This guide reads arXiv version 2 of the paper, dated 8 August 2019, published as Phys. Rev. Lett. 123, 111102 (2019). Isi, Giesler, Farr, Scheel and Teukolsky, Testing the no-hair theorem with GW150914, arXiv:1905.00869v2 We read the full text of the main paper, its four figures and the supplementary section on the likelihood. We did not compare it with the journal's typeset version. Our own fits, described in the method tutorial, use one detector and do not reproduce the paper.
Where each claim stands
The journal references come from the arXiv record of each paper. Giesler, Isi, Scheel and Teukolsky, Black hole ringdown: the importance of overtones, arXiv:1903.08284v2
Eq. (1): the model and its conventions
The paper writes the l=m=2 ringdown as
for times after a start time t0. Here ωn is a real angular frequency and τn a damping time, both set by the remnant's mass and spin; An and ϕn are free. Two details matter when you compare with other work.
- Printed vs used: the text defines Δt=t−t0 right after Eq. (1), but the equation is written in t. Fig. 2 says the amplitudes are defined at t=tpeak. Our code measures time from the fit start, which changes the amplitudes but not the mass and spin, because amplitudes are fitted away.
- Printed vs used: the paper keeps ωn real and writes the decay separately. We use one complex frequency, ω=2πf−i/τ with time dependence e−iωt, as in the course chapters. The two forms describe the same mode.
The start time is measured from tpeak=1126259462.423 GPS at Hanford.
Fig. 1: three models from the peak
Fig. 1 shows 90 percent regions in mass and spin for N=0, 1 and 2, all from Δt0=0, against the full-signal estimate, whose peak is at Mf=68.5M⊙, χf=0.69. With N=0 the region misses the full-signal one; the text calls it biased and does not give a direction. With N=1 the two agree, and the text quotes Mf=68±7M⊙ and χf=0.63±0.16 at 68 percent. With N=2 the region is slightly broader than with N=1, because two more free parameters do not buy a clearly detected second overtone.
Read this figure as a statement about model adequacy at the peak. In the Lab and in our Hanford-only slice, the N=0 fit from the peak also misses, once low and once high, so the direction depends on the record.
Fig. 2: how strong is the overtone?
Fig. 2 shows the amplitude posteriors. With N=1, the mean of A1 lies 3.6 standard deviations from zero. That number describes the width of a posterior; it is not a false-alarm probability from noise-only data. With N=2, the case A1=A2=0 is excluded at 90 percent credibility. The paper also states that the post-peak network signal-to-noise ratio is about 14 above 154.7Hz.
Fig. 3: waiting for the fundamental
Fig. 3 repeats the N=0 fit from 1, 3 and 5ms after the peak. Around 3ms it agrees with the full-signal estimate, with a larger region than the N=1 fit from the peak. This matches the earlier LIGO and Virgo analysis, where a single damped sinusoid started to overlap the prediction at t0=tM+3ms, about 10M. Abbott et al., Tests of general relativity with GW150914, arXiv:1602.03841v3 The Lab's "wait for the fundamental" preset makes the same trade-off on a synthetic record.
Fig. 4: the no-hair test itself
Here the overtone's frequency and damping time are freed, f1=f1Kerr(1+δf1) and τ1=τ1Kerr(1+δτ1), with flat priors δf1∈[−0.5,0.5] and δτ1∈[−1,1]. The result is δf1=−0.05±0.2 at 68 percent; δτ1 is largely unconstrained between about −0.06 and 1. The Bayes factor for the Kerr model over the free one is 1.75.
What this shows: a test with real data, passed at about the 20 percent level. What it does not show: a measurement precise enough to constrain specific alternative theories, or a test independent of the overtone's detection. If the overtone were absent, δf1 would be poorly constrained, and the test would mean little.
The supplementary likelihood
The supplement explains the data handling, and it is the part to read before trying to reproduce anything. The authors apply a fourth-order Butterworth high-pass at 20Hz, then work at 2048Hz with 0.5s of data from the peak. They estimate the noise autocovariance from 64s of off-source data and use the full Toeplitz covariance in the time domain, because a taper would require data before the peak. They project the complex strain onto both detectors with antenna patterns.
Our Hanford-only slice copies the filter, sample rate, segment and Toeplitz likelihood. It differs in detector count, off-source length, amplitude treatment and sampling, and the method tutorial lists each difference.
What came after
Cotesta, Carullo, Berti and Cardoso reanalysed the event and found no evidence for an overtone after the peak; support for a nonzero amplitude changed when the start time moved by much less than the overtone's damping time, and injections into nearby real noise produced false evidence. Cotesta, Carullo, Berti and Cardoso, Analysis of ringdown overtones in GW150914, arXiv:2201.00822v2 Isi and Farr answered that they could not reproduce those results and that the overtone amplitude stayed nonzero over reasonable start times. Isi and Farr, Revisiting the ringdown of GW150914, arXiv:2202.02941v2 The two groups disagree about the data analysis, and neither paper has a journal entry on arXiv.
Separately, two simulation studies found quadratic modes, sourced by the square of the 220 mode, in the (4,4) harmonic. Mitman et al., Nonlinearities in black hole ringdowns, arXiv:2208.07380v2 Cheung et al., Nonlinear effects in black hole ringdown, arXiv:2208.07374v3 They concern the (4,4) harmonic, not the (2,2) fit in Isi et al., but they show that a purely linear sum of modes is incomplete for higher harmonics in simulations.
Three questions remain open as of November 2026: whether nonlinear modes can be measured in any event, whether the GW150914 overtone is detected, and how early the linear description holds in the (2,2) harmonic of real signals.
Lab tasks
- Use the default preset as a stand-in for Fig. 1: compare the amber N=0 region and the mint N=1 region with the white truth cross.
- Use the "wait for the fundamental" preset for Fig. 3, and drag the start between 5M and 15M.
- Read the overtone frequency-shift curve as a synthetic Fig. 4 slice with δτ1=0. Remember that the Lab record is synthetic white-noise data, not LIGO data.
Exercises with explained answers
- The paper says A1=0 is disfavoured at 3.6σ. Why is this not the same as saying noise alone would produce such a value only once in several thousand trials?
- Why does the N=2 region come out slightly broader than the N=1 region?
- Which single figure would you check first to judge the no-hair claim, and what number on it?
AnswersShow the answer
- The 3.6σ is the posterior mean of A1 divided by its posterior standard deviation under the N=1 model. A false-alarm rate needs the distribution of that statistic in data without an overtone, which is a separate calculation. Cotesta and colleagues studied it with injections into nearby real noise.
- Adding A2 and ϕ2 adds freedom that the data cannot pin down, since the second overtone is not clearly detected. The extra freedom spreads the mass and spin posterior.
- Fig. 4, and the δf1 interval: −0.05±0.2 at 68 percent, consistent with zero. Then Fig. 2, because the test only means something if the overtone is present.
Try it in the Lab
Use the default preset as a stand-in for Fig. 1 and the wait-for-the-fundamental preset for Fig. 3; remember the Lab record is synthetic, not LIGO data. Open the black hole ringdown Lab
Sources
- Isi, Giesler, Farr, Scheel and Teukolsky, Testing the no-hair theorem with GW150914, arXiv:1905.00869v2
- Giesler, Isi, Scheel and Teukolsky, Black hole ringdown: the importance of overtones, arXiv:1903.08284v2
- Abbott et al., Tests of general relativity with GW150914, arXiv:1602.03841v3
- Cotesta, Carullo, Berti and Cardoso, Analysis of ringdown overtones in GW150914, arXiv:2201.00822v2
- Isi and Farr, Revisiting the ringdown of GW150914, arXiv:2202.02941v2
- Mitman et al., Nonlinearities in black hole ringdowns, arXiv:2208.07380v2
- Cheung et al., Nonlinear effects in black hole ringdown, arXiv:2208.07374v3