-
arX
iv:2
010.
0013
5v3
[m
ath.
FA]
10
Jan
2021
Blaschke–Santaló inequality for many functions and geodesic
barycenters of measures ∗
Alexander V. Kolesnikov†and Elisabeth M. Werner
‡
Abstract
Motivated by the geodesic barycenter problem from optimal
transportation theory,we prove a natural generalization of the
Blaschke–Santaló inequality and the affineisoperimetric
inequalities for many sets and many functions. We derive from it
anentropy bound for the total Kantorovich cost appearing in the
barycenter problem.We also establish a “pointwise Prékopa–Leindler
inequality” and show a monotonicityproperty of the multimarginal
Blaschke–Santaó functional.
1 Introduction
The Blaschke–Santaló inequality, see [4, 36], states that every
0-symmetric convex bodyK in Rn satisfies
voln(K)voln(K◦) ≤ (voln(Bn2 ))2,
where K◦ = {y ∈ Rn : 〈x, y〉 ≤ 1 ∀x ∈ K} is the polar body of K,
Bn2 = {x ∈ Rn : |x| ≤ 1}is the Euclidean unit ball and | · |
denotes the Euclidean norm on Rn. The left-hand sideof this
inequality is called the Mahler volume. The sharp lower bound for
the Mahlervolume is still open in dimensions 4 and higher. The
famous Mahler conjecture suggeststhat this functional is minimized
by the couple (Bn1 , B
n∞). Partial results can be found in
e.g., [23, 27, 33, 34].Here we ask: What is a natural
generalization of the bounds for the Mahler volume
for multiple sets? While this is not obvious from the geometric
viewpoint, we suggest inthis paper a reasonable extension, which is
naturally related to a functional counterpartof the
Blaschke–Santaló inequality.
The functional Blaschke–Santaló inequality was discovered by K.
Ball [6] and laterextended and generalized in [3], [20], [30]. In
its simplest form it states that for every twomeasurable even
functions V,W on Rn we have that
∫
e−V (x)dx
∫
e−W (y)dy ≤ (2π)n,
provided that V (x) +W (y) ≥ 〈x, y〉. Equality is attained if and
only if V (x) = |x|2
2 +
c,W (y) = |y|2
2 − c, where c is a constant. An interesting link to optimal
transportationtheory was discovered recently by M. Fathi [17]. He
showed that for probability measures
∗Keywords: Blaschke–Santaló inequality, optimal transport,
multimarginal Monge–Kantorovitch prob-
lem, geodesic barycenters of measures, Prekopa-Leindler
inequality, Kähler–Einstein equation. 2020 Math-
ematics Subject Classification: 52A20, 52A40, 60B.†Supported by
RFBR project 20-01-00432; Section 7 results have been obtained with
support of RSF
grant No 19-71-30020. The article was prepared within the
framework of the HSE University Basic Research
Program‡Partially supported by NSF grant DMS-1811146 and by a
Simons Fellowship
1
http://arxiv.org/abs/2010.00135v3
-
µ = f · γ, ν = g · γ, where γ is the standard Gaussian measure,
such that∫
xfdγ = 0, thefollowing inequality holds,
1
2W 22 (µ, ν) ≤ Entγ(µ) + Entγ(ν) (1.1)
and that this inequality is equivalent to the functional
Blaschke–Santaló inequality. Here,W 22 (µ, ν) is the L
2 Kantorovitch distance (see Section 2 for the definition)
and
Entγ(µ) =
∫
f log fdγ
is the relative entropy with respect to Gaussian measure.
Inequality (1.1) is a remarkablestrengthening of the Talagrand
transportation inequality and the starting point of ourpaper. We
refer to, e.g., [5] for Talagrand’s inequality and it’s fundamental
importancein probability theory. In this context, please also note
a very recent result of N. Gozlanabout a transportational approach
to the lower bound for the Blaschke–Santaló functional[26].
We would like to point out an important connection of the
Blaschke–Santaló inequalityto the Kähler–Einstein equation.
Fathi’s result establishes, in particular, that the func-tional f →
12W 22 (µ, ν)−Entγ(µ) is bounded from below. The minimum of this
functionalsolves the so-called Kähler–Einstein equation. This was
established by F. Santambrogio[35]. The form of the functional
presented here was considered in [29]. The well-posednessof the
Kähler–Einstein equation was proved by D. Cordero-Erausquin and B.
Klartag [14].Generalization to the sphere and relations to the
logarithmic Minkowski problem wereestablished in [28]. Other
related transportation inequalities can be found in [19].
To analyze the case of k functions with k > 2 we consider the
cost functional
c(x1, · · · , xn) =∑
i,j=1,i
-
1.1 The main results
In Section 2 we discuss some preliminary facts about Kantorovich
duality theory for manyfunctions (Legendre-type transforms) and
prove that our integral functional is boundedfor the case of
quadratic cost (1.2). We also show that for k > 2 our functional
has atrivial (zero) lower bound, unlike the case of two
functions.
In Section 3 we verify the above conjecture in the unconditional
case (see Section 3 forthe definition) and prove the following
theorem.
Theorem 3.1 Let fi : Rn → R+, 1 ≤ i ≤ k, be unconditional
functions satisfying
k∏
i=1
fi(xi) ≤ ρ(
k∑
i,j=1i
-
In particular, if ρ(t) = e−t
k−1 , then, if∑k
i=1,i
-
Proposition 2.1. Let Vi, 1 ≤ i ≤ k, be a family of Borel
functions on Rn. Then thefunctional
S(V1, · · · , Vk) =k∏
i=1
∫
e−Vi(xi) dxi
is bounded on the set
Ln,k ={
(V1, · · · , Vk) : Vi is even ∀i ∈ {1, · · · , k},k∑
i=1
Vi(xi) ≥k∑
i,j=1,i
-
Finally, we show that S is bounded. We observe that for every j
6= m
Vm(xm) ≥ supxi,i 6=m
(
k∑
i,j=1,i
-
At the end of this section we recall basic facts on duality
relations for the transportationcost appearing in the theory of
barycenters of measures. Recall that for a given familyof
probability measures µ1, · · · , µk and weights λi ∈ [0, 1]
satisfying
∑ki=1 λi = 1 its
barycenter µ is the minimum point of the functional
F(ν) = 12
k∑
i=1
λiW22 (µi, ν).
Here,
W 22 (ν1, ν2) = inf
{∫
|x− y|2dP : P ∈ P(Rn × Rn), P (·,Rn) = ν1, P (Rn, ·) = ν2}
is the L2 Kantorovitch distance of probability measures ν1, ν2.
It is well-known that thebarycenter problem is closely related to
the multimarginal (maximization) Kantorovichproblem with the cost
function
k∑
i,j=1,i 6=j
λiλj〈xi, xj〉
and marginals µi. Let π be the solution to this problem, i.e. a
measure that gives amaximum to the functional
P →∫ k∑
i,j,i 6=j
λiλj〈xi, xj〉dP (2.1)
among the measures on (Rn)k having µ1, . . . µk as
marginals.
The following facts are collected from [1] and [24].
Theorem 2.4. [1], [24] Assume that µi are absolutely continuous
measures with finitesecond moments and λi ∈ (0, 1) are numbers
satisfying
∑ki=1 λi = 1. Then the following
facts hold.
1. There exists a unique absolutely continuous solution µ to the
barycenter problem anda unique solution π to the problem (2.1).
2. The measure µ is the push-forward measure of π under the
mapping T (x1, · · · , xk) =∑k
i=1 λixi and the following relation holds:
k∑
i=1
λiW22 (µi, µ) =
∫ k∑
i=1
|xi − T (x)|2dπ.
3. The optimal transportation mappings ∇Φi of µ onto µi
satisfy
k∑
i=1
λi∇Φi(x) = x
for µ-a.e. x. and π is supported on the set (∇Φ1(x), · · ·
,∇Φk(x)).
7
-
4. There exists a unique (up to addition of constants and sets
of measure zero) dualsolution vi to the problem (2.1), i.e. a
k-tuple of functions satisfying
k∑
i=1
vi(xi) ≥k∑
i,j,i 6=j
λiλj〈xi, xj〉
with equality π-a.e. The following relation holds between vi and
Φi:
Φ∗i (xi) = λi|xi|22
+vi(xi)
λi+ Ci (2.2)
for µi-almost all xi.
Remark 2.5. The results of item 1. are obtained in Section 3 of
[1], item 2. is containedin Proposition 4.2 [1], item 3.
corresponds to Proposition 3.8 of [1]. Formula (2.2) needssome
explanations. It corresponds to formula (4.8) in [1], but in the
presentation in [1]there is no direct link to the optimal
transportation of the barycenter µ onto µi. Let usgive some
informal explanations.By the Kantorovich duality π is concentrated
of the zero set of the positive function
k∑
i=1
vi(xi)−k∑
i,j,i 6=j
λiλj〈xi, xj〉.
Thus for π-a.e. (x1, · · · , xk) and all 1 ≤ i ≤ k one has
∇vi(xi) =∑k
j 6=i λiλjxj. Equiva-lently,
λixi +∇vi(xi)λi
=
k∑
j=1
λjxj, (2.3)
π-a.e. It remains to note that µ is the image of π under T
=∑k
j=1 λjxj and µi is theprojection of π onto the i-th factor. Thus
relation (2.3) immediately implies that µ is the
image of µi under the mapping xi → λixi + ∇vi(xi)λi . Since the
latter is the gradient of theconvex function λi
|xi|2
2 +vi(xi)λi
, we conclude by uniqueness of the optimal transportaion
mapping that ∇Φ∗i = λixi +∇vi(xi)λi
.
3 The unconditional case
In this section we verify our conjecture (inequality part) for
the unconditional functions.A function f : Rn → R is called
unconditional, if
f(ε1x1, · · · , εnxn) = f(x1, x2, . . . , xn),
for every (ε1, · · · , εn) ∈ {−1, 1}n and every (x1, · · · , xn)
∈ Rn. We show that the conjec-ture is true in the unconditional
case.
Theorem 3.1. Let fi : Rn → R+, 1 ≤ i ≤ k, be measurable
unconditional functions
satisfyingk∏
i=1
fi(xi) ≤ ρ
k∑
i,j=1,i
-
where ρ is a positive non-increasing function such that∫
Rρ
1k (t2)dt
-
below. We get
(
k∏
i=1
∫
Rn+
fi(xi)dxi
)1k ≤
∫
Rn
supt= 1
k
∑ki=1 ti
k∏
i=1
(
f1k
i (eti)e
1k
∑nm=1(ti)m
)
dt
≤∫
Rn
supt= 1
k
∑ki=1 ti
[
ρ1k
(
k∑
i,j=1,i
-
and equality holds for λ 6= 0, if and only if there exists a ∈ R
and a positive definite matrixA such that V (x) = 〈Ax, x〉 + a, for
every x ∈ Rn. For λ = 0, equality holds trivially.
Remark. Theorem 3.2 (iii) is just a special case for λ = 12 of a
more general statementproved in [12].
We then get the following Proposition.
Proposition 3.3. Let Vi : Rn → R∪∞, 1 ≤ i ≤ k, be convex
unconditional functions and
let ρ is a positive non-increasing function such that∫
Rρ
1k (t2)dt
-
Proof. (i) We get immediately from Theorem 3.1 and inequality
(3.3) that for λ ∈ [0, 12 ],
k∏
i=1
asλ(Vi) ≤ (2π)knλ(
k∏
i=1
∫
e−Vi
)k(1−2λ)
≤ (2π)knλ(∫
Rn
ρ1k
(k(k − 1)2
|u|2)
du
)k(1−2λ)
.
If ρ(t) = e−t
k−1 , then
(2π)knλ(∫
Rn
ρ1k
(k(k − 1)2
|u|2)
du
)k(1−2λ)
= (2π)kn2 =
(
asλ
( | · |2
))k
,
which shows the second part of (i).
(ii) We use Theorem 3.2 (iii) and (iv) and Theorem 3.1 and get
that for λ ∈ [12 , 1],
k∏
i=1
asλ(Vi) =
k∏
i=1
as1−λ(V∗i ) ≤ (2π)kn(1−λ)
(
k∏
i=1
∫
e−V∗i
)k(2λ−1)
≤ (2π)kn(1−λ)(∫
Rn
ρ1k
(k(k − 1)2
|u|2)
du
)k(2λ−1)
.
The second part for ρ(t) = e−t
k−1 follows.
The equality characterizations follow immediately from the
results in the next section.
Remark 3.4. (i) Please note that for λ = 0, inequalities (3.4)
and (3.5) are just theinequalities of Theorem 3.1. For λ = 12 , we
do not need that the Vi are unconditional andthe inequalities are
just the inequalities of (3.3),
k∏
i=1
as 12(Vi) ≤ (2π)knλ.
See also Section 8 for more on as 12.
(ii) For λ > 1, we get an estimate from below with an
absolute constant c, see [12],
k∏
i=1
asλ(Vi) ≥ cknλ(∫
Rn
ρ1k
(k(k − 1)2
|u|2)
du
)k(1−2λ)
.
4 Characterization of the equality cases
In the proof of Theorem 3.1 we have used the Prékopa–Leindler
inequality which is aparticular case of the more general
Brascamp–Lieb inequality (see [10], [7]). To analyzethe equality
case we need the equality characterizations of the
Prékopa–Leindler inequality.We could not find those in the
literature, except in the case of two functions, establishedby
Dubuc [16]. We therefore give a proof of the equality
characterization.
Theorem 4.1 (Prékopa–Leindler). Let fi, 1 ≤ i ≤ k, and h be
nonnegative integrable realfunctions on Rn such that for all xi,
for all λi, 1 ≤ λ ≤ k, with
∑ki=1 λ1 = 1,
h
(
k∑
i=1
λixi
)
≥k∏
i=1
fλii (xi).
12
-
Thenk∏
i=1
(
∫
Rn
fidxi
)λi ≤∫
Rn
hdx. (4.1)
Equality holds in the Prekopa–Leindler inequality if and only if
there exist vectors y1, · · · , yksuch that for all x
f1(x− y1)∫
Rnf1dx
=f2(x− y2)∫
Rnf2dx
= · · · fk(x− yk)∫Rnfkdx
= e−ψ(x), (4.2)
where ψ is a convex function such that∫
R⋉e−ψ(x)dx = 1 and
h(x) = supx=
∑ki=1 λi xi
k∏
i=1
fλii (xi) =k∏
i=1
(
∫
Rn
fidxi
)λie−ψ
(
x+∑k
i=1 λiyi
)
.
Proof. It is clear that equality holds in inequality (4.1), if
the functions satisfy the condi-tion (4.2).The proof of the
inequality is well known and can be found in e.g., [25, 36]. We
give aproof of the inequality by induction on the number of
functions. This allows to establishthe equality characterizations,
as for two functions, those were established by Dubuc [16].We
have
supx=
∑ki=1 λi xi
k∏
i=1
fλii (xi) = supλ1x1+(1−λ1) y
fλ11 (x1)g1−λ1(y),
where
g(y) = supy= 1
1−λ1
∑ki=2 λi xi
k∏
i=2
f
λi1−λ1i (xi).
Applying the Prékopa–Leindler inequality for two functions
gives
∫
supx=
∑ki=1 λi xi
k∏
i=1
fλii (xi) ≥(∫
f1dx1
)λ1 (∫
gdy
)1−λ1
.
Applying the induction step, one gets
∫
gdy ≥k∏
i=2
(∫
fi(xi)dxi
)
λi1−λ1
.
This completes the proof of the inequality. The equality
characterization follows from theequality characterization for two
functions.
Theorem 4.2. Let fi : Rn → R+, 1 ≤ i ≤ k, be measurable
unconditional functions
satisfyingk∏
i=1
fi(xi) ≤ ρ
k∑
i,j=1,i
-
if and only if for all 1 ≤ i ≤ k, fi is log-concave and for
almost all x ∈ Rn,
fi(x) = ρ1k
(
k(k − 1)2
|x|2)
.
In particular, ρ is log-concave.
Proof. Suppose that all fi are log-concave and a.e. equal to
ρ1k
(
k(k−1)2 |x|2
)
. Then it is
obvious that equality holds in (3.1). Moreover, as then ρ is
also log-concave,
k∏
i=1
fi(xi) =k∏
i=1
ρ1k
(
k − 12
|xi|2)
≤ ρ(
k − 12
k∑
i=1
|xi|2)
≤ ρ
k∑
i,j=1,i
-
The last identity holds as yi + zi = x for all i. Therefore,
with (4.4),
e−ψ(x) =f1(x− y0)∫
Rnf1dx
=f2(x− y0)∫
Rnf2dx
= · · · fk(x− y0)∫Rnfkdx
=ρ
1k
(
k(k−1)2 |x− y0|2
)
∫
Rnρ
1k
(
k(k−1)2 |x|2
)
dx
or
fi(x)∫
Rnfidx
=ρ
1k
(
k(k−1)2 |x|2
)
∫
Rnρ
1k
(
k(k−1)2 |x|2
)
dx,
for all i.
5 The Blaschke–Santaló inequality and affine isoperimetric
inequality for many sets
The classical Blaschke–Santaló inequality for symmetric sets
can be stated in the followingway,
∫
Sn−1fndx
∫
Sn−1gndy ≤ n2 (voln(Bn2 ))2 =
(
voln−1(Sn−1)
)2,
where f, g are positive symmetric functions on Sn−1
satisfying
f(x)g(y) ≤ 1〈x, y〉+,
and where for a ∈ R, a+ = max{a, 0}. The latter inequality is
satisfied, in particular, if
f(x) = rK(x), g(y) =1
hK(y)= rK◦(y),
where rK(x) = max{λ ≥ 0 : λx ∈ K} is the radial function of the
convex body K,hK(y) = sup{〈x, y〉 : x ∈ K} is the support function
of K and where for a 0-symmetricconvex body K,
K◦ = {y ∈ Rn : 〈y, x〉 ≤ 1∀x ∈ K}is the polar body of K. We can
then write the above as follows,
voln(K1) voln(K2) ≤ (voln(Bn2 ))2,
provided〈x, y〉 ≤ 1, ∀x ∈ K1, ∀y ∈ K2. (5.1)
We now prove a Blaschke–Santaló inequality for multiple sets.
We recall that a subset Kin Rn is unconditional if its
characteristic function 1K is unconditional.
Theorem 5.1. Let Ki, 1 ≤ i ≤ k, be unconditional convex bodies
in Rn such that
k∏
i=1
e− 1
2‖xi‖
2Ki ≤ ρ
k∑
i=1,i
-
where ρ is a positive non-increasing function such that∫
Rρ
1k (t2)dt
-
where µK the usual surface area measure on ∂K, the boundary of
K, N(x) is the outerunit normal at x ∈ ∂K and κ(x) is the
generalized Gauss curvature at x ∈ ∂K. Note thatas0(K) = n voln(K),
and if K is C
2+, then as±∞(K) = n voln(K
◦).The Lp-affine isoperimetric inequalities state that for 0 ≤ p
≤ ∞,
asp(K)
asp(Bn2 )
≤(
voln(K)
voln(Bn2 )
)n−pn+p
(5.3)
and for −n < p ≤ 0,asp(K)
asp(Bn2 )
≥(
voln(K)
voln(Bn2 )
)n−pn+p
.
Equality holds trivially if p = 0. In both cases equality holds
for p 6= 0 if and only if K isan ellipsoid. If −∞ ≤ p < −n and K
is C2+, then
cnpn+p
(
voln(K)
voln(Bn2 )
)n−pn+p
≤ asp(K)asp(Bn2 )
. (5.4)
These inequalities were proved by Lutwak [32] for p > 1 and
for all other p by Werner andYe [39]. The case p = 1 is the
classical case.
Theorem 5.1 leads to a multi-set “affine” isoperimetric
inequality.
Proposition 5.2. Let Ki, 1 ≤ i ≤ k, be unconditional convex
bodies in Rn such that
k∏
i=1
e− 1
2‖xi‖
2Ki ≤ ρ
k∑
i=1,i
-
Proof. Let 0 ≤ p ≤ n. By the affine isoperimetric inequality and
Theorem 5.1 we getk∏
i=1
asp(Ki) ≤ (asp(Bn2 ))kk∏
i=1
(
voln(Ki)
voln(Bn2 )
)n−pn+p
≤ (asp(Bn2 ))k(
1
(2π)n2
∫
Rn
ρ1k
(k(k − 1)2
|x|2)
dx
)k
The first inequality shows that for p = n,
k∏
i=1
asn(Ki) ≤ (asn(Bn2 ))k .
If ρ(t) = e−t
k−1 , then we have for all 0 ≤ p ≤ n,k∏
i=1
asp(Ki) ≤ (asp(Bn2 ))kk∏
i=1
(
voln(Ki)
voln(Bn2 )
)n−pn+p
≤ (asp(Bn2 ))k
The equality characterizations follow from Section 4 and the
equality characterizations ofthe above affine isoperimetric
inequalities.
Remark 5.3. (i) For p = n, the inequality is just the affine
isoperimetric inequality (5.3).As as0(K) = nvoln(K), the
inequalities of the theorem for p = 0 are just the inequalitiesof
Theorem 5.1.
(ii) The corresponding inequalities for −∞ ≤ p < −n also
hold, using (5.4).(iii) It was shown in [32, 37] that for all p 6=
−n and for all invertible linear transforma-tions T : Rn → Rn
asp(T (K)) = |det(T )|n−pn+p asp(K). (5.7)
For p = 1, (5.7) even holds for affine T . Therefore we get that
for all invertible lineartransformations Ti : R
n → Rn,k∏
i=1
asp(Ti(Ki)) =
k∏
i=1
|det(Ti)|n−pn+p asp(Ki).
Hence∏ki=1 asp(Ki) and thus inequality (5.5) are invariant under
linear transformations
with determinant 1 and for p = 1 under affine transformations
with determinant 1 andfor p = n invariant under invertible linear
transformations.
A further multiple set version of the Blaschke–Santaló
inequality is given in the nextproposition.
Proposition 5.4. Let Ki, 1 ≤ i ≤ k, be unconditional convex
bodies in Rn with radialfunctions ri = rKi. Assume that for all xi
= ((xi)1, · · · , (xi)n) ∈ Rn,
k∏
i=1
ri(xi) ≤1
(
∑nj=1
(
|(x1)j|1k · · · |(xk)j |
1k
)2)
k2
. (5.8)
Thenk∏
i=1
voln(Ki) ≤ (voln(Bn2 ))k .
Equality holds if and only if Ki is a Euclidean ball for all 1 ≤
i ≤ k.
18
-
Proof. Let m ∈ R, 1 ≤ m < n and put xi = eti . Set w =
1k∑k
i=1 ti. Then
k∏
i=1
rmi (eti)1{|eti |≤1} e
∑i,j(ti)j ≤
1{|ew|≤1} e∑
i,j(ti)j
(
∑nj=1 e
2wj)km
2
=1{|ew|≤1} e
k∑n
j=1(w)j
(
∑nj=1 e
2wj)km
2
. (5.9)
We now apply again the change of variables xi = eti , 1 ≤ i ≤ k,
the Prékopa–Leindler
inequality and (5.9),
(
k∏
i=1
∫
Bn2 ∩Rn+
rmi dxi
)
1k
=
(
k∏
i=1
∫
Rn
rmi (eti)1{|eti |≤1} e
∑j(ti)j dti
)
1k
≤∫
Rn
supw= 1
k
∑ki=1 ti
[
k∏
i=1
rmk
i (eti)1{|eti |≤1} e
1k
∑i,j(ti)j
]
dw
≤∫
Rn
1{|ew|≤1} e∑n
j=1(w)j
(
∑nj=1 e
2wj)
m2
dw =
∫
Bn2 ∩Rn+
dx
|x|m .
Hence by symmetry(
k∏
i=1
∫
Bn2
rmi dxi
)
1k
≤∫
Bn2
dx
|x|m .
Next we observe that every radial function ri satisfies
ri(xi) = ri
(
xi
|xi|
)
1
|xi|.
For every 1 ≤ m < n, m ∈ R, we introduce the finite
probability measure dµm =1Bn2
(u)∫Bn2
du|u|m
du|u|m . The inequality above can then be rewritten as
follows,
k∏
i=1
∫
Bn2
rmi
(
xi
|xi|
)
dµm ≤ 1.
Since µm is rotational invariant, the above inequality can be
rewritten as
k∏
i=1
∫
Sn−1rmi (θ) dσ(θ) ≤ σ(Sn−1)k, (5.10)
where σ is the (n − 1)-dimensional Hausdorff measure. Passing to
the limit m → n andapplying the Fatou’s Lemma one gets that (5.10)
holds for m = n. On the other hand,for m = n one has for all i
∫
Sn−1rni (θ) dσ(θ) = σ(S
n−1)voln(Ki)
voln(Bn2 ). (5.11)
From this we derive the desired estimate.Now we address the
equality characterizations. If equality holds in the inequality,
thenequality holds everywhere and we get with (5.10) for m = 1,
k∏
i=1
∫
Bn2 ∩Rn
ri dxi =k∏
i=1
∫
Sn−1ri(θ) dσ(θ) = σ(S
n−1)k. (5.12)
19
-
Equality holds in the Prékopa–Leindler inequality. Thus there
exist a convex function ψand y1, · · · , yk such that for all
xr1(x− y1)1Bn2 (x− y1)
∫
Rn∩Bn2r1dx
=r2(x− y2)1Bn2 (x− y2)
∫
Rn∩Bn2r2dx
= · · ·rk(x− yk)1Bn2 (x− yk)
∫
Rn∩Bn2rkdx
= e−ψ(x).
(5.13)As Ki is in particular 0-symmetric, we have that ri(x −
yi) = 1‖x−yi‖Ki . We put Ri =∫
Bn2 ∩Rn ri dxi and thus we get for all i, j
Ri‖x− yi‖Ki1Bn2 (x− yj) = Rj‖x− yj‖Kj1Bn2 (x− yi).
We let x = yi in this inequality and get for all j 6= i,
0 = ‖yi − yj‖Kjor y1 = y2 = · · · = yk = y0. With z = x− y0, we
get by (5.13) for all z ∈ Bn2 ,
r1(z)
R1=r2(z)
R2= · · · rk(z)
Rk= e−ψ(z+y0). (5.14)
With (5.12)
k∏
i=1
r1k
i (z) =
(
k∏
i=1
Ri
)
1k
e−1k
∑ki=1 ψ(z+y0) = σ(Sn−1)e−ψ(z+y0).
Now we use (5.8) and get that∏ki=1 r
1k
i (z) =1|z| . Hence we have for all i,
1
σ(Sn−1) |z| = e−ψ(z+y0) =
ri(z)
Ri,
which means that Ki =Ri
σ(Sn−1)Bn2 for all i.
6 Prékopa–Leindler and displacement convexity inequali-
ties: refinement of the transportational argument
In this section we recall the transportational arguments of F.
Barthe [7] in his proof ofthe reverse Brascamp–Lieb inequality. We
show that the use of barycenters gives certainrefinements of the
Prekopa–Leindler inequality.
Let fi, 1 ≤ i ≤ k, be nonnegative integrable functions and λi ∈
[0, 1] be numbers suchthat
∑ki=1 λi = 1, and let dµ = p(x)dx be a probability measure. For
every i, ∇Φi is the
optimal transportation mapping that pushes forward µ onto
µi.
In what follows we apply the change of variables formula for the
optimal transportationmapping. In that form it was established by
R. McCann (see [38]),
p(x) =fi(∇Φi)∫
fidxidetD2aΦi(x),
where D2aΦi is the absolutely continuous part of the
distributional Hessian D2Φi of Φi.
This formula holds almost everywhere with respect to Lebesgue
measure. We will alsoapply below the following results
20
-
• The arithmetic-geometric mean inequality
k∏
i=1
(detAi)λi ≤ det
(
k∑
i=1
λiAi
)
,
where the Ai are symmetric nonnegative matrices, λi ≥ 0,∑k
i=1 λi = 1.
• The inequality between the distributional Hessian and its
absolutely continuous part
D2aΦi ≤ D2Φi.
First, we get by the arithmetic-geometric mean inequality
p(x) =
k∏
i=1
(
fi(∇Φi)(x)∫
fidxidetD2aΦi(x)
)λi
≤k∏
i=1
(
fi(∇Φi)(x)∫
fidxi
)λi
det
(
k∑
i=1
λiD2aΦi(x)
)
≤ sup{yi:∑i λiyi=
∑i λi∇Φi(x)}
k∏
i=1
(
fi(yi)∫
fidxi
)λi
det
(
k∑
i=1
λiD2aΦi(x)
)
. (6.1)
In the proof of Barthe one fixes an arbitrary measure µ and
integrates inequality (6.1).By the change of variables y =
∑
i λi∇Φi(x), we get the Prékopa–Leindler inequality
k∏
i=1
(∫
fidxi
)λi
≤∫
sup{yi:
∑i λiyi=y}
k∏
i=1
fλii (yi)dy.
If instead of an arbitrary measure µ, we apply this result to
the barycenter of the µ′is, weobtain the following pointwise
refinement of the Prékopa–Leindler inequality.
Theorem 6.1. (Pointwise Prékopa–Leindler inequality) Let µ be
the barycenter ofthe µi with weights λi. Then it has a density p
satisfying
k∏
i=1
(∫
fidxi
)λi
p(x) ≤ supx=
∑ki=1 λiyi
k∏
i=1
fλii (yi) (6.2)
Proof. By Theorem 2.4, 3. and the arithmetric-geometric mean
inequality one has
k∏
i=1
(
detD2aΦi(x))λi
≤ det(
k∑
i=1
λiD2aΦi(x)
)
.
Since∑
i=1 λiΦi(x) =|x|2
2 for p(x) dx-almost all x, then
k∑
i=1
λiD2aΦi(x) ≤ D2
(
∑
i=1
λiΦi(x))
= I
p(x) dx-a.e. Using this inequality and inequality (6.1) one gets
the result.
Let us rewrite (6.1) in terms of the standard Gaussian reference
measure dγ = e−
|x|2
2
(2π)n2dx.
21
-
Corollary 6.2. Let fidxi = ρi · dγ be probability measures and
dµ = ρ · dγ. Then µ-a.e.
ρ(x) e12
∑ki=1 λi|∇Φi(x)−x|
2 ≤k∏
i=1
ρλii (∇Φi). (6.3)
Proof. Applying the first inequality of (6.1) to fi = ρie−
|x|2
2
(2π)n2
and p = ρ e−
|x|2
2
(2π)n2, we get
ρ(x) e−|x|2
2 ≤k∏
i=1
ρλii (∇Φi)e−λi|∇Φi|
2
2 .
Also using Theorem 2.4, 3. we finally observe that
k∑
i=1
λi
( |∇Φi(x)|22
− |x|2
2
)
=
k∑
i=1
λi
( |∇Φi(x)|22
− |x|2
2− 〈∇Φi(x)− x, x〉
)
=1
2
k∑
i=1
λi|∇Φi(x)− x|2.
Integrating pointwise inequality (6.2) we get the
Prékopa–Leindler inequality. Takinglogarithm of (6.3) and
integrating we get the displacement convexity property of
theGaussian entropy,
Entγ(µ) +1
2
k∑
i=1
λiW22 (µ, µi) ≤
k∑
i=1
λiEntγ(µi). (6.4)
This result was proved in [1].
Mimicking the arguments that were used in the proof of (6.1),
leads to the following result.
Theorem 6.3. Let fi, 1 ≤ i ≤ k, be integrable functions
satisfyingk∏
i=1
fλii (xi) ≤ g(
k∑
i=1
λixi
)
, (6.5)
where λi ∈ [0, 1],∑k
i=1 λi = 1 and g is a nonegative function. Then for ρdx-almost
all x,
k∏
i=1
(∫
fidxi
)λi
ρ(x) ≤ g(x), (6.6)
where ρ(x)dx is the barycenter of the measures fi∫fidxi
dxi with weights λi.
Proof. Applying inequality (6.1) and the relation∑k
i=1 λi∇Φi(x) = x one immediatelygets
k∏
i=1
(
∫
fi(xi)dxi
)λiρ(x) ≤ sup
{yi:∑i λiyi=x}
k∏
i=1
fλii (yi) det(
k∑
i=1
λiD2Φi(x)
)
≤ g(x).
22
-
Remark 6.4. Assuming (6.5) and integrating (6.6) one gets the
inequality
∏
i=1
(
∫
fidxi
)λi≤∫
g(x)dx,
which can be considered as a weak form of the Blaschke–Santaló
functional inequality. Thisfollows, of course, directly from the
Prekopa–Leindler inequality.
In particular, assuming that the functions Vi satisfy
∑
i=1
λiVi(xi) ≥1
2
∣
∣
∣
k∑
i=1
λixi
∣
∣
∣
2,
one gets(
k∏
i=1
∫
e−Vi(xi) dxi
)λiρ(x) ≤ e−
|x|2
2 .
Rewriting this inequality with respect to the Gaussian reference
measure γ, one gets thefollowing equivalent formulation.
Corollary 6.5. Assume that the measurable functions Fi
satisfy
k∑
i=1
λiFi(xi) ≤1
2
[
k∑
i=1
λi|xi|2 −∣
∣
∣
k∑
j=1
λjxj
∣
∣
∣
2]
.
Then(
k∏
i=1
∫
eFidγ)λi
p(x) ≤ 1,
where p · γ is the barycenter of eFi∫eFidγ
· γ.
7 Talagrand-type estimates for the barycenter functional
In this section we show that a weak form of the
Blaschke–Santaló inequality is related todisplacement convexity
property of the Gaussian entropy. The conjectured strong formof the
Blaschke–Santaló inequality is equivalent to certain strong
entropy-W2-bound, aparticular case of this bound for two functions
was proved by M. Fathi in [17].
In what follows π denotes the solution to the multimarginal
Kantorovich problem withmarginals µi. Note that
k∑
i=1
∣
∣xi −1
k
k∑
j=1
xi∣
∣
2=
1
k
k∑
i,j=1,i
-
Theorem 7.1. Assume that for 1 ≤ i ≤ k, µi = ρi · γ are
probability measures and the ρiare unconditional. Then
F(µ) ≤ k − 1k2
k∑
i=1
∫
ρi log ρidγ =k − 1k2
k∑
i=1
Entγ(µi). (7.1)
Proof. By the Kantorovich duality (see e.g., [38]),
F(µ) = 12k2
∫
∑
i,j=1,i
-
Remark 7.2. This result is a generalization in the unconditional
setting of a result ofM. Fathi [17] for two functions:Let ρ0, ρ1 be
two Gaussian unconditional probability densities and ρ1/2 be the
correspondingbarycenter. Then inequality 7.1 implies
1
2W 22 (ρ0 · γ, ρ1 · γ) = 2W 22 (ρ0 · γ, ρ1/2 · γ) =W 22 (ρ0 · γ,
ρ1/2 · γ) +W 22 (ρ1 · γ, ρ1/2 · γ)
≤∫
ρ0 log ρ0dγ +
∫
ρ1 log ρ1dγ. (7.3)
This is a particular case of Fathi’s inequality.
Fathi has shown that in the class of symmetric functions
inequality (7.3) is equivalentto a Blaschke–Santaló inequality
involving two exponential functions. We follow his ap-proach in
[17] to show that the inequality of Theorem 7.1 is also equivalent
to a functionalBlaschke–Santaló for multiple exponential
functions.
Indeed, letting ρ(t) = e−t
k−1 in Theorem 3.1, we get the following multifunctional
Blaschke–Santaló inequality:Let fi : R
n → R+, 1 ≤ i ≤ k, be measurable unconditional functions such
thatk∑
i=1
fi(xi) ≤ −1
k − 1
k∑
i,j=1,i
-
By the Kantorovitch duality, the left hand side of this
inequality equals
−2kinfP
k∑
i,j=1,i
-
Theorem 7.6. Let µi = ρi ·γ be probability measures and fi(xi)
be the solution to the dualmultimarginal problem with marginals µi
and the cost function
12k
∑ki,j=1,i
-
Proposition 8.1. Let V be a strictly convex C2-function such
that e−V , e−V∗are inte-
grable functions. Let ∇Ψ be the optimal transportation of e−V
dx∫e−V dx
onto e−V ∗dx∫e−V ∗dx
. Then
BS(V ) ≤ J 2(Ψ) ≤ BS(Ψ).
Equivalently
∫
e−V dx
∫
e−V∗dx ≤
(
∫
e−12〈x,∇Ψ〉
√detD2Ψdx
)2≤∫
e−Ψdx
∫
e−Ψ∗dx.
Proof. The second inequality is just Theorem 3.2 (iii). To prove
the first inequality, weapply the change of variables formula
e−V∫
e−V dx=e−V
∗(∇Ψ)
∫
e−V∗dx
detD2Ψ.
Then∫
e−12〈x,∇Ψ〉
√detD2Ψdx =
√
∫
e−V∗dx
∫
e−V dx
∫
eV ∗(∇Ψ)−V (x)−〈x,∇Ψ〉
2 dx.
The result follows from the inequality V ∗(∇Ψ) + V (x) ≥
〈x,∇Ψ〉.
It can be easily seen from the proof that equality is attained
if and only if V = Ψ+ a forsome constant a. Thus, within a certain
appropriate class of functions, e.g., symmetric,
the maximum of the Blaschke–Santaló functional must satisfy
that the measure e−Ψ∗
∫e−Ψ∗dx
is the push-forward measure of e−Ψ
∫e−Ψdx
under the mapping ∇Ψ. This means that Ψ solvesthe following
Monge–Ampère equation
e−Ψ∫
e−Ψdx=e−Ψ
∗(∇Ψ)
∫
e−Ψ∗dx
detD2Ψ. (8.2)
It was shown in [13] that this equation admits the following
family of solutions, providede−Ψ∫e−Ψdx
has logarithmic derivatives,
Ψ =〈Ax, x〉
2+ c,
where A is a positive definite matrix and c is a constant. These
are exactly the maximizersof the Blaschke–Santaló functional.
Thus this observation suggests the following (so far heuristic)
approach to the Blaschke–Santaló inequality. Let Ψ0 = V , and
consider iterations Ψl, l ∈ N, where Ψl+1 is theoptimal
transportation potential pushing forward e
−Ψldx∫e−Ψldx
onto e−Ψ∗
l dx∫e−Ψ∗
l dx. By Proposition
8.1, one gets an increasing sequence BS(Ψl), l ∈ N. From this
one can try to extractconvergence of Ψl to a potential Ψ, which
gives a maximum to the Blaschke–Santalófunctional. Then prove that
Ψ solves (8.2), and by uniqueness deduce that Ψ is quadratic.
8.2 The multimarginal case
Next we generalize the previous result to the multimarginal
case, k > 2.
28
-
Theorem 8.2. Assume that Vi(xi), 1 ≤ i ≤ k, are measurable
functions satisfying
k∑
i=1
λiVi(xi) ≥ Ck∑
i 0 and λi ∈ (0, 1) with∑k
i=1 λi = 1.Let the tuple of functions λiUi(xi) be the solution
to the dual multimarginal maximization
problem with marginals e−Vidxi∫e−Vidxi
and the cost function C∑k
i
-
Let us informally analyze the equality case. Clearly, in this
case one has for almost all y,
∑
i=1
λiVi(∇Φi(y)) =∑
i=1
λiUi(∇Φi(y)).
Integrating over ρ dy we get that (λiVi) is a dual Kantorovich
solution as well. Hence, byuniqueness of the dual solution
Vi = Ui + Ci,
k∑
i=1
Ci = 0.
In addition, one has for all i that
e−Ui(∇Φi)∫
e−UidxidetD2Φi = ρ,
or, equivalently,e−Ui
∫
e−Uidxi= ρ(∇Φ∗i ) detD2Φ∗i .
In particular, since (see Theorem 2.4)
Φ∗i (xi) = λi|xi|22
+Ui(xi)
C+ Ci,
every function Ui must satisfy
e−Ui∫
e−Uidxi= ρ
(∇Ui(xi)C
+ λixi
)
det
(
D2Ui
C+ λiI
)
. (8.3)
Thus, a maximizer of the Blaschke–Santaló inequality, if it
exists, must satisfy the systemof equations (8.3), where every Ui
is convex.
Remark 8.3. Equation (8.3) is an equation of the
Kähler–Einstein type. We do notknow whether (8.3) admits a unique
solution. The well posedness of the classical Kähler–Einstein
equation
e−Φ∫
e−Φdx= ρ(∇Φ)detD2Φ
was proved under broad assumptions in [14].
References
[1] M. Agueh M and G. Carlier, Barycenters in the Wasserstein
space, SIAM J. Math.Anal. 43 (2011), 904–924.
[2] A.D. Alexandroff, Almost everywhere existence of the second
differential of aconvex function and some properties of convex
surfaces connected with it, (Russian)Leningrad State Univ. Annals
[Uchenye Zapiski] Math. Ser. 6, (1939), 3–35.
[3] S. Artstein-Avidan, B. Klartag and V. Milman, The Santaló
point of a function,and a functional form of Santaló inequality,
Mathematika 51 (2004), 33–48.
30
-
[4] S. Artstein-Avidan, A. Giannopoulos and V.D. Milman,
Asymptotic GeometricAnalysis, Part I, Mathematical Surveys and
Monographs Volume 202; Amer. Math.Soc., Providence, Rhode Island
(2015).
[5] D. Bakry, I. Gentil and M. Ledoux, Analysis and Geometry of
Markov DiffusionOperators, Springer, Berlin (2013).
[6] K. Ball, Isometric problems in lp and sections of convex
sets, PhD dissertation,University of Cambridge (1986).
[7] F. Barthe, On a reverse form of the Brascamp-Lieb
inequality, Invent. Math. 134(1998), 335–361.
[8] V. I. Bogachev and A. V. Kolesnikov, The Monge–Kantorovich
problem: achieve-ments, connections, and perspectives, Russian
Math. Surveys 67 no. (5), (2012), 785–890.
[9] J. M. Borwein and J.D. Vanderwerff, Convex Functions:
Constructions, Char-acterizations and Counterexamples, Cambridge
University Press 2010.
[10] H. Brascamp and E. H. Lieb, On extensions of the
Brunn–Minkowski andPrékopa–Leindler theorems, including
inequalities for log concave functions, and withan application to
the diffusion equation, J. Funct. Anal. 22 (1976), 366–389.
[11] H. Busemann and W. Feller, Kruemmungseigenschaften konvexer
Flächen, ActaMath. 66 (1935), 1–47.
[12] U. Caglar, M. Fradelizi, O. Guedon, J. Lehec, C. Schütt
and E. M.Werner, Functional versions of Lp-affine surface area and
entropy inequalities, Int.Math. Res. Not., vol. 2016, (2016),
1223–1250.
[13] U. Caglar, A.V. Kolesnikov, and E.M. Werner, Pinsker
inequalities and re-lated Monge-Ampère equations for log concave
functions, arXiv: 2005.07055.
[14] D. Cordero-Erausquin and B. Klartag, Moment measures. J.
Funct. Anal. 268no. 12, (2015), 3834–3866.
[15] G. Peyré and M. Cuturi, Computational Optimal Transport,
arXiv:1803.00567.
[16] S. Dubuc, Critères de convexité et inégalités
intégrales, Ann. Inst. Fourier Grenoble,27 (1977), 135–165.
[17] M. Fathi, A sharp symmetrized form of Talagrand’s
transport-entropy inequality forthe Gaussian measure, Electron.
Commun. Probab. 23 (2018), Paper No. 81, 9 pp.
[18] A. Figalli, The Monge–Ampère equation and its
applications, European Math. Soc.,Zürich, x+200 p, (2017).
[19] J. Fontbona, N. Gozlan, and J. F. Jabir, A variational
approach to sometransport inequalities, Ann. Inst. Henri Poincaré
Probab. Stat. 53, no. 4, (2017), 1719–1746.
[20] M. Fradelizi and M. Meyer, Some functional forms of
Blaschke-Santaló inequal-ity, Math. Z. 256 no. 2, (2007),
379–395.
31
http://arxiv.org/abs/1803.00567
-
[21] M. Fradelizi and M. Meyer, Increasing functions and inverse
Santaló inequalityfor unconditional functions, Positivity 12 no. 3
(2008), 407–420.
[22] M. Fradelizi and M. Meyer, Some functional inverse Santaló
inequalities, Adv.Math. 218, no. 5, (2008), 1430–1452.
[23] M. Fradelizi, A. Hubard, M. Meyer, E. Roldán-Pensado, A
Zvavitch,Equipartitions and Mahler volumes of symmetric convex
bodies, arXiv:1904.10765v3
[24] W. Gangbo and A. Świech, Optimal maps for the
multidimensional Monge–Kantorovich problem. Comm. Pure Appl. Math.
51, (1998), 23–45.
[25] R. J. Gardner, The Brunn-Minkowski inequality . Bull. Amer.
Math. Soc. (N.S.)39 (2002), 355–405.
[26] N. Gozlan, The Deficit in the Gaussian log-Sobolev
inequality and inverse Santalóinequalities arXiv:2007.05255.
[27] H. Iriyeh and M. Shibata, Symmetric Mahler’s conjecture for
the volume productin the three dimensional case, arXiv:1706.01749
(2019).
[28] A. V. Kolesnikov, Mass transportation functionals on the
sphere with applicationsto the logarithmic Minkowski problem. Mosc.
Math. J. 20 no. 1, (2020), 67–91.
[29] A.V. Kolesnikov and E.D. Kosov, Moment measures and
stability for Gaussianinequalities. Theory Stoch. Process. 22 no.
2, (2017), 47–61.
[30] J. Lehec, A direct proof of the functional Santaló
inequality, C. R. Math. Acad. Sci.Paris 347 (2009), 55–58.
[31] Y.J. Lin and G.S. Leng, On the functional Blaschke-Santaló
inequality, Math.Inequal. Appl. 18, no.1, (2015), 145–154.
[32] E. Lutwak, The Brunn-Minkowski-Firey theory. II. Affine and
geominimal surfaceareas, Adv. Math. 118 (1996) 244–294.
[33] F. Nazarov, F. Petrov, D. Ryabogin, and A. Zvavitch, A
remark on theMahler conjecture: local minimality of the unit cube,
Duke Math. J. 154 (2010), 419–430.
[34] S. Reisner, C. Schütt and E. M. Werner, Mahler’s
conjecture and curvatureInternational Mathematics Research Notices,
IMRN 2012 (2012), 1–16.
[35] F. Santambrogio, Dealing with moment measures via entropy
and optimal trans-port. J. Funct. Anal. 271 no. 2, (2016),
418–436.
[36] R. Schneider, Convex Bodies: The Brunn Minkowski Theory.
Second expandededition. Encyclopedia of Mathematics and its
Applications, 151. Cambridge UniversityPress (2014).
[37] C. Schütt and E.M. Werner, Surface bodies and p-affine
surface area, Adv. Math.187 (2004) 98–145.
[38] C. Villani, Topics in optimal transportation, Amer. Math.
Soc. Providence, 58,Rhode Island, (2003).
32
http://arxiv.org/abs/1904.10765http://arxiv.org/abs/2007.05255http://arxiv.org/abs/1706.01749
-
[39] E.M. Werner and D. Ye, New Lp affine isoperimetric
inequalities, Adv. Math. 218(2008) 762–780.
Alexander V. KolesnikovFaculty of Mathematics
National Research Institute Higher School of Economics
Moscow, Russian Federation
[email protected]
Elisabeth WernerDepartment of Mathematics Université de Lille
1
Case Western Reserve University UFR de Mathématique
Cleveland, Ohio 44106, U. S. A. 59655 Villeneuve d’Ascq,
France
[email protected]
33
1 Introduction1.1 The main results
2 Integral bounds and facts about barycenters3 The unconditional
case4 Characterization of the equality cases5 The Blaschke–Santaló
inequality and affine isoperimetric inequality for many sets 6
Prékopa–Leindler and displacement convexity inequalities:
refinement of the transportational argument7 Talagrand-type
estimates for the barycenter functional8 Monotonicity of the
Blaschke-Santaló functional8.1 The case of two functions8.2 The
multimarginal case