Prof.

Vincent Tassion HS 2024

PROBABILITY THEORY (D-MATH)

EXERCISE SHEET 10 – SOLUTION

Exercise 1. [R] Let A be an compact set in R2 and let (X, Y ) ∼ Unif(A). Compute
E X 2 |Y

in the following cases:

(1) A = [−1, 1]2 ,
(2) A = {(x, y) : |x| + |y| ≤ 1}.
Solution.
(1) In this case, X and Y are independent Unif[−1, 1] random variables. So
Z 1
2 2
E(X |Y ) = E(X ) = x2 /2 dx = 1/3.
−1
(2) In this case, (X, Y ) has density given by
1|x|+|y|≤1
f(X,Y ) (x, y) = .
2
Therefore, the marginal density of y is given by
Z 1−|y|
fY (y) = 1|y|≤1 1/2 dy = (1 − |y|)1|y|≤1 .
|y|−1

We compute the conditional expectation as in section 5 of chapter 10. Let y ∈ [−1, 1].
We get
Z 1 2
x 1|x|+|y|≤1
ϕ(y) = dx
−1 2(1 − |y|)
Z 1−|y|
1
= x2 dx
2(1 − |y|) |y|−1
(1 − |y|)2
= .
3
So
(1 − |Y |)2
E(X 2 |Y ) = a.s.
3

1
Exercise 2. Let X, Y be independent random variables and let ψ : R2 → R≥0 be a
measurable function such that


E |ψ(X, Y )| < ∞.
Define ϕ : R → [0, ∞] by

ϕ(y) = E ψ(X, y) .
Show that
E(ψ(X, Y )|Y ) = ϕ(Y ) a.s.
R
Solution. First, since ϕ(y) = R ψ(x, y)µX (dx). ϕ is a measurable function (this is part of
Fubini’s theorem for positive functions). Hence, h(Y ) is σ(Y )-measurable. Therefore, it
suffices to check that for every non-negative random variable Z which is σ(Y )-measurable,
we have


E ψ(X, Y )Z = E ϕ(Y )Z .
To this end, using Fubini’s theorem we get
Z


E ψ(X, Y )Z = ψ(x, y)zµ(X,Y,Z) (dxdydz)
R×R×R≥0
Z
(because X and (Y, Z) are independent) = ψ(x, y)zµX (dx)µ(Y,Z) (dydz)
R×R×R≥0
Z Z 
= z ψ(x, y)µX (dx) µ(Y,Z) (dydz)
R×R≥0 R
Z
= zϕ(y)µ(Y,Z) (dydz)
R×R≥0

= E(ϕ(Y )Z).

2
Exercise 3. [R] Let (Yn )n≥1 be iid random variables which are uniform in {−1, +1} and
let X be a random variable in L2 . Let [n] denote {1, . . . , n} and for a subset S ⊂ [n],
define Y
YS = Yi ,
i∈S
where Y∅ defined to be 1.
(1) Show that E(X|Y1 ) = E(X) + E(XY1 )Y1 .
(2) More generally, for all n ≥ 1 show that
X
E(X|Y1 , . . . , Yn ) = E(XYS )YS .
S⊂[n]

Solution. We use the notation Y = (Y1 , . . . , Yn ). We prove the general formula using the
L2 projection interpretation of conditional expectation, which is applicable because we
assume that X ∈ L2 . Let
F = σ(Y1 , . . . , Yn )
and let HF be the vector space of F-measurable random variables in L2 . We first establish
that HF has dimension at most 2n . Consider the random variables
{y ∈ {−1, 1}n : 1Y =y }.
Being a functions of Y and bounded, these random variables are in HG . We show that
they span HF . Let Z be a F-measurable random variable in L2 . Then we know that
there exists a measurable function f : Rn → R such that
Z = f (Y1 , . . . , Yn ).
But now X
f (Y1 , . . . , Yn ) = f (y)1Y =y .
y∈{−1,1}n
n
This shows that the dimension of HF is at most 2 . Next, we claim that
{S ⊂ [n] : YS }
is an orthonormal basis of HF . First, since YS ∈ {−1, 1} we have E(YS2 ) = 1. Next let
R, S be distinct subsets of [n]. Then
 Y  Y
E(YR YS ) = E(YR∆S ) = E Yk = E(Yk ) = 0,
k∈R∆S k∈R∆S

where we used independence in the second last inequality, and the facts that R∆S ̸= ∅
and E(Yk ) = 0 for all k in the last. This shows orthonormality. Since we have 2n
such functions, they form a basis. So finally, the formula we wanted to show is just the
projection formula with respect to this orthonormal basis.

3
Exercise 4. [R] Let X be a real-valued random variable defined on (Ω, F, P) that takes
values in [0, ∞] a.s. Let G ⊂ F be a sigma-algebra. Define E(X|G) and show that it is
unique (up to almost sure equivalence).
Solution. Refer to section 7 of chapter 10.