Predicting Collective Behavior from Pairwise Interactions

This direction is split into other directions: pairwise maximum entropy modeling. Its last version stays readable and citable here.

Version 1 of 130 papers34 conceptsfrozen 2026-10-04

When we observe a complex system—the firing of a hundred neurons in the retina, the co-activation of brain regions during a seizure, or the sudden surge of traffic in a city—we are faced with a "curse of dimensionality." If a system has $N$ binary components, it can exist in $2^N$ possible states. For just 100 neurons, this number is larger than the number of atoms in the universe.

How can we possibly model the "language" of such a system without an infinite amount of data? A growing body of research suggests a surprising answer: we don't need to measure every complex combination. Instead, the global behavior of these systems is often almost entirely determined by simple, local interactions between pairs of components.

The Principle of Maximum Entropy

To understand why this works, researchers like Gasper Tkacik and William Bialek turn to the Principle of Maximum Entropy. In Cell biology: Networks, regulation, pathways, they argue that the best model for a complex system is the one that is "least biased"—it matches the statistics we can measure (like the average activity of one neuron or the correlation between two neurons) but assumes nothing else.

This approach leads directly to the Ising model, a classic framework from statistical physics originally designed to explain magnetism. In an Ising model, the probability of the entire system being in a specific state is defined by the "energy" of that state, which is calculated solely from the activity of individual units and the strengths of the connections between pairs.

Why Pairwise Interactions are Sufficient

The most striking discovery in this field is how much information these pairwise models capture. In The Structure of Multi-Neuron Firing Patterns in Primate Retina, Greg D. Field and colleagues showed that pairwise interactions account for ~99% of the structured activity in the retina. Even though the neurons are part of a massive, interconnected network, their collective "codewords" emerge from simple two-cell correlations.

This suggests that biological networks are not mathematically "explosive." As Elad Schneidman and Michael J. Berry demonstrate in Ising models for networks of real neurons, the high-dimensional state space of neural networks is effectively constrained. By knowing how neuron A affects neuron B, and B affects C, we can accurately predict the probability that A, B, and C will all fire together, even if we never explicitly measured that triple-coincidence.

Mapping the Energy Landscape

If we treat a system's states as a map, the Ising model allows us to visualize an energy landscape. Stable, frequent states are "valleys," while rare, unstable states are "peaks."

This landscape has profound clinical and practical implications. Naoki Masuda and colleagues applied this to epilepsy in Energy landscape of resting magnetoencephalography reveals fronto-parietal network impairments in epilepsy. They found that the brains of patients with Juvenile Myoclonic Epilepsy have fewer stable "valleys" in their frontoparietal networks, making their brain states less stable and more prone to the transitions that trigger seizures.

Similarly, in urban planning, Shiyan Liu used maximum entropy to identify "hidden" high-risk traffic states in Hidden high-risky states identification from routine urban traffic. By modeling the pairwise dependencies between road segments, they could detect early warning signals of system-wide paralysis 30 minutes before it occurred.

Beyond Neurons: RNA and Social Surges

The power of this "pairwise sufficiency" extends far beyond neuroscience.

  • Molecular Biology: In 3D RNA from evolutionary couplings, Caleb Weinreb showed that by analyzing pairwise "evolutionary couplings" between nucleotides, researchers can predict the 3D shape of RNA molecules. The model distinguishes direct physical contacts from indirect correlations, filtering out the noise of the genetic code.
  • Social Dynamics: Christopher W. Lynn and Danielle S. Bassett demonstrated in Surges of collective human activity emerge from simple pairwise correlations that viral social surges and traffic jams aren't always caused by external events. Instead, they can emerge organically from simple interactions between pairs of individuals, which account for over 70% of the total correlation in a population.

Limits and Nuances

While pairwise models are incredibly powerful, they are not perfect. As networks scale up, new patterns can emerge. Gasper Tkacik and Olivier Marre found in Searching for Collective Behavior in a Large Network of Sensory Neurons that for very large populations, a "global constraint"—such as the total number of active units—must be added to the model to capture extreme collective events.

Furthermore, Hideaki Shimazaki suggests in Simultaneous silence organizes structured higher-order interactions in neural populations that we should look not just at who fires together, but who is silent together. This "simultaneous silence" can reveal hidden higher-order structures that simple pairwise models might overlook.

Conclusion: A Design Principle of Nature

Why does the universe allow us to describe such complex things with such simple models? The answer may lie in criticality. In Thermodynamics for a network of neurons: Signatures of criticality, Tkacik and Bialek argue that biological systems are often "poised" at a unique critical point between order and disorder. In this state, the system maximizes its sensitivity and dynamic range. Because the system is so finely tuned, its global state becomes a reflection of its local parts, allowing the simple rules of the Ising model to capture the profound complexity of life.

Papers in this version

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_predicting_collective_behavior_from_pair_4886ca82,
  title = {Predicting Collective Behavior from Pairwise Interactions},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/predicting-collective-behavior-from-pairwise-interactions/rdr_4886ca826393476f9df062c782b1de7b},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_predicting_collective_behavior_from_pair_4886ca82_v1,
  title = {Predicting Collective Behavior from Pairwise Interactions},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/predicting-collective-behavior-from-pairwise-interactions/rdr_4886ca826393476f9df062c782b1de7b/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}