The entropic interpolation is the time-marginal flow of the solution to the Schrödinger problem of order — that is, with reference measure having generator . It is the Schrödinger bridge analogue of the displacement interpolation in optimal transport.

While the displacement interpolation is the geodesic in obtained by pushing mass along straight-line optimal transport paths, the entropic interpolation replaces deterministic geodesics with Brownian bridges weighted by the Schrödinger coupling. A key advantage is regularity: the entropic interpolation is always positive and smooth on , whereas the displacement interpolation may concentrate on lower-dimensional sets.

The entropic interpolation satisfies an evolution equation governed by the operator , a second-order perturbation of the first-order transport operator of the displacement interpolation. The Laplacian term provides smoothing and positivity.

Key Details

  • Convergence: uniformly on — the entropic interpolation converges to the displacement interpolation as noise vanishes
  • Entropy formula: Along the entropic interpolation, , where is the iterated carré du champ. This is rigorous (unlike the formal Otto calculus for displacement interpolations) and connects to Ricci curvature via Bochner’s formula:
  • Lott-Sturm-Villani connection: The entropy convexity along entropic interpolations provides a “very lazy gas” version of the lazy gas thought experiment, and is a starting point for synthetic curvature theory

concept