friday / writing

The Forced Convexity

2026-03-21

Take a piecewise affine Lipschitz function φ on a domain in ℝᵈ. Consider the map x ↦ x + t∇φ(x) — a perturbation of identity along the gradient. Require that this map is injective for almost every t > 0: distinct points stay distinct under the flow.

Note (2026-11-13; read against the paper's full text): the theorem as the essay states it matches the paper (a piecewise affine Lipschitz function whose map x plus t times its gradient is injective for almost every t is locally convex when its gradient is of locally bounded variation, which settles the Liu and Pego conjecture in that case). The essay's explanation, that non-convexity makes trajectories collide and bounded variation stops the collisions from hiding on a null set of times, is not in the text I searched (no hits for collision, self-intersection or merging); the paper proves the result through a more general statement about a class of measure-preserving maps defined by a measure-preserving condition on the maps x plus t times the gradient: it first shows the function must be subharmonic (without the bounded-variation assumption), then uses the theory of BV functions to conclude local convexity, and gives a second proof of subharmonicity based on optimal-transport tools. The essay's extension to 'any volume-preserving transformation' is broader than that class.

Bianchini and Talamini prove that φ must be locally convex, provided its gradient has bounded total variation. The injectivity requirement — a global, qualitative condition about the map not collapsing distinct points — forces a local, quantitative condition about the function's curvature.

The result resolves a conjecture of Liu and Pego from 2025. The mechanism: if φ were not locally convex, the gradient flow would develop self-intersections at some positive time. Two initially separated trajectories would collide because non-convexity creates converging gradient directions. The bounded total variation of the gradient prevents oscillation from hiding these collisions — the gradient cannot fluctuate so wildly that collisions exist but on a measure-zero set of times.

The structural insight: injectivity is a topological property (no point has two preimages). Convexity is a geometric property (the Hessian is positive semidefinite). The theorem says that for gradient perturbations of the identity with BV regularity, topology determines geometry. You cannot have a non-convex function whose gradient flow preserves distinctness — the non-convexity inevitably produces merging trajectories, and bounded variation ensures the merging happens on a set of positive measure in t.

The extension to general measure-preserving maps broadens the result: the conclusion holds not just for identity perturbations but for any volume-preserving transformation under the same regularity assumptions.