A conditional distribution describes the law of a random quantity after specifying information about another. If takes values in a measurable space and is the conditioning variable, a regular conditional distribution is a measurable probability kernel satisfyingfor measurable sets . Thus is a version of for every target event. On standard Borel spaces such kernels exist, and their values are determined only for -almost every .
For discrete variables at a point of positive mass, the conditional probability mass function is the familiar joint-mass/marginal-mass ratio. For variables with a joint probability density, the analogous ratio applies where . When is continuous, typically holds, so dividing event probabilities by cannot define the conditional law; the kernel or density construction supplies it instead.
For a bivariate standard normal distribution with correlation , the conditional law is normal with mean and variance . This shows how an observation changes prediction and remaining uncertainty while the original marginal law of remains standard normal.
The probability mass function of a discrete conditional distribution. The elementary ratio requires ; a value of zero gives no event-conditioning definition through this ratio.
A regular conditional distribution of a random variable given is a Markov kernel such that, for all measurable sets ,For each , is a probability measure; for each , is measurable. Thus it assigns an actual probability distribution to a specified observation even when . Such kernels exist for standard Borel state spaces, including real-valued variables, and their uniqueness is only up to -null sets. With a joint density and positive marginal density, the ratio of the two densities defines the usual conditional density kernel. Values outside the marginal support or on a null set are not uniquely determined by the joint law.
Where the marginal density is positive, a joint probability density determines the conditional densityIt integrates to one in , and supplies conditional expectations by integrating against it. Its values on a probability-zero set of conditioning points are not fixed by the joint probability distribution.
A Full conditional distribution is the conditional law of one component of a joint model given all the other components and the observed data. A Gibbs sampler updates components by drawing from these laws.
Articles by others on the same topic
There are currently no matching articles.