Research Archive

Luminary Research ArchiveIndependent scholarship

Luminary Portal

The collection

All Research

Explore the Archive by subject, subfield, and paper type.

5 papers · Page 1 of 1

Statistical Theory

Edition 1

A sharp logarithmic inverse bound for Gaussian estimation

Scott Schubr

Primary subject
Statistics and Data Science
Subject subfields
Statistical Theory · Geometry and Optimal Transport
Paper type
Research Article
Manuscript stage
Preprint

At fixed dimension and positive Gaussian noise level, fix a centered isotropic reference whose twice continuously differentiable potential has Hessian between two positive constant multiples of the identity. Uniformly over isotropic log-concave inputs, we prove a sharp inverse-logarithmic Wasserstein modulus of order 3/2 from either noisy-output W1 distance or excess quadratic risk of the reference Bayes estimator. Explicit narrow-Gaussian perturbations with zero mass, mean and covariance corrections give the matching lower order at every permitted reference. The Gaussian-reference case yields a sharp total linear-MMSE gap modulus at every fixed positive SNR. Curvature-uniform constants yield sharp minimax W2 estimation rates of order (log n)^(-3/2) on two-sided curvature classes and fixed-reference testing at the same separation order. Inputs may have compact support or nonsmooth convex potentials.

    Geometry and Optimal Transport

    Edition 1

    Sharp Stability and Geometric Classification of a Vámos Quartic Deformation

    Scott Schubr

    Primary subject
    Mathematics
    Subject subfields
    Geometry and Optimal Transport
    Paper type
    Research Article
    Manuscript stage
    Preprint

    We study the quartics hλ = 4e₁e₃ + λa²b², where the elementary symmetric polynomials are taken in four variables. We prove that their real-stability interval is exactly [0, 64/27]. Throughout its positive part, the parameter is a genuine complex projective modulus, and the full linear automorphism group of the positive hyperbolicity cone consists of positive scalings and the independent interchanges of the two distinguished variable pairs. These cones are consequently pairwise linearly nonisomorphic and nonhomogeneous, including at the endpoint. The projective surfaces have eight singularities of type A₁ and one of type A₃ in the interior; one additional A₁ point appears at the endpoint. Correspondingly, the dual cone has exactly four Lorentz facets in the interior and exactly five at the endpoint. The arguments give an exact classification for this coefficient deformation of the specialized Vámos quartic, using classical hyperbolicity, singularity, and tangent-cone methods.

      Edition 1

      Frequency interactions and curvature in Fourier density models

      Scott Schubr

      Primary subject
      Mathematics
      Subject subfields
      Geometry and Optimal Transport
      Paper type
      Research Article
      Manuscript stage
      Preprint

      We classify the intrinsic sectional-curvature sign at the uniform density for every three-frequency aligned-cosine affine density model on the unit circle. All sectional curvatures are strictly negative except for the spectrum {m, 2m, 3m}, whose sharp attained range is π²m²[−25−√1649, −25+√1649]. A finite Fourier Gauss formula identifies the normal frequency interactions responsible for this exception. In arbitrary finite dimension, successive frequency ratios exceeding 1+√2 force a diagonal Fourier curvature operator and give an explicit attained negative curvature range; the constant is optimal for this universal diagonality assertion. Two four-frequency examples show that the triple criterion does not extend by inspecting subtriples and that a positive curvature-operator direction need not represent positive sectional curvature. As a computational application, we certify corrected heat-kernel Sinkhorn metrics and their parameter derivatives through order two on a fixed two-parameter box at five bandwidths. The smallest bandwidth has a uniform relative metric-error improvement greater than 870.37. We prove bandwidth-uniform magnitude bounds for the cancellation-preserving interval iteration and provide exact and interval reproducibility code. The curvature theorems concern intrinsic finite-dimensional model geometry; the finite-band certificates do not assert a new all-bandwidth convergence rate or a complexity bound.

        Edition 1

        Leading bias and positive correction of heat-kernel Sinkhorn metrics

        Scott Schubr

        Primary subject
        Mathematics
        Subject subfields
        Geometry and Optimal Transport
        Paper type
        Research Article
        Manuscript stage
        Preprint

        We identify the leading bias of the exact population heat-kernel Sinkhorn tensor on a flat torus and construct a compact, strictly positive correction with fourth-order smooth-input accuracy. The leading quadratic form is minus one forty-eighth of the Wasserstein Hessian of Fisher information, and an explicit nonconstant density exhibits both bias signs. The correction preserves the completed tangent form space and square-root range, with equivalent norms; full L² accuracy remains first order. An interval-certified nonconstant two-parameter example encloses continuum residual, differential-form and quadrature errors. For prescribed analytic density families sufficiently close to uniform, we construct fourth-order finite metric approximations, including normalization and every fixed parameter-derivative order. In each fixed dimension, per-coordinate Fourier resolution and iteration counts are logarithmic in inverse bandwidth; total modes and quadrature nodes grow as the dimension-th power of that logarithm. This is a resolution bound, not a running-time or optimality claim.

          Edition 1

          Sharp convergence of heat-kernel Sinkhorn metrics

          Scott Schubr

          Primary subject
          Mathematics
          Subject subfields
          Geometry and Optimal Transport
          Paper type
          Research Article
          Manuscript stage
          Preprint

          We study the Sinkhorn metric associated with the periodic heat kernel at a smooth strictly positive probability density. We prove convergence to the inverse weighted Laplacian in full L² operator norm at the optimal rate O(h). The sharper rate from Hʳ to L² is O(h¹⁺ʳᐟ²), 0 ≤ r ≤ 2; every exponent is optimal already at constant density. Parameter-differentiated estimates yield second-order convergence of smooth finite-dimensional pullback metrics and their intrinsic Christoffel symbols and curvature under nondegeneracy. A population normalization residual ρ gives projected tensor error O(h + ρ/h), with an obstruction at ρ comparable to h. We also give an explicit projected scale approximation whose tensor differs from the exact tensor by O(h²). The cost depends on the heat scale; fixed Euclidean squared-cost problems with boundaries are outside the scope.