Dense-Matter Multimessenger Inference

Dense-matter inference is an inverse problem with a shared physical object—the equation of state—but no shared raw observable. Radio arrival times constrain masses, X-ray photons constrain a surface-emission model, gravitational-wave strain constrains a tidal phase, and nuclear experiments constrain finite systems or transient matter. This chapter shows how to preserve those distinctions while combining their information. Its central rule is simple: carry each native likelihood, calibration, selection, and model domain far enough that agreement or discrepancy can be traced to a physical assumption. Dated evidence is assessed through 2026-08-18.

Helpful background. Stress–energy and conservation identifies the energy density and pressure being inferred; the Einstein equation maps them to stellar structure; relativistic matter and compact objects supplies the EOS, equilibrium, oscillation, and tidal framework; and gravitational radiation and waveforms connects stellar response to inspiral waveforms and detector measurements.

Enter the density–likelihood ladder

The forward direction is easier than the inverse one:

$$p(\varepsilon) \longrightarrow \{M(R),\Lambda(M),M_{\max},\text{composition-dependent response}\} \longrightarrow \text{timing, counts, strain, and laboratory observables}.$$

Inference reverses this map in the presence of noise, incomplete selections, nuisance parameters, and representations that can exclude possible EOS behavior. The data generally constrain overlapping density intervals rather than independent scalar properties. A heavy pulsar requires an EOS that supports its mass at the measured spin; a radius likelihood constrains an integrated stellar profile; a tidal likelihood is strongly mass weighted; and near-saturation nuclear information controls the low-density anchor and its covariance. Özel and Freire 2016, §§ 2–5 provides the observational mass–radius setting, while Kumar et al. 2024, §§ 2.3–2.7 maps the distinct nuclear and astrophysical domains.

For a first encounter with inference, keep three objects separate. The likelihood answers how probable the observed data are at a proposed mass, radius, or EOS. The prior states the allowed models and their relative weight before those data are used. Their normalized product is the posterior, which therefore changes when either the measurement model or the prior changes. More messengers can reduce uncertainty only when their likelihoods add information that has not already been used.

The chapter uses $c=\hbar=k_{\mathrm B}=1$ and keeps $G$ explicit. Mass is a gravitational mass unless stated otherwise; radius is the circumferential radius used by the stellar model; $\Lambda$ is dimensionless and tied to a declared Love-number convention. Observational intervals state whether they are Bayesian credible or frequentist confidence intervals. Source-frame and detector-frame masses are not mixed.

Six inference problems

The pages proceed in the order needed to build and then challenge a joint EOS result.

1. Mass Measurements and Population Structure derives radio and optical mass likelihoods, then shows how population families and selection alter an apparent upper cutoff. It separates a largest measured mass, a population edge, and the cold nonrotating maximum mass.
2. X-Ray Pulse-Profile Radius Inference maps local atmosphere emission through light bending, rotation, instrument response, and a Poisson count likelihood. Its central degeneracies are compactness, hot-region geometry, background, atmosphere, distance, and calibration.
3. Gravitational-Wave Tidal Inference follows $k_2$ and $\Lambda(M)$ into the binary combination $\widetilde\Lambda$ and the calibrated strain likelihood. Waveform, spin, frequency-band, event-prior, and population-selection effects remain visible.
4. Nuclear Experiments and Theory Priors organizes finite nuclei, PREX-II/CREX, chiral effective field theory, and heavy-ion observables by density and composition. Correlated theory error prevents a model ensemble from being treated as independent data.
5. Hierarchical Equation-of-State Reconstruction compares piecewise-polytropic, spectral, sound-speed, Gaussian-process, and explicit-transition representations. It imposes thermodynamic stability and causality while retaining branch and parameterization uncertainty.
6. Cross-Messenger Consistency Tests and Systematics closes the loop with prior prediction, held-out posterior prediction, leave-one-messenger-out tests, calibration alternatives, and synthetic recovery. “Tension” is reserved for discrepancies that survive those controls.

This order is pedagogical rather than chronological. A researcher may enter at one messenger, but a joint numerical result should be read through the reconstruction and consistency pages before being interpreted as a material property.

The common statistical spine

Each page uses the same hierarchy of questions:

1. What is the native data product and exact version?
2. Which calibration, response, or timing model maps instrument output to an observable?
3. Which objects, times, frequencies, energies, or kinematics were selected, and how does detection enter the normalization?
4. What forward model connects the EOS and object-specific parameters to the data?
5. Which priors and population assumptions were used, including interim priors embedded in posterior samples?
6. Which systematic alternatives are calibrated, and which are merely plausible?
7. Can a clean, finite reproduction recover a benchmark and a synthetic injection?

Bayesian factorization is useful only when conditional independence is justified. Shared radio masses, distances, EOS training ensembles, waveform products, or nuclear calibration data can couple nominally different messengers. Published posteriors are not likelihoods unless their interim priors are removed. Posterior-predictive checks ask whether the fitted model can reproduce native observables, not whether two derived one-dimensional intervals overlap; Gelman et al. 2013, ch. 6 develops that distinction.

The chapter does not prescribe one EOS parameterization. Instead, it demands prior-predictive comparison in pressure, sound speed, mass–radius, tidal response, maximum mass, and transition structure. A conclusion robust in spectral coefficients but absent under a sound-speed or Gaussian-process representation is a representation-conditioned result.

Suggested reading paths

For a first graduate route, read masses, X-ray radii, tidal inference, and then hierarchical reconstruction. The nuclear page explains where near-saturation information enters; the consistency page turns the sequence into a review checklist.

For observational work, begin with the native messenger page and reproduce its likelihood boundary before joining it to another channel. For nuclear theory, move from the density-aware evidence map to the EOS representation table and inspect how correlated truncation uncertainty is transported. For method development, start with hierarchical reconstruction, then use the cross-messenger page as an adversarial test suite.

This chapter develops the material interpretation and the joint EOS model. Detector calibration and production waveform likelihoods receive fuller treatment in gravitational-wave astronomy; orbital waveform construction belongs to gravitational radiation and two-body dynamics; production algorithms and executable large analyses belong to computational relativity. The instructions for reproducing an analysis do not constitute a newly executed analysis.

Raaijmakers et al. 2021, §§ 2–4 gives a concrete joint analysis in which data groups and EOS representations can be compared. It is an example, not a frozen consensus result: updated catalog, calibration, likelihood, and model releases must be attached to any current numerical claim.

Review the chapter

Suppose a NICER posterior for one pulsar already used a Gaussian radio mass prior from timing data. You are given posterior samples from that NICER run, the radio timing likelihood, and a GW tidal likelihood. How should these be combined without counting the radio data twice, and which two checks most directly test cross-messenger consistency?

Solution

Let the published NICER samples represent

$$p(M,R,\eta\mid \boldsymbol d) \propto \mathcal L_X(d_X\mid M,R,\eta) \mathcal L_R(d_R\mid M) \pi_0(M,R,\eta).$$

To recover an approximation to the X-ray likelihood, divide the sample density by both the documented interim prior $\pi_0$ and the radio likelihood $\mathcal L_R$—provided their support and normalizations are available. The joint EOS analysis then multiplies the reconstructed X-ray likelihood, the radio likelihood once, and the GW likelihood, while sharing the EOS map and marginalizing target-specific nuisances. If the original products cannot support that division, rerun the X-ray likelihood or keep the combined X-ray-plus-radio factor intact; do not multiply the radio term again.

The most direct consistency checks are a held-out posterior prediction—fit radio plus X-ray and predict the native GW tidal data, then rotate the held-out channel—and a leave-one-messenger-out reconstruction of pressure or radius by density. Both should be repeated under calibrated atmosphere, waveform, response, and EOS-parameterization alternatives. Overlap of published radius intervals alone is not such a check.

References

• Gelman, Andrew, John B. Carlin, Hal S. Stern, David B. Dunson, Aki Vehtari, and Donald B. Rubin. Bayesian Data Analysis. 3rd ed. Boca Raton: CRC Press, 2013. DOI.
• Kumar, Rajesh, et al. (MUSES Collaboration). “Theoretical and Experimental Constraints for the Equation of State of Dense and Hot Matter.” Living Reviews in Relativity 27 (2024): 3. DOI.
• Özel, Feryal, and Paulo Freire. “Masses, Radii, and the Equation of State of Neutron Stars.” Annual Review of Astronomy and Astrophysics 54 (2016): 401–440. DOI.
• Raaijmakers, Geert, et al. “Constraints on the Dense Matter Equation of State and Neutron Star Properties from NICER’s Mass–Radius Estimate of PSR J0740+6620 and Multimessenger Observations.” Astrophysical Journal Letters 918 (2021): L29. DOI.