FLAT SURFACES AND STABILITY STRUCTURES

by F. HAIDEN, L. KATZARKOV, and M. KONTSEVICH

CONTENTS
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247
1.1. Spaces of stability structures and flat surfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247
1.2. Fukaya categories of surfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 248
1.3. Finite vs. infinite area, finite length property . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249
1.4. Meromorphic connections on T∗ C and Nevanlinna’s Kernpolygon . . . . . . . . . . . . . . . . . . . . . . . 249
1.5. Comparison with the work of Bridgeland–Smith . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 250
2. Moduli spaces of flat surfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251
2.1. Grading of surfaces and curves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251
2.2. Flat surfaces with conical singularities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253
2.3. Saddle connections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 256
2.4. Horizontal strip decompositions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259
2.5. Complex-analytic point of view . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 260
2.6. Voronoi and Delaunay partitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267
3. The topological Fukaya category of a surface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 268
3.1. Preliminaries on A∞ -categories . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 268
3.2. Surfaces with marked boundary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271
3.3. Minimal A∞ -category of an arc system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 272
3.4. Formal generators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 276
3.5. Localization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 278
3.6. Cosheaf of categories . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 280
4. Tameness and geometricity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283
4.1. Twisted complexes from curves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283
4.2. Representations of nets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 286
4.3. Minimal twisted complexes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 293
4.4. Classification of objects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 294
5. Comparison of moduli spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 297
5.1. Stability structures on categories . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 297
5.2. Charge lattice . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 298
5.3. Statement of main theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 301
5.4. Support property . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 302
5.5. Harder–Narasimhan filtrations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 303
5.6. The map M(X) → Stab(F (X)) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 306
6. Cluster-like structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 308
6.1. S-graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 308
6.2. Example: Dynkin type An . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311
6.3. Example: affine type A n . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 313
7. Open problems and further directions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 315
Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 315
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 316

1. Introduction

1.1. Spaces of stability structures and flat surfaces

The space of stability structures, Stab(C ), of a triangulated category C was in-
troduced in seminal work of Bridgeland [11] based on a proposal of M.R. Douglas on
D-branes in super-conformal field theories. The data of a stability structure, σ , is a distin-
guished class of objects in C , the semistable ones, each with a phase φ ∈ R, and an additive

DOI 10.1007/s10240-017-0095-y
248 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH

map Z : K0 (C ) → C, the central charge. These have to satisfy a number of axioms, see
Section 5.1. Although the definition seems somewhat strange at first from a mathemat-
ical point of view, it leads to the remarkable fact, proven by Bridgeland, that Stab(C )
naturally has the structure of a complex manifold, possibly infinite-dimensional. Further-
more, Stab(C ) comes with an action of Aut(C ) and the universal cover of GL+ (2, R).
Thanks to the efforts of a number of people, the structure of Stab(C ) is understood in
many particular cases, see e.g. [7, 12, 13, 15, 29, 34].
Moduli spaces of very similar nature appear in the theory of complex curves
with a quadratic differential, also known as half-translation surfaces or simply flat surfaces,
as they carry a flat Riemannian metric with conical singularities. Each of these moduli
spaces, M(S), like Stab(C ), comes with a wall-and-chamber structure and an action of
GL+ (2, R). An explanation for this similarity was proposed by the third named author
and Soibelman in [31]: The spaces M(S) should be realized as moduli spaces of stability
structures for a suitably defined Fukaya category of the surface S. The main achievement
of the present paper is to make precise and verify this claim, leading to the following
result (Theorem 5.3 in the main text).

Theorem. — Let S be a marked surface of finite type, M(S) the space of marked flat structures
on S, and F (S) the Fukaya category of S. The there is a natural map
 
(1.1) M(S) → Stab F (S)
which is bianalytic onto its image, which is a union of connected components.

For now it remains an open question whether the image is in fact all of Stab(F (S)).
The method of proof is new and relies on a detailed understanding of the category F (S).

1.2. Fukaya categories of surfaces

In Section 3 we give a self-contained elementary definition of the Fukaya
A∞ -category F (S) of a surface, and provide arguments to show that it agrees with other
definitions found in the literature. Broadly speaking, these are either based on Floer in-
tersection theory (see [22], [44], and many others), sheaves on ribbon graphs [21, 36,
45], or more direct approaches [1]. Our own approach to define F (S) falls into the latter
category and is based on explicit A∞ -structures and contractibility of the classifying space
of arc systems on a marked surface.
The most significant result about F (S) which we prove is a complete classification
of objects (Theorem 4.3 in the main text).

Theorem. — Isomorphism classes of objects in F (S) are classified by isotopy classes of certain
immersed curves on S together with local system of vector spaces on them. In particular, these categories
are of tame representation type.
 FLAT SURFACES AND STABILITY STRUCTURES 249

Another way to say this is that all objects are “geometric”. This is not just a direct
consequence of the definition, since F (S) is required to be closed under taking cones.
We reduce the problem to the classification of certain pairs of filtrations on vector spaces
with subquotients identified.
An important tool, e.g. for proving the equivalence with other definitions and com-
puting K-theory, are formal generators of F (S) with endomorphism algebras given by
graded quivers with quadratic monomial relations. These exist only for some S, but the
other cases are obtained by localization of F (S). They are graded versions of the gentle
algebras introduced by Assem–Skowroński [3].

1.3. Finite vs. infinite area, finite length property

The flat surfaces of interest in ergodic theory are the ones with finite area, cor-
responding to quadratic differentials with arbitrary order zeros and simple poles. For a
generic flat surface with infinite area, on the other hand, the structure of the horizontal
foliation is very explicitly described by certain generalized ribbon graphs (see Section 6),
and the horizontal measure is not ergodic. In fact, the wall-and-chamber structure on
the space of flat surfaces of infinite area is described by mutations similar to those in the
theory of cluster algebras.
Another central tool in the study of flat surfaces of infinite area, apart from the hor-
izontal foliation, is the core, Core(S), which is defined to be the convex hull of the conical
points. It is a compact subset of S with geodesic boundary, and in fact S is completely de-
termined by its core. Changes to the structure of the core are related to wall-crossings of
the first kind, (central charges becoming colinear over R) while mutations are associated
with wall-crossings of the second kind (central charges becoming real).
In terms of the associated stability structures, this dichotomy manifests itself in the
finite length property of the t-structure, i.e. whether every descending chain of objects
is eventually constant. Any stability structure on a category C determines a family of t-
structures. If the surface has infinite area, the generic t-structure in this family is finite
length with a finite number of isomorphism classes of simple objects. If the surface has
finite area on the other hand, the t-structures are not finite length. Closely related is the
question of the distribution of phases of stable objects, in particular density, which was
investigated by Dimitrov and the authors in [18].

1.4. Meromorphic connections on T∗ C and Nevanlinna’s Kernpolygon

For a meromorphic quadratic differential, ϕ, on a compact Riemann surface, C,
√
there is a corresponding connection D on T∗ C so that the scalar multiples of ± ϕ are
its flat sections. This connection has a simple pole at the points where ϕ has a zero or
pole of any order. Note also that the monodromy of D is necessarily contained in {±1}.
250 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH

Conversely, given a meromorphic connection D on T∗ C with monodromy {±1} and

simple poles, we get a quadratic differential, up to scalar multiple.
If the connection D has higher order poles then the associated quadratic differen-
tials are not meromorphic but locally of the form
 
(1.2) exp f (z) g(z)dz2

where f and g are meromorphic and f has a pole of order one greater than D. We
call this an exponential singularity of a quadratic differential. In terms of flat geometry, this
singularity gives rise to n infinite-angle singularities of the metric, where n is the order
of the pole of f . Such a singularity is modeled on the metric completion of the universal
cover of the punctured Euclidean plane, and can be regarded as a limit of the conical
singularity when the cone angle goes to infinity. We show the following:

Theorem. — Let S be a half-translation surface of finite type with conical singularities, possi-
bly infinite-angle ones, then S comes from a compact Riemann surface with quadratic differential with
exponential singularities.

See Theorem 2.3 in the main text. Thus allowing exponential singularities is quite
natural from a geometric point of view, and turns out to be necessary to get stability
structures on partially wrapped Fukaya categories of surfaces.
The previous theorem is similar in nature and has some overlap with an old result
of Nevanlinna [38] on the class of functions which have Schwarzian derivative a polyno-
mial. Indeed, ideas of his are found in our proof. The analysis of the topological structure
of the branched covering defined by these functions leads him to a certain combinatorial
structure, the Kernpolygon, which translates into the t-structure of the stability structure for
a quadratic differential exp(P(z))dz2 .

1.5. Comparison with the work of Bridgeland–Smith

A connection between moduli spaces of quadratic differentials and stability struc-
tures is also made in the work of Bridgeland–Smith [15]. They do not use the Fukaya
category of the surface directly, but a certain CY3 category which is defined in terms of
quivers with potential attached to ideal triangulations. These categories are connected
to cluster theory via their quivers, and were shown to be subcategories of Fukaya cate-
gories of certain non-compact 6-folds by Smith [46]. Being CY3, they have generalized
Donaldson–Thomas invariants in the sense of [31]. The DT invariants and their wall-
crossing had previously appeared in work of Gaiotto–Moore–Neitzke [23], who also in-
troduce the “WKB triangulation”, a central tool in [15]. Indeed, one of the motivations
of Bridgeland–Smith is to develop a mathematical theory for [23].
The methods of [15] require one to restrict to meromorphic quadratic differentials
without higher order zeros, and, at a generic point in the moduli space, at least one higher
 FLAT SURFACES AND STABILITY STRUCTURES 251

order pole. Some quadratic differentials with simple poles do appear in the limit. It is
expected that their results extend in some form to holomorphic quadratic differentials
with simple zeros, but the stability structures are not described by quivers with potential
and would require different methods to construct. In this paper we do indeed construct
stability structures from these quadratic differentials, though not on CY3 categories.
It is somewhat puzzling that the same geometry should give rise to stability struc-
tures on, at least superficially, very different categories. The relation between the two
categories (the wrapped Fukaya category of the surface and the CY3 category) is still to
be clarified. It seems to be simpler in the case when all poles have order ≥ 3. One can
show that there is a non-commutative divisor in the sense of Seidel [42] with total space
the wrapped category and central fiber the CY3 category. We suspect that in the case of
double poles an additional deformation is necessary.
More recently, an extension of the results of Bridgeland–Smith to certain poly-
nomial quadratic differentials on C was carried out in [14, 27]. There, CYn categories
appear for n ≥ 2.

2. Moduli spaces of flat surfaces

In this section we define the moduli spaces M(X) of marked flat structures. Sec-
tion 2.1 discusses some topological preliminaries which are mostly relevant later when
we define Fukaya categories of surfaces, but also play a small role in the definition of
M(X). Next, in Section 2.2 we define flat surfaces with conical singularities and their
moduli spaces and state a theorem about their local structure. This theorem is proven
in the case of finite area in Section 2.3, using triangulations by saddle connections. Sec-
tion 2.4 deals with the case of infinite area, where the main tool is the horizontal strip
decomposition. In Section 2.5 we show that our class of flat surfaces may be described in
terms of quadratic differentials on compact Riemann surfaces. Finally, in Section 2.6 we
review the dual Voronoi and Delaunay partitions.

2.1. Grading of surfaces and curves

The notion of grading discussed in this subsection is essentially a special case of
the one for symplectic manifolds with almost complex structure and their Lagrangian
submanifolds, due to the third named author and Seidel [30, 43]. In the case of surfaces,
no symplectic geometry is involved.
We use the notation 1 (X, x, y) for the set of homotopy classes of paths from x
to y in a space X. Let X be a smooth oriented surface. A grading, η, on X is a foliation,
i.e. section of the projectivized tangent bundle, P(TX). The pair (X, η) is a graded surface.
A morphism of graded surfaces (X, η) → (Y, θ ) is a pair (f , f˜ ) where f : X → Y is an
orientation preserving local diffeomorphism, and f˜ is a homotopy from f ∗ θ to η, up to
252 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH

homotopy, i.e. f˜ ∈ 1 ((X, P(TX)), f ∗ θ, η). Composition is given by

   
(2.1) (f , f˜ ) ◦ (g, g̃) = f ◦ g, g ∗ f˜ · g̃

where α · β denotes concatenation of paths. Every graded surface X has a shift automor-
phism given by the pair (idX , σ ) where σ restricts, for every x ∈ X, to the generator of
π1 (P(Tx X)) given by the orientation of X, i.e. the path which rotates a line counterclock-
wise by an angle of π . Anticipating the connection with triangulated categories, we write
this automorphism as [1] and its integer powers as [n].
For an oriented surface X observe that the set
  
(2.2) G := π0  X, P(TX)

is a torsor over H1 (X, Z) = [X, S1 ]. The set G classifies gradings on X up to graded

diffeomorphism with underlying map on X the identity. The set of gradings of X up to
isomorphism is thus a quotient of G .

Example. — Consider the cylinder X = R × S1 . Then

     
(2.3) π0  X, P(TX) ∼ = π0 Maps S1 , S1 ∼ = Z.

The above isomorphism simply takes a foliation η and counts how many times it rotates
along the loop {0} × S1 ⊂ X. Call this number deg(η), then the map (x, θ ) → (−x, θ )
from the cylinder to itself acts on the foliation by switching the sign of deg(η). From this
one can see that |deg(η)| ∈ Z≥0 is a complete invariant of the graded surface.

Next, we consider immersed curves in graded surfaces. A graded curve in a graded

surface (X, η) is a triple (I, c, c̃) where I is a 1-manifold, possibly disconnected and with
boundary, c : I → S is an immersion, and c̃ ∈ 1 ((I, c∗ P(TX)), c∗ η, ċ). To spell this out,
c̃ is given by a homotopy class of paths in P(Tc(t) X) from the subspace given by the grad-
ing to the tangent space of the curve, varying continuously with t ∈ I. The pushforward
of a graded curve as above by a graded morphism (f , f˜ ) is given by (I, f ◦ c, (c∗ f˜ ) · c̃).
Note that graded curves are not naturally oriented.
A point of transverse intersection of a pair (I1 , c1 , c̃1 ), (I2 , c2 , c̃2 ) of graded curves
determines an integer as follows. Suppose ti ∈ Ii with

(2.4) c1 (t1 ) = c2 (t2 ) = p ∈ X, ċ1 (t1 ) = ċ2 (t2 ) ∈ P(Tp X).

We have the following homotopy classes of paths in P(Tp X): (1) c̃1 (t1 ) from η(p) to ċ1 (t1 ),
(2) c̃2 (t2 ) from η(p) to ċ2 (t2 ), (3) κ from ċ1 (t1 ) to ċ2 (t2 ) given by counterclockwise rotation
in Tp X by an angle < π . Define the intersection index of c1 , c2 at p
 
(2.5) ip (c1 , c2 ) = c̃1 (t1 ) · κ · c̃2 (t2 )−1 ∈ π1 P(Tp X) ∼= Z.
 FLAT SURFACES AND STABILITY STRUCTURES 253

FIG. 1. — Smoothing an intersection with ip (c1 , c2 ) = 1

More precisely, this notation is correct only if c1 , c2 pass through p exactly once, in par-
ticular if ci are in general position, otherwise the index may depend on the ti as well. We
note that

(2.6) ip (c1 , c2 ) + ip (c2 , c1 ) = 1

and
 
(2.7) ip c1 [m], c2 [n] = ip (c1 , c2 ) + m − n.

Another observation which will be used repeatedly, is that if graded curves c1 and c2
intersect in p with ip (c1 , c2 ) = 1, then we may perform a kind of smoothing near p which
produces again a graded curve (see Figure 1). (If the intersection index is not 1 then the
grading does not extend continuously along the modified curves.)
A graded surface (X, η) has a canonical double cover τ of X with fiber over p ∈ X
the set of orientations of η(p). We will consider τ as a locally constant sheaf and write

(2.8) Zτ := Z ⊗Z/2 τ

and similarly for Rτ and Cτ . Singular (co)homology with coefficients in any of these local
systems is defined, as is deRham cohomology with coefficients in Rτ or Cτ . Integration
provides a pairing

(2.9) H1 (X, ∂X; Zτ ) ⊗Z H1dR (X, ∂X; Cτ ) → C

where H∗dR (M, ∂M; E) is the cohomology of forms with coefficients in E which vanish
when pulled back to ∂M.

2.2. Flat surfaces with conical singularities

The following types of structures on an orientable surface X are in one-to-one
correspondence:
1. Flat Riemannian metrics together with a covariantly constant foliation η and
orientation.
254 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH

2. Complex structures together with a non-vanishing holomorphic quadratic dif-

ferential (section of the square of the holomorphic cotangent bundle).
3. Pairs of pointwise linearly independent commuting vector fields, defined up to
simultaneous change of sign, i.e. a section of the quotient of the frame bundle
of TX by Z/2.
On R2 the standard versions of these structures are: (1) the standard Euclidean metric
with the foliation by horizontal lines, (2) the standard complex structure R2 = C and
the quadratic differential dz2 , (3) the pair of vector fields (∂x , ∂y ). A general surface with
one of the structures above is locally isomorphic to standard R2 by charts obtained from:
(1) the exponential map of the metric, (2) the complex path integral of a square-root of
the quadratic differential, (3) the flows of the vector fields. The transition functions for
such charts are of the form
(2.10) v → ±v + c.
Indeed, these are the symmetries of R2 with any of the standard structures above. A smooth
flat surface is a surface with any of the above structures. The covariantly constant foliation
makes it a graded surface and is called the horizontal foliation. Another more precise but
longer term used for this kind of structure is half-translation surface, emphasizing the fact
that the holonomy of the metric must be a subgroup of {±1}.
The metric completion of a smooth flat surface may have fairly complicated sin-
gularities in general [10]. Here, we will only consider singularities which are conical with
cone angle a positive integer multiple of π or infinite. They may be described as follows.
For n ∈ Z let Cn be the space
(2.11) Cn := (Z/n) × R × R≥0 / ∼, (k, x, 0) ∼ (k + 1, −x, 0) for x ≤ 0
and
(2.12) C∞ := Z × R × R≥0 / ∼, (k, x, 0) ∼ (k + 1, −x, 0) for x ≤ 0.
Let C be one of the above spaces and 0 ∈ C the distinguished point (0, 0, 0) ∼ (k, 0, 0).
The standard flat structure on R2 gives C \ 0 the structure of a smooth flat surface and
C is its metric completion. The metric is singular unless C = C2 . The singularity is called
a conical singularity with cone angle nπ if C = Cn and an infinite angle singularity if C = C∞ .
See Figure 2 for a picture of C3 .
It is technically convenient to work not with the above spaces directly, but instead
their real blow-up at the origin, denoted  C. This construction produces a surface with
boundary in which the cone point 0 ∈ C is replaced by the set of geodesics starting at that
point, which is topologically S1 in the case of Cn and R in the case of C∞ . The complex
structure on C \ 0 does not extend to  C, but the horizontal foliation does. Also, locally,
any square-root of the quadratic differential extends as a closed complex-valued 1-form
to 
C which vanishes on vectors tangent to the boundary. Indeed, if we identify C  ∞ with
R≥0 × R the 1-form is given by e (dx + ixdy).
iy
 FLAT SURFACES AND STABILITY STRUCTURES 255

FIG. 2. — Conical singularity with cone angle 3π (“monkey saddle”)

Definition. — Let Y be a smooth surface with boundary. The structure of a real blow-up
of a flat surface on Y is given by the structure of a smooth flat surface on Y \ ∂Y such that each
component of ∂Y has a neighborhood diffeomorphic to a neighborhood of the boundary of some C n or C
 ∞
by a diffeomorphism preserving the flat structure on the interior, and satisfying the following completeness
condition: The space X obtained by identifying points lying in the same component of ∂Y is the metric
completion of Y \ ∂Y.

Definition. — A flat surface is a complete metric space together with a partition X = Xsm ∪ Xsg
into smooth and singular points and the structure of a smooth flat surface on Xsm (i.e. direction of
horizontal foliation) such that near each x ∈ Xsg the space is locally isometric to one of Cn or C∞ .

Note that in the above definition x ∈ Xsg is not necessarily a singular point of the
metric if near x the model is C2 . One can consider it as an arbitrary marked point in the
metrically smooth part. As the terminology suggests, the real blow-up of a flat surface has
the structure of a real blow-up of a flat surface. Conversely, one obtains a flat surface from the
real blow-up by taking the metric completion of the interior. (Topologically, this contracts
each boundary component to a point.)

Definition. — Fixing a graded surface S we consider the moduli space of marked flat sur-
faces M(S). An element of M(S) is represented by a real blow-up of a flat surface Y with underlying
oriented surface S together with a morphism of graded surfaces f : S → Y with underlying map the
identity on S. Two such pairs (Y, f ), (Z, g) are equivalent if there is a diffeomorphism h : Y → Z
preserving the flat structure so that h ◦ f and g are isotopic as maps of graded surfaces. We give flat
structures (before identification) the topology of pointwise convergence on the real blow-up, and M(S) the
quotient topology.

Note that the grading on S fixes the number and types of singular points and
marked points of any flat surfaces in M(S). Explicit descriptions of M(S) when S is a
disk or annulus are found in Section 6.
If Y is the real blow-up of a flat surface then the square-roots of the quadratic
differential on Y \ ∂Y may be viewed as a closed 1-form α with values in Cτ , and as such
it extends to the boundary. In particular we get a class [α] ∈ H1dR (Y, ∂Y; Cτ ) so using the
pairing (2.9) we get a homomorphism H1 (Y, ∂Y; Zτ ) → C which represents the periods
256 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH

of the flat surface. We say that a graded surface Y is of finite type if Y and ∂Y have a finite
number of connected components and H1 (Y, ∂Y; Zτ ) is finitely generated.

Theorem 2.1. — Let S be a graded surface of finite type. The map

 
(2.13)  : M(S) → Hom H1 (S, ∂S; Zτ ), C ,
assigning to each flat surface its periods, is a local homeomorphism.

The strategy of the proof will depend on whether the area of the flat surface is
finite or infinite and will be completed in the following two subsections.

Grading of curves on flat surfaces

Let X be a surface with flat structure defined in terms of a complex structure and
non-vanishing holomorphic quadratic differential ϕ ∈ (X, (T∗ X)⊗2 ). In this context,
grading of curves and intersection index can be defined somewhat more concretely. Re-
call that the holomorphic and real tangent bundles may be identified, and the horizontal
foliation is the unique section of P(TX) on which ϕ is real and positive. Moreover, we
have an identification of P(TX) with the bundle with constant fiber C∗ /R>0 ∼ = U(1)
under which this section corresponds to 1 ∈ U(1).
Given an immersed curve α : I → X a grading is defined by a function φ : I → R
such that
 
(2.14) ϕ α̇(t), α̇(t) ∈ R>0 e2πiφ(t) .
The corresponding path α ∗   α̇ is given by
(2.15) s → e2πiφ(t)s , s ∈ [0, 1]
under the identification as above. Note that α is geodesic if and only if φ is locally con-
stant. Suppose we have curves α1 , α2 with grading given by real-valued functions φ1 , φ2
intersecting transversely in p ∈ S. Then the intersection index is
 
(2.16) ip (α1 , α2 ) = φ1 (p) − φ2 (p) .

2.3. Saddle connections

Geodesics, in particular those of finite length, are fundamental in the study of flat
surfaces. The smooth part, Xsm , of a flat surface X is a Riemannian surface, and by our
completeness requirement any geodesic on Xsm can either be extended indefinitely or
converges towards a point in the complement Xsg of Xsm . A saddle connection is a geodesic
converging towards points in Xsg (“saddles”) in both directions. The endpoints are not
required to be distinct.
A geodesic arc system on a flat surface is a collection of saddle connections which
intersect at most in the endpoints.
 FLAT SURFACES AND STABILITY STRUCTURES 257

Proposition 2.1. — Let X be a flat surface of finite type, then the size of any geodesic arc system
is bounded above by a constant depending only on the topology of the real blow-up Y = X. 

Proof. — Let A be a finite geodesic arc system on X and E = |A| its cardinality.
Each saddle connection in A together with an arbitrary choice of orientation gives a non-
zero class in H1 (X, Xsg , Z). A complete set of relations between these classes is given by
the polygons cut out by A. Let F be the number of these polygons. Any bi-gon cut out
by A must contain a conical point with cone angle π in its interior, so their number, D,
is bounded by the number of compact boundary components of Y. Assigning boundary
edges to polygons we find that

2
(2.17) F − D ≤ E.
3
By assumption, B = rkH1 (X, Xsg , Z) is finite, and since E − F ≤ B we get

(2.18) E ≤ 3(D + B)

which completes the proof. 

Let A be a maximal geodesic arc system on X. By the previous proposition it is

finite. Let us study its local geometry. If p ∈ Xsg and D is the corresponding boundary
component of X,  then the set of maximal geodesics starting at p is identified with D.
A finite subset E ⊂ D corresponds to (ends of) saddle connections in A which converge
to p. For components of D \ E there are two possibilities: Either the maximal geodesics
in these directions intersect a saddle connection in A, or not. They cannot converge to
a singularity by maximality of A. In the first case there is a triangle cut out by A with p
as one of its vertices. In the second case there is an conical sector with cone angle ≥ π ,
possibly infinite, again by maximality.

Proposition 2.2. — Suppose A is a maximal geodesic arc system on a flat surfaces X. Let K be
the union of Xsg , the saddle connections in A, and the triangles cut out by A. Then K is the convex hull
of Xsg , thus independent of A.

Proof. — From the local geometry above we see that any geodesic which leaves K
stays in X \ K for all subsequent times, so K is convex. It contains Xsg by definition. If K
is convex and contains Xsg , then it must also contain all saddle connections, in particular
those in A, and thus all triangles cut out by A, so K ⊆ K . 

We will refer to the convex hull of Xsg as the core of the flat surface X, denoted
Core(X). The previous proposition shows that, for X of finite type, it is covered by a finite
number of singular points, saddle connections, and geodesic triangles, thus compact.
258 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH

Proposition 2.3. — Let X be a flat surface with the property that each component contains at
least one singular point, and let K be the core of X. Then K is a deformation retract of X.

Proof. — The assumption ensures that each p ∈ X has finite distance to Core(X).
A deformation retraction can be defined by moving p ∈ X along a geodesic path to the
unique point in Core(X) which is closest to p. 

Let X be a connected flat surface with finite area and assume that X is not smooth,
which excludes only the case of the torus. A maximal geodesic arc system on X gives a tri-
angulation with some triangles possibly degenerate. By definition, a triangle is degenerate
if two of its edges coincide, so that they meet in a conical singularity with cone angle π .
Using triangulations by saddle connections, we can now give a proof of Theorem 2.1 in
the case when the flat surface has finite area. This result is classical, see e.g. Veech [50],
but we include a proof here for completeness. The proof in the case of infinite area (but
still finite type) will be given in the next section.

Proof of Theorem 2.1, finite area case. — Fix a graded surface X of finite type admitting
flat structures with finite total area. This means that X is compact as a surface with
boundary. We also assume that there is at least one singularity, excluding the case of
the torus which is easily handled directly. Let  = H1 (X, ∂X; Zτ ) and  : M(X) →
Hom(, C) be the period map. We want to show that  is a local homeomorphism.
Let a0 ∈ M(X) and equip X with this flat structure. Pick a maximal collection
A of geodesic arcs on X. These arcs give a triangulation of X, since the flat metric has
finite area by assumption. Fix a grading on each arc in A, then each α ∈ A gives a class
[α] ∈ . A complete set of relations between these classes comes from the list of triangles.
The number (a0 )([α]) ∈ C records the length and slope of an arc α ∈ A.
Since the number of triangles is finite, there is a contractible neighborhood U of
(a0 ) ∈ Hom(, C) with the following property: For any b ∈ U and any three classes
α1 , α2 , α3 in  corresponding to the edges of a triangle of A, the vectors (b)([αi ]) form
the edges of a non-degenerate triangle in C. This ensures that for each b ∈ U we get
an actual flat structure on X essentially by direct construction. Namely, we cut X into its
geodesic triangles, deform each of the triangles according to the values (b)([α]), α ∈ A,
then glue the triangles back together in the same combinatorial manner. In this way, we
have defined a continuous section  of  over U.
To complete the proof, it remains to be shown that (U) is an open neighborhood
of a0 . Note that there is a neighborhood W of a0 ∈ M(X) such that all saddle connections
in A persist throughout W, i.e. do not break along a singularity. Making W sufficiently
small, we can assume that W ⊂ −1 (U). Then the same collection of arcs A, up to
homotopy, again form a maximal collection of geodesic arcs for any flat surface in W.
This shows that W ⊂ (U), so (U) is a neighborhood of a0 , in fact open by the same
arguments. 
 FLAT SURFACES AND STABILITY STRUCTURES 259

2.4. Horizontal strip decompositions

Given a complex number z ∈ C with Im(z) > 0 the corresponding horizontal strip of
finite height is the flat surface with boundary
(2.19) w ∈ C | 0 ≤ Im(w) ≤ Im(z)
with marked points {0, z} on the boundary. The horizontal strip of infinite height is just the
closure of the upper half-plane with the origin as marked point. It turns out that a generic
surface of finite type with infinite area is obtained, in a canonical way, by gluing these
types of pieces along their boundaries.

Proposition 2.4. — Let X be flat surface of finite type with infinite area. Assume that X is
connected and has at least one conical point. Then, after possibly rotating the horizontal direction, the
leaves converging towards a conical point cut the surface into horizontal strips as above.

Proof. — Assume that no horizontal leaf is a saddle connection. This can always
be achieved by rotating the horizontal direction, as there are at most countably many
slopes of saddle connections. In fact, the set of slopes of saddle connections is closed
under the present assumptions, but this is not needed. Note that no leaf can be a closed
geodesic either, as a cylinder foliated by closed geodesics is bounded by one or more
saddle connections. Arguments of Strebel [47] show that there are then only two types of
leaves. Generic leaves which are closed and intersect the core of X in a compact interval, and
Critical leaves which converge towards a conical point in one direction and eventually leave
the core in the other. Indeed, the boundary of the closure of a leaf remaining entirely in
the core would be a union of leaves which are saddle connections, which is impossible. In
particular, the closure of a critical leaf adds only a single conical point. The components
of the union of all generic leaves are open parts of horizontal strips. The exponential maps
at the various conical points give identifications of the closures of these open components
with the standard horizontal strips of finite or infinite height. 

Let us say a bit more about the horizontal strip decompositions appearing in the
previous proposition. A conical point with finite cone angle lies on finitely many hori-
zontal strips, while a conical point with infinite cone angle lies on infinitely many, all but
finitely many of which will have infinite height. Each horizontal strip of finite height con-
tains a unique saddle connection: the straight line segment connecting 0 and z. These
are exactly the saddle connections α1 , . . . , αm , which do not intersect any of the critical
leaves. We will refer to them as simple saddle connections. Following the leaves and collapsing
the horizontal strips of infinite height we construct a deformation retraction of the sur-
face to the union of the αi and the conical points. This shows that the αi give a basis of
H1 (X, ∂X; Zτ ), after choosing gradings.
The graph formed by the αi is identified with part of the Hausdorff version of the
leaf space, L, of the horizontal foliation. We may realize L inside X by adding the positive
260 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH

imaginary axis in each horizontal strip of infinite height. Combinatorially, L is a bit of a

generalization of a ribbon graph with some vertices connected to infinitely many half
edges (those corresponding to infinite-angle singularities). More precisely, the set of half
edges meeting a given vertex is a torsor over some cyclic group. This will be discussed in
more detail in Section 6.1.

Proof of Theorem 2.1, infinite area case. — Let a0 ∈ M(X) be a flat structure and assume
first that there are no leaves which are saddle connections. By the considerations above,
the critical leaves cut X into horizontal strips. Let α1 , . . . , αm be the saddle connections
corresponding to the horizontal strips of finite height. Also choose the grading on each
αi so that Im (a0 )(αi ) > 0. This is possible since αi is not horizontal and shifting the
grading flips the sign of (a0 )(αi ). We get a basis [α1 ], . . . , [αm ] of  = H1 (X, ∂X; Zτ ).
Consider the open subset U ⊂ Hom(, C) of maps sending each [αi ] to the upper half
plane. For each b ∈ U we construct a flat surface (b) with the same combinatorial type
of horizontal strip decomposition as a0 , but different slopes and lengths of the αi given
by b. The map  defines a section of M(X) → Hom(, C) over U. To see that (U) is
open, note that any flat surface sufficiently close to one in (U) has the same horizontal
strip decomposition.
If a0 has horizontal saddle connections, we argue as follows. By assumption  = 0,
so R ⊂ GL + (2, R) acts on M(X) by rotation of the horizontal direction and each orbit

contains a flat structure with only critical and generic leaves considered before. As 
is equivariant with respect to the action it must be a local homeomorphism near a0 as
well. 

2.5. Complex-analytic point of view

Meromorphic quadratic differentials
Let C be a compact Riemann surface and ϕ a non-zero meromorphic quadratic
differential on C, i.e. a meromorphic section of the square of the canonical bundle. If
D ⊂ C is the set of zeros and poles of ϕ, then Xsm = C \ D has the structure of a smooth
flat surface, possibly incomplete. The flat geometry near a zero or pole is analyzed by
Strebel [47] by finding a local holomorphic coordinate in which the quadratic differential
has a particularly simple form. Let us recall the results.
Zeros and simple poles. There is a local coordinate z such that ϕ = zn dz2 , n =
−1, 1, 2, 3, . . .. The metric has a conical singularity at z = 0 with cone angle (n + 2)π .
In particular, if ϕ has no higher order poles, then the corresponding flat surface has finite
total area.
Poles of order n ≥ 2. If n is odd then there is a local coordinate z with ϕ = z−n dz2 .
However, if n is even then ϕ admits a square root near z = 0, and the residue a ∈ C of
√
ϕ is an invariant up to sign. It turns out to be the only local invariant, so that there is a
 FLAT SURFACES AND STABILITY STRUCTURES 261

coordinate z with
√ a
(2.20) ϕ = z−n/2 + dz
z
√
when n ≥ 4 and ϕ = za dz when n = 2. In any case, we note that the flat metric near a
higher order pole has infinite area and is complete, i.e. higher order poles do not lead to
any additional singularities of the metric.

Theorem 2.2. — Let C be a compact Riemann surface with non-zero meromorphic quadratic
differential ϕ with set of zeros and poles D. Then |ϕ| gives C \ D the structure of the smooth part of a
flat surface of finite type with conical points which have cone angle an integer multiple of π . Conversely,
any flat surface of finite type without infinite-angle conical points is obtained in such a way.

Remark. — Let Tg,n be the Teichmüller space of compact Riemann surfaces of genus
g with n marked points. The cotangent space to Tg,n at C is naturally identified with the
space of quadratic differentials on C with at most simple poles at the marked points.
Thus every point in the cotangent bundle of Tg,n corresponds to some flat surface of
genus g with at most n conical points with cone angle π (and other conical points with
bigger cone angle). The moduli spaces M(S) defined above correspond to strata in the
cotangent bundle. The ribbon graph attached to a Strebel differential is a special case of
the core of a flat surface introduced here.

Proof. — It remains to show the second part, so let X be a flat surface of finite
type without infinite-angle conical points. First of all, we have a complex structure and
non-vanishing holomorphic quadratic differential on the smooth part Xsm . We know that
near a conical point the flat structure comes from a zero or first order pole of a quadratic
differential, so the complex structure extends to all of X and ϕ extends meromorphically.
If X has finite area, then it is compact and so C = X, D = Xsg , and we are done.
For the case of infinite area we need to find a compact K ⊂ X such that the com-
ponents of the complement X \ K are isometric to some punctured neighborhoods of
higher order poles. As a first step, take K to be the core of X. The boundary of K in X is
a sequence of saddle connections meeting in exterior angles φp ≥ π , by convexity of K.
The condition on the global monodromy of the metric (existence of horizontal foliation)
implies that

(2.21) (φp − π ) = nπ
p

where p runs over the corners of a fixed component of the boundary of K, and n ∈ Z≥0
depends on that component.
Points in the complement of K are parametrized by pairs (l, d) where l is a straight
line starting at some point in the boundary of K, meeting K only in the one endpoint,
262 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH

and meeting ∂K in angles ≥ π , together with a positive distance d ∈ R>0 from K. In

particular, each component of X \ K is topologically a cylinder.
We enlarge K so that its boundary is a sequence of closed geodesics and straight
edges (finite length geodesics) meeting in exterior angles 3π/2, not necessarily at conical
points. Consider some component B of the boundary of K. If B is a closed geodesic, then
the corresponding component of X \ K is identified with a flat cylinder which comes from
a second order pole of a meromorphic quadratic differential. Otherwise, B is a union of
an even number of edges with lengths a1 , . . . , a2n in counterclockwise order. If we parallel
transport some edge outwards from K by a distance d ≥ 0 then the lengths of the adjacent
edges increase by d. A simple proof by induction on n shows that when n is odd we can
parallel transport the edges outwards so that they all have the same length l, which can be
an arbitrary real number bigger than some constant depending on the original K. Hence,
in this case the local geometry is completely determined by n, and so must coincide with
the one coming from a pole of order n + 2.
When n is even we can arrange the sequence of lengths to be of the form
l1 , l2 , l3 , l3 , . . . , l3 after parallel transport of the edges, and add any positive real number
to all elements at once. Without loss of generality, the first edge (of length a) is horizontal,
then the integral of the square root of the quadratic differential around the cylinder is just
a = ±(l1 − l2 + i(l1 − l3 )), so the geometry must be locally isometric to the one coming
from the square of the abelian differential (z−n−2 + a/z)dz. 

Exponential-type singularities
An exponential-type singularity is a transcendental singularity of a quadratic differential
of the form

(2.22) ef (z) g(z)dz2 , f , g meromorphic

in some local coordinate z, where f has a pole of order n at z = 0. By change of coordi-

nates we may assume that f (z) = z−n . The following proposition describes the local flat
geometry near such a singularity.

Proposition 2.5. — Let

 
(2.23) ϕ = exp z−n zm g(z)dz2

with n ∈ Z>0 , m ∈ Z, g holomorphic, non-vanishing on a neighborhood U of 0 ∈ C. Set

(2.24) D = {0 < |z| ≤ r} ⊂ U

then the completion of D with respect to |ϕ| has n additional points, all of which are ∞-angle singular-
ities.
 FLAT SURFACES AND STABILITY STRUCTURES 263

Lemma 2.1. — Let ρ ≥ 1, π/2 < ε < π , A = {reiφ | r ≥ ρ, |φ| ≤ ε} ⊂ C, h(z) a

smooth function on A such that for some C1 , C2 > 0, λ ∈ R

(2.25) C1 |z|λ ≤ h(z) ≤ C2 |z|λ

Then the completion of A with respect to the metric

 −Re(z) 2
(2.26) e h(z) |dz|2

has a single additional point.

Proof. — 1. Define η(z) = (max{1, |Im(z)|})λ e−Re(z) , then for sufficiently small
ε > 0, the boundary of {η ≥ ε} in A is given by the curve
 
(2.27) Re(z) = λ log max 1, |Im(z)| − log(ε).

Moreover, for distinct values of ε, these curves have positive distance with respect to the
Euclidean metric.
2. We claim that e−Re(z) h(z) is bounded below on {η ≥ ε} for any ε > 0. By assump-
tion,

(2.28) e−Re(z) h(z) ≥ C1 |z|λ e−Re(z) .

Case λ ≥ 0: For |Im(z)| ≥ 1 we have

(2.29) C1 |z|λ e−Re(z) ≥ C1 |Im(z)|λ e−Re(z)

(2.30) = C1 η(z)
(2.31) ≥ C1 ε

Case λ < 0:
 λ
(2.32) C1 |z|λ e−Re(z) ≥ C1 |Re(z)|e−Re(z)/λ + |Im(z)|e−Re(z)/λ

The first term is bounded above, as Re(z) is bounded above by − log(ε), the second term
is bounded above by ε1/λ .
3. We claim that for every ε > 0, the set {η ≥ ε} is complete with respect to the
metric g = (e−Re(z) h(z))2 |dz|2 , thus any Cauchy sequence zj ∈ A without limit must sat-
isfy η(zj ) → 0. Namely, we have g ≥ Cgeucl on {η ≥ ε/2} be the previous step, and the
boundary curves {η = ε}, {η = ε/2} have some positive distance δ, so

(2.33) d(z1 , z2 ) ≥ C min{|z1 − z2 |, 2δ}

for z1 , z2 ∈ {η ≥ ε}. Hence, any sequence which is Cauchy with respect to g is Cauchy for
the standard metric.
264 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH

4. We claim that diam{η ≤ ε} → 0 as ε → 0. Assume 1 > ε > 0 is sufficiently small

so that {η ≤ ε} contains no z with Re(z) ≤ 0, |Im(z)| ≤ ρ, then for z ∈ {η ≤ ε} the path
α(t) = z + t, t ∈ [0, ∞) is contained in {η ≤ ε}. We compute the length of α,
 ∞
(2.34) l(α) = e−Re(z)−t h(z + t)dt
0
 ∞
(2.35) ≤ C2 e−Re(z)−t |z + t|λ dt =: C2 I
0

Case λ > 0:
 ∞ λ
(2.36) I≤ |Re(z) + t|e−(Re(z)+t)/λ + |Im(z)|e−(Re(z)+t)/λ dt
0

As η(z) ≤ ε, Re(z) ≥ − log(ε) > 0, the first summand is bounded above by (− log(ε) +
t)ε1/λ e−t/λ , and the second by ε1/λ e−t/λ , so
 ∞
 λ
(2.37) I≤ε 1 − log(ε) + t e−t dt
0

(2.38) ≤ Cεe(1−log(ε))/2
√
(2.39) = C eε

where C = C(λ) is a constant.

Case λ ≤ 0:
 ∞
 λ
(2.40) I≤ e−Re(z)−t max 1, |Im(z)| dt
0

(2.41) = η(z) ≤ ε

To prove the claim, it remains to show that for different values of z1 , z2 , the corresponding
horizontal curves starting at z1 , z2 have vanishing distance. Let β(t) = x + it, t ∈ [y1 , y2 ],
x > ρ, then
 y2
(2.42) l(β) = e−x h(x + it)dt
y1
 y2
−x x→∞
(2.43) ≤ C2 e |x + it|λ dt −−→ 0
y1

Which completes the proof of the lemma. 

Proof of Proposition 2.5. — After a change of coordinates w = eπi/n /z, ϕ is of the form

ϕ = e−w h(w)dw2
n
(2.44)
 FLAT SURFACES AND STABILITY STRUCTURES 265

with h meromorphic in a neighborhood of ∞. Let B = D−1 and fix ε ∈ ( 12 , 1). Define

sectors
   
 2π k  πε
(2.45) Vn = w ∈ B | arg(w) − <
n  n
and their complement


n−1
(2.46) A=B\ Vn .
k=0

Then for z = reiφ ∈ A we find cos(nφ) ≤ cos(π ε) < 0, hence

|ϕ| = e−r
n cos(nφ)
(2.47) |hdz2 | ≥ C|dz|2

for some C > 0. Since the sets Vk have positive distance with respect to the standard
metric, this shows that every Cauchy sequence in B without limit is eventually contained
in one of the Vk .
On Vk we perform a change of coordinates u = wn , so

(2.48) ϕ = e−2u f (u)du2

and there are C1 , C2 > 0 with

 m−n+1
(2.49) C1 |u|λ ≤ |f (u)| ≤ C2 |u|λ , λ=
2n
where m = ord∞ h. Applying the previous lemma, the first part of the theorem follows.
By the first part of the proof,

(2.50) B̄ =: B ∪ {a1 , . . . , an }

is a complete metric space. Let a = ak , the we must show that a is an ∞-angle singularity
of B̄.
As Vk is simply connected, we may choose a holomorphic f : Vk → C with
(df )2 = ϕ. Choose r > 0 such that D∗2r (0) ⊂ Vk . Then for any path α in D∗2r with endpoint
z1 , z2 ∈ D∗r (0) we compute
  
   z2 √  
(2.51) f (z2 ) − f (z1 ) =  ϕ  ≤ |ϕ|

z1 α

hence

(2.52) |f (z2 ) − f (z1 )| ≤ d(z1 , z2 )

so that f extends to a continuous map on Vk ∪ {a}.

266 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH

Without loss of generality, f (a) = 0. We claim that f restricts to a covering

D∗r (a)→ D∗r (0). By (2.52), the image of D∗r (a) under f is contained in Dr (0). The ra-
dius of injectivity at z ∈ D∗r (a) is r1 = d(z, a). It cannot be larger, as the singular point
a would have to be contained in any disk of radius > r1 centered at z, and it cannot be
smaller by completeness of B̄, and since any geodesic may be extended until it hits a or
leaves D2r (a). We conclude that Dr1 (z) is mapped isometrically to Dr1 (f (z)) by f , and
thus r1 = |f (z)|. This shows that f restricts to a covering D∗r (a) → D∗r (0).
From the proof of the lemma, it follows that a admits a fundamental system of
neighborhoods W with the property that W \ {a} is simply-connected. Thus D∗r (a) must
be simply connected, and f : D∗r (a) → D∗r (0) is a universal covering. 

Next we prove a kind of converse to the previous proposition, which is the main
result of this subsection.

Theorem 2.3. — Let X be a flat surface of finite type. Then there is a compact Riemann surface
C with quadratic differential ϕ of exponential type giving C \ D(ϕ) the structure of the smooth part of
a flat surface isomorphic to Xsm .

Proof. — Let K := Core(X), which we may assume to contain more than one point.
Define a boundary walk to be a piecewise geodesic path which follows the boundary of K
so that S \ K lies to the right (and possibly also to the left), takes the rightmost possible
direction at every singularity, and is maximal with these properties. By rightmost direction
we mean the following. From each singularity there is a finite set E of directions which lie
in ∂K, and they are either cyclically or totally ordered. So the direction from which we
approach the singularity has some successor, by definition the rightmost direction, or is a
maximal element, and the boundary walk ends. Hence a boundary walk is either closed
or starts and ends at infinite-angle singularities, and there is a unique boundary walk
starting/ending at each infinite-angle singularity. If the boundary walk starts and ends at
an infinite-angle singularity then t cuts X into two pieces, one of which is topologically a
disk lying completely outside the core K. This follows from the description of X in terms
of the core above.
Let si ∈ Xsg , i ∈ Z/k be a cyclic sequence of infinite-angle singularities so that there
is a boundary path from si to si+1 .
Choose a closed embedded curve α on Xsm which is a smoothing of the cyclic
sequence of boundary paths and cuts X into two pieces X , X so that X contains the si
and no other singularities. The smooth part of X is topologically an annulus. It consists,
up to modifications coming from the smoothing, of several pieces of the complement of
K which are topological disks. The smoothing cases these disks to be glued together at
infinite-angle singularities to form X . To complete the proof, it suffices to show that X
corresponds to some punctured neighborhood of an exponential singularity.
Fix some direction ri from the singularity si . For each integer n > 0 consider the
pair of geodesics Li , Ri which start at si in the directions ri + nπ/2 and ri − nπ/2 respec-
 FLAT SURFACES AND STABILITY STRUCTURES 267

tively. For n sufficiently large, all geodesics Li , Ri can be extended indefinitely in X . We

construct a flat surface Xn as follows. Cut X along Li , Ri and remove the (contractible)
pieces left of each Li and right of each Ri . Then glue each Li to Ri to obtain Xn . Note that
Xn has k conical singularities, all with cone angle nπ . Let Un ⊂ Xn be the complement of
the geodesic rays along which Xn was glued, including the singularities. If we also denote
the corresponding subset of X by Un , then

(2.53) Un ⊂ Un+1 , Un = Xsm .
n

By the previous results about flat surfaces without infinite-angle singularities, there
is a biholomorphic fn : Xn → D∗ and meromorphic quadratic differential ϕn on D with a
pole at 0 and k zeros, such that fn becomes an isometry of flat surfaces. Since D ⊂ C is
bounded, the fn form a normal family and, after passing to a subsequence, converge to
a holomorphic f : Xsm → C. For degree reasons, f cannot be constant and consequently
has image contained in D∗ . By a standard application of Hurwitz’s theorem, f is injective
as the limit of injective functions. We wish to show that the image of f is all of D∗ , and
that the ϕn converge to a quadratic differential with exponential singularity at the origin.
Let gn (z)dz2 = ϕn (z) and consider hn = (∂gn /∂z)/gn , then hn is meromorphic on the
unit disk with k + 1 simple poles, one of which is at the origin. Convergence of fn implies
convergence of hn , possibly after passing to a subsequence, to a meromorphic function
h with k + 1 poles, counted with multiplicity. As X has k infinite-angle singularities we
must have a single pole of order k + 1 at the origin, so that ϕn converge to ϕ of the
form ep(z) q(z)dz with p, q meromorphic and p having a pole of order k at the origin.
Completeness of X requires the image of f to be D∗ . 

2.6. Voronoi and Delaunay partitions

In this subsection we review the Voronoi and Delaunay partitions of a flat surface,
cf. Masur–Smillie [35]. They will be used later to construct convergent sequences of flat
surfaces.
Let X be a flat surface, not necessarily of finite type, but so that every compo-
nent contains a conical point or marked point (i.e. a point in Xsg ). For p ∈ X let ρp =
d(p, Xsg ) ≥ 0 be the distance to the closest conical point. Further, define μp ≥ 1 to be
the number of directions from p in which the geodesic of length ρp starting at p ends at
a conical point if ρp > 0, and μp = 1 if ρp = 0. In other words, ρp is the radius of the
largest isometrically immersed disk centered at p, and μp is the number of points on its
boundary mapped to Xsg . The Voronoi cells are the connected components of the level
sets of μp . Voronoi 2-cells, 1-cells, and 0-cells are components with μp = 1, μp = 2, and
μp ≥ 3 respectively. The 2-cells are contractible and in bijection with Xsg . The 1-cells are
geodesically embedded open intervals, possibly of infinite length. The union of the 0-cells
is a discrete subset of X.
268 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH

The Delaunay polygonal subdivision is in a sense dual to the Voronoi partition.

Given p ∈ X let D the largest isometrically immersed disk centered at p, and let C(p) ⊂ X
be the image of the convex hull (in D) of the points on the boundary of D which map
to Xsg . Note that when p is contained in a Voronoi 2-cell, 1-cell, or 0-cell, then C(p)
is a point, edge, or polygon respectively, and C(p) only depends on the component of
the Voronoi partition. The sets C(p) provide a polygonal subdivision of Core(X). For a
generic flat surface all polygons are triangles. Although the Voronoi and Delaunay cell
decompositions are combinatorially dual, a Delaunay edge need not always intersect its
Voronoi dual.
For a flat surface X of finite type consider the quantity
 
(2.54) ρ(X) := sup ρ(p) | p ∈ Core(X) = dH Xsg , Core(X)
where dH denotes the Hausdorff distance of closed subsets. We use ρ(X) to get an upper
bound on the lengths of Delaunay edges.

Lemma 2.2. — Let X be a flat surface of finite type, and C ≥ 0 the maximal length of any
saddle connection on the boundary of Core(X). Then the length of any Delaunay edge is bounded above
by max(C, 2ρ(X)).

Proof. — Let e be a Delaunay edge and p ∈ X with C(p) = e. The geodesic e is a

chord of the largest immersed disk, D, centered at p. If p ∈ Core(X), then length(e) ≤
2ρp ≤ 2ρ(X). On the other hand, if p ∈ / Core(X), then the preimage of the boundary of
Core(X) is a either one or two chords in D, one of which has length greater than or equal
to length(e), and less than or equal to C. 

3. The topological Fukaya category of a surface

In this section we give an elementary definition of the (topological) Fukaya cate-

gories of surfaces, and prove several results about them. First, in Section 3.1, we collect
some definitions concerning A∞ -categories. In the following subsection we introduce a
class of marked surfaces S which is convenient for the definition of F (S), which is given in
Section 3.3 in terms of explicit A∞ -structures. In Section 3.4 we describe F (S) in terms
of graded quivers with quadratic monomial relations, a central tool in what follows. This
is possible only for some S, but the remaining categories are obtained by localization, as
is discussed in Section 3.5. In the final subsection, we show that F (S) is the category of
global sections of a certain cosheaf of categories on a ribbon graph in S.

3.1. Preliminaries on A∞ -categories

We will need some basic definitions of the A∞ language, for more details see e.g.
[32, 44].
 FLAT SURFACES AND STABILITY STRUCTURES 269

An A∞ -category, A, consists of a set Ob(A) of objects, for each pair X, Y ∈ Ob(A)

a Z-graded vector space Hom(X, Y), and structure maps

μn : Hom(Xn−1 , Xn ) ⊗ · · · ⊗ Hom(X0 , X1 ) → Hom(X0 , Xn )[2 − n]

for each n ≥ 1, satisfying the A∞ -relations

 
(3.1) (−1)ak +···+a1  μi+1+k an , . . . , an−i+1 , μj (an−i , . . . , ak+1 ), ak , . . . , a1 = 0
i+j+k=n

where a = |a| − 1 is the reduced degree.

An A∞ -category A is strictly unital if for every object X there is a 1X ∈ Hom0 (X, X)
such that

(3.2) μ1 (1X ) = 0
(3.3) μ2 (a, 1X ) = (−1)|a| μ2 (1Y , a) = a, a ∈ Hom(X, Y)
(3.4) μk (. . . , 1X , . . .) = 0 for k ≥ 3

Strictly unital A∞ -categories with μk = 0 for k ≥ 3 correspond to small dg-categories with

(3.5) da = (−1)|a| μ1 (a), ab = (−1)|b| μ2 (a, b).

Indeed, in this case the first three A∞ -relations correspond to d 2 = 0, the Leibniz rule,
and associativity of the product.
An A∞ -functor F : A → B is given by a map ObA → ObB and multilinear maps

(3.6) Fd : HomA (Xd−1 , Xd ) ⊗ · · · ⊗ HomA (X0 , X1 ) → HomB (FX0 , FXd )[1 − d]

for every d ≥ 1, X0 , . . . , Xd ∈ ObA, satisfying the relations

 
(3.7) μrB Fsr (ad , . . . , ad−sr +1 ), . . . , Fs1 (as1 , . . . , a1 )
r≥1 s1 +···+sr =d
 
(3.8) = (−1)an +···+a1  Fd−m+1 ad , . . . , an+m+1 , μmA (an+m , . . . , an+1 ), an , . . . , a1 .
m,n

We will mostly consider strict A∞ -functors, i.e. those with Fd = 0 for d > 1.

Twisted complexes
We briefly recall the construction of the category of twisted complexes, TwA, over
an A∞ -category A. Twisted complexes can be thought of as formal vector bundles with
flat connection.
270 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH

The first step is to form addZA whose objects are formal sums of tensor products
of the form

(3.9) V= VX ⊗ X
X∈ObA

with VX finite-dimensional graded vector spaces, zero for all but finitely many X. Mor-
phism spaces are defined by

(3.10) Hom(V ⊗ X, W ⊗ Y) = Hom(V, W) ⊗ Hom(X, Y)

and additivity. When extending the μk a Koszul sign gets introduced:


|φi |aj 
(3.11) μk (φk ⊗ ak , . . . , φ1 ⊗ a1 ) = (−1) i<j φk · · · φ1 ⊗ μk (ak , . . . , a1 )

Note that addZA has a natural shift functor and formal finite direct sums. Strictly speak-
ing, Ob(addZA) is not a set, but since the category of finite-dimensional vector spaces is
essentially small, this is a non-issue.
An object in TwA is given by a pair (V, δ) with V ∈ addZA and δ ∈ Hom 1
 (V, V).
The first condition is that there is a direct sum decomposition of V (i.e. of VX as an
ObA × Z-graded vector space) so that δ is strictly upper triangular. This ensures that
μk (δ, . . . , δ) = 0 for k big, so that the second condition, which is

(3.12) μk (δ, . . . , δ) = 0
k≥1

makes sense. Morphism spaces are just

 
(3.13) Hom (V, δ), (W, ) = Hom(V, W)

and the structure maps are

(3.14) μk (ak , . . . , a1 ) = μk+n0 +···+nk (δk , . . . , δk , ak , . . . , a1 , δ0 , . . . , δ0 )

     
n0 ,...,nk ≥0 nk times n0 times

where ai ∈ Hom((Vi−1 , δi−1 ), (Vi , δi )). Again, this sum is actually finite by our require-
ment on the δi .
The main fact we need about TwA is that its homotopy category, H0 (TwA), is tri-
angulated. When we write K0 (TwA), we always mean the K0 -group of this triangulated
category.
 FLAT SURFACES AND STABILITY STRUCTURES 271

3.2. Surfaces with marked boundary

In this subsection we define a class of surfaces which will be convenient for our
definition of the Fukaya category, as well as cut and paste constructions. Topologically,
they are related to the real blow-ups of flat surfaces by gluing several infinite ends.
We will work with surfaces with corners, S, and denote their 0- and 1-dimensional
strata by ∂0 S and ∂1 S respectively. The boundary is ∂S = ∂0 S  ∂1 S set-theoretically.
A marked surface is a surface S with corners together with a subset M ⊂ ∂S which is the
closure of a collection of components of ∂1 S such that if B is a connected component of
∂S then the following holds: If B is smooth, then it belongs entirely to M, and if B con-
tains corners, then every other component of ∂1 S ∩ B belongs to M, i.e. ∂0 S is exactly the
boundary of M in ∂S. In particular, if S is compact then each component of ∂S is either
a circle which belongs entirely to M or a sequence of intervals which alternately belong
to M and its complement.
An arc in a marked surface S is an embedded closed interval intersecting M trans-
versely in the endpoints, and not isotopic to an embedded interval in M (the isotopy
keeping endpoints in M). We generally identify isotopic arcs. A boundary arc is an arc
which is isotopic to the closure of a component of ∂S \ M. Arcs which are not boundary
arcs are referred to as internal arcs. An arc system in (S, M) is a collection of pairwise disjoint
and non-isotopic arcs. The arc system is full if it includes all the boundary arcs and cuts
the surface into contractible, relatively compact components (polygons).
A map of marked surfaces f : (S1 , M1 ) → (S2 , M2 ) is an orientation preserving
immersion with f (M1 ) ⊂ M2 mapping boundary arcs of S1 to disjoint non-isotopic arcs
in S2 . The condition insures that a full arc system in S2 which includes all the images of
boundary arcs under f can be lifted to a full arc system on S1 . It is however not closed
under composition in general.

Combinatorial description: graded ribbon graphs

Let (S, M, η) be a graded marked surface and A a full system of graded arcs on
it. Such a collection of arcs is dual to a graph  on S having an edge for each element
of A and an n-valent vertex for each 2n-gon cut out by A (see Figure 3). By definition,
a half-edge is an edge of  with orientation pointing towards a vertex of the graph (but
not a boundary arc). Since the graph is embedded in an oriented surface, there is a
clockwise cyclic order on the set of half-edges pointing towards a given vertex (i.e.  has
the structure of a ribbon graph). Moreover, for any half-edge h the grading determines an
integer d(h) as follows. Let σ (h) be the half edge following h in the cyclic order. Then
h, σ (h) correspond to a pair of arcs α, β in A. Let c : [0, 1] → M be the embedded curve
which starts on α, ends on β, follows the boundary with S to the right, and bounds
the polygon corresponding to the vertex that h, σ (h) point to. Also choose an arbitrary
grading on c. Then
(3.15) d(h) = ic(0) (α, c) − ic(1) (β, c)
272 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH

FIG. 3. — Full system of arcs in a marked surface and dual ribbon graph (dashed)

is independent of the choice of grading on c. The numbers d(h) have the property that if
v is a vertex of , then

(3.16) d(h) = val(v) − 2

h

where the sum is over all half-edges pointing to v and val(v) is the valency of v.
Conversely, given a ribbon graph together with integers d(h) for every half-edge
h satisfying (3.16) one constructs a graded marked surface with arc system by gluing
polygons with suitable grading foliation which is tangent to the arcs. This is a convenient
way to specify any compact graded marked surface by a finite amount of data.

3.3. Minimal A∞ -category of an arc system

Fix a field of scalars K throughout. Let S = (S, M, η) is a graded marked surface
with system of graded arcs A. We define a strictly unital A∞ -category FA (S) with μ1 = 0.
• Objects: The set of arcs in A.
• Morphisms: A boundary path is a non-constant path in M which follows the re-
verse orientation of the boundary (i.e. the surface lies to the right). Given arcs
X and Y, a basis of morphisms from X to Y is given by boundary paths, up
to reparameterization, starting at an endpoint of X and ending at an endpoint
of Y, as well as the identity morphism if X = Y. The degree of a boundary path
a from p to q joining arcs X and Y is by definition

(3.17) |a| = ip (X, a) − iq (Y, a)

for arbitrary grading of a.

• Composition: Let a, b be boundary paths defining morphisms from X to Y and Y
to Z respectively. If a and b are composable then (−1)|a| μ2 (b, a) = a ·b, otherwise
μ2 (b, a) = 0.
 FLAT SURFACES AND STABILITY STRUCTURES 273

• Higher operations: Consider a marked surface (S , M ) which is topologically a

closed disk with M having n ≥ 3 components. Let a1 , . . . , an be the distinct
boundary paths (components of M ) ending in boundary arcs in clockwise or-
der. Given a map f : (S , M ) → (S, M) sending all boundary arcs of (S , M ) to
arcs in A we get a sequence f ◦ a1 , . . . , f ◦ an of boundary paths in (S, M), and
call any such sequence a disk sequence.
We define higher A∞ -operations so that if a1 , . . . , an is a disk sequence then

(3.18) μn (an , . . . , a1 b) = (−1)|b| b

for basis morphisms b with a1 b = 0, and

(3.19) μn (ban , . . . , a1 ) = b

for paths b with ban = 0, and μn vanishes on all sequences of paths not of the
above forms. The lemma below ensures that this is really well-defined.

Lemma 3.1. — For a sequence of composable basis morphisms an , . . . , a1 there is at most one
factorization a1 = a1 b with a1 , a2 , . . . , an a disk sequence. If such a factorization exists with b not
an identity, then there is no factorization an = can with a1 , . . . , an−1 , an a disk sequence and c not an
identity. The dual statement also holds.

Proof. — To see the first statement note that if we have such a factorization and αi is
the arc on which ai ends, then the concatenation a1 · α1 · · · an · αn is a null-homotopic loop.
Hence, b must be homotopic (relative endpoints) to the concatenation a1 · α1 · · · an · αn and
is thus uniquely determined as a morphism.
For the second statement, assume such a factorization exists. Then c = b by the
same argument as before, but b and an end at different endpoints of αn , contradicting
an = can . 

Proposition 3.1. — With the structure defined above, FA (S) is a strictly unital A∞ -category.

Proof. — Let an , . . . , a1 be a composable sequence of morphisms corresponding to

boundary paths. The claim is that the A∞ -equation (3.1) holds. For n = 1, 2, 3 there is
nothing to prove, so we may assume n ≥ 4. The following types of non-zero terms cancel
pairwise, ignoring signs for the moment:
 
μ2 μi (an , . . . , a1 b), c
 
μi an , . . . , a2 , μ2 (a1 b, c)
 
μ2 a, μi (bcn , . . . , c1 )
 
μi μ2 (a, bcn ), cn−1 , . . . , c1
 
μi ai , . . . , aj+1 , μk (aj bk , . . . , b1 ), aj−1 , . . . , a1 c
274 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH
 
μi+k−2 ai , . . . , aj bk , . . . , μ2 (b1 , aj−1 ), aj−2 , . . . , a1 c
 
μi ai , . . . , aj+1 , μk (bk , . . . , b1 aj ), aj−1 , . . . , a1 c
 
μi+k−2 ai , . . . , aj+2 , μ2 (aj+1 , bk ), . . . , b1 aj , aj−1 , . . . , a1 c
as well as variants of the previous two with c on the other side, and
 
μi ai , . . . , a2 , μj (a1 bcj , cj−1 , . . . , c1 )
 
μj μi (ai , . . . , a2 , a1 bcj ), cj−1 , . . . , c1
All non-zero terms which can appear must indeed belong to one of the above pairs. To
check that the signs are in fact opposite, one uses that
(3.20) ai  = −2
for any disk sequence a1 , . . . , an . 
There are two special cases for which the above description needs to be somewhat
amended, as disks also contribute to μ2 . The first is that of a square, in which the two
boundary arcs are isotopic. They represent the same object X, which has End(X) = K.
The second case is when S is a cylinder and M = ∂S. Up to isotopy there is again only
one arc, and the corresponding object X has End(X) = K[z±1 ], with the degree of z
depending on .
Remark. — The A∞ -category FA (S) was considered by Bocklandt [9], at least in the
case when S has only marked boundary circles. He shows that it computes the partially
wrapped Fukaya category of S as defined by Abouzaid–Seidel [2]. We expect the proof
to extend to the case of partially wrapped Fukaya categories of Auroux [5].
Certain maps between graded marked surfaces induce functors between the re-
spective Fukaya categories.
Proposition 3.2. — Let f : S1 → S2 be a map of graded marked surfaces and Ai and arc system
on Si such that f maps graded arcs in A1 to graded arcs in A2 . Then f induces a strict A∞ -functor f∗
from FA1 (S1 ) to FA2 (S2 ).
Proof. — Since f sends arcs in A1 to arcs in A2 and sends boundary arcs in S1
to boundary arcs in S2 , it is clear how to define f∗ on objects and morphisms (and f∗ is
faithful, but not full in general). To see that f∗ is a strict A∞ -functor note that if D is an
immersed disk D in S2 whose boundary consists of arcs and boundary curves, then if ∂D
lifts to S1 , so does D. 

The simplest example

Consider the case when S = (S, M, η) is a graded marked surface which is topo-
logically a disk and M has n ≥ 3 components, and A is the arc system containing exactly
 FLAT SURFACES AND STABILITY STRUCTURES 275

the boundary arcs, with some arbitrary grading. The A∞ -category FA (S) 
has objects Ek ,
k ∈ Z/n, and morphisms ak : Ek → Ek+1 of some degrees |ak | ∈ Z with |ak | = n − 2
which, together with the identity morphisms, form a basis of all morphisms. The only
non-zero A∞ -terms come from strict unitality and
(3.21) μn (ak+n−1 , . . . , ak ) = 1Ek .
See also [36] for a discussion of these categories.
We make a simple observation which will be essential in what follows. Namely, that
the twisted complex
a1 an−2
E1 −→ E2 [a1 ] −→ · · · −
−→ En−1 [a1  + · · · + an−2 ]
is isomorphic to En [−|an |], the inverse isomorphisms being given by an−1 and an . In other
words, n − 1 of the boundary arcs already generate all of Tw(FA (S)). This also shows that
Tw(FA (S)) is the bounded derived category of an An−1 quiver with all arrows oriented in
the same direction. To set up this equivalence send the objects E1 , . . . , En−1 to the simple
objects corresponding to vertices of the quiver, up to some shift.

Morita invariance
Fix a graded marked surface S. If A ⊂ B are full arc systems, i.e. all arcs in A
also belong to B, then it is evident from the definition that FA (S) is a full subcategory of
FB (S).

Lemma 3.2. — The inclusion functor FA (S) → FB (S) for full arc systems A ⊂ B is a
Morita equivalence.

Proof. — We can assume that B is obtained from A by adding a single graded arc Y.
Let D be one of the disks cut out by B and with Y on its boundary. Considering D as a
graded marked surface with minimal full arc system, we have a morphism f : D → S
inducing a functor of A∞ -categories. From the discussion above, we know that in F (D)
the object corresponding to Y is isomorphic to a twisted complex over the other boundary
arcs, Y1 , . . . Yn , of D. As A is a full arc system, there is exactly one boundary of D which
gets mapped to Y, hence Y1 , . . . , Yn get mapped to arcs in A. Applying the functor f∗ we
thus find that Y is isomorphic to a twisted complex over objects in FA (S). In particular,
the inclusion functor FA (S) → FB (S) is a Morita equivalence. 

Proposition 3.3. — The Morita equivalence class of FA (S) is independent of A. The equiva-
lences are canonical in the higher categorical sense (explained in the proof).

Proof. — Suppose A, B are full systems of graded arc. As established in the previous
lemma we get a canonical Morita equivalence when A ⊂ B. The same is also clearly true
when A and B just differ by grading shift.
276 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH

Consider the set A of full arc systems up to isotopy on (S, M), partially ordered
by inclusion. If we view A as a category in the usual way, then the arguments above
show that we get a functor A → FA (S) from A to the category of strictly unital A∞
categories and Morita equivalences. From contractability of the classifying space |A| of
full arc systems or ribbon graphs (see [25, 26]), it now follows that the various categories
FA (S) are canonically Morita equivalent. 

Thus, TwFA (S) is essentially independent of A and we may drop it from the nota-
tion and just write F (S), the (topological) Fukaya category of S.

3.4. Formal generators

Recall that an A∞ -algebra A is formal if it is quasi-isomorphic to its cohomology
H∗ A. A generator G of an A∞ -category is formal if its endomorphism algebra End(G) is.
We will see in this subsection that F (S) has a formal generator whenever S has at least
one boundary arc (unmarked part of ∂S) in each connected component. The generator
is just a direct sum of arcs, and its endomorphism algebra is the path algebra of a quiver
with quadratic monomial relations.
To set up some notation, a graded quiver Q with quadratic monomial relations is given by
• Q0 —set of vertices
• Q1 —set of arrows
• ∂0 , ∂1 : Q1 → Q0 —source and target maps
• |.| : Q1 → Z—grading of arrows
• R ⊂ Q1 ×Q0 Q1 —quadratic monomial relations, i.e. a collection of pairs of com-
posable arrows
Given a ground field K we may form the path category KQ = KQ/R, which is a graded
linear category with set of objects Q0 and basis of morphisms given by paths of arrows
not containing any of the relations.
Let S be a graded marked surface, compact for simplicity. Note first that if A is
any arc system on S, then the graded category obtained from FA (S) by forgetting the
higher μk is of the form KQ for some graded quiver with quadratic monomial relations.
Namely, take as arrows the boundary paths which start and end at arcs of A but do
not cross any other endpoints, and quadratic relations coming from composable arrows
which do not correspond to composable paths. Furthermore, the higher A∞ operations
of FA (S) vanish under the following condition on A: Any disk cut out by A is bounded by
a boundary arc of S which does not belong to A. We call such systems of arcs formal. Thus,
under this condition, the category FA (S) is of the form KQ for some graded quiver with
quadratic monomial relations Q. Call a formal system of arcs full if it cuts S into disks
each of which has exactly a single boundary arc (of itself) not belonging to A. For a full
formal system of arcs A the category Tw(FA (S)) is quasi-equivalent to F (S).
 FLAT SURFACES AND STABILITY STRUCTURES 277

Lemma 3.3. — Let S be a compact and connected graded marked surface. If S has at least one
boundary arc, then it has a full formal system of arcs.
Proof. — This follows from the fact that we can find a system of arcs on S which
cuts it into a single disk, the boundary of which must include all boundary arcs of S. This
in turn is seen from the dual ribbon graph. If it has more than one vertex we can contract
some edges to reduce the number of vertices to one eventually. 
We say a graded quiver with relations is of type F1 if it arises as above from a
full formal system of arcs. It is not difficult to see that a graded quiver with quadratic
monomial relations is of type F1 if and only if
1. There are no cycles a1 , . . . , an , n ≥ 1, with ai ai+1 = 0 for all i ∈ Z/n.
2. Each vertex has at most two incoming and outgoing arrows.
3. Let a, b = c be arrows. If ab, ac are defined, then ab = 0 or ac = 0, but not both.
If ba, ca are defined, then ba = 0 or ca = 0, but not both.

Resolution of the diagonal

Let Q be a graded quiver with quadratic monomial relations and A = KQ/R be
its path algebra. We want to investigate regularity properties of A. Denote the constant
path at the vertex i by ei . There is a splitting of A
 
(3.22) A= Aei , A= ei A
i∈Q0 i∈Q0

as a left (resp. right) module over A.

Define Aop ⊗ A-modules

(3.23) Mn = Ae∂0 α1 ⊗ e∂1 αn A, n≥1
α1 ,...,αn ∈Q1
(αi ,αi+1 )∈R

(3.24) M0 = Aei ⊗ ei A
i∈Q0

connected by maps fn : Mn → Mn−1 such that for a ⊗ b in the (α1 , . . . , αn )-component

of Mn
(3.25) fn (a ⊗ b) = aα1 ⊗ b + (−1)n a ⊗ αn b
and f0 : M0 → A, f0 (a ⊗ b) = ab.
Proposition 3.4. — The sequence of bimodules
f2 f1 f0
(3.26) ... M2 M1 M0 A 0

is exact.
278 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH

A proof can be found in [6].

Proposition 3.5. — Let Q be a graded quiver with quadratic monomial relations, A = KQ/R
its path algebra.
1. Suppose there are no cyclic paths α1 · · · αn in Q with (αi , αi+1 ) ∈ R, i ∈ Z/n, then A is
homologically smooth.
2. Suppose there are no cyclic paths α1 · · · αn in Q with (αi , αi+1 ) ∈
/ R, i ∈ Z/n, then A is
proper.

Proof. — 1. Under the stated condition on cyclic paths, we see that Mn = 0 for
n  0, hence A is perfect as an Aop ⊗ A-module.
2. The condition implies that only finitely many paths are non-zero in A, so A has
finite rank over K. 

We apply these results to graded linear models of F (S). In this case, the first condi-
tion of Proposition 3.5 is always satisfied, while the second is satisfied if S has no boundary
components without corners.

Corollary 3.1. — Let S be a compact graded marked surface without boundary components
diffeomorphic to S1 , then F (S) is homologically smooth and proper.

Question. — If S is a compact graded marked surface which has a full arc system
that gives an acyclic quiver, then the category F (S) has a full strong exceptional collection
formed by these arcs. Are there simple conditions on S which allows one to find such an
arc system?

3.5. Localization
We begin by recalling a version of Drinfeld’s construction for strictly unital
A∞ -categories, studied in detail in [33]. Let A be a strictly unital A∞ -category over a
field K and E ∈ Ob(A). It will be convenient to use the notation A(X, Y) for Hom(X, Y)
in this subsection. We will define the quotient category A/E = B of A by E, which will
again be a strictly unital A∞ -category. Informally, B is obtained by freely adjoining a
morphism ε ∈ B −1 (E, E) with μ1 (ε) = 1E . Set
(3.27) B (X, Y) = A(X, Y)⊕
   
 ⊗n
⊕ A(E, Y) ⊗ K[1] ⊗ A(E, E) ⊗ K[1] ⊗ A(X, E)
n≥0

as a graded vector space. Write generators of the summand

 ⊗n−2
A(E, Y) ⊗ K[1] ⊗ A(E, E) ⊗ K[1] ⊗ A(X, E)
 FLAT SURFACES AND STABILITY STRUCTURES 279

as

an εan−1 ε · · · a1

with an ∈ A(E, Y), a1 ∈ A(X, E), and ai ∈ A(E, E) for 2 ≤ i ≤ n − 1, then

(3.28) |an ε · · · εa1 | = |an | + · · · + |a1 | − n + 1

(3.29) an ε · · · εa1  = an  + · · · + a1 .

Structure maps are given by

(3.30) μr (anr ε · · · εanr−1 +1 , . . . , an1 ε · · · εa1 )

:= (−1)ak +···+a1  anr ε · · · εμj (aj+k , . . . , ak+1 )εak ε · · · εa1

i+j+k=nr
j+k>nr−1
k<n1

where 0 = n0 < n1 < · · · < nr , r ≥ 1. All this generalizes in a straightforward manner to

quotients by full subcategories E ⊂ A.
We return to the setting of surfaces with marked boundary where we consider the
following modification. If S is a marked surface and E a boundary arc, we get a new
marked surface S by adding E to the marked part of the boundary and smoothing the
two corners on which E ends. We do not quite get a morphism S → S , since E would
need to map to an arc which is isotopic to a path in M . For this reason we allow such arcs,
which we call null, in this subsection. The definition of the A∞ -category of a system of
arcs works as before with the following small change: The category is no longer minimal,
and the definition of μ1 is modeled on the one for the higher μi . This means that the
boundary path which is isotopic to a null arc E is a disk-sequence of length 1 and has
differential the identity morphisms of that arc.

Proposition 3.6. — Let A be an arc system on a graded marked surface S and E ∈ A a boundary
arc. By the modification as above we get S with arc system A containing a null arc. Then there is a
natural equivalence of A∞ -categories
 
(3.31) FA (S)/E ∼ = FA S

under FA (S).

Proof. — Comparing definitions, one notices an evident strict A∞ -functor

 
(3.32) G : FA S → FA (S)/E

compatible with the functors from FA (S). It is obtained by factoring a boundary path
in M into its pieces which are alternately contained in M and E. The functor G is not
280 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH

an isomorphism of A∞ -categories, but becomes one once we drop E from both the tar-
get and the source category. By general properties, E ∈ H∗ (FA (S)/E) is a zero-object,
and it is easily seen that the same is true in H∗ (FA (S )), implying that G is a quasi-
equivalence. 

Remark. — Localization of topological Fukaya categories of surfaces is also studied

in the recent works by Dyckerhoff [20] and Pascaleff–Sibilla [40]. The localizations they
consider depend more generally on an open subgraph of a ribbon graph on S.

3.6. Cosheaf of categories

Let S be a graded marked surface with full system of graded arcs A. Dual to A is
a ribbon graph G on S which is a deformation retract of the pair (S, ∂S \ M). A con-
structible cosheaf E of A∞ -categories on G is given, concretely, by assigning an A∞ -
category Cv (resp. Ce ) to each internal vertex v (resp. edge e) of G, together with functors
Ce → Cv for every half-edge. We understand this as a cosheaf with values in the higher
category of A∞ -categories with equivalences the Morita-equivalences, so when taking
global sections one needs to compute a homotopy colimit.
There is a canonical constructible cosheaf on G, defined as follows. For an edge
e, Ce is the category with a single object corresponding to the arc in A dual to e and
endomorphism algebra K. For a vertex v, let D be the dual polygon cut out by A, then Cv
is the category F (D) where we take the arc system consisting just of boundary arcs. Thus,
Cv is Morita-equivalent to the path-algebra of an An quiver, where n+1 is the valency of v.
The functors Ce → Cv are fully-faithful inclusions of boundary arcs of D. By construction,
we also have functors Ce → FA (S) and Cv → FA (S), and all these are compatible. Hence,
by universality there is a functor from global sections of E , (G, E ) → FA (S). The main
result of this subsection is that this is a Morita-equivalence.

Theorem 3.1. — Let S be a graded marked surface with ribbon graph G dual to a full graded
system of arcs. Then F (S) represents global sections of the constructible cosheaf E on G defined above.

The category of global sections (G, E ) is by definition the homotopy colimit of

the diagram formed by the Ce , Cv , and the functors between them. Thus we need to show
that F (S) represents this homotopy colimit. The strategy of the proof will be to first treat
the case with boundary arcs where we have a formal generator, then deduce the general
case by localization.
Let Qn be the quiver with n ≥ 1 vertices, arrows αi,i+1 from the i-th to the (i + 1)-st
vertex of degrees |αi,i+1 | ∈ Z, and quadratic relations αi,i+1 αi−1,i = 0 for 1 < i < n. Sup-
pose that Q = Qn1  · · ·  Qnk is a finite disjoint union quivers of this type, and that P
is a finite set with maps f1 , f2 : P → Q0 such that f1  f2 : P  P → Q is injective. Con-
struct a new quiver with relations R from Q by identifying the vertices f1 (p) and f2 (p) for
 FLAT SURFACES AND STABILITY STRUCTURES 281

each p ∈ P, and keeping the same arrows and relations. Passing to the associated path-
categories P , Q, R, where P is thought of as a quiver without arrows, we get a coequalizer
diagram

f1

(3.33) P Q R
f2

in the naive (1-categorical) sense.

Proposition 3.7. — The diagram (3.33) is a homotopy coequalizer in the category of dg-
categories.
 of Q making the
Proof. — To show this, we will make a cofibrant replacement Q
diagram

f1

(3.34) P 
Q
f2

cofibrant by the condition on f1  f2 , and verify and the coequalizer of (3.34) is quasi-
isomorphic to R.
The graded linear path-category A of Qn has a cofibrant replacement given by the
bar-cobar resolution B = BA. The dg-category B is described explicitly by a dg-quiver
with n vertices, arrows βij from the i-th to the j-th vertex, 1 ≤ i < j ≤ n, with degrees

(3.35) |βij | = 1 + (|αk,k+1 | − 1)

i≤k<j

and differential

(3.36) dβij = (−1)|βkj | βkj βik

i<k<j

extended to paths by the graded Leibniz rule. The functor P : B → A sends βi,i+1 to
αi,i+1 and all other arrows to zero. The right inverse I sending αi,i+1 to βi,i+1 is not mul-
tiplicative, but an inverse of P on the level of cohomology, as the homotopy H defined
by

(−1)|βik−1 ,ik | βik−2 ,ik · · · βi1 ,i2 if ik = ik−1 + 1, k ≥ 3
(3.37) H(βik−1 ,ik · · · βi1 ,i2 ) =
0 else

satisfying dH + Hd = 1 − IP, shows.

282 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH

We have a cofibrant replacement Q  = BQ which is described by a disjoint union

of the dg-quivers above. The coequalizer, R  of the diagram (3.34) is associated with
the dg-quiver we get by identifying pairs of vertices. There is a dg-functor R  → R,
which we claim is a quasi-isomorphism. To define the homotopy, let β1 , . . . , βn be non-
identity paths in the quiver of Q  which are not composable, but become so in R , so

that b := βn · · · β1 is a morphism in R. Note that any non-identity path in the quiver of
R canonically breaks up in such a way. Also define the length of a path, l(β), so that
l(βij ) = j − i and l(β2 β1 ) = l(β2 ) + l(β1 ). We define H extending (3.37) in the following
way. If l(βk ) = 1 for 1 ≤ k ≤ n, then H(βn · · · β1 ) = 0. Otherwise, there exists a largest k
with l(βk ) > 1 and

(3.38) H(βn · · · β1 ) := (−1)|βn |+···+|βk+1 | βn · · · H(βk ) · · · β1 .

Let us check the identity dH + Hd = 1 − IP in R . If l(βk ) = 1 for all k, then evidently

both sides vanish. Otherwise, we have a largest k with l(βk ) > 1 and the right-hand side
vanishes. On the other hand,

(3.39) (dH)(b) = (dH)(βk ) + (−1)|H(βk )|+···+|βj−1 | βn · · · H(βk ) · · · dβj · · · β1

j<k

and

(3.40) (Hd)(b) = (Hd)(βk ) + (−1)|βk |+···+|βj−1 | βn · · · H(βk ) · · · dβj · · · β1

j<k

so that the left-hand side also vanishes.

We conclude that the original coequalizer diagram (3.33) is quasi-isomorphic to a
cofibrant one, and hence also a homotopy coequalizer. 

Proof of Theorem 3.1. — Let S be a graded marked surface, which we may as well
assume to be connected. Suppose first that S has at least one boundary arc. Then S
has a full formal system of arcs A with dual ribbon graph G. Recall that (G, E ) is the
homotopy colimit of a diagram with the various categories Ce , Cv and functors between
them. Removing the categories Ce for e an edge ending in a boundary arc from the
diagram does not change the colimit, since such Ce are being included as subcategories
into some Cv only once. In the reduced diagram there is now, by formality of the arc
system, at least one object in each Cv which is not in the image of any functor Ce → Cv .
We remove such an object from each Cv , producing a Morita-equivalent diagram. Note
that the resulting diagram is now a diagram of graded linear categories. It is essentially
a diagram of the form (3.33) where P is the coproduct of the Ce (with P identified with
the set of arcs in A) and Q is the coproduct of the Cv . The coequalizer R is just the
graded linear model for F (S), and we know it is a homotopy coequalizer by the previous
proposition.
 FLAT SURFACES AND STABILITY STRUCTURES 283

The previous arguments also show that (G, E ) is independent of G, and hence
that the theorem holds in fact for any full system of arcs on a surface with boundary arc.
Namely, it suffices to check that global sections do not change when an edge of G is
contracted. This can be checked in a neighborhood of that edge, so we can assume that
G has only one internal edge and is dual to a full arcs system A on the disk D with single
internal arc. Let G be the graph obtained by contracting the internal edge in G, and E 
its cosheaf of categories. As it is already established that both (G, E ) and (G , E ) are
represented by F (D), they must be equivalent.
What remains is the case when S has no boundary arcs. As discussed in Section 3.5,
the category F (S) is a localization of a category F (S ) where S is obtained from S by
inserting a boundary arc on some boundary component. So to complete the proof of the
theorem it suffices to check that the corresponding categories of global sections have the
same relation.
Choose some ribbon graph G on S dual to a full system of graded arcs. To get a
ribbon graph G for S we just need to remove the edge ending on the unique boundary
arc. This means that in the diagram computing global sections we change some Cv from
type An to type An−1 , which is indeed just localization by the boundary arc. Finally, taking
the quotient by some object is a special case of a homotopy push out, thus commutes with
colimits, so that (G, E ) is the quotient of (G , E  ) by the boundary arc. 

4. Tameness and geometricity

In this section we deal with the problem of classifying objects in F (S). First, in
Section 4.1, we assign objects in F (S) to certain immersed curves in S with a local system
of vector spaces. The purpose of the rest of the section is to prove that we get all objects
in this way. In Section 4.2 we introduce nets and study their representations. Section 4.3
discusses a minimality condition on twisted complexes which ensures uniqueness up to
isomorphism. The proof of the classification is completed in Section 4.4.

4.1. Twisted complexes from curves

Fix a graded marked surface S and a ground field K. An immersed curve c in S is
unobstructed if it does not bound an immersed teardrop, which is a map from the closed
disk D to the surface which takes ∂D to c and which is a smooth immersion at every point
except one point of ∂D (see Figure 4). An admissible curve is an unobstructed graded curve
c such that one of the following holds:
1. The domain of c is S1 , the image of c is disjoint from ∂S, and c represents a
primitive class in π1 (S).
2. The domain of c is [0, 1], c intersects ∂S transversely in M and only in the
endpoints, and c is not homotopic relative endpoints to a path in M.
284 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH

FIG. 4. — Curve bounding a teardrop

Suppose that S is compact so that we have a category F (S), well defined up to canonical
equivalence. The purpose of this subsection is to show that an admissible curve together
with a local system of finite dimensional K-vector spaces (on its domain) gives an equiva-
lence class of objects in F (S).
First, in the case when S is topologically a disk any admissible curve c is a graded
arc. Thus we can find a full arc system A which includes c, so that c is an object of
FA (S) by definition. The isomorphism class of that object is clearly well-defined in F (S)
independently of the arc system.
Returning to the case of general S, we will first deal with admissible curves c which
have domain [0, 1]. For any such c we can find a graded marked surface S which is of
disk-type and with a map f to S so that c is the image of an admissible curve c̃ under f . To
see this, consider the universal cover  S of S and lift a full arc system A on S to 
A on 
S. Lift
  
c to c̃ on S and take as S a closed disk which is cut out by arcs in A and which contains
c̃. Now, c̃ gives an equivalence class of objects in F (S ), and the image under the functor
F (S ) → F (S) is independent of the choice of (S ). This follows from the fact that if we
have S , S as above, then the maps S → S, S → S both factor through a third S → S
as can be seen by looking at the universal cover again.
Suppose now instead that c is an admissible curve with domain S1 and local sys-
tem V of finite-dimensional vector spaces on it. We will follow the same strategy as be-
fore and assume first that (S, M) is of annular type, i.e. topologically a compact annulus
with corners on each boundary component. Choose a cyclic sequence of disjoint non-
isotopic arcs Xi , i ∈ Z/n so that at least one connects the two components of ∂S and such
that every component of M contains either exactly two endpoints of the arcs, belonging
Xi , Xi+1 for some i, or none of the endpoints (see Figure 5). Thus we get a sequence ai ,
i ∈ Z/n of distinct boundary paths so that ai connects endpoints of Xi , Xi+1 . If we fol-
low X0 , a0 , X1 , a1 , . . . we get a path which after suitable smoothing near the intersection
points becomes a simple closed loop isotopic to c. It is possible to choose grading and a
local systems on Xi , ai , so that the smoothed path is isotopic to c as a graded curve with
local system. As a result, each ai will be morphism of degree 1 either from Xi to Xi+1 or
in the other direction. Further we get a vector space Vi of sections of over Xi and parallel
transport Ti : Vi → Vi+1 along ai . Consider the twisted complex
 
(4.1) Vi ⊗ Xi , T±1
i ⊗ ai
 FLAT SURFACES AND STABILITY STRUCTURES 285

FIG. 5. — Example of allowed sequence of arcs (dashed)

where the signs are determined by the direction of the morphisms ai . This is the object
of F (S, M, ) we assign to c.

Lemma 4.1. — The equivalence class of the twisted complex constructed depends only on the
isotopy class of the graded curve c with local system.

Proof. — We claim first that we can replace X1 , . . . , Xn−1 by a single arc Y with
the pair X0 , Y giving an isomorphic twisted complex. Indeed, cutting S along X0 we
get a surface of disk type S in which the sequence of arcs X1 , . . . , Xn−1 concatenates
to a single graded arc Y. The corresponding isomorphism of twisted complexes formed
from X1 , . . . , Xn−1 and Y respectively was established in Section 3.3, and the claimed
isomorphism of twisted complexes formed from X0 , . . . , Xn−1 and X0 , Y follows.
We have reduced the problem to the case of two arcs. Any two pairs of arcs satisfy-
ing our requirements are related by Dehn twists (automorphisms) of S. It is clear that the
isotopy class of c is invariant under Dehn twists. To finish the proof we need to show that
the twisted complex associated with c is invariant under the induced autoequivalence, up
to isomorphism. This can be checked by direct computation, or by using the equivalence
    
(4.2) H0 Tw FX0 ,Y (S) = Db P1

under which X0 , Y correspond to O, O(1)[−1], the Dehn-twist to ⊗O(1), and the

twisted complex assigned to c to a torsion sheaf. 

We have excluded above the case when one or both boundary components are en-
tirely contained in M. These cases can be handled fairly easily directly, or alternatively by
the localization construction of the previous section. The case of general S is handled by
finding maps from surfaces of annular type. Here the argument uses the annular covering
associated with c instead of the universal covering.
286 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH

4.2. Representations of nets

In this subsection we classify representations of certain combinatorial structures,
called nets, based on ideas in the work of Nazarova–Roiter [37] on a class of tame linear
algebra problems. We will later use this result to derive the classification of objects in
Fukaya categories of surfaces with boundary. The authors are reluctant to introduce yet
another combinatorial structure, but it gives the necessary flexibility for the proof of the
classification.
A net is a quadruple X = (A, α, B, β) where
• A is a finite set,
• α is a fixed-point-free involution on A,
• B is a finite set with partition B = i∈A Bi into totally ordered sets,
• β is a fixed-point-free involution on Dom(β) ⊂ B.
A morphism of nets f : (A, α, B, β) → (A , α  , B , β  ) is given by maps f1 : A → A , f2 :
B → B such that

(4.3) f1 ◦ α = α  ◦ f1
(4.4) f2 (Domβ) ⊂ Domβ 
(4.5) f2 (B \ Domβ) ⊂ B \ Domβ 
(4.6) f2 |Domβ ◦ β = β  ◦ f2 |Domβ
(4.7) f2 (Bi ) ⊂ B f1 (i) and f2 |Bi is increasing

The height of a net X = (A, α, B, β) is defined as

(4.8) h(X ) = max |Bi |.

i∈A

Nets of height 1 are disjoint unions of the following types of nets:

α β α β α
(4.9) • • • • ··· •

and
α β α β α
(4.10) • • • • ··· •
β

where we have indicated elements of A = B by dots and the action of α, β by arrows. In

both cases |A| ≥ 2.
A representation of a net (A, α, B, β) is given by
 FLAT SURFACES AND STABILITY STRUCTURES 287

• a finite-dimensional vector space V with a direct sum decomposition


V= Vi ,
i∈A/α

• for each i ∈ A an increasing exhaustive filtration {Fj }j∈Bi on Vi (i.e. j ≤ k =⇒

Fj ⊂ Fk , Fmax Bi = Vi ),
• isomorphisms

φi : gri V → grβ(i) V, i ∈ Dom(β)

with φβ(i) = φi−1 . Here gri V = Fi /F<i is the associated graded.

In short, a representation of a net is a collection of vector spaces, each with two
filtrations, and with certain subquotients identified. Representations of a net form a linear
category with morphisms (V, Fi , φi ) → (W, Gi , ψi ) being linear maps f : V → W such
that f (Fi ) ⊂ Gi and

gri f
gri V gri W
φi ψi

grβ(i) f
grβ(i) V grβ(i) W

commutes for all i ∈ Dom(β). We note that the category of representations of a net as
in (4.9) is the category of finite dimensional vector spaces, while for (4.10) we get the
category of finite dimensional representations of Z (up to equivalence).
Given a morphism of nets f = (f1 , f2 ) : X = (A, α, B, β) → X  = (A , α  , B , β  )
with f2 injective on each Bi , there is an induced pushforward functor f∗ on the correspond-
ing categories of representations. If V = (V, Fi , φi ) is a representation of X , then we
produce a representation f∗ V = W = (W, Gi , ψi ) of X  with W = V as vector spaces and

(4.11) Wj = Vi , j ∈ A /α 
i∈A/α,f1 (i)=j


(4.12) Gl = Fk , where k = max r ∈ Bi | f2 (r) ≤ l
i∈A,l∈Bf (i)
1

for which we have


(4.13) grj W = gri V
i∈B,f2 (i)=j
288 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH

so that we can define


(4.14) ψj = φi .
i∈B,f2 (i)=j

If f2 fails to be injective on some Bi ’s, then one can still define f∗ V given additional choices
on V (i.e. in a non-functorial way). Namely, for each i ∈ A and j ∈ Bf1 (i) ∩ Dom(β  ) with

(4.15) (f2 |Bi )−1 (j) ∼

= {1, . . . , n}, n≥2
choose a splitting of the inclusions F1 V ⊂ · · · ⊂ Fn V. This gives us an isomorphism
(4.16) gr1 V ⊕ · · · ⊕ grn V ∼
= Fn V/F<1 V = grj W
allowing us to define ψj as direct sums of ψ1 ⊕ · · · ⊕ ψn for various i. Different choices of
splittings give isomorphic f∗ V.

Theorem 4.1. — Let X be a net, V a representation of X . Then there is a net X  of height 1,

a morphism f : X  → X , and representation V of X  such that f∗ V = V.

As a preliminary to the proof of the theorem, we discuss diagrams

A
X Y
B

of parallel surjective linear maps between finite dimensional vector spaces. First, we have
two increasing filtrations on X:
 k−1
(4.17) A−1 0 ⊂ A−1 BA−1 0 ⊂ A−1 BA−1 BA−1 0 ⊂ · · · ⊂ A−1 BA−1 0 ⊂ ···
  
=:Fk
 k−1
(4.18) B−1 0 ⊂ B−1 AB−1 0 ⊂ B−1 AB−1 AB−1 0 ⊂ · · · ⊂ B−1 AB−1 0 ⊂ ···
  
=:Gk

They necessarily stabilize with F∞ = FN , G∞ = GN for N big. In general, these filtrations

are not exhaustive, but
(4.19) F∞ = G∞ .
Furthermore, there are inverse isomorphisms

B−1 A
Mi,j Mi−1,j+1
A−1 B
 FLAT SURFACES AND STABILITY STRUCTURES 289

for i ≥ 0, j ≥ 1, where

Fi ∩ Gj
(4.20) Mi,j =
(Fi−1 ∩ Gj ) + (Fi ∩ Gj−1 )

for i, j ≥ 1. In particular,
• Ker(B) has an exhaustive filtration F ∩ Ker(B) with associated graded pieces
Mi,1 ,
• Ker(A) has an exhaustive filtration G ∩ Ker(A) with associated graded pieces
M1,j ,
• there are isomorphisms Mi,1 −→ M1,i .
We proceed with the proof of the theorem.

Proof. — Let (V, {Fi }, φi ) be a representation of a net X = (A, α, B, β). The proof
is by induction over

(4.21) dim Vi
|Bi |≥2

for all nets simultaneously. If h(X ) = 1 we are done, so let us assume that h(X ) ≥ 2. We
can also assume, by passing to subsets of A and the Bi , that all associated graded gri V are
non-zero and all Bi are non-empty.
Let r ∈ A with |Br | = h(X ). The goal is to separate a “strand” of the representation
going through r, thus decreasing the quantity (4.21). Let n = max Br and find the unique
k ∈ Bα(r) such that

(4.22) F<k + F<n  Vr , Fk + F<n = Vr .

Set X1 = F<k + F<n , X2 = Fk ∩ (F<k + F<n ) and consider the refinements of the two
filtrations of Vr :

(4.23) · · · ⊂ F<n ⊂ X1 ⊂ Fn = Vr , · · · ⊂ F<k ⊂ X2 ⊂ Fk ⊂ · · ·

Note that the identity map induces an isomorphism of associated graded

(4.24) Fn /X1 −→ Fk /X2

which follows from the general isomorphism theorem S/(S ∩ T) ∼ = (S + T)/T.

Case β(k) = n. The first step is to modify the net and its representation to de-
crease (4.21). The second step is to show that the pushforward of the modified represen-
tation gives the original one.
290 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH

• If n ∈ Dom(β), β(n) ∈ Bp , can refine the filtration on Vp to

· · · ⊂ F<β(n) ⊂ X3 ⊂ Fβ(n) ⊂ · · ·
where X3 is the preimage of φn (X1 /F<n ) under the quotient map Fβ(n) 
grβ(n) X.
• Similarly, if k ∈ Dom(β), β(k) ∈ Bp , can refine the filtration on Vp to
· · · ⊂ F<β(k) ⊂ X4 ⊂ Fβ(k) ⊂ · · ·
where X4 is the preimage of φk (X2 /F<k ) under the quotient map Fβ(k) 
grβ(k) X.
We construct a modified net X  = (A , α  , B , β  ) with
(4.25) A = A  {1, 2}, α  (1) = 2, α  |A = α
and B1 = {1}, B2 = {2}. Moreover, if n ∈ Domβ we insert an element “3” into B just
before β(n) with β  (1) = 3, and similarly if k ∈ Domβ we insert an element “4” before
β(k) with β  (2) = 4. Note that the new elements of B correspond to the additional pieces
of the filtrations described above. There is a morphism of nets f : X  → X with 1 → n,
2 → k, 3 → β(n), 4 → β(k).
V determines a representation V = (V , Fi , ψi ) of X  with
(4.26) V1 = V2 = Fn /X1 ∼
= Fk /X2 , Vr = Vα(r) = X1
where we suppress the canonical isomorphism Fn /X1 ∼ = Fk /X2 from now on. The filtra-
tions on Vr are obtained from those on Vr by restriction, and the filtrations containing
F3 , F4 by refinement. The isomorphisms ψn , ψk , ψ1 , ψ2 and their inverses are induced
by φn , φk and their inverses respectively. We claim that f∗ V is isomorphic to V. Choose
complements
(4.27) X1 /F<n ⊕ Y1 = grn V
(4.28) X2 /F<k ⊕ Y2 = grk V.
If n ∈ Domβ, then the choice of Y1 is equivalent to a choice of complement of
X3 /F<β(n) ⊂ grβ(n) V, and similarly for Y2 . We use these complements in the definition
of f∗ V .
We claim that we can find a U ⊂ Fk such that
(4.29) Fn  grn V induces an isomorphism U → Y1
(4.30) Fk  grk V induces an isomorphism U → Y2
Namely, choose a U ⊂ Fk with
(4.31) X2 ⊕ U = Fk
 FLAT SURFACES AND STABILITY STRUCTURES 291

such that (4.30) holds. We also have a direct sum X1 ⊕ U = Vr , but (4.29) may fail.
However, we can ensure (4.29) after shearing U in the directions of F<k , which does not
affect (4.30).
The choice of U gives an identification

(4.32) Vr = X1 ⊕ U = X1 ⊕ (Fn /X1 ) = Vr ⊕ V1

which we use to define the isomorphism V → f∗ V . Compatibility with filtrations follows

from U ⊂ Fk , and compatibility with φn , φk since they are block-diagonal with respect to
the chosen splittings of associated graded.
Case β(k) = n. The overall strategy is similar to the previous case. The main differ-
ence is that both Fn /X1 and Fk /X2 are now subquotients of the same vector space and
hence can have some overlap. As a consequence, we cannot just excise the subquotient
Fn /X1 ∼
= Fk /X2 as before, but must subdivide further.
We have a diagram of parallel surjections:
p1
grn V Fn /X1
∼
= φn ∼
=

p2
grk V Fk /X2

By the discussion preceding the proof we get first of all a pair of non-exhaustive filtrations

(4.33) 0 = G0 ⊂ G1 ⊂ · · · ⊂ Gm , 0 = H0 ⊂ H1 ⊂ · · · ⊂ Hm = Gm

of grn V, where G1 = Ker(p1 ) and H1 = φn−1 (Ker(p2 )). For the associated double graded
Mij , 1 ≤ i, j, i + j ≤ m + 1, we have isomorphisms Mij → Mi+1,j−1 induced by φn . Further,
we have restricted filtrations on Ker(p1 ), Ker(p2 ) lifting to refinements of the filtrations
on Vr :

(4.34) · · · ⊂ F<n = X1,0 ⊂ X1,1 ⊂ · · · ⊂ X1,m = X1 ⊂ Fn

(4.35) · · · ⊂ F<k = X2,0 ⊂ X2,1 ⊂ · · · ⊂ X2,m = X2 ⊂ Fk ⊂ · · ·

with

(4.36) X1,i /X1,i−1 ∼

= M1,i , X2,i /X2,i−1 ∼
= Mi,1
We construct a modified net X  = (A , α  , B , β  ) with

(4.37) A = A  {1, 2}  (l, i, j) | l ∈ {1, 2}, i ≥ 2, j ≥ 2, i + j ≤ m + 1

(4.38) α  (1) = 2, α  (1, i, j) = (2, i, j), α  |A = α
(4.39) B1 = {1}, B2 = {2}, Bl,i,j = {l, i, j}
292 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH
 
(4.40) Br = Br \ {n}  (1, 1, 1) < (1, 1, 2) < · · · < (1, 1, m)
 
(4.41) Bα(r) = Bα(r) \ {k}  (2, 1, 1) < (2, 2, 1) < · · · < (2, m, 1)
(4.42) β  (1) = 2, β  (1, i, j + 1) = (2, i + 1, j), i ≥ 1, j ≥ 1, i + j ≤ m

where in Br (resp. Bα(r) ) the new elements replace n (resp. k) in the total order. There is a
morphism of nets f : X  → X with 1, (1, i, j) → n, and 2, (2, i, j) → k.
Let G be the preimage of Gm = Hm under the projection Fn → grn V. V determines
a representation V = (V , Fi , ψi ) of X  with

(4.43) Vr = X1 , V1 = Fn /G, V1,i,j = Mi,j ,

the restrictions of the filtrations (4.34) to X1 , and ψ1 , ψ1,i,j induced by φn . We claim that
f∗ V is isomorphic to V. Choose a complement Y to G ⊂ Fn with Y ⊂ Fk . Further, let
Y1,1 , . . . , Y1,m ⊂ grn V with

(4.44) X1,j /F<n = Y1,1 ⊕ · · · ⊕ Y1,j

and define Yi+1,j = φn (Yi,j+1 ). We get Yij ∼

= Mij and

(4.45) G/X1 = Yij .
i,j≥2
i+j<m+1

Also we can use the X1,j and Xi,1 in the definition of f∗ V . Choose a complement Z to
X1 ⊂ G with Z ⊂ Fk , allowing us to lift Yi,j to Zi,j ⊂ Fk . Combining the various splittings
we obtain an isomorphism

(4.46) X1 ⊕ (Fn /G) ⊕ Mij → Vr
i,j≥2
i+j<m+1

which we use to identify f∗ V with V. 

Call a net as in (4.10) a cycle. Note that for any cycle and any n ≥ 2 there is a
morphism from another cycle which is n : 1, i.e. an n-fold “covering”. These, and iso-
morphisms, are the only morphisms between connected nets of height 1. We use this to
formulate a strengthening of the previous theorem.

Theorem 4.2. — Let X be a net, V an indecomposable representation of X . Then there exists

a connected net X  of height 1, a morphism f : X  → X which does not factor through a covering, and
an indecomposable representation V of X  such that f∗ V = V. Moreover, for any triple X  , f  , V
with these properties there is an isomorphism of nets g : X  → X  over X such that g∗ V ∼
= V .
 FLAT SURFACES AND STABILITY STRUCTURES 293

Proof. — Existence follows from Theorem 4.1, which gives us X  , f , V , with

f∗ V = V. Since V is indecomposable, V must be as well, and can only be supported
on a single component of X  so we can take X  to be connected. If f factors through
some covering, we just push V forward along it.
For uniqueness, consider the inductive procedure in the proof of Theorem 4.1.
After choosing a total order on A, the output X  , f , V depends, up to isomorphism, only
on the isomorphism class of V. Moreover, if V is of the form f∗ V for some connected net
of height 1, f : X  → X not factoring through a covering, and V indecomposable, then
the output of the procedure applied to V is isomorphic to the same (X  , f , V ). 

In applications of this theorem below it will be natural to consider nets which

are not finite (only individual Bi are). Still, all our results extend to this case, as finite
dimensional representations are supported on a finite subnet.

4.3. Minimal twisted complexes

Let A be a graded linear category which is augmented in the sense that there are
splittings
(4.47) A(X, Y) = Ae (X, Y) ⊕ Ar (X, Y)
such that Ae (X, Y) is K1X for X = Y and zero otherwise, and Ar (X, Y) are closed under
composition. We view Ar as a non-unital category with the same objects as A. Assume
for the rest of the subsection that Ar is nilpotent in the sense that there exists an integer
N > 0 such that any composition of N morphisms vanishes. Recall that a functor F is said
to reflect isomorphisms if f is invertible whenever F(f ) is.

Lemma 4.2. — The functor T : addZ(A) → addZ(Ae ) induced by the augmentation reflects
isomorphisms.

Proof. — Let M, N ∈ addZ(A) and φ : M → N such that T(φ) = φ is an iso-

morphism. Thus we have graded vector spaces M(X), N(X) for each X ∈ Ob(A), and
components of φ
     
φX ∈ Hom M(X), N(X) ⊕ Hom M(X), N(X) ⊗ Ar (X, X)
 
φX,Y ∈ Hom M(X), N(Y) ⊗ Ar (X, Y), X = Y

Let φ X be the component of φX in Hom(M(X), N(X)). By assumption, the φ X are iso-

−1 −1
morphisms. Composing φ with the morphism φ with components φ X , we may assume
that M(X) = N(X) and all φ X are identity morphisms. Thus,
(4.48) φ = 1 −  ∈ End0 (M), N = 0
which clearly has an inverse. 
294 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH

A two-sided twisted complex over A is given by a pair (M, δ) with M ∈

Ob(addZA) and δ ∈ End1 (M) with δ 2 = 0. We say that (M, δ) is minimal if the im-
age δ of δ under the functor addZA → addZAe vanishes, i.e. δ has components in (the
degree 1 part of)
 
(4.49) Hom M(X), M(Y) ⊗ Ar (X, Y).

Proposition 4.1.
1. Every twisted complex A ∈ TwA is isomorphic to a direct sum A = Am ⊕ Ac with Am
minimal and Ac contractible.
2. Any homotopy equivalence between minimal twisted complexes is an isomorphism.

Proof. — 1. Let (M, δ) be a twisted complex over A, (M, δ) its image in Tw(Ae ). By
semisimplicity of Ae , (M, δ) is isomorphic to a direct sum Bc ⊕ Bm where Bc is contractible
and Bm has trivial differential. Therefore, (M, δ) is isomorphic to a twisted complex of
the form
⎛ ⎛ ⎞⎞
δ11 δ12 δ13
⎝K ⊕ K[−1] ⊕ L, ⎝ 1K + δ21 δ22 δ23 ⎠⎠
δ31 δ32 δ33

with δ ij = 0. Using elementary row and column operations (automorphisms of K ⊕

K[−1] ⊕ L) one reduces the matrix to the form
⎛ ⎞
0 0 0
⎝ 1K 0 0 ⎠
0 0 δL

with δ L = 0, thus providing the direct sum decomposition.

2. Suppose (M, δ M ) and (N, δ N ) are minimal twisted complexes, f : M → N and
g : N → M closed morphisms of degree 0, G : M → M and H : N → N morphisms of
degree −1 such that

(4.50) 1M − gf = δ M G + Gδ M , 1N − fg = δ N H + Hδ N .

Using minimality, gf = 1M and fg = 1N , i.e. f , g are inverse isomorphisms. Hence, the

claim follows from the previous lemma. 

4.4. Classification of objects

We are now ready to state and prove the main result of this section.
 FLAT SURFACES AND STABILITY STRUCTURES 295

Theorem 4.3. — Let S be a compact graded marked surface, then the construction of Section 4.1
sets up a bijection between isomorphism classes of indecomposable objects in H0 (F (S)) and isotopy
classes of admissible curves with indecomposable local system.

Remark. — In the special case when H0 (F (S)) is the bounded derived category
of a gentle algebra the classification of indecomposable objects was found by Drozd–
Burban [16] and Bekkert–Merklen [8]. To classify objects in F (S) in general we need to
consider Z-graded algebras and the language of twisted complexes instead of projective
resolutions. Consider for example the Dynkin quiver with two parallel arrows which do
not have the same degree.

Proof. — Let Q be a graded quiver with quadratic monomial relations of type

F1, i.e. associated with some systems of arcs A on a compact graded marked surface
(S, M, ). The graded linear category KQ has a natural augmentation with (KQ)r gen-
erated by non-identity paths. Assume that sufficiently long paths are zero in KQ, equiv-
alently that no components of M are diffeomorphic to S1 , so that (KQ)r is nilpotent and
we can apply the results of the previous subsection.
Consider first the case when Q is of the form
(4.51) • ←− • ←− · · · ←− •
without relations and arbitrary grading. Let us number the vertices from left to right by
{1, . . . , n} and let αi denote the arrow from i +1 to i. A minimal twisted complex over KQ
is then given by finite-dimensional graded vector spaces V1 , . . . , Vn , and for 1 ≤ i < j ≤ n
a linear map
(4.52) δij : Vj → Vi , deg(δij ) = deg(αi ) + · · · + deg(αj−1 ) − 1
such that the matrix δ = (δij ) has square zero.
Consider the direct sum
 n
(4.53) V= Vi ,
i=1

which has the filtration

(4.54) V1 ⊂ V1 ⊕ V2 ⊂ · · · ⊂ V1 ⊕ · · · ⊕ Vn .
The endomorphism δ gives a three-step filtration
(4.55) Imδ ⊂ Kerδ ⊂ V
and an isomorphism of associated graded V/Kerδ ∼ = Imδ. We find that we have a repre-
sentation of the net
   
(4.56) {1, 2} × Z, (12) × id, {1, . . . , n}  {1, 2, 3} × Z, (13) × id
296 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH

with total order so that (i, k + d(αi ) − 1) < (i + 1, k) on {1, . . . , n} × Z and (1, n) <
(2, n) < (3, n) on {1, 2, 3} × Z. In fact, we get an embedding of the category of minimal
twisted complexes into the category of representations of the above net as a full subcate-
gory. We do not get an equivalence of categories, since we only consider δ which decrease
the filtration (4.54).
Turn now to the case of general Q. Let D be the set of maximal non-zero paths in
KQ. Then for each element of D there is a sub-quiver of Q which is of the simple form
above, and Q is obtained from their disjoint union by identifying some pairs of vertices.
The corresponding net is now given by

(4.57) A = D × {1, 2} × Z
(4.58) α = id × (12) × id
  
(4.59) B = (v, d) ∈ Q0 × D | v on d  D × {1, 2, 3} × Z =: Bg  Ba

with partial order defined so that for an arrow α in d ∈ D we have (∂1 (α), n + d(α) − 1) <
(∂0 (α), n) and (d, 1, n) < (d, 2, n) < (d, 3, n). Further,
   
(4.60) β = τ  id × (13) × id

where τ sends a pair (v, d1 ) ∈ Q0 × D with v on d1 , d2 ∈ D, d1 = d2 to (v, d2 ). The

definitions are set up so that the following proposition holds. The proof is just chasing
through the definitions.

Proposition 4.2. — There is a one-to-one correspondence between isotopy classes of admissible

curves on S with local system on the one hand, and morphisms from nets of height 1 with indecomposable
representation to the net (A, α, B, β) constructed above on the other. Under this correspondence, nets of
the type (4.9) (resp. (4.10)) correspond to admissible curves with domain [0, 1] (resp. S1 ).

There is a fully-faithful functor from the category of minimal twisted complexes

over Q to the category of representations of (A, α, B, β). Theorem 4.2 gives us a clas-
sification of indecomposable minimal twisted complexes over KQ, equivalently isomor-
phism classes in H0 (F (S)) by Proposition 4.1.
It remains to deal with the case when M has components diffeomorphic to S1 .
Let S be the surface obtained by adding two corners and a boundary arc on each
such component of M. Then H0 (F (S)) is a localization of the triangulated category
H0 (F (S )), and the localization functor is essentially surjective. Moreover, if the images
of E, F ∈ H0 (F (S)) corresponding to admissible curves become isomorphic under local-
ization, then E and F differ by extensions by boundary arcs, so map to isotopic curves
in S. 
 FLAT SURFACES AND STABILITY STRUCTURES 297

5. Comparison of moduli spaces

This section contains our main result, identifying M(S) with an open and closed
subset of Stab(F (S)). We begin by reviewing Bridgeland’s axioms in Section 5.1. In
Section 5.2 we compute K0 (F (S)) in terms of singular homology. The main theorems
are stated in Section 5.3 and proven in the remaining subsections. Sections 5.4 and 5.5
deal with the support property and Harder–Narasimhan filtrations, respectively. In the
final Section 5.6 we prove that the map on moduli spaces is complex bianalytic onto an
open and closed subset.

5.1. Stability structures on categories

Fix a triangulated category C and a homomorphism cl : K0 (C ) →  to a finitely
generated abelian group . A stability structure (cf. [11], [31], also called a stability condition)
on C is given by
• for each φ ∈ R a full additive subcategory C φ ⊂ C of semistable objects of phase φ,
and
• an additive map Z :  → C, the central charge.
This data has to satisfy the following axioms.
1. C φ [1] = C φ+1
2. If E ∈ C φ1 , F ∈ C φ2 , φ1 > φ2 , then Hom(E, F) = 0.
3. Every E ∈ C has a Harder–Narasimhan filtration: A tower of triangles

0 = E0 E1 ··· En−1 En = E

A1 An

with 0 = Ai ∈ C φi and φ1 > φ2 > · · · > φn . The Ai are called the semistable com-
ponents of E.
4. If 0 = E ∈ C φ then Z(E) := Z(cl([E])) ∈ R>0 eπiφ .
5. The support property: For some norm . on  ⊗R and C > 0 we have an estimate
(5.1) cl(E) ≤ C|Z(E)|
for E ∈ C φ .
The set Stab(C , ) of all stability structures has a natural topology which is induced by
the metric
  
 mσ2 (E) 
(5.2) d(σ1 , σ2 ) = sup |φσ2 (E) − φσ1 (E)|, |φσ2 (E) − φσ1 (E)|, log
− − + + 
0=E∈C m (E) 
σ1
298 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH

cf. [11]. Here, for an object E ∈ C  with semistable components A1 , . . . , An with phases
φ1 > · · · > φn one defines m(E) = |Z(Ai )|, φ − (E) = φn , and φ + (E) = φ1 . The main
result of [11] (see also [31]) is that the projection

Stab(C , ) → Hom(, C)

is a local homeomorphism, hence Stab(C , ) has the structure of a complex manifold of

dimension rk(), if it is non-empty.

5.2. Charge lattice

Recall that a graded surface has a double cover τ given by the orientations of
the foliation lines, and Zτ = Z ⊗Z/2 τ is the associated local system of abelian groups.
The following theorem describes K0 of the Fukaya category in terms of homology with
coefficients in Zτ .

Theorem 5.1. — Let S = (S, M, η) be a compact graded marked surface, then there is a
natural isomorphism of abelian groups
 
K0 F (S) ∼ = H1 (S, M; Zτ ).

Proof. — Choose
√ a full system A of graded arcs on S. Each arc defines a class in
H = H1 (S, M; Z ). From cellular homology it follows that H is the group generated
by these arcs and with relations of the form ±X1 ± · · · ± Xn = 0 where X1 , . . . , Xn are
the arcs bounding a disk cut out by A. Correspondingly, it is also clear that the arcs in A
generate
    
(5.3) K := K0 F (S) = K0 Tw FA (S)

and satisfy the same relations, which follows from the observation in Section 3.3. By
construction we get a surjective homomorphism H → K which is independent of the
choice of arcs and natural with respect to maps of graded marked surfaces. It remains to
show that no additional relations are needed to present K.
Suppose first that M has no components diffeomorphic to S1 . Choosing a formal
collection A of arcs, we can present F (S) as the category Tw(KQ) of twisted complexes
over the path category of a graded quiver Q with quadratic monomial relations. Note
that H is freely generated by A, so we need to verify that K is freely generated by the
vertices Q0 of Q. One way to see this is by using the explicit resolution of the diagonal
of KQ, showing that, as a bimodule, KQ is a repeated cone of Yoneda bimodules. As a
consequence, Tw(KQ) ∼ = Perf(KQ), and an inverse of the map
 
(5.4) ZQ0 → K0 Perf(KQ)
 FLAT SURFACES AND STABILITY STRUCTURES 299

sending a vertex to the corresponding simple module, is induced by the dimension vector
of a module.
For a general graded marked surface S there is a S obtained by adding two corners
to each component of ∂S diffeomorphic to S1 . Let N denote set of boundary arcs of S cre-
ated in this process, and N the corresponding full triangulated subcategory of H0 (F (S ))
generated by N. Each arc in N has endomorphism algebra of the form K[x]/x2 with
|x| ∈ Z, so N is a product of categories of the form H0 (Tw(K[x]/x2 )). We get a commu-
tative diagram

 √  √
(5.5) ZN H1 S , M , Z  H1 (S, M, Z ) 0

    
K0 ( N ) K 0 F S K0 F (S) 0

where we use Proposition 3.6 to get the bottom row. We claim that the rows are exact.
For the top row this is just the exact sequence of a triple. For the bottom row this uses the
fact that N is idempotent complete, thus a thick subcategory so that Proposition 3.1 in
[24] can be applied. Since the left and middle horizontal maps are isomorphisms, so is
the right one. 

When a flat surface S comes from a quadratic differential with higher order poles,
the category F (S) includes some objects which have infinite length since
 1
(5.6) x−n/2 dx = ∞
0

for n ≥ 2. We need to pass to the subcategory of F (S) which does not include these
objects. Formally, we want to allow S = (S, M) to have boundary circles which are
unmarked, i.e. not belonging to M. Thus, define F (S, M) as the full subcategory of
F (S, M ) of objects corresponding to immersed curves avoiding M \ M, where M is
the union of M and all boundary circles.

Example. — Consider the case when S is an annulus. By homological mirror sym-

metry, F (S) ∼ = Db Coh(Gm ). Objects in Coh(Gm ) are direct sums of line bundles and
torsion sheaves. The relevant subcategory is just the one of torsion sheaves. Coinciden-
tally, “finite-length” in the sense of abelian categories corresponds in this case to finite
length with respect to the flat metric.

Lemma 5.1. — The full subcategory F (S, M) ⊂ F (S, M ) is triangulated, i.e. closed under
shifts and cones.
300 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH

Proof. — For each component B of the boundary we will introduce a functor IB

from F (S, M ) to the category of complexes of vector spaces (possibly unbounded), such
that E ∈ F (S, M) if and only if IB (E) = 0 for all components B of M which are not in M,
which implies the claim. Roughly speaking, IB is Hom(B, _), but in general the curve B
does not have a grading and so is not an object of the Fukaya category as defined here.
As a workaround we define this functor explicitly by hand.
Fix a boundary component B ⊂ ∂S. The grading and vectors tangent to the
boundary define sections s1 , s2 of P(TS|B ). If B ∼ = S1 , then since S is oriented, B and
P(TS|B ) carry natural orientations, and so the intersection number, m, between s1 and s2
is well defined. If B ∼
= R we take m = 0. Let N = |m|. If N = 0, then B has a grading in
the sense of Section 2.1, and if N > 0 then B still has a Z/N-grading: A lift of s2 to the
fiberwise N-fold covering of P(TS|B ). In either case, choose an appropriate grading on B.
Then for every graded curve A ending on B transversely, we get an intersection number
i(B, A) ∈ Z/N.
The cohomology of IB (E) will be bounded if N = 0 and N-periodic if N > 0. It
suffices to define IB on generators of F (S, M ). Let A be an arc in S. If both endpoints of
A do not lie on B, then IB (A) = 0. If one endpoint of A lies on B, then

(5.7) IB (A) = K[−j].
j∈Z
j≡i(B,A)

Finally, if both endpoints of A lie on B, then IB (A) is a sum of two graded vector spaces
as above. The map on morphisms is defined in the obvious way so that IB (a) = 0 if a is a
boundary path which is not in B. One checks that this gives a well defined A∞ -functor. 

Proposition 5.1. — Let S be a compact graded marked surface which is allowed to have un-
marked boundary circles. Then there is a natural map
 
(5.8) K0 F (S) −→ H1 (S, M; Zτ ).

The group K0 (F (S)) is often infinitely generated, and so the above map will fail
to be injective. It may also fail to be surjective, e.g. for the cylinder with non-standard
grading, where F (S) is trivial if both boundary circles are unmarked.

Proof. — Let S be the graded marked surface obtained from S by replacing each
unmarked boundary circle with one marked and one unmarked boundary arc. Then
F (S) is also a subcategory of F (S ), as follows from localization. Note that
    
(5.9) ι : H1 (S, M; Zτ ) −→ H1 S , M ; Zτ ∼ = K 0 F S

is an inclusion and the image of ι contains the image of K0 (F (S)) −→ K0 (F (S )). Hence
we have the desired natural map. 
 FLAT SURFACES AND STABILITY STRUCTURES 301

5.3. Statement of main theorem

Let X be the real blow-up of a flat surface X of finite type. It contains a compact
surface with marked boundary S ⊂ X , canonical up to isotopy, as a deformation retract.
The complement X  \ S is a disjoint union of punctured open disks, one for each higher
order pole of the quadratic differential, and rectangles R>0 × [0, 1], one for each infinite-
angle singularity (more precisely, non-cyclic boundary walk). The marked boundary of S
 Each removed punctured open disk is bounded by an unmarked
is the part of ∂S in ∂ X.
boundary circle, while each removed rectangle is bounded on one end by an unmarked
part of ∂S. Thus, the Fukaya category F (X) = F (S) is defined. We also note that the
inclusion of pairs (S, M) ⊂ (X , ∂ X
) is a homotopy equivalence, so both give isomorphic
relative homologies.

Theorem 5.2. — Let X be a flat surface of finite type. Define subcategories C φ ⊂ H0 (F (X))
so that isomorphism classes of indecomposable objects correspond to unbroken graded geodesics of phase φ
with indecomposable local system, and let
 
(5.10) Z : K0 F (S) −→ H1 (X, ∂X; Zτ ) =:  → C

be the period map. Then this data satisfies the axioms of a stability structure. In particular, for a graded
surface of finite type X there is a continuous map
 
(5.11) M(X) −→ Stab F (X)

over Hom(, C).

Remark. — Let us emphasize that we put no restrictions on the ground field K of

F (S). In fact this category is defined over any ring, but already Perf(Z) does not admit
stability structures.

Theorem 5.3 (Main theorem). — Let X be a graded surface of finite type, then the map
M(X) → Stab(F (X)) is injective and its image open and closed, thus a union of components of
Stab(F (X)).

Remark. — When C is compact, then necessarily g(C) = 1. In this case one can
use homological mirror symmetry for elliptic curves [41] and Atiyah’s classification [4]
to derive a version of Theorem 5.2. The computation of Stab(Db (E)) in [11] shows that
M(X) = Stab(X).
In the case where C is a punctured torus, the mirror is a nodal elliptic curve E.
The space of stability conditions on Db (E) was computed by Burban–Kreußler [17].
Our theorem gives a different description of the same space.

The proofs of Theorem 5.2 and Theorem 5.3 will occupy the rest of this section.
302 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH

5.4. Support property

 its real blow-up. We have as before
Let X be a flat surface of finite type and Y = X
(5.12)  = H1 (Y, ∂Y; Zτ ), R =  ⊗Z R = H1 (Y, ∂Y; Rτ )

Proposition 5.2. — There is a norm . on R , C > 0 such that for γ ∈  the class of a
finite geodesic on X there is an estimate
(5.13) γ  ≤ C|Z(γ )|
where Z :  → C is the period map.

Proof. — The locally constant sheaf Rτ is given by flat sections of a flat line bundle
with metric for which we use the same notation. We have
(5.14) R∗ = H1 (Y, ∂Y; Rτ )∗
(5.15) = H1 (Y, ∂Y; Rτ )
(5.16) = H1dR (Y, ∂Y; Rτ )
 be the closure of
Choose forms ω1 , . . . , ωn representing a basis of H1dR (Y, ∂Y; Rτ ). Let K
the preimage of K = Core(X) under the inclusion Y \ ∂Y → X. Define a norm on R by

(5.17) γ  = |[ωi ](γ )|

i

and set
(5.18) C= ωi |K ∞
i

where the norm is taken with respect to the flat metric. Let α be a finite geodesic on X
with class γ ∈ , then
(5.19) l(α) = |Z(γ )|
where l(α) is the length. Thus

(5.20) γ  = |[ωi ](γ )|

i
 
 
(5.21) =  ωi 
 
i α

(5.22) ≤ l(α)ωi |K ∞

i

(5.23) = C|Z(γ )|. 

 FLAT SURFACES AND STABILITY STRUCTURES 303

5.5. Harder–Narasimhan filtrations

We continue with the proof of Theorem 5.2. The strategy is to first deal with the
special case of a disk and annulus, then deduce the general case by passing to the universal
and annular coverings. The idea for the proof of the Harder–Narasimhan property is to
use the fact that arbitrary indecomposable objects are represented by immersed curves
which we can deform to their geodesic representative: a closed loop or concatenation
of saddle connections. Each semistable component of the HN filtration corresponds to
several such finite length geodesics.
Let X be a flat surface with Xsm diffeomorphic to a disk and all conical singularities
with infinite angle. The corresponding graded marked surface S is homeomorphic to
a closed disk with n + 1 = |Xsg | marked and unmarked boundary arcs. Consider the
data of a stability structure on F (X) as in the statement of theorem. We first check that
Hom(A, B) = 0 for semistable objects A, B = 0 with φ(A) > φ(B). Note that, by the long
exact sequence for Hom, it suffices to do this for A, B stable, so that they correspond to
graded arcs. We consider the various cases.
1. If A and B differ only by a grading shift, the claim follows from Ext<0 (A, A) = 0.
2. If A and B are disjoint, then Ext(A, B) = 0.
3. If A and B intersect in a smooth point p of the surface, then Ext(A, B) is con-
centrated in degree
 
(5.24) ip (A, B) = φ(A) − φ(B)
so φ(A) > φ(B) implies Ext0 (A, B) = 0. This is clear from the alternative defi-
nition of the Fukaya category where morphisms are given by intersection points,
and can be verified here by direct computation for the Fukaya category of a disk
with four marked boundary components.
4. If A and B meet in an ∞-angle singularity p in an angle φ > 0, then
φ
(5.25) ip (A, B) = φ(A) − φ(B) +
π
and Ext(A, B) is either concentrated in degree ip (A, B) or zero, depending on
the order in which they meet p.
Next we check the HN-property. It suffices to do this for indecomposable objects E.
By the classification, E corresponds to a graded arc α connecting two ∞-angle singu-
larities. A geodesic representative of α is a concatenation of graded saddle connections
α1 , . . . , αk corresponding to objects A1 , . . . , Ak in F (X). Let φ1 > · · · > φl be the distinct
phases of the αi (generically, l = k). For 1 ≤ i ≤ l there is a semistable Bi given by a twisted
complex
⎛ ⎞

(5.26) Bi = ⎝ Aj , δi ⎠
φ(Aj )=φi
304 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH

where δi has non-zero coefficients corresponding to the singular points where the geodesic
representative of α passes through without changing phase, which is equal to φi . Also,

(5.27) E= Aj , δ
j

where δ has non-zero coefficients corresponding to all the singular points the geodesic
representative passes through (i.e. where αi meets αi+1 ). Note that this really gives mor-
phisms of degree 1 since the concatenation of the saddle connections smooths to a graded
curve. To show that the twisted complex representation of E in terms of the Bi is a HN-
tower, we need to check that components of δ increase phase. For this we use that fact
that the concatenation of the αi is a geodesic, and so αi , αi+1 necessarily meet at a singular
point in an angle φ ≥ π , thus φ > π if the phase jumps. The corresponding morphism,
without loss of generality from Ai to Ai+1 , has degree 1, and so
φ
(5.28) 1 = φ(Ai ) − φ(Ai+1 ) +
π
which implies φ(Ai ) < φ(Ai+1 ).
Suppose now that X is a general flat surface of finite type. The first claim is that
Hom(A, B) = 0 for a pair of graded saddle connections A, B with φ(A) > φ(B). To see
this, let Y be the flat surface so that its real blow-up   Note
Y is the universal cover of X.
that all singularities of Y have infinite cone angle. Lift A, B to graded curves  A, 
B on Y
with the same phases. We have

(5.29) Hom(A, B) = Hom( A, g
B)
)
g∈π1 (X

and hence it suffices to prove the claim for Y. But any two saddle connections in Y are
objects in some subcategory of F (Y) which is the Fukaya category of some surface of
disk-type as above, and so our previous arguments can be applied.
Second, we claim that the HN-property holds for admissible curves c with domain
[0, 1]. To see this, lift c to a graded curve c̃ in the universal cover Y. We work in the
category generated by all the saddle connections in a geodesic representative of c̃. It
corresponds to some surface of disk-type in the universal cover. We get a HN-tower for c̃
in this category and push it forward to F (X) to get a tower for c in that category.
We turn to the case when X is of annular type, i.e. Xsm is diffeomorphic to an an-
nulus with grading such that a simple closed loop around it is gradable, and all singular-
ities have infinite cone angles. Let S ⊂ X be the corresponding compact graded marked
surface. The category F (X) is the bounded derived category of finite-dimensional repre-
sentations of a quiver which is a cyclic chain of arrows, all in degree zero. The number of
arrows oriented in one or the other way is equal to the number of marked boundary arcs
 FLAT SURFACES AND STABILITY STRUCTURES 305

on the two components of ∂S, respectively. In arguments below we omit straightforward

computations of Ext∗ (A, B) in F (X).
First we check that Hom(A, B) = 0 for stable A, B with φ(A) > φ(B). It remains
to consider the cases when either A or B, or both, are closed geodesics with simple local
system.
1. If A and B both correspond to loops with simple local system, then they either
differ by a shift or else are orthogonal, i.e. Ext∗ (A, B) = Ext∗ (B, A) = 0. In
the first case it suffices to note that Ext<0 (A, A) = 0 since A lies in an abelian
category.
2. If one of A or B is a geodesic loop and the other a saddle connection tending in
both directions to the same component of ∂S, then A and B are orthogonal.
3. If one of A or B is a geodesic loop and the other a saddle connection from one
component of ∂S to the other, then A and B intersect in a unique point p and
Ext∗ (A, B) is concentrated in degree ip (A, B) = φ(A) − φ(B) > 0.
Recall that an annular covering of a surface is a covering corresponding to a subgroup of
π1 generated by a simple closed loop (cf. [47]). By passing to the annular covering (of X) 
we see, as in the case of two saddle connections, that the statement is true for general X
when at least one of A or B is a saddle connection.
For annular X there is, up to isotopy and grading, a unique admissible curve c
in X with domain S1 . We need to check that HN-filtrations exist for the corresponding
indecomposable objects X. If X has a geodesic loop (necessarily isotopic to c), then c
is semistable, so we are done. If not, then c has a geodesic representative which is a
concatenation of saddle connections. The corresponding collection of arcs satisfies the
properties we required when assigning a twisted complex to c in Section 3.1. Note that in
this case X cannot contain a semi-infinite flat cylinder, so S has no unmarked boundary
circles. The same argument as in the disk case shows that this twisted complex gives a
HN-filtration for X. Again, passing to the annular covering, we find that HN-filtration
exist for objects corresponding to admissible curves with local system and domain S1 on
general X. Since all isomorphism classes of indecomposable objects in H0 (F (X)) come
from admissible curves, this completes the proof of the HN-property.
Let X now be a general flat surface of finite type. It remains to show that for
two objects A, B ∈ F (X) corresponding to closed geodesics α, β with local systems and
φ(A) > φ(B) we have Hom(A, B) = 0. We may also assume that α and β are non-
isotopic as simple closed curves. Let Z denote the flat cylinder foliated by closed geodesics
isotopic to α. If Z has infinite height then α and β cannot intersect at all, so assume Z has
finite height with singular points on each boundary component, see Figure 6. The other
loop β can cross Z any number of times, but does so always with the same slope. Then we
can represent B as an extension of graded saddle connections with local systems Si , which
are contained in Z and connect the two components of its boundary, and arcs Ei which
are disjoint from the interior of Z. Moreover, these saddle connections can be chosen so
306 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH

FIG. 6. — The cylinder Z, cut along a dotted line, closed geodesics α, β, saddle connections Si , arcs Ei

that φ(Si ) < φ(B). Namely, there is an isotopy moving β to the concatenation of the Si
and Ei , and it should move any point of intersection with ∂Z so that Z lies to the right.
From what we have shown in the annular case, it follows that Hom(A, Si ) = 0 for all i.
Further, A is orthogonal to any Ei , so Hom(A, B) = 0 follows. This completes the proof
of Theorem 5.2.

5.6. The map M(X) → Stab(F (X))

We consider a construction for general categories with a stability structure. Let C
be a triangulated category with stability structure σ , and F : C → D(Mod(K)) an exact
functor to the derived category of chain complexes over a field K. We combine the two
to define
  
(5.30) CF,σ = [log |Z(E)|, +∞) × π φ(E) + k ⊂ R2
E∈C ss , k∈Z
Hk (F(E))=0

which is a union of horizontal rays starting at points of the form log(Z(E)), E semistable.
The support property ensures that the central charges Z(E) ∈ C, E semistable, form a
discrete subset, hence CF,σ ⊂ R2 is closed.

Lemma 5.2. — For σ1 , σ2 ∈ Stab(C ), C and F as above, we get

(5.31) dH (CF,σ1 , CF,σ2 ) ≤ d(σ1 , σ2 )

where dH is the Hausdorff distance on closed subsets induced by the max-metric on R2 .

Proof. — Let  = d(σ1 , σ2 ), then by definition

 −   + 
(5.32) φ (E) − φ − (E) ≤ , φ (E) − φ + (E) ≤ 
σ1 σ2 σ1 σ2
    
(5.33) log mσ (E) − log mσ (E)  ≤ 
1 2
 FLAT SURFACES AND STABILITY STRUCTURES 307

for all 0 = E ∈ C . For a subset A of a metric space let B (A) be the set of points with dis-
tance ≤  to A. To show that CF,σ2 ⊆ B (CF,σ1 ) it suffices to find, for every σ2 -semistable
object E with Hk (F(E)) = 0, a σ1 -semistable object E with Hk (F(E )) = 0 and
   
(5.34) |φσ2 (E) − φσ1 E | ≤ , log |Zσ1 E | ≤ log |Zσ2 (E)| + .

Let A1 , . . . , An be the σ1 -semistable components of E, n ≥ 1, then by exactness of F there

is an m with Hk (F(Am )) = 0, and we can take E = Am .
By symmetry we also have CF,σ1 ⊆ B (CF,σ2 ), thus dH (CF,σ1 , CF,σ2 ) ≤ . 

Proof of Theorem 5.3. — 1. Let B ⊂ ∂X be a connected component of the boundary.

As in the proof of Lemma 5.1 we choose a grading on B to get a functor IB : F (X) →
D(Mod(K)). Recall that if B is gradable in the usual sense, then IB (E) is perfect for
every E ∈ F (S). On the other hand, if B is only Z/n-gradable, n maximal, then IB (E) is
n-periodic for every E ∈ F (S).
Suppose X has a flat structure with corresponding stability structure σ ∈
Stab(F (X)). If B is a boundary component corresponding to a conical point s, then
geometrically CIB ,σ is just the complement of the domain of the exponential map at s,
Dom(exps ), in logarithmic coordinates. The Voronoi cell, Vs , around s is contained in
Dom(exps ), and in fact Vs is determined by Dom(exps ) alone, since by definition x ∈ Vs if
and only if the disk centered at x and with s on the boundary is contained in the domain.
This shows that the flat structure is recovered from σ , proving injectivity.
2. Let (Fi ) be a sequence of flat structures on X so that the corresponding se-
quence of stability structure (σi ) converges to σ ∈ Stab(F (X)). We want to show that the
sequence (Fi )i converges in M(X). The idea is to show that the corresponding sequence
of Voronoi diagrams eventually stabilizes, so we can then just take the limits of the lengths
of the edges.
Let β(Fi ) be the maximal length of a boundary edge of Core(Fi ). The first step
is to show that β(Fi ) ≤ β for some β > 0. Let E ∈ F (X) be the direct sum of objects
corresponding to all the boundary edges of Core(F1 ) with arbitrary fixed grading. An
edge should appear twice in the sum if it does not bound the interior of Core(F1 ). The
mass mσ1 (E) is then just the circumference of Core(F1 ). Since σi → σ , we have an upper
bound on mσi (E) ≥ β(Fi ), and the claim follows.
The second step is to show that ρ(Fi ) is bounded. Fix a conical point s of S. We
will show that d(s, x) is bounded for x ∈ Vs,i ∩ Core(Fi ), where Vs,i is the Voronoi cell
around s in Fi , and since
 
(5.35) ρ(Fi ) = max dH {s}, Vs,i ∩ Core(Fi )
s∈(Fi )sg

this will imply the claim. Consider the functor I : F (X) → D(Mod(K)) associated with
s as before. We have closed subsets Ci = CI,σi ⊂ R2 converging to C = CI,σ in the Haus-
dorff topology. Let D denote the set of directions (of geodesics) from s. The choice of
308 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH

grading of marked boundary as above gives an identification D ∼ = R/nπ when s has cone
∼
angle nπ and D = R when s has infinite cone angle. By the support of Ci , Supp(Ci ),
we mean the closure of its projection to D . It follows from compactness of Core(Fi )
that each Ci is compactly supported. By convergence, there is compact J ⊂ D contain-
ing all Supp(Ci ) and Supp(C). Let ri (φ) ∈ R≥0 ∪ {+∞} be the length of the maximal
geodesic starting at s in direction φ ∈ D and which is entirely contained in Vs,i . Note
that the functions ri are determined by Ci , thus σi alone, are continuous, and converge
pointwise to r : D → R≥0 ∪ {+∞} determined by σ . Define qi (φ) ∈ R≥0 similarly as the
length of the maximal geodesic starting at s in direction φ ∈ D and which is entirely con-
tained in Vs,i ∩ Core(Fi ). We need to shows that qi are uniformly bounded. By definition,
qi (φ) ≤ ri (φ), and
(5.36) (φ − π, φ + π ) ∩ J = ∅ =⇒ qi (φ) = 0.
Suppose, for contradiction, that there is a sequence φi ∈ D with qi (φi ) → ∞. By (5.36)
the φi are contained in a compact subset of D and we may assume φi → φ after passing
to a subsequence. Now since qi (φi ) → ∞ we must have r(φ) = ∞, which means that the
interval (φ − π, φ + π ) is disjoint from Supp(C). Thus also
% &
π π
(5.37) φ − ,φ + ∩ Supp(Ci ) = ∅
2 2
for sufficiently large i. The sector bounded by directions φ − π/2 and φ + π/2 in
Dom(exps ) cannot map entirely to Core(Fi ), as this contradicts ρ(Fi ) < ∞. Hence
the part of the sector which maps to Core(Fi ) is a triangle  or just the point s. The
length of the side of the triangle opposite the vertex s is bounded above by β. Thus also
qi |(φ − π/2, φ + π/2) has an upper bound in terms of β, which is independent of i, a
contradiction.
We have established that ρ(Fi ) and β(Fi ) have upper bounds independent of i. By
Lemma 2.2 the length the Delaunay edges on all the Fi is bounded above. Thus, there
are only a finite number of polygonal subdivisions up to isotopy arising as Delaunay
partitions in the sequence Fi . Passing to a subsequence, we may hence assume that the
Delaunay subdivisions for the Fi are all isotopic, i.e. combinatorially equivalent. Since
σi → σ , lengths of edges converge to positive numbers as i → ∞. It may still happen
that some polygons become degenerate in the limit, i.e. their area goes to zero. If we glue
back together the set of non-degenerate limiting polygons, we get a flat structure F which
is the limit of the Fi and corresponds to σ . 

6. Cluster-like structures

6.1. S-graphs
In this section we specialize to the case of flat surfaces with infinite area. It turns
out that for generic choice of horizontal direction the heart of the t-structure of the cor-
 FLAT SURFACES AND STABILITY STRUCTURES 309

responding stability structure is Artinian with a finite number of simple objects and de-
scribed by a certain type of finite graph. Tilting of the t-structure leads to mutation of the
graph.
Suppose X is a flat surface of finite type with infinite area. Recall from Section 2.4
that, for generic choice of horizontal direction, the leaves converging to the conical singu-
larities cut the surface into horizontal strips of finite or infinite height. The combinatorial
structure of this horizontal strip decomposition can be encoded in terms of a finite graph,
G, with set of vertices Xsg and set of edges the horizontal strips of finite height. An edge
is attached to the singularities which lie on the boundary of the corresponding horizontal
strip. We also need to record two additional pieces of data:
1. For each v ∈ Xsg with finite (resp. infinite) cone angle a cyclic (resp. total) order
on the set Hv of half-edges meeting that vertex given by the counter-clockwise
order of horizontal strips around s.
2. For each pair of successive half-edges a < b ∈ Hv a positive integer, d(a, b), so
that d(a, b) − 1 is the number of horizontal strips of infinite height between a
and b as we go around v.
We call a graph as above, i.e. a finite graph with orders on Hv and integers d(a, b),
an S-graph. In order to record the entire flat structure we attach a number Z(e) in the
upper half-plane, H, to each edge e, given by the vector between the singularities on the
boundary of the corresponding horizontal strip. We summarize the construction in the
following proposition.

Proposition 6.1. — There is a bijective correspondence between (isomorphism classes of)

1. Flat surfaces of finite type with infinite area with the property that there are no horizontal leaves
of finite length, and
2. S-graphs with choice of number Z(e) ∈ H for each edge e.

The data of an S-graph G may be encoded in terms of a graded quiver with rela-
tions Q. Vertices of Q are edges of G. Arrows of Q are given by pairs (h1 , h2 ) of successive
half edges h1 < h2 ∈ Hv in degree d(h1 , h2 ). Relations are quadratic: (h1 , h2 ) · (h3 , h4 ) = 0
when h2 = h3 belong to the same edge. Note that in the graded linear path category
A = KQ we have

K if A = B
(6.1) Hom (A, B) = 0,
<0
Hom (A, B) =
0
0 else

These properties ensure that the extension closure of the objects of A in TwA = PerfA
is the heart of an Artinian bounded t-structure, and objects of A are precisely the simple
ones.
310 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH

Proposition 6.2. — Let X be a flat surface with infinite area such that no horizontal leaves have
finite length (which is true for generic choice of horizontal direction), and A be the heart of the t-structure
for the stability structure on F (X) determined by the flat metric, then A is Artinian with a finite number of
simple objects. In terms of the S-graph, A is described by the graded quiver with relations considered above.
Proof. — Choose a family Xt , t ∈ [0, 1] of flat structures so that X0 = X, none of
the Xt have finite horizontal leaves, and the simple saddle connections of X1 are vertical.
All the Xt have horizontal strip decompositions described by the same S-graph, only the
parameters Z(e) change, and are in iR>0 for X1 . The corresponding stability structures
σt all have the same heart of the t-structure, as none of them have semistable objects with
real central charge (by the condition on the horizontal leaves), i.e. there is no wall-crossing
of the second kind.
The flat surface X1 has a purely one-dimensional core, which is just the union K of
the simple saddle connections. This is clear, since K contains all singularities, and exterior
angles are ≥ π . We conclude that X1 has no closed geodesics and all saddle connections
are simple. Thus, stable objects of σ1 in the heart of the t-structure are just the simple
geodesics. Also, the central charge of σ1 takes values in iR, so simple objects in the heart
of the t-structure are the same as stable ones.
The quiver determined by the S-graph is a special case of the graded quiver with
relations describing End(G) of a formal generator of F (X) (Section 3.4). 
Mutation
We have assumed above that there are no horizontal saddle connections. Points in
the moduli space where this does happen, i.e. some stable object has real central charge,
are by definition on walls of the second kind. Let us consider a generic path γ : (−, ) →
M(X) which crosses one of these walls at time t = 0. There are two possibilities, either
the foliation of γ (0) has a single horizontal saddle connection, e, or some of the leaves are
closed loops foliating a cylinder. The latter is a more complicated kind of wall-crossing,
and we will restrict attention to the former case.
Making  smaller, we may assume that the horizontal foliations of γ (t), t = 0, have
no finite-length leaves. Let G− , G+ be the S-graphs for γ (t) for t < 0, t > 0 respectively.
The saddle connection e becomes simple for small non-zero values of t, and so can be
identified with an edge of both G− and G+ . The graphs G− , G+ are related by mutation,
which we want to describe explicitly. There are two cases: Left mutation if the phase of
e is moving clockwise, right mutation if it is moving counterclockwise. Let us assume the
former. Looking at the horizontal strip decompositions one sees that S+ is obtained from
S− as follows. For a half-edge a denote by S(a) and S−1 (a) its successor and predecessor
respectively.
1. Let a, b be the half-edges of e. Decrement d(a, S(a)) and d(b, S(b)) by one.
2. If d(a, S(a)) = 0, then slide the end of S(a) along e so that it becomes the new
predecessor, c, of b with d(c, b) = 1. Otherwise, increment d(S−1 (b), b) by one.
3. Do the previous step with a and b switched.
 FLAT SURFACES AND STABILITY STRUCTURES 311

FIG. 7. — Left mutation of the S-graph

The process is illustrated in Figure 7. Right mutation is then simply the inverse opera-
tion.

Remark. — In the special case when all zeros are simple we get an ideal triangulation
(for generic horizontal direction) by picking a single horizontal leaf from the interior of
each horizontal strip as in [23]. A mutation of the S-graph corresponds then to a flip of
the triangulation. In general however the mutations described here are not mutations in
the sense of cluster theory.

6.2. Example: Dynkin type An

Let S be a graded surface with interior diffeomorphic to a disk and with n + 1 ≥ 2
marked boundary components. Such a graded surface is unique up to isomorphism. The
Fukaya category F (S) is equivalent to the bounded derived category of any An -type
quiver.

Theorem 6.1. — For S as above of An -type,

 
(6.2) M(S) = Stab F (S) ∼ = Cn .

The case n = 2 of this theorem is contained in [14].

Proof. — The proof consists of two parts, first to show that M(S) ∼ = Cn as complex
manifolds, and then to show that the map M(S) → Stab(F (S)) is surjective.
1. We may assume that the interior of S is identified with C, and the grading given
by the standard horizontal foliation on C. Define an analytic map u : Cn → M(S) by
   
(6.3) (a0 , . . . , an−1 ) → exp zn+1 + an−1 zn−1 + · · · + a0 dz2 , a0 .

This requires some explanation. By the results of Section 2.5, the quadratic differentials
above define a flat structure on C, the completion of which is, as a topological space,
independent of the parameters ai . We assume that S is the real blow-up of this comple-
tion, so that we get a flat structure on S. To determine a point in M(S) we also need a
312 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH

homotopy class of paths from the horizontal foliation of the quadratic differential to the
standard horizontal foliation on z, which is what a0 is used for. Since S is contractible,
such a homotopy class of paths is determined by its restriction to a single point, which we
take to be the origin, where a0 defines a path exp((1 − t)a0 )dz2 in (T∗0 C)⊗2 , t ∈ [0, 1].
An analytic automorphism of C acting on the set of polynomials of the form
z + an−1 zn−1 + · · · must be of the form z → ζ z, ζ n+1 = 1. All these extend to auto-
n+1

morphisms of S, but only for ζ = 1 is it isotopic to the identity. This implies injectivity
of u. For surjectivity, let S be a contractible flat surface with n + 1 conical points, all with
infinite cone angle. By the results of Section 2.5, S comes from a quadratic differential ϕ
with exponential singularities on a compact Riemann surface C. Since S is contractible,
C must be CP1 and ϕ must have a single exponential singularity. After applying a suit-
able Möbius transformation, the exponential singularity is at ∞ and ϕ is of the form
exp(zn+1 + an−1 zn−1 + · · · + a0 )dz2 .
2. Let σ ∈ Stab(F (S)) be a stability structure, A the heart of its t-structure. As
F (S) has only a finite number of indecomposable objects up to shift, A must be Artinian
with n = rkK0 (F (S)) simple objects E1 , . . . , En , up to isomorphism. Each Ei corresponds
to an isotopy class of graded arcs in S, which we denote by the same symbol. Since
Ext≤1 (Ei , Ej ) = 0 for i = j, we can choose representatives of the Ei which intersect only
at the endpoints. Because the Ei give a basis of H1 (S, ∂S; Zτ ), they necessarily form a
full formal system of graded arcs. Its structure is given by an S-graph which is a tree with
n edges. Rotating σ slightly by an angle , we may assume that no central charge lies
in R. Using Proposition 6.1 we get a flat surface X corresponding to σ , after rotating the
horizontal direction back by −. We already showed in the first part that X belongs to
M(S). 

Proposition 6.3. — Let S be the graded marked surface of An -type, n > 1, then the canonical
map MCG(S) → Aut(F (S)) from the mapping class group of graded automorphisms to the group
of autoequivalences up to natural equivalence, is an isomorphism. The group is abelian, isomorphic to
Z2 /Z(2, n + 1).

In the case n = 1 this map is surjective with kernel Z/2.

Proof. — The assumption n > 1 guarantees that distinct graded boundary arcs of
S correspond to distinct isomorphism classes of objects in F (S). Since any element of
MCG(S), which are all represented by rotations, is determined by its action on graded
boundary arcs, this implies injectivity.
Note that (isomorphism classes of) objects E ∈ F (S) corresponding to graded
boundary arcs are intrinsically characterized, e.g. by being indecomposables with F (S)/E
equivalent to the bounded derived category of an An−1 quiver. Thus an automorphism
 of F (S) induces a map on the set B of graded boundary arcs. B has a natural total
 FLAT SURFACES AND STABILITY STRUCTURES 313

FIG. 8. — Possible topologies of K when n = 3 for a generic stability structure. Note that the number of stable objects up
to shift is 6, 5, 4, 3 respectively

order, isomorphic to the total order on Z, so that dim Ext(Ei , Ej ) = 1 if j = i + 1, and 0

otherwise. This implies that the induced map on B is a shift, hence comes from the action
of an element of MCG(S). Finally, any autoequivalence of F (S) which fixes all objects is
seen to be naturally isomorphic to the identity functor. 

Remark. — Fix a grading on each boundary arc of S and let B1 , . . . , Bn+1 be the
Yoneda-duals of the corresponding objects in F (S). Their sum is a functor B : F (S) →
C∼= Perf(Kn+1 ). In order to make the statement of the above proposition work for n = 1
one needs to replace Aut(F (S)) by the group of pairs of autoequivalences, one of F (S),
one of C , which are compatible with B, up to natural equivalence. Something like this
also seems to be needed in order to generalize the proposition to other marked surfaces.
In the context of Fukaya–Seidel categories, B is the restriction to the distinguished fiber.

Corollary 6.1. — Let S be the graded marked surface of An -type, n > 1, then
   
(6.4) Stab F (S) /Aut F (S) = M(S)/MCG(S)

(6.5) = eP(z) dz2 | deg P = n + 1 /Aut(C)

n+1 +a n−1 +···+a
(6.6) = ez n−1 z 0
dz2 /Z/(n + 1)Z

as complex orbifolds.

In order to gain a better understanding of the wall and chamber structure on

M(S) = Stab(F (S)) one needs to study possible types of cores and S-graphs, giving
chambers between walls of the first and second kind respectively. We already know that
S-graphs which occur for S are those with underlying graph a tree with n edges and
total order on each set of half-edges. Possible types of cores, even generic ones, are more
difficult to enumerate except for small values of n. Figure 8 gives some examples for
n = 3.

n
6.3. Example: affine type A
Fix positive integers p, q and let n + 1 = p + q. Let S be a graded surface with
interior an annulus and p (resp. q) marked boundary components, each diffeomorphic
314 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH

to an open interval, on the left (resp. right) end, and grading so that a simple closed
loop around the annulus has vanishing Maslov index, i.e. is gradable. Such S is uniquely
determined up to isomorphism. For concreteness we take S to be the real blow-up of
the completion of C∗ with respect to the metric |ϕ|, where ϕ = exp(zp + z−q )dz2 , and
grading given by the horizontal foliation of ϕ. The Fukaya category F (S) is equivalent to
the bounded derived category of any quiver with underlying graph the extended Dynkin
diagram An and p (resp. q) arrows directed counterclockwise (resp. clockwise).

n -type,
Theorem 6.2. — For S as above of A
 
(6.7) M(S) = Stab F (S) ∼= Cn+1 .

When n = 1 the category F (S) is equivalent to the bounded derived category

of coherent sheaves on P1 and it was proven by Okada [39] that Stab(Db (P1 )) ∼ = C2 .
For n = 2 it was shown by different methods in [19] that Stab(F (S)) is contractible. By
homological mirror symmetry F (S) is equivalent to the bounded derived category of the
weighted projective line with weights (p, q) (see e.g. [45]) so our result gives a complete
description of their spaces of stability conditions.

Proof. — As in the Dynkin case the proof consists of two parts, first to show
that M(S) = Cn+1 as complex manifolds, and then to show that the map M(S) →
Stab(F (S)) is surjective.
1. To a point (ap−1 , . . . , a−q ) ∈ Cn+1 we assign the flat surface given by the quadratic
differential
 
(6.8) exp zp + ap−1 zp−1 + · · · + a−q+1 z−q+1 + exp(a−q )z−q dz2

on C∗ . Since Cn+1 is contractible and the dependence on the parameters continuous, we

can resolve the MCG(S)-ambiguity and get a map u : Cn+1 → M(S). Arguments as in
the proof of Theorem 6.1 show that u is an isomorphism.
2. Let σ ∈ Stab(F (S)). Unlike the Dynkin case, the heart of the t-structure of σ
may not be Artinian. We will show however that this happens non-generically and the
t-structure becomes Artinian after possibly rotating σ slightly. Let  = K0 (F (S)) ∼
= Zn+1 ,
and let N ⊂  be the subgroup generated by a simple curve around the annulus. Equiv-
alently, N is the kernel of the Euler form on . Let E ⊂  be the set of classes of in-
decomposable objects. It follows from the classification of indecomposable objects that
E/N ⊂ /N is finite. Hence, the image of E under Z :  → C has finitely many orbits
under translation by the vectors Z(N), and thus phases of stable objects can only ac-
cumulate towards Arg(Z(N)), or not at all if Z(N) = {0}. In particular, after possibly
rotating σ , 0 is not in the closure of the set of phases of stable objects, and thus the
 FLAT SURFACES AND STABILITY STRUCTURES 315

heart of the t-structure is Artinian. From here we can proceed as in the proof of Theo-
rem 6.1. 

7. Open problems and further directions

Moduli of objects
For a partially wrapped Fukaya category F of a surface S our classification de-
scribes Ob(F ) as a set in terms of curves on S. However, Ob(F ) has a much richer
structure of a derived stack.

Problem. — Describe the geometry of Ob(F ).

Also, the case of Fukaya categories of closed surfaces remains. This would be an
A-side generalization of Atiyah’s classification for the elliptic curve. However, for g > 1
the category is only Z/2-graded and there are no t-structures or stability structures.

Problem. — Classify objects of F (S) when S is a closed surface with g(S) > 1, assuming this
is a tame problem.

Higher dimensions
There is at present no complete conjectural picture extending the results of this
paper to higher dimensional symplectic manifolds. Investigations into this direction were
started by Thomas [48] and Thomas–Yau [49] with the aim of finding an algebraic no-
tion of stability describing special Lagrangian submanifolds and their mean curvature
flow. More recently, their proposal has been updated by Joyce [28], taking into account
recent developments on Fukaya categories, mean curvature flow, and Bridgeland’s ax-
iomatics. The conjecture is that given a compact Calabi–Yau M one has a stability struc-
ture on its Fukaya category, F (M), such that singular special Lagrangians in M, with ad-
ditional data to make them objects in F (M), are (all?) stable objects. The central charge
is given by integration of the holomorphic volume form. Mean curvature flow applied to
an object in F (S), at least with sufficiently small phase variation, should converge to the
components of the Harder–Narasimhan filtration. It is however still an open problem to
describe (a component of) Stab(F (M)) in general, even conjecturally.

Problem. — Find a geometric description of Stab(F (M)) for higher dimensional M.

Acknowledgements

The idea that quadratic differentials with exponential-type singularities can be

identified with stability structures was suggested to the authors by Alexander Polishchuk.
316 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH

We also thank Mohammed Abouzaid, Denis Auroux, Tom Bridgeland, Igor Burban,
Tobias Dyckerhoff, Alexander Efimov, Alexander Goncharov, Dominic Joyce, Mikhail
Kapranov, Gabriel Kerr, Andrew Neitzke, Tony Pantev, Yan Soibelman, Pawel Sosna,
and Zachary Sylvan for discussions relating to this paper. Our thanks also go to the
anonymous referees for carefully reading the manuscript and providing helpful sugges-
tions. The authors were partially supported by the FRG grant DMS-0854977, the re-
search grants DMS-0854977, DMS-0901330, DMS-1201321, DMS-1201475, DMS-
1265230, OISE-1242272 PASI from the National Science Foundation, the FWF grant
P24572-N25, by an ERC GEMIS grant, and by an RF Government grand, ag. No.
14.641.31.0001. We are grateful to the IHES for providing perfect working conditions.
The authors were supported by two Simons Foundation grants.

REFERENCES
1. M. ABOUZAID, On the Fukaya categories of higher genus surfaces, Adv. Math., 217 (2008), 1192–1235.
2. M. ABOUZAID and P. SEIDEL, An open string analogue of Viterbo functoriality, Geom. Topol., 14 (2010), 627–718.
3. I. ASSEM and A. SKOWROŃSKI, Iterated tilted algebras of type Ãn , Math. Z., 195 (1987), 269–290.
4. M. F. ATIYAH, Vector bundles over an elliptic curve, Proc. Lond. Math. Soc., 3 (1957), 414–452.
5. D. AUROUX, Fukaya categories and bordered Heegaard–Floer homology, in Proceedings of the International Congress of
Mathematicians, vol. II, pp. 917–941, Hindustan Book Agency, Gurugram, 2010.
6. M. BARDZELL, The alternating syzygy behavior of monomial algebras, J. Algebra, 188 (1997), 69–89.
7. A. BAYER, E. MACRÌ and Y. TODA, Bridgeland stability conditions on threefolds I: Bogomolov-Gieseker type inequali-
ties, J. Algebraic Geom., 23 (2014), 117–163.
8. V. BEKKERT and H. A. MERKLEN, Indecomposables in derived categories of gentle algebras, Algebr. Represent. Theory, 6
(2003), 285–302.
9. R. BOCKLANDT, Noncommutative mirror symmetry for punctured surfaces, Trans. Am. Math. Soc., 368 (2016), 429–469.
10. J. P. BOWMAN and F. VALDEZ, Wild singularities of flat surfaces, Isr. J. Math., 197 (2013), 69–97.
11. T. BRIDGELAND, Stability conditions on triangulated categories, Ann. Math., 166 (2007), 317–345.
12. T. BRIDGELAND, Stability conditions on K3 surfaces, Duke Math. J., 141 (2008), 241–291.
13. T. BRIDGELAND, Stability conditions and Kleinian singularities, Int. Math. Res. Not., 21 (2009), 4142–4157.
14. T. BRIDGELAND, Y. QIU and T. SUTHERLAND, Stability conditions on the A2 quiver, arXiv:1406.2566.
15. T. BRIDGELAND and I. SMITH, Quadratic differentials as stability conditions, Publ. Math. IHÉS, 121 (2015), 155–278.
16. I. BURBAN and Y. DROZD, On derived categories of certain associative algebras, in Representations of Algebras and Related
Topics, Fields Inst. Commun., vol. 45, pp. 109–128, Am. Math. Soc., Providence, 2005.
17. I. BURBAN and B. KREUSSLER, Derived categories of irreducible projective curves of arithmetic genus one, Compos.
Math., 142 (2006), 1231–1262.
18. G. DIMITROV, F. HAIDEN, L. KATZARKOV and M. KONTSEVICH, Dynamical systems and categories, in The Influence of
Solomon Lefschetz in Geometry and Topology: 50 Years of Mathematics at CINVESTAV, Con. Math., vol. 621, pp. 133–170, Am.
Math. Soc., Providence, 2014.
19. G. DIMITROV and L. KATZARKOV, Stability conditions on the acyclic triangular quiver, arXiv:1410.0904.
20. T. DYCKERHOFF, A1-homotopy invariants of topological Fukaya categories of surfaces, arXiv:1505.06941.
21. T. DYCKERHOFF and M. KAPRANOV, Triangulated surfaces in triangulated categories, arXiv:1306.2545.
22. K. FUKAYA, Y.-G. OH, H. OHTA and K. ONO, Lagrangian Intersection Floer Theory, Anomaly and Obstruction, Parts I, II,
AMS/IP Studies in Adv. Math., vol. 46.1, 2009.
23. D. GAIOTTO, G. W. MOORE and A. NEITZKE, Wall-crossing, Hitchin systems, and the WKB approximation, Adv. Math.,
234 (2013), 239–403.
24. A. GROTHENDIECK, Groupes de classes des categories abeliennes et triangulees. Complexes parfaits (Redige par I.
Bucur), in Semin. Geom. Algebr. Bois-Marie 1965–1966, SGA 5, Lect. Notes Math., vol. 589, Expose No. VIII, pp. 351–
371, 1977.
 FLAT SURFACES AND STABILITY STRUCTURES 317

25. J. L. HARER, Stability of the homology of the mapping class groups of orientable surfaces, Ann. Math. (2), 121 (1985),
215–249.
26. J. L. HARER, The virtual cohomological dimension of the mapping class group of an orientable surface, Invent. Math.,
84 (1986), 157–176.
27. A. IKEDA, Stability conditions on CYn categories associated to An -quivers and period maps, arXiv:1405.5492.
28. D. JOYCE, Conjectures on Bridgeland stability for Fukaya categories of Calabi–Yau manifolds, special Lagrangians,
and Lagrangian mean curvature flow, arXiv:1401.4949.
29. H. KAJIURA, K. SAITO and A. TAKAHASHI, Matrix factorization and representations of quivers. II. Type ADE case, Adv.
Math., 211 (2007), 327–362.
30. M. KONTSEVICH, Homological algebra of mirror symmetry, in Proceedings of the International Congress of Mathematicians,
vols. 1, 2, Zürich, 1994, pp. 120–139, Birkhäuser, Basel, 1995.
31. M. KONTSEVICH and Y. SOIBELMAN, Stability structures, motivic Donaldson-Thomas invariants and cluster transfor-
mations, arXiv:0811.2435.
32. M. KONTSEVICH and Y. SOIBELMAN, Notes on A∞ -algebras, A∞ -categories and non-commutative geometry, in Homo-
logical Mirror Symmetry, Lecture Notes in Phys., vol. 757, pp. 153–219, Springer, Berlin, 2009.
33. V. LYUBASHENKO and S. OVSIENKO, A construction of quotient A∞ -categories, Homol. Homotopy Appl., 8 (2006), 157–
203.
34. E. MACRÌ, Stability conditions on curves, Math. Res. Lett., 14 (2007), 657–672.
35. H. MASUR and J. SMILLIE, Hausdorff dimension of sets of nonergodic measured foliations, Ann. Math. (2), 134 (1991),
455–543.
36. D. NADLER, Cyclic symmetries of An -quiver representations, arXiv:1306.0070.
37. L. A. NAZAROVA and A. V. ROĬTER, A certain problem of I.M. Gel’fand, Funkc. Anal. Prilozh., 7 (1973), 54–69.
38. R. NEVANLINNA, Über Riemannsche Flächen mit endlich vielen Windungspunkten, Acta Math., 58 (1932), 295–373.
39. S. OKADA, Stability manifold of P1 , J. Algebraic Geom., 15 (2006), 487–505.
40. J. PASCALEFF and N. SIBILLA, Topological fukaya category and mirror symmetry for punctured surfaces, arXiv:1604.
06448.
41. A. POLISHCHUK and E. ZASLOW, Categorical mirror symmetry: the elliptic curve, Adv. Theor. Math. Phys., 2 (1998),
443–470.
42. P. SEIDEL, Fukaya A∞ structures associated to Lefschetz fibrations. II, arXiv:1404.1352.
43. P. SEIDEL, Graded Lagrangian submanifolds, Bull. Soc. Math. Fr., 128 (2000), 103–149.
44. P. SEIDEL, Fukaya Categories and Picard-Lefschetz Theory, Zurich Lectures in Advanced Mathematics, European Mathe-
matical Society (EMS), Zürich, 2008.
45. N. SIBILLA, D. TREUMANN and E. ZASLOW, Ribbon graphs and mirror symmetry, Sel. Math. New Ser., 20 (2014), 979–
1002.
46. I. SMITH, Quiver algebras as Fukaya categories, arXiv:1309.0452.
47. K. STREBEL, Quadratic Differentials, Ergebnisse der Mathematik und ihrer Grenzgebiete (3) [Results in Mathematics and
Related Areas (3)], vol. 5, Springer, Berlin, 1984
48. R. P. THOMAS, Moment maps, monodromy and mirror manifolds, in Symplectic Geometry and Mirror Symmetry, Seoul, 2000,
pp. 467–498, World Scientific, River Edge, 2001.
49. R. P. THOMAS and S.-T. YAU, Special Lagrangians, stable bundles and mean curvature flow, Commun. Anal. Geom., 10
(2002), 1075–1113.
50. W. A. VEECH, Flat surfaces, Am. J. Math., 115 (1993), 589–689.

F. H.
Department of Mathematics, Science Center,
Harvard University,
One Oxford Street,
Cambridge, MA 02138, USA
[email protected]
318 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH

L. K.
Fakultät für Mathematik,
Universität Wien,
Oskar-Morgenstern-Platz 1,
1090 Wien, Austria
[email protected]
and
HSE Moscow,
Moscow, Russia

M. K.
Institut des Hautes Études Scientifiques,
35 route de Chartres,
91440 Bures-sur-Yvette, France
[email protected]

Manuscrit reçu le 23 mai 2016

Manuscrit accepté le 14 novembre 2017
publié en ligne le 21 novembre 2017.