Manifold Hypothesis: Conditional Certificates and Consistent Intrinsic Dimension Estimation
Abstract
We study what noisy data can justify before a manifold and its intrinsic dimension are assumed. For repeated observations with independent isotropic Gaussian errors of unknown variance, we compose a support confidence event with a sound geometric verifier to certify a nearby smooth witness of arbitrary compact latent support. Whole-support inference requires an external lower mass condition; independent validation instead certifies proximity of probability mass. Separate all-scale hypotheses yield true-support structure and dimension recovery: a credited Reifenberg--Dini criterion and positive continuous density give local C1 structure and consistent dimension estimation, while two-sided Ahlfors regularity and positive reach give quantitative covariance separation under the stated sampling and observation-transfer conditions. An exact-arithmetic implementation verifies restricted witness families. A further matrix check certifies linear coordinates on a geometrically verified affine span containing the witness and propagates approximation and decoding errors. This restricted construction neither implements the general minimal-dimension Whitney theorem nor infers exact manifold support from finite approximation. Executed torus and Julia-set experiments test the observation-to-witness-to-coordinate route, including singular-support controls. Dynamical systems and chaos are the primary application context; state-identification and finite-horizon prediction results are conditional on observation embeddings and transition validation. Takens reconstruction is an optional scalar-delay application requiring additional smoothness, dynamics and observable conditions. Image comparisons are auxiliary reconstruction diagnostics. The finite-data verifier establishes neither those dynamical conditions nor the all-scale geometric premises for true-dimension recovery.
Related
- Error Amplification in Koopman Predictive Control: Computable Bounds, a Conditional Lyapunov Certificate, and Horizon-Consistent Estimation
- Work in progress: A Lean-verified conditional reduction of periodic Navier-Stokes regularity on T^3 to a single integrability hypothesis
- Hypothesis Testing with Conditional Queries: Learnability and the Value of Interaction
- Simulation-consistent Estimation of the Marginal Likelihood for Block Models
- Strongly Consistent Estimation of the Extended ℓ1-Sum of ϕ-Mixing Coefficients from a Single Trajectory
BibTeX
@misc{choi2026manifold,
title = {{Manifold Hypothesis: Conditional Certificates and Consistent Intrinsic Dimension Estimation}},
author = {U Jin Choi},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2609.21311},
eprint = {2609.21311},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2609.21311},
}
RIS
TY - UNPB
TI - Manifold Hypothesis: Conditional Certificates and Consistent Intrinsic Dimension Estimation
AU - U Jin Choi
PY - 2026
DA - 2026/10/16
PB - arXiv
DO - 10.48550/arxiv.2609.21311
UR - https://arxiv.org/abs/2609.21311
AB - We study what noisy data can justify before a manifold and its intrinsic dimension are assumed. For repeated observations with independent isotropic Gaussian errors of unknown variance, we compose a support confidence event with a sound geometric verifier to certify a nearby smooth witness of arbitrary compact latent support. Whole-support inference requires an external lower mass condition; independent validation instead certifies proximity of probability mass. Separate all-scale hypotheses yield true-support structure and dimension recovery: a credited Reifenberg--Dini criterion and positive continuous density give local C^1 structure and consistent dimension estimation, while two-sided Ahlfors regularity and positive reach give quantitative covariance separation under the stated sampling and observation-transfer conditions. An exact-arithmetic implementation verifies restricted witness families. A further matrix check certifies linear coordinates on a geometrically verified affine span containing the witness and propagates approximation and decoding errors. This restricted construction neither implements the general minimal-dimension Whitney theorem nor infers exact manifold support from finite approximation. Executed torus and Julia-set experiments test the observation-to-witness-to-coordinate route, including singular-support controls. Dynamical systems and chaos are the primary application context; state-identification and finite-horizon prediction results are conditional on observation embeddings and transition validation. Takens reconstruction is an optional scalar-delay application requiring additional smoothness, dynamics and observable conditions. Image comparisons are auxiliary reconstruction diagnostics. The finite-data verifier establishes neither those dynamical conditions nor the all-scale geometric premises for true-dimension recovery.
ER -