Deciphering Interactions in Moving Animal Groups

When a thousand fish turn together in a single fluid wave, with no leader and no signal passing between them, who is each fish actually watching and how closely? For decades, the honest answer was that we didn't know. Not because no one looked, but because the math to read individual decisions out of collective motion didn't exist yet. Gautrais and colleagues built it. The puzzle itself is old. Physicists have approached collective animal motion with minimal, top-down models, asking which simple rules are sufficient to generate the qualitative features of a flock or school. The canonical example is the Vicsek model, where point particles move at constant speed and, at each time step, set their new heading to the average heading of neighbors within a fixed radius. Elegant and surprisingly powerful, variations that allow speed fluctuations or add short-range repulsion leave many large-scale outcomes unchanged. High-density, high-order bands near the onset of collective motion and anomalously strong number fluctuations in the ordered phase are features that are robust across many versions of the model. But here is the tension. Almost every model in the literature is built from theoretical arguments or educated guesses, not from what actual animals do. Fitting parameters to collective behavior, or averaged, emergent behavior, tells you surprisingly little about the individual-level rules underneath. What Gautrais and colleagues did instead was start at the bottom: individual fish, individual decisions, individual data. They worked with barred flagtails, Kuhlia mugil, collected around Réunion Island and tested in a shallow circular tank roughly four meters across. The water was kept shallow enough that fish motion was effectively two-dimensional. Groups of one to thirty fish were tested, with five replicate trials per group size, recorded from a camera fixed five meters above the tank at twenty-four frames per second. Custom tracking software extracted smooth trajectories from the video; for pairs of fish, head position and orientation were tracked manually for higher precision. The central measurement was group polarization, P — a scalar that runs from zero when fish are pointing in random directions, up to one when everyone is swimming the same way. The first striking result came before any modeling at all. Small groups, two fish or five fish, showed high polarization — coordinated and school-like. But once group size reached ten or more, polarization dropped and stayed low. The group shifted from schooling to shoaling: less aligned and more milling. This density-driven transition in a confined tank was the first constraint any model would have to explain. To build that model, Gautrais and colleagues treated each fish as a self-propelled particle whose only dynamical variable is its turning speed — the rate at which it rotates. That turning speed is governed by a stochastic differential equation: at each moment, the fish steers toward a target turning rate determined by what it currently senses, plus random noise. Start with a single fish, where the target is just wall avoidance, a function of distance to the tank wall and the angle of approach. Fit that to the single-fish data, and it works well. Add a neighbor, and the target turning rate becomes the sum of three contributions: the wall term, a positional attraction term that depends on where the neighbor is relative to the focal fish, and an orientational alignment term that depends on which way the neighbor is facing. In plain language, the fish cares both about where you are and where you are going. The team tried more complex functional forms, such as higher harmonics and additional angular dependences, but the data didn't support them, so they were dropped. What remained was lean: the sine of the bearing for attraction and the sine of the heading difference for alignment. Parameters were estimated from two-fish time series using nonlinear least squares. The five key quantities are a relaxation time, a noise variance, a wall-avoidance strength, a positional interaction strength, and an orientational interaction strength. The estimates gave characteristic values: a persistence length of about 0.024 meters, a wall-avoidance coefficient of roughly 0.94, a positional interaction strength of about 0.41 per meter per second, and an orientational strength of about 2.7 per meter. Both social cues matter. Neither alone is enough. The most novel finding in the parameter estimation was speed dependency. The noise, wall-avoidance strength, and orientational interaction strength all scaled approximately in proportion to the fish's own swimming speed. The relaxation time scaled inversely with speed. In other words, a faster fish is noisier, more reactive to walls, and more sensitive to neighbors' headings — not in absolute terms, but in a way that makes sense if you think of interactions as being processed over distance rather than time. This speed dependence is not something any prior model had built in from data. Now comes the validation — and it is genuinely satisfying. The team took those parameter values, estimated entirely from pairs, and ran simulations in virtual tanks with no free parameters left to adjust. For groups of two and five fish, the model reproduces the alignment statistics and neighbor-distance distributions seen in real experiments. Agreement is described as very good. The approach works: a model calibrated on the simplest possible case predicts behavior in modestly larger groups without modification. Then the crack appears. For groups of ten fish and larger, the model's distance predictions remain accurate — it still gets the spacing between individuals right — but polarization diverges. The model predicts systematically higher coordination than the fish actually show, and the discrepancy grows with group size, most pronounced at high swimming speeds. When the team examined why, they found that the positional and orientational interaction strengths, kP and kV, decline with increasing group size. The fish in larger groups are less reactive to each other. The other parameters, including the persistence length and wall-avoidance coefficient, don't show any systematic trend with group size. The social ones do. Gautrais and colleagues treat this as evidence, not failure. Switching from Voronoi neighbors, where each fish responds to the individuals in its nearest spatial cell, to K-nearest neighbors doesn't explain the drop. What does explain it is a density-dependent behavioral change. In a crowded, confined tank, fish appear to reduce their sensitivity to neighbors. The same model run with the higher reactivity measured in small groups would, in open water, maintain high polarization and a typical neighbor distance of about two body lengths. That's a coherent school. The tank, at large group sizes, suppresses it. It is worth pausing on what Voronoi neighbors are, because the choice matters. Rather than responding to all fish within a fixed radius, or to a fixed number of nearest neighbors regardless of distance, the Voronoi rule means each fish responds to those neighbors that share a spatial boundary with it in a tessellation of the group. It's topological — it depends on who is adjacent, not on distance alone. The data favored this rule. The best predictions with K-nearest neighbors came at K values of roughly six to eight, consistent with first-shell Voronoi neighbors. This places barred flagtails in the same qualitative class as starlings, where topological interaction was first reported, and distinguishes them from mosquito fish, where metric interaction was found. Gautrais and colleagues are careful about what their framework can and cannot claim. Every parameter was estimated in a tank. Confinement likely suppressed interaction strengths in large groups, and whether those same parameters or similar ones govern behavior in open water remains to be tested. But the methodological contribution is clear and transferable. Start with single individuals. Build up incrementally through pairs to larger groups. Use nonlinear regression to let the data select among candidate functional forms. Check residuals. Fix parameters and validate without retuning. The procedure produced a compact, interpretable, fully data-grounded model of collective motion — one where the individual interaction rules are explicit, not assumed. That matters because it turns the question around. Once you can read interaction rules from tracking data, you can ask which rules produce which collective behaviors. You can compare species, test conditions, and make predictions about what happens when you change the environment. The fish in this study don't know they're doing anything remarkable. But the methodology for listening to them, carefully enough to hear individual decisions inside collective motion, is new. This lecture was created by ennepō. Go to https://ennepo.ai to Discover, Create and Follow the latest research in your field. Read when you can. Listen when you want to.