Conversion from linear to circular polarization in FPGA
K. Das^{1}  A. L. Roy^{1}  R. Keller^{1}  G. Tuccari^{2}
1  Max Planck Institute for Radio Astronomy (MPIFR),
Auf dem Hügel 69, Bonn, Germany
2 
Istituto di Radioastronomia, via P. Gobetti, 101 40129 Bologna, Italy
Received 31 August 2009 / Accepted 26 October 2009
Abstract
Context. Radio astronomical receivers are now expanding
their frequency range to cover large (octave) fractional bandwidths for
sensitivity and spectral flexibility, which makes the design of good
analogue circular polarizers challenging. Better polarization purity
requires a flatter phase response over increasingly wide bandwidth,
which is most easily achieved with digital techniques. They offer the
ability to form circular polarization with perfect polarization purity
over arbitrarily wide fractional bandwidths, due to the ease of
introducing a perfect quadrature phase shift. Further, the rapid
improvements in field programmable gate arrays provide the high
processing power, low cost, portability and reconfigurability needed to
make practical the implementation of the formation of circular
polarization digitally.
Aims. Here we explore the performance of a circular polarizer implemented with digital techniques.
Methods. We designed a digital circular polarizer in which the
intermediate frequency signals from a receiver with native linear
polarizations were sampled and converted to circular polarization. The
frequencydependent instrumental phase difference and gain scaling
factors were determined using an injected noise signal and applied to
the two linear polarizations to equalize the transfer characteristics
of the two polarization channels. This equalization was performed in
512 frequency channels over a 512 MHz bandwidth. Circular polarization
was formed by quadrature phase shifting and summing the equalized
linear polarization signals.
Results. We obtained polarization purity of 25 dB corresponding to a Dterm of 0.06 over the whole bandwidth.
Conclusions. This technique enables construction of broadband
radio astronomy receivers with native linear polarization to form
circular polarization for VLBI.
Key words: polarization  methods: statistical  techniques: photometric  instrumentation: polarimeters
1 introduction
Circular polarizers play important roles in modern communication systems including those in radio astronomy. To obtain higher sensitivity and frequency coverage for spectral line observations, the radio antennas are moving to broadband feeds and extremely broad bands are most easily realized with linearly polarized feeds due to the difficulty of producing phase shift accurately over wide bandwidth. However, circular polarization is simplest for the application of very long baseline interferometry (VLBI), which enables astronomical sources to be resolved with submilliarcsecond synthesized beam widths, since linear dipoles do not generally remain parallel to each other in a global array due to different parallactic angles at different stations when observing the same source, causing loss of coherence in the crosscorrelation products formed between stations. That loss could be recovered were one to compute also the crosspolarization cross correlation products to retain all information (doubling the correlator power needed), or one could rotate the receiver packages at each station to keep the dipoles parallel (requiring mechanical rotators). In contrast, use of circular polarization causes the parallactic angle differences between stations to add a simple phase rotation angle to the measured visibility, which can be predicted from the known observation geometry and subtracted in postprocessing.
Circular polarizers with broad bandwidths have been realized in the past with a number of methods. Most common are the septum polarizer by Chen & Tsandoulas (1973), corrugated waveguide phase shifter, Boifot junction (RuizCruiz et al. 2006), and the linear quadridge OMT followed by a hybrid junction. All are analogue techniques and produce a perfect phase shift and hence perfect polarization purity at only one, two, or three frequencies and the phase errors grow larger at frequencies away from those design points, which ultimately limits the bandwidth of the devices. In contrast, digital techniques offer the possibility to produce an accurate phase shift over broad bandwidths, but this potential has not yet been fully developed. One example is the Westerbork synthesis radio telescope, which converts native linear polarization to circular polarization by a combination of analogue and digital techniques for VLBI. During down conversion of the orthogonal linear signals, the shift is added to the (analogue) LO for one polarization. After analoguetodigital conversion, the (2 bit) signals are summed and differenced to form circular polarization, with a weight that corrects for the average receiver gain differences. The weights are determined by a separate measurement using a calibration noise source in each frontend and are updated every 10 s. This system operates on a bandwidth of 20 MHz, yielding one phase and amplitude correction for each 20 MHz of bandwidth (Boss 2007, private communication). The Westerbork system uses the preexisting correlator and analogue phase rotation in the LO system, but most VLBI stations lack this equipment and so another, more general, solution is needed.
We have developed a selfcontained digital processing system in which the correlator, channel equalization, phase rotation, gain scaling, quadrature phase shift and summation to form circular polarization are contained in a standalone unit. It accepts two intermediate frequency inputs with orthogonal linearly polarized signals, each of 500 MHz bandwidth and subdivides the band down to 1 MHz resolution. In each 1 MHz piece it measures the phase and amplitude differences between the orthogonally polarized channels using a calibration noise source in the front end that is common to both polarizations. It then uses those measurements to equalize the channel phases and amplitudes during observations. After equalization it introduces an ideal phase shift into one polarization channel and forms sum and difference outputs that respond to orthogonal circular polarizations at the input.
Digital systems offer the ability to process continuous data in real time implementing automatic data processing algorithms. Data transmitted digitally are more resistant to external interference and hence digital devices supersede analogue counterparts at least where speed and signal purity are matters of concern. With the advent of logic devices like ASICs and FPGAs, it is possible to implement many complex algorithms, which would have been impossible otherwise and they also offer the user ease of replication compared to analogue systems. For these reasons our system to produce circular polarization generates almost ideally perfect orthogonal field components in real time.
This paper acquaints the reader with the theory and derivation for obtaining instrumental phase and gain correction factors and applying them to the two received orthogonal linear polarizations to form phase and gain calibrated left hand circular (LHC) and right hand circular (RHC) polarizations. It explores limitations of system performance due to the most influential factors, which are the Dterms and receiver instabilities and explores the requirement for periodic recalibration to remove their detrimental effects on polarization purity. It describes the digital signal processing in FPGA and the experimental verification of the technique to show that good polarization purity is obtained.
2 Theoretical development
2.1 Overview of method
We measure the receiver transfer characteristics using a noise diode signal as a calibration source, injected into the feed as pure linear polarization in a plane oriented at to the two orthogonal linear dipoles that couple energy out of the waveguide. The noise diode injects signals into the x and y chains that are identical, having equal amplitude, zero phase difference, and 100% coherence with each other. These noise diode signals during calibration pass through the same x and y receiving chains as does the astronomical signal later, and are sampled at IF at 1 GS/s (i.e. 500 MHz contiguous Nyquist bandwidth). The sampled x(t) and y(t) signals are processed in an FX correlator on the FPGA, which transforms to frequency domain with 1 MHz channel widths, cross multiplies each spectrum against the corresponding spectrum and integrates for 8 s. We chose a spectral resolution of 1 MHz to allow for possibly rapidly changing channel phase differences with frequency. The result is a phase spectrum with low thermal phase noise that represents the phase difference between the x and y channels due to the transfer characteristics of the receiver chains. The phase spectrum is used during later astronomical observation for equalizing the (frequencydependent) phase lengths of the x and y receiver chains. The x and y bandpass amplitude shapes are also equalized, using gains derived during calibration from totalpower spectra of and accumulated during the calibration stage.
To form circular polarization from native linears during astronomical observations, we need likewise to transform the x(t) and y(t) time series to frequency domain in the same manner as during calibration, then equalize the transfer characteristics by applying a phase rotation to each frequency channel of one polarization and an amplitude scaling to each frequency channel of both polarizations in an equalizer stage, then simply add or subtract (equivalent to exchanging real and imaginary in the complex spectra), and summing to form results that respond to the two hands of circular polarization.
2.2 Instrumental phase and gain calibration
The orthogonal timedomain field components x(t) and y(t) received by the crossed dipoles undergo unequal phase and amplitude distortions due to different frequencydependent time delays and gains of the two receiving systems through which they pass. The phase and gain calibration aims to compensate the transfer characteristics of channel x and channel y to make them identical in x(t) and y(t), reducing instrumental artifacts to zero. This approach is already used in software for calibrating radio astronomical data though those operate on stored data rather than in real time. Our effort is to extend it to the digital domain processing sampled IF signals in real time, calibrating with fine frequency channels, to enable formation of circular polarization in real time with more accurate phase and magnitude response than the analogue techniques can achieve over broad bandwidths. We have not considered channel nonlinearities and multipath effects. Nonlinearity spreads the output spectrum beyond the input spectrum by introducing new frequency components and causes amplitude distortion. Therefore, radio astronomical receivers have to be designed to be linear. In the presence of strong RFI nonlinearity does occur and has to be blanked. Techniques for RFI mitigation are a sizable study in themselves and are beyond the scope of the present work. Nevertheless RFI mitigation techniques can easily be implemented in the same digital hardware. For the present development we assume linear transfer characteristics and Gaussian signal statistics. We have filtered the passband to avoid aliasing and kept signal levels in the linear regime. A linear time invariant system causes only pulse dispersion and amplitude scaling.
2.3 Phase equalization
Let us consider the noise diode signal during calibration as a
broadband source radiating Gaussian random signals continuously in
time, s(t), and we receive and sample a finite number, ,
of frames of timedomain samples each consisting of N samples spaced equally in time in two orthogonal linear polarization states, x(t) and y(t). Let the sampled time series be represented by x_{i}(t) and y_{i}(t) and the noise diode signal at the sample times be s_{i}(t) where
.
It is convenient to transform these time series into the frequency
domain since the transfer characteristic of the receiving system used
for calibration is frequency dependent. The Fourier transform produces
spectra, each consisting of N channels.
depends on the sampling rate, ,
the number of samples in a spectrum, N, and the total integration time,
, as
(1) 
Let a timedomain signal have length T_{0}, then the corresponding frequencydomain spectrum will have channels spaced at an interval f_{0} of frequency units. For a simplified analysis we consider one frequency component, the results from which hold good for all other spectral components in the band. The signals x_{i}(t) and y_{i}(t) are represented by the equations
where h_{x}(t) and h_{y}(t) are the transfer functions of channel x and channel y, t_{xy} is the xy propagation time difference from the dipoles to the receiver inputs (samplers for a digital receiver), and n_{xi}(t) and n_{yi}(t) are external unwanted signals (astronomical sources, thermal fluctuations and spurious sources). After transforming to the frequency domain these relations are
where and are the phases of and respectively and and are the phases of the transfer functions and respectively. The phase difference is due to the initial and instrumental phase difference between and . Now, let us consider the samples of and at uniform intervals of . If and are the rth (channel number) samples of and respectively, where , then Eqs. (4) and (5) can be rewritten in discrete form (Cooley & Tukey 1965) as
where K_{1} is the fractional sample delay caused by t_{xy} and is the sampling period. It is assumed that the system is linear. Thus the transfer function causes only linear distortion and no new frequency components are produced. The resulting pulse is dispersed in time and amplitude rescaled. The phase difference is obtained from the accumulated cross power spectrum of and in the frequency domain, which is expressed as
Substituting Eqs. (6) and (7) in Eq. (8) we obtain the product , which consists of the summation of the following contributing terms
where . is accumulated for time. Since S_{i} is incoherent with N_{Xi} and N_{Yi}, their cross product terms will average to zero in the summation and so are not shown here. We have chosen, in practice, 8 s integration time corresponding to frames of data so that the thermal fluctuations of the noise diode signal are averaged down to a fractional fluctuation of in 1 MHz channels that is 5% of the system temperature. Thus we attain our objective of measuring the transfer characteristic phase to precision. At such high levels of precision, the measurement is sensitive to corruption by possible external sources or by RFI via the second term Eq. (10). We cancel this effect on our measurement of the transfer characteristic, provided the corrupting source remains constant, by performing the summation in Eq. (8) twice, first in the presence of s_{i}(t) (noise diode switched on) and second with s_{i}(t) = 0 (noise diode switched off) and differencing. Hence, considering only the first term (Eq. (9)) we obtain the phase difference between the two polarization signals due to instrumental effects and due to any initial phase difference as follows
where, and ^{}. For accurate calibration, should be measured and used later to correct the signals for the effect of . This should reduce the phase difference to zero, requiring that the initial phase difference be zero so that corresponds to instrumental phase errors only. This requires that the two orthogonal dipoles receive signal from a common source placed at a location equidistant from both the dipoles, which is accomplished by placing a noise source at to the two dipoles. One could alternatively calibrate using an unpolarized astronomical source instead of the noise diode, and point on source and then off source for the two accumulations to be differenced. In the onsource position, the dish main beam must be pointed accurately to the source, placing the source on the focal axis. The dipoles are located in a plane accurately perpendicular to the focal axis and hence, the dipoles would be equidistant from the source to high accuracy, so would be close to zero and can be ignored. gives the angle through which one of the vectors say must be rotated to make the phase lengths of the two polarization channels equal. The frequencydependent rotation matrix elements and are obtained without trigonometric functions for computational efficiency using the relations
Let be the vector obtained after rotation of vector by the phase difference . Then the phase difference between and reduces to zero or
as required for a correctly calibrated system phase prior to formation of circular polarization.
2.4 Gain equalization
The amplitudes also undergo linear distortions due to different gains
in the two channels due to different passband characteristics. To
compensate the amplitude differences, all the spectral components of
and
in the passband are scaled to one same level, chosen to be the maximum signal level,
in the passbands of
and
.
Multiplying Eq. (4) with its complex conjugate, we obtain the power spectrum
and similarly Eq. (5) yields
.
The scaling factors to equalize the magnitudes of
and
are obtained from accumulating
and
respectively for
to reduce thermal noise fluctuations in the measurement of the passband shapes and are expressed by the following two equations:
Expanding these using Eqs. (6) and (7) we find that each contains the following contributing terms for summation:
where and are the phases of and respectively.
The first term (Eq. (17)) provides the measurement of the bandpass
shape that we seek (assuming that the noise diode produces white
noise). The second term (Eq. (18)) does not average to zero to the
extent that noise sources other than the noise diode (e.g. receiver
noise, radio sources, atmospheric emission, cosmic microwave
background, and RFI) are present. We cancel this term, provided the
corrupting source remains constant, by performing the summation in
Eqs. (15) and (16) twice, first in the presence of s_{i}(t) (noise diode switched on) and second with
s_{i}(t) = 0 (noise diode switched off) and differencing. Hence, we are left with only the first term (Eq. (17)) where
and
are the frequencydependent amplitude scaling factors. To determine the gains
and
that scale each
and
to equalize the passband amplitudes to the same level, the following two equations are used
in Eqs. (19) and (20) is obtained by comparing all and of Eqs. (15) and (16) for and finding the maximum power. The gains in Eqs. (19) and (20) are calculated using accumulated power spectra and are later applied to voltage spectra for equalization, which potentially introduces a small inaccuracy since the gains thus obtained are not the same as the actual gains obtained by accumulating individual voltage spectra consisting of absolute voltages in the denominators. However, if the absolute voltages, and , where for the subsequent spectra acquired during observations, are accumulated for sufficient integration time then
in Eqs. (21) and (22) is the maximum of all and for . Equations (21) and (22) are also confirmed in numerical simulation.
2.5 Windowing
A window function is derived for the X_{j} and Y_{j} spectra to truncate the possible analogue filter flanks to avoid scaling up, by large factors, signals that have been strongly attenuated by bandlimiting filters. The window function is conveniently obtained from the spectrum since the Z spectrum contains the band common to both X and Y spectra thereby providing necessary frequency shift and bandwidth information for the window function. Spectral channels in which the signal level is greater than one quarter of the maximum amplitude in the spectrum are given unit weight and all others are given zero weight resulting in a frequency shifted rectangular function of unit amplitude. A unit delta function is added to this since we want to pass undisturbed the DC signal produced by the A/D converter. This window function is applied to both X_{j} and Y_{j} spectra resulting in rectangular band shape for the frequency band that is in common. The effect of the window function on the timedomain is to convolve the signal with a sinc function. Since, the frequency spectra are already band limited, the window function does not in itself introduce any new waveform characteristics. Rather it prevents contamination of the waveform that would arise were one to scale up, by large factors, frequency channels that had little signal.
2.6 Forming circular polarization
The gains
and
,
the rotation parameters
and
,
and the window function
are applied to the spectral components
and
respectively. If
and
are the resulting vectors after calibration then
is related to
by the product of gain, window function and rotation matrix as follows
=  
(23) 
where , and , are real and imaginary components of and respectively. Similarly,
=  (24) 
The windowed, phase and gain calibrated and are added in quadrature ( ) to obtain LHC and RHC polarizations.
3 Performance limitations
In this section we will discuss the implications of possible contamination of phase caused by the cross polar component or leakage of unwanted orthogonal polarization component (Dterms) and by the temporal instability of receiver transfer characteristics and their effects on polarization purity with approximate quantitative results to estimate those errors. We will also discuss the requirement of frequent recalibration due to the variations in the transfer characteristics of the receiver by observing the drifts in the most sensitive parameters, which are the channel phases. Since the gain fluctuations are much smaller than phase fluctuations, we can ignore their effects.
3.1 Effect of phase error on polarization purity
Were one to introduce an imperfect
phase
shift into one channel when forming circular polarization from
perfectly orthogonal linearly polarized channels, the output voltages
and
would contain unwanted contributions from the opposite hand of polarization
(25)  
(26) 
where and are the fractional voltage leakage factors from unwanted polarizations (Dterms).
The larger the phase error, the greater the contribution from the
opposite hand. Consider the monochromatic case in which a linearly
polarized wave is incident normally on crossed linear dipoles with the
plane of the electric field oriented at
to the two dipoles. Then the voltages in the two dipoles are
V_{x}  (27)  
V_{y}  (28) 
After introducing an imperfect phase shift to the y channel, one has
V_{x }=_{ }  (29) 
V_{y' }=  (30) 
where is the error in the phase shift. Circular polarization is obtained by summing the x signal with the imperfectly phaseshifted y signal, giving
(31)  
(32) 
Had the phase shift been perfect, then giving and . Substituting Eqs. (31) and (32) into Eqs. (25) and (26) respectively, with and obtained by setting , we obtain the dependence of and , on the phase error, :
(33) 
This result is used in Sect. 3.3 for estimating the polarization purity.
3.2 Phase stability of the analogue receiver chain
The phase and amplitude transfer characteristics of the receiver chains for the orthogonal polarization channels are known to drift with time, due primarily to temperature changes of filters and cables used in the receiver chains. Fortunately, most of that change is common to both orthogonal polarization channels as the equipment for both channels is housed in close proximity to each other, and the relative changes are small compared to the total. The effect of drift in the relative transfer characteristic is a degradation of the polarization purity, since the equalizer weights that were determined prior to an observation would no longer perfectly correct the channel differences, by the amount of the relative drift since that determination was made. This translates into a requirement that the equalizer weights be redetermined periodically to ensure that polarization impurity due to drift remains below a predetermined level.
We have estimated how often such redetermination would need to be made by measuring the relative phase drift in some existing receiver chains at Effelsberg and the VLBA. The measurements were made using the VLBI phase calibration system (Thompson et al. 1991), which injects a pulse train in the front end and extracts them at the backend data acquisition rack or correlator, to monitor the phase length of the whole receiver system, from front end to the samplers.
The measurements show that indeed the phase changes in the orthogonallypolarized channels track each other well (Fig. 1 top and middle) and there are only occasional outliers, most likely related to phaselocked loop local oscillator used in the analogue baseband converters, which would not be present in a digital system. The drift in the relative phase is conveniently quantified using a structure function analysis, which converts the phase difference time series into the rms phase change as a function of timescale (Fig. 1 bottom). The result is that the rms phase difference due to drift is in the range to on timescales of 100 s to 9000 s.
3.3 Expected polarization purity
The polarization purity to be expected from polarization conversion performed at IF can be derived by combining the two results from Sects. 3.1 and 3.2  the sensitivity of polarization leakage to phase errors and the typical phase errors in existing analogue receiver chains ( to rms).
Figure 1: Top: calibration system phase vs. time for a single polarization channel for three different receivers at Effelsberg, showing phase changes of typically to over periods of hours. Middle: calibration system phase difference between orthogonal polarizations of the same receivers at Effelsberg, for the same experiments as in the top plot with an arbitrary offset. Bottom: structure functions constructed from the relative phases presented in the middle plot. These show the rms of the phase difference vs. timescale. 

Open with DEXTER 
The resulting D term is 0.006 (rms) for a rms phase error, for which one must recalibrate the equalizer every few minutes, rising to 0.025 (rms) for a rms phase error, which one would obtain were one to calibrate the equalizer once and leave it fixed for many hours. These are smaller than the leakage Dterms measured for existing radio telescopes, which are commonly 0.05 to 0.15.
However, the leakage in existing receivers is constant over long periods, since it occurs primarily due to tolerances in the manufacture of the analogue polarizers, and so can be calibrated using observations of astronomical polarization calibrators. That calibration reduces the effect of the leakage on the resulting polarization images by a factor of ten typically, and one typically sees residual polarization artifacts that are 0.005 to 0.015 rms times the peak flux density in the images. These values are comparable to those expected to be delivered from the digital polarization conversion without use of astronomical Dterm calibration.
However, the Dterm from IF polarization conversion, though small, is expected to drift with time between equalizer recalibrations due to drift in the relative phase of the orthogonal polarization receiver channels. Were one to want to improve on this by using astronomical calibration of the residual polarization leakage on timescales between the equalizer recalibration, then one must be able to derive the Dterm from a snapshot observation. Such an algorithm is available and requires the use of an unpolarized calibrator source. However, the changing Dterm requires that the postcorrelation analysis software be able to handle timevarying Dterms.
Figure 2: Logic block for conversion from linear to circular polarization. 

Open with DEXTER 
This will prevent the use of astronomical calibration of the residual polarization leakage, since the Dterm calibration in use assume that the leakage is constant on a 12 h timescale.
4 FPGA implementation details
Here we will describe the structural and functional details of the FPGA based circular polarizer (Fig. 2) for converting the two orthogonal input linear polarizations to the two circular polarization components in real time. The method of digital signal processing is based on the theory presented in Sect. 2. The sampled time domain signals, x and y are fed, at a rate of 1024 MS/s with 10 bits per sample representing positive integers, to the following logic elements.
1. Clock rate reduction logic. Since the sampled time series at 1024 MS/s is too fast for processing, serialtoparallel conversion of the streaming data has been implemented with a factoreight fanout, generating eight parallel sample streams clocked at 128 MHz. This block was developed in an earlier DBBC (digital baseband converter) project. The frontend of the system operates with 1024 MHz sampling rate and the IF band is 512 MHz. The samples taken at 1024 MHz in the AD device are transferred in progressive steps in parallel, to reduce the clock rate, still maintaining a formal sampling clock of 1024 MHz. So the data are demultiplexed to obtain from 8bit samples at 1024 MHz to 16bit at 512 MHz, 32bit at 256 MHz, 64bit at 128 MHz (8 samples at 128 MHz). This system is already commercially available (Tuccari 2004).
2. Serial frame generator. This is central block for enabling serial processing of real time data in parallel working at 128 MHz clock rate. To feed the eight parallel streaming FFT blocks with frames of 1024 real timedomain samples, intermediate logic takes in eight samples in parallel at every clock edge and outputs them to one of eight buffers, eight samples in parallel at each clock edge. Once a complete frame of 1024 samples have been loaded into the buffer, the next buffer is selected to receive the next 1024 samples. The first buffer is read out serially into its corresponding FFT block at one eighth of the rate (128 MHz) at which is was filled, starting at the next clock pulse after writing the first eight samples is complete, such that serial readout and parallel writing operations on the buffer occur simultaneously. When the register is ready to read out the last sample of the first data frame, the first eight samples of frame number nine is ready for writing and the two operations occur simultaneously at the same clock pulse. This continues and each register sends out data serially at a rate of 128 MHz. The length of the frame is equal to twice the number of spectral channels required for later operations.
3. FFT. The serial frame generator for the x polarization and for the y polarization each produce eight output lines to feed eight parallel FFT blocks that run continuously and independently of each other, each processing successive frames of data to keep up with the realtime sampling rate. The FFT is a complex transform, but the data are purely real, and so we used the relation for the FFT of two real functions simultaneously, to save a factor of two in device resources by feeding the x and y polarization data to the real and imaginary channels of each FFT engine input. The streaming pipelined FFT is generated conveniently using a Xilinx IP core. It is configured with 11 bits at the input and requires two's complement representation of the positive integers from the A/D converter. The real and imaginary terms of the X and Y spectra are retrieved from the FFT real and imaginary output channels by using a decoder at the output of each FFT. The method is equivalent to performing separate FFTs of the two real functions, but consumes half the space by performing them simultaneously in one transform. The FFT block outputs a 1024point complex spectrum every 10^{6} s in each polarization, keeping up with the realtime 1024 MS/s sample rate.
4. Power spectra accumulator. To calibrate the equalizer weights, we used totalpower spectrum accumulators when determining the amplitude scaling factors needed for bandpass calibration and we used crosscorrelation and accumulation to determine the phase difference between the X and Y polarizations in each frequency channel. To form the power spectra, each spectrum from each FFT (after decoding) is squared, forming and , and those are accumulated for nearly 8.4 s, which is determined by the goal of having thermal noise fluctuations that contribute at most rms phase error during equalization. The accumulator is dimensioned with enough bits to hold the accumulation result without overflow. To measure the (frequencydependent) phase difference between the X and Y polarizations, the crosspower spectra Z = XY^{*} are formed and the real and imaginary terms, and , are also accumulated for nearly 8.4 s. Each of the four quantities ( , , and ) are accumulated in eight pairs of accumulators, one following each FFT (after decoding), with each accumulator being paired to hold the noise diode on and off state results separately. After accumulation, the realtime processing of the calibration signal is stopped and calculations for the noise diode on and off states are performed sequentially on the accumulated results from all eight FFT engines. Differences are formed between the accumulators for the noise diode on and off states and the eight differences are summed to obtain the final integrated spectra, from which the equalizer gain and phase weights are calculated in the next block.
5. Equalization parameters. The phase and gain equalization parameters are calculated using Eqs. (12), (13) and (19), (20) respectively. The division and square root operations are implemented using the Xilinx floating point IP core. The floating point operations are performed with a clock frequency of 64 MHz due to the frequency limitations of the IP core. The resulting phase and gain calibration parameters are latched for realtime equalization of the X and Y data streams during subsequent observations.
Figure 3: Experimental setup in anechoic chamber to measure polarization purity. 

Open with DEXTER 
6. Window Function. A window function described in Sect. 2.5 is determined and latched along with the equalization parameters.
7. Synchronization. During observations the decoded FFT output samples are read out serially from the decoder and the corresponding gain, phase correction factors, and the window function are also read out serially from their respective latches and synchronization is required to ensure that the frequency channels of the spectrum and the equalization parameters are aligned. The synchronized values are passed to the next stage for phase and gain corrections.
8. Equalization. Phase and gain corrections are applied to each spectral channel of the outputs after synchronization in the equalizer block, which implements Eqs. (23) and (24).
9. Formation of circular polarization. After equalization X and phase shifted Y are added to form the LHC and RHC polarizations. The phase shifts of Y are implemented by exchanging the real and imaginary components of each spectral channel in the Y spectrum with a sign inversion of the real component for phase shift and of the imaginary component for phase shift. The number of bits of each spectral channel in each output power spectrum is 51.
The above logic blocks are tested separately since the design is too big to fit one FPGA chip. The design (Fig. 2) is also broken in parts to fit the commercially available DBBC boards and a practical implementation will be carried out in another future project by stacking DBBC boards each containing a part of the broken design. The simplest configuration requires fourteen such boards to keep the required precision and since DBBC has defined number of input/output pins. Since the DBBC already converts 1GS/s to 8 samples at 128 MHz and the logic blocks are tested to work at 128 MHz, we are certain that the developed circular polarizer will show expected performance in real time. The cost of implementation in the DBBC is around 50 000 euros and in custom designed boards is around 30 000 euros.
Figure 4: Power spectrum of the broadband noise measured at the input to the directional coupler by the spectrum analyzer shown in Fig. 3. The band was shaped using an antialiasing filter to ensure that the power level was low below 1000 MHz and above 1500 MHz to avoid aliasing of power from outside the third Nyquist zone during the later digital down conversion. 

Open with DEXTER 
Figure 5: Equalizer weights determined during calibration of the equalizer on the noise diode injected at to the two receiving dipoles. Top: phase difference between X and Y polarization channels. Middle: gains for X and Y polarization channels. Bottom: power spectrum of the broadband noise measured at the dipole outputs by the oscilloscope. The frequency range is labeled corresponding to the band after digital down conversion from the RF band of (1000 to 1500) MHz to baseband of (0 to 500) MHz. The gains are approximately the inverse of the square root of the noise spectrum in the bottom as expected. The gains and phases are set to zero by the window function where the power dropped 6 dB below the maximum, which happened below 150 MHz and above 450 MHz. This is as expected, given the shape of the broadband noise spectrum. 

Open with DEXTER 
5 Verification of polarization purity
We performed the following experiment to simulate the developed circular polarizer to quantify the polarization purity obtained by this digital technique. The experiment was performed in an anechoic chamber by coupling linearly polarized broadband noise into a circular waveguide and receiving with crossed linear dipoles at the end of the waveguide. The calibration coupler, which is responsible for polarization purity was tuned for equal amplitude and phase. The cross coupling was estimated to be better than 33.6 dB while doing the setup. The experimental setup is shown in Fig. 3.
The spectrum of the broadband noise that was coupled into the waveguide is shown in Fig. 4. The noise power spanned from 1160 MHz to 1462 MHz (6 dB down from peak), representing a fractional bandwidth of 23%. The power level at the edges of the Nyquist band, at 1000 MHz and 1500 MHz, were 37 dB and 30 dB below the peak in the band, and so very little power was aliased from outside the band. After receiving this signal with the crossed dipoles, the signal was undersampled at a sample rate of 1000 MS/s by a digital oscilloscope, causing digital down conversion to baseband. Since the sample rate was slightly less than 1024 MS/s used by the design, the frequency channel width in this experiment was 0.976 MHz instead of 1 MHz. The input voltage range of 250 mV to +250 mV was translated to 0 to 1024 effort was taken to match the complex gain or path length in the two channels from the receiving dipoles to the sampling oscilloscope. Thus, the RF band was translated to 160 MHz to 462 MHz, representing a fractional bandwidth of 97% that was presented to the polarization converter.
Figure 6: Power spectral densities for LHC and RHC for a single position angle. 

Open with DEXTER 
1. Measurement description. The outputs of the crossed linear receiving dipoles were sampled with a digital storage oscilloscope for 4 ms and saved to file for later processing by the design. A rotating waveguide joint allowed us to rotate the plane of linear polarization with respect to the crossed receiving dipoles. Any ellipticity of the resulting circular polarization would show up as a change in the power in the circular polarization as the linear polarization is rotated (see Sect. 6.2).
2. Equalizer calibration. For equalizer calibration, the linear polarization was aligned at to the two dipoles by connecting the dipole outputs to a twochannel power meter and rotated until equal power was measured in both channels. The resulting powers were +7.51 dBm and +7.50 dBm and drifted by dB during the measurement, corresponding to a polarization rotational position angle uncertainty of . For the noise diode off state, the data samples were simply set to zero rather than being measured, since the setup in Fig. 3 was well shielded from RFI, in which case measuring with the noise diode off gives almost same results as setting the offstate samples to zero (verified in a previous experiment). The signals were sampled by the digital oscilloscope with the noise diode on and processed by the design to obtain the equalizer amplitude and phase weights. These were loaded into the equalizer for calibration of subsequent measurements, and are shown in Fig. 5. The power spectrum in Fig. 5 (bottom) differs from the spectrum in Fig. 4 since the transfer characteristics of the sampled spectrum are modified by the frequency response of each component through which the signal is transferred namely the directional coupler, circular waveguide, dipoles and the sampling oscilloscope.
3. Polarization purity measurement. For polarization purity measurement, the plane of input linear polarization was rotated to five positions with respect to the receiving dipoles and the received signals from both dipoles were sampled in each position. To adjust the rotational position angle accurately, the plane of input linear polarization was rotated to either minimize the power in one of the receiving dipoles ( ) or to obtain equal power in both dipoles ( , ). The time elapsed between calibrating the equalizer and making all the measurements for determining the polarization purity was some 2 h. During this time, some drift in components might have occurred, but nevertheless good polarization purity was obtained.
4. Data processing to form circular polarization. We processed the data by running the design in software logic simulation, using ModelSim SE on suse 10.3 Linux machines having 16 GB of RAM and 2.7 GHz clock speed. Processing of 4 ms of data from each position required 4 days of elapsed time on a single computer so we processed for each position 20 sets of 1/20th of the data in parallel, which required one day.
6 Results
The resulting power spectra in LHC and RHC are shown in Fig. 6 for one of the five position angles of the input linear polarization on top of each other. These show that the gain equalization flattened the spectra and that the window function truncated the spectra where the filters roll off. The total power is found to change very little with rotation of the input linear polarization, signaling a high purity circular polarization. To quantify the purity, we measured the total power in LHC and RHC as a function of rotation angle, and show this in Fig. 7. This shows power level changes of around 1 part in 200 peaktopeak over all position angles as a function of rotation angle. We obtained an ellipticity of 0.9971 and 0.9976, axial ratio of 1.0029 and 1.0024 crosspolar response of 24.975 dB and 24.979 dB and Dterm of 0.05767 and .05764 for LHC and RHC polarizations respectively. Since the cross coupling of 33.6 dB caused by the directional coupler is better than the obtained crosspolar response, the limiting factor is most likely the repeatability of the connections between matings. The insertion loss is unlikely to be repeatable between matings to 34 dB that would cause the calibration signal to be the limiting factor. The uncertainty on the polarization purity measurement has contributions from thermal noise, mechanical tolerance in the rotating waveguide joint, and in the repeatability of RF connections that we had to disconnect and reconnect during the measurement for connecting alternately the power meter and the oscilloscope. These can be estimated from the small changes of the power in the circular polarization formed when the input linear polarization was at and at . Those two positions are symmetric and the resulting powers should be equal, regardless of the ellipticity of the transmitting or the receiving antenna. We found fractional changes of 0.001 in the LHC power and 0.0004 in the RHC power between these two position angles. These are comparable to the peak power variations seen as we rotated the input linear polarization, and so the polarization leakage measurement is limited by mechanical tolerances in the apparatus. The thermal noise contribution was minor  the fractional error due to thermal noise fluctuations in the total power measurement was only 0.0006.
Figure 7: Mean output powers in LHC and RHC as a function of the rotational position angle of the input linear polarization. 

Open with DEXTER 
Discussion
A small systematic offset is seen between LHC and RHC powers (Fig. 6). A small error in the magnitudes of the rotation matrix elements will cause the phase rotation of Y
to introduce an offset between the LHC and the RHC powers. This can
happen since truncation error can cause small differences in the
obtained phase difference and the actual phase difference. The
following equations show the dependence of the offset on the rotation
angle error
=  (34)  
=  
  (35)  
=  (36) 
Therefore,
(37) 
Similarly,
=  (38)  
=  (39)  
=  (40) 
Hence, the offset between the two power spectra is given by
(41) 
for a single channel.
Figure 8: Top: fractional deviations of LHC and RHC powers from their mean values across the band of interest. The spikes at 160 MHz in the two spectra (Fig. 6) are excluded. Bottom: same as top, but zoomed in to show the trends of the deviations in the two power spectra. 

Open with DEXTER 
The RHC power is almost flat across band but LHC showed a small change with frequency. To quantify this variation, we took equally spaced points from each power spectrum in the band of interest and plotted the deviation of those points from the mean of the respective power spectrum. Figure 8 shows the deviation of LHC and RHC power spectra from their mean values. The plots are derived from LHC and RHC power spectra obtained by applying the equalization parameters to the same data samples from which the equalization parameters are obtained to show the deviations caused by bit truncation. From around 300 MHz onwards (Fig. 8 bottom) the two polarizations deviate differently with increasing frequency. This difference can be explained in terms of truncation of numerical precision. We have truncated bits at stages before and after formation of LHC and RHC powers, determined by the input signal levels and by the aim to keep phase errors < . The truncation error will cause offsets in the two circular power spectra shown in Eqs. (37) and (40), to deviate unequally from their mean and hence, the deviation can be more in one than the other. We verified in simulations that the offsets and the distortions seen in LHC and RHC power spectra increases in proportion to the number of bits discarded. However, the effects are small and we nevertheless obtained good polarization purity.
The equalizer applies phase correction on 1 MHz channel spacing, so we must ensure that when the VLBI correlator later subdivides the spectrum into finer frequency channels that the phase equalization is smoothly interpolated and does not result in a sawtooth with 1 MHz spacing. We performed a numerical simulation that confirmed that when we applied equalizer weights at 1 MHz spacings, transformed back to time domain, then transformed to frequency domain with 0.5 MHz spacing (by adjoining two 1024 point frames of time domain data) then the intermediate points at n + 0.5 MHz, where n is an integer were seen also to have been phase corrected by the equalizer phase.
7 Conclusion
Though this technique was demonstrated using logic simulator in software, it has been implemented using Xilinx software generating firmware that can be loaded into FPGAs. An implementation on the digital baseband converter (Tuccari 2008) is in preparation, which will make this technique available for use at many radio observatories for VLBI, and with minor extension, for measuring Stokes parameters. A further extension offers the prospect of cancelling the remaining (already low) cross polarization response (Dterm). The idea is that one can use adaptive cancellation by forming a linear combination of the two polarizations and adapting the coefficients of the filter to minimize the output power. The result should be (almost) perfect polarization purity in circular polarization over large fractional bandwidths. This could enable the sensitive search for circular polarization in active galactic nuclei.
Later measurements at the telescope should confirm that excellent polarization purity is achieved in real applications, as it was in the anechoic chamber. One can also confirm the stability of the transfer characteristics and decide on a recalibration interval for operational use. One can also characterize the typical phase response of receivers and confirm that the choice of 1 MHz frequency spacing is well matched to the existing systems. The effect of RFI on the system can be explored, to give recommendations on tolerable RFI levels and required performance of mitigation strategies. The trend in nextgeneration receivers for radio astronomy is to move the samplers as close as possible to the front end, which will benefit this system of polarization conversion since the time variable path length changes due to analogue cables and filters and amplifiers will be much reduced, yielding even better polarization purity.
AcknowledgementsWe gratefully acknowledge Xilinx for their generous support, through their donation of FPGAs for this project and for providing prompt help with software issues as they arose. Koyel Das, as a member of the International Max Planck Research School for Astronomy and Astrophysics at the Universities of Bonn and Cologne, thankfully acknowledges IMPRS for funding her Ph.D. projects.
References
 Chen, M. H., & Tsandoulas, G. N. 1973, IEEE Transactions on Antennas and Propagation, 389 [NASA ADS] [CrossRef] (In the text)
 Cooley, J. W., & Tukey, J. W. 1965, Mathematics of Computation, 19, 297 [CrossRef] (In the text)
 RuizCruz, J. A., MontejoGarai, J. R., Rebollar, J. M., et al. 2006, IEEE MTTS International Microwave Symposium Digest, 1173 [CrossRef] (In the text)
 Thompson, A. R., & Bagri, D. S. 1991, ASP, 55 (In the text)
 Tuccari, G. 2004, ASP Conf. Ser., 306, 177 [NASA ADS] (In the text)
 Tuccari, G. 2008, Proc. 5th IVS General Assembly, St Petersberg (In the text)
Footnotes
 ...^{}
 We have not considered any frequency down conversion in Eqs. (6) and (7). A down conversion will cause a frequency shift in by an amount where is the mixing frequency. Then the samples of in Eqs. (6) and (7) or will represent the samples of . The equations will remain unchanged except for the initial phase, which will now be so that the initial phase is unaffected by down conversion.
All Figures
Figure 1: Top: calibration system phase vs. time for a single polarization channel for three different receivers at Effelsberg, showing phase changes of typically to over periods of hours. Middle: calibration system phase difference between orthogonal polarizations of the same receivers at Effelsberg, for the same experiments as in the top plot with an arbitrary offset. Bottom: structure functions constructed from the relative phases presented in the middle plot. These show the rms of the phase difference vs. timescale. 

Open with DEXTER  
In the text 
Figure 2: Logic block for conversion from linear to circular polarization. 

Open with DEXTER  
In the text 
Figure 3: Experimental setup in anechoic chamber to measure polarization purity. 

Open with DEXTER  
In the text 
Figure 4: Power spectrum of the broadband noise measured at the input to the directional coupler by the spectrum analyzer shown in Fig. 3. The band was shaped using an antialiasing filter to ensure that the power level was low below 1000 MHz and above 1500 MHz to avoid aliasing of power from outside the third Nyquist zone during the later digital down conversion. 

Open with DEXTER  
In the text 
Figure 5: Equalizer weights determined during calibration of the equalizer on the noise diode injected at to the two receiving dipoles. Top: phase difference between X and Y polarization channels. Middle: gains for X and Y polarization channels. Bottom: power spectrum of the broadband noise measured at the dipole outputs by the oscilloscope. The frequency range is labeled corresponding to the band after digital down conversion from the RF band of (1000 to 1500) MHz to baseband of (0 to 500) MHz. The gains are approximately the inverse of the square root of the noise spectrum in the bottom as expected. The gains and phases are set to zero by the window function where the power dropped 6 dB below the maximum, which happened below 150 MHz and above 450 MHz. This is as expected, given the shape of the broadband noise spectrum. 

Open with DEXTER  
In the text 
Figure 6: Power spectral densities for LHC and RHC for a single position angle. 

Open with DEXTER  
In the text 
Figure 7: Mean output powers in LHC and RHC as a function of the rotational position angle of the input linear polarization. 

Open with DEXTER  
In the text 
Figure 8: Top: fractional deviations of LHC and RHC powers from their mean values across the band of interest. The spikes at 160 MHz in the two spectra (Fig. 6) are excluded. Bottom: same as top, but zoomed in to show the trends of the deviations in the two power spectra. 

Open with DEXTER  
In the text 
Copyright ESO 2010