TRANSPORT TRANSFORM

AUTHORS,

1. intro
We consider the Monge Problem in 1-D. Let µ, ν be probability measures. We say
that a function T : R → R transport µ to ν if ν(A) = µ(T −1 (A)), and we call T a
transport map. We write ν = T# µ as a shorthand. The Monge problem is finding the
minimum among all transport maps from µ to ν of
Z
arg min |x − T (x)|2 dµ(x).
T : ν=T# µ
R

When µ = r0 (x)dx, and ν = p(x)dx, p, r0 ∈ L1 (R), by Brenier’s Theorem, there exists

a unique solution a.e. µ to Monge problem given by a function

n Zt Zx o
p̂(x) = sup t : p(ξ)dξ = r0 (ξ)dξ
∞ ∞

Remark 1.1.
(1) p̂ is monotone non-decreasing. Ivan: What monotone non decreasing means?
(2) p̂ is right continous.
(3) p̂ has countably many jump discontinuities, and these are the only discontinu-
ities of p̂.
(4) p̂ : R → R ∪ {−∞, ∞}
Some of the ideas are available in the pictures page.
Some observations about p̂ :
Observation 1.
(1) If p has an isolated zero, then p̂ has infinite derivative.
(2) If p has an interval of zeros, then p̂ has a jump
(3) If r has an isolated zero or an interval of zeros, then p̂=0.
We reach for pictures here.
Definition 1.2. A function f ∈ L1 (R) is non decreasing on R with respect to µ if
there exists a non-decreasing function g on R such that f = g µ-almost everywhere.
We define
I(µ) = {f ∈ L1 (R) : f is non decreasing respect to µ}.
2010 Mathematics Subject Classification.
Key words and phrases.
1
2 AUTHORS,

Ivan: Maybe we should mention something about the class of equivalence of function.
h∼h
Claim 1.3. Let h be an increasing function with respect to µ , the there exists e
1
and p ∈ L such that p̂ = h.
e

Future questions:
1. Trying approximations by these functions to generalize to a bigger class.
2. Can we get rid of non negativity conditions of p and r.
Theorem 1.4. If P is the space of probability density functions and I(µ) is the set of
non decreasing functions with respect to µ. Let
b : P → I(µ)
p → pb
be the transport transform. Then, it is a bijection.
We expand the domain of this transform to include all L1 -functions and not only
probability density functions.
Definition 1.5. Given f ∈ L1 (R) we define the transport transform
b : L1 (R) → I(µ) × I(µ) × R2
 + b  − b !
f f
f→ , , kf + k, kf − k .
kf + k kf − k
Ivan: Maybe I would separate the transform and norm parts as the direct sum of two
maps to work more comfortably.
Claim 1.6. The transport transform in the definition above is biyective.
Proposition 1.7. (Properties of the transport transform) Let a ∈ R and f ∈ L1 (R),
we denote
(τa f )(x) = f (x − a)
(Da f )(x) = af (ax).
Then,
(τa f )b = fb + (a, a, 0, 0)
 + b  − b !
1 f 1 f
(Da f )b = +
, −
, kf + k, kf − k .
a kf k a kf k

2. Defining the transport for measures in compact notation

Ivan: Add definition of cumulative distribution function(CDF) and of generalized
inverse. Remark about choosing bigger or bigger equal in the definitions. Properties
of CDF and of inverse. Right continuity, inverse of the inverse, composition of CDF
and generalized inverse inequalities.
 TRANSPORT TRANSFORM 3

Let µ, ν be such that µ is transported to ν. Ifµ does not give mass to atoms, then
the transport map is given by,
T = (Fν )−1 (Fµ )
Let dµ = r(x)dx,dν = p(x)dx
Proposition 2.1. (Properties of the transport transform) Let a ∈ R and f ∈ L1 (R),
we denote
1.ˆ(τa P ) = p̂ + a2.ˆ(Da )) = v̂/a.
\ −1 −1
Proposition 2.2. (τ a P ) = Fτa ◦ Fµ = Fν ◦ Fµ + a. = P + a This holds in general
b
even if P is not a pdf.
Proof. Claim: Ivan: This claim should be proved without using pdf just with
the definition of translation of a measureFτa ν (x) = τa Fν (x).
Rx
Fτa ν (x) = P (t − a)dt, using change of variable t − a = y we get
−∞
Zx Zx−a
P (t − a)dt = P (y)dy = τa Fν .
−∞ −∞
Now we want to show Fτ−1
aν
(x) = −1
(Fν + a)(x)

Fτ−1
aν
(x) = inf{t : Fτa ν (t) > x}
= inf{t : Fν (t − a) > x}
= inf{s + a : Fν (s) > x}
= inf{s : Fν (s) > x} + a
= Fν−1 (x) + a.

Dilations:

Da ν := da] (ν),
y
where da (x) = ax and d−1
a (y) = a
.
Zx
ν((−∞, x]) = P (s)ds.
−∞
Therefore,
x
Za
Da ν((−∞, x]) = ν(d−1
a (−∞, x]) = P (s)ds,
−∞
s = ta =⇒ ds = adt, therefore
4 AUTHORS,

Zt Zt
Da ν((−∞, x]) = P (ta)adt = a P (ta).dt
−∞ −∞
1
d(Da ν) = aP (ax)dx = a
P ( xa )dx, dν = P (x)dx.
−1
Proposition 2.3. D a ν = (Fν ◦ Fµ )a = ab
ν.
d

Proof.
FDa ν (x) = Da ν[−∞, x]
h x i
= ν −∞,
x a
= Fν .
a

FD−1a ν (x) = inf{t : FDa ν (t) > x}

 
t
= inf{t : Fν > x}
a
t
= inf{as : Fν (s) > x}, =s
a
= a inf{s : Fν (s) > x}
= aFν−1 (x).

Case when µ, ν 6= 0 and ν(R) 6= 1, µ(R) 6= 1 i.e ν, µ both are not probability
measures.
∂ν
∂Nν = ν(R) and FNν (x) = Fν(R)
ν (x)

FN−1ν (x) = inf{t ∈ R : FNν (x) > t}

= inf{t ∈ R : Fν (x) > tν(R)}
= Fν−1 (tν(R))
νb = Fν−1 (ν(R) · Fµ )
Fν : R → [0, ν(R)]
Fν−1 : [0, ν(R)] → R
Fµ : R → [0, 1]
ν(R) · Fµ: R → [0, ν(R)]
ν(R)
νb = Fν−1 µ(R) Fµ
(Since we are talking both µ and ν are not pdfs and therefore normalizing
both).
Remark 2.4. νb is the transport between the normalized µ and ν i.e. Nµ → Nν .
 TRANSPORT TRANSFORM 5

Proposition 2.5. (Translation  and dilation

 of a generalmeasure)
−1 Da ν(R) ν(R)
Dilation: (Da ν) = FDa ν µ(R) F µ = aFν−1 µ(R)
\ F µ ab
ν.
   
−1 τb ν(R) ν(R)
Transalation τcb ν = F τb ν µ(R)
Fµ = Fν−1 µ(R) Fµ + b = νb + b.

3. 0
Acknowledgment