Bourbaki

Annex

Complements on Hilbert spaces

1. Trace of a quadratic form with respect to another(1)^{(1)}

In this No., we denote by EE a real vector space and by Q,HQ, H two positive quadratic forms on EE. There exist two symmetric bilinear forms (x,y)↦(x∣y)Q(x, y) \mapsto (x|y)_Q and (x,y)↦(x∣y)H(x, y) \mapsto (x|y)_H characterized by

Q(x)=(x∣x)Q,H(x)=(x∣x)H Q(x) = (x|x)_Q , \quad H(x) = (x|x)_H
for all x∈Ex \in E.

One calls trace of Q with respect to H and one denotes by Tr⁡(Q/H)\operatorname{Tr}(Q/H) the positive real number, finite or not, defined as follows:

a) If there exists an x∈Ex \in E such that H(x)=0H(x) = 0 and Q(x)≠0Q(x) \neq 0, one sets Tr⁡(Q/H)=+∞\operatorname{Tr}(Q/H) = +\infty.

b) In the contrary case, Tr⁡(Q/H)\operatorname{Tr}(Q/H) is the supremum of the set of numbers of the form ∑i=1pQ(ei)\sum_{i=1}^p Q(e_i), where (e1,…,ep)(e_1, \ldots, e_p) runs over the set of finite sequences of elements of EE orthonormal for HH.

Let EE be a real Hilbert space and QQ a positive quadratic form on EE. Set H(x)=∥x∥2H(x) = \|x\|^2 for all x∈Ex \in E; then HH is a positive quadratic form on EE. One says that QQ is nuclear if Tr⁡(Q/H)\operatorname{Tr}(Q/H) is finite. For every x∈Ex \in E of norm 1, one has Q(x)≤Tr⁡(Q/H)Q(x) \leq \operatorname{Tr}(Q/H), whence Q≤Tr⁡(Q/H)⋅HQ \leq \operatorname{Tr}(Q/H) \cdot H; in particular, every nuclear form QQ is continuous.

Remark

— 1) For every subspace FF of EE, denote by QFQ_F the restriction of QQ to FF, and by HFH_F that of HH. Then Tr⁡(QF/HF)≤Tr⁡(Q/H)\operatorname{Tr}(Q_F/H_F) \leq \operatorname{Tr}(Q/H) and Tr⁡(Q/H)\operatorname{Tr}(Q/H) is the supremum of the numbers Tr⁡(QF/HF)\operatorname{Tr}(Q_F/H_F) for F⊂EF \subset E finite-dimensional.

2) Let E1E_1 be a real vector space, Q1Q_1 and H1H_1 two positive quadratic forms on E1E_1, and π:E→E1\pi : E \to E_1 a surjective linear mapping. If Q=Q1∘πQ = Q_1 \circ \pi and H=H1∘πH = H_1 \circ \pi, then Tr⁡(Q/H)=Tr⁡(Q1/H1)\operatorname{Tr}(Q/H) = \operatorname{Tr}(Q_1/H_1).

(1) Cf. TVS, V, §4, No. 9.

Proposition 1

— *Assume that E is finite-dimensional and H is nondegenerate.

a) There exists an endomorphism u of E characterized by (x∣y)Q=(u(x)∣y)H(x|y)_Q = (u(x)|y)_H for x, y in E.

b) Tr⁡(Q/H)=Tr⁡(u)\operatorname{Tr}(Q/H) = \operatorname{Tr}(u).

c) Tr⁡(Q/H)=∑i=1mQ(ei)\operatorname{Tr}(Q/H) = \sum_{i=1}^m Q(e_i) for every basis (e1,…,em)(e_1, \ldots, e_m) of E orthonormal for H.

a) follows from the fact that the bilinear form (x,y)↦(x∣y)H(x, y) \mapsto (x|y)_H is non-degenerate. Every sequence in E orthonormal for H may be completed to a basis of E orthonormal for H. Consequently, Tr⁡(Q/H)\operatorname{Tr}(Q/H) is the supremum of the set of numbers of the form ∑i=1mQ(ei)\sum_{i=1}^m Q(e_i) over all bases (e1,…,em)(e_1, \ldots, e_m) of E orthonormal for H. To prove b) and c), it suffices to show that ∑i=1mQ(ei)=Tr⁡(u)\sum_{i=1}^m Q(e_i) = \operatorname{Tr}(u) for every basis of this kind. Set aij=(u(ei)∣ej)H=(ei∣ej)Qa_{ij} = (u(e_i)|e_j)_H = (e_i|e_j)_Q for 1≤i,j≤m1 \leq i, j \leq m; then u(ei)=∑j=1maijeju(e_i) = \sum_{j=1}^m a_{ij} e_j for 1≤i≤m1 \leq i \leq m, whence

Tr⁡(u)=∑i=1maii=∑i=1m(ei∣ei)Q=∑i=1mQ(ei). \operatorname{Tr}(u) = \sum_{i=1}^m a_{ii} = \sum_{i=1}^m (e_i|e_i)_Q = \sum_{i=1}^m Q(e_i).

Q.E.D.

Proposition 2

— Assume that E is finite-dimensional. There exist a basis (e1,…,en)(e_1, \ldots, e_n) of E and an integer m with 0≤m≤n0 \leq m \leq n *such that

H⁡(∑i=1ntiei)=∑i=1mti2 \operatorname{H}\left( \sum_{i=1}^n t_i e_i \right) = \sum_{i=1}^m t_i^2

for t1,…,tnt_1, \ldots, t_n real. If, moreover, the relation H⁡(x)=0\operatorname{H}(x) = 0 implies Q(x)=0Q(x) = 0 for x∈Ex \in E, then Tr⁡(Q/H)=∑i=1mQ(ei)\operatorname{Tr}(Q/H) = \sum_{i=1}^m Q(e_i).

There exists a basis (e1′,…,en′)(e'_1, \ldots, e'_n) of E orthogonal for H. One can assume the basis to be indexed in such a way that H⁡(ei′)>0\operatorname{H}(e'_i) > 0 for 1≤i≤m1 \leq i \leq m and H⁡(ei′)=0\operatorname{H}(e'_i) = 0 for m<i≤nm < i \leq n. Then set ei=ei′/H⁡(ei′)1/2e_i = e'_i / \operatorname{H}(e'_i)^{1/2} for 1≤i≤m1 \leq i \leq m and ei=ei′e_i = e'_i for m<i≤nm < i \leq n; the relation (1) is satisfied.

Let F be the subspace of E generated by em+1′,…,en′e'_{m+1}, \ldots, e'_n; it is the set of x∈Ex \in E such that H⁡(x)=0\operatorname{H}(x) = 0. Denote by π\pi the canonical mapping of E onto E1=E/FE_1 = E/F. Since Q and H are zero on F, there exist two positive quadratic forms Q1Q_1 and H1H_1 on E1E_1 such that Q=Q1∘πQ = Q_1 \circ \pi and H=H1∘πH = H_1 \circ \pi. Moreover, (π(e1),…,π(em))(\pi(e_1), \ldots, \pi(e_m)) is a basis of E1E_1 orthonormal for H1H_1, and therefore H1H_1 is nondegenerate.

By Prop. 1 and Remark 2,

Tr⁡(Q/H)=Tr⁡(Q1/H1)=∑i=1mQ1(π(ei))=∑i=1mQ(ei). \operatorname{Tr}(Q/H) = \operatorname{Tr}(Q_1/H_1) = \sum_{i=1}^m Q_1(\pi(e_i)) = \sum_{i=1}^m Q(e_i).

Q.E.D.

Remark 3

— Assume E finite-dimensional and H nondegenerate. Let (e1,…,en)(e_1, \ldots, e_n) be a basis of E. Set q=((ei∣ej)Q)1≤i,j≤nq = ((e_i|e_j)_Q)_{1 \leq i,j \leq n} and h=((ei∣ej)H)1≤i,j≤nh = ((e_i|e_j)_H)_{1 \leq i,j \leq n}. With notations as in Prop. 1, the matrix of uu for the basis (e1,…,en)(e_1, \ldots, e_n) of E is equal to h−1qh^{-1}q, whence

(2)

Tr⁡(Q/H)=Tr⁡(h−1q)=Tr⁡(qh−1). \operatorname{Tr}(Q/H) = \operatorname{Tr}(h^{-1}q) = \operatorname{Tr}(qh^{-1}).

2. Hilbert–Schmidt mappings(2)^{(2)}

Let E be a real Hilbert space, in which the scalar product is denoted by (x∣y)(x|y). There exists an isometry jEj_E of E onto its dual, characterized by the formula

(3)

(x∣y)=⟨x,jE(y)⟩for x,y in E (x|y) = \langle x, j_E(y) \rangle \quad \text{for } x, y \text{ in E}
(TVS, V, §1, No. 7, Th. 3).

Let E1E_1 and E2E_2 be two real Hilbert spaces and uu a continuous linear mapping of E1E_1 into E2E_2. One calls adjoint of u the continuous linear mapping u∗=jE1−1∘tu∘jE2u^* = j_{E_1}^{-1} \circ {}^t u \circ j_{E_2} of E2E_2 into E1E_1. The mapping u∗u^* is characterized by the relation

(4)

(u(x1)∣x2)=(x1∣u∗(x2))for x1∈E1, x2∈E2. (u(x_1)|x_2) = (x_1|u^*(x_2)) \quad \text{for } x_1 \in E_1,\ x_2 \in E_2.

If vv is a continuous linear mapping of E2E_2 into a Hilbert space E3E_3, one has (v∘u)∗=u∗∘v∗(v \circ u)^* = u^* \circ v^*.

Let E1E_1 and E2E_2 be two real Hilbert spaces and uu a linear mapping of E1E_1 into E2E_2. One defines two positive quadratic forms H and Q on E1E_1 by the formulas

H(x)=∥x∥2,Q(x)=∥u(x)∥2(x∈E1). H(x) = \|x\|^2,\quad Q(x) = \|u(x)\|^2 \quad (x \in E_1).

Proposition 3

Suppose u is continuous. Let (ei)i∈I(e_i)_{i \in I} be an orthonormal basis of E1E_1 and (fj)j∈J(f_j)_{j \in J} an orthonormal basis of E2E_2. Then

Tr⁡(Q∣H)=∑i∈I∥u(ei)∥2=∑j∈J∥u∗(fj)∥2=∑i∈I∑j∈J(u(ei)∣fj)2. \operatorname{Tr}(Q|H) = \sum_{i \in I} \|u(e_i)\|^2 = \sum_{j \in J} \|u^*(f_j)\|^2 = \sum_{i \in I} \sum_{j \in J} (u(e_i)|f_j)^2.

(2) Cf. TVS, V, §4, No. 7.

For every x∈E1x \in E_1, we have ∥x∥2=∑i∈I(x∣ei)2\|x\|^2 = \sum_{i \in I} (x|e_i)^2, and similarly ∥y∥2=∑j∈J(y∣fj)2\|y\|^2 = \sum_{j \in J} (y|f_j)^2 for every y∈E2y \in E_2. Consequently,

∑i∈I∥u(ei)∥2=∑i∈I∑j∈J(u(ei)∣fj)2=∑j∈J∑i∈I(ei∣u∗(fj))2=∑j∈J∥u∗(fj)∥2. \sum_{i \in I} \|u(e_i)\|^2 = \sum_{i \in I} \sum_{j \in J} (u(e_i)|f_j)^2 \\ = \sum_{j \in J} \sum_{i \in I} (e_i|u^*(f_j))^2 \\ = \sum_{j \in J} \|u^*(f_j)\|^2 .

In particular, the number ∑i∈I∥u(ei)∥2\sum_{i \in I} \|u(e_i)\|^2 is independent of the orthonormal basis (ei)i∈I(e_i)_{i \in I} of E1E_1.

Set t=Tr⁡(Q∣H)t = \operatorname{Tr}(Q|H). For every finite subset I′I' of II, by definition

∑i∈I′∥u(ei)∥2=∑i∈I′Q(ei)≤t, \sum_{i \in I'} \|u(e_i)\|^2 = \sum_{i \in I'} Q(e_i) \leq t ,

whence ∑i∈I∥u(ei)∥2≤t\sum_{i \in I} \|u(e_i)\|^2 \leq t. Let (e1′,…,ep′)(e'_1, \ldots, e'_p) be a finite orthonormal sequence in EE. This sequence can be completed to an orthonormal basis (eα′)α∈A(e'_\alpha)_{\alpha \in A} of E1E_1. Then

∑α=1p∥u(eα′)∥2≤∑α∈A∥u(eα′)∥2=∑i∈I∥u(ei)∥2 \sum_{\alpha=1}^p \|u(e'_\alpha)\|^2 \leq \sum_{\alpha \in A} \|u(e'_\alpha)\|^2 = \sum_{i \in I} \|u(e_i)\|^2

and, passing to the supremum over all (e1′,…,ep′)(e'_1, \ldots, e'_p), one obtains the inequality t≤∑i∈I∥u(ei)∥2t \leq \sum_{i \in I} \|u(e_i)\|^2. We have thus established the equality t=∑i∈I∥u(ei)∥2t = \sum_{i \in I} \|u(e_i)\|^2.

Q.E.D.

One says that uu is a Hilbert–Schmidt mapping of E1E_1 into E2E_2 if the positive quadratic form Q:x↦∥u(x)∥2Q : x \mapsto \|u(x)\|^2 on E1E_1 is nuclear. When this is so, one has Q≤Tr⁡(Q/H)⋅HQ \leq \operatorname{Tr}(Q/H) \cdot H, therefore uu is continuous and

∥u∥≤Tr⁡(Q/H)1/2. \|u\| \leq \operatorname{Tr}(Q/H)^{1/2} .

Let u:E1→E2u : E_1 \to E_2 be a continuous linear mapping. By Prop. 3, uu is a Hilbert–Schmidt mapping if and only if there exists an orthonormal basis (ei)i∈I(e_i)_{i \in I} of E1E_1 such that ∑i∈I∥u(ei)∥2<+∞\sum_{i \in I} \|u(e_i)\|^2 < +\infty. When this is so, every orthonormal basis of E1E_1 has the same property. Moreover, if uu is a Hilbert–Schmidt mapping, then so is its adjoint u∗u^* by virtue of the formula ∑i∈I∥u(ei)∥2=∑j∈J∥u∗(fj)∥2\sum_{i \in I} \|u(e_i)\|^2 = \sum_{j \in J} \|u^*(f_j)\|^2 (Prop. 3).

Exercises

Exercises

See the exercises for Appendix 0.

Transcribed from the printed volume. Not translated.