Professional Documents
Culture Documents
net/publication/3451305
Article in IEEE Transactions on Circuits and Systems I Regular Papers · September 2006
DOI: 10.1109/TCSI.2006.877890 · Source: IEEE Xplore
CITATIONS READS
41 1,772
3 authors, including:
All content following this page was uploaded by Michael D. Godfrey on 15 February 2024.
Abstract— Reset noise sets a fundamental detection limit on required and provide references to extensive treatment of these
capacitive sensors. Many sensing circuits depend on accumulating methods [11]–[14].
charge on a capacitor as the sensing method. Reset noise is the Reset noise occurs in all sensing circuits which depend
noise that occurs when the capacitor is reset prior to the charge
accumulation cycle. Therefore, it is important to understand on measurement of the charge accumulated on a capacitor.
the factors which determine reset noise, and how this noise Such circuits operate by resetting the capacitor, allowing
may be mitigated. The purpose of this paper is to show how charge accumulation, and then sensing the charge value. Noise
capacitive reset noise can be reduced during the reset cycle. reduction requires trade-offs of space, speed, and cost. No
We present and analyze three circuits that implement the basic one solution will fit all needs. Of course, noise reduction at
methods for directly reducing capacitive reset noise. In addition,
we present a time-domain technique for analyzing the time– one point implies at least as much noise increase somewhere
varying statistics of these circuits. This technique makes use of else. We have not found a way to avoid the Second Law of
Itô calculus to obtain solutions to the time-varying stochastic Thermodynamics. In any case, it is necessary to distinguish
differential equations. Theoretical noise calculations and Monte tradeoffs associated with temporal noise, DC offset, stability,
Carlo simulation results are presented for each technique. We lag, and reset time within each circuit implementation. In this
show that theory and simulation yield similar results.
Finally, we show in the examples that reset noise may be paper we will concentrate on temporal noise.
reduced by a factor of 20 or more. We also refer to implemented This paper is organized as follows: Section II provides a
sensor arrays which achieve these results. summary of the theory of thermal noise in capacitive sensors.
Index Terms— Noise, Itô calculus, time-varying stochastic dif- Section III describes how thermal noise affects capacitive
ferential equations, CMOS sensors, capacitive sensor, imaging. sensors, and presents the notation used for the remainder
of the paper. Section IV presents a theoretical basis for our
noise analysis, and three specific noise reduction methods and
circuits. In Section V we present simulation results for the
I. I NTRODUCTION
three circuits presented in Section IV, and in Section VI we
noise is the noise that occurs when the capacitor is reset prior II. T HERMAL AND S HOT N OISE IN C APACITIVE S ENSOR
to the charge accumulation cycle. Therefore, it is important C IRCUITS
to understand the factors which determine reset noise, and
Reset noise in capacitive sensors is typically caused by
how this noise may be mitigated. The purpose of this paper
thermal and shot noise sources. 1/f noise is not typically a
is to show how capacitive reset noise can be reduced during
dominant reset noise component. Thermal and shot noise are
the reset cycle. We present and analyze three circuits that
intrinsic characteristics of all resistors, transistors, and diodes.
implement the basic methods for directly reducing capacitive
The physical basis for thermal and shot noise in CMOS
reset noise. In addition, we present a time-domain technique
circuits derives from classical Boltzmann thermodynamics, i.e.
for analyzing the time–varying statistics of these circuits. This
the uncertainty in a system at thermal equilibrium is 12 kT
technique makes use of Itô calculus to obtain solutions to
per degree of freedom, where k is Boltzmann’s constant and
the time-varying stochastic differential equations. Itô calculus
T is absolute temperature. A complete description of noise
and time domain solutions to stochastic differential equations
in semiconductor devices can be found in [15]. We will
are not widely used in circuit analysis. However, researchers
assume that all of the noise sources analyzed in this paper
are starting to use these techniques to analyze time-varying
can be modeled as zero mean wide–sense stationary random
stochastic circuits [1]–[3]. Moreover, [2] provides a relevant
processes [14]. It can be shown that the thermal noise voltage,
Itô calculus analysis of sampling mixer circuits.
Vn , in series with a resistor r has (two-sided) power spectral
There is a large body of literature which presents specific
density
circuits and methods for reducing reset noise [1], [4]–[10]. In
SVn (f ) = 2kT r ∀ − ∞ < f < ∞. (1)
this paper, we construct a general framework for understanding
these results. We outline the mathematical methods that are The derivation of this result was first given by Nyquist [16].
Note that the channel noise of a MOS transistor, operated
Manuscript received dd mm, 2005. This work was partially supported by
a research grant to Stanford University from Fairchild Imaging, Inc. above threshold (gate voltage such that the drift current is
Boyd Fowler and Steve Mims are at Fairchild Imaging. Michael D. Godfrey dominant, as opposed to subthreshold where diffusion current
is at Information Systems Lab, Stanford University. dominates) in the linear region, has the same power spectral
This work has been submitted to the IEEE for possible publication.
Copyright may be transferred without notice, after which this version will density as Equation 1 with r being the channel resistance of
be superseded. the MOS transistor. The gate referred noise power spectral
2 RESUBMITTED TO: IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS I. DRAFT 1.1: 5 DEC. 2005
density of a MOS transistor, operated above threshold in the and f (x, y) is the joint distribution of X and Y. If either X(t)
saturated region, is or Y (t) has zero mean then Equation 5 simplifies to
SVg (f ) =
2kT γ
∀ − ∞ < f < ∞, (2)
2
σXY (t) = E[X(t)Y (t)]. (7)
gm
The variance, i.e. power, of a zero mean random process X(t)
where gm is the transconductance of the MOS transistor, and at time t is defined as the cross correlation of the process with
γ is a constant based on the process and transistor size.1 The itself,
channel of a MOS transistor operated in subthreshold can 2
σX (t) = E[X(t)X(t)]. (8)
be modeled by two independent random diffusion processes
which yield a forward and a reverse shot noise current in the If X(t) and Y(t) are vectors of random processes, then the
channel. When an MOS transistor in subthreshold is operated cross correlation matrix is
below saturation, i.e. the drain to source voltage is much ΣXY (t) = E[X(t)Y(t)T ] − E[X(t)]E[Y(t)]T , (9)
less than the thermal voltage kTq , both forward and reverse
diffusion currents are present and the resulting drain current and the autocorrelation matrix is
noise power spectral density is
ΣX (t) = E[X(t)X(t)T ] − E[X(t)]E[X(t)]T . (10)
SId (f ) = 2qid ∀ − ∞ < f < ∞, (3)
We define W(t) as a vector of independent identically dis-
where q is the charge of an electron, and id is the average drain tributed Wiener processes [14]. Note that each Wiener process
current. As the transistor approaches saturation the reverse is normally distributed with W (0) = 0, zero mean, and
current goes to zero and the noise power is reduced by 1/2, variance t ∀ 0 ≤ t < ∞.
i.e.
SId (f ) = qid ∀ − ∞ < f < ∞. (4) SW R Vc
random variable X. larger capacitors reduce the voltage noise. Conversely, smaller
The cross correlation function of two random processes at capacitors reduce the charge noise.
any time t is defined as
2
σXY (t) = E[X(t)Y (t)] − E[X(t)]E[Y (t)], (5) C. Physical Analogy of Capacitive Reset Noise
where Z ∞ Z ∞
E[XY ] = xyf (x, y)dxdy, (6) In order to understand the source of capacitive reset noise
and how to reduce it, we present a physical analogy for the
−∞ −∞
1γ
is typically set to 23 for long channel devices. switched RC circuit in Figure 1. The analogy is depicted in
2 It
may be helpful to point out that equation 1 in [17] contains a Figures 2, 3, and 4. Each Figure consists of two buckets of
typographical error which the authors corrected before publication, but which ping–pong balls connected by a ramp. (Two analogies are
was not included in the published version. Equation (1) should read as follows:
V
− Us
V
− Ud
V
− Us
V −V
− dU s possible here: the ping-pong balls which we will use, and
If ∝ e , Ir ∝ e , I = I f − Ir , I ∝ e 1−e =
a hydraulic analogy with water as the natural medium. The
T T T T
V
use of water, and water vapor, is closer to the physics of
− Uds
.
Isat 1 − e T
FOWLER, GODFREY, AND MIMS: RESET NOISE REDUCTION IN CAPACITIVE SENSORS 3
capacitive sensor. Then an “error” amplifier, in a feedback via cf between t3 and t4 . t2 − t1 is selected such that Vse
loop, is connected via a capacitor to the sensor. After the reaches steady state during hard reset, and t4 − t3 is selected
feedback loop settles, attenuating noise on the capacitive such that the feedback loop reaches steady state during the
sensor, the loop is opened and the sensor is ready to detect reset noise reduction step. t3 − t2 only needs to be greater
signals again. The noise stored on the capacitive sensor is than zero.
attenuated because the feedback loop acts to minimize the
difference between the voltage on the sensor and a fixed
reference voltage, i.e. the error amplifier tries to minimize rst
vref-bit (see Figure 5). t t
1 2
Vreset
en
pixel
t t
en Vst
M1
3 4
c st Vse
Mf1
c
f c se
time
Vdd word
Fig. 7. Capacitive Control Timing Waveforms
Vo bit
vref
op1 When hard reset is complete and t is within t2 < t < t3 the
c
l 1pF capacitive sensor has a noise voltage power equal to kT /(cse +
2uA
cf cst /(cf + cst )). After the feedback loop is connected,
allowed to settle, and then opened, the voltage noise σV2 se , and
Fig. 5. Capacitive Control Schematic the charge noise σq2se = (cse + (cf cst /(cf + cst ))2 σV2 se on the
capacitive sensor can be determined by finding the covariance
matrix of
Vo (t)
V(t) = Vst (t) . (11)
op1
Vse (t)
Vo r en Vst c Vse
f
+ −
V V(t) satisfies the following set of SDEs
g n1
m g
o c
l c c dV(t) dW(t)
st se A = BV(t) + C (12)
dt dt
+ r rst where
V + −
n2 V cl 0 0
n3
A = 0 cf + cst −cf , (13)
0 −cf cse + cf
Fig. 6. Simplified Capacitive Control Model 1 1
− r − go r −gm
B= 1
r − 1r 0 , (14)
A circuit that implements capacitive control is shown in 0 0 0
Figure 5. This circuit is a single pixel in an image sensor with and
its associated column level reset amplifier. The photodiode √
q
2kT
2kT gmγ
capacitance cse is reset by controlling M1 via Vreset and qr
(15)
vref, while the switches en and word are closed. The reset C=
− 2kT
0
.
r
process is initiated by setting Vreset to ground while Vst 0 0
is held at about Vdd/2, this forces the voltage across cse
to ground. Then Vreset is raised to approximately Vdd, Note that W(t) is a column vector of two independent and
and after a short delay vref is lowered just enough, tens identically distributed 1-D Wiener processes. dW (t)/dt is
of millivolts, to turn off M1 and control cse via the parasitic a white noise process. The matrix C is derived from the
gate to source capacitance of M1 . After the feedback loop has noise sources Vn1 , Vn2 , and Vn3 in Figure 6. To simplify the
settled the switches en and word are opened. Before or after notation, let
reset the voltage across cse can be read–out by operating Mf 1 D = A−1 B, (16)
in source follower mode, while word is closed and en is and
open. E = A−1 C. (17)
A simplified linearized model of this circuit, for analyzing
reset noise, is shown in Figure 6, and the timing waveforms Then the expected value of Equation 12 is
for the model are shown in Figure 7. Hard reset is performed dE[V(t)]
between t1 and t2 , and the feedback loop is connected to cse = DE[V(t)]. (18)
dt
FOWLER, GODFREY, AND MIMS: RESET NOISE REDUCTION IN CAPACITIVE SENSORS 5
ΣV (0) = 0 0 0
. (24) cf , but it is much more difficult to control the final number of
kT
0 0 cf cst ping–pong balls in cse . Finding the optimal ratio of cse
cf
is an
cse + cf +cst
essential part of designing any circuit that utilizes this reset
After solving Equation 23, and taking the limit as t → ∞, it noise reduction technique. In addition, accurate measurement
can be shown that the noise voltage on the capacitive sensor of cse after hard reset, and accurate control of cf are important
is for optimal circuit performance.
c (rgo2 +γgm )+cl (go +αgm )
σV2 se = kT ctt(go +αg m )(ct (1+rgo )+cl )
α2 +
(cse +cf )go
2 (25) B. Bandwidth Control
kT cse go +c f go +cf gm
,
Reducing reset noise using bandwidth control is a two step
where ct = cst + αcse and α = cf /(cf + cse ). The use of the process. First, hard reset is performed on the capacitive sensor.
asymptotic limit t → ∞ in this result does not allow solution Then an error amplifier in a feedback loop is connected via
by standard frequency methods since the asymptotic behavior a time varying resistor to the capacitive sensor. Finally, the
depends on the initial conditions. Note that the first term in resistance of the time varying resistor is increased until the
Equation 25 is the noise added by the feedback amplifier and bandwidth of the error amplifier is much larger than the noise
the en switch, and the second term is the initial hard reset bandwidth of the resistor.4 Note that the noise bandwidth of the
noise on the capacitive sensor after it is attenuated by the time-varying resistance is approximately 1/(4r(t)(cse + cf )).
feedback loop. For example, if we assume that T = 300 K, In addition, it is necessary that r(t) → ∞ as t → ∞, i.e.
cl = 0.3 pF, γ = 28, cf = 0.4 fF, cse = 5.4 fF, cst = 5 fF, the resistor that connects the error amplifier to the capacitive
gm = 17.15 µS, go = 84.7 nS, and r = 10 kΩ then sensor must become disconnected by the end of the reset cycle.
σV2 se = 31 nV2 and σqse /q = 6.4 e- RMS.
A circuit that implements bandwidth control is shown in
Note that (kT cse )1/2 /q = 29.5 e- RMS. Figure 9. This circuit is a single pixel in an image sensor. The
In order to help clarify how this circuit reduces reset noise, circuit operates in two modes: the first is when the photodiode
we present an analogy based on Section III. This analogy is being reset and the second is when the photodiode is
is intended to help visualize the noise reduction process being read–out. During reset the photodiode capacitance c se is
described in this subsection. Imagine that the capacitive sensor, controlled by the reset transistor M1 and the common source
cse , is a bucket containing a finite number of ping–pong balls, transistor Mf 1 via reset, while the switches breset and
and that cf is another bucket used to either add or subtract word are closed. vref is typically held at 1V. The reset
ping–pong balls (by tipping the cf bucket) to cse (see Figure process is initiated by raising reset from ground to a voltage
8). At some time t, after hard reset, the number of ping–pong
4 If the resistor’s noise is to be attenuated, the bandwidth of the error
3 The initial noise voltages on Vo and Vst have almost no effect on the final amplifier must be larger than the noise bandwidth of the time varying resistor
steady state noise voltage on Vse . during a significant period within the reset cycle.
6 RESUBMITTED TO: IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS I. DRAFT 1.1: 5 DEC. 2005
Vdd
1/r(t)
breset
2uA
Vo time
pixel c breset
reset l
M1 c
f
Vse t t t
Mf1 1 2 3
c
se Fig. 11. Bandwidth Control Timing Waveforms
word
bit
and q √
2kT
breset r(t) 2kT γgm
1pF C= . (29)
2uA
q
2kT
− r(t) 0
vref
As in Subsection IV-A, it can be shown that V(t) is an Itô
Fig. 9. Bandwidth Control Schematic process, and assuming the initial condition
E[V(0)] = 0, (30)
c
f
op1
then
breset V
o r(t) Vse E[V(t)] = 0 ∀ t ≥ 0. (31)
+ −
V
n1
g
m g
o
Similarly, since E[V(t)] = 0, ΣV (t) = E[V(t)V(t)T ]. Using
c c
l breset se Equation 23 with the initial condition
0 0
+
V ΣV (0) = (32)
n2 0 csekT+cf
−6
BC Log Vse Noise Power vs Time C. Charge Control
10
tau=0.25us
Reducing reset noise using charge control is a three step
log(Vse noise power)
tau=0.5us
−7
10 tau=1us process. First, the voltage across the capacitive sensor is hard
tau=2us reset to Vreset. Then, a current source is connected to the
tau=4us
−8
10 capacitive sensor causing the voltage Vse to fall. Finally, when
the voltage across the capacitive sensor falls below a reference
−9
10 voltage the current source is disconnected from the capacitive
0 0.5 1 1.5 2 2.5
Time −5 sensor.
x 10
BC Log Vo Noise Power vs Time
−4
10 Vreset
tau=0.25us pixel
log(Vo noise power)
−8 bit
10 Vo
0 0.5 1 1.5 2 2.5 vref
Time x 10
−5
cmp1 1pF
2uA
Vreset reset
C
se r
− + V
dx se
V
t 0
n2 c
se
Fig. 13. Ping-pong Analogy: Bandwidth Control A circuit that implements charge control is shown in Fig-
ure 14. This circuit is a single pixel in an image sensor with
balls in cse is a random value. Assuming we can measure the a column level reset comparator. The circuit operates in two
number of ping–pong balls to a sufficient accuracy while the modes: the first is when the photodiode is being reset and the
width of the ramp is slowly being reduced, the difference in second is when the photodiode is being read–out. During reset
ping–pong ball levels between cse and the infinite bucket can the photodiode capacitance cse is first set to Vreset via M1
be used to control the final number of ping–pong balls in cse . while word is closed. vref is typically held at 1 V. Then
For example, when the ramp width is exactly wide enough vbias is raised to approximately 0.5 V and current flows
for one ping–pong ball, the slope of the ramp can be used to from cse to Vo, slowly discharging cse . When bit is less
control the average number balls that are either added to or than vref the output of the comparator switches high and M2
subtracted from the bucket. If we can adjust the slope of the is turned off. After reset is complete vbias is set to ground.
ramp faster than ping–pong balls can jump onto or off of the Before or after reset, the voltage across cse can be read–out by
ramp due to thermal energy, we can, on average, reduce the operating Mf 1 in source follower mode, while word is closed.
uncertainty in the number of ping–pong balls in cse . Finding A simplified model of the circuit, used for analyzing reset
the best waveform for r(t), i.e. the width of the ramp as a noise, is shown in Figure 15. We assume that the input-referred
function of time, is an essential part of designing any circuit noise of the comparator, Vn1 (t), is a Rwhite random process
with zero mean and variance σV2 n1 , and 0 Ir q(τ ) dτ , is a Poisson
t
that utilizes this reset noise reduction technique. In addition,
accurate measurement of cse after hard reset, and accurate process with parameter iqr . The voltage across the capacitive
control of r(t) are important for optimal circuit performance. sensor just after the current source is disconnected, i.e. t >
8 RESUBMITTED TO: IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS I. DRAFT 1.1: 5 DEC. 2005
into cse (see Figure 16). After hard reset, we fill cse up to
2
Volts
vref
1.5
I (t)
r Vse
1
0.5 en
0
0 5e-06 1e-05 1.5e-05 2e-05 2.5e-05
time
C
se Fig. 17. Capacitive Control Input/Output Waveforms
Vse_noise
1e-07
Circuit #1, Vse Noise Power
1e-06
Vse_noise
Volts^2
1e-08
1e-07
Volts^2
1e-08
1e-09
0 5e-06 1e-05 1.5e-05 2e-05 2.5e-05
time
Figure 19 shows the SPICE input waveforms used for Circuit #3, control and output signals
2.5
Vse
Circuit #2, control and output signals
breset 2
Vse
3 breset reset
Volts
1.5
2.5
reset
1
0.5
Volts
Vse
1.5
0.5
Fig. 21. Charge Control Input/Output Waveforms
0
0 5e-06 1e-05 1.5e-05 2e-05 2.5e-05
of the capacitor, transistor and amplifier characteristics for
time Figure 14. Figure 22 shows the transient noise simulation of
Fig. 19. Bandwidth Control Input/Output Waveforms the charge control circuit. The noise power σV2 se after 20 µs is
0.1 µV2 . Therefore the RNRF of this circuit is 0.62/0.1 = 6.2.
all of the capacitor, transistor and amplifier characteristics for
Figure 9. Figure 20 shows the transient noise simulation of A summary of the simulation results compared with the-
the bandwidth control circuit. The hard reset noise voltage oretical calculations from the previous Section are shown in
power on the capacitive sensor σV2 se at 5 µs is approximately Table IV. Note that CC is capacitive control, BC is bandwidth
0.4 µV2 . This result corresponds to about 60% of the predicted control, and CHC is charge control.
hard reset value csekT+cf . The noise power σVse after 25 µs is
2
Component Parameter Value units units Fowler [5] Pain [7] Lee [8] Kozlowski [10]
cse capacitance 6.7 fF RNRT CC BC & CHC CHC BC
M1 W/L 0.42/0.6 µm RNRF 18 40 3.9 14
M1 threshold voltage 0.45 V Reset time µs 25 5 40 10
Mf 1 W/L 0.7/1.2 µm Pixel size µm2 21×21 10×10 5.9×5.9 5×5
Mf 1 threshold voltage 0.45 V Tech µm 0.35 0.5 0.18 0.25
Mf 1 transconductance 21.3 µS Lag %FS 0.02 0 – 0.012
Mf 1 body effect conductance 6.2 µS Swing V 1 0.975 0.4 0.8
mf 1 output conductance 49 nS Vdd V 3.3 5.0 1.8 3.3
cmp1 gain bandwidth 167 MHz
cmp1 DC gain 2311 TABLE V
cmp1 input referred noise 5.79 nV2 M EASURED R ESULTS FROM THE L ITERATURE
vref voltage 0.85 V
Ir current 0.23 nA
TABLE III
C HARGE C ONTROL S IMULATION C OMPONENT PARAMETERS VII. C ONCLUSION
We have presented a theoretical framework and computa-
Circuit #3, Vse Noise Power
tional methods for understanding and modeling reset noise
0.0001
Vse_noise in capacitive sensors. The methods have been demonstrated
1e-05
on several circuits which are currently in use. It is clear that
there are many issues in deciding on a “best choice” for a
1e-06 specific sensor circuit. These include: reset noise reduction
factor (RNRF), reset speed, size/complexity, signal swing, and
Volts^2
1e-07
others. We have summarized the key characteristics of the
1e-08
three example circuits in Table IV.
It appears that circuits which provide noise reduction
1e-09
0 5e-06 1e-05 1.5e-05 2e-05 2.5e-05
through feedback are necessarily time varying and often non-
time
linear. The circuit’s noise behavior cannot be well-modeled
Fig. 22. Charge Control Transient Noise Waveforms
using linear time invariant approximations. This leads to the
conclusion that either Itô calculus or Monte Carlo simulation
methods are required for analysis. The Itô calculus is not
competing circuit parameters. These other parameters include widely used in circuit analysis, but it is particularly well-suited
fixed pattern noise (or voltage offset), lag, signal swing, circuit for the circuits presented here. Moreover, the Itô calculus
stability, and reset speed. Balancing these tradeoffs is the key method is computationally much more efficient than methods
to producing low noise capacitive sensor circuits which best based on Monte Carlo sampling for performing transient noise
meet specific needs . Since we have focused on understanding analysis. In addition, the Itô calculus method can be used to
the RNRF in this paper, measured results from the literature obtain both closed form and numerical noise power solutions
are presented in Table V to help the reader understand some of for time varying and/or non–linear circuits.
the design tradeoffs. In a future paper we plan to analyze these We hope that these results will contribute to the understand-
other design parameters. Table V presents circuit parameters ing, evaluation, and implementation of reset noise reduction
and measured results from four circuits that use different reset techniques.
noise reduction techniques (RNRT). Note that Lag is defined
in [5], and Swing is defined as the maximum voltage swing A PPENDIX
across a capacitive sensor. The results presented in Table V D ERIVATION OF C HARGE C ONTROL N OISE
clearly show that lag is not a problem for any of the reset noise
reduction techniques. Reset noise reduction techniques typ- In this Appendix we derive Equations 36, 37, and 38. This
ically increase reset time from nano/pico–seconds to micro– requires us to determine the mean and variance of Vse (t) when
seconds. These circuits can be implemented using a wide range t > Tsw + ∆T . The voltage across cse , when t > Tsw + ∆T ,
of CMOS geometries. Although it is not shown in Table V, is Z Tsw +∆T
1
our experience indicates that BC techniques often suffer from Vse = Vse (0) − Ir (τ )dτ, (40)
stability problems associated with the time varying nature of cse 0
the time constants in the feedback loop. where Tsw + ∆T is the duration of time the current source is
connected to the sensor, and Vse (0) is the initial voltage on
the capacitive sensor after hard reset. Note that Tsw and ∆T
CC BC CHC
Simulated RNRF 21.8 9.5 6.2 are assumed to be a random variables. This integral can be
Theoretical RNRF 23.2 11.7 7.0 separated into; the voltage at which the comparator switches,
%RNRF difference between 6.5% 23.1% 12.9% Vse (Tsw ), and the integral of the current source while the
simulation and theory values
comparator is switching,
TABLE IV
Z Tsw +∆T
S UMMARY OF S IMULATION R ESULTS 1
Vse = Vse (Tsw ) − Ir (τ )dτ. (41)
cse Tsw
FOWLER, GODFREY, AND MIMS: RESET NOISE REDUCTION IN CAPACITIVE SENSORS 11
We assume that the input-referred noise of the comparator, 2i2r ∆t2 i2r ∆t2
+ (59)
Vn1 (t), is aR white random process with zero mean and variance c2se c2se
σV2 n1 , and 0 Ir q(τ ) dτ , is a Poisson process with parameter iqr .
t
qir ∆t + i2r σ∆T
2
= σV2 n1 + , (60)
Note that Z t 2
cse
Ir (τ ) ir t
E dτ = (42) where equations 50 and 51 lead to equations 52 and 53 because
q q
Vse (Tsw ) and 0 Ir q(τ ) dτ are uncorrelated. Note that qse =
0
R ∆T
and "Z 2 # cse vse , and therefore the input referred charge noise is
t
Ir (τ ) ir t ir t
E dτ − = . (43) σq2se = c2se σV2 se (61)
0 q q q
qir ∆t + i2r σ∆T
2
= c2se σV2 n1 + (62)
Since Ir q(τ ) dτ is a Poisson process we can further simplify
R
c2se
Equation 41 using the independent increment property, i.e. 2 2
= cse σVn1 + qir ∆t + i2r σ∆T
2
. (63)
Z ∆T
q Ir (τ )
Vse = Vse (Tsw ) − dτ. (44) ACKNOWLEDGMENT
cse 0 q
We thank Tobi Delbrück, Abbas El Gamal, Dana How, John
Defining ∆t = E[∆T ] and σ∆T 2
= E[(∆T − ∆t)2 ], using the
Lazarro, Chiao Liu, Dick Lyon, and Sam Kavusi for their
assumption that all the random variables and processes are
helpful comments.
uncorrelated, and using conditional expectations, we find
h q Z ∆T I (τ ) i R EFERENCES
dτ (45)
r
E[Vse ] = E[Vse (Tsw )] − E
cse 0 q [1] H. Tian, B. Fowler, and A. El Gamal, “Analysis of Temporal Noise
= E[vref + Vn1 ] − (46) in CMOS APS,” in Proceedings of SPIE, vol. 3649, San Jose, January
1999.
[2] W. Yu and B. H. Leung, “Noise Analysis for Sampling Mixers Using
" #
h q Z ∆T I (τ ) i
(47)
r
E E dτ ∆T Stochastic Differential Equations,” IEEE Trans. Circuits and Systems II,
cse 0 q vol. 46, no. 6, pp. 699–704, June 1999.
" # [3] D. Ham and A. Hajimiri, “Complete Noise Analysis for CMOS
ir ∆T Switching Mixers via Stochastic Differential Equations,” in CICC2000,
= vref − E (48) Orlando, Florida, May 2000.
cse
[4] B. Pain et al., “Analysis and Enhancement of Low-light-level Per-
ir ∆t formance of Photodiode-type CMOS Active Pixel Imagers Operated
= vref − . (49) with Sub-threshold Reset,” in 1999 IEEE Workshop on CCDs and AIS,
cse Nagano, June 1999.
Note that equation 45 leads to equations 46 and 47 because [5] B. A. Fowler, M. D. Godfrey, J. Balicki, and J. Canfield, “Low-noise
Readout Using Active Reset for CMOS APS,” in Proceedings of SPIE,
E[Vse (Tsw )] is the voltage at which the comparator switches vol. 3965, San Jose, January 2000, pp. 126–135.
and E[g(X, Y )] = E[E[g(X, Y )|Y ]] (this can be shown [6] W. Loose et al., “2/3in CMOS Image Sensor for High Definition
using equation 6 and f (x, y) = f (x|y)f (y)). Using these Television,” in 2001 IEEE Workshop on CCDs and AIS, Lake Tahoe,
Nevada, June 2001.
assumptions we find: [7] B. Pain et al., “Reset Noise Suppression in Two-Dimensional CMOS
" Photodiode Pixels through Column-based Feedback-Reset,” in 2002
IEDM, 2002, pp. 809–811.
2
σVse = E Vse (Tsw ) − (50) [8] K. Lee and E. Yoon, “A CMOS Image Sensor with Reset Level Control
Using Current Source for Noise Suppression,” in 2004 ISSCC Digest of
Z ∆T 2 # Technical Papers, San Francisco, CA, February 2004, pp. 114–115.
q Ir (τ ) ir ∆t [9] Y. Chen and S. Kleinfelder, “CMOS Active Pixel Sensor Achieving 90
dτ − vref + (51)
cse 0 q cse dB Dynamic Range with Column-level Active Reset,” in Proceedings
of SPIE, vol. 5301, San Jose, January 2004, pp. 438–449.
[10] L. Kozlowski et al., “A Progressive 1920×1080 Imaging System–on–
h i
= E (Vse (Tsw ) − vref )2 + (52)
Chip for HDTV Cameras,” in ISSCC Digest of Technical Papers, San
h q
Z ∆T
Ir (τ ) ir ∆t 2 i Francisco, February 2005.
E − dτ + (53) [11] Z. Brzezniak and T. Zastawniak, Basic Stochastic Processes. London:
cse 0 q cse Springer, 1958.
[12] R. Durrett, Stochastic Calculus. Boca Raton, Florida: CRC Press, 1996.
" Z ∆T
h q Ir (τ ) [13] A. Demir, E. W. Y. Liu, and A. L. Sangiovanni-Vincentelli, “Time-
= σV2 n1 + E E − dτ + (54)
cse 0 q Domain Non-Monte Carlo Noise Simulations for Nonlinear Dynamic
Circuits with Arbitrary Excitations,” IEEE Transactions on Computer-
aided Design of Integrated Circuits and Systems, vol. 15, no. 5, pp.
#
ir ∆t 2 i
∆T (55) 493–505, May 1996.
cse [14] K. S. Shanmugan and A. M. Breipohl, Random Signals: Detection,
" Estimation and Data Analysis. New York: John Wiley & Sons, 1988.
1 [15] A. van der Ziel, Noise in Solid State Devices. New York: John Wiley
2
= σVn1 + E 2 (i2r ∆T 2 + qir ∆T ) − (56) & Sons, 1986.
cse [16] H. Nyquist, “Thermal Agitation of Electric Charge in Conductors,” Phys.
Rev., vol. 32, pp. 110–113, July 1928.
#
2i2r ∆t∆T i2r ∆t2 [17] R. Sarpeshkar, T. Delbrück, and C. A. Mead, “White Noise in MOS
+ 2 (57)
c2se cse Transistors and Resistors,” IEEE Circuits and Devices Mag., pp. 23–29,
November 1993.
qir ∆t + i2r (σ∆T
2
+ ∆t2 ) [18] C. Mead, Analog VLSI and Neural Systems. New York: Addison-
= σV2 n1 + − (58) Wesley, 1989.
c2se
12 RESUBMITTED TO: IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS I. DRAFT 1.1: 5 DEC. 2005
R ESPONSE TO R EVIEWERS OF D RAFT 1.0 attempted a full description of the Itô Calculus, but we have
tried to clarify specifically why it is required, and we have
A. Reviewer 1
included the useful reference [2], (2) we have clarified the
Reviewer Said: ping-pong analogy, but we stopped short of any implication
From the standpoint of someone not familiar with Itô that ping-pong balls are really like big electrons – we think
calculus, the presentation of this material is too brief to offer this analogy is helpful, but it is not exactly right, (3) we have
useful insight to the average reader. Particularly, the inclusion moved some of the mathematics to an Appendix.
of the Wiener vector W(t) lacks sufficient explanation, and 2) ¶1: Within the space limits, we have tried to clarify the
the entire workup hinges on the “Itô formula” of equation introduction of the Wiener process and the use of Itô Calculus.
20, which is pulled from a reference on the subject. As a 3) ¶2: We made the circuit descriptions more specific and
reader attempting to understand what this technique offers, precise – we think this will be found quite helpful. We did
even intuitively, the presentation of this material in its current not extend the analysis to such issues as timing, but we have
form is not very useful. tried to be more specific about what we do present.
The description of circuit operation is too vague. Without 4) ¶3: We have made the ping-pong ball analogy more
past familiarity with these techniques I would have had a diffi- specific to the extent that we think is possible. It is not good,
cult time understanding the circuits’ behavior without referring we think, to push this too far.
to the references. For example, in Fig. 5 the DC operating
point of Vst could lie anywhere between the rails depending B. Reviewer 2
on the exact circuit parameters and the result of the previous
noise-cancellation cycle. Would this then allow M1 to turn on Reviewer Said:
the next cycle? Also, how can we be sure that M1 operates Excellent work!
in the triode region without some limitations on Vreset and Response
Vst? Without this, you won’t necessarily get circuit settling Thanks.
between t1 & t2. There are many other small issues like this
that I ran across in the paper. In general, I think that the reason C. Reviewer 3
for this is that your circuits are specific topologies pulled from Reviewer Said:
photodetector applications, with corresponding assumptions on This paper is an interesting application of the time-domain
voltages, timing, etc. (that are not discussed here), and are stochastic differential equations (SDE) to the analysis of three
overly complicated for the concepts you are trying to convey. reset noise reduction techniques. The paper is concise and
I would recommend reducing their complexity, which would well-written.
allow you to keep your explanation of their operation as brief The main contribution of this paper is an SDE analysis
as possible without confusing the reader. for capacitive reset noise. The time-domain SDE approach
I do like the idea behind the ping-pong analogy, and think and the solution technique are essentially the same as that
that it could be useful in understanding circuit noise behavior. in reference [16] and the following paper (which should also
A few comments on this stuff for you. For Fig. 8, the analogy be referenced):
seems a little shaky because it does not allow for a sort of W. Yu and B. H. Leung, “Noise analysis for sampling mixers
capacitive current effect. What I mean, is that Cf can start with using stochastic differential equations,” IEEE Trans. Circuits
0 net charge and either add or subtract charge from Cse, but in and Systems II, vol.46, no.6, June 1999, pp.699-704.
your model Cf can only add if balls are present in it to begin Although the SDE techniques themselves are not new, the
with. Also, even if Cf is smaller than Cse, you can completely application of SDE to reset noise reduction technique is a very
fill or drain Cse of charge given large enough rails, but you interesting one. The paper represents a novel contribution.
cannot do this in your analogy. Also, I think there may be The paper analyzes three different reset noise reduction
something a little fishy in drawing a parallel between a finite techniques. My main technical comments are as below:
volume and an ideal capacitor since an ideal capacitor has 1. On the “Bandwidth Control” part of the paper in which
infinite charge capacity (without rail limitations) while your a time-varying SDE is used to characterize the reset noise: As
model affords it only finite ball capacity. You’d really need a the authors point out in equations (25)-(27), the coefficients
finite cross-section, infinite height box, right? Then, capping of the SDE depend on a time-varying parameter r(t). This by
the box off at two levels corresponds to your upper and lower itself does not represent a technical difficulty, as the covariance
rails, which don’t move (you can’t raise and lower boxes as process of a stochastic process governed by a time-varying
you did in some illustrations, you’d have to reduce the volume SDE can be converted into a deterministic ordinary differential
if you do). The ramp would also need to slide along the box equation in the same way as in equations (20)-(21), except in
to stay pinned at the height of the top ping-pong balls for the this case the coefficients (i.e. the equivalent D and E) are time
voltage analogy to hold, rather than moving the height of the varying. Such a time-varying ODE can be easily simulated
bottom of the boxes (moving the bottom would be like moving if not solved exactly. The simulation of the covariance ODE
ground). is much faster than a Monte Carlo simulation of the SDE.
Response Instead of taking this approach, the paper assumed that r(t) is
1) General: We have made numerous changes in response a constant and chose to instead present a closed-form solution,
to these comments. About the major points: (1) we have not which is not entirely accurate. Why? A revision of the paper
14 RESUBMITTED TO: IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS I. DRAFT 1.1: 5 DEC. 2005