Skip to main content
Sandbox Physics
Black hole ringdown and the no-hair test

Why a ringdown fit needs the overtone near the peak

Sandbox Physics ·

A fit that uses only the fundamental mode from the peak of a ringdown lands on the wrong mass and spin, because the fast-fading overtone makes the early signal decay faster than the fundamental alone can. Adding the overtone removes the bias at the peak; starting the fundamental-only fit later also removes it, at the price of a much wider answer. This chapter builds the fit step by step on a synthetic record, so the true answer is known.

Open the black hole ringdown Lab

The previous chapter, damped tones and the Kerr spectrum, supplies the frequencies used here. Least squares and the idea of a probability distribution are enough.

The record and its noise

Our test record is synthetic. It is the 220220 mode plus the 221221 overtone, both starting at t=0t=0, which we call the peak:

s(t)=e−t/τ0cos⁡(2πf0t)+A1A0 e−t/τ1cos⁡(2πf1t−Δϕ),s(t)=e^{-t/\tau_0}\cos(2\pi f_0t)+\frac{A_1}{A_0}\,e^{-t/\tau_1}\cos(2\pi f_1t-\Delta\phi),

with M=68.5 M⊙M=68.5\,\mathrm{M}_\odot, χ=0.69\chi=0.69, A1/A0=4A_1/A_0=4 and Δϕ=0\Delta\phi=0, sampled at 4096 Hz4096\,\mathrm{Hz} for 60M60M, about 20 ms20\,\mathrm{ms}. The overtone ratio follows a fit to a GW150914-like simulation. Giesler, Isi, Scheel and Teukolsky, Black hole ringdown: the importance of overtones, Table I We add white Gaussian noise of standard deviation σ\sigma, chosen so that the signal-to-noise ratio over the whole segment is

ρ=1σ∑is(ti)2=14.\rho=\frac{1}{\sigma}\sqrt{\sum_i s(t_i)^2}=14 .

That matches the post-peak signal-to-noise ratio Isi and colleagues quote for GW150914. Isi, Giesler, Farr, Scheel and Teukolsky, Testing the no-hair theorem with GW150914, arXiv:1905.00869v2 Real detector noise is coloured, and a real signal before the peak is more than a sum of modes; the method tutorial treats both on GW150914 data.

Remove the amplitudes with least squares

A model with modes at fixed mass and spin is linear in its amplitudes. Write each mode as ane−t/τncos⁡(2πfnt)+bne−t/τnsin⁡(2πfnt)a_n e^{-t/\tau_n}\cos(2\pi f_nt)+b_n e^{-t/\tau_n}\sin(2\pi f_nt) and collect these functions, sampled at the times in the fit window, as the columns of a matrix X(M,χ)X(M,\chi). For data dd and coefficients c=(a0,b0,a1,b1)c=(a_0,b_0,a_1,b_1),

−2ln⁡L=∣d−Xc∣2σ2+const.-2\ln L=\frac{\lvert d-Xc\rvert^2}{\sigma^2}+\mathrm{const}.

The best coefficients are c^=(XTX)−1XTd\hat c=(X^\mathsf TX)^{-1}X^\mathsf Td. Let P=X(XTX)−1XTP=X(X^\mathsf TX)^{-1}X^\mathsf T be the projection onto the model's columns. Then ∣d−Xc^∣2=∣d∣2−∣Pd∣2\lvert d-X\hat c\rvert^2=\lvert d\rvert^2-\lvert Pd\rvert^2, and the likelihood maximised over amplitudes and phases is

ln⁡Lprof(M,χ)=∣Pd∣22σ2+const.\ln L_{\mathrm{prof}}(M,\chi)=\frac{\lvert Pd\rvert^2}{2\sigma^2}+\mathrm{const}.

This is the profile likelihood. It turns a search over six parameters into a search over two. Its cost is that the widths come out a little too narrow, because amplitude uncertainty is maximised away instead of averaged; the coverage test below measures how much.

A grid posterior and its 90 percent region

Put flat priors on a grid, mass from 0.60.6 to 1.61.6 times the true mass in 81 steps and spin from 00 to 0.990.99 in steps of 0.0150.015, and set the posterior proportional to LprofL_{\mathrm{prof}}. The 90 percent region is the smallest set of grid cells that holds 90 percent of the weight.

To ask whether the truth is inside, compute the credible level of the true point: the share of posterior weight in cells more probable than the truth. A level below 0.90.9 means the truth lies inside the 90 percent region; a level near 11 means it is far outside.

What happens when the overtone is left out

Table 3.1 fits the same noisy record with the fundamental only (N=0N=0) and with the fundamental plus overtone (N=1N=1), starting at several times after the peak. Masses are medians with 5 to 95 percent ranges.

Table 3.1 Start-time scan on one synthetic record, true M=68.5 M⊙M=68.5\,\mathrm{M}_\odot, χ=0.69\chi=0.69, ρ=14\rho=14.
Start t0/Mt_0/MSamplesN=0N=0 massN=0N=0 levelN=1N=1 massN=1N=1 level
00838348.1  [42.2, 55.0]48.1\;[42.2,\,55.0]1.0001.00068.9  [49.5, 89.1]68.9\;[49.5,\,89.1]0.4970.497
22808062.3  [48.2, 88.6]62.3\;[48.2,\,88.6]0.8620.86271.7  [47.3, 99.0]71.7\;[47.3,\,99.0]0.4630.463
44777769.8  [49.2, 98.5]69.8\;[49.2,\,98.5]0.8340.83465.4  [43.2, 103.0]65.4\;[43.2,\,103.0]0.8000.800
1010696959.8  [42.5, 100.0]59.8\;[42.5,\,100.0]0.5020.50272.8  [43.4, 100.9]72.8\;[43.4,\,100.9]0.8520.852
2020555574.4  [44.6, 105.8]74.4\;[44.6,\,105.8]0.2750.27573.8  [45.2, 104.7]73.8\;[45.2,\,104.7]0.4150.415

Worked example 1. Why does N=0N=0 at the peak give a low mass? Early on, the overtone is four times stronger than the fundamental and dies about three times faster, so the total signal fades quickly. A single mode can only fade quickly if its QQ is small, which needs a low spin. The fit found χ=0.07\chi=0.07. At that spin Mω220=0.3825−0.0888iM\omega_{220}=0.3825-0.0888i. Keeping the frequency near the true 249 Hz249\,\mathrm{Hz} then forces M=0.3825/(2π×249 Hz×4.925 μs)=49.6 M⊙M=0.3825/(2\pi\times249\,\mathrm{Hz}\times4.925\,\mu\mathrm s)=49.6\,\mathrm{M}_\odot, close to the fitted 48.148.1. At 48.1 M⊙48.1\,\mathrm{M}_\odot and χ=0.07\chi=0.07 the mode has f=257 Hzf=257\,\mathrm{Hz} and τ=2.67 ms\tau=2.67\,\mathrm{ms}: almost the right pitch, but a much shorter ring than the true 4.16 ms4.16\,\mathrm{ms}.

The direction of this bias depends on the record. The same fundamental-only fit on real GW150914 data, in the method tutorial, gives a mass above the full-signal value. What is robust is that the fit is biased, and that the bias shrinks as the start moves later.

Choosing the start time

Starting later lets the overtone die before the fit begins. By t0=2Mt_0=2M the truth is already inside the N=0N=0 region in this record, and by 10M10M its level is 0.500.50. The price is width. The N=0N=0 range at 10M10M spans 58 M⊙58\,\mathrm{M}_\odot; the N=1N=1 range at the peak spans 4040. Fewer samples carry signal, and the loudest ones are thrown away.

Isi and colleagues saw the same trade-off in GW150914: the fundamental alone agreed with the full-signal estimate about 3 ms3\,\mathrm{ms} after the peak, with a larger uncertainty than the overtone fit from the peak. Isi, Giesler, Farr, Scheel and Teukolsky, Fig. 3 Our table is one noise draw. Another draw gives other numbers, and the coverage test below shows how often the truth lands inside.

How honest are the 90 percent regions?

Repeat the N=1N=1 fit from the peak on 100 independent noise draws and count how often the truth lands inside each region.

Table 3.2 Fraction of 100 draws with the truth inside the highest-density region.
Nominal levelρ=14\rho=14ρ=40\rho=40
0.500.500.400.400.370.37
0.680.680.580.580.600.60
0.900.900.890.890.880.88

The 90 percent regions cover the truth about as often as they should. The 50 and 68 percent regions are too small. With 100 draws the binomial 95 percent band is about 0.400.40 to 0.600.60 around 0.50.5 and 0.590.59 to 0.770.77 around 0.680.68, and three of the four low-level entries fall below their band. We suspect the profile likelihood, which does not average over amplitude uncertainty, but we have not tested that. Read the narrower ranges in the Lab with this in mind.

Lab tasks

  1. In the default preset, read the amber fundamental-only mass and the mint fundamental-plus-overtone mass. Compare both with Table 3.1.
  2. Drag the fit-start marker on the strain chart to 2M2M, 4M4M and 10M10M. Note when the amber region first covers the white cross.
  3. Open the advanced group and change the noise draw. Watch how far the N=1N=1 median moves compared with its width.

Exercises with explained answers

  1. Show that ∣d−Xc^∣2=∣d∣2−∣Pd∣2\lvert d-X\hat c\rvert^2=\lvert d\rvert^2-\lvert Pd\rvert^2.
  2. How many samples does the window hold when t0=10Mt_0=10M for M=68.5 M⊙M=68.5\,\mathrm{M}_\odot?
  3. If you lower the overtone ratio to 0.30.3 in the Lab, should the fundamental-only bias at the peak grow or shrink? What does the Lab show at ρ=30\rho=30?
  4. A credible level of 0.8620.862 was found for N=0N=0 at t0=2Mt_0=2M. Is the truth inside the 90 percent region?
Answers
  1. PP is symmetric and P2=PP^2=P. With Xc^=PdX\hat c=Pd, ∣d−Pd∣2=dTd−2dTPd+dTP2d=∣d∣2−∣Pd∣2\lvert d-Pd\rvert^2=d^\mathsf Td-2d^\mathsf TPd+d^\mathsf TP^2d=\lvert d\rvert^2-\lvert Pd\rvert^2.
  2. The window runs from 10M10M to 60M60M, that is 50×0.3374 ms=16.87 ms50\times0.3374\,\mathrm{ms}=16.87\,\mathrm{ms}. At 4096 Hz4096\,\mathrm{Hz} that is 6969 samples, as in Table 3.1.
  3. It should shrink, because a weaker overtone distorts the early decay less. The faint-overtone preset (A1/A0=0.3A_1/A_0=0.3, ρ=30\rho=30) still excludes the truth with the fundamental only, mass 56.7  [51.4, 62.5]56.7\;[51.4,\,62.5], because the higher signal-to-noise ratio also narrows the region. Bias matters when it is large compared with the width.
  4. Yes. 0.862<0.90.862<0.9, so the truth sits inside the 90 percent region, close to its edge.

Next, turn two fitted tones into a test of the no-hair theorem in from two tones to a no-hair test.

Try it in the Lab

Drag the fit start on the strain chart from 0 to 10 and watch the amber fundamental-only region move until it covers the white cross. Open the black hole ringdown Lab

Sources

  1. Giesler, Isi, Scheel and Teukolsky, Black hole ringdown: the importance of overtones, Table I
  2. Isi, Giesler, Farr, Scheel and Teukolsky, Testing the no-hair theorem with GW150914, arXiv:1905.00869v2

Keep reading

What this page links to directly

See this series on the content map →

Black hole ringdown and the no-hair test · Reading path

From a ringing black hole to a no-hair test

  1. ArticleWhy does a newly merged black hole ring like a bell?
  2. ArticleDamped tones and the Kerr spectrum
  3. You are hereWhy a ringdown fit needs the overtone near the peak
  4. ArticleFrom two tones to a no-hair test
  5. ArticleRecomputing ringdown tones, a synthetic signal and a slice of GW150914
  6. ArticleHow to read the 2019 GW150914 overtone paper and the debate that followed
Something broken, confusing, or missing on this page?Tell us
What kind of feedback?

Sent with this page’s address, its Lab ID, the site language and your device type. Please leave out personal details. Feedback is kept for up to 12 months.Privacy policy