Math paradoxes

Glossary / Common knowledge

Glossary · common knowledge

How many confirmations does it take to dare together?

Two generals stand on two hills with the enemy in the valley between them. Together they win, alone they die. They can agree only through messengers, and a messenger can be caught. The generals swap confirmations: "got it", "got your 'got it'"… It seems that after a dozen such notes no doubt would be left. Yet even a thousand confirmations don't give them what they really need.

In plain wordspictures, analogies, small numbers

“In plain words” is hidden on all puzzles.

The two generals paradox

General A sends a messenger: "We attack at dawn". General B replies with his own messenger: "Got it". A replies: "Got your 'got it'". And so on, taking turns. The enemy can catch any messenger, and then the sender never finds out.

After how many delivered messengers can a general attack and be certain he won't be fighting alone?

The last messenger is always in doubt

However many messengers the plan has, someone sends the last one and doesn't know whether it arrived. Finding out needs one more confirmation, and that brings the same trouble. So demanding "more confirmations" doesn't help: it only makes a joint attack rarer. And on the nights when a general does go into battle, the risk of being alone stays the same as the chance of losing a single messenger.

Each delivered messenger adds one "knows that" to the chain. The first: B knows the plan. The second: A knows that B knows. The third: B knows that A knows that B knows. But the generals need all of these floors built at once, without end. That is what is called common knowledge, and messengers never deliver it.

Whisper or say it out loud

Three friends are bored at a party. Each will leave only if the group leaves: standing at the door alone is awkward. You know a taxi is waiting outside. How do you get the news across so that they leave together? Try whispering to each of them separately, and then just say it out loud.

    What common knowledge is

    A fact is common knowledge in a group if everyone knows it, everyone knows that everyone knows it, everyone knows that everyone knows that everyone knows, and so on without end. It sounds like wordplay, but the difference is very practical. When people have to act together and nobody wants to go first, it isn't enough that each of them knows. Each must be sure that the others are sure too.

    The simplest way to create common knowledge is to say something in public, where everyone sees that everyone hears. A chain of private messages, however long, can't replace that.

    Rigoroustables, charts, formulas

    “Rigorous” is hidden on all puzzles.

    "Everyone knows" is not yet common knowledge

    Everyone knows

    Each person knows the fact but may think the others don't. Say each colleague was told separately that tomorrow is a holiday. Will someone still come to work tomorrow "just in case", because they aren't sure about the others?

    Common knowledge

    The same thing was announced at an all-hands meeting. Nobody heard a new fact, but now everyone knows that everyone knows, and that everyone knows that too. Everyone stays home.

    The classic example is Andersen's tale "The Emperor's New Clothes" (1837). Everyone in the crowd sees that the emperor is naked but keeps quiet: what if the others can see the clothes? A child shouts what everybody could already see. He reports nothing new, yet after his words everyone knows the others see it too, and the crowd laughs together.

    Almost common knowledge is no better than none

    It seems a thousand floors of "knows that" is almost infinity, and people should behave almost as if the knowledge were common. In 1989 the economist Ariel Rubinstein showed this is not so. In his "electronic mail game" two players' computers automatically send confirmations back and forth until some message gets lost. For rational players the outcome is the same as if there had been no e-mails at all: however many confirmations arrive, neither takes the risk.

    You can see why in the generals' simulation. General B has received exactly k messengers, and nobody else has come. Perhaps his own reply was caught, and then A also has k messengers and attacks. Or perhaps the messenger A should have sent next was caught, and then A has received only k − 1 and waits. Both are almost equally likely, so right at the threshold general B is guessing, as if tossing a coin.

    P(B alone | B received exactly k) = 1 / (2 − p)

    Here p is the chance of losing a messenger. With p = 10% that is 1/1.9 ≈ 52.6%, a little worse than a coin, and it doesn't depend on k. Play some nights in the simulation above, and your numbers will appear here.

    Rubinstein proved that the equilibrium is built out of such "coins": if it isn't worth the risk at the threshold, it isn't worth it one step higher either, and then one more step, and so on up to any number of e-mails.

    Where common knowledge decides everything

    • Meeting without a phoneIn his book "The Strategy of Conflict" (1960) Thomas Schelling asked students: where and when would you wait for a stranger in New York if you couldn't arrange it? The most common answer was noon at the information booth in Grand Central. Not because it is the best spot, but because each expects the other to think of it. Such points are called focal points.
    • Super Bowl advertisingIn his book "Rational Ritual" (2001) the political scientist Michael Chwe explained why companies pay huge sums for ads in the Super Bowl. Every viewer knows that millions of others are seeing the same thing, and knows that they know it too. That common knowledge is worth most for products people buy when others buy them.
    • Traffic rulesDriving on the right works only when everyone is sure that everyone else will also drive on the right. In his book "Convention" (1969) the philosopher David Lewis used common knowledge to explain how such agreements hold together.
    • Computer networksThe two generals problem is the problem of any two computers agreeing over a network where packets get lost. In 1990 Joseph Halpern and Yoram Moses proved that if messages can be lost, common knowledge cannot be reached. So network protocols don't look for certainty; they rely on time-outs and retries.
    • Games of the mindGame theorists often assume that the players' rationality is common knowledge: each is rational, knows the others are rational, and so on. Whether that assumption holds for real people decides how to play in puzzles such as "Guess 2/3 of the average" or "The traveler's dilemma".

    Two more strange consequences

    • You can't agree to disagreeIn 1976 Robert Aumann proved that if two rational people started from the same beliefs and their probability estimates for an event became common knowledge, those estimates must coincide, however different their information. An open disagreement that persists means one of them isn't quite rational or doesn't trust the other.
    • A sentence everyone already knewIf everyone in a group already knew a fact, a public announcement adds no new information about the world. It adds information about what the others know. Whether that can change anything, think it over in the puzzle "The blue-eyed islanders".
    Deepergeneralisations, theory, history

    “Deeper” is hidden on all puzzles.

    The formal definition

    E¹φ = "everyone knows φ",   Ek+1φ = E(Ekφ),   Cφ = E¹φ ∧ E²φ ∧ E³φ ∧ …

    Ekφ is k floors of "everyone knows that". Common knowledge Cφ requires all floors at once. It can also be defined without an infinite series, as a fixed point: Cφ is equivalent to E(φ ∧ Cφ). That is, a fact is common knowledge when everyone knows both the fact and that it is common knowledge. That is exactly how a public statement works: on hearing it, everyone knows that all the others heard it together with them.

    In the language of possible worlds Aumann described it like this: an event is common knowledge if it holds in every world reachable by any chain of "what if he thinks that…".

    Formulas for the generals

    P(exactly n messengers delivered) = (1 − p)n · p

    B receives the odd-numbered messengers, A the even ones. Under the rule "attack after k" B attacks once at least 2k − 1 messengers are delivered, and A once at least 2k are. Hence:

    together = (1 − p)2k  ·  alone = p · (1 − p)2k−1  ·  alone, given someone attacked = p

    The last share doesn't depend on k at all. A larger k only lowers the chance of attacking in the first place: with p = 10% and k = 10 the generals attack together on only 12.2% of nights, while the risk for whoever goes into battle stays at 10%. For it to vanish, k would have to be infinite, which means the generals would need common knowledge.

    Where the idea came from

    • 1969The philosopher David Lewis in his book "Convention" and, independently, the sociologist Morris Friedell first described common knowledge as an endless chain of "knows that".
    • 1975–1978Akkoyunlu, Ekanadham and Huber described the unreliable messengers problem, and in his 1978 notes on database systems Jim Gray retold it as the two generals problem.
    • 1976Robert Aumann gave the mathematical definition in his paper "Agreeing to Disagree". In 2005 he received the Nobel Prize in economics together with Thomas Schelling.
    • 1989–1990Ariel Rubinstein showed that "almost common" knowledge behaves like none, and Halpern and Moses proved that common knowledge cannot be reached over an unreliable channel.

    Where it shows up

    More on “Games and logic”

    All terms →