You are on page 1of 13

Simplicial Complexes and Their Applications

Abhineet Agarwal, Noah Goss


February 2019

1 A Review of Simplices
Def inition : Let S be a discrete set. An abstract simplicial complex is a col-
lection X of finite subsets of S, closed under restriction: ∀σ ∈ X, all subsets of
σ ⊂ X.

Let’s recall some properties of simplicial complexes:


• Every σ ∈ X is called a simplex.

• The k simplex corresponds to σ ⊂ X s.t. |σ| = k + 1


• For a given k simplex, the faces of the k simplex are the simplices corre-
sponding to all subsets of σ ⊂ X.
Def inition : For k+1 affinely independent points in Rk+1 , i.e. a set of points
u0 , u1 . . . uk , s.t. u1 −u0 , . . . , uk −u0 are linearly independent, we can topologize
a simplicial complex X as the quotient space built from topological simplices.
The standard K-simplex can then be defined as
k
X
k
∆ = {x0 uo + x1 u1 + . . . xk uk : xi = 1} (1)
i=0

Def inition : An independence complex, JO , is an abstract simplicial complex


defined on the vertex set O whose k-simplices are collections of k+1 independent
objects.

Applications of these independence complexes can be seen in the form of

• Linearly independent vectors


• Linearly independent Differential Equations
• Statistical independence of random variables

1
2 Vietoris-Rips Complexes and Point Clouds
It will be useful for the rest of this lecture to start by defining the concepts of
Homotopy Equivalence and of Convex Hulls.

Def inition : Two continuous functions from one topological space to another
are called homotopic if one can be continuously deformed into the other. Two
spaces X and Y are considered Homotopic Equivalent if there exists con-
tinous maps f : X → Y and g : Y → X s.t. g ◦ f is homotopic to the identity
mapping IdY and g ◦ f is homotopic to the identity mapping IdX .

Def inition : The Convex Hull of a set X of points defined in Rn , is the


smallest convex set that contains X. Recall that a set of points is convex if it
contains a line segment connecting each possible combinations of pairs of points
in X.

Figure 1: In the plane of R2 , a very useful way of thinking about the convex
hull of a set of points is a rubber band wrapping around the bordering elements

Def inition : Consider a discreet subset Q ⊂ Rn sampled from a submanifold

2
(recall all manifolds are locally homeomorphic to Rn ), we can define this discreet
subset as a point cloud. One could reasonably ask if there is some method by
which we can reconstruct a picture of our original submanifold just given the
datum contained in our point cloud.

Def inition : The Vietoris-Rips Complex of scale  defined on Q, V R (Q),


is the simplicial complex, whose simplices are the finite collections of points in
Q of pairwise distance ≤ . (Fun fact, the eponymous Austrian mathematician
Leopold Vietoris lived to 110!)

Now that we have constructed an abstract simplicial complex from our point
cloud, it is worth asking how useful this construct is in gleaning information
about our original submanifold. Some immediate questions we can ask are/con-
cerns we may have are :
• What is a good choice of  for our point cloud?
• What time complexity would an algorithm computing Vietoris-Rips com-
plexes for a point cluster have? (Exponential O(2n ))
• Although our point cloud lives in Rn , for sufficiently dense clusters of
points, we may construct simplices of order much greater than n.
We can define the projection map S : VR → Rn by taking the vertices (0 sim-
plices) in VR to Q and taking the K simplices in VR to the convex hulls of
the associated vertices in Q. The image of this projection S in Rn defines the
shadow of our Vietoris-Rips complex.

Can we generally expect our shadow to be homeomorphic to our Vietoris-Rips


complex? Unfortunately not, as generally speaking the domain of our projection
will likely be of higher dimension than our image in Rn . Unfortunately homo-
topy equivalence is also not something we can generally expect either, but in
general Vietoris-Rips complexes have applications to topological computation,
and have value in characterizing data with large holes.

3 The Čech Complex


One method to extract topological characteristics of a submanifold from a point
cloud that better avoids unwanted higher dimensional features appearing is the
Čech Complex.

Def inition : Given a point cloud Q the Čech Complex Č  is the simplicial
complex built on Q given by: a k simplex in Č  is a collection of k+1 distinct
elements xi of Q such that the net intersection of diameter  balls centered on
the xi ’s is non-empty.

3
Figure 2: A Vietoris-Rips complex defined in R2 for a point cloud. Red points
are 0-simplices, Edges are 1-simplices, Light Blue triangles define 2-simplices,
Dark Blue defines 3-simplices

One can immediately see that the algorithm for constructing Čech Complexes
will be more computationally intensive than their Vietoris-Rips counterparts, so
the question should be asked: is there any advantage to trying to characterize
our point cloud with Čech Complexes? Short answer: yes because Homotopy
equivalence is guaranteed by the Nerve Lemma between the union of the balls
and the Čech Complex.

Def inition : Given a collection U = {Uα } of compact subsets of a topological


space X, we can construct the nerve N (U) as follows: the k-simplices contained
in N (U) correspond to non-empty intersections of k+1 distinct elements in U.

Def inition : A topological space X is contractible if the identity map on X


is homotopic to some constant map. More intuitively, a topological space X is
contractible if it can be continuously shrunk to a point in space.

T he N erve Lemma : If U is a finite collection of open contractible sub-

4
Figure 3: An example of the Čech Complex constructed by points sampled from
a circle in the plane

S intersections of sub-collections of U contractible,


sets of X with all non-empty
then N (U) is homotopic to α Uα .

Figure 4: Which of the following spaces are contractible?

T heorem : ∀ > 0 Č  ⊂ VR ⊂ Č 2 and thus if we know we can piece together


a solid picture of our submanifold using data from the point cloud with Čech
Complexes of  and 2 then, we know that the Vietoris-Rips Complex of  will
be a good way of topologically characterizing our data as well.

5
4 Computational Applications
Now that we have defined abstract simplicial complexes, something that we are
interested in is the computational applications of simplicies – one of these ar-
eas is Machine Learning. Broadly in Machine Learning, we have the following
task: we are given some training data, and some test data that we would like
to classify . For example, we are given the heights and weights and from that
would like to classify if a person is male or female. The point of training data is
to train our machine learning algorithm so that it is able to perform well when
handed test data.

We can formalise this definition as follows; we are given labelled training data:
(−
→, y ), ...(−
x 1 1 x→, y )
n n (2)
We would use this training data to learn a f : X → Y , such that when given
some test data, we are able to predict it correctly.

4.1 Linearly Seperable Data and Perceptron


One such data set we can consider is the linearly seperable data set as we see in
the picture below. Therefore, we see there exists a linearly seperable boundary

Figure 5: Linearly Seperable Data Set

and that implies that we can simply have the following function to classify our
data:
g(→
−x)=→ −
w .→

x +b=0 (3)

6
f = linear classifier = sign(g(→

x )). Therefore, learning f implies learning w such
that we can classify our training data. We can do this using the perceptron
algorithm as follows:

Perceptron Algorithm
Given: labelled training data S = (− →,y ),...(−
x 1 1 x→
n ,yn )
−→
Step 1: Initialize wo = 0
For t = 1,2,3,....
If there exists (→

xi ,yi )) such that sign(→−
w t−1 .→

x 6= y) then do the following :

− →

w t = w t−1 + yx (4)
Terminate when there does not exist a mistake.

Claim: Perceptron algorithm will make at most T = ( R 2


γ ) , where γ is the


distance of the closest point to the boundary and R = max|| x ||

Figure 6: Perceptron Algorithm

Pf: Since the data is linearly seperarable, there exists a plane −


→ such that
w∗
the data is linearly seperable. Suppose the perceptron algorithm makes a mis-
take in iteration t, then


w t .→

w ∗ = (→

w t−1 + y →

x ).→

w∗ ≥ →

w t−1 .→

w∗ + γ (5)

7
||−
→||2 = ||→
wt

w t−1 +yx||2 = ||→

w t−1 ||2 +||yx||2 +2y(→

w t−1 .→

x ) ≤ ||→

w t−1 ||2 +R2 (6)
Therefore after T rounds:

Tγ ≤ →

w T .→

w ∗ ≤ ||→

w T ||||→

w ∗ || ≤ R T (7)

Therefore: T ≤ ( R
γ)
2

Non-Linearly Seperable Data and Simplices


Data is rarely Linearly Seperable, usually we require complex data boundaries
for classification as shown below:

Figure 7: NonLinear Decision Boundary for Classification

How do we classify this data? Clearly generating a linearly seperable boundary


is impossible. But what if there existed a representation space such that we
could linearly seperate our data? It turns out there is: lets simply map our
data points onto vertices of a simplex! Consider the mapping:
 
0
0
 .. 
 

− .
φ( xi ) = 
1
 (8)
 
.
 .. 
0

8
Then our linear seperation boundary simply becomes:
 
y1
 y2 

w→ ==  

(9)
∗ .
 .. 
yn

Therefore, we can map our data to another representation space. The geometry
of our points becomes a simplex and by working in this space, our data is now
linearly seperable! Of course, this is not the only mapping(it may not even be
the best mapping), but the point is that this is one example of a mapping we
can do to interesting geometries such that we can work in higher-dimensional
space that better respects our algorithms.

5 Strategy Complexes
Before we formally define strategy complexes, let us consider the following ex-
ample: Imagine an ambulance rescue in an old complicated city, perhaps after
an earthquake. In the figure below, we have an ambulance A trying to reach a
patient X, and they have a variety of paths. For example, from A → X directly
or A→B→C. We can represent the strategy described as the solid triangle in
the figure. In other words, this is a non-deterministic graph but it is also a
simplex! The vertices of the simplex are the individual actions to be executed
at any particular location.

Figure 8: Strategy Complex

We can also augment the graph above to include more actions:

9
Figure 9: Augmented Strategy Complex

Note that we do not want to include the cycle B to C since it is not a valid
strategy to reach X. This is because if we did then we would be caught in an
infinite cycle. Now that we have this example in mind, we can formally define
strategy complexes and non-deterministic graphs.

Def inition : A nondeterministic graph G = (V,A) is a set of states V and


a collection of (non-deterministic) actions A. Each a ∈ A is a nonempty set of
directed edges: {(v, u1 ), ....} with v and all possible ui in V. We refer to v as A’s
source and to each ui as a non-deterministic target of A. G is also considered
acyclic if for a sequence of states: v0 , ....vk in V such that (vi, vi+1 ) ∈ A, if none
of the possible paths have v0 = vj for j ≥ 1

Def inition : Given a non-deterministic graph G = (V,A), we let ∆G be the


simplicial complex whose simplices are the acyclic collections of actions B ⊂ A.
If V = Φ then we also let ∆G = Φ. We refer to ∆G as G’s strategy complex
and to every simplex in ∆G as a (non-deterministic) strategy
Examples: Consider the following graph.

Figure 10: The directed graph on the left defines the strategy complex on the
right

10
Note that while you would naively expect that this directed graph would
generate a simplex that is a tetrahedron because of all your four actions you
can take, this is not the case. Two of the actions 1→2 and 2→1, could give rise
to a cycle in the graph, so no simplex of the strategy complex can contain both
these actions. The complex is in fact generated by two triangles.

The two triangles, as well as three of the five edges in the complex, consti-
tute strategies for attaining state 3 in the graph. The central edge, consisting
of the actions {1 → 3, 2 → 3}. Also note that a strategy complex may consist of
strategies for a variety of goals. For instance, the top left edge of the complex
in Fig.5, comprising the actions {1 → 3, 1 → 2} is simply the strategy that says
move away from 1. Similarly, the edge below that is the strategy that says move
away from 2.

For contrast, consider the following figure.

Figure 11: The graph on the left

It contains two actions, ones each at state 1 and state 2. Each Each action has
two nondeterministic outcomes. The two actions cannot appear together asa
simplex since, depending on the actual nondeterministic transitions at runtime,
these actions could cause cycling in the graph between states 1 and 2. As a re-
sult, the strategy complex consists of two isolated vertices, representing the two
strategies “move away from state 1” and “move away from state 2.” In particu-
lar, there are no strategies guaranteed to attain state 3 from the other two states.

In fact, it was proved that any directed graph that can be written as the disjoint
union of its strongly connected components generate a strategy complex topo-
logically similar to a sphere of dimension n-k-1, where n is the number of states
in the graph and k is the number of strongly connected components. All other
directed graphs produce contractible strategy complexes – recall that means
roughly that the directed graph can produce strategy complexes that can be
shrunk to a point.

11
5.1 Loopback Graphs and Complexes
Let us modify the graph in Figure 9 by adding artificial deterministic transitions
from state 3 to each of states 1 and 2. We call these added transitions loopback
actions.

Figure 12: A loopback graph and loopback complex. The two triangular endcaps
outlined in dashed red are not part of the complex, since each give rise to a cycle.

The complex in this case looks roughly like a polygonal cylinder. The complex
is homotopic to a circle(we can also see that by an application of the previous
statement) Now we also add loopback graphs to the non-deterministic graph of

Figure 13: A loopback graph and loopback complex.

earlier. The complex is shown to the right, and it is homotopic to a point! This
gives us a slight hint to a later theorem.

In fact, no matter how complicated the non-deterministic graph, if we add


all loopback actions to it that transition from state s to the remaining states,
then the resulting loopback complex will always be homotpic either to a sphere
or a point. A phere tells us that there is a strategy guaranteed to attain state
s from all states in the graph; a point tell us no such strategy exists.

Def initions : Let G = (V,A) be a nondeterministic graph and suppose s


∈ V is some desired stop state. We make the following definitions:

12
• G contains a complete guaranteed strategy for attaining s if there is some
acyclic set of actions B ⊂ A such that B contains at least one actions
with source v for every v ∈ V \ {s}. Observe that any possible path in
the graph(V,B) terminates at some vk 6= s, may be extended to a path
converging at s. B is a complete guaranteed strategy for attaining
s.
• Define G←s to be the non-deterministic graph identical to G except the
source s have been discarded, replaced instead by (|V | − 1) many loopback
actions, each consisting of a single edge from s to some v, with v ranging
over V \ {s}

• ∆G←s is the strategy complex associated with G←s


Theorem: Let G = (V,A) be a non-deterministic graph and s ∈ V. If G contains
a complete guaranteed strategy for attaining s, then ∆G←s is homotopic to
sphere S n−2 , with n = |V |. Otherwise, G is contractible.

13

You might also like