Welcome to Scribd, the world's digital library. Read, publish, and share books and documents. See more
Download
Standard view
Full view
of .
Look up keyword
Like this
0Activity
0 of .
Results for:
No results containing your search query
P. 1
1401.7038

1401.7038

Ratings: (0)|Views: 0|Likes:
Published by crocoali
aaa
aaa

More info:

Categories:Types, Brochures
Published by: crocoali on Jan 29, 2014
Copyright:Attribution Non-commercial

Availability:

Read on Scribd mobile: iPhone, iPad and Android.
download as PDF, TXT or read online from Scribd
See more
See less

01/29/2014

pdf

text

original

 
A sparse representation of gravitational waves from precessing compact binaries
Jonathan Blackman,
1
Bela Szilagyi,
1
Chad R.Galley,
1
and Manuel Tiglio
2,1
1
Theoretical Astrophysics, California Institute of Technology, Pasadena, CA, 91125, USA
2
Center for Scientific Computation and Mathematical Modeling, and Joint Space Sciences Institute, Maryland Center  for Fundamental Physics, Department of Physics, University of Maryland, College Park, MD 20742, USA
Many relevant applications in gravitational wave physics share a significant common problem:the seven-dimensional parameter space of gravitational waveforms from precessing compact binaryinspirals and coalescences is large enough to prohibit covering the space of waveforms with suffi-cient density. We find that by using the reduced basis method together with a parametrization of waveforms based on their phase and precession, we can construct ultra-compact yet high-accuracyrepresentations of this large space. As a demonstration, we show that less than 100 judiciouslychosen precessing inspiral waveforms are needed for 200 cycles, mass ratios from 1 to 10 and spinmagnitudes
 
0
.
9. In fact, using only the first 10 reduced basis waveforms yields a maximum mis-match of 0
.
016 over the whole range of considered parameters. We test whether the parametersselected from the inspiral regime result in an accurate reduced basis when including merger andringdown; we find that this is indeed the case in the context of a non-precessing effective-one-bodymodel. This evidence suggests that as few as
 ∼
100 numerical simulations of binary black hole coa-lescences may accurately represent the seven-dimensional parameter space of precession waveformsfor the considered ranges.
Gravitational radiation produced by stellar-mass com-pact binaries of neutron stars and/or black holes are ex-pected to be the main signals detected by the advancedgeneration of gravitational wave detectors [14]. Detect- ing these signals and estimating the parameters of theirsources require the ability to sufficiently sample the spaceof precessing compact binary waveforms. A compactbinary intrinsically depends on its mass ratio and thespin angular momentum components of each body, whichforms a 7-dimensional space for gravitational waveforms[5].Much progress has been made in sampling the subspaceof non-spinning compact binary waveforms over the lastdecade. However, many relevant applications, from dataanalysis for gravitational wave searches and parameterestimation studies to numerical relativity simulations of binary black hole coalescences, face a common challenge.In particular, going from the non-spinning subspace tothe full 7d space naively requires prohibitively more sam-ples for the same coverage simply because the volumeof the space grows exponentially with dimension. As aresult, the general consensus in the gravitational wavecommunity is that the computational complexity asso-ciated with building template banks for matched-filtersearches, with making parameter estimation studies, andwith modeling precessing compact binaries by expensivenumerical simulations is intractable (e.g., see [6]). Thisphenomenon entails what is called the
 curse of dimen-sionality 
 [7].In this paper, we show how to beat the curse of dimen-sionality for precessing compact binary inspirals. We findthat only 50 judiciously chosen waveforms are neededto represent the entire 7
d
 space with an accuracy bet-ter than 10
7
for 200 cycles, mass ratios
 
 
 [1
,
10], di-mensionless spin magnitudes
 
 χ
1
,
2
 0
.
9, and through
 = 8 spherical harmonic modes. Using only the first10 of these select waveforms yields a maximum repre-sentation error
 
 1%. The results of this paper suggestthat for any given parameter range a remarkably smallnumber of numerical relativity simulations of precessingbinary black holes, if judiciously chosen, are sufficientto
 accurately represent any other precession waveform in that range 
. We expect these results to be useful also forgravitational wave matched-filter searches and parameterestimation studies for compact binary coalescences.
Beating the curse of dimensionality 
.– Previous studieshave shown that non-precession subspaces of the full 7dwaveforms space
 
 can be represented by linear spacesspanned by a relatively compact set of inspiral [8, 9], ringdown [10], and inspiral-merger-ringdown (IMR) [11] waveforms, which form a
 reduced basis 
 (RB). The re-duced basis waveforms are found by training a
 greedy algorithm 
 [12, 13] to learn from a given discretization of 
 
 which are the most relevant waveforms for repre-senting elements of 
 W 
 with regard to a given error mea-sure. See [8] for more details. The number of RB wave-forms for non-precessing inspirals hardly grows from twoto four parameter dimensions thereby explicitly beatingthe curse of dimensionality [9]. Of further interest is that
precession 
 subspaces of 
 
 carry significant redundancyand are amenable to dimensional reduction as found in[14]. Those results strongly suggest that one may beatthe curse of dimensionality in the full 7d waveform space.In this paper, we outline how to construct a very com-pact but highly accurate RB of precession waveforms. Weconsider the following specifications on the 7d waveformspace:
 ∈
[1
,
10]
 ,
 
 χ
1
,
2
[0
,
0
.
9]
 ,
 200 cycles
.
 (1)where
 
 =
 m
1
/m
2
 
 1. These were chosen based par-tially on practical limitations of binary black hole simu-lations. However, the general message of this paper does
  a  r   X   i  v  :   1   4   0   1 .   7   0   3   8  v   1   [  g  r  -  q  c   ]   2   7   J  a  n   2   0   1   4
 
2not depend on our choice.
Key ingredients 
.–Our construction of a very compact (orsparse) reduced basis representation of precession wave-forms depends on several key steps [15]: (1) A random-ized resampling strategy [16] for training the greedy algo-rithm on the 7d waveform space; (2) A frame that rotateswith the binary’s precession; and (3) A physically moti-vated parametrization of precession waveforms.The first key ingredient is a modification of the stan-dard greedy algorithm [8]. In its simplest inception, thegreedy algorithm learns which waveforms can linearlyspan the space of interest in a nearly optimal way [12, 13],starting from a sufficiently dense set of waveforms calleda training set or space. However, the curse of dimension-ality prevents us from sampling the waveform space withsufficient coverage to build a reliable training set. Toovercome this, we randomly resample the 7d space froma uniform distribution using a fixed number
 
 of wave-forms at each iteration of the greedy algorithm. Thesewaveforms constitute the training set at the current itera-tion. Because the 7d space is resampled at each iterationby different waveforms, the maximum error from pro- jecting waveforms onto the current basis at the
 j
th
stepis actually measuring this error over an effective trainingset with
 j
×
 randomly distributed waveforms. As moreiterations are made, more of the 7d space is sampled andthe more accurate the RB becomes. This is a simple im-plementation of more powerful techniques introduced inRef. [16].For our studies, we randomly and uniformly resampled
 
 36
,
000 waveforms at each iteration of the greedyalgorithm. We began our studies with small
 
 and in-creased each sample size up to
 K 
 = 36
,
000, for which thetotal number of RB waveforms was robust and indepen-dent of 
 
. The largest training set used in our studiesincluded more than 3
×
10
6
randomly selected waveforms.The second key ingredient is to work in the binary’sprecessing frame instead of the usual inertial one. Specifi-cally, we generate post-Newtonian (PN) precession wave-forms in the time-domain using the
 minimally rotating 
frame of Refs. [17, 18]. In this frame, a precession wave- form appears qualitatively similar to waveforms fromnon-spinning binaries in their inertial frame [6, 1822]. In the minimally rotating frame, waveforms have a weakerdependence on parameters than they do in the inertialframe. The rotation involved in going from the mini-mally rotating frame to the inertial one and vice versacan be accounted for by any convenient representation of the
 SO
(3) group.The third key ingredient, and perhaps the most crucial,is that we choose to parametrize precession waveformsby their phase instead of by time or frequency. To moti-vate this choice we momentarily consider the frequency-domain waveform (in the stationary phase approxima-tion) for a non-spinning binary inspiral at leading order(“0PN”) in the PN approximation,
h
(
;
M
) =
 A
M
5
/
6
7
/
6
e
i
Φ
0
(
;
M
)
,
 (2)where
 M
 =
 Mν 
3
/
5
is the chirp mass,
 
 is the totalmass,
 ν 
 is the symmetric mass ratio,
 A
 is a constantindependent of the binary’s intrinsic parameters, andΦ
0
(
;
M
)
 3128
π
M
5
/
3
.
 (3)Reparametrizing (2) by its phase, now taken as the inde-pendent variable, gives
(
ϕ
;
M
)
h
(
(
ϕ
);
M
) =
 A
M
2
ϕ
7
/
10
e
(4)with
 A
=
 
7
/
6
(128
/
3)
7
/
10
and
 
(
ϕ
) from solvingΦ
0
(
=
) =
 ϕ
. In this phase-domain, all waveformsare proportional to each other, thus constituting a 1dspace. In fact, performing the greedy algorithm analyti-cally (versus numerically) returns a single basis waveformthat
 exactly 
 represents all such waveforms in the contin-uum. This is the intrinsic dimensionality of the problemas has long been known because 0PN waveforms onlydepend on the chirp mass. Therefore, a single reducedbasis waveform spans the whole 0PN waveform space. Toclose the system, we also need to represent the mappingbetween the phase and frequency domains,
(
ϕ
;
M
) = 1
π
M
128
ϕ
3
3
/
5
,
 (5)using a separate basis. As we see again, the frequenciesfor different chirp masses are all proportional to eachother. Therefore, any 0PN waveform, as a function of frequency, is represented by
 one 
 reduced basis waveformthrough the non-linear transformation in (3).For the sake of comparison, we implemented a stan-dard greedy algorithm following [8] using 0PN waveformsparametrized by frequency (not phase) for binaries witha fixed total mass and with mass ratios and number of cycles as listed in (1). We found that 152 RB wave-forms are required to reach numerical round-off errorsin representing any waveform in this 1d space. Even toreach an error of about 1% requires 138 RB waveforms.Therefore, using the phase parametrization results in asingle RB waveform for
 exact 
 representation whereas afrequency parametrization can yield a much larger RBfor
 approximate 
 representation.Part of the reason why using waveforms in the phasedomain (or
 ϕ
-domain) is advantageous is because the os-cillations in two waveforms
 always 
 cancel in the scalarproduct used to measure the projection error onto theRB in the greedy algorithm,
M
1
,
M
2
ϕ
 ≡
 
 ϕ
max
ϕ
min
(
ϕ
;
M
1
)
(
ϕ
;
M
2
)
.
 (6)For 0PN waveforms this results in a very smooth depen-
 
3dence on the chirp masses since (6) is
 M
21
M
22
. Simi-larly, the waveform frequency as a function of phase (5)has a very smooth dependence on them as well.Higher PN orders include more physics, such as thenonlinearity of general relativity and spin-orbit, spin1-spin2, and self-spin interactions, that depend on all 7 in-trinsic parameters. These contributions add more struc-ture to the waveforms but only weakly depend on theparameters. This is especially true in the
 ϕ
-domain and,as discussed below, we also find this holds through themerger and ringdown phases where the PN expansionparameter is no longer small. Consequently, the inversefunction
 
(
ϕ
) (or
 
(
ϕ
) if in the time domain) retainsthe weak dependence on intrinsic parameters. As thereis thus only ever a weak parameter dependence, one mayexpect to find only a relatively small number
 
 of RBwaveforms, possibly as few as
 N 
 =
 O
(
d
).
Method outline 
.–In this paper, we use 3.5PN precess-ing inspiral waveforms. We solve the PN equations (seeRef. [23] and references therein) using the approach of Refs. [17, 18] where the waveforms themselves are solved in a frame that minimizes the binary’s precession, alongwith a rotation operator represented by unit quater-nions to track this frame relative to the fiducial inertialframe. All waveforms in this minimally rotating frameare normalized to unity, and the initial orbital phasesare aligned. It is convenient to decompose the waveforminto spin-weighted spherical harmonic modes [24] charac-terized by (
,m
). We use the phase associated with the(
,m
) = (2
,
2) mode to parametrize the waveform butother choices are possible. A precession waveform
 h
(
t
)in the inertial frame is thus decomposed in the followingway,
h
(
t
)
{
m
(
ϕ
)
}
,
(
ϕ
)
,Q
(
ϕ
)
 (7)where
 
m
 is a spin-weighted spherical harmonic modein the minimally rotating frame,
 
(
ϕ
) is the functionrelating the (2
,
2) phase to the time coordinate, and
 Q
 isthe unit quaternion describing the rotation back to theinertial frame. We take into account all modes up to
 
 =8 and cut all waveforms off at a dimensionless frequencyof 0
.
2 in the (2
,
2) modes. Finally, all waveforms contain200 wave cycles.We build a RB for each component in the decomposi-tion (7). It is natural to use the scalar product in (6) for the
 T 
 and
 Q
 functions but to integrate the minimally ro-tating waveform over the 2-sphere so that, upon using theorthogonality of the spin-weighted spherical harmonics,
λ
1
,
λ
2
ϕ
 ≡
,m
 
 ϕ
max
ϕ
min
m
(
ϕ
;
λ
1
)
m
(
ϕ
;
λ
2
)
,
 (8)where
 λ
i
 is a tuple of parameter values. Executing agreedy algorithm on each component in (7) would resultin a selection of parameter values that are different foreach element. In order to choose the same parameters
10    0    10    1   10    2   Basissize   10    -10    10    -8    10    -6    10    -4    10    -2   10    0    10    2   10    4    10    6    
      M   a   x    p   r   o  j   e   c   t  i   o    n   e   r   r   o   r
7dPN,
ϕ     
D    Time   Waveform     Quaternion    010203040506070    Basissize   7dPN,
ϕ     
D    7dchecks,TD    3dPN,
ϕ     
D    3dEOB,TD    
FIG. 1.
 Left
: Maximum
 ϕ
-domain (
ϕ
D) projection error(red) from (9) for 7d post-Newtonian precession waveformsversus basis size. The contributions from the time function(dotted), waveform in the minimally rotating frame (dashed),and quaternion (solid) are also shown.
 Right
: Maximumtime-domain, inertial frame mismatches from (10) for 10
7
ran-domly selected waveforms (+) using the first 10, 20, and 50reduced basis waveforms. Also plotted are
 ϕ
-domain projec-tion errors for non-precessing PN waveforms (dashed) andthe time-domain (TD) projection errors from using the latterparameter values selected by the greedy algorithm to repre-sent EOB waveforms (solid black), which additionally includemerger and ringdown phases.
for all three reduced bases, we define a total projectionerror
 
ϕ
 through,
ϕ
(
λ
)
8
×
10
6
δT 
λ
2
ϕ
 + 0
.
5
δH 
λ
2
ϕ
 + 0
.
0031
δQ
λ
2
ϕ
(9)so as to receive approximately equal contributions fromeach component. Here,
 λ
 = (
q, χ
1
, χ
2
) is a tuple of 7dparameter values,
 δX 
λ
 =
 X 
λ
X
[
λ
] with
 X 
 one of theelements in (7), and
 
X
 is the projection operator ontothe basis for
 
. The numerical coefficients are fixed togive approximately equal contributions to the mismatchin the time domain and inertial frame in the case of smallrandom perturbations. Binaries with periods near 200
lead to a small coefficient for the time function.
Results for precessing binary inspirals 
.–We implementeda greedy algorithm using the three key ingredients dis-cussed above to find RB representations for the space of precession waveforms for the ranges given in (1). Theleft panel of Fig. 1 shows the maximum of the total pro- jection error (9) found at each iteration of the greedyalgorithm. We observe a power-law decay with exponent
8. The total error is not monotonically decreasingbecause of the constant resampling at each iteration. Weobserve that the maximum normalized projection errorover the training set is 10
2
using 10 basis waveformsand
 ≈
4
×
10
8
for 50. Also shown are the contributionsto the total error from the projections onto the basis of 

You're Reading a Free Preview

Download
scribd
/*********** DO NOT ALTER ANYTHING BELOW THIS LINE ! ************/ var s_code=s.t();if(s_code)document.write(s_code)//-->