UNIVERSITY OF UTAH

Solving Cordial Minuet

A sketch of a game theory solution
Curtis Miller

Cordial Minuet is a gambling game played on a magic square of order 6 with two
players, with no random elements beyond the creation of the game board. In this
paper, I find an approximate solution to a version of the game played on a magic
square of order 4 using game theory techniques and fictitious play solution
approximation. The smaller version of the game likely has a value very close to
zero, and optimal strategies usually involve trying to guarantee at least a tie.

Curtis Miller
MATH 5750
September 8, y
Project

Solving Cordial Minuet

A sketch of a game theory solution
Last year, a blogger named Stephen Totilo [2] published an article on the
gaming website Kotaku about a multiplayer game in development called
Cordial Minuet by a game designer named Jason Rohrer. It is a gambling
game but with no element of chance beyond the initial creation of the game
board (and also the game employs an aesthetic referencing the occult, but
this has no impact on gameplay; notice that Cordial Minuet is an anagram for
"demonic ritual"). The game has strategic depth, but it appears to be a game
that could easily fit into a framework allowing for game theory analysis.
Every game is played on a magic square of order 6; the sum of each
row, column, and the two diagonals is 111, and the sum of all the numbers
on the game board is 666 (which fits nicely with the game's occult themes).
A new magic square is generated every game. (The number of magic

MATH 5750 PROJECT

squares of order 6 are surely finite and less than 36!, though I do not know
the method being used to generate the magic squares and how many magic
squares are possible with it. The only method I am aware of for generating
magic squares is the medjig method [3], but I do not know the number of
magic squares this method could generate.)
As the game is still in development, the rules of the game may change,
but I will use the rules described in [1] and [2]. The game has a minimum
pay-in of $2. The game was designed so that the player sees himself as the
column player, but I will stick with game theory convention and treat the row
player as player one (PI) and the column player as player two (PII). Players
do not make specific bets but instead wager a percentage of their pay-in.
The players start by betting 1% of their pay-in. (Opponents are paired at
random and anonymously, with opponents having similar pay-ins being more
likely to be paired, though they are not guaranteed to have the same pay-in;
this is not revealed to the players.)
The game is played in three rounds. In each round, PI will select a row
for himself and a row for his opponent. He can choose any combination of
rows so long as neither row was selected for either himself or his opponent in
a previous round. PII will do the same for the columns. The value of the
square where the row PI picked for himself and the column PII assigned to
him intersect will be added to PI's score. The value of the square where the
column PII picked for herself and the row PI assigned to her intersect will
likewise be added to PII's score. At the end of each round, the row (column)

MATH 5750 PROJECT

that PI (PII) assigned to PII (PI) will be revealed to PII (PI), so both players will
know their own scores. However, they will not know what the other player's
score is; the other column (or row) that the opponent selected is not
revealed. At the end of each round the players may wager any additional
percentage of their initial pay-in they desire. Betting rules are similar to the
rules of poker, where a player makes a bet, and the other player may call,
raise or fold.
At the very end of the game, after all rows and columns have been
assigned to PI or PII, both players will reveal one of the squares
corresponding to their own payoff to their opponent. A round of betting
follows. Finally each player's score is revealed. The winner at the end of the
game is the one with the highest cumulative score, and he earns the pot. If
there is a tie, each player receives the amount he or she bet (i.e. neither
player profits). (To see examples, both [1] and [2] demonstrate play.)
An important property of the game is that at the end, every row will
have been assigned to either PI or PII, with three rows being assigned to PI
and three rows assigned to PII. The same is true of the columns. So each
player's score will be the sum of three squares on the game board, with no
two squares sharing a row or column. If the row and column of a square is
represented as an ordered pair

( x , y ) , with

winning squares will share the same

a row and

or the same

a column, no

y .

The full game of Cordial Minuet is complex, since there are

6!

ways

for a player to assign his rows or columns, and decisions are made in three

MATH 5750 PROJECT

steps, not to mention the numerous betting strategies. I am not interested in

the betting strategies so much as in how to win the game, so I will ignore the
betting aspect. I will analyze a simpler version of the game, one played on a
magic square of order 4 and in two rounds of play. This preserves the
properties I described in the preceding paragraph and still allows for a large
number of tactics, preserving at least some of the depth that makes the full
version of the game so interesting.
In this reduced version of the game, PI picks a row for himself and a row
for PII, and PII does likewise for the columns. After PI and PII have made their
choices, PI sees the column PII assigned to him, and PII sees the row PI
assigned to her. Then the second round of the game is played. PI decides
which of the two remaining rows will be assigned to himself and which to PII,
and PII does likewise for the columns. After the players have made their
choices, the final scores are revealed, with PI's final score being the sum of
the cells of the square where the rows and columns assigned to him
intersect, and PII's score is found similarly. The player with the highest score
wins.
I begin by describing the game in extensive form. Despite being a
simplified version of Cordial Minuet, the game is still far from simple, and the
game tree is too large to draw. I feel the best way to describe the tree is
through words, so I will describe how the tree would be drawn if we could.
We will use the notation

iPr , jPr

to describe a strategy.

number of the round in which a strategy is picked, and

MATH 5750 PROJECT

is the

is the player

P
i r , is the row (or column) player

picking the strategy. The first coordinate,

P

picks for himself (or herself) in round

P
j r , is the row (or column) player

r , and the second coordinate,

picked for his or her opponent ( i

will always be used for the row or column player

j

picked for himself, and

will always be the row or column designated to the opponent). Notice

that for a player

P , all coordinates are unique; in other words,

i P1 j P1 i2P j P2 . For convenience, I may write

and

i , j instead of

iPr , jPr

will be clear from the context.

For a square on the game board itself, I use the notation

denote the value of the square at row
for PI, the value of the square

(i Ir , j IIr )

and for PII, the value of the square

PI's total score will be
S II ( j 1I ,i 1II ) +( j 2I ,i 2II ) . If

and column

( x , y ) to

y . This means that

will be added to his score in round r,

I

II

( j r ,i r )

S I ( i 1I , j II1 ) +(i 2I , j II2 )

S I > S II , PI wins; if

will be added to her score. So

and PII's total score will be

S I < S II , PII wins. Otherwise, there

is a tie.
We will let PI move first (the game tree will be designed so that it does
not matter who moves first). What choices are available to him? He needs to
select a row for himself and a row for PII, and there are

4 3=12

ways to do

so. We will label each of these pure strategies in round one with unique

i , j , each corresponding to the unique way to pick a row for PI and a row
for PII (PI's information set consists only of the top node).

MATH 5750 PROJECT

PI selects his strategy, and now it is PII's turn. PII does not know what PI
did in round one yet, so she will have only one information set consisting of
all the nodes corresponding to PI's choice in round one. In round one, PII also
has the choice of 12 pure strategies of the form

i , j , so from each node

spawning from PI's choice, PII will have 12 branches, each labeled with a
unique

i , j . There are now 144 nodes, each corresponding to a unique

combination of

iI1 , j1I

and

iII1 , j1II

PII selects a strategy and each player sees the j that was selected by
their opponent. Now it is PI's turn again. What are his information sets? Both
players have perfect recall, so their information sets will be limited to the
branch of the tree corresponding to the
what

II

j1

iP1 , jP1

selected. Since PI knows

is, his information set will be further reduced to those nodes with

iI1 , j1I

II
j 1 . So in each branch of the tree corresponding to

a common

there will be four information sets, each one corresponding to a common

II
j 1 . In round 2 there will be 48 information sets total for PI, and every

information set will contain three nodes.

In round two, PI has only two pure strategies available, since there are
only two ways to assign the remaining two rows (the ones not assigned in
round one) to himself and his opponent. So there will be two branches from
each node, each labeled with an
to either

i I1

or

MATH 5750 PROJECT

i , j where neither i nor

are equal

j 1I . There are now 288 nodes.

After PI makes his choice, PII now selects her round two strategy. What
are her information sets? PII has perfect recall, so all information sets are
limited to nodes with a common label representing PII's chosen strategy in
round one; in other words, all nodes in an information set have the same

iII1 , j1II

. PII also knows the row PI picked for her in round one,
I

nodes with a common

j 1I , so only

for PI will be in an information set. Since PII does

j1

not know what strategy PI chose in round two, all nodes that PI could have
taken in round two subject to the above restrictions will be included in an
information set. All told, there will be 48 information sets for PII, each
containing six nodes. Like PI, PII will have only two pure strategies available,
with labels subject to similar constraints PI's round two labels faced.
There are now 576 terminal nodes. In this game, a player can only win,
lose, or tie, and we can represent this with payoffs of

1 ,

1 , and

0 ,

respectively (we are ignoring betting). Recall that PI's final score is
S I ( i 1I , j II1 ) +(i 2I , j II2 ) , and
S I < S II

S II ( j 1I ,i 1II ) +( j 2I ,i 2II ) . PI wins if

S I > S II

and loses if

(in other words, PI wins if his score is higher than PII's, and loses if

his score is less than PII's). Otherwise, there is a tie. Using indicator
functions, I represent the payoff function as:
1

{( i , j )+( i , j ) >( j ,i ) +( j , i )}
I
1

II
1

I
2

II
2

I
1

II
1

I
2

II
2

{( i , j ) +( i , j ) <( j ,i ) +( j , i )}
I
1

II
1

I
2

II
2

I
1

II
1

I
2

II
2

=1{ S >S }1{ S < S }

I

II

II

This completely describes the game tree.

Using this description of the game in extensive form, we can then
represent the game in strategic form and find the matrix of the game. Each

MATH 5750 PROJECT

player has 49 information sets total. The procedure for converting a game
from extensive form to strategic form is to represent a pure strategy in
strategic form with a choice of a strategy from each of a player's information
sets. In this context, however, this procedure would result in dominated
strategies through strict equality. This results from the players' perfect recall
ability. Information sets in round two condition on a player's strategy chosen
in round one and what is known about the strategy the opponent chose in
round one. Since a player knows what strategy he or she chose in round one,
including rules of play that condition on that player choosing a different
strategy in round one from what that player actually chose (according to that
strategy) only result in redundant pure strategies. Therefore, I can simplify
the matrix by representing a pure strategy in strategic form by a choice of
strategy in row one, and a choice of strategy from each of the four
information sets following that choice of strategy. This makes intuitive sense
since it says a pure strategy is a combination of two rows (or columns, in the
case of PII) in round one, and then for each column (row) that PII (PI) could
assign to PI (PII) in round one, a response of a combination of the two
remaining rows (columns) is assigned. This results in a square matrix with
192 rows and 192 columns.
I wrote an R script that can process a game, create a game matrix
representing it in strategic form, and can find an approximate solution. In
practice, the resulting payoff matrix is singular due to rows or columns being
linearly dependent through equality, and the only way to solve this linear

MATH 5750 PROJECT

dependence is to eliminate dominated strategies (it does not check for

strategies that could be dominated by certain mixed strategies, since doing
so would be a nightmare to program, at least with my current programming
abilities). The resulting matrix is not square and therefore not invertible, so a
solution needs to be found using a method that does not require inverting
the matrix. I tried using the pivot method for solving a game, but the matrix
is too large to reach a solution in any reasonable time. The method I actually
use is the fictitious play method (with PI's first pure strategy being selected
at random), and my script can find an approximate solution after a specified
number of iterations or when a specified level of "accuracy" (defined as the
distance in absolute value between the lower bound and upper bound of the
value of the game) is first reached (be aware that the convergence of the
lower bound and upper bound of the value of the game found by the
fictitious play algorithm is not monotone).
I found approximate solutions of different games (i.e. different magic
square game boards) with accuracies of at least 0.001. The lower bound of
the value was always less than zero and the upper bound always greater
than zero, though their absolute values were always unequal. Optimal mixed
strategies also seemed to guarantee at least a tie for both players. It seems
that if the value of the game is not zero, it is very close to zero, and the
players' optimal strategies are to choose combinations of rows and columns
that, while not guaranteeing ties (there were no saddle points in the payoff

MATH 5750 PROJECT

matrix, so neither player has a pure strategy that will guarantee at least a
tie), make ties very likely when an opponent plays optimally.
Much more can be said about Cordial Minuet. I did not solve the full six
by six game, and I do believe a solution is possible. I believe the procedure I
used to analyze the simpler version of the game could be adapted to analyze
the full game. Cordial Minuet's third round of play will be the most difficult to
translate to strategic form, but I believe that doing so is possible, given a
good notation scheme.
Betting strategies would also be interesting to analyze. It may seem that
since the value of the game is zero, the best betting strategy would be to
match your opponent but not raise by more than what is necessary to
continue play (in practice, however, an aggressive player might assume that
his opponent is not playing optimally and therefore his expected payoff when
playing optimally is positive, so he will want to encourage his opponent to
invest as much of her pay-in as he can).
The exception to this is the final round of betting, when players begin to
reveal the squares they got. At this phase of the game, a player may decide
that winning is highly unlikely, and therefore fold. Perhaps the best way to
analyze this phase of the game is to imagine it being an entirely separate
game (similar to how insight is drawn from endgame poker).
For example, suppose two players are playing a card game where the
player holding the hand with the larger sum wins (all card have a numeric
value). There is a probability

MATH 5750 PROJECT

that PIs hand is a winning hand, a

10

probability

that PIs hand is a losing hand, and a probability

1 pq

that the hands are of equal value. The players see their own hands but not
their opponents hand. When the players receive their hands, there is a
round of betting. After this round of betting, each player reveals one card of
their choice from their hands. Another round of betting ensues, then the
players reveal their entire hands, with the winner taking the pot.
There are a number of questions for how this game is played. How do
the players decide to reveal their first card? Do they select their best, worst,
or middle card? How does a player react to the card his opponent revealed?
How do the players bet? Answering these questions may be enlightening to
playing the final round of Cordial Minuet, since the problem is similar.
Cordial Minuet has strategic depth. Game theory analysis could provide
answers to how to play, though probably not before the game is released.
Hopefully they are found after release, though; otherwise, interest in the
game might fade and it may never be released.

References
[1] Rohrer, S.C. Cordial Minuet: How to Play in Three Minutes. Retrieved
April 4th, 2015 from https://www.youtube.com/watch?
v=wZcmUGMj30k
[2] Totilo, S. A Mildly Satanic New Video Game That You Can Only Play
Online With Money. Retrieved April 4th, 2015 from
http://kotaku.com/a-mildly-satanic-new-video-game-that-you-mustplay-onli-1638083439
[3] Wikipedia. Magic Square. Retrieved April 4th, 2015 from
http://en.wikipedia.org/wiki/Magic_square

MATH 5750 PROJECT

11