Bayesian Forecasts on Cosmic Superstring Searches with LISA

Abstract: Cosmic superstrings are well-motivated early-Universe sources of stochastic gravitational waves, with phenomenology controlled by the string tension $Gμ$ and the intercommutation probability $P$. We study whether LISA can reconstruct these parameters and distinguish different superstring signal models in the presence of instrumental noise and astrophysical foregrounds. We consider two phenomenological models. In Model I, reduced intercommutation acts only as an amplitude enhancement of a cusp-dominated spectrum, $$Ω<em>{\rm SS}^{\rm I}=P^{-β}Ω</em>{\rm cusp}$$. In Model II, the signal is a cusp--kink mixture, $$Ω<em>{\rm SS}^{\rm II} =P^{-β}[p_cΩ</em>{\rm cusp}+(1-p_c)Ω_{\rm kink}]$$, so that $P$ controls both amplitude and spectral shape. Using simulated LISA data, we perform Bayesian inference with noise uncertainties, unresolved extragalactic compact-binary backgrounds, and a flexible Galactic double-white-dwarf foreground. We compare the models using Bayesian evidences and map the $(Gμ,P)$ posterior geometry with marginalized widths, correlations, covariance anisotropy, principal eigenvalues, posterior area, and bias. We find that reduced intercommutation generically produces strong $Gμ$--$P$ correlations, because the data often constrain an amplitude-like combination. However, when the cusp--kink spectral difference lies in the LISA band and the signal is sufficiently loud, Model II can be favored and the posterior can retain genuine shape information. As an optimistic foreground scenario, we also compute the Bayes factor using a reduced tanh Galactic foreground template, finding improved model discrimination when the foreground shape is constrained.

• The paper develops a Bayesian forecast showing that LISA can detect cosmic-superstring gravitational-wave backgrounds, but strong amplitude-driven correlations often limit separate recovery of string tension $G\mu$ and intercommutation probability $P$.
• The analysis finds that LISA can distinguish amplitude-only and cusp–kink spectral models mainly for small $P$, where kink emission changes the in-band spectrum; at large $P$, detectability does not imply model distinguishability.
• The study shows that flexible Galactic white-dwarf foreground models reduce discrimination power and emphasizes evaluating posterior precision, parameter correlation, and reconstruction bias together rather than relying on signal-to-noise or Bayes factors alone.

This paper presents a Bayesian forecast study assessing whether the Laser Interferometer Space Antenna (LISA) can not only detect a stochastic gravitational-wave background (SGWB) from cosmic superstrings, but also reconstruct its microscopic parameters and discriminate between competing phenomenological descriptions (2608.19406). The analysis is framed around two questions: whether LISA can separately infer the string tension $G\mu$ and the intercommutation probability $P$, and whether it can distinguish a purely amplitude-level effect of reduced intercommutation from one that also modifies the spectral shape through cusp–kink emission physics.

Motivation and model construction

The motivation stems from brane-inflation scenarios in which cosmic superstrings form with reconnection probabilities $P$ potentially far below unity, unlike field-theory strings for which $P \simeq 1$. Reduced intercommutation densifies the long-string network and enhances the loop population, boosting the SGWB amplitude. The paper deliberately abstracts away from realistic multi-species networks (F-strings, D-strings, $(p,q)$ bound states, junctions), adopting instead a reduced two-parameter description in which $\beta=1$ is fixed and the signal parameters are $(\log_{10}G\mu,\log_{10}P)$.

Two signal models are compared:

• Model I treats reduced intercommutation as an amplitude rescaling of a cusp-dominated spectrum, $$\Omega_{\rm SS}^{\rm I}=P^{-1}\Omega_{\rm cusp}(f;G\mu)$$. This is the same prescription used by NANOGrav when interpreting its 15-year common-spectrum process as a superstring signal. Because $P$ enters only as a normalization, the data constrain primarily the combination $(G\mu)^q P^{-1}$, so strong $P$0–$P$1 correlation is expected by construction.
• Model II extends this to a cusp–kink mixture, $P$2, motivated by suppression of cusps through motion in compact extra dimensions and enhancement of kink-like emission via junctions and small-scale structure. The cusp fraction uses a smooth transition ansatz $P$3 with $P$4 and $P$5.

A technically important point established early: the mixture cannot be reduced to a single effective harmonic index $P$6. The local index $P$7 interpolates between $P$8 and $P$9 as a function of harmonic number, and the full spectrum depends on both the harmonic weights and the loop density evaluated at the harmonic-dependent length $P$0. Consequently, shape sensitivity to $P$1 in Model II is localized near $P$2; outside this transition region, Model II collapses to either a cusp- or kink-dominated rescaled spectrum and $P$3 again behaves mainly as an amplitude parameter.

The loop spectra themselves follow the standard framework: loops of initial length $P$4 with $P$5 shrink under gravitational backreaction with $P$6, emitting into harmonics with $P$7, using $P$8 (cusps) and $P$9 (kinks).

Inference setup

Mock LISA data are generated in the noise-orthogonal TDI $P \simeq 1$0 basis, with the $P \simeq 1$1 channel treated as a null channel to help constrain instrumental noise. The likelihood assumes zero-mean complex Gaussian Fourier coefficients per segment and frequency bin. The full parameter vector comprises ten parameters: two instrumental-noise amplitudes ($P \simeq 1$2, $P \simeq 1$3), four parameters of a flexible Galactic double-white-dwarf (DWD) foreground template allowing both amplitude and shape freedom ($P \simeq 1$4), two extragalactic compact-binary foreground parameters (a power-law normalization $P \simeq 1$5 with spectral index $P \simeq 1$6), and the two signal parameters. The priors on the signal span $P \simeq 1$7 and $P \simeq 1$8. Posterior exploration uses dynesty nested sampling interfaced through Bilby.

The flexible DWD template is explicitly chosen as a conservative baseline: its four free shape/normalization parameters allow it to absorb broad mHz-band curvature, thereby weakening apparent reconstruction power. A complementary optimistic scenario replaces it with a reduced tanh template whose shape is fixed and only the normalization varies.

Posterior diagnostics

Beyond standard Bayesian evidences, the paper develops a systematic set of covariance-based diagnostics in the $P \simeq 1$9 plane: the marginalized widths $(p,q)$0, Pearson correlation coefficient $(p,q)$1, covariance anisotropy $(p,q)$2, best-constrained principal width $(p,q)$3, covariance area $(p,q)$4, and reconstruction bias (displacement between injected and recovered posterior location). These are combined into seven diagnostic patterns distinguishing, among others, compact correlated reconstruction, broad correlated degeneracy, axis-aligned loss of identifiability, balanced two-parameter reconstruction, prior-dominated regimes, and precise-but-biased recovery.

A central methodological claim of the paper is that large $(p,q)$5 does not imply poor reconstruction—an elongated but well-localized posterior still yields an accurate measurement of one parameter combination—and that precision (small $(p,q)$6) must be assessed independently of accuracy (small bias). This distinction matters because compact posteriors displaced from the injected value can arise from foreground degeneracy, non-Gaussian structure, or prior-boundary effects.

Results: model comparison and reconstruction

The Bayes-factor analysis generates mock data with Model II across a grid in the $(p,q)$7 plane and analyzes each realization with both models. Three findings structure the result:

First, at large intercommutation probability $(p,q)$8, Model II reduces exactly to the Model-I form ($(p,q)$9) and the Bayes factor satisfies $\beta=1$0 even when the signal is detectable. This yields a substantive negative result stated plainly in the paper: detectability does not imply distinguishability. A detected effectively cusp-dominated background would leave LISA without leverage on the underlying superstring microphysics.

Second, at small $\beta=1$1, the injected signal becomes kink dominated while Model I remains cusp dominated. The discriminating spectral difference scales as $\beta=1$2, so evidence favoring Model II concentrates where reduced intercommutation simultaneously amplifies the signal and makes the shape difference observable. Strongly negative $\beta=1$3 appears precisely in this regime—the Bayes factor traces detectable shape mismatch rather than total SNR.

Third, within-model reconstruction maps show that strong $\beta=1$4–$\beta=1$5 correlations persist over most of the detectable parameter space in both models, as expected from the shared $\beta=1$6-dependent amplitude response. In Model I, the data typically constrain only the trade-off direction $\beta=1$7, producing thin correlated ridges whose utility depends on whether $\beta=1$8 is small (compact correlated measurement) or large (broad degeneracy). In Model II, regions exist where the cusp–kink morphology changes measurably inside the band; there the widths shrink and $\beta=1$9 increases, marking the most favorable conditions for genuine two-parameter reconstruction of $(\log_{10}G\mu,\log_{10}P)$0.

Foreground assumptions and their impact

The comparison between the conservative flexible-DWD baseline and the reduced tanh foreground is the clearest illustration of how modeling choices control scientific reach. With the tanh template, the region in which Model II is favored expands toward larger $(\log_{10}G\mu,\log_{10}P)$1, because the fixed-shape foreground can no longer absorb broad spectral curvature associated with partially cusp-dominated mixtures. This result implies that the achievable model discrimination depends as much on prior knowledge of the Galactic foreground as on the intrinsic cosmological signal strength. The paper notes that only the Bayes-factor map was recomputed for the tanh case; the corresponding full covariance and bias diagnostics were not, leaving that extension open.

Limitations and open questions

Several limitations are acknowledged or evident. The two-parameter phenomenology fixes $(\log_{10}G\mu,\log_{10}P)$2 and $(\log_{10}G\mu,\log_{10}P)$3; alternative network-density scalings (e.g., those of Avgoustidis and Shellard) and different transition parameters would shift the location and strength of shape information. Realistic multi-tension networks with junctions are outside scope. The analysis assumes Gaussianity and stationarity of the total background, and the bias maps indicate that residual foreground–signal degeneracy can displace posteriors even where they appear precise. Whether the tanh-template improvement in model discrimination carries through to unbiased parameter reconstruction remains untested. Finally, the study is a mock-data forecast on a grid of injections; it does not address hierarchical inference over the full ten-dimensional parameter space jointly with foreground calibration from resolved binaries.

Conclusion

This work provides a concrete inference-level map of where LISA can move beyond detecting a cosmic-superstring SGWB toward reconstructing $(\log_{10}G\mu,\log_{10}P)$4 and $(\log_{10}G\mu,\log_{10}P)$5 and discriminating between amplitude-only and shape-extended superstring phenomenologies. Its main conclusions are that model distinguishability requires the cusp–kink difference to lie in-band and be amplified by small $(\log_{10}G\mu,\log_{10}P)$6; that $(\log_{10}G\mu,\log_{10}P)$7–$(\log_{10}G\mu,\log_{10}P)$8 correlations are generic because both parameters affect amplitude; and that reconstruction quality must be judged jointly by localization, correlation, and bias. The framework clarifies that a future LISA detection would be scientifically ambiguous unless accompanied by measurable spectral-shape information—a condition controlled substantially by how well the Galactic foreground is characterized.