Exploratory visualizations for arithmetic functions

Several experiments in this project are intentionally “EDA-like”: they produce heatmaps, atlases, and correlation matrices for many functions at once. This page collects best practices so those figures remain interpretable and comparable.

Core definitions

Given a function table \(F(n)\) for \(n=1,\dots,N\) and a list of arithmetic functions \(f_1,\dots,f_m\), two common derived views are:

  • Heatmap atlas: visualize values \(f_j(n)\) as an image (rows = functions, columns = \(n\) or blocks of \(n\)).

  • Correlation matrix: compute an empirical correlation between columns \(f_i(n)\) and \(f_j(n)\) (Pearson or rank-based).

What experiments usually visualize or measure

  • “Texture”: squarefree bands, prime powers, smooth numbers, record-holders, etc.

  • Clustering: which functions behave similarly on \(1..N\) (after normalization).

  • Stability: how patterns change when \(N\) grows or when you sample on a log grid.

Practical numerical caveats

  • Normalize first. Raw scales differ wildly (e.g. \(\sigma(n)\) vs \(\mu(n)\)). Typical choices: z-score, log1p, or scaling by \(n^\alpha\).

  • Beware heavy tails. Extremal values can dominate Pearson correlation; consider Spearman correlation or winsorization for robustness.

  • Sampling matters. Linear \(n\) emphasizes small structure; log-spaced \(n\) emphasizes asymptotic behavior. Record the sampling rule in the figure caption.

References

See References.

[Arnold, 2015, Borwein and Bailey, 2008]

Experiments in this repository

  • E122 — Heatmap atlas of μ, ω, Ω and related “squarefree texture” features.

  • E123 — Correlation matrix of arithmetic functions (with normalization variants).