Gláucio Terra
Departamento de Matemática IME - USP
October 21, 2019
Recall: Lipschitz constant
Given metric spacesX andY andf :X →Y Lipschitz, Lipf :=sup{dY f(x),f(y)
dX(x,y) |x 6=y ∈X}, is calledLipschitz constantoff.
Theorem (McShane’s lemma; 5.1)
Let A⊂Rnand f :A→Ra Lipschitz map. Define F :Rn→Rby:
F(x) :=inf{f(a) +Lipf· kx−ak |a∈A}. (1) Then F extends f andLipF =Lipf .
Recall: Lipschitz constant
Given metric spacesX andY andf :X →Y Lipschitz, Lipf :=sup{dY f(x),f(y)
dX(x,y) |x 6=y ∈X}, is calledLipschitz constantoff.
Theorem (McShane’s lemma; 5.1)
Let A⊂Rnand f :A→Ra Lipschitz map. Define F :Rn→Rby:
F(x) :=inf{f(a) +Lipf· kx−ak |a∈A}. (1) Then F extends f andLipF =Lipf .
Theorem (Kirszbraun’s theorem; 5.2)
Let A⊂Rnand f :A→Rma Lipschitz map. Then there exists a Lipschitz extension f :Rn→Rmof f such thatLipF =Lipf .
Motivation
IfΩis an open subset ofRnandX ∈C1c(Ω,Rn), Z
Ω
divXdLn =0.
Ifu∈C1(Ω)andϕ∈C1c(Ω,Rn), the previous equality applied to X =uϕyields theelementary Gauss-Green’s identity in divergence form:
Z
Ω
h∇u, ϕidLn=− Z
Ω
udivϕdLn.
Motivation
IfΩis an open subset ofRnandX ∈C1c(Ω,Rn), Z
Ω
divXdLn =0.
Ifu∈C1(Ω)andϕ∈C1c(Ω,Rn), the previous equality applied to X =uϕyields theelementary Gauss-Green’s identity in divergence form:
Z
Ω
h∇u, ϕidLn=− Z
Ω
udivϕdLn.
Definition (weak derivatives and gradients; 5.3)
LetΩbe an open subset ofRnandu ∈L1loc(Ln|Ω). We say that:
1 For 1≤i≤n,uhasweak i-th partial derivative vi ∈L1loc(Ln|Ω)if
∀ϕ∈C∞c (Ω), Z
Ω
viϕdLn =− Z
Ω
u∂ϕ
∂xi dLn.
2 uhasweak gradient v ∈L1loc(Ln|Ω,Rn)if∀ϕ∈C∞c (Ω,Rn), Z
Ω
hv, ϕidLn =− Z
Ω
udivϕdLn. (2)
Definition (weak derivatives and gradients; 5.3)
LetΩbe an open subset ofRnandu ∈L1loc(Ln|Ω). We say that:
1 For 1≤i≤n,uhasweak i-th partial derivative vi ∈L1loc(Ln|Ω)if
∀ϕ∈C∞c (Ω), Z
Ω
viϕdLn =− Z
Ω
u∂ϕ
∂xi dLn.
2 uhasweak gradient v ∈L1loc(Ln|Ω,Rn)if∀ϕ∈C∞c (Ω,Rn), Z
Ω
hv, ϕidLn =− Z
Ω
udivϕdLn. (2)
Exercise (weak gradients, bis; 5.4)
Weak gradients may be also characterized by means of Gauss-Green identity in gradient form. That is, letΩbe an open subset ofRnand u ∈L1loc(Ln|Ω); then u admits weak gradient v ∈L1loc(Ln|Ω,Rn) iff∀ϕ∈C∞c (Ω),
Z
Ω
ϕvdLn=− Z
Ω
u∇ϕdLn. (3)
Exercise (5.5)
LetΩbe an open subset ofRn, u∈L1loc(Ln|Ω)and1≤i≤n. If there exists ∂∂xwu
i ∈L1loc(Ln|Ω), then∀ϕ∈C1c(Ω), Z
Ω
∂wu
∂xi ϕdLn=− Z
Ω
u∂ϕ
∂xi dLn.
Exercise (weak gradients, bis; 5.4)
Weak gradients may be also characterized by means of Gauss-Green identity in gradient form. That is, letΩbe an open subset ofRnand u ∈L1loc(Ln|Ω); then u admits weak gradient v ∈L1loc(Ln|Ω,Rn) iff∀ϕ∈C∞c (Ω),
Z
Ω
ϕvdLn=− Z
Ω
u∇ϕdLn. (3)
Exercise (5.5)
LetΩbe an open subset ofRn, u∈L1loc(Ln|Ω)and1≤i≤n. If there exists ∂∂xwu
i ∈L1loc(Ln|Ω), then∀ϕ∈C1c(Ω), Z
Ω
∂wu
∂xi ϕdLn=− Z
Ω
u∂ϕ
∂xi dLn.
Proposition (vanishing weak gradient)
LetΩ⊂Rnbe a connected open set and u∈L1loc(Ln|Ω)such that
∀ϕ∈C∞c (Ω),R
Ωu∇ϕdLn=0. Then u coincidesLn-a.e. with a constant function.
Definition (5.8)
LetΩbe an open subset ofRn,u : Ω→Rand 1≤p≤ ∞. We say that
1 uis a(1,p)-Sobolev functionifu ∈Lp(Ln|Ω)and,∀1≤i ≤n,u has weak partial derivatives ∂x∂u
i ∈Lp(Ln|Ω). Notation: W1,p(Ω).
2 uis alocal(1,p)-Sobolev functionifu∈Lploc(Ln|Ω)and,
∀1≤i ≤n,u has weak partial derivatives ∂x∂u
i ∈Lploc(Ln|Ω).
Notation: W1,ploc(Ω).
Notation
Letu:Rn→Randτ ∈Sn−1. Forh∈R\ {0}, we denote by τhu:Rn →Rthe incremental ratio ofuin the directionτ:
τhu(x) := u(x +hτ)−u(x)
h .
By the invariance of the Lebesgue measure under translations, if u ∈L1loc(Rn),v :Rn→RboundedLn-measurable with compact support andh∈R\ {0}:
Z
u(x +hτ)v(x)dLn(x) = Z
u(x)v(x −hτ)dLn(x),
hence Z
τhu(x)v(x)dLn(x) =− Z
u(x)τ−hv(x)dLn(x). (4)
Notation
Letu:Rn→Randτ ∈Sn−1. Forh∈R\ {0}, we denote by τhu:Rn →Rthe incremental ratio ofuin the directionτ:
τhu(x) := u(x +hτ)−u(x)
h .
By the invariance of the Lebesgue measure under translations, if u ∈L1loc(Rn),v :Rn→RboundedLn-measurable with compact support andh∈R\ {0}:
Z
u(x +hτ)v(x)dLn(x) = Z
u(x)v(x −hτ)dLn(x),
hence Z
τhu(x)v(x)dLn(x) =− Z
u(x)τ−hv(x)dLn(x). (4)
Proposition (5.9 )
Let f :Rn→Rbe a Lipschitz function. Then f ∈W1,∞loc (Rn).
Fréchet Differentiability
LetU⊂Rnopen andf :U→R. Recall thatf is differentiable at x0∈U in the sense of Fréchet if there existsA∈L(Rn,R)such that
h→0lim
f(x0+h)−f(x0)−A·h
khk =0.
If that is the case,f has first order partial derivatives atx0,Asatisfying the above condition is unique and coincides withh∇f(x0),·i:Rn→R; Ais calledFréchet derivativeoff atx0and denoted by Df(x0).
Fréchet Differentiability
LetU⊂Rnopen andf :U→R. Recall thatf is differentiable at x0∈U in the sense of Fréchet if there existsA∈L(Rn,R)such that
h→0lim
f(x0+h)−f(x0)−A·h
khk =0.
If that is the case,f has first order partial derivatives atx0,Asatisfying the above condition is unique and coincides withh∇f(x0),·i:Rn→R; Ais calledFréchet derivativeoff atx0and denoted by Df(x0).
Exercise (characterization of Fréchet differentiability; 5.10)
Let U ⊂Rnopen and f :U→R. Then f is differentiable at x0∈U iffthere exists A∈L(Rn,R)such that
lim
t→0+
f(x0+tv)−f(x0)
t =A·v
uniformly in v ∈Sn−1. If so A=Df(x0).
Exercise (weak gradients under scaling and translations; 5.11)
Let x ∈Rn, h>0, T :Rn→Rngiven byτ 7→x+hτ and u∈W1,1loc(Rn).
Then u◦T ∈W1,1loc(Rn)and∇w(u◦T)(τ) =h∇wu(x +hτ).
Exercise (characterization of Fréchet differentiability; 5.10)
Let U ⊂Rnopen and f :U→R. Then f is differentiable at x0∈U iffthere exists A∈L(Rn,R)such that
lim
t→0+
f(x0+tv)−f(x0)
t =A·v
uniformly in v ∈Sn−1. If so A=Df(x0).
Exercise (weak gradients under scaling and translations; 5.11)
Let x ∈Rn, h>0, T :Rn→Rngiven byτ 7→x+hτ and u∈W1,1loc(Rn).
Then u◦T ∈W1,1loc(Rn)and∇w(u◦T)(τ) =h∇wu(x +hτ).
Theorem (Rademacher’s theorem; 5.12)
Let f :Rn→Rbe Lipschitz. Then f is differentiable in the sense of FréchetLn-a.e. and∇f =∇wf Ln-a.e.
Exercise (5.13)
Let f :Rn→Rbe Lipschitz. The set Df of points where f is differentiable in the sense of Fréchet is Borel measurable and Df :Df →L(Rn,R)is Borelian.
Corollary (5.14)
IfΩ⊂Rn open and f : Ω→Ris locally Lipschitz, then f isLn|Ω-a.e.
differentiable in the sense of Fréchet.
Theorem (Rademacher’s theorem; 5.12)
Let f :Rn→Rbe Lipschitz. Then f is differentiable in the sense of FréchetLn-a.e. and∇f =∇wf Ln-a.e.
Exercise (5.13)
Let f :Rn→Rbe Lipschitz. The set Df of points where f is differentiable in the sense of Fréchet is Borel measurable and Df :Df →L(Rn,R)is Borelian.
Corollary (5.14)
IfΩ⊂Rn open and f : Ω→Ris locally Lipschitz, then f isLn|Ω-a.e.
differentiable in the sense of Fréchet.
Theorem (Rademacher’s theorem; 5.12)
Let f :Rn→Rbe Lipschitz. Then f is differentiable in the sense of FréchetLn-a.e. and∇f =∇wf Ln-a.e.
Exercise (5.13)
Let f :Rn→Rbe Lipschitz. The set Df of points where f is differentiable in the sense of Fréchet is Borel measurable and Df :Df →L(Rn,R)is Borelian.
Corollary (5.14)
IfΩ⊂Rn open and f : Ω→Ris locally Lipschitz, then f isLn|Ω-a.e.
differentiable in the sense of Fréchet.
Corollary (5.16)
IfΩ⊂Rn open and f : Ω→Rmis locally Lipschitz, then f isLn|Ω-a.e.
differentiable in the sense of Fréchet.
Corollary (5.17)
1 Let f :Rn→Rmbe locally Lipschitz and Zf :={x ∈Rn|f(x) =0}.
ThenDf(x) =0forLn-a.e. x ∈Zf.
2 Let f,g :Rn→Rnbe locally Lipschitz and Y :={x ∈Rn |g f(x)
=x}. ThenDg f(x)
◦Df(x) =idRn for Ln-a.e. x ∈Y .
Corollary (5.16)
IfΩ⊂Rn open and f : Ω→Rmis locally Lipschitz, then f isLn|Ω-a.e.
differentiable in the sense of Fréchet.
Corollary (5.17)
1 Let f :Rn→Rmbe locally Lipschitz and Zf :={x ∈Rn|f(x) =0}.
ThenDf(x) =0forLn-a.e. x ∈Zf.
2 Let f,g :Rn→Rnbe locally Lipschitz and Y :={x ∈Rn |g f(x)
=x}. ThenDg f(x)
◦Df(x) =idRn for Ln-a.e. x ∈Y .
Definition (5.18)
Let V and W be finite-dimensional Hilbert spaces.
1 A linear mapO :V→W is called anorthogonal injectionif
∀x,y ∈V,hO·x,O·yi=hx,yi. Notation:O(V,W).
2 LetT :V→W be a linear map. We denote byT∗ theadjointofT with respect to the inner products on V and W. If V=W and T =T∗, we callT self-adjointorsymmetric(notation: Sym(V)).
3 We say that a linear mapT :V→V ispositiveif it is symmetric and∀x ∈V,hT ·x,xi ≥0.
Theorem (5.20)
LetVandWbe finite-dimensional Hilbert spaces and L:V→Wbe a linear map.
1 Ifdim V≤dim W, there exists a positive S∈Sym(V)and O∈O(V,W)such that
L=O◦S.
Moreover, in the above decomposition, S ∈Sym(V)positive is unique, and so is O∈O(V,W)if L is injective.
2 Ifdim V≥dim W, there exists a positive S∈Sym(W)and O∈O(W,V)such that
L=S◦O∗.
Moreover, in the above decomposition, S ∈Sym(W)positive is unique, and so is O∈O(W,V)if L is surjective.
Theorem (5.20)
LetVandWbe finite-dimensional Hilbert spaces and L:V→Wbe a linear map.
1 Ifdim V≤dim W, there exists a positive S∈Sym(V)and O∈O(V,W)such that
L=O◦S.
Moreover, in the above decomposition, S ∈Sym(V)positive is unique, and so is O∈O(V,W)if L is injective.
2 Ifdim V≥dim W, there exists a positive S∈Sym(W)and O∈O(W,V)such that
L=S◦O∗.
Moreover, in the above decomposition, S ∈Sym(W)positive is unique, and so is O∈O(W,V)if L is surjective.
Definition (Jacobian of a linear map; 5.21)
Let V and W be finite-dimensional Hilbert spaces andL∈L(V,W), with polar decompositionO◦Sif dim V≤dim W orS◦O∗if dim V>dim W.
We define theJacobianofLby:
JLK:=|detS|.
Remark (5.22)
1 Note thatJLKis well-defined, by the uniqueness ofSin the polar decomposition.
2 It is clear that
JLK=JL∗K= (√
detL∗L if dim V≤dim W
√
detLL∗ if dim V≥dim W.
Definition (Jacobian of a linear map; 5.21)
Let V and W be finite-dimensional Hilbert spaces andL∈L(V,W), with polar decompositionO◦Sif dim V≤dim W orS◦O∗if dim V>dim W.
We define theJacobianofLby:
JLK:=|detS|.
Remark (5.22)
1 Note thatJLKis well-defined, by the uniqueness ofSin the polar decomposition.
2 It is clear that
JLK=JL∗K= (√
detL∗L if dim V≤dim W
√
detLL∗ if dim V≥dim W.
Theorem (5.23)
LetVandWbe finite-dimensional Hilbert spaces with n=dim V≤dim W=m. If L∈L(V,W), then
JLK=
s X
B∈µ(m,n)
(detB)2,
whereµ(m,n)is the set of n×n minors in some matrix representation of L with respect to orthonormal bases onVandW.
Definition (5.25)
Letf :Rn→Rmbe Lipschitz.
We define, for each pointx wheref is differentiable, theJacobianoff atx,
Jf(x) :=JDf(x)K.
Note that Jf is a Borelian function defined on the complement of a Borel subset ofRnofLn-null measure.
Notation
For a Lipschitz mapf :Rn →Rm, we will use henceforth the following notation:
Df :={x ∈Rn| ∃Df(x)};
Jf+:={x ∈Df |Jf(x)>0};
Jf0:={x ∈Df |Jf(x) =0}.
Lemma (5.27)
If L:Rn→Rm is linear and n≤m, then∀A⊂Rn, Hn L(A)
=JLKLn(A). (5)
Exercise (5.28)
Let T ∈L(Rn,Rm), n≤m.
1 If R ∈L(Rn), thenJT ◦RK=JTKJRK.
2
JTK≤ kTkn. If T is 1-1, thenkT−1k−n≤JTK≤ kTkn.
3 If m≤k and R∈L(Rm,Rk), thenJR◦TK≤ kRknJTK. If R is 1-1, thenkR−1k−nJTK≤JR◦TK≤ kRknJTK.
Lemma (5.27)
If L:Rn→Rm is linear and n≤m, then∀A⊂Rn, Hn L(A)
=JLKLn(A). (5)
Exercise (5.28)
Let T ∈L(Rn,Rm), n≤m.
1 If R ∈L(Rn), thenJT ◦RK=JTKJRK.
2
JTK≤ kTkn. If T is 1-1, thenkT−1k−n≤JTK≤ kTkn.
3 If m≤k and R∈L(Rm,Rk), thenJR◦TK≤ kRknJTK. If R is 1-1, thenkR−1k−nJTK≤JR◦TK≤ kRknJTK.
Lemma (5.27)
If L:Rn→Rm is linear and n≤m, then∀A⊂Rn, Hn L(A)
=JLKLn(A). (5)
Exercise (5.28)
Let T ∈L(Rn,Rm), n≤m.
1 If R ∈L(Rn), thenJT ◦RK=JTKJRK.
2
JTK≤ kTkn. If T is 1-1, thenkT−1k−n≤JTK≤ kTkn.
3 If m≤k and R∈L(Rm,Rk), thenJR◦TK≤ kRknJTK. If R is 1-1, thenkR−1k−nJTK≤JR◦TK≤ kRknJTK.
Lemma (5.27)
If L:Rn→Rm is linear and n≤m, then∀A⊂Rn, Hn L(A)
=JLKLn(A). (5)
Exercise (5.28)
Let T ∈L(Rn,Rm), n≤m.
1 If R ∈L(Rn), thenJT ◦RK=JTKJRK.
2
JTK≤ kTkn. If T is 1-1, thenkT−1k−n≤JTK≤ kTkn.
3 If m≤k and R∈L(Rm,Rk), thenJR◦TK≤ kRknJTK. If R is 1-1, thenkR−1k−nJTK≤JR◦TK≤ kRknJTK.
Lemma (5.29)
Let f :Rn→Rmbe Lipschitz, with n≤m, and A⊂RnLn-measurable.
Then:
1 f(A)isHn-measurable.
2 The function N(f|A) :Rm→[0,∞]given by y 7→H0(A∩f−1{y})is Hn-measurable.
3 R
RmH0(A∩f−1{y})dHn(y)≤(Lipf)nLn(A).
Definition (multiplicity function;5.30)
With the notation from the previous lemma,
N(f|A) :y 7→H0(A∩f−1{y})is called themultiplicity functionoff|A.
Remark (5.31)
IfX is a complete, separable metric space,Y a Hausdorff topological space,µa Borel measure onY andf :X →Y continuous, then
∀A∈BX,f(A)isµ-measurable.
Definition (5.32)
Letf :Rn→Rmbe a Lipschitz map withn≤mandt >1. We say that (E,S)is at-linearizationforf ifE ∈BRn andS ∈Sym(n)∩GL(Rn) satisfy:
i) ∀x ∈E,f is differentiable atx and Jf(x)>0;
ii) ∀x,y ∈E,t−1kS·x−S·yk ≤ kf(x)−f(y)k ≤tkS·x −S·yk;
iii) ∀x ∈E,∀v ∈Rn,t−1kS·vk ≤ kDf(x)·vk ≤tkS·vk.
Proposition (5.33)
Let E ∈BRn such that condition i) in the previous definition holds and S ∈Sym(n)∩GL(Rn). Then(E,S)is a t-linearization for f ifff|E is 1-1 with Lipschitz inverse and satisfies:
ii’) Lipf|E◦S−1≤t andLipS◦(f|E)−1≤t;
iii’) ∀x ∈E ,kDf(x)◦S−1k ≤t andkS◦Df(x)−1k ≤t.
Corollary (5.34)
Let f :Rn→Rmbe a Lipschitz map with n≤m, t >1and(E,S)a t-linearization for f . Then∀x ∈E ,
t−n|detS| ≤Jf(x)≤tn|detS|. (6)
Proposition (5.33)
Let E ∈BRn such that condition i) in the previous definition holds and S ∈Sym(n)∩GL(Rn). Then(E,S)is a t-linearization for f ifff|E is 1-1 with Lipschitz inverse and satisfies:
ii’) Lipf|E◦S−1≤t andLipS◦(f|E)−1≤t;
iii’) ∀x ∈E ,kDf(x)◦S−1k ≤t andkS◦Df(x)−1k ≤t.
Corollary (5.34)
Let f :Rn→Rmbe a Lipschitz map with n≤m, t >1and(E,S)a t-linearization for f . Then∀x ∈E ,
t−n|detS| ≤Jf(x)≤tn|detS|. (6)
Theorem (5.35)
Let f :Rn→Rmbe a Lipschitz map with n≤m, t >1and
Jf+={x ∈Rn| ∃Df(x)and Jf(x)>0}. Then there exists a countable disjoint family(Ek)k∈NinBRn such that Jf+=∪_k∈NEk and,∀k ∈N, there exists Sk ∈Sym(n)∩GL(Rn)such that(Ek,Sk)is a
t-linearization for f .
Theorem (5.36)
Let f :Rn→Rmbe Lipschitz, n≤m. Then, for all A∈σ(Ln), Z
A
JfdLn = Z
Rm
H0(A∩f−1{y})dHn(y).
Corollary (5.37)
If f :Rn→Rm is Lipschitz, n≤m, then forHn-a.e. y ∈Rm, f−1{y}is countable.
Corollary (5.38)
Let f :Rn→Rmbe Lipschitz, n≤m. Then for all g :Rn →R Ln-measurable with g≥0or g summable,
Z
Rn
gJfdLn= Z
Rm
X
x∈f−1{y}
g(x)
dHn(y).
Corollary (5.39)
Let f :Rn→Rmbe Lipschitz 1-1, n≤m.
1 ∀A∈σ(Ln),Hn f(A)
=R
AJfdLn. In particular, we have
f#(Ln
x
Jf) =Hnx
Imf (7)(equality as Borel regular outer measures onRm).
2 If g:Rn→RisLn-measurable with g≥0or g ∈L1(Ln), then R
Imf g◦f−1dHn=R
RngJfdLn. In particular, if g :Imf →[0,∞]is Borelian, then
Z
Imf
gdHn= Z
g◦fJfdLn. (8)
Corollary (5.39)
Let f :Rn→Rmbe Lipschitz 1-1, n≤m.
1 ∀A∈σ(Ln),Hn f(A)
=R
AJfdLn. In particular, we have
f#(Ln
x
Jf) =Hnx
Imf (7)(equality as Borel regular outer measures onRm).
2 If g:Rn→RisLn-measurable with g≥0or g ∈L1(Ln), then R
Imf g◦f−1dHn=R
RngJfdLn. In particular, if g :Imf →[0,∞]is Borelian, then
Z
Imf
gdHn= Z
g◦fJfdLn. (8)
Example (5.40)
1 (length of a curve) Let−∞<a<b<∞andγ : [a,b]→Rmbe Lipschitz 1-1. We may extendγ to a Lipschitz function onR, which we still denote byγ. Then
Z b a
kγ0(t)kdt=H1 γ([a,b]) .
2 (area of a graph) Letg :Rn→Rbe Lipschitz andf :Rn →Rn+1 be given byf(x) := x,g(x)
.
For eachU⊂Rn open, the “surface area” of the graph ofg over U,Γ = Γ(g;U) :={ x,g(x)
|x ∈U}=f(U), is given by:
Hn(Γ) = Z
U
JfdLn= Z
U
q
1+k∇g(x)k2dx.
Example (5.40)
1 (length of a curve) Let−∞<a<b<∞andγ : [a,b]→Rmbe Lipschitz 1-1. We may extendγ to a Lipschitz function onR, which we still denote byγ. Then
Z b a
kγ0(t)kdt=H1 γ([a,b]) .
2 (area of a graph) Letg :Rn→Rbe Lipschitz andf :Rn →Rn+1 be given byf(x) := x,g(x)
.
For eachU⊂Rn open, the “surface area” of the graph ofg over U,Γ = Γ(g;U) :={ x,g(x)
|x ∈U}=f(U), is given by:
Hn(Γ) = Z
U
JfdLn= Z
U
q
1+k∇g(x)k2dx.
Exercise (5.41)
The area formula and its corollaries remain valid for locally Lipschitz maps defined on open subsets ofRn.
R
Exercise (5.42)
For any smooth embedded k -Riemannian submanifoldM⊂Rn, the measure induced by the Riemannian metric onM(i.e. the Lebesgue measure ofM) coincides with the traceHk|M. Conclude that
H-dimM =k and, ifMis closed (i.e. topologically closed),Hk
x
Misa Radon measure onRn.
Theorem (5.48)
Let f :Rn→Rmbe Lipschitz, n≥m. Then, for eachLn-measurable A⊂Rn,
Z
A
JfdLn= Z
Rm
Hn−m(A∩f−1{y})dLm(y). (9)
Remark (5.49)
1 Iff :Rn≡Rm×Rn−m →Rm is the projection on the first factor, we have Jf ≡1 and the coarea formula reduces to Fubini-Tonelli’s theorem.
2 Ifn=m, the coarea formula coincides with the area formula 22.
3 If we take the Borel set
A:= (Rn\Df)∪Jf0={x ∈Rn|@Df(x)or Jf(x) =0}in the coarea formula, we conclude thatHn−m(A∩f−1{y}) =0 for Lm-a.e. y ∈Rm.
That may be interpreted as a measure theoretic version of Morse-Sard’s theorem:Lm-a.e. y ∈Rmis a measure theoretic
“regular value” off, in the sense that, up toHn−mnull sets,f−1{y} lies in the setJf+of points where Df has maximal rank.
Remark (5.49)
1 Iff :Rn≡Rm×Rn−m →Rm is the projection on the first factor, we have Jf ≡1 and the coarea formula reduces to Fubini-Tonelli’s theorem.
2 Ifn=m, the coarea formula coincides with the area formula 22.
3 If we take the Borel set
A:= (Rn\Df)∪Jf0={x ∈Rn|@Df(x)or Jf(x) =0}in the coarea formula, we conclude thatHn−m(A∩f−1{y}) =0 for Lm-a.e. y ∈Rm.
That may be interpreted as a measure theoretic version of Morse-Sard’s theorem:Lm-a.e. y ∈Rmis a measure theoretic
“regular value” off, in the sense that, up toHn−mnull sets,f−1{y} lies in the setJf+of points where Df has maximal rank.
Remark (5.49)
1 Iff :Rn≡Rm×Rn−m →Rm is the projection on the first factor, we have Jf ≡1 and the coarea formula reduces to Fubini-Tonelli’s theorem.
2 Ifn=m, the coarea formula coincides with the area formula 22.
3 If we take the Borel set
A:= (Rn\Df)∪Jf0={x ∈Rn|@Df(x)or Jf(x) =0}in the coarea formula, we conclude thatHn−m(A∩f−1{y}) =0 for Lm-a.e. y ∈Rm.
That may be interpreted as a measure theoretic version of Morse-Sard’s theorem:Lm-a.e. y ∈Rmis a measure theoretic
“regular value” off, in the sense that, up toHn−mnull sets,f−1{y} lies in the setJf+of points where Df has maximal rank.
Remark (5.49)
1 Iff :Rn≡Rm×Rn−m →Rm is the projection on the first factor, we have Jf ≡1 and the coarea formula reduces to Fubini-Tonelli’s theorem.
2 Ifn=m, the coarea formula coincides with the area formula 22.
3 If we take the Borel set
A:= (Rn\Df)∪Jf0={x ∈Rn|@Df(x)or Jf(x) =0}in the coarea formula, we conclude thatHn−m(A∩f−1{y}) =0 for Lm-a.e. y ∈Rm.
That may be interpreted as a measure theoretic version of Morse-Sard’s theorem:Lm-a.e. y ∈Rmis a measure theoretic
“regular value” off, in the sense that, up toHn−mnull sets,f−1{y} lies in the setJf+of points where Df has maximal rank.
Corollary (curvilinear Fubini-Tonelli’s theorem; 5.50)
Let f :Rn→Rmbe Lipschitz, n≥m. Then for all g :Rn →R Ln-measurable with g≥0or g summable,
Z
Rn
gJfdLn= Z
Rm
Z
f−1{y}
g(x)dHn−m(x)
dLm(y), (10) meaning that the iterated integrals in second member make sense and the equality holds.
Exercise (Coarea Formula for locally Lipschitz maps; 5.51)
The coarea formula and its corollary remain valid for locally Lipschitz maps defined on open subsets ofRn.
Proposition (polar coordinates; 5.52)
If g :Rn →RisLn-measurable and g≥0or g∈L1(Ln), then Z
Rn
gdLn= Z ∞
0
Z
∂B(0,r)
gdHn−1
dr. (11)
Proposition (5.53)
LetΩ⊂Rnopen and f : Ω→Rbe locally Lipschitz. Then Z
Ω
k∇fkdLn= Z ∞
−∞
Hn−1({f =t})dt.
Proposition (polar coordinates; 5.52)
If g :Rn →RisLn-measurable and g≥0or g∈L1(Ln), then Z
Rn
gdLn= Z ∞
0
Z
∂B(0,r)
gdHn−1
dr. (11)
Proposition (5.53)
LetΩ⊂Rnopen and f : Ω→Rbe locally Lipschitz. Then Z
Ω
k∇fkdLn= Z ∞
−∞
Hn−1({f =t})dt.
Theorem (6.45)
Let n ≥2, f :Rn−1→RLipschitz andΩ :=epiS f (hence∂Ω =grf ).
Then
i) Hn−1
x
∂Ωis a Radon measure onRn;ii) there exists a Borel measurable unit vector fieldν :∂Ω→Rn, unique up toHn−1
x
∂Ω-null sets, such that, for allϕ∈C1c(Rn),Z
Ω
∇ϕdLn= Z
∂Ω
ϕνdHn−1, (12)
or, equivalently, such that, for allϕ∈C1c(Rn,Rn), Z
Ω
divϕdLn= Z
∂Ω
ϕ·νdHn−1. (13)
With the notation from the previous theorem,ν is calledouter unit normalto∂Ω.
Remark
Up toHn
x
∂Ω-null sets, on each point pointx = x0,f(x0) in∂Ω =grf whose abscissax0is a differentiability point off,
ν(x) = ∇f(x0),−1
p1+k∇f(x0)k2. (14) In particular, iff is C1,νcoincides with the usual outer unit normal from Differential Geometry.
With the notation from the previous theorem,ν is calledouter unit normalto∂Ω.
Remark
Up toHn
x
∂Ω-null sets, on each point pointx = x0,f(x0) in∂Ω =grf whose abscissax0is a differentiability point off, ν(x) = ∇f(x0),−1
p1+k∇f(x0)k2. (14) In particular, iff is C1,νcoincides with the usual outer unit normal from Differential Geometry.