Use of Mobius: THE Transformations IN Neural Networks and Signal Process1 N G
Use of Mobius: THE Transformations IN Neural Networks and Signal Process1 N G
IN N E U R A L N E T W O R K S A N D SIGNAL
PROCESS1N G
Danilo P. Mandic
School of Information Systems
University of East Anglia, Norwich, NR4 7TJ, UK
Phone: +44 1603 592 569
Fax: +44 1603 593 345
E-mail: d.mandic@uea.ac.uk
Web: sys.uea.ac.uk
INTRODUCTION
Mobius and related mappings have been used in signal processing [6], linear
circuits [18], microwave circuits [7], and neural filters [20, 5, 121. Hence, there
is a need for a unified approach that would connect some inherent properties
of neural networks, such as fixed points, invertibility and stability, digital
filters, and modular groups of composition of Mobius transformations. Let
us start with the definition of a Mobius transformation.
Definition 1 (Mobius mapping) Let U , b,c, and d denote complex con-
stants and ad # bc. The function
uz + b
w =f(2)= -
cz + d
is called a Mobius transformation, bilinear transformation, or linear frac-
tional transformation.
transformation w = S ( z ) = e
and its inverse is also a Mobius transformation. For example, the Mobius
maps the unit disk IzI < 1 one-to-one and
onto the upper half plane S{w} > 0, where S{.}is the imaginary part of a
complex number.
In this paper, we first identify the activation function of a neuron and a
single-pole all-pass digital filter section as Mobius transformations. Then,
in order to consider a general neural network and a general cascaded filter in
the same framework, the theory of modular groups and Blaschke products is
introduced. Some inherent properties of neural networks, such as fixed points
and invertibility, and group delay properties of cascaded all-pass filters, are
shown t o be the consequence of their Mobius representations. Finally, sta-
bility issues in some N N architectures are addressed via Routh stability and
Mobius mappings.
Two typical choices for activation functions employed in neural networks are
the logistic function [9] given by
(C)IEEE
0-7803-6278-O/OO$lO.OO 186
Proposition 1 The sigmoidal transformation f ( z ) performed by a neuron
i n a n NN o n a complex input signal z = a + j , B is a Mobius transformation.
@)(U)= -w - 2arctan
T sin (w- 0)
I - cos - e) 1 (7)
The group delay .(U) of (6) is the negative derivative of the phase a, and is
given by
1- r 2
=
1+ r2 - 2 r c o s ( w - 0)
.(W)
They are also called invariant points, or fixed elements. The Brower fixed
point theorem shows that a continuous function on an interval has a fixed
(C) I E E
0-7803-6278-0/00$10.00 187
point [23]. This means that, for a single-dimensional case, for a continuous
function g(z) E ( a ,b ) , Vx E [a,b], g has a fixed point in [a,b]. As a first
order all-pass transfer function is a Mobius transformation, its fixed points
are given by
z =f(z) =
-p*z +1
z-P
+
and can be found at z1,2 = S { p } f J ( ~ { P } ) ~1, wherg S { p } denotes the
imaginary part of p .
For first-order sections, p is always real, and z 1 , 2 = f l ,which correspond t o
w = 0, and w = n in frequency. Equation (8) shows that the fixed points of
(6) correspond to the maximum and minimum of the group delay at w = 0
and w = n.
Modular Groups
'This property is often used in the bilinear transformation for synthesis of IIR digital
filters. Namely, although the mapping from the s into z domain is defined as s = $6,
the constant is often skipped [21].
this context as
s.
by f - ' ( z ) = The Mobius transformation can be now represented in
az + b
AZ = -
cz + d
Modular group is generated by two transformations, T z = z + 1 and S z =
-5 [2], as shown in the following theorem.
Theorem 1 The modular group r is generated by the two matrices
(15)
and belongs t o r.
The Mobius mappings performed by H I and H Z are re-
spectively described by matrices A41 = bi
dl [: ]
and A 4 2 = [ : 2 1,
Recall that a point x* which is mapped onto itself under a map G so that
x* = G(z*) is called a fixed point. Fixed points are important in the analysis
of stability criteria, neural networks [4, lo], and learning algorithms [13, 111.
Therefore to obtain fixed points for a general NN (which is a composition of
Mobius transformations) we look at fixed points of the nonlinear activation
function of a neuron, which is a Mobius transformation itself.
Observation 1 Fixed points of a neural network are determined by fixed
points of the employed activation function.
Obviously, the existence for fixed points of an activation function are guar-
anteed by the underlying Mobius transformation (one or two fixed points).
As the global input-output relationship in an NN can be considered in the
framework of Mobius transformations, it has one or two fixed points.
Invertibility in NNs
The problem of inverting trained NNs is t o find the inputs which yield a given
output. This problem is an ill-posed problem because the mapping from the
output space t o the input space is a one-to-many mapping. The existence
and non-uniqueness conditions of an inverse of a neural network are provided
by the theory of modular groups of underlying Mobius transformations.
Existence: The input-output relationship between two neurons within an N N
is a Mobius transformation, whose inverse f - l ( z ) =
transformat ion.
s, is also a Mobius
(C) IEEE
0-7803-6278-O/OO$lO.OO 190
Routh stability criterion is given by
where C ( T ( z ) )is the system response transform, R ( T ( z ) )is the input trans-
form, W ( T ( z ) ) / Q ( T ( zis) )a transformation of the transfer function. Fol-
lowing the approach from [20], we introduce the Mobius transformation as a
DG - a(e~~*)+b
transformation T ( z ) in (19), T(e- ) - c ( e - D e ) + d l which can then be ap-
plied to (19) to give the Routh stability criterion based on the Hamiltonian
system of equations for a nonlinear neural field filter.
Thus, for input-to-output and output-to-input transfer functions of a neural
field filter, denoted respectively by H I and H2, we have the stability criterion
given by
which gives the stability defined by the extended nonlinear Routh stability
criterion which uses Mobius transformation operators instead of Laplace and
Fourier operators. These filters can be used as basic building blocks for
speech and vision recognition and image processing [l].
a* a--z
[ ( a , z )= -~
la1 1 - a*z
where [ ( a ,z ) = z if a = 0. Let {o,},~N be a sequence and let
is called the n-th finite Blaschke product with respect t o the sequence { a n } n E ~ ,
where B o ( z )= 1. (We assume that a0 = 0).
A normalised Blaschke term is given by
j=1
H(z)=
-p1z +
1 -pzz+ 1
(26)
z-Pl z-p2
with poles p 1 and pz, under the above conditions, such that the modules
represent Mobius transformations, H ( z ) represents a Blaschke product. The
matrices associated with H I and Hz are respectively M I =
[ -? ]
and M2 =
[ -”1 -P2 1-
The product HlH2 has the same form of critical points as H1( z ) and H 2 ( z ) .
This complies with theory of iterated Mobius transformations and their at-
tractors [3, 141.
Proposition 3 An n-th order cascaded all-pass filter consisting of first or-
der real-pole sections has the maxima of the group delay at w = 0 and w = T .
CONCLUSIONS
; 6 ~ ’
s
4-
2-
0-
05 1 15 2 25 3
Figure 1: Group delay for a cascaded all-pass structure with 1600 positive and 400
negative uniformly distributed real poles
networks, and cascaded all-pass filters has been provided. It has been shown
that using this framework, some inherent properties of cascaded systems and
layered (nested) neural networks can be derived as a consequence of belonging
t o group r. The Blaschke product of Mobius transformations has been further
introduced and shown t o represent cascaded digital filters. These results
apply for both simple elements and complex functions and help to unify the
theory of signal processing and neural networks.
REFERENCES