Yan Sheng's site A math blog

Measure Theory (VIII): The Lebesgue Measure on Rn\bb R^n

One of the most important applications of the Carathéodory construction of measures is to extend the notion of “length” or “volume” in Rn\bb R^n to the Lebesgue measure, yielding a more powerful integration theory. In this post, we cover the basic properties of the Lebesgue measure.

The Lebesgue measure

Recall that for a=(a1,,an)a=(a_1,\ldots,a_n), b=(b1,,bn)b=(b_1,\ldots,b_n), with akbk-\infty\leq a_k\leq b_k\leq\infty, the half-open interval in Rn\bb R^n is the set

[a,b)=[a1,b1)××[an,bn)Rn.[a,b)=[a_1,b_1)\times\cdots\times[a_n,b_n)\subseteq\bb R^n.

We write Hn\mc H_n for the collection of all half-open intervals in Rn\bb R^n. We have shown that Hn\mc H_n generates the Borel σ\sigma-algebra, ie. σ(Hn)=Bn\sigma(\mc H_n)=\mc B_n.

Define λ0:Hn[0,]\lambda_0:\mc H_n\to[0,\infty] by

λ0[a,b)=k=1n(bkak).\lambda_0[a,b)=\prod_{k=1}^n(b_k-a_k).

Then the Carathéodory construction gives the Lebesgue outer measure λ\lambda^* on Rn\bb R^n, which is given by

λ(A)=inf{k=1λ0(Ik):IkHn,Ak=1Ik}.\lambda^* (A)=\inf\left\{\sum_{k=1}^\infty\lambda_0(I_k)\,:\,I_k\in\mc H_n,\,A\subseteq\bigcup_{k=1}^\infty I_k\right\}.

The λ\lambda^* -measurable sets, called Lebesgue measurable sets, form a σ\sigma-algebra Σ\Sigma, and the corresponding measure λ=λΣ\lambda=\lambda^* \vert_\Sigma is called the Lebesgue measure on Rn\bb R^n. As we have shown in the previous post, (Rn,Σ,λ)(\bb R^n,\Sigma,\lambda) is a complete measure space.

We now consider some elementary properties of the Lebesgue measure. For any ARnA\subseteq\bb R^n, xRnx\in\bb R^n, and cR\{0}c\in\bb R\backslash\{0\}, define

A+x={ω+x:ωA},cA={cω:ωA}. \begin{aligned} A+x&=\{\omega+x\,:\,\omega\in A\},\\ cA&=\{c\omega\,:\,\omega\in A\}. \end{aligned}

Geometrically, these operations are translation by xx and scaling by cc, respectively. We first show that the Lebesgue measure is translation-invariant and homogeneous under scaling.

Proposition: Let ARnA\subseteq\bb R^n, xRnx\in\bb R^n, and c>0c>0. Then

λ(A+x)=λ(A),λ(cA)=cnλ(A). \begin{aligned} \lambda^* (A+x)&=\lambda^* (A),\\ \lambda^* (cA)&=c^n\lambda^* (A). \end{aligned}

Proof: First note that for all IHnI\in\mc H_n, we have

λ0(I+x)=λ0(I),λ0(cI)=cnλ0(I). \begin{aligned} \lambda_0(I+x)&=\lambda_0(I),\\ \lambda_0(cI)&=c^n\lambda_0(I). \end{aligned}

Let (Ik)(I_k) be a countable covering of AA with half-open intervals. Then (Ik+x)(I_k+x) and (cIk)(cI_k) are coverings of A+xA+x and cAcA, respectively. Hence

λ(A+x)kλ0(Ik+x)=kλ0(Ik),λ(cA)kλ0(cIk)=cnkλ0(Ik). \begin{aligned} \lambda^* (A+x)&\leq\sum_k\lambda_0(I_k+x)\\ &=\sum_k\lambda_0(I_k),\\ \lambda^* (cA)&\leq\sum_k\lambda_0(cI_k)\\ &=c^n\sum_k\lambda_0(I_k). \end{aligned}

Taking infimum over all (Ik)(I_k) gives

λ(A+x)λ(A),λ(cA)cnλ(A). \begin{aligned} \lambda^* (A+x)&\leq\lambda^* (A),\\ \lambda^* (cA)&\leq c^n\lambda^* (A). \end{aligned}

Now the same argument with x-x replacing xx and 1c\frac1c replacing cc gives

λ(A+x)λ((A+x)x)=λ(A),λ(cA)cnλ(1c(cA))=λ(A), \begin{aligned} \lambda^* (A+x)&\geq\lambda^* ((A+x)-x)\\ &=\lambda^* (A),\\ \lambda^* (cA)&\geq c^n\lambda^* (\frac1c(cA))\\ &=\lambda^* (A), \end{aligned}

and we are done.  \qed

Corollary: Let ARnA\subseteq\bb R^n, xRnx\in\bb R^n, and c>0c>0. Then AA is Lebesgue measurable if and only if A+xA+x is Lebesgue measurable, if and only if cAcA is Lebesgue measurable.

Proof: If AA is Lebesgue measurable, then for any SRnS\subseteq\bb R^n, we have

λ(S)=λ(Sx)=λ((Sx)A)+λ((Sx)Ac)=λ(S(A+x))+λ(S(A+x)c), \begin{aligned} \lambda^* (S)&=\lambda^* (S-x)\\ &=\lambda^* ((S-x)\cap A)+\lambda^* ((S-x)\cap A^c)\\ &=\lambda^* (S\cap(A+x))+\lambda^* (S\cap(A+x)^c), \end{aligned}

so A+xA+x is Lebesgue measurable. Conversely, if A+xA+x is Lebesgue measurable, then replacing xx by x-x gives (A+x)x=A(A+x)-x=A is measurable.

Similarly,

λ(S)=cnλ(1cS)=cn(λ((1cS)A)+λ((1cS)Ac))=λ(S(cA))+λ(S(cA)c), \begin{aligned} \lambda^* (S)&=c^n\lambda^* (\frac1cS)\\ &=c^n(\lambda^* ((\frac1cS)\cap A)+\lambda^* ((\frac1cS)\cap A^c))\\ &=\lambda^* (S\cap(cA))+\lambda^* (S\cap(cA)^c), \end{aligned}

so cAcA is Lebesgue measurable. Conversely, if cAcA is Lebesgue measurable, then replacing cc by 1c\frac1c gives 1c(cA)=A\frac1c(cA)=A is measurable.  \qed

Corollary: Let f:Rn[,]f:\bb R^n\to[-\infty,\infty], xRnx\in\bb R^n, and c>0c>0. Define g(ω)=f(ωx)g(\omega)=f(\omega-x) and h(ω)=f(ωc)h(\omega)=f(\frac\omega c). Then:

  1. ff is Lebesgue measurable if and only if gg is Lebesgue measurable, if and only if hh is Lebesgue measurable;
  2. ff is Lebesgue integrable if and only if gg is Lebesgue integrable, if and only if hh is Lebesgue integrable, in which case we have
 ⁣gdλ= ⁣fdλ, ⁣hdλ=cn ⁣fdλ.\int\!g\,d\lambda=\int\!f\,d\lambda,\qquad\int\!h\,d\lambda=c^n\int\!f\,d\lambda.

Proof: (1): This follows from g1(a,]=f1(a,]+xg^{-1}(a,\infty]=f^{-1}(a,\infty]+x and h1(a,]=cf1(a,]h^{-1}(a,\infty]=cf^{-1}(a,\infty].

(2): Note that g±(ω)=f±(ωx)g^\pm(\omega)=f^\pm(\omega-x) and h±(ω)=f±(ωc)h^\pm(\omega)=f^\pm(\frac\omega c). Hence it suffices to prove the identity for f0f\geq0.

For any sequence of step functions 0φ1φ0\leq\varphi_1\leq\varphi\leq\cdots with (φk)f(\varphi_k)\to f, let ψk(ω)=φk(ωx)\psi_k(\omega)=\varphi_k(\omega-x). Then 0ψ1ψ20\leq\psi_1\leq\psi_2\leq\cdots is a sequence of step functions with (ψk)g(\psi_k)\to g. Now

 ⁣ψkdλ=jajλ(ψk1(aj))=jajλ(φk1(aj)+x)=jajλ(φk1(aj))= ⁣φkdλ. \begin{aligned} \int\!\psi_k\,d\lambda&=\sum_ja_j\lambda(\psi_k^{-1}(a_j))\\ &=\sum_ja_j\lambda(\varphi_k^{-1}(a_j)+x)\\ &=\sum_ja_j\lambda(\varphi_k^{-1}(a_j))\\ &=\int\!\varphi_k\,d\lambda. \end{aligned}

Taking kk\to\infty gives  ⁣gdλ= ⁣fdλ\int\!g\,d\lambda=\int\!f\,d\lambda.

Similarly, let ηk(ω)=φk(ωc)\eta_k(\omega)=\varphi_k(\frac\omega c), so 0η1η20\leq\eta_1\leq\eta_2\leq\cdots is a sequence of step functions with (ηk)h(\eta_k)\to h. Now

 ⁣ηkdλ=jajλ(ηk1(aj))=jajλ(cφk1(aj))=cnjajλ(φk1(aj))=cn ⁣φkdλ. \begin{aligned} \int\!\eta_k\,d\lambda&=\sum_ja_j\lambda(\eta_k^{-1}(a_j))\\ &=\sum_ja_j\lambda(c\varphi_k^{-1}(a_j))\\ &=c^n\sum_ja_j\lambda(\varphi_k^{-1}(a_j))\\ &=c^n\int\!\varphi_k\,d\lambda. \end{aligned}

Taking kk\to\infty gives  ⁣hdλ=cn ⁣fdλ\int\!h\,d\lambda=c^n\int\!f\,d\lambda.  \qed

Note that we have omitted the case of c<0c<0: the main problem here is that IHnI\in\mc H_n does not imply that I=(1)IHn-I=(-1)I\in\mc H_n. To deal with this case, we will need stronger theorems about the structure of Lebesgue measurable sets, which we will cover in the next post.

Half-open intervals

Due to the indirect nature of the Carathéodory construction, it is not obvious that the Lebesgue measure λ\lambda on half-open intervals corresponds to their volume λ0\lambda_0; indeed, it is not obvious that half-open intervals are (λ\lambda^* -)measurable at all! The proof of these two statements are technical, and will take up the rest of the post.

Lemma: Let I,I1,I2,,InHnI,I_1,I_2,\ldots,I_n\in\mc H_n be half-open intervals.

  1. If II is the disjoint union of I1,,ImI_1,\ldots,I_m, then

    λ0(I)=k=1mλ0(Ik).\lambda_0(I)=\sum_{k=1}^m\lambda_0(I_k).
  2. If II is contained in the (not necessarily disjoint) union of I1,,ImI_1,\ldots,I_m, then

    λ0(I)k=1mλ0(Ik).\lambda_0(I)\leq\sum_{k=1}^m\lambda_0(I_k).

Proof: (1): Consider the special case of a “product partition” of the form

[a,b)=i1=0m11in=0mn1[xi11,xi1+11)××[xinn,xin+1n),[a,b)=\bigcup_{i_1=0}^{m_1-1}\cdots\bigcup_{i_n=0}^{m_n-1}[x^1_{i_1},x^1_{i_1+1})\times\cdots\times[x^n_{i_n},x^n_{i_n+1}),

where ak=x0k<x1k<<xmkk=bka_k=x^k_0< x^k_1<\cdots< x^k_{m_k}=b_k. Writing the half-open intervals in the partition as Si1inS_{i_1\ldots i_n}, we have

i1,,inλ0(Si1in)=i1,,in(xi1+11xi11)(xin+1nxin1)=(i1(xi1+11xi11))(in(xin+1nxin1))=(b1a1)(bnan)=λ0[a,b), \begin{aligned} \sum_{i_1,\ldots,i_n}\lambda_0(S_{i_1\ldots i_n})&=\sum_{i_1,\ldots,i_n}(x^1_{i_1+1}-x^1_{i_1})\cdots(x^n_{i_n+1}-x^1_{i_n})\\ &=\left(\sum_{i_1}(x^1_{i_1+1}-x^1_{i_1})\right)\cdots\left(\sum_{i_n}(x^n_{i_n+1}-x^1_{i_n})\right)\\ &=(b_1-a_1)\cdots(b_n-a_n)\\ &=\lambda_0[a,b), \end{aligned}

so the statement holds in this special case.

For a general partition of I=[a,b)I=[a,b), we “extend the sides” of each half-open interval to get a sub-partition of II, as follows. For each kk, list all lower and upper endpoints of the kk-th coordinates of every IjI_j, and arrange them in increasing order as ak=x0k<x1k<<xmkk=bka_k=x^k_0< x^k_1<\cdots< x^k_{m_k}=b_k.

Now consider the product partition of II associated to these xikx^k_i; by the special case, we have

i1,,inλ0(Si1in)=λ0(I).\sum_{i_1,\ldots,i_n}\lambda_0(S_{i_1\ldots i_n})=\lambda_0(I).

On the other hand, note that for each IjI_j, the Si1inS_{i_1\ldots i_n} which intersect IjI_j also form a product partition of IjI_j, and each Si1inS_{i_1\ldots i_n} is contained in exactly one IjI_j. Hence

λ0(I)=i1,,inλ0(Si1in)=jλ0(Ij),\lambda_0(I)=\sum_{i_1,\ldots,i_n}\lambda_0(S_{i_1\ldots i_n})=\sum_j\lambda_0(I_j),

and we are done.

(2): Without loss of generality, we may replace IjI_j by IIjI\cap I_j. Then the same operation as above gives a product partition of II. However, now each Si1inS_{i_1\ldots i_n} is contained in at least one IjI_j, so

λ0(I)=i1,,inλ0(Si1in)jλ0(Ij),\lambda_0(I)=\sum_{i_1,\ldots,i_n}\lambda_0(S_{i_1\ldots i_n})\leq\sum_j\lambda_0(I_j),

as desired.  \qed

Proposition: For any half-open interval IHnI\in\mc H_n, we have λ(I)=λ0(I)\lambda^* (I)=\lambda_0(I).

Proof: We first consider the case where II is finite, ie. bounded. Since {I}\{I\} is a covering of II by half-open intervals, we have λ(I)λ0(I)\lambda^* (I)\leq\lambda_0(I); in particular, λ(I)<\lambda^* (I)<\infty. Hence, we only need to consider countable coverings (Ik)(I_k) of II by half-open intervals where kλ0(Ik)<\sum_k\lambda_0(I_k)<\infty; this implies that each IkI_k is finite.

Fix ε>0\eps>0. By shrinking II slightly, we can get IHnI’\in\mc H_n such that III\supseteq\overline{I’}, the closure of II’, and

λ0(I)λ0(I)ε.\lambda_0(I')\geq\lambda_0(I)-\eps.

Similarly, by enlarging IkI_k slightly, we can get IkHnI_k’\in\mc H_n such that IkIkI_k\subseteq I_k’^\circ, the interior of IkI_k’, and

λ0(Ik)λ0(Ik)+ε2k.\lambda_0(I_k')\leq\lambda_0(I_k)+\frac\eps{2^k}.

Now (Ik)(I_k’) is an open covering of the compact set II’, so it has a finite subcovering, say Ik1IkmI_{k_1}’\cup\cdots\cup I_{k_m}’. By the previous lemma, we have

λ0(I)λ0(I)+εi=1mλ0(Iki)+εk=1λ0(Ik)+2ε, \begin{aligned} \lambda_0(I)&\leq\lambda_0(I')+\eps\\ &\leq\sum_{i=1}^m\lambda_0(I_{k_i}')+\eps\\ &\leq\sum_{k=1}^\infty\lambda_0(I_k)+2\eps, \end{aligned}

and taking ε0+\eps\to0^+ gives kλ0(Ik)λ0(I)\sum_k\lambda_0(I_k)\geq\lambda_0(I). Now taking infimum over all (Ik)(I_k) gives λ(I)λ0(I)\lambda^* (I)\geq\lambda_0(I), as desired.

Now suppose that I=[a,b)I=[a,b) is not finite, say bk=b_k=\infty (the argument for ak=a_k=-\infty is analogous). Then

S=[a1,b1)××[ak,ak+N)××[an,bn)S=[a_1,b_1)\times\cdots\times[a_k,a_k+N)\times\cdots\times[a_n,b_n)

is contained in II, so

λ(I)λ(S)=Nik(biai).\lambda^* (I)\geq\lambda^* (S)=N\prod_{i\neq k}(b_i-a_i).

Taking NN\to\infty gives λ(I)==λ0(I)\lambda^* (I)=\infty=\lambda_0(I), and we are done.  \qed

Proposition: Every half-open interval IHnI\in\mc H_n is Lebesgue measurable.

Proof: We need to show that

λ(S)λ(SI)+λ(SIc)\lambda^* (S)\geq\lambda^* (S\cap I)+\lambda^* (S\cap I^c)

for any SRnS\subseteq\bb R^n. If λ(S)=\lambda^* (S)=\infty there is nothing to prove, so assume that λ(S)<\lambda^* (S)<\infty.

Fix ε>0\eps>0, and take a countable cover (Rk)(R_k) of SS by half-open intervals, such that

k=1λ0(Rk)λ(S)+ε.\sum_{k=1}^\infty\lambda_0(R_k)\leq\lambda^* (S)+\eps.

Note that each RkR_k can be written as the disjoint union of finitely many half-open intervals,

Rk=R~ki=1NSki,R_k=\tilde R_k\cup\bigcup_{i=1}^NS_k^i,

where R~k=RkI\tilde R_k=R_k\cap I, by “extending the sides” into a product partition. By the lemma, we have

λ0(Rk)=λ0(R~k)+i=1Nλ0(Ski),\lambda_0(R_k)=\lambda_0(\tilde R_k)+\sum_{i=1}^N\lambda_0(S_k^i),

so

λ(S)+εk=1λ0(Rk)=k=1λ0(R~k)+k=1i=1Nλ0(Ski)λ(SI)+λ(SIc), \begin{aligned} \lambda^* (S)+\eps&\geq\sum_{k=1}^\infty\lambda_0(R_k)\\ &=\sum_{k=1}^\infty\lambda_0(\tilde R_k)+\sum_{k=1}^\infty\sum_{i=1}^N\lambda_0(S_k^i)\\ &\geq\lambda^* (S\cap I)+\lambda^* (S\cap I^c), \end{aligned}

since (R~k)(\tilde R_k) cover SIS\cap I and (Ski)(S_k^i) cover SIcS\cap I^c. Taking ε0+\eps\to0^+ gives

λ(S)λ(SI)+λ(SIc),\lambda^* (S)\geq\lambda^* (S\cap I)+\lambda^* (S\cap I^c),

as desired.  \qed

Corollary: Every Borel subset of Rn\bb R^n is Lebesgue measurable.

Proof: Since the collection Σ\Sigma of all Lebesgue measurable sets is a σ\sigma-algebra, HnΣ\mc H_n\subseteq\Sigma implies Bn=σ(Hn)Σ\mc B_n=\sigma(\mc H_n)\subseteq\Sigma.  \qed

Comments (0)

Be the first to comment!

Write a comment…