Pulsar Timing Array

The idea: turn the galaxy into an interferometer

A millisecond pulsar is a neutron star spun up to hundreds of rotations a second by binary accretion. The best of them have pulse-arrival timing precision of tens of nanoseconds over a decade — comparable to the world's best atomic clocks. A pulsar timing array (PTA) takes that fact seriously and asks: what if you treat a few dozen of these clocks, distributed across the sky, as a single instrument? The propagation time of pulses from any one pulsar to Earth is fixed by the spacetime metric along the line of sight. A passing gravitational wave at nanohertz frequencies stretches and compresses that path slightly, advancing or retarding pulse arrivals by tens to hundreds of nanoseconds. One pulsar's residual is degenerate with countless other things — clock errors, ephemeris errors, dispersion-measure drifts, intrinsic spin noise. But the cross-pulsar pattern is not.

The signature of a real isotropic stochastic gravitational-wave background is a precise angular correlation between pulsar pairs that depends only on their angular separation in the sky. That correlation, the Hellings-Downs curve, was derived in 1983 by Ronald Hellings and George Downs and is set entirely by the spin-2 transverse-traceless tensor structure of gravitational waves in general relativity. Detect the Hellings-Downs angular pattern in the cross-correlations and you have detected gravitational waves — independent of any pulsar's individual residuals.

The leverage is enormous. The baselines between pulsar pairs are thousands of light-years, so the detector "arm length" is roughly 10²² metres — fourteen orders of magnitude longer than LIGO's 4 km. That is what gives PTAs sensitivity at nanohertz frequencies: the longer the arm, the lower the frequency band you can probe.

The Hellings-Downs curve

Consider two pulsars separated by an angle θ on the sky. An isotropic stochastic gravitational-wave background sweeping across both lines of sight imprints correlated timing residuals. Averaging over wave directions and polarisations, the expected normalised cross-correlation is

ζ(θ) = (1/2) − (x/4) + (3/2) x ln x ,    x = (1 − cos θ)/2

The curve has a few signature features that distinguish it from any common-mode systematic:

• θ = 0°: ζ = 0.5 (the curve is normalised to 1 when including the pulsar's own auto-correlation; pair correlation peaks at 1/2)
• θ ≈ 80°: ζ reaches its minimum, ≈ -0.025 (slightly anti-correlated)
• θ = 180°: ζ ≈ 0.13 (positive again — antipodal pulsars are partially correlated because the wave passes through both)

Compare to systematic-error patterns: a clock drift produces a monopole (ζ = constant), a solar-system ephemeris error produces a dipole (ζ ∝ cos θ). Only the quadrupolar Hellings-Downs shape arises from spin-2 gravitational waves. This is why detecting the Hellings-Downs angular dependence is the gold-standard test, not just an excess of common-mode power.

The five PTA collaborations

Array	Telescope(s)	Pulsars timed	Started	Region
NANOGrav	Green Bank, VLA, (Arecibo until 2020)	~68	2007	North America
EPTA	Effelsberg, Lovell, Nançay, Westerbork, SRT	~25	2005	Europe
PPTA	Parkes 64 m	~25	2003	Australia
InPTA	upgraded GMRT	~10	2015	India
CPTA	FAST 500 m	~50	2019	China
IPTA combined	All of the above	~80 unique	2008	Global

Each array observes its set of pulsars roughly every two to four weeks, accumulating decade-plus timing baselines. NANOGrav, the longest-running, started with 17 pulsars in 2007 and has grown to 68. CPTA, leveraging the enormous collecting area of FAST (the world's largest single-dish radio telescope), is gaining sensitivity quickly despite its later start. The IPTA combines all the data into a single coherent analysis, with overlap between sites giving cross-checks on calibration and clock transfer.

From pulses to residuals: what is actually measured

The basic observable is the time of arrival (TOA) of each pulse at the telescope. To get nanosecond precision on a single observation, the radio data are folded over thousands of rotations using a known ephemeris, producing a pulse profile that is then cross-correlated against a template to extract the TOA. A typical NANOGrav observation gives a single-epoch TOA with rms uncertainty of order 100 ns; the best pulsars (J1909-3744, J0437-4715) reach below 50 ns.

Each TOA is fitted against a deterministic model:

• Spin period P and derivative Ṗ (with higher-order terms if needed)
• Astrometric parameters: position (α, δ), proper motion (μ_α, μ_δ), parallax π
• Time-varying dispersion measure DM(t) along the line of sight (frequency-dependent dispersion of pulses by free electrons)
• For binary pulsars: five Keplerian + up to seven post-Keplerian relativistic parameters
• Earth's motion around the solar-system barycentre using JPL ephemerides (DE440 or similar)
• Atomic-clock transfer from observatory clock to TT(BIPM) to TCB

What remains after subtracting this model from the observed TOAs is the timing residual. For a real PTA dataset, the residuals show a small per-pulsar "red noise" (excess power at low frequencies) plus a tiny common-spectrum process that, in the 2020 NANOGrav 12.5-yr paper, was the first hint of a possible GW background. The 2023 paper showed that the cross-correlations between pulsar pairs follow the Hellings-Downs curve.

June 2023: the announcement

On 28 June 2023, NANOGrav, EPTA together with InPTA, PPTA, and CPTA simultaneously released coordinated papers. The headline NANOGrav 15-year result, based on 68 pulsars timed over up to 17 years, reported:

• Common red spectrum. Power-law amplitude A ≈ 2.4 × 10⁻¹⁵ at f = 1/yr, spectral index γ ≈ 3.2 (close to the SMBH-binary expectation of γ = 13/3 ≈ 4.33 in the strain spectrum).
• Hellings-Downs angular correlation at 3.5σ in the Bayesian analysis (model comparison versus a common-spectrum-only model with no HD correlation), 4σ in the frequentist optimal-statistic test.
• No statistically significant individual SMBHB. The data are consistent with an unresolved stochastic background; no continuous-wave source from a single dominating binary stood out above the background.

EPTA+InPTA, with a comparable timing baseline but fewer pulsars, reported ~3σ for the Hellings-Downs correlation. PPTA reported ~2σ. CPTA reported the strongest single-array detection in some analyses by leveraging FAST's enormous SNR per pulse. The four results are consistent with each other within their respective uncertainties. The press releases were carefully worded as "evidence" rather than "discovery", which by particle-physics convention requires 5σ — but the cross-collaboration consistency is itself a powerful argument that the signal is real.

Worked example: NANOGrav's amplitude in physical units

The NANOGrav 15-yr power-law amplitude A = 2.4 × 10⁻¹⁵ is quoted at a reference frequency of f_ref = 1/yr ≈ 31.7 nHz, in the convention where the characteristic strain is

h_c(f) = A × (f / f_ref)^(-2/3)    for SMBHB inspiral spectrum

Convert to the timing residual induced on a single pulsar. The power spectral density of the residual is

S(f) = h_c²(f) / (12 π² f³)
     = A² × f_ref^(4/3) × f^(-13/3) / (12 π²)

Integrate over the observing band from f_low ≈ 1/(15 yr) to f_high ≈ 1/(2 weeks):

σ²_residual ≈ ∫ S(f) df ≈ A² × (15 yr)^(10/3) × ... ≈ (~100 ns)²

That order-of-magnitude answer — ~100 ns rms residual from the GW background — sits within the noise floor of an individual pulsar but is exactly the level that becomes visible when you cross-correlate dozens of pulsars and require the Hellings-Downs angular pattern. The signal is buried in single-pulsar noise but jumps out of the pair-correlation matrix.

What is sourcing the background?

The most natural astrophysical interpretation is the cosmic population of inspiralling supermassive black hole binaries (SMBHBs). Galaxy mergers throughout cosmic history leave behind binary SMBHs that sink to the centre of the merged system and inspiral by GW emission over Gyr timescales. The incoherent sum of waves from all these binaries — most of them unresolved, individually too faint to detect — produces a stochastic background. The expected spectrum, first calculated by Phinney in 2001, is h_c(f) ∝ f⁻²ᐟ³ in characteristic strain, or γ = 13/3 in the residual power spectrum.

The NANOGrav 15-yr spectrum is consistent with f⁻²ᐟ³ but slightly flatter — γ ≈ 3.2 rather than 13/3 ≈ 4.33. The tilt is not yet statistically significant, but it has opened serious discussion of alternative or additional sources:

• Inspirals with environmental coupling. Binaries embedded in dense stellar or gas environments lose energy to interactions other than GWs, modifying the spectrum at low frequencies.
• Cosmological first-order phase transitions. A symmetry-breaking transition in the early universe (e.g., at temperatures TeV–EeV) can produce a GW background peaked in the nanohertz band today after cosmological redshift.
• Cosmic string networks. Decaying cosmic strings from grand-unified-theory breaking produce a roughly scale-invariant background extending across many decades.
• Primordial GWs from inflation. Possible if the inflationary tensor power is enhanced at small scales.
• Domain walls. A late-time domain-wall network releasing energy into GWs.

The next decade of PTA data should distinguish these scenarios by measuring the spectrum precisely, searching for anisotropy on the sky, and looking for individual loud SMBHBs that would confirm the astrophysical origin.

PTAs in the gravitational-wave spectrum

Detector	Frequency	Wavelength	Best sources	Status
PTAs (NANOGrav, IPTA)	1 – 100 nHz	0.3 – 30 ly	SMBHB inspirals, cosmic strings	Operating, 2023 detection
LISA	0.1 – 100 mHz	10⁶ – 10⁹ km	Massive BH mergers, EMRIs, WD binaries	Launch ~2035
LIGO / Virgo / KAGRA	10 – 1000 Hz	300 – 30000 km	Stellar-mass BH and NS mergers	Operating since 2015
Cosmic Explorer / ET	1 – 10000 Hz	30 – 300000 km	All LIGO sources at much higher SNR	2030s
CMB B-modes	10⁻¹⁸ – 10⁻¹⁶ Hz	Hubble scale	Inflationary GWs	BICEP, LiteBIRD, ongoing

The frequency band sets the source population: stellar-mass mergers at LIGO; massive BH mergers and extreme mass ratio inspirals at LISA; SMBHB inspirals at PTAs; and inflationary tensor modes imprinted on the CMB at the longest possible wavelengths. PTAs are not in competition with LIGO and LISA — they probe a different and previously inaccessible regime.

Calibration and systematics

Distinguishing a real Hellings-Downs signal from systematics is the hardest part of PTA work. The leading systematic categories are:

• Solar-system ephemeris errors. If the model of the Sun and planets' positions is wrong, every pulsar's TOAs are shifted in a dipolar pattern (the Earth's orbital position moves in a coherent way). Modern JPL ephemerides DE440 are accurate to ~100 m for inner planets; outer-planet uncertainties are the main remaining contributor.
• Clock-transfer errors. Each observatory clock is referenced to TT(BIPM) via GPS or two-way satellite time transfer, with uncertainties of order ~10 ns per epoch.
• Pulsar intrinsic timing noise. Each MSP has its own "red noise" from magnetospheric drift, dispersion-measure variations, and possibly internal angular-momentum exchange. Modelled per-pulsar.
• Dispersion measure variations. Time-variable electron density along the line of sight (ISM clumps, solar wind) introduces frequency-dependent delays that must be modelled with multi-frequency observations.
• Scattering and scintillation. Multi-path propagation through the ISM broadens pulses and introduces frequency-dependent jitter.

The cross-array consistency of the 2023 NANOGrav, EPTA+InPTA, PPTA and CPTA results — all reporting a Hellings-Downs-like correlation despite using different telescopes, different pulsar samples, different ephemerides and different clocks — is the strongest evidence against any single systematic explanation.

Next decade: from evidence to discovery

The 3.5σ result is "evidence", not "discovery". Crossing 5σ requires more data, more pulsars, and tighter modelling. Several projects will deliver this:

• IPTA Data Release 3. Coordinated combination of all five regional arrays through about 2024 baselines, expected to deliver ~5σ Hellings-Downs detection by the late 2020s.
• FAST + MeerKAT. The world's largest single-dish and the largest interferometric array in the southern hemisphere are dramatically increasing per-pulsar SNR. MeerKAT's MPTA programme is timing ~80 pulsars at 200-ns precision over 5 years.
• SKA. The Square Kilometre Array, partially operational by 2030, will time hundreds of MSPs at nanosecond precision and almost certainly resolve individual SMBHB sources of GW emission, not just the background.
• Anisotropy mapping. The single-source resolved limit allows construction of a GW sky map, potentially revealing nearby SMBHB candidates that should also show optical/AGN signatures.

The science programme then extends from "is the background real?" to "what produces it?". Each spectrum slope, anisotropy detection, and individual binary resolution adds an independent measurement that tests SMBHB vs cosmological sources.

Common pitfalls

• Confusing common red noise with the GW background. NANOGrav 12.5-yr (2020) saw a common red process across pulsars but could not confirm the Hellings-Downs angular signature; the discovery claim came only with the 2023 paper. A common process is a necessary but not sufficient condition.
• Treating one pulsar's residual as a GW measurement. Single-pulsar residuals are degenerate with countless systematics. The information lives in the pair correlation matrix, not in individual signals.
• Conflating PTAs with LIGO physics. LIGO observes individual merger waveforms (deterministic chirps) at high SNR. PTAs detect a stochastic background — many unresolved sources adding up — via cross-correlations. Different sources, different math, different astrophysics.
• Reading too much into a few-σ shift in spectral index. The slope is poorly constrained at present; the apparent flattening from γ = 13/3 to γ ≈ 3.2 is not yet statistically significant. Wait for IPTA DR3 and the SKA before fitting models that depend on the exact slope.
• Expecting a continuous-wave detection of a single SMBHB. No PTA has yet found a statistically significant individual binary. The background detection came first; individual sources will emerge gradually as sensitivity improves through the late 2020s.
• Forgetting that 'nHz' means decades of observation. A 10 nHz signal has a period of about 3 years. You need at least one full cycle (~ a decade of timing data) to measure it, which is why PTAs are slow-cooking experiments rather than rapid-turnaround ones.