Sei sulla pagina 1di 5

158 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—II: EXPRESS BRIEFS, VOL. 57, NO.

3, MARCH 2010

Systematic Design Centering of Continuous


Time Oversampling Converters
Shanthi Pavan, Member, IEEE

Abstract—We address the practical problem of determining the


loop filter component values in a single-loop continuous-time delta
sigma modulator. Conventional techniques to design center the
converter to achieve a desired noise transfer function are cumber-
some and not numerically stable. We present a robust procedure
that can be used to determine the loop filter coefficients when real
opamps (with finite gain, arbitrary Digital to Analog Converter
(DAC) pulse, and multiple internal poles/zeros) are used. The
method can also account for excess loop delay. We illustrate our
technique with second-order low-pass and fourth-order bandpass
examples. Fig. 1. Block diagram of CTDSM with excess delay. The direct path with gain
k0 is intended to compensate for excess delay and finite-bandwidth effects in
Index Terms—A/D, analog to digital, compensation, continuous the opamps.
time, conversion, gain bandwidth, sigma-delta.

I. I NTRODUCTION
ONTINUOUS-TIME oversampling converters are be-
C coming increasingly important in analog signal process-
ing due to several attractive features like implicit anti-aliasing,
resistive input impedance, and low-power operation. The block
diagram of a continuous-time delta sigma modulator (CTDSM)
is shown in Fig. 1—it consists of a continuous-time loop filter,
whose output is sampled, quantized, and fed back through a
DAC, whose pulse shape is denoted by p(t). The excess loop
delay is τ , and is compensated by an additional DAC whose
strength is denoted by k0 . Given a noise transfer function (NTF)
to be realized, the transfer function of the continuous-time filter
(Lc (s)) is derived in the following manner [1]–[3], as explained Fig. 2. Determining CTDSM coefficients using the open-loop fitting ap-
proach. (a) Prototype discrete-time filter. (b) Continuous-time loop filter excited
using Fig. 2. Throughout this brief, we assume (without loss of by a DAC pulse (including excess delay). (c) Second-order example.
generality) that the sampling rate is 1 Hz.
The transfer function L(z) of the prototype discrete-time Approach 1: The desired loop filter transfer function is
modulator is first determined from the desired NTF using given by
L(z) = (1 − NTF(z))/NTF(z). Next, the impulse invariance
transformation (incorporating the pulse shape of the feedback z −1 (2 − z −1 )
L(z) = . (1)
DAC and excess delay τ ) is used to determine the transfer (1 − z −1 )2
function Lc (s) so that the samples of the continuous-time filter
The integrators are converted into their z-domain equivalents
coincide with those of the impulse response of L(z). Several ap-
using the impulse invariance transformation corresponding to
proaches have been proposed to accomplish this, as illustrated
a DAC pulse delayed by τ (we assume that τ < 1). From
with the second-order example shown in Fig. 2(c)—where a
[4, Tab. III], we obtain
cascade of integrators with feedforward (CIFF) loop filter with
a nonreturn-to-zero (NRZ) DAC is shown. Assuming that the 1 1−τ τ
desired NTF is (1 − z −1 )2 , the problem is to determine k0 , k1 , → + z −1 (2)
s z−1 z−1
and k2 .
1 (0.5 − τ + 0.5τ 2 )z + 0.5(1 − τ 2 )

s2 (z − 1)2
Manuscript received September 24, 2009; revised November 29, 2009.
Current version published March 17, 2010. This paper was recommended by τ (1 − 0.5τ )z + 0.5τ 2
Associate Editor P. Malcovati. + z −1 (3)
The author is with the Electrical Engineering Department, Indian Institute (z − 1)2
of Technology (IIT) Madras, Chennai 600 036, India (e-mail: shanthi@ee.iitm.
ac.in). k0 z −1 + k1 × (RHS of (2)) + k2 × (RHS of (3))
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
2z − 1
= (4)
Digital Object Identifier 10.1109/TCSII.2010.2041814 (z − 1)2

1549-7747/$26.00 © 2010 IEEE

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:52 UTC from IE Xplore. Restricon aply.
PAVAN: SYSTEMATIC DESIGN CENTERING OF CONTINUOUS TIME OVERSAMPLING CONVERTERS 159

where RHS denotes right-hand side. Going through the algebra


results in

{k0 , k1 , k2 } = {1.5τ + 0.5τ 2 , 1.5 + τ, 1}. (5)

For a different DAC pulse, the overall process remains the


same—only (2) and (3) will change. The algebra is already
tedious in a second-order example and gets even more so
for higher-order NTFs. Apart from this, the method becomes
impractical in a real design due to the following. The trans-
fer function of a practical integrator will have several extra
poles/zeros due to finite gain and bandwidth effects in the
opamp (assuming active-RC techniques are used to synthesize
the loop filter). Obtaining the locations of these poles/zeros
from circuit simulations is not a straightforward task, so some
authors suggest fitting the step response of the integrator to
a two-pole one-zero model [5]. After this, z-domain models
along the lines of (2) and (3) have to be derived and incorpo-
rated into (4) to determine k0 , k1 , and k2 . Equation (4) will not
have a solution when the integrators are nonideal since the poles
of L(z) are different from the poles of the integrator paths. For
example, if the integrators have finite gain, the poles of the real
loop filter will no longer be at z = 1, and (4) cannot be solved.
Approach 2: Alternatively, k0 , k1 , and k2 may be determined
numerically as follows. Let l[n] denote a column vector of N
samples of the impulse response of L(z). In our example, it is
given by l[n] = [ 0 2 3 · · · ]T .
The column vectors formed by N samples of the pulse re-
sponses of the direct path and integrator outputs are denoted as Fig. 3. (a) Second-order CTDSM using a CIFF loop filter—the component
values correspond to a dynamic range scaled prototype when ideal opamps are
l0 [n] =[ 0 1 0 · · · 0 ]T used. (b) Magnitudes of the NTF: the ideal NTF, the untuned NTF with real
opamps, and the NTFs with coefficients tuned as per (7) for N = 5, 15, 25.
l1 [n] = [ 0 (1 − τ ) 1 · · · 1 ]T
l2 [n] = [ 0 0.5(1−τ )2 (1.5−τ ) · · · (N −0.5−τ ) ]T . (6) (second-order) and bandpass (fourth-order) CTDSM examples.
Section IV concludes this brief.
N is chosen so that it is much larger than the number
of unknowns to be determined. We see that the weighting II. P ROBLEMS W ITH N UMERICAL C OEFFICIENT T UNING
coefficients K = [ k0 k1 k2 ]T can be determined by solving
The issues with the technique based on (7) are illustrated
[ l0 [n] l1 [n] l2 [n] ] K = l[n]. (7) below with a second-order modulator employing a CIFF loop
filter [Fig. 3(a)]. The excess delay in the loop is assumed to
Although there are more equations than unknowns, with ideal be τ = 0.25. The desired NTF is (1 − z −1 )2 . The first step
integrators, the above set of equations admits the unique solu- of the design is to determine K with ideal integrators, after
tion of (5). In other words, K is independent of N . which the summing coefficients and integrators are scaled for
Equation (7) is the time-domain equivalent of (4), but is dynamic range. The resulting component values are shown in
easier to use in practice, since it does away with tedious algebra. Fig. 3(a). The coefficient corresponding to the excess delay
Determining K using (7) remains just as simple even when real compensating DAC (DAC1 ) is denoted by k0 . The magnitude
integrators are used, since only the sampled pulse responses of the NTF is shown in part (b) of the figure. Then, the ideal
at the integrator outputs are needed. These are easily obtained opamps are replaced by their real counterparts. In this brief,
from simulation and naturally include the effects of nonideal we assume that the same feedforward-compensated opamp is
integrators and excess loop delay, as well as second-order used in all the integrators. The summing amplifier and the NRZ
effects like loading of the integrator by subsequent stages and DACs are assumed to be ideal. The topology and the parameters
finite rise/fall times of the DAC pulse. of the opamp (all normalized to an integrating resistor of 1 Ω
Unfortunately, the numerical method above has problems and sampling rate of 1 Hz) are shown in Fig. 4. The parasitic
when applied to a real CTDSM (with nonideal integrators). capacitors are somewhat exaggerated. As expected, replacing
In this brief, we determine the cause and provide a solution the ideal opamps with real ones results in significant peaking
that circumvents these difficulties. It is organized as follows. in the NTF (with a maximum gain of about 7), as shown in
In Section II, we illustrate the issues with coefficient tuning in Fig. 3(b). To tune the coefficients, the pulse responses at the
a practical second-order CTDSM. We show that tuning based direct path and integrator outputs (l0 , l1 , and l2 ) were found by
on (7) is not reliable, and explain the reasons for this. We exciting the loop filter in an open-loop fashion with an NRZ
present a technique that mitigates these problems in Section III, pulse. N samples of these responses (with N = 5, 15, and 25)
and demonstrate the efficacy of our technique with low-pass were used in (7), and K was determined using a least-squares

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:52 UTC from IE Xplore. Restricon aply.
160 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—II: EXPRESS BRIEFS, VOL. 57, NO. 3, MARCH 2010

Fig. 4. Topology of the two-stage feedforward-compensated opamp used in


CTDSM.
TABLE I
M ODULATOR C OEFFICIENTS O BTAINED BY S OLVING (7)
FOR D IFFERENT VALUES OF N

Fig. 5. Magnitudes of the NTF: the ideal NTF, the untuned NTF with real
TABLE II opamps, and the NTFs with coefficients tuned as per (10) for N = 5, 15, 25.
M ODULATOR C OEFFICIENTS O BTAINED BY S OLVING (10) The inset compares the in-band behavior of the tuned NTF with that of the ideal
FOR D IFFERENT VALUES OF N NTF—below (ω/π) ≈ 0.005), the tuned NTF exhibits first-order behavior due
to finite integrator gain.

is seen that they remain virtually unchanged. Fig. 5 shows the


magnitudes of the NTF calculated with coefficients obtained
using the proposed technique for N = 5, 15, and 25. They are
approach. The coefficients and the magnitudes of the NTF when almost indistinguishable. The magnitude of the ideal NTF is
(7) was solved with 5, 15, and 25 samples of li [n] are shown shown for comparison, indicating that the proposed technique
in Table I and Fig. 3(b), respectively. From these results, it is does a good job of approximating the desired NTF. A concern
seen that the coefficients change with N , causing the resulting with this approach is the resulting behavior of the NTF at in-
NTF to vary significantly. This puts the designer in a quandary band frequencies (after coefficient tuning). To see that this is
over the values of the coefficients that need to be used in not an issue, one can interpret (10) as choosing K to minimize
the modulator. For instance, using K obtained for N = 25 is the L2 norm of the difference between the frequency responses
clearly not an appropriate choice, as seen from the catastrophic of the two sides of the equation. The magnitude of δ[n] − h[n]
behavior of the NTF. is 1 at dc and becomes 3 at ω = π. (In general, it will be
|NTF(ejπ )| − 1.) At low frequencies, (ω ≈ 0), |H0 (ejω )| ∝
ω 2 , |H1 (ejω )| ∝ ω, and |H2 (ejω )| ∝ ω 0 . It is thus seen that
III. P ROPOSED T ECHNIQUE the largest term in the frequency response of the left-hand
The idea behind the proposed technique is to attempt to fit side of (10) (at low frequencies) is contributed by k2 h2 . The
NTF(z)(1 + L(z)) to unity rather than the loop filter response, least-square solution will therefore attempt to choose k2 so
as indicated below. The NTF and the loop filter transfer function that k2 |H2 (ejω )| approximates unity at low frequencies—to
are related by the “best” extent possible, given that the integrators have finite
gain. The inset in Fig. 5 compares the in-band behavior of the
1 tuned NTF with that of the ideal NTF (whose slope must be
NTF(z) = . (8)
1 + L(z) 40 dB/dec). Below (ω/π) ≈ 0.005), the tuned NTF exhibits
first-order behavior due to integrator finite gain. When go1 of
In the discussion to follow, we denote the impulse response
the opamp was decreased by a factor of 10, the resulting in-
corresponding to NTF(z) by h[n]. Further, let h0 [n] = l0 [n] ∗
band behavior of the tuned NTFs (shown in grey, in the inset)
h[n], h1 [n] = l1 [n] ∗ h[n], and h2 [n] = l2 [n] ∗ h[n], where
follows the ideal behavior from (ω/π) ≈ 0.0005), confirming
li [n] represents the sampled pulse response of the ith integrator
the intuition above.
in the continuous-time loop filter. Equation (8) can be recast in
the time domain as
A. Discussion
h[n] + h[n] ∗ l[n] = δ[n]. (9)
We now give intuition for the poor numerical stability of
Since l[n] = [l0 [n]1 [n]2 [n]]K, the equation above can be the technique in Section I and the improved performance of
written as the proposed method. Ideally, the loop filter of an mth-order
modulator has m poles on the unit circle, whereas the NTF
[h0 h1 h2 ]K = δ[n] − h[n]. (10) should have m poles inside the circle. Fig. 6(a) illustrates this
in a third-order example. The NTF poles are marked by ×, and
The set of equations in (10) can be solved to determine K. The the poles of the loop filter are marked by ⊗. If the opamps were
coefficients obtained for different N are shown in Table II. It ideal, and there was no excess delay, then the only poles (of

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:52 UTC from IE Xplore. Restricon aply.
PAVAN: SYSTEMATIC DESIGN CENTERING OF CONTINUOUS TIME OVERSAMPLING CONVERTERS 161

Fig. 6. (a) Pole locations of the NTF (×) and the “discretized” loop filter (⊗).
(b) Effect of perturbing a pole of L(z) on an NTF pole.

the NTF and L) would be those enclosed in the region labeled Fig. 7. l2 [n] and h2 [n] for ideal and real opamps. Although l2 [n] changes
“ideal.” Finite gain and bandwidth effects in the opamps result significantly due to opamp nonidealities, h2 [n] changes very little.
in extra poles in each integrator and cause those poles of L that
were on the unit circle to move from their ideal locations by
small amounts. For well-designed integrators, the extra poles
of L(z) due to opamp bandwidth limitations lie well within the
unit circle, and it is possible to mitigate the effect of these poles
(and excess loop delay of less than one clock cycle) by adding
an additional coefficient in the form of a direct path across the
loop filter, as shown in Fig. 2(b). Since the modulator is now
a high-order system, one must not expect to be able to achieve
the ideal NTF. The aim of the design centering process is to
determine the coefficients so as to restore the magnitude of the
NTF to something that bears as close a resemblance to the ideal
as possible.
Let z1 denote one of the poles of L(z) near the unit circle,
as shown in Fig. 6(b). Further, let p1 be a pole of the NTF.
Consider the effect of a small change in z1 on p1 . L(z) can be Fig. 8. Diagram of the fourth-order bandpass CTDSM. The intended NTF is
expressed as L1 (z)/(z − z1 ), and p1 satisfies (1 + z −2 )2 .

L1 (p1 ) + p1 − z1 = 0. (11) h0 , h1 , and h2 in (10) are less sensitive to changes in the pole
positions of L(z) due to the following. For simplicity, consider
If z1 is perturbed by a small amount ∆z1 , then the change in p1 an NTF with all its zeros at z = 1. If the integrators were ideal,
is given by then hi [n] = li [n] ∗ h[n] would have a finite impulse response,
   since the zeros of the NTF would cancel the poles of Li (z). If
dL1  the locations of those poles of Li (z) near the unit circle were
∆p1 = ∆z1 1 − . (12)
dz p1 perturbed by ∆z, then the pole-zero cancellation is not exact,
but the change in hi is negligible (although the effect on li is
Since L1 (z) is analytic at p1 , we see that small changes in the dramatic). Fig. 7 shows l2 and h2 in the second-order example
location of z1 cause small changes in the NTF pole locations. of Fig. 3 for two cases: one where the opamp gain is infinite, and
This means that the coefficients of the modulator (K) should another where it is 35. Although there is a significant difference
remain substantially the same if z1 was perturbed by a very in l2 , h2 is virtually unchanged.
small amount. On the other hand, K computed from (7) varies
significantly with ∆z1 , as seen from the following. l0 [n], l1 [n],
B. Fourth-Order Bandpass CTDSM Example
and l2 [n] are very sensitive to the position of those poles of L(z)
that are close to the unit circle. For instance, if the integrators As a final example, we apply our technique to the design
were ideal, l2 [n] ∝ n for large n, whereas finite-gain integrators of a fourth-order bandpass CTDSM—the desired NTF is (1 +
result in l2 [n] → 0 (for large n). Since the least-square solution z −2 )2 . The modulator block diagram is shown in Fig. 8. The
of (7) minimizes the norm of [ l0 l1 l2 ]K − l, and the error sampling rate is 1 Hz. Ideally, the loop filter has repeated poles
in l2 due to finite integrator gain increases greatly with n, the at ±j(π/2) radians. The opamps are feedforward compensated,
coefficient k2 increases with N (confirmed by the trend in as shown in Fig. 4: in view of the bandpass operation, the
Table I). To reduce the error for large n (by using a large k2 ), opamps used have a three times higher bandwidth—gm1 =
k1 and k0 have to also change with N . It is thus seen that the 15, gm2 = gm3 = 18. The finite opamp gain–bandwidth prod-
primary reason for the undesirable behavior of the coefficients uct causes the center frequencies of the resonators to shift
extracted using (7) is that l0 , l1 , and l2 are very sensitive to the and makes the quality factors negative. To avoid this, the
locations of those poles of L(z) that are close to the unit circle. capacitors are made lossy by inserting small series resistors

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:52 UTC from IE Xplore. Restricon aply.
162 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—II: EXPRESS BRIEFS, VOL. 57, NO. 3, MARCH 2010

Fig. 9. Responses of the loop filter to NRZ pulses injected at the inputs of the
first and third integrators. Fig. 10. Magnitudes of the ideal NTF and those obtained using our de-
sign centering technique for N = 10, 20, 30. The inset shows the NTF near
TABLE III ω = π/2.
M ODULATOR C OEFFICIENTS O BTAINED FOR D IFFERENT VALUES OF N
loop filter poles are inside the unit circle) results in coefficient
values that are close to the ones in Table III(b) when the
proposed technique is used (as one should expect), but results
in vastly different coefficients when (7) is employed, giving
further evidence of the robustness of the proposed method.

IV. C ONCLUSION
(rx = 0.02 Ω). A direct path (with a gain ko ) is added around
the quantizer in an additional degree of freedom intended to Conventional techniques for design centering CTDSMs rely
compensate for the excess delay caused by the opamps. In on fitting the sampled pulse response of the continuous-time
practice, it will not be possible to ensure an infinite quality loop filter to the impulse response of the prototype discrete-
factor for the resonators—and the open-loop poles could move time loop filter. This is inaccurate when practical nonidealities
into the right-half s-plane. The responses of the loop filter like excess loop delay, finite dc gain, and high-order integrator
obtained by injecting NRZ pulses at the inputs of the first and poles are considered. We presented a numerical technique to
third integrators, denoted by l4 (t) and l2 (t), are shown in Fig. 9. determine the loop filter coefficients. Our method does away
Notice that the envelopes of the waveforms are exponentially with the cumbersome algebra involved in conventional methods
increasing, indicating right-half-plane poles. and does not make any assumptions on the integrator gain,
The modulator coefficients were tuned using (7) and (10) in the number of high-order integrator poles/zeros, or the DAC
an attempt to restore the NTF to (1 + z −2 )2 . The solutions of pulse shape. It is more robust since open-loop fitting is avoided.
(7) and (10), evaluated for different values of N , are shown Illustrative design examples and the intuition behind our
in Table III. Significant variations are seen in the coefficients method were given.
determined using (7), whereas those obtained using (10) seem
robust. The modulator poles, computed using the coefficients ACKNOWLEDGMENT
obtained using (7), were outside the unit circle for N = 20 and
The author would like to thank N. Muthusubramanian for
N = 30. The NTF magnitudes computed using the solutions
useful discussions.
of (10) for N = 10, 20, 30 are shown in Fig. 10. A good
approximation to the desired NTF is seen. The inset shows
the behavior around ω = π/2. The slight shift in the center R EFERENCES
frequency due to finite opamp bandwidth effects is apparent. [1] P. Benabes, M. Keramat, and R. Kielbasa, “A methodology for designing
A finite resonator Q results in a finite notch depth. continuous-time sigma–delta modulators,” in Proc. Eur. Conf. Design Test,
Washington, DC, 1997, pp. 46–50.
The alert reader would have noticed that the coefficients in [2] J. Cherry and W. Snelgrove, Continuous-Time Delta–Sigma Modulators for
Table III(b) vary more than those in the low-pass example of High-Speed A/D Conversion: Theory, Practice, and Fundamental Perfor-
Table II. This is due to the following. From (10), we see that mance Limits. Berlin, Germany: Springer-Verlag, 1999.
h0 , h1 , and h2 will tend to ∞ when the loop filter has poles [3] R. Schreier and G. Temes, Understanding ∆Σ Data Converters.
Piscataway, NJ: IEEE Press, 2005.
outside the unit circle (as is the case in our bandpass example). [4] J. Cherry and W. Snelgrove, “Excess loop delay in continuous-time
Hence, as N increases indefinitely, K determined using (10) delta–sigma modulators,” IEEE Trans. Circuits Syst. II, Analog Digit. Sig-
will be in error. However, thanks to pole-zero cancellation, it nal Process., vol. 46, no. 4, pp. 376–389, Apr. 1999.
[5] M. Ortmanns, F. Gerfers, and Y. Manoli, “Compensation of finite gain-
will still be far more robust than using (7) for practical values bandwidth induced errors in continuous-time sigma–delta modulators,”
of N (≈15). Space limitations prevent us from giving the results IEEE Trans. Circuits Syst. I, Reg. Papers, vol. 51, no. 6, pp. 1088–1099,
here, but simulations show that using rx = 0.03 Ω (so that the Jun. 2004.

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:52 UTC from IE Xplore. Restricon aply.

Potrebbero piacerti anche