Approximate Bayesian Computation

A family of likelihood-free Bayesian methods that simulates data under proposed parameters and approximates a posterior from closeness to observed summaries.

Version

v1 · 2026-09-28 · History

Domain-specific #

7998

Domain group

Formal Sciences

Origin domain

Experimental Design & Statistics

Subdomains

Bayesian Computation, Likelihood Free Inference → Experimental Design & Statistics

Aliases

ABC

Core Idea

Approximate Bayesian computation replaces likelihood evaluation with a simulation-and-comparison loop. Parameters are proposed from a prior or proposal distribution, the model generates synthetic data, and summaries of those data are compared with observed summaries. Proposals producing sufficiently close simulations are retained or weighted to form an approximate posterior.

Three approximations must remain visible. Summary statistics may discard information; a nonzero tolerance admits mismatches; and finite simulation limits Monte Carlo accuracy. ABC therefore expands the class of usable generative models without making inference assumption-free. Identifiability, prior sensitivity, model misspecification, and model comparison can remain difficult or become more pronounced.

How would you explain it like I'm…

Guess the Secret Dials

Imagine a toy machine with secret dials that makes a pattern of marbles, and you want to guess how the dials were set. You try lots of dial settings, run the machine each time, and keep only the settings that made a pattern close to the real one. The settings you kept show your best guesses. That's the idea of approximate Bayesian computation, and it's a good guess, not a perfect answer.

Simulate, Compare, and Keep

Scientists often have a model, like a computer game version of how a disease spreads, with settings they don't know. Normally they'd use a math formula to check how well each setting fits the real data, but for complicated models that formula is too hard to work out. Approximate Bayesian computation skips it: try a setting, run the model to make pretend data, and compare a few key numbers from the pretend data with the real data. Settings whose pretend data come out close enough get kept, and the kept settings show which values are believable. It's approximate because "close enough" isn't exact, the key numbers leave some information out, and you can only run so many tries.

Simulation-Based Bayesian Inference

Approximate Bayesian computation (ABC) is a statistical method for estimating a model's parameters when the likelihood -- the probability of the observed data given the parameters -- is too hard to calculate but the model can still simulate data. It works in a loop: propose parameter values from a prior distribution, simulate a dataset, compute summary statistics, and compare them with the real data's summaries. Values whose simulations come within a tolerance are kept (or weighted), and together they approximate the posterior distribution. Three approximations are involved: summaries may lose information, a nonzero tolerance accepts imperfect matches, and a finite number of simulations adds random error. ABC lets you use more realistic models, but it doesn't remove problems like prior sensitivity, parameters that can't be told apart, or a wrong model.

 

Approximate Bayesian computation replaces likelihood evaluation with a simulation-and-comparison loop, making Bayesian inference possible for generative models whose likelihoods are intractable. Parameters are drawn from the prior or a proposal distribution, the model generates synthetic data, and summary statistics of the synthetic data are compared with the observed summaries using a distance measure. Draws whose simulations fall within a tolerance are retained (as in rejection ABC) or weighted (as in more elaborate sampling schemes) to form an approximate posterior. Three distinct approximations must be tracked. Non-sufficient summary statistics can discard information, so the result approximates the posterior given the summaries rather than the full data. A nonzero tolerance admits mismatches between simulated and observed summaries. Finite simulation budgets limit Monte Carlo accuracy. ABC therefore enlarges the class of usable generative models without making inference assumption-free: identifiability, prior sensitivity, model misspecification, and model comparison can remain difficult or even become harder.

Structural Signature

Sig role-phrases:

• prior distribution — proposes parameter values and supplies Bayesian baseline weight It is essential. Counterfactual: Without a prior the accepted sample is not the ABC posterior approximation described.
• generative simulator — produces synthetic data under a proposed parameter without computing likelihood It is essential. Counterfactual: If the model cannot simulate, the defining likelihood-free comparison cannot run.
• observed data or summaries — provide the empirical target against which simulations are judged It is essential. Counterfactual: No target means acceptance has no inferential direction.
• distance function — quantifies discrepancy between simulated and observed summaries It is essential. Counterfactual: Acceptance cannot be ordered without a comparison rule.
• tolerance or weighting kernel — turns discrepancy into acceptance or importance It is essential. Counterfactual: An arbitrary retained set has no defined approximation.
• posterior approximation — represents the distribution of retained or weighted parameters It is essential. Counterfactual: One best-fitting simulation is not a Bayesian posterior sample.

What It Is Not

• It is not exact Bayesian inference merely because it avoids likelihood evaluation.
• It is not any simulation-based parameter search.
• It is not posterior predictive checking, which evaluates a fitted model for a different purpose.
• It is not guaranteed to recover parameters when summaries are uninformative.
• Closest near-miss. Synthetic likelihood is a near neighbor that models a likelihood for summaries rather than accepting simulations directly through a tolerance rule.

Scope of Application

• Population genetics. Complex demographic simulators support posterior approximation.
• Ecology and epidemiology. Mechanistic stochastic models can be used without tractable likelihoods.
• Systems biology. Simulator outputs are matched through chosen summaries.
• Methodological research. Tolerance, summary, regression, and sequential corrections are compared.

Clarity

Specify prior, simulator, observed summaries, distance scaling, tolerance or kernel, proposal scheme, simulation count, acceptance rate, posterior correction, and validation. Distinguish error from summaries, tolerance, finite simulation, and model misspecification.

Manages Complexity

ABC moves complexity from symbolic likelihood derivation into repeated forward simulation and discrepancy design. This makes mechanistic models accessible while creating a new inferential interface whose summary geometry can dominate the result. The simulator is necessary but not self-validating.

Abstract Reasoning

1. Specify the generative model, parameters, and prior.
2. Choose observed summaries with an argument for relevance or sufficiency.
3. Define and scale a discrepancy measure.
4. Propose parameters and simulate synthetic datasets.
5. Accept or weight proposals according to discrepancy and tolerance.
6. Construct the approximate posterior and diagnose Monte Carlo quality.
7. Test sensitivity to summaries, tolerance, prior, and model misspecification.

Knowledge Transfer

ABC transfers across domains when a trustworthy stochastic simulator exists but a tractable likelihood does not. It stops at black-box curve fitting without Bayesian proposals or posterior weighting. The cargo is simulation-conditioned approximate inference; summaries and distance remain problem-specific.

Examples

Applied / In Practice

Parameter draws from the prior are kept when simulated summary statistics fall within a chosen tolerance of observed statistics.

Mapped back: simulator → Each proposal generates synthetic data.; acceptance → Distance and tolerance filter proposals..

Applied / In Practice

A population of particles is moved through decreasing tolerances, with reweighting between rounds.

Mapped back: efficiency → Later proposals concentrate where earlier simulations matched..

Applied / In Practice

An optimizer chooses the parameter whose simulated curve looks closest, but returns no weighted distribution.

Mapped back: boundary → Simulation matching alone is not posterior approximation..

Structural Tensions

T1 — Computational Reach versus Approximation Bias. Simulation admits complex models, while tolerance and insufficient summaries can move the target away from the true posterior.

Diagnostic: Perform sensitivity checks across summaries, tolerances, and simulation budgets.

T2 — Small Tolerance versus Acceptance Efficiency. Closer matches improve fidelity but become rare, especially in high-dimensional summaries.

Diagnostic: Balance error against effective sample size and use dimension reduction only with justified information retention.

Structural–Framed Character

Proposal, simulation, discrepancy, and weighting are structural; summary choice and acceptable tolerance are inferentially framed. A mathematically coherent ABC run can approximate the wrong target if its compressed observations are inadequate.

Structural Core vs. Domain Accent

The skeleton is indirect conditioning through simulated resemblance. Bayesian statistics supplies priors and posteriors; computation supplies simulation, distances, tolerances, and particle methods. Those commitments define ABC.

Instantiates / Related Primes

This entry typically is a kind of Approximation.

• Approved root. Frozen DAG placement is unparented.

• Related — likelihood-free inference, synthetic likelihood, and posterior predictive checking. They overlap in simulation but use different targets and weighting rules.

Relationships to Other Abstractions

Current abstraction Approximate Bayesian Computation Domain-specific

Parents (1) — more general patterns this builds on

• Approximate Bayesian Computation is a kind of, typical Approximation Prime

  ABC approximates an intractable Bayesian posterior with simulated-data acceptance in place of an unavailable likelihood.

Hierarchy path (1) — routes to 1 parentless root

• Approximate Bayesian Computation → Approximation → Representation → Abstraction

Neighborhood in Abstraction Space

Approximate Bayesian Computation sits in a crowded region of the domain-specific corpus (25^th percentile for distinctiveness): several abstractions share nearly its structure, so a description that fits it tends to fit its neighbors too.

Family — Statistical Hypothesis Tests & Diagnostics (9 abstractions)

Nearest neighbors

• Probability matching — 0.90
• In Silico Experimentation — 0.90
• Misuse of p-values — 0.89
• Neural modeling fields — 0.89
• Fitness-Proportionate Selection — 0.89

Computed from structural-signature embeddings · 2026-10-08

Not to Be Confused With

• Markov chain Monte Carlo. Tell: Usually evaluates a likelihood or unnormalized target.
• Posterior predictive checking. Tell: Simulates from a fitted posterior to assess fit.
• Method of simulated moments. Tell: Matches moments as an estimator without necessarily forming a Bayesian posterior.
• Synthetic likelihood. Tell: Approximates a likelihood for summaries rather than using direct tolerance acceptance.

References

• Frozen Wikipedia discovery revision: https://en.wikipedia.org/wiki/Approximate_Bayesian_computation (revision 1366296011).
• Preserved source candidate: https://www.ssrn.com/abstract=3785580
• Preserved source candidate: https://www.ssrn.com/abstract=3841065
• Preserved source candidate: https://www.ssrn.com/abstract=3785582
• Preserved source candidate: https://www.sciencedirect.com/science/article/pii/S0167947310003786
• Preserved source candidate: https://www.sciencedirect.com/science/article/pii/S0167668710000351
• Preserved source candidate: https://www.risk.net/journal-of-operational-risk/2160915/bayesian-inference-monte-carlo-sampling-and-operational-risk
• Preserved source candidate: http://github.com/icb-dcm/pyabc
• Preserved source candidate: http://www1.montpellier.inra.fr/CBGP/diyabc/

The frozen Wikipedia revision is discovery provenance. The retained source set was reviewed for identity, formal or operational relation, and scope. The encyclopedia's structural synthesis is bounded to those claims; a thin authority surface is recorded as a nonblocking source-strengthening repair rather than concealed.