proof. a) is immediate from hw, F (w)i = −hF (w), wi. For b), let v ∈ ker F and w be of the form w = F (u) for
some u ∈ V , then
hw, vi = hF (u), vi = −hu, F (v)i = 0
the last equality following because F (v) = 0. c) is a consequence of b) because w belongs both to ker F and to
its orthogonal, so in particular it must be orthogonal to itself, hence null. d) is immediate from
0 = hw, F (w)i = λhw, wi

4
so if w is non-null, its eigenvalue λ must be zero. e) is a corollary of b) because by hypothesis w ∈ ker F so

Im F ⊂ (ker F )⊥ ⊂ w⊥

the last inclusion being a consequence of the general fact U1 ⊂ U2 ⇒ U2⊥ ⊂ U1⊥ . Finally, f ) is true because for
any u in a F -invariant subspace U and w ∈ U ⊥

0 = hF (u), wi = − hu, F (w)i .

Another well-known property of skew-symmetric endomorphisms that we will use is that dim Im F is always
even. Equivalently, dim ker F has the same parity than dim V . To see this, consider the 2-form F assigned to
every F ∈ SkewEnd (V ) by the standard relation

F (e, e′ ) = he, F (e′ )i , ∀e, e′ ∈ V.

The matrix representing F is skew in the usual sense, hence the dimension of Im F ⊂ V ⋆ (the dual of V ) is the
rank of that matrix, which is known to be even (see e.g. [15]) and clearly dim Im F = dim Im F .
The strategy that we will follow to classify skew-symmetric endomorphisms of V with g Lorentzian is via
F -invariant spacelike planes. Conditions for F -invariance of spacelike planes are stated in the following lemma:
Lemma 2.2. Let F ∈ SkewEnd (V ). Then F has a F -invariant spacelike plane Πs if and only if

F (u) = µv, F (v) = −µu, (1)

for Πs = span{u, v} with u, v ∈ V spacelike, orthogonal, unit and µ ∈ R. Moreover, (1) is satisfied for µ 6= 0 if
and only if ±iµ are eigenvalues of F with (null) eigenvectors u ± iv, for u, v ∈ V spacelike, orthogonal with the
same square norm.
Proof. If (1) is satisfied for u, v ∈ V spacelike, orthogonal, unit, then Πs = span{u, v} is obviously F -invariant
spacelike. On the other hand, if Πs is F -invariant, then it must hold that

F (u) = a1 u + a2 v, F (v) = b1 u + b2 v, a1 , a2 , b1 , b2 ∈ R,

for a pair of orthogonal, unit, spacelike vectors u, v spanning Πs . Using skew-symmetry and the orthogonality
and unitarity of u, v, the constants are readily determined: a2 = b2 = 0 and a2 = −b1 =: µ, which implies (1).
This proves the first part of the lemma.
For the second part, it is immediate that if (1) holds with µ 6= 0, then ±iµ are eigenvalues of F with respective
eigenvectors u ± iv. The orthogonality of u, v follows from hF (u), ui = 0 = µ hv, ui and the equality of norm
from skew-symmetry hF (u), vi = − hu, F (v)i ⇒ µ hv, vi = µ hu, ui. Assume now that F has an eigenvalue iµ 6= 0
with (necessarily null) eigenvector w = u + iv, for u, v ∈ V . Since F is real, neither u nor v can be zero. From
the nullity property hw, wi = 0, it follows that hu, ui − hv, vi = 0 and hu, vi = 0. Hence, u, v are orthogonal with
the same norm, so they are either null and proportional, which can be discarded because it would imply that u
(and v) is a real eigenvector with complex eigenvalue; or otherwise u, v are spacelike, thus the lemma follows.

There is an analogous result for F -invariant timelike planes:

5
Lemma 2.3. Let F ∈ SkewEnd (V ). Then F has a F -invariant timelike plane Πt if and only if

F (e) = µv, F (v) = µe, (2)

for Πt = span{e, v} with e, v ∈ V for e timelike unit orthogonal to v spacelike, unit and µ ∈ R. Moreover, (1) is
satisfied for µ 6= 0 if and only if ±µ are eigenvalues of F with (null) eigenvectors e ± v, for e, v ∈ V orthogonal,
timelike and spacelike respectively with opposite square norm.
Proof. For the first claim, repeat the first part of the proof of Lemma 2.2 assuming u = e timelike.
For the second claim, assume (1) is satisfied with µ 6= 0. Then it is immediate that hF (e), ei = 0 = µ hv, ei,
hence e, v are orthogonal and by skew-symmetry hF (e), vi = − he, F (v)i ⇒ µ hv, vi = −µ he, ei, i.e. must have
opposite square norm. Conversely, let ±µ 6= 0 be a pair of eigenvalues with respective null eigenvectors q± , that
w.l.o.g can be chosen future directed. Then e := q+ + q− and v := q+ − q− are orthogonal, with opposite square
norm he, ei = 2 hq+ , q− i = − hv, vi < 0, and they satisfy (2).
The F -invariant spacelike or timelike planes will be often be refered to as “eigenplanes” and µ will be denoted
as the “eigenvalues” of Π. Notice that a simple change of order in the vectors switches the sign of the eigenvalue
µ. Thus, unless otherwise stated, we will consider the eigenvalues of eigenplanes (both spacelike and timelike)
non-negative by default.
The first question we address here is under which conditions such a plane exists (cf. Proposition 2.1). But
before doing so, we need to prove some results first.
Lemma 2.4. Let V be a Lorentzian vector space F ∈ SkewEnd (V ). Then there exist two vectors x, y ∈ V ,
with x 6= 0, such that one of the three following exclusive possibilities hold
(i) x is a null eigenvector of F .
(ii) x is a non-null eigenvector (with zero eigenvalue).
(iii) x, y are orthogonal, spacelike and with the same norm, and define an eigenplane of F with non-zero
eigenvalue, i.e.

F (x) = µy, F (y) = −µx, µ ∈ R\ {0} .

If, instead, V is riemannian, only cases (ii) and (iii) can arise.
Proof. From the Jordan block decomposition theorem we know that there is at least one, possibly complex,
eigenvalue s1 + is2 with eigenvector x + iy, that is, F (x + iy) = (s1 + is2 )(x + iy), or equivalently:

F (x) = s1 x − s2 y, (3)
F (y) = s2 x + s1 y. (4)

This system is invariant under the interchange (x, y) → (−y, x), so without loss of generality we may assume
x 6= 0. The respective scalar products of (3) and (4) with x, y yield
     
s1 hx, xi − s2 hx, yi = 0 hx, xi − hx, yi s1 0
⇐⇒ = . (5)
s1 hy, yi + s2 hx, yi = 0 hy, yi hx, yi s2 0

Observe that if s1 + is2 6= 0 the determinant of the matrix must vanish. i.e. hx, yi (hx, xi + hy, yi) = 0. Hence,
we can distinguish the following possibilities:

6
(a) s1 = s2 = 0. Then x is an eigenvector of F with vanishing eigenvalue so we fall into cases (i) or (ii).
(b) s1 + is2 6= 0. From hx, yi (hx, xi + hy, yi) = 0 we distinguish two cases:
(b.1) hx, yi = 0. If s1 6= 0 then (5) forces x and y to be both null and, being also orthogonal to each other,
there is a ∈ R such that y = ax and we fall into case (i). So, we can assume s1 = 0 (and then s2 6= 0). Let
µ := −s2 , thus (iii) follows from equations (3), (4) and Lemma 2.2.
(b.2) hx, yi =
6 0. Then hx, xi = −hy, yi and the matrix problem (5) reduces to

s1 hx, xi − s2 hx, yi = 0.

In addition, (3) and (4) imply

hF (x), yi = s1 hx, yi − s2 hy, yi = s1 hx, yi + s2 hx, xi = hF (y), xi .

But skew-symmetry requires hF (x), yi = − hF (y), xi, so hF (y), xi = 0 and we conclude

s1 hx, yi + s2 hx, xi = 0.

Combining with (2) yields

    
hx, xi − hx, yi s1 0
= .
hx, yi hx, xi s2 0

The determinant of this matrix is non-zero which yields a contradiction with s1 + is2 6= 0. So this case is
empty.
To conclude the proof, we must consider the case when the vector space V is riemannian. The proof is
identical except from the fact that all cases involving null vectors are imposible from the start.

Remark 2.1. One may wonder why the lemma includes the possibility of having a spacelike eigenplane (case
(iii)), but not a timelike eigenplane. The reason is that invariant timelike planes, which are indeed possible, fall
into case (i) by Lemma 2.3, because e ± v are null eigenvectors.
In the case of Riemmanian signature, Lemma 2.4 can be reduced to the following single statement:
Corollary 2.4.1. Let V be Riemannian of dimension d and F ∈ SkewEnd (V ). If d = 1 then F = 0 and if
d ≥ 2 then there exist two orthogonal and unit vectors u, v satisfying

F (u) = µv, F (v) = −µu, µ∈R (6)

Proof. The case d = 1 is trivial, so let us assume d ≥ 2. By the last statement of Lemma 2.4 either there exists
an eigenvector x with zero eigenvalue or the pair {u, v} claimed in the corollary exists. In the former case,
we consider the vector subspace x⊥ . Its dimension is at least one and F restricts to this space so again either
the pair {u, v} exists or there is y ∈ x⊥ satisfying F (y) = 0. But then {x, y} are orthogonal and non-zero.
Normalizing we find a pair {u, v} that satisfies (6) with µ = 0,
Lemma 2.4 lists a set of cases, one of which must always occur. However, we now show that, if the dimension
is sufficiently high, case (i) of that lemma implies one of the other two:

7
Lemma 2.5. Let F ∈ SkewEnd (V ), with V Lorentzian of dimension at least four. If F has a null eigenvector,
then it also has either a spacelike eigenvector or a spacelike eigenplane.
Proof. Let k ∈ V be a null eigenvector of F . The space A := k ⊥ ⊂ V is a null hyperplane and F restricts to A.
On this space we define the standard equivalence relation y0 ∼ y1 iff y0 − y1 = ak, a ∈ R. The quotient A/ ∼
(which has dimension at least two) inherits a positive definite metric g and F also descends to the quotient.
More precisely, if we denote the equivalence class of any y ∈ A by y, then for any y ∈ A/ ∼ and any y ∈ y the
expression F (y) = F (y) is well-defined (i.e. independent of the choice of representative y) and hence defines an
endomorphism F of A/ ∼ which, moreover, satisfies

hF (y1 ), y2 ig = −hy1 , F (y2 )ig .

In other words F is a skew-symmetric endomorphism in the riemannian vector space A/ ∼. By Corollary 2.4.1
(here we use that the dimension of A/ ∼ is at least two) there exists a pair of orthogonal and g-unit vectors
{e1 , e2 } satisfying

F (e1 ) = a e2 , F (e2 ) = −a e1 , a ∈ R.

Select representatives e1 ∈ e1 and e2 ∈ e2 . In terms of F , the condition (2) and the fact that k is eigenvector
require the existence of constants σ, a, λ1 and λ2 such that

F (k) = σk, F (e1 ) = ae2 + λ1 k, F (e2 ) = −ae1 + λ2 k.

Whenever a2 + σ 2 6= 0 the vectors

1 1
u := e1 − (aλ2 + σλ1 ) k, v := e2 + (aλ1 − σλ2 ) k
a2 + σ 2 a2 + σ 2
satisfy F (u) = av and F (v) = −au. Since u and v are spacelike, unit and orthogonal to each other the claim
of the proposition follows (with µ = a). If σ = a = 0, then either λ1 = λ2 = 0 and then {e1 , e2 } are directly
the vectors {u, v} claimed in the proposition (with µ = 0), or at least one of the λs (say λ2 ) is not zero. Then
e := e1 − λλ12 e2 is a spacelike eigenvector of F .
Now we have all the ingredients to show one of the main results of this section, that will eventually allow us
to classify skew-symmetric endomorphisms of Lorentzian vector spaces.
Proposition 2.1. Let V be a Lorentzian vector space of dimension at least five and F ∈ SkewEnd (V ). Then,
there exists a spacelike eigenplane.
Proof. We examine each one of the three possibilities described in Lemma 2.4. Case (iii) yields the result
trivially, so we can assume that F has an eigenvector x.
If we are in case (ii), the vector x is either spacelike or timelike. If it is timelike we consider the riemannian
space x⊥ where F restricts. We may apply Corollary 2.4.1 (note that x⊥ has dimension at least four) and
conclude that the vectors {u, v} exist. So it remains to consider the case when x is spacelike and F admits no
timelike eigenvectors. We restrict to x⊥ which is Lorentzian and of dimension at least four. Applying again
Lemma 2.4, either there exists a spacelike eigenplane, or a second eigenvector y ∈ x⊥ , which can only be
spacelike or null. If y is spacelike, {u := x, v := y} span a spacelike eigenplane with µ = 0. If y is null, we
may apply Lemma 2.5 to F |x⊥ to conclude that either a spacelike eigenplane exists, or there is a spacelike
eigenvector e ∈ x⊥ , so the pair {u := e, v := x} satisfies (1) with µ = 0. This concludes the proof of case (ii).

8
 In case (i), i.e. when there is a null eigenvector x we can apply Lemma 2.5 and conclude that either {u, v}
exist, or there is a spacelike eigenvector e ∈ V , in which case we are into case (ii), already solved. This completes
the proof.

Proposition 2.1 provides the basic tool to classify systematically skew-symmetric endomorphisms if the
dimension d is at least five. The idea is to start looking for a first spacelike eigenplane Π. Then, we restrict to
Π⊥ , that is Lorentzian of dimension d − 2. If d − 2 ≥ 5, Proposition 2.1 applies again and we can keep on going
until we reach a subspace of dimension three if d odd or dimension four if d even. Therefore, for a complete
classification it only remains to solve the problem in three and four dimensions. This has already been done
in [24], where a canonical form based on the classification of skew-symmetric endomorphisms is introduced.
The results from [24] that we shall need are summarized in Proposition 2.2 and Corollary 2.5.1 and their main
consequences in the present context are discussed in Remarks 2.2 and 2.3 below, where we also relate the
canonical form with the classification of skew-symmetric endomorphisms. For a proof and extended discussion,
we refer the reader to [24]. In the remainder, when we explicitly write a matrix of entries F α β , where α is the
d−1
row and β the column, we refer to a linear transformation expressed in a vector basis {eα }α=0 acting on the
vectors v = v α eα ∈ V by
F (v) = F α β v β eα .
Proposition 2.2. For every non-zero F ∈ SkewEnd (V ), with V Lorentzian four-dimensional, there exists an
orthonormal basis B := {e0 , e1 , e2 , e3 }, with e0 timelike future directed, into which F is
 
0 0 −1 + σ4 τ
4
 0 0 −1 − σ4 − τ4 
F =  −1 + σ 1 + σ
, σ, τ ∈ R, (7)
4 4 0 0 
τ τ
4 4 0 0

where σ := − 12 TrF 2 and τ 2 := −4 det F , with τ ≥ 0. Moreover, if τ = 0 the vector e3 can be taken to be any
spacelike unit vector lying in the kernel of F .
Corollary 2.5.1. For every non-zero F ∈ SkewEnd (V ), with V Lorentzian three-dimensional, there exists an
orthonormal basis B := {e0 , e1 , e2 }, with e0 timelike future directed, into which F is
 
0 0 −1 + σ4

1
F = 0 0 −1 − σ4  , σ := − Tr F 2 ∈ R. (8)
2
−1 + σ4 1 + σ4 0

Remark 2.2. A classification result follows because only two exclusive possibilities arise:
1. If either σ or τ do not vanish, F has a timelike eigenplane and an orthogonal spacelike eigenplane with
respective eigenvalues
p p p
µt := (−σ + ρ)/2 and µs := (σ + ρ)/2 for ρ := σ 2 + τ 2 ≥ 0. (9)

The inverse relation between µt , µs and σ, τ is σ = µ2s − µ2t and τ = 2µt µs .

2. Otherwise, σ = τ = 0 if and only if ker F is degenerate two-dimensional. Equivalently, F has null
eigenvector orthogonal to a spacelike eigenvector both with vanishing eigenvalue.

9
One can easily check that when τ = 0, the sign of σ determines the causal character of ker F , namely σ < 0
if ker F is spacelike, σ = 0 if ker F is degenerate and σ > 0 if ker F is timelike. Obviously, τ 6= 0 implies
ker F = {0}. The characteristic polynomial of F is directly calculated from (7)
PF (x) = (x2 − µ2t )(x2 + µ2s ).
Remark 2.3. For a classification result in the three dimensional case, one can see by direct calculation that
q := (1 + σ/4)e0 + (1 − σ/4)e1 generates ker F and furthermore hq, qi = −σ. Hence, the sign of σ determines the
causal character of ker F , namely it is spacelike if σ < 0, degenerate if σ = 0 and timelike
p if σ > 0. Moreover,
when σ 6= 0, F has an eigenplane with opposite causal character than q and eigenvalue |σ|. The characteristic
polynomial of F reads
PF (x) = x(x2 + σ).
We have now all the necessary ingredients to give a complete classification of skew-symemtric endomor-
phisms of Lorentzian vector spaces. In what follows we identify Lorentzian (sub)spaces of d-dimension with the
Minkowski space M1,d−1 . Also, for any real number x ∈ R, [x] ∈ Z denotes its integer part.
Theorem 2.6 (Classification of skew-symmetric endomorphisms in Lorentzian spaces). Let F ∈ SkewEnd (V )
with V Lorentzian of dimension d > 2. Then V has a set of [ d−1 2 ] − 1 mutually orthogonal spacelike eigenplanes
{Πi }, i = 1, · · · , [ d−1
2 ] − 1, so that V admits one of the following decompositions into direct sum of F -invariant
subspaces:
a) If d even V = M1,3 ⊕ Π d−4 ⊕ · · · ⊕ Π1 and either F |M1,3 = 0 or otherwise one of the following cases holds:
2

a.1) F |M1,3 has a spacelike eigenvector e orthogonal to a null eigenvector with vanishing eigenvalue and then
M1,3 = M1,2 ⊕ span{e}.
a.2) F |M1,3 has a spacelike eigenplane Π d−2 (as well as a timelike eigenplane M1,1 orthogonal to Π d−2 ) and
2 2
then M1,3 = M1,1 ⊕ Π d−2 .
2

1,2
b) If d odd V = M ⊕ Π d−3 ⊕ · · · ⊕ Π1 and either F |M1,2 = 0 or otherwise one of the following cases holds:
2

b.1) F |M1,2 has a spacelike eigenvector e and then M1,2 = M1,1 ⊕ span{e}.
b.2) F |M1,2 timelike eigenvector t and then M1,3 = span{t} ⊕ Π d−1 .
2

b.3) F |M1,2 has a null eigenvector with vanishing eigenvalue.

Proof. The proof is a simple combination of the previous results. First, if d ≥ 5, we can apply Proposition 2.1
to obtain the first spacelike eigenplane Π1 . Then Π⊥ 1 is Lorentzian of dimension d − 2. If d − 2 ≥ 5, we can
apply again Proposition 2.1 to obtain a second eigenplane Π2 . Continuing with this process, depending on d,
two things can happen:


a) If d even, we get d−4 2 = [ d−1
2 ] − 1 spacelike eigenplanes, until we eventually reach a Lorentzian vector
subspace of dimension four, M1,3 , where Proposition 2.1 cannot be applied. In M1,3 , either F |M1,3 = 0 or
otherwise cases a.1) and a.2) follow from Remark 2.2, cases 2 and 1 respectively.


b) If d odd, we get d−3 2 = [ d−1
2 ] − 1 spacelike eigenplanes, until we reach a Lorentzian vector subspace of
dimension three, M1,2 . In M1,2 , either F |M1,2 = 0 or by Remark 2.3 there exists a unique eigenvector σ with
vanishing eigenvalue. If σ null, case b.3) follows. If it is spacelike e := σ, F restricts to e⊥ = M1,1 ⊂ M1,2
and b.1) follows. If σ timelike, the same argument applies with t := σ and t⊥ ⊂ M1,2 defines the remaining
spacelike plane Π d−1 .
2

10
3 Canonical form for skew-symmetric endomorphisms
Our aim here is to extend the results in Proposition 2.2 and Corollary 2.5.1 to arbitrary dimensions. To do
that, we will employ the classification Theorem 2.6 derived in Section 2, from which it immediately follows a
decomposition of any F ∈ SkewEnd (V ) into direct sum of skew-symmetric endomorphisms of the subspaces
that F restricts to, namely

[ d−1
2 ]−1
M
F = F |M1,3 F |Π i if d even, (10)
i=1
[ d−1
2 ]−1
M
F = F |M1,2 F |Π i if d odd, (11)
i=1

where Πi are spacelike eigenplanes. In what follows, we will denote

p := [(d − 1)/2] − 1.

Notice that the blocks F |M1,3 and F |M1,2 may also admit different subdecompositions depending on the case,
but our purpose is to remain as general as possible, so we leave this part unaltered. It will be convenient for
the rest of the paper to give a name to the decompositions (10) and (11):
Definition 3.1. Let F ∈ SkewEnd (V ) non-zero for V Lorentzian d-dimensional. Then, a decomposition of the
form (10) or (11) is called block form of F . A basis that realizes a block form is called block form basis.
Writing F in block form form allows us to work with F as a sum of skew-symmetric endomorphisms of
riemmanian two-planes plus one skew-symmetric endomorphism of a three or four dimensional Lorentzian
vector space. For the latter we will employ the canonical forms in Proposition 2.2 and Corollary 2.5.1, and for
the former, it is immediate that in every (suitably oriented) orthonormal basis of Πi
 
0 −µi
F |Π i = , 0 ≤ µi ∈ R. (12)
µi 0

Having defined a canonical form for four, three and two dimensional endomorphisms (i.e. matrices (7), (8)
and (12) respectivley), the idea is to extend this result to arbitrary dimensions finding a systematic way to
construct a block form (10), (11) such that each of the blocks are in canonical form. This is not immediate,
firstly, because the block form does not require the blocks F |M1,3 or F |M1,2 to be non-zero and secondly, because,
unlike in the four and three dimensional cases, the parameters σ, τ of the four and three dimensional blocks
cannot be invariantly defined as, for example, traces of F 2 or determinant of F . The first of these concerns is
easily solved by suitably choosing a block form:
Lemma 3.1. Let F ∈ SkewEnd (V ) be non-zero for V Lorentzian of dimension d. Then there exists a block
form (10) and (11) such that F |M1,3 and F |M1,2 are non-zero and they either contain no spacelike eigenplanes
or they contain one with largest eigenvalue (among all spacelike eigenplanes of F ). In addition, the rest of
spacelike eigenplanes Πi are sorted by decreasing value of µ2i , i.e. µ21 ≥ µ22 ≥ · · · ≥ µ2p .
Proof. If ker F is degenerate, it must correspond with cases a.1) (d even) or b.3) (d odd) of Theorem 2.6. Hence,
in any block form the blocks F |M1,3 and F |M1,2 are non-zero and they do not contain any spacelike eigenplane,

11
as claimed in the lemma. So let us assume that ker F is non-degenerate or zero, which discards cases a.1) and
b.3) of Theorem 2.6. In all possible cases, any block form admits the following splitting in

F |M1,3 = F |Πt ⊕ F |Πs , F |M1,2 = F |span{v} ⊕ F |v⊥ , (13)

with Πs , Πt spacelike and timelike eigenplanes with (possibly zero) respective eigenvalues µs and µt , v a timelike
or spacelike eigenvector (in ker F ) and v ⊥ ⊂ M1,2 an eigenplane with opposite causal character than v. If v is
spacelike, then either F |v⊥ is non-zero, in which case F |M1,2 6= 0 and clearly contains no spacelike eigenplanes
(which is one of the possibilities in the lemma), or F |v⊥ = 0 and then F |M1,2 = 0, so we can rearrange the
decomposition (13) using some timelike vector v ′ ∈ v ⊥ instead of v, i.e. F |M1,2 = F |span{v′ } ⊕ F |v′⊥ . Hence,
in the case of d odd, we may assume that v is timelike and v ⊥ ⊂ M1,2 is a spacelike eigenplane. Let Πµ be a
spacelike eigenplane of F with largest eigenvalue µ among Πs (d even) or v ⊥ (d odd) and Π1 , · · · , Πp . Then,
switching F |Πs or F |Π ⊥ by F |Πµ we construct
v

F̂ |M1,3 := F |Πt ⊕ F |Πµ , F̂ |M1,2 := F |span{v} ⊕ F |Πµ .

The resulting matrix is still in block form and has non-zero blocks F̂ , F̂ containing a spacelike
M1,3 M1,2
eigenplane with largest eigenvalue, which is the other possibility in the lemma. The last claim follows by simply
rearranging the remaining spacelike eigenplanes Πi by decreasing order of µ2i .
With a skew-symmetric endomorphism F in the block form given in Lemma 3.1 we can take each of the
blocks to its respective canonical form. Let us denote Fστ := F |M1,3 (if d even), Fσ := F |M1,2 (if d odd) and
Fµi := F |Πi when written in the canonical forms (7), (8) and (12) respectively. Consequently
p
M p
M
F = Fστ Fµi (d even), F = Fσ Fµi (d odd), (14)
i=1 i=1

where, notice, each of the blocks is written in an orthonormal basis of the corresponding subspace, which
moreover is future directed if the subspace is Lorentzian, i.e. M1,3 or M1,2 (c.f. Proposition 2.2 and Corollary
2.5.1). Hence, the form given in (14) corresponds to a future directed, orthonormal basis of M1,d−1 .
Our aim now is to give an invariant definition of σ, τ, µi . A possible way to do this is through the eigenvalues
of F 2 . One may wonder why not to use directly the eigenvalues of F . One reason is that since we are interested
in real Lorentzian vector spaces V (although, for practical reasons, we may rely on the complexification VC for
some proofs), it is more consistent to give our canonical form in terms of real quantities, while the eigenvalues
of F may be complex. In addition, the canonical form will require to sort them in some way, for which using
real numbers is better suited.
The characteristic polynomial of F is known (e.g. [23]) to possess the following parity:

PF (x) = (−1)d PF (−x). (15)

Thus, a simple calculation relates the characteristic polynomials of F and F 2

√
√

PF 2 (x) = det(xIdd − F 2 ) = det xIdd − F det xIdd + F
√ √ √
2 (16)
= (−1)d PF ( x)PF (− x) = PF ( x) ,
√
x being any of the square roots of x in C and Idd the d × d identity matrix. We can extract some conclusions
from (16):

12
Lemma 3.2. Let F ∈ SkewEnd (V ) for V Lorentzian of dimension d. Then the non-zero eigenvalues of F 2
have even multiplicity ma and the zero eigenvalue has multilplicity m0 with the parity of d. In addition, F
possesses pa (resp. exaclty one) spacelike (resp. timelike) eigenplanes with eigenvalue µ 6= 0 if and only if F 2
has a negative (resp. positive) non-zero eigenvalue −µ2 (resp. µ2 ) with multiplicity ma := 2pa (resp. exactly
two).
Proof. It is an immediate consequence of equation (16) that non-zero eigenvalues of F 2 must have even multi-
plicity ma . Moreover, since the sum of all multiplicites adds up to the dimension d, the multiplicity of the zero
m0 has the parity of d.
Combining Lemma 2.2 and equation (16), F has a spacelike eigenplane Π with non-zero eigenvalue µ if and
only if F 2 has a negative double1 eigenvalue −µ2 . If d ≤ 4, there

cannot be any other spacelike eigenplanes in
Π⊥ , so applying the same argument to F |Π⊥ ∈ SkewEnd Π⊥ , the multiplicity ma of −µ2 must be ma = 2.
If d > 4 and ma ≥ 4, then −µ2 is an eigenvalue of ( F |Π⊥ )2 with multiplicity ma − 2, thus F has a second
spacelike eigenplane with eigenvalue µ in Π⊥ . Repeating this argument, F 2 has a negative eigenvalue −µ2 with
multilplicity ma if and only if F has pa = ma /2 spacelike eigenplanes with eigenvalue µ.
Finally, by Lemma 2.3 and equation (16), F has a timelike eigenplane Π with non-zero eigenvalue µ if and
only if F 2 has a positive double eigenvalue µ2 . Obviously, the maximum number of timelike eigenplanes that
F can have is one. Thus, F |Π⊥ cannot have timelike eigenplanes and hence ( F |Π⊥ )2 has no additional positive
eigenvalues. Consequently, the multiplicity of µ2 is exactly two.

Taking into account Lemma 3.2, we will employ the eigenvalues of −F 2 rather than those of F 2 , so we
assign positive eigenvalues of F 2 with spacelike eigenplanes and negative eigenvalues to timelike eigenplanes.
This amounts to employ the roots of the characteristic polynomial PF 2 (−x).
We now discuss how to invariantly define the parameters σ, τ, µi for d even and σ, µi for d odd. The result
of the argument is formalized below in Definition 3.2. Recall that the characteristic polynomial of a direct sum
of two or more endomorphisms is the product of their individual characteristic polynomials, in particular, the
characteristic polynomial of −F 2 equals to the product of the characteristic polynomials of −Fστ
2
or −Fσ2 times
2
those of each −Fµi (c.f. equation (14)). Let us define:
 1/2
1/2 PF 2 (−x)
QF 2 (x) := (PF 2 (−x)) (d even), QF 2 (x) := (d odd), (17)
x

Starting with d even, from formula (14) it is immediate that µ2i arepdouble roots of PF 2 (−x2 ), p which by Lemma
3.1 satisfy µ21 ≥ · · · ≥ µ2p ≥ 0. On the other hand, let µt := (−σ + ρ)/2 and iµs := i (σ + ρ)/2 with
√
ρ := σ 2 + τ 2 ≥ 0, that by Remark 2.2,  are roots of PFστ (x), thus roots of PF (x) . By equation (16), −µ2t , µ2s
are double roots of PF 2 (−x). The set −µt , µ2s , µ21 , · · · , µ2p are in total p + 2 = [(d − 1)/2] + 1 = d/2 elements,
2

each of which is a double root of PF 2 (−x). In other words, −µ2t , µ2s , µ21 , · · · , µ2p is the set of all roots of the
polynomial2 QF 2 (x). If ker F is degenerate, then ker Fστ is degenerate and by Remark 2.2 it must happen
µt = µs = 0. Hence µ21 ≥ µ22 ≥ · · · µ2p ≥ µ2s = −µ2t = 0. Otherwise, also by Remark 2.2, Fστ contains a
spacelike eigenplane with eigenvalue µs (which by Lemma 3.1 is the largest) as well as a timelike eigenplane
with eigenvalue µt . In this case µ2s ≥ µ21 ≥ · · · µ2p ≥ 0 ≥ −µ2t .
2 2
We next discuss σ, µi for d odd. Again, from (14) we have √ that µi are double roots of PF 2 (−x ), which by
Lemma 3.1 also satisfy µ21 ≥ · · · ≥ µ2p ≥ 0. By Remark 2.3, σ is a root of PFσ (x), thus a root of PF (x), so by
1 We adopt the convention that a root with multiplicity m ≥ 2 is also double
2Q
F 2 (x)
is a polynomial because all the roots of PF 2 (−x) are double.

13
formula (16), σ is a double root of PF 2 (−x). Also, PF 2 (−x) has at least one zero root and hence, PF 2 (−x)/x
is a polynomial with d − 1 roots (counting multiplicity). Then, the set σ, µ21 , · · · , µ2p are all double roots of
PF 2 (−x)/x, which
 are p + 1 = [(d − 1)/2] = (d − 1)/2 elements. Therefore QF 2 as defined in (17) is also a
polynomial and σ, µ21 , · · · , µ2p is the set of all its roots. If ker F is timelike, then ker Fσ is timelike, which
p
happens if and only if σ > 0 (c.f. Remark 2.3) and also Fσ has a spacelike eigenplane with eigenvalue |σ|,
that by Lemma 3.1 is the largest eigenvalue among spacelike eigenplanes. Thus σ ≥ µ21 ≥ · · · ≥ µ2p . In the case
ker F not timelike, the inequalities become µ21 ≥ · · · ≥ µ2p ≥ 0 ≥ σ.
Summarizing, the paramaters σ, τ, µi correspond to the set of all roots of QF 2 sorted in a certain order fully
determined by the causal character of ker F . This allows us to put forward the following definition:
Definition 3.2. Let Roots (QF 2 ) denote the set of roots of QF 2 (x) repeated as many times as their multiplicity.
Then

a) If d odd, σ; µ21 , · · · , µ2p := Roots (QF 2 ) sorted by σ ≥ µ21 ≥ · · · ≥ µ2p if ker F is timelike and µ21 ≥ · · · ≥
µ2p ≥ 0 ≥ σ otherwise.

b) If d even, σ := µ2s − µ2t , τ := 2|µt µs | with −µ2t , µ2s ; µ21 , · · · , µ2p := Roots (QF 2 ) sorted by µ21 ≥ · · · ≥
µ2p ≥ µ2s = −µ2t = 0 if ker F is degenerate and µ2s ≥ µ21 ≥ · · · ≥ µ2p ≥ 0 ≥ −µ2t otherwise.
In addition, we also summarize the results concerning the canonical form in the following Theorem:
Theorem 3.3. Let F ∈ SkewEnd (V ) non-zero, with V Lorentzian of dimension d ≥ 3 and p := [(d − 1)/2] − 1.
Then there exists an orthonormal, future oriented basis such that F is given (14) where Fστ := F |M1,3 , Fσ :=
F |M1,2 , Fµi := F |Πi are given by (7), (8), (12) respectively and σ, τ, µi are given in Definition 3.2. In particular,
Fστ , Fσ are non-zero and they either do not contain a spacelike eigenplane or they contain one with maximal
eigenvalue (among all spacelike eigenplanes of F ) and the eigenvalues µi are sorted by µ21 ≥ µ22 ≥ · · · µ2p .
Definition 3.3. For any F ∈ SkewEnd (V ), for V Lorentzian d-dimensional, the form of F given in Theorem
3.3 is called canonical form and the basis realizing it is called canonical basis.
The first and obvious reason why the canonical form is useful is that it allows one to work with all elements
F ∈ SkewEnd (V ) at once. The fact that we can give a canonical form for every element without splitting into
cases is a great strenght, since we can perform a general analysis just in terms of the parameters that define the
canonical form. Moreover, as we will show in Section 5, this form is the same for all the elements in the orbit
generated by the adjoint action of the orthochronous Lorentz group O+ (1, d − 1). Thus, the canonical form is
specially suited for problems with O+ (1, d − 1) invariance (or covariance) which, as discussed in Section 6, is
directly related to certain conformally covariant problems in general relativity.
We finish this section with two corollaries that will be useful later. The first one is trivial from the canonical
form (14)
Corollary 3.3.1. The characteristic polynomial of F ∈ SkewEnd (V ) is
p
Y p
Y
PF (x) = (x2 − µ2t )(x2 + µ2s ) (x2 + µ2i ) (d even), PF (x) = x(x2 + σ) (x2 + µ2i ) (d odd), (18)
i=1 i=1
√ √
where −2µ2t := σ − σ 2 + τ 2 , 2µ2s := σ + σ 2 + τ 2 .
The second gives a formula for the rank of F . We base our proof in the canonical form (14) because it is
straightforward. However, we remark that this corollary can also be regarded as a consequence of Theorem 2.6.

14
Corollary 3.3.2. Let F ∈ SkewEnd (V ), with V Lorentzian of dimension d and m0 the multiplicity of the zero
eigenvalue. Then, only of the following exclusive cases hold:
a) ker F is non-degenerate or zero if and only if rank F = d − m0 .
b) ker F is degenerate if and only if m0 > 2 and rank F = d − m0 + 2.
Proof. Consider F in canonical form (14) and let k ∈ N be the number of parameters µi that vanish. For d even
we have dim ker F = 2k + dim ker Fστ . On the one hand, ker F degenerate implies ker Fστ degenerate, which
by Remark 2.2 happens if and only if σ = τ = 0 and in addition dim ker Fστ = 2. Therefore dim ker F = 2k + 2
and by (18), m0 = 2k + 4 (> 2). Thus rank F = d − dim ker F = d − m0 + 2. On the other hand, ker F
non-degenerate if at most one of σ or τ vanish. If τ 6= 0 (so that µs 6= 0 and µt 6= 0), dim ker Fστ = 0 and
m0 = 2k = dim ker F . Consequently rank F = d− m0 . If τ = 0 (and σ 6= 0, so that exactly one of µs , µt vanish),
by Remark 2.2 dim ker Fστ = 2 and by (18) m0 = 2k + 2. Hence dim ker F = 2k + 2 and rank F = d − m0 .
For d odd, we have dim ker F = 2k + dim ker Fσ = 2k + 1, because dim ker Fσ = 1 (c.f. Remark 2.3). ker F
is degenerate if and only if ker Fσ is degenerate, which by Remark 2.3 occurs if and only if σ = 0. Hence, by
equation (18), m0 = 2k + 3 (> 2) and rank F = d − dim ker F = d − m0 + 2. For the ker F non-degenerate case,
σ 6= 0 and also by (18) m0 = 2k + 1 = dim ker F . Therefore rank F = d − m0 .

4 Simple endomorphisms
By simple skew-symmetric endomorphism we mean a G ∈ SkewEnd (V ) satisfying rank G = 2. As usual
e♭ ≡ he, ·i is the one-form obtained by lowering index to a vector e ∈ V . Then, a simple skew-symmetric
endomorphism can be always written as
G = e ⊗ v♭ − v ⊗ e♭
for two linearly independent vectors e, v ∈ V and its action on any vector w ∈ V is

G(w) = hv, wi e − he, wi v.

Since the two-fom associated to a simple endomorphism is G = e♭ ∧ v♭ , it follows from elementary algebra that
two simple skew-symmetric endomorphisms G = e ⊗ v♭ − v ⊗ e♭ and G′ = e′ ⊗ v♭′ − v ′ ⊗ e′♭ are proportional if
and only if span{e, v} = span{e′ , v ′ }. This freedom in the pair {e, v} defining G can be used to choose them
orthogonal.
Lemma 4.1. Let G ∈ SkewEnd (V ) be simple. Then there exist two non-zero orthogonal vectors e, v ∈ V such
that G = e ⊗ v♭ − v ⊗ e♭ with v spacelike.
Proof. By definition G = ẽ ⊗ ṽ♭ − ṽ ⊗ ẽ♭ for two linearly indepedent vectors ẽ, ṽ ∈ V . If one of them is non-null,
we set ṽ := v and decompose V = span{v} ⊕ v ⊥ . Thus ẽ = av + e with a ∈ R and e ∈ v ⊥ and G takes the form
G = (av +e)⊗v♭ −v ⊗(av +e)♭ = e⊗v♭ −v ⊗e♭ , as claimed. If ẽ and ṽ are both null, consider V = span{ẽ} ⊕ e (ẽ)c
(we use ⊕e because this direct sum is not by orthogonal spaces) where (ẽ) is a spacelike complement of span{ẽ}.
c

Then we can write ṽ = aẽ + v ′ , with a ∈ R and v ′ ∈ ẽc non-null. Thus G = ẽ ⊗ v♭′ − v ′ ⊗ ẽ♭ , with v ′ non-null
and we fall into the previous case. All in all, G = e ⊗ v♭ − v ⊗ e♭ with e, v orthogonal. Consequently, either one
of the vectors is spacelike or both are null and proportional which would imply G = 0, against our hypothesis
rank G = 2.

15
 The decomposition G = e ⊗ v♭ − v ⊗ e♭ is not unique even with the restriction of v being spacelike unit and
orthogonal to e. One can easily show that the remaining freedom is given by the transformation e′ = ae−b he, ei v,
v ′ = be + av with a, b ∈ R restricted to a2 + b2 he, ei = 1. Nevertheless, the square norm he′ , e′ i is invariant
under this change, so the following definition makes sense:
Definition 4.1. Let G ∈ SkewEnd (V ) be simple, with G = e ⊗ v♭ − v ⊗ e♭ , e, v ∈ V orthogonal with v spacelike
unit. Then G is said to be spacelike, timelike or null if the vector e is spacelike, timelike or null respectively. In
the non-null case, G is called spacelike (resp. timelike) unit whenever he, ei = +1 (resp. he, ei = −1).
By Lemma 4.1, it is immediate that Definition 4.1 comprises any possible simple endomorphism (up to a
multiplicative factor).
We next obtain the necessary and sufficient conditions for a simple endomorphism G to commute with a
given F ∈ SkewEnd (V ). We first make the simple observation that the composition of a one-form e♭ and a
skew-symmetric endomorphism F satisfies (simply apply for sides to any w ∈ V )

e♭ ◦ F = −F (e)♭ .

An immediate consequence is that for any pair of vectors e, v ∈ V and F ∈ SkewEnd (V ) it holds

F ◦ (e ⊗ v♭ ) = F (e) ⊗ v♭ , (e ⊗ v♭ ) ◦ F = −e ⊗ F (v)♭ (19)

The following commutation result will be used later.

Lemma 4.2. Let F, G ∈ SkewEnd (V ) with G = e ⊗ v♭ − v ⊗ e♭ simple and e, v ∈ V as in Definition 4.1. Then
[F, G] = 0 if and only if there exist µ ∈ R such that:

F (e) = he, ei µv, F (v) = −µe. (20)

Proof. The commutator is [F, G] = F ◦ G − G ◦ F

[F, G] = F ◦ G − G ◦ F = F ◦ (e ⊗ v♭ − v ⊗ e♭ ) − (e ⊗ v♭ − v ⊗ e♭ ) ◦ F
= F (e) ⊗ v♭ − F (v) ⊗ e♭ + e ⊗ F (v)♭ − v ⊗ F (e)♭ , (21)

where we have used (19). The “if” part is obtained by direct calculation inserting (20) in (21). To prove the “only
if” part, the condition [F, G] = 0 requires the two endomorphisms F (e) ⊗ v♭ − v ⊗ F (e)♭ and F (v) ⊗ e♭ − e ⊗ F (v)♭
to be equal. One such endomorphism is either identically zero or simple. This implies that span{F (e), v} and
span{e, F (v)} are either both one dimensional or both two-dimensional and equal. In the first case, F (v) = −µe
and F (e) = αv for µ, α ∈ R, which are determined by skew-symmetry to satisfy α = µ he, ei, so the lemma
follows. The second case is empty, for it is necessary that v = ae + bF (v) with a, b ∈ R, which implies
hv, vi = hae + bF (v), vi = b hF (v), vi = 0, against the hypothesis of v being spacelike.
Corollary 4.2.1. Let G, G′ ∈ SkewEnd (V ) be simple, spacelike and linearly independent. Let {e, v}, {e′ , v ′ }
be orthogonal spacelike vectors such that G = e ⊗ v♭ − v ⊗ e♭ and G′ = e′ ⊗ v♭′ − v ′ ⊗ e′♭ . Then [G, G′ ] = 0 if and
only if {e, v, e′ , v ′ } are mutually orthogonal.
Proof. By the previous lemma [G, G′ ] = 0 if and only if there exist µ ∈ R such that

G(e′ ) = he′ , vi e − he′ , ei v = µv ′ , G(v ′ ) = hv ′ , vi e − hv ′ , ei v = −µe′ . (22)

If µ 6= 0, then span{e, v} = span{e′ , v ′ } and G and G′ are proportional, agains hypothesis. Thus, µ = 0 and by
(22) the set {e, v, e′ , v ′ } is mutually orthogonal.

16
5 O+ (1, d − 1)-classes
In this section we use the canonical form of Section 3 to characterize skew-symmetric endomorphisms of V under
the adjoint action of the orthochronous Lorentz group O+ (1, d−1). Recall that this is the subgroup of O(1, d−1)
preserving time orientation. The corresponding classes of skew-symmetric endomorphisms are also known as
the adjoint orbits or conjugacy classes and we denote them by [F ]O+ for a given element F ∈ SkewEnd (V ).
The charaterization of these orbits by a set of independent invariants is known and it can be found in [23] in
terms of two-forms and other references such as [5]. What we do here is, first, to give an alternative way to
characterize the orbits [F ]O+ by a convenient set of invariants and second, to show that the canonical form
is the same for every element in a given orbit. This makes the canonical form specially useful as a tool for
problems with O+ (1, d − 1) invariance.
Although we restrict here to the orthochronous component O+ (1, d−1) because of its relation with conformal
transformations of the sphere Sd−2 (see Section 6), from this case, the orbits of the full group O(1, d − 1) are
easy to determine. Recall that the time-reversing component O− (1, d − 1) is one-to-one with O+ (1, d − 1). In an
orthonormal basis, we can map elements Λ− ∈ O− (1, d−1) to elements in Λ+ ∈ O+ (1, d−1) by e.g. Λ+ := Λ− η,
where η = diag(−, +, · · · , +) is an element in O− (1, d − 1) with the same matrix form as the metric. Then

Λ+ F (Λ+ )−1 = Λ− ηF η(Λ− )−1 = −Λ− F (Λ− )−1 ,

where the last equality follows from skew-symmetry. Hence, the elements F and −F belong to the same orbit
under the action of O(1, d − 1). If we denote the space of orbits of the orthochronous component and full group
respectively by [O+ ] = SkewEnd (V )/O+ (1, d − 1) and [O] = SkewEnd (V )/O(1, d − 1), this is expressed as
[O] = [O+ ]/Z2 . This fact is of course well-known and can be inferred from general references e.g. [18].
A consequence of equation (15) is that the characteristic polynomial of F ∈ SkewEnd (V ) must have the
form
Xq
PF (x) = xd + cb xd−2b , (23)
b=1

where we have introduced q := [ d2 ].

The coefficients cb can be obtained using the Fadeev-LeVerrier algorithm,
summarized by the following matrix determinant [15]:

Tr F 2b − 1 0 ··· 0
Tr F 2 Tr F 2b − 2 ··· 0
1 .. .. ..
cb = . . . .
(2b)!
Tr F 2b−1 Tr F 2b−2 ··· ··· 1
Tr F 2b Tr F 2b−1 ··· ··· Tr F

Since the traces of odd powers vanish by skew-symmetry, the coefficients cb depend on the entries of F only
through the traces of the squared powers of F :
1
Ib = Tr (F 2b ), b = 1, · · · , q.
2
Recall that the adjoint respresentation Ad of a matrix Lie group G is a linear representation of G on its Lie

17
algebra automorphisms Aut(g) given by

Ad : G −→ Aut(g)

g −→ Ad(g) := Adg : g → g .
X → gXg −1

The traces Ib are obviously invariant under the adjoint action of O+ (1, d − 1) and so are the coefficients cb .
Another invariant that plays an important role in the classification of conjugacy classes is the rank of F . Since
this is always even, we denote it by
rank F = 2r,
and clearly r ≤ q. From now we say rank parameter to refer to r. In the following proposition we show that
this set of invariants actually identifies the canonical form.
Proposition 5.1. Let F, Fe ∈ SkewEnd (V ), for V Lorentzian of dimension d. Then the invariants {cb , r} and
cb , re} of F and Fe respectively are equal if and only if their canonical forms given by Theorem 3.3 are the same.
{e
Proof. The “if” part (⇐) is trivial, because the invariants cb , r are independent on the basis, so they can be
calculated in a canonical basis. Hence, same canonical form implies same invariants. For the “only if” part (⇒),
we notice that if the coefficients cb and ecb of PF and PFe are equal, so are their characterisic polynomials, the
multiplicities of their zero eigenvalue and the polynomials QF 2 and QFe2 (equation (17)). Since rank F = rank Fe ,
Corollary 3.3.2 implies that ker F and ker Fe must have the same causal character. The canonical form only
depends on the roots QF 2 and the causal character of ker F through Definition 3.2. Thus, F and Fe must have
the same canonical form.
We now characterize the classes [F ]O+ in terms of the same invariants given in Proposition 5.1. As mentioned
above, this result is known [23], but we give here an alternative and very simple proof based on our canonical
form:
Theorem 5.1. [23] Let F, Fe ∈ SkewEnd (V ), for V Lorentzian of dimension d. Then their invariants {cb , r}
cb , re} are the same if and only if F and Fe are O+ (1, d − 1)-related.
and {e
Proof. The if (⇐) part is immediate, since it is trivial from their definitions that the quantities {cb , r} are
Lorentz invariant. To prove the “only if” (⇒), by Proposition 5.1, F and Fe have the same canonical form in
canonical bases B and Be respectively. By definition (c.f. Theorem 3.3), these bases are unit, future oriented
and orthonormal. Thus, the transformation taking B to B e transforms F into Fe and both must be O+ (1, d − 1)-
related.

Theorem 5.1 stablishes the necessary and sufficient conditions for two endomorphisms to be O+ (1, d − 1)-
related. Combining this result with Proposition 5.1, we find that the canonical form (hence the parameters
σ, µ2i or σ, τ, µ2i ) totally define the equivalence class of skew-symmetric endomorphisms up to O+ (1, d − 1)
transformations. Moreover, we emphasize that this form is the same for every equivalence class, unlike other
canonical (or normal) forms based on the classification of SkewEnd (V ), such as the one in [9], where they seek
irreducibility of the blocks, so they must give two different forms to cover every case.
Next, we discuss some facts about the coefficients of the characteristic polynomial, also stated in [23], where
the proof is only indicated, and which can now be easily proven using the canonical form.

18
Lemma 5.2. Let F ∈ SkewEnd (V ) be non-zero and let 2r = rank F . Then cr > 0, cr = 0, cr < 0 if and only
if ker F is timelike, null or spacelike (or zero) respectively. Moreover, if r < q, cq = cq−1 = · · · = cr+1 = 0.
Proof. Taking into account that the parities of d and m0 are equal (Lemma 3.2), q − [ m20 ] = [ d2 ] − [ m20 ] = d−m0
2 ,
so equation (23) can be rewritten
   d−m0 
q−[m0 /2]
X X 2

PF (x) = xm0 xd−m0 + cb xd−m0 −2b  = xm0 xd−m0 + cb xd−m0 −2b  , (24)
b=1 b=1

where we have explicitly substituted all zero coefficients by extracting the common factor xm0 , thus the remaining
coefficients cb 6= 0 for b = 1, · · · , (d − m0 )/2. By Corollary 3.3.2, ker F degenerate if and only if 2r = d − m0 + 2
and m0 > 2, so the sum in (24) runs up to (d − m0 )/2 = r − 1, which means cr = cr+1 = · · · = cq = 0, as stated
in the lemma. Also by Corollary 3.3.2, ker F non-degenerate if and only if 2r = d − m0 . In this case, the sum in
(24) runs up to (d−m0 )/2 = r, hence cr 6= 0 and if r < q, the next coefficients vanish cr+1 = cr+2 = · · · = cq = 0.
In addition cr is the independent term in the polynomial in parentheses. Let µ1 , · · · , µλ be all the non-zero
parameters among the {µi } of the canonical form of F given in (14). By equation (18), cr can be written for d
odd:
cr = σµ21 · · · µ2λ .
Then, the sign of σ determines the sign of cr and, by Remark 2.3, also the causal character of ker Fσ , hence,
the causal character of ker F in accordance with the stament of the lemma. For d even, also from (18) we have

τ2 2
cr = − µ · · · µ2λ < 0 (τ 6= 0), cr = σµ21 · · · µ2λ (τ = 0),
4 1
where the expression for τ = 0 follows because in this case either µt or µs (or both) vanish, hence either
cr = µ2s µ21 · · · µ2λ or cr = −µ2t µ21 · · · µ2λ and σ equals µ2s in the first situation and −µ2t in the second. By Remark
2.2, when τ 6= 0 we have ker Fστ = {0} and hence ker F is always spacelike or zero and when τ = 0, the causal
character of ker Fστ (and that of ker F ) is determined by the sign of σ in accordance with the statement of the
lemma.

Remark 5.1. A converse version of Lemma 5.2 also holds, in the sense that the number ν of last vanishing
coefficients restricts the allowed rank parameters r. Let ν be defined by ν = 0 if cq 6= 0 and, otherwise, by the
largest natural number satisfying cq = cq−1 = · · · cq−ν+1 = 0. By equation (24) it follows ν = [m0 /2], and since
the dimension d and m0 have the same parity (cf. Lemma 3.2), d − m0 = 2[d/2] − 2[m0 /2] = 2(q − ν) which in
particular shows that ν determines m0 uniquely. If m0 > 2, by Corollary 3.3.2 the rank parameter admits two
possibilities r = {q − ν, q − ν + 1}, each of which determined by the causal character of ker F . If m0 ≤ 2, also
by Corollary 3.3.2 the ker F degenerate case cannot occur and r = (q − ν) is uniquely determined. In particular,
if d = 4, r is always determined by c1 , c2 , because r = 2 happens if and only if ν = 0 and otherwise r = 1 (unless
F is identically zero, in which case r = 0).

5.1 Structure of SkewEnd (V )/O + (1, d − 1)

By Theorem 5.1, the q-tuple (c1 , · · · , cq ) corresponding to the coefficients of the characteristic polynomial of a
skew-symmetric endomorphism, does not suffice to determine a point in the quotient space SkewEnd (V )/O+ (1, d−
1), since generically two ranks are possible (dimensions three and four are an exception). As dicussed in Remark

19
5.1, for a number ν of last vanishing coefficients cb , the allowed rank parameters are r ∈ {q − ν, q − ν + 1},
and r = q − ν + 1 is only possible provided m0 > 2 (in particular, when cq 6= 0 then necessarily r = q). One
says that there is a degeneracy for the value of the rank at certain points in the space of coefficients cb . In the
submanifold {cq = · · · = cq−ν+1 = 0, cq−ν 6= 0}, the possible rank parameters are r ∈ {q − ν, q − ν + 1}. When a
boundary point where the number of last vanishing coefficients increases by exactly one is approached, the rank
parameter may remain equal to q − ν or jump to q − ν − 1 (note that while the coefficients ci are continuous
functions of F , the rank is only lower semicontinuous, e.g. [21]). As we shall see in this section, this behaviour
gives rise to special limit points which make the space of parameters defining the canonical form (i.e. the space
of conjugacy classes) a non-Hausdorf topological space, when endowed with the natural quotient topology. Let
us start by locating these limit points using the canonical form. Degeneracies can only occur in dimensions
d = 5 or larger because in dimension three the rank is two for any non-trivial F and in dimension four the rank
is uniquely determined by the invariants (c.f. Remark 5.1). We thus consider first the case d = 5 and then
extend to all values d ≥ 5. In d = 5 the space of parameters A defining the [F ]O+ classes is (see fig. 1 )

A := (σ, µ2 ) ∈ R × R+ | σ ≥ µ2 if σ > 0 .

Consider a [F ]O+ in the region R+ := σ ≥ µ2 > 0 and let F be a representative of [F ]O+ in a canonical basis
B = {eα }α=0,··· ,4 , that is
 
0 0 −1 + σ4  
 σ  0 −µ
F = 0 0 −1 − 4 ⊕ . (25)
µ 0
−1 + σ4 1 + σ4 0
Let us define the functions C± (x) := x1 ± x4 . Then, the following change of basis to B ′ = {e′α } is well defined in
R+ :
√ √ √ √
e′0 = C+ (µ) (C+ ( √
σ)e0 + C− ( √σ)e1 ) − C− (µ)e4 , e′2 = −e3 , e′4 = C− ( σ)e0 + C+ ( σ)e1
(26)
e′1 = −C− (µ) (C+ ( σ)e0 + C− ( σ)e1 ) + C+ (µ)e4 , e′3 = −e2 .
By direct calculation, F is written in basis B ′ as
 
µ2
0 0 −1 + 4
 √ 
F =

0 0 −1 − µ2 
⊕ √0 − σ
. (27)
2 2
4 σ 0
−1 + µ4 1 + µ4 0
The basis B ′ is non-canonical because µ2 < σ. However, if we vary the parameters so that µ → 0 (keeping σ
unchanged), the matrix (27) becomes canonical (i.e. of the form (14)) in the limit and the class [limµ→0 F ]O+
is given by l1 = (0, σ). On the other hand, F in canonical form (25) also admits a limit µ → 0, which is
also canonical and whose representative [limµ→0 F ]O+ is given by l2 = (σ, 0). Both limits are defined by the
same sequence of points, because the transformation (26) is invertible in R+ . However this sequence has two
different limit points. As a consequence, the space of canonical matrices, and therefore the quotient space
SkewEnd (V )/O+ (1, d − 1), inherits a non-Hausdorff topology.
Something similar happens in the region R− := {σ < 0, µ > 0}. Let F be a representative
p in canonical form
of a point [F ]O+ in this region. Then, F has a timelike eigenplane Πt with eigenvalue |σ| (c.f. Remark 2.3),
a spacelike eigenvector e as well as spacelike eigenplane Πs with eigenvalue µ. Thus V = Πt ⊕ span{e} ⊕ Πs
and there exist a (non-canonical) basis B ′ adapted to this decomposition, into which F takes the form
 p 
 
p0 |σ| 0
0 −µ
F =  |σ| 0 0  ⊕ . (28)
µ 0
0 0 0

20
Keeping µ unchanged, expression (28) has a limit σ → 0, which has a spacelike eigenplane Πs of eigenvalue µ
and it is identically zero on Π⊥ . Hence, ker F is timelike and using Definition 3.2, the canonical form of this
limit limσ→0 F is given by σ ′ = µ2 and µ′ = 0. Thus [limσ→0 F ]O+ is represented by the point l2 = (µ2 , 0). On
the other hand, in a canonical basis (25), F also admits a limit σ → 0, whose class [limσ→0 F ]O+ is obviously
represented by the point l1 = (0, µ2 ).
µ2

PSfrag replacements

l2
R− R0 R+
l1
σ

Figure 1: Representation of SkewEnd (V )/O+ (1, d − 1) in the region A ⊂ R2 . The shadowed region is not included.

The same reasoning can be carried out to arbitrary odd dimension. First, define the regions
(d,0)  (d,0) 
R+ := σ ≥ µ21 ≥ · · · ≥ µ2p > 0 and R− := σ < 0, µ21 ≥ · · · ≥ µ2p > 0

and also the limit regions

(d,0)  (d,1) 
R0 := σ = 0, µ21 ≥ · · · ≥ µ2p > 0 and R+ := σ ≥ µ21 ≥ · · · ≥ µ2p−1 > µ2p = 0 .

Consider representatives F+ and F− (in canonical form) of points (σ + , (µ+ 2 + 2 − − 2 − 2

1 ) , · · · , (µp ) ) and (σ , (µ1 ) , · · · , (µp ) )
(d,0) (d,0)
in the regions R+ and R− respectively. Then F+ has a spacelike eigenplane Π+ +
s with eigenvalue µp as
+ +
√
+
well as timelike eigenvector e and spacelike eigenplane Πt with eigenvalue σ . Restricting to the sub-
space W + = span{e+ } ⊕ Π+ +
t ⊕ Πs we can repeat the procedure followed for the five dimensional case and
(d,1)
conclude that [limµ+p →0
F+ ] has simultaneously limits on the points (σ + , (µ+ 2 + 2
1 ) , · · · , (µp−1 ) , 0) ∈ R+ and
(d,0)
(0, (µ+ 2 + 2
1 ) , · · · , (µp ) ) ∈ R0 . Analogously F− has a spacelike eigenplane Π− −
s with eigenvalue µp as well as
′−
p
−
spacelike eigenvector e and timelike eigenplane Π s with eigenvalue |σ |. Restricting to the subspace W − =
−
′−
Π− −
s ⊕span{e }⊕Π s , the above arguments for the five dimensional case show that [limσ− →0 F− ] has simultaneous
− 2 − (d,1) (d,0)
limits on the points ((µ− 2 2
p ) , (µ1 ) , · · · , (µp−1 ) , 0) ∈ R+ and (0, (µ− 2 − 2
1 ) , · · · , (µp ) ) ∈ R0 . Thus the regions
(d,0) (d,0) (d,1) (d,0)
R+ and R− limit simultaneously with R+ and R0 as µp and σ tend to zero respectively. Indeed, the
(d,1) (d,1) 
same ideas can be applied again to R+ and R− := σ < 0, µ21 ≥ · · · ≥ µ2p−1 > µ2p = 0 , so that they also
(d,1) 
limit simultaneously, as µp−1 and σ go to zero respectively, with R0 := σ = 0, µ21 ≥ · · · ≥ µ2p−1 > µ2p = 0
(d,2)  (d,i)
and R+ := σ > 0, µ21 ≥ · · · ≥ µ2p−1 > µ2p−1 = µ2p = 0 . In general, the regions R± analogously defined,
i.e. where i gives the number of vanishing parameters µp = · · · = µp−i = 0 and the subindex ± gives the sign
(d,i) (d,i+1)
of σ, have simultaneous limits in R0 and R+ , where the subindex 0 stands for vanishing σ.
For the even dimensional case (with d ≥ 6), notice that the canonical form (14) with τ = 0 is equivalent to
the odd dimensional case direct sum with a one dimensional zero endomorphism (of a Riemannian line). Hence,

21
the previous reasoning for odd dimensions  also applies for even dimensions  and τ = 0. For example, consider
in d = 6 dimensions the regions R+ = τ = 0, σ ≥ µ2 > 0 and R− = τ = 0, σ < 0, µ2 > 0 . Then they both
 (d,1) 
assume limit in R0 = τ = 0, σ = 0, µ2 > 0 and R+ = τ = 0, σ > µ2 = 0 . Notice that if we keep τ 6= 0
no degenerate limits of this kind occur. This can be justified as follows. Let µt , µs be defined as in (9). Then,
it can be readily checked that det F = −µ2t µ2s µ2 , so if σ, τ, µ 6= 0, then rank F = 6. If we keep τ 6= 0 (thus both
µs , µt are different from zero), µ 6= 0 and make σ → 0, the limit must have always rank F = 6. Hence, it is not
possible that a limit σ → 0 ends at two points with different rank. Similarly, keeping τ 6= 0, the limit µ → 0
always has rank F = 4 and therefore, µ → 0 limits cannot be degenerate either. The generalization to arbitrary
even dimensions with τ = 0 is also straightforward from the odd dimensional case discussed above, which we
now summarize in the following remark:
Remark 5.2. In the case of d odd, consider the subset of Rq given by


p
A(odd) := { σ, µ21 , · · · , µ2p ∈ R × R+ | µ21 ≥ · · · ≥ µ2p−1 and if σ > 0, σ ≥ µ21 ≥ · · · ≥ µ2p }.

Define also the subsets of A(odd) given by

n o
(d,i)
R+ := (σ, µ21 , · · · , µ2p ) ∈ A(odd) | σ ≥ µ21 ≥ · · · ≥ µ2p−i > µ2p−i+1 = · · · = µ2p = 0 ,
n o
(d,i)
R− := (σ, µ21 , · · · , µ2p ) ∈ A(odd) | σ < 0, µ21 ≥ · · · ≥ µ2p−i > µ2p−i+1 = · · · = µ2p = 0 ,
n o
(d,i)
R0 := (σ, µ21 , · · · , µ2p ) ∈ A(odd) | σ = 0, µ21 ≥ · · · ≥ µ2p−i > µ2p−i+1 = · · · = µ2p = 0 .

(p−i) (d,i)
Then in the quotient topology of SkewEnd (V )/O+ (1, d − 1) the sequences of R± with limit at R0 also
(d,i+1)
have limit at R+ .
In the case of d even, first define µs as in (9) and let A(even) be the subspace of Rq given by:


p
A(even) := { σ, τ, µ21 , · · · , µ2p ∈ R × R+ | µ21 ≥ · · · ≥ µ2p and if τ 6= 0 or σ > 0, µ2s ≥ µ21 ≥ · · · ≥ µ2p }.

Define also the following subsets of A(even)

n o
(d,i)
R+ := (σ, µ21 , · · · , µ2p ) ∈ A(even) | τ = 0, σ ≥ µ21 ≥ · · · ≥ µ2p−i > µ2p−i+1 = · · · = µ2p = 0 ,
n o
(d,i)
R− := (σ, µ21 , · · · , µ2p ) ∈ A(even) | τ = 0, σ < 0, µ21 ≥ · · · ≥ µ2p−i > µ2p−i+1 = · · · = µ2p = 0 ,
n o
(d,i)
R0 := (σ, µ21 , · · · , µ2p ) ∈ A(even) | τ = 0, σ = 0, µ21 ≥ · · · ≥ µ2p−i > µ2p−i+1 = · · · = µ2p = 0 .

(d,i) (d,i)
Then in the quotient topology of SkewEnd (V )/O+ (1, d − 1) the sequences of R± with limit at R0 also have
(d,i+1)
limit at R+ .

6 Conformal vector fields

One interesting applications of our previous results is based on the relation between skew-symmetric endomor-
phisms and the set of conformal Killing vector fields (CKVFs) of the n-sphere, CKill(Sn ). These vector fields
are the generators of the conformal transformations of the n-sphere Conf (Sn ), i.e. the group of transformations

22
 ⋆
ψΛ that scale the spherical metric γ, ψΛ (γ) = Ω2 γ, where Ω is a smooth positive function of Sn . A standard
technique to describe these transformations consists in viewing Sn as the (real) projectivization of the null future
cone in M1,n+1 , in such a way that Conf (Sn ) is induced from the isometries of M1,n+1 . This is discussed in
detail for the four dimensional case in [26] and in arbitrary dimensions in [23] and in [28] (the latter considers
arbitrary signature and the projectivization of the null ”cone” in Mp+1,q+1 , giving Sp × Sq ). This procedure
stablishes a group homomorphism ψ : O(1, n + 1) → Conf(Sn ), Λ 7→ ψΛ , which is one-to-one when restricted
to the orthochronous component O+ (1, n + 1) ⊂ O(1, n + 1).
The Euclidean space En = (Rn , gE ) and Sn are well-known to be conformally related via the stereographic
projection StN : Sn \{N } → En , where N denotes the point w.r.t. which the projection is taken. Observe that
the stereographic projection depends not only on the point N but also on the (signed) distance between N
and the plane onto which the projection is performed. We do not reflect this dependence in the notation for
simplicity.
Hence, the composition of a transformation ψΛ ∈ Conf (Sn ) with the stereographic projection yields StN ◦ψΛ ◦
−1
StN =: φΛ ∈ Conf (En ), which is a conformal transformation of En . Strictly speaking, these transformations are
not diffeomorphisms of En , as they require to remove the two points p1 , p2 ∈ En satisfying ψΛ ◦St−1
N (p1 ) = N and
−1
ψΛ ◦ St−1
N (p 2 ) = N , which are the “preimage” and the “image” of infinity under φΛ respectively. Nevertheless,
since ψ : O(1, n + 1) → Conf(Sn ) is a group homomorphism, so is φ : O(1, n + 1) → Conf(En ), Λ 7→ φΛ as
well as the map which assigns ψΛ 7→ φΛ . In that sense Conf (Sn ) and Conf (E
n ) are the same. These group
homomorphisms, induce Lie algebra homomorphisms between SkewEnd M1,n+1 , CKill(Sn ) and CKill(En ) (the
vector fields generating Conf (En )). The precise form of these maps depends, firstly, on the representative used
to describe the projective cone (i.e. Sn ) and secondly on the point N as well as on the signed distance from
this point to the plane. In [23], the morphism


ξ := φ⋆ : SkewEnd M1,n+1 −→ CKill(En ),
(29)
F 7−→ ξ(F ) =: ξF ,

is constructed3 related to each other using the representative with {x0 = 1} ∩ {xα xα = 0} for the projective
cone, where {xα } (α, β = 0, · · · , n + 1) are Minkowskian coordinates of M1,n+1 , N is the point with coordinates
{x0 = −x1 = 1, xA+1 = 0} (A, B = 1, · · · , n) and the image plane for the stereographic projection is {x0 =
x1 = 1}. The result is a representation of CKill(En ) where the vector vector fields are expressed in Cartesian
coordinates {y A } induced from the Minkowskian coordinates by means of {x0 = x1 = 1, xA+1 = y A }.
Theorem 6.1. [[23]]

Let M1,n+1 endowed with Minkowskian coordinates {xα } and consider any element F ∈
SkewEnd M1,n+1 written in the basis {∂xα } in the form
 
0 −ν −at + bt /2
F =  −ν 0 −at − bt /2 , (30)
−a + b/2 a + b/2 ω
−ω

where a, b ∈ Rn are column vectors, t stands for the transpose and ω is a skew-symmetric n×n matrix (ω ω t ).
ω = −ω
A n n 1,n+1
Then, in the Cartesian coordinates {y } of E defined by the embedding i : E ֒→ M , i(E ) = {x = x1 =
n 0
A+1 A n
1, x = y }, the CKVFs of E are
 
1
ξF = bA + νy A + (aB y B )y A − (yB y B )aA − ω A B y B ∂yA . (31)
2
3 The method in [23] is based on the unit spacelike hyperboloid in Minkowski instead of on the null cone. However, the two

methods are easily seen to be equivalent to the one we describe

23
Moreover, ξAdΛ (F ) = φΛ⋆ (ξF ) for every Λ ∈ O+ (1, n + 1) and ξ is a Lie algebra antihomomorphism, i.e.
[ξF , ξG ] = −ξ[F,G] .
Remark 6.1. For later use, we write explicitly the parameters of the vector field ν, aA , bA , ω A B in terms of the
entries F α β of the endomorphism F :
1

ν = −F 0 1 , aA = − F 0 A+1 + F 1 A+1 ,
2 (32)
1 0 1


bA = F A+1 − F A+1 , ω B = −F A+1 B+1 .
A
2
where capital Latin indices are lowered with the Kronnecker δAB . Unless otherwise stated, ξ without subindex
refers to the map ξ given in (29) while ξF refers to the CKVF which is image under ξ of the skew-symmetric
endomorphism F .
The freedom of choosing a representative for Sn (as well as the point N and the projection stereographic
plane) can be also seen in a more “passive” picture. Consider two different sets of Minkowskian coordinates
{xα } and {x′α } related by a O+ (1, n + 1) transformation Λ, x′α = Λα β xβ . Using Theorem 6.1, we obtain two
different embeddings i, i′ : En ֒→ M1,n+1 associated to {xα } and {x′α } respectively, for which i(En ) = {x0 =
x1 = 1, xA+1 =: y A } and
i′ (En ) = {x′0 = x′1 = 1, x′A+1 =: y ′A }, as well as two associated maps ξ, ξ ′ . Let
1,n+1
F ∈ SkewEnd M , defined by (30) with parameters {ν, aA , bA , ω A B } and {ν ′ , a′A , b′A , ω ′A B } in the bases
{∂xα } and {∂x′α } respectively. Then, F can be associated to two vector fields
1

ξF = bA + νy A + (aB y B )y A − (yB y B )aA − ω A B y B ∂yA ,
2
A B 1 ′ ′B ′A A B

ξF′ = b′ + ν ′ y ′A + (a′B y ′ )y ′A − (yB y )a − ω ′ B y ′ ∂y′A ,
2
n
which are equal in the following sense. If we transform the representative S′ = {x′0 = 1} ∩ {x′α x′α = 0}
with Λ, we obtain a new representative of the projective cone which in coordinates xα is precisely Sn = {x0 =
1} ∩ {xα xα = 0}. Abusing the notation, the map χΛ := StN ◦ Λ ◦ St−1 ′
N ′ is such that χΛ⋆ (ξF ) = ξF . Then,
considering i(E ) and i (E ) as respresentations of the same space in two different global charts (y A , Rn ) and
n ′ n

(y ′A , Rn ), χΛ can be seen as a change of coordinates y A = (χΛ (y ′ ))A , with the property that the Euclidean
metric in coordinates {y ′A } transforms as
gE = δAB dy ′A dy ′B = Ω2 (y)δAB dy A dy B
for a smooth positive function Ω. In other words, changing to different Minkowskian coordinates in M1,n+1
induces a change of coordinates in En in such a way that the form (31) of the map ξ is preserved. Notice that a
similar result holds if we change the point w.r.t. which we take the stereographic projection, because any two
N, N ′ ∈ Sn must be related by a SO(n) ⊂ O+ (1, n + 1) transformation.
Therefore, for the rest of this section, we will often adapt our choice of Minkowskian coordinates {xα } of
1,n+1
M
 to simplify the problem at hand. With this choice, it comes a corresponding set of cartesian coordinates
y A of En such that ξF is given by equation (31) and the Euclidean metric is gE = Ω(y)2 δAB dy A dy B . Which
coordinates are adequate obviously depends on the problem. For example, from the block form (10) and (11)
of skew-symmetric endomorphisms, consider each of the blocks F |M1,3 F |M1,2 as endomorphisms of M1,n+1 ,
extended as the zero map in (M1,3 )⊥ and (M1,2 )⊥ respectively, and similarly for each F |Πi . If we denote by
ξ F |M1,3 , ξ F |M1,2 and ξ F |Π the corresponding images by ξ, one readily gets following decomposition:
i

p
X p
X
ξF = ξ F |M1,3 + ξ F |Π (n even), ξF = ξ F |M1,2 + ξ F |Π (n odd), (33)
i i
i=1 i=1

24
where in terms of n, p is given by
 
n+1
p= −1 (34)
2

(because the dimension of the Minkowski space where F is defined is d = n + 2, cf. Theorem 6.1). The explicit
form of each of the terms in (33) is direct from (32). Namely, the terms ξ F |M1,3 and ξ F |M1,2 are given by (31)
with vanishing parameters aA , bA , ω A B for A, B ≥ 3 and A, B ≥ 2 respectively, and each ξ F |Π is proportional
i
to a vector field of the form
η := y A0 ∂yB0 − y B0 ∂yA0 (35)
with A0 , B0 ∈ {1, · · · , n} such that A0 6= B0 . More specifically, ξ F |Π = µi ηi , where ηi is given by equation (35)
i
with B0 = A0 + 1 and A0 = 2i if n even while A0 = 2i + 1 if n odd. Vector fields of the form (35) will play an
important role in the following analysis. They have the form of axial Killing vector fields, although in general
they are CKVFs because of the conformal factor in gE = Ω(y)2 δAB dy A dy B . From the previous discussion, it
follows that there exists a conformal transformation χΛ ∈ Conf (En ) such that gE ′
:= χ⋆Λ (gE ) = δAB dy A dy B .
Then by the properties of the Lie derivative it is immediate

0 = Lη χ⋆Λ (gE ) = LχΛ⋆ η gE .

′
In other words, η is an axial Killing vector of gE and χΛ⋆ η is an axial Killing vector of gE . Thus, we define:
Definition 6.1. A CKVF of an Euclidean metric gE , η, is said to be a conformally axial Killing vector field
(CAKVF) if and only if the exist a χΛ ∈ Conf (En ) such that χΛ⋆ (η) is an axial Killing vector field of gE .
Equivalently, η is a CAKVF if and only if it is an axial Killing vector field of χ⋆Λ (gE ).
Remark 6.2. Using Theorem 6.1, it is immediate to verify that a CKVF is a CAKVF if and only if it is the
image under ξ of a simple unit spacelike endomorphism G.
Notice that the terms in (33) form a commutative subset of CKill (En ). This is an immediate consequence of
the fact that ξ is a Lie algebra antihomomorphism (c.f. Theorem 6.1) and the blocks F |M1,2 (resp. F |M1,3 ) and
F |Πi are pairwise commuting. In addition, a straightforward calculation shows that they form an orthogonal
set
e ηi ) = 0,
gE (ξ, gE (ηi , ηj ) = 0 (i 6= j)
where ξe := ξ F |M1,3 for n even and ξe := ξ F |M1,2 for n odd. In fact, as we show next, orthogonality of two CKVFs
implies commutativity provided one of them is a CAKVF. If both are CAKVF, then orthogonality turns out to
be equivalent to commutativity.
Lemma 6.2. Let η, η ′ be non-proportional CAKVFs and ξF a CKVF. Then [η, η ′ ] = 0 if and only if there exist
cartesian coordinates such that η = y n−2 ∂yn−3 − y n−3 ∂yn−2 and η ′ = y n−1 ∂yn − y n ∂yn−1 . Equivalently [η, η ′ ] = 0
if and only if gE (η, η ′ ) = 0. In addition, [ξF , η] = 0 if gE (ξF , η) = 0.


Proof. Let G, G′ ∈ SkewEnd M1,n+1 be such that ξ(G) = η, ξ(G′ ) = η ′ . Since G and G′ are simple, spacelike
and unit (cf. Remark 6.2), we can write G = e ⊗ v♭ − v ⊗ e♭ and G′ = e′ ⊗ v♭′ − v ′ ⊗ e′♭ for spacelike, unit
vectors {e, e′ , v, v ′ }, such that 0 = he, vi = he′ , v ′ i. By Corollary 4.2.1, it follows that [G, G′ ] = 0 if and only
if {e, e′ , v, v ′ } are mutually orthogonal. Let us take cartesian coordinates
 A of M1,n+1 such that e = ∂xn−2 , v =
′ ′
∂xn−1 , e = ∂xn , v = ∂xn+1 . Then, in the associated coordinates y of E it follows η = y n−2 ∂yn−3 −y n−3 ∂yn−2
n

25
and η ′ = y n−1 ∂yn −y n ∂yn−1 . This proves the first part of the lemma. From this result, it is trivial that [η, η ′ ] = 0
implies gE (η, η ′ ) = 0.
To prove that gE (η, ξF ) = 0 implies [η, ξF ] = 0 (which  A in particular establishes the converse gE (η, η ′ ) =
′ n−1 n
0 =⇒ [η, η ] = 0 for CAKVFs), let us take coordinates y such that η = y ∂y − y n ∂yn−1 . Then, writing
ξF as a general CKVF (31), we obtain by direct calculation:
 
yB y B n n−1
gE (η, ξF ) = Ω2 y n bn−1 − y n−1 bn − (a y − an−1 y n ) + ω n−1 B y B y n − ω n B y B y n−1 = 0.
2

Therefore an , an−1 , bn , bn−1 , ω n B , ω n−1 B must vanish. This implies that the associated endomorphisms G and
F to η and ξF adopt a block structure from which it easily follows that [G, F ] = 0 and hence [η, ξF ] = 0.
P
Definition 6.2. Let ξF ∈ CKill (En ). Then a decomposed form of ξF is ξF = ξe+ pi=1 µi ηi for an orthogonal

e ηi }, where ηi are CAKVFs, µi ∈ R for i = 1, · · · , p. A set of cartesian coordinates y A such that
subset {ξ,
ηi = y Ai ∂yAi +1 − y Ai +1 ∂yAi , for Ai = 2i for n odd and Ai = 2i + 1 for n even, is called a set of decomposed
coordinates.
Remark 6.3. Observe that the ξe is a CKVF. By Lemma 6.2 and its proof, the parameters {ν, a, b, ω} defining
ξe in a set of decomposed coordinates must all vanish except possibly {ν, a1 , a2 , b1 , b2 , ω 1 2 = −ω12 } when n is
even or {ν, a1 , b1 } when n is odd. This means that there is a skew-symmetric endomorphism Fe with restricts
to M1,3 ⊂ M1,n (n even) or M1,2 ⊂ M1,n (n odd) and vanishes identically on their respective orthogonal
complements such that ξe = ξFe . We will exploit this fact in an essential way below.
With the definition of decomposed form of CKVFs, we can reformulate Theorem 2.6 in terms of CKVFs.
Proposition 6.1. Let ξF ∈ CKill (En ). Then there exist an orthogonal set {ηi }pi=1 of CAKVFs such that
[ξF , ηi ] = 0. For every such a set {ηj }pj=1 and i ∈ {1, · · · , p} there exist µi ∈ R such that gE (ηi , ηi )µi =
P e P µi ηi provides a decomposed
gE (ξF , ηi ). In addition,with the definition ξe := ξF − µi ηi the expression ξF = ξ+
form of ξF .
Proof. The existence of p commuting CAKVFs is a direct consequence of decompositions (10) and (11) of the
associated skew-symmetric endomorphism F , for n even and odd respectively. Indeed, for each such decom-
position of F , it follows a set of p CAKVFs commuting with ξF . Let us denote {ηi } any such set. Each ηi is
associated to a simple, spacelike unit endomorphism Gi that commutes with F . By Lemma 4.2, Gi defines a
spacelike eigenplane Πi of F . The orthogonality of any two such eigenplanes Πi , Πj , i 6= j is a consequence of
Corollary 4.2.1 because [Gi , Gj ] = 0. In other words, given a set of p CAKVFs commuting with ξF , we have a
P P
block form of F , thus, defining ξe := ξF − µi ηi , it is immediate that ξF = ξe + µi ηi is a decomposed form
with gE (ηi , ηi )µi = gE (ξF , ηi ).
The next step now is to give a definition of canonical form for CKVFs, which we induce from the canonical
form of the associated skew-symmetric endomorphism.
Definition 6.3. A CKVF ξF is in Pcanonical form if it is the image of a skew-symmetric endomorphism F
in canonical form, i.e. ξF = ξe + µi ηi such that ξe is given, in a cartesian set of coordinates {y A } denoted
canonical coordinates, by the parameters a1 = 1, b1 = σ/2, a2 = 0, b2 = τ /2 if n even and a1 = 1, b1 = σ/2
if n odd (the non-specified parameters all vanish) and ηi are CAKVFs ηi = y Ai ∂yAi +1 − y Ai +1 ∂yAi , for Ai = 2i
for n odd and Ai = 2i + 1 for n even, and where σ, τ, µi are given by Definition 3.2.

26
 Given a CKVF ξF , the existence of a canonical form and canonical coordinates are guaranteed by Theorem
3.3. By Theorem 6.1, the conformal class [ξF ]Conf of a CKVF ξF is equivalent to the equivalence class [F ]O+
of F under the adjoint action of O+ (1, n + 1), and this is determined by the canonical form of F (c.f. Theorem
5.1). This argument together with the results of Section 5 yield the following statement.
Theorem 6.3. Let ξF ∈ CKill (Sn ) be in canonical form. Then its conformal class [ξF ]Conf is determined by
(σ, τ, µ2i ) if n even and (σ, µ2i ) if n odd. Moreover, the structure of CKill(Sn )/Conf (Sn ) corresponds with that
of Remark 5.2.
P
Given a canonical form ξF = ξe + e ηi } are pairwise commuting and linearly
µi ηi the set of vectors {ξ,
independent. As we will next prove, in the case of odd dimension this set is a maximal (linearly indepen-
dent) pairwise commuting set of CKVFs commuting with ξ (i.e. it is not contained in a larger set of lin-
early independent vectors commuting with ξ). In the case of even dimension it is not maximal. By Remark
6.3, ξe = ξ(ν,e a1 , a2 , b1 , b2 , ω), where the right-hand side denotes a CKVF of the form (35) whose parameters
vanish, except possibly {ν, a1 , a2 , b1 , b2 , ω := ω 1 2 }. As mentioned in the Remar, the corresponding skew-


symmetric endomorphism Fe satisfying ξFe = ξe can be understood as an element Fe ∈ SkewEnd M1,3 , with

⊥
M1,3 = span{e0 , e1 , e2 , e3 }, that is identically zero in M1,3 . Fix the orientation in M 1,3 so that the basis
{e0 , e1 , e2 , e3 } is positively oriented. The Hodge star maps two-forms into two-forms. This defines a natural
map


⋆ : SkewEnd M1,3 −→ SkewEnd M1,3 ,
Fe 7−→ Fe⋆ .

From standard properties of two-forms, (see also [24]) it follows that Fe⋆ commutes with Fe . We may extend Fe⋆
to an endomorphism on M1,n+1 that vanishes identically on (M1,3 )⊥ , just as Fe . It is clear that the commutation
property is preserved by this extension. The image of Fe⋆ under ξ is the vector field
 ⋆
e
e a1 , a2 , b1 , b2 , ω) = ξ(−ω,
ξe⋆ := ξ(ν, a2 , −a1 , −b2 , b1 , ν),

which e e
P by construction commutes with ξ. In the case that ξ is the first element in a decomposede form ξF =
ξe + µi ηi , it is immediately true that ξe⋆ also commutes with all of the CAKVFs ηi . Hence, {ξ, ξe⋆ , ηi } is a
pairwise commuting set, all of them commuting with ξ. This set can be proven to be maximal:
P
Proposition 6.2. Let ξF = ξe+ µi ηi be a CKVF in canonical form. If n is odd, {ξ, e ηi } is a maximal linearly
independent pairwise commuting set of elements that commute with ξF . If n is even, {ξ, e ξe⋆ , ηi } is a maximal
linearly independent pairwise commuting set of elements that commute with ξF .
Proof. Suppose that there is an additional CKVF ξ ′ commuting with each element in {ξ, e ηi } if n odd or {ξ,
e ξe⋆ , ηi }
′
if n even (in either case ξ clearly commutes with ξF also). Since it commutes with each ηi , by Proposition
Pp
6.1, it admits a decomposed form ξ ′ = ξe′ + i=1 µ′i ηi , where ξe′ is a CKVF orthogonal to each ηi and which
must verify [ξe′ , ξ]
e = 0. Equivalently, their associated endomorphisms satisfy Fe ′ ∈ C(Fe), where C(Fe ) denotes the
centralizer of F , i.e. the set of all skew-symmetric endomorphisms that commute with F . From the results in
[24], C(Fe |M1,2 ) = span{Fe |M1,2 } when n is odd and C(Fe |M1,3 ) = span{Fe |M1,3 , Fe ⋆ |M1,3 } when n is even. Here,
Fe ⋆ is the skew-symmetric endomorphim associated with ξe⋆ and we restrict to M1,3 because the action of the
endomorphisms is identically zero in (M1,3 )⊥ . Thus ξe′ = aξ, e a ∈ R, if n odd and ξe′ = bξe + cξe⋆ , b, c ∈ R if n
even.

27
7 Adapted coordinates
In the previous section we obtained a canonical form for each CKVF of euclidean space based on the canonical
form of skew-symmetric endomorphisms in Section 3. As an application, we consider in this section the problem
of adapting coordinates in En to a given CKVF ξF . The use of the canonical form will allow us to solve
the problem for every possible ξF essentially in one go. Actually it will suffice to consider the case of even
dimension n and assume that at least one of the parameters σ, τ in the canonical form of ξF is non-zero. The
case where both σ and τ vanish will be obtained as a limit (and we will check that this limit does solve the
required equations). The case of odd dimension n wil be obtained from the even dimensional one by expliting
the property that E2m+1 can be viewed as a hyperplane of E2m+2 in such a way that the given CKVF ξF
in E2m+1 extends conveniently to E2m+2 . Restricting the adapted coordinates already obtained in the even
dimensional case to the appropriate hyperplane we will be able to infer the odd dimensional case. As we will
justify the process of adapting coordinates is different for n = 2 and n ≥ 4 even. The case n = 2 has been
treated in detail in [24], so it will suffice to consider even n ≥ 4 here.
Consider En endowed with a CKVF ξF . First of all we adapt the Cartesian coordinates of En so that ξF
takes its canonical form and we fix the metric of En to take the explictly flat form in these coordinates. We
further assume (for the moment) that n is even. For notational reasons it is convenient to rename the canonical
coordinates4 as z1 := y 1 , z2 := y 2 and xi := y 2i+1 , yi := y 2i+2 for i = 1, · · · , p, where in the even case
case p = n/2 − 1 (see (34)). By Proposition 6.1, ξF can be decomposed as a sum of CKVFs ξe and ηi and,
additionally one can construct canonically yet another CKVF ξe⋆ . This collection of CKVFs defines a maximal
commutative set. Moreover, {ηi } are all mutually orthogonal and perpendicular to ξe and ξe⋆ . It is therefore
most natural to try and find coordinates adapted simultaneously to the whole family {ξ, e ξe⋆ , ηi }. This will lead
a (collection of) coordinate systems where the components of ξF are simply constants. From here one can
immediately find coordinates that rectify ξF , if necessary. It is important to emphasize that selecting the whole
e ξe⋆ , ηi } to adapt coordinates provides enough restrictions so that the coordinate change(s) can be fully
set {ξ,
determined. Imposing the (much weaker) condition that the system of coordinates rectifies only ξF is just a
too poor condition to solve the problem. This is an interesting example where the structure of the canonical
decomposition of ξF (or of F ) is exploited in full.
By Theorem 6.1, the explicit form of {ξ, e ξe⋆ , ηi } in the canonical coordinates is
!!
σ 1 Xp τ  Xp
e
ξ= + 2 2
z1 − z2 − 2
(xi + yi )2
∂z1 + + z1 z2 ∂z2 + z1 (xi ∂xi + yi ∂yi ) (36)
2 2 i=1
2 i=1
!!
τ  σ 1 Xp Xp
e ⋆
ξ =− + z1 z2 ∂z1 + − 2
z2 − z1 −2 2 2
(xi + yi ) ∂z2 − z2 (xi ∂xi + yi ∂yi )
2 2 2 i=1 i=1
ηi = xi ∂yi − yi ∂xi .

We are seeking coordinates {t1 , t2 , φi , vi } adapted to these vector fields, i.e. such that ∂t1 = ξ, e ∂t2 = ξe⋆ , ∂φi = ηi .
It is clear that if {t1 , t2 , φi , vi } is an adapted coordinate system, so it is {t1 − t0,1 (v), t2 − t0,2 (v), φi − φ0,i (v), vi }
for arbitrary functions t0,1 (v), t0,2 (v) and φ0,i (v), where v = (v1 , · · · , vp ). This will be used to simplify the
4 The fact that we tag the coordinates {z , z , x , y } with lower indices has no particular meaning. It is simply to avoid a
1 2 i i
notational clash of upper indices and powers that will appear later

28
process of integration. This freedom, may be restored at the end if so desired. Hence
p
!
∂z1 σ 1 X ∂z2 τ ∂xi ∂yi
2 2 2 2
= + z1 − z2 − (xi + yi ) , = + z1 z2 , = z1 xi , = z1 yi , (37)
∂t1 2 2 i=1
∂t1 2 ∂t1 ∂t1
p
!
∂z2 σ 1 X ∂z1 τ ∂xi ∂yi
2 2 2 2
= − z2 − z1 − (xi + yi ) , = − − z1 z2 , = −z2 xi , = −z2 yi , (38)
∂t2 2 2 i=1
∂t 2 2 ∂t2 ∂t2
∂z1 ∂z2 ∂xi ∂yi
=0 =0 = −yi = xi (39)
∂φi ∂φi ∂φi ∂φi
The additional p coordinates vi , will appear through functions of integration. It is clear that the structure of
the equations is different for n = 2, where there are no {xi , yi }, which implies that the process of integration
follows a different route. The case n = 2 has been treated in full detail in [24], where the the complex structure
of S2 can be exploited to simplify the problem. Here we adress the problem for n ≥ 4 which we assume from
now on.
We may start by integrating (39). The first pair gives z1 = z1 (t1 , t2 , v), z2 = z2 (t1 , t2 , v), so that the second
pair becomes a harmonic oscillator in xi , yi , whose solution is
xi = ρi (t1 , t2 , v) cos(φi − φ0,i (t1 , t2 , v)), yi = ρi (t1 , t2 , v) sin(φi − φ,0,i (t1 , t2 , v)), (40)
where ρi and φ0,i are arbitrary functions (depending only on the variables indicated) and ρi is not identically
zero.
Inserting (40) in any of the two right-most equations of (37) and (38) and equating terms multiplying
sin(φi + φ(0)i ) and cos(φi + φ(0)i ) yields:
1 ∂ρi 1 ∂ρi ∂φ(0)i ∂φ(0)i
z1 = , z2 = − , = 0, = 0.
ρi ∂t1 ρi ∂t2 ∂t1 ∂t2
Thus, φ0,i is a function only of v, which may be absorbed on the coordinate φi as discussed above. The two
first equations imply
1 ∂ρi 1 ∂ρj 1 ∂ρi 1 ∂ρj
= , = ⇐⇒ ρi = α̂i (v)ρ̂(t1 , t2 , v),
ρi ∂t1 ρj ∂t1 ρi ∂t2 ρj ∂t2
p
 p

2
P P
for arbitrary (non-zero) functions α̂i and ρ̂. Defining ρ := ρ2i = α̂2i ρ̂2 we can write
i=1 i=1

α̂i ǫ
ρi = α̂i ρ̂ = qP ρ = αi ρ,
p
j=1 α̂2j
qP p
P
p
where αi := α̂i ǫ/ j=1 α̂2j , with ǫ2 = 1, form a set of arbitrary (non-zero) functions of v such that α2i = 1.
i=1
The function ρ satisfies
1 ∂ρ 1 ∂ρ
z1 = , z2 = − . (41)
ρ ∂t1 ρ ∂t2
Inserting (41) in the two left-most equations in (37) and (38), with the change of variable U = ρ−1 , we obtain
after some algebra the following covariant system of PDEs (indices a, b = 1, 2 refer to {t1 , t2 })
1 1
∇a ∇b U = U Aab + (1 + ∇c U ∇c U )gab with A= (−σdt21 + σdt22 + 2τ dt1 dt2 ), g = dt21 + dt22 , (42)
2U 2

29
and where ∇ is the Levi-Civita covariant derivative of g.
Lemma 7.1. Up to shifts t1 → t1 − t0,1 (v) and t1 → t1 − t0,1 (v), the general solution of (42) with either σ or
τ non-zero is given by
q
ǫ
U= 2 (β cosh(t + ) − α cos(t − )) with β = α2 + µ2t + µ2s (43)
µt + µ2s
where α is a function of integration (depending on v), ǫ2 = 1 and t+ := µt t1 + µs t2 , t− := µt t2 − µs t1 , with
µs , µt given by (9). The solution (43) admits a limit σ = τ = 0 (i.e. µt = µs = 0) provided α > 0, which is
α ǫ
lim U = ǫ (t21 + t22 ) + . (44)
µs µt →0 2 2α
Up to shifts t1 → t1 − t0,1 (v) and t2 → t2 − t0,2 (v), this function is the general solution of (42) for σ = τ = 0.
Proof. The coordinates t+ , t− defined in the lemma diagonalize A and g simultaneously and yield
1 2 1
A= (dt − dt2− ), g= (dt2+ + dt2− ).
2 + µ2s + µ2t
From this and equation (42) it follows that ∂ 2 U/∂t+ ∂t− = 0 or, equivalently, U (t+ , t− ) = U+ (t+ ) + U− (t− ).
Substracting the {t+ , t+ } and {t− , t− } components of (42) one obtains
d2 U+ d2 U− d2 U+ d2 U−
− = U = U+ + U− =⇒ − U + = + U− = â
dt2+ dt2− dt2+ dt2−
for an arbitrary separation function â(v). The general solution is clearly
U+ = −â + a cosh(t+ ) + b sinh(t− ) U− = â + c cos(t− − δ), (45)
where a, b, c, δ are also functions of v. Since â drops out in U = U+ + U− we may set â = 0 w.l.o.g. Inserting
(45) in (any of) the diagonal terms of (42) and one simply gets
1
a2 − b 2 = + c2 .
µ2s + µ2t
Hence |a| > |b| and we may use the freedom of translating t+ by a function of v to write U+ = a cosh(t+ )
(i.e. b = 0). A similar translation in t− sets δ = 0. Rescaling the functions a, c as a = (µ2s + µ2t )−1 β and
c = −(µ2s + µ2t )−1 α we get
β α
U = U− + U− = cosh(t+ ) − 2 cos(t− ), β 2 = µ2s + µ2t + α2 . (46)
µ2s + µ2t µs + µ2t
p
It is obvious that sign(U ) = sign(β). Thus taking β as the positive root β = α2 + µ2s + µ2s and adding
a multiplicative sign ǫ in (46), we obtain (43). To evaluate the convergence as both σ, τ tend to zero, or
equivalently µs , µt → 0, consider the series expansion
  
µ2s + µ2t (4) (µs t2 + µt t1 )2 (4)
β cosh(t+ ) = |α| + + oµt ,µs 1+ + oµt ,µs ,
2|α| 2
(µt t2 − µs t1 )2
α cos(t− ) = α − α + o(4)
µt ,µs ,
2

30
 (4)
where oµt ,µs denotes a sum of homogeneous polynomials in µt , µs starting at order four, whose coeficients may
depend on t1 , t2 and α. Then, the expansion of U is
 
ǫ |α|µ2s + αµ2t 2 |α|µ2t + αµ2s 2 µ2s + µ2t (4)
U= 2 (|α| − α)(1 + µs µt t1 t2 ) + t2 + t1 + + oµt ,µs .
µs + µ2t 2 2 2|α|
(4)
It is clear that limµs ,µt →0 oµt ,µs /(µ2s + µ2t ) = 0 and the rest of the equation converges if and only if α > 0 in
which case the limit is (44). An easy calculation shows that this limit is (up to shifts in t1 , t2 ) is the general
solution of (42) when σ, τ = 0.
Having the general general solution (43) of (42) we can give the expression of the adapted coordinates

1 ∂U 1 αµs sin(t− ) − βµt sinh(t+ )

z1 = − = , (47)
U ∂t1 U µ2s + µ2t
1 ∂U 1 αµt sin(t− ) + βµs sinh(t+ )
z2 = = , (48)
U ∂t2 U µ2s + µ2t
αi αi
xi = cos(φi ), yi = sin(φi ), (49)
U U
where no sign of α is in principle assumed5 , except for the case µs = µt = 0, where U must be understood as
the limit (with α > 0) (44) and z1 = −U −1 ∂U/∂t1 , z2 = U −1 ∂U/∂t2 . This coincides with the limit of the RHS
expressions (47), (48), which is

−2α2 t1 2α2 t2
z1 = , z2 = . (50)
1 + α2 (t21 + t22 ) 1 + α2 (t21 + t22 )

From equations (47), (48) and (49) it is obvious that the sign ǫ is not relevant in the definition of the adapted
coordinates. This is because the two branches ǫ = 1 and ǫ = −1 correspond with U > 0 and U < 0 respectively,
which in terms of the adapted coordinates, is equivalent to a rotation of π in the φi angles. Hence, w.l.o.g. we
consider ǫ = 1, Pi.e. U > 0. Also notice that the dependence on the variables vi appears through the functions
p
αi and α, with i=1 α2i = 1. The set {αi , α} define p independent arbitrary functions of the Ppvariables vi , so it
is natural to use as coordinates {αi , α} themselves, provided they are restricted to satisfy i=1 α2i = 1.
We now calculate the region of En covered by the adapted coordinates. It is clear that in no case this region
can include neither the zeros of the vector fields ξe and ξe⋆ and ηi nor the points where these p + 2 vectors are
linearly dependent. We therefore start by locating those points. Denoting the loci of the zeros of ξe and ξe⋆ and
e Z(ξ)
ηi by Z(ξ), e ⋆ and Z(ηi ) respectively, a simple calculation gives

 \p   p
X 
e = {
Z(ξ) {xj = yj = 0}} ∩ {z1 = ±µt , z2 = ∓µs } ∪ {z1 = 0} ∩ {z22 + (x2j + yj2 ) = µ2s − µ2t } if µs µt = 0 ,
j=1 j=1
(51)
 \p   p
X 
Z(ξe⋆ ) = { {xj = yj = 0}} ∩ {z1 = ±µt , z2 = ∓µs } ∪ {z2 = 0} ∩ {z12 + (x2j + yj2 ) = µ2t − µ2s } if µs µt = 0 ,
j=1 j=1
(52)
Z(ηi ) = {xi = yi = 0} .
5 The domain of definition of α will be later restricted under the condition that the adapted coordinates define a one to one map.

31
These e e⋆
 Tp expressions are valid for every value of µs , µt and imply that in the case µs = µt = 0, Z(ξ) = Z(ξ ) =
j=1 {xj = yj = 0} ∩ {z1 = z2 = 0}, which is contained in each Z(ηi ) = {xi = yi = 0}.
On the other hand, since {ξ, e ηi } is an orthogonal set of CKVFs (cf. Lemma 6.2), they are pointwise linearly
independent at all points where they do not vanish. Similarly, {ξe⋆ , ηi } is also an orthogonal set, so linear
independence is guaranteed away from the zero set. Away from this set, the set of vectors {ξ, e ξe⋆ , ηi } is linearly
e e ⋆
dependent only at points where ξ and ξ are proportional to each other with a non-zero proportionality factor,
ξe = aξe⋆ , a 6= 0. One easily checks that, away from Z(ξ) e and Z(ξe⋆ ), the set of point where ξe − aξe⋆ vanishes is
µt
empty except when µs 6= 0, µt 6= 0 and a = µs . It turns out to be useful to determine the set of points where
µs ξe − µt ξe⋆ = 0 when at least one of {µs , µt } is non-zero. We call this set Z(µs ξe − µt ξe⋆ ), and a straightforward
analysis gives
 p
 P

 {µs z1 = −µt z2 } ∩ {(µ2s + µ2t )z22 + µ2s (x2i + yi2 ) = (µ2s + µ2t )µ2s } if µs 6= 0
Z(µs ξe − µt ξe⋆ ) = i=1
Pp (53)

 2 2 2 2 2 2 2 2 2
 {µ z
s 1 = −µ z
t 2 } ∩ {(µ s + µ )z
t 1 + µ t (x i + y i ) = (µ s + µ t )µ t } if µ t =
6 0
i=1

Obviously, the two expressions are equivalent when both µs and µt are non-zero. The interest of this set is that
it happens to always contain Z(ξ)e and Z(ξe⋆ ). This, together with the fact that when µs = µt = 0 these sets
are contained in the axes Z(ηi ) will allow us to ignore them altogether.

Lemma 7.2. Assume that at least one of {µs , µt } is non-zero. Then Z(ξ), e Z(ξe⋆ ) ⊂ Z(µs ξe − µt ξe⋆ ).
 Tp
Proof. Consider first µs , µt 6= 0. Then at Z(µs ξe − µt ξe⋆ ) ∩ j=1 {xj = yj = 0} we have that z1 = ±µt and
z2 = ∓µs which establishes Z(ξ), e Z(ξe⋆ ) ⊂ Z(µs ξe − µt ξe⋆ ) in this case. When µt = 0, µs 6= 0, by definition of
the respective sets we have Z(ξ) e = Z(µs ξe − µt ξe⋆ ). Moreover, directly from (52) one finds
p
\
Z(ξe⋆ ) = {xj = yj = 0} ∩ {z1 = 0, z2 = ±µs } ,
j=1

which (cf. the first expression in (53)) is clearly contained in Z(µs ξe − µt ξe⋆ ). An analogous argument applies in
the case µt 6= 0, µs = 0.
Let us define the following auxiliary coordinates
µs z 1 + µt z 2 µs z 2 − µt z 1
ẑ+ := pPp 2 2
, ẑ− := pPp 2 2
, x̂i := xi , ŷi := yi .
i=1 (xi + yi ) i=1 (xi + yi )

Except
 Tpfor the case µs = µt = 0 (which will be analyzed later) the coordinates {ẑ+ , ẑ− , x̂i , ŷi } obviously cover
Rn \ j=1 {xj = yj = 0} . In terms of the adapted coordinates, they read

αi αi
ẑ+ = α sin(t− ), ẑ− = β sinh(t+ ) x̂i = cos(φi ), ŷi = sin(φi ). (54)
U U
Let us analyze the points where (54) fails to be a change of coordinates and hence restrict the domain of definition
of {α, t− , t+ , αi , φi }. The first thing to notice is that a change of sign in the coordinate αi is equivalent to a
rotation of angle π in the coordinate φi . Moreover, S at points where αi = 0, i.e. the axis of ηi , the coordinate φi
p
is completely degenerate, which obviously excludes j=1 {xj = yj = 0} from the region covered by the adapted

32
coordinates. To avoid duplications, we must restrict αi ∈ (0, 1) and φ ∈ [−π, π) or alternatively αi ∈ (−1, 1)\{0}
and φi ∈ [0, π). We choose the former for definiteness.
The hypersurface {α = const,P t− = const, t+ = const} is an n − 3 dimensional sphere of radius U −1 , namely
p
{ẑ− = const, ẑ+ = const}T∩ { i=1 (x2i + yi2 ) = U −2 = const}. This gives a straightforward splitting of
p
R \{0n−2 }, with 0n−2 := { j=1 {xj = yj = 0} , into R2 × (Rn−2 \{0n−2 }), where Rn−2 \{0n−2} is foliated by
n

n − 3 dimensional spheres. The set Z(µs ξe − µt ξe⋆ ) respects this foliation, so it descends to R2 × R+ (the last
factor is the radius of the n − 3 sphere). To avoid extra notation we also use Z(µs ξe − µt ξe⋆ ) to denote this
quotient set. We nextSshow that the adapted coordinates actually cover the largest possible domain, namely
p
Rn \ {Z(µs ξe− µt ξe⋆ ) ∪ j=1 {xj = yj = 0}}. From the previous discussion, this is a consequence of the following
result.
Lemma 7.3. Assume that at least one of {µs , µt } is not zero. Then. the transformation


(ẑ+ , ẑ− , U ) : R × [−π, π) × R+ −→ R2 × R+ \Z(µs ξe − µt ξe⋆ )
(55)
(t+ , t− , α) 7→ (ẑ+ , ẑ− , U ).
is a diffeomorphism.
Proof. The determinant of the jacobian of (55) reads
∂(ẑ+ , ẑ− , U )
= αU.
∂(t+ , t− , α)
Since U is strictly positive (cf. (43) and recall that we chose ǫ = 1 w.l.o.g.), the conflictive points are α = 0.
To calculate the locus {α = 0} we obtain the inverse transformation of α in terms of U, ẑ+ , ẑ− by solving (43)
and the first two in (54). The result is, after a straightforward computation,
 1/2
2 1 2 2 2 2 2 2 2 2 2
α = ± ẑ+ + (ẑ + ẑ− − U (µs + µt ) + (µs + µt )) . (56)
4U 2 (µ2s + µ2t )2 +
2
It follows that α = 0 is equivalent to ẑ+ = 0 and ẑ− + µ2s + µ2t = U 2 (µ2s + µ2t )2 . When translated into the original
coordinates{z1, z2 , xi , yi } this set is precisely Z(µs ξe− µt ξe⋆ ). Also, from (56) it is obvious that α is multivalued,
which also implies that t− is multivalued after substituting α as a function of ẑ+ , ẑ− , U in the first equation in
(54)6 . We solve this issue by restricting α to be strictly positive and let t− take values in [−π, π).
Sp
We have shown that the adapted coordinates cover all Rn except j=1 Z(ηj ) ∪ Z(µs ξe− µt ξe⋆ ) . The domain
of definition of the coordinates t1 , t2 depends on µt and µs , because −π ≤ t− = µt t2 − µs t1 < π. This defines
a band B(µs , µt ) := {−π ≤ t− = µt t2 − µs t1 < π}, whose width and tilt is determined by σ, τ through µs , µt
(see figure 2). Nevertheless, the coordinate change is well defined for all values of t1 and t2 and involves only
periodic functions of t− . Thus, we can extend the domain of definition of t1 , t2 to all of R2 . This defines a
Sp
covering of the original space Rn \( j=1 Z(ηj ) ∪ Z(µs ξe − µt ξe⋆ )) which unwraps completely the orbits of ξe and
ξe⋆ . It is not the universal covering because it does not unwrap the orbits of the axial vectors. This result is a
generalization to higher dimensions of the covering dicussed in detail in [24].
The limit case µs = µt = 0 (that is σ = τ = 0) corresponds with a band of infinite Sp width, i.e. B(µs , µt ) = R2 .
n
In this case, the adapted coordinates also cover the largest possible set R \( j=1 Z(ηj )). Recall that in this
case the only points where {ξ, e ξe⋆ , ηi } is not a linearly independent set is the union of Z(ξ), e Z(ξe⋆ ), and Z(ηi )
and we have already seen that in this case Z(ξ) e = Z(ξe⋆ ) ⊂ Z(ηi ), for i = 1, · · · , p. This limit case is the same
result that we would have obtained, had we performed a direct analysis using U as given by (44).
6 This was already evident by observing that a change of sing in α is cancelled by a rotation of π in t−

33
 t1

t2

PSfrag replacements w

 
µs
Figure 2: Band B(µs , µt ) where the coordinates t1 , t2 are defined. The tilt is given by θ = arctan µt and the
width w is 2π/µt if µt 6= 0, 2π/µs if µt = 0, µs 6= 0 and w → ∞ if µs = µt = 0.

Once we have determined the adapted coordinates and the region they cover, we may proceed to calculate
the expression of the Euclidean metric
p
X

gE = dz12 + dz22 + dx2i + dyi2 . (57)
i=1
p
P

in adapted coordinates. We start with the term dx2i + dyi2 , which is straightforward
i=1
p p p
!
X
dU2
1 X
2dU X dU 1
dx2i + dyi2 = 4 + 2 dα2i + α2i dφ2i Pp
α2 =1 − αi dαi = + 2 γSn−3 , (58)
i=1
U U i=1
i=1 i U2 i=1
U 4 U
Pp Pp
where in the last equality we used i=1 αi dαi = 0, which follows from i=1 α2i = 1 and we have defined
p
X

γSn−3 := dα2i + α2i dφ2i Pp
α2i =1
. (59)
i=1
i=1

The notation is P justified because the
right-hand side corresponds to the standard
Pp unit metric on Sn−3 . This
p 2 2 2 2
follows because i=1 dαi + αi dφi is obviously flat and the restriction i=1 αi = 1 defines a unit sphere.
We emphasize, however that the notation γSn−3 refers to the quadratic form above, not to the spherical metric
in any other coordinate system. Observe also that dU in (58) should be understood as a short name for the
explicit differential of U in terms of dt1 , dt2 , dα. Using (57) and (58), we have
 2  2  2
∂z1 ∂z2 1 ∂U
gt1 t1 = + + 4 ,
∂t1 ∂t1 U ∂t1

34
which after a explicit calculation reduces to

α2 + µ2t
gt1 t1 = .
U2

Notice that gt1 t1 = gE (ξ, e ξe⋆ ). From the expressions in cartesian

e gt t = gE (ξe⋆ , ξe⋆ ) and gt ,t = gE (ξ,
e ξ),
2 2 1 2
coordinates it is straightforward to show
p
X p
e ξ)
e = gE (ξe⋆ , ξe⋆ ) − σ σ τ X 2 τ
gE (ξ, (x2i + yi2 ) = gE (ξe⋆ , ξe⋆ ) − , e ξe⋆ ) =
gE (ξ, (x + yi2 ) =
i=1
U2 2 i=1 i 2U 2

p
P
where we have used U −2 = (x2i + yi2 ) (see (49)). Thus
i=1

σ α2 + µ2s τ µs µt
gt2 t2 = gt1 t1 + = , gt1 t2 = = .
U2 U2 2U 2 U2
The remaining terms are rather long to calculate. With the aid of a computer algebra system one gets
2  2  2
∂z1 ∂z2 1 ∂U 1
gαα = + + 4 = 2 2
∂α ∂α U ∂α β U
∂z1 ∂z1 ∂z2 ∂z2 1 ∂U ∂U
gαt1 = + + 4 = 0,
∂α ∂t1 ∂α ∂t1 U ∂α ∂t1
∂z1 ∂z1 ∂z2 ∂z2 1 ∂U ∂U
gαt2 = + + 4 = 0.
∂α ∂t2 ∂α ∂t2 U ∂α ∂t2
Notice that no terms in dαi , dφi appear but those in γSn−3 , since neither U nor z1 , z2 depend on αi , φi . Putting
all these results together we obtain the following expression:
Lemma 7.4. In adapted coordinates {t1 , t2 , α, αi , φi }, the Euclidean metric gE takes the form
 
1 dα2
gE = 2 (α2 + µ2t )dt21 + (α2 + µ2s )dt22 + 2µs µt dt1 dt2 + 2 + γ Sn−3 . (60)
U α + µ2s + µ2t

We would like to stress the simplicity of this result. Except in the a global conformal factor, the metric does
not depend in t1 and t2 (so, both ξe and ξe⋆ are Killing vectors of U 2 gE ). The dependence in the coordinate α
and the conformal class constants {µs , µt } is also extremely simple. Even more, the fact that all dependence in
{αi , φi } arises only in γSn−3 allows us to use any other coordinate system on the unit Sn−3 . Any such coordinate
system is still adapted to ξe and ξe⋆ but (in general) no longer to {ηi }. This enlargement to partially adapted
coordinates is an interesting consequence of the foliation of Rn by (n − 3)-spheres described above.
We now work out the odd n case. As already discussed, we will base the analysis on the even dimensional
case by restricting to a suitable a hyperplane. The underlying reason why this is possible is given in the following
lemma.
Lemma 7.5. Fix n ≥ 3 odd. Let ξF be a CKVF of En in canonical form and let {z1 , xi , yi } be canonical
coordinates. Consider the embedding En ֒→ En+1 where En is identified with the hyperplane {z2 = 0}, for a
cartesian coordinate z2 of En+1 . Then ξF extends to a CKVF of En+1 with the same value of σ, µi and τ = 0.

35
Proof. By Remark 6.3 and Theorem 6.1, the expression of ξF in the canonical coordinates {z1 , xi , yi } is
p
!! p p p
σ 1 X X X X
ξF = + 2
z1 − 2 2
(xi + yi ) ∂z1 + z1 (xi ∂xi + yi ∂yi ) + µi (xi ∂yi − yi ∂xi ) := ξe + µi ηi .
2 2 i=1 i=1 i=1 i=1
p
P
Define ξF′ on En+1 in cartesian coordinates {z1 , z2 , xi , yi } by ξF′ = ξe′ + µi (xi ∂yi − yi ∂xi ) where ξe′ is given
i=1
by (36) with τ = 0. It is clear that this vector is a CKVF of En+1 written in canonical form, that it is tangent
to the hyperplane z2 = 0 and that it agrees with ξF on this submanifold.
Consequently, introducing adapted coordinates for the extended CKVF and restricting to {z2 = 0} will
provide adapted coordinates for ξF . The restriction will obviously reduce the domain of definition of the
adapted coordinates (t1 , t2 , α, αi , φi ) to a hypersurface. It is straightforward from equation (48) and the second
equation in (50) that for the three cases σ > 0, σ = 0 or σ < 0, the hyperplane {z2 = 0} corresponds to
{t2 = 0}. It follows that the remaining coordinates {t1 , α, αi , φi } are adapted to ξe and all ηi . Their domain of
definition is t1 ∈ R, α ∈ R+ , αi ∈ (0, 1), φi ∈ [−π, π) and the coordinate change is given by (47) (or the first in
(50)) together with (49) after setting τ = 0 and t2 = 0. Depending on the sign of σ one gets for z1
 −1 α sin(√σt1 )

 √ , σ>0
 |U + | √ 2 σ √
−1 α +|σ| sinh( |σ|t1 )
z1 = √ , σ<0 , (61)

 |U − | |σ|
 −1
|U 0 | αt1 , σ=0
where
1 p 2 √ 1 p 2 √ 1 1
U + := ( α + σ − α cos( σt1 )), U − := ( α − σ cosh( −σt1 ) − α), U 0 := (αt21 + ),
σ −σ 2 α
and for all three cases
αi αi
xi = cos(φi ), yi = ǫ sin(φi ), (62)
Uǫ U
where we write U ǫ for the function U + , U − or U 0 according with sign of σ.
The range of variation of {t1 , α, αi , φi } was inferred before from the corresponding range of variation of
{t1 , t2 , α, αi , φi } in En+1 . It may happen, however, that when we restrict to the hyperplane {z2 = 0}, the range
gets enlarged and additional points get covered by the adapted coordinate system. The underlying reason is
that, in effect, we are no longer adapting coordinates to ξe′⋆ , so the points on z2 = 0 where this vector is linearly
dependent to ξe′ (or zero) are no longer problematic. When τ = 0, one has
√ p
(µs = σ, µt = 0) if σ ≥ 0, (µs = 0, µt = |σ|) if σ ≤ 0.

We may ignore the case σ = 0 because Z(ξe′ ) = Z(ξe′⋆ ). It follows from (51) and (53) that
  p 
 P 2
 {z1 = 0} ∩ 2
(xi + yi ) = σ if σ > 0
Z(ξe′ ) = T
i=1 n p o
z2 =0 
 p {xj = yj = 0} ∩ z1 = ± |σ|
j=1 if σ < 0
  p 
 P 2

 {z1 = 0} ∩ (xi + yi2 ) = σ if σ > 0
Z(µs ξe − µt ξe )
′ ′⋆
= p
P
i=1
z2 =0 

 {z12 + (x2i + yi2 ) = |σ|} if σ < 0.
i=1

36
When σ > 0, the two sets are the same and no extension of the coordinates {t1 , α, αi , φi } is possible. However,
when σ < 0, the set Z(µs ξe′ − µt ξe′⋆ )|z2 =0 is strictly larger than Z(ξe′ )|z2 =0 . From expressions (61) and (62) one
checks that Z(µs ξe′ − µt ξe′⋆ )|z2 =0 \ Z(ξe′ )|z2 =0 corresponds exactly to the value α = 0 and that Z(ξ)
e = Z(ξe′ )|z =0
2
is at the limit t1 → ±∞. Thus, a priori there is the possibility that the adapted coordinates {t1 , α, αi , φi } can
be extended regularly to α = 0 when σ < 0. It follows directly from (61) that this is indeed the case (observe
that, to the contrary, the limit α → 0 in (61) is singular when σ ≥ 0, in agreement with the previous discussion).
Thus, the range of definition of α is [0, ∞) when σ < 0. The conclusion is that, irrespectively of the value of
σ, the adapted coordinates {t1 , α, αi , φi } cover the largest possible domain of En , namely all points where ξe is
non-zero away from the axes of {ηi }.
To obtain the Euclidean metric in En for n odd in adapted coordiantes we simply restrict (60) (with n → n+1)
to the hypersurface t2 = 0, and get
  
ǫ 1 2 (1 − ǫ)|σ| 2 dα2
gE = α + dt1 + 2 + γSn−2 , (63)
(U ǫ )2 2 α + |σ|

where ǫ = −1, 0, 1 respectively if σ < 0, σ = 0, σ > 0.

Remark 7.1. The three odd dimensional cases can be unified into one. The function U 0 coincides with the
limits of U + and U − when σ → 0. However, the analytical continuation of U + to negative values of σ does not
directly yield U − . To solve this we introduce the function
1 p 2 √

W1 (y) = y + σ − y cos σt1 ,
σ
which is analytic in σ and takes real values forpreal σ. We observe that U + (α = y) = W1 (y) for σ > 0,
U 0 (α = y) = W1 (y) (σ = 0) and U − (α =p+ y 2 + σ) = W1 (y) (σ < 0). This suggests introducing the
α = + y 2 + σ for σ < 0. From the domain of α, it follows that y takes
coordinate change α = y for σ ≥ 0 and √
values in y > 0 when σ ≥ 0 and y ≥ −σ when σ < 0. In terms of y, the three metrics metric g ǫ take the
unified form
 
ǫ 1 2 2 dy 2
gE = y dt1 + 2 + γSn−2 .
W1 (y)2 y +σ

The function W1 is the analytic continuation of U + to negative values of σ. We could have started with U −
and continued analytically√to positive values of σ. Instead of repeating
√ the argument, we simply introduce a new
variable z defined by y = z 2 − σ with range of variation z > σ for σ ≥ 0 and z ≥ 0 for σ < 0. The metric
takes the (also unified and even more symmetric) form
 
ǫ 1 2 2 dz 2 1 p
2 − σ cos
√

gE = (z − σ)dt 1 + + γ Sn−2 , W 2 (z) := z − z σt1 .
W2 (z)2 z2 − σ σ

The function W2 (z) is again analytic in σ, takes real values on the real line, and √now it extends U − . More
specifically, U − (α = z) = W2 (z) (σ < 0), U 0 (α = z) = W2 (z) (σ = 0) and U + (α = z 2 − σ) = W2 (z) (σ > 0).
Remark 7.1 allows us to work with all the odd dimensional cases at once, which will be useful for Section 8.
However, this unified form does not arise naturally when the odd dimensional case is viewed as a consequence
of the n + 1 even dimensional case. So, leaving aside this remark for Section 8, we summarize the results of this
section in the following Theorem.

37
 e Pp
Theorem 7.6. Given a CKVF ξF P of En , with n ≥ 4 even, in canonical form ξF = ξ+ i=1 µi ηi , the coordinates
p
t1 , t2 , φi , α, αi , for i = 1, · · · p and i=1 α2i = 1, defined by

1 ∂U 1 ∂U αi αi
z1 = − , z2 = xi = cos(φi ), yi = sin(φi )
U ∂t1 U ∂t2 U U
with p
α2 + µ2t + µ2s cosh(µt t1 + µs t2 ) − α cos(µt t2 − µs t1 )
U= ,
µ2t + µ2s
which admits a limit limµs µt →0 U = α 2
+ t22 ) + 1 e e⋆
2 (t1 S adapted coordinates to ξ = ∂t1 ξ = ∂t2 ηi = ∂φi ,
furnish 2α ,
p e e⋆
which cover the maximal possible domain, namely En \ j=1 Z(ηj ) ∪ Z(µs ξ − µt ξ ) for t1 , t2 ∈ B(µs , µt ),
φi ∈ [−π, π), αi ∈ (0, 1) and α ∈ R+ . Moreover, the metric gE , which is flat in canonical cartesian coordinates,
is given by

1  2 
X p
2 2 2 2 2 dα2

gE = (α + µ t )dt 1 + (α + µ s )dt 2 + 2µ s µ t dt 1 dt 2 + + dα2i + α2i dφ2i Pp 2 . (64)
U2 α2 + µ2s + µ2t i=1 αi =1
i=1

Pp
If n ≥ 3 is odd and ξF is in canonical form, ξF = ξe+ i=1 µi ηi , the coordinates {t1 , φi , α, αi } adapted to ξe = ∂t1
ηi = ∂φi are given by the case of n+1 (even) dimensions, for τ = 0 restricted to t2 = 0 (which S defines the embed- 
n n+1 n p e
ding E = {z2 = 0} ⊂ E ) and cover again the maximal possible domain, given by E \ Z(ηj ) ∪ Z(ξ)
j=1
for t1 ∈ R, φi ∈ [−π, π), αi ∈ (0, 1) and α ∈ R+ when σ ≥ 0 and α ∈ R+ ∪ {0} when σ < 0. Moreover,
the metric gE , which is flat in canonical cartesian coordinates, is given by the pull-back of (64) at t2 = 0 after
setting τ = 0. Explicitly gE is, depending on the sign of σ, given by (63) with γSn−2 as in (59).

8 TT-Tensors
The adapted coordinates derived in Section 7 provide a useful tool to solve geometric equations involving
CKVFs. In this section we give an example of this in the context of Λ-vacuum spacetimes admitting a smooth
null conformal infinity.
Recall that for such spacetimes the data at I is a conformal class [g] of riemannian metrics and a conformal
class of transverse and divergence-free tensors. More specifically, for a representative metric g in the conformal
class, there is associated a symmetric tensor DAB satisfying gAB DAB = 0 (divergence-free) and ∇A DAB = 0
(transverse). For any other metric g̃ = Ω2 g in the conformal class, the associated tensor is Ω−(n+2) DAB , which
is again a TT tensor with respect to g̃. In dimension n = 3, it has been shown in [25] that the spacetime
generated by the Cauchy data at I admits a Killing vector if and only if the metric g admits a CKV ξ (which
is the restriction of the Killing vector to I ) and D satisfies the so-called Killing initial data (KID) equation.
This equation admits a natural generalization to arbitrary dimension which is
n+2
Lξ DAB + divg ξDAB = 0, (65)
n
where divg ξ is the divergence of ξ. Equation (65) reduces to the KID equation of Paetz in dimension n = 3 and
it is conformally covariant, i.e. if {gAB , DAB , ξ A } is a solution, then so it is {Ω2 gAB , Ω−(n+2) DAB , ξ A }. We
emphasize however, that in higher dimension (n ≥ 4) it is not known whether a spacetime admiting a smooth

38
I such that the corresponding data at null infinity solves the KID equation for some CKV ξ, must necessarily
admit a Killing vector.
A CKVF satisfying (65) will be called KID vector for short. An important property of KID vectors is that
they form a Lie subalgebra of CKVFs, i.e. if ξ, ξ ′ are KIDs for a given T T tensor D, then [ξ, ξ ′ ] is also a KID
for D. The problem of obtaining all TT-tensors with generality for a given conformal structure is hard, even
in the conformally flat case (see e.g. [3]). In this section we exploit the results above to obtain the general
solution of the KID equations in dimension n = 3 for spacetimes which possess two commuting symmetries, one
of which is axial. This case is specially relevant since n = 3 corresponds to the physical case of four spacetime
dimensions and the class necessarily contains the Kerr-de Sitter family of spacetimes, which is a particularly
interesting explicit familiy of spacetimes. Our strategy is to take an arbitrary CKVF ξ, derive its canonical
form ξF = ξe + µη, adapt coordinates to ξe and η and impose the KID equations to ξe and η. The problem
simplifies notably in the conformal gauge to g := (U ǫ )2 gEǫ
because both ξe and η become Killing vector fields.
From Remark 7.1, we may treat all cases σ < 0, σ = 0, σ > 0 at the same time by using the form of the metric
dz 2
g= + (z 2 − σ)dt2 + dφ2 , ξe = ∂t , η = ∂φ . (66)
z2−σ
We remark that even though we solve the problem by fixing the coordinates and conformal gauge, we shall
write the final result in fully covariant form (cf. Theorem 8.2 below).
In the conformal gauge of g, the condition that a TT-tensor D satisfies KID equations for both ξe and η
(which is equivalent to imposing that ξ and η are KID vectors) is trivial in the adapted coordinates obtained
in the previous section:
LξeDAB = ∂t DAB = 0, Lη DAB = ∂φ DAB = 0.
Thus, DAB are only functions of z. The transversality condition is also quite simple in adapted coordinates:

 
dDzz Dzz 2 tt
−z + (z − σ)D = 0, (67)
dz z2 − σ
dDzt 2z
+ 2 Dzt = 0 (68)
dz z −σ
dDzφ
= 0, (69)
dz
while the traceless condition imposes
Dzz
gAB DAB = + (z 2 − σ)Dtt + Dφφ = 0. (70)
z2 − σ
There are no equations for Dtφ so Dφt = h(z) with h(z) an arbitrary function. The general solution of equations
(68) and (69) is obtained at once and reads
K1
Dzt = , Dzφ = K2 , K1 , K2 ∈ R.
z2 − σ
For equations (67) and (70), we let Dzz =: f (z) be an arbitrary function and obtain the remaining components
1 df 1 df f
Dφφ = − , Dtt = − 2 .
z dy z(z 2 − σ) dz (z − σ)2
Summarizing

39
Lemma 8.1. In the three-dimensional conformally flat class [g], let ξF be a CKVF. Decompose ξ in canonical
form ξ = ξe+µη and fix the conformal gauge so that g given by (66). Then the most general symmetric TT-tensor
D satifying the KID equations for ξ and η simultaneously is, in adapted coordinates {z, t, φ}, a combination
(with constants) of the following tensors
 
1 df f 1 df
Df := f ∂z ⊗ ∂z + − ∂t ⊗ ∂t − ∂φ ⊗ ∂φ ,
z(z 2 − σ) dz (z 2 − σ)2 z dz
1
Dh := h(∂t ⊗ ∂φ + ∂φ ⊗ ∂t ), Dξe := 2 (∂z ⊗ ∂t + ∂t ⊗ ∂z ), Dη = ∂z ⊗ ∂φ + ∂φ ⊗ ∂z ,
z −σ
where f and h are arbitrary functions of z.
Having obtained the general solution in a particular gauge, our next aim is to give a (diffeomorphism and
conformal) covariant form of the generators in Lemma 8.1. From [23], we know that, for any CKV ξ of any
n-dimensional metric g (not necessarily conformally flat) the following tensors are TT w.r.t. to g and satisfy
the KID equation with respect to ξ.
!
1 |ξ|2g ♯
D(ξ) = n+2 ξ ⊗ ξ − g ,
|ξ|g n

where | · |g denotes the norm w.r.t. g and g ♯ the contravariant form of g. Thus, we can rewrite Df as
   
(z 2 − σ)3/2 df e − f 1 df
Df = −2(z 2 − σ)1/2 f + D(ξ) + D(η).
z dz z 2 − σ z dz

We now restore the conformal gauge freedom by considering the metric b b f = Df /Ω5 , for any
g = Ω2 g and D
e D(η) are already conformal and diffeomorphism covariant,
(positive) conformal factor Ω. Since the tensors D(ξ),
b
we must impose their multiplicative factors in Df to be conformal and diffeomorphism invariant. With the gauge
freedom restored, the norms of the CKVFs now are
p
e bg = Ω z 2 − σ,
|ξ| |η|bg = Ω.
√ √
Then, considering f =: Xfb(X) as function of the conformal invariant quantity X = |ξ|
e bg /|η|bg = z 2 − σ, one
b f in the following form:
can directly cast D
!
d b(X)
f  
Db f = X4 D( e − 1 d X3/2 fb(X) D(η),
ξ)
dX X3/2 X2 dX

which is a conformal and diffeomorphism covariant expression. Notice that the expression is symmetric under
the interchange ξe ↔ η because the coefficient of D(η) expressed in the variable Y = X−1 is identical in form to
the coefficient of D(ξ).
For the tensor D b h := Dh /Ω5 , redifining h =: b e −5/2 , it is immediate to write
h|ξ|

b
h
bh = D
D b b :=
h 5/2 e 5/2
(ξe ⊗ η + η ⊗ ξ),
e (71)
|η|bg |ξ| g
b

40
which is obviously conformal and diffemorphism covariant if and only if b h is conformal invariant, e.g. considering
b
h≡b h(X). We remark that the form (71) already appeared (with different powers due to the different dimension)
in the classification [24] of TT tensors in dimension two satisfying the KID equation.
For the remaining tensors D b e := D e/Ω5 and D b η := Dη /Ω5 , we define a conformal class of vector fields χ,
ξ ξ
which in the original gauge coincides with χ := ∂z . This vector is divergence-free ∇A χA = 0, and this equation
is conformally invariant provided the conformal weight of χ is −3 (i.e. for b g = Ω2 g, the corresponding vector is
−3 7
χ
b = Ω χ). We therefore impose this conformal behaviour of χ. The direction of χ is fixed by orthogonality
to ξe and η. The combination of norms that has this conformal weight and recovers the appropriate expression
e −1 |η|−2 (note that the orthogonality and norm conditions fix χ uniquely
in the gauge of Lemma 8.1 is |χ|bg := |ξ| g
b g
b
up to an irrelevant sign in any gauge). Thus, we may write
1 1
Dξe = (χ ⊗ ξe + ξe ⊗ χ), Dη = (χ ⊗ η + η ⊗ χ),
e2
|ξ| |η|2bg
g
b

which are conformally covariant expressions. Therefore, we get to the final result:
Theorem 8.2. Let ξ be a CKVF of the class of three dimensional conformally flat metrics and let ξ = ξe+ µη a
canonical form. For each conformal gauge, let us define a vector field χ with norm |χ|bg := |ξ| e −1 |η|−2 , orthogonal
g
b g
b
to ξe and η. Then, any TT-tensor satisfying the KID equations (65) for ξe and η is a combination (with constants)
of the following tensors:
!
d fb(X)   b
bb= X
D 4 e − 1 d X3/2 fb(X) D(η),
D(ξ) Dbb = h
(ξe ⊗ η + η ⊗ ξ),
e
f dX X 3/2 2
X dX h 5/2 e 5/2
|η|bg |ξ| g
b
1 1
Dξe = (χ ⊗ ξe + ξe ⊗ χ), Dη = (χ ⊗ η + η ⊗ χ),
e2
|ξ| |η|2bg
g
b

for arbitrary functions fb and b e bg /|η|bg .

h of X = |ξ|
Remark 8.1. The vector field χ defined in this Theorem is divergence-free. This property would have been
difficult to guess (and even to prove) in the original Cartesian coordinate system.
Remark 8.2. A corollary of this theorem is that the general solution of the Λ-vacuum Einstein field equation
in four dimensions with a smooth conformally flat null infinity and admitting an axial symmetric and a second
commuting Killing vecor can be parametrized by two functions of one variable and two constants. Recall that
in the Λ = 0 case, the general asympotically flat stationary and axially symmetric solution of the Einstein field
equations can be parametrized (in a neighbourhood of spacelike infinity, by two numerable sets of mass and
angular multipole moments (satisfying appropriate convergence properties), see [1], [4], [6] for details. There
is an intriguing paralelism between the two situations, at least at the level of crude counting of degrees of
freedom. This suggests that maybe in the Λ > 0 case it is possible to define a set of multipole-type moments that
characterizes de data at null infinity (and hence the spacetime), at least in the case of a conformally flat null
infinity. This is an interesting problem, but well beyond the scope of the present paper.
7 This choice may appear somewhat ad hoc at this point. However, the condition of vanishing divergence appears naturaly when

studying (for more general metrics) under which conditions a tensor ξ ⊗ W + W ⊗ ξ is a TT tensor satisfying the KID equation for
ξ. We leave this general analysis for a future work.

41
Remark 8.3. It is natural to ask whether Theorem 8.2 is general for TT-tensors admitting two commuting
e η, without the condition of η being conformally axial. In Appendix C of [23] one can explicitly find,
KIDs, ξ,
for an arbitrary CKVF ξ, the set C(ξ) of elements that commute with ξ. Then, from a case by case analysis,
one concludes that except in one special situation, for any linearly independent pair ξ, ξ ′ , with ξ ′ ∈ C(ξ) it is the
case that there is a CAKVF η ∈ C(ξ) such that span{ξ, η} = span{ξ, ξ ′ }. Thus, all these cases are covered by
Theorem 8.2. The exceptional case is when ξ, ξ ′ are conformal to translations. It is immediate to solve the TT
and KID equations for such a case directly in Cartesian coordinates.
The solution given in Theorem 8.2 provides a large class of initial data, which we know must contain the
so-called Kerr-de Sitter-like class with conformally flat I (see [23] for precise definition and properties of this
class), which in turn contains the Kerr-de Sitter family of spacetimes. It is interesting to identify this class
within the general solution given in Theorem 8.2. The characterizing property of the Kerr-de Sitter-like class
in the conformally flat case is D = D(ξ) for some CKVF ξ, where moreover, only the conformal class of ξ
matters to determine the family associated to the data. Decomposing canonically ξ = ξe+ µη, a straightforward
computation yields

X5 e + µ2 µX5/2 bb ,
D(ξ) = D(ξ) D(η) + D
(X2 + µ2 )5/2 (X2 + µ2 )5/2 (X2 + µ2 )5/2 h=1

which comparing with Theorem 8.2 yields the following corollary:

Corollary 8.2.1. The Kerr-de Sitter-like class with conformally flat I is determined by the TT-tensor DKdS =
bf + D
D b b with
h
1 X3/2 X5/2
fb = − 2 2 3/2
, b
h=µ 2 .
3 (X + µ ) (X + µ2 )5/2
It is also of interest to identify the the Kerr-de Sitter family. To that aim we combine the results in [23]
to those in the present paper to show that this family corresponds to σ < 0. The classification of conformal
classes of ξ in [23] is done in terms of the invariants bc = −c1 and b k = −c2 together with the rank parameter
r, where c1 and c2 are the coefficients of the characteristic polynomial of the skew-symmetric endomorphism F
associated to ξ. In terms of these objects, it is shown in [23] that the Kerr-de Sitter family corresponds to either
S1 = {b k > 0, bc ∈ R and r = 2}, or S2 = {b k = 0, b
c > 0 and r = 1}, the latter defining the Schwarzschild-de
Sitter family. It is immediate to verify that, since (cf. Corollary 3.3.1) bk = −σµ2 < 0 and b c = −σ − µ2 , then
S1 = {σ < 0, µ 6= 0} and S2 = {σ < 0, µ = 0} (the condition µ 6= 0 implies r = 2 and µ = 0 implies r = 1).
Thus, in terms of the classification developed in this paper, the Kerr-de Sitter family corresponds to σ < 0. It is
interesting that in the present scheme we no longer need to specify the rank parameter to identify the Kerr-de
Sitter family (unlike in [23]) and that the whole family is represented by an open domain. We emphazise that
the dependence in σ in the solutions given in Theorem 8.2 and Corollary 8.2.1 is implicit through the norm of
e
ξ.

Acknowledgements
The authors acknowledge financial support under the projects PGC2018-096038-B-I00 (Spanish Ministerio de
Ciencia, Innovación y Universidades and FEDER) and SA083P17 (JCyL). C. Peón-Nieto also acknowledges the
Ph.D. grant BES-2016-078094 (Spanish Ministerio de Ciencia, Innovación y Universidades).

42
References
[1] A. E. Aceña. Convergent null data expansions at space-like infinity of stationary vacuum solutions. Annales
Henri Poincaré, 10, 275–337, 2009.
[2] M. T. Anderson. Existence and stability of even-dimensional asymptotically de Sitter spaces. Annales
Henri Poincaré, 6, 801–820, 2005.
[3] R. Beig. TT-tensors and conformally flat structures on 3-manifolds. Banach Center Publications, 41,
109–118, 1997.
[4] R. Beig and B. G. Schmidt. Time independent gravitational fields. Lect. Notes Phys., 540, 325–372, 2000.
[5] N. Burgoyne and R. Cushman. Conjugacy classes in linear groups. Journal of Algebra, 44, 339 – 362, 1977.
[6] T. Bäckdahl. Axisymmetric stationary solutions with arbitrary multipole moments. Classical and Quantum
Gravity, 24, 2205–2215, 2007.
[7] Y. Choquet-Bruhat. General Relativity and the Einstein Equations. Oxford Mathematical Monographs.
OUP Oxford, 2008.
[8] R. Cushman and W. van der Kallen. Adjoint and coadjoint orbits of the poincaré group. Acta Applicandae
Mathematicae, 90, 65–89, 2006.
[9] D. Z. Djoković, J. Patera, P. Winternitz, and H. Zassenhaus. Normal forms of elements of classical real
and complex lie and jordan algebras. Journal of Mathematical Physics, 24, 1363–1374, 1983.
[10] F. Fernández-Álvarez and J. M. M. Senovilla. Gravitational radiation condition at infinity with a positive
cosmological constant. Phys. Rev. D, 102, 101502, 2020.
[11] Y. Fourès-Bruhat. Théorème d’existence pour certains systèmes d’équations aux dérivées partielles non
linéaires. Acta Math., 88, 141–225, 1952.
[12] H. Friedrich. Existence and structure of past asymptotically simple solutions of Einstein’s field equations
with positive cosmological constant. Journal of Geometry and Physics, 3, 101 – 117, 1986.
[13] H. Friedrich. Conformal Einstein Evolution. Springer Berlin Heidelberg, Berlin, Heidelberg, 2002.
[14] H. Friedrich and A. Rendall. The cauchy problem for the einstein equations. In Bernd G. Schmidt,
editor, Einstein’s Field Equations and Their Physical Implications, Berlin, Heidelberg, 2000. Springer Berlin
Heidelberg.
[15] F.R. Gantmacher. The theory of matrices. Vol. 1 The Theory of Matrices. Chelsea Pub. Co., 1960.
[16] R. Goodman and N.R. Wallach. Symmetry, Representations, and Invariants. Graduate Texts in Mathe-
matics. Springer New York, 2009.
[17] B.C. Hall. Lie Groups, Lie Algebras, and Representations: An Elementary Introduction. Graduate Texts
in Mathematics. Springer, 2003.
[18] G. S. Hall. Symmetries And Curvature Structure In General Relativity. World Scientific Lecture Notes In
Physics. World Scientific Publishing Company, 2004.

43
[19] D. Ida. Complete integrability of cohomogeneity-one strings in Rn,1 and canonical form of Killing vector
algebra. Journal of Mathematical Physics, 61, 082501, 2020.
[20] A.W. Knapp. Lie Groups Beyond an Introduction. Progress in Mathematics. Birkhäuser Boston, 2002.
[21] K. Lange. Optimization. Springer Texts in Statistics. Springer New York, 2013.
[22] A. Lichnerowicz. Théories relativistes de la gravitation et de l’électromagnétisme: relativité générale et
théories unitaires. Collection d’ouvrages de mathématiques à l’usage des physiciens. Masson, 1955.
[23] M. Mars, T-T Paetz, and J. M. M. Senovilla. Classification of Kerr–de Sitter-like spacetimes with confor-
mally flat I . Classical and Quantum Gravity, 34, 095010, 2017.
[24] M. Mars and C. Peón-Nieto. Skew-symmetric endomorphisms in M1,3 : A unified canonical form with
applications to conformal geometry. Classical and Quantum Gravity, 2020. DOI: 10.1088/1361-6382/abc18a
[25] T-T Paetz. Killing Initial Data on spacelike conformal boundaries. Journal of Geometry and Physics, 106,
51 – 69, 2016.
[26] R. Penrose and W. Rindler. Spinors and Space-Time: Volume 1, Two-Spinor Calculus and Relativistic
Fields. Cambridge Monographs on Mathematical Physics. Cambridge University Press, 1984.
[27] G. Y. Rainich. Electrodynamics in the general relativity theory. Proceedings of the National Academy of
Sciences, 10, 124–127, 1924.
[28] M. Schottenloher. A Mathematical Introduction to Conformal Field Theory. Lecture Notes in Physics.
Springer Berlin Heidelberg, 2008.
[29] J.L. Synge. Relativity: The Special Theory. Series in physics. North-Holland Publishing Company, 1956.

[30] L. B. Szabados and P. Tod. A review of total energy–momenta in gr with a positive cosmological constant.
International Journal of Modern Physics D, 28, 1930003, 2019.

44