Source-linked AI summary
Common-Center Geometry and Certified Radial Reconstruction for Energy-Form Full Conformal Regions
Yiheng Feng
TL;DR
The paper addresses whether convex candidate scores suffice for connected FullCP regions and develops a score-specific geometric reconstruction instead. It derives common-center comparison sets and radial regularity for power distances, yielding exact star-shaped geometry and certified low-dimensional radial approximations within stated scope.
Problem
Candidate-score convexity alone does not guarantee connected FullCP regions, even for empirical averages of losses convex in the candidate argument.
Method
The paper derives exact pairwise-dissimilarity comparisons, identifies a common Fréchet-type center, and combines radial order statistics with score-specific Lipschitz bounds and classical extension machinery.
Results
For power distances, exact common-center star-shaped geometry holds for β ≥1; for 1 < β < 2, radial exits admit Lipschitz control enabling certified inner and outer radial envelopes.
Takeaways & Limitations
The resulting certificate preserves potentially nonconvex star-shaped FullCP geometry and is intended for low-dimensional multivariate outputs.
Takeaways & Limitations
The approach is score-specific, has an unconditional reconstruction range 1 < β < 2 with m ≥2, and makes no claim of high-dimensional scalability or runtime improvement.
Abstract
from arXiv · showhide
This note studies the geometry of full conformal prediction (FullCP) regions generated by an empirical energy-form pairwise score. Candidate-score convexity alone does not guarantee connected FullCP regions, even when the candidate score is an empirical average of a loss convex in its first argument. Direct expansion of the leave-one-out scores shows that each training-point comparison for the energy-form score is exactly a pairwise-dissimilarity sublevel condition. Under symmetry, a constant diagonal, a diagonal lower bound, and attainment of the associated Fréchet-type objective, every comparison region contains a common minimizer; when the comparison regions are convex, the nontrivial exact conformal region is therefore star-shaped about that same point. For power distances $ρ_β(x,y)=\|x-y\|^β$, this deterministic geometry holds for $β\ge1$, while the conventional energy score is strictly proper for $0<β<2$. In the univariate $β=1$ specialization, every nontrivial empirical-CRPS FullCP region is a nonempty closed interval, possibly $\mathbb R$ in the $m=1$ degeneracy. On the unconditional reconstruction range $1<β<2$ and $m\ge2$, explicit data-checkable derivative bounds yield Lipschitz control of the comparison-set radial exits and hence of the exact conformal radial function. These score-specific bounds permit existing directional root-search ideas and classical Lipschitz-extension machinery to yield certified inner and outer radial envelopes with width at most $δ+2Lh_{\mathcal U}$ and corresponding same-ray Hausdorff guarantees. An analytic two-dimensional example shows why retaining star-shaped but nonconvex geometry can matter. The resulting reconstruction perspective is intended for low-dimensional multivariate outputs rather than high-dimensional scaling or runtime improvement.
1 Problem, Positioning, and Scope
The paper derives FullCP geometry from an empirical energy-form pairwise score, showing that convexity of the candidate score alone is insufficient and that score-specific structure enables certified radial reconstruction.
- Candidate-score convexity alone does not ensure connected FullCP regions, even for empirical averages of losses convex in the candidate argument.
- Energy-form comparisons reduce exactly to pairwise-dissimilarity sublevel sets sharing a common minimizer under stated diagonal and attainment conditions.
- For power distances, the deterministic common-center geometry holds for β ≥1, while conventional energy scoring is strictly proper for 0 < β < 2.
- Explicit score-derived radial regularity supports inherited directional root search and McShane–Whitney extension machinery for deterministic inner, full, and outer certificates.
- The certificate is specific to power-distance geometry and targets low-dimensional multivariate outputs, with no claim of runtime superiority or high-dimensional scalability.
2 Exact Score-Induced Full-Conformal Geometry
This section shows that candidate-score convexity does not guarantee connected FullCP regions, then derives the stronger common-center and star-shaped structure of empirical energy-form comparisons.
- Symmetry and a constant diagonal reduce each leave-one-out energy-form comparison exactly to a pairwise-dissimilarity sublevel condition.
- A convex candidate score can still produce a disconnected FullCP region, as shown by the explicit m = 2 example with acceptance outside an interval.
- The counterexample establishes that convexity in the candidate argument is insufficient even within empirical-average convex-loss constructions.
- A diagonal lower bound and an attained Fréchet-type minimum place the same minimizer in every comparison region.
- When comparison regions are convex, the strict vote-threshold aggregate is star-shaped about the common minimizer, although it need not be convex.
3 Power Distances and the Radial Interface
For power distances, deterministic common-center geometry holds for β ≥1, while strict propriety occupies 0 < β < 2; the univariate β = 1 case yields interval-valued FullCP regions.
- Deterministic geometry: For β ≥1, power-distance comparison sets are convex, share every minimizer of the Fréchet objective, and induce common-center star-shaped FullCP regions.The objective is coercive and attains a global minimum; for m ≥2, comparison sets are bounded with finite radial exits.
- Scoring interpretation: The conventional power-distance energy score is strictly proper for 0 < β < 2, loses strictness at β = 2, and is improper for β > 2.At β = 2, the expected score depends only on the forecast mean.
- Univariate CRPS: In one dimension with β = 1, every nontrivial empirical-CRPS FullCP region is a nonempty closed interval containing the full empirical median set.For m ≥2 it is compact; for m = 1, the region is R.
- Range boundaries: The deterministic geometry threshold β = 1 is not interchangeable with the 1 < β < 2 range used for radial certification.The star-shaped conclusion should not be extended below β = 1.
- Radial interface: The common-center star-shaped structure supports an exact radial order-statistic representation without requiring the aggregate prediction region to be convex.Along each ray, the conformal radial boundary is the kth-largest comparison-set exit under the stated convention.
4 Certified Radial Reconstruction
For 1 < β < 2, explicit score-specific bounds control comparison-set radial exits and the exact conformal radial function, enabling certified reconstruction from finitely many directions.
- Scope and assumptions: The reconstruction certificate is specific to power-distance geometry and controls radial and same-ray Hausdorff discrepancies rather than claiming generic FullCP coverage.Its purpose is an explicit common Lipschitz bound for radial exits followed by directional propagation.
- Certified envelopes: The resulting radial inequalities directly imply same-ray Hausdorff bounds by projecting points along rays from the common center.This conversion requires no external result.
- Radial regularity: For 1 < β < 2, derivative and strong-convexity bounds yield explicit Lipschitz constants for each comparison-set radial exit.The argument handles observation foci by segment partitioning and continuity of the gradient.
- Radial regularity: The exact conformal radial function inherits the comparison-set Lipschitz bound because it is an order statistic of the individual radial exits.Order statistics are 1-Lipschitz under the ℓ∞ norm.
- Certified envelopes: Finite certified directional intervals can be extended to global inner and outer radial envelopes using McShane–Whitney-type Lipschitz formulas.The envelope construction uses sampled directions and their certified radial intervals.
- Computation: Existing directional root or bisection procedures can obtain the certified intervals because ray membership is initial-interval valued and the vote count is nonincreasing.A finite search bound is available from Hmax := maxj Hj.
5 Why Preserving Nonconvexity Matters
An analytic two-dimensional example shows that the exact FullCP region can be star-shaped yet nonconvex, with convexification introducing positive radial and Hausdorff discrepancies.
- Analytic example: The one-vote prediction region is exactly the union of the two ball components generated by the comparison sets.The construction uses y1 = (−1, 0), y2 = (1, 0), and ε = 1/2.
- Convexification: Its convex hull is the stadium [−e1, e1] + B(0, 2), replacing the exact nonconvex geometry with a convex superset.The convex hull fills the region between the two radius-2 components.
- Convexification: The example has exact positive radial and Hausdorff discrepancies between the convex hull and the exact region.The Hausdorff discrepancy is attained at the midpoint of the top or bottom flat portion of the stadium boundary, at distance 5 − 2 from either component ball.
- Interpretation: These discrepancies demonstrate why preserving star-shaped but nonconvex geometry can matter, without implying that convex reconstruction is inappropriate for every conformal problem.The example is analytic and specific to the stated configuration.
- Analytic example: For m = 2, d = 2, and β = 3/2, the exact prediction region is the union of two radius-2 balls and is star-shaped about 0 but nonconvex.The points (−1, 2) and (1, 2) are accepted while (0, 2) is rejected.
- Interpretation: The same nonconvexity phenomenon persists in a three-point example, where two symmetric endpoints are accepted while their midpoint is rejected.The appendix example uses m = 3, β = 3/2, and k = 2, with nonzero comparison margins.
6 Computational Scope and Endpoint Limitations
The certification framework uses finite directional sampling and covering bounds, but its unconditional regularity statement excludes β = 1 and its practical scope remains low-dimensional.
- Directional certification: For fixed d and fine resolution, the directional covering law determines how many sampled directions are needed for global certification.Corollary 2 combines certified values on a finite directional set with a sphere-covering radius.
- Directional certification: The fixed-d covering notation is not a dimension-growth statement, and the corresponding uniform metric-entropy form applies when h is fixed and small.The distinction matters when interpreting directional sampling requirements.
- Scope: The intended practical scope is low-dimensional multivariate outputs; no claim is made for high-dimensional scalability, dimension-independent certification, or runtime improvement.Directional covering requirements create the stated scope boundary.
- Endpoint limitations: At β = 1, exact common-center star-shaped geometry still holds, but the global radial function need not be continuous or Lipschitz.Therefore the unconditional radial-reconstruction certificate is restricted to 1 < β < 2.
- Endpoint limitations: The certified range β < 2 also aligns with strict propriety of the conventional energy score, while deterministic convex geometry remains valid for every β ≥ 1.The geometric and scoring-rule boundaries are treated as separate statements.
7 Discussion
The discussion separates the geometry guaranteed by the empirical energy-form construction from the limits of generic convexity and radial certification.
- Candidate-score convexity alone does not explain connected FullCP geometry, whereas energy-form comparisons yield common-center star-shaped regions under convexity.The common center comes from the exact pairwise-dissimilarity comparison identity and an attained Fréchet-type minimizer.
- For power distances, deterministic common-center geometry applies for β ≥1, while the regular reconstruction range is 1 < β < 2 with m ≥2.In that range, supporting-ball radii and radial Lipschitz bounds feed directional root search and Lipschitz-extension machinery.
- The certification tools are inherited machinery rather than claims of new general approximation theory.The paper presents them as consequences of score-specific geometric bounds.
- At β = 1, exact star-shapedness can remain valid even when unconditional global Lipschitz radial control fails.The intended scope is low-dimensional multivariate outputs, without claims of generic FullCP geometry, universal superiority, or computational speed.
A Exact β = 1 Segment Pathology
The β = 1 construction exhibits a segment-shaped comparison geometry whose radial function can be discontinuous away from the common center.
- The example uses y1 = −e1, y2 = 0, y3 = e1 with β = 1.
- The middle comparison region contains the segment joining the two foci because equality in the triangle inequality holds precisely on that segment.
- The common Fréchet minimizer is c = 0, and the accepted comparison geometry preserves the whole segment through that center.
- For directions not collinear with e1, the ray leaves the segment immediately, so the radial function is discontinuous when d ≥2.Thus exact common-center star-shaped geometry at β = 1 does not imply an unconditional global Lipschitz radial certificate.
B A Three-Point Nonconvex Robustness Witness
A three-point β = 3/2 example provides a strict accepted-endpoints/rejected-midpoint witness of nonconvex FullCP geometry.
- The witness uses a common Fréchet minimizer c = (0, 0.5247333004905787534 . . .).
- The listed points are A = (0.58, −0.94), B = (−0.58, −0.94), and M = (0, −0.94).
- For β = 3/2, the comparison margins and vote counts accept A and B but give M only one vote.
- Because k = 2 and equality counts as a vote, the strict margins establish accepted endpoints with a rejected midpoint.The witness therefore supplies a nonzero-margin nonconvexity check.
C Technical Proof Details
The technical details establish convex comparison geometry, supporting-ball representations, radial Lipschitz control, and Hausdorff bounds through same-ray projection.
- For β ≥1, convexity of z 7→∥t −z∥β and Jensen’s inequality provide a ball containment for each comparison set.When Hj = 0, the radius is zero and the comparison set reduces to the common point {c}.
- Hessian eigenvalue bounds and integration across focus crossings establish the required strong-convexity inequality without treating the Hessian as finite at observations.
- Normal-cone identities and tangent-ball arguments yield a uniformized supporting-ball radius for the comparison sets.
- The exact radial function is an infimum of supporting-ball exits, so the ball exit bound transfers to a radial Lipschitz bound.The scalar exit calculation covers both positive and zero residual-radius cases.
- ℓ∞ stability of order statistics and sampled Lipschitz bounds produce inner and outer radial envelopes.
- Same-ray projection converts radial envelope gaps into the three Hausdorff inequalities of Corollary 2.