Measurement 115 (2018) 95–103

Contents lists available at ScienceDirect

Measurement
journal homepage: www.elsevier.com/locate/measurement

Maximum likelihood estimation of aperture jitter using sinusoidal excitation MARK

Tamás Virosztek
Budapest University of Technology and Economics, Department of Measurement and Information Systems, H-1117 Magyar tudósok körútja 2, Budapest, Hungary

A R T I C L E I N F O A B S T R A C T

Keywords: This paper proposes a solution to estimate the aperture jitter of analog-to-digital (ADCs) converters using the
Nonideal sampling maximum likelihood (ML) approach. The idea is to extend the model built for the ML estimation of ADC and
Aperture jitter excitation signal parameters to handle the uncertainty of sampling. Using the Gaussian model of aperture jitter
Parameter estimation this extension requires only one additional parameter. This research article investigates the feasibility of this
Maximum likelihood
extension, gives the extended likelihood function in closed form and proposes a slight approximation to decrease
ADC testing
computational demand significantly. The increased explanatory power of the extended model is shown via
evaluation of simulated and real measurements.

1. Introduction
There are several methods to quantify the imperfection of real
analog-to-digital converters. These techniques are estimating some
parameters of the device under test and then calculate the static or
dynamic quality measures which describe the converter from a certain
point of view. Many of these methods are canonized and appear in
standards e.g. in [1] or [2]. There are other techniques which can be
advantageous in multiple use cases, but are not standardized yet. The
maximum likelihood estimation of ADC and excitation signal para-
meters is one of these. This approach has several attractive properties
such as consistency, effectiveness and asymptotic normality [3,4]. On
the other hand – owing to its probabilistic approach and the more de-
tailed model – the computation demand is usually larger and the re-
quirements regarding the design of the measurement are more strict in
the case of ML estimation compared to the standardized techniques. In
conclusion ML estimation is recommended in those cases where the
accuracy of estimation is crucial and the number of measurements is
large enough to utilize the good asymptotic properties of the ML
method. [5] provides a comparative performance analysis of ML and
other methods to emphasize the use cases where ML estimation is re-
commended to achieve satisfying ADC test results.
The model developed for ML estimation of ADC and excitation
signal parameters [6] assumes ideal sampling, thus does not handle
aperture jitter. This approach can be useful in several cases where the
stimulus is slow and noisy thus the effect of aperture jitter can be ne-
glected compared to the effect of the additive noise. However, in the
case of a faster and less noisy signal the aperture jitter becomes sig-
nificant and results a time-variant, amplitude-modulated noise com-
ponent. Modeling the uncertainty of sampling has two important
advantages: on the one hand it increases the explanatory power of the
model, on the other hand it provides an ML estimate for the aperture
jitter which can be important for the users of the ADC under test. The
aperture jitter can be estimated different ways besides the ML method
as it is described in [7,8] or [9]. Standard IEEE-1241 [1] also proposes a
method to measure aperture jitter, but it requires a special measure-
ment setup dedicated to estimate the amount of jitter. Nevertheless
these estimation methods do not provide the attractive properties that
the ML estimation does. Furthermore the suggested ML method esti-
mates the jitter using a simple sinusoidal measurement record while
also estimates multiple other parameters of the stimulus and the device
under test – thus makes possible to calculate dynamic quality measures
such as ENOB or SINAD. This paper proposes an extension of the model
developed for ML estimation which handles – and thus estimates in ML
sense – the aperture jitter while the parameter space grows only by one
additional parameter. The details of this extension are itemized in the
following sections.
2. The model
The measurement setup for sinusoidal ADC testing is simple. A
dedicated sine wave generator is connected to the input of the ADC
under test that records this stimulus. It is recommended to use a slow
( f ≈ 10 Hz…100 Hz ) sine wave, the length of the record is usually a few
seconds. Depending on the sampling rate this measurement leads to a
digital record containing a few hundred thousand or a few million
samples. Since noise is part of the model – the amount of noise is also a
parameter to be estimated – EMC protection and avoiding disturbances
is not particularly important. Naturally more noise leads to larger
variance in the estimation, thus it is recommended to pursue low-noise

E-mail address: virosztek@mit.bme.hu.

http://dx.doi.org/10.1016/j.measurement.2017.10.026
Received 17 May 2017; Received in revised form 9 October 2017; Accepted 11 October 2017
Available online 12 October 2017
0263-2241/ © 2017 Elsevier Ltd. All rights reserved.
T. Virosztek Measurement 115 (2018) 95–103

environment, but not crucial. The measurement noise is consequence of The objective function to be optimized is usually the negative log-
different phenomena. The main components of it are. likelihood function which is also known as cost function:

• the electronic noise of the sine wave generator,

M
⎡ ⎛ Ty [k] + 1−x [k ] ⎞
CF(p) = −lnL(p) = M ln2− ∑ ln ⎢erf ⎜
• the electromagnetic disturbances superimposed to the signal due to
⎟
k=1 ⎣ ⎝ σ ( 2) ⎠
the cable and
• the elctronic noise of the amplifier in the sample and hold subcircuit ⎛ Ty [k]−x [k ] ⎞ ⎤
−erf ⎜ ⎟ ⎥.
of the ADC. ⎝ σ ( 2) ⎠ ⎦ (6)
Using the log-likelihood function is advantageous regarding the
Note that these three categories cover many individual noise
optimization process since the calculation of the derivatives becomes
sources: the electronic noise is result of several different semiconductor
less complex and numerically less tense.
components and the EM disturbances are due to multiple devices
(transformers, switch mode power supplies, etc.). Owing to the dif-
ferent sources of disturbance the noise profile of the measurements is 3. Extension of the ML estimation framework to model nonideal
largely balanced: the presence or absence of a certain noise source does sampling
not have large influence on the overall noise profile.
The record of the sampled and quantized noisy sine wave is being The phenomena owing to nonideal sampling are not considered in
processed during the evaluation. The parameters of the model — ac- the model originally for ML estimation of sine wave and ADC para-
cording to the notation used in [6] – are the following: meters. The original article [6] introduces the additive noise on the
analog excitation as a parameter to estimate and the effect of nonideal

• Parameters of the sinusoidal excitation: x sampling – which can be handled using the term aperture jitter – is not
separated from the additive noise on the analog signal. The uncertainty
(t ) = Acos(2πft ) + B sin(2πft ) + C and x n (t ) = x (t ) + n (t ) , where A
is the cosine coefficient, B is the sine coefficient of the sine wave, C of sample y [k ] owing to the aperture jitter is
stands for the DC component of the stimulus and the frequency is dx (t )
denoted by f. The additive noise n (t ) is assumed to be Gaussian Δy [k ] ≃ ·Δtk,
dt t = tk,ideal (7)
white noise [6] with zero mean and σ standard deviation. Note that
the frequency is usually expressed as sampled angular frequency: where x (t ) is the noiseless sinusoidal excitation signal, tk,ideal is the ideal
θ = 2πfTs , where Ts is the nominal sampling time. sampling time corresponding to the kth sample, tk is the real sampling
• Parameters of the quantizer: the nonideal quantizer is described moment of tk and Δtk is the difference of the real and the ideal moment
of sampling for the kth sample: Δtk = tk−tk,ideal . Assuming that the Δtk
by its code transition levels. The kth code transition level is the DC
voltage which results output code k−1 with 50 % probability and values can be characterized by the same probability distribution the
code k with 50 % probability when applied to the input of the additive noise owing to aperture jitter can be modeled as amplitude-
quantizer. An N-bit quantizer has 2N −1 code transition levels, these modulated noise. Fig. 1 illustrates this model: the Δtk values are in-
are denoted by T1,T2,…,T 2N − 1. dependent realizations of a random variable and Δy [k ] values are cal-
culated as described in (7).
Therefore the parameter vector is The standard deviation of the additive noise owing to the jitter can
be calculated as
pT = [ A B C θ σ T1 … T 2N − 1]. (1)
dx (t )
To express the likelihood of a given parameter set we have to in- σn,t = ·σt ,
dt rms (8)
troduce a discrete random variable Y [k ]. This random describes gives
the probability distribution of the kth sample assuming the parameters where σt is the standard deviation of the distribution that provides the
contained by p . The P[Y [k ] = l] quantity is the probability of the in- Δt values. Let us introduce μt for the expected value of this distribution.
cidence the kth sample noisy sine wave x n (t ) is between Tl and Tl + 1. Since the sampling usually does not have systematic error components,
Owing to the Gaussian noise model we have to use the error function to dx (t )
we assume μt = 0 . In (8) dt is the root mean square value of the
express these probabilities: rms
dx (t )
x
slope of the noiseless excitation signal. Since is sinusoidal as well,
2 dt
erf(x ) = ∫0 e−t dt . (2) its RMS value is

Hence the probability distribution of the discrete random variable dx (t ) 2πfA

= ,
Y [k ] can be expressed as dt rms 2 (9)

1 ⎡ ⎛ Tl + 1−x [k ] ⎞ where A is the amplitude and f is the frequency of x (t ) . Since the ex-
⎛ Tl−x [k ] ⎞ ⎤
P[Y [k ] = l] = erf ⎜ ⎟−erf ⎜ ⎟ ⎥, citation signals used for ADC testing are usually slow sine waves
2 ⎢ ⎝ σ ( 2) ⎠ ⎝ σ ( 2) ⎠ ⎦ (3)
⎣ ( f < 100 Hz) and the aperture jitter of the modern S/H circuits is low
where x [k ] is the kth sample of the noiseless sine wave: (typically a few picoseconds and very rarely larger than 50 ps, see [10])
x [k ] = Acos(kθ) + B sin(kθ) + C . Note that this model assumes ideal, the additive noise owing to jitter can be majorized as
equidistant sampling. Exploiting the independence of the noise samples dx (t ) 2π·100Hz·10V
the likelihood function of the measurement is σn,t ,max = ·σt ,max = ·50ps = 0.2221μ V.
dt rms,max 2 (10)
M
L(p) = ∏ P[Y [k ] = y [k ]], This amount of noise is usually negligible compared to the analog
k=1 (4) noise owing to the semiconductor devices and the EM interference. This
where y [k ] is the kth sample of the digital measurement record. Using way on the one hand it is valid to assume that the effect of jitter can be
(3) and (4) the likelihood function can be written as neglected – it is a component of the additive noise thus it can be in-
tegrated into the additive noise model. On the other hand these ADCs
M
1 ⎡ ⎛ Ty [k] + 1−x [k ] ⎞ ⎛ Ty [k]−x [k ] ⎞ ⎤. shall also measure signals which are faster by several orders of mag-
L(p) = ∏ erf ⎜
2 ⎢ ⎝ σ ( 2) ⎠
⎟−erf ⎜ ⎟⎥ nitude and jitter is not only a random error component but can cause
k=1 ⎣ ⎝ σ ( 2) ⎠ ⎦ (5)
systematic errors (as it is shown in Section 3.1). Therefore it worths to

96
T. Virosztek Measurement 115 (2018) 95–103

Fig. 1. Amplitude-modulated Gaussian white noise.

extend the model developed for the ML estimation of ADC and signal ∞ dx (t ) ∞

parameters to handle the effect of jitter separately.

E{y [k ]} ≃ x (tk,ideal )· ∫−∞ fΔt (Δt )dΔt + dt
· ∫−∞ ΔtfΔt (Δt )dΔt
t = tk ,ideal

1 d2x (t ) ∞
+
2 dt 2
· ∫−∞ (Δt )2fΔt (Δt )dΔt
3.1. Notation and definitions t = tk ,ideal

dx (t ) 1 d2x (t )
= x (tk,ideal ) + ·μt + (μt2 + σt2).
The aperture jitter can be modeled by a linear time-variant system dt t = tk ,ideal 2 dt 2 t = tk ,ideal
which has the impulse response
(16)
h (t ) = δ (t −Δt ), (11) Assuming that there is no systematic error in the sampling (i.e.
μt = 0 )
where Δt is a random variable with expected value μt and standard
deviation σt . Note that the distribution of Δt is not necessarily Gaussian, 1 d2x (t )
E{y [k ]} ≃ x (tk,ideal ) + ·σt2.
nevertheless it is usually assumed [11,12]. Let x (t ) denote the input and 2 dt 2 t = tk,ideal (17)
y [k ] the output of this system. Therefore
In the frequency domain the transfer function of this linear time-
y [k ] = x (tk,ideal−Δtk ). (12) variant system can be written as
E{Y (jω)} 1
In other terms the weighting function of this linear time-variant H (jω) = ≃ 1 + (jω) μt + (jω)2 (μt2 + σt2),
X (jω) 2 (18)
system for the kth sample is
where ω = 2πf . Using the assumption μt = 0 again it leads to
hk (t ) = δ (t −Δtk ). (13)
1
H (jω) ≃ 1− ω2σt2.
2 (19)
Let fΔt (Δt ) denote the probability density function of Δt . This no-
tation is largely compatible with the probabilistic approach of [11]. The This means that the effect of jitter can be modeled by a low-pass
expected value of y [k ] can be expressed as follows: filter which causes systematic error in the amplitude estimation of a
sine wave. Using another heuristic the power of the sine wave (which is
∞
E{y [k ]} = ∫−∞ x (tk,ideal + Δt ) fΔt (Δt )dΔt. (14) originally concentrated to a certain frequency) is spread in the fre-
quency domain owing to the jitter. The power of noise owing to the
Expanding the Taylor series of x (t ) around tk,ideal we achieve jitter – in case of sinusoidal excitation – can be expressed as

Pn,jitter = A2 (1−|H (jω)|2 ) ≃ A2 ω2σt2, (20)

∞ ⎡ dx (t ) 1 d2x (t )
E{y [k ]} = ∫−∞ ⎢x (tk,ideal ) + dt ·Δt +
2 dt 2
·(Δt )2 where A is the amplitude and ω is the angular frequency of the sine
t = tk ,ideal t = tk ,ideal
⎣ wave. Note that the above approximations are valid in those – very
⎤ usual – cases when the sine wave can be approximated by its second
+ …⎥ fΔt (Δt )dΔt . order Taylor series in the [tk,ideal−Δt ,tk,ideal + Δt ] interval (i.e.
⎦ (15) dx (t )
dt
·Δt ≪ 1 or in other terms ω·σt ≪ 1). The low-pass filtering proper-
dx (t ) ties of the aperture jitter are detailed in [12].
Assuming dt ·Δt ≪ 1 – which is a greatly established assumption
according to (10) – the first few elements of the Taylor series approx-
imates function x (t ) satisfactorily. Using the second order Taylor series 3.2. Merging jitter into the model developed for ML estimation
of x (t ) leads to
In that model the sampling is considered to be ideal, however

97
T. Virosztek Measurement 115 (2018) 95–103

sampling nonidealities can be merged into this model. There are at least The slope of the sine wave at the kth sample is
two ways to consider the nonideality of sampling.
dx (t )
= 2πf (B cos(2πftk )−Asin(2πftk )).
dt t = tk (24)
3.2.1. Estimation of the value of each sampling moment
Using this approach the moments of sampling (tk , where k = 1…M ) Since the jitter is usually small compared to the sampling time, for
are parameters to estimate as well. In this case the θ = 2πfTs parameter the calculation of the slope we can use the following approximation:
cannot be used since the tk = k ·Ts assumption is not held. The kth value tk ≃ k ·Ts . This way the slope of the pure sine wave can be expressed in
of the noiseless sine wave is terms of θ = 2πfTs
x [k ] = Acos(2πftk ) + B sin(2πftk ) + C . (21) dx (t ) θ (B cos(kθ)−Asin(kθ))
= .
dt t = k·Ts Ts (25)
This approach leads to multiple consequences.
Note that the nominal sampling time Ts is a well-known global
• Instead of the sampled angular frequency (θ) the frequency (f) of the constant of the measurement. This way using parameters σv and σt the
sine wave shall be used and estimated time-variant standard deviation of noise can be calculated for all the
• The number of the parameters increases largely: the parameter samples.
vector contains the 5 signal parameters, the 2N −1 code transition
2
levels and the M sampling moments. Formally P = 2N + 4 + M . dx (t )
σ [k ] = σv2 + ⎛⎜ ·σt ⎞⎟ .
Since the recommended number of the samples is multiple times
⎝ dt t = k·Ts ⎠ (26)
larger than the number of the code transition levels of the quantizer
under test, this extension increases the length of the parameter Using this extension the likelihood function is
vector very largely. The computational demand of the calculation of M
1 ⎡ ⎛ Ty (k ) + 1−x [k ] ⎞ Ty (k )−x [k ] ⎞ ⎤
the gradient vector and the Hessian matrix becomes such large that L(p) = ∏ erf ⎜ −erf ⎛
⎟ ⎜ ⎟ ,
2⎢⎣ ⎝ 2 σ [k ] ⎠ ⎝ 2 σ [k ] ⎠ ⎥ ⎦ (27)
it cannot be served in PC environment. k=1

• The sensitivity of the cost function is low to a certain sampling where

moment. A sampling moment affects only one element of the cost
function and the sensitivity to tk also depends on the slope of the x [k ] = Acos(kθ) + B sin(kθ) + C (28)
sine wave near tk . This is largely similar to the problem regarding and
the low sensitivity of the ML cost function to the code transition
2
levels. The parameterization of quantizer nonidealities [13] is ne- θ (B cos(kθ)−Asin(kθ)) ⎞
σ [k ] = σv2 + ⎛ ⎜ ·σt . ⎟

cessary because a single code transition level influences the like- ⎝ Ts ⎠ (29)
lihood of the entire measurement only a little bit. This low sensi-
This way the time-variant noise deviation σ [k ] also depends on
tivity leads to uncertain code transition level estimates and
parameters A,B,C ,θ and σv . The sensitivities can be described via the
uncompressed information in a large amount of parameters.
first order partial derivatives of σ [k ], see (30)–(34).
Nevertheless the sensitivity of the cost function to a sampling mo-
ment is usually even lower. The information regarding jitter shall be ∂σ [k ] −θ 2 (B cos(kθ)−Asin(kθ))·sin(kθ)·σt2
= ,
stored in a more compressed way. ∂A σ [k ]·Ts2 (30)

3.2.2. Using an additional parameter to describe the jitter ∂σ [k ] θ 2 (B cos(kθ)−Asin(kθ))·cos(kθ)·σt2

= ,
The aperture jitter is very commonly described by one parameter: ∂B σ [k ]·Ts2 (31)
the standard deviation of the sampling time around the ideal sampling
∂σ [k ]
moment. = 0,
∂C (32)
σt2 = E{(Δtk )2} = E{(tk,real−tk,ideal )2}. (22)
θ (B cos(kθ)−Asin(kθ))·σt2·(B cos(kθ)−Asin(kθ)
The most common one-parameter model for the jitter [11,14] uses ∂σ [k ] + θ (−A·k ·cos(kθ)−B·k ·sin(kθ)))
the following assumptions: = ,
∂θ σ [k ]·Ts2 (33)

• The difference between the real and the ideal sampling moment ∂σ [k ]
=
σv
.
follows Gaussian distribution with zero mean and σt standard de- ∂σv σ [k ] (34)
viation.
• All instances of the sampling error are from the same distribution:
Using this σ [k ] makes the derivatives of the likelihood function
much more complex. In (27) both the numerator and the denominator
Δtk ∼ N (0,σt2),k = 1…M .
• The instances of the sampling error are independent and un-
of the argument of the error function depends on parameters A and B
and θ . This way even the first order derivatives of the likelihood
correlated.
function (or the cost function) become very complex. The complexity
and the computational demand of these shall not be increased if not
These assumptions for the aperture jitter are very common and
necessary. To avoid such challenges in the implementation the fol-
widely used [12,15,16]. This model leads to the amplitude-modulated
lowing approximation can be used.
Gaussian white noise conception of jitter. Let σv denote the standard
deviation of the additional noise on the analog signal which is not re-
lated to jitter (the noise owing to the semiconductor components, etc.). • The initial estimates for the sine wave parameters – achieved via
four parameter least squares fit – are good enough to describe the
Assuming that the noise owing to jitter is independent from the other
slope of the sine wave and thus to parameterize the amplitude-
noise sources we can express the standard deviation of the noise (in-
modulated Gaussian white noise model for the jitter. It is not ne-
cluding its jitter-related component) as a function of time:
cessary to adjust the slope in each iteration by using the actual ML
2 estimates of the signal parameters.
dx (t ) ⎞
σ (t ) = σv2 + ⎛
⎝ dt
·σt .
⎠ (23) • Hence σ [k ] can be written as

98
T. Virosztek Measurement 115 (2018) 95–103

2 that is built using the initially estimated code transition levels. Note
θ ̂ (B
 cos(kθLŜ )−A
LS sin(kθLŜ )) ⎞
σ [k ] = σv2 + ⎛⎜ LS LS ·σt ⎟ , that not all the samples shall be used to calculate this initial esti-
⎝ Ts ⎠ (35) mate: Mσ ⩽ M .
where A LS and θLŜ are the initial least squares (LS) estimates for
LS ,B • Use an initial estimate for the aperture jitter σt . This initial estimate
signal parameters A,B and θ . can be the quantity that appears in the datasheet of the ADC under

• Since A LS and θLŜ are constant in the ML estimation procedure,

LS ,B test or can be a value which is typical for the examined type of
sample-and hold (S/H) circuits.
σ [k ] depends only on parameters σv and σt . The sensitivity of σ [k ] to
σv is described in (34) and the sensitivity of σ [k ] to σt is • Optimize the cost function with respect to the parameters given in
̂
(37). σt,ML is the maximum likelihood estimate of the standard de-
∂σ [k ] (θ ̂ (B
 cos(kθLŜ )−A
LS sin(kθLŜ )))2 ·σt viation of the aperture jitter.
= LS LS .
∂σt σ [k ]·Ts2 (36)
4. Experimental results

4.1. Evaluation of simulated measurements

The parameter vector of the extended model is
To illustrate the process of ML estimation including jitter, two si-
⎡ A ⎤
⎢ B ⎥ mulated measurements have been generated. In the first one the ad-
⎢ C ⎥ ditive noise and the jitter parameters are typical, in the second one the
⎢ θ ⎥
⎢ σv ⎥ jitter has been increased artificially to make the difference more visible.
pext = ⎢ σt ⎥. The common parameters of the generated records are the following:
⎢ ⎥
⎢ T1 ⎥
⎢ T2 ⎥ • number of samples: 200000,
⎢
⎢ ⋮ ⎥
⎥
• sampling frequency: 200 kHz,
T N (37) • signal frequency: 97 Hz,
• number of bits: 8,
⎣ 2 − 1⎦
The sensitivity of the cost function to parameters σv and σt can also
• INL < ±1 LSB,
be described:
• amplitude: 107.53% FSR,
∂CF
=−∑
1
·
M
∂arg(k ) ∂σ [k ]
· ,
• DC: 50% FSR.
∂σv arg(k ) ∂σ [k ] ∂σv (38)
k=1 The parameters of the first simulated measurement record with ty-
where pical jitter are.

• standard deviation of additive noise: σ

2 2
Ty (k ) − x (k ) Ty (k ) + 1− x (k ) ⎞
∂arg(k ) ∂σ [k ] 2 ⎛ −⎛ 2 σ [k] ⎞ Ty (k )−x (k ) −⎛ v = 0.1 LSB,
• aperture jitter: σ = 10 ·T = 5 ps.
⎜ ⎟ ⎜ ⎟

· = ⎠ · −e ⎝ 2 σ [k ] ⎠ ·
⎜e ⎝ 2 t
−6
s
∂σ [k ] ∂σv π⎜ 2 (σ [k ])
⎝
The parameters of the second simulated measurement record with
Ty (k ) + 1−x (k ) ⎞ σv
⎟· large jitter are.
2 (σ [k ])2 ⎟ σ [k ]
⎠ (39)
• standard deviation of additive noise: σ v = 0.1 LSB,
and
M
• aperture jitter: σ = 0.1·T = 500 ns.
t s

∂CF 1 ∂arg(k ) ∂σ [k ]
=−∑ · · , The estimation results appear in Table 1.
∂σt arg(k ) ∂σ [k ] ∂σt (40)
k=1 The 2nd and the 3rd rows contain the real parameters which were
where used to generate the records. The 4th and 5th rows contain the estimates
2 2
of the additive noise and the jitter using the extended ML estimation
Ty (k ) − x (k ) Ty (k ) + 1− x (k ) ⎞
∂arg(k ) ∂σ [k ] 2 ⎛ −⎛ 2 σ [k] ⎞ Ty (k )−x (k ) −⎛
⎜ ⎟ ⎜
2 σ [k ]
⎟ that handles jitter and the 7th row show the estimated standard devia-
· = ⎜e ⎝ ⎠ · −e ⎝ ⎠ ·
∂σ [k ] ∂σt π⎜ 2 (σ [k ])2 tion of the additive noise using the ML estimation without modeling the
⎝ jitter separately. The results lead to two main consequences:
Ty (k ) + 1−x (k ) ⎞ (θLŜ (B
LScos(kθLŜ )−A
LS sin(kθLŜ )))2 ·σt
2 (σ [k ])2 ⎟
⎟·
σ [k ]·Ts2
. • The cost function is larger when the jitter is not estimated sepa-
⎠ rately. This means that separating the jitter provides a more so-
(41) phisticated model of the reality which has larger explanatory power
and leads to better goodness of fit. The difference between the cost
The ML estimation in the extended parameter space can be per-
formed this way:
Table 1
• Perform a 4-parameter least squares fit to achieve initial estimates Evaluation results of simulated measurement records.
for parameters A,B,C and θ .
• Perform a histogram test to get initial estimates for the code tran- Quantity 1st record 2nd record

sition levels. σv 0.1 LSB 0.1 LSB

• Calculate an initial estimate for σ . This can be e.g.
v σt 10−6·Ts = 5 ps 0.1·Ts = 500 ns
̂ (with modeling jitter)
σv,1 0.099876 LSB 0.099920 LSB
1 M
σv̂ ,initial =
Mσ −1
∑k =σ1 (y [k ]−q (xLS [k ]))2 , σt ̂ (with modeling jitter) 9.979043·10−7·Ts 0.099777·Ts
(42) Cost Function 27974.8 29268.6
̂ (without modeling jitter)
σv,2 0.100067 LSB 0.100262 LSB
where q (xLS [k ]) is the kth sample of the pure sine wave assembled
Cost Function 27975.4 29328.0
using the initial LS estimates and then quantized by the quantizer

99
T. Virosztek
function values is larger when the amount of jitter is larger: in case
of 5 ps jitter the difference is only 0.6 and in case of 500 ns jitter the
difference is 59.4. Not surprisingly the larger explanatory power of
the more detailed model is more visible if the effect of jitter is larger.
The time variant noise owing to the aperture jitter – if not modeled
separately – increases the variance of the other parameter estimates
and even can cause systematic error in the estimation of them (see
(18) and (19)). Furthermore in special cases the non-modeled jitter
can lead the ML optimization into local extrema that mislead the
entire parameter estimation. Note that if the effect of jitter is neg-
ligible, this extension only causes a minor computational overhead
and leads to the result that can be achieved without modeling jitter.
• The estimated deviation of noise is σv̂ is larger in the case when the
jitter is not separated. When jitter is not treated separately, the es-
timated σv,2̂ contains the noise owing to the real additive noise and
owing to the jitter:
2
dx (t )
σv,2 = σv2,1 + ⎜⎛
⎝ dt
·σt ⎟⎞ .
rms ⎠ (43)
̂ must be larger that the value of σv,1
This way the value of σv,2 ̂ . These
experimental results are in accordance with (43). Note that if an
experimental result shows σv,2̂ < σv,1̂ , that can be caused by a rare
coincidence when the realization of the noise owing to jitter and the
realization of the noise owing to other sources more or less cancel
each other. In this case the experiment shall be repeated a few times
to avoid coincidence. Furthermore in case of long measurement
records (containing a few hundred thousand samples) this coin-
cidence is largely unlikely.
The results of the evaluation of these measurement records are
shown in Figs. 2 and 3. These figures show the measurement record and
the fitted sine wave (1st subplot), the time-varying noise parameter σ [k ]
(2nd subplot) and the components of the likelihood function – the
probabilities of the samples assuming the estimated parameters (3rd
subplot). To provide readable figures, these quantities are displayed for
the first 2500 samples.
Note that σ [k ] is approximately a sinusoidal with DC offset; the
100
Measurement 115 (2018) 95–103
amplitude of this sine wave is larger if the amount of jitter is larger. The
probability of a recorded sample – assuming the parameters stored in
parameter vector pext , see (37) – is reaching 1 if the stimulus overdrives
the quantizer and remains between 0 and 1 otherwise.
4.2. Evaluation results of real measurements
A measurement has been performed with similar parameters using a
Bruel & Kjaer 1051 sine wave generator, a Bruel & Kjaer 2638 wideband
signal conditioning amplifier and an NI-9201 data acquisition board.
This DAQ board utilizes an analog multiplexer and a 12-bit SAR ADC
manufactured by Texas Instruments. The measurement has been per-
formed in standard laboratory environment without special EMC pre-
cautionary measures. The cables were standard RG-58 coaxial cables
with 50 Ω characteristic impedance. The nominal parameters of the
measurement are.
• number of samples: 100000,
• sampling frequency: 100 kHz,
• signal frequency: 97 Hz,
• amplitude: 120.4% FSR,
• Number of bits: 12,
• INL < ±1.5 LSB according to [17].
The results of the evaluation are collected in Table 2.
It is important to remark that σv̂ is lower when jitter is modeled
̂ = 1.199825 LSB and σv,2
separately (σv,1 ̂ = 1.200036 LSB). Quantity σv,2 ̂
estimates the sum of the noise owing to the aperture jitter and other
noise components while σv,1 ̂ estimates the additive noise components
separated from the amplitude-modulated noise owing to the jitter. The
cost function is lower in the first case: it indicates better goodness of fit
when jitter is separated.
Other results of the evaluation are shown in Fig. 4. This figure
displays the measured and the fitted sine wave (1st subplot), the σ [k ]
values which show the time-varying property of the noise parameter
(2nd subplot) and the probabilities of the recorded samples which de-
termine the likelihood of the measurement (3rd subplot). To be readable,
Fig. 2. Evaluation results of the simulated
measurement record with typical jitter.
T. Virosztek Measurement 115 (2018) 95–103

Fig. 3. Evaluation results of the simulated

measurement record with large jitter.

Table 2 Fig. 4 shows the first 2500 elements of these vectors.

Evaluation results of a real measurement record. Subplot 2 shows the time-variant noise estimate:

Quantity With modeling jitter Without modeling jitter 2

dx (t ) ⎞
σ (̂ t ) = (σv̂ )2 + ⎛ ·σt ̂ .
σv̂ 1.199825 LSB 1.200036 LSB ⎝ dt ⎠ (44)
σt ̂ 0.6083·10−6·Ts = 6.0831 ps N/A
dx (t )
Cost Function 124223.5 135538.3 Since σv̂ is a large constant compared to dt
·σt ̂
the result is a si-
nusoidal with DC offset. The probabilities of the samples in the record –
assuming the parameters in parameter vector pext , defined by (37) –

Fig. 4. Evaluation results of a real mea-

surement record.

101
T. Virosztek
Table 3
Evaluation results of a population containing N = 20 real measurement records.
Quantity With modeling jitter Without modeling jitter
σv̂ 1.2074 LSB (3.8291·10−5 LSB) 1.2078 LSB (7.1770·10−5 LSB)
σt ̂ 5.9162 ps (1.8438 fs) N/A
Cost Function 128374.9 (1.2214) 136849.7 (4.0835)
appear in subplot 3. The probability reaches 1 when the stimulus is
overdrives the DUT and stays below 0.4 when the excitation signal is
within the full scale range of the ADC. These values are likely con-
sidering that the standard deviation of the noise is nearly 1.2 LSB.
To gain an estimate population the measurement has been repeated
with the same nominal parameters N = 20 times. The 20 measurements
were not consecutive but they were close enough to each other in time
to assume that the parameters of the excitation signal and the en-
vironment noise did not change significantly. The different measure-
ments provided different noise realizations but the amount of noise was
the same according to our assumption. This way Table 2 has been re-
produced using populations. In each cell of Table 3 the first value is the
expected value and the second is the empirical standard deviation of the
corresponding quantity (in parentheses).
This 20-element population draws a figure similar to the previous
measurement results. The estimated deviation of noise is larger if jitter
is not treated separately (in this case σv̂ also contains the effect of jitter)
and the goodness of fit is better (the cost function is lower) using the
extended model. The empirical standard deviation of the jitter estimate
is relatively large (312 ppm of the expected value). It is important to
remark that σt ̂ estimates a parameter of a small component in a rela-
tively large noise.
The scripts generating the simulated measurements, the real mea-
surement records and the evaluation routines are available at [18].
4.3. Comparison with other jitter estimation methods
Since the proposed method uses a sinusoidal measurement record it
can only be compared to techniques that also use this kind of excitation
and measurement setup. There are several techniques that use other
measurement setups e. g. the one proposed in [19] or the double beat
frequency technique (see [20]). The locked histogram test [20] is one of
the most feasible ones: it requires a sinusoidal stimulus and uses the
samples near the positive zero-crossing of the signal. It assumes that the
slope of the sine wave near the zero-crossing is
dx (t )
= 2πfx A,
dt t = k·Tx
where Tx is the period of the stimulus, fx = 1 and A is the amplitude of
Tx
the excitation signal. The histogram of the samples closest to the po-
sitive zero crossing is recorded in a histogram, the standard deviation of
them (σhist ) can be calculated. Therefore – assuming that the stimulus
signal drives the full scale range of the ADC under test – the aperture
jitter can be calculated:
σhist
̂
σt ,locked hist = , [2] IEEE Waveform Generation, Measurements and Analysis Committee (TC-10), IEEE
πfx 2N
̂ is expressed in LSB and N is the number of bits. The 20-
where σhist
element measurement population described in Section 4.2 has been re-
evaluated: the aperture jitter has been estimated using the locked his-
togram method (the zero-crossing has been detected by frequency es-
timation instead of hardware lock.) The estimation results of the ex-
isting and the proposed method have been compared in Table 4.
The notation is similar to that of Table 3: the first value is the ex-
pected value of the estimate over the entire population and the second
value (in parentheses) is the empirical standard deviation of the esti-
mate. There is a significant difference between the expected value of
Measurement 115 (2018) 95–103
Table 4
Comparison of estimation results of the ML and the locked histogram method.
Quantity ML method Locked histogram technique
σt ̂ 5.9162 ps (1.8438 fs) 44.1394 ps (104.37 fs)
estimates achieved via the ML method and the locked histogram. One
source of the difference is the uncertainty of estimating the samples
nearest to the positive zero-crossings of the stimulus. The other source
of it is more straightforward: since (46) considers all the noise on the
samples nearest to the positive zero-crossings as result of jitter, this
estimate contains the effect of other noise sources as well.
5. Conclusions
The model developed for the maximum likelihood estimation of
ADC and excitation signal parameters can be extended to handle the
effect of aperture jitter. This extension makes the standard deviation of
the noise time-variant owing to the varying slope of the sine wave. The
parameter space increases only by one additional parameter (the
standard deviation of the aperture jitter: σt ). Furthermore the extremum
of the extended likelihood function gives a maximum likelihood esti-
̂ ). The introduction of the additional parameter
mate for σt as well (σt,ML
increases the complexity and the computational demand of the opti-
mization, but this problem can be solved by a simple and slight ap-
proximation: the slope of the sine wave can be fixed during the opti-
mization using the initial LS estimates of the sine wave parameters. This
way the maximum likelihood estimation of aperture jitter is feasible
and the required efforts do not increase significantly compared to the
ML parameter estimation without modeling jitter. The experimental
results show that extending the model to handle jitter enhances its
explanatory power, thus increases the quality of the estimation (e.g. in
terms of cost function) in those cases where the effect of aperture jitter
is relevant.
The main advantages of the proposed method are:
• it provides consistent and effective ML estimates for aperture jitter,
• it can be evaluated using a sinusoidal measurement record (while
estimating other parameters in ML sense as well),
• it does not require special measurement setup and
• it enhances the performance of the sine-wave based ML estimation
and helps to avoid systematic errors if the amount of jitter is sig-
nificant.
Acknowledgement
(45)
This work was supported by the National Research, Development
and Innovation Office (NKFIH) under Grant No. K-115820.
References
[1] IEEE Waveform Generation, Measurements and Analysis Committee (TC-10), IEEE
standard for terminology and test methods for analog-to-digital converters, IEEE-
1241-2010, 2011, pp. 1–139, http://dx.doi.org/10.1109/IEEESTD.2011.5692956.
(46)
standard for digitizing waveform recorders, IEEE-1057-2007, 2008, pp. 1–210,
http://dx.doi.org/10.1109/IEEESTD.2008.4494996.
[3] S.S. Wilks, Mathematical Statistics, John Wiley and Sons, 1962.
[4] D. Wackerly, W. Mendenhall, R.L. Scheaffer, Mathematical Statistics with
Applications, seventh ed., Duxbury Press, 2008.
[5] J. Saliga, L. Michaeli, J. Busa, J. Liptak, I. Kollár, T. Virosztek, A comparison of least
squares and maximum likelihood based sine fittings in ADC testing, Measurement
46 (10) (2013) 4362–4368, http://dx.doi.org/10.1016/j.measurement.2013.05.
004.
[6] L. Balogh, I. Kollár, A. Sárhegyi, Maximum likelihood estimation of ADC para-
meters, in: IEEE Instrumentation and Measurement Technology Conference
Proceedings, 2010, pp. 24–29, http://dx.doi.org/10.1109/IMTC.2010.5488286.
[7] V. Srinivasan, S.K. Islam, G.T. Hendrickson, A method for the estimation of aperture
uncertainty in A/D converters, in: Proceedings of the IEEE International Symposium
102
T. Virosztek
on Circuits and Systems, ISCAS 2002, 2002, http://dx.doi.org/10.1109/ISCAS.
2002.1010176.
[8] G.K.D. Teyou, H. Petit, P. Loumeau, H. Fakhoury, Y.L. Guillou, S. Paquelet,
Statistical analysis of noise in broadband and high resolution ADCs, in: Proceedings
of the 21st IEEE International Conference on Electronics, Circuits and Systems
(ICECS), 2014, http://dx.doi.org/10.1109/ICECS.2014.7050029.
[9] S. Chaganti, L. Xu, D. Chen, A low-cost method for separation and accurate esti-
mation of ADC noise, aperture jitter, and clock jitter, in: Proceedings of the IEEE
35th VLSI Test Symposium (VTS), 2017, http://dx.doi.org/10.1109/VTS.2017.
7928950.
[10] B. Brannon, A. Barlow, Aperture uncertainty and ADC system performance, Analog
Dev. - Appl. Notes AN-501 (2006).
[11] T.M. Souders, D.R. Flach, C. Hagwood, G.L. Yang, The effects of timing jitter in
sampling systems, IEEE Trans. Instrument. Meas. 39 (1) (1990) 80–85, http://dx.
doi.org/10.1109/19.50421.
[12] H. Kobayashi, M. Morimura, K. Kobayashi, Y. Onaya, Aperture jitter effects in
wideband sampling systems, in: Proceedings of the 16th IEEE Instrumentation and
Measurement Technology Conference, 1999. IMTC/99, IEEE Instrumentation and
Measurement Society, 1999, http://dx.doi.org/10.1109/IMTC.1999.776991.
[13] T. Virosztek, I. Kollár, Parameterization of nonideal quantizers for simultaneous
103
Measurement 115 (2018) 95–103
estimation of quantizer and excitation signal parameters, Measurement 50 (2017)
27, http://dx.doi.org/10.1016/j.measurement.2017.04.020.
[14] M. Shinagawa, Y. Akazawa, T. Wakimoto, Jitter analysis of high-speed sampling
systems, IEEE J. Solid-State Circ. 25 (1) (1990) 220–224.
[15] R. Tangelder, H. de Vries, R. Rosing, H. Kerkhoff, M. Sachdev, Jitter and decision-
level noise separation in a/d converters, in: Proceedings of the 16th IEEE
Instrumentation and Measurement Technology Conference, 1999. IMTC/99, 1999,
http://dx.doi.org/10.1109/IMTC.1999.776087.
[16] C.-L. Chang, P.S. Huang, T.-M. Tu, Aperture jitter of sampling system in AWGN and
fading channels, IEEE Trans. Instrument. Meas. 56 (3) (2007) 831–839.
[17] Datasheet – NI9201 8AI, 10V, 12Bit, 500 kS/s, <http://www.ni.com/pdf/
manuals/373783a_02.pdf> (accessed: 2017-01-28).
[18] T. Virosztek, Raw data regarding ML estimation of jitter, <http://home.mit.bme.
hu/%7Evirosztek/publicationdata/jitter>.
[19] G. Chiorboli, M. Fontanili, C. Morandi, A new method for estimating the aperture
uncertainty of A/D converters, IEEE Trans. Instrument. Meas. 47 (1) (1998), http://
dx.doi.org/10.1109/19.728790.
[20] Y. Langard, J.L. Balat, J. Durand, An improved method of adc jitter measurement,
in: Proceedings of the International Test Conference 1994, 1994, pp. 763–770,
http://dx.doi.org/10.1109/TEST.1994.528023.