the large deviation principle
The large deviation principle is the central organising statement of large deviations theory. The law of large numbers tells you that an empirical mean concentrates at its expected value, and the central limit theorem describes Gaussian-sized fluctuations of order 1/sqrt(n) around that value. But neither says anything quantitative about genuinely rare events — the chance that the empirical mean of n iid variables sits a fixed distance away from its mean, which decays exponentially fast in n. The LDP is the precise machinery that captures this exponential decay and identifies its exact rate.
Formally, a family of probability measures (mu_n) on a topological space X satisfies an LDP with rate function I and speed n if, for every Borel set A, the probability mu_n(A) decays like e^(-n inf over A of I). Because the inf over the interior and the closure can differ, the statement is split into two halves: the lower bound mu_n(G) >= roughly e^(-n inf_{x in G} I(x)) for open sets G, and the upper bound mu_n(F) <= roughly e^(-n inf_{x in F} I(x)) for closed sets F, where 'roughly' means after taking 1/n times the log and a limit. The function I, the rate function, measures the exponential cost of each outcome x; the cheapest point of A dominates the probability of A.
The LDP is the right language because it composes: it passes through continuous maps (the contraction principle), it can be read off from a limiting log-moment generating function (Gartner-Ellis), and it turns weighted exponential integrals into optimisation problems (Varadhan). The speed need not be n — it can be any sequence a_n tending to infinity (n/2 for Schilder's theorem at noise level 1/sqrt(n), for instance). One must always state the topology on X: the same sequence can satisfy an LDP in a weak topology but fail it in a stronger one, so the choice of space is part of the theorem, not a detail.
Toss a fair coin n times and let S_n/n be the fraction of heads. The LLN says S_n/n -> 1/2. The LDP refines this: P(S_n/n >= 3/4) decays like e^(-n I(3/4)) with I(x) = x log(2x) + (1-x) log(2(1-x)), so the probability of seeing 75% heads in 1000 tosses is astronomically small, about e^(-1000 * 0.13), and the rate function tells you exactly how small.
The LDP turns 'rare' into a precise exponential rate set by the rate function.
The LDP gives only the exponential rate (the leading order of the log); it does not pin down the polynomial prefactor. P(S_n/n >= x) ~ C(x) n^(-1/2) e^(-n I(x)) needs the sharper Bahadur-Rao analysis to recover C(x).