betaFeedback

WorksPhysics and Astronomy

Joint Angular-Power Distributions of Interfering Nanohertz Gravitational Waves

Abstract

Finite populations of supermassive black-hole binaries produce fluctuations in both gravitational-wave power and pulsar timing correlations. We characterize the realization-dependent, orientation-averaged Earth-term angular correlation through its frequency-resolved Legendre coefficients. A hybrid model of bright sources and a Gaussian faint background extends the conditional Gamma distribution in the many-weak-source limit to non-central χ2χ^2 statistics, yielding correlated multipoles after population marginalization. Our analytical variance decomposition and SMBHB population simulations reveal a population-dependent transition from interference dominated variance at low frequencies to population-power-dominated variance at high frequencies. In the many-weak-source regime, a multivariate lognormal approximation captures the main features of the joint distribution using only the first two log-cumulants. These results provide statistical benchmarks for fluctuations in both the overall amplitude and angular structure of the realized correlation.

Introduction

Pulsar timing arrays (PTAs) probe nanohertz gravitational waves (GWs) through correlated variations in the pulse arrival times of Galactic millisecond pulsars [1]. Current PTA analyses commonly model the gravitational-wave background (GWB) as a stationary, isotropic, unpolarized Gaussian process, whose ensemble spatial correlations follow the Hellings–Downs (HD) curve. This Gaussian approximation is well motivated when each frequency bin contains sufficiently many weak sources with independent random phases and no dominant contribution. However, an astrophysical GWB generated by a finite population of supermassive black hole binaries (SMBHBs) need not satisfy these conditions.

SMBHBs are discrete, slowly evolving GW sources in the PTA band. The discreteness of their population leads to realization-to-realization fluctuations in the total GW power, with rare nearby or massive binaries producing a heavy high-power tail that becomes more prominent at high frequencies. At low frequencies, a single Fourier bin may contain ≳O(103)\gtrsim\mathcal{O}(10^3) binaries [2]. Their radiation-reaction timescales, tc∼O(106) yrt_c \sim\mathcal{O}(10^6)\,\mathrm{yr}, far exceed typical PTA observing baselines, Tobs∼O(10) yrT_{\mathrm{obs}} \sim\mathcal{O}(10)\,\mathrm{yr} [3], allowing multiple sources to remain within the same Fourier bin throughout the observation. Their complex amplitudes then add coherently, producing interference-induced fluctuations in the angular correlation pattern. Consequently, a single astrophysical realization need not reproduce either the ensemble-averaged GW spectrum or the HD correlation.

These effects have been studied from several complementary perspectives. The variance and cosmic variance of the HD correlation have been derived for Gaussian and discrete-source ensembles [2, 4, 5], while interference between discrete sources has been shown to generate frequency-dependent multipolar departures from the HD curve [6, 7, 8]. More recently, non-Gaussian statistics of the strain power, timing residuals, and higher-order correlators have been investigated [9, 10, 11, 12, 13, 14, 15, 16, 17]. Together, these studies highlight that a realistic astrophysical GWB is not fully characterized by its ensemble-averaged two-point function. A more complete statistical description should therefore characterize the probability distribution of the realized angular covariance itself, accounting for both interference-induced fluctuations in the angular correlation pattern and population-induced variations in the total GW power.

In this Letter, we develop a statistical description of the realized angular correlation using frequency-resolved Legendre multipoles of the pulsar-averaged Earth-term correlation. We construct a hybrid model of discrete bright sources and a Gaussian faint background, extending the conditional Gamma distribution in the many-weak-source limit to non-central χ2\chi^2 statistics when bright sources become important. Marginalizing over source-population fluctuations yields the joint multipole distributions and correlations between different angular modes. We further quantify the contributions of interference and total-power fluctuations to the multipole variance, and derive a compact multivariate lognormal approximation in the many-weak-source regime. We test these predictions against Monte Carlo SMBHB populations.

Frequency-resolved angular spectrum. Consider a frequency bin pp centered at fpf_p, with complex source amplitudes sjp=2/3Ajeiϕjsinc⁡[πT(fj−fp)]s_j^p = \sqrt{2/3}A_j e^{i\phi_j}\operatorname{sinc}[\pi T(f_j-f_p)], where Aj=2/5AjA_j = \sqrt{2/5}A_j is the effective circularly polarized amplitude and AjA_j is the physical face-on amplitude. In the narrow-bin approximation, we replace the sinc response by a delta function, sinc⁡[πT(fj−fp)]∝δ(fj−fp)\operatorname{sinc}[\pi T(f_j-f_p)] \propto\delta(f_j-f_p), neglecting spectral leakage and inter-bin correlations. We henceforth consider a single frequency bin and suppress the index pp.

Let z~a(t)\tilde{z}_a(t) be the complex Earth-term redshift, expanded as z~a(t)=∑ℓmzℓmYℓm∗(p^a)\tilde{z}_a(t) = \sum_{\ell m} z_{\ell m}Y_{\ell m}^{*}(\hat{p}_a). For a given GW realization, averaging pulsar pairs over all orientations at fixed angular separation γab\gamma_{ab} defines the angular correlation [18]

C(γab)=⟨z~az~b∗⟩γab=∑ℓ=2∞SℓPℓ(cos⁡γab).(1)C(\gamma_{ab}) = \langle\tilde{z}_a \tilde{z}_b^{*}\rangle_{\gamma_{ab}} = \sum_{\ell=2}^{\infty} S_{\ell}P_{\ell}(\cos\gamma_{ab}). \tag*{(1)}

Using the spin-weighted harmonic representation of the GW field, the Legendre coefficients are given by

Sℓ=cℓHD4πnℓ∑m=−ℓℓ∣2aℓm∣2,(2)S_{\ell} = c_{\ell}^{\mathrm{HD}}\frac{4\pi}{n_{\ell}}\sum_{m=-\ell}^{\ell}|{}_{2}a_{\ell m}|^{2}, \tag*{(2)}

where nℓ=2ℓ+1n_{\ell} = 2\ell+ 1, cℓHD=3nℓ/[2(ℓ−1)ℓ(ℓ+1)(ℓ+2)]c_{\ell}^{\mathrm{HD}} = 3n_{\ell}/[2(\ell-1)\ell(\ell+1)(\ell+2)], and 2aℓm=(−1)ℓ∑jsj2Yℓm(Ω^j){}_{2}a_{\ell m} = (-1)^{\ell}\sum_j s_j{}_{2}Y_{\ell m}(\hat{\Omega}_j). The HD coefficients satisfy ∑ℓcℓHDPℓ(cos⁡γab)=ΓHD(γab)\sum_{\ell}c_{\ell}^{\mathrm{HD}}P_{\ell}(\cos\gamma_{ab}) = \Gamma_{\mathrm{HD}}(\gamma_{ab}), with ΓHD(0)=1/2\Gamma_{\mathrm{HD}}(0) = 1/2.

Substituting the source expansion yields

Sℓ=cℓHDH2+Iℓ,Hn=∑j∣sj∣n,(3)S_{\ell} = c_{\ell}^{\mathrm{HD}}H_2 + I_{\ell}, \qquad H_n = \sum_j |s_j|^n, \tag*{(3)}

where

Iℓ=8πcℓHDnℓℜ∑m∑j<ksjsk∗2Yℓm(Ω^j)2Yℓm∗(Ω^k).(4)I_{\ell} = \frac{8\pi c_{\ell}^{\mathrm{HD}}}{n_{\ell}}\Re\sum_m\sum_{j<k}s_js_k^{*}{}_{2}Y_{\ell m}(\hat{\Omega}_j){}_{2}Y_{\ell m}^{*}(\hat{\Omega}_k). \tag*{(4)}

The common total-power term cℓHDH2c_{\ell}^{\mathrm{HD}}H_2 preserves the HD multipole proportions, while the interference term IℓI_{\ell} introduces multipole-dependent deviations, with C(γab)=H2ΓHD(γab)C(\gamma_{ab}) = H_2\Gamma_{\mathrm{HD}}(\gamma_{ab}) in the absence of interference. Thus, fluctuations in SℓS_{\ell} arise from both population-power fluctuations and wave interference.

To isolate relative interference-induced angular fluctuations, we define the power-normalized multipole Qℓ≡Sℓ/(cℓHDH2)=1+Iℓ/(cℓHDH2)Q_{\ell} \equiv S_{\ell}/(c_{\ell}^{\mathrm{HD}}H_2) = 1 + I_{\ell}/(c_{\ell}^{\mathrm{HD}}H_2). For many weak sources with isotropic directions and independent uniform phases, the proper-complex Gaussian approximation gives Qℓ∼Gamma⁡(nℓ,nℓ)Q_{\ell} \sim\operatorname{Gamma}(n_{\ell},n_{\ell}), or 2nℓQℓ∼χ2nℓ22n_{\ell}Q_{\ell} \sim\chi^{2}_{2n_{\ell}} [19], where the Gamma parameters denote shape and rate. This approximation can fail when a few bright SMBHBs dominate.

To account for bright sources, we construct a hybrid model in which bright sources are retained explicitly and the faint population is approximated by an isotropic, unpolarized, proper-complex Gaussian field. The harmonic mode vector is

Bℓ=4πnℓ(2aℓℓ,2aℓℓ−1,⋯ ,2aℓ−ℓ)T=bℓB+uℓ,(5)B_{\ell} = \sqrt{\frac{4\pi}{n_{\ell}}}( {}_{2}a_{\ell\ell},{}_{2}a_{\ell\ell-1},\cdots,{}_{2}a_{\ell-\ell})^{T} = b_{\ell}^{B}+u_{\ell}, \tag*{(5)}

where uℓ∼CN(0,HuI/nℓ)u_{\ell}\sim\mathcal{CN}(0,H_uI/n_{\ell}), HuH_u is the faint-source power, and Sℓ=cℓHD∥Bℓ∥2S_{\ell}=c_{\ell}^{\mathrm{HD}}\lVert B_{\ell}\rVert^{2}. The Gaussian vectors uℓu_{\ell} are independent across ℓ\ell at fixed HuH_u. For a fixed bright-source realization, let WℓB=∥bℓB∥2W_{\ell}^{B}=\lVert b_{\ell}^{B}\rVert^{2}. The conditional Laplace transform of SℓS_{\ell}, Lℓ(t)=E[e−tSℓ∣bℓB,Hu]\mathcal{L}_{\ell}(t)=\mathbb{E}[e^{-tS_{\ell}}\mid b_{\ell}^{B},H_u], follows from Gaussian integration:

Lℓ(t)=exp⁡[−tcℓHDWℓB1+tcℓHDHu/nℓ](1+tcℓHDHu/nℓ)nℓ.(6)\mathcal{L}_{\ell}(t) = \frac{\exp\left[-\dfrac{tc_{\ell}^{\mathrm{HD}}W_{\ell}^{B}}{1+tc_{\ell}^{\mathrm{HD}}H_u/n_{\ell}}\right]} {\left(1+tc_{\ell}^{\mathrm{HD}}H_u/n_{\ell}\right)^{n_{\ell}}}. \tag*{(6)}

For multiple bright sources, WℓBW_{\ell}^{B} includes their mutual interference and generally depends on ℓ\ell. We focus on the single-bright-source limit, motivated by individual-source searches [20] and the possible dominance of one binary in high-frequency bins [21]. In this limit, WℓB=HBW_{\ell}^{B}=H_B for all ℓ\ell. Defining the total power H2=Hu+HBH_2=H_u+H_B and bright-source fraction fB=HB/H2f_B=H_B/H_2, Eq. (6) yields a scaled non-central χ2\chi^2 distribution for QℓQ_{\ell},

Qℓ∣fB∼1−fB2nℓχ2nℓ2(2nℓfB1−fB).(7)Q_{\ell}\mid f_B \sim\frac{1-f_B}{2n_{\ell}}\chi^{2}_{2n_{\ell}}\left(\frac{2n_{\ell}f_B}{1-f_B}\right). \tag*{(7)}

At fixed total power, increasing fBf_B suppresses interference-induced angular scatter without changing the mean, as E[Qℓ∣fB]=1\mathbb{E}[Q_{\ell}\mid f_B]=1 and Var⁡[Qℓ∣fB]=(1−fB2)/nℓ\operatorname{Var}[Q_{\ell}\mid f_B]=(1-f_B^2)/n_{\ell}. The distribution reduces to the central Gamma law at fB=0f_B=0 and approaches δ(Qℓ−1)\delta(Q_{\ell}-1) as fB→1f_B\to1, as shown in Figure 1.

Image not extracted

Figure 1. Conditional distributions of the power-normalized quadrupole Q2Q_2 for different bright-source fractions fBf_B. The gray dashed curve represents the central Gamma limit (fB=0f_B = 0), while the blue and orange solid curves show the non-central χ2\chi^2 distributions at fB=0.5f_B = 0.5 and 0.90.9, respectively. Increasing fBf_B narrows the distribution while preserving its unit mean, indicating suppressed interference-induced angular fluctuations.

For a Poisson source population, the total GW power H2H_2 follows a compound-Poisson distribution [12, 13, 22, 23]. Given the joint population distribution pH(H2,fB)p_H(H_2,f_B), marginalizing over both H2H_2 and fBf_B yields

p(Sℓ)=∫dH2∫01dfB pH(H2,fB) p(Sℓ∣H2,fB).(8)p(S_{\ell})=\int dH_2\int_0^1 df_B\,p_H(H_2,f_B)\,p(S_{\ell}\mid H_2,f_B). \tag*{(8)}

where the conditional distributions follow from Eq. (7) through Sℓ=cℓHDH2QℓS_{\ell}=c_{\ell}^{\mathrm{HD}}H_2Q_{\ell}. This expression applies to the positive-power component. Allowing both HuH_u and HBH_B to fluctuate, the law of total covariance gives

Cov⁡(Sℓ,Sℓ′)=cℓHDcℓ′HD[δℓℓ′nℓE(Hu2+2HuHB)+Var⁡(H2)].(9)\operatorname{Cov}(S_{\ell},S_{\ell'})=c_{\ell}^{\mathrm{HD}}c_{\ell'}^{\mathrm{HD}}\left[\frac{\delta_{\ell\ell'}}{n_{\ell}}\mathbb{E}(H_u^2+2H_uH_B)+\operatorname{Var}(H_2)\right]. \tag*{(9)}

The first term describes interference-induced angular fluctuations at fixed source powers, while the second arises from fluctuations in the total power H2H_2 and correlates different multipoles. In the single-source limit, angular fluctuations vanish, but fluctuations in H2H_2 can still contribute to the covariance. Equation (9) recovers known finite-source covariances under the corresponding assumptions [24] (see Supplemental Material).

We test these predictions using Monte Carlo realizations of SMBHB populations based on semi-analytic galaxy and black-hole evolution prescriptions [25, 26, 27] (see

Supplemental Material). At f=2f = 2 nHz and 30 nHz, we fix the source amplitudes and randomly sample their phases and sky directions, keeping H2H_2 and fBf_B unchanged. Figure 2 compares the resulting distributions of Q2Q_2 with the central Gamma and conditional non-central χ2\chi^2 predictions. Both models agree with simulations at low frequencies, whereas the non-central χ2\chi^2 model provides a better description at high frequencies, where bright sources become important.

Image not extracted

Figure 2. Conditional distributions of the power-normalized quadrupole Q2Q_2 at low and high frequencies, obtained by randomizing source phases and sky directions at fixed source amplitudes. Gray histograms show simulations, while blue dashed and red solid curves represent the central Gamma and conditional non-central χ2\chi^2 predictions, respectively. Both models agree at low frequencies, but the noncentral χ2\chi^2 model performs better at high frequencies.

We then allow the source population to vary between realizations, introducing fluctuations in both H2H_2 and fBf_B. Marginalizing over their joint distribution pH(H2,fB)p_H(H_2, f_B) yields the distributions of S2S_2, incorporating both wave interference and total-power fluctuations. The left panels of Figure 3 compare the distributions with and without interference, the latter corresponding to S2=c2HDH2S_2 = c_2^{\mathrm{HD}} H_2. Interference broadens the distributions without changing

Image not extracted

Figure 3. Left panels: Distributions of the angular power S2S_2 at different frequencies, with (blue) and without (orange) wave interference. Interference broadens the distributions without changing their means, with a diminishing effect at higher frequencies. Right panels: Total variance of S2S_2 and its contributions from interference and population-power fluctuations as a function of frequency. Dots show simulations, while dashed and solid curves show predictions based on the Gamma and non-central χ2\chi^2 models for Q2Q_2, respectively. The non-central χ2\chi^2 model agrees with simulations across the frequency range, whereas the Gamma approximation deviates at high frequencies. Interference dominates the variance at low frequencies, while population-power fluctuations dominate at high frequencies.

their means, with a diminishing effect at higher frequencies. The right panels show the total variance of S2S_2 and its decomposition into interference and population-power contributions. Interference dominates at low frequencies, whereas population-power fluctuations dominate at high frequencies as rare bright binaries become more influential. This crossover depends on the population model. The non-central χ2\chi^2 model reproduces the simulated variances across the frequency range, while the Gamma approximation deviates at high frequencies.

Lognormal approximation

At low frequencies, where numerous weak sources contribute without a single dominant source, the harmonic modes are approximately Gaussian. The resulting conditional Gamma distribution allows a compact description of the population-marginalized Legendre coefficients in logarithmic space.

Writing Sℓ=cℓHDH2QℓS_\ell= c_\ell^{\mathrm{HD}} H_2 Q_\ell, with Qℓ∼Gamma⁡(nℓ,nℓ)Q_\ell\sim\operatorname{Gamma}(n_\ell,n_\ell) independent of H2H_2, the cumulant-generating function of ln⁡Sℓ\ln S_\ell is

Kℓ(t)≡ln⁡E[Sℓt]=tln⁡(cℓHDH‾2)+Kln⁡h(t)+ln⁡Γ(nℓ+t)Γ(nℓ)−tln⁡nℓ.(10)\begin{aligned} K_\ell(t) \equiv\ln\mathbb{E}[S_\ell^t] \\ &= t \ln(c_\ell^{\mathrm{HD}}\overline{H}_2) + K_{\ln h}(t) + \ln\frac{\Gamma(n_\ell+t)}{\Gamma(n_\ell)} - t\ln n_\ell. \tag*{(10)} \end{aligned}

where H‾2=E[H2]\overline{H}_2 = \mathbb{E}[H_2] is the ensemble-averaged total GW power, h=H2/H‾2h = H_2/\overline{H}_2 is the ensemble-normalized power, and Kln⁡h(t)=ln⁡E[ht]K_{\ln h}(t) = \ln\mathbb{E}[h^t]. Differentiating Eq. (10) at t=0t = 0 gives

κ1(ln⁡Sℓ)=ln⁡(cℓHDH‾2)+κ1(ln⁡h)+ψ(nℓ)−ln⁡nℓ,κn(ln⁡Sℓ)=κn(ln⁡h)+ψn−1(nℓ),n≥2,(11)\begin{aligned} \kappa_1(\ln S_\ell) &= \ln(c_\ell^{\mathrm{HD}}\overline{H}_2) + \kappa_1(\ln h) + \psi(n_\ell) - \ln n_\ell,\\ \kappa_n(\ln S_\ell) &= \kappa_n(\ln h) + \psi_{n-1}(n_\ell), \qquad n \ge2, \tag*{(11)} \end{aligned}

where ψ\psi and ψn\psi_n denote the digamma and polygamma functions. The higher-order log-cumulants separate the two sources of fluctuations: κn(ln⁡h)\kappa_n(\ln h) describes fluctuations in the total source power, while ψn−1(nℓ)\psi_{n-1}(n_\ell) describes angular realization fluctuations at fixed power.

Retaining the first two log-cumulants yields

ln⁡Sℓ≃N(μℓ,σℓ2),μℓ=κ1(ln⁡Sℓ),σℓ2=κ2(ln⁡Sℓ).(12)\ln S_\ell\simeq\mathcal{N}(\mu_\ell,\sigma_\ell^2), \qquad\mu_\ell= \kappa_1(\ln S_\ell), \quad\sigma_\ell^2 = \kappa_2(\ln S_\ell). \tag*{(12)}

These parameters match the logarithmic rather than ordinary moments of SℓS_\ell. At large ℓ\ell, the angular log-cumulants satisfy ψ1(nℓ)∼nℓ−1\psi_1(n_\ell) \sim n_\ell^{-1}, ψ2(nℓ)∼−nℓ−2\psi_2(n_\ell) \sim-n_\ell^{-2}, and ψ3(nℓ)∼2nℓ−3\psi_3(n_\ell) \sim2n_\ell^{-3}, so the standardized skewness and excess kurtosis vanish as O(nℓ−1/2)\mathcal{O}(n_\ell^{-1/2}) and O(nℓ−1)\mathcal{O}(n_\ell^{-1}). Thus, the angular contribution approaches Gaussianity, while the full lognormal approximation also depends on the higher log-cumulants of H2H_2.

At fixed H2H_2, different multipoles are independent in the conditional Gamma model, but share the same random power after population averaging. Their leading joint approximation is therefore multivariate lognormal,

(ln⁡S2,ln⁡S3,…)≃N(μ,Σ),Σℓℓ′=Var⁡[ln⁡h]+δℓℓ′ψ1(nℓ),(13)(\ln S_2,\ln S_3,\ldots) \simeq\mathcal{N}(\boldsymbol{\mu},\boldsymbol{\Sigma}), \qquad\Sigma_{\ell\ell'} = \operatorname{Var}[\ln h] + \delta_{\ell\ell'}\psi_1(n_\ell), \tag*{(13)}

where μℓ=κ1(ln⁡Sℓ)\mu_\ell= \kappa_1(\ln S_\ell). The common covariance term reflects population-power fluctuations, while the diagonal term describes independent interference-induced angular fluctuations. The accuracy of the lognormal approximation ultimately depends on the higher log-cumulants of H2H_2.

Figure 4 compares the simulated joint distribution of ensemble-normalized Legendre coefficients S2/(c2HDHˉ2)S_2/(c_2^{\mathrm{HD}}\bar{H}_2) and S3/(c3HDHˉ2)S_3/(c_3^{\mathrm{HD}}\bar{H}_2) at f=2 nHzf = 2\,\mathrm{nHz} with predictions obtained by marginalizing the conditional Gamma and non-central χ2\chi^2 models over the source-power distribution, as well as the multivariate lognormal approximation. While less accurate than the full non-central χ2\chi^2 model, the lognormal approximation captures the main features of the joint distribution, including correlations between multipoles, using only the first two log-cumulants. It thus provides a compact alternative to explicit population marginalization. Neglecting interference, by contrast, yields substantially narrower distributions.

Image not extracted

Figure 4. Joint distribution of the ensemble-normalized Legendre coefficients S2/(c2HDHˉ2)S_2/(c_2^{\mathrm{HD}}\bar{H}_2) and S3/(c3HDHˉ2)S_3/(c_3^{\mathrm{HD}}\bar{H}_2) at f=2 nHzf = 2\,\mathrm{nHz}. The central panel shows the joint PDF, with the corresponding marginal distributions shown in the top and right panels. Contours enclose the 68% and 95% probability regions. Predictions based on Gamma (blue dashed lines) and non-central χ2\chi^2 (red solid lines) models for QℓQ_\ell both agree well with the simulations in this low-frequency regime. The lognormal approximation (green dot-dashed) is similarly accurate, whereas neglecting wave interference (orange solid lines) substantially underestimates the realization-to-realization scatter.

Conclusions

PTAs probe the GW background through correlated variations in pulsar timing signals. In this Letter, we have characterized the realization-dependent Earth-term angular correlation through its frequency-resolved Legendre coefficients SℓS_\ell. These coefficients separate into a common total-power term cℓHDH2c_\ell^{\mathrm{HD}} H_2 and a wave-interference term IℓI_\ell, describing fluctuations in the overall amplitude and angular structure of the correlation.

In the low-frequency regime dominated by numerous weak sources, the power-normalized multipoles Qℓ=Sℓ/(cℓHDH2)Q_\ell= S_\ell/(c_\ell^{\mathrm{HD}} H_2) follow a conditional Gamma distribution with shape and rate nℓ=2ℓ+1n_\ell= 2\ell+ 1. When bright sources become important, a hybrid model yields non-central χ2\chi^2 statistics that capture their effects on angular fluctuations. Marginalizing over source-population fluctuations gives the joint distributions of SℓS_\ell and correlations between different multipoles. Our SMBHB population simulations reveal a transition from interference-dominated variance at low frequencies to population-power-dominated variance at high frequencies. In the many-weak-source regime, we further derive a multivariate lognormal approximation that captures the main features of the joint distributions using only the first two log-cumulants.

These results provide statistical benchmarks for the realized angular correlation, complementing the ensemble-averaged GW spectrum and Hellings–Downs correlation. Extending this framework to fixed, noisy pulsar arrays requires incorporating their response and measurement uncertainties, which we leave for future work.

Acknowledgments

QGH is supported by the National Natural Science Foundation of China (Grant No. 12547110, 12475065, 12447101). YMW is supported by the National Natural Science Foundation of China (Grant No. 12505086).

References

SUPPLEMENTAL MATERIAL

Yan-Chen Bi; Yu-Mei Wu; Qing-Guo Huang

SMBHB population model

We summarize the population prescription used for the Monte Carlo calculation. The ensemble-mean squared characteristic strain, hˉc2(f)\bar{h}_{c}^{2}(f), from SMBHBs is

hˉc2(f)=∫dM dq dz d4Ndz dM dq dln⁡frh2(fr),(14)\bar{h}_{c}^{2}(f)=\int dM\,dq\,dz\,\frac{d^{4}N}{dz\,dM\,dq\,d\ln f_{r}}h^{2}(f_{r}), \tag*{(14)}

where MM, qq and zz are the primary black-hole mass, mass ratio, and redshift, and fr=(1+z)ff_{r}=(1+z)f is the rest-frame GW frequency. The sky- and polarization-averaged GW spectral strain of a single circular binary is

h(fr)=810(GM)5/3c4dc(πfr)2/3,(15)h(f_{r})=\frac{8}{\sqrt{10}}\frac{(\mathcal{G}M)^{5/3}}{c^{4}d_{c}}(\pi f_{r})^{2/3}, \tag*{(15)}

with chirp mass M\mathcal{M} and comoving distance dcd_{c}. The number of binaries per logarithmic rest-frame frequency interval is related to the galaxy differential merger rate by

d4Ndz dM dq dln⁡fr=d3nGdz dMG dqGdMGdMdqGdqτ(fr)4πc(1+z)dc2.(16)\frac{d^{4}N}{dz\,dM\,dq\,d\ln f_{r}}=\frac{d^{3}n_{G}}{dz\,dM_{G}\,dq_{G}}\frac{dM_{G}}{dM}\frac{dq_{G}}{dq}\tau(f_{r})4\pi c(1+z)d_{c}^{2}. \tag*{(16)}

Here, cc is the speed of light, MGM_{G} and qGq_{G} are the primary galaxy mass and the mass ratio between the paired galaxy. The adopted black-hole-bulge relation gives q=qGα∗q=q_{G}^{\alpha_{*}} with α∗\alpha_{*} a model parameter. For binaries evolving only through GW emission, the binary hardening timescale is given by

τ(fr)=frdtrdfr=596(GMc3)−5/3(πfr)−8/3.(17)\tau(f_{r})=f_{r}\frac{dt_{r}}{df_{r}}=\frac{5}{96}\left(\frac{\mathcal{G}M}{c^{3}}\right)^{-5/3}(\pi f_{r})^{-8/3}. \tag*{(17)}

Following [25, 28], the differential galaxy merger rate is

d3nGdz dMG dqG=Φ(MG,z′)MGln⁡10F(MG,z′,qG)τ(MG,z′,qG)∣dtrdz′∣,(18)\frac{d^{3}n_{G}}{dz\,dM_{G}\,dq_{G}}=\frac{\Phi(M_{G},z')}{M_{G}\ln10}\frac{\mathcal{F}(M_{G},z',q_{G})}{\tau(M_{G},z',q_{G})}\left|\frac{dt_{r}}{dz'}\right|, \tag*{(18)}

where z′z' is the redshift at which the galaxy pair forms. It is related to the binary-merger redshift zz by

∫zz′∣dtdz~∣dz~=τ(MG,z′,qG),(19)\int_{z}^{z'}\left|\frac{dt}{d\tilde{z}}\right|d\tilde{z}=\tau(M_{G},z',q_{G}), \tag*{(19)}

where the relationship between time and redshift assumes a flat Λ\LambdaCDM Universe. The merger timescale is parameterized as [29]

τ(MG,z′,qG)=τ0(MGbMG0)ατ(1+z′)βτqGγτ,(20)\tau(M_{G},z',q_{G})=\tau_{0}\left(\frac{M_{G}}{bM_{G0}}\right)^{\alpha_{\tau}}(1+z')^{\beta_{\tau}}q_{G}^{\gamma_{\tau}}, \tag*{(20)}

where τ0\tau_{0} represents merger time normalization with unit of Gyr. Meanwhile, bMG0=0.4×1011M⊙/h0bM_{G0}=0.4\times10^{11}M_{\odot}/h_{0} is an reference mass.

Model ComponentParametersDescriptionFiducial values
GSMF [25]Φ0\Phi_{0}GSMF norm-2.78
GSMF [25]ΦI\Phi_{I}GSMF norm redshift evolution-0.27
GSMF [25]MG0M_{G0}GSMF scaling mass+11.3
GSMF [25]α0\alpha_{0}GSMF mass slope-1.24
GSMF [25]αI\alpha_{I}GSMF mass slope redshift evolution-0.03
Pair fraction [25]f0′f'_{0}pair fraction norm0.04
Pair fraction [25]αf\alpha_{f}pair fraction mass slope+0.07
Pair fraction [25]βf\beta_{f}pair fraction redshift slope1.20
Pair fraction [25]γf\gamma_{f}pair fraction mass ratio slope0.002
Merger timescale [25]τ0\tau_{0}merger time norm0.67
Merger timescale [25]ατ\alpha_{\tau}merger time mass slope-0.12
Merger timescale [25]βτ\beta_{\tau}merger time redshift slope-1.69
Merger timescale [25]γτ\gamma_{\tau}merger time mass ratio slope+0.007
Mbulge−MBHM_{\mathrm{bulge}} - M_{\mathrm{BH}} [33]α∗\alpha_{*}Mbulge−MBHM_{\mathrm{bulge}} - M_{\mathrm{BH}} relation slope+1.02
Mbulge−MBHM_{\mathrm{bulge}} - M_{\mathrm{BH}} [33]M∗M_{*}Mbulge−MBHM_{\mathrm{bulge}} - M_{\mathrm{BH}} relation norm8.38
Mbulge−MBHM_{\mathrm{bulge}} - M_{\mathrm{BH}} [33]f∗,bulgef_{*,\mathrm{bulge}}MG−MbulgeM_{G} - M_{\mathrm{bulge}} relation ratio0.615
Mbulge−MBHM_{\mathrm{bulge}} - M_{\mathrm{BH}} [33]ϵ\epsilonMbulge−MBHM_{\mathrm{bulge}} - M_{\mathrm{BH}} relation scatter0.37

Table 1. TABLE I. Fiducial values of parameters used in the analyses.

The galaxy stellar mass function (GSMF) is represented by a single Schechter form [30, 31]

Φ(MG,z′)=10Φ(z′)(ln⁡10)(MGMG0)1+α(z′)exp⁡(−MGMG0),(21)\Phi(M_{G},z') = 10^{\Phi(z')}(\ln10)\left(\frac{M_{G}}{M_{G0}}\right)^{1+\alpha(z')} \exp\left(-\frac{M_{G}}{M_{G0}}\right), \tag*{(21)}

where Φ(z′)=Φ0+z′ΦI\Phi(z') = \Phi_{0} + z'\Phi_{I}, and α(z′)=α0+z′αI\alpha(z') = \alpha_{0} + z'\alpha_{I}. The differential pair fraction is [25]

F(MG,z′,qG)=f0′(MGaMG0)αf(1+z′)βfqGγf,(22)\mathcal{F}(M_{G},z',q_{G}) = f'_{0}\left(\frac{M_{G}}{aM_{G0}}\right)^{\alpha_{f}}(1+z')^{\beta_{f}}q_{G}^{\gamma_{f}}, \tag*{(22)}

where aMG0=1011M⊙aM_{G0} = 10^{11}M_{\odot} is an reference mass [25]. The galaxy mass MGM_{G} is mapped to the primary black-hole mass MBHM_{\mathrm{BH}} through bulge relation [32, 33]. We use Mbulge=f∗,bulgeMGM_{\mathrm{bulge}} = f_{*,\mathrm{bulge}}M_{G} and

MBH=N{M∗(Mbulge1011M⊙)α∗,ϵ},(23)M_{\mathrm{BH}} = \mathcal{N}\left\{M_{*}\left(\frac{M_{\mathrm{bulge}}}{10^{11}M_{\odot}}\right)^{\alpha_{*}},\epsilon\right\}, \tag*{(23)}

where M∗M_{*}, α∗\alpha_{*}, and ϵ\epsilon are the normalization, slope, and logarithmic scatter. N{x,y}\mathcal{N}\{x,y\} a log normal distribution with mean xx and standard deviation yy. The fiducial parameters, calibrated to the NANOGrav 15-year data set, are listed in Table. I.

Monte Carlo Realizations

The SMBHB populations are generated with holodeck [27], which follows semi-analytic prescriptions prescriptions above and the calibration of [25, 26]. These populations produce an expectation value for the number of binaries as a function of their parameters. To generate individual realizations of the Universe, we sample from a Poisson distribution around this expected value

hc2(fp)=∑M,q,z,fP(ΔMΔqΔzΔln⁡fd4NdzdMdqdln⁡fr)h2(fp)Δln⁡f,(24)h_{c}^{2}(f_{p}) = \sum_{M,q,z,f} \mathcal{P}\left(\Delta M\Delta q\Delta z\Delta\ln f\frac{d^{4}N}{dzdM dqd\ln f_{r}}\right)\frac{h^{2}(f_{p})}{\Delta\ln f}, \tag*{(24)}

where P\mathcal{P} represents the Poisson sampling process. This sampling produces one realization of the binary population. Source directions and initial phases are then drawn independently and isotropically.

Legendre representation of the GWB in PTAs

The GW contribution to pulsar-timing data can be described either by the redshift Za(t)Z_a(t) of pulsar aa, or equivalently by the timing residual ra(t)=∫tdt′Za(t′)r_a(t)=\int^t dt' Z_a(t'). Throughout this work, we use the complex analytic redshift, with the physical redshift given by its real part. For a superposition of discrete sources,

Za(t)=∑jAjei(2πfjt+Φj)Ra(Ω^j,fj),(25)Z_a(t)=\sum_j A_j e^{i(2\pi f_jt+\Phi_j)}R_a(\hat{\Omega}_j,f_j), \tag*{(25)}

where fjf_j, Φj\Phi_j, and Ω^j\hat{\Omega}_j denote the observed GW frequency, initial phase, and propagation direction of GW source jj, respectively. Each binary is taken to be monochromatic over the PTA observing baseline, hj=Ajexp⁡[i(2πfjt+Φj)]h_j=A_j\exp[i(2\pi f_jt+\Phi_j)], with polarization-averaged amplitude

Aj=325(GMz,j)5/3(πfj)2/3c4dL,j,(26)A_j=\sqrt{\frac{32}{5}}\frac{(G\mathcal{M}_{z,j})^{5/3}(\pi f_j)^{2/3}}{c^4d_{L,j}}, \tag*{(26)}

where Mz,j=(1+zj)Mj\mathcal{M}_{z,j}=(1+z_j)\mathcal{M}_j and dL,jd_{L,j} are the redshifted chirp mass and luminosity distance of the source. The full pulsar response can be written as

Ra(Ω^j,fj)=[1−e−i2πfjLa(1+Ω^j⋅p^a)/c]Fa(Ω^j),(27)R_a(\hat{\Omega}_j,f_j)=\left[1-e^{-i2\pi f_jL_a(1+\hat{\Omega}_j\cdot\hat{p}_a)/c}\right]F_a(\hat{\Omega}_j), \tag*{(27)}

with the antenna pattern function, Fa(Ω^j)F_a(\hat{\Omega}_j) is given in terms of complex polarization tensor, ϵαβ+,×\epsilon_{\alpha\beta}^{+,\times}, as

Fa(Ω^j)=p^aαp^aβ2(1+Ω^j⋅p^a)(ϵαβ+−iϵαβ×).(28)F_a(\hat{\Omega}_j)=\frac{\hat{p}_a^\alpha\hat{p}_a^\beta}{2(1+\hat{\Omega}_j\cdot\hat{p}_a)}(\epsilon_{\alpha\beta}^{+}-i\epsilon_{\alpha\beta}^{\times}). \tag*{(28)}

where LaL_a and p^a\hat{p}_a are the distance and sky direction of pulsar aa. The polarization tensors have the standard normalization ϵαβAϵBαβ=2δBA\epsilon_{\alpha\beta}^{A}\epsilon_{B}^{\alpha\beta}=2\delta^{A}_{B}. For an observing baseline TT, we define the Fourier coefficient

Z~ap=1T∫−T/2T/2Za(t)e−i2πfptdt=∑jAjeiΦjsinc⁡[πT(fj−fp)]Ra(Ω^j,fj),(29)\tilde{Z}_a^p=\frac{1}{T}\int_{-T/2}^{T/2}Z_a(t)e^{-i2\pi f_pt}dt=\sum_j A_je^{i\Phi_j}\operatorname{sinc}[\pi T(f_j-f_p)]R_a(\hat{\Omega}_j,f_j), \tag*{(29)}

where fp=p/Tf_p=p/T, and p∈{1,2,⋯ ,Nf}p\in\{1,2,\cdots,N_f\} denotes Fourier modes with NfN_f is the number of frequency bins. In the basis of spherical harmonics, the Earth term of redshift z~a\tilde{z}_a could be further expressed as

z~ap=4π∑ℓmCℓHDnℓ 2aℓmpYℓm∗(p^a),(30)\tilde{z}_a^p=4\pi\sum_{\ell m}\sqrt{\frac{C_\ell^{\mathrm{HD}}}{n_\ell}}\,{}^2a_{\ell m}^pY_{\ell m}^{*}(\hat{p}_a), \tag*{(30)}

where nℓ=2ℓ+1n_\ell=2\ell+1, and 2aℓmp=(−1)ℓ∑j2sjpYℓm(Ω^j){}^2a_{\ell m}^p=(-1)^\ell\sum_j{}^2s_j^pY_{\ell m}(\hat{\Omega}_j) with 2sjp=2/3AjeiΦjsinc⁡[πT(fj−fp)]{}^2s_j^p=\sqrt{2/3}A_je^{i\Phi_j}\operatorname{sinc}[\pi T(f_j-f_p)]. In the narrow-bin approximation, we can safely replace the sinc response by an idealized frequency selection, sinc⁡[πT(fj−fp)]∝δ(fj−fp)\operatorname{sinc}[\pi T(f_j-f_p)]\propto\delta(f_j-f_p), neglecting spectral leakage and correlations between frequency bins. We henceforth consider a single bin and suppress the index pp. Noted that the factor 2/3\sqrt{2/3} keeps the HD normalization. The HD multipole coefficients are

CℓHD={0,for ℓ=0,1,3(2ℓ+1)(ℓ−2)!2(ℓ+2)!,for ℓ≥2.(31)C_\ell^{\mathrm{HD}}= \begin{cases} 0, & \text{for }\ell=0,1,\\ \dfrac{3(2\ell+1)(\ell-2)!}{2(\ell+2)!}, & \text{for }\ell\ge2. \end{cases} \tag*{(31)}

Averaging over rotations of a pulsar pair at fixed angular separation γab\gamma_{ab} gives

⟨Yℓm(p^a)Yℓ′m′∗(p^b)⟩γab=δℓℓ′δmm′4πPℓ(cos⁡γ).(32)\left\langle Y_{\ell m}(\hat{p}_a)Y_{\ell' m'}^{*}(\hat{p}_b)\right\rangle_{\gamma_{ab}}=\frac{\delta_{\ell\ell'}\delta_{mm'}}{4\pi}P_\ell(\cos\gamma). \tag*{(32)}

Consequently, the orientation-averaged Earth-term angular correlation reads

C(γab)=ℜ⟨z~az~b∗⟩γab=∑ℓ≥2SℓPℓ(cos⁡γ),(33)C(\gamma_{ab})=\Re\left\langle\tilde{z}_a\tilde{z}_b^{*}\right\rangle_{\gamma_{ab}}=\sum_{\ell\ge2}S_\ell P_\ell(\cos\gamma), \tag*{(33)}

where the angular-correlation multipoles are

Sℓ=14π∑m∣zℓm∣2≥0.(34)S_{\ell} = \frac{1}{4\pi}\sum_m \lvert z_{\ell m}\rvert^2 \ge0. \tag*{(34)}

A useful normalized vector and equivalent spin notation are

bℓm=zℓm4πcℓHD=4πnℓ 2aℓm.(35)b_{\ell m} = \frac{z_{\ell m}}{\sqrt{4\pi c_{\ell}^{\rm HD}}} = \sqrt{\frac{4\pi}{n_{\ell}}}\,{}^{2}a_{\ell m}. \tag*{(35)}

The Legendre coefficients can then be written as

Sℓ=cℓHD∥Bℓ∥2=cℓHD4πnℓ∑m∣2aℓm∣2=cℓHD4πnℓ∑m∣∑jsj2Yℓm(Ω^j)∣2.(36)S_{\ell} = c_{\ell}^{\rm HD}\lVert B_{\ell}\rVert^2 = c_{\ell}^{\rm HD}\frac{4\pi}{n_{\ell}}\sum_m \lvert{}^{2}a_{\ell m}\rvert^2 = c_{\ell}^{\rm HD}\frac{4\pi}{n_{\ell}}\sum_m\left\lvert\sum_j s_j{}^{2}Y_{\ell m}(\hat{\Omega}_j)\right\rvert^2. \tag*{(36)}

If wave interference is neglected, the coherent sum is replaced by an incoherent sum of individual source powers,

Sℓnon−intf=cℓHD4πnℓ∑m∑j∣sj2Yℓm(Ω^j)∣2=cℓHD∑j∣sj∣2.(37)S_{\ell}^{\rm non-intf} = c_{\ell}^{\rm HD}\frac{4\pi}{n_{\ell}}\sum_m\sum_j\lvert s_j{}^{2}Y_{\ell m}(\hat{\Omega}_j)\rvert^2 = c_{\ell}^{\rm HD}\sum_j\lvert s_j\rvert^2. \tag*{(37)}

Thus, the distinction between the interference and non-interference cases is whether source amplitudes are summed coherently before taking the power or their powers are summed incoherently.

Angular-power distribution of an interfering GWB

We define the source-power moments within a frequency bin as Hn≡∑j∣sj∣nH_n \equiv\sum_j \lvert s_j\rvert^n. For isotropically distributed GW sources with independent random phases,

∫dΦ2π∫dΩ^4π(2aℓm2aℓ′m′∗)=H24πδℓℓ′δmm′.(38)\int\frac{\mathrm{d}\Phi}{2\pi}\int\frac{\mathrm{d}\hat{\Omega}}{4\pi}\left({}^{2}a_{\ell m}{}^{2}a_{\ell' m'}^*\right) = \frac{H_2}{4\pi}\delta_{\ell\ell'}\delta_{mm'}. \tag*{(38)}

If the effective number of sources, NeffN_{\rm eff}, contributing to the frequency bin is large and no single source dominates, max⁡j∣sj∣2/H2≪1\max_j\lvert s_j\rvert^2/H_2 \ll1, the central-limit theorem implies Gaussian approximation

2aℓm∣H2≃CN(0,H24π).(39){}^{2}a_{\ell m}\mid H_2 \simeq\mathcal{CN}\left(0,\frac{H_2}{4\pi}\right). \tag*{(39)}

Importantly, this is a Gaussian approximation for the angular coefficients at fixed total power. It does not imply that H2H_2 itself is Gaussian, the single-source power distribution is heavy tailed and can generate strongly non-Gaussian population-to-population fluctuations in H2H_2. For the properly normalized complex Gaussian distribution in Eq. (39), it follows

4πH2∣2aℓm∣2∼Gamma⁡(1,1).(40)\frac{4\pi}{H_2}\lvert{}^{2}a_{\ell m}\rvert^2 \sim\operatorname{Gamma}(1,1). \tag*{(40)}

where the Gamma distribution is parameterized by shape and rate. Summing over the nℓ=2ℓ+1n_{\ell}=2\ell+1 independent mm modes therefore gives the power-normalized multipole Qℓ=Sℓ/(cℓHDH2)Q_{\ell}=S_{\ell}/(c_{\ell}^{\rm HD}H_2) as

Qℓ∣H2∼Gamma⁡(nℓ,nℓ).(41)Q_{\ell}\mid H_2 \sim\operatorname{Gamma}(n_{\ell},n_{\ell}). \tag*{(41)}

Equivalently,

p(Sℓ∣H2)=nℓnℓSℓnℓ−1Γ(nℓ)(cℓHDH2)nℓexp⁡[−nℓSℓcℓHDH2],(42)p(S_{\ell}\mid H_2) = \frac{n_{\ell}^{n_{\ell}}S_{\ell}^{n_{\ell}-1}}{\Gamma(n_{\ell})(c_{\ell}^{\rm HD}H_2)^{n_{\ell}}}\exp\left[-\frac{n_{\ell}S_{\ell}}{c_{\ell}^{\rm HD}H_2}\right], \tag*{(42)}

with Γ(nℓ)\Gamma(n_{\ell}) denotes the gamma function. This distribution also implies that the Sℓ>0S_{\ell}>0 should be fulfilled. Thus, only positive values on SℓS_{\ell} are physically valid. Moreover, the corresponding skewness and (excess) kurtosis are

Skew⁡(Qℓ∣H2)=2nℓ,(43)\operatorname{Skew}(Q_{\ell}\mid H_2)=\frac{2}{\sqrt{n_{\ell}}}, \tag*{(43)}
Kurt⁡(Qℓ∣H2)=6nℓ.(44)\operatorname{Kurt}(Q_{\ell}\mid H_2)=\frac{6}{n_{\ell}}. \tag*{(44)}

Bright-source conditional distribution

Realistic SMBHB populations span a broad range of amplitudes, and one or a few binaries can contribute an appreciable fraction of the GW power in a frequency bin. We therefore separate the coherent field into an explicit bright-source contribution and an unresolved faint component, the latter being approximated as a proper complex Gaussian field. For a single explicit bright source, 2aℓm=(−1)ℓ[scw2Yℓm(Ω^cw)+∑jsj2Yℓm(Ω^j)]{}_2a_{\ell m}=(-1)^\ell\left[s_{\rm cw}{}_2Y_{\ell m}(\hat{\Omega}_{\rm cw})+\sum_j s_j{}_2Y_{\ell m}(\hat{\Omega}_j)\right]. The former one denotes contributions from bright source, while the latter still denotes unresolvable GWBs. The Legendre coefficients can correspondingly be decomposed as

Sℓ=SℓB+Sℓu+Iℓ.(45)S_\ell=S_\ell^{\rm B}+S_\ell^{\rm u}+I_\ell. \tag*{(45)}

The first two terms denote the bright-source and unresolved-background contributions, while the wave-interference term IℓI_\ell contains their interference. Thus, a bright source does not simply add power to the background. It interferes coherently with the unresolved field.

At fixed complete bright-source catalog bℓBb_\ell^{\rm B} and unresolved-background power HuH_u, write Bℓ=bℓB+uℓB_\ell=b_\ell^{\rm B}+u_\ell, with uℓ∼CN(0,HuI/nℓ)u_\ell\sim\mathcal{CN}(0,H_uI/n_\ell), and define WℓB=∥bℓB∥2W_\ell^{\rm B}=\lVert b_\ell^{\rm B}\rVert^2. The conditional proper-complex Gaussian density is

p(Bℓ∣bℓB,Hu)=(πHu/nℓ)−nℓexp⁡[−∥Bℓ−bℓB∥2nℓ/Hu].(46)p(B_\ell\mid b_\ell^{\rm B},H_u)=(\pi H_u/n_\ell)^{-n_\ell}\exp\left[-\lVert B_\ell-b_\ell^{\rm B}\rVert^2n_\ell/H_u\right]. \tag*{(46)}

Utilizing the polynomial variation as

∥B−b∥2ν+tℓcHD∥B∥2=(ν−1+tℓcHD)∥B−b1+tℓcHDν∥2+tℓcHD∥b∥21+tℓcHDν.(47)\frac{\lVert B-b\rVert^2}{\nu}+t_\ell^{\rm cHD}\lVert B\rVert^2=(\nu^{-1}+t_\ell^{\rm cHD})\left\lVert B-\frac{b}{1+t_\ell^{\rm cHD}\nu}\right\rVert^2+\frac{t_\ell^{\rm cHD}\lVert b\rVert^2}{1+t_\ell^{\rm cHD}\nu}. \tag*{(47)}

The integral of the shifted complex Gaussian gives Laplace transform

Lℓ(t∣bℓB,Hu)=(1+tℓcHDHu/nℓ)−nℓexp⁡[−tℓcHDWℓB1+tℓcHDHu/nℓ].(48)\mathcal{L}_\ell(t\mid b_\ell^{\rm B},H_u)=\left(1+t_\ell^{\rm cHD}H_u/n_\ell\right)^{-n_\ell}\exp\left[-\frac{t_\ell^{\rm cHD}W_\ell^{\rm B}}{1+t_\ell^{\rm cHD}H_u/n_\ell}\right]. \tag*{(48)}

Each real or imaginary part of uℓmu_{\ell m} has variance Hu/(2nℓ)H_u/(2n_\ell). Consequently,

2SℓnℓcℓHDHu ∣ bℓB,Hu∼χ2nℓ2(2nℓWℓBHu).(49)\frac{2S_\ell n_\ell}{c_\ell^{\rm HD}H_u}\,\bigg|\,b_\ell^{\rm B},H_u\sim\chi^2_{2n_\ell}\left(\frac{2n_\ell W_\ell^{\rm B}}{H_u}\right). \tag*{(49)}

Thus, the Gamma distribution is recovered as the zero-non-centrality limit of the more general non-central family. Differentiating the log transform gives conditional cumulants,

κn(Sℓ∣bℓB,Hu)=(n−1)!(cℓHD)n[nνℓn+nWℓBνℓn−1].(50)\kappa_n(S_\ell\mid b_\ell^{\rm B},H_u)=(n-1)!(c_\ell^{\rm HD})^n\left[n\nu_\ell^n+nW_\ell^{\rm B}\nu_\ell^{n-1}\right]. \tag*{(50)}

In particular, the mean is E(Sℓ∣bℓB,Hu)=cℓHD(Hu+WℓB)\mathop{\rm E}(S_\ell\mid b_\ell^{\rm B},H_u)=c_\ell^{\rm HD}(H_u+W_\ell^{\rm B}) and the variance is Var(Sℓ∣bℓB,Hu)=(cℓHD)2(Hu2+2HuWℓB)/nℓ\mathop{\rm Var}(S_\ell\mid b_\ell^{\rm B},H_u)=(c_\ell^{\rm HD})^2(H_u^2+2H_uW_\ell^{\rm B})/n_\ell. Defining θℓ=cℓHDHu/nℓ\theta_\ell=c_\ell^{\rm HD}H_u/n_\ell and ηℓ=WℓBnℓ/Hu\eta_\ell=W_\ell^{\rm B}n_\ell/H_u, the conditional density may also be written in the standard non-central Gamma expansion,

p(Sℓ∣bℓB,Hu)=e−ηℓ∑k=0∞ηℓkk!Sℓnℓ+k−1e−Sℓ/θℓΓ(nℓ+k)θℓnℓ+k.(51)p(S_\ell\mid b_\ell^{\rm B},H_u)=e^{-\eta_\ell}\sum_{k=0}^{\infty}\frac{\eta_\ell^k}{k!}\frac{S_\ell^{n_\ell+k-1}e^{-S_\ell/\theta_\ell}}{\Gamma(n_\ell+k)\theta_\ell^{n_\ell+k}}. \tag*{(51)}

The Poisson index kk in this expansion is an auxiliary mathematical variable.

At fixed complete bright-source catalog, the unresolved Gaussian components are independent between multipoles. The population-averaged joint Laplace transform is therefore

L(t)=EbℓB,Hu[∏ℓLℓ(tℓ∣bℓB,Hu)].(52)\mathcal{L}(t)=\mathop{\rm E}_{b_\ell^{\rm B},H_u}\left[\prod_\ell\mathcal{L}_\ell(t_\ell\mid b_\ell^{\rm B},H_u)\right]. \tag*{(52)}

Importantly, we need to emphasis that conditioning only on the individual bright-source powers is insufficient when several explicit sources are retained. Because their contributions add coherently in bℓBb_\ell^{\rm B}, the quantity WℓB=∥bℓB∥2W_\ell^{\rm B}=\lVert b_\ell^{\rm B}\rVert^2 also depends on their sky directions and relative phases. These variables must therefore be marginalized jointly over the complete bright-source catalog. Independently marginalizing the multipole densities would erase this shared catalog dependence.

RegimeNeffN_{\mathrm{eff}}fBf_{\mathrm{B}}SℓS_{\ell} statistics
Many faint sources≫1\gg1≪1\ll1Gamma
Bright + faint≫1\gg10<fB<10 < f_{\mathrm{B}} < 1non-central χ2\chi^2
Bright dominated—∼1\sim1CW-like limit

Table II. Schematic regimes of the bright-source and background model.

For one source, WℓB=HBW_{\ell}^{\mathrm{B}} = H_{\mathrm{B}} for every ℓ\ell. Therefore, their physical meaning is more clear and practical. If HuH_{\mathrm{u}} is fixed, Eq. (52) reduces to

L(t)=∏ℓ(1+tϵℓHDνℓ)−nℓLHB(∑ℓtϵℓHDνℓ1+tϵℓHDνℓ).(53)\mathcal{L}(t) = \prod_{\ell}(1+t\epsilon_{\ell}^{\mathrm{HD}}\nu_{\ell})^{-n_{\ell}\mathcal{L}_{H_{\mathrm{B}}}}\left(\sum_{\ell}\frac{t\epsilon_{\ell}^{\mathrm{HD}}\nu_{\ell}}{1+t\epsilon_{\ell}^{\mathrm{HD}}\nu_{\ell}}\right). \tag*{(53)}

Defining the total power H2=Hu+HBH_{2}=H_{\mathrm{u}}+H_{\mathrm{B}} and bright-source fraction fB=HB/H2f_{\mathrm{B}}=H_{\mathrm{B}}/H_{2}, the power-normalized multipole gives

Qℓ∣fB∼1−fB2nℓ χ2nℓ2(2nℓfB1−fB).(54)Q_{\ell}\mid f_{\mathrm{B}}\sim\frac{1-f_{\mathrm{B}}}{2n_{\ell}}\,\chi^{2}_{2n_{\ell}}\left(\frac{2n_{\ell}f_{\mathrm{B}}}{1-f_{\mathrm{B}}}\right). \tag*{(54)}

Its mean and variance are E[Qℓ]=1\mathbb{E}[Q_{\ell}]=1, and Var⁡[Qℓ]=(1−fB2)/nℓ\operatorname{Var}[Q_{\ell}]=(1-f_{\mathrm{B}}^{2})/n_{\ell}. An alternative parameter is ηB=HB/Hu\eta_{\mathrm{B}}=H_{\mathrm{B}}/H_{\mathrm{u}}, with fB=ηB/(1+ηB)f_{\mathrm{B}}=\eta_{\mathrm{B}}/(1+\eta_{\mathrm{B}}). The limit fB→0f_{\mathrm{B}}\to0 recovers the Gamma distribution, whereas at the limit fB→1f_{\mathrm{B}}\to1, Eq. (54) approaches δ(Q−1)\delta(Q-1) corresponding to the deterministic single-source limit (CW-like) after pulsar averaging. Figure 5 illustrates how the angular variance evolves with the bright-source fraction fBf_{\mathrm{B}}, showing its progressive suppression as the signal becomes increasingly dominated by a single coherent source. Moreover, the higher moments are

Image not extracted

Figure 5. Demonstration of cosmic variance for different values of fBf_B. Effect of a bright source on the realization-dependent angular correlation. Left panel: Normalized correlation Γ(γ)/H2\Gamma(\gamma)/H_2 for representative bright-source fractions fBf_B values, compared with the HD expectation. Right panel: Variance of Γ/H2\Gamma/H_2 (black solid line), normalized to the Gaussian-background value (grey dotted line), together with its background self-term (red dashed line) and wave-interference term (blue dash-doted line). Increasing fBf_B suppresses the total angular realization variance, which vanishes in the deterministic single-source limit.

Skew⁡(Qℓ∣fB)=2nℓ1−fB(1+2fB)(1+fB)3/2.(55)\operatorname{Skew}(Q_{\ell}\mid f_{\mathrm{B}})=\frac{2}{\sqrt{n_{\ell}}}\frac{\sqrt{1-f_{\mathrm{B}}}(1+2f_{\mathrm{B}})}{(1+f_{\mathrm{B}})^{3/2}}. \tag*{(55)}
Kurt⁡(Qℓ∣fB)=6nℓ(1−fB)(1+3fB)(1+fB)2.(56)\operatorname{Kurt}(Q_{\ell}\mid f_{\mathrm{B}})=\frac{6}{n_{\ell}}\frac{(1-f_{\mathrm{B}})(1+3f_{\mathrm{B}})}{(1+f_{\mathrm{B}})^{2}}. \tag*{(56)}

If both HuH_{\mathrm{u}} and HBH_{\mathrm{B}} fluctuate across population realizations and may be statistically dependent, the law of total covariance gives

Cov⁡(Sℓ,Sℓ′)=cℓHDcℓ′HD[δℓℓ′nℓE(Hu2+2HuHB)+Var⁡(H2)].(57)\operatorname{Cov}(S_{\ell},S_{\ell'})=c_{\ell}^{\mathrm{HD}}c_{\ell'}^{\mathrm{HD}}\left[\frac{\delta_{\ell\ell'}}{n_{\ell}}\mathbb{E}(H_{\mathrm{u}}^{2}+2H_{\mathrm{u}}H_{\mathrm{B}})+\operatorname{Var}(H_{2})\right]. \tag*{(57)}

The first term represents interference-induced angular fluctuations at fixed source powers, whereas the second arises from population-to-population fluctuations in the total GW power and correlates different multipoles.

Conditional narrowing therefore does not necessarily imply a smaller population variance. At fixed total power H2H_{2}, the interference-induced variance relative to the Gaussian case is suppressed by the factor 1−fB21-f_{\mathrm{B}}^{2}. By contrast, if HuH_{\mathrm{u}} is held fixed while a bright source is added, the absolute conditional variance is enhanced relative to the unresolved-background value by the factor 1+2ηB1+2\eta_{\mathrm{B}}. A deterministic single source has no pulsar-averaged interference-induced angular fluctuations, although its response to a fixed finite pulsar array remains direction dependent.

For several explicit sources with fixed powers and with their phases and sky directions averaged over, define H=Hu+∑j∣sj∣2H=H_{\mathrm{u}}+\sum_{j}|s_{j}|^{2} and H4=∑j∣sj∣4H_{4}=\sum_{j}|s_{j}|^{4}, where the sums run over the explicit discrete sources. The normalized multipole variance then becomes

Var⁡(Qℓ∣⟨bℓB⟩Φ,Ω^,Hu)=1−rnℓ,r=H4H2.(58)\operatorname{Var}(Q_{\ell}\mid\langle b_{\ell}^{\mathrm{B}}\rangle_{\Phi,\hat{\Omega}},H_{\mathrm{u}})=\frac{1-r}{n_{\ell}},\qquad r=\frac{H_{4}}{H^{2}}. \tag*{(58)}

where ⟨ ⟩Φ,Ω^\langle\ \rangle_{\Phi,\hat{\Omega}} denotes the average over phases and sky directions. For a single explicit source, r=fB2r=f_{\mathrm{B}}^{2}, recovering Eq. (54). The Gaussian unresolved component contributes no discrete ∣sj∣4|s_{j}|^{4} term in this hybrid model. Replacing a genuinely finite faint population by a Gaussian remainder therefore neglects its source-level connected higher-order cumulants.

Autocorrelations

The pulsar autocorrelation requires separate treatment from the cross-correlation between distinct pulsars. Restoring the pulsar term and averaging over pulsar directions and the rapidly varying pulsar phase gives

⟨z~az~a∗⟩=∑ℓ=2∞Sℓ+H22,(59)\langle\tilde{z}_a\tilde{z}_a^* \rangle= \sum_{\ell=2}^{\infty} S_\ell+ \frac{H_2}{2}, \tag*{(59)}

in agreement with Ref. [26]. The additional term H2/2H_2/2 arises from the pulsar contribution to the autocorrelation. It modifies the diagonal pulsar covariance but is not part of the Legendre expansion describing the cross-correlation between distinct pulsars.

Within the many-source approximation and at fixed H2H_2, Eq. (59) is a shifted sum of independent Gamma or noncentral-χ2\chi^2 distributed multipoles with different scales. It therefore does not, in general, reduce to a single Gamma distribution. After marginalizing over the population-power distribution pH(H2)p_H(H_2), the autocorrelation distribution can be evaluated numerically, for example by Monte Carlo sampling or numerical inversion of its Laplace transform.

References

  1. [1]R. S. Foster and D. C. Backer, “Constructing a Pulsar Timing Array,” Astrophys. J. 361, 300 (1990).
  2. [2]Bruce Allen, “Variance of the Hellings-Downs correlation,” Phys. Rev. D 107, 043018 (2023), arXiv:2205.05637 [gr-qc].DOI
  3. [3]Vincent Corbin and Neil J. Cornish, “Pulsar Timing Array Observations of Massive Black Hole Binaries,” (2010), arXiv:1008.1782 [astro-ph.HE].arxiv.org/abs/1008.1782
  4. [4]Bruce Allen and Joseph D. Romano, “Hellings and Downs correlation of an arbitrary set of pulsars,” Phys. Rev. D 108, 043026 (2023), arXiv:2208.07230 [gr-qc].DOI
  5. [5]Reginald Christian Bernardo and Kin-Wang Ng, “Pulsar and cosmic variances of pulsar timing-array correlation measurements of the stochastic gravitational wave background,” JCAP 11, 046 (2022), arXiv:2209.14834 [gr-qc].DOI
  6. [6]Yu-Mei Wu, Yan-Chen Bi, and Qing-Guo Huang, “Spatial correlations between pulsars for interfering gravitational-wave sources in pulsar timing array,” Phys. Rev. D 112, 123034 (2025), arXiv:2407.07319 [astro-ph.CO].
  7. [7]Yu-Mei Wu, Yan-Chen Bi, and Qing-Guo Huang, “Spatial correlation between pulsars from interfering gravitational-wave sources in massive gravity,” Phys. Rev. D 112, 064053 (2025), arXiv:2506.07679 [gr-qc].DOI
  8. [8]Yan-Chen Bi, Yu-Mei Wu, and Qing-Guo Huang, “Harmonic Analysis on Correlation for Gravitational-Wave Backgrounds of Arbitrary Polarization from Interfering Sources in Generic Dispersion Relation,” (2026), arXiv:2602.13621 [gr-qc].arxiv.org/abs/2602.13621
  9. [9]Xiao Xue, Zhen Pan, and Liang Dai, “Non-Gaussian statistics of nanohertz stochastic gravitational waves,” Phys. Rev. D 111, 043022 (2025), arXiv:2409.19516 [astro-ph.CO].arxiv.org/abs/2409.19516
  10. [10]123065 (2026), arXiv:2511.09659 [gr-qc].DOI
  11. [11]Valerie Domcke, Gabriele Franciolini, and Mauro Pieroni, “Cosmic Variance in Anisotropy Searches at Pulsar Timing Arrays,” Phys. Rev. Lett. 137, 061003 (2026), arXiv:2508.21131 [astro-ph.CO].arxiv.org/abs/2508.21131
  12. [12]Yacine Ali-Haïmoud, “A practical theorem on gravitational-wave background statistics,” (2026), arXiv:2604.19701 [astro-ph.CO].
  13. [13]Juhan Raidal, Juan Urrutia, Ville Vaskonen, and Hardi Veermäe, “The Heavy Tailed Non-Gaussianity of the Supermassive Black Hole Gravitational Wave Background,” (2026), arXiv:2604.08506 [astro-ph.CO].arxiv.org/abs/2604.08506
  14. [14]Adrien Kuntz, Clemente Smarra, and Massimo Vaglio, “Looking for non-gaussianity in Pulsar Timing Arrays through the four point correlator,” (2026), arXiv:2603.12311 [gr-qc].DOI
  15. [15]Hinano Hisamatsu and Koutarou Kyutoku, “Higher-order statistics of the stochastic gravitational wave background from supermassive black hole binaries,” (2026), arXiv:2605.17983 [astro-ph.HE].
  16. [16]Chiara Cecchini, Jonas El Gammal, Gabriele Franciolini, and Mauro Pieroni, “Are PTA measurements sensitive to gravitational wave non-Gaussianities?” (2026), arXiv:2605.05157 [astro-ph.CO].arxiv.org/abs/2605.05157
  17. [17]Mikel Falxa, “Approximating the statistics of a gravitational wave background,” (2026), arXiv:2609.07686 [astro-ph.IM].arxiv.org/abs/2609.07686
  18. [18]Bruce Allen, “Pulsar timing array harmonic analysis and source angular correlations,” Phys. Rev. D 110, 043043 (2024), arXiv:2404.05677 [gr-qc].DOI
  19. [19]Bruce Allen and Joseph D. Romano, “Optimal Reconstruction of the Hellings and Downs Correlation,” Phys. Rev. Lett. 134, 031401 (2025), arXiv:2407.10968 [gr-qc].DOI
  20. [20]Boris Goncharov, Gabriela Sato-Polito, Xiaoming Bi, and Matias Zaldarriaga, “A Joint Optimal Search for Gravitational Waves from Resolved and Unresolved Supermassive Binary Black Holes with Pulsar Timing Arrays,” (2026), arXiv:2606.18241 [astro-ph.HE].arxiv.org/abs/2606.18241
  21. [21]John Ellis, Malcolm Fairbairn, Gert Hütsi, Juhan Raidal, Juan Urrutia, Ville Vaskonen, and Hardi Veermäe, “Gravitational waves from supermassive black hole binaries in light of the NANOGrav 15-year data,” Phys. Rev. D 109, L021302 (2024), arXiv:2306.17021 [astro-ph.CO].DOI
  22. [22]Gabriela Sato-Polito and Matias Zaldarriaga, “Distribution of the gravitational-wave background from supermassive black holes,” Phys. Rev. D 111, 020343 (2025), arXiv:2406.17010 [astro-ph.CO].
  23. [23]Jiming Yu, Zhen Pan, Xiao Xue, and Liang Dai, “Population statistics of nanohertz gravitational wave sources,” (2026), arXiv:2607.07477 [astro-ph.HE].arxiv.org/abs/2607.07477
  24. [24]Bruce Allen, Deepali Agarwal, Joseph D. Romano, and Serena Valintoliņa, “Source anisotropies and pulsar timing arrays,” Phys. Rev. D 110, 123507 (2024), arXiv:2406.16031 [gr-qc].DOI
  25. [25]Siyuan Chen, Alberto Sesana, and Christopher J. Conselice, “Constraining astrophysical observables of Galaxy and Supermassive Black Hole Binary Mergers using Pulsar Timing Arrays,” Mon. Not. Roy. Astron. Soc. 488, 401–418 (2019), arXiv:1810.04184 [astro-ph.GA].DOI
  26. [26]Yan-Chen Bi, Yu-Mei Wu, Zu-Cheng Chen, and Qing-Guo Huang, “Implications for the supermassive black hole binaries from the NANOGrav 15-year data set,” Sci. China Phys. Mech. Astron. 66, 120402 (2023), arXiv:2307.00722 [astro-ph.CO].DOI
  27. [27]Gabriella Agazie et al. (NANOGrav), “The NANOGrav 15 yr Data Set: Constraints on Supermassive Black Hole Binaries from the Gravitational-wave Background,” Astrophys. J. Lett. 952, L37 (2023), arXiv:2306.16220 [astro-ph.HE].DOI
  28. [28]A. Sesana, A. Vecchio, and M. Volonteri, “Gravitational waves from resolvable massive black hole binary systems and observations with Pulsar Timing Arrays,” Mon. Not. Roy. Astron. Soc. 394, 2555 (2009), arXiv:0809.3412 [astro-ph].DOI
  29. [29]Gregory F. Snyder, Jennifer M. Lotz, Vicente Rodriguez-Gomez, Renato da Silva Guimarães, Paul Torrey, and Lars Hernquist, “Massive close pairs measure rapid galaxy assembly in mergers at high redshift,” Monthly Notices of the Royal Astronomical Society 468, 207–216 (2017).DOI
  30. [30]Alice Mortlock et al., “Deconstructing the Galaxy Stellar Mass Function with UKIDSS and CANDELS: the Impact of Colour, Structure and Environment,” Mon. Not. Roy. Astron. Soc. 447, 2–24 (2015), arXiv:1411.3339 [astro-ph.GA].DOI
  31. [31]Christopher J. Conselice, Aaron Wilkinson, Kenneth Duncan, and Alice Mortlock, “THE EVOLUTION OF GALAXY NUMBER DENSITY AT z< 8 AND ITS IMPLICATIONS,” Astrophys. J. 830, 83 (2016).DOI
  32. [32]A. Sesana, “Insights into the astrophysics of supermassive black hole binaries from pulsar timing observations,” Class. Quant. Grav. 30, 224014 (2013), arXiv:1307.2600 [astro-ph.CO].DOI
  33. [33]John Kormendy and Luis C. Ho, “Coevolution (Or Not) of Supermassive Black Holes and Host Galaxies,” Ann. Rev. Astron. Astrophys. 51, 511–653 (2013), arXiv:1304.7762 [astro-ph.CO].DOI

Paper details

vlnc:W_6c4hdgxx
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents