IPT Interactive exploration
An illustrative mathematical model

One system, different views

Watch a distribution spread, then compare the whole system with a smaller observer region. Every rule used in this example is stated below.

Set up an experiment

Ready. No simulation is running.

Changing the pattern or seed resets the experiment. Changing rate or radius keeps the current distribution.

A heat map of the distribution. Numeric results are listed below.
Darker: less weightBrighter: more weight
Update steps
0
Total normalized weight
1.000000
Global entropy (bits)
0.000
Observer entropy (bits)
0.000

The CSV records settings, observer size, and each sampled result. Heat-map brightness uses the largest value in the current frame.

What this demonstrates

This is classical diffusion on a supplied grid, with Shannon entropy calculated from normalized nonnegative weights. It helps explain the effect of spatial distribution and observer selection. It is not a microscopic IPT derivation or an experimental test of the theory.

The model, fully specified

Grid and update rule

The model uses a 32 × 32 lattice with periodic boundaries: a value leaving one edge can enter at the opposite edge. At each update, each cell retains part of its weight and receives a contribution from its four nearest neighbors.

p′[i,j] = (1 − 4α)p[i,j] + α(p[i−1,j] + p[i+1,j] + p[i,j−1] + p[i,j+1])

For 0 ≤ α ≤ 0.25, every coefficient is nonnegative and the coefficients sum to one. Starting from nonnegative normalized weights, the update conserves their total, apart from floating-point roundoff.

At exactly α = 0.25, a localized spike on this even-sized periodic lattice alternates between two checkerboard subsets. Its long-run entropy approaches 9 bits rather than the uniform-grid maximum of 10. Values below 0.25 retain a self-weight and avoid that particular parity effect.

What “entropy” means here

H(p) = −Σ pᵢ log₂(pᵢ), with zero-weight terms omitted.

The global calculation uses the whole lattice, whose maximum entropy is log₂(1024) = 10 bits. The observer calculation uses a centered circular mask and renormalizes the selected weights to sum to one. The observer’s maximum depends on the number of selected cells, so its raw entropy is not directly comparable to the global value as a fraction of the same maximum.

If the observer region has zero weight, its conditional distribution is undefined and the interface displays “N/A.” The observer region itself is a fixed, ordinary circular selection; it does not wrap around the boundaries.

Limits and reproducibility

The example supplies its space, neighbor relationships, time steps, diffusion rule, and observer mask. It does not infer those ingredients from data. The random case uses a fixed seeded generator. CSV exports record the model version, initial settings, elapsed steps, diffusion rate, radius, selected-cell count, weight, and entropies.

The model stops automatically after 2,000 steps to keep a session bounded. It starts paused and pauses when the tab is hidden.

Read the numerical model source · Download the Python illustrations