ROBIN FORMAN
0.
Introdu tion
ROBIN FORMAN
and all asso
iated pro
esses will be dis
rete. Hen
e, we have
hosen the name dis
rete
Morse Theory for the ideas we will des
ribe.
Of
ourse, these dierent approa
hes to
ombinatorial Morse Theory are not distin
t. One
an sometimes translate results from one of these theories to another by
\smoothing out" a dis
rete Morse fun
tion, or by
arefully repla
ing a
ontinuous
fun
tion by a dis
rete set of its values. However, that is not the path we will follow
in this paper. Instead, we show that even without introdu
ing any
ontinuity, one
an re
reate, in the
ategory of
ombinatorial spa
es, a
omplete theory that
aptures
many of the intri
a
ies of the smooth theory, and
an be used as an ee
tive tool for
a wide variety of
ombinatorial and topologi
al problems.
The goal this paper is to present an overview of the subje
t of dis
rete Morse
Theory that is su
ient both to understand the major appli
ations of the theory
to
ombinatori
s, and to apply the the theory to new problems. We will not be
presenting theorems in their most re
ent or most general form, and simple examples
will often take the pla
e of proofs. We hope to
onvey the fa
t that the theory is really
very simple, and there is not mu
h that one needs to know before one
an be
ome a
\user". Those interested in a more
omplete presentation of the theory
an
onsult
the referen
e [10. Earlier surveys of this work have appeared in [9 and [13.
1.
CW Complexes
The main theorems of dis
rete (and smooth) Morse Theory are best stated in
the language of CW
omplexes, so we begin with an overview of the basi
s of su
h
omplexes. J.H.C. Whitehead introdu
ed CW
omplexes in his foundational work on
homotopy theory, and all of the results in this se
tion are due to him. The reader
should
onsult [28 for a very
omplete introdu
tion to this topi
. In this paper we
will
onsider only nite CW
omplexes, so many of the subtleties of the subje
t will
not appear.
The building blo
ks of CW
omplexes are
ells. Let B d denote the
losed unit ball
in d-dimensional Eu
lidean spa
e. That is, B d = fx 2 E d s:t:jxj 1g. The boundary
(i)
(ii)
(i). A 1- ell atta hed to a ir le. (ii). This is not a 1- ell atta hed to a ir le
Figure 1.1.
We are now ready for our main denition. A nite CW
omplex is any topologi
al
spa
e X su
h that there exists a nite nested sequen
e
(1.1)
; X0 X1 Xn = X
ROBIN FORMAN
X0
X1
X2
X3
A CW de omposition of a 2-torus
Figure 1.2.
It may seem that quite a bit has been lost in the transition from simpli
ial
omplexes
to general CW
omplexes. After all, a simpli
ial
omplex is
ompletely des
ribed by
a nite amount of
ombinatorial data. On the other hand, the
onstru
tion of a CW
de
omposition requires the
hoi
e of a number of
ontinuous maps. However, if one
is only
on
erned with the homotopy type of the resulting CW
omplex, then things
begin to look a bit more manageable. Namely, the homotopy type of X [f depends
only on the homotopy type of X and the homotopy
las of f .
An important spe
ial
ase is when h is the identity map. We state this
ase separately
for future referen
e.
(See Theorem 2.3 on page 120 of [28.) Therefore, the homotopy type of a CW
omplex
is determined by the homotopy
lass of the atta
hing maps. Sin
e homotopy
lases
are dis
rete obje
ts, we have now re
aptured a bit of the
ombinatorial atmosphere
that we seemingly lost when generalizing from simpli
ial
omplexes to CW
omplexes.
Let us now present some examples.
1) Suppose X is a topologi
al spa
e whi
h has a CW de
omposition
onsisting of
exa
tly one 0-
ell and one d-
ell. Then X has a CW de
omposition ; X0 X1 = X .
The spa
e X0 must be the 0-
ell, and X = X1 is the result of atta
hing the d-
ell
to X0 . Sin
e X0
onsists of a single point, the only possible atta
hing map is the
onstant map. Thus X is
onstru
ted from taking a
losed d-ball and identifying all
of the points on its boundary. One
an easily see that this implies that the resulting
spa
e is a d-sphere.
2) Suppose X is a topologi
al spa
e whi
h has a CW de
omposition
onsisting of
exa
tly one 0-
ell and n d-
ells. Then X has a CW de
omposition as in (1.1) su
h
that X0 is the 0-
ell, and for ea
h i = 1; 2; : : : ; n Xi is the result of atta
hing a d-
ell
to X(i 1) . From the previous example, we know that X1 is a d-sphere. The spa
e X2
is
onstru
ted by atta
hing a d-
ell to X1 . The atta
hing map is a
ontinuous map
from a (d 1)-sphere to X1 . Every map of the (d 1)-sphere into X1 is homotopi
to
a
onstant map (sin
e (d 1) (X1 )
= 0). If the atta
hing map is a
tually
= (d 1) (S d )
a
onstant map, then it is easy to see that the spa
e X2 is the wedge of two d-spheres,
denoted by S d ^ S d . (The wedge of a
olle
tion of topologi
al spa
es is the spa
e
ROBIN FORMAN
resulting from
hoosing a point in ea
h spa
e, taking the disjoint union of the spa
es,
and identifying all of the
hosen points.) Sin
e the atta
hing map must be homotopi
to a
onstant map, Corollary 1.4 implies that X2 is homotopy equivalent to a wedge
of two d-spheres.
When
onstru
ting X3 by atta
hing a d-
ell to X2 , the relevant information is a
map from S d 1 to X2 , and the homotopy type of the resulting spa
e is determined
by the homotopy
lass of this map. All su
h maps are homotopi
to a
onstant map
(sin
e d 1 (X2 )
= 0). Sin
e X2 is homotopy equivalent to a wedge
= d 1 (S d ^ S d )
of two d-spheres, and the atta
hing map is homotopi
to a
onstant map, it follows
from Theorem 1.3 that X3 is homotopy equivalent to the spa
e that would result from
atta
hing a d-
ell to S d ^ S d via a
onstant map, i.e. X3 is homotopy equivalent to a
wedge of three d-spheres.
Continuing in this fashion, we
an see that X must be homotopy equivalent to a
wedge of n d-spheres.
The reader should not get the impression that the homotopy type of a CW
omplex
is determined by the number of
ells of ea
h dimension. This is true only for very
few spa
es (and the reader might enjoy
oming up with some other examples). The
fa
t that wedges of spheres
an, in fa
t, be identied by this numeri
al data partly
explains why the main theorem of many papers in
ombinatorial topology is that a
ertain simpli
ial
omplex is homotopy equivalent to a wedge of spheres. Namely
su
h
omplexes are the easiest to re
ognize. However, that does not explain why so
many simpli
ial
omplexes that arise in
ombinatori
s are homotopy equivalent to a
wedge of spheres. I have often wondered if perhaps there is some deeper explanation
for this.
3) Suppose that X is a CW
omplex whi
h has a CW de
omposition
onsisting
of exa
tly one 0-
ell, one 1-
ell and one 2-
ell. Let us
onsider a CW de
omposition
for X with these
ells: ; X0 X1 X2 = X: We know that X0 is the 0-
ell.
Suppose that X1 is the result of atta
hing the 1-
ell to X0 . Then X1 must be a
ir
le,
and X2 arises from atta
hing a 2-
ell to X1 . The atta
hing map is a map from the
; X0 X2 Xn = X
ROBIN FORMAN
as in (1.2), and with the
ells atta
hed so that their dimensions form a nonde
reasing
sequen
e.
I rst learned of simpli
ial
omplexes in an algebrai
topology
ourse. They were
introdu
ed as a
ategory of topologi
al spa
es for whi
h it was rather easy to dene
homology and
ohomology, i.e. in terms the simpli
al
hain- and
o
hain-
omplexes.
One might be
on
erned that in the transition from simpli
ial
omplexes to CW
omplexes we have lost this ability to easily
ompute the homology. In fa
t, mu
h of
this
omputability remains. Let X be a CW
omplex with a xed CW de
omposition.
Suppose that in this de
omposition X is
onstru
ted from exa
tly
d
ells of dimension
d for ea
h d = 0; 1; 2; : : : ; n = dim(K ), and let Cd (X; Z) denote the spa
e Z
d (more
pre
isely, Cd (X; Z) denotes the free abelian group generated by the d-
ells of X ). The
following is one of the fundamental results in the theory of CW
omplexes.
Theorem 1.6. There are boundary maps d : Cd (X; Z) ! Cd 1 (X; Z), for ea
h d, so
that
d
d = 0
n
! Cn(X; Z) !
Cn 1 (X; Z) n! ! C0 (X; Z) ! 0
1
Hd (C; )
= Hd (X; Z)
The a
tual denition of the boundary map is slightly nontrivial and we will not
go into it here (see Ch. V se
. 2 of [28 for the details). At rst it may seem that
without knowing this boundary map, there is little to be gained from Theorem 1.6.
In fa
t, mu
h
an be learned from just knowing of the existen
e of su
h a boundary
map. For example, let us
hoose a
oe
ient eld F , and tensor everything with F
to get a dierential
omplex
n
! Cn(X; F ) !
Cn 1 (X; F ) n! ! C0 (X; F ) ! 0
whi
h
al
ulates H (X; F ), where now Cp(X; F )
= F
d . From basi
linear algebra. we
(X ) = b0
b1 + b2
::::
(X ) = 0
1 + 2
::::
Theorem 1.8 (The Strong Morse Inequalities). With all notation as in Theorem 1.7,
for ea
h d = 0; 1; 2; : : :
d
d 1 + d
+(
1)d 0 bd
bd 1 + bd
+(
1)d b0 :
Comparing Strong Morse Inequalities for
onse
utive values of d yields Theorem 1.7.
We mentioned earlier that a great benet of passing from simpli
ial
omplexes to
the more general CW
omplexes is that one often
an use many fewer
ells. Let us
take another look at this phenomenon in light of the Morse inequalities. Consider the
10
ROBIN FORMAN
ase where X is a two-dimensional torus, so that with respe
t to any
oe
ient eld
b0 = 1; b1 = 2; b2 = 1. From the weak Morse inequalities, we have that for any CW
de
omposition,
0 b 0 = 1
1 b 1 = 2
2 b2 = 1:
A simpli
ial de
omposition is a spe
ial
ase of a CW de
omposition, so these inequalities are satised when
d denotes the number of d-simpli
es in a xed simpli
ial
de
omposition. However, every simpli
ial de
omposition has at least 7 0-simpli
es,
21 1-simpli
es and 14 2-simpli
es, so these inequalities are far from equality. It is
generally the
ase that for a simpli
ial de
omposition these inequalities are very far
from optimal, and hen
e are generally of little interest. On the other hand, earlier
we demonstrated a CW de
omposition of the two-torus with exa
tly one 0-
ell, two
1-
ells and one 2-
ell. The inequalities tell us, in parti
ular, that one
annot build a
two-torus using fewer
ells.
2.
The dis
ussion in the previous se
tion leads us to an important question. Suppose
one is given a nite simpli
ial
omplex X . Typi
ally, we
an expe
t that X has a CW
de
omposition with many fewer
ells than in the original simpli
ial de
omposition.
How
an one go about nding su
h an \e
ient" CW de
omposition for X ? In
this se
tion we present a te
hnique, dis
rete Morseh Theory, whi
h
an be useful
in su
h an investigation. (We note that the ideas we will des
ribe
an be applied
with no modi
ation at all to any nite regular CW
omplex, and with only minor
modi
ations to a general nite CW
omplex. However, for simpli
ity, in this paper
we will restri
t attention to simpli
ial
omplexes.)
We begin by re
alling that a nite simpli
ial
omplex is a nite set of verti
es V ,
along with a set of subsets K of V . The set K satises two main properties:
1) V K
11
2) If 2 K and then 2 K .
By a slight abuse of notation, we will refer to the simpli
ial
omplex simply as K .
The elements of K are
alled simpli
es. If 2 K , and
ontains p + 1 verti
es, then
we say that the dimension of is p, and we will sometimes denote this by (p) . For
simpli
es and we will use the notation < or > to indi
ate that is
a subset of (thinking of and as subsets of V ), and say that is a fa
e of .
We emphasize that at this point we will not be pla
ing any restri
tions on the nite
simpli
ial
omplexes under investigation. In parti
ular, the
omplexes need not be
manifolds (even though many of our examples will be). In se
tion 9 we will brie
y
indi
ate how some our
on
lusions
an be strengthened in the
ase that the
omplexes
are assumed to have additinal stru
ture.
A dis
rete Morse fun
tion on K is a fun
tion whi
h, roughly speaking, assigns
higher numbers to higher dimensional simpli
es, with at most one ex
eption, lo
ally,
at ea
h simplex. More pre
isely,
!R
is a dis
rete Morse fun
tion if for every (p) 2 K
(1) #f (p+1) > j f ( ) f ()g 1;
f :K
and
(2) #f (p
1)
< j f ( ) f ( ) g 1:
A simple example will serve to illustrate the denition. Consider the two
omplexes
shown in Figure 2.2. Here we indi
ate fun
tions by writing next to ea
h simplex the
value of the fun
tion on that simplex. The fun
tion (i) is not a dis
rete Morse fun
tion
as the edge f 1 (0) violates rule (2), sin
e it has 2 lower dimensional \neighbors" on
whi
h f takes on higher values, and the vertex f 1 (5) violates rule (1), sin
e it has 2
higher dimensional \neighbors" on whi
h f takes on lower values. The fun
tion (ii)
is a Morse fun
tion. Note that a dis
rete Morse fun
tion is not a
ontinuous fun
tion
on K . Rather, it is an assignment of a single number to ea
h simplex.
12
ROBIN FORMAN
5
4
0
(ii)
(i)
(i). This is not a dis rete Morse fun tion. (ii). This is a dis rete Morse fun tion.
Figure 2.2.
The other main ingredient in Morse Theory is the notion of a
riti
al point.
(2) #f (p
1)
< j f ( ) f ( ) g = 0:
For example, Figure 2.2(ii), the vertex f 1 (0) and the edge f 1 (5) are
riti
al, and
there are no other
riti
al simpli
es.
We mention for later use that it follows from the axioms that a simplex
annot
simultaneously fail both
onditions in the test for
riti
ality.
Lemma 2.4. If K is a simpli
ial
omplex with a Morse fun
tion f , then for any
simplex , either
(1) #f (p+1) > j f ( ) f ()g = 0;
or
(2) #f (p
1)
< j f ( ) f ( ) g = 0:
(See Lemma 2.5 of [10.) This lemma will play a
ru
ial role in Se
tion 3.
We
an now state the main theorem of dis
rete Morse Theory.
13
Theorem 2.5. Suppose M is a simpli
ial
omplex with a dis
rete Morse fun
tion.
Then M is homotopy equivalent to a CW
omplex with exa
tly one
ell of dimension p
for ea
h
riti
al simplex of dimension p.
Rather than present a proof of this theorem, we will
ontent ourselves here with
a brief dis
ussion of the main ideas. A dis
rete Morse fun
tion gives us a way to
build the simpli
ial
omplex by atta
hing the simpli
es in the order pres
ribed by the
fun
tion, i.e. adding rst the simpli
es whi
h are assigned the smallest values. More
pre
isely, for any simpli
ial
omplex K with a dis
rete Morse fun
tion f , and any real
number
, dene the level sub
omplex K (
) by
K (
) = [f ()
[ :
That is, K (
) is the sub
omplex
onsisting of all simpli
es of K su
h that f ()
along with all of their fa
es.
Theorem 2.5 follows from two basi
lemmas.
Lemma 2.6. If there are no
riti
al simpli
es with f () 2 (a; b, then M (b) is
homotopy equivalent to M (a). (In fa
t, M (b)
ollapses to M (a) - this will be explained
later.)
Lemma 2.7. If there is a single
riti
al simplex with f () 2 (a; b then there is a
map F : S (d 1) ! M (a), where d is the dimension of , su
h that M (b) is homotopy
equivalent to M (a) [F B d .
In Figure 2.8 we illustrate all of the level sub
omplexes in the
ase that K is the
ir
le triangulated with 3 edges and 3 verti
es, and f is the Morse fun
tion shown in
Figure 2.2 (ii). Here we
an see why these lemmas are true.
14
ROBIN FORMAN
5
2
1
0
K(0)
1
0
K(1)=K(2)
K(3)=K(4)
3
0
K(5)=K
The level sub omplexes of the dis rete Morse fun tion shown in Figure 2.2(ii)
Figure 2.8.
Let us begin with Lemma 2.6. Consider the transition from K (0) to K (1). We
have not added any
riti
al simpli
es, and, just as the lemma predi
ts, K (0) and K (1)
are homotopy equivalent. Let us try to understand why the homotopy type did not
hange. To
onstru
t K (1) from K (0), we rst have to add the edge f 1 (1). This
edge is not
riti
al be
ause it has a
odimension-one fa
e whi
h is assigned a higher
value, namely the vertex f 1 (2). In order to have K (1) be a sub
omplex, we must
also add this vertex. Thus we see that the edge f 1 (1) in K (1) has a free fa
e, i.e.
a fa
e whi
h not the fa
e of any other simplex in K (1). We
an deformation retra
t
K (1) to K (0) by "pushing in" the edge f 1 (1) starting at the vertex f 1 (2).
This is a very general phenomenon. That is, it follows from the axioms for a dis
rete
Morse fun
tion that for any simpli
ial
omplex with any dis
rete Morse fun
tion, when
passing from one level sub
omplex to the next the non
riti
al simpli
es are added in
pairs, ea
h of whi
h
onsists of a simplex and a free fa
e. Suppose that K2 K1 are
simpli
ial
omplexes, and K1 has exa
tly two simpli
es and that are not in K2 ,
where is a free fa
e of . Then it is easy to see that K2 is a deformation retra
t
of K1 , and hen
e K1 and K2 are homotopy equivalent (see Figure 2.9). This spe
ial
sort of
ombinatorial deformation retra
t is
alled a simpli
al
ollapse. If one
an
transform a simpli
ial
omplex K1 into a sub
omplex K2 by simpli
ial
ollapses, then
15
we say that K1
ollapses to K2 , and we indi
ate this by K1 & K2 . Figure 2.10 shows
a 2-dimensional simplex
ollapsing to one of its verti
es.
1
A simpli
ial
ollapse.
Figure 2.9.
Figure 2.10.
The pro
ess of simpli
ial
ollapse was studied by J.H.C. Whitehead, and he dened
simple homotopy equivalen
e to be the equivalen
e relation generated by simpli
ial
ollapse. This indi
ates that dis
rete Morse Theory may be parti
ularly useful when
working in the
ategory of simple homotopy equivalen
e.
Now let us turn to Lemma 2.7 and investigate what happens when one adds a
riti
al simplex, for example when making the transition from K (4) to K (5). In this
ase we are adding a
riti
al edge. We
an see
learly from the illustration that we
pass from K (4) to K (5) by atta
hing a 1-
ell, just as predi
ted by Lemma 2.7. To
see why this works in general,
onsider a
riti
al d-simplex . It follows from the
denition of a
riti
al simplex that ea
h fa
e of is assigned a smaller value than
16
ROBIN FORMAN
+(
1)n mn = b0 b1 + b2
+(
mp 1 + m0 bp
3.
bp 1 + b0 :
Any ambitious reader who has already started trying some examples will have
noti
ed that the theory as presented in the previous se
tion
an be a bit unwieldy.
After all, how is one to go about assigning numbers to ea
h of the simpli
es of a
simpli
ial
omplex so that they satisfy the axioms of a dis
rete Morse fun
tion?
17
Fortunately, in pra
ti
e one need not a
tually nd a dis
rete Morse fun
tion. Finding
the gradient ve
tor eld of the Morse fun
tion is su
ient. This requires a bit of
explanation.
Let us now return to the example in Figure 2.2(ii). Non
riti
al simpli
es o
ur in
pairs. For example, the edge f 1 (1) is not
riti
al be
ause it has a "lower dimensional
neighbor" whi
h is assigned a higher value, i.e. the vertex f 1 (2). Similarly, the vertex
f 1 (2) is not
riti
al be
ause it has a "higher dimensional neighbor" whi
h is assigned
a lower value, i.e. the edge f 1 (1). We indi
ate this pairing by drawing an arrow from
the vertex f 1 (2), pointing into the edge f 1 (1). Similarly, we draw an arrow from
the vertex f 1 (4) pointing into the edge f 1 (3). (See Figure 3.1.) One
an think of
these arrows as pi
torially indi
ating the simpli
ial
ollapse that is referred to in the
proof of Lemma 2.6.
5
2
3
0
The gradient ve tor eld of the Morse fun tion shown in Figure 2.2 (ii).
Figure 3.1.
We
an apply this pro
ess to any simpli
ial
omplex with a dis
rete Morse fun
tion.
The arrows are drawn as follows. Suppose (p) is a non-
riti
al simplex with (p+1) >
satisfying f ( ) f (). We then draw an arrow from to . Figure 3.2 illustrates
a more
ompli
ated example. Note that the dis
rete Morse fun
tion drawn in this
gure has one
riti
al vertex, f 1 (0), and one
riti
al edge, f 1 (11). Theorem 2.5
18
ROBIN FORMAN
implies this simpli
ial
omplex is homotopy equivalent to a CW
omplex with exa
tly
one 0-
ell and one 1-
ell, i.e. a
ir
le.
It follows from Lemma 2.4 that that every simplex satises exa
tly one of the
following:
(i) is the tail of exa
tly one arrow.
(ii) is the head of exa
tly one arrow.
(iii) is neither the head nor the tail of an arrow.
Note that a simplex is
riti
al if and only if it is neither the tail nor the head of any
arrow. These arrows
an be viewed as the dis
rete analogue of the gradient ve
tor
eld of the Morse fun
tion. (To be pre
ise, when we say \gradient ve
tor eld" we
are really referring to the negative of the gradient ve
tor eld.)
6
7
10
14
16 17
5
12 13
3
11
4
(i)
15
1 2
(ii)
Figure 3.2.
As we will see in examples later, these arrows are mu
h easier to work with than
the original dis
rete Morse fun
tion. In fa
t, this gradient ve
tor eld
ontains all of
the information that we will need to know about the fun
tion for most appli
ations.
The upshot is that if one is given a simpli
ial
omplex and one wishes to apply the
theory of the previous se
tion, one need not
onstru
t nd a dis
rete Morse fun
tion.
One \only" needs to nd a gradient ve
tor eld.
This leads us to the following question. Suppose we atta
h arrows to the simpli
es
so that ea
h simplex saties exa
tly one of properties (i),(ii),(iii) above. Then how do
19
we know if that set of arrows is the gradient ve
tor eld of a dis
ret Morse fun
tion?
This is the question we will answer in the remainder of this se
tion.
Let K be a simpli
ial
omplex with a dis
rete Morse fun
tion f . Then rather than
thinking about the dis
rete gradient ve
tor eld V of f as a
olle
tion of arrows,
we may equivalently des
ribe V as a
olle
tion of pairs f(p) < (p+1) g of simpli
es
of K , where f(p) < (p+1) g is in V if and only if f ( ) f (). In other words,
f(p) < (p+1) g is in V if and only if we have drawn an arrow that has as its
tail, and as its head. The properties of a dis
rete Morse fun
tion imply that ea
h
simplex is in at most one pair of V . This leads us to the following denition.
Denition 3.3. A dis
rete ve
tor eld V on K is a
olle
tion of pairs f(p) < (p+1) g
of simpli
es of K su
h that ea
h simplex is in at most one pair of V .
Su
h pairings were studied in [41 and [8 as a tool for investigating the possible
f -ve
tors for a simpli
ial
ompex. Here we take a dierent point of view. If one has a
smooth ve
tor eld on a smooth manifold, it is quite natural to study the dynami
al
system indu
ed by
owing along the ve
tor eld. One
an begin the same sort of study
for any dis
rete ve
tor eld. In [12 we present a study of the dynami
s asso
iated
to a dis
rete ve
tor eld. Here, we present just enough to
ontinue our dis
ussion of
dis
rete Morse Theory.
Given a dis
rete ve
tor eld V on a simpli
ial
omplex K , a V path is a sequen
e
of simpli
es
(3.1)
Theorem 3.4. Suppose V is the gradient ve
tor eld of a dis
rete Morse fun
tion f .
Then a sequen
e of simpli
es as in (3.1) is a V -path if and only if i < i > i+1 for
20
ROBIN FORMAN
ea h i = 0; 1; : : : ; r, and
Theorem 3.5. A dis
rete ve
tor eld V is the gradient ve
tor eld of a dis
rete
Morse fun
tion if and only if there are no non-trivial
losed V -paths.
We will not prove this theorem here. However, many readers may noti
e the similarity with the following standard theorem from the subje
t of dire
ted graphs.
Theorem 3.6. Let G be a dire
ted graph. Then there is a real-valued fun
tion of the
verti
es that is stri
tly de
reasing along ea
h dire
ted path if and only if there are no
dire
ted loops.
We will show in se
tion 6 that, in fa
t, Theorem 3.6 implies Theorem 3.5. The
power of Theorem 3.5 is indi
ated in the next two se
tions in whi
h we
onstru
t
some dis
rete ve
tor elds and use Theorem 3.5 to verify that they are gradient
ve
tor elds.
4.
Figure 4.1 (i) shows a triangulation of the real proje
tive plane P2 . Note that the
verti
es along the boundary with the same labels are to be identied, as are the edges
whose endpoints have the same labels. In Figure 5(ii) we illustrate a dis
rete ve
tor
eld V on this simpli
ial
omplex. One
an easily see that there are no
losed V -paths
(sin
e all V -paths go to the boundary of the gure and there are no
losed V -paths
on the boundary), and hen
e is a gradient ve
tor eld. The only simpli
es whi
h are
neither the head nor the tail of an arrow are the vertex labelled 1, the edge e, and
the triangle t. Thus, by Theorem 2.5, the proje
tive plane is homotopy equivalent
21
to a CW
omplex with exa
tly one 0-
ell, one 1-
ell and one 2-
ell. (Of
ourse, we
already knew this from our dis
ussion of example 3 in se
tion 2.)
3
e
2
1
t
3
(i)
(ii)
(i) A triangulation of the real proje tive plane. (ii) A dis rete gradient ve tor eld on P2 .
Figure 4.1.
This example gives rise to two potential
on
erns. The rst is that from the main
theorem we learn only a statement about "homotopy equivalen
e". This is su
ient
if one is only interested in
al
ulating homology or homotopy groups. However, one
might be interested in determining the (PL-)homeomorphism type of the
omplex.
This is possible, in some
ases, using deep results of J.H.C. Whitehead. We revisit
this topi
in se
tion 8.
The se
ond potential point of
on
ern is that as we saw in se
tion 2 there are an
innite number of dierent homotopy types of CW
omplexes whi
h
an be built from
exa
tly one 0-
ell, one 1-
ell and one 2-
ell. One might wonder if Morse Theory
an
give us any additional information as to how the
ells are atta
hed. In fa
t, one
an
dedu
e mu
h of this information if one has enough information about the gradient
paths of the Morse fun
tion. This point is dis
ussed further in se
tion 7, where we
will return to this example of the triangulated proje
tive plane.
22
ROBIN FORMAN
5.
A number of fas
inating simpli
ial
omplexes arise from the study of monotone
graph properties. Let Kn denote the
omplete graph on n verti
es, and suppose we
have labelled the verti
es 1,2,. . . ,n. Let Gn denote the spanning subgraphs of Kn , that
is, the subgraphs of Kn that
ontain all n verti
es. A subset P Gn is
alled a graph
property of graphs with n verti
es if in
lusion in P only depends on the isomorphism
type of the graph. That is, P is a graph property if for all pairs of graphs G1 ; G2 2 Gn ,
if G1 and G2 are isomorphi
(ignoring the labellings on the verti
es) then G1 2 P
if and only if G2 2 P . A graph property P of graphs with n verti
es is said to be
monotone de
reasing if for any graphs G1 G2 2 Gn , if G2 2 P then G1 2 P .
Monotone de
reasing properties abound in the study of graph theory. Here are
some typi
al examples: graphs having no more than k edges (for any xed k), graphs
su
h that the degree of every vertex is less that (for any xed ), graphs whi
h are
not
onne
ted, graphs whi
h are not i-
onne
ted (for any xed i), graphs whi
h do
not have a Hamiltonian
y
le, graphs whi
h do not
ontain a minor isomorphi
to H
(for any xed graph H ), graphs whi
h are r-
olorable (for any xed r), and bipartite
graphs.
Any monotone de
reasing graph property P gives rise to a simpli
ial
omplex K
where the d-simpli
es of K are the graphs G 2 P whi
h have d +1 edges. In parti
ular,
if G is a d-simplex in K, then the fa
es of G are all of the nontrivial spanning subgraphs
of G (the monotoni
ity of P implies that ea
h of these graphs is in K). Said in another
way, if P is nonempty, then the verti
es of K are the edges of Kn , and a
olle
tion of
verti
es in K span a simplex if the spanning subgraph of Kn
onsisting of all edges
whi
h
orrespond to these verti
es lies in P .
The simpli
ial
omplexes indu
ed by many of the above-mentioned monotone de
reasing graph properties have been studied using the te
hniques of this paper. See
for example [6, [7,[21,[22,[27,[37. These papers
ontain some beautiful mathemati
s in whi
h the authors
onstru
t, "by hand", expli
it dis
rete gradient ve
tor
23
elds, along the way illuminating some of the intri
ate ner stru
tures of the graph
properties.
Some monotone graph properties have re
ently been the fo
us of intense interest
be
ause of their relation to knot theory. Unfortunately this is probably not a good
time for an in depth dis
ussion of this fas
inating topi
. We will mention only that
Vassiliev has shown how one
an derive "nite type knot invariants" from the study
of the spa
e of \singular knots" (i.e. maps from S 1 to R 3 whi
h are not embeddings).
The homology of the simpli
ial
omplexes of not
onne
ted and not 2-
onne
ted
graphs show up in his spe
tral sequen
e
al
ulation of the homology of this spa
e.
This is explained in [43, where Vassiliev derives the homotopy type of the
omplex
of not
onne
ted graphs. In [42 and [1, the topology of the spa
e of not 2-
onne
ted
graphs is determined, with dis
rete Morse Theory playing a minor role in the latter
referen
e. This topi
is reexamined in [37, in whi
h the entire investigation is framed
in the language of dis
rete Morse Theory. Dis
rete Morse Theory is used to determine
the topology of not 3-
onne
ted graphs in [21.
In this se
tion, we will provide an introdu
tion to this work by taking a look at the
simpler
ase of the
omplex of not
onne
ted graphs. We will show how the ideas of
this paper may be used to determine the topology of Nn , the simpli
ial
ompex of
not
onne
ted graphs on n verti
es. Let me begin by pointing out that this
omplex
an be well studied by more
lassi
al methods, and the answer has also been found by
Vassiliev in [43. The only novelty of this se
tion is our use of dis
rete Morse Theory.
Our goal is to
onstru
t a dis
rete gradient ve
tor eld V on Nn , the simpli
ial
omplex of all not-
onne
ted graphs with the vertex set f1; 2; 3; : : : ; ng. The
onstru
tion will be in steps. Let V12 denote the dis
rete ve
tor eld
onsisting of all
pairs fG; G + (1; 2)g, where G is any graph in Nn whi
h does not
ontain the edge
(1; 2) and su
h that G + (1; 2) 2 Nn. Another way of des
ribing V12 is that if G is
any graph in Nn whi
h
ontains the edge (1; 2), then G (1; 2) and G are paired in
V12 . A
tually, there is one ex
eption to this rule. Let G denote the graph
onsisting
of only the single edge (1,2). Then G (1; 2) is the empty graph, whi
h
orresponds
24
ROBIN FORMAN
to the empty simplex in Nn, and may not be paired in a dis
rete ve
tor eld. Thus,
G is unpaired in V12 .
The graphs in Nn other than G whi
h are unpaired in V12 are those that do
not
ontain the edge (1; 2) and have the property that G + (1; 2) 62 Nn . That is,
those dis
onne
ted graphs G with the property that G + (1; 2) is
onne
ted. Su
h a
graph must have exa
tly two
onne
ted
omponents, one of whi
h
ontains the vertex
labelled 1, and one whi
h
ontains the vertex labelled 2. We denote these
onne
ted
omponents by G1 and G2 , resp. See Figure 5.1.
a
connected
graph
G1
a
connected
graph
G2
The graphs other than G whi h are unpaired in the ve tor eld V12 .
Figure 5.1.
Let G be a graph other than G whi
h is unpaired in V12 , and
onsider vertex 3.
This vertex must either be in G1 or G2 . Suppose that vertex 3 is in G1 . If G does
not
ontain the edge (1; 3) then G + (1; 3) is also unpaired in V12 , so we
an pair G
with G + (1; 3). If vertex 3 is in G1 , then the graph G is still unpaired if and only if
G
ontains the edge (1,3) and G (1; 3) is the union of three
onne
ted
omponents,
one
ontaining vertex 1, one
ontaining vertex 2, and one
ontaining vertex 3.
25
Similarly, if vertex 3 is in G2 and G does not
ontain the edge (2; 3), then pair G
with G + (2; 3). Let V3 denote the resulting dis
rete ve
tor eld.
The unpaired graphs in V3 are G and those that either
ontain the edge (1,3) and
have the property that G (1; 3) is the union of three
onne
ted
omponents, one
ontaining vertex 1, one
ontaining vertex 2, and one
ontaining vertex 3, or
ontain
the edge (2,3) and have the property that G (2; 3) is the union of three
onne
ted
omponents, one
ontaining vertex 1, one
ontaining vertex 2, and one
ontaining
vertex 3. We illustrate these graphs in Figure 5.2. The
ir
les in this gure indi
ate
onne
ted graphs.
The graphs other than G whi h are unpaired in the ve tor eld V3 .
Figure 5.2.
Now
onsider the lo
ation of the vertex labelled 4, and pair any graph G whi
h is
unpaired in V3 with G + (1; 4),G + (2; 4), or G + (3; 4) if possible (at most one of these
graphs is unpaired in V3 ). Call the resulting dis
rete ve
tor eld V4 . We
ontinue
in this fashion,
onsidering in turn the verti
es labelled 5; 6; : : : ; n. Let Vi denote
the dis
rete ve
tor eld that has been
onstru
ted after the
onsideration of vertex
i, and V = Vn the nal dis
rete ve
tor eld. When we are done the only unpaired
26
ROBIN FORMAN
graphs in V will be G and those graphs that are the union of two
onne
ted trees,
one
ontaining the vertex 1 and one
ontaining the vertex 2. In addition, both trees
have the property that the vertex labels are in
reasing along every ray starting from
the vertex 1 or the vertex 2. There are pre
isely (n 1)! su
h graphs, and they ea
h
have n 2 edges, and hen
e
orrespond to an (n 3)-simplex in Nn .
It remains to see that the dis
rete ve
tor eld V is a gradient ve
tor eld, i.e. that
there are no
losed V -paths. We rst
he
k that V12 is a gradient ve
tor eld. Let
= 0(p) ; 0(p+1) ; 1(p) denote a V12 -path. Then 0 must be the "tail of an arrow",
i.e. the smaller graph of some pair in V12 , with 0 being the head of the arrow, i.e.
0 = 0 + (1; 2). The simplex 1 is a
odimension-one fa
e of 0 other than 0 . Thus,
1
orresponds to a graph of the form 0 + (1; 2) e, where e is an edge of 0 other
than (1,2). Sin
e 1
ontains the edge (1; 2) it is the "head of an arrow" in V12 , i.e.
the larger graph of some pair in V12 , whi
h implies that
annot be
ontinued to a
longer V12 -path. This
ertainly implies that there are no
losed V12 -paths.
The same sort of argument will work for V . Re
all that V is
onstru
ted in stages,
by rst
onsidering the edge (1,2) and then the verti
es 3,4,5,. . . in order. Let
=
0 ; 0 ; 1 denote a V -path. In parti
ular, 0 and 0 must be paired in V . The reader
an
he
k that if 0 and 0 are rst paired in Vi , i 3, then either 1 is the head
of an arrow in Vi , in whi
h
ase the V -path
annot be
ontinued, or 1 is paired in
Vi 1 . It follows by indu
tion that there
an be no
losed V -paths.
In summary, V is a dis
rete gradient ve
tor eld on Nn with exa
tly one unpaired
vertex, and (n 1)! unpaired (n 3)-simpli
es. We
an now apply Theorem 2.5 to
on
lude
Theorem 5.3 ([43). The
omplex Nn of not
onne
ted graphs on n-verti
es is homotopy equivalent to the wedge of (n 1)! spheres of dimension (n 3).
6.
The notion of a gradient ve
tor eld has a very ni
e purely
ombinatorial des
ription
due to Chari [6, using whi
h we
an re
ast the Morse Theory in an appealing form.
27
We begin with the Hasse diagram of M , that is, the partially ordered set of simpli
es
of M ordered by the fa
e relation. Consider the Hasse diagram as a dire
ted graph.
The verti
es of the graph are in 1-1
orresponden
e with the simpli
es of M , and
there is a dire
ted edge from to if and only if is a
odimension-one fa
e of
. (See Figure 6.1 (i).) Now let V be a
ombinatorial ve
tor eld. We modify the
dire
ted graph as follows. If f < g 2 V then reverse the orientation of the edge
between and , so that it now goes from to . (See Figure 6.1(ii).) A V -path
an be thought of as a dire
ted path in this modied graph. There are some dire
ted
paths in this modied Hasse diagram whi
h are not V -paths as we have dened them.
However, the following result is not hard to
he
k.
t
v1
e3
v3
e1
e2
e1
e2
e3
v1
v2
v3
(i)
v1
e3
v3
t
e2
t
e1
v2
e1
e2
e3
v1
v2
v3
(ii)
Figure 6.1.
Theorem 6.2. There are no nontrivial
losed V -paths if and only if there are no
nontrivial
losed dire
ted paths in the
orresponding dire
ted Hasse diagram.
28
ROBIN FORMAN
Thus, in this
ombinatorial language, a dis
rete ve
tor eld is a partial mat
hing of
the Hasse diagram, and a dis
rete ve
tor eld is a gradient ve
tor eld if the partial
mat
hing is a
y
li
in the above sense. Note that using Theorem 6.2, we
an see that
Theorem 3.5 does follow from Theorem 3.6.
We
an now restate some of our earlier theorems in this language. There is a very
minor
ompli
ation in that one usually in
ludes the empty set as an element of the
Hasse diagram (
onsidered as a simplex of dimension -1) while we have not
onsidered
the empty set previously.
Theorem 6.3. Let V be an a
y
li
partial mat
hing of the Hasse diagram of K (of the
sort des
ribed above - assume that the empty set is not paired with another simplex).
Let up denote the number of unpaired p-simpli
es. Then M is homotopy equivalent to
a CW-
omplex with exa
tly up
ells of dimension p, for ea
h p 0.
An important spe
ial
ase is when V is a
omplete mat
hing, that is, every simplex
(this time in
luding the empty simplex) is paired with another simplex. In this
ase,
Lemma 2.9 implies the following result.
In this se
tion we will see how knowledge of the gradient paths of a dis
rete Morse
fun
tion on a spa
e K
an allow one to strengthen the
on
lusions of the main theorems. In parti
ular, rather than just knowing the number of
ells in a CW deomposition for K , one
an
al
ulate the homology exa
tly.
Let K be a simpli
ial
omplex with a Morse fun
tion f . Let Cp(X; Z) denote the
spa
e of p-simpli
ial
hains, and Mp Cp (X; Z) the span of the
riti
al p-simpli
es.
We refer to M as the spa
e of Morse
hains. If we let mp denote the number of
29
Mp = Zmp :
Sin
e homotopy equivalent spa
es have isomorphi
homology, the following theorem
follows from Theorems 2.5 and 1.6.
Theorem 7.1. There are boundary maps ~d : Mp ! Md 1 , for ea
h d, so that
~d
~d = 0
n
Mn
! Mn ~!
~n
! ~! M0 ! 0
1
30
ROBIN FORMAN
the gradient path shown in Figure 7.2, the indi
ated orientation on indu
es the
indi
ated orientation on .
Figure 7.2.
We are now ready to state the desired formula.
Theorem 7.3. Choose an orientation for ea
h simplex. Then for any
riti
al (p +1)simplex set
(7.1)
=
riti al
where
; =
;
(p)
2 (;)
m( )
A proof this theorem appears in se
tion 8 of [10. We refer to this
omplex (7.1)
with the dierential (7.1) as the Morse
omplex (it goes by many dierent names in
the litereature). An extensive study of the Morse
omplex in the smooth
ategory
appears in [36
We end this se
tion with a demonstration of how the ideas of this se
tion may
be applied to the example of the real proje
tive plane P2 as illustrated in Figure
4.1(ii). We saw in se
tion 2 how dis
rete Morse Theory
an help us see that P2
has a CW de
omposition with exa
tly one 0-
ell, one 1-
ell and one 2-
ell. Here we
31
will see how Morse Theory
an distinguish between the spa
es whi
h have su
h a
CW de
omposition. In Figure 7.4 we redraw the gradient ve
tor eld, and indi
ate a
hosen orientation on the
ritial edge e and the
riti
al triangle t. Let us now
al
ulate
the boundary map in the Morse
omplex. To
al
ulate ~(e), we must
ount all of
the gradient paths from the boundary of e to v . There are pre
isely two su
h paths.
Namely, following the unique gradient path beginning at ea
h endpoint of e leads us
to v . (The gradient path beginning at the head of e is the trivial path of 0 steps.)
Sin
e the orientation of e indu
es a + orientation on the head of e, and a - orientation
on the tail of e, adding these two paths with their
orresponding signs leads us to the
formula that ~(e) = 0. It
an be seen from the illustration that there are pre
isely two
gradient paths from the boundary of t to e, and, with the illustrated orientation for t,
both indu
e the
hosen orientation on e, so that ~(t) = 2e. Therefore the homology
of the real proje
tive plane
an be
al
ulated from the following dierential
omplex.
2
0
!Z !
Z ! 0:
H0 (P2 ; Z)
= 0:
= Z=2Z; H2 (P2 ; Z)
= Z; H1 (P2 ; Z)
3
e
2
1
t
Figure 7.4.
32
ROBIN FORMAN
8.
Sphere Theorems
As mentioned in our dis
ussion at the end of se
tion 4, one
an sometimes use
dis
rete Morse Theory to make statements about more than just the homotopy type
of the simpli
ial
omplex. One
an sometimes
lassify the
omplex up to homeomorphism or
ombinatorial equivalen
e. This will be a very short se
tion, as this topi
seems a bit far from the main thrust of this paper. In addition, some terms will
unfortunately have to be dened only
ursorily or not at all. So far, we have not
pla
ed any restri
tions on the simpli
ial
omplexes under
onsideration. The main
idea of this se
tion is that is if our simpli
ial
omplex has some additional stru
ture,
then one may be able to strengthen the
on
lusion. This idea rests on some very deep
work of J.H.C. Whitehead [44.
Re
all that a simpli
ial
omplex K is a
ombinatorial d-ball if K and the standard
d-simplex d have isomorphi
subdivisions. A simpli
ial
omplex K is a
ombinatorial (d 1)-sphere, if K and _d have isomorphi
subdivisions (where _d denotes the
boundary of d with its indu
ed simpli
ial stru
ture). A simpli
ial
omplex K is a
ombinatorial d-manifold with boundary if the link of every vertex is either a
ombinatorial (d 1)-sphere or a
ombinatorial (d 1)-ball. The following is a spe
ial
ase
of the main theorem of [44.
Theorem 8.1. Let K be a
ombinatorial d-manifold with boundary whi
h simpli
iallly
ollapses to a vertex. Then K is a
ombinatorial d-ball.
It is with this theorem (and its generalizations) that one
an strengthen the
on
lusion of Theorem 2.5 beyond homotopy equivalen
e. We present just one example.
Theorem 8.2. Let X be a
ombinatorial d-manifold with a dis
rete Morse fun
tion
with exa
tly two
riti
al simpli
es. Then X is a
ombinatoral d-sphere.
The proof is quite simple (given Theorem 8.1). If X is a
ombinatorial d-manifold
with a dis
rete Morse fun
tion f with exa
tly two
riti
al simpli
es, then the
riti
al simpli
es must be the minimum of f , whi
h must o
ur at a vertex v , and the
maximum of f , whi
h must o
ur at a d-simplex . Then X is a
ombinatorial
33
d-manifold with boundary with a dis
rete Morse fun
tion with only a single
riti
al
simplex, namely the vertex v . It follows from Lemma 2.6 that X
ollapses to v .
Whitehead's theorem now implies that X is a
ombinatorial d-ball, whi
h implies
that X is a
ombinatorial d-sphere.
9.
One of the main problems in Morse Theory, whether in the
ombinatorial or smooth
setting, is to nd a Morse fun
tion for a given spa
e with the fewest possible
riti
al
points (mu
h of the book [38 is devoted to this topi
). In general this is a very
di
ult problem, sin
e, in parti
ular, it
ontains the Poin
are
onje
ture { spheres
an be re
ognized as those spa
es whi
h have a Morse fun
tion with pre
isely 2
riti
al
points. In [31, Milnor presents Smale's proof [40 of the higher dimensional Poin
are
onje
ture (in fa
t, a proof is presented of the more general h-
obordism theorem)
ompletely in the language of Morse Theory. Drasti
ally oversimplifying matters,
the proof of the higher Poin
are
onje
ture
an be des
ribed as follows. Let M be a
smooth manifold of dimension 5 whi
h is homotopy equivalent to a sphere. Endow
M with a (smooth) Morse fun
tion f . One then pro
eeds to show that the
riti
al
points of f
an be
an
elled out in pairs until one is left with a Morse fun
tion with
exa
tly two
riti
al points, whi
h implies that M is a (topologi
al) sphere.
A key step in this proof is the "
an
ellation theorem" whi
h provides a su
ient
ondition for two
riti
al points to be
an
elled (see Theorem 5.4 in [31, whi
h Milnor
alls "The First Can
ellation Theorem", or the original proof in [33). In this se
tion
we will see that the analogous theorem holds for dis
rete Morse fun
tions. Moreover,
in the
ombinatorial setting the proof is mu
h simpler. The main result is that if (p)
and (p+1) are 2
riti
al simpli
es, and if there is exa
tly 1 gradient path from the
boundary of to , then and
an be
an
elled. More pre
isely,
Theorem 9.1. Suppose f is a dis
rete Morse fun
tion on M su
h that (p+1) and
(p) are
riti
al, and there is exa
tly one gradient path from the boundary to .
Then there is another Morse fun
tion g on M with the same
riti
al simpli
es ex
ept
34
ROBIN FORMAN
that and are no longer
riti
al. Moreover, the gradient ve
tor eld asso
iated to
g is equal to the gradient ve
tor eld asso
iated to f ex
ept along the unique gradient
path from the boundary to .
In the smooth
ase, the proof, either as presented originally by Morse in [33 or as
presented in [31, is rather te
hni
al. In our dis
rete
ase the proof is simple. If, in
the top drawing in Figure 9.2, the indi
ated gradient path is the only gradient path
from the boundary of to , then we
an reverse the gradient ve
tor eld along this
path, repla
ing the gure by the ve
tor eld shown in the botton drawing in Figure
9.2.
Figure 9.2.
The uniqueness of the gradient path implies that the resulting dis
rete ve
tor eld
has no
losed orbits, and hen
e, by Theorem 3.5, is the gradient ve
tor eld of some
Morse fun
tion. Moreover, and are not
riti
al for this new Morse fun
tion, while
the
riti
ality of all other simpli
es is un
hanged. This
ompletes the proof.
35
The proof in the smooth
ase pro
eeds along the same lines. However, in addition
to turning around those ve
tors along the unique gradient path from to , one must
also adjust all nearby ve
tors so that the resulting ve
tor eld is smooth. Moreover,
one must
he
k that the new ve
tor eld is the gradient of a fun
tion, so that, in
parti
ular, modifying the ve
tors did not result in the
reation of a
losed orbit. This
is an example of the sort of
ompli
ations whi
h arise in the smooth setting, but
whi
h do not make an appearan
e in the dis
rete theory.
This theorem was re
ently put to very good use in [2, in whi
h dis
rete Morse
Theory is used to determine the homotopy type of some simpli
ial
omplexes arising
in the study of partitions. It is fas
inating, and quite pleasing, to see the same idea
play a
entral role in two subje
ts, the Poin
are
onje
ture and the study of partitions,
whi
h seem to have so little to do with one another.
10.
So far, we have indi
ated some appli
ations of dis
rete Morse Theory to
ombinatori
s and topology. We now present an appli
ation to
omputer s
ien
e. The reader
should see the referen
e [14 for a more
omplete treatement of the
ontent of this
se
tion.
The problem we study is a topologi
al version of a standard type of \sear
h problem". The generalized version that we will present rst appeared in [35. Let S be an
n-dimensional simplex, with verti
es v0 ; v1 ; : : : ; vn , and K a sub
omplex of S whi
h
is known to you. Let be a fa
e of S whi
h is not known to you. Your goal is to
determine if is in K . In parti
ular, you need not determine the fa
e , just whether
or not it is in K . You are permitted to ask questions of the form \Is vi in ?". You
may use the answers to the questions you have already asked in determining whi
h
vertex to ask about next. Of
ourse, you
an determine if is in K by asking n + 1
questions, sin
e by asking about all n + 1 verti
es you
an
ompletely determine .
You win this game if you answer the given question after asking fewer that n + 1
questions.
36
ROBIN FORMAN
Say that K is nonevasive if there is a winning strategy for this game, i.e there
is a question algorithm that determines whether or not 2 K in fewer than n + 1
questions, no matter what is. Say K is evasive otherwise.
Kahn, Saks and Sturtevant proved the following relationship between the evasiveness of K and its algebrai
topology.
Theorem 10.1. If H~ (K ) 6= 0, where H~ (K ) denotes the redu
ed homology of K ,
then K is evasive.
In fa
t, they proved something stronger, and we will
ome ba
k to this point later.
We illustrate the previous theorem with a simple example. Let S be the 2-simplex
shown in Figure 10.2, spanned by the verti
es v0 ; v1 and v2 , with K the sub
omplex
onsisting of the edge [v0 ; v1 together with the vertex v2 .
v2
v0
v1
Figure 10.2.
A possible guessing algorithm is shown in Figure 10.3. Dene an evader of a
guessing algorithm to be a fa
e of S with the property that when questions are asked
in the order determined by the algorithm one must ask all three questions before it is
known whether or not is in K . In parti
ular, the evaders of the illustrated guessing
algorithm are:
= [v2 ; [v0 ; v2
37
Note that the sub
omplex K has nonzero redu
ed homology, so the theorem of Kahn,
Saks and Sturtevant guarantees that every guessing algorithm has some evaders.
v2
no
yes
v0
no
v1
no
[]
yes no
v1
yes
no
v1
v0
yes
yes
v0
no
yes
no
yes
A guessing algorithm
Figure 10.3.
Morse Theory
omes to the fore when one observes that a guessing algorithm
indu
es a dis
rete ve
tor eld on S . For example, the guessing algorithm shown in
Figure 10.3 indu
es the ve
tor eld
V=
That is, V
onsists of those pairs of fa
es of S whi
h are not distinguished by the
guessing algorithm until the last question. There is slight subtlety here in that a
guessing algoithm pairs a vertex with the empty simplex ;, while in our original
denition, it was not permitted to pair a simplex with ;. Thus, to get a true dis
rete
ve
tor eld, we must remove this pair from V . (It is pre
isely this subtle point
that results in the redu
ed homology of K being the relevant measure of topologi
al
omplexity, rather than the nonredu
ed homology.) However, for simpli
ity, from now
on we will simply ignore this te
hni
al point.
38
ROBIN FORMAN
v2
v0
v1
The ve tor eld indu ed by the guessing algorithm shown in Figure 10.3.
Figure 10.4.
Theorem 10.5. This indu
ed ve
tor eld is always a gradient ve
tor eld.
We will postpone the proof of this result until the end of this se
tion.
Now restri
t V to K (by taking only those pairs in V s.t. both simpli
es are in K ).
For example, in our example, this results in the ve
tor eld
39
2 dim H(K )
Suppose that K is nonevasive. Then there is some guessing algorithm whi
h has no
evaders. From our above dis
ussion we seen that this implies that K has a gradient
ve
tor eld with no
riti
al simpli
es. A
tually, this is not quite true. The gradient
ve
tor eld must have a
riti
al vertex - the vertex that is paired with the empty set
- this is that minor te
hni
ality that we have been ignoring. Applying Lemma 2.6
yields the following strengthening of Theorem 10.1.
40
ROBIN FORMAN
Further Thoughts
We
lose this paper with some additional thoughts on the subje
ts dis
ussed in this
paper.
I would like to begin by en
ouraging the reader to take a look at the papers [24,[25,
and [3. In these papers, dis
rete Morse Theory is used to investigate quite interesting
problems. These referen
es were not mentioned earlier only be
ause they did not
easily t into any of the previous se
tions of this paper.
There are a number of dire
tions in whi
h dis
rete Morse Theory
an be extended
and generalized. Here we mention a few su
h possibilities. In [16 we show how one
an re
over the ring stru
ture of the
ohomology of a simpli
ial
omplex from the
point of view of dis
rete Morse Theory (this follows work of Betz and Cohen [4 and
Fukaya [17[18 in the smooth setting). In [34, Novikov presents a generalization
of standard smooth Morse Theory in whi
h the role of the Morse fun
tion is now
played by a
losed 1-form (the
lassi
al
ase arises when the
losed 1-form is exa
t).
In [15 we present the analogous generalization for dis
rete Morse Theory. In [45,
Witten shows how smooth Morse Theory
an be seen as arising from
onsiderations
of supersymmetry in quantum physi
s. In [11 we present a
ombinatorial version of
Witten's derivation. We believe that this latter work may have greater signi
an
e.
At
ru
ial points in [45, Witten appeals to path integral arguments whi
h are rather
standard in quantum physi
s, but are ill-dened mathemati
ally. In the
orresponding
moments in [11 what arises is a well-dened dis
rete sum. Perhaps the approa
h in
[11
an nd uses in the analysis of other quantum eld theories.
One topi
whi
h we have only tou
hed upon is the study of the dynami
s asso
iated
to
owing along the gradient ve
tor eld asso
iated to a dis
rete Morse fun
tion. In
fa
t, an understanding of the dynami
s is
ru
ial to the proof of theorem 7.3, for
41
example. The relevant study takes pla
e in se
tion 6 of [10. In [12 we study the
dynami
al properties of the
ow asso
iated to a general dis
rete ve
tor eld.
One area in whi
h mu
h work remains to be done is the investigation of dis
rete
Morse Theory for innite simpli
ial
omplexes. The theory as desribed in this paper
an be applied without
hange to an innite simpli
ial
omplex K endowed with a
dis
rete Morse fun
tion f whi
h is proper, i.e. one in whi
h for ea
h real number
the level sub
omplex K (
) is a nite
omplex. Unfortunately, properness is often an
unnatural requirement when
onsidering the innite simpli
ial
omplexes whi
h arise
in pra
ti
e. In the interesting paper [29, dis
rete Morse Theory is applied to the
investigation of innite simpli
ial
omplexes K whi
h arise as a
overing spa
e of a
nite simpli
ial
omplex K 0 . In this
ase, the authors restri
t attention to dis
rete
Morse fun
tions whi
h are lifts of a Morse fun
tion on K 0 , and
ompare the number
of
riti
al simpli
es to the L2 -Betti numbers of K . While it appears to be too mu
h
to hope that one
an develop a useful theory that applies to all innite simpli
ial
omplexes with no restri
tions on the dis
rete Morse fun
tion, it seems likely that
there is room for very useful investigations of large
lasses of
omplexes and fun
tions
with restri
tions dierent than those already
onsidered.
I will
lose these notes with some
omments of a less rigorous nature. Whether
in the smooth
ategory or the
ombinatorial
ategory, Morse Theory is not essential
to any problem, it is usually "only" a
onvenient and e
ient language. Anything
that
an be done with Morse Theory
an be done without it. It seems to me that
Morse Theory takes on a spe
ial signi
an
e in three dierent
ases. First are the
ases in whi
h Morse Theory is not intrinsi
to the problem, but where the existen
e
of su
h an e
ient language may make the dieren
e between whether or not one
is able to see the way to the end of a problem. The best example of this in the
smooth setting, I think, is the proof of the higher dimensional Poin
are
onje
ture
([39,[31). Most of the appli
ations of dis
rete Morse Theory mentioned in se
tion
5, for example, seem to fall into this
ategory. Se
ond are the
ases in whi
h the
spa
e one is studying
omes naturally endowed with a Morse fun
tion, or a gradient
42
ROBIN FORMAN
ve
tor eld. Here the prime example is Bott's proof of Bott periodi
ity ([5, see also
Part IV of [30), resting on the fa
t that the loop spa
e of a Riemannian manifold
is endowed with a natural Morse fun
tion. In the
ombinatorial setting, I would
pla
e the Morse-theoreti
examination of evasiveness of the previous se
tion in this
ategory. Third are the
ases in whi
h the obje
ts under investigation
an be naturally
identied as the
riti
al points of a Morse fun
tion on a larger spa
e. Examples of
this phenomenon abound in dierential geometry, where one often studies extremals
of energy fun
tionals. In parti
ular, Morse's rst great triumph with Morse Theory
was his investigation of the set of geodesi
s between two points in a Riemannian
manifold ([32, see also Part III pf [30). The geodesi
s are pre
isely the
riti
al
points of the natural Morse fun
tion on the path spa
e, and Morse used the Morse
inequalities, along with a knowledge of the topology of the path spa
e, to dedu
e
the existen
e of many
riti
al points. It is intriguing to this author that there are as
yet no
orresponding examples in the
ombinatorial setting. I know of no examples
in whi
h a
olle
tion of
lassi
ally studied obje
ts in
ombinatori
s
an be naturally
identied with the
riti
al simpli
es of a Morse fun
tion on some larger
omplex.
Indeed, I believe that soon
ombinatorial examples of interest will be found that t
into this third
ategory. I wonder if appli
ations of dis
rete Morse theory will be found
that approa
h the beauty, depth and fundamental signi
an
e of the appli
ations of
smooth Morse Theory mentioned in this paragraph.
On a broader note, I believe that dis
rete Morse Theory is only a small part of what
someday will be a more
omplete theory of \
ombinatorial dierential topology",
although I hesitate to predi
t (at least in print) what form su
h a theory will take.
43
Referen es
44
ROBIN FORMAN
[28 A. Lundell and S. Weingram, The Topology of CW Complexes, Van Nostrand Reinhold
Company, New York, 1969.
[29 V. Mathai and S.G. Yates, Dis
rete Morse theory and extended L2 homology, preprint,
1999.
[30 J. Milnor Morse Theory, Annals of Mathemati
s Study No. 51, Prin
eton University Press,
1962.
Le
tures on the h-Cobordism Theorem Prin
eton Mathemati
al Notes, Prin
eton
[31
University Press, 1965.
[32 M. Morse, The Cal
ulus of Variations in the Large, Amer. Math. So
. Colloquium Pub.
18, Amer. Math. So
., Providen
e R.I., (1934).
[33
, Bowls of a Non-Degenerate Fun
tion on a Compa
t Dierentiable Manifold, in
Dierential and Combinatorial Topology (A Symposium in Honor of M. Morse), Prin
eton
University Press (1965), pp. 81-104.
[34 S. Novikov, Multivalued fun
tions and fun
tions: An analogue of the Morse theory, Soviet
Math. Dokl., 24 (1981), pp. 222-226.
[35 R.L. Rivest and J. Vuillemin On re
ognizing graph properties from adja
en
y matri
es,
Theor. Comp. S
i, 3 (1976) pp. 371-384.
[36 M. S
hwartz Morse Homology Progress in Mathemati
s, 111, Birkhauser Verlag, Basel
(1993).
[37 J. Shareshian Dis
rete Morse Theory for Complexes of 2-Conne
ted Graphs, to appear in
Topology.
[38 V. Sharko, Fun
tions on Manifolds, Algebrai
and Topologi
al Aspe
ts, Trans. of Math.
Monographs 131, Amer. Math. So
., Providen
e, R.I., (1993).
[39 S. Smale, On Gradient Dynami
al Systems, Annals of Math. 74 (1961) pp. 199-206.
The Generalized Poin
are Conje
ture in Dimensions Greater than Four, Annals of
[40
Math. 74 (1961) pp. 391-406.
[41 R. Stanley, A Combinatorial De
omposition of A
y
li
Simpli
ial Complexes Dis
rete
Math. 118 (1993)
[42 V. Tur
hun Homology of Complexes of Bi
onne
ted Graphs Uspekhi Mat. Nauk. 52 (1997)
pp. 189-190.
[43 V. Vassiliev Complexes of Conne
ted Graphs in The Gelfand Mathemati
al Seminars,
1990{1992, pp. 223{235. Birkhauser Boston, Boston, (1993).
[44 J.H.C. Whitehead, Simpli
ial Spa
es, Nu
lei, and m-Groups, Pro
. London Math. So
.
45 (1939) pp. 243-327.
[45 E. Witten Supersymmetry and Morse Theory J. Di. Geom. 17 (1982) pp. 661{692.
Department of Mathemati
s, Ri
e University, Houston, TX 77251