From Constraint to Physical Conclusion

Scrollable formula. Use Left and Right arrow keys to reveal hidden content.

Scrollable table. Use Left and Right arrow keys to reveal additional columns.

Scrollable figure. Use Left and Right arrow keys to inspect hidden content.

A fitted interval on a deviation coefficient is the beginning of an interpretation, not its end. To reach a physical conclusion one must identify what the data constrained, show how a specified theory maps into that quantity in the probed regime, propagate shared nuisance and selection effects, and verify that the theory is stable and predictive throughout the inferred region. Skipping any arrow can turn a precise statistical statement into an overbroad claim about gravity.

Required background. The Einstein Equation supplies the Einstein equations defining the GR reference model. Probability and Inference for Gravity supplies likelihood, posterior, confidence, evidence, and model-checking concepts.

Helpful background. Weak Metric Perturbations and Gauge Freedom supplies the weak-field variables often used for constraints. The Tests of Gravity overview places the inference among complementary regimes.

The chain of conditional statements

Suppose an analysis introduces a phenomenological coefficient $\delta$ with GR value $\delta=0$ . The data produce a likelihood

\mathcal L(\delta,\eta)=p(d\mid\delta,\eta,M),

where $M$ contains the waveform, timing, lens, source, propagation, instrument, and noise assumptions relevant to the experiment. After profiling or marginalizing nuisance parameters $\eta$ , one may report a confidence or credible interval on $\delta$ . The honest statement is conditional: within model $M$ , its calibration and selections, and the chosen interval construction, the data restrict $\delta$ to this region.

For a theory with parameters $\alpha$ , interpretation requires a controlled map

\delta=F(\alpha;\mathcal R,\mathcal E,\mathcal B),

where $\mathcal R$ denotes the approximation regime, $\mathcal E$ the environment and source state, and $\mathcal B$ the boundary or cosmological conditions. The map may be nonlinear, non-injective, branch-dependent, or undefined where the effective description fails. A one-dimensional limit on $\delta$ does not define a universal limit on $\alpha$ until these dependencies are fixed.

The map can also be many-to-one in the other direction: several theories may share the same leading coefficient. Then agreement at that order constrains an equivalence class, while higher-order or differently scaled observables are needed to distinguish its members. This is why the regime analysis and parameter-measure analysis are parts of the physical result rather than statistical appendices.

Results and limitations through 2026-08-19

The evidence corpus cited in this volume through 2026-08-19 supports a mature discipline of constructing stringent, mutually complementary gravity tests. It does not support a model-independent theorem that every viable departure from GR is uniformly small in every regime.

Inference layer	Evidence or result at the cutoff	Limitation or scope boundary
Experiment-level coefficient	Many calibrated analyses find deformation parameters compatible with their GR values	The result is conditional on release, calibration, selection, nuisance model, and the fitted template
Framework-level consistency	Solar-System, pulsar, wave, and cosmological frameworks constrain different combinations and scales	A coefficient name need not denote the same theory quantity across environments or approximation orders
Complete-theory interpretation	Specified theories can be mapped to multiple observables and jointly constrained	Stability, screening, strong coupling, branch selection, and source physics can remove or reshape the mapped domain
Global statement about gravity	The named analyses constrain particular models and regimes; a global theory ranking requires its own joint analysis	Weakly coupled, screened, degenerate, or untested sectors remain open; this page does not survey every viable alternative

The conservative interpretation is that GR remains a successful reference model across an exceptionally broad but finite set of controlled tests. A stronger competing position claims that a recurring residual or coefficient shift is evidence for new gravity. That position becomes persuasive only when the effect survives independent calibration and nuisance models, appears with the correlated scaling predicted by a viable theory, and improves predictions outside the data that selected it. Negative evidence includes the disappearance of an effect under a revised calibration, failure in an independent channel, or a theory map that crosses an instability or cutoff before reaching the fitted coefficient.

Translating an interval without losing information

The likelihood is a function of the parameters, not a probability density over them. Consequently its transformation is simple substitution:

\mathcal L_\alpha(\alpha)=\mathcal L_\delta(F(\alpha)).

There is no parameter Jacobian in this equation. A Jacobian is needed when transforming a probability density, such as an existing posterior on a one-to-one branch:

p_\alpha(\alpha\mid d) =p_\delta(F(\alpha)\mid d)\left|\frac{dF}{d\alpha}\right|.

That transformed posterior inherits the original prior; it need not equal the posterior obtained by assigning a new theory prior $\pi_\alpha$ and using $p_\alpha\propto\mathcal L_\delta(F(\alpha))\pi_\alpha$ . This distinction follows directly from probability conservation and the definition of a likelihood Particle Data Group 2024, §40.1, PDF. If several $\alpha$ values map to one $\delta$ , the original one-dimensional posterior does not determine their relative weights: a theory prior must supply them. If $F$ depends on source parameters already inferred from the same data, use a joint likelihood; multiplying a published $\delta$ posterior by an independent source posterior generally duplicates priors and discards covariance.

Population and selection effects demand the same care. A theory can change which sources enter a catalog, so the detection efficiency may depend on $\alpha$ . The normalized hierarchical likelihood must then include that dependence Mandel, Farr, and Gair 2019, §§2–4. Combining experiments also requires identifying shared ephemerides, calibration standards, astrophysical simulations, cosmological assumptions, and theory priors. Treating those inputs as independent can distort the combined uncertainty or apparent tension; same-sign shared calibration errors, for example, cannot be averaged down like independent noise.

Bayesian and frequentist summaries should retain their meanings. A credible interval is probability under a stated prior and model; a confidence construction is calibrated under repeated samples from a stated data-generating family. A Bayes factor compares complete prior-predictive models. None is, by itself, the probability that an unspecified class called “modified gravity” is true.

Independent checks that could change the conclusion

An adversarial interpretation test starts from a candidate theory rather than the reported upper limit. First solve for the background and relevant source in each regime; then compute at least two observables with distinct nuisance directions; finally pass simulated signals through the released analysis and selection pipeline. The null, theory injection, and a deliberately misspecified source or calibration model should be recovered or rejected at their calibrated rates.

A useful counterexample to universal translation is an environmentally screened scalar: a small Solar-System charge can coexist with a different charge in an unscreened galaxy, while a compact star may develop another strong-field branch. Conversely, an unscreened weak-field coefficient may predict a correlated pulsar or wave effect that is absent. The discriminating calculation is the theory-wide map, not an overlay of independently labeled exclusion curves Burrage and Sakstein 2018, §§2–5.

The principal unresolved question is therefore model-dependent: which theoretically consistent departures retain observable effects after all known regimes and systematics are imposed? It remains unknown whether future data will expose such a sector or continue to shrink it toward limits that are experimentally inaccessible.

What could change the conclusion. The interpretation could change after a superseding major data or calibration release, a corrected analysis that changes a consequential coefficient, or a viable theory calculation that supplies a new cross-regime mapping and passes the independent injection-and-prediction check described above.

Check your understanding

An experiment reports $|\delta|<10^{-3}$ , while a theory predicts $\delta=\alpha^2 q(\rho)$ with an environment-dependent charge $q$ . Why does the published interval not imply $|\alpha|<10^{-3/2}$ ?

Solution

The algebraic square root is valid only if $q=1$ , the branch is stable, and the experiment’s template matches the theory over its data domain. Here $q$ depends on density and possibly geometry or boundary conditions; it may be small, change sign, or be correlated with source parameters. Therefore one must solve for $q(\rho)$ in the experimental environment, insert the resulting signal into the likelihood, and retain all admissible branches with the declared parameter measure. A second environment provides an independent check because the theory predicts how $q$ changes rather than allowing it to float freely.

References

Burrage, C., and Sakstein, J. (2018), “Tests of chameleon gravity,” Living Reviews in Relativity 21, 1. Burrage and Sakstein 2018, §§2–5.
Mandel, I., Farr, W. M., and Gair, J. R. (2019), “Extracting distribution parameters from multiple uncertain observations with selection biases,” Monthly Notices of the Royal Astronomical Society 486, 1086–1093. Mandel, Farr, and Gair 2019, §§2–4.
Particle Data Group (2024), “Statistics,” in Review of Particle Physics, Physical Review D 110, 030001. PDF.