Proc.

X National Conference with International Participation Conference "Electronica 2019", May 16 - 17, 2019, Sofia, Bulgaria

Efficient FIR Filter Synthesis Using Sigmoidal

Function
Peter Stoyanov Apostolov* Alexey Kostadinov Stefanov* and
Mariya Simeonova Bagasheva*
*South-West University, Department of Communication and Computer Engineering, Blagoevgrad, 2700, Ivan Mihaylov str. 66, E-mail:
[email protected]
*South-West University, Department of Communication and Computer Engineering, Blagoevgrad, 2700, Ivan Mihaylov str. 66, E-mail:
[email protected]
*South-West University, Faculty of Philology, Blagoevgrad, 2700, Ivan Mihaylov str. 66, E-mail: [email protected]

cases the oscillations increase near to the function transition

Abstract – The graph of the sigmoidal function represents a
transfer function of a low pass filter. Such function is the band. This is due to the Gibbs’ phenomenon [1, 2].
Complementary integral Gaussian error function - erfc. In Filter design using L1 and L2 norms has been studied in
this paper, FIR filter synthesis using erfc is examined. details in many publications [3 – 10]. Various methods for
Analytical expressions of the filter parameters are proposed. reduction of Gibbs’ phenomenon have been proposed: with
A comparison of filter selectivity and computational costs with window functions, "don’t care” transition band, change the
Parks & McClellan FIR filter, and analysis of the results is
transition band, etc. Some methods use iterative
made. The conclusion that the proposed FIR filter has better
selectivity and faster computational time has been arrived at.
optimization algorithms to derive the polynomial
coefficients, which leads to greater computational
Keywords – approximation; computational cost; FIR filter; complexity.
sigmoidal function; transfer function.
The L (Chebyshev’s) norm is the most suitable one for
I. INTRODUCTION the filter design problem. The design is performed with the
The Finite Impulse Response (FIR) filters are basic well-known Parks-McClellan’s (P&M) method [11, 12].
technical tools in signal processing. Their main advantages This is an equiripple approximation. The design uses
are the linear phase response and the absolute stability. The Remez’ exchange algorithm [13]. With this method, the
FIR filter synthesis is based on approximation of ideal low transition band can be accurately defined and arbitrarily
pass filter (LPF) response. The aim is to obtain a low narrow. However, this increases the approximation error
power trigonometric polynomial, which approximates this and the amplitude of oscillations (Fig.1). It has been found
ideal function with minimal error. Most of the FIR filter out that in equal conditions the Parks-McClellan’s method
design methods use approximations in L1, L2 and L norms. produces trigonometric polynomial of the lowest degree.
That means that the filters will be implemented with the
least number of coefficients.
The filter selectivity depends on four parameters: pass-
band ripple, stop band attenuation, transition bandwidth,
and filter length. A filter has a better selectivity when the
three parameters are equal, while the fourth one has to have
Magnitude responses

smaller pass-band ripple, greater stop band attenuation,

narrower transition bandwidth, and smaller filter length.
When all the four parameters are equal, the digital filter
efficiency depends on the shorter computational time.
The goal in the design is to obtain a rectangular shape of
the ideal LPF response, i.e. maximally flat pass band and
stop band, and the narrowest possible transition band.
An approximation with sigmoidal function for two-
dimensional FIR filter design is proposed in [14]. In the
present paper, the same approximation for one-dimensional
FIR filter design is applied. As a result, FIR filters with
Fig. 1. Approximations of ideal LPF response in L1, L2 and L
norms better selectivity and fast computational time have been
obtained.
Figure 1 shows a comparison of 32-degree
approximating polynomials using L1, L2 and L norms. As II. THEORETICAL BASIS
can be seen the FIR filters design requires a suitable trade- The sigmoidal functions possess S-shaped graphs. Such
off between the flatness and the transition bandwidth. In all functions are arcos tangent, sigmoid, tangent hyperbolic,
the criteria, the functions have the oscillations in the pass integral Gaussian error function – erf (.), etc.
band and the stop band. These oscillations determine the
approximation error and they are undesirable. In L1 and L2
978-1-7281-3622-6/19/$31.00 ©2019 IEEE
 H(f)
H

Fig. 2. Magnitude response of LPF Fig. 3. The improvement of the selectivity of magnitude responses
depending on parameter 
After some modifications, their graphics become a shape
of magnitude response of a LPF – Figure 2. The graph III. COMPARISON WITH EQUIRIPPLE APPROXIMATION
shows fpass - pass-band cut-off frequency, fstop - stop band The polynomial approximation methods use
cut-off frequency, and ft - middle frequency of the trigonometric functions: sinc or cosine. They are periodic
transition band. The difference in the cut-off frequencies functions and the approximations possess their
defines the transition bandwidth: f = fstop – fpass. Parameter characteristics – periodic local extremes (Fig. 1). Unlike
 determines the attenuations in the stop band DS (dB) and them, the proposed method uses a function with smooth
the pass-band DA (dB). The following dependencies are graph without local extremes in the pass band and stop
valid: band. Parameter  rapidly increases the function gradient in
the transition band f. Furthermore, the function decreases
DS = 20 lg ( δ ) , δ ∈ ( 0, 0.5 ) ; (1) monotonically. When  = , then f = 0 (see eq. 7) and the
graph of the function coincides with the rectangular shape
DA = 20 lg (1 − δ ) ; (2) of the ideal LPF response - Fig.3. In this sense, the offered
sigmoidal function (6) represents simple and efficient
δ = 10
DS 20
; (3) approximation. This approximation has better properties
than equiripple - Parks & McClellan. Figure 4 depicts a
H ( f pass ) = 1 − δ ; (4) comparison between them. In most parts of the pass band
and the stop band, the approximation error is practically
H ( f stop ) = δ . (5) zero. Evidently, the proposed approximation is closer to the
ideal LPF response.
In [15], it is shown that the greatest slope of the graph
possesses the integral Gaussian error function. In this way
of reasoning, we offer the sigmoidal function

H ( f ) = 0.5 erfc ( 2 β ( f − ft ) ) ; f ∈ [ 0,1] , (6)

where f is normalized frequency and erfc(.) is a

complementary integral Gaussian error function. Parameter
H(f)

 > 0 changes the function slope gradient, as shown in Fig.

3. This is equivalent to the improvement of the filter
selectivity.
Parameter  is obtained using equations (4), (5) and erfc–
1
(– x) = – erfc– 1(x)

erfc
−1
( 2δ )
β= , (7)
Δf
Fig. 4. Comparison of magnitude responses: f = 0.2;  = 0.01
where erfc–1(.) is the inverse complementary integral
Gaussian error function. Furthermore, the proposed approximation possesses a
steeper slope in the transition band with the associated
The FIR filter impulse response is determined from the greater attenuation in the interval from ft = 0.5 to fstop.
uniformly sampled function graph using shifted Inverse Fast
Fourier Transform (IFFT). This technique is also called
“Frequency sampling method”.
 Derivatives stop band.

Magnitude, V
Fig. 5. Comparison of the derivatives of the functions: f = 0.2;
 = 0.01
For a clearer distinction between the proposed Fig. 7. Filtration of test signal with proposed filter; f = 30Hz; DS
function (6) and Parks&McClellan, Figure 5  –40dB, filter length 512
depicts a comparison between their derivatives.
The derivative of (6) for the inflexed point ft =
0.5 is greater in absolute value. As a result, the
proposed function has a better selectivity.
IV. DESIGN EXAMPLE
Magnitude, V

All examples are implemented in MATLAB®, and all

designs are done on a Lenovo – Laptop, Intel® Core™ i7-
7500U CPU @ 2.70GHz RAM 12 GB.
A design example of LPF described as: ft = 2000 Hz; f
= 30 Hz; fsampl = 8000 Hz,   0.01 (DS  –40dB) and filter
length 256 will be demonstrated.
From (7)  = 219.3302 is calculated. From (6) 256
values of the magnitude response are obtained. The
impulse response is derived using shifted Inverse Fast
Fourier Transform (IFFT). Figure 6 shows the filter
magnitude and phase responses. Fig.8. Filtration of test signal with P&M filter; f = 30Hz; DS  –
40dB, filter length 512
To examine the compared filters calculation costs,
Matlab function “Run and Time” is applied.
Figures 9 and 10 show the calculation times of the
compared filters. There are two important times.
The Self Time represents the filtration duration time. It
is seen that the compared filters have equal Self Times
(0.01 s.). This is due to the equal filter lengths.

Fig. 6. Magnitude and phase responses: f = 30Hz; DS  –40dB,

filter length 512
Figure 7 depicts the filtration of a sweep sine signal with
an amplitude of 1 V and a frequency from 0 to 4000 Hz.
The signal duration is 10 seconds, with sampling frequency
of 8000 Hz. The simulation has been made with Matlab®
function filter (.).
Figures 7 and 8 depict the simulation results. As can be Fig. 9. Calculation times of the proposed filter: f = 30Hz; DS  –
seen the proposed filter possesses greater attenuation in the 40dB, filter length 512
 [1] J. W. Gibbs, (1898), "Fourier's Series", Nature, 59 (1522):
200.
[2] J. W. Gibbs, (1899), "Fourier's Series", Nature, 59 (1539):
606.
[3] D. Liron, Grossmann and Yonina C. Eldar, “An L1-Method
for the Design of Linear-Phase FIR Digital Filters,”Signal
Process., vol. 55, no. 11, 2007, pp. 5253–5265.
[4] D. W. Tufts and J. T. Francis, “Designing digital low-pass
filters—comparison of some methods and criteria,” IEEE
Trans. Audio Electroacoust., vol. AE-18, 1970, pp. 487–
494.
[5] V. R. Algazi and M. Suk, “On the frequency weighted least-
squares design of finite duration filters,” IEEE Trans.
Circuits Syst., vol. CAS-22, no. 12, Dec. 1975, pp. 943–
953.
[6] C. S. Burrus, A. W. Soewito, and R. A. Gopinath, “Least
squared error FIR filter design with transition bands,” IEEE
Trans. Signal Process., vol. 40, no. 6, Jun. 1992, pp. 1327–
1340.
Fig. 10. Calculation times of P&M filter: f = 30Hz; DS  – [7] G. Cherneva, E. Dimkina. „Simulation and Examination of
40dB, filter length 512 a Signal Masking Chaotic Communication System, Based
on the Duffing Oscillator“, Journal „Communications”.
The main advantage of the proposed filter is the shorter Zilina, Slovakia. ISSN 1335-4205.  2A/2013.Vol. 15, pp.
Total Time – 0.012 s., vs 0.214 s. to P&M filter. The P&M 6-10.
method uses Remez’ exchange algorithm. From the Profile [8] E. Z. Psarakis, “A weighted -based method for the design of
Summary in Figure 10 it can be seen that the Remez’ arbitrary one-dimensional FIR digital filters,” Signal
algorithm is called 1806 times, while the coefficients of the Process., vol. 86, 2006, pp. 937–950.
proposed filter are calculated by formulas (6), (7) and [9] I. W. Selesnick, M. Lang and C. S. Burrus “Constrained
shifted IFFT. Least Square Design of FIR Filters without Specified
Transition Bands,” IEEE Trans. on Signal Processing, vol.
V. CONCLUSION SP- 44, Aug. 1996, , pp. 1879-1892.
The paper examines one-dimensional FIR filter [10] G. Cherneva, E. Dimkina. „Chaotic Masking Approach
Based on the Duffing Oscillator“. Proceedings of the
synthesis. An approximation of an ideal BPF response with
International Scientific Conference ELEKTRO 2012. Zilina,
complementary integral Gaussian error function is Slovakia. 21-22.05.2012, pp.42-45.
performed. Analytical dependencies for magnitude [11] J. H. McClellan and T. W. Parks, “A unified approach to the
response, parameter  and transition bandwidth f are design of optimum FIR linear-phase digital filter,” IEEE
obtained. The theory is confirmed by  filter design Trans. Circuit Theory, vol. CT-20, Nov. 1973, pp. 697–701.
example. A comparison with P&M filters is made. As a [12] L. R. Rabiner, J. H. McClellan, and T. W. Parks, “FIR
result, FIR filters with better selectivity and lower digital filter design techniques using weighted Chebyshev
calculation time than P&M filters are obtained. The lower approximations,” Proc. IEEE, vol. 63, Apr. 1975, pp. 595–
calculation time implies application of the proposed filters 610.
[13] E. Ya. Remez, “Fundamentals of numerical methods for
in cases which require fast change of the parameters in real
Chebyshev approximations, “ Naukova Dumka, Kiev, 1969.
time. [14] P. S. Apostolov, B. Yurukov, A. Stefanov, „An Easy and
Efficient Method for Synthesizing Two-Dimensional Finite
ACKNOWLEDGMENT
Impulse Response Filters with Improved Selectivity“, IEEE
This paper is supported by the National Scientific Signal Processing Magazine, Sept. 2017, pp.180-183.
Program "Information and Communication Technologies [15] P. S. Apostolov, A. Stefanov, S. P. Apostolov. „A Study of
for a Single Digital Market in Science, Education and Filters Selectivity with Maximally Flat Responses with
Security (ICTinSES)", financed by the Ministry of Respect to Hausdorff Distance“, 9th National Conference
Education and Science. with International Participation, ELECTRONICA 2018, pp.
1-4.
REFERENCES