and AppliedLogic 28 (1985) 137-161

Annals of Pure 137

North-Holland

PROBABILm LOGIC WITH CONDITIONAL EXPECTATION*

Sergio FAJARDO
Department of Mathematics, University of Wisconsin, Madison, WI 53705, USA
(Current address: Dept. of Math., Uniu. of Colorado, Boulder, CO 80309, USA)

Communicated by K. Kunen
Received 19 August 1983

0. Introduction

Probability logics are logics adequate for the study of structures arising in
Probability Theory. The newest member of this family is the Probability Logic
with Conditional Expectation, denoted L,,, recently introduced by Keisler in
[16]. In this paper we develop the model theory of this logic. We answer
affirmatively problems 5 and 6 of Keisler’s paper, regarding the existence and
uniqueness of hyperfinite (‘saturated’) models and the validity of the Robinson
consistency property and Craig interpolation theorem for L,,. Here is a summary
of the history of Probability Logics and a description of the contents of this paper.
The development of the model theory of first-order logic has brought up the
need for the study of logics with a stronger expressive power. This allows us to
incorporate into the realm of logic certain common mathematical structures and
concepts that have been left out of first-order logic due to its limited scope. Most
of the new logics that have been studied focus on structures arising in algebra and
set theory. Topological Logic is an example of a well behaved logic with
‘enriched’ structures, its models besides a first order part also have a topology on
the universe (see Zeigler’s survey paper [28]). An up to date account of the
research done during the last twenty years in several extensions of first order logic
can be found in the forthcoming book [5].
Keisler in [17] began the study of the Probability Logics L, and L,. These
are logics where the quantifiers Vx and 3x are not allowed and instead the
quantifiers (Px > r) and J dx are respectively incorporated. The most interesting
structures (models) for LAP and L, are of the form (A, X, p), where p is a
probability measure on A and X is a random variable defined on A. The model
theory of the logics has been developed further by Hoover in [ll-141 and Keisler
in [16]. Later on, inspired by Keisler’s work on stochastic differential equations

*I wish to thank my advisor Professor H.J. Keisler for his continuous interest in my work and
invaluable comments. This research was partially supported by a University of Wisconsin research
assistantship.

0168-0072/85/$3.30 0 1985, Elsevier Science Publishers B.V. (North-Holland)

138 S. Fajardo

[lS], Hoover and Keisler (see [16], [19]) introduced Adapted Probability Logic
denoted L,. This is a logic adequate for the study of stochastic processes. Its
structures are of the form (A, X= (X(t)),,,,, rl, (%t)teC,,,il, p), where p is a
probability measure on A, X is a measurable stochastic process defined on A with
time parameters in [0, l] and ($t)tEl,,, i, is an increasing sequence of w-algebras of
t.~-measurable sets. For information about the role of these structures in the
theory of stochastic processes see [8] and [9]. Hoover and Keisler in [15]
introduced in a probabilistic context two notions of elementary equivalence for
Lad; using nonstandard analysis tools and model theoretic techniques they gave
direct applications to problems in Probability Theory. Rodenhausen in his thesis
[25] presented a set of axioms for L,, and proved a completeness theorem for this
logic. His proof was very long and complicated due to the apparent lack of
connection between L,, and L,, with Lad.
With the introduction of L,, the gap was filled. This logic extends L, with the
addition of a new operator E[]] that allows one to talk about conditional
expectations of random variables with respect to o-algebras. Its simplest struc-
tures are of the form (A, X, 9, FL) with (A, X, CL)as for L,, and % a o-algebra of
F-measurable sets. Keisler in [16], [20] proved a completeness theorem for L,,
introduced Lad in a two-sorted form of L,, and using the completeness theorem
for this logic gave a simpler and more natural proof of Rodenhausen’s complete-
ness theorem for Lad.
In this paper we continue the study of LA,. In Section 1 we give a direct
definition of L, without assuming a previous knowledge of either of L, or L,,
and list some basic facts about L,,. The reader is assumed to be familiar with
some elementary probabilistic concepts such as the definition of conditional
expectation and some of its properties; these can be found in any introductory
probability textbook (see for example [3] or [6]). We also indicate in this section
some possible generalizations of L,, by allowing more than one conditional
expectation operator symbol; this permits us to study in LAE stochastic processes
with discrete time without having to go to Lad.
Sections 2 and 3 contain the main theorems of this paper. We introduce a
notion of hyperfinite model for LAE that naturally extends the corresponding
notion of hyperfinite model for L, as presented in [16]. We prove that these
models exist (Section 2) and for a special type of hyperfinite models that we call
uniform we prove that they are unique (Section 3). Most of our proofs make use
of arguments and constructions from Nonstandard Analysis; in particular, we
assume the reader is familiar with the definition and main features of a nonstan-
dard universe and also has a knowledge of the Loeb measure construction. The
books [27] and [l] have good presentations of nonstandard analysis, [26] concen-
trates on the study of Probability Theory from a hyperfinite point of view and
Cutland’s survey paper [7] has a nice introduction to non-standard measure
theory aimed at the general mathematician.
In Section 4 we use the existence and uniqueness theorems of hyperfinite
 Probability logic with conditional expectation 139

models in order to give proofs of the Robinson consistency property and Craig
interpolation theorem for LAE.
A knowledge of basic model theory is desirable; with respect to Probability
Logic not much is needed but some familiarity with [16] will help to have a better
understanding of the whole subject of Probability Logics and in particular of the
results presented here. Our results have been previously announced in [lo] and
were presented at the 1983-84 ASL Annual Meeting held in Boston.

1. Basic definitions and background

L, was introduced by Keisler in [16] as an extension of the logic L,, which

allows us to express in a natural way probabilistic notions involving conditional
expectations of random variables with respect to a-algebras. In this section we
present the basic definitions for both logics and all those results about them that
are relevant to our exposition. A is an admissible set contained in HC (Hereditar-
ily countable sets) with w E A. For simplicity we restrict our attention to the case
A = HC and denote the logic so obtained by Lo,E. All our results can be readily
generalized to L,.

Definition 1.1. Throughout this paper we work with a countable set L =

{X> : i E I} of random variable symbols. The ni represents the number of argu-
ments of Xi and we will not bother to write it unless it is strictly necessary.
The Logical Symbols of L,,,+ are:
(a) A countable list of individual variables: u, U, x, y, z, . . .
(b) Connectives:
(1) First-order Connectives: 1, /\.
(2) Function Connectives: The symbol F, for each n EN and FE C(W) = the
set of continuous functions F : R” + R.
(c) Integral Quantifier Symbol: J.
(d) Conditional Expectation Symbol: E[ I].

Definition 1.2. The set of L,,,-atomic terms is defined by:

(a) For each random variable symbol Xi, u an n, -tuple of variables and r E Q’,
[X,(u) r r] is an L,,,-atomic term.
(b) If u1 and u2 are variables, then l(u, = UJ is an atomic term.
The set of LolE- terms is the least set such that:
(a) Every atomic term is a term.
(b) Each real number r is a term.
(c) If 71, . . . ) T, are terms and FE C(R”), then F(T~, . . . , T,,) is a term.
(d) If 7 is a term and u is a variable, then J T du is a term.
(e) If ~(u, V) is a term, V& u and x is a variable, then E[T(u, v) ( v](x) is a term.
140 S. Fajardo

Due to the introduction of the conditional expectation operator we have to be

careful in the way we handle the variables that appear in terms. Intuitively, the
quantifier J binds the variable u in j du and in E[T ) U] (x) the occurrence of n is
bound and that of x is free. But when confronted with terms of the form
E[T(u) 1u] (u) we have to decide whether the variable u is bound or free. We are
going to use the concept ‘free’ variable for those variables ‘with at least one free
occurrence’, this idea is made precise with the following definition.

Definition 1.3. Given a term T and a variable v we say that v is free in T

according to the following inductive definition:
(a) If T(U)is [X(u) 1 I], then u is free in T if and only if v E u.
(b) If T(u~, u2) is l(ur = uJ, then u is free in T if and only if u is ur or v is u*.
(c) If 7 is F(rr, ...,T,,), then v is free in T if and only if, for some i, v is free
in TV.
(d) If T is 5 13du, then n is free in T if and only if v is not u and v is free in 8.
(e) If r is E[0 ( z] (x), then n is free in T if and only if v is x or v is not z and it
is free in 19.

If v is not free in T we say that v is bound in T.With these definitions we can

see that u is free in E[T(u) 1u]( u ) an d u is bound in E[T(v, u) ) u](x). A closed
term is a term with no free variables.

Definition 1.4. The set of Lo,n -formulas is the least set such that:
(a) For each L,,,,n- term 7,[T2 0] is an (atomic) formula.
(b) If C$ is a formula, so is 14.
(c) If r is a countable set of formulas with finitely many free variables, then
/j r is a formula.
A sentence is a formula with no free variables.

Remark. In the previous definitions when we delete those clauses that involve
conditional expectations (i.e. (d) in Definition 1.1 and (e) in Definition 1.2), we
obtain the corresponding notions for L,,,,s. Given the way the conditional expecta-
tion operator is interpreted (see Definition 1.8) it turns out that it is important to
distinguish two types of variables among the free variables that appear on terms.
Intuitively we divide them in two groups: In one we have the ‘E-bound’ variables
which are those that when interpreted produce an S-measurable random variable
(see Definition 1.8 for undefined concepts) and in the other group we have the
‘E-free’ variables which are those that when interpreted produce random vari-
ables that are not g-measurable. These informal notions are made rigorous with
our next definition and the examples following it illustrate different cases.

Definition 1.5. Given an Lw,E-term T and a variable v, we say “v is E-free in 7”

according to the following inductive definition:
(a) If r(u) is [X(u) 1 I], then v is E-free in r if and only if for some i, v is 4.
 Probability logic with conditional expectation 141

(b) If 7(u1, u2) is l(u, = u,), then IJ is E-free in r if and only if ‘u is u1 or 2) is uz.
(c) If 7 is F(7t,... , T,), then u is E-free in T if and only if for some i, TVis
E-free in ri.
(d) If T is J 8 du, then u is E-free in 7 if and only if 21is not u and u is E-free
in 8.
(e) If 7 is E[8 ( z](x), then u is E-free in T if and only if u is not z and u is in 8.
If v is free in 7 and it is not E-free in T, we say that “v is E-bound in 7”.

Examples. (1) z is E-bound in E[[X(u, II) 1 r]) u](z) and u is E-free.

(2) If we let T(U, v) be E[[X(u, V) 1 r]) v](u), then u is E-free in T and z, is
E-bound in T.
(3) If we let T(U, u’, u) be [X(u, V) 1 r]+E[[X(u’, z) r r] 1z](v), then u is E-free
in T, u' is E-free in T and v is E-free in T.
Notice also that v is E-bound in E[[X(u’, z) 1 r]) z](v) and it is E-free in
[X(4 u) 1 rl.
(4) v is E-free in E[[X(v, z) 1 r] ) z](u).

As mentioned in the introduction we assume the reader is familiar with the

elementary probabilistic concepts that appear on the following definitions.

Definition 1.6. (a) A triple (A, 3, ZJ) is a probability space if $9 is a c-algebra on

A such that each singleton belongs to 3. Usually we just write (A, v). % is called
the set of v-measurable sets.
(b) (Keisler [16]). Let (A, 3, v) be a probability space and n E N. (A”, ‘SC”‘,v(“))
is the probability space where 3’“’ is the a-algebra on A” generated by the
rectangles of the form Gr x * * . x G,, with Gi E 93 for each i and the diagonal sets
of the form Dij = {x E A” : xi = xi}. d”) is the unique extension of the product
measure v” = Y x - - . x I, to SC”) such that v’“‘(Dij) = I,..* ~({x})“.
In general when we say that a random variable is v’“‘-measurable, we mean it is
measurable with respect to SC”).

Definition 1.7. (a) A random variable model for L is a structure % = (A, X”, v)iGI
where (A, v) is a probability space and each XT is a real-valued v’“i’-measurable
random variable (i.e. for every Bore1 subset B of R, (X”)-‘(B) belongs to %‘“~‘).
(b) A conditional expectation model for L is a structure of the form % =
(9&,, S”) where a0 is a random variable model for L and 8% is a u-algebra on A
contained in 3. We write % = (A, XF, .F%, v)ier.

Definition 1.8. An interpretation of LolE in a conditional expectation model % is

a function that assigns to each term T(u~, . . . , u,,) a v’“‘-measurable function
142 S. Fajardo

7% : A” -FL! such that:

(a)
r if X”(a) > r,

(l) [X,(u) 1 rlq(u)= e-r if X”(a) 6 -r,

X”(a) if -r <X”(a) s r.
1 if a,=az,
(2) ll(u, = %)PL% 4 = { if a,#a,.
0

(3) F(% . . . 3?JIIpL(~)= F(T?b), . . . ,7:(u)).

(4) [ I%, v) dull(u) = j 8%(u, b) du(b).

A

(b) (E[du, v) 1VI(X)>" (a, b) is 7~“‘)x CT-measurable and for each a E A”,
(E[T(u, u)g u](x))” (a, b) = E[r% (a, a) 1P] (b) for v-almost all b.

The following lemma is an immediate consequence of the definition of interpre-

tation.

Lemma 1.9. For every conditional expectation model % there exists an interpretation
of Lo,E in %. Two interpretations agree almost surely on each tern. The values of
closed terms and sentences are the same in all interpretations.

The reader can notice that in Definition 1.8 (a.1) the values of the random
variables are truncated. This is a technical point that makes the interpretation of
each term a bounded random variable in the model and consequently an integra-
ble random variable. Moreover, we can give uniform bounds for the interpreta-
tions of terms in arbitrary models as the following definition and lemma show.
Hereafter 7% is some interpretation in 021.

1.10. The bound ()7((of an LwlE-term T is defined by:

Definition
(a) Il[X, (u) I rlll = r,
(b) Ill(u = v)ll= 1,
(c) 115r
dxll= l]rll,
(4 llE[~ 1ul(v)ll= kdl,
(4 b%, . . . >7n>II = SUP{IF(SI, * . * 3sn>I
: ISiISIITiII),
(0 lIdI=I4

Lemma 1.11.Given a structure % for all a in A we have: (~~(a)1 G ~~T~~.

Proof. Induction on the complexity of terms. Cl

In [16, Section 3.51, we can find a set of axioms for I,,, and a proof of the
completeness theorem for L,,,. These axioms express natural properties of the
 Probabilityfogic with conditional expectation 143

integral and instead of writing them down we refer the reader to that section for
complete details.
In the same paper, Section 4.2, we find the following set of
axioms and rules of inference for LwqE (we shorten E[T ( u](u) to E[T ( u]).

Definition 1.12. (a) Axioms for L_+.

(AO) All axiom schemes for LwlJ.
(Al) E[T(u, x) I xl(u) = ECdu, Y>1YI(u> where x and y do not occur in u.
(A2) J E[a(u, u) ( U]T(U) du = J E[a(u, u) ( u] . E[T(u) ) u] du.
(b) Rules of inference for Lo+:
(Rl) Modus Ponens: 4, 4 + y tr.
(R2) Conjunction: (4 -+ y : y E r}k C#I+ AT.
(R3) Generalization: C#J
-+ [T(U, x) Z= 0] !- C#J
+ D T(U, x) du 2 01 where u is not
free in 6.

Theorem 1.13 (Keisler [16]). Soundness and completeness for Lo,E: A countable
set of sentences of Lw,E has a model if and only if it is consistent.

Definition 1.14 (Generalizations). In probabilistic practice we usually find situa-

tions where it is important to consider two or more u-algebras in the same model.
A particularly well known case is that of discrete stochastic processes where we
have structures of the form % = (A, (X,,)nEN, (g,,)neN, FL) such that for all n EN:
(a) X,, :A +LQ is a random variable.
(b) F,, L 9n+1 5 p-measurable sets.
L w,E can be extended in such a way that this type of model is covered. For each
n EN introduce a conditional expectation operator E,, where the interpretation of
E,,[T(u, u) ) u](x) in “11is E[T%((Q, a) 1 S,,](b) a.s.

If for each n EN we add to the list of axioms given in Definition 1.12 the new
axiom:

(A3), E,,[T(~, u) iul= E,+~[En[du, u> !ul! ~1,

then we have:

Theorem 1.15 (Generalized soundness and completeness for L_&.

Let r be a
countable set of sentences in LwlE with conditional expectation operators E, for each
nEN.
r has a model Ou= (A, (XF)iSl, (.Fn)nGN,p ) with 9,, G S,,+1 for each n EN if and
only if r is consistent in LolE with (A3),, n EN.

Proof. Soundness as usual is easy. For the other direction we just add a few
things to Keisler’s proof of the completeness theorem for LolE. Instead of defining
a single c-algebra $, for each n EN let 9,, be the a-algebra generated by the sets
144 S. Fajardo

of the form {d E A4 :X,(c, d)& 2 r} where T(U, V) is of the form E,,[u(u, w) ) w](v)
with v E-bound in r and c is in M.
The proof is finished as before using the axioms (A3),, n EN to guarantee that
the .5F,,‘s form an increasing sequence of u-algebras. Cl

Definition 1.16. Let % and 8 be Lo+ -models. Ou is said to be elementarily

equivalent to 93 if for each Lo,E -sentence 8, % b 0 if and only if 9 !=0. We denote
this relation by % =w,E .%. A trivial but useful observation is

Proposition 1.17. %!.l=o,E 93 if and only if for each closed term r, 7% = 7%.

It is clear from Definitions 1.2 and 1.4 that the set of L_,-formulas is
uncountable, but a result of Keisler shows that for most arguments it is enough to
consider a countable ‘dense’ set of Lw,E-terms. Here are some details.

Definition 1.18. A countable set D of terms is said to be dense in LolE if for

every term 7(x) of Lw,E and every E >O there is a term T&) ED such that
117-roll < e.

A particular instance of the famous Stone-Weierstrass theorem which is useful

to us is the following:

m-rem 1.19 (Stone-Weierstrass). The set 9 of polynomials with rational coefi-

cients is such that for each n EN, [r, s], FE C(W) and E > 0 there is F0 E 9 n C(R”)
such that:

sup{jF(x)-F,(x)l:x~[r, s]n}<~.

The next two results are applications of this theorem to Lo,+

Theorem 1.20 (Keisler [17]). If 9 is as above, then the set D of LolE-terms where
the only function connectives allowed come from 9 is a countable dense set of terms
in Lo,E.

Corollary 1.21 (Keisler [17]). Let % and 93 be L,>,-models. If D is a dense set of

terms such that for each closed term r E D, 7% = TV, then % =w,E 3.

2. Existence of hyperfinitemodels

Keisler has pointed out that hyperfmite models for LW,P and Lo,, can be
considered as the Probability Logic notion analogous to that of saturated models
for first-order logic. In this section we give definitions of hyperfinite models
 Probability logic with conditional expectation 145

(internal and external) for Lw,n and prove, see Theorems 2.11 and 2.12, that
given any model Ou for Lo+ we can find a (uniform) hyperhnite model which is
elementarily equivalent to %. The main tools that we use in our proofs come from
basic nonstandard analysis; [7] and [22] are short introductions to this subject. We
also make use of previous existence and uniqueness results of Keisler for
L,,l-hyperfinite models. For detailed proofs of these theorems see [16] and [17].
From now on we use a fixed or-saturated nonstandard universe. We also
simplify notation by giving definitions and proofs for L = {X} where X is a
2-argument random variable; of course, all results are valid for L as considered in
Definition 1.1.

Definition 2.1. (a) A model K = (N, XiN, I, c) is a *-finite model for Lo+ if:
(1) N is an internal set such that its internal cardinality IN] is hyperlinite (i.e.
IN] E “N ,N).
(2) X : N X N + “R is an internal function.
(3) I is an *-algebrac *P(N) = set of internal subsets N.
(4) @ is the *-counting measure on N.
(b) X as above is said to be a uniform *-finite model for Lo,E if:
(1) N=HxK with H, K internal such that IH], IK(E*N-~V.
(2) I is the *-algebra of internal subsets of N of the form A X K with
A E *S(H).
(c) A model Jbc = (N, X, 5, p) is said to be a (Uniform) Hyperfinite Model for
L w,E if there exists a (uniform) *-finite model .,V such that:
(1) X is a lifting of X. This means “X(w, w’) = X(0, w’) a.s. (&, where pz is
the *-counting measure on N x N.
(2) p is the Loeb measure obtained from fi.
(3) 9 is the u-algebra a(I) generated by I.
Given this situation we say that X is a lifting of JU and write “N=Ju.

Remark. Observe that if in part (a) of the above definition we leave out the
*-algebra I we obtain the so-called *-finite models for Lw,J. LmlJ-hyperfinite
models are defined as in part (c) but leaving out (3).

It is important to notice that for a *-finite model X the random variable X”

takes values in “R and not necessarily in R. We introduce the notion of
*-interpretation for *-finite models which is analogous to the one given in
Definition 1.8 for standard L,,,-models.

Definition 2.2. Let X = (fi, X, I, fi) be a *-finite model for Lo+. A *-

interpretation of Lo+ in K is a function that assigns to each L,,,-term r an
146 S. Fajardo

internal function i” satisfying the following conditions:

(1) If T(U) is [X(u) 1 r], r EQ+, then:

r if Xi-“(a) 2 r,
Y(a)= -r if Xi”(a) s-r,

1 8”(a) if -r Sri”(a) G r.

(2) If r(u, U) is l(u = v), then

1 if a = b,
?“(a, b) =
0 ifafb.

(3) If r is F(rl,. . . , T,), then

;“(a) = *F(+f(a), . . . , ;:(a)).

(4) If r(u) is j e(u, V) dv, then

;“(a) = c 8”(a, b) l/]N(.

beN

(5) If T(U, x) is E[8(u, v) ( v](x), then

?“(a, b) = E[oN(u, *) ) I](b).

Observe that in this definition C is internal summation and E[ *) I] is internal

conditional expectation with respect to I.
The following lemma is an immediate consequence of the definitions of the
nonstandard integral and nonstandard conditional expectation for hyperfinite
Loeb spaces (see for example Section 3 of [7]).

Lemma 2.3. For every *-finite model X there exists a *-interpretation of Lo,E in JY.

Lemma 2.4. If X is a lifting of 4, then for all Lul,-terms r we have

(Y(w)) = T”yO) a.s.

Proof. Induction on the complexity of terms. For the integral quantifier step use
Loeb’s Lifting Theorem (Proposition 2, p. 118 in [23]) and for the conditional
expectation operator use Theorem 12 (ii) from [2]. 0

Before we start proving the main theorems of this section, let’s take a closer
look to the elementary equivalence relation for L,,,-models introduced in
Definition 1.16. This gives us an opportunity to present results about G,,,,~ that
are important in their own right and to introduce some concepts that appear in
later proofs.

Definition 2.5. Let 9 be as in the Stone-Weierstrass Theorem 1.19. Let (Y,,),,~~

 Probability logic with conditional expectation 147

be an enumeration of the dense set D of L,,+ -terms obtained from 9 (Theorem

1.20).
(a) For each r,,,(u) let X,(u) be a new random variable symbol. Let L” =
{X,&_:m <n} and L”={X,&_:~E~J}.
(b) Let % = (A, XWL,9, u) be an &,-model.
(1) The Lz,J-model associated to 9.l is the model ‘-‘U,= (A, X%, (y%,+,, V>
where for each m <n rz is the interpretation of q”(u) in Q1,.
(2) The Lz,,-model associated to Ou is the model %,= (A, XQL, (Y:),,,~~, u>
where for each m EN 72 is the interpretation of q”(u) is “&.
(3) Let ?/mbe a term from our list.
(i) If y,,, is of the form E[O(u, x) 1x](v) with v E-bound in -y,,, let ST,, be
the c-algebra on A generated by the sets of the form {d E A : ?:(a, d) 2 r, Q E
A, ~EQ}.
(ii) If ?/mis not of this form let Y&_ = (4, A}.
(iii) For each neN. Let %,,=~(u,,,~,,%~~) and %=(+(UneN%,,).
(4) (i) The jk+- model associated to %, is the model % L = (A, XQL, ‘S,,,,v).
(ii) The &,-model associated to Ou, is the model %L = (A, XQL, 59, v).
(c) If instead of 011as in (b) we have an *-finite model X = (N, X”, I, V) we can
define corresponding internal notions as follows:
(l), (2): JV,, and .JV~ are defined in the obvious way.
(3) (i) If y,,, is of the form E[B(u, x) ( x](v) with ZJ E-bound in 3/m.Let .Jymbe
the *-algebra on N generated by the sets of the form {d E N : qc(a, d) 2 r, a E
N, I E *a}.
(ii) If Y,,, is not of this form let Jym= (4, N}.
(iii) For each n EMUlet J,, be the *-algebra generated by U ,,,+, J,,.
(4) The Lo,~- model associated to JY,, is the model X,!, = (N, R”, J,,, fi).

Observe that we do not define an Lo,E -model associated to ~~ because the

algebra generated by countably many internal algebras is not necessarily an
internal algebra.

Proposition 2.6. Let du and 53 be Lo,E -models. The following are equivalent.
(a) GJl=W,E 9%
(b) %, EL,,,, %k.
(c) For each n EN, %, =L;,,i9,,.

Proof. It follows easily from the definitions. 0

Proposition 2.7. Let % and 5% be L,,+-models. Then

(a) % =,,E%L.
(b) du~“,~93 if and only if %L =o,E%A.

Proof. (a) It is similar to Keisler’s proof (Section 4.2, [16]) of the completeness
148 S. Fajardo

theorem for L_,n, we just have to observe the following:

Given a term T(U, u) of the form E[O(u, V) ) v](v), the u-algebra ST generated
by the sets {d E A : 7% (a, d) 2 r, r E Q, a E A} is contained in 3. This is because the
y,,‘s form a dense set of Lw,E-terms, by Theorem 1.10 and Lemma 1.11 we have:
For each m EN and L_,n -term T we can find ?/Tfrom our list such that for all a,
d in A

17%(a, 4 - ~:(a, 41 s 117- ?/,]I< l/m.

Therefore ST s a( U mENST,,) = %.
(b) follows immediately from (a). q

Here is an interesting convergence result for I,,,,,-models.

Proposition 2.8. Let 92 be an L,+-model. The sequence (Q,‘,),,, converges almost

surely to %L. This means: For each L,,,,,-term T(U),

T%d) = Tq’(W) a.s.

Proof. By induction on the complexity of T.

(i) Trivial for atomic terms.
(ii) If 7 is F(T~, . . . , T,,) use the continuity of F, Fubini’s theorem and the
hypothesis of induction.
(iii) If T(U) is 5 0(u, V) dv use the Dominated Convergence Theorem with
bound 1]0\1,Fubini’s theorem and the hypothesis of induction.
(iv) If T(U, v) is E[B(u, V) ) v]( v ) use Hunt’s Convergence Theorem (for the
statement and proof of this theorem see [9, p. 391). We have to recall that the
%,,‘s form an increasing sequence of a-algebras on A and that ]181\ is a uniform
bound for each 7%:. As before use Fubini’s Theorem and the hypothesis of
induction Cl

fioposi&n 2.9. Let “21and W be L,,,-models. If for each n EN, %,!,-,,,3~, then
%!=@ 93.

Proof. By Proposition 2.8 for each closed L,,,-term 7 we have:

By Proposition 1.17,
'PL'_ 33,
for each nEN: 7 I~- T n. (2)

Therefore, from (1) and (2) we have T%‘= 7%: and so by Proposition 2.7(b) we
have 7% = 7% and by Proposition 1.17, % =_,nS% 0
 Probability logic with conditional expectation 149

A partial converse to this Proposition is given by

Proposition 2.10. If JH = (M, X&, 9, F) and 9 = CM, X9, % EL) are hwerfinite
models with JU =w,E9. Then for all n EN, JU,!,=~,~CY’L.

Proof. Let n EN be fixed. By Proposition 2.6(c), JII, = (M, XA, (~z)~~,,, p) and
g’, = 0% X9, (Y:)~~~, CL) are LZ,, -hyperfinite models such that A,, sL;,,,gn. By
the uniqueness theorem for Lz,,- hyperfinite models, see [16], we know the
following:
There exists an internal bijection h : M + M such that:
(i) X&(h(w), h(w’)) = Xp(w, 0’) a.s., and
(ii) for all m s n, y$(h(w)) = y:(o) a.s.
From (i) and (ii) and the form we defined the u-algebras that appear in .4; and
9: we obtain an extra property:
(iii) For all m s n, h(%z) = 925, where %z(%<) is the c-algebra in &,!,(9~).
Using properties (i)-(iii) it is easy to see that

We are now ready for the existence theorems of this section. We first prove the
existence of hyperfinite models for Lw,E. The method of proof is suggested by
Keisler’s proof of the completeness theorem for L++ The idea is to ‘translate’
problems about Lw,E -models into problems about integral models and use the
tools already developed for logic with integral quantifier.

Theorem 2.11
(Existence of Lw,E -hyperfinite models). Let % = (A, Y%, 93, v) be
an Lw,E -model and P E “N -N. Then there exists a hypefinite model JU =
(N, X4, 9, p) with lN( = P and such that 4!l=,,,+U.

Proof. From our fixed dense set of LwlE -terms we pick those terms T of the form
E[G, x) 1 xl (u) and let (d,,,N be an enumeration of them such that if T is a
subterm of 8 then T is listed before 8. For each such T,,,(u, v) let XJu, v) be a
new random variable symbol and let K” = {X, X, : m G n} and K” = UnErmK”.
We can inductively define a correspondence between L,,+,-terms and K”,ll-terms
as follows:
(a) If T is [X 1 r], then the Kz,l-translation a, of T is [X 1 r].
(b) If T is l(u = v), then mT is l(u = v).
(c) If 7 is F(T1,. . . , T,,), then a, is F(a_, . . . , uTn).
(d) If T is J f3(u, u) dv, then cT is 5 CQ(U,v) dv.
(e) If T is E[B(u, x) 1x](v), then a, is x(u, v).
This correspondence can be extended to the L+-formulas in the obvious way.
Clearly, if 4 E LwJ, then a, = 4.
For each n EN let Q, = (A, Y”, (X,“,,,,,,, v) be the K&,-model obtained from
du by taking X, % = 7%m. By the existence theorem of K&J-hyperfinite models, see
150 S. Fajardo

[16], we can find A,, = (N, X&n, (X”,->,,,, EL) hyperfinite such that:

%l,=K~,,& and jNI=P. (1)

By the lifting theorem for KzlJ-hyperfinite models we can find a *-finite JV’,,=
(N, X”-, (Xc) ,,_,, fi) such that A’,, lifts &, i.e.:

(a) R”- lifts X&m-,

(2)
(b) for all m <II, Xc lifts X$.

By the Robinson consistency theorem for the logic with integral quantifier
(Keisler [16]) we can assume the &‘s and .M,,‘s are increasing, i.e.,

for all n~lV, &+r ~Kzl,=&, and &+, r,;,,=.M,,. (3)

Therefore for each n EN, X&m= X&n+1= X and X8m = 8”-+I = X. For each n EN
let I,, be the *-algebra generated by the internal sets

{dEN:X$(c,d)%,rE*Q,m+cEN}.

Let n EN, then Q,, satisfies the translations crT of all those L,l,-sentences T such
that a, belongs to KE1, and % !=T. In particular, if y and E[8 ( x] E (T,,,),~,, then

au, l= x;XECB 1x1= X&

I I

By (l), A,, and %!,, satisfy the same translations and by (2), JV,, satisfies these
translations with = instead of =. By this we mean the following: If for two
K:,l-terms T and 8, +rJll”= 8”*, then rN~= O&m.This is an immediate consequence
of (2), since Y- = r&m= BALn = i”-. So, for example instead of (4) for Qn, in JV~ we
have : .

c RN”~~e ,x, z_?pn *N, *

y co (5)

Using or-saturation we can find HE “N -N such that .MH= (N, X, (Xc)msH, fi)
satisfies the translations uT of all the *L,,+-axioms T such that 0; E Kzl~ with =
instead of =. For this H we have:

Claim. The hypefinite model JU = (N, X, u(IH), p) is L,,,-elementarily equivalent

to ‘II.

Proof. Let Xf,= (N, 2, IHH,$). We first prove that the function that assigns to
each Lw,E-term T the function cT AX, is almost surely infinitely close to <.K~, the
*-interpretation of LwlE in .hr;I. We do this by induction on the complexity of T.
(a) If 7 iS [x 1 I] Or l(U = V), it iS easy.
(b) If T is F.(T~, . . . , T,) we show c+~=*F(+, . . . , +$”’ ) a.s.

S? = @(&, . . . , &))“H = *F(&$, . . . , Gc) = *F(+, . . . , tf ;I) a.s.

 Probability logic with conditional expectation 151

(c) If T is j 0(u, V) du we show &?(a) =CbGN &“;(a, b)(l/P) as.

(d) If T(U, v) is E[8(u, x) \x](v), we want to show that

(~?~((a,b))“~ = X$(fz, b) -E[fL%(a, .) ( I,](b) a.s.

AN~(a,b) + l?[e^“h(a, 0) 1I,](b)}

Suppose this is false, i.e., {(a, b):c+, has positive
measure. Then by Keisler’s Fubini Theorem (Section 1 of [IS]) we can find n EIV
such that the internal set

U’={b:G+(a, b)-E[9”$a, *) (I,](b)al/n or

B[f%qa, *) ( &J(b) - c+(a, b) 2 l/n>

has positive measure for a set of a’s of positive measure. Without loss of
generality we can assume:

U={b:a, ^“$a, b) - l?[i”A(a, .) ) I,](b) L l/n} has positive

(6)
measure for a set of a’s of positive measure.

Observe that UE IH since both G+(Q) *) and ~‘?[d”A(a, *) 1&I(*) are IH-
measurable.
Now we prove that for a set of a’s of positive measure
0

(j (2-Q”+, b) - e^“+, b)) db) > 0. (7)

u

This is clearly a contradiction to the translation of the conditional expectation

axiom in Xr$ (i.e. (5) does not hold in JY~).

Since U E IH we can find K, [b,lH such that:

(i) U= UssK[bJH, KE*N--N.
(ii) For each s, [b,lH is an &-atom.
(iii) If s # s’, then [bJH n[b,,], = 8.
Then

(by definition of j?‘[ ( IJ>

152 S. Fajardo

(because _%$(a, .) is constant in [b,],).

and by (6), “(Iu(/P)>O, so (l/nP) (uJ>O. This proves (7) and we obtain the
contradiction.
To complete the proofs of the claim and the theorem observe the following:
Given a closed LwlE- term 7 there is some n EN such that its translation a, is in
%,I. Then 7% = a:- and by (1) we have:

a7% = (+K
7 . (8)

The lifting property (2) implies a$-= afm-, by the definition of the *-
interpretation for XL, cV- = FH and by the definition of “JYA we have T& =
0 (;“A) = &-“. Then from (8) we can conclude that:
7 o = 7M and since this is true for each &,+,-closed term T, % E,,,,~JU as we
wanted to show. Cl

Given an L,,,,E -model

% we want to find a uniform hyperfinite model Y
elementarily equivalent to it. We first use the previous theorem to get a hyperfi-
nite model J4’ elementarily equivalent to 4% and then use Theorem 3.15 from
Rodenhausen’s thesis [25] conveniently changed to fit our setting in order to get
& uniform hyperfinite elementarily equivalent to JU’. The idea is simple but the
details are rather technical.

Theorem 2.12 (Existence of uniform hyperfinite models for Lo+). Let Lz1be an
L _+-model and P, Q E “N -NJ. Then we can find a uniform hyperjinite model &
with universe H x K such that (HI = P, (KJ = Q and % =o,E~.

Proof. Let INI E *kJ -kJ be such that INI <P and (NI/Q = 0. As in the proof of
Theorem 2.11 we can find a *-finite model X’ = (IV, ?, J, cN) such that
“JV =,,E%. (& is the *-counting measure on IV.)
Now we modify the proof of Theorem 3.15 in [25] to get:
(2) There exists X such that the uniform *-finite model x = (H X K, -%,I, &K)
satisfies “.,VEw+Ofl.
To prove this let {A, : s s B} with I3 E *N be an *-enumeration of the J-atoms.
These atoms are the equivalence classes of the internal equivalence relation -I
definedonNasfollows:Givenn,n’ENn-In’iffforallAEJ,nEAiffn’EA.
 Probability logic with conditional expectation 153

Recall that I is the *-algebra of internal subsets of HX K of the form C x K

with C E *9(H). The I-atoms are of the form {h} x K with h E K, there are P of
them and each one has internal cardinality Q.
For each s s B let a, E *[O, 11 be defined by

a, = iMAs) = lA,]lN. (3)

Let {Es : s <I?} be an internal partition of I such that for all s =sB

(4)

Therefore for each s =sP, Es = IJ i=l Ek where for each i, Et is an I-atom and
]Ei] = Q. For each s <B we can find an internal surjection h, : Es + A, such that:

(i) for all a, U’E A,, &(h;‘(a>)= L(h;‘(a’)),

(ii) for all i, i, (h;‘(a) nEi]= (h;l(a) nE& (5)
(iii) for all a, U’E A,, h;l(a) n hF’(a’) = (a.

Let’s give an idea of how to find h, : Given a E A, and i we can find aEf c Ef such
that

and if a # a’, then aEd fl a’Ed = $3.

This is possible because IA, I s (NJ and (NJ/Q = 0. Then we can define h, on aEi
as follows: If x E uE~ let h,(x) = a.
Now we define h: HxK+iV by h= lJsss h, and J~:((HxK)x(HxK)-+*R!
*
by X(w, w’) = ?(h(o), h(o’)) and claim:

for every LolE -term, 7(n), [r(u)lN(a) = [+)Y’(h(a)). (7)

We prove this statement by induction on the complexity of T. The only
non-trivial steps are those that involve integrals and conditional expectation. We
give the proof of the integral case. The other case is similar and we leave it to the
reader. If T(U) is 5 e(u, V) du we want to show

T”(a) = ,.gxK ~“(a, b) &=,FN

W’(h(a>,b’) j$,=T”‘(hb)h (8)

By induction hypothesis 6”(& b)=i”‘(h(a), h(b)) for all b EHX K. Taking the
maximum difference we can find R E “N -N such that

for all b E Hx K, IQ”(a, b) - @‘(h(a), h(b))1 C$. (9)

Then
P(a)= c
beHxK
d”(a, b)&- ,,zxK~“‘UG),h(b)) & (10)
154 S. Fajardo

so

I;“(a) - ~“‘(h(a))l= IbE&

f%,b) &- c @(h(a),6’) $-,( b’csN

l
b’eN

=
I c
b’EN
@‘(h(a), b’)+fi- c
b’eN
@‘(h(a), b’)l
tNt .

This is because for each b’E A, there are = P(Q + ]A, l)/lNl elements in H X K
mapped onto b’. The last term is

= c @-‘(/.+), b’) <(Nlsup ,$“‘M

tNt
=w@‘tAstzo
Ib’sN

sin~e!A!<!.E!
Q Q-Q
20 and 6” is bounded.

This proves (8) and completes the proof of (7).

(2) follows directly from (7) and if we put together (1) and (2) we have:
Jl = “X solE OJV=_,~% as we wanted to show. Cl

In Definition 1.14 we pointed out some possible ways of generalizing LolE in

order to allow more than one conditional operator in the language. There are
natural ways of extending the definitions and theorems of this section to cover
those generalizations. We now give indications on how to handle the case
involving two conditional expectation operators E, and E, and leave to the
interested reader the details of the other cases.

Definition 2.13 (Generalizations). Let L,+ contain the conditional expectation

symbols El and Ez.
(a) *-finire models are of the form X = (N, 2, I, J, G) with I c J *-algebras
contained in *9(N) and everything else as in Definition 2.1(a).
(b) Uniform *-finite models are *-finite models such that:
(1) N=HxKxL with IHJ, IKJ, (L.tE*N-N.
(2) I is the *-algebra of internal subsets of N of the form A X K XL with
A Ebb’.
(3) .7 is the *-algebra of internal subsets of N of the form A X B XL with
AEON and BEAM.
(c) (Unifom) Hyperfinite models are defined in the obvious way.

We can sum things up with

Theorem 2.14. The corresponding versions of Theorems 11 and 12 of this section

hold for the E,, E, case.
 Probability logic with conditional expectation 155

Proof. The only place where the generalization may not be immediate is in the
definition of the function h in the proof of Theorem 12. For this observe that if N
is a uniform *-finite model each I” splits into IKl many JN-atoms. The idea is to
define first ho : IN-atoms + IN’-atoms and then define h, as before from JN-atoms
onto J”‘-atoms in such a way that h, preserves ho. Again the details are technical.
See Rodenhausen [25] for a more complete description of this construction in a
slightly different setting.

3. Uniqueness of uniform hypeate models

Pursuing the analogy with first order logic, once we have proved the existence
of hyperfinite models (‘saturated’) we go on to prove their uniqueness. This in the
sense that if two of them with the same hyperfmite cardinality are elementarily
equivalent then they are a.s. isomorphic. In this context a.s. isomorphic means
that there exists a probabilistic a.s. isomorphism as in Definition 3.7.
In this section we are going to be dealing with certain definable sets and their
probabilities; as in the case of L,,, and Lolp, it is sometimes more convenient to
have at hand both quantifiers Jdx and (Pear). We know that there was no
problem in doing so by Theorem 5.7 in [17]. We can naturally extend this idea to
L ,_+ as the following definition and theorem indicate.

Definition 3.1. Lw,EP is the logic obtained from LolE by allowing the quantifiers
(Pn z=r), formally:
(i) The atomic formulas of Lo,EP are the expressions of the form [Tao]
where T(U) is an LU1,-term.
(ii) The set of LolEp- formulas is the least set containing the L,,,,-atomic
formulas and closed under 1, /\, (Px 2 r).
(iii) LolEP- models and interpretations are defined in the obvious way.

In this paper we are not interested in studying Lo,EP by itself. We just want to
use the following theorem in order to give a natural proof of the main theorem of
this section.

Theorem 3.2. Lo,Ep is a conservative definitional extension of LwlE, that is:

(a) For any sentence 8 in LwlE:

L o,Ep I=8 if and only if LolE t=8.

(b) For any L,,,-formula 4(u) there is an L,l,-forrnula -y(u) such that

L 0,EP Mu) * Y(U).

Proof. It follows from Keisler’s quantifier elimination theorem (Theorem 5.7 in

 156 S. Fajardo

[17]) that shows how to eliminate the probability quantifiers (fi 2 r) by means of
the integral quantifier. •!

Lemma 3.3. Let T(U, v) by an Lo,E-term. If v is E-bound in T, then

FL,,, EL+, v) I VI = T(U, v).

Proof. Intuitively the lemma says: “Once we have conditioned one variable, we
get nothing new by conditioning the same variable later, as long as it remains
E-bound.” The examples following Definition 1.5 illustrate the situation. It is an
easy consequence of the definition of conditional expectation in probability that if
v is E-bound in 7, then E[T(u, v) 1v] = ~(u, v) is true in all L,,,-models. There-
fore, by the completeness theorem for LwIE, l-L,,, E[T(u, v) 1 v] = ~(u, v). q

Lemma 3.4. Let T(ZJ, v) be an L+- term with v E-bound in T. If % = (A, X, 9, v)

is an Lo,E- model and S c[w is a Bore1 set, then {(a, b) E A x A: ~(a, b) E S} is
v(“) x g-measurable.

Proof. It is sufficient to prove the lemma for S of the form (c, m) with c EC&‘.By
Lemma 3.3, l-L,,e E[T(u, v) 1v] = T(U, v), then for r E R:

T~(u, b)ar if and only if E[T~(~, *)I 91(b)ar,

so
{(a,b)~AxA:~~(a,b)~c}={(a,b)~A~A:E[~~(~;)(~l(~)~c~

and this last set is vCn)x g-measurable by Definition 1.7(b). 0

A natural question is: What happens if in Lemma 3.4 we consider more than
one E-bound variable in 7 at a time? Let’s take a look at a simple case, say
T(U, v) is E[E[X(u, v) I u]v] both u and v are E-bound in T, is 7 9~5%
measurable? We do not know. Applying the lemma on each component of 7, we
can see that T is v x %-measurable and 9 x v-measurable but in general 9 x 9 is
strictly contained in v X 9n 9~ v. For example, if singletons do not belong to B
we can have {a}xAEvXg and Ax{a}$9Xv and {(a,a)}~~X~ The lesson
is: “E[E[X) 9]( 91 (0, .) is not necessarily equal to E[X ) 9x S](., e).”

Theorem 3.5 (Rectangle approximation for Lo,=). Let T(U, v), % and S be as in
Lemma 3.4. Then for all m EN there exists a finite union of rectangles IJ f-1 A,i x
Azix. . . x A,i x Bi such that for all i = 1, . . . , r,,,, j = 1, . . . , n, Aii is v-measurable,
Bi is .CF-measurable and

v(“+l) (T)-‘(S)n 6 Ali x . . . x Ani x Bi <+

( i=l ) m’
where

(7=)-l(S) = {(a, b) E A” x A : ~~((a, b) E S}.

 Probability logic with conditional expectation 157

Proof. By Lemma 3.4 we know that (T%)-‘(S) is v(“)x %-measurable. Then for
real 6 > 0 we can find rectangles Ai X Bi, i = 1, . . . , m,, such that

p+l)
(ill A: X Bia(T”)-‘(S)) < 6

with AI #“-measurable and Bi E 9.

For each i, by Proposition 1.7.2 in [16], there exist Ai zY’-measurable such that
v’“‘(Ai A Ai) = 0. Therefore in (1) we can replace each A: by Ai to get:

A(Tu)‘(s))<~
v(n+l)
(icl Ai X6

with Ai u”-measurable and Bi E 9.

Now, for each v”-measurable Ai and each real E > 0 we can find rectangles
A;$ . . . x A:; with A& u-measurable (k = 1,. . . , n; j = 1,. . . , rf) such that:

. . .xA$AAi CE.

We can then consider approximations to (7%)-l(S) of the form

i;l.&ix. . *XALi)XBiA(T’)-l(S))

and if we take 6 and E sufficiently small we can make sure that (4) is less than
llm. q

Theorem 3.6. Let % = (A, X, 9, v) and 93 = (A, Y, 9, u) be LW1,-models, r(u, v)

an L+ -term, S E R a Bore1 set and n EN. If “u‘_,E 8, then we can find rectangle
approximations to (7%)-l(S) and (?-l(S) as given by the previous theorem, say

V(n+l) (7%)-1(S) A ; Ali x - . . x A,i x Bi) c1

i=l m

and
,(“+l) ij Cli x . . . x C,,i x Di -&
i=l m

such that:
(1) r, = r:.
(2) For all i, j, V(Aji) = U(Cii) and v(Bi) = v(Q).

Proof. As usual it is sufficient to prove the theorem for Bore1 sets S of the form
(c, w) with c EC!. It is for this theorem that it is more convenient to work with
L o&P than with Lo+. For each c E Q, [T(u, v) - c 2 0] is an atomic formula and for
each r E [0, 11,(Pu > r)[T(u) a c] is an Lo,EP -sentence. By Theorem 3.2 we have
158 S. Fajardo

for each r~[0, l] and I,,,,-term T(U):

%~=(Pu>r)[r(u)>c] if and only if %?~(Pu>~)[T(u)~c].

Therefore ~(~+l)((r~)-l(S)) = ~(“+l)((r~))~(S)). Using j&P sentences we ensure

that the projections of (I-~>-‘(S) and (r%))l(S) have the same probabilities. For
example:
(i) ~({a E A : ~%(a, . . . , a) > c}) = ~({a : 7%(a, . . . , a) a c})
(i.e. the diagonals
have the same probability). To see this we just consider the sentences

e=(Pu>r)[T(u,..., u)ac] with r~[O,l].

Again we have % k 8 iff 9 I=8 and this implies what we want.

(ii) Consider the sentence

This type of sentence allows us to show that the i-th projections of (7%)-l(S)and
(TB)-1(S) have the same probabilities.
Using conjunctions and negations of sentences like this one we can make sure
that corresponding Boolean combinations of the projections have the same
probabilities. The theorem follows from these considerations. 0

Definition 3.7. Let Ju = (H x K, X, o(I), w) and JU’ = (Hx K, Y, a(l), p) be un-

iform hyperfinite models. .n?tand .4t’ are said to be as. isomorphic if there exists
an internal bijection h: HX K +- HX K such that:
(i) X(h(m)) = Y(w) a.s.
(ii) h(u(l)) = a(l).
We denote this by JI =_.N. Finally, the main theorem of this section:

Theorem 3.8. (Uniqueness of Uniform Hyperfinite Models). Let Ju =

(H x K, X, a(I), p) and At’ = (Hx K, Y, a(I), FL)be uniform hyperfinite models for
L w,E. Then:

Al =_,,&t’ if and only if .d =a.s,AC).

Proof. (3) Easy.

(e) Let {Sm.. m EN} be a countable basis for the Bore1 o-algebra on R. Recall
the following key properties of hyperfinite models:
(i) If D’ is p-measurable, then there exists an internal D such that p(D AD’) =
0.
(ii) If D’ E a(I), then there exists D E I such that p (D AD’) = 0.
Using these two properties we can strengthen the conclusions of Theorem 3.5 and
3.6 for the case of two elementary equivalent hyperfinite models JU and JU’.
Following the notation of those theorems with Ju and 4’ in place of % and 53
respectively, we have:
(1) We can assume the Aji ‘s and Cii‘s are internal. The R’s and Di ‘s are
 Probability logic with conditional expectation 159

internal and belong to I. And

(2) Corresponding Boolean combinations of ‘s and Ci‘s (Bji‘s and Q’s) have
Aji
the same internal cardinality.
Now using (1) and (2) we can show: For each (I, m, n) E N3 there exists an
internal bijection h;“” :HxK-+HxK such that:
(3) h;““(I)=I. And
(4) If (m)ncrm is our fixed dense set of Lw,E- terms as described in Definition 2.5
then we have: For each j s n and p s m and approximations satisfying (1) and (2)
and

(r~‘)-‘(Q n fj A’,, x . -
i=l

and
J Cj,i)
<f
w(rJ
(
(.yF’)p’(SJ n
i=l
Cii X . ’ . X

where (ri) is the number of free variables in rj, we have

h;““(A$) = Cl,, s = 1, . . . , ri.

Notice that for each term rj(u, V) with n E-bound in yj in the above approxima-
tions we have that the A$ (C$) corresponding to 2, is an element of I.
The proof of (3) and (4) is technical. By a saturation argument we can find
h : H X K + H x K internal bijection satisfying (3) and (4) for all 1, m, n EN. For
each rr EN h satisfies: y$(h(w)) = -y$‘( o ) a.s. and in particular X(h(w)) = Y(w)
as.; therefore h is an a.s. isomorphism between J% and &l”’ and this is precisely
what we wanted to show. 0

At the end of the previous section we indicated how to extend the existence
theorems for hyperfinite models to the case where we allowed more than one
conditional expectation operator. Again, all the theorems in this section can be
naturally extended to cover those cases. Details are left to the reader.

4. Applications: Robinson consistency and Craig interpolation for LolE

In first-order logic there are several proofs of the Robinson Consistency

theorem, among them a particularly simple one uses recursively saturated models
(see for example [21]). For the probability logics Lo+ I+_, and Lad Hoover in
[ll, 121, Keisler in [16], Hoover and Keisler in [15,16] have given proofs that
resemble that construction. Using the existence and uniqueness results of the
previous sections we can give a proof of Robinson consistency for I,,,+ along the
same lines. As before, the generalizations of the theorems to Lo+ with more than
one conditional expectation symbol go through in a natural way.
160 S. Fajardo

Theorem 4.1 (Robinson Consistency for L wlE). Let L’ and L2 be two languages
with L = L’ U L2 and Lo = L1 fl L2. Let Q.l be an L&-structure and 9 an L&-
structure such that Ou1 L~=w,EB ILo. Then there exists an L,,,-model C such that:

C rL~=,+% and C rL~=o,ES3.

Proof. By Theorem 2.12 we can find L& and LilE uniform hyperfinite models
A& and Jcc2 such that:

(9 Jtc, = (Hx K, . . . , 40, EL), Al,=(HxK,. . .,(+(I), p).

(ii) A&=~+% and J~X~=_,~%.

We then have &I lL~=o,E~2 ILo and using the uniqueness Theorem 3.8 we can
assume At, lL~=o,EA 2 1LO,so we can take as C the common expansion to L of JU,
and A2. Cl

We cannot give the usual first-order logic argument in order to prove the Craig
interpolation theorem for LolE from the Robinson Consistency theorem for Lw+
Keisler in [16] explains the difference between Probability Logics and extensions
of first order logic as considered by Mundici in [24]. Nevertheless, using Robinson
Consistency and a Henkin construction Hoover in [ll, 121 proved an interpola-
tion theorem for Lo,p and Hoover and Keisler, see [16], gave a proof of
Interpolation for L,. Hoover in [14] proves an almost sure interpolation theorem
L alp for which we have no analog in Lo+.

Theorem 4.2 (Craig Interpolation for Lo,&. If 4 is an L&-sentence and y is an

Lt,,-sentence such that I=4 + y, then there is an LzlE = L& fl L&E sentence e
such that k4+tl and Fe+y.

Proof. Use a Henkin argument in conjunction with Theorem 4.1 as in [16]. Cl

To conclude, we remind the reader that all the above results hold for the logics
L WE with A an admissible set and w E A.

References

[l] S. Alveberio, et al., Nonstandard Methods in Stochastic Analysis and Mathematical Physics
(Academic Press, New York, to appear).
[2] R.M. Anderson, Star finite representations of measure spaces, Trans. AMS (271) 667-687.
[3] R. Ash, Real Analysis and Probability (Academic Press, New York, 1972).
[4] J. Barwise, Admissible Sets and Structures (Springer, Berlin, 1975).
[5] J. Barwise and S. Feferman, Abstract Model Theory and Logics of Mathematical Concepts
(Springer, Berlin, to appear).
[6] K.L. Chung, A course in Probability Theory (Academic Press, New York, 1974).
 Probability logic with conditional expectation 161

[7] N. Cutland, Nonstandard measure theory and its applications, Bull. London Math. Sot. 15 (1983)
529-589.
[8] C. Dellacherie and P.A. Meyer, Probabilities and Potential, A (North-Holland, Amsterdam,
1978).
[9] C. Dellacherie and P.A. Meyer, Probabilities and Potential, B (North-Holland, Amsterdam,
1982).
[lo] S. Fajardo, Probability Logic with Conditional Expectation, Abstracts AMS.
[ll] D. Hoover, Model Theory of Probability Logic. Ph.D. Thesis, University of Wisconsin, 1978.
[12] D. Hoover, Probability logic, Ann. Math. Logic 14 (1978) 287-313.
[13] D. Hoover, A normal form theorem for L,,r, with applications, J. Symbolic Logic 47 (1982)
605624.
[14] D. Hoover, A probabilistic interpolation theorem, to appear.
[15] D. Hoover and H.J. Keisler, Adapted probability distributions, to appear.
[16] H.J. Keisler, Probability Quantifiers, to appear in: J. Barwise and S. Feferman, eds., Abstract
Model Theory and Logics of Mathematical Concepts (Springer, Berlin).
[17] H.J. Keisler, Hyperfinite model theory, in: R.O. Gandy and J.M.E. Hyland, eds., Logic
Colloquium 76 (North-Holland, Amsterdam, 1977) 5-110.
[ 181 H.J. Keisler, An infinitesimal approach to stochastic analysis, Mem. AMS 48 (1984) 297.
[19] H.J. Keisler, Hyperfinite Probability Theory and Probability Logic, Lecture Notes, Univ. of
Wisconsin (unpublished), 1979.
[20] H.J. Keisler, A completeness proof for adapted probability logic, to appear.
[21] H.J. Keisler, Fundamentals of Model Theory, in: J. Barwise, ed., Handbook of Mathematical
Logic (North-Holland, Amsterdam, 1978) 47-103.
[22] P.A. Loeb, An introduction to nonstandard analysis and hyperfinite probability theory, in: A.T.
Bharucha-Reid, ed., Probabilistic and Related Topics 2 (Academic Press, New York, 1979)
105-142.
[23] P.A. Loeb, Conversion from non-standard to standard measure spaces and applications in
probability theory, Trans. AMS 211 (1975) 113-122.
[24] D. Mundici, Chapter VI, to appear in: J. Barwise and S. Fefferman, eds., Abstract Model Theory
and Logics of Mathematical Concepts (Springer, Berlin).
[25] H. Rodenhausen, The Completeness Theorem for Adapted Probability Logic, Ph.D. thesis, Univ.
of Heidelberg, 1982.
[26] K. Stroyan and J. Bayod, Foundations of Infinitesimal Stochastic Analysis (North-Holland,
Amsterdam, to appear).
[27] K. Stroyan and W.A.J. Luxemburg, Introduction to the Theory of Infinitesimals (Academic
Press, New York, 1976).
[28] M. Ziegler, Topological Model Theory, to appear in: J. Barwise and S. Feferman, eds., Abstract
Model Theory and Logics of Mathematical Concepts (Springer, Berlin).