1 DEFINITION OF CONDITIONAL EXPECTATION 4

1 Denition of Conditional Expectation

x
P(A, X = x) =

x
P(A|X = x)P(X = x)
and the resulting formula for conditional expectation
E(Y |X = x) =
_

Y ()P(dw|X = x)
=
_
X=x
Y ()P(dw)
P(X = x)
=
E(Y 1
(X=x)
)
P(X = x)
We would like to extend this to handle more general situations where densities
dont exist or we want to condition on very complicated sets.
Denition 1 Given a random variable Y with E|Y | < dened on a prob-
ability space (, A, P) and some sub--eld G A we will dene the con-
ditional expectation as the almost surely unique random variable E(Y |G)
which satises the following two conditions
1. E(Y |G) is G-measurable
2. E(Y Z) = E(E(Y |G)Z) for all Z which are bounded and G-measurable
Remark: one could replace 2. in the previous denition with:
G G, E(Y 1
G
) = E(E(Y |G)1
G
).
1 DEFINITION OF CONDITIONAL EXPECTATION 5
Proof of existence and unicity
Existence Using linearity, we need only consider X 0. Dene a
measure Q on F by Q(A) = E[X1
A
] for A F. This is trivially
absolutely continuous with respect to P
|F
, the restriction of P to F.
Let E[X|F] be the Radon-Nikodym derivative of Q with respect to
P
|F
. The Radon-Nikodym derivative is F-measurable by construction
and so provides the desired random variable.
Unicity: If Y
1
, Y
2
are two F-measurable random variables with E[Y
1
1
A
] =
E[Y
2
1
A
] for all A F, then Y
1
= Y
2
, a.s., or conditional expectation is
unique up to a.s. equivalence.
For G = (X) when X is a discrete variable, the space is simply partitioned
into disjoint sets = G
n
. Our denition for the discrete case gives
E(Y |(X)) = E(Y |X)
=

n
E(Y 1
X=xn
)
P(X = x
n
)
1
X=xn
=

n
E(Y 1
Gn
)
P(G
n
)
1
Gn
which is clearly G-measurable. In general for G = (X):
Denition 2 Conditional expectation of Y given X
Let (, A, P) be a probability space, Y L
1
(, A, P) and X another
random variable dened on (, A, P). Dene then E(Y | X) the conditional
expectation of Y given X as E(Y | (X)).
Proposition 3 Let (, A) be a measurable space,
Y L
1
(, A, P)
and X another real-valued random variable dened on (, A, P). As
X = f(Y ), where f is measurable, real-valued function if and only if (X)
(Y ), we get that E(Y | X) is a measurable function of X.
Proposition 4 Let (, A, P) be a probability space, and X and Y two in-
dependent random variables such that Y is P-integrable. Then E(Y | X) =
E(Y ), P-almost surely.
Do not mix this notion with the following:
2 PROPERTIES OF CONDITIONAL EXPECTATION 6
1.2 Couples of random variables with p.d.f.
Proposition 5 Let (X, Y ) be a couple of real-valued random variables with
p.d.f. f
X,Y
(x, y) w.r.t. the Lebesgue measure on R
2
. Denote the respective
marginal p.d.f. of X and Y as f
X
(x) and f
Y
(y). Consider f
X|Y
(x | y) =
f
X,Y
(x,y)
f
Y
(y)
. Then almost surely
C B, P(X C | Y = y) =
_
C
f
X|Y
(x | y)dx.
If besides X is P-integrable, then
E(X | Y = y) =
_
R
xf
X|Y
(x | y)dx.
If g : R
2
R is a measurable function such that g(X, Y ) is integrable, then
E(g(X, Y ) | Y = y) =
_
R
g(x, y)f
X|Y
(x | y)dx.
Remarks: As soon as f
Y
(y) > 0, this denes the distribution of X given
that Y = y, described by p.d.f f
X|Y
(x | y), which is nonnegative and of
integral 1.
If X and Y are independent, f
X|Y
= f
X
and f
Y |X
= f
Y
. To make the link
with E[X|Y ] would require to introduce the concept of regular conditional
distribution.
Equation (5) may be useful to compute the mathematical expectation of
g(X, Y ) as
E(g(X, Y )) =
_
R
__
R
g(x, y)f
X|Y
(x | y)dx
_
f
Y
(y)dy.
2 Properties of Conditional Expectation
2.1 Conditional expectation
E(|G) may be seen as an operator on random variables that transforms A-
measurable variables into G-measurable ones.
Let us recall the basic properties of conditional expectation:
2 PROPERTIES OF CONDITIONAL EXPECTATION 7
1. E(|G) is positive:
Y 0 E(Y |G) 0)
2. E(|G) is linear:
E(aX +bY |G) = aE(X|G) +bE(Y |G)
3. E(|G) is a projection:
E(E(X|G)|G) = E(X|G)
4. More generally, the tower property. If H G then
E(E(X|G)|H) = E(X|H) = E(E(X|H) | G)
Proof: The right equality holds because E[X|H] is H- measurable,
hence G-measurable. To show the left equality, let A H. Then since
A is also in G,
E[E[E[X|G]|H]1
A
] = E[E[X|G]1
A
] = E[X1
A
] = E[E[X|H]1
A
].
Since both sides are H- measurable, the equality follows.
5. E(|G) commutes with multiplication by G-measurable variables:
E(XY |G) = E(X|G)Y for E|XY | < and Y Gmeasurable
Proof: If A G, then for any B G,
E[1
A
E[X|G]1
B
] = E[E[X|G]1
AB
] = E[X1
AB
] = E[(1
A
X)1
B
].
Since 1
A
E[X|G] is G-measurable, this shows that the required equality
holds when Y = 1
A
and A G. Using linearity and taking limits shows
that the equality holds whenever Y is G-measurable and X and XY are
integrable.
2 PROPERTIES OF CONDITIONAL EXPECTATION 8
6. E(|G) respects monotone convergence:
0 X
n
X = E(X
n
|G) E(X|G)
7. If is convex (in particular if (x) = x
2
) and E|(X)| < then a
conditional form of Jensens inequality holds:
(E(X|G) E((X)|G)
8. E(|G) is a continuous contraction of L
p
for p 1:
E(X|G)
p
X
p
and
X
n
L
2
X implies E(X
n
|G)
L
2
E(X|G)
9. Repeated Conditioning. For G
0
G
1
. . ., G

= (G
i
), and X L
p
with p 1 then
E(X|G
n
)
a.s.
E(X|G

)
E(X|G
n
)
L
p
E(X|G

)
10. Best approximation property:
Suppose that the random variable X is square-integrable, but not mea-
surable with respect to G. That is, the information in G does not
completely determine the values of X. The conditional expectation,
Y = E[X | G], has the property that it is the best approximation to
X among functions measurable with respect to Y , in the least squares
sense. That is, if

Y is G-measurable, then
E
_
(

Y X)
2
_
E
_
(Y X)
2

.
It thus realizes the orthogonal projection of X onto a convex closed
subset of a Hilbert space. This predicts the variance decomposition
theorem that we shall see in a further section.
2 PROPERTIES OF CONDITIONAL EXPECTATION 9
2.2 Conditional variance
Denition 6 Let X be a square-integrable, real-valued random variable de-
ned on a probability space (, A, P), and let F be a sub--algebra of A.
Dene the conditional variance of X given F (denoted by Var(X | F))
as the random variable E((X E(X | F))
2
| F).
Dene also the conditional variance of X given a real-valued random vari-
able Y dened on (, A, P) (denoted by Var(X | Y )) as the random variable
E((X E(X | Y ))
2
| Y ).
Proposition 7 Var(X | F) and Var(X | Y ) are well- dened, almost surely
nonnegative and nite.
Var(X | F) = E(X
2
| F) E(X | F)
2
,
and
Var(X | Y ) = E(X
2
| Y ) E(X | Y )
2
.
Proposition 8 Variance decomposition formula
Let (X, Y ) be a couple of random variables dened on a probability space
(, A, P), such that X is square-integrable. Then
Var(X) = E(Var(X | Y )) + Var(E(X | Y )).
This may be very useful in non-life insurance to nd the variance of a com-
pound distribution.
Proof:
Var(X | Y ) = E(X
2
| Y ) (E(X | Y ))
2
.
E[Var(X | Y )] = E[E(X
2
| Y )] E[(E(X | Y ))
2
].
E[E(X
2
| Y )] = E[X
2
].
E[Var(X | Y )] = E[X
2
] E[(E(X | Y ))
2
].
Var(E(X | Y )) = E[(E(X | Y ))
2
] (E[E(X | Y )])
2
.
E[E(X | Y )] = E[X].
Hence Var(E(X | Y )) = E[(E(X | Y ))
2
] (E[X])
2
.