Digital Signal Processing 22 (2012) 859–868

Contents lists available at SciVerse ScienceDirect

Digital Signal Processing

www.elsevier.com/locate/dsp

A novel two-stage nonrecursive architecture for the design of generalized comb

filters
Gordana Jovanovic Dolecek a,∗ , Massimiliano Laddomada b
a
Department of Electronics, Institute INAOE, Puebla, Mexico
b
Electrical Engineering Department, Texas A&M University–Texarkana, USA

a r t i c l e i n f o a b s t r a c t

Article history: This paper presents a novel two-stage class of decimation filters with superior spurious signal rejection
Available online 12 April 2012 performance around the so-called folding bands, i.e., frequency intervals whose signals get folded down
to baseband due to decimation. The key idea to enhance signal rejection in the frequency domain lies
Keywords:
on an effective way to place the zeros of a classical comb filter in the aforementioned folding bands. On
A/D converter
CIC filter
the other hand, the paper provides a mathematical framework for designing two-stage multiplierless and
Comb nonrecursive structures of the proposed filters.
Decimation Examples are provided to highlight the key steps in the design of the proposed filters. Moreover, the
Decimation filter frequency behavior of the proposed filters in both baseband and stopband is compared with classical and
Generalized comb filters generalized comb filters, and a droop compensator is proposed to counteract the passband distortion of
the proposed filters.
© 2012 Elsevier Inc. All rights reserved.

1. Introduction

Multirate architectures for sampling rate conversion have been

under investigation for many years [1]. From a practical point of
view, the decimation of a highly oversampled signal is accom-
plished using a cascade of two (or more) stages of decimation [1].
The first stage is usually a comb filter of order N decimating by a
factor M, while the last stage employs an FIR filter providing the
required selectivity on the sampled signal down-converted to base-
band. Moreover, the design of a decimation stage in a multistage
architecture imposes stringent constraints only on the shape of the
filter frequency response over the so-called folding bands.
Let us briefly highlight the meaning of folding bands in connec-
tion to the first decimation stage in a multistage architecture. Con-
sider an analog signal x(t ) with baseband bandwidth [− B x , + B x ]
sampled by an A/D converter at the frequency f s = ρ 2B x (ρ is
the oversampling ratio; by definition, it is ρ  1). The discrete-
time sampled signal, x[nT s ], presents normalized frequency band-
width in the dimensionless frequency set [− f 0 , + f 0 ], with f 0 =
B x / f s = B x /(ρ 2B x ) = 1/(2ρ ). With this setup, the frequency re-
sponse of the anti-aliasing filter employed in the first decimation
stage should attenuate the quantization noise (QN) and any other
undesired signal falling inside the frequency intervals defined as
*
Corresponding author.
E-mail addresses: gordana@ieee.org (G.J. Dolecek), mladdomada@tamut.edu
(M. Laddomada).
Fig. 1. Pictorial representation of the frequency intervals (folding bands) that should
be carefully considered for the design of the first decimation stage.
 
k k
− f 0, + f0 , for k = 1, . . . ,  M . (1)
M M
The reason follows upon noting that the signals within these fre-
quency bands will fold down to baseband because of the sampling
rate reduction by M in the first decimation stage, thus irremedia-
bly affecting the signal resolution after the multistage decimation
chain [1]. These frequency ranges are pictorially shown in Fig. 1.
Classical comb filters with z-transfer function [2]
1 1 − z− M
H ( z) = (2)
M 1 − z −1
have many advantages over other filters. Among them, the main
advantage stems from the fact that their system function is ex-
tremely simple to implement and it does not require any multi-
plier. However, the magnitude response exhibits low attenuation

1051-2004/$ – see front matter © 2012 Elsevier Inc. All rights reserved.
http://dx.doi.org/10.1016/j.dsp.2012.04.007
860 G.J. Dolecek, M. Laddomada / Digital Signal Processing 22 (2012) 859–868

in the folding bands, as well as a considerable passband droop that

deteriorates the sampled signal.
With this background, let us provide a survey of the recent lit-
erature related to the problem addressed in this paper. Tutorials
focusing on the design of multirate filters can be found in [1,3–7].
The design of optimized multistage decimation and interpolation
filters has been recently addressed by Coffey in [8], while the de-
sign of multistage decimation architectures relying on constituent
cyclotomic polynomial filters has been presented in [9,10]. Imple-
mentation aspects of comb filters with special emphasis on power
consumption and overflow have been addressed in [11,12] where
nonrecursive architectures relying on the polyphase decomposition
were introduced. Methods [13,14] proposed compensation filters
aimed to decrease the passband droop of comb filters.
 In [15], the
authors proposed novel decimation schemes for Δ A/D convert-
ers based on Kaiser and Hamming sharpened filters. These filters
were improved in [16], then generalized in [17] for higher order
decimation filters, and in [19] for wideband receivers. In [20], the
authors introduced a two-stage architecture presenting better mag-
nitude characteristics than the structure proposed in [16].
A 3rd-order modified decimation sinc filter was proposed in
[21], and developed in [22]. The class of comb filters was gen-
eralized in [24], whereby the authors proposed an optimization
framework for deriving the optimal zero rotations of Generalized
Comb filters (GCFs) for any filter order and decimation  factor. GCFs Fig. 2. Magnitude response (in dB) of the 3rd-order GCF filter in (3) compared with
in [24] were proposed with the aim of increasing the Δ QN re- the magnitude response of a 3rd-order comb. Design parameters are as follows:
M = 16, ρ = 64, and α = 0.03828.
jection around the folding bands with respect to classical comb
filters
 of equivalent order. In spite of the particular application to
Δ A/D converters, GCFs are pretty general in that they can be choice for α is α = q2π f 0 , with q ∈ [−1, +1]: this choice is such
used in place of classical comb filters in multistage rate conversion that the zeros of H GCF3 ( z) fall within the folding bands in (1). The
architectures as the ones employed for signal extraction in broad- magnitude of a 3rd-order GC filter is compared with the magni-
band and reconfigurable digital receivers. Works [25,26] proposed tude response of a classical 3rd-order comb filter in Fig. 2 for the
compensation filters to decrease the passband droop of GC filters, following parameters: M = 16, ρ = 64, α = 0.03828, and q = 0.78.
while [27] addressed the design of efficient GC filters exploiting The drawback stemming from the zero-optimization is the in-
recursive architectures. troduction of multipliers in the first stage of decimation, as well
The considerations drawn above in connection to the frequency as the impossibility to implement multiplierless recursive architec-
behavior of a classical comb filter around the folding bands, par- tures due to the instability arising from imperfect zero-pole can-
ticularly around the first folding band, represent the very starting cellation when coefficients quantization is exploited. In addition,
point toward the ideas proposed in this work. The main ratio- GC filters exhibit passband droops of the same extent of classical
nales are discussed in Section 2 where we also point out the main comb filters.
contributions of this work. The rest of the paper is organized as The main contributions of this work stem from solving the
follows. In Section 3, we derive the z-transfer function of the pro- aforementioned problems, while retaining the advantages of GC fil-
posed decimation filters. Section 4 presents a variety of design ters. We propose a novel class of decimation filters featuring the
issues for improving the frequency response of the filters under following advantages hereafter identified as conditions.
investigation. Moreover, a mathematical framework for optimizing
the zero location of the z-transfer function is developed. In Sec- 1. No need of a programmable FIR filter at low rate to compen-
tion 5, we present some design examples in order to highlight the sate the passband droop introduced by comb filters.
choice of the parameters defining the behavior of the proposed fil- 2. Improved spurious signal rejection in the folding bands com-
ters. Finally, Section 6 draws the conclusion. pared to traditional Nth-order comb filters (in this respect the
proposed filters are superior as compared to comb filters).
2. Motivation 3. Nonrecursive implementation not suffering from the instability
problems arising from imperfect pole-zero cancellation due to
One of the advantages of the decimation filters proposed in finite precision effects.
[24] stems from the optimization of the zeros locations of a clas- 4. Overall multiplierless filter design.
sical Nth-order comb filter within the folding bands defined in (1).
A nice consequence of this optimization consists in an increased Let us briefly discuss in which way the present work differs
attenuation in the folding bands. Given a decimation factor M, from the previous work relevant to this one, developed by the
and neglecting the normalization constant ensuring unity gain at same authors. Papers [9,10] addressed the design of multistage
baseband, the z-transfer function of a 3rd-order GCF filter can be decimation filters based on cyclotomic polynomials. The problem
written as was to solve an optimization problem for finding a proper combi-
nation of cyclotomic polynomials to meet a set of filter specifica-
1 − z− M 1 − z− M e j α M 1 − z− M e − j α M
H GCF3 ( z) = , (3) tions. The design in [9,10] is different from the focus of this work
1 − z −1 1 − z −1 e j α 1 − z −1 e − j α since this very work proposes improved decimation filters stem-
where α is the extent of the rotation undergone by the zeros of ming from classical comb filters. Paper [16] presented a sharpened
a classical comb filter. Using the notation summarized in the pre- comb decimator structure consisting of a cascade of a comb-filter
vious section about the folding bands, we note that a convenient based decimator and a sharpened comb decimator. In the attempt
 G.J. Dolecek, M. Laddomada / Digital Signal Processing 22 (2012) 859–868 861

to reduce the passband droop of classical comb filters, the sharp-

ened architecture was used. One of the main drawbacks of the
sharpening technique is the use of multiple instances of the same
basic cell just for reducing the passband droop of the overall filter.
This work, on the other hand, attacks the passband droop problem
by cascading a very efficient cell at the end of the decimation ar-
chitecture without increasing the overall filter complexity. Similar
considerations can be derived for work [17] where the sharpen-
ing technique was employed for the sake of reducing the passband
droop of typical comb filters. Unlike this work, notice that the
filter architectures proposed in [16] and [20] did not present im-
proved signal rejection around the folding bands. Papers [18,23,24]
addressed the design of generalized comb filters for decimation
purposes but the proposed architectures included multipliers and
did not consider the passband droop compensation. As explained
above in this section, the main contributions of this work com-
pared to the previous ones is the proposal of filter architectures
featuring the four conditions listed above, which improve the pre-
vious work from the same authors.

Fig. 3. Magnitude characteristics of | H 1 (e j ω )| (curve labeled H1) and | H 2 (e j ω )|

3. The proposed decimation filters (curve labeled H2) for M = 16.

This section presents the rationales at the very basis of the pro-
posed class of decimation filters, derives the z-transfer function,
and discusses the frequency behavior of the proposed two-stage
decimation filters.
Fig. 4. Architecture 1.

3.1. Derivation of the two-stage architecture

  1 1 − z− M e j α M
H 21 z M 1 = ;
We assume that the decimation factor M is an even integer that 2 1 − z− M 1 e j α M 1
can be represented as follows:   1 1 − z− M e − j α M
H 22 z M 1 = . (8)
2 1 − z− M 1 e − j α M 1
M = 2M 1 . (4)
Eq. (8) can be rewritten as follows:
With this setup, the transfer function in (2) can be rewritten as      
H 2r z M 1 = H 21 z M 1 H 22 z M 1
1 1 − z− M 1 1 − z− M 1 1 1 − z−2M 1
H ( z) = = 1 1 − 2 cos(α M ) z− M + z−2M
M 1 − z −1 M 1 − z −1 2 1 − z − M 1
1 = . (9)
 M  4 1 − 2 cos(α M 1 ) z− M 1 + z−2M 1
= H 1 ( z) H 2 z 1
(5)
The overall z-transfer function is obtained by replacing H 2 ( z M 1 ) in
where (5) with H 2r ( z M 1 ) in (9):
1 1 − z− M 1 1   1 − z−2M 1       −1
H 1 ( z) = ; H 2 z M1 = ; H r ( z) = H 12 ( z) H 2r z M 1 = H 12 ( z) N 2r z M D 2r z M 1 , (10)
M 1 1 − z −1 2 1 − z− M 1
where:
H 2 ( z ) = 1 + z −1 . (6)
  1
N 2r z M = 1 − 2 cos(α M ) z− M + z−2M
The magnitude characteristics of the z-transfer functions in (6) can 4
be derived upon replacing z with e j ω : 1
= 1 − 2 cos(2α M 1 ) z− M + z−2M ,
  4
  j ω   1 sin( M21 ω )   
H1 e  =   D 2r z M 1 = 1 − 2 cos(α M 1 ) z− M 1 + z−2M 1 .
 M sin( ω ) , (11)
1 2
  Some considerations are in order. From (10), it is easy to note the
  j ω M   1 sin( M2ω ) 
H2 e 1 
=  . (7) presence of a classical 2nd-order comb filter H 12 ( z) with decima-
2 sin( ω M 1 )  tion M 1 in the first stage of the proposed filter. This stage can be
2
implemented in nonrecursive form by using polyphase decomposi-
As an example, the magnitude characteristics of both | H 1 (e j ω )| and tion as explained in [11].
| H 2 (e j ω )| are shown in Fig. 3 for M = 16. From Fig. 3 we can rec- On the other hand, using the commutative properties of mul-
ognize that, upon applying zero rotation only to the second filter tirate systems [1], and recalling the relation M = 2M 1 , the second
| H 2 (e j ω M 1 )| in the cascade, rotation would occur in all zeros but stage H 2r ( z M 1 ) can be moved after the decimation by M 1 yielding
the last one. [ D 2r (z)]−1 before the decimator by 2 and N 2r (z) after the over-
all decimator by M. With this said, we notice the presence of two
3.2. Moving zero rotation to lower rate multipliers: the first operates at M 1 lower rate while the second
one at M times lower rate. The architecture of the decimation filter
Let us focus on the benefits deriving from applying zero rota- with z-transfer function in (10) is depicted in Fig. 4. This architec-
tion only to the second stage H 2 ( z M 1 ). Applying zero rotation to ture can also be generalized by cascading an Nth-order comb filter.
H 2 ( z M 1 ), we get: By doing so, the following decimation filter H r ( z) easily follows:
862 G.J. Dolecek, M. Laddomada / Digital Signal Processing 22 (2012) 859–868
Fig. 5. Architecture 2 for implementing the proposed multirate filter.
Fig. 6. Architecture 3 for implementing the proposed multirate filter.
   
H r ( z) = H 12 ( z) H 2r z M 1 H N ( z) = H 1N +2 ( z) H 2N z M 1 H 2r z M 1
  N 2r ( z M )
= H 1N +2 ( z) H 2N z M 1 . (12)
D 2r ( z M 1 )
Additionally, this structure works well for infinite precision coeffi-
cients because the poles on the unit circle belonging to D 2r ( z M 1 )
are perfectly canceled by zeros belonging to N 2r ( z M ).
Upon employing the commutative properties of multirate sys-
tems [1,3], the z-transfer function in (12) can be implemented with
the architecture shown in Fig. 5. The first stage is a classical comb
filter that can be implemented in nonrecursive, mixed nonrecur-
sive/recursive form, or by resorting to polyphase decomposition.
The z-transfer function H 2r ( z M 1 ) can be implemented effi-
ciently as discussed in the next section.
3.3. Nonrecursive implementation of the zero rotation term
Upon considering (4), (9) can be rewritten as
 M1
 1 1 − 2 cos(2α M 1 ) z− M + z−2M
H 2r z = .
4 1 − 2 cos(α M 1 ) z− M 1 + z−2M 1
Using the well-known trigonometric identity
cos(2α M 1 ) = 2 cos2 (α M 1 ) − 1,
Eq. (13) becomes:
  1 1 − 2[2 cos2 (α M 1 ) − 1] z−2M 1 + z−4M 1
H 2r z M 1 =
4 1 − 2 cos(α M 1 ) z− M 1 + z−2M 1
The z-transfer function in (15) is the recursive representation of
the FIR filter:
  1
H 2r z M 1 = 1 + 2 cos(α M 1 ) z− M 1 + z−2M 1 ,
4
as it can be verified by accomplishing the polynomial division in
(15) against the variable z. Upon employing (16), we obtain the
more efficient structure depicted in Fig. 6. Note that the filter (16)
is a linear phase filter with only one multiplier. Moreover, unlike
(15), H 2r ( z M 1 ) in (16) does not suffer from any instability problem
since there is no pole-zero cancellation involved in it.
4. Design issues
This section proposes the use of a simple filter to improve the
passband characteristics and spurious signal rejection in the fold-
ing bands, and to design multiplierless architectures.
4.1. Passband droop compensation
Let us focus on the compensation of the passband droop of
the proposed filters. Since the proposed filters present monoton-
ically decreasing behavior from the zero frequency through the
frequency corresponding to the signal bandwidth f 0 , the passband
Table 1
Values of the parameter b.
Parameters N + 2 Parameter b
2 2
3 1
4 1
5 0
6 0
7 0
8 0
9 −1
 
droop can be evaluated as the value of the filter frequency re-
sponse in f 0 . The goal is to cascade a computationally efficient,
yet simple, multiplierless filter compensating the value of the filter
frequency response H r (e j ω ) evaluated in f 0 . To achieve this goal,
we adopt the simple compensator proposed for comb compensa-
tion in [13]:
 
H C e j ω M = 1 + 2−b sin2 (ω M /2), (17)
where b is an integer that depends only on the value N + 2. How-
ever, for the GCF compensation we have to define the new values
of b. The values of b for different values of N + 2, obtained by the
computer simulations, are given in Table 1.
From (17), we have the following z-transfer function of the
passband droop compensator
  
H C z M = A 1 + B z− M + z−2M , (18)
where
 
A = −2−(b+2) , B = − 2b+2 + 2 . (19)
(13)
Note that the compensator is multiplierless and requires only two
additions. The effect of the compensation filter on the frequency
response of the overall decimation filter is illustrated in Fig. 7
(14) where we compare the frequency response of the proposed pass-
band droop compensated filter with the frequency response of a
classical comb filter of the same order, as well as with the fre-
quency response of the filter H r ( z) without droop compensation.
. (15) As far as the practical implementation of the droop compen-
sator is concerned, basic multirate identity allows the compensator
in (17) to be moved to lower rate as shown in Fig. 8.
The z-transfer function of the proposed compensated filter is
 
(16) H rr ( z) = H r ( z) H C z M , (20)
where H r ( z) and H C ( z M ) are given in (12) and (18), respectively.
We notice in passing that the use of the droop compensator allows
the condition 1 to be fulfilled.
4.2. Optimization of the proposed decimation filters
In this section we propose a theoretical framework for optimiz-
ing the zero locations of the proposed decimation filters in such
a way to increase spurious signal attenuation around the folding
bands compared to classical comb filters.
To get started, consider the z-transfer function H rr ( z) of the
overall decimation filter
     
H rr ( z) = H 12+ N ( z) H 2N z M 1 H 2r z M 1 H C z M . (21)
The frequency response of this filter is:
         
H rr e j ω = H 12+ N e j ω H 2N e j ω M 1 H 2r e j ω M 1 H C e j ω M , (22)
where the function
 G.J. Dolecek, M. Laddomada / Digital Signal Processing 22 (2012) 859–868 863

Notice that the normalization constant of the 2nd-order rotated

cell is:
1 − cos(α M )
. (26)
1 − cos(α M 1 )
Therefore, it is:
    1 1 − cos(α M 1 )
H rr e j ω M = H C e j ω M
M 12 M N 1 − cos(α M )
− j ω[ N M2−1 +( M −1)] N
×e
   
sin( ω2M ) N sin( ω M
2
1
) 2
×
sin( ω
2
) sin( ω
2
)
cos(ω M ) − cos(α M )
× , (27)
cos(ω M 1 ) − cos(α M 1 )
where
1 1 − cos(α M 1 )
(28)
M 12 M N 1 − cos(α M )

is the normalization constant of the overall decimation filter in

(21) assuring unity gain at baseband. The magnitude squared of
the frequency response is
 
  j ω M 2   j ω M 2  1 1 − cos(α M 1 ) 2
 H rr e  = H C e   
 M 2 M N 1 − cos(α M ) 
1
   
sin( ω2M ) 2N sin( ω M 2
1
) 4
×
sin( ω
2
) sin( ω
2
)
[cos(ω M ) − cos(α M )]2
× (29)
[cos(ω M 1 ) − cos(α M 1 )]2
where we have omitted the magnitude of the Euler exponential
e − j Ω in (27) being equal to one regardless of Ω .
Let us address the choice of the parameter α in (29). The main
objective of the proposed class of decimation filters is the increase
of the filter attenuation around the folding bands; therefore, a con-
venient choice for α is

α = q2π f 0 , q ∈ [−1, +1]. (30)

Fig. 7. Magnitude responses of the proposed filter compensated (a), and behavior of
the magnitude responses in the passband (b).
This definition is such that the couple of zeros belonging to H 2r ( z)
falls inside each folding band. We notice in passing that the inde-
pendent parameter appearing in the definition of α is q.
The optimal α , identified by αopt , can be evaluated by solving
the optimization strategy discussed in the following. Find αopt as
the argument of the optimization problem:

  
S n ( f ) H rr e j2π f  df ,
2
Fig. 8. Architecture of the proposed filter embedding the droop compensator in the αopt = arg min (31)
last stage of the two-stage architecture. α
Bω
 
    1 M −1 sin( ω2M ) N where the integral is extended over the digital bandwidth B w cor-
H 1N e j ω H 2N e j ω M 1 = N e − j ω N 2 (23) responding to the first folding band defined as
M sin( ω )
2
 
1 1 1 1
is the frequency response of a classical comb filter of order N, Bω = − , + , (32)
M 2ρ M 2ρ
while the other constituent functions are defined as follows:
1
with lower edge f le = M − 21ρ . The function S n ( f ) is the power
 2
  1 sin( ω M 1
) spectrum density of the quantization noise. For oversampled A/D
H 12 e j ω = e − j ω ( M 1 −1 ) 2
, (24) converters without noise shaping, S n ( f ) is a constant function over
M 12 sin( ω
2
) the whole frequency range; therefore, it does not have any effect
  1 − cos(α M 1 ) − j ω( M − M 1 ) on the optimization problem and it can be neglected. On the other
H 2r e j ω M 1 = e hand, for noise-shaped A/D converters, such as sigma-delta, it is a
1 − cos(α M )
function of the digital frequency f , but it can be very well con-
cos(ω M ) − cos(α M ) sidered constant over the first folding band [28]; therefore, it can
× . (25) also be neglected since multiplicative constant terms do not affect
cos(ω M 1 ) − cos(α M 1 )
864 G.J. Dolecek, M. Laddomada / Digital Signal Processing 22 (2012) 859–868

the optimization problem derived above. Therefore, the optimiza- Table 2

tion problem (31) can be solved as follows: Optimal values of α for M 2 = 2 and various choices of M 1 and ρ .

qopt αopt Denoising
d   
S n ( f ) H rr e j2π f  df = 0.
2 −[dB]
(33)
dα ρ = 128, N = 3
Bω M1 = 32; M 2 = 2; 0.911 0.02235930396422 10.73
M1 = 16; M 2 = 2; 0.896 0.02199114857513 12.57
Upon omitting the constant function S n ( f ), it is M1 = 8; M 2 = 2; 0.887 0.02177025534167 12.95

= 4; M 2 = 2;
d    M1 0.884 0.02169662426385 13.03
 H rr e j2π f 2 df = 0, (34) M1 = 2; M 2 = 2; 0.883 0.02167208057125 13.05
dα
Bω ρ = 64, N = 3
M1 = 16; M 2 = 2; 0.911 0.04471860792844 10.73
where | H rr (e j2π f )|2 is defined in (29) with ω = 2π f . Therefore, M1 = 8; M 2 = 2; 0.896 0.04398229715026 12.56
M1 = 4; M 2 = 2; 0.887 0.04354051068335 12.94
(34) can be rewritten as M1 = 2; M 2 = 2; 0.883 0.04334416114250 13.02

  
d   j2π M 2  1 1 − cos(α M 1 ) 2 ρ = 32, N = 3
HC e    M 1 = 8; M 2 = 2;
dα  M 2 M N 1 − cos(α M )  0.911 0.08943721585688 10.71
1 M 1 = 4; M 2 = 2; 0.895 0.08786641953009 12.52
Bω
M 1 = 2; M 2 = 2; 0.886 0.08698284659627 12.89
 2N  4
sin(π f M ) sin(π f M 1 ) ρ = 16, N = 3
× M 1 = 4; M 2 = 2; 0.910 0.17867808217292 10.66
sin(π f ) sin(π f )
 M 1 = 2; M 2 = 2; 0.893 0.17534013997848 12.39
[cos(2π f M ) − cos(α M )] 2
× df = 0 (35) ρ = 8, N = 3
[cos(2π f M 1 ) − cos(α M 1 )]2 M 1 = 2; M 2 = 2; 0.908 0.35657076618244 10.42

which simplifies to
 2 Therefore, it is the noise power rejection gain of the proposed filter
1 1 − cos(α M 1 ) with compensator compared to a classical comb filter.
M 12 M N 1 − cos(α M ) From the results shown in the table above, we notice a rejection

    gain of more than 10 dB compared to classical 5th-order comb
  j2π f M 2 sin(π f M ) 2N sin(π f M 1 ) 4 d
× H C e  filters. Therefore, condition 2 is met.
sin(π f ) sin(π f ) dα
Bω
4.3. Multiplierless, nonrecursive architecture for implementing the filter
 
[cos(2π f M ) − cos(α M )]2 cell H 2r ( z)
× df = 0. (36)
[cos(2π f M 1 ) − cos(α M 1 )]2
This section presents an effective algorithm guaranteeing mul-
Upon neglecting the constant multiplicative terms, after some al- tiplierless architectures of the proposed class of decimation filters.
gebra to find the derivative against the variable α in the integral, Let us quantize the multiplier 2 cos(α M 1 ) in (16) as
the following equation can be obtained:

N = 2−k I ,
 (40)
F ( f ) cos(2π f M ) − cos(α M )
where I is an integer that can be represented as a sum of power-
Bω
  of-2 terms. Upon using (40), H 2r ( z) = 0.25[1 + 2 cos(α M 1 ) z−1 +
M 2 [cos(2π f M 1 ) − cos(α M 1 )] sin(α M ) + [cos(2π f M ) − cos(α M )] sin(α M 1 )
× df z−2 ] can be rewritten as
[cos(2π f M 1 ) − cos(α M 1 )]3

=0 (37) 
H 2r ( z) = 2−2 1 + 2−k+1 I z−1 + z−2 . (41)
where F ( f ) is defined as follows
Next line of pursuit consists in defining the integer k used above.
   
  j2π f M 2 sin(π f M ) 2N sin(π f M 1 ) 4 This is addressed in the next subsection.

F ( f ) = HC e  . (38)
sin(π f ) sin(π f )
4.4. Choice of the integer k
The integral in (37) cannot be solved in closed-form, and numer-
ical integration has to be employed to find the optimal solution We choose a tolerance Δ in such a way that the actual value
(we have used trapezoidal numerical integration implemented in of α , called α1 , differs from αopt by the chosen value Δ, i.e., α1 =
Matlab). The optimal values of α (and q = α /2π f 0 ) are noted in αopt + Δ. Basically, α1 is the actual value of αopt resulting from
Table 2 for the parameters shown in the first column. We chose the quantization of 2 cos(α1 M 1 ), while Δ is the maximum error
N = 3 since such value let H rr ( z) in (21) be equivalent to a 5th- tolerable.
order comb filter embedded in practical Digital Down Converters A convenient choice justified by simulation results is the value
(DDCs). Δ = 0.001. By this setup, we have cos(α1 M 1 ) = 2−k I , while I max =
The term Denoising in the last column of Table 2 is defined as 2k − 1. Therefore, it is
follows:
 

 cos(α1 M 1 )  2−k 2k − 1 = 1 − 2−k . (42)
  2
10 log10 S n ( f ) H comb, N +2 e j2π f  df
From the previous relation, the following inequality easily follows:
Bω

  
   2k  1/ 1 − cos(α1 M 1 ) . (43)
S n ( f ) H rr e j2π f  df .
2
− 10 log10 (39)
Bω Upon solving for k, we obtain:
 G.J. Dolecek, M. Laddomada / Digital Signal Processing 22 (2012) 859–868 865

 
1 log10 ( 1−cos(1α )
1 M1 )
k  log2 →k . (44)
1 − cos(α1 M 1 ) log10 (2)
The latter can be rewritten as follows:

log10 ( 1−cos(1α )
1 M1 )
k= , (45)
log10 (2)
where the function x returns the integer greater than or equal to
the real number x.
Next section presents two design examples to clarify the choice
of the aforementioned parameters.

5. Design examples and comparisons

This section presents two design examples in order to empha-

size the steps for the choice of the parameters appearing in the
proposed decimation filters. Comparisons are also provided with
respect to classical comb filters.

5.1. Design example 1

Fig. 9. Magnitude response of the proposed filter with quantized multipliers com-
pared to a 5th-order comb filter.
In this design example we consider a decimation factor M =
16 and an oversampling factor ρ = 64. This example accounts for
moderately oversampled signals.
The normalized bandwidth of the sampled signal is f 0 = 21ρ =
0.0078125 (this means that the analog bandwidth B x of the signal
is mapped to the digital frequency interval [− f 0 , + f 0 ] after A/D
conversion with sample rate f s ), while the first folding band (from
(1)) is
 
1 1
− f 0, + f 0 = [0.0546875, 0.0703125].
M M
Since ω = 2π f , when the x-axis represents the variable ω/π the
ω
aforementioned intervals are mapped to the frequencies π0 = ρ1 =
0.015625.
 
2 2
− 2 f 0, + 2 f = [0.109375, 0.140625].
M M
Let us discuss the design steps of the proposed decimation filter.

1. Given M = 16, it is M 1 = 8. The order N = 3 is chosen in or-

der to compare the proposed filter to a 5th-order comb filter Fig. 10. Magnitude response of the proposed filter compared to a 5th-order comb
employed in commercial products. filter around the first folding band.
2. The parameter b belonging to the droop compensation filter is
b = 0 from Table 1.
3. From Table 2, the optimal values of the zeros rotations are
qopt = 0.896 and αopt = 0.0439823, thus yielding an additional
noise power rejection of 12.56 dB compared to a 5th-order
comb filter.
4. Considering Δ = 0.001 in α1 = αopt + Δ = 0.0449823, and us-
ing (46), the value k = 4 easily follows.
5. Finally, I = 15 from I = 2k − 1.

Fig. 9 compares the magnitude responses of the proposed filter

H rr ( z) with that of a 5th-order comb filter. A key observation here
is the presence of the rotated zeros spanning the folding band
[0.109375, 0.140625] as clearly highlighted in Fig. 10 around the
first folding band. The passband behavior of the proposed filter
H rr ( z) is compared to the one of a classical 5th-order comb filter
in Fig. 11.
Notice that the proposed filter introduces a maximum droop
of 0.06 dB, while the 5th-order comb filter presents a signal dis-
tortion as high as 1.1 dB. The latter value must be compensated
by the FIR filter following the comb filter in any multistage deci- Fig. 11. Magnitude response of the proposed filter compared to a 5th-order comb
mation chain, thus increasing the computational complexity of the filter at baseband. The signal bandwidth extends from 0 to 0.03125 in terms of
overall decimation filter. normalized frequencies.
866 G.J. Dolecek, M. Laddomada / Digital Signal Processing 22 (2012) 859–868

Fig. 12. Magnitude response of the proposed filter with quantized multipliers com- Fig. 13. Magnitude response of the proposed filter with quantized multipliers com-
pared to the frequency response of the filter employing real multipliers. Notice that pared to the magnitude response of a 5th-order comb filter.
the two curves are superimposed.

Finally, Fig. 12 compares the frequency behavior of the pro-

posed filter H rr ( z) employing real multipliers with the frequency
response of the filter embedding quantized multipliers. A funda-
mental observation here is the perfect match between the two
magnitude responses, confirming once again that the quantization
of the multipliers does not affect the noise rejection performance
of the proposed class of decimation filters.

5.2. Design example 2

In this design example, we consider a decimation factor M = 16

and an oversampling factor ρ = 128. This setup accounts for highly
oversampled digital signals. The normalized bandwidth of the sam-
pled signal is f 0 = 21ρ = 0.0039, while the first folding band is:
 
1 1
− f 0, + f 0 = [0.0586, 0.0664].
M M
Since ω = 2π f , we notice in passing that the aforementioned fre- Fig. 14. Magnitude response of the proposed filter with quantized multipliers com-
ω
quencies are mapped to the frequencies π0 = ρ1 = 0.0078. pared to a 5th-order comb filter around the first folding band.

 
2 2
− 2 f 0, + 2 f = [0.1172, 0.1328],
M M
when the x-axis represents the variable ω/π .
Let us discuss the design steps of the proposed decimation fil-
ter.
1. Given M = 16, it is M 1 = 8. The order N = 3 is once again
chosen in order to compare the proposed filter to 5th-order
comb filters.
2. The parameter b belonging to the droop compensation filter is
b = 0 from Table 1.
3. From Table 2, the optimal values of the zeros rotations are
qopt = 0.887, and αopt = 0.02177025534167, thus yielding an
additional noise power rejection of 12.95 dB compared to a
5th-order comb filter.
4. Given Δ = 0.001, it is α1 = αopt + Δ = 0.02212596078600.
From (46), it is k = 6 and I = 63 (from I = 2k − 1).
The magnitude response of the proposed filter H rr ( z) is com-
pared with the one of a 5th-order comb filter in Fig. 13. Notice
the presence of the rotated zeros spanning the first folding band
[0.1172, 0.1328], as clearly emphasized in Fig. 14.
The passband behavior of the magnitude responses of the pro-
posed filter H rr ( z) and the one of a 5th-order CIC filter is depicted
in Fig. 15. Notice that the proposed filter introduces a maximum
droop lower that 0.05 dB, while the 5th-order comb filter presents
a signal distortion as high as 0.28 dB.
Finally, Fig. 16 compares the frequency behavior of the pro-
posed filter H rr ( z) employing real multipliers with the frequency
response of the filter embedding approximated multipliers in the
first folding band.
5.3. Design example 3
Unlike the previous two design examples, the aim of this last
example is to apply the proposed decimation filters to ΣΔ A/D
converters. Upon using Matlab, we have simulated a 2nd-order
ΣΔ A/D converter with a two-level quantizer and a sampling fre-
quency f s = 25.6 kHz. The input signal is a band-limited signal
with bandwidth B x = 100 Hz. With this setup, it is ρ = f s /2B x =
 G.J. Dolecek, M. Laddomada / Digital Signal Processing 22 (2012) 859–868
Fig. 15. Magnitude response of the proposed filter with quantized multipliers com-
pared to a 5th-order comb filter at baseband. The plot also shows the droops
introduced by the two filters in the frequency domain.
Fig. 16. Magnitude response of the proposed filter with quantized multipliers com-
pared to the one embedding real multipliers around the first folding band.
128, while the normalized frequency bandwidth of the input sig-
nal is f 0 = 1/2ρ = 1/256. For decimating oversampled ΣΔ A/D
converters, the parameter N must be chosen in such a way that
N + 2 equals the ΣΔ A/D order plus one [28]. Therefore, for a
2nd-order ΣΔ A/D modulator, it is N + 2 = 2 + 1, or N = 1. The
oversampled signal at the output of the ΣΔ A/D modulator is dec-
imated through the filter in Fig. 8 using M 1 = 16, M = 32. Notice
that this kind of filters can be used to decimate down to 4 times
the Nyquist frequency [28] as for regular comb filters. Therefore,
a more frequency selective filter must be employed after the cas-
cade in order to reduce the sampling rate to the Nyquist frequency
and select the useful signal. This filter is usually a cascade of two
half-band filters each decimating by 2 [1,3]. The power spectrum
of the oversampled digital signal at the output of the ΣΔ A/D
modulator is shown in the upper subplot of Fig. 17. As expected,
this plot shows the presence of the useful signal with bandwidth
f 0 = 1/2ρ = 1/256 shrunk at baseband, and the noise spectrum
moved outside the useful signal bandwidth. The power spectrum
of the decimated signal at the output of the proposed filter in
867
Fig. 17. Power spectrum of the digital signal at the output of a 2nd-order ΣΔ A/D
converter (upper subplot), and power spectrum of the signal decimated by M = 32
with the proposed filter employing N = 1 (lower subplot).
Fig. 8, is shown in the lower subplot of Fig. 17. Notice that the
useful signal bandwidth is now around M f 0 = 32/256 = 0.125, and
the ΣΔ modulator noise has been filtered greatly so that the re-
maining noise is below −50 dB.
6. Conclusion and discussion
This paper proposed novel reprogrammable, multiplierless,
multirate filter architectures, especially tailored to Digital Sam-
ple Rate Conversion in Oversampled Transceivers, Software Radio
Transceivers, and Sigma-Delta Converters. The proposed solution
consists of novel two-stage multiplierless and nonrecursive dec-
imation architectures with improved performance compared to
classical comb filters employed in commercial products. Like comb
filters, the proposed architecture presents a reprogrammable order
N in order to attain the desired specifications for the intended sig-
nals, as well as a reprogrammable decimation factor M. Compared
to comb filters, the proposed filters enjoy a number of advan-
tages. Among them, we mention the improved spurious signal
rejection in the folding bands compared to classical comb filters,
and a reduced passband distortion. When compared to GC Filters,
the proposed filters are multiplierless, present reduced passband
distortion and, additionally, do not have problems of pole-zero can-
cellation.
One key application of the proposed filters is for decimating
oversampled ΣΔ A/D converters. In this scenario, the parameter
N must be chosen in such a way that N + 2 equals the ΣΔ A/D
order plus one.
We have provided a number of design examples to show the
way the proposed filters can be applied to meet the required spec-
ifications in general multirate applications as well as for ΣΔ A/D
converters.
References
[1] R.E. Crochiere, L.R. Rabiner, Multirate Digital Signal Processing, Prentice Hall,
NJ, 1983.
[2] E.B. Hogenauer, An economical class of digital filters for decimation and inter-
polation, IEEE Trans. Acoust. Speech Signal Process. 29 (4) (1981) 155–162.
[3] F.J. Harris, Multirate Signal Processing for Communication Systems, Prentice
Hall, NJ, 2004.
[4] P.P. Vaidyanathan, Multirate Systems and Filter Banks, Prentice Hall, 1992.
[5] F. Mintzer, On half-band, third-band, and Nth-band FIR filters and their design,
IEEE Trans. Acoust. Speech Signal Process. 30 (5) (1982) 734–738.
[6] F. Mintzer, B. Liu, Aliasing error in the design of multirate filters, IEEE Trans.
Acoust. Speech Signal Process. 26 (1) (1978) 76–88.
868 G.J. Dolecek, M. Laddomada / Digital Signal Processing 22 (2012) 859–868
[7] H.G. Gockler, G. Evangelista, A. Groth, Minimal block processing approach to
fractional sample rate conversion, EURASIP Signal Process. 81 (2001) 673–691.
[8] M.W. Coffey, Optimizing multistage decimation and interpolation processing –
part II, IEEE Signal Process. Lett. 14 (1) (2007) 24–26.
[9] M. Laddomada, Design of multistage decimation filters using cyclotomic poly-
nomials: optimization and design issues, IEEE Trans. Circuits Syst. I 55 (7)
(2008) 1977–1987.
[10] M. Laddomada, D. Troncoso, G.J. Dolecek, Design of multiplierless decimation
filters using an extended search of cyclotomic polynomials, IEEE Trans. Circuits
Syst. II 58 (2011) 115–119.
[11] H. Aboushady, et al., Efficient polyphase decomposition of comb decimation
filters in sigma-delta analog-to-digital converters, IEEE Trans. Circuits Syst.
II 48 (10) (2001) 898–903.
[12] T.K. Shahana, et al., Polyphase decomposition of non-recursive comb decima-
tors for sigma-delta A/D converters, in: International IEEE EDSSC Conference,
2007, pp. 825–882.
[13] G.J. Dolecek, S.K. Mitra, Simple method for compensation of CIC decimation
filter, Electron. Lett. 44 (19) (2008) 1162–1163.
[14] W. Kim, et al., Design of CIC roll-off compensation filter in a W-CDMA digital
receiver, Digital Signal Process. 16 (2006) 846–854.
[15] A.Y. Kwentus, Z. Jiang, A.N. Willson Jr., Application of filter sharpening to cas-
caded integrator-comb decimation filters, IEEE Trans. Signal Process. 45 (2)
(1997) 457–467.
[16] G.J. Dolecek, S.K. Mitra, A new two-stage sharpened comb decimator, IEEE
Trans. Circuits Syst. I 52 (7) (2005) 1414–1420.
[17] M. Laddomada, M. Mondin, Decimation schemes for sigma-delta A/D convert-
ers based on Kaiser and Hamming sharpened filters, IEE Proc., Vis. Image Signal
Process. 151 (4) (2004) 287–296.
[18] M. Laddomada, Comb-based decimation filters for sigma-delta A/D convert-
ers: novel schemes and comparisons, IEEE Trans. Signal Process. 55 (5) (2007)
1769–1779.
[19] G.J. Dolecek, F. Harris, Design of wideband compensator filter for a digital IF
receiver, Digital Signal Process. 19 (5) (2009) 827–837.
[20] G.J. Dolecek, S.K. Mitra, Two-stage CIC-based decimator with improved charac-
teristics, IET Signal Process. 4 (2010) 22–29.
[21] L. Lo Presti, Efficient modified-sinc filters for sigma-delta A/D converters, IEEE
Trans. Circuits Syst. II 47 (2000) 1204–1213.
[22] L. Lo Presti, A. Akhdar, Efficient antialiasing decimation filter for sigma-delta
converters, in: Proc. of ICECS98, 1998, pp. 367–370.
[23] M. Laddomada, et al., An efficient decimation sinc-filter design for software
radio applications, in: Proc. of IEEE SPAWC, 2001.
[24] M. Laddomada, Generalized comb decimation filters for sigma-delta A/D con-
verters: analysis and design, IEEE Trans. Circuits Syst. I 54 (2007) 994–1005.
[25] A. Fernandez Vazquez, G.J. Dolecek, A general method to design GCF compen-
sation filter, IEEE Trans. Circuits Syst. II 56 (2009) 409–413.
[26] A. Fernandez Vazquez, G.J. Dolecek, An L2 design of GCF compensation filter,
Signal Process. 91 (5) (2011) 1143–1149.
[27] G.J. Dolecek, M. Laddomada, An economical class of droop-compensated gener-
alized comb filters: analysis and design, IEEE Trans. Circuits Syst. II 57 (2010)
275–279.
[28] R. Schreier, G.C. Temes, Understanding Delta-Sigma Data Converters, 1st ed.,
Wiley–IEEE Press, 2004.
Gordana Jovanovic Dolecek received a BS degree
from the Department of Electrical Engineering, Uni-
versity of Sarajevo, an MSc degree from University of
Belgrade, and a PhD degree from the Faculty of Elec-
trical Engineering, University of Sarajevo. She was a
full professor at the Faculty of Electrical Engineering,
University of Sarajevo until 1993, and in 1993–1995
she was with the Institute Mihailo Pupin, Belgrade. In
1995 she joined Institute INAOE, Department for Elec-
tronics, Puebla, Mexico, where she works as a full professor.
During 2001–2002 and 2006 she was with Department of Electrical &
Computer Engineering, University of California, Santa Barbara, and during
2008–2009 she was with San Diego State University, as visiting scholar.
She was invited to hold lectures and short courses a number of times
across the world.
She is a reviewer for 13 SCI journals and many indexed international
conferences, and she was a member of program committee for several in-
ternational conferences. She was a guest editor of the journal ETRI Signal
Processing, Special Issue on Advanced Techniques on Multirate Signal Pro-
cessing for Digital Information Processing.
She is the author/co-author of four books, the editor of one book, and
the author of 18 book chapters and more than 300 journal and conference
papers.
Her research interests include digital signal processing and digital com-
munications. She is a Senior member of IEEE, the member of Mexican
Academy of Sciences, and the member of National Researcher System (SNI)
Mexico.
Massimiliano Laddomada, Associate Professor of
Electrical Engineering, received a PhD in Communi-
cations and Electronics Engineering from Polytechnic
University of Turin in 2003. He is also an adjunct
professor at California State University, Los Angeles
since 2006. Prior to joining the Texas A&M University–
Texarkana faculty on 2008, he worked as a visiting
assistant professor at Polytechnic University of Turin
(Italy) in 2003–2008, and as a senior engineer at
Technoconcepts, Inc., Los Angeles in 2000–2001.
His main areas of research are digital signal processing and wireless
communications, especially modulation and coding, including turbo codes
and, more recently, network and distributed coding in sensor networks.
Currently, he is an associate editor of IEEE Transactions on Circuits and
Systems – I: Regular Papers and IEEE Communications Surveys and Tuto-
rials.