At the end, we come to the observation and analysis techniques used for discovering new pulsars. Pulsar searches fall into one of two broad categories : targeted and untargeted searches. In an untargeted search (or survey) for pulsars, the idea is to uniformly cover a large area of the sky with a desired sensitivity in flux level. In targeted searches, one is searching a limited area of the sky where there is a higher than normal possibility of finding a pulsar (for example, the region in and around a supernova remnant or a steep spectrum point source identified in mapping studies). Here some of the parameters of the search can be tailored to suit the a priori knowledge about the search region.
For a pulsar survey, the choice of (i) the range of directions to search in, (ii) the frequency of observations, (iii) the bandwidth and number of spectral channels, (iv) the sampling interval and (v) the duration of the observations are some of the critical items that need to be chosen carefully. The choice of these parameters is interlinked in many cases.
Analysis of pulsar search data is an extremely compute intensive task. For each position in the sky for which data is recorded, the analysis technique needs to search for the presence of a periodic signal in the presence of system noise. However, from the discussion in section 3, it is clear that if appropriate dispersion correction is not done, the sensitivity to the presence of a periodic signal can be reduced significantly. Since a pulsar can be located at any distance (and hence DM) along a given direction in the sky, the search has to be carried out in (at least) two dimensions : DM and period. For this, the data is dedispersed for different trial dispersion measures. For each choice of DM, the dedispersed data is search for a periodic signal.
To reduce the computational load for search data analysis, several optimised algorithms are used. For example, when dedispersing for a range of DM values, it is possible to use the results from the computations for some DM values to compute part of the results for some other DM values. This saves a lot of redundant calculations. This method, known as Taylor's Dedispersion Algorithm, is used quite often. Similarly, there are optimised techniques for searching for periodic signals in the presence of noise. The simplest method is to fold the dedispersed data for each choice of possible period and examine the resulting profile for the presence of a significant peak that is well above the noise level. Once again, computations done for folding at a given period can be used for folding at other periods. This redundancy is exploited by the Fast Folding Algorithm. A signal containing a periodic train of pulses gives a well defined signature in the Fourier domain - its spectrum consists of peaks at the frequency corresponding to the periodicity, and harmonics thereof. It can be shown that it is possible to detect the periodic signal by searching for harmonically related peaks in the spectral domain. It turns out that it is more economical to implement the FFT followed by harmonic search technique compared to the folding search techniques.
Additional complications are introduced in the search algorithm when one allows the parameter space to cover pulsars in binary orbits as the period can actually change during the interval of observation. Special processing techniques are needed to handle such requirements.