Results of the IGEC2 search for gravitational wave bursts during 2005
Abstract
The network of resonant bar detectors of gravitational waves resumed coordinated observations within the International Gravitational Event Collaboration (IGEC2). Four detectors are taking part in this collaboration: ALLEGRO, AURIGA, EXPLORER and NAUTILUS. We present here the results of the search for gravitational wave bursts over 6 months during 2005, when IGEC2 was the only gravitational wave observatory in operation. The network data analysis implemented is based on a time coincidence search among AURIGA, EXPLORER and NAUTILUS, keeping the data from ALLEGRO for followup studies. With respect to the previous IGEC 19972000 observations, the amplitude sensitivity of the detectors to bursts improved by a factor and the sensitivity bandwidths are wider, so that the data analysis was tuned considering a larger class of detectable waveforms. Thanks to the higher duty cycles of the single detectors, we decided to focus the analysis on threefold observation, so to ensure the identification of any single candidate of gravitational waves (gw) with high statistical confidence. The achieved false detection rate is as low as 1 per century. No candidates were found.
pacs:
04.80.Nn, 95.30.Sf, 95.85.SzI Introduction
The search for transient gravitational waves (gws) requires the use of a network of detectors. In fact, the analysis of simultaneous data from more detectors at different sites allows an efficient rejection of the spurious candidates, either caused by transient local disturbances or by the intrinsic noise of the detectors. Moreover, the false alarm probability of the network due to uncorrelated noise sources at the different sites can be reliably estimated.
The first long term search for bursts gw by a network of detectors has been performed by the five resonant bars ALLEGRO, AURIGA, EXPLORER, NAUTILUS and NIOBE, within the International Gravitational Event Collaboration (IGEC) Allen et al. (2000). The search consisted in a time coincidence analysis over a 4year period, from 1997 to 2000, and set an upper limit on the rate of impulsive gravitational waves as a function of the gw amplitude threshold of the data analysis Astone et al. (2003a). However, the overlap in observation time of the detectors was modest: three or more detectors were in simultaneous validated observation for , of the time, and twofold observations covered an additional period of , . Moreover, since in the twofold coincidence searches at the lowest amplitude thresholds some false alarms were expected, the IGEC 19972000 observation was not able to discriminate a single gw candidate from the accidental coincidences for most of the time. The target gw signals were short transients showing a flat Fourier component around 900 Hz, like pulses of duration or oscillating signals with a few cycles of period.
The same class of signals was targeted in the searches performed with the EXPLORER and NAUTILUS data in 2001 Astone et al. (2001) and 2003 Astone et al. (2006a). These searches, being based on twofold coincidences, could not aim at the identification of single GW candidates. They addressed the study of a possible excess of coincidences taking advantage of sidereal time analysis.
Subsequent searches for burst gws have been performed also by networks of interferometric detectors, which feature a better sensitivity in a wider frequency bandwidth. In particular, the LIGO searches demonstrated a significant improvement in sensitivity from the 2003 observation Abbott et al. (2005) to the 2005 observation LSC (2006). Due to the shorter duration of these searches, an improvement on the limit set by IGEC at higher amplitudes on the rate of millisecond gw signals was not possible. However, In November 2005, the LIGO observatory started its first long term scientific observation at its design sensitivity LIG ().
In 2004 four bar detectors resumed simultaneous operation: ALLEGRO McHugh et al. (2005), AURIGA Zendri et al. (2002); Vinante and (for the AURIGA Collaboration ) (2006), EXPLORER and NAUTILUS Astone and (for the ROG Collaboration) (2004); Astone et al. (2006b). A new long term gw search started under the IGEC2 Collaboration, whose primary goal is to identify any single candidate of burst gw with high statistical confidence. This coordinated observation is still running and targets to a broader signal class than the previous IGEC search, as for instance binary BH mergers and ringdowns Campanelli et al. (2006) and longer transients recently predicted for Supernova core collapses Ott et al. (2006).
This paper is the first report on the IGEC2 observations and describes the results of the analysis of 6 months of data, from May the to November the 2005, when IGEC2 was the only gravitational wave observatory in operation. The AURIGA, EXPLORER and NAUTILUS data are actually used to search for gravitational wave candidates showing up as triple time coincidences. Due to a delay in the validation of the ALLEGRO data, we agreed to use the data from this detector for followup investigations on possible signal candidates identified by the other resonant bars.
Ii Characteristics and goals of the IGEC2 observatory
As for the previous IGEC search in 19972000, the detectors are aligned within a few degrees and so feature the same directional sensitivity at any time. The spectral sensitivities of the resonant bar detectors during 2005 is shown in Fig. 1. The minima of the noise power spectral densities are very close, within , as the four detectors share a similar design (i.e. cylindrical Al5056 bar with a mass of cooled at liquid He temperature, resonant transducers, similar mechanical quality factors, dcSQUID signal amplifier). With respect to the previous IGEC1 network Astone et al. (2003a), all detectors have been upgraded and exhibit now wider bandwidths. EXPLORER and NAUTILUS were improved respectively in 2000 and 2002. Some modifications on the cryogenics apparatus and mechanical filters, new transducers and dcSQUIDs were adopted Astone et al. (2003b). The upgrade of AURIGA, completed in 2003, concerned most of the apparatus, from seismic isolation system Bignotto et al. (2005) to the readout Vinante et al. (2001); in particular, a better coupling between the transducer and the signal amplifier was achieved by tuning the electric resonance of the signal transformer to the mechanical modes of the antenna and transducer and the signal amplifier is now based on a two stage dcSQUID. The result was a very large increase in the detector bandwidth Baggio et al. (2005); Vinante and (for the AURIGA Collaboration ) (2006). Additional upgrades of the room temperature suspensions during 2005 led to a significant improvement of dutycycle and data quality. ALLEGRO resumed operation in early 2004, after changing both the resonant transducer and the readout electronics McHugh et al. (2005).
The primary scientific interest of the IGEC2 observations is the ability of identifying any single candidate of gravitational wave signal with high statistical confidence. Moreover, the results hereby presented refer to a period when only IGEC2 was surveying gws.
Iii Characteristics of the exchanged data
During the 180 days considered in this analysis, the AURIGA, EXPLORER and NAUTILUS detectors show a high duty cycle, see Tab. 1. In particular the validated data of AURIGA, EXPLORER and NAUTILUS overlap by 130.71 days in threefold coincidence (corresponding to 73%) and 45 days more are covered as two fold coincidences (about 25%).
The noise stability of the detectors is remarkable, either if compared to past performances or to data analysis requirements. As shown in Fig. 2, the standard deviation of AURIGA noise shows a slow systematic dependence on the liquid He levels in the cryostats with peaktopeak variations of the order of 10 %. Minimum noise levels of EXPLORER and NAUTILUS are higher by a factor of in terms of equivalent amplitude of a millisecond gravitational wave burst.
Each group validates its data and tunes its searches for gravitational wave candidates independently. These analyses are based on linear filters matched to signals. The algorithms implemented for the AURIGA filter and for the EXPLORER and NAUTILUS filter are different and have been independently developed. In both pipelines, the filtered data stream is calibrated to give the reconstructed Fourier component of the strain waveform of a short ( gravitational wave burst at input.
A candidate event is identified by detecting a local maximum in the absolute value of the filtered data stream: the occurrence time of the maximum and the corresponding amplitude are the estimates of the arrival time and of the Fourier component of the gw . Since these estimates refer to the , they are consistent only for gws of short duration with a flat Fourier transform over the detection bandwidth. For waveforms of colored spectral structure, the filter mismatch leads to nonoptimal SNRs (Signal to Noise Ratios) for the candidate events and to biases in their amplitude and time estimates.
As an example, in the case of signals shaped as damped sinusoids with damping time , the typical reconstructed by filters is of the of the signalmatched filter for , and for AURIGA, EXPLORER and NAUTILUS respectively. For such classes of waveforms, the implemented reconstruct the arrival times with relative systematic errors .
A cross validation has been performed on the different analysis pipelines. A sample day of raw data of EXPLORER and of NAUTILUS was processed by AURIGA data analysis, using the same epoch vetoes, but with a different implementation of the data validation and conditioning. The comparison of the candidate events found by AURIGA and ROG pipelines show a good consistency for with some unavoidable differences at lower .
AURIGA  EXPLORER  NAUTILUS  

AURIGA  172.9 d  
EXPLORER  151.8 d  158.0 d  
NAUTILUS  150.2 d  135.3 d  155.0 d 
Event lists of each detector are exchanged according to the protocol of the previous IGEC 19972000 observations IGE (), and include information on amplitude and time uncertainties and on the amplitude threshold used to select the events.
A rigid time offset, chosen within , is added to all the event lists prior to the exchange and is kept confidential, so that all the tuning of the analysis is performed without knowledge of the true coincidences. When the network analysis is completely defined, these confidential time shifts are disclosed to draw the final results. This procedure, referred to as blind analysis, ensures an unbiased statistical interpretation of the results.
The choice of the most suitable exchange threshold is left to each group. The thresholds used to select the exchanged events are for AURIGA and for EXPLORER and NAUTILUS. These are considered as the minimal thresholds that allow the identification of a candidate event by each detector with reasonable confidence, according to the results of tests carried out with hardware and software injections. In particular, at lower the timing uncertainty related to the candidate events increases rapidly.
For the AURIGA events, the conservative estimates of the timing uncertainty () range from a maximum of at the threshold to a minimum of at , as computed assuming signals. For EXPLORER and NAUTILUS, the timing uncertainty was conservatively set to .
The amplitude distribution of the exchanged events corresponding to the period of threefold observations is shown in Fig. 3. The amplitude distribution of the AURIGA events is very close to that expected for Gaussian noise up to , while non Gaussian outliers are dominating at at higher s. The number of candidate events above the minimal thresholds is listed in Tab. 2; the mean event rate is for AURIGA, while it is larger for EXPLORER and NAUTILUS, and respectively, due to the lower thresholds.
data cut AU  

EX  
NA  
AURIGA  186911  34598  790 
EXPLORER  489103  29217  245000 
NAUTILUS  679775  42028  351375 
Iv Network data analysis
The network data analysis consists of a time coincidence search among the exchanged events. The coincidence time window is set accordingly to the same procedure previously implemented within IGEC 19972000 search Astone et al. (2003a). Two events are defined in coincidence if their arrival times and are compatible within their variances, and :
(1) 
where is set to , as in ref. Astone et al. (2003a). According to the ByenaiméTchebychev inequality (see for example Papoulis (1984)), this choice limits the maximum false dismissal probability of the above coincidence condition to 5% regardless of the statistical distribution of arrival time uncertainties. For the threefold coincidence search considered here, the same condition is required per each detector pair, leading to a maximum false dismissal probability . The resulting coincidence windows are between EXPLORER and NAUTILUS and when AURIGA is considered. In the previous searches performed between Explorer and Nautilus a fixed time window of 30 ms was adopted by the ROG group. This value ensured a low false dismissal in the case of deltalike signals considering the measured time response to excitations due to cosmic ray showers Astone and (for the ROG Collaboration) (2004).
Here we have neglected the effect of the propagation time of the gws among the different sites, since it is quite small, . Moreover, in case the signal duration is not small with respect to the coincidence window, the cited false dismissals are no more strictly ensured, because the systematic uncertainties on the arrival time can be different in different detectors (see the previous section).
The coincidence search is tuned to ensure a high statistical confidence in case of detection of any single gravitational wave: 1 false alarm per century. To meet this requirement, we analyze only the 130.71 days of threefold observation by AURIGA, EXPLORER and NAUTILUS, since the twofold coincidence search cannot reach such a low rate of accidental coincidences without sacrificing too much on the sensitivity side.
iv.1 Accidental coincidence estimates
The threefold accidental coincidences has been investigated with high statistics: about 20 millions independent, offsource, resampling of the counting experiment have been performed by applying relative shifts at the times of two detectors within in steps. The resulting changes in the overlap time of the resamplings with respect to the actual observation time are negligible: the mean observation time of the resamplings is less than the actual observation time and the largest difference is at most . Fig. 4 shows the histogram of the counts of the accidental coincidences using the whole set of the exchanged events. The histogram is very well in agreement with a Poisson distribution of mean equal to 2.16 counts per observation time.
Crosschecks on the accidental coincidences rate estimate have been performed with other independent algorithms and different choices of the relative shifts, giving in all cases results well within the expected statistical fluctuations. A further check was pursued with a method based on an analytic estimate of the random coincidence rates (see Appendix). The results were in very good agreement with the values obtained with the timeshifts technique.
iv.2 Data selection
In order to achieve the goal of 1 false alarm per century, a data selection is necessary to reduce by a factor the accidental coincidences found on the exchanged data set. In general, the data selection has to be tuned with the aim of preserving the detection efficiency of the gw survey. In our case the balance between false alarms and detection efficiency has been addressed from first principles, since measurements of average efficiency during the observation time were not available. We decided to perform three searches based on different data selection procedures:

A) fixed and equal thresholds on the of the events of each detector. Its motivation relies on setting a minimal comparable significance for the considered events as well as setting a similar events rate for each detector. Given the different spectral sensitivities, this search is more sensitive to colored signals that show the largest fraction of their power in the overlapping part of the bandwidths (e.g. , see Fig. 1) rather than in the remaining part of the AURIGA bandwidth. Such signals would produce similar in the filtered data of all detectors.

B) fixed thresholds on the of the events, but chosing different thresholds for the detectors so that they correspond to comparable levels of absolute gw amplitude in all detectors. This search is targeted to short signals which feature a flat Fourier transform within the AURIGA bandwidth and therefore appear at higher in AURIGA with respect to EXPLORER and NAUTILUS. It allows to use lower thresholds for EXPLORER and NAUTILUS than the previous data selection procedure.

C) common absolute amplitude thresholds: same procedures used in the IGEC 19972000 search. The different data sets are selected according to a common gw amplitude Astone et al. (2003a): the coincidence search is performed only during the periods when the exchange thresholds of all detectors were lower than and considering only the events whose amplitude is larger than . This procedure is repeated for a grid of selected values. This search, as the previous one, is targeted to short bursts. Differently from the two previous procedures, this search not only selects the events, but also the effective observation time as a function of . Its main advantages are to keep under control the false dismissal probability of the observatory and therefore to make possible an interpretation in terms of rate of gw candidates as well as a straightforward comparison with the previous IGEC upper limit results.
We decided to consider the union of these three searches, i.e. to perform one composite search made by an ”OR” of the three data selections procedures. This new approach simplifies the statistical analysis, since it takes care of the correlations expected in our multiple trials. In fact, any accidental coincidence occurring in more trials is counted only once and the expected overall distribution of accidental coincidences is estimated by histogramming the union of the found accidental coincidences on offsource samples.
Our tuning led to the following choices of data selections: A) ; B) AURIGA , EXPLORER and NAUTILUS , C) common search thresholds . Tab. 2 reports the number of considered events of each detector for data selections A) and B). In particular in A) the event rate is similar in all detectors even though AURIGA features a wider bandwidth. Instead in B), the number of AURIGA events is a few hundred times smaller that that of EXPLORER and NAUTILUS.
There is a significant correlation in false alarms between data selections B) and C), which show a large fraction of common accidental coincidences. Tab. 3 summarizes the numbers of accidental coincidences found for the three data selections on the same offsource resampling considered in Sec. IV.1.
AU  common  
EX  search  
NA  threshold  
data cut  A  B  C 
A  27368  
B  515  39507  
C  147  5177  9280 
The final histogram of the accidental coincidences is plotted in Fig. 5: the probability of getting a non zero number of accidental coincidences is 0.00363 during the observation time, corresponding to 1.01 false alarms per century. The estimated statistical uncertainty on this probability is . This uncertainty has been empirically determined by grouping the offsource samples in many disjoint subsets of equal size. The standard deviation of the number of accidental coincidences in these subsets has been propagated to the mean, calculated on the entire offsource dataset. The resulting is only slightly higher, by a factor , than the one expected from a purely Poisson model. Independent checks with different pipelines and on different sets of offsource samples limit the systematic uncertainty on the probability to .
iv.3 Plan of the statistical data analysis
Before looking at the true coincidences in the onsource data set, we finalize a priori the plan for the statistical data analysis. Two steps are planned: the test of the null hypothesis and the setting of confidence intervals.
We chose to reject the null if at least one triple coincidence is found in the onsource data set of the composite search. This corresponds to a significance of the test of with a statistical uncertainty of . Therefore, if at least one coincidence is found, the collaboration excludes it is an accidental coincidence with the above stated confidence: In fact, the rejection of the null points out a correlation in the observatory at the true time (i.e. not consistent with the measured accidental noise at different time lags). The source of correlation may be gws or disturbances affecting distant detectors (e.g. instrumental correlations).
The final result on the estimated number of coincidences, related to any source of correlated noise or gws, is given by confidence intervals ensuring a minimum coverage, i.e. the probability that the true value is included in that interval. We decided to set confidence intervals according to the standard confidence belt construction of Feldman and Cousins Feldman and Cousins (1998). The noise model for the number of coincidences is the Poisson distribution shown in Fig. 5. To take into account its uncertainties, we consider the union of the confidence belts given by the mean noise , i.e. . Thanks to this low false alarm rate, the chosen confidence belt detaches from 0 when at least one coincidence is found (provided that the coverage is lower than the significance of the null hypothesis test). The final result cannot be easily interpreted in terms of gw source models, since IGEC2 is lacking a measurement of the detection efficiency.
Any triple coincidence found would then be investigated a posteriori using also the data set of ALLEGRO, whenever possible. These followup results would be interpreted in terms of likelihood or subjective confidence by the collaboration and would not affect the significance of the rejection of the null hypthesis. The main goal of the followup investigation will be to discriminate among known possible sources, e.g. gravitational waves, electromagnetic or seismic disturbances, etc.. An exchange of the raw data and gw transfer function would allow to implement more advanced network analyses, as searches based on crosscorrelation. Additional complementary information could come from electromagnetic and neutrino detectors as well as from environmental monitors.
V Results
Once the network analysis was tuned, the groups exchanged the confidential time shifts necessary to reconstruct the onsource data set. This blind procedure makes the statistical interpretation of any result unambiguous.
No triple coincidences are found in the composite search described in the previous Sections and therefore the null hypothesis is not rejected.
The upper limit set by the full search is given in terms of the number of detectable gravitational wave candidates, since the false dismissal of the composite search has not been directly measured for any model of gw source. According to the chosen confidence belt, the upper limits are and events at 90% and 95% coverages respectively. For a gw waveform with a flat Fourier transform over the bars bandwidths, the efficiency of this search is mainly contributed by the data selection B (see sec. IV.2). In this case, according to back of the envelope calculations, IGEC2 features a low false dismissal, , at Fourier amplitudes for optimally oriented sources.
Outside the planned composite search for gws, we checked a posteriori the number of coincidences among all exchanged events. Three coincidences were found, well in agreement with the expected Poisson distribution of mean 2.16, as presented in subsection IV.1. All the events associated with these threefold coincidences were at close to the thresholds and therefore no followup investigation has been implemented for diagnostic purposes.
v.1 Comparison with IGEC previous results
Using the subset of the current results relative to the data cut C, we can compute the upper limit on the rate of millisecond bursts as a function of the amplitude threshold. This upper limit is uninterpreted, i.e. it is set in terms of detectable gws, and is done mainly for comparison with the previous IGEC 19972000 search Astone et al. (2003a), see Fig. 6. The new upper limit improves the old one at lower amplitudes thanks to the better sensitivity of current detectors. The current asymptotic rate, at 95% coverage, is higher than in the previous search because of the shorter observation time, but it is reached at much lower signal amplitudes. In fact, the current detectors feature much more stationary performances and the current search is free from false alarms, while the 19972000 result was dominated at low amplitudes by twofold observations, which typically show several false alarms per year.
As a general remark, the main improvement of the current result is the capability of identification of any single candidate gw, while the previous upper limit was mostly contributed by coincidence searches with much higher false alarm rates. An additional improvement of the current search comes from the new data selections procedures (i.e. data selections A and B), which extend the target towards a broader class of signals.
Vi Final Remarks
The IGEC2 observatory is currently surveying for gw transients. Our results show that for a plain time coincidence search at least threefold bar observations are necessary to identify any single candidate gw with satisfactory statistical confidence. The role of the resonant bar observatory is significative to search for signals occurring whenever the network of the more sensitive interferometric detectors is not fully operative and therefore not able to issue an autonomous detection of a gw candidate. In fact, since the spectral sensitivity achieved by the LIGO instruments is better than a factor at the narrower bandwidths of the IGEC2 detectors, LIGO is nowadays able to perform surveys and, eventually, set upper limits at lower amplitudes and on a wider class of gws signals than IGEC2 LSC (2006). In this framework, IGEC2 can collaborate with the other observatories to extend the time coverage of current gw surveys and can contribute to the identification of rare gw events. In addition, if a candidate gw will be identified by the interferometric observatory, a joint investigation barinterferometer could increase the information on the gw candidate, for instance on the signal direction and polarization amplitudes. To take the most from an hybrid barinterferometer observatory, the data analysis methodology should overcome the intrinsic limitations of a time coincidence search and exploit the phase information of the data streams provided by the different detectors, aiming at the solution of the inverse problem for the wave tensor. Tests of such methodologies are ongoing using short periods of real data sets.
Appendix A analytical estimate of the accidental coincidences
The rate of accidental coincidences of the IGEC2 observatory has been empirically estimated by shifting the time of the detectors’ data. The results have been checked by comparison with the following analytical method, based on the common assumption that the exchanged events are Poisson point processes with a slowly variable rate.
a.1 Analytical model
The expected number of accidental coincidences in the simpler case of a constant time coincidence window and constant event rate is
(2) 
where is the number of detectors, is the common observation time, and the number of events in the detector.
In our case, the coincidence window depends on the detector pair and changes for the different AURIGA events (see eq.1 and sec.III). Therefore, the above expression for has to be modified as shown in the following.
Given a time of an event of the detector #1, the probability that detector #2 has an event at a time such that is
(3) 
and similarly for detector #3. These two probabilities are independent, so that the probability that both occur is . When both occurs (necessary condition for a triple coincidence) we can find the distribution of the variable by considering that the variables have uniform distributions in the intervals respectively. Their probability density functions (p.d.f.) are then
(4) 
with j=2 or 3, and their characteristic functions (Fourier transform of the p.d.f.) are
(5) 
The characteristic function of the variable is
(6) 
and its p.d.f. F(x) is then given by the inverse Fourier transform of
(7) 
yielding the trapezoidal shape shown in Fig.7 and described by
(8) 
where we have assumed .
The Probability that also detectors #2 and #3 are in coincidence, i.e. , is a fraction of the area of this trapezium:
and in the intermediate range
The probability of a triple coincidence at each event of the detector #1 is given by the product . The number of accidental triple coincidences is obtained by further multiplying by
(10) 
which turns to eq.2 when all coincidence windows are equal to .
In our case the coincidence window is set from the timing uncertainties of the single events, according to eq.1. Then, it is easy to verify that is bounded between the difference and the sum of the other two ’s so that is given by eq.LABEL:eq_area. The resulting analytical estimate for the accidental coincidences in case of constant event rates and different time windows is
(11) 
a.2 Implementation
The common observation time of the three detectors has been divided in short subintervals with a duration randomly chosen within a selected range, e.g. from to . The minimum and maximum duration must be chosen to meet the assumptions required by eq.A10. In particular, the event rate should be stationary, the coincidence window much smaller than the average distance between events and the number of accidental (background) events much larger than the number of signal (foreground) events.
Since the AURIGA events had a variable time uncertainty, we computed eq.11 for each of them using different time windows. The prediction of is obtained by summing the result over all the AURIGA events in the subinterval
(12) 
where is the interval duration, are the number of events of Ex, Na, Au, is the ExNa (fixed) coincidence window, is the coincidence window Au(Ex or Na) computed with the of the Auriga event, and is the combination of windows in eq.11. The total result for the whole overlapping period is then obtained summing over all the subintervals.
This procedure can be repeated with a different choice of the minimum and maximum intervals duration and/or with a different random initialization, in order to evaluate the fluctuations in the numerical value of the final result.
Footnotes
 preprint: APS/123QED
References
 Z. Allen, P. Astone, L. Baggio, D. Busby, M. Bassan, D. Blair, M. Bonaldi, P. Bonifazi, P. Carelli, M. Cerdonio, et al., Physical Review Letters 85, 5046 (2000).
 P. Astone, D. Babusci, L. Baggio, M. Bassan, D. G. Blair, M. Bonaldi, P. Bonifazi, D. Busby, P. Carelli, M. Cerdonio, et al., Phys. Rev. D 68, 022001 (2003a).
 P. Astone, M. Bassan, P. Bonifazi, P. Carelli, E. Coccia, C. Cosmelli, S. D’Antonio, V. Fafone, G. Federici, A. Marini, et al., Classical and Quantum Gravity 18, 243 (2001), URL http://stacks.iop.org/02649381/18/243.
 P. Astone, D. Babusci, R. Ballantini, M. Bassan, P. Carelli, G. Cavallari, F. Cavanna, A. Chincarini, E. Coccia, C. Cosmelli, et al., Classical and Quantum Gravity 23, S169 (2006a), URL http://stacks.iop.org/02649381/23/S169.
 B. Abbott, R. Abbott, R. Adhikari, A. Ageev, J. Agresti, B. Allen, J. Allen, R. Amin, S. Anderson, W. Anderson, et al., Physical Review D 72, 62001 (2005).
 LSC Collaboration, http://arxiv.org/abs/0704.0943 (2007).
 see for instance http://www.ligo.caltech.edu/.
 M. P. McHugh, W. W. Johnson, W. O. Hamilton, J. Hanson, I. S. Heng, D. McNeese, P. Miller, D. Nettles, J. Weaver, and P. Zhang, Classical and Quantum Gravity 22, S965 (2005), URL http://stacks.iop.org/02649381/22/S965.
 J. Zendri, L. Baggio, M. Bignotto, M. Bonaldi, M. Cerdonio, L. Conti, M. De Rosa, P. Falferi, P. Fortini, M. Inguscio, et al., Class. Quantum Grav 19, S1925 (2002).
 A. Vinante and (for the AURIGA Collaboration ), Classical and Quantum Gravity 23, S103 (2006), URL http://stacks.iop.org/02649381/23/S103.
 P. Astone and (for the ROG Collaboration), Classical and Quantum Gravity 21, S1585 (2004), URL http://stacks.iop.org/02649381/21/S1585.
 P. Astone, R. Ballantini, D. Babusci, M. Bassan, P. Carelli, G. Cavallari, F. Cavanna, A. Chincarini, E. Coccia, C. Cosmelli, et al., Classical and Quantum Gravity 23, S57 (2006b), URL http://stacks.iop.org/02649381/23/S57.
 M. Campanelli, C. Lousto, and Y. Zlochower, Physical Review D 73, 61501 (2006).
 C. Ott, A. Burrows, L. Dessart, and E. Livne, Physical Review Letters 96, 201102 (2006).
 P. Astone, D. Babusci, M. Bassan, P. Carelli, G. Cavallari, E. Coccia, C. Cosmelli, S. D’Antonio, V. Fafone, A. C. Fauth, et al., Phys. Rev. Lett. 91, 111101 (2003b).
 M. Bignotto, M. Bonaldi, M. Cerdonio, L. Conti, F. Penasa, G. Prodi, G. Soranzo, L. Taffarello, and J. Zendri, Review of Scientific Instruments 76, 084502 (2005).
 A. Vinante, R. Mezzena, G. A. Prodi, S. Vitale, M. Cerdonio, P. Falferi, and M. Bonaldi, Applied Physics Letters 79, 2597 (2001).
 L. Baggio, M. Bignotto, M. Bonaldi, M. Cerdonio, L. Conti, P. Falferi, N. Liguori, A. Marin, R. Mezzena, A. Ortolan, et al., Physical Review Letters 94, 241101 (2005).
 http :// igec.lnl.infn.it/ fundamentals/ protocol/ igecprotocol2000vdec4.pdf.
 A. Papoulis, Probability, Random Variables and Stochastic Processes (McGraw Hill, New York, 1984), p. 115, 2nd ed.
 G. Feldman and R. Cousins, Physical Review D 57, 3873 (1998).