MaThCryst summer school 15-20 July 2007, Havana, Cuba

Space groups
Bernd Souvignier

Definition 1 A crystal pattern is a set of points in Rn such that the translations leaving
it invariant form a (vector) lattice in Rn .

Definition 2 A space group is a group of isometries of Rn (i.e. of mappings of Rn

preserving all distances) which leaves some crystal pattern invariant.
A typical example of a 2-dimensional crystal pattern is displayed in Figure 1. Of
course, the figure only displays a finite part of the pattern which is assumed to be
infinite, but the continuation of the pattern should be clear from the displayed excerpt.
◦ ◦ ◦ ◦ ◦
◦ ◦ ◦ ◦ ◦
◦ ◦ ◦ ◦ ◦
◦ ◦ ◦ ◦ ◦

◦ ◦ ◦ ◦ ◦
◦ ◦ ◦ ◦ ◦
◦ ◦ 6 ◦ ◦ ◦
◦ ◦ ◦ ◦ ◦

◦ ◦ ◦ ◦ ◦
◦ ◦ - ◦ ◦ ◦
◦ ◦ ◦ ◦ ◦
◦ ◦ ◦ ◦ ◦

Figure 1: Crystal pattern in 2-dimensional space.

Remark: The pattern in Figure 1 was actually obtained as the orbit of some point
under a space group G which in turn is just the group of isometries of this pattern. This
observation already indicates that space groups can be investigated without explicit
retreat to a crystal pattern, since a crystal pattern for which a space group is its group
of isometries can always be constructed as the orbit of a (suitably chosen) point.
It is fairly obvious that the space group of the crystal pattern in Figure 1 contains trans-
lations along the indicated vectors and that it also contains fourfold rotations around
the centers of each block of 4 points.
It is the purpose of this and the following sessions, to find an appropriate description
of space groups which on the one hand reflects the geometric properties of the group
elements and on the other hand allows to classify space groups under various aspects.
Although the application to 2- and 3-dimensional crystal patterns is the most interest-
ing, it costs almost no extra effort to develop the concepts for arbitrary dimensions
n. We will therefore formulated most statements for general dimension n, but will
illustrate them in particular for the cases n = 2 and n = 3.

1
1 Space group elements
Before we have a closer look at the elements of space groups, we briefly review some
concepts from linear algebra.

1.1 Linear mappings

Definition 3 A linear mapping g on the n-dimensional space Rn is a map that respects
the sum and the scalar multiplication of vectors in Rn , i.e. for which:

(i) g(v + w) = g(v) + g(w) for all v, w ∈ Rn ;

(ii) g(α · v) = α · g(v) for all v ∈ Rn , α ∈ R.

Note: Since a linear mapping g respects linear combinations, it is completely deter-

mined by the images g(v1 ), . . . , g(vn ) on a basis (v1 , . . . , vn ):

g(α1 · v1 + α2 · v2 + . . . + αn · vn ) = α1 · g(v1 ) + α2 · g(v2 ) + . . . + αn · g(vn )

Moreover, once the images of the basis vectors are known, the image of an arbitrary
linear combination w = α1 · v1 + α2 · v2 + . . . + αn · vn only depends on its coordinates
with respect to the basis.

Definition 4 Let (v1 , . . . , vn ) be a basis of Rn and let w = α1 ·v1 +α2 ·v2 +. . .+αn ·vn
be an arbitrary vector of Rn , written as a linear combination of the basis vectors.
Then the α called the coordinates of w with respect to the basis (v1 , . . . , vn ) and
i are
α1
the vector  ...  is called the coordinate vector of w with respect to this basis.
 

αn
     
1 1 2 x
Example: Choose , as basis of R . Then the coordinate vector of
0 1 y
 
x−y
is , since
y
         
x 1 1 x−y y
= (x − y) · +y· = +
y 0 1 0 y

Note: If we choose the standard basis

      
1 0 0
 0 1 0
e1 =  ..  , e2 =  ..  , . . . , en =  .. 
      
 . .  . 
0 0 1

for Rn , then each column vector coincides with its coordinate vector.
For every basis, the coordinate vectors of the basis are the vectors of the standard basis,
since vi = 0 · v1 + . . . + 0 · vi−1 + 1 · vi + . . . + 0 · vn . Therefore it is useful to work
with coordinate vectors, since that turns every basis into the standard basis.

2
Since linear mappings are determined by their images on basis vectors, it is very conve-
nient to describe them by matrices which provide the coordinate vectors of the images
of the basis vectors.

Definition 5 Let (v1 , . . . , vn ) be a basis of Rn and let g be a linear mapping of Rn .

Then g can be described by the n × n matrix
 
a11 a12 . . . a1n
 a21 a22 . . . a2n 
A= .
 
.. .. .. 
 .. . . . 
an1 an2 . . . ann

which has as its j-th column the coordinate vector of the image g(vj ) of the j-th basis
vector, i.e.
g(vj ) = a1j v1 + a2j v2 + . . . + anj vn
If w = α1 · v1 + α2 · v2 + . . . + αn · vn is an arbitrary vector of Rn , then the coordinate
vector of its image under g is given by the product of the matrix A with the coordinate
vector of w:    
α1 β1
 ..   .. 
A· . = . 
αn βn
denotes that g(w) = β1 · v1 + β2 · v2 + . . . + βn · vn .
Examples:
2 the first is the standard basis
(1) The
 following
  figure shows
 two bases of  R ,   
1 0 ′ 1 ′ −1
v1 = , v2 = the second is v1 = , v2 = .
0 1 1 1

v2′ v2 v1′
I 6 

-
v1

We consider the linear mapping g which is the reflection in the dashed line (the
x-axis).Since v
1 7→ v1 , v2 7→ −v2 , with respect to the standard basis g has the
1 0
matrix .
0 −1
On the other hand, we have v1′ → 7→ −v1′ , hence with respect to the
7  −v2′ , v2′ 
0 −1
alternative basis, g has the matrix .
−1 0

3
 (2) The hexagonal lattice has a threefold rotation g as symmetry operation.

v2
v2′ 6
K

- v1

With respect to the standard basis (v1 , v2 ), this rotation has the matrix
√ !
1 3
−√2
− 2
3
.
2 − 21
!!
1
−
 
1 ′ 2
However, if a symmetry adapted basis v1 = , v2 = √3 is chosen,
0 2
the matrix of g becomes much simpler, since g(v1 ) = v2′ and g(v2′ ) = −v1 − v2′ .
The matrix of g with respect to this basis is thus
 
0 −1
.
1 −1

In the context of symmetry operations, we have to make sure that a transformation can
be reversed, i.e. that it has an inverse transformation such that the composition of the
two mappings is the identity operation.

Definition 6 A linear mapping g is called invertible if there is a linear mapping g−1

such that gg−1 = g−1 g = id, where id denotes the identity mapping leaving every
vector unchanged, i.e. id(v) = v for all vectors v ∈ Rn .

Lemma 7 A linear mapping g is invertible if and only if the images g(v1 ), . . . , g(vn )
of a basis (v1 , . . . , vn ) of Rn form again a basis of Rn , i.e. are linearly independent.

Definition 8 The set of invertible linear mappings on Rn forms a group. The group of
corresponding n × n matrices is denoted by GLn (R) (for general linear group).

1.2 Affine mappings

The following argument shows that elements from space groups have to be mappings
of a special kind, namely affine mappings:
Let o be the (chosen) origin of Rn and let s be an isometry in a space group, then we
denote by t the translation by the vector s(o) − o. Since a translation is an isometry,
the mapping s − t is also an isometry and by construction it fixes the origin o.

4
It is an elementary (but not so well-known) fact that an isometry fixing the origin
actually has to be an invertible linear mapping g, hence the isometry s is what is called
an affine mapping: the sum of an invertible linear mapping and a translation.

Lemma 9 Each element of a space group is the sum of an invertible linear mapping
and a translation, i.e. an affine mapping.
Since the elements of space groups are affine mappings, we will now investigate in
some more detail the properties of groups of affine mappings.

Definition 10 The affine group An of degree n is the group of all mappings {g | t} (in
Seitz notation) consisting of a linear part g ∈ GLn (R) (i.e. an invertible n × n matrix)
and a translation part t ∈ Rn .
The elements of An act act as
{g | t}(v) := g · v + t
on the vectors v of the vector space Rn .
Note that the linear part of an element {g | t} has to be an invertible matrix, since
otherwise the element would not have an inverse.
Examples:
(1) In dimension 2, a reflection in the line x = y is given by
   
0 1 0
{g | t} =
1 0 0
which acts as
           
0 1 0 x y 0 y
= + = .
1 0 0 y x 0 x
1
(2) A glide reflection with shift 2 along the x-axis is given by
   1 
1 0 2
{g | t} =
0 −1 0
and acts as
  1       1  
x + 21
 
1 0 2 x x 2
= + = .
0 −1 0 y −y 0 −y
1
(3) In 3-dimensional space, a fourfold screw rotation with a shift of 4 around the
z-axis is given by
   
 0 −1 0 0 
{g | t} = 1 0 0  0 
1 
0 0 1

4
and acts as
           
 0 −1 0 0  x −y 0 −y
1 0 0  0   y  =  x  +  0  =  x  .
1  1
0 0 1 z z z + 14

4 4

5
Since we are only interested in isometries, we only have to deal with the subgroup of
the affine group An which consists of the elements preserving all distances. Clearly,
translations keep all distances and a linear mapping is an isometry if and only if its
matrix g is an orthogonal matrix (with respect to the standard basis), i.e. gtr · g = id.

Definition 11 The group En := {{g | t} ∈ An | gtr = g−1 } of affine mappings with

orthogonal linear part is called the Euclidean group.
In particular, every space group is a subgroup of the Euclidean group. Before we focus
on space groups, we state some important facts about the affine group, which will be
used in the discussion of space groups.

1.3 Basic properties of the affine group

It is a fruitful exercise to compute the product of two affine mappings explicitly. For
that, we apply the product of the two elements {g | t} and {h | u} to an arbitrary
vector v:
({g | t} · {h | u})(v) = {g | t}(h · v + u) = g · (h · v + u) + t = gh · v + g · u + t
= {gh | g · u + t}(v).

Lemma 12 The product of two affine mappings {g | t} and {h | u} is given by

{g | t} · {h | u} = {gh | g · u + t}.

The short computation above thus shows, that the linear parts of two affine mappings
are simply multiplied, but that the translation part is not just the sum of the two transla-
tion parts, but that the translation part u of the second element is twisted by the action
of the linear part g of the first element.
By Lemma 12 it is also easy to derive the inverse of an element, since for {h | u}
being the inverse of {g | t} we require h = g−1 and g · u = −t, thus u = −g−1 · t.

Lemma 13 The inverse of the affine mapping {g | t} is given by

{g | t}−1 = {g−1 | −g−1 · t}.

An important way of investigating subgroups of the affine group hinges on the fact that
the linear parts are just multiplied. This means that forgetting about the translation part
results in a homomorphism from the affine group to the matrix group GLn (R).

Theorem 14 Let Π be the mapping Π : An → GLn (R) : {g | t} 7→ g which forgets

about the translation part of an affine mapping.

(i) The mapping Π is a group homomorphism from An onto GLn (R) with kernel
T := {{id | t} | t ∈ Rn } and image GLn (R).

(ii) An contains a subgroup isomorphic to the image GLn (R) of Π, namely the
group G = {{g | 0} | g ∈ GLn (R)} of elements with trivial translation part.

6
 (iii) Every element {g | t} can be written as {g | t} = {id | t} · {g | 0}, thus
An = T · G. Since on the other hand the intersection T ∩ G consists only
of the identity element {id | 0}, the affine group An is the semidirect product
T ⋊ GLn (R) of T and GLn (R).
The homomorphism Π can be applied to every subgroup G ≤ An of the affine group,
it has the group of linear parts as image and the normal subgroup of translations in G
as its kernel. The homomorphism Π therefore allows to split a space group G into two
parts.

Definition 15 Let G be a space group and let Π be the homomorphism defined in

Theorem 14.
(i) The translation subgroup T := {{id | t} ∈ G} is the kernel of the restriction of
Π to G.

(ii) The group P := Π(G) of linear parts in G is called the point group of G. It is
isomorphic to the factor group G/T .
Note: In general, a subgroup G ≤ An is not the semidirect product of its translation
subgroup and its point group. For space groups, only the symmorphic groups are
semidirect products, whereas groups containing e.g. glide reflections with a glide not
contained in their translation subgroup do not contain their point group as a subgroup.
Exercise 1.
Prove that two affine mappings {g | t} and {h | u} commute (i.e. {g | t} · {h | u} =
{h | u} · {g | t}) if and only if
(i) the linear parts g and h commute;

(ii) the translation parts fulfill (g − id) · u = (h − id) · t.

Conclude that an arbitrary affine mapping {g | t} commutes with a translation {id | u}
if and only if u is fixed under the action of g.

1.4 Matrix notation

A very convenient way of representing affine mappings are the so-called augmented
matrices.

Definition 16 The augmented matrix of an affine mapping {g | t} with linear part

g ∈ GLn (R) and translation part t ∈ Rn is the (n + 1) × (n + 1) matrix
 
g t
.
0...0 1

In order to apply such an augmented matrix to a vector v ∈ Rn , the vector is also

augmented by an additional component of value 1. The usual left-multiplication of a
vector by a matrix now gives
     
g t v g·v+t
· =
0 1 1 1

7
and ignoring the additional component yields the desired result.
We also check that the product of the augmented matrices coincides with the product
of affine mappings as given in Lemma 12. By usual matrix multiplication we get:
     
g t h u gh g · u + t
· = ,
0 1 0 1 0 1
thus the linear part of the product is gh and the translation part is g · u + t as required.
Note: In view of the representation of affine mappings by augmented matrices, the
homomorphism Π becomes very natural, it just picks the upper left n × n submatrix
of an (n + 1) × (n + 1) augmented matrix.
Examples
(1) p4mm
If we take as crystal pattern the lattice points of a common square lattice, the
group of isometries of this pattern is the group generated by a rotation of order
4, the reflection in the x-axis and the two unit translations along the x- and
y-axis. These four elements are given by the matrices
       
0 −1 0 1 0 0 1 0 1 1 0 0
 1 0 0  ,  0 −1 0  ,  0 1 0  ,  0 1 1  .
0 0 1 0 0 1 0 0 1 0 0 1

(2) c2mm
If the crystal pattern consists of the lattice points of a rectangular lattice and
the centers of the rectangles, the space group of this pattern is generated by two
reflections in the x- and y-axis and translations to the centers of two adjacent
rectangles. These generators are given by the matrices
1 0 12 1 0 12
       
1 0 0 −1 0 0
 0 −1 0  ,  0 1 0  ,  0 1 1  ,  0 1 − 1  .
2 2
0 0 1 0 0 1 0 0 1 0 0 1

(3) P41
In this example a 3-dimensional crystal pattern is assumed that in addition to
the translations only allows a fourfold screw rotation which after 4 applications
results in a unit translation along the z-axis. This space group is generated by
the matrices
       
0 1 0 0 1 0 0 1 1 0 0 0 1 0 0 0
−1 0 0 0   0 1 0 0   0 1 0 1   0 1 0 0 
 0 0 1 1 ,  0 0 1 0 ,  0 0 1 0 ,  0 0 1 1  .
       
4
0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1

Exercise 2.
Two space group elements are given by the following transformations:
z + 21
       
x x −y
g : y  → x + 21  , h : y  → x + 12  .
z −y z z + 21
Determine the augmented matrices for g and h and compute the products g · h and h · g.

8
2 Analysis of space groups
We have already noted that every space group is a subgroup of the Euclidean group
En and that it can be split into its translation subgroup T and its point group P via the
homomorphism Π. We will now deduce some more properties of the point group P .
Since the observation of the following theorem is crucial for the analysis of space
groups we include its proof (which is very short).

Theorem 17 Let G be a space group, let P = Π(G) be its point group and denote by
L the (vector) lattice
L = {v | {id | v} ∈ T }
of translation vectors in T . Then P acts on the lattice L, i.e. for v ∈ L and g ∈ P one
has g · v ∈ L.
 
id v
Proof: Since T is a normal subgroup of G, conjugating the element by an
0 1
 −1 
g t
element ∈ G gives again an element of T . Working out this conjugation
0 1
explicitly gives:
−1 
g−1 t g−1 t g−1 t + v
     
id v g −g · t
=
0 1 0 1 0 1 0 1 0 1
 
id g · v
= ∈ T.
0 1

This shows that indeed g · v ∈ L and hence the point group P acts on the lattice L. 

Note: The distinction between the translation subgroup T of a space group and its
translation lattice L may seem somewhat artificial, since by the mapping {id | t} →
t the two groups are clearly isomorphic. However, since we multiply space group
elements, but add lattice vectors, it is good practice to keep the two notions apart.
By now we have worked out that the point group P of G is a group of isometries and
that is acts on the lattice L of translations in G. This means that P is a subgroup of the
automorphism group

Aut(L) := {g ∈ GLn (R) | gtr = g−1 , g(L) = L}

of L. From this fact we now can prove that P is a finite group.

Theorem 18 The point group P of a space group G is finite.

Proof: We fix some lattice basis (v1 , . . . , vn ) of L and assume that vn is the longest of
these basis vectors (they may of course all have the same length). Since L is a lattice,
it is in particular discrete, hence it contains only finitely many vectors of length at most
kvn k. Since an automorphism of L preserves lengths, it can only permute vectors of
the same length. But for a finite set of m elements there are at most m! permutations,
and since every element of P is determined by its action on the lattice basis, there are
only finitely many possibilities for the elements of P . 

9
2.1 Transformation of a space group to a lattice basis
The observation that the point group P acts on the translation lattice L gives rise to a
change of perspective:
New point of view: Instead of writing all vectors and matrices (and hence the aug-
mented matrices) with respect to the standard basis of Rn (as we did so far) it is con-
venient to transform the elements of a space group to a lattice basis of its translation
lattice.

Lemma 19 Let G be a space group written with respect to some basis B of Rn (e.g.
the standard basis). Let X be the matrix of a basis transformation to a new basis B ′ of
Rn , i.e. the columns of X are the coordinate vectors of the vectors in B ′ with respect
to the basis B.
Then writing out the conjugation by X:
 −1   −1
X gX X −1 · t
  
X 0 g t X 0
=
0 1 0 1 0 1 0 1

shows that with respect to the new basis B ′ the element {g | t} of G is transformed to
the element
{g′ | t′ } = {X −1 gX | X −1 · t}.
In particular, if (v1 , . . . , vn ) is a lattice basis of the translation lattice of G and X is the
transformation matrix to this lattice basis, then the translation {id | vi } is transformed
to
{id | X −1 · vi } = {id | ei }
where ei is the i-th unit vector having 1 in its i-th coordinate and 0 else.
Writing a space group with respect to a lattice basis (v1 , . . . , vn ) of its translation
lattice L has the following consequences:

• All vectors v ∈ Rn are given as coordinate  vectors with respect to the basis
x1
 .. 
(v1 , . . . , vn ), i.e. the coordinate vector  .  denotes the vector v = x1 v1 +
xn
. . . + xn vn .

• In particular, the translation lattice L becomes L = Zn , since the lattice vectors

are precisely the integral linear combinations of a lattice basis.

• The translation subgroup T of G becomes T = {{id | t} | t ∈ Zn }.

• The point group P becomes a subgroup of GLn (Z), since the images of the
vectors in the lattice basis are again lattice vectors and thus integral linear com-
binations of the lattice basis.

The price we pay for this transformation to a lattice basis is that the point group no
longer consists of orthogonal matrices for which gtr = g−1 holds, but that they fix the
metric tensor of the lattice basis.

10
Definition 20 For a basis B = (v1 , . . . , vn ) the metric tensor of B is the matrix F ∈
Rn×n of dot products of the basis vectors, i.e. Fij = vi ◦ vj .
If X is the matrix with vi as i-th column, then the metric tensor is given by F =
X tr · X.

Theorem 21 If a space group G is written with respect to a basis (v1 , . . . , vn ), then the
metric tensor F of this basis is invariant under transformations from the point group P
of G, i.e.
gtr F g = F for each g ∈ P.
In particular, if G is written with respect to a lattice basis of its translation lattice, the
point group elements fix the metric tensor of the lattice basis.
Proof: Let g ∈ P be an element from the point group of G written with respect to
the basis (v1 , . . . , vn ) and denote by g′ the same element written with respect to the
standard basis. Let X be the matrix of the basis transformation from the standard
basis to the new basis, i.e. the matrix with vi as i-th column. Then the rules for basis
transformations state that g = X −1 g′ X and g′ = XgX −1 .
For the orthogonal matrix g′ we know that g′tr g′ = id and replacing g′ by XgX −1
gives

gtr g′ = id ⇒ X −tr gtr X tr XgX −1 = id ⇒ gtr X tr Xg = X tr X.

The metric tensor F = X tr X of the basis (v1 , . . . , vn ) is thus preserved by g. 

By a slight variation of the above proof one deduces how a metric tensor is transformed
under a basis transformation.

Corollary 22 If F is the metric tensor of a basis B = (v1 , . . . , vn ) and X is the basis

transformation to a new basis (v1′ , . . . , vn′ ), then the metric tensor of B ′ is given by

F ′ = X tr F X.

In particular, if the metric tensor F is invariant under a point group P and P is trans-
formed to a new basis by the basis transformation X, i.e. to P ′ = {X −1 ·g·X | g ∈ P },
then P ′ fixes the metric tensor X tr F X.
Exercise 3.
Prove the above corollary, i.e. show that if gtr F g = F for all g ∈ P and P ′ =
{X −1 · g · X | g ∈ P }, then g′tr X tr F Xg′ = X tr F X for all g′ ∈ P ′ .
Note: It is often the case that a space group is neither given with respect to the stan-
dard basis of Rn nor with respect to a lattice basis of its translation lattice, but with
respect to another convenient basis. This is for example the case for the group c2mm
in
theexamples
  above. The matrices there are given with respect to the (obvious) basis
a 0
, , a 6= b of a rectangular lattice.
0 b
 2 
a 0
In this case the point group fixes the metric tensor .
0 b2
Example: In the examples above we gave the space group c2mm of the centered
rectangular lattice with respect to a basis of the rectangular lattice. This resulted in

11
translations
 a with
 anonintegral
 coordinates. If we transform this group to the lattice
basis 2 , 2 , the generators given above are transformed to
b b
2 − 2
       
0 1 0 0 −1 0 1 0 1 1 0 0
 1 0 0  ,  −1 0 0  ,  0 1 0  ,  0 1 1 .
0 0 1 0 0 1 0 0 1 0 0 1
Written with respect to the lattice basis, the point group fixes the metric tensor
1 a2 + b2 a2 − b2
 
.
4 a2 − b2 a2 + b2

Important note: The transformation to a lattice basis is not a standard point of

view taken in crystallography. Here, one often distinguishes between primitive
lattices, where the conventional cell spanned by the basis vectors contains just
one lattice point and centered lattices, where it contains more than one lattice
point. In a centered lattice, not all translations have integral coordinates and
the translation lattice thus is actually larger than the lattice generated by the
chosen basis.
The reason for this distinction between primitive and centered lattices is that it
is often convenient to work with special kinds of bases which are regarded as
particularly nice and simple. In dimensions 2 and 3 it is indeed the case that
for families of lattices which are contained in each other (like the cubic lattice
and its centerings), one of these lattices has a particularly nice basis (such as the
standard basis for the cubic lattice). Moreover, in these dimensions the primitive
lattice and its centerings almost always have isomorphic automorphism groups
(the only exception being the hexagonal lattice and the rhombohedral lattice as
a centering of it).
However, the same is no longer true in higher dimensions. There it is often
impossible to distinguish one of the lattices in a family as primitive lattice, since
none of the lattices may have a basis with particular nice geometric properties.
Moreover, the automorphism groups of the lattices in one family may differ sub-
stantially, and not always the most simple one has the largest symmetry group.
Two examples may illustrate this:
– In dimension 4, the standard lattice Z4 generated by the standard basis has
a symmetry group of order 384. It has a sublattice of index 2 which has a
symmetry group of order 1152, i.e. larger by a factor of 3. (This sublattice
is the so-called root lattice of type F4 and has the corners of the regular
polytope called the 24-cell as vectors of minimal length.)
– In dimension 8 the situation is even more intriguing: The root lattice of
type E8 might be regarded as a centering of the 8-dimensional checker-
board lattice D8 , which in turn is the sublattice of all vectors with even
coordinate sum in the standard lattice Z8 . Both the standard lattice and the
checkerboard lattice have a symmetry group of order 10321920, whereas
the E8 lattice has a much larger symmetry group of order 696729600.

12
 Finally, for the interplay between the translation subgroup and the point group
of a space group, it is extremely convenient to use the property that the trans-
lations are integral vectors, and we therefore will always assume that a space
group is written with respect to a lattice basis.
From the perspective of classical crystallography, this may look like we are only
dealing with primitive on not with centered lattices, but in our approach sim-
ply all lattices are regarded as primitive, since there is no reasonable general
concept that allows to distinguish between primitive and centered lattices.

Exercise 4.
The point group P (in the arithmetic class 3m1P) is generated by the matrices
   
0 1 0 0 1 0
g = −1 1 0  , h = 1 0 0  .
0 0 −1 0 0 −1
 
2a −a 0
(i) Check that P fixes the metric tensor F = −a 2a 0. It thus acts on a
0 0 b
hexagonal lattice.

(ii) P also acts on a rhombohedral lattice, which is obtained from the above hexag-
onal lattice by the basis transformation
   
−1 2 −1 0 −1 1
1
X = −2 1 1  with inverse transformation X −1 =  1 0 1
3
1 1 1 −1 1 1

Transform the metric tensor F of the hexagonal lattice to the metric tensor of
the rhombohedral lattice (with the columns of X as lattice basis).

(iii) Transform P to the rhombohedral lattice (thus obtaining a point group P ′ in the
arithmetic class 3mR) and check that the so obtained point group fixes the metric
tensor computed in (ii).

2.2 Systems of nonprimitive translations

We already remarked that in general a space group is not a semidirect product of its
translation subgroup and its point group, since it does not necessarily contain a sub-
group isomorphic to the point group.
Example: The smallest example for a space group that is not a semidirect product is
the space group G with point group of order 2 acting on a rectangular lattice such that
the nontrivial element of the point group is induced by a glide reflection g:

1 0 12
 

g =  0 −1 0  .
0 0 1

13
Any product of g with a translation has  component along the x-axis of
 a translation
1 0 1
the form 12 + k with k ∈ Z and g2 =  0 1 0  is itself also a translation. Hence
0 0 1
the space group G has besides the identity element no elements of finite order and in
particular no subgroup of order 2, isomorphic to its point group.
Although the point group P may not be found as a subgroup, it still plays an important
role for the description of the elements of G, since P is isomorphic to the factor group
G/T .

Definition 23 For a subgroup T ≤ G, a (right) coset of T is a set of the form

T g = {tg | t ∈ T } for some g ∈ G.
Two cosets are either equal or disjoint.
A set {g1 , . . . , gr } of elements in G is called a set of coset representatives or transver-
sal for T in G if G is the disjoint union of the cosets T g1 , T g2 , . . . , T gr , i.e. if
G = T g1 ∪˙ T g2 ∪˙ . . . ∪˙ T gr .
In the case of space groups, one has {id | v} · {g | t} = {g | t + v}, hence all elements
in a coset of T have the same linear part. This implies that every transversal of T in G
has to contain each linear part of P precisely once.

Lemma 24 Every transversal of the translation subgroup T in a space group G with

point group P is of the form {{g | tg } | g ∈ P }. It contains precisely one element for
each element g in the point group P of G.
Remark: A transversal of T in G is quite useful to construct (a reasonable part of) an
orbit of G on Rn which in general will be a crystal pattern having G as its space group.
For that, choose a point p ∈ Rn and apply all elements of the transversal to p. If not
all of the so obtained points are different or if two of these points differ by a lattice
vector, the point p is in special position and its orbit may have a space group differing
from G. Otherwise, the point p is in general position and the full orbit of p under G is
obtained by translating the |P | points obtained from the transversal by lattice vectors.
The space group of the orbit of a point in general position is precisely G.
Example: Figure 2 below displays the orbit of the space group G = p2gg which has
a transversal
1 0 12 −1 0 12
       
1 0 0 −1 0 0
 0 1 0  ,  0 −1 1  ,  0 1 1  ,  0 −1 0 
2 2
0 0 1 0 0 1 0 0 1 0 0 1
with respect to its translation subgroup T corresponding to a rectangular lattice.
It is convenient to plot an asymmetric symbol at the positions of the orbit points instead
of a point, since this allows to recognize reflections and rotations moreeasily.

0.2
As point p of which the orbit is calculated, we choose the point p = and we
0.15
plot the symbol ◦⌊ at each position in the appropriate orientation (the point p has the
symbol •⌊).

14
 ◦⌈ ◦⌈ ◦⌈
◦⌊ ◦⌊ ◦⌊
◦⌉ ◦⌉ ◦⌉
◦⌋ ◦⌋ ◦⌋
◦⌈ ◦⌈ ◦⌈
◦⌊ ◦⌊ ◦⌊
◦⌉ 6 ◦⌉ ◦⌉
◦⌋ ◦⌋ ◦⌋
◦⌈ ◦⌈ ◦⌈
•⌊ ◦⌊ ◦⌊
-
◦⌉ ◦⌉ ◦⌉
◦⌋ ◦⌋ ◦⌋

Figure 2: Orbit of a point in general position under space group p2gg.

Definition 25 Let {{g | tg } | g ∈ P } be a transversal of T in G. Then the set

{tg | g ∈ P } of translation parts in this transversal is called a system of nonprimitive
translations or translation vector system which we will abbreviate as SNoT.
Of course, the transversal and thus the SNoT is by no means unique, since each tg
may be altered by a vector from the translation lattice. This means in particular that an
element tg which lies in T can be replaced by the 0-vector.
This also explains the term ’nonprimitive translation’, since one may assume that the
elements of the SNoT lie inside the unit cell of the lattice, and are therefore vectors
with non-integral coordinates (or 0).
From the multiplication rule of affine mappings we can deduce an important property
of a SNoT.

Theorem 26 The product {g | tg } · {h | th } = {gh | g · th + tg } lies in the same coset

of T as the element {gh | tgh }, therefore the elements of a SNoT conform with

tgh = g · th + tg + t for some t ∈ T

which we call the product condition, abbreviated as

tgh ≡ g · th + tg mod T.

In particular, a SNoT is completely determined by its values on generators of the point

group, since the value on products follows via the product condition.
If we assume that a space group is written with respect to a lattice basis, we can assume
that the elements of its SNoT have coordinates 0 ≤ xi < 1, since adjusting them by
lattice vectors means to alter their coordinates by values in Z. This actually makes the
SNoT unique.

Definition 27 A space group G that is written with respect to a lattice basis of its
translation lattice is determined by:

15
 • the metric tensor F of the lattice basis;

• a finite group P ≤ GLn (Z) fixing the metric tensor F ;

• a SNoT {tg | t ∈ P } with coordinates in the interval [0, 1)

The space group can then be written as:

  
g tg + t n
G= | g ∈ P, t ∈ Z .
0 1

A space group given in this form is said to be given in standard form.

Exercise 5.
A space group G is generated by the elements

1 0 14 −1 0 32
       
1 0 1 1 0 0
g =  0 −1 0  , h =  0 1 1  ,  0 1 0  ,  0 1 1  .
0 0 1 0 0 1 0 0 1 0 0 1

The point group P of G has 4 elements, the identity element and the linear parts of g,
h and g · h.

(i) Determine the translation subgroup of G (which is not the standard lattice),
transform G to a lattice basis of its translation lattice and write G in standard
form. (Hint: g2 and h2 are translations.)

(ii) The elements g · h and h · g have the same linear part. Check that their translation
part only differs by a lattice vector of the translation lattice.

16
3 Construction of space groups
So far we have analyzed what a space group G looks like. We have seen that G
contains a translation subgroup T as a normal subgroup and that the factor group by
this normal subgroup is (isomorphic to) the group of linear parts of the space group,
and is a finite group called the point group P . The way in which G is built from T and
P is controlled by a system of nonprimitive translations.
We will now investigate the somewhat opposite problem, how for a given translation
lattice L and a point group P acting on L, a space group G can be built that has trans-
lation subgroup T ∼ = L and point group P ∼ = G/T and what the different possibilities
are.
We will always assume that we write a space group with respect to a lattice basis of its
translation lattice, hence we have L = Zn and P ≤ GLn (Z).
Since we have seen that a space group is completely determined by its translation
subgroup T , its point group P and a SNoT, the question boils down to finding the
different possible SNoTs for a point group P ≤ GLn (Z).
One possible solution to our question always exists, namely the trivial SNoT which
has tg = 0 for all g ∈ P .

Definition 28 For a given point group P ≤ GLn (Z), the space group
  
g t n
G= | g ∈ P, t ∈ Z
0 1

with trivial SNoT is called the symmorphic space group with point group P . It is the
semidirect product of Zn and P .

3.1 Shift of origin

Before we address the question how nontrivial SNoTs can be found, we first note a
slight complication.
Example: The1-dimensional  example of the  space group
 generated by the ’glide-
−1 12 1 1
reflection’ g = and the translation is a space group in standard
0 1 0 1
 
1 0
form, since g2 = . But of course, there is nothing like a glide-reflection in
0 1
1-dimensional space, there are only two space groups, one with trivial point group and
the other with a point group of order 2 and both are symmorphic.
If we check how g acts, we see that 0 is mapped to 12 and 12 is mapped to 0, but 41
remains fixed. We therefore have a reflection in the point 14 which means that our
space group is indeed symmorphic, but that the origin is not chosen in a clever way.
What we have seen in the example above is that a shift of the origin alters the SNoT
of a space group. We can actually compute quite easily how the SNoT is changed by
a shift of the origin by a vector v. To compute how the matrices change, we have to

17
  
id v
conjugate with the matrix :
0 1
      
id −v g tg id v id −v g g · v + tg
=
0 1 0 1 0 1 0 1 0 1
   
g g · v + tg − v g tg + (g − id) · v
= = .
0 1 0 1

The translation part tg from the SNoT is thus changed by (g − id) · v.

Definition 29 A SNoT of the form {(g − id) · v | g ∈ P } for some vector v is called
an inner derivation.
The strange term ’inner derivation’ has its origin in differential geometry and is com-
monly used in cohomology theory. We only remark that a SNoT can actually be re-
garded as an element of a cohomology group.
Note: The inner derivations form a vector space, since for tg = (g − id) · v and
t′g = (g − id) · v ′ we have tg + t′g = (g − id) · (v + v ′ ).

Theorem 30 A space group with SNoT {tg | g ∈ P } is symmorphic if and only if

each tg is of the form tg = (g − id) · v for some fixed vector v, i.e. if the SNoT is an
inner derivation.
If the SNoTs {tg | g ∈ P } and {t′g | g ∈ P } of two space groups with the same point
group P differ only by an inner derivation (i.e. tg − t′g = (g − id) · v for some vector
v), then the space groups are actually the same, only written with respect to different
origins (differing by the vector v).
Example: Let P be the point group 2mm generated by
   
1 0 −1 0
g= , h= .
0 −1 0 1
     
x 0 −2x
For a vector v = we get (g − id) · v = and (h − id) · v = as
y −2y 0
inner derivations.      
a c ′ a
An arbitrary SNoT {tg = , th = } can thus be changed to {tg = , t′h =
b d 0
   
0 1 c
} by an inner derivation by choosing v = 2 .
d b
Exercise 6.
Show that an inner derivation {tg = (g − id) · v | g ∈ P } fulfills the product condition
tgh ≡ g · th + tg mod T by showing that even the equality tgh = g · th + tg holds.
The following theorem (which is not hard to prove) states that by an appropriate shift
of the origin, the coordinates of a SNoT become rational numbers with denominators
at most the order |P | of the point group. This immediately shows that there are only
finitely many different space groups for a given point group and lattice, since there are
only finitely many rational numbers 0 ≤ pq < 1 with denominator at most |P |.

18
Theorem 31 Let {tg | g ∈ P } be the SNoT of a space group. If the origin is shifted
by the vector
1 X
v= tg ,
|P |
g∈P
1
then for the SNoT {t′g | g ∈ P } with respect to the new origin one has t′g ∈ n
|P | Z , i.e.
the denominators of the coordinates of each t′g are divisors of |P |.

3.2 Frobenius congruences

We have seen that the different possible space groups built from T and P are de-
termined by the different SNoTs modulo inner derivations. The possible SNoTs are
restricted by:

(1) the product condition tgh ≡ g · th + tg mod Zn ;

(2) the translation part t of {id | t} has to be an integral vector, i.e. t ∈ Zn .

The product condition reduces the determination of the SNoT to generators of the point
group P . But even then the second restriction - although appearing fairly innocent -
amounts in a seemingly infinite task:
Problem: If an arbitrary product in the generators of P gives the identity element of
P , then the translation part of the corresponding product in the space group has to be
an integral vector. In principle these are infinitely many different products which one
would have to check.
Fortunately, the question of describing all products in the generators of a group which
result in the identity is a classical problem in group theory and actually was one of the
first problems to be addressed computationally. The idea is to use a presentation of the
point group by abstract generators and defining relators.

Definition 32 A group P = hg1 , . . . , gs i has the presentation

hx1 , . . . , xs | r1 , . . . , rt i

with abstract generators xi and defining relators rj = rj (x1 , . . . , xs ) which are prod-
ucts in the xi and their inverses x−1
i , if the following hold:

• substituting gi for xi in the relators yields the identity element of P ;

• all products of the gi giving the identity can be derived from the relators rj by
the following transformations:

– insertion or deletion of a relator in a product;

– conjugation with a generator xi or its inverse x−1
i ;
– insertion or deletion of subterms of the form xx−1 and x−1 x.

Examples:

(1) The cyclic group Cn of order n has the presentation hx | xn i.

19
 (2) The symmetry group Dn of a regular n-gon has the presentation

hx, y | xn , y 2 , (xy)2 i

where x represents a rotation of order n and y a reflection.

The first two relators allow to reduce the powers of x and y by n and 2, respec-
tively. The third relator can be read as xy = yx−1 and allows to collect all
powers of x to the left and all y to the right. The relators thus allow to reduce
every product in x and y to one of the 2n products xi y j with 0 ≤ i < n and
0 ≤ j < 2. These 2n products correspond to the 2n elements of Dn .

(3) The symmetric group S4 of all permutations of 4 symbols has the presentation

hx, y, z | x2 , y 2 , z 2 , (xy)3 , (yz)3 , (xz)2 i

where x, y, z represent the permutations (1, 2), (2, 3), (3, 4), respectively. In
this example it is slightly harder to check that the given relations are actually
sufficient.

(4) The symmetry group Oh of the cube has generators

   
0 0 1 0 −1 0
g = 1 0 0 , h = 1 0 0 .
0 −1 0 0 0 1

A presentation for this group with x and y representing g and h is given by

hx, y | x6 , y 4 , (xyx)2 , (xy −1 )2 i.

Remarks:
(1) For a finite group it is always possible to find defining relators. For small groups
this can usually be done by hand, but often it is more convenient to use standard
tools from computer algebra packages.

(2) The opposite problem, to identify a group given by a presentation is much

harder. In general, it is even impossible to decide whether a product in a group
given by abstract generators and defining relations is the identity element of the
group.
The application of group presentations to the problem of determining SNoTs is based
on the following observation.

Theorem 33 Let g1 , . . . , gs be generators of a point group P and let hx1 , . . . , xs |

r1 , . . . , rt i be a presentation
 ofP .
gi ti
Assume that gi = are augmented matrices for 1 ≤ i ≤ s such that
0 1
substituting xi by gi in the relators of P gives translations with translation vector in
Zn .
Then all products in the gi which have the identity of P as linear part have translation
parts in Zn .

20
This theorem is proved by checking that the transformations given in Definition 32 by
which the products evaluating to the identity may be manipulated do not change the
property of having a translation part in Zn .

Corollary 34 Let P be a point group with presentation as above and let gi be aug-
mented matrices such that the relators of P evaluate to translations with translation
vectors in Zn .
Then, extending the translations ti for the generators of P to all elements of P via the
product condition tgh = g · th + tg gives a SNoT for P .
We are thus reduced to the problem of choosing translation parts for the generators
of P such that evaluating the relators of P on these elements gives translations with
integral coordinates. But this means just to solve a (finite) system of linear congruences
modulo Z, which are called the Frobenius congruences.

Definition 35 Let g1 , . . . , gs be generators of a point group P and let hx1 , . . . , xs |

r1 , . . . , rt
i be a presentation
 of P .
gi ti
Let gi = be augmented matrices for 1 ≤ i ≤ s where the coordinates of
0 1
the translation vectors ti are indeterminates.
Then evaluating the relators of P in the augmented matrices gi and equating the re-
sult with 0 mod Z gives rise to a system of linear congruences which are called the
Frobenius congruences.
Every solution of the Frobenius congruences gives rise to a SNoT for P .
Since we already know that SNoTs differing only by an inner derivation represent the
same space group with respect to a different origin, in order to determine the different
space groups with point group P and translation lattice Zn , we only have to consider
representatives of the solutions of the Frobenius congruences up to inner derivations.
To whom it may concern: We are by now heavily busy with cohomology theory. The
solutions of the Frobenius congruences modulo inner derivations are nothing but the
first cohomology group H 1 (P, Rn /Zn ) which is isomorphic to the second cohomol-
ogy group H 2 (P, Zn ).
Example: We consider the point group 2mm generated by
   
1 0 −1 0
g= , h=
0 −1 0 1

which has presentation hx, y | x2 , y 2 , (xy)2 i.

Evaluating the relators on the augmented matrices
   
1 0 a −1 0 c
g =  0 −1 b  , h =  0 1 d 
0 0 1 0 0 1
gives the three matrices
     
1 0 2a 1 0 0 1 0 0
 0 1 0  ,  0 1 2d  ,  0 1 0  .
0 0 1 0 0 1 0 0 1

21
The Frobenius congruences are thus

2a ≡ 0 mod Z and 2d ≡ 0 mod Z.

We have already seen that the inner derivations for this group allow to set b = 0 and
c = 0, and it is indeed a good idea to first compute the inner derivations and eliminate
as many of the indeterminates as there are linearly independent inner derivations before
evaluating the relations.
Thus, modulo the inner derivations we have the possible solutions a ∈ {0, 21 } and
d ∈ {0, 21 } which give rise to the following four SNoTs:
 
0
(1) tg = th = : this is the symmorphic space group.
0
1  
2 0
(2) tg = , th = : the space group has a glide reflection along the x-axis
0 0
and an ordinary reflection along the y-axis.
   
0 0
(3) tg = , th = 1 : the space group has an ordinary reflection along the
0 2
x-axis and a glide reflection along the y-axis.
1  
2 0
(4) tg = , th = 1 : the space group has glide reflections along the x- and
0 2
y-axis.

Exercise 7.
Compute the inner derivations and the solutions of the Frobenius congruences modulo
the inner derivations for the following point groups P :

(1) P is generated by    
0 1 0 −1
g= , h=
1 0 −1 0
and has presentation hx, y | x2 , y 2 , (xy)2 i.

(2) P is generated by    
0 −1 1 0
g= , h=
1 0 0 −1
and has presentation hx, y | x4 , y 2 , (xy)2 i.

Example: In order to show that the concept of finding SNoTs via Frobenius congru-
ences carries over to higher dimensions, we consider a 4-dimensional example.
The symmetry group of a regular octagon is the dihedral group of order 16, which is
generated by the matrices
   
0 0 0 −1 0 0 0 1
1 0 0 0  0 0 1 0
g= 0 1 0 0  , h = 0 1 0 0
  

0 0 1 0 1 0 0 0

22
and has presentation hx, y | x8 , y 2 , (xy)2 i. Note that representing
√ the group by 2 × 2
matrices is possible, but involves irrational numbers like 2 and thus results in a non-
crystallographic group.
We first determine the inner derivations. Since g − id is an invertible matrix, letting v
run over R4 results in (g − id) · v running over all vectors of R4 . Thus the translation
part of g can be chosen as the 0-vector and only the translation part of h has to be
considered in indeterminates.
The first relator is now superfluous. Evaluating the other two relators on the matrices
   
0 0 0 −1 0 0 0 0 1 a
 1 0 0 0 0   0 0 1 0 b 
   
g=  0 1 0 0 0 , h =  0 1 0 0 c 
  
 0 0 1 0 0   1 0 0 0 d 
0 0 0 0 1 0 0 0 0 1
gives the two matrices
   
1 0 0 0 a+d 1 0 0 0 0
 0 1 0 0 b+c   0 1 0 0 a+c
  
h2 =  2
 
 0 0 1 0 b+c  , (gh) =  0 0 1 0 2b 
  
 0 0 0 1 a+d   0 0 0 1 a+c 
0 0 0 0 1 0 0 0 0 1
The Frobenius congruences are thus:
a + d ≡ 0 mod Z, b + c ≡ 0 mod Z, a + c ≡ 0 mod Z, 2b ≡ 0 mod Z
1
We either have b = 0 which implies c = 0, a = 0, d = 0 or b = 2 which implies
c = 21 , a = 12 , d = 12 .
The only nontrivial SNoT is thus given by
   
0 1
0 1 1
tg = 
0 , th = 2 1 .
  

0 1

3.3 Normalizer action

There is still one issue we have to consider in order to arrive at truly different space
groups. So far, we have regarded the point group P as the set of linear parts of the
space group G. However, these elements can be permuted by an automorphism of the
point group. Since we are dealing with space groups, we can only apply such automor-
phisms which respect that the space group has translation lattice Zn . In particular, an
automorphism has to map the standard basis of Zn to another lattice basis of Zn and
therefore must be given by conjugation with an element of GLn (Z).

Definition 36 For a point group P ≤ GLn (Z) the group

N := NGLn (Z) (P ) := {a ∈ GLn (Z) | a−1 ga ∈ P for all g ∈ P }
is called the integral normalizer of P .
It is the group of automorphisms of P which additionally map the lattice Zn to itself.

23
Remark: It can in general be a fairly difficult task to determine the integral normalizer
of a point group P . However, in low dimensions the point groups are well-known
groups and also their automorphisms can be computed easily. It then remains to check
whether an abstract automorphism is induced by conjugation with an integral matrix.
Examples:
(1) The group P = {id, −id} has N = GLn (Z) as its integral normalizer, since
±id commutes with any matrix. This shows that the integral normalizer is not
necessarily a finite group. However, since the finite group P has only finitely
many different automorphisms, there are only finitely many different conjuga-
tion actions on P . The subgroup of N which fixes P elementwise, i.e. for which
a−1 ga = g holds for all g ∈ P is called the integral centralizer of P . It is a
subgroup of finite index in the integral normalizer.
(2) The point group P generated by the matrices
   
1 0 −1 0
g= , h=
0 −1 0 1
has an integral normalizer which is generated by g, h and the additional element
 
0 1
a=
1 0
which interchanges the two basis vectors.
Note that the group P has an abstract automorphism ϕ of order 3 which cycli-
cally interchanges the elements g, h and gh. But since gh has trace −2, whereas
g and h have trace 0, an automorphism which is given by matrix conjugation
has to fix gh and can only interchange g and h, since the trace is invariant under
matrix conjugation.
(3) The full symmetry group P of the square lattice generated by the matrices
   
0 −1 −1 0
g= , h=
1 0 0 1
has an abstract automorphism which interchanges the two types of reflections
(reflections in x- and y-axis vs. diagonal reflections). This automorphism is
induced by conjugation with the matrix
 
1 1
a=
1 −1
which is an element of GLn (Q) but not of GLn (Z) and thus is not contained in
the integral normalizer of P . The integral normalizer NGL2 (Z) (P ) is thus just P
itself.

Lemma 37 Assume that a ∈ NGLn (Z) (P ) and that {g | tg } ∈ G. The action of a on

{g | tg } is given by
 −1   −1
a ga a−1 · tg
    
a 0 g tg a 0
· · = .
0 1 0 1 0 1 0 1

24
In particular, if g′ ∈ P such that g = a−1 g′ a, then conjugation by a maps {g | tg } to
{g | a−1 · tg′ }.
The element tg of a SNoT is thus changed by the action of a, namely according to

tg 7→ a−1 · taga−1 .

We have just seen that transforming a space group with an element from the integral
normalizer will in general change the SNoT. However, conjugation by a matrix cer-
tainly is an isomorphism of groups, and hence the space group which is obtained via
the action of the integral normalizer should not be regarded as a new space group.
Important note: The integral normalizer reveals an inherent ambiguity in the geo-
metric situation. In example (2) above we have seen that the integral normalizer of the
group P generated by
   
1 0 −1 0
g= , h=
0 −1 0 1
contains the transformation which interchanges the two basis vectors. This means that
after interchanging the basis vectors, the group P remains the same. But this means,
that the two basis vectors are geometrically indistinguishable. The crucial point is
that g and h are reflections in two perpendicular lines, but none of these lines can be
distinguished geometrically as belonging to the first basis vector.
Example: We have already computed that there are four SNoTs modulo inner deriva-
tions for the point group P = 2mm generated by
   
1 0 −1 0
g= , h= :
0 −1 0 1
 
0
(1) tg = th = ;
0
1  
0
(2) tg = 2 , th = ;
0 0
   
0 0
(3) tg = , th = 1 ;
0 2
1  
0
(4) tg = 2 , th = 1 .
0 2

Since the normalizer element  

0 1
a=
1 0
interchanges g and h, its action on the SNoTs can be seen immediately.
 1 a to the 
Applying SNoTs
 (1) and (4) does not change them, but for the SNoT (2) with
0
tg = 2 , th = we get
0 0
   
0 0
tg 7→ a−1 · taga−1 = a · th = , th 7→ a−1 · taha−1 = a · tg = 1
0 2

25
and this is precisely the SNoT (3).
The two SNoTs (2) and (3) are thus interchanged by the integral normalizer and give
rise to the same space group.
Discussion: It is worthwhile to discuss this example in full detail: The point group
2mm is the symmetry group of a rectangular lattice. It fixes a metric tensor of the form
 
a 0
F = , a, b > 0, a 6= b.
0 b

However, from the point group it can not be concluded whether a < b or a > b, i.e.
whether the first or the second basis vector is the short one. If we thus have a space
group with a reflection along one of the axes and a glide reflection along the other one,
we can not tell whether the glide is along the short or the long side. Thus, the two
space groups with a glide for the first and for the second basis vectors are regarded as
equivalent.
Note: The algorithm consisting of:

• finding the inner derivations;

• setting up and solving the Frobenius congruences;

• finding orbit representatives for the action of the integral normalizer modulo the
inner derivations

was described by H. Zassenhaus in 1948 and is therefore often called the Zassenhaus
algorithm.
Exercise 8.
A certain point group P (known as m3) is generated by
   
0 0 −1 −1 0 0
g = −1 0 0  , h =  0 1 0
0 −1 0 0 0 1

and has presentation hx, y | x6 , y 2 , (xy)3 , (x3 y)2 i.

Since g − id is invertible, (g − id) · v runs over all vectors in R3 , hence by a shift of
origin the translation part of g may be assumed to be 0. 
1 0 0
The integral normalizer of P contains the matrix 0 0 1 which interchanges the
0 1 0
second and third basis vector.
Determine the solutions of the Frobenius congruences for P (assuming that tg = 0)
and check which of the resulting SNoTs lie in one orbit under the integral normalizer
of P .

26
4 Space group classification
In this section we will consider various aspects under which space groups may be
grouped together. We will start with the finest notion of equivalence, which are the
space group types and will end with the coarsest, the notion of crystal family.

4.1 Space group types

By the approach via translation lattices and point groups which are glued together
to a space group via a SNoT, we can (in principle) determine all space groups up to
isomorphism, provided the possible lattices and point groups are known.
By a famous theorem of Bieberbach (1911) isomorphism of space groups is the same
as affine equivalence.

Theorem 38 Two space groups in n-dimensional space are isomorphic if and only if
they are conjugate by an affine mapping from An .
In crystallography, usually a slightly different notion of equivalence than affine equiv-
alence is used. Since crystals occur in physical space and physical space can only be
transformed by orientation preserving mappings, space groups are only regarded as
equivalent if they are conjugate by an orientation preserving affine mapping, i.e. by an
affine mapping that has linear part with positive determinant.

Definition 39 Two space groups are said to belong to the same space group type if
they are conjugate under an orientation preserving affine mapping.
Thus, although space groups generated by a fourfold right-handed screw and by a
fourfold left-handed screw are clearly isomorphic, they do not belong to the same
space group type.

Definition 40 Two space groups G and G′ are said to form an enantiomorphic pair
if they are conjugate under an affine mapping, but not under an orientation preserving
affine mapping.
If G is the group of isometries of some crystal pattern, then its enantiomorphic coun-
terpart G′ is the group of isometries of the mirror image of this crystal pattern.
The number of space group types is thus the number of isomorphism classes plus the
number of enantiomorphic pairs. For dimensions up to 6, these numbers are displayed
in Table 1.

dimension 1 2 3 4 5 6
isomorphism classes 2 17 219 4783 222018 28927922
enantiomorphic pairs 0 0 11 111 79 7052
space group types 2 17 230 4894 222097 28934974

Table 1: Number of space group types in dimensions up to 6.

27
4.2 Arithmetic classes
Starting from the space groups, it is natural to collect those space groups together
which only differ by their SNoTs. Assuming that the space groups are given in stan-
dard form, i.e. with respect to a lattice basis of their translation subgroups, this means
that two groups are regarded as equivalent if they only differ by the choice of the lattice
basis.

Definition 41 Two space groups lie in the same arithmetic class if their point groups
P and P ′ are conjugate by an integral basis transformation, i.e. if P ′ = {X −1 gX |
g ∈ P } for some X ∈ GLn (Z).
We will also say that two point groups P, P ′ ≤ GLn (Z) lie in the same arithmetic
class if they are conjugate by a matrix in GLn (Z).
Point groups in the same arithmetic class act on the same lattice and differ only by the
choice of the lattice basis.
The numbers of arithmetic classes of space groups are given in Table 2.

dimension 1 2 3 4 5 6
arithmetic classes 2 13 73 710 6079 85311

Table 2: Number of arithmetic classes in dimensions up to 6.

We have seen that the point group P of a space group G is a subgroup of the full
automorphism group Aut(L) of the translation lattice L of G. But Aut(L) is a finite
subgroup of GLn (Z), hence it is a point group itself, namely of the symmorphic space
group with point group Aut(L) and translation lattice L.
This shows that some of the arithmetic classes are distinguished, because their groups
are full automorphism groups of their lattices, while others are proper subgroups.

Definition 42 A point group P acting on a lattice L is called a Bravais group if it is

the full automorphism group of L.
The arithmetic class containing P is then called a Bravais class.
Since the groups in one Bravais class act on the same lattice, but groups from different
Bravais classes act on different lattices, the Bravais classes correspond to the different
Bravais types of lattices or lattice types for short.
There are now two obvious directions in which arithmetic classes can be merged into
larger classes. The word ’direction’ can be taken literally, if groups are considered to
be positioned in a plane, where groups of the same order are on the same horizontal
level and subgroups thus lie below their supergroups.

Vertically: Starting with a Bravais group P , we can join the arithmetic class of P with
the arithmetic classes of its subgroups. However, since P acts on a particular
lattice, we will only consider those subgroups of P which do not act on a more
general lattice, i.e. on a lattice which has a smaller Bravais group than P .
This direction of joining arithmetic classes leads to the notion of Bravais flocks.

28
Horizontally: Suppose that P is a point group acting on some lattice L. We assume as
always that P is written with respect to a lattice basis of L, thus P ≤ GLn (Z).
But P also acts on other lattices than L, obvious examples are scalings like 2L,
3L, or 12 L. The interesting cases are those lattices L′ which lie between L and
one of its scalings, these are the centerings of L.
In general, the action of P on L′ gives rise to a point group P ′ which does not lie
in the same arithmetic class as P , but is isomorphic with P ′ and it is worthwhile
to join the arithmetic classes of P and P ′ .
This direction of joining arithmetic classes leads to the notion of geometric
classes.

4.3 Bravais flocks

We have already seen that a lattice can be characterized by its metric tensor containing
the dot products of a lattice basis. If a point group P acts on a lattice L, it fixes the
metric tensor of L. However, a point group in general fixes not only a single metric
tensor (or multiples thereof), but it actually fixes all metric tensors from a vector space.

Definition 43 Let P ≤ GLn (Z) be a finite integral matrix group. Then

F(P ) := {F ∈ Rn×n | F = F tr , gtr F g = F for all g ∈ P }
is called the space of metric tensors of P .
The dimension of F(P ) is called the number of parameters for the metric tensors of
P.
If P is generated by the matrices g1 , . . . , gr , the space F(P ) of metric tensors can be
computed as the space of solutions of a system of linear equations in the entries of F ,
namely
gitr F gi − F = 0, 1 ≤ i ≤ r.

Examples:
(1) Let P = 2mm be the group generated by
   
1 0 −1 0
g= , h=
0 −1 0 1
 
a c
and let F = . Then
c b
     
tr a −c a c 0 −2c
g Fg − F = − = ,
−c b c b −2c 0
     
tr a −c a c 0 −2c
h Fh − F = − = ,
−c b c b −2c 0
hence c = 0 and a and b are arbitrary, thus the number of parameters is 2 and
  
a 0
F(P ) = | a, b ∈ R .
0 b
This space of metric tensors characterizes the rectangular lattice.

29
    
0 −1 a c
(2) Let P = 4 be the group generated by g = and let F = .
1 0 c b
Then
     
tr b −c a c b − a −2c
g Fg − F = − = ,
−c a c b −2c b − a

hence c = 0 and a = b is arbitrary, thus the number of parameters is 1 and

  
a 0
F(P ) = |a∈R .
0 a

This space of metric tensors characterizes the square lattice.

The space of metric tensors is useful to decide whether a subgroup of a Bravais group
acts on a more general lattice than the Bravais group. For example, the group 4 from
example (2) above has the same space of metric tensors as the Bravais group 4mm of
the square lattice. However, the subgroup 2 of 4 (generated by g2 ) has a space of
metric tensors of dimension 3. It acts on the oblique lattice, which is more general
than the square lattice.

Definition 44 Let P be a Bravais group. Then the Bravais flock of P consists of the
arithmetic classes of subgroups of P , which have the same space of metric tensors as
P.
The Bravais flocks collect together those arithmetic classes which genuinely act on
the same lattice. They are thus in correspondence with the lattice types and Bravais
classes, since each Bravais flock contains exactly one Bravais class.
The numbers of Bravais flocks, and thus also of Bravais classes and lattice types are
given in Table 3.

dimension 1 2 3 4 5 6
lattice types 1 5 14 64 189 841

Table 3: Number of lattice types in dimensions up to 6.

4.4 Geometric classes

Let P be a point group acting on a lattice L and written with respect to a lattice basis
of L. Assume that P also acts on a lattice L′ which is different from L and let X be
the transformation matrix from the lattice basis of L to a lattice basis of L′ . Written
with respect to that basis of L′ the action of P is given by

P ′ = {X −1 gX | g ∈ P }.

Since L 6= L′ , we have that X 6∈ GLn (Z), but clearly X ∈ GLn (R).

Definition 45 Two space groups lie in the same geometric class if their point groups
P and P ′ are conjugate by a real basis transformation, i.e. if P ′ = {X −1 gX | g ∈ P }
for some X ∈ GLn (R).

30
We will also say that two point groups P, P ′ ≤ GLn (Z) lie in the same geometric
class if they are conjugate by a matrix in GLn (R).
Point groups in the same geometric class are the actions of a matrix group on different
lattices.
Historically, the geometric classes in dimension 3 were determined much earlier than
the space groups, because they can be obtained from the face normals of crystal faces
and thus describe the morphological symmetry of macroscopic crystals.
The numbers of geometric classes of space groups are given in Table 4.

dimension 1 2 3 4 5 6
geometric classes 2 10 32 227 955 7104

Table 4: Number of geometric classes in dimensions up to 6.

Note: It is common to speak of the geometric classes as the types of point groups. This
emphasizes the point of view to regard a point group as the group of linear parts of a
space group, written with respect to an arbitrary basis of Rn (not necessarily a lattice
basis).
Starting with the space group types, we therefore get the classification into arithmetic
classes if we keep the information about the point groups and lattices and forget about
the SNoTs, and we get the classification into geometric classes if we also forget about
the lattices, thus keeping only the point group information:

space group types −→ arithmetic classes −→ geometric classes

forget SNoT forget lattice

Diagram of arithmetic classes

In Figure 3 the subgroup diagram of arithmetic classes in the hexagonal crystal family
(we will explain this term below) in dimension 3 is displayed.
This diagram illustrates the different possibilities of moving between arithmetic classes
discussed so far:

• the boxes represent the arithmetic classes;

• the thick boxes represent the Bravais classes;

• if boxes are joined by a line, the lower group is a maximal subgroup of the higher
group;

• the Bravais flock of a Bravais class consists of those boxes which can be joined
by a chain to the box of the Bravais class (note that in this diagram all groups
have spaces of metric tensors of dimension 2);

• boxes which are directly joined together lie in the same geometric class and are
thus actions of the same group on different lattices;

◦ for the sake of clearness, some boxes are slightly lowered (the boxes with sym-
bols ending on R) in order to emphasize that the action is on a different lattice.

31
 6/mmmP

3mR 31mP 3m1P 6/mP 622P 6mmP 62mP 6m2P

3R 3P 32R 312P321P 3mR 31mP3m1P 6P 6P

3R 3P

Figure 3: Arithmetic classes in the hexagonal crystal family.

In particular, we can read off that the 21 arithmetic classes fall into 12 geometric
classes and 2 Bravais flocks, the Bravais flock of Bravais class 6/mmmP contains all
arithmetic classes with symbols ending on P and contains the groups genuinely acting
on a hexagonal lattice, the Bravais flock of Bravais class 3mR contains all classes with
symbols ending on R and contains the groups acting on a rhombohedral lattice.

4.5 Lattice systems

The idea by which arithmetic classes are joined into geometric classes can analogously
be applied to Bravais classes and Bravais flocks. If two Bravais groups for different lat-
tices are conjugate by a basis transformation X ∈ GLn (R), the corresponding Bravais
flocks may be joined into a larger class.

Definition 46 Two Bravais flocks are said to belong to the same lattice system if their
Bravais classes belong to the same geometric class.
Analogously, we will say that two lattice types belong to the same lattice system if
their Bravais groups belong to the same geometric class.
On the one hand every lattice system contains a Bravais class, on the other hand all the
Bravais classes in a lattice system lie in the same geometric class, hence there are as
many lattice systems as there are geometric classes containing Bravais classes.

Definition 47 A geometric class is called a holohedry if at least one of the arithmetic

classes contained in it is a Bravais class.
Every holohedry belongs to precisely one lattice system and every lattice system con-
tains precisely one holohedry.

32
Note: In the hexagonal crystal family displayed in Figure 3 every lattice system con-
sists just of a single Bravais flock, since both holohedries contain only one Bravais
class. This is not a typical situation, usually a holohedry contains more than one Bra-
vais class the Bravais flocks of which are then joined into a lattice system.
The numbers of lattice systems are given in Table 5.

dimension 1 2 3 4 5 6
lattice systems 1 4 7 33 57 220

Table 5: Number of lattice systems in dimensions up to 6.

4.6 Crystal systems

For the geometric class of a point group P , the arithmetic classes contained in it deter-
mine on which lattices P acts. A further possibility to classify point groups therefore
is given by joining those geometric classes which act on the same set of lattices.

Definition 48 Two geometric classes belong to the same crystal system if the arith-
metic classes contained in them belong to the same set of Bravais flocks.
Example: In the hexagonal crystal family displayed in Figure 3, the dashed line sepa-
rates the two crystal systems. The geometric classes below the dashed line act both on
the hexagonal and on the rhombohedral lattice, this crystal system is called the trigonal
crystal system. The geometric classes above the dashed line only act on the hexagonal
lattice and belong to the hexagonal crystal system.
A crystal system can contain at most one holohedry, and in the example above it does
so. Indeed, all crystal systems in dimensions up to 4 contain a holohedry, but for higher
dimensions this is no longer true.
Figure 4 displays a part of the arithmetic classes in a crystal family in 5-dimensional
space. There are six Bravais classes, indicated by the bold boxes and only the geomet-
ric classes in the oval frame act on all the six different lattices, whereas the holohedries
only act on four or two of the different lattices.
The numbers of lattice systems are given in Table 6.

dimension 1 2 3 4 5 6
crystal systems 1 4 7 33 59 251

Table 6: Number of crystal systems in dimensions up to 6.

Note that in dimension 6 there are already 31 crystal systems that do not contain a
holohedry (251 crystal classes vs. 220 holohedries).

4.7 Crystal families

The coarsest classification level for space groups (and point groups) collects all arith-
metic classes together which can be reached by moving inside Bravais flocks and inside
geometric classes.

33
 G34 G56

H1 H2 H3 H4 H5 H6 H3′ H4′ H5′ H6′ H7′ H8′

U1 U2 U3 U4 U5 U6 U7 U8

M1 M2 M3 M4 M5 M6 M7 M8 M1′ M2′ M3′ M4′ M5′ M6′ M7′ M8′

Figure 4: Crystal system without a holohedry in 5-dimensional space.

Definition 49 The crystal family of a space group G is the smallest set of arithmetic
classes containing G which contains full Bravais flocks and full geometric classes.
Thus, if we graph all arithmetic classes of dimension n in the way shown in Figures 3
and 4, the crystal families are the connected components if we regard boxes joined by
lines or directly joined as being linked.
The numbers of crystal families are given in Table 7.

dimension 1 2 3 4 5 6
crystal families 1 4 6 23 32 91

Table 7: Number of crystal families in dimensions up to 6.

Up to dimension 3 it seems exceptional that a crystal family splits into different crystal
systems, since the only instance of this phenomenon is the splitting of the hexagonal
crystal family into the trigonal and the hexagonal crystal systems. However, in higher
dimensions it becomes rare that a crystal family consists of a single crystal system,
hence this is actually the exceptional case and the splitting into several crystal systems
is the rule.

Summary
We finish this section by collecting together the numbers of classes on the different
classification levels for dimensions up to 6 in Table 8.

34
 dimension 1 2 3 4 5 6
crystal families 1 4 6 23 32 91
lattice systems 1 4 7 33 57 220
crystal systems 1 4 7 33 59 251
lattice types 1 5 14 64 189 841
geometric classes 2 10 32 227 955 7104
arithmetic classes 2 13 73 710 6079 85311
space group types 2 17 230 4894 222079 28934974

Table 8: Number of classes on different classification levels in dimensions up to 6.

Some useful literature

• T. Hahn (ed.), International Tables for Crystallography, Vol. A, 5th ed., Springer,
Dordrecht, 2005.

• H. Zassenhaus, Über einen Algorithmus zur Bestimmung der Raumgruppen,

Comm. Math. Helv. 21, pp. 117-141, 1948.

• T. Janssen, Crystallographic Groups, North-Holland, Amsterdam, 1973.

• J. Opgenorth, W. Plesken, T. Schulz, Crystallographic Algorithms and Tables,

Acta Cryst. A54, pp. 517-531, 1998.

• B. Souvignier, Enantiomorphism of crystallographic groups in higher dimen-

sions with results in dimensions up to 6, Acta Cryst. A59, pp. 210-220, 2003.

35