Are posterior samples independent?

Not necessarily. MCMC samples are commonly correlated, while some direct samplers can produce independent draws.

Is one sample the most likely answer?

No. A draw represents one outcome from the distribution and need not be its mode.

Does sampling quantify all uncertainty?

It quantifies uncertainty represented by the model and method; omitted model errors or assumptions may remain unaccounted for.

AI Glossary / Computer vision and multimodal AI / Intermediate

Posterior sampling

Definition

Posterior sampling generates draws from, or approximately from, a posterior probability distribution. The draws represent possible values of unknown quantities after conditioning on observed data and can be used to summarize uncertainty, calculate expectations or generate plausible reconstructions.

What it means

An inference problem may have many explanations consistent with the evidence. Returning only one estimate hides that variety. Posterior sampling instead produces a collection of candidate parameter values, states or images, weighted through the sampling process according to the target posterior. A sample is one possible explanation, not a verified ground truth. In image restoration, multiple reconstructions can reveal ambiguity, but they reflect the prior, observation model and sampling approximation. The phrase is also used in decision methods such as Thompson sampling; the focus here is sampling for inference and reconstruction.

How it works

Specify the posterior using a prior and likelihood, then choose a sampling method. Simple distributions may permit direct independent draws. More complex problems can use Markov chain Monte Carlo, sequential methods or specialized approximate generative approaches. A Markov chain typically produces correlated samples, so the number of stored draws is not the same as the amount of independent information.
Diagnostics assess whether a method appears to explore the target adequately, including convergence and effective sample size where appropriate. Samples can estimate means, quantiles and event probabilities. Posterior predictive sampling adds another step: it generates possible observations conditional on sampled unknowns. That is different from drawing the unknown quantities themselves.

Three practical examples

An image-restoration study generates several candidate clean images consistent with a degraded observation under its model.
A Bayesian calibration analysis samples possible camera parameters to propagate uncertainty into later calculations.
A model-analysis workflow estimates the probability of a decision-relevant event by counting how often it occurs across suitable posterior draws.

Think of it this way

Think of drawing many candidate explanations from a weighted collection rather than selecting only the favorite. Frequent appearances indicate greater probability under the model. The analogy is imperfect because practical algorithms may produce correlated or approximate draws, and an apparently diverse collection can still omit an important region of the true target distribution.

Worked example

Suppose a discrete posterior assigns probabilities 0.2, 0.5 and 0.3 to depths of 2, 3 and 4 meters. Its exact posterior mean is 0.2 × 2 + 0.5 × 3 + 0.3 × 4 = 3.1 meters. Independent posterior draws will select those depths with the corresponding long-run frequencies.
The probability that depth exceeds 3 meters is 0.3. With 1,000 genuinely independent draws, the Monte Carlo standard error of the sampled event fraction is approximately √(0.3 × 0.7/1,000) = 0.0145. Correlated MCMC draws can have a much smaller effective sample size, so treating every stored sample as independent would overstate precision.

Illustrative case study

In a fictional reconstruction project, a team restores a blurred image containing a repeating pattern. Several posterior samples show different plausible alignments of that pattern. The team uses the variation to explain that the observation does not uniquely determine the fine detail. It also tests the sampling procedure on controlled examples and checks whether the generated candidates actually fit the observation model. This case illustrates uncertainty communication, not proof that the sample collection exhausts every valid reconstruction.

Common misconceptions and mistakes

More samples do not fix an incorrect prior or likelihood. A visually varied set is not automatically a valid posterior sample, and repeatedly producing similar outputs does not prove that the posterior is narrow. An algorithm may be stuck in one mode. A posterior mean formed by averaging samples is also different from an individual sample and can look unlike any plausible image. Finally, an approximate method called posterior sampling should be evaluated for how closely its outputs match the claimed target distribution.

Business use

For applied systems, decide which uncertainty summaries help the user or downstream process. Track computational cost, sampling diagnostics and sensitivity to modeling choices. If image samples are shown, identify them as model-based alternatives rather than additional observations. Evaluate measurement consistency and relevant task outcomes, not diversity alone. Where decisions depend on rare events or distribution tails, ensure the sampling evidence is adequate for those specific quantities rather than relying on a reassuring overall sample count.

Check your understanding

A chain stores 10,000 highly correlated draws. Should uncertainty in an estimated mean automatically be calculated as though there were 10,000 independent observations?

Answer: No. Correlation reduces effective information, so appropriate diagnostics and effective sample-size estimates are needed.

Related terms

Markov chain Monte Carlo
Effective sample size
Posterior distribution
Monte Carlo method
convergence diagnostic
posterior predictive sampling
Bayesian inverse problem
uncertainty quantification

Learn next

Learn Monte Carlo error and effective sample size next. Compare a direct sampler with a correlated chain on a simple distribution, then distinguish uncertainty in the unknown quantity from numerical uncertainty in the estimated summary.

Sources

Stan, official MCMC sampling documentation: https://mc-stan.org/docs/reference-manual/mcmc.html Stan, posterior diagnostics and Monte Carlo summaries: https://mc-stan.org/docs/reference-manual/analysis.html Primary documentation search excerpts consulted; full source and editorial review remain pending.

Previous: Posterior probability Next: Postings