Michal–Bastiani calculus on locally convex spaces

15 Krein spaces

Indefinite inner product spaces after Bognár [ Bog74 ] : ortho-complemented subspaces of a Krein space (Chapter V), the majorant topology and the intrinsic dimension (Chapter IV), and continuous families of Krein spaces. This is the linear analysis behind the pseudo-Riemannian metrics of infinite index (Krein type) in lc-manifolds.

15.1 Complexification of real Hilbert spaces

For a real (or complex, regarded as real) inner product space \(E\), the complexification \(E_{\mathbb {C}} = E \times E\), with \((x, y)\) standing for \(x + iy\), is a complex inner product space with \(\langle (x, y), (u, v)\rangle = \langle x, u\rangle + \langle y, v\rangle + i(\langle x, v\rangle - \langle y, u\rangle )\) (real parts of the inner products of \(E\)), and \(\| (x, y)\| ^2 = \| x\| ^2 + \| y\| ^2\); it is complete when \(E\) is. A continuous linear operator \(T\) extends to \(T_{\mathbb {C}}(x, y) = (Tx, Ty)\); \(T \mapsto T_{\mathbb {C}}\) is multiplicative and preserves adjoints, and the conjugation \((x, y) \mapsto (x, -y)\) commutes with every \(T_{\mathbb {C}}\).

This is the standard device that brings the continuous functional calculus of the complex C\(^*\)-algebra of operators of \(E_{\mathbb {C}}\) to bear on real Hilbert spaces: it gives the spectral projections and the positive and negative parts used below over \(\mathbb {R}\). Mathlib has no complexification of inner product spaces in the version we use.

15.2 Ortho-complemented subspaces

Following [ Bog74 , § I.11, § III.8, § V.1 ] , a subspace \(L\) of a space with a Hermitian form \([\cdot , \cdot ]\) is a Krein space if it admits a fundamental decomposition \(L = L_+ \oplus L_-\), a \([\cdot , \cdot ]\)-orthogonal direct sum with \([\cdot , \cdot ]\) positive definite on \(L_+\) and negative definite on \(L_-\), whose components are intrinsically complete: complete for the norms \(\sqrt{[x, x]}\) on \(L_+\) and \(\sqrt{-[x, x]}\) on \(L_-\). No topology enters. The ambient space is a Hilbert space \(H\) over \(\mathbb {R}\) or \(\mathbb {C}\) with a fundamental symmetry \(J\), a self-adjoint involution, and \([x, y] = \langle Jx, y\rangle \); every Krein space is of this form [ Bog74 , Corollary V.1.2 ] . A subspace \(L \subseteq H\) is ortho-complemented if \(L + L^{[\perp ]} = H\) [ Bog74 , § I.9, p. 18 ] (projectively complete in [ AI89 , Definition 1.7.8 ] ).

Let \(H\) be a Hilbert space with a fundamental symmetry \(J\), and \(L \subseteq H\) a subspace.

  1. If \(L\) is ortho-complemented, it is closed and nondegenerate [ Bog74 , Corollaries III.2.5 and I.9.5 ] .

  2. \(L\) is ortho-complemented if and only if there is a continuous \(J\)-orthogonal projector \(Q\) onto \(L\), i.e. \([Qw, t] = [w, t]\) for \(t \in L\); such a projector is unique. For closed \(L\) this happens if and only if the Gram operator \(G_L = P_L J|_L\) is bijective.

  3. A closed subspace which is a Krein space with the restricted form is ortho-complemented.

  4. A subspace is ortho-complemented if and only if it is closed and a Krein space (over \(\mathbb {R}\) and \(\mathbb {C}\)).

Proof ▶

Items 1 and 2 use no completeness beyond the orthogonal projection onto \(L\). For 3, Bognár goes through Hilbert majorants and majorant topologies [ Bog74 , Theorem V.1.3 ] ; here the closed graph theorem shows that the intrinsic norms dominate the norm of \(H\), and the Riesz representation theorem in \(L_+\) and \(L_-\) produces the \(J\)-orthogonal projection, the route c \(\Rightarrow \) d \(\Rightarrow \) a of [ AI89 , Theorem 1.7.16 ] . For 4, \(L_\pm \) are the spectral subspaces of \(G_L\) for \((0, \infty )\) and \((-\infty , 0)\), obtained from the continuous functional calculus in the C\(^*\)-algebra \(\mathcal{L}(L)\); they are closed with \(\pm [x, x] \ge c\| x\| ^2\), hence intrinsically complete. Over \(\mathbb {R}\) the continuous functional calculus is applied to the complexification \(G_{\mathbb {C}}\) of \(G_L\) on \(L \times L\): its spectral projection \(P\) commutes with \(G_{\mathbb {C}}\), hence with every operator commuting with \(G_{\mathbb {C}}\); the conjugate \(CPC\) (for the conjugation \(C(x, y) = (x, -y)\)) is an idempotent commuting with \(P\) with the same sign bounds, so \(CPC = P\) and \(P\) restricts to a real projection on \(L\). We do not claim this argument is new.

A positive definite subspace \(L\) of the Krein space \(H\) is ortho-complemented if and only if it is closed and intrinsically complete, if and only if it is closed and uniformly positive: \([x, x] \ge c\| x\| ^2\) on \(L\) for some \(c {\gt} 0\). A closed positive definite subspace which is not uniformly positive is nondegenerate and decomposable (\(L = L \oplus 0\)) but neither intrinsically complete, nor a Krein space, nor ortho-complemented.

Proof ▶

A Krein decomposition of a positive definite \(L\) has \(L_- = 0\), so Theorem 15.3 gives the first equivalence. If \(L\) is closed and intrinsically complete, the identity of \(L\) onto its intrinsic Hilbert space is continuous (\([x, x] \le \| J\| \| x\| ^2\)) and bijective between complete spaces, so its inverse is bounded by the open mapping theorem; the converse is a Cauchy sequence argument. Uses the completeness of \(H\).

Example 15.5 [ Bog74 , Example V.3.3 ]
✓
#

On \(\ell ^2\) (over \(\mathbb {R}\) or \(\mathbb {C}\)) let \(J x = (s_i x_i)\) with \(s_i = -1\) for even and \(s_i = 1\) for odd \(i\), and \(L = \{ x \mid x_{2k+1} = \tfrac {2k+2}{2k+1}\, x_{2k}\} \), the subspace of [ Bog74 , Example I.9.7, pp. 19–20 ] . Then \(L\) is closed, positive definite, nondegenerate and decomposable, but not intrinsically complete, not a Krein space, and not ortho-complemented.

Proof ▶

For \(x \in L\), \([x, x] = \sum _k (d_k^2 - 1)|x_{2k}|^2\) with \(d_k = \tfrac {2k+2}{2k+1}\), which is positive for \(x \ne 0\), while \(d_k^2 - 1 \le 3/(2k+1) \to 0\); apply Proposition 15.4. Bognár exhibits instead an intrinsic Cauchy sequence whose would-be limit is not in \(\ell ^2\).

Let \(G\) be a bounded self-adjoint operator on a real or complex Hilbert space \(E\). Then \(G = G^+ - G^-\) with \(G^\pm \) positive and \(G^+ G^- = G^- G^+ = 0\), and \(E = N_0 \oplus N_+ \oplus N_-\) with \(N_0 = \ker G\), \(N_+ = (\ker G^+)^\perp \), \(N_- = (\ker G^-)^\perp \); the three components and the sums of any two of them are closed, \(\langle Gx, x\rangle \) is positive definite on \(N_+\) and negative definite on \(N_-\), and \(\langle G N_+, N_-\rangle = 0\).

Proof ▶

Over \(\mathbb {C}\), \(G^\pm \) are the positive and negative parts from the continuous functional calculus. Over \(\mathbb {R}\), take those of the complexification \(G_{\mathbb {C}}\); \(C G_{\mathbb {C}}^\pm C\) is again a decomposition of \(G_{\mathbb {C}}\) into positive operators with product \(0\), so it equals \(G_{\mathbb {C}}^\pm \) by uniqueness, and \(G_{\mathbb {C}}^\pm \) restrict to \(E\). Since \(G^+ G^- = 0\), \(G^- y \in \ker G^+\), so \(\langle y, G^- y\rangle = 0\) for \(y \in N_+\), whence \(G^- y = 0\) (a positive operator vanishes where its quadratic form does): \(N_+ \subseteq \ker G^-\), and likewise \(N_- \subseteq \ker G^+\). On \(N_+\), \(\langle Gx, x\rangle = \langle G^+x, x\rangle {\gt} 0\) for \(x \ne 0\), since \(G^+ x = 0\) would put \(x\) in \(\ker G^+ \cap N_+ = 0\). Decomposing along \(\ker G^+\) and then \(\ker G^-\) gives \(E = N_0 + N_+ + N_-\), with \(N_0 + N_+ = \ker G^-\), \(N_0 + N_- = \ker G^+\), \(N_+ + N_- = N_0^\perp \). These are the spectral subspaces of \(G\) for \(\{ 0\} \), \((0, \infty )\), \((-\infty , 0)\), but no spectral measure is used.

Every closed subspace \(L\) of the Krein space \(H\) has a fundamental decomposition \(L = L^0 \dotplus L_+ \dotplus L_-\) [ Bog74 , § I.11, p. 24 ] (\(L^0\) the isotropic part, \(L_+\) positive definite, \(L_-\) negative definite, mutually orthogonal) whose three components and the sums of any two of them are closed. In particular a closed nondegenerate subspace has a fundamental decomposition \(L = L_+ \oplus L_-\) with closed components.

Proof ▶

Apply Lemma 15.6 to the Gram operator \(G_L\) on the Hilbert space \(L\); \(\ker G_L\) is the isotropic part of \(L\). Bognár deduces the theorem from his Theorem IV.5.2 (p. 89), whose proof uses the spectral function of the Gram operator; since \(0\) may lie in the spectrum of \(G_L\), the corresponding spectral projections are not continuous functions of \(G_L\), and Lemma 15.6 replaces them.

Proof ▶

“Only if”: \(L\) is closed and nondegenerate by Theorem 15.3, and \(L^{[\perp ]}\) is ortho-complemented, hence a Krein space with a decomposition \(M_+ \oplus M_-\); \(H_\pm = L_\pm + M_\pm \), as in Bognár. “If”: \(L\) has a fundamental decomposition by Theorem 15.7. Every fundamental decomposition of \(H\) has uniformly definite components (they are ortho-complemented and definite, Proposition 15.4); \(L_+ = L \cap L_-^{[\perp ]}\) is closed and, inside \(H_+\), uniformly positive, hence ortho-complemented; likewise \(L_-\); and an orthogonal sum of ortho-complemented subspaces is ortho-complemented [ Bog74 , Lemma I.9.2 ] . Bognár instead ortho-complements \(L_+\) inside the Hilbert space \((H_+, [\cdot , \cdot ])\). We do not claim this argument is new.

15.3 The majorant topology

For a fundamental decomposition \(E = E_+ \oplus E_-\) of a Hermitian form \(B\), with projections \(P_\pm \) and fundamental symmetry \(J = P_+ - P_-\), the \(J\)-inner product \([x, y]_J = B(Jx, y) = B(P_+ x, P_+ y) - B(P_- x, P_- y)\) is positive definite. Its norm defines the decomposition majorant [ Bog74 , § IV.5–6 ] .

Let \((E, B)\) be a Krein space and \(K\) a Krein decomposition. Then \(E\) with \([\cdot , \cdot ]_K\) is a Hilbert space, every fundamental decomposition \(D\) of \((E, B)\) has a \(J\)-norm equivalent to that of \(K\), and \(D\) is a Krein decomposition. On a Hilbert space with \(B = \langle G \cdot , \cdot \rangle \), \(G\) invertible and self-adjoint, every fundamental decomposition has closed components, a continuous fundamental symmetry, and a \(J\)-norm equivalent to the norm.

Proof ▶

\(E_\pm = E_\mp ^{[\perp ]}\) is closed, so \(J\) is continuous by the closed graph theorem. The operator \(U = GJ\) is self-adjoint, positive (\(\langle Ux, x\rangle = [x, x]_J\)) and invertible, so \(\| Ux\| ^2 \le \| U\| \langle Ux, x\rangle \) (Cauchy–Schwarz for the positive form) and \(\| x\| \le \| U^{-1}\| \| Ux\| \) give \(c\| x\| ^2 \le [x, x]_J \le \| U\| \| x\| ^2\). The algebraic case is reduced to this one on the Hilbert space of \(K\), whose Gram operator is \(J_K\). Bognár proves Theorem 6.4 for an arbitrary Hilbert majorant by an iteration; for Krein spaces the positivity and invertibility of \(GJ\) suffice. We do not claim this argument is new.

15.4 The intrinsic dimension

The dense-span dimension of a topological vector space is the least cardinality of a set with dense span; the intrinsic dimension of a definite subspace is its dense-span dimension for the intrinsic norm. For two fundamental decompositions of a Krein space, \(P_{1+} \colon E_{2+} \to E_{1+}\) is an isomorphism of topological vector spaces, so the intrinsic dimensions of \(E_+\) and of \(E_-\) (the index) do not depend on the decomposition.

Proof ▶

For \(x \in E_{2+}\), \(B(x, x) = B(P_{1+}x, P_{1+}x) + B(P_{1-}x, P_{1-}x) \le B(P_{1+}x, P_{1+}x)\), so \(P_{1+}\) is bounded below; it is bounded by Theorem 15.10. Its range is closed, and its orthogonal in \(E_{1+}\) lies in \(E_{2+}^{[\perp ]} = E_{2-}\), hence is zero. Bognár deduces Corollary 7.4 from his Lemma 7.3 instead.

In an inner product space, no orthonormal family is larger than a set with dense span. Hence the dense-span dimension of an inner product space is the cardinality of any orthonormal family with dense span; in particular all Hilbert bases of a Hilbert space have the same cardinality, and the ranks of positivity and of negativity of a Krein space are the Hilbert dimensions of \(E_+\) and \(E_-\).

Proof ▶

Let \((v_i)\) be orthonormal and \(s\) have dense span. Each \(v_i\) has a nonzero inner product with some element of \(s\), otherwise \(v_i\) would be orthogonal to the closure of the span of \(s\). If \(s\) is infinite, each element of \(s\) has nonzero inner products with countably many \(v_i\) only (Bessel’s inequality), so the index set has cardinality at most \(|s| \cdot \aleph _0 = |s|\). If \(s\) is finite, its span is closed, hence everything, and the \(v_i\) are linearly independent.

15.5 Continuous families of Krein spaces

Let \(G(x)\) be a norm-continuous family of invertible self-adjoint operators on a real or complex Hilbert space. The spectral projections of \(G(x)\) onto the positive part depend continuously on \(x\), so fundamental decompositions of \((H, \langle G(x)\cdot , \cdot \rangle )\) can be chosen continuously; and the index of any choice of fundamental decompositions is a locally constant function of \(x\).

Proof ▶

Over \(\mathbb {C}\) the projection is \(\chi _{(0,\infty )}(G(x))\), and the continuous functional calculus is continuous in the operator when the spectra stay in a fixed open set on which the function is continuous (upper hemicontinuity of the spectrum); over \(\mathbb {R}\) use the complexification. If \(P\), \(Q\) are orthogonal projections with \(\| P - Q\| {\lt} 1\), then \(v \mapsto v - Qv\) is an isomorphism \(\ker P \cong \ker Q\). The index is computed on the spectral decomposition, where the intrinsic norm on \(\ker P\) is equivalent to the norm.