Bourbaki

Proposition 10

Tag 09SS, permanent. Cite it as [Bourbaki, Tag 09SS].

Integration, chapter IX, § 6. Promeasures and measures on a locally convex space

— Under the hypothesis Φ(0)−ℜΦ≤ε+q\Phi(0) - \Re \Phi \leq \varepsilon + q, we have

μ(T′−Cr)≤3(ε+r−2Tr⁡(q/h)) \mu(T' - C_r) \leq 3 (\varepsilon + r^{-2} \operatorname{Tr}(q/h))

for every r>0r > 0.

One writes Tr⁡(q/h)\operatorname{Tr}(q/h) for the trace of q with respect to h (cf. Annex, No. 1). The formula (55) is trivial when Tr⁡(q/h)\operatorname{Tr}(q/h) is infinite. We assume henceforth that Tr⁡(q/h)\operatorname{Tr}(q/h) is finite, hence that h(x)=0h(x) = 0 implies q(x)=0q(x) = 0 for x∈Tx \in T.

Let a1,…,ana_1, \ldots, a_n be elements of TT, and DD the set of x′∈T′x' \in T' such that ∑j=1n⟨aj,x′⟩2>1\sum_{j=1}^n \langle a_j, x' \rangle^2 > 1. For every real t≥0t \geq 0 we have 3(1−e−t/2)≥03(1 - e^{-t/2}) \geq 0, and we even have

3(1−e−t/2)≥3(1−e−1/2)≥3(1−(94)−1/2)=1 3(1 - e^{-t/2}) \geq 3(1 - e^{-1/2}) \geq 3\left(1 - \left(\frac{9}{4}\right)^{-1/2}\right) = 1

for t>1t > 1, because e>94e > \frac{9}{4}. Applying these inequalities to t=∑j=1n⟨aj,x′⟩2t = \sum_{j=1}^n \langle a_j, x' \rangle^2, we obtain

μ(D)≤3∫T′(1−exp⁡(−12∑j=1n⟨aj,x′⟩2))dμ(x′). \mu(D) \leq 3 \int_{T'} \left(1 - \exp \left(-\frac{1}{2} \sum_{j=1}^n \langle a_j, x' \rangle^2\right)\right) d\mu(x').

Let γ\gamma be the measure on R\mathbf{R} having density t↦(2π)−1/2e−t2/2t \mapsto (2\pi)^{-1/2} e^{-t^2/2} with respect to Lebesgue measure. By Lemma 2 of No. 4,

∫Reiutdγ(t)=e−u2/2 \int_{\mathbf{R}} e^{iut} d\gamma(t) = e^{-u^2/2}

for all real uu. Consequently,

1−exp⁡(−12∑j=1n⟨aj,x′⟩2)=∫⋯∫(1−ei∑j=1n⟨aj,x′⟩tj)dγ(t1)⋯dγ(tn) \begin{align*} 1 - \exp \left(-\frac{1}{2} \sum_{j=1}^n \langle a_j, x' \rangle^2\right) \\ &= \int \cdots \int \left(1 - e^{i \sum_{j=1}^n \langle a_j, x' \rangle t_j}\right) d\gamma(t_1) \cdots d\gamma(t_n) \end{align*}

for all x′∈T′x' \in T'. The function of x′,t1,…,tnx', t_1, \ldots, t_n to be integrated in the second member is continuous and is bounded above in absolute value by 2, and the measures μ\mu and γ\gamma are bounded; one can therefore integrate the two members of (57) with respect to dμ(x′)d\mu(x') and interchange the integrations with respect to μ\mu and γ\gamma; one obtains

∫T′(1−exp⁡(−12∑j=1n⟨aj,x′⟩2))dμ(x′)=∫⋯∫(Φ(0)−Φ(∑j=1ntjaj))dγ(t1)⋯dγ(tn). \begin{align*} \int_{T'} \left(1 - \exp \left(-\frac{1}{2} \sum_{j=1}^n \langle a_j, x' \rangle^2\right)\right) d\mu(x') \\ &= \int \cdots \int \left(\Phi(0) - \Phi\left(\sum_{j=1}^n t_j a_j\right)\right) d\gamma(t_1) \cdots d\gamma(t_n). \end{align*}

Since qq is a quadratic form on TT, there exist real numbers qjkq_{jk} such that

q(∑j=1ntjaj)=∑j,kqjktjtk q \left( \sum_{j=1}^n t_j a_j \right) = \sum_{j,k} q_{jk} t_j t_k
for t1,…,tnt_1, \ldots, t_n real; in particular, qjj=q(aj)q_{jj} = q(a_j) for 1≤j≤n1 \leq j \leq n. Moreover, the integral ∫Rtndγ(t)\int_{\mathbf{R}} t^n d\gamma(t) has the values 1, 0, 1 for n=0,1,2n = 0, 1, 2, respectively (No. 4, Lemma 1). From this, one deduces immediately
∫⋯∫(ε+q(∑j=1ntjaj))dγ(t1)…dγ(tn)=ε+∑j=1nq(aj). \int \cdots \int \left( \varepsilon + q \left( \sum_{j=1}^n t_j a_j \right) \right) d\gamma(t_1) \ldots d\gamma(t_n) = \varepsilon + \sum_{j=1}^n q(a_j).
Now, the first member of (58) and Φ(0)\Phi(0) are real numbers; one can therefore replace Φ\Phi by ℜΦ\Re \Phi in the second member of (58). The inequality Φ(0)−ℜΦ≤ε+q\Phi(0) - \Re \Phi \leq \varepsilon + q and the formulas (56), (58) and (59) then imply
μ(D)≤3(ε+∑j=1nq(aj)). \mu(D) \leq 3 \left( \varepsilon + \sum_{j=1}^n q(a_j) \right).
Let us fix the number r>0r > 0. Since the quadratic form hh is positive, there exist a basis (e1,…,en)(e_1, \ldots, e_n) of TT and an integer mm between 0 and nn such that
h(∑j=1ntjej)=∑j=1mtj2 h \left( \sum_{j=1}^n t_j e_j \right) = \sum_{j=1}^m t_j^2
for t1,…,tnt_1, \ldots, t_n real (Annex, No. 1, Prop. 2). It is then immediate that CrC_r consists of the x′∈T′x' \in T' such that
∑j=1m⟨ej,x′⟩2≤r2,∑j=m+1n⟨ej,x′⟩2=0. \sum_{j=1}^m \langle e_j, x' \rangle^2 \leq r^2, \qquad \sum_{j=m+1}^n \langle e_j, x' \rangle^2 = 0.
For every integer l≥1l \geq 1, let DlD_l be the set of x′∈T′x' \in T' satisfying the inequality
∑j=1m⟨r−1ej,x′⟩2+∑j=m+1n⟨lej,x′⟩2>1. \sum_{j=1}^m \langle r^{-1} e_j, x' \rangle^2 + \sum_{j=m+1}^n \langle l e_j, x' \rangle^2 > 1.
One sees easily that the sequence (Dl)l≥1(D_l)_{l \geq 1} is increasing with union T′−CrT' - C_r, whence
μ(T′−Cr)=lim⁡l→∞μ(Dl). \mu(T' - C_r) = \lim_{l \to \infty} \mu(D_l).

But by (60),

μ(Dl)≤3(ε+∑j=1mr−2q(ej)+∑j=m+1nl2q(ej)); \mu(D_l) \leq 3 \left( \varepsilon + \sum_{j=1}^m r^{-2} q(e_j) + \sum_{j=m+1}^n l^2 q(e_j) \right);

for j=m+1,…,nj = m+1, \ldots, n we have h(ej)=0h(e_j) = 0, therefore q(ej)=0q(e_j) = 0. Moreover, Tr⁡(q/h)=∑j=1mq(ej)\operatorname{Tr}(q/h) = \sum_{j=1}^m q(e_j) (Annex, No. 1, Prop. 2). The relation (55) then follows from (61) and (62).

Q.E.D.

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.