Contraste con la vegetación real y potencial

The data were analysed using the Max-Planck-Gesellschaft Supercomputer Hydra2_and

a pulsar searching pipeline based on the presto3 _{software (Ransom,2001). The accel-}

eration search step, which dominates the CPU time, is performed on the GPU-enabled version of the programme accelsearch4_{. The data processing was as follows.}

First, the observations at 4.85 and 8.35 GHz are downsampled in time by a factor 4 and 2, to 262.144 and 131.072µs, respectively, because of the pulse broadening expected due to the scattering toward the GC (Spitler et al.,2014). The downsampled data keeps the maximum achivable time resolution (dominated by the expected scattering), while the reduced data size speeds up the processing of these bands considerably (roughly by a factor of 4 and 2, respectively).

2_{http://www.mpcdf.mpg.de/services/computing/hydra}

3_{https://github.com/scottransom/presto}

Next, a full-length observation (usually of∼1.2or∼2.4 hr) is recursively split into segments of half the observing time down to a minimum segment length of ∼4.5 min. These segmentations are necessary to be sensitive to pulsars with short orbital periods because the algorithms utilized by presto can best recover periodic signals with an acceleration when this acceleration is constant over the length of the observation. A good rule of thumb, applicable only for circular orbits and less valid if the orbit is eccentric or the acceleration of the pulsar is very high5_{, is to assume that the accel-}

eration of a pulsar in a binary orbit will remain constant within portions of ∼10 per

cent of the orbit (seeJohnston & Kulkarni,1991;Ransom et al.,2003;Ng et al.,2015). Following this rule, we are most sensitive to pulsars with orbital periods longer than

∼0.75, 1.5, 3, 6, 12, or 24 hr, depending on the data segment searched (see Table 3.3).

The increase in sensitivity to shorter orbital periods by reducing the observation length comes at the cost of an increase in the minimum ux densities that can be detected, as the observing length in the shorter segments is smaller (assuming radiometer noise, Smin ∝ Tobs−1/2). This partially-coherent segmented acceleration search scheme is a good way to cover a large range of dierent accelerated systems and orbital periods. It was rst applied to the Parkes Multibeam Pulsar Survey (Eatough et al., 2013b), and later to the the Galactic Plane region of the High-Time-Resolution-Universe South survey (Ng et al.,2015), both with time domain acceleration searches.

For each of the segments, a mask in time and frequency across the lterbank of data potentially aected by Radio Frequency Interference (RFI) is created with the programme rfifind. Additionally, a list of periodicities to be excluded from the anal- ysis is prepared, including the mains frequency fundamental at 50 Hz with some strong harmonics and the magnetar SGR J1745−2900 spin frequency with its rst 32 integer

harmonics. The data are then barycentred and dedispersed to time series assuming dierent dispersive delays starting at a trial dispersion measure (DM) of 800cm−3pc, with the nal trial DM value, and the number and size of the steps dependent on the frequency of observation. The dedispersion plans are computed using the DDplan.py tool. At this stage, a search for single pulses above a threshold of 6σ is performed on the longest available dedispersed time series using single_pulse_search.py.

All the dedispersed time series are then Fourier transformed using the Fast Fourier Transform (FFT) algorithm, corrected for an excess of red noise and searched for periodic signals. To be sensitive to accelerated signals, presto uses an estimation of the spread of the power of a periodic signal in the Fourier domain due to the acceleration produced by an orbital motion. The total power (or most of it) is recovered by applying the inverse eect using a matched lter (Ransom et al.,2002). The maximum of the z−value, or zmax parameter, of the routine accelsearch, which controls the maximum

number of Fourier bins where to apply the matched lters, is dened asz=a Tobs2/P c, where a is the maximum acceleration of the pulsar on the line-of-sight, Tobs is the observation length,P is the shortest period to be detected (i.e. the period of the highest harmonic), and c is the speed of light. This parameter is set to 1200 in our pipeline,

5_{A very high acceleration translates into a high jerk in some orbital phases that can break down the} assumption of the constant acceleration of the 10 per cent rule (see Section3.2.2.2for more details).

Table 3.2: Summary of the data properties and searching parameters at each frequency band. νc is the central observing frequency, ∆ν the total bandwidth, tsamp is the sampling time, Nch the number of frequency channels, DM range shows the DM values covered, and z−value the maximum number of Fourier bins used with the matched

lters to recover accelerated signals (see text).

νc ∆ν tsamp Nch DM range z−value (GHz) (GHz) (µs) (cm−3pc)

4.85 0.5 262.144 128 800−11900 1200

8.35 0.5 131.072 128 800−15080 1200

14.60 0.5 65.536 128 800−10800 1200

18.95 2.0 128 256 800−21200 1200

the maximum allowed by the software. During the periodicity search step, harmonic summing is applied, with up to 16 harmonics summed for non-accelerated signals, and up to 8 harmonics in the other case. The lter matching algorithm of presto has the advantage of being sensitive simultaneously to slow spinning pulsars with high accelerations, and fast spinning pulsars with lower acceleration ranges, because of the trade-o in the z−value between these quantities. We note that the large acceleration

ranges used in this search, that are key to be sensitive to the most extreme pulsar binaries, were mainly possible thanks to the newly developed GPU-powered pipeline. Otherwise, with the computational resources available at the time6 _{the computation}

would have taken too long to be processed in time for this thesis. Table 3.2 shows a summary of the data and searching set-up.

After the periodicity and acceleration searches, all the detected signals are sifted to select only those above a threshold of harmonically summed power of6σ, removing in the process the duplicated (i.e. detected at dierent DMs and accelerations) and the harmonically related candidates.

Finally, the folding and creation of pulsar candidate evaluation plots is done for up to 150 candidates per segment. The folding and plotting step is done twice per candidate, one time letting the programme prepfold to optimise the parameters of the candidate and a second time without optimisation. This is necessary because the parameter optimization step can improve the detection signicance of a candidate, especially those accelerated, but it has the risk of being confused by an interference signal close to the candidate spin period, in which case the candidate is lost. This makes the folding and plotting process more robust against RFI.

3.2.2.2 Analysis of the acceleration detection capabilities of the pipeline