Study Plan

Source of truth for scope: CLAUDE.md / ntu_proposal_080526.md. Materials: resources.md. How to study / benchmark test: study_methods.md. Learned/not tracking: topic_coverage.md. Weekly record: log.md.

Edit this file in place as you progress — update checkboxes and status inline rather than creating new dated copies of the plan.

Cadence

• ~12–15 hrs/week alongside full-time work.
• Each week: some reading + one worked example/exercise set + a log.md entry.
• Don’t create a notes/ file until you’re actually mid-topic.

Phases

Priority order follows how central each area is to the proposal (see CLAUDE.md): comparative-judgement statistics and optimisation are core investigative content; linear algebra/calculus is support for reading papers, not a goal in itself.

Phase 1 — Discrete math, logic & proof technique

• Concrete Mathematics Ch.1–2 (recurrences, sums)
• Refresh proof technique (induction, direct, contradiction) — via First Principles of Mathematics if needed
• Read Hokamp & Liu (2017) and Post & Vilar (2018) with this foundation in hand

Phase 2 — Algorithms & complexity

• Concrete Mathematics Ch.9 (asymptotics)
• Practice multi-parameter worst-case bounds (not just single-variable Big-O)
• Draft own correctness/complexity notes for a toy constrained-decoding scheme

Phase 3 — Probability & statistics (comparative judgement priority)

• Introductory Statistics Ch.3, 7, 8, 9–10 (probability, CLT, CI, hypothesis testing)
• Read Thurstone (1927), Bradley & Terry (1952), Kendall & Babington Smith (1939) in the original — no textbook substitute exists in current library, see resources.md §3
• Review prior wiki work: reference/estimate-bradley-terry-strengths-from-paired-comparison-data.md and reference/choose-wald-vs-wilson-vs-exact-for-a-proportion-interval.md before rebuilding
• Read Verhavert et al. (2019) meta-analysis on CJ reliability

Phase 4 — Optimisation

• Altman (1999) Constrained MDPs — key chapters TBD once sourced
• Achiam et al. (2017) Constrained Policy Optimization (paper)
• Rafailov et al. (2023) DPO (paper)
• Pareto efficiency / multi-objective optimisation basics

Phase 5 — Linear algebra & calculus (as needed)

• Hefferon Ch.1–3 if working knowledge is rusty
• Mathematics for Machine Learning — dip into relevant chapters per paper being read, not cover-to-cover

Status

Last updated: 2026-08-14 — plan created, no phase started yet. CLAUDE.md reviewed and rewritten (math_wiki references removed, study procedure and document roles clarified) — no change to phases/scope.