Signal & Image Processing: An International Journal (SIPIJ) Vol.15, No.

4, August 2024

A PRONY METHOD VARIANT WHICH SURPASSES

THE ADAPTIVE LMS FILTER IN THE
PRECISION OF THE OUTPUT SIGNAL’S
REPRESENTATION OF THE INPUT
Parthasarathy Srinivasan

Bee Hive Software Solutions, USA

ABSTRACT
The Prony method for approximating signals comprising sinusoidal/exponential components is known
through the pioneering work of Prony in his seminal dissertation in the year 1795. However, the Prony
method saw the light of real world application only upon the advent of the computational era, which made
feasible the extensive numerical intricacies and labor which the method demands inherently. The Adaptive
LMS Filter which has been the most pervasive method for signal filtration and approximation since its
inception in 1965 does not provide a consistently assured level of highly precise results as the extended
experiment in this work proves. As a remedy this study improvises upon the Prony method by observing
that a better (more precise) computational approximation can be obtained under the premise that
adjustment can be made for computational error , in the autoregressive model setup in the initial step of
the Prony computation itself. This adjustment is in proportion to the deviation of the coefficients in the
same autoregressive model. The results obtained by this improvisation live up to the expectations of
obtaining consistency and higher value in the precision of the output (recovered signal) approximations as
shown in this current work and as compared with the results obtained using the Adaptive LMS Filter.

KEYWORDS
Prony Method, Fourier Series, Auto Regression, imprecision, Adaptive LMS Filter.

1. INTRODUCTION
The Prony method is an effective observation on the transformation of an exponential expression
to a mathematically convenient and tractable polynomial form.

The cornerstone of the method is to setup the autoregressive model where the subsequent values
of the input signal (are assumed to) depend regressively on the prior values and then solve the
model (by dividing the Toeplitz matrix with the input signal and obtaining the remainder) to
obtain the autoregression coefficients(a).

The characteristic polynomial summing the autoregression coefficients (a) is then solved to
obtain the roots. The roots obtained directly provide the value of the damping factor and
frequency of the output approximating signal. The roots obtained are then also, used (in the form
of knowns) to again setup and solve the original autoregressive model to obtain the Amplitude,
and phase, of the approximating (or smoothened for lack of a better term) output signal. As
obvious the procedure described above is mathematically (computationally) intensive and precise
only up to the precision allowed for in the computational process.

DOI: 10.5121/sipij.2024.15401 1
 Signal & Image Processing: An International Journal (SIPIJ) Vol.15, No.4, August 2024
Thus, the Prony method was practicable only with modern computers came to aide. This made
the computing task easier however introduced a limit (albeit smaller than manual computation)
on the precision to which the resultant (approximating) signal resembled the original input signal.

In this current work, the author makes the empirical premise that the computer introduces an
imprecision proportional to the variance of the signals’ autoregression coefficients and therefore
can be thus corrected to yield a more precise output estimate of the Amplitude, frequency,
dampener, and phase.

The Adaptive LMS filter method can be described in simple terms as a technique in which the
filter coefficients(weights) are adjusted to converge to (theoretical) optimal values where the
filter approximates to the desired characteristics. The value of the weight adjustment is obtained
by computing the difference between the desired output and the Adaptive Filter output at each
iteration and taking a Mean Square of the difference as the quantum of weight adjustment in
multiplicative conjunction with the constant convergence multiplier, whose value is chosen
somewhat arbitrarily following a suitable procedure described later in this work. Experimentation
with this setup of the Adaptive LMS Filter and evaluation of the results using the Precision
Measure (again defined later in this work) , show unequivocally that the results obtained for
precision do not provide any guarantee or assurance on the lower bound for precise signal
representation of the output signal viz. a viz. the input signal. This is in sharp contrast to the
adaptation of the Prony Method found as part of this work, which does provide that assured lower
bound on the precision making this a superior alternative candidate.

2. CONTEXT OF CURRENT WORK IN REGARD TO RECENT LITERATURE OF

RELATED WORKS
In the recent (century) time frame (~2014,2022) there have been various attempts to comprehend,
adapt, implement, and optimize the Prony method with modern electronic and computing
platforms doing the heavy duty tasks. These adaptations take the form of varying the parameters
in the core Prony Algorithm (with optimization in perspective)[7], generating usable intermediary
output byproduct from the core algorithm (for use in related settings and contexts)[5], evaluating
the utility of the transformative power (in the co-domain) of the algorithm for ease of scientific
analysis[6] etc. Also, as described above the Adaptive LMS Filter has been employed albeit there
is a lacuna as proven experimentally by this work. However, the objective of this present work
is more subtle and profound, in that the adaptation proposed here-in has wide-reaching
consequences (in terms of results) in critical applications across diverse domains and application
functional areas.

3. SCOPE AND SIGNIFICANCE OF WORK

This work is focused on obtaining reliable bounds in the computational precision of the output
from the Prony method. The bounds so obtained are significant in the practical application of the
method owing to the requirement that, for the diverse domains in which the method is applied,
such as biomedical engineering, power systems etc. the level of precision significantly impacts
the ends for which the method is used , for example the monetary cost effectiveness in the case of
power systems and the effectiveness of diagnosis in the case of bio-medical engineering. The
work is unique in that though works with similar objectives exist , the particular premise and
methodology described are not found in existing literature and this work surpasses the existing
methods , including the Adaptive LMS Filter method.

2
 Signal & Image Processing: An International Journal (SIPIJ) Vol.15, No.4, August 2024

4. FUNDAMENTAL PREMISE IN THIS ADAPTATION OF PRONY’S METHOD

The implementation of the Prony’s method in MATLAB is described in [2]. As mentioned above,
in this adaptation of the Prony method, the coefficients of the characteristic polynomial,
computed in the first step autoregressive model are self-adjusted by a constant multiple of the
standard deviation of their own values and this assuredly compensates for the imprecision in the
computation due to the computing device’s limitation in capacity to perform the computation to
an arbitrary precision.

5. METHODOLOGY EMPLOYED IN THIS ADAPTATION OF PRONY’S METHOD

Step 1 :

This step consists of setting up an Auto Regressive Model with the random sample observations,
where the assumption is that the observations are naturally organized such that the nth
observation is linearly dependent on the n-1 pre-ceding observations thus yielding a model which
regressively relies on computing the Toeplitz matrix set up to relate those observations in
accordance with the above assumption.

Step 2 :

Solving the above auto-regressive model yields coefficients of the Prony characteristic
polynomial which is pre-requisite to obtain the Amplitude and Frequency of the final
approximating signal output as yielded by the polynomial roots.

Step 3 : (Nuanced Prony method proposed by this work)

Adjust the coefficients of the characteristic polynomial by a standard deviation of their own
values. This work proves that this adjustment helps overcome the computational imprecision
introduced due to limitations of the computing devices’ arbitrary precision arithmetic.

Step 4 : Work with the adjusted polynomial coefficients to obtain the Amplitude and Frequency
of the output signal components and also compute the damping factor and phase by yet another
derivative system of equations obtained by substitution of obtained roots in the original equation.

Step 5 : Compute the Precision Metric defined and observe that the nuance proposed here-in
lives up to the expectation of yielding constant lower bound in the precision of the output signal
approximation.

6. THE COMPARATIVE, ADAPTIVE LMS EXPERIMENT

In order to compare the Adaptation of the Prony Method found in this work, the following
Adaptive LMS filter was setup in MATLAB(SIMULINK) (based upon the standard example
provided by the Simulink software MATLAB/SIMULINKexample:
dsp\ViewTheCoefficientsOfYourAdaptiveFilterExample) :

3
 Signal & Image Processing: An International Journal (SIPIJ) Vol.15, No.4, August 2024

Fig 6(a) : Shows the Model setup for the Adaptive LMS Filter (Refer Introduction Section [1])

Fig 6(b) : Shows the Input Sine wave “Summed” with the Gaussian Random Noise (The Input Signal)

Fig. 6 (c) : Shows the Output Signal from the Adaptive LMS Filter (The Output Representative of the Input
Signal)

4
 Signal & Image Processing: An International Journal (SIPIJ) Vol.15, No.4, August 2024

7. FOLLOWING INITIALIZATION PARAMETERS WERE USED FOR THE

ADAPTIVE LMS FILTER SETUP

Fig. 7 : Initialization of Adaptive LMS.

8. PRECISION MEASURE OBTAINED FOR THE ADAPTIVE LMS

Figure 8

5
 Signal & Image Processing: An International Journal (SIPIJ) Vol.15, No.4, August 2024

9. ADVANTAGES OF THE PRONY ADAPTATION FOUND HERE COMPARED TO

THE EXISTINGADAPTIVE LMS METHOD

1) The Adaptive LMS Filter is Heavily dependent on the hugely arbitrary choice of convergence
coefficient (mu) for signals with low and high pass characteristics which are increasingly
constituting a large chunk of signals now appearing in several practical applications.

2) Even with a proper choice of (mu) the precision of signal recovery(output viz. a viz. input) has
been found to vary hugely without any assurance/guarantee on the recovery precision with the
Adaptive LMS Filter method.

3) The New Prony Adaptation Approach found here completely eliminates any need for arbitrary
choices for parameters for (weight) adjustments , since the coefficients are now adjusted by their
own (known) standard deviation with a known multiplier.

4) A guaranteed assurance on the lower bound of the precision is obtained with the Adapted
Prony method, as proven by experimentation as against the Adaptive LMS method which does
not yield that precision guarantee across varying parameters, as proven by experiment.

5) Also, the Prony Method Adaptation is computationally lightweight with a one step adjustment
to the coefficients as opposed to the Adaptive LMS Filter which requires coefficient computation
and adjustment at each iteration.

10. DETAILS OF HARDWARE AND SOFTWARE USED IN THE EXPERIMENTAL

EMULATION
Hardware Operating System Software Library
PC 64 bit (AMD64) Windows 10 MATLAB R2024a

11. DEFINITION OF METRIC TO QUANTIFY THE PRECISION OF THE

RECOVERED OUTPUT SIGNAL
To quantify the improvement in precision with this adaptation to the Prony method the following
formulation (similar to [1]) is used as a measure of the comparative similarity in the input and
output signals in both the unchanged Prony method and the adaptation introduced in this work.

Precision Measure,

PM = N – (norm(grecons[i]-g[i])/norm(grecons[i]-mean(g[i]))),

where N=size(g[i]), g[i] = input signal vector, grecons[i] = reconstructed output signal vector,
norm is the standard 2 norm, and mean is the usual mean definition.

6
 Signal & Image Processing: An International Journal (SIPIJ) Vol.15, No.4, August 2024

12. EXPERIMENTAL RESULTS OF THE PRECISION MEASURE FOR THIS

PRONY METHOD ADAPTATION
The results are divided into 3 distinct tiers based on the size of the input signal set vector, to the
computation, as follows :

Input Vector Coefficient Adjustment Factor Precision Measure Value

Size (consistently for this adaptation of
Prony method) (constant across
several program runs)
100 1*sigma(coefficients) 90
1000 10*sigma(coefficients) 968.3772
10000 100*sigma(coefficients) 9900

The intuitive rigor behind the empirical constant results obtained above by experiment seems to
be the fact that the amount memory required for the computer’s mathematical operations in terms
of precision grows linearly with the variance of the values being computed itself. This is likely an
aggregate manifestation of the optimizations of the memory architecture implementation.

13. EXPERIMENTAL RESULTS OF THE PRECISION MEASURE FOR THE LMS

FILTER
Input Vector Size Precision Measure Value
(consistently for this adaptation of Prony
method) (constant across several program runs)
225 174.79
202 193.35

14. CONCLUSION AND FUTURE WORK

This work has proven that with the adaptation of the Prony method set forth by the Author there
is a consistent improvement of the Lower Bound in the recovered signal precision, thus
improvising the Prony Method for advanced use in diverse and sensitive fields such as Bio-
Medical engineering and Electrical Power Systems.

Also , by experimentation described above the Prony Method Adaptation has been proved to
surpass the existing Adaptive LMS filter method in the consistent Bound obtained for the
Precision of the Output Signal with respect to the Input.

The Bio-Medical applications include improvement to disease detection by signal analysis as in

[1].

There is scope for rigorous Stability Analysis (of recovered signal) with the adjustment founded
by this work in place.

Also, there is scope to apply this work in the realm of Quantum Computing to evaluate whether
the same empirical results are upheld in that realm as well.

7
 Signal & Image Processing: An International Journal (SIPIJ) Vol.15, No.4, August 2024
REFERENCES

[1] Fernández Rodríguez et al. BMC Bioinformatics (2018) 19:451 https://doi.org/10.1186/s12859-018-

2473-y
[2] 80 IEEE Transactions on Power Systems, Vol. 5, No. I, February 1990 INITIAL RESULTS IN
PRONY ANALYSIS OF POWER SYSTEM RESPONSE SIGNALS J.F. Hauer Senior Member C•
J. Demeure Member L.L. Scharf Fellow
[3] Parameter Estimation in Electrical Power Systems Using Prony's Method Claudius Ortbandt et al
2015 J. Phys.: Conf. Ser. 659 012013
[4] https://amses-journal.springeropen.com/fourier-based- computational-approaches-pdes
[5] Research on Fourier and Prony analysis algorithm of inter-harmonic in power system
Hong Lei https://www.sciencedirect.com/science/article/pii/S2352484722004528
[6] Online identification and suppression of low frequency oscillation in power system based on
WAMS(Conference Paper) https://www.scopus.com/record/display.uri?eid=2-s2.0-
84897735036&origin=inward&txGid=c7699b92fff1ead1b2030efc1fc825f7
[7] A Fault Signal Processing Method Based on An Improved Prony Algorithm
https://www.frontiersin.org/journals/energy-research/articles/10.3389/fenrg.2021.831347/full
[8] MATLAB/SIMULINKexample: dsp\ViewTheCoefficientsOfYourAdaptiveFilterExample
[9] A computationally empirical adaptation of the Prony method which assures consistent higher
precision and stability in the reconstructed output signal components Parthasarathy Srinivasan DOI :
https://doi.org/10.5121/csit.2024.141418

Appendix A - MATLAB Code of modified Prony Method

%Code 1 Polynomial Implementation
function [Amp,alfa,freq,theta]=polynomial_method (x,p,Ts,method)
% method: 'classic', 'Is' or 'tls' (case insensitive)
% define the solving methods
CLASSIC = 0;
LS = 1;
TLS = 2;

N = length(x);

if strcmpi(method,'classic')
if N ~= 2*P
disp ('ERROR: length of x must be 2*p samples in classical method.');
Amp = [];
alfa = [];
freq = [];
theta = [];
return;
end
else solve_method = CLASSIC;
end

%% step 1

T = toeplitz(x(p:N-1) ,x(p:-1:1));

switch solve_method
case {CLASSIC, LS}
a = -T\x(p+1:N);

case TLS
a= tls(T,-x(p+1:N));

8
 Signal & Image Processing: An International Journal (SIPIJ) Vol.15, No.4, August 2024
end

%% step 2
c = transpose([1; a]);
r = roots(c);
rprim = r - (std(r)/n);
%rprim = r - std(r); %TBTD

alfa = log(abs(r))/(2*pi*Ts);

% In case alfa equals to +/-Inf the signal will not be recovered for n=0
%(Inf*0 = Nan). Making alfa = +/-realmax that indeterminance will be solved
alfa(isinf(alfa))=realmax*sign(alfa(isinf(alfa)));

%% step3
switch solve_method
case CLASSIC
len_vandermonde = p; % exact case (N=2p) find h with p samples
case LS
len_vandermonde = N; % overdetermined case (N>2p) find h with N samples
case TLS
len_vandermonde =N; % overdetermined case (N>2p) find h with N samples
end

Z = zeros(len_vandermonde,p);
for i=1:lenght(r)

z(:,i) = transpose(r(i).^(0:len_vandermonde-1));
end

rZ = real(Z);
iZ = imag (Z);
% here Inf values are substituted by realmax values
rZ(isinf(rZ))=realmax*sign(rZ(isinf(rZ)));
iZ(isinf(iZ))=realmax*sign(iZ(isinf(iZ)));

z = rZ+1i*iZ

switch solve_method
case {CLASSIC,LS}
h = Z\x(1:len_vandermonde);
case TLS
% if exists nan values SVD won't work
indeterminate_form = sum(sum(isnan(Z) | isinf(Z)));
if (indeterminate_form)
Amp = []; alfa = []; freq = []; theta = [];
return;
else
h = tls(Z,x(1:len_vandermonde))
end
end
Amp = abs(h);
theta = atan2(imag(h),real(h));