The paper is NANOGrav Collaboration, 2023, arXiv version 1, "The NANOGrav 15-year Data Set: Evidence for a Gravitational-Wave Background". Section numbers below refer to that version. Read the optimal-statistic chapter first if Bayes factors and phase shifts are new to you.
Three nested questions
The paper answers three questions in order, and each answer has its own statistic.
- Is there a common red signal with the same spectrum in all pulsars? This compares a model with only intrinsic pulsar noise against a common uncorrelated red process. The Bayes factor exceeds 1014, so the answer is yes by a wide margin. That signal was already strongly supported in the 12.5-year data.
- Is that common signal correlated between pulsars in the Hellings–Downs pattern? This compares the Hellings–Downs model against the common uncorrelated model. It is the gravitational-wave question.
- Could a clock monopole or an ephemeris dipole explain the correlations instead? This compares other correlation patterns.
Most headlines quote question 1 or mix it with question 2. The evidence for gravitational waves rests on question 2, which is much weaker. NANOGrav Collaboration, 2023, Abstract and Section 3
The Bayesian evidence
Figure 2 lists Bayes factors between models. The Hellings–Downs model is favoured over the common uncorrelated model by 200 when 14 frequencies are included and by 1000 with 5 frequencies. Monopole and dipole models are strongly disfavoured against the common uncorrelated model, and adding a monopole or dipole on top of Hellings–Downs gives roughly even odds.
A Bayes factor is not a probability that the hypothesis is true. To turn it into a false-alarm rate, the authors recomputed it on 5000 data sets in which random phase shifts removed the inter-pulsar correlations. Five of them exceeded the observed value, so p=10−3, about three standard deviations. Figure 3, left panel, shows that distribution. NANOGrav Collaboration, 2023, Section 3 and Figure 3
The frequentist evidence
Section 4 builds the optimal statistic. Figure 4 shows its formal S/N, 5±1 for a free spectral index and 4±1 for γ=13/3, spread over the noise-parameter posterior. The paper states plainly that this S/N is not a sigma level.
The p-values come from Figure 3, right panel: 19 of 400 000 phase shifts beat the observed S/N, p=5×10−5. Simulations of the uncorrelated model give 1.8×10−4 and an analytic calculation 1.9×10−4. Sky scrambles, which keep correlations but remove their dependence on angle, give p<10−4.
Figure 1(c) is the binned correlation plot that most people remember. It averages the 2211 pair correlations in 15 bins of roughly equal pair counts, chosen to avoid the zeros of the curve, at the maximum a posteriori noise parameters. Its error bars include the covariance between pairs that a real signal creates. The 15 bins give χ2=8.1 against the curve, a p-value of 0.75. Figure 1(c) pictures the fit; a different binning or noise draw would move the points, and the tests above carry the evidence.
The amplitude and the spectral index
With γ fixed at 13/3 the amplitude at one cycle per year is A=2.4−0.6+0.7×10−15, median and 90 % credible interval. Letting γ vary gives A=6.4−2.7+4.2×10−15 and γ=3.2−0.6+0.6, with 13/3 at the 99 % credible boundary. The two amplitudes look very different only because the reference frequency of one year sits far above the sensitive band; at a reference of one cycle per ten years the correlation between A and γ largely disappears. NANOGrav Collaboration, 2023, Section 3 and Figure 1(b)
The tension in γ matters for interpretation. Binaries that lose energy to stars and gas as well as to gravitational waves would flatten the low-frequency spectrum. Population models of these binaries can reproduce both the amplitude and the shape, but the amplitude needs either several parameters near the edges of expected values or a few notably different from standard expectations. The origin stays open until higher-order features of the spectrum can be measured. NANOGrav Collaboration, 2023c, Abstract
The monopole at 4 nHz
Section 5.3 is the part most worth reading slowly. A Legendre fit with the multiple-component optimal statistic (Figure 7) shows the quadrupole structure expected from gravitational waves, plus a small but significant monopole. Simulations of pure Hellings–Downs data produce an apparent monopole that large with p=0.11. A Bayesian free-spectrum search finds monopole power only in the second frequency bin, 3.95nHz; a model with an extra sinusoid there has a Bayes factor of about one. The authors audited the Arecibo and Green Bank time corrections and found nothing. They conclude that they cannot say whether the signal is present or what causes it. NANOGrav Collaboration, 2023, Section 5.3
Our own simplified recomputation also finds a monopole S/N larger than the Hellings–Downs S/N in the three-shape fit, 0.66 against 0.39 in formal units. That agrees with the paper, but our noise model is too crude to separate the two shapes, so it adds no independent weight.
What the evidence does not yet show
The anisotropy limit comes from the companion search, which found no significant evidence for anisotropy and set Cl>0/Cl=0<20%. NANOGrav Collaboration, 2023b, Abstract Allen's analysis explains why no data set will ever match the curve perfectly: the interference of the particular sources in our sky leaves a cosmic variance. Allen, 2022, Section V
Reading checklist
When you read a pulsar-timing result, check five things. Which question does each number answer: common signal, correlation, or source? Is a significance a Bayes factor, a formal S/N or a calibrated p-value? How were the nulls built, and how many trials were run? Which normalisation of the curve and which reference frequency are used? And what extra correlated signals were fitted at the same time? The recomputation guide lets you test each of these on the public data.