Вы находитесь на странице: 1из 59

Universit`a Commerciale Luigi Bocconi

Graduate School
Master of Science in Economics and Social Sciences
Equivalent Extensive Game Forms
Advisor: prof. Pierpaolo Battigalli
Discussant: prof. Fabio Angelo Maccheroni
Master of Science Thesis of: Paolo Leonetti
Student ID: 1302108
Academic year: 2013-2014
Ringraziamenti
Iniziai questo lavoro un anno fa, con lobiettivo di produrre qualcosa di originale, in termini
di esposizione e di contenuto. E sono soddisfatto, nalmente ci sono riuscito. Credo che la
tesi segni la ne di anni di studio e molte amicizie, forse dei miei anni migliori; e mi sembra
giusto ringraziare qui le tante persone che ho incrociato lungo questo percorso perche, se
sono arrivato a questo punto, `e soprattutto merito loro.
Il grazie pi` u grande va a mamma e pap` a. Non perche si deve fare cos`. Ma per avermi
incoraggiato nei momenti peggiori, e creduto in me, anche quando non ci avrei creduto
neanche io. Se `e vero che buon sangue non mente, da grande voglio diventare come te, pap`a.
E grazie a Marco, riesci a stare simpatico a tutti: non potrei avere un fratello migliore.
Grazie al prof. Pierpaolo Battigalli, che mi ha seguito in questi ultimi anni, per aver avuto
tanta pazienza nel correggere i miei errori. Non mi riferisco a quelli della tesi, dove alla ne ho
fatto un po di testa mia, al solito. Ma per avermi fatto capire che la coerenza e la correttezza
dellimpostazione concettuale `e fondamentale. E per i tanti pomeriggi passati a spiegarmi
in dettaglio il perch`e delle cose da fare, per averci tenuto davvero ad insegnarmi qualcosa.
Grazie al prof. Fabio Maccheroni, per essere sempre stato sempre disponibile con me, ed
aver tolto tempo prezioso alle sue attivit` a per seguirmi, forse perch`e ha davvero creduto che
sarei stato capace di fare qualcosa. Per avermi trattato alla pari, ed avermi lasciato sempre
lultima parola su cosa scegliere.
Grazie ad Alessio, per essere sempre daccordo con me, per le serate al Gelatiere, e per
cercare di organizzare i miei programmi, anche quando non vanno come previsto. Grazie a
Veronica, per essere stata sempre simpatica e sorridente, e trovare le parole giuste. Grazie
a Ugo e a Saverio per non esserci persi di vista, anche a distanza. Grazie a Cristina, per
essermi stata a sentire parecchie volte.
Un ringraziamento particolare alla prof.ssa Carmen Gim`enez, per avermi aiutato a super-
are quellultimo esame di spagnolo, e a Salvatore Tring ali, che mi ha insegnato davvero tanto
negli ultimi anni.
Contents
Abstract vii
1 Preliminaries 1
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Literature review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2 Extensive Game Forms 3
2.1 Possibly innite trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.2 Active players and available actions . . . . . . . . . . . . . . . . . . . . . . . 20
2.3 Information sets partitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.4 :-reduced normal forms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3 A characterization of Behavioral Equivalence 37
3.1 The set of Invariant transformations . . . . . . . . . . . . . . . . . . . . . . . 37
3.2 Necessary and sucient conditions . . . . . . . . . . . . . . . . . . . . . . . 41
3.3 Extension to chance moves . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
4 Invariance of Solution Concepts 49
4.1 Renements of equilibrium concepts . . . . . . . . . . . . . . . . . . . . . . . 49
4.2 Rationalizability in extensive games . . . . . . . . . . . . . . . . . . . . . . . 49
v
5 Conclusions 50
Bibliography 51
Abstract
The main result of this thesis is to provide a complete characterization of behavioral
equivalence between possibly innite extensive game forms with imperfect information. Its
shown that two extensive game forms belong to the same subset of the quotient space
with respect to behavioral equivalence, i.e. they share the same reduced normal forms
with respect to terminal paths, if and only if one can be transformed into the other one
through a composition of two elementary transformations, commonly known as Interchanging
of Simultaneous Moves and Coalescing Moves / Sequential Agent Splitting. Then, the result
is extended to a more general theoretical framework, where extensive game forms allow
chance moves. Finally, we study the invariance of these trasformations, with respect to
known solution concepts, such as some variants of sequential equilibrium and extensive form
rationalizability.
CHAPTER 1
Preliminaries
1.1 Introduction
In the context of non cooperative game theory, where players make decision independently,
two dierent representations have been widely used to describe interactions in dynamic
games: the extensive form and the normal form, together with their variants. The main
dierence is that the former provides a richer structure, specifying the information available
to each active player and making explicit the order of moves, i.e. assigning a kind of value to
the temporal variable. It does not happen in the last representation, since it seems that each
player chooses instantaneously his strategy, i.e. the prole of actions he would take whenever
he could have the possibility to do it. The problem is that there exist situations where the
richer information structure of the extensive form is unuseful from some point of view, as in
the case of dynamic games where some players have to choose their moves, one by one, and
ignoring all previous actions chosen by his opponents. Although the order of playing may not
matter, e.g. from a solution mode perspective, the extensive form representation require its
specication.
1
Nevertheless, there is no representation which is actually better than the other
1
In the following Section, it will be developed a more general framework where at each vertex of the tree,
two or more players can choose simultaneously their actions.
1
2 Chapter 1. Preliminaries
one: if every dynamic game is going to be represented in normal form, then every situation
will look like a simultaneous moves game. Thats why, a coherent theory should explicit
highlights the connection between these two representations. The main question is, indeed,
to provide a characterization of behavioral equivalence between possibly innite extensive
game forms with imperfect information. In rough words, to characterize all the situations
where two extensive game forms are essentially the same, meaning that a rational player
would make eventually the same actions, according to the same lines of reasoning.
This paper is concerned not with rationality, or strategic considerations, but with the
descriptivec omponento f the theory of non-cooperativeg ames.I t is therefore in the same
spirit as the papers by Dalkey [3], Elmes and Reny [4], Krentel et al. [7], and Thompson [9].
1.2 Literature review
Much ink has been spilt in literature over the correct analysis of game equivalence. Our view is
that there is room in game theory for many notions of equivalence; ideally, they will eventually
be organized into a hierarchy in terms of the increasingly strict transformations under
which game structure remains invariant. From this perspective Thompsons transformations
provide a fruitful starting point for the development of more rened equivalence notions.
Contributions to this project can be found scattered throughout the literature. Kohlberg
and Mertens (1986) supplement Thompsons transformations with two which respectively
introduce superuous chance moves
CHAPTER 2
Extensive Game Forms
2.1 Possibly innite trees
As R. Stanley recently observed in [11] the number of systems of terminology presently used
in graph theory is equal, to a close approximation, to the number of graph theorists. Hence,
the set of notations that is going to be used throughout this thesis do not pretend to be
better than any alternative choice of terminology, as far as its only a matter of conventions.
To start with, the blackboard letters R, Q, Z, and N stand for
R Ordered eld of real numbers,
Q Field of rational numbers,
Z Subring of integers,
N Monoid of non-negative integers.
For any collection C of sets, dene the subcollections C

and C

by
C

= C C : [C[ 1 and C

= C C : [C[ 2.
Given two non-empty sets A, B,dene AB as x A: x / B and their symmetric dierence
3
4 Chapter 2. Extensive Game Forms
by A B = x A B: x / A B; moreover
1
2
A
Power set of A,
(A) Set of probability measures on the measurable space (A, B),
B
A
Set of functions f : A B,
Part(A) Set of partitions of A.
For any pair of partitions P
1
, P
2
Part(A), the expressions
P
1
is ner than P
2
or P
2
is coarser than P
1

mean that every member of the partition P


1
is a subset of some member of P
2
.
Given (possibly empty) sets A, B R, we will shorten A B with B
A
: for example
Q
(a,b]
= (a, b]
Q
= Q (a, b]
for all reals a, b such that a < b. Whenever the set A = (a, ) or [a, ) for some a R the
set B
A
will be written equivalently as B
>a
or B
a
respectively: thats why, for example, the
following representations are equivalent for the set of non-negative integers:
a R
[1,0)
, b R
(1,0]
, N = Z
>a
= Z
b
.
A relation r: A B is a map from each element a A to some (possibly empty) subset
of B, i.e. r(a) B; for all (possibly empty) subsets X A dene:
r(X) = b B: x X, b r(x),
I
r
(X) = Y 2
B
: x X, Y = r(x),
D
r
(X) = x X: b B, b r(x).
If X = A then r(X), I
r
(X) and D
r
(X) represent the range, image and domain of such relation,
1
A measurable space is a pair (A, B) where B is a -algebra dened on A; in the case A is a nite set in
this thesis its assumed B = 2
A
; in the case A is a topological space, B is assumed to be the Borel--algebra
of A.
2.1. Possibly innite trees 5
respectively. In case we would not allow r(a) to be empty for some a A, i.e. r(a)
_
2
B
_

for all a A, then D


r
(A) = A, and the relation r will be termed correspondence.
The correspondence map will be replaced with

to mean strong injectivity


2
and

to
mean weak surjectivity,
3
which in turn are simplied to and for the case of functions.
4
To conclude this preliminary section, given two real-valued functions f, g : R R, the
Bachmann-Landau notation f = O(g) represents the existence of a constant c R
0
such
that [f(x)/g(x)[ is upper bounded by c whenever x is suciently large (see for example [5]),
that is
limsup
x

f(x)
g(x)

c.
To be coherent through notations, capital letters A, B, . . . will be reserved for sets, ghotic
letters a, b, . . . for correspondence maps and A, B, . . . for collections of sets. Three important
exceptions will be made: G, T and V will be used for the class of all extensive games (possibly
with imperfect information), the set of all invariant trasformations and the set of rooted
trees, respectively. Calligraphic letters /, B, . . . will be used only to represents objects with
particular characteristics, e.g. a subcollections of extensive games.
Denition 1. A poset (partially ordered set) is an ordered pair (V, _) of a (possibly innite)
non-empty set V of vertices endowed with a binary relation _ contained in V V such that
the relation _ is:
1. reexive, i.e. V, _ ;
2. transitive, i.e. ,
t
,
tt
V, ( _
t

t
_
tt
) = _
tt
;
3. antisymmetric, i.e. ,
t
V, ( _
t

t
_ ) = =
t
.
2
Recall that a correspondence r: A B has the property of strong injectivity whenever r(a) r(a

) ,=
if and only if a = a

, meanwhile it has the property of weak injectivity whenever r(a) = r(a

) if and only if
a = a

, for all choices of a, a

A. Clearly, if r(a) is a singleton for all a A (i.e. r can be seen as a function)


then it has the strong injectivity property if and only if it has the weak injectivity property.
3
Recall that a correspondence r: A B has the property of strong surjectivity whenever I
r
(A) = (2
B
)

,
and it has the property of weak surjectivity whenever r(A) = B. Again, if r(a) is a singleton for all a A
then it has the strong surjectivity property if and only if it has the weak surjectivity property.
4
Notice that, formally, a function is not a correspondence r: A B such that [r(a)[ = 1 for all a A.
Indeed r(a) is a singleton in the collection of non-empty subsets
_
2
B
_

, that is a set which contains a unique


element. If f : A B is a function then f(a) B, hence it is not a set. Obviously, there exists a isomorphism
between the class of correspondences r: A B such that r(a) is a singleton for all a A and the class of
functions f B
A
.
6 Chapter 2. Extensive Game Forms
Notice that, according to [2], a binary relation _ which satises only 1 and 2 will be called
(partial) pre-order, while if it saties 1, 2 and 3 it will be called equivalence, where
3. symmetric, i.e. ,
t
V, _
t
=
t
_ .
Since it will be convenient in the description of the model, also the following notation is going
to be used: the rst one represents the strict predecessor relation, while the second one
stands for the immediate predecessor.

t
( _
t
) ( ,=
t
),

t
((
t
) (
tt
V, _
tt
_
t
)) =
tt
,
t
.
Notice that the a poset is only a particular instance of the notion of graph, intended
as a triple (V, E, ) where V and E are (possibly innite) sets of vertices and edges, and
the function : E
__
V
2
__
maps each edge to an unordered set of vertices ,
t
, where
and
t
are not necessarily distinct.
5
The most important dierence arises not on the fact
that the image (e) is unordered (indeed a triple (V, E,
t
) with the ordering requirement

t
: E V V is just called directed graph), but on the fact that the map is not necessarily
injective, i.e. we can have (e) = (e
t
) for some distinct e, e
t
E.
Denition 2. A tree is poset (V, _) which veries the following conditions:
4. If
t
and
tt
for some ,
t
,
tt
V then
t
=
tt
.
5. There exists a root
r
such that
r
_ for all V .
According to these denitions, a tree (V, _) can be equivalently seen as a directed graph
(V, E,
t
) such that the map
t
is injective, (E) = V , there do not exist cycles,
6
and there
do not exist proper subsets V
t
V and E
t
E such that (V
t
, E
t
,
t
[
E
) with the same
properties. To make things clear, Figure 2.1 shows four examples of (nite) graphs: bullets
stand for vertices and lines stand for edges; they can be modied (each one in exactly 2
[E[
ways) to be directed graphs, as shown for example in Figure 2.2.
5
Recall that if A is a set such that [A[ 2 and b a positive integer such that b [A[ then the multibinomial
__
A
b
__
represents the collection of sets of the forma
1
, . . . , a
k
for some a
1
, . . . , a
b
A not necessarily distinct.
In particular if A is nite then

__
A
b
__

= [A[
b
.
6
Given a directed graph (V, E,

), a cycle is a sequence (
1
, e
1
,
2
, e
2
, . . . ,
k
, e
k
) such that

(e
i
) =
(
i
,
i+1
) for all i [1, k 1]
Z
and

(e
k
) = (
k
,
1
), for some edges e
1
, . . . , e
k
E and some vertices

1
, . . . ,
k
V .
2.1. Possibly innite trees 7

Figure 2.1: Examples of graphs

Figure 2.2: Examples of directed graphs


All directed graphs in Figure 2.2 cannot be seen as trees, each one for a dierent reason.
The rst three ones cannot be even considered partially ordered sets: in the rst one the map

t
is not injective, i.e.
t
: E , V ; in the second one there exists a cycle, and in particular it
violates the antysimmetric property of the binary relation _; in the third one the transitive
property is not satised; the last one is a poset, but there exists non-empty proper subset
V
t
V such that (V
t
, _ [
V
) is still a poset (and moreover there does not exist a root
r
).
An example of tree can be seen in Figure 2.3.
8 Chapter 2. Extensive Game Forms

Figure 2.3: An example of tree.


Such a tree can be drawn equivalent as in Figure 2.4: the vertex
r
on the top precedes all
other ones, i.e.
r
_ for all V . Lets dene V the set of rooted trees.
Lemma 1. The root
r
of (V, _) V is unique.
Proof. On the one hand, the existence of such a vertex is guaranteed by the condition 5
of Denition 2. On the other hand, the root has to be unique: suppose for the sake of
contradiction that there exists another
t
r
such that
t
r
_ for all V . Then
r
_
t
r
and
t
r
_
r
. By the antisymmetric property of the relation _ given in the condition 3 of
Denition 1 we must have
r
=
t
r
.

r
Figure 2.4: The same tree of Figure 2.3.
2.1. Possibly innite trees 9
Lemma 2. Given (V, _) V with root
r
such that the set V :
r
is well-dened,
there exists a unique P Part (V
r
) such that if P P then (P, _ [
P
) V.
Proof. First, lets prove the existence of at least one such partition P of the set V
r
:
dene the set of vertices V
t
= V :
r
, and also V

= V :
t
_ for each

t
V
t
. Clearly P = V

V
Part (V
r
) and (V

, _ [
V

) V for all
t
V
t
.
Suppose that there exists another partition P
t
dierent from the previous one P, which
satises the assumption as well. Then there exist P P
t
and distinct
t
,
tt
V
t
such that
(P V

,= ) (P V

,= ) .
Hence some distinct vertices
1
,
2
P such that
1
P V

and
2
P V

are well-
dened. Since (P, _ [
P
) belongs to V by assumption, there exists a root
p
P such that

p
_ for all P, which is unique according to Lemma 1. In particular
p
_
1
and

p
_
2
, and at the same time
(
r

t
_
1
) (
r

tt
_
2
) .
As far as
t
,=
tt
, by construction every pair of trees with roots
t
and
tt
have no elements
in common, that is V

= . Hence it must hold also that


p

t
and
p

tt
. The
unique possible case is
p
=
r
, which is a contradiction since it was assumed that

p
P P
t
Part (V
r
) .
Let s be the correspondence such that s() is the (non-empty) subset of vertices which
weakly precedes ,
7
that is
s: V V

t
V :
t
_ .
Morever, let s be the function such that s() is the vertex such that s() , whenever it
7
Indeed
r
s() for all V , hence s is a correspondence.
10 Chapter 2. Extensive Game Forms
can be dened. Call D
s
(V ) V the domain of such function,
8
then
s: D
s
(V ) V
s() such that s() .
Denition 3. Given a tree (V, _), dene V the collection of all sets V
t
(2
V
)

which can
be completely ordered by the immediate relation , i.e. such that there exists a path
9

t
. . . . . .
tt
with
t
, . . . , . . . ,
tt
= V
t
.
In particular it implies that all vertices dierent from the root
r
which belong to some
V
t
V have to have a well-dened predecessor, so that the following chain of inclusions
holds:

V

V
V
t
D
s
(V )
r
V.
Lemma 3. If (V, _) V and V
t
V then V
t
is nite or at most countable.
Proof. If V
t
is a nite set then the claim is trivial. Otherwise [V
t
[ = , and according to
Denition 3 the condition V
t
V implies that there exists a path
t
. . . . . .
tt
such that
t
, . . . , . . . ,
tt
= V
t
. Hence
10
[N[ [V
t
[ 2[N[ = [N[ = [V
t
[ = [N[.
Lemma 4. Fix a tree (V, _) V and a vertex V
r
. Then
s() V s() (D
s
(V )
r
) (D
s
1(V ) ) .
8
In particular the domain of such function veries D
s
(V ) V
r
. Its possible also that D
s
(V ) is a
proper subset (possibly the empty-set) of V
r
, as its going to be shown in Remark 2, or simply choosing
(V, _) = ([0, 1]
R
, ).
9
Notice that the lenght of a path, that is [V

[, is allowed to be also (possibly uncountable) innite.
10
Recall the smallest order of innity is the one of the set of (positive) integers, so that [X[ [N[ for all
sets X such that [X[ = .
2.1. Possibly innite trees 11
Proof. The if part trivially follows by denitions. For the only if part, suppose that s()
belongs to V , i.e. it can be completely ordered by the immediate relation , in a path of
the form
11

r
. . . . . . . (2.1)
Then s(
t
) is well-dened for all
t
D
s
(V )
r
, so that (s
n
())
nN
represents the
(countable) sequence of vertices in such path starting from the right. Also s
1
(
t
) s()
is non-empty
12
(or better, it is a singleton) for all
t
D
s
1(V ) . It means that iterating
such function we can construct the countable sequence starting from the left at the root

r
. It implies that we can order with countably many points of s() in the path (2.1) at
the left and at the right. Adding the fact that s() can be only nite or at most countable
by Lemma 3, the proof is complete.
Moreover, such path (2.1) has to be unique. Suppose for the sake of contradiction that
there exists a vertex V
r
with (at least) two distinct paths from
r
to itself, i.e.
there exist ,
t
,
tt
V such that
t
,=
tt
,

r
. . .
t
_ and
r
. . .
tt
_ .
In particular it implies that
t
V :
t
_
t

tt
V :
tt
_
tt
, which is impossible,
as far as at the same time by Lemma 2

t
V :
t
_
t

tt
V :
tt
_
tt
= .
As immediate corollary, the following inclusion holds:
s (D
s
(V )) (D
s
(V )
r
) (D
s
1(V ) ) .
Special labelings are going to be assigned to some subsets of vertices with useful properties,
the rst one related to the closeness to (nite) terminal nodes, the second one with respect
to the tree representation in Hasse diagram (see e.g. [1]).
11
Indeed
r
, which is unique according to Lemma 1, belongs to s().
12
Recall that if f : A B is a function then the inverse map f
1
: B A is a correspondence if and only
if f is surjective, i.e. f : A B.
12 Chapter 2. Extensive Game Forms
For all (V, _) V, dene the (possibly nite) sequence of sets
_
V
(n)
_
nN
by
V
(0)
= V,
V
(n+1)
= V
(n)
:
t
V
(n)
,
t
, for all n N
whenever possible, and in particular when [V
(n)
[ 2.
Moreover, dene the sequence of sets
_
Z
(n)
_
nN
by
Z
(0)
= V : (
t
V _
t
) = ( =
t
),
Z
(n+1)
= V :
t
Z
(n)
,
t
, for all n N
whenever possible, i.e. whenever Z
(n)
D
s
(V ). Notice that from the denition there is no
reason to state that Z
(0)
is well-dened. In rough words, the sequence
_
V
(n)
_
nN
aims to
represent the whole set of vertices V after exactly n consecutive deletion of terminal nodes
(whenever such operation is possible); for example V
(1)
stands for non-terminal nodes, which
is a well-dened sets for a tree (V, _) V if and only if Z
(0)
D
s
(V ). Meanwhile, the
sequence
_
Z
(n)
_
nN
represents the set of nodes V such that there are exactly other n
distinct vertices which follow itself.
Remark 1. The identity

nN
Z
(n)
= V is false in general, even in the case that the whole
sequence (Z
(n)
)
nN
is well-dened. As example, choose the tree (V, _) such that
V = [0, 1]
R

_
2 +
1
n + 1
_
nN
,
where _
t
for some ,
t
V if and only if
t
. On the one hand, the set Z
(n)
is
well-dened for all n N, indeed Z
(n)
= 2 +
1
n+1
; on the other hand

nN
Z
(n)
[2, 3]
Q
,
which has to be clearly dierent from V itself.
Looking (V, _) V such that Z
(0)
is well-dened as an Hasse diagram, call V
Z-n
the greatest
subset of V made by all vertices in some nite terminal paths, i.e.

Z
(0)
: [s()[N
s(),
2.1. Possibly innite trees 13
and V
Z-inf
the greatest subset of V made by innite terminal paths such that the set of
weakly predecessors s() can be ordered by the immediate relation . Hence V
Z-n
, V
Z-inf

represents a partition of the vertices which belong to paths of terminal nodes which can be
completely ordered by . Formally
V
Z-n
= V :
t
Z
(0)
, s(
t
), [s(
t
)[ N,
V
Z-inf
= V V
Z-n
:
t
Z
(0)
, s(
t
) V .
Then, let us partition V (V
Z-n
V
Z-inf
) in V
Ninf
and V
Rinf
such that the former represents
the set of vertices V belonging to some non-terminal path which can be completely
ordered by the immediate relation . More precisely
V
Ninf
= V (V
Z-n
V
Z-inf
):
(
n
)
nN
V
N
,
n

nN
, s() V ,
n

nN
Z
(0)
= ,
and
V
Rinf
= V (V
Z-n
V
Z-inf
V
Ninf
).
Notice that these sets V
Z-n
, V
Z-inf
, V
Ninf
and V
Rinf
are well-dened if and only if the
set of terminal nodes Z
(0)
is well-dened as well. Intuitively, the set V
Z-n
represents the
set of all vertices which belongs to nite path of terminal nodes Z
(0)
, i.e. s() =
, s(), s
2
(), . . . , s
n
() for some non-negative integer n, where as usual s
n
= s . . . s
. .
n times
represents the n-th iterate of the function s. The reason underlying the denition of these
sets is to classify the vertices V primarly to the existence of paths which can be ordered by
(i.e. sets V
t
V ), then to the existence of a terminal node in the path and nally to the
lenght of the path itself.
Lemma 5. Given (V, _) V such that Z
(0)
is well-dened, then:
V
Z-n
= s() V .
Proof. By assumption there exists an non-negative integer n such that s() =
14 Chapter 2. Extensive Game Forms
, s(), s
2
(), . . . , s
n
(). Recalling that
r
s(), the nite path is of the form

r
= s
n
() s
n1
() . . . s() .
In particular it means that s()
r
= s
n
()
[0,n1]
Z
D
s
(V ). Since the set s() has to
be nite, then also the following inclusion holds
s() (D
s
(V )
r
) (D
s
1(V ) ) .
The claim follows by the equivalence representation given in Lemma 4.
This Lemma allows us to describe this set equivalently as
V
Z-n
= V :
t
Z
(0)
, s(
t
) V , [s(
t
)[ N,
claring the dierence with V
Z-inf
: assumed that the set Z
(0)
can be identied, the latter
allows the set s() to be not nite, for some vertex Z
(0)
V D
s
1(V ).
Remark 2. If (V, _) V is tree such that Z
(0)
is dened, then the costructed sets are
pairwise disjoint and
V
Z-n
, V
Z-inf
, V
Ninf
, V
Rinf
Part(V )
Moreover, if V
Rinf
,= then it cannot be a nite set. Nevertheless, there is no reason to
state that V
n
is a nite set. A simple counterexample is the rooted tree (V, _) such that V
is a (possibly uncountable) innite set and the _ relation is dened only by
r
_ for all
V .
Remark 3. Although Lemma 3 proves that s() is an innite countable set for all
V
Z-inf
Z
(0)
, the same does not hold for the whole set V
Z-n
. Let us make an example: dene
the tree (V, _) with (unique) root
r
, and call T
t

tR
0
Part(V ) such that T
0
=
r
and
T
t
D
s
(V ) a countable innite set for all positive reals t. The relation _, except for the
root
r
which preceeds every V , will be choosen such that _
t
only if there exists a
t R
>0
such that ,
t
T
t
. According to this construction, V
Z-n
= V
Ninf
= V
Rinf
=
and V
Z-inf
= V . Moreover
[V
Z-n
[ [N R[ [R[,
2.1. Possibly innite trees 15
implying that V
Z-n
can be also an uncountable set. Such costruction is really possibile, e.g.
choosing
t (0, 1)
R
, T
t
= t + z
zZ
and _ coincides with on each set T
t
.
Also, for all trees (V, _) V dene (whenever possible) the following sequence of sets (V
n
)
nN
by
V

0
=
r
,
V

n+1
= V :
t
V
n
,
t
, for all n N.
Again, from the denition of tree (V, _), it can be the case that V

1
is not well-dened.
Remark 4. There is no reason to state that

zZ
V
z
= V : a simple counterexample is the
rooted tree (V, _) such that V = [0, 1]
R
and _
t
if and only if
t
. Notice that in this
example V
n
is well-dened only if n = 0. Even if it was possible to well-dene such subsets,
they would be singletons for all n N, implying that V would be made by at most countably
many vertices, which is not the case: a explicit example of this type can be made on the the
same line of Remark 1.
The main results of this thesis will be shown to hold conditioning on the fact that (V, _) V
belongs to a particular subcollection of trees, which is going to be called V
std
.
Denition 4. The set V
std
is composed by all trees (V, _) V such that
1. Z
(0)
is well-dened;
2. V
Rinf
= ;
3. V Z
(0)
is at most countable.
In other words, a standard tree (V, _) is tree such that the terminal nodes are well-
dened (such set is allowed to be also uncountable), and such that the subset of non-terminal
nodes V Z
(0)
is nite or at most countable (hence, the same holds for V
(1)
, as far as
V
(1)
V Z
(0)
); in particular for all vertices V the path from the root
r
to itself can
16 Chapter 2. Extensive Game Forms
be explicitely written only with the immediate predecessor relation . Hence, the domain
of the predecessor function s veries
(V, _) V
std
= D
s
(V ) = V
r
.
Observe that if (V, _) is a tree and V = V
Z-n
then each set above is easily well-dened. Let
us see what happen in other cases.
Lemma 6. If (V, _) V
std
and V ,= V
Z-n
then V
n
is well-dened and non-empty for all
n N and
V
n

nN
Part(V ).
Proof. As far as (V, _) V
std
, the set V
Rinf
has to be empty, so that
V = V
Z-n
V
Z-inf
V
Ninf
.
Then, according to Lemma 5, if V is a vertex then s() V , i.e. the sets of (weakly)
predecessors of can be ordered by the immediate predecessor relation . If s() is nite
then there exists n N such that
s() = , s(), . . . , s
n
().
Otherwise s() is countable innite set, according to Lemma 3, and there exists a (unique)
ordering by in the form

r
. . . s
n
() s
n1
() . . . s() .
It means that its uniquely dened a vertex
1
s() such that
r
= s(
1
). Suppose that

n
is uniquely identied for some positive integer n; then
n+1
s() is the unique vertex
such that
n
= s(
n+1
). By the principle of mathematical induction (which can be shown to
be exactly equivalent to the axiom of well-ordering) the vertex
n
is uniquely dened for all
positive integers n. Its clear now how to construct the sets V
n
for all n N.
In particular it implies that the collection of sets V
n

nN
has to be nite
13
or at
13
Clearly, the collection V
n

nN
is nite if and only if V = V
Z-n
.
2.1. Possibly innite trees 17
most countable innite. Moreover, its going to use explicitely the following well-known
assumption:
AC (Axiom of Choice). For each family of non-empty sets (X
i
)
iI
, there exists (x
i
)
iI
such that x
i
X
i
for all i I.
Such axiom was formulated rst in 1904 by Ernst Zermelo in order to formalize his proof
of the well-ordering theorem : see for example [10]. The following Lemma will be very useful
for the proof of the main result:
Lemma 7. Let X be a non-empty set and (P
i
)
iI
a family of partitions, i.e. P
i
Part(X)
for all i I. Under AC then
14

iI
P
i
Part(X).
Proof. For each x X dene the family of sets (P
i
(x))
iI
such that x P
i
(x) P
i
for all
i I. And dene P(x) =

iI
P
i
(x), for each x X. Consider the collection P(x)
xX
and suppose that P(x) P(x
t
) ,= for some x, x
t
X, i.e. there exists x
tt
X such that
x
tt
P(x) P(x
t
). If there exists j I such that P
j
(x) ,= P
j
(x
t
) then
x
tt
P(x) P(x
t
) =

iI
(P
i
(x) P
i
(x
t
)) P
j
(x) P
j
(x
t
) = ,
which is impossible. It implies that P
i
(x) = P
i
(x
t
) for all i I, i.e.
P(x) P(x
t
) ,= P(x) = P(x
t
).
Hence, its well-dened a equivalence relation on the collection P(x)
xX
such that
P(x) P(x
t
) if and only if P(x) = P(x
t
). By AC its possible to choose a representative
element of each equivalence class, let us say P(x)
xX
for some non-empty X
t
X.
By construction the collection P(x)
xX
is made by pairwise disjoint sets such that

xX

P(x) = X, hence

iI
P
i
= P(x)
xX
Part(X).
14
For all P
1
, P
2
Part(X), the intersection P
1
P
2
is dened as the coarset partition P Part(X)
such that P is ner than P
1
and P
2
. Explicitely, it represents the inferior limit with respect to the partial
order of the collection of partitions, i.e. P = P
1
P
2
: P
1
P
1
, P
2
P
2
.
18 Chapter 2. Extensive Game Forms
Since it will be used often later, its convenient also to dene the following sets, whenever
(V, _) V
std
and n N:
1
(n+1)
= V Z
(n)
.
Remark 5. According to the previous denitions, the inclusion V
(n)
1
(n)
holds for all
positive integers n. Indeed:
V
(n)
= s (V
r
) = (V
Z-n
V
Z-inf
) Z
(n1)
(V
Z-n
V
Z-inf
V
Ninf
) Z
(n1)
= V Z
(n1)
= 1
(n)
.
Finally, the set of sequences of (possible innite) terminal paths : is partitioned in :
n
and :
inf
, where
:
n
=
_
(
n
)
n[0,m]
Z


nN
V
n
:
t
Z
(0)
V
Z-n
, s(
t
) =
n

n[0,m]
Z
_
and
:
inf
=
_
(
n
)
nN
V
N
:
t
Z
(0)
V
Z-inf
, s(
t
) =
n

nN
_

_
(
n
)
nN
V
N
: m N, s(
n

n[0,m]
Z
) =
n

n[0,m]
Z

n

nN
Z
(0)
=
_
.
According to these denitions, :
n
represents the set of terminal and nite paths with can
be completely ordered by . Meanwhile, :
inf
stands for the set of innite (countable) paths
which can be also ordered by (such that they can end in some terminal Z
(0)
or
not). Clearly, if V = V
Z-n
then : = :
n
and it can be isomorphically represented by Z
(0)
.
Remark 6. Apart from the denition of the set V
Z-n
, which is mathematically clear, it
should be not obvious to state that such set can actually exist in some tree (V, ) V
std
.
According to Lemma 3, if Z
(0)
V
Z-n
then s() has to be an innite countable set, i.e.
[s()[ = [N[. On the one hand, the rst example that one can think is ([0, 1]
Q
, ), but the
problem is that it does not belong to V
std
, as far as s(1) / V .
15
On the other hand, such
example really exists: its enough to choose (V, _), where
V =
_
1
2

1
n + 2
_
nN
and _
t

t
.
15
The reason underlying this fact is that [0, 1]
Q
is dense in [0, 1]
R
.
2.1. Possibly innite trees 19
Its straightforward to check that (V, _) V
std
and V = V
Z-inf
.
For convenience, it will be useful to dene the function which associate to each tree
(V, _) V
std
the maximal lenght of paths which belong to :
n
, which has to be a nite
non-negative integer, that is
v : V
std
N,
(V, _) 1 + max
(n)
n[0,m]
Z
:
n

n[0,m]
Z

.
Hence it represents the longest nite lenght between all terminal paths, as its going to be
seen in Section 2.3.
Remark 7. The reason underlying the denition of the collection of trees V
std
is that a
construction level-by-level is needed, in the sense that (a possibly innite sequence of) games
with simultaneous moves will be played, one for each level V

0
, V

1
, . . .. This sequence of
games stops if a terminal node is reached, i.e. such that if a game is played at some vertex
V
(1)
. It would be rather reasonable to assume that if the sequence of games is not nite
and the game with simultaneous moves at level V
n
is going to be played at time t
n
, for some
(strictly increasing) sequence of reals (t
n
)
nN
, then
liminf
nN
t
n+1
t
n
> 0.
In the real world, indeed, a minimum amount of time is always available between two
consecutive games. Thats why the addition assumption on the description of the set V
std
could be given in Denition 4:
V
Z-inf
= .
As far as it does not allow to improve the results of the following sections, and neither
to obtain any technical simplication in their proofs, its chosen to keep the more general
theoretical framework which is already dened.
20 Chapter 2. Extensive Game Forms
2.2 Active players and available actions
Given a tree (V, _) V
std
, a game with simultaneous moves
16
is going to be played at each
non-terminal vertex 1
(1)
.
17
Let I be the non-empty greatest set of players who are
potentially active at least once in the tree (V, _). One the one hand, it means that it is
well-dened a weakly surjective correspondence
p: 1
(1)

I
such that p() represents the set of players who are active at the node 1
(1)
. On the other
hand, also its inverse correspondence is weakly surjective,
18
let us say
v: I

1
(1)
,
i 1
(1)
: i p().
Clearly, for each player i I the set v(i) stands for the (non-empty) greatest subset of
vertices 1
(1)
such that i is potentially active at the game with simultaneous moves which
is played at vertex .
Lemma 8. If (V, _) V
std
and sup
1
(1) [p()[ is nite then I is nite or at most countable.
Proof. Since the correspondence p is weakly surjective by construction then
I = p
_
1
(1)
_
=

1
(1)
p().
Also by assumption there exists a positive integer n such that [p()[ n for all 1
(1)
. It
follows that
[I[ =

1
(1)
p()



1
(1)
p() n[1
(1)
[.
As far as (V, _) V
std
then 1
(1)
is nite or at most countable, according to Denition 4.
The claim follows.
16
According to literature, sometimes games with simultaneous moves are called with the misleading term
strategic games.
17
Notice that if (V, _) belongs to V
std
, then 1
(1)
= V
(1)
if and only if V
Ninf
= , according to Remark 5.
18
Recall that if r: A B is a correspondence, then its inverse r
1
: B A is not in general a relation.
Moreover, r
1
is a correspondence if and only if r is weakly surjective, that is r(A) = B; in such case, r
1
is
weakly surjective too.
2.2. Active players and available actions 21
Dene / the subset of I 1
(1)
such that (i, ) / if and only if i p(), or equivalently
/ = (i, ) I 1
(1)
: v(i).
In other words, / represents the collection of all pairs (i, ) such that the player i I is
active
19
in the game with simultaneous moves played at node 1
(1)
. In particular the set
/ has to be non-empty, as its shown in the following Lemma.
Lemma 9. If (V, _) V
std
then [/[ max[I[, [1
(1)
[.
Proof. By construction the relation p: 1
(1)

I is a correspondence, implying that [/[


[1
(1)
[. Adding the fact that its also weakly surjective (or equivalently that also its inverse v
is a correspondence), then also [/[ [I[.
For all pairs (i, ) /, dene A
i,
a non-empty set of actions: it represents the set of
available actions at player i I when the game with simultaneous moves is played at the
non-terminal node 1
(1)
. For all non-empty subsets B / dene the set
A
B
=

(i,)B
A
i,
. (2.2)
Notice that A
B
is well-dened even if B is not a cartesian product. To ease the notation, the
sections-type sets are
i I, A
i
= A
iv(i)
=

(i,),(i1
(1)
)
A
i,
,
and
1
(1)
, A

= A
p()
=

(i,),(I)
A
i,
.
In few words, the denition of the sets A
i,
is the natural way to assign labellings to edges
going from 1
(1)
to
t
V such that
t
. It follows that
1
(1)
, A


= s
1
().
In particular, the game with simultaneous moves at vertex 1
(1)
, which is going to be
termed

from here later, allow each player i p() to choose in the set of available actions
19
Formally, a player i I is said to be active at node 1
(1)
if and only if (i, ) /.
22 Chapter 2. Extensive Game Forms
A
i,
. Hence, a sequence
(a
i,
)
ip()
A

(2.3)
uniquely identies a vertex
t
V such that
t
.
Remark 8. From here later it can be assumed without loss of generality that
i I, ,
t
v(i), A
i,
A
i,
,= A
i,
= A
i,
. (2.4)
Indeed, in the case that A
i,
A
i,
is non-empty, but the two sets are dierent, the player
i can realize that the set of actions available at node is dierent than the set of actions
available at
t
. Hence, he can assign new labels to one of these sets so that condition (2.4)
holds.
2.2.1 Truncated trees
Denition 5. Let (V, _) V
std
be a tree and x
t
1
(1)
and non-empty set V
t
s
1
(
t
).
20
Then
(V [

,V
, _ [

,V
) V
std
is the truncated tree such that
V [

,V
=
t

tt
V : _
tt

and

tt
_ [

,V

ttt

tt
_
ttt
.
Remark 9. Notice that if (V, _) V
std
then the identity
(V, _) = (V [

,V
, _ [

,V
)
holds if and only if

t
=
r
V
t
= V

1
.
According to Denition 5, this statement holds because V

1
= s
1
(
r
).
20
Recall that if f : A B is a function then its inverse f
1
: B A: b a A: f(a) = b is in general
a relation; moreover, it is a correspondence if and only if the function f : A B is surjective, i.e. f(A) = B.
2.2. Active players and available actions 23
Denitions for sets given in Section 1.1 for the whole tree (V, _) are extended for the
truncated trees (V [

,V
, _ [

,V
): if n is a non-negative integer then
V
(n)
[

,V
= V
(n)
V [

,V
,
Z
(n)
[

,V
= Z
(n)
V [

,V
,
1
(n)
[

,V
= 1
(n)
V [

,V
.
The classication of truncated trees is in this Section because the denition of the set of
players I can be extended in a similar way. For all
t
1
(1)
and non-empty V
t
s
1
(), let
us write
I[

,V
=

V [

,V

p(),
and
I|

,V
=

V [

,V
\

p()
Clearly, I[

,V
represents the set of players who are active at least once in the truncated tree
(V [

,V
, _ [

,V
), while I|

,V
does not require that such players are active in the root
t
of
the truncated tree.
Remark 10. For all trees (V, _) V
std
and
t
, V
t
such that
t
1
(1)
and 1
(1)
s
1
(
t
),
the set of players I[

,V
has to be non-empty. Indeed

t
V [

,V
= p(
t
) I[

,V
.
Conversely, there is no reason to state that I|

,V
,= , and it should be clear that
I|

,V
=
t
V
(1)
.
Moreover, the trivial inclusion I|

,V
I[

,V
does not have to be necessarily strict, indeed,
the equality can be reached if and only if p(
t
) I|

,V
, that is, if each player active in the
root
t
is active at least another time in the truncated tree (V [

,V
, _ [

,V
).
Lemma 10. Given a tree (V, _) V
std
such that 1
(1)
is nite, then
V
z

z[0,v(V,)]
Z
Part(V )
24 Chapter 2. Extensive Game Forms
Proof. If V
Z-inf
V
Ninf
,= then there exists at least a innite countable path (which can
be terminal in some Z
(0)
or not), i.e. :
inf
,= , implying that
(V
Z-inf
Z
(0)
) V
Ninf
V Z
(0)
= 1
(1)
,
which is impossible. Hence
[1
(1)
[ < V = V
Z-n
[V
(1)
[ < ,
and in such case 1
(1)
= V
(1)
. In particular the collection of set of vertices V
z

z[0,v(V,)]
Z
is
a partition of V .
Lemma 11. Given a tree (V, _) V
std
such that V is nite, then
[V [ = 1 +

z[1,v(V,)]
Z

Z
(z)

(i,),(I)
[A
i,
[.
Proof. Since 1
(1)
V then 1
(1)
has to be nite as well, and Lemma 10 applies. Hence
V

0
, . . . , V

v(V,)
Part(V ). Observe now that
z [0, v(V, _)]
Z
, Z
(z)
= V

v(V,)z
.
In particular
V =
_
z[0,v(V,)]
Z
V
z
=
_
z[0,v(V,)]
Z
Z
(z)
=
r

_
z[0,v(V,)1]
Z
Z
(z)
.
Counting the number of elements on each member of this identity and considering that
1
(1)
, [A

[ = [s
1
()[,
we can conclude
[V [ = 1 +

z[0,v(V,))
Z
[Z
(z)
[ = 1 +

z[1,v(V,)]
Z

Z
(z)
[A

[.
The claim follows from the Equation (2.2).
2.2. Active players and available actions 25
Lemma 12. Given a tree (V, _) V
std
and two partitions P, P
t
Part(/), then
21

1P
A
1

=

A
1
.
Proof. Such isomorphism, formally between sequences of cartesian sets, easily holds as far as
P Part(/), A
,

=

1P
A
1
.
As immediate corollary, if (V, _) V is a tree such that V is nite, then

iI
[A
i
[ =

V
(1)
[A

[.
2.2.2 Best-control sets
For each player i I and (pre-non-terminal), a vertex
t
1
(2)
and a non-empty set
V
t
s
1
(
t
) (that is, a set of vertices such that each one of them immediately follows
t
), a
set of non-terminal vertices
C v(i) (V [

,V

t
)
is said to be a control-set if every terminal path allowed by the truncated tree (V [

,V
, _
[

,V
) has at least one vertex of C along its path. Formally it means that
(
n
) : (V [

,V
, _ [

,V
) =
n
C ,= . (2.5)
Here, : (V [

,V
, _ [

,V
) stands for the set of terminal paths allowed by the truncated tree
(V [

,V
, _ [

,V
) and the sequence (
n
) is intentionally not indexed to mean that it can be
nite or countable innite.
Observe that it is not assumed that player i is necessarily active at vertex , i.e. it can be
the case i / p(). Moreover such denition is meaningful only if the set 1
(2)
is well-dened,
which happens if and only if v(V, _) 2.
21
Informally, an isomorphism

= is a map that preserves sets and relations among elements.
26 Chapter 2. Extensive Game Forms
Denition 6. For each tree (V, _) V
std
, player i I, vertex
t
1
(2)
and non-empty set
V
t
s
1
(
t
), let C

,V

,i
be the collection of all control-sets. If this collection is non-empty,
the best-control set c

,V

,i
is dened by
c

,V

,i
= arg min
CC

,V

,i
[C[.
Intuitively, such set, whenever its dened, represents the group of simultaneous moves games
which are nearest to the vertex 1
(2)
and where player i I is always active such that each
terminal path allowed by the truncated tree has to pass also through one of such vertices.
The following Lemma shows that the Denition 6 is well-posed.
Lemma 13. For each tree (V, _) V
std
, the following holds
[C

,V

,i
[ 1 =

arg min
CC

,V

,i
[C[

= 1.
Proof. It is enough to construct the set c

,V

,i
in an alternative way, and check that this
construction denes this set in a unique way. According to (2.5), for each allowed terminal
path (
n
) : (V [

,V
, _ [

,V
), let ((
n
)) dened by
((
n
)) = (
n
) : (V [

,V
, _ [

,V
) :
t
i p(),
that is, the nearest vertex along the path (
n
) which (stricly) follows
t
and such that player
i is active. Then its clear that
c

,V

,i
=
_
(n):(V [

,V
,[

,V
)
((
n
)).
The condition C

,V

,i
,= in Lemma 13 is clearly necessary, indeed it can be the case that
such collection of sets is empty, e.g. whenever i / I|

,V
, or equivalently
v(i) V [

,V

t
.
With the same line of reasoning of the above proof, it should be clear that if the greatest set
2.2. Active players and available actions 27
V
t
s
1
(
t
) such that the best-control set c

,V

,i
exists (or equivalently such that C

,V

,i
,= )
then is uniquely dened as well; moreover, it is non-empty if and only if there exists a vertex

tt
1
(1)
such that
t

tt
and the collection C

,i
is non-empty. From here later, such
greatest subset V
t
s
1
(
t
) will be called V
max

,i
,
22
that is
V
max

,i
= arg max
V

s
1
(

): C

,V

,i
,=
[V
t
[.
Hence, for all players i I and vertices
t
1
(2)
, if there exists a vertex
tt
s
1
(
t
)
such that the collection C

,i
is non-empty, then its well-dened uniquely the greatest
V
max

,i

_
2
s
1
(

)
_

such that the best-control set c

,V
max

,i
,i
exists.
22
The notation V
max

,i
highlights the fact that this (non-empty) set is maximal and it denition depends on
the choice of the vertex

1
(2)
and player i I|

,s
1
(

)
.
28 Chapter 2. Extensive Game Forms
2.3 Information sets partitions
Let us study now the problem of information sets, i.e. to formalize the situation where at
some point of the game the actions of some players could not be observed by a subset of other
players, not necessarily distinct from the rst ones, according to the rules of the game.
23
Denition 7. Given a tree (V, _) V
std
, dene 1 the set of histories, which is partitioned
into
1
n
=
_
(
n
)
n[0,m]
Z


nN
V
n
: s
_

n[0,m]
Z
_
=
n

n[0,m]
Z
_
and
1
inf
=
_
(
n
)
nN
V
N
: m N, s(
n

n[0,m]
Z
) =
n

n[0,m]
Z
_
.
In few words, 1
n
represents the set of nite histories h

, where h

is the sequence of
vertices which goes from the root
r
to the vertex itself (indeed, such sequence is unique
according to Lemma 4); and 1
inf
stands for the set of countable innite sequences (
n
)
nN
such that
n

n+1
for all non-negative integers n and
0
=
r
. Notice that, according to
denitions of terminal paths, the inclusion :
n
1
n
holds; moreover, under the assumption
(V, _) V
std
, also 1
inf
= :
inf
. Hence it makes sense to dene the set of non-terminal histories
1 as
1 = 1 : = 1
n
:
n
.
Remark 11. If (V, _) V
std
then V coincides exactly with (2
V
)

, and in particular
[1[ [1
n
[ [1[ [V [.
Moreover, there exists a trivial isomorphism 1
n

= V such that if V there exists a unique
associated history h

1, and conversely for each h 1 there exists a unique associated


vertex
h
V . Indeed, the following characterization can be stated
h

1
h
1
(1)
.
Finally, observe that the implication
V
Z-n
V
Ninf
= h

1
fin
23
Notice that it has nothing to do with features of players, e.g. bounded rationality, forgetfulness, etc.
2.3. Information sets partitions 29
holds in general, but the same cannot be said for vertices V
Z-inf
: according to the example
shown in Remark 6, it can be the case that
h

V
Z-inf 1
n
,= h

V
Z-inf 1
inf
,= .
It should be clear now that for all trees (V, _) V the following identity holds
v(V, ) = 1 + max
V
Zn [h

[.
and it can be the case that v(V, ) ,= 1 + max
V
[h

[.
The set of non-terminal histories of each player i I can be dened at the set of non-terminal
histories where at the last node he is active, that is
i I, 1
i
= h 1: i p(
h
).
Informally, for each player i I, a subset of vertices V
t
v(i) belongs to one of his
information set if such player is not able to realize at which particular node he is playing: at
rst sight whenever

V

A
i,
,= .
According to Remark 8, this notation can be notably simplied.
Denition 8. For each player i I in a tree (V, _) V
std
let H
i
Part(1
i
) be the
collection of information sets where each H
i
H
i
represents a class of the quotient space
1
i
/ , where
h

A
i,
= A
i,
.
Since H
i
is a partition of the set of non-terminal histories 1
i
then easily
i I, [H
i
[ [1
i
[ [1[ = [1
n
:
n
[ [V Z
(0)
[.
Considering that 1
(1)
is at most countable (indeed the tree (V, _) belongs to V
std
), the
collection of information sets has to be nite or at most countable as well. In the second
30 Chapter 2. Extensive Game Forms
case, it makes sense to index each H
i
H
i
with a subscript j J, that is
H
i
= H
i,j

jJ
,
where J is a non-empty set which is nite or at most countable. A straightforward assumption
about the structure of the collection H
i
is going to be made, which is common in literature
(see for example [3], [12] or [4]):
Perfect recall. Each player knows and remembers everything he did in prior moves.
In rough words, it states that players never forget information once it is acquired.
24
Formally
a extensive game played on a tree (V, _) V
std
satises perfect recall if
i I,
t
,
tt
v(i), ((
t
_
tt
) (h

H
i,j
H
i
)) = (
t
=
tt
). (2.6)
A special class of games are the one with perfect information, i.e. games such that
i I, H
i
= h

v(i)
,
i.e. the nest partition of 1
i
(see for example [9]). Obviously, every game with perfect
information satises also perfect recall.
According to the above construction and observations, a (possibly innite) extensive game
form with imperfect information can be represented by a tuple
V, _, I, p, (H
i
)
iI
, (

)
1
(1) , (2.7)
24
Until recently, the framework of this thesis allowed the structure of information sets H
i
to violate perfect
recall. Following a mathematical point of view, I was convinced that this additional assumption was in
some sense restrictive, as far as the main result holds only in a proper subset of possible games. As my
advisor P. Battigalli noticed me, my belief was not correct: historically, the violation of perfect recall was
introduced in the rules of the game to represent with a technical trick the popular card game of bridge
between two pairs of players as a two-players game, allowing forgetfulness (see [13]). The fact is that, apart
from the technical trick, the violation of this property has nothing to do with the rules of the game, i.e.
depending on last instance only on the features of the players. Nevertheless, it can be shown without much
diculty that all the results which are going to be proved in the following sections hold also for games
without perfect recall, conditioning on the fact that it is assumed the knowledge of the sequence of partitions
(P
i
)
iI


iI
Part (v(i)) which represents the points where such property does not hold, relying on the
trivial isomorphism 1
i

= v(i) for all i I.
2.4. :-reduced normal forms 31
such that (V, _) V
std
and the property of perfect recall given in (2.6) is satised. The
collection of such extensive games will be denoted with G.
2.4 :-reduced normal forms
In game theory, the term strategy is sometimes wrongly confused with move. On the
one hand, the latter is an action taken at the some point of the game, that is a
i,
A
i,
for
some (i, ) /. On the other hand, the former is a complete algorithm which completely
denes the evolution at every possibile situation where the players could be active. As far
as by construction each player i I is not able to distinguish between histories in the same
information set, i.e. histories which belong to the same element of the quotient space 1
i
/ ,
then a strategy has to be dened a ordered sequences of actions, one for each vertex associated
with a histories in each information set.
25
Denition 9. Given a game G, the set of strategies of player i I is dened as
S
i
=

(h
i,j
)
jJ

jJ
H
i,j
A
i,
h
i,j
. (2.8)
According to the construction of the information partition H
i
= H
i,j

jJ
given in Denition
8, the cartesian product S
i
is invariant to the choice of the sequence (h
i,j
)
jJ


jJ
H
i,j
:
again, such choice is possible thanks to the AC. Denote also with S the set of sequences of
strategies of each player, that is

iI
S
i
.
26
Lemma 14. Given a game G and a prole of strategies s o, the terminal path z :
is uniquely dened.
Proof. Suppose for the sake of contradiction that there exists a strategy s = (s
i
)
iI


iI
S
i
such that it allows (at least) two distinct terminal paths z, z
t
:, let us say z = (
n
) and
25
That is the reason why in literature games with simultaneous are (wrongly, again) dened strategic
games.
26
Notice that the isomorphism between cartesian products S

=

iI

(hi,j)
jJ

jJ
Hi,j
A
i,
h
i,j
is not
simply an identity.
32 Chapter 2. Extensive Game Forms
z
t
= (
t
n
).
27
Then the set
V
t
=
n

t
n

has to be non-empty (indeed the root


r
has to belong to V
t
) and nite (in the opposite case
it should be
n
=
t
n
= z = z
t
). In particular V
t
belongs to V , meaning that it can
be completely ordered by the immediate relation . Call the vertex such that _ for
all V
t
, which exists and it is uniquely dened. Then by construction
s
1
()
n
, = s
1
()
t
n
.
It means that the prole of strategies s S allows the game with simultaneous moves

to
be played, and that (at least) two distinct vertices will be identied there by a sequence of
actions (a
i,
)
ip()
A

. This is impossibile according to (2.3).


Then, Lemma 14 explains that the map : S :, which associates each strategy s S
into the terminal path z : is well-dened. From here later, it will be termed outcome
function.
Denition 10. For each extensive game form = V, _, I, p, (H
i
)
iI
, (

)
1
(1) G,
dene its :-normal form as the normal form game
28
nf
:
() = I, (S
i
)
iI
, :, .
As far as the aim of this thesis is to obtain a complete characterization of behavioral
equivalence, it is natural to consider the sequence of partitions
(S
i
)
iI


iI
Part(S
i
)
obtained by behaviorally equivalent strategies. Here, for each player i I, two strategies
27
Again, the subscript is hidden to mean that such sequence can be nite or at most innite countable.
28
The normal form game is a description of a game: it is not graphical per se, but rather represents
the extensive game by way of a (possibly innite) matrix. Usually it is dened with respect to conceivable
strategies of each player, and associated payos. As far as we want to take into consideration only the rules
of the game, the denition of normal form has been restricted to terminal paths.
2.4. :-reduced normal forms 33
s
i
, s
t
i
S
i
are said to be behaviorally equivalent, if and only if
j I i, s
j
S
j
,
_
s
i
, (s
j
)
jI\i
_
=
_
s
t
i
, (s
j
)
jI\i
_
. (2.9)
Let us dene the equivalence relation
29
on S
i
such that s
i
s
t
i
if and only if they satisfy
(2.9).
Denition 11. For each extensive game form = V, _, I, p, (H
i
)
iI
, (

)
1
(1) G,
dene its :-reduced normal form as the normal form game
rnf
:
() = I, (S
i
)
iI
, :, ,
where S
i
= S
i
/ is the quotient space of behaviorally equivalent strategies.
Notice that the function has to be surjective by construction: indeed for each terminal
path (
n
) :, a prole of strategies (s
i
)
iI
S can be constructed such that each game with
simultaneous move
n
uniquely identies the vertex
n+1
by a prole of actions (a
i,n
)
ip(n)
.
It allows to conclude that its inverse
1
and projections proj
S
i
have to be correspondences.
Lemma 15. Given a game G which can be represented by (2.7), the following identities
hold
i I, S
i
=

(s
j
)
jI\{i}

jI\{i}
S
j
_
_
proj
S
i

_
1
(z)
_
z:
.
Proof. For each prole of strategies of the opponents (s
j
)
jI\i


jI\i
S
j
, clearly
_
_
proj
S
i

_
1
(z)
_
z:
Part(S
i
).
Indeed the inverse map of the function proj
S
i
: S
i
: allow us to obtain the obtain
the partition of S
i
which is behaviorally equivalent to the xed prole of strategies of the
opponents (s
j
)
jI\i
. Under the AC, the claim follows by Lemma 7.
Now, its presented an algorithm to explicitely calculate the partition S
i
, assumed that
S
i
is a nite or at most countable set in a game G:
29
Recall that an equivalence is relation which satises the reexive, transitive and symmetric properties.
34 Chapter 2. Extensive Game Forms
1. Choose randomly a partition P Part(S
i
).
2. Denote r(P) the reduction of the independent sets P P with respect to behavioral
equivalence, i.e.
r(P) = P/
PP
Part(S
i
).
3. According to Lemma 7, its well-dened the partition
r

(P) =

nN
r
n
(P),
where as usual r
0
(P) = P for all P Part(S
i
). For example, the sequence of
partitions (r
n
(P))
nN
becomes denitively constant whenever S
i
is a nite set: indeed
P is coarser than r(P), which is in turn coarser than s
i

s
i
S
i
.
4. Let P
j

jJ
Part(S
i
) the collection of all partitions of S
i
such that P
j
is coarser
than r

(P) and s
i
, s
t
i
P
j
P
j
= s
i
s
t
i
.
30
Then
S
i
=

jJ
P
j
.
It is clear that, if S
i
is a nite set, then this algorithm ends in a nite number of steps.
More precisely, r

(P) = r
n
(P) for all integers n [S
i
[, and we can obtain r

(P) in at
most [S
i
[ steps. Moreover, the collection of (P
j
)
jJ
at point 4 has a number of elements [J[
which is smaller than the number of partitions of [1, [r

(P)[]
Z
, which is in turn smaller than
2
[r

(P)[
1 (that is, the number of non-empty subsets of [1, [r

(P)[]
Z
). Considering that
[r

(P)[ [S
i
[, the inequality [J[ < 2
[S
i
[
holds. In conclusion, if S
i
is a nite set then the
quotient space S
i
can be calculated in less than [S
i
[ +2
[S
i
[
reduction processes. Let us rene
a bit this upper bound.
Lemma 16. The following inequality holds for all positive integers n:
[Part ([1, n]
N
)[
_
n + 1
2
_
n
.
30
In other words, it means that P
j
is a partition of S
i
such that it is coarser than r

(P), and ner than


S
i
. The advantage of this algorithm is that the reduction processes can be applied to smaller subsets of S
i
,
i.e. to each P P, allowing probably a smaller amount of computations.
2.4. :-reduced normal forms 35
Proof. Let us verify the claim by induction. It is trivially veried for n = 1. Suppose that
it has been proved that [Part ([1, n]
N
)[ n! for some positive integer n. Then each partition
in Part ([1, n + 1]
N
) can be in form P, n where P Part ([1, n]
N
) or n belongs to some
P P. Considering that [P[ n then
[Part ([1, n + 1]
N
)[ (n + 1) [Part ([1, n]
N
)[ (n + 1)!.
Then it could be enough to prove that 2
n
n! (n + 1)
n
for all positive integers n. And this
is case: if n is odd then
n! =

j[
1n
2
,
n1
2
]
Z
n + 2j + 1
2
=
n + 1
2

j[1,
n1
2
]
Z
_
n + 1
2
_
2
j
2
,
otherwise
n! =

j[1n,n1]
2Z+1
n + j + 1
2
=
n + 1
2

j[1,n1]
2Z+1
_
n + 1
2
_
2
j
2
.
It completes the proof, since they are clearly not greater than 2
n
(n + 1)
n
.
Of course, heavy machinery like partial Abel summation, Laplace approximation or
asymptotic formulae of Hardy and Ramanujan, the upper bound given in Lemma 16 can
be considerably improved. Nevertheless, its enough to prove that, if G is taken such
that I and S
i
are nite sets, then the number of steps required by such algorithm to compute
the rnf
:
() from the nf
:
() is at most

iI
[S
i
[ +
_
[S
i
[+1
2
_
[S
i
[
[I[ +

iI
[S
i
[
[S
i
[
.
Lemma 17. For each game G such that I and S
i
are nite sets, call c() the number of
comparisons between pair of strategies s
i
, s
t
i
S
i
to check if they are behaviorally equivalent.
Then
[I[ +

iI
[S
i
[ c()

iI
[S
i
[
2
/2.
Proof. As far as S
i
is a nite set for all i I, then its enough to check which strategies
s
i
S
i
are behaviorally equivalent to a xed s
i
S
i
. In the best case S
i
/ = S
i
and its
36 Chapter 2. Extensive Game Forms
enough to check exactly [S
i
[ 1 equivalence relations. In the worst case S
i
/ = s
i

s
i
S
i
so that exactly
_
[S
i
[
2
_
comparisons are needed. The claim follows considering that
_
n
2
_
is not
greater than n
2
/2 for all integers n 2.
Remark 12. It should be clear now, according to this construction, why the propery of
perfect recall has been assumed in the collection of games G: its not possible from the
:-reduced normal form rnf
:
() to infer if this property has been violated or not. Take as
example the games in Figure 2.5.
A
B A
z
2
z
3
z
4
z
1
a
1
a
2
b
1
b
2
a
1
a
2
A
z
1 B
z
2
z
3
a
1
a
2
b
1
b
2
Figure 2.5: Extensive game forms and
t
, respectively.
Clearly
t
satisfy the property of perfect recall given in (2.6), while does not. Nevertheless
nf
:
() = rnf
:
()

= rnf
:
(
t
) = nf
:
(
t
).
CHAPTER 3
A characterization of Behavioral Equivalence
3.1 The set of Invariant transformations
As far as the main result of this work aims to provide a complete characterization of behavioral
equivalence between possibly innite extensive game forms with imperfect information, it is
nally time to introduce our equivalence relation = contained in GG. Two extensive game
forms ,
t
G are said to be behaviorally equivalent if they share the same :-reduced
normal form,
1
i.e.
=
t
rnf
:
()

= rnf
:
(
t
).
Let G Part(G) be the partition of the collection of extensive game forms which is obtained
by the quotient space G/ =. It is going to be proved that two extensive game forms ,
t
belong to the same member G G if and only if one can be transformed into the other one
through a sequence of predened transfomations.
1
Here, the adjective same stands for up to isomorphism, which in turn is not equivalent to up to
relabellings of players, action and strategies. Indeed two extensive game forms could dier only by the
existence of an additional player who is forced to play an action. There will be a trivial isomorphism between
such pair of :-reduced normal forms, but there does not exist any relabelling.
37
38 Chapter 3. A characterization of Behavioral Equivalence
Denition 12. Given a non-empty set ( G, dene T(() the set of all invariant
transformations T : D
T
(() I
T
(() such that
rnf
:
() is isomorphic to rnf
:
(T ()) for all D
T
((),
where D
T
(() and I
T
(() are two non-empty subsets of (.
Remark 13. Without loss of generality the invariant transformation T (() is surjective.
Indeed its enough to restrict T (D
T
(()) to
I
T
(() :=
_
D
T
()
T ().
For all vertices
R
1 dene
/

R
:= i T

R
,,
R
P(
R
): A
i,
R
= /

R
(3.1)
as the set of players active at least once in the truncated tree (Z

R
, c

R
,

R
), not active at
the root
R
and such that they are not able to distinguish between any action choosen at
vertex
R
by the subset of players P(
R
). Observe that there is no reason to infer that such
set is non-empty.
Remark 14. Notice that, according to (3.1), we must have
/
i,
= /
i,
for all
t
,
tt
V

R
,,
R
,i
and all players i /

R
.
Remark 15. According to (3.1) the inclusion (3.2) necessarily holds, while (3.3) not:
/

R

_
1Z
(
R
,A
R
)
l
1
T
,,
, (3.2)
/

1Z
(
R
,A
R
)
l
1
P(). (3.3)
3.1. The set of Invariant transformations 39
Lets see two examples of invariant transformations; the next section will show their
importance.
Denition 13. The transformation T
0
. The domain of such transformation is dened as
D
T
0
(G) := G:

1
[/

[ 1,
and call
R
one of such vertices.
Then T
0
() =
t
:=
_
Z
t
, c
t
,
t
, Q
t
p
, P
t
, (
t

)
1
, (1
i,j
)
t
ip,j
i
_
such that
Z
t
:= Z V

R
,,
R
,i
: P() = /

R
;
Q
t
p
= Q
p
;
P
t
() = P() for all Z
_

R
V

R
,,
R
,i
_
; P
t
(
R
) = /

R
P(
R
) and
P
t
() = P() /

R
whenever V

R
,,
R
,i
such that P() ,= /

R
;
/
t
i,
= /
i,
whenever i P() P
t
(). Otherwise /
i,
R
= /
i,
for all i /

R
and
a vertex V

R
,,
R
,i
;
the set of edges c
t
and the map
t
change uniquely accordingly to previous points;
g

= g
t

for all vertices 1 1


t
;
P
f,i
changes according to previous points.
In rough words, transformation T
0
applies to all extensive games such that there exists a
vertex
R
1 with a non-empty set /

R
. Then it movesall actions of players i /

R
(who cannot distinguish by construction any chosen move at vertex
R
) to the vertex
R
itself. Vertices V

R
,,
R
,i
1 disappearif and only if the subset of players of

is
exactly /

R
. The partition P
f,i
and the collection g

1
is uniquely dened by P
h,i
, which
in turn is isomorphic to 1
i,j

ip,j
i
thanks to Proposition ??; then, the partition P
t
f,i
and
the collection g
t

1
changes according to Denition 13, dening uniquely the collection of
information sets 1
t
i,j

i
p
,j

i
, again uniquely dened by Proposition ??.
40 Chapter 3. A characterization of Behavioral Equivalence
Remark 16. Notice that according to (3.1) for all i /

R
there exists a integer t
i
Q

i
such that
(, . . . , )
V

R
,A
R
,i
1
i,t
i
.
Denition 14. The transformation T
1
. The domain of such transformation is dened as
D
T
1
(G) := G:

ip

v(i)
(1 +[/
i,
[) = 0,
so that if D
T
1
(G) then a pair (i
C
,
C
) Q
p
1 can be xed so that
C
v(i
C
) and
/
i
C
,
C
is a singleton.
Then T
1
() =
t
:=
_
Z
t
, c
t
,
t
, Q
t
p
, P
t
, (
t

)
1
, (1
i,j
)
t
ip,j
i
_
such that
Z
t
:= Z
C
if P(
C
) is a singleton, and Z
t
:= Z otherwise;
Q
t
p
= Q
p
i
C
if [v(i)[ = 1, and Q
t
p
= Q
p
otherwise;
P
t
(
C
) = Pi
C
, and P
t
() = P() otherwise;

t

C
=

P
t
(
C
), (/
i,
C
)
iP(
C
).i
C

,
t
_
, and
t

otherwise;
the set of edges c
t
and the map
t
change uniquely accordingly to previous points;
P
t
f,i
= P
f,i
whenever i Q
p
i
C
, otherwise P
t
f,i
C
is identical to P
f,ic
, except for the
cancellation of the vertex
C
from v(i
C
);
P
t
g,i
= P
g,i
whenever i Q
p
i
C
, otherwise P
t
g,i
C
is identical to P
g,ic
, except for the
cancellation of the vertex
C
from v(i
C
).
Loosely speaking, transformation T
1
is dened on all extensive games such that there
exists a vertex
C
1 with [/
i,
C
[ = 1 for some player i P(
C
). Then T
1
just deletes
the forcedaction of that player at such vertex. Thanks again to Proposition ??, new
information sets are uniquely dened. Moreover, the collection of information sets essentially
do not change, as far as the new unordered pair of partition P
t
f,i
, P
t
g,i
is isomorphic to the
old one P
f,i
, P
g,i
for all players i Q
t
p
.
3.2. Necessary and sucient conditions 41
3.2 Necessary and sucient conditions
Denition 15. The transformation T

. Given a integer n 1 and a vector =


(
1
, . . . ,
n
) 1
n
1
, the domain of such transformation is dened recursively as
D
T
(() := D
Tn
(() T
n
_
D
T
(
1
,...,
n1
)
(()
_
,
so that if D
T
(() then
T

() =
_
T
n
T

n1
. . . T

1
_
().
Proposition 1.
_
n1
_
1
n
1
T

T(() for all non-empty ( G.


Proof. Lets show the claim by induction on n. First, it has be shown that

1
1
T

T(().
For all D
T
0
(() the transformed extensive game T
0
() has exactly the same set of
strategies o
i
for all players i Q
p
, implying that
nf
:
() = nf
:
(T
0
()) ,
that is even stronger than the claim T
0
T((). At the same time for all D
T
1
(() the
transformed extensive game T
1
() has not the same set of strategies: more precisely o
i
= o
t
i
for all players i Q
p
i
C
and
o
t
i
C
=

v(i).
C

/
i
C
,
,
that is isomorphic to o
i
=

v(i)
/
i
C
,
, as far as /
i
C
,
C
is a singleton. Then
_
o
(red)
i
C
_
t
is isomorphic to o
(red)
i
C
, implying that

ip
_
o
(red)
i
_
t
is isomorphic to

ip
o
(red)
i
as well.
Following Denition 12, we obtain T
0
, T
1
T((). Suppose now for a integer n 1, its
proved that
_
mN[1,n]
_
1
m
1
T

T(().
Lets show then that the inclusion (3.2) holds also for n + 1. Given e 1
1
and
42 Chapter 3. A characterization of Behavioral Equivalence
= (
1
, . . . ,
n
) 1
n
1
, we have that D
T
(
1
,...,n,e)
(() D
Te
(() and T
(
1
,...,n,e)
() = T
e
(T

()).
By the inductive hypothesis we have that rnf
:
() is isomorphic to rnf
:
(T

()). But also


T

() D
Te
((), hence rnf
:
(T

()) is isomorphic to rnf


:
(T
e
(T

())) as well. Since the


existence of isomorphism is a equivalence relation (and in particular it is transitive), the
inclusion (3.2) holds for all integers n 1.
In few words, its going to be shown that T
0
and T
1
work independetly from each other,
and they can be applied only a nite number of times, reaching a transformed extensive
game that leaves unchanged its :-reduced normal form and its invariant with respect to the
choice of the order of such transformations. Lets rst dene the functions

0
: G N:

ip

j
i

/
i,v
(h
i,j
)
(h
i,j
)

,
and

1
: G N:

iP()
[/
i,
[.
Proposition 2. For all extensive games D
T
0
(G) D
T
1
(G) consider the non-empty set



n1
1
n
1
such that

if and only if
D
T
(G)
_
D
T
(G)
_
e1
1
(D
Te
(G) T
e
(I
T
(G)))
_
.
Then T

1
() = T

2
() for all
1
,
2

.
Proof. If a extensive game belongs to D
T
0
(G) D
T
1
(G) then at least one invariant
transformation between T
0
and T
1
can be applied. According to Denitions 13 and 14,
observe that

0
() =
0
(T
0
()) and
1
()
1
(T
0
()) + 1,
for all games D
T
0
(G) and

0
()
0
(T
1
()) + 1 and
1
() =
1
(T
1
()) ,
3.2. Necessary and sucient conditions 43
for all games D
T
1
(G). It implies that if D
T
(G) for some 1
m
1
then
(
0
+
1
) () (
0
+
1
) (T

()) m.
Once is xed, then (
0
+
1
) () is xed as well and (
0
+
1
) (T

) 1: hence the positive


integer m is upper bounded by (
0
+
1
) () 1, i.e. only a nite number of transformations
T
0
and T
1
can be applied to each game . We can do better: each transformation T
0
and T
1
applies independently from each other in the sense that the application of one of them has no
eects on the potential application of the other one. Moreover T
0
has to be applied exactly

1
[/

[ times, and T
1
has to be applied exactly

iP()
([/
i,
[ 1) times, where
is the Dirac-delta function [6]. The result is that T

() is invariant for all choices of in

.
Proposition 2 explains why the following Denition is well-posed.
Denition 16. Dene the map : G G such that () := T

() whenever belongs to
D
T
0
(G) D
T
1
(G) for some in

, otherwise () := .
According to Propositions ?? and ??, the following Denition is well-posed too.
Denition 17. Dene the map with domain G such that
() :=
_
Z, c, , Q
p
, P, (

)
1
, (P
g,i
)

ip
_
for all G.
Denition 18. Dene P(G) as the unique partition of P(G) such that two extensive games
with imperfect information
t
,
tt
belong to the same set of the partition P(G) if and only if
((
t
)) is isomorphic to ((
tt
)). Finally, dene ( the member of P(G) such that
( := P(G): P
g,i
= v(i), i Q
p
.
Remark 17. Notice that the set ( is strictly larger than the class of extensive game with
imperfect information and perfect recall.
44 Chapter 3. A characterization of Behavioral Equivalence
Denition 19. Dene the map
R: rnf
:
(G) rnf
:
(G):

Q
p
, (o
i
)
ip
,
_

Q
p
i : [o
i
[ = 1, (o
i
)
ip.i : [S
i
[=1
,
_
.
Remark 18. Observe that rnf
:
(
t
) is isomorphic to rnf
:
(
tt
) if and only if R(rnf
:
(
t
)) is
isomorphic to R(rnf
:
(
tt
)) for all extensive games
t
,
tt
G.
Proposition 3. R(rnf
:
(G)) and (() are isomorphic, i.e. there exists a bijective map
: R(rnf
:
(G)) (().
Proof. Once a triple

Q
p
, (o
i
)
ip
,
_
R(rnf
:
(G)) has been xed, it has to be shown that
there exists a unique (() such that
rnf
:
() =

Q
p
, (o
i
)
ip
,
_
and [o
i
[ 2 for all i Q
p
.
The assumption (() is equivalent to (P
g,i
)

= for all i Q
p
and , D
Te
(G) for all
e 1
1
. For all players i Q
p
dene the unique partition S
i
P(o
i
) such that
(S
i
, o
i
)
S
i
S
i
P(:) .
If S
i
= o
i
then the player i Q
p
is not active at the root ; otherwise i P() and his
set of action /
i,
will be isomorphic to Q
[S
i
[
. At this point vertex , the subset of players
P() and their potential actions /
i,
are completely dened, up to isomorphisms. Notice
that P() has to be non-empty, according to the Denition of :-reduced normal form.
For each vertex Z

1
its possible to dene its :-reduced normal form as the triple

Q
p
, (S
i
)
iP()
,
_
,
for each choice of (S
i
)
iP()


iP()
S
i
. Observe also that [Z
l
1
[ =

ip
[S
i
[. Since
such algorithm will terminate in a nite number of steps, then its uniquely dened a game
=
_
Q
p
, (o
i
)
ip
,
__
, under the assumptions that (() and [o
i
[ 2 for all
i Q
p
.
3.2. Necessary and sucient conditions 45
Remark 19. Given a strategic game R(rnf
:
(()), its uniquely dened a extensive game
= ()
_
(
_
, but no information about the set of partitions P
g,i

ip
can be inferred
from the :-reduced normal form. It means that all results that we are going to obtain
are conditioned to the membership of some xed xed member of the partition P(G) of the
Denition 18.
Proposition 4. Fix a class ( P(G) and two extensive games
t
,
tt
(. Then rnf
:
(
t
)
is isomorphic to rnf
:
(
tt
) if and only if (
t
) is isomorphic to (
tt
).
Proof. Fix
t
,
tt
(, i.e. such that ((
t
)) is isomorphic to ((
tt
)), and particular
(P
t
g,i
)

is isomorphic to (P
tt
g,i
)

for all players i; recall that since Q


p
is isomorphic to Q
p

then the number of players has to be the same.


Suppose that (
t
) is isomorphic to (
tt
). Since the map : G G is just a
composition of a nite number of transformation T
0
and T
1
according to Denition 16, then
R(rnf
:
()) = rnf
:
(()) for all G, and in particular they are isomorphic, according
to Remark 18. Hence rnf
:
((
t
)) is isomorphic to rnf
:
((
tt
)), which in turn implies that
rnf
:
(
t
) is isomorphic to rnf
:
(
tt
). Figure 3.1 shows a graphical representation of the proof
of the if part, where red lines stands for isomorphic to.
(
t
) rnf
:
((
t
)) R(rnf
:
(
t
)) rnf
:
(
t
)
(
tt
) rnf
:
((
tt
)) R(rnf
:
(
tt
)) rnf
:
(
tt
)
=
Assumption
Figure 3.1: Structure of the proof of if part of Proposition 4
Suppose on the contrary that rnf
:
(
t
) is isomorphic to rnf
:
(
tt
). Then also R(rnf
:
(
t
)) is
isomorphic to R(rnf
:
(
tt
)) according to Remark 18. Then also (R(rnf
:
(
t
))) is isomorphic
to (R(rnf
:
(
tt
))) according to Proposition 3. They uniquely dene elements in ((), hence
they uniquely dene a elements in (() as far as ( is isomorphic to ( for all classes ( P(G).
46 Chapter 3. A characterization of Behavioral Equivalence
Since by assumption
t
and
tt
belong to the same of the partition P(G), then also (
t
) is
isomorphic to (
tt
). Figure 3.2 shows a graphical representation of the proof of the only if
part, where red lines stands again for isomorphic to. It concludes the proof.
rnf
:
(
t
) R(rnf
:
(
t
)) (R(rnf
:
(
t
)))
rnf
:
(
tt
) R(rnf
:
(
tt
)) (R(rnf
:
(
tt
)))
Assumption
=
=
=
( (()
( (()
(
t
)
(
tt
)
Figure 3.2: Structure of the proof of only if part of Proposition 4
Corollary 1. Let
t
,
tt
be two extensive games such that P
t
g,i
= v
t
(i) for all i Q
p
and
P
tt
g,i
= v
tt
(i) for all i Q
p
. Then rnf
:
(
t
) is isomorphic to rnf
:
(
tt
) if and only if (
t
) is
isomorphic to (
tt
).
Proof. The assumption is equivalent to choose
t
,
tt
in the class ( P(G). The result
follows from Proposition 4.
Proposition 5. Fix a class ( P(G). Then T(() =
_
n1
_
1
n
1
T

.
Proof. Fix a game (, for some ( P(G). Recall that T(() represents the set
of transformations that are invariant with respect to the :-reduced normal form, up to
3.2. Necessary and sucient conditions 47
isomorphisms. It means that if T T(() then rnf
:
() is isomorphic to rnf
:
(T ()), whenever
D
T
(() (. According to Proposition 4, it happens if and only if () is isomorphic to
(T ()), but the map is just a composition of a nite number of transformation T
0
and
T
1
, from also the inclusion T(()
_
n1
_
1
n
1
T

holds true. The claim follows adding the


result of Proposition 1.
Corollary 2. Let
t
,
tt
be two extensive games such that P
t
g,i
= v
t
(i) for all i Q
p
and
P
tt
g,i
= v
tt
(i) for all i Q
p
. Then rnf
:
(
t
) is isomorphic to rnf
:
(
tt
) if and only if there
exists a nite sequence of games
0
,
1
, . . . ,
n
such that
0
:=
t
,
n
is isomorphic to
tt
and
j
= T
e
(
j1
) or
j1
= T
e
(
j
) for all j Q
n
, and some e 1
1
.
Proof. Its enough to set ( = ( in Proposition 5.
.
Corollary 3. Let
t
,
tt
be two extensive games with imperfect information and perfect recall.
Then rnf
:
(
t
) is isomorphic to rnf
:
(
tt
) if and only if there exists a nite sequence of games

0
,
1
, . . . ,
n
such that
0
:=
t
,
n
is isomorphic to
tt
and
j
= T
e
(
j1
) or
j1
= T
e
(
j
)
for all j Q
n
, and some e 1
1
.
Proof. It follows from Corollary 2 and Remark 17.
.
Remark 20. Since transformations T
0
and T
1
can be applied at most once for every vertex
1 and player i P(), it follows that the number n of the nite sequence of games of
Corollaries 2 and 3 veries the following chain of inequalities:
n 4

1
[P()[ 4p[1[ 4p([Z[ 2).
It implies that if
t
and
tt
belong to ( (and in particular when they have perfect recall)
and
t
does not reach a game structure isomorphic to
tt
with at most 4p([Z[ 2) through
48 Chapter 3. A characterization of Behavioral Equivalence
only transformation T
0
and T
1
then they do not have a isomophic :-reduced normal form.
In particular, the maximal number M of tries is upper bounded by
M

j1
4p(|Z|2)
2
j
2
4p([Z[2)+1
16
p[Z[
.
3.3 Extension to chance moves
CHAPTER 4
Invariance of Solution Concepts
Complete.
4.1 Renements of equilibrium concepts
4.2 Rationalizability in extensive games
49
CHAPTER 5
Conclusions
One may ask what is the additional value to a well-dened notion of quotient space with
respect to some equivalence relation by the explicit specication of all transformations under
which it remains invariant.
One the one hand, the rst reason could be precision. Transformations on extensive
games can give us insight into the nature of strategic structure, to make explicit the
logical equivalence of transformation rules. On the other hand, another answer could
be transparency. For example, topological structure remains invariant under arbitrary
stretching of the plane, rotations, and translations. In geometry, similarity is invariant under
translations, dilations and rotations, and congruence is not invariant to second one of them.
These various equivalence relations becomes more transparent stated in terms of permissible
transformations.
50
Bibliography
[1] G.Birkho, Lattice theory, American Mathematical Society, 1948.
[2] R.Diestel, Graph theory, Second Edition, Springer, 2000.
[3] S.Elmes and P.Reny, On the Strategic Equivalence of Extensive Form Games, Journal
of Economic Theory, 1994.
[4] D.M.Kreps, A course in microeconomic theory, Princeton University Press, 1990.
[5] E.Landau, Vorlesungen uber Zahlentheorie, Vol.2, Leipzig, 1927.
[6] M.J.Lighthill, Introduction to Fourier Analysis and Generalized functions, Cambridge
University Press, 1958.
[7] H.Minkowski, Geometrie der Zahlen, Chelsea.
[8] T.Nagell, Introduction to Number Theory, New York: Wiley, 1951.
[9] M.J.Osborne and A.Rubinstein, A course in game theory, MIT press, 1994.
[10] H.L.Royden, Real Analysis, Second Edition, Macmillan, 1968.
[11] R.Stanley, Enumerative Combinatorics, Volume I, Second Edition, Cambridge
University Press, 2011.
51
52 Chapter 5. Conclusions
[12] F.B.Thompson, Equivalence of Games in Extensive Form, The RAND Corporation,
1952.
[13] J.von Neumann and O.Morgenstern, Theory of Games and Economic Behavior,
Princeton University Press, 1953.

Вам также может понравиться