Binary black hole spectroscopy Chris Van Den Broeck1,∗ and Anand S. Sengupta1,† 1School of Physics and Astronomy, Cardiff University, Queen’s Buildings, The Parade, Cardiff CF24 3AA, United Kingdom Abstract Westudyparameterestimationwithpost-Newtonian(PN)gravitationalwaveformsforthequasi- circular, adiabatic inspiral of spinning binary compact objects. In particular, the performance of amplitude-corrected waveforms is compared with that of the more commonly used restricted waveforms, in Advanced LIGO and EGO. With restricted waveforms, the properties of the source canonlybeextractedfromthephasing. Inthecaseofamplitude-correctedwaveforms,thespectrum encodes a wealth of additional information, which leads to dramatic improvements in parameter 7 estimation. At distances of 100 Mpc, the full PN waveforms allow for high-accuracy parameter 0 ∼ 0 extraction for total mass up to several hundred solar masses, while with the restricted ones the 2 errors are steep functions of mass, and accurate parameter estimation is only possiblefor relatively n light stellar mass binaries. At the low-mass end, the inclusion of amplitude corrections reduces the a J error on the time of coalescence by an order of magnitude in Advanced LIGO and a factor of 5 5 in EGO compared to the restricted waveforms; at higher masses these differences are much larger. 1 The individual component masses, which are very poorly determined with restricted waveforms, 2 become measurable with high accuracy if amplitude-corrected waveforms are used, with errors as v low as afewpercentinAdvancedLIGO andafew tenths ofapercentinEGO.Theusualspin-orbit 6 parameter β is also poorly determined with restricted waveforms (except for low-mass systems in 2 1 EGO), but the full waveforms give errors that are small compared to the largest possible value 0 consistent withthe Kerrbound. Thissuggests a way of findingoutif one or bothof thecomponent 1 6 objects violate this bound. On the other hand, we find that the spin-spin parameter σ remains 0 poorly determined even when the full waveform is used. Generally, all errors have but a weak / c dependence on the magnitudes and orientations of the spins. We also briefly discuss the effect of q - amplitude corrections on parameter estimation in Initial LIGO. r g : PACS numbers: 04.25.Nx, 04.30.-w,04.80.Nn, 95.55.Ym v i X r a ∗Electronic address: [email protected] †Electronic address: [email protected] 1 I. INTRODUCTION Inspiralingcompact binaryobjects(neutronstarsand/orblackholes) areamongthemost promising sources for existing and planned interferometric gravitational wave observatories [1]. Efforts are in progress to detect their gravitational wave signals in LIGO, VIRGO, GEO600 and TAMA [2]. When the gravitational wave signal enters the bandwidth of these detectors, typically the orbits of the bodies will have approximately circularized, and the subsequent inspiral is described well by an adiabatic approximation in which the radii of the orbits are assumed not to change significantly over a single period [3]. In this regime the frequency and amplitude of the signal increase steadily in a “chirp” until a last stable orbit (LSO) is reached. The amplitude andphasing of signals arising from quasi-circular adiabatic inspiral have been computed to great accuracy using post-Newtonian (PN) methods (see [4] for a review), which lead to expressions that take the form of series expansions in the orbital velocity v. Apart from their obvious astrophysical importance, binary inspirals are potential “lab- oratories” for testing general relativity [5, 6, 7]. This is because at least in the adiabatic phase they are relatively “clean” and reasonably well-understood systems. However, it is important to find out how good a testing ground they offer. Precisely what information do the waveforms carry? Which parameters associated with the binary can be determined from them, and to what accuracy can this be achieved? In other words, to what extent will it be possible to reconstruct binary inspiral events from the gravitational radiation they emit? The most accurate waveforms currently available for the quasi-circular, adiabatic inspiral of compact objects are of order (v5) in amplitude [8] and (v7) in phase [9], at least in the O O case of non-spinning binaries; in the usual notation this corresponds to 2.5PN and 3.5PN orders, respectively. An efficient way of searching for inspiral signals in data is matched filtering [10], which involves a bank of templates. Post-Newtonian waveforms describing adiabatic inspiral are linear combinations of harmonics in the orbital phase. In most of the literature, simulated signals as well as detection templates are taken to be restricted post-Newtonian waveforms [11], which only have the dominant harmonic at twice the orbital phase, withPNcorrectionstothephasingbutnocorrectionstotheamplitude. Theeffectson signal-to-noise ratio (SNR) and redshift reach as a result of including amplitude corrections in templates and simulating signals have been investigated in [12, 13] in the context of both initial and advanced detectors, and it was found that the consequences are considerable. In initial detectors the use of restricted PN waveforms can lead to a significant overestimation of SNR. More interestingly, in advanced detectors the higher harmonics can give rise to considerable SNRs even when the dominant harmonic (and hence the restricted waveform) does not even enter the detector’s bandwidth. As a result, the mass range of such detectors increases dramatically with the inclusion of amplitude corrections in template waveforms. Also in the case of masses for which the dominant harmonic reaches the bandwidth in only a small frequency interval, taking into account all known harmonics greatly increases the SNR. Thus, for high stellar mass to intermediate mass binaries (100M . M . 1000M ), ⊙ ⊙ detection rates increase dramatically when amplitude corrections are taken into account. This also opens up the possibility of high-quality parameter estimation up to much higher masses than would be thought possible on the basis of the restricted PN approximation. More generally, to properly address the questions raised above, we will see that it is crucial to use the amplitude-corrected waveforms to highest available post-Newtonian order. With restricted waveforms h , almost any information one can extract about the source has 0 2 ˜ ˜ to come from the phasing, i.e. arg(h (f)), with h the Fourier transform of the time domain 0 0 waveform. In the case of amplitude-corrected waveforms h, on top of the information carried ˜ by arg(h(f)) one can exploit the intricate structure of the spectrum, i.e. the power per frequency bin as a function of frequency. As we shall see, the additional information carried by the spectrum of the full waveform has dramatic consequences for parameter estimation, whence the title of this paper. (The study of black hole ringdown in terms of quasi-normal modes is often also referred to as “spectroscopy” [14]; in the case of inspiral the spectrum is of course continuous.) Most of the research to date on parameter estimation with inclusion of additional har- monics has been of a preliminary nature. For Initial LIGO there is the work of Sintes and Vecchio [15], who used waveforms at 0.5PN order in amplitude and 2PN order in phase. Very recently, R¨over et al. [16] considered a network of initial detectors (the Initial LIGO detectors and VIRGO), taking the sources to be binary neutron star inspirals and using waveforms at 2PN in amplitude and 2.5PN in phase. For the proposed space-based detector LISA there have been studies by Sintes and Vecchio [17] (again at 0.5PN in amplitude and 2PN in phase) and by Hellings and Moore [18, 19] (2PN in both amplitude and phase). None of these works took the effects of spin into account. In this paper we focus on ad- vanced ground-based interferometric gravitational wave detectors. We will mostly consider parameter estimation with Advanced LIGO as well as a European third-generation detector that is currently under consideration [20]; the latter will tentatively be called the Euro- pean Gravitational-wave Observatory (EGO). The waveforms used will be of 2.5PN order in amplitude and phase, with inclusion of spin-related contributions up to 2PN in the phase. Up to 2.5PN order in amplitude, the waveforms in the two polarizations take the general form [8] 2Mη h = x H(0) +x1/2H(1/2) +xH(1) +x3/2H(3/2) +x2H(2) +x5/2H(5/2) (1.1) +,× r +,× +,× +,× +,× +,× +,× n o where r is the distance to the binary, M its total mass, and η the ratio of reduced mass to total mass. We have set G = c = 1, as we will do throughout this paper. The post- Newtonian expansion parameter is defined as x = (2πMF)2/3, with F(t) the instantaneous (p/2) orbital frequency. The coefficients H , p = 0,...,5, are linear combinations of various +,× harmonics with prefactors that depend on the inclination angle ι of the angular momentum of the binary with respect to the line of sight as well as on η; their explicit expressions can be found in [8]. The measured signal also depends on the polarization angle and the position in the sky through the detector’s beam pattern functions F : +,× h(t) = F h (t)+F h (t). (1.2) + + × × Note that for ground-based detectors, which are the ones we will be concerned with, it is reasonable to approximate F as being constant over the duration of the signal. They +,× depend on angles (θ,φ,ψ), where (θ,φ) determine sky position while ψ is the polarization angle. The signal (1.2) is a linear combination of harmonics of the orbital phase Ψ(t) with offsets ϕ : (k,m/2) N 5 h(t) = A (t)cos(kΨ(t)+ϕ ), (1.3) (k,m/2) (k,m/2) Xk=1mX=0 where k runs over harmonics while m/2 is PN order in amplitude. The coefficients A (k,m/2) are functions of (r,M,η,θ,φ,ψ,ι) multiplied by x(m+2)/2, with x depending on time through 3 F(t). The number of harmonics N depends on the PN order in amplitude; at 2.5PN one has N = 7. Spin-related effects will be taken into account in the following way. In the phasing, spin- orbit effects are included at 1.5PN and 2PN orders, and spin-spin effects at 2PN order, as in [21]. (The phasing formula with the inclusion of spin-orbit interactions up to 2.5PN has recently been computed [22], but the research we report on here was already well underway when these results appeared.) Wewill use the stationaryphase approximation to the Fourier transform [23], in which the coefficients of the harmonics are proportional to the inverse square root of the time derivative of the associated instantaneous frequency. We let spin- relatedeffectsenter theamplitudes throughthese“frequency sweeps”, atthesamerespective orders. Since we want to view the full waveform as a model for an actual signal as well as a search template, for consistency we take both amplitude and phase to be of 2.5PN order. This is a choice we make, motivated by conceptual considerations; there are no technical problems in taking the phasing up to 3.5PN. Both the expressions for the amplitudes in (1.3) and the expression for the phase can be truncated to a lower post-Newtonian order. Such truncated waveforms will be denoted as (p,q)PN, where p refers to the PN order of the amplitude and q that of the phase. Thus, the restricted PN waveform will be referred to as (0,2.5)PN and the full waveform as (2.5,2.5)PN. For the Initial LIGO detector, Sintes and Vecchio [15] discussed the improvement in error estimationgainedbyusingamplitude-correctedwaveformsbymeansofthecovariancematrix formalism [24]; in the present paper the same techniques will be employed. Concretely, the following questions will be addressed: (i) Using the best available waveforms, which parameters can be determined? (ii) How do the estimation accuracies improve as the PN order of the amplitude is increased? and (iii) What is the effect of the magnitudes and orientations of the components’ spins?1 This paper is structured as follows. In section II we first briefly review the covariance ma- trix formalism for parameter estimation. The Fourier transform of the amplitude-corrected waveforms is computed in the stationary phase approximation and its spectrum is discussed. We then apply the covariance matrix formalism to (p,2.5)PN waveforms with p 0.5. In ≥ section III we compare the performance of restricted and amplitude-corrected waveforms in EGO and Advanced LIGO for a variety of systems. We also study the improvement in error estimation by using waveforms of order (p,2.5)PN as p varies from zero to 2.5, and the dependence of errors on the spins of the component objects. Finally, we present results for some representative systems with Initial LIGO. Section IV provides a summary and conclusions. 1 Throughout this paper we will include a spin-orbit and a spin-spin parameter as variables in the error analysis so that even when spins are set to zero, the corresponding errors are never neglected. 4 II. AMPLITUDE-CORRECTED POST-NEWTONIAN WAVEFORMS AND PA- RAMETER ESTIMATION A. Overview of the covariance matrix formalism We begin by recalling some basic facts of the covariance matrix formalism for parameter estimation with matched filtering [24]. To search for signals buried in the noise, the data analyst uses families of waveforms h˜[θa](f), determined by a finite number of parameters θa that characterize the binary. The measured values of the parameters, θˆa, depend on the realization of the detector noise at the time the signal was present, so that in general they do not correspond to the “true” values θ¯a. For large SNR, the measurement errors ∆θa = θ¯a θˆa follow a multivariate Gaussian probability distribution [24], − P(∆θa) = P exp( Γ (θˆa)∆θb∆θc/2) (2.1) 0 bc − with P a normalization factor, and Γ (θa) is the Fisher information matrix. The latter 0 ab may be thought of as a metric on the differentiable manifold naturally defined by the family of waveforms h[θa] , and it is given by { } Γ = (h h ), (2.2) ab a b | where h ∂h/∂θa, and (. .) is the usual product defined by the one-sided noise power a ≡ | spectral density S (f): h fmax Re(x˜∗(f)y˜(f)) (x y) 4 df . (2.3) | ≡ Z S (f) fmin h The integration domain [f ,f ] is determined both by the detector and by the nature min max of the signal. As explained in [13] (and in subsection IIB below), it is reasonable to cut off the kth harmonic in the waveform at a frequency kf , where f may be viewed as the LSO LSO orbital frequency at the last stable orbit. Since the amplitude-corrected waveform contains 7 harmonics, in that case we will effectively have f = 7f , while for the restricted max LSO waveform f = 2f . As to the lower cut-off frequency, power spectral densities S (f) max LSO h tend to rise very quickly below a certain frequency f where they can be considered infinite s for all practical purposes; accordingly, we take f = f . min s The signal-to-noise ratio is also defined in terms of the product (2.3), ρ[h] (h h). (2.4) ≡ | p The inverse of the information matrix is the covariance matrix Σab, and one has Σab (Γ−1)ab = ∆θa∆θb (2.5) ≡ h i where . denotes the average with respect to the probability distribution P(∆θa) (2.1). h i The root-mean-square error in the measurement of the parameters is then σa (∆θa)2 = √Σaa, (2.6) ≡ h i p and with slight abuse of notation these will simply be denoted ∆θa in the rest of the paper. Finally, the correlation coefficients are defined by Σab cab . (2.7) ≡ √ΣaaΣbb 5 (Note that repeated indices in Eqs. (2.6) and (2.7) are not meant to be summed over.) By definition one has 1 c 1. If the correlation coefficient between two parameters is ab − ≤ ≤ close to 1 (resp. 1), they are strongly correlated (resp. anticorrelated), meaning that one of − themis redundant. Acorrelationcoefficient being closeto zero indicates thattheparameters are largely unrelated. Usually the Fisher information matrix, and hence errors and correlations, can not be calculated exactly because the integrals defining Γ can not be performed analytically. This ab is also the case here, and most of the results in this paper will refer to numerical calculations with the software package Mathematica. The matrix Σ = Γ−1 was obtained using the built- in matrix inversion routine. In that regard it is worth mentioning that problems can arise when numerically inverting large matrices such as the Fisher matrices encountered here; we will come back to this point in subsection IID below. The covariance formalism is only valid for sufficiently high SNR; at low SNR it tends to underestimate errors, as seen in Monte Carlo treatments involving large numbers of simulations with different realizations of the noise [25]. However, at distances of a few hundred Mpc the SNRs in Advanced LIGO and EGO will be rather high, as was made evident in [13]; see also subsection IIIA below. B. Amplitude-corrected PN waveforms The Fisher matrix (2.2) is defined in terms of (derivatives of) the Fourier transform of the waveform h[θa](t). We will find it convenient to use the well-known stationary phase approximation (SPA) [23] to the Fourier transform in order to compute the scalar products involved. The resulting waveforms have already been discussed in detail in [13]; here we only give a quick overview. As mentioned in the Introduction, the waveforms h[θa](t) take the form 7 h(t) = h(k)(t) (2.8) X k=1 where the h(k)(t) involve harmonics of the orbital phase with constant offsets: 5 h(k)(t) = A (t)cos(kΨ(t)+ϕ ). (2.9) (k,m/2) (k,m/2) mX=0 Theseharmonicsarefoundfrom(1.1)and(1.2); skyposition(θ,φ)andthepolarizationangle ψ enter the coefficients A andoffsets ϕ throughthebeampatternfunctions F : (k,m/2) (k,m/2) +,× 1 F (θ,φ,ψ) = 1+cos2(θ) cos(2φ)cos(2ψ) cos(θ)sin(2φ)sin(2ψ), + 2 − (cid:0) (cid:1) 1 F (θ,φ,ψ) = 1+cos2(θ) cos(2φ)sin(2ψ)+cos(θ)sin(2φ)cos(2ψ), (2.10) × 2 (cid:0) (cid:1) which in the case of ground-based detectors can be considered constant for the duration of the observed part of the signal. During the inspiral phase one has dlnA /dt 1 and kd2Ψ/dt2 (kdΨ/dt)2, in (k,s) | | ≪ | | ≪ which case we can use the stationary phase approximation [23] to the Fourier transform of 6 (2.9). For positive frequencies we have 5 A t(1f) 1 1 h˜(k)(f) m=0 (k,m/2) k exp i 2πft f kΨ t f ϕ π/4 ≃ P 2 kF˙ t 1(cid:0)f (cid:1) (cid:20) (cid:18) (cid:18)k (cid:19)− (cid:18) (cid:18)k (cid:19)(cid:19)− (k,m/2) − (cid:19)(cid:21) k q (cid:0) (cid:0) (cid:1)(cid:1) 5 A t(1f) e−iϕ(k,m/2) 1 = m=0 (k,m/2) k exp i 2πft π/4+kψ f . (2.11) P 2 kF˙(cid:0) t(1f(cid:1)) (cid:20) (cid:18) c − (cid:18)k (cid:19)(cid:19)(cid:21) k q (cid:0) (cid:1) A dot denotes derivation with respect to time and t is the coalescence time. c We recall that the function F(t) is the instantaneous orbital frequency. The function t(f) is defined implicitly by F(t(f)) = f. In the expressions (2.11) for the h˜(k) we take the “frequency sweep” to be [8, 9, 26] 96 743 11 F˙ = (2π F)11/3 1 + η (2πMF)2/3 +(4π β)(2πMF) 5π 2 M (cid:20) −(cid:18)336 4 (cid:19) − M 34103 13661 59 4159π 189π + + η + η2 +σ (2πMF)4/3 + η (2πMF)5/3 , (cid:18)18144 2016 18 (cid:19) −(cid:18) 672 8 (cid:19) (cid:21) (2.12) where = Mη3/5 isthechirpmass. Theparameterβ occurringat1.5PNorderencapsulates M the leading-order spin-orbit effects, while the parameter σ at 2PN order also encodes spin- spin effects; at 2.5PN order we neglect all spin-related effects. β and σ are given by [30]: 2 1 β = 113(m /M)2 +75η Lˆ χ¯ , (2.13) i i 12 · Xi=1 (cid:2) (cid:3) η ˆ ˆ σ = 247(χ¯ χ¯ )+721(L χ¯ )(L χ¯ ) , (2.14) 1 2 1 2 48 − · · · h i with χ¯ = S~ /m2 and S~ , i = 1,2, the spins of the binary’s components, while Lˆ is the unit i i i i vector in the direction of orbital angular momentum. To a reasonable approximation, β and σ may be considered constant [33]. The maximum possible value of σ for sub-extremal | | black holes is 79/32 2.5 and the maximum value of β is 113/12 9.4. ≃ | | ≃ To 2.5PN order, the phase ψ(f) is given by 5 3 ψ(f) = ψ + ψ (2πMf)i/3, (2.15) − c 256(π f)5/3 i M Xi=0 where ψ is the orbital phase at coalescence. We take the coefficients ψ to be [8, 9, 26] c i ψ = 1, 0 ψ = 0, 1 20 743 11 ψ = + η , 2 9 (cid:20)336 4 (cid:21) ψ = 4(4π β), 3 − − 3058673 5429 617 ψ = 10 + η+ η2 σ , 4 (cid:20)1016064 1008 144 − (cid:21) 38645 38645 f 65 f ψ = π + ln η 1+ln . (2.16) 5 (cid:20) 756 756 (cid:18)f (cid:19)− 9 (cid:18) (cid:18)f (cid:19)(cid:19)(cid:21) LSO LSO 7 The SPA for the (p,2.5)PN waveform is then Np 2p A t 1f e−iϕ(k,m/2) 1 h˜ (f) =  m=0 (k,m/2) k  exp i 2πft π/4+kψ f , SPA X P 2 kF˙(cid:0) (cid:0)t 1f(cid:1)(cid:1) (cid:20) (cid:18) c − (cid:18)k (cid:19)(cid:19)(cid:21) k=1  q k p (cid:0) (cid:0) (cid:1)(cid:1) (2.17) where N is the number of harmonics, and [.] denotes consistent truncation to pth post- p p Newtonian order (i.e., the “Newtonian” prefactor f−7/6 is taken outside and the remaining expression is Taylor-expanded in (2πMf)1/3 up to order (2πMf)5/3). For more explicit expressions we refer to [13]. In the time domain it is reasonable to cut off the waveform at a time corresponding to F(t) = f , where f is the orbital frequency at “last stable orbit”. For simplicity we LSO LSO use the expression corresponding to the extreme mass ratio: 1 f = , (2.18) LSO 63/22πM where M is the total mass of the system. In the frequency domain this roughly corresponds to terminating the harmonics at multiples of f ; in practice we multiply the truncated LSO kth harmonic by θ(kf f) where θ(x) is the usual Heaviside function (θ(x) = 1 if x > 0 LSO − and θ(x) = 0 otherwise) [13]. Finally, in [28, 29] the stationary phase approximation was compared with an alternative way of approximating the Fourier transform, the so-called fast Fourier transform (FFT), for initial detectors and relatively small masses. In that case the SPA and FFT differ by only a few percent, and the situation improves dramatically with increased detector sensitivity at low frequencies. We also note that the conditions for the applicability of the SPA as spelled out above are satisfied more easily with increasing k. C. Spectrum as observed in a detector Ourgoalistocomparethequalityofparameterestimationwithrestricted andamplitude- corrected waveforms. As we now explain, the extent to which the full waveform will perform better is determined by the information content of its spectrum. Consider the contribution to the SNR squared per logarithmic frequency bin, as a function of frequency. The squared signal-to-noise ratio of a waveform h in a detector with noise power spectral density S (f) h is fmax h(f) 2 ρ2[h] = 4 | | df Z S (f) fmin h fmax f h(f) 2 = 4 | | dln(f). (2.19) Z S (f) fmin h We now define the “observed spectrum” as f h(f) 2 (f) | | . (2.20) P ≡ S (f) h Thenameisapt; theobservedspectrumbearsadirectrelationshiptothewaysignalsareseen in a detector, which depends both on the waveform and on the sensitivity of the detector. In 8 [29] the quantity (f) was used (under another name) as a diagnostic to compare different P (restricted) waveform approximants. When using restricted waveforms h for parameter estimation, of necessity almost all 0 ˜ information one can extract about the source has to come from the phasing, arg(h (f)); 0 the spectrum does not play much of a role. The situation is very different when amplitude- ˜ corrected waveforms h are used. In that case we can exploit not only arg(h(f)); on top of that there is the wealth of information contained in h˜(f) 2, or more directly in the observed | | spectrum (f). P x 104 x 104 5 10 Full Full Restricted Restricted 4 8 (f)h3 (f)h 6 S S 2/ 2/ f |h(f)| 2 f |h(f)| 4 1 2 0 0 0 200 400 600 800 1000 1200 1400 20 40 60 80 100 120 140 f (Hz) f (Hz) Figure 1: The observed spectrum in Advanced LIGO for a (1,10)M system (left) and a ⊙ (10,100)M system(right). Distanceis100Mpcandangleswerechosenarbitrarilyasθ = φ= π/6, ⊙ ψ = π/4, ι = π/3. Spins are zero. InFig.1 we have plotted (f) forbothrestricted andamplitude-corrected waveforms, for P two different systems in Advanced LIGO.With a (1,10)M system, the restricted waveform ⊙ h will penetrate a detector’s bandwidth up to a relatively high frequency, as 2f = 0 LSO 1/(63/2πM) will be large. The higher harmonics that are present in the full waveform continue to still higher frequencies, but their contribution beyond 2f is suppressed by LSO the denominator S (f) in (2.20); Advanced LIGO has its peak sensitivity at about 215 Hz. h On the other hand, for a (10,100)M system, h will deposit power in a much smaller ⊙ 0 frequency interval, with an upper cut-off frequency much below the frequency where the detector is most sensitive. Therefore, for the restricted waveform, the quantity (f) will P be non-zero in only a small interval in which the denominator S (f) in (2.20) is large. By h contrast, the higher harmonics in the amplitude-corrected waveform h can penetrate the bandwidth up to frequencies where the detector is much more sensitive; note the distinctive features at 40 Hz, 60 Hz, 80 Hz, .... For both systems, (f) for the full waveform exhibits a great deal of structure due to P interferencebetweenthevariousharmonics. Asaconsequence, wemayexpecttheamplitude- corrected waveforms to allow for parameter extraction with significantly smaller errors than therestricted waveforms, withthedifference beingmuch morepronouncedforhighermasses. 9 D. Parameter estimation with amplitude-corrected waveforms: Overview Before going into details, we begin with an overview and discuss parameter estimation with the waveforms of the previous subsection in a qualitative way. For any family of waveforms of order (p,q)PN with 0.5 p < , the metric Γ has a ab ≤ ∞ coordinate singularity at η = 1/4. This can readily be inferred from the explicit expression for (0.5,q)PN waveforms in [15]: h blows up as η 1/4 and the same is true of Γ = ln(η) ln(η)b → (h h ). The problem does not occur if the usual coordinate ln(η) is replaced by δ, where ln(η) b | m m 1 2 δ − . (2.21) ≡ (cid:12)m +m (cid:12) (cid:12) 1 2(cid:12) (cid:12) (cid:12) (cid:12) (cid:12) (This new parameter was already introduced in [18], but there the motivation was left unclear.) For this reason, when considering amplitude-corrected waveforms we will use the set of parameters (θa) = (ln(1/r),ln( ),δ,t ,ψ ,cos(ι),cos(θ),φ,ψ,β,σ). (2.22) c c M Generally, not all parameters affecting the waveform can be extracted a posteriori, and this will also be the case here. As the PN order in amplitude is increased, one has the following: The restricted (p = 0) waveform, whose SPA takes the form • f−7/6 exp[i(2πft π/4+2ψ(f/2))], (2.23) c A − only allows for the determination of [26, 31] (ln( ),ln( ),δ,t ,ψ ,β,σ), (2.24) c c A M where the amplitude depends on r, , δ, and the four angular parameters. A M Going to 0.5PN in amplitude, two additional harmonics appear and it becomes possi- • ble in principle to discriminate between three variables appearing in the amplitudes: The distance r, the inclination angle ι, and e.g. the quotient F /F of the beam × + pattern functions [15]. Below we will instead include (ln(1/r),cos(ι),cos(θ)), as these parameters have a direct physical meaning. At 1PN in amplitude there are four harmonics in all so that one might expect either φ • or ψ to be an additional measurable parameter. However, whichever one chooses, the errors on (cos(θ),φ,ψ) tend to be extremely large and any of these parameters will be unmeasurable in practice. The same goes for the distance r. This remains true up to 2.5PN order in amplitude. Ideally one would like to work with the full 11 11 Fisher matrix based on all of the × variables in (2.22). However, inversion of such a large matrix tends to fail when using the built-in inversion routine of Mathematica. The reason is that the matrix will often be “ill- conditioned”; this problem was also encountered in [5], and a good discussion can be found in Appendix B of that paper. We were able to avoid this problem by instead considering a Fisher matrix that takes into account a smaller number of variables but leads to accurate 10