the Cheeger-Gromoll splitting theorem
/ CHAY-gur GROH-mol /
Suppose a complete space with no negative curvature contains a single geodesic that is shortest forever — a straight 'line' you can extend infinitely in both directions and that always realizes the distance between its points. That one infinite line forces the whole space to split: it must be a product, an exact copy of the real line glued onto something one dimension lower. The Cheeger-Gromoll splitting theorem is this striking rigidity statement, the manifold version of the intuition that a perfectly straight infinite road can only exist in a space that is itself flat in that direction.
Precisely: let M be a complete Riemannian manifold with nonnegative Ricci curvature, Ric >= 0, that contains a line — a geodesic gamma : R -> M that is minimizing between every pair of its points (d(gamma(s), gamma(t)) = |s - t| for all s, t). Then M is isometric to a Riemannian product R x N, where N is a complete manifold also with Ric >= 0, and the line is the R-factor. The proof is a tour de force of the Bochner technique: from the two ends of the line one builds Busemann functions b_+ and b_-, whose sum is subharmonic by a Laplacian comparison from Ric >= 0 yet has a minimum, forcing it to be harmonic and then (by a Bochner identity) parallel; a parallel function's gradient is a parallel vector field, and a parallel field splits the manifold isometrically.
The splitting theorem is a cornerstone of the structure theory of nonnegative Ricci curvature, used to peel off flat factors and prove finiteness and classification results (for instance, that a compact manifold with Ric >= 0 has a finite cover that splits as a torus times a simply connected factor — the Cheeger-Gromoll structure theorem). Honest caveats. First, the hypothesis is the WEAKER Ric >= 0, not K >= 0 — this is a major strengthening over the earlier sectional-curvature splitting (Toponogov), and getting it down to Ricci was the hard achievement. Second, you genuinely need a line, not merely a ray; a ray that minimizes only in one direction does not split anything (the paraboloid has rays but no line and does not split). Third, the conclusion is an isometric, not merely diffeomorphic, product — the splitting is metric.
The flat cylinder S^1 x R has Ric = 0 >= 0 and contains lines — each generating line {pt} x R is minimizing forever — and indeed it splits isometrically as (the circle factor) times R, exactly matching the theorem. By contrast the paraboloid of revolution has Ric > 0 but contains NO line: every geodesic eventually stops minimizing (it curls), only rays survive, and the paraboloid does not split (it is not a metric product). The clean lesson: a single honest line plus Ric >= 0 is enough to peel off an R-factor, but a ray is not.
A line + Ric >= 0 forces an isometric R-factor (cylinder splits); the paraboloid has only rays and does not split.
The hypothesis is Ric >= 0, weaker than K >= 0 — that is the whole achievement. And you need a full line, not a ray: rays minimize in one direction only and split nothing.