Abstract
We derive the exact finite-sample distribution function, density and moments (when they exist) of the least squares estimator of the coefficient in a Gaussian firstorder autoregression. The intercept may be known or unknown and the initial observation fixed or stationary. With a fixed initial observation, the results cover every real coefficient, including the unit root and the explosive region. The joint characteristic function of the numerator and denominator of the estimation error follows from a bordered, rank-one-corrected tridiagonal Toeplitz determinant and a noncentral exponential factor. Five scalar primitives give this function, and all the derivatives that the inversion and moment formulas require, in closed form, at a cost independent of sample size. Chebyshev expansions resolve confluence near the unit root, accretivity settles the branch of the complex square root, and rearranged closed forms avoid cancellation in the explosive region. The same tools evaluate the local-to-unity limits of all four cases, including the Dickey-Fuller limits, and permit comparison with the closed-form explosive limit. In our benchmarks, exact p-values take milliseconds, and exact critical values and moments under a tenth of a second, at sample sizes from 100 to 1000000 .