What Play Conceals: Identification Limits of Learning Dynamics in Two-Population Games

Imagine watching a game of strategy unfold, not to see who wins, but to figure out the hidden rules that made the players move the way they did. This is the challenge of reverse-engineering a game: an observer watches the flow of play, tracking how choices shift over time, and tries to work backward to discover the exact rewards and penalties that drove those decisions. In the world of game theory, this is a fundamental question. If we can see how people learn and adapt, can we always reconstruct the incentives that shaped their behavior? For decades, researchers have assumed that with enough data, the answer is yes. They believed that if you watch long enough, the path the players take will reveal the map of the game they are playing.

A new study challenges this assumption, showing that the very act of learning can hide the truth. The research focuses on a specific, well-understood model of how players adjust their strategies over time, a process where individuals gradually shift their choices toward options that have paid off better in the past. The researchers asked a simple but profound question: if an outsider watches this learning process from start to finish, what can they actually learn about the game's underlying structure? They found that the answer depends entirely on how the game ends. If the players eventually settle into a stable pattern where no one wants to change their strategy, the observer hits a permanent wall. No matter how long they watch, they can never fully recover the true rewards of the game. However, if the players keep moving in a persistent loop, never settling down, the observer can learn with surprising speed, uncovering details much faster than standard statistical rules would predict.

The study begins by defining exactly what is invisible to the observer. It turns out that certain changes to the game's rules leave the players' behavior completely unchanged. If you add a constant bonus to every possible outcome for a specific player, regardless of what the other person does, the players will not alter their strategy. Similarly, if you speed up or slow down the entire game by a uniform factor, the path the players take remains the same, only the timing changes. The researchers proved that these two types of changes—adding non-strategic bonuses and rescaling time—are the only things that can be hidden. Any other difference in the game's rules will eventually show up in the players' movements. This means that an observer can never know the absolute value of the rewards, only the relative differences between them, and they can never know the exact speed of the game, only the shape of the path.

The most striking discovery concerns what happens when the game reaches a conclusion. In many strategic situations, learning leads to a stable state where players stop changing their minds. The researchers showed that once the players reach this calm, the observer's ability to learn the game's rules stops improving. Even if the observer continues to watch for years, the information they gather does not grow; it hits a hard ceiling. This is a permanent limitation. It means that for games that end in a stable agreement, it is mathematically impossible to perfectly reconstruct the strategic incentives, no matter how much data is collected. The learning process itself destroys the very information needed to understand the game, because the players stop moving, and without movement, there is no new signal to decode.

In contrast, the study found a different reality for games that never settle. Some games, particularly those with a specific type of balance where one player's gain is another's loss, lead to players circling endlessly without ever finding a stable point. In these persistent loops, the observer's ability to learn accelerates dramatically. The researchers discovered that the information gathered grows at a rate far faster than usual. While standard learning usually improves at a steady, linear pace, these looping games allow the observer to uncover the rules at a rate that cubes with time. This means that doubling the observation time does not just double the knowledge; it multiplies it by a much larger factor. The mechanism behind this is that the continuous motion of the players acts like a magnifying glass, making the subtle differences in the game's rules increasingly visible as time goes on.

To confirm these theoretical findings, the researchers ran thousands of computer simulations using random games. They watched how well an observer could guess the game's rules under different conditions. The results matched the theory perfectly. For games that settled down, the error in the observer's guess stopped improving after a certain point, confirming the existence of a permanent blind spot. For games that kept moving, the error dropped rapidly, following the predicted super-fast curve. The simulations also showed that the ability to learn depends heavily on the nature of the game. Games that are close to the "balanced" type that causes endless looping are the ones where learning is fastest, while games that naturally lead to a stable agreement are the ones where learning hits a wall.

The study also explored the geometry of the game space, showing that the ability to learn is not uniform. There are specific directions in the game's rules that are impossible to learn, no matter what happens. These are the non-strategic parts of the game, the bonuses that do not change the relative value of choices. The researchers proved that these parts are completely invisible to the observer. Furthermore, they showed that the center of the game, where players are indifferent between all options, is a place of maximum confusion. If the players are at this center, the observer cannot tell if the game is a balanced loop or a stable point; the information is completely erased.

This work reshapes our understanding of what can be learned from observing behavior. It suggests that the success of learning is not just about how much data we have, but about how the system behaves. If a system settles, the truth becomes permanently obscured. If a system keeps moving, the truth becomes clearer at an accelerating rate. The researchers did not just find a new way to estimate game rules; they identified the fundamental limits of what can ever be known. They showed that the dynamic of the game itself acts as a filter, either preserving the signal of the rules or washing it away. This has deep implications for anyone trying to understand human or artificial intelligence behavior, reminding us that the path to understanding is not always a straight line, and sometimes, the very act of reaching a conclusion hides the answer we are looking for.

The study concludes by placing these findings in the broader context of how we study games. It challenges the common assumption that more data always leads to better understanding. Instead, it shows that the type of data matters. A long, static record of a settled game is less informative than a shorter record of a dynamic, looping game. The researchers suggest that future work should explore how different learning rules might change these limits, but for the specific model they studied, the boundaries are now clear. The game reveals its secrets only when it refuses to stay still.