PATTERN MATCHING PROBABILITIES AND

PARADOXES AS A NEW VARIATION ON

PENNEYS COIN GAME
Yutaka Nishiyama
Department of Business Information,
Faculty of Information Management,
Osaka University of Economics,
2, Osumi Higashiyodogawa Osaka, 533-8533, Japan
[email protected]
April 1, 2010
Abstract: This paper gives an outline of an interesting probability game related to coin arrangements discovered by Walter Penney in 1969, and explains
John Conways concept of leading numbers and his algorithm for using them
to automatically calculate mutual odds of winning. I will show that Conways
elegant algorithm is correct by using average waiting times as proposed by
Stanley Collings, and discuss the eect of applying the rules of this coin game
to a card game instead.
AMS Subject Classification: 60C05, 91A60, 97A20
Key Words: Penney Ante, pattern matching, odds, non-transitive relation,
Conways Algorithm, average waiting time

1. Penney Ante
One day I received an email from a British acquaintance, Steve Humble. In
some excitement, he told me that a magician named Derren Brown was introducing an interesting game on the television. I was a little dubious upon
hearing the word magician , but after close examination I realized that the
game had a mathematical background and was an interesting exercise in probability. The game is not a well known one, so I would like to introduce it to
you.
Assume a coin flip with equal probability of coming up heads or tails. The
game is played by Players A and B each selecting a sequence of three flips,
flipping the coin repeatedly, and seeing whose sequence comes up first. For
1

example, assume that Player A selected heads-heads-heads (HHH) and

Player B selected tails-heads-heads (THH). Then, assume that the coin is
flipped repeatedly, resulting in a sequence like the following:
HTHTHHHHTHHHTTTTHTHH
The player whose sequence showed up first, HHH or THH, would be declared
the winner.
Both heads and tails have an equal probability of appearing. There are
23 = 8 possible sequences of three flips, and so each sequence has a 1 in 8
chance of appearing. It would seem that both players have an equal chance
of winning, regardless of the sequence chosen, but would you believe that the
player who selected THH is seven times more likely to win than the player who
selected HHH?

2. Non-transitive relations
This problem in probability has been around for some time, but is not widely
known. Walter Penney first presented it in an article, only ten lines long, in
the Journal of Recreational Mathematics in 1969 [1]. Martin Gardner later
provided a more detailed description in his Mathematical Games column
in the October 1974 issue of Scientific American [2].
When the game is played using patterns of length 3, no matter what sequence Player A chooses, Player B can always make a winning selection. Table
1 shows the eight possible selections for Player A and the winning selection
for Player B. According to the table, selecting THH in response to HHH gives
7:1 odds of winning, selecting THH in response to HHT gives 3:1 odds, and
selecting HHT in response to HTH gives 2:1 odds.
As choice
HHH
HHT
HTH
HTT
THH
THT
TTH
TTT

Bs choice
THH
THH
HHT
HHT
TTH
TTH
HTT
HTT

Odds in favor of B
7 to 1
3 to 1
2 to 1
2 to 1
2 to 1
2 to 1
3 to 1
7 to 1

Table 1. Odds in favor of B.

Table 1 shows that THH is stronger than HHH, TTH is stronger than THH,
HTT is stronger than TTH, HHT is stronger than HTT, and THH is stronger
than HHT. In other words, Player A does not have a strongest selection. This

Figure 1: Relationship between the eight patterns.

relationship is shown in Figure 1 as a looped relationship among the four inner
patterns.
Mathematically speaking, the following is referred to as a transitive relationship:
If A B and B C then A C.
Rephrasing this in terms of a game, then we would have If player A beats
player B and player B beats player C, then player A beats player C . But is
this a true proposition? Gardner poses that this transitive relationship does
not hold in the Penney Ante game, that it forms non-transitive relations [2].
Such three-way relationships may be unfamiliar in the West, but they are well
known in Japan, for example in the old game of rock-scissors-paper. In that
game, it does not follow that because rock beats scissors and scissors beat paper
that therefore rock beats paper. Instead, rock loses to paper, therefore failing
to uphold the transitive relationship. Similarly, transitive relationships do not
hold in Penney Anteas in rock-scissors-paper, there is no best play and
the second player can always win.
The following rules exist for Player Bs selection, based on the selection
made by Player A. Assume that Player A chooses HHH. In that case, the
second call is for H. Flip that selection to a T, and move it to the beginning of
Player Bs sequence. Next, take the first two calls in Player As sequence, and
take those as the second and third call in Player Bs. Doing so results in THH
for Player B, a selection that will win against Player A. Player As third call is
not considered.
HHH THH
Readers should confirm that each of the selections for Player B in Table 1
follow this rule. This rule for the selection by Player B based on the selection
by Player A works for runs of length 3, but only for runs of length 3the same
algorithm will not necessarily work for other lengths.
Let us confirm the probabilities listed in Table 1. We will first calculate the
probability that THH will appear before HHH. Assume that Player A selects
HHH. Should that sequence appear in the first three tosses, then Player A will
win. In any other situation, Player B will win. The probability that the first

 1 3

1
= ; thus,
2
8
1
7
P (THH before HHH)=1 = .
8
8
7 1
Player Bs odds are therefore = 7, or 7:1.
8 8
Let us next calculate the probability that THH will appear before HHT.
The probability of flipping HHT, or HHHT, or HHHHT is
1
1
1
1/8
1
+
+
+ =
= .
8 16 32
1 1/2
4
This gives the probability of HHT winning. The probability of THH winning
is therefore the complementary event, and thus
1
3
P (THH before HHT)=1 = .
4
4
3 1
Player Bs odds are therefore = 3, or 3:1.
4 4
Let us next calculate the probability that HHT will appear before HTH.
x=P (HHT before HTH)
We can ignore all initial flips of T. Taking the first flip as H, if the next flip is
also H then the probability of HHT appearing is 1/2. If a T is flipped, however,
the sequence is reset and must wait for the next H flip. That then is equivalent
to x. Shown as an equation, we have
1 1 1
x = + x,
2 2 2
and solving for x gives
2
x= .
3
2 1
Player Bs odds are therefore = 2, or 2:1. Continuing in this manner,
3 3
the individual odds for each pair can be calculated, but Conways Algorithm
provides a general solution. The following section describes that algorithm.
three tosses will come up HHH is

3. Conways Algorithm
John Conway proposed an algorithm that can be used to easily calculate the
odds of Player B beating Player A. Conway introduced leading numbers as an
index indicating the degree of pattern overlap. Leading numbers are also an
index of the level of repetition of a given pattern within a preceding pattern,
and are called Conway Numbers after their inventor.
Let us now examine Conways algorithm, taking as an example the case
where Player A selected HHH, and Player B selected THH.
A HHH
B T HH

Let us first find the Conway Number for Player A with regards to Player
As own selection. First, we place AHHH over AHHH. Next, if the HHH of
the upper sequence is the same as the lower sequence, then we place a 1 over
the first element and a 0 otherwise.
1
A HHH
A HHH

Next, remove the leading H from the upper HHH sequence, and shift it to
the left, aligning the leading elements. Then, compare the HH to the first two
elements in the lower sequence, and if they are the same place a 1 above the
leading element of the HH, or a 0 otherwise.
1

HH

HHH

HHH

Repeat this procedure through to the last element of the upper sequence,
and compile the results as follows.
1 1 1
A HHH
A HHH

The resulting binary number 111 is an index of the overlap of A with regards
to A, and the notation for the Conway number is AA=111. There are a total of
2 2 = 4 permutations of Conway numbers for a given pair of upper and lower
sequences. The following are the Conway numbers for pairs AA, AB, BB, and
BA.
111 = 7

000 = 0

A HHH

A HHH

A HHH

B T HH

100 = 4

011 = 3

B T HH

B T HH

B T HH

A HHH

Conway numbers are derived as binary numbers, but converting to decimal

gives
5

AA = 7, AB = 0, BB = 4, BA = 3.
Using the four Conway numbers, the odds of Player B winning are given by
the equation
(AA AB)/(BB BA).
Replacing the above four values in the equation, we have
AA AB
70
=
= 7.
BB BA
43
Player Bs odds are given as 7, matching with the first entry in Table 1.
Taking Player As probability of winning as p, and Player Bs probability
of winning as q, we obtain
1
1
7
7
p=
= , q=
= .
1+7
8
1+7
8
Both Player A and Player B have eight possible selections. Because the
players make their selections independently, there are 8 8 = 64 possible
matches, and the odds for each are automatically generated by the Conway
algorithm as shown in Table 2. Taking the best odds for Player B with respect
to Player As selection gives the results listed in Table 1. Conways algorithm is
very powerful, in that it can give probabilities not only for sequences of length
3, but for those of any length, or even for sequences of dissimilar length.
A
HHH
HHH

HHT

HTH

HTT

THH

THT

TTH

TTT

1/2

2/5

2/5

1/8

5/12

3/10

1/2

2/3

2/3

1/4

5/8

1/2

7/10

1/2

1/2

1/2

3/8

7/12

1/2

1/2

3/4

7/8

1/2

1/3

3/5

1/3

3/5

HHT

1/2

HTH

3/5

1/3

HTT

3/5

1/3

1/2

THH

7/8

3/4

1/2

1/2

THT

7/12

3/8

1/2

1/2

1/2

TTH

7/10

1/2

5/8

1/4

2/3

2/3

TTT

1/2

3/10

5/12

1/8

2/5

2/5

1/2
1/2

Table 2. Probabilities of winning for players A and B

(for sequences of length 3).

4. Proof according to average wait time

The algorithm for determining odds using Conway Numbers is indeed elegant
and powerful. Let us consider precisely why one can calculate win probabilities from Conway Numbers. As elegant an algorithm as it is, Conway himself
did not propose a proof in the paper in which he demonstrated it. Searching
through the literature, I came across an easily understood proof that uses Stanley Collingss concept of average waiting time [3]. Collingss proof is developed
from three theorems.
Theorem X
The expected waiting time to get a given sequence of integers from the set
{1, 2, ..., k} is k.AA, where AA stands for the corresponding Conway Number
considering entirely of 0s and 1s and interpreted in the scale of k.
The faces of a coin {H, T } can be mapped to {1, 2}. AHHH, BTHH is
therefore mapped to A111, B211, and their Conway Numbers are AA = 111,
BB = 100. Converting the binary numbers to decimal we have AA = 7, BB =
4, and doubling them we have average wait times of 14 and 8.
Theorem Y
Given a sequence B to start with, the expected further number of rolls required to complete or produce the A sequence is k.AAk.BA.
For example, given the sequence B211, because AA = 7, BB = 3, the
average wait time for A111 to appear is k.AA k.BA = 2 7 2 3 = 8.
Theorem Z
The odds that the B sequence precedes the A sequence are given by (AAAB):(BB-BA).
When the average wait time of A is larger than that of B, then the dierence
is expressed as
k.(AA BB).
On the other hand, from Theorem Y we have that the average wait time for A
to appear when B precedes A is
k.AA k.BA.
Similarly, the average wait time for B to appear when A precedes B is
k.BB k.AB.
Taking the probability of Player A winning as p and the probability of
Player B winning as q, q will occur with probability k.AA k.BA, and p will
occur with probability k.BB k.AB, and so the three terms above are related
by
k(AA BB) = q(k.AA k.BA) p(k.BB k.AB).
7

Dividing both sides of the equation by k, we obtain

AA BB = q.AA q.BA p.BB + p.AB
and
(1 q)AA (1 p)BB = p.AB q.BA.
Because p + q = 1, we have
p.AA q.BB = p.AB q.BA,
p(AA AB) = q(BB BA),
and
q
AA AB
=
.
p
BB BA
The last equation gives the odds of Player B beating Player A, and proves
Conways algorithm.

5. Application to a card game

In the preceding sections, we examined Penneys Ante using heads and tails in
a coin flipping game; however, Steve Humble and I have been discussing how
this might be applied to a more familiar context, such as a card game. While
coin tosses give equal probabilities of heads or tails, we cannot guarantee that
the probabilities are exactly 1/2. The situation is the same with the red/black
results of a roulette table. Coin flips are also problematic in that the action
of flipping a coin is somewhat dicult, and one must maintain a record of the
results. We therefore turned our consideration to packs of cards.
There are 52 cards in a deck, and the 26 black (spades and clubs) and 26
red (hearts and diamonds) cards can be used to substitute for the heads or
tails results of a coin flip. Let B represent black cards, and R red cards, so
that HHH versus THH when flipping coins could be represented as BBB versus
RBB when using cards.
The advantage of using cards lies in their ease of manipulation, and the
fact that one does not need to keep track of results. They can also be well
randomized. If all 26 red and black cards are used, then there is a 1/2 chance
of choosing either color from the 52 total cards. The 52 cards also allow an
appropriate number of trials. It is dicult to determine the end of a game that
uses coin flips, but one can simply declare an end when all 52 cards are used.
The only problem to consider is if they will form a game. The number of trials
that can be performed with 52 cards is given by Collingss theory [3], as shown
below, which states that the average wait time E(N) before sequence A or B
will appear is
E(N ) = k(p.AA + q.BA)
or
E(N ) = k(p.AB + q.BB).
The average wait time before A or B appears can be given as a group A B.

As shown in Table 3, the values of E(N) dier according to matches, giving

7 for BBB versus RBB, and 16/3 for BRR versus BBR. Here, the average
wait time E(N) refers to the number of tricks played per match. Dividing this
number by the number of cards (52) therefore gives the number of trials. The
theoretical values give a distribution of 7.49.8 trials, which matches well with
the statistical values we found when we tried this with actual cards (710).
A
BBB
BBR
BRB
BRR
RBB
RBR
RRB
RRR

B
RBB
RBB
BBR
BBR
RRB
RRB
BRR
BRR

AA
7
4
5
4
4
5
4
7

BB
4
4
4
4
4
4
4
4

AB
0
1
1
0
0
1
1
0

BA
3
3
2
2
2
2
3
3

p
1/8
1/4
1/3
1/3
1/3
1/3
1/4
1/8

q
7/8
3/4
2/3
2/3
2/3
2/3
3/4
7/8

E(N)
7
13/2
6
16/3
16/3
6
13/2
7

Trials
7.4
8
8.7
9.8
9.8
8.7
8
7.4

Table 3. Average trials using a pack of cards.

(E(N): Average wait times)
From Table 3 we see that the probabilities for BBB and RBB are 1/8 and
7/8, giving RBB overwhelming odds of 7:1 in a match versus BBB. In the
match RBR versus RRB, however, the odds for RRB are only 2:1. Such odds
imply no guarantee of winning in a single match. We would therefore like to
perform n trials to increase the odds of RRB winning. For example, in a match
of three tricks, Player B will win by taking two tricks, and in a match of seven
tricks Player B will win by taking four. Showing this as an equation gives us
the following. Taking P (RRB, n) over n trials, the probability of an RRB win
is
P (RRB,7)=
2 7
2 6 1
2 5 1 2
2 4 1 3
+ 7 C6
+ 7 C5
+ 7 C4
= 0.827,
7 C7
3
3
3
3
3
3
3
a clear advantage for Player B over the 0.667 probability for a single trial. Playing the game using cards over 79 trials, we can expect a very high probability
that Player B will win.
We have shown in this article that Penneys game can be applied to the use
of cards, and that cards are better suited to its application than coin tosses. We
would like to present this new card-based version as the Humble-Nishiyama
Randomness Game, and hope that our readers will try it with a deck of
cards at home.

References
[1] Walter Penney, 95. Penney-Ante, Journal of Recreational Mathematics, 2,
(1969), 241.
[2] Martin Gardner, Mathematical Games, Scientific American, October
1974, 231, No.4, 120-125.
[3] Stanley Collings, Coin Sequence Probabilities and Paradoxes, Bulletin of the Institute of Mathematics and its Applications, 18, November/December 1982, 227-232.

10