Common Knowledge of Rationality

In a world of strategic choices, from a boardroom negotiation to a global political standoff, success often depends not on your own move, but on your anticipation of others' moves. This introduces a complex web of second-guessing: what should you do, given what you think they will do, based on what they think you will do? This dizzying recursive logic seems impossibly complex, yet game theory offers a powerful, elegant concept to cut through the noise: ​​common knowledge of rationality​​. It is the bedrock assumption that everyone involved is not just rational, but that everyone shares the knowledge of this universal rationality, ad infinitum.

This article delves into this foundational principle, exploring the logical machinery it sets in motion. It addresses a core gap in our intuitive understanding of strategy: how can a shared, abstract belief system lead to concrete, predictable outcomes? By unpacking this concept, you will gain a new lens for viewing competition and cooperation.

First, in ​​Principles and Mechanisms​​, we will dissect the concept itself, exploring how it drives processes like the iterated elimination of dominated strategies and backward induction. We will see how this logic can unravel a game to a single, inevitable conclusion. Then, in ​​Applications and Interdisciplinary Connections​​, we will witness this theory in action, seeing how it explains phenomena ranging from stock market bubbles in economics to desperate arms races and the cat-and-mouse games of cybersecurity.

By the end, you'll understand not only the profound power of assuming perfect logic but also its surprising limitations, revealing the fascinating frontier where game theory meets human psychology. Let’s begin by exploring the dizzying spiral of "I think that you think that I think..."

Imagine you are standing at a crossroads. Your choice of path depends not on the path itself, but on which path you expect a friend, who is also trying to meet you, will choose. But your friend’s choice depends on which path they expect you to choose. And you, knowing this, consider their expectation of your choice in your own deliberation. Suddenly, a simple decision becomes a dizzying spiral of "I think that you think that I think..." This is the world of strategic interaction, and at its heart lies a powerful and elegant concept: ​​common knowledge of rationality​​.

This isn't just about being smart. It's about a shared understanding that everyone involved is smart, and everyone knows that everyone knows—all the way down. In this chapter, we'll unpack this profound idea. We won't just define it; we'll see it in action, as a real, mechanical process that chisels away at uncertainty, revealing the logical core of a strategic situation.

Let's start with a simple, foundational idea. A ​​rational​​ player is someone who tries to get the best possible outcome for themselves. But in a game, "best" depends on what others do. So, rationality isn't just about wanting to win; it's about choosing your best move given a belief about what others will do.

What's the most basic form of rational behavior? It's avoiding moves that are just plain bad. We call a strategy ​​strictly dominated​​ if there's another available strategy that gives you a better payoff, no matter what anyone else does. A rational person would never play a strictly dominated strategy. It's like having a choice between getting $5 or getting$10—you always take the $10.

That's Level 1 of rationality: "I won't play a dominated strategy."

But the magic begins when you assume others are also rational.

That's Level 2: "I know you won't play your dominated strategies." Knowing this, you can effectively cross out those options from the list of things your opponent might do. The game has shrunk. And in this new, smaller game, some of your strategies that looked perfectly fine before might now be dominated. So, you eliminate them.

And so it continues. Level 3 is "I know that you know that I won't play my dominated strategies," and this knowledge once again allows you to prune the tree of possibilities. When this "I know you know..." chain goes on infinitely, we have reached ​​common knowledge of rationality​​. It’s an assumption that everyone is a perfect logician, and everyone knows it.

This might seem abstract, but it has a surprisingly concrete effect. Let's see how a simple piece of information, when it becomes common knowledge, can cause a game to completely unravel.

Imagine a game between two players, with the following payoffs:

If you analyze this game, you'll find that no strategy is strictly dominated. Player 1 might play $T$ if they think Player 2 will play $L$, but might play $M$ if they expect $C$. Nothing can be ruled out. The game is stuck.

Now, imagine there's a bit of pre-game chatter—what economists call "cheap talk." Player 1 announces, "I have no intention of playing strategy $T$." Crucially, let's suppose this announcement is truthful and becomes common knowledge. It's not a contract, just a shared piece of truth. The game board, in everyone's mind, is now smaller:

Look what happens. Player 2, knowing that $T$ is off the table, re-evaluates her options. Playing strategy $C$ now yields a payoff of 2 (if Player 1 plays $M$) or 1 (if Player 1 plays $B$). But look at strategy $R$: it yields payoffs of 4 and 2, respectively. No matter what Player 1 does (between $M$ and $B$), $R$ is strictly better for Player 2 than $C$. So, a rational Player 2 will eliminate $C$.

This is the first domino. Now, Player 1, knowing that Player 2 is rational and will never play $C$, faces an even smaller game:

Player 1's turn to reason. Playing $M$ yields 1 or 2. Playing $B$ yields 3 or 4. Clearly, $B$ is now strictly better than $M$. So, Player 1 eliminates $M$.

One final step. Player 2, knowing that Player 1 will surely play $B$, has a simple choice between $L$ (payoff 3) and $R$ (payoff 2). She chooses $L$.

The entire game has unraveled to a single, inevitable outcome: $(B,L)$. A cascade of eliminations, triggered by a single piece of shared knowledge, solved the game completely. This mechanical process is called ​​Iterated Elimination of Strictly Dominated Strategies (IEDS)​​, and it is the physical manifestation of common knowledge of rationality in action.

We've been talking about one strategy being "always worse" than another. But there's a beautiful subtlety here. Sometimes, a strategy isn't dominated by any single alternative, but by a mix of them.

Imagine Player 1 has three choices: $U$, $M$, and $D$. Player 2 can play either $L$ or $R$. Let's focus on Player 1's payoffs:

Is strategy $D$ dominated? Let's check. It's not dominated by $U$, because if Player 2 plays $R$, $D$ (payoff 1) is better than $U$ (payoff 0). It's not dominated by $M$ either, because if Player 2 plays $L$, $D$ (payoff 1) is better than $M$ (payoff 0). So, it seems $D$ is a reasonable choice.

But what if Player 1 doesn't have to choose just one strategy? What if they can delegate the choice to a coin flip? Consider a ​​mixed strategy​​: flip a fair coin, and play $U$ if it's heads, $M$ if it's tails. Let's call this mix $\sigma$. What's the expected payoff of this coin-flip strategy?

• If Player 2 plays $L$, the payoff is $0.5 \times u_1(U,L) + 0.5 \times u_1(M,L) = 0.5 \times 4 + 0.5 \times 0 = 2$.
• If Player 2 plays $R$, the payoff is $0.5 \times u_1(U,R) + 0.5 \times u_1(M,R) = 0.5 \times 0 + 0.5 \times 4 = 2$.

The mixed strategy $\sigma$ gives a guaranteed payoff of 2. But strategy $D$ gives a guaranteed payoff of only 1. Since $2 > 1$, the mixed strategy of flipping a coin between $U$ and $M$ strictly dominates the pure strategy $D$. A rational player, therefore, would never play $D$. A strategy that can't be beaten by any single opponent, but can be beaten by a roll of the dice!

This brings us to one of the most elegant results in game theory. The set of strategies that a player might play under common knowledge of rationality—the so-called ​​rationalizable strategies​​—is precisely the set of strategies that survive the process of IEDS when we allow for domination by mixed strategies. The abstract, epistemic notion of infinitely nested beliefs finds its perfect, operational equivalent in this mechanical algorithm. This is the kind of profound unity that makes science beautiful.

So far, we've looked at games where everyone moves at once. What about games that unfold over time, like chess or a business negotiation? Here, common knowledge of rationality takes on a different form: ​​backward induction​​.

The logic is simple: to decide what to do today, you must first figure out what you will do tomorrow. To solve a dynamic game, you go to the very last possible decision and figure out what a rational player would do. Then, knowing the outcome of that final stage, you step back to the second-to-last decision and solve that, and so on, all the way back to the beginning.

Consider a multi-stage game where Player 1 first chooses a path, then Player 2 does, leading to a final stage where they play a simultaneous game. How does Player 1 make her initial choice? She can't do so in a vacuum. A rational Player 1 must look ahead. She anticipates that if she sends the game down a certain path, the final stage will be resolved by the logic of IEDS we just discussed. She calculates the payoff she would get in that future. She does the same for the other path. Only then, by comparing the anticipated future outcomes, can she make a rational choice at the very beginning.

The logic of the end of the game propagates backward, determining the rational choices at every prior step. Your best move today is determined by the inevitable logic of the game's final move.

This brings us to a famous and unsettling puzzle: the ​​Centipede Game​​. Two players take turns deciding whether to Take a growing pot of money or Pass it to the other player, which makes the pot even bigger. Let's say the game has four rounds. If you Pass, the other player gets a chance at an even larger pot. But if they Take it, you might end up with less than you started with. The biggest prize is at the very end, if both players always choose Pass.

So, what does our powerful tool of backward induction say? Let's go to the last move (Node 4). Player 2 can either Take a payoff of $6$ or Pass and get $5$. A rational Player 2 will Take. Knowing this, we step back to Node 3. Player 1 can Take a payoff of $5$ or Pass, which he knows will lead to Player 2 taking the pot and leaving him with only $4$. A rational Player 1 will Take. This logic continues, unraveling the game backward. At Node 2, Player 2 will Take (getting $4$ instead of the $3$ she'd get if she passed). This means that at the very first move, Player 1 faces a choice: Take the pot and get a payoff of $3$, or Pass, knowing that Player 2 will immediately Take the pot, leaving him with only $2$. The prediction is inescapable: a rational Player 1 will Take the money at the very first opportunity. The game ends before it even begins.

Here is the paradox. This perfectly logical train of thought, built on the assumption of common knowledge of rationality, leads to a paltry outcome for both players. A little trust, a little hope that the other person might not be a perfect logician, could have made both of them much richer.

And when this game is played in experiments, that's exactly what we see. Most people do not Take the money on the first round. They Pass. They try to cooperate. This tells us something incredibly important. The assumption of common knowledge of rationality is immensely powerful, but it is also incredibly fragile. For backward induction to hold, it's not enough for you to be rational. You must believe that I am rational. And you must believe that I believe that you believe that I am rational... all the way to the end of the game. A single shred of doubt—"Maybe she'll make a mistake," or "Maybe she thinks I'm not a ruthless maximizer"—is enough to break the chain.

This is not a failure of game theory. On the contrary, it's where the theory becomes most interesting. It reveals the precise boundary between the world of perfect logic and the complex, messy, and fascinating world of human psychology. And it has spawned new theories, like ​​level-k thinking​​ or ​​Quantal Response Equilibrium (QRE)​​, that try to model players with this limited depth of reasoning, giving us a richer picture of strategic behavior. The cold, hard logic of common rationality provides the essential baseline against which we can measure and understand the hopeful, trusting, and sometimes beautifully "irrational" nature of real human interaction.

Now that we’ve taken apart the clockwork of rationality and seen how its gears—the assumptions and the logic—fit together, it's time for the real fun. Let's wind it up and watch it tick. Where in our world does this intricate machine of “I know that you know that I know…” actually run? You might guess it’s confined to the chalkboard of a game theory class, a niche intellectual plaything. But you’d be guessing wrong. The principle of common knowledge of rationality is a ghost in the machine of human affairs, a silent, powerful force shaping outcomes in arenas as diverse as financial markets, corporate strategy, and global security. Its logic doesn't just predict behavior; in many cases, it creates it.

Let’s take this beautiful theoretical lens we’ve ground and polished—the iterated elimination of dominated strategies—and point it at the world. What we’ll see isn’t just a new perspective, but a deeper understanding of the hidden currents that guide strategic interaction.

Imagine you’re a judge in a peculiar "beauty contest." Your task is not to pick the face you find most beautiful, but to pick the face that you think the other judges, on average, will find most beautiful. What do you do? Your personal preference is suddenly irrelevant. You must ascend to a second level of thinking: what will they think? But wait—every other judge is just as clever as you are. They aren't voting their preference either. They are also trying to guess what the average opinion will be. So you must guess what they think the average opinion will be. This is the famous analogy the great economist John Maynard Keynes used to describe stock market investment. Most professional investors, he argued, aren't trying to pick companies with the best fundamental value. They're trying to pick the stocks that they believe other investors will favor, which will then drive up the price.

We can see this logic in its purest form in a simple game. Imagine a large group of people are asked to pick a number between 0 and 100. The winner is the person whose number is closest to, say, two-thirds of the average of all numbers chosen. Let's assume everyone is rational, and everyone knows everyone else is rational, and so on. What should you choose?

A first-level thinker might reason: "Well, people will pick numbers randomly, so the average might be around 50. Two-thirds of 50 is about 33. I'll pick 33." That’s a good start, but it's not enough. A second-level thinker says, "Hold on. Everyone else will do that same calculation. So they'll all pick numbers around 33. But if everyone picks 33, the average will be 33, and two-thirds of that is 22. I should pick 22!"

You see the cascade beginning. This is common knowledge of rationality in action. The first thing a rational player realizes is that the target number can never be greater than $\frac{2}{3} \times 100 \approx 66.7$. So, picking a number like 70 is a dominated strategy—you can always do better by picking a smaller number. But since this is common knowledge, every rational player knows that no one will pick a number above 66.7. This fact becomes part of the game. The new effective maximum is 66.7. A rational player, knowing this, realizes the target can't possibly be higher than $\frac{2}{3} \times 66.7 \approx 44.4$. Any guess above 44.4 is now a dominated strategy.

The ceiling keeps getting lower. As we iterate this logic—as our shared knowledge of each other's rationality deepens—the range of possible rational choices shrinks relentlessly. This intellectual race to the bottom spirals downward, until it converges on the only number that can survive infinite rounds of this argument: zero. If everyone chooses 0, the average is 0, and two-thirds of the average is 0. It is a self-fulfilling, stable prediction. This remarkable result holds whether the choices are any real number on a continuum or just integers. In the real world, experiments show that while not everyone makes the full logical leap to zero, the winning number is always very low. The game reveals the powerful gravitational pull that shared rationality exerts, drawing behavior away from naive first guesses and toward a much sharper, more subtle equilibrium.

This cascade of logic isn't limited to guessing games. It dictates the brutal, forward march of competition. Consider a technology "arms race" in the world of high-frequency trading (HFT), where milliseconds mean millions. Imagine two trading firms can choose from a set of algorithms: some are legacy systems, and one is a brand-new, lightning-fast "smart-router".

Let's say the firms could, in theory, achieve a reasonably profitable outcome if they both stuck with a slightly older, "good-enough" algorithm. Could they implicitly collude to do so? Common knowledge of rationality says no.

Let’s analyze it from one firm's perspective. They look at their oldest, slowest algorithm. They run the numbers and see that, no matter what their competitor does, a slightly faster legacy algorithm always yields a better profit. The slowest one is strictly dominated. A rational firm would never use it. Out it goes.

But—and here is the crucial step—they know their competitor is also rational and is performing the exact same calculation. So, the slowest algorithm is off the table for both of them. The game has shrunk. Now, in this new, smaller world of strategies, they look again. Perhaps that "good-enough" algorithm now looks weak compared to an even faster one, because the slowest option against which it performed well is no longer a possibility. It too becomes a dominated strategy.

This process continues. One by one, the inferior strategies are peeled away by the sheer force of logic. The introduction of the new, superior smart-router N creates a domino effect. Even if using N against another N is less profitable than some other hypothetical pairings, it survives each round of elimination while its predecessors fall. Why? Because at each step, N provides a better outcome against the surviving strategies of the opponent. The chilling conclusion is that the only strategy profile that can survive this logical gauntlet is for both firms to adopt the new technology. Rationality forces their hand. They don't switch to N because they necessarily want to, but because it's the only choice that survives the assumption that their competitor is also rational. This explains the relentless pressure in so many fields—from military hardware to consumer electronics—to constantly invest in the next big thing, not just to get ahead, but to avoid being the one left holding a dominated strategy.

The world of strategic conflict is murkier still. Here, the best move is rarely obvious; it depends entirely on out-thinking your opponent. Consider a cybersecurity game between an attacker and a network's defender. The attacker has several vectors—Phishing, Ransomware, DDoS—and the defender has a suite of countermeasures.

A first glance at the attacker's payoffs might not reveal any purely dominated strategy. Against one defense, Phishing is best; against another, Ransomware is. But the logic of rationality can be more subtle. What if we consider not just pure strategies, but mixed strategies—a probabilistic blend of options? It's like a pitcher in baseball who doesn't just throw fastballs, but mixes them with curveballs and sliders to keep the batter guessing.

Upon closer inspection, the attacker might realize that their DDoS attack, while sometimes effective, is always worse than a particular coin-flip combination of Phishing and Ransomware. For any defense the defender might mount, a 50/50 mix of PH and RW (for example) yields a higher expected payoff than a pure DDoS attack. Thus, for a rational attacker, DDoS is a dominated strategy. It's a tool they will never use.

This is the first domino to fall in a cross-player cascade. The defender, being rational, knows this. They can now plan their defense knowing that DDoS is off the table. Suddenly, the strategic landscape has changed for the defender. With one major threat gone, they re-evaluate their own options. They may find that some of their own countermeasures, which were only useful against a DDoS attack, are now strictly dominated by others. They discard them.

The attacker, anticipating the defender's rational response, knows which defenses will be discarded. This, in turn, may make one of the attacker's remaining strategies (say, Ransomware) look weaker. This chain of "he knows that she knows that he knows" continues until only one strategy for the attacker and one for the defender remain. It is the only pair of choices that can withstand this intense logical scrutiny. What starts as a complex web of possibilities is whittled down to a single, sharp point of equilibrium. This is the essence of rationalizability—finding the actions that are justifiable under the most rigorous assumptions about your opponent's intellect.

From the abstract beauty of a number game to the cold realities of market competition and digital warfare, the principle of common knowledge of rationality acts as a powerful, unifying thread. It teaches us that in any strategic situation, we must look beyond our own choices and even our opponent's choices, to the shared understanding that binds us. Of course, the real world is messy. Humans are driven by emotion and prone to error. But these elegant models are not meant as perfect descriptions of reality. They are what a physicist might call a "spherical cow"—a simplification that reveals a fundamental force. They expose the powerful, underlying current of logic that flows beneath the turbulent surface of human interaction. To understand this current is the first, and most crucial, step in learning how to navigate it.