BLOWING-UP HERMITIAN YANG–MILLS CONNECTIONS

arXiv:2301.00525v2 [math.DG] 6 Nov 2023

ANDREW CLARKE AND CARL TIPLER

Abstract. We investigate hermitian Yang–Mills connections for pullback vec-

tor bundles on blow-ups of Kähler manifolds along submanifolds. Under some
technical asumptions on the graded object of a simple and semi-stable vec-
tor bundle, we provide a necessary and sufficent numerical criterion for the
pullback bundle to admit a sequence of hermitian Yang–Mills connections for
polarisations that make the exceptional divisor sufficiently small, and show
that those connections converge to the pulled back hermitian Yang-Mills con-
nection of the graded object.

1. Introduction
A cornerstone in gauge theory is the Hitchin–Kobayashi correspondence ([9,
12, 20, 7]). This celebrated generalisation of the Narasimhan and Seshadri theo-
rem asserts that a holomorphic vector bundle over a Kähler manifold carries an
Hermite–Einstein metric if and only if it is polystable in the sense of Mumford
and Takemoto ([14, 19]). The interplay between the differential geometric side, in
the form of the hermitian Yang–Mills connections (HYM for short) that originated
from physics, and the algebro-geometric side, that of stability notions coming from
moduli constructions, has had many applications and become a very fertile source
of inspiration. Given that HYM connections are canonically attached to polystable
vector bundles, it is natural to investigate their relations to natural maps between
vector bundles, such as pullbacks. In this paper, we address the problem of pulling
back HYM connections along blow-ups. While the similar problem for extremal
Kähler metrics has seen many developments in the past ten years [1, 2, 3, 18, 17, 5],
relatively little seems to be known about the behaviour of HYM connections under
blow-ups [4, 6]. In this paper, under some mild asumptions, we solve the problem
for pullback of semi-stable vector bundles on blow-ups along smooth centers.
Let π : X ′ → X be the blow-up of a polarised Kähler manifold (X, [ω]) along a
submanifold Z ⊂ X, and E ′ = π ∗ E the pullback of a holomorphic vector bundle
E → X. For 0 < ε ≪ 1, Lε := π ∗ [ω] − εc1 (Z ′ ) defines a polarisation on X ′ ,
where we denote by Z ′ = π −1 (Z) the exceptional divisor. There are obstructions
for E ′ to admit HYM connections with respect to ωε ∈ c1 (Lε ), with 0 < ε ≪ 1.
In particular, E should be simple and semi-stable with respect to [ω]. In the latter
case, E admits a Jordan–Holder filtration by semi-stable sheaves with polystable
graded object Gr(E) (see Section 2.2 for definitions). A further obstruction comes
then from subsheaves of E arising from Gr(E). While those sheaves have the same
slope as E, their pullbacks to X ′ could destabilise E ′ . Our main result asserts that,
under some technical asumptions on Gr(E) and on the dimension of Z, these are
actually the only obstructions for E ′ to carry a HYM connection. More precisely,

2010 Mathematics Subject Classification. Primary: 53C07, Secondary: 53C55, 14J60.

1
2 A. CLARKE AND C. TIPLER

from now on, and until the end of the paper, we will assume that Z ⊂ X satisfies
codim(Z) ≥ 3.
Recall that a semi-stable holomorphic vector bundle E → (X, [ω]) is said to be
sufficiently smooth if its graded object Gr(E) is locally free. We will assume further
that the Jordan–Hölder filtration with locally-free subsheaves
0 = F0 ⊂ F1 ⊂ . . . ⊂ Fℓ = E
c (π ∗ F )·Ln−1
is unique. For 1 ≤ i ≤ ℓ, denote by µLε (Fi ) = 1 rank(Fi ε
i)
the slope of π ∗ Fi on
′
(X , Lε ). In the following statements, when referring to a HYM connection A on E
(resp. on Gr(E), or π ∗ E), we will implicitly assume that A0,1 is gauge equivalent
to the holomorphic connection, or Dolbeault operator, of E (resp. of Gr(E), or
π ∗ E).
Theorem 1.1. Let E → X be a sufficiently smooth semi-stable holomorphic vector
bundle on (X, [ω]). Assume that the stable components of Gr(E) are pairwise non-
isomorphic and that E admits a unique Jordan–Holder filtration by locally free sub-
sheaves. Then, there exists ε0 > 0 and a sequence of HYM connections (Aε )ε∈(0,ε0 )
on π ∗ E with respect to (ωε )ε∈(0,ε0 ) if and only if
(1.1) ∀ i ∈ [[1, ℓ − 1]], µLε (Fi ) < µLε (E).
ε→0

In that case, if A denotes a HYM connection on Gr(E) with respect to ω, then

(Aε )ε∈(0,ε0 ) can be chosen so that Aε −→ π ∗ A in any Sobolev norm.
ε→0

The expression µLε (F ) < µLε (E) means that the first non-zero term in the
ε→0
ε-expansion for µLε (E)−µLε (F ) is strictly positive. The conclusion of Theorem 1.1
can be rephrased as follows. If ∂ Gr(E) (resp. ∂ E ) stands for the Dolbeault operator
of Gr(E) (resp. of E) and if h is a Hermite–Einstein metric on Gr(E) with respect
to ω, then there exist gauge transformations (fε )0<ε<ε0 on X ′ such that the Chern
connections of (fε∗ π ∗ ∂ E , π ∗ h) are HYM with respect to (ωε )0<ε<ε0 and converge to
the Chern connection of (π ∗ ∂ Gr(E) , π ∗ h) in any Sobolev norm.
Remark 1.2. The simplicity of E is implied by the hypotheses made on Gr(E) (see
[16, Lemma 34]). Semi-stability and condition (1.1) are also necessary to produce
the connections (Aε ) from Theorem 1.1. The other three asumptions on Gr(E) are
technical. Assuming Gr(E) to be locally free enables one to regard E as a smooth
complex deformation of Gr(E) and to work with the various connections on the same
underlying complex vector bundle. We should warn the reader though that if one
drops this asumption, Condition (1.1) may not be enough to ensure semi-stability
of π ∗ E on (X ′ , Lε ) (see the extra conditions in [15, Theorem 1.10]). On the other
hand, the asumption on Gr(E) having no pairwise isomorphic components is purely
technical, and ensures that its automorphism group, which will provide obstructions
in the perturbative theory, is abelian. Finally, the fact that the Jordan–Holder
filtration is unique provides a special shape in the extension matrix from Gr(E) to
E, with non-zero terms above the diagonal. This is used to produce approximate
solutions by induction on the number of stable components of Gr(E).
We now list some corollaries of Theorem 1.1. First, the stable case :
 BLOWING-UP HYM CONNECTIONS 3

Corollary 1.3. Let E → X be a stable holomorphic vector bundle on (X, [ω]) and
let A be a HYM connection on E with respect to ω. Then, there exists ε0 > 0 and a
sequence of HYM connections (Aε )ε∈(0,ε0 ) on π ∗ E with respect to (ωε )ε∈(0,ε0 ) such
that Aε −→ π ∗ A in any Sobolev norm.
ε→0

For the semi-stable case, Condition (1.1) reduces to a finite number of intersec-
tion product computations. One interesting feature comes from the second term in
the expansion of µLε (E). It is the opposite of the slope of the restriction of E to
Z. The following formula is proved in [15, Section 4.1], where m = dim(Z) :
 
n−1
(1.2) µLε (E) = µL (E) − µL|Z (E|Z )εn−m + O(εn−m+1 ).
m−1
We then have :
Corollary 1.4. Let E → X be a sufficiently smooth semi-stable holomorphic vec-
tor bundle on (X, [ω]). Assume that the stable components of Gr(E) are pairwise
non-isomorphic and that the Jordan–Holder filtration with locally free subsheaves is
unique. Denote by A an HYM connection on Gr(E) with respect to ω. If
(1.3) ∀ i ∈ [[1, ℓ − 1]], µL|Z (E|Z ) < µL|Z ((Fi )|Z ),
then, there exists ε0 > 0 and a sequence of HYM connections (Aε )ε∈(0,ε0 ) on π ∗ E
with respect to (ωε )ε∈(0,ε0 ) converging to π ∗ A in any Sobolev norm.
Condition (1.3) was checked on explicit examples in [15, Section 4.5] to pro-
duce stable perturbations of tangent sheaves by blow-ups, and our result provides
information on the associated connections and their asymptotic behaviour. Note
that by the Mehta–Ramanathan theorem [13], if [ω] = c1 (L) is integral, and if Z
is a generic intersection of divisors in linear systems |Lk |, then E|Z is semi-stable
as soon as E is. In that case, Condition (1.3) cannot be satisfied, and it seems
unlikely that Condition (1.1) will hold true. Hence, blowing-up such subvarieties
tend to destabilise a semi-stable bundle.
In general, we expect that it should not be too hard to obtain stability of suffi-
ciently smooth pulled back bundles under condition (1.1) with purely algebraic
methods. However, we emphasize that the Hitchin–Kobayashi correspondence
doesn’t provide any information on the asymptotic behaviour of the associated
HYM connections, which is then the main content of Theorem 1.1. Nevertheless, we
state the following corollary, that extends [15, Theorem 1.10] to a non-equivariant
situation:
Corollary 1.5. Let E → X be a sufficiently smooth semi-stable holomorphic vec-
tor bundle on (X, [ω]). Assume that the stable components of Gr(E) are pairwise
non-isomorphic and that the Jordan–Holder filtration with locally free subsheaves is
unique. Then, there exists ε0 > 0 such that π ∗ E → (X ′ , Lε ) is
(i) stable if and only if for all i ∈ [[1, ℓ − 1]], µLε (Fi ) < µLε (E),
ε→0
(ii) semi-stable if and only if for all i ∈ [[1, ℓ − 1]], µLε (Fi ) ≤ µLε (E),
ε→0
(iii) unstable otherwise.
Finally, we comment on previous related works. Theorem 1.1 extends results
from [4, 6] where blow-ups of HYM connections along points are considered. In the
present paper, we consider blow-ups along any smooth subvariety, and also cover
4 A. CLARKE AND C. TIPLER

the semi-stable situation, which is technically more involved due to the presence
of automorphisms of the graded object that obstruct the linear theory. While [6]
is a gluing construction of a similar vein to the various solutions to the analogous
problem of producing extremal Kähler metrics on blow-ups [2, 3, 18, 17, 5], one
of the key feature in our approach is to directly apply the quantitative implicit
function theorem, closely following the method developed in [16]. The main new
technical input is in Section 3, where we perform a precise study of the linear
operators involved in relation to the geometry of the blow-up. The rest of the
argument, in Sections 4 and 5, is more standard, and follows the work of Sektnan
and the second author [16].

Outline: In Section 2, we recall basic material about HYM connections and stabil-
ity. We then perform in Section 3 the analysis of the linear theory on the blow-up.
Relying on this, in Section 4 we explain how to produce approximate solutions to
the HYM equations. This section, and the one that follows, rely on the treatment
in [16]. Then, we perturb those approximate solutions in Section 5 to actual solu-
tions, which concludes the proof of Theorem 1.1. The corollaries are addressed in
Section 5.3.

Acknowledgments: The authors benefited from visits to LMBA and Gotheborg

University; they would like to thank these welcoming institutions for providing
stimulating work environments. The idea of this project emerged from discussions
with Lars Martin Sektnan, whom we thank for sharing his ideas and insight. AC
is partially supported by the grants BRIDGES ANR–FAPESP ANR-21-CE40-0017
and Projeto CAPES - PrInt UFRJ 88887.311615/2018-00. CT is partially sup-
ported by the grants MARGE ANR-21-CE40-0011 and BRIDGES ANR–FAPESP
ANR-21-CE40-0017.

2. Preliminaries
In Sections 2.1 and 2.2 we introduce the notions of HYM connections and slope
stability, together with some general results, and refer the reader to [10] and [8].
From Section 3 we start to specialise the discussion to blow-ups. In particular, in
Section 3.2, we provide various asymptotic expressions for the linearisation of the
HYM equation on the blow-up. Those results will be used in Section 5.

2.1. The hermitian Yang–Mills equation. Let E → X be a holomorphic vector

bundle over a compact Kähler manifold X. A hermitian metric on E is Hermite–
Einstein with respect to a Kähler metric with Kähler form ω if the curvature
Fh ∈ Ω2 (X, End E) of the corresponding Chern connection satisfies
(2.1) Λω (iFh ) = c IdE
for some real constant c. Equivalently, if h is some hermitian metric on the smooth
complex vector bundle underlying E, a hermitian connection A on (E, h) is said to
be hermitian Yang–Mills if it satisfies
FA0,2

= 0,
Λω (iFA ) = c IdE .
The first equation of this system implies that the (0, 1)-part of A determines a
holomorphic structure on E, while the second that h is Hermite–Einstein for this
 BLOWING-UP HYM CONNECTIONS 5

holomorphic structure. We will try to find hermitian Yang–Mills connections within

the complex gauge group orbit, which we now define. The complex gauge group is
G C (E) = Γ (GL (E, C)) .
We note that the set of unitary gauge transformations preserves the metric h so,
taking account of the fibre-wise polar decomposition of an element of G C (E), we
consider the bundle EndH (E, h) of Hermitian endomorphisms of E and the set
G C (E, h) := G C (E) ∩ Γ(EndH (E, h)) = {es : s ∈ Γ(EndH (E, h))}.
If ∂¯ is the Dolbeault operator defining the holomorphic structure on E, then f ◦
∂¯ ◦ f −1 defines a biholomorphic complex structure on E. Let dA = ∂A + ∂¯A be the
Chern connection of (E, h) with respect to the original complex structure (that is
∂¯A = ∂).
¯ Then the Chern connection Af of h with respect to f ◦ ∂¯ ◦ f −1 is

dAf = (f ∗ )−1 ◦ ∂A ◦ (f ∗ ) + f ◦ ∂¯ ◦ f −1 .
Solving the hermitian Yang–Mills equation is equivalent to solving
Ψ(s) = c IdE
where
Ψ : Γ(EndH (E, h)) −→ Γ(EndH (E, h)),
s 7−→ iΛω (FAexp(s) ),
and where Γ(EndH (E, h)) is the tangent space to G C (E, h) at the identity. For a
connection A on E, the Laplace operator ∆A is
∆A = iΛω ∂¯A ∂A − ∂A ∂¯A .


(2.2)
If AEnd E denote the connection induced by A on End E, then :
Lemma 2.1. If A is the Chern connection of (E, ∂, h), the differential of Ψ at
identity is
dΨIdE = ∆AEnd E .
If moreover A is assumed to be hermitian Yang–Mills, then the kernel of ∆AEnd E
acting on Γ(End(E)) is given by the Lie algebra aut(E) of the space of automor-
phisms Aut(E) of (E, ∂).
The last statement about the kernel follows from the Kähler identities and the
Akizuki-Nakano identity that imply ∆AEnd E = ∂A ∗
∂A + ∂¯A
∗ ¯
∂A , the two terms of
which are equal if A is Hermitian Yang-Mills. The operator ∆AEnd E being elliptic
and self-adjoint, aut(E) will then appear as a cokernel in the linear theory for
perturbations of hermitian Yang–Mills connections.

2.2. Slope stability. We recall some basic facts about slope stability, as intro-
duced by [14, 19], and refer the interested reader to [8] for a detailed treatment.
We denote here L := [ω] the polarisation of the n-dimensional Kähler manifold X.
Definition 2.2. For E a torsion-free coherent sheaf on X, the slope µL (E) ∈ Q
(with respect to L) is given by the intersection formula
degL (E)
(2.3) µL (E) = ,
rank(E)
6 A. CLARKE AND C. TIPLER

where rank(E) denotes the rank of E while degL (E) = c1 (E) · Ln−1 stands for its
degree. Then, E is said to be slope semi-stable (resp. slope stable) with respect to
L if for any coherent subsheaf F of E with 0 < rank(F ) < rank(E), one has
µL (F ) ≤ µL (E) ( resp. µL (F ) < µL (E)).
A direct sum of slope stable sheaves of the same slope is said to be slope polystable.
In this paper, we will often omit “slope” and simply refer to stability of a sheaf,
the polarisation being implicit. We will make the standard identification of a holo-
morphic vector bundle E with its sheaf of sections, and thus talk about slope sta-
bility notions for vector bundles as well. In that case slope stability relates nicely
to differential geometry via the Hitchin–Kobayashi correspondence :
Theorem 2.3 ([9, 12, 20, 7]). There exists a Hermite–Einstein metric on E with
respect to ω if and only if E is polystable with respect to L
We will be mostly interested in semi-stable vector bundles. A Jordan–Hölder
filtration for a torsion-free sheaf E is a filtration by coherent saturated subsheaves:
(2.4) 0 = F0 ⊂ F1 ⊂ . . . ⊂ Fℓ = E,
such that the corresponding quotients,
Fi
(2.5) Gi = ,
Fi−1
for i = 1, . . . , ℓ, are stable with slope µL (Gi ) = µL (E). In particular, the graded
object of this filtration
l
M
(2.6) Gr(E) := Gi
i=1

is polystable. From [8, Section 1], we have the standard existence and uniqueness
result:
Proposition 2.4. Any semi-stable coherent torsion-free sheaf E on (X, L) admits
a Jordan–Hölder filtration, and the double dual of the graded object Gr(E)∗∗ of such
filtrations is unique up to isomorphism.
When E is locally-free and semi-stable, we say that it is sufficiently smooth if
Gr(E) is locally-free. In that case, we denote E[ω] the set of holomorphic subbundles
of E built out of successive extensions of some of the stable components of Gr(E).
Equivalently, E[ω] is the set of holomorphic subbundles of E arising in a Jordan-
Holder filtration for E. Finally, we recall that a necessary condition for E to be
stable is simplicity, that is Aut(E) = C∗ · IdE .

3. Geometry of the blow-up

We consider now Z ⊂ X a m-dimensional complex submanifold of codimension
r = n − m ≥ 2 and the blow-up map
π : BlZ (X) → X.
We will denote by X = BlZ (X) the blown-up manifold and by Z ′ = π −1 (Z) the
′

exceptional divisor. We denote by

Lε := π ∗ L − ε[Z ′ ]
 BLOWING-UP HYM CONNECTIONS 7

a polarisation on X ′ , for 0 < ε ≪ 1. Let E → X be a holomorphic vector bundle,

and denote by E ′ = π ∗ E the pulled back bundle. For any holomorphic subbundle
F ⊂ E, the intersection numbers µLε (π ∗ E)− µLε (π ∗ F ) admit expansions in ε, with
first term given by µL (E) − µL (F ). For that reason, given the Hitchin–Kobayashi
correspondence in Theorem 2.3, semi-stability of E on (X, L) is a necessary con-
dition for its pullback E ′ to admit an HYM connection with respect to a Kähler
metric in Lε , for all 0 < ε ≪ 1. Another necessary condition is simplicity of E ′ ,
which, by Hartogs’ theorem, is equivalent to simplicity of E. Then, natural can-
didates to test for stability of E ′ are given by the pullbacks of elements in E[ω] ,
and Condition (1.1) clearly is necessary for E ′ to be stable in the polarisations
we consider, and thus to admit an HYM connection. Hence, we will assume E to
be simple, semi-stable, and to satisfy (1.1). We now turn back to the differential
geometry of the blow-up.

3.1. Decomposition on spaces of sections. We have a commutative diagramm:

ι
Z′ −→ X ′
↓ ↓
ι0
Z −→ X
where ι0 and ι denote the inclusions, while the vertical arrows are given by the
projection map π. We then have a pullback map on sections
π ∗ : Γ(X, End(E)) −→ Γ(X ′ , End(π ∗ E))
as well as a restriction map :
ι∗ : Γ(X ′ , End(π ∗ E)) −→ Γ(Z ′ , End(ι∗ π ∗ E)).
Our goal now is to fit those maps in a short exact sequence, that will in the end split
the space Γ(X ′ , End(π ∗ E)). If NZ = T X|Z /T Z denotes the normal bundle of Z in
X, then Z ′ ≃ P(NZ ), and we can fix a (1, 1)-form λ ∈ c1 (OP(NZ ) (1)) that restricts
to Kähler metrics on the fibers of P(NZ ) → Z. We also fix a Kähler form ω ∈ c1 (L)
on X, and consider its restriction to Z. We then have a Kähler CPr−1 -fibration :
π : (Z ′ , λ) −→ (Z, ω).
By averaging along fibers as described in [16, Section 2.3], we obtain a splitting
(3.1) Γ(Z ′ , End(ι∗ π ∗ E)) = π ∗ (Γ(Z, End(ι∗0 E))) ⊕ Γ0 (Z ′ , End(ι∗ π ∗ E)).
We will omit the ι∗ and π ∗ to simplify notation. Using the projection on the second
factor
p0 : Γ(Z ′ , End(E)) → Γ0 (Z ′ , End(E))
in (3.1), we deduce a short exact sequence :
π∗ p0 ◦ι∗
0 −→ Γ(X, End(E)) −→ Γ(X ′ , End(E)) −→ Γ0 (Z ′ , End(E)) −→ 0.
We can actually split this sequence by mean of a linear extension operator
ι∗ : Γ0 (Z ′ , End(E)) −→ Γ(X ′ , End(E))
such that
p0 ◦ ι∗ ◦ ι∗ = Id.
8 A. CLARKE AND C. TIPLER

This can be done using bump functions and a standard partition of unity argument.
The outcome is an isomorphism :
Γ(X ′ , End(E)) −→ Γ(X, End(E)) ⊕ Γ0 (Z ′ , End(E))
(3.2)
s 7−→ (s − ι∗ ◦ p0 ◦ ι∗ s , p0 ◦ ι∗ s),
with inverse map (sX , sZ ) 7→ (π ∗ sX + ι∗ sZ ). This splits the Lie algebra of gauge
transformations, and will be used to identify contributions coming from X and from
Z ′ in the ε-expansion of the linearisation, which we describe in the next section.
From now on, by abuse of notations, we will consider the spaces Γ(X, End(E))
and Γ0 (Z ′ , End(E)) as subspaces of Γ(X ′ , End(π ∗ E)), and denote s = sX + sZ the
decomposition of an element s ∈ Γ(X ′ , End(E)).

3.2. Decomposition of the Laplace operator. We extend λ to a closed (1, 1)-

form over X ′ as in [21, Section 3.3] and consider the family of Kähler metrics on
X ′:
ωε = π ∗ ω + ελ ∈ c1 (Lε ), 0 < ε ≪ 1.
Let A be a Hermitian connection on E, which we pull back to X ′ and extend to
the bundle End(π ∗ E). We will now study the Laplace operator
∆ε s = iΛε (∂¯A ∂A − ∂A ∂¯A )s
acting on the various components of s = sX + sZ ∈ Γ(X ′ , End(E)), where Λε is the
Lefschetz operator for the metric ωε . For this, we need to introduce an operator on
Z ′.
Definition 3.1. The vertical Laplace operator, denoted
∆V : Γ0 (Z ′ , End(E)) → Γ0 (Z ′ , End(E)) ,
is the operator
∆V = iΛV ∂F ∂¯F − ∂¯F ∂F ,

where
ΛV : Ω1,1 (Z ′ , End(E)) → Γ0 (Z ′ , End(E))
is the vertical contraction operator defined on each fiber Fz = π −1 (z) of π : Z ′ → Z
by
(ΛV α)|Fz λr−1 r−2
|Fz = (r − 1)α|Fz ∧ λ|Fz

and where dF = ∂F + ∂¯F is the flat connection along the fibres of π : Z ′ → Z.

The following lemma relies on [11] and was already observed in [16, Lemma 4.6].
Lemma 3.2. The vertical Laplacian
∆V : Γ0 (Z ′ , End(E)) → Γ0 (Z ′ , End(E))
is an invertible linear map.
In the following statements, an expression of the form O(εj ) is to be understood
as holding pointwise. Convergence considerations of those expressions with respect
to various Sobolev space norms will be addressed in Section 5.
Lemma 3.3. If sZ = ι∗ σZ for σZ ∈ Γ(Z ′ , End(E)), then
(p0 ◦ ι∗ )∆ε (ι∗ σZ ) = ε−1 ∆V σZ + O(1).
 BLOWING-UP HYM CONNECTIONS 9

Proof. We introduce the operator D given by

DsZ = i(∂¯A ∂A − ∂A ∂¯A )sZ .
The Laplacian ∆ε satisfies on X ′ :
∆ε sZ ωεn = nDsZ ∧ ωεn−1 ,
or equivalently
n DsZ ∧ (ω + ελ)n−1
∆ε sZ = .
(ω + ελ)n
We note that ω is a Kähler form on X, but on X ′ is degenerate along the fibre
directions of the submanifold Z ′ . Then (i∗ ω)m+1 = 0 ∈ Ω2(m+1) (Z ′ ), and at
x ∈ Z ′ ⊆ X ′ , ω m+2 = 0. Then, expanding (ω + ελ)n−1 and (ω + ελ)n gives
DsZ ∧ ω m+1 ∧ λn−m−2
ι∗ ∆ε sZ = (n − m − 1)ε−1 + O(1).
ω m+1 ∧ λn−m−1
¯ Z , acting
Restricting to Z ′ , the connection 1-forms of A vanish, so ι∗ DsZ = i∂ ∂σ
on the coefficient functions of σZ . On the other hand, by considering a convenient
orthonormal frame at x ∈ Z ′ , we see that ι∗ ∆ε ι∗ σZ = ε−1 ∆V σZ + O(1). 
In the next lemma, we denote ∆ε sZ = (∆ε sZ )X + (∆ε sZ )Z the decomposition
according to (3.2).
Lemma 3.4. For sZ = ι∗ σZ with σZ ∈ Γ(Z ′ , End(E)), we have
(∆ε sZ )X = O(1).
Proof. By definition, (∆ε sZ )X = π ∗ φ for some φ ∈ Γ(X, End(E)). As we also have
(∆ε sZ )X = (Id − ι∗ (p0 ◦ ι∗ ))Λε DsZ ,
we deduce that the section φ is the continuous extension of π∗ (Id−ι∗ (p0 ◦ι∗ ))Λε DsZ
across Z ⊆ X. On X ′ \ Z ′ we have
DsZ ∧ (ω n−1 + O(ε))
Λε DsZ = n = O(1).
ω n + O(ε)
As π∗ (Id − ι∗ (p0 ◦ ι∗ )) is O(1), the result follows. 
From the previous two lemmas, in the decomposition
s = sX + sZ ,
∆ε sZ also lies in the subspace Γ0 (Z ′ , End(E)) ⊆ Γ(X ′ , End(E)) to higher order in
ε. For sX ∈ Γ(X, End(E)),
∆ε sX = (∆ε sX )X + (∆ε sX )Z
where (∆ε sX )Z = ι∗ (p0 ◦ ι∗ )∆ε sX . We first consider ι∗ ∆ε sX .
Lemma 3.5. For sX = π ∗ σX ∈ Γ(X, End(E)) ⊆ Γ(X ′ , End(E)),
DsX ∧ ω m ∧ λn−m−1
ι∗ ∆ε sX = (m + 1) + O(ε).
ω m+1 ∧ λn−m−1
Proof. Firstly, sX = π ∗ σX , and the connection A is pulled back from X, so DsX
is basic for the projection to X and Ds ∧ ω m+1 = 0 at points in Z ′ . Secondly, we
note that ω m+1 ∧ λn−m−1 is a volume form on X ′ , in a neighbourhood of Z ′ . Then,
the result follows similarly to the previous lemma. 
10 A. CLARKE AND C. TIPLER

For the final term (∆ε sX )X , we introduce ∆X the Laplace operator of A on

End(E) → (X, ω):
∆X : Γ (X, End(E)) → Γ (X, End(E))
σ 7→ iΛω (∂¯A ∂A − ∂A ∂¯A )σ.
Lemma 3.6. For sX = π ∗ σX ∈ Γ(X, End(E)) ⊆ Γ(X ′ , End(E)),
(∆ε sX )X = π ∗ (∆X σX ) + O(ε).
Proof. There is φ ∈ Γ(X, End(E)) such that (∆ε sX )X = π ∗ φ. The element φ can be
identified as the lowest order term in the asymptotic expansion in ε of (∆ε π ∗ σX )X .
However, we have at x ∈ X ′ \ Z ′ :
Dπ ∗ σX ∧ (ω + ελ)n−1 ∗ DσX ∧ ω
n−1
∆ε π ∗ σX = n = nπ + O(ε)
(ω + ελ)n ωn
so we see that the lowest order term in the expansion of (∆ε π ∗ σX )X is ∆X σX . 

Summarizing the above calculations, with respect to the decomposition s =

sX + sZ produced by (3.2), the operator ∆ε takes the form
 
∆X 0
(3.3)
L ε−1 ∆V
plus higher order terms, for some second order operator L.

4. The approximate solutions

The goal of this section is to produce approximate solutions to the HYM equa-
tions. As in last section, expressions of the form O(εj ) are to be understood as
holding pointwise until Section 5. As from now on the methods and proofs fol-
low very closely [16]. We will only sketch most of the steps, giving details when
major differences occur and precisely quoting results in [16] otherwise. Note that
in our work, the parameter ε plays the role of k −1 in [16], and in the decompo-
sition Γ(X ′ , End E) = Γ(X, End E) ⊕ Γ0 (Z ′ , End E), the space Γ(X, End E) (resp.
Γ0 (Z ′ , End E)) plays the role of Γ(B, End(E)) (resp. of Γ0 (X, End E)) in [16].

4.1. Fixing the complex deformation parameters. We start from a semi-

stable and sufficiently smooth holomorphic vector bundle E on (X, L), with L = [ω].
Lℓ
Denote by Gr(E) = i=1 Gi the associated polystable graded object, with stable
components Gi . We also assume that Gr(E) has non-isomorphic stable quotients,
and that the Jordan–Holder filtration with locally-free quotients is unique. We now
explain the technical implications of these hypotheses. First, by [16, Lemma 5.8],
E is simple. The automorphism group G := Aut(Gr(E)) is a reductive Lie group
with Lie algebra g := aut(Gr(E)) and compact form K ⊂ G, with k := Lie(K). We
have M
g= H 0 (X, Hom(Gi , Gj ).
i,j

As the Gi ’s are all stable, and non isomorphic, we deduce that

ℓ
M
g= C · IdGi ,
i=1
 BLOWING-UP HYM CONNECTIONS 11

and in particular g is abelian. The upshot for us is that elements in k that will
obstruct the linear theory all live on the diagonal in the matrix block decomposi-
tion induced by the decomposition Gr(E) = ⊕ℓi=1 Gi . We let ∂ 0 be the Dolbeault
operator of Gr(E). The Dolbeault operator ∂ E on E is given by
∂E = ∂0 + γ
where γ ∈ Ω0,1 (X, Gr(E)∗ ⊗ Gr(E)) can be written
X
γ= γij
i<j
0,1
for (possibly vanishing) γij ∈ Ω (X, Gj∗
⊗ Gi ). From the uniqueness hypothesis on
the Jordan–Hölder filtration, we deduce that the terms γi,i+1 , for 1 ≤ i ≤ ℓ − 1,
are all non-zero, see [16, Lemma 5.4]. This will enable us to use induction on ℓ, the
number of stable components, in the construction of the approximate solutions.
Our starting point to produce HYM connections on π ∗ E → (X ′ , ωε ) will be a
product Hermite–Einstein metric h = h1 ⊕ . . . ⊕ hℓ on Gr(E) (we can fix one by
polystability), and the associated HYM connection A0 on (Gr(E), ∂ 0 ) with con-
tracted curvature form
Λω iFA0 = c0 · Id .
2
We then introduce the L projection
(4.1)
Πik : Γ(X, EndH (E, h)) → ik
ℓ
1 1
X Z
s 7→ ( traceGi (s) ω n ) IdGi
Vol(ω) i=1 rank(Gi ) X
and the induced orthogonal decomposition:
Γ(X, EndH (E, h)) = ik ⊕ Γik⊥ (X, EndH (E, h))
with respect to the pairing
Z
(s1 , s2 ) 7→ trace(s1 · s2 ) ω n .
X

Denoting by ∆Gr(E),0 the Laplace operator of the Chern connection of (∂ 0 , h) with

respect to ω, from Lemma 2.1 and self-adjointness, we have
Lemma 4.1. The following operator is invertible :
∆Gr(E),0 : Γik⊥ (X, EndH (E, h)) → Γik⊥ (X, EndH (E, h)).
We will now make use of the gauge action of G to calibrate the complex de-
formation parameter γ with the metric variation parameter ε. We gauge fix γ by
imposing
∗
(4.2) ∂ 0 γ = 0,
where the adjoint is with respect to the Kähler structure ω. Elements
g := g1 IdG1 + . . . , +gℓ IdGℓ ∈ G,
∗ ℓ
for (gi ) ∈ (C ) , act on ∂ E and produce isomorphic holomorphic vector bundles in
the following way :
X
(4.3) g · ∂E = ∂0 + gi gj−1 γij .
i<j
12 A. CLARKE AND C. TIPLER

We assume from now on that for 1 ≤ i ≤ ℓ − 1,

µLε (Fi ) < µLε (E).
ε→0

We introduce qi to be the discrepancy order of Fi for each i ∈ [[1, ℓ − 1]], that is

the order qi ∈ N such that µLε (E) − µLε (Fi ) = νi εqi + O(εqi +1 ) for a constant
νi > 0. The discrepancy order q of E will be q = max(qi )1≤i≤ℓ−1 . Define now
m = (mi )1≤i≤ℓ−1 where for all i, mi satisfies1
2mi = qi .
We then define gυ,m ∈ G by2
gυ,m = IdG1 +υ1 ε−m1 IdG2 +υ1 υ2 ε−m1 −m2 IdG3 + . . . + (Πℓ−1
i=1 υi )ε
−m1 −...−mℓ−1
IdGℓ ,
where the constants υ = (υi )1≤i≤ℓ−1 ∈ (R∗ )ℓ−1 will be determined soon. Denote
by
(4.4) γε := gυ,m · γ.
Then, the operators ∂ E and ∂ ε = ∂ 0 + γε are gauge equivalent for any υ. We denote
by Aε the Chern connection on (π ∗ E, π ∗ (∂ ε ), π ∗ h) (and we will drop from now the
π ∗ to ease notations). We will consider projections onto various components of g.
For ψ ∈ g and s = sX + sZ ∈ Γ(X ′ , End(E)) = Γ(X, End(E)) ⊕ Γ0 (Z ′ , End(E)),
we will denote by Πhψi (s) the orthogonal projection (as in Equation (4.1)) of sX
onto the subspace spanned by ψ. Then, exactly as in [16, Proposition 5.25], using
Chern-Weil theory together with the facts that the constants νi are all strictly
positive and the γi,i+1 all non zero, we obtain :
Proposition 4.2. There exist υ = (υ1 , . . . , υℓ−1 ) ∈ (R∗ )ℓ−1 and a positive constant
C such that for all j ∈ [[1, ℓ − 1]],
1
ΠhIdFj i Λωε (iFAε ) = C µLε (E) IdFj +O(εqj + 2 ).

In what follows, we assume that the constants υ are fixed to satisfy the conclu-
sion in Proposition 4.2. The next step is to produce, for any order p ≥ q, gauge
perturbations fε · Aε of Aε satisfying
Λωε (iFfε ·Aε ) = C µLε (E) IdE +O(εp ).

4.2. Inductive process and approximate solutions. To construct the approx-

imate solutions to any desired order, we will perturb iteratively the connections
Aε by gauge transformations of the form exp(εi si ) and use the properties of the
associated Laplace operators. We denote by ∆ε (resp. ∆Gr(E),ε ) the associated
Laplacian of Aε (resp. of the Chern connection of (π ∗ ∂ 0 , π ∗ h)) with respect to ωε .
We also set
∆X := ∆Gr(E),0
the Laplacian of (∂ 0 , h) on X and L the second order operator associated to (∂ 0 , h)
as in the matrix Expression (3.3).

1The formula differs from [16], where it was 2m + 1 = q , as here, when contracting upon Λ ,
i i ε
we don’t obtain an extra ε-term (as was the case when contracting basic terms in [16])
2Note that our convention for the gauge action is opposite to the one in [16], which accounts
for the negative powers in ε here.
 BLOWING-UP HYM CONNECTIONS 13

Proposition 4.3. Under the above hypotheses, there are expansions

(4.5) ∆ε = ∆Gr(E),ε + O(ε),
and for sX + sZ ∈ Γ(X, End E) ⊕ Γ0 (Z ′ , End E),
(4.6) ∆Gr(E),ε (sX + εsZ ) = ∆X (sX ) + L(sX ) + ∆V (sZ ) + O(ε).
The same expansions also hold at a Chern connection on π ∗ E → X ′ coming
from a complex structure fε · π ∗ (∂ 0 + γε ) provided fε = IdE +sε for some sε ∈
Γ (X ′ , End π ∗ E) whose X-component sX,ε satisfies sX,ε = O(ε) and whose Z ′ -
component sZ,ε satisfies sZ,ε = O(ε2 ).
Proof. For the proof of (4.5), we use the formula for the change of curvature induced
by the change in complex structure ∂ 0 7→ ∂ 0 + γε :
(4.7) FAε = FA + dA0 (γε − γε∗ ) + (γε − γε∗ ) ∧ (γε − γε∗ ).
Upon contraction with ωε , we obtain
Λε (iFAε ) = Λε (iFA0 ) + Λε ∂0 γε − ∂ 0 γε∗ − 2Λε (γε ∧ γε∗ ) .

Now, by choice of gυ,m and definition of γε in Equation (4.4), we see that

γε = O(εmin(mi ) ).
By definition, mi = q2i . By Formula (1.2), for all i ∈ [[1, ℓ − 1]], we have qi ≥ n − m,
and as we are blowing-up a submanifold, we must have m ≤ n − 2. Then, qi ≥ 2,
and thus mi ≥ 1 for all i. Then,
γε = O(ε),
which gives
Λε (iFAε ) = Λε (iFA0 ) + O(ε).
Note that all the terms in ∂0 γε − ∂ 0 γε∗ − 2 (γε ∧ γε∗ ) are pulled back terms, and

arguing as in Section 3.2, we obtain that this estimate is preserved for first order
variations, which gives
∆ε = ∆Gr(E),ε + O(ε).
The proof of (4.6) is a direct consequence of the various lemmas in Section 3.2.
Finally, we consider the perturbed connections under gauge transformations fε as
described in the statement of the lemma. The expansion (4.5) is trivially preserved
for such connections, as fε = IdE +O(ε). Then, the X-part of ∂Aε ∂¯Aε − ∂¯Aε ∂Aε
changes at order ε while its Z ′ -part changes at order ε2 , which is enough to com-
pensate for the ε−1 -contribution that comes upon restriction to Z ′ and contraction,
as in the proof of Lemma 3.3. This concludes the proof. 

When producing the approximate HYM connections, we will use the mapping
properties of ∆X and ∆V . This will allow us to remove all error terms modulo
elements in the cokernel of ∆X = ∆Gr(E),0 , which is g. By Chern-Weil theory,
those remaining terms will be controlled by the expansions of µLε (Fi ), and it is
important in the argument that upon removing the other errors, they remain fixed.
This is the content of the following lemma.
14 A. CLARKE AND C. TIPLER

Lemma 4.4. Let s ∈ Γ(End(Gi )) and j ∈ [[1, ℓ]]. Then

ΠhIdFj i ∆ε (s) = O(εqj )
and
ΠhIdFj i Λε (∂ε s − ∂¯ε s) ∧ (∂ε s − ∂¯ε s) = O(εqj ).

Let σ ∈ Γ(Hom(Gp , Gi )), with i < p, and j ∈ [[1, ℓ]]. Then

ΠhIdFj i ∆ε (εmi +...+mp−1 +1 σ) = O(εqj +1 ).
and


ΠhIdFj i Λε ε2(mi +...+mp−1 +1) (∂ε σ − ∂¯ε σ) ∧ (∂ε σ − ∂¯ε σ) = O(εqj +1 ),
and similarily for p < i.
Proof. Note that we only need to consider the Γ(X, End E) components of the
sections in the splitting (3.2), as we are only interested in the projection onto IdFj .
Then, the proof follows exactly as in [16, Lemma 5.20 and Lemma 5.21]. 
The only errors that will then remain to remove are those in g. This is where
we will use the following lemma, whose proof is a direct adaptation of [16, proof of
Lemma 5.23].
Lemma 4.5. For all j ∈ [[1, ℓ − 1]], there is a negative constant aj,j+1 such that
(4.8) Πg ∆ε (IdFj ) = aj,j+1 εqj Idj,j+1 +O(εqj +1 ),
where for (p, l) ∈ [[1, ℓ]]2 :
1 1
Idpl = IdGp − IdGl .
rank Gp rank Gl
Note that as in Proposition 4.3, the conclusions of Lemma 4.4 and 4.5 also
hold for perturbed connections of the form fε · Aε for fε = IdE +sε,X + sε,Z with
sε,X = O(ε) and sε,Z = O(ε2 ). We are now ready to prove the main proposition of
this section.
Proposition 4.6. Let p ≥ q. Then, there exists gauge transformations fεp =
IdE +spε,X + spε,Z with spε,X = O(ε) and spε,Z = O(ε2 ), and constants cpε , such that
for all 0 < ε ≪ 1, we have
Λε Ffεp ·Aε = cpε IdE +O(εp ).

Proof. The proof follows the strategy developped in [16, Section 5.2 and Section
5.3.2], so we refer to those relevant sections and will only sketch the argument. We
can expand in ε each term on the right hand side of the following identity :
Λε (iFAε ) = Λε (iFA0 ) + Λε ∂0 γε − ∂ 0 γε∗ − 2Λε (γε ∧ γε∗ ) .

As A0 is HYM, we have
Λε (iFA0 ) = c IdE +O(ε).
∗
By the gauge fixing condition ∂ 0 γ = Λ0 ∂0 γ = 0, the following off-diagonal terms
have entries
(Λε ∂0 γε − ∂ 0 γε∗ )ip = O(εmi +...+mp−1 +1 ),

while the remaining terms will contribute to higher orders. Following [16, Proposi-
tion 4.9 and Section 5.2], we start by perturbing inductively Aε to remove all errors
that live in Γik⊥ (X, EndH (E, h)) ⊕ Γ0 (Z ′ , EndH (E, h)). This can be achieved by
 BLOWING-UP HYM CONNECTIONS 15

considering perturbations exp(εi (sX +εsZ )·Aε ) thanks to Proposition 4.3, together
with Lemmas 3.2 and 4.1. What remains are errors in ik. By Lemma 4.4, and from
the formula


ΛFf ·A = ΛFA + ∆A (f ) + Λ (∂A f − ∂ A f ) ∧ (∂A f − ∂ A f ) ,
the previous perturbations won’t affect any of the IdFj projections of the contracted
curvatures. By proposition 4.2, we then obtain a connection whose contracted
curvature satisfies, for all j ∈ [[1, ℓ − 1]],
1
Λωε (iFAε ) = C µLε (E) IdFj +O(εqj + 2 ).
The end of the proof is then done by induction on ℓ the number of stable components
of Gr(E). As discussed in [16, Section 5.3.2], using Lemma 4.5, we can remove errors
in g beyond the orders qj , by induction on ℓ, and eventually obtain the result. 

5. The perturbation argument

We keep notations from the last section. We will now prove Theorem 1.1, and
its corollaries. Again, we follow closely [16]. The first step is to obtain an upper
bound for the operator norm of the inverse of ∆ε . This is where the argument
from [16] has to be adapted, given the different geometric settup that we address.
Then, based on this bound and on the construction of approximate solutions, a
quantitative version of the implicit function theorem is used in Section 5.2.
For a Riemannian metric g on X ′ , let L2d (g) denote the Sobolev space W d,2 (X ′ , g)
of order 2 and d derivatives, with respect to the metric g. If g is Kähler with Kähler
form η, we may write L2d (η) instead of L2d (g). When d = 0, we omit the subscript.
For Sobolev spaces associated to π ∗ E or End π ∗ E, we use the metric h (and the
metric it induces on End π ∗ E) on the bundle.

5.1. Estimating the linearisation. Our main goal in this section is to prove:
Proposition 5.1. There exists C > 0 such that for all s ∈ Γ(X ′ , End E) whose
trace is average zero with respect to ωε ,
(5.1) k∆ε (s)kL2d (ωε ) ≥ Cεq kskL2d+2 (ωε ) .
The same estimate also holds for the Laplacian of the perturbed connections built
in Proposition 4.6.
This will be done by steps. We first work out the case of a single stable component
G := Gi ⊂ Gr(E).
Lemma 5.2. There exists C > 0 such that for all s ∈ Γ(X ′ , End G) that is L2 (ωε )-
orthogonal to IdG ,
(5.2) k∆ε (s)kL2d (ωε ) ≥ CεkskL2d+2(ωε ) .

The proof of Lemma 5.2 will require to compare various L2 -norms on X ′ , X and
′
Z . We gather those comparisons in the following lemmas.
Lemma 5.3. There are positive constants C and C ′ independent on ε such that
for any σ ∈ Γ(X, End G) (or σ ∈ Ω1 (X, End G)), we have
(5.3) Ckπ ∗ σkL2 (X ′ ,ωε ) ≤ kσkL2 (X,ω) ≤ C ′ kπ ∗ σkL2 (X ′ ,ωε ) .
16 A. CLARKE AND C. TIPLER

Proof. The result follows from the facts that the 2-forms ωε vary in a bounded
family (ε ∈ [0, ε0 ]), and that a pulled back section is constant on the fibers of
π : Z ′ → Z. 
Lemma 5.4. There are positive constants C and C ′ independent on ε such that
for any σ ∈ Γ0 (Z ′ , End G) (or σ ∈ Ω1 (Z ′ , End G)), we have
(5.4) Ckι∗ σk2L2 (X ′ ,ωε ) ≤ kσk2L2 (Z ′ ,ωε ) ≤ C ′ kι∗ σk2L2 (X ′ ,ωε ) .
Proof. This can be done by a local and then patching argument. We may fix
W ⊂ X ′ an open set with Z ′ ⊂ W , and such that ι∗ σ vanishes away from W .
Then, Z
kι∗ σk2L2 (X ′ ,ωε ) = kι∗ σk2ε ωεn .
W
We may also consider an open finite covering of W by sets U of the form U ≃
B(0, 1) × V where the sets V ⊂ Z ′ cover Z ′ and such that ι∗ σ is given locally by
ρ·σ for a bump function ρ on B(0, 1). Then, from the facts that ε varies in a compact
family, and that for any v ∈ V , the metrics (ωε (b, v))b∈B(0,1) on B(0, 1) × {v} are
mutually bounded, by applying Fubini’s theorem, we obtain bounds :
Z Z Z
C kι∗ σk2ε ωεn ≤ kσk2ε ωεn−1 ≤ C ′ kι∗ σk2ε ωεn .
U V U
A patching argument then provides the result. 
Lemma 5.5. There exists a positive constant C independent on ε such that for
any sX ∈ Γ(X, End G) and sZ ∈ Γ0 (Z ′ , End G), we have
|hsX , sZ iL2 (ωε ) | ≤ Cε3 ||sX ||L2 (ωε ) ||sZ ||L2 (ωε ) .
Proof. We keep the notations from the proof of Lemma 5.4. Note that we can
restrict to the open set W as sZ vanishes away from W . Then, locally, we estimate
Z
trace(sX · sZ ) ωεn .
U
Take coordinates (b, f, v) ∈ B(0, 1) × F × VZ so that U ≃ B(0, 1) × F × VZ where
this time F denote fibers of π : Z ′ → Z, VZ ⊂ Z is an open set in Z and B(0, 1)
stands for normal coordinates. We may assume that sX is independent on f ∈ F :
sX = sX (b, v)
and sZ is of the form
sZ = ρ(b)sZ (f, w)
for ρ some cut off function used to produce the extension operator. Then, we
compute with Fubini’s :
Z Z
trace(sX · sZ ) ωεn = ρ(b) trace(sX (b, w) · sZ (f, w)) ωεn
U Z B×F ×VZ Z

= ρ(b) trace(sX (b, w) · sZ (f, w)) ωεn .

B×VZ F
But to higher order in ε,
Z
trace(sX (b, w) · sZ (f, w)) ωεn =
F
Z
n−m−1 m+1
ε ω trace(sX (b, w) · sZ (f, w)) λn−m−1 + O(εn−m )
F
 BLOWING-UP HYM CONNECTIONS 17

and as sZ is of average zero w.r.t. λn−m−1 , we get

Z
trace(sX (b, w) · sZ (f, w))ωεn = O(εn−m ),
F

which is O(ε3 ) as codim(Z) = n − m ≥ 3. The result then follows from Cauchy-

Schwarz inequality, for a constant C that depends on ω, λ and ε0 . 
Proof of Lemma 5.2. The positive constants C, C ′ , Ci used in the proof might vary
at several stages. We follow the strategy from [16, Section 4.2]. The key to obtain
this estimate is the analogue of [16, Lemma 4.12], as the rest of the argument
follows as in [16, Section 4.2]. So our goal is to obtain the following Poincaré type
inequality: there exists C > 0 such that for any ε ∈ [0, ε0 ], and all s ∈ Γ(X ′ , End G)
that is L2 (ωε )-orthogonal to IdG , we have :
kdAε (s)k2L2 (ωε ) ≥ Cεksk2L2 (ωε ) .
As Aε = A0 + O(ε), it is actually enough to obtain this Poincaré inequality for A0 :
kdA0 (s)k2L2 (ωε ) ≥ Cεksk2L2 (ωε ) .
We will then prove this in three steps. First, for sections s ∈ Γ(X, End G), then for
sections s ∈ Γ0 (Z ′ , End G) and last for sums of such sections.
Step 1 : For σ ∈ Γ(X, End G) and s = π ∗ σ ∈ Γ(X ′ , End G), let
1 1
Z
α = trace(σ) ω n ,
rank G vol(X, ω) X
1 1
Z
′
α = trace(s) ωεn .
rank G vol(X ′ , ωε ) X ′
Then, s is orthogonal to Id on X ′ if and only if α′ = 0. Moreover,
Z
α′ − α = O(ε) trace(σ) ω n .
X
Then, noting that dA0 σ = dA0 (σ − α Id) as Id is parallel, we have
kdA0 (π ∗ σ)k2L2 (X ′ ,ωε ) ≥ C1 kπ ∗ dA0 σk2L2 (X ′ ,ωε ) ,
≥ C2 kdA0 σk2L2 (X,ω) ,
≥ C3 kσ − α Id k2L2 (X,ω) ,
≥ C4 kσk2L2 (X,ω) ,
≥ C5 ksk2L2 (X ′ ,ωε ) ,
where the second and fifth inequalities follow from Lemma 5.3, the third one from
Poincaré inequality for A0 on (X, ω), and the fourth inequality follows since α is
O(εkσkL2 (ω) ).
Step 2 : We next consider s = ι∗ σ ∈ Γ0 (Z ′ , End G). We set this time α′′ to be
the constant given by the L2 (Z ′ , ωε )-orthogonal projection of σ onto IdE :
1
Z
α′′ = trace(σ) ωεn−1 .
rank G vol(Z ′ , ωε ) Z ′
In general, for a given σ ∈ Γ(Z ′ , End G), α′′ = O(1) with respect to ε. However, by
definition of the subspace Γ0 (Z ′ , End G) ⊂ Γ(Z ′ , End G), since σ ∈ Γ0 (Z ′ , End G) is
orthogonal to the identity on all fibres of π : Z ′ → Z, an argument similar to that
18 A. CLARKE AND C. TIPLER

of Lemma 5.5, using the same system of coordinates (b, f, v) ∈ B(0, 1) × F × VZ ,

shows that α′′ = O(εkσkL2 (Z ′ ,ωε ) ). Then we find positive constants Ci such that
kdA0 (ι∗ σ)k2L2 (ωε ) ≥ C1 kι∗ (dA0 σ)k2L2 (X ′ ,ωε )
≥ C2 kdA0 σk2L2 (Z ′ ,ωε )
≥ ε C3 kσ − α′′ Idk2L2 (Z ′ ,ωε )
≥ ε C4 kσk2L2 (Z ′ ,ωε ) ,
≥ ε C5 kι∗ σk2L2 (X ′ ,ωε ) ,
where this time the first inequality follows from construction of the operator ι∗ ,
the second and last inequalities come from Lemma 5.4, the third one follows from
the corresponding [16, Lemma 4.12], and the fourth one from the fact that α′′ =
O(εkσkL2 (Z ′ ,ωε ) ).
Step 3 : We need now to obtain similar estimates for sections s = sX + sZ . Note
however that the splitting of (3.2) is not orthogonal with respect to the L2 (ωε )
inner product. What remains is to estimate
hdA0 sX , dA0 sZ iL2 (ωε ) = h∆A0 ,ε (sX ), sZ iL2 (ωε ) = hsX , ∆A0 ,ε (sZ )iL2 (ωε ) ,
for ∆A0 ,ε the Laplacian of A0 with respect to ωε . We use the expansion of ∆ε from
Section 3.2. From Lemma 3.5, Lemma 3.6 and formula (3.3) :
h∆A0 ,ε (sX ), sZ iL2 (ωε ) = h∆X sX , sZ iL2 (ωε ) + hL(sX ), sZ iL2 (ωε ) + O(ε)
Then, from Lemma 5.5,
|h∆X sX , sZ iL2 (ωε ) | ≤ C1 ε3 ||∆X sX ||L2 (ωε ) ||sZ ||L2 (ωε ) .
Setting sX = π ∗ σ, and using Lemma 5.3, together with the continuity of the Green
operator of ∆X , we obtain
|h∆X sX , sZ iL2 (ωε ) | ≤ C2 ε3 ||∆X σ||L2 (ω) ||sZ ||L2 (ωε ) ,
≤ C3 ε3 ||σ||L2 (ω) ||sZ ||L2 (ωε ) ,
≤ C4 ε3 ||sX ||L2 (ωε ) ||sZ ||L2 (ωε ) ,
where again we dealt with the L2 (X, ω)-projection of σ onto IdG as in Step 1. For
the term hL(sX ), sZ iL2 (ωε ) , we use local coordinates (b, f, v) ∈ B × F × VZ as in
the proof of Lemma 5.5. Then, the expression of L can be locally written
DsX ∧ ω m ∧ λn−m−1
 
∗
L(sX ) = ι∗ ◦ p0 ◦ ι (m + 1)
ω m+1 ∧ λn−m−1
Db sX ∧ ω ∧ λn−m−1
m
= (m + 1)
ω m+1 ∧ λn−m−1
where Db stands for derivatives in the B(0, 1)-direction. We then compute the local
contribution on U of the term hL(sX ), sZ iL2 (ωε ) :
Z Z
trace(L(sX ) · sZ ) ωεn = ρ(b) trace(L(sX (b, w)) · sZ (f, w)) ωεn ,
U B×F ×VZ

and by Fubini’s, Stokes theorem, and integration by parts, the relevant higher order
term in ε is Z
Db (ρ(b)) trace(sX · sZ )ωεn ,
U
 BLOWING-UP HYM CONNECTIONS 19

which is O(ε3 ||sX ||L2 (ωε ) ||sZ ||L2 (ωε ) ) by a similar argument as in Lemma 5.5. Gath-
ering those estimates, we see that the mixed terms satisfy
hdA0 sX , dA0 sZ iL2 (ωε ) = h∆A0 ,ε (sX ), sZ iL2 (ωε ) = O(ε3 ||sX ||L2 (ωε ) ||sZ ||L2 (ωε ) ),
and together with steps 1 and 2, using Lemma 5.5 again, we obtain
kdA0 (sX + sZ )k2L2 (ωε ) ≥ CεksX + sZ k2L2 (ωε ) .
Then, from this Poincaré inéquality, the result follows as in [16, Section 4.2],
using an analogous uniform Schauder estimate as in [16, Proposition 4.13]. The
latter estimate can be obtained by patching local Schauder estimates (as in [16,
Lemma 4.14]) that are easily derived away from the exceptional divisor, and can
be obtained around a point in the exceptional divisor by adapting [16, Proof of
Lemma 4.14] using local coordinates as in Lemma 5.5. 
Once this bound settled, we can deal with the general case.
Proof of Proposition 5.1. The proof for the estimate (5.1) can be established ex-
actly as in [16, Proposition 5.27], following [16, Lemma 5.28, Lemma 5.29]. First,
to obtain the estimate for elements in g⊥ , one uses the full expansion in the lin-
earisation :
∆ε = ∆Gr(E),ε
iΛε ∂ 0 ([γε∗ , ·]) − [γε∗ , ∂ 0 ·]


+
(5.5)
+ iΛε (∂0 ([γε , ·]) − [γε , ∂0 ·])
+ iΛε ([γε , [γε∗ , ·]] − [γε∗ , [γε , ·]]).
As γε is a pulled back term, arguing as in Section 3.2, we see that the X and
Z ′ contributions of the operator Λε (∂0 ([γε , ·])) in the splitting (3.2) will be of the
same order as γε . As argued before, from the choice of gυ,m and definition of γε in
Equation (4.4), we see that
(γε )ip = O(εmi ).
Now, mi = q2i ≥ codim(Z)
2 by Formula (1.2). Arguing similarily for the other terms
in (5.5), we see that the contributions from ∆ε − ∆Gr(E),ε will all come at order
codim(Z)
at least ε 2 , and will be absorbed by the estimate in Lemma 5.2, because we
assumed codim(Z) ≥ 3. Arguing as in [16, Lemma 5.28], we obtain the bound :
k∆ε (s)kL2d (ωε ) ≥ CεkskL2d+2(ωε ) ,
for s ∈ g⊥ . Then, from here, the proof for the estimate for sections in g, or for
sums of sections, follows as in [16, proof of Lemma 5.29 and proof of Proposition
5.27]. Finally, the result for perturbed connections follows as in [16, Proposition
5.30]. 
5.2. Perturbing to solutions. To conclude the proof of Theorem 1.1, we refer to
[16, Section 4.2 and Section 5.4.2]. It relies on a quantitative version of the implicit
function theorem, applied to the operator
Ψp,ε : L2d+2 (X ′ , ωε ) × R → L2d (X ′ , ωε )
(s, α) 7→ iΛωε (Fexp(s)·fεp ·Aε ) − α IdE ,
where fεp · Aε are the connections built in Proposition 4.6. Proposition 4.6 is the
analogue of [16, Proposition 4.9 and Section 5.3.2] where approximate solutions to
Ψp,ε = 0 are constructed, while Proposition 5.1 plays the role of [16, Propositions
5.27 and 5.30] and provides the required estimate on the linearisation of Ψp,ε , for p
20 A. CLARKE AND C. TIPLER

large enough, at the approximate solutions. Then, the implicit function theorem,
as stated in [16, Theorem 4.10], enables to conclude the existence of zeros for Ψp,ε ,
for p large and ε small, which ends the proof of Theorem 1.1.

5.3. Proof of the corollaries. We comment now on the various corollaries stated
in the introduction. First, Corollary 1.3 is a direct application of Theorem 1.1,
where E = Gr(E) as a single stable component. Corollary 1.4 also follows directly,
using Formula (1.2). What remains is to show Corollary 1.5. The only remaing
case to study is when for all i ∈ [[1, ℓ − 1]], µLε (Fi ) ≤ µLε (E), with at least one
ε→0
equality. This case can be dealt with exactly as for [16, Corollary 5.2], proved in
[16, Section 5.5].

References
[1] Claudio Arezzo and Frank Pacard. Blowing up and desingularizing constant scalar curvature
Kähler manifolds. Acta Math., 196(2):179–228, 2006. 1
[2] Claudio Arezzo and Frank Pacard. Blowing up Kähler manifolds with constant scalar curva-
ture. II. Ann. of Math. (2), 170(2):685–738, 2009. 1, 4
[3] Claudio Arezzo, Frank Pacard, and Michael Singer. Extremal metrics on blowups. Duke Math.
J., 157(1):1–51, 2011. 1, 4
[4] Nicholas P. Buchdahl. Blowups and gauge fields. Pacific J. Math., 196(1):69–111, 2000. 1, 3
[5] Ruadhaı́ Dervan and Lars Martin Sektnan. Extremal Kähler metrics on blow-ups. ArXiv
preprint arXiv:2110.13579, 2021. 1, 4
[6] Ruadhaı́ Dervan and Lars Martin Sektnan. Hermitian Yang-Mills connections on blowups. J.
Geom. Anal., 31(1):516–542, 2021. 1, 3, 4
[7] S. K. Donaldson. Infinite determinants, stable bundles and curvature. Duke Math. J.,
54(1):231–247, 1987. 1, 6
[8] Daniel Huybrechts and Manfred Lehn. The geometry of moduli spaces of sheaves. Cambridge
Mathematical Library. Cambridge University Press, Cambridge, second edition, 2010. 4, 5, 6
[9] Shoshichi Kobayashi. Curvature and stability of vector bundles. Proc. Japan Acad. Ser. A
Math. Sci., 58(4):158–162, 1982. 1, 6
[10] Shoshichi Kobayashi. Differential geometry of complex vector bundles. Princeton Legacy
Library. Princeton University Press, Princeton, NJ, [2014]. Reprint of the 1987 edition [
MR0909698]. 4
[11] Kunihiko Kodaira and D. C. Spencer. On deformations of complex analytic structures. III:
Stability theorems for complex structures. Ann. Math. (2), 71:43–76, 1960. 8
[12] Martin Lübke. Stability of Einstein-Hermitian vector bundles. Manuscripta Math., 42(2-
3):245–257, 1983. 1, 6
[13] V. B. Mehta and A. Ramanathan. Restriction of stable sheaves and representations of the
fundamental group. Invent. Math., 77(1):163–172, 1984. 3
[14] David Mumford. Projective invariants of projective structures and applications. In Proc.
Internat. Congr. Mathematicians (Stockholm, 1962), pages 526–530. Inst. Mittag-Leffler,
Djursholm, 1963. 1, 5
[15] Achim Napame and Carl Tipler. Toric sheaves, stability and fibrations. ArXiv preprint
arXiv:2210.04587, 2022. 2, 3
[16] Lars Martin Sektnan and Carl Tipler. Hermitian Yang–Mills connections on pullback bundles.
To appear in Calculus of Variations and PDEs, arXiv:2006.06453, 2020. 2, 4, 7, 8, 10, 11, 12,
14, 15, 17, 18, 19, 20
[17] Reza Seyyedali and Gábor Székelyhidi. Extremal metrics on blowups along submanifolds. J.
Differential Geom., 114(1):171–192, 2020. 1, 4
[18] Gábor Székelyhidi. Blowing up extremal Kähler manifolds II. Invent. Math., 200(3):925–977,
2015. 1, 4
[19] Fumio Takemoto. Stable vector bundles on algebraic surfaces. Nagoya Math. J., 47:29–48,
1972. 1, 5
 BLOWING-UP HYM CONNECTIONS 21

[20] K. Uhlenbeck and S.-T. Yau. On the existence of Hermitian-Yang-Mills connections in stable
vector bundles. volume 39, pages S257–S293. 1986. Frontiers of the mathematical sciences:
1985 (New York, 1985). 1, 6
[21] Claire Voisin. Hodge theory and complex algebraic geometry. I, volume 76 of Cambridge
Studies in Advanced Mathematics. Cambridge University Press, Cambridge, english edition,
2007. Translated from the French by Leila Schneps. 8

Andrew Clarke, Instituto de Matemática, Universidade Federal do Rio de Janeiro,

Av. Athos da Silveira Ramos 149, Rio de Janeiro, RJ, 21941-909, Brazil
Email address: [email protected]

Carl Tipler, Univ Brest, UMR CNRS 6205, Laboratoire de Mathématiques de Bre-
tagne Atlantique, France
Email address: [email protected]