Instant TravelLuminary Oddities — Enter the Portal
Research Archive

Research Article · Preprint Mathematics

First-Hit Stopping and Bias-Corrected Score Inference in Dimension Three

Scott Schubr

Luminary Oddities LLC

  • Mathematics
  • Probability and Statistics
Primary subject
Mathematics
Subject subfield
Probability and Statistics
Paper type
Research Article
Manuscript stage
Preprint
Manuscript
Archived
Edition
2 · Permanent link
Read / download PDF Publication filesCite this edition

Overview

Abstract

We study the actual first residual crossing of the symmetric half-Sinkhorn iteration for observations from a fixed positive C⁸ density on the three-dimensional torus. At two heat scales, computed from the same sample, a linear combination removes the first-order population score bias. For ε = n^(−κ) with 2/17 < κ < 4/9 and every fixed 0 < γ < 3/4, stopping each run at its first projective mass residual below n^(−1/2) h^(−γ) preserves the corrected estimator’s fixed-query Gaussian limit. The stopping index is asymptotic to (3/4 − γ) log(1/h)/log 2, and the attained residual lies between c times the threshold and the threshold with probability tending to one for every fixed 0 < c < 1/2. The covariance retains the dependence between the two scales. An explicit cubic population coefficient gives a fixed analytic density for which the limit is shifted at κ = 2/17 and the normalized bias diverges below it, also under actual stopping. The formal Gaussian range extends to κ < 4/9; consistency and shrinking uncertainty require κ < 2/5. A sharp heat-spectral trace determines logarithmic corrections to the stopping count and phase-dependent residual limits, proving that the constants 1/2 and 1 in the residual bracket are sharp. Necessity or equality at the upper exponent is not asserted.

Keywords entropic self-transport · symmetric Sinkhorn · score estimation · first-hit stopping · bias correction

Scope of the work

Fixed positive C⁸ density on the unit three-dimensional torus, a fixed deterministic query, exact periodic heat kernel, the same observations at both scales, and fixed power exponents. Corrected first-hit Gaussian inference holds for 2/17 < κ < 4/9 and every fixed 0 < γ < 3/4; consistent estimation and shrinking uncertainty require 2/17 < κ < 2/5. Sharp stopping-clock and residual-phase geometry has the separate range 2/19 < κ < 2/3 and does not extend the score-inference range. The strict lower 2/17 obstruction concerns this specified correction and stopping rule. Equality or necessity at the upper 4/9 endpoint, the high-probability residual inequality at exactly 1/2, and one-sided integer-phase rounding remain unresolved. No finite-sample coverage, data-selected query, simultaneous spatial band, minimax optimality, nonpower bandwidth, γ = 0 result or measured runtime gain is claimed.

Reproduction notes

Manuscript version 2.1, October 20, 2026; next edition of the existing Dimension Three paper. The thirteen-member package contains the complete 31-page manuscript PDF, nine LaTeX components, a portable Python build helper, README and citation metadata. The proof is analytic; finite diagnostic computations are not premises. Rebuild using Python 3.9 or later and Tectonic with its LaTeX resources installed; the README describes a new output directory, optional compiler/cache paths and offline mode. Rebuilt PDF bytes may differ with compiler resources or metadata. Version 1.1 contained the corrected estimator, same-sample covariance, lower obstruction and earlier stopping results; version 2.0 added the wider range, complete second-order trajectory and sharp clock/phase law. Version 2.1 adds notation and revision-history clarification only. Prior public editions and the separate dimensions-one-and-two paper remain preserved.

The manuscript

Read the paper

Open PDF

The complete manuscript, as published in this edition.

If the reader is unavailable on your device, open or download the PDF directly.

Author disclosures

AI tools assisted mathematical development, literature discovery, drafting and checking. The author is responsible for the content. Internal AI Publisher assessment is not external human refereeing, journal acceptance, formal verification or exhaustive priority certification. Existing score central limit theorems, entropic bias correction, operator expansions and statistical stopping methods are credited. Novelty is supported narrowly against inspected predecessors; worldwide priority remains provisional.

References and related publications

  1. S. Schubr. Residual stopping and pointwise inference for nonuniform entropic self-transport. Luminary Research Archive, edition 1, 18 October 2026.
  2. S. Schubr. Residual certificates and empirical sharpness for entropic self-transport scores. Luminary Research Archive, edition 1, 18 October 2026.
  3. G. Mordant. The statistics of entropic optimal transport with decreasing regularisation. Preprint, arXiv:2412.12007v2, 29 April 2026.
  4. A. Kokot. Variational Problems in Feature Selection and Data Compression. Ph.D. dissertation, University of Washington, 2026.
  5. C. L. Wormell and S. Reich. Spectral convergence of diffusion maps: improved error bounds and an alternative normalization. SIAM Journal on Numerical Analysis 59(3) (2021), 1687–1734. doi:10.1137/20M1344093.
  6. X. Cheng and B. Landa. Bi-stochastically normalized graph Laplacian: convergence to manifold Laplacian and robustness to outlier noise. Information and Inference: A Journal of the IMA 13(4) (2024), iaae026. doi:10.1093/imaiai/iaae026.
  7. I. Maeda, R. Yao and A. Nitanda. Statistical analysis of the Sinkhorn iterations for two-sample Schrödinger bridge estimation. Preprint, arXiv:2510.22560v1, 26 October 2025.
  8. N. Kessler, R. Magnet and J. Feydy. Sinkhorn normalization of diffusion kernels. In Proceedings of the 43rd International Conference on Machine Learning, Proceedings of Machine Learning Research 306 (2026), 56764–56795.
  9. J. Feydy, T. Séjourné, F.-X. Vialard, S.-i. Amari, A. Trouvé and G. Peyré. Interpolating between optimal transport and MMD using Sinkhorn divergences. In Proceedings of the 22nd International Conference on Artificial Intelligence and Statistics, Proceedings of Machine Learning Research 89 (2019), 2681–2690.
  10. L. Chizat, P. Roussillon, F. Léger, F.-X. Vialard and G. Peyré. Faster Wasserstein distance estimation with the Sinkhorn divergence. In Advances in Neural Information Processing Systems 33 (2020).
  11. J. A. Tropp. An introduction to matrix concentration inequalities. Foundations and Trends in Machine Learning 8(1–2) (2015), 1–230. doi:10.1561/2200000048.
  12. S. Minsker. On some extensions of Bernstein's inequality for self-adjoint operators. Statistics & Probability Letters 127 (2017), 111–119. doi:10.1016/j.spl.2017.03.020.
  13. Z. Goldfeld, K. Kato, G. Rioux and R. Sadhu. Limit theorems for entropic optimal transport maps and the Sinkhorn divergence. Electronic Journal of Statistics 18(1) (2024), 980–1041. doi:10.1214/24-EJS2217.
  14. R. Yao, Y.-H. Kim and G. Schiebinger. Quantitative stability of many-marginal Schrödinger bridge. Preprint, arXiv:2604.15191v1, 16 April 2026.
  15. E. Pauwels and S. Vaiter. The derivatives of Sinkhorn–Knopp converge. Preprint, arXiv:2207.12717v2, 3 August 2022.
  16. X. Li, F. Lu, M. Tao and F. X.-F. Ye. Robust first and second-order differentiation for regularized optimal transport. Preprint, arXiv:2407.02015v2, 21 October 2024.
  17. D. Scieur, Q. Bertrand, G. Gidel and F. Pedregosa. The curse of unrolling: rate of differentiating through optimization. Preprint, arXiv:2209.13271v2, 19 May 2023.
  18. L. Liu, R. Mehta, S. Pal and Z. Harchaoui. The benefits of balance: from information projections to variance reduction. Preprint, arXiv:2408.15065v2, 11 February 2025.
  19. G. Blanchard, M. Hoffmann and M. Reiss. Optimal adaptation for early stopping in statistical inverse problems. SIAM/ASA Journal on Uncertainty Quantification 6(3) (2018), 1043–1075. doi:10.1137/17M1154096.
  20. A. Celisse and M. Wahl. Analyzing the discrepancy principle for kernelized spectral filter learning algorithms. Journal of Machine Learning Research 22(76) (2021), 1–59.
  21. L. Hucker and M. Reiss. Early stopping for conjugate gradients in statistical inverse problems. Numerische Mathematik 157 (2025), 1739–1791. doi:10.1007/s00211-025-01469-4.

A stable scholarly record

Cite this edition

Scott Schubr. (2026). First-Hit Stopping and Bias-Corrected Score Inference in Dimension Three (Version 2). Luminary Research Archive. https://research.luminaryoddities.com/first-hit-stopping-bias-corrected-score-inference-dimension-three/v/2

BibTeX
@techreport{luminary_first_hit_stopping_bias_corrected_score_inference_dimension_three_v2,
  title = {First-Hit Stopping and Bias-Corrected Score Inference in Dimension Three},
  author = {Scott Schubr},
  year = {2026},
  institution = {Luminary Research Archive},
  type = {Research Article},
  note = {Version 2},
  url = {https://research.luminaryoddities.com/first-hit-stopping-bias-corrected-score-inference-dimension-three/v/2}
}

Loading discussion…