Volterra Filters

P.NAGASREENIVASARAO
09B21D5704
 INTRODUCTION
In this dissertation we deal with

• discrete-time, finite-order, time-invariant Volterra filters.

• We know that the unit impulse response is insufficient to fully

characterize a nonlinear system unlike linear time-invariant systems.

• The identification method might be considered as a successful extension

of the impulse response of the linear, time-invariant systems to the realm
of nonlinear systems.

• The developed method indeed includes identification using the unit

impulse response as a sub-case when the system under consideration is a
linear system.

P.NAGASREENIVASARAO 09B21D5704 2
 Volterra Filters
• The use of linear system models has been well established with successful applications.

• However, there are still a large number of problems where one has to resort to nonlinear system models.

• Linear systems are fully described by their impulse response.

• There is no such unified framework for the representation of nonlinear systems. There are various categories for
modeling nonlinear systems.

• In this dissertation we will be dealing with nonlinear polynomial system models based on the Volterra series
representation.

• Volterra filters based on the Volterra series have been an attractive nonlinear system class due to some desired
properties.

• Volterra filters bear similarities to the well-developed linear system theory.

• Volterra filters can approximate a large class of nonlinear systems with a finite number of coefficients.

• Many real world processes lend themselves to get modeled naturally by polynomial systems.

P.NAGASREENIVASARAO 09B21D5704 3
 APPLICATIONS
• ACOUSTIC ECHO CANCELLATION
• ACTIVE NOISE CONTROL
• NONLINEAR SYSTEM MODELING AND IDENTIFICATION
• SATELLITE CHANNEL EQUALIZATION
• SIGNAL COMPANDING
• AMPLIFIER SATURATION
• MULTIPLICATIVE INTERACTION BETWEEN GAUSSIAN SIGNALS
• HIGH DATA RATE TRANSMISSIONS (e.g. copper lines, satellite links)
• TELECOMMUNICATIONS, IMAGE PROCESSING, IN GEOPHYSICAL AND
BIOMEDICAL SIGNAL PROCESSING(e.g. neuronal spike detection)

P.NAGASREENIVASARAO 09B21D5704 4
 Overview of the Volterra series
Representation
• The Taylor series expansion with memory is
known as the Volterra series.
• The naming is due to Vito Volterra, the Italian
mathematician who introduced this
polynomial series.
• The LMS and RLS adaptive algorithms are
suitable for practical implementation.

P.NAGASREENIVASARAO 09B21D5704 5
For a general continuous-time nonlinear system the input-output relationship
is represented by the following infinite continuous-time Volterra series integral.

The equivalent discrete-time Volterra series sum is given as follows.

P.NAGASREENIVASARAO 09B21D5704 6
Truncated Volterra Filters
The truncated or doubly finite Volterra series is obtained by
confining the infinite summations to finite values.

The truncated Volterra series is suitable for the modeling of a

wide variety of nonlinearities encountered in real-life systems.

The truncated Volterra series is given as

P.NAGASREENIVASARAO 09B21D5704 7
P.NAGASREENIVASARAO 09B21D5704 8
Volterra Filter - Figure

Volterra Filter as
Sum of Subsystems

P.NAGASREENIVASARAO 09B21D5704 9
Volterra Filter Architecture

P.NAGASREENIVASARAO 09B21D5704 10
 Nonlinear Adaptive Filtering
A nonlinear filter

Figure : Adaptive nonlinear filter

The Vector Form of Truncated Volterra Series Expansion

 Linear combination of nonlinear functions of the input signal, the

input-output relationship can be expressed easily in a vector form

P.NAGASREENIVASARAO 09B21D5704 11
P.NAGASREENIVASARAO 09B21D5704 12
P.NAGASREENIVASARAO 09B21D5704 13
P.NAGASREENIVASARAO 09B21D5704 14
P.NAGASREENIVASARAO 09B21D5704 15
1. a new implementation of the third order RLS Volterra filter based on the
decomposition of the input vector.
2.

The nonlinear model described by the relations (1) and (2) is called a third
order Volterra model.
P.NAGASREENIVASARAO 09B21D5704 16
P.NAGASREENIVASARAO 09B21D5704 17
 (3)

(4)

P.NAGASREENIVASARAO 09B21D5704 18
The aim of this section is to discuss the efficient implementation of the RLS
adaptive algorithm on a third order Volterra filter.
For simplicity we have considered only odd order nonlinearities, up to the
third order, in the system being identified.

P.NAGASREENIVASARAO 09B21D5704 19
 The RLS algorithm updates the filter coefficients according to the following steps:
 
I. Initialization:
 
define the filter memory( length of or H(n) and X(n))
H(0) = [0 0 ....0];
CXX (0) =d ∗ I where d is a small positive constant;
 
II. Operations: for n = 1,nr.o f iterations
 
1. Create the input vector:
X (n)
 
2. Compute the error:
e (n/n−1) = d(n)−H(n−1) ∗X′(n);
 
3. Compute the scalar:
m(n) = X (n) ∗CXX (n−1) ∗X′
(n);
 
4. Compute the matrix:
G(n) = (CXX (n−1) ∗H′(n−1))/(l +m);
 
5. Updates the filter vector:
H(n) = H(n−1)+e(n/n−1) ∗G′(n);
 
6. Updates the matrix CXX :
CXX =l −1 ∗ (CXX (n−1)−G(n) ∗X (n) ∗CXX (n−1));
 
In the relations above CXX denotes the inverse autocorrelation matrix of the extended input signal. Inversion was done
according to the matrix inversion lemma

P.NAGASREENIVASARAO 09B21D5704 20