The NANOGrav 15 yr Data SetEvidence for a Gravitational-wave Background

Somewhere in the universe right now, two black holes the size of solar systems are slowly spiraling toward each other, a collision billions of years in the making. As they orbit, they ripple spacetime itself. And those ripples — stretched so long they take years to complete a single oscillation — are washing through our galaxy at this moment. In 2023, the North American Nanohertz Observatory for Gravitational Waves, or NANOGrav, reported strong evidence that they have detected exactly this kind of signal. Here's how they did it. The instrument isn't a telescope or a detector in the conventional sense. It's a galaxy-scale network of natural clocks. Millisecond pulsars — rapidly spinning neutron stars that flash radio pulses with extraordinary regularity — are among the most stable timekeepers in the known universe. A gravitational wave passing between Earth and a pulsar stretches and squeezes that intervening space, which means the pulsar's pulses arrive slightly early or slightly late depending on the wave's phase. The effect is tiny: a fractional change in distance of about 2.4 parts in ten to the fifteenth power. But accumulated over thousands of light-years, it becomes measurable — if you're patient enough and precise enough. NANOGrav has been patient. The collaboration collected fifteen years of data, from July 2004 through August 2020, timing sixty-seven millisecond pulsars using three major facilities: the three hundred five meter Arecibo Observatory, the one hundred meter Green Bank Telescope, and, since 2015, the Very Large Array. The key question running through all of this: is there a faint, correlated disturbance buried in those arrival times — one that shows up across many pulsars simultaneously, in a pattern that only gravitational waves could produce? That pattern has a name. Back in 1983, Ronald Hellings and George Downs worked out what a true, isotropic gravitational wave background should look like across a network of pulsars. Their prediction is now called the Hellings-Downs curve, and it's the conceptual heart of this result. Here's the key idea: if a gravitational wave background is washing over our galaxy from all directions, it should produce correlated timing disturbances between pairs of pulsars. But the strength of that correlation shouldn't be random — it should depend on the angular separation between the pulsars on the sky in a very specific, predictable way. Pulsar pairs that are close together on the sky show positive correlations: the wave affects them similarly. As the angular separation grows, the correlation falls, eventually crossing zero, then dipping negative, then rising back toward a slight positive value at the largest separations. This particular curved shape — quasi-quadrupolar, meaning it has a specific angular structure that's neither a simple monopole nor a dipole — is the fingerprint of gravitational waves. Why does that shape matter so much? Because other noise sources produce very different patterns. A global clock error would shift every pulsar the same way — a monopole signature. Errors in our models of the solar system produce dipolar correlations. The interstellar medium affects pulsars differently depending on their direction and distance. None of these mimic the Hellings-Downs curve. Finding that curve in the data is the smoking gun. NANOGrav's fifteen-year dataset contains two thousand two hundred eleven pulsar pairs. When the team reconstructed the interpulsar correlations empirically — binned by angular separation — the shape they found matched the Hellings-Downs prediction. The chi-squared of that reconstruction relative to the template was eight point one with a p-value of about zero point seventy five, meaning the data are thoroughly consistent with the expected curve. Now comes the statistics, and there are two independent lines of evidence worth walking through. The first is Bayesian. A Bayes factor is essentially a ratio of how well two models explain the data — a Bayes factor of one hundred means one model is one hundred times more probable than the other given what you observed. The NANOGrav team compared a model with a gravitational wave background to a model with only independent pulsar noise, and the Bayes factor exceeded ten to the fourteenth power. That is a staggeringly decisive number. But the more targeted test is the one that specifically asks: are the correlations following the Hellings-Downs pattern, versus just being some common uncorrelated noise process? That comparison — Hellings-Downs model versus an uncorrelated common-spectrum model — yielded Bayes factors of two hundred to one thousand depending on modeling choices. A predictive cross-validation measure called the pseudo-Bayes factor came in at one thousand four hundred in favor of the Hellings-Downs model. The team then stress-tested those Bayes factors by building empirical null distributions — essentially asking what Bayes factors you would get if the Hellings-Downs correlations weren't real. They did this using a clever trick called phase shifts: they scrambled the Fourier components of the common signal across pulsars, destroying the real interpulsar coherence while leaving each pulsar's individual power intact. Running this scramble thousands of times, only five out of five thousand trials produced a Bayes factor as large as the one observed in real data. That gives an empirical p-value of ten to the minus three, or roughly three sigma. The second line of evidence is frequentist. The team built what's called a noise-marginalized optimal statistic: a weighted, template-matched sum of all the pairwise cross-correlations, tuned specifically to the Hellings-Downs pattern. This statistic reached a mean signal-to-noise ratio of about five when marginalizing over the common-spectrum parameters, and about four when fixing the spectral index. To calibrate significance, they ran four hundred thousand phase-shift trials. Only nineteen produced a signal-to-noise as high as observed — yielding a frequentist p-value of five times ten to the minus five. Independent simulation-based and analytic calculations gave consistent values around one point nine times ten to the minus four. Sky scrambles, which randomize the pulsar positions to destroy the angular dependence of the correlations, gave p-values below ten to the minus four. Collectively, these correspond to statistical significance between three point five and four sigma — strong evidence, though not yet the gold standard five sigma of particle physics. So what is making this gravitational wave hum? The leading candidate is a cosmic population of supermassive black hole binaries — pairs of black holes, each millions or billions of times the mass of the sun, brought together when their host galaxies merged. As each pair slowly inspirals, it radiates gravitational waves. Across billions of galaxies and cosmic time, those individual signals blur together into a stochastic background, a constant hum from every direction. The expected spectral signature from such a population follows a clean power law: characteristic strain scales as frequency to the negative two-thirds power. That's the theoretical prediction, and it fits. Assuming that fiducial spectrum, the collaboration measured a characteristic strain amplitude of two point four times ten to the minus fifteen at a reference frequency of one cycle per year, with a ninety percent credible interval running from one point eight to three point one times ten to the minus fifteen. The spectral slope inferred from the data is consistent with the minus two-thirds prediction. Both the amplitude and the shape are in line with astrophysical expectations for a supermassive black hole binary population. That said, the team is careful not to overstate this attribution. Exotic sources — cosmic strings from the early universe, cosmological phase transitions — cannot yet be ruled out. The spectral shape alone isn't sufficient to uniquely identify the source. Future measurements of anisotropy in the background, or detections of individual loud binary systems, will be needed to nail down the origin. What makes this result especially compelling is that NANOGrav didn't find it alone. The European Pulsar Timing Array, the Parkes Pulsar Timing Array in Australia, the Indian Pulsar Timing Array, and the Chinese Pulsar Timing Array all reported similar findings simultaneously in 2023. Independent instruments, independent datasets, independent analysis pipelines — all converging on the same signal. The International Pulsar Timing Array's forthcoming combined dataset, with roughly eighty pulsars and timing baselines up to twenty-four years, will push sensitivity further. This is the second confirmed method for detecting gravitational waves, complementing LIGO and Virgo — but at frequencies roughly a billion times lower. LIGO hears the final milliseconds of black hole mergers. Pulsar timing arrays hear the slow, years-long orbital cycles of the most massive objects in the universe. Together, they're opening different octaves of a gravitational symphony that we're only just beginning to listen to. This lecture was created by ennepō. Go to https://ennepo.ai to Discover, Create and Follow the latest research in your field. Read when you can. Listen when you want to.