GAME THEORY

JENNIFER YUAN AND SHREYA SINHA

Abstract. This paper is an exploration of game theory. We will focus on combinatorial

games and the strategies players may use within the game. In section 2, we will define
concepts essential to our study of combinatorial games. We will then explore normal-play
games and broad generalizations we can make of all such games in section 3. In section 4,
we explore a game called Nim and the Sprauge-Grundy theorem, which allows us to explore
an important property of impartial games.

1. Combinatorial Games
1.1. Combinatorial Games.
Definition 1.1. A combinatorial game is a game played between 2-players, Louise and
Richard. The game consists of the following:
(1) A set of possible positions. These are the states of the game (e.g. the starting board
of a chess game).
(2) A move rule indicating what positions Louise can move to and what positions Richard
can move to on their turn.
(3) A win rule indicating a set of terminal positions where the game ends. Each terminal
position has an associated outcome, either Louise wins and Richard loses (denoted
+-), Louise loses and Richard wins (-+), or the game is a draw (00).
To play a combinatorial game, a starting position is chosen with a designated player mak-
ing the first move. The starting position indicates how the game will look like before it is
played. Then, the two players alternate in making moves until a terminal position is reached,
indicating the end of the game. A terminal position is reached when one of the two players
has won the game or when no possible moves can be made and the game ends with a draw.
Combinatorial games may range from simple games such as Tic-Tac-Toe to more compli-
cated ones such as Chess. However, these games have no element of randomness, so games
such as Monopoly that require the use of a dice or a spinner are not considered to be com-
binatorial games. Combinatorial games thus rely on the player’s strategy and moves to win,
whereas games with probabilistic elements depends on chance and not wholly in the player’s
strategy.
Example. A common combinatorial game is Pick-Up-Bricks. In this game, a pile of bricks
is placed in the center. Each move consists of a player removing 1 or 2 bricks from the pile.
The game ends when the pile is empty, and the last player to take a brick wins.

1.2. Game Trees. A game tree is a helpful tool to diagram combinatorial games. A branch
is a line that connects two nodes, and each branch represents a player’s move. The point
that branches connect are called nodes, and each models a new state of the game. Terminal
nodes - ones that only have one branch connecting it to another node - indicate the possible
outcomes.
1
2 JENNIFER YUAN AND SHREYA SINHA

Figure 1. Pick-Up-Bricks

The topmost node is called the root node, and represents the beginning state of the game.
Each branch node is annotated with an L or an R to indicate whose turn it is to play, with L
for Louise and R for Richard.

Figure 2. Game Tree Of A Tic-Tac-Toe Game Row

The depth of a game tree is defined as the maximum number of possible moves from the
start to the end of the game (i.e. the longest path from the root node to a terminal node).
The path that leads to decisive victory for a player may be called a winning strategy, and
ones that lead to a draw outcome is called a drawing strategy.s
The power of game tree lies in the fact that it allows us to diagram all possible combinatorial
games regardless of how they are played or any other detail specific to them. This allows us
to explore broad properties of games, as we will see in the next section.
1.3. Zermelo’s Theorem. An important theorem relating to combinatorial games is Zer-
melo’s Theorem, which states that in all such games, either Louise has a winning strategy,
Richard has a winning strategy, or both players have a strategy that guarantees a draw. There
are no combinatorial games in which one player doesn’t have a winning strategy or the game
cannot end in a draw definitively.
The power of this theorem is not very intuitive, because it means there exists a winning or
drawing strategy in games like chess, where people play professionally. Of course, the reason
chess is still played is because there are so many possible positions a player can reach within
the game that we have no way of documenting all of them, and no way of finding that strategy
at the moment.
 GAME THEORY 3

Theorem 1.2 (Zermelo’s Theorem). In all combinatorial games, at exactly one player has a
winning strategy or both players have a draw strategy.
Proof of Zermelo’s Theorem. We will prove this by induction. For the base case of depth 0,
the game’s outcome is already decided as a win for Louise (+-), a win for Richard (-+), or a
draw (00).
For the inductive step, assume for a tree of depth n > 0 that this assumption holds for
all trees with smaller depth. Let N1 , N2 , . . . N` denote the nodes that the starting player can
reach in one move. For 1 ≤ i ≤ ` let Ti be the subtree with root node Ni . Each game Ti
where 1 ≤ i ≤ ` must thus be type +-, -+,or 00 as per the assumption. We can thus label
each game path as strategies Li for Louise and Ri for Richard such that either one of these is
winning or both are drawing. A strategy L thus exists for Louise by combining all individual
strategies L1 , ..., L` , and the same for Richard denoted R by combining R. Assume Richard
is the first to move. We then must consider the following cases:
(1) Case 1: At least one of T1 , T2 , ..., T` is type -+. Let Ti be type -+; Richard thus has
a strategy Ri from position Ni , so the winning strategy is to have Richard move to
position Ni from the starting position R.
(2) Case 2: All of T1 , ..., T` are type -+. Here, every L1 , ..., L` is a winning strategy for
Louise, so L is a winning strategy for her.
(3) Case 3: None of T1 , ..., T` is -+, but at least one Ti is 00. This means that Richard’s
best strategy is Ri because he will not lose, and draws instead. Thus, R is a drawing
strategy and Richard must move to position Ni to play by it. In addition, each
L1 , ..., L` must then be a drawing or winning strategy, so L is a drawing strategy for
Louise.
When Louise is the first to move, the parallel argument holds. 

2. Normal-Play Games
Some combinatorial games can be classified into more specific types. One such is a normal
play game, where the win rule indicates that the winner of the game is the last player to make
a move (i.e. there cannot be a draw). Some normal play games can be categorized further as
impartial if the set of moves that either player can make is the same, such as in the Pick-up
Bricks game. If a game is not impartial, it is partizan, meaning that the player’s respective
set of moves are not the same. For example, chess is partizan because each player has their
respective set of pieces that they may move.
2.1. Positions and Their Types. In this section, we will define new notation to represent
positions, classify and describe the different types of positions, and figure out how to determine
the type of a position in order to explore more general properties of games.
Positions are denoted as α, β, and γ, where each notation represents a state of a game.
The purpose of position notation is to emphasize the relation between the current state, and
the possible future states of the game. Thus, for a possible game, its position notation may
look like this: γ = {α1 , α2 , ..., αn |β1 , β2 , ..., βn }, where γ is the current state of the game, αi
represents all the possible positions Louise may bring the game to if it her turn using her move
rules (γ → αi ), and βi represents all the possible positions Richard may move to (γ → βi ).
Each position can be classified into specific types. Defining types and their properties will
make it easier to predict the outcomes of several simultaneous games at once. There are four
potential types of a game:
(1) L : Louise has a winning strategy from this position regardless of whose turn it is.
4 JENNIFER YUAN AND SHREYA SINHA

(2) R: Richard has a winning strategy from this position regardless of whose turn it is.
(3) N: The current player whose turn it is has the winning strategy.
(4) P: The player whose turn it is not has the winning strategy.
This also allows us to analyze what type of position the game is by analyzing what kind of
positions Richard and Louise can move to:

Figure 3. Determining the Current Type Using the Types of Possible Sub-
sequent Positions

In this paper, we will be focused on impartial games. In these games, each position is
always type N or P.
2.2. Sum of Positions. In a game of Pick-Up Bricks, the pile of bricks may be broken down
into many different-sized piles over the course of the game. If we call each piece a component
of the game, then we can view a Pick-Up Bricks position as a sum of these components.
Definition 2.1. If α and β are positions in normal-play games, then we define α + β as the
starting position in a combined game where game α and game β are played simultaneously.
To move in the game α + β, a player chooses to make a move in either game α or game β.
For example, if a player makes a move in game α, then the player is moving from α to α0 so
the new position of the game may be described as α0 + β. Similarly, if a player makes a move
in game β, then the player is moving from β to β 0 so the new position may be described as α
+ β0.

Figure 4. Example Game Tree of Combined Game α + β

If β is type P, then α and α + β are the same type. To understand the reasoning behind
this, let’s look at an example. Suppose position β is type P and position α is type L. In a case
 GAME THEORY 5

where it is Louise’s turn, she can start the game by making a move in game α. If Richard also
makes a move in game α, then Louise can continue to make a move in game α. If Richard
makes a move in game β, then Louise will proceed to make a move in game β as well. Since
game α is type L, Richard will eventually run out of moves in game α and will be forced to
make the first move in game β. This means that Louise will always have a strategy to win
game β because she will always be the player moving second in the game making α + β also
type L. The same strategy can be followed for a game where Richard makes the first move.
If α and β are both type L, then α + β is also type L. Similarly, if α and β are both type
R, then α + β is also type R. The reason for this is because if both game α and β were type
L or type R, it means that no matter what the other person does in either of the two games,
he or she will still not be able to win either game α or game β. Therefore, α + β would be
of the same type.

Figure 5. Types of sum

How do we determine the type of a game? Looking at a game of Domineering may help.
Domineering is a normal-play game played using some squares from a rectangle array. On
Louise’s turn, she may place a 2 x 1 domino over two unoccupied squares. On Richard’s turn,
he may place a 1 x 2 domino over two unoccupied squares. Since this is a normal-play game,
the last player to move wins.

Figure 6. Domineering

A easy way to look at a game of domineering is to remove any unusable components after a
player makes a move. For example, if Louise was the first player and chooses the first option
as shown in Figure 6, then the two squares in the center will be removed. The game would
be left with one 1 x 1 piece and one 2 x 1 piece. Since no move can be made with the 1 x 1
6 JENNIFER YUAN AND SHREYA SINHA

piece, the piece may be removed as well. The result would be a new position with only one 2
x 1 piece remaining. The remaining 2 x 1 piece is a type L, meaning that Louise will win the
game.
2.3. Equivalence of Positions.
Definition 2.2 (Equivalence). We say that position α and β in (possibly two different)
normal-play games are equivalent if they have the same type for any position γ added. In
other words, if α + γ and β + γ has the same type for any position of γ, then α ≡ β
There are some fundamental properties of equivalence that are useful to know. If α, β, and
γ are positions in normal-play games, then:
(1) α ≡ α (reflexivity),
(2) α ≡ β implies that β ≡ α (symmetry),
(3) α ≡ β and β ≡ γ implies that α ≡ γ (transitivity)
(4) α + β ≡ β + α (communtativity)
(5) (α + β) + γ ≡ α + (β + γ) (associativity)
It is also important to note that if α ≡ α’, then α and α’ have the same type.

However, when two positions have the same type, it does not mean that the two positions
are equivalent.

Figure 7. Two sums in Domineering with different types

When determining if two (or more) positions are equivalent, type P serves as the number
zero. In other words, if β is type P, then α + β ≡ α. The reasoning behind this is explained
in the sum of positions. Any position added with a position of type P will result in the sum
of the two positions to be the same as the first position. So, if β is type P, then α + β is the
same type as α. Two positions are equivalent when they have the same type when a third
position is added, so α + β ≡ α.

3. Nim and the Sprauge-Grundy Theorem

Nim is an impartial game in which positions consist of ` piles of stones sizes α1 , α2 , ..., α` .
A player moves by choosing a pile and removing any positive number of stones from that pile.
The winner is the person who picked up the last stone.
The notation ∗a represents a Nim position consisting of one pile of a stones. We call ∗a a
nimber. A game with piles of sizes a1 , a2 , ...a` can be written as ∗a1 + ∗a2 + ... + ∗a` .
 GAME THEORY 7

Figure 8. A Position in Nim

Each pile in a game of Nim can be broken into subpiles based on the binary expansion of
the number of stones in the pile.

Figure 9. Subgroups Based on Binary Expansion

For example, in the figure above, ∗11 is broken into subgroups of size 1, 2, and 8 because
the binary expansion of 11 is 1+2+8 = 11. This allows us to introduce a new property of the
game Nim: balance. Each position ∗a1 + ∗a2 , ... + ∗a` is balanced if, for every power of 2, the
total number of subgroups of that size across the entire game is even. If this is not true of
the position, then it is an unbalanced position.
The ability to get from an unbalanced position to a balanced one allows us to find a winning
strategy.
Theorem 3.1 (Balancing Nim). A player who moves on an unbalanced position can always
move it back to a balanced one.
Proof. Consider an unbalanced game ∗a1 + ∗a2 , ... + ∗a` , and assume that 2m is the largest
power of 2 for which there are an odd number of subgroups. Suppose that a pile ∗ai is a pile
with a subpile size 2m . Pick up all the subgroups of ∗ai that are less than or equal to 2m in
size. This will be at most 2m stones, because in the worst case of picking up piles of all sizes
less than 2m , that is still 20 + 21 + ... + 2m−1 = 2m − 1 stones. For all the subpiles of size 2j
such that j < m, if the number of subpiles is now odd, put the 2j stones initially picked up
back in ∗ai to leave it even. This leaves the final position balanced. 
Proposition 3.2 (Types of Nim Game Positions). Another important property of Nim games
is that each balanced Nim position is type P and each unbalanced Nim position is type N.

Proof. Assume a balanced game ∗a1 + ∗a2 , ... + ∗a` . Any move the player makes leaves the
game unbalanced, because that player only has access to one pile, and thus only one subpile
of the sizes in that pile. The second player may use the balancing procedure to re-balance the
game. This cycle continues, and the end position of the game - one where there are no more
8 JENNIFER YUAN AND SHREYA SINHA

Figure 10. Position Types of Nim

stones - is a balanced one, so the second player has the winning strategy in this case, and it
is type P. If the game starts unbalanced, the first player has the ability to balance it and thus
has a winning strategy, making it type N. 
The idea of balancing the game also allows us to think about the idea of equivalent positions
in the game. As we found earlier, any type P positions of a game are equivalent. Since ∗0
is type P, it follows that any balanced game is equivalent to ∗0. This allows us to simplify
games down to single nimbers.
After learning that every Nim position is equivalent to a nimber, it is actually true that
every position in any impartial game is equivalent to a nimber! This is known as the Sprague-
Grundy Theorem.
To begin, let’s define what MEX is.
Definition 3.3 (MEX Value). For a set S = {α1 , α2 , ..., αn } of nonnegative integers, we
define the Minimal Excluded value, abbreviated MEX, of S to be the smallest nonnegative
integer b which is not one of {α1 ...αn }. For instance, the MEX of the set {0,1,2,5,8} is 3.
This leads us to the MEX Principle:
Theorem 3.4 (MEX Theorem). For a position α = {α1 , α2 , ..., αk }, if αi ≡ ∗ai for each
1 ≤ i ≤ k, then α ≡ ∗b when b is the MEX of {a1 , a2 , ..., ak }.
Proof. Assume a combined position of α + ∗b.

Figure 11. Visualization of the Combined Positions α and *b

Suppose the first player moves ∗b to ∗b0 where b0 < b. By definition of a MEX value b, there
must be some ai = b0 where 1 ≤ i ≤ k. The second player can change α to αi so that the
position of the entire game is ∗b0 + αi ≡ ∗b0 + ∗ai ≡ 0. Thus, the second player has a winning
strategy.
If, however, the first player moves α to αj , then because αj 6= b (by the definition of a MEX
value), this position is equivalent to ∗(b ⊕ αj ) where ∗(b ⊕ αj ) 6≡ ∗0. The second player may
 GAME THEORY 9

then change ∗b to ∗b0 such that b0 = aj , so that ∗b0 + αj ≡ ∗b0 + ∗aj ≡ ∗0. The second player
again has a winning strategy, so α + ∗b must be type P.
Then, because α + ∗b is type P, α + ∗b ≡ ∗0. Adding ∗b to both sides gets α + ∗b + ∗b ≡ ∗b,
and because ∗b + ∗b must be equivalent to ∗0, α ≡ ∗b. 
We can use this to prove the Sprauge-Grundy theorem by induction based on the depth of
a game.
Theorem 3.5 (Sprauge-Grundy Theorem). Every position in any impartial game is equiva-
lent to a nimber.
Proof of the Sprauge-Grundy Theorem. Assume α to be a impartial game position. The base
case is when α has depth d = 0, so α ≡ ∗0. For the inductive step, assume d > 0 and
the theorem holds for all positions with depth less than d. Since α = {α1 , α2 , ..., α` }, then
each αi has depth less than d and is equivalent to some ∗ai . Thus, α = {α1 , α2 , ..., α` } ≡
∗a1 , ∗a2 , ..., ∗a` . Assume b to be the MEX of {a1 , a2 , ..., a` }. Then α = {α1 , α2 , ..., α` } ≡
{∗a1 , ∗a2 , ..., ∗a` } ≡ ∗b. 

References
[1] DeVos, M., Kent, D. Game Theory: A Playful Introduction. American Mathematical Society, 2016.