Literature cutoff: 24 October 2026.

On 28–29 June 2023, four collaboration papers appeared together: NANOGrav (North American Nanohertz Observatory for Gravitational Waves) [[1]], the EPTA (European Pulsar Timing Array) with InPTA (Indian Pulsar Timing Array) data [[2]], the PPTA (Parkes Pulsar Timing Array, Australian) [[3]], and the CPTA (Chinese Pulsar Timing Array) [[4]]. They reported evidence for a stochastic gravitational-wave background at nanohertz frequencies. The four analyses used different data sets and pipelines; the EPTA and InPTA result was a joint analysis, not a fifth independent measurement.

The detector for all four collaborations was not a machine. It was the Milky Way — specifically, a collection of millisecond pulsars scattered across the galaxy, separated by thousands of light-years. The arms of this instrument make LIGO look microscopic by comparison, and they are pointed in every direction at once.

There is a professional habit, well established in physics, of finding the scale of things either comforting or terrifying. This one is genuinely awe-inspiring.

The scale problem

LIGO detects gravitational waves in the frequency range of roughly 10–1000 Hz. A wave at 100 Hz has a wavelength of

$$\lambda = \frac{c}{f} = \frac{3 \times 10^8 \text{ m/s}}{100 \text{ Hz}} \approx 3000 \text{ km},$$

about a quarter of the Earth’s diameter. LIGO’s 4-km Fabry-Pérot arm cavities are short compared to the wavelength, but the cavities (finesse about 450) fold the light so that it effectively traverses the arm roughly 280 times, an effective path length of about 1,100 km [[5]], and the interferometer is exquisitely sensitive to differential arm-length changes of order $10^{-19}$ m — a fraction of the proton radius. The physics is heroic. But it works at the scale of stellar-mass binary mergers: neutron star–neutron star, neutron star–black hole, stellar-mass black hole–black hole. Those systems radiate at audible frequencies. LIGO famously converts its strain signals to sound, and the chirps do sound like something from a science fiction film.

Now consider pulsar timing arrays. The targets are nanohertz gravitational waves: frequencies of order $f \sim 10^{-9}$ to $10^{-8}$ Hz. At $10^{-9}$ Hz, the wavelength is

$$\lambda = \frac{c}{f} \approx \frac{3 \times 10^8 \text{ m/s}}{10^{-9} \text{ Hz}} = 3 \times 10^{17} \text{ m} \approx 9.7 \text{ pc} \approx 32 \text{ light-years}.$$

There is no general rule that a gravitational-wave detector must be as large as the wavelength it observes, and neither instrument is. LIGO’s arms are $10^{-4}$ to $10^{-2}$ of its wavelengths; its sensitivity comes from interferometry, long effective optical paths, and extraordinary displacement precision. A pulsar timing array errs the other way. The nearest pulsar in NANOGrav’s set, PSR J0437$-$4715, is 157 pc away, about 510 light-years; the farthest are around 7 kpc, more than 20,000 light-years [[6]]. Over the 2–28 nHz band of NANOGrav’s gravitational-wave fit [[1]], that makes every arm, by my arithmetic, about 32 to 20,000 wavelengths long (Fig. 1).

Figure 1. Interferometers are shorter than the waves they detect; a pulsar timing array is far longer. Bars span each band at the arm length; the PTA box spans the nearest and farthest NANOGrav pulsars [[6]] over the 2–28 nHz band of its gravitational-wave fit [[1]].

An arm of about a wavelength already gives the full response, of order $h/(2\pi f)$; longer arms do not enlarge it. What the extra length buys is a clean correlation: the far ends of the arms, where the pulsar terms arise, lie many wavelengths apart, so those terms average out between pulsars and only the Earth ends share the wave (see the Hellings-Downs section). By accident of the galaxy’s size and the distribution of recycled pulsars within it, we happen to live inside an instrument whose arms are comfortably longer than the waves it is after. The galaxy did not plan this. We got lucky.

Millisecond pulsars as cosmic clocks

A gravitational wave detector is only as good as its clock. LIGO measures differential length changes in its arm cavities using the interference of laser light — effectively using the constancy of the speed of light as a metronome. Pulsar timing arrays use a different clock: the rotation of a neutron star.

Ordinary pulsars are the collapsed remnants of massive stars, born in supernova explosions. A neutron star of roughly 1.4 solar masses is compressed into a sphere about 25 km across (NICER X-ray timing gives radii of 12.7 and 13.0 km for PSR J0030+0451 [[7]] [[8]]), rotating rapidly and radiating beams of radio waves from near its magnetic poles. The rotation periods of young pulsars are typically of order seconds, and they spin down on timescales of millions of years as they lose rotational energy to magnetic dipole radiation. They are not particularly stable clocks — the spin-down is uneven, and they occasionally “glitch,” suddenly spinning up by a small amount.

Millisecond pulsars (MSPs) are a different beast entirely. They have been recycled: spun up to near-millisecond rotation periods by accreting matter from a binary companion star. The accretion process deposits angular momentum onto the neutron star, increasing its spin rate dramatically, while simultaneously burying and weakening its magnetic field. The typical MSP has a surface magnetic field of order $10^8$–$10^9$ G, three to four orders of magnitude weaker than the roughly $10^{12}$ G typical of a young pulsar. Since the spin-down torque scales as $B^2$, the weaker field means the MSP loses rotational energy far more slowly. Once the accretion stops — when the companion has exhausted its transferable mass — the MSP is left spinning rapidly and stably, with a spin-down rate of order $10^{-20}$ s/s.

The result is rotational stability competitive with terrestrial atomic clocks. PSR J0437$-$4715, one of the best-timed MSPs, has a rotation period of $P \approx 5.76$ ms and a timing residual — the scatter of individual pulse arrival times around the best-fit timing model — of order 100 ns over years of observation [[9]]. For a pulsar completing about 174 rotations per second, a residual of 100 ns over a baseline of years is remarkable. Held for a decade, $3 \times 10^8$ s, that would be a fractional stability of about $3 \times 10^{-16}$, my extrapolation from 4.8 years of NANOGrav timing. These are not merely good clocks; they are among the most stable periodic phenomena known to physics.

The timing model accounts for everything we know about the pulsar’s environment: its spin period and spin-down rate, its proper motion across the sky, parallax (from which we get the distance), Shapiro delay from any companions, dispersion measure variations in the interstellar medium, and more. After subtracting all modelled effects, what remains are timing residuals — small, unexplained deviations in the pulse arrival times. If a gravitational wave passes through, it will appear in those residuals.

How a gravitational wave shifts a pulse arrival time

A gravitational wave is a propagating perturbation of the spacetime metric. In the transverse-traceless (TT) gauge, a wave propagating along the $z$-direction perturbs the metric as

$$ds^2 = -c^2 dt^2 + \bigl(1 + h_+\bigr)dx^2 + \bigl(1 - h_+\bigr)dy^2 + 2h_\times \, dx \, dy + dz^2,$$

where $h_+$ and $h_\times$ are the two polarisation amplitudes, both functions of $t - z/c$. As the wave passes, the proper distances in the $x$- and $y$-directions oscillate: space itself is being stretched and squeezed at the wave frequency.

A radio pulse travelling from a pulsar to Earth traverses this oscillating spacetime. The proper path length changes, and so does the travel time. For a pulsar at distance $L$ in the direction of the unit vector $\hat{p}$ and a plane wave travelling in direction $\hat{k}$, the Fourier component of the timing residual at frequency $f$ is [[10]]

$$\tilde{R}(f) = \frac{\hat{p}^i \hat{p}^j \tilde{h}_{ij}(f)}{2} \, \frac{1 - e^{-2\pi i f L (1 + \hat{k}\cdot\hat{p})/c}}{2\pi i f \, (1 + \hat{k}\cdot\hat{p})}, \tag{1}$$

where $\tilde{h}_{ij}(f)$ is the transverse-traceless strain at Earth and $\hat{p}^i \hat{p}^j \tilde{h}_{ij}$ (summed over $i, j$) its projection on the line of sight. The factor $1/(2\pi i f)$ integrates the fractional frequency shift $\Delta\nu/\nu$ into a time delay. The 1 in the numerator is the Earth term, the strain at Earth when the pulse arrives; the exponential is the pulsar term, the strain at the pulsar when it left.

For $fL/c \ll 1$ the residual is of order $hL/c$ and grows with the arm: LIGO’s regime, and LISA’s over most of its band. For a pulsar, $fL/c$ is about 32 to 20,000, the numerator never exceeds 2, and a sinusoidal wave of strain amplitude $h$ leaves a residual of order $h/(2\pi f)$ whatever the distance. The observing span determines which low frequencies can be separated from fitted timing-model trends; it is not an extra factor in this single-wave amplitude.

For a single pulsar, a timing residual tells you that something disturbed the spacetime between Earth and the pulsar — but it could be a systematic in the timing model, interstellar medium fluctuations, or intrinsic pulsar noise. You cannot claim a gravitational wave detection from one pulsar alone. What you need is the correlation between many pulsars.

The Hellings-Downs curve

Here is the central idea of pulsar timing array science. Consider two pulsars in directions $\hat{n}_a$ and $\hat{n}_b$, separated on the sky by an angle $\theta$ such that $\cos\theta = \hat{n}_a \cdot \hat{n}_b$. Both are embedded in the same stochastic gravitational wave background — a superposition of waves arriving from all directions, at all frequencies in the nanohertz band, with random phases and amplitudes. The timing residuals of both pulsars will be perturbed by this background. The question is: what is the expected cross-correlation between their residuals?

Hellings and Downs (1983) computed this [[11]], assuming the background is isotropic (equal power from all directions), unpolarised, and stationary, and that the pulsars are many wavelengths away: the pulsar terms of Eq. (1) then oscillate rapidly across the sky and drop out of the cross-correlation of two distinct pulsars [[10]]. The answer is now called the Hellings-Downs (HD) curve:

$$\Gamma(\theta) = \frac{3}{2}\, x \ln x - \frac{x}{4} + \frac{1}{2}, \qquad x = \frac{1-\cos\theta}{2}. \tag{2}$$

This is NANOGrav’s normalisation [[1]]. A pulsar correlated with itself gets an extra $1/2$ from its own pulsar term, making the autocorrelation one [[10]]. Let me unpack the features of this function:

  • At $\theta = 0$ for two distinct pulsars in the same direction: $\Gamma(0) = 1/2$, the maximum cross-correlation.
  • At $\theta \approx 49.3°$: the curve crosses zero and turns negative (anticorrelated).
  • At $\theta \approx 82.5°$, where $d\Gamma/dx = 0$ gives $x = e^{-5/6}$: the minimum, $\Gamma \approx -0.152$; at right angles, $\Gamma(\pi/2) \approx -0.145$.
  • At $\theta \approx 121.8°$: the curve crosses zero again.
  • At $\theta = \pi$ (antipodal pulsars, opposite directions on the sky): $\Gamma(\pi) = 1/4$. Positive correlation even for pulsars in opposite directions — counterintuitive but correct.

The shape is dominated by the quadrupole: its Legendre series in $\cos\theta$ starts at $\ell = 2$ [[10]], so it contains no monopole and no dipole. Under the usual assumptions of an isotropic, unpolarised tensor background in general relativity, it arises from spin-2 gravitational waves. A clock error common to all pulsars would produce a flat, angle-independent correlation; an error in Earth’s ephemeris would produce an approximately dipolar, $\cos\theta$ pattern (Fig. 2). These are useful competing models, which is why PTA analyses fit them rather than merely looking for low-frequency noise.

Figure 2. The Hellings-Downs curve changes sign twice and ends positive; it has no monopole or dipole component, so its pattern differs from both a clock error and an ephemeris error (NANOGrav’s data show an inconclusive hint of extra monopole power near 3.95 nHz [[1]]). Drawn: Eq. (2) for distinct pulsars, and the monopole ($\Gamma = 1$) and dipole ($\cos\theta$) alternatives NANOGrav tests [[1]] [[12]], scaled to 1/2 at zero separation because each has a free amplitude in a fit.

This is why the HD curve is the central discriminant. Cross-correlations with this specific angular dependence are evidence for the tensor-background model, but their strength and robustness still depend on noise modelling, sampling, and the alternatives tested.

The difficulty is statistical. You need enough pulsars, spread over a wide range of sky angles, each timed with sufficient precision, over a long enough baseline, that you can measure this correlation function with confidence. That has been the programme of the PTA collaborations since the 1990s. In June 2023, they had enough.

The 2023 evidence

The NANOGrav 15-year dataset [[9]] contains observations of 68 millisecond pulsars obtained between July 2004 and August 2020; the gravitational-wave-background analysis uses the 67 of them with a timing baseline of at least three years [[1]]. (The data set is called “15-year” although the first and last observations are 16.03 years apart, because no single pulsar exceeds 16 years of observation.) The dataset represents an enormous investment of telescope time — primarily at the Arecibo Observatory, whose data end in August 2020, four months before its collapse, and the Green Bank Telescope, with some from the Very Large Array [[9]].

The analysis found an excess of low-frequency noise common to many pulsars, consistent with a power-law spectral shape. More importantly, when the cross-correlations between all 2,211 pairs of pulsars were computed and binned by angular separation, the result was consistent with the Hellings-Downs curve. The evidence that the correlations follow this pattern rather than spatially uncorrelated common noise came in several forms [[1]]: a Bayes factor of 200 to 1,000, depending on the spectral modelling; a false-alarm probability of $10^{-3}$ (about $3\sigma$) for that Bayes factor; and $p = 5 \times 10^{-5}$ to $1.9 \times 10^{-4}$ (about 3.5–4$\sigma$) for a frequentist statistic built from the weighted pair correlations. Posterior-predictive versions of the optimal statistic and of the Bayesian model comparison later gave $3.2\sigma$ and $3\sigma$ [[12]]. The collaboration carefully described this as “evidence for” rather than “detection of” a gravitational wave background, following community conventions (detection would conventionally require $\geq 5\sigma$).

Simultaneously, the EPTA published its second data release, with InPTA data for ten pulsars [[2]], the PPTA published its results [[3]], and the Chinese Pulsar Timing Array (CPTA), the fourth collaboration, published results from its first data release [[4]]. Each group used its own datasets, different telescopes, different software pipelines, different statistical methodologies. All found a common low-frequency signal; the Hellings-Downs evidence varied. The EPTA reported a Bayes factor of 60 and a false-alarm probability of about 0.1% ($\gtrsim 3\sigma$) in the 10.3 years from its newer observing systems, but only a Bayes factor of four (false-alarm probability 4%) in its full 24.7 years [[2]]; the PPTA, 30 pulsars over 18 years, $p \lesssim 0.02$, about $2\sigma$ [[3]]; the CPTA, 57 pulsars timed with FAST for 3.4 years, $4.6\sigma$ for Hellings-Downs against a constant correlation, in one frequency bin near 14 nHz [[4]].

The agreement among the four contemporaneous analyses is important, but it does not turn their evidence into a single independent four-way measurement: their data volumes, pulsar samples, telescopes, and analysis choices differ, and the EPTA–InPTA analysis is joint. The result is therefore best described, as the collaborations did, as convergent evidence for a background rather than a completed proof of its origin.

The combined interpretation is that the timing data contain a common low-frequency signal with spatial correlations consistent with a tensor gravitational-wave background. The Milky Way is a gravitational-wave detector, and it has measured something.

What is making the noise?

The million-dollar question. A population of supermassive black hole binaries (SMBHBs) on circular orbits, driven purely by gravitational-wave emission and spread across a wide range of masses and redshifts, predicts the spectral shapes of the characteristic strain $h_c$, the timing-residual power spectral density $S$ and the energy density $\Omega_{\rm GW}$:

$$h_c(f) = A \left(\frac{f}{1 \text{ yr}^{-1}}\right)^{-2/3}, \qquad S(f) = \frac{h_c(f)^2}{12\pi^2 f^3} \propto f^{-13/3}, \qquad \Omega_{\rm GW}(f) = \frac{2\pi^2 f^2 h_c(f)^2}{3 H_0^2} \propto f^{2/3},$$

where $A$ is the strain amplitude at $f = 1 \text{ yr}^{-1}$ and $H_0$ the Hubble constant. With this slope fixed, NANOGrav measures $A = 2.4^{+0.7}_{-0.6} \times 10^{-15}$ [[1]]. Left free, the residual spectrum is shallower, $S \propto f^{-\gamma}$ with $\gamma = 3.2 \pm 0.6$ (90% credible intervals), which puts $13/3$ on the boundary of the 99% region: a “moderate tension” that could reflect the binaries’ interactions with stars and gas, and that depends on the noise modelling of a few pulsars [[1]].

Here is the astrophysical picture. When two massive galaxies merge — an event that happens billions of times over cosmic history — their central supermassive black holes (each typically $10^7$–$10^{10}$ solar masses) do not immediately merge. First, they sink toward the centre of the merger remnant by dynamical friction against the stellar background, forming a loosely bound binary on scales of a parsec or so. The binary then hardens: three-body interactions with individual stars passing close to the binary extract orbital energy, driving the pair to smaller separations. Eventually, when the binary has hardened to a separation where gravitational wave emission dominates the energy loss, the pair inspirals and merges on a timescale of millions to billions of years.

The incoherent superposition of gravitational wave emission from the large population of SMBHB systems across the observable universe — at all masses, at all orbital frequencies, at all redshifts — produces a stochastic background. It is, in a sense, cosmic traffic noise: no individual merger is detectable as a discrete event, but the combined hum from all of them is. The predicted spectral amplitude and shape from this population are broadly consistent with the observed signal, though the uncertainties on the astrophysical model are large enough that the agreement is not a precise test.

Alternative sources are possible. A first-order phase transition in the early universe (such as a QCD or electroweak phase transition) would produce a background of gravitational waves with a different spectral shape — more peaked and potentially harder than the SMBHB prediction. A network of cosmic strings, topological defects from symmetry-breaking phase transitions in the early universe, would produce yet another spectrum, approximately flat in $\Omega_{\rm GW}(f)$. Primordial gravitational waves from inflation are expected at far lower frequencies, near the CMB scale, but with a nearly scale-invariant spectrum that could contribute at nanohertz frequencies in some models. The 2023 data do not single out the SMBHB interpretation. In NANOGrav’s comparison, every cosmological model tested except stable field-theory cosmic strings could reproduce the signal, and many fitted better than the standard binary population, with Bayes factors of 10 to 100 — a preference that, the collaboration stresses, depends on assumptions about the binaries and is not evidence for new physics [[13]]. SMBHBs remain the expected source, but early-universe contributions — or combinations of both — cannot be ruled out. Future data with longer baselines and more pulsars will tighten the spectral measurements and may reveal deviations from the SMBHB prediction.

There is also the tantalising prospect of eventually resolving individual SMBHB systems above the stochastic background — the gravitational wave equivalent of resolving individual radio sources above the extragalactic background. No individual SMBHB has yet been identified in PTA data; a targeted NANOGrav search of 114 active galactic nuclei proposed as binary hosts found none, and narrowed the parameters allowed for a putative binary in 3C 66B [[14]].

Multi-frequency gravitational wave astronomy

We are now, unambiguously, in the era of multi-frequency gravitational wave astronomy. The universe produces gravitational waves across an enormous span of frequencies, and different frequency windows probe fundamentally different source populations:

  • Primordial / CMB B-modes (~$10^{-17}$ Hz): Wavelengths comparable to the Hubble scale. Primordial tensor perturbations from inflation would imprint a distinctive B-mode polarisation pattern in the cosmic microwave background. No confirmed detection yet; the BICEP/Keck programme and future experiments such as LiteBIRD are sensitive to this regime.

  • Pulsar timing arrays (~$10^{-9}$–$10^{-8}$ Hz): The nanohertz band, just accessed by NANOGrav and partners. Sources: SMBHB inspirals, possibly early-universe phase transitions and cosmic strings. Arm length: hundreds to tens of thousands of light-years.

  • LISA (~$10^{-4}$–$10^{-1}$ Hz): The millihertz band. The LISA space interferometer (planned launch in the 2030s) will have arm lengths of 2.5 million km and will detect SMBHB mergers directly as they happen, stellar-mass compact binary inspirals years before their LIGO-band merger, and possibly signals from extreme mass-ratio inspirals. LISA will see lighter massive black-hole binaries (roughly $10^4$–$10^7$ solar masses) than the $10^8$–$10^{10}$ solar-mass binaries that dominate the PTA signal, through their final inspiral and merger.

  • LIGO/Virgo/KAGRA (~10–$10^3$ Hz): The audible band. Stellar-mass black hole and neutron star mergers. 90 probable compact-binary merger candidates in the first three observing runs [[15]], with many more in O4. Source masses: $\sim$1–100 $M_\odot$.

This is the gravitational wave analogue of multi-wavelength astronomy. Just as the universe looks completely different in radio, infrared, optical, X-ray, and gamma-ray light — each wavelength band revealing different physical processes and source populations — the universe sounds completely different at each gravitational wave frequency. PTAs hear the rumble of cosmic structure formation; LISA will hear the whisper of the final inspiral and merger of lighter massive black holes; LIGO hears the sharp crack of stellar-mass collisions.

The 2023 announcements represent the opening of the nanohertz window. We have gone from one gravitational wave frequency band (LIGO/Virgo) to two. The next decade, with LISA launching and PTAs continuing to accumulate data, will see the opening of a third.

A brief detour: LIGO’s O4 run and the mass-gap object

While the PTA collaborations were making their June 2023 announcement, LIGO’s fourth observing run (O4, 24 May 2023 to 18 November 2025) was proceeding at an extraordinary rate. The network issued roughly 250 public candidate alerts during the run. This is now a production science rather than an exploration.

Among the most scientifically interesting events was GW230529, detected on 29 May 2023 by LIGO Livingston alone and published by the LIGO, Virgo and KAGRA collaborations [[16]]. The signal is consistent with the merger of a neutron star with a compact object whose mass was measured to be approximately $2.5$–$4.5\ M_\odot$. That puts it in, or at the lower edge of, what theorists call the “mass gap”, which the paper takes as about 3–5 $M_\odot$: heavier than the heaviest neutron stars ($\lesssim 2.3\ M_\odot$, though this upper limit is uncertain), lighter than the lightest stellar-mass black holes inferred from X-ray binaries ($\gtrsim 5\ M_\odot$).

Whether GW230529’s heavier component was the heaviest neutron star ever observed, or the lightest black hole, is not settled. Given current estimates of the maximum neutron-star mass, the collaborations’ most probable interpretation is a black hole in the gap, but they state that the gravitational-wave data alone cannot determine whether either component is a neutron star or a black hole [[16]]. The distinction matters enormously for nuclear physics: if it is a neutron star, it constrains the nuclear equation of state at supranuclear densities. If it is a black hole, it means the mass gap is narrower than X-ray observations suggested, and our understanding of compact object formation needs revision. Tides could tell them apart, since a black hole’s tidal deformability is zero; the heavier component’s tidal deformability peaks at zero but is also consistent with neutron-star matter. The paper reports no electromagnetic counterpart. This question will likely not be settled until we have electromagnetic constraints from similar systems, or until we accumulate enough mass-gap events to understand the population statistically.

The instrument we did not build

I want to return to the pulsar timing array concept, because I think it deserves more than a passing technical description. The idea is this: nature has distributed a set of extremely stable clocks across the galaxy. We did not put them there. We did not design them. We simply discovered that neutron stars, after a particular evolutionary pathway involving mass transfer from a binary companion, achieve a rotational stability that happens to be sufficient to detect perturbations of spacetime at cosmological scales.

The instrument is the galaxy itself — or rather, our ability to model it. We build a timing model for each pulsar: a comprehensive description of every known effect that influences pulse arrival times. We subtract the model. What remains, the residuals, contains the signal we cannot yet explain. We cross-correlate the residuals of 67 (or 25, 30, or 57, depending on the collaboration) pulsars in pairs, compute the correlation as a function of angular separation, and compare to the Hellings-Downs prediction.

The engineering challenge is not building the detector. It is characterising it. Understanding the noise. Modelling the interstellar medium, which disperses radio pulses in a frequency-dependent way and varies as the pulsar moves through clouds of ionised gas. Accounting for clock errors in the terrestrial time standards used to timestamp observations. Dealing with instrumental noise in each of the different radio telescopes that contribute data. Building a Bayesian framework that can simultaneously fit the timing model parameters, the pulsar noise properties, and the GWB parameters for dozens of pulsars.

This is painstaking, years-long work. The timing spans of the 68 pulsars in the 15-year NANOGrav dataset add up to about 580 pulsar-years [[9]]. The evidence is earned.

The Hellings-Downs correlation — the specific pattern sought in those residuals, consistent with the quadrupolar fingerprint of general relativity’s spin-2 gravitational waves — is one of the more beautiful results I have seen in recent astrophysics. It is evidence for that tensor background at frequencies far below LIGO’s band, using a detector assembled by the Milky Way over the course of 10 billion years of stellar evolution and galaxy mergers.

We are in an age of gravitational wave astronomy. I find that remarkable.

If you are interested in the broader theme of using astronomical observations as physics experiments rather than just cataloguing the sky, the posts on transit photometry and the gift of transits and on smartphone-based exoplanet observations cover similar ground at different scales — planetary radii measured in units of the star’s radius, by timing the dimming of a star as a planet crosses its face. The underlying logic is the same: precision timing plus a physical model plus statistics equals a measurement of something you could not directly touch.

Literature and calculations checked through 2026-10-24.


References

[1] Agazie, G., et al. (NANOGrav Collaboration). (2023). The NANOGrav 15 yr Data Set: Evidence for a Gravitational-wave Background. The Astrophysical Journal Letters, 951, L8. DOI: 10.3847/2041-8213/acdac6

[2] Antoniadis, J., et al. (EPTA Collaboration and InPTA Collaboration). (2023). The second data release from the European Pulsar Timing Array — III. Search for gravitational wave signals. Astronomy & Astrophysics, 678, A50. DOI: 10.1051/0004-6361/202346844

[3] Reardon, D. J., et al. (PPTA Collaboration). (2023). Search for an isotropic gravitational-wave background with the Parkes Pulsar Timing Array. The Astrophysical Journal Letters, 951, L6. DOI: 10.3847/2041-8213/acdd02

[4] Xu, H., et al. (CPTA Collaboration). (2023). Searching for the nano-Hertz stochastic gravitational wave background with the Chinese Pulsar Timing Array Data Release I. Research in Astronomy and Astrophysics, 23(7), 075024. DOI: 10.1088/1674-4527/acdfa5

[5] Aasi, J., et al. (LIGO Scientific Collaboration). (2015). Advanced LIGO. Classical and Quantum Gravity, 32, 074001. DOI: 10.1088/0264-9381/32/7/074001

[6] Manchester, R. N., Hobbs, G. B., Teoh, A., & Hobbs, M. (2005). The Australia Telescope National Facility Pulsar Catalogue. The Astronomical Journal, 129, 1993–2006. DOI: 10.1086/428488. Catalogue version 2.6.1, atnf.csiro.au/research/pulsar/psrcat, queried 24 October 2026.

[7] Riley, T. E., Watts, A. L., Bogdanov, S., et al. (2019). A NICER view of PSR J0030+0451: Millisecond pulsar parameter estimation. The Astrophysical Journal Letters, 887, L21. DOI: 10.3847/2041-8213/ab481c

[8] Miller, M. C., Lamb, F. K., Dittmann, A. J., et al. (2019). PSR J0030+0451 mass and radius from NICER data and implications for the properties of neutron star matter. The Astrophysical Journal Letters, 887, L24. DOI: 10.3847/2041-8213/ab50c5

[9] Agazie, G., et al. (NANOGrav Collaboration). (2023). The NANOGrav 15 yr Data Set: Observations and Timing of 68 Millisecond Pulsars. The Astrophysical Journal Letters, 951, L9. DOI: 10.3847/2041-8213/acda9a

[10] Romano, J. D., & Cornish, N. J. (2017). Detection methods for stochastic gravitational-wave backgrounds: a unified treatment. Living Reviews in Relativity, 20, 2. DOI: 10.1007/s41114-017-0004-1

[11] Hellings, R. W., & Downs, G. S. (1983). Upper limits on the isotropic gravitational radiation background from pulsar timing analysis. The Astrophysical Journal, 265, L39–L42. DOI: 10.1086/183954

[12] Agazie, G., et al. (NANOGrav Collaboration). (2025). The NANOGrav 15 yr dataset: Posterior predictive checks for gravitational-wave detection with pulsar timing arrays. Physical Review D, 111, 042011. DOI: 10.1103/PhysRevD.111.042011

[13] Afzal, A., et al. (NANOGrav Collaboration). (2023). The NANOGrav 15 yr Data Set: Search for Signals from New Physics. The Astrophysical Journal Letters, 951, L11. DOI: 10.3847/2041-8213/acdc91

[14] Agarwal, N., et al. (NANOGrav Collaboration). (2026). The NANOGrav 15 yr Dataset: Targeted Searches for Supermassive Black Hole Binaries. The Astrophysical Journal Letters, 998, L11. DOI: 10.3847/2041-8213/ae3719

[15] Abbott, R., et al. (LIGO Scientific, Virgo and KAGRA Collaborations). (2023). GWTC-3: Compact Binary Coalescences Observed by LIGO and Virgo during the Second Part of the Third Observing Run. Physical Review X, 13, 041039. DOI: 10.1103/PhysRevX.13.041039

[16] Abac, A. G., et al. (LIGO Scientific, Virgo and KAGRA Collaborations). (2024). Observation of Gravitational Waves from the Coalescence of a 2.5–4.5 $M_\odot$ Compact Object and a Neutron Star. The Astrophysical Journal Letters, 970, L34. DOI: 10.3847/2041-8213/ad5beb


Changelog

  • 2025-12-01: Corrected the summary from “four independent pulsar timing arrays” to “five” — the CPTA (Chinese Pulsar Timing Array) also published consistent results in June 2023 and is counted as the fifth group in the body text.
  • 2026-07-11: Corrected the 2023-publication account: four papers reported results from NANOGrav, EPTA/InPTA jointly, PPTA, and CPTA. Replaced several “detection” and “direct measurement” formulations with the collaborations’ evidence-level language; corrected the single-wave timing-residual scaling and the O4 end date.
  • 2026-10-03: Corrected the millisecond-pulsar surface magnetic field from $10^4$–$10^5$ G to $10^8$–$10^9$ G (and the contrast with young pulsars to three to four orders of magnitude, against about $10^{12}$ G); removed the leftover “fifth collaboration” and “five-way” wording for CPTA, which contradicted the corrected four-paper account; and corrected the LISA description, since LISA targets lighter ($10^4$–$10^7$ solar-mass) massive black-hole binaries than the PTA band, not the same sources at a later stage. Corrected the LIGO effective arm length from about 1600 km to about 1,100 km (arm-cavity finesse about 450, roughly 280 passes of 4 km; Aasi et al., 2015); stated the NANOGrav 15-year figures exactly (68-pulsar data set, 67 pulsars in the gravitational-wave-background analysis, 16.03 years between first and last observation) and removed the unsourced “roughly 2200 observations per pulsar”.
  • 2026-10-24: Reconciled the scale argument, which said both that no detector must match its wavelength and that a pulsar timing array needs arms of “approximately the right dimensions”: arms longer than about a wavelength saturate the response at order $h/(2\pi f)$ and let the pulsar terms average out of the cross-correlations. The schematic $R \propto \int h\,dt$ is replaced by the explicit response, Eq. (1) (Romano & Cornish 2017), and Fig. 1 compares arm lengths with wavelengths. Corrected the 1-nHz wavelength (10 pc ≈ 33 → 9.7 pc ≈ 32 light-years), the pulsar distances (“a few hundred” to “several thousand” → about 510 to more than 20,000 light-years, ATNF catalogue) and the neutron-star size (“about 10 km across” → about 25 km; NICER radii of 12.71 and 13.02 km for PSR J0030+0451, Riley et al. and Miller et al. 2019). J0437’s 100 ns residual now holds “over years”, not “decades”, and “δP/P ∼ 10⁻¹⁴ or better” is replaced by about 3 × 10⁻¹⁶, marked as an extrapolation.
  • 2026-10-24: Gave the Hellings-Downs zeros and minimum (49.3°, 121.8°; −0.152 at 82.5°) and drew the curve against monopole and dipole patterns (Fig. 2). Replaced “3–4σ” with NANOGrav’s itemised significances (Bayes factor 200–1,000; about 3σ and 3.5–4σ) and “All found the same thing” with each collaboration’s, from marginal evidence in EPTA’s full 24.7-year set (false-alarm probability 4%) to 4.6σ in one CPTA frequency bin. Corrected the Arecibo date (its data end in August 2020, not at the December collapse). Replaced “the observed signal” scaling as $f^{-13/3}$ with NANOGrav’s fits (γ = 3.2 ± 0.6, with 13/3 at the 99% boundary); “most consistent with the SMBHB interpretation” with its new-physics comparison (Bayes factors 10–100, not evidence); “many billions” of binaries with “the large population”; and the unsourced “upper limits on the most massive known candidate systems” with the targeted search of 114 active galactic nuclei.
  • 2026-10-24: Cited the 2025 posterior-predictive checks under their correct DOI (042011, not 024039, an unrelated paper). Corrected the GW230529 citation (Abac et al., not Abbott et al.); replaced an unsourced spin argument with the paper’s reading (a mass-gap black hole is most probable, tides do not decide it); and changed “sits squarely in” the mass gap to “in, or at the lower edge of” it (2.5–4.5 against about 3–5 $M_\odot$). Also corrected “over 90” O3 events to 90 candidates (GWTC-3), about 600 to about 580 pulsar-years, the pulsar counts (the PPTA’s 30 were missing) and “stellar radii measured in units of stellar radii”; removed CMB-S4, which lost DOE and NSF support in July 2025. Citations are now numbered and linked.