Professional Documents
Culture Documents
Z −|w|2 /2
α·w−|α|2 /2 e
= e dw
A+α (2π)d/2
2
e−|w−α| /2
Z
= d/2
dw
A+α (2π)
Z −|u|2 /2
e
= d/2
du
A (2π)
= P(W ∈ A)
where the second to last line follows from the change of variables u = w − α.
Alternatively, since moment generating functions, when finite on an open set, determine
the distribution of a random variable, we could also argue as follows:
2 /2+θ·(W −α)
EQ [eθ·Ŵ ] = EP [eα·W −|α| ]
−θ·α−|α|2 /2
=e EP [e(α+θ)·W ]
2 /2+|α+θ|2 /2
= e−θ·α−|α|
2 /2
= e|θ|
In particular, an arbitrage can be identified with a vector π ∈ Rm with (AT π)j ≥ 0 for all
j = 1, . . . , n such that AT π 6= 0.
4
An equivalent martingale measure (with respect to the cash numéraire) is a probability
measure Q on {1, . . . , n} such that Q{j} > 0 for all j and that EQ (Y S1 ) = 0. Letting
qj = Q{j} we have
n
X
EQ S1i = aij qj = (Aq)i
j=1
n
Pn martingale can be identified with a q ∈ R with qj > 0 for all
In this context an equivalent
j such that Aq = 0 and j=1 qj = 1.
The first fundamental theorem of asset pricing is that there exists an equivalent martingale
measure if and only if no arbitrage exists.
We have now seen one proof of the 1FTAP in lecture. For the sake of developing intuition,
here’s yet another one: As before, the easy direction is to show that the existence of an
equivalent martingale measure implies no arbitrage: Suppose that there exists an x ∈ Rn
with xi > 0 for all i = 1, . . . , n such that Ax = 0. If there exists a y ∈ Rm with (AT y)i ≥ 0
for all i = 1, . . . , n then x · (AT y) = (Ax) · y = 0. Hence AT y = 0.
Now we prove the harder direction using a version of the separating hyperplane theorem.
Suppose that there is no arbitrage: If there exists a y ∈ Rm with (AT y)i ≥ 0 for all
i = 1, . . . , n then AT y = 0. Let
S = {AT y : y ∈ Rm } ⊆ Rn
and let ( )
n
X
n
C= u ∈ R : ui ≥ 0 for all i = 1, . . . , n and ui = 1 ⊂ Rn .
i=1
By assumption, the subspace S and the convex compact set C are disjoint. Indeed, if v ∈ S,
= AT y for some y. If vi ≥ 0 for all i, then v = 0 by the no-arbitrage assumption.
then v P
Hence i vi = 0 6= 1 and v is not in C.
The situation is illustrated by the figure. A version of the separating hyperplane theorem,
says there exists a vector λ ∈ Rn such that
λ·v =0 for all v ∈ S
λ·u>0 for all u ∈ C.
(Try proving this!) We will be done once we show that λj > 0 for all j. So fix a j ∈ {1, . . . , n}
and let e ∈ Rn be given by ej = 1 and ei = 0 for all i 6= j. Then e ∈ C and λ · e = λj > 0 as
desired.
Problem 7. Consider a discrete time model with n asset with prices (Pt )t≥0 and dividends
(δt )t≥1 . Explain why an appropriate self-financing condition is
Ht · (Pt + δt ) = Ht+1 · Pt + ct .
Let Z be a positive process such that the process
t
!
X
Zt Pt + Zs δs
s=1 t≥0
Note that B is predictable since rt is Ft−1 -measurable. Let Hti = 0 if i 6= t and Htt = Bt , i.e.
Ht = (0, . . . , 0, Bt , 0, . . . , 0).
(In the notation above, we are pretending that the matured bonds continue to exist. But
since we are not trading in them, their prices are irrelevant.) Then H is predictable and
self-financing since
Ht+1 · Pt = Bt+1 Ptt+1
= Bt
= Ht · Pt .
and its corresponding wealth process is B, the money market account.
Problem 11. Let X and Y be martingales (with respect to the same filtration). Show that
if XT = YT almost surely for some non-random T > 0, then Xt = Yt almost surely for all
0 ≤ t ≤ T.
Solution 11. If XT = YT almost surely, then
Xt = E(XT |Ft ) = E(YT |Ft ) = Yt
for all 0 ≤ t ≤ T .
Problem 12. (Bayes’s formula) Let P and Q be equivalent probability measures defined on
(Ω, F) with density Z = dQ
dP
. Let G ⊆ F be a sigma-field. Prove the identity:
EP (ZX|G)
EQ (X|G) =
EP (Z|G)
for each bounded random variable X.
P
Solution 12. Let Y = EEP(ZX|G)
(Z|G)
. Note that Y is G-measurable. Hence, we need only verify
the equation E (Y 1G ) = E (X 1G ) for all G ∈ G; equivalently, we must verify EP (ZY 1G ) =
Q Q
8
EP (ZX 1G ) for all G ∈ G.
EP (1G Y Z) = EP [E(1G Y Z|G)] tower property
= EP [1G Y E(Z|G)] taking out what’s known
= EP [1G EP (XZ|G)]
= EP [EP (1G XZ|G)] pulling in what’s known
= EP (1G XZ) tower property
Problem 13. Consider a single period market with two assets. The first asset is a riskless
bond with prices B0 = 1 and B1 = 1 + r for a constant r. The second asset is a stock with
prices (St )t∈{0,1} .
Let (φ∗ , π ∗ ) be the optimal solution to the problem
maximise E U (φB1 + πS1 ) subject to φB0 + πS0 = x
for a given concave increasing utility function U . Prove that the investor is holds a non-
negative number of shares of the stock if
E S1 > (1 + r)S0
Does this agree with your intuition?
Solution 13. We can solve for φ and rephrase the problem as
maximise E V (πξ)
where
V (y) = U [(1 + r)x + y]
and ξ = S1 − (1 + r)S0 .
Let π ∗ be the choice of π ∈ R that maximizes E V (πξ). On the one hand, we have
V (0) ≤ E V (π ∗ ξ)
by the optimality of π = π ∗ as compared to π = 0.
On the other hand, we have
E V (π ∗ ξ) ≤ V (π ∗ E ξ)
by Jensen’s inequality. Since U , and hence V , is increasing, we have
0 ≤ π ∗ Eξ.
If E ξ > 0 then π ∗ ≥ 0 as desired.
Another way of looking at this problem is to consider the function π 7→ E[V (πξ)]. Since
this function is maximized at π ∗ , the gradient E[V 0 (πξ)ξ] is non-negative if and only if
π ≤ π ∗ . Letting π = 0 above and noting that V 0 (0) = U 0 [(1 + r)x] is positive and non-
random completes the argument.
Problem 14. (Tower property of conditional expectation) Let X and Y be identically
distributed random variables taking values in the set {2n : n ≥ 0} such that X/Y ∈ {1/2, 2}
almost surely and
1
P(X = 2n , Y = 2n+1 ) = 2−n = P(X = 2n+1 , Y = 2n ) for n ≥ 0.
4
9
1
(a) Show that P(X = 1) = 4
and
3
P(X = 2n ) = 2−n for n ≥ 1.
4
(b) Show that P(Y = 2|X = 1) = 1 and
1
P(Y = 2n+1 |X = 2n ) = = 1 − P(Y = 2n−1 |X = 2n ) for n ≥ 1.
3
(c) Let Z = Y − X. Show that E(Z|X = 1) = 1 and
E(Z|X = 2n ) = 0 for n ≥ 1.
1
E(Z) = E[E(Z|X)] = > 0.
4
However, by symmetry we also have E(Z|Y ) = −1{Y =1} and
1
E(Z) = E[E(Z|Y )] = − < 0.
4
What has gone wrong?!
P(X = 2n , Y = 2n+1 )
P(Y = 2n+1 |X = 2n ) =
P(X = 2n )
1
=
1 + 21{n≥1}
by part (a).
(c) Using the formula from part (b) yields
2In fact, it is possible to define a generalised notion of conditional expectation so that E(Z|X) = 1{X=1} .
But this definition is fickle, and in particular, it would not obey the tower propery of conditional expectation.
11