Vector Bundles on Rational Homogeneous Spaces

Rong Du ∗, Xinyi Fang †and Yun Gao ‡

Dedicate to people who devote their lives to fight against coronaviruses

arXiv:2007.06816v1 [math.AG] 14 Jul 2020

Abstract
We consider a uniform r-bundle E on a complex rational homogeneous space X and show
that if E is poly-uniform with respect to all the special families of lines and the rank r is
less than or equal to some number that depends only on X, then E is either a direct sum of
line bundles or δi -unstable for some δi . So we partially answer a problem posted by Muñoz-
Occhetta-Solá Conde( [19] ). In particular, if X is a generalized Grassmannian G and the rank
r is less than or equal to some number that depends only on X, then E splits as a direct sum
of line bundles. We improve the main theorem of Muñoz-Occhetta-Solá Conde ( [18] Theorem
3.1) when X is a generalized Grassmannian by considering the Chow ring. Moreover, by
calculating the relative tangent bundles between two rational homogeneous spaces, we give
explicit bounds for the generalized Grauert-Mülich-Barth theorem on rational homogeneous
spaces.

Key words: vector bundle, generalized Grassmannian, rational homogeneous space

1 Introduction
Algebraic vector bundles on a projective variety X over complex number field C are fundamental
research objects in algebraic geometry. However, up to now, algebraic vector bundles are still
mysterious for general projective varieties. According to Serre (GAGA), the classification of
algebraic vector bundles over C is equivalent to the classification of holomorphic vector bundles.
So there are not only algebraic ways but also analytic ways to handle the problems of vector
bundles. For simplicity, we just call vector bundles of rank r or r-bundles in the context.
If X is P1 , the structure of a vector bundle on X is quite clear because Grothendieck tells
us that it splits as a direct sum of line bundles. However, if X is a projective space and the
dimension of it is bigger than or equal to two then the structures of vector bundles on X are
not so easy to be determined. Since projective spaces are covered by lines, it is a natural way to
consider the restriction of vector bundles to lines. From Grothendieck’s result, they split after
∗ School of Mathematical Sciences Shanghai Key Laboratory of PMMP, East China Normal University, Rm.

312, Math. Bldg, No. 500, Dongchuan Road, Shanghai, 200241, P. R. China, [email protected]. The
Research is Sponsored by National Natural Science Foundation of China (Grant No. 11531007), Natural Science
Foundation of China and the Israel Science Foundation (Grant No. 11761141005) and Science and Technology
Commission of Shanghai Municipality (Grant No. 18dz2271000).
† School of Mathematical Sciences Shanghai Key Laboratory of PMMP, East China Normal University, No.

500, Dongchuan Road, Shanghai, 200241, P. R. China, [email protected]. The Research is Sponsored by
National Natural Science Foundation of China (Grant No. 11531007) and Science and Technology Commission
of Shanghai Municipality (Grant No. 18dz2271000).
‡ School of Mathematical Sciences, Shanghai Jiao Tong University, Shanghai 200240, P. R. of China, gaoyun-

[email protected]. The Research is Sponsored by National Natural Science Foundation of China (Grant No.
11531007).

1
being restricted to P1 . By semicontinuity theorem, for “almost all” lines, r-bundle E has the
constant splitting type. This means that the lines to which the vector bundle restricts having
different splitting type consist a closed subset of a Grassmannian. If the closed subset is empty,
such bundles are called uniform vector bundles. Uniform bundles are widely studied not only
on projective spaces ( [25] [27] [24] [9] [10] [11] [3] ) but also on special Fano manifolds of Picard number
one ( [2] [15] [12] [6] [18] ). Please see Introduction in [6] for the details.
Instead of considering vector bundles, Occhetta-Solá Conde-Wiśniewski ( [22] ) studied flag
bundles which are constructed upon the action of the defining group G on the flag manifold
G/B. Recently, Muñoz-Occhetta-Solá Conde ( [17] ) studied uniform principle G-bundles with
G semisimple over Fano manifolds. They present a number of theorems that are flag bundle’s
versions of some of the central results in the theory of uniform vector bundles. More precisely,
they paid special attention to homogeneous filtrations of relative tangent bundles between flag
manifolds and generalized the different standard decomposability notions of vector bundles. They
used an interesting concept of "tag" of a G/B-bundle to describe diagonalizability of any uniform
flag bundle of low rank. But they mainly focus on X with Picard number one. In [19] , the authors
proposed a problem as follows.
Problem 1.1. Classify low rank uniform principle G-bundles (G semisimple algebraic group)
on rational homogeneous spaces.
In this paper, we consider uniform bundles on generalized flag varieties or even rational
homogeneous spaces with arbitrary Picard numbers and give partial answers to this problem.
Let
X = G/P ≃ G1 /PI1 × G2 /PI2 × · · · × Gm /PIm ,
where Gi is a simple Lie group with Dynkin diagram Di whose set of nodes is Di and PIi is a
parabolic subgroup of Gi corresponding to Ii ⊂ Di . We set F (Ii ) := Gi /PIi by marking on the
Dynkin diagram Di of Gi the nodes corresponding to Ii . Let δi be a node in Di and N (δi ) be
the set of nodes in Di that are connected to δi .
If δi ∈ Ii , we call
δc δc
Mi i := Gi /Pi i × G\i /PIi (1 ≤ i ≤ m),
c
the i-th special family of lines of class δ̌i , where Pi i := PIi \δi ∪N (δi ) and G\
δ
i /PIi is G1 /PI1 ×
G2 /PI2 × · · · × Gm /PIm by deleting i-th term Gi /PIi . Denote by
δc
Ui i := Gi /PIi ∪N (δi ) × G\
i /PIi

δc
the i-th universal family of class δ̌i , which has a natural P1 -bundle structure over Mi i .
We separate our discussion into two cases:
δc δc
Case I: N (δi ) ⊆ Ii , then Ui i = X and we have the natural projection X → Mi i ;
Case II: N (δi ) * Ii , then we have the standard diagram

δc q2
/ Mδi
c
Ui i i (1.1)
q1

X.

For the definition of "poly-uniform", "δi -stable", "δi -unstable", "ς(G)" and "ν(X)", please
see Section 3.

2
Theorem 1.2. On X, if an r-bundle E is poly-uniform with respect to all the special families
of lines and r ≤ ν(X) − 2, then E is δi -unstable for some δi (1 ≤ i ≤ m) or E splits as a direct
sum of line bundles.
In particular, if the Picard number of X is one, we improve an interesting theorem of Muñoz-
Occhetta-Solá Conde ( [18] Theorem 3.1) a little bit (see Table 2).
Theorem 1.3. Suppose that E is a uniform r-bundle on a generalized Grassmann G. If r ≤ ς(G),
then E splits as a direct sum of line bundles.
By calculating the relative tangent bundles and using Descent Lemma, we can have explicit
bounds for the generalized Grauert-Mülich-Barth theorem on rational homogeneous spaces.
Theorem 1.4. Fix δi ∈ Ii and assume that αδi is not an exposed short root. Let E be a
(δ ) (δ ) (δ ) (δ ) (δ )
holomorphic r-bundle over X of type aE i = (a1 i , . . . , ar i ), a1 i ≥ · · · ≥ ar i with respect to
c
δ
Mi i . If for some t < r,

(δ ) (δi ) 1, and N (δi ) fits Case I
at i − at+1 ≥
2, and N (δi ) fits Case II,
then there is a normal subsheaf K ⊂ E of rank t with the following properties: over the open set
(δ ) (δ ) c
VE = q1 (q2−1 (UE i )) ⊂ X, where UE i is an open set in Mδi , the sheaf K is a subbundle of E,
(δi )
which on the line L ⊂ X given by l ∈ UE has the form
K|L ∼
= ⊕ts=1 OL (as(δi ) ).
Corollary 1.5. With the same assumption as Theorem 1.4. For a δi -semistable r-bundle E over
(δ ) (δ ) (δ ) (δ ) (δ ) δc
X of type aE i = (a1 i , . . . , ar i ), a1 i ≥ · · · ≥ ar i with respect to Mi i , we have
(δ )
as(δi ) − as+1
i
≤ 1 for all s = 1, . . . , r − 1.
(δ )
In particular, if N (δi ) fits Case I, then we have as i ’s are constant for all 1 ≤ s ≤ r.
Theorem 1.6. Fix δi ∈ Ii and assume that αδi is an exposed short root. Let E be a holomorphic
(δ ) (δ ) (δ ) (δ ) (δ ) δc
r-bundle over X of type aE i = (a1 i , . . . , ar i ), a1 i ≥ · · · ≥ ar i with respect to Mi i . If for
some t < r, 
(δi ) (δi ) 1, and N (δi ) fits Case I
at − at+1 ≥
4, and N (δi ) fits Case II,
then there is a normal subsheaf K ⊂ E of rank t with the following properties: over the open set
(δ ) (δ ) c
VE = q1 (q2−1 (UE i )) ⊂ X, where UE i is an open set in Mδi , the sheaf K is a subbundle of E,
(δ )
which on the line L ⊂ X given by l ∈ UE i has the form
K|L ∼
= ⊕ts=1 OL (as(δi ) ).
Corollary 1.7. With the same assumption as Theorem 1.6. For a δi -semistable r-bundle E over
(δ ) (δ ) (δ ) (δ ) (δ ) δc
X of type aE i = (a1 i , . . . , ar i ), a1 i ≥ · · · ≥ ar i with respect to Mi i , we have
(δ )
as(δi ) − as+1
i
≤ 3 for all s = 1, . . . , r − 1.
(δ )
In particular, if N (δi ) fits Case I, then we have as i ’s are constant for all 1 ≤ s ≤ r.
If X = G/B, where G is a semi-simple Lie group and B is a Borel subgroup of G, then we
have the following result.
Corollary 1.8. If an r-bundle E on X is δi -semistable for all i and δi (1 ≤ i ≤ m), then E
splits as a direct sum of line bundles.

3
2 Preliminaries
Throughout this paper, all algebraic varieties and morphisms will be defined over complex number
field C.

2.1 Semisimple Lie groups and algebras

In mathematics, Lie group–Lie algebra correspondence allows one to study Lie groups, which are
geometric objects, in terms of Lie algebras, which are linear objects. Let G be a semi-simple Lie
group. Assume V to be a nonzero finite dimensional complex vector space, fix a maximal torus
H ⊂ G, and let ϕ : G → GL(V ) be a representation of G. It is well known that V decomposes
into a direct sum of simultaneous eigenspaces
M
V = Vλ ,

where the direct sum run over λ in the character group of H, which is the set of all holomorphic
homomorphisms λ from H to C∗ , and

Vλ = {v ∈ V |ϕ(h) = λ(h)v, for all h ∈ H}.

Since V is a finite dimensional vector space, we have Vλ = 0 for all but finitely many values of
λ. Those values of λ for which Vλ 6= 0 are called the weights of V , and Vλ is called the weight
space.
In the Lie algebra side, let g be the associated semi-simple Lie algebra of G. The maximal
torus corresponds to h ⊂ g which is an abelian subalgebra of maximal dimension, i.e. Cartan
subalgebra. A holomorphic representation ϕ : G → GL(V ) of a complex Lie group G gives rise
to a complex linear representation ϕg : g → gl(V ) of the Lie algebra g of G. Similarly, every
finite dimensional representation of g admits a decomposition
M
V = Vλ ,
λ∈{λ∈h∨ |Vλ 6=0}

where
Vλ = {v ∈ V |[h, v] = λ(h)v, for all h ∈ h}.
Those λ are still called the weights of V , and Vλ is called the weight space. If we apply the above
decomposition to V = g and ϕg the adjoint representation, we have Cartan decomposition of g:
M
g=h⊕ gα ,
α∈h∨ \{0}

where
gα := {g ∈ g|[h, g] = α(h)g, for all h ∈ h}.
The elements α ∈ h∨ \{0} for which gα 6= 0 are called roots of g, and the set of these elements
will be denoted by Φ and be called root system. For every α ∈ Φ, gα is called root space which
is one dimensional.
Fix a linear functional
f : spanR Φ → R
whose kernel does not intersect Φ. Let

Φ+ := {α ∈ Φ|f (α) > 0} and Φ− := {α ∈ Φ|f (α) < 0}.

4
Then Φ+ is called positive system of roots and Φ− is called negative system of roots. Given such
a positive system Φ+ , we define the fundamental system Π ⊂ Φ+ as follows: α ∈ Π if and only if
α ∈ Φ+ and α cannot be expressed as the sum of two elements of Φ+ . A non-zero representation
V of g is called a highest
L weight representation if it is generated by a vector v ∈ Vλ such that
gv = 0 for all g ∈ α∈Φ gα . In this case, v is called the highest weight vector, and λ is the
+

highest weight of V .
For each α ∈ Φ there is a unique element hα ∈ h such that

α(h) =< hα , h > for all h ∈ h.

Pl
The vectors hα for α ∈ Φ span h. We denote by hR the set of all elements of form i=1 ai hα for
ai ∈ R.
The Killing form < α, β >:= tr(adα ◦ adβ ) defines a nondegenerated bilinear form on h,
where α, β ∈ g and ad is the adjoint representation. It can be shown that h∨ R is a Euclidean
space with respect to <, >. Set n := dimC (h) and D := {1, 2, . . . , n}. We identify D with the set
of fundamental roots with respect to a choice of maximal torus T and fixed Borel subgroup B.
It is known that every fundamental system Π ⊂ Φ can form a basis of h∨ R . Let Π = {α1 , . . . , αn }
be a fundamental system. Then we define Aij by
< αi , αj >
Aij = 2 ∈ Z, i, j = 1, . . . n.
< αi , αi >

The matrix A = (Aij ) is called the Cartan matrix of g.

The Dynkin diagram of G, which we denoted by D := D(G), is determined by the Cartan
matrix. It consists of a graph whose set of nodes is D and where the nodes i and j are joined by
Aij Aji edges. When two nodes i and j are joined by a double or triple edge, we add to it an arrow
pointing to i if |Aij | > |Aji |. We call αi a short root of D and αj a non-short (or long) root of D.
(Sometimes, for the sake of narrative convenience, we freely interchange the terminology "node"
and "root".) One may prove that there is a one to one correspondence between isomorphism
classes of semisimple Lie algebras and Dynkin diagrams of reduced root systems. Moreover, every
reduced root system is a disjoint union of mutually orthogonal irreducible root subsystems, each
of them corresponding to one of the connected finite Dynkin diagrams An , Bn , Cn , Dn (n ∈ Z>0 ),
E6 , E7 , E8 , F4 , G2 :

1 2 n−2 n−1 n 1 2 n−2 n−1 n

An : Bn :

1 2 n−3 n−2 1 2 n−2 n−1 n

n−1
Dn : Cn :
n

E6 : F4 :
1 3 4 5 6 1 2 3 4

2
1 2
E7 : G2 :
1 3 4 5 6 7

5
 2

E8 :
1 3 4 5 6 7 8

The connected components of the Dynkin diagram D determine the simple Lie groups that
are factors of the semisimple Lie group G, each of them corresponding to one of the Dynkin
diagrams above.

2.2 Parabolic Subgroup and Subalgebra

A closed subgroup P of G is called parabolic if the quotient space G/P is complete, hence
projective. A maximal connected solvable subgroup B of G is called a Borel subgroup. We fix
a Cartan subalgebra h. Let M
b=h⊕ gα
α∈Φ−

be a fixed Borel subalgebra. It is easy to determine the parabolic subalgebras containing b. They
are all of the form M
p=b⊕ gα ,
α∈Φ+
P

where Φ+P is a subset of Φ

+
that is closed under the addition of roots. Hence the parabolic
subalgebras containing b lie in bijection with the subsets of D. For some subset I of D, we write
pI the parabolic subalgebra corresponding to I. We define PI by the parabolic subgroup of G
such that its Lie algebra is pI . Therefore the parabolic subgroup PI corresponds to the subset
I (so a maximal parabolic subgroup is defined by a single root). Then G/PI has a minimal
homogeneous embedding inP projective space of the highest weigh module Vλ of G corresponding
to the highest weight λ = i∈I ωi , where ωi is the i-th fundamental weight dual to the roots
αi ∈ Π, by a very ample line bundle Lλ .

2.3 Rational homogeneous spaces

It is well known that G/P carries a transitive G-action, it is a smooth projective variety. Borel
and Remmert’s classical theorem ( [4] ) states that a projective complex manifold which admits
a transitive action of its automorphism group is a direct product of an abelian variety and a
rational homogeneous space G/P , where G is a semi-simple algebraic group and P is a parabolic
subgroup.
Every rational homogeneous space G/P can be decomposed into a product
G/P ≃ G1 /PI1 × G2 /PI2 × · · · × Gm /PIm
of rational homogeneous spaces with simple algebraic group G1 , · · · , Gm . Each rational homo-
geneous space Gi /PIi , called the generalized flag manifold, only depends on the Lie algebra gi of
Gi , which is classically determined by the marked Dynkin diagram ( [16] ). In the most common
notation, we set FI := G/PI by marking on the Dynkin diagram D of G the nodes corresponding
to I. For instance, numbering the nodes of An , the usual flag manifold F (d1 , . . . , ds ; n + 1)
corresponds to the marking of I = {d1 , . . . , ds } (sometimes we omit the braces and just write as
Pd1 ,...,ds ).

6
 The two extremal cases correspond to the generalized complete flag manifolds (all nodes
marked), and the generalized Grassmannian (only one node marked).
In [16] , the authors explicitly describe the lines through a point of a rational homogeneous
space G/PI , where G is a simple Lie group. Let j ∈ I and N (j) be the set of nodes in D that
are connected to j.
Definition 2.1. We call αj (j ∈ I) an exposed short root if the connected component of j in
D\(I\j) contains root longer than αj , i.e., if an arrow in D\(I\j) points towards j.
Remark 2.2. Obviously, long roots of D in I are not exposed short roots. If I is a set of single
point, i.e. X = G/PI is the generalized Grassmannian, then the exposed short root is just the
usual short root. It’s worth mentioning that if I contains all long roots of D, then short roots of
D in I are not exposed short roots.
Theorem 2.3. ( [16] Theorem 4.3) Let I ⊆ D = {1, . . . , n}. Suppose G to be a simple Lie group.
Consider X = G/PI in its minimal homogeneous embedding. Denote by F1 (X) the space of P1, s
in X. Then
`
1. F1 (X) = j∈I F1j (X), where F1j (X) is the space of lines of class α̌j ∈ H2 (X, Z).

2. If αj is not an exposed short root, then F1j (X) = G/PI\j∪N (j) .

3. If αj is an exposed short root, then F1j (X) is the union of two G-orbits, an open orbit and
its boundary G/PI\j∪N (j) .
Remark 2.4. If I = {j} and αj is a long root, then F1 (X) is just the variety of lines on X.
Example 2.5. Let’s consider the Dynkin diagram An , i.e. X = SLn+1 /PI is the generalized
flag manifold.
1) For I = {k}, X is the usual Grassmannian and F1 (X) is just the variety of lines on X.

1 2 k−1 k k+1 n−2 n−1 n

X:

1 2 k−1 k k+1 n−2 n−1 n

F1 (X) :

2) For I = {d1 , d2 }, X is the usual flag manifold F (d1 , d2 ; n + 1) and F1 (X) is the disjoint union
of F1d1 (X) and F1d2 (X).

1 2 d1 −1 d1 d1 +1 d2 −1 d2 d2 +1 n−2 n−1 n
X:

1 2 d1 −1 d1 d1 +1 d2 −1 d2 d2 +1 n−2 n−1 n
F1d1 (X) :

1 2 d1 −1 d1 d1 +1 d2 −1 d2 d2 +1 n−2 n−1 n
X:

1 2 d1 −1 d1 d1 +1 d2 −1 d2 d2 +1 n−2 n−1 n
F1d2 (X) :

7
 There is a similar statement for Cx ⊆ PTx X, the set of tangent directions to lines on X
passing through a fixed point x. It is a disjoint union of spaces of lines of class α̌ through x.
Theorem 2.6. ( [16] Theorem 4.8) Let I ⊆ D = {1, . . . , n} and j ∈ I. Suppose G to be a
simple Lie group. Consider X = G/PI in its minimal homogeneous embedding. Let H be the
semisimple part of PI and D(H) be the components of (D\I)\j containing an element of N (j),
where D\I means D\I plus any nodes of I attached to a node of D\I. Denote by Cxj the space
of lines of class α̌j through x. Then
1. If αj is not an exposed short root, then Cxj = H/PN (j) .
2. If αj is an exposed short root, then Cxj is a union of an open PI -orbit and its boundary
H/PN (j) .
Remark 2.7. If I = {j}, then the set of nodes of the Dynkin diagram H is D(H) = D\j. PN (j)
is a parabolic subgroup of H by marking in D(H) the nodes in D that are connected to j and
Cxj = H/PN (j) . Moreover, if αj is a long root then Cxj is just the variety of lines through fixed
points, i.e. so-called VMRTs. We refer to [14] for a complete account on VMRTs.
Example 2.8. Let’s consider the Dynkin diagram An , i.e. X = SLn+1 /PI is the generalized
flag manifold.
1) For I = {k}, X is the usual Grassmannian and Cx = Pk−1 × Pn−k is just the variety of
lines through x.

1 2 k−1 k k+1 n−2 n−1 n

X:

1 2 k−1 k+1 n−2 n−1 n

Cx : ×

2) For I = {d1 , d2 }, X is the usual flag manifold F (d1 , d2 ; n + 1) and Cx is the disjoint union of
Cxd1 and Cxd2 .
Cxd1 = Pd1 −1 × Pd2 −d1 −1 , Cxd2 = Pd2 −d1 −1 × Pn−d2 .

1 2 d1 −1 d1 d1 +1 d2 −1 d2 d2 +1 n−2 n−1 n
X:

1 2 d1 −1 d1 +1 d2 −2 d2 −1
Cxd1 : ×

1 2 d1 −1 d1 d1 +1 d2 −1 d2 d2 +1 n−2 n−1 n
X:

d1 +1 d1 +2 d2 −1 d2 +1 n−1 n
Cxd2 : ×

Not only P1 but also all linear spaces can be read from the marked Dynkin diagrams.
Theorem 2.9. ( [16] Theorem 4.9, 4.14) Let G be a simple group and X = G/PS is a rational
homogeneous space. Let Fkα (X) denote the variety parameterizing the α-class Pk ’s on X.

8
 • If α ∈ S is not an exposed short root, then for all k, Fkα (X) is the disjoint union of
homogeneous spaces G/PP βj , where {βj } is a set of positive roots such that the component
of D\{βj } containing α is isomorphic to D(Ak ), intersects S only in α, and α is an extremal
node of this component.
• If α ∈ S is an exposed short root, then for all k, Fkα (X) consists of a finite number of
G-orbits.

3 Uniform vector bundles

Given a smooth projective variety X and a vector bundle E on X, we denote M to be an unsplit
family of rational curves on X. M is called unsplit if M is a proper C-scheme. We say that E is
uniform with respect to M if the restriction of E to the normalization of every curve in M splits
as a direct sum of line bundles with the same splitting type. If X is a generalized Grassmannian
G/Pk , then we just call E uniform without mention the unspilt family M.

3.1 Uniform vector bundles on generalized Grassmannians

Along this section we will work on uniform vector bundles on rational homogeneous spaces of
Picard number one, i.e. generalized Grassmannians. Let G be a simple Lie group and D =
{1, 2, . . . , n} be the set of nodes of the Dynkin diagram D of G. Denote by Pk the parabolic
subgroup of G corresponds to the node k. Consider the generalized Grassmannian G = G/Pk or,
for brevity, D/Pk . Denote by M := G/PN (k) the generalized flag manifold defined by the marked
Dynkin diagram (D, N (k)) and by U := G/Pk,N (k) = G/(PN (k) ∩ Pk ) the universal family, which
has a natural P1 -bundle structure over M, i.e. we have the natural diagram
q
U /M (3.1)
p

G = G/Pk .

Remarkably, M defined above is indeed an unsplit family of rational curves on G. Given x ∈ G,

Mx = q(p−1 (x)), which we call the special family of lines of class αˇk through x, coincides with
H/PN (k) by Remark 2.7, where the set of nodes of the Dynkin diagram H is D(H) = D\k.
When k is an extremal node, that is, the subdiagram D(H) is connected. Remarkably, in
the case D = Dn , i.e. G = SO(2n), since G/Pn−1 ∼ = G/Pn , we only need to think about the
extremal node n. Similarly, since E6 /P1 ∼
= E6 /P6 , we just consider the the extremal node 1 in
E6 . According to Theorem 2.6, Mx has the following possibilities:
• Projective spaces or smooth quadrics,
• Grassmannians,
• Spinor varities,
• E6 /P6 , E7 /P7 , C3 /P3 .
The possibilities are list in Table 1 below.
We observe that for x ∈ G, the morphism from Mx to Grassmannian plays a critical role in
determining whether a uniform vector bundle can split as a direct sum of line bundles. Let ς
be a positive integer smaller than or equal to dim Mx . As long as we show that the morphism

9
 Table 1: Mx corresponding to an extremal node
Mx D
An Bn Cn Dn En (n = 6, 7, 8) Fn (n = 4) Gn (n = 2)
node
1 Pn−1 Q2n−3 P2n−3 Q2n−4 Sn−2 C3 /P3 P1
n Pn−1 Pn−1 Pn−1 G(2, n) En−1 /Pn−1 (n 6= 6) S3 P1
2 G(3, n)

Mx → G(t, ς) can only be constant for any integer 1 ≤ t ≤ [ 2ς ] and every x ∈ G, then every
uniform r-bundle on G splits for r ≤ ς. We suggest that interested readers refer to Theorem 3.1
in paper [18] for details. Now, let’s analyze the morphism Mx → G(t, ς) one by one according to
the probabilities of Mx .

Case I. When Mx is a projective space PN or a smooth quadric QN (N = 2m + 1), then

their Chow rings have the form
Z[H]/(HN +1 ),
where H is a hyperplane section. In particular, dimH 2t (Mx , C) = 1 for every t ≤ [ N2 ]. By the
proof of Lemma 3.4 in paper [18] , the only morphisms Mx → G(t, N ) are constant for any integer
1 ≤ t ≤ [ N2 ].
When Mx is a smooth quadric QN (N = 2m), since

A(Q2m ) = Z[H, U]/(H2m+1 , 2HU − Hm+1 , Hm U − U 2 ),

where H is a hyperplane section and U is a subvariety of codimension m, then we get that

Mx → G(t, N − 1) can only be constant map similarly.

Case II. Mx is Grassmannian G(d, n) (2 ≤ d ≤ n − d). We claim that the only morphisms
G(d, n) → G(t, n − d + 1) are constant for any integer 1 ≤ t ≤ [ n−d+1
2 ].
Lemma 3.1. There are no nonconstant maps from G(d, n) (2 ≤ d ≤ n − d) to G(t, n − d + 1)
for any integer 1 ≤ t ≤ [ n−d+1
2 ].
Proof. Assume that we have a nonconstant morphism φ := G(d, n) → G(t, n− d+ 1). Then there
exists a maximal linear subspace Pn−d such that φ restricts to it is also nonconstant. Denote by
ψ to be the restriction map. Let’s consider ψ ∗ Ht , ψ ∗ Qn−d+1−t and φ∗ Ht , φ∗ Qn−d+1−t (the pull
back of universal bundle Ht and universal quotient bundle Qn−d+1−t under ψ, φ). Denote by
c1 , . . . , ct and d1 , . . . , dn−d+1−t the Chern classes of ψ ∗ Ht and ψ ∗ Qn−d+1−t , respectively, and by
C1 , . . . , Ct and D1 , . . . , Dn−d+1−t the Chern classes of φ∗ Ht and φ∗ Qn−d+1−t , respectively.
On G(d, n) (2 ≤ d ≤ n − d), we have an exact sequence
⊕n−d+1
0 → φ∗ Ht → OG(d,n) → φ∗ Qn−d+1−t → 0

which is the pull back of the universal exact sequence

⊕n−d+1
0 → Ht → OG(t,n−d+1−t) → Qn−d+1−t → 0

on G(t, n − d + 1). Then

c(φ∗ Ht ) · c(φ∗ Qn−d+1−t ) = 1,

10
i.e.
(1 + C1 + · · · + Ct ) · (1 + D1 + · · · + Dn−d+1−t ) = 1.
Since n − d + 1 < d(n − d) = dimG(d, n), obviously we have Ct · Dn−d+1−t = 0.
On the other hand, on Pn−d ⊂ G(d, n), we also have an exact sequence
0 → ψ ∗ Ht → OP⊕n−d+1
n−d
→ ψ ∗ Qn−d+1−t → 0.
Then
c(ψ ∗ Ht ) · c(ψ ∗ Qn−d+1−t ) = 1,
i.e.
(1 + c1 + · · · + ct ) · (1 + d1 + · · · + dn−d+1−t ) = 1.
Combining the above equation and the ψ nonconstant assumption, i.e.
c1 = degψ ∗ OG(t,n−d+1) (1) 6= 0,
we get ct 6= 0 and dn−d+1−t 6= 0.
In order to show that this contradicts that Ct · Dn−d+1−t = 0, we need to use some Schubert
classes in G(d, n) (see [8] Chapter 4). Let’s review some basic facts and fix some notation first.
Choose a complete flag V in Cn , that is, a nested sequence of subspaces
0 ⊂ V1 ⊂ · · · ⊂ Vn−d ⊂ Vn−d = Cn
with dim Vi = i. For a sequence a = (a1 , . . . , ad ) with n − d ≥ a1 ≥ . . . ≥ ad ≥ 0, we define the
Schubert cycle Σa (V) ⊂ G(d, n) to be the closed subset
Σa (V) = {Λ ∈ G(d, n)|dim(Vn−d+i−ai ∩ Λ ≥ i) for all i}.
The class σa := [Σa (V)] ia called Schubert class. By Theorem 4.1 in [8] , σa is of codimension
a1 + · · · + ad . To simplify notation, we generally suppress trailing zeros in the indices and write
σa1 ,...,as in place of σa1 ,...,as ,0,...,0 . With this notation, Pn−d can be represented as σn−d,...,n−d
and we may write
X X
(−1)t Ct = xa1 ,...,ad σa1 ,...,ad and Dn−d+1−t = ya1 ,...,ad σa1 ,...,ad ,
a1 +···+ad =t a1 +···+ad =n−d+1−t

where the xa1 ,...,ad s and ya1 ,...,ad s are non negative integers by the nefness of Ht∨ (the dual of
′ ′

Ht ) and Qn−d+1−t .
Pieri’s formula (see Proposition 4.9 in [8] ) tells us that for any integers b1 , . . . , bd with b1 +
· · · + bd = l ≤ n − d,
σb1 · · · σbd ∩ Pn−d = σb1 · · · σbd · σn−d,...,n−d,0 = σn−d,...,n−d,l .
Using Giambelli’s formula (see Proposition 4.16 in [8] ) for σa1 ,...,ad ’s, expanding the determinants
and then intersecting with σn−d,...,n−d,0 , we can immediately get the following identities:
ct = Ct ∩ Pn−d
= (−1)t xt,0,...,0 σt · σn−d,...,n−d,0
= (−1)t xt,0,...,0 σn−d,...,n−d,t ;
dn−d+1−t = Dn−d+1−t ∩ Pn−d
= yn−d+1−t,0,...,0 σn−d+1−t · σn−d,...,n−d,0
= yn−d+1−t,0,...,0 σn−d,...,n−d,n−d+1−t .

11
Hence xt,0,...,0 yn−d+1−t,0,...,0 6= 0. But

0 = Ct · Dn−d+1−t = xt,0,...,0 yn−d+1−t,0,...,0 σt · σn−d+1−t + . . . ,

where the summation is linear combination of Schubert cycles with non-negative coefficients by
the Littlewood-Richardson formula. Therefore

xt,0,...,0 yn−d+1−t,0,...,0 = 0,

a contradiction.
When G = Dn /Pn , Mx = G(2, n). By the above Lemma, we obtain the only morphisms
G(2, n) → G(t, n − 1) are constant for any integer 1 ≤ t ≤ [ n−1
2 ].
When G = En /P2 (n = 6, 7, 8), Mx = G(3, n) (n = 6, 7, 8). By the above Lemma, we
obtain the only morphisms G(3, n) → G(t, n − 2) (n = 6, 7, 8) are constant for any integer
1 ≤ t ≤ [ n−2
2 ]. Remarkably, in these cases, the value of ς can be appropriately enlarged. Since
G(3, n) (n = 6, 7, 8) is 3(n − 3) dimensional, we can easily know

dim(G(3, 6)) = 9 > dimG(t, 5),

dim(G(3, 7)) = 12 > dimG(t, 6)

and
dim(G(3, 8)) = 15 > dimG(t, 7) for all t ≥ 1.
Since the Picard number of G(3, n) (n = 6, 7, 8) is one, the only morphisms

G(3, 6) → G(t, 5),

G(3, 7) → G(t, 6)
and
G(3, 8) → G(t, 7)
are all constant for any integer t ≥ 1.

Case III. Mx is spinor variety Sn (n = 3, 4, 5, 6). The Chow ring of Sn is presented as a

quotient of Z[X1 , . . . , Xn ] module the relations
s−1
X
Xs2 + 2 (−1)i Xs+i Xs−i + (−1)s X2s = 0
i=1

for 1 ≤ s ≤ n, where Xj, sare the Schubert classes of codimension j, X0 = 1 and Xj = 0 for j < 0
or j > n (see Section 3.2 in [26] ). In particular, dimH 2t (Mx , C) = 1 for every t ≤ [ 52 ]. Hence,
the only morphisms Mx → G(t, 5) are constant for any integer 1 ≤ t ≤ [ 25 ]. Remarkably, for
n = 4, 5, 6, the value of ς can be appropriately enlarged. Due to the dimension of Sn is n(n+1) 2 ,
one can check that
dim(S4 ) = 10 > dimG(t, 6),
dim(S5 ) = 15 > dimG(t, 7)
and
dim(S6 ) = 21 > dimG(t, 9) for all t ≥ 1.

12
Since the Picard number of Sn is one, the only morphisms

S4 → G(t, 6),

S5 → G(t, 7)
and
S6 → G(t, 9)
are all constant for any integer t ≥ 1.

Case IV. Mx = E6 /P6 . The Chow ring of E6 /P6 have the following form (see [7] Theorem
5). Let y1 , y4 be the Schubert classes on E6 /P6 . Then

A(E6 /P6 ) = Z[y1 , y4 ]/(r9 , r12 ),

where
r9 = 2y19 + 3y1 y42 − 6y15 y4 ;
r12 = y43 − 6y14 y42 + y112 .
Lemma 3.2. There are no nonconstant maps from E6 /P6 to G(t, 10) for any integer 1 ≤ t ≤ 5.
Proof. (i). 1 ≤ t ≤ 3. Since dimH 2t (Mx , C) = 1, by the proof of Lemma 3.4 in paper [18] , the
only morphisms E6 /P6 → G(t, 10) are constant.
(ii). t = 4. Let φ be a morphism from E6 /P6 to G(4, 10). On E6 /P6 , we have an exact
sequence
⊕10
0 → φ∗ H4 → OG(d,n) → φ∗ Q6 → 0.
Then c(φ∗ H4 ) · c(φ∗ Q6 ) = 1. According to the Chow ring of E6 /P6 , we can expand the equation
into the following form:

(1 + a1 y1 + a2 y12 + a3 y13 +a4 y14 + ae4 y4 ) · (1 + b1 y1 + b2 y12 +

b3 y 3 + b4 y 4 + be4 y4 + b5 y 5 + be5 y1 y4 + b6 y 6 + be6 y 2 y4 ) = 1.
1 1 1 1 1

Since A4 (E6 /P6 ) is freely generated by the classes y14 , y4 , the above equation implies that the
coefficient of y4 is 0, i.e. ae4 + be4 = 0. On the other hand, A8 (E6 /P6 ) is freely generated by
the classes y18 , y14 y4 , y42 , the above equation implies that the coefficient of y42 is also zero, i.e.
ae4 · be4 = 0. Hence ae4 = be4 = 0. Therefore, this case can boil down to case (i).
(iii). t = 5. By iterating the previous process, we get c(φ∗ H5 ) · c(φ∗ Q5 ) = 1, i.e.

(1 + a1 y1 + a2 y12 + a3 y13 + a4 y14 + ae4 y4 + a5 y15 + ae5 y1 y4 )

·(1 + b1 y1 + b2 y12 + b3 y13 + b4 y14 + be4 y4 + b5 y15 + be5 y1 y4 ) = 1.

In a similar way, we can prove ae4 = be4 = 0. Next, let’s consider the vanishing of ae5 . Since
A5 (E6 /P6 ) is freely generated by the classes y15 , y1 y4 , the above equation implies that the coeffi-
cient of y1 y4 is 0, i.e. ae5 + be5 = 0. On the other hand, A6 (E6 /P6 ) is freely generated by the classes
y16 , y12 y4 , the above equation implies that the coefficient of y12 y4 is also zero, i.e. a1 be5 + b1 ae5 = 0.
Combining these equations with a1 = −b1 , we obtain that ae5 = be5 = 0. Therefore, this case can
boil down to case (i).

13
 Case V. Mx = E7 /P7 . The Chow ring of E7 /P7 have the following form (see [7] Theorem 6).
Let y1 , y5 , y9 be the Schubert classes on E7 /P7 . Then

A(E7 /P7 ) = Z[y1 , y5 , y9 ]/(r10 , r14 , r18 ),

where
r10 = y52 − 2y1 y9 ;
r14 = 2y5 y9 − 9y14 y52 + 6y19 y5 − y114 ;
r1 8 = y92 + 10y13 y53 − 9y18 y52 + 2y113 y5 .
Lemma 3.3. There are no nonconstant maps from E7 /P7 to G(t, 13) for any integer 1 ≤ t ≤ 6.
Proof. (i). 1 ≤ t ≤ 4. Since dimH 2t (Mx , C) = 1, by the proof of Lemma 3.4 in paper [18] , the
only morphisms E7 /P7 → G(t, 13) are constant.
(ii). t = 5. Let φ be a morphism from E7 /P7 to G(5, 13). On E7 /P7 , we have an exact
sequence
⊕13
0 → φ∗ H5 → OG(d,n) → φ∗ Q8 → 0.
Then c(φ∗ H5 ) · c(φ∗ Q8 ) = 1. According to the Chow ring of E7 /P7 , we can expand the equation
into the following form:

(1 + a1 y1 +a2 y12 + a3 y13 + a4 y14 + a5 y15 + ae5 y5 ) · (1 + b1 y1 + b2 y12 + b3 y13

+ b4 y 4 + b5 y 5 + be5 y5 + b6 y 6 + be6 y1 y5 + b7 y 7 + be7 y 2 y5 + b8 y 8 + be8 y 3 y5 ) = 1.
1 1 1 1 1 1 1

Since A5 (E7 /P7 ) is freely generated by the classes y15 , y5 , the above equation implies that the
coefficient of y5 is 0, i.e. ae5 + be5 = 0. On the other hand, A10 (E7 /P7 ) is freely generated by
the classes y110 , y15 y5 , y52 , the above equation implies that the coefficient of y52 is also zero, i.e.
ae5 be5 = 0. Hence ae5 = be5 = 0. Therefore, this case can boil down to case (i).
(iii). t = 6. By iterating the previous process, we get c(φ∗ H6 ) · c(φ∗ Q7 ) = 1, i.e.

(1 + a1 y1 + a2 y12 + a3 y13 + a4 y14 + a5 y15 + ae5 y5 + a6 y16 + ae6 y1 y5 ) · (1 + b1 y1 +

b2 y 2 + b3 y 3 + b4 y 4 + b5 y 5 + be5 y5 + b6 y 6 + be6 y1 y5 + b7 y 7 + be7 y 2 y5 ) = 1.
1 1 1 1 1 1 1

In a similar way, we can prove ae5 = be5 = 0. Next, let’s consider the vanishing of ae6 . Since
A6 (E7 /P7 ) is freely generated by the classes y16 , y1 y5 , the above equation implies that the coef-
ficient of y1 y5 is 0, i.e. ae6 + be6 = 0. On the other hand, A12 (E7 /P7 ) is freely generated by the
classes y112 , y12 y52 , y17 y5 , the above equation implies that the coefficient of y12 y52 is also zero, i.e.
ae6 be6 = 0. Hence, ae6 = be6 = 0. Therefore, this case can boil down to case (i).
Case VI. Mx = C3 /P3 . The Chow ring of C3 /P3 is

A(C3 /P3 ) = Z[y1 , y3 ]/(y14 − 8y1 y3 , x23 ),

where y1 , y3 are the Schubert classes on C3 /P3 (see Section 3.1 in [26] ). Since dimH 2t (C3 /P3 , C) =
1 for 1 ≤ t ≤ 2, the only morphisms C3 /P3 → G(t, 5) are constant for any integer 1 ≤ t ≤ 2.
Summing up, we have obtained the following Table 2.
Theorem 3.4. Let E be a uniform r-bundle on a generalized Grassmannian G with extremal
node marked. If r ≤ ς(G), then E splits as a direct sum of line bundles.

14
 Table 2: ς(G)

Dn /P1 E6 /P1 E6 /P2 E6 /P6

An /P1 Bn /P1 Cn /P1 F4 /P1 G2 /P1
G Dn /Pn−1 E7 /P1 E7 /P2 E7 /P7
An /Pn Bn /Pn Cn /Pn F4 /P4 G2 /P2
Dn /Pn E8 /P1 E8 /P2 E8 /P8
2n − 5 6 5 6
n−1 2n − 3 2n − 2 5 3
ς(G) n−1 7 6 10
n−1 n−1 n−1 5 1
n−1 9 7 13

Remark 3.5. One can find the value of ς(G) is equal to or bigger than the value of ς that we
analyzed such that the morphism Mx → G(t, ς) can only be constant. ς(G) is equal to ς except
G = Cn /P1 and G2 /P1 . The reason is that uniform 2n − 2 bundles split on Cn /P1 ∼
= P2n−1 and
uniform 3 bundles split on G2 /P1 ∼
= Q5 by the previous arguments. In all cases, however, the
morphism Mx → G(t, ς(G) − 2) can only be constant.
Corollary 3.6. If G is a generalized Grassmannian with marked point k, k is not extremal, then
Mx is a product of rational homogeneous spaces. If

Mx = A1 × · · · × At (2 ≤ t ≤ 3)

and
r ≤ ς(G) := min{ς(A′1 ), . . . , ς(A′t )},
where Ai (1 ≤ i ≤ t) is the special family of lines of class αˇk through x on the generalized
Grassmannian A′i with extremal node marked. Then uniform r-bundle E on G splits as a direct
sum of line bundles.
Proof. In order to prove E splits, we just need to prove that the only morphisms Mx → G(t, r) are
constant. Let φ be a morphism from Mx to G(t, r). Because every Ai (1 ≤ i ≤ t) can be regarded
as the subspace of Mx corresponding to the special family of lines of class αˇk through x on A′i ,
we can consider the restriction of φ to Ai (1 ≤ i ≤ t). By assumption r ≤ min{ς(A′1 ), . . . , ς(A′t )},
all the restriction maps are constant. Hence, φ is also constant.
Remark 3.7.
• For the case where Mx = P1 or Mx = P1 × · · · , we can say nothing about the splitting
result according to our theorem.
• In some cases, the value of ς(G) cannot be expanded anymore, which means that there
exist uniform but nonsplitting ς(G) + 1-bundles. For instance, Grassmannian An−1 /Pd =
G(d, n) (d ≤ n − d) has uniform but nonsplitting d-bundle Hd (the universal bundle of
G(d, n)); Spinor variety Dn+1 /Pn+1 = Sn has uniform but nonsplitting n + 1-bundle Qn+1
(the universal quotient bundle of Sn ).
• Compare to the main theorem (Theorem 3.1) in [18] , we improve their results. In particular,
we enlarge the splitting threshold for uniform bundles on Hermitian symmetric spaces
E6 /P6 from ≥ 5 to ≥ 6 and E7 /P7 from ≥ 7 to ≥ 10 (see Table 1 in [18] ).
Corollary 3.8. Let G be a generalized Grassmannian covered by linear projective subspaces of
dimension 2 and E be an r-bundle on G. If E splits as a direct sum of line bundles when it
restricts to every P2 ⊆ G, then E splits as a direct sum of line bundles on G.

15
Proof. If G is a projective space, then the result holds ( [23] Theorem 2.3.2). Suppose G = G/Pk
which is not a projective space and k is the unique black node, where G is a simple Lie group
and Pk is a maximal parabolic subgroup of G. The condition implies that E is uniform. The
reason for this is the fact every line L is contained in two different P2 by Theorem 2.9.
We prove the corollary by induction on r. If we have the exact sequence of vector bundles

0→M →E→N →0 (3.2)

on G, where the rank of M and N are smaller than r, such that

r−t
M
⊕t
M |Z = OZ (at+i ), N |Z = OZ ,
i=1

for every Z ≃ P2 , then by the induction hypothesis, M and N split. Since H 1 (G, N ∨ ⊗ M ) = 0,
the above exact sequence splits and hence also E.
Similar to the proof of Theorem 3.3 in [6] , on U = G/(Pk ∩ PN (k) ), we can obtain an exact
sequence
0→M f → q1 ∗ E → N e → 0.
If we prove that the morphism ϕ is constant for every x ∈ G, then there exist two bundles M ,
N over G with Mf = q1 ∗ M, N
e = q1 ∗ N . By projecting the bundle sequence

0 → q1 ∗ M → q1 ∗ E → q1 ∗ N → 0

onto G, we can get the desired exact sequence (3.2). Thus, to prove the existence of the above
exact sequence, it suffices to show that the map

ϕ : Mx → Gk (t − 1, Pk (Ex∨ ))

is constant for every x ∈ G . Given a projective subspace Z of dimension 2 and a line L ⊆ Z,

we take any point x ∈ L and denote by Z ′ the subspace of Mx corresponding to the tangent
directions to Z at x. By the hypothesis, E|Z is a direct sum of line bundles, so

ϕ|Z ′ : Z ′ → Gk (t − 1, Pk (Ex∨ ))

is constant. Since G covered by linear projective subspaces of dimension 2 and Mx is chain-

connected by P1 , ϕ is constant for every x ∈ G.

3.2 Uniform vector bundles on rational homogeneous spaces

Let
X = G/P ≃ G1 /PI1 × G2 /PI2 × · · · × Gm /PIm ,
where Gi is a simple Lie group with Dynkin diagram Di whose set of nodes is Di and PIi is a
parabolic subgroup of Gi corresponding to Ii ⊂ Di . We set F (Ii ) := Gi /PIi by marking on the
Dynkin diagram Di of Gi the nodes corresponding to Ii . Let δi be a node in Di and N (δi ) be
the set of nodes in Di that are connected to δi .
If δi ∈ Ii , we call
δc δc
Mi i := Gi /Pi i × G\i /PIi (1 ≤ i ≤ m),
c
the i-th special family of lines of class δ̌i by Theorem 2.3, where Pi i := PIi \δi ∪N (δi ) and G\
δ
i /PIi
is G1 /PI1 × G2 /PI2 × · · · × Gm /PIm by deleting i-th term Gi /PIi . For i = 1 and δ ∈ I, we will
use the notation Mδ to denote the special family of lines of class δ̌.

16
 For x ∈ X, we call
δc
SPVMRT(δ
x
i)
= {L ∈ Mi i |x ∈ L}
the δi -th special part of variety of minimal rational tangents at x (sometimes we just write
SPVMRTx if there is no confusion).
Fix δi ∈ Ii . SPVMRTx(δi ) is just the special family of lines of class δ̌i through x on the
generalized Grassmannian G δi whose Dynkin diagram Dδi is the maximal sub-diagram of (Di , Ii )
with the only marked point δi . Denote ν(X, δi ) := ς(G δi ). Let

ν(X) := mini {minδi ∈Ii {ν(X, δi )}}.

δc
Definition 3.9. A vector bundle E on X is called poly-uniform with respect to Mi i for every
δc
i (1 ≤ i ≤ m) and every δi ∈ Ii if the restriction of E to every line in Mi i has the same splitting
type. We also call that E poly-uniform with respect to all the special families of lines.
Let F be a torsion free coherent sheaf of rank r over X. Fix integer i (1 ≤ i ≤ m) and
δc
δi ∈ Ii . Since the singularity set S(F ) of F has codimension at least 2, there are lines L ∈ Mi i
which do not meet S(F ). If
(δ )
F |L ∼
= OL (a1 i ) ⊕ · · · ⊕ OL (ar(δi ) ).
Let
(δ ) (δ )
c1 i (F ) = a1 i + · · · + ar(δi ) ,
which is independent of the choice of L. We set
(δ )
c1 i (F )
µ(δi ) (F ) = .
rk(F )
Definition 3.10. A torsion free coherent sheaf E over X is δi -semistable (δi -stable) if for every
coherent subsheaf F ⊆ E with 0 < rk(F ) < rk(E), we have

µ(δi ) (F ) ≤ (<) µ(δi ) (E).

If E is not δi -semistable, then we call E is δi -unstable.

Proposition 3.11. Fix integer i (1 ≤ i ≤ m) and δi ∈ Ii . Let E be a uniform r-bundle on X
(δ ) (δ ) (δ ) (δ ) δc
of type (a1 i , . . . , ar i ), a1 i ≤ . . . ≤ ar i with respect to Mi i . If r ≤ ν(X, δi ) − 2 and these
(δ )
aj i ’s are not all same, then E can be expressed as an extension of uniform bundles with respect
δc
to Mi i . In particular, E is δi -unstable.
Proof. After twisting with an appropriate line bundle, we can assume that E has the splitting
type
(δ ) (δi ) (δi )
aE i = (0, . . . , 0, at+1 , . . . , a(δi)
r ), at+i > 0, for i = 1, . . . , r − t.
δc
with respect to Mi i .
Let’s consider the standard diagram
δc q2 c
Ui i = Gi /PIi ∪N (δi ) × G\
i /PIi
/ Mδi
i (3.3)
q1

X.

17
For L ∈ Mδi i , the q2 -fiber
c
e = q2 −1 (L) = {(x, L) ∈ X × Mδi |x ∈ L},
L i

is mapped under q1 to the line L identically in G/P and we have

q1 ∗ E|Le ∼
= E|L .
For x ∈ G/P , the q1 -fiber q1 −1 (x) is mapped isomorphically under q2 to the subvariety
δc
SPVMRTx = {L ∈ Mi i |x ∈ L}.

Because
r−t
M (δ ) (δ )
E|L ∼ ⊕t
= OL ⊕ i
OL (at+i i
), at+i > 0,
i=1
 
0
h q2 −1 (L), q1 ∗ (E ∨ )|q−1 (L) = t
2

δc ∗ ∨ δc
for all L ∈ Thus the direct image q2 ∗ q1 (E ) is a vector bundle of rank t over Mi i . The
Mi i .
canonical homomorphism of sheaves

q2 ∗ q2 ∗ q1 ∗ (E ∨ ) → q1 ∗ (E ∨ )

makes N e the
e ∨ := q2 ∗ q2 ∗ q1 ∗ (E ∨ ) to be a subbundle of q1 ∗ (E ∨ ). Because over each q2 -fiber L,
evaluation map
e ∨ | e = H 0 (L,
N e q1 ∗ (E ∨ )| e ) ⊗k O e → q1 ∗ (E ∨ )| e
L L L L

e ∨ | e with O ⊆ O ⊕
⊕t ⊕t L r−t (δ ) ∨ δ c
identifies N L L L i=1 OL (−at+i ) = E |L . Over Ui we thus obtain an exact
i i

sequence
0→M f → q1 ∗ E → N e→0
of vector bundles, whose restriction to q2 -fibers L e looks as follows:

0 f| e
/M / q1 ∗ E| e e |e
/N /0
L L L

∼
= ∼
= ∼
=
  
0 / Lr−t OL (a(δi ) ) / O⊕t ⊕ Lr−t OL (a(δi ) ) / O⊕t / 0.
i=1 t+i L i=1 t+i L

e ∨ is a subbundle of q1 ∗ (E ∨ ) of rank t, for every point x ∈ G, it provides a morphism

Because N

ϕ : SPVMRTx → Gk (t − 1, Pk (Ex∨ )).

Since r ≤ ν(X, δi ) − 2, morphism ϕ is constant by Remark 3.5 and the definition of ν(X, δi ). It
follows that Mf and N e are trivial on all q1 -fibers. So the canonical morphisms q1 ∗ q1 Mf→M
f
∗
∗ e e f
and q1 q1 ∗ N → N are isomorphisms. Hence there are uniform bundles M = q1 ∗ M , N = q1 ∗ N e
δic
with respect to Mi over X with
f = q1 ∗ M, N
M e = q1 ∗ N.

By projection formula and the rationality of the q1 fiber SPVMRTx , we project the bundle
sequence
0 → q1 ∗ M → q1 ∗ E → q1 ∗ N → 0

18
onto X to get the exact sequence
0 → M → E → N → 0. (3.4)
(δi ) (δi )
So µ (N ) > µ (E) and thus E is δi -unstable
Proposition 3.12. On X, if an r-bundle E is poly-uniform with respect to all the special families
δc
of lines on X such that the splitting type with respect to Mi i is (a(δi ) , . . . , a(δi ) ) for each i and
δi , then E splits as a direct sum of line bundles.
Proof. After twisting with an appropriate line bundle, we can assume that E is trivial on all the
special families of lines on X. We are going to show that E is trivial.
Let’s first consider the case that X is a generalized flag manifold, which corresponds to a
connected marked Dynkin diagram D(X) with l black nodes. We prove the lemma by induction
on l. For l = 1, X is just a generalized Grassmannian. Then the result holds by [1] Proposition 1.2.
Suppose the assertion is true for all generalized flag manifolds with connected marked Dynkin
diagram and l′ black nodes (1 ≤ l′ < l). Let’s consider the natural projection
π : X → X ′,
where X ′ is corresponding to the marked Dynkin diagram D(X) by changing the first black
node δ to white. It’s not hard to see that every π-fiber π −1 (x) is isomorphic to the generalized
Grassmannian G δ with the only marked point δ. Since the restriction of E to every line in Mδ
is trivial, so is to every line in π −1 (x). Thus E is trivial on all π-fibers by [1] Proposition 1.2. It
follows that E ′ = π∗ E is an algebraic vector bundle of rank r over X ′ and E ∼ = π∗ E ′ .
′ ′
Claim. E is trivial on all the special families of lines on X .
In fact, let γ be a black node in D(X) that is different from δ and L be a line in Mγ . Then
π(L) is a line in the family of lines of X ′ . When γ runs through all black nodes except δ and L
runs through all lines in Mγ in X, π(L) also runs through all lines in all the special families of
lines of X ′ . The projection π induces an isomorphism
E ′ |π(L) ∼
= π ∗ E ′ |L ∼
= E|L .
We identify L with π(L). Since E|L is trivial for every line L in Mγ by assumption, E ′ |L is
trivial for every line L in all the special families of lines of X ′ . By the induction hypothesis, E ′
is trivial. Thus E ∼= q ∗ E ′ is trivial.
Now let’s think about the general case, where the marked Dynkin diagram of X is not
connected. Assume X can be decomposed into a product
X = G/P ≃ G1 /PI1 × G2 /PI2 × · · · × Gm /PIm ,
where m ≥ 2 and Gi (1 ≤ i ≤ m) is a simple Lie group with connected Dynkin diagram. We
prove E is trivial by induction on m. For m = 1, the result holds from the previous analysis.
Consider the natural projection
f : X → X ′ := G2 /PI2 × · · · × Gm /PIm ,
it’s easy to see that every f -fiber f −1 (x) is isomorphic to G1 /PI1 . By assumption, E is trivial
on all the special families of lines on f −1 (x) ∼= G1 /PI1 . Thus E is trivial on all f -fibers by the
previous analysis. It follows that E ′ = f∗ E is an algebraic vector bundle of rank r over X ′ and
E∼ = f ∗ E ′ . Similarly, we can prove that E ′ is trivial and thus E is trivial.
Theorem 3.13. On X, if r-bundle E is poly-uniform with respect to all the special families of
lines and r ≤ ν(X) − 2, then E is δi -unstable for some δi (1 ≤ i ≤ m) or E splits as a direct
sum of line bundles.
Proof. It is obviously from Proposition 3.11 and 3.12.

19
4 Semistable vector bundles on rational homogeneous spaces
Let G be a simple Lie group and D = {1, 2, . . . , n} be the set of nodes of the Dynkin diagram D of
G. Denote by Pk the parabolic subgroup of G corresponding to the node k. Consider X = G/Pk
in its minimal homogeneous embedding. Denote by M := G/PN (k) the generalized flag manifold
defined by the marked Dynkin diagram (D, N (k)) and by U := G/Pk,N (k) = G/(PN (k) ∩ Pk )
the universal family, which has a natural P1 -bundle structure over M, i.e. we have the natural
diagram:
q
U /M (4.1)
p

X = G/Pk .

Given x ∈ X, Mx = q(p−1 (x)) coincides with H/PN (k) by Remark 2.7 where the set of nodes of
the Dynkin diagram H is D(H) = D\k .
In the above setting, we will show that the splitting type of TU /X |q−1 (l) take the form
(−1, . . . , −1), (−1, . . . , −1, −2, . . . , −2) or (−3). Building upon this assert, we will generalize
the Grauert-Mülich-Barth theorem to any rational homogeneous spaces.

4.1 The classical simple Lie algebras

We say G is of classical type when its Dynkin diagram is of type An , Bn , Cn or Dn . Because
X = G/Pk has clear geometric explanations, we can even write down the specific form of the
relative tangent bundle TU /X . In the case of type An , the relative tangent bundle TU /X is known
(see [6] Lemma 5.6) and the splitting type of TU /X |q−1 (l) is (−1, . . . , −1). So let’s just consider
the remaining three types.

4.1.1 The marked Dynkin diagram (Bn , k)

In this section, we consider the Dynkin diagram Bn , which corresponds to the classical Lie group
SO2n+1 . Denote by Bn /PI := SO2n+1 /PI the generalized flag manifold with I = {i1 , . . . , is } ⊆
D. Let V = C2n+1 be a vector space equipped with a nondegenerate symmetric bilinear form
Q. Then Bn /PI is actually the odd Orthogonal flag manifold OG(i1 , . . . , is ; 2n + 1), which
parametrizes flags
Vi1 ⊂ Vi2 ⊂ · · · ⊂ Vis ⊂ V,
where each Vit (1 ≤ t ≤ s) is an it -dimensional isotropic subspace in V .
There is a universal flag of subbundles
0 = H0 ⊂ Hi1 ⊂ · · · ⊂ His ⊂ Hi⊥s ⊂ · · · ⊂ Hi⊥1 ⊂ OBn /PI × V
on Bn /PI , where Hi⊥t is the Q-orthogonal complement of Hit , rank Hit = it and rank Hi⊥t =
2n + 1 − it (1 ≤ t ≤ s).
(1) For k = 1, the odd Orthogonal Grassmannian Bn /P1 := OG(1, 2n + 1) is just the quadric
Q2n−1 . In this case, M = Bn /P2 and U = Bn /P1,2 . We have the natural diagram
q
U = Bn /P1,2 / M = Bn /P2 (4.2)
p

X = Bn /P1

20
and Mx is Bn−1 /P1 , i.e. the quadric Q2n−3 .
e = q −1 (l) ⊂ U for l ∈ M. For the relative tangent bundle TU /X , we have
Lemma 4.1. Let L

TU /X |Le = OLe (−1)⊕2n−3 .

Proof. For x ∈ X, the p-fiber p−1 (x) = {(x, l)|x ∈ L} is isomorphic to Mx = Q2n−3 . Over
p−1 (x) ∼
= Q2n−3 , we have the universal bundle sequence

0 → OQ2n−3 (−1) → OQ2n−3 (−1)⊥ → OQ2n−3 (−1)⊥ /OQ2n−3 (−1) → 0, (4.3)

where OQ2n−3 (−1) is the rank 1 tautological bundle over Q2n−3 , which is the pull back of
OP2n−2 (−1) under the embedding Q2n−3 ֒→ P2n−2 and OQ2n−3 (−1)⊥ denotes its Q-orthogonal
complement. Notice that the tangent bundle of Q2n−3 can be represented as

TQ2n−3 = OQ2n−3 (1) ⊗ OQ2n−3 (−1)⊥ .

Let’s consider the exact sequence

0 → H2 /H1 → H2⊥ /H1 → q ∗ (H2⊥ /H2 ) → 0,

of vector bundles on U = Bn /P1,2 . By restricting it to the the p-fiber p−1 (x), we obtain the
universal bundle sequence (4.3) on p−1 (x). Therefore, TU /X = (H2 /H1 )∨ ⊗ q ∗ (H2⊥ /H2 ) and

TU /X |Le = OLe (−1)⊕2n−3 .

(2) For k (2 ≤ k ≤ n − 1), the odd Orthogonal Grassmannian Bn /Pk := OG(k, 2n + 1)

parametrizes the k-dimensional isotropic subspaces in V . In this case, M = Bn /Pk−1,k+1 and
U = Bn /Pk−1,k,k+1 . We have the natural diagram:
q
U = Bn /Pk−1,k,k+1 / M = Bn /Pk−1,k+1 (4.4)
p

X = Bn /Pk

and Mx is Pk−1 × Bn−k /P1 = Pk−1 × Q2(n−k)−1 .

e = q −1 (l) ⊂ U for l ∈ M. For the relative tangent bundle TU /X , we have
Lemma 4.2. Let L

TU /X |Le = OLe (−1)⊕2n−k−2 .

Proof. It’s not hard to check that over U, we have the following two exact sequences of vector
bundles:

0 → (Hk /Hk−1 )∨ → p∗ Hk∨ → q ∗ Hk−1

∨
→ 0,

⊥
0 → Hk+1 /Hk → Hk+1 /Hk → q ∗ (Hk+1
⊥
/Hk+1 ) → 0.

21
We will consider the following diagram
π1 π2
U1 = Bn /Pk−1,k o U = Bn /Pk−1,k,k+1 / U2 = Bn /Pk,k+1 (4.5)
❖❖❖ ♦♦
❖❖❖ ♦♦♦
❖❖❖ ♦♦
❖❖❖ ♦♦
❖❖ p ♦♦♦♦
p1 ❖❖❖ ♦ p
❖❖❖ ♦♦ 2
❖❖❖ ♦♦♦♦♦
❖❖❖  ♦♦♦
' w♦
X.
All the morphisms in the above diagram are projections. For any x ∈ X, the p1 -fiber p−1 1 (x) is
isomorphic to Pk−1 and the p2 -fiber p−1
2 (x) is isomorphic to Q 2(n−k)−1
. Note that for x ∈ X,
the projection q induces an isomorphism
q|p−1 (x) : p−1 (x) → Mx = Pk−1 × Q2(n−k)−1 .
Hence p−1 (x) = p1−1 (x) × p2−1 (x). So we get
TU /X = π1∗ (TU1 /X ) ⊕ π2∗ (TU2 /X ).

We mimic the proof of Lemma 5.6 in [6] to conclude that

π1∗ (TU1 /X ) ∼
= Hk /Hk−1 ⊗ q ∗ Hk−1
∨ ∼ (Hk+1 /Hk )∨ ⊗ q ∗ (H ⊥ /Hk+1 ).
, π2∗ (TU2 /X ) = k+1

L

Therefore, TU /X ∼ ∗
= Hk /Hk−1 ⊗ q Hk−1 ∨ ∨ ∗ ⊥
(Hk+1 /Hk ) ⊗ q (Hk+1 /Hk+1 ) and

TU /X |Le = OLe (−1)⊕2n−k−2 .

(3) For k = n (corresponding to the short root αn ), the odd Orthogonal Grassmannian
Bn /Pn := OG(n, 2n + 1) parametrizes the n-dimensional isotropic subspaces in V . In this case,
M = Bn /Pn−1 and U = Bn /Pn−1,n . We have the natural diagram
q
U = Bn /Pn−1,n / M = Bn /Pn−1 (4.6)
p

X = Bn /Pn

and Mx is Pn−1 .
e = q −1 (l) ⊂ U for l ∈ M. For the relative tangent bundle TU /X , we have
Lemma 4.3. Let L

TU /X |Le = OLe (−2)⊕n−1 .

Proof. It’s not hard to check that over U, we have the following exact sequence of vector bundles:
0 → (Hn /Hn−1 )∨ → p∗ Hn∨ → q ∗ Hn−1
∨
→ 0.
Restricting it to the p-fiber p−1 (x) ∼
= Pn−1 is just the Euler sequence on Pn−1 . In fact, the
projection p identifies U in a canonical fashion with the projective bundle P(Hn ) of X. Hence
TU /X ∼
= Hn /Hn−1 ⊗ q ∗ Hn−1
∨
and

TU /X |Le = OLe (−2)⊕n−1 .

22
4.1.2 The marked Dynkin diagram (Cn , k)
In this section, we consider the Dynkin diagram Cn , which corresponds to the classical Lie group
Sp2n . Denote by Cn /PI := Sp2n /PI the generalized flag manifold with I = {i1 , . . . , is } ⊆ D.
Let V = C2n be a vector space equipped with a nondegenerate skew-symmetric bilinear form
Ω. Then Cn /PI is actually the Lagrangian flag manifold LG(i1 , . . . , is ; 2n), which parametrizing
flags
Vi1 ⊂ Vi2 ⊂ · · · ⊂ Vis ⊂ V,
where each Vit (1 ≤ t ≤ s) is an it -dimensional isotropic subspace in V .
There is universal flag of subbundles
0 = H0 ⊂ Hi1 ⊂ · · · ⊂ His ⊂ Hi⊥s ⊂ · · · ⊂ Hi⊥1 ⊂ OCn /PI × V

on Cn /PI , where Hi⊥t is the Ω-orthogonal complement of Hit , rank Hit = it and rank Hi⊥t =
2n − it (1 ≤ t ≤ s).
(1) For k = 1 (corresponding to the short root α1 ), the Lagrangian Grassmannian Cn /P1 :=
LG(1, 2n) is just the projective space P2n−1 . In this case, M = Cn /P2 and U = Cn /P1,2 . We
have the natural diagram
q
U = Cn /P1,2 / M = Cn /P2 (4.7)
p

X = Cn /P1

and Mx is Cn−1 /P1 , i.e. the the projective space P2n−3 .

e = q −1 (l) ⊂ U for l ∈ M. For the relative tangent bundle TU /X , we have
Lemma 4.4. Let L

TU /X |Le = OLe (−1)⊕2n−4 ⊕ OLe (−2).

Proof. It’s not hard to check that over U, we have the following exact sequence of vector bundles:
0 → H2 /H1 → p∗ (H1⊥ /H1 ) → H1⊥ /H2 → 0.

Restricting it to the p-fiber p−1 (x) ∼

= P2n−3 is just the Euler sequence on P2n−3 . In fact, the
projection p identifies U in a canonical fashion with the projective bundle P(H1⊥ /H1 ) of X. Hence
TU /X ∼
= (H2 /H1 )∨ ⊗ H1⊥ /H2 and

TU /X |Le = OLe (−1)⊕2n−4 ⊕ OLe (−2).

(2) For k (2 ≤ k ≤ n − 1) (corresponding to the short root αk ), the Lagrangian Grassmannian

Cn /Pk := LG(k, 2n + 1) parametrizes the k-dimensional isotropic subspaces in V . In this case
M = Cn /Pk−1,k+1 and U = Cn /Pk−1,k,k+1 . We have the natural diagram
q
U = Cn /Pk−1,k,k+1 / M = Cn /Pk−1,k+1 (4.8)
p

X = Cn /Pk

and Mx is Pk−1 × Cn−k /P1 = Pk−1 × P2(n−k)−1 .

23
 e = q −1 (l) ⊂ U k for l ∈ M. For the relative tangent bundle TU /X , we have
Lemma 4.5. Let L

TU /X |Le = OLe (−1)⊕2n−k−3 ⊕ OLe (−2).

Proof. It’s not hard to check that over U, we have the following two exact sequences of vector
bundles:

0 → (Hk /Hk−1 )∨ → p∗ Hk∨ → q ∗ Hk−1

∨
→ 0,

0 → Hk+1 /Hk → p∗ (Hk⊥ /Hk ) → Hk⊥ /Hk+1 → 0.

We will consider the following diagram

π1 π2
U1 = Cn /Pk−1,k o U = Cn /Pk−1,k,k+1 / U2 = Cn /Pk,k+1 (4.9)
❖❖❖ ♦
❖❖❖
❖❖❖ ♦♦ ♦♦♦
❖❖❖ ♦♦♦
❖❖❖ p ♦♦♦♦
p1 ❖❖❖ ♦♦ p2
♦
❖❖❖
❖❖❖ ♦♦♦♦♦
❖'  w♦♦♦♦
X.

All the morphisms in the above diagram are projections. For any x ∈ X, the p1 -fiber p−1 1 (x) is
isomorphic to Pk−1 and the p2 -fiber p2−1 (x) is isomorphic to P2(n−k)−1 . Note that for x ∈ X, the
projection q induces an isomorphism

q|p−1 (x) : p−1 (x) → Mx = Pk−1 × P2(n−k)−1 .

Hence p−1 (x) = p1−1 (x) × p2−1 (x). So we get

TU /X = π1∗ (TU1 /X ) ⊕ π2∗ (TU2 /X ).

We mimic the proof of Lemma 5.6 in [6] to conclude that

π1∗ (TU1 /X ) ∼
= Hk /Hk−1 ⊗ q ∗ Hk−1
∨
, π2∗ (TU2 /X ) ∼
= (Hk+1 /Hk )∨ ⊗ Hk⊥ /Hk+1 .

L

Therefore, TU /X ∼= Hk /Hk−1 ⊗ q ∗ Hk−1
∨
Hk+1 /Hk )∨ ⊗ Hk⊥ /Hk+1 and

TU /X |Le = OLe (−1)⊕2n−k−3 ⊕ OLe (−2).

(3) For k = n (corresponding to the long root αn ), the Lagrangian Grassmannian Cn /Pn :=
LG(n, 2n) parametrizes the n-dimensional isotropic subspaces in V . In this case M = Cn /Pn−1
and U = Cn /Pn−1,n . We have the natural diagram
q
U = Cn /Pn−1,n / M = Cn /Pn−1 (4.10)
p

X = Cn /Pn

and Mx is Pn−1 .

24
 e = q −1 (l) ⊂ U for l ∈ M. For the relative tangent bundle TU /X , we have
Lemma 4.6. Let L

TU /X |Le = OLe (−1)⊕n−1 .

Proof. It’s not hard to check that over U, we have the following exact sequence of vector bundles:

0 → (Hn /Hn−1 )∨ → p∗ Hn∨ → q ∗ Hn−1

∨
→ 0.

Restricting it to the p-fiber p−1 (x) ∼

= Pn−1 is just the Euler sequence on Pn−1 . In fact, the
projection p identifies U in a canonical fashion with the projective bundle P(Hn ) of X. Hence
TU /X ∼
= Hn /Hn−1 ⊗ q ∗ Hn−1
∨
and

TU /X |Le = OLe (−1)⊕n−1 .

4.1.3 The marked Dynkin diagram (Dn , k)

In this section, we consider the Dynkin diagram Dn , it corresponds to the classical Lie group
SO2n . Denote by Dn /PI := SO2n /PI the generalized flag manifold with I = {i1 , . . . , is } ⊆ D.
Let V = C2n be a vector space equipped with a nondegenerate symmetric bilinear form Q. Then
Case 1. n−1, n ∈ I. Dn /PI is the even Orthogonal flag manifold OG(i1 , . . . , is−2 , n−1, n; 2n)
which parametrizes two families of flags
′
Vi1 ⊂ · · · ⊂ Vis−2 ⊂ Vn−1 ⊂ V and Vi1 ⊂ · · · ⊂ Vis−2 ⊂ Vn−1 ⊂V
′
where each Vit (1 ≤ t ≤ s − 2) is an it -dimensional isotropic subspace in V , Vn−1 and Vn−1 are
the (n − 1)-dimensional isotropic subspaces in V .
There are universal flags of subbundles

0 = H0 ⊂ Hi1 ⊂ · · · ⊂ Hn−1 ⊂ ODn /PI × V,

′
0 = H0 ⊂ Hi1 ⊂ · · · ⊂ Hn−1 ⊂ ODn /PI × V
′
on Dn /PI , where rank Hit = it (1 ≤ t ≤ s − 2) and rank Hn−1 = rank Hn−1 = n − 1.
Case 2. n − 1 ∈ I, n ∈
/ I or n − 1 ∈
/ I, n ∈ I. Dn /PI is one of the two irreducible components
of the even Orthogonal flag manifold OG(i1 , . . . , is−1 , n; 2n) (is = n), which parametrizes flag

Vi1 ⊂ Vi2 ⊂ · · · ⊂ Vn ⊂ V,

where each Vit (1 ≤ t ≤ s) is an it -dimensional isotropic subspace in V .

There is a universal flag of subbundles

0 = H0 ⊂ Hi1 ⊂ · · · ⊂ Hn ⊂ ODn /PI × V

on Dn /PI , where rank Hit = it (1 ≤ t ≤ s).

Case 3. n−1, n ∈/ I. Dn /PI is actually the even Orthogonal flag manifold OG(i1 , . . . , is−1 , is ; 2n),
which parametrizes flag
Vi1 ⊂ Vi2 ⊂ · · · ⊂ Vis ⊂ V,
where each Vit (1 ≤ t ≤ s) is an it -dimensional isotropic subspace in V .
There is a universal flag of subbundles

0 = H0 ⊂ Hi1 ⊂ · · · ⊂ His ⊂ Hi⊥s ⊂ · · · ⊂ Hi⊥1 ⊂ ODn /PI × V

25
on Dn /PI , where Hi⊥t is the Q-orthogonal complement of Hit , rank Hit = it and rank Hi⊥t =
2n − it (1 ≤ t ≤ s).
(1) For k (1 ≤ k ≤ n − 3), the even Orthogonal Grassmannian Dn /Pk := OG(k, 2n)
parametrizes the k-dimensional isotropic subspaces in V . In this case, M = Dn /Pk−1,k+1 and
U = Dn /Pk−1,k,k+1 , i.e. we have the natural diagram:
q
U = Dn /Pk−1,k,k+1 / M = Dn /Pk−1,k+1 (4.11)
p

X = Dn /Pk

and Mx is Pk−1 × Dn−k /P1 = Pk−1 × Q2(n−k−1) .

e = q −1 (l) ⊂ U for l ∈ M. For the relative tangent bundle TU /X , we have
Lemma 4.7. Let L

TU /X |Le = OLe (−1)⊕2n−k−3 .

Proof. It’s not hard to check that over U, we have the following two exact sequences of vector
bundles:

0 → (Hk /Hk−1 )∨ → p∗ Hk∨ → q ∗ Hk−1

∨
→ 0,

⊥
0 → Hk+1 /Hk → Hk+1 /Hk → q ∗ (Hk+1
⊥
/Hk+1 ) → 0.

We will consider the following diagram

π1 π2
U1 = Dn /Pk−1,k o U = Dn /Pk−1,k,k+1 / U2 = Dn /Pk,k+1 (4.12)
❖❖❖ ♦♦
❖❖❖ ♦♦♦
❖❖❖ ♦♦
❖❖❖ ♦♦♦
❖❖
p1 ❖❖❖
p
♦♦♦p♦2♦
❖❖❖ ♦♦
❖❖❖ ♦♦♦♦♦
❖❖❖  ♦♦♦
' w♦
X.

All the morphisms in the above diagram are projections. For any x ∈ X, the p1 -fiber p−1 1 (x) is
isomorphic to Pk−1 and the p2 -fiber p−1
2 (x) is isomorphic to Q 2(n−k−1)
. Note that for x ∈ X,
the projection q induces an isomorphism

q|p−1 (x) : p−1 (x) → Mx = Pk−1 × Q2(n−k−1) .

Hence p−1 (x) = p1−1 (x) × p2−1 (x). So we get

TU /X = π1∗ (TU1 /X ) ⊕ π2∗ (TU2 /X ).

We mimic the proof of Lemma 5.6 in [6] to conclude that

π1∗ (TU1 /X ) ∼
= Hk /Hk−1 ⊗ q ∗ Hk−1
∨
, π2∗ (TU2 /X ) ∼
= (Hk+1 /Hk )∨ ⊗ q ∗ (Hk+1
⊥
/Hk+1 ).

L

Therefore, TU /X ∼= Hk /Hk−1 ⊗ q ∗ Hk−1∨
(Hk+1 /Hk )∨ ⊗ q ∗ (Hk+1
⊥
/Hk+1 ) and

TU /X |Le = OLe (−1)⊕2n−k−3 .

26
 (2) For k = n−2, the even Orthogonal Grassmannian Dn /Pn−2 := OG(n−2, 2n) parametrizes
the n − 2-dimensional isotropic subspaces in V . In this case M = Dn /Pn−3,n−1,n and U =
Dn /Pn−3,n−2,n−1,n . We have the natural diagram
q
U = Dn /Pn−3,n−2,n−1,n / M = Dn /Pn−3,n−1,n (4.13)
p

X = Dn /Pn−2

and Mx is Pn−3 × P1 × P1 .
e = q −1 (l) ⊂ U for l ∈ M. For the relative tangent bundle TU /X , we have
Lemma 4.8. Let L

TU /X |Le = OLe (−1)⊕n−1 .

Proof. It’s not hard to check that over U, we have the following three exact sequences of vector
bundles:

0 → (Hn−2 /Hn−3 )∨ → p∗ Hn−2

∨
→ q ∗ Hn−3
∨
→ 0,

0 → Hn−1 /Hn−2 → OU⊕n /Hn−2 → q ∗ (OM

⊕n
/Hn−1 ) → 0.

′
0 → Hn−1 /Hn−2 → OU⊕n /Hn−2 → q ∗ (OM
⊕n ′
/Hn−1 ) → 0.

We will consider the following diagram

π1 π2
U1 = Dn /Pn−3,n o U = Dn /Pn−3,n−2,n−1,n / U2 = Dn /Pn−2,n−1,n
PPP ♠♠
PPP
PPP ♠ ♠♠♠♠
PPP ♠♠♠
PPP p ♠♠♠♠
p1 PPP ♠ p
♠
PPP ♠♠♠ 2
PPP ♠♠♠♠♠
PPP  ♠♠♠♠
( v♠
X.
(4.14)

All the morphisms in the above diagram are projections. For any x ∈ X, the p1 -fiber p−1 1 (x) is
isomorphic to Pn−3 and the p2 -fiber p−1
2 (x) is isomorphic to P 1
× P 1
. Note that for x ∈ X, the
projection q induces an isomorphism

q|p−1 (x) : p−1 (x) → Mx = Pn−3 × P1 × P1 .

Hence p−1 (x) = p1−1 (x) × p2−1 (x). So we get

TU /X = π1∗ (TU1 /X ) ⊕ π2∗ (TU2 /X ).

We mimic the proof of Lemma 5.6 in [6] to conclude that

π1∗ (TU1 /X ) ∼
= Hn−2 /Hn−3 ⊗ q ∗ Hn−3
∨

and

π2∗ (TU2 /X ) ∼ ⊕n
= (Hn−1 /Hn−2 )∨ ⊗ q ∗ (OM ′
/Hn−1 ) ⊕ (Hn−1 ⊕n
/Hn−2 )∨ ⊗ q ∗ (OM ′
/Hn−1 ).

27
Therefore,

M
TU /X ∼
= Hn−2 /Hn−3 ⊗ q ∗ Hn−3
∨

⊕n

M ⊕n


(Hn−1 /Hn−2 )∨ ⊗ q ∗ (OM /Hn−1 ) ′
(Hn−1 /Hn−2 )∨ ⊗ q ∗ (OM ′
/Hn−1 )

and TU /X |Le = OLe (−1)⊕n−1 .

(3) For k = n−1 or n, Dn /Pk is one of the two irreducible components of the even Orthogonal
Grassmannian OG(n, 2n). In this case M = Dn /Pn−2 and U = Dn /Pn−2,k . We have the natural
diagram
q
U = Dn /Pn−2,k / M = Dn /Pn−2 (4.15)
p

X = Dn /Pk

and Mx is Grassmannian G(2, n).

e = q −1 (l) ⊂ U for l ∈ M. For the relative tangent bundle TU /X , we have
Lemma 4.9. Let L

TU /X |Le = OLe (−1)⊕2(n−2) .

Proof. It’s not hard to check that over U, we have the following exact sequence of vector bundles
0 → (Hn /Hn−2 )∨ → p∗ Hn∨ → q ∗ Hn−2
∨
→ 0.
Restricting it to the p-fiber p−1 (x) ∼
= G(2, n) is just the Euler sequence on G(2, n). In fact, the
projection p identifies U in a canonical fashion with the Grassmannian bundle G(2, Hn ) of X.
Hence TU /X ∼= Hn /Hn−2 ⊗ q ∗ Hn−2
∨
and

TU /X |Le = OLe (−1)⊕2(n−2) .

4.2 The exceptional simple Lie algebra

We say G is of exceptional type if its Dynkin diagram is of type En (n = 6, 7, 8), F4 or G2 . Due
to the complexity of the geometry of X = G/Pk , it is difficult to write down the specific form
of the relative tangent bundle TU /X , so here we use the method in [17] to calculate the splitting
type of TU /X |q−1 (l) .
When G is simple, the fundamental root system will be written as in the standard reference [5] .
In the case of En (n = 6, 7, 8) type, the mark of the nodes of the Dynkin diagam in the reference
is different from our paper, so is the expression of the fundamental root system. So we write
down the fundamental root system corresponding to the diagram of En (n = 6, 7, 8)-type in our
paper as follow:
Let V be a real vector space with dim V = 8 with orthogonal basis ei , i = 1, . . . , 8. Then the
vectors
1
α1 = − (e1 + · · · + e8 ), α2 = e6 − e7 ,
2
α3 =e6 + e7 , α4 = e5 − e6 ,
α5 =e4 − e5 , α6 = e3 − e4 ,
α7 =e2 − e3 , α8 = e1 − e2

28
form a fundamental root system of type E8 . Since E6 , E7 can be identified canonically with
subsystem of E8 , so α1 , . . . , αn form a fundamental root system of type En (n = 6, 7).
It’s well known that homogeneous vector bundles especially the tangent bundle on a rational
homogeneous space G/P are determined by representations of the Lie algebra p. Restricting
the relative tangent bundle TU /X to a line L e = q −1 (l), we will get a flag bundle H/PN (k) and
D(H) = D\k with tag, where the tag is the set of intersection numbers of a minimal section
of the associated H/PN (k) -flag bundle with the relative canonical divisors (see [22] Proposition
3.17). We refer to [20,22] for a complete account on the tag of a flag bundle. Assume H/PN (k) :=
G1 /PI1 × G2 /PI2 × · · · × Gm /PIm (m ≤ 3) where Gi is simple Lie group. Calculating weights
that give the tangent bundle of Gi /PIi (1 ≤ i ≤ m), which are all the positive roots of the Lie
algebra of Gi that contain the root adjacent to αk . Putting together the above tag and weights,
we will get the splitting type of TU /X |Le . We refer to ( [20] , page 5-7) for a complete account. By
calculation, we have the tables (Table 3, 4, 5, 6) in Appendix.
Professor L. E. Solá Conde told us the following example for calculating the relative tangent
bundle restricting to an isotropic line by personal correspondance.
Example 4.10. Let X = E7 /P7 and U = E7 /P6,7 . Then H/PN (7) = E6 /P6 . Let E7 /PI :=
E7 /P{1,2,3,4,5,6,7} . The isotropic lines in E7 /Pi are the image of the fibers of the P1 -bundle

pi := E7 /PI → E7 /PI\i , (i = 1, 2, . . . 7)

into E7 /P7 via the natural map

π i : E7 /PI → E7 /Pi .
Let −Ki (i = 1, 2, . . . 7) be the relative anticanonical bundles of pi and Ci be a fiber of pi . When
mapped to the varieties of Picard number one E7 /Pi ’s, Ci ’s are the isotropic lines, where 1 ≤
i ≤ 7. The C7 can be regarded as a minimal section of the fibration morphism E7 /P6,7 → E7 /P7
when it restricts to π 7 (C7 ). By [21] Proposition 2.13, the matrix (−Ki Cj ) is the Cartan matrix
of E7 . The tag is exactly
(K1 C7 , ..., K6 C7 ) = (0, 0, 0, 0, 0, 1),
which is the 7-th column of the Cartan matrix without the last term and with the signs changed.
The weights are coefficients of all the positive roots of the Lie algebra of E6 in term of the linear
combinations of the fundamental roots that contain the 6-th root as a summand; that is the
roots of the form m1 α1 + · · · + m6 α6 with mi ≥ 0, m6 > 0, where αi ’s are the fundamental roots
(i = 1, . . . , 6) (See Table 4 in Appendix for the values of mi ). So

TU /X |Le = OLe (−1)⊕16 .

Combining Table 3, 4, 5, 6 we can get the following Proposition.

Proposition 4.11. Suppose G to be of exceptional type and X = G/Pk . Let L e = q −1 (l) ⊂ U
for l ∈ M. When X = F4 /P3 , F4 /P4 , G2 /P1 , TU /X |Le has the following form respectively:

OLe (−2)⊕2 ⊕ OLe (−1), OLe (−2)⊕3 ⊕ OLe (−1)⊕3 , OLe (−3).

Otherwise
TU /X |Le = OLe (−1)N , where N = dim U − dimX.
Summing up, we have obtained the following

29
Proposition 4.12. Let G be a simple Lie group with the Dynkin diagram D and αk is a long
root of D. Denote by X the generalized Grassmannian by marking on D the node k. Then for
the relative tangent bundle TU /X (Notations as Section 4), we have

TU /X |Le = OLe (−1)N , where N = dim U − dimX.

Proof. It is obviously from Lemma 4.1,4.2,4.6,4.7,4.8,4.9 and Proposition 4.11.

4.3 The generalization of the Grauert-Mülich-Barth theorem

The construction of subsheaves in holomorphic vector bundles plays an important role in the proof
of the generalized Grauert-Mülich-Barth theorem. A generalization of the Grauert-Mülich-Barth
theorem to normal projective varieties in characteristic zero is proved (see [13] Theorem 3.1.2).
Since the theorem in the book is for any normal projective variety, the bound is pretty coarse.
In this section, we will find the explicit bound for rational homogeneous spaces.
The following Descent Lemma provides a way for us to prove the existence of subsheaves.
Lemma 4.13. (Descent Lemma [23] ) Let X, Y be nonsingular varieties over k, f : X → Y be a
e ⊂ f ∗E
surjective submersion with connected fibers and E be an algebraic r-bundle over Y . Let K
∗ e ∗ e
be a subbundle of rank t in f E and Q = f E/K be its quotient. If
e Q))
Hom(TX/Y , Hom(K, e = 0,

e is the form K
then K e = f ∗ K for some algebraic subbundle K ⊂ E of rank t.

Follow the previous notations. Let G be a simple Lie group with the Dynkin diagram D and
αk is a root of D. Let’s consider the standard diagram associated to X
q
U /M (4.16)
p

X.
Theorem 4.14. Let X be a generalized Grassmannian by marking on D the node k and αk be a
long root of D. Let E be a holomorphic r-bundle over X of type aE = (a1 , . . . , ar ), a1 ≥ · · · ≥ ar .
If for some t < r,
at − at+1 ≥ 2 f or some t < r,
then there is a normal subsheaf K ⊂ E of rank t with the following properties: over the open set
VE = p(q −1 (UE )) ⊂ X, where UE is an open set in M, the sheaf K is a subbundle of E, which
on the line L ⊂ X given by l ∈ UE has the form

K|L ∼
= ⊕tj=1 OL (aj ).

Proof. The Proposition 4.12 and the Lemma 4.13 play important roles in our proof and our proof
applies almost verbatim the proof of Theorem 2.1.4 given in [23] .
This theorem has far reaching consequences. We give first a series of immediate deductions.
Corollary 4.15. Let X be a generalized Grassmannian with long root αk . For a semistable
r-bundle E over X of type aE = (a1 , . . . , ar ), a1 ≥ · · · ≥ ar . we have

ai − ai+1 ≤ 1 f or i = 1, . . . , r − 1.

30
Proof. If for some t < r, we had at − at+1 ≥ 2, then we could find a normal subsheaf K ⊂ E
which is of the form
K|L ∼= ⊕tj=1 OL (aj )
over the general line L ⊂ X. Then we would have µ(E) < µ(K) contrary to hypothesis.
In particular, we get the generalization theorem of Grauert-Mülich-Barth :
Corollary 4.16. Let X be a generalized Grassmannian with long root αk . The splitting type of
a semistable normalized 2-bundle E over generalized Grassmannian X is

(0, 0) if c1 (E) = 0
aE =
(0, −1) if c1 (E) = −1.

Corollary 4.17. Let X be a generalized Grassmannian with long root αk . For a uniform r-
bundle E (r = ς(X) + 1) over X of type (see Section 3 for the notation ς(X))

aE = (a1 , . . . , ar ), a1 ≥ · · · ≥ ar ,

which does not split, we have

ai − ai+1 ≤ 1 f or i = 1, . . . , r − 1.

Proof. If for some t < r, we had at − at+1 ≥ 2, then we could find a uniform subbundle K ⊂ E
of type which is of aK = (a1 , . . . , at ) (because VE = X). Then quotient bundle Q = E/K would
be uniform of type (as+1 , . . . , ar ). According to Theorem 3.4 the bundle K and Q must be direct
sums of line bundles. The exact sequence

0→K→E→Q→0

would therefore split and hence E would be a direct sum of line bundles contrary to hypothesis.

When αk is a short root of D, M is not the variety of lines on X, but only a closed G-orbit.
Therefore we’re just going to think about the splitting type and semistability of vector bundles
with respect to M.
Theorem 4.18. Let X be a generalized Grassmannian by marking on D the node k and αk be a
short root of D. Let E be a holomorphic r-bundle over X of type aE = (a1 , . . . , ar ), a1 ≥ · · · ≥ ar
with respect to M. If for some t < r,

3, if D 6= G2
at − at+1 ≥
4, if D = G2 ,

then there is a normal subsheaf K ⊂ E of rank t with the following properties: over the open set
VE = p(q −1 (UE )) ⊂ X, where UE is an open set in M, the sheaf K is a subbundle of E, which
on the line L ⊂ X given by l ∈ UE has the form

K|L ∼
= ⊕tj=1 OL (aj ).

Proof. The Lemma 4.13, 4.3, 4.4, 4.5 and Proposition 4.11 play important roles in our proof and
our proof applies almost verbatim the proof of Theorem 2.4 given in [23] .
Similarly, we have the following corollary.

31
Corollary 4.19. Let X be a generalized Grassmannian with short root αk . For a semistable
r-bundle E over X of type aE = (a1 , . . . , ar ), a1 ≥ · · · ≥ ar with respect to M. we have

ai − ai+1 ≤ 3 f or i = 1, . . . , r − 1.

In particular, if D 6= G2 , ai − ai+1 ≤ 2 f or i = 1, . . . , r − 1.
From now, Let
X = G/P ≃ G1 /PI1 × G2 /PI2 × · · · × Gm /PIm ,
where Gi is a simple Lie group with Dynkin diagram Di whose set of nodes is Di and PIi is a
parabolic subgroup of Gi corresponding to Ii ⊂ Di . We set F (Ii ) := Gi /PIi by marking on the
Dynkin diagram Di of Gi the nodes corresponding to Ii . Let δi be a node in Di and N (δi ) be
the set of nodes in Di that are connected to δi .
If δi ∈ Ii , we call
δc δc
Mi i := Gi /Pi i × G\i /PIi (1 ≤ i ≤ m),
c
the i-th special family of lines of class δ̌i , where Pi i := PIi \δi ∪N (δi ) and G\
δ
i /PIi is G1 /PI1 ×
G2 /PI2 × · · · × Gm /PIm by deleting i-th term Gi /PIi . Denote by
δc
Ui i := Gi /PIi ∪N (δi ) × G\
i /PIi

δc
the i-th universal family of class δ̌i , which has a natural P1 -bundle structure over Mi i .
We separate our discussion into two cases:
δc δc
Case I: N (δi ) ⊆ Ii , then Ui i = X and we have the natural projection X → Mi i ;
Case II: N (δi ) * Ii , then we have the standard diagram:

δc q2 c
Ui i / Mδi (4.17)
i

q1

X.
δc
Notice that for x ∈ X, Mxi = q2 (q1 −1 (x)) coincides with H/PN (δi ) where D(H) is the
components of (Di \Ii )\δi containing an element of N (δi ) by Theorem 2.6.
Let G δi be a generalized Grassmannian whose Dynkin diagram Dδi is the maximal sub-
diagram of (Di , Ii ) with the only marked point δi . Let’s consider the standard diagram associated
to G δi
q
U /M (4.18)
p

G δi .
δc
It’s not hard to see that Mxi is isomorphic to My = q(p−1 (y)) for every y ∈ G δi and we have
the following Lemma.
c c
e = q −1 (l) ⊂ U δi for l ∈ Mδi . For the relative tangent bundle T
Lemma 4.20. Let L δc , we
2 i i Ui i /X
have
T δc |e = TU /G δi |Le .
Ui i /X L

32
 If δi is an exposed short root of (Di , Ii ), then δi is a short root of Dδi and vice versa.
Combining this fact and the above Lemma, we get the following Proposition:
c c
Proposition 4.21. Let L e = q −1 (l) ⊂ U δi for l ∈ Mδi . If δi is not an exposed short root, then
2 i i
for the relative tangent bundle T δic , we have
Ui /X

δc
T δc |Le = OLe (−1)N , N = dim Ui i − dimX.
Ui i /X

Proof. It’s obviously from Proposition 4.12.

Similarly, by combining with Lemma 4.3, 4.4, 4.5, 4.20 and Proposition 4.11, we can imme-
diately draw the following proposition.
δc δc
e = q −1 (l) ⊂ U i for l ∈ M i . If δi is an exposed short root, then for the
Proposition 4.22. Let L 2 i i
relative tangent bundle T δic , the splitting type of T δic takes several forms: (−1, . . . , −1, −2),
Ui /X Ui /X
(−2, . . . , −2), (−2, −2, −1), (−2, −2, −2, −1, −1, −1) or (−3).
δc
According to Theorem 2.3, 2.6, if δi is not an exposed short root, then Mi i is the space of
δc
of lines of class δ̌i and Mxi is the space of lines of class δ̌i through x.
Using the above propositions, we have the similar results for rational homogeneous spaces.
Theorem 4.23. Fix δi ∈ Ii and assume that δi is not an exposed short root. Let E be a
(δ ) (δ ) (δ ) (δ ) (δ )
holomorphic r-bundle over X of type aE i = (a1 i , . . . , ar i ), a1 i ≥ · · · ≥ ar i with respect to
c
δ
Mi i . If for some t < r,

(δ ) (δi ) 1, and N (δi ) fits Case I
at i − at+1 ≥
2, and N (δi ) fits Case II,
then there is a normal subsheaf K ⊂ E of rank t with the following properties: over the open set
(δ ) (δ ) c
VE = q1 (q2−1 (UE i )) ⊂ X, where UE i is an open set in Mδi , the sheaf K is a subbundle of E,
(δ )
which on the line L ⊂ X given by l ∈ UE i has the form

K|L ∼
= ⊕ts=1 OL (as(δi ) ).
Proof. The Proposition 4.21 and the Lemma 4.13 play important roles in our proof and our proof
applies almost verbatim the proof of Theorem 5.7 given in [6] .
Corollary 4.24. With the same assumption as Theorem 4.23. For a δi -semistable r-bundle E
(δ ) (δ ) (δ ) (δ ) (δ ) δc
over X of type aE i = (a1 i , . . . , ar i ), a1 i ≥ · · · ≥ ar i with respect to Mi i , we have
(δ )
as(δi ) − as+1
i
≤ 1 for all s = 1, . . . , r − 1.
(δ )
In particular, if N (δi ) fits Case I, then we have as i ’s are constant for all 1 ≤ s ≤ r.
δc
When αj is an exposed short root of (D, I), Mi i is not the space of of lines of class δ̌i , but
only a closed G-orbit.
Theorem 4.25. Fix δi ∈ Ii and assume that δi is an exposed short root. Let E be a holomorphic
(δ ) (δ ) (δ ) (δ ) (δ ) δc
r-bundle over X of type aE i = (a1 i , . . . , ar i ), a1 i ≥ · · · ≥ ar i with respect to Mi i . If for
some t < r, 
(δi ) (δi ) 1, and N (δi ) fits Case I
at − at+1 ≥
4, and N (δi ) fits Case II,

33
then there is a normal subsheaf K ⊂ E of rank t with the following properties: over the open set
(δ ) (δ ) c
VE = q1 (q2−1 (UE i )) ⊂ X, where UE i is an open set in Mδi , the sheaf K is a subbundle of E,
(δ )
which on the line L ⊂ X given by l ∈ UE i has the form

K|L ∼
= ⊕ts=1 OL (as(δi ) ).

Proof. The Proposition 4.22 and the Lemma 4.13 play important roles in our proof and our proof
applies almost verbatim the proof of Theorem 5.7 given in [6] .
(δ ) (δ ) (δ ) (δ )
Corollary 4.26. For a δi -semistable r-bundle E over X of type aE i = (a1 i , . . . , ar i ), a1 i ≥
(δ ) δc
· · · ≥ ar i with respect to Mi i , we have
(δ )
as(δi ) − as+1
i
≤ 3 for all s = 1, . . . , r − 1.
(δ )
In particular, if N (δi ) fits Case I, then we have as i ’s are constant for all 1 ≤ s ≤ r.
From Proposition 3.12, Corollary 4.24 and Corollary 4.26, we can have the following result.
Corollary 4.27. Let X = G/B, where G is a semi-simple Lie group and B is a Borel subgroup
of G. If an r-bundle E is δi -semistable for all i and δi (1 ≤ i ≤ m) over X, then E splits as a
direct sum of line bundles.
δc
Proof. The assumption tells us that E is poly-uniform with respect to Mi i for each i and δi and
the splitting type is (a(δi ) , . . . , a(δi ) ) by Corollary 4.24 and Corollary 4.26. We therefore conclude
by Proposition 3.12.

Acknowledgements
We would like to show our great appreciation to professor B. Fu, L. Manivel, R. Muñoz, G.
Occhetta, L. E. Solá Conde and K. Watanabe for useful discussions. In particular, we would like
to thank professor L. E. Solá Conde for recommending us their interesting papers and explaining
the details in their papers with patience, especially for the method of calculating the relative
tangent bundles between two generalized flag manifolds.

5 Appendix

34
 Table 3: E6 -type
G node k H/PN (k) tag weights
{(12211), (11211), (01211), (11111), (11101),
E6 1 D5 /P5 (00001)
(01111), (01101), (00111), (00101), (00001)}
{(11100), (11110), (11111), (01100), (01110),
E6 2 A5 /P3 (00100)
(01111), (00100), (00110), (00111)}
{(1)} × {(1100), (1110), (1111),
E6 3 A1 /P1 × A4 /P2 (1, 0100)
(0100), (0110), (0111)}
E6 4 A2 /P2 × A1 /P1 × A2 /P1 (01, 1, 10) {(01), (11)} × {(1)} × {(10), (11)}
{(1110), (1111), (0110),
E6 5 A4 /P3 × A1 /P1 (0010, 1)
(0111), (0010), (0011)} × {(1)}
{(12211), (11211), (01211), (11111), (11101),
E6 6 D5 /P5 (00001)
(01111), (01101), (00111), (00101), (00001)}

Table 4: E7 -type
G node k H/PN (k) tag weights
{(122211), (112211), (012211), (111111),
(111101), (011111), (001101), (000111),
E7 1 D6 /P6 (000001)
(000101), (000001), (111211), (011211),
(011101), (001211), (001111)}
{(111000), (111100), (111110), (111111),
E7 2 A6 /P3 (001000) (011000), (011100), (011110), (011111),
(001000), (001100), (001110), (001111)}
{(1)} × {(11000), (11100),
E7 3 A1 /P1 × A5 /P2 (1, 01000) (11110), (11111), (01000),
(01100), (01110), (01111)}
E7 4 A2 /P2 × A1 /P1 × A3 /P1 (01, 1, 100) {(01), (11)} × {(1)} × {(100), (110), (111)}
{(1110), (1111), (0110), (0111)
E7 5 A4 /P3 × A2 /P1 (0010, 10)
(0010), (0011)} × {(10), (11)}
{(12211), (11211), (01211), (11111),
E7 6 D5 /P5 × A1 /P1 (00001, 1) (11101), (01111), (01101), (00111),
(00101), (00001)} × {(1)}
{(000001), (000011), (000111), (010111),
(001111), (101111), (011111), (111111)
E7 7 E6 /P6 (000001)
(011211), (111211), (011221), (112211),
(111211), (112221), (112321), (122321)}

35
 Table 5: E8 -type
G node k H/PN (k) tag weights
{(1222211), (1122211), (1112211),
(1111211), (1111111), (1111101),
(0122211), (0112211), (0111211),
E8 1 D7 /P7 (0000001) (0111111), (0111101), (0012211),
(0011211), (0011111), (0011101),
(0001211), (0001111), (0001101),
(0000111), (0000101), (0000001)}
{(1110000), (1111000), (1111110),
(1111110), (1111111), (0110000),
E8 2 A7 /P3 (0010000) (0111000), (0111100), (0111110),
(0111111), (0010000), (0011000),
(0011100), (0011110), (0011111)}
{(1)} × {(110000), (111000),
(111100), (111110), (111111),
E8 3 A1 /P1 × A6 /P2 (1, 010000)
(010000), (011000), (011100),
(011110), (011111)}
{(01), (11)} × {(1)} × {(1000),
E8 4 A2 /P2 × A1 /P1 × A4 /P1 (01, 1, 1000)
(1100), (1110), (1111)}
{(1110), (1111), (0110), (0111),
E8 5 A4 /P3 × A3 /P1 (0010, 100)
(0010), (0011)} × {(100), (110), (111)}
{(12211), (11211), (01211),
(11111), (11101), (01111),
E8 6 D5 /P5 × A2 /P1 (00001, 10)
(01101), (00111), (00101),
(00001)} × {(10), (11)}
{(000001), (000011), (000111),
(010111), (001111), (101111),
(011111), (111111), (011211),
E8 7 E6 /P6 × A1 /P1 (000001, 1)
(111211), (011221), (112211),
(111211), (112221), (112321),
(122321)} × {(1)}
{(000001), (000011), (000111),
(0101111), (0001111), (0112221),
(0112211), (0112111)(0111111),
(0011111), (1112221), (1112211),
E8 8 E7 /P7 (0000001) (1112111), (1111111), (1011111),
(1224321), (1223321), (1123321),
(1223221), (1123221), (1122221),
(1223211), (1123211), (1122211),
(1122111), (1234321), (2234321)}

36
 Table 6: F4 , G2 -type
G node k H/PN (k) tag weights
F4 1 C3 /P3 (001) {(121), (111), (011), (221), (021), (001)}
F4 2 A1 /P1 × A2 /P1 (1, 10) {(1)} × {(10), (11)}
F4 3 A2 /P2 × A1 /P1 (02, 1) {(01), (11)} × {(1)}
F4 4 B3 /P3 (001) {(122), (112), (012), (111), (011), (001)}
G2 1 A1 /P1 (3) {(1)}
G2 2 A1 /P1 (1) {(1)}

References
[1] M. Andreatta and J. Wiśniewski. On manifolds whose tangent bundle contains an ample
locally free subsheaf. Invent. Math., 146:209–217, 2001.
[2] E. Ballico. Uniform vector bundles on quadrics. Ann. Univ. Ferrara Sez. VII (N.S.),
27(1):135–146, 1982.
[3] E. Ballico. Uniform vector bundles of rank (n + 1) on Pn . Tsukuba J. Math., 7(2):215–226,
1983.
[4] A. Borel and R. Remmert. Über kompakte homogene kählersche mannigfaltigkeiten. Math.
Ann., 145:429–439, 1961/1962.
[5] R. Carter. Lie Algebras of Finite and Affine Type. Cambridge studies in advanced mathe-
matics, 2005.
[6] R. Du, X. Fang, and Y. Gao. Vector bundles on flag varieties. arXiv:1905.10151.
[7] H. Duan and X. Zhao. The chow ring of generalized grassmannians. Foundations of Com-
putational Mathematics, 10(3):245–274, 2005.
[8] D. Eisenbud and J. Harris. 3264 All that-Interesection Theory in Algebraic Geometry. Cam-
bridge University Press, 2011.
[9] G. Elencwajg. Les fibrés uniformes de rang 3 sur P2 (C) sont homogénes. Math.Ann, 231:217–
227, 1978.

[10] G. Elencwajg, A. Hirschowitz, and M. Schneider. Les fibrés uniformes de rang n sur Pn (C)
sont ceux qu’on croit. Progr. Math., 7:37–63, 1980.
[11] P. Ellia. Sur les fibrés uniformes de rang (n + 1) sur Pn . Mém. Soc. Math. France (N.S.),
7:1–60, 1982.
[12] M. Guyot. Caractérisation par l’uniformité des fibrés universels sur la grassmanienne. Math.
Ann., 270(1):47–62, 1985.

[13] D. Huybrechts and M. Lehn. The geometry of moduli spaces of sheaves, second ed. Cambridge
University Press, 2010. xviii+325 pp.
[14] J-M. Hwang. Geometry of minimal rational curves on fano manifolds. School on Vanishing
Theorems and Effective Results in Algebraic Geometry (Trieste, 2000), ICTP Lect. Notes,
vol. 6, Abdus Salam Int. Cent. Theoret. Phys., Trieste, 2001, pages 335–393, 2001.

37
[15] Y. Kachi and E. Sato. Segre’s reflexivity and an inductive characterization of hyperquadrics.
Mem. Amer. Math. Soc., 160(763), 2002.
[16] J.M. Landsberg and L. Manivel. On the projective geometry of rational homogeneous
varieties. Comment. Math. Helv., 78:65–100, 2003.
[17] R. Muñoz, G. Occhetta, and L. E. Solá Conde. On uniform flag bundles on fano manifolds.
arXiv:1610.05930.
[18] R. Muñoz, G. Occhetta, and L. E. Solá Conde. Uniform vector bundles on fano manifolds
and applications. J. Reine Angew. Math. (Crelles Journal), 664:141–162, 2012.
[19] R. Muñoz, G. Occhetta, and L. E. Solá Conde. Splitting conjectures for uniform flag bundles.
European Journal of Mathematics, 6:430–452, 2020.
[20] R. Muñoz, G. Occhetta, L. E. Solá Conde, and K. Watanabe. Rational curves,dynkin
diagrams and fano manifolds with nef tangent bundle. Math.Ann., 361:583–609, 2015.
[21] G. Occhetta, L. E. Solá Conde, K. Watanabe, and J. A. Wiśniewski. Fano manifolds whose
elementary contractions are smooth P1 -fibrations: a geometric characterization of flag vari-
eties. Ann. Sc. Norm. Super. Pica. Cl. Sci.
[22] G. Occhetta, L. E. Solá Conde, and J. A. Wiśniewski. Flag bundles on fano manifolds.
Journal de Mathmatiques Pures et Appliques, 106(4):651–669, 2016.
[23] C. Okonek, M. Schneider, and H. Spindler. Vector bundles on complex projective spaces.
Birkhäuser/Springer Basel AG, Basel, 2011. viii+239 pp.
[24] E. Sato. Uniform vector bundles on a projective space. J. Math. Soc. Japan, 28(1):123–132,
1976.
[25] R. L. E. Schwarzenberger. Vector bundles on the projective plane. Proc. London Math.
Soc., 11(3):623–640, 1961.
[26] H. Tamvakis. Quantum cohomology of isotropic grassmannians. Geometric Methods in
Algebra and Number Theory, 235:311–338, 2006.
[27] A. Van de Ven. On uniform vector bundles. Math. Ann., 195:245–248, 1972.

38