SEMISTABILITY, MODULAR LATTICES, AND

ITERATED LOGARITHMS

F. Haiden, L. Katzarkov, M. Kontsevich, P. Pandit

Abstract
We provide a complete description of the asymptotics of the
gradient flow on the space of metrics on any semistable quiver
representation. This involves a recursive construction of approx-
imate solutions and the appearance of iterated logarithms and a
limiting filtration of the representation. The filtration turns out to
have an algebraic definition which makes sense in any finite length
modular lattice. This is part of a larger project by the authors
to study iterated logarithms in the asymptotics of gradient flows,
both in finite and infinite dimensional settings.

Contents
1. Introduction 2
1.1. Background 2
1.2. Weight filtration 3
1.3. Modular lattices and iterated weight filtration 4
1.4. Asymptotics of the gradient flow 6
1.5. Outline 7
2. Zig-zag example 8
2.1. Weight filtration 8
2.2. Asymptotics 9
3. Weights on directed acyclic graphs 10
3.1. Weight grading 11
3.2. Gradient flow 13
3.3. From DAGs to lattices 16
4. Weight filtrations in modular lattices 16
4.1. Some lattice theory basics 17
4.2. Harder–Narasimhan filtration and mass 18
4.3. Paracomplemented R–filtrations 21
4.4. Weight filtrations 25
4.5. Iterated weight filtration 30
5. Gradient flow on quiver representations 32

1
2 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH, P. PANDIT

5.1. Kähler geometry of quiver representations 32

5.2. Star-algebras and bimodules 34
5.3. Monotonicity and homogeneity 37
5.4. Asymptotic solution 39
References 45

1. Introduction
This paper consists of two parts. The first is lattice-theoretic (lat-
tice in the sense of partial order) and the main result is the existence
of a weight-type filtration, depending on finitely many real parameters,
in any finite-length modular lattice. In the second part we study the
asymptotic behavior of the gradient flow on the space of Hermitian met-
rics on a quiver representation, which involves iterated logarithms, i.e.
the functions log t, log log t, log log log t, . . . , and turns out to be con-
trolled by the filtration defined in the first part applied to the lattice of
subrepresentations.
Evidence that the case of quiver representations is just one example
of a more general theory of asymptotics of certain gradient flows and it-
erated logarithms can be found in our companion paper [8]. The natural
context for the considerations here should be some form of “categorical
Kähler geometry”, a geometric enhancement of Bridgeland’s notion of
stability [3], which the authors are developing in an ongoing project [7].
1.1. Background. An n × n-matrix, A, with complex entries is diag-
onalizable if and only if there is a Hermitian metric (inner product),
h, on Cn such that A is normal, i.e. AA∗ = A∗ A, when the adjoint
is taken with respect to h. A generalization of this fact to quiver rep-
resentations was discovered by A.D. King [12]. A quiver, Q, is just a
finite graph with oriented edges (arrows), and a representation assigns
a finite-dimensional vector space Ei over C to every vertex i and a lin-
ear map φα : Ei → Ej to every arrow α : i → j. Representations of
a quiver form an abelian category, so in particular the notions of sim-
ple and semisimple representation are defined in the usual way. King’s
theorem then states that E is semisimple if and only if one can find a
metric on each Ei such that
X
(1.1) [φ∗α , φα ] = 0.
α
This result can be seen as a finite-dimensional analog of the celebrated
Donaldson and Uhlenbeck–Yau theorem [5, 17] which relates existence
of Hermitian Yang–Mills metrics on a holomorphic vector bundle to
slope stability of that bundle. Both are a special instance of the gen-
eral Kempf–Ness principle [11] which relates quotient constructions in
geometric invariant theory and symplectic geometry.
SEMISTABILITY, MODULAR LATTICES, AND ITERATED LOGARITHMS3

The solutions to (1.1) are the minima of the function

X
(1.2) S(h) := tr(h−1 ∗
i φ α hj φ α )
α:i→j

on the space of metrics on the given representation, where h−1 ∗

i φα hj is
∗
the adjoint of φα with respect to h and φα is the adjoint with respect to
some arbitrary reference metric. Provided E is semisimple we can thus
follow a trajectory of the gradient flow
dhi X X
(1.3) mi h−1
i = h−1 ∗
i φα hj φα − φα h−1 ∗
j φα hi
dt
α:i→j α:j→i

to arrive at a solution to (1.1). Here the mi > 0 are parameters of the

homogeneous Riemannian metric
X
mi tr h−1 −1


hv, wi := i vhi w
i∈Q0
on the space of Hermitian metrics, h, on E. Thus, a natural question
arises: Suppose E is not semisimple, then what is the asymptotic behavior
of the gradient flow of the function S? Roughly what happens is that
the metric grows/decays at different rates on different vectors in E and
this determines a filtration by subrepresentations. It turns out that
just as the existence of a minimizing metric is controlled by an algebraic
criterion, semisimplicity, the asymptotic filtration, which we call iterated
weight filtration, has a purely algebraic definition which depends only on
the partially ordered set of subrepresentations of E together with a finite
number of real parameters, and can be generalized to any finite length
modular lattice.
1.2. Weight filtration. Before giving the general definition of the weight
filtration we describe it in the case of a representation of the quiver with a
single loop, i.e. just a vector space V with an endomorphism φ. Without
loss of generality we may assume that φ is nilpotent, otherwise consider
each algebraic eigenspace separately. The weight filtration is then the
unique filtration V≤λ of V by λ ∈ 21 Z such that φ(V≤λ ) ⊂ V≤λ−1 and
φk induces an isomorphism Grk/2 V → Gr−k/2 V for any positive integer
k. This is, up to relabeling, what Griffiths calls the Picard–Lefschetz
filtration induced by φ in [6], see also [15]. When φ is the logarithm of
the unipotent part of the monodromy, it gives the weight filtration on
the limiting mixed Hodge structure on the vanishing cohomology of an
isolated hypersurface singularity. This is the origin of our terminology.
For the reader who is more familiar with abelian categories than mod-
ular lattices we state our first main result, the definition of the weight
filtration (Theorem 4.9 in the main text), in this context.
Theorem/Definition. Suppose A is an artinian (finite length) abelian
category and X : K0 (A) → R a homomorphism which is positive on each
4 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH, P. PANDIT

class of a non-zero object. For each object E ∈ A there exists a unique

filtration
(1.4) 0 = E0 ⊂ E1 ⊂ . . . ⊂ En = E
with subquotients Ek /Ek−1 6= 0 labeled by real numbers λ1 < . . . < λn
such that the following conditions are satisfied.
1) The subquotient El /Ek−1 is semisimple for any 1 ≤ k ≤ l ≤ n with
λl − λk < 1.
2) The balancing condition
n
X
(1.5) λk X(Ek /Ek−1 ) = 0
k=1

holds.
3) For any collection of objects Fk with Ek−1 ⊆ Fk ⊆ Ek , k =
1, . . . , n, such that Fl /Fk is semisimple for 1 ≤ k < l ≤ n with
λl − λk ≤ 1, the inequality
n
X
(1.6) λk X(Fk /Ek−1 ) ≤ 0
k=1

holds.
The uniquely defined filtration, depending on X, is called the weight
filtration on E.
There is always a canonical choice for X, which is to assign to each
object its length (of a Jordan–Hölder filtration). The filtration is trivial
precisely when E is semisimple.
We emphasize that in the case of quiver representations the filtration
defined above gives only the first approximation to the iterated weight
filtration which determines the asymptotics of the gradient flow. As
the name suggests, this is a refinement of the weight filtration which is
constructed by iteratively applying the above theorem/definition some
finite number of times. In order to define it, it will be much more
convenient to use the language of lattices.

1.3. Modular lattices and iterated weight filtration. Suppose E

is an object in an artinian abelian category, e.g. a quiver representation.
The set L of subobjects of E, partially order by inclusion, enjoys the
following properties crucial for our purposes:
• L is a lattice: Any two elements a, b have a least upper bound
a ∨ b and greatest lower bound a ∧ b.
• modularity: If a ≤ b then (x ∧ b) ∨ a = (x ∨ a) ∧ b for all x ∈ L.
• finite length: There is a global upper bound on the length of any
chain a1 < a2 < . . . < an of elements in L.
SEMISTABILITY, MODULAR LATTICES, AND ITERATED LOGARITHMS5

In the first part of the paper we will work in the general context of finite
length modular lattices. Besides providing a natural level of generality,
there are interesting examples of modular lattices which do not come
from abelian categories, for instance normal subgroups of a finite group,
or semistable subbundles (of the same slope) of a semistable Arakelov
bundle.
There is an analog of the Grothendieck group K0 for modular lattices.
Let [a, b] := {x ∈ L | a ≤ x ≤ b} be the interval from a to b. If L is a
modular lattice then denote by K(L) the abelian group with generators
[a, b], a ≤ b, and relations
[a, b] + [b, c] = [a, c], [a, a ∨ b] = [a ∧ b, b]
We let K + (L) ⊂ K(L) be the sub-semigroup generated by elements
[a, b], a < b. If L is moreover finite length then it is a consequence of the
Jordan–Hölder–Dedekind theorem that K(L) (resp. K + (L)) is a free
abelian group (resp. semigroup).
Note that an object in an abelian category is semisimple if and only if
the corresponding lattice of subobjects is complemented : For any a ∈ L
there is a b ∈ L with a ∧ b = 0 and a ∨ b = 1, where 0 (resp. 1) is the
minimum (resp. maximum) of L.
With the above definitions we are ready to state our first main result
in full generality. This is Theorem 4.9 in the main text with some of the
definitions unwrapped.
Theorem/Definition. Let L 6= ∅ be a finite length modular lattice
and X : K + (L) → R>0 an additive map. Then there exists a unique
filtration (=chain)
0 = a0 < a1 < . . . < an = 1
with intervals [ak−1 , ak ] 6= 0 labeled by real numbers λ1 < . . . < λn such
that the following conditions are satisfied.
1) The interval [ak−1 , al ] is complemented for 1 ≤ k ≤ l ≤ n with
λl − λk < 1.
2) The balancing condition
n
X
λk X([ak−1 , ak ]) = 0
k=1
holds.
3) For any collection of elements bk ∈ [ak−1 , ak ], k = 1, . . . , n, such
that [bk , bl ] is complemented for 1 ≤ k < l ≤ n with λl − λk ≤ 1,
the inequality
X n
λk X([ak−1 , bk ]) ≤ 0
k=1
holds.
6 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH, P. PANDIT

The uniquely defined filtration, depending on X, is called the weight

filtration on L.
Given L with its weight filtration ai , λi , i = 1, . . . , n we may con-
struct a new finite length modular lattice L0 which appeared implicitly
in the theorem/definition above. Namely an element b ∈ L0 is given by
elements bk ∈ [ak−1 , ak ], k = 1, . . . , n, such that
P [bk , bl ] is complemented
for 1 ≤ k < l ≤ n with λl − λk ≤ 1 and nk=1 λk X([bk−1 , bk ]) = 0.
Moreover there is a map X 0 : K + (L0 ) → R>0 given by
n
X
0
X ([b, c]) := X([bk , ck ]).
k=1

Applying the theorem/definition above to (L0 , X 0 ) we get a filtration of

L0 which in particular gives a filtration on each interval [ak−1 , ak ] by
definition of L0 , thus a refinement of the weight filtration on L indexed
by R2 with the lexicographical order. This inductive process continues
building a sequence (L, X), (L0 , X 0 ), (L00 , X 00 ), . . . until we reach a lattice
which is complemented, thus has trivial weight filtration. We call the
refinement of the weight filtration constructed in this way the iterated
weight filtration, and it is indexed by Rn ⊂ R∞ for some n with the
lexicographical order.
At this point, the reader may want to skip ahead to Section 2 for
a discussion of the simplest example where this refinement occurs. In
examples, the refinement appears not for generic choice of X, but along
“walls” in Hom(K + (L), R>0 ) described by real algebraic varieties defined
over Z. The general properties of these walls warrant further study.
We note that for any stability condition on a triangulated category
in the sense of Bridgeland [3] the full subcategory of semistable objects
of a fixed phase is finite-length abelian and the restriction of the central
charge determines a suitable map X. Thus, our theory gives a canoni-
cal refinement of the Harder–Narasimhan filtration of any object. One
application of this refinement is perhaps to define stratifications of the
stack of semistable objects by type of the weight-filtration. Refinements
of this sort were defined and studied by Kirwan [13], in particular for
vector bundles on a curve. We do not consider stratifications in the
present paper, but hope to return to this problem in the future.

1.4. Asymptotics of the gradient flow. For the dynamical interpre-

tation of the iterated weight filtration we identify its indexing set, R∞ ,
with the space of functions
R log t ⊕ R log log t ⊕ R log log log t ⊕ . . .
defined for t  0, as totally ordered sets.
As above, let Q be a quiver, E a representation of Q given by vector
spaces Ei for each vertex i and linear maps φα for each arrow α, and hi
SEMISTABILITY, MODULAR LATTICES, AND ITERATED LOGARITHMS7

a Hermitian metric on Ei , which will be allowed to vary. Fix a positive

real number mi for each vertex i of Q. These P determine an additive
map X : K0 (Rep(Q)) → R via X(E) := i mi dim Ei , as well as a
Riemannian metric on the space of metrics on any representation of Q,
see (1.2). Our general theory defines a unique iterated weight filtration
Fλ of E labeled by λ ∈ R∞ . Our second main result is the following,
which is Theorem 5.11 in the main text.
Theorem. Let E = (Ei , φα ) be a representation of a quiver Q over
C, then the limiting filtration of the flow (1.3) on metrizations of E
coincides with the iterated weight filtration on E as an object in the
category of representations of Q over C with X determined by the mi .
Moreover on the piece Fλ of the filtration, (λ1 , . . . , λn ) ∈ Rn ⊂ R∞ , any
trajectory h of the flow satisfies
(1.7) log |h(t)| = λ1 log t + λ2 log log t + · · · + λn log(n) t + O(1)
where log(k) is the k-times iterated logarithm.
The proof involves an inductive procedure which produces explicit
solutions of (1.3) up to terms in L1 . A crucial property of the flow is
monotonicity: If g, h are solutions with g(0) ≤ h(0), then g(t) ≤ h(t) for
all t ≥ 0.

1.5. Outline. The text is organized as follows. In Section 2 we discuss

in detail an example which exhibits many of the general features. This
should give the reader a good idea of the practical content of our theory
before diving deeper into it. In section 3 we look at the special case when
all Ei , i ∈ Q0 , are one-dimensional, where the weight filtration can be
defined much more easily as a solution to a convex optimization problem.
Section 4 concerns the purely lattice theoretic part or the work. After
reviewing some basics, the main goal is proving existence and uniqueness
of the weight filtration in any finite-length modular lattice. In Section 5
we construct asymptotic solutions to (1.3) and prove our second main
result. For this, the language of ∗-algebras and ∗-bimodules provides a
useful tool.
Acknowledgments: We thank S. Donaldson and C. Simpson for
useful discussions. We also thank anonymous referees for carefully proof-
reading the text and providing suggestions to help improve the exposi-
tion. The authors were supported by a Simons research grant, NSF DMS
150908, ERC Gemis, DMS-1265230, DMS-1201475 OISE-1242272 PASI,
Simons collaborative Grant - HMS, HSE Megagrant, Laboratory of Mir-
ror Symmetry NRU HSE, RF government grant, ag. 14.641.31.000,
Simons Principle Investigator Grant, CKGA VIHREN grant КП-06-
ПВ/16. Much of the research was conducted while the authors enjoyed
the hospitality of the IHES, the IMSA Miami, and the Laboratory of
Mirror Symmetry HSE Moscow
8 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH, P. PANDIT

2. Zig-zag example
In this section we discuss in detail the simplest example which ex-
hibits many of the general features: refinement of the weight filtration,
wall-crossing, and iterated logarithms. This is the four-dimensional rep-
resentation of the quiver
• •
(2.1)
• •
which assigns C to each vertex and the identity map to each arrow. We
hope this section will be aid the reader in following the more general
discussion in subsequent sections. In particular it would be useful to
read the second subsection below before attempting Subsection 5.4.
2.1. Weight filtration. The lattice L of subrepresentations of the rep-
resentation of the zig–zag quiver above is the set of order ideals in the
partially order set P := {1 > 2 < 3 > 4}, i.e. P has four elements and
Hasse diagram which looks like (2.1) and L is the set of subsets I ⊂ P
with the property that if a ∈ I, b ∈ P , b ≤ a, then b ∈ I, and has Hasse
diagram
P

124 234

12 24

2 4

∅
where we use the notation 12 := {1, 2} and so on. We have K + (L) = ZP>0
where e.g. [4, 124] = (1, 1, 0, 0), so X ∈ R4>0 under this identification.
• X1 X4 < X2 X3 : We claim that the weight filtration is ∅ < 24 < P
with labels λ1 = λ and λ2 = λ + 1 where
X1 + X3
λ=− ∈ (−1, 0)
X1 + X2 + X3 + X4
as follows from the balancing condition. This is verified by going
through the 16 possibilities for b1 ∈ [∅, 24], b2 ∈ [24, P ]. The con-
dition X1 X4 ≤ X2 X3 is needed only in the case b1 = 2, b2 = 124.
The strict inequality X1 X4 < X2 X3 ensures that L0 ∼ = {0, 1}, so
there is no refinement and the iterated weight filtration is equal to
the weight filtration.
SEMISTABILITY, MODULAR LATTICES, AND ITERATED LOGARITHMS9

• X1 X4 = X2 X3 : The weight filtration is ∅ < 24 < P as in the

previous case, however
L0 = {(∅, 24) < (2, 124) < (24, P )}
is not complemented and has weight filtration (∅, 24) < (2, 124) <
(24, P ) with labels µ, µ + 1 where
X3 + X4
µ=− .
X1 + X2 + X3 + X4
There is no further refinement and the iterated weight filtration
is thus ∅ < 2 < 24 < 124 < P with labels (λ, µ) < (λ, µ + 1) <
(λ + 1, µ) < (λ + 1, µ + 1).
• X1 X4 > X2 X3 : The weight filtration is ∅ < 2 < 24 < 124 < P
with labels
X1 X3
λ=− <µ=− <λ+1<µ+1
X1 + X2 X3 + X4
where λ < µ is equivalent to X1 X4 > X2 X3 . The lattice L0 has
four elements and is isomorphic to the lattice of subsets of a two-
element set, in particular complemented, so there is again no re-
finement of the weight filtration.
To summarize the situation we have a “wall” X1 X4 = X2 X3 dividing
the space R4>0 of parameters into two chambers. The filtration is, up
to relabeling, the same across a given open chamber. The refinement
occurs only along the wall.

2.2. Asymptotics. Let us look at the gradient flow (1.3) in our 4-

dimensional example. For convenience, write the ODE in terms of vari-
able xi = log hi ∈ R, then we get
m1 ẋ1 = ex2 −x1 , m2 ẋ2 = −ex2 −x1 − ex2 −x3 ,
(2.2)
m3 ẋ3 = ex2 −x3 + ex4 −x3 , m4 ẋ4 = −ex4 −x3 .
We try the ansatz
xi = ai log t + log bi
which we want to solve the above equations up to error terms in L1 (R0 ).
This is indeed possible as long as m1 m4 6= m3 m2 and discussed in detail
in Section 3 below. The numbers ai come from the labels of the weight
filtration for Xi = mi by our general theory. Here we consider instead
the more interesting case where m1 m4 = m3 m2 . For concreteness we
take m1 = m2 = m3 = m4 = 1.
Start by refining the first ansatz as follows:
1 1
x1 = log t + y1 (log t), x2 = − log t + y2 (log t),
(2.3) 2 2
1 1
x3 = log t + y3 (log t), x4 = − log t + y4 (log t)
2 2
10 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH, P. PANDIT

This gives a system of ODEs in new dependent variables yi and inde-

pendent variable s = log t,
1 1
ẏ1 = ey2 −y1 − , ẏ2 = −ey2 −y1 − ey2 −y3 + ,
(2.4) 2 2
1 1
ẏ3 = ey2 −y3 + ey4 −y3 − , ẏ4 = −ey4 −y3 + .
2 2
Note that y1 and y4 are fixed for
(2.5) y2 − y1 = y4 − y3 = − log 2.
Assuming (2.5), what remains is the system
(2.6) ẏ2 = −ey2 −y3 , ẏ3 = ey2 −y3
which has the same general form as the original one, (2.2). This self-
reproducing feature of this class of equations is completely parallel to the
passage from L to L0 in the construction of the iterated weight filtration
on a finite length modular lattice.
Returning to (2.6) we easily find the explicit solution
1 1
(2.7) y2 = − log(2s), y3 = log(2s).
2 2
Of course (2.5) assumed that y1 and y4 are fixed, which contradicts (2.7),
so we do not get a solution of the original system (2.4). However, it turns
out that (2.5) and (2.7) still give the correct asymptotic behavior up to
bounded terms. This follows from the general theory developed in the
subsequent sections. The proof involves construction a solution of (2.4)
up to error terms in L1 (R0 ). This is achieved by combining (2.5), (2.7),
and adding terms of the form C/s. Explicitly, the solution of (2.4) up
to terms in L1 is
1 s 1 1
y1 = − log + , y2 = − log(2s),
2 2 s 2
1 1 s 1
y3 = log(2s), y4 = log −
2 2 2 s
which may be substituted into (2.3) to give a solution to the original
system (2.2) up to terms in L1 .
It turns out that the strategy above, with some modifications, pro-
vides solution of the gradient flow up to terms in L1 for any quiver
representation, see Subsection 5.4. Another key ingredient, monotonic-
ity, will be discussed there.

3. Weights on directed acyclic graphs

For a special class of modular lattices constructed from directed acyclic
graphs by taking the set of closed subgraphs, the weight filtration has
a simpler definition avoiding the language of lattice theory. This corre-
sponds to the case of representations of acyclic quivers which assign a
SEMISTABILITY, MODULAR LATTICES, AND ITERATED LOGARITHMS
11

one-dimensional space to each vertex. The discussion in this section is

essentially subsumed by the later ones, and the reader is free to skip it,
but we hope this section will help motivate the general theory and make
it more accessible.

3.1. Weight grading. A directed acyclic graph (DAG) is an ori-

ented graph without multiple edges or oriented cycles. (For us, the terms
oriented graph and quiver are synonymous.) If G is a DAG, we write
G0 for the set of vertices and G1 ⊂ G0 × G0 for the set of edges/arrows.
Also write α : i → j to indicate that α is an edge from a i to j where
i, j ∈ G0 . We assume throughout that the graph is finite.
An R-grading on a DAG, G, is a choice of number vi ∈ R for every
vertex i ∈ G0 which decreases at least by one on each edge, i.e.
(3.1) vi − vj ≥ 1
if there is an edge α : i → j. R-gradings form a closed convex subset in
RG0 .
There is a canonical “energy minimizing” R-grading depending only
on (arbitrary) masses mi > 0, i ∈ G0 . More precisely, we define the
weight grading on G for given choice of the mi to be the R-grading v
which minimizes
X
(3.2) mi vi2 .
i∈G0

Since we are minimizing essentially the length squared on a closed convex

subset, existence and uniqueness of a minimizer follow for very general
reasons. The method of Lagrange multipliers (Karush–Kuhn–Tucker
conditions) gives the following equivalent definition of the weight grad-
ing.
Lemma 3.1. Let G be a DAG and mi ∈ R>0 for i ∈ G0 arbitrary,
then an R-grading, v, is the weight grading if it satisfies the following
condition: There are numbers uα ≥ 0, α ∈ G1 , such that uα = 0 for any
edge α : i → j with vi − vj > 1 and
X X
(3.3) mi vi = uα − uα
i−
→ j k−
→ i
α α

for i ∈ G0 .
The Lagrange multipliers uα are in general not unique unless G is
a tree. As a simple consequence of the lemma we see that the weight
grading v satisfies the balancing condition
X
(3.4) mi vi = 0.
i∈G0
12 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH, P. PANDIT

Furthermore, suppose E ⊂ G0 is a set of vertices with the following

property: If i ∈ E and α : i → j with vi − vj = 1 then j ∈ E. Then for
such subsets
X X
(3.5) mi vi = − uα ≤ 0.
i∈E α:k→i
k∈E,i∈E
/
It turns out these properties characterize the weight grading uniquely,
providing a convenient way of checking that a certain R-grading is in
fact the weight grading.
Proposition 3.2. Let G be a DAG with choice of mi > 0, then an
R-grading v is the weight grading if and only if
X
mi vi = 0.
i∈G0
and for every subset E ⊂ G0 such that if i ∈ E and α : i → j with
vi − vj = 1 then j ∈ E, then
X
mi vi ≤ 0.
i∈E

Proof. One implication is clear from the discussion above. Suppose

then that v satisfies the two conditions stated in the theorem. To show
that v is the weight grading it suffices to verify for δ in the tangent cone
at v to the space of R-gradings, C, that
X
(3.6) mi vi δi ≥ 0
i∈G0

i.e. the variation of (3.2) in the direction δ is non-negative. Note that

C consists of δ ∈ RG0 such that if there is an arrow α : i → j and
vi − vj = 1 then δi ≥ δj . It follows that C is generated by vectors 1G0
and −1E where E ranges over subsets of G0 such that if i ∈ E and
α : i → j with vi − vj = 1 then j ∈ E. By the first assumption on v,
the balancing condition, the variation vanishes in the direction 1G0 , and
by the second assumption it is non-negative in the directions −1E . This
shows that v is a minimum of (3.2). q.e.d.
Example 3.3. As a basic example, consider the following DAG with
n ≥ 1 vertices:
• −→ • −→ · · · −→ •
m1 m2 mn
The weight grading is given by (λ, λ − 1, . . . , λ − n + 1) where the highest
weight λ is determined by (3.4) to be
m2 + 2m3 + . . . + (n − 1)mn
(3.7) λ=
m1 + m2 + . . . + mn
Note that if mi = 1 for all i then the weights are integers or half-integers.
For this particular graph, the only effect of changing the parameters mi
SEMISTABILITY, MODULAR LATTICES, AND ITERATED LOGARITHMS
13

is to shift the overall grading. We will see below that in general more
interesting changes can occur along codimension one walls.
Remark 3.4. One can also consider graphs with infinitelyP many ver-
tices and parameters mi decaying sufficiently fast so that 2
m i vi < ∞
for some R-grading vi ∈ R. Elementary Hilbert space theory then im-
plies existence and uniqueness of an R-grading which minimizes total
mi vi2 .
P
energy

3.2. Gradient flow. The weight grading on a DAG has a dynamical

interpretation, describing the asymptotics of a certain gradient flow. Let
G, mi be as before and fix also constants cα > 0, α ∈ G1 . Consider the
function
X
(3.8) S : RG0 → R, S(x) = cα exj −xi .
α:i→j

The negative gradient flow of S with respect to the flat metric

X
(3.9) mi (dxi )2
i∈G0

is given by
X X
(3.10) mi ẋi = cα exj −xi − cα exi −xk
i−
→j k−
→i
α α

We can also write the flow in terms of variables attached to the edges
instead of the vertices. Set
(3.11) yα = −(xj − xi + log cα )
for each arrow α : i → j, then
(3.12)    
1  X X  1  X X
ẏα = e−yβ − e−yβ − e−yβ − e−yβ  .

mi mj
 
←−
α
i−
→ ←−
α
i←
− −
→
α
j−
→ −
→
α
j←
−
β β β β

The right hand side of the system of equations can be interpreted as

∆e−y , where ∆ is a graph Laplacian. In terms of variables pα = e−yα
the system of equations becomes a special case of the higher–dimensional
Lotka–Volterra equations which have the general form
 
X
(3.13) ṗα = pα  aαβ pβ + bα  .
β

This system provides a basic model for population dynamics, see for ex-
ample Hofbauer–Sigmund [10]. The asymptotic behavior in the general
case can be significantly more complicated than in our case — one need
not have convergence to a stable equilibrium.
14 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH, P. PANDIT

Example 3.5. Consider the simplest non-trivial DAG:

• −→ •
m1 m2

The system of ODEs (3.10) is

(3.14) m1 ẋ1 = cex2 −x1 , m2 ẋ2 = −cex2 −x1
with explicit solution
m2 m1
(3.15) x1 = log t + log C1 , x2 = − log t + log C2
m1 + m2 m1 + m2
where C1 , C2 > 0 are chosen so that
C2 m1 m2
(3.16) = .
C1 c(m1 + m2 )
Note that the coefficients of log t in the solution coincide with the weight
grading.
In general, the ODE (3.10) does not have an explicit solution, however
it turns out that we can always find an explicit asymptotic solution which
solves the equation up to terms in L1 . Such a solution will differ from an
actual solution by a bounded error term, thus have the same asymptotics
up to O(1).
We begin with the following ansatz for the solution xi (t).
(3.17) xi = vi log t + bi
with vi , bi ∈ R. Plugging this into (3.10) gives
mi vi X X
(3.18) = cα tvj −vi ebj −bi − cα tvi −vk ebi −bk .
t
i−
→ j k−
→ i
α α

For this equation to be true up to terms in L1 , it is necessary that only

t≤−1 appear on the right hand side, i.e. vj − vi ≤ −1 whenever there is
an edge α : i → j. Then, comparing coefficients of t−1 (other terms are
in L1 ) we need to solve
X X
(3.19) mi vi = cα ebj −bi − cα ebi −bk .
α:i→j α:k→i
vi −vj =1 vk −vi =1

Comparing this with Lemma 3.1, we see that the vi are necessarily the
weight grading on G. Furthermore, if (3.19) has a solution, b, then we
can evidently choose Lagrange multipliers uα such that uα > 0 whenever
α : i → j is an edge with vi − vj = 1, and uα = 0 otherwise. It turns
out that this is not always possible. We will see below that in the case
where we cannot solve (3.19) it is necessary to refine the original ansatz
with terms involving iterated logarithms.
SEMISTABILITY, MODULAR LATTICES, AND ITERATED LOGARITHMS
15

Note that (3.19) is the equation for a critical point of the function
X X
(3.20) S(x)
e = cα exj −xi + mi vi xi
α:i→j i∈G0
vi −vj =1

Suppose that we can find uα > 0 such that

X X
(3.21) mi vi = uα − uα
α:i→j α:k→i
vi −vj =1 vk −vi =1

then
X
cα exj −xi − uα (xj − xi )


(3.22) S(x)
e =
α:i→j
vi −vj =1

hence Se is the composition of a linear map (the differential d : RG0 →

RG1 ) with a proper strictly convex function, thus its critical locus is an
affine subspace of RG0 . We summarize the result so far in the following
theorem.
Theorem 3.6. Let G be a DAG, mi > 0, and v ∈ RG0 the weight
grading on G. Suppose that we can find uα > 0 for each edge α : i → j
with vi − vj = 1 such that (3.21) holds. Then the flow (3.10) admits an
asymptotic solution up to terms in L1 of the form
xi = vi log t + bi
and thus for an actual solution x(t) we have
xi = vi log t + O(1).
The claim about asymptotics of actual solutions could be verified here
directly without difficulty, however we will show it for a more general
setup in Subsection 5.3. We conclude this subsection with an example
where Theorem 3.6 is not applicable. Consider the following DAG which
is an orientation (zig-zag) of the A4 Dynkin diagram. Masses mi indicate
the labeling of the vertices.
m1 m3
(3.23)
m2 m4
The weight grading depends on the choice of m1 , m2 , m3 , m4 > 0.
Case m1 m4 > m2 m3 : In this region there are four distinct weights
λ
(3.24) µ
λ−1
µ−1
where
m4 m2
(3.25) λ= , µ= .
m3 + m4 m1 + m2
16 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH, P. PANDIT

The Lagrange multipliers which certify v are

m1 m2 m3 m4
(3.26) u1 = , u2 = 0, u3 =
m1 + m2 m3 + m4
hence Theorem 3.6 can be applied.
Case m1 m4 ≤ m2 m3 : In this region there are two distinct weights
λ = v1 = v3 and λ − 1 = v2 = v4 where
m2 + m4
(3.27) λ= .
m1 + m2 + m3 + m4
The Lagrange multipliers which certify v are
(3.28)
m1 (m2 + m4 ) m2 m3 − m1 m4 (m1 + m3 )m4
u1 = , u2 = , u3 =
M M M
where M = m1 + m2 + m3 + m4 . Note that u2 > 0 if and only if
m1 m4 < m2 m3 , so if m lies on the quadric m1 m4 = m2 m3 then the
condition of Theorem 3.6 is not satisfied.

3.3. From DAGs to lattices. Given a directed acyclic graph G con-

sider the collection L of subsets of E ⊂ G0 which span closed subgraphs,
i.e. no arrows lead out of E. Note that L is closed under unions and
intersections, thus a sublattice of the boolean lattice of all subsets of G0 .
We can almost recover G from the partially ordered set L. For example
the DAGs

(3.29) • • • • • •
have the same lattices of closed subgraphs. However, this does not affect
the weight grading.
The lattice of subrepresentations of a finite-dimensional representa-
tion is in general more complicated than the lattices constructed from
graphs, in that complements, if they exist, need not be unique. However,
such a lattice is still modular which leads to a good theory of filtrations.
In the next section we will generalize the notion of weight filtration from
DAGs to finite length modular lattices.

4. Weight filtrations in modular lattices

This section contains the proof of our first main result, the existence
and uniqueness of weight filtrations in finite-length modular lattices.
The reader interested mainly in the case of quiver representations and
willing to take this result on faith can skip this entire section on first
reading.
In the first subsection we review some definitions and results from
lattice theory. In Subsection 4.2 we define the Harder–Narasimhan fil-
tration of a finite length modular lattice with polarization, as well as
SEMISTABILITY, MODULAR LATTICES, AND ITERATED LOGARITHMS
17

its mass, and prove a triangle inequality for mass. Subsection 4.3 intro-
duces the concept of a paracomplemented R-filtration, which is essential
for the proof in Subsection 4.4. In the final subsection we discuss the
iterated weight filtration and provide examples to show that it can have
arbitrary depth.

4.1. Some lattice theory basics. In this subsection we recall some

basic notions from lattice theory, in particular modular lattices as intro-
duced by Dedekind. We learned this material in part from G. Birkhoff’s
classic textbook [2] and J.B. Nation’s online notes [14], which are ex-
cellent sources for more background.
A lattice is a partially ordered set, L, in which any two elements
a, b ∈ L have a least upper bound a ∨ b and greatest lower bound a ∧ b.
When L contains both a least element 0 ∈ L and greatest element 1 ∈ L,
then L is called a bound lattice. Given elements a ≤ b in L, the
interval from a to b is the bound lattice
(4.1) [a, b] := {x ∈ L | a ≤ x ≤ b}.
In a general lattice there are two ways of projecting an arbitrary element
x ∈ L to the interval [a, b], given by the left and hand right side of the
following inequality:
(4.2) (x ∧ b) ∨ a ≤ (x ∨ a) ∧ b.
The defining property of a modular lattice is that the above inequality
becomes an equality, hence
(4.3) a ≤ b =⇒ (x ∧ b) ∨ a = (x ∨ a) ∧ b for all x ∈ L.
The basic example of a modular lattice is the lattice of subobjects in a
given object of an abelian category.
There is an equivalence relation on the set of intervals in a modular
lattice L generated by
(4.4) [a, a ∨ b] ∼ [a ∧ b, b].
Modularity is equivalent to the condition that the maps
(4.5) [a, a ∨ b] → [a ∧ b, b], x 7→ x ∧ b
(4.6) [a ∧ b, b] → [a, a ∨ b], x 7→ x ∨ a
are inverse isomorphisms for all a, b ∈ L. Thus, equivalent intervals are
isomorphic lattices.
a∨b

(4.7)
⇐

a b
⇒

a∧b
18 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH, P. PANDIT

A lattice L is finite length if there is an upper bound on the length

n of any chain
(4.8) a0 < a1 < . . . < an
of elements in L. A finite length lattice is complete in the sense that any
(not necessarily finite) collection of elements has a least upper bound
and greatest lower bound. In particular, unless L = ∅, there are least
and greatest elements 0 and 1 in any finite length lattice. We say a
lattice is artinian if it is modular and has finite length. In an artinian
lattice, any two maximal chains have the same length, in fact:
Theorem 4.1 (Jordan–Hölder–Dedekind). Suppose
0 = a0 < a1 < . . . < am = 1, 0 = b0 < b1 < . . . < bn = 1
are maximal chains in a modular lattice. Then m = n and there is a
permutation σ of the set {1, . . . , n} such that there are equivalences of
intervals
[ak−1 , ak ] ∼ [bσ(k)−1 , bσ(k) ]
for k = 1, . . . , n.
The proof is essentially the same as for the classical Jordan–Hölder
theorem, but translated into the setting of modular lattices. See for
example the texts mentioned at the beginning of this subsection.
Let L be an artinian lattice, then we denote by K(L) the abelian
group with generators [a, b], a ≤ b, and relations
(4.9) [a, b] + [b, c] = [a, c]
(4.10) [a, a ∨ b] = [a ∧ b, b]
We let K + (L) ⊂ K(L) be the sub-semigroup generated by elements
[a, b], a < b. It is a direct consequence of Theorem 4.1 that K(L) (resp.
K + (L)) is the free abelian group (resp. semigroup) generated by the set
of equivalence classes of intervals of length 1 in L.

4.2. Harder–Narasimhan filtration and mass. Harder–Narasimhan

filtrations were originally defined for vector bundles on an algebraic
curve. The notion admits a straightforward generalization to modu-
lar lattices, which we include here for the sake of completeness and
to fix terminology. We also prove a triangle inequality for the notion
of mass coming from the HN filtration. Cornut [4] has also recently
studied Harder–Narasimhan filtrations in modular lattices by attaching
building-like spaces to them.
Consider a sub-semigroup of (C, +) of the form
(4.11) H = {reiφ | r > 0, φ ∈ I}
where I ⊂ R is a half-open interval of length π, e.g. I = [0, π). A
polarization on an artinian lattice L is a homomorphism Z : K(L) → C
SEMISTABILITY, MODULAR LATTICES, AND ITERATED LOGARITHMS
19

such that Z(K + (L)) ⊂ H. The pair (L, Z) is a polarized lattice. For
each a < b ∈ L we get a well-defined phase
(4.12) φ([a, b]) := Arg(Z([a, b])) ∈ I.
A polarized lattice is stable (resp. semistable) if
(4.13) x 6= 0, 1 =⇒ φ([0, x]) < φ(L) (resp. φ([0, x]) ≤ φ(L)).
Note that since Z(L) = Z([0, x]) + Z([x, 1]) one has φ([0, x]) < φ(L)
iff φ([x, 1]) > φ(L), which gives an equivalent condition for stability.
Theorem 4.2. Let L be a polarized lattice, then there is a unique
chain
0 = a0 < a1 < . . . < an = 1
such that [ak−1 , ak ] is semistable for k = 1, . . . , n and
φ([ak−1 , ak ]) > φ([ak , ak+1 ]).
The uniquely defined chain in the theorem above is the Harder–
Narasimhan filtration. (The terms chain and filtration are used in-
terchangeably here.)
Proof. We first show uniqueness, which does not require the finite
length hypothesis on L. Suppose
(4.14) 0 = a0 < a1 < . . . < am = 1, 0 = b0 < b1 < . . . < bn = 1
are Harder–Narasimhan filtrations. If n = 0, then 0 = 1 in L and so
m = 0 also. Otherwise, let k be such that b1 ≤ ak but b1  ak−1 . This
means that
(4.15) ak−1 < ak−1 ∨ b1 ≤ ak , ak−1 ∧ b1 < b1 .
By semistability of [b0 , b1 ] and [ak−1 , ak ] we get
(4.16)
φ([b0 , b1 ]) ≤ φ([ak−1 ∧ b1 , b1 ]) = φ([ak−1 , ak−1 ∨ b1 ]) ≤ φ([ak−1 , ak ])
hence, by the assumption on the slopes of the intervals, φ([b0 , b1 ]) ≤
φ([a0 , a1 ]). By symmetry we have equality, but then k = 1 in the above
argument and thus b1 ≤ a1 . Again, by symmetry, it must be that
a1 = b1 , so the proof follows by induction on max(m, n) applied to the
lattice L0 = [a1 , 1] = [b1 , 1].
Next we show existence, excluding the trivial case where 0 = 1. It
follows from the finite length hypothesis that the set of complex numbers
Z([0, a]), a > 0, is finite, so let
(4.17) φ := max{φ([0, a]) | a > 0}
and a1 be the join of all a > 0 with φ([0, a]) = φ. By construction
[0, a1 ] is semistable, and furthermore any interval [a1 , a], a1 < a, must
satisfy φ([a1 , a]) < φ by maximality. Thus, if the process is continued
inductively with [a1 , 1], then the φ([ak , ak+1 ]) are strictly decreasing.
q.e.d.
20 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH, P. PANDIT

If (L, Z) is a polarized lattice with HN-filtration a0 < a1 < . . . < an

then the mass of L is defined as
Xn
(4.18) m(L) = |Z([ak−1 , ak ])|.
k=1
It follows from the triangle inequality that
(4.19) m(L) ≥ |Z(L)|
with equality if and only if L is semistable. The mass satisfies the
following triangle inequality.
Theorem 4.3. If (L, Z) is a polarized lattice then
m(L) ≤ m([0, x]) + m([x, 1])
for any x ∈ L. More generally, by induction, if 0 = a0 < a1 < . . . <
an = 1 is any chain in L then
n
X
(4.20) m(L) ≤ m([ak−1 , ak ]).
k=1

Proof. First consider the case when [0, x] and [x, 1] are semistable.
Let φ1 = φ([0, x]) and φ2 = φ([x, 1]). If φ1 > φ2 , then 0 < x < 1 is
a HN filtration, and there is nothing to show. If φ1 = φ2 , then L is
semistable and m(L) = m([0, x]) + m([x, 1]). If φ1 < φ2 let 0 = a0 <
a1 < . . . < an = 1 be the HN filtration of L. In this case
(4.21) φ2 ≥ φ([a0 , a1 ]) > . . . > φ([an−1 , an ]) ≥ φ1 .
Indeed to see that φ2 ≥ φ([a0 , a1 ]) suppose x < x ∨ a1 , then by
semistablity
(4.22) φ2 = φ([x, 1]) ≥ φ([x, x ∨ a1 ]) = φ([x ∧ a1 , a1 ]) ≥ φ([0, a1 ])
and otherwise a1 ≤ x so φ([0, a1 ]) ≤ φ([0, x]) = φ1 < φ2 .
To show the inequality, let z = Z([0, x]) and w = Z([x, 1]) which form
an R-basis of C by assumption. If we write
(4.23) Z([ak−1 , ak ]) = λk z + µk w
then λk , µk ≥ 0 by the bound on the phases. Thus
(4.24) |Z([ak−1 , ak ])| ≤ λk |z| + µk |w|
and taking the sum over all k we get
(4.25) m(L) ≤ |z| + |w| = m([0, x]) + m([x, 1])
P P
since k λk = k µk = 1.
The general case is equivalent to the claim that if 0 = a0 < a1 < . . . <
an = 1 is any chain in L with [ak−1 , ak ] semistable, then
Xn
(4.26) m(L) ≤ |Z([ak−1 , ak ])| =: M
k=1
SEMISTABILITY, MODULAR LATTICES, AND ITERATED LOGARITHMS
21

since we get such a chain by concatenating the HN filtrations of [0, x] and

[x, 1]. The strategy is to modify the chain step-by-step until it becomes
the HN filtration, with M getting smaller each time.
If there are two consecutive intervals in the chain with the same phase,
then they can be combined to a single interval, decreasing the length of
chain by one. If after this the chain ak is not the HN filtration, then
there must be consecutive intervals with
(4.27) φ([ak−1 , ak ]) < φ([ak , ak+1 ]).
If ak−1 < ak < ak+1 is replaced by the HN filtration of [ak−1 , ak+1 ], then
by the first part of the proof M either gets strictly smaller or the length
of the chain stays the same. Either way we must eventually reach the
HN filtration, since the possible values of M form a discrete subset of
R≥0 , and if M remains constant then the phases will eventually be in
the right order. q.e.d.

4.3. Paracomplemented R–filtrations. Let L be an artinian lattice.

An R-filtration in L is a strictly increasing sequence of elements in L
ending with 1 and labeled by real numbers. The following notation
will be convenient. Given a finite subset X of R let I0 , . . . , In be the
connected components of the complement R \ X in their natural order.
Any chain
(4.28) 0 = a0 < a1 < . . . < an = 1
in L defines a locally constant increasing function a : R \ X → L. Let
a+ , a− : R → L be the upper-/lower-semicontinuous extensions of a,
then we call this pair of increasing functions an R-filtration in L. Thus
an R-filtration in L is a pair of increasing functions a± : R → L with
a+ upper-semicontinuous, a− lower semicontinuous, a+ = a− outside a
finite set, and a± (λ) = 0 for λ  0 and a± (λ) = 1 for λ  0. Of course
any one of a± determines the other, but it will be convenient to have
both. The support of an R-filtration is the finite set
(4.29) supp(a) = {λ ∈ R | a+ (λ) 6= a− (λ)}.
A lattice L with 0, 1 is complemented if any a ∈ L has a comple-
ment: An element b ∈ L with
(4.30) a ∧ b = 0, a ∨ b = 1.
Note that for the lattice of subobjects in a given object E of an artinian
category, the property being complemented means that E is semisimple.
We call an R-filtration, a, paracomplemented if all the intervals
[a+ (λ), a+ (λ + 1)], λ ∈ R, are complemented lattices. Equivalently, all
intervals [a− (λ), a− (λ + 1)] are complemented lattices.
Let B(L) be the set of all paracomplemented R-filtrations in L. De-
note by C0 (R; K(L)) the abelian group of finite K(L)-linear combina-
tions of points in R, with the obvious topology coming from R. We can
22 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH, P. PANDIT

introduce a topology in B(L) such that the map

X
(4.31) cl : B(L) → C0 (R; K(L)), a 7→ [a− (λ), a+ (λ)]λ
λ∈R

is continuous. A neighborhood basis at a ∈ B(L), is given by sets

(4.32) Ur (a) = {b ∈ B(L) | dist(λ, supp(a)) ≥ r =⇒ a± (λ) = b± (λ)}
where r > 0. This topology is Hausdorff, but generally not locally
compact for infinite L.
We will describe the local structure around a ∈ B(L) in terms of an-
other artinian lattice, Λ(a). By definition, an element x ∈ Λ(a) is given
by xλ ∈ [a− (λ), a+ (λ)] such that a+ (λ) ∈ [xλ , xλ+1 ] has a complement
for every λ ∈ R. Thus, Λ(a) is a subset of
Y
(4.33) [a− (λ), a+ (λ)]
λ∈R
which is an artinian lattice as an essentially finite product of such. How-
ever, it is not obvious that Λ(a) is closed under ∨ and ∧, i.e. is a
sublattice. Showing this will require a lemma about complements.
We draw a diagram
d

(4.34) a b

c
to represent the statement that c = a ∧ b and d = a ∨ b, i.e. that a has
complement b in [c, d]. These diagrams satisfy cut and paste rules:
d e e
(4.35) a b and d f =⇒ a f

c b c
(4.36)
d a∨x d
a b and c ≤ x ≤ b =⇒ a x a∨x b
c c x
In the following we will not draw all the diagrams for practical reasons,
but they proved a useful device to avoid getting lost in formulas.
Lemma 4.4. Let L be a modular lattice with 0. Suppose x, a, b ∈ L
such that x ≤ a ∧ b, x has a complement in [0, a] and [0, b], and [x, a] is
complemented, then x has a complement in [0, a ∨ b].
SEMISTABILITY, MODULAR LATTICES, AND ITERATED LOGARITHMS
23

Proof. Let a0 be a complement of a∧b in [x, a], then we have a diagram

a∨b
(4.37) a0 b
x

and x has a complement c in [0, a0 ]. If d is a complement of x in [0, b],

then we have a diagram

a∨b
(4.38) c∨d b

which shows that c ∨ d is a complement of x in [0, a ∨ b]. q.e.d.

Q
Proposition 4.5. If a ∈ B(L) then Λ(a) ⊂ [a− (λ), a+ (λ)] is a
sublattice, hence an artinian lattice.

Proof. Suppose x, y ∈ Λ(a), then a+ (λ) has a complement in both

[xλ ∨ yλ , xλ+1 ] and [xλ ∨ yλ , yλ+1 ], and [a+ (λ), xλ+1 ] is complemented.
By the lemma, a+ (λ) has a complement in [xλ ∨ yλ , xλ+1 ∨ yλ+1 ], hence
x ∨ y ∈ Λ(a). By the dual argument, Λ(a) is also closed under ∧. q.e.d.

The defining condition for elements in Λ(a) can be reformulated.

Q
Lemma 4.6. Let a ∈ B(L), x ∈ [a− (λ), a+ (λ)], λ ∈ R, then the
following are equivalent:
1) a+ (λ) has a complement in [xλ , xλ+1 ].
2) a− (λ + 1) has a complement in [xλ , xλ+1 ].
3) [xλ , xλ+1 ] is complemented.
These conditions hold for all λ ∈ R if and only if x ∈ Λ(a).

This is a direct consequence of the following lemma.

Lemma 4.7. Let L be a bound modular lattice, x ∈ L such that [0, x]

and [x, 1] are complemented, and x has a complement in L. Then L is
complemented.

Proof. Let a ∈ L, b a complement of a ∧ x ∈ [0, x], c a complement of

a ∨ x ∈ [x, 1], and d a complement of x in L, then we have the following
24 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH, P. PANDIT

diagram.
1

a∨x c
(4.39) a x b ∨ (c ∧ d)

a∧x b

0
This shows that a has complement b ∨ (c ∧ d) in L. q.e.d.
Note that the defining condition for x to be in Λ(a) only relates xλ
and xλ+1 , so xλ and xµ are completely independent if λ − µ is not an
integer. Hence Λ(a) splits as a product
Y
(4.40) Λ(a) = Λτ (a)
τ ∈R/Z
Y
(4.41) Λτ (a) ⊂ [a− (λ), a+ (λ)].
λ∈τ

For a ∈ B(L) define

(4.42) ρ1 (a) = min {|λ − µ| | λ 6= µ, λ, µ ∈ supp(a)}
(4.43) ρ2 (a) = min {|λ − µ| − 1 | |λ − µ| > 1, λ, µ ∈ supp(a)}
1
(4.44) ρ(a) = min(ρ1 (a), ρ2 (a), 1) > 0.
2
The following gives a local description of B(L).
Proposition 4.8. Let a ∈ B(L), ρ = ρ(a), then there is a canonical
bijection between Uρ (a) and the set of R-filtrations in Λ(a) with support
in (−ρ, ρ).
Proof. Denote the set of R-filtrations in Λ(a) with support in (−ρ, ρ)
by V . The map Uρ (a) → V sends b ∈ Uρ (a) to the R-chain x ∈ Λ(a)
with
(4.45) x± (α)λ = b± (α + λ) ∈ [a− (λ), a+ (λ)]
for α ∈ [−ρ, ρ] and λ ∈ supp(a). To see that x± (α) ∈ Λ(a) note that if
λ, λ + 1 ∈ supp(a) then
(4.46) [x+ (α)λ , x+ (α)λ+1 ] = [b+ (α + λ), b+ (α + λ + 1)]
is complemented since b is paracomplemented by assumption. For this
part we only used ρ ≤ ρ1 /2, not ρ ≤ ρ2 /2.
The inverse map V → Uρ (a) sends x ∈ V to b ∈ Uρ (a) with the
same relation (4.45). We need to check that b is paracomplemented. So
SEMISTABILITY, MODULAR LATTICES, AND ITERATED LOGARITHMS
25

suppose that
(4.47) α + λ + 1 = α 0 + λ0 , λ, λ0 ∈ supp(a), α, α0 ∈ (−ρ, ρ).
Because of ρ ≤ 1/2 we have λ < λ0 also. We need to show that
(4.48) [b+ (α + λ), b+ (α0 + λ0 )] = [x+ (α)λ , x+ (α0 )λ0 ]
is complemented. If λ0 −λ > 1 the by definition of ρ we get 2ρ ≤ λ0 −λ−1
hence
(4.49) α + λ + 1 < ρ + λ + 1 ≤ −ρ + λ0 < α0 + λ0
which is a contradiction, thus λ0 − λ ≤ 1. If λ0 − λ < 1, then
(4.50) [x+ (α)λ , x+ (α0 )λ0 ] ⊂ [a− (λ), a+ (λ0 )]
which is complemented since a is paracomplemented. Otherwise λ0 =
λ + 1 so α = α0 , but then the interval is complemented because x+ (α) ∈
Λ(a). q.e.d.
If a ∈ B(L) and x is an R-filtration in Λ(a) with support in (−ρ, ρ),
ρ = ρ(a), and b ∈ B(L) corresponds to x, then Λ(b) splits as a product
Y
(4.51) Λ(b) = [x− (λ), x+ (λ)]
λ∈R
which follows from (4.40) and the definition of ρ. Essentially, as a is
deformed to b classes of the support in R/Z = S 1 split but do not
collide.

4.4. Weight filtrations. In this section we define a weight-type filtra-

tion in any finite length modular lattice by proving an existence and
uniqueness theorem.
Let L be an artinian lattice and let X : K(L) → R be a homomor-
phism with X(K + (L)) ⊂ R>0 . For any a ∈ B(L) the lattice Λ(a) has a
canonical polarization given by
X
(4.52) Z([x, y]) = (1 + λi)X([xλ , yλ ])
λ∈supp(a)

for x, y ∈ Λ(a), x ≤ y. The main result of this section is the following.

Theorem 4.9. Let L be an artinian lattice and X : K + (L) → R>0
a semigroup homomorphism. Then there exists a unique a ∈ B(L) such
that Λ(a) is semistable with phase φ(Λ(a)) = 0.
We call the paracomplemented R-filtration in L which is uniquely
determined by the theorem the weight filtration in L.
The theorem stated in the introduction is the special case where L is
the lattice of subobjects of a fixed object E ∈ A in an artinian abelian
category A. Then K + (L) is the sub-semigroup of K(A) generated by
simple objects which appear as sub-quotients of E and we can obtain
X as in the theorem above by restriction. The first condition of the
26 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH, P. PANDIT

theorem in the introduction is that the filtration is paracomplemented,

the second that φ(Λ(a)) = 0, and the third that Λ(a) is semistable.

Proof. Uniqueness. Suppose a and b are R-filtrations in L. Combine

these to
(4.53) c+ (α, β) := a+ (α) ∧ b+ (β)
(4.54) c− (α, β) := a+ (α) ∧ b+ (β) ∧ (a− (α) ∨ b− (β)).
We claim that a = b if and only if c+ (α, β) = c− (α, β) for all α 6= β. In
one direction, if a = b, α < β say, then
(4.55) a+ (α) ∧ a+ (β) ∧ (a− (α) ∨ a− (β)) = a+ (α) = a+ (α) ∧ a+ (β).
On the other hand, if a 6= b then there is an α ∈ R with a− (α) = b− (α)
but a+ (α) 6= b+ (α). By symmetry, we may assume that a+ (α)  b+ (α),
so there is a β > α with
(4.56) a+ (α) ≤ b+ (β), a+ (α)  b− (β).
We have
(4.57)
a+ (α)∧b+ (β) = a+ (α) > a+ (α)∧b− (β) = a+ (α)∧b+ (β)∧(a− (α)∨b− (β))
thus c+ (α, β) > c− (α, β).
Now suppose a, b ∈ B(L) are both semistable of phase 0, and a = 6 b
for contradiction. Then there are α 6= β with c− (α, β) < c+ (α, β) and
we may assume by symmetry that such a pair exists with α > β. Let
δ > 0 be maximal such that there exists an α with with
(4.58) c− (α, α − δ) < c+ (α, α − δ)
i.e. the most off–diagonal. Such a δ exists because of finiteness of the
filtrations. We claim that
(4.59) xα = a− (α) ∨ b+ (α − δ)
defines an element of Λ(a). First, by choice of δ, we have b+ (α − δ) ≤
a+ (α) thus xα ∈ [a− (α), a+ (α)]. We need to show that a+ (α) has a
complement in [xα , xα+1 ]. Consider the following diagrams

a+ (α) ∨ b+ (α − δ + 1)
a− (α + 1)
a+ (α) b+ (α − δ + 1)
(4.60) a+ (α) e
a+ (α) ∧ b+ (α − δ + 1) d
a− (α) ∨ b+ (α − δ) = xα
b+ (α − δ)
SEMISTABILITY, MODULAR LATTICES, AND ITERATED LOGARITHMS
27

a+ (α) ∨ b+ (α − δ + 1)

(4.61) a+ (α) a− (α) ∨ d

a− (α) ∨ b+ (α − δ) = xα
where existence of complements e, d follows from the assumption that
a, b are paracomplemented, and the third is obtained from the first.
Since
(4.62) a+ (α) ∨ a− (α + 1) ∨ b+ (α − δ + 1) = xα+1
the claim follows from Lemma 4.4.
In a similar way one shows that
(4.63) yβ = a− (β + δ) ∧ b+ (β)
defines an element y ∈ Λ(b). We compute
X
(4.64) ImZ([0, x]) = αX([a− (α), xα ])
α
X
(4.65) > βX([a− (β + δ), xβ+δ ])
β
X
(4.66) = βX([a− (β + δ) ∧ b+ (β), b+ (β)])
β
X
(4.67) = β(X([b− (β), b+ (β)]) − X([b− (β), yβ ]))
β
(4.68) = 0 − ImZ([0, y])
which implies that at least one of ImZ([0, x]), ImZ([0, y]) is positive.
This contradicts the assumption that both Λ(a) and Λ(b) are semistable.
Existence. Consider the function
(4.69) m : B(L) → R, a 7→ m(Λ(a))
sending a paracomplemented R-filtration, a, to the mass of the associ-
ated lattice Λ(a). By (4.19) we have
(4.70) m(Λ(a)) ≥ |Z(Λ(a))| ≥ ReZ(Λ(a)) = X(L)
with equality if and only if Λ is semistable of phase 0, i.e. the weight
filtration.
We claim that if a ∈ B(L) is a local minimum of m, then a is a weight
filtration, thus a global minimum. Suppose x0 < x1 < . . . < xn is the
HN filtration in Λ(a), φk := φ([xk−1 , xk ]). We want to show that n = 1
and φ1 = 0 if a is a local minimum. The idea is to deform a using its HN
filtration. Let t > 0 and consider the R-filtration in Λ(a) with support
(4.71) − φ1 t < . . . < −φn t
28 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH, P. PANDIT

and values x0 < x1 < . . . < xn . For sufficiently small t > 0 this
R-filtration has support in (−ρ(a), ρ(a)), so let at ∈ B(L) be the corre-
sponding paracomplemented R-filtration given by Proposition 4.8. We
have at → a as t → 0. The mass of at is given by
n
X
(4.72) m(Λ(at )) = |zk (t)|
k=1
X
(4.73) zk (t) := (1 + (λ − φk t)i)X([xk−1,λ , xk,λ ])
λ∈R

which also gives the mass of a for t = 0. Note that Re(zk (t)) is indepen-
dent of t and
dIm(zk ) X
(4.74) = −φk X([xk−1,λ , xk,λ ])
dt
λ∈R

which has the opposite sign of φk , if φk 6= 0. But φk = Arg(Z([xk−1 , xk ]))

has the same sign as
(4.75) Im(zk (0)) = Im(Z([xk−1 , xk ]))
hence a cannot be a local minimum unless φk = 0 for all k, i.e. n = 1
and Λ(a) is semistable of phase 0.
In preparation for what follows, we want to show that there is a C > 0
such that
(4.76) max{|λ|, λ ∈ supp(a)} ≤ Cm(Λ(a))
for any a ∈ B(L). The argument is that the cardinality of supp(a) is
bounded above by the length, n, of L, so if the diameter of supp(a)
becomes larger than n − 1, then there is a gap of length > 1 and Λ(a)
splits as a product corresponding to points on the left and right of the
gap. Thus, if the left hand side of (4.76) is larger than n − 1, then there
must be a factor of Λ(a) (possibly everything) supported entirely on one
side of 0 ∈ R. The mass of this factor is bounded above by Xmin times
the distance of its support to 0, where Xmin is the minimum of X on
K + (L). Note also that since m is bounded below by a positive constant,
any additive constant in the estimate can be absorbed into C.
Now recall from (4.31) that there is continuous map
X
(4.77) cl : B(L) → C0 (R; K(L)), a 7→ [a− (λ), a+ (λ)]λ
λ∈R

whose image is contained in the homology class in H0 (R; K(L)) = K(L)

given by [L]. By (4.76) the infimum of m : B(L) → R stays the same
if we restrict to a subset V ⊂ B(L) given by R-filtrations supported
in [−M, M ] for some sufficiently large M  0. The image of V under
cl is contained in the set W of 0-chains supported in [−M, M ], with
SEMISTABILITY, MODULAR LATTICES, AND ITERATED LOGARITHMS
29

coefficients in K + (L), and with class [L] ∈ H0 (R; K(L)), which is com-
pact. In fact, cl(V ) ⊂ W is closed, hence compact. To see this, suppose
xn ∈ V with cl(xn ) → y ∈ W . If Y = supp(y) let
(4.78)
1
κ = min{|λ − µ| | λ, µ ∈ Y } ∪ {1 − |λ − µ| | λ, µ ∈ Y, |λ − µ| < 1} > 0
2
There is some N such that every point in supp(xN ) has distance less
than κ from supp(y). From xN we get a coarser R-filtration x with
support Y and values
(4.79) x± (λ ± κ) = xN,± (λ ± κ), λ∈Y
which is paracomplemented by definition of κ and satisfies cl(x) = y.
We claim that m takes only finitely many values on each fiber of cl.
Indeed, if a0 < . . . < an is the HN filtration of a ∈ B(L), then m(Λ(a))
only depends on the partition
(4.80) cl(a) = [a0 , a1 ] + . . . + [an−1 , an ]
of cl(a) into 0-chains with positive coefficients, and there are only finitely
many such partitions. Taking fiberwise minimum of m gives a function
(4.81) f : cl(V ) → R, f (x) = min{m(a) | cl(a) = x}.
Since it has already established that cl(V ) is compact, we can conclude
that m has a global minimum, and thus the existence of a weight filtra-
tion, if we show that f is lower semicontinuous.
Let x ∈ cl(V ), then ρ = ρ(a) is the same for all a with cl(a) = x, since
it only depends on the support. After possibly shrinking ρ we also have
ρ ≤ κ(x), where κ(x) = κ is defined as in (4.78) with Y = supp(x). Let
Oρ be the neighborhood of x consisting of 0-chains which differ from x
by a 1-chain with support in a ρ-neighborhood of supp(x). This is in
complete analogy with the definition of Uρ (a) for a ∈ B(L), and we get
[
(4.82) cl−1 (Oρ ) = Uρ (a)
cl(a)=x

where the inclusion ⊇ is clear and the inclusion ⊆ follows from ρ ≤ κ

by the same argument which showed that cl(V ) is closed.
Suppose b ∈ Uρ (a) corresponds to an R-filtration w, then by the
triangle inequality for mass, Theorem 4.3, and (4.51) we get
X
(4.83) m(a) ≤ m([w− (λ), w+ (λ)]).
λ∈R

Let cλ,0 < . . . < cλ,nλ be the HN filtration in [w− (λ), w+ (λ)], then
nλ X
X
(4.84) m([w− (λ), w+ (λ)]) = (1 + µi)X([cλ,k−1,µ , cλ,k,µ ])
k=1 µ∈R
30 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH, P. PANDIT

where [w− (λ), w+ (λ)] gets its polarization from Λ(a) and
nλ X
XX
(4.85) m(b) = (1 + (λ + µ)i)X([cλ,k−1,µ , cλ,k,µ ]) .
λ∈R k=1 µ∈R

The difference between the right hand side of (4.83) and (4.85) can be
made smaller than some given ε by suitable choice of ρ, which does not
depend on the particular a or b but only a partition of x, of which there
are finitely many. This shows that f is lower semicontinuous. q.e.d.
Besides the weight filtration, any artinian lattice has two other canon-
ically defined filtrations 0 = a0 < a1 < . . . < an = 1 such that the
intervals [ak−1 , ak ] are complemented lattices. The socle filtration is de-
fined inductively by the property that ak ∈ [ak−1 , an ] is maximal such
that [ak−1 , ak ] is complemented. Dually, the cosocle filtration is defined
inductively by the property that ak−1 ∈ [a0 , ak ] is minimal such that
[ak−1 , ak ] is complemented. Both are examples of a Loewy filtration:
A filtration of minimal length such that [ak−1 , ak ] are complemented
lattices. These filtrations are typically considered in the context of rep-
resentations of finite–dimensional algebras, see for example [1].

4.5. Iterated weight filtration. If (L, Z) is a semistable polarized

lattice, then we can consider the subset L0 ⊂ L given by
(4.86) L0 = {x ∈ L | x = 0 or φ([0, x]) = φ(L)}
which is a sublattice, hence artinian and there is a homomorphism
(4.87) X : K(L0 )+ → R>0 , X([x, y]) = e−iφ(L) Z([x, y]).
Moreover, L0 has strictly smaller length than L, unless the image of Z
is contained in a single ray. If L0 is complemented, then L is called
polystable.
We apply the above to the following situation. Suppose L is an ar-
tinian lattice with homomorphism X : K + (L) → R>0 and let a ∈ B(L)
be the weight filtration in L. By definition, Λ(a) is semistable, so we
can consider L(2) = Λ(a)0 which has a weight filtration b. The filtra-
tion b gives a filtration in Λ(a) ⊃ L(2) , hence a refinement of a to an
R2 -filtration a(2) with
(2)
(4.88) a± (λ1 , λ2 ) = b± (λ2 )λ1 ∈ [a− (λ1 ), a+ (λ1 )]
where R2 is given the lexicographical order. By induction we get lattices
L(n) and Rn -filtration a(n) . The lengths of L(n) are strictly decreasing
until some L(N +1) is complemented and thus its weight filtration trivial,
so the process stops after finitely many steps. This shows that there is
a canonical R∞ -filtration in L, the iterated weight filtration, defined
to be a(N ) . We refer to N as the depth of the iterated weight filtration.
SEMISTABILITY, MODULAR LATTICES, AND ITERATED LOGARITHMS
31

• •

• •

• • • •

• • • •

• • • • • • • •

• • • • • • • •

Figure 1. The graphs G(1) , G(2) , G(3) , G(4) .

We will construct a series of examples generalizing the one in Section 2

to show that the depth can be any non-negative integer. The lattices
will be obtained as lattices of closed subgraphs of oriented trees with the
canonical homomorphism X : K + (L) → R>0 given by the length of an
interval.
Let G(0) be the graph with a single vertex and no edges and G(1) be
the directed graph with two vertices and a single arrow between them.
Inductively define G(n+1) to be the directed graph obtained from G(n)
by adding an outgoing arrow from each source to a new vertex and an
incoming arrow to each sink starting at a new vertex. More formally,
(n+1) (n+1)
define vertices G0 = G(n) × {0, 1} and arrows G1 to include
(n)
(i, 0) → (i, 1) for each i ∈ G0 and (i, 0) → (j, 1) for each arrow i → j
(n)
in G1 (see Figure 1).
The weight grading on G(n) is just vi = 12 if i is a source and vi = − 21
if i is a sink. This follows from Lemma 3.1 with Lagrange multipliers
uα = 12 if α is a new arrow in G(n) and uα = 0 otherwise. To compute
the iterated weight filtration we should next look at the lattice L(2) of
closed subgraphs of G(n) such that the sum of vi is zero, i.e. which
include an equal number of sinks and sources. It is easy to see that this
coincides with the lattice of closed subgraphs of G(n−1) . This shows that
the iterated weight filtration on the lattice of closed subgraphs of G(n)
has depth n.
32 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH, P. PANDIT

5. Gradient flow on quiver representations

The purpose of this section is to show that the iterated weight fil-
tration has a dynamical interpretation, describing the asymptotics of
certain gradient flows which appear in the study of quiver representa-
tions. We start by providing background on the Kähler geometry of
spaces of quiver representations in the first subsection. Subection 5.2
gives an alternative description of the flow in the language of ∗-algebras
and ∗-bimodules, which is more invariant and simplifies formulas. Gen-
eral properties of the flow are discussed in Subsection 5.3. The final
subsection completes the proof of our second main theorem by giving a
construction of asymptotic solutions.

5.1. Kähler geometry of quiver representations. Many problems

in linear algebra are instances of the following general one. Given a
quiver
Q1
s t
(5.1)
Q0 Q0

where Q0 is the set of vertices, Q1 the set of arrows, and s and t assign to
each arrow its starting and target vertex, classify all the ways in which
such a diagram can be realized (represented) using finite-dimensional
vector spaces and linear maps. The space of representations for fixed
vector spaces Ei , i ∈ Q0 , is a quotient
M  Y
(5.2) Hom(Ei , Ej ) GL(Ei ) =: V /G
α:i→j i∈Q0

of a vector space by a reductive group.

If the ground field is C then V /G is approximated by a Kähler man-
ifold. To construct it, choose a Hermitian metric on each Ei , then the
norm-squared
X
(5.3) S(φ) = tr (φ∗α φα )
α:i→j

where φα ∈ Hom(Ei , Ej ), is a Kähler potential for the flat metric on

V . We can look for points in V which minimize S on a given G-orbit.
These are representations with
X
(5.4) [φ∗α , φα ] = 0.
α:i→j

Such a minimum can be found if and only if the G-orbit corresponds to

a semisimple representation. This is an application of the Kempf–Ness
SEMISTABILITY, MODULAR LATTICES, AND ITERATED LOGARITHMS
33

theorem. A Kähler manifold is then obtained as the quotient of the set

of minimizers by the unitary subgroup K ⊂ G preserving the metric on
each Ei , with the potential which is the restriction of S.
If Q has no oriented cycles then the only semisimple representations
are those with φ = 0. Following ideas from geometric invariant theory
A. King [12] shows how to obtain non-trivial spaces by generalizing the
above construction. They depend on a choice of polarization, which is
in this context just a real number θi ∈ R for each vertex i ∈ Q0 . They
allow us to extend the action of G to V × C by letting g ∈ G act on
z ∈ C by multiplication with
Y
(5.5) (det gi )θi .
i∈G0

(Strictly speaking, this is ill-defined if θi are not integers and we should

work with virtual line bundles.) On V × C∗ consider the potential
X
(5.6) S(φ, z) = tr (φ∗α φα ) + log |z|.
α∈Q1

Fixing φ ∈ V , we can consider the (K-invariant) restriction of S to

the orbit G(φ, 1) as a function on the homogeneous space G/K. A
point in G/K corresponds to a choice of positive definite Hermitian
endomorphism hi on each Ei , and
X
X
(5.7) S(h) = tr h−1 ∗
i φ α hj φ α + θi log det hi .
α:i→j i∈Q0

The equation for h ∈ G/K to be a critical point of S is

X X
(5.8) [h−1
i φ ∗
α h j , φα ] = θi prEi .
α:i→j i∈Q0

To describe those representations for which the above equation has a

solution, we need to recall some terminology. For any representation E
of Q we define
X
(5.9) θ(E) = θi dim Ei
i∈Q0

and say that E is semistable if θ(E) = 0 and any subrepresentation

F ⊂ E satisfies θ(F ) ≤ 0. If in addition θ(F ) < 0 whenever 0 6= F ( E,
then E is called stable. Finally, E is polystable if it is a direct sum
of stable representations. Note that for θ = 0 all representations are
semistable and polystable=semisimple.
Theorem 5.1 (King). S is bounded below on the G-orbit through
(φ, z) ∈ V × C if and only if φ defines a semistable representation.
Moreover, there is a solution to (5.8), i.e. a minimum point of S, if and
only if the representation is polystable.
34 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH, P. PANDIT

The set of polystable representations (up to isomorphism) thus has

the structure of a Kähler manifold. (More precisely a stratified Kähler
manifold, see [16].)
From a dynamical point of view, polystability means that the gradient
flow of S on G/K has the simplest possible asymptotics: exponentially
fast convergence to a fixed point, which is a solution of (5.8). One can
study the asymptotic behavior of the flow for non-polystable represen-
tations and see if this yields more information about V /G.
To define a gradient of S we need to choose a Riemannian metric on
G/K. We consider metrics of the form
X
mi tr h−1 −1


(5.10) hv, wi = i vhi w , v, w ∈ Th (G/K)
i∈Q0
where mi > 0, i ∈ Q0 , are some fixed positive numbers. The negative
gradient flow is then
dhi X X
(5.11) mi h−1
i = h−1 ∗
i φα hj φα − φα h−1 ∗
j φα hi − θi .
dt
α:i→j α:j→i
We will show in this section that in the semistable case the asymptotics
of this flow are completely described by the iterated weight filtration.
More precisely, on the Eλ piece of the filtration, λ = (λ1 , . . . , λn ), we
have
(5.12) log(|h(t)|) = λ1 log t + λ2 log log t + . . . + λn log(n) t + O(1).
5.2. Star-algebras and bimodules. In order to simplify formulas like
(5.11) and all calculations below, it is useful to adopt the more invariant
language of ∗-algebras and ∗-bimodules. This offers perhaps also a more
algebraic point of view on the Kähler geometry discussed in the previous
subsection. To motivate the general definitions below, we first describe
the structure in the case of quiver representations.
To begin, note that
Y
(5.13) B := End(Ei )
i∈Q0
is a finite-dimensional ∗
C -algebra, with ∗-structure determined by the
choice of metrics on the vector spaces Ei . It follows from the classifi-
cation of type I factors, or more directly using the Artin–Wedderburn
theorem, that every finite-dimensional C ∗ -algebra is of this form. Finite-
dimensional C ∗ -algebras are also precisely those ∗-algebras which have
a faithful finite-dimensional ∗-representation on an inner product space.
Recall that a ∗-algebra over C is a C-algebra, A, together with a map
A → A, a 7→ a∗ such that
(5.14)
a∗∗ = a, (a + b)∗ = a∗ + b∗ , (λa)∗ = λa∗ , (ab)∗ = b∗ a∗
for a, b ∈ A, λ ∈ C.
SEMISTABILITY, MODULAR LATTICES, AND ITERATED LOGARITHMS
35

The masses mi > 0, i ∈ Q0 , determine a positive trace

X
(5.15) τ : B → C, b 7→ mi tr(bi ).
i∈Q0

Functionals τ obtained in such a way are characterized by the properties

(5.16) τ (a∗ ) = τ (a), τ (ab) = τ (ba), τ (aa∗ ) > 0 for a 6= 0
which imply that
(5.17) ha, bi := τ (a∗ b)
defines a Hermitian inner product on A.
The space of representations
M
(5.18) M := Hom(Ei , Ej )
α:i→j

has the structure of a B–B bimodule. Additionally, there are two B-

valued inner products
X 1
(5.19) (φ, ψ) 7→ φψ ∗ = φα ψα∗
mj
α:i→j
X 1
(5.20) (φ, ψ) 7→ φ∗ ψ = φ∗ ψα
mi α
α:i→j

where the normalization is chosen so that

(5.21) τ (φψ ∗ ) = τ (φ∗ ψ).
More generally, suppose A, B are finite dimensional C ∗ -algebras with
trace. If M is an A–B bimodule, then M is the complex conjugate
vector space with identity map M → M , m 7→ m∗ and B–A bimodule
structure given by
(5.22) bm∗ a := (a∗ mb∗ )∗
for a ∈ A, b ∈ B, m ∈ M . We say M is a ∗-bimodule if it is equipped
with homomorphisms of bimodules
(5.23) M ⊗B M → A, m ⊗ n∗ 7→ mn∗
(5.24) M ⊗A M → B, m∗ ⊗ n 7→ m∗ n
which are algebra-valued inner products on M in the sense that
(5.25) (mn∗ )∗ = nm∗ , mm∗ ≥ 0, mm∗ = 0 =⇒ m = 0
(5.26) (m∗ n)∗ = n∗ m, m∗ m ≥ 0, m∗ m = 0 =⇒ m = 0
and are related by
(5.27) τA (mn∗ ) = τB (n∗ m).
Caution: In general one has (ab∗ )c 6= a(b∗ c) for a, b, c ∈ M .
36 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH, P. PANDIT

The finite-dimensional B–B ∗-bimodules are, up to isomorphism, all

obtained as above from quivers. The following table summarizes our
setup and the dictionary between the two languages.
notation type in terms of quiver Q
finite-dim. C ∗ -algebra
L
B i∈Q0 End(Ei )
P
τ trace B → C i∈Q0 mi tr(bi ), mi > 0

ρ ρ ∈ center(B), ρ = ρ∗ ρi = θi /mi ∈ R, i ∈ Q0
L
M B–B ∗-bimodule α:i→j Hom(Ei , Ej )

φ element of M φα : Ei → Ej , α : i → j
For example the equation (5.11) for the flow now takes the form
dh
(5.28) h−1 = [h−1 φ∗ h, φ] − ρ
dt
where h ∈ B moves in the cone
(5.29) P := {h ∈ B | h∗ = h, Spec(h) ∈ (0, ∞)} ⊂ B
of self-adjoint operators with strictly positive spectrum (which was writ-
ten as G/K before).
For the remainder of this subsection we show how to obtain from a
triple (B, τ, M ) a new one (B 0 , τ 0 , M 0 ) by deforming (“twisting”) along
an element φ ∈ M .
Lemma 5.2. Let φ ∈ M , then the adjoint of [φ, _] : B → M is
[φ∗ , _] : M → B
Proof.
(5.30) h[φ, b], mi = τ ([b∗ , φ∗ ]m)
(5.31) = τ (b∗ φ∗ m − φ∗ b∗ m)
(5.32) = τ (b∗ [φ∗ , m])
(5.33) = hb, [φ∗ , m]i
Note the use of (5.27). q.e.d.
The kernel of [φ, _] is a subalgebra in general, but it need not be closed
under the operation ∗. The following proposition states that, under the
condition of centrality of [φ∗ , φ], passing to the “harmonic part” of the
complex B → M produces another pair (B 0 , M 0 ) of the same sort.
Proposition 5.3. Suppose that
(5.34) [φ∗ , φ] ∈ center(B)
and let
(5.35) B 0 = {b ∈ B | [φ, b] = 0}, M 0 = {m ∈ M | [φ∗ , m] = 0}
SEMISTABILITY, MODULAR LATTICES, AND ITERATED LOGARITHMS
37

then B 0 is a ∗-subalgebra of B and M 0 a B 0 –B 0 ∗-bimodule with B 0 -valued

inner products given by composition of those of M with the orthogonal
projection B → B 0 .
Proof. We have
(5.36) h[φ, b], [φ, b]i = hb, [φ∗ , [φ, b]]i
(5.37) = hb, [φ, [φ∗ , b]]i
(5.38) = h[φ∗ , b], [φ∗ , b]i
but [φ∗ , b] = −[φ, b∗ ]∗ , so [φ, b] = 0 implies [φ, b∗ ] = 0. Thus, B 0 is a
∗-subalgebra.
If b ∈ B 0 , m ∈ M 0 , then
(5.39) [φ∗ , bm] = [φ∗ , b]m + b[φ∗ , m] = 0
so bm ∈ M 0 and similarly mb ∈ M 0 . Thus M 0 is a B 0 –B 0 bimodule.
Next, let P : B → B 0 be the orthogonal projection. By Lemma 5.2 it
is characterized by P (b) ∈ B 0 and P (b) − b = [φ∗ , m] for some m ∈ M .
This also shows that τ (P (b)) = τ (b). We claim that if a ∈ B 0 , b ∈ B,
then P (ab) = aP (b). To see this, let P (b) − b = [φ∗ , m], then
(5.40) aP (b) − ab = a[φ∗ , m] = [φ∗ , am].
As a consequence, we see that
(5.41) M 0 ⊗ M 0 → B0, m ⊗ n∗ 7→ P (mn∗ )
(5.42) M 0 ⊗ M 0 → B0, m∗ ⊗ n 7→ P (m∗ n)
are maps of bimodules. Also, (5.27) for M 0 follows from the correspond-
ing identity for M and τ (P (b)) = τ (b). q.e.d.

5.3. Monotonicity and homogeneity. A key property of the flow

(5.28), for our purposes, is a certain kind of monotonicity.
Proposition 5.4 (Monotonicity). Let h1 (t), h2 (t) be solutions of (5.28)
with h1 (0) ≤ h2 (0), then h1 (t) ≤ h2 (t) for all t ≥ 0.
Proof. Consider
(5.43) A := {(g, h) ∈ P × P | g ≤ h}
which is a manifold with corners. To prove the proposition it suffices to
show that the flow on pairs (g, h) ∈ P × P is pointing inwards or in a
tangential direction on the boundary ∂A, which is the subset where g −h
is not invertible. Assume, for convenience, that B is given concretely as
(5.44) B = End(V1 ) × . . . × End(Vn )
where Vi are finite-dimensional Hermitian spaces. Then the claim to
check is that
b := h h−1 φ∗ h, φ − g g −1 φ∗ g, φ − (h − g)ρ
   
(5.45)
38 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH, P. PANDIT

is non-negative on Ker(g − h) for (g, h) ∈ ∂A.

Since the flow is coordinate-independent,
L we may assume that h is
the identity. So let v ∈ V := Vi with g(v) = v, then
v ∗ bv = v ∗ φ∗ φ − φφ∗ − φ∗ gφ + gφg −1 φ∗ g v


(5.46)
= (φv)∗ (1 − g)φv + (φ∗ v)∗ g −1 − 1 φ∗ v ≥ 0


(5.47)
since 1 − g ≥ 0 and thus g −1 − 1 ≥ 0. q.e.d.

As a first consequence we see that any two solutions have the same
asymptotics by a “sandwiching” argument.
Corollary 5.5. Let h1 , h2 be solutions of (5.28) for t ≥ 0. Then
there is a constant C > 0 such that
1
(5.48) h1 (t) ≤ h2 (t) ≤ Ch1 (t)
C
for t ≥ 0.
A related result is established by Harada–Wilkin [9] who show that
the flow is distance decreasing.

Proof. We can find a C > 0 such that the inequality holds for t = 0.
By monotonicity, it holds for all t ≥ 0. q.e.d.

Call h an asymptotic solution of (5.28) if for some (hence any)

actual solution g there is a C > 0 such that C −1 g(t) ≤ h(t) ≤ Cg(t)
for sufficiently large t. We will find that (5.28) always admits explicit
asymptotic solutions in terms of iterated logarithms, and these are gen-
erally not actual solutions.
We return to the point of view that the flow is changing coordinates
on the Ei ’s instead of the metric. Write h = x∗ x, then (5.28) implies
that

∗ h
∗ i
(5.49) ẋx−1 + ẋx−1 = xφx−1 , xφx−1 − ρ.

Note that this equation only determines the selfadjoint part of ẋx−1 ,
which corresponds to the fact that x is determined only up to multipli-
cation by unitary elements on the left.
Proposition 5.6 (Homogeneity). Let x be a solution of (5.49) and
f : R → R a continuous function, then
 Z 
1
(5.50) y := x exp f
2
solves

∗ h
∗ i
(5.51) ẏy −1 + ẏy −1 = yφy −1 , yφy −1 − ρ + f.
SEMISTABILITY, MODULAR LATTICES, AND ITERATED LOGARITHMS
39

Proof. Let
Z
1
(5.52) F = f
2
then
 
(5.53) ẏy −1 = ẋeF + xḞ eF e−F x−1 = ẋx−1 + Ḟ

and the right hand side of (5.49) remains unchanged if x is replaced by

y. q.e.d.
The following gives a sufficient criterion to recognize asymptotic so-
lutions. It relies on monotonicity and homogeneity.
Proposition 5.7. Suppose

∗ h
∗ i
(5.54) ẋx−1 + ẋx−1 = xφx−1 , xφx−1 − ρ + s

with s = s(t) an absolutely integrable function with values in selfadjoint

elements of B. Then h = x∗ x is an asymptotic solution of (5.28).
Proof. In the special case when s is scalar-valued (i.e. takes values
in R · 1) the claim follows immediately from Proposition 5.6, since the
absolute value of an antiderivative of s is bounded by assumption. For
the general case it suffices to show (by symmetry) that if f is scalar-
valued with f ≥ s and y solves

∗ h
∗ i
(5.55) ẏy −1 + ẏy −1 = yφy −1 , yφy −1 − ρ + f

with (x∗ x)(0) ≤ (y ∗ y)(0), then (x∗ x)(t) ≤ (y ∗ y)(t) for t ≥ 0. This is
a strengthening of the monotonicity property, and proven in much the
same way as Proposition 5.4. The only modification is the following:
Assuming x(0) = 1 after a change of coordinates, the additional term is
(5.56) v ∗ (y(0))∗ f (0)y(0)v − v ∗ s(0)v = v ∗ (f (0) − s(0))v ≥ 0,
where we use the fact that f (0) is a scalar and (y ∗ y)(0)v = v. q.e.d.

5.4. Asymptotic solution. The goal of this subsection is to construct

an asymptotic solution of (5.28) using the iterated weight filtration on
the lattice of subrepresentations, which is described in terms of our ∗-
data as follows. The selfadjoint elements of B are partially ordered by
a ≤ b iff b − a is a non-negative operator. In particular, we get a partial
order on projectors, those p ∈ B with p2 = p∗ = p. Because B is finite-
dimensional, the poset Λ(B) of projectors is an artinian modular lattice.
To see this, identify B with a product of matrix algebras and projectors
with their images. The lattice of subrepresentations of φ ∈ M is the
sublattice
(5.57) Λ(B, φ) := {p ∈ Λ(B) | pφp = φp}
40 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH, P. PANDIT

of projectors which are compatible with φ. The trace τ on B together

with ρ ∈ B provide a polarization
(5.58) Z : K(Λ(B, φ)) → C, [p, q] 7→ τ ((1 + ρi)(q − p))
which sends positive classes to the right half-plane. By the general the-
ory, Λ(B, φ) has a HN filtration, and each semistable interval is further
refined by a balanced filtration, perhaps iterated. Since we are mainly
interested in the refinement of the HN filtration, we assume that
(5.59) Λ(B, φ) is semistable of phase 0
which means that τ (ρ) = 0 and τ (ρp) ≤ 0 for all p ∈ Λ(B, φ), i.e. just
semistability with respect to θ.
By our general theory for modular lattices there is a canonical iterated
weight filtration p± (λ) in the sublattice Λ(B, φ)0 of semistables of phase
0, as in (4.86). Let
(5.60) pλ := p+ (λ) − p− (λ)
which is a projector, though usually not in Λ(B, φ). Since p± (λ) is an
R-filtration, the pλ are mutually orthogonal and sum to 1. Split φ into
its λ-components
X
(5.61) φλ = pµ+λ φpµ
µ∈R

then φλ = 0 for λ > 0 since p± (λ) ∈ Λ(B, φ).

Since each interval
(5.62) [p± (λ), p± (λ + 1)] ⊂ Λ(B, φ)0
is complemented by assumption, we can ensure, after conjugating φ by
a suitable invertible element b ∈ B, that
(5.63) φλ = 0, λ ∈ (−1, 0)
where b is chosen to take a splitting to an orthogonal one. Also, by
definition of Λ(B, φ)0 and the assumption that each [p+ (λ), p− (λ)] is
complemented, applying Theorem 5.1 and further conjugating φ we have
(5.64) [φ∗0 , φ0 ] = ρ.
Furthermore, we can choose harmonic representatives of the φλ , λ ≤ −1,
meaning we conjugate φ to get
(5.65) [φ∗0 , φλ ] = 0, λ ≤ −1.
Let
X
(5.66) r= λpλ
λ∈R
SEMISTABILITY, MODULAR LATTICES, AND ITERATED LOGARITHMS
41

and
(5.67) B 0 := {b ∈ B | [r, b] = 0, [φ0 , b] = 0}
(5.68) M 0 := {m ∈ M | [r, m] = −m, [φ∗0 , m] = 0}
which are the harmonic r-degree 0 part of B and harmonic r-degree −1
part of M respectively. A slight extension of the proof of Proposition 5.3
shows that B 0 is a ∗-subalgebra and M 0 is a B 0 –B 0 ∗-bimodule. The
point is that if m, n ∈ M 0 then [r, m∗ n] = 0 automatically, but we
still need to project to the harmonic part to get an element of B 0 as in
Proposition 5.3. Note also that
(5.69) ρ0 := r ∈ center(B 0 ), φ0 := φ−1 ∈ M 0
by definition and (5.65). The defining property of the iterated weight
filtration guarantees that the new quadruple (B 0 , ρ0 , M 0 , φ0 ) is semistable
of phase 0.
Let x be a solution of the flow (5.49) for (B 0 , θ0 , M 0 , φ0 ), so

∗ h
∗ i
(5.70) ẋx−1 + ẋx−1 = P xφ−1 x−1 , xφ−1 x−1 − r

where P : B → B 0 is the orthogonal projection. Note that (B 0 , θ0 , M 0 , φ0 )

is polystable, i.e. Λ(B 0 , φ0 )0 complemented, if and only if there exists a
constant solution x. It follows from the calculation below and induction
that x grows at most polynomially in general.
Lemma 5.8. Suppose x : [0, ∞) → B 0 is a solution to (5.70), then
(5.71) y(t) := tr/2 x(log t)
satisfies

∗ h
∗ i
(5.72) ẏy −1 + ẏy −1 =P yφ−1 y −1 , yφ−1 y −1 .

Proof. Indeed,
(5.73)
 
dy r r/2−1 dx
(t)(y(t))−1 = t x(log t) + tr/2 (log t)t−1 (x(log t))−1 t−r/2
dt 2 dt
 
−1 r dx −1
(5.74) =t + (log t)(x(log t))
2 dt
and
(5.75) y(t)φ−1 (y(t))−1 = t−1/2 x(log t)φ−1 (x(log t))−1
since φ−1 has r-degree −1. We use here the fact that x commutes with
r, thus tr/2 , as x(t) ∈ B 0 by definition. q.e.d.
Let
(5.76) ∆ : B → B, ∆(b) = [φ∗0 , [φ0 , b]]
42 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH, P. PANDIT

and
(5.77) G : B → B, 1 = P + ∆G = P + G∆, P G = GP = 0
the “Green’s operator”. All three P, ∆, G are endomorphisms of B as a
B 0 –B 0 bimodule and commute with the ∗ operation.
Let x, y be as in the previous lemma and consider
h i
−1 ∗ −1


(5.78) k := yφ−1 y , yφ−1 y
 
1 −1
(5.79) z := y 1 + G(y ky) .
2
Lemma 5.9. The function z above is a solution of (5.49) up to terms
in L1 , i.e. satisfies the hypothesis of Proposition 5.7 and is thus an
asymptotic solution.
It is important to note that the factor y −1 z = (1 + 21 G(y −1 ky)) is
bounded for large t, hence does not change the asymptotics. It is only
needed to get a solution up to terms in L1 .
Proof. We write O(tα L) for terms which are O(tα ) up to logarithmic
corrections, e.g. O(tα log t), O(tα log t log log t), and so on. For instance,
since φ−1 has r-degree −1 and k has r-degree 0 we have
(5.80)
yφ−1 y −1 = O(t−1/2 L), k = O(t−1 L), G(y −1 ky) = O(t−1 L).
Consequently,
 −1  
1 −1 1
(5.81) 1 + G(y ky) = 1 − G(y ky) + O(t−2 L)
−1
2 2
and
 
1
(5.82) ż = ẏ 1 + G(y ky) + yO(t−2 L)
−1
2
where the terms in O(t−2 L) are of r-degree 0, hence
(5.83) żz −1 = ẏy −1 + O(t−2 L).
Recall the splitting
(5.84) φ = φ0 + φ−1 + φ<−1
where φ<−1 collects components of r-degree < −1 −  for some  > 0.
We have
(5.85)
   
1 1
zφ0 z = y 1 + G(y ky) φ0 1 − G(y ky) y −1 + O(t−2 L)
−1 −1
2 2
1
(5.86) = φ0 + [G(k), φ0 ] + O(t−2 L)
2
SEMISTABILITY, MODULAR LATTICES, AND ITERATED LOGARITHMS
43

and
(5.87) zφ−1 z −1 = yφ−1 y −1 + O(t−3/2 L)
(5.88) zφ<−1 z −1 = O(t(−1−)/2 L)
hence
(5.89)
[(zφ0 z −1 )∗ , zφ0 z −1 ] =
1 1
(5.90) = [φ∗0 , φ0 ] + [φ∗0 , [G(k), φ0 ]] + [[G(k), φ0 ]∗ , φ0 ] + O(t−2 L)
2 2
(5.91) = θ − ∆G(k) + O(t−2 L)
(5.92) = θ + (P − 1)(k) + O(t−2 L)
by (5.64). Furthermore, by (5.65),
(zφ0 z −1 )∗ , zφ−1 z −1 = O(t−3/2 L)
 
(5.93)
(zφ−1 z −1 )∗ , zφ0 z −1 = O(t−3/2 L)
 
(5.94)
and
(zφ−1 z −1 )∗ , zφ−1 z −1 = (yφ−1 y −1 )∗ , yφ−1 y −1 + O(t−2 L).
   
(5.95)
Finally, combining the above we get
(zφz −1 )∗ , zφz −1 = θ + P (k) + O(t−1− L)
 
(5.96)
thus

∗
∗
(5.97) żz −1 + żz −1 = ẏy −1 + ẏy −1 + O(t−2 L)
(5.98) = P (k) + O(t−2 L)
= (zφz −1 )∗ , zφz −1 − θ + O(t−1− L)
 
(5.99)
which completes the proof. q.e.d.

Let
X
(5.100) 1= pλ
λ∈R∞

be the orthogonal splitting of the identity in B given by the iterated

weight filtration. Disregarding multiplicatively bounded terms coming
from the Green’s operator, the asymptotic solution of (5.28) constructed
in the proof above is
X  λ n
(5.101) tλ1 (log t)λ2 · · · log(n−1) t pλ
λ=(λ1 ,...,λn )∈R∞

where log(k) is the k-times iterated logarithm.

44 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH, P. PANDIT

Corollary 5.10. After conjugating φ by a suitable invertible element

in B, the function [C, ∞) → B given by (5.101) is an asymptotic solution
of (5.28).
Theorem 5.11. Suppose B, τ , θ, M , φ are as in Section 5.2 and h
a solution of (5.28). For λ = (λ1 , . . . , λn ) ∈ R∞ let p± (λ) ∈ Λ(B, φ) be
the projector which is the λ-step of the iterated weight filtration in the
polarized lattice Λ(B, φ), then
  λ n 
λ1 λ2 (n−1)
(5.102) p+ (λ)hp+ (λ) = O t (log t) · · · log t

and
  λn 
(5.103) p− (λ)hp− (λ) = o tλ1 (log t)λ2 · · · log(n−1) t .

Example 5.12. The simplest non-trivial example is the representa-

tion
√
1/ 2
(5.104) C −−−→ C
of the A2 quiver. The normalization is chosen so that (5.64) holds at
the second level. Also, we set θ = 0 and τ = tr. The equations (5.28)
are
1 1
(5.105) h˙1 /h1 = h2 /h1 , h˙2 /h2 = − h2 /h1
2 2
with asymptotic solution
(5.106) h1 = t1/2 , h2 = t−1/2
which happens to be an exact solution.
Example 5.13. Let us look at an example where the weight filtration
is iterated. Namely, take the representation
√ √
1/ 2 1 1/ 2
(5.107) C −−−→ C ←−−−− C −−−→ C
of the A4 zig-zag quiver, again with τ = tr, θ = 0. The equations (5.28)
are
1 1
(5.108) h˙1 /h1 = h2 /h1 , h˙2 /h2 = − h2 /h1 − h2 /h3
2 2
1 1
(5.109) h˙3 /h3 = h4 /h3 + h2 /h3 , h˙4 /h4 = − h4 /h3
2 2
with asymptotic solution
(5.110) h1 = t1/2 (log t)−1/2 1 + (log t)−1 , h2 = t−1/2 (log t)−1/2

h3 = t1/2 (log t)1/2 , h4 = t−1/2 (log t)1/2 1 + (log t)−1

(5.111)
which is not an exact solution, but solves (5.28) up to terms in L1 .
This is what we found in Section 2 but with slightly different constants
resulting from the change of basis on the representation.
SEMISTABILITY, MODULAR LATTICES, AND ITERATED LOGARITHMS
45

References
[1] I. Assem, D. Simson, and A. Skowroński, Elements of the representation the-
ory of associative algebras. Vol. 1, volume 65 of London Mathematical Society
Student Texts. Cambridge University Press, Cambridge, 2006, MR2197389, Zbl
1092.16001. 30
[2] G. Birkhoff, Lattice Theory, 3rd ed., volume 25 of American Mathematical Soci-
ety Colloquium Publications. American Mathematical Society, 1967, MR0227053,
Zbl 0505.06001. 17
[3] T. Bridgeland, Stability conditions on triangulated categories, Ann. of Math.
166 (2007) 317–345, MR2373143, Zbl 1137.18008. 2, 6
[4] C. Cornut, On Harder-Narasimhan filtrations and their compatibility with tensor
products, Confluentes Math. (2) 10 (2018) 3–49, MR3928223. 18
[5] S.K. Donaldson, Anti self-dual Yang–Mills connections over complex algebraic
surfaces and stable vector bundles, Proc. London Math. Soc. (3) 50 (1985) 1–26,
MR0765366, Zbl 1427.14003. 2
[6] P.A. Griffiths, Periods of integrals on algebraic manifolds: Summary of main
results and discussion of open problems, Bull. Amer. Math. Soc. 76 (1970) 228–
296, MR0258824, Zbl 0214.19802. 3
[7] F. Haiden, L. Katzarkov, M. Kontsevich, and P. Pandit, Categorical Kähler
geometry, in preparation. 2
[8] F. Haiden, L. Katzarkov, M. Kontsevich, and P. Pandit, Iterated logarithms and
gradient flows, preprint, arXiv:1802.04123. 2
[9] M. Harada and G. Wilkin, Morse theory of the moment map for representations
of quivers, Geom. Dedicata 150 (2011) 307–353, MR2753709, Zbl 1226.53081.
38
[10] J. Hofbauer and K. Sigmund, Evolutionary Games and Population Dynamics,
Cambridge University Press, 1998, MR1635735, Zbl 0914.90287. 13
[11] G. Kempf and L. Ness, The length of vectors in representation spaces, Algebraic
geometry (Proc. Summer Meeting, Univ. Copenhagen, Copenhagen, 1978), pp.
233–243, Lecture Notes in Math., 732, Springer, Berlin, 1979, MR0555701, Zbl
0407.22012. 2
[12] A.D. King, Moduli of representations of finite-dimensional algebras, Quart. J.
Math. Oxford Ser. (2) 45 (1994), no. 180, 515–530, MR1315461, Zbl 0837.16005.
2, 33
[13] F. Kirwan, Refinements of the Morse stratification of the normsquare of the
moment map, The breadth of symplectic and Poisson geometry, 327–362, Progr.
Math., 232, Birkhäuser Boston, Boston, MA, 2005, MR2103011, Zbl 1073.32015.
6
[14] J.B. Nation, Notes on Lattice Theory, Available at the author’s webpage:
http://www.math.hawaii.edu/∼jb/. 17
[15] W. Schmid, Variation of Hodge structure: the singularities of the period map-
ping. Invent. Math. 22 (1973), 211–319, MR0382272, Zbl 0278.14003. 3
[16] R. Sjamaara and E. Lerman, Stratified symplectic spaces and reduction. Ann.
of Math. (2) 134 (1991), no. 2, 375–422, MR1127479, Zbl 0759.58019. 34
[17] K. Uhlenbeck and S.-T. Yau, On the existence of Hermitian–Yang–Mills con-
nections in stable vector bundles. Frontiers of the mathematical sciences: 1985
46 F. HAIDEN, L. KATZARKOV, M. KONTSEVICH, P. PANDIT

(New York, 1985). Comm. Pure Appl. Math. 39 (1986), no. S, suppl., S257–
S293, MR0861491, Zbl 0615.58045. 2

(F. Haiden) University of Oxford, Mathematical Institute, Andrew

Wiles Building, Woodstock Road, Oxford OX2 6GG, UK
E-mail: [email protected]
(L. Katzarkov) Fakultät für Mathematik, Universität Wien, Oskar-Morgenstern-
Platz 1, 1090 Wien, Austria, HSE Moscow, and CMS Institute of Math-
ematics and Informatics, BAS Sofia, Bulgaria
E-mail: [email protected]
(M. Kontsevich) Institut des Hautes Études Scientifiques, 35 route de
Chartres, 91440 Bures-sur-Yvette, France
E-mail: [email protected]
(P. Pandit) International Centre for Theoretical Sciences (ICTS-TIFR),
Survey No. 151, Shivakote, Hesaraghatta Hobli, Bengaluru North 560089,
India
E-mail: [email protected]