Dionysus Slides

A Practical Guide to Persistent
Homology
Dmitriy Morozov
Lawrence Berkeley National Lab
A Practical Guide to Persistent
Homology
(Dionysus edition)
from dionysus import *
Code snippets available at: from dionysus.viewer import *
http://hg.mrzv.org/Dionysus-tutorial from readers import *
Dmitriy Morozov
Lawrence Berkeley National Lab
Dionysus
C++ library
Implements various algorithms that Ive found interesting over the years:
ordinary persistence
vineyards
image persistence
zigzag persistence
persistent cohomology
circular coordinates
alpha shapes
Vietoris-Rips complexes
bottleneck and wasserstein distances between diagrams
To make life easier, added Python bindings
This talk exclusively in Python
Python
Good news: You already know Python! Its just like pseudo-code in
your papers, but cleaner. ;-)
Lists and list comprehensions
lst1 = [1,3,5,7,9,11,13]
lst2 = [i for i in lst1 if i < 9]
print lst2 # [1,3,5,7]
Functions
def pow(x):
def f(y):
return y**x
return f
Loops and conditionals
for i in lst1:
if i % 3 == 0 and i > 5:
print square(i)
Lots of extra functionality in modules
from math import sqrt
from dionysus import *
Persistent Homology
Over a decade old now. Introduced as a way to detect prominent topo-
logical features in point clouds. Since then evolved into a rich theory with
many applications.
What is the homology of this point cloud?

Persistent Homology
many applications.
Squint our eyes

Persistent Homology
many applications.
Squint our eyes

Persistent Homology
many applications.
Squint our eyes

Persistent Homology
many applications.
Squint our eyes

Persistent Homology
many applications.
Squint our eyes

Persistent Homology
many applications.
Squint our eyes no natural fixed scale persistent homology

Eye Squinting
P point set in Rn Pr = pP Br (p)
Eye Squinting
Eye Squinting
Eye Squinting
Eye Squinting
Eye Squinting
0 H(Pr1 ) H(Pr2 ) . . . H(Rn )

Eye Squinting
Death
10 points
Dgm1
Birth
0 H(Pr1 ) H(Pr2 ) . . . H(Rn )

Eye Squinting
Death
10 points
Dgm1
Birth
0 H(Pr1 ) H(Pr2 ) . . . H(Rn )

Eye Squinting
Death
10 points
Dgm1
Birth
Squinting our eyes gives us a

continuous function. Algorithms
work with (discrete) simplicial
0 H(Pr1 ) H(Pr2 ) . . . H(Rn ) complexes.
Simplices and Complexes
0 (Geometric) k-simplex: convex hull of (k + 1) points.
(Abstract) k-simplex: subset of (k + 1) elements of a
universal set.
2 1
Boundary: [v0 , . . . , vk ] = i (1)i [v0 , . . . , vi , . . . , vk ]
P
s = Simplex([0,1,2])
print "Dimension:", s.dimension Dimension: 2
Vertices:
print "Vertices:" 0
for v in s.vertices: 1
print v 2
Boundary:
print "Boundary:" <1, 2>
for sb in s.boundary: <0, 2>
print sb <0, 1>
universal set.
2 1
P
Simplicial complex: collection of simplices closed under

face relation.
1 3
2 4
not a simplicial
complex:
universal set.
2 1
P

face relation.
1 3
complex = [Simplex(vertices) for vertices in
[[0], [1], [2], [3], [4], [5],
0
[0,1], [0,2], [1,2], [0,1,2],
[1,3], [2,4], [3,4]]]
2 4
not a simplicial
complex:
universal set.
2 1
P

face relation.
1 3
complex = [Simplex(vertices) for vertices in
[[0], [1], [2], [3], [4], [5],
0
[0,1], [0,2], [1,2], [0,1,2],
[1,3], [2,4], [3,4]]]
2 4
not a simplicial simplex9 = Simplex([0,1,2,3,4,5,6,7,8,9])

complex: sphere8 = closure([simplex9], 8)
print len(sphere8)
1022
Homology
over Z2 , a set of
k-chain = formal sum of k-simplices simplices
k-cycle = chain without a boundary
k-boundary = boundary of an (k + 1)-dimensional chain
two cycles are homologous

Z = cycle group
if they differ by a boundary
B = boundary group
H = Z/B
homology: count cycles up to

differences by boundaries
Homology in Dionysus
Dionysus doesnt compute homology directly, but we can get it as a by-
product of persistent homology.
complex = sphere8
Dimension: 0
0 inf
f = Filtration(complex, dim_cmp)
Dimension: 1
p = StaticPersistence(f) Dimension: 2
p.pair_simplices() Dimension: 3
Dimension: 4
dgms = init_diagrams(p,f, lambda s: 0) Dimension: 5
Dimension: 6
for i, dgm in enumerate(dgms): Dimension: 7
print "Dimension:", i Dimension: 8
print dgm 0 inf
03-complex.py
Persistent Homology (pipeline)
Filtration of a simplicial complex:
K1 K2 . . . Kn
(w.l.o.g. assume Ki+1 = Ki + i ).

so, really, an ordering of simplices
1 2 3 2 4 2 5 2 6
1 1 1 1 1
0 0 0 0 0 0
K1 K2 . . . Kn

1 2 3 2 4 2 5 2 6
1 1 1 1 1
0 0 0 0 0 0
simplices = [([0], 1), ([1], 2), ([0,1], 3), ([2], 4), \

([1,2], 5), ([0,2], 6)]
f = Filtration()
for vertices, time in simplices:
f.append(Simplex(vertices, time))
f.sort(dim_data_cmp)
for s in f:
04-1-filtration.py
print s, s.data # s.data is the time
K1 K2 . . . Kn

1 2 3 2 4 2 5 2 6
1 1 1 1 1
0 0 0 0 0 0
H1 :
H0 :
H(K1 ) H(K2 ) . . . H(Kn )

p = StaticPersistence(f)
p.pair_simplices()
dgms = init_diagrams(p,K1 f)
K2 . . . Kn
for i, dgm in enumerate(dgms):
print
(w.l.o.g. "Dimension:",
assume Ki+1 = Kii + i ).
print dgm 04-2-persistence.py
1 2 3 2 4 2 5 2 6
1 1 1 1 1
0 0 0 0 0 0
H1 :
H0 :
H(K1 ) H(K2 ) . . . H(Kn )

Filtrations: -shapes
Kr = Nrv{Br (u) Vor u}
P : Kr ' pP Br (p)
Kr1 Kr2 . . . Kr . . .
P : Kr ' pP Br (p)
Kr1 Kr2 . . . Kr . . .
P : Kr ' pP Br (p)
Kr1 Kr2 . . . Kr . . .
P : Kr ' pP Br (p)
Kr1 Kr2 . . . Kr . . .
P : Kr ' pP Br (p)
Kr1 Kr2 . . . Kr . . .
P : Kr ' pP Br (p)
Kr1 Kr2 . . . Kr . . .
P : Kr ' pP Br (p)
Kr1 Kr2 . . . Kr . . .
P : Kr ' pP Br (p)
Kr1 Kr2 . . . Kr . . .
r = min dP (x)
xVor
P : Kr ' pP Br (p)
Kr1 Kr2 . . . Kr . . .
r = min dP (x)
xVor

points = read_points(data/trefoil.pts)
f = Filtration()
fill_alpha_complex(points, f)
show_complex(points, [s for s in f if sqrt(s.data[0]) < 1])
Fills f with all the simplices of the Delaunay triangulation

(thanks to CGALs Delaunay package).
The data field of each simplex is set to a pair (r2 , Vor 6= ).

P : Kr ' pP Br (p)
Kr1 Kr2 . . . Kr . . .
r = min dP (x)
xVor

points = read_points(data/trefoil.pts) an alpha shape is a one-liner thanks
f = Filtration() to list comprehensions
Fills f with all the simplices of the Delaunay triangulation

(thanks to CGALs Delaunay package).
The data field of each simplex is set to a pair (r2 , Vor 6= ).

P : Kr ' pP Br (p)
Kr1 Kr2 . . . Kr . . .
r = min dP (x)
xVor

f = Filtration()
05-alpha-shapes.py
p.pair_simplices()
dgms = init_diagrams(p, f, lambda s: sqrt(s.data[0]))

show_diagram(dgms)
Filtrations: Vietoris-Rips
VR(r) = { P | |u v| < r u, v }
(clique complex of r-nearest neighbor graph)
NB: only pairwise distances matter

VR(r) = { P | |u v| < r u, v }

distances = PairwiseDistances(points)
distances = ExplicitDistances(distances)
rips = Rips(distances)
f = Filtration()
rips.generate(2, 1.7, f.append)
print "Number of simplices:", len(f)
show_complex(points, f)
show_complex(points, [s for s in f if rips.eval(s) < 1.6])
VR(r) = { P | |u v| < r u, v }

rips = Rips(distances) skeleton
f = Filtration() cutoff
VR(r) = { P | |u v| < r u, v }

rips = Rips(distances) skeleton
f = Filtration() cutoff
06-rips.py
f.sort(rips.cmp)
p.pair_simplices()
dgms = init_diagrams(p, f, rips.eval)

show_diagram(dgms[:2])
Filtrations: Lower-Star
f
f : Vrt K R
f : |K| R linearly interpolated
a |K|a = f 1 (, a]
Interested in the filtration:
|K|a1 |K|a2 . . . |K|an

f
f : Vrt K R
a |K|a = f 1 (, a]
|K|a1 |K|a2 . . . |K|an
Ka = { K | max f(v) a}
v
(changes only as a passes vertex values)
|K|a ' Ka
So, instead, we can compute:
Ka1 Ka2 . . . Kan

f : Vrt K R
f
|K|a = f 1 (, a]
|K|a1 |K|a2 . . . |K|an
v
|K|a ' Ka
Ka1 Ka2 . . . Kan

f : Vrt K R
f
|K|a = f 1 (, a]
a
|K|a1 |K|a2 . . . |K|an
v
|K|a ' Ka
Ka1 Ka2 . . . Kan

elephant_points, elephant_complex = read_off(data/cgal/elephant.off)
elephant_complex = closure(elephant_complex, 2)
show_complex(elephant_points, elephant_complex)
def pojection(points, axis = 1): # projection onto a coordinate axis

def value(v):
return points[v][axis]
return value
value = projection(elephant_points, 1)

def value(v):
return value
def max_vertex_compare(value):
def max_vertex(s):
return max(value(v) for v in s.vertices)
def compare(s1, s2):
return cmp(s1.dimension(), s2.dimension()) or \
cmp(max_vertex(s1), max_vertex(s2))
return compare
f = Filtration(elephant_complex, max_vertex_compare(value))

def value(v):
return value
def max_vertex_compare(value):
def max_vertex(s): 07-ls-filtration.py
return max(value(v) for v in s.vertices)
def compare(s1, s2):
return cmp(s1.dimension(), s2.dimension()) or \
cmp(max_vertex(s1), max_vertex(s2))
return compare
f = Filtration(elephant_complex, max_vertex_compare(value))
p = DynamicPersistenceChains(f)
p.pair_simplices()
dgms = init_diagrams(p, f, lambda s: max(value(v) for v in s.vertices))
show_diagrams(dgms)
Extended Persistence
Extended persistence was introduced as a way to measure the essential
persistence classes:
H(Xa1 ) H(Xa2 ) ... H(Xan )

H(X)

H(X, Xa1 ) H(X, Xa2 ) . . . H(X, Xan ) H(X, )
Extended Persistence
Extended persistence was introduced as a way to measure the essential
persistence classes:
H(Xa1 ) H(Xa2 ) ... H(Xan ) H(X)

H(X, Xa1 ) H(X, Xa2 ) . . . H(X, Xan ) H(X, )
H(X, Y) ' H(X w Y, w)
execfile(08-extended-persistence.py)
Persistent Homology
Filtration D, ordered boundary matrix (indexed by simplices)
D[i, j] = index of i in boundary of j
Persistence Decomposition R = DV , where R is reduced, meaning low-
est ones are in unique rows, and V is upper-triangular.
R D V
Persistent Homology

0 =
R D V
cycle
Persistent Homology

0 =
R D V
boundary cycle chain
Persistent Homology

0 =
R D V
StaticPersistence computes just R, enough for the pairing.
Iterating over StaticPersistence, we can access columns of R, through cycle attribute.
(Also pair(), sign(), unpaired().)
smap = p.make_simplex_map(f)
for i in p:
if not i.sign():
print [smap[j] for j in i.cycle]
Persistent Homology

0 =
R D V
Iterating over StaticPersistence, we can access columns of R, through cycle attribute.
(Also pair(), sign(), unpaired().)
smap = p.make_simplex_map(f)
for i in p:
if not i.sign():
print [smap[j] for j in i.cycle]
DynamicPersistenceChains computes matrices R and V .
Access columns of V through chain. (E.g., gives access to the infinitely persistent classes.)
Persistent Homology

0 =
R D V
while True:
pt =over
Iterating show_diagram(dgms)
StaticPersistence, we can accessexecfile(08-cycle-chain.py)
columns of R, through cycle attribute.
(Also if not pt:
pair(), break
sign(), unpaired().)
print pt
smap i= =p.make_simplex_map(f)
pt[2]
for ismap = persistence.make_simplex_map(f)
in p:
chain
if not =i.sign():
[smap[ii] for ii in i.chain]
pair_cycle = [smap[ii]
print [smap[j] for jfor
in ii in i.pair().cycle]
i.cycle]
pair_chain = [smap[ii] for ii in i.pair().chain]
DynamicPersistenceChains computes matrices R and V .
show_complex(elephant_points, subcomplex = chain)
Access columns of V through chain. (E.g., gives access to the infinitely persistent classes.)
show_complex(elephant_points, subcomplex = pair_cycle + pair_chain)
Diagrams, Stability, and Distances
Dgm(f )
Bottleneck distance:
W (Dgm(f ), Dgm(g)) = inf kx (x)k

Dgm(f )
Dgm(g)

Dgm(f )
Dgm(g)

bottleneck_distance(dgm1, dgm2)
Dgm(f )
Dgm(g)

Dgm(f )
Dgm(g) Stability Theorem:
W (Dgm(f ), Dgm(g)) kf gk

Dgm(f )
Dgm(g) Stability Theorem:
W (Dgm(f ), Dgm(g)) kf gk
Wasserstein distance: (More sensitive to the entire diagram.)

X
Wqq (Dgm(f ), Dgm(g)) = inf kx (x)kq

wasserstein_distance(dgm1, dgm2, q)
Wasserstein Stability Theorem: For Lipschitz functions f and g,
under some technical conditions on the domain,
Wq (Dgm(f ), Dgm(g)) C kf gkk

Circle-Valued Coordinates
How to get a tangible feel for the topological features that we find?
H1 (X; Z)
= [X, S 1 ] Maps into circles, natural for:
Phase coordinates for waves
Angle coordinates for directions
Periodic data
H1 (X; Z)
Start with the canonical isomorphism Phase coordinates for waves

between 1-dimensional cohomology Angle coordinates for directions
classes and homotopy classes of maps
Periodic data
into a circle.
Algorithm:
1. Compute persistent cohomology classes
2. Turn each representative cocycle z into a map, X S 1
3. Smooth that map (minimize variation across edges),
staying within the same cohomology/homotopy class
(equivalently, find the harmonic cocycle)
H1 (X; Z)

Periodic data
into a circle.
Dgm1
Algorithm:
H1 (X; Z)

Periodic data
into a circle.
+2
-1 Vertices map to 0;
edges wind with the
degree given by z (e).
Algorithm:
H1 (X; Z)

Periodic data
into a circle.
Algorithm:
Persistent Cohomology in Dionysus
points = read_points(data/annulus.pts)
execfile(10-circular.py)
f = Filtration()
p = StaticCohomologyPersistence(f, prime = 11)

p.pair_simplices()
dgms = init_diagrams(p,f, lambda s: sqrt(s.data[0]), lambda n: n.cocycle)
while True:
pt = show_diagram(dgms)
if not pt: break
rf = Filtration((s for s in f if sqrt(s.data[0]) <= (pt[0] + pt[1])/2))
values = circular.smooth(rf, pt[2])
cocycle = [rf[i] for (c,i) in pt[2] if i < len(rf)]
show_complex(points, subcomplex = cocycle)
show_complex(points, values = values)
Persistent Cohomology in Dionysus
points = read_points(data/annulus.pts)
execfile(10-circular.py)
f = Filtration()
p = StaticCohomologyPersistence(f, prime = 11)

p.pair_simplices()
dgms = init_diagrams(p,f, lambda s: sqrt(s.data[0]), lambda n: n.cocycle)
while True:
pt = show_diagram(dgms)
if not pt: break
rf = Filtration((s for s in f if sqrt(s.data[0]) <= (pt[0] + pt[1])/2))
values = circular.smooth(rf, pt[2])
cocycle = [rf[i] for (c,i) in pt[2] if i < len(rf)]
show_complex(points, subcomplex = cocycle)
show_complex(points, values = values)
Image Persistence
Noisy domains: instead of f : X R, we have a function f : P R
P a sample of X
For suitably-chosen parameters and :
a1 a2 an
H(K ) H(K ) ... H(K )

H(Ka1 ) H(Ka2 ) . . . H(Kan )
Ka = alpha shape or Vietoris-Rips complex with parameter built
on f1 (, a]
Image Persistence
Noisy domains: instead of f : X R, we have a function f : P R
P a sample of X
For suitably-chosen parameters and :
a1 a2 an
H(K ) H(K ) ... H(K )

H(Ka1 ) H(Ka2 ) . . . H(Kan )
Ka = alpha shape or Vietoris-Rips complex with parameter built
on f1 (, a]
# assume parallel lists points and values
f = Filtration()
f = fill_alpha_complex(points, f)
# use persistence of f to choose alpha and beta chosen
f = Filtration([s for s in f if sqrt(s.data[0]) <= beta])

f.sort(max_vertex_compare(values))
p = ImagePersistence(f, lambda s: sqrt(s.data[0]) <= alpha)
p.pair_simplices()
dgms = init_diagrams(p, f, lambda s: max(values(v) for v in s.vertices))

show_diagrams(dgms)
Conclusions
Persistence is easy to use. Dionysus can help you try out new ideas.
Conclusions
Practice reinforces theory. For example, persistent cohomology algorithm,
in practice, is the fastest way I know to compute persistence diagrams.
(This realization is a pure accident of experimental work with circular
coordinates.) Studying why this is the case has lead to Dualities in
Persistent (Co)Homology.
Conclusions
Python bindings were one of the best decisions. (Hint, hint, CGAL.)
However, sometimes much slower than the C++ counter-parts. A lot of
the common functionality is available as examples in C++; dont overlook
them.
Conclusions
Python bindings were one of the best decisions. (Hint, hint, CGAL.)
However, sometimes much slower than the C++ counter-parts. A lot of
the common functionality is available as examples in C++; dont overlook
them.
Dionysus includes significant chunks of open-source code by the following
people (many thanks to them):
Jeffrey Kline (LSQR port to Python)
Bernd Gaertner (implementation of Miniball algorithm used for Cech complexes)
John Weaver (Hungarian algorithm used for Wasserstein distances)
Arne Schmitz (PyGLWidget.py)
Thank you for your
time and attention!
Title

Dionysus Slides

Uploaded by

Copyright:

Available Formats

Dionysus Slides

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Dionysus Slides

Uploaded by

Copyright:

Available Formats

A Practical Guide to Persistent

What is the homology of this point cloud?

What is the homology of this point cloud?

Squint our eyes

What is the homology of this point cloud?

Squint our eyes

What is the homology of this point cloud?

Squint our eyes

What is the homology of this point cloud?

Squint our eyes

What is the homology of this point cloud?

Squint our eyes

What is the homology of this point cloud?

Squint our eyes no natural fixed scale persistent homology

0 H(Pr1 ) H(Pr2 ) . . . H(Rn )

0 H(Pr1 ) H(Pr2 ) . . . H(Rn )

0 H(Pr1 ) H(Pr2 ) . . . H(Rn )

Squinting our eyes gives us a

Simplicial complex: collection of simplices closed under

Simplicial complex: collection of simplices closed under

Simplicial complex: collection of simplices closed under

not a simplicial simplex9 = Simplex([0,1,2,3,4,5,6,7,8,9])

two cycles are homologous

homology: count cycles up to

(w.l.o.g. assume Ki+1 = Ki + i ).

(w.l.o.g. assume Ki+1 = Ki + i ).

simplices = [([0], 1), ([1], 2), ([0,1], 3), ([2], 4), \

(w.l.o.g. assume Ki+1 = Ki + i ).

H(K1 ) H(K2 ) . . . H(Kn )

H(K1 ) H(K2 ) . . . H(Kn )

from math import sqrt

Fills f with all the simplices of the Delaunay triangulation

The data field of each simplex is set to a pair (r2 , Vor 6= ).

from math import sqrt

Fills f with all the simplices of the Delaunay triangulation

The data field of each simplex is set to a pair (r2 , Vor 6= ).

from math import sqrt

dgms = init_diagrams(p, f, lambda s: sqrt(s.data[0]))

NB: only pairwise distances matter

NB: only pairwise distances matter

NB: only pairwise distances matter

NB: only pairwise distances matter

dgms = init_diagrams(p, f, rips.eval)

|K|a1 |K|a2 . . . |K|an

|K|a1 |K|a2 . . . |K|an

Ka1 Ka2 . . . Kan

|K|a1 |K|a2 . . . |K|an

Ka1 Ka2 . . . Kan

Ka1 Ka2 . . . Kan

def pojection(points, axis = 1): # projection onto a coordinate axis

def pojection(points, axis = 1): # projection onto a coordinate axis

def pojection(points, axis = 1): # projection onto a coordinate axis

H(Xa1 ) H(Xa2 ) ... H(Xan )

H(Xa1 ) H(Xa2 ) ... H(Xan ) H(X)

H(X, Y) ' H(X w Y, w)

W (Dgm(f ), Dgm(g)) = inf kx (x)k

W (Dgm(f ), Dgm(g)) = inf kx (x)k

W (Dgm(f ), Dgm(g)) = inf kx (x)k

W (Dgm(f ), Dgm(g)) = inf kx (x)k