Postprocessed time-delay interferometry for LISA

Transcription

1 PHYSICAL REVIEW D, VOLUME 70, (R) Postrocessed time-delay interferometry for LISA D. A. Shaddock,* B. Ware, R. E. Sero, and M. Vallisneri Jet Proulsion Laboratory, California Institute of Technology, Pasadena, California 91109, USA (Received 28 June 2004; ublished 27 October 2004) High-recision interolation of LISA hase measurements allows signal reconstruction and formulation of time-delay interferometry (TDI) combinations to be conducted in ostrocessing. The reconstruction is based on hase measurements made at aroximately 10 Hz (for a 1 Hz signal bandwidth) at regular intervals indeendent of the TDI delay times. Interolation introduces an error less than with continuous data segments as short as 2 s in duration. The 10 Hz samling rate reresents an increase from the 2 Hz samling rate needed for the original imlementation of TDI. The advantages of this technique include increased flexibility of the data analysis and significantly simlified hardware. DOI: /PhysRevD I. INTRODUCTION The Laser Interferometer Sace Antenna (LISA) is a mission to detect gravitational waves in the frequency band from 0.1 mhz to 1 Hz. The LISA constellation consists of three sacecraft flying in a heliocentric, Earth-trailing orbit, with searations of L m. Each sacecraft contains two roof masses that are shielded from external disturbances. To detect a assing gravitational wave, the change in searation L of the roof masses in different sacecraft must be monitored with a recision of L=L & = Hz using laser interferometry. This fractional length stability is far better than the fractional frequency stability of the laser source, which is exected to be = * = Hz. Degradation in sensitivity due to laser frequency noise could be avoided by oerating the constellation as a Michelson interferometer with equal arm lengths. Unfortunately, the orbital dynamics of the constellation make it imracticable to equalize the LISA arm lengths accurately enough to cancel the excess frequency noise. Time-delay interferometry (TDI) [1] is a technique to remove the otherwise overwhelming laser frequency fluctuations. TDI cancels laser frequency noise by combining hase measurements made at different times. The required timing of the measurements is set by the light travel times between the LISA sacecraft, and it must be accurate to 100 ns to meet the laser frequency noise suression requirements [2]. One obvious method to achieve this timing accuracy is to measure the hase with a 10 MHz samling frequency. Selecting the nearest-neighbor samles would then rovide the requisite 100 ns timing resolution. This aroach, however, would require data to be transmitted between sacecraft or back to Earth at the rate of aroximately 10 9 bits=s. The current design for TDI is to samle the hase at a much lower data rate, in the range of 2 to 10 Hz, with 100 ns accuracy triggering of the hasemeters [2,3]. *Electronic address: Daniel.Shaddock@jl.nasa.gov PACS numbers: Nn, Ly, Ym This aroach oses a number of technical challenges. To ensure a timing accuracy of 100 ns, the absolute lengths of the arms must be known to an accuracy of 30 m when the measurement is made. For some TDI combinations, each sacecraft must have knowledge of its nonadjacent arm s length. Also, the clocks on different sacecraft must be synchronized at the 100 ns level. Errors in arm length knowledge or clock synchronization would lead to an irreversible corrution of the TDI combinations. An alternative aroach is to samle the hase with a low rate at equally saced times, and to reconstruct the hase at intermediate times by interolation. Interolation must be imlemented with excetional accuracy for effective cancellation of laser frequency noise by subsequent TDI rocessing. Tinto and colleagues [2] examined one ossible method of interolation and found that months of uninterruted data around the time of interest are needed to achieve the necessary accuracy. This imlies that months of data would be unusable at the beginning and end of a measurement, and levies extreme requirements on instrument reliability and oerating duty cycle. The interolation technique was deemed infeasible, and the triggered measurement aroach was adoted. In this article, we demonstrate that interolation is feasible and that it can roduce the required accuracy with less than 2 s of data. We discuss the significant simlification in the design and oeration of the LISA mission resulting from this change. The method is based on fractional-delay filtering [4], a mature technique in digital signal rocessing. II. INTERPOLATION BY FRACTIONAL-DELAY FILTERING We secify that the interolation error be less than cycles= Hz for frequency comonents from 1 mhz to 1 Hz. This noise level is aroximately a factor of 10 below the hase noise contribution of shot noise. Below 1 mhz the requirement is relaxed, as the 1=f 2 roof mass dislacement noise dominates shot noise and a larger =2004=70(8)=081101(5)$ The American Physical Society

2 SHADDOCK, WARE, SPERO, AND VALLISNERI PHYSICAL REVIEW D interolation error can be tolerated. Assuming that the laser frequency noise roduces aroximately 100 cycles= Hz cycles= Hz at the hasemeter outut, interolation must have a fractional error of less than for frequency comonents in the 1 mhz to 1 Hz range. We assume that the LISA hase measurements will be recorded with a f s 10 Hz samling rate. The samling rate must be high enough to accurately reroduce the hase information in the LISA signal band, and to avoid adding noise from aliasing of higher-frequency hase noise, at the 10 6 level. Moreover, the erformance of the interolation schemes considered below imroves with oversamling. Ultimately, the samling rate will be determined by filtering requirements on the hasemeter and by the availability and cost of telemetry bandwidth to Earth. A. Perfect interolation and fractional-delay filters Interolation is the rocess of reconstructing the amlitude of a regularly samled signal between samles. Shannon [5] roved that a bandlimited signal samled at a sufficiently high frequency can be reconstructed erfectly by convolving the discrete time series with a continuous sine cardinal function sinc f s t sin f s t = f s t. Sinc interolation can also be viewed as alying an acausal finite-imulse-resonse (FIR) filter to the samled time series. The filter kernel (imulse resonse) is a samled version of the sinc function. In effect, instead of interolating the signal we interolate the filter kernel. As the sinc function is a known analytic function, the filter kernel can be interolated with arbitrary accuracy simly by time shifting the argument of the sinc. In general, the interolated signal s n D is the discrete convolution of the original signal s n with the shifted kernel: s n D s n h n D ; (1) where n is the samling index, D is the delay in samles 1 2 D<1 2, and h n is the filter kernel. For sinc interolation, h n D sinc n D. With zero delay (D 0), sinc interolation corresonds to a FIR filter with delta function imulse resonse [see Fig. 1(a)], since for integer n sinc n n0, where nk is the Kronecker delta function. If D 0, we obtain a FIR filter with a nondelta imulse resonse [see Fig. 1(b)], which has the effect of alying the fractional delay D to the original time series. Errors in fractionaldelay filtering are caused by the finite-length aroximation of the infinitely long delayed-sinc filter. B. Truncated-sinc fractional-delay filters The simlest finite-length aroximation to the ideal delayed-sinc filter is obtained by truncating the kernel. FIG. 1 (color online). Sinc FIR filter kernel values (circles) with delay set to (a) D 0, and (b) D 0:3. Filtering by a truncated-sinc of kernel length N can be written as s N n D for odd N : N 1 =2 X k N 1 =2 s n k sinc D k In the following discussion we restrict ourselves to filters where N is odd for simlicity. Although for a given filter order N the truncated-sinc is otimal in a least-squares sense [4], its frequency resonse is far from ideal (unity magnitude), exhibiting significant rile even at low frequencies. This is unaccetable for TDI, where very high fidelity is required in the 1 mhz to 1 Hz measurement band. In fact, Ref. [2] showed that truncated-sinc interolation becomes sufficiently accurate only for very large N. Figure shows the interolation error versus N for several filters, including the truncated-sinc. The interolation error " is defined as the maximum difference of the filter s frequency resonse and the ideal frequency resonse, e i2 fd=f s for frequencies between 1 mhz and 1 Hz: (2) " max jh f e i2 fd=f s j1 mhz f 1Hz ; (3) where H f is the Fourier transform of h n. We used D 0:5, which is exected from theory to be the worst case. With truncated-sinc interolation, " 1=N. Samling at 10 Hz, this filter would require a kernel almost four months long, N * 10 8, to achieve "<10 8. This means that two months of data at the beginning and end of each measurement eriod would be unusable. C. Windowed-sinc fractional-delay filters The rile in the frequency resonse of the truncatedsinc filter can be significantly reduced by windowing the filter,

3 POST-PROCESSED TIME-DELAY INTERFEROMETRY FOR LISA PHYSICAL REVIEW D s N n D N 1 =2 X k N 1 =2 s n k w k sinc D k ; (4) where the window w k goes smoothly to zero for k N 1 =2, so that the end oints are taered to zero instead of abrutly truncated. Of several conventional windows [6] tested, the Blackman function n 2 n w b n 0:42 0:5 cos 0:08 cos (5) N 1 N 1 was best, roducing "<10 8 for N 345 (see Fig. 2) corresonding to a loss of 14.4 s of data at the beginning and end of each measurement eriod. In comarison, the TDI combinations need several L=c arm travel times, or at least 65 s, to gather enough data to cancel laser noise. As seen in Fig. 2, " 1=N 3 for Blackman windowed-sinc filters. One simle modification to the Blackman windowedsinc filter kernel is to aly the Blackman function more than once. Our tests showed that using w 3 b n (alying the Blackman three times) roduced "<10 8 for N 21, corresonding to a loss of 1.1 s of data at the beginning and end of each measurement eriod. D. Lagrange filter A more accurate filter at low frequencies can be found by requiring a maximally flat frequency resonse near f 0 [4]. This filter kernel is equal to the Lagrange olynomial [7] N 1 =2 h L n Y t D k n k ; (6) k N 1 =2 k n where t D N 1 =2 D. The Lagrange filter can be FIG. 2 (color online). Comarison of interolation error for four interolation methods; truncated-sinc filter kernel, Blackman windowed-sinc filter kernel, Blackman 3 windowedsinc filter kernel, and Lagrange interolation. comared to filters in Sec. II C by exressing its kernel as a windowed-sinc function [Eq. (4)], with the window w L n N t D N 1 ; (7) sin t D N n N 1 =2 where the binomial coefficient is extended to noninteger arguments by the generalized factorial function (Gamma function) [8]. The erformance of the Lagrange filter for this alication is excellent, meeting the requirements with N 15 (1.5 s) as shown in Fig. 2. Although Lagrange interolation is known to roduce large surious oscillations at the ends of the interolation interval, this does not occur when the kernel is fully immersed in the signal. By acceting only data where the kernel is comletely immersed, we obtain excellent erformance at the exense of losing N=2 oints at the beginning and end of each measurement eriod. Lagrange interolation is related also to the Thiran infinite-imulse-resonse fractional-delay filter [4], which has nominally flat frequency resonse. The erformance of the Thiran filter in our alication is comarable to the Lagrange windowed-sinc FIR of the same order; but the latter is favored on grounds of simlicity, esecially for time-deendent delays. The filters were tested both by interolating known analytic functions and interolating bandlimited white noise. The bandlimited noise was generated with a samling rate of 10 MHz and resamled at times t n=f s, f s 10 Hz. The 10 Hz signal was delayed by D samles and comared to the original 10 MHz signal resamled at times t D=f s. The test results agreed with the calculated fractional error shown in Fig. 2. E. Further tests We have characterized the error of fractional-delay filtering with fixed delays. In orbit, the delays will slowly vary due to the changing arm lengths, and so we also tested Lagrange filtering with varying delay. For this test, we generated a time series of white noise, bandlimited to 2.5 Hz and samled at 10 Hz; we then used Lagrange filters of increasing order to interolate the noise to the original samling times shifted by delays ranging linearly in time from D 0:5to D 0 (no delay), during a eriod of s. This arrangement aroximately simulates the slow variation in the LISA arm lengths (which determine the TDI-mandated delays). Figure 3 shows sectra of the interolation error, along with the sectrum of the original white noise. The required interolation accuracy is achieved at all frequencies in the measurement band for N 16 (window length of 1.6 s). If the requirements become more stringent, for examle due to increased laser frequency noise, it is easy to imrove the erformance to the desired level by increasing the length of the filter kernel. By setting the filter

4 SHADDOCK, WARE, SPERO, AND VALLISNERI PHYSICAL REVIEW D FIG. 4. Real-time TDI, in which the delays must be known at the time of measurement, and the TDI combinations are comuted before transonding to Earth. TDI combination X t shown [Eq. (8)]. FIG. 3. Amlitude sectral density of interolation error for Lagrange filters, shown with the sectral density of the initial 2.5 Hz-bandlimited noise. Sectral density is estimated by a triangle-windowed, averaged eriodogram. to N 31 (3 s in length at 10 Hz), " can be reduced to III. IMPLICATIONS FOR LISA To illustrate the design simlifications enabled by highrecision interolation, consider the simlest TDI combination the Michelson combination X t [1]: X t s 21 t s 31 t s 21 t 2L 3 =c s 31 t 2L 2 =c ; (8) where s m1 is the hase measurement made at Sacecraft 1 of the light received from Sacecraft m, L n is the length of the arm oosite Sacecraft n, and we are assuming that the six LISA lasers are hase locked [2]. The first two terms reresent the otical hase of two arms of a Michelson interferometer, and the second two terms reresent the same quantity with the secified delays. The imlementation of TDI based on timed triggering [2], which we designate real-time TDI, calls for all four terms in Eq. (8) to be exlicitly measured, combined on-board, and sent to Earth. This is illustrated in Fig. 4 for the measurement of X t aboard Sacecraft 1. The blocks labeled d 21 and d 31 reresent the radiofrequency beat signals from the hotodiode oututs, containing laser frequency noise suerimosed on the gravitational wave signal. The Ranging block reresents the system that measures distances between sacecraft, and comutes the delays 2L 3 =c and 2L 2 =c required to assemble X t. The telemetry inuts to the ranging blocks (not shown) contain ranging data measured on Sacecraft 1 and Sacecraft 2. Phase measurements are made by the hasemeter (PM) blocks. In Fig. 4, the hase signals are delayed electronically by the Delay blocks, which imlement a variable delay as controlled by the ranging system. Equivalently, identical hase signals can be fed to two PM blocks, and the timing of the hase measurements can be set by adjustable triggers from the ranging system oututs. The final outut X t is free of laser frequency noise for fixed arm lengths. For time-deendent arm lengths, we exect [9] that velocity-correcting or second generation TDI combinations will be required. They have roughly double the measurements of length-correcting or first generation TDI combinations such as that shown in Fig. 4. As we have demonstrated in this aer, the delayed hase measurements required for TDI can alternatively be inferred by interolation from an equally saced sequence at a relatively low rate. This imlementation is referred to as ost-rocessed TDI and is shown in Fig. 5. The elimination of the need for ranging knowledge at the time of measurement simlifies the imlementation, as is evident comaring Figs. 4 and 5. Ranging information will still be needed as inut to signal reconstruction, but it can be transmitted to Earth indeendently of the hasemeter signals. Alternatively, ost-rocessed TDI can be imlemented without exlicit ranging by determining the delays using autocorrelations [3], or by adjusting the delays in ostrocessing for minimum sensitivity to laser frequency variations. With ost-rocessed TDI it is no longer necessary to synchronize the clocks on different sacecraft. Clock synchronization error can be corrected FIG. 5. X t. Post-rocessed imlementation of TDI combination

5 POST-PROCESSED TIME-DELAY INTERFEROMETRY FOR LISA PHYSICAL REVIEW D simly by time shifting the data in ostrocessing. The decouling of ranging from hase measurements and the removal of the need for clock synchronization allows reduction or ossibly elimination of intersacecraft communications. Post-rocessed TDI also allows comlete flexibility in combining hasemeter signals. All the raw data are available for rocessing by any TDI algorithm, including ones not develoed until after the data are in hand. The delays can be adjusted to otimize the suression of laser frequency noise; by contrast, if there is an error in triggering in real-time TDI, noise is irrevocably added. Postrocessed TDI simlifies the hase measurement hardware, allowing all ossible TDI combinations to be constructed from one constant-rate hase measurement er hotodetector. A significant oerating cost of LISA will be telemetry to Earth of science data. Real-time TDI requires one data stream er TDI combination. Post-rocessed TDI requires one data stream er hasemeter; more recisely, er hasemeter that does not have its outut held fixed by a high-gain control system. We exect that ost-rocessed TDI will require fewer telemetry signals than real-time TDI, but this deends on details of hardware design and on data requirements that are currently under consideration. Other factors influencing overall telemetry costs are the data rate and number of bits er datum. The ultimate LISA data will have a signal bandwidth of 1 Hz and a dynamic range large enough to encomass both the largest exected gravitational wave signal (or erhas the largest instrumental effect) and shot-noise limited sensitivity. Each outut of the real-time TDI signal chain [X t for examle] will have essentially eliminated laser frequency noise before data are transonded to Earth, reducing the dynamic range requirement. A nominal datum size is 20 bits er samle. The data rate for each TDI combination is 2 samles=s, in keeing with the 1 Hz requirement and the Nyquist limit. In comarison, ost-rocessed TDI must transond large laser frequency noise suerimosed on the small gravitational wave signal. Larger dynamic range is required, erhas 30 bits er samle. The samle rate for ost-rocessed TDI is likely to be greater than 2 samles=s for two reasons. First, a larger samling frequency may be needed to rovide the more stringent antialiasing filtering needed when laser frequency noise is resent. Second, an oversamling factor may be needed for the interolation rocedure. The minimum samle rate imosed by ost-rocessed TDI is still under study; our initial estimate is 10 samles=s or less ( 5 times higher than needed for real-time TDI). If the hase measurements were bandlimited closer to 1 Hz, the samling rate could be reduced. Although the interolation algorithms resented here may not be otimal, they demonstrate the feasibility of ost-rocessed TDI. They serve as a roof of rincile, and rovide guidance for the design of LISAwith significant simlification in several resects over real-time TDI. The secific imlementation details of ost-rocessed TDI (such as the samle rate, number of bits er samle, and window arameters) will be refined in concert with the LISA hasemeter develoment. ACKNOWLEDGMENTS We thank John Armstrong for many useful discussions and for hel with tests of the interolation rocedure. M.V. was suorted by the LISA Mission Science Office at JPL. This research was erformed at the Jet Proulsion Laboratory, California Institute of Technology, under contract with the National Aeronautics and Sace Administration. [1] M. Tinto and J.W. Armstrong, Phys. Rev. D 59, (1999); J.W. Armstrong, F. B. Estabrook, and M. Tinto, Classical Quantum Gravity 18, 4059 (2001); M. Tinto, F. B. Estabrook, and J.W. Armstrong, Phys. Rev. D 65, (2002). [2] M. Tinto, D. A. Shaddock, J. Sylvestre, and J.W. Armstrong, Phys. Rev. D 67, (2003). [3] R.W. Hellings, Phys. Rev. D 64, (2001). [4] T. I. Laakso et al., IEEE Signal Process. Mag. 13, 30 (1996). [5] C. E. Shannon, in Proceedings of the Institution of Radio Engineers, 1949, Vol. 37, [6] S.W. Smith, The Scientist and Engineer s Guide to Digital Signal Processing (California Technical Publishing, San Diego, CA, 1997). [7] P. J. Kootsookos and R. C. Williamson, IEEE Trans. Circuits Syst. II Analog Digital Signal Processing 43, 269 (1996). [8] W. H. Press, B. P. Flannery, S. A. Teukolsky, and W. T. Vetterling, Numerical Recies in FORTRAN: The Art of Scientific Comuting (Cambridge University Press, Cambridge, England, 1992), 2nd ed., ; E.W. Weisstein, from MathWorld, htt://mathworld. wolfram.com/binomialcoefficient.html. [9] D. A. Shaddock, M. Tinto, F. B. Estabrook, and J.W. Armstrong, Phys. Rev. D 68, (2003)