You are on page 1of 7

Circuits, Systems, and Signal Processing

https://doi.org/10.1007/s00034-018-0964-8

SHORT PAPER

A Versatile Hyperbolic Cosine Window for Spectral


Resolution

Raghavendra G. Kulkarni1

Received: 13 June 2018 / Revised: 6 October 2018 / Accepted: 9 October 2018


© Springer Science+Business Media, LLC, part of Springer Nature 2018

Abstract
A simple but versatile hyperbolic cosine window function is presented in this paper,
which has two terms in the time domain. A parameter p in the window can be varied to
make the mainlobe width of the window function to approach that of a rectangular win-
dow (±1/T ,) while maintaining higher sidelobe decay (12 dB/octave.) Even though
such behavior has been demonstrated by the two-term polynomial window, it suffers
from the limitation that only some discrete values of mainlobe widths can be achieved
in the range, ±1/T to ±1.5/T . The proposed hyperbolic cosine window has no such
limitation; one can achieve any desired value of mainlobe width in the above range.
The proposed window can be employed for applications involving spectral resolution.

Keywords Hyperbolic cosine window · Polynomial window · Spectral resolution ·


Decay of sidelobes · Mainlobe width

1 Introduction

Window functions find use in various situations requiring—spectral resolution, sup-


pression of harmonics, reducing co-channel interference, minimizing time-bandwidth
product, etc. A good spectral resolution of frequencies requires windows with mini-
mum possible mainlobe width, while suppression of harmonics requires windows with
fast decaying sidelobes, and applications requiring less co-channel interference need
windows with well suppressed nearby sidelobes [2]. Applications such as encoding
of audio signals in MP3 or some other formats and beam forming using apodization
function in a medical ultrasound system need windows with minimum possible root
mean square (RMS) time-bandwidth product [9,10]. However, the uncertainity prin-
ciple puts a limit on minimum achievable RMS time-bandwidth product (t ×  f )

B Raghavendra G. Kulkarni
raghavendrakulkarni@pes.edu ; dr_rgkulkarni@yahoo.com

1 Department of Electronics and Communication Engineering, PES University, 100 Feet Ring Road,
BSK III Stage, Bengaluru 560085, India
Circuits, Systems, and Signal Processing

as 1/4π , which is of course achieved by the Gaussian window [2]. Unfortunately,


Gaussian window extends up to ±∞ in both time and frequency domains and thus
has less practical use. The approximate confined Gaussian windows described in [10]
have RMS time-bandwidth product very close to the minimum limit, with mainlobe
widths greater than ±2.5/T , where T is the duration of the window.
The classic cosine-sum windows like Blackman window, Blackman–Harris win-
dow, and Blackman–Nuttall window given in [8] have highly suppressed sidelobes,
with mainlobe widths ranging from ±2.5/T to ±4/T . These windows are useful in
minimizing the co-channel interference. Closely related to the cosine-sum windows are
the odd and even series cosine windows [3,4,7], which have varying mainlobe widths
depending on the number of terms (N ) in the series. The mainlobe widths obtainable
for odd series are: ±2/T , ±3/T , ±4/T , ... and for even series the mainlobe widths
are: ±2.5/T , ±3.5/T , ±4.5/T , ... for N = 2, 3, 4, ... The odd and even cosine series
windows can be designed either to minimize nearby sidelobes [3] or to maximize
the sidelobe decay [4] and thus are handy for applications requiring suppression of
co-channel interference or suppression of harmonics, respectively.
The basic rectangular window has minimum possible mainlobe width (±1/T ), so it
appears to be ideal window for spectral resolution, but its high sidelobe level (−13 dB)
and slow decay of sidelobes (6 dB/octave) limit its usefulness for this application. Thus
any window with mainlobe width close to ±1/T and better performance in sidelobe
level or decay of sidelobes is naturally preferred.
The Reisz window [2], which is a second-order polynomial window, is a reasonably
good choice for spectral resolution as it has a mainlobe width of ±1.43/T and decays
at 12 dB/octave rate. It has its highest sidelobe level at −22 dB. However, applications
requiring still narrower mainlobe widths cannot use this window. The polynomial
windows presented in [5,6] have fast decay of sidelobes, and it can be noticed that as
the order of the window (2N ) increases, the mainlobe width shrinks and approaches
±1/T , while maintaining higher (≥ 12 dB/octave) decay of sidelobes, rendering them
useful for applications requiring good spectral resolution. The only limitation with the
polynomial windows is that—only some discrete values of mainlobe widths can be
achieved as the order of the windows is in discrete numbers (even integers).
This paper presents a simple hyperbolic cosine window, which does not have the
above mentioned limitation of the polynomial window. The hyperbolic cosine window
contains a parameter p, which when varied yields any desired value of mainlobe width
in the range, ±1/T to ±1.5/T , with a sidelobe decay of 12 dB/octave.

2 Hyperbolic Cosine Window

The hyperbolic cosine window function h(t) in time domain is an even, continuous,
positive, and finite function in t, which exists for the duration from −T /2 to +T /2,
and is defined as:

h(t) = a0 + a1 cosh( pt/T ), for |t| ≤ T /2,


(1)
= 0, for |t| > T /2,
Circuits, Systems, and Signal Processing

where a0 and a1 are the coefficients in (1), and p is a positive real number. The
corresponding frequency domain function, H ( f ), is obtained by taking the Fourier
transform of h(t),
 +T /2
H( f ) = [a0 + a1 cosh( pt/T )]e− j2π f t dt. (2)
−T /2

Evaluation of the integral in (2) and further simplification yields an expression for
H ( f ) given as:
 
a0 sin(x) 2a1 [ p sinh( p/2) cos(x) + 2x cosh( p/2) sin(x)]
H( f ) = T + , (3)
x p 2 + 4x 2

where x = π f T . Notice that the coefficients, a0 and a1 , are unknowns requiring two
expressions to determine them. At t = 0, the time function h(t) is assumed to attain
a maximum value of unity, so from (1) we obtain,

a0 + a1 = 1. (4)

It is desired to have maximum possible decay of sidelobes for the proposed window.
The criteria for maximum sidelobe decay is that—the time function h(t) and its first
n derivatives have to vanish at ±T /2, if the window has to decay as 1/ f n+2 at the
rate 6(n + 2) dB/octave [4,8]. Since only one more equation is needed to determine
the coefficients, the above criteria leads to h(±T /2) = 0, which further yields,

a0 + a1 cosh( p/2) = 0, (5)

Using (4) and (5) a0 and a1 are determined as,

a0 = cosh( p/2)/[cosh( p/2) − 1],


a1 = −1/[cosh( p/2) − 1]. (6)

Use of the above expressions in (3) yields,

T p{ p cosh( p/2)[sin(x)/x] − 2 sinh( p/2) cos(x)}


H( f ) = . (7)
[cosh( p/2) − 1]( p 2 + 4x 2 )

One can notice from (7) that the frequency function H ( f ) decays as 1/ f 2 with 12
dB/octave decay of sidelobes. Observe that H ( f ) is a function of the parameter p; in
order to study the behavior of H ( f ) with respect to variation of p, we first normalize
H ( f ) with respect to its amplitude at f = 0. Therefore, let H ( f )nor = H ( f )/H (0),
where H (0) is obtained from (7) as,

T [ p cosh( p/2) − 2 sinh( p/2)]


H (0) = . (8)
p[cosh( p/2) − 1]
Circuits, Systems, and Signal Processing

Fig. 1 Frequency and time plots of hyperbolic cosine windows. a Frequency plots. b Time plots

Table 1 Salient features of hyperbolic cosine window

p a0 a1 Mainlobe width Sidelobe level Sidelobe decay

0.25 128.8334635 −127.8334635 ±1.43/T −21.3 dB 12 dB/oct


5 1.194844816 −0.194844816 ±1.326/T −18.8 dB 12 dB/oct
25 1.000007453 −0.000007453 ±1.0847/T −13.89 dB 12 dB/oct

Hence H ( f )nor is obtained as:

p 2 { p cosh( p/2)[sin(x)/x] − 2 sinh( p/2) cos(x)}


H ( f )nor = . (9)
[ p cosh( p/2) − 2 sinh( p/2)]( p 2 + 4x 2 )

The frequency function, H ( f )nor , is now plotted versus frequency as shown in Fig. 1a
for various values of p. These plots reveal that as p is increased from 0.25 to 25,
the mainlobe width decreases from ±1.43/T to ±1.0847/T , and the sidelobe level
increases from −21.3 dB to −13.89 dB, while asymptotic decay of sidelobes remains
at 12 dB/octave. A proper choice of p yields the desired mainlobe width. If the window
has to provide a mainlobe width of say, B, then its frequency function, H ( f ), has to
be zero at f = ±B/2, and this yields [see (7)],

π BT
p= tanh( p/2), (10)
tan(π BT /2)

which is a transcendental equation in p, required to be solved numerically. When B =


2/T (i.e., for mainlobe width of ±1/T ), p = ∞; and when B = 3/T (i.e., for mainlobe
width of ±1.5/T ), p = 0. Hence, mainlobe widths in the range, ±1/T to ±1.5/T ,
only can be achieved by this window. Table 1 lists several values of mainlobe widths
and the corresponding sidelobe levels obtainable, with p as a parameter. Figure 1b
shows the hyperbolic cosine window in time domain for p = 0.25, 5, and 25; we
notice that the shape of window function approaches that of the rectangular window
as the value of p is increased.
Circuits, Systems, and Signal Processing

3 Comparison with Polynomial Window

Since the two-term polynomial window described in [5] also achieves a variation
in mainlobe width similar to the proposed window, it is worthwhile to compare the
two windows. For the two-term polynomial window of order 2N , the time function
[ p2N (t)] and the frequency function [P2N ( f )] are given by:

p2N (t) = 1 + C2N (t/T )2N , for |t| ≤ T /2,


= 0, for |t| > T /2, and
P2N ( f ) = I0 + C2N I2N , (11)

where I0 and I2N are given by,

sin(x) T N (2N − 1)
I0 = T , I2N = N 2 [x sin(x) + 2N cos(x)] − I2(N −1) , (12)
x 4 x 2x 2
and x = π f T . Notice from Table 1 given in [5] that—as the order of the window (2N )
increases, the mainlobe width approaches ±1/T . From (11) and (12), we observe that
in order to obtain the expression for the frequency function, P2N ( f ), we need the
expressions for I0 and I2N ; to get I2N , we need to determine I2(N −1) , and to obtain
I2(N −1) , one has to determine I2(N −2) , and so on.
Therefore, as we try to achieve a mainlobe width closer to ±1/T by increasing
the order of the window, we need to derive more and more expressions. Thus the
derivation of the frequency function P2N ( f ) becomes cumbersome for large N ; we
take an example to illustrate this.
Let us derive an expression for the frequency function, P10 ( f ). From (11), we
get P10 ( f ) = I0 + C10 I10 . Referring to Table 1 given in [5], we note that C10 =
−1024. The expression for I2N [see (12)] implies that determination of I10 requires
determination of I8 and that of I8 requires determination of I6 , and so on as shown
below.
T (x sin x + 10 cos x) 45I8 T (x sin x + 8 cos x) 14I6
I10 = 2
− 2
, I8 = − 2
1024x 2x 256x 2 x
T (x sin x + 6 cos x) 15I4 T (x sin x + 4 cos x) 3I2
I6 = − , I4 = − 2
64x 2 2x 2 16x 2 x
T (x sin x + 2 cos x) I0 sin(x)
I2 = − 2 , and I0 = T . (13)
4x 2 2x x
With the use of expressions in (13), P10 ( f ) is derived after much effort as:

10T 
P10 ( f ) = (9 sin x/x)(x 8 − 56x 6 + 1680x 4 − 20160x 2 + 40320)
x 10 
− cos x(x 8 − 72x 6 + 3024x 4 − 60480x 2 + 362880) . (14)

The normalized frequency function of P10 ( f ) is: P10 ( f )nor = P10 ( f )/P10 (0), where
P10 (0) is the value of P10 ( f ) at f = 0, and it is evaluated using the Fourier transform
Circuits, Systems, and Signal Processing

Fig. 2 Comparative frequency plots of hyperbolic cosine windows and two-term polynomial windows. a
Mainlobe width: ±1.09796/T . b Mainlobe width: ±1.2319/T

 +∞
Theorem [1], F(0) = −∞ f (t)dt, as: P10 (0) = T /1.1. Notice the amount of effort
involved in deriving the frequency function P10 ( f ).
For the proposed hyperbolic cosine window, no such cumbersome derivations are
involved to obtain the frequency function—one just need to resort to the expression
(9) and vary the parameter p to obtain any value of mainlobe width within the range,
±1/T to ±1.5/T . Let us compare the frequency response performance (mainlobe
width, sidelobe levels and sidelobe decay) of hyperbolic cosine window with that of
two-term polynomial window, for two values of mainlobe widths.
In the first case, a mainlobe width of ±1.09796/T is chosen, since it is achieved
by the two-term polynomial window, P10 ( f ) (see Table 1 in [5]). To obtain the same
mainlobe width for the proposed window, we use (10) and determine p numerically as,
p = 21.7043. Figure 2a shows comparative normalized frequency plots of hyperbolic
cosine window with p = 21.7043 and P10 ( f ); notice that both the frequency plots are
almost identical with the same sidelobe levels, decay of sidelobes, and the mainlobe
width. In the second case, a mainlobe width of ±1.2319/T is chosen, as this is the
value achieved by the two-term polynomial window, P4 ( f ) (see Table 1 in [5]). To
get the same mainlobe width with the proposed window, we solve (10) and obtain
p = 8.67169. The frequency plots of the two windows are compared in Fig. 2b; one
can notice that the frequency responses of the two windows match very well.
We note from Table 1 in [5] that the two-term polynomial window achieves only
some discrete values of mainlobe widths corresponding to 2N = 2, 4, 6, 8, 10, etc.,
whereas for the proposed window there is no such limitation; practically any value of
mainlobe width can be achieved within ±1/T to ±1.5/T . The only limitation of this
window is—its decay cannot be increased beyond 12 dB/octave as it contains only
two terms in the time domain like the two-term polynomial window.

4 Conclusions

This paper has described a hyperbolic cosine window function, which is capable of
achieving any desired mainlobe width between ±1/T and ±1.5/T , while maintaining
Circuits, Systems, and Signal Processing

a sidelobe decay of 12 dB/octave, by just varying a parameter p in the window, unlike


the earlier reported two-term polynomial window, which has the drawback of achieving
only some discrete values of mainlobe widths in the range, ±1/T and ±1.5/T . The
proposed window is well suited for applications requiring spectral resolution.

Acknowledgements The author thanks the management of PES University for supporting this work. The
valuable comments of Associate Editor and the anonymous referees improved the manuscript considerably.

References
1. R.N. Bracewell, The Fourier Transform and Its Applications, 3rd edn. (Tata McGraw-Hill, New Delhi,
2003), p. 152
2. F.J. Harris, On the use of windows for harmonic analysis with discrete Fourier transform. Proc. IEEE
66(1), 51–83 (1978)
3. R.G. Kulkarni, S.K. Lahiri, Improved sidelobe performance of cosine series functions. IEEE Trans.
Ultrason. Ferroelectr. Freq. Control UFFC–46(2), 464–466 (1999)
4. R.G. Kulkarni, Asymptotic behavior of cosine windows. Microwav. J. 43(10), 96–104 (2000)
5. R.G. Kulkarni, Polynomials for narrowband signal processing. IEE Proc. Vis. Image Signal Process.
149(3), 159–161 (2002)
6. R.G. Kulkarni, Polynomial windows with fast decaying sidelobes for narrow-band signals. Signal
Process. 83(6), 1145–1149 (2003)
7. D.C. Malocha, C.D. Bishop, The classical truncated cosine series functions with applications to SAW
filters. IEEE Trans. Ultrason. Ferroelectr. Freq. Control UFFC–34, 75–85 (1987)
8. A.H. Nuttall, Some windows with very good sidelobe behavior. IEEE Trans. Acoust. Speech Signal
Process. ASSP – 29(1), 84–91 (1981)
9. K.J. Parker, Apodization and windowing functions. IEEE Trans. Ultrason. Ferroelectr. Freq. Control
60(6), 1263–1271 (2013)
10. S. Starosielec, D. Hagele, Discrete-time windows with minimal RMS bandwidth for given RMS tem-
poral width. Signal Process. 102, 240–246 (2014)

You might also like