Searching for Pulsars

with PRESTO

By Scott Ransom
NRAO / UVa
 Getting PRESTO
● Homepage:
http://www.cv.nrao.edu/~sransom/presto/

● PRESTO is freely available from github

https://github.com/scottransom/presto

● You are highly encouraged to fork your own

copy, study / modify the code, and make bug-
fixes, improvements, etc....
 For this tutorial...
● You will need a fully working version of PRESTO (including
the python extensions)
● If you have questions about a command, just try it out!
Typing the command name alone usually gives usage info.
● You need at least 1GB of free disk space
● Linux users: if you have more than that amount of RAM, I
encourage you to do everything in a subdirectory under
/dev/shm
● Commands will be > typewriter script
● The sample dataset that I'll use is here (25MB)
http://www.cv.nrao.edu/~sransom/GBT_Lband_PSR.fil
 Outline of a PRESTO Search
1) Examine data format (readfile)
2) Search for RFI (rfifind)
3) Make a topocentric, DM=0 time series (prepdata and exploredat)
4) FFT the time series (realfft)
5) Identify “birdies” to zap in searches (explorefft and accelsearch)
6) Make zaplist (makezaplist.py)
7) Make De-dispersion plan (DDplan.py)
8) De-disperse (prepsubband)
9) Search the data for periodic signals (accelsearch)
10) Search the data for single pulses (single_pulse_search.py)
11) Sift through the candidates (ACCEL_sift.py)
12) Fold the best candidates (prepfold)
13) Start timing the new pulsar (prepfold and get_TOAs.py)
 Examine the raw data
> readfile GBT_Lband_PSR.fil
● readfile can
automatically identify
most of the datatypes
that PRESTO can
handle (in PRESTO
v2, though, this is
only SIGPROC
filterbank and
PSRFITs)
● It prints the meta-data
about the observation
 Search for prominent RFI: 1
> rfifind -time 2.0 -o Lband GBT_Lband_PSR.fil
● rfifind identifies strong
narrow-band and/or short
duration broadband RFI
● Creates a “mask” (basename
determined by “-o”) where RFI is
replaced by median values
● PRESTO programs automatically
clip strong, transient, DM=0
signals (turn off using -noclip)
Usually a good thing!
● Typical integration times (-time)
should be a few seconds
● Modify the resulting mask using
“-nocompute -mask ...” and
the other rfifind options
Search for prominent RFI: 2
● Check the number of bad
intervals. Usually should be
less than ~20%
● Most significant and most
numbers birdies are listed (to
see all, use -rfixwin)
● Makes a bunch of output files
including “...rfifind.ps” where
colors are bad (red is periodic
RFI, blue/green are time-
domain statistical issues)
● Re-run with “-time 1” or re-
compute with “-nocompute”
in this case
 Search for prominent RFI: 3
This is not so great... too much color, and randomly arranged!
Usually we see bad channels or bad time intervals.
Random red color probably means we are masking a bit too much data.
Search for prominent RFI: 4

This is after using “-time 1” and it looks slightly better.

 Look for persistent low-level RFI
> prepdata -nobary -o Lband_topo_DM0.00 \
-dm 0.0 -mask Lband_rfifind.mask \
-numout 530000 GBT_Lband_PSR.fil

● prepdata de-disperses a single

time-series. The “-nobary” flag
tells PRESTO not to barycenter
the time series.
● If you need to de-disperse
multiple time-series, use
prepsubband
● Since we will search these data
(and FFT them), make sure that
the resulting time-series has a
“good” (i.e. easily factorable)
number of points (-numout)
Explore and FFT the time-series
> exploredat Lband_topo_DM0.00.dat
> realfft Lband_topo_DM0.00.dat
> explorefft Lband_topo_DM0.00.fft
● exploredat and
explorefft allow you
to interactively view a
time-series or its power
spectrum (for finding RFI)
● changing the power
normalization (key ‘n’) in
explorefft is often
very helpful
● realfft requires that
the time-series is easily
factorable (and at least
has 1 factor of '2').
Check using “factor”.
 Find the periodic interference
> accelsearch -numharm 4 -zmax 0 \
Lband_topo_DM0.00.dat

● We “trick” accelsearch into finding periodic interference (it found 4

candidates, with several harmonics in each)
● That information will be used to create a “birds” file
● “.inf” file is human readable ASCII (it is also found in the ACCEL file).
 Make a “birds” file
● What the heck is a “birds” file?
● “birds” are pulsar astronomer jargon for periodic interference that
shows up in our power spectra. We usually “zap” them by zeroing
them out before we search the power spectrum.
● In PRESTO, a .birds file is a simple ASCII text file with 5 columns
● The fundamental frequency of the periodic interference in Hz
● The width of the interference in Hz (power lines RFI at 50 or 60 Hz is
often quite wide, but some interference is only a single FFT bin wide)
● The number of harmonics of the fundamental to zap, and then 0/1
(no/yes) for whether the width of the harmonics should grow with
harmonic number and whether the freqs are barycentric or not (e.g. the
ATNF database freq for a strong pulsar in the data is barycentric)
● A row starting with a “#” is a comment
● Here is an example .birds file:
 Make a “birds” file
● Use explorefft and the *ACCEL_0 files to identify the main periodic
signals. Since these are DM=0, they are almost certainly RFI.
● Edit the .birds file with a text editor
● Given the results of our earlier accelsearch run, here is an example
(where I examined the signals with explorefft to check their widths):

● Notes:
● Don’t stress out too much over getting a perfect .birds file (especially
about high frequency not-too-strong signals – they will be smeared out
at high DMs). You mainly want to get the really strong stuff, with
Fourier powers more than 50 or so.
● Usually I make a .birds file only for a certain type of data (like once for
a whole project where the data are all the same) or for really important
single pointings.
 Convert the “birds” file to a zaplist
● Make an associated “.inf” file for the “.birds” file
> cp Lband_rfifind.inf Lband.inf
● Now convert all of the “birds” and harmonics into individual freqs/widths
> makezaplist.py Lband.birds
● The resulting “Lband.zaplist” is ASCII and can be edited by hand
● It can also be loaded into explorefft so you can see if you are zapping
everything you need (see the explorefft help screen)
● Apply the zaplist using “zapbirds”:
> zapbirds -zap -zapfile Lband.zaplist \
Lband_topo_DM0.00.fft
● Zapping barycentric time-series requires “-baryv” to convert topocentric
RFI freqs to barycentric. Get that by running prepdata or prepfold on
raw data (you can ctrl-c to stop them). As an example:
> prepdata -o tmp GBT_Lband_PSR.fil | grep Average
Average topocentric velocity (c) = -5.697334e-05
 Determining a De-Dispersion Plan
> DDplan.py -d 500.0 -n 96 -b 96 -t 0.000072 \
-f 1400.0 -s 32 -r 0.5

“-r” reduces the effective time

resolution to speed up search

● DDplan.py determines near-optimal ways to de-disperse your data

to maintain sensitivity to fast pulsars yet save CPU and I/O time
● Assumes using prepsubband to do multiple-passes through the
data using “subband” de-dispersion
● Specify command line information from readfile
Determining a De-Dispersion Plan
 Subband De-Dispersion 1
● Incoherent de-dispersion
requires you to shift the arrival
times of each input channel for
a particular DM
● This can be made much
quicker by partially shifting
groups of channels (subbands)
to some nominal DM
● The resulting subband dataset
can then be de-dispersed
around neighboring DMs with
many fewer calculations
● In PRESTO, we do this
subband de-dispersion with
prepsubband and
mpiprepsubband

From Magro and Zarb Adami, MNRAS in press

 Subband De-Dispersion 2
> prepsubband -nsub 32 -lodm 0.0 -dmstep 2.0 -numdms
24 -numout 132500 -downsamp 4 -mask
Lband_rfifind.mask -o Lband GBT_Lband_PSR.fil
● That command comes from the first call of the first plan line:

● Run prepsubband as many times as there are “calls” in the plan

● Accepted file formats to run prepsubband on are SIGPROC
filterbank (“.fil”) and PSRFITS (“.sf” or “.fits”)
● If you have a parallel computer (and very long observations), you
can use the fully parallel program mpiprepsubband to have one
machine read the data, broadcast it to many other CPUs and then
each CPU effectively makes a “call”
● The dedisp.py script in $PRESTO/python can help you automate
this process (and generates subbands as well, which can be used to
fold candidates and see the DM-curve much faster than by folding
raw data). When the file has been edited, do: python dedisp.py
 Prepare for Searching the Data
● First we'll clean up this directory but putting the subband files in their
own directory and getting rid of the temporary topocentric files
> mkdir subbands
> mv *.sub* subbands/
> rm -f Lband*topo*
● Use xargs (awesome Unix command) to fft and zap the *.dat files
> ls *.dat | xargs -n 1 realfft
> ls *.fft | xargs -n 1 zapbirds -zap \
-zapfile Lband.zaplist -baryv -5.69733e-05
● Remember that we can get the barycentric value (i.e. average
topocentric velocity) by running a fake prepdata or prepfold
command on the raw data:
> prepdata -o tmp GBT_Lband_PSR.fil | grep Average
● Now we are ready to run accelsearch on the *.fft files
● If your time series are short (like these), you can use accelsearch
to do its own FFTing and zapping by calling it on the “.dat” file. See
the -zaplist and -baryv options for accelsearch.
 Searching for Periodic Signals
> accelsearch -zmax 0 Lband_DM0.00.fft
● Accelsearch conducts Fourier-domain acceleration (or not, if
zmax=0) searches for periodic signals using Fourier interpolation and
harmonic summing of 1, 2, 4, 8 and/or 16.
● “zmax” is the max number of Fourier bins the highest harmonic for a
particular search (i.e. fundamental or 1st harm. for a 1 harm. search, 8th
harm. for a 8 harm. search) can linearly drift in the power spectrum (i.e.
due to orbital motion). Sub-harmonics drift proportionally less (i.e. if 2nd
harmonic drifts 10 bins, the fundamental will drift 5).
● The time that the searches take doubles for each additional level of
harmonic summing, and is linearly proportional to zmax.
● For MSPs, 8 harmonics is almost always enough. And zmax < 200 or
so (beyond that non-linear acceleration start to creep in).
● You can use xargs: ls *.fft | xargs -n 1 accelsearch …
● For this tutorial data, which is short, you might want to use “-flo 8”
so that the rednoise at the very lowest freq bins aren’t detected
 Sifting the periodic candidates
> python ACCEL_sift.py > cands.txt
● ACCEL_sift.py is in $PRESTO/python and can be edited and
tweaked on an observation specific basis
● It uses several heuristics to reject bad candidates that are unlikely to
be pulsars. And it combines multiple detections of the same candidate
signals over various DMs (and harmonics as well).
● The output is a human-readable ranked list of the best candidates
● ASCII “plots” in the cands.txt file allow you to see rough signal-to-noise
versus DM (if there is a peak at DM != 0, that is good)
● The format for the “candidate” is the candfile:candnum (as you would
use them with prepfold)
● You can also look through the ACCEL files themselves. The ones
ending in numbers are human readable (use less -S). Summaries of
the candidates are at top and details of their harmonics at bottom.
● For large single ACCEL files, you can use quick_prune_cands.py
 Folding Pulsar Candidates
> prepfold -accelcand 2 -accelfile \
Lband_DM62.00_ACCEL_0.cand Lband_DM62.00.dat
● prepfold can fold time-series (*.dat files), subbands (*.sub?? files),
or rawdata files. Many ways to specify period (-p) / freq (-f) etc.
● Folding time-series is very fast and is useful to decide which
candidates to fold the raw data
● When you fold subbands and/or the raw data, make sure that you
specify the DM (and choose the set of subbands with closest DM).
● For modern raw data, using 64 or more subbands (-nsub) is a good
idea for folding (to see narrow band RFI and scintillation better)
● If RFI is bad, can zap it using show_pfd or re-fold using -mask
> prepfold -dm 62.0 -accelcand 2 -accelfile \
Lband_DM62.00_ACCEL_0.cand \
subbands/Lband_DM72.00.sub??
> prepfold -n 64 -nsub 96 -p 0.004621638 -dm 62.0 \
GBT_Lband_PSR.fil
 Pulsar! (timeseries)
> prepfold -accelcand 2 -accelfile \
Lband_DM62.00_ACCEL_0.cand Lband_DM62.00.dat
 Pulsar! (raw data)
> prepfold -n 64 -nsub 96 -p 0.004621638 -dm 62.0 \
GBT_Lband_PSR.fil
 Searching for Transient Bursts
> single_pulse_search.py *.dat
● single_pulse_search.py conducts matched-filtering single-
pulse searches using “boxcar” templates.
● --fast can make things about a factor of 2 faster, but only use it if
the data are well-behaved (relatively constant power levels)
● If there are very strong pulses in your data, they can look like RFI.
For those cases, turn off bad-block finding (--nobadblocks)
● Generates *.singlepulse files that are ASCII and a single-pulse plot
● Can regenerate a plot using (for instance)
> single_pulse_search.py *DM1??.??*.singlepulse
● Can choose start and end times as well (--start and --end)
Searching for Transient Bursts
Making TOAs from the discovery obs

● get_TOAs.py needs to be run on a prepfold file of either a

topocentric time series or a fold of raw data. The fold must have
been made either using a parfile (use -timing) or with the (-
nosearch) option.
● The must be either a single gaussian (-g FWHM), an ASCII profile
(i.e. a bestprof file from prepfold) or a multi-gaussian-template
(derived using pygaussfit.py: “-g template.gaussian”)
● -n is the number of TOAs (and must factor the number of parts (-
npart) from the prepfold file
● -s is the number of subband TOAs to generate (1 is default)
> get_TOAs.py -g 0.1 -n 20 newpulsar.pfd
 Now try it from scratch...

● There is another sample data set (with mystery pulsar) here:

http://www.cv.nrao.edu/~sransom/Parkes_70cm_PSR.fits

● Command history for this tutorial can be found here:

http://www.cv.nrao.edu/~sransom/GBT_Lband_PSR_cmd_history.txt

● Let me know if you have any problems or suggestions!

Scott Ransom <[email protected]>