Does reversible jump mean every move is accepted?

No; reversibility concerns probability balance, not automatic acceptance.

Must model dimensions differ?

The framework can include equal-dimensional moves too.

Can it choose the true model with certainty?

No; its output reflects the specified target, available data, and sampling quality.

AI Glossary / Statistics and probability / Intermediate

Reversible-jump Markov chain Monte Carlo

Definition

Reversible-jump Markov chain Monte Carlo, or RJMCMC, is a simulation method for sampling a target distribution over states whose parameter dimension can change. In Bayesian applications, a state typically includes a model indicator and that model's parameters. Carefully constructed proposals and acceptance probabilities allow moves between models while preserving the desired joint posterior distribution.

What it means

Sometimes uncertainty concerns both parameter values and how many parameters are needed. A mixture model may have an unknown number of components, or a signal model an unknown number of change points. Ordinary fixed-dimensional updates alone cannot explore all such alternatives.
RJMCMC extends the Metropolis - Hastings idea to this setting. It is a sampling framework, not an automatic guarantee that the selected model is correct or that the chain explores every important possibility efficiently.

How it works

Specify a coherent target distribution over model indicators and parameters. In a Bayesian model-selection problem this combines model probabilities, model-specific parameter priors, and likelihoods. Priors and their normalizing constants across models matter; arbitrary omitted constants can distort model probabilities.
Design within-model updates to explore parameters and between-model proposals to change the model. Examples include adding or removing a component or splitting one component into two and merging two back into one.
A move between different dimensions uses auxiliary random variables to match the total dimensions on the forward and reverse sides. If the current parameter vector has dimension d and the proposal draws r auxiliary quantities, the reverse representation must have the same combined dimension d + r.
Define an invertible mapping between the augmented forward and reverse states. Its Jacobian determinant accounts for how the transformation changes volume. Dimension matching alone is insufficient if the map or inverse is incorrectly specified.
Construct the acceptance ratio from the target-density ratio, forward and reverse move-selection probabilities, auxiliary proposal densities, and the absolute Jacobian determinant. Accept with the smaller of one and that ratio. A rejected proposal leaves the state unchanged and remains part of the chain's sampling history.
The acceptance construction supports detailed balance, but useful sampling also requires the chain to reach the relevant state space and explore it adequately. Poor proposals can make cross-model transitions extremely rare even when the formula is correct.
Check coding with simple targets whose probabilities are known, and monitor multiple chains, model occupancy, transition behavior, and effective sample sizes for quantities of interest. A long run with almost no model changes can be misleading.
After adequate sampling, posterior model probabilities can be estimated by the proportion of retained iterations in each model. Model-averaged predictions require integrating parameter and model uncertainty; selecting only the most frequently visited model discards part of that uncertainty.

Three practical examples

Signal analysis: researchers infer an uncertain number of change points and their associated parameters.
Mixture modeling: an analyst explores models with different numbers of mixture components.
AI-assisted science: a workflow propagates uncertainty over competing structural representations into predictive summaries.

Think of it this way

Think of moving between rooms with different numbers of coordinates. Auxiliary variables provide the extra coordinates needed for a reversible doorway, while the acceptance calculation accounts for both the destination's weight and how that doorway changes volume.

Worked example

Use an invented target with two models. Model A has one parameter theta uniformly distributed on (0, 1) and total target probability 0.6. Model B has two parameters theta and u uniformly distributed on the unit square and total target probability 0.4. Assume no additional likelihood variation.
From A, propose B by drawing u uniformly on (0, 1) and appending it to theta. From B, propose A by dropping u and retaining theta. Use equal probabilities for selecting the corresponding forward and reverse move types. The augmented transformation is the identity on (theta, u), so its absolute Jacobian is 1, and the auxiliary density is 1.
The A-to-B acceptance ratio is 0.4/0.6 = 2/3, giving acceptance probability 2/3. The reverse ratio is 0.6/0.4 = 1.5, so B-to-A acceptance is 1.
The model-level probability flows match: 0.6 × 2/3 = 0.4 × 1 = 0.4, before any common move-selection factor. Within-model updates would still be needed to explore theta rather than preserve its initial value forever. This example verifies one move pair, not a complete general-purpose sampler.

Illustrative case study

Fictional case: Willow Signal Lab compares change-point models with uncertain dimension. Its team first checks an RJMCMC implementation on targets with known model probabilities, then examines cross-model mixing on the real analysis. When chains remain trapped in different structures, the analysts improve proposals and report unresolved uncertainty instead of presenting iteration counts as reliable posterior evidence.

Common misconceptions and mistakes

Do not omit the Jacobian merely because the target ratio resembles ordinary Metropolis - Hastings. It is one only for transformations that justify that value.
Do not compare raw likelihoods alone when the target is a posterior over models. Model priors, parameter priors, and proposal probabilities matter.
Correct acceptance arithmetic does not guarantee convergence or efficient exploration. Rejected iterations must not simply be discarded as though they were not part of the chain.
Model occupancy from a poorly mixing chain is not a trustworthy model probability estimate.

Business use

RJMCMC can support specialized analyses where structural uncertainty affects decisions. Its implementation and diagnostic burden is substantial, so organizations should require transparent target definitions, independently checked reversible moves, simulation validation, and evidence of adequate exploration before relying on model probabilities or averaged predictions.

Check your understanding

Why is an auxiliary variable drawn in the example?

Answer: It supplies the extra coordinate needed for the move to a two-parameter model.

Why are within-model updates still necessary?

Answer: The illustrated jump pair alone leaves theta unchanged and cannot explore its distribution.

Related terms

Markov chain Monte Carlo
Detailed balance
Jacobian
Metropolis–Hastings
Bayesian model selection
model averaging
mixture model
dimension matching

Learn next

Learn next how to construct split-and-merge mappings, how prior normalization affects cross-model inference, and how diagnostics reveal poor exploration of model space.

Model Inference

Sources

Sources: Peter J. Green and David I. Hastie, Reversible jump MCMC, https://people.maths.bris.ac.uk/~mapjg/papers/GreenHastie.pdf ; Green, original 1995 article record, https://doi.org/10.1093/biomet/82.4.711 . The author-hosted chapter provides the general-state-space and change-of-variable derivation. The two-model target and fictional case are original.

Previous: Reverse star schema Next: Reward