648 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—II: EXPRESS BRIEFS, VOL. 55, NO.

7, JULY 2008

Statistical Spectra and Distortion Analysis of

Time-Interleaved Sampling Bandwidth Mismatch
Sai-Weng Sin, Member, IEEE, U.-Fat Chio, Student Member, IEEE, Seng-Pan U, Senior Member, IEEE, and
R. P. Martins, Fellow, IEEE

Abstract—Time interleaving is one of the most efficient tech-

niques employed in the design of high-speed sampled-data systems.
However, the mismatches appearing among different channels will
create distortion tones that will degrade the system performance.
This paper presents a detailed analysis of the effect of sampling
bandwidth mismatches in the statistical behavior of time-inter-
leaved sampling systems. Closed-form and handy signal-to-noise
distortion ratio formulas will also be derived which allow quick
evaluation of the system performance, and the MATLAB simu-
lations are provided in order to verify the effectiveness of those
formulas. Finally, the corresponding design procedure extracted
from the formulas will be further addressed through a design
example.
Index Terms—Bandwidth mismatches, sampled-data systems,
Fig. 1. Time-interleaved ADC.
time-interleaved (TI) analog-to-digital converter (ADC).

and characterized in the very early design phase, allowing

I. INTRODUCTION an accurate estimation of sampling circuit components, oth-
erwise post-calibration (either analog or digital) [6]–[8] is
unavoidable, which may lead to additional power and silicon
H IGH-SPEED sampled-data systems, e.g., analog-to-dig-
ital converters (ADCs), have found increasing impor-
tance in various electronics circuits and systems, such as
consumption with complex hardware and software (algorithms)
for measurement and calibration.
in wireless communications, flat-panel displays, video and In this paper, a thorough spectra-domain analysis of the band-
imaging signal processing applications. Time-interleaving (TI) width mismatch effect will be presented and based on this a
in ADCs [1], as shown in Fig. 1, is one of the effective ways to closed-form SNDR expression will be derived. Also, it will be
boost the speed of sampled-data systems beyond the limitation proven that the most important contribution for the bandwidth
imposed by the technology. However, special attention must be mismatch is the phase mismatch originated from the phase-shift
paid to avoid various types of mismatches arising in parallel of the RC-filtering by the sampling branch, which causes per-
identical channels, e.g., offset, [2], gain, [2], sampling-time formance degradation even for signal frequencies smaller than
[2]–[5], as well as bandwidth mismatches (from the sampling the bandwidth of such sampling branch. Although the analysis
RC time constant) [2], [6]–[8]. These mismatches create dis- of deterministic mismatches had already been addressed in [2],
tortion tones at frequency locations which are multiples of [6] and [7], this paper will present a comprehensive and thor-
, where is the overall sampling frequency of ough study with a new statistical approach that comprises the
the system and the number of TI channels. following improved features over previous works.
To characterize conveniently the performance degradation a) A handy closed-form SNDR expression is derived.
due to various types of mismatches, handy and closed-form b) The formula can be generalized for any number of
expressions for the signal-to-noise-and-distortion ratio (SNDR) TI channels.
are highly desirable. Thus, the mismatches can be predicted c) Since intrinsic circuit imbalance (being highly circuit and
architecture dependent) or inadequate layout design will
introduce systematic mismatches, and considering that
Manuscript received February 6, 2007; revised December 22, 2007. First pub-
lished May 7, 2008; last published July 16, 2008 (projected). This work was sup-
only standard deviations of various mismatch parameters
ported by the University of Macau under Research Grant RG058/06-07S/MR/ are given by the foundries process datasheets, a statistical
FST. This paper was recommended by Associate Editor G. Manganaro. method dealing with random mismatches is derived.
S.-W. Sin and U-F. Chio are with the Analog and Mixed-Signal VLSI Labo- The formula derived here will simplify the performance eval-
ratory, Faculty of Science and Technology, University of Macau, Macao, China
(e-mail: [email protected]).
uation of TI sampling systems under bandwidth mismatches,
S.-Pan U is with the Chipidea Microelectronics (Macao) Limited, Macao, and will also allow the prior determination of the specifications
China (e-mail: [email protected]). of sampling switches and capacitors in the design phase for a
R. P. Martins is with the Analog and Mixed-Signal VLSI Laboratory, Faculty specifically targeted value of SNDR. The analysis in [8] focuses
of Science and Technology, University of Macau, Macao, China, and also with
the Instituto Superior Técnico/TUL, Lisbon, Portugal (e-mail: rmartins@umac. on the design methodology for digital post calibration assuming
mo). small input signal bandwidth, while this paper is targeting a
Digital Object Identifier 10.1109/TCSII.2008.921600 prior analysis to suppress the bandwidth mismatch within a full
1549-7747/$25.00 © 2008 IEEE
SIN et al.: STATISTICAL SPECTRA AND DISTORTION ANALYSIS OF TIME-INTERLEAVED SAMPLING BANDWIDTH MISMATCH 649

This leads to an output spectrum that can be expressed by [7]

(5)

where

Fig. 2. Typical front-end S/H circuit. (5a)

input signal bandwidth at a much earlier design phase that is

Equations (5) and (5a) completely describe the spectrum of a
well suited for usual top-down analog design flows.
TI sampling system with bandwidth mismatches. Furthermore,
they reveal that the characteristic of the bandwidth mismatch
II. SPECTRA ANALYSIS OF BANDWIDTH MISMATCHES is similar to a pure gain - and timing mismatch, also creating
A typical sample-and-hold (S/H) circuit widely used in the weighted modulation sidebands at multiples of
front-end of sampled-data systems is shown in Fig. 2, which [with weights specified by in (5a)].
must also be part of each channel of the ADC front-end of Fig. 1.
In phase 1 the continuous-time input signal is passively sampled III. CLOSED-FORM FORMULA FOR SNDR
into the capacitor C and also RC-filtered by the finite bandwidth In this section, a closed-form formula for SNDR will be
imposed by the sampling switch’s on-resistance R and the sam- derived for TI systems with bandwidth mismatches. Recalling
pling capacitor C. The mismatches occurring in the sampling that (5a) represents the weights of the various modulation
switches and the capacitors among various channels will lead sidebands which allows SNDR’s computation. For an input
to the overall bandwidth mismatch. This will only modulate the sinusoidal signal with frequency , the sidebands
continuous-time input signal, and after the signal being sampled contain tones located at . Due to the
into C the operation of the remaining part of the S/H is in dis- symmetry property of the Fourier Transform of real signals,
crete-time and all the mismatches will appear as offset and pure the calculation can be simplified by evaluating the sideband
gain-mismatches that were already fully characterized before, components at only over the range
[2]. Assuming that and are the switch’s on-resistance (or equivalently to find the SNDR (with
and the sampling capacitance of the th channel, respectively, respect to the signal only centered at ) over the range of
then, the sampling time constant would be .
where corresponds to its mean/nominal value and Simplifying (5a) using and (1) it
is the percentage of the time-constant mismatch. Based on yields
this, the frequency response in the th channel, from the input
to the sampled signal , can be defined as (6)

(1) where for corresponds to the signal component,

and for to correspond to different
sideband components. The SNDR, over the range of or
containing both magnitude and phase contributions. On the equivalently for to , can be expressed as [9]
other hand, the sampled time-domain signal in capacitor C can
be represented as
dB (7)

(2) In order to evaluate the SNDR by statistical means it would be

necessary to assume now as
independent identically distributed (i.i.d) random variables with
where is the impulse response of the RC-filter in the
Gaussian distribution of zero mean and standard deviation .
th channel and denotes the convolution operation. For a TI
Then, the expected values of the signal component
system with channels the following will hold:
can be evaluated by substituting into (6) and multiplying
it by its complex conjugate as follows:
for integer (3)

and (2) becomes

(8)

which shows that the signal component is approximately scaled

by the RC-filter for small values of . The expected
(4)
value of the sideband components can also be calculated from
650 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—II: EXPRESS BRIEFS, VOL. 55, NO. 7, JULY 2008

(6) leading to (9), shown at the bottom of the page, where, in since and are random variables with zero means. Further-
this case, the second-order term in (9) cannot be neglected more, for the second-order term
since in the following derivations it will be required to use the
statistical values of and and, besides
this, it will constitute the most significant contribution to the
sideband components. However, in (8) such approximation is
valid because the amplitude of the signal component is, in prin-
ciple, much larger than that of the sideband. The evaluation of (14)
the expected value of the sideband components in (9) is quite
complex and only key steps will be addressed here. The double
summation in (9) can be analyzed in two parts for (con-
taining M terms) and (contain terms) in the
(15)
following way:
Since and have zero mean and are independent (uncorre-
lated) (15) can be evaluated considering
and which will lead to

(10) (16)

Then, substituting (12) and (14) into (10) it yields

(17)

(11) and, similarly, by substituting (13) and (16) into (11)

where

(18)

should be small for leading to a valid Taylor Series where

expansion in (10) and (11). The expansion can also be evalu-
ated up to the second order because and contain at least
a third-order term like which is negligible compared to the
main contribution from . The expected values can be calcu-
lated as

(19)
(12)
leading to the following simplification of part 2:

(13)
(20)

(9)
SIN et al.: STATISTICAL SPECTRA AND DISTORTION ANALYSIS OF TIME-INTERLEAVED SAMPLING BANDWIDTH MISMATCH 651

M versus SNDR [simulated and calcu-

!  = 0:8  = 1%.
Fig. 3. Plot of number of TI channels
lated from (22)] with and !
M =4  =1
Fig. 4. Plot of normalized frequency versus SNDR [simulated and calcu-
lated from (22) and (24)] with and %.
Finally, the sideband components can be calculated by the ad-
dition of both parts 1 and 2 Then, following the same procedure it can be shown that :

(21)

From (21) it can be concluded that the expected values of the

various sidebands are independent of the frequency index , i.e., dB (24)
all the sidebands have the same expected distortion power. As a
result, the total distortion power should be equal to (21) multi- which differs from (22) only in the last term. Fig. 4 presents
plied by , and the final SNDR can be calculated by sub- a plot of (22) and (24) as a function of with and
stituting (8) and (21) into (7) %. It clearly shows that the effect of the phase mismatch
is dominant especially at low frequency, which will lead to (e.g.,
25 dB) overestimation of SNDR for the TI systems if only gain
mismatch is considered as a dominating factor around the corner
dB (22) frequency.

IV. SIMULATION RESULTS AND DESIGN FLOW

Although the previous analysis can be considered relatively
complex, the final SNDR formula is simple enough for quick To verify the accuracy of the formula, MATLAB FFT model
evaluation of the performance of TI sampled-data systems under simulations will be presented. Fig. 5(a) shows plots of the simu-
the influence of bandwidth mismatches. Important information lated SNDR versus normalized signal frequency and by
can be extracted directly from (22), namely that SNDR is in- 10000 times Monte Carlo simulations under and
versely proportional to the signal frequency, as expected. The , with the absolute error (in decibels) between the
1st term in (22) is reflecting the RC filtering effect by the finite simulated and calculated SNDR presented in Fig. 5(b). From
bandwidth, since this will degrade the signal amplitude and thus this figure the absolute error between the simulated and calcu-
the SNDR. Also, the formula can be generalized for any number lated SNDR is well below 0.5 dB for the normal range of
of TI-channels, and actually the SNDR only degrades by and . The error starts to increase with higher values of the
an amount of 3 dB from to , as shown in Fig. 3, a fact mismatch and high input frequency, since an approximation
that was not explained before when the phenomenon of band- of is considered in the previous formula derivations,
width mismatch was previously analyzed [2], [6], [7] (actually but such a large value of mismatches (e.g., %) are not typ-
a similar conclusion can be found in the gain- and timing-mis- ical in CMOS which will lead to an SNDR of only dB.
match analysis, e.g., from, [3], and [5]). Fig. 6(a) shows another 3-D plots of the simulated SNDR versus
The first-order filtering effect caused by the sampling branch the number of TI channels and by 10000 times Monte
will usually have more influence when the signal frequency is Carlo simulations under and ,
near or greater than its bandwidth, since before the 3-dB fre- with the absolute error (in decibels) between the simulated and
quency the filter provides only negligible attenuation. However, calculated SNDR presented in Fig. 6(b). It is also demonstrated
the bandwidth mismatches among TI channels can have a large that the formula can accurately predict systems’ performance
impact on the performance even when the signal frequency is under any number of TI channels .
well below the corner frequency (quite unexpectedly), since the On the other hand, to demonstrate the corresponding design
phase shift starts to increase at of the corner-frequency. To flow using (22), a 10-bit 6-channel 360 MHz pipelined ADC
demonstrate this effect it should be considered only the magni- was implemented in 0.18- m CMOS [10], where the SNDR
tude response of an idealized RC-filter specification for distortions imposed by signal-dependent mod-
ulation sidebands was designed to be 60 dB (including all error
(23) sources), and the SNDR contribution from pure bandwidth mis-
matches should be below 65 dB. With the signal frequency of
652 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—II: EXPRESS BRIEFS, VOL. 55, NO. 7, JULY 2008

Fig. 5. (a) Simulated SNDR. (b) Absolute error between the simulated and calculated SNDR of TI systems under bandwidth mismatches versus normalized
frequency !  and mismatch standard derivation  by 10000 times Monte Carlo Simulations (M = 4; ! T = 2 0:2495). 2

Fig. 6. (a) Simulated SNDR. (b) Absolute error between the simulated and calculated SNDR of TI systems under bandwidth mismatches versus no. of TI channels
M and mismatch standard derivation  by 10000 times Monte Carlo Simulations (!  = 0:8; ! T = 2 0:2495). 2
25.2 MHz and sampling bandwidth of 100 MHz, the mismatch formula), in the dimensioning of the sampling network compo-
in the sampling bandwidth should be smaller than % nents, thus leading to the suppression of such mismatches.
from (22), which is achievable in current standard CMOS tech-
nology. According to this specification and the statistical param- REFERENCES
eters from the foundry’s datasheet (with design margins), 1-pF [1] L. Sumanen et al., “A 10-bit 200 MS/s CMOS parallel pipeline A/D
converter,” in Proc. ESSCIRC, Sep. 2000, pp. 440–443.
sampling capacitor and 1.6-k resistance were chosen which [2] N. Kurosawa, H. Kobayashi, K. Maruyama, H. Sugawara, and K.
will imply ns. This chip was successfully implemented Kobayashi, “Explicit analysis of channel mismatch effects in time-in-
with good silicon performance, and without any type of band- terleaved ADC systems,” IEEE Trans. Circuits Syst. I, Fundam. Theory
width mismatch calibration, showing the importance of mini- Appl., vol. 48, no. 3, pp. 261–271, Mar. 2001.
[3] Y. C. Jenq, “Digital spectra of nonuniformly sampled signals: Fun-
mizing the bandwidth mismatches according to the formula and damentals and high-speed waveform digitizers,” IEEE Trans. Instrum.
following an analog design flow leading to precise values of the Meas., vol. 37, no. 2, pp. 245–251, Jun. 1988.
sampling circuit components. [4] Y. C. Jenq, “Digital-to-analog (D/A) converters with nonuniformly
sampled signals,” IEEE Trans. Instrum. Meas., vol. 45, no. 1, pp.
56–59, Feb. 1996.
[5] S.-P. U, S.-W. Sin, and R. P. Martins, “Exact spectra analysis of sam-
V. CONCLUSION pled signal with jitter-induced nonuniformly holding effects,” IEEE
Trans. Instrum. Meas., vol. 53, no. 4, pp. 1279–1288, Aug. 2004.
[6] T.-H. Tsai, P. J. Hurst, and S. H. Lewis, “Bandwidth mismatch and
This paper presents a comprehensive spectrum analysis of its correction in time-interleaved analog-to-digital converters,” IEEE
the bandwidth mismatch in time-interleaved sampling systems. Trans. Circuits Syst. II, Exp. Briefs, vol. 53, no. 10, pp. 1133–1137,
Due to the statistical nature of mismatches a statistical anal- Oct. 2006.
[7] C. Vogel, D. Draxelmayr, and F. Kuttner, “Compensation of timing
ysis is presented together with the derivation of a handy and mismatches in time-interleaved analog-to-digital converters through
closed-form SNDR formula which can be used to quickly eval- transfer characteristics tuning,” in Proc. of 2004 Midwest Symp.
uate the system performance. The analysis was further general- Circuits Syst., Jul. 2004, vol. 1, pp. I-341–I-344.
ized to any number of TI channels with an elegant and simple [8] Z. Liu, M. Furuta, and S. Kawahito, “Simultaneous compensation of
RC mismatch and clock skew in time-interleaved S/H circuits,” IEICE
expression. To verify the accuracy of the formula MATLAB Trans. Electron., vol. E89-C, no. 6, pp. 710–716, Jun. 2006.
simulations were presented showing that the absolute error in [9] Y. C. Jenq, “Direct digital synthesizer with jittered clock,” IEEE Trans.
performance prediction was as low as 0.5 dB only for a rea- Instrum. Meas., vol. 46, no. 3, pp. 653–655, Jun. 1997.
[10] S.-W. Sin, S.-P. U, and R. P. Martins, “A 1.2-V 10-bit 60–360 MS/s
sonable range of bandwidth mismatches and signal frequencies. time-interleaved pipelined ADC in 0.18-m CMOS with low-voltage
An IC design example is also briefly referred to show the cor- demultiplexing and gain-and-offset compensation,” IEEE J. Solid-State
rectness of following an analog design procedure (based on the Circuits, to be published.