The Astrophysical Journal, 716:746–757, 2010 June 10 doi:10.

1088/0004-637X/716/1/746

C 2010. The American Astronomical Society. All rights reserved. Printed in the U.S.A.

AMiBA WIDEBAND ANALOG CORRELATOR

Chao-Te Li1 , Derek Y. Kubo1 , Warwick Wilson2 , Kai-Yang Lin1,3 , Ming-Tang Chen1 , P. T. P. Ho1,4 ,
Chung-Cheng Chen1 , Chih-Chiang Han1 , Peter Oshiro1 , Pierre Martin-Cocher1 , Chia-Hao Chang1 , Shu-Hao Chang1 ,
Pablo Altamirano1 , Homin Jiang1 , Tzi-Dar Chiueh3 , Chun-Hsien Lien3 , Huei Wang3 , Ray-Ming Wei3 ,
Chia-Hsiang Yang3 , Jeffrey B. Peterson5 , Su-Wei Chang1 , Yau-De Huang1 , Yuh-Jing Hwang1 , Michael Kesteven2 ,
Patrick Koch1 , Guo-Chin Liu1,6 , Hiroaki Nishioka1 , Keiichi Umetsu1 , Tashun Wei1 , and Jiun-Huei Proty Wu3
1 Institute of Astronomy and Astrophysics, Academia Sinica, P.O. Box 23-141, Taipei 106, Taiwan
2 Australia Telescope National Facility, Epping, NSW 1710, Australia
3 National Taiwan University, Taipei 106, Taiwan
4 Harvard-Smithsonian Center for Astrophysics, 60 Garden Street, Cambridge, MA 02138, USA
5 Carnegie-Mellon University, Pittsburgh, PA 15213, USA
6 Tamkang University, 251-37 Tamsui, Taipei County, Taiwan; [email protected]
Received 2009 August 18; accepted 2010 January 6; published 2010 May 21

ABSTRACT
A wideband analog correlator has been constructed for the Yuan-Tseh Lee Array for Microwave Background
Anisotropy. Lag correlators using analog multipliers provide large bandwidth and moderate frequency resolution.
Broadband intermediate frequency distribution, back-end signal processing, and control are described. Operating
conditions for optimum sensitivity and linearity are discussed. From observations, a large effective bandwidth
of around 10 GHz has been shown to provide sufficient sensitivity for detecting cosmic microwave background
variations.
Key words: cosmic background radiation – instrumentation: interferometers – telescopes
Online-only material: color figures

1. INTRODUCTION ate frequency (IF) is from 2 to 18 GHz and is matched with a

wideband analog correlator.
Interferometric observations have gained much popularity The strong interest in CMB observations has motivated the
in the study of the cosmic microwave background (CMB) development of very broadband correlators with a limited spec-
anisotropy (White et al. 1999; Carlstrom et al. 2002; Padin et al. tral resolution. Utilizing a filter bank and complex correlators,
2002; Leitch et al. 2002; Zwart et al. 2008), due to their advan- Padin et al. (2002) built an analog correlator with 10 GHz band-
tage in stability and phase preserving characteristic via the het- width. Harris & Zmuidzinas (2001) worked on a different ap-
erodyne scheme for correlation and polarization observations. proach toward broadband operations by adopting an analog lag
The cross-correlations used in an interferometric array can ef- correlation scheme to build an auto correlator with a 4 GHz
fectively suppress many systematics. To achieve high brightness bandwidth. The AMiBA correlator is also based on the con-
sensitivity, broadband low noise receivers and matching corre- cept of the lag correlator and is designed to correlate the entire
lators are the two most important components for a continuum 16 GHz bandwidth. Recently, Holler et al. (2007) also reported
interferometer. The advance in millimeter and microwave de- a lag correlator for the Arcminute Microkelvin Imager with a
tector technologies in recent years has produced very broadband 6 GHz bandwidth.
components. Low-noise amplifiers (LNAs) with bandwidths of In principle, analog correlators can achieve better sensitivity
10 GHz or more are easily accessible with noise performance over their digital counterparts due to the absence of the digiti-
comparable to bolometric direct detectors, e.g., ACBAR (Run- zation process. Analog multipliers can easily achieve high sen-
yan et al. 2003) with bandwidths of 30–40 GHz and an equiv- sitivity over multi-octave frequency ranges. In comparison the
alent noise temperature or RJ temperature loading TRJ between use of analog-to-digital converters (ADCs) is expensive and has
40 K and 70 K. The CMB interferometers are therefore very limited bandwidth. Thus, the analog approach is preferred for
competitive at millimeter wavelengths as compared to bolome- observations which require high sensitivity but modest spectral
ters which are usually preferred at shorter wavelengths. resolution. However, the major difficulty in making a broad-
The Yuan-Tseh Lee Array for Microwave Background band system lies in the distribution and processing of the multi-
Anisotropy (AMiBA; Lo et al. 2001; Ho et al. 2009) is a ra- octave signals coming from the antennas. Broadband impedance
dio interferometric array for the study of the CMB at 3 mm matching between components presents a major technical chal-
wavelength. AMiBA detects the minute deviations of the nearly lenge for integrating a large-scale microwave system. Due to the
constant CMB temperature over the sky and can study the spatial nonlinear responses of an analog system, applying appropriate
variation of this temperature fluctuation. In particular, AMiBA drive power levels and modulation/demodulation techniques to
is imaging galaxy clusters via the Sunyaev–Zel’dovich effect minimize the effects of the spurious terms are also very im-
(SZE; Sunyaev & Zel’dovich 1970, 1972; Birkinshaw 1999) for portant. The AMiBA correlator is our attempt to address these
the first time at 3 mm wavelength. The array presently con- technical issues. The four-lag analog correlator has a nominal
sists of 13 elements of 1.2 m reflectors distributed on a 6 m 16 GHz bandwidth. This is currently the only correlator in oper-
diameter platform. The receiver attached to each reflector is ation with an effective correlation bandwidth of around 10 GHz.
dual-polarization and equipped with cryogenically cooled LNAs The scientific goals and design philosophy of AMiBA are
operating in the 84–104 GHz frequency range. The intermedi- presented in Ho et al. (2009). A broader description of the

746
No. 1, 2010 AMiBA WIDEBAND ANALOG CORRELATOR 747

Figure 1. Block diagram of the AMiBA correlator. Signal flow for one baseline (1X2X) is presented. Following the correlator module, the signal flow for a particular
lag (lag 4) is depicted.

AMiBA detection system is given in Chen et al. (2009), while signals from the CCC, it also generates control signals for phase
a detailed description of the AMiBA hexapod mount can be switching, demodulation, and the readout process.
found in Koch et al. (2009). From 2007 to 2008, observations The CCC coordinates all the activities in the correlator, as
were carried out with the seven-element array equipped with well as archiving the data. Equipped with three special purpose
60 cm dishes (Koch et al. 2006). Details on the observations cards from the Australia Telescope National Facility (ATNF),
and analysis of six massive galaxy clusters are presented in the CCC provides timing, data interface, and event signals for
Wu et al. (2009). Subaru weak lensing data of four galaxy correlator operations. Further processing of the archived data
clusters were analyzed with the SZE data to derive the baryon is performed offline. This includes the processing required to
fraction (Umetsu et al. 2009). This paper describes in detail the transform the four lag domain measurements from each lag
instrumental design and testing of the AMiBA correlator. We correlator into two complex channels in the frequency domain.
provide a system overview in Section 2 and major components In a digital lag correlator, the conversion from the lag domain
of the system are described in Sections 3–7. In particular, we to the frequency domain is a simple discrete Fourier transform.
discuss various aspects of the correlator module in Section 4. In In the analog lag correlator, variations in gain and bandpass
Section 8, we outline the testing and data processing. Finally, a of the individual multipliers complicate the transform process
conclusion on the system is given in Section 9. considerably. Calibration with strong point sources such as
planets becomes necessary. By observing a strong point source,
2. SYSTEM OVERVIEW the expected signal from each frequency bin of each baseline
can be calculated. Correction to the transformation can then
As shown in Figure 1, the AMiBA correlator consists of be extracted and applied to the following observations. A brief
five parts, namely the IF distribution, correlation, readout, data description of the relevant data processing and a demonstration
acquisition, and control. For the 13-element array, there would are presented in Section 8.
be a total of (13 × 12/2) × 4 = 312 correlations between
the two polarizations of each receiver in order to obtain the 3. IF DISTRIBUTION
four Stokes parameters for polarization measurements. In the
AMiBA correlator, the number of correlator modules required is An IF from 1 GHz to 21 GHz was proposed in the beginning
reduced by a factor of 2 by introducing two-way switches which of the project (Lo et al. 2001), but was subsequently changed
choose between the two parallel-hand products, XX and YY, to 2 GHz–18 GHz after taking into consideration the avail-
or the cross polarization products, XY and YX. For example, able commercial microwave components,7 the complexity of
observations of the CMB spatial intensity variations require the multi-octave circuit design, and the physical sizes of compo-
measurement of the XX and YY products only. nents. The IF signal distribution uses a mixture of off-the-shelf
The network to distribute the 2–18 GHz IF from the receivers and specially designed microwave components.
to the correlators is implemented by cascading three sections of After passing through the radio frequency (RF) LNAs and
four-way power dividers. The broadband analog lag correlator sub-harmonically pumped mixers (SHMs), which have a con-
modules were designed using double-balanced diode mixers version loss of around 12 dB, the faint signals from the receiver
as the multipliers. The subsequent amplifier provides low-pass inputs reach a level of −45 to −40 dBm. At this point, the noise
filtering with a 3 dB cutoff frequency at around 10 KHz. figure of the subsequent components has a negligible contribu-
In the readout electronics, we use voltage-to-frequency con- tion to the overall noise temperature, as long as the IF power level
verters (VFCs) and counters to provide integrating analog to can be kept well above the thermal noise of about −71 dBm.
digital conversion of the multiplier outputs. This style of ADC Multiple broadband IF amplifiers are used along the IF chain to
has a slow response but a high dynamic range making them well
suited to this application. The data acquisition electronics serve 7 Commercial broadband microwave components usually come with a
as a memory buffer between the readout electronics and the frequency range up to 18 GHz, or up to 26 GHz for the next available
correlator control computer (CCC). In response to timed event frequency range.
748 LI ET AL. Vol. 716

Figure 2. IF power profile from the receiver output to correlator module input (prior to 2009).

Figure 3. Schematic of the AMiBA receiver, including the RF, LO, and IF sections.

compensate for the divider loss. Attenuators are inserted along IF sections are shown in Figures 3 and 4, respectively, for
the IF paths to adjust the IF power level in order not to satu- reference.
rate the amplifiers and also to improve the matching between
components. The input power level to the correlator modules is 3.1. First Section
chosen to optimize the output signal-to-noise (S/N) ratio. The
In the first section, an IF amplifier8 with a gain of 34 dB
diagram of IF power level settings from receiver to correlator is and output 1 dB compression point (P1dB) of 20 dBm is used
shown in Figure 2.
in conjunction with a four-way power divider.9 A directional
Another issue is to keep the IF power stable over a period
coupler10 sends part of the IF signal into a total power detector11
greater than the integration time of each data point. It was found to monitor the IF power after the receiver. The reading can be
that most of the gain variations can be attributed to ambient tem-
used to adjust the variable gain amplifier12 (VGA) in the receiver
perature changes (Nishioka et al. 2009). A ventilation system
using fans and a feedback proportional-integral-derivative con- 8 CMA-18-2004, Teledyne Microwave.
trol was installed to minimize the temperature variations within 9 PDM-44M-10G, Merrimac.
the electronic cabinets (Jiang et al. 2009). 10 1822, Krytar.
The following sections provide a description of each IF 11 302A, Krytar.
12
section. Schematics of the AMiBA receiver and correlator AVG4-02001800-40, Miteq.
No. 1, 2010 AMiBA WIDEBAND ANALOG CORRELATOR 749

Figure 4. Schematic of the correlator IF sections. The gain, insertion loss (I.L.), noise figure (N.F.), and output 1 dB compression point (P1dB) of each component are
listed where applicable.

IF to maintain the power level from the receiver. In addition to

monitoring the receiver gain variations, the total power reading
can also be used for sky dip or hot/cold load experiments to
determine the sky and receiver noise temperature. The input
of the total power readout electronics is switched between the
signal and ground for alternating data. Common mode noise
such as ground noise13 is reduced after subtracting the ground
reference from the signal.
An 18 GHz low-pass filter14 determines the overall system
bandwidth and also filters out the 21 GHz Local Oscillater (LO)
leakage from the mixer. A 13 dB negative slope equalizer15
compensates for the gain slope arising from the following
components. Figure 5. Modular layout of the correlator third section power dividers,
correlator modules, DC amplifiers, and readout boards. The central portion
consists of an array of lag correlators and DC amplifiers. The correlators are
3.2. Second Section fed by power dividers at both ends. Readout boards underneath the horizontal
There are two versions of the second section—one with a power dividers are responsible for VFC ADC and integration.
two-way switch,16 a delay trimmer,17 an IF amplifier, and a
four-way power divider; the other with an IF amplifier, a power 3.3. Third Section
divider, and a delay compensation cable. During polarization
observations, the two-way switches can be utilized to obtain The third section consists of a four-way power divider
four cross-correlations. The delay trimmer has a range of ±90 ps with built-in amplifiers in front. Their dimensions are custom
for fine delay tuning. The coarse delay adjustment is done designed in order to feed the wideband IF signals into an array of
by installing delay compensation cables of designated lengths. correlator modules in a compact way as shown in Figure 5. The
More details about delay trimming can be found in Lin et al. two stages of the MMIC IF amplifiers18 with a cascaded gain of
(2009). 14 dB are placed in front of the power divider to compensate for
the loss in the power divider. A millimeter-wave DC-blocking
13 As the total power readout electronics input is connected to ground through capacitor19 required for bias is placed in front of the first MMIC
a 1 MΩ resistor, the fluctuations on the reading can be seen. Similar effect can and rejects signals below 0.5 GHz. The power divider utilizes
be observed as the readout is connected to the total power signal while the
input of the 1st section is either terminated with a 50 Ω resistor or connected to a Cu-clad substrate with an embedded resistive layer20 which
a stable IF signal. allows thin-film resistors to be etched into the circuit as part of
14 11SL10-1800/X24000-O/O, K&L Microwave.
15 EQ1251-13M, Aeroflex/Inmet. 18 TGA8300-SCC, Triquint Semiconductor.
16 521-420803A, Dow-Key Microwave. 19 500S100GT 50 XT, American Technical Ceramics.
17 981, Aeroflex/Weinshel. 20 Arlon 25N + OhmegaPly.
750 LI ET AL. Vol. 716

Figure 6. Picture of the MMIC IF amplifiers and power dividers used in the
third section of the correlator IF. Bias circuit boards are not shown.
(A color version of this figure is available in the online journal.)

Figure 8. Layouts of the custom four-way power dividers.

(A color version of this figure is available in the online journal.)

To limit the gain loss from bandwidth smearing to an

acceptable level, for an efficiency of sin(x)/x = 0.9 at the
primary beam half-power points, the fractional bandwidth must
Figure 7. Assembly diagram for the IF amplifiers used in the third section of the be at most
correlator IF as recommended by the manufacturer, except the 100 pF bypass δν 1D
capacitors used to remove the oscillation at high frequencies. = , (3)
ν 2B
where B is the baseline length and D is the dish diameter. Given
the fabrication process. However, the same resistive layer causes the final configuration of the array (1.2 m dishes with longest
significant loss and gain slope along a long transmission line. baselines of about 6 m), a frequency resolution of 8 GHz is
For cavity resonance suppression, metallic irises were extended chosen. With a bandwidth of 16 GHz, the number of lags in
down from the cover to confine the circuits into small channels. the lag correlator is then set to four to provide two frequency
A picture of the custom power divider is shown in Figure 6. The channels. Correlator modules with more (e.g., eight) lags would
component layout of the amplification stage and layouts of the not furnish much more information but would minimize the
two types of power dividers are also shown in Figures 7 and 8, bandwidth smearing for the case of smaller diameter dishes.
respectively, for illustration. Details on the circuit design and To examine the bandwidth smearing from another perspective,
test results can be found in Li et al. (2004). Figure 9 shows the calculated bandwidth patterns of the four lag
outputs with an effective bandwidth of 16 GHz. The lag span
4. CORRELATOR
is between ±60 ps, enough to cover the delay for 1.2 m dishes
4.1. Design and Specifications with 6 m baselines.
For the AMiBA correlator, it is important to have a very
In a lag correlator system as for AMiBA, a number of wide bandwidth for CMB observations. To obtain the wide
multipliers are used to measure the correlation as a function bandwidth, analog multipliers in the form of balanced mixers
of the time offset or “lag” between two signals, namely, were employed. The use of passive multipliers circumvents the
 T problem of 1/f noise and other noise from the bias circuitry
1
r(τ ) = V1 (t)  V2 (t) = lim V1 (t)V2∗ (t − τ )dt, (1) usually associated with active multipliers. A flat amplitude
T →∞ 2T −T response and a linear phase response (non-dispersive) are
essential to achieve a large effective bandwidth. From tests with
where τ is the lag between the two signals. The cross-correlation a translating noise source, the entire signal path with the four-
is represented by the pentagram symbol () and V2∗ denotes the lag correlator module has an effective bandwidth of around
complex conjugate of V2 . The cross power spectrum of V1 (t) 10 GHz (Lin et al. 2009). A brief list of the AMiBA correlator
and V2 (t) can be derived via Fourier transforming the cross- specifications is given in Table 1.
correlation Within each correlator module, two stages of two-way power
V1 (t)  V2 (t)  V̂1 (ν)V̂2∗ (ν). (2) dividers are cascaded to split the IF signals to feed four
In a lag correlator of the type described here, the correlation multipliers. The lags for each multiplier are specified to be 37.5,
function is measured at discrete values of lags. The bandwidth 12.5, −12.5, and −37.5 ps, respectively. The 25 ps lag spacing
of the correlator, BW, is determined by the delay increment is designed to provide Nyquist sampling of the IF up to 20 GHz.
δτ between measurements according to the Nyquist sampling The output voltages from the four lags are transformed into two
theorem (Harris & Zmuidzinas 2001). The frequency resolution, complex data points in the cross power spectrum. The double-
δf , is determined by the number of lag measurements, N, such balanced mixers are comprised of low-barrier Silicon Schottky
that δf = BW/(N/2). diode ring quads mounted between two wideband baluns. To
No. 1, 2010 AMiBA WIDEBAND ANALOG CORRELATOR 751

Figure 9. Upper left: recovered spectrum for a four-lag correlator with CW input signals at the center of two channels. Upper right: spectrum for an eight-lag correlator.
Bottom left: bandwidth pattern for a four-lag correlator and delay range for 1.2 m dishes with 6 m baseline. Bottom right: bandwidth pattern for an eight-lag correlator.
(A color version of this figure is available in the online journal.)

Table 1
AMiBA Four-lag Correlator Specifications

Characteristics Specifications
Input frequency range 2–18 GHz
Responsivity 80 Vrms /W minimum
Responsivity variation vs. frequency < 3 dBpp
Phase response < 30 degree peak-to-peak deviation from linear
Delay increment accuracy 25.0 ± 2.5 ps
Input 1 dB compression point >−5.0 dBm
Squared term contribution <5%
Output impedance <100 kΩ

avoid reflections due to discontinuities such as wire bonding

between circuits, the entire circuit, including power dividers
and baluns, is manufactured on a single microwave substrate.
A picture of the four-lag correlator module manufactured by
Marki Microwave Inc. is shown in Figure 10.
Regarding the channel isolation, for an ideal four-lag cor-
relator, the recovered spectrum is Σi ci e−j ω τi , where ci is the
correlation measurement at each lag and τi is the nominal lag.
If the input signals are continuous-wave (CW) signals at the
center of each channel, namely 6 GHz and 14 GHz in our case,
the recovered spectrum is shown in Figure 9 with a FWHM of
about 12 GHz for each CW signal. Therefore, the leakage be-
tween channels is quite severe for the four-lag correlators. For
comparison, the recovered spectrum of an ideal eight-lag cor-
relator with input CW signals at the center of its four channels
is also shown. Although the leakage between adjacent channels Figure 10. Four-lag correlator module manufactured by Marki Microwave Inc.
is still severe for an eight-lag module, the isolation between The power dividers and multipliers can be seen.
non-adjacent channels is acceptable. To improve on the channel (A color version of this figure is available in the online journal.)
isolation, one possibility is to apply a window function during
the lag-to-visibility transformation. However, constrained by the 4.2. Optimum Input Power
small number of lags that we have, another possibility might be
to apply bandpass filtering or a filter bank before correlation. During observations, the correlator input signals consist of
The filter bank might also reduce the bandpass variation within large uncorrelated noise signals from the receivers and tiny
each channel and maximize the effective bandwidth. correlated signals from the sky. These large uncorrelated noise
752 LI ET AL. Vol. 716
100 to 700 counts, including the back-end noise, as shown in
Figure 12.
As the correlators are operated in the linear regime, it is
seen that the correlator output fluctuations are proportional
√
to pi pj , where pi,j are the input power of each baseline
(Wrobel & Walker 1999). Since there is no total power detector
at the correlator module inputs, the input power cannot be
measured directly, but can be inferred from the measured rms
of each baseline. In the compressed regime, the output rms
grows monotonically, although not linearly, as the input power
increases. From the rms value, we can identify which of the
three regimes (under-pumped, linear, and compressed) we are
operating in. The gain can then be adjusted to optimize the S/N.

4.3. Spurious and Back End Noise

An ideal double-balanced mixer used as a multiplier can sup-
press even harmonics of both input signals and their products,
i.e., v1n , v2m , and v1n v2m , where either n or m is even and vi (i = 1, 2)
is the input voltage (Maas 1993). From simulations, it has been
found that with a double-balanced mixer, the spurious terms
such as v12 , v22 , and v12 v22 are suppressed by over 40 dB below the
desired product v1 v2 . Experiments show that residuals of these
spurious terms can still be seen. They can be further suppressed
by the use of a phase switching/demodulation process which
Figure 11. Upper: with v1 and v2 as the large uncorrelated noises, the output allows the introduction of a DC blocking capacitor at the input
from a multiplier can be represented as the summation of various possible terms of the low-frequency amplifier following the correlator module.
v1m , v2n , and v1m v2n , where m and n are integers. Each output term can contribute
to correlator output fluctuations. On the other hand, v3 and v4 present the small
Phase switching is achieved by switching a suitable delay into
correlated signals and v3 v4 is the expected product. A current noise source is the LO path. However, as the LO signal is switched through
used to represent the back-end noise from the DC amplifier and the VFC ADC different paths, a small amplitude modulation of the IF power
at the input of the DC amplifier. Bottom: in the simulation of a double-balanced level is introduced. This modulation survives the demodulation
mixer with four tones, when the large signals have power below a certain level process and appears as an offset in the v12 v22 term. To minimize
(−2 dBm), the product of two small signals (red curve) drops dramatically. As
the input power of the large signals increases above the threshold, products of offset, variable attenuators are used to adjust the 42 GHz LO
both small signals and large signals (black curve) increase linearly. As the input power to the SHM to minimize the IF power modulation.
power keeps increasing (above 8 dBm), eventually both small-signal and large- S/N is optimized by operating the correlator in a slightly com-
signal products get compressed. Since a standard diode model is used during pressed regime, thereby ensuring that front-end noise dominates
the simulations, the input power levels of the large signals where the diodes are
sufficiently pumped or compressed are different from our measurements.
back-end noise from the DC amplifier and readout electronics.
Figure 12 shows the output S/N estimated from the Jupiter
(A color version of this figure is available in the online journal.)
fringes as the input power were varied. Most of the lag outputs
show an increase in S/N as the input power (inferred from the
signals tend to pump the correlator diodes in much the same output rms) increases, even when some of them were driven into
way as the local oscillator signal in a mixer application. This the compressed regime.
can lead to nonlinearity of the correlator multipliers and an A correlation interferometer where a multiplier is employed,
excess of noise. We adjust the correlator input power level to the minimum detectable temperature or sensitivity is
reduce this degradation.
In a simulation of a double-balanced mixer with four tones Tsys
ΔTmin = √ , (4)
(two with large power and two with 28 dB smaller power) 2Δντ
as shown in Figure 11, when the large signals have a power
below a certain level, the product of the two small signals drops where Tsys is the system noise temperature, Δν is the bandwidth,
dramatically. As the input power of the large signals increases and τ is the integration time, assuming 100% efficiency (Kraus
above the threshold, the products of both small signals
√ and large 1986). From the correlator output fluctuations, we can estimate
signals increase linearly, i.e., pout proportional to p1 p2 . As the noise contributions from each part of the system, namely the
the input power keeps increasing, eventually both small-signal front end (v1m , v2n , and v1m v2n terms, where m and n are integers,
and large-signal products become compressed. We can refer v1 and v2 represent the large uncorrelated noises.) and back
the large signals in the simulation to the uncorrelated noise end (Figure 11), assuming all noises are uncorrelated, i.e., their
from the receivers, and their product as the output fluctuations variances are additive. Typical correlator output fluctuations
of the correlator, as the output fluctuation of the correlator is measured under different conditions are listed in Table 2 to
indeed due to the beating or mixing of uncorrelated signals illustrate how the output noise might increase due to the spurious
(Kraus 1986). The small signals in the simulation are regarded terms. As a result, noise contributions from the back end, as well
as the correlated signals in the system. The simulation defines as the spurious v1m , v2n terms can be estimated. Currently these
three regimes of operation, namely, underpumped, linear, and spurious terms and the back-end noise reduce the S/N by 20%.
compressed. From experiments, our correlator modules have An analysis of the system efficiency based on the observations
a linear range for input power from −20 dBm to −12 dBm, with the seven-element array can be found in Lin et al.
corresponding to output fluctuations with an rms from around (2009).
No. 1, 2010 AMiBA WIDEBAND ANALOG CORRELATOR 753

Figure 12. To determine the IF power for optimum S/N at correlator outputs, drift scans of Jupiter were taken with different input power levels to the correlator. The
input power were varied by adjusting the VGAs along the IF paths. The input S/N was fixed and could be referred to the values at the receiver inputs. Output from
the four-lag correlator module, usually referred to as the lag output, can be fitted to estimate the signal strength. During data processing, each of 168 lag outputs of the
seven-element array is designated with a trace number. Upper left: one lag output (trace 41, labeled as tr41) is plotted (green crosses) and curve-fitted (red line) in terms
of counts after the VFC ADC. Upper right: high-pass filtering is applied to the data to remove the low-frequency terms and the signal. The noise is estimated from the
rms of the remaining fluctuations (green crosses). The responses of the lag correlators vary. Lower left: the S/N of several lag outputs (circles) is plotted against the
rms values. Lines are drawn through data points of each output, and different colors are used for distinction. Lower right: for four lag outputs of one baseline, output
fluctuations (rms in counts), including the back-end noise, are plotted (circles) as a function of the input power.
(A color version of this figure is available in the online journal.)

Table 2
Typical Results of the Correlator Output Fluctuations or rms Measured Under Four Conditions

Conditions Measurements Estimations

1 2 3 4
Ant 2 (v1 ) Off off on on
Ant 3 (v2 ) Off on off on
YY lag 3 rms (counts) 95 783 805 1850 1120 1473
Noise Back end v1m terms, v2n terms, Overall v1m terms, v1m v2n terms
composition back end back end v2n terms,
back end

Notes. When both receivers were off, the output noise could be attributed to the back-end noise from the DC amplifier
and the ADC. When only one antenna was on, e.g., Ant 2, the output of the multiplier would include various terms
from V1 . The situation was similar when only Ant 3 was on. For an ideal multiplier, the v1m and v2n terms are considered
spurious. From the measurements, these terms also contributed to the output noise so that correlator output rms increased.
Assuming that noise from each term can be summed in quadrature, we can estimate the output fluctuation when there is
no v1m or v2n terms at the output and no back-end noise either, so as to determine how much the S/N has been degraded.

4.4. Low-frequency Amplifier the phase switching frequency. A DC blocking capacitor is used
in front of the amplifier to remove any DC term from the mixer
For the low-frequency LNA following the lag correlator output. As a result, the “DC” amplifier has a 3 dB passband
module, because of the high output impedance of the mixers, from 0.1 to 9 kHz. The schematic of the DC amplifier is shown
low noise current amplifiers21 were chosen to minimize the in Figure 13.
back-end noise. The bandwidth of the amplifier is limited in
the feedback loop in order to reduce the output noise. Due to 5. READOUT ELECTRONICS
the phase switching, the correlated signals are square waves, at
The correlator readout circuit uses a VFC plus a 24 bit counter
21 OPA627AU. as the ADC. The VFC generates pulse sequences at a frequency
754 LI ET AL. Vol. 716

Figure 13. Schematic of the low-frequency or “DC” amplifier.

linearly proportional to the analog signal from the DC amplifier correlation and total power data. There are two steps involved in
output. The up/down counter accumulates the pulses, acting as transferring the data from the readout circuits to the CCC, scan,
a long-term integrator as well as a phase switching demodulator. and DMA. Each step is triggered by events originating from the
The up/down function is controlled by the demodulation signal. event generator (EG) in the CCC. The DMA process is ideal
In AMiBA correlator, there are four layers of processes used for transferring large volumes of data but is more efficient if the
to remove systematics. The correlation process suppresses the data are stored in consecutive addresses.
uncorrelated receiver noise and gain fluctuations. The phase For the seven-element array, Walsh functions of 64 intervals
switching and demodulation process reduces the offsets or gain per cycle are used as the phase switching signals. Currently there
drifts that do not have the characteristics of the demodulation are five cycles in each integration of 0.226 s. This corresponds
signals. One additional phase switching layer is implemented by to a fastest switching frequency of around 700 Hz. For the
inverting the sign of the demodulation signals for every other in- 13-element array, Walsh functions of 128 intervals per cycle
tegration. Every two data points are then subtracted with respect are used. Since the corner frequency of the 1/f noise from the
to each other offline in the CCC. From analysis, the correlator passive diode multiplier we use is low, phase switching at rather
output spectra show a white noise signature between 10−4 and low frequencies is feasible.
1 Hz, with an increase in power at lower frequency due to slow
gain drifting (Nishioka et al. 2009). For the remaining offsets or 7. CONTROL
false signals due to ground pick up, a two-patch scheme (Padin
The operation of the AMiBA correlator is controlled by
et al. 2002) is adopted. Observations are taken with, for our
the CCC—an industrial grade PC running Linux. The CCC
case, 3 minute tracking on the main field, and another 3 minute
is equipped with three special cards—EG, Australia Telescope
tracking on a trailing or leading field at the same declination,
Distributed Clock (ATDC), and PCI Correlator Data Bus In-
separated by a few minutes in right ascension. The main and the
terface (PCIIF) from ATNF.23 The function of the EG is to
trailing/leading fields then share the same azimuth-elevation
generate events with precise timing. The PCIIF acts as an inter-
track with identical ground contamination of the data.
face between the data acquisition circuit and the CCC to receive
At the end of an integration interval, the contents of the
the DMA data. The interface appears as a 256 kbyte block of
counters are dumped into shift registers and serially scanned
out by the data acquisition electronics. For the 13-element ar- memory on the PCI bus while the PCIIF also assigns mem-
ory addresses and several control signals for the DMA process.
ray, instead of the custom readout ICs used previously in the
Interrupts from the EG are relayed to the PCIIF for DMA tim-
seven-element array (Li et al. 2004), discrete VFC components22
are used. The digital section (counters/shift registers) is imple- ing. The ATDC provides precise timing for all the correlator
operations and also generates an 8 MHz reference clock for dig-
mented in a field programmable gate array (FPGA). The new
ital processing in the data acquisition and readout circuits. The
VFC is synchronous and has better linearity than the previous
readout ICs. A schematic of the readout electronics with the ATDC can phase-lock to both a 5 (or 10) MHz sine wave and a
1 pulse-per-second signal from a GPS receiver and is scheduled
timing diagram of the control signals is shown in Figure 14.
to synchronize the system clock of the CCC periodically. An
alternative is to synchronize the system clock of the CCC to the
6. DATA ACQUISITION GPS receiver via the network time protocol. A block diagram of
The data acquisition circuit is used to store data from the correlator control components and signals is shown in Figure 15.
readout circuits before they are transferred to the CCC via a
direct memory access (DMA) process. Control signals, such 8. TESTING AND DATA PROCESSING
as readout control signals and phase switching/demodulation To test the response of each baseline, a W-band noise source
signals are also generated here. RAM blocks configured as is set up to translate between the two receiver inputs. The data of
128 × 24 bit RAMs within the FPGAs are allocated to store the lag sequence can be Fourier transformed to obtain the bandpass
22 AD7741, Analog Devices. 23 http://www.atnf.csiro.au/technology/electronics/
No. 1, 2010 AMiBA WIDEBAND ANALOG CORRELATOR 755

Figure 14. Upper: the timing diagram of the control signals and the schematic of the VFC/counter ADC used in the correlator for the 13-element array. Signals (dump,
scan, reset, and demod—demodulation) are applied to the counter only when counting is stopped (“freezed”). “Freeze” needs to be buffered before being applied to
the counter to avoid glitches. Since AD7741 is a synchronous VFC, i.e., the output pulse is initiated by the edge of the clock (CLKIN), “freeze” signal is buffered with
a flip-flop triggered by the clock of the VFC. Bottom: schematic of the scan block used to transform 24 bit data into series of bits for the scan-out process.

Figure 15. AMiBA correlator control block diagram.

756 LI ET AL. Vol. 716

Figure 16. Cleaned images of Jupiter. Left: the image was formed with uncalibrated visibilities directly transformed from correlator outputs. Right: the constituent
visibilities of the image have been calibrated by another set of Jupiter data. Both images are plotted with the same residual noise after CLEAN. The green circle
indicates the FWHM of the primary beam, and the blue shaded ellipse at the bottom right corner represents the synthesized beam of the seven-element array in the
compact configuration.
(A color version of this figure is available in the online journal.)

response. The bandpass shows a two-hump gain response and fraction of the synthesized beam) around the location at which
some scatter from a linear phase response. From the bandpass the calibration source is observed. At other locations within the
response, we can estimate the effective bandwidth Be as field of view, there are gain and phase errors set by uncertain-
  ties in the lag-to-visibility transformation that affect the quality
 W (f ) df 2 of the image reconstruction, although further calibration should
Be =  , (5) reduce these errors. However, further simulations show that ap-
|W (f )|2 df
plying the calibration to the entire field of view only contributes
where W (f ) is the complex bandpass response. The baselines about 2% in rms to the recovered point source flux after com-
of the seven-element array were estimated to have effective bining data from all baselines (Lin et al. 2008).
bandwidths ranging from 9 to 13 GHz (Lin et al. 2009). Although using a translating noise source at the receiver
With the derived bandpass responses, a transform matrix K inputs to create artificial fringes is a useful method to measure
can be created so that the transfer function of the correlator, it is hard to run the test
inline with observations. The spectral resolution obtained with
[K][S] = [R], (6) our current setup was not sufficient. This approach becomes
more tedious as the number of baselines increases and we would
where S is the complex column vector representing the input still need to use the planet data to calibrate the gain and phase
spectrum and R is the real column vector for the correlator of each baseline. An alternative would be to derive the response
output. K and S include both positive frequency terms and matrix K in two bands, assuming nominal responses of an ideal
their complex conjugate at negative frequencies to yield real lag correlator. Data from tracking of strong point sources such
correlation products (Li et al. 2004). as Jupiter, Saturn, or Mars, performed regularly at an interval
With the correlation outputs from each baseline, we can invert of approximately 3 hr during the observations, can be used for
K and then derive the spectra of the input signals or visibilities calibration. All data analysis and calibrations were restricted
S via to observations made at the phase center at the moment.
[S] = [K]−1 [R]. (7) Figure 16 shows the images of Jupiter from visibilities with and
without calibration. The calibrator was another set of data of
However, since the four lag outputs only give us two independent Jupiter taken several minutes apart. The uncalibrated visibilities
complex data points in the frequency spectrum, K is rather suffer from errors in the transformation as mentioned above.
singular due to degeneracy. The singular value decomposition Forming an image directly results in very strong cancellation.
method can be used to invert K. Once K −1 is obtained, the After calibration, the visibilities add up coherently and form a
spectra of n (n > 2) channels can be obtained. Eventually, we strong point source at the phase center. The images have been
have to consolidate those n channels into two bands. Another deconvolved using the CLEAN algorithm (Hogbom 1974) and
approach would be to integrate the complex response matrix plotted with the same dynamical range. More details about data
K in frequency into two bands first. The integrated matrix K̄ processing can be found in Wu et al. (2009) and Lin et al. (2009).
can be easily inverted, and the two-band visibility output can be
written as [S̄] = [K̄]−1 [R]. 9. CONCLUSION
From simulations, Lin et al. (2009) showed that given an
accurate and high spectral resolution measurement of K, the From observations, the AMiBA correlator has proven to have
visibility can be recovered regardless of the source offset from the sensitivity required for CMB detection. The analog multi-
the phase center. On the other hand, inaccurate estimate of K pliers used in the lag correlators provide the wide bandwidth
results in errors in the recovered visibility. The error varies with required for high sensitivity. The inherent noise rejection of
the source offset and can be calibrated at any given offset by ob- the interferometer is also very beneficial. Compared with the
serving a point source at the same offset from the phase center. filter bank scheme used by CBI (Padin et al. 2002), the lag cor-
The calibration is strictly valid only in a small region (a small relator design is simpler and more compact. Large bandwidth
No. 1, 2010 AMiBA WIDEBAND ANALOG CORRELATOR 757
with a small number of lags does not present a significant chal- REFERENCES
lenge for the lag-to-visibility transformation after proper cali-
brations. At the moment, the effective bandwidth we can achieve Birkinshaw, M. 1999, Phys. Rep., 310, 97
Carlstrom, J. E., et al. 2002, ARA&A, 40, 643
is limited by bandpass variations due to wideband impedance Chen, M.-T., et al. 2009, ApJ, 694, 1664
matching. However, a significant portion (60%) of the nomi- Harris, A. I., & Zmuidzinas, J. 2001, Rev. Sci. Instrum., 77, 1531
nal bandwidth has been achieved. To improve on the bandpass Ho, P. T. P., et al. 2009, ApJ, 694, 1610
variations, it is possible to operate the correlator at a higher IF, Hogbom, J. A. 1974, A&AS, 15, 147
Holler, C. M., et al. 2007, A&A, 464, 795
since the response of the analog multipliers is not limited to Jiang, H., et al. 2009, I2MTC 2009 (Los Alamitos, CA: IEEE), 1268
frequencies of few GHz. Thus, with a similar fractional band- Koch, P., et al. 2006, in Proc. EuCAP 2006, ed. H. Lacoste & L. Ouwehand
width, a larger bandwidth could be achieved. Wideband com- (ESA SP-626; Noordwijk : ESA), 668.1
plex correlators in conjunction with bandpass filters can also Koch, P., et al. 2009, ApJ, 694, 1670
be considered. For AMiBA, complex correlators with a band- Kraus, J. D. 1986, Radio Astronomy (2nd ed.; Durham, N. H.: Cygnus-Quasar)
Leitch, E. M., et al. 2002, ApJ, 568, L28
width of 8 GHz are suitable. By interleaving a number of them Li, C.-T., et al. 2004, Proc. SPIE, 5498, 455
in frequency, a large bandwidth is feasible. The filters would Lin, K.-Y., et al. 2008, Proc. SPIE, 7012, 701207
also improve the isolation between channels. Similar analog Lin, K.-Y., et al. 2009, ApJ, 694, 1629
interferometric systems can be constructed for high sensitiv- Lo, K. Y., et al. 2001, in AIP Conf. Proc. 586, Relativistic Astrophysics, 20th
Texas Symp., ed. J. C. Wheeler & H. Martel (Melville, NY: AIP), 172
ity, high angular resolution, and moderate frequency resolution Maas, S. A. 1993, Microwave Mixers (2nd ed.; Norwood, MA: Artech House)
observations. Nishioka, H., et al. 2009, ApJ, 694, 1637
Padin, S., et al. 2002, PASP, 114, 83
We thank the administrative staff for their support over the Runyan, M. C., et al. 2003, ApJS, 149, 265
Sunyaev, R. A., & Zel’dovich, Y. B. 1970, Comments Astrophys. Space Phys.,
years. We thank A. Harris for useful discussions and notes. We 2, 66
thank the Ministry of Education, the National Science Council, Sunyaev, R. A., & Zel’dovich, Y. B. 1972, Comments Astrophys. Space Phys.,
and the Academia Sinica for their support of this project. We 4, 173
thank the NOAA for accommodating the AMiBA project on Umetsu, K., et al. 2009, ApJ, 649, 1643
their site on Mauna Loa. We thank the Hawaiian people for White, M., et al. 1999, ApJ, 514, 12
Wrobel, J. M., & Walker, R. C. 1999, in ASP Conf. Ser. 180, Synthesis Imaging
allowing astronomers to work on their mountains in order to in Radio Astronomy II, ed. G. B. Taylor, C. L. Carilli, & R. A. Perley (San
study the universe. This work is partially supported by the Francisco, CA: ASP), 171
National Science Council of Taiwan under the grant NSC98- Wu, J.-H. P., et al. 2009, ApJ, 694, 1619
2119-M-001-024-MY4. Zwart, J. T. L., et al. 2008, MNRAS, 391, 1545