Peculiar Modules For 4-Ended Tangles

PECULIAR MODULES FOR 4-ENDED TANGLES
CLAUDIUS BODO ZIBROWIUS
Abstract. With a 4-ended tangle T , we associate a Heegaard Floer invariant CFT (T ), the
peculiar module of T . We classify peculiar modules in terms of immersed curves on the 4-
punctured sphere, based on an algorithm due to Hanselman, Rasmussen and Watson [HRW16].
We deduce that the peculiar module of any tangle T can be computed combinatorially using
nice diagrams and illustrate this result by computing the peculiar modules of the 2-stranded
pretzel tangles T2n,(2m+1) for n, m > 0.
arXiv:1712.05050v1 [math.GT] 13 Dec 2017
Based on Zarevs bordered sutured Heegaard Floer theory [Zar11], we prove a glueing for-
mula for peculiar modules which recovers link Floer homology HFL. [ We then study some
applications: Firstly, we show that peculiar modules detect rational tangles. Secondly, we give
short proofs of various skein exact triangles. Finally, we observe certain symmetries in the
peculiar modules of the pretzel tangles T2n,(2m+1) and deduce that mutation of those tangles
preserves -graded, and for some orientations even bigraded link Floer homology.
Table of contents
Introduction 1
1. Preliminaries: Algebraic structures from dg categories 5
2. The invariant CFT 14
3. Pairing 4-ended tangles 25
4. Curved complexes for marked surfaces 33
5. Applications 54
References 66
Introduction
0.1. Peculiar modules. Let L be a link in the 3-sphere S 3 . Consider an embedded closed
3-ball B 3 S 3 whose boundary intersects L transversely. Then, modulo a parametrization of
the boundary B 3 , the embedding L B 3 , B 3 is essentially what we call a tangle. In [Zib16],
I introduced a set of Alexander polynomials sT and a Heegaard Floer theory HFT(T
[ ) for such
tangles T . They should be regarded as generalisations of the classical multivariate Alexander
polynomial [Ale28] and Ozsvth and Szabs and J. Rasmussens knot and link Floer homology
[OS03a, Ras03, OS05], respectively. Indeed, both tangle invariants have similar properties to
their corresponding link invariants. In particular, the graded Euler characteristic of HFT(T
[ )
recovers the polynomial invariants T . Moreover, the polynomials T satisfy a simple glueing
s s
theorem which allows one to prove results about the classical multivariate Alexander polynomial
of links, such as invariance under Conway mutation [Zib16, corollary 3.6]. Unfortunately, we
do not have a similar glueing theorem for the categorified invariants HFT(T
[ ).
1
2 CLAUDIUS BODO ZIBROWIUS
The main objective of this paper is to resolve this problem in the case of 4-ended tangles,
ie tangles with four ends on B 3 , see figure 1. For this purpose, we upgrade the tangle Floer
homology HFT(T
[ ) of 4-ended tangles T to an invariant which we call the peculiar module
of T and denote by CFT (T ). This is done by adding some more structure maps. In fact,
we construct an even more general invariant CFT (T ) := CFT (T, M ), a generalised peculiar
module, for tangles T in homology 3-balls M with spherical boundary. As algebraic objects, both
generalized and ordinary peculiar modules are curved type D structures over certain algebras,
the generalized and ordinary peculiar algebras A n and A , respectively. The algebra A is a

quotient of A n , obtained by setting certain variables equal

to 0. Similarly, CFT (T ) can be recovered from CFT (T ),
just as the hat version CFL
[ of link Floer homology can be
recovered from its -version CFL .
Both generalised and ordinary peculiar modules are Hee-
gaard Floer type invariants and, as such, rely on some choice
of Heegaard diagram. So the first goal is to prove that the in-
variants are independent of this choice. However, this follows
essentially from multi-pointed Heegaard Floer theory.
Theorem 0.1 (2.17). Given a 4-ended tangle T in a homology Figure 1. A diagram of a 4-

3-ball with spherical boundary, the bigraded chain homotopy ended tangle; in this case, the
(2, 3)-pretzel tangle T2,3
types of CFT (T ) and CFT (T ) are invariants of T .
We do not have a glueing theorem for the generalised peculiar modules CFT (T ), except
that one can recover CFL of certain closures of the tangle T from CFT (T ) (see remark 2.28).
Nonetheless, we do have a glueing formula for peculiar modules CFT (T ). Its proof relies on
Zarevs glueing theorem for his bordered sutured invariants [Zar11] and an identification of some
structure maps of certain bordered sutured invariants for tangles and peculiar modules. The
precise statement of the glueing theorem uses the -tensor product between type A and type D
structures familiar from bordered Heegaard Floer homology; for details, see definition 1.19.
Theorem 0.2 (3.6). Let T1 and T2 be two 4-ended tangles
and L the link obtained by glueing them together as illustrated
in figure 2. Then the link Floer homology HFL(L)
[ can be com- T1 T2
puted from CFT (T1 ) and CFT (T2 ). More precisely, there

exists a bounded, strictly unital type AA structure P such that
CFL(L)
[ V i = CFT (T1 ) P CFT (T2 )
Figure 2. A link obtained
where V is some 2-dimensional vector space and i {0, 1}. from two 4-ended tangles
0.2. Classification of peculiar modules. In section 4, we classify peculiar modules in terms

of immersed curves on the 4-punctured sphere.
Definition 0.3 (4.23). A loop on the 4-punctured sphere S = S 2 r 4D2 is a pair (, X), where
is an immersion of an oriented circle into S representing a non-trivial primitive element of
1 (S) and X GLn (F2 ) for some integer n. For such loops (, X), we call X the local system
of the loop. A collection of loops is a set of loops {(i , Xi )}iI such that the immersed curves
i are pairwise non-homotopic.
PECULIAR MODULES FOR 4-ENDED TANGLES 3
Theorem 0.4 (4.31, 4.35). With every peculiar module (C, ), we can associate a collection of
loops L(C, ) = {(i , Xi )}iI such that if (C 0 , 0 ) is another peculiar module with L(C 0 , 0 ) =
{(j0 , Xj0 )}jJ , (C, ) and (C 0 , 0 ) are homotopic iff there is a bijection : I J such that i is
homotopic to (i) 0 and Ai is similar to A0(i) for all i I.
Definition 0.5 (4.34). In particular, with any 4-ended

tangle T in a homology 3-ball M with spherical boundary, we
can associate a collection of loops, denoted by LT := LT,M ,
which is a tangle invariant up to homotopy of the underlying
curves and similarity of the local systems.
Question 0.6. The number of curves in LT is obviously a

tangle invariant. What is its geometric meaning?
The proof of theorem 0.4 is based on an algorithm due to

Hanselman, Rasmussen and Watson [HRW16] which they use
Figure 3. The three loops
to classify bordered Heegaard Floer invariants for 3-manifolds of L
T2,3 (with the unique
with torus boundary. There are striking similarities between 1-dimensional local systems)
their bordered Heegaard Floer invariants and peculiar mod- on the 4-punctured sphere for
ules, and it would be interesting to see if there exists a closer the (2, 3)-pretzel tangle T2,3
from figure 1.
connection between them apart from their formal properties.
Another important ingredient for the classification of peculiar modules is a classification of
morphisms between them in terms of the intersection theory of the corresponding collections of
loops. As a corollary of this second classification, we obtain the following two reformulations of
the glueing theorem.
Corollary 0.7 (4.44). With the same notation as in theorem 0.2,
CFL(L)
[ V i = HF(LT1 , Lm(T2 ) ) = H (Mor(CFT (T1 ), CFT (m(T2 )))),
where m(T2 ) denotes the mirror of T2 and HF denotes Lagrangian intersection Floer homology.
We actually prove theorem 0.4 for all categories of curved complexes over arbitrary marked
surfaces, see definition 4.7. As a consequence, we can show that by passing to certain quotients
of the peculiar algebra A , we do not lose information. This rather abstract observation has
the following very practical consequence.
Theorem 0.8 (4.38). Peculiar modules for 4-ended tangles can be computed combinatorially.
We illustrate this result in section 5.3 by computing CFT (T ) for an infinite family of 2-
stranded pretzel tangles. The following definition is inspired by [HW15, definition 3.2].
Definition 0.9. A tangle T is called loop-type if all local systems in LT are similar to per-
mutation matrices.
All tangles for which I have so far computed the invariants are loop-type. This provokes
the following question, whose corresponding counterpart for 3-manifolds with torus boundary
is open as well [HRW16, section 2.4].
Question 0.10. Are all tangles loop-type?

0.3. Applications. Peculiar modules detect rational tangles. In fact, this is true for any invari-
ant of 4-ended tangles for which there exists a glueing theorem recovering link Floer homology.
This is an easy consequence of unlink detection of link Floer homology and should be regarded
as its corresponding analogue for invariants of 4-ended tangles. However, the description of pe-
culiar modules of rational tangles is particularly simple, and hence so is rational tangle detection
for our invariants:
Theorem 0.11 (5.2). A 4-ended tangle T in the 3-ball is rational iff LT is a single embedded
loop with the unique 1-dimensional local system.
Likewise, any tangle invariant like CFT also detects orientable tangle genera, defined in
a suitable sense. For peculiar modules, this will be discussed in a forthcoming, separate pa-
per [WZ].
As another application, we reprove a result originally due to Manolescu [Mnl06], namely
the existence of an unoriented skein exact sequence, see theorem 5.6. Similarly, we obtain the
following slight generalisation of Ozsvth and Szabs oriented skein exact sequence [OS03a].
Theorem 0.12 (5.4, see also 5.5). Let Tn be the positive n-
twist tangle, Tn the negative n-twist tangle and T0 the trivial
tangle, see figure 4. Then there is an exact triangle

.. ..
CFT (Tn ) CFT (Tn ) n . n .

CFT (T0 ) V (a) Tn (b) Tn (c) T0
where V is a 2-dimensional vector space. If the tangles are Figure 4. The basic tangles
oriented and coloured consistently, one obtains (bi)graded ver- for the skein exact triangle from
sions of this triangle. Furthermore, it gives rise to an exact theorem 0.12
triangle relating the (appropriately stabilised) link Floer ho-
mologies of links that differ in these three tangles.
Definition 0.13 (Conway mutation). Given a link L in S 3 ,

let L0 be the link obtained by cutting out a tangle diagram R
R
R
with four ends from a diagram of L and glueing it back in
after a half-rotation, see figure 5 for an illustration. We say
Figure 5. Conway mutation
L0 is a Conway mutant of L and we call R the mutating
tangle in this mutation. If L is oriented, we choose an orientation of L0 that agrees with the
one for L outside of R. If this means that we need to reverse the orientation of the two open
components of R, then we also reverse the orientation of all other components of R during the
mutation; otherwise we do not change any orientation.
Ozsvth and Szab showed in [OS03b] that knot and link Floer homology is, in general, not
invariant under mutation. However, in [BL11, conjecture 1.5], Baldwin and Levine conjectured
the following.
Conjecture 0.14. Let L be a link and let L0 be obtained from L by Conway mutation. Then
HFL(L)
[ and HFL(L
[ 0 ) agree after collapsing the bigrading to a single Z-grading, known as the -
grading. More colloquially, -graded link Floer homology is mutation invariant.
In this paper, we prove a slightly stronger version of this

conjecture for an infinite family of mutating tangles.

2m + 1

.. ..
Theorem 0.15 (5.12). Mutation of (2n, (2m + 1))-pretzel . .
2n

tangles for n, m > 0, oriented as in figure 6, preserves bi-
graded link Floer homology, after identifying the Alexander

gradings of the two open strands. If we reverse the orientation
of one of the two strands, mutation of these tangles preserves Figure 6. An infinite family of
-graded link Floer homology. pretzel tangles for n, m > 0
This generalises an earlier result from my thesis [Zib17] for the (2, 3)-pretzel tangle. The
result again simply follows from an observation that the peculiar invariants for the mutating
tangles have a certain symmetry. However, the calculation of the invariants for general n, m > 0
is more involved and relies on theorem 0.8.
0.4. Similar work by other people. It is interesting to compare the ideas described in this
paper to the tangle Floer theories defined in [PV14, OS16, OS17] and their corresponding
decategorifications in terms of the representation theory of Uq (gl(1|1)) [EPV15, Mnn16]. In
fact, the definition of our generalised peculiar modules CFT is primarily inspired by Ozsvth
and Szabs invariants from [OS17], see remark 2.19. I also want to mention some impressive
work of Lambert-Cole [Lam16, Lam17], where he confirms conjecture 0.14 for various families
of mutant pairs (different from the one in theorem 0.15), using entirely different techniques.
0.5. Acknowledgements. This paper grew out of the final chapter of my PhD thesis [Zib17].
I would therefore like to take the opportunity to thank my PhD supervisor Jake Rasmussen for
his generous support. I consider myself very fortunate to have been his student.
My PhD was funded by an EPSRC scholarship covering tuition fees and a DPMMS grant for
maintenance, for which I thank the then Head of Department Martin Hyland. Some parts of
this paper were written during my stay at the Isaac Newton Institute during the programme
Homology Theories in Low Dimensions (EPSRC grant number EP/K032208/1). The paper was
completed during my time as a CIRGET postdoctoral fellow at the Universit de Sherbrooke.
I thank my PhD examiners Ivan Smith and Andrs Juhsz for many valuable comments on
and corrections to my thesis. I also thank Jonathan Hanselman, Robert Lipshitz, Andy Manion,
Ina Petkova, Vera Vrtesi and Marcus Zibrowius for helpful conversations. My special thanks
go to Liam Watson for his continuing interest in my work.
1. Preliminaries: Algebraic structures from dg categories
In this paper, we often work in categories of various algebraic structures, namely type D and
curved type D structures, but also type A structures and various bimodules. In all settings, we
often want to simplify these structures by replacing them by homotopy equivalent ones. The
main goal of this section is to develop some tools for dealing with this problem, namely the
cancellation lemma (1.23) and the clean-up lemma (1.25). The former can be used to reduce
the number of generators of an algebraic structure, essentially by doing Gaussian elimination
as in [BB14, lemma 3.2], the latter for making the structure maps look nicer, essentially by
changing the basis.
In the category of ordinary chain complexes, both tools will be familiar to the reader as
easy exercises in linear algebra. So it might not be too surprising that they also work in quite
general settings. We will spend the first part of this section explaining a general construction
which turns any differential graded category into another such category in which the lemmas
hold in some generality sufficient for our purposes, see definitions 1.4 and 1.5. Next, we show
that the various different algebraic structures mentioned above arise naturally from this general
construction. We also study its functoriality properties and interpret the box tensor product
between type A and type D structures in this framework. Finally, we state and prove the
cancellation and clean-up lemmas.
For simplicity, we only work over the field F2 = Z/2, so we do not need to keep track of signs.
However, with the correct sign conventions, most (if not all) statements should also hold over
fields of arbitrary characteristic.
1.1. The general construction.
Definition 1.1. Let Com be the category of Z-graded chain complexes over F2 and grading
preserving chain maps between them. A differential graded (dg) category C over F2 is an
enriched category over Com. To spell this out more explicitly, the hom-objects are Z-graded
F2 -vector spaces,
M
Mor(A, B) = Mori (A, B)
iZ
endowed with differentials
i : Mori (A, B) Mori1 (A, B),
ie vector space homomorphisms satisfying i1 i = 0 and
(1) m = m ( id + id ),
where
m : Morj (B, C) Mori (A, B) Mori+j (A, C)
denotes composition in C, which is associative and unital. For more details on enriched categor-
ies, see for example [Rie14]. Note that the identity morphisms have degree zero and lie in the
kernel of .
Definition 1.2. [Rie14, definition 3.4.5]. Given an enriched category C over some monoidal
category V, the underlying ordinary category C0 of C has the same objects as C and its
hom-sets are defined by
C0 (A, B) := MorV (1V , C(A, B)).
Example 1.3. Let C be a dg category. The unit in Com is the complex 0 F2 0, supported
in homological degree 0, and the morphisms in Com are grading preserving. Hence, the hom-sets
of C0 are those elements in the kernel of 0 . Next, consider the enriched category H (C) over the
category of graded vector spaces and grading preserving morphisms between them, obtained
from C by replacing the hom-objects by their homologies with respect to the differential . By
passing to the underlying ordinary category, we pick out the degree 0 morphisms in H (C).
Therefore, we denote this category by H0 (C). Since the hom-sets in H0 (C) are just quotients of
those in C0 , we now get the usual notions of chain homotopies between morphisms and objects.
The reason why we need to pass to the underlying category is that otherwise, two objects could
be (chain) isomorphic through grading shifting morphisms.
Definition 1.4. (cp. [Bar04, section 6] and [Sei06, section I.3k]) Given a dg category C, we
define another dg category Mat(C) as follows. Its objects are formal direct sums
M
Oi [ni ],
iI
where I is some finite index set and Oi [ni ] denotes the object Oi ob(C) with a formal grading
shift by an integer ni . Morphisms are given by
M M M
Morn ( Oi [ni ], Oj [nj ]) := Morn+ni nj (Oi , Oj ).
iI jJ (i,j)IJ
Compositions and differentials in Mat(C) are induced by those in C.
Definition 1.5. (cp. [Sei06, section I.3l]) Given a differential graded category C, we define an
auxiliary category Cxpre (C), the category of pre-complexes, which is an enriched category
over the category of Z-graded vector spaces and grading preserving morphisms between them.
Its objects are pairs (O, dO ), where O ob(C) and dO Mor1 (O, O). The hom-objects are
the same as in C,
Mor((O, dO ), (O0 , dO0 )) = Mor(O, O0 ),
viewed as Z-graded vector spaces. On these, we can define a map
D : Mori ((O, dO ), (O0 , dO0 )) Mori1 ((O, dO ), (O0 , dO0 ))
by setting
D(f ) := dO0 f + f dO + (f ).
We would like D to be a differential in order to turn Cxpre (C) into a dg category. However,
this only works in general if we restrict ourselves to a full subcategory of Cxpre (C). It is easy to
check that D is always compatible with multiplication in the sense of (1). So D is a differential
iff
D2 (f ) = (d2O0 + (dO0 )) f + f (d2O + (dO ))
vanishes. This is, of course, the case for the full subcategory Cx0 (C) of Cxpre (C) consisting of
those objects (O, dO ) for which
() d2O + (dO )
vanishes. However, in some situations, other conditions on () also work. For example, if we
replace Com by the category of Z/2-graded chain complexes, we can restrict to those objects
(O, dO ) for which () is equal to the identity. Also, if the hom-objects are bimodules over an
algebra A, we can ask () to be equal to a. idO for a fixed central algebra element a (of degree
2) which commutes with all morphisms f . In both cases, D will be a differential. We call any
such full subcategory a category of complexes, denoted by Cx (C), where {0, 1, a} is the
value of (). By construction, Cx (C) is a dg category.
Remark 1.6. As usual, we can associate a directed graph to a category, where objects cor-
respond to vertices and arrows to morphisms. In the same way, we can think of complexes in
Cx (Mat(C)) as graphs.
The point of the construction above is that after choosing a basis, we can interpret the
categories of type D, type A, type AA and curved type D structures as instances of Cx (Mat(C))
for suitable choices of relatively simple differential graded categories C. But let us start with
an even simpler example: ordinary chain complexes.
Note of warning. In the following examples, our definitions only coincide with the usual
ones after passing to the underlying ordinary categories, see example 1.3. The advantage of our
point of view is that the conditions we usually impose on morphism and chain homotopies for
various algebraic structures arise naturally by viewing those morphisms as elements of chain
complexes.
Example 1.7 (ordinary chain complexes over F2 ). Let C be the category with a single object
in grading 0, Mor(, ) = F2 and vanishing differential. Then (the underlying ordinary category
of) Cx0 (Mat(C)) is Com.
Example 1.8 (type D modules over dg F2 -algebras). Let A be a differential graded algebra
over F2 . Let C be the category with a single object and morphisms being elements in A.
Composition is multiplication in A and the differential is induced by the differential on A.
We define the category of type D modules by Cx0 (Mat(C)). (Again, note that we need to pass
to the underlying ordinary category to obtain the definitions in [Zar09] and [LOT08].)
Example 1.9 (type D modules over dg I-algebras). Let us assume that A is an algebra over
some ring I A of idempotents and fix a basis {ij }jJ of idempotents of I, where J is some
index set. Let CAD be the category with one object for each basis element of I, and for any
two such elements i1 and i2 , let Mor(i1 , i2 ) := i2 .A.i1 . Again, composition is multiplication
in A and the differential is induced by the differential on A. We call Cx0 (Mat(CA D )) the
category of (right) type D modules over dg I-algebras. To obtain the category of left type D
modules, we only need to change the morphism spaces to Mor(i1 , i2 ) := i1 .A.i2 with the obvious
multiplication; however, we usually work with right type D modules in this paper, since we
interpret algebra elements as functions and thus read them from right to left.
Remark 1.10. A priori, the definition in the previous example depends on a choice of basis
for I. In all examples in this paper, there is a natural choice of such a basis, so this is not an
issue. However, we can replace CAD above by the enlarged category C D , where there is an object
A
for every element in I. Then CA D is a full subcategory of C D and it is not hard to see that
A
Mat(CA D ) and Mat(C D ) are equivalent. Now, the construction of the category of complexes is
A
functorial (in the category of dg categories), so after all, the definition above does not depend
on a basis for I.
Example 1.11 (curved type D modules over dg I-algebras). We start with the same category
D as in the previous example, but we fix a central element a Z(A), the curvature, and
CA c
define the category of curved (right) type D modules with curvature ac as Cxac (Mat(CA
D )). For
a more explicit, but less concise definition, see definition 2.13.
Remark 1.12. In the Heegaard Floer community, the term curved seems to be the accepted
attribute for algebraic structures for which some differential is non-vanishing; however, the first
written reference (that I am aware of) in which this terminology is used is of very recent date
[Zem16].
Example 1.13 (type A structures over an A -algebra over I). Let A be an A -algebra over
a ring of idempotents I over F2 . As in example 1.9, fix a basis {ik }kI of idempotents of I,
where I is some index set. Let CA
A be the category with one object for each basis element in I,
just as for type D structures. However, a morphism in a hom-object Mor(i1 , i2 ) of CA

A is given
by a sequence of vector space homomorphisms
(fi : i2 .Ai .i1 F2 )i0
where composition is defined by

X
(f g)i (ai a1 ) := fk (ai aj+1 ).gj (aj a1 ).
j+k=i
The differential is given by

X ik
X
((f ))i (ai a1 ) := fj (ai k (al+k al+1 ) a1 ).
j+k=i+1 l=0
We call Cx0 (Mat(CAA )) the category of (left) type A structures. We define the category of strictly
unital type A structures by restricting to those objects (O, dO ) such that
dO (, 1) = idO
and
dO (, ai a1 ) = 0 if i > 1 and aj = 1 for some j = 1, . . . , i
and morphisms to those satisfying
f (, ai a1 ) = 0 if i > 0 and aj = 1 for some j = 1, . . . , i.
Right type A structures are defined in an analogous way; we only define the hom-objects of the
underlying dg category to be given by sequences of vector space homomorphisms
(fi : i1 .Ai .i2 F2 )i1 ,
and adapt the multiplication maps accordingly. In this paper, we restrict ourselves to left type A
structures.
Remark 1.14. When we describe type A structures as directed graphs, it is useful to fix a
basis of the algebra A. Then, we label an arrow corresponding to a morphism f by the formal
sum of those tuples/tensor products of basis elements of the algebra A on which f is non-zero.
In this language, composition of two morphisms f and g can be described as the sum of all
concatenations of labels for f and g (modulo 2).
Now consider a morphism fa whose only label is a tuple of basic algebra elements a =
(ai , . . . , a1 ). Then the arrow of (fa ) is labelled by the formal sum (modulo 2) of all labels
obtained from a by replacing a single entry ai by sequences (a0j , . . . , a01 ) such that
ai = j (a0j a01 ).
Example 1.15 (bimodules of various kinds). We can form the categories of type DD, type DA,
type AD and type AA bimodules as follows. We start with the dg category where objects
correspond to idempotents as before, but where the hom-objects are defined as the corresponding
products of the hom-objects of CAA and C D in examples 1.9 and 1.13. Multiplication is defined
A
as the product on the two factors and the differential is defined as usual by the Leibniz rule.
Likewise, multi-modules can be defined, but we will not need those in the current paper.
We sometimes include the algebras over which the modules are defined in our notation,
following the usual convention to use subscripts for type A and superscripts for type D sides.
For example, the notation A M B means that M is a type AD A-B-bimodule where A acts on
the left and B on the right.
1.2. Functoriality and pairing for type A and type D structures. In the proofs of the
glueing formula in section 3, we need to change the underlying algebras and rings of idempotents
of the algebraic structures involved.
Definition 1.16. Suppose, we have a subring J of the ring of idempotents I. Let us also fix
an F2 -basis {i |i J} of J and extend it to a basis {i |i I} of I. Let A be a dg I-algebra
and B be a dg J -algebra. Note that via the inclusion J , I, we can regard A also as a dg
J -algebra. Let : B A be a dg J -algebra homomorphism.
Let us first consider the construction for (curved) type D structures: Consider the category
CA corresponding to the I-algebra A from example 1.9. Similarly, let CBD be the category
D
corresponding to the J -algebra B. The inclusion J , I and the J -algebra homomorphism

induce a functor
FD : CBD CA
D
and hence also a functor
FD : Cx (Mat(CBD )) Cx (Mat(CA
D
)),
both of which respect the differentials on both sides.

There is also a dual construction for type A structures: consider the category CA A corres-
ponding to the I-algebra A from example 1.13. Similarly, let CBA be the category corresponding
to the J -algebra B. However, to define an induced functor, we need to slightly modify CBA by
adding a zero object. Let us call this new category CBA,0 . Then, we can define a functor
FA : CA
A
CBA,0
as follows: an object i is sent to i if i J and to the zero-object otherwise. A basic morphism

f Mor(1 , 2 ) of the form 2 .Ai .1 F2 is sent to

f 2 . i .1 : 2 .B i .1 2 .Ai .1 F2 Mor(1 , 2 )

if 1 and 2 are in J and to 0 Mor FA (1 ), FA (2 ) otherwise. This functor is well-defined
and respects the differential if A and B are A -algebras over I and J , respectively, and is
an A -algebra homomorphism. Thus, it induces a functor
FA : Cx0 (Mat(CA
A
)) Cx0 (Mat(CBA,0 ))
that likewise respects the differentials on both sides. Note that the category Cx0 (Mat(CBA,0 ))
agrees with the category of type A structures Cx0 (Mat(CBA )), since adding a zero-object in the
underlying dg category does not change the result after applying Cx0 (Mat()).
Similarly, we may define functors for bimodules of various types.
Remark 1.17. In the previous definition, non-injective homomorphisms J I also induce

functors
FD : Cx (Mat(CBD )) Cx (Mat(CAD
))
between the categories of (curved) type D structures. All we need to change in its construction
is to add a zero-object to CA
D . Note, however, that in all examples that we are concerned with
in this paper, either J = I and is a quotient map, or B = J .A.J and is the inclusion.
Observation 1.18. Let J , I, A, B and : B A be as in definition 1.16. Choose a basis of

the kernel of and extend it to B. This induces a basis on the image of , which we can then
extend it to a basis of A. Suppose we have an oriented labelled graph representing a type D
structure N over B (with respect to the basis chosen above). Then the graph representing
FD (N ) is obtained by replacing all algebra elements by their images under . Similarly, if we
have an oriented labelled graph representing a type A structure M over A, FA (M ) has the
effect of replacing each label a = (an , . . . , a1 ) by the formal sum of all labels whose images
under n are equal to a.
Definition 1.19 (pairing type D and type A structures). Let M be a (right) type D structure
and N a (left) type A structure over the same dg algebra A over a ring I of idempotents,
together with a fixed basis of I and A. We now reformulate the definition of the chain complex
(M N, ) from [Zar09, definition 7.4] and [LOT08, section 2.4] in terms of the graphs
associated with M and N . The generators of M N are defined by pairs of vertices in M and
N labelled by the same idempotents. Given two such pairs (v1 , w1 ) and (v2 , w2 ), the (v2 , w2 )-
component of (v1 , w1 ) is equal to the number of sequences s of labels of consecutive arrows
along a path from v1 to v2 in M such that s agrees with a label on the arrow from w1 to w2
in N , all modulo 2. Note that for the differential to be well-defined, we need to make sure that
this number is finite. This is usually done by requiring that at least one of M or N is bounded:
for type D structures, this means that there are no loops in its graph; for type A structures,
this means that there are only finitely many labels.
If we start with a type AD or type DD bimodule M and a type AA or type AD bimodule
N , the pairing is defined in the same way, except that we need to record the labels for the
remaining type A or type D sides: the first component is equal to the product (ie algebra
product or concatenation) of the first components of the arrows along the corresponding path
from v1 to v2 in M ; the second component of a label on the arrow from (v1 , w1 ) to (v2 , w2 ) is
equal to the second component of the corresponding label of the arrow from w1 to w2 in N .
Similarly, we can define a pairing between other types of bimodules, as long as we pair type A
sides with type D sides and left structures with right structures. For details, see [LOT10a,
definition 2.3.9].
Remark 1.20. The result of the pairing operation described above is a well-defined object in
the corresponding category, ie a chain complex, type A, type D structure or a bimodule, see
[LOT10a, proposition 2.3.10]. Furthermore, pairing type D and type A structures is invariant
under homotopy up to homotopy since it is functorial, see [LOT10a, lemma 2.3.13].
Questions 1.21. Is there a way to see that the statements in the previous remark are obvious?
D and C A ?
How do I interpret the pairing on the level of the categories CA A
Theorem 1.22 (Pairing adjunction). Let J , I, A, B and : B A be as in definition 1.16.

Let M be a (right) type D structure over B and N a (left) type A structure over A. Then we
have an identification
FD (M )A A N
= M B B FA (N ).
The same holds true for bimodules of various types.
Proof. This follows almost tautologically from the interpretation of the induced functors in terms
of oriented labelled graphs in observation 1.18. On the level of generators this identity is clear,
since by definition, M , FD (M ) and FA (N ) only possess generators belonging to idempotents
in J , so only those generators of N that belong to idempotents in J survive the pairing. Next,
fix a label a = (am , . . . , a1 ) from a vertex w1 to w2 in N , along with two vertices v1 and v2 in M
such that v1 and w1 belong to the same idempotent in J and so do v2 and w2 . The contribution
of a to the differential from (v1 , w1 ) to (v2 , w2 ) on the left-hand side is equal to the number of
all sequences b = (bn , . . . , b1 ) of labels from v1 to v2 such that (b) := ((bn ), . . . , (b1 )) = a.
The label a in N corresponds to labels b0 = (b0n , . . . , b01 ) from w1 to w2 in FA (N ) such that
(b0 ) := ((b0n ), . . . , (b01 )) = a. Those labels contribute 1 to the differential from (v1 , w1 ) to
(v2 , w2 ) on the right-hand side iff they agree with some b and do not contribute otherwise. So
the contributions agree.
For bimodules, note that the functors only act on one component of the morphism spaces.
1.3. Cancellation and Cleaning-up. We now state and

(Z, )
prove the two central lemmas mentioned at the beginning of
this section. b c
a d
Lemma 1.23 (Cancellation Lemma). Let (X, ) be an ob- f
ject of Cx (Mat(C)) for some differential graded category C (Y1 , 1 ) (Y2 , 2 )

e
and suppose it has the form shown in figure 7, where (Y1 , 1 ),
(Y2 , 2 ), (Z, ) ob(Cxpre (Mat(C))) and f is an isomorphism Figure 7. The object (X, )
with inverse g. Then (X, ) is chain homotopic to (Z, +bgc). for lemma 1.23
Remark 1.24. We usually apply this lemma to the case where (Y1 , 1 ) = (Y2 , 2 ) and f is the
identity map.
Proof. First of all, let us check that (Z, + bgc) is indeed an object of Cx (Mat(C)):
( + bgc)2 + ( + bgc) = + bgc + bgc + bgcbgc + () + (b)gc + b(g)c + bg(c)

= + (b + (b))gc + bg(c + (c)) + () + b(g)c + bgcbgc
= + (b1 + df )gc + bg(2 c + f a) + () + b(g)c + bgcbgc
= + dc + ba + () + (
b (D(g)
(((+ c = .
(
((
(gcbg)
(
For the last step, we observe that
gcbg = gD(f )g = D(gf g) + D(g)f g + gf D(g) = D(g).
Next, we consider the two chain maps
F : (Z, + bgc) (X, ) and G : (X, ) (Z, + bgc)

1 1
(Z, + bgc) (Z, ) (Z, ) (Z, + bgc)
c c bg
d d
b a b a
gc (Y2 , 2 ) (Y2 , 2 )
f f
e e
(Y1 , 1 ) (Y1 , 1 )
(a) The chain map F (b) The chain map G
1 1
(Z, ) (Z, + bgc) (Z, )
c c
bg
d d
b a b a
(Y2 , 2 ) gc (Y2 , 2 )
f f
e e
g
(Y1 , 1 ) (Y1 , 1 )
(c) The composition of F and G and the homotopy H (dashed arrow).
Figure 8. Maps for the proof of lemma 1.23
defined in figure 8a and 8b, respectively. One easily checks that indeed D(F ) = 0 and D(G) = 0.
Indeed, the only non-trivial terms we need to compute are
gc( + bgc) + 1 gc + (gc) + a = g(c + (c)) + (1 g + (g))c + a + gcbgc

= g(f a + 2 c) + (1 g + (g))c + a + gcbgc
= (D(g) + gcbg) c = 0
for the first identity and similarly
( + bgc)bg + bg2 + (bg) + d = (b + (b))g + b(g2 + (g)) + d + bgcbg

= (df + b1 )g + b(g2 + (g)) + d + bgcbg
= b (D(g) + gcbg) = 0.
for the second identity. Now, GF = 1Z and conversely, it is not hard to check that
F G = 1Z + D(H),
where H is the homotopy given by the dashed line in figure 8c.
Lemma 1.25 (Clean-up Lemma). Let (O, dO ) be an object in Cx (C) for some differential
graded category C. Then for any morphism h Mor0 ((O, dO ), (O, dO )) for which
h2 , hD(h) and D(h)h
vanish, (O, dO ) is chain homotopic to (O, dO + D(h)).

Proof. We can easily check that (O, dO + D(h)) is an object in Cx (Mat(C)):

(dO + D(h))2 + (dO + D(h)) = d2O + (dO ) + dO D(h) + D(h)dO + (D(h))
+ D(h)D(h)
The first term on the right gives () and the last term vanishes, which can be seen by applying
the differential D to hD(h) = 0. The middle term also vanishes, which can be seen by expanding
D(h) = dO h + hdO + (h) and using the fact that a term () commutes with any morphism.
The chain isomorphisms between the two objects are given by
1+h 1+h
(O, dO ) (O, dO + D(h)) and (O, dO + D(h)) (O, dO ).
Indeed, these two morphisms lie in the kernel of D, since hD(h) and D(h)h vanish. Their
composition is equal to 1 + h2 = 1.
2. The invariant CFT
2.1. Heegaard diagrams for tangles. Let us start by re-

d
calling the basic definitions from [Zib16, section 4], adapted 4
to 4-ended tangles. 1 b
b
Definition 2.1. A 4-ended tangle T in a homology 3- a c

ball M with spherical boundary is an embedding
a
T: S 1 , , M, S 1 S 2 = M ,
b
I qI q 2 b
3
such that the endpoints of the two intervals lie on a fixed circle b
S 1 on the boundary of M , together with a labelling of the arcs
S 1 r im(T ) by some index set; we usually use {a, b, c, d}, but Figure 9. 3 The (2, 3)-pretzel
tangle in S
sometimes it is more conventient to write s1 for a, s2 for b, s3
for c and s4 for d. We call a choice of a single arc a site of the tangle T . An orientation on
a tangle is an choice of orientation on the two intervals and the circles. We consider tangles
up to ambient isotopy which keeps track of the labelling of the arcs. The images of the two
intervals are called the open components, the images of any circles are called the closed
components of the tangle. We number the ends of the open components such that the first
endpoint is the point where the arcs d and a meet, the second for a and b, etc. We also label
these tangle components by variables t1 and t2 for the open components and t01 , t02 , . . . for the
closed components. We call those variables the colours of T .
Definition 2.2. A Heegaard diagram HT for a tangle T with n closed components in a

Z-homology 3-ball with spherical boundary M is a tuple (g , Z, = c a , ), where
g is an oriented surface of genus g with 2n + 4 boundary components, denoted by Z,
which are partitioned into (n + 2) pairs,
c is a set of (g + n) pairwise disjoint circles 1 , . . . , g+n on g ,
a is a set of 4 pairwise disjoint arcs on g , labelled a, b, c, d, which are disjoint from c
and whose endpoints lie on Z, and
is a set of (g + n + 1) pairwise disjoint circles 1 , . . . , g+n+1 on g .
back qi
i
pi
front
(a) An original tangle end (b) A new tangle end
Figure 10. The difference between peculiar and ordinary Heegaard diagrams for
tangles. The former are obtained from the latter by local changes near the boundary of
the Heegaard surface.
We impose the following condition on the data above: The 3-manifold obtained by attaching
2-handles to g [0, 1] along c {0} and {1} is equal to the tangle complement M r (T )
such that under this identification,
each pair of circles in Z is a pair of meridional circles for the same tangle component,
and each tangle component belongs to exactly one such pair, and
a is equal to S 1 r (T ) M .
Remark 2.3. As in [Zib16], our convention on the orientation of the Heegaard surface
is that its normal vector (determined using the right-hand rule) points in the positive direction,
ie in the direction of the -curves. However, we usually draw the Heegaard surfaces such that
the normal vector points into the projection plane.
Definition 2.4. Let T be a 4-ended tangle. A peculiar Heegaard diagram for T is obtained
from a tangle Heegaard diagram for T by a local modification around the punctures, as illus-
trated in figure 10: We collapse the four boundary components of which meet the -arcs,
thereby joining the four -arcs to a single -circle S 1 . Then we add a marked point for each
tangle end on either side of S 1 , pi on the front and qi on the back, and connect these two points
by an arc which intersects S 1 exactly once and no other curve. For each closed component, we
also contract the corresponding boundary components to points zj and wj . Thus, we obtain a
multi-pointed Heegaard diagram, whose underlying Heegaard surface is now closed and carries
basepoints pi , qi , zj and wj . As in [Zib16], we need to restrict ourselves to admissible dia-
grams, ie diagrams whose non-zero periodic domains avoiding all basepoints have both negative
and positive multiplicities, see [Zib16, definition 4.7].
Remark 2.5. It is obvious that we can go from a peculiar Heegaard diagram back to an ordinary
tangle Heegaard diagram. The only reason for introducing peculiar Heegaard diagrams is to
avoid any bordered Heegaard Floer theory, so the proof of invariance of the algebraic structures
we are about to define is a minor adaptation of the one for link Floer homology. In particular,
note that the number of -circles and -circles in a peculiar Heegaard diagram is the same.
Obviously, the Heegaard moves from [Zib16, lemma 4.6] are equivalent to the following moves
for peculiar Heegaard diagrams:
isotopies of the - and -curves away from the marked points and the connecting arcs,
handleslides of -curves over -curves and handleslides of -curves over -curves other
than S 1 , and
stabilisation.
Remark 2.6. The attribute peculiar should be considered as a homophone of p-q-lier, a

reference to the labels pi and qi we have chosen for the marked points. This comes from the
notation used in [AAEKO] in the context of the wrapped Fukaya category of the n-punctured
sphere.
Let us briefly recall the definition of the tangle Floer homology HFT(T
[ ) from [Zib16, sec-
tion 5], adapted to 4-ended tangles.
Definition 2.7. Given a peculiar Heegaard diagram HT for a 4-ended tangle T , let T := T(HT )
be the set of tuples of intersection points, also called generators, such that each - and each
-curve is occupied by exactly one point. We partition T into four sets Ts , where s is a site of
the tangle T , depending on which of the former four arcs of the special -circle S 1 are occupied.
For i = 1, 2, 3, 4, let Ii = F2 [i ]/(2i = i ) and
I := I1 I2 I3 I4
be the ring of idempotents, with one idempotent for each of the four sites of a 4-ended tangle.
Then, we define a right I -module CFT(T
[ ) by letting CFT(T [ ).i be the F2 -vector space freely
generated by elements in Tsi . For any two generators x and y of T, one considers the space
of domains from x to y, denoted by 2 (x, y). For each element 2 (x, y), one can define
a moduli space M() of holomorphic curves in [0, 1] R representing . (Alternatively,
one may count holomorphic discs in Symg+n+1 (g ); however, by the main result of [Lip05], this
gives the same theory, see also [OS05, section 5.2].) The dimension of this moduli space is given
by the Maslov index (). The differential in CFT(T
[ ) is defined as
X X
M()
x = # y,
y 0 (x,y) R
2
()=1

where 20 (x, y) denotes the domains avoiding all basepoints and # M()
R denotes the number
of points in the quotient of the 1-dimensional moduli spaces M() by a natural R-action. The
tangle Floer homology HFT(T
[ ) is defined to be the homology of CFT(T
[ ).
Definition 2.8. A matching P is a partition {{i1 , o1 }, {i2 , o2 }} of {1, 2, 3, 4} into pairs. An

ordered matching is a matching in which the pairs are ordered. A 4-ended tangle T gives rise
to a matching PT as follows: The first pair consists of the two endpoints of the open component
with colour t1 , the second consists of the two endpoints of the second open component of T ,
the one labelled t2 . Given an orientation of the two open components of T , we order each pair
of points such that the inward pointing end comes first, the outward pointing end second.
Definition 2.9. Given a domain 2 (x, y) between two generators x and y in T and a
basepoint x = pi , qi , zj , wj , let x() denote the multiplicity of at x. Then, we define three
gradings on the generators of CFT(T [ ): The -grading is a relative 1 Z-grading and defined by
2
X n
X
1
(y) (x) = () 2 (pi () + qi ()) (zi () + wi ()).
i=1,2,3,4 i=1
When comparing this to [Zib16, definition 4.19], note that we now use peculiar Heegaard dia-
grams to compute the Maslov index . Furthermore, for every component t of the tangle, there
is a relative Z-grading At , which is called the Alexander grading, see [Zib16, definition 4.14].
Given a ordered matching P = {{i1 , o1 }, {i2 , o2 }}, we define Atk for k = 1, 2 by
Atk (y) Atk (x) := Atk () := pok () + qok () pik () qik ().
For k = 1, . . . , n, we define At0k by
At0k (y) At0k (x) := At0k () := wk () zk ().
By taking the sum of all Alexander gradings, we obtain a relative Z-grading, the reduced
Alexander grading Ar . Finally, the homological grading h, a relative Z-grading, is defined as
h = 12 Ar .
We sometimes denote the Alexander grading on generators by a superscript list of integers

(or half-integers, if eg, we want to achieve the same symmetry present in the decategorified
3 1
invariants [Zib16]), like a+1 for the single-variate or a( 2 , 2 ) for the multivariate Alexander
grading.
2.2. Peculiar algebras.
Definition 2.10. For n 0, let Rpre n be the free polynomial ring generated by the variables
pi , qi and Uj and Vj , where i = 1, 2, 3, 4 and j = 1, . . . , n. Let Apre n be the I -I -algebra
whose underlying I -I -bimodule structure is given by s0 Apre n .s := Rn for pairs (s, s ) of
0
sites and whose algebra multiplication is defined by the unique I -I -bimodule homomorphism
Apre pre pre
n I An An which, for all triples (s, s , s ) of sites, restricts to the multiplication map
0 00
in Rpre
n :
s00 . Apre pre pre
n .s0 I s0 . An .s s00 . An .s .
| {z } | {z } | {z }
Rpre
n Rpre
n Rpre
n
We define a 12 Z-grading on pre

An , called the -grading, by setting
1
(i ) := 0, (pi ) = (qi ) := 2 and (Uj ) = (Vj ) := 1,
where i = 1, 2, 3, 4 and j = 1, . . . , n, and then extending linearly to all of Apre

n . Similarly, given
1
an ordered matching P = {{i1 , o1 }, {i2 , o2 }}, we define relative 2 Z-gradings Atk for k = 1, 2,
called Alexander gradings, by
1
Atk (s ) := 0, Atk (pik ) = Atk (qik ) := 2 and Atk (pok ) = Atk (qok ) := 21 ,
and similarly Alexander gradings At0k for k = 1, . . . , n by
At0k (s ) := 0, At0k (Uk ) := 1 and At0k (Vk ) := 1,
and then extend linearly to Apre

n . These gradings give rise to a reduced Alexander grading and
a homological grading as in definition 2.9.
pre
Definition 2.11. For each n 0, we define A
n to be the subalgebra of An generated as an
algebra by the following elements:
pi := i1 .pi .i , qi := i .qi .i1 , Ui := 1.pi qi .1, Uj0 := 1.Uj0 .1 and Vj0 := 1.Vj0 .1,
where we take the indices i = 1, 2, 3, 4 modulo 4 with an offset of 1, j = 1, . . . , n and where

1 = 1 + 2 + 3 + 4 is the identity in I . Naturally, the various gradings on Apre restrict to
well-defined gradings on this subalgebra.
Note that A n is a faithful Rn -module, where Rn is the free polynomial ring in the variables
Ui , Uj and Vj for i = 1, 2, 3, 4 and j = 1, . . . , n. Thus any element in A
0 0
n can be uniquely
written as a sum of elements of the form
pi pi+1 . . . pj1 pj .r and qi qi1 . . . qj+1 qj .r,
where r Rn is a monomial. This the standard basis on A n as a vector space over F2 . For con-
venience, we sometimes write the elements pi pi+1 . . . pj1 pj as pi(i+1)(j1)j and qi qi1 . . . qj+1 qj
as qi(i1)(j+1)j , where again, we take the indices modulo 4 with an offset of 1. Furthermore,
to simplify notation, we set
p = p1 + p2 + p3 + p4 A
n and q = q1 + q2 + q3 + q4 A
n,
so we can write for example p4 = p1234 + p2341 + p3412 + p4123 . We call A

n the generalised
peculiar algebra.
For most part of this paper, we will only be concerned with 4

a certain quotient of A
n , which was already defined in [Zib17].
q1 q4
p1 p4
1 3
Definition 2.12. Let A be the quotient of A n by the re- p2 p3
lations Ui = 0, Uj0 = 0 and Vj0 = 0. This algebra can be q2 q3

interpreted as the path algebra of the quiver in figure 11 with 2
relations pi qi = 0 = qi pi , see also example 4.8 and figure 23. Figure 11. The quiver for an
We call A the peculiar algebra. alternative definition of A
2.3. Peculiar modules. For the differential in CFT, [ we only consider holomorphic curves
which stay away from the basepoints in our peculiar Heegaard diagrams. We claim that we also
obtain a tangle invariant if we add those curves to our differential, recording their multiplicities
at the basepoints by elements of the algebras A n or A . However, the resulting complex does

not satisfy the relation 2 = 0. Instead, we obtain a slightly modified 2 -relation which enables
us to promote CFT
[ to more sophisticated homological invariants, namely certain curved type
D modules. As abstract algebraic structures, we defined curved type D modules in example 1.9.
Let us recall this definition here in slightly more down-to-earth terms.
Definition 2.13. Let I be a ring of idempotents and A a Z-graded algebra over I. Also fix
a central element ac Z(A) of degree 2. A (right) curved type D structure over A is a
Z-graded I-module M together with a (right) I-module homomorphism : M M I A of
degree 1 satisfying
(1M ) ( 1A ) = 1M ac ,
where denotes composition in A. We call ac the curvature of M . A morphism
between two curved type D structures (M, M ) and (N, N ) is an I-module homomorphism
M N I A. For two such morphisms f and g, their composition is defined as
(g f ) = (1M 2 ) (g 1A ) f.
We endow the space of morphisms Mor(M, N ) with a differential D defined by
D(f ) = N f + f M .
Then indeed D2 = 0, since we have chosen ac to be central. This gives us an enriched category
over Com, the category of ordinary chain complexes over F2 . The underlying ordinary category is
obtained by restricting the morphism spaces to degree 0 elements in the kernel of D, giving us the
usual notions of chain homotopy and homotopy equivalence, see definition 1.2 and example 1.3.
Remark 2.14. It is interesting to compare curved type D structures to matrix factorisations

as studied by Khovanov-Rozansky [KR04]. Given an algebra A over some field k, a matrix
factorisation of a potential w A consists of two free A-modules M0 and M1 with two maps
d0
(2) M0 M1
d1
such that d1 d0 = w. idM0 and d0 d1 = w. idM1 . If M0 and M1 denote the k-vector spaces
generated by an A-basis of M0 and M1 , respectively, we can regard d0 and d1 as maps
d0 : M0 M1 I A and d1 : M1 M0 I A.
Then (M0 M1 , d0 + d1 ) defines a curved type D structure over the k-algebra A.

In general, we cannot go in the other direction. For example curved complexes associated
with manifolds with torus boundary do not, in general, admit a splitting of the form (2). This is
for the simple reason that the total number of generators can be odd, see for example [HRW16,
p. 9 fig. 5]. However, for the curved type D structure invariants of tangles, such splittings exist,
which is an easy corollary of the classification in terms of immersed curves on the 4-punctured
sphere, see section 4.
Definition 2.15. Given an (ordered) matching P = {{i1 , o1 }, {i2 , o2 }}, let pqMod be the
category of -graded (and Alexander graded) curved complexes over A with curvature
p4 + q 4 .
We call the objects of this category peculiar modules. Furthermore, let gpqModP,n be the
category of -graded (and Alexander graded) curved complexes over A
n with curvature
(3) p4 + q 4 + Ui1 Uo1 + Ui2 Uo2 .
We call the objects of this category generalised peculiar modules.
Definition 2.16. Given a 4-ended tangle T with n closed components in a Z-homology 3-ball
M with spherical boundary and an (admissible) peculiar Heegaard diagram for T , let use define
a generalised peculiar module CFT (T ) := CFT (T, M ) in gpqModPT ,n whose underlying right
I -module agrees with CFT(T
[ ). However, the differential on CFT (T ) is defined by
X X
M()
(4) x = # y I a(),
yT 2 (x,y)
R
()=1
where for 2 (x, y), a() is the preimage of

Y n
Y
p () q () z () wj ()
s(y) . pi i qi i Uj j Vj .s(x) ,
i=1,2,3,4 j=1
pre
under the inclusion map An , An . We call (CFT (T ), ) the generalised peculiar module

of T . Its image under the functor
gpqModPT ,n pqMod
induced by the quotient map A

n A is denoted by CFT (T ), which we call the peculiar

module of T .
Theorem 2.17. CFT (T ) is indeed a well-defined generalised peculiar module. Furthermore,

its bigraded chain homotopy type is an invariant of the tangle T . Hence CFT (T ) is a well-
defined a peculiar module, whose bigraded chain homotopy type is also an invariant of the
tangle T .
Remark 2.18. On link Floer homology, one can promote both Alexander and -gradings to
absolute gradings via symmetries and the spectral sequence to HF(S d 3 ) [OS05]. I expect that
something similar can be done for our tangle invariants. Alternatively, one could simply fix
absolute gradings on a specific test tangle, say a trivial tangle, and then define absolute gradings
on all other tangles via the pairing with this test tangle (using theorem 4.44) and the absolute
gradings on HFL.
[ However, in this paper, we are only working with the relative gradings on
CFT and CFT inherited from those on CFT
[ which were defined in [Zib16]. So, throughout
this paper, all gradings on CFL should be regarded as relative, too.
[
Remark 2.19. The generalized algebra A n and the generalised invariant CFT are inspired

by Ozsvth and Szabs algebra and tangle invariant from [OS17]. Computations suggest that
their invariant for one-sided 4-ended tangles is closely related to CFT . Conceptually, it might
also be interesting to set up their theory for an odd number of tangle strands and then compare
CFT to their invariants of (1, 3)-tangles.
pre
Lemma 2.20. 2 (x, y), a() lies in the image of inclusion map A
n , An .
Proof. This follows from the observation that intersected with the -circle S 1 is a path on
S 1 connecting the two points of x and y on S 1 .
Lemma 2.21. increases the -grading by 1 and preserves the Alexander grading. (As usual,
the grading on a tensor product is given by the sum of the gradings of the tensor factors.)
Proof. Both statements follows directly from the definitions of the gradings of generators and
algebra elements.
Lemma 2.22. For each x CFT , the sum on the right-hand side of (4) is finite.
Proof. The proof is essentially the same as in link Floer homology, see [OS05, lemma 4.2]. Since
CFT is finitely generated, it is sufficient to show that the coefficient of each y CFT is
a finite sum. Note that by the previous lemma, the difference of the -gradings of x and y
determines the -grading of a(). Thus, there are only finitely many choices for the coefficients
a(). So let us also fix the multiplicities of at the basepoints. We can now argue as in the
proof of [OS01, lemma 4.13], using admissibility of the underlying Heegaard diagram.
In the following, we need two analytical facts from [OS05].
Fact 2.23. [OS05, lemma 5.4] Given a homology class of an -injective boundary degen-
eration, write as a linear combination of connected components of r . Then its Maslov
index () is equal to twice the sum of the coefficients. The same holds for -injective boundary
degenerations.
Fact 2.24. [OS05, theorem 5.5] Given a surface of genus g, equipped with a set of (g + r)
attaching circles, where r 1, and a pseudo-holomorphic curve of Maslov index () = 2 and
with non-negative multiplicities, then is equal to one of the connected components of r
and the number of pseudo-holomorphic boundary degenerations in the homology class of is
odd. The same holds for -injective boundary degenerations.
Proof of theorem 2.16. Checking the 2 -identity is analogous to the link case; we can follow
[OS05, proof of lemma 4.3] and count ends of moduli spaces of Maslov index 2 curves. We
fix two generators x and z and consider the disjoint union of moduli spaces M(), where
varies over those curves in 2 (x, z) with () = 2 and a() = a for some fixed a A n.
(In particular, this fixes the multiplicities of at the basepoints.) If there are no boundary
degenerations, there is an even number of ends, so the z a-component of 2 x vanishes. If
there are boundary degenerations, then by fact 2.23 above, they contribute at least 2 to the
Maslov index, so the remaining curve has to be constant, hence x = z. By fact 2.24, we get a
boundary degeneration for each component of r and r . For closed tangle components,
these boundary degenerations come in pairs which cancel each other. The remaining two -
injective boundary degenerations contribute the first two terms of (3) and the remaining two
-injective boundary degenerations contribute the last two terms. All other ends appear in
pairs again, so their contributions cancel.
It remains to show that the peculiar module is an invariant of the tangle T . However,
CFT (T ) is essentially the chain complex associated with a multi-pointed Heegaard diagram in
Heegaard Floer theory. Thus, we obtain invariance as a (curved) type D module over the free
polynomial ring Rpre
n . However, the same proof also works if we work over An , since we only

allow handleslides of -curves over -curves other than S 1 .
Example 2.25 (rational tangles). Figure 12 shows the peculiar modules of some very simple 4-
ended tangles. As shown in [Zib16, example 4.5], every rational tangle T has a tangle Heegaard
diagram with just a single -curve. Thus, we only count bigons in the differential of the peculiar
invariant, and only those that do not occupy both pi and qj . By tightening the -curve, we
can assume that there are no honest differentials in CFT (T ), ie that every bigon covers some
pi or qi . Then CFT (T ) can be read off from this single -curve as follows: the vertices of
the graph of CFT (T ) correspond to intersection points of this -curve with the -arcs. Its
arrows come in pairs labelled by powers of p or q. More precisely, for each component of the
-curve minus the -arcs, we obtain an arrow pair connecting the vertices corresponding to the
ends of this component; this arrow pair is labelled by powers of p if the component goes via
the front component of the 4-punctured sphere minus the -arcs and by powers of q otherwise.
Conversely, the -curve can be read off from CFT (T ). So in this case, we might actually view
the -curve as the invariant associated with the tangle.
d
b
d b
d
b
a b b
c a b b c
b b b
b b b
1 1
p1 1 1 1 1 1 q1 1 1 1
0 a( 2 , 2 ) 2 d( 2 , 2 ) 0 a( 2 , 2 ) 2 d( 2 , 2 )
p432 q234
p43 +q12
0 b(0,0) 0 d(0,0) q2 q341 q123 q4 p2 p143 p321 p4
p21 +q34
1 1 1
p214 1 1 1 1 1 q412 1 1
2 b( 2 , 2 ) 0 c( 2 , 2 ) 2 b( 2 , 2 ) 0 c( 2 , 2 )
p3 q3
Figure 12. Basic rational tangles, their Heegaard diagrams and peculiar modules. The
superscripts of the generators specify the Alexander grading.
Example 2.26 (the (2, 3)-pretzel tangle). Figure 13 shows the computation of CFT (T )
for the (2, 3) pretzel tangle from figure 9. In subsection 5.3, we will compute the peculiar
modules for more general pretzel tangles, using some more advanced methods, which we develop
in section 4 as a corollary of the general classification of peculiar modules; here, we compute
everything from definition, which works surprisingly well.
First, we compute the generators of the complex. They are ordered according to their Al-
exander grading on an infinite chessboard, where the generators in each of its fields have the
same Alexander grading and where moving one field down, respectively to the right, increases
the Alexander grading corresponding to the colour t1 , respectively t2 , by 1. Next, we compute
bigons and squares. Those correspond to the labelled arrows in figure 13c. But there are also
other contributing domains. Grading constraints tell us that we can only get additional morph-
isms between those generators which are connected by the other arrows. In principle, those
could point in both directions. However, in each case, the connecting domains in one direction
either have negative multiplicities or occupy both pi s and qi s, so we can only get arrows in one
direction. From this and the 2 -relation, we can deduce that all solid arrows contribute. There
are only eight remaining arrows (the dotted ones) and they can only appear in pairs. But it is
easy to see that we can homotope those dotted arrows away (using the clean-up lemma 1.25 for
curved type D modules), so in any case, the complex is homotopic to the invariant consisting of
the solid arrows only. We can then apply the cancellation lemma 1.23. We obtain a complex in
which every arrow is paired with another one going in the opposite direction and every generator
is connected along the arrows to exactly two other generators just as for rational tangles! A
schematic picture of this complex is shown in figure 14a, where these arrow pairs have been
replaced by single unoriented edges, such that we obtain a collection of loops.
q1 q4
d d
b
b
d t2 p1 p4
t1 3 3
b
b
4
x1 y1 4
b
b
a1 6 c1
b
a c p2 y2 b
b
b
a2 x2 y3 c2
p3
b
b
c3
2 2 q3
b b
b
b
q2 1
t1
b t2 1 5
(a) A tangle diagram for the (2, 3)-pretzel tangle (b) A Heegaard diagram for the tangle on the left
p2143 (46)
p4 p214 (4) p214 (46)
3 3 q23 (1345) 3
2 x2 d0 2 x2 c1 1 by1 2 x2 c2 1 by2 2 x2 c3 2 x2 b0 1 by3
q41 (1236) q3
q2 (126) p3 (26) q341 (133455) p3 (2) 1(2)
q2
q234 (15) q1 (3) q41 (35)
q41 (3) q2 q2
q23 (15) q23 (1)
3 3 3 3 3 3
2 a1 y1 2 a1 y2 2 a2 y1 2 a1 y3 2 a2 y2 2 a2 y3
p14 (4) p1 p14 (46)

p1
p2 (2) p143 (46)
PECULIAR MODULES FOR 4-ENDED TANGLES
p1 p32 (26)
p32 (2)
1(3) p1 (345)
p4 q4 (3) p432 (122466) q4 (35)
3
1 dy1 2 x d0 32 1 32 p32 (2345)
1 x1 c1 dy2 p14 (1246) x1 c2 1 dy3 2 x1 c3 2 x1 b0
q123 (15) q123 (1) q3
q1234 (15)
(c) The peculiar invariant for the pretzel tangle above
Figure 13. A computation of a peculiar invariant for a non-rational tangle, see example 2.26
23
2 23 1 23 1 23
b b b b b b
32 32 23 32
23 32
b b b b b b
b b b b b b
32 1 32 1 32 2
(a) A schematic picture of the result. Generators correspond to vertices, arranged according to their
Alexander grading and labelled by their -grading. The dotted edges correspond to pairs of arrows
labelled by powers of q, the solid ones to pairs of arrows labelled by powers of p.
b b b
b b
b b b
b
b b b
b b
b b b
(b) (c) (d)
Figure 14. The final result of the computation from example 2.26 and figure 13. The
subfigures (b)(d) show the three loops from (a) separately on 4-punctured spheres.
In figures 14b-d, the loops have been transferred onto separate 4-punctured spheres in such a
way that the vertices lie on the four arcs that connect the punctures and the unoriented edges
lie on the front or back of the spheres depending on whether they correspond to arrow pairs
labelled by powers of p or q.
The meaning of both of these representations as loops will be discussed in section 4. For the
moment, they are just a convenient way to see certain symmetries.
Proposition 2.27. Let T be a 4-ended tangle and L the link

obtained by closing T at the sites a and c as shown in fig- a T c
ure 15. Then
CFL(L)
[ = F1 (CFT (T )) =
Vi F2 (CFT (T ))
Figure 15. A link obtained as
where i = 0 or 1, depending on whether L has two or one the closure of a tangle T
closed component(s) more than T , respectively, and where F1 and F2 are the functors induced
by the quotient maps
f1 : A F2 , p1 , p2 , q3 , q4 , 7 1, q1 , q2 , p3 , p4 7 0
f2 : A F2 , p1 , p2 , q3 , q4 , 7 0, q1 , q2 , p3 , p4 7 1,
respectively. For the other two opposite sites, we obtain a similar formula by a cyclic permutation
of the indices.
Proof. If we delete those basepoints in a peculiar Heegaard diagram of T that correspond to

the variables that we set equal to 1, we obtain a Heegaard diagram for L. In CFL(L),
[ we only
count those holomorphic curves that stay away from the remaining basepoints, so we need to
set those algebra elements equal to 0.
Remark 2.28. One can obtain the minus version of link Floer homology from the generalised
peculiar modules from the same idea as in the proof of the previous proposition.
We end this section with some simple observations about CFT .
Observation 2.29. By definition, the Alexander grading corresponding to a closed tangle

component vanishes on A . Also, the differential of a peculiar module preserves the Alexander
grading by lemma 2.21. Thus, CFT (T ) decomposes into the direct sum over the Alexander
gradings of closed tangle components.
Observation 2.30. In [OS02], Ozsvth and Szab showed that the link Floer homology of
alternating links is completely determined by the Alexander polynomial (and, to be precise, the
signature, but this is only needed to fix the absolute grading). The proof generalises immediately
to HFT,
[ using the generalised clock theorem [Zib15, theorem 1.13].
Question 2.31. Given an alternating tangle T , is the bigraded chain homotopy type of
CFT (T ) determined by sT ?
3. Pairing 4-ended tangles
In this section, we prove a glueing formula for CFT : Given the peculiar modules of two
4-ended tangles T1 and T2 , we compute the Heegaard Floer homology of the link L obtained
by glueing T1 to T2 , up to at most one stabilisation. The proof is essentially a calculation of a
bordered sutured type AA bimodule. So let us start by recalling Zarevs Heegaard Floer theory
for bordered sutured manifolds.
3.1. Review of bordered sutured Heegaard Floer theory. We will assume some famili-
arity with [Zar09, Zar10, Zar11] and only give a short review of the basic geometric objects
involved.
Definition 3.1. An arc diagram Z is a triple (Z, a, M ), where Z is a (possibly empty) set of
oriented line segments, a an even number of points on Z and M a matching of points in a. The
graph G(Z) of an arc diagram Z is the graph obtained from the line segments Z by adding
an edge between matched points in a. Given an arc diagram Z, Z is defined to be the same
arc diagram, except that the orientation of the line segments is reversed.
Definition 3.2. A bordered sutured manifold with - and -arcs is a tuple
(Y, , Z , , Z , ),
where
Y is a sutured manifold with sutures ; more precisely, Y is an oriented manifold and
Y are embedded oriented simple closed curves, dividing Y into two oriented open
surfaces-with-boundary R and R+ such that R is equal to as embedded oriented
1-manifolds;
Z = (Z , a , M ) and Z = (Z , a , M ) are arc diagrams;
is an embedding of G(Z ) into the closure of R such that (Z ) ;
is an embedding of G(Z ) into the closure of R+ such that (Z ) and (Z )

(Z ) = ;
such that the map
(5) 0 ( r ( (Z ) (Z ))) 0 (Y r (im( ) im( )))
is surjective.
Remark 3.3. Condition (5) is called homological linear independence. If we drop this
condition, Zarevs invariants fail to be well-defined in general. Note that unlike Zarev, we
allow the sutured surfaces of bordered sutured manifolds to be degenerate in the sense that
surgery along all edges between matched points may contain closed components. This allows us
to consider more general bordered sutured manifolds. If we restrict to non-degenerate sutured
surfaces, homological linear independence is automatically satisfied, see [Zar09, proposition 3.6].
Definition 3.4. A Heegaard diagram of a bordered sutured manifold is obtained from a

Heegaard diagram of the underlying sutured manifold by adding the graphs of the arc diagrams
to it. To be more precise, consider a Heegaard diagram of the underlying sutured manifold.
Then we can embed the graph G(Z ) into R in such a way that it misses the 2-handles
corresponding to the -curves, simply by sliding them off those 2-handles. This gives us an
embedding of G(Z ) into the Heegaard surface such that its image does not intersect the -
curves. We view the images of the edges connecting points in a as -arcs. We proceed similarly
for G(Z ) in R+ and the -curves.
The image of Z lies on the boundary of the Heegaard diagram, ie the sutures, which we
usually draw in green. We put a marked point, a basepoint, in every open component of the
boundary minus the image of Z.
Given an arc diagram Z, Zarev defines a moving strands algebra A(Z), and given a bordered
sutured Heegaard diagram, Zarev defines various bimodules over the strands algebras corres-
ponding to its arc diagrams. Each arc diagram can either play the role of a type D or type A
side of the bimodule; in fact, this is true for each connected component of an arc diagram, in
which case, one obtains multimodules. For a discussion of type A and D structures as algebraic
objects, we refer the reader to section 1, in particular examples 1.9 and 1.13.
A central result of Zarevs work is a pairing theorem, which is phrased in terms of the -
tensor product (see definition 1.19). The following theorem summarizes his theory for bimodules.
However, it is true for general multimodules.
Theorem 3.5 ([Zar09, theorem 3.10] or [Zar11, theorem 12.3.2]). Let Y be a bordered sutured
manifold, bordered by two arc diagrams Z1 and Z2 . Then there are bimodules, well defined up
to homotopy equivalence:
A(Z1 )
A(Z1 ) BSAA(Y )A(Z2 ) BSDA(Y
\ \ )A(Z )
2
A(Z1 ) BSAD(Y
\ )A(Z2 ) A(Z1 ) \ )A(Z2 )
BSDD(Y
Let Y1 and Y2 be two such manifolds, bordered by Z1 and Z2 , and Z2 and Z3 , respectively,
and let Y1 Z2 Y2 be the 3-manifold obtained by glueing Y1 and Y2 together along tubular neigh-
bourhoods of the images of Z2 on Y1 , respectively Z2 on Y2 . Then there are homotopy
equivalences
A(Z1 ) \ )A(Z ) A(Z2 ) BSDA(Y
A(Z1 ) BSAA(Y1 Z2 Y2 )A(Z3 ) BSAA(Y
\ = \ )A(Z ) ,
2 3
A(Z1 ) \ )A(Z2 ) A(Z ) BSAA(Y

A(Z1 ) BSDA(Y1 Z2 Y2 )A(Z3 ) BSDD(Y
\ = \ )A(Z ) ,
2 3
etc. Any combination of bimodules for Y1 and Y2 can be used, where one is type A for A(Z2 ),
and the other is type D for A(Z2 ).
3.2. A first glueing formula.
Theorem 3.6. Let T1 and T2 be two 4-ended tangles and L

the link obtained by glueing them together such that their sites d d
match up, as illustrated in figure 16. Let P be the strictly a T1 c a T2 c
unital type AA structure over A defined in figure 17. Then
b b
CFL(L)
[ V i = CFT (T1 ) P CFT (T2 )
where V is the 2-dimensional vector space supported in Alex- Figure 16. The link obtained
ander gradings t and t1 and identical -gradings and where by glueing two 4-ended tangles
i = |T1 | + |T2 | |L| 2 {0, 1}. together along matching sites
(q3 |p1 )
0 ba 0 cd
(q2 |) (q23 |p1 ) (q34 |p1 ) (q4 |)
(q3 |p21 ) (q3 |p14 )

(|p2 ) (|p4 )
12 (p1 |q234 )+(q234 |p1 ) 1
aa 2 dd
(p21 |q34 )+(q34 |p21 )
(p14 |q23 )+(q23 |p14 )

1 1
2 bb 2 cc
(p214 |q3 )+(q3 |p214 )
(|q2 ) (|q4 )
(p1 |q23 ) (p1 |q34 )
(p2 |) (p21 |q3 ) (p14 |q3 ) (p4 |)
0 ab 0 dc
(p1 |q3 )
Figure 17. The type AA structure P for theorem 3.6. The identity action is implicit.
The idempotent of a generator xy is x in the first component and y in the second.
The algebra elements in the first and second components are also coloured red and blue,
respectively. This is the same convention that we use in the proof of theorem 3.6, where
we identify P with the bordered sutured type AA structure P 0 illustrated in figure 19c.
The -grading of each generator is specified by the exponent of . Similarly, for any given
ordered matching, there is a unique way to define a relative Alexander grading on P
which is preserved by the differentials after inverting the algebra Alexander grading.
d D
a T1 c A T2 C
b B
(a) The parametrization on XT1 (b) The parametrization on XT2
D F
A a c C
E
b
e
(c) The parametrization on the boundary of the thickened 4-punctured sphere X
Figure 18. A decomposition of the complement of the link from figure 16

f d

p4 F
b
D
b
P1 q4 c
b
d P4
b
p4
b
p1 D Q4
f P4
b
C F
b
c
a C
A
E q2 a
b
p2 E
b
Q3
Q2 b
q2
b
B
b
A P2 q3

b
e Q2 e b
b
b
(a) A Heegaard diagram for X. The orientation on the surface is such that the normal vector (de-
termined by the right-hand rule) points out of the projection plane.
aa : cCE ab : bcAC cc : aAF cd : adAC
aa : bcBC ba : acBC cc : adAD dc : acAD
aa : ceC bb : acAC cc : af A dd : acAC
(b) Generators in idempotents I0 and I0 for the Heegaard diagram above. The idempotent of a
generator xy is x I0 and y I0 .
{q3 /P1 }
ba cd
{q2 , } {q2 ,q3 /P1 , } {q4 ,q3 /P1 } {q4 }
{q3 /P1 ,P2 } {q3 /P1 ,P4 , }

{P2 } {P4 , }
aa {p1 /Q3 ,Q2 ,Q4 , }+{q2 ,q4 ,q3 /P1 , } dd
{p2 ,p1 /Q3 ,Q4 }+{q4 ,q3 /P1 ,P2 }
{p4 ,p1 /Q3 ,Q2 , , }+{q2 ,q3 /P1 ,P4 , , }
bb {p4 ,p2 ,p1 /Q3 , }+{q3 /P1 ,P2 ,P4 , } cc

{Q2 , } {Q4 }
{p1 /Q3 ,Q2 , } {p1 /Q3 ,Q4 }
{p2 } {p2 ,p1 /Q3 } {p4 ,p1 /Q3 , } {p4 , }
ab dc
{p1 /Q3 }
(c) The domains that contribute to the type AA structure P 0
Figure 19. A Heegaard diagram for the bordered sutured manifold X from figure 18c
and some computations of generators and domains
Proof of theorem 3.6. The strategy of the proof is to glue the tangle complements XT1 and XT2
to opposite sides of a thickened 4-punctured sphere X = I (S 2 r4D2 ), ie along {0}(S 2 r4D2 )
and {1} (S 2 r 4D2 ), respectively. For this, we equip XT1 , XT2 and X with the structures of
bordered sutured manifolds specified by the arc diagrams in figure 18. Note that each closed
component of T1 and T2 carries two oppositely oriented meridional sutures. By glueing XT1 to
the inside and XT2 to the outside of the thickened 4-punctured sphere X as shown in figure 18c,
we obtain the complement XL of the link from figure 16. In addition to those sutures on closed
components of T1 and T2 , the sutured manifold XL carries one meridional suture at each of the
four places where the tangles have been glued together.
Let A be the bordered sutured algebra corresponding to the arc diagram on X consisting of
-arcs and I the corresponding ring of idempotents. Let I0 be the subring of idempotents
occupying the -arcs e and f and : A0 , A the subalgebra I0 .A.I0 of A. Similarly define
I , I0 and : B 0 , B.
Zarevs glueing theorem (theorem 3.5) tells us that
(6) SFC(XL ) [ T1 )A A,B BSAA(X)

= BSD(X \ [ T2 )B
BSD(X
The left-hand side agrees with the left-hand term of the identity from theorem 3.6. So the goal
is to identify the three tensor-factors on the right-hand sides with each other.
First of all, observe that we can choose a Heegaard diagram for XT1 where the two -arcs
that have ends on the same suture do not intersect any -curve. Thus, generators of its type D
structure belong to the idempotents that occupy one of the four arcs a, b, c or d. The same is
true for XT2 and its -arcs. Moreover, the labels of the type D structures for XT1 and XT2 are
contained in A0 . In other words, BSD(X
[ T1 ) and BSD(X[ T2 ) lie in the images of the functors FD

and FD induced by the inclusions and , respectively (see definition 1.16). Thus, by the
pairing adjunction (theorem 1.22), the right hand side of (6) is equal to
(7) BSD(X
[ T1 )A0 A0 ,B0 FAA
,
(BSAA(X))
\ BSD(X
[ T2 )B0
where FAA ,
is the functor induced by the inclusions and .
So let us compute FAA ,
(BSAA(X))
\ first. A Heegaard diagram for X is shown in figure 19a.
The regions adjacent to a basepoint are shaded light green (). Intersection points are labelled
by black Greek and Roman letters. - and -arcs are labelled by a, b, c, d, e, f and A, B,
C, D, E, F , respectively, as in figure 18c. Furthermore, the two square regions with vertices
{d, , D, f } and {, b, e, B} are labelled by and , respectively. Finally, all other regions
in the Heegaard diagram have at least one boundary component and are labelled by Qi , Pi ,
pi and qi . There are two domains that have two labels, namely q3 /P1 and p1 /Q3 . There are
four pairs of regions that have the same label. Note however, that the coefficients of any paired
regions agree in any domains connecting generators in FAA ,
(BSAA(X)),
\ ie those generators
that occupy e, f , E and F . Thus, we may use these labels to describe all domains that contribute
to FAA
,
(BSAA(X)).
\ The generators of FAA
,
(BSAA(X))
\ are shown in figure 19b.
In the following, let us write domains D as formal differences D+ D of unordered sets of
regions D+ and D with D+ D = such that
X X
D= r r.
rD+ rD
Let us calculate some connecting domains between the generators. First of all, here are some
bigons with a single boundary puncture:
{p2 } : bb ab, {P2 } : bb ba, {q4 } : cd dd, {Q4 } : dc dd.

The following domains consist of two bigons, each with a single boundary puncture:
{q2 , } : aa ba, {Q2 , } : aa ab, {p4 , } : dc cc, {P4 , } : cd cc.

The polygonal regions q3 /P1 and p1 /Q3 connect the following generators:
{q3 /P1 } : ba cd, {p1 /Q3 } : ab dc.

All those domains above contribute to the type AA structure. From these, we can compute
the connecting domains shown in figure 19c. We claim that these are in fact all domains that
connect these eight generators and have non-negative multiplicities. To show this, we can argue
as follows: Pick any two generators x and y and calculate a connecting domain between them,
eg by following along the arrows in figure 19c. We observe that if x and y are among those
eight generators from figure 19c, we can always choose a domain
X = (X+ X ) : x y
with multiplicities in {1, 0, +1}. We can obtain any other connecting domain between x and
y by adding a periodic domain P = P+ P to X. Suppose this new connecting domain
(X+ + P+ ) (X + P ) has non-negative (respectively non-positive) coefficients only. Then
X+ + P+ X + P , so P+ X , X+ P , or respecively
X+ + P+ X + P , so P+ X , X+ P .
In particular, since neither X nor X+ contain any region more than once, we only need to
consider periodic domains which have multiplicities 1 or +1.
Let us compute the group of periodic domains of our Heegaard diagram. It is easy to see
that it is freely generated by the following three domains:
D1 = {p2 , p1 /Q3 , Q4 } {q4 , q3 /P1 , P2 },

D2 = {p2 , q2 } {P2 , Q2 },
D3 = {p4 , q4 } {P4 , Q4 }
The periodic domains which have multiplicities 1 or +1 are given by D1 , D2 , D3 ,
D1 D2 = {p1 /Q3 , Q4 , Q2 } {q4 , q2 , q3 /P1 },

D1 + D3 = {p4 , p2 , p1 /Q3 } {q3 /P1 , P4 , P2 },
D1 D2 + D3 = {p4 , p1 /Q3 , Q2 } {q2 , q3 /P1 , P4 },
D2 + D3 = {p2 , q2 , p4 , q4 } {P2 , Q2 , P4 , Q4 },
D2 D3 = {p2 , q2 , P4 , Q4 } {p4 , q4 , P2 , Q2 }
and their negatives. It is now elementary to check that indeed, figure 19c shows all connecting
domains with non-negative multiplicities.
There are four generators that we have not considered yet, namely aa, aa, cc and cc. They
are connected by the following two contributing domains
{ } : aa aa, { } : cc cc.
Thus the two generator pairs can be cancelled. Let us connect these two generators to those
from figure 19c. There are for example the two domains
{P2 , Q2 } : aa aa, {P4 , Q4 } : cc cc.

We can use the same arguments as above to verify that there are no domains with non-negative
domains leaving aa or terminating at cc other than { } and { }. Hence, after cancellation,
the remaining complex is the same as before. We can now use the A -relations to deduce
that all domains from figure 19c contribute and thus define a type AA structure P 0 which, by
definition, is homotopic to FAA
,
(BSAA(X)).
\ Thus, the expression (7) simplifies to
(8) BSD(X [ T2 )B0

[ T1 )A0 A0 ,B0 P 0 BSD(X
Let A0 be the quotient algebra obtained from A by setting p3 = 0 = q1 . Now identify I with
I0 and I0 such that an idempotent for site s corresponds to an idempotent which does not
occupy the - and respectively -arc labelled s in figure 18c. Under this identification, there
are unique I -algebra epimorphisms
: A0 A0 and : B 0 A0 .
Note that, as the notation for our domains suggests, each domain from figure 19c is recorded
in P 0 by the algebra elements in A0 and B 0 corresponding to the algebra elements in A0 and
A0 obtained by the product of all red and blue labels of the domain, respectively. In fact, P 0 is
equal to the image of the type AA structure P from figure 17 (viewed as a bimodule over A0 )
under the induced functor FAA
,
. Thus, by the pairing adjuction (theorem 1.22), the expression
(8) is equal to
(9) F (BSD(X
[ T1 ))A0 P F (BSD(X
[ T2 ))A0

A ,A 0 0
Let : A

A0 be the quotient map defining A0 . Then by a final application of the pairing
adjunction, it is sufficient to identify F (BSD(X [ T1 )) and F (BSD(X

[ T2 )) with F (CFT (T1 ))
and F (CFT (T2 )), respectively.
The first identification is immediate from lemma 3.7 below and the observation that the
bordered sutured Heegaard diagram for XT1 can be regarded as a tangle Heegaard diagram for
T1 , up to a minor modification at the tangle ends. The second identification follows likewise,
observing that switching the roles of - and -curves while at the same time switching the
orientation of the Heegaard surface leaves the resulting complex unchanged.
Finally, the -grading on P is calculated as usual, see [Zib16, definition 4.19]. By the addit-
ivity of the -grading under glueing, the -gradings on both sides of our glueing formula agree.
Similarly, we can define a relative Z4 -grading Auni , a universal Alexander grading: For fixed
i, j {1, 2, 3, 4} and x {pi , qi , Pi , Qi }, let

1 if j = i,
Auni
j (x) = 0 if j 6= i.
D
D
Figure 20. A domain with multiplicity 1 near a tangle end
Since the three generating periodic domains have vanishing Alexander gradings, we can also
define such a grading Auni on P which is preserved by the arrows in P. Now, given an ordered
matching, we can specialize this universal Alexander grading to a relative Z2 -grading which
corresponds to the Alexander grading induced by that ordered matching. Again, by its additivity
under glueing, the Alexander grading on both sides of the glueing formula agree.
Lemma 3.7. Let D and D0 be two domains which only differ in a small region of multiplicity
1 as shown in figure 20. Then, for a suitable choice of complex structure, D contributes iff D0
does.
Proof. This follows from the same arguments as [Han13, proposition 2.7].
4. Curved complexes for marked surfaces
Definition 4.1. A marked surface is a pair (S, M ) where S is an oriented surface with
boundary and without closed components and M is a (possibly empty) set of basepoints on
S. An arc a of a marked surface (S, M ) is (the image of) an embedding of an oriented closed
interval
(I, I) , (S, S r M )
such that its class [a] 1 (S, S r M ) is non-trivial. Suppose, A is a set of pairwise disjoint
arcs on a marked surface (S, M ). For each arc a A, choose a closed neighbourhood N (a) of
a such that N (a) N (a0 ) = for all a0 A r {a}. Let s1 (a) and s2 (a) be the two components
of N (a) in S such that s1 (a) lies to the right of the oriented arc a and s2 (a) to its left; we call
s1 (a) and s2 (a) the two sides of a. We also fix a foliation Fa = I I of N (a) such that s1 (a),
s2 (a) and a are leaves of Fa .
S
Furthermore, we call the closures of the connected components of S r aA N (a) faces and
denote the set of all faces by F (S, M, A). We call A (together with a choice of fixed N (a) and
foliation Fa as above) an arc system if all faces f F (S, M, A) are topological discs containing
at most one point in M .
We now introduce some more notation. Given an arc system A on a marked surface (S, M ),
we denote the arc corresponding to a side s by a(s). A side s also lies on the boundary of some
face in F (S, M, A); we denote this face by f (s). The set of all sides s with f (s) = f is denoted
by S(f ). For each face f in F (S, M, A), let nf be the number of components of f S. Then, by
definition, nf 2 or f contains a basepoint in M , since arcs in A give rise to nontrivial elements
in 1 (S, S r M ). In the first case, we call f a closed face, in the second case an open face.
Let F c (S, M, A) be the subset of those f F (S, M, A) for which f is closed. Furthermore, let
n = lcm(nj , j F (S, M, A)).
p1
s2 (a5 ) a5 s1 (a5 )
s2 (a1 ) p2 s2 (a2 ) p5 p4 s2 (a3 )

a1 a2 a3 p7 p8
s1 (a1 ) p3 s1 (a2 ) p6 m2 s1 (a3 )
s2 (a4 ) a4 s1 (a4 )
m1
(a) An arc system A on the marked surface (S, M ) = (D2 r3D2 , {m1 , m2 }). The boundary components
are labelled by the elementary algebra elements pi , which correspond to the arrows in the quiver
Q(S, M, A) below.
p1 a5
p5 p4
p2
a1 p3
a2 a3 p7 p8
p6
a4
(b) The quiver Q(S, M, A) for the arc system A on the marked surface (S, M ) above
Figure 21. An illustration of the definitions 4.1 and 4.4
Remark 4.2. Our conventions are slightly different from those in [HKK14]: Their markings
M correspond to the subsets S r M . Also, their arc systems allow faces of arbitrary genus.
Our arc systems correspond to their admissible arc systems, except that they require nf 3.
Example 4.3. Figure 21a shows an example of a marked surface (S, M ) with an arc system A.
We usually draw the boundary of the surface in green and mark elements in M by dashes
through the boundary, in this case labelled by m1 and m2 . The oriented arcs in A are drawn
in red, labelled by ai , i = 1, . . . , 5. The neighbourhood of each arc ai is shaded in light red,
bounded by solid arrows representing the two sides of ai . In this particular example, the arcs
cut the surface into three components, which are the faces in F (S, M, A). For each face f , we
have labelled the components of S f which do not contain the two points m1 and m2 by
variables pi . These correspond to the basic algebra elements from the next definition.
Note that this particular marked surface will be irrelevant for our tangle invariants; it simply
serves as an example to illustrate the generality of the arguments in this section.
Definition 4.4. Given an arc system A on a marked surface p1 p2

(S, M ), consider the graph Q(S, M, A) obtained by contract- a
S S
ing the closed neighbourhood of each arc in A to a single q1 q2
point (which defines the vertices of Q(S, M, A)) and remov-
ing the interior of each face f as well as any component of Figure 22. A typical neigh-
S f containing a basepoint in M (which defines the edges bourhood of an arc a
of Q(S, M, A)). By choosing the induced boundary orientation on f , the 1-cells inherit an
orientation from (S, M ), which turns Q(S, M, A) into a quiver. Geometrically, we think of
Q(S, M, A) as a deformation retract of the surface S with a single puncture in every closed face.
A typical neighbourhood of the image of an arc a A is shown in figure 22; as we can see,
each arc a is the starting (ending) vertex of at most two arrows. We associate with the triple
(S, M, A) the following two path algebras with relations:
A := FQ(S, M, A)/R and A := FQ(S, M, A)/R ,
where
R = {p1 q1 = 0 = q2 p2 | arcs a A} and R = {p1 p2 = 0 = q2 q1 | arcs a A}.
Note that we follow the convention to read algebra elements from right to left. Every arrow in
the quiver corresponds an algebra element which we call the elementary algebra elements.
For each arc a A, denote the idempotent corresponding to a by a and let I be the ring of all
idempotents. We can consider A and A as two categories: The underlying objects are given
by the arcs in A. Then for any a, b A,
MorA (a, b) := b .A.a and MorA (a, b) := b . A .a
and multiplication is given by algebra multiplication.
Definition 4.5. Let A be an arc system on a marked surface (S, M ) and f F (S, M, A).
Given two sides s and t in S(f ), consider a path ts on f connecting s to t when oriented
according to the boundary orientation of f and missing some point on S f . If ts contains
a point in M , we define pst = 0. Otherwise, ts is homotopic to a path of arrows in Q(S, M, A)
and thus to an element in the algebra A, which we denote by pst . Note that if s = t, pst = a(s) .
Given a side s of a closed face f , define Us to be the algebra element corresponding to the
path that starts at the side s and traverses the boundary of f exactly once. Let Uf be the sum
of the algebra elements Us over all s S(f ). For open faces f , set Uf = 0.
The standard basis of A as a vector space over F2 is given by non-zero elements of the form
pt Ufk , where s and t are two sides of a face f and k is some non-negative integer. For fixed f ,
s
we call these elements of A f -pure, or generally for all faces, pure. The f -pure elements which
are not equal to the identity are called strictly f -pure.
Given two f -pure elements p and p0 for the same face f , we say p is longer than p0 if the
former is divisible by the latter, ie if there exists some algebra element q such that p = qp0 or
p = p0 q. We say p is strictly longer than p0 if we additionally require that q does not act as
the identity. This defines a partial order on all f -pure algebra elements.
Definition 4.6. Let A be an arc system on a marked surface (S, M ). An R0 -grading on A

is called a -grading, if for all f F c , Uf is homogeneous of degree 2 and the subalgebra of
grading 0 is I. Such gradings always exist: For example, we can define a n2 Z-grading by
setting (p) = n2f for any elementary algebra elements p with f (p) = f .
Definition 4.7. Let A be an arc system on a marked surface (S, M ). Let

X
U= Uf .
f F c (S,M,A)
We denote the dg category CxU (Mat(A(S, M, A))) by CC(S, M, A) and call it the category of
curved complexes associated with (S, M, A).
Example 4.8. Let (S, M ) = (S 2 r 4D2 , ) and let A be q1 4 q4

the arc system consisting of four arcs which divide the sur- 1 d 4
face into exactly two faces each of which has four bound- p1 p4
ary components, as shown in figure 23. Then A = A and
CC(S, M, A) = pqMod. 1 a c 3
Remark 4.9. Usually, we fix a -grading on A in the defin-
p2 p3
ition of the category CC(S, M, A) of curved complexes. This
plays the role of the Z-grading in section 1, except that the 2 b 3
q 2 q3
differential here increases -grading by 1. So actually, with 2
respect to the -grading, one should call CC(S, M, A) the cat-

Figure 23. The marked sur-
egory of curved cocomplexes. We have chosen this convention, face and arc system which are
because, as seen in the previous example, the category pqMod relevant for peculiar modules,
of peculiar modules is a special case of CC(S, M, A), and for see example 4.8
pqMod, there is also a homological grading h, which decreases along differentials by 1. The grad-
ing h is defined as a combination of the -grading and the Alexander grading, see definition 2.9.
The latter is preserved by the differentials, so for the present section it is irrelevant.
Remark 4.10. In [HKK14], an A -structure is defined on A , which then gives rise to the
notion of twisted complexes using an A -version of Cx0 (Mat()).
4.1. Curved complexes as immersed curves with local systems. It is worthwhile to

compare our definition of curved complexes associated with marked surfaces to extendable
type D structures studied in [HRW16]. In fact, we will make use of their algorithm below
in theorem 4.26, which is the main goal of this subsection and an adaptation of [HRW16,
theorem 32] to our setting. The main technical difference is that in [HRW16], the algebra A is
truncated in the sense that for any face f , there is a special elementary algebra element p such
that any f -pure algebra element in A that corresponds to a path in Q(S, M, A) traversing p more
than once is set to zero. We work instead with the full algebra A. A posterior, we will see that
these two perspectives coincide and that it is actually sufficient to work with marked surfaces
and arc systems in which each face has a basepoint, see corollary 4.33 and corollary 4.38.
Remark 4.11. In what follows, we will think about objects and morphisms in CC(S, M, A) in
terms of labelled graphs as in section 1. However, in this particular case, the graphs can be
immersed into the marked surface (S, M ) with arc system A as follows: vertices corresponding to
generators of an object are mapped to pairwise disjoint points on the arcs in A corresponding
to the idempotents of the generators. Arrows labelled by idempotents are mapped into the
neighbourhoods of the corresponding arcs in A. For each face f , each arrow labelled by a
strictly f -pure element a = pst Ufk is mapped to an oriented path in S connecting the images of
the endpoints such that the induced path in Q(S, M, A) corresponds to the algebra element a.
In particular, the path can be drawn such that it only intersects the face f and only the two
sides s and t and that it only intersects A in the endpoints.
Definition 4.12. Let A be an arc system on a marked surface (S, M ). A pre-curve L is a

pair ({Pa }aA , {f }f F (S,M,A) ), where
for each arc a A, Pa GLna (F2 ) for some fixed na Z0 ,
for each face f F (S, M, A), there is a map f : B(f ) B(f ) (M f ), where
B(f ) := {(s, i)|s S(f ), 1 i na(s) },
satisfying the following properties:

for B0 (f ) := f1 (B(f )), f (B0 (f )) = B0 (f ) and f |B0 (f ) : B0 (f ) B0 (f ) is a fixed
point free involution;
if f ((s, i)) = (s0 , i0 ) for some (s, i) B0 (f ), then s 6= s0 .
Given a -grading on the algebra A, a compatible -grading on such a pre-curve is an R-
grading on the sets B(f ) satisfying the following properties:
(Pa )ij = 0 if (s1 (a), j) 6= (s2 (a), i) and
(t, j) (s, i) = 1 (pst ) for (s, i) B0 (f ) and f (s, i) = (t, j).
We denote the set of all -graded pre-curves of (S, M, A) by Cpre (S, M, A).
Remark 4.13. Given a pre-curve L = ({Pa }aA , {f }f F (S,M,A) ), we usually represent the
elements in B(f ) by pairwise different dots (s, i) on the sides specified by the first components
and ordered by the second components according to the orientation of the respective side. Then,
the map f can be interpreted as a set of paths on the face f as follows: We draw an embedded
path between each pair of points matched by the fixed point free involution f |B0 (f ) . If f is a
closed face, B0 (f ) = B(f ), so there are no unmatched points left. Otherwise, there is a single
point m M f . In this case, we draw an embedded path from each of the remaining points
in B(f ) to m. For an illustration, see figure 24a. We call the paths matching point in B(f ) to
point in B(f ) (M f ) f -joins, or if we do not specify f , we just talk about joins.
With this graphical interpretation, the second condition in the previous definition says that
no path returns to the same side. However, the corresponding arcs might be the same, as
illustrated in figure 24a.
Definition 4.14. Given an arc system A on a marked surface (S, M ), we define a map
0 : Cpre (S, M, A) ob(CC(S, M, A))
as follows: Given a pre-curve ({Pa }, {f }), let the a -component of its image under 0 be given
by the vector space Va = Fna . We identify the elements (s1 (a), i), i = 1, . . . , na in B(f (s1 (a)))
with the standard basis vectors of Va , which we denote by
s (a)
{eai := ei 1 }i=1,...,na .
This is the preferred basis on Va . We then regard the invertible matrix Pa as a base change
matrix which gives rise to a second basis
s (a) s (a)
{ei 2 := Pa1 .ei 1 }i=1,...,na

1 0
p1 Pa5 =
0 1
b b
Pa5
b b
p2
b
b
p5 p4 b b
Pa1
Pa2
Pa3
p7 p8
b b b
b p6 2
m
p3
b
Pa4 b

1 1
Pa3 =
1
m 1 0
(a) An example of a pre-curve for the triple (S, M, A) from figure 21

ea25
p1 p2
p5
p4 p7
p4
p1 ea15
p4
p2 p5
p7
ea11 ea12 ea13 ea23
p7
p6 p3
ea14
(b) The curved complex corresponding to the pre-curve shown above. Note that any -grading of the
algebra needs to take the value 1 on p7 .
Figure 24. An illustration of the map 0 from definition 4.14
of Va , which we identify with the elements (s2 (a), i), i = 1, . . . , na in B(f (s2 (a))). The -grading
on the sets B(f ) induces a -grading on the bases {esi }i=1,...,na . The first condition on the -
grading for pre-curves ensures that the -gradings induced on each vector space Va = Fna by
the bases of the two sides s1 (a) and s2 (a) agree.
The differential of the curved complex is now easy to describe in terms of these two bases
of the underlying I-module: Let esi be a generator. If f (s) (s, i) M , the pss0 -component of
(esi ) for s0 ranging over all elements in S(f (s)) vanishes. Otherwise, there is exactly one non-
vanishing such component, namely esi0 pss0 , where (s0 , i0 ) is the image of (s, i) under f (s) . Any
0
components involving non-trivial powers of Uf , f F (S, M, A), vanish.

It is now straightforward to check that this defines a curved complex in CC(S, M, A): Indeed,
the only non-zero compositions of arrows in this pre-complex correspond to the fixed point free
involutions f for closed faces f , so the 2 -identity holds. Finally, the second condition on the
-grading for pre-curves ensures that the differential increases -grading by 1.
Example 4.15. An example illustrating the definition above is shown in figure 24.
Proposition 4.16. Every (-graded) curved complex (X, ) in CC is homotopic to one in the
image of the map 0 from the previous definition.
Proof. First of all, we may assume without loss of generality that the differential does not
contain any identity components; otherwise, since X is -graded, any arrow labelled by an
idempotent does not have any other labels, so we can reduce the number of generators by
cancellation (lemma 1.23). Then, for every arc a A, set na = dim(a .X).
Step 1. Fix a face f . Let us choose a -homogeneous basis {esi } of s .X for all sides s.
Interpreting (X, ) as a graph, we can consider the subgraph G of (X, ) consisting of the
vertices corresponding to {esi } with s S(f ) and those arrows labelled by f -pure elements
pst Ufn for s, t S(f ) and n 0. Consider an arrow in G labelled by a shortest algebra element
p = pst Ufn 6= 0 among all labels of arrows in G. Let us call the source of this arrow esi and
the target etj . For open faces, clearly n = 0 and since there are no identity labelled arrows,
s 6= t. For closed faces f , 2 (esi ) contains a term Uf a, so a shortest arrow leaving esi has to
be labelled by an element strictly shorter than Uf . Thus, in both cases, p = pst with s 6= t.
Suppose there is another arrow in G, labelled by psr Ufm
esi etj
leaving esi and going to a generator erk . Then by assumption, pst
b
there exists some algebra element q such that qpst = psr Ufm .
Consider an arrow h1 labelled by q going from etj to erk . Then
h1 is a morphism of -grading 0 and the only algebra elements q
s m
p r Uf
of -grading 0 are the idempotents, so h21 = 0. Furthermore,
we see that the compositions with the differential on X have
erk
b
-grading 1. Hence, if there were an arrow from erk to etj
labelled by prt Ufm for some integer m0 , the composition with
0
q would be a power of Uf and thus have an even -grading. So Figure 25. The homotopy h1
h1 h1 = 0 as well, so h1 satisfies the hypothesis of lemma 1.25.
If q 6= t , we apply this lemma to h1 , which removes the arrow esi to erk . Note that this only
affects arrows in the subgraph G of the complex (X, ). If t = r and q = t , such a homotopy
would also affect arrows in the rest of the complex (X, ). So instead, we consider such a
homotopy as a base change of {eti }. We now repeat this precedure until there are no arrows in
G leaving esi .
Similarly, we can achieve that there are no arrows labelled
esi etj
by prt Ufm going into etj . Indeed, any such arrows have to be at pst
b
least as long as as p, so by the same argument as above, we

can either perform a similar homotopy h2 , see figure 26, or do
a corresponding base change. q
psr Ufm
Finally, note that there is at most one arrow going into
esi , namely one labelled by pts , starting at etj , because of the
erk
b
2 -relation. In fact, there is such an arrow, iff the face f is

closed. We can argue similarly that there is at most one arrow
leaving etj , namely the one going to esi labelled by pts and that Figure 26. The homotopy h2
there is such an arrow iff f is closed.
We have now split off a graph with two vertices and one or two arrows between them. We
can now apply induction on the number of vertices in G to see that we can split G into pairs of
vertices and, once there are no arrows left, singletons. This defines the map f . Note that if f
is closed, there are no singletons, so the whole map f is a fixed point free involution.
Step 2. After doing step 1 for all faces f , we can now define the matrices Pa . These are simply
the base change matrices corresponding to the bases for the two sides s1 (a) and s2 (a) for each
arc a. (Note that these need not be the same as the ones chosen at the beginning of step 1.)
It is now straightforward to check that ({Pa }, {f }) is a well-defined pre-curve. A -grading
is induced by the -grading on the basis elements. Finally, by construction, the image of this
pre-curve under 0 is homotopic to (X, ).
Remark 4.17. In some sense, to made precise in the proof of theorem 4.26 and definition 4.19
below, our pre-curves correspond to train-tracks from [HRW16, definition 17] of a special form.
Proposition 4.16 above corresponds to propositions 22 and 24 in [HRW16]. While their proofs
rely on the fact that extendable type D structures are defined over the truncated algebra A (see
remark 4.10), our proof relies in essential points on the -grading of our curved complexes over
the full algebra A: Indeed, in the proof above, we first reduce the complex, using the fact that
any arrow labelled by an idempotent is not labelled by another algebra element. If we dropped
the -grading, we might also see arrows labelled by 1 + Uf , which do not have an inverse in A!
(One might be able to rectify this by using power series.) Secondly, in step 1 of the proof, we
repeatedly apply the clean-up lemma. To verify that all hypotheses of the lemma are satisfied,
we use the -grading again.
Definition 4.18. Given three positive integers i, j, n satisfying i, j n and i 6= j, we define

the elementary matrix Eij GLn (F2 ) by
i j

1

..
i . 1
Eij := (i0 j 0 + ii0 jj 0 )i0 j 0 = ..
.

j
.

1
Furthermore, let Pij GLn (F2 ) be the permutation matrix corresponding to the permutation
(i, j).
Definition 4.19. We now introduce a graphical notation for

s2 Pij Eij s1
the matrices Pa in a pre-curve on a marked surface (S, M ) b b
with an arc system A. For each arc a A, choose a decom-

position of Pa into elementary matrices i
b b
>
b b
j
Pa := Pala . . . Pa1
b b
for some la 0, where for each k = 1, . . . , la , Pak is either

equal to some Eij or Pij , for some i = i(k) and j = j(k).
(Such a decomposition always exists; this is basically Gaussian Figure 27. A train-track in a
elimination.) Divide the neighbourhood N (a) of a along some neighbourhood of an arc
leaves of Fa into la segments and order them along a path from s1 := s1 (a) to s2 := s2 (a). Thus
we can label the k th segment by the matrix Pak . Next, connect the dots on the two sides s1
and s2 corresponding to points in B(f (s1 )) and B(f (s2 )) by pairwise disjoint arcs which are
(T1) (T3)
Figure 28. Some train-track moves corresponding to a change of matrix decomposition

in definition 4.19
transverse to the foliation Fa . If Pak = Pij for some i 6= j, change the ith and j th arcs in the
k th segment such that they cross once and remain transverse to the foliation, while leaving the
other arcs unchanged. We call this a switch. If Pak = Eij for some i 6= j, we draw a pair of
arrows in the k th segment from the j th arc to the ith arc, whose ends are tangential to the two
strands and which intersect each leaf of Fa at most once. We call this a double-arrow. Both
double-arrows and switches are illustrated in figure 27.
If we add the joins to this picture, the result is an immersion of a graph with both oriented
and unoriented edges. We call it the train-track associated with a pre-curve. It is well-defined
up to homotopy through train-tracks and the following moves, which are illustrated in figure 28:
(T1) A switch can be expressed in terms of double-arrows, since Pij = Eij Eji Eij .
(T2) Two adjacent identical double-arrows can be cancelled, since Eij Eij is the identity.
(T3) Any two double-arrows can be moved past one another unless one strand is the start
of one and the end of the other; in such a case, a double-arrow connecting the other two
strands needs to be added, as shown in figure 28. This corresponds to the identities

E 0 0E if i 6= j 0 and j 6= i0
i j ij
Eij Ei0 j 0 = E 0 0E E 0
i j ij i j if i = j 0 and j 6= i0

E 0 0 E E 0
ij ij ij if i 6= j 0 and j = i0 .
Lemma 4.20. Let L = ({Pa }, {f }) be a pre-curve on a marked surface (S, M ) with arc
system A. Let f F (S, M, A), s S(f ), a = a(s) and fix some distinct integers i and j with
1 i, j na . If L is -graded, assume that (s, i) and (s, j) have the same -grading. Then
0 (L) is homotopic to 0 (L0 ), where L0 is the pre-curve obtained from L by one of the following
three moves:
(M1) Multiply Pa on the right/left by Pij , depending on whether s = s1 (a) or s = s2 (a),
and precompose f with the map that permutes (s, i) and (s, j).
(M2) Suppose neither f (s, i) nor f (s, j) lie in M and they belong to the same side s0 of
some arc a0 , so f (s, i) = (s0 , i0 ) and f (s, j) = (s0 , j 0 ) for some integers i0 and j 0 . Then
multiply Pa on the right/left by Eij , depending on whether s = s1 (a) or s = s2 (a), and
multiply Pa0 on the right/left by Ei0 j 0 , depending on whether s0 = s1 (a0 ) or s0 = s2 (a0 ).
(M3) Suppose at least one of f (s, i) and f (s, j) lies in M or they both belong to different
sides of f . Without loss of generality, assume also that if we follow along a path along
the boundary of f (s), starting at s, we do not meet f (s, i) before f (s, j). Then multiply
Pa on the right/left by Eij depending on whether s = s1 (a) or s = s2 (a).
Pij j b b
j b
j j b

b b b b
i i i i
(M1) (M2)
mM
b b
b
j b
j b
j b
j
mM
b b b b
i i i i
b b
mM
(M3a) (M3b) (M3c) (M3d)
Figure 29. Graphical calculus for lemma 4.20. (M1) shows a local picture of the pre-
curve L on the left and of L0 on the right. For the other moves, only the local pictures
of L0 are shown; L is obtained by removing all double-arrows.
Definition 4.21. As a consequence of the previous lemma, the map 0 induces a well-defined
map
0 : Cpre (S, M, A)/ ob(CC(S, M, A))/ ,
where the equivalence on the left hand side is given by the moves (M1) to (M3) of the previous
lemma and on the right hand side denotes homotopy equivalence.
Remark 4.22. In terms of train-tracks from definition 4.19, we can interpret the three moves
from the lemma above graphically, as illustrated in figure 29. Using (T2), we see that we
can reformulate (M2) as follows: if there is only one double-arrow in the picture for (M2) in
figure 29, then we may remove that double-arrow and replace it by the other double-arrow,
thus, pushing the double-arrow along parallel joins. For (M3), there are four different cases to
consider, depending on whether the joins end on sides or points in M .
proof or lemma 4.20. All three parts of the lemma can be

t
shown in the same way, namely by doing a base change, which b
in our graphical notation shifts the outermost segment of N (a)

adjacent to s into the face f . For (M1), the result after such a pts b
s h = ptt
base change obviously agrees with 0 (L). For (M2), we do two pst b
such base changes, one for s and one for s0 , and again 0 (L0 )
agrees with 0 (L). For L00 , we do one such base change; the b
t
result then differs in at most two arrows from the one for
0 (L), namely the ones shown in figure 30. However, we can
Figure 30. The homotopy for
homotope such arrows away using the homotopy h from the the proof of invariance of (M2)
same figure.
Definition 4.23. Let (S, M ) be a marked surface with an arc system A. A curve on (S, M, A)
is a pair (, X), where either
(1) is an immersion of an oriented circle into S, representing a non-trivial primitive element
of 1 (S) and X GLn (F2 ) for some n; or
(2) is an immersion of an interval (I, I) into (S, M ), defining a non-trivial element of
1 (S, M ) and X = id GLn (F2 ) for some positive integer n
restricted to each component of the preimage of each face is an embedding and
restricted to each component of the preimage of the neighbourhood N (a) of each arc
a is an embedding, intersecting each leaf of Fa exactly once.
In case (1), we call (, X) a loop, in case (2) a path. We say that a curve (, X) is supported
on the immersed curve and call the underlying curve and X its local system. Note
that the local system of paths only records some positive integer n. We consider loops up to
homotopy of the underlying immersed curves through curves. Furthermore, we consider the
local systems of loops up to matrix similarity.
A -grading on a curve (, X) is an R-grading on the set of intersection points of the
underlying curve with arcs in A satisfying the following property: Let x and y be two intersection
points of -grading (x) and (y), respectively. Suppose x and y are joined by a component of
r A. Then such a component is mapped to a path in Q(S, M, A) corresponding to a pure
algebra element of the form pst , from x to y, say. Then we ask that 1 (pst ) = (y) (x).
A collection of (-graded) curves is a finite set of (-graded) curves such that all underly-
ing curves are pairwise non-homotopic as unoriented curves. We denote the set of all collections
of -graded curves up to equivalence by C(S, M, A).
Remark 4.24. The definition of C(S, M, A) is essentially independent of A except for the
-grading.
Definition 4.25. Given an arc system A on (S, M ), we define a map
00 : C(S, M, A) Cpre (S, M, A)/
as follows: Given a single curve (, X), choose a small immersed tubular neighbourhood of
and replace by dim X parallel copies thereof in this neighbourhood. Then, for each face
f F (S, M, A), the joins of 00 (, X) are given by the intersection of these curves with f . Then
pick an intersection point x of an arc a with . Let the matrix Pa be the diagonal block matrix
with blocks of dimension dim X such that all blocks are equal to the identity matrix except the
one corresponding to the intersection point x. We define this block to be equal to X if goes
through x from the right of a to its left, and set it equal to X 1 otherwise. On all other arcs,
we choose the identity matrix. Finally, we extend 00 to collections of curves by taking unions
on both sides.
We also define the map
: C(S, M, A) ob(CC(S, M, A))/
as the composition of 00 and 0 . Note that the definition of this composition is independent
of the choice of a and x up to homotopy in CC(S, M, A), which can be seen by repeatedly
applying (M2). Similarly, we see that conjugation of the local systems X of a loop (, X) does
not change the homotopy type of the image under .
Theorem 4.26. Any peculiar module is homotopic to a peculiar module which is in the image
of .
Proof. Train-tracks associated with pre-curves ({Pa }, {f }) in our setting for which the maps
f are all involutions correspond to certain types of train-tracks in the language of [HRW16],
namely those in which all arrows come in pairs and only sit in a neighbourhood of the arcs.
Then the moves from [HRW16, proposition 24] on train-tracks correspond exactly to our
moves (T1) to (T3) and (M1) to (M3) on our train-tracks. So we can apply the algorithm
explained in [HRW16, section 3.7] for simplifying train-tracks, since the geometric object agree,
even though they represent algebraic objects defined over slightly different algebras. The output
of the algorithm is train-tracks whose double-arrows only connect parallel immersed curves, ie
loops.
The same algorithm also works without any changes for pre-curves ({Pa }, {f }) for which
the maps f are not necessarily involutions, since the additional moves for such pre-curves from
lemma 4.20 namely (M3b), (M3c) and (M3d) from figure 29 can be regarded as special cases
of (M3a). Once all arrows only connect parallel immersed curves, we can remove all arrows on
paths by applying moves (M2) followed by (M3d).
4.2. Classification of morphisms between curved complexes.
Definition 4.27. Let L = (, X) and L0 = ( 0 , X 0 ) be two -graded curves on a marked surface

with arc system (S, M, A) with dim X =: n and dim X 0 =: n0 . Assume, that if and 0 are
parallel, their orientations agree. Let M 0 be the set of basepoints on S obtained from M by
small pushoffs in the direction of the boundary orientation on S and require that if 0 is a
path, its ends lie on points in M 0 . Homotope and 0 such that they intersect minimally,
transversely and away from the neighbourhoods N (a) of the arcs a A. (Minimality can be
achieved through a sequence of homotopies of the underlying curves which strictly reduce the
number of intersection points by removing any immersed bigon in the sense of definition 4.42.)
Let t (, 0 ) denote the vector space spanned by intersection points between and 0 . Each
intersection point can be -graded as follows: follow the first curve from the intersection point
to the first point x that lies on A. Then go back along that path to the intersection point,
turn left onto the second curve and follow it to the first point y that lies on A. Both x and y
are -graded. Define the -grading of this intersection point to be (y) (x) + (ps(y) ). Note
s(x)
that there are two choices for x, but the -grading does not depend on this choice. If and 0
are parallel, let (, 0 ) be the unique real number one needs to add to the -grading of each
intersection point of with arcs in A such that and 0 agree as -graded curves.
For any non-negative integer m, let V (m) be an m-dimensional vector space in -grading
R. If and 0 are parallel, we define HF(L, L0 ) to be

(10) V0 (n n0 ) t (, 0 ) (V0 (1) V1 (1)) V(, 0 ) dim ker (X 1 )t X 0 id .
Otherwise, we set HF(L, L0 ) equal to
(11) V0 (n n0 ) t (, 0 ).
Theorem 4.28. Let L = (, A) and L0 = ( 0 , B) be two curves. Then
H (Mor((L), (L0 )))

= HF(L, L0 ).
Remark 4.29. The notation HF(L, L0 ) will be justified in subsection 4.4. As a consequence
of the theorem above, we obtain that HF(L, L0 ) is independent of the representatives of the
underlying immersed curves of L and L0 , as long as they intersect transversely and minimally.
In subsection 4.4, we will get rid of the minimality condition.
Proof. Let f be a face. Consider a single f -join of 00 (L) between s and t and a single f -join of
00 (L0 ) between s0 and t0 where s and s0 are sides of f , and t and t0 are either sides of f or points
in M f or M 0 f , respectively. Moreover, consider all possible morphisms represented by
arrows labelled by f -pure algebra elements between the generators corresponding to s, s0 and t
and t0 (if they do not correspond points in M f or M 0 f ). Let us restrict the differential D
defined on the morphism spaces (see definition 1.5) to the face f and consider the morphisms
in the kernel of this restricted differential modulo the images of morphisms labelled by strictly
f -pure algebra elements. Then, for the various configurations of s, t, s0 and t0 , these morphisms
are illustrated in figure 31. Let us discuss these pictures in some more detail. They fall into
four different classes, indicated by the colours of the arrows:
empty: For configurations shown in figures 31a, f, j and o, all morphisms in the ker-
nel of the restricted differential lie in the image of morphisms labelled by strictly
f -pure algebra elements. These configurations do not contribute any morphisms to
H (Mor((L), (L0 ))).
black: For configurations shown in figures 31b, g, k and n, the vector space of such
morphisms is generated by one element, namely the morphism represented by the black
arrow(s). These morphisms also lie in the kernel of the unrestricted differential D and
not in the image thereof. These configurations contribute exactly one generator to
H (Mor((L), (L0 ))).
red: For configurations shown in figures 31c, h, l and p, the vector space of such morphisms
is also generated by one element; the difference to the previous case is that at least one
of the arrows of the morphism is labelled by an idempotent. Thus, such a morphism
does not necessarily lie in the kernel of D. However, it does not lie in the image of D.
blue: Finally, for configurations shown in figures 31d, e, i and m, the vector space of such
morphisms is generated by one element. In this case, the morphisms also lie in the kernel
of the unrestricted D, but possibly in the image thereof, as they can be homotoped away
along identity arrows.
If we replace the joins in each picture by parallel copies thereof, we simply need to multiply
the number of morphisms by the multiplicities of the joins. We can now piece those local
pictures together. Morphisms of red type only constitute a morphism in H (Mor((L), (L0 )))
iff the red arrows on one side of an arc are matched up by the local system on that arc with
red arrows on the other side of the arc. Thus, if and 0 are not parallel loops, the red arrows
contribute iff the two segments of and 0 intersect once, and this intersection is counted
with the multiplicity of the product of the dimensions of the local systems. Similarly, the blue
arrows only contribute to H (Mor((L), (L0 ))) iff the blue dotted arrows match up. Again,
this corresponds to a a single intersection point between and 0 .
s
b
t
b
tb
t
b
L L L L
L s b b
t
s b
s b
t b b
t t b b
s
L s b
s b
L s b b
t
b b
L b
L b L
s s t t
(a) (b) (c) (d) (e)
s
b
t
b
M t b
L L L
L s b
s b
t b
M s b
M
L s b
s b
L
b b
L b
L
s s t M
(f) (g) (h) (i)
s b
M t
b
M
L L L
L s b
s b
M b
t s b b
t
L s b
s b
L
b b
L L b
s s M t
(j) (k) (l) (m)
s b
t
b
L L
L M b
t
M M
M M t
L L M b
b b L
s t
(n) (o) (p)
Figure 31. Generators of morphism spaces. The first row shows all configurations
without an end of a path. (However, there might be a point of M in the face, in which
case, some of those arrows are actually zero.) The second and third rows show those
configurations with either an end of a path on M or M 0 , respectively. The last row
shows those configurations with two ends of paths.
If and 0 are parallel loops, there are potentially some

L
additional morphisms that do not correspond to intersection b b
points, namely those corresponding to local pictures of the X X X X

form illustrated in figure 31e. In figure 32, we have redrawn
this picture; here, however, the curves carry local systems and
b b
L
the arrows labelled by X represent unions of arrows corres- A
ponding to some matrix X in Matmn (F2 ) =: G. The red B
arrows only contribute iff XA = BX. Two pairs of blue ar-
Figure 32. Additional morph-
rows are homotopic iff their difference lies in the subgroup
isms for parallel loops
H := {Y A + BY |Y G} of G. We can also regard H as a
quotient of G via the map : Y 7 Y A + BY . Thus, the number of homotopy classes of blue
arrows is equal to # H
G
= # ker , so it its equal to the number of contributing red arrows. Now
we conclude with the following lemma.
Lemma 4.30. Given two matrices A GLn (F2 ) and B GLm (F2 ) for some positive integers n
and m, then
(12) dim{X Matmn (F2 )|XA = BX} = dim(ker((A1 )t B + id))
Proof. We can regard any X Matmn (F2 ) as a linear transformation V W between two
vector spaces V and W of dimensions n and m, respectively, and thus as an element in V W ,
where V denotes the dual vector space of V . The left hand side of (12) is therefore equal to the
dimension of the kernel of the linear automorphism of V W that sends X to BXA1 . So it
suffices to show that this linear transformation can be identified with the map (A1 )t B. For
this, fix some basis {vi } of V and let {vi } be the dual basis in V . Then X V W can be
P
written as i vi Xvi , as shown in figure 33a. Using the basis {Avi }, we can express BXA1
P
as i (Avi ) BXvi . As the diagram in figure 33b shows, (Avi ) = (A1 ) vi . Dualizing a
linear transformation corresponds to transposing the corresponding matrix, so the diagram in
figure 33a indeed commutes.
4.3. Classification of curved complexes. We are now ready to prove theorem 0.4 as a special
case of the following general classification.
Theorem 4.31. Let L = {(i , Ai )}iI and L0 = {(j0 , A0j )}jJ be two collections of loops.
Then 0 (L) is homotopic to 0 (L0 ) iff there is a bijection : I J such that i is homotopic
0
to (i) and Ai is similar to A0(i) .
Remark 4.32. We expect the same theorem to hold for general curves ie loops and paths. The
arguments that we use in the proof of theorem 4.31 rely on different growth properties of the
dimensions of the two summands in (10) from definition 4.27 under pairing with particular test
curves. However, the second term simply vanishes for paths and thus, a separate argument
would be needed in this case.
Corollary 4.33. Consider a marked surface (S, ) with an arc system A. Let M be a set of
points on S, such that every face f (S, , A) contains at most one point in M . Then, two
P P
i (A
Xvi 1 )t v BXvi
i vi i

(A1 )t B A1
V W V W Avj V V vj
vi
(Avi )

F2
P

i (A
VW VW 1 ) v BXvi
i ij

P P

i vi Xvi = X BXA1 = i (Avi )
BXvi (b)
(a)
Figure 33. Commutative diagrams for the final identification in the proof of lemma 4.30
objects in CC(S, , A) are homotopic iff their images under the induced functor
CC(S, , A) CC(S, M, A)
are homotopic.
Proof of corollary 4.33. The induced functor is a functor of dg categories, as it is induced by

the quotient map
A(S, , A) A(S, , A)/{pm = 0|m M } = A(S, M, A),
where pm is the algebra element corresponding to the boundary component in which m M

lies. Thus the images of two homotopic objects are homotopic. So we may assume that the two
objects are direct sums of loops with local systems. The images of such objects are represented
by the same loops with local systems. By theorem 4.31, these represent homotopic objects
in both CC(S, , A) and CC(S, M, A) iff the curves are the same and the local systems are
equivalent.
Before we prove theorem 4.31, let us pause for a moment to consider what the previous two
results tell us about our tangle invariant CFT .
Definition 4.34. Given a 4-ended tangle T in a homology 3-ball M with spherical boundary,
let LT := LT,M denote the collection of loops associated with CFT (T ) by theorem 4.26, in
conjunction with example 4.8.
Theorem 4.35. LT is a well-defined tangle invariant up to homotopy of the underlying im-

mersed curves and similarity of the local systems.
Definition 4.36. Given a tangle T , pick two basepoints pi and qj for some i, j {1, 2, 3, 4} as
well as one of zj and wj for each closed component of T . A peculiar Heegaard diagram for T
is nice with respect to this choice of special basepoints, if all regions except those containing
these basepoints are bigons or squares.
Theorem 4.37. Every 4-ended tangle T has a nice peculiar Heegaard diagram with respect to
any choice of basepoints.
Proof. This is an application of Sarkar and Wangs main result in [SW06], where they describe
an algorithm for niceifying any pointed Heegaard diagram with one basepoint in each component
of the Heegaard surface minus the -circles.
Corollary 4.38. Peculiar modules for 4-ended tangles can be computed combinatorially.
Proof. We can use a nice peculiar Heegaard diagram for a tangle to compute the peculiar
module CFT (T ). Since all regions away from the special basepoints are bigons or squares,
the calculation of all domains that miss those basepoints is purely combinatorial. The complex
CFT (T ) corresponding to those domains is exactly the image of CFT (T ) under the functor
induced by the quotient map A A /(pi = 0 = qj ). So by corollary 4.33, we can recover
the homotopy type of CFT (T ) from CFT (T ). This can be done algorithmically, by finding a
curve with local system representing CFT (T ) as described at the beginning of this section.
We now turn to the proof of theorem 4.31.
Definition 4.39. Given a polynomial

n1
X
f = xn + ai xi F2 [x],
i=0
define the companion matrix Cf of f to be the matrix

0 a0

..
1 . a1
Cf :=
.. ..
GLn (F2 ).

. 0 .

1 an1
Note that Cf is invertible iff a0 6= 0. A diagonal block matrix of the form

Cf 1
..
, where f1 , . . . , fr F2 [x],

.
C fr
is in Frobenius normal form if fi+1 |fi for all i = 1, . . . , r 1.
Theorem 4.40. Every matrix is similar to a matrix in Frobenius normal form. Two matrices
are similar, if they have the same Frobenius normal form.
Proof. This is standard linear algebra.
Lemma 4.41. Given a polynomial f F2 [x], and A GLm (F2 ) for some integer m,
dim(ker((A1 )t Cf + id)) = dim(ker(f (A))),
where Cf is the companion matrix of f from definition 4.39.
Proof. This follows from the same arguments as [HRW16, proposition 36]. Let n = deg f . By
performing row and column operations, we can bring (A1 )t Cf + id) into block diagonal form
with the first block of dimension (n 1)m equal to the identity matrix and the second block of
dimension m equal to the expression
(13) id +an1 (A1 )t + a0 ((A1 )t )n ,
where the ai are the coefficients of f as in definition 4.39. The kernel of the matrix (13) has
the same dimension as the kernel of ((A1 )t Cf + id). Now multiply (13) by (At )n to obtain
f (At ). Transposing a matrix does not change the dimension of its kernel, so we are done.
Proof of theorem 4.31. For every j J such that there is no i I with j = i , add a formal
curve to L which is supported on j and has a 0-dimensional local system. Do the same for L0 .
Note that this does not change HF(L, L00 ) nor HF(L0 , L00 ) for any curve L00 with local system.
So by allowing 0-dimensional local systems, we may assume without loss of generality that there
exists a bijection : I J such that i = (i) . Fix such a bijection .
Suppose there exists some i I such that (i , Ai ) is a loop such that dim Ai 6= dim A(i) .
Let p be the minimal polynomial of the matrix Ai and p0 be the minimal polynomial of the
matrix A0(i) . Then for N > deg p + deg p0 , let
fN (x) := (xN deg pdeg p + 1) p(x) p0 (x).

0
By lemma 4.41, the dimensions of the morphism spaces from the test curve (i , CfN ) to L
and L0 are equal to
X
#i i0 dim Ai0 N + 2 dim Ai
i0 I
and
X
#i j0 dim A0j N + 2 dim A0(i) ,
jJ
respectively. By considering these two terms as linear functions in N , we see that they coincide
iff their coefficients coincide. Hence, in particular dim Ai = dim A(i) .
So we have now established that if L and L0 represent homotopic curved complexes, we can
find a bijection : I J such that for all i I, i is homotopic to (i) 0 and dim Ai = dim A(i) .
So it remains to show that then also Ai and A(i) are similar for all i I.
For this, let us assume that both the two matrices are in Frobenius normal form defined by
polynomials f1 , . . . , fr and f10 , . . . , fr0 0 such that fl+1 |fl and fl00 +1 |fl00 for all l = 1, . . . , r 1 and
l0 = 1, . . . , r0 1. Then Ai and A(i) are similar, if r = r0 and fj = fj0 for all j = 1, . . . , r.
Suppose this is not the case. Then there exists some minimal m such that fm 6= fm 0 , where
fm = 0 if m > r and fm 0 = 0 if m > r 0 . Assume without loss of generality that f 0 6 |f . Then

m m
consider the test curve given by (i , Cfm ). Let N = deg fm . Then the space of morphisms
from L and L0 to (i , Cfm ) has dimension

X m1
X r
X
(14) #i i0 dim Ai0 N + 2 dim ker(fm (Cfj )) + dim ker(fm (Cfj ))
i0 I j=1 j=m
and

X m1
X r
X
(15) #i i0 dim A0i0 N + 2 dim ker(fm (Cfj0 )) + dim ker(fm (Cfj0 )) ,
i0 I j=1 j=m
respectively. The first sums coincide by the results that we have already established. The
second sums agree by minimality of m. The summands in the third sum of (14) are equal to
dim Cfj = deg fj , since fj |fm for j > m. Now,
r r 0
X X
dim Cfj = dim A0i = dim A(i) = dim Cfj0 .
j=1 j=1
By minimality of m, we obtain
r r 0
X X
dim Cfj = dim Cfj0 .
j=m j=m
Hence, the third sum in (15) is at most as large as the third sum in (14). However, fm
0 6 |f ,
m
hence fm (Cfm
0 ) 6= 0, so
dim ker(fm (Cfm

0 )) < dim Cf 0 .
m
Contradiction.
4.4. The glueing formula revisited. In theorem 4.28, we gave a formula for the homology of
the morphism space between two objects in CC(S, M, A) for a marked surface (S, M ) with arc
system A, in terms of the collections of curves L and L0 representing those objects. We denoted
this space by HF(L, L0 ), because it is also equal to the homology of a finite-dimensional chain
complex associated directly with L and L0 , namely the Lagrangian intersection Floer homology
of (the train-tracks associated with) L and L0 .
The following discussion is an adaptation of the arguments in [Abo06] to our setting, using
the language of [HRW16]. However, note that [HRW16] use slightly different orientation con-
ventions: because of the way they express their glueing theorem, they find it more convenient
to interpret the two collections of curves L and L0 differently, namely one in terms of a type D
structure and the other in terms of a type A structure; we treat both curves the same, which
follows more standard conventions in Lagrangian intersection Floer theory. For example, in
their setting, HF(L, L) vanishes for some objects L; so in particular, there would be no identity
morphism for such L. This does not happen with our conventions.
Definition 4.42. Let (S, M ) be a marked surface with arc

system A. Since the train-track associated with a curve with
x y
local system of (S, M, A) only has double-arrows between par-
b b
allel curves, we may, in the following, consider these train-

tracks up to homotopy and the moves (M2), (T1), (T2) and
(T3) only. We may then also allow double-arrows to lie in the Figure 34. A bigon x y
faces of (S, M, A). So let and 0 be train-tracks associated with two curves with local systems
L = (, X) and L0 = ( 0 , X 0 ) on (S, M, A). Let us assume that all intersection points between
and 0 are transverse. Let CF(, 0 ) be the F2 vector space generated by the intersection points
between and 0 .
Given two such intersection points x, y, let M(x, y) be the set of orientation-preserving maps
: (D2 , ) (S, 0 )

is an immersion away from i D2 C;
the restriction of to the non-negative real part of D2 is a path from x = (i) to
y = (+i) on such that the orientation agrees with orientation of any double-arrows
in ;
the restriction of to the non-positive real part of D2 is a path from y to x on 0 such
that the orientation agrees with orientation of any double-arrows in 0 ;
x and y are convex corners of the image of .
We call an element M(x, y) an immersed bigon. An immersed annulus is an orientation-
preserving immersion
: S 1 [0, 1] S,
which restricts to S 1 {0}, respectively S 1 {1}, to an immersed circle on , respectively 0 ,
whose orientation agrees with the orientation of the double-arrows in , respectively 0 .
If there are no immersed annuli, let d be the endomorphism on CF(, 0 ) defined by
X
d(x) = #M(x, y) y.
y0
(a) A finger move (b) A bridge move
Figure 35. Two moves for which the homotopy type of (CF(, 0 ), d) remains invariant
Theorem 4.43. If there are no immersed annuli, the endomorphism d is a well-defined differ-
ential. Furthermore, the homotopy type of (CF(, 0 ), d) is an invariant of L and L0 , so we may
write CF(L, L0 ). Finally, the homology of this space is given by HF(L, L0 ) from definition 4.27.
Proof. For any two intersection points x and y of the two train-tracks and 0 , define a map
: M(x, y) H2 (S, 0 ) as follows: Given a bigon , define the coefficient of each component
of Sr(0 ) in () to be the number of points in the preimage of some basepoint in r. This map
is well-defined, injective and independent of the basepoints. Since train-tracks are not allowed
to back-track, the two paths from x to y obtained by restricting a bigon to the boundary of
the disc intersect the same arcs in A in the same order. Since is injective, having an infinite
number of bigons between two fixed intersection points x and y would imply that some bigon
can be restricted to a square by removing neighbourhoods of i such that the images of the new
boundary map to the same path, along which we can glue the square to obtain an immersed
annulus, contradiction! So d is a well-defined homomorphism. The proof that d is a differential
is the same as in [Abo06].
Furthermore, one can easily check that the chain complex (CF(, 0 ), d) does not change
under the moves on each train-track that take place locally, away from the other train-track.
It remains to check that the homotopy type of (CF(, 0 ), d) does not change under finger and
bridge moves, see figure 35. The first creates creates/kills a bigon region, the second moves a
single double-arrow past a strand of the other train-track. The first corresponds to a cancellation
and the other to a base change in (CF(, 0 ), d). We leave the details to the reader.
The last part follows from homotoping the train-tracks such that they intersect minimally
without any immersed annuli. Consider two parallel components and assume without loss of
generality that their orientations agree with those in figure 36 and that all double-arrows on

i i
j X j

b y
x b

i i
j X j

Figure 36. An illustration of the identification of bigons between parallel curves and
the matrix (X 1 )t X 0 . Up to conventions, this is the same as [HRW16, figure 43].
these components are contained in the boxes labelled by the corresponding local systems X
and X 0 . Let x = (j, j 0 ) and y = (i, i0 ) be two intersection points on either sides of the local
systems in this picture. Then, modulo 2, (X 1 )ij counts the number of paths from x to y on
and Xj0 0 i0 counts the number of paths from y to x on 0 , so (X 1 )ij Xj0 0 i0 counts the number
of bigons from x to y within this local picture. Finally, there is a bigon from (j, j 0 ) to (i, i0 )
outside this area if i0 = i and j 0 = j. Thus X 1 (X 0 )t id is the differential, which has the
same rank as its transpose (X 1 )t X 0 id.
Theorem 4.44. With the same notation as in theorem 3.6,
CFL(L)
[ V i = HF(LT1 , Lm(T2 ) ) = H (Mor(CFT (T1 ), CFT (m(T2 )))),
where m(T2 ) denotes the mirror image of T2 .
Proof. The second equality is an application of theorem 4.28. By theorem 4.43, we can use any
representatives of the curves in LT1 and Lm(T2 ) , as long as there are no immersed annuli. For
example, we may homotope the curves in LT1 and Lm(T2 ) into neighbourhoods of the blue and
red curves in figure 37, respectively. The first equality is then seen by identifying the intersection
points between those curves and the connecting bigons with the generators and differentials of
CFT (T1 ) P CFT (T2 )
from theorem 3.6.
q1 q4
dd
d d
1 4
a q1 q4 dc c
q2 q3
aa p1 p4 cd
a ab p2 p3 c cc
b b
2 3
bb p3 p4
p2 p1
ba
q2 q3
Figure 37. Geometric interpretation of the type AA structure P for pairing in the
wrapped Fukaya category of the 4-punctured sphere. The boundary of the picture is
identified to a point. The blue curves denote a 1-skeleton and the red ones a Hamiltonian
translate thereof with reversed roles of pi and qi . The orientation is chosen such that
the normal vector (determined by the right-hand rule) points into the projection plane.
5. Applications
5.1. Rational tangle detection.
Observation 5.1. The Alexander grading on CFT (T, M ) implies that for a tangle T , each
loop in LT lies in the kernel of
1 (M r T ) 1 (M r (T )) H1 (M r (T )),
where the first map is induced by the inclusion and the second is the Abelianization map.
Theorem 5.2. A 4-ended tangle T in the 3-ball is rational iff LT is a single embedded loop
with the unique 1-dimensional local system.
Proof. The only-if direction is simply a calculation, see example 2.25. Conversely, suppose LT
is a single loop which corresponds to an embedded loop on the 4-punctured sphere. It divides
the sphere into two disc components, each of which has at least one puncture, since the loop
is not nullhomotopic. By observation 5.1, there are exactly two punctures in each disc, so LT
agrees with LT 0 for some rational tangle T 0 . Then also Lm(T ) agrees with Lm(T 0 ) . Let L be the
link obtained by pairing T with its mirror m(T ) as in figure 16. If we glue T 0 to m(T 0 ), we
obtain the 2-component unlink q , so by theorem 4.44
HFL(L)
[ = HF(LT , Lm(T ) ) = HF(LT 0 , Lm(T 0 ) ) = HFL(
[ q ).
We now apply the fact that link Floer homology detects unlinks [OS06a], so L is the 2-component
unlink. The following lemma finishes the proof.
Lemma 5.3. Let T1 and T2 be two 4-ended tangles without closed components that glue together
to the 2-component unlink. Then either T1 or T2 is a rational tangle.
Proof. Let S be the 4-punctured sphere along which we glue T1 and T2 . Let U be the sphere
that separates the two unknot components and assume that S and U intersect transversely in
a disjoint union of circles. We now proceed by induction on the number of circles in S U .
First of all, this intersection is non-empty, since U is separating. So we can always find a curve
that bounds a disc D in U which does not contain any other curves in U S. If bounds a
disc D0 in S, D D0 bounds a 3-ball, which we can use as a homotopy for U to remove (along
with any other components of S U in D0 ), so we are done by the induction hypothesis. If
does not bound a disc in S, it separates two punctures from the other two. So D separates the
two strands in T1 or T2 . They must obviously be unknotted, since the connected sum of two
knots is the unknot iff both knots are unknots. Thus either T1 or T2 is rational.

5.2. Skein exact sequences. We start with a slight gener-

alisation of Ozsvth and Szabs exact triangle [OS03a] which

.

.
..
categorifies the oriented skein relation for the Alexander poly- n . n

.

nomial. However, we remind the reader that all gradings on

link Floer homology should be regarded as relative, see re-
mark 2.18, so the graded version of the following theorem is (a) Tn (b) Tn (c) T0
not quite as strong as Ozsvth and Szabs result in the case Figure 38. The basic tangles
n = 1. for theorem 0.12
CFT (Tn )
n
CFT (Tn ) HFL(L
[ n ) V ln HFL(L
[ n ) V ln
CFT (T0 ) V [ 0 ) V l0 +1
HFL(L
(a) (b)
Figure 39. The skein exact triangles from theorem 5.4
Theorem 5.4 (n-twist skein exact triangle). Let Tn be the positive n-twist tangle, Tn the
negative n-twist tangle and T0 the trivial tangle, see figure 38. Furthermore, let V be a 2-
dimensional vector space supported in degrees 0 tn and 0 tn , where t is the colour of the two
open strands. Then there is an exact triangle shown in figure 39a. n preserves the (single-
variate) Alexander grading and changes - and homological gradings by +1 and 1, respectively;
the other two maps preserve the all three gradings. Moreover, given three links Ln , Ln and L0
in S 3 , which agree outside a closed 3-ball and in this closed 3-ball agree with the 4-ended tangles
Tn , Tn and T0 , respectively, then the above triangle together with the glueing theorem induces
an exact triangle shown in figure 39b, where for i {n, n, 0}, li is either 0 or 1, depending
on whether the two strands in Ti belong to different or the same components in Li , respectively.
Remark 5.5. Similar results hold for other orientations; for n even, also multivariate Alexander
gradings are preserved.
Proof. It is straightforward to compute CFT (Tn ) and CFT (Tn ) from genus 0 Heegaard
diagrams. They are both shown in figure 40, the former on the left, the latter on the right. If n
is odd, the dashed lines denote a sequence of alternating generators in sites a and c, connected
by pairs of morphisms, labelled alternatingly by pi s and qi s. For even n, the two components
are connected by similar sequences along the dotted lines. The horizontal arrows in figure 40
1 1 1
p3 2 c1n p321
2 c1n
p21 +q34
0 bn 0 dn
p214 p4
q341 q2 q1 q234
1 1 1
2 a1n 2 a1n
1 1 1
2 cn1 2 cn1
q4 q123 q412 q3
p43 +q12
p432 0 dn p2 0 bn
1
p1 1
1 p143
2 an1 2 an1
Figure 40. The morphism n : CFT (Tn ) CFT (Tn )


CFT ( ) CFT ( ) HFL(L
[ 0 ) V l0 HFL(L
[ 1 ) V l1
CFT ( ) HFL(L
[ X ) V lX
(a) (b)
Figure 41. The skein exact triangles from theorem 5.6
describe n . By cancelling all identity components of the mapping cone of n , we get two copies
of
p43 +q12
CFT (T0 ) = 0 b0 0 d0 .
p21 +q34
Thus, we can write CFT (T0 ) V as a cone of CFT (Tn ) and CFT (Tn ), which gives rise to
the exact triangle of the required form.
Next, we give a new proof of a theorem by Manolescu [Mnl06, theorem 1]. Like Manolescus
triangle, ours does not preserve any gradings. Note that Manolescu uses slightly different
conventions from ours, so the two triangles only look the same after reversing the direction of
the three arrows.
Theorem 5.6 (resolution skein exact triangle). There is an exact triangle shown in figure 41a.
Moreover, given three links L0 , L1 and LX in S 3 , which agree outside a closed 3-ball and in this
closed 3-ball agree with the 4-ended tangles T0 = , T1 = and TX = , respectively, then
the above triangle, together with the glueing theorem induces the exact triangle from figure 41b,
where for i {0, 1, X}, li is either 0 or 1, depending on whether the two strands in Ti belong
to different or the same components in Li , respectively.
Proof. The map is given by (the horizontal arrows in) the following diagram on the left:
q412
q3
b c b c
q3
p2
p2 p321
p21 +q34 p43 +q12 p32 +q41 p14 +q23 =

p4 p143
p4
q1
q1
d a d a
q234
Using lemma 1.25, we see that it is homotopic to the diagram on the right, which is CFT ( ).
Now apply the same arguments as in the proof of theorem 5.4.
Proposition 5.7. There are two morphisms

1

2 t1 CFT CFT
whose mapping cones are homotopic to the peculiar modules represented by the figure-8 loops
shown in figure 42. Since these loops are invariant under taking the mirror, they agree with the
mapping cones of maps from the negative crossing to the trivial tangle.
p2 p432
1 1
0 a0 2 b+1 0 a0 2 d+1
p143 b b p1
b b
q2 q341 q3 q412 q234 q1 q123 q4
p3 p321
1 1
0 c0 c0
b b
b b
2 b1 2 d1
p214 p4
(a) (b) (c) (d)
Figure 42. Two figure-8 loops (a) and (c) and their peculiar modules (b) and (d)
1 1
1
2 d+1 1 1
2 d+1 2 d1 2 d+1
p432 p4 p432
q1 p1
0 a0
p21 +q34 p1
0 a0 q4 p4
q123
p43 +q12
q123 p43 +q12
q341 q341
p21 +q34
0 c0
q2
0 c0
p2 q2
p3 p3
q3 p214
p214
1 1
12 1
2 b1 b1
b+1 b1
2 1
2
(a) (b)
Figure 43. The two maps from proposition 5.7
Proof. The two maps between the peculiar invariants of the two tangles are shown in figure 43.
After cancelling the identity arrows in both mapping cones, we obtain the two peculiar modules
shown in figure 42. Also, reversing all arrows, swapping pi and qi and reversing the Alexander
grading leaves both of them invariant. Doing this to the mapping cone gives us maps between
the mirrors of the two tangles, but in the opposite direction.
Remark 5.8. It is interesting to compare the figure-8 curve to the local Heegaard diagram for
a singular crossing in [OSS07], as the number of generators agree for the second loop, up to
b
an additional tensor factor. Also note that the proposition above gives rise to an exact triangle
similar to the one in [OS07]. Moreover, we can write the n-twist tangle Tn from figure 38a, with
both strands oriented upwards, as a complex in the objects and . Indeed, cancelling the b
identity components in the complex from figure 44 gives us a loop representing Tn .
b2n b4n
p1
1 1
p2
bn 1
dn
p2 p2
q412 q412
a1n a3n
p143 p143 p143 q4
p14 p14
an1
q412
q3 q3 q3
q2 q341 q2 q341
c1n c3n
q341 p43 +q12
q2
p3 q41 p3 q41 p3
cn1 p21 +q34
p214 p214 p214

bn b2n bn2 bn
Figure 44. A complex of peculiar modules homotopic to CFT (Tn )

Similarly, we can obtain a complex for Tn by applying the mirror operation. Furthermore,
it is also easy to find such complexes for other orientations of Tn and Tn . Note that these
complexes look very much like the ones we get in Bar-Natans Khovanov homology of tangles
[Bar04]. We assume that every tangle can be written as a complex in the two objects and b
, or the two objects and , depending on the orientation. In fact, we can iteratively use
the type AA glueing structure from theorem 3.6 together with the skein exact sequence from
theorem 5.4 to locally modify tangles until we obtain a complex of peculiar modules of trivial
and 1-crossing tangles, up to a large number of tensor factors from glueing. Then, one (only)
needs to get rid of these extra factors. For example, in the case of the (2, 3)-pretzel tangle,
this is indeed possible.
It is also interesting to compare our figure-8 curve to the curve that Hedden, Herald, Kirk
associate with a trivial tangle in [HHK13, figure 10] in the context of instanton tangle Floer
homology in the pillowcase.
5.3. Peculiar modules of (2n, (2m+1))-pretzel tangles.

t1 t2

Theorem 5.9. The peculiar modules of (2n, (2m + 1))-

2m + 1

pretzel tangles for n, m > 0, oriented as in figure 45, are ... ..
.
2n

represented by the curves shown in figure 46.
Remark 5.10. For n = m = 1, this calculation was already t1 t2

done in example 2.26 directly from the definition of peculiar
modules. The general case uses the combinatorial algorithm Figure 45. The pretzel tangle
for computing peculiar modules from corollary 4.38. from theorem 5.9
Theorem 5.11. Let T be a tangle in the closed 3-ball B 3 and T 0 the tangle obtained by relabelling
the sites such that a and c, and b and d are interchanged. If T is oriented, orient T 0 such that
the orientation at the first tangle ends of T and T 0 (and hence any others) agree, by either
changing the orientation on all strands or leaving them all the same. Then, if CFT (T ) and
CFT (T 0 ) are (graded) chain-homotopic, mutation of these tangles preserve (graded) link Floer
homology.
Proof. This follows directly from the definition of mutation and the glueing theorem.
Corollary 5.12. Mutation about (2n, (2m + 1))-pretzel tangles for n, m > 0, oriented as in
figure 45, preserves bigraded link Floer homology, after identifying the Alexander gradings of the
two open strands. If we reverse the orientation of one of the two strands, mutation in general
only preserves -graded link Floer homology.
Proof of corollary. The invariants of the (2n, (2m+1))-pretzel tangles simply have the desired
symmetry. This can be seen as follows: In terms of the loops on our infinite chessboard, a
relabelling of the sites as in theorem 5.11 corresponds to a recolouring of the vertices: and
. An orientation reversal of both tangle strands corresponds to a rotation of the chessboard
by . After identifying the Alexander gradings of the two tangle strands, all generators on the
diagonals from bottom-left to top-right have the same Alexander grading. Finally, if we reverse
the orientation of one strand, we do not need to rotate the curves, but the generators in the
same Alexander gradings now sit on the diagonals that go from top-left to the bottom-right.
tn n2m2
1 t2 tn n+2m
1 t2
b b b b b b b b b b b b b b b b b
b b b b b b b b b b b b b b b b
b b b b b b b b b b b b b b
b b b b b b b b b b b b b b b b b b
tn1 tn2m
2 tn1 tn+2m+2
2
(a) n m + 1
tn n2m2
1 t2 tn n+2m
1 t2
b b b b b b b b b b b
b b b b b b b b b b b b
t1n2m2 tn
2
tn+2m+2
1 tn2
tn1 tn2m
2 tn1 tn+2m+2
2
(b) n > m + 1
Figure 46. The peculiar module of the (2n, (2m + 1))-pretzel tangle, shown in fig-
ure 45. We use the same conventions as in example 2.26, see also figures 13c and 14a. All
generators for site a and c, ie the red and green vertices, are in the same -grading. The
diagonals connecting pairs of red and green generators of the same Alexander gradings
should be continued in such a way that they do not intersect each other.
Proof of theorem 5.9. The generators of the peculiar module can already determined from the
decategorified invariants and from the observation of two obvious differentials that can be can-
celled as in example 2.26. Thus, the vertices of the graphs in figure 46 are fixed. What we need
to decide is how they are connected. Because of the restrictions given by the gradings, for most
cases, there is only one way to connect them such that the result is a peculiar module. The only
d d d
q1 d q4
|
|
p4
a1 c1
|
|
a2 x1 c1
|
|
x2 c1
|
|
y1 c2
|
y1 c2
|
y1 c2
|
y2
|
y2
|
y2
|
c2m
|
c2m
|
y 2m c2m
|
c2m+1
|
|
y2m
c2m+1
|
|
y 2m
y 2m+1 c2m+1
|
|
a2n1
|
a2n x2n1 y2m+1

y 2m+1
|
x2n
p2 p3 q3
|
b
|
b b b
(a) A niceified Heegaard diagram for a (2n, (2m + 1))-pretzel tangle with n, m > 0
ai yj b0 yj xi b xi cj d0 yj xi d
byj y j b cj yk y k cj dyj yj d
byj y j b cj yk y k cj dyj yj d
(b) Generators of the Heegaard diagram above, where 1 i 2n and 1 j, k 2m + 1. The

generators of the second and third row can be cancelled.
Figure 47. The first step of the calculation of the peculiar modules of (2n, (2m + 1))-
pretzel tangles
xi+1 cj1 xi+1 d xi+1 c2m

q14 q14
q1

ay
ai yj
ay
i 1 i 2m+1
p23 p23 p2
xi1 cj+1 xi1 c2 xi1 b
(a) 1 < i < 2n and 1 < j < 2m + 1 (b) 1 < i < 2n and j = 1 (c) 1 < i < 2n and j = 2m + 1
ai+1 yj1 ai+1 y2m
p23 p23
p4 q3
xi cj xi d xc xc xi b
i 1 i 2m+1
q14 q14
ai1 yj+1 ai1 y2
(d) 1 < i < 2n and 1 < j < 2m + 1 (e) 1 < i < 2n and j = 1 (f) 1 < i < 2n and j = 2m + 1
ai+1 y2m+1 p4
xi d

xi c1
p
2
q3 q1
xi c2m+1 xb
i ai1 y1
(g) 1 i < 2n (h) 1 < i 2n
Figure 48. Some differentials for the computation of the (2n, (2m+1))-pretzel tangle
in non-extremal t1 -Alexander grading
question is how the diagonal strings of red and green generators connect the generators on the
top left to the generators on the bottom right of each of the subfigures of figure 46. For this,
we are going to apply the algorithm from corollary 4.38, setting p1 = 0 and q2 .
We start with a Heegaard diagram obtained by glueing two Heegaard diagrams for the rational
tangles with 2n twist, respectively (2m + 1) twist together. We can niceify the diagram by
doing two handleslides of the -curve for the 2n-twist rational tangle across the other -curve.
The result is shown in figure 47a. The generators of this diagram are shown in figure 47b.
The generators of the second and third row are the ones that were created during the first and
second handleslide, respectively, and thus can be cancelled along the identity arrows connecting
those generators of the same site and with the same indices.
Note that the nice Heegaard diagram has the same symmetry as the tangles, and thus the
complex inherits this symmetry. More precisely, the Heegaard diagram remains invariant under
the following operation: in the names of the generators, exchange the letters d and b, exchange
underlining and overlining, replace i by 2n + 1 i, and j and k by 2m + 2 j and 2m + 2 k,
respectively. Finally, in the algebra, exchange p1 and q2 , as well as p3 and q4 , and q3 and
p4 . This symmetry corresponds to mutation about the horizontal axis, which leaves the tangle
invariant up to exchanging the two sites b and d. We will use this symmetry in the following to
simplify some parts of the computation.
x2 c2m x3 c2m1 y 2m+1 c2m1

q14 q14
q14

a y a y
1 2m+1 2 2m
p2 p23

q4 1 p3 q3
x1 c2m d0 y2m+1 y1b by1 x c x1 b
1 2m+1
(a)
x2 c2j x3 c2j1 y 2m+1 c2j1

q14 q14
q14
q4
a1 y2j+1 a2 y2j dy2j+3

p23 p23
q4 1 1 q4 0
x1 c2j d0 y2j+1 y c c2j+2 y1 x c d y2j+2
1 2j+2 1 2j+1
(b) 1 j < m
x3 c2j2 y 2m+1 c2j2
q14 q14

a1 y2j a2 y2j1

p23
q4 q4 0
x1 c2j1 d0 y 2j x1 c2j d y2j+1

(c) 1 j m
x2 d
q1

a y
1 1
p23

y 1 c2

1
p
0 4
dy c y
1 2 1
1 1
p4 q4
x
1 d
x c
1 1
d0 y2
(d)
at generators in maximal t1 -Alexander grading before cancellation. The dotted arrows
appear iff n = 1, the dashed arrows iff n > 1.
In figures 48 and 49, we compute all differentials that start/end at some selected generators,
which are enclosed in those figures by boxes. Note that in all figures, generators in the same
shaded regions share the same Alexander bigrading. Since the Heegaard diagram is nice, the
only contributing domains are bigons and squares, so the computation is purely combinatorial
and rather straightforward. We therefore ask the reader to check for themselves that indeed,
the differentials starting/ending at all marked generators in those figures are included. (There
are two observations that one might find useful when determining the contributing differentials:
firstly, the only bigons in the diagram contribute arrows labelled by the elementary algebra
elements q3 and p4 . Secondly, all other contributions come from squares, which necessarily have
boundary on both -curves.)
p4 p3
x2n d x2n c1 b0 y1 x2n c2
q14
q143
q14
a2n1 y2 a2n y1
p23
p23
x2n2 c3 x2n1 c2
(a)
p3 p3
b0 y2j x2n c2j+1 b0 y2j+1 x2n c2j+2
q14
q143
q14
a2n1 y2j+2 a2n y2j+1
p23
p23
x2n2 c2j+3 x2n1 c2j+2
(b) 1 j < m
p3 p3
b0 y2j1 x2n c2j b0 y2j x2n c2j+1
q14
q143
a2n1 y2j+1 a2n y2j

p23
x2n2 c2j+2
(c) 1 j m
p3
b0 y2m x2n c2m+1
q14
a2n y2m+1
p2
x2n1 b
(d)
at generators in minimal t1 -Alexander grading after cancellation. The dashed arrows
appear iff n > 1.
e b b
o
b b
e
b b
o b b
e b b
o b b
e
b b
oo b oe b b eo oo b b ee oe b b eo oo b b ee oe b b eo oo b b ee b
eo o
e oo o
b b b b b b b b b b b eo
oo b
e o
oo
eb eo ob
b b b b b b b b b b eo
oo b b b b b b b b b b b
eo
b
ob b b b b b b
e oo
b b b b b b b b
b b b b b b b b b b b eo
oo b b b b b b b
e o
oo
b b b b b
eb eo ob
b b b b b b b b b b eo
oo b b b b b b b b b b b
eb eo
b ob b b b b b b b b b b b b b b
e oo
b oe b b eo oo b b ee oe b b eo oo b b ee oe b b eo oo b b ee b eo
e o e o e o o
Figure 51. The last step of the calculation of the peculiar modules of (2n, (2m + 1))-
pretzel tangles. Some of the vertices are labelled according to the parity of the indices
of the generators they correspond to, where e stands for even and o for odd.
Next, we consider the effect of cancelling generators, first those corresponding to undoing
the handleslides and then any remaining identity arrows. Obviously, the pictures in figure 48
do not change. In figure 49a, cancellation only contributes an arrow p23 : x1 c2m+1 a1 y2m+1 .
The only possible arrow labelled by a power of p leaving d0 y2m+1 can go to a1 y2m+1 . Because of
the 2 -relation in the peculiar module, this arrow has to be there. Similarly, we can argue for
figure 49b. Cancellation contributes an arrow p23 : x1 c2j+1 a1 y2j+1 . It might also contribute
another arrow, p3 : x1 c2j+1 d0 y2j+2 , stemming from the arrow q4 : y 1 c2j+2 dy2j+3 ; however,
the 2 -relation at x1 c2j+1 in the peculiar module tells us that this does not happen. Again, there
has to be a contribution p234 : d0 y2j+1 a1 y2j+1 . In figure 50c, there is no arrow x1 c2j a1 y2j ,
h=q
p2
q3
b b b b b b b b
= =
p3
q2
b b b b b b b b
h=p
(a) (b)
q
p2
b b
h=p
b b b b b b
q
= p2 = q2
p h=q
b b b b b b q2 b b
(c) (d)
Figure 52. Basic homotopies for the final step of the proof of theorem 5.9
but again, the cancellation must contribute an arrow p234 : d0 y2j a1 y2j . Finally, in figure 49d,
we cancel two arrows, namely x1 d d0 y1 and y 1 c2 c2 y1 . This only contributes one arrow,
namely p23 : x1 c1 a1 y1 .
Similarly, we argue for those generators in minimal Alexander grading corresponding to
the colour t1 . Alternatively, we may apply the symmetry of the Heegaard diagram. The
corresponding subcomplexes after cancellation are shown in figure 50.
We have now obtained a reduced complex. To this, we add some arrows labelled by basic
algebra elements that lie in the kernel of A A /(p1 = 0 = q2 ) to turn the complex into the
peculiar module shown in figure 51. To see that this is indeed the result of putting all pieces
together, we suggest starting at the ends of figure 49d and 50d which sit at the bottom left and
top right corners of figure 51, respectively, and then connecting these subcomplexes. To see
how the diagonal strings of red and green generators connect the generators on the top left to
the generators on the bottom right of figure 51, we have indicated the parity of the generator
indices for some generators, see the caption of the figure.
We can now do some basic homotopies shown in figure 52 to obtain the desired result.
References
[Abo06] M. Abouzaid, On the Fukaya Categories of Higher Genus Surfaces, arXiv: 0606598
[AAEKO] M. Abouzaid, D. Auroux, A. I. Efimov, L. Katzarkov, D. Orlov, Homological mirror symmetry for
punctured spheres, J. Amer. Math. Soc. 26 (2013), 1051-1083 (arXiv: 1103.4322) A2311
[Ale28] J. W. Alexander, Topological invariants of knots and links, Trans. Amer. Math. Soc. 30 (1928),
275306
[Bar04] D. Bar-Natan, Khovanovs homology for tangles and cobordisms, Geom. Topol. 9 (2005), 14431499
(arXiv: 0410495v2)
[BB14] D. Bar-Natan, H. Burgos-Soto, Khovanov homology for alternating tangles, arXiv: 1305.1695
[BL11] J. A. Baldwin, A. S. Levine, A combinatorial spanning tree model for knot Floer homology, Adv.
Math. 231 (2012), 18861939 (arXiv: 1105.5199v2)
[EPV15] A. P. Ellis, I Petkova, V. Vrtesi: Quantum gl(1|1) and tangle Floer homology, arXiv: 1510.03483v1
[HKK14] F. Haiden, L. Katzarkov, M. Kontsevich, Flat surfaces and stability structures, arXiv: 1409.8611v2
[Han13] J. Hanselman, Bordered Heegaard Floer homology and graph manifolds, arXiv: 1310.6696
[HRW16] J. Hanselman, J. A. Rasmussen, L. Watson, Bordered Floer homology for manifolds with torus
boundary via immersed curves, arXiv: 1604.03466v2
[HW15] J. Hanselman, L. Watson, A calculus for bordered Floer homology, arxiv: 1508.05445v1
[HHK13] M. Hedden, C. Herald, P. Kirk, The pillowcase and perturbations of traceless representations of knot
groups, Geom. Topol. 18 (2014), 211287 (arXiv: 1301.0164v1)
[HHK15] , The pillowcase and traceless representations of knot groups II: a Lagrangian-Floer theory in
the pillowcase, arXiv: 1501.00028v1
[KR04] M. Khovanov, L. Rozansky, Matrix factorizations and link homology, arXiv: 0401268v2
[Lam16] P. Lambert-Cole, Twisting, mutation and knot Floer homology, arXiv: 1608.02011
[Lam17] , On Conway mutation and link homology, arXiv: 1701.00880
[Lip05] R. Lipshitz, A cylindrical reformulation of Heegaard Floer homology, Geometry & Topology 10
(2006) 9551096, (arXiv: 0502404v2)
[LOT08] R. Lipshitz, P. Ozsvth, D. Thurston, Bordered Heegaard Floer homology: Invariance and pairing,
arXiv: 0810.0687v5
[LOT10a] , Bimodules in bordered Heegaard Floer homology, Geom. Topol. 19 (2015) 525724,
(arXiv: 1003.0598v4)
[Mnn16] A. Manion, On the decategorification of Ozsvth and Szabs bordered theory for knot Floer homology,
arXiv: 1611.08001
[Mnl06] C. Manolescu, An unoriented skein exact triangle for knot Floer homology, Math. Res. Lett. 14
(2007), 839852 (arXiv: 0609531v3)
[OS01] P. Ozsvth, Z. Szab, Holomorphic discs and topological invariants of closed 3-manifolds, Ann. Math.
159 (2004), 10271158 (arXiv: 0101206v4)
[OS02] , Heegaard Floer homology and alternating knots, Geom. Topol. 7 (2003), 225254
(arXiv: 0209149v3)
[OS03a] , Holomorphic disks and knot invariants, Adv. Math. 186 (2004), no. 1, 58116
(arXiv: 0209056v4)
[OS03b] , Knot Floer homology, genus bounds, and mutation, arXiv: 0303225v2
[OS05] , Holomorphic disks and link invariants, Algebr. Geom. Topol. 8 (2008), 615692
(arXiv: 0512286v2)
[OS06a] , Link Floer homology and the Thurston norm, arXiv: 0601618v3
[OS07] , A cube of resolutions for knot Floer homology, J. Topol. 2 (2009), no. 4, 865910
(arXiv: 0705.3852v1)
[OS16] , Kauffman states, bordered algebras, and a bigraded knot invariant, arXiv: 1603.06559v1
[OS17] , Bordered knot algebras with matchings, arXiv: 1707.00597v1
[OSS07] P. Ozsvth, A. I. Stipsicz, Z. Szab, Floer homology and singular knots, J. Topol. 2 (2009), 380404
(arXiv: 0705.2661v3)
[PV14] I. Petkova, V. Vrtesi, Combinatorial tangle Floer homology, arXiv: 1410.2161v2

[Ras03] J. A. Rasmussen, Floer homology and knot complements, PhD thesis (2003), Harvard
(arXiv: 0306378v1)
[Rie14] E. Riehl, Categorical homotopy theory, Cambridge University Press (2014)
[SW06] S. Sarkar, J. Wang, An algorithm for computing some Heegaard Floer homologies, Ann. of Math. 171
(2010), no. 2, 12131236 (arXiv: 0607777v4)
[Sei06] P. Seidel, Fukaya Categories and Picard-Lefschetz Theory, Zrich lectures in advanced mathematics,
European Math. Soc., 2008
[WZ] L. Watson, C. Zibrowius, Orientable genera of 4-ended tangles, in preparation
[Zar09] R. Zarev, Bordered Floer homology for sutured manifolds, arXiv: 0908.1106v2
[Zar10] , Joining and gluing sutured Floer homology, arXiv: 1010.3496v1
[Zar11] , Bordered Sutured Floer Homology, PhD thesis (2011), Columbia University
(doi: 10.7916/D83R10V4)
[Zem16] I. Zemke, Link cobordisms and functoriality in link Floer homology, arXiv: 1610.05207v1
[Zib15] C. B. Zibrowius, On a polynomial Alexander invariant for tangles and its categorification, Essay for
the Smith-Knight & Rayleigh-Knight Prize Competition 2015, Cambridge (arXiv: 1601.04915v1)
[Zib16] , Heegaard diagrams for tangles, arXiv: 1601.04915
[Zib17] , On a Heegaard Floer theory for tangles, University of Cambridge, 2017 (arXiv: 1610.07494)

Peculiar Modules For 4-Ended Tangles

Hochgeladen von

Dokumentinformationen

Originaltitel

Copyright

Verfügbare Formate

Dieses Dokument teilen

Dokument teilen oder einbetten

Freigabeoptionen

Stufen Sie dieses Dokument als nützlich ein?

Sind diese Inhalte unangemessen?

Copyright:

Verfügbare Formate

Peculiar Modules For 4-Ended Tangles

Hochgeladen von

Copyright:

Verfügbare Formate

PECULIAR MODULES FOR 4-ENDED TANGLES

CLAUDIUS BODO ZIBROWIUS

quotient of A n , obtained by setting certain variables equal

Theorem 0.1 (2.17). Given a 4-ended tangle T in a homology Figure 1. A diagram of a 4-

exists a bounded, strictly unital type AA structure P such that

0.2. Classification of peculiar modules. In section 4, we classify peculiar modules in terms

Definition 0.5 (4.34). In particular, with any 4-ended

Question 0.6. The number of curves in LT is obviously a

The proof of theorem 0.4 is based on an algorithm due to

Corollary 0.7 (4.44). With the same notation as in theorem 0.2,

Question 0.10. Are all tangles loop-type?

CFT (T0 ) V (a) Tn (b) Tn (c) T0

Definition 0.13 (Conway mutation). Given a link L in S 3 ,

In this paper, we prove a slightly stronger version of this

graded link Floer homology, after identifying the Alexander

1. Preliminaries: Algebraic structures from dg categories

1.1. The general construction.

Compositions and differentials in Mat(C) are induced by those in C.

D : Mori ((O, dO ), (O0 , dO0 )) Mori1 ((O, dO ), (O0 , dO0 ))

a more explicit, but less concise definition, see definition 2.13.

just as for type D structures. However, a morphism in a hom-object Mor(i1 , i2 ) of CA

by a sequence of vector space homomorphisms

(fi : i2 .Ai .i1 F2 )i0

where composition is defined by

The differential is given by

unital type A structures by restricting to those objects (O, dO ) such that

f (, ai a1 ) = 0 if i > 0 and aj = 1 for some j = 1, . . . , i.

(fi : i1 .Ai .i2 F2 )i1 ,

corresponding to the J -algebra B. The inclusion J , I and the J -algebra homomorphism

and hence also a functor

both of which respect the differentials on both sides.

as follows: an object i is sent to i if i J and to the zero-object otherwise. A basic morphism

Remark 1.17. In the previous definition, non-injective homomorphisms J I also induce

Observation 1.18. Let J , I, A, B and : B A be as in definition 1.16. Choose a basis of

Theorem 1.22 (Pairing adjunction). Let J , I, A, B and : B A be as in definition 1.16.

1.3. Cancellation and Cleaning-up. We now state and

ject of Cx (Mat(C)) for some differential graded category C (Y1 , 1 ) (Y2 , 2 )

( + bgc)2 + ( + bgc) = + bgc + bgc + bgcbgc + () + (b)gc + b(g)c + bg(c)

For the last step, we observe that

gcbg = gD(f )g = D(gf g) + D(g)f g + gf D(g) = D(g).

Next, we consider the two chain maps

F : (Z, + bgc) (X, ) and G : (X, ) (Z, + bgc)

Figure 8. Maps for the proof of lemma 1.23

gc( + bgc) + 1 gc + (gc) + a = g(c + (c)) + (1 g + (g))c + a + gcbgc

for the first identity and similarly

( + bgc)bg + bg2 + (bg) + d = (b + (b))g + b(g2 + (g)) + d + bgcbg

where H is the homotopy given by the dashed line in figure 8c.

h2 , hD(h) and D(h)h

vanish, (O, dO ) is chain homotopic to (O, dO + D(h)).

Proof. We can easily check that (O, dO + D(h)) is an object in Cx (Mat(C)):

2. The invariant CFT

2.1. Heegaard diagrams for tangles. Let us start by re-

Definition 2.1. A 4-ended tangle T in a homology 3- a c

Definition 2.2. A Heegaard diagram HT for a tangle T with n closed components in a

(a) An original tangle end (b) A new tangle end

Remark 2.6. The attribute peculiar should be considered as a homophone of p-q-lier, a

Definition 2.8. A matching P is a partition {{i1 , o1 }, {i2 , o2 }} of {1, 2, 3, 4} into pairs. An

For k = 1, . . . , n, we define At0k by

At0k (y) At0k (x) := At0k () := wk () zk ().

We sometimes denote the Alexander grading on generators by a superscript list of integers

2.2. Peculiar algebras.

We define a 12 Z-grading on pre

where i = 1, 2, 3, 4 and j = 1, . . . , n, and then extending linearly to all of Apre