Vela pulsar observations with the Ooty Radio Telescope

Kishalay De
Indian Institute of Science

1 Introduction
We have been given a sample of the voltage signal obtained from an observation of the
Vela pulsar (PSR B0833-45) with the Ooty Radio Telescope. The Ooty Radio Telescope
consists of a cylindrical paraboloid reflecting surface which is 530 m long and 30 m wide,
placed on a slope of 11.2 degrees in the north-south direction in Muthorai village near
Ooty. The signal detectors consist of an array of 1056 half-wave dipoles which are oper-
ated to produce phased array beams in the sky.

Over the course of this analysis, we will try to investigate some of the statistical,
time-domain and frequency-domain properties of the signal, and extract some information
about the observed pulsar. The observation has been conducted at 326.5 MHz with a
bandwidth of 16.5 MHz, and the voltage signals from the northern and southern half of
the telescope is available to us. Note that we will not use any specific set of units to
quantify the voltage or the intensity signals, since no calibration data has been provided.

2 Signal statistics
2.1 Voltage signal characteristics
We would first like to investigate the statistical properties of the voltage signal received
at the telescope to verify some of its expected properties. We expect the voltage signal
to have a Gaussian distribution with some mean and standard deviation. Let us have
a look at the histograms of the voltage received at the northern and southern feeds of
the telescope. To ensure that the presence of the pulsar signal within this time series
(where one would expect increased levels of power) does not affect the general statistical
conclusions, we start with a set of 100,000 voltage samples which have been randomly
and uniformly selected from the data set (a thousandth fraction of the entire sample).
While it is evident that the distributions have a characteristic Gaussian shape, we
will verify this further. We estimate the mean and variance of the two voltage samples
using the unbiased estimators for the parameters. For the northern half voltage, these
parameters are:
µN orth = 3.445 σN orth = 28.142 (1)
The same set of parameters for the voltage obtained from the southern half is:

µSouth = 0.610 σSouth = 29.838 (2)

With these parameter estimates, we scale the entire data set to a normalized distribution
using,
Xold − µ
Xnew = (3)
σ
where Xnew is the scaled voltage, Xold is the original randomly sampled voltage, µ is
the distribution mean and σ is the standard deviation for the northern and southern

1
(a) Histogram of the voltage at the northern half (b) Histogram of the voltage at the southern half

Figure 1: Histograms for the voltage distributions for a total of 100,000 randomly selected
samples

half voltages respectively. We now test the hypothesis that these distributions follow a
standard normal distribution with the Student’s t-test, which gives a p-value of ∼ 10−14
less than 1 in favor of the null hypothesis, for both the northern and southern feed
voltages. Clearly, it can be concluded with a high level of confidence that these voltages
follow a normal distribution.
We can now proceed to estimate the statistical parameters and their confidence in-
tervals for the voltage signals. We then get,

µN orth = 3.445 ± 0.174 σN orth = 28.142 ± 0.124

µSouth = 0.610 ± 0.185 σSouth = 29.838 ± 0.131

where the intervals correspond to 95% confidence bounds. Note that the voltage values
do not have a zero mean as one would have expected. This may result from small bias
voltages in the telescope back-end, which would add an offset to all voltages.

2.2 Power signal characteristics

We now investigate the properties of the intensity signal, i.e., the square of the voltage
signal. We expect the signal to have an exponential distribution with equal mean and
standard deviation. The histogram of the power signals of the northern and southern
feeds are given in Figure 2 (for the same set of samples). The distributions are well fit
with an exponential curve. We now estimate the mean and standard deviation of this
distribution. Note that X
E[X] = p i Xi (4)
where pi is the probability (obtained from the histogram by dividing the number of counts
in a bin by the total number of counts) and Xi is the value of the variable. Again,

V ar[X] = E[X 2 ] − E[X]2 (5)

4
p relations, we get E[X] = 832.43 and V ar[X] = 122.71 × 10 , which gives
Using these
σ[X] = V ar[X] = 1107.75. Though σ[X] and E[X] have values which are fairly close

2
(a) Intensity signal distribution for the northern (b) Intensity signal distribution for the southern
half. half.

Figure 2: Distributions of the intensity signal for the two halves of the telescope, along
with best-fit exponential distributions.

as expected, they are not equal. This may be an effect of the quantization of a continuous
distribution by binning into a histogram and from the earlier observed offset from zero
mean in the voltage distributions, which would create deviations from a pure exponential
distribution.

3 Time-Frequency domain properties

3.1 The voltage power spectrum
We would now like to analyze the signal in the frequency domain. It has been told that
the observations have been carried out with a bandwidth of 16.5 MHz. Consequently, it
would be interesting to see the power spectrum of the voltage signal to identify any signs
of aliasing in the frequency domain. Using the Fast Fourier Transform (FFT) with 256
frequency channels (a sufficient frequency resolution of ∼ 125 kHz; power of 2 increases
the speed of the FFT), we find the average power spectrum of the voltage signal by
averaging the power spectrum obtained from all 512-point (time samples) FFTs. The
spectra for the northern and southern half are given below.
A number of interesting points can be observed from these spectra. Firstly, note
that the 0 frequency channels has a large signal (much larger than its nearby channels),
indicating that there is a DC offset in the voltage time series. This reinforces the claim
made at the end of Section 2.1 to explain the non-zero mean in the voltage. Secondly,
it is easy to see that the DC channel power is much larger in the northern signal, than
the southern signal which is consistent with the fact that the norther signal voltage mean
is much larger than the southern signal mean, which is closer to zero. Finally, we can
clearly see that the power spectrum smoothly tapers off to zero at both the edges of the
band, indicating that aliasing is very minimal for these receivers. The sharp peaks in the
spectra may be signatures of local Radio Frequency Interference (RFI).

3
 (a) Power spectrum from the northern half (b) Power spectrum from the southern half

Figure 3: Average power spectrum from the two feeds in the telescope. The frequency
axis goes from 0 to 16.5 MHz, corresponding to the bandwidth of the receiver.

3.2 Dynamic spectrum

We now try to detect the first signs of the pulsar signal by looking at the dynamic
spectrum of the signal, i.e., the power in the signal as a function of time and frequency.
As a trial, we examine ∼350 ms of the data to detect a pulsed signal, if present. The
dynamic spectrum is shown below, where the y-axis is frequency in MHz, x-axis is time
in ms and the color code indicates the signal power for a particular frequency and time
sample. To improve the Signal to Noise Ratio (SNR) of the signal, the power from the
two halves of the array have been added (Incoherent addition). Note that the maximum
time delay between the northern and the southern half of the array is of the order of 500
m / c ∼ 1 µs, which is much smaller than the time resolution we have used and hence
delay correction will not affect our analysis.
Clearly, there is an interesting source visible in the plot. A pulsed signal appears
earliest at the high frequencies, and gradually appears later at lower frequencies. Four
such pulses are visible in the dynamic spectrum. This frequency dependent delay is a
characteristic sign of a signal dispersed in the interstellar medium. The observed delays
can be used to estimate the Dispersion Measure (DM) along the line of sight to the pulsar.

4 Discovering the pulsar

4.1 Pulse time of arrival and the Dispersion Measure
Given the limited amount of data, estimating pulse time arrivals as a function of frequency
is a fairly difficult task because of a couple of reasons:

• Single pulses in individual frequency channels have very low SNR to allow a useful
calculation. Hence, we need to increase the bandwidth of individual channels to get
better SNR.

• We cannot increase the channel-width as much as we would like to, since remaining
dispersion within the channels would spoil the signal.

4
Figure 4: The dynamic spectrum of the signal for the first ∼ 350 ms. Frequency is plotted
on the y-axis, and increases upwards from the lower edge of the band to the upper edge.
Frequency resolution is 125 kHz. Time is plotted on the x-axis for a duration of 200 ms,
and a time resolution of ∼ 490 µs. The power from the northern and southern half of
the telescope have been added to increase the SNR.

Hence, we now use 64 channels (corresponding to channel width of 500 kHz) across the
band to increase the SNR of single pulses. Given that we do not know the period of
the pulsar, we cannot fold the profile to obtain better SNR either. Hence, we use one
strong pulse in the data set to estimate the time at which it arrives at three different
frequencies, and subsequently obtain the Dispersion Measure along the given direction.
Note that we are limited to using the frequency channels which have the strongest signals
(verified visually), ans other frequency channels have significantly poorer SNR for single
pulses.

To obtain the time of arrival of the pulse in different frequency channels, we fit the
single pulse to a Gaussian curve, and use the mean of the Gaussian to estimate the arrival
time with respect to the start of the voltage stream. The fit Gaussians and the relevant
parameters are given in Figure 4. It should be noted that the pulse arrival times are
essentially arrival times for the pulse across a 500 kHz bandwidth, and by approximating
this channel by its center frequency, we may overestimate the DM. Furthermore, we are
assuming that the pulse shapes are well-fit with a Gaussian curve. While this is true for
a lot of pulsars, it may not be completely accurate in this case.

Now, the pulse arrival time at a frequency ν (in MHz) depends on the DM (in pc/cc)
of the pulsar as (see NRAO website in References),
t ≈ t∞ + 4.149 × 103 × DM × ν −2 (6)
where t is the pulse arrival time in seconds and t∞ is the pulse arrival time in seconds
at infinite frequency. Clearly, a plot of t vs. ν −2 would provide us with the information

5
(a) Single pulse and best fit Gaussian at 327.918 (b) Single pulse and best fit Gaussian at 327.209
MHz across 500 kHz channel width. Pulse ar- MHz across 500 kHz channel width. Pulse ar-
rival time estimated to be 279.3 ± 0.8 ms. rival time estimated to be 292.6 ± 1.1 ms.

(c) Single pulse and best fit Gaussian at 326.500 (d) Plot of arrival time (s) with respect to start
MHz across 500 kHz channel width. Pulse ar- of observation vs. ν − 2 (in MHz− 2). Measured
rival time estimated to be 300.4 ± 1.0 ms. slope is (2.636 ± 0.635) × 105 s MHz−2 .

Figure 5: Estimation of pulse arrival times across the band and the DM along the line of
sight to the pulsar. Time resolution used is ∼ 490 µs.

required to calculate the DM. From the information given in, we plot time of arrival as
a function of ν −2 and fit it with a straight line, as shown in Figure 4. The slope of the
line is (2.636 ± 0.635) × 105 s MHz−2 where the uncertainty is at the 68 % confidence
interval. Hence, the DM estimate for the pulsar is 63.5 ± 15.3 pc/cc based on Equation 6.

Assuming the mean electron density in the galactic medium to be ne = 0.03 cc−1 , we
estimate a distance of 2.11 ± 0.51 kpc to the pulsar.

4.2 Dedispersion
We can now dedisperse this signal by correcting for the frequency dependent delays using
our knowledge of the DM. Going back to our original 256 channel resolution (to minimize
intra-channel smearing), and following Equation 6, we correct for the frequency dependent
term by applying the required time delays to the frequency channels. From Equation 6,

6
it immediately follows that the delay between two frequency channels at ν1 and ν2 MHz
will be,
τ (s) = 4.149 × 103 × DM × (ν1−2 − ν2−2 ) (7)
We adjust the time-domain position of all lower frequency channels to align them with
the pulse arrival time at the highest channel and then add the intensity in all the channels
to obtain the dedispersed signal intensity. The result of the procedure is shown in Figure
6.

Figure 6: The dedispersed time series of the voltage stream provided with 256 frequency
channels. The DM used is 63.5 pc/cc (as obtained earlier) and time resolution used is ∼
490 µs.

It is evident that the single pulses now have significantly higher SNR, and are clearly
visible above the background noise. Note that the length of the time series has been
shortened because time delays could not be applied to channels where the pulse did not
arrive at the lower edge of the band before the end of the observation.

4.3 The pulsar period

With a number of high SNR single pulses available, we are now in a position to estimate
the periodicity of the signal, i.e., the rotation period of the pulsar. We require the arrival
times of the individual pulses to fit a period solution, and hence, we go back to an earlier
technique of fitting Gaussian curves to individual pulses, to estimate the arrival times.
An example of a fit pulse is shown in Figure 7 (a). The obtained arrival times for each
of the pulses with the respect to the start of the observation (starting from the second
visible pulse in Figure 6), along with their uncertainties are given in Table 1. A linear fit
to the arrival times is shown in Figure 7 (b).
The best-fit linear parameters for the given fit is P = 89.46 ± 0.14 ms, where P is the
pulsar period. Note that a second order polynomial fit yields a Ṗ which is consistent with
0, and hence Ṗ cannot be estimated with such a short length of data. Finally, based on
the period estimated, we can fold the entire time series with the pulsar period to obtain
an average profile for the pulsar. This is shown in Figure 8. Note that the average profile
indicates that our initial assumption for a Gaussian pulse shape is fairly accurate.

7
(a) A dedispersed single pulse with a Gaussian (b) A linear fit to the pulse arrival times to es-
fit. timate the rotation period.

Figure 7: Estimating the pulsar period by fitting single pulses with Gaussian curves,
followed by fitting the estimated arrival times with a linear curve.

Pulse Number Arrival time (ms) Uncertainty (ms)

1 88.09 0.50
2 177.50 0.45
3 266.80 0.55
4 356.80 0.66
5 445.50 0.55
6 535.10 0.50
7 625.20 0.60

Table 1: Pulse arrival times for the given data set

Figure 8: The average profile of the pulsar, folded with the estimated pulsar period.

8
5 Conclusions
After this long effort, we have managed to investigate a number of interesting characteris-
tics of the signal. Starting from a statistical examination of the properties of the voltage
signal to estimating the DM of the pulsar, up to estimating the period of the pulsar, we
have recovered a number of useful parameters with this rather small data set, albeit with
a fair amount of uncertainty. Note that all quantitative estimates, including that of the
pulsar period may be corrupted by the large uncertainties in other parameters, especially
the DM. Indeed, longer observations would help to significantly reduce these error bars,
and possibly better knowledge of the observing time and configuration would allow one
to phase the two halves of the array to increase sensitivity.
The visibilities of the source from the two arrays have not been calculated, primarily
because interferometry cannot be done using the limited information available. More
specifically, the observing time, and the exact location of the phase centers for these two
halves would be required to calculate the projected baselines with respect to the source.
However, we can expect that the visibilities would remain constant as a function of time,
because the source is a compact object, i.e., the equivalent of a delta function in the sky.
Hence, the Fourier transform of this brightness distribution would be a constant across
all baselines, and therefore, will not change even as the projected baseline changes with
time.

6 References
• Pulsars. Website of the National Radio Astronomy Observatory. Available at
www.cv.nrao.edu/course/astr534/Pulsars.html. Accessed 15 Nov, 2015.

• J. N. Chengalur, Y. Gupta & K. S. Dwarkanath, Low Frequency Radio Astronomy,

2003.

• D. R. Lorimer & M. Kramer, Handbook of Pulsar Astronomy, Vol.4, Cambridge

University Press (2004).