This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.Lecture 4 The Extended Fillmore–Springer–Cnops Construction
Cycles—circles, parabolas and hyperbolas—are invariant families
under their respective Möbius transformations. We will now proceed
with a study of the invariant properties of cycles according to the
Erlangen programme. A crucial step is to treat cycles not as subsets
of ℝn but rather as points of some projective space of
higher dimensionality, see [33]*Ch. 3
[59] [276]. It is common in algebra to
represent quadratic forms by square matrices, thus circles and their
Möbius transformations are naturally described by 2× 2
matrices, see [306]*Ch. 9
[300]*Ch. 1. Yet one important component is
still missing in those works: the invariant cycle product. This
enhancement was independently introduced by several researchers
in hypercomplex context [65, 100], see
also [163]. This chapter introduces a further extension:
the signature of the cycle project can vary even for the same type of
point space.
4.1 Invariance of Cycles
K-orbits, shown in Fig. 1.2, are
K-invariant in a trivial way. Moreover, since actions of both A
and N for any σ are extremely ‘shape-preserving’, see
Exercise 12, we meet natural invariant objects of the
Möbius map:
Definition 1
The common name cycle
[339] is
used to denote circles, parabolas with horizontal directrices and
equilateral hyperbolas with vertical axes of symmetry (as well as
straight lines as the limiting cases of any from above) in the
respective EPH case.
It is known, from analytic geometry, that a cycle is defined
by the equation
k(u2−σ v2)−2lu−2nv+m=0,
(1) |
where σ = ι2 and k, l, n, m are real parameters, such that
not all of them are equal to zero. Using hypercomplex numbers, we
can write the same equation as,
cf. [339]*Suppl. C(42a),
where w= u+ι v, K=ι k, L=n+ι l,
M=ι m and conjugation is defined by w= u−ι v.
Exercise 2
Check that such cycles result in straight lines for certain k,
l, n, m and, depending from the case, one of the
following:
-
In the elliptic case: circles with centre
(l/k,n/k) and squared radius
(l2+n2−mk)/k2.
- In the parabolic case: parabolas with horizontal directrices and
focus at (l/k,
m/2n−l2/2nk+n/2k).
- In the hyperbolic case: rectangular hyperbolas with centre
(l/k,−n/k) and
vertical axes of symmetry.
Hereafter, the words parabola and
hyperbola always assume only the above described
types. Straight lines are also called flat cycles
.
All three EPH types of cycles enjoy many common properties,
sometimes even beyond those which we normally expect. For example, the
following definition is quite intelligible even when extended from the
above elliptic and hyperbolic cases to the parabolic one:
Definition 3
σ
r-Centre
of the
σ
-cycle (1) for any EPH case is the
point (
l/
k,
−σ
rn/
k)∈ℝ
σ. Notions of
e-centre, p-centre,
h-centre are used according the adopted EPH notations.Centres of straight lines are at infinity, see Subsection 8.1.
Remark 4
Here, we use a signature σ
r=−1
, 0
or 1
of
a number system which does not coincide with the signature
σ
of the space ℝ
σ. We will also need a
third signature σ
c to describe the geometry of cycles
in
Definition 11.
Exercise 5
Note that some quadruples (
k,
l,
n,
m)
may correspond through the
Equation (1) to the empty set on a certain point
space ℝ
σ. Give the full description of parameters
(
k,
l,
n,
m)
and σ
which leads to the empty set.
Hint:
Use the coordinates of the cycle’s centre and completion to full squares
to show that the empty set may appear only in the elliptic point
space. Such circles are usually called imaginary.
⋄
The usefulness of Defn. 3, even in the parabolic case, will be
justified, for example, by
- the uniformity of the description of relations between the centres of
orthogonal cycles, see the next subsection and
Fig. 6.3,
- the appearance of concentric parabolas in
Fig. 10.3(Pe:N and Ph:N).
Here is one more example of the natural appearance of concentric parabolas:
Exercise 6
Show that, in all EPH cases, the locus of points having a fixed
power
with respect to a given cycle C is a
cycle concentric with C.
This property is classical for
circles [71]*§ 2.3 and is also known for
parabolas [339]*§ 10. However, for parabolas, Yaglom
used the word ‘concentric’ in quotes, since he did not define the centres
of a parabola explicitly.
The family of all cycles from Definition 1 is invariant
under Möbius transformations (4) in all EPH
cases, which was already stated in
Theorem 2. If a cycle does not intersect
the real axis then the proof of Theorem 2
has the only gap: a demonstration that we can always transform the
cycle to a K-orbit.
Exercise 7
Let C be a cycle in ℝ
σ with
its centre on the vertical axis. Show that there is the unique
scaling w ↦
a2w which maps C into a K-orbit.
Hint:
Check that a cycle in ℝσ with its centre belonging
to the vertical axis is completely defined by the point of
its intersection with the vertical axis and its curvature at this
point. Then find the value of a for a scaling such that the
image of C will satisfy the
relation (7).
⋄
To finish the proof of Theorem 2 we need
to consider two more cases which are also related to the structure of
the group SL2(ℝ):
Exercise 7(a)
-
An arbitrary cycle may have only two, one or none common
point(s) with the real axis. This number is invariant under
Möbius transformations.
- If a cycle has two, one or none common
point(s) with the real axis and is fixed by a Möbius
transformation, then the map is generated by a matrix which is
conjugated to an element of subgroup A, N or K.
Hint:
Use the Exercise 3.
⋄
If all the above details are taking into consideration the proof of
Theorem 2 may not appear as simple as we
may initially thought. However, it still highlights important
connections between the invariance of cycles and the structure of the
group SL2(ℝ). Furthermore, we fully describe how cycles are
transformed by Möbius transformations in
Theorem 13.
4.2 Projective Spaces of Cycles
Figure 1.3 suggests that we may obtain a unified
treatment of cycles in all EPH cases by consideration of higher-dimension
spaces. The standard mathematical method is to declare objects under
investigation to be simply points of some larger space.
Example 8
In functional analysis, sequences or functions are considered as
points (vectors) of certain linear spaces. Linear operations (addition and multiplication
by a scalar) on vectors (that is, functions) are defined point-wise.
If an object is considered as a point (in a new space) all information
about its inner structure is lost, of course. Therefore, this space
should be equipped with an appropriate structure to hold information
externally which previously described the inner properties of our
objects. In other words, the inner structure of an object is now revealed
through its relations to its peers1.
Example 9
Take the linear space of continuous real-valued functions on the
interval [0,1]
and introduce the inner product
of two functions by the formula:
It allows us to define the norm
of a function and the orthogonality
of two functions. These are the building blocks of
Hilbert space
theory, which recovers much of Euclidean
geometry in terms of the spaces of functions.
In the geometry, the similar ideas are known as representational
geometry,
see [101, 99] and references
there for a discussion. We will utilise the this fundamental approach
for cycles. A generic cycle from Definition 1 is the set of points
(u,v)∈ℝσ defined for the respective values of
σ by the equation
k(u2−σ v2)−2lu−2nv+m=0.
(3) |
This equation (and the corresponding cycle) is completely determined
by a point (k, l, n, m)∈ ℝ4. However, this is not a
one-to-one correspondence. For a scaling factor λ ≠ 0,
the point (λ k, λ l, λ n, λ m) defines an
equivalent equation to (3). Thus, we prefer to
consider the projective space ℙ3, that is,
ℝ4 factorised by the equivalence relation (k, l, n,
m)∼ (λ k, λ l, λ n, λ m) for any real
λ ≠ 0. A good introductory read on projective spaces
is [296]*Ch. 10, see also Ch. 13
in [311]. The chapter in the later book is titled
“Projective Geometry Is All Geometry”, which is a quote from Arthur
Cayley and our consideration below is a further illustration to this claim.
Definition 10
We call ℙ
3 the cycle space
and refer to the initial
ℝ
σ as the point space
. The correspondence which associates a
point of the cycle space to a cycle equation (3) is
called map Q.
Exercise 10(a)
Which cycles correspond to the quadruples (1,0,0,0)
, (0,1,0,0)
, (0,0,1,0)
and (0,0,0,1)
?
Hint:
for the last cycle see Defn. 1 and the following Ex. 2.
We also note that the Equation (2) of a cycle
can be written as a quadratic form
K w1w1−Lw1w2+Lw1w2+Mw2w2=0
(4) |
in the homogeneous coordinates (w1,w2), such that
w=w1/w2. Since quadratic forms are related to square
matrices, see Section 4.4, we define another map on
the cycle space as follows.
Definition 11
We arrange numbers (
k,
l,
n,
m)
into the cycle matrix
with a new hypercomplex unit ι
c and an
additional real parameter s, which is usually equal to ±
1
. If we omit it in the cycle notation Cσc, then the
value s=1
is assumed.The values of σc:=ιc2 are −1,
0 or 1, independent of the value of σ. The
parameter s=±1 often (but not always) is equal to
σ. Matrices which differ by a real non-zero factor are
considered to be equivalent.
We denote by M such a map from
ℙ3 to the projective space of 2× 2 matrices.
The matrix (5) is the cornerstone of the
(extended) Fillmore–Springer–Cnops construction
(FSCc) [65, 100] and is closely
related to the technique recently used by A.A. Kirillov to study the
Apollonian gasket [163]. A hint for the composition of
this matrix is provided by the following exercise.
Exercise 12
Let the space A
2 be equipped with a product of
symplectic type
where w(′)=(
w1(′),
w2(′))∈A
2.
Letting C=(
)
, check that the equation of the
cycle (4) is given by the expression
[
w,
Cw]=0
. Identifications of both Q and M are straightforward. Indeed,
a point (k, l, n, m)∈ℙ3 equally well represents (as
soon as σ , σc and s are already fixed) both the
Equation (3) and the line of
matrix (5). Thus, for fixed σ ,
σc and s, one can introduce the correspondence between
quadrics and matrices shown by the horizontal arrow on the following
diagram:
ℙ3 <−>[dl]Q <−>[dr]M Quadrics on ℝσ
<−>[rr](.55)Q∘ MM2(A)
(6) |
which combines Q and M.
4.3 Covariance of FSCc
At first glance, the horizontal arrow in (6)
seems to be of little practical interest since it depends on too
many different parameters (σ, σc and s). However,
the following result demonstrates that it is compatible with simple
calculations of cycles’ images under the Möbius transformations.
Theorem 13
The image S
σcs of a cycle Cσcs
under transformation (4) in ℝ
σ
with g∈
SL2(ℝ)
is given by similarity of the matrix
(5):
In other word, FSCc (5)
intertwines
Möbius action (4) on cycles with
linear map (7). Explicitly, it means:
| ⎛
⎜
⎝ | l′+ιc s n′ | −m′ |
k′ | −l′+ιc s n′
|
| ⎞
⎟
⎠ |
| =
| |
| |
| | .
(8) |
Proof.
There are several ways to prove (7),
but for now we present a brute-force calculation, which can fortunately
be performed by a
CAS [191]. See Appendix C for information
on how to install (Section C.2) and start
(Section C.3) the software, which will be
sufficient for this proof. Further use of the software would be
helped by learning which methods are available from the library
(Section C.4) and which predefined objects exist
upon initialisation (Section C.5).
Assuming the above basics are known we proceed as follows.
First, we build a cycle passing a given point P=[u, v].
For this, a generic cycle C with parameters
(k,l,n,m) is bounded by the corresponding condition:
In [2]: C2=C.subject_to(C.passing(P))
Then, we build the conjugated cycle with a generic g=
(
)∈SL2(ℝ) and a hypercomplex unit es and
parameter s=±1:
In [3]: C3=C2.sl2_similarity(a, b, c, d, es, matrix([[1,0],[0,s]])
We also find the image of P under the Möbius
transformation with the same element of g∈SL2(ℝ) but a different
hypercomplex unit e:
In [4]: P1=clifford_moebius_map(sl2_clifford(a, b, c, d, e), P, e)
Finally, we check that the conjugated cycle C3 passes the
Möbius transform P1. A simplification based on the
determinant value 1 and s=±1 will be helpful:
In [5]: print C3.val(P1).subs([a==(1+b*c)/d,pow(s,2)==1]) \
.normal().is_zero()
Out[5]: True
Thus, we have confirmation that the theorem is true in the stated
generality. One may wish that every mathematical calculation can be
done as simply.
Remark 14
There is a bit of cheating in the above proof. In fact, the library
does not use the hypercomplex form (5) of
FSCc matrices. Instead, it operates with non-commuting Clifford
algebras, which makes it usable at any
dimension—see Section B.5
and [191, 186].
The above proof cannot satisfy everyone’s aesthetic feeling. For this
reason, an alternative route based on orthogonality of cycles [65] will be given later—see
Exercise 7.
It is worth noticing that the image S
σcs under
similarity (7) is independent of the
values of s and σc. This, in particular, follows from the
following exercise.
Exercise 15
Check that the image (
k′,
l′,
n′,
m′)
of the cycle (
k,
l,
n,
m)
under similarity with g=
(
)∈
SL2(ℝ)
is
(k′,l′,n′,m′)=
(k d2+2 l c d+m c2, k b d+l( bc+a d)+m a c, n, k b2+2 l a b+m a2).
|
This can also be presented through matrix multiplication:
| | =
| ⎛
⎜
⎜
⎜
⎝ | d2 | 2 c d | 0 | c2 |
b d | bc+a d | 0 | a c |
0 | 0 | 1 | 0 |
b2 | 2 a b | 0 | a2
|
| ⎞
⎟
⎟
⎟
⎠ |
|
| | .
|
Now we have an efficient tool for investigating the properties of some
notable cycles, which have appeared before.
Exercise 16
Use the similarity formula (7) for the
following:
-
Show that the real axis v=0 is represented by the line passing through
(0,0,1,0) and a matrix
(). For any
()∈ SL2(ℝ), we have
i.e. the real line is SL2(ℝ)-invariant.
-
Write matrices which describe cycles represented by A, N and
K-orbits shown in Figs. 1.1
and 1.2. Verify that matrices representing
these cycles are invariant under the similarity with elements of the
respective subgroups A, N and K.
- Show that cycles (1,0,n,σ), which are orbits of
isotropy groups as described in
Exercise 2, are invariant under the
respective matrix similarity for the respective values of
σ=ι2 and any real n.
-
Find the cycles which are transverse to orbits of the isotropy
subgroups, i.e. are obtained from the vertical axis by the
corresponding actions.
These easy examples also show that the software is working as expected.
4.4 Origins of FSCc
The Fillmore–Springer–Cnops construction, in its generalised form, will
play a central rôle in our subsequent investigation. Thus, it is
worth looking at its roots and its origins before we begin using
it. So far, it has appeared from the thin air, but can we intentionally
invent it? Are there further useful generalisations of FSCc? All these
are important questions and we will make an attempt to approach them
here.
As is implied by its name, FSCc was developed in stages. Moreover, it
appeared independently in a different form in the recent work of
Kirillov [163]. This indicates the naturalness and
objectivity of the construction. We are interested, for now, in the
flow of ideas rather than exact history or proper credits. For the
latter, the reader may consult the original
works [100] [64]
[65] [163] [289]*Ch. 18 as
well as references therein. Here, we treat the simplest
two-dimensional case. In higher dimensions, non-commutative Clifford
algebras are helpful with some specific adjustments.
4.4.1 Projective Coordiantes and Polynomials
An important old observation is that Möbius maps appear
from linear transformations of homogeneous (projective) coordinates,
see [269]*Ch. 1 for this in a context of invariant
theory. This leads to the FSCc in several steps:
- Take a real projective space ℙ1 as a quotient of
ℝ2 by the equivalence relation (x,y)∼ (tx,ty)
for t≠ 0. Then, any line with y≠ 0 can be represented by
(x/y,1). Thus, the invertible linear transformation
will become the Möbius transformation (3) on the
representatives (). Similarly, we can consider Möbius
actions on complex numbers w=u+i v.
- The next observation [64] is that, if we replace the
vector
(), w∈ℂ by a 2× 2 matrix
(), then the matrix multiplication with g∈SL2(ℝ)
will transform it as two separate copies of the vector
in (9). The “twisted square” of this
matrix is
Then, the linear action (9) on vectors is
equivalent to the similarity g Z g−1 for the respective
matrix Z from (10).
- Finally, one can link matrices Z
in (10) with zero-radius circles, see
Exercise 1, which are also in one-to-one
correspondences with their centres. Then, the above similarity
Z↦ g Z g−1 can be generalised to the
action (7).
Another route was used in a later book of Cnops [65]: a
predefined geometry of spheres, specifically their orthogonality, was
encoded in the respective matrices of the
type (5). Similar connections between
geometry of cycles and matrices lead Kirillov,
see [163]*§ 6.3 and the end of this section. He
arrived at an identification of disks with Hermitian matrices (which
is similar to FSCc) through the geometry of Minkowski space-time and the intertwining property of the
actions of SL2(ℂ).
There is one more derivation of FSCc based on projective coordinates.
We can observe that the homogeneous
form (4) of cycle’s
equation (2) can be written using matrices as
follows:
K w1w1−Lw1w2+Lw1 w2+Mw2w2=
| |
| |
| | =0.
(11) |
Then, the linear action (9) on vectors
(
)
corresponds to
conjugated action on 2× 2 matrices (
) by the intertwining identity:
where the respective actions of SL2(ℝ) on vectors and matrices are
In other words, we obtained the usual FSCc with the intertwining
property of the type (7). However, the
generalised form FSCc does not result from this consideration yet.
Alternatively, we can represent the
Equation (4) as:
K w1w1−Lw1w2+Lw1w2+Mw2w2=
| |
| |
| | =0.
|
Then, the intertwining relation similar to (12)
holds if matrix similarity is replaced by the matrix
congruence:
This identity provides a background to the Kirillov
correspondence between circles and matrices,
see [163]*§ 6.3
and [306]*§ 9.2. Clearly, it is essentially
equivalent to FSCc and either of them may be depending on a
convenience. It does not hint at the generalised form of FSCc either.
We will mainly work with the FSCc, occasionally stating equivalent
forms of our results for the Kirillov correspondence.
4.4.2 Co-Adjoint Representation
In the above construction, the identity (11)
requires the same imaginary unit to be used in the quadratic form (the
left-hand side) and the FSCc matrix (the right-hand side). How can we
arrive at the generalised FSCc directly without an intermediate step
of the standard FSCc with the same type of EPH geometry used in the point
cycle spaces? We will consider a route originating from the
representation theory.
Any group G acts on itself by the conjugation g: x ↦ g x g−1,
see 1. This map obviously fixes the group
identity e. For a Lie group G, the tangent space at e can
be identified with its Lie algebra g, see
Subsection 2.3.1. Then, the derived map for the
conjugation at e will be a linear map g →
g. This is an adjoint representation of a Lie group G on its Lie
algebra. This is the departure point for Kirillov’s orbit method, which is closely connected to induced representations,
see [159, 161].
Example 17
In the case of G=
SL2(ℝ)
, the group operation is the multiplication
of 2× 2
matrices. The Lie algebra g can be
identified with the set of traceless matrices, which we can write in
the form
(
)
for k, l, m∈ℝ
. Then, the
co-adjoint action of SL2(ℝ)
on sl2 is:
We can also note that the above transformation fixes matrices
(
) which are scalar multiples of the identity
matrix. Thus, we can consider the conjugated
action (13) of SL2(ℝ) on the pairs of
matrices, or intervals in the matrix space of the type
| ⎧
⎨
⎩ | | , | | ⎫
⎬
⎭ | , or, using hypercomplex numbers,
| | .
|
In other words, we obtained the generalised FSCc,
cf. (5), and the respective action of SL2(ℝ).
Furthermore, a connection of the above pairs of matrices with the cycles on
a plane can be realised through the Poincaré extension, see Ch. 14.
Remark 18
Note that scalar multiplies of the identity matrix, which are
invariant under similarity, correspond in FSCc to the real line,
which is also Möbius invariant.
4.5 Projective Cross Ratio
An important invariant of Möbius transformations in complex numbers
is the cross ratio of four distinct
points, see [26]*§ 13.4:
[z1 , z2 , z3 , z4 ] =
| (z1 − z3 )(z2 − z4 ) |
|
(z1 − z2 )(z3 − z4) |
| .
(14) |
Zero divisors make it difficult to handle a similar
fraction in dual and double numbers. Projective coordinates are again
helpful in this situation—see [52]. Define the
symplectic form, cf. [11]*§ 41, of two
vectors by
ω(z,z′)=xy′−x′y, where z=(x,y), z′=(x′,y′)
∈ ℝσ2.
(15) |
The
one-dimensional projective space
ℙ1(ℝσ) is the quotient of
ℝσ2 with respect to the equivalence
relation: z∼ z′ if their symplectic form
vanishes. We also say [52] that two points in
ℙ1(ℝσ) are
essentially distinct if there exist their representatives
z, z′∈ℝσ2 such that
ω(z,z′) is not a zero divisor.
Definition 19
[52] The projective cross ratio
of four essentially distinct points
zi∈
ℙ
1(ℝ
σ)
represented by (
xi,
yi)∈ℝ
σ2, i=1,…,4
, respectively, is defined by:
[z1 , z2 , z3 , z4 ]=
| ⎛
⎜
⎝ | (x1 y3 − x3 y1 )(x2 y4 − x4 y2 ) |
(x1 y2 − x2 y1 )(x3 y4 − x4 y3 )
|
| ⎞
⎟
⎠ |
| ∈ ℙ1(ℝσ).
(16) |
Exercise 20
Check that:
-
Let the map S: ℝσ →
ℙ1(ℝσ) be defined by z↦
(z,1). Then, it intertwines the Möbius transformations on
ℝσ with linear maps (9).
- The map S intertwines cross ratios in the
following sense:
S([z1,z2,z3,z4])=[S(z1),S(z2),S(z3),S(z4)],
(17) |
where zi∈ℝe, and the left-hand side
contains the classic cross ratio (14) while the
right-hand side the projective (16).
- The symplectic form (15) on
ℝσ2 is invariant if vectors are multiplied
by a 2× 2 real matrix with the unit determinant.
-
The projective cross ratio (16)
is invariant if points in ℝσ2 are multiplied
by a matrix from SL2(ℝ). Then, the classic
cross ratio (14) is invariant under the Möbius
transformations, cf. (17).
Exercise 21
[52] Check further properties of the projective cross ratio.
-
Find transformations of
[z1,z2,z3,z4] under all permutations of points.
- Demonstrate the cancellation formula for cross ratio:
[z1 , z2 , z3 , z4 ][z1 , z3 , z5 , z4 ] = [z1 , z2 , z5 , z4 ],
(18) |
where, in the left-hand side, values in ℝσ2 are multiplied
component-wise. Such a multiplication is commutative but not
associative on ℙ1(ℝσ).
We say that a collection of points of
ℙ1(ℝσ) is
concyclic if all their representatives in
ℝσ2 satisfy to
Equation (11) for some FSCc matrix
(
)
.
Exercise 22
[52] Show that:
-
Any collection of points in ℝσ belonging
to some cycle is mapped by S from
Exercise 1 to concyclic points in
ℙ1(ℝσ).
- For any three essentially distinct points z1, z2 and
z3∈ ℙ1(ℝσ) there is a fixed
2× 2 matrix A such that [z1,z2,z,z3]=Az for
any z∈ ℙ1(ℝσ). Moreover, the
matrix A has a determinant which is not a zero divisor.
-
Any four essentially distinct points in
ℙ1(ℝσ) are concyclic if and only if
their projective cross ratio is S(r) for some real number r.
Hint:
Let [z1,z2,z,z3] correspond through S to a real
number. We know that [z1,z2,z,z3]=Az for an invertible A,
then Āz=Az or z=Ā−1Az.
Multiply both sides of the last identity by row vector
(−ȳ,x), where z=
(). The final step is to verify that
Ā−1A has the FSCc structure,
cf. [52]*§ 4.
⋄
We have seen another way to obtain FSCc—from the projective cross ratio.
Last modified: October 28, 2024.