I
The photograph
Both panels show the same people and the same links. Nothing in the picture says which one remembers through its people and which through its friendships.
Press play. The contacts start to flow, at the same average rate in both. Then click a node to drop one infection into both worlds at once.
Toy model for illustration. 48 people, a fixed neighbourhood graph, SIR spreading.
Same photograph. Press play.
II
Shuffle the beat
Each row is one person, each tick a contact. Shuffling the order of a person's waiting times keeps the rhythm texture and destroys any memory of what came before.
A number fitted to the waiting times barely moves. A statistic that looks at order collapses. In the paper's real records the fitted node memory sits near 0.9, and shuffling changes it by under 10%.
The first dial is a simple stand-in estimator that reads only waiting times, not the paper's fit. Synthetic bursty records.
- Memory read from waiting times
- —
- Order memory, model-free
- —
Original order.
III
Who called whom
Every contact has an initiator. When it is recorded, the timing of events and the choice of partner can be studied separately, like two clean strands.
When it is not, as in most proximity data, the strands tangle. A missing-information term couples them, and tie reinforcement looks stronger than it is, by 28% to 45% in the paper's tests.
Schematic. The slider is the share of contacts whose initiator is recorded.
IV
The forecast
On a high-school contact record, a model built from the fitted node memory was used to forecast outbreaks. Move the slider to change how contagious the disease is.
The node-memory forecast falls short, by almost a factor of 2 at its worst. Keeping tie reinforcement alone does better when the disease is weak and overshoots when it is strong.
These are the paper's numbers (Sec. II H), not a toy: 314 students, 6 transmissibility values.
What to take away
- Record who initiates each contact.
- Before reading a fitted node memory as memory, test it against shuffled and model-free benchmarks.
M. Tizzani, Fundamental limits to identifying node and tie memory in temporal networks: marginal artefacts and spreading dynamics, Physical Review Research, doi 10.1103/n24h-kkc6 · arXiv:2609.10399 · code · publications
Scenes I to III are illustrative models. Scene IV plots the paper's published values.