“…For simplicity, the integration notation ∫ ⋅ dy in this paper will refer to the Lebesgue integral ∫ M ⋅ dy over the manifold, instead of the whole space R n . Further, while (for simplicity) such integrals are written without a specific measure one can equivalently, w.l.o.g., replace dx with an appropriate measure representing data sampling distribution over M. Let g(x, y) ≜ exp −∥x − y∥ 2 /ε , x, y ∈ M, ε > 0, define the Gaussian kernel Gf (x) = ∫ g(x, y)f (y)dy used in [3] to capture local neighborhoods from data sampled from M. Following [3] and related work, we define the Gaussian degree q(x) = ∥g(x, ⋅)∥ 1 = ∫ g(x, y)dy and assume it provides a suitable approximation of the distribution (or local density) of data over the manifold M. Finally, given a measure µ over the manifold, an MGC kernel [1], [2] is defined as k µ (x, y) = ∫ g(x, r)g(y, r)dµ(r). Note that while we use a Gaussian kernel for the remainder of this work, the definitions and theorems to follow do not depend on the choice of g, so long as it is a kernel function.…”