Mathematical theory that characterizes the exponential decay rates of probabilities of rare events for a family or sequence of stochastic processes or random variables, typically by establishing a Large Deviations Principle (LDP) that provides asymptotic upper and lower bounds of the form P(X_n ∈ A) ≍ exp(-n I(A)) where I is a rate function describing the dominant exponential rate.

Principle

Rate-function principle: under an LDP the probability of atypical outcomes decays on an exponential scale determined by a lower semicontinuous rate function I; minimizers of I over an event describe the most likely ways the rare event occurs (the 'most probable path').

Demonstration

Illustrative scenario → For i.i.d. real-valued variables with finite moment-generating function, Cramér's theorem gives an LDP for the empirical mean: P( (1/n)∑X_i ∈ A ) ≍ exp(-n inf_{x∈A} I(x)), where I is the Legendre–Fenchel transform of the log-moment-generating function; this identifies how probabilities of large deviations of the sample mean scale exponentially with n.

Misapplication

Using the LDP exponent as an accurate finite-sample probability estimate (ignoring multiplicative prefactors and finite-n corrections) or applying LDP results without verifying required technical conditions (e.g., existence of moment-generating function, exponential tightness), which can render the asymptotic approximation invalid.

Consequence

When applicable, large deviations give a principled asymptotic tool to approximate tail probabilities, design importance-sampling schemes, and identify dominant failure modes; misapplied LDP-based estimates can severely under- or overstate actual finite-sample probabilities and misguide risk assessments or algorithm design.

Reversal

LDP conclusions do not hold for heavy-tailed distributions lacking exponential moments or when the chosen scaling is inappropriate; in such regimes algebraic (polynomial) decay or other non-exponential asymptotics dominate and different techniques are required.

Boundary

Clearly within: sequences of random variables/processes satisfying the technical conditions of an LDP with a well-defined good rate function under a natural scaling (e.g., sample averages scaled by n). Boundary case: moderate deviations bridging CLT-scale fluctuations and large deviations; Clearly outside: heavy-tailed variables whose tails decay sub-exponentially and do not admit a standard LDP.

Semantic Tension

Asymptotic clarity (precise exponential rate information as n→∞) versus finite-sample accuracy (lack of reliable prefactors and corrections for moderate n), requiring care when transferring asymptotic statements to practical estimates.

Synthesis

Large deviations theory isolates the exponential scale governing rare events and identifies the most probable ways those events occur; it is a powerful asymptotic lens but must be combined with finite-sample analysis or alternative tail methods when exponential-moment conditions fail or n is not in the asymptotic regime.