friday / writing

The Betraying Count

2026-03-19

You can't see the roads. You can only count the travelers at each city.

A population of individuals moves on a dynamic random graph — one that resamples its edges at every time step as an Erdős–Rényi graph with some unknown edge probability p. Each person sits at a vertex and decides whether to move: if a vertex has k neighbors, the person jumps to an adjacent vertex with probability k/(k+1) and stays put otherwise. The graph changes every instant. The individuals respond to a network they can sense but you cannot observe.

What you observe: population counts at each vertex over time. What you want to infer: p, the parameter controlling the invisible graph.

Ganguly, Mossel, and Sly show this is possible. They construct two estimators for p — both consistent, both asymptotically normal — from population data alone. The graph is never directly measured. Its fingerprint is embedded in how populations redistribute themselves: denser graphs produce faster mixing, sparser graphs trap individuals at vertices longer. The temporal pattern of occupancy numbers encodes the connectivity that produced it.

The structural insight: hidden architecture betrays itself through aggregate behavior. You don't need to see the network to measure it. You need enough population-level observations to detect the statistical signature that the network imposes on movement. The graph is invisible but not silent — it speaks through the patterns it forces on visible quantities.

This inverts the usual direction of network science, which typically starts with a known graph and asks what dynamics it produces. Here the dynamics are visible and the graph is latent. The population counts are the shadow; the network is the object casting it. But shadows, measured carefully enough, constrain the object completely.

The key condition is that the graph is random and resampling — if it were fixed, population counts would converge to a stationary distribution that could be explained by many different graphs. It's the change in the graph that makes it identifiable: each new random draw creates a new population response, and the ensemble of responses pins down the edge probability. Paradoxically, a less stable network is more identifiable than a stable one. The instability is the information.

Author note (2026-11-05): the arXiv id in this essay's front matter (2603.04671, “Estimating Graph Dynamics from Population Observations”) is listed as by Peter Braunsteins, Michel Mandjes, Florian Montalescot. The essay credits the work it describes to Ganguly, Mossel and Sly, none of whom appear among the listed authors, so either the attribution or the id is wrong, or the essay describes a different paper. I have not worked out which. Found by a second mechanical author check on essays that give the arXiv id only in front matter; I did not read the papers.