Common-Spectrum and Correlated-Process Inference

A common red spectrum is the first population-level clue in a pulsar timing array, not the final gravitational-wave signature. A process can have similar low-frequency power in many pulsars while remaining uncorrelated between them. The decisive additional evidence is the cross-pulsar spatial covariance—especially Hellings–Downs structure after clock, ephemeris, intrinsic, and chromatic noise are included. Only then does source-population interpretation begin.

Helpful background. Hellings–Downs Correlations derives the tensor angular pattern. Pulsar Noise and Solar-System Ephemerides develops the principal alternatives.

Nested levels of evidence

Consider a common residual spectrum

$$P_{\rm c}(f) =\frac{A_{\rm c}^2}{12\pi^2} \left(\frac{f}{f_{\rm yr}}\right)^{-\gamma} f_{\rm yr}^{-3}.$$

Four hypotheses can share this spectrum but differ spatially:

$$C_{ab}^{\rm c}(f)=\Gamma_{ab}P_{\rm c}(f),$$

with $\Gamma_{ab}=\delta_{ab}$ for an uncorrelated common-spectrum process, approximately $1$ for a monopole, and $\boldsymbol p_a\cdot\boldsymbol p_b$ for an idealized dipole. In the normalization of the preceding page, the full isotropic tensor covariance uses

$$\Gamma_{ab}^{\rm tensor} =\Gamma_{\rm HD}(\zeta_{ab}) +\frac12\delta_{ab}.$$

The added term is the pulsar-term auto-power: for distinct pulsars $\Gamma_{ab}^{\rm tensor}=\Gamma_{\rm HD}(\zeta_{ab})$, while the same-pulsar value is one.

The observations support successively stronger conclusions:

1. red noise appears in individual pulsars;
2. a shared spectral process improves the array likelihood;
3. cross-correlations favor a spatial family over an uncorrelated model;
4. tensor Hellings–Downs structure is favored over clock and ephemeris alternatives;
5. spectrum, anisotropy, non-Gaussianity, and external astrophysics constrain the source class.

Skipping from step 2 to “supermassive binaries detected” confuses common power with correlated gravitational radiation and then with its origin.

Bayesian and frequentist statistics

Stack all pulsars’ compressed residuals into $\boldsymbol r$ after timing-model marginalization. If their joint process is Gaussian, each spatial hypothesis supplies a different positive-definite covariance $C$, and

$$\ln\mathcal L=-\frac12\boldsymbol r^TC^{-1}\boldsymbol r -\frac12\ln\det C-\frac{n}{2}\ln(2\pi).$$

The quadratic term rewards a covariance that explains the observed pairs; the determinant prevents arbitrarily increasing every variance from becoming a good fit. Models can have identical auto-spectra and differ only in off-diagonal blocks. That is how correlation evidence adds information beyond a common spectrum Agazie et al. 2023, §3.

A Bayesian analysis marginalizes timing, pulsar noise, ephemeris, and common-process parameters and compares evidences. Bayes factors depend on prior measures, parameter bounds, and model completeness. Product-space samplers can estimate odds efficiently but require adequate switching and occupancy.

A frequentist optimal statistic cross-correlates residuals with inverse-noise weighting and a target overlap function. In the weak-signal limit it estimates a background amplitude and signal-to-noise ratio. Noise-marginalized variants propagate uncertainty in pulsar noise. Neither framework is intrinsically superior; agreement under calibrated injections is powerful.

Sky-scramble and phase-shift tests construct surrogate spatial or temporal patterns. They are useful null diagnostics, but scrambles are correlated for a finite array and must not be counted as unlimited independent trials.

Results through the stated cutoff

The NANOGrav 15-year analysis reports strong evidence for a common red process and multiple lines of evidence for Hellings–Downs-compatible correlations. Its frequentist and Bayesian results, alternative spatial patterns, ephemeris treatment, and simulations are documented together Agazie et al. 2023, §§3–5 and appendices.

EPTA/InPTA, PPTA, and CPTA data releases reported contemporaneous evidence for nanohertz common processes and gravitational-wave-compatible spatial structure, with materially different strengths and test statistics. EPTA’s modern-backend subset carries stronger Hellings–Downs support than its longer span Antoniadis et al. 2023, §§4.2–5 and Table 5; PPTA finds insignificant full-likelihood model odds but an approximately $2\sigma$ cross-correlation-only result Reardon et al. 2023, §§3.6–3.7 and 5; CPTA’s $4.6\sigma$ result is a single-frequency Hellings–Downs-versus-constant comparison that does not exclude a dipole Xu et al. 2023, §§4–5 and Appendices A and C. Different data spans, pulsar sets, noise models, and evidence conventions prevent naive multiplication of headline significances or Bayes factors.

Evidence line at 2026-08-18	Supported reading	Remaining question
Multiple PTAs find low-frequency common power	The effect is not peculiar to one telescope or analysis	Cross-release covariance and combined calibration
NANOGrav and other analyses favor tensor-like correlations	A nanohertz gravitational-wave background is the leading interpretation	Exact isotropy, Gaussianity, and spectral shape
Monopole/dipole and ephemeris alternatives are less favored in tested models	Simple clock or barycenter error is insufficient	More complex systematics are bounded, not metaphysically excluded
Spectrum is broadly compatible with massive-black-hole binaries	That population is a plausible dominant source	Environmental evolution, source discreteness, and nonastrophysical contributions

The strongest credible conservative position is not “there is only common red noise.” Spatial-correlation evidence has moved beyond that. It is that the source composition and detailed random-field model remain uncertain even if a nanohertz gravitational-wave background is accepted.

From background to source population

For circular, gravitational-radiation-driven supermassive-black-hole binaries, an idealized ensemble gives characteristic strain

$$h_c(f)=A\left(\frac{f}{f_{\rm yr}}\right)^{-2/3},$$

Because $P_{\rm gw}=h_c^2/(12\pi^2f^3)$, this gives $P_{\rm gw}\propto f^{-4/3-3}=f^{-13/3}$ and hence $\gamma=13/3$. The time integration from redshift to residual is already included in this relation; applying another factor $f^{-2}$ would double count it. Stellar/gas coupling, eccentricity, finite source counts, and mass/redshift demographics can bend or fluctuate the spectrum. Cosmic strings or primordial processes predict other shapes but can overlap across a limited band.

Source attribution therefore needs spectral freedom, anisotropy and non-Gaussian tests, galaxy/black-hole population priors, and independent cosmological constraints. A fitted slope near $13/3$ is consistency, not unique identification.

Reproducing the result and testing its robustness

A bounded rerun pins the PTA data release, timing solutions, ephemeris, pulsar/noise model, Fourier basis, spatial templates, priors, sampler, evidence estimator, and software commits. It reproduces common-uncorrelated, monopole, dipole, and Hellings–Downs odds; a noise-marginalized optimal statistic; chain and Fourier-basis convergence; and tensor/clock/ephemeris injection recovery on the real cadence.

Adversarial checks allow anisotropy, free spectral bins, non-Gaussian source realizations, alternative ephemerides, and leave-one-pulsar subsets. Refresh after a new public PTA/IPTA release, corrected noise/timing products, or a cross-correlation reanalysis that changes the evidence ranking.

Exercises

1. Common but uncorrelated. Each pulsar has an identical red-noise spectrum drawn from an independent realization. What are the off-diagonal covariances?

Solution

They vanish in the ensemble: $C_{ab}=0$ for $a\ne b$. The auto-spectra can share amplitude and slope while providing no spatial gravitational-wave signature. This is why a common-spectrum process is a distinct hypothesis from a correlated background.

2. Prior sensitivity. Why can widening the allowed amplitude range change a Bayes factor even if the likelihood peak is unchanged?

Solution

Evidence averages likelihood over the prior measure. Widening a prior adds low-likelihood volume and changes the Occam factor. The posterior peak can stay fixed while the model evidence changes, so prior ranges and measures must accompany reported odds.

References

• Agazie, Gabriella, et al. “The NANOGrav 15-Year Data Set: Evidence for a Gravitational-Wave Background.” Astrophysical Journal Letters 951 (2023): L8. DOI. Open PDF.
• Antoniadis, J., et al. “The Second Data Release from the European Pulsar Timing Array. III. Search for Gravitational Wave Signals.” Astronomy & Astrophysics 678 (2023): A50. DOI. Open PDF.
• Reardon, Daniel J., et al. “Search for an Isotropic Gravitational-Wave Background with the Parkes Pulsar Timing Array.” Astrophysical Journal Letters 951 (2023): L6. DOI. Open PDF.
• Xu, Heng, et al. “Searching for the Nano-Hertz Stochastic Gravitational Wave Background with the Chinese Pulsar Timing Array Data Release I.” Research in Astronomy and Astrophysics 23 (2023): 075024. DOI. Open PDF.