You are on page 1of 11

12 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS–I: REGULAR PAPERS, VOL. 67, NO.

1, JANUARY 2020

Distortion Analysis of RC Integrators


With Wideband Input Signals
Nikolaus Hammler , Student Member, IEEE, and Boris Murmann , Fellow, IEEE

Abstract— The integrator is an important building block


in circuit design and its transfer characteristic is inherently
frequency-dependent. When it is implemented as an RC op-amp
circuit, a nonlinear element (the amplifier) is placed within
a frequency-dependent feedback loop. Conventional techniques
allow us to obtain the nonlinear error easily at certain frequencies
only. Furthermore, when the integrator is periodically reset,
the circuit is not in steady-state and conventional techniques
assuming a steady-state behavior do not apply. It is not obvious
how the actual error relates to the circuit parameters when
the integrator is driven with a wideband signal. In this paper
we present an approach to quickly estimate the normalized
mean squared error (NMSE) as a function of the circuit and
signal parameters. We show that, to first-order, the NMSE
is proportional to the ratio of the signal bandwidth and the
dominant pole of the integrator. For an integrator with reset,
the signal builds up from zero and results in an additional, Fig. 1. Error caused by a nonlinearity in an RC integrator when driven by
integration-time dependent NMSE reduction. a wideband input signal.

Index Terms— Integrator, nonlinearity, feedback, non-


stationary.
different amounts of error in different frequency ranges.
I. I NTRODUCTION Sometimes this problem can be mitigated by looking at the

C ONVENTIONALLY, the nonlinear behavior of circuits is


described with metrics such as harmonic distortion (HD),
total harmonic distortion (THD), intermodulation distortion
worst-case nonlinearity across the frequency range. However,
integrators are inherently frequency-dependent and hence such
an approximation does not make sense.
(IM) or intercept point (IIP) [1], [2]. However, these metrics In the case of an RC integrator, the nonlinearity is wrapped
are only meaningful for sinusoidal or narrow-band signals. within a frequency-dependent feedback loop which results in
In contrast, many real world signals are wideband, and can a frequency-dependent loop gain. This causes not only the
often be approximated as Gaussian signals (e.g., OFDM or signal path to be frequency-dependent but also the nonlinearity
WCDMA signals). Conventional wideband metrics such as coefficients can be interpreted to be frequency-dependent.
multi-tone power ratio (MTPR) or adjacent channel leakage Conventional methods may give an estimate of the error at
ratio (ACLR) are only meaningful to assess the amount of a narrow frequency band but often the error for a wideband
error in simulations or measurements but do not relate to signal is of interest. Finally, integrators are often periodically
circuit parameters such as the nonlinearity coefficients. Hence, reset in which case the system is not in steady-state and
they cannot be used easily to design a circuit based on these conventional steady-state circuit analysis does not apply in the
metrics. When the circuit has a static nonlinearity (as it is first place. Examples include the random demodulator [3] or
often the case for generic amplifiers), some estimates can be the digital predistortion feedback receiver presented in [4].
made on how much error the nonlinearity adds to a wideband In principle, the system can be analyzed using Volterra
signal (e.g., based on HD3 or IIP3 and input signal power). series [5] and the total error can be found by integrating
The situation becomes harder when the system is the Volterra kernels. However, even if the series can be
frequency-dependent because then the nonlinearity adds truncated to include only the third-order kernel, the three-
Manuscript received May 14, 2019; revised August 2, 2019; accepted fold integration required is hard to evaluate analytically and
October 9, 2019. Date of publication October 22, 2019; date of current interpret intuitively.
version January 15, 2020. This article was recommended by Associate Editor In this paper we assess the error of such systems with
A. M. A. Ali. (Corresponding author: Nikolaus Hammler.)
The authors are with the Department of Electrical Engineering, Stan- the normalized mean squared error (NMSE), which gives a
ford University, Stanford, CA 94305 USA (e-mail: nhammler@stanford.edu; concise estimate of the actual error that the circuit nonlin-
murmann@stanford.edu). earity adds to a wideband signal. This setup is represented
Color versions of one or more of the figures in this article are available
online at http://ieeexplore.ieee.org. in Fig. 1. We assume a wideband input signal v i characterized
Digital Object Identifier 10.1109/TCSI.2019.2946974 by its probability density function (PDF) and power spectral
1549-8328 © 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

Authorized licensed use limited to: National Institute of Technology- Meghalaya. Downloaded on July 03,2020 at 10:24:49 UTC from IEEE Xplore. Restrictions apply.
HAMMLER AND MURMANN: DISTORTION ANALYSIS OF RC INTEGRATORS WITH WIDEBAND INPUT SIGNALS 13

nonlinear behavior via the ratio a1 /a3 . However, the IIP3 is


independent of the input signal x and hence is not useful for
obtaining the error at the output. The third-order harmonic
distortion (HD3) [1] quantifies the error level in the output
signal:
⎛ ⎞
 
Fig. 2. Block diagram of a system with third-order nonlinearity. 
1 a3  1
HD3 = 20 log ⎝     A2 ⎠ , (3)
4 a1 1 +  3a3  A2
4a1
density (PSD). We then express the NMSE as a function of  
 
the circuit (and signal) parameters. for x(t) = A sin(ωt). The term 1+  34 aa31  A2 is due to the gain
The rest of this paper is organized as follows: Section II error at the fundamental frequency. For low input signals (i.e.,
defines the error metric that will be used for the rest of Pin = 10 log(10 A2 ) < IIP3 [dBm]) this term can be neglected
the paper. In particular, we define the NMSE and apply it and by expressing (2) as a3 /a1 , (3) can be written in terms of
to a conventional system with static nonlinearity and flat input power (in dBm) and IIP3:
frequency response. Section III generalizes the results from
Section II to open-loop implementations such as the gm C HD3 ≈ 2(Pin −IIP3−10 log 3) ≈ 2(Pin −IIP3)−4.77 dB. (4)
integrator [6] in which the nonlinearity is followed by a linear For example, a circuit with IIP3 = 10dBm, driven by a
filter function (Hammerstein system). In Section IV, we review sinusoidal input signal with power Pin = −20dBm, contains
the conventional nonlinearity analysis based on loop gain and an error signal (due to harmonic distortion) −65 dB below the
Volterra series. Section V derives the NMSE for a first-order output signal.
model of the RC integrator in a wideband scenario. It is The HD3 (and similar metrics like IM3 [1]) are defined only
shown that the error is directly proportional to the ratio of the for sinusoidal signals. In practice, many real-world signals are
signal bandwidth B and the dominant pole of the integrator. wideband and can be reasonably well described as Gaussian.
Section VI extends the model to include second-order effects A meaningful way to describe the error introduced by the
such as input and load capacitance or finite bandwidth of nonlinearity is the normalized mean squared error (NMSE).
the amplifier. A few examples are provided on how these Fig. 2 shows the block diagram of a system with third-order
parasitic effects affect the NMSE. Section VII adds a reset nonlinearity (1). The ideal output is given as y1 = a1 x and
to the integrator. A non steady-state analysis extends the pre- the error signal is e = a3 x 3 and hence the NMSE becomes:
vious results. It will be shown that integration-time dependent
 

signal build-up after reset results in further reduction of the Var {e} a32 E(x 6 )
NMSE = 10 log = 10 log . (5)
NMSE. Section VIII confirms our results with transistor level Var {y1 } a12 E(x 2 )
simulations. Finally, conclusions will be given in Section IX.
If we assume that the input signal is zero-mean Gaussian with
II. E RROR M ETRICS variance σ 2 , then
E(x 2 ) = σ 2 and E(x 6 ) = 15σ 6 . Solving
Pin = 10 log σ 2 /50 · 1000 for σ , (5) can be expressed in
For the rest of this paper we assume a weakly nonlinear sys- terms of IIP3 (2) and Pin :
tem that is dominated by the third-order nonlinearity. While the  
√ 50 40
presented approach can be extended to the second-order term NMSE = 2(Pin − IIP3) + 20 log 15
or higher-order terms, we focus on the third-order term since 1000 3
even-order terms can often be avoided using a fully differen- ≈ 2(Pin − IIP3) + 8.24 dB. (6)
tial circuit implementation and weakly nonlinear systems are
Note that (6) does not take the gain error into account.
dominated by the third-order distortion term [1]. Furthermore,
However, for a Gaussian signal the NMSE is within a few dB
adding a less significant second-order term would obscure
of the actual error and hence we use NMSE as error metric.
our otherwise readable and interpretable results. When the
This is further discussed in Appendix A.
nonlinearity is static (frequency independent), as for example
to first-order in a conventional amplifier, the output signal y
can be represented in terms of the input signal x as: III. O PEN -L OOP I NTEGRATOR
Consider an open-loop integrator, for example implemented
y = a1 x + a3 x 3 . (1)
using a gm C topology [6] as in Fig. 3. We assume that the
The third-order input-referred intercept point (IIP3) [7] is a distortion due to the gm cell is dominated by the third-order
popular metric to describe the nonlinear behavior of such a term and that the output distortion due to nonlinear R = ro
circuit. It can be shown to expand to (assuming all signals in can be neglected.
rms voltages) The important question is which error level (due to the
   nonlinear gm cell) does this circuit add to the output signal in
4  a1 
IIP3 = 10 log + 10 [dBm]. (2) terms of NMSE, as discussed in Section II? Since the response
3  a3  is frequency-dependent, HD3 or THD do not answer this
All logarithms in this paper are taken to the base 10 question since the harmonics are attenuated by the integrator
(log = log10 ). Note that the IIP3 directly encodes the static transfer function. In practice, the circuit is often driven by

Authorized licensed use limited to: National Institute of Technology- Meghalaya. Downloaded on July 03,2020 at 10:24:49 UTC from IEEE Xplore. Restrictions apply.
14 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS–I: REGULAR PAPERS, VOL. 67, NO. 1, JANUARY 2020

Fig. 6. Block diagram of the integrator in Fig. 5.

equations describing the circuit are:


v o = −a1 v x − a3 v x3 , (7)
Fig. 3. gm C integrator.
1 s RC
v x = vi + vo . (8)
1 + s RC 1 + s RC
Using these equations, the result can be represented with the
block diagram shown in Fig. 6.
Before we continue, assume for a moment that instead of the
capacitor we have a feedback resistor R f . This corresponds to
the case of a conventional amplifier with constant frequency
response. In this case it is well known [1] that the linear gain
a1 is attenuated by (1 + T ) and the third-order coefficient
Fig. 4. Block diagram of gm C integrator. a3 is attenuated by (1 + T )4 where T denotes the loop gain.
Hence, (6) can be used to describe the NMSE of this system
where the static IIP3 is corrected by the loop gain:
   
 4 a1  Rf

IIP3 = 10 log   +10−30 log |1+T |+20 log .
3 a3  R + Rf
 
IIP3 from (2)
(9)
The last term comes from the attenuation at the input of
the actual feedback due to loading of the feedback network.
When we generalize this concept to the integrator, we notice
that the loop gain and the pre-factor become frequency-
dependent. Using the magnitude, this can be interpreted as
frequency-dependent IIP3.
The Volterra series can be used to accurately describe a
frequency-dependent, nonlinear system. Reference [5, Ch. 2.5]
describes how to obtain the Volterra kernels in a feedback
Fig. 5. RC integrator. configuration. For the described system, the first-order kernel
is the linear response given by:
a1 a1
a wideband signal. The block diagram of Fig. 3 is shown H1(s) = − =− , (10)
1 + s RC(a1 + 1) 1 + sτint
in Fig. 4 along with the signal PSDs. In general, the signals i 1
where we use the notation 1/(2πτint ) for the dominant pole
and i e have different bandwidths (due to the nonlinearity) and
of the integrator. The third-order kernel is given as:
hence are filtered differently. However, the signal bandwidth
is usually much larger than the dominant pole 1/(2π RC) H3(s1 , s2 , s3 ) = P(s1 )P(s2 )P(s3 )a3 ·
and hence, both signals are approximately subject to the R(s1 )R(s2 )R(s3 )R(s1 + s2 + s3 ),
same filtering, in which case the filter action cancels in the 1 + s RC 1
error calculation. For that reason, (6) remains valid also for R(s) = , P(s) = . (11)
1 + sτint 1 + s RC
an open-loop integrator for which the nonlinearity can be
The third-order intermodulation products are given as
input-referred.
H3( j ω1 , j ω1 , − j ω2 ) from which an expression for the
IIP3 can be derived as:
 
IV. C ONVENTIONAL A NALYSIS OF AN RC I NTEGRATOR 4 H1( j ω1 ) 
IIP3 = 10 log    + 10 [dBm]. (12)
Consider the RC integrator shown in Fig. 5. Compared to 3 H ( jω , jω , − jω )
3 1 1 2
an ideal RC integrator, this implementation has finite gain a1 By making the two input frequencies closely spaced,
as well as a third-order nonlinearity described by a3 . The two the IIP3 can be measured as a function of frequency in the

Authorized licensed use limited to: National Institute of Technology- Meghalaya. Downloaded on July 03,2020 at 10:24:49 UTC from IEEE Xplore. Restrictions apply.
HAMMLER AND MURMANN: DISTORTION ANALYSIS OF RC INTEGRATORS WITH WIDEBAND INPUT SIGNALS 15

Fig. 9. Block diagram of Fig. 8.

Fig. 7. IIP3 vs. frequency.

Fig. 10. IIP3 vs. frequency including parasitics. C L = 10 pF, τint = 0 (solid)
and 1/(2π τint ) = 300 MHz (dotted).

is modeled using a resistor R1 and C1 after the (static,


nonlinear) voltage source. This filter can be used to model
the frequency dependent nonlinearity of the amplifier core.
The block diagram is shown in Fig. 9. The Volterra kernels
can be extended to include all effects and the result is shown
Fig. 8. RC integrator including second-order effects. in Fig. 10. Again, the results are obtained from Cadence
Spectre pss analysis, but this time using the behavioral model
in Fig. 8. IIP3( f ) is shown for varying Ro , C L = 10 pF
lab or simulated using periodic steady-state (pss) analysis. For and IIP30 = −10dBm. The results are shown for an amplifier
ω2 → ω1 the result coincides with the loop gain method with infinite bandwidth (solid) and 300 MHz (dashed). With
as discussed above and is shown in Fig. 7. This plot has parasitics, IIP3( f ) is not increasing with frequency any more.
been generated by sweeping the two closely spaced input A combination of finite amplifier bandwidth and nonzero
frequencies over the frequency range, and performing a pss Ro creates a pair of complex conjugated zeros. It is not
analysis in Cadence Spectre. The circuit uses the behavioral clear how the NMSE behaves for a wideband signal. Since
model from Fig. 5. At very low frequencies, the capacitor H3(s1 , s2 , s3 ) represents the added error, in principle, the first
is an open circuit and the output voltage given as v o = and third-order Volterra kernels could be integrated. However,
−a1 v i −a3 v i3 . We denote this value as IIP30 which coincides the three-fold integration required for the third-order kernel is
with (2). Using (12) and (6) the NMSE can be found for hard to evaluate analytically. Hence we present an alternative
a narrowband signal in a small frequency band (in which approach in the next sections.
IIP3( f ) is assumed to be constant). However, the NMSE is
not evident for a wideband signal. Hence this method is not V. RC I NTEGRATOR IN S TEADY-S TATE
suitable to analyze systems driven by wideband signals. If a The goal is to find the NMSE between v o in Fig. 6 and
lower frequency limit is known, a lower bound on the NMSE the same system with a3 = 0 as a function of the circuit
could be found since IIP3( f ) is monotonically increasing. parameters (IIP3, R and C, a1 ). Conventional circuit analysis
Furthermore, IIP3( f ) starts to rise at the frequency 1/(2πτint ). or Volterra series easily allow us to obtain distortion products
Hence the NMSE can be minimized by maximizing a1 for a at certain frequencies only. Our approach for wideband signals
given unity gain frequency. first uses a method proposed by Bussgang in 1974 [8] to
A more comprehensive version of the RC integrator is unroll the feedback loop. The basic idea is shown in Fig. 11.
shown in Fig. 8 which adds input capacitance Ci , nonzero By replacing the nonlinearity with its linear part a1 v x ,
output resistance Ro , load capacitance C L and a parasitic we obtain the linear output, as shown in Fig. 11(a). The signal
resistance R p . The finite bandwidth of the amplifier 1/(2πτ1 ) seen in front of this block (v i,3 ) is the signal which was

Authorized licensed use limited to: National Institute of Technology- Meghalaya. Downloaded on July 03,2020 at 10:24:49 UTC from IEEE Xplore. Restrictions apply.
16 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS–I: REGULAR PAPERS, VOL. 67, NO. 1, JANUARY 2020

If v o,1 is the ideal integrator output and v o,3 the error term,
the NMSE is given as:
 

Var v o,3
NMSE = 10 log   . (14)
Var v o,1
From basic signal theory, the ideal output power can be
obtained by integrating the second-order transfer function (the
input PSD is normalized) [10]:
 ∞
  σ2 a12
Var v o,1 = |H0(s)H1 (s)|2 d s =
0 4B τint + 4B1

σ 2 a12
≈ . (15)
4B τint
Note that for B  1/(4τint ), the effect of the noise shaping
can be ignored (second line). This is the case for all prac-
tical applications (otherwise the system would operate as an
amplifier rather than integrator). Hence we can for simplicity
assume a white input signal whose PSD is scaled such that its
variance is σ 2 given a certain bandwidth B; i.e., H0(s) = √σ .
  B
Obtaining Var v o,3 is much harder since the the output
Fig. 11. Unwrapping the feedback loop. consists of a Wiener-Hammerstein system of (H0 H1→3)
and H3. For that reason we use the framework around
Best Linear Approximation (BLA) of nonlinear systems [11].
BLA replaces the nonlinear system with the best possible
approximation of a linear system. For a Wiener-Hammerstein
system, it can be shown that this approximation is given
by cbla H0 H1→3(s)H3 (s) = cbla G(s), where cbla is a con-
stant. The error caused by this approximation is discussed in
Appendix B. G(s) is the cascade between the Wiener- and
Fig. 12. Block diagram of the integrator with unwrapped feedback loop. the Hammerstein part. An analytic expression for cbla can be
found [12]:

σ2 ∞ 3 a3 σ 2
sent through the feedback loop and will subsequently hit the cbla = 6 a3 |H1→3(s)|2 d s = . (16)
third-order nonlinearity as shown in Fig. 11(b). The output 2B 0 4 Bτint
of Fig. 11(b) with the input set to zero (v i = 0) represents Theresulting
 simplified model is shown in Fig. 13. Now
the output due to the third-order nonlinearity. While the RC Var v o,3 can be calculated in the same way as Var v o,1 (15)
integrator is assumed to only contain the third-order coefficient and the NMSE (14) results in:
 
(see Fig. 5), the feedback still generates higher-order distortion f p,1,eff √
terms. However, for a weakly nonlinear system, the approxi- NMSE=2(Pin −IIP30 )+20 log +20 log 2, (17)
B
mation is fairly accurate (see Appendix C).
where f p,1,eff = π2 f p,1 = 1/(4τint ) and we used the approx-
We call the transfer functions v o,1 /v i = H1(s), v i,3 /v i =
 imation (a1 + 1)2  1 and the white noise approximation
H1→3(s) and v o,3 /v i,3 = H3(s) which results in the block
H0(s) = √σ from above. The result makes it clear that for a
diagram in Fig. 12. The total output is given as v o = v o,1 +v o,3 . B
given IIP3, Pin and unity crossover frequency, the NMSE is
Note that we model the wideband signal by passing a mean
minimized by maximizing a1 , which increases the loop gain
free Gaussian independent and identically distributed (i.i.d)
and hence linearity. While this result was intuitively clear, (17)
process through a signal shaping filter H0 . While in general
provides a concise relationship.
H0 can be arbitrary to accommodate flexible evaluation for
It is interesting to note that as a first estimate in Fig. 12,
different wideband signals, we use the following filter:
H3 can be ignored (H3(s) = 1), resulting in a Wiener system.
σ 1 The output of H0 H1→3 is then given by
H0(s) = √ , (13)
B 1 + s/(4B)  ∞
σ2 1
which creates a bandlimited white Gaussian noise input signal |H0 (s)H1→3(s)|2 d s = , (18)
0 4B τint
v i with first-order rolloff in frequency, effective noise band-
which is, apart from a1 , the same
 result
 as for (15). Pretending
width B and variance σ 2 at the input of the integrator [9].
this signal is still white, Var v o,3 = 15 a32 σ 6 /(4 Bτint )3 .
For our transient simulations in Cadence Spectre, we create
Combining this with (15) results in
a noise resistor with R = 1/(4 kT ) followed by an ideally  
f p,1,eff
buffered RC lowpass √ filter with cutoff frequency ωc = 4 B NMSE = 2(Pin −IIP30 )+20 log +8.24 dB. (19)
and gain A0 = σ/ B. B

Authorized licensed use limited to: National Institute of Technology- Meghalaya. Downloaded on July 03,2020 at 10:24:49 UTC from IEEE Xplore. Restrictions apply.
HAMMLER AND MURMANN: DISTORTION ANALYSIS OF RC INTEGRATORS WITH WIDEBAND INPUT SIGNALS 17

Fig. 13. Simplified linear error model for the RC integrator with first-order
effects.

TABLE I
C OEFFICIENTS FOR D(s) P OLYNOMIAL

Fig. 14. NMSE vs. input capacitance (α = Ci /C) for different a1 .


Solid: approximation, cross: simulation.

The calculation of the integrals such as in (15) becomes


much more involved, in particular for G(s), since the order
It is interesting to compare this result to the wideband of the polynomials doubles due to the multiplication of H1→3
result (6): the NMSE of the RC integrator is smaller by a and H3. Hence we employ the approach from [14], which
factor of f p,1,eff /B. This makes intuitively sense. Based on allows us to obtain the integrals of transfer functions with an
Fig. 12, the input signal is first attenuated by the dominant arbitrary number of poles.
pole of the integrator and hence it excites the nonlinearity As an example, consider adding an input capacitance Ci
with a much weaker signal. and let Ci = αC. All other second order effects are set
We note that in the literature the BLA framework is com- to zero to keep the expressions manageable and useful for
monly used to model a linear system with nonlinear distortion interpretation. The NMSE evaluates to:
as a linear model. For example, [13] applies the framework to  
1
feedback systems. However, in our case we are interested in NMSE = 2(Pin − IIP30 ) + 20 log
RC(a1 + α + 1)B
describing the error caused by the nonlinearity in terms of the  
(1 + α)2 √
circuit parameters. Hence we only apply BLA to the nonlinear +10 log 1 + − 20 log(2 2). (21)
part of the system to get an estimate of the error variance. (1 + a1 + α) 2

It is interesting to see that a larger Ci (larger α) decreases


VI. RC I NTEGRATOR AND S ECOND -O RDER E FFECTS the√NMSE. If α is much larger than a1 , the error scales by
α/ 2. The result is shown in Fig. 14 for different gains.
The RC integrator including second-order effects (Fig. 8) The lines show the NMSE according to (21). The crosses
was introduced in Section IV. Fortunately, its block diagram show simulation results from a transient simulation in Cadence
(Fig. 9) maps to the same form as in Figs. 12 and 13 and Spectre of the behavioral model from Fig. 8. The NMSE is
hence, its transfer functions can be used in the same model obtained by comparing the output of the behavioral model
(Fig. 13) that was introduced in Section V with: with nonzero a3 to the same model with a3 = 0 (see Fig. 1).
The improvement is fairly small (a few dB) and becomes
H1 (s)
Ro C
negligible for large a1 . Intuitively, Ci adds additional high
1 + s(R p C − 2 Ro Cτ1
a1 ) − s ( a1 )) frequency filtering (attenuation) and hence the nonlinearity is
= a1
D(s) excited less. The effect on the NMSE is only pronounced when
H1→3 (s) Ci is large and hence also the low frequencies (close to the
1 + s(Ro (C + C L ) + R p C + τ1 ) dominant pole of the integrator) are attenuated.
= Another interesting result can be obtained from the model:
D(s)
s 2 (Ro R p CC L+τ1 (Ro (C +C L )+ R p C))+s 3 Ro R p CC L τ1 Consider a finite bandwidth of the amplifier described by
+ τ1 = R1 C1 (see Fig. 8). We assume that the signal band-
D(s)
width is set to the bandwidth of the amplifier, i.e. we set
H3 (s)
τ1 = 1/(4B). The resulting expression can be simplified
1 + s(R(C + Ci ) + C R p ) + s 2 R R p CCi assuming a large amplifier gain a1 (in particular a1 +2 ≈ a1 +1
=
D(s) and a1 + 3/2 ≈ a1 + 1) and the resulting NMSE is:
D(s)  
1 4
= 1 + d1 s + d2 s 2 + d3 s 4 + d4 s 4 . (20) NMSE = 2(Pin − IIP3) + 20 log
2 1 + 4Bτint
 
The coefficients for the denominator polynomial are given 1 + 2Bτint
+10 log . (22)
in Table I. 1 + 4Bτint

Authorized licensed use limited to: National Institute of Technology- Meghalaya. Downloaded on July 03,2020 at 10:24:49 UTC from IEEE Xplore. Restrictions apply.
18 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS–I: REGULAR PAPERS, VOL. 67, NO. 1, JANUARY 2020

to comparator-based circuits in a similar way as the noise


analysis was performed in [15].
We add a non-stationary analysis [16], [17] to the model
from the previous sections, similarly as it was also done
in [15]. If we assume a white Gaussian input PSD, the output
variance of an LTI system h at time Tint is given as:
 Tint
1
Var {y} (Tint ) = N0 |h(ξ )|2 d ξ. (24)
2 0
If we assume that the reset presets all dynamic states (for
example, by setting capacitor voltages to their operating point
voltages), the system starts from zero initial conditions after
each reset. Since the system is not yet in steady-state, the vari-
ance of the output signal is time-dependent, as evident by (24).
We are interested in the NMSE at t = Tint :
 

Fig. 15. RC integrator with reset. Var v o,3 (Tint )


NMSE = 10 log   . (25)
Var v o,1 (Tint )
 
With the approximation Bτint  1 (which holds for an In order to obtain Var v o,1 (Tint ), we first invoke the inverse
integrator), (22) approaches (17). Hence we can conclude that Laplace transform to obtain h 1 (t) and then apply (24). The
as long as the open-loop amplifier bandwidth is at least as impulse response is given as:
wide as the signal bandwidth, the effects of nonlinearity are 1 −t /τint
negligible. When the signal bandwidth is much larger than h 1 (t) = L−1 {H1(s)} = −a1 e . (26)
τint
than the largest pole of the system (such that √ we can again
use the white noise approximation H0 (s) = σ/ B), and the After squaring and integrating, the result is
amplifier bandwidth is much larger than the dominant pole,   σ 2 a12  
i.e., B  a1 /(4τ1 ) and τint  τ1 , the NMSE increases with the Var v o,1 (Tint ) = 1 − e−2Tint /τint . (27)
B 4τint
ratio of the dominant pole divided by the unity gain frequency,  
as compared to (17): For Var v o,3 (Tint ), we assume that cbla is not a constant but
 τ1  time-dependent as well:
NMSE ≈ 20 log 1 + . (23) 
RC σ2 t σ2 1
cbla (t) = 6 a3 |h 1→3 (τ )|2 d τ = 3a3
We note that the influence of additional effects may not be neg- 2B 0 B 4τint
 
ligible but we assume them to be zero to examine the impact
1 − e−2t /τint . (28)
of a single parameter. With less simplifications, the resulting
expressions would not provide any useful insights. However, After repeating the same analysis for g(t), the resulting
as will be shown in Section VIII, the error can be dominated NMSE (30), shown at the bottom of the next page. Eq. (30)
by the first-order effects. Furthermore, we emphasize that the allows us to obtain the NMSE exactly but it is too complicated
model shown in Fig. 13 in combination with the full model to get first-order insights. The expressions under the square
still permits a quick way to numerically evaluate the design roots can be reasonably well approximated with the function
α
trade-offs with multiple second-order effects included. 1 − e−(γ Tint /τint ) where γ and α denote fudge factors. This
is shown in Fig. 16 where both versions are compared. It is
VII. RC I NTEGRATOR W ITH R ESET important to note that γ and α are not really constants; they
The analysis so far has assumed that the system is in are only valid in certain regions of the design parameters.
steady-state. If the integrator settles, i.e., we wait much longer However, it allows us to represent the NMSE in a more
than τint , the system acts as a normal filter which attenu- compact form:
ates the higher frequencies but passes the lower frequencies  
f p,1,eff
unchanged. For that reason, a commonly employed topology NMSE = 2(Pin − IIP30 ) + 20 log
B
resets the integrator as shown in Fig. 15 where the value at  
T
α 
− γ τ int
an instant before the reset (at t = Tint ) is of interest. For +10 log 1 − e int . (29)
simplicity, we consider the first-order model of the integrator
in this section (consisting of only R, C, a1 , a3 ). Note that Compared to (17), the expression in (29) adds an additional
this model applies to many applications which do not have term which decreases the NMSE: The shorter the integration
explicit reset switches. For example, consider a wideband interval (and the farther away the system is from steady-state),
signal which is modulated (multiplied) with a long pulse train the smaller the error. This result has an intuitive explanation.
(a signal which is one for Ton and zero for Toff ). This action Due to the integrator reset, the input signal builds up from
can be interpreted as a switch where the wideband signal is zero after each reset and hence never reaches the full signal
integrated during Ton . The method could also be applied swing. Hence, the nonlinearity is excited with a smaller signal,

Authorized licensed use limited to: National Institute of Technology- Meghalaya. Downloaded on July 03,2020 at 10:24:49 UTC from IEEE Xplore. Restrictions apply.
HAMMLER AND MURMANN: DISTORTION ANALYSIS OF RC INTEGRATORS WITH WIDEBAND INPUT SIGNALS 19

Fig. 18. Transistor level implementation of the amplifier core.

Fig. 16. Approximation of the square root in (30).

Fig. 19. Comparison between transistor level implementation and behavioral


model (ac analysis).

reset with a differential implementation of the integrator by


shorting the input and output terminals.

Fig. 17. NMSE for different values of a1 . VIII. T RANSISTOR -L EVEL S IMULATION
In order to verify our technique, we performed transistor
level (TL) simulations in ST 28 nm technology. The amplifier
resulting in a lower error. Fig. 17 compares the theory (30) and core is implemented as a single stage amplifier with a source
the approximation (29) with simulation results. The simulation follower. The schematic is shown in Fig. 18. The small signal
results have again been obtained using a transient simulation parameters are extracted and a model (as shown in Fig. 8) is
in Cadence Spectre of the behavioral model from Fig. 5. The created. The extracted parameters are: a1 = 22.45, IIP30 =
theory (30) matches the simulated values fairly well. −13.15dBm, τ1 = 1/(2 π 550.36 MHz), Ro = 43.32
, Ci =
The analysis presented here can be extended to include 147.43 fF, and C L = 245.64 fF. For the closed loop integrator
higher-order effects as well, but the expressions become configuration, additional C L = Ci = 20 pF are added, as well
intractable. In this case, the block diagram from Fig. 13 can as R = 300
and C = 100 pF. Fig. 19 compares the
again be used for numerical sweeps in the design space. frequency response of the transistor level implementation of
Furthermore, it may not be possible to reset the states of the amplifier with the behavioral model (solid: model, dotted:
all dynamics (for example, finite bandwidth of the amplifier transistor level implementation). Furthermore, the frequency
caused by higher-order poles). A combined steady-state and response of the closed-loop system is shown. Fig. 20 shows
non steady-state analysis is possible as well. If the parasitic the IIP3 versus frequency as explained in Section IV (Fig. 7
poles and zeros due to Ci and C L are significant, they can be and Fig. 10).

 
3 a3 2 1
NMSE = 20 log σ
4 a1 4Bτint (a1 + 1)



  a12 a12 Tint a12 2
−2T /τ 
+20 log 1 − e int int + 20 log 1 + a1 + −2T /τ
− e int int 1 + a1 + + a1 (a1 + 2) + 2 Tint (30)
2 2 τint τint

Authorized licensed use limited to: National Institute of Technology- Meghalaya. Downloaded on July 03,2020 at 10:24:49 UTC from IEEE Xplore. Restrictions apply.
20 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS–I: REGULAR PAPERS, VOL. 67, NO. 1, JANUARY 2020

metrics such as THD, HD3, IIP3, MTPR or ACLR. These


metrics are either only useful for sinusoidal or narrowband
signals and can only be extrapolated to wideband signals in
systems with flat frequency response or they do not provide
a link to circuit parameters. These methods fail when the
nonlinear behavior is frequency-dependent or the system is
not in steady-state.
The presented approach works for wideband signals as they
often occur and uses the NMSE as error metric. The NMSE
is a natural error metric that gives the error level due to the
nonlinearity as compared to an ideal system, measured in dB.
The presented equations relate the NMSE to circuit-related
properties such as R, C, a1 and a3 .
The analysis shows that the NMSE in an RC integrator
with reset can be substantially lower than the NMSE of the
Fig. 20. Comparison between transistor level implementation and behavioral
model (IIP3 vs. frequency). open-loop amplifier. We identified two primary mechanisms
that decrease the error:
1) The RC feedback network causes the signal at the non-
linearity to be decreased by the ratio between the dom-
inant pole and the signal bandwidth . The conventional
loop gain analysis, which is based on sinusoidal analysis,
does not easily provide this insight for wideband signals.
2) The periodic reset causes the signal to build up from
zero and hence excites the nonlinearity less.
Other large signal effects may affect the nonlinear behavior
as well. Care has to be taken to avoid slewing in the inte-
grator. Charge injection is an effect which has to be analyzed
separately and is not within the scope of our analysis. In our
simulations (see Section VIII), the error is dominated by the
small signal nonlinearity. The presented results can be used
as first-order design equations and can serve as sanity checks.
The simplified model can be used for more accurate error
Fig. 21. NMSE of the transistor-level implementation versus RC estimations when evaluated numerically.
time-constant.
A PPENDIX
A 400 MHz bandlimited white Gaussian noise signal with A. Error Due to Gain Normalization
Pin = −10dBm is used as system input. As before, the signal As mentioned in Section II, the nonlinearity causes a gain
is generated with a resistor R = 1/(4 kT ) and filtering it error. Hence it is meaningful to normalize the output of the
with a first-order
√ filter with cutoff frequency ωc = 4 B and nonlinear system to have the original gain a1 (or vice versa)
gain σ/ B. The NMSE is obtained using a transient noise before obtaining the error. The error can also be obtained by
simulation (averaged over 50 runs) and compared to a linear calculating the difference between the ideal (y1 = a1 x) and
version of the system. Since it is not possible to selectively non-ideal (y = a1 x + a3 x 3 ) outputs normalized to unit
disable nonlinear behavior in a transistor-level simulation, norm which we call unit-norm normalized mean squared error
the linear version is emulated by attenuating the input of (UNMSE):
the transistor-level implementation by 30 dB and amplifying  
y1 y
it again at the output. Fig. 21 shows the obtained simulation UNMSE = 10 log Var std(y 1)
− std(y) . (31)
results along with the theory (30) and the approximation (29),
Using the identity Var {a − b} = Var {a} + Var {b} − 2E(ab),
showing a good match.
this becomes:
While Section VI introduced 2nd order effects,   
1 + 3α
the transistor-level simulations show that the error is UNMSE = 10 log 2 1 − √
dominated by the first-order effects as long as the amplifier   1 + 15α  + 6α
2

bandwidth is sufficiently large and the output resistance is ≈ 10 log 2 1 − (1 − 3α ) ,2


(32)
sufficiently small. The presented technique was also used to
design the integrator in [4]. where α = σ 2 a3 /a1 and the second line follows from a Taylor
expansion. Finally, the UNMSE can be represented again using
IX. C ONCLUSION Pin and IIP3:

In this paper we presented a new way to assess the nonlinear UNMSE ≈ 2(Pin − IIP3) + 10 log(24/9)
error of RC integrators. Conventional methods rely on error ≈ 2(Pin − IIP3) + 4.26 dB. (33)

Authorized licensed use limited to: National Institute of Technology- Meghalaya. Downloaded on July 03,2020 at 10:24:49 UTC from IEEE Xplore. Restrictions apply.
HAMMLER AND MURMANN: DISTORTION ANALYSIS OF RC INTEGRATORS WITH WIDEBAND INPUT SIGNALS 21

Comparing (6) and (33) it can be seen that the difference is


about 4 dB. Since the UNMSE is much harder to obtain and
results in more complicated expressions, we use the NMSE
as a pessimistic estimate of the UNMSE. Simulations confirm
that the difference is always within a few dB (see for example
Fig. 17).
Note that due to the gain error, the NMSE is only a suitable
error metric for wideband signals. For a sinusoidal input signal
the actual error of interest is the HD3. However, it can be
shown that NMSE = HD3 + 10 dB. In contrast, the UNMSE
equals HD3 in this case.
To summarize, while conventional metrics such as HD3,
THD or IM3 are only meaningful for sinusoidal or narrowband
signals, the UNMSE generalizes its concept for wideband
signals. The UNMSE reduces to HD3 for a sinusoidal signal
Fig. 22. NMSE for a closes-loop integrator and the approximation with
applied to a static nonlinearity while also describing the error Bussgang’s method.
on the output signal for a Gaussian signal. Furthermore,
the NMSE ignores the gain normalization but follows the
UNMSE up to a few dB for a wideband signal. Therefore In order to obtain an exact match, the method has to be
the NMSE acts as a good proxy for the UNMSE in wideband repeated an infinite amount of time since the third-order output
conditions. cycles through the loop and creates a fifth-order term and
so on.
B. Error Due to the Best Linear Approximation However, for a weakly nonlinear system, we assume that the
The motivation for applying the BLA approximation is that third-order contribution is dominant. Although the feedback
it is easy to find the output variance for a linear system generates higher order distortion terms, they are much lower
(and generally hard for a nonlinear system). The question than the third-order term for a weakly nonlinear system [1].
is how much error does the BLA approximation add? The Hence it is reasonable to approximate even the unrolled
advantage of our method is that the actual approximation error system with the third-order term only. This approach was
is irrelevant; we are only interested on how the output variance employed in [19] as well. Fig. 22 shows the NMSE from
compares between the actual system and the BLA. This is a transient simulation from a closed-loop integrator and the
because in Fig. 13, the lower branch (the BLA approximation) approximation versus Pin . It can be seen that they match very
is only used to determine the variance of the error (not the well for small input signals (i.e., when the system is weakly
error itself). nonlinear). A discrepancy is only visible for larger Pin in
If the output variance for the lower branch (the Wiener which case the NMSE is high (and this operating condition is
Hammerstein system) could be easily obtained (to compare usually not of interest).
it with the BLA), we would not need the approximation in
the first place. Hence it is hard to answer this question for R EFERENCES
the general case. However, consider the case when we assume [1] P. Wambacq and W. Sansen, Distortion Analysis of Analog Integrated
H1→3(s) = α and H3(s) = β. This can be interpreted as Circuits. New York, NY, USA: Springer, 1998.
[2] B. Razavi, RF Microelectronics, 2nd ed. Upper Saddle River, NJ, USA:
examining the problem at one frequency. Then the output Prentice-Hall, 2012.
is given as y = a3 β(αx)3 = ã3 x 3 . The BLA is given by [3] J. N. Laska, S. Kirolos, M. F. Duarte, T. S. Ragheb, R. G. Baraniuk, and
Bussgang’s theorem [18]: Y. Massoud, “Theory and implementation of an analog-to-information
converter using random demodulation,” in Proc. IEEE Int. Symp. Circuits
E(x y) Syst. (ISCAS), May 2007, pp. 1959–1962.
gbla = = 15ã32σ 6 , (34)
E(x 2 ) [4] N. Hammler, A. Cathelin, P. Cathelin, and B. Murmann, “A spectrum-
sensing DPD feedback receiver with 30× reduction in ADC acquisition
for ybla = gbla x. The error difference in our approximation is bandwidth and sample rate,” IEEE Trans. Circuits Syst. I, Reg. Papers,
then given by vol. 66, no. 9, pp. 3340–3351, Sep. 2019.
    [5] W. J. Rugh, Nonlinear System Theory: The Volterra/Wiener
Var {gbla x} 9 Approach. Baltimore, MD, USA: The John Hopkins Univ. Press,
NMSE = 10 log = 10 log , (35) 1981.
Var {y} 15 [6] H. Khorramabadi and P. R. Gray, “High-frequency CMOS continuous-
i.e., the difference is only 2.2 dB. Our simulations suggest that time filters,” IEEE J. Solid-State Circuits, vol. SSC-19, no. 6, pp. 939–
948, Dec. 1984.
this result extends to the Wiener-Hammstein case. As can be [7] S. C. Blaakmeer, E. A. M. Klumperink, D. M. W. Leenaerts, and
seen for example in Fig. 14 or Fig. 17, the error compared to B. Nauta, “Wideband balun-LNA with simultaneous output balancing,
simulation is only around 2 dB. noise-canceling and distortion-canceling,” IEEE J. Solid-State Circuits,
vol. 43, no. 6, pp. 1341–1350, Jun. 2008.
[8] J. J. Bussgang, L. Ehrman, and J. W. Graham, “Analysis of nonlinear sys-
C. Error Due to Bussgang’s Method tems with multiple inputs,” Proc. IEEE, vol. 62, no. 8, pp. 1088–1119,
Aug. 1974.
Section V explained how we use Bussgang’s method [8] to [9] A. Papoulis, Probability, Random Variables, and Stochastic Processes,
unroll the feedback loop and to obtain a feedforward system. 4th ed. New York, NY, USA: McGraw-Hill, 2002.

Authorized licensed use limited to: National Institute of Technology- Meghalaya. Downloaded on July 03,2020 at 10:24:49 UTC from IEEE Xplore. Restrictions apply.
22 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS–I: REGULAR PAPERS, VOL. 67, NO. 1, JANUARY 2020

[10] A. Jeffrey and D. Zwillinger, Table of Integrals, Series, and Products. Boris Murmann (S’99–M’03–SM’09–F’15) rece-
New York, NY, USA: Elsevier, 2007. ived the Dipl.Ing. (FH) degree in communications
[11] J. Schoukens, M. Vaes, and R. Pintelon, “Linear system identification in engineering from Fachhochschule Dieburg, Dieburg,
a nonlinear setting: Nonparametric analysis of the nonlinear distortions Germany, in 1994, the M.S. degree in electrical engi-
and their impact on the best linear approximation,” IEEE Control Syst., neering from Santa Clara University, Santa Clara,
vol. 36, no. 3, pp. 38–69, Jun. 2016. CA, USA, in 1999, and the Ph.D. degree in electrical
[12] R. Pintelon and J. Schoukens, System Identification: A Frequency engineering from the University of California at
Domain Approach, 2nd ed. Hoboken, NJ, USA: Wiley, 2012. Berkeley, Berkeley, CA, USA, in 2003.
[13] R. Pintelon and J. Schoukens, “The best linear approximation of From 1994 to 1997, he was with Neutron
nonlinear systems operating in feedback,” in Proc. IEEE Int. Instrum. Mikrolektronik GmbH, Hanau, Germany, where he
Meas. Technol. Conf., May 2012, pp. 2092–2097. was involved in the development of low-power and
[14] A. Dastgheib and B. Murmann, “Calculation of total integrated noise smart-power application-specified integrated circuits (ASICs) in automotive
in analog circuits,” IEEE Trans. Circuits Syst. I, Reg. Papers, vol. 55, CMOS technology. Since 2004, he has been with the Department of Electrical
no. 10, pp. 2988–2993, Nov. 2008. Engineering, Stanford University, Stanford, CA, USA, where he is currently a
[15] T. Sepke, P. Holloway, C. G. Sodini, and H.-S. Lee, “Noise analysis for Full Professor. His current research interests include the area of mixed-signal
comparator-based circuits,” IEEE Trans. Circuits Syst. I, Reg. Papers, integrated circuit design, with a special emphasis on data converters, sensor
vol. 56, no. 3, pp. 541–553, Mar. 2009. interfaces, and circuits for embedded machine learning.
[16] A. Papoulis, Probability, Random Variables, and Stochastic Processes, Dr. Murmann served as the Data Converter Subcommittee Chair and the
3rd ed. New York, NY, USA: McGraw-Hill, 1991. 2017 Program Chair for the IEEE International Solid-State Circuits Confer-
[17] A. Leon-Garcia, Probability and Random Processes for Electrical Engi- ence (ISSCC). He was a co-recipient of the Best Student Paper Award at
neering, 2nd ed. Reading, MA, USA: Addison-Wesley, 1994. the Very Large-Scale Integration (VLSI) Circuits Symposium in 2008 and a
[18] J. J. Bussgang, “Crosscorrelation functions of amplitude-distorted recipient of the Best Invited Paper Award at the IEEE Custom Integrated
Gaussian signals,” Dept. Res. Lab. Elect. Eng., Massachusetts Inst. Circuits Conference (CICC) in 2008, the Agilent Early Career Professor
Technol., Cambridge, MA, USA, Tech. Rep. 216, Mar. 1952. Award in 2009, and the Friedrich Wilhelm Bessel Research Award in 2012.
[19] S. Pavan, “Efficient simulation of weak nonlinearities in continuous- He served as an Associate Editor for the IEEE J OURNAL OF S OLID -S TATE
time oversampling converters,” IEEE Trans. Circuits Syst. I, Reg. Papers, C IRCUITS .
vol. 57, no. 8, pp. 1925–1934, Aug. 2010.

Nikolaus Hammler (S’12) received the B.S. and


M.Sc. degrees in telematics from the Graz Uni-
versity of Technology, Graz, Austria, in 2008 and
2011, respectively, and the Ph.D. degree in electrical
engineering from Stanford University, Stanford, CA,
USA, in 2019.
From 2011 to 2012, he was with NXP Semicon-
ductors, Gratkorn, Austria, where he was involved
in the development of smartcards and security
application-specified integrated circuits (ASICs).
From summer 2013 to summer 2017, he has held
internship positions at Maxim Integrated, STMicroelectronics, Xilinx, and
Apple. His research interests include circuit and system-level design of
wireless systems, digital predistortion, and signal processing.

Authorized licensed use limited to: National Institute of Technology- Meghalaya. Downloaded on July 03,2020 at 10:24:49 UTC from IEEE Xplore. Restrictions apply.

You might also like