Large Perturbations of Nest Algebras

If P \wedge Q = R \ne 0, we can replace P and Q by P-R and Q-R. This does not affect the problem unless P=Q, which is a trivial case. Let P' = (P \vee Q) - P. Then P'QP is injective on PH and has range dense in P'H. There is no loss in assuming that P+P'=I.

Let C = (PQP)^{1/2} and S = (1-C^2)^{1/2}. The Halmos model [Halmos] for two subspaces in generic position (i.e., P\wedge Q = 0 and P \vee Q = I) shows that P, Q and U are unitarily equivalent to

Then since the two columns of P-Q have orthogonal ranges,

So \|S\| = \sin\theta. Thus C = (1-S^2)^{1/2} \ge \cos\theta I and \| I - C \| = 1-\cos\theta. Therefore

Hence \|U-P\| = \sqrt{2(1-\cos\theta)} = 2 \sin \frac\theta2.

We may suppose that \operatorname{rank} (P_1+P_2)^\perp \le \operatorname{rank} (Q_1+Q_2)^\perp. If both are infinite, there is nothing to prove. Thus we may assume that \operatorname{rank} (P_1+P_2)^\perp is finite. Form the polar decomposition

Then U = U_1+U_2 is a partial isometry of (P_1+P_2)H onto (Q_1+Q_2)H. Thus U is left semi-Fredholm since it has finite nullity and closed range. However by the lemma,

because the two terms in the 1 \times 2 matrix have orthogonal domains. If \operatorname{rank} (Q_1+Q_2)^\perp = \infty, then \pi(U) would be a proper isometry, and thus \sigma_e(U) = \overline{{\mathbb{D}}} is the closed unit disc. If \operatorname{rank} (Q_1+Q_2)^\perp < \infty, then \pi(U) is unitary. If also \operatorname{ind} U \ne 0, then \sigma_e(U) = \overline{{\mathbb{T}}} is the unit circle. In either case,

Hence \operatorname{ind} U = 0, which means \operatorname{rank} (P_1+P_2)^\perp = \operatorname{rank} (Q_1+Q_2)^\perp.

If this result were false, we could have \|P - Q_i \| < 1 for i=1,2. Using P, P^\perp, Q_1 and Q_2^\perp in Theorem 2.2, we would obtain

This contradiction establishes the result.

First of all, P-Q = PQ^\perp - P^\perp Q is the difference of two operators with pairwise orthogonal domains and ranges. Thus

We make use of the general form for a pair of projections [Halmos]. There is a canonical decomposition {\mathcal{H}} = {\mathcal{H}}_{00} \oplus {\mathcal{H}}_{10} \oplus {\mathcal{H}}_{01} \oplus {\mathcal{H}}_{11} \oplus {\mathcal{H}}_2 \oplus {\mathcal{H}}_2 and positive injective operators C,S \in {\mathcal{B}}(H_2) such that C^2+S^2 = I and

Any of the spaces {\mathcal{H}}_{ij} and {\mathcal{H}}_2 can be vacuous. It follows that

Now

If \| P - Q \| < 1, then {\mathcal{H}}_{10} and {\mathcal{H}}_{01} must be vacuous. Hence

However if either {\mathcal{H}}_{10} or {\mathcal{H}}_{01} is non-trivial, then \|PQ^\perp\| = 1 or \|PQ^\perp\| = 1, respectively. If say {\mathcal{H}}_{10} \ne \{0\} and {\mathcal{H}}_{01} = \{0\}, then \|PQ^\perp\|=1 and \|P^\perp Q\| = \|S\|. So if \|S\|<1, the two norms differ. It is similar if {\mathcal{H}}_{10} = \{0\} and {\mathcal{H}}_{01} \ne \{0\}.

For each M\in {\mathcal{M}}, there is an N \in {\mathcal{N}} with \| P_M - P_N \| \le \gamma. By Corollary 2.3, the subspace N is unique. Define \theta(M) = N. Reversing the roles of {\mathcal{M}} and {\mathcal{N}} shows that \theta is a bijection. It is evident that this map must preserve order, so \theta is an order isomorphism.

Suppose that M_1 < M_2, P_i := P_{M_i} and Q_i = P_{\theta(M_i)} for i=1,2. Then \|P_1 - Q_1\| < 1, \| P_2^\perp - Q_2^\perp\| = \|P_2 - Q_2 \| < 1, P_1 and P_2^\perp are orthogonal and Q_1 < Q_2, so that Q_1 and Q_2^\perp are orthogonal. Hence by Theorem 2.2,

This means that \theta preserves dimension.

By Theorem 3.1, \theta preserves dimension. Hence by the Similarity theorem [Dav_sim], \theta is implemented by a similarity. (See the introduction for some discussion of this.) When \gamma < \frac12, [DavNestAlgs]*Lemma 18.3 shows that one can choose S so that \| S - I \| < 2 \gamma. For \gamma \ge \frac12, one can take S = U+K with U unitary and \|K\| < \varepsilon < \frac12. Hence

Let \varepsilon > 0. We may choose a unit vector \eta \in N^\perp so that \|P_M \eta \| > \|P_M P_N^\perp \| - \varepsilon. Likewise choose a unit vector \zeta \in N_+ so that \| P_M^\perp \zeta \| > \| P_M^\perp P_{N_+} \| - \varepsilon. The rank one operator \zeta \eta^* lies in {\mathcal{T}}({\mathcal{N}}) and has norm 1. Therefore

As \varepsilon>0 is arbitrary, the result follows.

For the contrapositive, we suppose that d({\mathcal{M}},{\mathcal{N}}) = 1 and show that d({\mathcal{T}}({\mathcal{M}}), {\mathcal{T}}({\mathcal{N}})) = 1.

There are two cases to consider. In the first case, there are projections M_i\in{\mathcal{M}} and {\mathcal{N}}_i \in {\mathcal{N}} such that \| P_{M_i} - P_{N_i} \| = t_i < 1 for i\ge1 with \sup t_i = 1. In this case, Proposition 2.4 shows that \| P_{M_i}^\perp P_{N_i} \| = \| P_{M_i} P_{N_i}^\perp \| = t_i. Thus by Corollary 4.3,

In the second case, there is a projection in one nest at distance 1 from the other. Interchanging the two nests if required, we may suppose that there is M \in {\mathcal{M}} such that \| P_M - P_N \| = 1 for all N\in {\mathcal{N}}. Let

Again Corollary 4.3 and Proposition 2.4 show that d({\mathcal{T}}({\mathcal{M}}), {\mathcal{T}}({\mathcal{N}})) \ge \delta. Hence if \delta = 1, we are done.

So we assume that \delta < 1. If \| P_M^\perp P_N \| \le \delta, the same holds for all N'<N in {\mathcal{N}}. Define N_0 = \bigvee \{ N : \| P_M^\perp P_N \| \le \delta \}. There are N_i increasing to N_0 with \| P_M^\perp P_{N_i} \| \le \delta. Since P_{N_i} converges in the strong operator topology to P_{N_0}, we have \| P_M^\perp P_{N_0} \| \le \delta. That is, N_0 is the largest element of {\mathcal{N}} with this property. Hence \|P_M P_{N_0}^\perp \| = 1.

If N>N_0, then \| P_M^\perp P_N \| = 1, and thus \|P_M P_N^\perp \| \le \delta. Recall that N_{0+} = \bigwedge \{ N \in {\mathcal{N}} : N > N_0 \}. Arguing as above, we obtain \|P_M P_{N_{0+}}^\perp \| \le \delta. Thus \| P_M^\perp P_{N_{0+}} \| = 1. This shows that P_{N_{0+}} > P_{N_0}. Now an application of Lemma 4.2 shows that