Theory, Observable, and Measurement

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 test of gravity does not compare a field equation directly with a number printed by an instrument. It compares a forward model—gravity plus sources, propagation, apparatus, calibration, and data reduction—with recorded data. Keeping those layers distinct is the central discipline of this volume: a discrepancy at the data level identifies a failed composite model before it identifies which layer failed.

Required background. The Einstein Equation supplies the Einstein equations relating geometry to stress-energy and used to construct the null prediction. Probability and Inference for Gravity supplies conditional probability, likelihoods, marginalization, and model comparison.

Helpful background. Weak Metric Perturbations and Gauge Freedom develops the weak-field variables used in many examples. The Tests of Gravity overview locates this forward-model logic among the experimental regimes of the volume.

The forward map from dynamics to data

Let $H$ denote a gravitational hypothesis, $\vartheta$ its physical parameters, $\eta$ source and environmental parameters, and $\lambda$ instrument and analysis parameters. A useful schematic chain is

(H,\vartheta) \longrightarrow g_{\mu\nu},\,\text{fields} \longrightarrow O_{\rm phys}(\vartheta,\eta) \longrightarrow s(\vartheta,\eta,\lambda) \longrightarrow d.

The first arrow solves an initial- or boundary-value problem. The second constructs an operational observable: a proper-time ratio, an orbital phase, a pulse arrival time, an image visibility, or a detector strain. The third models how an apparatus converts that observable into a signal. The last adds noise, selection, digitization, and data-processing effects.

An observable must be tied to a realizable protocol. Coordinate components such as $g_{00}$ are not measurements by themselves; a clock comparison measures accumulated proper-time or frequency ratios along specified worldlines. Likewise, a gravitational-wave detector records calibrated differential arm response, not a curvature scalar in isolation. Diffeomorphism invariance therefore enters twice: the field calculation must be consistent, and the final prediction must be expressible in terms of the actual protocol Will 2014, §2.

For data $d$ , first specify the full sampling model $p(d\mid H,\vartheta,\eta,\lambda)$ . A Bayesian treatment with stated nuisance priors then forms the marginalized likelihood

p(d\mid H,\vartheta) =\int p(d\mid H,\vartheta,\eta,\lambda)\, \pi(\eta,\lambda\mid H,\vartheta)\, d\eta\,d\lambda .

Marginalizing nuisance parameters does not make them unimportant. It propagates their stated uncertainty and correlations into the gravity conclusion. A frequentist analysis may instead profile over them and calibrate the resulting procedure; the integral is not the definition of every possible likelihood analysis. An omitted nuisance process is absent from either model and can reappear as an apparent deviation.

Four statements that must not be conflated

As a concrete example, imagine comparing two identical clocks held at fixed heights in a weak, stationary field. Restore $c$ and write $g_{00}=-(1+2\Phi/c^2)$ , with Newtonian potential $\Phi$ increasing upward. For a clock at rest,

d\tau=\sqrt{-g_{00}}\,dt \simeq\left(1+\frac{\Phi}{c^2}\right)dt.

An ideal comparison referred to the same coordinate-time interval gives the fractional rate difference $y=(\Phi_B-\Phi_A)/c^2\simeq g\Delta h/c^2$ . For the illustrative inputs $g=9.81\,\mathrm{m\,s^{-2}}$ , $\Delta h=1\,\mathrm m$ , and $c=299\,792\,458\,\mathrm{m\,s^{-1}}$ , this is $1.09\times10^{-16}$ : the higher clock accumulates more proper time. That is the physical prediction. Actual recorded frequency ratios also require a transfer link and corrections for motion, tides, and clock offsets. If a model adds a redshift coefficient $\beta$ , a toy measurement is

y_{\rm meas}=(1+\beta)\frac{g\Delta h}{c^2}+b+n,

where $b$ is an unknown offset and $n$ is noise. At one fixed height difference, changing $\beta$ can be canceled exactly by changing $b$ . Varying $\Delta h$ while controlling the offset separates the slope from the intercept. This is why a convincing experiment varies the response, rather than merely collecting more samples in an unchanged configuration. The complete clock protocol is developed in Gravitational Redshift Experiments.

A theoretical prediction is conditional on equations, solution branch, initial or boundary data, and an approximation regime. General relativity plus a vacuum Kerr metric, for example, is a narrower hypothesis than general relativity with an uncertain matter environment.

A physical observable is the invariant or operational quantity predicted under those conditions. It may depend on source structure: compact-body sensitivities can enter a binary observable even when the underlying coupling is universal at the field level.

A measured quantity is the estimator produced from finite, calibrated data. Its sampling distribution depends on noise, selection, and the analysis pipeline. A posterior median is neither raw data nor a theory prediction.

A physical conclusion states what the comparison excludes or supports, with a declared confidence or probability convention and a validity domain. A tight interval around a phenomenological coefficient is not automatically a tight interval on every theory that can generate that coefficient.

This separation is especially important when the map is many-to-one. Different actions can predict the same leading post-Newtonian coefficient, while one theory can predict different effective coefficients in screened and unscreened environments. The data constrain the image of a theory in observable space; interpreting that image requires the theory-to-observable map.

Identifiability and degeneracy

Near a reference model, write the predicted data vector as

\delta \boldsymbol{s} =J_{\vartheta}\,\delta\boldsymbol{\vartheta} +J_{\eta}\,\delta\boldsymbol{\eta} +J_{\lambda}\,\delta\boldsymbol{\lambda}.

If a column of $J_{\vartheta}$ lies in the span of nuisance columns, that gravity direction is locally degenerate with source or instrument physics. More data of the same type may reduce random error without resolving the degeneracy. A second observable helps only if its response changes the rank or conditioning of the combined Jacobian.

This linear test is diagnostic, not definitive. Nonlinear branches, parameter boundaries, multimodal posteriors, and selection effects can defeat a Fisher-matrix argument. Simulation-based calibration, injected-signal recovery, posterior-predictive checks, and analysis with alternative nuisance models probe the full pipeline more honestly.

Independent checks and failure boundaries

A trustworthy test closes several loops. The forward model should recover the Newtonian or other controlled limit; dimensions and signs should survive a convention change; synthetic data generated under the null should yield calibrated coverage; and an independent implementation should reproduce a consequential observable. Real-data residuals should also be tested against known environmental and instrumental channels.

None of these checks proves the gravitational hypothesis. They establish that a stated comparison means what it claims to mean. If the source model is inadequate, calibration covariance is omitted, or the approximation leaves its domain, the appropriate conclusion is that the composite prediction failed—not that a new gravitational interaction was discovered.

Check your understanding

Suppose an orbital analysis finds a residual precession $\Delta\dot\omega$ . Explain why the residual alone cannot be interpreted as a non-GR parameter, and identify a minimal forward model for the test.

Solution

The residual is defined relative to a fitted model, so it can absorb errors in the mass distribution, reference-frame realization, time standard, propagation correction, and gravity calculation. A minimal forward model specifies the gravitational equations and solution regime; maps masses, spins, multipoles, and environmental forces to an orbit; maps that orbit to timed observables; and includes the instrument, calibration, noise, and selection likelihood. Therefore $\Delta\dot\omega$ can constrain a declared deformation only after nuisance parameters and their covariance are fitted or marginalized, because otherwise the same residual has several physical explanations. The resulting bound remains conditional on the deformation-to-theory map and the approximation domain.

References

Will, C. M. (2014), “The confrontation between general relativity and experiment,” Living Reviews in Relativity 17, 4. Will 2014, §§2–3.