Stochastic Target Problems with Controlled Loss

by B. Bouchard, R. Elie, and N. Touzi

Chris Hammond

March 7, 2009
Motivation: Quantile Hedging

Suppose we have d risky assets X and wealth process Y . We must

pay out a contingent claim g (X ) at time T . We fix p ∈ [0, 1]. We
would like to find the minimum initial wealth y such that we can
find an investment strategy that will allow us to cover our
contingent claim at time T with probability p.
Motivation: Quantile Hedging

Suppose we have d risky assets X and wealth process Y . We must

pay out a contingent claim g (X ) at time T . We fix p ∈ [0, 1]. We
would like to find the minimum initial wealth y such that we can
find an investment strategy that will allow us to cover our
contingent claim at time T with probability p.
More precisely Xx,t ν is an Rd -valued process with X ν (t) = x and ν
x,t
ν
tells us how much is invested in asset i. Also, Yt,x,y is our wealth
ν
process with Yt,x,y (t) = y . Given p ∈ [0, 1], we wish to find
ν
V̄ (t, x, p) = inf{y ≥ 0 : P[Yt,x,y ν (T ))] ≥
(T ) ≥ g (Xt,x
p for some admissible control ν}.
Motivation: Quantile Hedging

Suppose we have d risky assets X and wealth process Y . We must

pay out a contingent claim g (X ) at time T . We fix p ∈ [0, 1]. We
would like to find the minimum initial wealth y such that we can
find an investment strategy that will allow us to cover our
contingent claim at time T with probability p.
More precisely Xx,t ν is an Rd -valued process with X ν (t) = x and ν
x,t
ν
tells us how much is invested in asset i. Also, Yt,x,y is our wealth
ν
process with Yt,x,y (t) = y . Given p ∈ [0, 1], we wish to find
ν
V̄ (t, x, p) = inf{y ≥ 0 : P[Yt,x,y ν (T ))] ≥
(T ) ≥ g (Xt,x
p for some admissible control ν}.
For p = 1, this problem was solved by Soner and Touzi when the
controls are bounded.
Outline

I ν
Characterize V̄ (t, x, p) = inf{y ≥ 0 : P[Yt,x,y (T ) ≥
ν
g (Xt,x (T ))] ≥ p for some control ν}
Outline

I ν
Characterize V̄ (t, x, p) = inf{y ≥ 0 : P[Yt,x,y (T ) ≥
ν
g (Xt,x (T ))] ≥ p for some control ν}
I Reduce this to the case where p = 1 but with unbounded
controls ν
Outline

I ν
Characterize V̄ (t, x, p) = inf{y ≥ 0 : P[Yt,x,y (T ) ≥
ν
g (Xt,x (T ))] ≥ p for some control ν}
I Reduce this to the case where p = 1 but with unbounded
controls ν
I Extend existing characterization of V (t, x) = inf{y ≥ 0 :
ν (T ), Y ν
G (Xt,x t,x,y (T )) ≥ 0 for some control ν} to the case of
unbounded controls
Outline

I ν
Characterize V̄ (t, x, p) = inf{y ≥ 0 : P[Yt,x,y (T ) ≥
ν
g (Xt,x (T ))] ≥ p for some control ν}
I Reduce this to the case where p = 1 but with unbounded
controls ν
I Extend existing characterization of V (t, x) = inf{y ≥ 0 :
ν (T ), Y ν
G (Xt,x t,x,y (T )) ≥ 0 for some control ν} to the case of
unbounded controls
I Apply this to the quantile hedging problem
Problem Formulation and MANY definitions
I Let T > 0 a finite time horizon, W = {Wt : 0 ≤ t ≤ T } a
d-dimensional Brownian motion on (Ω, F, P) with F = {Ft }
the P-augmentation of the filtration generated by W
Problem Formulation and MANY definitions
I Let T > 0 a finite time horizon, W = {Wt : 0 ≤ t ≤ T } a
d-dimensional Brownian motion on (Ω, F, P) with F = {Ft }
the P-augmentation of the filtration generated by W
I Let U0 be a subset of the F-progressively measurable
processes ν which are in L2 ([0, T ]) P-almost surely and taking
values in a closed (possibly unbounded) subset U of Rd
Problem Formulation and MANY definitions
I Let T > 0 a finite time horizon, W = {Wt : 0 ≤ t ≤ T } a
d-dimensional Brownian motion on (Ω, F, P) with F = {Ft }
the P-augmentation of the filtration generated by W
I Let U0 be a subset of the F-progressively measurable
processes ν which are in L2 ([0, T ]) P-almost surely and taking
values in a closed (possibly unbounded) subset U of Rd
I For t ∈ [0, T ], z = (x, y ) ∈ Rd × R and ν ∈ U0 define
ν = (X ν , Y ν ) as the Rd × R-valued solution of the SDE:
Zt,z t,x t,z

dX (r ) = µ(X (r ), νr )dr + σ(X (r ), νr )dW (r ) (1)

dY (r ) = µY (Z (r ), νr )dr + σY (Z (r ), νr ) (2)
for t ≤ r ≤ T with Z (t) = (X (t), Y (t)) = (x, y ), where
(µY , σY ) : Rd × R × U → R × Rd and
(µ, σ) : Rd × U → Rd × Md are locally Lipschitz and satisfy
|µY (x, y , u)|+|µ(x, u)|+|σY (x, y , u)|+|σ(x, u)| ≤ K (x, y )(1+|u|)
where K is locally bounded
Problem Formulation and MANY definitions (continued)

I Let U denote the subset of U0 allowing a strong solution of

(1) and (2) for all initial data
Problem Formulation and MANY definitions (continued)

I Let U denote the subset of U0 allowing a strong solution of

(1) and (2) for all initial data
I Let X be the interior of the support of X
Problem Formulation and MANY definitions (continued)

I Let U denote the subset of U0 allowing a strong solution of

(1) and (2) for all initial data
I Let X be the interior of the support of X
I Given u ∈ U, let Lu denote the Dynkin operator, Lu φ(t, x) =
∂t φ(t, x) + µ(x, u)Dφ(t, x) + 12 Tr(σσ T (x, u)D 2 φ(t, x))
Problem Formulation and MANY definitions (continued)

I Let U denote the subset of U0 allowing a strong solution of

(1) and (2) for all initial data
I Let X be the interior of the support of X
I Given u ∈ U, let Lu denote the Dynkin operator, Lu φ(t, x) =
∂t φ(t, x) + µ(x, u)Dφ(t, x) + 12 Tr(σσ T (x, u)D 2 φ(t, x))
I N u (x, y , q) = σY (x, y , u) − σ T (x, u)q and
N0 (x, y , q) = {u ∈ U : N u (x, y , q) = 0}
Problem Formulation and MANY definitions (continued)

I Let U denote the subset of U0 allowing a strong solution of

(1) and (2) for all initial data
I Let X be the interior of the support of X
I Given u ∈ U, let Lu denote the Dynkin operator, Lu φ(t, x) =
∂t φ(t, x) + µ(x, u)Dφ(t, x) + 12 Tr(σσ T (x, u)D 2 φ(t, x))
I N u (x, y , q) = σY (x, y , u) − σ T (x, u)q and
N0 (x, y , q) = {u ∈ U : N u (x, y , q) = 0}
I Let G : Rd+1 → R be measurable with for all x ∈ Rd the
function y 7→ G (x, y ) is non-decreasing and right continuous
Problem Formulation and MANY definitions (continued)

I Let U denote the subset of U0 allowing a strong solution of

(1) and (2) for all initial data
I Let X be the interior of the support of X
I Given u ∈ U, let Lu denote the Dynkin operator, Lu φ(t, x) =
∂t φ(t, x) + µ(x, u)Dφ(t, x) + 12 Tr(σσ T (x, u)D 2 φ(t, x))
I N u (x, y , q) = σY (x, y , u) − σ T (x, u)q and
N0 (x, y , q) = {u ∈ U : N u (x, y , q) = 0}
I Let G : Rd+1 → R be measurable with for all x ∈ Rd the
function y 7→ G (x, y ) is non-decreasing and right continuous
I Stochastic Target Problem is V (t, x) = inf{y ∈ R :
ν (T ), Y ν
G (Xt,x t,x,y (T )) ≥ 0 for some ν ∈ U}, equivalently
ν
V (t, x) = inf{y ∈ R : Yt,x,y ν (T )) for some ν ∈
(T ) ≥ g (Xt,x
U} where g (x) = inf{y : G (x, y ) ≥ 0}
What is known

Soner and Touzi

When U is bounded, Soner and Touzi proved that V is a
discontinuous viscosity solution of
sup{µY (x, v (t, x), u) − Lu v (t, x) : u ∈ N0 (x, v (t, x), Dv (t, x))} =
0
What is known

Soner and Touzi

When U is bounded, Soner and Touzi proved that V is a
discontinuous viscosity solution of
sup{µY (x, v (t, x), u) − Lu v (t, x) : u ∈ N0 (x, v (t, x), Dv (t, x))} =
0

What is a viscosity solution?

Let D(x, u(x), Du(x), D 2 u(x)) be a second-order elliptic
differential operator, i.e., if A − B is positive-definite, then
D(x, y , v , A) ≤ D(x, y , v , B). Then u is a viscosity supersolution
of D(x, u(x), Du(x), D 2 u(x)) ≥ 0 if for any x0 and for any smooth
φ for which x0 is a min of φ − u, D(x0 , u(x0 ), Dφ(x0 ), D 2 φ(x)) ≥ 0
Dynamic Programming Equation-Auxiliary Definitions and
Assumptions
Since U is not necessarily bounded, we introduce relaxed
semilimits.
I Let Θ = (x, y , q, A) ∈ Rd × R × Rd × Sd and let
N (x, y , q) = {u ∈ U : |N u (x, y , q)| ≤ }
Dynamic Programming Equation-Auxiliary Definitions and
Assumptions
Since U is not necessarily bounded, we introduce relaxed
semilimits.
I Let Θ = (x, y , q, A) ∈ Rd × R × Rd × Sd and let
N (x, y , q) = {u ∈ U : |N u (x, y , q)| ≤ }
I For  > 0 let
F (Θ) = sup{µY (x, y , u) − µ(y , u)q − 21 Tr(σσ T (x, u)A) : u ∈
0
N (x, y , q)} Then F ∗ (Θ) = lim sup&0,Θ0 →Θ F (Θ ) and
0
F∗ (Θ) = lim inf&0,Θ0 →Θ F (Θ )
Assumption 2.1(needed for subsolution property) Let
B ⊂ X × R × Rd such that N0 6= ∅ on B. Then for each
 > 0, (x0 , y0 , q0 ) ∈ intB and u0 ∈ N0 (x0 , y0 , q0 ), there exists an
0
open neighborhood B of (x0 , y0 , q0 ) and a locally Lipschitz map ν̂
0 0
on B such that |ν̂(x0 , y0 , q0 ) − u0 | <  and ν̂ ∈ N0 (x, y , q) on B
Dynamic Programming Equation-Auxiliary Definitions and
Assumptions
Since U is not necessarily bounded, we introduce relaxed
semilimits.
I Let Θ = (x, y , q, A) ∈ Rd × R × Rd × Sd and let
N (x, y , q) = {u ∈ U : |N u (x, y , q)| ≤ }
I For  > 0 let
F (Θ) = sup{µY (x, y , u) − µ(y , u)q − 21 Tr(σσ T (x, u)A) : u ∈
0
N (x, y , q)} Then F ∗ (Θ) = lim sup&0,Θ0 →Θ F (Θ ) and
0
F∗ (Θ) = lim inf&0,Θ0 →Θ F (Θ )
Assumption 2.1(needed for subsolution property) Let
B ⊂ X × R × Rd such that N0 6= ∅ on B. Then for each
 > 0, (x0 , y0 , q0 ) ∈ intB and u0 ∈ N0 (x0 , y0 , q0 ), there exists an
0
open neighborhood B of (x0 , y0 , q0 ) and a locally Lipschitz map ν̂
0 0
on B such that |ν̂(x0 , y0 , q0 ) − u0 | <  and ν̂ ∈ N0 (x, y , q) on B
Assume also that V is locally bounded
Dynamic Programming Equation

Theorem (2.1)
V∗ is a viscosity supersolution of −∂t V∗ + F ∗ V∗ ≥ 0 on [0, T ) × X
If Assumption 2.1 holds, then V ∗ is a viscosity subsolution of
−∂t V ∗ + F∗ V ∗ ≤ 0 on [0, T ) × X
Dynamic Programming Equation

Theorem (2.1)
V∗ is a viscosity supersolution of −∂t V∗ + F ∗ V∗ ≥ 0 on [0, T ) × X
If Assumption 2.1 holds, then V ∗ is a viscosity subsolution of
−∂t V ∗ + F∗ V ∗ ≤ 0 on [0, T ) × X
Let N(x, y , q) = {r ∈ Rd : r = N u (x, y , q) for u ∈ U} and let
δ = dist(0, Nc ) − dist(0, N)
Dynamic Programming Equation

Theorem (2.1)
V∗ is a viscosity supersolution of −∂t V∗ + F ∗ V∗ ≥ 0 on [0, T ) × X
If Assumption 2.1 holds, then V ∗ is a viscosity subsolution of
−∂t V ∗ + F∗ V ∗ ≤ 0 on [0, T ) × X
Let N(x, y , q) = {r ∈ Rd : r = N u (x, y , q) for u ∈ U} and let
δ = dist(0, Nc ) − dist(0, N)
Theorem (2.2)
The function x ∈ X 7→ V∗ (T , x) is a viscosity supersolution of
min{(V∗ (T , •) − g∗ )1{F ∗ V∗ (T ,•)<∞} , δ ∗ V∗ (T , •)} ≥ 0 on X and if
Assumption 2.1 holds, x 7→ V ∗ (T , x) is a viscosity subsolution of
min{(V ∗ (T , •) − g ∗ ), δ∗ V ∗ (T , •)} ≤ 0 on X
Problem Reduction

I For α an F-Progressively measurable Rd -valued P-almost

surely square integrable process and p ∈ [0, 1], define
α (t) = p, dP α (s) = P α (s) = P α (s)(1 − P α (s))α dW
Pt,p t,p t,p t,p t,p s s
for s ∈ [t, T ]
Problem Reduction

I For α an F-Progressively measurable Rd -valued P-almost

surely square integrable process and p ∈ [0, 1], define
α (t) = p, dP α (s) = P α (s) = P α (s)(1 − P α (s))α dW
Pt,p t,p t,p t,p t,p s s
for s ∈ [t, T ]
I Let X̄ = (X , P), X̄ = X × (0, 1), Ū = U × Rd

Theorem
For t ∈ [0, T ] and x̄ = (x, p) ∈ X̄, V̄ (t, x̄) = inf{y ∈ R+ :
ν̄ (T ), Y ν
Ḡ (X̄t,x̄ t,x,y (T )) ≥ 0 for some ν̄ ∈ Ū}
Problem Reduction

I For α an F-Progressively measurable Rd -valued P-almost

surely square integrable process and p ∈ [0, 1], define
α (t) = p, dP α (s) = P α (s) = P α (s)(1 − P α (s))α dW
Pt,p t,p t,p t,p t,p s s
for s ∈ [t, T ]
I Let X̄ = (X , P), X̄ = X × (0, 1), Ū = U × Rd
I Define Ḡ (x̄, y ) = 1{G (x,y )≥0} − p.

Theorem
For t ∈ [0, T ] and x̄ = (x, p) ∈ X̄, V̄ (t, x̄) = inf{y ∈ R+ :
ν̄ (T ), Y ν
Ḡ (X̄t,x̄ t,x,y (T )) ≥ 0 for some ν̄ ∈ Ū}
Proof of Theorem

I Suppose y > (RHS), then for some ν̄ ∈ Ū we have

ν̄ (T ), Y ν
Ḡ (X̄t,x̄ t,x,y (T )) ≥ 0.
Proof of Theorem

I Suppose y > (RHS), then for some ν̄ ∈ Ū we have

ν̄ (T ), Y ν
Ḡ (X̄t,x̄ t,x,y (T )) ≥ 0.
I Applying E to the above inequality and using the fact that
α is a martingale, we have
Pt,p
E[1{G (Xt,x
ν (T ),Y ν
t,x,y (T ))≥0}
]−p ≥0
Proof of Theorem

I Suppose y > (RHS), then for some ν̄ ∈ Ū we have

ν̄ (T ), Y ν
Ḡ (X̄t,x̄ t,x,y (T )) ≥ 0.
I Applying E to the above inequality and using the fact that
α is a martingale, we have
Pt,p
E[1{G (Xt,x
ν (T ),Y ν
t,x,y (T ))≥0}
]−p ≥0
I Suppose y > V̄ (t, x, p), then there exists ν ∈ U with
ν (T ), Y ν
P[G (Xt,x t,x,y (T )) ≥ 0] ≥ p
Proof of Theorem

I Suppose y > (RHS), then for some ν̄ ∈ Ū we have

ν̄ (T ), Y ν
Ḡ (X̄t,x̄ t,x,y (T )) ≥ 0.
I Applying E to the above inequality and using the fact that
α is a martingale, we have
Pt,p
E[1{G (Xt,x
ν (T ),Y ν
t,x,y (T ))≥0}
]−p ≥0
I Suppose y > V̄ (t, x, p), then there exists ν ∈ U with
ν (T ), Y ν
P[G (Xt,x t,x,y (T )) ≥ 0] ≥ p
I By the Stochastic Integral Representation Theorem, there
exists φ with 1{G (Xt,x ν (T ),Y ν
t,x,y (T ))≥0}
=
RT
E[1{G (Xt,x
ν (T ),Y ν
t,x,y (T ))≥0}
] + t φs dWs
Proof of Theorem

I Suppose y > (RHS), then for some ν̄ ∈ Ū we have

ν̄ (T ), Y ν
Ḡ (X̄t,x̄ t,x,y (T )) ≥ 0.
I Applying E to the above inequality and using the fact that
α is a martingale, we have
Pt,p
E[1{G (Xt,x
ν (T ),Y ν
t,x,y (T ))≥0}
]−p ≥0
I Suppose y > V̄ (t, x, p), then there exists ν ∈ U with
ν (T ), Y ν
P[G (Xt,x t,x,y (T )) ≥ 0] ≥ p
I By the Stochastic Integral Representation Theorem, there
exists φ with 1{G (Xt,x ν (T ),Y ν
t,x,y (T ))≥0}
=
RT
E[1{G (Xt,x
ν (T ),Y ν
t,x,y (T ))≥0}
] + t φs dWs
φ
I Let τ be the first time that Pt,p hits 0 and let α = φ1[t,τ ] .
− α
Then 1{G (Xt,x
ν (T ),Y ν
t,x,y (T ))≥0}
P t,p (T ) ≥ 0
Dynamic Programming Equation

I µ̄(x̄, ū) = (µ(x, u), 0)T σ̄(x̄, ū) = (σ(x, u), αT )T

Dynamic Programming Equation

I µ̄(x̄, ū) = (µ(x, u), 0)T σ̄(x̄, ū) = (σ(x, u), αT )T

I For (y , q, A) ∈ R × Rd+1 × Sd+1 , and for ū = (u, α) ∈ Ū
N̄ ū (x̄, y , q) = σY (x, y , u) − σ̄(x̄, ū)T q = N u (x, y , qx ) − qp α
Dynamic Programming Equation

I µ̄(x̄, ū) = (µ(x, u), 0)T σ̄(x̄, ū) = (σ(x, u), αT )T

I For (y , q, A) ∈ R × Rd+1 × Sd+1 , and for ū = (u, α) ∈ Ū
N̄ ū (x̄, y , q) = σY (x, y , u) − σ̄(x̄, ū)T q = N u (x, y , qx ) − qp α
I N̄ (x̄, y , q) = {ū ∈ Ū : |N̄ ū (x̄, y , q)| ≤ }
Dynamic Programming Equation

I µ̄(x̄, ū) = (µ(x, u), 0)T σ̄(x̄, ū) = (σ(x, u), αT )T

I For (y , q, A) ∈ R × Rd+1 × Sd+1 , and for ū = (u, α) ∈ Ū
N̄ ū (x̄, y , q) = σY (x, y , u) − σ̄(x̄, ū)T q = N u (x, y , qx ) − qp α
I N̄ (x̄, y , q) = {ū ∈ Ū : |N̄ ū (x̄, y , q)| ≤ }
I F̄ (x̄, y , q, A) =
supū∈N̄ (x̄,y ,q) {µY (x, y , u)− µ̄(x̄, u, α)q − 21 Tr(σ̄σ̄ T (x̄, u, α)A)}
Dynamic Programming Equation

I µ̄(x̄, ū) = (µ(x, u), 0)T σ̄(x̄, ū) = (σ(x, u), αT )T

I For (y , q, A) ∈ R × Rd+1 × Sd+1 , and for ū = (u, α) ∈ Ū
N̄ ū (x̄, y , q) = σY (x, y , u) − σ̄(x̄, ū)T q = N u (x, y , qx ) − qp α
I N̄ (x̄, y , q) = {ū ∈ Ū : |N̄ ū (x̄, y , q)| ≤ }
I F̄ (x̄, y , q, A) =
supū∈N̄ (x̄,y ,q) {µY (x, y , u)− µ̄(x̄, u, α)q − 21 Tr(σ̄σ̄ T (x̄, u, α)A)}
I N̄(x̄, y , q) = {N ū (x̄, y , q) : ū ∈ Ū} and
δ̄ = dist(o, Nc ) − dist(0, N̄)
Dynamic Programming Equation (continued)

Assumption 3.1
Let B ⊂ X × [0, 1] × R × Rd+1 such that N̄0 6= ∅ on B. Then for
every  > 0, (x0 , p0 , y0 , q0 ) ∈ intB, ū0 ∈ N̄0 (x0 , p0 , y0 , q0 ), there
0
exists an open neighborhood B of (x0 , p0 , y0 , q0 ) and a locally
0
Lipschitz map ν̂ on B such that |ν̂(x0 , p0 , y0 , q0 ) − ū0 | ≤  and
0
ν̂(x, p, y , q) ∈ N̄0 (x, p, y , q) on B
Dynamic Programming Equation (continued)

Assumption 3.1
Let B ⊂ X × [0, 1] × R × Rd+1 such that N̄0 6= ∅ on B. Then for
every  > 0, (x0 , p0 , y0 , q0 ) ∈ intB, ū0 ∈ N̄0 (x0 , p0 , y0 , q0 ), there
0
exists an open neighborhood B of (x0 , p0 , y0 , q0 ) and a locally
0
Lipschitz map ν̂ on B such that |ν̂(x0 , p0 , y0 , q0 ) − ū0 | ≤  and
0
ν̂(x, p, y , q) ∈ N̄0 (x, p, y , q) on B

Corollary 3.1
The functtion V̄∗ is a viscosity supersolution of −∂t V̄∗ + F̄ ∗ V̄∗ ≥ 0
on [0, T ) × X̄. Under the additional Assumption 3.1, V̄ ∗ is a
viscosity subsolution of min{V̄ ∗ , −∂t V̄ ∗ + F̄∗ V̄ ∗ } on [0, T ) × X̄.
Boundary Conditions

Remark
Note that V̄ (•, 1) = V and V̄ (•, 0) = 0. Since G is non-decreasing
in y , 0 ≤ V̄∗ (•, 0) ≤ V̄ ∗ (•, 1) ≤ V ∗
Boundary Conditions

Remark
Note that V̄ (•, 1) = V and V̄ (•, 0) = 0. Since G is non-decreasing
in y , 0 ≤ V̄∗ (•, 0) ≤ V̄ ∗ (•, 1) ≤ V ∗

Assumption 3.2
For all (x, y , q) ∈ X × (0, ∞) × Rd , we have N0 (x, y , q) is a proper
subset of U
Boundary Conditions

Remark
Note that V̄ (•, 1) = V and V̄ (•, 0) = 0. Since G is non-decreasing
in y , 0 ≤ V̄∗ (•, 0) ≤ V̄ ∗ (•, 1) ≤ V ∗

Assumption 3.2
For all (x, y , q) ∈ X × (0, ∞) × Rd , we have N0 (x, y , q) is a proper
subset of U

Assumption 3.3
For all compact subsets A of Rd × R × Rd × Sd , there exists
C > 0 such that F (Θ) ≤ C (1 + 2 ) for all  ≥ 0 and Θ ∈ A.
Boundary Conditions (continued)

Theorem3.1
Assume the function supu∈U |σ(•, u)| is locally bounded on X and
that Assumption 3.1 holds, then (i) Under Assumption 3.2,
V̄ ∗ (•, 0) = 0 on [0, T ) × X and V̄∗ (•, 0) = 0 on [0, T ] × X (ii)
Under Assumption 3.3, V̄ ∗ (•, 1) is a viscosity supersolution of
(2.10)-(2.19) on [0, T ] × X. In particular, if Assumption 2.2 is
satisfied, then V̄ ∗ (•, 1) = V̄∗ (•, 1) = V∗ = V ∗ on [0, T ] × X
Quantile Hedging Problem

I X = (0, ∞)d
Quantile Hedging Problem

I X = (0, ∞)d
I µ(x, u) = µ(x), σ(x, u) = σ(x) with σ invertible and
supx∈X |λ(x)| < ∞ where λ = σ −1 µ
Quantile Hedging Problem

I X = (0, ∞)d
I µ(x, u) = µ(x), σ(x, u) = σ(x) with σ invertible and
supx∈X |λ(x)| < ∞ where λ = σ −1 µ
I µY (x, y , u) = uµ(x), σY (x, y , u) = σ T (x)u
Quantile Hedging Problem

I X = (0, ∞)d
I µ(x, u) = µ(x), σ(x, u) = σ(x) with σ invertible and
supx∈X |λ(x)| < ∞ where λ = σ −1 µ
I µY (x, y , u) = uµ(x), σY (x, y , u) = σ T (x)u
I G (x, y ) = y − g (x) for g Lipschitz
Quantile Hedging Problem

I X = (0, ∞)d
I µ(x, u) = µ(x), σ(x, u) = σ(x) with σ invertible and
supx∈X |λ(x)| < ∞ where λ = σ −1 µ
I µY (x, y , u) = uµ(x), σY (x, y , u) = σ T (x)u
I G (x, y ) = y − g (x) for g Lipschitz
I dXt,x (s) = µ(Xt,x (s))ds + σ(Xt,x (s))dWs where Xt,x (t) = x.
Quantile Hedging Problem

I X = (0, ∞)d
I µ(x, u) = µ(x), σ(x, u) = σ(x) with σ invertible and
supx∈X |λ(x)| < ∞ where λ = σ −1 µ
I µY (x, y , u) = uµ(x), σY (x, y , u) = σ T (x)u
I G (x, y ) = y − g (x) for g Lipschitz
I dXt,x (s) = µ(Xt,x (s))ds + σ(Xt,x (s))dWs where Xt,x (t) = x.
ν
Rs
I Yt,x,y (s) = y + t νr dXt,x (r )
Quantile Hedging Problem

I X = (0, ∞)d
I µ(x, u) = µ(x), σ(x, u) = σ(x) with σ invertible and
supx∈X |λ(x)| < ∞ where λ = σ −1 µ
I µY (x, y , u) = uµ(x), σY (x, y , u) = σ T (x)u
I G (x, y ) = y − g (x) for g Lipschitz
I dXt,x (s) = µ(Xt,x (s))ds + σ(Xt,x (s))dWs where Xt,x (t) = x.
ν
Rs
I Yt,x,y (s) = y + t νr dXt,x (r )
I V (t, x) = EQt,x [g (Xt,x (T ))], where Qt,x is the P-equivalent
Martingale measure
R defined by 2
−1 T
dQt,x RT
dP = exp( 2 t |λ(Xt,x (s))| ds − t λ(Xt,x (s))dWs )
Quantile Hedging Problem

I X = (0, ∞)d
I µ(x, u) = µ(x), σ(x, u) = σ(x) with σ invertible and
supx∈X |λ(x)| < ∞ where λ = σ −1 µ
I µY (x, y , u) = uµ(x), σY (x, y , u) = σ T (x)u
I G (x, y ) = y − g (x) for g Lipschitz
I dXt,x (s) = µ(Xt,x (s))ds + σ(Xt,x (s))dWs where Xt,x (t) = x.
ν
Rs
I Yt,x,y (s) = y + t νr dXt,x (r )
I V (t, x) = EQt,x [g (Xt,x (T ))], where Qt,x is the P-equivalent
Martingale measure
R defined by 2
−1 T
dQt,x RT
dP = exp( 2 t |λ(Xt,x (s))| ds − t λ(Xt,x (s))dWs )
R•
I Define W Qt,x = W + t λ(Xt,x (s))ds, the Qt,x -Brownian
motion defined on [t, T ]
Quantile Hedging Problem

DPE
By Corollary 3.1, Theorem 3.1 and (i) of Proposition 3.2, one has
that V̄∗ is a viscosity supersolution on [0, T ) × X̄ of
0 ≤ −∂t V̄∗ + F̄ ∗ V̄∗ = −∂t V̄∗ − 21 Tr(σσ T Dxx V̄∗ ) −
infα∈Rd −α(Dp V̄∗ )T σ −1 µ + Tr(σαDxp V̄∗ ) + 21 |α|2 Dpp V̄∗


Quantile Hedging Problem

DPE
By Corollary 3.1, Theorem 3.1 and (i) of Proposition 3.2, one has
that V̄∗ is a viscosity supersolution on [0, T ) × X̄ of
0 ≤ −∂t V̄∗ + F̄ ∗ V̄∗ = −∂t V̄∗ − 21 Tr(σσ T Dxx V̄∗ ) −
infα∈Rd −α(Dp V̄∗ )T σ −1 µ + Tr(σαDxp V̄∗ ) + 21 |α|2 Dpp V̄∗

Boundary Conditions
V̄∗ (•, 1) = V and V̄∗ (•, 0) = 0 on [0, T ] × X, V̄∗ (T , x, p) ≥ pg (x)
on X × [0, 1]. We define V̄∗ (•, p) = 0 for p < 0 and V̄∗ (•, p) = ∞
for p > 1.
Quantile Hedging Problem

DPE
By Corollary 3.1, Theorem 3.1 and (i) of Proposition 3.2, one has
that V̄∗ is a viscosity supersolution on [0, T ) × X̄ of
0 ≤ −∂t V̄∗ + F̄ ∗ V̄∗ = −∂t V̄∗ − 21 Tr(σσ T Dxx V̄∗ ) −
infα∈Rd −α(Dp V̄∗ )T σ −1 µ + Tr(σαDxp V̄∗ ) + 21 |α|2 Dpp V̄∗

Boundary Conditions
V̄∗ (•, 1) = V and V̄∗ (•, 0) = 0 on [0, T ] × X, V̄∗ (T , x, p) ≥ pg (x)
on X × [0, 1]. We define V̄∗ (•, p) = 0 for p < 0 and V̄∗ (•, p) = ∞
for p > 1.

How do we solve this system?

They use the (standard) trick of taking the Legendre transform (in
the p-variable) to convert this to a linear problem. Let
v (t, x, q) = supp∈R {pq − V̄∗ (t, x, p)} for
(t, x, q) ∈ [0, T ] × (0, ∞)d × R
Transformed Problem

I v (•, q) = ∞ for q < 0 and

v (•, q) = supp∈[0,1] {pq − V̄∗ (•, p)} for q > 0.
Transformed Problem

I v (•, q) = ∞ for q < 0 and

v (•, q) = supp∈[0,1] {pq − V̄∗ (•, p)} for q > 0.
I (4.11) v is an upper-semicontinuous viscosity subsolution of
−∂t v − 21 Tr(σσ T Dxx v ) − 12 |λ|2 q 2 Dqq v − Tr(σλDx,q v ) ≤ 0
Transformed Problem

I v (•, q) = ∞ for q < 0 and

v (•, q) = supp∈[0,1] {pq − V̄∗ (•, p)} for q > 0.
I (4.11) v is an upper-semicontinuous viscosity subsolution of
−∂t v − 21 Tr(σσ T Dxx v ) − 12 |λ|2 q 2 Dqq v − Tr(σλDx,q v ) ≤ 0
I (4.12) v (T , x, q) ≤ (q − g (x))+
Transformed Problem

I v (•, q) = ∞ for q < 0 and

v (•, q) = supp∈[0,1] {pq − V̄∗ (•, p)} for q > 0.
I (4.11) v is an upper-semicontinuous viscosity subsolution of
−∂t v − 21 Tr(σσ T Dxx v ) − 12 |λ|2 q 2 Dqq v − Tr(σλDx,q v ) ≤ 0
I (4.12) v (T , x, q) ≤ (q − g (x))+
I Uppersemicontinuity of v follows from the lower
semi-continuity of V̄∗ .
Transformed Problem

I v (•, q) = ∞ for q < 0 and

v (•, q) = supp∈[0,1] {pq − V̄∗ (•, p)} for q > 0.
I (4.11) v is an upper-semicontinuous viscosity subsolution of
−∂t v − 21 Tr(σσ T Dxx v ) − 12 |λ|2 q 2 Dqq v − Tr(σλDx,q v ) ≤ 0
I (4.12) v (T , x, q) ≤ (q − g (x))+
I Uppersemicontinuity of v follows from the lower
semi-continuity of V̄∗ .
I To prove (4.11), let (t0 , x0 , q0 ) ∈ [0, T ) × (0, ∞)d × (0, ∞)
and let φ be a smooth local maximizer of v − φ at (t0 , x0 , q0 )
with (v − φ)(t0 , x0 , q0 ) = 0
Transformed Problem

I v (•, q) = ∞ for q < 0 and

v (•, q) = supp∈[0,1] {pq − V̄∗ (•, p)} for q > 0.
I (4.11) v is an upper-semicontinuous viscosity subsolution of
−∂t v − 21 Tr(σσ T Dxx v ) − 12 |λ|2 q 2 Dqq v − Tr(σλDx,q v ) ≤ 0
I (4.12) v (T , x, q) ≤ (q − g (x))+
I Uppersemicontinuity of v follows from the lower
semi-continuity of V̄∗ .
I To prove (4.11), let (t0 , x0 , q0 ) ∈ [0, T ) × (0, ∞)d × (0, ∞)
and let φ be a smooth local maximizer of v − φ at (t0 , x0 , q0 )
with (v − φ)(t0 , x0 , q0 ) = 0
I WLOG, q 7→ φ(•, q) is strictly convex. Since v is convex
Dqq φ(t0 , x0 , q0 ) ≥ 0. Given , η > 0, define φ,η (t, x, q) =
φ(t, x, q)+|q −q0 |2 +η|q −q0 |2 (|q −q0 |2 +|t −t0 |2 +|x −x0 |2 )
Transformed Problem (continued)

I (t0 , x0 , q0 ) still maximizes v − φ,η and Dqq φ(t0 , x0 , q0 ) ≥ 0

implies Dqq φ,η (t0 , x0 , q0 ) ≥ 2 > 0 Since φ has bounded
derivatives, we can choose η large enough that
Dqq φ,η (t, x, q) > 0
Transformed Problem (continued)

I (t0 , x0 , q0 ) still maximizes v − φ,η and Dqq φ(t0 , x0 , q0 ) ≥ 0

implies Dqq φ,η (t0 , x0 , q0 ) ≥ 2 > 0 Since φ has bounded
derivatives, we can choose η large enough that
Dqq φ,η (t, x, q) > 0
I (4.11) holds for φ,η implies it for φ since we have
convergence in C 2
Transformed Problem (continued)

I (t0 , x0 , q0 ) still maximizes v − φ,η and Dqq φ(t0 , x0 , q0 ) ≥ 0

implies Dqq φ,η (t0 , x0 , q0 ) ≥ 2 > 0 Since φ has bounded
derivatives, we can choose η large enough that
Dqq φ,η (t, x, q) > 0
I (4.11) holds for φ,η implies it for φ since we have
convergence in C 2
I Assume q 7→ φ(•, q) is strictly convex. Let
φ̃(t, x, p) = supq∈R {pq − φ(t, x, q)}. Since φ is strictly
convex in q and smooth, φ̃ is strictly convex in p and smooth
Transformed Problem (continued)

I φ(t, x, q) = supp∈R {pq − φ̃(t, x, p)} =

J(t, x, q)q − φ̃(t, x, J(t, x, q)) on
(0, T ) × (0, ∞)d × (0, ∞) ⊂ int(dom(φ)) where q 7→ J(•, q)
denotes the inverse of p 7→ Dp φ̃(•, p)
Transformed Problem (continued)

I φ(t, x, q) = supp∈R {pq − φ̃(t, x, p)} =

J(t, x, q)q − φ̃(t, x, J(t, x, q)) on
(0, T ) × (0, ∞)d × (0, ∞) ⊂ int(dom(φ)) where q 7→ J(•, q)
denotes the inverse of p 7→ Dp φ̃(•, p)
I Since q0 > 0 we can find p0 ∈ [0, 1] such that
v (t0 , x0 , q0 ) = p0 q0 − V̄∗ (t0 , x0 , p0 ) which implies that
(t0 , x0 , p0 ) is a minimizer of V̄∗ − φ̃ such that
(V̄∗ − φ̃)(t0 , x0 , p0 ) = 0 and
φ(t0 , x0 , q0 ) = supp∈R {pq0 − φ̃(t0 , x0 , p)} = p0 q0 − φ̃(t0 , x0 , p0 )
with p0 = J(t0 , x0 , q0 )
Solution of Transformed Problem

I Recall (4.11) v is an upper-semicontinuous viscosity

subsolution of
−∂t v − 12 Tr(σσ T Dxx v ) − 12 |λ|2 q 2 Dqq v − Tr(σλDx,q v ) ≤ 0
Solution of Transformed Problem

I Recall (4.11) v is an upper-semicontinuous viscosity

subsolution of
−∂t v − 12 Tr(σσ T Dxx v ) − 12 |λ|2 q 2 Dqq v − Tr(σλDx,q v ) ≤ 0
I Since (4.11) is linear, Feynman-Kac yields an upper bound,
namely
v (t, x, q) ≤ v̄ (t, x, q) = EQt,x [(Qt,x,q (T ) − g (Xt,x (T )))+ ],
where dQ(s)
Qt,x
Q(s) = λ(Xt,x (s))dWs and Qt,x,q (t) = q
Solution of Transformed Problem

I Recall (4.11) v is an upper-semicontinuous viscosity

subsolution of
−∂t v − 12 Tr(σσ T Dxx v ) − 12 |λ|2 q 2 Dqq v − Tr(σλDx,q v ) ≤ 0
I Since (4.11) is linear, Feynman-Kac yields an upper bound,
namely
v (t, x, q) ≤ v̄ (t, x, q) = EQt,x [(Qt,x,q (T ) − g (Xt,x (T )))+ ],
where dQ(s)
Qt,x
Q(s) = λ(Xt,x (s))dWs and Qt,x,q (t) = q
I v̄ provides a lower bound for V̄
Solution (continued)

I (4.15) v̄ is convex in q and there is a unique solution q̂ of

∂v̄ Qt,x [Q
∂q (t, x, q̄) = E t,x,1 (T )1{Qt,x,q̂ (T )≥g (Xt,x (T ))} ] =
P[Qt,x,q̂ (T ) ≥ g (Xt,x (T ))] = p, where we used
dP
dQt,x = Qt,x,1 (T )
Solution (continued)

I (4.15) v̄ is convex in q and there is a unique solution q̂ of

∂v̄ Qt,x [Q
∂q (t, x, q̄) = E t,x,1 (T )1{Qt,x,q̂ (T )≥g (Xt,x (T ))} ] =
P[Qt,x,q̂ (T ) ≥ g (Xt,x (T ))] = p, where we used
dP
dQt,x = Qt,x,1 (T )
I V̄ (t, x, p) ≥ V̄∗ (t, x, p) ≥ pq̂(t, x, p) − v̄ (t, x, q̂(t, x, p)) =
q̂(t, x, p)(p − EQt,x [Qt,x,1 (T )1{q̂(t,x,p)Qt,x,1 (T )≥g (Xt,x (T ))} ]) +
EQt,x [g (Xt,x (T ))1{q̂(t,x,p)Qt,x,1 (T )≥g (Xt,x (T ))} ] =
EQt,x [g (Xt,x (T ))1{q̂(t,x,p)Qt,x,1 (T )≥g (Xt,x (T ))} ] =: y (t, x, p)
Solution (continued)

I (4.15) v̄ is convex in q and there is a unique solution q̂ of

∂v̄ Qt,x [Q
∂q (t, x, q̄) = E t,x,1 (T )1{Qt,x,q̂ (T )≥g (Xt,x (T ))} ] =
P[Qt,x,q̂ (T ) ≥ g (Xt,x (T ))] = p, where we used
dP
dQt,x = Qt,x,1 (T )
I V̄ (t, x, p) ≥ V̄∗ (t, x, p) ≥ pq̂(t, x, p) − v̄ (t, x, q̂(t, x, p)) =
q̂(t, x, p)(p − EQt,x [Qt,x,1 (T )1{q̂(t,x,p)Qt,x,1 (T )≥g (Xt,x (T ))} ]) +
EQt,x [g (Xt,x (T ))1{q̂(t,x,p)Qt,x,1 (T )≥g (Xt,x (T ))} ] =
EQt,x [g (Xt,x (T ))1{q̂(t,x,p)Qt,x,1 (T )≥g (Xt,x (T ))} ] =: y (t, x, p)
I By Martingale Representation Theorem, there exists ν ∈ U
ν
with Yt,x,y (t,x,p) (T ) ≥ g (Xt,x (T ))1{q̂(t,x,p)Qt,x,1 (T )≥g (Xt,x (T ))}
Solution (continued)

I (4.15) v̄ is convex in q and there is a unique solution q̂ of

∂v̄ Qt,x [Q
∂q (t, x, q̄) = E t,x,1 (T )1{Qt,x,q̂ (T )≥g (Xt,x (T ))} ] =
P[Qt,x,q̂ (T ) ≥ g (Xt,x (T ))] = p, where we used
dP
dQt,x = Qt,x,1 (T )
I V̄ (t, x, p) ≥ V̄∗ (t, x, p) ≥ pq̂(t, x, p) − v̄ (t, x, q̂(t, x, p)) =
q̂(t, x, p)(p − EQt,x [Qt,x,1 (T )1{q̂(t,x,p)Qt,x,1 (T )≥g (Xt,x (T ))} ]) +
EQt,x [g (Xt,x (T ))1{q̂(t,x,p)Qt,x,1 (T )≥g (Xt,x (T ))} ] =
EQt,x [g (Xt,x (T ))1{q̂(t,x,p)Qt,x,1 (T )≥g (Xt,x (T ))} ] =: y (t, x, p)
I By Martingale Representation Theorem, there exists ν ∈ U
ν
with Yt,x,y (t,x,p) (T ) ≥ g (Xt,x (T ))1{q̂(t,x,p)Qt,x,1 (T )≥g (Xt,x (T ))}
I By (4.15), this implies that V̄ (t, x, p) = y (t, x, p)
Derivation of DPE for Singular Stochastic Target Problems

Theorem (Geometric Dynamic Programming Principle)

Let (t, x) ∈ [0, T ) × X and θ be a [s, T ]-valued stopping time.
ν
Then, V (t, x) = inf{y ∈ R : Yt,x,y (θ) ≥
ν
V (θ, Xt,x (θ)) a.s. for some ν ∈ U}.
Derivation of DPE for Singular Stochastic Target Problems

Theorem (Geometric Dynamic Programming Principle)

Let (t, x) ∈ [0, T ) × X and θ be a [s, T ]-valued stopping time.
ν
Then, V (t, x) = inf{y ∈ R : Yt,x,y (θ) ≥
ν
V (θ, Xt,x (θ)) a.s. for some ν ∈ U}.
One obtains the DPE by using GDPP and choosing stopping times
appropriately. The arguments adapt those in Soner and Touzi.
Open Questions and Future Work

I The authors find a way to solve a control problem where the

constraints are satisfied with fixed probability. Can this
method be used to solve the corresponding problem for other
constrained control problems?
Open Questions and Future Work

I The authors find a way to solve a control problem where the

constraints are satisfied with fixed probability. Can this
method be used to solve the corresponding problem for other
constrained control problems?
I Are there other interesting applications of DPE with
unbounded controls?
Open Questions and Future Work

I The authors find a way to solve a control problem where the

constraints are satisfied with fixed probability. Can this
method be used to solve the corresponding problem for other
constrained control problems?
I Are there other interesting applications of DPE with
unbounded controls?
I Can the regularity assumptions be further relaxed?