Gravitational-Wave Background Signal & PPL Models

• GWB signal is an unresolved superposition of gravitational waves from diverse astrophysical and cosmological sources with distinct spectral features.
• The piecewise power-law framework models the GWB spectrum through segmented power laws, enabling robust spectral reconstruction and model flexibility.
• Bayesian model selection and averaging are applied to evaluate and compare competing spectral models in pulsar timing array analyses.

A gravitational-wave background (GWB) signal is an unresolved, stochastic superposition of gravitational waves emitted by a large number of independent sources, both astrophysical (such as supermassive black-hole binaries) and cosmological (such as processes in the early Universe). The detection, spectral inference, and modeling of the GWB represents a central challenge for pulsar timing array (PTA) experiments, such as those conducted by the NANOGrav collaboration. The analysis and characterization of the GWB rely heavily on robust spectral models, Bayesian inference techniques, and flexible parametrizations that can accommodate the broad phenomenology expected from theoretical models and data.

1. Mathematical Formulation of the GWB Spectrum

The energy-density spectrum of the GWB, typically written as $h^2\Omega_{\rm GW}(f)$, encodes the mean-square amplitude of gravitational-wave strain per logarithmic interval in frequency $f$. Spectral reconstruction for PTA experiments requires a parametrization capable of capturing both smooth and broken (feature-rich) spectra, as different theoretical origins predict markedly different $f$-dependencies.

The piecewise power-law (PPL) ansatz provides a mathematically explicit and flexible model for this purpose (Agazie et al., 14 Jan 2026):

$$h^2\Omega_{\rm GW}(f) = \begin{cases} A_1\bigl(f/f_0\bigr)^{5-\gamma_1}, & f_0\le f< f_1 \ A_2\bigl(f/f_1\bigr)^{5-\gamma_2}, & f_1\le f< f_2 \ \vdots \ A_N\bigl(f/f_{N-1}\bigr)^{5-\gamma_N}, & f_{N-1}\le f\le f_N \end{cases}$$

Here, $N$ is the number of segments, $f_0=1/T_{\rm obs}$ (minimum accessible frequency), $f_N=N_{\rm bins}/T_{\rm obs}$ (maximum frequency), and the internal nodes $f_1,\ldots,f_{N-1}$ may be fixed (binned) or treated as free parameters.

Continuity is enforced by requiring:

$$A_{i+1} = A_i (f_i/f_{i-1})^{5-\gamma_i} \,, \quad i=1,\ldots,N-1\,.$$

This framework subsumes the canonical single power-law (constant $\gamma$), as well as broken and multiply-broken models, which are required to test for features as predicted by specific astrophysical or cosmological GWB scenarios.

2. Family of Piecewise Power-Law Submodels

The practical deployment of the PPL framework involves a hierarchy of submodels that vary in flexibility:

• Constant Power Law (PPL 1): One segment, two parameters $(A, \gamma)$.
• Broken Power Law (PPL 2): Two segments, parameterized by low/high-frequency indices and the breakpoint.
• Doubly/Triply Broken (PPL 3,4,...): Higher-order models for increased spectral complexity.

Two versions are constructed:

• Binned: Breakpoints fixed at logarithmically equispaced locations.
• Free: Breakpoints are free, with ordering constraints.

The number of free parameters (for $n$ segments) is $n+1$ in the binned case and $2n$ in the free-node case (amplitudes, indices, breakpoint locations).

3. Bayesian Model Selection and Averaging

Spectral inference is performed within a fully Bayesian context. The PTA likelihood is marginalized over all timing-residual cross-spectra, including pulsar-intrinsic white and red noise, and a common-spectrum red process with Hellings–Downs angular correlations:

$$\mathcal{L}(D|\theta) \propto \exp\left[-\tfrac12\,\delta t^T C^{-1}(\theta) \delta t\right]/\sqrt{\det C(\theta)}$$

where $C$ encodes all noise and signal covariances (Agazie et al., 14 Jan 2026).

Model comparison utilizes the marginal evidence (integrated likelihood) $\mathcal{Z}_k$ for each submodel $M_k$, computed efficiently via product-space sampling and likelihood reweighting. Bayes factors $\mathcal{B}_{n1} = \mathcal{Z}_n/\mathcal{Z}_1$ weigh complex PPL$n$ models against the reference constant power law.

The Bayesian model-averaged (BMA) spectrum is constructed as:

$$\left\langle h^2\Omega_{\rm GW}(f)\right\rangle = \sum_n P_n\, [h^2\Omega_{\rm GW}(f)]_{M_n}$$

where $P_n$ is the posterior weight for model $n$. This averaging incorporates model-selection uncertainty into credible intervals for $h^2\Omega_{\rm GW}(f)$ and derived quantities such as the running index $\gamma(f) = 5 - \frac{d\ln\Omega}{d\ln f}$.

4. Prior Choices and Hyperparametric Structure

All PTA spectral reconstructions must carefully specify parameter priors. For the NANOGrav analysis (Agazie et al., 14 Jan 2026):

• Pulsar red-noise amplitudes $A_a^{\rm red}$: log-uniform over $[10^{-20},10^{-11}]$
• Pulsar red-noise indices $\gamma_a^{\rm red}$: uniform over $[0,7]$
• PPL segment amplitude $A_1$: log-uniform over $[10^{-12},10^{-8}]$
• PPL indices $\gamma_i$: uniform over $[-3,12]$
• Node frequencies $f_i$: log-uniform within $[f_0, f_N]$ with sorting constraints

No additional hierarchical or hyperpriors are employed. These choices balance coverage of physically plausible parameter space and computational tractability.

5. Comparison with Alternative GWB Spectrum Models

Three primary spectral ansätze are used in PTA GWB studies:

Model	Flexibility	Number of Parameters	Regularization
Constant Power Law	Minimal (rigid)	2	Slope fixed globally
Free Spectrum	Maximal (unregular.)	$N_{\rm bins}$	No smoothness, prone to noise
PPL (Piecewise)	Intermediate	$n+1$ (binned)/$2n$	Local smoothness via segments

PPL models occupy a middle ground: they enforce smooth, segment-wise power laws (suppressing wild local variation) while allowing for multiple slopes and breaks as required by the data. This improves spectral adaptivity over the constant power-law model, while avoiding the unconstrained amplitude fluctuations of the free spectral model—which yields wider credible intervals and reduced interpretability.

PPL reconstructions are thus particularly well-suited to distinguish between classical astrophysical GWB spectra (smooth, single-index) and more exotic, cosmological-origin backgrounds (which may feature bends and sharp roll-offs).

6. Implementation, Data Products, and Downstream Applications

For the NANOGrav 15-yr data, the PPL methodology is implemented in the PTArcade, ENTERPRISE, and ENTERPRISE_EXTENSIONS software stacks, leveraging MCMC for parameter inference and product-space methods for evidence calculation (Agazie et al., 14 Jan 2026).

Released data products include:

• Posterior parameter chains for every PPL$n$ model.
• Model weights $P_n$.
• BMA medians and 68%/95% credible intervals for $h^2\Omega_{\rm GW}(f)$ and $\gamma(f)$.
• PPL-based "violin" plots at every PTA frequency bin.

A crucial application is the "fast refit": the PPL posterior can be KDE-interpolated to provide an induced likelihood over PPL parameter space. Any theoretical GWB spectrum can then be mapped to its best-fit PPL form (by minimizing matched-filter $\chi^2$ distance), and reweighted against the empirical PPL posterior. This enables rapid evaluation of new theoretical models without resampling the full PTA likelihood.

7. Context and Significance in Gravitational-Wave Astrophysics

The PPL reconstruction framework is tailored to meet the physical and statistical demands of PTA gravitational-wave background inference:

• It provides the flexibility needed to test for both canonical and non-standard spectral shapes as predicated by emerging evidence (e.g., the NANOGrav 15yr signal).
• Model comparison and BMA rigorously account for model-selection uncertainty.
• Data products are designed for modular reuse, downstream model comparison, and rapid update as new theoretical or empirical developments arise.

Overall, the PPL approach sets a new standard for robust, interpretable, and flexible spectral reconstruction of astophysical stochastic backgrounds, and is expected to remain central as the PTA dataset volume and fidelity continue to increase (Agazie et al., 14 Jan 2026).