A framework describing how probabilities of rare events decay exponentially with a scaling parameter, characterized by a lower semicontinuous rate function I that governs asymptotic upper and lower exponential bounds for families of probability measures.

Principle

Asymptotic exponential scaling: probabilities of atypical outcomes scale like exp(−n I(x)) for large n, where I measures the unlikeliness of states and encapsulates the dominant cost of a deviation.

Demonstration

For independent identically distributed samples, the empirical mean satisfies exponential bounds: the probability that the empirical mean falls in a set A decays approximately like exp(−n inf_{x∈A} I(x)), where I is the Legendre-type rate function obtained from the cumulant generating function. This yields precise asymptotic estimates for tail probabilities beyond central limit scales.

Misapplication

Applying large deviations estimates at finite or small sample sizes without error control, or using an incorrect rate function when dependencies or scaling differ; assuming pointwise probabilities equal the exponential approximation rather than asymptotic rates.

Consequence

Permits rigorous quantification of extremely unlikely events, justifies exponential error estimates in statistical mechanics and information theory, and guides importance sampling and rare-event simulation by identifying dominant contribution regions.

Reversal

At the diffusion or central-limit scale probabilities behave like Gaussian tails with polynomial prefactors rather than pure exponential rates; thus LDP describes the far tail regime complementary to moderate deviation regimes.

Boundary

Requires a scaling regime (often n→∞), exponential tightness or compactness conditions, and a well-defined rate function; it may fail or lose informativeness for finite n, nonexponential scaling, or in settings without a good coarse-grained cost function.

Semantic Tension

Competes conceptually with central limit and moderate deviation results that describe fluctuations on smaller scales; LDP emphasizes exponentially small probabilities and a variational cost, while CLT emphasizes Gaussian approximation and variance.

Synthesis

The Large Deviations Principle gives a variational, exponential-rate description of rare events: a rate function assigns a cost to each atypical state, and probabilities concentrate exponentially around minimizers of that cost as the scaling parameter grows.