Tests of Gravity

Tests of gravity asks a deceptively difficult question: when an observation disagrees with a prediction, what exactly has been tested? This path is for advanced students and researchers entering gravitational tests from theory, observation, or data analysis. It develops a common language for weak- and strong-field experiments, while keeping theory assumptions, astrophysical modeling, calibration, selection effects, and statistical conclusions visibly separate.

Helpful background. Weak-field gravity supplies metric linearization, gauge control, the Newtonian limit, and leading observables. Probability and statistical inference supplies likelihoods, priors, posterior checks, nuisance uncertainty, and selection effects. Einstein equations supplies the geometric field equation, Bianchi consistency, and stress–energy coupling. These are recommendations rather than barriers: the check below identifies the most useful topic to review.

Turning observations into tests of gravity

By the end of the path, you should be able to:

• formulate a gravity test as a null test or a model comparison with its observables defined;
• use parameterized frameworks without treating their coefficients as complete theories;
• assess likelihoods, priors, calibration, nuisance models, systematics, and selection effects; and
• synthesize laboratory, Solar-System, pulsar, electromagnetic, gravitational-wave, lensing, and cosmological evidence without counting shared information twice.

A useful discipline runs through every phase:

$$\begin{aligned} \text{theory assumptions}&\longrightarrow\text{source and propagation model} \longrightarrow\text{observable}\\ &\longrightarrow\text{data model} \longrightarrow\text{bounded conclusion}. \end{aligned}$$

A discrepancy late in this chain is not automatically evidence for a change in the gravitational field equations. It may instead expose an incomplete source model, a calibration error, a selection effect, or a failed approximation.

Check your preparation

Try these before following the route.

1. In a weak, slowly varying field, which quantities are invariantly measurable, and why is a coordinate component of the metric not by itself an observable?
2. If a likelihood contains a gravity-deviation parameter and a calibration offset, what goes wrong if the offset is fixed without justification?
3. Two analyses use the same catalog and the same calibration model. May their published likelihoods simply be multiplied?

Answers and review

1. Proper-time ratios, relative accelerations, lensing angles, timing residuals, and suitably defined waveform features are examples of observables. A metric component changes under a coordinate transformation, so it acquires operational meaning only after the measurement protocol and gauge-invariant or convention-fixed mapping are specified. If this distinction is not secure, review weak-field observables in the first graduate general relativity path.
2. The offset can mimic or absorb the deviation, biasing the estimate and understating its uncertainty. It should be constrained, calibrated, or marginalized within the data model. Review nuisance parameters and marginalization in probability and inference.
3. Not in general. Shared events, selection functions, waveform approximations, or calibration priors induce dependence. A joint likelihood must represent that covariance or be built from conditionally independent data. Use the same inference preparation if this answer was uncertain.

If all three answers are secure, begin with the logic phase and move briskly. If one is uncertain, read the linked review before the corresponding inference phase. If two or more are not yet available, complete both reviews first; later comparisons assume that foundation.

Route through the evidence

The phases are ordered to prevent a catalog of bounds from replacing an argument.

1. Define the claim. Study the logic of testing gravity and then parameterized frameworks and null tests. Decide whether each question is a null test, parameter estimate, or comparison among complete models.
2. Build the controlled weak-field baseline. Continue through equivalence and free fall, clocks, redshift, and local symmetries, laboratory and short-range gravity, and Solar-System and spacecraft tests.
3. Enter relativistic and astrophysical systems. Work through binary pulsars and relativistic timing, galactic and astrophysical dynamics, gravitational-lensing tests, and electromagnetic strong-field tests. Track environmental and source-model degeneracies explicitly.
4. Use radiative and cosmological probes. Study gravitational-wave tests and cosmological tests of gravity. Record which assumptions transport a parameter from one regime to another.
5. Synthesize rather than stack. Finish the conceptual route with cross-regime evidence, then test prior sensitivity and posterior-predictive residuals with all likelihood and population assumptions stated.

Worked reasoning: one residual is not one theory

Suppose a measured vector $d$ is modeled as

$$d=m_{\mathrm{GR}}(\theta)+\alpha s+q u+n, \qquad n\sim\mathcal N(0,C).$$

Here $\theta$ denotes ordinary source parameters, $s$ is the response associated with a parameterized departure from general relativity, $u$ is a calibration or environmental template, and $q$ is its amplitude. The Gaussian log-likelihood is, up to a constant,

$$-2\log \mathcal L=(d-m_{\mathrm{GR}}-\alpha s-q u)^{\mathsf T} C^{-1}(d-m_{\mathrm{GR}}-\alpha s-q u).$$

If $s$ and $u$ are nearly parallel in the covariance-weighted inner product, the data constrain a combination of $\alpha$ and $q$, not $\alpha$ alone. Fixing $q=0$ can manufacture a sharp nonzero estimate of $\alpha$; allowing a physically justified prior for $q$ can broaden it or move it back toward the null value. A posterior for $\alpha$ therefore tests the declared response model and nuisance treatment. It does not yet identify a unique modified-gravity theory.

Now imagine a second experiment reports a coefficient $\alpha_2$. Combining it with $\alpha$ requires a map from a common theory parameter $\lambda$ to both observables,

$$\alpha=F_1(\lambda;\mathcal A_1), \qquad \alpha_2=F_2(\lambda;\mathcal A_2),$$

where $\mathcal A_i$ denotes regime and modeling assumptions. Without those maps, the two coefficients may merely share a name. With them, a joint analysis must still account for shared data and systematics. This is why cross-regime synthesis is a modeling problem, not a multiplication exercise.

Checkpoints

Checkpoint 1. A pulsar-timing residual is consistent with both dipole radiation and an unmodeled propagation delay. What must be added before calling it a gravity test?

Conceptual hint

Separate the gravity response, environmental propagation, measurement noise, and their independently constrained parameters.

Setup hint

Write separate terms for the putative dipole response and the propagation delay in one timing model, each with its own parameters.

Next-step hint

Ask which frequency or orbital signatures, or which auxiliary measurements of the propagation medium, can make those parameters identifiable.

Solution

Write a generative timing model containing both effects, state which frequencies and orbital dependencies distinguish them, include independent information about the propagation medium, and marginalize over its uncertainty. The conclusion should be conditional on that model and should report whether the gravity parameter remains identifiable.

Checkpoint 2. A parameterized gravitational-wave analysis gives a narrow posterior centered on the general-relativistic value. Name two questions that must be answered before interpreting this as a strong test.

Conceptual hint

Ask both whether the data could reveal a deviation and whether the waveform and population models could hide one.

Setup hint

List the tested deviation parameter, its prior, the analysis’s injection or recovery sensitivity, and the waveform, calibration, and population nuisance models.

Next-step hint

Check whether plausible changes to those models or to event selection would broaden or shift the reported posterior.

Solution

At minimum, check the analysis sensitivity through injections or posterior-predictive tests, and test robustness to waveform error, calibration, event selection, and prior choice. A narrow posterior can otherwise reflect a restrictive model or prior rather than informative data.

Transfer task

Choose one proposed strong-field test and one weak-field test. For each, write a one-page chain from theory assumptions to observable to likelihood. Then identify a single theory parameter that could consistently generate both effects—or explain why no justified common map exists. Your comparison must mark shared data, shared nuisance models, and regime-dependent approximations. A successful transfer does not need a new numerical bound; it needs a defensible statement of what the evidence can and cannot compare.

Capstone: a cross-regime evidence brief

Prepare a compact evidence assessment for one gravitational hypothesis. It should:

• define the hypothesis, null, observables, and applicable regimes;
• derive or document the map from theory parameters to every reported test coefficient;
• specify likelihoods, priors, nuisance parameters, calibration, selection effects, and shared covariance;
• compare at least three genuinely different regimes without double counting; and
• end with a bounded conclusion that separates exclusion, consistency, anomaly, and unresolved modeling.

Use four criteria to judge the result: physical validity of the theory-to-observable map, inferential validity of the likelihood and uncertainty treatment, cross-regime coherence of the combination, and epistemic restraint of the conclusion. Strong work makes every assumption inspectable; exemplary work also shows how the conclusion changes under a plausible alternative nuisance or prior model.

When you are ready to move on

You have completed this path when you can pass the preparation check, explain every link in the theory-to-data chain, solve both checkpoints without opening the hints, complete the two-regime transfer, and produce a capstone that meets all four criteria. From here, use Gravity Beyond General Relativity to examine candidate theories, Relativistic Cosmology to deepen cosmological observables and inference, or the Research destination to follow dated evidence assessments.