Upload
donhu
View
220
Download
1
Embed Size (px)
Citation preview
Lipschitz Changes of Variables between Perturbations ofLog-concave Measures
Yash Jhaveri
University of Texas at Austin
Prairie Analysis SeminarKansas State University
September 25, 2015
Joint work with Maria Colombo and Alessio Figalli.
Yash Jhaveri (UT Austin) Lipschitz Changes of Variables September 25, 2015 1 / 16
An Old Result
Theorem (Caffarelli)
Let Q be a quadratic polynomial with D2Q > 0 and V be a convexfunction such that e−Q , e−Q−V ∈ P(Rn). Then, there exists a1-Lipschitz change of variables T taking e−Q to e−Q−V .
Yash Jhaveri (UT Austin) Lipschitz Changes of Variables September 25, 2015 2 / 16
A New Result
Theorem (Colombo, Figalli, -)
Let V be a convex function with V (0) = inf V , R > 0, q ∈ Cc(BR), and cq ∈ Rbe such that e−V , e−V+cq−q ∈ P(Rn). Assume that λ Id ≤ D2V ≤ Λ Id andD2q ≥ −λq Id for 0 < λ,Λ, λq <∞ in the sense of distributions. Then, thereexists a change of variables T taking e−V to e−V+cq−q that satisfies
‖∇T‖L∞(Rn) ≤ C with C = C (R, λ,Λ, λq) > 0
Yash Jhaveri (UT Austin) Lipschitz Changes of Variables September 25, 2015 3 / 16
Existence: The Optimal Transport Problem
- The map T is the unique solution of the Monge problem forquadratic cost:
min
{∫Rn
|x − T (x)|2 dµ(x) : T#µ = ν
},
where µ = e−V (x) dx , ν = e−V (y)+cq−q(y) dy ∈ P(Rn).
Yash Jhaveri (UT Austin) Lipschitz Changes of Variables September 25, 2015 4 / 16
Brenier’s Theorem
Theorem (Brenier)
Let µ, ν ∈ P(Rn) such that µ = f (x) dx and∫Rn
|x |2 dµ(x) +
∫Rn
|y |2 dν(y) <∞.
Then, there exists a unique, optimal transport T taking µ to ν.Moreover, there is a convex function φ : Rn → R such that
T = ∇φ.
Yash Jhaveri (UT Austin) Lipschitz Changes of Variables September 25, 2015 5 / 16
Regularity: OT and the Monge-Ampere Equation
- Let µ = f (x) dx , ν = g(y) dy ∈ P(Rn), ϕ ∈ C∞c (Rn), andassume that the OT map T taking f to g is a smoothdiffeomorphism.
Then,∫ϕ(T (x))f (x) dx
T#µ=ν=
∫ϕ(y)g(y) dy
y=T (x)=
∫ϕ(T (x))g(T (x))| det∇T (x)| dx .
- Since ϕ ∈ C∞c (Rn) was arbitrary,
f (x) = g(T (x))| det∇T (x)|.
- Recalling Brenier, if g > 0,
detD2φ =f
g(∇φ).
Yash Jhaveri (UT Austin) Lipschitz Changes of Variables September 25, 2015 6 / 16
Regularity: OT and the Monge-Ampere Equation
- Let µ = f (x) dx , ν = g(y) dy ∈ P(Rn), ϕ ∈ C∞c (Rn), andassume that the OT map T taking f to g is a smoothdiffeomorphism. Then,∫
ϕ(T (x))f (x) dxT#µ=ν
=
∫ϕ(y)g(y) dy
y=T (x)=
∫ϕ(T (x))g(T (x))| det∇T (x)| dx .
- Since ϕ ∈ C∞c (Rn) was arbitrary,
f (x) = g(T (x))| det∇T (x)|.
- Recalling Brenier, if g > 0,
detD2φ =f
g(∇φ).
Yash Jhaveri (UT Austin) Lipschitz Changes of Variables September 25, 2015 6 / 16
Regularity: OT and the Monge-Ampere Equation
- Let µ = f (x) dx , ν = g(y) dy ∈ P(Rn), ϕ ∈ C∞c (Rn), andassume that the OT map T taking f to g is a smoothdiffeomorphism. Then,∫
ϕ(T (x))f (x) dxT#µ=ν
=
∫ϕ(y)g(y) dy
y=T (x)=
∫ϕ(T (x))g(T (x))| det∇T (x)| dx .
- Since ϕ ∈ C∞c (Rn) was arbitrary,
f (x) = g(T (x))| det∇T (x)|.
- Recalling Brenier, if g > 0,
detD2φ =f
g(∇φ).
Yash Jhaveri (UT Austin) Lipschitz Changes of Variables September 25, 2015 6 / 16
Regularity: OT and the Monge-Ampere Equation
- Let µ = f (x) dx , ν = g(y) dy ∈ P(Rn), ϕ ∈ C∞c (Rn), andassume that the OT map T taking f to g is a smoothdiffeomorphism. Then,∫
ϕ(T (x))f (x) dxT#µ=ν
=
∫ϕ(y)g(y) dy
y=T (x)=
∫ϕ(T (x))g(T (x))| det∇T (x)| dx .
- Since ϕ ∈ C∞c (Rn) was arbitrary,
f (x) = g(T (x))| det∇T (x)|.
- Recalling Brenier, if g > 0,
detD2φ =f
g(∇φ).
Yash Jhaveri (UT Austin) Lipschitz Changes of Variables September 25, 2015 6 / 16
Some Heuristics
- Since T = ∇φ and φ is convex, we know that φαα ≥ 0 for everydirection α.
- So, we just need to find an upper bound on φαα.
- Let’s assume that φαα attains a global maximum at some pointx0 ∈ Rn, and without loss of generality, assume that this happensin the e1-direction.
Yash Jhaveri (UT Austin) Lipschitz Changes of Variables September 25, 2015 7 / 16
Some Heuristics
- Since T = ∇φ and φ is convex, we know that φαα ≥ 0 for everydirection α.
- So, we just need to find an upper bound on φαα.
- Let’s assume that φαα attains a global maximum at some pointx0 ∈ Rn, and without loss of generality, assume that this happensin the e1-direction.
Yash Jhaveri (UT Austin) Lipschitz Changes of Variables September 25, 2015 7 / 16
Some Heuristics
- Since T = ∇φ and φ is convex, we know that φαα ≥ 0 for everydirection α.
- So, we just need to find an upper bound on φαα.
- Let’s assume that φαα attains a global maximum at some pointx0 ∈ Rn, and without loss of generality, assume that this happensin the e1-direction.
Yash Jhaveri (UT Austin) Lipschitz Changes of Variables September 25, 2015 7 / 16
Some Heuristics
- Since our potential φ solves a MA equation,
log detD2φ = −V + V (∇φ)− cq + q(∇φ).
- Differentiating the left side twice in the e1-direction andevaluating at x0,
∂11(LHS) = − tr([
[D2φ]−1D2φ1
]2)+ tr
([D2φ]−1D2φ11
)≤ 0
- On the right side,
∂11(RHS) = −V11 + 〈∇V (∇φ),∇φ11〉+ 〈D2V (∇φ)∇φ1,∇φ1〉+ 〈∇q(∇φ),∇φ11〉+ 〈D2q(∇φ)∇φ1,∇φ1〉
≥ −Λ + (λ− λq)|∇φ1(x0)|2.
- Hence, if λ− λq > 0, then |∇φ1(x0)| ≤√
Λ/(λ− λq).
Yash Jhaveri (UT Austin) Lipschitz Changes of Variables September 25, 2015 8 / 16
Some Heuristics
- Since our potential φ solves a MA equation,
log detD2φ = −V + V (∇φ)− cq + q(∇φ).
- Differentiating the left side twice in the e1-direction andevaluating at x0,
∂11(LHS) = − tr([
[D2φ]−1D2φ1
]2)+ tr
([D2φ]−1D2φ11
)≤ 0
- On the right side,
∂11(RHS) = −V11 + 〈∇V (∇φ),∇φ11〉+ 〈D2V (∇φ)∇φ1,∇φ1〉+ 〈∇q(∇φ),∇φ11〉+ 〈D2q(∇φ)∇φ1,∇φ1〉
≥ −Λ + (λ− λq)|∇φ1(x0)|2.
- Hence, if λ− λq > 0, then |∇φ1(x0)| ≤√
Λ/(λ− λq).
Yash Jhaveri (UT Austin) Lipschitz Changes of Variables September 25, 2015 8 / 16
Some Heuristics
- Since our potential φ solves a MA equation,
log detD2φ = −V + V (∇φ)− cq + q(∇φ).
- Differentiating the left side twice in the e1-direction andevaluating at x0,
∂11(LHS) = − tr([
[D2φ]−1D2φ1
]2)+ tr
([D2φ]−1D2φ11
)≤ 0
- On the right side,
∂11(RHS) = −V11 + 〈∇V (∇φ),∇φ11〉+ 〈D2V (∇φ)∇φ1,∇φ1〉+ 〈∇q(∇φ),∇φ11〉+ 〈D2q(∇φ)∇φ1,∇φ1〉
≥ −Λ + (λ− λq)|∇φ1(x0)|2.
- Hence, if λ− λq > 0, then |∇φ1(x0)| ≤√
Λ/(λ− λq).
Yash Jhaveri (UT Austin) Lipschitz Changes of Variables September 25, 2015 8 / 16
Some Heuristics
- Since our potential φ solves a MA equation,
log detD2φ = −V + V (∇φ)− cq + q(∇φ).
- Differentiating the left side twice in the e1-direction andevaluating at x0,
∂11(LHS) = − tr([
[D2φ]−1D2φ1
]2)+ tr
([D2φ]−1D2φ11
)≤ 0
- On the right side,
∂11(RHS) = −V11 + 〈∇V (∇φ),∇φ11〉+ 〈D2V (∇φ)∇φ1,∇φ1〉+ 〈∇q(∇φ),∇φ11〉+ 〈D2q(∇φ)∇φ1,∇φ1〉
≥ −Λ + (λ− λq)|∇φ1(x0)|2.
- Hence, if λ− λq > 0, then |∇φ1(x0)| ≤√
Λ/(λ− λq).
Yash Jhaveri (UT Austin) Lipschitz Changes of Variables September 25, 2015 8 / 16
Some Heuristics
- If OT map takes x0 outside of BR , then the right side becomes,
∂11(RHS) = −V11 + 〈D2V (∇φ)∇φ1,∇φ1〉+ 0 + 0
≥ −Λ + λ|∇φ1(x0)|2.
- And so, |∇φ1(x0)| ≤√
Λ/λ.
Yash Jhaveri (UT Austin) Lipschitz Changes of Variables September 25, 2015 9 / 16
Some Heuristics
- If OT map takes x0 outside of BR , then the right side becomes,
∂11(RHS) = −V11 + 〈D2V (∇φ)∇φ1,∇φ1〉+ 0 + 0
≥ −Λ + λ|∇φ1(x0)|2.
- And so, |∇φ1(x0)| ≤√
Λ/λ.
Yash Jhaveri (UT Austin) Lipschitz Changes of Variables September 25, 2015 9 / 16
Where Does Our (Inverse) Transport Take Points?
- Question: Can we quantify how far points travel under our OT orits inverse?
- Answer: Yes.
Lemma
For any P ≥ R , there exists a constant P ′ = P ′(P , λ,Λ, λq) > 0 suchthat
T−1(BP) ⊆ BP′ .
Yash Jhaveri (UT Austin) Lipschitz Changes of Variables September 25, 2015 10 / 16
Where Does Our (Inverse) Transport Take Points?
- Question: Can we quantify how far points travel under our OT orits inverse?
- Answer: Yes.
Lemma
For any P ≥ R , there exists a constant P ′ = P ′(P , λ,Λ, λq) > 0 suchthat
T−1(BP) ⊆ BP′ .
Yash Jhaveri (UT Austin) Lipschitz Changes of Variables September 25, 2015 10 / 16
Where Does Our (Inverse) Transport Take Points?
- Question: Can we quantify how far points travel under our OT orits inverse?
- Answer: Yes.
Lemma
For any P ≥ R , there exists a constant P ′ = P ′(P , λ,Λ, λq) > 0 suchthat
T−1(BP) ⊆ BP′ .
Yash Jhaveri (UT Austin) Lipschitz Changes of Variables September 25, 2015 10 / 16
A More Refined A Priori Estimate
Lemma
If h(x , α) := φαα(x)eψ(|∇φ(x)|) attains a maximum at some point(x0, α0) among all (x , α) ∈ Rn × Sn−1, then
h(x0, α0) ≤ C with C = C (R , λ,Λ, λq) > 0.
R P
ψ(t)
Yash Jhaveri (UT Austin) Lipschitz Changes of Variables September 25, 2015 11 / 16
A More Refined A Priori Estimate
Lemma
If h(x , α) := φαα(x)eψ(|∇φ(x)|) attains a maximum at some point(x0, α0) among all (x , α) ∈ Rn × Sn−1, then
h(x0, α0) ≤ C with C = C (R , λ,Λ, λq) > 0.
R P
ψ(t)
Yash Jhaveri (UT Austin) Lipschitz Changes of Variables September 25, 2015 11 / 16
Proof (Sketch) of Theorem: Two Cases
1: Our auxiliary function h attains a maximum at some point(x0, α0) ∈ Rn × Sn−1.
2: Our auxiliary function h is not guaranteed to achieve a maximumin Rn × Sn−1.
Yash Jhaveri (UT Austin) Lipschitz Changes of Variables September 25, 2015 12 / 16
Proof (Sketch) of Theorem: Two Cases
1: Our auxiliary function h attains a maximum at some point(x0, α0) ∈ Rn × Sn−1.
2: Our auxiliary function h is not guaranteed to achieve a maximumin Rn × Sn−1.
Yash Jhaveri (UT Austin) Lipschitz Changes of Variables September 25, 2015 12 / 16
Proof (Sketch) of Theorem: Two Cases
1: Our auxiliary function h attains a maximum at some point(x0, α0) ∈ Rn × Sn−1.
2: Our auxiliary function h is not guaranteed to achieve a maximumin Rn × Sn−1.
Yash Jhaveri (UT Austin) Lipschitz Changes of Variables September 25, 2015 12 / 16
Proof (Sketch) of Theorem: Case 2
- First, we consider the approximating OT maps ∇φj = T j takinge−V to ecq,j−V
j+cq−q, where for j � R
V j(x) =
{V (x) x ∈ Bj(0)
+∞ x ∈ Bcj (0)
and cq,j > 0 are normalizing constants.
Yash Jhaveri (UT Austin) Lipschitz Changes of Variables September 25, 2015 13 / 16
Proof (Sketch) of Theorem: Case 2
- Second, we use that the functions
hjε(x , α) := (φj(x + εα) + φj(x − εα)− 2φj(x))eψ(|∇φj (x)|)
achieve maxima outside BR for all ε� 1.
- Third, we formalize our original heuristics replacing pure secondderivatives with second order difference quotients to conclude thatφjαα ≤
√Λ/λ.
- Finally, we pass to the limit via stability of OT.
Yash Jhaveri (UT Austin) Lipschitz Changes of Variables September 25, 2015 14 / 16
Proof (Sketch) of Theorem: Case 2
- Second, we use that the functions
hjε(x , α) := (φj(x + εα) + φj(x − εα)− 2φj(x))eψ(|∇φj (x)|)
achieve maxima outside BR for all ε� 1.
- Third, we formalize our original heuristics replacing pure secondderivatives with second order difference quotients to conclude thatφjαα ≤
√Λ/λ.
- Finally, we pass to the limit via stability of OT.
Yash Jhaveri (UT Austin) Lipschitz Changes of Variables September 25, 2015 14 / 16
Proof (Sketch) of Theorem: Case 2
- Second, we use that the functions
hjε(x , α) := (φj(x + εα) + φj(x − εα)− 2φj(x))eψ(|∇φj (x)|)
achieve maxima outside BR for all ε� 1.
- Third, we formalize our original heuristics replacing pure secondderivatives with second order difference quotients to conclude thatφjαα ≤
√Λ/λ.
- Finally, we pass to the limit via stability of OT.
Yash Jhaveri (UT Austin) Lipschitz Changes of Variables September 25, 2015 14 / 16
Other Theorems: 1-d and the Radially Symmetric Case
Theorem (Colombo, Figalli, -)
Let V : R→ R ∪ {∞} be a convex function and q be a boundedfunction such that e−V , e−V−q ∈ P(R). Then, the optimal transportT that takes e−V to e−V−q is Lipschitz and satisfies
‖ logT ′‖L∞(R) ≤ ‖q+‖L∞(R) + ‖q−‖L∞(R).
Theorem (Colombo, Figalli, -)
Let V : Rn → R ∪ {∞} be a convex, radially symmetric function andq be a bounded, radially symmetric function such thate−V , e−V−q ∈ P(Rn). Then, the optimal transport T that takes e−V
to e−V−q is Lipschitz: for a.e. x ∈ Rn
e−‖q+‖L∞(Rn)−‖q−‖L∞(Rn) Id ≤ ∇T (x) ≤ e‖q
+‖L∞(Rn)+‖q−‖L∞(Rn) Id .
Yash Jhaveri (UT Austin) Lipschitz Changes of Variables September 25, 2015 15 / 16
Other Theorems: 1-d and the Radially Symmetric Case
Theorem (Colombo, Figalli, -)
Let V : R→ R ∪ {∞} be a convex function and q be a boundedfunction such that e−V , e−V−q ∈ P(R). Then, the optimal transportT that takes e−V to e−V−q is Lipschitz and satisfies
‖ logT ′‖L∞(R) ≤ ‖q+‖L∞(R) + ‖q−‖L∞(R).
Theorem (Colombo, Figalli, -)
Let V : Rn → R ∪ {∞} be a convex, radially symmetric function andq be a bounded, radially symmetric function such thate−V , e−V−q ∈ P(Rn). Then, the optimal transport T that takes e−V
to e−V−q is Lipschitz: for a.e. x ∈ Rn
e−‖q+‖L∞(Rn)−‖q−‖L∞(Rn) Id ≤ ∇T (x) ≤ e‖q
+‖L∞(Rn)+‖q−‖L∞(Rn) Id .
Yash Jhaveri (UT Austin) Lipschitz Changes of Variables September 25, 2015 15 / 16
The End
Thank you.
Yash Jhaveri (UT Austin) Lipschitz Changes of Variables September 25, 2015 16 / 16