Lemma 1

Let \(\mathcal{A}\) be a Banach algebra with unit \(\mathbf{1}\). Suppose \(x\) and \(y\) are idempotent and

\[\lVert x - y \rVert < \frac{1}{\lVert 2x - \mathbf{1} \rVert}.\]

Then \(x\) and \(y\) are conjugate, i.e., \(x = aya^{-1}\) for some \(a \in \mathcal{A}\).

In other words, conjugacy classes partition the set of idempotent elements into open subsets.

Proof

Displayed below is our candidate \(a\). \(a = \mathbf{1} - x - y + 2xy.\) Observe

\[xa = x - x^{2} - xy + 2x^{2}y = xy\]

and

\[ay = y - xy - y^{2} + 2xy^{2} = xy.\]

If \(a\) is invertible, then \(x = aya^{-1}.\) It therefore remains only to prove the invertibility of \(a\), and we do so by invoking Proposition 1 of the previous post.

I compute:

\[\mathbf{1} - a = x + y - 2xy = (x - y)(2x - \mathbf{1}).\]

This readily implies

\[\lVert a - \mathbf{1} \rVert \leq \lVert x - y \rVert \lVert 2x - \mathbf{1} \rVert < 1.\]

Now apply Proposition 1 as planned. \(\tag*{$\blacksquare$}\)

Note that \(\lVert 2x - \mathbf{1} \rVert\) is nonzero for idempotent \(x\), since \(x = (2x - \mathbf{1})x\) implies \(1 \leq \lVert 2x - \mathbf{1} \rVert.\)

I now specialize to the following case. Let \(X\) be a compact Hausdorff space. Let \(M(n,C(X))\) be the Banach algebra of \(n \times n\) matrices with entries in the ring of continuous \(\mathbf{C}\)-valued functions on \(X\) equipped with the supremum norm.

Lemma 2

Elements \(E\) of \(M(n,C(X))\) determine continuous functions \(E\colon X \to M(n,\mathbf{C})\).

Proof

Fix \(A \in M(n,\mathbf{C})\) and \(\epsilon > 0\). Consider the open ball \(U_{\epsilon}(A)\). I claim \(E^{-1}(U_{\epsilon}(A))\) is open in \(X\). To this end, fix \(p \in E^{-1}(U_{\epsilon}(A))\).

\[\lVert E(p) - A \lVert < \epsilon.\]

Define

\[\delta(p) = \frac{\epsilon - \lVert E(p) - A \lVert}{n^{2}}.\]

I claim that

\[V(p) = \bigcap_{1 \leq i,j \leq n} E_{ij}^{-1}(U_{\delta(p)}(E_{ij}(p)))\]

is an open neighborhood of \(p\) contained within \(E^{-1}(U_{\epsilon}(A))\).

Certainly \(p\) lies in \(V(p)\). Suppose \(q \in X\) too lies in \(V(p)\). For any vector \(v \in \mathbf{C}^{n}\) of unit length,

\[\left| (E(q) - E(p))v \right| \leq n \max_{1 \leq i \leq n} \left| \sum_{j=1}^{n} (E_{ij}(q) - E_{ij}(p))v_{j} \right|\] \[\leq n \max_{1 \leq i \leq n} \sum_{j=1}^{n} \left| E_{ij}(q) - E_{ij}(p) \right| |v_{j}|\] \[< n \max_{1 \leq i \leq n} \sum_{j=1}^{n} \delta(p) |v_{j}|\] \[\leq \epsilon - \lVert E(p) - A \rVert.\]

This computation reveals

\[\lVert E(q) - A \rVert \leq \lVert E(q) - E(p) \rVert + \lVert E(p) - A \rVert < \epsilon.\]

That is, \(q \in E^{-1}(U_{\epsilon}(A))\). \(\tag*{$\blacksquare$}\)

Proposition 1

Let \(E\colon X \to \operatorname{M}(n,\mathbf{C})\) be an idempotent element of \(\operatorname{M}(n,C(X))\). Then \(\operatorname{Ran} E\) is a subbundle of \(\Theta^{n}(X)\).

Proof

Fix \(x \in X\). Define

\[V = \left\{ F \in \operatorname{M}(n,\mathbf{C}) : \lVert E(x) - F \rVert < \frac{1}{\lVert 2E(x) - I_{n} \rVert} \right\}.\]

Define \(U = E^{-1}(V)\). By Lemma 2, \(U\) is an open subset of \(X\). We prove \(U\) is a trivial neighborhood of \(x\).

Define \(S\colon U \to \operatorname{GL}(n,\mathbf{C})\) by the formula below.

\[S(y) = I_{n} - E(x) - E(y) + E(x)E(y).\]

By Lemma 1, \(E(y) = S(y)^{-1}E(x)S(y).\) Hence, the function \(S(y)\colon (\operatorname{Ran} E)_{y} \to (\operatorname{Ran} E)_{x}\) is an isomorphism. \(\tag*{$\blacksquare$}\)

The obvious generalization to endomorphisms of vector bundles does not hold. Consider for instance the trivial bundle over the unit interval together with the endomorphism

\[(t,\lambda) \mapsto (t,t\lambda),\]

where \((t,\lambda) \in [0,1] \times \mathbf{R}\). The stalk at \(0\) is zero-dimensional.