EXPERIMENTAL SET-UP FOR DETECTING VERY FAST AND DISPERSED MILLISECOND PULSARS B. ROUGEAUX1∗ , G. PETIT1∗∗ , T. FAYARD2 and E. DAVOUST3 1 BIPM, Pavillon de Breteuil, F-92312 Sèvres, France 2 CNES, 18 Av. E. Belin, F-31055 Toulouse, France 3 OMP, 14 Av. E. Belin, F-31000 Toulouse, France

∗ Present address: National Physical Laboratory (NPL), Teddington, Middlesex, UK ∗∗ author for correspondence, e-mail: [email protected]

(Received 3 November 1999; accepted in revised form 24 March 2000)

Abstract. We have developed observational and data processing techniques for detecting millisecond pulsars. The method of data processing consists in correlating time series of data by folding them in two steps according to a trial period. While very time consuming, especially for very short periods, this methods allows us in principle to detect very fast and dispersed millisecond pulsars. The experimental set-up has been integrated into a data acquisition system developed at Centre national d’études spatiales (CNES). Observations of known millisecond pulsars at Nançay radio observatory are used to validate the system and assess the sensitivity of the survey, which is 3.5 mJy for 240 sec of integration. Future improvements in the radio-telescope and the system should bring it down to below 1 mJy. The application of the method to a millisecond pulsar survey is under way at Nançay radio observatory in collaboration between the Bureau international des poids et mesure (BIPM), Observatoire Midi-Pyrénées (OMP), and CNES. Keywords: data processing technique, instrumentation, millisecond pulsars, radio astronomy, survey

1. Introduction In the last ten years, millisecond pulsar timing in the radio frequency band has evolved into a major field of research for the study of pulsars and their emission processes. It has yielded important results in fundamental physics, cosmology, and astrometry (Blandford, 1992). Discovering new pulsars is essential for gaining on the existing results. In addition to the better statistics of a larger data set, it can be expected that pulsars with better intrinsic rotation stability will be discovered. Presently, the hints are that old millisecond pulsars with very short period P and low period derivative P˙ would be the most plausible candidates for best stability (Arzoumanian et al., 1994). Several such pulsars could be used to build a dynamic time scale with a very good long term stability (Petit and Tavella, 1996). In addition, the possible discovery of millisecond pulsars with very short period would allow one to set new constraints on the state equations of matter (Friedman et al., 1989; Kulkarni, 1992; Burderi and D’Amico, 1997). To design a radio pulsar search, trade-offs are inevitable between conflicting Experimental Astronomy 10: 473–485, 2000. © 2000 Kluwer Academic Publishers. Printed in the Netherlands.

474

B. ROUGEAUX, G. PETIT, T. FAYARD AND E. DAVOUST

parameters, technical capabilities and cost or complexity (see, e.g., D’Amico, 1998). Partly as a consequence, only a small fraction of the expected galactic population of millisecond pulsars has been discovered so far (Lyne and Graham-Smith, 1998) and PSR B1937+21 still holds the record for the pulsar with the shortest rotation period (1.56 ms), despite subsequent surveys theoretically sensitive to even shorter period objects. One reason is that the sensitivity limit of all pulsar surveys that have been conducted so far deteriorates dramatically when searching for very short period pulsars, especially for large values of the dispersion measure. Indeed on the one hand such signals are significantly affected by dispersion smearing across individual filterbank channels and, on the other hand, the duration of the pulse may be significantly shorter than the sampling rate which results in a loss of signal-to-noise ratio. Therefore only the very nearby bright pulsars (i.e. those with a very small dispersion measure) can be detected (Lorimer, 1997), so that the present limits on the existence of sub-millisecond pulsars are rather poor. It is thus expected that new surveys consider ever faster sampling rates and narrower filterbank channels (see, e.g., D’Amico, 1997). Along this line of thought, we have set up a system to carry out a millisecond pulsar survey with the special goal of obtaining a detection sensitivity optimised for short periods and large values of the dispersion measure. We emphasize that, whenever possible, we favour detection sensitivity for this class of pulsars with respect to other parameters such as processing load. For this purpose, observational techniques have been developed and integrated into a data acquisition system developed at CNES. A survey has been started at the Nançay radio observatory using this system. In Section 2, we describe the method of data processing we have used. The third section develops the application to our survey: The experimental set-up, including the acquisition system and custom designed hardware, and the implementation of the processing system. Finally we estimate the sensitivity of our survey through the observations of known pulsars.

2. Method of data processing Because the individual pulses are very weak, it is necessary to fold the signal at the period of the pulsar to detect its emission. The interstellar medium causes the pulsar signal to be dispersed in frequency, an effect that is quantified by the dispersion measure (DM). In the folding procedure, the total band of the system is generally divided into smaller passbands to take into account the dispersion of the signal; appropriate delays are thus applied to each passband before adding them. In a survey, the folding technique must be performed over a range of pulsar periods and dispersion measures. When the rotation period of the searched pulsar is very short (a few milliseconds), a high sampling rate of the signal is required and consequently a large quantity of data must be recorded and processed.

DETECTING VERY FAST AND DISPERSED MILLISECOND PULSARS

475

We consider the observation of one point of the sky for a duration T (typically a few minutes), during which the single sideband voltage is digitised and recorded. While our hardware has been developed to perform one-bit and two-bit digitisation, it has so far only been used in one-bit sampling, so only this case is considered in this article. The basis of the processing is to calculate the power spectrum of the digitised signal over time intervals known as ‘windows’. A window is a time interval of duration L of the order of the expected pulse width, in other words a small fraction (tens of µs) of the trial period. The resulting power spectra are then folded for many trial periods, and the results from the different spectral bands added for many trial values of the dispersion measure. A pulse will manifest itself as a strong signal for a given window, period and dispersion measure. 2.1. D ETECTION

TECHNIQUE

The technique of detection is based on the comparison of power spectra between windows with and without signal. Following the well-known method of auto-correlation spectroscopy (Cooper, 1976; Thompson et al., 1986), the power spectrum P is the Fourier transform of the auto-correlation function R. Starting from the digitised data x, we first calculate the digital auto-correlation function over each window for Nl lags as: ! NX w −1 1 R(n) = xm xm+n , Nw m=0 where Nw is the number of samples (i.e. bits) in a window and n = 0, 1, . . ., Nl . When the results of the whole time series have been folded for a given pulsar period (see Section 2.2) we are left with a set of values Rnj for lag n (from 0 to Nl ) and window j (from 1 to Nf , with Nf windows covering the period). From these we compute the continuous auto-correlation function estimates ρnj . The Fourier transform yields the powerg estimates Pij for frequency band i (from 1 to Nl + 1) and for window j : "  # Nl X 1 n(j − 1) Pij = ρnj cos π ρ0 + 2 . Nl + 1 Nl + 1 n=1 The power spectrum over a window could equivalently be estimated by means of first obtaining voltage spectra, converting them into power spectra and averaging these over the window. Given the relatively small number of spectral channels (typically 32), and owing to the fact that auto-correlation sums may be treated by a computer as short integers, it has been found computationally more efficient to choose the auto-correlation approach. From the power spectrum over all the windows forming the trial period, it is possible to detect a pulsar signal by summing over the frequency bands, applying

476

B. ROUGEAUX, G. PETIT, T. FAYARD AND E. DAVOUST

appropriate time delays to each band to take the effect of dispersion into account. Such a sum yields Nf values of the total spectral power Pi for a given folding period and dispersion measure. These spectral power values may be expressed in terms of S/N ratio by estimating the mean M and the standard deviation σ of the values Pi for a large number of trial periods and dispersion measures. The S/N ratio is calculated for every window i as SNRi = (Pi − M)/σ . It should be noted that if the dispersion measure of a pulsar is low enough for its signal to appear in the same window j for all frequency bands, such a pulsar will not be detectable (see Section 4 for the implications on the sensitivity limits of the survey). Indeed, in this case, Pi is the same whether a signal is present or not because for each window the PNl +1 power spectrum is normalised ( i=1 Pij = 1). 2.2. S EARCH

TECHNIQUE

Pulsar searches have been initiated at the end of the 1960s after the first radio pulsar was discovered in 1967. The basic techniques for performing such a search with digitised data, taking into account the periodicity and the dispersion of the signal, were described at that time (see, e.g., Lovelace et al., 1969; Burns and Clark, 1969; Hamilton et al., 1973). Later on, these methods were refined and adapted into integrated systems for implementation at radio-observatories (see, e.g., Backer et al., 1990). There are two classes of methods for detecting pulses in the observed signal. In the first class, correlation of the time series of data is performed, folding the data in time according to a trial period. Fast folding methods using a binary tree algorithm are used to try all periods up to some multiple 2l of the elementary sample L with a minimum number of operations. In the second class, a Fourier transform of the time series of data provides a high-resolution spectrum in which peaks, corresponding to the period of the pulsar and its harmonics, are searched for. A technique of harmonic averaging is used to enhance the detection of the peaks. In both classes of methods, the unknown dispersion of the pulsar signal, which is necessary for folding the data in frequency, has to be searched for by trials. This frequency folding may also be performed using a binary tree algorithm to minimise the number of operations. The second class (FFT method) is generally much more efficient in terms of execution time (see, e.g., Burns and Clark, 1969). We nevertheless chose a method which belongs to the first class for the following reasons. First, the computational advantage of the FFT method is mainly important when searching in a very large range of periods (in principle one may search for periods from 2L to T /2) while we restrict our search to a limited range of periods where the sensitivity of existing surveys is the lowest. Indeed we typically search from the lowest expected period of a pulsar (∼0.5 ms) to a couple of milliseconds, i.e. only over two octaves. The second reason is that, in the case of weak signals, the sensitivity of methods in the first class, where the total power of all the pulses is added coherently, is in

DETECTING VERY FAST AND DISPERSED MILLISECOND PULSARS

477

principle better than those in the second class where the total power is present in different Fourier components which are added incoherently. This advantage is offset by the fact that, in the second class of methods, the sensitivity is less dependent on the pulse shape which only affects the way in which the power is distributed among the different harmonics. By choosing a method in the first class, we optimise our survey for finding pulsars with clean pulses, covering only a small fraction of the period. Finally, in the first class of methods, the phase of the signal is obtained from the analysis. In general, this is of little importance for a survey, however we take advantage of this to perform the search in several adjacent or nearly adjacent frequency bands independently. We then compare the results of the independent searches to test if the phases of statistically significant signals may actually correspond to the same pulsar. Consequently, the search requires somewhat less computation and is less sensitive to interferences which, in general, are present in only one frequency band. As noted above, one drawback of our choice is that it provides optimal sensitivity only for clean and narrow pulses. If the pulse profile is complex or if the pulse if wider, the sensitivity decreases. In a recent study (Kramer et al., 1998, 1999), millisecond pulsars pulse profiles mostly taken at 1.4 GHz are examined. Over a sample of about 30 millisecond pulsars, the reported pulse width for 50% intensity level would be smaller than the size of one window in our processing (i.e. 5 to 7% of the period) for about one out of three pulsars. It would also be smaller than two adjacent windows in about two out of three pulsars. Even for cases where this is not the case, a very significant fraction of the power is generally still present in two adjacent windows, the only notable exception being J0218+4232. It is obviously difficult to extend these conclusions to the class of very short period millisecond pulsars and hypothetical sub-millisecond pulsars that we search for, however there is no reason to expect that their pulses should be significantly wider or more complex (Kramer et al., 1998). For this reason, we examine for each trial, in the way indicated in Section 2.1, the maximum value of the power present in one window, and also in a set of two adjacent windows. 2.2.1. Period trials: time folding strategy The binary tree algorithm is the most efficient procedure for folding the data in time, however it has some drawbacks. First, since it folds a number of samples which is always a power of 2, it does not make optimal use of all the data for all the searched periods, thus resulting in a loss of sensitivity occurring at each step of the algorithm. Also, when applied to large data sets (in our case T /L ∼ several 106 and Nl is typically 32), it requires more memory than was (and still is) usually available on standard computers. Finally, it does not readily allow us to consider the case of binary pulsars for which the period varies significantly over the duration T (see Section 2.3). We have therefore implemented a two-level folding algorithm, with an intermediate interval (known as a ‘model’) of duration a fraction of the total

478

B. ROUGEAUX, G. PETIT, T. FAYARD AND E. DAVOUST

duration T . On the first level, the time folding is performed within each model with a variation of the binary tree algorithm (Petit, 1994). On the second level, the results of the different models are simply added (there is no special addition algorithm), however different addition procedures allow us to search for the timevarying period of binary pulsars (see Section 2.3). 2.2.2. Dispersion measure trials: frequency folding strategy In the time-frequency plane, and over a bandwidth small enough with respect to the observation frequency, the dispersion delay is linear in frequency and its slope may be expressed by a parameter known as the dispersion slope K (in µs/MHz) that is related to the dispersion measure DM (in cm−3 pc) by the relation K=

8.30×DM , ν3

where ν is the observation frequency in GHz. Since the basic unit of time in our processing is the window and the basic unit of frequency is the frequency band (the recorded bandwidth divided by Nl + 1), it is convenient to express K as a slope p in units of window per frequency band. Indeed, between two trial slopes, the appropriate step corresponds to one window over the total pass band, i.e., δp = 1/(Nl + 1). The maximum trial slope is also conveniently expressed in window per frequency band: any slope p larger than about 1 window per frequency band will smear the pulse over about p adjacent windows, therefore increasing the minimum detectable flux density by almost a factor p. The implications on the sensitivity of the survey will be detailed in Section 4. If we choose to keep p lower than 2, and if Nl is 32, the number of trial slopes is 64 = 26 and a binary tree algorithm can easily be implemented to perform the frequency folding. 2.3. T HE

CASE OF A BINARY PULSAR

For pulsars in tight binary systems, folding in time according to a constant period may induce a loss of sensitivity with long integration times because the apparent pulse period may vary significantly over the integration duration due to the variation of the Doppler shift. At the cost of additional computing time, it is possible to take into account this effect by folding with an acceleration parameter. This question has already been studied by Johnston and Kulkarni (1991), particularly for binary millisecond pulsars in clusters. They defined a factor γ to describe the efficiency of detection of binary millisecond pulsars with and without acceleration parameter. In our implementation, we use as an acceleration parameter the displacement of the pulse x, expressed in windows, due to the variation of the period over the total duration T . This parameter x can be linked to orbital parameters: Porb , the orbital period and a sin i, the projected semi-major axis (in second).

DETECTING VERY FAST AND DISPERSED MILLISECOND PULSARS

479

Supposing a circular orbit, the radial velocity has a sinusoidal variation and the maximum value of x in this case is given by the relation (Rougeaux, 1999): 2 Porb 4π 2 T 2 a sin i < xL . For example, taking a data length of 240 s and a window duration of 50 µs, 3 of the 31 known millisecond pulsars in circular binary system (Lyne and Graham-Smith, 1998) may cause a maximum displacement x of 16 windows, for which a significant degradation of sensitivity will occur if this effect is not taken into account. Indeed, for a given value of√x, the coherent integration time without acceleration search is approximately T / x so that the sensitivity is divided by x 1/4 . 3. Application to a survey 3.1. E XPERIMENTAL

SET- UP

3.1.1. Acquisition system The experimental set-up of the survey for very fast and dispersed millisecond pulsars, initiated at the Nançay radio-observatory, is the following: Four 6.4 MHz RF bands are provided by the radio-telescope and are recorded with a sampling interval of 62.5 ns (16 Mbps on each of four channels) using the S2 VLBI data acquisition system (Wietfeldt et al., 1996). The S2 is a new-generation very long base interferometry (VLBI) tape recorder designed at the Institute for Space and Terrestrial Science (ISTS) in Toronto. It sustains a maximum data rate of 128 megabits per second i.e. a maximum pass band of 64 MHz in the Nyquist limit. Eight transports allow recording 2 terabits of data in 240 minute’s videotape. The S2 unit used for this experiment is the property of OMP and has been installed at Nançay for temporary use. To adapt the S2 to the radio-telescope, CNES designed a digitiser. It takes analogue data from the four available channels and digitises in two bits. Four transports of the S2 record the amplitude bit, the others the sign bit. Each channel is configured differently for the observations: the usual configuration is to record two frequency bands (at 1400 and 1410 MHz) and two polarisations for each frequency. The data are recorded on videotape. They are subsequently transferred to computer disks using a specially designed interface developed at CNES. It is based on PC and a modified SCSI technology. It operates under Windows95 and sustains a rate of 4 Mbytes per second without data loss. This corresponds to the data rate of two S2 transports and allows one to retrieve either one channel with two-bit sampling or two channels with one-bit sampling. 3.1.2. Processing implementation The implementation of the processing method must use a rather powerful system because of the ∼1012 operations presently needed to test periods between 650 and

480

B. ROUGEAUX, G. PETIT, T. FAYARD AND E. DAVOUST

2600 µs and dispersion slopes p lower than 2 (see Section 4 for the corresponding DM limits). A network of PCs installed at BIPM has been used to implement the processing. The network is managed by a workstation which automatically retrieves the results of each personal computer when the calculations are finished. The 2-level time folding described above has been implemented, with the first level performed by the network of PCs and the second by the workstation. The four channels initially available are processed in pairs, where the two channels with the different polarisations of the same frequency are merged together to reduce processing time. With this implementation, a network of seven personal computers (typical for the year 1997) can calculate one point of the sky survey (4 min duration) in about six hours. This use is limited to night processing because the computers are devoted to other tasks during the day. However during a weekend, it is possible to process several such points of the survey. An important part of the processing is the graphic display of all processing results. This has been developed in C under UNIX system on the workstation. After the second level of time folding has been performed, S/N ratios are obtained for all ‘window’ time intervals, trial periods and dispersion measures. Next, these results are sorted out and compared. This part of software can be launched automatically when each computer has finished its part of the processing. There is a probable detection when a signal appears with sufficient SNR in each of the two frequency bands. If it appears in only one, it is most probably an interference. However it may also be a pulsar whose flux density is highly variable with frequency and some checks have to be made. It is difficult to implement a fully automatic decision criterion and the display must show the results of comparisons and all S/N ratios likely to be pulsar signal. Points with colours depending on their S/N ratio value for the two frequency bands are plotted on a chart with period on the x axis and dispersion measure expressed by the slope p (as defined in Section 2.2) on the y axis. A small number of such charts allows one to visualise all the results of the processing for one point of the sky survey. Figure 1 shows one such chart displaying the results with direct comparison between the two channels. The search processing will subsequently be carried out on another computer system. For a more extensive processing, it is expected that a new parallel computer based on the Sharc processor will be used. 3.2. O BSERVATION

OF KNOWN PULSARS

We have carried out observations of four known millisecond pulsars: B1937+21, B1643-12, B0613-02 and J1713+0747. These were selected from the set of millisecond pulsars in the Princeton catalogue (Taylor et al., 1993, 1995) as having a short period and a relatively large flux density and dispersion measure. The data were processed with the method described above (survey mode) but searching only

DETECTING VERY FAST AND DISPERSED MILLISECOND PULSARS

481

Figure 1. Example of a display for a part of the processing for one point of the sky (detection of B1937+21). High SNR values obtained on two different frequency bands are represented for a set of trial values of (period, dispersion slope p).

a small interval of periods centred on the approximate value of each pulsar period. A sample plot emphasizing the detection of B1937+21 is shown in Figure 1. B1937+21 (P = 1.56 ms, DM = 71.0 cm−3 pc) has the characteristics of the typical target of our survey. Since it also has the largest average flux density of all the known millisecond pulsars, it has been observed many times since the installation of the system in November 1996, and has also been used as a calibration reference (see Section 4). B1643-12 (P = 4.62 ms, DM = 62.4 cm−3 pc) has been observed in November 1996: Since its flux density is relatively stable over time and frequency (Maitia, 1998) it can also be used as a calibration reference. B0613-02 (P = 3.06 ms, DM = 38.8 cm−3 pc) is part of a binary system in a 1.2 day orbit. We selected it for testing the capacity of the processing system to account for the variation of the pulsar period due to its orbital motion. It was observed in December 1998, but, unfortunately, only one successful observation was obtained and the period variations were negligible because of the orbital phase of the pulsar at the time. In addition, its wide pulse (Lorimer et al., 1995) does not make it optimally detectable for our survey method. J1713+0747 (P = 4.57 ms, DM = 16.0 cm−3 pc) has been detected in observations carried out in December 1998. However, the results cannot yield significant

482

B. ROUGEAUX, G. PETIT, T. FAYARD AND E. DAVOUST

conclusions for our survey. Its flux density is highly variable in time and frequency, so that it cannot be used as a calibration reference. Furthermore, the dispersion across the bandwidth is so small compared to the period that it would not be detected in standard survey processing.

4. Sensitivity of the survey and its domain of application In order to determine the sensitivity of our survey method under operating conditions, we carried out observations of B1937+21 that we processed using the standard procedures described above. We derive the average flux density corresponding to the detection threshold (SNR = 7) from the value of the detection SNR and the knowledge of the flux density of the main pulse of B1937+21 during our observations. Ten minutes of observation carried out on 16 October, 1998 with two 6.4 MHz frequency bands (1400 and 1410 MHz LSB) were used for this analysis. The flux density of the main pulse of B1937+21 has been determined by Foster et al. (1991). From Table 4 in that reference, we estimate that the main pulse of B1937+21 observed at 1400 MHz is a source with an average flux density of about 9 mJy (the total average flux density of B1937+21 is larger but also includes the interpulse). Since the flux density of B1937+21 varies significantly at this RF frequency, over time scales of several minutes and frequency bands of a few MHz due to interstellar scattering (Cognard, 1993), we first have to estimate the actual flux density during our observations on the frequency bands used. For this purpose, simultaneous recording of B1937+21 data was performed at Nançay with the NBPP system (Backer et al., 1997). This system recorded 96 channels of 1.5 MHz spanning the band 1288–1432 MHz during this experiment. Since the channels have a bandwidth of 1.5 MHz, a pack of four adjacent channels is about equivalent to one of our bands. After evaluation of the data quality, 14 packs of four adjacent channels were kept for a total bandwidth of 84 MHz. Over such a large bandwidth, we estimate that the observed flux density is indeed the average flux density of the pulsar. By comparing the average SNR of the fourteen 4-channel packs with the SNR obtained with the two 4-channel packs closest to our observing bands, we estimate that the observed flux density over our bands was actually slightly lower than average, about 8 mJy. Our detection SNR for B1937+21 was of the order of 11 for each band and 4 min of integration. We have shown that its average flux density was 8 mJy during the observations, so we estimate that the detection threshold (7σ ) of the survey is about 3.5 mJy average flux density, corresponding to a SNR of 5 in each of the two bands for four minutes of integration. Assuming a sensitivity of 1 K/Jy and a system temperature of 50 K, we would expect to obtain a detection threshold (7σ ) of about 2 mJy for four 6.4 MHz frequency bands and 240 s of observation. This corresponds to the mean flux density of a pulsar whose pulse duration is 5% of the period, accounting for a 25% loss due to rounding errors in the windowing process.

DETECTING VERY FAST AND DISPERSED MILLISECOND PULSARS

483

Some factors may explain this discrepancy, such as some aliasing in the digitising interface, which is still under study. However the most important one may be an increased noise level at Nançay due to interferences. In experiments conducted with the NBPP system, similar discrepancies were found between expected and observed noise levels (Cognard, 1999, personal communication). The above result is valid when the energy of the pulse is kept within one of the windows, i.e. for a narrow pulse (5 to 7% of the period) with little smearing over one frequency band (dispersion slope p smaller than about one window per band) and when the dispersion measure is not so small as to render the signal undetectable (see Section 2.1), i.e. for p larger than about 2–3/Nl window per band. For values of p larger than 1, the pulse will be smeared over about p adjacent windows so that the detection threshold increases about linearly with p. When we search for signal in two adjacent windows (pulsars with pulse width of order 10 to 15% of the period), the detection threshold (minimum average flux density) is larger by about √ 2 for p < 1, but loss due to smearing is comparatively less important than for the one-window case so that the difference is somewhat reduced for larger values of p. Another factor to consider is the pulse broadening due to interstellar scattering. From Taylor and Cordes (1993) we estimate that the pulse broadening at 1.4 GHz will remain, on average, at or below half a window in all cases considered by our present survey program. Only over a few hundred square degrees of sky area towards the galactic center would the average broadening (as computed by the Taylor and Cordes model) reach or exceed the duration of one window, and restrict our sensitivity for a part of the search parameter space (see Figure 2). Apart from this effect, we can derive the sensitivity limits of the survey from the above simple considerations, taking into account the parameters chosen in the search procedure. For trial periods between 0.65 and 2.6 ms, we take the window as 7 to 5% of the period (7% for the shortest periods, 5% for the longest), i.e. a window duration of 46 µs (P = 0.65 ms) to 130 µs (P = 2.6 ms). Adopting a frequency band of 250 kHz, we derive values for the DM corresponding to p = 1 and for the minimum DM allowing detection (p∼0.1). These are 60 cm−3 pc (respectively 6 cm−3 pc) for the shortest period, and 173 cm−3 pc (respectively 17 cm−3 pc) for the longest period. Since we currently limit the search to p = 2 windows per band, the largest DM values searched are about 120 cm−3 pc for the shortest period (0.65 ms) and about 350 cm−3 pc for the longest period (2.6 ms). The corresponding sensitivity limit of the survey is indicated in Figure 2. In addition to the area searched (highlighted in the figure), it is also possible to detect pulsars with parameters outside this region when an harmonic falls in the searched area. The sensitivity is then reduced accordingly.

484

B. ROUGEAUX, G. PETIT, T. FAYARD AND E. DAVOUST

Figure 2. Present (1998) 7σ sensitivity limit (mJy) of the survey in the parameter space (Period, DM) for the one-window/two-window cases (solid lines). The parameter area presently searched for is highlighted. The dashed lines correspond to the indicated pulse broadening (in windows) when observing towards the galactic center. The sensitivity may then be reduced accordingly in the portion of the parameter space above these lines. The parameters of some field millisecond pulsars are also indicated.

5. Conclusion

We have described a processing method to detect very fast and dispersed millisecond pulsars. This method has been set up with the hardware system presently operating in Nançay. This millisecond pulsar survey is optimised for pulsars with very short period (from a few hundred microseconds up to a few milliseconds) and dispersion measures between 10-20 and 120–350 cm−3 pc. Moreover if the pulsar is in a close binary system, the search software can take into account variations of the period due to Doppler shift. Observations of 615 scans (with a resolution of 40 (α) × 220 (δ)) in the galactic plane (between 60 and 100◦ of galactic longitude) were conducted at the Nançay radio observatory in 1997 and 1998. By observing known millisecond pulsars, we have shown that the detection sensitivity is currently of order 3.5 mJy for the average flux density of the searched pulsars (period between 0.65 and 2.6 ms, dispersion measure limits varying with the period, see Section 4). As of January 1999, about one square degree has been processed without detection of new pulsars. Future improvements in the hardware of the experiment and in the Nançay system should yield a detection sensitivity below 1 mJy for this class of pulsars.

DETECTING VERY FAST AND DISPERSED MILLISECOND PULSARS

485

Acknowledgments We are grateful to Ismaël Cognard and Jean-François Lestrade for discussions and for providing the NBPP data and to the Nançay staff for their help in this work. We thank three anonymous referees for critical comments that helped to improve the manuscript.

References Arzoumanian, Z., Nice, D. J., Taylor, J. H. and Thorsett, S. E.: 1994, Ap. J. 422, 671. Backer, D. C., Clifton, T. R., Kulkarni, S. R. and Wertheimer, D. J.: 1990, Astron. Astrophys. 232, 292. Backer, D. C., Dexter, M. R., Zepka, A., Ng, D., Werthimer, D. J., Ray, P. S. and Foster, R. S.: 1997, PASP v 109, p. 61. Blandford, R. D.: 1992, Phil. Trans. R. Soc. Lond. A 341, 177. Burderi, L. and D’Amico, N.: 1997, ApJ 490, 343. Burns, W. R. and Clark, B. G.: 1969, Astron. Astrophys. 2, 280. Cognard, I.: 1993, Thèse de Doctorat, University of Paris VI. Cooper, B. F. C.: 1976, in M. L. Meeks (ed.), Methods of Experimental Physics, Vol. 12B, Academic Press, p. 280. D’Amico, N.: 1997, in A. Alpar, R. Buccheri and J. van Paradijs (eds.), The Many Faces of Neutron Stars, Kluwer. D’Amico, N.: 1998, Adv. Space Res. 21, 125. Foster, R. S., Fairhead, L. and Backer D. C.: 1991, Ap. J. 378, 687. Friedman, J. L., Ipser, J. R. and Parker, L.: 1989, Phys. Rev. Lett. 62, 3015. Hamilton, P. A., McCulloch, P. M., Ables, J. G., Komesaroff, M. M. and Cooke, D. J.: 1973, Astroph. Lett. 15, 163. Johmston, H. M. and Kulkarni, S. R.: 1991, Ap. J. 368, 504. Kramer, M., Xilouris, K. M., Lorimer, D. R. et al.: 1998, Ap. J. 501, 270. Kramer, M, Lange, C., Lorimer, D. R. et al.: 1999, Ap. J. 526, 957. Kulkarni, S. R.: 1992, Phil. Trans. R. Soc. Lond. A 341, 77. Lorimer, D. R., Nicastro, L., Lyne, A. G. et al.: 1995, Ap. J. 439, 933. Lorimer, D. R.: 1997, Proceedings of the 2nd Gravitationnal Wave Data Analysis Workshop, Paris, Ed. Hello. Lovelace, R. V. E., Sutton, J. M. and Salpeter, E. E.: 1969, Nature 222, 231. Lyne, A. G. and Graham-Smith, F.: 1998, Pulsar Astronomy, Cambridge University Press. Maitia, V.: 1998, Ph.D. Thesis, University of Paris VII. Petit, G.: 1994, Ph.D. Thesis, Paris Observatory. Petit, G. and Tavella, P.: 1996, Astron. Astrophys. 308, 290. Rougeaux, B.: 1999, Ph.D. Thesis, University Paul Sabatier Toulouse. Taylor, J. H. and Cordes, J. M.: 1993, Ap. J. 411, 674. Taylor, J. H., Manchester, R. N. and Lyne, A. G.: 1993, Ap. J. 88, 529. Taylor, J. H., Manchester, R. N., Lyne, A. G. and Camilo, F.: 1995, Catalog available on internet at pulsar.princeton.edu. Thompson, A. R., Moran, J. M. and Swenson Jr., G. W.: 1986, in Interferometry and Synthesis in Radio Astronomy, John Wiley & Sons, Chapter 8. Wietfeldt, R. D., Boer, D., Cannon, W. H., Feil, G., Jakovina, R., Leone, P., Newby, P. S. and Tan, H.: 1996, IEEE Trans. IM 45(6), 923.