DataTerminology.comNEW
Also known as: Burn-in Period, Warm-up Period
An initial period discarded or stabilized before measurement, used to reach steady state in MCMC or verify reliability in hardware.
DESCRIPTION
Burn-in refers to discarding early MCMC samples before the chain mixes to the target distribution or running hardware under stress to precipitate early failures. In MCMC, warm-up adapts step sizes and tuning parameters, and initial samples are excluded from inference. In hardware, burn-in seeks to surface infant mortality defects to improve field reliability. Choosing adequate burn-in length or duration relies on diagnostics, domain experience, and cost-benefit trade-offs.
ALTERNATIVE DEFINITIONS
Wikipedia
In statistics, burn-in is the number of samples to discard at the beginning of a Markov chain Monte Carlo simulation.
Directly defines the statistical MCMC usage.
Stan Development Team
During warmup, adaptation occurs and the draws are not used for inference.
Clarifies that warm-up (burn-in) samples are excluded from posterior estimates in Stan.
PyMC
The tuning samples are discarded by default and are not used for posterior estimates.
Describes automatic discarding of tuning (burn-in) samples in PyMC.
RELATED TERMS
Markov Chain Monte Carlo Convergence Diagnostics Autocorrelation Gibbs Sampling Hamiltonian Monte Carlo Soak Testing Reliability Testing Initialization Simulated Annealing
COMPETING TERMS
RELATIONS
- Broader: Initialization
KNOWLEDGE AREAS
Data Science AI Mathematics Infrastructure Testing & Validation
TARGET AUDIENCE
Data Scientist MLOps Engineer AI Engineer Hardware Engineer
EXTERNAL SOURCES
- Wikipedia - Burn-in
- Stan - Reference manual covering warm-up details
- PyMC - Sampling and tuning concepts in PyMC
- ReliaSoft - Hardware burn-in testing concepts
TOP QUESTIONS
Why is Burn-in used in MCMC?
Early samples reflect initialization and adaptation rather than the target distribution. Discarding them reduces bias in posterior estimates.
Category: Explanation
How long should MCMC Burn-in be?
It depends on sampler and model; use diagnostics like trace plots, R-hat, and effective sample size to assess convergence.
Category: Requirements
Is Burn-in the same as Thinning?
No. Burn-in discards initial samples; thinning keeps every k-th sample to reduce autocorrelation.
Category: Misunderstanding
What is Burn-in in hardware testing?
Operating components under stress to precipitate early failures and improve reliability before deployment.
Category: Example
Can Burn-in be harmful?
Excessive hardware burn-in may reduce remaining lifespan; in MCMC, overlong burn-in wastes computation.
Category: Risk
Does every MCMC require Burn-in?
Most samplers benefit from warm-up, especially adaptive ones; well-initialized chains may need less.
Category: Opinion
How do frameworks handle Burn-in automatically?
Stan and PyMC separate tuning/warm-up and exclude those draws from posterior summaries by default.
Category: Example
What’s a common mistake with Burn-in?
Assuming fixed burn-in length regardless of diagnostics can leave residual bias.
Category: Mistakes
Is Burn-in relevant for online inference?
Streaming or amortized methods reduce warm-up costs, but initial calibration still matters.
Category: Relevancy