Back to search

Boston University

A theory of logifold learning and entropy filtrations

Abstract

dc:description.abstract

This dissertation develops a geometric and entropy-based theory for learning as the formation of a logifold: a fuzzy local-to-global structure whose local charts are models and whose incompatibility is measured by entropy. A learning system does not only fit a function. It also organizes the data space into local domains on which different models provide local descriptions. On overlaps of these domains, disagreement among models defines fuzziness. Entropy therefore becomes a geometric quantity measuring how far local descriptions are from gluing into a coherent global object. The dissertation develops this framework in four stages. First, it introduces fuzzy logical functions and logifold structures as local chart systems for learning. A logifold is not assumed to be globally coherent at the beginning. Instead, its local charts may give different probability-valued descriptions of the same point, and this disagreement is part of the structure. Second, it defines entropy for a chart collection. This entropy is not the entropy of one averaged predictor, but a measurement of the failure of local charts to glue. When the total entropy vanishes, the fuzzy local system becomes a strict global structure up to measure zero. Pointwise entropy then defines cores: regions where the learned structure is sufficiently coherent. Third, it studies learning dynamics in terms of entropy, cross-entropy, and truth functions. Conservation-type identities relate the work of learning to the difference between cross-entropy and entropy, while entropy-decrease principles show that, under convergence to the truth and suitable compatibility assumptions, fuzziness vanishes. In this sense, learning is viewed not only as optimization, but also as the formation of a coherent local-to-global geometry. Finally, it connects the entropy cores to topological data analysis. Varying the entropy threshold produces a filtration of the learned space. Persistent homology and persistent Laplacians are then used to describe the multiscale shape of the coherent and partially coherent regions. Together, these constructions present learning as a process in which fuzzy local models organize into geometric structure, entropy measures coherence, and topology records the shape revealed by this organization.

Degree

thesis:*
Name thesis:degree_name
Doctor of Philosophy
Level thesis:degree_level
doctoral
Discipline thesis:degree_discipline
Mathematics & Statistics
Grantor
Boston University
Year dc:date.issued
2026

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Jung, Inkee
Advisor dc:contributor.advisor
  • Lau, Siu-Cheong
ORCID dc:identifier.orcid

Subjects

dc:subject × 6

Rights

dc:rights
Statement dc:rights
  • Attribution 4.0 International
Language dc:language.iso
en_US

Identifiers

dc:identifier.*
Handle dc:identifier.uri
https://hdl.handle.net/2144/53582

Chain of custody

source
Harvested from
Boston University
Base URL
open.bu.edu/oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Jung, Inkee. A theory of logifold learning and entropy filtrations. doctoral thesis, Boston University, 2026. https://hdl.handle.net/2144/53582