Compactness results for neckstretching

Similar documents
INSTANTON MODULI AND COMPACTIFICATION MATTHEW MAHOWALD

Math. Res. Lett. 13 (2006), no. 1, c International Press 2006 ENERGY IDENTITY FOR ANTI-SELF-DUAL INSTANTONS ON C Σ.

Donaldson Invariants and Moduli of Yang-Mills Instantons

Morse theory and stable pairs

SYMPLECTIC MANIFOLDS, GEOMETRIC QUANTIZATION, AND UNITARY REPRESENTATIONS OF LIE GROUPS. 1. Introduction

SYMPLECTIC GEOMETRY: LECTURE 5

LECTURE 6: J-HOLOMORPHIC CURVES AND APPLICATIONS

REAL INSTANTONS, DIRAC OPERATORS AND QUATERNIONIC CLASSIFYING SPACES PAUL NORBURY AND MARC SANDERS Abstract. Let M(k; SO(n)) be the moduli space of ba

THE HODGE DECOMPOSITION

MATH 263: PROBLEM SET 1: BUNDLES, SHEAVES AND HODGE THEORY

September 27, :51 WSPC/INSTRUCTION FILE biswas-loftin. Hermitian Einstein connections on principal bundles over flat affine manifolds

Lecture 5: Hodge theorem

LECTURE 5: COMPLEX AND KÄHLER MANIFOLDS

Higgs Bundles and Character Varieties

Notes by Maksim Maydanskiy.

CHARACTERISTIC CLASSES

Mathematical Research Letters 2, (1995) A VANISHING THEOREM FOR SEIBERG-WITTEN INVARIANTS. Shuguang Wang

Scalar curvature and the Thurston norm

L19: Fredholm theory. where E u = u T X and J u = Formally, J-holomorphic curves are just 1

Determinant lines and determinant line bundles

As always, the story begins with Riemann surfaces or just (real) surfaces. (As we have already noted, these are nearly the same thing).

Banach space valued Cauchy-Riemann equations with totally real boundary conditions corrected

CONSIDERATION OF COMPACT MINIMAL SURFACES IN 4-DIMENSIONAL FLAT TORI IN TERMS OF DEGENERATE GAUSS MAP

LAGRANGIAN BOUNDARY CONDITIONS FOR ANTI-SELF-DUAL INSTANTONS AND THE ATIYAH-FLOER CONJECTURE. Katrin Wehrheim

Quasi Riemann surfaces

Donaldson and Seiberg-Witten theory and their relation to N = 2 SYM

APPROXIMATE YANG MILLS HIGGS METRICS ON FLAT HIGGS BUNDLES OVER AN AFFINE MANIFOLD. 1. Introduction

The Yang-Mills equations over Klein surfaces

Lagrangian Intersection Floer Homology (sketch) Chris Gerig

Quasi Riemann surfaces II. Questions, comments, speculations

Cohomology of the Mumford Quotient

Riemannian Curvature Functionals: Lecture III

CALCULUS ON MANIFOLDS. 1. Riemannian manifolds Recall that for any smooth manifold M, dim M = n, the union T M =

HYPERKÄHLER MANIFOLDS

Elliptic Regularity. Throughout we assume all vector bundles are smooth bundles with metrics over a Riemannian manifold X n.

Hyperkähler geometry lecture 3

ON NEARLY SEMIFREE CIRCLE ACTIONS

fy (X(g)) Y (f)x(g) gy (X(f)) Y (g)x(f)) = fx(y (g)) + gx(y (f)) fy (X(g)) gy (X(f))

Moment map flows and the Hecke correspondence for quivers

Stable complex and Spin c -structures

L E C T U R E N O T E S O N H O M O T O P Y T H E O R Y A N D A P P L I C AT I O N S

ERRATA FOR INTRODUCTION TO SYMPLECTIC TOPOLOGY

B 1 = {B(x, r) x = (x 1, x 2 ) H, 0 < r < x 2 }. (a) Show that B = B 1 B 2 is a basis for a topology on X.

The Calabi Conjecture

Stable bundles on CP 3 and special holonomies

Part V. 17 Introduction: What are measures and why measurable sets. Lebesgue Integration Theory

Introduction (Lecture 1)

arxiv:math/ v1 [math.ag] 18 Oct 2003

EXISTENCE THEORY FOR HARMONIC METRICS

LECTURE 2: SYMPLECTIC VECTOR BUNDLES

Summary of Prof. Yau s lecture, Monday, April 2 [with additional references and remarks] (for people who missed the lecture)

Hodge theory for bundles over C algebras

RIEMANN S INEQUALITY AND RIEMANN-ROCH

Kähler manifolds and variations of Hodge structures

NOTES ON FIBER BUNDLES

Lemma 1.3. The element [X, X] is nonzero.

Orientation transport

Compact 8-manifolds with holonomy Spin(7)

Vortex Equations on Riemannian Surfaces

Two simple ideas from calculus applied to Riemannian geometry

k=0 /D : S + S /D = K 1 2 (3.5) consistently with the relation (1.75) and the Riemann-Roch-Hirzebruch-Atiyah-Singer index formula

THE VORTEX EQUATION ON AFFINE MANIFOLDS. 1. Introduction

Geometric Composition in Quilted Floer Theory

GEOMETRIC QUANTIZATION

CW-complexes. Stephen A. Mitchell. November 1997

LECTURE 11: SYMPLECTIC TORIC MANIFOLDS. Contents 1. Symplectic toric manifolds 1 2. Delzant s theorem 4 3. Symplectic cut 8

1. Geometry of the unit tangent bundle

Dirac Operator. Göttingen Mathematical Institute. Paul Baum Penn State 6 February, 2017

J-holomorphic curves in symplectic geometry

Holomorphic line bundles

Math 249B. Nilpotence of connected solvable groups

NOTES ON HOLOMORPHIC PRINCIPAL BUNDLES OVER A COMPACT KÄHLER MANIFOLD

Max-Planck-Institut für Mathematik Bonn

Math 550 / David Dumas / Fall Problems

Mirror Symmetry: Introduction to the B Model

Instanton homology and symplectic fixed points

Constructing compact 8-manifolds with holonomy Spin(7)

Introduction to Index Theory. Elmar Schrohe Institut für Analysis

An Invitation to Geometric Quantization

MASLOV INDEX. Contents. Drawing from [1]. 1. Outline

We have the following immediate corollary. 1

Nodal symplectic spheres in CP 2 with positive self intersection

Broken pencils and four-manifold invariants. Tim Perutz (Cambridge)

Clifford Algebras and Spin Groups

Finite-dimensional representations of free product C*-algebras

Formality of Kähler manifolds

THE EULER CHARACTERISTIC OF A LIE GROUP

Linear connections on Lie groups

RIMS-1897 On multiframings of 3-manifolds By Tatsuro SHIMIZU December 2018 RESEARCH INSTITUTE FOR MATHEMATICAL SCIENCES KYOTO UNIVERSITY, Kyoto, Japan

LECTURE 10: THE ATIYAH-GUILLEMIN-STERNBERG CONVEXITY THEOREM

Instanton Floer homology with Lagrangian boundary conditions

Math 210C. A non-closed commutator subgroup

Math 350 Fall 2011 Notes about inner product spaces. In this notes we state and prove some important properties of inner product spaces.

HARMONIC MAPS INTO GRASSMANNIANS AND A GENERALIZATION OF DO CARMO-WALLACH THEOREM

The Atiyah bundle and connections on a principal bundle

ON ALGEBRAIC ASPECTS OF THE MODULI SPACE OF FLAT CONNECTIONS

LECTURE 3 Functional spaces on manifolds

Exercises in Geometry II University of Bonn, Summer semester 2015 Professor: Prof. Christian Blohmann Assistant: Saskia Voss Sheet 1

Theorem 2. Let n 0 3 be a given integer. is rigid in the sense of Guillemin, so are all the spaces ḠR n,n, with n n 0.

NOTES ON FORMAL NEIGHBORHOODS AND JET BUNDLES

Transcription:

Compactness results for neckstretching limits of instantons David L. Duncan Abstract We prove that, under a suitable degeneration of the metric, instantons converge to holomorphic quilts. To prove the main results, we develop estimates for the Yang-Mills heat flow on surfaces and cobordisms. Contents 1 Introduction 1 2 Background, notation and conventions 3 2.1 Gauge theory................................. 4 2.2 The complexified gauge group....................... 7 3 Compactness for products S Σ 11 3.1 Small curvature connections in dimension 2................ 15 3.2 Proof of Lemma 3.5.............................. 31 3.3 Proof of Theorem 3.3............................. 34 4 Compactness for cylinders R Y 36 4.1 The heat flow on cobordisms........................ 41 4.2 Uniform elliptic regularity.......................... 46 4.3 Proof of Lemma 4.2.............................. 65 4.4 Proof of Theorem 4.1............................. 67 5 Perturbations 73 1 Introduction In Donaldson theory one obtains invariants of 4-manifolds Z by counting instantons on a suitable bundle over Z. On the other hand, in symplectic geometry one can obtain invariants of Z by counting J-holomorphic curves in a suitable symplectic manifold associated to Z. For example, when Z = S Σ is a product of surfaces, then the relevant J-holomorphic curves are maps from S into the moduli space of 1

flat connections on Σ. There is a heuristic argument, dating back to Atiyah [1], that suggests these invariants are the same. Indeed, consider the case Z = S Σ, and fix a metric in which the S-fibers are very large equivalently, the Σ-fibers are very small. Motivated by Atiyah s terminology, we refer to this type of metric as a neckstretching metric, and use ɛ 2 to denote the volume of S. Then with respect to such a metric, the instanton equation splits into two equations: the first is essentially the holomorphic curve equation, and the second shows that the curvature of the instanton is bounded by ɛ 2 in the Σ-directions. In particular, taking ɛ to zero, the instanton equation formally recovers the J-holomorphic curve equation. In this paper we formalize the heuristic of the previous paragraph in the case when Z = S Σ, and also when Z = R Y, where Y has positive first Betti number. In each case, we prove that if ɛ ν is a sequence converging to zero and A ν is an instanton with respect to the ɛ ν -metric, then a subsequence of the A ν converges, in a suitable sense, to a J-holomorphic curve. The main convergence results are stated precisely in Theorem 3.3 for S Σ, and Theorem 4.1 for R Y. When Z = R Y our main theorem extends results of Dostoglou-Salamon in [8] [9], where they consider the special case where Y is a mapping torus. The case Z = S Σ serves as a model for the technically more difficult analysis necessary for R Y. The results of this paper can be extended quite naturally to more general 4-manifolds, however we leave a full treatment of such extensions to future work. The proofs of the main results proceed roughly as follows: As mentioned, when ɛ is small each instanton A has curvature that is small in the Σ-direction. Motivated by Donaldson [5], our strategy is to use an analytic Narasimhan-Seshadri correspondence on Σ to map each restriction A {s} Σ to some nearby flat connection NSA {s} Σ on Σ. This correspondence preserves the equations in the sense that the limiting connection NSA {s} Σ is holomorphic. Then the convergence result for the case Z = S Σ essentially follows from Gromov s compactness theorem for holomorphic curves. In the case when Z = R Y, the situation is considerably more difficult. To describe the difficulty we need to digress to discuss the neck-stretching metric used. Fix a circle-valued Morse function for Y. Then we stretch the metric only in a fixed region that is bounded away from the critical points. In this set-up, we expect that the limiting J-holomorphic curve will now have Lagrangian boundary conditions, where the Lagrangians are associated to the Morse critical points in a natural way. Now suppose A is an instanton with respect to the metric just described. Then NSA {s} Σ continues to be holomorphic in this case, but the essential difficulty is that it only has approximate Lagrangian boundary conditions. The standard Gromov compactness theorem breaks down when one does not have Lagrangian boundary conditions on the nose. We get around this using the following ingredients. First, we establish several C 1 -estimates for the map NS see Section 3.1.3, and then W 2,2 - estimates for instantons A see Section 4.2. These allow us to control the behavior of the holomorphic curves NSA {s} Σ near the boundary. The last ingredient is provided by the Yang-Mills heat flow on 3-manifolds with boundary. This gives us candidates for what the boundary conditions should be, if they were Lagrangian. Combining this with the C 1 - and W 2,2 -estimates allows us to reprove a version of 2

Gromov s compactness theorem for almost Lagrangian boundary conditions. In Section 2 we introduce our notation and conventions. We begin Section 3 by precisely stating our compactness result for S Σ. We then discuss the Narasimhan- Seshadri correspondence and develop the necessary estimates. We conclude Section 3 with the proof of the compactness result in the case of S Σ. Section 4 opens by stating the compactness result for R Y. We then discuss the heat flow on 3- manifolds with boundary, and prove the compactness theorem for R Y. The choice of neck-stretching metric used for R Y is by no means new. Indeed, the use of a real-valued Morse function on Y and the associated metric is again due to Atiyah [1]. In this case, the conjecture relating the instanton and symplectic invariants of R Y came to be known as the Atiyah-Floer conjecture due to considerations on the instanton side, one typically requires that Y is a homology 3-sphere. Unfortunately, at the time of writing, the Atiyah-Floer conjecture is illposed due to complications on the symplectic side arising from reducible connections; however, see [28]. Circle-valued Morse functions on Y were introduced as a mechanism for ruling out these troublesome reducible connections the existence of a suitable circle-valued Morse function requires that Y have positive first Betti number. This approach appears, in various ways, in work by Dostoglou-Salamon [7], and then later in Wehrheim [33] and Wehrheim-Woodward [35]. The upshot with this circle-valued approach is that the symplectic Floer homology and instanton Floer homology of Y are both well-defined; see [15], [35], [13], [14] and [3]. The relevant conjecture identifying these homology theories has come to be known as the quilted Atiyah-Floer conjecture [10], with the term quilted reflecting the possible existence of Morse critical points. Note: Upon completion of this project, the author learned of an earlier work [25] wherein Nishinou proves the result of Theorem 3.3 using essentially the same techniques used here. Included here are many details that are not present in [25]. Moreover, the technically more difficult Theorem 4.1 appears to be new. Acknowledgments: The author is grateful to his thesis advisor Chris Woodward for his insight and valuable suggestions. In addition, the author benefited greatly from discussions with Katrin Wehrheim, via Chris Woodward, of her unpublished work [33] on the Atiyah-Floer conjecture. This unpublished work analyzes various bubbles that appear in the limit of instantons with Lagrangian boundary conditions with degenerating metrics, and it outlines the remaining problems in such an approach. The present paper takes a different approach that avoids instantons with Lagrangian boundary conditions. This work was partially supported by NSF grants DMS 0904358 and DMS 1207194. 2 Background, notation and conventions Throughout this section X will denote an oriented manifold. Given a fiber bundle F X, we will use ΓF to denote the space of smooth sections. If V X is a vector bundle, then we will write Ω k X, V := ΓΛ k T X V for the space of k-forms with values in V, and we set 3

Ω X, V := k Ω k X, V. If V X is the trivial rank-1 bundle then we will write Ω k X for Ω k X, V. Given a metric on X and a connection on V, we can define Sobolev norms on ΓV in the usual way. We will use W k,p X, V to denote the completion of ΓV with respect to the W k,p -norm. We note that the usual Sobolev embedding and compactness statements for W k,p X, R hold equally well for W k,p X, V when V is finite-dimensional. If V is infinite-dimensional then we assume V is a Banach bundle and note that the usual embeddings hold here as well, but the compactness of these embeddings is very subtle question. We will typically not keep track of bundle V in the notation for the Sobolev norms. For example, we will use the same symbol L 2 X for the norm on ΓTX as for the norm on ΓT X V. Now suppose P X is a principal G-bundle, where G is a Lie group with Lie algebra g. Given a manifold M and homomorphism G DiffM, we can define the associated bundle P G M := P M/G. This is naturally a fiber bundle over X with fiber M. If M has additional structure, and the image of G DiffM respects this structure, then this additional structure is passed to the fiber bundle P G M. For example, when V is a vector space and G GLV DiffV is a representation, then PV := P G V is a vector bundle. The most important example for us comes from the adjoint representation G GLg Diffg. This respects the Lie algebra structure, and so the adjoint bundle Pg := P G g is a vector bundle with a Lie bracket on the fibers. This fiber-lie bracket combines with the wedge product to determine a graded Lie bracket on the vector space Ω X, Pg, and we denote this by µ ν [µ ν]. Similarly, if g is equipped with an Adinvariant inner product,, then this determines an inner product on the fibers of Pg and moreover combines with the wedge to form a graded bilinear map Ω X, Pg Ω X, Pg Ω X that we denote by µ ν µ ν. 2.1 Gauge theory Let P X be a principal G-bundle. We will write { AP = A Ω 1 P, g g P A = Adg 1 A, ι ξp A = ξ, g G ξ g } for the space of connections on P. Here g P resp. ξ P is the image of g G resp. ξ g under the map G DiffP resp. g VectP afforded by the group action. It follows that AP is an affine space modeled on Ω 1 X; Pg, and we denote the affine action by V, A A + V. In particular, AP is a smooth infinite dimensional manifold with tangent space Ω 1 X, Pg. Each connection A AP determines a covariant derivative d A : Ω X, Pg Ω +1 X, Pg and a curvature 2-form F A Ω 2 X, Pg. These satisfy d A+V = d A + [V ] and F A+V = F A + d A V + 1 2 [V V]. We say that a connection A is irreducible if d A is injective on 0-forms. 4

Given a metric on X, we can define the formal adjoint d A, which satisfies d A V, W L 2 = V, d A W L 2 for all compactly supported V, W Ω X, Pg. Here, L 2 is the L 2 -inner product coming from the metric on X. A connection A is flat if F A = 0. We will denote the set of flat connections on P by A flat P. If A is flat then im d A ker d A and we can form the harmonic spaces HA k := ker d A Ω k X, Pg Hk A X, Pg := im d A Ω k 1 X, Pg, H A := HA k. k Suppose X is compact with possibly empty boundary, and let denote the restriction Ω X, Pg Ω X, Pg X. Then the Hodge isomorphism [31, Theorem 6.8] says H A = kerd A d A, Ω X, Pg = H A imd A im d A, 1 for any flat connections A on X, where the summands on the right are L 2 -orthogonal. We will treat these isomorphisms as identifications. From the first isomorphism in 1 we see that H A is finite dimensional since d A d A is elliptic. We will use proj A : Ω X, Pg H A to denote the projection; see Lemma 3.10 for an extension to the case where A is not flat. Example 2.1. Suppose X = Σ is a closed, oriented surface equipped with a metric. Then the pairing ωµ, ν := Σ µ ν is a symplectic form on the vector space Ω1 X, Pg. Note that changing the orientation on Σ replaces ω by ω. On surfaces, the Hodge star squares to -1 on 1-forms and so defines a complex structure on Ω 1 Σ, Pg. It follows that the triple Ω 1 Σ, Pg,, ω is Kähler. If α AP is flat, then H 1 α Ω 1 Σ, Pg is a finite-dimensional Kähler subspace. Now suppose X is 4-manifold. Then on 2-forms the Hodge star squares to the identity, and it has eigenvalues ±1. Denoting by Ω ± X, Pg the ±1 eigenspace of, we have an L 2 -orthogonal decomposition Ω 2 X, Pg = Ω + X, Pg Ω X, Pg. The elements of Ω X, Pg are called anti-self dual 2-forms. A connection A X is said to be anti-self dual ASD or an instanton if its curvature F A Ω X, Pg is an anti-self dual 2-form; that is, if F A + F A = 0. Similarly, a connection A is self dual if F A Ω + X, Pg. The space of connections AP admits a function YM P : AP R, A 1 2 F A 2 L 2 called the Yang-Mills functional. Obviously, the global minimizers of this function are the flat connections, when they exist. However, in high dimensions the existence 5

of flat connections is rare due to topological obstructions. In dimension four the minimizers are the self dual or ASD connections. A gauge transformation is an equivariant bundle map U : P P covering the identity. The set of gauge transformations on P forms a Lie group, called the gauge group, and is denoted GP. This is naturally a Lie group with Lie algebra Ω 0 X, Pg under the map R exp R, where exp : g G is the Lie-theoretic exponential. The gauge group acts on the left on the space AP by pulling back by the inverse: U, A UA := U 1 A, 2 for U GP, A AP. In terms of a faithful matrix representation of G we can write this as U A = U 1 AU + U 1 du, where the concatenation appearing on the right is just matrix multiplication, and du is the linearization of U when viewed as a G-equivariant map P G. The infinitesimal action of GP at A AP is Ω 0 X, Pg Ω 1 X, Pg, R d A R 3 More generally, the derivative of 2 at U, A is UΩ 0 X, Pg Ω 1 X, Pg Ω 1 M, Pg UR, V AdUd A R + AdUV. where, in writing this expression, we have chosen a faithful matrix representation G GLV. This allows us to view GP and its Lie algebra Ω 0 X, Pg as subspaces of the same algebra ΓPEndV, and so it makes sense to identify the tangent space of GP at U with UΩ 0 X, Pg, as we have done in 4. The gauge group also acts on the left on Ω X, Pg by the pointwise adjoint action. The curvature of A AP transforms under U GP by F U A = AdU 1 F A. This shows that GP restricts to an action on A flat P and, in 4-dimensions, the instantons. In general, we will tend to use capital letters A, U to denote connections and gauge transformations on 4-manifolds Z, lower case letters a, u to denote connections and gauge transformations on 3-manifolds Y, and lower case Greek letter α, µ to denote gauge transformations on surfaces Σ. We will be interested in the case G = PUr for r 2. We equip the Lie algebra g = sur EndC r with the inner product µ, ν := κ r trµ ν; here κ r > 0 is arbitrary, but fixed. On manifolds X of dimension at most 4, the principal PUr- bundles P X are classified, up to bundle isomorphism, by two characteristic classes t 2 P H 2 X, Z r and q 4 P H 4 X, Z. These generalize the 2nd Stiefel- Whitney class and 1st Pontryagin class, respectively, to the group PUr; see [38]. When X is a closed, oriented 4-manifold, there is also a Chern-Weil formula q 4 P = r 4π 2 F A F A 5 κ r X which holds for any connection A AP. 4 6

Consider a principal PUr-bundle P X where we assume dimx 3. Then there are maps η : GP H 1 X, Z r, deg : GP H 3 X, Z called the parity and degree. These detect the connected components of GP in the sense that u can be connected to u by a path if and only if ηu = ηu and degu = degu. We denote by G 0 P the identity component of GP. See [11]. Suppose X = Σ is a closed, connected, oriented surface, and P Σ is a principal PUr-bundle with t 2 P [Σ] Z r a generator. It can be shown that all flat connections on P on irreducible, and that G 0 P acts freely on A flat P. Moreover, one has that MP := A flat P/G 0 P is a compact, simply-connected, smooth manifold with tangent space at [α] MP canonically identified with H 1 α for any choice of representative α [α]. It follows from Example 2.1 that MP is a symplectic manifold, and any metric on Σ determines an almost complex structure J Σ on MP that is compatible with the symplectic form. See [35] for more details regarding these assertions. Now suppose Y ab is an oriented elementary cobordism between closed, connected, oriented surfaces Σ a and Σ b. Fix a PUr-bundle Q ab Y ab with t 2 Q ab [Σ a ] a generator of Z r. Then the flat connections on Q ab are irreducible, and the quotient A flat Q ab /G 0 Q ab is a finite-dimensional, simply-connected, smooth manifold. Restricting to the two boundary components induces an embedding A flat Q ab /G 0 Q ab MQ Σa MQ Σb, and we let LQ ab denote the image. It follows that LQ ab MQ Σa MQ Σb is a smooth Lagrangian submanifold [35], where the superscript in MQ Σa means that we have replaced the symplectic form with its negative. More generally, if G is a compact Lie group, and P X is a principal G-bundle, then we can consider the space A flat P/GP. Note that we are not quotienting by the identity component here, so this will not be MP when G = PUr and X = Σ. This space has a natural topology, and it follows from Uhlenbeck s compactness theorem that this space is compact when X is compact. However, it is rarely a smooth manifold. 2.2 The complexified gauge group Let G be a compact, connected Lie group and fix a faithful Lie group embedding ρ : G Un for some n. We identify G with its image in Un. Using the standard representation of Un on C n, define E := P G C n. The standard real inner product and complex structure on C n are preserved by Un, and hence by G. It follows that there is an induced real inner product, and a complex structure J E on E, and that these are compatible in the sense that J E ξ, J E η = ξ, η for all ξ, η Ω 0 Σ, E. 7

Suppose Σ, j Σ is a Riemann surface. We will write Ω k,l Σ, E for the smooth E-valued forms of type k, l. Observe that j Σ acts by the Hodge star on 1-forms. Consider the space { CE := D : Ω 0 Σ, E Ω 0,1 Σ, E D f ξ = f Dξ + f ξ, for ξ Ω 0 Σ, E, f Ω 0 Σ of Cauchy-Riemann operators on E. This can be naturally identified with the space of holomorphic structures on E see [23, Appendix C]. Each element D CE has a unique extension to an operator D : Ω j,k Σ, E Ω j,k+1 Σ, E satisfying the Leibniz rule. Consider the space of C-linear covariant derivatives on E: { D : Ω 0 Σ, E Ω 1 Σ, E D f ξ = f Dξ + d f ξ, for ξ Ω 0 Σ, E, f Ω 0 Σ There is a natural map from this space onto CE defined by D 1 2 D + J ED j Σ 6 Here and below we are using the symbol to denote composition of operators. For example, if M : ΩΣ, E ΩΣ, E is a derivation we define M j Σ : ΩΣ, E ΩΣ, E to be the derivation given by the formula ι X M j Σ ξ := ι jσ X Mξ; here ι X is contraction with a vector X. The map 6 is surjective, and restricts to a C-linear isomorphism on the set AE := D : Ω0 Σ, E Ω 1 Σ, E } D f ξ = f Dξ + d f ξ, d ξ, η = Dξ, η + ξ, Dη for ξ, η Ω 0 Σ, E, f Ω 0 Σ of Hermitian C-linear covariant derivatives on E. Let Pg C denote the complexification of the vector bundle Pg. Then we have bundle inclusions Pg Pg C EndE, where EndE is the bundle of complex linear automorphisms of E and the latter inclusion is induced by the embedding ρ. Each connection α AP induces a covariant derivative d α,ρ : Ω k Σ, E Ω k+1 Σ, E, and so we have a map AP AE. Furthermore, this map is an embedding of Ω 1 Σ, Pg-affine spaces, where Ω 1 Σ, Pg acts on AE via the inclusion Ω 1 Σ, Pg Ω 1 Σ, EndE. In particular, restricting to the image of AP in AE, the map 6 becomes an embedding AP CE, α α := 1 2 dα,ρ + J E d α,ρ j Σ The image of 7 is the set of covariant derivatives that preserve the G-structure, and we denote this image by CP. See [23, Appendix C] for the case when G = Un. The space CP is an affine space modeled on Ω 0,1 Σ, Pg C. Similarly, AP is } 7 8

an affine space modeled on Ω 1 Σ, Pg, and 7 intertwines these affine actions, where we identify Ω 1 Σ, Pg with Ω 0,1 Σ, Pg C by sending µ to its anti-linear part µ 0,1 := 1 2 µ + J Eµ j Σ. To summarize, we have a commutative diagram = AP CP AE = CE where the vertical arrows are inclusions and everything is equivariant with respect to the action of Ω 1 Σ, Pg. Let α AP be a connection on P with curvature F α Ω 2 Σ, Pg. Consider the associated covariant derivative d α,ρ AE as well as its curvature F α,ρ = d α,ρ d α,ρ Ω 2 Σ, EndE. Since the representation ρ is faithful, we have pointwise estimates of the form c F α,ρ F α C F α,ρ ; this allow us to discuss curvature bounds in terms of either F α or F α,ρ. To define the complexified gauge group, we first recall some basic properties of the complexification of compact Lie groups. See [19] or [17] for more details on this material. Since G is compact and connected, there is a connected complex group G C and an embedding G G C such that G is a maximal compact subgroup of G C, and the Lie algebra g C = LieG C is the complexification of g = LieG. This group G C is unique up to natural isomorphism and is called the complexification of G. We may assume that the representation ρ : G Un from above extends to an embedding G C GLC n, and we identify G C with its image see [17, Proof of Theorem 1.7]. Then we have G = { u G C u u = Id }, where u denotes the conjugate transpose on GLC n. It follows from the standard polar decomposition in GLC n that we can write G C = {g expiξ g G, ξ g }, and this decomposition is unique. The same holds true if we replace g expiξ by expiξg. It is then immediate that g expiξ = expiadgξg 8 for all g G and all ξ g. We can now define the complexified gauge group on P to be GP C := ΓP G G C. As in the real case, we may identify Ω 0 Σ, P g C with the Lie algebra of GP C via the map ξ exp ξ, 9 hence the Lie group theoretic exponential map on GP C is given pointwise by the exponential map on G C. It follows by the analogous properties of G C that each element of GP C can be written uniquely in the form 9

g expiξ 10 for some g GP and ξ Ω 0 Σ, Pg, and 8 continues to hold with g, ξ interpreted as elements of GP, Ω 0 Σ, Pg, respectively. The complexified gauge group acts on CP by GP C CP CP, µ, D µ D µ 1 11 Viewing GP as a subgroup of GP C in the obvious way, then the identification 7 is GP-equivariant. We can then use 7 and 11 to define an action of the larger group GP C on AP, extending the GP-action. We denote the action of µ GP C on α by µ 1 α or simply µα when there is no room for confusion. Explicitly, the action on AP takes the form d µ 1 α,ρ = µ 1 α µ + µ α µ 1. where the dagger is applied pointwise. In particular, the infinitesimal action at α AP is Ω 0 Σ, Pg C Ω 1 Σ, Pg, ξ + iζ d α,ρ ξ + d α,ρ ζ 12 More generally, the derivative of the map µ, α µ 1 α at µ, α with µ GP an element of the real gauge group is the map µ Ω 0 Σ, Pg iω 0 Σ, Pg Ω 1 Σ, Pg Ω 1 Σ, Pg given by µξ + iζ, η Adµ d α ξ + d α ζ + η } = {d µ 1 α µξ + d µ 1 α µζ µ 1 + Adµη 13 Compare with 4. Here we are using the fact that GP C and its Lie algebra both embed into the space ΓP G EndC n, and so it makes sense to multiply Lie group and Lie algebra elements. The curvature transforms under µ GP C by µ 1 F µ 1 α,ρ µ = F α,ρ + α h 1 α h, 14 where we have set h = µ µ. We will mostly be interested in this action when µ = expiξ for ξ Ω 0 Σ, Pg, in which case the action can be written as where we have set exp iξ F exp iξ α,ρ expiξ = Fα, ξ, 15 Fα, ξ := F α,ρ + α exp 2iξ α exp2iξ. 16 10

It will be useful to define the real gauge group on E and the complexified gauge group on E by, respectively, GE := ΓP G Un, GE C := ΓP G GLC n. Note that the complexification of Un is GLC n, so this terminology is consistent, and in fact motivates, the terminology above. These are both Lie groups with Lie algebras LieGE = ΓP G un and LieGE = ΓP G EndC n, respectively, where we are identifying EndC n with the Lie algebra of GLC n. We have the obvious inclusions The space GE C acts on CE by the map GP GP C GE GE C GE C CE CE, µ, D µ D µ 1 17 Using 6, this induces an action of GE C on AE hence an action of GE on AE, though, neither GE C nor GE restrict to actions on AP, unless G = Un. Finally, we mention that the vector spaces LieGE, LieGE C and LieGP C admit Sobolev completions. For example, the space LieGP C k,q is the W k,q -completion of the vector space ΓP G Pg C. When we are in the continuous range for Sobolev embedding i.e., when kq > 2 then these are Banach Lie algebras. Similarly, when we are in the continuous range we can form the Banach Lie groups G k,q E, G k,q E C and G k,q P C by taking the W k,q -completions of the groups of smooth functions GE, GE C and GP C, which we view as lying in the vector space ΓP G EndC n k,q. The complexified gauge action extends to a smooth action of G k,q E C on A k 1,q E, and this restricts to a smooth action of G k,q P C on A k 1,q P. See [34, Appendix B] for more details regarding the Sobolev completions of these spaces. 3 Compactness for products S Σ In this section we consider Z = S Σ, where S, g S, Σ, g Σ are closed, connected, oriented Riemannian surfaces. We also assume that Σ has positive genus. We will work with the metric g = proj S g S + proj Σ g Σ on Z; from now on we will typically drop the projections from the notation. For ɛ > 0, define a new metric by g ɛ := ɛ 2 g S + g Σ. Fix a principal PUr-bundle P Σ such that t 2 P [Σ] Z r is a generator, and let R Z be the pullback bundle under Z Σ. 11

Given orthonormal coordinates U, x = s, t for S, any connection A on R can be written as A {s,t} Σ = αs, t + φs, t ds + ψs, t dt, where αs, t is a connection on P and φs, t, ψs, t Ω 0 Σ, Pg. In fact, α can be defined in a coordinate-independent way as x αx := ι x A AP, where ι x : Σ = {x} Σ Z is the inclusion. We will say that a connection A on R is ɛ-asd or an ɛ-instanton if it is an instanton with respect to the metric g ɛ ; that is, if F A = ɛ F A, where ɛ is the Hodge star on Z coming from g ɛ. This can be written explicitly in terms of local coordinates as s α d α φ + Σ t α d α ψ = 0 s ψ t φ [ψ, φ] + ɛ 2 Σ F α = 0 18 where Σ is the Hodge star associated to the S-dependent metric g Σ. The ɛ-energy of A is Eɛ inst A := 1 F 2 A 2 ɛdvol ɛ = 1 F Z 2 A ɛ F A, Z where the norm and volume form are the ones induced by g ɛ ; this is exactly the Yang-Mills functional on AR with the metric g ɛ. If A is ɛ-asd, then Eɛ inst A = 2π 2 κ r r 1 q 4 R is a topological invariant by the Chern-Weil formula 5. On the symplectic side, we will be considering maps v : S MP, where MP = A flat P/G 0 P. Any such map v has a lift α : S A flat P if and only if the pullback G 0 P-bundle v A flat P S is trivial. When S is not closed, all maps v have smooth lifts since G 0 P is connected and S retracts to its 1-skeleton; however, when S is closed there are maps v that do not lift. Given α : S A flat P there is a unique section χ of the bundle T S Ω 1 Σ, Pg over S such that, for all x S and v T x S, the 1-form D x αv d α χv H 1 αx Ω1 Σ, Pg is αx-harmonic; here D x α : T x S T αx AP is the push-forward of α : S AP. This data naturally determines a connection A 0 AR by setting A 0 {x} S := αx + χx. If α is the lift of some v, then we call A 0 a representative of v, and we call A 0 a holomorphic curve representative if v is holomorphic. Example 3.1. Let U, s, t be local orthonormal coordinates for S. Then Dα U = ds s α + dt t α and χ U = φ ds + ψ dt, where φ, ψ : U Ω 0 Σ, Pg are the unique sections such that s αs, t d αs,t φs, t, t αs, t d αs,t ψs, t 12

are αs, t-harmonic. It follows that A 0 is a holomorphic curve representative if and only if s α d α φ + Σ t α d α ψ = 0 F α = 0, 19 where the equations should be interpreted as being pointwise in s, t S. We define the energy E symp A 0 of a representative A 0 to be the energy of the associated map v : S MP. Lemma 3.2. Given any ɛ > 0, the energy of a holomorphic curve representative A 0 is E symp A 0 = 2π 2 κ r r 1 q 4 R. Proof. Suppose A 0 represents v. Then the energy of v is 2 1 S Dv 2 dvol S, where Dv is the push-forward of v : S MP. In local coordinates, write A 0 = α + φ ds + ψ dt. Then Dv = ds s v + dt t v = ds β s + dt β t, where we have set β s := s α d α φ and β t := t α d α ψ. On the other hand, F A0 = ds β s + dt β t + ds dt s ψ t φ [ψ, φ] where we have used F α = 0. Then by 19 we have 1 2 Σ F A0 F A0 = Σ β s β t ds dt = β s 2 L 2 Σ ds dt = 1 2 Dv 2 dvol S. Integrating the right over S gives E symp A 0, and integrating the left over S gives 2π 2 κ r r 1 q 4 R by the Chern-Weil formula. Now we can state the main theorem of this section. Theorem 3.3. Fix 2 < q < and let R Z be as above. Suppose ɛ ν ν N is a sequence of positive numbers converging to 0, and assume that, for each ν, there is an ɛ ν -ASD connection A ν A 1,q loc R. Then there is i a finite set B S; ii a subsequence of the A ν still denoted A ν ; iii a sequence of gauge transformations U ν G 2,q loc R; and iv a holomorphic curve representative A A 1,q loc R 13

such that the restrictions sup ι x U ν ν A ν A C 0 Σ 0 x K converge to zero for every compact K S\B. The gauge transformations U ν can be chosen so that they restrict to the identity component G 0 P on each slice {x} Σ Z. Moreover, for each b B there is a positive integer m b > 0 such that for any ν, E symp A = Eɛ inst ν A ν 4π 2 κ r r 1 m b. 20 b B Throughout we will use notation of the form A {x} Σ = αx + χx, and U {x} Σ = µx for connections and gauge transformations, respectively. Then the conclusion of the theorem says that µ να ν converges to α in C 0 on compact sets in S\B, where these are viewed as maps from S to AP, with the C 0 -topology on AP. Here, the action of µ = µx is the gauge action on surfaces not 4-manifolds. Remark 3.4. a If one allows S to be a compact manifold-with-boundary, then the proof we give here remains valid, except the equality in 20 should be replaced by. This is due to holomorphic disk bubbles occurring at the boundary the convergence we prove here is not strong enough to show that these are non-trivial; see Remark 3.18. In Section 4 we provide tools for recovering the equality in a certain setting. b This theorem has a straight-forward extension to the case where S has cylindrical ends assuming one knows flat connections on the ends are non-degenerate; see Section 5. In this case, there can be energy loss at the ends, and so one obtains a limiting holomorphic curve on S with a finite number of broken cylindrical trajectories on the ends. Then Theorem 3.3 continues to hold provided one accounts for the energies of the broken trajectories on the the left-hand side 20. See also Theorem 4.1. c Suppose, for each ν, we have an open set S ν S that is a deformation retract of S, and with the further property that the S ν are increasing and exhausting: S ν S ν+1 and S = ν S ν. Then the statement of Theorem 3.3 continues to hold if we assume that A ν is defined on S ν Σ. d It is possible to show that the C 0 -convergence of the α ν to α implies Wɛ 1,2 ν -convergence of the A ν to A, where Wɛ 1,2 is the Sobolev norm defined with respect to the ɛ-dependent metric. A similar statement holds in the case of Theorem 4.1, as well. We defer the details to a future paper. As a stepping stone to Theorem 3.3, we first prove the following lemma; the assumptions allow us to rule out bubbling a priori. Lemma 3.5. Fix 2 < q <, and a submanifold S 0 S possibly with boundary. Let R 0 := R S0 Σ denote the restriction. Suppose ɛ ν ν N is a sequence of positive numbers not necessarily converging to zero, and suppose that for each ν there is an ɛ ν -ASD connection A ν A 1,q R 0 satisfying the following conditions for each compact K S 0. 14

i The slice-wise curvatures converge to zero: ii There is some constant C with ν sup F αν x L Σ 0. x K sup sup proj αν x D xα ν L 2 Σ C, ν x K where proj αν x is the harmonic projection, and D xα ν : T x S T αx A 1,q P is the push-forward. Then there is a subsequence of the connections still denoted A ν, a sequence of gauge transformations U ν on R 0, and a holomorphic curve representative A on R 0 such that α x µ ν x α ν x 0 21 sup x K for every compact K S 0, and sup x K C 0 Σ proj α x D xα Adµ 1 ν xproj αν x D L xα ν 0 22 p Σ for any compact K int S 0 and any 1 < p <. The connections A ν from Theorem 3.3 satisfy the same type of convergence as in 22, for compact K S\B. We also point out that the projection operator proj αν appearing in 22 can be removed by weakening the C 0 -convergence to L r -convergence. The proofs of Lemma 3.5 and Theorem 3.3 will appear in Sections 3.2 and 3.3, respectively. First we develop some machinery that will allow us to pass from ɛ- instantons to holomorphic curves. 3.1 Small curvature connections in dimension 2 In our proof of the various convergence results, we will encounter connections on surfaces that have small curvature. Here we develop a strategy for identifying nearby flat connections. The idea is to use the well-known fact that quotienting the subset of small curvature connections by the action of the complexified gauge group recovers the moduli space of flat connections called a Narasihman-Seshadri correspondence. The details of this procedure were originally carried out by Narasimhan and Seshadri [24]. They worked with unitary bundles, and this allowed them to use algebraic techniques. Later, their techniques were extended to more general structure groups by Ramanathan in his thesis [27]. See Kirwan s book [20] for a finitedimensional version. In preparation for a boundary-value problem, we need to work in an analytic category. Consequently, we adopt an approach of Donaldson [5], and use an implicit function theorem argument to arrive at a Narasimhan-Seshadri correspondence in our setting. This allows us to establish several C 1 -estimates that will be needed for our proof of the main theorems. 15

3.1.1 The analytic Narasimhan-Seshadri correspondence The goal of this section is to define a gauge-equivariant deformation retract NS : A ss A flat and establish some of its properties. Here A ss is a suitable neighborhood of A flat the superscript stands for semistable. The relevant properties of the map NS are laid out in Theorem 3.6, below. The proof will show that for each α A ss there is a purely imaginary complex gauge transformation µ such that µ α a flat connection, and µ is unique provided it lies sufficiently close to the identity. We then define NSα := µ α. After proving Theorem 3.6 below, where the map NS is formally defined, we spend the remainder of this section establishing useful properties and estimates for NS. For example, in the proof of Lemma 3.15 we establish the Narasimhan-Seshadri correspondence A ss /G0 C = A flat /G 0, and in Proposition 3.14 and Corollary 3.17 we show that, to first order, the map NS is the identity plus the L 2 -orthogonal projection to the tangent space of flat connections. Theorem 3.6. Suppose G is a compact, connected Lie group, Σ is a closed Riemannian surface, and P Σ is a principal G-bundle such that all flat connections are irreducible. Then for any 1 < q <, there are constants C > 0 and ɛ 0 > 0, and a G 2,q P-equivariant deformation retract { } NS P : α A 1,q P F α L q Σ < ɛ 0 A 1,q flat P 23 that is smooth with respect to the W 1,q -topology on the domain and codomain. Moreover, the map NS P is also smooth with respect to the L p -topology on the domain and codomain, for any 2 < p <. Remark 3.7. The restriction in the second part of the theorem to 2 < p < is merely an artifact of our proof, and it is likely that the conclusion holds for, say, 1 < p 2 as well. See Lemma 3.16. Proof of Theorem 3.6. Suppose we can define NS P on the set { } α A 1,q P dist W 1,q α, A 1,q flat P < ɛ 0, 24 for some ɛ 0 > 0, and show that it satisfies the desired properties on this smaller domain. Then the G 2,q -equivariance will imply that it extends uniquely to the flowout by the real gauge group: { } µ α A 1,q P µ G 2,q P, dist W 1,q α, A 1,q flat P < ɛ 0, and continues to have the desired properties on this larger domain. The next claim shows that this flow-out contains a neighborhood of the form appearing in the domain in 23, thereby reducing the problem to defining NS P on a set of the form 24. Claim: For any ɛ 0 > 0, there is some ɛ 0 > 0 with 16

{ α A 1,q } P F α L q { < ɛ 0 µ α A 1,q P µ G 2,q P, dist W 1,q α, A 1,q flat P < ɛ 0 }. For sake of contradiction, suppose that for every ɛ > 0 there is some connection α with F α L q < ɛ, but where µ α α 0 W 1,q ɛ 0, µ G 2,q P, α 0 A 1,q flat P 25 Then we can find a sequence of connections α ν with F αν L q 0, but 25 holds with α ν replacing α. By Uhlenbeck s weak compactness theorem, there is a sequence of gauge transformations µ ν G 2,q P such that, after possibly passing to a subsequence, µ να ν converges weakly in W 1,q to a limiting connection α. The condition on the curvature implies that α A 1,q flat P is flat. Moreover, the embedding W 1,q L 2q is compact, so the weak W 1,q -convergence of µ να ν implies that µ να ν converges strongly to α in L 2q. By redefining µ ν if necessary, we may suppose that µ να ν is in Coulomb gauge with respect to α, d α 0 µ να ν α = 0, and still retain the fact that µ να ν converges to α strongly in L 2q. This gives µ να ν α q W 1,q = µ να ν α q L q + d α µ να ν α q L q + d α µ να ν α q L q µ να ν α q L q + F αν q L q + 2 1 µ να ν α q L 2q C µ να ν α ql + F 2q αν q L q where we have used F α +ν = d α ν + 1 2 [ν ν]. Hence µ να ν α q W 1,q 0, in contradiction to 25. This proves the claim. To define NS P, it therefore suffices to show that for α sufficiently W 1,q -close to A flat P there is a unique Ξα Ω 0 Σ, Pg close to 0, with F expiξα α,ρ = 0. Once we have shown this, then we will define NS P α := expiξα α. Recall the definition of F in 16. In light of 15, finding Ξα is equivalent to solving for ξ in Fα, ξ = 0, for then Ξα := ξ. To solve for ξ we need to pass to suitable Sobolev completions. It follows from 14, and the Sobolev embedding and multiplication theorems, that F extends to a smooth map A 1,q P LieGP 2,q LieGP 0,q, whenever q > 1. Suppose α 0 is a flat connection. The linearization of F at α 0, 0 in the direction of 0, ξ is D α0,0f0, ξ = 2J E α0 α0 ξ, where we have used the fact that d α,ρ commutes with J E = i this is because the complex structure J E is constant and the elements of AP are unitary. Observe that j Σ acts by the Hodge star on vectors, so 17

d α0,ρd α0,ρ j Σ = d α0,ρ d α0,ρ, d α0,ρ j Σ d α0,ρ = F α0,ρ j Σ, Id. 26 Using this and the fact that F α0,ρ = 0, we have D α0,0f0, ξ = 1 2 α 0,ρξ where α0,ρ = d α 0,ρd α0,ρ + d α0,ρd α 0,ρ is the Laplacian. By assumption, all flat connections are irreducible, so Hodge theory tells us that the operator α0,ρ : LieGP 2,q LieGP 0,q is an isomorphism. Since α 0 is flat, the pair α 0, 0 is clearly a solution to Fα, ξ = 0. It therefore follows by the implicit function theorem that there are ɛ α0, ɛ α 0 > 0 such that, for any α A 1,q with α α 0 W 1,q < ɛ α0, there is a unique Ξ = Ξα LieGP 2,q with Ξα W 2,q < ɛ α 0 and Fα, Ξα = 0. The implicit function theorem also implies that Ξα varies smoothly in α in the W 1,q -topology. Moreover, by the uniqueness assertion, it follows that Ξα = 0 if α is flat. We need to show that ɛ α0 and ɛ α 0 can be chosen to be independent of α 0 A 1,q flat P. Since the moduli space A flat/g of flat connections on P is compact, it suffices to show that ɛ α0 = ɛ µ α 0, for all real gauge transformations µ G 2,q P, and likewise for ɛ α 0. Fix µ G 2,q P and α a connection that is W 1,q -close to α 0, then find Ξα as above. By 8 and the statement following 10 we have expiξαµ = µ expiadµ 1 Ξα. 27 Since the curvature is G 2,q P-equivariant, we also have 0 = Adµ 1 F expiξ α = F expiξµ α = F expiadµ 1 Ξ µ α. It follows that Ξµ α = Adµ 1 Ξα since Ξµ α is uniquely defined by F expiξµ α µ α = 0. We therefore have ɛ µ α 0 = ɛ α0 and ɛ µ α 0 = ɛ α 0, so we can take ɛ 0 to be the minimum of inf ɛ α 0 > 0 and inf [α 0 ] A flat /G [α 0 ] A flat /G ɛ α 0 > 0. This argument also shows that NS P is G 2,q P-equivariant. Finally, we show that NS P α depends smoothly on α in the L p -topology for p > 2. It suffices to show that α Ξα extends to a map A 0,p P LieGP 1,p that is smooth with respect to the specified topologies. To see this, note that F from 16 is well-defined as a map A 0,p P LieGP 1,p LieGP 1,p. and is smooth with respect to the specified topologies the restriction to p > 2 is required so that Sobolev multiplication is well-defined. Then the implicit function theorem argument we gave above holds verbatim to show that for each α sufficiently 18

L p -close to A 0,p flat P, there is a unique W1,p -small Ξα LieGP 1,p such that expi Ξα α is flat. Moreover, the assignment A 0,p flat P LieGP1,p, α Ξα is smooth. The uniqueness of Ξα and Ξα ensures that the former is indeed an extension of the latter. Remark 3.8. Let Π : A 1,q flat P A1,q flat P/G2,q P denote the projection. The above proof shows that the composition Π NS P is invariant under a small neighborhood of G 2,q P in G 2,q P C. Indeed, α and expiξ α both map to the same flat connection under NS P whenever they are both in the domain of NS P. More generally, differentiating the identities NS P exptφ α = exptφ NS P α, NS P expitφ α = NS P α at t = 0 gives D α NS P d α φ = d NSP α φ, D αns P d α φ = 0, where D α NS P is the linearization of NS P at α. 3.1.2 Analytic properties of almost flat connections This section is of a preparatory nature. The results extend several elliptic properties, which are standard for flat connections, to connections with small curvature. The following lemma addresses elliptic regularity for the operator d α on 0-forms. Lemma 3.9. Suppose G is a compact Lie group, Σ is a closed oriented Riemannian surface, and P Σ is a principal G-bundle such that all flat connections are irreducible. Let 1 < q <. Then there are constants C > 0 and ɛ 0 > 0 with the following significance. i Suppose that either α A 1,q P with F α L q Σ < ɛ 0, or α A 0,q P with α α L 2q Σ < ɛ 0 for some α A 0,2q flat P. Then the map d α : W 1,q Pg L q Pg is a Banach space isomorphism onto its image. Moreover, for all f W 1,q Pg the following holds f W 1,q Σ C d α f L q Σ. 28 ii For all α A 1,q P with F α L q Σ < ɛ 0, the Laplacian d αd α : W 2,q Pg L q Pg is a Banach space isomorphism. Moreover, for all f W 2,q Pg the following holds f W 2,q Σ C d α d α f L q Σ. 29 19

Proof. This is basically the statement of [8, Lemma 7.6], but adjusted a little to suit our situation. We prove ii, the proof of i is similar. The assumption that all flat connections α are irreducible implies that the kernel and cokernel of the elliptic operator d α d α : W 2,q Pg L q Pg are trivial. In particular, we have an estimate f W 2,q C d α d α f L q for all f W 2,q Pg, so the statement of the lemma holds when α = α is flat. Next, fix α A 1,q P and α A 1,q flat P. Then, by the above discussion, and the relation d α f = d α f + [α α, f ], we have f W 2,q is bounded by { C d αd α f L p C d α d α f L q + d α [ α α, f ] L q { C d α d α f L q + C f W 2,q + [α α [ α α, f ]] L q } d α α α L q + α α L 2q, for all f W 2,q Pg, where we have used the embeddings W 2,q W 1,q and W 2,q L in the last step. Now suppose that α α L 2q < 1/2CC is small. Then by composing α with a suitable gauge transformation, we may suppose α is in Coulomb gauge with respect to α, and still retain the fact that α α L 2q < 1/2CC. Then the above gives f W 2,q C d α d α f L q + 1 2 f W 2,q, which shows that d αd α is injective when sufficiently L 2q -close to the space of flat connections. Now we prove the lemma. Suppose ii in the statement of the lemma does not hold. Then there is some sequence of connections α ν with F αν L q 0, but the estimate 29 does not hold for any C > 0. By Uhlenbeck s weak compactness theorem, after possibly passing to a subsequence, there is some sequence of gauge transformations u ν, and a limiting flat connection α, such that α ν u να L 2q 0. So the discussion of the previous paragraph shows that, for ν sufficiently large, the estimate 29 holds with α replaced by α ν. This is a contradiction, and it proves the lemma. Now we move on to study the action of d α on 1-forms. First we establish a Hodge-decomposition result for connections with small curvature. For 2 q < and k Z, we will use the notation V k,q to denote the W k,q -closure of a vector subspace V W k,q T Σ Pg. The standard Hodge decomposition says W k,q T Σ Pg = H 1 α im d α k,q im d α k,q, 30 for any flat connection α. Here H 1 α is finite dimensional this dimension is independent of α A, and so is equal to its own W k,q -closure. Furthermore, the direct } 20

sum in 30 is L 2 -orthogonal, even though the spaces need not be complete in the L 2 -metric. We have a similar situation whenever α has small curvature, as the next lemma shows. Lemma 3.10. Assume that P Σ satisfies the conditions of Lemma 3.9, and let 1 < q < and k 0. Then there are constants ɛ 0 > 0 and C > 0 with the following significance. If α A 1,q P has F α L q Σ < ɛ 0, then H 1 α := ker d α k,q ker d α k,q W k,q T Σ Pg has finite dimension equal to dim H 1 α, for any flat connection α. Furthermore, the space H 1 α equals the L 2 -orthogonal complement of the image of d α d α: H 1 α = and so we have a direct sum decomposition W k,q T Σ Pg = H 1 α In particular, the L 2 -orthogonal projection im d α k,q im d α k,q, im d α k,q im d α k,q. 31 is well-defined. proj α : W k,q T Σ Pg H 1 α 32 Remark 3.11. It follows by elliptic regularity that the space ker d α k,q ker d α k,q consists of smooth forms. Moreover, when k 2/q k 2/q, the inclusion W k,q W k,q restricts to an inclusion of finite-dimensional spaces ker d α k,q ker d α k,q ker d α k,q ker d α k,q, and this map is onto by dimensionality. Hence, the definition of H 1 α is independent of the choice of Sobolev constants k, q. Proof of Lemma 3.10. We first show that, when F α L q Σ is sufficiently small, we have a direct sum decomposition W k,q T Σ Pg = H 1 α im d α k,q im d α k,q. We prove this in the case k = 0; the case k > 0 is similar but slightly easier. By definition of H 1 α, it suffices to show that the images of d α and d α intersect trivially. Towards this end, write d α f = d α g for 0-forms f, g of Sobolev class L q = W 0,q. Acting by d α and then d α gives [F α, f ] = d α d α g, [F α, g] = d α d α f. A priori, d α d α g and d α d α f are only of Sobolev class W 2,q, however, the left-hand side of each of these equations is in L r, where 1/r = 1/q + 1/p. So elliptic regularity implies that f and g are each W 2,q. This bootstrapping can be continued to show 21

that f, g are smooth, but we will see in a minute that they are both zero. By Lemma 3.9 and the embedding W 2,q L, it follows that, whenever F α L q is sufficiently small, we have f L C d α d α f L q = C [F α, g] L q 2C F α L q g L. Similarly, g L 2C F α L q g L, and hence f L 4C 2 F α 2 L q f L. If F α 2 L q < 2C 2, then this can happen only if f = g = 0. This establishes the direct sum 31. Now we prove that the dimension of Hα 1 is finite and equals that of Hα 1 for any flat connection α. It is well-known that the operator d α d α : W k+1,q Pg W k+1,q Pg W k,q T Σ Pg is elliptic, and hence Fredholm, whenever α is flat. The irreducibility condition implies that it has trivial kernel, and so has index given by dimhα 1, which is a constant independent of α. Then for any other connection α, the operator d α d α differs from d α d α by a compact operator, and so d α d α is Fredholm with the same index dimhα 1 [23, Theorem A.1.5]. It follows from Lemma 3.9 that the bounded operator d α d α : W k+1,q Pg W k+1,q Pg W k,q T Σ Pg is injective whenever F α L q Σ is sufficiently small, and hence the cokernel has finite dimension dimhα 1. That is, dimhα 1 = dimhα 1, so this finishes the proof of Lemma 3.10. Next we show that the L 2 -orthogonal projection to H 1 α = ker d α ker d α depends smoothly on α in the L q -topology. Proposition 3.12. Suppose that P Σ and ɛ 0 > 0 are as in Lemma 3.10, and let 1 < q <. Then the assignment α proj α is affine-linear and bounded proj α proj α op,l q C α α L q Σ, 33 provided F α L q, F α L q < ɛ 0, where op,l q is the operator norm on the space of linear maps L q T Σ Pg L q T Σ Pg. Proof. We will see that defining equations for proj α are affine linear, and so the statement will follow from the implicit function theorem in the affine-linear setting. First, we introduce the following shorthand: W k,q Ω j := W k,q Λ j T Σ Pg, L q Ω j := W 0,q Ω j. Next, we note that for µ L q Ω 1, the L 2 -orthogonal projection proj α µ is uniquely characterized by the following properties: 22

Property A: u, v W 1,q Ω 0 W 1,q Ω 2, µ proj α µ = d α u + d αv, Property B: a, b W 1,q Ω 0 W 1,q Ω 2, proj α µ, d α a + d αb = 0, where q is the Sobolev dual to q: 1/q + 1/q = 1. Here and below, we use the notation µ, ν to denote the L 2 -pairing on forms. Note that by Lemma 3.9 the operators d α and d α are injective on 0- and 2-forms, respectively, so any pair u, v satisfying Property A is unique. Consider the map A 0,q L q Ω 1 L q Ω 1 W 1,q Ω 0 W 1,q Ω 0 W 1,q Ω 0 W 1,q Ω 2 L q Ω 1 34 defined by α, µ; ν, u, v ν, d α, ν, d α, µ ν d α u d αv The key point is that a tuple α, µ; ν, u, v maps to zero under 34 if and only if this tuple satisfies Properties A and B above. By the identification W k,q = W k,q, we can equivalently view 34 as a map A 0,q L q Ω 1 L q Ω 1 W 1,q Ω 0 W 1,q Ω 0 W 1,q Ω 0 W 1,q Ω 2 L q Ω 1 35 defined by α, µ; ν, u, v d αν, d α ν, µ ν d α u d αv. Claim 1: The map 35 is bounded affine linear in the A 0,q -variable, and bounded linear in the other 4 variables. Claim 2: The linearization at α, 0; 0, 0, 0 of 35 in the last 3-variables is a Banach space isomorphism, provided α α L q is sufficiently small for some flat connection α. Before proving the claims, we describe how they prove the lemma. Observe that α, 0; 0, 0, 0 is clearly a zero of 35 for any α. Claim 1 implies that 35 is smooth, and so by Claim 2 we can use the implicit function theorem to show that, for each pair α, µ A 0,q L q Ω 1, with α α L q sufficiently small, there is a unique ν, u, v L q Ω 1 W 1,q Ω 0 W 1,q Ω 0 such that α, µ; ν, u, v is a zero of 35. A priori this only holds for µ in a small neighborhood of the origin, but since 35 is linear in that variable, it extends to all µ. It will then follow that ν = proj α µ depends smoothly on α in the L q -metric. In fact, 35 is affine linear in α and linear in the other variables, so the uniqueness assertion of the implicit function theorem implies 23