Introduction
The relationship between the existence of Hermitian-Einstein metrics and the stability of holomorphic vector bundles has a long and profound history. Starting with the work of Narasimhan and Seshadri, this culminated in the works of Donaldson [15 , 16 ] and Uhlenbeck-Yau [46 ] , where it was proved that the slope (poly)stability of a holomorphic vector bundle is a necessary and sufficient condition for the existence of the Hermitian-Einstein metric. This result was extended by many other authors, most notably by Simpson in [42 ] to Higgs bundles, and Bando and Siu [4 ] for stable sheaves on compact manifolds. Furthermore, Li and Yau [29 ] extended the original Donaldson-Uhlenbeck-Yau theorem to non-Kähler manifolds.
For non-stable bundles the research was started by Daskalopoulos and Wentworth [11 , 12 ] , and then later expanded by Jacob [23 , 24 ] , Sibley [38 ] , Collins-Jacob [10 ] and Sibley-Wentworth [39 ] . Their combined result shows that outside an analytic subvariety
Z a l g = { x ∈ X ∣ Gr h n s ( E ) x is not free } , Z_{alg} = \{x \in X \mid\operatorname{Gr}^{hns}(E)_x \text{ is not free}\}, Z a l g = { x ∈ X ∣ Gr hn s ( E ) x is not free } ,
there is a Yang-Mills connection on E ∣ X ∖ Z a l g E|_{X\setminus Z_{alg}} E ∣ X ∖ Z a l g .
However, the notion of the slope stability depends on the metric on the base manifold X X X . Moreover, it is known that a bundle E E E may be stable with respect to one metric ω 1 \omega_1 ω 1 , but fail to be stable with respect to another one ω 0 \omega_0 ω 0 . It is natural to ask what happens with the corresponding Hermitian-Einstein metrics, if one deforms the underlying metric on X X X in an unstable direction. This question has natural parallels with convergence of Einstein metrics on Riemannian manifolds (see [1 , 2 , 5 ] ), and cscK metrics on compact Kähler manifolds when the Kähler class approaches the canonical class (see [21 , 7 ] ).
Another way to look at the question is through the variation of symplectic (or GIT) quotients. Let ( X , ω ) (X,\omega) ( X , ω ) be a compact Kähler manifold, and let ( E , H ) (E,H) ( E , H ) be a hermitian vector bundle over X X X . Let A A A be an integrable Hermitian connection on E E E . The set of integrable connections A H 1 , 1 \mathcal{A}^{1,1}_H A H 1 , 1 can be written as A + a A+a A + a , where a ∈ Γ ( X , u ( E ) ⊗ Λ X 1 ) a \in\Gamma(X,\mathfrak{u}(E)\otimes\Lambda^1_X) a ∈ Γ ( X , u ( E ) ⊗ Λ X 1 ) , and a a a satisfies the Maurer-Cartan equation
∂ ˉ A a 0 , 1 + a 0 , 1 ∧ a 0 , 1 = 0. \bar{\partial}_A a^{0,1} + a^{0,1} \wedge a^{0,1} = 0. ∂ ˉ A a 0 , 1 + a 0 , 1 ∧ a 0 , 1 = 0.
One can define a symplectic form on A H 1 , 1 \mathcal{A}^{1,1}_H A H 1 , 1 as:
Ω ( a , b ) = − ∫ X Tr ( a ∧ b ) ∧ ω n − 1 ( n − 1 ) ! . \Omega(a,b)=-\int_X \operatorname{Tr}(a\wedge b)\wedge\frac{\omega^{n-1}}{(n-1)!}. Ω ( a , b ) = − ∫ X Tr ( a ∧ b ) ∧ ( n − 1 )! ω n − 1 .
The complex gauge group G C \mathcal{G}_{\mathbb{C}} G C acts on A H 1 , 1 \mathcal{A}^{1,1}_H A H 1 , 1 in the standard way, and the real group G R ( E ) \mathcal{G}_{\mathbb{R}}(E) G R ( E ) , which consists of the isometries of ( E , H ) (E,H) ( E , H ) , preserves the symplectic form. In this setting, one can define the corresponding moment map, and show, that the zeroes of this moment map in a given orbit of ∂ ˉ A \bar{\partial}_A ∂ ˉ A are the Hermitian Yang-Mills connections, i.e. the connections, that satisfy the equation
− 1 Λ ω F A − μ ω ( E ) Id = 0. \sqrt{-1}\Lambda_\omega F_A-\mu_\omega(E)\operatorname{Id}=0. − 1 Λ ω F A − μ ω ( E ) Id = 0.
If such A A A exists, then it is standard that one can find a Hermitian-Einstein metric on E E E .
Now if we vary the symplectic form, then the corresponding moment map changes, and at some point, it may happen, that the moment map has no zeroes. A similar variation question in the case of actions of reductive groups on projective manifolds was addressed in [14 ] , and in our framework these questions were considered in [19 , 37 ] . In these works, the authors studied wall-crossing and stability chambers, mostly on projective manifolds. Also, the related questions for the Gieseker stable bundles were considered in [20 , 32 ] . However, the differential-geometric side of the picture was much less studied.
However, it is easy to see, that we must impose additional assumptions on the Hermitian-Einstein metrics, as one can see, that for different Kähler metrics in a given class [ ω ] [\omega] [ ω ] the corresponding Hermitian-Einstein metrics may be different as well. Also, one must suitably normalize Hermitian-Einstein metrics, as multiplication by a positive number does not change the curvature of the metric, while it clearly may affect convergence.
Throughout this paper we work under the following set of assumptions:
Assumptions:
( X , ω 0 ) (X, \omega_0) ( X , ω 0 ) is a compact Kähler manifold of complex dimension n n n ;
ω s \omega_s ω s , s ∈ ( 0 ; 1 ] s \in(0; 1] s ∈ ( 0 ; 1 ] is a smooth family of uniformly equivalent Kähler metrics. This means that each ω s \omega_s ω s is smooth, they depend smoothly on s s s , and there is a uniform constant C > 1 C > 1 C > 1 , such that for all s ∈ [ 0 ; 1 ] s \in[0; 1] s ∈ [ 0 ; 1 ] we have
C − 1 ω 0 ≤ ω s ≤ C ω 0 ; C^{-1}\omega_0 \le\omega_s \le C\omega_0; C − 1 ω 0 ≤ ω s ≤ C ω 0 ;
We also require that ω s \omega_s ω s tends to ω 0 \omega_0 ω 0 uniformly as s → 0 s \to0 s → 0 . Later, in § 4 , we also need the additional assumption that curvatures of ω s \omega_s ω s are uniformly bounded.
E E E is a holomorphic vector bundle of rank r r r , which is stable with respect to all ω s \omega_s ω s , s ∈ ( 0 ; 1 ] s \in(0; 1] s ∈ ( 0 ; 1 ] ;
Let H s H_s H s be a corresponding family of Hermitian-Einstein metrics (HE), depending smoothly on s ∈ ( 0 ; 1 ] s \in(0;1] s ∈ ( 0 ; 1 ] (this should be possible to arrange, via the implicit function theorem, [43 iv] ). More precisely:
− 1 Λ ω s F H s = μ ω s ( E ) Id , \sqrt{-1}\Lambda_{\omega_s}F_{H_s}=\mu_{\omega_s}(E)\operatorname{Id}, − 1 Λ ω s F H s = μ ω s ( E ) Id ,
where
μ ω s ( E ) = 2 π r Vol ( X , ω s ) ∫ X c 1 ( E ) ∧ ω s n − 1 ( n − 1 ) ! . \mu_{\omega_s}(E)=\frac{2\pi}{r\operatorname{Vol}(X,\omega_s)}\int_X c_1(E)\wedge\frac{\omega_s^{n-1}}{(n-1)!}. μ ω s ( E ) = r Vol ( X , ω s ) 2 π ∫ X c 1 ( E ) ∧ ( n − 1 )! ω s n − 1 .
We also introduce the relative endomorphism h s = H − 1 H s h_s=H^{-1}H_s h s = H − 1 H s , and we can choose H s H_s H s such that ∫ X log det h s ω s n = 0 \int_X\log\det h_s\omega_s^n=0 ∫ X log det h s ω s n = 0 . We will show that under this assumption, the determinant of h s h_s h s is uniformly bounded above and below.
A few comments about our assumptions. As we discussed earlier, the definition of μ ω ( E ) \mu_\omega(E) μ ω ( E ) , and hence, the stability notion, depend on the Kähler class rather than the metric itself. However, if one changes the metric within the Kähler class, then the corresponding Hermitian-Einstein metric also changes. If we have a smooth family [ ω s ] [\omega_s] [ ω s ] of Kähler classes, then the corresponding family of Kähler metrics, and the corresponding family of Hermitian-Einstein metrics H s H_s H s need not be smooth. Hence, it is necessary to vary the Kähler metrics themselves. Moreover, if we want to study compactness properties of the Hermitian-Einstein metrics H s H_s H s , it is necessary to impose a certain normalization on these metrics, and we later show that our normalization for the relative endomorphism h s h_s h s is natural to assume, and can always be achieved. It is not hard to see that our assumptions on Kähler metrics are quite natural too, and do not impose any additional restrictions on their Kähler classes. Indeed, if [ ω 0 ] [\omega_0] [ ω 0 ] lies on the boundary of a stability chamber, and [ ω 1 ] [\omega_1] [ ω 1 ] is the class, lying within this chamber and sufficiently close to [ ω 0 ] [\omega_0] [ ω 0 ] , then the family ω s = ( 1 − s ) ω 0 + s ω 1 \omega_s=(1-s)\omega_0+s\omega_1 ω s = ( 1 − s ) ω 0 + s ω 1 satisfies all our regularity assumptions. One can construct other examples of such families of Kähler metrics.
Our main result here is the following.
Theorem 1.1. Let X X X be a compact complex manifold, and suppose that the Assumptions above hold. Let H s H_s H s be the corresponding family of Hermitian-Einstein metrics, such that the endomorphism h s = H 1 − 1 H s h_s = H_1^{-1}H_s h s = H 1 − 1 H s is normalized by ∫ X log det h s ω s n = 0 \int_X \log\det h_s\omega_s^n = 0 ∫ X log det h s ω s n = 0 . Then
If Tr h s \operatorname{Tr} h_s Tr h s is uniformly bounded, then E E E is ω 0 \omega_0 ω 0 -stable.
If Tr h s \operatorname{Tr} h_s Tr h s is unbounded, then the bundle E E E is strictly ω 0 \omega_0 ω 0 -semistable. The projector π F \pi_{\mathcal{F}} π F onto a ω 0 \omega_0 ω 0 -destabilizing sheaf F \mathcal{F} F can be constructed as a weak limit of endomorphisms Id − h ~ s o = Id − ( h s sup X Tr h s ) o \operatorname{Id}-\tilde{h}_s^o = \operatorname{Id}-\left(\frac{h_s}{\sup_X\operatorname{Tr}h_s}\right)^{o} Id − h ~ s o = Id − ( s u p X Tr h s h s ) o .
If we have a uniform bound on Rm ( ω s ) \operatorname{Rm}(\omega_s) Rm ( ω s ) , then there is a subsequence s j s_j s j , and a subset Z a n Z_{an} Z an , such that the Chern connections weakly converge (after a suitable gauge transform) in L 1 , l o c p ( X ∖ Z a n ) L^p_{1,\mathrm{loc}}(X\setminus Z_{an}) L 1 , loc p ( X ∖ Z an ) , p > 2 n p>2n p > 2 n , and blow up along Z a n Z_{an} Z an . If, in addition, we have C k C^k C k bounds on Rm ( ω s ) \operatorname{Rm}(\omega_s) Rm ( ω s ) , then the convergence of connections can be taken in C l o c k ( X ∖ Z a n ) C^k_{\mathrm{loc}}(X\setminus Z_{an}) C loc k ( X ∖ Z an ) . Also, the limiting bundle E ∞ E_\infty E ∞ can be extended over all of X X X as a reflexive coherent sheaf, and we have equality
E ∞ ≅ ( Gr ω 0 S ( E ) ) ∗ ∗ . E_\infty\cong\left(\operatorname{Gr}_{\omega_0}^{S}(E)\right)^{**}. E ∞ ≅ ( Gr ω 0 S ( E ) ) ∗∗ . In addition, under the same assumptions as previously, the set Z a n Z_{an} Z an is an analytic subvariety, and the sets Z a n Z_{an} Z an and Z a l g Z_{alg} Z a l g coincide. Any irreducible component Z ⊂ Z a n Z\subset Z_{an} Z ⊂ Z an of codimension 2 lies in Z a l g Z_{alg} Z a l g , and the multiplicities of Z Z Z in Z a n Z_{an} Z an and Z a l g Z_{alg} Z a l g are the same.
The theorem says, that, as we approach the wall of the stability chamber, the appropriate choice of Hermitian-Einstein metrics for [ ω s ] [\omega_s] [ ω s ] detects instability of our bundle E E E with respect to [ ω 0 ] [\omega_0] [ ω 0 ] . Moreover, the destabilizing sheaves and the Seshadri filtration emerge directly from these metrics.
Theorem 1.1 should be compared with the results from [9 ] . They establish results similar to part (3) of Theorem 1.1 , under the additional assumption that E E E is sufficiently smooth, i.e. Gr ω 0 S ( E ) \operatorname{Gr}_{\omega_0}^{S}(E) Gr ω 0 S ( E ) is locally free, and has abelian automorphism group. Our result does not require such restrictions.
It is also worth mentioning, that the results (1)-(2) of Theorem 1.1 hold if our metrics ω s \omega_s ω s are balanced or Gauduchon. The Uhlenbeck compactness from (3) also can be proved, under additional assumptions on the curvatures and torsion of ω s \omega_s ω s . The part (4) of the theorem cannot be proven for non-Kähler manifolds via our methods, as it requires certain homological properties of compact Kähler manifolds. The paper [13 ] also gives results similar to part (3) of Theorem 1.1 in the setting of balanced metrics, and also removes the abelian automorphism assumption. However, the assumption on sufficient smoothness remains in that paper.
The organization of the paper is as follows: in Section 2 we review all the necessary background material, set up the notation, and establish the normalization for h s h_s h s . We also establish in Proposition 2.7 the first part of Theorem 1.1 . In Section 3 we adapt the method of Uhlenbeck and Yau from [46 ] , and prove the second part of Theorem 1.1 . Section 4.1 is devoted to preliminaries for Uhlenbeck compactness for Hermitian Yang-Mills connections in the case of varying metrics ω s \omega_s ω s . In Section 4.2 we establish Uhlenbeck compactness in our setting, and show part (3) of Theorem 1.1 . In Section 4.3 we show that Z a n Z_{\mathrm{an}} Z an is rectifiable and analytic. Finally, in Section 4.4 we establish the last part of Theorem 1.1 .
Acknoledgements: The author is grateful to Prof. Yingying Zhang and to Prof. Shing-Tung Yau for suggesting the problem, and their encouragement.
AI disclosure: The author used AI to check the language, exposition, literature review, and to serve as a proofreader. During this review, AI found a gap in the initial proof of Proposition 2.7 , as well as several other minor issues throughout the text. The correction of these issues was done by the author. All other proofs were also done by the author, and the author bears exclusive responsibility for the correctness of the results.
Preliminaries.
Harder-Narasimhan and Seshadri filtrations.
Let ( X , ω ) (X,\omega) ( X , ω ) be a compact Kähler manifold with volume V = ∫ X ω n n ! V = \int_X \frac{\omega^n}{n!} V = ∫ X n ! ω n , and E → X E \to X E → X be a holomorphic vector bundle over X X X of rank r r r . We define the slope of E E E by the formula
μ ω ( E ) = 2 π r V ∫ X c 1 ( E ) ∧ ω n − 1 ( n − 1 ) ! . (1) \mu_{\omega}(E)=\frac{2\pi}{rV}\int_X c_1(E)\wedge\frac{\omega^{n-1}}{(n-1)!}. \tag*{(1)} μ ω ( E ) = r V 2 π ∫ X c 1 ( E ) ∧ ( n − 1 )! ω n − 1 . (1)
This definition is also valid for coherent sheaves over X X X .
The bundle is called (semi)stable if for any proper and nonzero torsion-free coherent subsheaf F \mathcal{F} F of E E E with 0 < rank ( F ) < r 0<\operatorname{rank}(\mathcal{F})<r 0 < rank ( F ) < r there is an inequality
μ ( F ) ( ≤ ) < μ ( E ) . \mu(\mathcal{F})(\leq)<\mu(E). μ ( F ) ( ≤ ) < μ ( E ) .
Clearly the definition of stability depends on the Kähler class, rather than the Kähler form itself. Moreover, if E E E is stable with respect to [ ω ] [\omega] [ ω ] , then it is also stable with respect to any [ ω ~ ] [\widetilde{\omega}] [ ω ] sufficiently close to [ ω ] [\omega] [ ω ] . Hence, the set
Stab ( E ) = { [ ω ] ∈ H 1 , 1 ( X , R ) ∣ [ ω ] is K a ¨ hler and E is [ ω ] -stable } \operatorname{Stab}(E)=\{[\omega]\in H^{1,1}(X,\mathbb{R})\mid[\omega]\text{ is Kähler and }E\text{ is }[\omega]\text{-stable}\} Stab ( E ) = {[ ω ] ∈ H 1 , 1 ( X , R ) ∣ [ ω ] is K a ¨ hler and E is [ ω ] -stable }
is open and stable under dilations [ ω ] → λ [ ω ] [\omega]\to\lambda[\omega] [ ω ] → λ [ ω ] , λ > 0 \lambda>0 λ > 0 .
Next, we recall the notion of Harder-Narasimhan (HN) filtration.
Definition 2.1. (see [28 Theorem 5.7.15] , [11 Definition Proposition 2.2] ) Given any torsion free coherent sheaf E \mathcal{E} E over ( X , ω ) (X,\omega) ( X , ω ) , there exists a filtration
0 = E 0 ⊂ E 1 ⊂ ⋯ ⊂ E m = E , 0=\mathcal{E}_0\subset\mathcal{E}_1\subset\cdots\subset\mathcal{E}_m=\mathcal{E}, 0 = E 0 ⊂ E 1 ⊂ ⋯ ⊂ E m = E , such that each Q j + 1 = E j + 1 / E j \mathcal{Q}_{j+1}=\mathcal{E}_{j+1}/\mathcal{E}_j Q j + 1 = E j + 1 / E j is semistable and torsion-free with μ ω ( Q j ) > μ ω ( Q j + 1 ) \mu_{\omega}(\mathcal{Q}_j)>\mu_{\omega}(\mathcal{Q}_{j+1}) μ ω ( Q j ) > μ ω ( Q j + 1 ) . Such a filtration is called the Harder-Narasimhan filtration of E \mathcal{E} E . The associated graded object Gr H N ( E ) = ⨁ j = 1 m Q j \operatorname{Gr}^{HN}(\mathcal{E})=\bigoplus_{j=1}^{m}\mathcal{Q}_j Gr H N ( E ) = ⨁ j = 1 m Q j is uniquely defined by the isomorphism class of E \mathcal{E} E .
For semistable sheaves there is another filtration, which one may consider as a decomposition of such sheaves into stable parts.
Definition 2.2. (see [11 Proposition 2.5] ) Given a semistable sheaf Q \mathcal{Q} Q over ( X , ω ) (X,\omega) ( X , ω ) , there is a filtration
0 = F 0 ⊂ ⋯ ⊂ F l = Q , 0=\mathcal{F}_0\subset\cdots\subset\mathcal{F}_l=\mathcal{Q}, 0 = F 0 ⊂ ⋯ ⊂ F l = Q , such that for any j ∈ { 1 , … , l } j\in\{1,\ldots,l\} j ∈ { 1 , … , l } the sheaf F j / F j − 1 \mathcal{F}_j/\mathcal{F}_{j-1} F j / F j − 1 is stable and torsion-free, and μ ω ( F j / F j − 1 ) = μ ω ( Q ) \mu_{\omega}(\mathcal{F}_j/\mathcal{F}_{j-1})=\mu_{\omega}(\mathcal{Q}) μ ω ( F j / F j − 1 ) = μ ω ( Q ) . Such a filtration is called a Seshadri filtration.
The double dual of the associated graded object Gr S ( Q ) = ⨁ j F j / F j − 1 \operatorname{Gr}^{S}(\mathcal{Q}) = \bigoplus_j F_j/F_{j-1} Gr S ( Q ) = ⨁ j F j / F j − 1 is uniquely determined by the isomorphism class of Q \mathcal{Q} Q (see [6 ] ).
For a semistable sheaf Q \mathcal{Q} Q one can define the notion of algebraic singular set.
Definition 2.3: Let Q \mathcal{Q} Q be a semistable torsion-free sheaf over a compact Kähler manifold ( X , ω ) (X,\omega) ( X , ω ) . The algebraic singular set is defined as
Z a l g = Z a l g ( Q ) = { x ∈ X ∣ Gr S ( Q ) x is not free } . Z_{alg}=Z_{alg}(\mathcal{Q})=\{x\in X\mid\operatorname{Gr}^{S}(\mathcal{Q})_x\text{ is not free}\}. Z a l g = Z a l g ( Q ) = { x ∈ X ∣ Gr S ( Q ) x is not free } . It is an analytic subvariety of complex codimension at least two.
We conclude this section by an observation: as we pointed out before, the notions of stability, as well as Harder-Narasimhan and Seshadri filtrations depend on the background Kähler class [ ω ] [\omega] [ ω ] . With a different choice of [ ω ] [\omega] [ ω ] the stability may change, as well as the filtrations above. There are examples (see [17 p.99] ), that already on algebraic surfaces there exist bundles, which are stable with respect to one polarization, and not stable with respect to another.
Hermitian-Einstein metrics, stability and the Einstein constant.
Let ( X , ω ) (X,\omega) ( X , ω ) be a compact Kähler manifold, and E E E be a holomorphic bundle over X X X with a Hermitian metric H H H . Recall that the metric is called Hermitian-Einstein, if
− 1 Λ ω F H = λ Id . \sqrt{-1}\Lambda_\omega F_H=\lambda\operatorname{Id}. − 1 Λ ω F H = λ Id .
A direct computation shows that λ = μ ω ( E ) \lambda=\mu_\omega(E) λ = μ ω ( E ) in our case. There is the following fundamental theorem, which relates the existence of Hermitian-Einstein metric to the questions of stability.
Theorem 2.4: The bundle E E E over ( X , ω ) (X,\omega) ( X , ω ) admits a Hermitian-Einstein metric if and only if it is polystable, i.e. it is a direct sum of stable vector bundles with the same slope.
The proof can be found in [15 , 16 ] for the case of projective manifolds, and in [46 , 42 ] for general Kähler manifolds.
We now briefly discuss the formalism, that leads to one direction of Theorem 2.4 . Assume that we have a short exact sequence of bundles over X X X :
0 → S → E → Q → 0. 0 \to S \to E \to Q \to0. 0 → S → E → Q → 0.
The metric H H H induces metrics on S S S and Q Q Q in the usual way. The curvatures are related by the following identity:
− 1 F E = ( − 1 F S + − 1 γ ∧ β − 1 ∂ ˉ γ − 1 ∂ β − 1 F Q + − 1 β ∧ γ ) (2) \sqrt{-1}F_E =
\begin{pmatrix}
\sqrt{-1}F_S+\sqrt{-1}\gamma\wedge\beta& \sqrt{-1}\bar{\partial}\gamma\\
\sqrt{-1}\partial\beta& \sqrt{-1}F_Q+\sqrt{-1}\beta\wedge\gamma
\end{pmatrix} \tag*{(2)} − 1 F E = ( − 1 F S + − 1 γ ∧ β − 1 ∂ β − 1 ∂ ˉ γ − 1 F Q + − 1 β ∧ γ ) (2)
Here β = − γ † = ∇ π S \beta=-\gamma^\dagger=\nabla\pi_S β = − γ † = ∇ π S is the second fundamental form of S S S . The ( 0 , 1 ) (0,1) ( 0 , 1 ) -form γ \gamma γ lies in Γ ( X , Hom ( Q , S ) ⊗ Λ X 0 , 1 ) \Gamma(X,\operatorname{Hom}(Q,S)\otimes\Lambda_X^{0,1}) Γ ( X , Hom ( Q , S ) ⊗ Λ X 0 , 1 ) , and satisfies the equation ∂ ˉ γ = 0 \bar{\partial}\gamma=0 ∂ ˉ γ = 0 . Since the formula above is local, we obtain the following useful statement.
Proposition 2.5: (Chern-Weil formula, see [42 Lemma 3.2] ) Let F \mathcal{F} F be a saturated torsion-free subsheaf of a holomorphic bundle E E E , and let π = π F ∈ L 1 2 ( X , ω ) \pi=\pi_{\mathcal{F}}\in L_1^2(X,\omega) π = π F ∈ L 1 2 ( X , ω ) be an orthogonal projector onto F \mathcal{F} F . We have the following formula for the slope μ ω ( F ) \mu_\omega(\mathcal{F}) μ ω ( F ) :
μ ω ( F ) = 1 rank ( F ) V ω ∫ X ( − 1 Tr ( ( Λ ω F E ) π ) − ∣ ∇ π ∣ 2 ) ω n n ! . \mu_\omega(\mathcal{F})=
\frac{1}{\operatorname{rank}(\mathcal{F})V_\omega}
\int_X\left(\sqrt{-1}\operatorname{Tr}((\Lambda_\omega F_E)\pi)-|\nabla\pi|^2\right)\frac{\omega^n}{n!}. μ ω ( F ) = rank ( F ) V ω 1 ∫ X ( − 1 Tr (( Λ ω F E ) π ) − ∣∇ π ∣ 2 ) n ! ω n . Moreover, the integral on the right hand side may be computed over X ∖ Z X\setminus Z X ∖ Z , where Z Z Z is a codimension at least two analytic set, where F ↪ E \mathcal{F}\hookrightarrow E F ↪ E is not a subbundle.
The Chern-Weil formula also implies the ”easy” direction in Theorem 2.4 .
We now return to our setting. Assume ( X , ω s , E , H s ) (X,\omega_s,E,H_s) ( X , ω s , E , H s ) as in the statement of Theorem 1.1 . We will denote H 1 H_1 H 1 by H H H , and use it as a given reference metric. We have the following important proposition.
Proposition 2.6: Consider the relative endomorphism h s = H − 1 H s h_s=H^{-1}H_s h s = H − 1 H s , normalized by
∫ X log det h s ω s n = 0. \int_X \log\det h_s\,\omega_s^n=0. ∫ X log det h s ω s n = 0.
Then det h s \det h_s det h s is uniformly bounded, i.e. there exists C > 1 C>1 C > 1 , such that
C − 1 ≤ det h s ≤ C . C^{-1}\leq\det h_s\leq C. C − 1 ≤ det h s ≤ C .
Proof: Note that
Δ s log det h s = − 1 Tr Λ ω s F H − r μ ω s ( E ) . \Delta_s\log\det h_s=\sqrt{-1}\operatorname{Tr}\Lambda_{\omega_s}F_H-r\mu_{\omega_s}(E). Δ s log det h s = − 1 Tr Λ ω s F H − r μ ω s ( E ) .
By our assumption, the RHS is uniformly bounded. By the result of [22 ] , the Green functions of the metrics ω s \omega_s ω s are uniformly bounded from below, with the bound independent of s s s . Let u s = ± log det h s u_s=\pm\log\det h_s u s = ± log det h s . The Green formula
u s = 1 Vol ( X , ω s ) ∫ X u s ω s n + ∫ X G ω s ( x , y ) ( − Δ s u s ) ( y ) ω s n ( y ) u_s=\frac{1}{\operatorname{Vol}(X,\omega_s)}\int_Xu_s\omega_s^n+\int_XG_{\omega_s}(x,y)(-\Delta_su_s)(y)\omega_s^n(y) u s = Vol ( X , ω s ) 1 ∫ X u s ω s n + ∫ X G ω s ( x , y ) ( − Δ s u s ) ( y ) ω s n ( y )
implies that u s ≤ C u_s\leq C u s ≤ C , where C > 0 C>0 C > 0 depends on the lower bound on the Green function, μ ω s ( E ) \mu_{\omega_s}(E) μ ω s ( E ) and F H F_H F H . ■
Note that in our estimate we only need the uniform equivalence of ω s \omega_s ω s . Of course, one can do the usual Moser iteration argument, but the estimates from [22 ] allow us to give a particularly streamlined proof in the Kähler case. The next proposition is important, and deals with the first part of Theorem 1.1 .
Proposition 2.7: If the endomorphisms h s = H − 1 H s h_s=H^{-1}H_s h s = H − 1 H s have uniformly bounded eigenvalues for s ∈ ( 0 ; 1 ] s\in(0;1] s ∈ ( 0 ; 1 ] , then there is a subsequence { s j } \{s_j\} { s j } , such that H j = H s j → H 0 H_j=H_{s_j}\to H_0 H j = H s j → H 0 as s j → 0 s_j\to0 s j → 0 , and H 0 H_0 H 0 is a Hermitian-Einstein metric. Moreover, E E E is ω 0 \omega_0 ω 0 -stable.
Proof:
Step 1: First of all, we need to show the following statement
Suppose E E E is a trivial holomorphic bundle over C n \mathbb{C}^n C n . Let H H H be a hermitian metric on E E E , which solves Λ ω e u c F H = 0 \Lambda_{\omega_{\mathrm{euc}}}F_H=0 Λ ω euc F H = 0 on C n \mathbb{C}^n C n , and h = H f l a t − 1 H h=H_{\mathrm{flat}}^{-1}H h = H flat − 1 H is the relative endomorphism with respect to the standard metric on E E E . If h h h has uniformly bounded eigenvalues, then H H H is constant. Let S = ∇ h h − 1 S = \nabla h h^{-1} S = ∇ h h − 1 , where ∇ \nabla ∇ is the Chern connection of H H H . The standard formula, together with upper and lower bounds on h h h gives:
Δ Tr h ≥ c 0 ∣ ∇ h h − 1 ∣ 2 \Delta\operatorname{Tr} h \ge c_0|\nabla h h^{-1}|^2 Δ Tr h ≥ c 0 ∣∇ h h − 1 ∣ 2
Here the norm is taken with respect to H H H .
Then
Δ ∣ S ∣ 2 = ∣ ∇ S ∣ 2 + ∣ ∇ ‾ S ∣ 2 + 2 Re ∇ ‾ k ˉ ∇ k ˉ S ⋅ S ≥ 0 , \Delta|S|^2 = |\nabla S|^2 + |\overline{\nabla}S|^2 + 2\operatorname{Re}\overline{\nabla}^{\bar{k}}\nabla_{\bar{k}}S \cdot S \ge0, Δ∣ S ∣ 2 = ∣∇ S ∣ 2 + ∣ ∇ S ∣ 2 + 2 Re ∇ k ˉ ∇ k ˉ S ⋅ S ≥ 0 ,
because flatness of C n \mathbb{C}^n C n together with Λ ω e u c F H = 0 \Lambda_{\omega_{\mathrm{euc}}}F_H = 0 Λ ω euc F H = 0 imply ∇ ‾ k ˉ ∇ ‾ k ˉ S = ∇ ‾ k ∇ ‾ k ˉ S = 0 \overline{\nabla}^{\bar{k}}\overline{\nabla}_{\bar{k}}S = \overline{\nabla}_{k}\overline{\nabla}^{\bar{k}}S = 0 ∇ k ˉ ∇ k ˉ S = ∇ k ∇ k ˉ S = 0 .
Thus, ∣ S ∣ 2 |S|^2 ∣ S ∣ 2 is subharmonic.
Using the standard cutoff function χ R \chi_R χ R , χ R ≡ 1 \chi_R \equiv1 χ R ≡ 1 on B R B_R B R , supported on B 2 R ⊂ C n B_{2R} \subset\mathbb{C}^n B 2 R ⊂ C n , and ∣ Δ χ R ∣ ≤ C R − 2 |\Delta\chi_R| \le CR^{-2} ∣Δ χ R ∣ ≤ C R − 2 , we obtain
∫ B R ∣ S ∣ 2 ≤ 1 c 0 ∫ B 2 R χ R Δ Tr h ≤ C R 2 n − 2 . \int_{B_R}|S|^2 \le\frac{1}{c_0}\int_{B_{2R}}\chi_R\Delta\operatorname{Tr}h \le CR^{2n-2}. ∫ B R ∣ S ∣ 2 ≤ c 0 1 ∫ B 2 R χ R Δ Tr h ≤ C R 2 n − 2 .
Hence, by the mean value inequality for subharmonic functions, we obtain
∣ S ∣ 2 ( 0 ) ≤ C R 2 n ∫ B R ( 0 ) ∣ S ∣ 2 ≤ C R − 2 . |S|^2(0) \le\frac{C}{R^{2n}}\int_{B_R(0)}|S|^2 \le CR^{-2}. ∣ S ∣ 2 ( 0 ) ≤ R 2 n C ∫ B R ( 0 ) ∣ S ∣ 2 ≤ C R − 2 .
Hence ∣ S ∣ 2 ( 0 ) = 0 |S|^2(0) = 0 ∣ S ∣ 2 ( 0 ) = 0 . Since the same argument applies to any point, this means ∇ h = 0 \nabla h = 0 ∇ h = 0 , hence H H H is constant.
Step 2: The metrics H s H_s H s satisfy
− − 1 Λ ω s F H s = μ s ( E ) Id E , -\sqrt{-1}\,\Lambda_{\omega_s}F_{H_s} = \mu_s(E)\operatorname{Id}_E, − − 1 Λ ω s F H s = μ s ( E ) Id E ,
and with our assumptions, the eigenvalues of H s H_s H s with respect to H = H 1 H = H_1 H = H 1 are uniformly bounded above and below.
Then we may adapt the argument from [15 Lemma 19] or [43 ] , to obtain the uniform bound on ∇ H h s \nabla^H h_s ∇ H h s . Suppose the gradients are unbounded. Since the metrics ω s \omega_s ω s converge to ω 0 \omega_0 ω 0 uniformly, we can find a sequence h j h_j h j , and corresponding sequence of metrics ω j \omega_j ω j , such that
sup X ∣ ∇ H h j ∣ = ∣ ∇ H h j ∣ ( x j ) = M j → + ∞ , x j → x 0 . \sup_X|\nabla^H h_j| = |\nabla^H h_j|(x_j) = M_j \to+\infty,\quad x_j \to x_0. X sup ∣ ∇ H h j ∣ = ∣ ∇ H h j ∣ ( x j ) = M j → + ∞ , x j → x 0 .
We rescale the metrics in the usual way
h ^ j ( x ) = h j ( x j + x M j ) , ω ^ j ( x ) = ω j ( x j + x M j ) . \hat{h}_j(x) = h_j\left(x_j+\frac{x}{M_j}\right),\qquad\hat{\omega}_j(x) = \omega_j\left(x_j+\frac{x}{M_j}\right). h ^ j ( x ) = h j ( x j + M j x ) , ω ^ j ( x ) = ω j ( x j + M j x ) .
Note that ω ^ j \hat{\omega}_j ω ^ j converge uniformly to ω e u c \omega_{\mathrm{euc}} ω euc .
Now, the uniform equivalence and uniform convergence of the metrics ω s \omega_s ω s allow us to obtain uniform L 2 L^2 L 2 -estimates for h ^ j \hat{h}_j h ^ j (see [18 Theorem 9.11] ). Then h ^ j \hat{h}_j h ^ j (up to taking a subsequence) converge in C 1 C^1 C 1 to some metric H ^ 0 \hat{H}_0 H ^ 0 , which satisfies
Λ ω e u c F H ^ 0 = 0. \Lambda_{\omega_{\mathrm{euc}}}F_{\hat{H}_0}=0. Λ ω euc F H ^ 0 = 0.
Step 1 guarantees that H ^ 0 \hat{H}_0 H ^ 0 is constant. Hence ∇ H ^ 0 = 0 \nabla\hat{H}_0=0 ∇ H ^ 0 = 0 , which contradicts the fact that ∣ ∇ H h ^ j ∣ ( 0 ) = 1 \lvert\nabla^H\hat{h}_j\rvert(0)=1 ∣ ∇ H h ^ j ∣ ( 0 ) = 1 . Hence, the original family has bounded gradients, and there is a subsequence H j H_j H j , which converges to H 0 H_0 H 0 in C 1 C^1 C 1 .
Let H 0 = H h 0 H_0=Hh_0 H 0 = H h 0 . Then H 0 H_0 H 0 satisfies − 1 Λ ω 0 F H 0 = μ ω 0 ( E ) Id \sqrt{-1}\Lambda_{\omega_0}F_{H_0}=\mu_{\omega_0}(E)\operatorname{Id} − 1 Λ ω 0 F H 0 = μ ω 0 ( E ) Id weakly. Now the standard elliptic bootstrap implies that H 0 H_0 H 0 is smooth and Hermitian-Einstein. Clearly E E E is ω 0 \omega_0 ω 0 -stable, because it is stable with respect to other metrics ω s \omega_s ω s . ▪
Yang-Mills, Hermitian-Yang-Mills functionals, and their generalizations.
Let A A A be an integrable connection on a hermitian bundle ( E , H ) (E,H) ( E , H ) over a compact Kähler manifold. We define the Yang-Mills (YM) functional as
YM ω ( A ) : = ∫ X ∣ F A ∣ H 2 ω n n ! . \operatorname{YM}_{\omega}(A) := \int_X |F_A|_H^2 \frac{\omega^n}{n!}. YM ω ( A ) := ∫ X ∣ F A ∣ H 2 n ! ω n .
The Hermitian Yang-Mills (HYM) functional is defined as follows:
HYM ω ( A ) : = ∫ X ∣ Λ ω F A ∣ H 2 ω n n ! . \operatorname{HYM}_{\omega}(A) := \int_X |\Lambda_{\omega}F_A|_H^2 \frac{\omega^n}{n!}. HYM ω ( A ) := ∫ X ∣ Λ ω F A ∣ H 2 n ! ω n .
Hence YM ω \operatorname{YM}_{\omega} YM ω and HYM ω \operatorname{HYM}_{\omega} HYM ω have the same critical points. They are related by the following identity (see [28 ] ):
YM ω ( A ) = HYM ω ( A ) + 4 π 2 ( n − 2 ) ! ∫ X ( 2 c 2 ( E ) − c 1 2 ( E ) ) ∧ ω n − 2 . (3) \operatorname{YM}_{\omega}(A) = \operatorname{HYM}_{\omega}(A) + \frac{4\pi^2}{(n-2)!}\int_X \left(2c_2(E)-c_1^2(E)\right)\wedge\omega^{n-2}. \tag*{(3)} YM ω ( A ) = HYM ω ( A ) + ( n − 2 )! 4 π 2 ∫ X ( 2 c 2 ( E ) − c 1 2 ( E ) ) ∧ ω n − 2 . (3)
In fact, we can reformulate this as follows:
YM ω ( A ) = ∥ − 1 Λ ω F A − μ ω ( E ) Id ∥ L 2 ( X , ω ) 2 + r μ ω ( E ) 2 V ω (4) \operatorname{YM}_{\omega}(A) = \left\|\sqrt{-1}\Lambda_{\omega}F_A-\mu_{\omega}(E)\operatorname{Id}\right\|^2_{L^2(X,\omega)}+r\mu_{\omega}(E)^2V_{\omega} \tag*{(4)} YM ω ( A ) = − 1 Λ ω F A − μ ω ( E ) Id L 2 ( X , ω ) 2 + r μ ω ( E ) 2 V ω (4)
+ 4 π 2 ( n − 2 ) ! ∫ X ( 2 c 2 ( E ) − c 1 2 ( E ) ) ∧ ω n − 2 . (5) +\frac{4\pi^2}{(n-2)!}\int_X \left(2c_2(E)-c_1^2(E)\right)\wedge\omega^{n-2}. \tag*{(5)} + ( n − 2 )! 4 π 2 ∫ X ( 2 c 2 ( E ) − c 1 2 ( E ) ) ∧ ω n − 2 . (5)
Moreover, critical points of YM ω \operatorname{YM}_{\omega} YM ω on the set A H 1 , 1 \mathcal{A}^{1,1}_H A H 1 , 1 of integrable hermitian connections are generalized Hermitian Yang-Mills connections (see [28 ] ). In the next sections we will need the following useful lemma.
Proposition 2.9: Let ω s \omega_s ω s be a family of Kähler metrics, that satisfy our assumptions, and let H s H_s H s be the family of the Hermitian-Einstein metrics on E E E with respect to ω s \omega_s ω s . Then
lim s → 0 YM ω s ( A H s ) = lim s → 0 YM ω 0 ( A H s ) \lim_{s\to0}\operatorname{YM}_{\omega_s}(A_{H_s})=\lim_{s\to0}\operatorname{YM}_{\omega_0}(A_{H_s}) s → 0 lim YM ω s ( A H s ) = s → 0 lim YM ω 0 ( A H s ) = r μ ω 0 2 ( E ) V ω 0 + 4 π 2 ( n − 2 ) ! ∫ X ( 2 c 2 ( E ) − c 1 2 ( E ) ) ∧ ω 0 n − 2 , =r\mu_{\omega_0}^2(E)V_{\omega_0}+\frac{4\pi^2}{(n-2)!}\int_X\left(2c_2(E)-c_1^2(E)\right)\wedge\omega_0^{n-2}, = r μ ω 0 2 ( E ) V ω 0 + ( n − 2 )! 4 π 2 ∫ X ( 2 c 2 ( E ) − c 1 2 ( E ) ) ∧ ω 0 n − 2 ,
and
lim s → 0 HYM ω s ( A H s ) = lim s → 0 HYM ω 0 ( A H s ) = r μ ω 0 2 ( E ) V ω 0 . \lim_{s\to0}\operatorname{HYM}_{\omega_s}(A_{H_s})=\lim_{s\to0}\operatorname{HYM}_{\omega_0}(A_{H_s})=r\mu_{\omega_0}^{2}(E)V_{\omega_0}. s → 0 lim HYM ω s ( A H s ) = s → 0 lim HYM ω 0 ( A H s ) = r μ ω 0 2 ( E ) V ω 0 .
In particular,
− 1 Λ ω 0 F A H s → s → 0 L 2 μ ω 0 ( E ) Id , \sqrt{-1}\Lambda_{\omega_0}F_{A_{H_s}}\xrightarrow[s\to0]{L^2}\mu_{\omega_0}(E)\operatorname{Id}, − 1 Λ ω 0 F A H s L 2 s → 0 μ ω 0 ( E ) Id ,
and
lim s → 0 YM ω s ( A H s ) = inf A ∈ A H 1 , 1 YM ω 0 ( A ) . \lim_{s\to0}\operatorname{YM}_{\omega_s}(A_{H_s})=\inf_{A\in\mathcal{A}^{1,1}_{H}}\operatorname{YM}_{\omega_0}(A). s → 0 lim YM ω s ( A H s ) = A ∈ A H 1 , 1 inf YM ω 0 ( A ) .
Proof: By 2.3 we have
YM ω s ( A H s ) = r μ ω s 2 ( E ) V ω s + 4 π 2 ( n − 2 ) ! ∫ X ( 2 c 2 ( E ) − c 1 2 ( E ) ) ∧ ω s n − 2 , \operatorname{YM}_{\omega_s}(A_{H_s})=r\mu_{\omega_s}^{2}(E)V_{\omega_s}+\frac{4\pi^2}{(n-2)!}\int_X\left(2c_2(E)-c_1^2(E)\right)\wedge\omega_s^{n-2}, YM ω s ( A H s ) = r μ ω s 2 ( E ) V ω s + ( n − 2 )! 4 π 2 ∫ X ( 2 c 2 ( E ) − c 1 2 ( E ) ) ∧ ω s n − 2 ,
and
YM ω 0 ( A ) = HYM ω 0 ( A ) + 4 π 2 ( n − 2 ) ! ∫ X ( 2 c 2 ( E ) − c 1 2 ( E ) ) ∧ ω 0 n − 2 . \operatorname{YM}_{\omega_0}(A)=\operatorname{HYM}_{\omega_0}(A)+\frac{4\pi^2}{(n-2)!}\int_X\left(2c_2(E)-c_1^2(E)\right)\wedge\omega_0^{n-2}. YM ω 0 ( A ) = HYM ω 0 ( A ) + ( n − 2 )! 4 π 2 ∫ X ( 2 c 2 ( E ) − c 1 2 ( E ) ) ∧ ω 0 n − 2 .
The integral ∫ X ( 2 c 2 ( E ) − c 1 2 ( E ) ) ∧ ω s n − 2 \int_X\left(2c_2(E)-c_1^2(E)\right)\wedge\omega_s^{n-2} ∫ X ( 2 c 2 ( E ) − c 1 2 ( E ) ) ∧ ω s n − 2 and μ ω s ( E ) \mu_{\omega_s}(E) μ ω s ( E ) are both continuous in s s s . Next, YM ω s ( A H s ) \operatorname{YM}_{\omega_s}(A_{H_s}) YM ω s ( A H s ) is uniformly equivalent to YM ω 0 ( A H s ) \operatorname{YM}_{\omega_0}(A_{H_s}) YM ω 0 ( A H s ) . More precisely, there are ε s > 0 \varepsilon_s>0 ε s > 0 , ε s → 0 \varepsilon_s\to0 ε s → 0 , such that
( 1 − ε s ) YM ω s ( A H s ) ≤ YM ω 0 ( A H s ) ≤ ( 1 + ε s ) YM ω s ( A H s ) . (1-\varepsilon_s)\operatorname{YM}_{\omega_s}(A_{H_s})\leq\operatorname{YM}_{\omega_0}(A_{H_s})\leq(1+\varepsilon_s)\operatorname{YM}_{\omega_s}(A_{H_s}). ( 1 − ε s ) YM ω s ( A H s ) ≤ YM ω 0 ( A H s ) ≤ ( 1 + ε s ) YM ω s ( A H s ) .
Now, the statement follows from the fact that
lim s → 0 YM ω s ( A H s ) = r μ ω 0 2 ( E ) V ω 0 + 4 π 2 ( n − 2 ) ! ∫ X ( 2 c 2 ( E ) − c 1 2 ( E ) ) ∧ ω 0 n − 2 , \lim_{s\to0}\operatorname{YM}_{\omega_s}(A_{H_s})=r\mu_{\omega_0}^{2}(E)V_{\omega_0}+\frac{4\pi^2}{(n-2)!}\int_X\left(2c_2(E)-c_1^2(E)\right)\wedge\omega_0^{n-2}, s → 0 lim YM ω s ( A H s ) = r μ ω 0 2 ( E ) V ω 0 + ( n − 2 )! 4 π 2 ∫ X ( 2 c 2 ( E ) − c 1 2 ( E ) ) ∧ ω 0 n − 2 ,
and 2.4.
Alternatively,
∥ − 1 Λ ω 0 F H s − μ ω 0 ( E ) Id ∥ L 2 ≤ C ∥ ω s − ω 0 ∥ C 0 ∥ F H s ∥ L 2 + C ∣ μ ω 0 ( E ) − μ ω s ( E ) ∣ , \left\|\sqrt{-1}\Lambda_{\omega_0}F_{H_s}-\mu_{\omega_0}(E)\operatorname{Id}\right\|_{L^2}\leq C\|\omega_s-\omega_0\|_{C^0}\|F_{H_s}\|_{L^2}+C\left|\mu_{\omega_0}(E)-\mu_{\omega_s}(E)\right|, − 1 Λ ω 0 F H s − μ ω 0 ( E ) Id L 2 ≤ C ∥ ω s − ω 0 ∥ C 0 ∥ F H s ∥ L 2 + C ∣ μ ω 0 ( E ) − μ ω s ( E ) ∣ ,
which tends to zero as s → 0 s\to0 s → 0 . Hence
− 1 Λ ω 0 F A H s → s → 0 L 2 μ ω 0 ( E ) Id . \sqrt{-1}\Lambda_{\omega_0}F_{A_{H_s}}\xrightarrow[s\to0]{L^2}\mu_{\omega_0}(E)\operatorname{Id}. − 1 Λ ω 0 F A H s L 2 s → 0 μ ω 0 ( E ) Id .
The last statement immediately follows from this observation. ∎
Finally, we recall the definition of generalized Hermitian-Yang-Mills functionals, introduced in [3 , 11 ] .
Definition 2.10: Let α ∈ [ 1 ; + ∞ ) \alpha\in[1;+\infty) α ∈ [ 1 ; + ∞ ) , and Φ α ( μ ) = ∑ γ = 1 r ∣ μ γ ∣ α \Phi_\alpha(\mu) = \sum_{\gamma=1}^{r} |\mu_\gamma|^\alpha Φ α ( μ ) = ∑ γ = 1 r ∣ μ γ ∣ α . We define HYM α , ω ( A ) \operatorname{HYM}_{\alpha,\omega}(A) HYM α , ω ( A ) as
HYM α , ω ( A ) : = ∫ X Φ α ( − 1 Λ ω F A ) ω n n ! . \operatorname{HYM}_{\alpha,\omega}(A) := \int_X \Phi_\alpha\left(\sqrt{-1}\Lambda_\omega F_A\right)\frac{\omega^n}{n!}. HYM α , ω ( A ) := ∫ X Φ α ( − 1 Λ ω F A ) n ! ω n . For any N ∈ R N \in\mathbb{R} N ∈ R we similarly define HYM α , N , ω ( A ) \operatorname{HYM}_{\alpha,N,\omega}(A) HYM α , N , ω ( A ) as
HYM α , N , ω ( A ) : = ∫ X Φ α ( − 1 Λ ω F A + N Id ) ω n n ! . \operatorname{HYM}_{\alpha,N,\omega}(A) := \int_X \Phi_\alpha\left(\sqrt{-1}\Lambda_\omega F_A + N\operatorname{Id}\right)\frac{\omega^n}{n!}. HYM α , N , ω ( A ) := ∫ X Φ α ( − 1 Λ ω F A + N Id ) n ! ω n . These functionals clearly behave like L α L^\alpha L α -norm of Λ ω F A \Lambda_\omega F_A Λ ω F A . The main use of HYM α , N , ω \operatorname{HYM}_{\alpha,N,\omega} HYM α , N , ω functionals is the identification of the Harder-Narasimhan-Seshadri type of A A A .
Reduction to the method of Uhlenbeck and Yau
In this section we adapt the arguments of Uhlenbeck and Yau and construct a ω 0 \omega_0 ω 0 -destabilizing subsheaf of E E E with the slope equal to the slope of E E E . We start with the standard estimate:
Lemma 3.1: Under the conditions above there is a constant C > 0 C > 0 C > 0 , independent of s s s , such that the following holds:
Δ s log Tr h s ≥ − ∣ Λ s F H s ∣ − ∣ Λ s F 1 ∣ ≥ − C \Delta_s \log\operatorname{Tr} h_s \geq-|\Lambda_s F_{H_s}| - |\Lambda_s F_1| \geq-C Δ s log Tr h s ≥ − ∣ Λ s F H s ∣ − ∣ Λ s F 1 ∣ ≥ − C
Proof: The first inequality is a calculation in Siu’s book [43 1.9.2] . The second one follows from the fact that the metrics ω s \omega_s ω s are comparable, and H s H_s H s are Hermitian-Einstein. ■
We immediately obtain (via Moser iteration), that there is C > 0 C > 0 C > 0 , independent of s s s , such that
sup X Tr h s ≤ C ∫ X Tr h s ω 0 n n ! . (6) \sup_X \operatorname{Tr} h_s \leq C\int_X \operatorname{Tr} h_s\frac{\omega_0^n}{n!}. \tag*{(6)} X sup Tr h s ≤ C ∫ X Tr h s n ! ω 0 n . (6)
We rescale our relative endomorphism h s h_s h s :
h ~ s = h s sup X Tr h s . \widetilde{h}_s = \frac{h_s}{\sup_X \operatorname{Tr} h_s}. h s = sup X Tr h s h s .
Proposition 3.2: There is a constant C > 0 C > 0 C > 0 , such that for every s ∈ ( 0 ; 1 ] s \in(0;1] s ∈ ( 0 ; 1 ] and any σ ∈ ( 0 ; 1 ] \sigma\in(0;1] σ ∈ ( 0 ; 1 ] the endomorphisms h ~ s σ \widetilde{h}_s^\sigma h s σ are uniformly bounded by C C C in L 1 2 ( X , End ( E ) ) L^2_1(X,\operatorname{End}(E)) L 1 2 ( X , End ( E )) , i.e.
∥ h ~ s σ ∥ L 1 2 ( X , ω 0 ) ≤ C . \|\widetilde{h}_s^\sigma\|_{L^2_1(X,\omega_0)} \leq C. ∥ h s σ ∥ L 1 2 ( X , ω 0 ) ≤ C .
Proof: We recall a general identity relating the curvatures F = F H s F = F_{H_s} F = F H s and F 1 F_1 F 1 :
F j k ˉ − ( F 1 ) j k ˉ = − ∂ k ˉ ( h s − 1 ∇ j h s ) (7) F_{j\bar{k}} - (F_1)_{j\bar{k}} = -\partial_{\bar{k}}(h_s^{-1}\nabla_j h_s) \tag*{(7)} F j k ˉ − ( F 1 ) j k ˉ = − ∂ k ˉ ( h s − 1 ∇ j h s ) (7)
where ∇ \nabla ∇ is (always) the Chern connection for H : = H 1 H := H_1 H := H 1 (we omit the bundle indices).
Taking the H H H -inner product with h s σ h_s^\sigma h s σ , σ ∈ ( 0 ; 1 ] \sigma\in(0;1] σ ∈ ( 0 ; 1 ] gives:
− g s j k ˉ ⟨ ∂ k ˉ ( h s − 1 ∇ j h s ) , h s σ ⟩ = ⟨ − 1 Λ ω s F s − μ ω s Id , h s σ ⟩ − ⟨ − 1 Λ ω s F 1 − μ ω s Id , h s σ ⟩ (8) -g_s^{j\bar{k}}\left\langle\partial_{\bar{k}}(h_s^{-1}\nabla_jh_s),h_s^\sigma\right\rangle
=\left\langle\sqrt{-1}\Lambda_{\omega_s}F_s-\mu_{\omega_s}\operatorname{Id},h_s^\sigma\right\rangle-\left\langle\sqrt{-1}\Lambda_{\omega_s}F_1-\mu_{\omega_s}\operatorname{Id},h_s^\sigma\right\rangle \tag*{(8)} − g s j k ˉ ⟨ ∂ k ˉ ( h s − 1 ∇ j h s ) , h s σ ⟩ = ⟨ − 1 Λ ω s F s − μ ω s Id , h s σ ⟩ − ⟨ − 1 Λ ω s F 1 − μ ω s Id , h s σ ⟩ (8)
The first summand on the right hand side is zero. The left hand side can be rearranged as follows:
− g s j k ˉ ⟨ ∂ k ˉ ( h s − 1 ∇ j h s ) , h s σ ⟩ = − g s j k ˉ ∂ k ˉ ⟨ h s − 1 ∇ j h s , h s σ ⟩ + g s j k ˉ ⟨ h s − 1 ∇ j h s , ∇ k h s σ ⟩ (9) -g_s^{j\bar{k}}\left\langle\partial_{\bar{k}}(h_s^{-1}\nabla_jh_s),h_s^\sigma\right\rangle
=-g_s^{j\bar{k}}\partial_{\bar{k}}\left\langle h_s^{-1}\nabla_jh_s,h_s^\sigma\right\rangle+g_s^{j\bar{k}}\left\langle h_s^{-1}\nabla_jh_s,\nabla_kh_s^\sigma\right\rangle \tag*{(9)} − g s j k ˉ ⟨ ∂ k ˉ ( h s − 1 ∇ j h s ) , h s σ ⟩ = − g s j k ˉ ∂ k ˉ ⟨ h s − 1 ∇ j h s , h s σ ⟩ + g s j k ˉ ⟨ h s − 1 ∇ j h s , ∇ k h s σ ⟩ (9)
= − g s j k ˉ ∂ k ˉ ⟨ ∇ j h s , h s σ − 1 ⟩ + g s j k ˉ ⟨ h s − 1 ∇ j h s , ∇ k h s σ ⟩ (10) =-g_s^{j\bar{k}}\partial_{\bar{k}}\left\langle\nabla_jh_s,h_s^{\sigma-1}\right\rangle+g_s^{j\bar{k}}\left\langle h_s^{-1}\nabla_jh_s,\nabla_kh_s^\sigma\right\rangle \tag*{(10)} = − g s j k ˉ ∂ k ˉ ⟨ ∇ j h s , h s σ − 1 ⟩ + g s j k ˉ ⟨ h s − 1 ∇ j h s , ∇ k h s σ ⟩ (10)
≥ − 1 σ Δ s Tr h s σ + ∣ h s − σ / 2 ∇ h s σ ∣ ω s 2 , (11) \geq-\frac{1}{\sigma}\Delta_s\operatorname{Tr}h_s^\sigma+\left|h_s^{-\sigma/2}\nabla h_s^\sigma\right|_{\omega_s}^{2}, \tag*{(11)} ≥ − σ 1 Δ s Tr h s σ + h s − σ /2 ∇ h s σ ω s 2 , (11)
and the last inequality follows from calculations of Uhlenbeck and Yau.
We obtain the following inequality:
− 1 σ Δ s Tr h s σ + ∣ h s − σ / 2 ∇ h s σ ∣ ω s 2 ≤ − ⟨ − 1 Λ ω s F H 1 − μ ω s Id , h s σ ⟩ , (12) -\frac{1}{\sigma}\Delta_s\operatorname{Tr}h_s^\sigma+\left|h_s^{-\sigma/2}\nabla h_s^\sigma\right|_{\omega_s}^{2}
\leq-\left\langle\sqrt{-1}\Lambda_{\omega_s}F_{H_1}-\mu_{\omega_s}\operatorname{Id},h_s^\sigma\right\rangle, \tag*{(12)} − σ 1 Δ s Tr h s σ + h s − σ /2 ∇ h s σ ω s 2 ≤ − ⟨ − 1 Λ ω s F H 1 − μ ω s Id , h s σ ⟩ , (12)
and, by the rescaling,
− 1 σ Δ s Tr h ~ s σ + ∣ h ~ s − σ / 2 ∇ h ~ s σ ∣ ω s 2 ≤ − ⟨ − 1 Λ ω s F H 1 − μ ω s Id , h ~ s σ ⟩ . (13) -\frac{1}{\sigma}\Delta_s\operatorname{Tr}\widetilde{h}_s^\sigma+\left|\widetilde{h}_s^{-\sigma/2}\nabla\widetilde{h}_s^\sigma\right|_{\omega_s}^{2}
\leq-\left\langle\sqrt{-1}\Lambda_{\omega_s}F_{H_1}-\mu_{\omega_s}\operatorname{Id},\widetilde{h}_s^\sigma\right\rangle. \tag*{(13)} − σ 1 Δ s Tr h s σ + h s − σ /2 ∇ h s σ ω s 2 ≤ − ⟨ − 1 Λ ω s F H 1 − μ ω s Id , h s σ ⟩ . (13)
Integrating against ω s n \omega_s^n ω s n , and using the equivalence of Kähler metrics together with the fact that the eigenvalues of h ~ s σ \tilde h_s^\sigma h ~ s σ are at most 1, we get the following inequality:
∫ X ∣ ∇ h ~ s σ ∣ 2 ω 0 n n ! ≤ C , (14) \int_X \lvert\nabla\tilde h_s^\sigma\rvert^2 \frac{\omega_0^n}{n!} \leq C, \tag*{(14)} ∫ X ∣ ∇ h ~ s σ ∣ 2 n ! ω 0 n ≤ C , (14)
where C > 0 C>0 C > 0 depends only on ( X , ω 0 , ω s , H 1 ) (X,\omega_0,\omega_s,H_1) ( X , ω 0 , ω s , H 1 ) (and note that ω 0 \omega_0 ω 0 and ω s \omega_s ω s are equivalent). Q.E.D.
■
The result of Proposition 3.2 enables the argument of Uhlenbeck and Yau. Let s j ⟶ j → + ∞ 0 s_j \longrightarrow_{j\to+\infty} 0 s j ⟶ j → + ∞ 0 be any sequence, such that sup X Tr h s j → + ∞ \sup_X \operatorname{Tr} h_{s_j} \to+\infty sup X Tr h s j → + ∞ . Let h j = h s j h_j=h_{s_j} h j = h s j , and ω j = ω s j \omega_j=\omega_{s_j} ω j = ω s j . We have the following:
Proposition 3.3: There is a nonnegative self-adjoint element H ∈ L 1 2 ( X , End ( E ) ) H\in L_1^2(X,\operatorname{End}(E)) H ∈ L 1 2 ( X , End ( E )) , such that for any rational σ ∈ ( 0 ; 1 ] \sigma\in(0;1] σ ∈ ( 0 ; 1 ] the sequence (maybe after taking a subsequence) h ~ j σ \tilde h_j^\sigma h ~ j σ converges to H σ H^\sigma H σ in L 2 L^2 L 2 (and a.e.), and ∇ h ~ j σ \nabla\tilde h_j^\sigma ∇ h ~ j σ converges to ∇ H σ \nabla H^\sigma ∇ H σ in L 2 L^2 L 2 weakly. Moreover, there exists a decreasing sequence σ ∈ Q ∩ ( 0 ; 1 ] \sigma\in\mathbb{Q}\cap(0;1] σ ∈ Q ∩ ( 0 ; 1 ] , such that π = lim σ → 0 ( Id − H σ ) \pi=\lim_{\sigma\to0}(\operatorname{Id}-H^\sigma) π = lim σ → 0 ( Id − H σ ) . The limit π \pi π lies in L 1 2 ( X , ω 0 ) L_1^2(X,\omega_0) L 1 2 ( X , ω 0 ) , and satisfies the following identities
π = π 2 = π ∗ \pi=\pi^2=\pi^* π = π 2 = π ∗ ( Id − π ) ∂ ˉ π = 0 a.e. (\operatorname{Id}-\pi)\bar\partial\pi=0\ \text{a.e.} ( Id − π ) ∂ ˉ π = 0 a.e. and defines a proper coherent subsheaf F \mathcal{F} F of $E.
Proof: This is a standard fact, but we include the proof for the reader’s convenience. The author learned it from Professor Duong H. Phong.
The convergence statements and equality π = π 2 = π ∗ \pi=\pi^2=\pi^* π = π 2 = π ∗ are trivial. The difficulty is the second one. Note that ( Id − π ) ∂ ˉ π = − ( π ∇ ( Id − π ) ) ∗ (\operatorname{Id}-\pi)\bar\partial\pi=-(\pi\nabla(\operatorname{Id}-\pi))^* ( Id − π ) ∂ ˉ π = − ( π ∇ ( Id − π ) ) ∗ , so we prove that the right hand side is zero.
We show, that
π ∇ ( Id − π ) = lim ( Id − h ~ j σ ) ∇ h ~ j σ . \pi\nabla(\operatorname{Id}-\pi)=\lim(\operatorname{Id}-\tilde h_j^\sigma)\nabla\tilde h_j^\sigma. π ∇ ( Id − π ) = lim ( Id − h ~ j σ ) ∇ h ~ j σ .
Here, of course, the equality is vague. We shall take three limits, and the order of limits matters.
First of all, from the Hölder inequality we have the weak convergence of ( Id − h ~ j s ) ∇ h ~ j σ (\operatorname{Id}-\widetilde{h}_j^s)\nabla\widetilde{h}_j^\sigma ( Id − h j s ) ∇ h j σ to ( Id − H s ) ∇ H σ (\operatorname{Id}-H^s)\nabla H^\sigma ( Id − H s ) ∇ H σ . Indeed, for any ψ ∈ L 2 ( X , Λ X 1 ⊗ End ( E ) ) \psi\in L^2(X,\Lambda_X^1 \otimes\operatorname{End}(E)) ψ ∈ L 2 ( X , Λ X 1 ⊗ End ( E )) we have
⟨ ( Id − h ~ j s ) ∇ h ~ j σ , ψ ⟩ L 2 = ⟨ ( Id − H s ) ∇ h ~ j σ , ψ ⟩ L 2 + ⟨ ( H s − h ~ j s ) ∇ h ~ j σ , ψ ⟩ L 2 , \left\langle(\operatorname{Id}-\widetilde{h}_j^s)\nabla\widetilde{h}_j^\sigma,\psi\right\rangle_{L^2}
=
\left\langle(\operatorname{Id}-H^s)\nabla\widetilde{h}_j^\sigma,\psi\right\rangle_{L^2}
+
\left\langle(H^s-\widetilde{h}_j^s)\nabla\widetilde{h}_j^\sigma,\psi\right\rangle_{L^2}, ⟨ ( Id − h j s ) ∇ h j σ , ψ ⟩ L 2 = ⟨ ( Id − H s ) ∇ h j σ , ψ ⟩ L 2 + ⟨ ( H s − h j s ) ∇ h j σ , ψ ⟩ L 2 ,
hence we have the weak convergence, by the dominated convergence theorem.
Next, since every eigenvalue μ α \mu_\alpha μ α of h ~ j \widetilde{h}_j h j lies between 0 and 1, it satisfies the inequality
1 − μ α s ≤ 2 s 2 s + σ μ α − σ / 2 , α = 1 , … , r . 1-\mu_\alpha^s \le\frac{2s}{2s+\sigma}\mu_\alpha^{-\sigma/2}, \qquad\alpha=1,\ldots,r. 1 − μ α s ≤ 2 s + σ 2 s μ α − σ /2 , α = 1 , … , r .
This implies
∫ X ∣ ( Id − h ~ j s ) ∇ h ~ j σ ∣ 2 ω 0 n n ! ≤ ( 2 s 2 s + σ ) 2 ∫ X ∣ h ~ j − σ / 2 ∇ h ~ j σ ∣ 2 ω 0 n n ! ≤ C ( 2 s 2 s + σ ) 2 (15) \int_X \left|(\operatorname{Id}-\widetilde{h}_j^s)\nabla\widetilde{h}_j^\sigma\right|^2\frac{\omega_0^n}{n!}
\le
\left(\frac{2s}{2s+\sigma}\right)^2
\int_X \left|\widetilde{h}_j^{-\sigma/2}\nabla\widetilde{h}_j^\sigma\right|^2\frac{\omega_0^n}{n!}
\le
C\left(\frac{2s}{2s+\sigma}\right)^2 \tag*{(15)} ∫ X ( Id − h j s ) ∇ h j σ 2 n ! ω 0 n ≤ ( 2 s + σ 2 s ) 2 ∫ X h j − σ /2 ∇ h j σ 2 n ! ω 0 n ≤ C ( 2 s + σ 2 s ) 2 (15)
by 3.8.
Hence, we have
∫ X ∣ ( Id − H s ) ∇ H σ ∣ 2 ω 0 n n ! ≤ C ( 2 s 2 s + σ ) 2 , (16) \int_X \left|(\operatorname{Id}-H^s)\nabla H^\sigma\right|^2\frac{\omega_0^n}{n!}
\le
C\left(\frac{2s}{2s+\sigma}\right)^2, \tag*{(16)} ∫ X ∣ ( Id − H s ) ∇ H σ ∣ 2 n ! ω 0 n ≤ C ( 2 s + σ 2 s ) 2 , (16)
which implies, after taking the limit s → 0 s\to0 s → 0 ,
∫ X ∣ π ∇ H σ ∣ 2 ω 0 n n ! ≤ 0 (17) \int_X \left|\pi\nabla H^\sigma\right|^2\frac{\omega_0^n}{n!}\le0 \tag*{(17)} ∫ X ∣ π ∇ H σ ∣ 2 n ! ω 0 n ≤ 0 (17)
and
∫ X ∣ π ∇ ( Id − π ) ∣ 2 ω 0 n n ! ≤ 0. (18) \int_X \left|\pi\nabla(\operatorname{Id}-\pi)\right|^2\frac{\omega_0^n}{n!}\le0. \tag*{(18)} ∫ X ∣ π ∇ ( Id − π ) ∣ 2 n ! ω 0 n ≤ 0. (18)
This is because ∇ H σ \nabla H^\sigma ∇ H σ tends weakly to ∇ ( Id − π ) \nabla(\operatorname{Id}-\pi) ∇ ( Id − π ) . The fact that π \pi π defines a coherent subsheaf follows from the separate meromorphicity theorem (see [40 , 34 , 46 , 47 ] ). The fact that F \mathcal{F} F is nontrivial follows from the fact that det h ~ j \det\widetilde{h}_j det h j tends to zero, and F ≠ E \mathcal{F}\ne E F = E follows from 3.1. ■
Finally we need a version of the theorem of Uhlenbeck and Yau, which shows us that our π \pi π defines a subsheaf F \mathcal{F} F of E E E , which destabilizes E E E (with respect to ω 0 \omega_0 ω 0 ). The difficulty comes from the fact that our metric is not fixed. Moreover, it varies with the metric H j H_j H j in the fiber, so we cannot take the limit directly, as Uhlenbeck and Yau did.
Let ω j \omega_j ω j be the metric, corresponding to h j h_j h j , and V j = Vol ( X , ω j ) V_j = \operatorname{Vol}(X,\omega_j) V j = Vol ( X , ω j ) . We start with the standard Chern-Weil formula:
μ ω j ( F ) = 1 rank ( F ) V j { ∫ X ∖ Z Tr ( ( − 1 Λ ω j F H 1 − μ ω j ( E ) Id ) π ) ω j n n ! − ∫ X ∖ Z ∣ ∇ π ∣ ω j 2 ω j n n ! } + μ ω j ( E ) . (19) \mu_{\omega_j}(\mathcal{F}) = \frac{1}{\operatorname{rank}(\mathcal{F})V_j}\left\{\int_{X\setminus Z}\operatorname{Tr}\left((\sqrt{-1}\Lambda_{\omega_j}F_{H_1}-\mu_{\omega_j}(E)\operatorname{Id})\pi\right)\frac{\omega_j^n}{n!}-\int_{X\setminus Z}|\nabla\pi|_{\omega_j}^2\frac{\omega_j^n}{n!}\right\}+\mu_{\omega_j}(E). \tag*{(19)} μ ω j ( F ) = rank ( F ) V j 1 { ∫ X ∖ Z Tr ( ( − 1 Λ ω j F H 1 − μ ω j ( E ) Id ) π ) n ! ω j n − ∫ X ∖ Z ∣∇ π ∣ ω j 2 n ! ω j n } + μ ω j ( E ) . (19)
and Z = Z F Z=Z_{\mathcal{F}} Z = Z F is an analytic subset with codim C Z ≥ 2 \operatorname{codim}_{\mathbb{C}}Z\ge2 codim C Z ≥ 2 , where F ↪ E \mathcal{F}\hookrightarrow E F ↪ E fails to be a subbundle.
We have for all j j j :
Tr ( ( − 1 Λ ω j F H 1 − μ ω j ( E ) Id ) π ) = Tr ( ( − 1 Λ ω j F H j − μ ω j ( E ) Id ) π ) + Tr ( g j a b ˉ ∂ b ˉ ( h j − 1 ∇ a h j ) π ) = Tr ( g j a b ˉ ∂ b ˉ ( h j − 1 ∇ a h j ) ( π − Id ) ) + Tr ( g j a b ˉ ∂ b ˉ ( h j − 1 ∇ a h j ) ) = Tr ( g j a b ˉ ∂ b ˉ ( h j − 1 ∇ a h j ) ( π − Id ) ) + g j a b ˉ ∂ b ˉ Tr ( h j − 1 ∇ a h j ) = Tr ( g j a b ˉ ∂ b ˉ ( h j − 1 ∇ a h j ) ( π − Id ) ) + Δ j log det h j . \begin{aligned}
\operatorname{Tr}\left((\sqrt{-1}\Lambda_{\omega_j}F_{H_1}-\mu_{\omega_j}(E)\operatorname{Id})\pi\right)
&=\operatorname{Tr}\left((\sqrt{-1}\Lambda_{\omega_j}F_{H_j}-\mu_{\omega_j}(E)\operatorname{Id})\pi\right)+\operatorname{Tr}\left(g_j^{a\bar b}\partial_{\bar b}(h_j^{-1}\nabla_a h_j)\pi\right)\\
&=\operatorname{Tr}\left(g_j^{a\bar b}\partial_{\bar b}(h_j^{-1}\nabla_a h_j)(\pi-\operatorname{Id})\right)+\operatorname{Tr}\left(g_j^{a\bar b}\partial_{\bar b}(h_j^{-1}\nabla_a h_j)\right)\\
&=\operatorname{Tr}\left(g_j^{a\bar b}\partial_{\bar b}(h_j^{-1}\nabla_a h_j)(\pi-\operatorname{Id})\right)+g_j^{a\bar b}\partial_{\bar b}\operatorname{Tr}\left(h_j^{-1}\nabla_a h_j\right)\\
&=\operatorname{Tr}\left(g_j^{a\bar b}\partial_{\bar b}(h_j^{-1}\nabla_a h_j)(\pi-\operatorname{Id})\right)+\Delta_j\log\det h_j.
\end{aligned} Tr ( ( − 1 Λ ω j F H 1 − μ ω j ( E ) Id ) π ) = Tr ( ( − 1 Λ ω j F H j − μ ω j ( E ) Id ) π ) + Tr ( g j a b ˉ ∂ b ˉ ( h j − 1 ∇ a h j ) π ) = Tr ( g j a b ˉ ∂ b ˉ ( h j − 1 ∇ a h j ) ( π − Id ) ) + Tr ( g j a b ˉ ∂ b ˉ ( h j − 1 ∇ a h j ) ) = Tr ( g j a b ˉ ∂ b ˉ ( h j − 1 ∇ a h j ) ( π − Id ) ) + g j a b ˉ ∂ b ˉ Tr ( h j − 1 ∇ a h j ) = Tr ( g j a b ˉ ∂ b ˉ ( h j − 1 ∇ a h j ) ( π − Id ) ) + Δ j log det h j .
The second summand has zero integral over X ∖ Z X\setminus Z X ∖ Z .
So, we have the following formula:
μ ω j ( F ) = 1 rank ( F ) V j ∫ X ∖ Z { − Tr ( g j a b ˉ ∂ b ˉ ( h j − 1 ∇ a h j ) ( Id − π ) ) − ∣ ∇ π ∣ ω j 2 } ω j n n ! + μ ω j ( E ) . (20) \mu_{\omega_j}(\mathcal{F}) = \frac{1}{\operatorname{rank}(\mathcal{F})V_j}\int_{X\setminus Z}\left\{-\operatorname{Tr}\left(g_j^{a\bar b}\partial_{\bar b}(h_j^{-1}\nabla_a h_j)(\operatorname{Id}-\pi)\right)-|\nabla\pi|_{\omega_j}^2\right\}\frac{\omega_j^n}{n!}+\mu_{\omega_j}(E). \tag*{(20)} μ ω j ( F ) = rank ( F ) V j 1 ∫ X ∖ Z { − Tr ( g j a b ˉ ∂ b ˉ ( h j − 1 ∇ a h j ) ( Id − π ) ) − ∣∇ π ∣ ω j 2 } n ! ω j n + μ ω j ( E ) . (20)
Thus, we need to show that the following proposition holds:
Proposition 3.4: The following inequality holds:
lim inf j → + ∞ ∫ X { ⋯ } ≥ 0. \liminf_{j\to+\infty}\int_X\{\cdots\}\ge0. j → + ∞ lim inf ∫ X { ⋯ } ≥ 0.
Proof:
We start with a simple rearrangement:
− ∫ X \ Z Tr ( g j a ˉ b ˉ ∂ b ˉ ( h j − 1 ∇ a h j ) ( Id − π ) ) ω j n n ! = − ∫ X \ Z Tr ( g j a ˉ b ˉ ∂ b ˉ ( h j − 1 ∇ a h j ) ( h ~ j σ ) ) ω j n n ! + ∫ X \ Z Tr ( g j a ˉ b ˉ ∂ b ˉ ( h j − 1 ∇ a h j ) ( h ~ j σ − Id + π ) ) ω j n n ! = ( I ) + ( I I ) . (21) \begin{aligned}
-\int_{X\backslash Z}\operatorname{Tr}(g_j^{\bar a\bar b}\partial_{\bar b}(h_j^{-1}\nabla_a h_j)(\operatorname{Id}-\pi))\frac{\omega_j^n}{n!}
=-\int_{X\backslash Z}\operatorname{Tr}(g_j^{\bar a\bar b}\partial_{\bar b}(h_j^{-1}\nabla_a h_j)(\widetilde h_j^\sigma))\frac{\omega_j^n}{n!} \\
&+\int_{X\backslash Z}\operatorname{Tr}(g_j^{\bar a\bar b}\partial_{\bar b}(h_j^{-1}\nabla_a h_j)(\widetilde h_j^\sigma-\operatorname{Id}+\pi))\frac{\omega_j^n}{n!}=(\mathbf{I})+(\mathbf{II}). \tag*{(21)}
\end{aligned} − ∫ X \ Z Tr ( g j a ˉ b ˉ ∂ b ˉ ( h j − 1 ∇ a h j ) ( Id − π )) n ! ω j n = − ∫ X \ Z Tr ( g j a ˉ b ˉ ∂ b ˉ ( h j − 1 ∇ a h j ) ( h j σ )) n ! ω j n + ∫ X \ Z Tr ( g j a ˉ b ˉ ∂ b ˉ ( h j − 1 ∇ a h j ) ( h j σ − Id + π )) n ! ω j n = ( I ) + ( II ) . (21)
Let us deal with (I) . First, note that
− ∫ X \ Z Tr ( g j a ˉ b ˉ ∂ b ˉ ( h j − 1 ∇ a h j ) ( h ~ j σ ) ) ω j n n ! ≥ ∫ X \ Z ∣ ∇ h ~ j − σ / 2 ∇ h ~ j σ ∣ 2 ω j n n ! ≥ ∫ X \ Z ∣ ∇ h ~ j σ ∣ 2 ω j n n ! . -\int_{X\backslash Z}\operatorname{Tr}(g_j^{\bar a\bar b}\partial_{\bar b}(h_j^{-1}\nabla_a h_j)(\widetilde h_j^\sigma))\frac{\omega_j^n}{n!}\geq\int_{X\backslash Z}|\nabla\widetilde h_j^{-\sigma/2}\nabla\widetilde h_j^\sigma|^2\frac{\omega_j^n}{n!}\geq\int_{X\backslash Z}|\nabla\widetilde h_j^\sigma|^2\frac{\omega_j^n}{n!}. − ∫ X \ Z Tr ( g j a ˉ b ˉ ∂ b ˉ ( h j − 1 ∇ a h j ) ( h j σ )) n ! ω j n ≥ ∫ X \ Z ∣∇ h j − σ /2 ∇ h j σ ∣ 2 n ! ω j n ≥ ∫ X \ Z ∣∇ h j σ ∣ 2 n ! ω j n .
by the proof of Proposition 3.2 .
Lemma 3.5: We have the following inequality:
lim σ → 0 inf lim inf j → + ∞ ∫ X \ Z ∣ ∇ h ~ j σ ∣ ω j 2 ω j n n ! ≥ ∫ X \ Z ∣ ∇ π ∣ ω 0 2 ω 0 n n ! . (22) \lim_{\sigma\to0}\inf\liminf_{j\to+\infty}\int_{X\backslash Z}|\nabla\widetilde h_j^\sigma|_{\omega_j}^2\frac{\omega_j^n}{n!}\geq\int_{X\backslash Z}|\nabla\pi|_{\omega_0}^2\frac{\omega_0^n}{n!}. \tag*{(22)} σ → 0 lim inf j → + ∞ lim inf ∫ X \ Z ∣∇ h j σ ∣ ω j 2 n ! ω j n ≥ ∫ X \ Z ∣∇ π ∣ ω 0 2 n ! ω 0 n . (22)
Proof: (of Lemma 3.5 ) This is the place where we need the uniform convergence of ω s \omega_s ω s to ω 0 \omega_0 ω 0 as s → 0 s\to0 s → 0 . The uniform convergence guarantees, that there are positive ε j → 0 \varepsilon_j\to0 ε j → 0 , such that
( 1 − ε j ) ∫ X \ Z ∣ ∇ h ~ j σ ∣ ω 0 2 ω 0 n n ! ≤ ∫ X \ Z ∣ ∇ h ~ j σ ∣ ω j 2 ω j n n ! ≤ ( 1 + ε j ) ∫ X \ Z ∣ ∇ h ~ j σ ∣ ω 0 2 ω 0 n n ! . (23) (1-\varepsilon_j)\int_{X\backslash Z}|\nabla\widetilde h_j^\sigma|_{\omega_0}^2\frac{\omega_0^n}{n!}\leq\int_{X\backslash Z}|\nabla\widetilde h_j^\sigma|_{\omega_j}^2\frac{\omega_j^n}{n!}\leq(1+\varepsilon_j)\int_{X\backslash Z}|\nabla\widetilde h_j^\sigma|_{\omega_0}^2\frac{\omega_0^n}{n!}. \tag*{(23)} ( 1 − ε j ) ∫ X \ Z ∣∇ h j σ ∣ ω 0 2 n ! ω 0 n ≤ ∫ X \ Z ∣∇ h j σ ∣ ω j 2 n ! ω j n ≤ ( 1 + ε j ) ∫ X \ Z ∣∇ h j σ ∣ ω 0 2 n ! ω 0 n . (23)
We immediately have
∫ X \ Z ∣ ∇ H σ ∣ ω 0 2 ω 0 n n ! ≤ lim inf j → + ∞ ∫ X \ Z ∣ ∇ h ~ j σ ∣ ω j 2 ω j n n ! . (24) \int_{X\backslash Z}|\nabla H^\sigma|_{\omega_0}^2\frac{\omega_0^n}{n!}\leq\liminf_{j\to+\infty}\int_{X\backslash Z}|\nabla\widetilde h_j^\sigma|_{\omega_j}^2\frac{\omega_j^n}{n!}. \tag*{(24)} ∫ X \ Z ∣∇ H σ ∣ ω 0 2 n ! ω 0 n ≤ j → + ∞ lim inf ∫ X \ Z ∣∇ h j σ ∣ ω j 2 n ! ω j n . (24)
Taking the limit lim σ → 0 \lim_{\sigma\to0} lim σ → 0 , we obtain the desired result. ■
Now we estimate (II):
∣ ( I I ) ∣ ≤ ∫ X ∖ Z ∣ Tr ( g j a b ˉ F a b ˉ ( h ~ j σ − Id + π ) ) ∣ ω j n n ! + ∫ X ∖ Z ∣ Tr ( g j a b ˉ F 1 b ˉ ( h ~ j σ − Id + π ) ) ∣ ω j n n ! (25) |(II)| \le\int_{X\setminus Z} \left|\operatorname{Tr}\left(g_j^{a\bar b}F_{a\bar b}(\widetilde h_j^\sigma-\operatorname{Id}+\pi)\right)\right|\frac{\omega_j^n}{n!}+\int_{X\setminus Z}\left|\operatorname{Tr}\left(g_j^{a\bar b}F_{1\bar b}(\widetilde h_j^\sigma-\operatorname{Id}+\pi)\right)\right|\frac{\omega_j^n}{n!} \tag*{(25)} ∣ ( I I ) ∣ ≤ ∫ X ∖ Z Tr ( g j a b ˉ F a b ˉ ( h j σ − Id + π ) ) n ! ω j n + ∫ X ∖ Z Tr ( g j a b ˉ F 1 b ˉ ( h j σ − Id + π ) ) n ! ω j n (25)
≤ C ∫ X ∖ Z ∣ h ~ j σ − Id + π ∣ ω 0 n n ! , (26) \le C\int_{X\setminus Z}\left|\widetilde h_j^\sigma-\operatorname{Id}+\pi\right|\frac{\omega_0^n}{n!}, \tag*{(26)} ≤ C ∫ X ∖ Z h j σ − Id + π n ! ω 0 n , (26)
and the RHS tends to zero, because h ~ j σ \widetilde h_j^\sigma h j σ tends a.e. to H σ \mathbb{H}^\sigma H σ as j → + ∞ j\to+\infty j → + ∞ , and H σ \mathbb{H}^\sigma H σ tends a.e. to Id − π \operatorname{Id}-\pi Id − π , as σ → 0 \sigma\to0 σ → 0 . This concludes the proof of Proposition 3.4 . ▪
We have shown that
lim j → + ∞ μ ω j ( F ) ≥ μ ω 0 ( E ) . \lim_{j\to+\infty}\mu_{\omega_j}(\mathcal{F})\ge\mu_{\omega_0}(E). j → + ∞ lim μ ω j ( F ) ≥ μ ω 0 ( E ) .
In fact, in our setting, we also have
lim j → + ∞ μ ω j ( F ) ≤ μ ω 0 ( E ) , \lim_{j\to+\infty}\mu_{\omega_j}(\mathcal{F})\le\mu_{\omega_0}(E), j → + ∞ lim μ ω j ( F ) ≤ μ ω 0 ( E ) ,
and this inequality holds for any subsheaf of E E E , since for any s > 0 s>0 s > 0 our bundle E E E is ω s \omega_s ω s -stable. Hence we have
μ ω 0 ( F ) = μ ω 0 ( E ) . \mu_{\omega_0}(\mathcal{F})=\mu_{\omega_0}(E). μ ω 0 ( F ) = μ ω 0 ( E ) .
Thus, combining these inequalities, we obtain the following proposition, which is the second part of Theorem 1.1 .
Proposition 3.6: If the family of relative endomorphisms h s h_s h s , normalized by
∫ X log det h s ω s n = 0 , \int_X\log\det h_s\omega_s^n=0, ∫ X log det h s ω s n = 0 , has unbounded traces Tr h s \operatorname{Tr}h_s Tr h s , then the bundle E E E is strictly ω 0 \omega_0 ω 0 -semistable. Moreover, a projection onto a destabilizing sheaf may be constructed by the Uhlenbeck-Yau construction.
d ω s n − 1 = 0 d\omega_s^{n-1}=0 d ω s n − 1 = 0 ) or Gauduchon (i.e. ∂ ∂ ˉ ω s n − 1 = 0 \partial\bar{\partial}\omega_s^{n-1}=0 ∂ ∂ ˉ ω s n − 1 = 0 ) metrics. The only place that should be modified is the proof of Proposition 2.6 . Since the uniform equivalence allows us to control the Sobolev constant, the usual Moser iteration argument shows that log det h s \log\det h_s log det h s is uniformly bounded.
Remark 3.8: Our arguments show, that the metrics H ~ j = H j sup X Tr h j \widetilde{H}_j=\frac{H_j}{\sup_X\operatorname{Tr}h_j} H j = s u p X Tr h j H j degenerate to some limiting metric on E / F E/F E / F over X ∖ Z X\setminus Z X ∖ Z . However, the convergence of these metrics is very weak, as one can see directly from our arguments above. In particular, H j H_j H j and H ~ j \widetilde{H}_j H j have the same Chern connections A H j A_{H_j} A H j , and curvatures F H j F_{H_j} F H j . We address the behaviour of the connections in the next section.
Uhlenbeck limits of connections
In this section we discuss the Uhlenbeck limits of the Chern connections of metrics H s H_s H s . While we have shown that the metrics themselves blow up, their connections still may have a limit (modulo gauge). However, there is a delicate point here: while there is a well-developed compactness theory for the Yang-Mills connections on Riemannian manifolds, this theory works under the assumption that the background Kähler (or just a Riemannian) metric is fixed. Our setting is different, as we allow the base metrics ω s \omega_s ω s to vary, so we shall show that Uhlenbeck compactness and other results hold in the setting of varying metrics as well. In order to do it, we need to impose the following conditions on ω s \omega_s ω s :
ω s \omega_s ω s are uniformly equivalent to ω 0 \omega_0 ω 0 ;
ω s \omega_s ω s converge in C 0 C^0 C 0 to ω 0 \omega_0 ω 0 , as s → 0 s\to0 s → 0 ;
There is a constant C > 0 C>0 C > 0 , such that ∣ Rm ( ω s ) ∣ ω s ≤ C \lvert\operatorname{Rm}(\omega_s)\rvert_{\omega_s}\le C ∣ Rm ( ω s ) ∣ ω s ≤ C .
The only new condition here is the bound for curvature. At some points, the bound for the curvature enters our estimates. We also need to control the uniform lower bound for the injectivity radii of ω s \omega_s ω s , and the bound for the curvature together with the first condition (which translates to uniform volume and diameter bounds) allow us to obtain such a bound.
Also, for higher regularity estimates, we need to control ∣ ∇ k Rm ( ω s ) ∣ |\nabla^{k}\operatorname{Rm}(\omega_s)| ∣ ∇ k Rm ( ω s ) ∣ . We will always state explicitly, when such bounds are required for our purposes.
Preliminaries: ε \varepsilon ε -regularity, monotonicity formula and the estimates of curvature
In this section we turn our attention to the limiting behaviour of the Chern connections of our Hermitian-Einstein metrics. First of all, we need to recall the monotonicity formula for the Yang-Mills connections, which we slightly reformulate for our purposes.
Theorem 4.1: (see [44 Thm. 2.1.2] and [36 ] ) Let ( X , ω s ) (X,\omega_s) ( X , ω s ) be a Kähler manifold with a family of Kähler metrics ω s \omega_s ω s . Assume that the metrics are equivalent, and their curvatures are uniformly bounded. Let ( E , H ) (E,H) ( E , H ) be a hermitian vector bundle over X X X , and A s A_s A s be a family of the integrable Hermitian Yang-Mills connections with respect to ω s \omega_s ω s . Then there is a positive number R 0 R_0 R 0 and a ≥ 0 a \ge0 a ≥ 0 , such that for all 0 < r ≤ R ≤ R 0 0 < r \le R \le R_0 0 < r ≤ R ≤ R 0 and for any s s s the following holds:
R 4 − 2 n e a R 2 ∫ B R , s ( p ) ∣ F A s ∣ 2 ω s n n ! − r 4 − 2 n e a r 2 ∫ B r , s ( p ) ∣ F A s ∣ 2 ω s n n ! (27) R^{4-2n}e^{aR^2}\int_{B_{R,s}(p)} |F_{A_s}|^2\frac{\omega_s^n}{n!}-r^{4-2n}e^{ar^2}\int_{B_{r,s}(p)} |F_{A_s}|^2\frac{\omega_s^n}{n!} \tag*{(27)} R 4 − 2 n e a R 2 ∫ B R , s ( p ) ∣ F A s ∣ 2 n ! ω s n − r 4 − 2 n e a r 2 ∫ B r , s ( p ) ∣ F A s ∣ 2 n ! ω s n (27) ≥ 2 ∫ B R , s ( p ) ∖ B r , s ( p ) ρ s 4 − 2 n e a ρ s 2 ∣ ι ∂ ρ s F A s ∣ 2 ω s n n ! (28) \ge2\int_{B_{R,s}(p)\setminus B_{r,s}(p)}\rho_s^{4-2n}e^{a\rho_s^2}|\iota_{\partial_{\rho_s}}F_{A_s}|^2\frac{\omega_s^n}{n!} \tag*{(28)} ≥ 2 ∫ B R , s ( p ) ∖ B r , s ( p ) ρ s 4 − 2 n e a ρ s 2 ∣ ι ∂ ρ s F A s ∣ 2 n ! ω s n (28) where B r , s ( p ) B_{r,s}(p) B r , s ( p ) denotes the ball of radius r r r in ( X , ω s ) (X,\omega_s) ( X , ω s ) , and ρ s \rho_s ρ s is the distance function with respect to ω s \omega_s ω s . Moreover, the equality holds if X = C n X=\mathbb{C}^n X = C n with standard flat metric and a = 0 a=0 a = 0 .
Next, we need the ε \varepsilon ε -regularity statement.
Proposition 4.2: Let ( X , ω s , E , H s ) (X,\omega_s,E,H_s) ( X , ω s , E , H s ) be as in our setting. There are ε 0 > 0 \varepsilon_0 > 0 ε 0 > 0 , C > 0 C > 0 C > 0 and R 0 > 0 R_0 > 0 R 0 > 0 , such that if for ε ≤ ε 0 \varepsilon\le\varepsilon_0 ε ≤ ε 0 and 0 < r < R 0 0 < r < R_0 0 < r < R 0 we have
r 4 − 2 n ∫ B r , s ( p ) ∣ F H s ∣ 2 ω 0 n n ! < ε , r^{4-2n}\int_{B_{r,s}(p)} |F_{H_s}|^2\frac{\omega_0^n}{n!}<\varepsilon, r 4 − 2 n ∫ B r , s ( p ) ∣ F H s ∣ 2 n ! ω 0 n < ε , then
∣ F H s ( p ) ∣ ≤ C r 2 ( r 4 − 2 n ∫ B r , 0 ( p ) ∣ F H s ∣ 2 ω 0 n n ! ) 1 2 \lvert F_{H_s}(p)\rvert\leq\frac{C}{r^2}\left(r^{4-2n}\int_{B_{r,0}(p)}\lvert F_{H_s}\rvert^2\frac{\omega_0^n}{n!}\right)^{\frac{1}{2}} ∣ F H s ( p )∣ ≤ r 2 C ( r 4 − 2 n ∫ B r , 0 ( p ) ∣ F H s ∣ 2 n ! ω 0 n ) 2 1
Proof: Since F H s F_{H_s} F H s satisfies the (Hermitian-)Yang–Mills equations (with respect to ω s \omega_s ω s ), and we have the uniform control of ∣ Rm ( ω s ) ∣ \lvert\operatorname{Rm}(\omega_s)\rvert ∣ Rm ( ω s )∣ , we have
Δ ω s ∣ F H s ∣ ≥ − C 1 ( ∣ F H s ∣ 2 + ∣ F H s ∣ ) , \Delta_{\omega_s}\lvert F_{H_s}\rvert\geq-C_1(\lvert F_{H_s}\rvert^2+\lvert F_{H_s}\rvert), Δ ω s ∣ F H s ∣ ≥ − C 1 (∣ F H s ∣ 2 + ∣ F H s ∣) ,
for some universal C 1 > 0 C_1>0 C 1 > 0 independent of s s s . Also, the monotonicity formula for the Yang–Mills connections holds (for fixed s s s ), so does the ε \varepsilon ε -regularity. Since ω s \omega_s ω s is uniformly equivalent to ω 0 \omega_0 ω 0 , we can replace ω s \omega_s ω s by ω 0 \omega_0 ω 0 by adjusting the constant on the right hand side. ■
Uhlenbeck compactness.
In this section we investigate the limits of Chern connections ∇ s \nabla^s ∇ s of the Hermitian-Einstein metrics H s H_s H s , as s s s approaches 0. First of all, we note that the Uhlenbeck compactness theorem does not work in our setting, since each ∇ H s \nabla^{H_s} ∇ H s is hermitian, but with respect to its own metric H s H_s H s . To fix this problem, we introduce w s = h s 1 / 2 w_s=h_s^{1/2} w s = h s 1/2 , and introduce new connections
∇ A s = d + A s = w s ∘ ∇ H s ∘ w s − 1 . \nabla_{A_s}=d+A_s=w_s\circ\nabla^{H_s}\circ w_s^{-1}. ∇ A s = d + A s = w s ∘ ∇ H s ∘ w s − 1 .
These connections are integrable and unitary with respect to fixed background metric H = H 1 H=H_1 H = H 1 . Moreover, each A s A_s A s is Hermitian Yang-Mills with respect to ω s \omega_s ω s . These connections enjoy the following obvious properties.
Lemma 4.3. We have the following formulas
∣ F H s ∣ H s 2 = ∣ F A s ∣ H 2 \lvert F_{H_s}\rvert_{H_s}^2=\lvert F_{A_s}\rvert_H^2 ∣ F H s ∣ H s 2 = ∣ F A s ∣ H 2 ∣ Λ ω F H s ∣ H s 2 = ∣ Λ ω F A s ∣ H 2 , \lvert\Lambda_{\omega}F_{H_s}\rvert_{H_s}^2=\lvert\Lambda_{\omega}F_{A_s}\rvert_H^2, ∣ Λ ω F H s ∣ H s 2 = ∣ Λ ω F A s ∣ H 2 , and the same holds for covariant derivatives of F A s F_{A_s} F A s . Here ω \omega ω is any fixed Kähler form on X X X .
The proof follows from the fact that w s w_s w s is a bundle isometry from ( E , H s ) (E,H_s) ( E , H s ) to ( E , H ) (E,H) ( E , H ) , so we omit it. In the following we drop subscript H H H from ∣ F A s ∣ H |F_{A_s}|_H ∣ F A s ∣ H to lighten the notation.
However, another difficulty arises in our course. In our setting, we have uniform control of ∣ Λ ω s F A s ∣ |\Lambda_{\omega_s}F_{A_s}| ∣ Λ ω s F A s ∣ , but we do not control ∣ Λ ω 0 F A s ∣ |\Lambda_{\omega_0}F_{A_s}| ∣ Λ ω 0 F A s ∣ . Even if we have fairly good control of our family of metrics, we cannot directly convert this into the estimate for ∣ Λ ω 0 F A s ∣ |\Lambda_{\omega_0}F_{A_s}| ∣ Λ ω 0 F A s ∣ in terms of ω s , ω 0 \omega_s,\omega_0 ω s , ω 0 and ∣ Λ ω s F A s ∣ |\Lambda_{\omega_s}F_{A_s}| ∣ Λ ω s F A s ∣ . So we cannot apply the Uhlenbeck compactness theorem directly in the form as it is stated, for example, in [38 Theorem 2.17] . So, we need to show, that for the family { A s } \{A_s\} { A s } there is a version of Uhlenbeck compactness theorem.
Before we proceed, we need the following definition.
Definition 4.4: Let { s j } \{s_j\} { s j } be a sequence, such that s j → 0 s_j \to0 s j → 0 as j → + ∞ j \to+\infty j → + ∞ . Write A j = A s j A_j=A_{s_j} A j = A s j , and the corresponding Kähler metric ω j = ω s j \omega_j=\omega_{s_j} ω j = ω s j .
Z a n = { p ∈ X | lim r → 0 lim inf j → + ∞ r 4 − 2 n ∫ B r , j ( p ) ∣ F A j ∣ 2 ω j n n ! ≥ ε 0 } (29) Z_{an}=\left\{p\in X\ \middle|\ \lim_{r\to0}\liminf_{j\to+\infty}r^{4-2n}\int_{B_{r,j}(p)}|F_{A_j}|^2\frac{\omega_j^n}{n!}\ge\varepsilon_0\right\} \tag*{(29)} Z an = { p ∈ X r → 0 lim j → + ∞ lim inf r 4 − 2 n ∫ B r , j ( p ) ∣ F A j ∣ 2 n ! ω j n ≥ ε 0 } (29) Here B r , j ( p ) B_{r,j}(p) B r , j ( p ) is a ball in ( X , ω j ) (X,\omega_j) ( X , ω j ) of radius r r r .
Of course, a priori, Z a n Z_{an} Z an depends on the chosen sequence. However, we omit the indication of such dependence, because it will be clear, which sequence we are working with.
We have the following proposition.
Proposition 4.5: One can replace ω s \omega_s ω s in this definition with ω 0 \omega_0 ω 0 . Moreover, Z a n Z_{an} Z an is closed and has Hausdorff dimension at most 2 n − 4 2n-4 2 n − 4 .
Proof: The statement follows from our assumption, that any sequence ω j \omega_j ω j converges to ω 0 \omega_0 ω 0 uniformly, and metrics are uniformly equivalent. In particular, B r , j ( p ) B_{r,j}(p) B r , j ( p ) converges to B r , 0 ( p ) B_{r,0}(p) B r , 0 ( p ) for any p p p . In particular, there are sequences of positive constants { ε j } \{\varepsilon_j\} { ε j } , { δ j } \{\delta_j\} { δ j } , both tending to zero as j → + ∞ j\to+\infty j → + ∞ , such that
( 1 − ε j ) ∫ B r − δ j , j ( p ) ∣ F A j ∣ 2 ω 0 n n ! ≤ ∫ B r , 0 ( p ) ∣ F A j ∣ 2 ω 0 n n ! ≤ ( 1 + ε j ) ∫ B r + δ j , j ( p ) ∣ F A j ∣ 2 ω j n n ! . (1-\varepsilon_j)\int_{B_{r-\delta_j,j}(p)}|F_{A_j}|^2\frac{\omega_0^n}{n!}\leq\int_{B_{r,0}(p)}|F_{A_j}|^2\frac{\omega_0^n}{n!}\leq(1+\varepsilon_j)\int_{B_{r+\delta_j,j}(p)}|F_{A_j}|^2\frac{\omega_j^n}{n!}. ( 1 − ε j ) ∫ B r − δ j , j ( p ) ∣ F A j ∣ 2 n ! ω 0 n ≤ ∫ B r , 0 ( p ) ∣ F A j ∣ 2 n ! ω 0 n ≤ ( 1 + ε j ) ∫ B r + δ j , j ( p ) ∣ F A j ∣ 2 n ! ω j n .
The statement about the Hausdorff dimension of Z a n Z_{an} Z an is standard, see for instance [33 , 44 ] . Q.E.D. ■
Now we can state the version of Uhlenbeck compactness theorem, which works in our setting.
Theorem 4.6. Let X X X be a compact complex manifold, and let { ω s } s ∈ [ 0 ; 1 ] \{\omega_s\}_{s\in[0;1]} { ω s } s ∈ [ 0 ; 1 ] be a family of Kähler metrics, which are uniformly equivalent to ω 0 \omega_0 ω 0 , converge uniformly to ω 0 \omega_0 ω 0 , and have uniformly bounded curvature. Let ( E , H ) (E,H) ( E , H ) be a hermitian vector bundle, and suppose { A s } \{A_s\} { A s } is a family of integrable hermitian connections. Assume that ∥ F A s ∥ L 2 ( X , ω 0 ) \lVert F_{A_s}\rVert_{L^2(X,\omega_0)} ∥ F A s ∥ L 2 ( X , ω 0 ) and ∣ Λ ω s F A s ∣ \lvert\Lambda_{\omega_s}F_{A_s}\rvert ∣ Λ ω s F A s ∣ are uniformly bounded. Fix p > 2 n p>2n p > 2 n . Then, there is a sequence s j → 0 s_j\to0 s j → 0 , such that the following holds for A j = A s j A_j=A_{s_j} A j = A s j :
There is a C 1 C^1 C 1 hermitian vector bundle ( E ∞ , H ∞ ) (E_\infty,H_\infty) ( E ∞ , H ∞ ) on X ∖ Z a n X\setminus Z_{an} X ∖ Z an , where Z a n Z_{an} Z an is defined in Definition 4.4 ;
A hermitian connection A ∞ A_\infty A ∞ on E ∞ E_\infty E ∞ with finite YM ( A ∞ ) \operatorname{YM}(A_\infty) YM ( A ∞ ) , such that A j A_j A j converge to A ∞ A_\infty A ∞ modulo gauge transforms weakly in L 1 , loc p ( X ∖ Z a n ) L^p_{1,\operatorname{loc}}(X\setminus Z_{an}) L 1 , loc p ( X ∖ Z an ) .
Proof: As shown in [45 pages 4 and 13] , the metric ω \omega ω enters in estimates for connection by the lower bound to its injectivity radius from below, and the bound for its curvature. By our assumptions on metrics ω s \omega_s ω s , we have the uniform control of their curvatures. Also, by uniform equivalence, we control volumes and diameters of ω s \omega_s ω s . By Cheeger Finiteness Theorem, we have uniform lower bound on the injectivity radii of ω s \omega_s ω s , hence estimates from [45 Chapter 2] are uniform in s s s . Now we apply arguments of Uhlenbeck and the standard elliptic theory and obtain the sequence of A j A_j A j with stated properties. ■
In our setting, where A s A_s A s are Hermitian Yang-Mills connections constructed from the Chern connections of Hermitian-Einstein metrics, we can say more.
Corollary 4.7. If in addition − 1 Λ ω s F A s = μ ω s ( E ) Id \sqrt{-1}\Lambda_{\omega_s}F_{A_s}=\mu_{\omega_s}(E)\operatorname{Id} − 1 Λ ω s F A s = μ ω s ( E ) Id , then the Uhlenbeck limit A ∞ A_\infty A ∞ is Hermitian Yang-Mills with respect to ω 0 \omega_0 ω 0 . Moreover, A j A_j A j converge
to A ∞ A_{\infty} A ∞ in C l o c ∞ ( X ∖ Z a n ) C^{\infty}_{\mathrm{loc}}(X \setminus Z_{\mathrm{an}}) C loc ∞ ( X ∖ Z an ) , if for every k k k we have the uniform C k C^{k} C k -bounds for Rm ( ω s ) \operatorname{Rm}(\omega_s) Rm ( ω s ) , and E ∞ E_{\infty} E ∞ extends over all of X X X as a reflexive sheaf.
Proof: The fact that the limit is Yang-Mills follows from taking the weak limit of connections (modulo gauge) in L 1 p L^p_1 L 1 p , p > 2 n p > 2n p > 2 n . Indeed, the assumption − 1 Λ ω j F A j = μ ω j ( E ) Id \sqrt{-1}\Lambda_{\omega_j}F_{A_j} = \mu_{\omega_j}(E)\operatorname{Id} − 1 Λ ω j F A j = μ ω j ( E ) Id implies that
− 1 Λ ω 0 F A ∞ = μ ω 0 ( E ) Id \sqrt{-1}\Lambda_{\omega_0}F_{A_\infty} = \mu_{\omega_0}(E)\operatorname{Id} − 1 Λ ω 0 F A ∞ = μ ω 0 ( E ) Id
in the sense of distributions. The statement about the C ∞ C^\infty C ∞ convergence can be shown along the lines of [33 ] , and the fact that A ∞ A_\infty A ∞ is Hermitian Yang-Mills follows from it directly. The fact that E ∞ E_\infty E ∞ extends to a reflexive sheaf over all of X X X follows from [4 ] . ■
However, we can say more about A ∞ A_\infty A ∞ . Now we show the main result of this section.
Theorem 4.8: The Uhlenbeck limit A ∞ A_\infty A ∞ defines a reflexive sheaf E ∞ E_\infty E ∞ , which is isomorphic to Gr ω 0 s ( E ) ∗ ∗ \operatorname{Gr}^{s}_{\omega_0}(E)^{**} Gr ω 0 s ( E ) ∗∗ .
Proof: Consider the following YM flow on ( X , ω 0 ) (X,\omega_0) ( X , ω 0 ) :
{ ∂ t A t , j = − d A t , j ∗ F A t , j , A t , j ∣ t = 0 = A j . (30) \begin{cases}
\partial_t A_{t,j}=-d^{*}_{A_{t,j}}F_{A_{t,j}},\\
A_{t,j}|_{t=0}=A_j. \tag*{(30)}
\end{cases} { ∂ t A t , j = − d A t , j ∗ F A t , j , A t , j ∣ t = 0 = A j . (30)
Since lim j → + ∞ YM ω j ( A j ) = lim j → + ∞ YM ω 0 ( A j ) \lim_{j\to+\infty}\operatorname{YM}_{\omega_j}(A_j)=\lim_{j\to+\infty}\operatorname{YM}_{\omega_0}(A_j) lim j → + ∞ YM ω j ( A j ) = lim j → + ∞ YM ω 0 ( A j ) , we know that YM ω 0 ( A j ) \operatorname{YM}_{\omega_0}(A_j) YM ω 0 ( A j ) tends to the infimum of YM ω 0 \operatorname{YM}_{\omega_0} YM ω 0 by Proposition 2.9. Now the same argument, as in [11 ] and [38 Proposition 7.2] shows that
∥ A t , j − A j ∥ L 2 ( X , ω 0 ) ≤ ∫ 0 t ∥ ∂ s A s , j ∥ L 2 ( X , ω 0 ) d s ≤ t ( ∫ 0 t ∥ d A s , j ∗ F A s , j ∥ L 2 ( X , ω 0 ) 2 d s ) 1 / 2 ≤ t 2 ( YM ω 0 ( A j ) − YM ω 0 ( A t , j ) ) → j → + ∞ 0 , (31) \begin{aligned}
\lVert A_{t,j}-A_j\rVert_{L^2(X,\omega_0)}
\leq\int_0^t \lVert\partial_s A_{s,j}\rVert_{L^2(X,\omega_0)}\,ds \\
&\leq\sqrt{t}\left(\int_0^t \lVert d^{*}_{A_{s,j}}F_{A_{s,j}}\rVert^2_{L^2(X,\omega_0)}\,ds\right)^{1/2} \\
&\leq\sqrt{\frac{t}{2}\left(\operatorname{YM}_{\omega_0}(A_j)-\operatorname{YM}_{\omega_0}(A_{t,j})\right)} \xrightarrow[j\to+\infty]{} 0, \tag*{(31)}
\end{aligned} ∥ A t , j − A j ∥ L 2 ( X , ω 0 ) ≤ ∫ 0 t ∥ ∂ s A s , j ∥ L 2 ( X , ω 0 ) d s ≤ t ( ∫ 0 t ∥ d A s , j ∗ F A s , j ∥ L 2 ( X , ω 0 ) 2 d s ) 1/2 ≤ 2 t ( YM ω 0 ( A j ) − YM ω 0 ( A t , j ) ) j → + ∞ 0 , (31)
because YM ω 0 ( A j ) → inf A ∈ A H 1 , 1 YM ω 0 ( A ) \operatorname{YM}_{\omega_0}(A_j) \to\inf_{A\in\mathcal{A}^{1,1}_H} \operatorname{YM}_{\omega_0}(A) YM ω 0 ( A j ) → inf A ∈ A H 1 , 1 YM ω 0 ( A ) as j → + ∞ j \to+\infty j → + ∞ by Proposition 2.9, and YM ω 0 \operatorname{YM}_{\omega_0} YM ω 0 is decreasing along the Yang-Mills flow. In particular, after taking a further subsequence, we have
∥ d A s , j ∗ F A s , j ∥ L 2 ( X , ω 0 ) ⟶ j → + ∞ 0 \left\|d^*_{A_{s,j}}F_{A_{s,j}}\right\|_{L^2(X,\omega_0)} \longrightarrow_{j\to+\infty} 0 d A s , j ∗ F A s , j L 2 ( X , ω 0 ) ⟶ j → + ∞ 0
for almost all s s s . Hence we showed, that the Uhlenbeck limit of A t 0 , j A_{t_0,j} A t 0 , j is the same as the limit of A j A_j A j , and we again showed that the limit of A j A_j A j is Hermitian Yang-Mills with respect to ω 0 \omega_0 ω 0 . Also, the standard argument shows that
∣ Λ ω 0 F t , j ∣ 2 ( x ) ≤ ∫ X K t ( x , y ) ∣ Λ ω 0 F A j ∣ 2 ω 0 n n ! , \left|\Lambda_{\omega_0}F_{t,j}\right|^2(x) \leq\int_X K_t(x,y)\left|\Lambda_{\omega_0}F_{A_j}\right|^2\frac{\omega_0^n}{n!}, ∣ Λ ω 0 F t , j ∣ 2 ( x ) ≤ ∫ X K t ( x , y ) Λ ω 0 F A j 2 n ! ω 0 n ,
where K t ( x , y ) K_t(x,y) K t ( x , y ) is the heat kernel of the Laplacian on X X X with respect to ω 0 \omega_0 ω 0 . Since K t ( x , y ) ≤ C ( 1 + t − n ) K_t(x,y) \leq C(1+t^{-n}) K t ( x , y ) ≤ C ( 1 + t − n ) for any t > 0 t>0 t > 0 , the function ∣ Λ ω 0 F t , j ∣ 2 \left|\Lambda_{\omega_0}F_{t,j}\right|^2 ∣ Λ ω 0 F t , j ∣ 2 is uniformly bounded above by a constant times ∥ Λ ω 0 F A j ∥ L 2 ( X , ω 0 ) 2 \left\|\Lambda_{\omega_0}F_{A_j}\right\|^2_{L^2(X,\omega_0)} Λ ω 0 F A j L 2 ( X , ω 0 ) 2 . Also, note that for any j j j the operators ∂ ˉ A j \bar\partial_{A_j} ∂ ˉ A j lie in the same complex gauge orbit. Hence, the operators ∂ ˉ A t 0 , j \bar\partial_{A_{t_0,j}} ∂ ˉ A t 0 , j also lie in the same gauge orbit.
Finally, from
YM ω 0 ( A ) = HYM ω 0 ( A ) + 4 π 2 ( n − 2 ) ! ∫ X ( 2 c 2 ( E ) − c 1 2 ( E ) ) ∧ ω 0 n − 2 , \operatorname{YM}_{\omega_0}(A)=\operatorname{HYM}_{\omega_0}(A)+\frac{4\pi^2}{(n-2)!}\int_X\left(2c_2(E)-c_1^2(E)\right)\wedge\omega_0^{n-2}, YM ω 0 ( A ) = HYM ω 0 ( A ) + ( n − 2 )! 4 π 2 ∫ X ( 2 c 2 ( E ) − c 1 2 ( E ) ) ∧ ω 0 n − 2 ,
we see that
lim j → + ∞ HYM ω j ( A t , j ) = lim j → + ∞ HYM ω 0 ( A j ) = HYM ω 0 ( μ 0 ) , \lim_{j\to+\infty}\operatorname{HYM}_{\omega_j}(A_{t,j})=\lim_{j\to+\infty}\operatorname{HYM}_{\omega_0}(A_j)=\operatorname{HYM}_{\omega_0}(\mu_0), j → + ∞ lim HYM ω j ( A t , j ) = j → + ∞ lim HYM ω 0 ( A j ) = HYM ω 0 ( μ 0 ) ,
where μ 0 = ( μ 0 ( E ) , … , μ 0 ( E ) ) \mu_0=(\mu_0(E),\ldots,\mu_0(E)) μ 0 = ( μ 0 ( E ) , … , μ 0 ( E )) – the Seshadri type of E E E . Hence, the same is true for HYM α , N , ω 0 ( A j ) \operatorname{HYM}_{\alpha,N,\omega_0}(A_j) HYM α , N , ω 0 ( A j ) for any α ∈ [ 1 ; 2 ] \alpha\in[1;2] α ∈ [ 1 ; 2 ] , and any N > 0 N>0 N > 0 . In other words, for any α ∈ [ 1 ; 2 ] \alpha\in[1;2] α ∈ [ 1 ; 2 ] we have
1 r ∑ γ = 1 r ∣ μ γ ( A ∞ ) + N ∣ α = ∣ μ 0 ( E ) + N ∣ α . \frac{1}{r}\sum_{\gamma=1}^{r}\left|\mu_\gamma(A_\infty)+N\right|^\alpha=\left|\mu_0(E)+N\right|^\alpha. r 1 γ = 1 ∑ r ∣ μ γ ( A ∞ ) + N ∣ α = ∣ μ 0 ( E ) + N ∣ α .
Hence for any γ \gamma γ we obtain μ γ ( A ∞ ) = μ ω 0 ( E ) \mu_\gamma(A_\infty)=\mu_{\omega_0}(E) μ γ ( A ∞ ) = μ ω 0 ( E ) .
We have shown that the limit of A j A_j A j is the same as the limit of A t 0 , j A_{t_0,j} A t 0 , j for some t 0 t_0 t 0 . Moreover, the Harder-Narasimhan type of ( E ∞ , A ∞ ) (E_\infty,A_\infty) ( E ∞ , A ∞ ) is the same as μ 0 \mu_0 μ 0 . Now the arguments from [15 ] , or [38 Section 7.1] , or invoking [39 Theorem 2.7] show the isomorphism of E ∞ E_\infty E ∞ to Gr ω 0 S ( E ) ∗ ∗ \operatorname{Gr}^{S}_{\omega_0}(E)^{**} Gr ω 0 S ( E ) ∗∗ . ▪
Rectifiability and analyticity of the blow-up set.
At this moment, we do not know that Z a n Z_{an} Z an is indeed an analytic subvariety of X X X . We do not even know whether it is rectifiable. So, in this part we are going to establish these properties of Z a n Z_{an} Z an . For the rectifiability part, we will mimic the exposition in [8 Chapter 3] , and for analyticity we will follow the arguments of Hong and Tian from [26 ] and [44 ] . First of all, we need the following definition
Definition 4.9: Let { s j } \{s_j\} { s j } be a sequence, which tends to zero, and ω j \omega_j ω j with A j A_j A j are the corresponding sequences of metrics and unitary connections. Then one can define a Radon measure μ \mu μ as
μ = lim j → + ∞ ∣ F A j ∣ 2 ω j n n ! = ∣ F A ∞ ∣ 2 ω 0 n n ! + ν . \mu= \lim_{j\to+\infty} |F_{A_j}|^2\frac{\omega_j^n}{n!} = |F_{A_\infty}|^2\frac{\omega_0^n}{n!}+\nu. μ = j → + ∞ lim ∣ F A j ∣ 2 n ! ω j n = ∣ F A ∞ ∣ 2 n ! ω 0 n + ν . The measure ν \nu ν is called the defect measure.
Of course, this measure depends on the sequence A j A_j A j .
We need the following proposition
Proposition 4.10:
Let ( X , ω j , E , A j ) (X,\omega_j,E,A_j) ( X , ω j , E , A j ) be as above. Let B r ( x ) B_r(x) B r ( x ) be the ball of radius r r r with respect to ω 0 \omega_0 ω 0 . Then the following holds
For 0 < r ≪ 1 0<r\ll1 0 < r ≪ 1 the quantity e a r 2 r 4 − 2 n μ ( B r ( x ) ) e^{ar^2}r^{4-2n}\mu(B_r(x)) e a r 2 r 4 − 2 n μ ( B r ( x )) is decreasing as r → 0 r\to0 r → 0 . Hence, the function
Θ ( μ , x ) = lim r → 0 r 4 − 2 n e a r 2 μ ( B r ( x ) ) \Theta(\mu,x)=\lim_{r\to0}r^{4-2n}e^{ar^2}\mu(B_r(x)) Θ ( μ , x ) = r → 0 lim r 4 − 2 n e a r 2 μ ( B r ( x )) is well-defined, upper semi-continuous and H 2 n − 4 \mathcal{H}^{2n-4} H 2 n − 4 approximately continuous at H 2 n − 4 \mathcal{H}^{2n-4} H 2 n − 4 a.e. x ∈ Z a n x\in Z_{an} x ∈ Z an ;
The point x x x belongs to Z a n Z_{an} Z an if and only if Θ ( μ , x ) ≥ ε 0 \Theta(\mu,x)\ge\varepsilon_0 Θ ( μ , x ) ≥ ε 0 . Moreover, for H 2 n − 4 \mathcal{H}^{2n-4} H 2 n − 4 a.e. x ∈ Z a n x\in Z_{an} x ∈ Z an we have that
lim sup r → 0 r 4 − 2 n e a r 2 ∫ B r ( x ) ∣ F A ∞ ∣ 2 ω 0 n = 0. \limsup_{r\to0}r^{4-2n}e^{ar^2}\int_{B_r(x)}|F_{A_\infty}|^2\omega_0^n=0. r → 0 lim sup r 4 − 2 n e a r 2 ∫ B r ( x ) ∣ F A ∞ ∣ 2 ω 0 n = 0.
The defect measure ν \nu ν is H Z a n 2 n − 4 \mathcal{H}^{2n-4}_{Z_{an}} H Z an 2 n − 4 absolutely continuous and is equal to Θ ( μ , x ) d H Z a n 2 n − 4 \Theta(\mu,x)d\mathcal{H}^{2n-4}_{Z_{an}} Θ ( μ , x ) d H Z an 2 n − 4 .
Proof: The arguments from [8 , 26 , 30 , 31 , 44 ] can be carried over almost verbatim here. We just outline the proof, and explain the main differences.
This follows from the observation, that ∣ F A j ∣ 2 ω j n |F_{A_j}|^2\omega_j^n ∣ F A j ∣ 2 ω j n and ∣ F A j ∣ 2 ω 0 n |F_{A_j}|^2\omega_0^n ∣ F A j ∣ 2 ω 0 n have the same weak limit n ! μ n!\mu n ! μ . Indeed, our assumptions on ω s \omega_s ω s imply that ω j → ω 0 \omega_j\to\omega_0 ω j → ω 0 uniformly. Hence, there are δ j → 0 \delta_j\to0 δ j → 0 , such that:
lim j → + ∞ ∫ B r + δ j , j ( x ) ∣ F A j ∣ 2 ω j n = lim j → + ∞ ∫ B r ( x ) ∣ F A j ∣ 2 ω 0 n , \lim_{j\to+\infty}\int_{B_{r+\delta_j,j}(x)} |F_{A_j}|^2\omega_j^n
=
\lim_{j\to+\infty}\int_{B_r(x)} |F_{A_j}|^2\omega_0^n, j → + ∞ lim ∫ B r + δ j , j ( x ) ∣ F A j ∣ 2 ω j n = j → + ∞ lim ∫ B r ( x ) ∣ F A j ∣ 2 ω 0 n ,
again because ω j \omega_j ω j are uniformly equivalent and converge to ω 0 \omega_0 ω 0 uniformly. By Theorem 4.1 we see that r 4 − 2 n e a r 2 μ ( B r , ω 0 ( x ) ) r^{4-2n}e^{ar^2}\mu(B_{r,\omega_0}(x)) r 4 − 2 n e a r 2 μ ( B r , ω 0 ( x )) is non-decreasing for every x x x , and the limit exists. We only need to ensure, that μ ( ∂ B r , 0 ( x ) ) = 0 \mu(\partial B_{r,0}(x))=0 μ ( ∂ B r , 0 ( x )) = 0 , as measure may concentrate on the spheres as j → + ∞ j\to+\infty j → + ∞ . However, this problem does not exist for almost all radii.
The upper-semicontinuity is a standard fact, and the approximate continuity follows from the argument of [44 Lemma 3.2.2] almost verbatim.
The statement follows from standard covering arguments, see [44 ] or [31 ] .
The proof is similar to [8 Lemma 3.3] .
■ \blacksquare ■
We also need the following proposition, which is an analogue of decomposition from [8 Lemma 3.3] .
Proposition 4.11: The set Z a n Z_{an} Z an can be decomposed as
Z a n = Z b , a n ∪ Sing ( A ∞ ) , Z_{an}=Z_{b,an}\cup\operatorname{Sing}(A_\infty), Z an = Z b , an ∪ Sing ( A ∞ ) ,
where Z b , a n = supp ( ν ) Z_{b,\mathrm{an}}=\operatorname{supp}(\nu) Z b , an = supp ( ν ) , and
Sing ( A ∞ ) = { x ∈ X | lim sup r → 0 r 4 − 2 n ∫ B r , 0 ( x ) ∣ F A ∞ ∣ 2 ω 0 n ≥ ε 0 } \operatorname{Sing}(A_{\infty})=\left\{x\in X\mathrel{}\middle|\mathrel{}\limsup_{r\to0}r^{4-2n}\int_{B_{r,0}(x)}\lvert F_{A_{\infty}}\rvert^{2}\,\omega_{0}^{n}\ge\varepsilon_{0}\right\} Sing ( A ∞ ) = { x ∈ X r → 0 lim sup r 4 − 2 n ∫ B r , 0 ( x ) ∣ F A ∞ ∣ 2 ω 0 n ≥ ε 0 }
is an analytic subset of codimension at least 3.
Proof: This follows essentially from (3) of Proposition 4.10 , and it can be proved along the lines of [8 Lemma 3.3] , by using the fact that metrics ω j \omega_j ω j are uniformly equivalent to ω 0 \omega_0 ω 0 . The fact that Sing ( A ∞ ) \operatorname{Sing}(A_{\infty}) Sing ( A ∞ ) is analytic and of codimension at least 3 follows from the result of Bando and Siu [4 ] . ■
It can be shown, that one can find suitable local gauge transformations u u u , such that u ∗ A ∞ u^{*}A_{\infty} u ∗ A ∞ extends across Z b , a n ∖ Sing ( A ∞ ) Z_{b,\mathrm{an}}\setminus\operatorname{Sing}(A_{\infty}) Z b , an ∖ Sing ( A ∞ ) .
Next, we need to consider tangent cone measures. Before we proceed, we need to set some notation.
Let τ λ : T x X → T x X \tau_{\lambda}:T_xX\to T_xX τ λ : T x X → T x X be a dilation by λ \lambda λ . Let μ x , λ ( E ) = λ 4 − 2 n μ ( exp x , g 0 ( τ λ ( E ) ) ) \mu_{x,\lambda}(E)=\lambda^{4-2n}\mu(\exp_{x,g_0}(\tau_{\lambda}(E))) μ x , λ ( E ) = λ 4 − 2 n μ ( exp x , g 0 ( τ λ ( E ))) for any E ⊂ T x X E\subset T_xX E ⊂ T x X . Here exp x , g 0 \exp_{x,g_0} exp x , g 0 is the exponential map of the metric ω 0 \omega_0 ω 0 . Denote also by g j g_j g j the Riemannian metric for ω s j \omega_{s_j} ω s j , and by g 0 g_0 g 0 the Riemannian metric given by ω 0 \omega_0 ω 0 . We have the following proposition
Proposition 4.12: There is a sequence λ j → 0 + \lambda_j\to0^{+} λ j → 0 + , such that μ x , λ j → η \mu_{x,\lambda_j}\to\eta μ x , λ j → η , where η \eta η is a cone measure on T x X T_xX T x X , i.e.
η ( λ E ) = λ 2 n − 4 η ( E ) . \eta(\lambda E)=\lambda^{2n-4}\eta(E). η ( λ E ) = λ 2 n − 4 η ( E ) .
Proof: This essentially follows from Proposition 4.10 , together with Preiss’ theorem from [35 ] . However, we can give a short proof, following [8 Proposition 3.4] .
Consider the following connections:
A x , j , λ = τ λ ∗ exp x , g 0 ∗ A s j . A_{x,j,\lambda}=\tau_{\lambda}^{*}\exp_{x,g_0}^{*}A_{s_j}. A x , j , λ = τ λ ∗ exp x , g 0 ∗ A s j .
Let g x , j , λ = 1 λ 2 τ λ ∗ exp x , g 0 ∗ g j g_{x,j,\lambda}=\frac{1}{\lambda^{2}}\tau_{\lambda}^{*}\exp_{x,g_0}^{*}g_j g x , j , λ = λ 2 1 τ λ ∗ exp x , g 0 ∗ g j , and let d V x , j , λ dV_{x,j,\lambda} d V x , j , λ be the volume form with respect to this metric.
Let η = lim λ k → 0 μ x , λ k \eta= \lim_{\lambda_k \to0} \mu_{x,\lambda_k} η = lim λ k → 0 μ x , λ k . This limit exists because of the monotonicity formula from Theorem 4.1 . There is a subsequence A x , j , λ j A_{x,j,\lambda_j} A x , j , λ j of A x , j , λ k A_{x,j,\lambda_k} A x , j , λ k , such that
∣ F x , j , λ j ∣ 2 d V x , j , λ j ⟶ j → + ∞ η . |F_{x,j,\lambda_j}|^2 dV_{x,j,\lambda_j} \longrightarrow_{j \to+\infty} \eta. ∣ F x , j , λ j ∣ 2 d V x , j , λ j ⟶ j → + ∞ η .
Note that g x , j , λ j g_{x,j,\lambda_j} g x , j , λ j and g x , 0 , λ j g_{x,0,\lambda_j} g x , 0 , λ j have the same limit, as metrics are uniformly equivalent, and g j g_j g j tends to g 0 g_0 g 0 as j → + ∞ j \to+\infty j → + ∞ . The fact that η \eta η is a cone measure now follows from the monotonicity formula Theorem 4.1 , just as in [8 Proposition 3.4] or [44 Lemma 3.2.1] , but with a little modification. For fixed j j j , and any radially invariant function φ ≥ 0 \varphi\ge0 φ ≥ 0 , we have
R 4 − 2 n ∫ B R , g j ( x ) φ ∣ F x , j , λ j ∣ 2 d V x , j , λ j − r 4 − 2 n ∫ B r , g j ( x ) φ ∣ F x , j , λ j ∣ 2 d V x , j , λ j = (32) R^{4-2n}\int_{B_{R,g_j}(x)}\varphi|F_{x,j,\lambda_j}|^2dV_{x,j,\lambda_j}-r^{4-2n}\int_{B_{r,g_j}(x)}\varphi|F_{x,j,\lambda_j}|^2dV_{x,j,\lambda_j} = \tag*{(32)} R 4 − 2 n ∫ B R , g j ( x ) φ ∣ F x , j , λ j ∣ 2 d V x , j , λ j − r 4 − 2 n ∫ B r , g j ( x ) φ ∣ F x , j , λ j ∣ 2 d V x , j , λ j = (32)
∫ r R d d t ( t 4 − 2 n ∫ B 1 , g j ( x ) φ ∣ τ t ∗ F x , j , λ j ∣ 2 τ t ∗ d V x , j , λ j ) (33) \int_r^R\frac{d}{dt}\left(t^{4-2n}\int_{B_{1,g_j}(x)}\varphi|\tau_t^*F_{x,j,\lambda_j}|^2\tau_t^*dV_{x,j,\lambda_j}\right) \tag*{(33)} ∫ r R d t d ( t 4 − 2 n ∫ B 1 , g j ( x ) φ ∣ τ t ∗ F x , j , λ j ∣ 2 τ t ∗ d V x , j , λ j ) (33)
As in [8 Proposition 3.4] , the expression on the right hand side of 4.5 is bounded by a constant times the integral of
r 4 − 2 n ∣ ι ∂ r F x , j , λ j ∣ ∣ F x , j , λ j ∣ . r^{4-2n}|\iota_{\partial_r}F_{x,j,\lambda_j}||F_{x,j,\lambda_j}|. r 4 − 2 n ∣ ι ∂ r F x , j , λ j ∣∣ F x , j , λ j ∣.
Here r = r j r = r_j r = r j – the distance with respect to g j g_j g j . As λ j → 0 \lambda_j \to0 λ j → 0 , we know that g x , j , λ j g_{x,j,\lambda_j} g x , j , λ j tend uniformly to the Euclidean metric δ = g x , 0 \delta= g_{x,0} δ = g x , 0 , so the integral of ∣ ι ∂ r F x , j , λ j ∣ |\iota_{\partial_r}F_{x,j,\lambda_j}| ∣ ι ∂ r F x , j , λ j ∣ tends to zero by Theorem 4.1 .
□
Next we show the existence of weak complex tangent planes.
Proposition 4.13. The set Z a n Z_{an} Z an is rectifiable, and for H 2 n − 4 \mathcal{H}^{2n-4} H 2 n − 4 a. e. x ∈ Z a n x \in Z_{an} x ∈ Z an the tangent plane to Z a n Z_{an} Z an at x x x is a complex subspace in T x X T_xX T x X , calibrated by ω 0 n − 2 ( n − 2 ) ! \frac{\omega_0^{n-2}}{(n-2)!} ( n − 2 )! ω 0 n − 2 .
Proof: The rectifiability follows from Proposition 4.10 together with the main result of [35 ] . For the last statement, that tangent planes are complex, the arguments from [44 Section 4] can be carried over word by word in our setting.
Let γ y 1 , y 2 δ j ( x ) = δ j x + ( y 1 , y 2 ) \gamma^{\delta_j}_{y_1,y_2}(x)=\delta_jx+(y_1,y_2) γ y 1 , y 2 δ j ( x ) = δ j x + ( y 1 , y 2 ) . Here y 1 ∈ V y_1\in V y 1 ∈ V , and y 2 ∈ V ⊥ y_2\in V^\perp y 2 ∈ V ⊥ . The forms
Ω j = δ j 4 − 2 n τ y 1 , y 2 δ j , ∗ exp ∗ ω j n − 2 ( n − 2 ) ! \Omega_j=\delta_j^{4-2n}\tau^{\delta_j,*}_{y_1,y_2}\exp_* \frac{\omega_j^{n-2}}{(n-2)!} Ω j = δ j 4 − 2 n τ y 1 , y 2 δ j , ∗ exp ∗ ( n − 2 )! ω j n − 2
converge to Ω 0 = ω 0 n − 2 ( n − 2 ) ! ∣ T x X \Omega_0=\frac{\omega_0^{n-2}}{(n-2)!}\big|_{T_xX} Ω 0 = ( n − 2 )! ω 0 n − 2 T x X . Tian’s argument shows that there exists Ω 0 \Omega_0 Ω 0 -anti-self-dual instanton B B B with F B ≠ 0 F_B\ne0 F B = 0 , Λ F B = 0 \Lambda F_B=0 Λ F B = 0 , and ι v F B = 0 \iota_vF_B=0 ι v F B = 0 for any v ∈ V v\in V v ∈ V . We now can repeat the argument of [44 Corollary 4.2.2] , that Ω 0 ∣ T x Z a n \Omega_0|_{T_xZ_{\mathrm{an}}} Ω 0 ∣ T x Z an is equal to the volume form on T x Z a n T_xZ_{\mathrm{an}} T x Z an , defined by the metric g x , 0 g_{x,0} g x , 0 . In other words, T x Z a n T_xZ_{\mathrm{an}} T x Z an is Ω 0 \Omega_0 Ω 0 -calibrated, and by the Wirtinger inequality, it is a complex subspace.
Now we are in a position to prove the analogue of the theorem of Hong and Tian.
Theorem 4.14: (see [26 Theorem 11] and [44 Theorem 4] ). Let X X X be a compact Kähler manifold, such that the Kähler metrics ω j \omega_j ω j uniformly converge to ω 0 \omega_0 ω 0 , and have uniformly bounded curvature. Let ( E , H ) (E,H) ( E , H ) be a hermitian vector bundle and suppose A j ∈ A H 1 , 1 ( E ) A_j\in\mathcal{A}^{1,1}_H(E) A j ∈ A H 1 , 1 ( E ) are the integrable Hermitian Yang-Mills connections with respect to ω j \omega_j ω j . Then, up to taking a subsequence, A j A_j A j converge weakly to the ω 0 \omega_0 ω 0 Hermitian Yang-Mills connection with the blow-up locus Z a n Z_{\mathrm{an}} Z an . The subset Z a n Z_{\mathrm{an}} Z an is rectifiable and admits a decomposition Z a n = ⋃ α Z α ∪ Sing ( A ∞ ) Z_{\mathrm{an}}=\bigcup_\alpha Z_\alpha\cup\operatorname{Sing}(A_\infty) Z an = ⋃ α Z α ∪ Sing ( A ∞ ) , where each Z α Z_\alpha Z α is a codimension 2 analytic subvariety. Moreover, for each smooth ( n − 2 , n − 2 ) (n-2,n-2) ( n − 2 , n − 2 ) form φ \varphi φ
lim j → + ∞ ∫ X c 2 ( A j ) ∧ φ = ∫ X c 2 ( A ∞ ) ∧ φ + ∑ α m α ∫ Z α φ . \lim_{j\to+\infty}\int_X c_2(A_j)\wedge\varphi=\int_X c_2(A_\infty)\wedge\varphi+\sum_\alpha m_\alpha\int_{Z_\alpha}\varphi. j → + ∞ lim ∫ X c 2 ( A j ) ∧ φ = ∫ X c 2 ( A ∞ ) ∧ φ + α ∑ m α ∫ Z α φ . Here for any connection A ∈ A H 1 , 1 A\in\mathcal{A}^{1,1}_H A ∈ A H 1 , 1 we define c 2 ( A ) c_2(A) c 2 ( A ) as
c 2 ( A ) = 1 8 π 2 ( Tr ( F A ∧ F A ) − Tr F A ∧ Tr F A ) . c_2(A)=\frac{1}{8\pi^2}\left(\operatorname{Tr}(F_A\wedge F_A)-\operatorname{Tr}F_A\wedge\operatorname{Tr}F_A\right). c 2 ( A ) = 8 π 2 1 ( Tr ( F A ∧ F A ) − Tr F A ∧ Tr F A ) .
Proof: The proof is essentially the same, as [44 Theorems 4.2.3-4.3.3] . We outline the main idea, and the modifications to add in our setting, so Tian’s arguments would work. First of all, we need to show, that Tr F A j ∧ Tr F A j \operatorname{Tr} F_{A_j} \wedge\operatorname{Tr} F_{A_j} Tr F A j ∧ Tr F A j converge to Tr F A ∞ ∧ Tr F A ∞ \operatorname{Tr} F_{A_\infty} \wedge\operatorname{Tr} F_{A_\infty} Tr F A ∞ ∧ Tr F A ∞ , and the latter extends to the whole X X X . Note that α j = − 1 Tr F A j \alpha_j=\sqrt{-1}\operatorname{Tr} F_{A_j} α j = − 1 Tr F A j is ω j \omega_j ω j -harmonic, and lies in 2 π c 1 ( E ) 2\pi c_1(E) 2 π c 1 ( E ) . The Bochner formula and the Moser iteration imply that C 0 C^0 C 0 -norm of α j \alpha_j α j is controlled by its L 2 ( X , ω 0 ) L^2(X,\omega_0) L 2 ( X , ω 0 ) -norm. Moreover, for any ω \omega ω the formula
α 2 ∧ ω n − 2 ( n − 2 ) ! = ( ∣ Λ ω α ∣ 2 − ∣ α ∣ ω 2 ) ω n n ! \alpha^2\wedge\frac{\omega^{n-2}}{(n-2)!}=\left(\lvert\Lambda_\omega\alpha\rvert^2-\lvert\alpha\rvert_\omega^2\right)\frac{\omega^n}{n!} α 2 ∧ ( n − 2 )! ω n − 2 = ( ∣ Λ ω α ∣ 2 − ∣ α ∣ ω 2 ) n ! ω n
together with Proposition 2.9 implies that the L 2 L^2 L 2 -norm of α j \alpha_j α j is uniformly bounded. Thus, by the standard elliptic theory, the C 1 C^1 C 1 -norm of α j \alpha_j α j is uniformly bounded by a constant, independent of j j j .
Let α 0 \alpha_0 α 0 be a ω 0 \omega_0 ω 0 -harmonic form in 2 π c 1 ( E ) 2\pi c_1(E) 2 π c 1 ( E ) . The same argument as above implies that
∥ α j − α 0 ∥ L 2 ( X , ω 0 ) 2 = ∥ Λ ω 0 α j − Λ ω 0 α 0 ∥ L 2 ( X , ω 0 ) 2 ⟶ j → + ∞ 0. \lVert\alpha_j-\alpha_0\rVert_{L^2(X,\omega_0)}^2=\lVert\Lambda_{\omega_0}\alpha_j-\Lambda_{\omega_0}\alpha_0\rVert_{L^2(X,\omega_0)}^2\underset{j\to+\infty}{\longrightarrow}0. ∥ α j − α 0 ∥ L 2 ( X , ω 0 ) 2 = ∥ Λ ω 0 α j − Λ ω 0 α 0 ∥ L 2 ( X , ω 0 ) 2 j → + ∞ ⟶ 0.
Clearly α 0 = − 1 Tr F A ∞ \alpha_0=\sqrt{-1}\operatorname{Tr} F_{A_\infty} α 0 = − 1 Tr F A ∞ on X ∖ Z a n X\setminus Z_{\mathrm{an}} X ∖ Z an . Hence, there is a bounded extension of Tr F A ∞ \operatorname{Tr} F_{A_\infty} Tr F A ∞ on the whole X X X , and Tr F A j ∧ Tr F A j \operatorname{Tr} F_{A_j}\wedge\operatorname{Tr} F_{A_j} Tr F A j ∧ Tr F A j converge to Tr F A ∞ ∧ Tr F A ∞ \operatorname{Tr} F_{A_\infty}\wedge\operatorname{Tr} F_{A_\infty} Tr F A ∞ ∧ Tr F A ∞ as currents.
Consider the currents T j = c 2 ( A j ) − c 2 ( A ∞ ) T_j=c_2(A_j)-c_2(A_\infty) T j = c 2 ( A j ) − c 2 ( A ∞ ) . The previous discussion and the same arguments as [44 Proposition 2.3.1] imply, that T j T_j T j are closed. The uniform L 2 L^2 L 2 -bound on curvatures F A j F_{A_j} F A j implies that T j T_j T j have bounded mass. Thus, up to taking a subsequence, there is a limit current
T = lim j → + ∞ T j = lim j → + ∞ 1 8 π 2 ( Tr ( F A j ∧ F A j ) − Tr ( F A ∞ ∧ F A ∞ ) ) . T=\lim_{j\to+\infty}T_j=\lim_{j\to+\infty}\frac{1}{8\pi^2}\left(\operatorname{Tr}(F_{A_j}\wedge F_{A_j})-\operatorname{Tr}(F_{A_\infty}\wedge F_{A_\infty})\right). T = j → + ∞ lim T j = j → + ∞ lim 8 π 2 1 ( Tr ( F A j ∧ F A j ) − Tr ( F A ∞ ∧ F A ∞ ) ) .
which is closed and has bounded mass. Up to replacing g x , λ j g_{x,\lambda_j} g x , λ j by g x , j , λ j g_{x,j,\lambda_j} g x , j , λ j , and τ λ j ∗ Ω \tau^*_{\lambda_j}\Omega τ λ j ∗ Ω by Ω j \Omega_j Ω j from Proposition 4.13, we can repeat Tian’s arguments, and show that T T T is integer.
Now, Simon’s rectifiability theorem (see [41 Theorem 32.1] ) guarantees, that there is a rectifiable set Σ \Sigma Σ of Hausdorff codimension 4 4 4 , together with an orientation multivector η \eta η , and an integer-valued function θ \theta θ , such that
T ( φ ) = ∫ Σ ⟨ φ , η ⟩ θ d H 2 n − 4 . T(\varphi)=\int_\Sigma\langle\varphi,\eta\rangle\theta\,d\mathcal{H}^{2n-4}. T ( φ ) = ∫ Σ ⟨ φ , η ⟩ θ d H 2 n − 4 .
On the other hand, for φ = f ω 0 n − 2 ( n − 2 ) ! \varphi=f\frac{\omega_0^{n-2}}{(n-2)!} φ = f ( n − 2 )! ω 0 n − 2 we clearly have
T ( φ ) = 1 8 π 2 ∫ Z a n f Θ d H Z a n 2 n − 4 . T(\varphi)=\frac{1}{8\pi^2}\int_{Z_{an}}f\Theta\,dH_{Z_{an}}^{2n-4}. T ( φ ) = 8 π 2 1 ∫ Z an f Θ d H Z an 2 n − 4 .
Finally, as in Tian’s paper, the results from [27 , 25 ] imply, that there are holomorphic subvarieties Z α Z_\alpha Z α , and positive integers m α m_\alpha m α , such that
T ( φ ) = ∑ α m α ∫ Z α φ . T(\varphi)=\sum_\alpha m_\alpha\int_{Z_\alpha}\varphi. T ( φ ) = α ∑ m α ∫ Z α φ .
Hence, Z b , a n = ⋃ α Z α Z_{b,an}=\bigcup_\alpha Z_\alpha Z b , an = ⋃ α Z α , and Z a n = Z b , a n ∪ Sing ( A ∞ ) Z_{an}=Z_{b,an}\cup\operatorname{Sing}(A_\infty) Z an = Z b , an ∪ Sing ( A ∞ ) is analytic. ▪
Definition 4.15: We will refer to the numbers m α m_\alpha m α that appear in the proof of Theorem 4.14 as analytic multiplicities of Z α Z_\alpha Z α .
The analog of Sibley-Wentworth theorem
In this section we will prove the analogue of [39 Theorem 1.1] . This result is the content of the last part of Theorem 1.1 . In fact, we just explain how the arguments of Sibley and Wentworth can be carried over in our framework.
Before we proceed, we recall the notion of algebraic multiplicities. We need to introduce the notion of associated graded object of a coherent sheaf (see [39 Erratum pp 27-28] ).
Let F \mathcal{F} F be a coherent sheaf over X X X , and J = Ann ( F ) \mathcal{J}=\sqrt{\operatorname{Ann}(\mathcal{F})} J = Ann ( F ) be the radical of its annihilator. We define gr ( F ) \operatorname{gr}(\mathcal{F}) gr ( F ) as
gr ( F ) = ⨁ k ≥ 0 J k F / J k + 1 F . \operatorname{gr}(\mathcal{F})=\bigoplus_{k\geq0}\mathcal{J}^k\mathcal{F}/\mathcal{J}^{k+1}\mathcal{F}. gr ( F ) = k ≥ 0 ⨁ J k F / J k + 1 F .
Note that this sum is clearly finite.
Definition 4.16: (see [39 Erratum on pages 27-28] ) Let Gr ω 0 S ( E ) \operatorname{Gr}^{S}_{\omega_0}(E) Gr ω 0 S ( E ) be an associated graded sheaf, and Z a l g Z_{alg} Z a l g be the analytic set defined as in Definition 2.3 . Let ι : Z ↪ Z a l g \iota:Z\hookrightarrow Z_{alg} ι : Z ↪ Z a l g be an irreducible component of Z a l g Z_{alg} Z a l g of codimension 2 2 2 . The algebraic multiplicity n Z n_Z n Z of Z Z Z is the generic rank of ι ∗ gr ( ( Gr ω 0 S ( E ) ) ∗ ∗ Gr ω 0 S ( E ) ) \iota^*\operatorname{gr}\left(\frac{\left(\operatorname{Gr}^{S}_{\omega_0}(E)\right)^{**}}{\operatorname{Gr}^{S}_{\omega_0}(E)}\right) ι ∗ gr ( Gr ω 0 S ( E ) ( Gr ω 0 S ( E ) ) ∗∗ ) .
This definition appears in [39 Errata, pp.27-28] as a replacement of [39, Definition 2.4.
Theorem 4.17. The subset Z a n Z_{an} Z an coincides with Z a l g Z_{alg} Z a l g as sets. Moreover, for any codimension two irreducible subvariety Z α ⊂ Z a n Z_{\alpha} \subset Z_{an} Z α ⊂ Z an , we have Z α ⊂ Z a l g Z_{\alpha} \subset Z_{alg} Z α ⊂ Z a l g , and its algebraic multiplicity coincides with m α m_{\alpha} m α .
Proof. The proof is essentially the same, as in [39 ] . We just outline the main differences here.
Step 1: The slicing lemma [39 Lemma 4.1] ensures that for an embedded 4-disc Σ ⊂ X \Sigma\subset X Σ ⊂ X , which is a transverse slice for Z α ⊂ Z b , a n Z_{\alpha} \subset Z_{b,an} Z α ⊂ Z b , an (i.e. Σ \Sigma Σ intersects transversely Z α Z_{\alpha} Z α at one point), one has
m α = lim j → + ∞ ∫ Σ c 2 ( A j ) − c 2 ( A ∞ ) = lim j → + ∞ 1 8 π 2 ∫ Σ Tr ( F A j ∧ F A j ) − Tr ( F A ∞ ∧ F A ∞ ) m_{\alpha} = \lim_{j \to+\infty} \int_{\Sigma} c_{2}(A_{j}) - c_{2}(A_{\infty}) = \lim_{j \to+\infty} \frac{1}{8\pi^{2}} \int_{\Sigma} \operatorname{Tr}(F_{A_{j}} \wedge F_{A_{j}}) - \operatorname{Tr}(F_{A_{\infty}} \wedge F_{A_{\infty}}) m α = j → + ∞ lim ∫ Σ c 2 ( A j ) − c 2 ( A ∞ ) = j → + ∞ lim 8 π 2 1 ∫ Σ Tr ( F A j ∧ F A j ) − Tr ( F A ∞ ∧ F A ∞ )
The cancellation of Tr F A j ∧ Tr F A j − Tr F A ∞ ∧ Tr F A ∞ \operatorname{Tr} F_{A_{j}} \wedge\operatorname{Tr} F_{A_{j}} - \operatorname{Tr} F_{A_{\infty}} \wedge\operatorname{Tr} F_{A_{\infty}} Tr F A j ∧ Tr F A j − Tr F A ∞ ∧ Tr F A ∞ is justified in the proof of Theorem 4.14 , and the proof of the equality above follows from the same argument, which Sibley and Wentworth use to show that the integral ∫ ∂ Σ CS ( A j , A ∞ ) \int_{\partial\Sigma} \operatorname{CS}(A_{j}, A_{\infty}) ∫ ∂ Σ CS ( A j , A ∞ ) tends to zero as j → + ∞ j \to+\infty j → + ∞ .
Step 2: Next, [39 Lemma 4.3] is a general complex-analytic fact, and does not require any information about our connections A j A_{j} A j . Also, we can use [39 Theorem 1.3] . This theorem says, that for a holomorphic vector bundle E E E with metric H H H , the following holds:
ch 2 ( Gr ( E ) , H ~ ) − ch 2 ( E , H ) = ∑ α n α W α + d d c Ψ \operatorname{ch}_{2}(\operatorname{Gr}(E), \widetilde{H}) - \operatorname{ch}_{2}(E,H) = \sum_{\alpha} n_{\alpha} W_{\alpha} + dd^{c}\Psi ch 2 ( Gr ( E ) , H ) − ch 2 ( E , H ) = α ∑ n α W α + d d c Ψ
Here H ~ \widetilde{H} H is an admissible metric on the free part of Gr S ( E ) ∗ ∗ \operatorname{Gr}^{S}(E)^{**} Gr S ( E ) ∗∗ (i.e. with L 2 L^{2} L 2 -curvature and with bounded ∣ Λ ω 0 F H ~ ∣ \lvert\Lambda_{\omega_{0}}F_{\widetilde{H}}\rvert ∣ Λ ω 0 F H ∣ ; see [4 ] ), W α W_{\alpha} W α is a codimension 2 component of Z a l g Z_{alg} Z a l g , and n α n_{\alpha} n α are defined in Definition 4.16 .
Step 3: Finally, the proof of [39 Proposition 4.2] carries over verbatim. Let Z α ⊂ Z a l g Z_{\alpha} \subset Z_{alg} Z α ⊂ Z a l g with algebraic multiplicity n α n_{\alpha} n α . Upon taking a modification of X X X (see [39 Lemma 4.3] ), we can take a closed 4-dimensional submanifold Σ ⊂ X \Sigma\subset X Σ ⊂ X , such that [ Σ ] ⋅ [ Z α ] ≠ 0 [\Sigma] \cdot[Z_{\alpha}] \ne0 [ Σ ] ⋅ [ Z α ] = 0 , and Σ \Sigma Σ intersects Z α Z_{\alpha} Z α transversely at finitely many points z 1 , … , z k z_{1}, \ldots, z_{k} z 1 , … , z k , and avoids other components. Then, using Step 2, we can repeat arguments of Sibley and Wentworth, and show that
n α [ Σ ] ⋅ [ Z α ] = 1 8 π 2 ∫ Σ ∩ ( ⋃ a = 1 k B ε ( z a ) ) ( Tr ( F A j ∧ F A j ) − Tr ( F A ∞ ∧ F A ∞ ) ) − ∫ Σ ∩ ∂ ( ⋃ a = 1 k B ε ( z a ) ) CS ( A j , A ∞ ) . n_{\alpha}[\Sigma]\cdot[Z_{\alpha}]=\frac{1}{8\pi^{2}}\int_{\Sigma\cap\left(\bigcup_{a=1}^{k}B_{\varepsilon}(z_{a})\right)}\left(\operatorname{Tr}(F_{A_j}\wedge F_{A_j})-\operatorname{Tr}(F_{A_{\infty}}\wedge F_{A_{\infty}})\right)-\int_{\Sigma\cap\partial\left(\bigcup_{a=1}^{k}B_{\varepsilon}(z_{a})\right)}\operatorname{CS}(A_j,A_{\infty}). n α [ Σ ] ⋅ [ Z α ] = 8 π 2 1 ∫ Σ ∩ ( ⋃ a = 1 k B ε ( z a ) ) ( Tr ( F A j ∧ F A j ) − Tr ( F A ∞ ∧ F A ∞ ) ) − ∫ Σ ∩ ∂ ( ⋃ a = 1 k B ε ( z a ) ) CS ( A j , A ∞ ) .
By Step 1, we can take the limit of the right hand side, and obtain that it is equal to m α [ Σ ] ⋅ [ Z α ] m_{\alpha}[\Sigma]\cdot[Z_{\alpha}] m α [ Σ ] ⋅ [ Z α ] .
Consider the current
∑ α m α Z α − ∑ β n β W β . \sum_{\alpha}m_{\alpha}Z_{\alpha}-\sum_{\beta}n_{\beta}W_{\beta}. α ∑ m α Z α − β ∑ n β W β .
Here Z α ⊂ Z a n Z_{\alpha}\subset Z_{an} Z α ⊂ Z an , and W β ⊂ Z a l g W_{\beta}\subset Z_{alg} W β ⊂ Z a l g are codimension 2 irreducible components of Z a n Z_{an} Z an and Z a l g Z_{alg} Z a l g . As in [39 ] we see, that this current is nonnegative and exact. Hence it is zero, and any codimension 2 component of Z a n Z_{an} Z an lies in Z a l g Z_{alg} Z a l g with the same multiplicity.
Step 4: We showed, that the codimension 2 components of Z a n Z_{an} Z an and Z a l g Z_{alg} Z a l g coincide, and their multiplicities are the same. Thus, the remaining part is to show that the higher codimension components coincide as sets. This can be done again as Sibley and Wentworth did after the proof of [39 Proposition 4.2] . ▪
Beijing Institute of Mathematical Sciences and Applications, China nklemyatin@bimsa.cn