Chapter 0: Preliminaries
Introduction
Warm-up Game: Rock-Paper-Scissors
Two players simultaneously pick Rock, Paper or Scissors: Rock beats Scissors, Scissors beats Paper, Paper beats Rock; equal picks tie. The winner gains 1 point and the loser pays 1.
Rock-Paper-Scissors is often played against a computer that plays at random or learns the frequency or the pattern of your moves.
Is there a way of playing that a learning computer cannot beat in the long run?
Someone always plays Rock. Once the opponent notices, how many points per round does this player lose on average?
Introduction
Learning Outcomes
Read sets, Cartesian products and notation like u1(s1, s2).
Evaluate sums and geometric series; find a maximum and its argmax, with derivatives (chain rule, partials) for continuous choices.
Find probabilities by counting, multiplying and conditioning (Bayes); use U[0, 1].
Expected payoff of a lottery or a mixed strategy; read exponents, ln and softmax.
Read a pair (a, b) and mark best responses in a matrix.
Solve quadratics and indifference conditions; solve a game tree from the leaves; read O and Θ.
Sets and Functions
Sets and Cartesian Products
A set is a collection of options; the Cartesian product of two sets is the set of all ordered pairs formed by taking one element from each.
Ayşe's strategy set is S1 = {Rock, Paper, Scissors}; Rock ∈ S1 ("Rock is an element of S1") and |S1| = 3.
Bora's is also S2 = {Rock, Paper, Scissors}. The possible outcomes form S1 × S2: |S1 × S2| = 3 · 3 = 9 pairs. Each pair is a strategy profile.
Sets and Functions
Function Notation
A function is a machine: it takes an input and returns exactly one output. The payoff function ui takes a strategy profile and returns player i's payoff (a number).
u1(Rock, Scissors) = 1: if Ayşe plays Rock and Bora Scissors, Ayşe gains 1 point.
u2(Rock, Scissors) = −1: in the same profile Bora pays 1 point.
u1(Paper, Paper) = 0: a tie.
Notation: u1 : S1 × S2 → ℝ ("from profiles to real numbers").
Sets and Functions
Counting: Product Rule, n! and C(n, k)
For successive choices, the numbers of options multiply. n distinct objects can be ordered in n! = n · (n − 1) · … · 1 ways. The number of ways to choose k of n objects, order ignored, is C(n, k) = n! / (k! (n − k)!).
Three players with 2, 3 and 4 strategies: 2 · 3 · 4 = 24 strategy profiles.
Three people can stand in a line in 3! = 3 · 2 · 1 = 6 ways.
A guard protects 2 of 4 targets: C(4, 2) = 4!/(2! · 2!) = 6 pure strategies. 5 of 20 targets: C(20, 5) = 15,504.
Sums and Maxima
The Summation Symbol Σ
Σ means "add them all up". The index below says where to start, the number above where to stop: 3Σi=1 xi = x1 + x2 + x3.
3Σi=1 xi = 4 + 1 + 6 = 11
On a fair die each face has probability 1/6: 6Σi=1 pi = 6 · 1/6 = 1. Probabilities always sum to 1.
An expected value is a sum as well: nΣi=1 xi · pi.
Sums and Maxima
Geometric Series and the Discount Factor δ
A game is repeated every round; a payoff t rounds ahead is multiplied by δt (0 ≤ δ < 1): a payoff of 1 today is worth 1, one round later δ, two rounds later δ². Earning 1 every round gives 1 + δ + δ² + … = 1/(1 − δ).
If S = 1 + δ + δ² + …, then δS = δ + δ² + …; subtracting gives S − δS = 1 ⇒ S = 1/(1 − δ).
Example: δ = 0.9 and a payoff of 3 per round: 3/(1 − 0.9) = 30. δ = 1/2 and 1 per round: 1 + 1/2 + 1/4 + … = 2.
Sums and Maxima
max and argmax
max answers "what is the largest value?"; argmax answers "which option gives the largest value?". In game theory a player chooses an action, not a value, so argmax is used more often.
You pick one of four face-down cards; their prizes are A: 3, B: 7, C: 2, D: 5 chips.
max{3, 7, 2, 5} = 7 (a number)
argmax = B (an option): "pick B".
Sums and Maxima
Trace: Finding the argmax with a Loop
action = ["A", "B", "C", "D"] payoff = [3, 7, 2, 7] best = 0 # index of the best for i in range(1, 4): if payoff[i] > payoff[best]: best = i print(payoff[best], action[best])
Start: take the first card (A, 3 chips) as the best.
i = 1: payoff[1] = 7 > 3? Yes.
best = 1: new best is B (7 chips).
i = 2: payoff[2] = 2 > 7? No; the best stays.
i = 3: 7 > 7? No: with strict ">" a tie keeps the first one found.
max = 7 and the code returns B as the argmax; since D also gives 7, the argmax set is really {B, D}.
Orange bar: the card being examined. Green frame: the best card so far.
Sums and Maxima
Maximizing a Function: Derivative = 0
When the choice is continuous (a quantity q, say), the argmax cannot be read off a table. At the top of the payoff curve the slope is zero: positive to the left of the peak, negative to the right. The slope is the derivative f′(q).
(qk)′ = k · qk−1 · derivative of a constant is 0 · (c · f)′ = c · f′ · derivative of a sum is the sum of derivatives.
Example: f(q) = q(10 − q) = 10q − q². f′(q) = 10 − 2q = 0 ⇒ q = 5, f(5) = 25. The second derivative f″ (derivative of the derivative) tells how the slope changes: f″ = −2 < 0, the slope decreases, so the point is a peak.
Sums and Maxima
Trace: Best Quantity by Derivative
Price = 12 − (q + r); q: the firm's quantity, r: the rival's; no cost. The plot is for r = 4.
Profit = quantity × price: u(q) = q · (12 − q − r).
If the rival produces r = 4: u(q) = q(8 − q) = 8q − q².
Derivative: u′(q) = 8 − 2q.
Set it to zero: 8 − 2q = 0 ⇒ q = 4; profit u(4) = 4 · 4 = 16.
Check: u″ = −2 < 0; neighbors u(3) = u(5) = 15 < 16 ✓
General r: u′(q) = 12 − 2q − r = 0 ⇒ q = (12 − r)/2. Quantities cannot be negative: if r ≥ 12 it is best not to produce, so the full formula is q*(r) = max{0, (12 − r)/2}. Giving the best quantity for every rival quantity, it is a best response in continuous form.
Exponents, Logarithms and Derivative Rules
Exponent Rules and Fractional Exponents
For an integer a, xa is x multiplied by itself a times. Rules: xa · xb = xa+b, (xa)b = xa·b, x0 = 1, x−a = 1/xa, x1/2 = √x. Fractional exponents (such as x0.88) obey the same rules.
δ² · δ³ = δ5; 2−1 = 1/2; 91/2 = √9 = 3; 50 = 1.
Utility x0.88: 1000.88 ≈ 57.5 and 2000.88 ≈ 105.9. Doubling the money multiplies utility by 20.88 ≈ 1.84: with an exponent below 1, growth slows down.
Exponents, Logarithms and Derivative Rules
The Number e and the Natural Logarithm ln
e ≈ 2.718 is a special base; ex is positive for every x. ln (natural logarithm) undoes ex: ln(ex) = x and eln y = y for y > 0. Rules: ln(a · b) = ln a + ln b, ln(ak) = k · ln a, ln 1 = 0. Derivatives: (ex)′ = ex, (ln x)′ = 1/x.
ln 8 = ln 23 = 3 · ln 2 ≈ 3 · 0.693 = 2.079; back again: e2.079 ≈ 8.
Utility u(x) = ln x: wealth going from 10 to 20 raises utility by ln 20 − ln 10 = ln 2 ≈ 0.69; going from 20 to 40 again adds 0.69. u′(x) = 1/x: the richer you are, the less an extra lira is worth.
Exponents, Logarithms and Derivative Rules
The Product Rule and the Chain Rule
Product: (f · g)′ = f′ · g + f · g′. Chain (nested functions): [f(g(x))]′ = f′(g(x)) · g′(x); differentiate the outer function first (leaving the inner one as is), then multiply by the derivative of the inner one.
Profit u(q) = q · (12 − q − r): f = q, g = 12 − q − r. u′ = 1 · (12 − q − r) + q · (−1) = 12 − 2q − r; expanding the product first gives the same result.
Chain: [ln(1 + 2x)]′ = 2/(1 + 2x); (e−x)′ = −e−x; [(1 − p)²]′ = 2(1 − p) · (−1) = −2(1 − p).
Exponents, Logarithms and Derivative Rules
Partial Derivatives: Hold the Others Fixed
If a function depends on several variables (such as u(q, r)), ∂u/∂q (partial derivative) is the derivative with respect to q alone, treating r as a constant. In a game each player chooses only its own variable, so it sets the partial derivative with respect to that variable to zero.
u1(q, r) = q(12 − q − r) = 12q − q² − qr: ∂u1/∂q = 12 − 2q − r, ∂u1/∂r = −q.
The rival's profit is u2 = r(12 − q − r): ∂u2/∂r = 12 − q − 2r. Both zero: 12 − 2q − r = 0 and 12 − q − 2r = 0 ⇒ q = r = 4.
Probability and Expected Value
Probability by Counting
If outcomes are equally likely, P(event) = favorable outcomes / all outcomes. For independent events (rolls that do not affect each other) probabilities multiply.
Even number on a die: {2, 4, 6} → 3/6 = 1/2.
Two heads in two coin tosses: 1/2 · 1/2 = 1/4.
At least one 6 with two dice: use the complement. No 6 at all has probability (5/6)·(5/6) = 25/36, so 1 − 25/36 = 11/36. The table has 11 orange cells.
Probability and Expected Value
Probability Distributions
A probability distribution is a table assigning a probability to each possible outcome. Every probability lies between 0 and 1, and they sum to 1.
There are 36 equally likely outcomes. The pairs summing to 7 are (1,6), (2,5), (3,4), (4,3), (5,2), (6,1): P(7) = 6/36. A sum of 2 only comes from (1,1): P(2) = 1/36.
Check: 1 + 2 + 3 + 4 + 5 + 6 + 5 + 4 + 3 + 2 + 1 = 36, so the total probability is 36/36 = 1.
Probability and Expected Value
Conditional Probability: Information Changes Probabilities
Once we learn that event B occurred, only the outcomes in B remain. A's share of this smaller world is the conditional probability:
P(A | B) = P(A ∩ B) / P(B)
A ∩ B: both A and B occur.
The light picks one of three states with probability 1/3 each and shows every driver only their own signal. Ayşe sees "Stop". P(Ayşe Stop) = 2/3, P(Ayşe Stop ∩ Bora Go) = 1/3.
P(Bora Go | Ayşe Stop) = (1/3) / (2/3) = 1/2.
Probability and Expected Value
Bayes' Rule: Updating Beliefs after an Observation
We start with a belief (the prior, P(A)); after seeing a signal we update it (the posterior, P(A | B)):
P(A | B) = P(B | A) · P(A) / P(B)
20% of players are aggressive. An aggressive player bets big with probability 0.8, a calm one with probability 0.2. We see a big bet.
P(big) = 0.8 · 0.2 + 0.2 · 0.8 = 0.32 ⇒ P(aggressive | big) = 0.16 / 0.32 = 1/2.
Counting with 100 players gives the same answer: 16 of the 32 big bettors are aggressive.
Probability and Expected Value
Expected Value
X is a number that depends on chance (e.g. the face of a die). Its expected value is the average per round X would have if the experiment were repeated many times: multiply each value x of X by its probability and add.
E[X] = Σx x · P(X = x)
E[X] = 1·(1/6) + 2·(1/6) + 3·(1/6) + 4·(1/6) + 5·(1/6) + 6·(1/6)
= 21/6 = 3.5
Probability and Expected Value
Trace: Expected Value of a Lottery
prize = [100, 20, 0] # TL prob = [0.05, 0.25, 0.70] E = 0 for i in range(3): E = E + prize[i] * prob[i] print("E =", E) print("net =", E - 12) # ticket 12 TL
Three prizes and their probabilities; total probability 0.05 + 0.25 + 0.70 = 1. Accumulator E = 0.
Add 100 · 0.05 = 5: E = 5.
Add 20 · 0.25 = 5: E = 10.
0 · 0.70 = 0: a losing ticket does not change E, but its probability still counts.
Expected prize per ticket: 10 TL.
The ticket costs 12 TL: net expected payoff 10 − 12 = −2 TL. Someone who plays many times loses 2 TL per ticket on average.
Probability and Expected Value
Example: Is This Dice Game Worth Playing?
A fairground game: you pay 10 TL and roll a die. A 6 pays 40 TL, a 5 pays 10 TL; any other number pays nothing.
| Die | Prize x | P(x) | x · P(x) |
|---|---|---|---|
| 6 | 40 | 1/6 | 40/6 |
| 5 | 10 | 1/6 | 10/6 |
| 1, 2, 3, 4 | 0 | 4/6 | 0 |
| Total | 6/6 = 1 | 50/6 ≈ 8.33 |
Expected prize ≈ 8.33 TL against a 10 TL fee: net ≈ −1.67 TL per game. Ignoring risk, playing does not pay.
Let the prize for a 6 be y: (y + 10)/6 = 10 ⇒ y = 50 TL. This is an equation in one unknown.
Probability and Expected Value
The Continuous Uniform Distribution U[0, 1]
X ~ U[0, 1] reads "X is uniformly distributed between 0 and 1": every value in the interval is equally likely. The density f(x) = 1 is the height of the graph; probability is the area under it. With infinitely many values, a single point has probability 0; an interval has probability equal to its length: P(a ≤ X ≤ b) = b − a, in particular P(X ≤ x) = x.
A rival's value is U[0, 1]. A bid of 0.6 beats it with probability P(X ≤ 0.6) = 0.6.
Two independent rivals: beating both has probability 0.6 · 0.6 = 0.36.
P(0.2 ≤ X ≤ 0.5) = 0.3. The mean is E[X] = 1/2.
Probability and Expected Value
Integrals: Area under a Curve
∫ab f(x) dx is the area under the graph of f between a and b: the continuous version of Σ. Cut the area into thin rectangles and add them up; as they get thinner, the sum approaches the integral.
∫01 x dx = area of the triangle with base 1 and height 1 = 1/2. In general ∫0b x dx = b²/2.
For U[0, 1], probability and expectation are areas: P(X ≤ b) = ∫0b 1 dx = b, E[X] = ∫01 x · 1 dx = 1/2.
Probability and Expected Value
Expected Value and Tail Probabilities
P(X > t), the probability that X exceeds t, is a tail probability. If X ≥ 0, the expected value is the area under the tail probability:
E[X] = ∫0∞ P(X > t) dt
X ~ U[0, 1]: P(X > t) = 1 − t (0 ≤ t ≤ 1), and 0 for t > 1. E[X] = ∫01 (1 − t) dt = area of the triangle with base 1 and height 1 = 1/2 ✓
For integer X the sum form is E[X] = P(X ≥ 1) + P(X ≥ 2) + … For a die, 6/6 + 5/6 + … + 1/6 = 21/6 = 3.5 ✓
Mixed Strategies
The Idea of a Mixed Strategy
A mixed strategy (randomized strategy) is a probability distribution a player places over its own actions: "let a die or a coin decide my move."
Roll a die: 1–2 → Rock, 3–4 → Paper, 5–6 → Scissors. This is the mixed strategy (1/3, 1/3, 1/3).
Toss a coin: heads, Rock; tails, toss again for Paper or Scissors: (1/2, 1/4, 1/4).
Playing one action for sure is a pure strategy: (1, 0, 0) = always Rock.
Mixed Strategies
Expected Payoff against a Mixed Strategy
Bora plays (1/2 Rock, 1/4 Paper, 1/4 Scissors). The expected payoff of each of Ayşe's pure moves: multiply the row by Bora's probabilities and add.
| Ayşe ↓ / Bora → | Rock (1/2) | Paper (1/4) | Scissors (1/4) | Expected payoff |
|---|---|---|---|---|
| Rock | 0 | −1 | +1 | 0 − 1/4 + 1/4 = 0 |
| Paper | +1 | 0 | −1 | 1/2 + 0 − 1/4 = 1/4 |
| Scissors | −1 | +1 | 0 | −1/2 + 1/4 + 0 = −1/4 |
argmax = Paper: against a Rock-heavy opponent, Paper earns 1/4 point per round on average. Expected value + argmax together give the "best response".
Mixed Strategies
Exponential Weights: Logit and Softmax
A smooth way to assign probabilities to options with values v1, …, vk is softmax: P(i) = eλvi / Σj eλvj. With two options it is called logit: P(1) = 1 / (1 + e−λ(v1 − v2)). λ ≥ 0 is the sensitivity: at λ = 0 all are equal; as λ grows, probability piles up on the argmax.
Expected payoffs Rock 0, Paper 1/4, Scissors −1/4 and λ = 4: exponents 0, 1, −1.
e0 = 1, e1 ≈ 2.718, e−1 ≈ 0.368; sum ≈ 4.086. P ≈ (0.24, 0.67, 0.09).
Payoff Matrices
Reading a Payoff Matrix
The row player picks a row, the column player a column; the cell where they meet is the outcome. A cell (a, b): a is the row player's payoff, b the column player's.
Ayşe holds a red or a black card; Bora guesses the color. If Bora guesses red correctly he takes 3 chips from Ayşe, if black, 1 chip; a wrong guess costs him 1 chip. (Red, Red) = (−3, 3): Ayşe held red and Bora guessed it; Ayşe pays 3, Bora gets 3. In every cell a + b = 0: the game is zero-sum.
Payoff Matrices
Practice: Reading the Matrix
- Ayşe holds Black and Bora says "Red". Who gets how many chips?
- When Ayşe holds Black, which payoffs can Bora get?
- If Bora says "Red" every round, which card should Ayşe hold?
- What are Ayşe's largest and smallest possible payoffs?
Show answers
- (Black, Red) = (1, −1): Bora guessed wrong; Ayşe gets 1 chip, Bora pays 1.
- Second entries in the Black row: −1 (if he says "Red") and 1 (if he says "Black").
- Ayşe's numbers in the "Red" column: Red −3, Black 1. 1 > −3 ⇒ Black.
- Ayşe's numbers (first entries): −3, 1, 1, −1. Largest 1, smallest −3.
Payoff Matrices
Expected Payoff Is a Matrix–Vector Product: xᵀAy
A mixed strategy is a probability vector: x = (x1, …, xm), each xi ≥ 0, summing to 1. Let A be the row player's payoff matrix. Ay: multiply each row of A by y and add; each resulting number is that row's expected payoff against y. xᵀAy: weight these row values by the probabilities in x and add. (ᵀ means x is written as a row.)
A = [[−3, 1], [1, −1]], Ayşe x = (1/2, 1/2), Bora y = (1/4, 3/4).
Ay = (−3/4 + 3/4, 1/4 − 3/4) = (0, −1/2); xᵀAy = ½ · 0 + ½ · (−½) = −1/4.
Best Response
Best Response
The answer to "what would I do if I knew what my opponent will do?" is a best response (BR). Fix the opponent's choice, look at your own payoffs, take the argmax.
Bora always says "Red": Ayşe's first entries in that column are Red −3, Black 1 ⇒ best response Black.
Bora always says "Black": Red 1, Black −1 ⇒ best response Red.
If Bora mixes: compute each row's expected payoff and pick the larger.
Best Response · Preview
Trace: Best Responses in a 2×2 Matrix
Ayşe picks a row (Up/Down), Bora a column (Left/Right). Nothing marked yet.
If Bora plays Left, Ayşe compares the first entries: Up 1, Down 2 ⇒ Down; mark 2.
If Bora plays Right: Up 3, Down 0 ⇒ Up; mark 3.
If Ayşe plays Up, Bora compares the second entries: Left 2, Right 1 ⇒ Left.
If Ayşe plays Down: Left 3, Right 0 ⇒ Left. Left is better for Bora either way.
The only cell with both entries marked is (Down, Left) = (2, 3): both play best responses, so neither wants to deviate alone. This is a Nash equilibrium.
Equations and Indifference
Quadratic Equations and the Discriminant
ax² + bx + c = 0 (a ≠ 0). The discriminant Δ = b² − 4ac gives the number of roots: two if Δ > 0, one if Δ = 0, no real root if Δ < 0. The roots are:
x = (−b ± √Δ) / (2a)
For a probability p we got 2p² + p − 1 = 0. Δ = 1² − 4 · 2 · (−1) = 9, √Δ = 3.
p = (−1 ± 3)/4 ⇒ p = 1/2 or p = −1. A probability lies between 0 and 1: p = 1/2. Check: 2 · 1/4 + 1/2 − 1 = 0 ✓
Equations and Indifference
Two Linear Equations in Two Unknowns
Two unknowns need two equations: eliminate one, solve for the other, substitute back.
(1) 2z + d = 50 (2) z + d = 35 z: die, d: deck
(1) − (2): (2z + d) − (z + d) = 50 − 35 ⇒ z = 15
Substitute into (2): 15 + d = 35 ⇒ d = 20
Check (1): 2 · 15 + 20 = 50 ✓
Equations and Indifference
The Idea of Indifference
If a player genuinely mixes between two actions, both must give her the same expected payoff; if one were better, she would always play it. This is indifference.
If Bora says "Red" with probability q, Ayşe gets E[R] = −3q + 1 · (1 − q) = 1 − 4q and E[B] = 1 · q − 1 · (1 − q) = 2q − 1.
q < 1/3: Red is better. q > 1/3: Black is better. Only at q = 1/3 is Ayşe indifferent.
Equations and Indifference · Preview
Trace: The Indifference Condition
p: probability Ayşe holds Red. q: probability Bora says "Red".
Let Bora say "Red" with probability q and "Black" with 1 − q (total 1).
Ayşe holds Red: E[R] = −3 · q + 1 · (1 − q) = 1 − 4q
Ayşe holds Black: E[B] = 1 · q + (−1) · (1 − q) = 2q − 1
Indifference: 1 − 4q = 2q − 1 ⇒ 2 = 6q ⇒ q = 1/3
Check: E[R] = 1 − 4/3 = −1/3 and E[B] = 2/3 − 1 = −1/3 ✓ With q = 1/3, Ayşe gets the same expected payoff whichever card she holds.
Equations and Indifference
Back to the Game: Rock-Paper-Scissors
Yes: (1/3, 1/3, 1/3). Whatever the opponent plays, the expected payoff is 0; e.g. against Paper: 1/3 · (−1) + 1/3 · 0 + 1/3 · (+1) = 0. Every reply earns the same, so there is nothing to exploit.
The opponent's best response is always Paper: −1 per round. Even a milder bias loses: playing (1/2, 1/4, 1/4) earns 1/2 · (−1) + 1/4 · 0 + 1/4 · (+1) = −1/4 against Paper.
| Your mix (R, P, S) | Opponent's best response | Your expected payoff |
|---|---|---|
| (1, 0, 0): always Rock | Paper | −1 |
| (1/2, 1/4, 1/4) | Paper | −1/4 |
| (1/3, 1/3, 1/3) | all equal | 0 |
Game Trees
Reading a Game Tree
Sequential games use a game tree: a circled node shows whose turn it is, branches are actions, and the leaves at the bottom give the (Ayşe, Bora) payoffs when the game ends.
Ayşe hides or shows her card. Bora, seeing what Ayşe did, plays on or folds.
Path Hide → Fold: leaf (3, 0); Ayşe gets 3, Bora 0.
Bora has two nodes and decides separately at each.
Game Trees
Trace: Backward Induction
Idea: start with the player who moves last. Once Bora's choice at each node is known, Ayşe's decision is easy.
After Hide, Bora compares the second entries: Play 2, Fold 0 ⇒ Play. This node is worth (1, 2).
After Show, Bora: Play 1, Fold 3 ⇒ Fold. This node is worth (2, 3).
At the root, Ayşe anticipates Bora's replies: Hide → 1, Show → 2 ⇒ Show.
The resulting path is Show → Fold with payoff (2, 3). Leaf (3, 0) is Ayşe's largest number, but Bora never goes there.
Game Trees
Growth Rates: O and Θ Notation
How does an algorithm's step count grow with the input size n? f(n) = O(g(n)): for large n, f stays below a constant multiple of g (an upper bound). f(n) = Θ(g(n)): f grows at the same rate as g (squeezed between constant multiples of g). Constant factors and lower-order terms are dropped: 3n² + 5n = Θ(n²).
Order: 1 < log log n < log n < n < n log n < n² < 2n.
If every node of a game tree has 2 branches and the depth is d, there are 2d leaves: backward induction takes Θ(2d) steps, about 109 for d = 30.
| n | log₂ log₂ n | log₂ n | n² | 2n |
|---|---|---|---|---|
| 16 | 2 | 4 | 256 | 65,536 |
| 256 | 3 | 8 | 65,536 | ≈ 1077 |
| 65,536 | 4 | 16 | ≈ 4.3 · 109 | huge |
Summary
Common Mistakes
| Topic | Mistake | Correct |
|---|---|---|
| Profile | (Rock, Paper) = (Paper, Rock) | Order matters: Ayşe first, then Bora |
| Σ | (Σ xi)² = Σ xi² | Evaluate the terms first, then add |
| argmax | "argmax = 7" | argmax is an option (B); max is a value (7) |
| Probability | At least one 6: 1/6 + 1/6 | 1 − (5/6)² = 11/36 |
| Expected value | "E = 3.5 will come up" | Long-run average; a single roll may never show it |
| Mixed strategy | Cycling through moves | One random pure move each round |
| Matrix | Giving b to the row player | (a, b): a is the row player's, b the column player's |
| Best response | Looking at the opponent's numbers | Fix the opponent's choice, compare your own numbers |
| Indifference | Getting p from Ayşe's condition | p comes from the condition that makes Bora indifferent |
| Game tree | Starting at the root | From leaves to root: last move first |
Summary
Summary
S1 × S2: all profiles; ui: i's payoff.
Σ: evaluate, add. max: a value; argmax: the option giving it.
xa · xb = xa+b, x0 = 1; ln inverts ex; (ln x)′ = 1/x.
(fg)′ = f′g + fg′; chain: outer × inner; ∂u/∂q: others fixed.
Count and divide; multiply if independent; P(A | B) = P(A ∩ B)/P(B).
E[X] = Σ x · P(x); if X ≥ 0, E[X] = ∫ P(X > t) dt.
Distribution over actions; softmax: P(i) ∝ eλvi.
(a, b): row payoff a, column payoff b.
Fix the opponent; argmax of your payoff.
x = (−b ± √Δ)/(2a); indifference: actions played in equilibrium pay the same.
Leaves to root; the mover picks the best.
log log n < log n < n < n² < 2n; drop constants.
Summary
Test Yourself
- Two coins are tossed. What is the probability of at least one head?
- A lottery pays 50 TL with probability 0.2 and 0 TL with probability 0.8; a ticket costs 8 TL. What is the net expected payoff?
- Payoffs: Rock 2, Paper −1, Scissors 2. What are max and argmax?
- Bora plays (1/4 Rock, 1/2 Paper, 1/4 Scissors). What is Ayşe's best response?
- If x + y = 1 and 3x = y, what are x and y?
- A die is known to show an even number. What is the probability it is a 6?
- At which q does f(q) = q(6 − q) reach its maximum?
Show answers
- 1 − P(two tails) = 1 − 1/4 = 3/4.
- E = 50 · 0.2 = 10 TL; net 10 − 8 = +2 TL.
- max = 2; argmax = {Rock, Scissors} (a tie: two options).
- E[Rock] = −1/2 + 1/4 = −1/4; E[Paper] = 1/4 − 1/4 = 0; E[Scissors] = −1/4 + 1/2 = 1/4 ⇒ Scissors.
- x + 3x = 1 ⇒ x = 1/4, y = 3/4.
- P(6 | even) = (1/6)/(1/2) = 1/3.
- f′(q) = 6 − 2q = 0 ⇒ q = 3, f(3) = 9.
The End
Next: Chapter 1: Decision Theory