IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS, VOL. 54, NO.

3, MARCH 2007 643

Closed-Form Design of Digital IIR Integrators

Using Numerical Integration Rules
and Fractional Sample Delays
Chien-Cheng Tseng, Senior Member, IEEE

Abstract—In this paper, the numerical integration rules and digital integrator, the digital differentiator can be obtained. Ref-
fractional sample delays will be used to obtain the closed-form erences [6]–[8] have demonstrated the effectiveness of this idea.
design of infinite-impulse response (IIR) digital integrators. There On the other hand, fractional sample delay has become an
are two types of numerical integration rules to be investigated. One
is Newton–Cotes quadrature rule, the other is Gauss–Legendre important device in the applications of time adjustment in dig-
integration rule. Although the proposed IIR digital integrators ital receiver, antenna array processing, speech coding and syn-
will involve the implementation of fractional sample delays, this thesis, modeling of music instruments and analog–digital (A/D)
problem is easily solved by applying well-documented design tech- conversion, etc. So far, several methods have been presented to
niques of the finite-impulse response Lagrange and IIR allpass design fractional sample delay filters [9]–[17]. In this paper, we
fractional delay filters. Several design examples are illustrated to
demonstrate the effectiveness of the proposed method. will design IIR integrators using numerical integration rules and
fractional sample delay filters. There are two types of numerical
Index Terms—Digital integrator, finite-impulse response (FIR)
integration rules to be studied. One is Newton–Cotes quadrature
filter, fractional sample delay, Gauss–Legendre integration rule,
infinite-impulse response (IIR) filter, Newton–Cotes quadrature (NCQ) rule, the other is Gauss–Legendre integration rule. The
(NCQ) rule. proposed IIR digital integrators will involve the implementa-
tion of fractional sample delay. However, this problem is easily
solved by applying well-documented design techniques of the
I. INTRODUCTION FIR Lagrange and IIR allpass fractional delay filters in [9]. De-
HE DIGITAL integrator is an important device in the areas
T of control, biomedical engineering, and radar [1]–[8]. The
design methods of digital integrator generally can be classi-
sign examples show that the proposed integrators have smaller
approximation errors than the conventional IIR integrators given
in (1)–(3) especially at high-frequency range.
fied into two categories. One is the linear phase finite-impulse This paper is organized as follows. In Section II, the NCQ
response (FIR) filter approach in which the filter coefficients rule is first reviewed briefly. Then, this rule is applied to design
are determined by using maximal flatness constraints [1], [2], several IIR integrators with closed-form transfer function. In
the other is the infinite-impulse response (IIR) filter method Section III, we will describe the conventional Gauss–Legendre
in which the filter coefficients are obtained directly from well- integration rule and use it to get closed-form design of digital
known rectangular, trapezoidal and Simpson methods of numer- IIR integrators. Finally, a conclusion is made.
ical integration [3]–[8]. In the literature, some typical transfer
functions of IIR integrators are given by II. DESIGN BASED ON NCQ RULE
In this section, we first review the NCQ rule which can be
used to estimate the following definite integral numerically:
Rectangular (1)

Trapezoidal (2) (4)

Simpson (3) where interval is the integral region. To achieve this es-
timation, the interval is uniformly partitioned into
subintervals , as shown in Fig. 1.
where is the sampling period. Any the above two integra-
tors can be combined by weighting sum to obtain a new IIR The length of each subinterval is and the
sampling point is equal to . Given the sam-
integrator which has better accuracy at high frequency [3]–[5].
pling data in which we define , then a Newton
Moreover, by inverting the transfer function of minimum phase
polynomial with degree can be found to pass
through these sampling points
Manuscript received September 28, 2004; revised August 29, 2005, and June
5, 2006. This work was supported by the National Science Council, R.O.C.,
under Contract NSC94-2213-E-327-005. This paper was recommended by As-
sociate Editor Y. Lian.
The author is with the Department of Computer and Communication Engi-
neering, National Kaohsiung First University of Science and Technology, Kaoh-
siung 811, Taiwan, R.O.C. (e-mail: [email protected]). (5)
Digital Object Identifier 10.1109/TCSI.2006.887641

1549-8328/$25.00 © 2007 IEEE

644 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS, VOL. 54, NO. 3, MARCH 2007

Now, let us discuss some special cases of NCQ rule next. If we

choose , then the estimation value (2) is given by

(11)

Fig. 1. Interval [a; b] is uniformly partitioned into m-1 subintervals [x ; x ]; Clearly, this is the conventional trapezoidal rule. Similarly,
, and [x
[x ; x ]; . . . ; x ].
when we choose , the estimation value
are given by

where coefficients can be computed by divided difference

table method [18]. If we define the operator as
then coefficients can be written as

(6) (12)

Based on this formula, some typical coefficients are given by Obviously, is the conventional Simpson’s rule,
is the conventional Simpson’s rule, and
is the conventional Boole’s rule. Based on (12), it is clear that
the NCQ rule can be rewritten as the following form:

(7) (13)

Because the polynomial approximates the function For example, when we choose order , the parameters
well in the interval , the definite integral in (4) can and
be estimated as . Moreover, the coefficients satisfy the even symmetric
condition

(14)

and the following equality is also valid:

(8) (15)

So far, the NCQ rule has been described briefly. Now, let us
where is given by
use these results to design digital IIR integrators. One is the de-
sign based on the same sampling period for various quadra-
(9) ture rules , the other is the design based on the dif-
ferent sampling periods for various quadrature
rules . The details are described next.
Recipes for the can also be derived. The following are a
few examples: A. Design Based on the Same Sampling Period for Various
Orders
Given the discrete-time signal with sampling period ,
let us use the composite NCQ rule with to
design digital IIR integrators. In this case, the output of
the integrator is given by

(10) (16)
TSENG: CLOSED-FORM DESIGN OF DIGITAL IIR INTEGRATORS USING NUMERICAL INTEGRATION RULES 645

Taking Z-transform at both sides of (16) and using the sym- Fact 2: The frequency response is almost equal to
metric condition in (14), we have when is small. That is, is an ideal integrator
at low-frequency region.
Proof: When the frequency is small, we
have and
. Thus, (21) can be ap-
proximated by

(22)
(17)

Thus, the transfer functions for quadrature rules are at low-frequency range. Using (14) and (15), we get
given by

(18) (23)

Based on (12) and (18), some typical transfer functions with

are given by Combining (20) and (23), it yields . That
is, is an ideal digital integrator at low-frequency re-
gion for odd-order . Similarly, it is easy to show
at low-frequency region for even-order . The proof
is completed.
Now, let us compare the design errors of integrators
for various orders . To evaluate the performance, the fre-
quency-response error is defined as

Note that the numerator of is even symmetric and the (24)

denominator is odd symmetric. Taking and =1, we
have the frequency responses of the IIR integrators Fig. 2 shows the magnitude responses of the integrators
for and . The dashed line is the ideal
(19) magnitude response . It is clear that the ideal response is
only fitted well at low-frequency region. The errors of
Now, two useful properties of frequency response are integrators for various orders and are given
described below. by
Fact 1: The phase response of integrator is exactly
equal to or , that is, the angle of is
or for all .
Proof: If order is an odd integer, the frequency response
in (19) can be rewritten as (20), shown at the bottom of the page,
(25)
by using symmetric condition in (14), where the function
is given by (21), shown at the bottom of the page. Because From these results, it can be observed that the errors can not
is a real-valued function of , the angle of is be reduced by increasing order . The reason to explain this
for and for . Thus, the result is described as follows. When the NCQ rules
phase response of integrator is exactly equal to or with sampling period are used to design integrators ,
. Similarly, it is easy to show the angle of is the size of integral interval shown in Fig. 1 is equal to
or for even-order . The proof is completed. . So, the size of integral interval is different

(20)

(21)
646 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS, VOL. 54, NO. 3, MARCH 2007

Fig. 2. Magnitude responses jH (e )j for various orders m. The dashed line is the ideal response (1=!). (a) jH (e )j. (b) j H (e )j. (c) j H (e )j. (d)
jH (e )j .

for various orders . The larger is chosen, the larger interval integrators. Given signal , the output of the integrator
is used. Thus, the design errors can not be reduced. How- is given by
ever, it is well known that if the size of integral interval
is fixed, the error of numerical integral can be reduced by in-
creasing order . In next subsection, the size of integral in-
terval to apply composite rule will be fixed to be the sam-
pling period for various orders . That is, the sampling pe-
riod for quadrature rule is rather than (27)
. By using this method, the designed errors can be reduced
by increasing order . Moreover, it is worth mentioning that
Fig. 2 shows that the high-order integrators perform well at the Taking Z-transform on both sides of (27) and using the sym-
low-frequency band , that is, a band below the metric condition in (14), we have
angle of the first zero of their denominator of . To illus-
trate this result, let us choose for various
. The errors in this case are given by

(26)

Clearly, the errors are all small for various orders . And, the (28)
error is reduced by increasing order .
Thus, the transfer functions for quadrature rules are
B. Design Based on Different Sampling Period for Various given by
Orders
Now, let us study how to use the composite Newton–Cotes
(29)
rules with to design digital IIR
TSENG: CLOSED-FORM DESIGN OF DIGITAL IIR INTEGRATORS USING NUMERICAL INTEGRATION RULES 647

Based on (12) and (29), some typical transfer functions with at low-frequency range. Using (14) and (15), it yields
are given by
(34)

Combining (31) and (34), it yields .

That is, is an ideal digital integrator at low-fre-
quency region for odd-order . Similarly, it is easy to show
at low-frequency region for even-order
. The proof is completed.
The Fact 4 can be also shown by the following procedure.
Compared (18) with (29), we have the following equality:

Taking and , we have the frequency responses of (35)

the IIR integrators
Using Fact 2, it yields the result

(30)
(36)
Now, two important properties of frequency response
are described as follows. Although scaling relation in (35) can be used to obtain
Fact 3: The phase response of integrator is exactly from easily, this relation does not provide physical
equal to or , that is, the angle of is meaning given in (27). Moreover, the technique used in (27)
or for all . can be easily extended to design IIR integrator by Gauss–Le-
Proof: If order is an odd integer, the frequency response gendre integration rule described in next section, but scaling
in (30) can be rewritten as (31), shown at the bottom of the page relation in (35) can not achieve the same purpose. Now, let us
by using symmetric condition in (14), where the function compare the design errors of IIR integrators for various
is given by (32), shown at the bottom of the page. Because orders . Fig. 3 shows the magnitude responses of the integra-
is a real-valued function of , the angle of is tors for and . The dashed line is
for and for . Similarly, it the ideal magnitude response . It is clear that the ideal
is easy to show the angle of is or for response is approximated very well, so the solid and dashed
even-order . The proof is completed. lines have overlapped together in Fig. 3(b)–(d). To evaluate
Now, one remark on Fact 3 is made. In practical implementa- performance, the frequency-response error is defined by
tion, the fractional delay elements in are always approx-
imated, therefore, the phase response is approximately equal to (37)
or . Fact 3 is only valid when all fractional delay
The errors of for various orders and
elements in are perfectly realized.
are given by
Fact 4: The frequency response is almost equal to
when is small. That is, is an ideal integrator
at low-frequency region.
Proof: When the frequency is small,
and . This ap-
proximation yields the result (38)
From these results, it is clear that the errors can be reduced by
(33) increasing order . Compared (25) with (38), it can be seen that
the design errors of are smaller than those of for

(31)

(32)
648 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS, VOL. 54, NO. 3, MARCH 2007

Fig. 3. Magnitude responses j F (e )j for various orders m. The dashed line is the ideal response (1=! ). (a) j F (e )j. (b) j F (e )j. (c) j F (e )j .
(d) jF (e )j.

. This is because the size of integral interval in Fig. 1 methods are presented. One concerns FIR design case, the other
is for and for . Moreover, compared is IIR allpass filter case.
(26) with (38), we have . This Case 1) FIR filter case
result is easy to show by using relation in (35). Although the In this case, we use the Lagrange interpolation
improved design of integrators can be obtained, there are frac- method to design an FIR filter for approximating a
tional delay elements needed to be implemented in the transfer given fractional delay [9]. In this method,
function for . Fortunately, this problem is easily the delay is approximated by
solved by applying well-documented design techniques of the
FIR Lagrange and IIR allpass fractional delay filters in [9]. In (40)
next subsection, the implementation issue of transfer function
will be studied in details. where the filter coefficients can be computed
by the following formula:
C. Implementation of IIR Integrator
To apply the tools in [9] to design fractional sample delays in (41)
integrator, a pure integer delay is cascaded with the inte-
grator . So, we have Based on this fractional delay filter, there are two
methods to implement IIR integrators . One
is filter bank approach, the other is conventional IIR
filter method. The details are described below.
(39) Without losing generality, let us use integrator
as an example to study the filter bank implementation
Thus, the problem is reduced to design the fractional delay ele- method. The transfer function of is given by
ments with . In the following, two (42), shown at the bottom of the page. Fig. 4 shows

(42)
TSENG: CLOSED-FORM DESIGN OF DIGITAL IIR INTEGRATORS USING NUMERICAL INTEGRATION RULES 649

ement) is high. So, another implementation method

with low complexity needs to be developed. In what
follows, the other realization method will be pro-
posed to achieve this purpose.
In the second implementation method, the Lagrange
FIR fractional delay filter in (40) is sub-
stituted into IIR integrator in (39) to get the
combined transfer function

(44)
Fig. 4. Filter bank implementation structure for IIR integrator U (z ).
where coefficients in the numerator are given by
the filter bank implementation of . Clearly,
there are three fractional delay elements in this (45)
structure. After replacing three fractional delays in
the structure with FIR filters in (40), the structure Clearly, the transfer function of in (44) is a
is only composed of integer delays. So, it is easy conventional IIR filter with integer delays. Thus, the
to be implemented. Now, let us study the perfor- direct-form or lattice-form realizations in the text-
mance of this approach. When the parameters are book [19] can be used to implement it. The output
chosen as and , the group of digital integrator is delayed approximately
delay responses of the designed FIR filters with by , where is the order of the FIR
, and are shown in Fig. 5(a). filter used to produce the fractional delayed samples.
Clearly, the responses fit ideal delay values 5.25, Now, let us use integrator as a numerical ex-
5.5, and 5.75 well except at high-frequency re- ample to study this implementation method. When
gion. This is because the Lagrange FIR fractional the parameters are chosen as and ,
delay filter is a maximally flat delay at the zero three Lagrange FIR fractional delay filters are given
frequency . Inserting these three FIR filters by
into the structure in Fig. 4, the magnitude response
(solid line) of the designed integrator is
depicted in Fig. 5(b), where the dashed line is ideal
response . Because two lines almost overlap
together except at high-frequency region, the ideal
response is approximated well by the proposed (46)
design method. Now, let us study how to determine Substituting the above three filters into the following
order of FIR fractional delay filter such that integrator:
the design error of integrator is as small as
possible. The criterion used here is to find order
to minimize the frequency-response error
(47)
(43) we get the final transfer function
Fig. 6 shows the error curves for various or- (48)
ders and . The integers in
Fig. 6(a) and (b) is chosen as 5 and 10. It is clear that Fig. 7(a) shows the magnitude response of the above
the error curve has minimum value at . integrator for . The dashed line is the
Thus, due to symmetry reasons, when delay is ideal magnitude response . It is clear that the
specified, it is suggested to choose ideal response is fitted well. Fig. 7(b) shows the
for odd and for even. Although absolute magnitude error between ideal and actual
the above filter bank method can be used to im- responses. It is obvious that the error is very small at
plement IIR integrator , its implementation low-frequency range. The frequency-response error
complexity (number of multiplier and unit delay el- of is 0.0686 on frequency band .
650 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS, VOL. 54, NO. 3, MARCH 2007

Fig. 5. (a) Group delay responses of the designed Lagrange FIR filters with L = 5; M = 10; d = 1=4; 1=2 and 3=4. (b) Magnitude response jU (e )j of the
designed integrator. The dashed line is the ideal response (1=! ).

Fig. 6. Error curves E (5) for various orders M; :

= 0 95 and T = 1. (a) L = 5 (b) L = 10.

For comparison, the frequency-response error of over, the order of denominator of in (48) is
Simpson integrator in (3) is also computed also smaller than that of denominator of , so
for and frequency band . As a re- the implementation complexity (number of mul-
sult, the error in the Simpson case is 0.3275. Thus, tiplier and unit delay) of is less than that
the error of proposed integrator is smaller of Simpson integrator . Thus, the integrator
than that of conventional Simpson integrator. More- is better than by considering accuracy
TSENG: CLOSED-FORM DESIGN OF DIGITAL IIR INTEGRATORS USING NUMERICAL INTEGRATION RULES 651

Fig. 7. (a) Magnitude response of the integrator U (z ) for T = 1. Dashed line is the ideal magnitude response (1=! ). (b) Absolute magnitude error between
ideal and actual responses.

and complexity. Finally, let us compare integrator pass filters into the transfer function in (42),
with trapezoidal integrator in (2) the magnitude response (solid line) of re-
whose frequency-response error on band sultant integrator is shown in Fig. 8(b). The ideal re-
is . As a result, the error 0.0686 of is sponse (dashed line) is also included. Because two
smaller than the error 0.0755 of . However, lines almost overlap together, the ideal response is
integrator has higher numerator implementa- approximated well by proposed integrator. Note that
tion complexity than integrator . the design error can be reduced by increasing filter
Case 2) Allpass filter case order . However, the complexity of implementa-
In this case, we use the maximally flat group delay tion will also increase. Moreover, the filter bank ap-
allpass filter to approximate the given fractional proach and conventional method in FIR filter case
delay [9]. The transfer function chosen to can be easily used to implement the integrators in
approximate is given by allpass filter case.
Finally, two remarks are made. First, because integrator
is an IIR filter with single pole , initial value
must be specified in advance when is used to
(49) filter the input signal . If and , then
The above filter coefficients are given by it can be shown that output signal for all .
In order to let integrator only compute the integral of
the input signal without being affected by initial value
(50) , it is suggested to set initial value . The above
remark on initial condition applies only for the integrators
where is a binomial co- designed with fractional delays implemented by FIR filters.
efficient. Now, let us use integrator in (42) to When we use allpass filters to implement fractional delays, then
demonstrate the performance of allpass filter case. more zero initial values need to be specified. Second, when the
When we choose , Fig. 8(a) shows the group Newton–Cotes rule is applied to design IIR integrators. The
delay responses of the designed allpass filters with fractional sample delays always distribute uniformly because
and . Clearly, the responses fit of uniform partition of interval. Thus, if nonuniform fractional
ideal delay values 5.25, 5.5, and 5.75 very well at sample delays want to be used, the NCQ rule can not be used
low-frequency range. This is because allpass frac- to achieve the purpose. In the next section, the Gauss–Le-
tional delay filter is a maximally flat group delay gendre integration rule will be described to get nonuniformly
design at frequency . Inserting these three all- distributed fractional sample delays.
652 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS, VOL. 54, NO. 3, MARCH 2007

Fig. 8. (a) Group delay responses of the designed IIR allpass filters with L = 5; d = 1=4; 1=2 and 3=4. (b) Magnitude response jU (e )j of the designed
integrator. The dashed line is the ideal response (1=! ).

III. DESIGN BASED ON GAUSS–LEGENDRE INTEGRATION RULE For the three-point case, i.e., , the abscissas and weights
are given by
In this section, we first review the Gauss–Legendre integra-
tion rule which can be used to estimate the following definite
integral numerically:

(51) (55)

For the four-point case, i.e., , the abscissas and weights

The -point Gauss–Legendre approximation formula for this
are given by
definite integral is given by

(52)

where abscissas and weights must satisfy the fol- (56)

lowing constraints:
After studying the error analysis [21], it can be shown that the
following expression is valid:
(53)

(57)
for . By solving nonlinear equations
in (53), the abscissas and weights can be obtained
It is not immediately obvious how rapid this decays, be-
and tabulated. Reference [20, Table 7.9] gives the values up to
cause of the competing effects of the factorials in the numerator
. Now, three typical cases are listed next. For two-point
and denominator. Thus, the Stiring’s formula is used to obtain
Gauss–Legendre rule, i.e., , the abscissas and weights are
the estimated value of

(54) (58)
TSENG: CLOSED-FORM DESIGN OF DIGITAL IIR INTEGRATORS USING NUMERICAL INTEGRATION RULES 653

Fig. 9. (a) Magnitude response jG (e )j (b) Magnitude response j G (e )j. The dashed line is the ideal response (1=! ).

where and lies between 0.7 to 1.04 [21]. When and , the transfer function can be written as
Clearly, this estimate says that the error goes down exponen- follows:
tially with . So far, Gauss–Legendre rule has been reviewed
briefly. (62)
Now, let us use Gauss–Legendre integration rule to design
IIR digital integrators. Given the discrete-time signal with (63)
sampling period the output of the Gauss–Legendre in-
tegrator is given by

(64)
(59)
Now, let us study the frequency response of Gauss–Legendre in-
tegrator. For simplicity, we only study the case of . Taking
Taking the Z-transform on both sides of (59), we have and , we have the frequency response of the
Gauss–Legendre integrator

(65)

When the frequency is very small, we have and

(60) . This approximation yields the result

Thus, the transfer function of the proposed -point Gauss–Le-

gendre integrator is given by at low-frequency range. That is, is an ideal digital in-
tegrator at low-frequency region. Fig. 9(a) and (b) depicts the
magnitude responses and of the above de-
(61) signed Gauss–Legendre integrators with . The dashed line
is the ideal response of integrator. Compared Fig. 9(b)
654 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS, VOL. 54, NO. 3, MARCH 2007

 (e
Fig. 10. (a) Group delay responses of the designed Lagrange FIR filters. The ideal delay values are 4.2254 and 5.7746. (b) Magnitude response jG )j of the
designed Gauss–Legendre integrator. The dashed line is the ideal response (1=! ).

with Fig. 2(b), it is clear that the proposed Gauss–Legendre inte- use as an example to illustrate this design method. The
grator has smaller approximation error than the conven- is given by
tional Simpson integrator at high-frequency range.
In order to compare the performance of the above Gauss–Le-
gendre integrators, let us define the error function of frequency
response (69)

(66) When the FIR Lagrange filter in (40)–(41) with and

is used to design two fractional sample delays in
(69), the group delay responses of the designed FIR filters are
The errors for and are given by
shown in Fig. 10(a). Clearly, the responses fit ideal delay values
and well.
Inserting these two FIR filters into the transfer function
(67) in (69), the magnitude response (solid line) of the
designed Gauss–Legendre integrator is depicted in Fig. 10(b),
From this result, it is clear that the larger is, the smaller ap- where the dashed line is ideal response . Because two
proximation error integrator . Also, the error decay rate lines almost overlap together at low-frequency region, the ideal
is very fast. However, the implementation of involves response is approximated well by proposed design method.
the fractional delay elements. To apply the tools in [9] to design Finally, one remark is made. A better Gauss–Legendre inte-
these fractional sample delays, a pure integer delay is cas- grator can be obtained by doubling the sampling rate, i.e., re-
caded with the Gauss–Legendre integrator in (61). So, ducing sampling period from to . In this case, the transfer
we have function of new Gauss–Legendre integrator is given by

(68)

Thus, the problem is reduced to design the fractional delay el-

ements which are easily solved by using FIR La- (70)
grange and IIR allpass fractional delay filters in [9]. Now, let us
TSENG: CLOSED-FORM DESIGN OF DIGITAL IIR INTEGRATORS USING NUMERICAL INTEGRATION RULES 655

When and , the transfer function can be written as [8] M. A. Al-Alaoui, “A class of second-order integrators and lowpass dif-
follows: ferentiators,” IEEE Trans. Circuits Syst. I, Fundam. Theory App., vol.
42, no. 4, pp. 220–223, Apr. 1995.
[9] T. I. Laakso, V. Valimaki, M. Karjalainen, and U. K. Laine, “Split-
(71) ting the unit delay: Tool for fractional delay filter design,” IEEE Signal
Process. Mag., vol. 44, no. 1, pp. 30–60, Jan. 1996.
[10] W. S. Lu and T. B. Deng, “An improved weighted least-squares de-
sign for variable fractional delay FIR filters,” IEEE Trans. Circuits Syst.
II, Analog Digit. Signal Process., vol. 46, no. 8, pp. 1035–1040, Aug.
1999.
[11] T. B. Deng and W. S. Lu, “Weighted least-squares method for de-
signing variable fractional delay 2-D FIR filters,” IEEE Trans. Circuits
(72) Syst. II, Analog Digit. Signal Process., vol. 47, no. 2, pp. 114–124, Feb.
2000.
To evaluate the performance, the error in (66) is computed [12] J. Vesma and T. Saramaki, “Design and properties of polynomial-based
after is replaced by . When we choose fractional delay filters,” in Proc. Int. Symp. Circuits Syst., May 2000,
vol. 1, pp. 104–107.
, the errors for are given by [13] T. B. Deng, “Discretization-free design of variable fractional-delay FIR
filters,” IEEE Trans. Circuits Syst. II, Analog Digit. Signal Process.,
vol. 48, no. 6, pp. 637–647, Jun. 2001.
(73) [14] C. C. Tseng, “Design of 1-D and 2-D variable fractional delay allpass
filters using weighted least-squares method,” IEEE Trans. Circuits Syst.
Comparing (67) and (73), it is clear that the frequency-response I, Fundam. Theory App., vol. 49, no. 10, pp. 1413–1422, Oct. 2002.
error can be reduced by using double sampling technique. [15] C. C. Tseng, “Eigenfilter approach for the design of variable fractional
delay FIR and allpass filters,” in Proc. IEE Vis Image Signal Process,
Oct. 2002, vol. 149, pp. 297–303.
IV. CONCLUSION [16] T. B. Deng, “Closed-form design and efficient implementation of
In this paper, the IIR digital integrators have been designed by variable digital filters with simultaneously tunable magnitude and
fractional-delay,” IEEE Trans. Signal Process., vol. 52, no. 6, pp.
using numerical integration rules and fractional sample delays. 1668–1681, Jun. 2004.
Two types of numerical integration rules have been investigated. [17] T. B. Deng and Y. Nakagawa, “SVD-based design and new struc-
One is NCQ rule, the other is Gauss–Legendre integration rule. tures for variable fractional-delay digital filters,” IEEE Trans. Signal
Process., vol. 52, no. 9, pp. 2513–2527, Sep. 2004.
Although the proposed IIR digital integrators will involve the [18] C. F. Van Loan, Introduction to Scientific Computing, 2nd ed. Engle-
implementation of fractional sample delays, it is easily solved wood Cliffs, NJ: Prentice-Hall, 2000.
by applying well-documented design techniques of the FIR La- [19] A. V. Oppenheim and R. W. Schafer, Discrete-Time Signal Pro-
cessing. Englewood Cliffs, NJ: Prentice-Hall, 1989.
grange and IIR allpass fractional delay filters. Several design [20] J. H. Mathews and K. D. Fink, Numerical Methods Using MATLAB,
examples are also illustrated to demonstrate the effectiveness 4th ed. Englewood Cliffs, NJ: Prentice-Hall, 2004.
of the proposed method. However, only NCQ rule and Gauss– [21] J. F. Epperson, An Introduction to Numerical Methods and Analysis.
New York: Wiley, 2002.
Legendre integration rule are studied here. Thus, it is interesting
to apply other integration rules to design digital integrators in
the future. Chien-Cheng Tseng (S’90–M’95–SM’01) was born
in Taipei, Taiwan, R.O.C., in 1965. He received the
REFERENCES B.S. degree (with honors) from Tatung Institute of
[1] B. Kumar, D. R. Choudhury, and A. Kumar, “On the design of linear Technology, Taipei, Taiwan, R.O.C., in 1988, and the
phase FIR integrators for midband frequencies,” IEEE Trans. Signal M.S. and Ph.D. degrees from the National Taiwan
Process., vol. 44, no. 10, pp. 2391–2395, Oct. 1996. University, Taipei,Taiwan, R.O.C., in 1990 and 1994,
[2] B. Kumar and A. Kumar, “FIR linear-phase approximations of fre- respectively, all in electrical engineering.
quency response 1=(j! ) for maximal flatness at an arbitrary frequency From 1995 to 1997, he was an Associate Re-
! ; (0 < ! <  ),” IEEE Trans. Signal Process., vol. 47, no. 6, pp. search Engineer at Telecommunication Laboratories,
1772–1775, Jun. 1999. Chunghwa Telecom Company, Ltd., Taoyuan,
[3] M. A. Al-Alaoui, “Novel digital integrator and differentiator,” Electron. Taiwan, R.O.C. He is currently a Professor and
Lett., vol. 29, pp. 376–378, Feb. 1993. Chairman of the Department of Computer and Communication Engineering,
[4] J. Le Bihan, “Novel class of digital integrators and differentiators,” National Kaohsiung First University of Science and Technology, Kaohsiung,
Electron. Lett., vol. 29, pp. 971–973, May 1993. Taiwan, R.O.C. His research interests include digital signal processing, pattern
[5] N. Papamarkos and C. Chamzas, “A new approach for the design of recognition, quantum computation. He has over 100 published journal and
digital integrators,” IEEE Trans. Circuits Syst. I, Fundam. Theory App., conference papers in these fields.
vol. 43, no. 9, pp. 785–791, Sep. 1996. Dr. Tseng is a member of the Chinese Image Processing and Pattern Recog-
[6] M. A. Al-Alaoui, “Novel approach to designing digital differentiators,” nition (IPPR) Society and Color Association of Taiwan. He was the recipient of
Electron. Lett., vol. 28, pp. 1376–1378, Jul. 1992. research awards from the National Science Council in 1998, 1999, and 2001,
[7] M. A. Al-Alaoui, “Novel IIR differentiator from the Simpson integra- the Outstanding Paper Award from the IPPR Society of Taiwan in 1993, and the
tion rule,” IEEE Trans. Circuits Syst. I, Fundam. Theory App., vol. 41, Best Paper Award from the Symposium of Telecommunications held in Taiwan
no. 2, pp. 186–187, Feb. 1994. 1994.