Skip to main content
Sandbox Physics
Gravitational-wave standard sirens

Dark sirens, galaxy catalogs and the selection normalisation

Sandbox Physics ·

A dark siren replaces the one known host by a weighted sum over every catalog galaxy the signal allows, and divides the result by the fraction of sources that would have been detected for each trial Hubble constant. For nearby sources that fraction grows as H03H_0^3. Leaving it out biases the answer, and in our synthetic tests it pushes a 40-event dark-siren result to the edge of the prior. With it, both bright and dark posteriors are calibrated. A missing host galaxy breaks the calibration unless the analysis models the galaxies the catalog did not see.

Open the standard sirens Lab

This chapter follows from one distance and one redshift to the Hubble constant. It uses conditional probability and integrals over a volume.

The catalog sum

Suppose the localisation volume of one event contains NN catalog galaxies with measured velocities viv_i. Each galaxy's true redshift is uncertain because of its peculiar velocity, so it carries a normalised redshift distribution

p(z∣vi)∝N(vi; cz,σv) p0(z),p(z\mid v_i)\propto\mathcal N(v_i;\,cz,\sigma_v)\,p_0(z),

where p0(z)p_0(z) is the distribution of galaxies in redshift before any measurement. With equal prior weight on each galaxy, the likelihood of the gravitational-wave data xx is

p(x∣H0)∝1N∑i=1N∫p ⁣(x∣DL(z,H0))p(z∣vi) dz.p(x\mid H_0)\propto\frac1N\sum_{i=1}^{N}\int p\!\left(x\mid D_L(z,H_0)\right)p(z\mid v_i)\,\mathrm dz .

Each galaxy implies its own value H0≈vi/DH_0\approx v_i/D. The true host appears in every event; impostors appear at random. Chen, Fishbach and Holz write the same sum with weights that can follow galaxy luminosity, and Gray et al. implement it in the gwcosmo code. Chen, Fishbach and Holz, A two per cent Hubble constant measurement from standard sirens within five years, Methods Gray et al., Cosmological inference using gravitational wave standard sirens: a mock data analysis

The normalisation of p(z∣vi)p(z\mid v_i) matters. An unnormalised Gaussian times p0(z)∝z2p_0(z)\propto z^2 gives distant galaxies more total weight than near ones, and the posterior drifts high. Our first Lab model made exactly that mistake.

Worked example 1. Our reference event has a measured distance of 77.9−10.3+7.6 Mpc77.9^{+7.6}_{-10.3}\,\mathrm{Mpc}. Its true host has v=5300 km s−1v=5300\,\mathrm{km\,s^{-1}}, and nine other catalog galaxies have velocities between 31003100 and 18300 km s−118300\,\mathrm{km\,s^{-1}}. They imply H0=5300/77.9=68H_0=5300/77.9=68 for the host, and 4040, 5454, 8989, 9898, 126126, 141141, 159159, 194194 and 235235 for the others. The last two lie outside the prior range from 2020 to 160160. The bright-siren posterior, knowing the host, is 70.2−7.3+11.370.2^{+11.3}_{-7.3}; the dark-siren posterior is 90−41+4990^{+49}_{-41}, broad and many-peaked. These numbers come from the published inputs of our method guide.

Why the likelihood must be normalised

Only events above the detection threshold enter the analysis. Mandel, Farr and Gair show that the correct likelihood for one detected event, given population parameters λ\lambda, is

p(d∣λ)=∫p(d∣θ) ppop(θ∣λ) dθα(λ),α(λ)=∫pdet(θ) ppop(θ∣λ) dθ,p(d\mid\lambda)=\frac{\int p(d\mid\theta)\,p_{\mathrm{pop}}(\theta\mid\lambda)\,\mathrm d\theta}{\alpha(\lambda)},\qquad\alpha(\lambda)=\int p_{\mathrm{det}}(\theta)\,p_{\mathrm{pop}}(\theta\mid\lambda)\,\mathrm d\theta ,

and that independent events multiply. Here θ\theta are the source parameters and pdetp_{\mathrm{det}} is the probability that a source with those parameters is detected. Mandel, Farr and Gair, Extracting distribution parameters from multiple uncertain observations with selection biases, Eqs. 5 to 7

For a siren, λ=H0\lambda=H_0 and the population is the galaxy distribution. The normalisation is usually written

β(H0)=∫Pdet ⁣(DL(z,H0))p0(z) dz,\beta(H_0)=\int P_{\mathrm{det}}\!\left(D_L(z,H_0)\right)p_0(z)\,\mathrm dz ,

with Pdet(D)P_{\mathrm{det}}(D) the detection probability averaged over orientation and sky position. Chen, Fishbach and Holz, A two per cent Hubble constant measurement from standard sirens within five years, Methods, Eq. 9 Chen et al. replace the galaxy distribution in β\beta by a smooth isotropic one. In the Lab and in our tests the catalog is small and fully known, so the exact dark-siren normalisation is the average of PdetP_{\mathrm{det}} over the listed galaxies' redshift distributions. That exact average is our modelling choice.

Why the normalisation grows as the cube of the Hubble constant

Let galaxies fill space uniformly out to zmax⁡z_{\max}, so p0(z)=3z2/zmax⁡3p_0(z)=3z^2/z_{\max}^3, and use the nearby relation D=cz/H0D=cz/H_0. Changing variables to DD,

β(H0)=3zmax⁡3(H0c)3∫0czmax⁡/H0Pdet(D) D2 dD.\beta(H_0)=\frac{3}{z_{\max}^3}\left(\frac{H_0}{c}\right)^3\int_0^{cz_{\max}/H_0}P_{\mathrm{det}}(D)\,D^2\,\mathrm dD .

If the detection horizon lies well inside the catalog, the upper limit can be taken to infinity and the integral no longer depends on H0H_0. Then β∝H03\beta\propto H_0^3.

The physical reason is simple. Raising H0H_0 squeezes the same galaxies, listed by redshift, into a smaller distance, so more of them fall inside the fixed volume the detectors can reach. Without β\beta, the catalog sum also grows as H03H_0^3 near the measured distance, for the same reason, and the posterior climbs toward large H0H_0.

Table 4.1 Checks of the normalisation in our synthetic study. All events are simulated; nothing here is detector data.
CheckExpectedFound
Detected fraction at H0=70H_0=70, ρ100=15\rho_{100}=15quadrature 0.021790.02179Monte Carlo 0.02189±0.000230.02189\pm0.00023
β(80)/β(40)\beta(80)/\beta(40), horizon inside catalog887.927.92
β(140)/β(40)\beta(140)/\beta(40)42.942.941.841.8

The ratios fall slightly short of H03H_0^3 because DL(z)D_L(z) curves upward at larger zz, by the factor 1+(1−q0)z/21+(1-q_0)z/2. At H0=140H_0=140 the detectable galaxies sit at larger redshift, and the shortfall reaches 2.6 percent.

Testing calibration by injection

A posterior is calibrated if, over many universes drawn from the prior, the true value falls in its central 68 percent interval 68 percent of the time, and the posterior quantile of the true value is uniformly distributed. We drew H0H_0 from the flat prior, drew events in that universe until one was detected, analysed it, and repeated. A Kolmogorov-Smirnov test checks the uniformity of the quantiles.

Table 4.2 Simulation-based calibration of single synthetic events. Expected coverage is 0.6830.683; the binomial standard error is 0.0230.023 for 400 trials and 0.0290.029 for 250.
AnalysisTrialsCoverageKS ppVerdict
Bright, with β\beta4004000.7000.7000.0830.083calibrated
Bright, without β\beta4004000.6530.6531.2×10−91.2\times10^{-9}biased
Dark, with catalog β\beta2502500.7280.7280.850.85calibrated
Dark, without β\beta2502500.3840.3848×10−428\times10^{-42}biased

In 51 of the 250 dark trials the first catalog contained no galaxy that could produce a detection, so the catalog was redrawn. Stacking 40 events at H0=70H_0=70 gives 71.9−1.6+1.971.9^{+1.9}_{-1.6} for bright sirens with β\beta and 77.6−2.0+2.377.6^{+2.3}_{-2.0} without it. Dark sirens give 67.9−2.6+2.767.9^{+2.7}_{-2.6} with β\beta; without it the result sits at the edge of the prior, near 159159. Each stacked number comes from one random realisation that we did not repeat, so an offset of about one interval width, as in the bright result, is within the expected scatter.

Galaxies the catalog does not contain

Real catalogs miss faint and distant galaxies. Gray et al. split the likelihood into the case where the host is in the catalog, GG, and the case where it is not, Gˉ\bar G:

p(x∣D,H0)=p(x∣G,D,H0) p(G∣D,H0)+p(x∣Gˉ,D,H0) p(Gˉ∣D,H0).p(x\mid D,H_0)=p(x\mid G,D,H_0)\,p(G\mid D,H_0)+p(x\mid\bar G,D,H_0)\,p(\bar G\mid D,H_0).

The second term spreads the missing galaxies smoothly according to a luminosity function and the catalog's magnitude limit. In their mock data, combining 249 simulated binary neutron stars, a known host gave 1.13 percent, a complete catalog 1.84 percent, and catalogs 75, 50 and 25 percent complete gave 2.21, 2.48 and 3.20 percent. Their simulated universe had no redshift uncertainties or clustering, and the catalog was about three times denser than the local universe. Gray et al., Cosmological inference using gravitational wave standard sirens: a mock data analysis, Table II

We ran the opposite experiment: delete the host from the catalog and keep the in-catalog likelihood only. Coverage fell to 0.550.55 with KS p=1.9×10−8p=1.9\times10^{-8}, and one posterior had no support at all. The analysis was confidently wrong. A second failure case underestimated the peculiar-velocity spread, 100100 instead of 500 km s−1500\,\mathrm{km\,s^{-1}}, with loud events. Coverage fell only to 0.640.64 and the KS test did not flag it. The likely reason is that at these distances a 500 km s−1500\,\mathrm{km\,s^{-1}} error is a few percent of the velocity, smaller than most of the distance errors. That case shows a test can miss a real mistake.

Lab tasks

  1. Load the selection bias preset, where β\beta is off. Note both medians, then turn β\beta on in the advanced group. Which posterior moves more, and in which direction?
  2. Use the crowded catalog preset and drag galaxies in the cone toward the host's distance. Watch the dark-siren peaks follow them.
  3. Set the candidate count to 1010, then to 200200. Explain the change in the dark-siren width with the catalog sum.

Exercises with explained answers

  1. In the nearby limit, by what factor does β\beta change between H0=60H_0=60 and 7575?
  2. A siren has D=60±6 MpcD=60\pm6\,\mathrm{Mpc}, and the catalog lists two galaxies with v=4200v=4200 and 6300 km s−16300\,\mathrm{km\,s^{-1}}. Where does the single-event dark posterior peak?
  3. Why does a dark posterior without β\beta drift to large H0H_0, while a bright posterior without β\beta is biased only mildly?
  4. Why does deleting the host from the catalog produce over-confident errors rather than just wider posteriors?
Answers
  1. (75/60)3=1.95(75/60)^3=1.95.
  2. Near 4200/60=704200/60=70 and 6300/60=1056300/60=105, each with a width of about 10 percent from the distance error. With equal weights and similar redshift weight, the two peaks have comparable height; the data cannot choose between them.
  3. Without β\beta the catalog sum counts how many galaxies sit near the measured distance, and that count grows as H03H_0^3 until the catalog edge. A bright siren uses one fixed redshift, so the missing factor only reweights a single narrow likelihood; the bias comes from the threshold favouring sources that look nearer.
  4. The posterior is still a sum of narrow peaks at the impostors. One of them takes all the weight, and its interval is as narrow as a real detection, so the true value falls outside more often than the width suggests.

Try it in the Lab

Load the selection bias preset, where the correction is off, then turn the selection normalisation on in the advanced group and watch both posteriors move. Open the standard sirens Lab

Sources

  1. Chen, Fishbach and Holz, A two per cent Hubble constant measurement from standard sirens within five years, Methods
  2. Gray et al., Cosmological inference using gravitational wave standard sirens: a mock data analysis
  3. Mandel, Farr and Gair, Extracting distribution parameters from multiple uncertain observations with selection biases, Eqs. 5 to 7

Keep reading

What this page links to directly

See this series on the content map →

Gravitational-wave standard sirens · Reading path

From a wave amplitude to the Hubble constant

  1. ArticleHow can a gravitational wave tell us how far away it came from?
  2. ArticleWaveform amplitude, orbit tilt and luminosity distance
  3. ArticleFrom one distance and one redshift to the Hubble constant
  4. You are hereDark sirens, galaxy catalogs and the selection normalisation
  5. ArticleRecomputing a bright-siren and a dark-siren Hubble constant
  6. ArticleReading the GWTC-3 dark-siren cosmology paper
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