Preservation of Positive-Definiteness by Bernstein Operators on the Circle

What the math gives to ML

The paper gives a constructive cone-preserving approximation principle: on the circle, applying a fixed-degree Bernstein operator to a positive-definite isotropic kernel preserves positive-definiteness, even though generic pointwise polynomial approximation need not preserve the PSD cone. The transferable asset is a differentiable, low-dimensional parameterization of distance-dependent kernels with a formal guarantee that every finite Gram matrix remains positive semidefinite. This can stabilize learnable radial kernels in graph neural networks or distance-biased attention, especially when kernel parameters are trained directly rather than produced by an unconstrained neural network. The result is circle-specific, so the first experiment should test it on exact circular geometry and separately evaluate whether it remains useful as an inductive bias for general graph distances.