Bourbaki

Proposition 3

Tag 08XQ, permanent. Cite it as [Bourbaki, Tag 08XQ].

Integration, chapter IV, § 6. Convexity inequalities

— Let μ\mu be a positive measure on X, F a real or complex Banach space, F' its strong dual, and (z,z′)↦⟨z,z′⟩(z, z') \mapsto \langle z, z' \rangle the canonical bilinear form on F×F′F \times F'.

1° For every function f∈LFpf \in \mathscr{L}_F^p (1≤p≤+∞1 \leq p \leq +\infty),

Np(f)=sup⁡∣∫⟨f,g⟩ dμ∣ N_p(f) = \sup \left| \int \langle f, g \rangle \, d\mu \right|

as g runs over the set of functions in LF′q\mathscr{L}_{F'}^q, such that Nq(g)≤1N_q(g) \leq 1.

2° For every function g∈LF′qg \in \mathscr{L}_{F'}^q (1≤q≤+∞1 \leq q \leq +\infty),

Nq(g)=sup⁡∣∫⟨f,g⟩ dμ∣ N_q(g) = \sup \left| \int \langle f, g \rangle \, d\mu \right|

as f runs over the set of functions in LFp\mathscr{L}_F^p such that Np(f)≤1N_p(f) \leq 1.

Let us first prove the relation (9); we distinguish between two cases.

(i) 1≤p<+∞1 \leq p < +\infty. The relation (9) being trivial when Np(f)=0N_p(f) = 0 (because f and ⟨f,g⟩\langle f, g \rangle are then negligible), we can always suppose, on multiplying f by a scalar, that Np(f)=1N_p(f) = 1. Suppose first that f is an integrable step function, f=∑k=1nakφAkf = \sum_{k=1}^n a_k \varphi_{A_k}, where the AkA_k are pairwise disjoint (§ 4, No. 9, Lemma). Thus ∑k=1n∣ak∣pμ(Ak)=1\sum_{k=1}^n |a_k|^p \mu(A_k) = 1 by hypothesis. For every ε>0\varepsilon > 0, there exists (for every index k) a vector ak′∈F′a'_k \in F' such that ∣ak′∣q=∣ak∣p|a'_k|^q = |a_k|^p if p>1p > 1 (resp. ∣ak′∣=1|a'_k| = 1 if p=1p = 1) and ⟨ak,ak′⟩≥(1−ε)∣ak∣⋅∣ak′∣\langle a_k, a'_k \rangle \geq (1 - \varepsilon)|a_k| \cdot |a'_k| (TVS, IV,

§1, No. 3, Prop. 8). Setting g=∑k=1nak′φAkg = \sum_{k=1}^n a_k' \varphi_{A_k}, we have ∑k=1n∣ak′∣qμ(Ak)=1\sum_{k=1}^n |a_k'|^q \mu(A_k) = 1 if p>1p > 1 (resp. sup⁡1≤k≤n∣ak′∣=1\sup_{1 \leq k \leq n} |a_k'| = 1 if p=1p = 1), thus Nq(g)=1N_q(g) = 1; on the other hand,

∫⟨f,g⟩dμ=∑k=1n⟨ak,ak′⟩μ(Ak)≥(1−ε)∑k=1n∣ak∣⋅∣ak′∣μ(Ak) \int \langle f, g \rangle d\mu = \sum_{k=1}^n \langle a_k, a_k' \rangle \mu(A_k) \geq (1 - \varepsilon) \sum_{k=1}^n |a_k| \cdot |a_k'| \mu(A_k)

and, since ∣ak′∣=∣ak∣p/q=∣ak∣p−1|a_k'| = |a_k|^{p/q} = |a_k|^{p-1} if p>1p > 1 (resp. ∣ak′∣=1|a_k'| = 1 if p=1p = 1), we have

∫⟨f,g⟩dμ≥(1−ε)∑k=1n∣ak∣pμ(Ak)=(1−ε)Np(f)=1−ε, \int \langle f, g \rangle d\mu \geq (1 - \varepsilon) \sum_{k=1}^n |a_k|^p \mu(A_k) = (1 - \varepsilon) N_p(f) = 1 - \varepsilon,

which proves the relation (9) in this case.

Let us pass to the case that ff is any element of LFp\mathscr{L}_F^p such that Np(f)=1N_p(f) = 1. For every ε>0\varepsilon > 0, there exists a step function f1∈LFpf_1 \in \mathscr{L}_F^p such that Np(f−f1)≤εN_p(f - f_1) \leq \varepsilon (§ 4, No. 10, Cor. 1 of Prop. 19). By what we have just seen, there exists a function g∈LFqg \in \mathscr{L}_F^q, such that Nq(g)=1N_q(g) = 1 and

∫⟨f1,g⟩dμ≥Np(f1)−ε≥1−2ε. \int \langle f_1, g \rangle d\mu \geq N_p(f_1) - \varepsilon \geq 1 - 2\varepsilon.

Now,

∫⟨f,g⟩dμ=∫⟨f1,g⟩dμ+∫⟨f−f1,g⟩dμ \int \langle f, g \rangle d\mu = \int \langle f_1, g \rangle d\mu + \int \langle f - f_1, g \rangle d\mu

and, by (6),

∣∫⟨f−f1,g⟩dμ∣≤Np(f−f1)Nq(g), \left| \int \langle f - f_1, g \rangle d\mu \right| \leq N_p(f - f_1) N_q(g),

whence

∣∫⟨f,g⟩dμ∣≥1−3ε, \left| \int \langle f, g \rangle d\mu \right| \geq 1 - 3\varepsilon,

which proves (9).

(ii) p=+∞p = +\infty. We may again restrict ourselves to the case that N∞(f)>0N_\infty(f) > 0. Let α\alpha be any number such that 0<α<N∞(f)0 < \alpha < N_\infty(f); by hypothesis, the set of x∈Xx \in X such that ∣f(x)∣>α|f(x)| > \alpha is measurable and is not locally negligible, therefore it contains a compact set KK of measure >0> 0. Since ff is measurable, there exists a compact set K1⊂KK_1 \subset K of measure >0> 0, such that the restriction of ff to K1K_1 is continuous. It follows that, for every ε>0\varepsilon > 0, there exists a partition of K1K_1 into a finite number of integrable sets, in each of which the oscillation of ff is ≤ε\leq \varepsilon; at least one of these sets AA has measure > 0. Let aa be one of the values of ff in AA; then ∣a∣>α|a| > \alpha and ∣f(x)−a∣≤ε|f(x) - a| \leq \varepsilon for all x∈Ax \in A. There exists a vector a′∈F′a' \in F' such that ∣a′∣=1|a'| = 1 and ∣⟨a,a′⟩∣≥∣a∣−ε|\langle a, a' \rangle| \geq |a| - \varepsilon; the function g=φA⋅a′/μ(A)g = \varphi_A \cdot a'/\mu(A) is integrable and N1(g)=1N_1(g) = 1; on the other hand,

∫⟨f,g⟩dμ=1μ(A)∫⟨f,a′⟩φAdμ. \int \langle f, g \rangle d\mu = \frac{1}{\mu(A)} \int \langle f, a' \rangle \varphi_A d\mu.

Now, one can write

∫⟨f,a′⟩φAdμ=⟨a,a′⟩μ(A)+∫⟨f−a,a′⟩φAdμ, \int \langle f, a' \rangle \varphi_A d\mu = \langle a, a' \rangle \mu(A) + \int \langle f - a, a' \rangle \varphi_A d\mu,

and since

∣⟨f−a,a′⟩φA∣≤εφA, |\langle f - a, a' \rangle \varphi_A| \leq \varepsilon \varphi_A,

we see that

∣∫⟨f,g⟩dμ∣≥∣⟨a,a′⟩∣−ε≥∣a∣−2ε>α−2ε; \left| \int \langle f, g \rangle d\mu \right| \geq |\langle a, a' \rangle| - \varepsilon \geq |a| - 2\varepsilon > \alpha - 2\varepsilon;

since ε\varepsilon is arbitrary and α\alpha is any number < N∞(f)N_\infty(f), the relation (9) is also verified in this case.

One argues in exactly the same manner to prove the relation (10), on considering separately the case 1≤q<+∞1 \leq q < +\infty and the case q=+∞q = +\infty, and using the fact that for every z′∈F′z' \in F', ∣z′∣=sup⁡∣z∣≤1∣⟨z,z′⟩∣|z'| = \sup_{|z| \leq 1} |\langle z, z' \rangle| by the definition of the norm in F′F'.

Bourbaki puts the proof directly after the statement, so what is above is the statement and its proof together.

Cited by

Cites

Transcribed from the printed volume. Not translated.