Jordan Normal Forms of Endomorphisms of Vector Bundles over Curves and Applications to Moduli Space Automorphisms

Abstract

Let X be a compact connected Riemann surface of genus $g\ge 2$ and let E be a holomorphic vector bundle of rank n over X. The compactness and connectedness of X imply that the characteristic polynomial of any holomorphic endomorphism $\phi \in H^0(X,\mathcal{E}nd\left(E\right))$ has constant coefficients, a fact we call the Principle of Spectral Constancy. As a consequence, the eigenvalues of $\phi$ are globally constant over X, the primary decomposition of E with respect to $\phi$ consists of globally defined holomorphic subbundles, and the Jordan decomposition $\phi ={\phi }_s+{\phi }_n$ into semisimple and nilpotent parts is globally well defined as a decomposition of sections of $\mathcal{E}nd\left(E\right)$. This paper provides a systematic analysis of Jordan normal forms for endomorphisms of holomorphic vector bundles over X, relating the Jordan type of $\phi$ to the stability properties of E. In particular, it is proved that endomorphisms of stable bundles are necessarily scalar, that the Jordan decomposition of an endomorphism of a polystable bundle is determined componentwise by the classical Jordan normal forms of matrices in the associated endomorphism algebra, and that finite-order endomorphisms are always semisimple. These results are applied to the study of fixed points of automorphisms of the moduli space ${\mathcal{B}}_X\left(\mathrm{SL}(n,\mathbb{C})\right)$ of rank n and trivial determinant polystable vector bundles over X. Specifically, a new result establishes that the commutative subalgebra of $H^0(X,\mathcal{E}nd\left(E\right))$ generated by the endomorphism associated with a fixed-point condition is semisimple, so nilpotent endomorphisms of E are precisely those incompatible with the fixed-point structure.

1. Introduction

In linear algebra the Jordan normal form theorem asserts that every endomorphism of a finite-dimensional complex vector space decomposes uniquely as the sum of a semisimple part and a nilpotent part, and that the associated generalized eigenspaces give a direct sum decomposition of the space. This important result underlies applications in analysis, geometry, and related fields, ranging from the solution of systems of linear differential equations and the analysis of Markov chains to the study of dynamical systems and signal processing [1].

A natural generalization arises when one replaces a single vector space by a holomorphic family of vector spaces parametrized by a compact complex manifold. Such families are precisely holomorphic vector bundles, and the natural notion of a linear map in this context is a global section of the endomorphism bundle. This paper develops a Jordan normal form theory for holomorphic endomorphisms of vector bundles over a compact Riemann surface and applies it to the geometry of the associated moduli spaces.

The central new input is the Principle of Spectral Constancy: if X is a compact connected Riemann surface of genus $g\ge 2$ and E is a holomorphic vector bundle of rank n over X, then the characteristic polynomial of any holomorphic endomorphism $\phi \in H^0(X,\mathcal{E}nd\left(E\right))$ has globally constant coefficients. This follows from the fact that the coefficients of the characteristic polynomial of ${\phi }_x$ are holomorphic functions of $x\in X$ and hence constant by compactness. As a consequence, the eigenvalues of $\phi$ are globally defined complex numbers, the generalized eigenspace subsheaves are globally defined holomorphic subbundles, and the Jordan decomposition $\phi ={\phi }_s+{\phi }_n$ into semisimple and nilpotent parts is globally well defined as a decomposition of sections of $\mathcal{E}nd\left(E\right)$. The moduli space of holomorphic vector bundles over X was first constructed by Narasimhan and Seshadri [2,3] and generalized by Ramanathan [4,5,6] to the context of principal bundles with a reductive complex structure group; the Jordan theory developed here provides new structural information about the bundles parametrized by these spaces.

The relationship between Jordan type and stability is a central theme. Endomorphisms of stable bundles are necessarily scalar multiples of the identity—a bundle-theoretic analogue of Schur’s Lemma—so stable bundles admit no nontrivial Jordan structure. For polystable bundles, the endomorphism algebra $H^0(X,\mathcal{E}nd\left(E\right))$ is identified with a direct sum of matrix algebras ${⨁}_i\mathrm{M}(a_i,\mathbb{C})$, and the Jordan normal form of any endomorphism is determined componentwise by the classical Jordan normal forms of the corresponding matrices, with the stable summands playing the role of the scalar field. A key new result shows that any finite-order endomorphism, i.e., one satisfying ${\phi }^m=\theta {\mathrm{Id}}_E$ for some $m\ge 1$ and $\theta \in {\mathbb{C}}^{\ast }$, is necessarily semisimple. This follows from the squarefreeness of $t^m-\theta$ when $\theta \ne 0$ and provides the main tool for the applications.

As an application, we study fixed points of automorphisms of the moduli space ${\mathcal{B}}_X\left(\mathrm{SL}(n,\mathbb{C})\right)$ of rank n and trivial determinant polystable vector bundles over X. By the work of Kouvidakis and Pantev [7], the automorphism group of this space is generated for $n\ge 3$ by: the unique nontrivial outer automorphism $\sigma$ of $\mathrm{SL}(n,\mathbb{C})$ of order 2; tensorization by line bundles $L\in H^1(X,\mathbb{Z}/\left(n\right))$; and pull-back by elements of $\mathrm{Aut}\left(X\right)$. This was extended by Baraglia [8] to Higgs bundles, by Biswas, Gómez, and Muñoz to vector bundles with fixed determinant [9] and symplectic bundles [10]; and by Fringuelli [11] for principal G-bundles, where G is any semisimple complex Lie group. The fixed-point subvarieties for these automorphisms were studied in [12,13,14]; injective maps between moduli spaces whose images consist of fixed points were constructed by Serman [15] and extended to exceptional groups in [16]. The present paper provides the systematic linear-algebraic foundation for the geometric results of [12]. In each case, the endomorphism arising from the fixed-point condition satisfies a finite-order condition and is therefore semisimple by the results above, so that the primary decomposition applies and yields the geometric decompositions of [12]. A new result, Theorem 6, shows that the commutative subalgebra of $H^0(X,\mathcal{E}nd\left(E\right))$ generated by any fixed-point endomorphism consists entirely of semisimple elements, so that nilpotent endomorphisms of E are precisely those incompatible with the fixed-point structure.

The paper is organized as follows. Section 2 establishes the algebraic foundations: the Principle of Spectral Constancy, the primary decomposition, and the structure of endomorphism algebras. The Jordan decomposition theory and its relationship with stability is developed in detail in Section 3. In Section 4, we apply the theory to fixed-point analysis on ${\mathcal{B}}_X\left(\mathrm{SL}(n,\mathbb{C})\right)$. Finally, we draw the main conclusions of the paper and discuss open problems and lines for future research.

2. Preliminaries on Vector Bundles, Stability, and Endomorphisms

Throughout this paper, X denotes a compact connected Riemann surface of genus $g\ge 2$, and ${\mathcal{O}}_X$ denotes the sheaf of holomorphic functions on X.

Recall that a holomorphic vector bundle of rank n over X is a complex manifold E together with a holomorphic surjection $\pi :E\to X$ such that each fiber $E_x={\pi }^{-1}\left(x\right)$ is a complex vector space of dimension n, and E is locally holomorphically trivial: every point $x\in X$ admits an open neighborhood $U\subset X$ and a biholomorphism

$${\pi }^{-1}\left(U\right)\cong U\times {\mathbb{C}}^n$$

compatible with the projections and linear on each fiber. Equivalently, a holomorphic vector bundle of rank n over X is determined by an open cover $\left\{U_{\alpha }\right\}$ of X and holomorphic transition functions

$$g_{\alpha \beta }:U_{\alpha }\cap U_{\beta }\to \mathrm{GL}(n,\mathbb{C})$$

satisfying the cocycle condition

$$g_{\alpha \beta }{g}_{\beta \gamma }=g_{\alpha \gamma }$$

on triple overlaps. Two such collections define isomorphic bundles if and only if they are cohomologous [17].

The degree $\mathrm{deg}\left(E\right)$ of a holomorphic vector bundle E over X is defined as the degree of its determinant line bundle $\det \left(E\right)={\bigwedge }^{n}E$. The slope of E is

$$\mu \left(E\right)={\displaystyle \frac{\mathrm{deg}\left(E\right)}{\mathrm{rk}\left(E\right)}}.$$

The slope is additive in exact sequences: if

$$0\to F\to E\to Q\to 0$$

is an exact sequence of holomorphic vector bundles, then

$$\mathrm{rk}\left(E\right)\mu \left(E\right)=\mathrm{rk}\left(F\right)\mu \left(F\right)+\mathrm{rk}\left(Q\right)\mu \left(Q\right).$$

(1)

A holomorphic vector bundle E has trivial determinant if $\det \left(E\right)\cong {\mathcal{O}}_X$.

Definition 1.

Let E be a holomorphic vector bundle of rank n over X.

(i) 

E is stable if $\mu \left(F\right)<\mu \left(E\right)$ for every holomorphic subbundle $F⊊E$ with $0<\mathrm{rk}\left(F\right)<n$.

(ii) 

E is semistable if $\mu \left(F\right)\le \mu \left(E\right)$ for every such F.

(iii) 

E is polystable if

$$E\cong F_1\oplus \cdots \oplus F_s$$

where each $F_j$ is stable with $\mu \left(F_j\right)=\mu \left(E\right)$.

Remark 1.

Every stable bundle is semistable and polystable (with $s=1$). A polystable bundle with $s\ge 2$ is strictly polystable. Not every semistable bundle is polystable: a nontrivial extension of stable bundles of equal slope is semistable but neither stable nor polystable.

Theorem 1

(Narasimhan–Seshadri [3]; Seshadri [18]). Let X be a compact Riemann surface of genus $g\ge 2$ and $n\ge 1$, $d\in \mathbb{Z}$. There exists a coarse moduli space $\mathcal{M}(n,d)$ parametrizing S-equivalence classes of semistable holomorphic vector bundles of rank n and degree d over X. It is a normal projective variety of complex dimension $n^2(g-1)+1$, with the stable locus ${\mathcal{M}}^s(n,d)$ a smooth open dense subvariety. Every S-equivalence class contains a unique polystable representative.

In particular, for $n\ge 3$ and $d=0$, the moduli space

$${\mathcal{B}}_X\left(\mathrm{SL}(n,\mathbb{C})\right)=\left\{E\mathrm{polystable}:\mathrm{rk}\left(E\right)=n,\det \left(E\right)\cong {\mathcal{O}}_X\right\}/\cong$$

is a normal projective variety of complex dimension $(n^2-1)(g-1)$ [3,18].

Let E be a holomorphic vector bundle of rank n over X. The endomorphism bundle is $\mathcal{E}nd\left(E\right)=E\otimes E^{\ast }$, where $E^{\ast }=\mathcal{H}om(E,{\mathcal{O}}_X)$. Its fiber over $x\in X$ is

$${\mathrm{End}}_{\mathbb{C}}\left(E_x\right)\cong \mathrm{M}(n,\mathbb{C}),$$

and it has rank $n^2$. A global section $\phi \in H^0(X,\mathcal{E}nd\left(E\right))$ is a holomorphic endomorphism of E, assigning to each $x\in X$ a linear map

$${\phi }_x:E_x\to E_x$$

holomorphically. The space $H^0(X,\mathcal{E}nd\left(E\right))$ is a finite-dimensional complex algebra under pointwise composition.

Lemma 1 (Principle of Spectral Constancy).

Let E be a holomorphic vector bundle of rank n over X and $\phi \in H^0(X,\mathcal{E}nd\left(E\right))$. Then:

(i) 

The characteristic polynomial of φ is independent of $x\in X$: there exist $c_0,\dots ,c_{n-1}\in \mathbb{C}$ such that

$$\det (t{\mathrm{Id}}_{E_x}-{\phi }_x)=t^n+c_{n-1}{t}^{n-1}+\cdots +c_0$$

for all $x\in X$.

(ii) 

The eigenvalues of φ (with algebraic multiplicities) are constant over X.

(iii) 

φ satisfies its characteristic polynomial globally:

$${\phi }^n+c_{n-1}{\phi }^{n-1}+\cdots +c_0{\mathrm{Id}}_E=0$$

in $H^0(X,\mathcal{E}nd\left(E\right)).$

Proof. 

Part (i): The coefficient

$$c_{n-k}\left(x\right)={(-1)}^k\mathrm{tr}\left(\stackrel{k}{\bigwedge }{\phi }_x\right)$$

is holomorphic in x (as a polynomial in the entries of $\phi$ in any local trivialization). Since X is compact and connected, every holomorphic function on X is constant by the maximum modulus principle [17]. Hence $c_{n-k}\left(x\right)$ is independent of x.

Part (ii): The eigenvalues of ${\phi }_x$ are the roots of

$$\det (t\mathrm{Id}-{\phi }_x),$$

which is independent of x by part (i).

Part (iii): The Cayley–Hamilton theorem gives ${\chi }_{\phi }\left({\phi }_x\right)=0$ for every $x\in X$. Since ${\chi }_{\phi }$ has constant coefficients by part (i), the corresponding section of $\mathcal{E}nd\left(E\right)$ vanishes identically. □

Remark 2.

The compactness and connectedness of X are both essential. On a non-compact Riemann surface, holomorphic functions need not be constant, and endomorphisms of the trivial bundle ${\mathcal{O}}^{\oplus n}$ can have non-constant eigenvalues.

Corollary 1.

The minimal polynomial ${\mu }_{\phi }\in \mathbb{C}\left[t\right]$ of φ, defined as the monic polynomial of least degree with ${\mu }_{\phi }\left(\phi \right)=0$ in $H^0(X,\mathcal{E}nd\left(E\right))$, exists, is unique, and divides ${\chi }_{\phi }$, and every root of ${\mu }_{\phi }$ is an eigenvalue of φ.

Proof. 

The set

$$\left\{p\in \mathbb{C}\left[t\right]:p\left(\phi \right)=0\right\}$$

is a nonzero ideal of $\mathbb{C}\left[t\right]$ (nonzero by Lemma 1(iii)), hence principal, generated by a unique monic ${\mu }_{\phi }$. Since ${\chi }_{\phi }\in \mathcal{I}$, we have ${\mu }_{\phi }\mid {\chi }_{\phi }$. If $\lambda$ is a root of ${\mu }_{\phi }$, then ${\mu }_{\phi }\left({\phi }_x\right)=0$ for every x, so the minimal polynomial of ${\phi }_x$ divides ${\mu }_{\phi }$, hence $\lambda$ is an eigenvalue of ${\phi }_x$ for every x. □

Lemma 2 (Constant-rank theorem for bundle maps,

([19], Lemma 1.1.4)). Let $f:E\to F$ be a holomorphic morphism between holomorphic vector bundles over a complex manifold M. If the rank of $f_x$ is constant for all $x\in M$, then $\ker \left(f\right)$ and $\mathrm{Im}\left(f\right)$ are holomorphic subbundles of E and F respectively.

Proposition 1.

Let E be a holomorphic vector bundle of rank n over X, $\phi \in H^0(X,\mathcal{E}nd\left(E\right))$, and ${\lambda }_1,\dots ,{\lambda }_k$ the distinct eigenvalues of φ with algebraic multiplicities $m_1,\dots ,m_k$ and ${\sum }_j{m}_j=n$. Set

$${\mathcal{F}}_{{\lambda }_j}=\ker \left({\left(\phi -{\lambda }_j{\mathrm{Id}}_E\right)}^{m_j}\right).$$

Then:

(i) 

Each ${\mathcal{F}}_{{\lambda }_j}$ is a holomorphic vector subbundle of E of rank $m_j$.

(ii) 

The vector bundle E admits a decomposition

$$E={\mathcal{F}}_{{\lambda }_1}\oplus \cdots \oplus {\mathcal{F}}_{{\lambda }_k}$$

into a direct sum of holomorphic vector bundles.

(iii) 

The endomorphism φ preserves each summand, and

$$(\phi -{\lambda }_j\mathrm{Id}){|}_{{\mathcal{F}}_{{\lambda }_j}}$$

is nilpotent of index at most $m_j$.

Proof. 

Part (i): The algebraic multiplicity $m_j$ of ${\lambda }_j$ is constant over X by Lemma 1(ii). It is a standard linear algebra fact that the dimension of

$$\ker \left({({\phi }_x-{\lambda }_j\mathrm{Id})}^{m_j}\right)$$

equals $m_j$ for every x. Hence ${\left(\phi -{\lambda }_j\mathrm{Id}\right)}^{m_j}$ has constant rank $n-m_j$, and Lemma 2 gives that ${\mathcal{F}}_{{\lambda }_j}$ is a holomorphic subbundle of rank $m_j$.

Part (ii): Since the polynomials ${(t-{\lambda }_j)}^{m_j}$ are pairwise coprime and

$${\chi }_{\phi }=\prod _j{(t-{\lambda }_j)}^{m_j},$$

the Chinese Remainder Theorem in $\mathbb{C}\left[t\right]$ yields polynomials $p_j\in \mathbb{C}\left[t\right]$ with

$$p_j\equiv 1\left(\mathrm{mod}{(t-{\lambda }_j)}^{m_j}\right)$$

and

$$p_j\equiv 0\left(\mathrm{mod}{(t-{\lambda }_i)}^{m_i}\right)$$

for $i\ne j$. Explicitly, with

$$q_j=\prod _{i\ne j}{(t-{\lambda }_i)}^{m_i},$$

choose $u_j,v_j\in \mathbb{C}\left[t\right]$ with

$$u_j{q}_j+v_j{(t-{\lambda }_j)}^{m_j}=1$$

(Bézout) and set $p_j=u_j{q}_j$. The sections

$${\pi }_j=p_j\left(\phi \right)\in H^0(X,\mathcal{E}nd\left(E\right))$$

are orthogonal idempotents with ${\sum }_j{\pi }_j={\mathrm{Id}}_E$, and $\mathrm{Im}\left({\pi }_j\right)={\mathcal{F}}_{{\lambda }_j}$, giving the direct sum decomposition.

Part (iii): Since ${\pi }_j$ commutes with $\phi$ (being a polynomial in $\phi$), each ${\mathcal{F}}_{{\lambda }_j}$ is $\phi$-invariant, and

$${(\phi -{\lambda }_j\mathrm{Id})}^{m_j}{|}_{{\mathcal{F}}_{{\lambda }_j}}=0$$

by definition. □

Remark 3.

The decomposition

$$E=\underset{j}{⨁}{\mathcal{F}}_{{\lambda }_j}$$

is the bundle-theoretic primary decomposition theorem. Its proof is global because the idempotents ${\pi }_j=p_j\left(\phi \right)$ are polynomial expressions in φ with constant coefficients, which follows from the Principle of Spectral Constancy.

Proposition 2.

Let E be a stable holomorphic vector bundle over X. Then

$$H^0(X,\mathcal{E}nd\left(E\right))=\mathbb{C}·{\mathrm{Id}}_E.$$

Proof. 

Let $\phi \in H^0(X,\mathcal{E}nd\left(E\right))$ with eigenvalue $\lambda$. Set

$$\mathcal{K}=\ker \left(\phi -\lambda {\mathrm{Id}}_E\right)$$

and

$$\mathcal{I}=\mathrm{Im}(\phi -\lambda {\mathrm{Id}}_E).$$

Both have constant rank by Lemmas 1(ii) and 2, with

$$\mathrm{rk}\left(\mathcal{K}\right)+\mathrm{rk}\left(\mathcal{I}\right)=n$$

and

$$\mathrm{deg}\left(\mathcal{K}\right)+\mathrm{deg}\left(\mathcal{I}\right)=\mathrm{deg}\left(E\right).$$

If $\mathcal{K}⊊E$, then both $\mathcal{K}$ and $\mathcal{I}$ are proper nonzero subbundles of E, so $\mu \left(\mathcal{K}\right)<\mu \left(E\right)$ and $\mu \left(\mathcal{I}\right)<\mu \left(E\right)$ by stability, giving

$$\mathrm{deg}\left(\mathcal{K}\right)+\mathrm{deg}\left(\mathcal{I}\right)<n\mu \left(E\right)=\mathrm{deg}\left(E\right),$$

which is a contradiction. Hence $\mathcal{K}=E$ and $\phi =\lambda {\mathrm{Id}}_E$, concluding the result. □

Remark 4.

Proposition 2 is the bundle-theoretic Schur’s Lemma: a stable bundle is a simple object whose only endomorphisms are scalar multiples of the identity. In particular, every stable bundle is indecomposable.

Proposition 3.

Let

$$E=F_1^{\oplus a_1}\oplus \cdots \oplus F_s^{\oplus a_s}$$

be polystable, with $F_1,\dots ,F_s$ pairwise non-isomorphic stable bundles and $\mu \left(F_i\right)=\mu \left(E\right)$ for all i. Then

$$H^0(X,\mathcal{E}nd\left(E\right))\cong \mathrm{M}(a_1,\mathbb{C})\oplus \cdots \oplus \mathrm{M}(a_s,\mathbb{C})$$

as complex algebras.

Proof. 

By the Künneth decomposition,

$$H^0(X,\mathcal{E}nd\left(E\right))\cong \underset{i,j}{⨁}{H}^0{(X,\mathcal{H}om(F_j,F_i))}^{\oplus a_i{a}_j}.$$

For $i\ne j$: any nonzero $f:F_j\to F_i$ is injective (since $\ker \left(f\right)$ is a proper subbundle of $F_j$ with

$$\mu \left(\ker f\right)<\mu \left(F_j\right)=\mu \left(F_i\right),$$

contradicting stability of $F_i$ via the slope formula) and surjective (since $\mathrm{Im}\left(f\right)$ is a proper subbundle of $F_i$ with

$$\mu \left(\mathrm{Im}f\right)=\mu \left(F_j\right)=\mu \left(F_i\right),$$

which by stability of $F_i$ forces $\mathrm{Im}\left(f\right)=F_i$), hence an isomorphism, contradicting $F_i\ncong F_j$. So $H^0(X,\mathcal{H}om(F_j,F_i))=0$ for $i\ne j$.

For $i=j$: $H^0(X,\mathcal{E}nd\left(F_i\right))\cong \mathbb{C}$ by Proposition 2. Then, the result follows. □

Remark 5.

By the Artin–Wedderburn theorem, $H^0(X,\mathcal{E}nd\left(E\right))$ is a semisimple $\mathbb{C}$-algebra, being a direct sum of matrix algebras.

Corollary 2.

Let E be a holomorphic vector bundle over X and $\phi \in H^0(X,\mathcal{E}nd\left(E\right))$. If ${\phi }^m=\theta {\mathrm{Id}}_E$ for some $m\ge 1$ and $\theta \in {\mathbb{C}}^{\ast }$, then φ is semisimple.

Proof. 

The minimal polynomial ${\mu }_{\phi }$ divides $t^m-\theta$. We claim that $t^m-\theta$ is squarefree. To see this, note that any common factor of $t^m-\theta$ and its formal derivative $m{t}^{m-1}$ in $\mathbb{C}\left[t\right]$ must divide both $t^{m-1}$ and $t^m-\theta$; subtracting, it divides $-\theta$, which is a nonzero constant since $\theta \ne 0$. Equivalently, $\gcd (t^m-\theta ,m{t}^{m-1})=1$ in $\mathbb{C}\left[t\right]$, so $t^m-\theta$ has no repeated roots in $\mathbb{C}$; i.e., it is squarefree. Hence ${\mu }_{\phi }$ is squarefree and $\phi$ is semisimple. □

Corollary 3.

Let

$$E=F_1^{\oplus a_1}\oplus \cdots \oplus F_s^{\oplus a_s}$$

be polystable as above, and

$$\phi =({\phi }_1,\dots ,{\phi }_s)$$

with ${\phi }_i\in \mathrm{M}(a_i,\mathbb{C})$. Then

• φ is semisimple if and only if each ${\phi }_i$ is diagonalizable.
• φ is nilpotent if and only if each ${\phi }_i$ is nilpotent.
• φ is an automorphism if and only if each ${\phi }_i$ is invertible.

Proof. 

All three statements follow from the identification

$$H^0(X,\mathcal{E}nd\left(E\right))\cong \underset{i}{⨁}\mathrm{M}(a_i,\mathbb{C})$$

of Proposition 3 and the corresponding classical characterizations in matrix algebras over $\mathbb{C}$. For the first statement, notice that the minimal polynomial of $\phi$ is ${\mathrm{lcm}}_i\left({\mu }_{{\phi }_i}\right)$, which is squarefree if and only if each ${\mu }_{{\phi }_i}$ is squarefree. □

3. Jordan Decomposition for Endomorphisms of Vector Bundles

The goal of this section is to develop the Jordan decomposition theory for holomorphic endomorphisms of vector bundles over X, building on the foundations of Section 2. Throughout, X is a compact connected Riemann surface of genus $g\ge 2$.

Theorem 2 (Jordan Decomposition for Bundle Endomorphisms).

Let E be a holomorphic vector bundle of rank n over X and $\phi \in H^0(X,\mathcal{E}nd\left(E\right))$. There exist unique holomorphic endomorphisms ${\phi }_s,{\phi }_n\in H^0(X,\mathcal{E}nd\left(E\right))$ such that:

(i) 

$\phi ={\phi }_s+{\phi }_n$;

(ii) 

${\phi }_s$ is semisimple;

(iii) 

${\phi }_n$ is nilpotent;

(iv) 

${\phi }_s\circ {\phi }_n={\phi }_n\circ {\phi }_s$;

(v) 

Both ${\phi }_s$ and ${\phi }_n$ are polynomials in φ with coefficients in $\mathbb{C}$.

Proof. 

Let ${\lambda }_1,\dots ,{\lambda }_k$ be the distinct eigenvalues of $\phi$ with multiplicities $m_1,\dots ,m_k$, and let

$$E=\underset{j}{⨁}{\mathcal{F}}_{{\lambda }_j}$$

with idempotents ${\pi }_j=p_j\left(\phi \right)$ as in Proposition 1. Set

$${\phi }_s=\sum _j{\lambda }_j{\pi }_j$$

and ${\phi }_n=\phi -{\phi }_s$. Both are polynomials in $\phi$ with constant coefficients, establishing $\left(v\right)$. The endomorphism ${\phi }_s$ acts as ${\lambda }_j\mathrm{Id}$ on ${\mathcal{F}}_{{\lambda }_j}$, so its minimal polynomial is

$$\prod _j(t-{\lambda }_j),$$

squarefree, giving part (ii). On ${\mathcal{F}}_{{\lambda }_j}$,

$${\phi }_n=\phi -{\lambda }_j\mathrm{Id}$$

is nilpotent of index at most $m_j$, giving part (iii). Properties (i) and (iv) are clear.

For uniqueness, suppose

$$\phi ={\phi }_s^{\prime }+{\phi }_n^{\prime }$$

is another such decomposition. By property (v), ${\phi }_s^{\prime }$ and ${\phi }_n^{\prime }$ are polynomials in $\phi$, hence commuting with ${\phi }_s$ and ${\phi }_n$. Setting

$$A={\phi }_s-{\phi }_s^{\prime }={\phi }_n^{\prime }-{\phi }_n,$$

since ${\phi }_s$ and ${\phi }_s^{\prime }$ commute (both polynomials in $\phi$), they are simultaneously diagonalizable, so

$$A={\phi }_s-{\phi }_s^{\prime }$$

is semisimple; and

$$A={\phi }_n^{\prime }-{\phi }_n$$

is nilpotent by the binomial theorem applied to commuting nilpotents. A semisimple nilpotent endomorphism has minimal polynomial dividing both

$$\prod _j(t-{\lambda }_j)$$

(squarefree) and $t^N$, hence dividing t, so $A=0$. □

Remark 6.

Theorem 2 is the global bundle-theoretic analogue of the additive Jordan–Chevalley decomposition ([20], Chapter I). The globalness of ${\phi }_s$ and ${\phi }_n$ follows from the constancy of the interpolating coefficients, itself a consequence of the Principle of Spectral Constancy (Lemma 1).

Proposition 4.

Let E be a holomorphic vector bundle of rank n over X and ${\phi }_n\in H^0(X,\mathcal{E}nd\left(E\right))$ nilpotent of index $N\ge 1$. Set

$${\mathcal{K}}_{\ell }=\ker \left({\phi }_n^{\ell }:E\to E\right)$$

with ${\phi }_n^0={\mathrm{Id}}_E$. Then:

(i) 

Each ${\mathcal{K}}_{\ell }$ is a holomorphic subbundle of E, with $\mathrm{rk}\left({\mathcal{K}}_0\right)=0$ and $\mathrm{rk}\left({\mathcal{K}}_N\right)=n$.

(ii) 

The subbundles form a strictly increasing filtration

$$0={\mathcal{K}}_0⊊{\mathcal{K}}_1⊊\cdots ⊊{\mathcal{K}}_N=E,$$

called the Jordan filtration of ${\phi }_n$.

(iii) 

We have

$${\phi }_n\left({\mathcal{K}}_{\ell }\right)\subset {\mathcal{K}}_{\ell -1}$$

for $\ell =1,\dots ,N$.

(iv) 

For $\ell =2,\dots ,N$, ${\phi }_n$ induces a well-defined injective holomorphic bundle map

$${\overline{\phi }}_n:{\mathcal{G}}_{\ell }={\mathcal{K}}_{\ell }/{\mathcal{K}}_{\ell -1}\hookrightarrow {\mathcal{G}}_{\ell -1}={\mathcal{K}}_{\ell -1}/{\mathcal{K}}_{\ell -2}.$$

Proof. 

Part (i): Since ${\phi }_n^{\ell }\in H^0(X,\mathcal{E}nd\left(E\right))$, Lemma 1(i) applied to ${\phi }_n^{\ell }$ gives that its characteristic polynomial has constant coefficients, so its rank is constant over X. By Lemma 2,

$${\mathcal{K}}_{\ell }=\ker \left({\phi }_n^{\ell }\right)$$

is a holomorphic subbundle.

Part (ii): The inclusions

$${\mathcal{K}}_{\ell }\subseteq {\mathcal{K}}_{\ell +1}$$

are immediate. For strictness: if

$$\ker \left({\phi }_n^{\ell }\right)=\ker \left({\phi }_n^{\ell +1}\right),$$

then by induction on j one shows

$$\ker \left({\phi }_n^{\ell }\right)=\ker \left({\phi }_n^{\ell +j}\right)$$

for all $j\ge 0$ (if $v\in \ker \left({\phi }_n^{\ell +j+1}\right)$, then ${\phi }_n^{\ell +j}\left({\phi }_{n}v\right)=0$, so ${\phi }_{n}v\in \ker \left({\phi }_n^{\ell }\right)$ by hypothesis, giving ${\phi }_n^{\ell +1}v=0$; hence $v\in \ker \left({\phi }_n^{\ell +1}\right)=\ker \left({\phi }_n^{\ell }\right)$). Taking $j=N-\ell$ gives $\ker \left({\phi }_n^{\ell }\right)=E$; hence ${\phi }_n^{\ell }=0$, contradicting ${\phi }_n^{N-1}\ne 0$ when $\ell \le N-1$.

Part (iii): If $v\in {\mathcal{K}}_{\ell }$, then

$${\phi }_n^{\ell -1}\left({\phi }_{n}v\right)={\phi }_n^{\ell }v=0,$$

so ${\phi }_{n}v\in {\mathcal{K}}_{\ell -1}$.

Part (iv): By part (iii), ${\phi }_n$ maps

$${\mathcal{K}}_{\ell }\to {\mathcal{K}}_{\ell -1}$$

and

$${\mathcal{K}}_{\ell -1}\to {\mathcal{K}}_{\ell -2},$$

so $\left[v\right]\mapsto \left[{\phi }_{n}v\right]$ is well defined on ${\mathcal{G}}_{\ell }$.

To prove injectivity, suppose that ${\phi }_{n}v\in {\mathcal{K}}_{\ell -2}$. Then

$${\phi }_n^{\ell -1}v={\phi }_n^{\ell -2}\left({\phi }_{n}v\right)=0,$$

so $v\in {\mathcal{K}}_{\ell -1}$ and $\left[v\right]=0$. The rank of ${\overline{\phi }}_n$ is constant by the constancy of the $r_{\ell }$, so Lemma 2 gives the injective bundle map. □

Remark 7.

The Jordan filtration is the bundle-theoretic analogue of the standard flag of subspaces associated with a nilpotent linear map. Each step of the filtration yields a short exact sequence of holomorphic vector bundles

$$0⟶{\mathcal{K}}_{\ell -1}⟶{\mathcal{K}}_{\ell }⟶{\mathcal{G}}_{\ell }⟶0,$$

which defines an extension class $\left[{\epsilon }_{\ell }\right]\in {\mathrm{Ext}}^1({\mathcal{G}}_{\ell },{\mathcal{K}}_{\ell -1})$. The vanishing of all these classes is precisely the condition for the Jordan filtration to split holomorphically, i.e., for E to admit a global Jordan basis: a decomposition $E\cong {⨁}_{\ell }{\mathcal{G}}_{\ell }$ of holomorphic vector bundles compatible with ${\phi }_n$. When some $\left[{\epsilon }_{\ell }\right]\ne 0$, no such global splitting exists and ${\phi }_n$ cannot be brought to block-diagonal form by a holomorphic gauge transformation. This is analogous to the situation for indecomposable bundles over elliptic curves studied by Atiyah [21].

Theorem 3 (Jordan Normal Form for Polystable Bundles).

Let

$$E=F_1^{\oplus a_1}\oplus \cdots \oplus F_s^{\oplus a_s}$$

be polystable with $F_i$ pairwise non-isomorphic stable bundles, and identify

$$H^0(X,\mathcal{E}nd\left(E\right))\cong \underset{i}{⨁}\mathrm{M}(a_i,\mathbb{C})$$

via Proposition 3. Write

$$\phi =({\phi }_1,\dots ,{\phi }_s)$$

and let ${\phi }_i=D_i+N_i$ be the Jordan decomposition in $\mathrm{M}(a_i,\mathbb{C})$. Then:

(i) 

${\phi }_s=(D_1,\dots ,D_s)$ and ${\phi }_n=(N_1,\dots ,N_s)$.

(ii) 

Both ${\phi }_s$ and ${\phi }_n$ preserve each isotypic component $F_i^{\oplus a_i}$.

(iii) 

For each i, writing $E_i=F_i\otimes {\mathbb{C}}^{a_i}$, the Jordan filtration of ${\phi }_n{|}_{E_i}$ has subbundles

$${\mathcal{K}}_{i,\ell }=F_i\otimes \ker \left(N_i^{\ell }\right)$$

and successive quotients

$${\mathcal{G}}_{i,\ell }\cong F_i^{\oplus d_{i,\ell }},$$

where

$$d_{i,\ell }=\#\{j:b_{i,j}\ge \ell \}$$

and

$$b_{i,1}\ge \cdots \ge b_{i,c_i}\ge 1$$

are the Jordan block sizes of $N_i$.

Proof. 

Part (i): By Corollary 3, $(D_1,\dots ,D_s)$ is semisimple, $(N_1,\dots ,N_s)$ is nilpotent, their sum is $\phi$, and they commute componentwise. Uniqueness in Theorem 2 gives the identification.

Part (ii): Each component acts on $F_i^{\oplus a_i}$ through the i-th block.

Part (iii): Under the identification $E_i=F_i\otimes {\mathbb{C}}^{a_i}$, the transition functions of $E_i$ have the form $g_{\alpha \beta }\otimes I_{a_i}$, acting nontrivially on the $F_i$-factor and as the identity on ${\mathbb{C}}^{a_i}$. Hence $N_i$ acts as ${\mathrm{Id}}_{F_i}\otimes N_i$, and

$$\ker \left({\left({\phi }_n{|}_{E_i}\right)}^{\ell }\right)=F_i\otimes \ker \left(N_i^{\ell }\right)$$

is a well-defined holomorphic subbundle of $E_i$. The rank formula

$$d_{i,\ell }=\dim \ker \left(N_i^{\ell }\right)-\dim \ker \left(N_i^{\ell -1}\right)=\#\{j:b_{i,j}\ge \ell \}$$

is classical linear algebra. □

Remark 8.

The stable summand $F_i$ plays the role of the scalar field: Jordan blocks $J_b\left(\mu \right)\in \mathrm{M}(b,\mathbb{C})$ acting on ${\mathbb{C}}^b$ are replaced by $J_b\left(\mu \right)\otimes {\mathrm{Id}}_{F_i}$ acting on $F_i^{\oplus b}=F_i\otimes {\mathbb{C}}^b$.

Proposition 5.

Let E be semistable and ${\phi }_n\in H^0(X,\mathcal{E}nd\left(E\right))$ nilpotent of index N, with Jordan filtration $\left\{{\mathcal{K}}_{\ell }\right\}$ and quotients

$${\mathcal{G}}_{\ell }={\mathcal{K}}_{\ell }/{\mathcal{K}}_{\ell -1}.$$

Then

$$\mu \left({\mathcal{G}}_{\ell }\right)=\mu \left(E\right)$$

for all ℓ,

$$\mu \left({\mathcal{K}}_{\ell }\right)=\mu \left(E\right)$$

for all ℓ, and each ${\mathcal{K}}_{\ell }$ is semistable.

Proof. 

The upper bound follows since, composing the injective maps of Proposition 4(iv) gives

$${\overline{\phi }}_n^{\ell -1}:{\mathcal{G}}_{\ell }\hookrightarrow {\mathcal{K}}_1.$$

The image is a subbundle of ${\mathcal{K}}_1\subset E$, hence a subbundle of E, giving

$$\mu \left({\mathcal{G}}_{\ell }\right)\le \mu \left(E\right)$$

by semistability.

Degree additivity gives

$$\mathrm{deg}\left(E\right)=\sum _{\ell }\mathrm{rk}\left({\mathcal{G}}_{\ell }\right)\mu \left({\mathcal{G}}_{\ell }\right).$$

Since $\mathrm{rk}\left({\mathcal{G}}_{\ell }\right)\ge 1$ (by strict inclusions) and $\mu \left({\mathcal{G}}_{\ell }\right)\le \mu \left(E\right)$:

$$n\mu \left(E\right)=\sum _{\ell }\mathrm{rk}\left({\mathcal{G}}_{\ell }\right)\mu \left({\mathcal{G}}_{\ell }\right)\le \mu \left(E\right)\sum _{\ell }\mathrm{rk}\left({\mathcal{G}}_{\ell }\right)=n\mu \left(E\right).$$

Equality forces $\mu \left({\mathcal{G}}_{\ell }\right)=\mu \left(E\right)$ for all ℓ. Then

$$\mathrm{deg}\left({\mathcal{K}}_m\right)={\displaystyle \sum _{\ell =1}^m}\mathrm{rk}\left({\mathcal{G}}_{\ell }\right)\mu \left(E\right)=\mathrm{rk}\left({\mathcal{K}}_m\right)\mu \left(E\right),$$

giving $\mu \left({\mathcal{K}}_m\right)=\mu \left(E\right)$.

To prove semistability of ${\mathcal{K}}_{\ell }$, let $F\subset {\mathcal{K}}_{\ell }\subset E$ be any subbundle. Since F is also a subbundle of E, semistability of E gives

$$\mu \left(F\right)\le \mu \left(E\right)=\mu \left({\mathcal{K}}_{\ell }\right).$$

Hence ${\mathcal{K}}_{\ell }$ is semistable, concluding the proof. □

Theorem 4.

Let E be a holomorphic vector bundle over X and $\phi \in H^0(X,\mathcal{E}nd\left(E\right))$ with Jordan decomposition $\phi ={\phi }_s+{\phi }_n$.

(i) 

If E is stable, then ${\phi }_n=0$ and ${\phi }_s=\lambda {\mathrm{Id}}_E$ for some $\lambda \in \mathbb{C}$.

(ii) 

If E is polystable, then ${\phi }_s$ and ${\phi }_n$ preserve every isotypic component.

(iii) 

If ${\phi }^m=\theta {\mathrm{Id}}_E$ for some $m\ge 1$ and $\theta \in {\mathbb{C}}^{\ast }$, then ${\phi }_n=0$ and φ is semisimple with eigenvalues among the m-th roots of θ.

Proof. 

Part (i) follows directly from Proposition 2. Part (ii) is Theorem 3(i)–(ii). Part (iii) is a consequence of Corollary 2: ${\mu }_{\phi }\mid t^m-\theta$, which is squarefree since $\theta \ne 0$, so $\phi$ is semisimple and ${\phi }_n=0$. □

Remark 9.

The key result for this section is Theorem 4(iii): the endomorphisms arising from fixed-point conditions on the moduli space satisfy ${\phi }^N=\theta {\mathrm{Id}}_E$ for some N and $\theta \in {\mathbb{C}}^{\ast }$, so their semisimplicity—and hence the applicability of the primary decomposition of Proposition 1—follows automatically.

4. Applications to Fixed Points of Moduli Space Automorphisms

Let X be a compact connected Riemann surface of genus $g\ge 2$ and $n\ge 3$. By [7,12], the automorphism group of ${\mathcal{B}}_X\left(\mathrm{SL}(n,\mathbb{C})\right)$ is generated by: the unique nontrivial outer automorphism $\sigma$ of $\mathrm{SL}(n,\mathbb{C})$ of order 2; tensorization by $L\in H^1(X,\mathbb{Z}/\left(n\right))$; and pull-back by elements of $\mathrm{Aut}\left(X\right)$. This section provides the linear-algebraic framework for the fixed-point analysis of [12].

Let $L\in H^1(X,\mathbb{Z}/\left(n\right))$ be nontrivial of order m, so $L^m\cong {\mathcal{O}}_X$ and $L^k\ncong {\mathcal{O}}_X$ for $0<k<m$. Let E be a polystable rank n and trivial determinant vector bundle with

$$f:E\stackrel{\sim }{\to }E\otimes L.$$

Define $f^{\left[1\right]}=f$ and

$$f^{\left[j\right]}=(f\otimes {\mathrm{Id}}_{L^{j-1}})\circ f^{[j-1]}:E\to E\otimes L^j,j=2,\dots ,m.$$

Since $H^0(X,L^m)=\mathbb{C}$, any nonzero $\omega \in H^0(X,L^m)$ gives

$${\iota }_{\omega }:E\otimes L^m\stackrel{\sim }{\to }E,$$

and we set

$${\mathsf{\Phi }}_f={\iota }_{\omega }\circ f^{\left[m\right]}\in H^0(X,\mathcal{E}nd\left(E\right)).$$

(2)

Replacing $\omega$ by $\lambda \omega$ replaces ${\mathsf{\Phi }}_f$ by ${\lambda }^{-1}{\mathsf{\Phi }}_f$; semisimplicity is independent of this choice.

Proposition 6.

The endomorphism ${\mathsf{\Phi }}_f$ introduced in (2) is semisimple.

Proof. 

If E is stable, then ${\mathsf{\Phi }}_f=\theta {\mathrm{Id}}_E$ by Proposition 2, so it is semisimple.

Suppose that E is strictly polystable. Then, by ([12], Proposition 3.1), each component ${\mathsf{\Phi }}_{f,i}\in \mathrm{M}(a_i,\mathbb{C})$ satisfies

$${\mathsf{\Phi }}_{f,i}^{N_i}={\theta }_i{I}_{a_i}$$

for some $N_i\ge 1$ and ${\theta }_i\in {\mathbb{C}}^{\ast }$. Corollary 2 gives each ${\mathsf{\Phi }}_{f,i}$ semisimple, and Corollary 3(i) gives ${\mathsf{\Phi }}_f$ semisimple. □

Let p be a prime divisor of m, $k={\displaystyle \frac{m}{p}}$, $L_0=L^k$, and $\delta :Y\to X$ the p-to-1 cyclic covering associated with $L_0$ ([12], Section 3). The curve Y is a compact connected Riemann surface. Set

$$f_k=f^{\left[k\right]}:E\to E\otimes L_0.$$

Since $L_0^p\cong {\mathcal{O}}_X$ and $L_0\ncong {\mathcal{O}}_X$, the pull-back ${\delta }^{\ast }\left(E\right)$ carries a holomorphic endomorphism

$${\tilde{f}}_k\in H^0(Y,\mathcal{E}nd\left({\delta }^{\ast }\left(E\right)\right))$$

induced by $f_k$ and the Galois action of $\mathbb{Z}/\left(p\right)$, satisfying

$${\tilde{f}}_k^p=\theta {\mathrm{Id}}_{{\delta }^{\ast }\left(E\right)}$$

for some $\theta \in {\mathbb{C}}^{\ast }$ ([12], Proposition 3.1).

Theorem 5.

Let E be a stable rank n and trivial determinant vector bundle with $E\cong E\otimes L$ via f. Then:

(i) 

${\tilde{f}}_k$ is semisimple with p distinct eigenvalues ${\mu }_0,\dots ,{\mu }_{p-1}$, each of multiplicity $n/p$, and the primary decomposition of Proposition 1 gives

$${\delta }^{\ast }\left(E\right)={\displaystyle \underset{j=0}{\overset{p-1}{⨁}}}{\mathcal{F}}_j$$

with $\mathrm{rk}\left({\mathcal{F}}_j\right)={\displaystyle \frac{n}{p}}$.

(ii) 

The Galois group $\mathbb{Z}/\left(p\right)$ cyclically permutes the summands.

(iii) 

There exists a rank ${\displaystyle \frac{n}{p}}$ bundle ${\mathcal{L}}_E$ over Y with $E\cong {\delta }_{\ast }{\mathcal{L}}_E$.

Proof. 

Part (i): Since ${\tilde{f}}_k^p=\theta \mathrm{Id}$ with $\theta \in {\mathbb{C}}^{\ast }$ and Y is compact connected, Theorem 4(iii) gives ${\tilde{f}}_k$ semisimple with eigenvalues among the p-th roots of $\theta$. Since p is prime and $L_0\ncong {\mathcal{O}}_X$, these p roots are distinct ([12], Proposition 3.1). The Galois symmetry forces equal ranks, so each ${\mathcal{F}}_j$ has rank ${\displaystyle \frac{n}{p}}$ (using $p\mid n$ from ([12], Lemma 3.1)). Proposition 1 gives the decomposition.

Part (ii): The deck generator of $\mathbb{Z}/\left(p\right)$ multiplies each eigenvalue ${\mu }_j$ by

$${\zeta }_p=e^{2\pi {\displaystyle \frac{i}{p}}}$$

by ([12], Section 3), hence permuting

$${\mathcal{F}}_j\mapsto {\mathcal{F}}_{j+1\mathrm{mod}p}.$$

Part (iii): By transitivity, ${\delta }_{\ast }{\mathcal{F}}_0\cong E$ ([12], Proposition 3.1); set ${\mathcal{L}}_E={\mathcal{F}}_0$. This satisfies the stated properties. □

Remark 10.

Theorem 5 reinterprets ([12], Theorem 3.1): the decomposition of E as a push-forward from a cyclic covering is the primary decomposition of ${\delta }^{\ast }\left(E\right)$ with respect to the semisimple endomorphism ${\tilde{f}}_k$. The idempotents ${\pi }_j=p_j\left({\tilde{f}}_k\right)$ of Proposition 1(ii) produce the splitting

$${\delta }^{\ast }\left(E\right)=\underset{j}{⨁}{\mathcal{F}}_j.$$

Theorem 6.

Let E be polystable with $E\cong E\otimes L$ via f, and let

$$\mathcal{A}\left({\mathsf{\Phi }}_f\right)\subset H^0(X,\mathcal{E}nd\left(E\right))$$

be the commutative subalgebra generated by ${\mathsf{\Phi }}_f$. Then:

(i) 

Every element of $\mathcal{A}\left({\mathsf{\Phi }}_f\right)$ is semisimple.

(ii) 

$\mathcal{A}\left({\mathsf{\Phi }}_f\right)$ contains no nonzero nilpotent endomorphism.

(iii) 

If $\phi \in H^0(X,\mathcal{E}nd\left(E\right))$ has ${\phi }_n\ne 0$, then $\phi \notin \mathcal{A}\left({\mathsf{\Phi }}_f\right)$.

In particular, the nonzero nilpotent elements of $H^0(X,\mathcal{E}nd\left(E\right))$ (present when $a_i\ge 2$ for some i by Corollary 3(ii)) lie strictly outside $\mathcal{A}\left({\mathsf{\Phi }}_f\right)$.

Proof. 

Part (i): Any $q\left({\mathsf{\Phi }}_f\right)\in \mathcal{A}\left({\mathsf{\Phi }}_f\right)$ acts as $q\left({\mu }_j\right)$ on each eigenspace subbundle ${\mathcal{F}}_{{\mu }_j}$ of the primary decomposition of ${\mathsf{\Phi }}_f$, so its minimal polynomial divides

$$\prod _j(t-q\left({\mu }_j\right)),$$

a product of linear factors. Hence $q\left({\mathsf{\Phi }}_f\right)$ is semisimple.

Part (ii): A semisimple nilpotent endomorphism has minimal polynomial dividing both ${\prod }_j(t-q\left({\mu }_j\right))$ (squarefree) and $t^N$; their gcd divides t, so the endomorphism is zero.

Part (iii): If ${\phi }_n\ne 0$, $\phi$ is not semisimple, so $\phi \notin \mathcal{A}\left({\mathsf{\Phi }}_f\right)$ by part (i). □

Remark 11.

Theorem 6 proves that the subalgebra $\mathcal{A}\left({\mathsf{\Phi }}_f\right)$ is semisimple and commutative; nilpotent endomorphisms of E are precisely those incompatible with the fixed-point structure encoded by ${\mathsf{\Phi }}_f$. This is a new result that further develops the results of [12].

Recall that $\sigma$ acts on

$$Z\left(\mathrm{SL}\right(n,\mathbb{C}\left)\right)\cong \mathbb{Z}/\left(n\right)$$

by inversion [12], so $\sigma \left(L\right)\cong L^{-1}$ and $\sigma \left(L\right)\otimes L\cong {\mathcal{O}}_X$.

Proposition 7.

Let E be polystable rank n and trivial determinant, $L\in H^1(X,\mathbb{Z}/\left(n\right))$, and $f:E\stackrel{\sim }{\to }\sigma \left(E\right)\otimes L$. Since $\sigma \left(f\right):\sigma \left(E\right)\to E\otimes L^{-1}$, the map

$$\sigma \left(f\right)\otimes {\mathrm{Id}}_L:\sigma \left(E\right)\otimes L\to E$$

is well defined. Set

$${\mathsf{\Psi }}_f=(\sigma \left(f\right)\otimes {\mathrm{Id}}_L)\circ f:E\to E.$$

Then:

(i) 

${\mathsf{\Psi }}_f^2={\mathrm{Id}}_E$.

(ii) 

${\mathsf{\Psi }}_f$ is semisimple with eigenvalues in $\{+1,-1\}$.

(iii) 

$E=E_{+}\oplus E_{-}$ where $E_{\pm }=\ker ({\mathsf{\Psi }}_f\mp {\mathrm{Id}}_E)$ are holomorphic subbundles.

(iv) 

E admits a holomorphic nondegenerate orthogonal or symplectic bilinear form ([12], Theorem 4.1).

Proof. 

Notice that

$$\sigma \left(f\right):\sigma \left(E\right)\to {\sigma }^2\left(E\right)\otimes \sigma \left(L\right)=E\otimes L^{-1},$$

so

$$\sigma \left(f\right)\otimes {\mathrm{Id}}_L:\sigma \left(E\right)\otimes L\to E,$$

and ${\mathsf{\Psi }}_f:E\to E$.

Part (i): Since E is polystable, ${\mathsf{\Psi }}_f\in H^0(X,\mathcal{E}nd\left(E\right))$ preserves each isotypic component $F_i^{\oplus a_i}$ (Proposition 3), and Proposition 2 gives

$${\mathsf{\Psi }}_f{|}_{F_i}={\lambda }_i{\mathrm{Id}}_{F_i}$$

for some ${\lambda }_i\in {\mathbb{C}}^{\ast }$. Hence

$${\mathsf{\Psi }}_f^2{|}_{F_i}={\lambda }_i^2{\mathrm{Id}}_{F_i}.$$

By ([12], Theorem 4.1), choosing a representative s of $\sigma$ of order 2 and $g_e\in \mathrm{SL}(n,\mathbb{C})$ with $f\left(e\right)=e·g_e\otimes l$, the relation $s\left(g_e\right)={\lambda }_i{g}_e^{-1}$ holds, and

$$g_e=s^2\left(g_e\right)=s\left({\lambda }_i{g}_e^{-1}\right)={\lambda }_i^{-1}{\left({\lambda }_i{g}_e^{-1}\right)}^{-1}={\lambda }_i^{-2}{g}_e,$$

so ${\lambda }_i^2=1$. Hence ${\mathsf{\Psi }}_f^2={\mathrm{Id}}_E$.

Part (ii): By Theorem 4(iii) with $m=2$, $\theta =1$: ${\mathsf{\Psi }}_f$ is semisimple with eigenvalues in $\{\pm 1\}$.

Part (iii) is Proposition 1 applied to ${\mathsf{\Psi }}_f$ and part (iv) is proved in ([12], Theorem 4.1). □

Remark 12.

Parts (ii)–(iii) of Proposition 7 follow entirely from Theorem 4(iii) and Proposition 1: the condition ${\mathsf{\Psi }}_f^2={\mathrm{Id}}_E$ and our spectral theory suffice. Part (iv) requires the additional representation-theoretic structure of $\mathrm{SL}(n,\mathbb{C})$ considered in [12].

Definition 2

([12], Definition 5.2). Let ${\sigma }_X:X\to X$ be an involution. A polystable rank n and trivial determinant vector bundle E admits a Galois structure if there exists a holomorphic isomorphism $f:E\stackrel{\sim }{\to }{\sigma }_X^{\ast }\left(\sigma \left(E\right)\right)$ with

$$\left({\sigma }_X^{\ast }f\right)\circ f={\mathrm{Id}}_E.$$

Theorem 7.

Let ${\sigma }_X$ be an order-2 involution of X, σ the outer automorphism of $\mathrm{SL}(n,\mathbb{C})$ of order 2, and E polystable rank n and trivial determinant admitting a Galois structure via f. Suppose $\omega \in \mathrm{Aut}\left(E\right)$ satisfies $\omega \ne c{\mathrm{Id}}_E$ for all $c\in {\mathbb{C}}^{\ast }$ and

$${\sigma }_X^{\ast }\left(\sigma \left(\omega \right)\right)\circ f=f\circ \omega .$$

(3)

Then:

(i) 

ω is semisimple.

(ii) 

The primary decomposition

$$E={\mathcal{F}}_{{\mu }_1}\oplus \cdots \oplus {\mathcal{F}}_{{\mu }_r}$$

satisfies

$${\sigma }_X^{\ast }{\mathcal{F}}_{{\mu }_j}\cong {\mathcal{F}}_{{\mu }_j}$$

for all j.

(iii) 

$r\ge 2$, so E is strictly polystable.

Proof. 

Part (i): By ([12], Theorem 5.1), the element $g_e\in \mathrm{SL}(n,\mathbb{C})$ defined by $\omega \left(e\right)=e·g_e$ is semisimple. By Lemma 1(i), the characteristic polynomial of $\omega$ has constant coefficients; since $g_e$ is semisimple at every fiber, the minimal polynomial ${\mu }_{\omega }\in \mathbb{C}\left[t\right]$ has no repeated roots. Hence $\omega$ is semisimple in $H^0(X,\mathcal{E}nd\left(E\right))$.

Part (ii): From (3), for $v\in {\left({\mathcal{F}}_{{\mu }_j}\right)}_x$:

$${\sigma }_X^{\ast }{\left(\sigma \left(\omega \right)\right)}_x\left(f_x\left(v\right)\right)=f_x\left({\mu }_{j}v\right)={\mu }_j{f}_x\left(v\right),$$

so f maps ${\mathcal{F}}_{{\mu }_j}$ into the ${\mu }_j$-eigenspace of ${\sigma }_X^{\ast }\left(\sigma \left(\omega \right)\right)$. By ([12], Theorem 5.1),

$$\sigma \left({\mathcal{F}}_{{\mu }_j}\right)\cong {\mathcal{F}}_{{\mu }_j},$$

so

$${\sigma }_X^{\ast }\left(\sigma \left({\mathcal{F}}_{{\mu }_j}\right)\right)\cong {\sigma }_X^{\ast }\left({\mathcal{F}}_{{\mu }_j}\right).$$

Since

$$f:{\mathcal{F}}_{{\mu }_j}\stackrel{\sim }{\to }{\sigma }_X^{\ast }\left({\mathcal{F}}_{{\mu }_j}\right)$$

is an isomorphism, it follows that

$${\sigma }_X^{\ast }{\mathcal{F}}_{{\mu }_j}\cong {\mathcal{F}}_{{\mu }_j}.$$

Part (iii): A stable bundle has

$$H^0(X,\mathcal{E}nd\left(E\right))=\mathbb{C}·{\mathrm{Id}}_E$$

by Proposition 2, so $\omega \ne c{\mathrm{Id}}_E$ forces E to be strictly polystable. The semisimple $\omega$ has $\ge 2$ distinct eigenvalues (otherwise $\omega =\mu {\mathrm{Id}}_E$, which is a contradiction), giving $r\ge 2$. □

Remark 13.

Theorem 7 provides the linear-algebraic framework for ([12], Theorem 5.1). The novel contribution is the identification of the fiberwise semisimplicity of $g_e$ (from [12]) with the global semisimplicity of ω as a section of $\mathcal{E}nd\left(E\right)$, via the Principle of Spectral Constancy.

5. Conclusions

This paper has developed new results describing the Jordan normal forms for holomorphic endomorphisms of vector bundles over compact Riemann surfaces, and has applied it to the geometry of moduli spaces of vector bundles. The central analytical input is the Principle of Spectral Constancy: the compactness and connectedness of the base curve force the characteristic polynomial of any holomorphic endomorphism to have constant coefficients, which in turn makes the entire Jordan theory global. The main contributions are the following.

The primary decomposition theorem establishes that the generalized eigenspace subsheaves of any holomorphic endomorphism are globally defined holomorphic subbundles, with the direct sum decomposition produced by idempotent endomorphisms constructed via the Chinese Remainder Theorem in $\mathbb{C}\left[t\right]$ with constant coefficients. The Jordan decomposition theorem proves that every holomorphic endomorphism of a vector bundle over a compact connected Riemann surface admits a unique global additive Jordan decomposition into semisimple and nilpotent parts, both of which are polynomial expressions in the original endomorphism with constant coefficients. This is the bundle-theoretic analogue of the Jordan–Chevalley decomposition in Lie theory. The nilpotent part defines a Jordan filtration by holomorphic subbundles, with injective bundle maps between successive quotients, providing the bundle-theoretic analogue of the flag of generalized eigenspaces.

The relationship between the Jordan type and the stability properties of the bundle is completely described. Endomorphisms of stable bundles are necessarily scalar multiples of the identity, so stable bundles admit no nontrivial Jordan structure. For polystable bundles, the Jordan normal form is determined componentwise by the classical Jordan normal forms of matrices in the direct sum of matrix algebras isomorphic to the endomorphism algebra, with the stable summands playing the role of the scalar field. For semistable bundles, any nilpotent endomorphism induces a Jordan filtration in which all steps have the same slope as the bundle. A central new result shows that any finite-order endomorphism is necessarily semisimple: if ${\phi }^m=\theta {\mathrm{Id}}_E$ for some $\theta \in {\mathbb{C}}^{\ast }$, then the minimal polynomial of $\phi$ divides the separable polynomial $t^m-\theta$, forcing $\phi$ to be semisimple.

In the context of moduli spaces, this last result is applied systematically as follows. The endomorphisms arising from fixed-point conditions—whether from tensorization by a line bundle of finite order, from the outer involution of the structure group, or from automorphisms of the base curve—all satisfy finite-order conditions, and are therefore automatically semisimple. The primary decomposition then produces, in each case, the geometric decompositions described by Antón-Sancho for fixed points of moduli space automorphisms. The decomposition of a vector bundle fixed by tensorization with a finite-order line bundle is identified as the primary decomposition of the pull-back bundle with respect to a semisimple endomorphism on a cyclic covering. The decomposition of a bundle fixed by the outer involution is the primary decomposition of the endomorphism ${\mathsf{\Psi }}_f$ satisfying ${\mathsf{\Psi }}_f^2={\mathrm{Id}}_E$, producing the orthogonal-symplectic splitting. The strict polystability of a Galois bundle admitting a nontrivial automorphism commuting with the Galois structure follows from the semisimplicity of that automorphism and the fact that a semisimple non-scalar endomorphism has at least two distinct eigenvalues. As a new result of the paper, not previously explicit in the literature, we show that the commutative subalgebra generated by any fixed-point endomorphism consists entirely of semisimple elements, so that nilpotent endomorphisms of the bundle are precisely those incompatible with the fixed-point structure.

Several natural directions for further research arise from this work. The extension to Higgs bundles is an open problem: for a Higgs bundle $(E,\phi )$ where

$$\phi :E\to E\otimes K$$

is the Higgs field and K is the canonical bundle of X, the characteristic polynomial of $\phi$ defines the spectral curve in the total space of K, and the associated eigenspace decomposition is the Hitchin fibration. A Jordan-type theory for Higgs fields—relating the nilpotent part of $\phi$ to the geometry of the spectral curve and the structure of the Hitchin fibers—would be a natural continuation of the present work. In the stable case, the Higgs field is not necessarily scalar (in contrast to endomorphisms), and the relationship between the Jordan type of the Higgs field and the stability of the Higgs bundle is more subtle.

The extension to principal bundles with structure group a general semisimple complex Lie group G is another natural direction. The results of this paper rely on the identification of rank n vector bundles with principal $\mathrm{SL}(n,\mathbb{C})$-bundles; for other classical groups such as $\mathrm{Sp}(2n,\mathbb{C})$ and $\mathrm{SO}(n,\mathbb{C})$, the endomorphism algebra must be replaced by the appropriate subalgebra of invariant endomorphisms, and the Jordan theory would need to be developed in that constrained setting.

The relationship between the Jordan filtration of a nilpotent endomorphism and the Harder–Narasimhan filtration of a vector bundle that is not semistable is an open problem. For semistable bundles, the present paper shows that the Jordan filtration has all steps of equal slope; for non-semistable bundles, the interaction between the two filtrations—both defining flags of subbundles, but governed by different algebraic constraints—is not yet understood.

Finally, the computational aspects of the theory developed here offer opportunities for applications in numerical linear algebra and mathematical physics. The explicit construction of the idempotent projections ${\pi }_j=p_j\left(\phi \right)$ via the Chinese Remainder Theorem provides an algorithmic procedure for computing the primary decomposition of a vector bundle from knowledge of a global endomorphism. The numerical stability and conditioning of this procedure in families of bundles parametrized by a moduli space is an open problem with potential applications to the computational study of gauge theories and integrable systems.