THE ENDGAME IN POKER

Chris Ferguson, Full Tilt Poker

Tom Ferguson, Mathematics, UCLA
Abstract. The simple two-person poker model, known as Basic Endgame, may be
described as follows. With a certain probability known to both players, Player I is dealt
a hand which is a sure winner. If it is not a sure winner, then Player II is sure to have a
better hand. Player I may either check, in which case the better hand wins the antes, or
bet. If Player I bets, Player II may either fold, conceeding the antes to Player I, or call,
in which case the better hand wins the antes plus the bets. This model is reviewed and
extended in several ways. First, several rounds of betting are allowed before the hands
are compared. Optimal choice of the sizes of the bets are described. Second, Player II
may be given some hidden information concerning the probability that Player I has a sure
winner. This information is given through the cards Player II receives. Third, the model
is extended to the case where the hand Player I receives only indicates the probability it
is a winner. This allows for situations in which cards still to be dealt may inuence the
outcome.
Introduction.
In this investigation, we treat from a game-theoretic point of view several situations
that occur in the game of poker. The analysis of models of poker has a long and distin-
guished existence in the game theory literature. Chapter 5 of the book of

Emile Borel,
Applications aux jeux de hasard (1938), and Chapter 19 of the seminal book on game
theory by von Neumann and Morgenstern (1944) are devoted to the topic. In the 1950s,
others developed further certain aspects of modeling poker. Kuhn (1950) treats three
card poker. Nash and Shapley (1950) treat a three person poker model. Bellman and
Blackwell (1949), Bellman (1952), Gillies, Mayberry and von Neumann (1953), Karlin and
Restrepo (1957), Goldman and Stone (1960ab), and Pruitt (1961) extend various aspects
of the poker models of Borel and of von Neumann-Morgenstern further. Chapter 9 of the
textbook of Karlin (1959) summarizes this development. For a more recent treatment of
the models of Borel and von Neumann, see Ferguson and Ferguson (2003) and Ferguson,
Ferguson and Gawargy (2007).
Generally, the aim of such research is to analyze a simplied model of the game of
poker completely, with the hope of capturing the spirit of poker in a general sense. Others
have tried to analyze specic situations or aspects of the real game with the hope that
1
it may improve ones play. Papers of Newman (1959), Friedman (1971), Cutler (1975,
1976), and Zadeh (1977), and the book of Ankeny (1981) are of this category. On the
other hand there are also books that contain valuable information and recommendations
on how to play the real game of poker. The books of Brunson (1978) and Sklansky (1987)
on general games of poker and of Sklansky and Malmuth (1988), Sklansky, Malmuth and
Zee (1989), Zee (1992) and Harrington and Robertie (2004-2006) on specic games of
poker can be recommended. Unusual in its treatment mixing mathematical analysis, game
theoretic ideas and the real game of poker, the book of Chen and Ankenman (2006) may
be especially recommended.
One of the simplest and most useful mathematical models of a situation that occurs in
poker is called the classical betting situation by Friedman (1971) and basic endgame
by Cutler (1976). These papers provide explicit situations in the game of stud poker and of
lowball stud for which the model gives a very accurate description. This model is also found
in the exercises of the book of Ferguson (1967). Since this is a model of a situation that
occasionally arises in the last round of betting when there are two players left, we adopt
the terminology of Cutler and call it Basic Endgame in poker. This will also emphasize
what we feel is an important feature in the game of poker, that like chess, go, backgammon
and other games, there is a distinctive phase of the game that occurs at the close, where
special strategies and tactics that are analytically tractable become important.
1. Basic Endgame.
Basic Endgame is played as follows. Two players, Player I and Player II, both put
an ante of a dollars into a pot (a > 0). Player I then draws a card from a deck of cards
that gives him a winning card with probability (w.p.) P and a losing card w.p. 1 P,
0 < P < 1. Both players know the value of P, but only Player I knows if the card he
received is a winning card or not. Player I may then check (also called pass) or bet b
dollars (b > 0). If Player I checks, the game is over and the antes goes to Player I if he
has a winning card and to Player II otherwise. If Player I bets, Player II may then fold, or
she may call by also putting b dollars in the pot. If she folds, then Player I wins the ante
whatever card he has. If Player II calls, then the ante plus the bet is won by Player I if
he has a winning card and by Player II otherwise. Only the ratio of b to a is signicant,
but we retain separate symbols for the ante and the bet so that the results are easier to
understand.
Situations of the form of Basic Endgame arise in poker. For example, in the last round
of betting in a game of ve card stud poker, Player Is cards are the 5 of diamonds, the
6 of spades, the 7 of diamonds, the 8 of hearts and a hidden hole card. Player IIs cards
are the 2 of hearts, the 3 of spades, the king of spades, the king of clubs and a hidden
hole card. No matter what card Player II has in the hole, Player I will win if and only if
he has a 4 or a 9 in the hole. Since Player II has the higher hand showing, she must act
rst by betting or checking. In this situation, it is optimal for her to check. Assuming she
does check, it then becomes Player Is turn to act, and we have a situation close to Basic
Endgame described in the previous paragraph. The number a may be taken to be half the
size of the present pot, and b will be the maximum allowable bet. The number P is taken
2
to be the probability that Player I has a winning hole card given the past history of the
game.
There are several reasons why Basic Endgame is not a completely accurate description
of this poker situation. In the rst place, the probability P, calculated on the basis of three
rounds of betting before the nal round and the actions of the players in these rounds, is
an extraordinarily complex entity. It would be truly remarkable if both players arrived at
the same evaluation of P as required by Basic Endgame. Secondly, Player IIs hole card
gives her some secret information unknown to Player I that inuences her estimation of P,
and both Player I and Player II must take this into account. In Section 5, we extend the
model to allow for this hidden type of information. Thirdly, a player may unknowingly give
away information through mannerisms, hesitations, nervousness, etc. This type of hidden
information, called tells, (see Caro (1984)) is of a dierent category than information
given by a hidden hole card. The player who gives away such information has no way
of taking this into account since he is unaware of its existence. Another way this type
of hidden information can arise is through cheating. For example, someone may gather
information about the cards through a hole in the ceiling and pass this information to
one of the players at the table. (Dont laughthis happened at one of the casinos in
California.) Such games in which one of the players does not know all the rules of the
game, are called pseudo-games, and have been studied by Ba nos (1968) and Megiddo
(1980). The extensive literature on repeated games of incomplete information (e.g. see
Aumann and Mashler (1995) or Sorin (2000)) is also an attempt to treat this problem.
If Player I receives a winning card, it is clear that he should bet: If he checks, he wins
a net total of a dollars, whereas if he bets he will win at least a dollars and possibly more.
In the analysis of the game below, we assume that Player I will bet with a winning card.
The rules of the game may be summarized in a diagram called the Kuhn tree, a device
due to Kuhn (1953). Figure 1 gives the Kuhn tree of Basic Endgame. It is to be read from
the top down. The rst move is a chance move with probabilities P and 1 P attached to
the edges. Then Player I moves, followed by Player II. The payos to Player I are attached
to each terminal branch of the tree. The only features not self-explanatory are the circle
and the long oval. These represent information sets. The player whose turn it is to move
from such a set does not know which node of the set the previous play has led to. Thus
the long oval indicates that Player II does not know which of the two nodes she is at when
she makes her choice, whereas the circle indicates that Player I does learn the outcome of
the chance move.
In this situation, Player I has two possible pure strategies: (a) the blu strategybet
with a winning card or a losing card; and (b) the honest strategybet with a winning card
and check with a losing card. Player II also has two pure strategies which are (a) the call
strategyif Player I bets, call; and (b) the fold strategyif Player I bets, fold. The payo
matrix is the two by two matrix of expected winnings of Player I,
_
fold call
honest (2P 1)a (2P 1)a + Pb
blu a (2P 1)(a + b)
_
3
P
1P
bet bet pass
call fold call fold
a
a+b a a (a+b)
Chance
II
I
I
Figure 1.
We state the solution to Basic Endgame by giving the value and optimal (i.e. minimax)
strategies for the players. There are two cases. In the rst case, for values of P close to
one, there is a saddle-point the players have optimal pure strategies. The main case
is for small values of P, when both players use both strategies with positive probabilities.
This is called the all-strategies-active case.
If P (2a + b)/(2a + 2b), then there is a saddlepoint.
(i) Is optimal strategy is to bet always.
(ii) IIs optimal strategy is to fold always.
(iii) The value of the game is a.
If P < (2a + b)/(2a + 2b), then we are in the all-strategies-active case.
(i) Is optimal strategy is to blu w.p. := (b/(2a + b))(P/(1 P)).
(ii) IIs optimal strategy is to fold w.p. := b/(2a + b).
(iii) The value of the game is 2aP[(2a + 2b)/(2a + b)] a.
These strategies have an easy derivation and interpretation using one of the basic
principles of game theory called The Indierence Principle: In those cases where your
opponent, using an optimal strategy, is mixing certain pure strategies against you, play to
make your opponent indierent to which of those strategies he uses.
In Basic Endgame, this principle may be used as follows. In the all-strategies-active
case, II plays to make I indierent to betting or checking with a losing card. If I passes,
he wins a. If he bets, he wins a w.p. and wins (a + b) w.p. 1 , where is the
(unknown) probability that II folds. His expected return in this case is a(1)(a+b).
Player I is indierent if a(1)(a+b) = a. This occurs if = b/(2a+b). Therefore,
II should choose to fold w.p. b/(2a + b).
Similarly, I chooses the probability of betting with a losing card to make II indierent
to calling or folding. If II calls, she loses (a+b) w.p. P, while w.p. 1P she wins (a+b) w.p.
and wins a w.p. 1; her expected loss is therefore, P(a+b)(1P)(a+b)(1P)(1
)a. If she folds, she loses a w.p. P, while w.p. 1 P, she loses a w.p. and wins a w.p.
4
1 ; in this case her expected loss is a[P +(1 P) (1 P)(1 )]. She is indierent
between calling and folding if these are equal, namely if = (P/(1 P))(b/(2a + b)).
Therefore, I should blu with this probability.
It is interesting that Player IIs optimal strategy does not depend on P in the all-
strategies-active case. In particular in pot limit poker where b = 2a, her optimal strategy
is to call w.p. 1/2 and fold w.p. 1/2. However, Player II must check the condition for the
all-strategies-active case before using this strategy. This condition is automatic for I. If I
computes and it is greater than 1, then he should blu w.p. one.
Choosing the Size of the Bet.
In Basic Endgame, the size of the bet of Player I was xed at some number b > 0.
The general situation where I is allowed to choose any positive bet size, b, less that some
maximum amount, B, was investigated in an unpublished paper by Cutler (1976). The
conclusion is that Player I may as well always bet the maximum; in fact, in the all-
strategies-active case it is a mistake for I to bet less than the maximum. (By a mistake
for Player I, we mean that Player II can take advantage of such a bet without risk and
achieve a strictly better result than she could against optimal play of the opponent.) It
is dangerous for I to let the size of the bet depend on whether he has a winning card or
not. II may be able to take advantage of this information. Also, if I bets the same amount
regardless of his hand, he might as well bet the maximum since the value to him is a
nondecreasing function of the bet size.
Cutler gives optimal strategies for Player II which allow her to take advantage of any
variation in Is bet sizes without incurring any risk. His general result is as follows, where
B represents the maximum allowable bet size.
If P B/(2a + B), Player I should always blu with a losing hand. Player II should
always fold no matter how big or small Is bet is.
If P < B/(2a+B), Player I should blu by betting B w.p. (B/(2a+B))(P/(1 P)).
If Player II hears a bet of size b, 0 < b B, then she should call w.p. p(b) where
2a
2a + b
p(b) min{1,
2aB
(2a + B)b
}.
Any such p(b) is minimax for Player II. In particular, she may pretend that the bet
size was xed at b and use the solution to Basic Endgame, namely call w.p. 2a/(2a + b).
When b < B, this gives her an improved expected payo against all strategies of Player
I except strategies that only bet less than B when I has a losing card. To obtain an
improved expected payo against all strategies, she should call a little more often, but
with probability still less than 2aB/((2a + B)b).
2. Basic Endgame with Two Rounds of Betting.
It sometimes happens that Player I will face an endgame situation with two or more
rounds of betting yet to take place, in which the cards to be dealt between rounds do not
aect the outcome. For two rounds, this is modelled as follows.
5
Two players both ante a units into the pot. Then Player I receives a winning card
w.p. P 0 and a losing card w.p. 1 P 0. It is assumed that I knows which card he
has whenever he makes a decision, and II does not know which card I holds whenever she
makes a decision. Player I then either passes or bets an amount b
1
> 0. If he passes, the
game is over and he wins a if he holds the winning card and loses that amount if he holds
the losing card. If I bets, Player II may call or fold. If II folds the game is over and I wins
a. If II calls, I may either pass or bet b
2
> 0. If he passes, the game is over and he wins
a + b
1
if he holds the winning card and loses that amount if he holds the losing card. If
he bets, then II may call or fold. If II folds, then the game is over and I wins a +b
1
. If II
calls, then the game is over and I wins a + b
1
+ b
2
if he holds the winning card and loses
that amount if he holds the losing card.
If I receives a winning card, it is clear he should never pass. We assume the rules of
the game require him to bet in this situation. With such a stipulation, the game tree is
displayed in Figure 2.
P 1P
Chance
bet bet pass
call fold call fold
bet bet pass
call fold call fold
a
a a
(a+b
1
)
a+b
1
+b
2
a+b
1
(a+b
1
+b
2
) a+b
1
I
I
I
I
II
II
Figure 2.
If I chooses to pass at the rst round, then it does not matter what he does in the
second round. So I has just three pure strategies, pass, bet-pass, and bet-bet. Similarly, II
has just three strategies, fold, call-fold and call-call. The resulting three by three matrix
of expected payos is
6
_
_
fold call-fold call-call
pass a(2P 1) a(2P 1) +Pb
1
a(2P 1) + P(b
1
+ b
2
)
bet-pass a (2P 1)(a + b
1
) (2P 1)(a + b
1
) + Pb
2
bet-bet a a + b
1
(2P 1)(a + b
1
+ b
2
)
_
_
We state the solution to this game. Since this is a special case of the problem treated
in the next section, we omit the proof. Let
P
0
:=
(2a + b
1
)(2a + 2b
1
+ b
2
)
(2a + 2b
1
)(2a + 2b
1
+ 2b
2
)
.
If P > P
0
, then
(i) the value is V = a,
(ii) it is optimal for Player II to fold on the rst round, and
(iii) it is optimal for Player I to bet on the rst round, and to bet w.p. (P/(1P))(b
2
/(2a+
2b
1
+ b
2
)) (or w.p. 1 if this is greater than 1) on the second round.
If P P
0
, then all strategies are active,
(i) the value is V = a(2P P
0
)/P
0
(ii) it is optimal for Player II to fold on the rst round w.p. b
1
/(2a + b
1
), and to fold on
the second round w.p. b
2
/(2a + 2b
1
+ b
2
), and
(iii) with a winning card, Player I always bets; with a losing card, he bets on the rst round
w.p.
P
1 P

1 P
0
P
0
, and on the second round w.p.
b
2
(2a +b
1
)
b
2
(2a +b
1
) +2b
1
(a + b
1
+ b
2
)
.
Note the following features. The cuto-point, P
0
, between the two cases is just the
product of the cuto points of the two rounds treated separately, that is (2a+b
1
)/(2a+2b
1
)
for the rst round and (2a + 2b
1
+ b
2
)/(2a + 2b
1
+ 2b
2
) for the second round. The rst
case, P > P
0
, occurs if and only if the lower right 2 by 2 submatrix of the payo matrix
has value at least a. If Player I uses the optimal strategy for this 2 by 2 submatrix, then
his expected payo is at least a, and since he can get no more than a if Player II always
folds, the value must be a. In this sense, the rst case is easy.
In the all-strategies-active case where P P
0
, Player IIs optimal strategy is just the
strategy that uses her optimal strategy for Basic Endgame in both the rst and second
round. In the rst round, Player II sees a pot of size 2a + b
1
and is required to invest b
1
to have a chance to win it. Therefore, she folds w.p. b
1
/(2a + b
1
). In the second round,
she sees a pot of size 2a + 2b
1
+ b
2
and is required to call with b
2
to continue. Therefore,
she folds with conditional probability b
2
/(2a + 2b
1
+ b
2
).
Player Is optimal strategy in the all-strategies-active case makes Player II indierent
to folding or calling in both the rst and the second round. It is interesting to note that
Player Is behavioral strategy in the second round is independent of P.
7
Choosing the Sizes of the Bets.
In choosing the sizes of the bets in the all-strategies-active case, three cases deserve
special mention. The rst case is the case of limit poker, where there is a xed upper limit,
B, on the size of every bet. The value, V , is an increasing function of both b
1
and b
2
, so
the optimal choice of bet size for Player I is b
1
= b
2
= B. The formulas for the optimal
strategies and the value do not simplify signicantly in this case.
The second case is the pot-limit case. Since V is increasing in both b
1
and b
2
, the
optimal choices are b
1
= 2a and b
2
= 6a, the pot-limit bets. The basic formulas for the
solution simplify in this case. The inequality dening the all-strategies-active case reduces
in this case to P 4/9. The value is V = (9P 2)a/2. The optimal strategy of Player II
is: Call w.p. 1/2 on both the rst and second rounds. The optimal strategy for Player I is:
Bet with a winning card; with a losing card bet on the rst round w.p. (5/4)(P/(1 P)),
and on the second round w.p. 2/5.
The third case is no-limit (table-stakes) poker, in which a player may bet as much
as he likes but no more than he has placed on the table when play begins. In addition,
if the amount bet exceeds that amount he has left, he may call that part of the bet up
to the amount he has left. If there are two players with remaining resources, B
1
and B
2
,
the maximum bet size is for all practical purposes B = min{B
1
, B
2
}. As in Section 1, it
is optimal for Player I to bet the maximum eventually, but the question remains of how
much of it to bet on the rst round.
Suppose therefore that the sum of the bets, b
1
+ b
2
= B, is xed, and Player I is
allowed to choose the size of the rst bet, b
1
, subject to 0 b
1
B. Then in the
all-strategies-active case, the optimal value of b
1
is to maximize
V = a
_
2P
(2a + 2b
1
)(2a + 2b
1
+ 2b
2
)
(2a + b
1
)(2a + 2b
1
+ b
2
)
1
_
= a
_
2P
(2a +2b
1
)(2a + 2B)
(2a + b
1
)(2a + b
1
+ B)
1
_
.
The value of b
1
that maximizes V is easily found by setting the derivative of V with respect
to b
1
to zero. Solving the resulting equation reveals the optimal choice of b
1
to be
b
1
=
_
aB + a
2
a.
3. Basic Endgame With Many Rounds of Betting.
We may extend Basic Endgame to allow an arbitrary nite number, n, of betting
rounds.
Two players both ante a units into the pot, a > 0. Then Player I receives a winning
card w.p. P 0 and a losing card w.p. 1 P 0. It is assumed that I knows which card
he has whenever he makes a decision, and II does not know which card I holds whenever
she makes a decision. Player I then either passes or bets an amount b
1
> 0. If he passes,
the game is over and he wins a if he holds the winning card and loses that amount if he
holds the losing card. If I bets, Player II may call or fold. If II folds the game is over and
8
I wins a. If II calls, the game enters round 2. In round k where 2 k < n, I may either
pass or bet b
k
> 0. If he passes, the game is over and he wins a + b
1
+ + b
k1
if he
holds the winning card and loses that amount if he holds the losing card. If he bets, then
II may call or fold. If II folds, then the game is over and I wins a + b
1
+ + b
k1
. If II
calls, then the game enters round k +1. If II calls in round n, the game is over and I wins
a +b
1
+ +b
n
if he holds the winning card and loses that amount if he holds the losing
card. We rst assume the b
k
are xed numbers.
Below, we derive the value optimal strategies of the players. We summarize the
solution as follows.
Summary of Solution. Let
r
k
:=
b
k
2(a + b
1
+ + b
k1
) + b
k
and let P
0
= (

n
1
(1 + r
j
))
1
.
If P > P
0
, then it is optimal for Player I to bet on the rst round and for Player II
to fold. The value is V = a.
If P P
0
, then
(1) the value is V = a(2P P
0
)/P
0
,
(2) Player IIs optimal strategy is at each stage k to fold w.p. r
k
, and
(3) Player Is optimal strategy is to bet with a winning card; with a losing card, to bet
on the rst stage w.p. p
1
=
P
1 P

1 P
0
P
0
, and if stage k > 1 is reached, to bet w.p.
p
k
=

n
k
(1 +r
j
) 1

n
k1
(1 + r
j
) 1
.
We notice some remarkable features of this solution in the all-strategies-active case,
P (

n
1
(1 +r
j
))
1
. First note that r
k
is just the amount bet at stage k divided by the
new pot size. This means that Player IIs optimal strategy is just the repeated application
of Player IIs optimal strategy for Basic Endgame. In addition, Player Is optimal strategy
depends on P only at the rst stage; thereafter his behavior is independent of P. After
the rst stage, he will bet with the same probabilities if P is very small, say P = .001, as
he would if P = .1. This means that his main decision to blu is taken at the rst stage;
thereafter all blus are carried through identically.
Derivation. If I receives a winning card, it is clear he should never pass. We assume
the rules of the game require him to bet in this situation.
Player I has n +1 pure strategies, i = 0, 1, . . . , n, where i represents the strategy that
blus exactly i times. Similarly, there are n+1 pure strategies for Player II, j = 0, 1, . . . , n,
where j represents the strategy that calls exactly j times. Let A
ij
denote the expected
payo to I if I uses i and II uses j. Let
s
j
:= a + b
1
+ + b
j
9
denote half the size of the pot after round j; in particular, s
0
= a. Then
A
ij
=
_
Ps
j
(1 P)s
i
for 0 i j n
s
j
for 0 j < i n.
(1)
Let (
0
,
1
, . . . ,
n
) denote the mixed strategy for Player II in which
j
is the proba-
bility that II calls exactly j times. If II uses this strategy and Player I uses i, the average
payo is
V
i
:=
n

j=0
A
ij

j
=
i1

j=0
s
j

j
+ P
n

j=i
s
j

j
(1 P)s
i
n

j=i

j
= P
n

j=0
s
j

j
+(1 P)
i1

j=0
s
j

j
(1 P)s
i
n

j=i

j
.
(2)
We search for a strategy (
0
,
1
, . . . ,
n
) to make V
i
independent of i. Such a strategy
would guarantee that Player IIs average loss would be no more than the common value of
the V
i
. We look at the dierences
V
k
V
k1
= (1 P)[(s
k
+ s
k1
)
k1
(s
k
s
k1
)
n

j=k1

j
] (3)
This is zero for k = 1, . . . , n if

k1

n
j=k1

j
=
s
k
s
k1
s
k
+ s
k1
= r
k
(4)
This denes the
k
. In fact, the left side represents the probability II folds in round k
given that it has been reached, and so is the behavioral strategy for II in round k. The
equalizing value of the game may be found as follows.
V
0
= P
n

j=0
s
j

j
(1 P)s
0
V
n
=
n1

j=0
s
j

j
+ Ps
n

n
(1 P)s
n

n
=
n

j=0
s
j

j
2(1 P)s
n

n
(5)
From V
0
= V
n
, we see that

n
j=0
s
j

j
= 2s
n

n
s
0
. This gives the value as V
0
=
2Ps
n

n
s
0
. Finally, repeatedly using (4) in the form 1 r
k
=

n
k

j
/

n
k1

j
, we nd
that

n
i=1
(1 r
i
) =
n
. Hence the value is
V
0
= 2Ps
n
n

i=1
(1 r
i
) s
0
.
10
Noting that 1 r
k
= 2s
k1
/(s
k
+ s
k1
) and 1 + r
k
= 2s
k
/(s
k
+ s
k1
), we nd that
s
n

n
1
(1 r
k
) = s
0

n
1
(1 + r
k
). This, with s
0
= a, gives an alternate form of V
0
, namely
V
0
= a
_
2P
n

1
(1 + r
k
) 1
_
= a
_
2
P
P
0
1
_
. (6)
Player II can keep the value of the game to be at most V
0
. But Player II can also keep
the value to be at most a by folding always. We shall now see by examining Player Is
strategies that the value of the game is the minimum of (6) and a.
Let (
0
,
1
, . . . ,
n
) denote a mixed strategy of Player I, where
i
is the probability
of making exactly i bets. If Player I uses this strategy and Player II uses column j, the
average payo is for 0 j n,
W
j
:=
n

i=0

i
A
ij
=
j

i=0

i
(Ps
j
(1 P)s
i
) +
n

i=j+1

i
s
j
= Ps
j
+ (1 P)s
j
n

i=j+1

i
(1 P)
j

i=0
s
i

i
.
(7)
Equating W
j
and W
j1
leads to the following simultaneous equations for 1 j n,
P(s
j
s
j1
) +(1 P)(s
j
s
j1
)
n

i=j+1

i
(1 P)(s
j
+ s
j1
)
j
= 0 (8)
Solving for
j
yields the equations,

j
=
P
1 P
r
j
+ r
j
n

i=j+1

i
. (9)
which denes
n
, . . . ,
1
by backward induction. We nd
n
= (P/(1 P))r
n
, and
n

i=j+1

i
=
P
1 P
_
_
n

i=j+1
(1 +r
i
) 1
_
_
, (10)
so that Player Is behavioral strategy at stages 2 j n is
p
j
=

n
j

i

n
j1

i
=

n
j
(1 + r
i
) 1

n
j1
(1 +r
i
) 1
. (11)
The behavioral strategy at the rst stage is
p
1
= 1
0
=
n

i
=
P
1 P
_
n

1
(1 + r
i
) 1
_
. (12)
11
Assuming the resulting p
1
1, we can evaluate the common value of the W
j
using
W
0
= Ps
0
+ (1 P)s
0
n

i=1

i
(1 P)s
0

0
= 2Ps
0
n

1
(1 + r
i
) s
0
= V
0
. (13)
Therefore, W
0
= V
0
is the value of the game provided p
1
1, or equivalently, provided
P

n
1
(1 + r
i
) 1.
Finally it is easily checked that p
1
1 if and only if V
0
is not greater than s
0
.
Choosing the sizes of the bets.
Suppose that the initial size of the pot, 2b
0
= 2s
0
, is xed, and that the total amount
to be bet, b
1
+ + b
n
= s
n
s
0
, is xed, where n is the number of rounds of betting.
This situation occurs in no-limit, table stakes games, where s
n
s
0
is the minimum of
the stakes that Player I and Player II have in front of them when betting begins. In the
last round, Player I should certainly bet the maximum amount possible. The problem for
Player I is to decide how much of the total stakes to wager on each intervening round.
This is equivalent to nding the choices of s
1
, s
2
, . . . , s
n1
that maximize the value, V
0
, of
(6) subject to the constraints,
s
0
s
1
s
2
s
n
. (14)
Maximizing V
0
is equivalent to maximizing
n

i=1
(1 + r
i
) =
n

i=1
2s
i
s
i
+ s
i1
. (15)
As a function of s
i
for xed s
1
, . . . , s
i1
, s
i+1
, . . . , s
n1
, this proportional to
s
i
(s
i
+ s
i1
)(s
i+1
+ s
i
)
(16)
This is unimodal in s
i
on (0, ) with a maximum at (s
i
+ s
i1
)(s
i+1
+ s
i
) = s
i
[s
i+1
+
2s
i
+ s
i1
], or equivalently, at
s
i
=

s
i1
s
i+1
(17)
Note that s
i
is the geometric mean of s
i1
and s
i+1
so that s
i
is between s
i1
and s
i+1
.
Thus, the global maximum of (15) subject to (14) occurs when (17) is satised for all i =
1, 2, . . . , n1. Inductively using (17) from i = 1 to n1, we can nd s
i
in terms of s
0
and
s
i+1
to be s
i
= s
1/(i+1)
0
s
i/(i+1)
i+1
for i = 1, . . . , n. For i = n 1, this is s
n1
= s
1/n
0
s
(n1)/n
n
.
We may now work back to nd
s
i
= s
(ni)/n
0
s
i/n
n
= s
0
(s
n
/s
0
)
i/n
for i = 1, . . . , n 1. (18)
12
From this we may nd the folding probabilities for Player II.
r
i
=
s
1/n
n
s
1/n
0
s
1/n
n
+ s
1/n
0
for i = 1, . . . , n 1. (19)
Note that this is independent of i. Since this is the ratio of bet size to the size of the new
pot, one sees that the optimal bet size must be a xed proportion of the size of the pot!
This proportion is easily computed to be
r
1 r
=
(s
n
/s
0
)
1/n
1
2
, (20)
where r denotes the common value of the r
i
of (19). Note that this is independent of P!
Therefore it is optimal for Player I to bet this proportion of the pot at each stage.
Let us compute Player Is optimal blung probabilities. At the initial stage, I should
bet (with a losing card) w.p.
p
1
=
P
1 P
[(r + 1)
n
1], (21)
If p
1
> 1, then Player I should bet w.p. 1, and Player II should fold. In terms of P, this
inequality becomes P (r +1)
n
, so if P (r +1)
n
, Player I should blu initially w.p.
p
1
and subsequently, if stage j +1 is reached, he should blu w.p.
p
j+1
=
(1 + r)
nj
1
(r + 1)
nj+1
1
. (22)
The value of the game when the optimal bet sizes are used is
V
0
= a
_
P2
n+1
s
n
(s
1/n
n
+ s
1/n
0
)
n
1
_
(23)
Example.
Suppose there are $20 in the pot, so s
0
= 10, and suppose there are going to be three
rounds of betting, so n = 3. If one player has $260 in front of him and the other has $330,
then since s
3
s
0
is the minimum of these two quantities, we have s
3
= 260 + 10 = 270.
Then s
3
/s
0
= 27 and 27
1/3
= 3, so from (18), s
i
= 10 3
i
, so s
1
= 30 and s
2
= 90.
Therefore, Player I should bet s
1
s
0
= $20 on the rst round. If II calls I should bet
$60= s
2
s
1
on the second round. In this example, I bets the size of the pot at each
round (including the third round), the same amount as in pot limit poker. To be in the
all-strategies-active case, we require P < 8/27. If this is satised, Player II calls each bet
of Player I w.p. r = 1/2; this agrees with (19). To nd Player Is blung probabilities,
13
we must specify P. If P = 1/4 for example, then p
1
= (1/3)[(3/2)
3
1] = 19/24. The
subsequent betting probabilities (22) are
p
2
=
(3/2)
2
1
(3/2)
3
1
= 10/19 and p
3
=
(3/2) 1
(3/2)
2
1
= 2/5.
Note that these probabilities are decreasing in the later rounds.
The proportion of the pot Player I bets on each round depends on the amount of the
minimum table stakes. If this was $70 instead of $260, then the optimal bet would be half
the size of the pot at each stage.
4. Basic Endgame With a Continuum Number of Rounds.
We see that Player I gets a denite advantage if he is allowed to split his betting
over two rounds rather than betting the entire amount in one round. To see what sort of
advantage I gets from a large number of rounds, we model the game as having a continuum
number of rounds. We suppose that both players ante 1 unit each into the pot, and that
the total amount to be bet is B, with an innitesimal bet of dt being placed at time t for
0 < t < B. Before play begins, Player I receives a winning card w.p. P and a losing card
w.p. 1 P. If I has a winning card, he bets continuously throughout the whole interval
[0, B]. If he has a losing card, he chooses a time x [0, B] at which to stop betting. If
x = 0, he passes initially, and if x = B he bets throughout the whole interval. Player
II chooses a time y [0, B] at which to stop calling. For xed choices of x and y, the
expected payo is
W(x, y) =
_
1 + y if 0 y < x B
P(1 + y) (1 P)(1 + x) if 0 x y B.
(1)
Thus if II stops calling before I stops betting, I wins the ante plus the total amount bet,
namely 1 + y. If I stops betting (with a losing card) before II stops calling, I wins 1 + y
if he has the winning card and loses 1 + x if he has the losing card. We assume that x
represents the last time Player I bets and y represents the last time player 2 calls, so that
if x = y, the payo is P(1 + y) (1 P)(1 + x).
Let us analyze this game assuming the general principle that at all stages Player II
will fold w.p. equal to the amount bet divided by the present pot size, which at time y is
dy/(2 + 2y + dy) dy/(2 + 2y). We will see that this strategy makes I indierent. If we
let Y denote the random time at which Player II folds, then the general principle above
implies that P{Y = y|Y y} = dy/(2 + 2y). The quantity on the left is known as the
failure rate of the distribution of Y . It is equal to G

(y)/(1G(y)), where G(y) denotes the

distribution function of Y . The above equation becomes G

(y)/(1 G(y)) = 1/(2(1 +y)),

which may also be written
d
dy
log(1 G(y)) =
1
2
d
dy
log(1 + y). Solving this dierential
equation for G(y) and using the boundary condition G(0) = 0, we nd
G(y) = 1
1

1 + y
for 0 < y < B. (2)
14
Since G(B) = 1, we see that G gives probability P
0
:= 1/

1 + B to the point B. This is

the probability that Player II never folds. If we take the expectation of W(x, Y ) for xed
x we nd
_
W(x, y) dG(y) =
_
x
0
(1 + y) dG(y) +
_
B
x
[P(1 + y) (1 P)(1 + x)] dG(y)
+
1

1 + B
[P(1 + B) (1 P)(1 +x)]
= (

1 + x 1) +P(

1 + B

1 +x)
(1 P)(1 +x)
_
1

1 +x

1 + B
_
+
1

1 + B
[P(1 + B) (1 P)(1 +x)]
= 2P

1 +B 1 = 2
P
P
0
1
(3)
independent of x. Since Player II can always keep her losses to be at most 1 by folding
immediately, we suspect that the value of the game is the minimum of these, namely
V =
_
2(P/P
0
) 1 if P P
0
1 if P P
0
(4)
To show that this is in fact the value, we consider a similar strategy for Player I. Let
F(x) denote the distribution function of random time at which Player I stops betting with
a losing card, and take F to have a mass
0
at x = 0, zero probability on the interval
(z, B], for some z B, and a positive density on the interval (0, z) of the following form:
F(x) =
0
+ c(1
1

1 + x
) for 0 x z, (5)
where the parameters
0
0, c > 0 and z > 0 are restricted so that F(z) = 1. For xed y
with 0 y z, the expectation of W(X, y) is
_
W(x, y) dF(x) = (1 + y)(1 (1 P)(
0
+ c)) + (1 P)(c
0
) (6)
and for z y B, we have
_
W(x, y) dF(x) = P(1 + y) (1 P)(
0
+ c(

1 + z 1)). (7)
Since (7) is increasing in y, Player II will never choose a y larger than z. Expression (6) is
a constant in y [0, z] provided (1 P)(
0
+ c) = 1 so Player I can keep the value of the
game to at least (1P)(c
0
). So I chooses
0
= (1/(1P))c and proceeds to choose c
and z to make c as large as possible, subject to the condition that F, which is now F(x) =
15
(1/(1 P)) (c/

1 + x) for 0 x z, is still a distribution function. If P

1 + B 1,
we may choose z = B and c = (P/(1 P))

1 + B, since
0
= (1 P

1 + B)/(1 P)
is still nonnegative. In this case the value is (1 P)(c
0
) = 2P

1 + B 1, the same
as found in (4). But if P

1 + B 1, we must choose
0
= 0 and c = 1/(1 P), when
z = (1/P
2
) 1. In this case the value is (1 P)(c
0
) = 1, also found in (4).
Summary of the Solution. Let P
0
= 1/

B + 1.
If P P
0
, then the value is 2(P/P
0
) 1. Player I has an optimal strategy,
F(x) =
1
1 P

P
1 P

1 + B

1 + x
for 0 x B,
with mass (1 P

1 + B)/(1 P) at x = 0. Player II has an optimal strategy

G(y) = 1
1

1 + y
for 0 y B,
with mass P
0
at y = B.
If P P
0
, then the value is 1. Player I has an optimal strategy,
F(x) =
1
1 P
_
1
1

1 +x
_
for 0 x
1
P
2
1.
Player II has an optimal strategy that gives mass 1 to the point y = 0 (i.e. fold immedi-
ately).
It may be noted that when P

1 + B 1, the hazard rate of Player Is optimal

strategy (i.e. F

(x)/(1 F(x))) is independent of P. Of special interest is the fact that if

Player I is blung, he will stop betting before reaching B; that is his strategy gives zero
probability to the point B.
5. Endgame with Information for Player II.
We modify Basic Endgame by allowing Player II to receive information on Player Is
card. This provides a model of situations that occur regularly in poker. For example,
suppose Player I needs a four or a nine to complete a straight, or a spade to complete a
ush. Player IIs hand may contain a four or a spade unknown to Player I, giving Player
II some hidden information on the chances that Player I has a straight or a ush. This
situation is modeled as follows.
The game is two-person, zero-sum. Player I receives a winning card, W, w.p. P, and
a losing card, L, w.p. 1 P, where 0 < P < 1. Then Player I observes his card and must
either check or bet a xed amount b > 0. If Player I checks, the game ends and Player I
wins or loses the ante a > 0, depending on whether he has W or L.
Suppose Player I bets. Then Player II is allowed to observe a random variable Y
whose distribution depends on the card received by Player I. We may assume without loss
16
of generality that the density of Y with respect to a -nite measure, , exists. Let f
W
(y)
denote the density of Y if I has W, and f
L
(y) denote the density of Y if I has L. Based on
the observation Y , II must either call or fold. If Player II folds, the game ends and I wins
a from II. If Player II calls, the game ends and Player I wins or loses a + b depending on
whether I has W or L. It is assumed that both players know a, b, P and f
W
, f
L
and .
It is clear that Player I may as well bet whenever he holds W. We assume without
loss of generality that the rules of the game require this. Then a pure strategy for I is
just a rule telling him what to do when he receives an L. Therefore, I has just two pure
strategies. He may bet with a losing card, the blu strategy, or he may check with a losing
card, the honest strategy.
A pure strategy for II is a rule telling her whether to fold or call for each possible
value of Y that may be observed. A behavioral strategy for Player II is a function (y)
satisfying 0 (y) 1 for all y, with the understanding that if II observes Y = y she folds
w.p. (y) and calls with probability 1 (y). In statistical parlance, is called a test.
Let E
W
(Y ) =
_
(y)f
W
(y) d(y). Then E
W
(Y ) represents the probability that II
folds given I has a high card and bets. Similarly, E
L
(Y ) =
_
(y)f
L
(y) d(y) represents
the probability that II folds given I has a low card and bets. The expected payo to Player
I if he uses one of his pure strategies and II uses is
V (blu , ) = P[(a + b)E
W
(1 (Y )) + aE
W
(Y )]
+ (1 P)[(a + b)E
L
(1 (Y )) + aE
L
(Y )]
= (2P 1)(a + b) PbE
W
(Y ) + (1 P)(2a + b)E
L
(Y )
V (honest , ) = P[(a + b)E
W
(1 (Y )) + aE
W
(Y )] (1 P)a (1)
= (2P 1)a + Pb PbE
W
(Y ).
If E
L
(Y ) is xed equal to some number , then II will choose subject to this
constraint to maximize E
W
(Y ) since both V (blu , ) and V (honest , ) are decreasing in
E
W
(Y ). Such a is called a best test of size for testing the hypothesis H
0
: f
L
(y)
against H
1
: f
W
(y). Player II may restrict her attention to such . The Neyman-Pearson
Lemma states that for xed 0 < 1, the test of the form

(y) =
_
_
_
0 if f
W
(y) < kf
L
(y)
if f
W
(y) = kf
L
(y)
1 if f
W
(y) > kf
L
(y)
(2)
is a best test of size when k 0 and are chosen so that E
L
(Y ) = . In addition,
corresponding to k = , the test

0
(y) =
_
1 if f
L
(y) = 0
0 if f
L
(y) > 0
(3)
is a best test of size = 0. Player II may restrict attention to the tests

for 0 1.
17
Player IIs optimal strategy is to choose so that the maximum of V (blu ,

) and
V (honest ,

) is a minimum. The dierence in the payos is V (blu ,

)V (honest ,

) =
(1 P)((2a + b) b). This is linear increasing in with a unique root,

0
= b/(2a + b) (4)
in (0, 1). Moreover, E
W

(Y ) is nondecreasing in (in fact concave, and increasing as

long as it is less than 1), so V (honest ,

) is nonincreasing in . Hence, there are only two

cases, depending on the sign of s
r
, the right slope of V (blu ,

) at =
0
.
Case 1: s
r
> 0. In this case,

0
is optimal for Player II. It is interesting to note that
this strategy does not depend on P. If s
l
denotes the left slope of V (honest ,

), then
s
l
0 and an optimal strategy for Player I is to mix blu and honest in the proportions
|s
l
| and s
r
.
Case 2: s
r
0. In this case, it is optimal for Player I to blu all the time. If
1
denotes
the value of that minimizes V (blu ,

), then
1
>
0
and

1
is an optimal strategy
for Player II. In this case, the optimal strategy for II may be taken to be nonrandomized.
As noted in Ferguson (1967), a sucient condition for s
r
to be positive is that P < 1/2.
This may be seen as follows. Let g() = E
W

(Y ). Using E
L

(Y ) = in (1), we have
V (blu ,

) = (2P 1)(a+b) Pbg() +(1 P)(2a+b). Letting g

() denote the right

slope of g(), we have s
r
= Pbg

(
0
)+(1P)(2a+b) = P(2a+b)[(1P)/P
0
g

(
0
).
But since g is concave and g(0) 0, we have g

(
0
) g(
0
)/
0
. Therefore, if P < 1/2,
we have s
r
> 0.
It is interesting to compare this solution with that of Basic Endgame. In the latter, the
cuto between Case 1 and Case 2 is P = (2a +b)/(2a +2b), and while the cuto between
Case 1 and Case 2 in the above solution may be greater or less than this, it is always at least
1/2. In Case 1 of Basic Endgame, Player II always folds with probability
0
= b/(2a +b),
while in the above solution, II folds w.p.
0
when Player I has L (i.e. E
L

0
(Y ) =
0
),
and w.p. E
W

0
(Y ) >
0
when Player I has W. In other words, Player I can detect that
Player II is using his information only by noting that II is calling Is winning hands less
often.
Although Player Is strategy depends on P, one feature of this strategy is independent
of P, namely the probability that he has L given that he has bet in Case 1. Using
s
r
= Pbg

(
0
) +(1 P)(2a +b) and s
l
= Pbg

(
0
), we nd the probability that I bets
with a L is |s
l
|/(s
r
+ |s
l
|) = Pbg

(
0
)/((1 P)(2a + b)) = P
0
g

(
0
)/(1 P). Thus we
have P(I has L|I bets) =
0
g

(
0
)/(1 +
0
g

(
0
)), independent of P.
The Binary Case.
As a simple illustration of the computations involved, we consider the case where
the observation Y takes only two values, say 0 and 1, and f
L
(1) = p
L
, f
L
(0) = 1 p
L
,
f
W
(1) = p
W
and f
W
(0) = 1 p
W
. If p
L
= p
W
, then the observation of Y gives no
information about the card Player I has, and the problem reduces to Basic Endgame.
Otherwise, we may assume without loss of generality that p
W
> p
L
. Then Player II
18
prefers to see Y = 0 since then it is less likely Player I has W. We refer to Y = 0 as good
hands of Player II, and Y = 1 as poor hands.
First we nd the best test of size . For 0 < p
L
, we must set k = p
W
/p
L
and we
have

(y) =
_
0 if y = 0
if y = 1
(5)
where for E
L

(Y ) = p
L
to be equal to , we require = /p
L
. This is the strategy:
Call with a good hand. With a poor hand, fold w.p. = /p
L
and call w.p. 1 .
For p
L
< 1, we must set k = (1 p
W
)/(1 p
L
) and we have

(y) =
_
if y = 0
1 if y = 1
(6)
where for E
L

(Y ) = p
L
+ (1 p
L
) to be equal to , we require = ( p
L
)/(1 p
L
).
This is the strategy: Fold with a poor hand. With a good hand, fold with probability
= ( p
L
)/(1 p
L
) and call w.p. 1 .
For = p
L
, we may put k = 1 and nd

(y) =
_
0 if y = 0
1 if y = 1.
(7)
This is the strategy: Call with a good hand. Fold with a poor hand.
To implement the above solution, we need s
r
, the right slope of V (blu ,

), and s
l
,
the left slope of V (honest ,

). We rst nd
E
W

(Y ) =
_
p
W
/p
L
for 0 p
L
p
W
+ ( p
L
)(1 p
W
)/(1 p
L
) for p
L
< 1.
(8)
From this we may nd
s
r
() =
_
Pbp
W
/p
L
+ (1 P)(2a +b) for 0 < p
L
Pb(1 p
W
)/(1 p
L
) + (1 P)(2a + b) for p
L
< 1.
(9)
and
s
l
() =
_
bPp
W
/p
L
for 0 < p
L
bP(1 p
W
)/(1 p
L
) for p
L
< 1.
(10)
We may state the solution as follows. Let

0
=
b
2a + b

0
=
p
L
p
W
1 P
P

1
=
1 p
L
1 p
W
1 P
P
and note that
0
<
1
since we are assuming p
L
< p
W
.
19
1. If
0

1
, then
it is optimal for I to blu, and
it is optimal for II to fold.
The value is a.
2. If
0
<
1
and
0
p
L
, then
it is optimal for I to blu w.p.
0
/
1
, and
it is optimal for II to fold with a poor hand and
to fold w.p. (
0
p
L
)/(1 p
L
) with a good hand.
The value is a[1 2(1 P)(1 (
0
/
1
))].
3. If
0

0
<
1
and
0
< p
L
, then
it is optimal for I to blu, and
it is optimal for II to fold with a poor hand and
to call with a good hand.
The value is (2P 1)(a + b) p
L
(1 P)b((
0
/
0
) 1)(2a +b).
4. If
0
<
0
and
0
< p
L
, then
it is optimal for I to blu w.p.
0
/
0
, and
it is optimal for II to call with a good hand and
to fold w.p.
0
/p
L
with a poor hand.
The value is (2P 1)(a + b) +b(1 P)(1 (
0
/
0
)).
Case 1
Case 2 Case 4
Case 3
0 1
0
1
P
p
L

0
Figure 3.
Figure 3 shows the regions of (
0
, P) for the four cases when p
L
= .35 and p
W
= .85.
The mixed strategy cases (Case 2 and Case 4) may be derived from the solution to
Basic Endgame as follows. Player I will be indierent between betting and folding with an L
if Player II folds with overall probability
0
. So II should choose her strategy in such a way
that her overall probability of folding is
0
. In Case 4 if I has an L, II folds w.p. p
L
(
0
/p
L
)
which is
0
. Also in Case 2 if I has an L, II folds w.p. p
L
+ (1 p
L
)(
0
p
L
)/(1 p
L
)
which is also
0
.
Similarly for Player I. In Case 4, he should play to keep II indierent if she has
a poor hand, while in Case 2 he should play to keep her indierent if she has a good
hand. If Player II has a poor hand, she evaluates the probability that I has a W as
20
P
1
= Pp
W
/(Pp
W
+ (1 P)p
L
) = 1/(1 +
0
), so I should replace P in his strategy for
Basic Endgame with P
1
. This is in fact what he does since
0
P
1
/(1 P
1
) =
0
/
0
.
If Player II has a good hand, she evaluates the probability that I has a W as P
1
=
P(1 p
W
) +(1 P)(1 p
L
), so I should blu w.p.
0
P
0
/(1 P
0
) =
0
/
1
, which is what
he does.
Examples.
Example 1. Consider the introductory example of the last round of stud poker with Player
I having a 5, 6, 7 and 8, not suited, showing and Player II having 2, 3, K and K showing.
For the purposes of this example we shall take the probabilities, P, p
L
, and p
W
, to be those
assuming that the down card was dealt last. Since there are exactly 8 cards out of the
remaining 44 cards that can give Player I a straight, we have P = 8/44 = 2/11. Similarly,
p
L
= 35/43 and p
W
= 36/43. Then we may compute
0
= (35/36) (9/2) = 35/8, and

1
= (8/7) (9/2) = 36/7. Suppose we are playing pot limit poker so that b = 2a and

0
= 1/2. Then
0
<
0
and
0
< p
L
so we are in Case 4. If he doesnt have the straight,
Player I should blu w.p.
0
/
0
= 4/35. With a good card (4 or 9), Player II should call.
With a poor card, she should fold w.p.
0
/p
L
= 43/70.
Example 2. An interesting feature of this game is that, unlike Basic Endgame, it may be
optimal to bet less than the maximum allowable bet. One can guess that this will occur in
situations where Player II can sometimes, though rarely, get sure information that Player
I has a losing card. I can gain by blung reasonably often, but with a high enough bet
size this cannot be optimal because Player II will call only when she knows she will win.
Here is an example.
Suppose P = 1/2, p
W
= 1 and p
L
= 1 for some small > 0. This means that
Player II will occasionally (probability ) get accurate information that Player I has a
losing card. Then
1
= so case 1 does not occur. Moreover
0
= p
L
so Case 3 does not
occur. If
0
p
L
(Case 2), then I does not blu and the value is 0. If
0
< p
L
(Case 4),
then I blus w.p.
0
/p
L
, and II calls with a good hand and folds w.p.
0
/p
L
with a poor
hand. The value is (b/2)(1 (
0
/p
L
)) = b(p
L
(b/(2a +b))/(2p
L
). This is positive for
0
in the interval [0, p
L
], and has a maximum at b = 2a((1/

) 1). For example, if = 1/4,

then the optimal bet size is b = 2a, the bet size for pot limit poker.
Example 3. The phenomenon of nite optimal bet size is frequent in this model. We take
another example with continuous observations for Player II. Suppose the random variable
Y has a uniform distribution when Player I has a losing card, i.e. f
L
(y) = 1 on the set
(0, 1), and suppose Y has a triangular distribution with f
W
(y) = 2(1 y) on (0, 1) when
I has a winning card. The best tests of size are simply

(y) =
_
1 if y
0 if y > .
Then g() = E

(Y ) = (2 ). We are in the case of a positive slope of V (blu ,

)
at =
0
= b/(2a + b), so

0
is optimal for Player II. The value may be computed
as V (honest ,

0
) = (2P 1)a + 4a
2
Pb/(2a + b)
2
. As a function of b, the value has a
maximum at b = 2a, again pot limit poker.
21
6. Endgame with Imprecise Information for Player I.
We consider the generalization of basic endgame achieved by allowing the card received
by Player I to indicate only his probability of winning. This models situations in which
there are still cards to be drawn, which with some probability may change a winning hand
into a losing one. We model this by denoting the card Player I receives by the probability of
win. Thus, if he receives a card marked p, then p is his probability of win. The distribution
of p on [0, 1] is arbitrary however, and may be discrete or continuous. The distribution
function of p is denoted by F. In Basic Endgame, p is either 0 or 1, and the probability
that p = 1 is P. It is assumed that both players know F but only Player I learns the value
of p. Other than this, the form of the game with one possible bet and then a call or fold
is the same as it was in Basic Endgame.
Player II has the same two pure strategies of fold or call. But now Player I may use
a distinct strategy for each card he receives. A mixed strategy for Player I is a function,
(p), that denotes the probability that Player I bets when he observes that his probability
of win is p. We have 0 (p) 1 for all p. The expected payo to Player I if he uses
and II uses one of her pure strategies is
V (, fold) = E[(p)a (1 (p))a(2p 1)] = 2aE(1 p)(p) + a(2 1)
V (, call ) = E[(p)(2p 1)(a + b) +(1 (p))a(2p 1)]
= bE(p) 2bE(1 p)(p) + a(2 1)
where = Ep denotes the overall probability that I wins. Out of the class of (with
0 (p) 1) such that E(1p)(p) is held xed, I will choose to maximize E(p), since
that will maximize V (, call ) with V (, fold) held xed. This is equivalent to maximizing
Ep(p) out of the class of such that E(p) is held xed equal to some number . By the
Neyman-Pearson Lemma, we may nd such a of the form

(p) =
_
1 if p > k
if p = k
0 if p < k
where k and are chosen so that E(p) = . Player I may restrict his choice of strategy to
the class of

. This shows that Player I will bet with cards having a high value of p and
check with cards having a low value of p. Any randomization will occur only for those p on
the boundary between betting and checking. If the distribution function F is continuous,
Player I will have an optimal pure strategy. Note that both V (, fold) and V (, call ) are
increased by putting (p) = 1 for p 1/2. Therefore in the model, Player I should always
bet if his probability of win is at least 1/2.
Let g() = Ep

(p). Then g() is continuous, concave and increasing on [0, 1]. It is

apparent that V (

, fold) is a nondecreasing convex function of in [0,1], while V (

, call )
is concave and increasing in as long as k > 1/2 and nonincreasing thereafter. Moreover,
these functions are equal at = 0 and there is at most one other value of in (0, 1] at
which they are equal.
22
From these observations, we may write down the solution of the game. The form of
the optimal strategies diers in three dierent regions of the space of parameters.
Case 1. (2a +b)/(2a + 2b). In this case, V (

, fold) V (

, call ) for all values of .

Therefore, folding is optimal for II, betting always is optimal for I and the value is a.
Let
0
= P(p 1/2) and if
0
> 0, let
0
= E(p|p > 1/2) = Ep

0
(p)/
0
. Note that
V (

, call ) takes on its maximum value at =

0
and that

0
is the indicator of the set
{p 1/2}. Also note that
0
.
Case 2.
0
= 0, or
0
> 0 and
0
(2a + b)/(2a + 2b). In this case, V (

0
, call )
V (

0
, fold), so that the maximum of the minimum of these two functions occurs at
0
.
Therefore, calling is optimal for II, and

0
is optimal for I. If
0
= 0, then V (

, call ) is
decreasing in , so that checking always is optimal for I. Otherwise, the strategy that bets
if and only if p 1/2 is optimal for I. The value is
0
(2b
0
1) + a(2 1).
The last case is the main case, requiring mixed strategies for Player II. The minimax
value of occurs at the point of intersection of V (

, fold) and V (

, call ), call it
1
.
Then
1
satises the equation g(
1
)/
1
= (2a + b)/(2a + 2b).
Case 3. < (2a + b)/(2a + 2b) <
0
. In this case, the maximin occurs at
1
. Therefore,
mixing fold and call in the proportions b 2bg

(
1
) : 2a 2ag

(
1
) is optimal for II, and

1
is optimal for I. (Here g

() represents any value between the left derivative and the

right derivative at .) The value is 2a(
1
g(
1
)) + a(2 1).
The Binary Case.
As an illustration, consider the case where F gives mass to only two points. Suppose
F gives mass to p
H
and mass 1 to p
L
, where p
L
< p
H
. Thus, Player I receives one
of two cards, a high card giving him probability p
H
of winning, and a low card giving him
probability p
L
of winning. The probability he receives a high card is 0 < < 1. The
probability that I wins with the card he receives is = p
H
+ (1 )p
L
. Basic Endgame
occurs when p
L
= 0, p
H
= 1, and = = P. We let Q
0
= (2a +b)/(2a + 2b).
If both p
L
and p
H
are in [0, 1/2], then it is optimal for I to check with either card and
it is optimal for II to call (Case 2 with
0
= 0). The value is a(2 1).
If both p
L
and p
H
are in [1/2, 1], then
0
= P(p 1/2) = 1 and it is optimal for I to
bet with either card. It is optimal for II to fold if Q
0
(Case 1, value = a) and to call
if Q
0
(Case 2, value = (a + b)(2 1)).
Suppose now that p
L
< 1/2 < p
H
. Then
0
= and
0
= E(p|p > 1/2) = p
H
.
If Q
0
, then it is optimal for II to fold and for I to bet with either card (Case
1, value = a). If p
H
Q
0
, then it is optimal for II to call and it is optimal for I to
bet with a high card and check with a low card (Case 2 with 0 <
0
< 1 and value
= a(2 1) + b(2p
H
1)).
There remains to consider < Q
0
< p
H
(Case 3). This requires computation of

23
and the payo functions. The strategies

are easily found.

For ,

(p) =
_
/ if p = p
H
0 if p = p
L
For ,

(p) =
_
1 if p = p
H
( )/(1 ) if p = p
L
From this we may nd g() = Ep

(p).
g() =
_
p
H
for
p
H
+( )p
L
for
and from this we may nd the expected payos.
V (

, fold) = 2a( g()) + a(2 1)

V (

, call ) = b(2g() ) + a(2 1).

The value,
1
, at which these intersect is that value of in [, 1] such that 2a(g()) =
b(2g() ). This reduces to

1
=
p
H
p
L
Q
0
p
L
.
Is optimal strategy is

1
. This strategy calls for betting with a high card and randomizing
with a low card, betting w.p. (
1
)/(1) and checking otherwise. The optimal strategy
for Player II is to mix calling and folding in proportions |s
c
| : s
f
, where s
c
and s
f
are the
slopes at
1
of the payo functions for calling and folding. Since s
1
= 2a(1 p
L
) and
s
2
= b(2p
L
1), we have that II should call with probability
b(1 2p
l
)
(2a +b) (2a + 2b)p
l
=
1 Q
0
Q
0
1 2p
L
1 p
L
and fold otherwise. The value is
2a(
1
g(
1
)) +a(2 1) = 2a
(p
H
p
L
)(1 Q
0
)
Q
0
p
L
+a(2 1).
It is interesting to note that IIs optimal strategy in Case 3 (the general case) does
not depend on or p
H
. Player IIs strategy depends only on Q
0
and p
L
. It should also be
noted that in all cases Player Is optimal strategy does not randomize with a high card.
24
References.
Ankeny, Nesmith C. (1981) Poker Strategy: Winning with Game Theory, Basic Books,
New York.
R. J. Aumann and M. B. Mashler (1995) Repeated Games with Incomplete Information,
The MIT Press, Cambridge Mass.
A. Ba nos (1968) On pseudo-games. Ann. Math. Statist. 39, 1932-1945.
R. Bellman and D. Blackwell (1949) Some two-person games involving blung Proc. Nat.
Acad. Sci. 35, 600-605.
R. Bellman (1952) On games involving blung, Rendiconti del Circolo Math. di Palermo
Ser. 2, Vol. 1 139-156.

Emile Borel (1938) Traite du Calcul des Probabilites et de ses Applications Volume IV,
Fascicule 2, Applications aux jeux de hasard, Gauthier-Villars, Paris.
Doyle Texas Dolly Brunson (1978) Super/System, A Course in Power Poker, B & G
Publishing Co. Inc., Las Vegas.
Mike Caro (1984) Mike Caros Book of Tells, Gambling Times Incorporated.
Bill Chen and Jerrod Ankenman (2006) The Mathematics of Poker, ConJelCo LLC.
W. H. Cutler (1975) An optimal strategy for pot-limit poker, Amer. Math. Mo. 82, 368-
376.
W. H. Cutler (1976) End-Game Poker, Preprint.
T. S. Ferguson (1967) Mathematical Statistics, A Decision Theoretic Approach, Academic
Press, New York.
C. Ferguson and T.S. Ferguson (2003) On the Borel and von Neumann poker models,
Game Theory and Applications, 9, 17-32, Nova Sci. Publ., New York.
C. Ferguson, T.S. Ferguson and C. Gawargy (2007) U(0,1) two-person poker models, Game
Theory and Applications, 12, Nova Sci. Publ., New York, to appear.
L. Friedman (1971) Optimal blung strategies in poker, Man. Sci. 17, B764-B771.
D. B. Gillies, J. P. Mayberry and J. von Neumann (1953) Two variants of poker, Contrib.
Theory of Games II 13-50.
A. J. Goldman and J. J. Stone (1960a) A symmetric continuous poker model, J. Res. Nat.
Bur. Standards 64B, 35-40.
A. J. Goldman and J. J. Stone (1960b) A continuous poker game, Duke Math. J. 27, 41-53.
Dan Harrington and Bill Robertie (2004-2006) Harrington on Holdem: Expert Strategy
for No-Limit Tournaments, vol. 1, 2 and 3, Two Plus Two Publishing LLC.
S. Karlin and R. Restrepo (1957) Multistage poker models, Contrib. Theory of Games III,
337-363.
25
S. Karlin (1959) Mathematical Methods and Theory in Games, Programming and Eco-
nomics, Vol. II, Addison-Wesley, 238-278.
H. W. Kuhn (1950) A simplied two-person poker, Contrib. Theory of Games I, 97-104.
H. W. Kuhn (1953) Extensive games and the problem of information, Contrib. Theory
of Games II, 193-216. (Reprinted in Classics in Game Theory, (1997), H. W. Kuhn
(ed.), Princeton University Press.)
N. Megiddo (1980) On repeated games with incomplete information played by non-
Bayesian players, Int. J. Game Theory 9, 157-167.
J. F. Nash and L. S. Shapley (1950) A simple 3-person poker game, Contrib. Theory of
Games I 105-116.
D. J. Newman (1959) A model for real poker, Oper. Res. 7, 557-560.
W. E. Pruitt (1961) A class of dynamic games, Nav. Res. Log. Quart. 8, 55-78.
David Sklansky (1987) The Theory of Poker, Two Plus Two Publishing, Nevada.
David Sklansky and Mason Malmuth (1988) Holdem Poker for Advanced Players, Two
Plus Two Publishing, Nevada.
D. Sklansky, M. Malmuth and R. Zee (1989) Seven Card Stud for Advanced Players, Two
Plus Two Publishing, Nevada.
Sylvain Sorin (2000) A First Course on Zero-Sum Repeated Games, Mathematiques &
Applications, 37, Springer.
J. von Neumann and O. Morgenstern (1944) Theory of Games and Economic Behavior,
John Wiley 1964, 186-219.
Norman Zadeh (1977) Computation of optimal poker strategies, Operations Research 25,
541-561.
Ray Zee (1992) High-Low-Split Poker for Advanced Players, Two Plus Two Publishing,
Nevada.
26