The second data release from the European Pulsar Timing Array

Millisecond pulsars are nature's most precise celestial clocks. These rapidly spinning neutron stars emit radio pulses with such regularity that tiny deviations in their arrival times — measured against atomic clocks tied to international time standards — can be detected with extraordinary sensitivity. When a gravitational wave passes between a pulsar and Earth, it stretches and squeezes spacetime, producing small correlated advances or delays in those arrival times. Observe enough of these clocks scattered across the sky, and the network behaves like a single, galaxy-scale gravitational wave detector. This detector is sensitive to nanohertz frequencies — cycles per decade to cycles per year — far below what ground-based interferometers like LIGO can hear. The nanohertz regime is where supermassive black hole binaries, or SMBHBs, are expected to be loudest. In hierarchical galaxy formation, mergers between galaxies carrying supermassive black holes at their centers inevitably produce these massive pairs. The incoherent superposition of gravitational waves from countless such binaries across cosmic history generates a stochastic gravitational wave background, a kind of cosmic hum whose amplitude and spectral shape encode information about galaxy merger history and binary dynamics. Foster and Backer proposed the modern pulsar timing array in 1990 precisely to exploit this idea across many pulsars simultaneously. The definitive fingerprint of this background is the Hellings-Downs curve, first derived by Hellings and Downs in 1983. It is a specific relationship between the angular separation of two pulsars and the expected correlation of their timing residuals. Pulsar pairs at different angular separations on the sky should be correlated in a precise, predictable way — and that pattern is quadrupolar, meaning its power sits in the second and higher Legendre modes, not the zeroth or first. A clock error would corrupt all pulsars equally, producing a monopole correlation. Errors in the solar system ephemeris create a dipole. Only a genuine gravitational wave background produces the Hellings-Downs quadrupolar curve. Finding a common timing signal across all pulsars is necessary but not sufficient — you also have to find this specific spatial pattern. That is the threshold for a detection. The European Pulsar Timing Array, or EPTA, was formed in 2004 to clear that threshold using the continent's largest radio telescopes: the Lovell Telescope at Jodrell Bank, the Nançay decimetric radio telescope, the Westerbork Synthesis Radio Telescope, the Effelsberg one hundred meter dish, the Sardinia Radio Telescope, and the Large European Array for Pulsars, or LEAP, a coherently phased synthesis of these telescopes equivalent to a dish up to one hundred ninety-four meters across. The second data release from EPTA, presented by Antoniadis and more than one hundred collaborators in 2023, brings together timing data for twenty-five millisecond pulsars selected to capture roughly ninety-five percent of the expected sensitivity of a larger forty-two pulsar array. The collaboration analyzed four dataset combinations. DR2full spans the complete twenty-four point seven years of EPTA data, stretching back to 1994. DR2new is a deliberately trimmed ten point three year subset covering only observations from modern wideband, coherently dedispersing backends — cleaner hardware that produces more uniform data. Then there are two extended combinations that fold in observations from the Indian Pulsar Timing Array, or InPTA, which uses the upgraded Giant Metrewave Radio Telescope and contributes simultaneous multi-frequency coverage that helps disentangle interstellar medium effects from genuine gravitational wave signals. Noise modeling is central to all four. Each pulsar's data is carefully characterized for white noise, spin noise, dispersion measure variations, and scattering — each process modeled as a power law in frequency, with the number of Fourier components chosen pulsar by pulsar through Bayesian model selection rather than fixed in advance. Now for the results. Using DR2full, the twenty-four point seven year dataset, the evidence for a gravitational wave background is marginal: a Bayes factor of four in favor of a gravitational wave background over a common uncorrelated red noise alternative, and a false alarm probability of about four percent. A Bayes factor of four means the data are roughly four times more likely under the gravitational wave background hypothesis than under the null. That is suggestive but not convincing. The situation changes considerably with DR2new. The ten point three year modern backend subset yields a Bayes factor of about sixty and a false alarm probability of roughly zero point one percent — more than three sigma. That is a meaningful result. The spatial correlations support this. An optimal statistic analysis, which provides a frequentist-style cross-check on the Bayesian Bayes factors, finds a median signal-to-noise ratio for the Hellings-Downs correlation of about three point five in DR2new and only about one point three in DR2full. The DR2new result was reproduced across several independent analysis pipelines — ENTERPRISE and FORTYTWO — returning Bayes factors of sixty-six, fifty-six, and sixty-two in alternate runs. That is a strong consistency check. The team also used null distribution tests: phase-shifting the Fourier components of the data and scrambling pulsar sky positions to destroy cross pulsar correlations while preserving single pulsar noise. Those procedures placed the DR2new signal comfortably in the multi-sigma regime, though the exact p-values vary depending on the construction — ranging from about zero point zero zero zero five to zero point zero zero eight across different methods. The authors are candid about this spread and about limitations in the null distributions from incomplete simulations. The conservative quoted false alarm probability for DR2new is around zero point one percent. There is a genuine tension between the two datasets that deserves attention. When the spectral index of the common red process is left free, DR2full and DR2new prefer different slopes: DR2full gives an index of about four point nineteen with an uncertainty of plus zero point seventy-three minus zero point sixty-three, while DR2new returns a flatter index around two point seventy-one with larger errors. These are in mild disagreement. But here is the important clue: when the spectral index is fixed to thirteen divided by three — the canonical value expected for a population of circular, gravitational-wave-driven supermassive black hole binaries — the two datasets give consistent amplitudes. At that fixed index, the inferred characteristic strain amplitude at a reference frequency of one cycle per year is two point five times ten to the minus fifteen, with an uncertainty of plus or minus zero point seven times ten to the minus fifteen. The fact that a physically motivated spectral index reconciles the two datasets is telling. Adding InPTA data produces results broadly consistent with the EPTA-only analyses and improves noise modeling, particularly for characterizing interstellar propagation effects, but does not resolve the spectral tension between DR2full and DR2new. The collaboration treats this honestly: the full dataset is marginal, the modern backend subset is significant, and the spectral differences need further investigation before firm astrophysical conclusions are warranted. The leading interpretation is a stochastic gravitational wave background from a cosmic population of supermassive black hole binaries. Two features of the analysis support this. First, the spatial correlations in DR2new follow the Hellings-Downs curve. Second, the signal's spectral shape is consistent with the thirteen-over-three power law expected from circular, gravitational-wave-driven binaries in the inspiral phase. The spectral index discrepancy between datasets could arise from data quality differences, environmental coupling of some binaries, eccentric binary populations, or the presence of a few individually loud binaries — none of which rules out the SMBHB picture. Alternative explanations exist: early universe processes and certain dark matter models can also produce a nanohertz gravitational wave background, and the paper notes these briefly. But the spectrum and spatial pattern both point toward merging black holes as the dominant source. The natural next step is combining all global pulsar timing array efforts under the International Pulsar Timing Array, bringing together EPTA, NANOGrav, the Parkes Pulsar Timing Array, MeerKAT, and others into a single dataset with greater baseline and more pulsars. More data, more pulsars, and continued improvements in noise characterization will sharpen the Hellings-Downs measurement and constrain the spectral shape well enough to distinguish between source populations. What the EPTA second data release represents, in the broadest terms, is the first tentative hearing of a hum woven into spacetime by the mergers of the universe's most massive objects — a signal that has been traveling toward us across billions of years, written in the slow, correlated delays of twenty-five spinning neutron stars scattered across the Milky Way. The evidence is not yet definitive. But it is non-negligible, it is spatially structured the way it should be, and it points unmistakably in one direction. 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.