Journal of Topology 1 (2008) 285–305

c 2008 London Mathematical Society

doi:10.1112/jtopol/jtn002

Cobordism of knots on surfaces

Vladimir Turaev

Abstract

We introduce a relation of cobordism for knots in thickened surfaces and study cobordism
invariants of such knots.

Introduction
This paper is concerned with knots on oriented surfaces, that is, with knots in 3-manifolds of
type (an oriented surface) × R. The class of such knots appears to be a natural intermediate
between the class of classical knots in R3 = R2 × R and the class of knots in arbitrary 3-
manifolds. The diagrammatical methods used for knots in R3 extend to knots on surfaces. On
the other hand, one encounters new phenomena, which are absent in the classical case. The
study of knots on surfaces has a long history; see for instance [3, 5, 7, 13]. This area has been
especially active in the context of the Kauffman theory of virtual knots; see [9, 10, 11]. For
other approaches, see [1, 4, 6].
The principal aim of this paper is to introduce and to study an equivalence relation of
cobordism for knots on surfaces. Briefly speaking, a knot K1 on a surface Σ1 is cobordant to
a knot K2 on a surface Σ2 if there is an oriented 3-manifold M with Σ1
(−Σ2 ) ⊂ ∂M such
that the 1-manifold K1
(−K2 ) ⊂ ∂M × R bounds an annulus in M × R. We introduce several
nontrivial cobordism invariants of knots. This includes geometric invariants (slice genera) and
algebraic invariants: polynomials u+ , u− , algebraic genera, and a so-called ‘graded matrix’. The
algebraic invariants of a knot on a surface Σ are derived from an arbitrary diagram D of this
knot on Σ. The key idea is that every crossing of D splits the underlying loop of D into two
‘half-loops’, and we can select one of them in a canonical way. This gives a finite family of
loops on Σ. The homological intersection numbers of these loops with each other and with the
underlying loop of D form a skew-symmetric square matrix over Z. It will be the main source
of our invariants.
It is natural to compare the knot cobordism defined above with the standard concordance
for knots in R3 ; see [12] for a recent survey. It is obvious that concordant classical knots are
cobordant in our sense. Besides this fact, essentially nothing seems to be known about our
relation of cobordism for classical knots. To the best of the author’s knowledge, it is possible
that all classical knots are cobordant to each other. It is also conceivable that two classical
knots are cobordant if and only if they are concordant in the standard sense. All knot invariants
defined in this paper are trivial for the classical knots.
One way to obtain invariants of a knot K ⊂ Σ × R is to project K to Σ and to study
homotopy invariants of the resulting loop K of K. Loops on surfaces and their cobordisms
were studied in [14]. The present paper refines the methods of [14] and lifts them to knots.
Preliminary knowledge of [14] can be useful to the reader but is not required.
The content of the paper is as follows. As a warm-up, we introduce in Section 1 the knot
polynomials u+ and u− . In Section 2, we discuss cobordisms of knots and show that u+ and

Received 14 May 2007; published online 2 April 2008.

2000 Mathematics Subject Classification 57M27.
286 VLADIMIR TURAEV

u− are cobordism invariants. In Section 3, we study graded matrices and introduce graded
matrices of knots. In Section 4, we analyze an equivalence relation of cobordism on the class of
graded matrices. The relationship with knot cobordisms is discussed in Section 5. Miscellaneous
remarks and open questions are collected in Section 6.
Throughout the paper we work in the smooth category, though all the results can be
reformulated in PL and topological categories. All surfaces and 3-manifolds in this paper are
oriented.

1. Polynomials u+ and u−

1.1. Knots and knot diagrams

By a knot, we mean a pair consisting of an (oriented) surface Σ and an oriented embedded
circle K ⊂ Σ×R disjoint from ∂Σ×R. We say that K is a knot on Σ. This knot will be denoted
K ⊂ Σ × R or (Σ, K), or simply K. Note that Σ may be noncompact or nonconnected.
Two knots K1 , K2 ⊂ Σ × R on a surface Σ are isotopic if K1 can be deformed into K2 in
the class of knots in Σ × R (keeping the orientation of the knot). Two knots K1 ⊂ Σ1 × R and
K2 ⊂ Σ2 ×R are diffeomorphic if there is an orientation-preserving diffeomorphism f : Σ1 → Σ2 ,
such that the knot (f × idR )(K1 ) ⊂ Σ2 × R is isotopic to K2 . Isotopic knots are necessarily
diffeomorphic.
A knot diagram on a surface Σ is an oriented closed curve on Σ − ∂Σ with only double
transversal crossings such that at each double point, one of the branches of the curve passing
through this point is distinguished. The distinguished branch is said to be overgoing, while the
second branch passing through the same double point is undergoing. A knot diagram D on
Σ = Σ×{0} determines a knot in Σ×R by pushing the overgoing branches of D into Σ×(0, ∞).
It is clear that every knot in Σ × R is isotopic to a knot presented by a knot diagram on Σ.
Two knot diagrams on Σ present isotopic knots if and only if these diagrams are related by a
finite sequence of standard Reidemeister moves (briefly, R-moves) applied in small disks in Σ
while keeping the rest of the diagram.

1.2. Halves of knot diagrams

Let D be a knot diagram on a surface Σ. The set of double points of D will be denoted


(D). With each double point x ∈
(D), we associate a loop Dx on Σ as follows. Let A and
B be the branches of D passing through x, where the notation is chosen so that the pair (a
positive tangent vector of A at x, a positive tangent vector of B at x) is a positive basis in
the tangent space of Σ at x. Denote by Dx the loop on Σ starting at x, going along A in the
positive direction, and then going along the underlying loop of D until the first return to x.
We call Dx the distinguished half of D at x. Note that the definition of Dx does not use the
over/under data at x. To keep track of this data, we define the sign sign(x) of x to be +1 if A
is overgoing and −1 if B is overgoing. The set

(D) splits as a union of two disjoint subsets

+ (D) = {x ∈

(D) | sign(x) = +1} and

−(D) = {x ∈

(D) | sign(x) = −1}.

1.3. Definition of u±
Let K ⊂ Σ×R be a knot presented by a diagram D on Σ. Let [D] ∈ H1 (Σ) = H1 (Σ; Z) be the
homology class of the (oriented) closed curve underlying D. For a double point x ∈
(D), we
can consider the homology class [Dx ] ∈ H1 (Σ) of the loop Dx and the homological intersection
number [Dx ] · [D] ∈ Z. For brevity, we shall write Dx · D for [Dx ] · [D].
 COBORDISM OF KNOTS ON SURFACES 287

We define two one-variable polynomials u+ (K), u− (K) ∈ Z[t]. For ε = ±1, set


uε (K) = sign(x) t| D x ·D | ,
x∈ 

(D ), D x ·D = 0, sign(D x ·D ) = ε sign(x)

where sign(Dx · D) ∈ {+1, −1} and | Dx · D | are respectively the sign and the absolute value
of the nonzero integer Dx · D. In other words,



u+ (K) = tD x ·D − t−(D x ·D ) ,
x∈ 

+ (D ), D x ·D > 0 x∈ 

−(D ), D x ·D < 0



u− (K) = t−(D x ·D ) − tD x ·D .
x∈ 

+ (D ), D x ·D < 0 x∈ 

−(D ), D x ·D > 0

Both u+ (K) and u− (K) are isotopy invariants of K. To see this, it is enough to check that they
are preserved under the R-moves on D. The invariance under the first R-move creating a new
crossing x0 follows from the fact that [Dx 0 ] = 0 or [Dx 0 ] = [D]; in both cases Dx 0 · D = 0. The
invariance under the second R-move creating two new crossings x1 , x2 follows from the fact
that sign(x1 ) = − sign(x2 ) and [Dx 1 ] + [Dx 2 ] = [D], so that Dx 1 · D = −(Dx 2 · D). If Dx 1 · D =
Dx 2 · D = 0, then x1 , x2 do not contribute to u+ (K), u− (K). If Dx 1 · D = −(Dx 2 · D) = 0, then
the contributions of x1 , x2 cancel each other. The invariance under the third R-move follows
from the fact that the set of pairs {([Dx ], sign(x))}x is the same before and after the move.
Thus, the polynomials u+ , u− are isotopy invariants of knots. It is also clear that they are
diffeomorphism invariants of knots.

1.4. Properties of u±
We point out several properties of the polynomials u+ (K), u− (K) of a knot K on a surface Σ.
It is clear that both polynomials have zero free terms. The difference u(K) = u+ (K) − u− (K)
is computed from a diagram D of K by


u(K) = sign(Dx · D) t|D x ·D | .
(D ), D x ·D = 0
x∈

It follows from this formula that u(K) does not depend on the over/under data in D. Therefore,
the polynomial u = u(K) ∈ Z[t] is a homotopy invariant of K. This polynomial is the invariant
of the underlying loop of D on Σ introduced in [14, Section 3]. One property of u established
there computes the value of its derivative u = du/dt at t = 1, namely u (1) = 0. We conclude
that
u+ (K)(1) = u− (K)(1). (1.4.1)
The reader can prove this equality directly as an exercise.
We denote by K the same knot K ⊂ Σ × R with opposite orientation on Σ (keeping that
on K). Similarly, −K denotes the same knot K ⊂ Σ × R with opposite orientation on K
(keeping that on Σ). It is easy to check that u± (K) = −u∓ (K) and u± (−K) = u∓ (K).
If two knots K1 , K2 ⊂ Σ × R are presented by disjoint diagrams D1 , D2 on Σ and a knot
K1 #K2 ⊂ Σ × R is presented by the connected sum of D1 , D2 along an embedded arc on Σ
relating a point on D1 to a point on D2 and disjoint from D1 , D2 otherwise, then u± (K1 #K2 ) =
u± (K1 ) + u± (K2 ).

1.5. Examples
1. For any knot K ⊂ R3 = R2 × R, we have u+ (K) = u− (K) = 0. Indeed, the homological
intersection number of any two loops on R2 is zero.
2. Let D be the knot diagram on the torus S 1 × S 1 obtained from the standard diagram of
a trefoil on S 2 by replacing the overgoing branch at one crossing with a branch going along a
288 VLADIMIR TURAEV

1-handle attached to S 2 . The diagram D has only two double points x, y. A direct computation
shows that for appropriate orientations on D and on the torus, sign(x) = sign(y) = Dx · D =
−Dy ·D = 1. Then for the knot K ⊂ S 1 ×S 1 ×R presented by D, we have u+ (K) = u− (K) = t.
Note that u+ (K) = u− (K) for an arbitrary knot K ⊂ S 1 × S 1 × R; this follows from the fact
that any loop on the torus may be deformed inside an annulus and the intersection number of
any two loops on an annulus is zero.

1.6. Realization
We say that a pair of polynomials p+ , p− ∈ Z[t] is realized by a knot K if u+ (K) = p+ and
u− (K) = p− . If a pair (p+ , p− ) is realized by a knot, then it can be realized by a knot on a
closed oriented surface. This follows from the facts that the computation of u+ , u− proceeds
entirely inside a neighborhood of a knot diagram in the ambient surface, and that any knot
diagram on a surface has a neighborhood that embeds into a closed surface.
The following theorem solves the realization problem for u+ , u− .

Theorem 1.6.1. A pair of polynomials p+ , p− ∈ Z[t] can be realized by a knot if and only
if p+ (0) = p− (0) = 0 and p+ (1) = p− (1).

Proof. We need only to prove the sufficiency. Set u = p+ − p− . The conditions on p+ , p−

imply that u(0) = 0 and u (1) = 0. It follows from [14, Theorem 3.4.1] that there is a knot
diagram D on an oriented surface Σ presenting a knot K0 ⊂ Σ × R with u(K0 ) = u+ (K0 ) −
u− (K0 ) = u. The knot K0 and the knot K0 #(S 1 × x) on Σ#(S 1 × S 1 ), where x ∈ S 1 , have
the same u± . Therefore, replacing if necessary K0 by K0 #(S 1 × x), we can assume that there
is a loop α on Σ such that α · D = 1. We claim that for any knot K  ⊂ Σ × R homotopic to K0
and any nonzero integer N , there is a diagram D of K  on Σ and a double point x ∈
(D )
such that
sign(x) = sign(N ) and Dx · D = N.
This claim implies the theorem. Indeed, exchanging the over/undercrossings at such an x ∈
 (D ), we obtain a new knot diagram on Σ presenting a knot K  ⊂ Σ × R homotopic to


K0 . Clearly, u+ (K  ) = u+ (K  ) − sign(N ) t|N | . Starting from K0 and applying this procedure
recursively, we can obtain a knot K ⊂ Σ×R homotopic to K0 and such that u+ (K) = p+ . Then
u− (K) = u+ (K) − u(K) = p+ − u(K0 ) = p+ − u = p− .
To prove the claim above, we take a small arc on D and push it with a finger along a generic
loop on Σ homologous to αN and passing everywhere over (or everywhere under) D. When the
finger comes back to D near the original arc, we obtain two new self-crossings; one of them
will be the required x. We leave the details to the reader.

1.7. Coverings of knots and higher polynomials

The polynomials u± give rise to a family of polynomial invariants of knots numerated by
finite sequences of positive integers. We first define the mth covering of a knot (Σ, K) for any
integer m
1. Projecting the homology class [K] ∈ H1 (Σ × R; Z) = H1 (Σ; Z) to H1 (Σ; Z/mZ)
and applying the Poincaré duality H1 (Σ; Z/mZ) ∼ = H 1 (Σ; Z/mZ), we obtain a cohomology
class in H (Σ; Z/mZ). This class determines an m-fold covering Σ(m ) → Σ. It is clear that K
1

lifts to a knot K (m ) ⊂ Σ(m ) × R. Different lifts of K to Σ(m ) × R give diffeomorphic knots, so

that K (m ) is determined by m and the knot (Σ, K) uniquely up to diffeomorphism. Iterating
this construction, we define for any finite sequence of integers m1 , . . . , mk
1, a knot
 (m 2 )
K (m 1 ,...,m k ) = (. . . K (m 1 ) . . .)(m k ) .
 COBORDISM OF KNOTS ON SURFACES 289

Set
um
±
1 ,...,m k
(K) = u± (K (m 1 ,...,m k ) ) ∈ Z[t].

By the results above, this polynomial is a diffeomorphism invariant of (Σ, K). Note that
u1+ (K) = u+ (K) and u1− (K) = u− (K).

2. Cobordism and sliceness of knots

2.1. Cobordism of knots

Two knots (Σ1 , K1 ) and (Σ2 , K2 ) are cobordant if there is an oriented 3-manifold M and
an embedding Σ1
(−Σ2 ) → ∂M such that the image of K1
(−K2 ) under the induced
embedding (Σ1 × R)
(Σ2 × R) → ∂M × R bounds an embedded oriented annulus in M × R.
It is understood that the orientation of M and that of the annulus induce, respectively, the
given orientations on Σ1 and K1 , and orientations opposite to the given ones on Σ2 and K2 .
A standard argument shows that cobordism of knots is an equivalence relation.
It follows from the definitions that diffeomorphic knots are cobordant. If Σ is a subsurface of
a surface Σ, then any knot in Σ ×R is cobordant to the same knot viewed in Σ×R. (To see this,
take M = Σ × [0, 1] and consider the embeddings Σ = Σ × 0 → ∂M and Σ = Σ × 1 → ∂M .)
In the language of [3], these facts may be reformulated by saying that stably equivalent knots
are cobordant.

(m ) (m )
Lemma 2.1.1. If two knots K1 and K2 are cobordant, then the knots K1 and K2 are
cobordant for any m
1.

Proof. Let M be an oriented 3-manifold as in the definition of cobordism, so that the

1-manifold K1
(−K2 ) ⊂ ∂M × R bounds an embedded oriented annulus F ⊂ M × R. Let
p : M × R → M be the projection. Replacing, if necessary, M by a closed neighborhood of
p(F ), we can assume that M is compact. By the Poincaré duality, there is a cohomology class
a ∈ H 1 (M ; Z) dual to the homology class of p(F ) in H2 (M, ∂M ; Z). Let M (m ) → M be the m-
fold covering determined by a (mod m) ∈ H 1 (M ; Z/mZ). The annulus F lifts to an embedded
(m ) (m )
annulus in M (m ) × R; the latter yields a cobordism between K1 and K2 .

2.2. Slice knots

A knot K ⊂ Σ × R cobordant to a trivial knot in R3 = R2 × R is said to be slice. It is easy
to check that a knot K ⊂ Σ × R is slice if there is an oriented 3-manifold M and an embedding
Σ ⊂ ∂M, such that the image of K under the induced embedding Σ × R ⊂ ∂M × R bounds an
embedded disk in M × R.
One should be careful in applying this terminology to knots in R3 . A knot in R3 that is
slice in the usual sense of the word is slice in our sense, but the converse is not known and is
possibly untrue. In what follows, we use the term ‘slice’ exclusively in the sense defined above.
It is clear that knots cobordant to a slice knot are slice. By Lemma 2.1.1, the coverings of
slice knots are slice. Note also, that the usual notion of ribbon knots in R3 directly extends to
knots on surfaces and yields a wide class of slice knots.

2.3. Polynomials of cobordant knots

We state the main result of this section.
290 VLADIMIR TURAEV

Theorem 2.3.1. The polynomials u+ and u− are cobordism invariants of knots.

This theorem follows from the identity u± (−K) = −u± (K) and the following more general
lemma.

Lemma 2.3.2. Let M be an oriented 3-manifold and let K1 , . .  . , Kr be knots in ∂M × R

r
with pairwise disjoint projections to ∂M . If the oriented 1-manifold i=1 Ki ⊂ ∂M × R bounds
in M × R an embedded compact oriented surface of genus 0, then


r
u± (Ki ) = 0. (2.3.1)
i=1

Proof. We need some terminology concerning maps from surfaces to M . We say that a map
ω from a surface F to M is generic if
(i) card(ω −1 (x))  3 for all x ∈ M ;
(ii) ω −1 (∂M ) = ∂F and the restriction of ω to ∂F is an immersion into ∂M with only
double transversal crossings; any point of ω(∂F ) has a neighborhood V ⊂ M such that the
pair (V, V ∩ ω(F )) is diffeomorphic to (R2 , R × 0) × R+ or to (R2 , R × 0 ∪ 0 × R) × R+ , where
R+ = {r ∈ R | r
0};
(iii) there is a finite set B ⊂ F ◦ = F − ∂F, such that the restriction of ω to F ◦ − B is an
immersion into M ◦ = M − ∂M and any distinct points a, a ∈ F ◦ − B with ω(a) = ω(a ) have
disjoint neighborhoods in F with transversal images under ω;
(iv) for any b ∈ B, there is a closed 3-ball D ⊂ M ◦ , such that ω(F ) ∩ ∂D is a figure-8 curve
(a closed curve with one transversal self-crossing) and ω(F ) ∩ D is the cone over this curve
with cone point ω(b) ∈ D − ∂D.

The points of the set B = B(ω) are called the branch points of ω. Conditions (i) and (iii)
imply that any point of ω(F ◦ −B) has a neighborhood V ⊂ M ◦ such that the pair (V, V ∩ω(F ))
is diffeomorphic to the pair (R3 , a union of i coordinate planes) for i = 1, 2, 3. For more on
this, see [8, Chapter 7]. r
Now let F ⊂ M × R be an embedded compact surface of genus 0 bounded by i=1 Ki . We
can slightly deform F in the class of embedded surfaces with boundary on ∂M × R to ensure
that F ∩ (∂M × R) = ∂F and the restriction to F of the projection M × R → M is a generic
map. We can still assume that the components of ∂F in ∂M × R project to disjoint subsets of
∂M . These components are isotopic to the knots K1 , . . . , Kr and have the same polynomials
u± . Thus, to prove the lemma, we can use the same symbols K1 , . . . , Kr for the components
of ∂F . The generic map from F to M obtained as the restriction of the projection M × R → M
will be denoted ω. Set B = B(ω) ⊂ F ◦ and

C = B ∪ {a ∈ F | card (ω −1 (ω(a)))
2} ⊂ F.

It is clear from conditions (i)–(iv) above that C consists of a finite number of immersed circles
and intervals in F meeting ∂F exactly at the endpoints of the intervals. These endpoints are
the preimages of the double points of ω|∂ F : ∂F → ∂M . The set of these endpoints coincides
with C ∩ ∂F and is denoted ∂C. The immersed circles and intervals forming C have only
double transversal crossings and self-crossings, all lying in F ◦ . The set of these crossings and
self-crossings is denoted S; it consists precisely of the preimages of the triple points of ω. The
subsets B, S, ∂C of C are finite and pairwise disjoint.
Let C  be an abstract 1-dimensional manifold parametrizing C. The projection p : C  → C is
 −1
2-to-1 over S and 1-to-1 over C − S. We shall identify C − p (S) with C − S via p.
 COBORDISM OF KNOTS ON SURFACES 291

For any point a ∈ C − (S ∪ B), there is exactly one other point a ∈ C − (S ∪ B) such that
ω(a) = ω(a ). The correspondence a
→ a on
 − p−1 (S ∪ B)
C − (S ∪ B) = C
extends by continuity to an involution on C,  denoted τ . The set of fixed points of τ is B. It is
clear that ω p τ = ω p : C  → M . For a ∈ ∂C = ∂ C,  the point τ (a) is the unique other point of
∂C such that ω(a) = ω(τ (a)). Applying ω, we can identify the quotient set ∂C/τ with the set
(ω|∂ F ) ⊂ ∂M of the self-crossings of ω|∂ F .


We define an involution µ : ∂C → ∂C. For a ∈ ∂C = ∂ C,  let Ia ⊂ C  denote the interval
adjacent to a and let µ(a) ∈ ∂C be its endpoint distinct from a. Clearly, µ2 = id and µ is
fixed-point-free. Note that µ commutes with τ |∂ C . Indeed, if τ (Ia ) = Ia , then τ exchanges the
endpoints of Ia so that τ = µ on ∂Ia ⊂ ∂C. (In this case, τ must have a unique fixed point on
Ia , so that Ia contains a unique branch point of ω.) If τ (Ia ) = Ia , then τ (Ia ) has the endpoints
τ (a), τ (µ(a)) so that µ(τ (a)) = τ (µ(a)). Since µ and τ |∂ C commute, µ induces an involution on
∂C/τ =
(ω|∂ F ). The latter involution is denoted ν. It has a simple geometric interpretation.
Start in a point x ∈
 (ω|∂ F ) ⊂ ∂M and move along the smooth immersed interval in M
adjacent to x and formed by double points of ω(F ). There are two possibilities for the second
endpoint y of this interval: either y ∈
(ω|∂ F ) and then ν(x) = y or y ∈ B and then ν(x) = x.
By the assumptions, ω maps the components K1 , . . . , Kr of ∂F to disjoint subsets of ∂M.
Therefore,


 (ω|∂ F ) =
ri=1

 (ω(Ki )), (2.3.2)
where
(ω(Ki )) is the set of all self-crossings of the loop ω(Ki ). Pick an arbitrary point
x∈
(ω|∂ F ). We associate with x a certain homology class as follows. By equation (2.3.2), x
is a self-crossing of ω(Ki ) for a unique i ∈ {1, . . . , r}. Endow ω(Ki ) with over/undercrossing
information, so that the resulting knot diagram on ∂M presents Ki ⊂ ∂M × R. Let ω −1 (x) =
{a, b}, where a, b ∈ Ki and we choose the notation so that the pair (the positive tangent vector
of Ki at a, the positive tangent vector of Ki at b) is transformed by (the differential of) ω into
a positive basis in the tangent space of ∂M at x. Let γx ⊂ Ki be the positively oriented arc
on Ki leading from a to b. The loop ωx = ω(γx ) is the ‘distinguished half’ of the diagram of
Ki at x. Let [ωx ] ∈ H1 (∂M ) be the homology class of ωx and let in : H1 (∂M ) → H1 (M ) be
the inclusion homomorphism. We now study the class [ωx ] in more detail.
Consider first the case where ν(x) = x. We claim that then
in([ωx ]) ∈ ω∗ (H1 (F )) ⊂ H1 (M ). (2.3.3)
Indeed, the equality ν(x) = x means that µ(a) ∈ {a, b}. Since µ is fixed point-free, µ(a) = b.
By the definition of the involution τ : C → C, we have τ (a) = b and τ (b) = a. Since τ preserves
 we necessarily have τ (Ia ) = Ia . The image
the set of endpoints {a, b} of the interval Ia ⊂ C,

of Ia under the projection p : C → C ⊂ F is an immersed interval on F connecting a and b.
The product of the path γx ⊂ Ki ⊂ ∂F with the immersed interval p(Ia ) oriented from b to a
is a loop in F , denoted by ρ. The loop ω(ρ) in M is a product of the loops ω(γx ) = ωx and
ωp|I a . The latter loop has the form δδ −1 , where δ is the path in M obtained by restricting
ω p to the arc in Ia leading from b to the unique branch point of ω on Ia . Since the loop δδ −1
is contractible in M , the loops ωx and ωρ are homologous in M . Since ρ is a loop on F , we
conclude that in([ωx ]) ∈ ω∗ (H1 (F )).
Suppose now that y = ν(x) = x. We shall establish two properties of the points x, y ∈

(ω|∂ F ). The first property is the inclusion
in([ωx ] + [ωy ]) ∈ ω∗ (H1 (F )) ⊂ H1 (M ), (2.3.4)
 (ω(Kj )) for some j ∈ {1, . . . , r} (possibly j = i).
which we now verify. As we know, y ∈


Since the involution ν on ∂C/τ = ∂ C/τ is induced by the involution µ : ∂C → ∂C and
292 VLADIMIR TURAEV

ω −1 (x) = {a, b}, we must have ω −1 (y) = {µ(a), µ(b)} with µ(a), µ(b) ∈ Kj . It is easy to
check that the pair (the positive tangent vector of Kj at µ(a), the positive tangent vector of
Kj at µ(b)) is transformed by ω into a negative basis in the tangent space of ∂M at y (this
change of orientation was first pointed out in [2]; cf. the argument in the next paragraph).
Therefore, the oriented arc γy ⊂ Kj begins at µ(b) and terminates at µ(a). Consider the loop
ρ = γx p(Ib ) γy (p(Ia ))−1 in F based at a. Here the intervals Ib , Ia are oriented from b to µ(b)
and from a to µ(a), respectively. Then ω(ρ) is the product of the loop ωx based at x, the
path ωp(Ib ) beginning in x and ending in y, the loop ωy based at y, and the path (ωp(Ia ))−1
beginning in y and ending in x. The paths ωp(Ib ) and (ωp(Ia ))−1 are mutually inverse since
Ib = τ (Ia ) and ωpτ = ωp. Hence, in([ωx ] + [ωy ]) = [ω(ρ)]. This implies equation (2.3.4).
To state the second property of the pair x, y = ν(x) = x, recall the signs of the self-crossings
of a knot diagram introduced in Subsection 1.2. By definition, sign(x) = +1 if the pair (the
overgoing branch at x, the undergoing branch at x) determines a positive basis in the tangent
space of ∂M at x and sign(x) = −1 otherwise. We claim that sign(x) = − sign(y). Indeed,
consider the smooth immersed interval I = ωp(Ia ) in M formed by double points of ω(F )
and connecting x to y. This interval is an intersection of two branches of ω(F ). Each of these
branches is the projection of an immersed ribbon on F with core that projects to I and with
bases that are short intervals on ∂F parametrizing certain branches of ω(∂F ) passing through
x and y. Since the cores of the two ribbons in question project to the same interval I in M , one
of these ribbons in F ⊂ M × R has to lie above the second one with respect to the projection
to R. This allows us to push the pair (the overgoing branch at x, the undergoing branch at
x) along I, so that each branch is pushed along the corresponding ribbon transversally to the
core. We obtain in the end a pair (the overgoing branch at y, the undergoing branch at y).
If the first pair is positive with respect to the orientation of ∂M induced from M , then the
second one is negative and vice versa. Hence, sign(x) = − sign(y).
We can now prove the lemma for u+ ; the claim concerning u− can be proved similarly, or
deduced by reversing the orientation in M . Set si = [ω(Ki )] ∈ H1 (∂M ) for i = 1, . . . , r and
s = s1 + s2 + · · · + sr ∈ H1 (∂M ). By definition,


r

r

u+ (Ki ) = sign(x) t|ω x ·s i | . (2.3.5)
i=1 (ω (K i )), ω x ·s i = 0, sign(ω x ·s i )=sign(x)
i=1 x∈


For x ∈
(ω(Ki )), the loop ωx lies on ω(Ki ) and is disjoint from j = i ω(Kj ). Hence ωx · s =
ωx · si . Using equation (2.3.2), we can rewrite formula (2.3.5) as


r

u+ (Ki ) = sign(x) t|ω x ·s| . (2.3.6)
i=1 (ω |∂ F ), ω x ·s= 0, sign(ω x ·s)=sign(x)
x∈

We now show that each orbit of the involution ν on
 (ω|∂ F ) contributes 0 to the right-
hand side of equation (2.3.6). This will imply the claim of the lemma. Consider first an orbit
consisting of one element x ∈
 (ω|∂ F ) such that ν(x) = x. It suffices to prove that ωx · s =
0 because this would mean that x does not contribute to the right-hand side of equation
(2.3.6). Since the genus of F is 0, the group H1 (F ) is generated by the homology classes of the
boundary components. Formula (2.3.3) implies that there is an integral linear combination h of
s1 , . . . , sr ∈ H1 (∂M ) such that in([ωx ]) = in(h). Then [ωx ]−h ∈ K, where K is the kernel of the
inclusion homomorphism in : H1 (∂M ) → H1 (M ). The sum s = s1 + . . . + sr , being represented
by the boundary of ω(F ), also lies in K. It is well known that the homological intersection form
on H1 (∂M ) restricts to 0 on K; that is, K · K = 0. Hence ([ωx ] − h) · s = 0. Since the curves
ω(K1 ), . . . , ω(Kr ) are pairwise disjoint and the intersection form is skew-symmetric, h · s = 0.
Therefore [ωx ] · s = 0.
Consider an orbit of ν consisting of two distinct points x, y ∈
(ω|∂ F ). As we know, sign(x) =
− sign(y). Arguing as in the previous paragraph with [ωx ] replaced by [ωx ] + [ωy ] and using
 COBORDISM OF KNOTS ON SURFACES 293

equation (2.3.4), we conclude that ωx · s = −ωy · s. This implies that either ωx · s = ωy · s = 0,

so that x, y do not contribute to the right-hand side of equation (2.3.6) or the integers ωx · s,
ωy · s are nonzero and differ by sign. If sign(ωx · s) = − sign(x), then sign(ωy · s) = − sign(y)
and x, y do not contribute to the right-hand side of (2.3.6). If sign(ωx · s) = sign(x), then
sign(ωy · s) = sign(y) and the contributions of x and y cancel each other.

(m 1 ,...,m k )
Corollary 2.3.3. For any integers m1 , . . . , mk
1, the polynomials u± are cobor-
dism invariants of knots.
This follows from Lemma 2.1.1 and Theorem 2.3.1.
Corollary 2.3.4 For any slice knot K and any integers m1 , . . . , mk
1, we have
(m ,...,m k )
u± 1 (K)
= 0. In particular, u± (K) = 0.
For example, the knot K ⊂ S 1 × S 1 × R constructed in Subsection 1.5(2) is nonslice, since
u+ (K) = t = 0.

3. Graded matrices and knots

We introduce abstract graded matrices and define the graded matrices of knots.

3.1. Graded matrices

A graded matrix over an abelian group A is a triple (G, s, b), where G is a finite set with
distinguished element s, such that the set G − {s} is partitioned as a union of two disjoint
subsets G+ , G− (possibly empty) and b is an arbitrary mapping G × G → A. The partition
G−{s} = G+
G− can be equivalently described in terms of the function sign : G−{s} → {±1}
sending G+ to +1 and G− to −1. An example of a graded matrix is provided by the trivial
graded matrix (G, s, b), where G = {s} and b(s, s) = 0.
Two graded matrices (G, s, b) and (G , s , b ) over A are isomorphic if there is a bijection
G → G transforming s into s , b into b , and G± into G± .
To a graded matrix T = (G, s, b) over A, we associate two graded matrices −T and T − over A.
By definition, −T = (−G, s, −b), where −G = G with (−G)± = G∓ and (−b)(g, h) = −b(g, h)
for g, h ∈ G. The graded matrix T − is defined by T − = (G, s, b− ), where b− (s, h) = −b(s, h),
b− (h, s) = −b(h, s) for h ∈ G, b− (g, h) = b(g, h) − b(g, s) − b(s, h) for g, h ∈ G − {s}, and the
partition G − {s} = G+
G− is preserved. The transformations T
→ −T and T
→ T − are
commuting involutions on the class of graded matrices.

3.2. Skew-symmetric graded matrices

A graded matrix (G, s, b) over A is skew-symmetric if the mapping b : G × G → A is skew-
symmetric in the sense that b(g, h) = −b(h, g) for all g, h ∈ G and b(g, g) = 0 for all g ∈ G.
Given a skew-symmetric graded matrix (G, s, b), we say that an element g ∈ G − {s} has
type 1 if b(g, h) = 0 for all h ∈ G. We say that g ∈ G − {s} has type 2 if b(g, h) = b(s, h) for
all h ∈ G. We call two elements g1 , g2 ∈ G − {s} complementary if sign(g1 ) = − sign(g2 ) and
b(g1 , h) + b(g2 , h) = b(s, h) for all h ∈ G.
We define three transformations (or moves) M1 , M2 , M3 on skew-symmetric graded matrices.
The moves M1 , M2 , M3 delete a type 1 element, a type 2 element, and a pair of complementary
elements, respectively. The inverse moves M1−1 , M2−1 , M3−1 add to a skew-symmetric graded
matrix (G, s, b) a type 1 element, a type 2 element, and a pair of complementary elements,
respectively. More precisely, the moves M1−1 and M2−1 transform (G, s, b) into a skew-symmetric
graded matrix (G = G
{g}, s, b) such that b : G × G → A extends b and, respectively,
b(g, h) = 0 and b(g, h) = b(s, h) for all h ∈ G. The function sign on G − {s} extends the given
294 VLADIMIR TURAEV

function sign on G − {s} and takes an arbitrary value ±1 on g. The move M3−1 transforms
(G, s, b) into a graded matrix (G = G
{g1 , g2 }, s, b), where b : G
×G
 → A is any skew-
  
symmetric map extending b and such that b(g1 , h) + b(g2 , h) = b(s, h) for all h ∈ G.  The

function sign on G − {s} extends the given function sign on G − {s}, and takes an arbitrary
value ±1 on g1 and the opposite value on g2 . Although we shall not need it, note that the move
M2−1 can be expanded as a composition of M3−1 and M1 .
Two skew-symmetric graded matrices are homologous if they can be obtained from each other
by a finite sequence of transformations M1±1 , M2±1 , M3±1 and isomorphisms. The homology is
an equivalence relation on the class of skew-symmetric graded matrices over A.

3.3. Primitive graded matrices

A graded matrix over an abelian group is primitive if it is skew-symmetric and has no
elements of type 1 or 2 and no complementary pairs of elements. For instance, the trivial
graded matrix is primitive.
Starting from an arbitrary skew-symmetric graded matrix T and recursively deleting ele-
ments of types 1, 2, and complementary pairs of elements, we eventually obtain a primitive
graded matrix T• . The following lemma shows that T• is determined by T uniquely up to
isomorphism.

Lemma 3.3.1. Two homologous primitive graded matrices are isomorphic.

Proof. We begin with the following assertion:

(∗) a move Mi−1 followed by Mj with i, j ∈ {1, 2, 3} yields the same result as an isomorphism,
or a move Mk±1 , or a move Mk followed by Ml−1 with k, l ∈ {1, 2, 3}.
This assertion will imply the lemma. Indeed, suppose that two primitive graded matrices
T, T  are related by a finite sequence of transformations M1±1 , M2±1 , M3±1 and isomorphisms.
An isomorphism followed by Mi±1 can be also obtained as Mi±1 followed by an isomorphism.
Therefore, all the isomorphisms in our sequence can be accumulated at the end. The claim
(∗) implies that T, T  can be related by a finite sequence of moves consisting of several moves
of type Mi followed by several moves of type Mi−1 and isomorphisms. However, since T is
primitive, we cannot apply to it a move of type Mi . Hence, there are no such moves in the
sequence. Similarly, since T  (and any isomorphic graded matrix) is primitive, it cannot be
obtained by an application of Mi−1 . Therefore, our sequence consists solely of isomorphisms so
that T is isomorphic to T  .
Let us prove (∗). For i, j ∈ {1, 2}, the move Mi−1 on a skew-symmetric graded matrix (G, s, b)
adds one element g and then Mj deletes an element g  ∈ G
{g} of type j. If g  = g, then
Mj ◦ Mi−1 = id. If g  = g, then the transformation Mj ◦ Mi−1 can be achieved by first applying
the Mj that deletes g  ∈ G and then applying the Mi−1 that adds g.
Let i ∈ {1, 2} and j = 3. The move Mi−1 on (G, s, b) adds an element g of type i and Mj
deletes complementary elements g1 , g2 ∈ G
{g}. If g1 , g2 ∈ G, then Mj ◦ Mi−1 can be achieved
by first deleting g1 , g2 and then adding g. If g1 = g, then g2 ∈ G has type 3 − i and Mj ◦ Mi−1
is the move M3−i deleting g2 . The case g2 = g is similar.
Let i = 3 and j ∈ {1, 2}. The move Mi−1 on (G, s, b) adds complementary elements g1 , g2
and Mj deletes a certain g ∈ G
{g1 , g2 } of type j. If g ∈ G, then Mj ◦Mi−1 can be achieved by
first deleting g and then adding g1 , g2 . If g = g1 , then g2 has type 3 − j and Mj ◦ Mi−1 = M3−j −1
.
The case g = g2 is similar.
Let i = j = 3. The move Mi−1 on (G, s, b) adds complementary elements g1 , g2 and Mj
deletes complementary elements g1 , g2 ∈ G
{g1 , g2 }. If the pairs g1 , g2 and g1 , g2 are disjoint,
then Mj ◦ Mi−1 can be achieved by first deleting g1 , g2 ∈ G and then adding g1 , g2 . If these
 COBORDISM OF KNOTS ON SURFACES 295

two pairs coincide, then Mj ◦ Mi−1 is the identity. It remains to consider the case where these
pairs have one common element, say g1 = g1 , while g2 = g2 . Then g2 ∈ G and for all h ∈ G,

b(g2 , h) = b(s, h) − b(g1 , h) = b(s, h) − b(g  , h) = b(g  , h) = b(g  , h).

1 2 2

Therefore, the move Mj ◦ Mi−1 produces a graded matrix isomorphic to (G, s, b). The isomor-
phism (G − {g2 }) ∪ {g2 } ≈ G is the identity on G − {g2 } and sends g2 into g2 . Note that
sign(g2 ) = − sign(g1 ) = − sign(g1 ) = sign(g2 ).

Note finally that the transformations T

→ −T , T
→ T − preserve the class of skew-symmetric
graded matrices and are compatible with the relation of homology on this class. Clearly,
(−T )• = −T• and (T − )• = (T• )− .

3.4. Graded matrices of knots

To a knot diagram D on a surface Σ, we associate a skew-symmetric graded matrix T (D) =
(G, s, b) over Z as follows. Set G = {s}


(D) and provide G−{s} =
(D) with the bipartition
G+ =
+ (D) and G− =
− (D). To define the pairing b : G × G → Z, consider the mapping
α : G → H1 (Σ) sending s to [D] ∈ H1 (Σ) and sending any x ∈
(D) to [Dx ] ∈ H1 (Σ). The
pairing b is defined by b(g, h) = α(g)·α(h) ∈ Z for g, h ∈ G. It is clear that b is skew-symmetric.

Lemma 3.4.1. If two knot diagrams on Σ present isotopic knots in Σ×R, then their graded
matrices are homologous.

Proof. It suffices to verify that, if knot diagrams D, D on Σ are related by Reidemeister

moves, then their graded matrices are homologous. The arguments given in Section 1.3 in the
proof of isotopy invariance of u± show that if D is obtained from D by the first R-move, then
T (D) is obtained from T (D ) by M1 or M2 . Likewise, if D is obtained from D by the second
R-move, then T (D) is obtained from T (D ) by M3 . If D is obtained from D by the third
R-move, then T (D) is isomorphic to T (D ).

This lemma allows us to define the graded matrix of a knot K ⊂ Σ × R. Present K by a

diagram D on Σ. By Lemma 3.4.1, the homology class of the skew-symmetric graded matrix
T (D) does not depend on the choice of D. By Lemma 3.3.1, the primitive graded matrix over Z
defined by T• (K) = (T (D))• is determined by K uniquely up to isomorphism. The isomorphism
class of T• (K) = (G• , s• , b• ) is a diffeomorphism invariant of K. Note one application of this
invariant: any diagram of K on Σ must have at least card(G• ) − 1 double points. A related
fact is that if K is presented by a diagram D such that T (D) is primitive, then any diagram
of K on Σ has at least as many double points as D. Indeed, in this case T• (K) = T (D) and
card(G• ) − 1 = card(
(D)).
It is easy to check that T• (−K) = (T• (K))− and T• (K) = −(T• (K))− . If K bounds a disk
in Σ × R, then T• (K) is the trivial graded matrix.
More generally, for a finite sequence of positive integers m1 , . . . , mk , we define the higher
graded matrix of K by

T•m 1 ,...,m k (K) = T• (K (m 1 ,...,m k ) ).

This primitive graded matrix over Z is a diffeomorphism invariant of K.

296 VLADIMIR TURAEV

3.5. Polynomials u± re-examined

For a graded matrix T = (G, s, b) over Z and ε = ±1, set


uε (T ) = sign(x) t|b(g ,s)| ∈ Z[t] .
x∈G , b(g ,s)= 0, sign(b(g ,s))=ε sign(x)

It is clear that uε (−T ) = −uε (T ) and uε (T − ) = u−ε (T ). Both u+ and u− are homology
invariants of skew-symmetric graded matrices.
The polynomial invariants u± (K) of any knot K can be computed from T• (K) by u± (K) =
u± (T• (K)). Indeed, for a diagram D of K,
u± (K) = u± (T (D)) = u± ((T (D))• ) = u± (T• (K)).
More generally, um
±
1 ,...,m k
(K) = u± (T•m 1 ,...,m k (K)) for any integers m1 , . . . , mk
1.

3.6. Remark
By [14], any loop on a surface gives rise to a based matrix (based matrices are defined as
graded matrices (G, s, b), but no bipartition of G − {s} is distinguished). The based matrix of
the loop underlying a knot diagram D is obtained from T (D) by ignoring the bipartition.

3.7. Examples
(1) Consider the knot K on S 1 × S 1 presented by the diagram D with two crossings x, y
from Subsection 1.5(2). Then T (D) = (G, s, b), where G = {s, x, y} with G+ = {x, y}, G− = ∅
and the mapping b : G × G → Z is given by the matrix
⎡ ⎤
0 −1 1
⎢ ⎥
⎣ 1 0 1 ⎦, (3.7.1)
−1 −1 0
where the rows and the columns correspond to s, x, y, respectively. This matrix is primitive
and therefore T• (K) = (T (D))• = T (D). The equalities Dx · D = −Dy · D = 1 imply that the
knot K (m ) with m
2 is presented by a diagram without self-crossings; therefore the higher
graded matrices of K are trivial.
(2) For any integer p, q
1, the author defined in [14, Sections 3.3 and 4.3] a closed curve
αp,q on a closed (oriented) surface Σ. This curve is defined by its Gauss diagram consisting
of a circle in R2 with p horizontal chords directed leftward and q vertical chords directed
upward (each horizontal chord should cross each vertical chord inside the circle). The genus
of Σ is equal to 1 if p = q = 1, is equal to 3 if min(p, q)
3, and is equal to 2 in all the
other cases. The curve αp,q has p + q double points x1 , . . . , xp+q corresponding to the chords
of the Gauss diagram. To transform αp,q into a diagram of a knot K on Σ, it is enough to fix
a function sign : {x1 , . . . , xp+q } → {±1}. The corresponding skew-symmetric graded matrix
T = (G, s, b : G × G → Z) is computed as follows (cf. [14]): G = {s, x1 , . . . , xp+q }; b(xi , s) = q
for i ∈ {1, . . . , p}; and b(xp+j , s) = −p for j ∈ {1, . . . , q}; b(xi , xi
) = b(xp+j , xp+j
) = 0 for
i, i ∈ {1, . . . , p}, j, j  ∈ {1, . . . , q}; and b(xi , xp+j ) = p + q + 1 − i − j for i ∈ {1, . . . , p},
j ∈ {1, . . . , q}. The partition G − {s} = G+ ∪ G− is determined by the function sign in the
usual way. It is easy to check from the definitions that the graded matrix T is primitive except
in the case where p = q = 1 and sign(x1 ) = − sign(x2 ). Excluding this case, we find that
T• (K) = T . This implies, in particular, that K is nontrivial and any diagram on Σ presenting
a knot isotopic to K has at least p + q double points. To compute the polynomial u± (K), set
a± = card{1  i  p, sign(xi ) = ±1},
b± = card{1  j  q, sign(xp+j ) = ±1}.
 COBORDISM OF KNOTS ON SURFACES 297

Then
u± (K) = u± (T ) = ±(a± tq − b∓ tp ).
This formula easily implies that such knots K corresponding to different pairs (p, q) are never
diffeomorphic and that two knots corresponding to the same pair (p, q) and different functions
{x1 , . . . , xp+q } → {±1} may be isotopic only if these two functions take the value +1 the same
number of times and the restrictions of these functions to the set {x1 , . . . , xp } take the value
+1 the same number of times. In the case where p = q = 1 and sign(x1 ) = sign(x2 ) = 1,
we recover the knot K from the previous example. In the exceptional case p = q = 1 and
sign(x1 ) = − sign(x2 ), the knot K is isotopic to a knot on Σ presented by a diagram without
self-crossings.
In this example, the higher graded matrix T•m (K) is nontrivial if the integer m divides p or
q. For instance, if m divides both p and q, then K (m ) = K and T•m (K) = T . If m divides p
but not q, then T•m (K) is obtained from T by deleting x1 , . . . , xp .

4. Genus and cobordism for graded matrices

We introduce a numerical genus for finite families of graded matrices and use it to define a
relation of cobordism for graded matrices (not necessarily skew-symmetric). This relation will
be confronted with knot cobordism in the next section.
Throughout this section, the symbol R denotes a domain, that is, a commutative ring without
zero divisors. By a graded matrix over R, we mean a graded matrix over the additive group
of R.

4.1. The genus

Consider a family of r
1 graded matrices T1 = (G1 , s1 , b1 ), . . . , Tr = (Gr , sr , br ) over R. We
define a numerical invariant σ(T1 , . . . , Tr ) of this family called the graded genus. First of all,
replacing T1 , . . . , Tr by
risomorphic graded rmatrices, we can assume that the sets G1 , . . . , Gr
are disjoint. Set G = t=1 Gt and G± = t=1 (Gt )± ⊂ G. Denote RG the free R-module with
basis G. The maps {bt : Gt × Gt → R}t induce a bilinear form b = ⊕t bt : RG × RG → R such
that b(g, h) = bt (g, h) for g, h ∈ Gt and b(Gt , Gt
) = 0 for t = t .
Let S be the submodule of RG generated by s1 , . . . , sr . We call a vector x ∈ RG short if
x ∈ S or x ∈ g + S for some g ∈ G − {s1 , . . . , sr } or x ∈ g + h + S for g, h ∈ G − {s1 , . . . , sr } of
opposite sign (that is, g ∈ G+ , h ∈ G− or g ∈ G− , h ∈ G+ ). A (graded) filling of T1 , . . . , Tr is a
finite family {λi }i of short vectors in RG such that i λi = g ∈G g (mod S) and one of the λi
is equal to s1 + s2 + . . . + sr . Note that each basis vector g ∈ G − {s1 , . . . , sr } appears in exactly
one λi with coefficient +1 and does not appear in other λi . The basis vectors s1 , . . . , sr may
appear in several λi with nonzero coefficients. For example, the set consisting of all elements
of G and the vector s1 + s2 + . . . + sr is a filling of T1 , . . . , Tr .
The matrix of a filling λ = {λi }i is the square matrix (b(λi , λj ))i,j over R. Let σ(λ) be half
of its rank. This number is equal to half of the rank of the restriction of b to the submodule
of RS generated by the vectors {λi }i . The number σ(λ)
0 is an integer or a half-integer; it is
certainly an integer if T1 , . . . , Tr (and then b) are skew-symmetric. Set
σ(T1 , . . . , Tr ) = min σ(λ)
0,
λ

where λ runs over all fillings of T1 , . . . , Tr . Clearly, σ(T1 , . . . , Tr ) = 0 if and only if (T1 , . . . , Tr )
has a filling with zero matrix. In this case, we call the family T1 , . . . , Tr hyperbolic.
It is obvious that σ(T1 , . . . , Tr ) is preserved when T1 , . . . , Tr are permuted or replaced with
isomorphic graded matrices. It is easy to check that
σ(−T1 , . . . , −Tr ) = σ(T1 , . . . , Tr ).
298 VLADIMIR TURAEV

If Tr is a trivial graded matrix, then σ(T1 , . . . , Tr ) = σ(T1 , . . . , Tr −1 ) (because then the vector
sr ∈ S lies in the annihilator of the form b).
For r = 1, the definitions above apply to a single graded matrix T and yield the genus
σ(T )
0. By definition, T is hyperbolic if and only if σ(T ) = 0.
We state a key property of the genus.

Lemma 4.1.1. For any 1  t  r and any graded matrices T0 , T1 , . . . , Tr ,

σ(T1 , . . . , Tr )  σ(T1 , . . . , Tt , T0 ) + σ(−T0 , Tt+1 , . . . , Tr ).

Proof. Consider for concreteness the case where t = 1 and r = 2; the general case is similar.
We must prove that σ(T1 , T2 )  σ(T1 , T0 ) + σ(−T0 , T2 ). Let Ti = (Gi , si , bi ) for i = 0, 1, 2 and
let T0 = (G0 , s0 , b0 ) be a copy of T0 , where G0 = {g  | g ∈ G0 }, s0 = (s0 ) , and b0 is defined
by b0 (g  , h ) = b0 (g, h) for g, h ∈ G0 . We can assume that the sets G1 , G0 , G0 , G2 are disjoint.
Let Λ1 , Λ0 , Λ0 , Λ2 be free R-modules freely generated by G1 , G0 , G0 , G2 , respectively, and let
Λ = Λ1 ⊕Λ0 ⊕Λ0 ⊕Λ2 . There is a unique skew-symmetric bilinear form B = b1 ⊕b0 ⊕(−b0 )⊕b2
on Λ such that the sets G1 , G0 , G0 , G2 ⊂ Λ are mutually orthogonal and the restrictions of B
to these sets are equal to b1 , b0 , −b0 , b2 , respectively.
Let Φ be the submodule of Λ0 ⊕ Λ0 generated by the vectors {g + g  }g ∈G 0 . Set L = Λ1 ⊕
Φ ⊕ Λ2 ⊂ Λ. The projection p : L → Λ1 ⊕ Λ2 along Φ transforms B into b1 ⊕ b2 . Indeed, for
any g1 , h1 ∈ G1 , g, h ∈ G0 , g2 , h2 ∈ G2 , we have
B(g1 + g + g  + g2 , h1 + h + h + h2 ) = b1 (g1 , h1 ) + b0 (g, h) + (−b0 )(g  , h ) + b2 (g2 , h2 )
= b1 (g1 , h1 ) + b2 (g2 , h2 ).
Pick a filling {λi }i ⊂ Λ1 ⊕ Λ0 of the pair (T1 , T0 ) whose matrix has rank 2σ(T1 , T0 ). This
means that the restriction of B to the module V1 ⊂ Λ1 ⊕ Λ0 generated by {λi }i has rank
2σ(T1 , T0 ). Similarly, pick a filling {ϕj }j ⊂ Λ0 ⊕Λ2 of the pair (−T0 , T2 ) such that the restriction
of B to the module V2 ⊂ Λ0 ⊕ Λ2 generated by {ϕj }j has rank 2σ(−T0 , T2 ). We claim that
there is a finite set ψ ⊂ (V1 + V2 ) ∩ L such that the set p(ψ) ⊂ Λ1 ⊕ Λ2 is a filling of the
pair (T1 , T2 ). Denoting by V the submodule of Λ1 ⊕ Λ2 generated by p(ψ), we obtain then the
desired inequality
σ(T1 , T2 )  σ(p(ψ)) = (1/2) rank((b1 ⊕ b2 )|V ) = (1/2) rank(B|p −1 (V ) )
 (1/2) rank(B|(V 1 +V 2 )∩L )  (1/2) rank(B|V 1 +V 2 )
= (1/2) rank(B|V 1 ) + (1/2) rank(B|V 2 ) = σ(T1 , T0 ) + σ(−T0 , T2 ).
Here the second inequality follows from the inclusion
p−1 (V ) ⊂ (V1 + V2 ) ∩ L + Ker p
and the fact that Ker p = Φ lies in the annihilator of B|L .
To construct ψ, we first modify {λi }i as follows. Let λ1 be the vector of this filling equal to
s1 + s0 . Adding appropriate multiples of λ1 to other λi , we can ensure that the basis vector
s0 ∈ G0 appears in all {λi }i= 1 with coefficient 0. This transforms {λi }i into a new filling of
(T1 , T0 ) which will from now on be denoted λ = {λi }i . This transformation does not change
the module V1 generated by the vectors of the filling. Similarly, we can assume that a vector
ϕ1 of the filling ϕ = {ϕj }j is equal to s0 + s2 and that the basis vector s0 ∈ G0 appears in all
{ϕj }j = 1 with coefficient 0.
The filling λ gives rise to an oriented 1-dimensional manifold Γλ with boundary (G1 −
{s1 }) ∪ (G0 − {s0 }). It is defined as follows. Each λi having the form g + h (mod Rs1 ) with
g, h ∈ (G1 − {s1 }) ∪ (G0 − {s0 }) gives rise to a component of Γλ diffeomorphic to [0, 1] and
connecting g with h. By the definition of a filling, the elements g, h have opposite signs. We
orient the component in question so that it leads from the element with sign −1 to the element
 COBORDISM OF KNOTS ON SURFACES 299

with sign +1. Each λi having the form g (mod Rs1 ) with g ∈ (G1 − {s1 }) ∪ (G0 − {s0 })
gives rise to a component of Γλ which is a copy of [0, ∞), where 0 is identified with g. We
orient this component toward g if sign(g) = +1 and out of g if sign(g) = −1. All the other
λi (in particular λ1 ) do not contribute to Γλ . The definition of a filling implies that ∂Γλ =
(G1 −{s1 })∪(G0 −{s0 }). Similarly, the filling ϕ gives rise to an oriented 1-dimensional manifold
Γϕ with boundary (G0 − {s0 }) ∪ (G2 − {s2 }). We can assume that Γλ and Γϕ are disjoint.
Gluing Γλ to Γϕ along the canonical identification G0 − {s0 } → G0 − {s0 }, g
→ g  , we obtain
a 1-dimensional manifold Γ with ∂Γ = (G1 − {s1 }) ∪ (G2 − {s2 }). Note that the orientations
of Γλ , Γϕ at G0 − {s0 } = G0 − {s0 } are compatible because (G0 )± = (G0 )∓ (cf. the definition
of the transformation T
→ −T ). Therefore, these orientations extend to an orientation of Γ.
Each component C of Γ is glued from several components of Γλ
Γϕ associated with certain
vectors
λi ∈ V1 ⊂ Λ 1 ⊕ Λ 0 ⊂ Λ or ϕj ∈ V2 ⊂ Λ0 ⊕ Λ2 ⊂ Λ.

Let ψC ∈ V1 + V2 ⊂ Λ be the sum of these vectors over all components of Γλ
Γϕ contained

in C. The following two facts imply that ψC ∈ L:
(1) each point of C ∩ (G0 − {s0 }) ≈ C ∩ (G0 − {s0 }) is adjacent to one component of Γλ and
to one component of Γϕ ; and
(2) s0 does not show up in {λi }i= 1 and s0 does not show up in {ϕj }j = 1 .
Set ψ1 = λ1 + ϕ1 = s1 + s0 + s0 + s2 ∈ Λ. Clearly, ψ1 ∈ (V1 + V2 ) ∩ L. Set ψ = {ψ1 } ∪ {ψC }C ,
where C runs over the components of Γ with nonvoid boundary. As we know, ψ ⊂ (V1 ⊕V2 )∩C.
Let us check that p(ψ) ⊂ Λ1 ⊕ Λ2 is a filling of (T1 , T2 ). Observe that for a compact component
C of Γ with endpoints g, h ∈ G1 ∪ G2 , we have p(ψC ) = g + h (mod Rs1 + Rs2 ). If C is
oriented, say, from g to h, then necessarily sign(g) = −1 and sign(h) = +1. For a noncompact
component C of Γ with one endpoint g ∈ G1 ∪ G2 , we have p(ψC ) = g (mod Rs1 + Rs2 ).
both cases, the vector p(ψC ) is short. The sum of all vectors in the family ψ is equal to
In
g ∈G 1 ∪G 2 g (mod Rs1 +Rs2 ). Also p(ψ1 ) = s1 +s2 . This means that p(ψ) is a filling of (T1 , T2 ),
so that ψ satisfies all the required conditions.

4.2. Cobordism of graded matrices

Two graded matrices T1 , T2 over R are cobordant if σ(T1 , −T2 ) = 0.

Theorem 4.2.1. (i) Cobordism is an equivalence relation on the class of graded matrices.
(ii) Isomorphic graded matrices are cobordant.
(iii) The graded genus of a family of graded matrices is a cobordism invariant.
(iv) A graded matrix is cobordant to the trivial graded matrix if and only if it is hyperbolic.
(v) Homologous skew-symmetric graded matrices are cobordant.

Proof. (i) and (ii) For a graded matrix T = (G, s, b), the graded matrix −T is isomorphic
to the triple (G , s , b ), where G = {g  | g ∈ G} is a disjoint copy of G, (G )± = G∓ , and
b (g  , h ) = −b(g, h) for any g, h ∈ G. The vectors {g + g  }g ∈G form a filling of the pair (T, −T ).
The matrix of this filling is 0. Therefore σ(T, −T ) = 0, so that T is cobordant to itself. A
similar argument proves part (ii). That the relation of cobordism is symmetric follows from the
equalities σ(T2 , −T1 ) = σ(−T2 , T1 ) = σ(T1 , −T2 ). The transitivity follows from the inequalities

0  σ(T1 , −T3 )  σ(T1 , −T2 ) + σ(T2 , −T3 ),

where the second inequality is a special case of Lemma 4.1.1.

300 VLADIMIR TURAEV

(iii) We need to prove that σ(T1 , . . . , Tr ) is preserved when T1 , . . . , Tr are replaced with
cobordant graded matrices. By induction, it suffices to prove that

σ(T1 , . . . , Tr −1 , Tr ) = σ(T1 , . . . , Tr −1 , Tr ) (4.2.1)

for any graded matrix Tr cobordant to Tr . By Lemma 4.1.1,

σ(T1 , . . . , Tr −1 , Tr )  σ(T1 , . . . , Tr −1 , Tr ) + σ(−Tr , Tr ) = σ(T1 , . . . , Tr −1 , Tr ).

Similarly, σ(T1 , . . . , Tr −1 , Tr )  σ(T1 , . . . , Tr −1 , Tr ). This implies equation (4.2.1).

(iv) If a graded matrix T is cobordant to the trivial graded matrix T0 = ({s0 }, s0 , b = 0),
then σ(T ) = σ(T0 ) = 0 and T is hyperbolic. Conversely, if T = (G, s, b) is hyperbolic, then it
has a filling λ = {λi }i with zero matrix. Since s = λi for some i, we necessarily have b(s, λi ) = 0
for all i (and in particular, b(s, s) = 0). Adding to λ the vector s + s0 , we obtain a filling of
the pair (T, T0 ) with zero matrix. Hence, T is cobordant to −T0 = T0 .
(v) Let a skew-symmetric graded matrix T  = (G , s , b ) be obtained from a skew-symmetric
graded matrix T = (G, s, b) by a move Mi−1 with i = 1, 2, 3. We can assume that G is a union
of a disjoint copy {h | h ∈ G} of G and one new element g in the case i = 1, 2 or two new
elements g1 , g2 in the case i = 3. For i = 1, 2, 3, the vectors {h + h }h∈G and the vectors g,
g − s and g1 + g2 − s , respectively form a filling of the pair (T, −T  ). The matrix of this filling
is zero. Hence σ(T, −T  ) = 0, so that T is cobordant to T  .

Corollary 4.2.2. For any skew-symmetric graded matrices T1 , . . . , Tr , we have

σ((T1 )• , . . . , (Tr )• ) = σ(T1 , . . . , Tr ).

4.3. Remarks
We point out a few further properties of graded matrices.
(1) Let us call a filling of a graded matrix T = (G, s, b) simple if all vectors of the filling are
pairwise distinct and either belong to the set of basis vectors G or have the form g + h with
g, h ∈ G − {s} and sign(g) = − sign(h). Any filling λ of T can be transformed into a simple
one by adding vectors proportional to s to the vectors of λ distinct from s (and eliminating
repetitions). This transformation does not change the module generated by the vectors of λ
and does not change σ(λ). Thus, to compute σ(T ), we can restrict ourselves to the simple
fillings. The simple fillings of T = (G, s, b) are finite in number; they bijectively correspond to
involutions ν on G − {s} such that every free orbit of ν meets both G+ and G− . In particular,
if one of the sets G+ , G− is empty, then T has only one simple filling consisting of all elements
of G and then σ(T ) is half of the rank of the matrix (b(g, h))g ,h∈G . Such a graded matrix T is
hyperbolic if and only if b = 0. For example, the graded matrix of Subsection 3.7(1) satisfies
G− = ∅, b = 0 and therefore it is not hyperbolic.
(2) The function (T1 , T2 )
→ σ(T1 , −T2 ) defines a metric on the set of cobordism classes of
graded matrices over a domain.
(3) For any graded matrices T1 , . . . , Tr , T1 , . . . , Tq with q
1 and any 1  t  r,

σ(T1 , . . . , Tr )  σ(T1 , . . . , Tt , T1 , . . . , Tq ) + σ(−T1 , . . . , −Tq , Tt+1 , . . . , Tr ) + q − 1.

(4) For a graded matrix T = (G, s, b) over a domain R, the element b(s, s) of R is a cobordism
invariant of T . We call T normal if b(s, s) = 0. For example, all skew-symmetric graded matrices
are normal. For normal graded matrices over Z, the polynomials u+ and u− are cobordism
invariants. This follows from the following more general fact: if a family T1 , . . . , Tr of normal
graded matrices over Z is hyperbolic, then u± (T1 ) + · · · + u± (Tr ) = 0.
 COBORDISM OF KNOTS ON SURFACES 301

5. Cobordism of knots versus cobordism of matrices

5.1. The matrices of cobordant knots

Setting R = Z, we can apply the definitions and results of Section 4 to the graded matrices
of knots.The following theorem relates cobordisms of knots to cobordisms of their matrices.

Theorem 5.1.1. The graded matrices of cobordant knots are cobordant.

The proof of this theorem will be given in Subsection 5.3 using the notion of a slice genus
introduced in Subsection 5.2.

Corollary 5.1.2. If two knots K1 , K2 are cobordant, then for any m1 , . . . , mk
1, the
higher graded matrices T•m 1 ,...,m k (K1 ) and T•m 1 ,...,m k (K2 ) are cobordant.

Corollary 5.1.3. The graded matrices and the higher graded matrices of slice knots are
hyperbolic.

5.2. Slice genus

The slice genus sg(K1 , . . . , Kr ) of r
1 knots K1 ⊂ Σ1 × R, . . ., Kr ⊂ Σr × R is the minimal
integer k
0, satisfying the following condition: there is an oriented 3-manifold M such that

ri=1 Σi ⊂ ∂M and the 1-manifold

ri=1 Ki ⊂
ri=1 Σi × R ⊂ ∂M × R
bounds in M × R an embedded compact (oriented) surface of genus k. If there is no such k,
then sg(K1 , . . . , Kr ) = +∞. Note that we do not require M or the surface in M × R to be
connected, although it is always possible to achieve this by taking connected sum. The genus
of a disconnected surface is by definition the sum of the genera of its components.
Computing the slice genus of a family of knots is an interesting geometric problem. The
following lemma estimates the slice genus from below.

Lemma 5.2.1. For any knots K1 , . . . , Kr ,

2 sg(K1 , . . . , Kr )
σ(T• (K1 ), . . . , T• (Kr )).

Proof. Consider a 3-manifold M as in the definition of the slice genus and an embedded
compact surface F ⊂ M × R of genus k = sg(K1 , . . . , Kr ) with ∂F =
ri=1 Ki ⊂ ∂M × R.
By assumption, the components of ∂F project to disjoint subsets of ∂M . As in the proof of
Lemma 2.3.2, we can additionally assume that F ∩ (∂M × R) = ∂F and the restriction to F
of the projection M × R → M is a generic map. Denote this map F → M by ω.
The group H1 (F ) = H1 (F ; Z) is generated by the homology classes of r boundary compo-
nents of F and a subgroup H ⊂ H1 (F ) isomorphic to Z2k . Set
L = in−1 (ω∗ (H)) ⊂ H1 (∂M ),
where ω∗ : H1 (F ) → H1 (M ) is the homomorphism induced by ω and in : H1 (∂M ) → H1 (M ) is
the inclusion homomorphism. Denote the homological intersection form H1 (∂M ) × H1 (∂M ) →
Z by B. Since B annihilates the kernel of in,
rank (B|L : L × L → Z)  2 rank ω∗ (H)  2 rank H = 4k. (5.2.1)
For t = 1, . . . , r, consider the graded matrix Tt = (Gt , st , bt ) of the diagram of Kt associated
with the projection of Ki to ∂M . We derive from ω : F → M a filling λ of the tuple T1 , . . . , Tr .
302 VLADIMIR TURAEV

Set G =
t Gt . As in the proof of Lemma 2.3.2, the map ω gives rise to an involution ν on the
set
(ω|∂ F ) =
rt=1 (Gt − {st }) = G − {s1 , . . . , sr }.


Each orbit X of this involution gives rise to a vector λX in the lattice ZG freely generated by
G. This vector is defined as follows. Set


[X] = [ωx ] ∈ H1 (∂M ),
x∈X

where ωx is the loop on ∂M associated with the self-crossing x of ω|∂ F as in the proof of
Lemma 2.3.2. This proof shows that in([X]) ∈ ω∗ (H1 (F )) or, equivalently,
[X] ∈ in−1 (ω∗ (H1 (F ))).
Therefore, there is a linear combination


r
nX = nX ,t [ω(Kt )] ∈ H1 (∂M )
t=1

of the homology classes [ω(K1 )], . . . , [ω(Kr )] ∈ H1 (∂M ) with nX ,t ∈ Z such that [X] + nX ∈ L.
(Such nX may be nonunique; we take any.) Set



r
λX = x+ nX ,t st ∈ ZG.
x∈X t=1

The vectors {λX }X corresponding to all orbits X of ν together with the vector s1 + . . . + sr ∈
ZG form a filling of the tuple T1 , . . . , Tr . The matrix of this filling is obtained by evaluating
the form B on the homology classes
[X] + nX ∈ H1 (∂M ) and [ω(K1 )] + . . . + [ω(Kr )] ∈ H1 (∂M ).
Since these homology classes belong to L, formula (5.2.1) implies that the rank of this matrix
is smaller than or equal to 4k. Thus,
σ(T1 , . . . , Tr )  2k = 2 sg(K1 , . . . , Kr ).
By definition, T• (Kt ) = (Tt )• for t = 1, . . . , r. By Corollary 4.2.2,
σ(T1 , . . . , Tr ) = σ((T1 )• , . . . , (Tr )• ) = σ(T• (K1 ), . . . , T• (Kr )).
Hence, σ(T• (K1 ), . . . , T• (Kr ))  2 sg(K1 , . . . , Kr ).

5.3. Proof of Theorem 5.1.1

If two knots K1 , K2 are cobordant, then we have sg(K1 , −K 2 ) = 0. By Lemma 5.2.1,
σ(T• (K1 ), T• (−K 2 )) = 0. As we know, T• (−K 2 ) = −T• (K2 ). Thus, σ(T• (K1 ), −T• (K2 )) = 0
so that T• (K1 ) is cobordant to T• (K2 ).

6. Miscellaneous remarks and open questions

Theorem 5.1.1 leads to a number of interesting questions, which we briefly discuss.

6.1. Invariants of graded matrices

Theorem 5.1.1 suggests that we should search for a cobordism classification of skew-symmetric
graded matrices over Z. We list here several constructions of cobordism invariants of graded
matrices (not necessarily skew-symmetric) that may help.
 COBORDISM OF KNOTS ON SURFACES 303

As we know, the genus σ(T ) of a graded matrix T over a domain R is a cobordism invariant
of T . Moreover, for any finite family of graded matrices T1 , . . . , Tr of graded matrices over R,
the genus σ(T, T1 , . . . , Tr ) is a cobordism invariant of T . Further invariants may be obtained
by ring replacements. Given a ring homomorphism ϕ from R to a domain R , we can derive
from T = (G, s, b) the graded matrix Tϕ = (G, s, ϕ ◦ b) over R . The cobordism invariants of Tϕ
will be cobordism invariants of T . For example, let R = Z, R = Z/pZ, where p
2 is a prime
integer, and let ϕ : R → R be the projection. Then the p-genus σp (T ) = σ(Tϕ ) is a cobordism
invariant of a graded matrix T over Z. It is clear that 0  σp (T )  σ(T ) and that σp (T ) = σ(T )
for any given T and all sufficiently big prime integers p. Thus, σ(T ) = maxp σp (T ).
More invariants of graded matrices over a domain R can be obtained using the following
transformations of graded matrices. Pick a set A ⊂ R such that −A = A, where −A =
{−a | a ∈ A}. For a graded matrix T = (G, s, b) over R, set GA = {g ∈ G | b(g, s) ∈ A}. The
bipartition of G induces a bipartition of GA by (GA )± = GA ∩G± . We say that T is A-normal,
if b(s, s) ∈ A or equivalently, if s ∈ GA . If T is A-normal, then γA (T ) = (GA , s, b|G A ) is a graded
matrix over R. It is easy to see that if two A-normal graded matrices T1 , T2 are cobordant, then
the graded matrices γA (T1 ), γA (T2 ) are cobordant. Thus, the formula T
→ γA (T ) induces a
transformation on the set of cobordism classes of A-normal graded matrices. This construction
can be applied to a skew-symmetric graded matrix T over Z and A = (−B) ∪ {0} ∪ B, where
B is an arbitrary set of positive integers. In particular, the genus and the p-genera of γA (T )
are cobordism invariants of T parametrized by B.

6.2. Realization problem

It would be interesting to give an algebraic description of all the primitive graded matrices
over Z that can be realized as the graded matrices of knots. Formula 1.4.1 yields a necessary
condition.

6.3. Secondary obstructions to sliceness

As we know, the graded matrix and the higher graded matrices of knots yield obstructions
to the sliceness of knots. We outline a Casson–Gordon style construction of secondary obstruc-
tions.
Let K be a knot on a closed connected (oriented) surface Σ and let K be the projection of
K to Σ. Fix an integer m
2 and set H = H1 (Σ; Z/mZ). For any h ∈ H such that h · K = 0,
consider the covering Σh → Σ corresponding to the normal subgroup of π1 (Σ) consisting of the
homotopy classes of loops on Σ whose intersection number with h is 0. The knot K ⊂ Σ × R
lifts to a knot Kh ⊂ Σh × R and the diffeomorphism type of Kh does not depend on the choice
of the lift. The invariants of Kh can be viewed as invariants of K parametrized by m and h.
In particular, we can consider the polynomials u± (Kh ), the graded matrix T• (Kh ), the slice
genus of Kh , and so on. In the next lemma, a Lagrangian is an additive group L ⊂ H equal to
its annihilator Ann(L) = {g ∈ H | g · L = 0}.

Lemma 6.3.1. If K is slice, then there is a Lagrangian L ⊂ H satisfying the following

conditions: [K] ∈ L; the knot Kh is slice for all h ∈ L; for any diagram D of K on Σ, there is
involution on the set

an (D) such that for any orbit of this involution X ⊂
(D), we have
x∈X [Dx ] ∈ L.

This lemma implies that for all h ∈ L, the graded matrix T• (Kh ) is hyperbolic.

Proof. If K is slice, then there exist an oriented 3-manifold M and an embedding Σ ⊂ ∂M

such that the image of K under the induced embedding Σ × R ⊂ ∂M × R bounds an embedded
disk B ⊂ M × R. Replacing, if necessary, M by a compact submanifold containing both Σ and
304 VLADIMIR TURAEV

the projection of B to M , we can assume that M is compact. Gluing handlebodies to all the
components of ∂M distinct from Σ, we can additionally assume that ∂M = Σ.
Set R = Z/mZ. Consider the boundary homomorphism ∂ : H2 (M, ∂M ; R) → H1 (∂M ; R) =
H and the inclusion homomorphism i : H = H1 (∂M ; R) → H1 (M ; R). Set L = Im(∂) = Ker(i).
It is well known that L is a Lagrangian. For completeness, we outline a proof. An element g ∈ H
belongs to Ann(L) if and only if g · ∂a = 0 for every a ∈ H2 (M, ∂M ; R). Given a, the Poincaré
duality says that there is a unique  a ∈ H 1 (M ; R) = Hom(H1 (M ; R), R) such that a = a ∩ [M ].
Then g · ∂a = i(g) · a = a(i(g)). Therefore, g ∈ Ann(L) if and only if i(g) is annihilated by all
homomorphisms H1 (M ; R) → R. This holds if and only if i(g) = 0, that is if and only if g ∈ L.
Pick h ∈ L and pick any a in ∂ −1 (h) ⊂ H2 (M, ∂M ; R). The cohomology class  a ∈ H 1 (M ; R)
defines a cyclic m-fold covering M  → M . The disk B ⊂ M × R lifts to a disk B  ⊂M  × R.
By   → M restricted to Σ = ∂M can be
a(i(g)) = g · ∂a = g · h for all g ∈ H, the covering M
identified with the covering Σh → Σ considered above. Then Kh = ∂ B  ⊂ Σh × R is a slice
knot. Constructing an involution on
 (D) as in the proof of Lemma 2.3.2 (for F = B), we
obtain the last claim of the lemma.

6.4. More on the slice genus

Consider again the slice genus sg(K1 , . . . , Kr )
0 of a family of knots K1 , . . . , Kr . The slice
genus does not depend on the order in the sequence K1 , . . . , Kr and is preserved if K1 , . . . , Kr
are replaced with cobordant knots. If sg(K1 , . . . , Kr ) = 0, then we say that the family of knots
K1 , . . . , Kr is slice. One immediate corollary of Lemma 5.2.1 is the following theorem.

Theorem 6.4.1. If a finite family of knots is slice, then the family of their graded matrices
is hyperbolic.

Observe, in generalization of Lemma 2.1.1, that if the family K1 , . . . , Kr is slice, then

(m ) (m )
the family of the mth coverings K1 , . . . , Kr is slice for any m
1. By Theorem 6.4.1,
if K1 , . . . , Kr is slice, then for any finite sequence of integers m1 , . . . , mk
1, the family of
(m ,...,m k ) (m ,...,m k )
graded matrices T• 1 (K1 ), . . . , T• 1 (Kr ) is hyperbolic.
For r = 2, we can rewrite the slice genus in the equivalent form sg (K1 , K2 ) = sg(K1 , −K2− ).
The number sg (K1 , K2 ) depends only on the cobordism classes of K1 , K2 and defines a metric
on the set of cobordism classes of knots.
In [14], the author defined a slice genus sg for any finite family of disjoint loops on surfaces.
It follows from the definitions that sg(K1 , . . . , Kr )
sg(K 1 , . . . , K r ).

6.5. Long knots

The definitions and results of this paper can be generalized to links and to long knots on
surfaces. The context of long knots is especially interesting since their cobordism classes form
a group. It is a challenging question to compute this group. Is it abelian?

References
1. M. Asaeda, J. Przytycki and A. Sikora, ‘Categorification of the Kauffman bracket skein module of
I-bundles over surfaces’, Algebr. Geom. Topol. 4 (2004) 1177–1210.
2. J. S. Carter, ‘Closed curves that never extend to proper maps of disks’, Proc. Amer. Math. Soc. 113
(1991) 879–888.
3. J. S. Carter, S. Kamada and M. Saito, ‘Stable equivalence of knots on surfaces and virtual knot
cobordisms’, J. Knot Theory Ramifications 11 (2002) 311–322.
4. D. Cimasoni and V. Turaev, ‘A generalization of several classical invariants of links’, Osaka J. Math. 44
(2007) 531–561.
5. P. Cotta-Ramusino and M. Rinaldi, ‘On the algebraic structure of link-diagrams on a 2-dimensional
surface’, Comm. Math. Phys. 138 (1991) 137–173.
 COBORDISM OF KNOTS ON SURFACES 305

6. R. Fenn and V. Turaev, ‘Weyl algebras and knots’, J. Geom. Phys. 57 (2007) 1313–1324.
7. T. Fiedler, Gauss diagram invariants for knots and links, Mathematics and its Applications 532 (Kluwer
Academic Publishers, Dordrecht, 2001).
8. S. Kamada, Braid and knot theory in dimension four, Mathematical Surveys and Monographs 95 (American
Mathematical Society, Providence, RI, 2002).
9. L. Kauffman, ‘Virtual knots theory’, European J. Combin. 20 (1999) 663–690.
10. L. Kauffman and V. Manturov, ‘Virtual knots and links’, Tr. Mat. Inst. Steklova 252 (2006), Geom.
Topol., Diskret. Geom. i Teor. Mnozh., 114–133 (Russian); Proc. Steklov Inst. Math. 252 (2006), 104–121
(English).
11. G. Kuperberg, ‘What is a virtual link?’, Algebr. Geom. Topol. 3 (2003) 587–591.
12. C. Livingston, ‘A survey of classical knot concordance’, Handbook of knot theory (Elsevier, Amsterdam,
2005) 319–347.
13. V. Turaev, ‘Skein quantization of Poisson algebras of loops on surfaces’, Ann. Sci. École Norm. Sup. (4)
24 (1991) 635–704.
14. V. Turaev, ‘Virtual strings’, Ann. Inst. Fourier (Grenoble) 54 (2004) 2455–2525.

Vladimir Turaev
IRMA, Université Louis Pasteur–CNRS
7 Rue René Descartes
F-67084 Strasbourg
France
and
Department of Mathematics
Indiana University
Rawles Hall, 831 East 3rd St
Bloomington, IN 47405
USA
[email protected]