Statistics for a Computational Topologist - Part I · Wasserman. All of Statistics: a Concise Course in Statistical Inference. Springer, 2010. Givens and Hoeting. Computational Statistics

Statistics for a Computational TopologistPart I

Brittany Terese FasyTA: Samuel A. Micka

School of Computing and Dept. of Mathematical SciencesMontana State University

August 14, 2018

B. Fasy (MSU) Statistics for a Computational Topologist August 14, 2018 1 / 25

Why Topological Data Analysis?“Data has shape and the shape matters.” - Gunnar Carlsson

Today, Data is high-dimensional,

HUGE, present everywhere

———————Nicolaua, Levine, andCarlsson, PNAS 2011

———————http://astrobites.com/ ———————

www.mapconstruction.org

... and needs to be summarized, analyzed, and compared!




Today, Data is high-dimensional, HUGE,

present everywhere


———————http://astrobites.com/

———————www.mapconstruction.org





Today, Data is high-dimensional, HUGE, present everywhere


———————http://astrobites.com/ ———————





What questions do we ask in data anaylsis?

Think! Write down one question (2 min)

Pair! Share with partner, and add more questions to your list (5 min)

Share! Raise hands please! (5 min)

More ideas? [email protected]


[email protected]







[email protected]







[email protected]







[email protected]







[email protected]

Data Analysis Questions

Summarize and Analyze

What is this shape?

How many components / populations?

Can we categorize? (Classification)

What are the parameters? (Inference: Point Estimation)

How far do parameters likely lie from estimates? (Confidence Sets)

Compare

Are these the same? In distribution?

Has something changed? If so, what has changed?

Which is bigger?

Can we retain the null hypothesis? (Inference: Hypothesis Testing)

What is the relationship between X and Y ? (Regression)




What is this shape? How many components / populations?




Compare



Which is bigger?










Compare



Which is bigger?










Compare



Which is bigger?










Compare



Which is bigger?










Compare

Are these the same?

In distribution?


Which is bigger?










Compare



Which is bigger?










Compare


Has something changed?

If so, what has changed?

Which is bigger?










Compare



Which is bigger?










Compare



Which is bigger?










Compare



Which is bigger?










Compare



Which is bigger?




Most Important Questions

1. Which descriptor best captures our data?

Descriptors

Confidence Sets

2. How do we measure distance between descriptors?

Distances

Clustering




Descriptors

Confidence Sets


Distances

Clustering




Descriptors

Confidence Sets


Distances

Clustering




Descriptors

Confidence Sets


Distances

Clustering




Descriptors

Confidence Sets


Distances

Clustering


Descriptors

Topological Descriptors


Descriptors

Stat Reverences

Wasserman. All of Statistics: a Concise Course in Statistical Inference.Springer, 2010.

Givens and Hoeting. Computational Statistics. Wiley, 2013.


Descriptors

Stat Slide: The Basics

Let F be a probability distribution with density f .

X ∼ F reads “X has distribution F”.

Here, X is called a random variable.

Expectation: E(X ) =∫x dF (x).

Quantile Function CDF−1(q).

−4 −2 0 2 4

0.00.1

0.20.3

0.4

−4 −2 0 2 4

0.00.2

0.40.6

0.81.0


Descriptors







−4 −2 0 2 4

0.00.1

0.20.3

0.4

−4 −2 0 2 4

0.00.2

0.40.6

0.81.0


Descriptors







−4 −2 0 2 4

0.00.1

0.20.3

0.4

−4 −2 0 2 4

0.00.2

0.40.6

0.81.0


Descriptors







−4 −2 0 2 4

0.00.1

0.20.3

0.4

−4 −2 0 2 4

0.00.2

0.40.6

0.81.0


Descriptors

Prob/Stat Slide: Descriptors and Limit Theory

Let F be some distribution.

Let X1,X2, . . . ,Xn ∼ F . (The data).

A statistic or descriptor is a function of the data:T (X1,X2, . . . ,Xn) or T (X n).

Sample average: X n = 1n

∑Xi .

Law of Large Numbers

X n converges to E(Xi ) in probability:

∀ε > 0, limn→∞

(|P(X n − E(Xi )| > ε))→ 0.

Central Limit Theorem√n(X n − E(Xi )) converges in distribution to a Normal distribution, i.e.,

sample average is approximately Normal for large enough samples.


Descriptors






∑Xi .



∀ε > 0, limn→∞

(|P(X n − E(Xi )| > ε))→ 0.




Descriptors






∑Xi .



∀ε > 0, limn→∞

(|P(X n − E(Xi )| > ε))→ 0.




Descriptors






∑Xi .



∀ε > 0, limn→∞

(|P(X n − E(Xi )| > ε))→ 0.




Descriptors






∑Xi .



∀ε > 0, limn→∞

(|P(X n − E(Xi )| > ε))→ 0.




Descriptors






∑Xi .



∀ε > 0, limn→∞

(|P(X n − E(Xi )| > ε))→ 0.

Central Limit Theorem√n(X n − E(Xi )) converges in distribution to a Normal distribution,

i.e.,sample average is approximately Normal for large enough samples.


Descriptors






∑Xi .



∀ε > 0, limn→∞

(|P(X n − E(Xi )| > ε))→ 0.




Descriptors

Data as Point Clouds


Descriptors

Data as Point Clouds

big loop

noise

pinch


Descriptors

Data as Persistence Diagrams


Confidence Sets

Confidence Sets for Persistence Diagrams:Analyzing Descriptors


Confidence Sets

Objective

To Find a Threshold

Given α ∈ (0, 1), we will find qα > 0 such that

P(W∞(D, Dn) ≤ qα) ≥ 1− α.

References

BTF, Lecci, Rinaldo, Wasserman, Balakrishnan, and Singh.Confidence sets for persistence diagrams. Annals of Stat., 2014.

Chazal, BTF, Lecci, Rinaldo, Singh, and Wasserman. On theBootstrap for Persistence Diagrams and Landscapes. Modeling andAnalysis of Information Systems, 2013.

Chazal, BTF, Lecci, Michel, Rinaldo, and Wasserman. RobustTopological Inference: Distance To a Measure and Kernel Distance,JMLR 18(159):1–40, 2018.


Confidence Sets

Objective

To Find a Threshold

Given α ∈ (0, 1), we will find qα > 0 such that

P(W∞(D, Dn) ≤ qα) ≥ 1− α.

References

BTF, Lecci, Rinaldo, Wasserman, Balakrishnan, and Singh.Confidence sets for persistence diagrams. Annals of Stat., 2014.

Chazal, BTF, Lecci, Rinaldo, Singh, and Wasserman. On theBootstrap for Persistence Diagrams and Landscapes. Modeling andAnalysis of Information Systems, 2013.

Chazal, BTF, Lecci, Michel, Rinaldo, and Wasserman. RobustTopological Inference: Distance To a Measure and Kernel Distance,JMLR 18(159):1–40, 2018.


Confidence Sets

Stat Slide: Bootstrapping

Old idiom: “pull yourself up by your bootstraps”

Want: a parameter of an unknown distribution F .

Try: estimate using empirical distribution F .

Nonparametric technique!


Confidence Sets







Confidence Sets







Confidence Sets







Confidence Sets

Bottleneck Bootstrap

We have a point cloud sample:Sn = {X1, . . . ,Xn}; Xi ∼ P.

Subsample (with replacement),obtaining: X = {X ∗1 , . . . ,X ∗b }

Compute Θ∗b(X ∗) = W∞(X ∗,Sn)using KDE or DTM.

Consider all possible outcomes:

{Θ∗b(X ∗)}X∗⊂Sn

Mimics:

{Θ(X ) = W∞(Sn,M)}Sn⊂M


Confidence Sets







Mimics:



Confidence Sets







Mimics:



Confidence Sets







Mimics:



Confidence Sets







Mimics:



Confidence Sets







Mimics:



Confidence Sets

Confidence Sets for Persistent Diagrams

Cα = {D ∈ DT : W∞(D, Dn) ≤ qα}


Confidence Sets

Confidence Sets for Persistent Diagrams

Cα = {D ∈ DT : W∞(D, Dn) ≤ qα}


Confidence Sets

Example

Birth

Dea

th

Noisy GridNoisy Grid KDE h=0.05●

●●●●●●●●

●●●●●●●●●●●●

●●●●●●●●

0.0 0.5 1.0 1.5

0.0

0.5

1.0

1.5

Death

Birth

KDE h=0.05●

●●●●●●●●

●●●●●●●●●●●●

●●●●●●●●

0.0 0.5 1.0 1.5

0.0

0.5

1.0

1.5

Death

Birth

DTM m=0.01●

●●●●●●●●●●●●

●● ●●●●●●●●●●

●●●●●●●●●●

●●●

0.05 0.10 0.15

0.05

0.10

0.15

dim 0dim 1

Birth

Dea

th

DTM m=0.01●

●●●●●●●●●●●●

●● ●●●●●●●●●●

●●●●●●●●●●

●●●

0.05 0.10 0.15

0.05

0.10

0.15

dim 0dim 1

Birth

Dea

th


Confidence Sets

Challenges

Techniques

Prove limit theorems.

Determine suitableassumptions on input.

Use the geometry of input(e.g., properties of anunderlying smoothmanifold).

Questions

These results are in thelimit. When is n big enough?

What confidence sets can weconstruct in the multi-dsetting?

What is the optimalthreshold for particularfiltrations?

Power analysis: are therejected points topologicallyinsignificant? (Type IIerrors)


Confidence Sets

Challenges

Techniques




Questions






Confidence Sets

Challenges

Techniques




Questions






Confidence Sets

Challenges

Techniques




Questions






Confidence Sets

Challenges

Techniques




Questions






Confidence Sets

Challenges

Techniques




Questions






Confidence Sets

Challenges

Techniques




Questions






Confidence Sets

Challenges

Techniques




Questions






Confidence Sets

Challenges

Techniques




Questions






Distances

Distance Measures:Comparing Descriptors


Distances

?=


Distances

?=


Distances

Distances Between Diagrams

Bottleneck d∞.

Interleaving distance.

Wasserstein dp.

Erosion distance.

Question

Can we define a centroid /Frechet mean?

arg minD

∑i

W 2∞(D,Di )

1. Turner, Mileyko, Mukherjee, and Harer. Frechet Meansfor Distributions of Persistence Diagrams. DCG, 2014.2. Munch, Tuner, Bendich, Mukherjee, Mattingly, andHarer. Probabilistic Frechet Means for Time VaryingPersistence Diagrams. Electronic Journal of Statistics,2015.


Distances


Bottleneck d∞.


Wasserstein dp.

Erosion distance.

Question


arg minD

∑i

W 2∞(D,Di )

1. Turner, Mileyko, Mukherjee, and Harer. Frechet Meansfor Distributions of Persistence Diagrams. DCG, 2014.2. Munch, Tuner, Bendich, Mukherjee, Mattingly, andHarer. Probabilistic Frechet Means for Time VaryingPersistence Diagrams. Electronic Journal of Statistics,2015.


Distances


Bottleneck d∞.


Wasserstein dp.

Erosion distance.

Question


arg minD

∑i

W 2∞(D,Di ) 1. Turner, Mileyko, Mukherjee, and Harer. Frechet Means

for Distributions of Persistence Diagrams. DCG, 2014.2. Munch, Tuner, Bendich, Mukherjee, Mattingly, andHarer. Probabilistic Frechet Means for Time VaryingPersistence Diagrams. Electronic Journal of Statistics,2015.


Distances


Bottleneck d∞.


Wasserstein dp.

Erosion distance.

Question


arg minD

∑i

W 2∞(D,Di ) 1. Turner, Mileyko, Mukherjee, and Harer. Frechet Means

for Distributions of Persistence Diagrams. DCG, 2014.2. Munch, Tuner, Bendich, Mukherjee, Mattingly, andHarer. Probabilistic Frechet Means for Time VaryingPersistence Diagrams. Electronic Journal of Statistics,2015.


Distances


Distances


Distances

Clustering


Distances

Clustering

... and Classification

Clustering (Unsupervised Learning)

Heirarchical: agglomerative or divisive.

k-means: NP-hard, so algorithms find a local minimum.

Distribution- and density-based clustering: e.g., DBSCAN.

Fuzzy clustering: membership is not binary.

Classification (Supervised Learning)

input data (training sample): D = {(Xi ,Yi )}ni=1

k-nn clustering: for new X , we predict Y by majority vote of the k nearestneighbors of the covariates (features) in D.


Distances

Clustering











Distances

Clustering











Distances

Clustering











Distances

Clustering ... and Classification










Distances

Homework!

Curate a list of topological descriptors. For each, we are looking for:

Name of descriptor.

List of distances that can be used between descriptors.

Short explanation (very short).

Reference to where first used, or a good use of it.

Pros: What is it good for?

Cons: Where / when is it insufficient?

https://github.com/compTAG/ima-multid


https://github.com/compTAG/ima-multid

Documents

Statistics for a Computational Topologist - Part I · Wasserman. All of Statistics: a Concise Course in Statistical Inference. Springer, 2010. Givens and Hoeting. Computational Statistics