the Ornstein-Uhlenbeck semigroup and hypercontractivity
/ OR-n-stine OO-len-bek; hyper-con-TRAK-tiv-it-ee /
On the Gaussian measure there is a canonical dynamics that interpolates between any function and its mean — the Ornstein-Uhlenbeck semigroup. It is the heat flow of the Gaussian world, and its single most remarkable property, hypercontractivity (Nelson), is that it does not merely smooth functions, it actually improves their integrability: it maps an L^p function into a strictly higher L^q for the right times. This is one of the deepest functional inequalities in Gaussian analysis, equivalent to the log-Sobolev inequality.
Let gamma be the standard Gaussian measure. The OU semigroup acts by Mehler's formula: (P_t f)(x) = integral f( exp(-t) x + sqrt(1 - exp(-2t)) y ) gamma(dy), which averages f over a Gaussian re-randomization that decays the input by exp(-t) and tops it up with fresh independent noise. As t -> 0 it is the identity; as t -> infinity it sends every f to its Gaussian mean E_gamma[f]; it is a contraction semigroup on every L^p(gamma) with stationary (invariant and reversible) measure gamma. Its generator is the OU operator L f = Delta f - x . grad f, whose eigenfunctions are the Hermite polynomials: L H_n = -n H_n, so P_t acts as multiplication by exp(-n t) on the n-th Wiener chaos — the chaos decomposition diagonalizes the semigroup. Nelson's hypercontractivity theorem: for 1 < p <= q < infinity, the map P_t : L^p(gamma) -> L^q(gamma) has norm exactly 1 (is a contraction into the larger space) IF AND ONLY IF exp(2t) >= (q - 1)/(p - 1). So waiting time t buys you an integrability upgrade from p to as large a q as that sharp inequality permits.
Why it matters: hypercontractivity is equivalent (Gross) to the Gaussian logarithmic Sobolev inequality, and through it controls eigenfunction estimates, concentration, mixing rates, the analysis of Boolean functions (via the discrete-cube analogue), and Nelson's original use — proving stability bounds in quantum field theory. The honest subtleties: the threshold exp(2t) >= (q-1)/(p-1) is SHARP and is false at p = 1 (no improvement from L^1); and hypercontractivity is genuinely stronger than ordinary contractivity (P_t shrinks L^p norms for free) — the content is the jump UP from p to q, which is special to log-concave measures like the Gaussian.
Take p = 2 and ask for the largest q reachable at time t: the threshold gives q = 1 + exp(2t). So P_t maps L^2(gamma) into L^(1 + exp(2t))(gamma) as a contraction. On a single Hermite mode, P_t H_n = exp(-n t) H_n shows the decay rate grows with the chaos order n — high-frequency (high-degree) components are smoothed away fastest.
Hypercontractivity: at time t the OU semigroup lifts L^p into L^q whenever exp(2t) >= (q-1)/(p-1); chaos H_n decays at rate exp(-nt).
Hypercontractivity is the integrability JUMP from L^p to a strictly larger L^q, with a sharp time threshold; it fails at p = 1 and is equivalent to the Gaussian log-Sobolev inequality.