computing persistent homology afra zomorodian and gunnar carlsson

45
Computing Persistent Homology Afra Zomorodian and Gunnar Carlsson

Upload: dylon-inglesby

Post on 01-Apr-2015

258 views

Category:

Documents


4 download

TRANSCRIPT

Page 1: Computing Persistent Homology Afra Zomorodian and Gunnar Carlsson

Computing Persistent Homology Afra Zomorodian and Gunnar

Carlsson

Page 2: Computing Persistent Homology Afra Zomorodian and Gunnar Carlsson

Motivation•Persistence complexes arise naturally whenever one is attempting to study topological invariants of a space computationally.•Given a complex K, we are interested in it’s kth homology and it’s persistence.•Computers are needed to handle high dimensional spaces.•This presentation will demonstrate an algorithm for computing persistent homology over a field.

Page 3: Computing Persistent Homology Afra Zomorodian and Gunnar Carlsson

Filtered Complex•A filtration of a complex K is a nested sequence of subcomplexes, ∅ = K0 ⊆ K1 ⊆ K2 ⊆ . . . ⊆ Km = K.

Persistence Complex•A filtered simplicial complex, along with its associated chain and boundary maps.

Page 4: Computing Persistent Homology Afra Zomorodian and Gunnar Carlsson

Boundary Operator

Page 5: Computing Persistent Homology Afra Zomorodian and Gunnar Carlsson

Cycles and Boundaries•A chain is a cycle when its boundary is zero•The cycles form a subgroup Zk(К) of chain group Ck(К), which is the kernel of boundary operator: Zk(К) =ker(∂k)

•The elements in Im(∂k+1) are called boundaries.•The k-boundary group of К is the set of boundaries of (k+1)-chains in К, i.e. Its the Image of the (k+1)-chain group• Bk(К)= Im(∂k+1)

Page 6: Computing Persistent Homology Afra Zomorodian and Gunnar Carlsson

Cycles and Boundaries

Page 7: Computing Persistent Homology Afra Zomorodian and Gunnar Carlsson

Homology•A certain general procedure used to associate Modules with a Topological Space

•The kth homology Module is .

•Also depicts the “holes” in the shape created by the Complex K.

/k k kH Z B

Page 8: Computing Persistent Homology Afra Zomorodian and Gunnar Carlsson

Modules•Module over a ring is a generalization of the notion of vector space, wherein the corresponding scalars are allowed to lie in an arbitrary ring.

r(x + y) = rx + ry(r + s)x = rx + sx

(rs)x = r(sx)1Rx = x if R has multiplicative identity 1R

, , ,r s R x y M

Page 9: Computing Persistent Homology Afra Zomorodian and Gunnar Carlsson

Standard matrix representation (1)•The boundary operator performs the following manipulation on the vertices of k-simplices in K

•Which can be written in the form of a matrix with entries {-1,0,1}.

Page 10: Computing Persistent Homology Afra Zomorodian and Gunnar Carlsson

Standard matrix representation (2)• has columns and rows representing the number k and k-1 simplices in K.•The null space (Kernel) of corresponds to . •The Range space corresponds to .•Linear operations on the matrix rows cause basis change in

kmkM 1km

kM kZ

1kB

1kC

Page 11: Computing Persistent Homology Afra Zomorodian and Gunnar Carlsson

Standard matrix representation (3)•The reduction algorithm transforms into its diagonal form with the elementary operations:

•The above mentioned operators also work on columns that correspond to a base change in .

kC

,

1

i j

i i

i j i

r r i j

r r

r zr r z

kM

Page 12: Computing Persistent Homology Afra Zomorodian and Gunnar Carlsson

Algorithm – Methodology•The homology of a Complex K is a vector space of rank=β.

•The standard method for computing homology is the reduction algorithm for the representing matrix .

•Exploring the properties of the Matrices will allow us to deduct the properties of the topological space.

kM

kM

Page 13: Computing Persistent Homology Afra Zomorodian and Gunnar Carlsson

Algorithm – principles•The oriented k-simplices form the standard basis for the kth complex in the chain C.

• Matrix representation reduction.

•Matrix rank Betti numbers

kC

k

Page 14: Computing Persistent Homology Afra Zomorodian and Gunnar Carlsson

Example of a basis•The two solid 1-cycles form a basis for the first homology class of the torus.•The two solid 1-cycles form a basis for the first homology class of the torus

Page 15: Computing Persistent Homology Afra Zomorodian and Gunnar Carlsson

Diagonal form (1)•After reduction we get the following form of the matrix M, now called :•

kM

ˆk k kl rank M rank M

Page 16: Computing Persistent Homology Afra Zomorodian and Gunnar Carlsson

Diagonal form (2)•The basis spans the kernel Z, whereas spans the range B.

| 1i k ke l i m

ˆ |1i i kb e i l

1k k k k k k krank H rankZ rankB m l l

Page 17: Computing Persistent Homology Afra Zomorodian and Gunnar Carlsson

Example (1)

Page 18: Computing Persistent Homology Afra Zomorodian and Gunnar Carlsson

Example (2)•Matrix representation of .•This is the matrix that takes us from 1-simplices to 0-simplices (vertices) in K.

1

Page 19: Computing Persistent Homology Afra Zomorodian and Gunnar Carlsson

Example (3)•The reduced representation for , where:

1M

11

2

0

ad bc cd ab

ac bc ab

, ,

zZ span

z

B span d c c b b a

Page 20: Computing Persistent Homology Afra Zomorodian and Gunnar Carlsson

Algorithm downsides•The reduction algorithm requires elementary operations, where m is the number of simplices in K.

•The entries of the intermediate matrices may become extremely large while maintaining exact values (requires long integer storage word).

•We need to simplify the process

3O m

Page 21: Computing Persistent Homology Afra Zomorodian and Gunnar Carlsson

Homology of Filtration•Given a filtered complex, the ith complex has associated boundary operators , matrices , and groups for all i, k ≥0.

•i is called the filtration index.

iKik

ikM

, , ,i i i ik k k kC Z B H

Page 22: Computing Persistent Homology Afra Zomorodian and Gunnar Carlsson

Persistence•assuming a k-cycle z, created at time i with the arrival of simplex σ into the complex.•The homology class of this cycle, [z], is an element of .•arrival of simplex τ at time j ≥ I, merges [z] with an older class, turning a cycle z’ in [z] into a boundary. •Rank of homology group decreases.

ikH

' ikz B

Page 23: Computing Persistent Homology Afra Zomorodian and Gunnar Carlsson

P Interval•A P-interval is an ordered pair (i, j) with 0 ≤ i < j ∈ Z∞ = Z ∪ {+∞}.

•Describes the the appearance time of a simplex, i, and the disappear time j-1.

•The P interval is the lifetime (j-i-1) of this class in the filtration.

Page 24: Computing Persistent Homology Afra Zomorodian and Gunnar Carlsson

Persistent Homology

•The p-persistent kth homology group of is

•By increasing p sufficiently we “kill” short-term changes – the noise. Over-increasing p will cover the attributes of the space.

iK

,

, ,

/i p i i p ik k k k

i p i pk k

H Z B Z

rank H

Page 25: Computing Persistent Homology Afra Zomorodian and Gunnar Carlsson

Persistence Complex•Moving to the right increases filtration index.•Moving down decreases dimensions

Page 26: Computing Persistent Homology Afra Zomorodian and Gunnar Carlsson

Algorithm over fields•Simplify the standard reduction algorithm by using “The Persistence Module”.

•We compute the P-intervals directly over the field without computing the Persistence Module.

Omer
Define!!!
Page 27: Computing Persistent Homology Afra Zomorodian and Gunnar Carlsson

Mid-conclusion•Add 3.3 from the paper.

Page 28: Computing Persistent Homology Afra Zomorodian and Gunnar Carlsson

Example’ (1)•Compute over .•Persistence Module corresponds to .

•Definition: • is the basis for • is the basis for

2 2 t

ˆ

i

j

e

e1

k

k

C

C

Page 29: Computing Persistent Homology Afra Zomorodian and Gunnar Carlsson

Example’ (2)•Algorithm:•We start from the trivial case ofusing the standard basis to represent .•Extract the matrix rep. to

0 0 00, Z C

1

kM k

je

Page 30: Computing Persistent Homology Afra Zomorodian and Gunnar Carlsson

Example’ (3)•Degrees of simplices in the filtration.

•Where:

ˆdeg deg , degi k je M i j e

Page 31: Computing Persistent Homology Afra Zomorodian and Gunnar Carlsson

Example’ (4)•We bring the Matrix to column echelon form using Gaussian elimination.

1

1 2 22

ad cd t·bc t·ab

z = ac t ·bc t ·ab

zZ

Page 32: Computing Persistent Homology Afra Zomorodian and Gunnar Carlsson

Algorithm•Lemma 4.1 (Echelon Form) The pivots in column echelon form are the same as the diagonal elements in normal form. •Moreover, the degree of the basis elements on pivot rows is the same in both forms.•Let be the column-echelon form for ∂k relative to bases {ej} and for Ck and Zk−1, respectively.

kM ie

Page 33: Computing Persistent Homology Afra Zomorodian and Gunnar Carlsson

Meaning•If row i has pivot then it contributes a P-interval of to . •otherwise, the interval is

•In our case - giving an interval of (1,2) to the description of H0.

kM (i, j) tn

iˆ ˆdeg e , deg e ni 1kH

ˆdeg ,ie

1M (1, 1) = t

Page 34: Computing Persistent Homology Afra Zomorodian and Gunnar Carlsson

Example’ (5)•Next, we wish to find the k+1 matrix using the The base we computed for , since:•

•As we perform the column opertors on to get it to echelon form, we perform the same operators on rows in which causes rows to zero out.

1 10 0k k k kM M

kM

1kM

kZ

Page 35: Computing Persistent Homology Afra Zomorodian and Gunnar Carlsson

Example’ (6)•Since the corresponding rows even out, we might as well erase them manualy and replace the wors with the basis of :

1

1 2 22

ad cd t·bc t·ab

z = ac t ·bc t ·ab

zZ

1Z

Page 36: Computing Persistent Homology Afra Zomorodian and Gunnar Carlsson

Conclusions from example•Matrix in Echelon form gives us all the information we need.•Next dimension matrix is partly eliminated by column operators from the current dimension.

•There is no need for row operators

•There is no need for matrix representation

Page 37: Computing Persistent Homology Afra Zomorodian and Gunnar Carlsson

Data structure•From now on we work with an array T with a slot for each simplex.•we compute homology in all dimensions incrementally and concurrently.

Page 38: Computing Persistent Homology Afra Zomorodian and Gunnar Carlsson

COMPUTE INTERVALS (K)The algorithm stores the list of P-intervals for Hk in Lk.

Page 39: Computing Persistent Homology Afra Zomorodian and Gunnar Carlsson

COMPUTE INTERVALS (K)•New simplex is passed through “RPR” routine to check whether its boundary d corresponds to zero ( ) or pivot. If “zero” – we mark so corresponding row doesn’t get eliminated in k+1.•If is a pivot, we look for the pivot value (maxindex(d)).•We store index j and chain d representing the column in T[i].• •We then perform another pass through the filtration in search of infinite P-intervals: marked simplices whose slot is empty.

jkZ

j

j

int deg ,degi jP erval

Page 40: Computing Persistent Homology Afra Zomorodian and Gunnar Carlsson

REMOVE PIVOT ROWS

•If T[i] is non-empty, a pivot was already stored and we eliminate the row. Otherwise, we found a pivot

Gaussian elimination

Page 41: Computing Persistent Homology Afra Zomorodian and Gunnar Carlsson

Algorithm - results• For our example, the marked 0-simplices {a, b, c, d} and 1-simplices {ad, ac} generate P-intervals L0 = ={(0,∞), (0, 1), (1, 1), (1, 2)} andL1 = {(2, 5), (3, 4)}

,i j

Page 42: Computing Persistent Homology Afra Zomorodian and Gunnar Carlsson

Conclusion•When a 0-simplex is created – H0 is incremented.•When a 1-simplex is created, either H1 is increased or H0 is decreased. Etc…

•each P-interval (i, j) describes a basis element for the homology vector spaces starting at time i (in which he joins ) until time j − 1.•This basis element is a non-bounding k-cycle until filtration index j, till he joins .

jkB

jkZ

i i ik k krank Z rank B

Page 43: Computing Persistent Homology Afra Zomorodian and Gunnar Carlsson

Conclusion•L0 ={(0,∞), (0, 1), (1, 1), (1, 2)}

L1 = {(2, 5), (3, 4)}

Page 44: Computing Persistent Homology Afra Zomorodian and Gunnar Carlsson

Multidimensional FiltrationCOMPUTING MULTIDIMENSIONAL PERSISTENCE

Gunnar Carlsson, Gurjeet Singh, and Afra Zomorodian

Page 45: Computing Persistent Homology Afra Zomorodian and Gunnar Carlsson

SourcesCOMPUTING MULTIDIMENSIONAL PERSISTENCE

Gunnar Carlsson, Gurjeet Singh, and Afra Zomorodian