convexity and asymptotic geometric...
TRANSCRIPT
What is the perimeter of a log–concave function?
Andrea Colesanti, in collaboration with Ilaria Fragala
Convexity and asymptotic geometric analysis
Centre de Recherches MathematiquesMontreal, 16–20 April 2012
Log–concave functions
We consider functions f : Rn → R+ of the form
f = e−u ,
whereu : Rn → (−∞,∞] is convex.
(with u 6≡ ∞). We will always assume
lim|x |→∞
f (x) = 0 ⇔ lim|x |→∞
u(x) =∞ .
Example.
f (x) = χK (x) =
{1 if x ∈ K ,0 if x 6∈ K ,
where K ⊂ Rn is a convex body. In this case
u(x) = IK (x) =
{0 if x ∈ K ,∞ if x 6∈ K ,
Log–concave functionsWe consider functions f : Rn → R+ of the form
f = e−u ,
whereu : Rn → (−∞,∞] is convex.
(with u 6≡ ∞).
We will always assume
lim|x |→∞
f (x) = 0 ⇔ lim|x |→∞
u(x) =∞ .
Example.
f (x) = χK (x) =
{1 if x ∈ K ,0 if x 6∈ K ,
where K ⊂ Rn is a convex body. In this case
u(x) = IK (x) =
{0 if x ∈ K ,∞ if x 6∈ K ,
Log–concave functionsWe consider functions f : Rn → R+ of the form
f = e−u ,
whereu : Rn → (−∞,∞] is convex.
(with u 6≡ ∞). We will always assume
lim|x |→∞
f (x) = 0
⇔ lim|x |→∞
u(x) =∞ .
Example.
f (x) = χK (x) =
{1 if x ∈ K ,0 if x 6∈ K ,
where K ⊂ Rn is a convex body. In this case
u(x) = IK (x) =
{0 if x ∈ K ,∞ if x 6∈ K ,
Log–concave functionsWe consider functions f : Rn → R+ of the form
f = e−u ,
whereu : Rn → (−∞,∞] is convex.
(with u 6≡ ∞). We will always assume
lim|x |→∞
f (x) = 0 ⇔ lim|x |→∞
u(x) =∞ .
Example.
f (x) = χK (x) =
{1 if x ∈ K ,0 if x 6∈ K ,
where K ⊂ Rn is a convex body. In this case
u(x) = IK (x) =
{0 if x ∈ K ,∞ if x 6∈ K ,
Log–concave functionsWe consider functions f : Rn → R+ of the form
f = e−u ,
whereu : Rn → (−∞,∞] is convex.
(with u 6≡ ∞). We will always assume
lim|x |→∞
f (x) = 0 ⇔ lim|x |→∞
u(x) =∞ .
Example.
f (x) = χK (x) =
{1 if x ∈ K ,0 if x 6∈ K ,
where K ⊂ Rn is a convex body. In this case
u(x) = IK (x) =
{0 if x ∈ K ,∞ if x 6∈ K ,
Log–concave functionsWe consider functions f : Rn → R+ of the form
f = e−u ,
whereu : Rn → (−∞,∞] is convex.
(with u 6≡ ∞). We will always assume
lim|x |→∞
f (x) = 0 ⇔ lim|x |→∞
u(x) =∞ .
Example.
f (x) = χK (x) =
{1 if x ∈ K ,0 if x 6∈ K ,
where K ⊂ Rn is a convex body.
In this case
u(x) = IK (x) =
{0 if x ∈ K ,∞ if x 6∈ K ,
Log–concave functionsWe consider functions f : Rn → R+ of the form
f = e−u ,
whereu : Rn → (−∞,∞] is convex.
(with u 6≡ ∞). We will always assume
lim|x |→∞
f (x) = 0 ⇔ lim|x |→∞
u(x) =∞ .
Example.
f (x) = χK (x) =
{1 if x ∈ K ,0 if x 6∈ K ,
where K ⊂ Rn is a convex body. In this case
u(x) = IK (x) =
{0 if x ∈ K ,∞ if x 6∈ K ,
Operations on log–concave functions
Let f and g be log–concave and s > 0. We set
(f ⊕ g)(z) := supx+y=z
f (x)g(y) ,
(s · f )(x) := f s(x
s
).
I These operations preserve log–concavity: f ⊕ g and s · f arelog–concave.
I If K and L are convex bodies and t, s > 0, then
s · χK ⊕ t · χL = χsK+tL ,
wheresK + tL = {sx + ty : x ∈ K , y ∈ L} .
Operations on log–concave functions
Let f and g be log–concave and s > 0.
We set
(f ⊕ g)(z) := supx+y=z
f (x)g(y) ,
(s · f )(x) := f s(x
s
).
I These operations preserve log–concavity: f ⊕ g and s · f arelog–concave.
I If K and L are convex bodies and t, s > 0, then
s · χK ⊕ t · χL = χsK+tL ,
wheresK + tL = {sx + ty : x ∈ K , y ∈ L} .
Operations on log–concave functions
Let f and g be log–concave and s > 0. We set
(f ⊕ g)(z) := supx+y=z
f (x)g(y) ,
(s · f )(x) := f s(x
s
).
I These operations preserve log–concavity: f ⊕ g and s · f arelog–concave.
I If K and L are convex bodies and t, s > 0, then
s · χK ⊕ t · χL = χsK+tL ,
wheresK + tL = {sx + ty : x ∈ K , y ∈ L} .
Operations on log–concave functions
Let f and g be log–concave and s > 0. We set
(f ⊕ g)(z) := supx+y=z
f (x)g(y) ,
(s · f )(x) := f s(x
s
).
I These operations preserve log–concavity: f ⊕ g and s · f arelog–concave.
I If K and L are convex bodies and t, s > 0, then
s · χK ⊕ t · χL = χsK+tL ,
wheresK + tL = {sx + ty : x ∈ K , y ∈ L} .
Operations on log–concave functions
Let f and g be log–concave and s > 0. We set
(f ⊕ g)(z) := supx+y=z
f (x)g(y) ,
(s · f )(x) := f s(x
s
).
I These operations preserve log–concavity: f ⊕ g and s · f arelog–concave.
I If K and L are convex bodies and t, s > 0, then
s · χK ⊕ t · χL = χsK+tL ,
wheresK + tL = {sx + ty : x ∈ K , y ∈ L} .
Operations on log–concave functions
Let f and g be log–concave and s > 0. We set
(f ⊕ g)(z) := supx+y=z
f (x)g(y) ,
(s · f )(x) := f s(x
s
).
I These operations preserve log–concavity: f ⊕ g and s · f arelog–concave.
I If K and L are convex bodies and t, s > 0, then
s · χK ⊕ t · χL = χsK+tL ,
wheresK + tL = {sx + ty : x ∈ K , y ∈ L} .
The integral functional and its derivative
For f log–concave (verifying the decay condition at infinity), weconsider
I (f ) =
∫Rn
f (x) dx ∈ [0,∞) .
I If f = χK (K convex body), I (f ) = V (K ) (the volume of K ).
We are interested in the limit
δI (f , g) := limε→0+
I (f ⊕ ε · g)− I (f )
ε.
I If f = χK and g = χL (K and L convex bodies) then
δI (f , g) = limε→0+
V (K + εL)− V (K )
ε= n V (K , . . . ,K , L)
=
∫Sn−1
hL(ξ) dσK (ξ) =
∫∂K
hL(νK (x)) dHn−1(x) .
The integral functional and its derivativeFor f log–concave (verifying the decay condition at infinity), weconsider
I (f ) =
∫Rn
f (x) dx
∈ [0,∞) .
I If f = χK (K convex body), I (f ) = V (K ) (the volume of K ).
We are interested in the limit
δI (f , g) := limε→0+
I (f ⊕ ε · g)− I (f )
ε.
I If f = χK and g = χL (K and L convex bodies) then
δI (f , g) = limε→0+
V (K + εL)− V (K )
ε= n V (K , . . . ,K , L)
=
∫Sn−1
hL(ξ) dσK (ξ) =
∫∂K
hL(νK (x)) dHn−1(x) .
The integral functional and its derivativeFor f log–concave (verifying the decay condition at infinity), weconsider
I (f ) =
∫Rn
f (x) dx ∈ [0,∞) .
I If f = χK (K convex body), I (f ) = V (K ) (the volume of K ).
We are interested in the limit
δI (f , g) := limε→0+
I (f ⊕ ε · g)− I (f )
ε.
I If f = χK and g = χL (K and L convex bodies) then
δI (f , g) = limε→0+
V (K + εL)− V (K )
ε= n V (K , . . . ,K , L)
=
∫Sn−1
hL(ξ) dσK (ξ) =
∫∂K
hL(νK (x)) dHn−1(x) .
The integral functional and its derivativeFor f log–concave (verifying the decay condition at infinity), weconsider
I (f ) =
∫Rn
f (x) dx ∈ [0,∞) .
I If f = χK (K convex body), I (f ) = V (K ) (the volume of K ).
We are interested in the limit
δI (f , g) := limε→0+
I (f ⊕ ε · g)− I (f )
ε.
I If f = χK and g = χL (K and L convex bodies) then
δI (f , g) = limε→0+
V (K + εL)− V (K )
ε= n V (K , . . . ,K , L)
=
∫Sn−1
hL(ξ) dσK (ξ) =
∫∂K
hL(νK (x)) dHn−1(x) .
The integral functional and its derivativeFor f log–concave (verifying the decay condition at infinity), weconsider
I (f ) =
∫Rn
f (x) dx ∈ [0,∞) .
I If f = χK (K convex body), I (f ) = V (K ) (the volume of K ).
We are interested in the limit
δI (f , g) := limε→0+
I (f ⊕ ε · g)− I (f )
ε.
I If f = χK and g = χL (K and L convex bodies) then
δI (f , g) = limε→0+
V (K + εL)− V (K )
ε= n V (K , . . . ,K , L)
=
∫Sn−1
hL(ξ) dσK (ξ) =
∫∂K
hL(νK (x)) dHn−1(x) .
The integral functional and its derivativeFor f log–concave (verifying the decay condition at infinity), weconsider
I (f ) =
∫Rn
f (x) dx ∈ [0,∞) .
I If f = χK (K convex body), I (f ) = V (K ) (the volume of K ).
We are interested in the limit
δI (f , g) := limε→0+
I (f ⊕ ε · g)− I (f )
ε.
I If f = χK and g = χL (K and L convex bodies)
then
δI (f , g) = limε→0+
V (K + εL)− V (K )
ε= n V (K , . . . ,K , L)
=
∫Sn−1
hL(ξ) dσK (ξ) =
∫∂K
hL(νK (x)) dHn−1(x) .
The integral functional and its derivativeFor f log–concave (verifying the decay condition at infinity), weconsider
I (f ) =
∫Rn
f (x) dx ∈ [0,∞) .
I If f = χK (K convex body), I (f ) = V (K ) (the volume of K ).
We are interested in the limit
δI (f , g) := limε→0+
I (f ⊕ ε · g)− I (f )
ε.
I If f = χK and g = χL (K and L convex bodies) then
δI (f , g) = limε→0+
V (K + εL)− V (K )
ε
= n V (K , . . . ,K , L)
=
∫Sn−1
hL(ξ) dσK (ξ) =
∫∂K
hL(νK (x)) dHn−1(x) .
The integral functional and its derivativeFor f log–concave (verifying the decay condition at infinity), weconsider
I (f ) =
∫Rn
f (x) dx ∈ [0,∞) .
I If f = χK (K convex body), I (f ) = V (K ) (the volume of K ).
We are interested in the limit
δI (f , g) := limε→0+
I (f ⊕ ε · g)− I (f )
ε.
I If f = χK and g = χL (K and L convex bodies) then
δI (f , g) = limε→0+
V (K + εL)− V (K )
ε= n V (K , . . . ,K , L)
=
∫Sn−1
hL(ξ) dσK (ξ) =
∫∂K
hL(νK (x)) dHn−1(x) .
The integral functional and its derivativeFor f log–concave (verifying the decay condition at infinity), weconsider
I (f ) =
∫Rn
f (x) dx ∈ [0,∞) .
I If f = χK (K convex body), I (f ) = V (K ) (the volume of K ).
We are interested in the limit
δI (f , g) := limε→0+
I (f ⊕ ε · g)− I (f )
ε.
I If f = χK and g = χL (K and L convex bodies) then
δI (f , g) = limε→0+
V (K + εL)− V (K )
ε= n V (K , . . . ,K , L)
=
∫Sn−1
hL(ξ) dσK (ξ)
=
∫∂K
hL(νK (x)) dHn−1(x) .
The integral functional and its derivativeFor f log–concave (verifying the decay condition at infinity), weconsider
I (f ) =
∫Rn
f (x) dx ∈ [0,∞) .
I If f = χK (K convex body), I (f ) = V (K ) (the volume of K ).
We are interested in the limit
δI (f , g) := limε→0+
I (f ⊕ ε · g)− I (f )
ε.
I If f = χK and g = χL (K and L convex bodies) then
δI (f , g) = limε→0+
V (K + εL)− V (K )
ε= n V (K , . . . ,K , L)
=
∫Sn−1
hL(ξ) dσK (ξ) =
∫∂K
hL(νK (x)) dHn−1(x) .
Representation formulas for δI (f , g) (1)
Let f = e−u, g = e−v . The equality
δI (f , g) =
∫Rn
v∗(∇u(x)) f (x) dx ,
where v∗ is the usual conjugate of v :
v∗(y) = supx∈Rn
x · y − v(x) ,
has been proved in some special cases.
I Klartag & Milman (2005), Rotem (2011): f = density of theGaussian measure.
I C. & Fragala (2011): general case, but with several additionalassumptions concerning smoothness, strict convexity, decay atinfinity of f and g .
I see also: Klartag (2007), who proved analogous results in thecontext of p–concave functions.
None of them includes the case: f = χK , g = χL, K and L convexbodies.
Representation formulas for δI (f , g) (1)Let f = e−u, g = e−v . The equality
δI (f , g) =
∫Rn
v∗(∇u(x)) f (x) dx ,
where v∗ is the usual conjugate of v :
v∗(y) = supx∈Rn
x · y − v(x) ,
has been proved in some special cases.
I Klartag & Milman (2005), Rotem (2011): f = density of theGaussian measure.
I C. & Fragala (2011): general case, but with several additionalassumptions concerning smoothness, strict convexity, decay atinfinity of f and g .
I see also: Klartag (2007), who proved analogous results in thecontext of p–concave functions.
None of them includes the case: f = χK , g = χL, K and L convexbodies.
Representation formulas for δI (f , g) (1)Let f = e−u, g = e−v . The equality
δI (f , g) =
∫Rn
v∗(∇u(x)) f (x) dx ,
where v∗ is the usual conjugate of v :
v∗(y) = supx∈Rn
x · y − v(x) ,
has been proved in some special cases.
I Klartag & Milman (2005), Rotem (2011): f = density of theGaussian measure.
I C. & Fragala (2011): general case, but with several additionalassumptions concerning smoothness, strict convexity, decay atinfinity of f and g .
I see also: Klartag (2007), who proved analogous results in thecontext of p–concave functions.
None of them includes the case: f = χK , g = χL, K and L convexbodies.
Representation formulas for δI (f , g) (1)Let f = e−u, g = e−v . The equality
δI (f , g) =
∫Rn
v∗(∇u(x)) f (x) dx ,
where v∗ is the usual conjugate of v :
v∗(y) = supx∈Rn
x · y − v(x) ,
has been proved in some special cases.
I Klartag & Milman (2005), Rotem (2011): f = density of theGaussian measure.
I C. & Fragala (2011): general case, but with several additionalassumptions concerning smoothness, strict convexity, decay atinfinity of f and g .
I see also: Klartag (2007), who proved analogous results in thecontext of p–concave functions.
None of them includes the case: f = χK , g = χL, K and L convexbodies.
Representation formulas for δI (f , g) (1)Let f = e−u, g = e−v . The equality
δI (f , g) =
∫Rn
v∗(∇u(x)) f (x) dx ,
where v∗ is the usual conjugate of v :
v∗(y) = supx∈Rn
x · y − v(x) ,
has been proved in some special cases.
I Klartag & Milman (2005), Rotem (2011): f = density of theGaussian measure.
I C. & Fragala (2011): general case, but with several additionalassumptions concerning smoothness, strict convexity, decay atinfinity of f and g .
I see also: Klartag (2007), who proved analogous results in thecontext of p–concave functions.
None of them includes the case: f = χK , g = χL, K and L convexbodies.
Representation formulas for δI (f , g) (1)Let f = e−u, g = e−v . The equality
δI (f , g) =
∫Rn
v∗(∇u(x)) f (x) dx ,
where v∗ is the usual conjugate of v :
v∗(y) = supx∈Rn
x · y − v(x) ,
has been proved in some special cases.
I Klartag & Milman (2005), Rotem (2011): f = density of theGaussian measure.
I C. & Fragala (2011): general case, but with several additionalassumptions concerning smoothness, strict convexity, decay atinfinity of f and g .
I see also: Klartag (2007), who proved analogous results in thecontext of p–concave functions.
None of them includes the case: f = χK , g = χL, K and L convexbodies.
Representation formulas for δI (f , g) (1)Let f = e−u, g = e−v . The equality
δI (f , g) =
∫Rn
v∗(∇u(x)) f (x) dx ,
where v∗ is the usual conjugate of v :
v∗(y) = supx∈Rn
x · y − v(x) ,
has been proved in some special cases.
I Klartag & Milman (2005), Rotem (2011): f = density of theGaussian measure.
I C. & Fragala (2011): general case, but with several additionalassumptions concerning smoothness, strict convexity, decay atinfinity of f and g .
I see also: Klartag (2007), who proved analogous results in thecontext of p–concave functions.
None of them includes the case: f = χK , g = χL, K and L convexbodies.
Representation formulas for δI (f , g) (1)Let f = e−u, g = e−v . The equality
δI (f , g) =
∫Rn
v∗(∇u(x)) f (x) dx ,
where v∗ is the usual conjugate of v :
v∗(y) = supx∈Rn
x · y − v(x) ,
has been proved in some special cases.
I Klartag & Milman (2005), Rotem (2011): f = density of theGaussian measure.
I C. & Fragala (2011): general case, but with several additionalassumptions concerning smoothness, strict convexity, decay atinfinity of f and g .
I see also: Klartag (2007), who proved analogous results in thecontext of p–concave functions.
None of them includes the case: f = χK , g = χL, K and L convexbodies.
Representation formulas for δI (f , g) (2)
Let f = e−u and g = e−v and set
Sf = {x ∈ Rn : f (x) > 0} , Sg = {x ∈ Rn : g(x) > 0} .
Sf and Sg are convex sets; let Hf and Hg be their supportfunctions. Then the equality
δI (f , g) =
∫Sf
v∗(∇u(x)) f (x) dx +
∫∂Sf
Hg (ν(x)) f (x) dHn−1(x) ,
where ν is the outer unit normal to Hf , has been proved underseveral additional assumptions on f , g , Sf and Sg .(C. & Fragala, 2011)
Representation formulas for δI (f , g) (2)
Let f = e−u and g = e−v and set
Sf = {x ∈ Rn : f (x) > 0} , Sg = {x ∈ Rn : g(x) > 0} .
Sf and Sg are convex sets; let Hf and Hg be their supportfunctions. Then the equality
δI (f , g) =
∫Sf
v∗(∇u(x)) f (x) dx +
∫∂Sf
Hg (ν(x)) f (x) dHn−1(x) ,
where ν is the outer unit normal to Hf , has been proved underseveral additional assumptions on f , g , Sf and Sg .(C. & Fragala, 2011)
Representation formulas for δI (f , g) (2)
Let f = e−u and g = e−v and set
Sf = {x ∈ Rn : f (x) > 0} , Sg = {x ∈ Rn : g(x) > 0} .
Sf and Sg are convex sets; let Hf and Hg be their supportfunctions.
Then the equality
δI (f , g) =
∫Sf
v∗(∇u(x)) f (x) dx +
∫∂Sf
Hg (ν(x)) f (x) dHn−1(x) ,
where ν is the outer unit normal to Hf , has been proved underseveral additional assumptions on f , g , Sf and Sg .(C. & Fragala, 2011)
Representation formulas for δI (f , g) (2)
Let f = e−u and g = e−v and set
Sf = {x ∈ Rn : f (x) > 0} , Sg = {x ∈ Rn : g(x) > 0} .
Sf and Sg are convex sets; let Hf and Hg be their supportfunctions. Then the equality
δI (f , g) =
∫Sf
v∗(∇u(x)) f (x) dx +
∫∂Sf
Hg (ν(x)) f (x) dHn−1(x) ,
where ν is the outer unit normal to Hf , has been proved underseveral additional assumptions on f , g , Sf and Sg .(C. & Fragala, 2011)
Representation formulas for δI (f , g) (2)
Let f = e−u and g = e−v and set
Sf = {x ∈ Rn : f (x) > 0} , Sg = {x ∈ Rn : g(x) > 0} .
Sf and Sg are convex sets; let Hf and Hg be their supportfunctions. Then the equality
δI (f , g) =
∫Sf
v∗(∇u(x)) f (x) dx +
∫∂Sf
Hg (ν(x)) f (x) dHn−1(x) ,
where ν is the outer unit normal to Hf ,
has been proved underseveral additional assumptions on f , g , Sf and Sg .(C. & Fragala, 2011)
Representation formulas for δI (f , g) (2)
Let f = e−u and g = e−v and set
Sf = {x ∈ Rn : f (x) > 0} , Sg = {x ∈ Rn : g(x) > 0} .
Sf and Sg are convex sets; let Hf and Hg be their supportfunctions. Then the equality
δI (f , g) =
∫Sf
v∗(∇u(x)) f (x) dx +
∫∂Sf
Hg (ν(x)) f (x) dHn−1(x) ,
where ν is the outer unit normal to Hf , has been proved underseveral additional assumptions on f , g , Sf and Sg .
(C. & Fragala, 2011)
Representation formulas for δI (f , g) (2)
Let f = e−u and g = e−v and set
Sf = {x ∈ Rn : f (x) > 0} , Sg = {x ∈ Rn : g(x) > 0} .
Sf and Sg are convex sets; let Hf and Hg be their supportfunctions. Then the equality
δI (f , g) =
∫Sf
v∗(∇u(x)) f (x) dx +
∫∂Sf
Hg (ν(x)) f (x) dHn−1(x) ,
where ν is the outer unit normal to Hf , has been proved underseveral additional assumptions on f , g , Sf and Sg .(C. & Fragala, 2011)
Perimeter
If K ⊂ Rn is a convex body, its perimeter, i.e. Hn−1(∂K ), can bedefined as
P(K ) = limε→0+
V (K + εB)− V (K )
ε,
where B is the unit ball of Rn.
I Define a corresponding notion of perimeter of a log-concavefunction f , as
P(f ) = limε→0+
I (f ⊕ ε · g)− I (f )
ε= δI (f , g) ,
with an appropriate choice of a unit ball g in the space oflog-concave functions.
Question: is there a natural candidate for g?
Perimeter
If K ⊂ Rn is a convex body, its perimeter, i.e. Hn−1(∂K ), can bedefined as
P(K ) = limε→0+
V (K + εB)− V (K )
ε,
where B is the unit ball of Rn.
I Define a corresponding notion of perimeter of a log-concavefunction f , as
P(f ) = limε→0+
I (f ⊕ ε · g)− I (f )
ε= δI (f , g) ,
with an appropriate choice of a unit ball g in the space oflog-concave functions.
Question: is there a natural candidate for g?
Perimeter
If K ⊂ Rn is a convex body, its perimeter, i.e. Hn−1(∂K ), can bedefined as
P(K ) = limε→0+
V (K + εB)− V (K )
ε,
where B is the unit ball of Rn.
I Define a corresponding notion of perimeter of a log-concavefunction f , as
P(f ) = limε→0+
I (f ⊕ ε · g)− I (f )
ε= δI (f , g) ,
with an appropriate choice of a unit ball g in the space oflog-concave functions.
Question: is there a natural candidate for g?
Perimeter
If K ⊂ Rn is a convex body, its perimeter, i.e. Hn−1(∂K ), can bedefined as
P(K ) = limε→0+
V (K + εB)− V (K )
ε,
where B is the unit ball of Rn.
I Define a corresponding notion of perimeter of a log-concavefunction f , as
P(f ) = limε→0+
I (f ⊕ ε · g)− I (f )
ε= δI (f , g) ,
with an appropriate choice of a unit ball g in the space oflog-concave functions.
Question: is there a natural candidate for g?
Perimeter
If K ⊂ Rn is a convex body, its perimeter, i.e. Hn−1(∂K ), can bedefined as
P(K ) = limε→0+
V (K + εB)− V (K )
ε,
where B is the unit ball of Rn.
I Define a corresponding notion of perimeter of a log-concavefunction f ,
as
P(f ) = limε→0+
I (f ⊕ ε · g)− I (f )
ε= δI (f , g) ,
with an appropriate choice of a unit ball g in the space oflog-concave functions.
Question: is there a natural candidate for g?
Perimeter
If K ⊂ Rn is a convex body, its perimeter, i.e. Hn−1(∂K ), can bedefined as
P(K ) = limε→0+
V (K + εB)− V (K )
ε,
where B is the unit ball of Rn.
I Define a corresponding notion of perimeter of a log-concavefunction f , as
P(f ) = limε→0+
I (f ⊕ ε · g)− I (f )
ε= δI (f , g) ,
with an appropriate choice of a unit ball g in the space oflog-concave functions.
Question: is there a natural candidate for g?
Perimeter
If K ⊂ Rn is a convex body, its perimeter, i.e. Hn−1(∂K ), can bedefined as
P(K ) = limε→0+
V (K + εB)− V (K )
ε,
where B is the unit ball of Rn.
I Define a corresponding notion of perimeter of a log-concavefunction f , as
P(f ) = limε→0+
I (f ⊕ ε · g)− I (f )
ε= δI (f , g) ,
with an appropriate choice of a unit ball g in the space oflog-concave functions.
Question: is there a natural candidate for g?
Perimeter
If K ⊂ Rn is a convex body, its perimeter, i.e. Hn−1(∂K ), can bedefined as
P(K ) = limε→0+
V (K + εB)− V (K )
ε,
where B is the unit ball of Rn.
I Define a corresponding notion of perimeter of a log-concavefunction f , as
P(f ) = limε→0+
I (f ⊕ ε · g)− I (f )
ε= δI (f , g) ,
with an appropriate choice of a unit ball g in the space oflog-concave functions.
Question: is there a natural candidate for g?
The unit ball in the class of log–concave functions
I A natural possibility is to chose
g(x) = γn(x) = e−|x |2/2 ;
this choice was made also by Klartag & Milman and Rotem intheir definition of mean width of a log–concave function, andin [C. & Fragala, 2012].
I A drastically different choice is
g = χB = characteristic function of the unit ball of Rn,
which is what we consider here.
I There are clearly many other choices, (e.g.
g(x) = e−|x |p/p , p ≥ 1 .)
The unit ball in the class of log–concave functions
I A natural possibility is to chose
g(x) = γn(x) = e−|x |2/2 ;
this choice was made also by Klartag & Milman and Rotem intheir definition of mean width of a log–concave function, andin [C. & Fragala, 2012].
I A drastically different choice is
g = χB = characteristic function of the unit ball of Rn,
which is what we consider here.
I There are clearly many other choices, (e.g.
g(x) = e−|x |p/p , p ≥ 1 .)
The unit ball in the class of log–concave functions
I A natural possibility is to chose
g(x) = γn(x) = e−|x |2/2 ;
this choice was made also by Klartag & Milman and Rotem intheir definition of mean width of a log–concave function, andin [C. & Fragala, 2012].
I A drastically different choice is
g = χB = characteristic function of the unit ball of Rn,
which is what we consider here.
I There are clearly many other choices, (e.g.
g(x) = e−|x |p/p , p ≥ 1 .)
The unit ball in the class of log–concave functions
I A natural possibility is to chose
g(x) = γn(x) = e−|x |2/2 ;
this choice was made also by Klartag & Milman and Rotem intheir definition of mean width of a log–concave function, andin [C. & Fragala, 2012].
I A drastically different choice is
g = χB = characteristic function of the unit ball of Rn,
which is what we consider here.
I There are clearly many other choices, (e.g.
g(x) = e−|x |p/p , p ≥ 1 .)
The unit ball in the class of log–concave functions
I A natural possibility is to chose
g(x) = γn(x) = e−|x |2/2 ;
this choice was made also by Klartag & Milman and Rotem intheir definition of mean width of a log–concave function, andin [C. & Fragala, 2012].
I A drastically different choice is
g = χB = characteristic function of the unit ball of Rn,
which is what we consider here.
I There are clearly many other choices, (e.g.
g(x) = e−|x |p/p , p ≥ 1 .)
The unit ball in the class of log–concave functions
I A natural possibility is to chose
g(x) = γn(x) = e−|x |2/2 ;
this choice was made also by Klartag & Milman and Rotem intheir definition of mean width of a log–concave function, andin [C. & Fragala, 2012].
I A drastically different choice is
g = χB = characteristic function of the unit ball of Rn,
which is what we consider here.
I There are clearly many other choices,
(e.g.
g(x) = e−|x |p/p , p ≥ 1 .)
The unit ball in the class of log–concave functions
I A natural possibility is to chose
g(x) = γn(x) = e−|x |2/2 ;
this choice was made also by Klartag & Milman and Rotem intheir definition of mean width of a log–concave function, andin [C. & Fragala, 2012].
I A drastically different choice is
g = χB = characteristic function of the unit ball of Rn,
which is what we consider here.
I There are clearly many other choices, (e.g.
g(x) = e−|x |p/p , p ≥ 1 .)
Definition of P(f ) and a useful formula
Let B be the unit ball of Rn. For a log–concave function f = e−u
define:
P(f ) := δI (f , χB) = limε→0+
I (f ⊕ ε · χB)− I (f )
ε.
Thm. (C. & Fragala, 2012) The function ε→ I (f ⊕ ε · χB) is apolynomial of degree n w.r.t. ε. Moreover, its (right) derivative atε = 0, i.e. P(f ), is
P(f ) =
∫ ∞0
[perimeter of the level set {f > t}] dt .
Note that {f > t} is convex for every t, by the log–concavity of f .
Definition of P(f ) and a useful formula
Let B be the unit ball of Rn.
For a log–concave function f = e−u
define:
P(f ) := δI (f , χB) = limε→0+
I (f ⊕ ε · χB)− I (f )
ε.
Thm. (C. & Fragala, 2012) The function ε→ I (f ⊕ ε · χB) is apolynomial of degree n w.r.t. ε. Moreover, its (right) derivative atε = 0, i.e. P(f ), is
P(f ) =
∫ ∞0
[perimeter of the level set {f > t}] dt .
Note that {f > t} is convex for every t, by the log–concavity of f .
Definition of P(f ) and a useful formula
Let B be the unit ball of Rn. For a log–concave function f = e−u
define:
P(f ) := δI (f , χB) = limε→0+
I (f ⊕ ε · χB)− I (f )
ε.
Thm. (C. & Fragala, 2012) The function ε→ I (f ⊕ ε · χB) is apolynomial of degree n w.r.t. ε. Moreover, its (right) derivative atε = 0, i.e. P(f ), is
P(f ) =
∫ ∞0
[perimeter of the level set {f > t}] dt .
Note that {f > t} is convex for every t, by the log–concavity of f .
Definition of P(f ) and a useful formula
Let B be the unit ball of Rn. For a log–concave function f = e−u
define:
P(f ) := δI (f , χB) = limε→0+
I (f ⊕ ε · χB)− I (f )
ε.
Thm. (C. & Fragala, 2012) The function ε→ I (f ⊕ ε · χB) is apolynomial of degree n w.r.t. ε.
Moreover, its (right) derivative atε = 0, i.e. P(f ), is
P(f ) =
∫ ∞0
[perimeter of the level set {f > t}] dt .
Note that {f > t} is convex for every t, by the log–concavity of f .
Definition of P(f ) and a useful formula
Let B be the unit ball of Rn. For a log–concave function f = e−u
define:
P(f ) := δI (f , χB) = limε→0+
I (f ⊕ ε · χB)− I (f )
ε.
Thm. (C. & Fragala, 2012) The function ε→ I (f ⊕ ε · χB) is apolynomial of degree n w.r.t. ε. Moreover, its (right) derivative atε = 0, i.e. P(f ), is
P(f ) =
∫ ∞0
[perimeter of the level set {f > t}] dt .
Note that {f > t} is convex for every t, by the log–concavity of f .
Definition of P(f ) and a useful formula
Let B be the unit ball of Rn. For a log–concave function f = e−u
define:
P(f ) := δI (f , χB) = limε→0+
I (f ⊕ ε · χB)− I (f )
ε.
Thm. (C. & Fragala, 2012) The function ε→ I (f ⊕ ε · χB) is apolynomial of degree n w.r.t. ε. Moreover, its (right) derivative atε = 0, i.e. P(f ), is
P(f ) =
∫ ∞0
[perimeter of the level set {f > t}] dt .
Note that {f > t} is convex for every t, by the log–concavity of f .
Definition of P(f ) and a useful formula
Let B be the unit ball of Rn. For a log–concave function f = e−u
define:
P(f ) := δI (f , χB) = limε→0+
I (f ⊕ ε · χB)− I (f )
ε.
Thm. (C. & Fragala, 2012) The function ε→ I (f ⊕ ε · χB) is apolynomial of degree n w.r.t. ε. Moreover, its (right) derivative atε = 0, i.e. P(f ), is
P(f ) =
∫ ∞0
[perimeter of the level set {f > t}] dt .
Note that {f > t} is convex for every t, by the log–concavity of f .
Perimeter and total variation
Let f = e−u : Rn → R be log–concave. By the previous result andcoarea formula, if moreover f ∈ C 1(Rn), we have
P(f ) =
∫ ∞0
[perimeter of the level set {f > t}] dt
=
∫Rn
|∇f | dx .
In the general case
P(f ) = |Df |(Rn) = total variation of f in Rn
=
∫Sf
|∇f | dx +
∫∂Sf
f dHn−1 .
These equalities confirm the validity of the representation formulasfor δI (f , g) seen before when f is arbitrary and g = χB .
Perimeter and total variation
Let f = e−u : Rn → R be log–concave.
By the previous result andcoarea formula, if moreover f ∈ C 1(Rn), we have
P(f ) =
∫ ∞0
[perimeter of the level set {f > t}] dt
=
∫Rn
|∇f | dx .
In the general case
P(f ) = |Df |(Rn) = total variation of f in Rn
=
∫Sf
|∇f | dx +
∫∂Sf
f dHn−1 .
These equalities confirm the validity of the representation formulasfor δI (f , g) seen before when f is arbitrary and g = χB .
Perimeter and total variation
Let f = e−u : Rn → R be log–concave. By the previous result andcoarea formula,
if moreover f ∈ C 1(Rn), we have
P(f ) =
∫ ∞0
[perimeter of the level set {f > t}] dt
=
∫Rn
|∇f | dx .
In the general case
P(f ) = |Df |(Rn) = total variation of f in Rn
=
∫Sf
|∇f | dx +
∫∂Sf
f dHn−1 .
These equalities confirm the validity of the representation formulasfor δI (f , g) seen before when f is arbitrary and g = χB .
Perimeter and total variation
Let f = e−u : Rn → R be log–concave. By the previous result andcoarea formula, if moreover f ∈ C 1(Rn), we have
P(f ) =
∫ ∞0
[perimeter of the level set {f > t}] dt
=
∫Rn
|∇f | dx .
In the general case
P(f ) = |Df |(Rn) = total variation of f in Rn
=
∫Sf
|∇f | dx +
∫∂Sf
f dHn−1 .
These equalities confirm the validity of the representation formulasfor δI (f , g) seen before when f is arbitrary and g = χB .
Perimeter and total variation
Let f = e−u : Rn → R be log–concave. By the previous result andcoarea formula, if moreover f ∈ C 1(Rn), we have
P(f ) =
∫ ∞0
[perimeter of the level set {f > t}] dt
=
∫Rn
|∇f | dx .
In the general case
P(f ) = |Df |(Rn) = total variation of f in Rn
=
∫Sf
|∇f | dx +
∫∂Sf
f dHn−1 .
These equalities confirm the validity of the representation formulasfor δI (f , g) seen before when f is arbitrary and g = χB .
Perimeter and total variation
Let f = e−u : Rn → R be log–concave. By the previous result andcoarea formula, if moreover f ∈ C 1(Rn), we have
P(f ) =
∫ ∞0
[perimeter of the level set {f > t}] dt
=
∫Rn
|∇f | dx .
In the general case
P(f ) = |Df |(Rn)
= total variation of f in Rn
=
∫Sf
|∇f | dx +
∫∂Sf
f dHn−1 .
These equalities confirm the validity of the representation formulasfor δI (f , g) seen before when f is arbitrary and g = χB .
Perimeter and total variation
Let f = e−u : Rn → R be log–concave. By the previous result andcoarea formula, if moreover f ∈ C 1(Rn), we have
P(f ) =
∫ ∞0
[perimeter of the level set {f > t}] dt
=
∫Rn
|∇f | dx .
In the general case
P(f ) = |Df |(Rn) = total variation of f in Rn
=
∫Sf
|∇f | dx +
∫∂Sf
f dHn−1 .
These equalities confirm the validity of the representation formulasfor δI (f , g) seen before when f is arbitrary and g = χB .
Perimeter and total variation
Let f = e−u : Rn → R be log–concave. By the previous result andcoarea formula, if moreover f ∈ C 1(Rn), we have
P(f ) =
∫ ∞0
[perimeter of the level set {f > t}] dt
=
∫Rn
|∇f | dx .
In the general case
P(f ) = |Df |(Rn) = total variation of f in Rn
=
∫Sf
|∇f | dx +
∫∂Sf
f dHn−1 .
These equalities confirm the validity of the representation formulasfor δI (f , g) seen before when f is arbitrary and g = χB .
Perimeter and total variation
Let f = e−u : Rn → R be log–concave. By the previous result andcoarea formula, if moreover f ∈ C 1(Rn), we have
P(f ) =
∫ ∞0
[perimeter of the level set {f > t}] dt
=
∫Rn
|∇f | dx .
In the general case
P(f ) = |Df |(Rn) = total variation of f in Rn
=
∫Sf
|∇f | dx +
∫∂Sf
f dHn−1 .
These equalities confirm the validity of the representation formulasfor δI (f , g) seen before when f is arbitrary and g = χB .
Properties of the perimeter
Recall:
P(f ) =
∫ ∞0
[perimeter of the level set {f > t}] dt .
I Monotonicity:f ≤ g ⇒ P(f ) ≤ P(g) .
I Valuation property: for every f and g log–concave such thatf ∧ g is log–concave:
P(f ∧ g) + P(f ∨ g) = P(f ) + P(g) .
I Brunn-Minkowski inequality: for every f and g log–concave
P((1− t) · f ⊕ t · g) ≥ (P(f ))1−t (P(g))t .
This result was proved by Bobkov (2007).
Properties of the perimeterRecall:
P(f ) =
∫ ∞0
[perimeter of the level set {f > t}] dt .
I Monotonicity:f ≤ g ⇒ P(f ) ≤ P(g) .
I Valuation property: for every f and g log–concave such thatf ∧ g is log–concave:
P(f ∧ g) + P(f ∨ g) = P(f ) + P(g) .
I Brunn-Minkowski inequality: for every f and g log–concave
P((1− t) · f ⊕ t · g) ≥ (P(f ))1−t (P(g))t .
This result was proved by Bobkov (2007).
Properties of the perimeterRecall:
P(f ) =
∫ ∞0
[perimeter of the level set {f > t}] dt .
I Monotonicity:f ≤ g ⇒ P(f ) ≤ P(g) .
I Valuation property: for every f and g log–concave such thatf ∧ g is log–concave:
P(f ∧ g) + P(f ∨ g) = P(f ) + P(g) .
I Brunn-Minkowski inequality: for every f and g log–concave
P((1− t) · f ⊕ t · g) ≥ (P(f ))1−t (P(g))t .
This result was proved by Bobkov (2007).
Properties of the perimeterRecall:
P(f ) =
∫ ∞0
[perimeter of the level set {f > t}] dt .
I Monotonicity:f ≤ g ⇒ P(f ) ≤ P(g) .
I Valuation property: for every f and g log–concave such thatf ∧ g is log–concave:
P(f ∧ g) + P(f ∨ g) = P(f ) + P(g) .
I Brunn-Minkowski inequality: for every f and g log–concave
P((1− t) · f ⊕ t · g) ≥ (P(f ))1−t (P(g))t .
This result was proved by Bobkov (2007).
Properties of the perimeterRecall:
P(f ) =
∫ ∞0
[perimeter of the level set {f > t}] dt .
I Monotonicity:f ≤ g ⇒ P(f ) ≤ P(g) .
I Valuation property: for every f and g log–concave such thatf ∧ g is log–concave:
P(f ∧ g) + P(f ∨ g) = P(f ) + P(g) .
I Brunn-Minkowski inequality: for every f and g log–concave
P((1− t) · f ⊕ t · g) ≥ (P(f ))1−t (P(g))t .
This result was proved by Bobkov (2007).
Properties of the perimeterRecall:
P(f ) =
∫ ∞0
[perimeter of the level set {f > t}] dt .
I Monotonicity:f ≤ g ⇒ P(f ) ≤ P(g) .
I Valuation property: for every f and g log–concave such thatf ∧ g is log–concave:
P(f ∧ g) + P(f ∨ g) = P(f ) + P(g) .
I Brunn-Minkowski inequality: for every f and g log–concave
P((1− t) · f ⊕ t · g) ≥ (P(f ))1−t (P(g))t .
This result was proved by Bobkov (2007).
A Cauchy type formula
Let f be log–concave, ξ ∈ Sn−1 and Eξ = ξ⊥. We define theprojection of f onto Eξ as
fξ(x′) := sup
t∈Rf (x ′ + tξ) , x ′ ∈ Eξ .
Put simply, the graph of fξ is the projection of the graph of f ontoEξ × R. Let
In−1(fξ) :=
∫Eξ
fξ(x′) dx ′ .
Then
P(f ) = c(n)
∫Sn−1
In−1(fξ) dHn−1(ξ) .
A Cauchy type formula
Let f be log–concave, ξ ∈ Sn−1 and Eξ = ξ⊥.
We define theprojection of f onto Eξ as
fξ(x′) := sup
t∈Rf (x ′ + tξ) , x ′ ∈ Eξ .
Put simply, the graph of fξ is the projection of the graph of f ontoEξ × R. Let
In−1(fξ) :=
∫Eξ
fξ(x′) dx ′ .
Then
P(f ) = c(n)
∫Sn−1
In−1(fξ) dHn−1(ξ) .
A Cauchy type formula
Let f be log–concave, ξ ∈ Sn−1 and Eξ = ξ⊥. We define theprojection of f onto Eξ as
fξ(x′) := sup
t∈Rf (x ′ + tξ) , x ′ ∈ Eξ .
Put simply, the graph of fξ is the projection of the graph of f ontoEξ × R. Let
In−1(fξ) :=
∫Eξ
fξ(x′) dx ′ .
Then
P(f ) = c(n)
∫Sn−1
In−1(fξ) dHn−1(ξ) .
A Cauchy type formula
Let f be log–concave, ξ ∈ Sn−1 and Eξ = ξ⊥. We define theprojection of f onto Eξ as
fξ(x′) := sup
t∈Rf (x ′ + tξ) , x ′ ∈ Eξ .
Put simply, the graph of fξ is the projection of the graph of f ontoEξ × R.
Let
In−1(fξ) :=
∫Eξ
fξ(x′) dx ′ .
Then
P(f ) = c(n)
∫Sn−1
In−1(fξ) dHn−1(ξ) .
A Cauchy type formula
Let f be log–concave, ξ ∈ Sn−1 and Eξ = ξ⊥. We define theprojection of f onto Eξ as
fξ(x′) := sup
t∈Rf (x ′ + tξ) , x ′ ∈ Eξ .
Put simply, the graph of fξ is the projection of the graph of f ontoEξ × R. Let
In−1(fξ) :=
∫Eξ
fξ(x′) dx ′ .
Then
P(f ) = c(n)
∫Sn−1
In−1(fξ) dHn−1(ξ) .
A Cauchy type formula
Let f be log–concave, ξ ∈ Sn−1 and Eξ = ξ⊥. We define theprojection of f onto Eξ as
fξ(x′) := sup
t∈Rf (x ′ + tξ) , x ′ ∈ Eξ .
Put simply, the graph of fξ is the projection of the graph of f ontoEξ × R. Let
In−1(fξ) :=
∫Eξ
fξ(x′) dx ′ .
Then
P(f ) = c(n)
∫Sn−1
In−1(fξ) dHn−1(ξ) .
Isoperimetric inequalities
I Using the definition of P(f ) and Prekopa–Leindler inequalitywe find:
P(f ) ≥ c(n)I (f ) + Ent(f ) , (1)
where: Ent(f ) =
∫Rn
f log(f ) dx − I (f ) log(I (f )) ,
and equality holds iff f = χB (up to a translation).
I The Sobolev inequality in Rn for functions of boundedvariation yelds
P(f ) ≥ nω1/nn ‖f ‖ n
n−1, (2)
where ωn is the volume of the unit ball. Equality holds ifff = χB (up to translation).
Isoperimetric inequalities
I Using the definition of P(f ) and Prekopa–Leindler inequalitywe find:
P(f ) ≥ c(n)I (f ) + Ent(f ) , (1)
where: Ent(f ) =
∫Rn
f log(f ) dx − I (f ) log(I (f )) ,
and equality holds iff f = χB (up to a translation).
I The Sobolev inequality in Rn for functions of boundedvariation yelds
P(f ) ≥ nω1/nn ‖f ‖ n
n−1, (2)
where ωn is the volume of the unit ball. Equality holds ifff = χB (up to translation).
Isoperimetric inequalities
I Using the definition of P(f ) and Prekopa–Leindler inequalitywe find:
P(f ) ≥ c(n)I (f ) + Ent(f ) , (1)
where: Ent(f ) =
∫Rn
f log(f ) dx − I (f ) log(I (f )) ,
and equality holds iff f = χB (up to a translation).
I The Sobolev inequality in Rn for functions of boundedvariation yelds
P(f ) ≥ nω1/nn ‖f ‖ n
n−1, (2)
where ωn is the volume of the unit ball. Equality holds ifff = χB (up to translation).
Isoperimetric inequalities
I Using the definition of P(f ) and Prekopa–Leindler inequalitywe find:
P(f ) ≥ c(n)I (f ) + Ent(f ) , (1)
where: Ent(f ) =
∫Rn
f log(f ) dx − I (f ) log(I (f )) ,
and equality holds iff f = χB (up to a translation).
I The Sobolev inequality in Rn for functions of boundedvariation yelds
P(f ) ≥ nω1/nn ‖f ‖ n
n−1, (2)
where ωn is the volume of the unit ball. Equality holds ifff = χB (up to translation).
Isoperimetric inequalities
I Using the definition of P(f ) and Prekopa–Leindler inequalitywe find:
P(f ) ≥ c(n)I (f ) + Ent(f ) , (1)
where: Ent(f ) =
∫Rn
f log(f ) dx − I (f ) log(I (f )) ,
and equality holds iff f = χB (up to a translation).
I The Sobolev inequality in Rn for functions of boundedvariation yelds
P(f ) ≥ nω1/nn ‖f ‖ n
n−1, (2)
where ωn is the volume of the unit ball. Equality holds ifff = χB (up to translation).
Isoperimetric inequalities
I Using the definition of P(f ) and Prekopa–Leindler inequalitywe find:
P(f ) ≥ c(n)I (f ) + Ent(f ) , (1)
where: Ent(f ) =
∫Rn
f log(f ) dx − I (f ) log(I (f )) ,
and equality holds iff f = χB (up to a translation).
I The Sobolev inequality in Rn for functions of boundedvariation yelds
P(f ) ≥ nω1/nn ‖f ‖ n
n−1, (2)
where ωn is the volume of the unit ball. Equality holds ifff = χB (up to translation).
Isoperimetric inequalities
I Using the definition of P(f ) and Prekopa–Leindler inequalitywe find:
P(f ) ≥ c(n)I (f ) + Ent(f ) , (1)
where: Ent(f ) =
∫Rn
f log(f ) dx − I (f ) log(I (f )) ,
and equality holds iff f = χB (up to a translation).
I The Sobolev inequality in Rn for functions of boundedvariation yelds
P(f ) ≥ nω1/nn ‖f ‖ n
n−1, (2)
where ωn is the volume of the unit ball.
Equality holds ifff = χB (up to translation).
Isoperimetric inequalities
I Using the definition of P(f ) and Prekopa–Leindler inequalitywe find:
P(f ) ≥ c(n)I (f ) + Ent(f ) , (1)
where: Ent(f ) =
∫Rn
f log(f ) dx − I (f ) log(I (f )) ,
and equality holds iff f = χB (up to a translation).
I The Sobolev inequality in Rn for functions of boundedvariation yelds
P(f ) ≥ nω1/nn ‖f ‖ n
n−1, (2)
where ωn is the volume of the unit ball. Equality holds ifff = χB (up to translation).