Polarized and Non-Gaussian Backgrounds

The standard stochastic search assumes an unpolarized, Gaussian tensor field and therefore uses only its second-order cross spectrum. Real astrophysical populations can be anisotropic, circularly polarized, intermittent, or non-Gaussian, while alternative theories can introduce additional propagation or polarization responses. These possibilities require more observables and stronger identifiability checks; excess cross power alone cannot say which extension produced it.

Helpful background. Cross-Correlation for Stochastic Backgrounds defines the baseline statistic and overlap response, and Anisotropic Stochastic Mapping develops the inverse problem and its null space.

Polarization in correlation data

For tensor waves, introduce right- and left-handed amplitudes $h_R$ and $h_L$. The intensity and circular-polarization spectra are proportional to

$$I \propto \left\langle\lvert h_R\rvert^2+\lvert h_L\rvert^2\right\rangle, \qquad V \propto \left\langle\lvert h_R\rvert^2-\lvert h_L\rvert^2\right\rangle .$$

Fix $h_R=(h_+-ih_\times)/\sqrt2$ and $h_L=(h_++ih_\times)/\sqrt2$ in a declared oriented transverse basis. Then $I$ is parity even and $V$ is parity odd; positivity requires $|V|\leq I$. Reversing which helicity is called right-handed reverses the sign assigned to $V$, without changing any observable. Linear-polarization Stokes fields $Q$ and $U$ depend on a sky basis and are naturally treated in an anisotropic spin-weighted expansion. For an isotropic tensor analysis, a baseline cross spectrum can be written schematically as

$$\left\langle C_b(f)\right\rangle = H_I(f)\gamma_b^I(f)I +H_V(f)\gamma_b^V(f)V.$$

At one frequency a single baseline generally measures only one combination. Nondegenerate baselines or sufficiently different frequency responses can separate $I$ and $V$ under specified spectral assumptions. Earth’s rotation helps anisotropic reconstruction, but does not change an isotropic overlap function for a rigid Earth-fixed baseline: rotating the entire geometry leaves an isotropic sky invariant. Nearly parallel response columns make a sharp polarization fraction prior-driven even if intensity is well measured Seto and Taruya 2007.

Searches for scalar or vector responses introduce corresponding overlap functions. Calling them “polarizations” is only part of the physical model: a theory must also specify generation, propagation speed and dispersion, correlations among modes, and detector coupling. A phenomenological response constraint should not be advertised as a bound on every theory carrying the same label Callister et al. 2017.

Gaussianity and duty cycle

A Gaussian background is characterized by its two-point function. That approximation is expected when many independent unresolved signals overlap in each time–frequency cell. A sparse population instead produces intermittent or “popcorn” data with nonzero higher connected moments.

An elementary mixture for one segment is

$$p(d\mid\xi,\Lambda) = (1-\xi)p_N(d) +\xi\,p_S(d\mid\Lambda),$$

Here $p_S$ is the likelihood of signal plus noise, marginalized over source parameters, and $\xi$ is the probability that a segment contains signal. For Poisson occupancy $\lambda$, $\xi=1-e^{-\lambda}$, but the occupied component must include multiple sources when $\lambda$ is not small. Many independent weak contributions can approach a Gaussian if no source dominates the variance. At small duty cycle, segment likelihoods, higher moments, or coherent population methods retain information discarded by averaging Drasco and Flanagan 2003, §§II–III.

Non-Gaussianity is not automatically astrophysical. Glitches, wandering lines, nonstationary spectra, and environmental transients also create heavy tails. A valid statistic must compare astrophysical coherence across detectors with those noise alternatives.

Identifiability and calibrated tests

For parameters $\boldsymbol a=(I,V,\ldots)$ and baseline-frequency data $\boldsymbol C$, the linear response is

$$\langle\boldsymbol C\rangle = \boldsymbol R\boldsymbol a.$$

The singular values of the noise-weighted $\boldsymbol R$ reveal which combinations are measured. Quoting separate limits after assuming every other component is zero gives conditional single-component bounds; a simultaneous analysis generally broadens them.

For non-Gaussian searches, the null distribution must include the observed noise tails and all selection. Useful checks include:

• inject pure $I$, pure $V$, and mixed signals and recover their covariance;
• reverse the helicity convention and verify that both $V$ and its response coefficient change sign, leaving predicted data invariant;
• exchange detector labels and verify $C_{JI}=C_{IJ}^*$; this conjugation is not a parity reversal and does not itself change the physical helicity of the sky;
• inject tensor, scalar, and vector responses and measure leakage among them;
• inject intermittent populations across duty cycle and compare Gaussian and mixture likelihoods;
• run time shifts, detector scrambles, environmental injections, and single-detector glitch populations;
• refine segment duration, frequency resolution, sky basis, and response singular-value cutoff until inference stabilizes.

The software record should pin response tensors, polarization basis, Fourier and baseline-order conventions, priors, segment list, calibration, environmental vetoes, random seeds, and convergence tolerances.

Frontier evidence through 2026-08-18

Evidence stream	What is supported	What is not established
O4a isotropic tensor search	No significant isotropic tensor background is reported; the release gives spectral-model-dependent upper limits LVK 2025.	An upper limit is not evidence for nonzero $I$, and it does not by itself measure $V$ or alternative responses.
O4a directional search	No significant persistent directional signal is reported, with morphology- and sky-dependent limits LVK 2025, directional search.	Null tensor maps do not universally exclude polarized or intermittent populations outside the analyzed response model.
Polarization-response methodology	Network overlap functions make simultaneous tensor/vector/scalar or $I/V$ inference possible in identifiable response directions Callister et al. 2017.	Feasibility and upper limits are not detections; degeneracies can remain prior dominated.
Non-Gaussian methodology	Mixture and higher-order statistics can outperform Gaussian averaging for sufficiently sparse populations Drasco and Flanagan 2003.	Current ground-based data through the cutoff do not establish an astrophysical polarized or non-Gaussian stochastic detection.

This matrix must be updated when an official polarization or non-Gaussian release, new observing data, revised calibration, or materially improved correlated-noise model appears.

Exercises

1. Two baseline measurements satisfy $C_1=I+V$ and $C_2=I-V$. Solve for $I$ and $V$. What happens if both baselines instead satisfy $C_b=I+V$?

Solution

For the first response matrix,

$$I=\frac{C_1+C_2}{2}, \qquad V=\frac{C_1-C_2}{2}.$$

If both rows are identical, the response matrix has rank one. Only $I+V$ is measured; $I-V$ lies in the null space. No amount of lower noise in those same two responses separates the components.

2. A time series contains a signal in a fraction $\xi$ of segments. Conditional signal segments have zero mean and variance $\sigma_s^2$, while noise has variance $\sigma_n^2$. Find the total variance.

Solution

Both mixture components have zero mean, so the law of total variance gives

$$\operatorname{Var}(d) = (1-\xi)\sigma_n^2 +\xi(\sigma_n^2+\sigma_s^2) = \sigma_n^2+\xi\sigma_s^2.$$

Different pairs $(\xi,\sigma_s^2)$ can produce the same variance. Higher-order or segment-level information is needed to distinguish rare loud signals from frequent weak ones.

References

• Callister, T., Biscoveanu, A. S., Christensen, N., Isi, M., Matas, A., Minazzoli, O., Regimbau, T., Sakellariadou, M., Tasson, J., and Thrane, E. (2017). “Polarization-Based Tests of Gravity with the Stochastic Gravitational-Wave Background.” Physical Review X 7, 041058. DOI.
• Drasco, S., and Flanagan, É. É. (2003). “Detection methods for non-Gaussian gravitational wave stochastic backgrounds.” Physical Review D 67, 082003. DOI.
• LIGO Scientific Collaboration, Virgo Collaboration, and KAGRA Collaboration (2025). “Directional Search for Persistent Gravitational Waves: Results from the First Part of LIGO–Virgo–KAGRA’s Fourth Observing Run.” arXiv:2510.17487.
• LIGO Scientific Collaboration, Virgo Collaboration, and KAGRA Collaboration (2025). “Upper Limits on the Isotropic Gravitational-Wave Background from the first part of LIGO, Virgo, and KAGRA’s fourth Observing Run.” arXiv:2508.20721.
• Seto, N., and Taruya, A. (2007). “Measuring a Parity-Violation Signature in the Early Universe via Ground-Based Laser Interferometers.” Physical Review Letters 99, 121101. DOI.