© All Rights Reserved

2 views

© All Rights Reserved

- Borel Sets and Lebesgue Measure
- Vlsi Final Notes Unit3
- lab manual MCSE 101.docx
- Rudin Real Complex Solutions
- Kramers Kronig Relations Using a Pictorial Proof
- Isomorphism of Interval Valued Neutrosophic Hypergraphs
- Music Through Fourier Space
- Image-Space Modal Bases for Plausible Manipulation of Objects in Video_MIT
- series de fourier
- chap1(2)
- Lab-07
- Reducing FPGA Algorithm Area by Avoiding Redundant Computation.pdf
- 1st & 2nd Order Edge Detection
- The Subtle Art of Not Giving a F*ck: A Counterintuitive Approach to Living a Good Life
- Sapiens: A Brief History of Humankind
- Hidden Figures: The American Dream and the Untold Story of the Black Women Mathematicians Who Helped Win the Space Race
- Shoe Dog: A Memoir by the Creator of Nike
- The Unwinding: An Inner History of the New America
- The Little Book of Hygge: Danish Secrets to Happy Living
- Never Split the Difference: Negotiating As If Your Life Depended On It

You are on page 1of 5

Dominik Flegel

1 Introduction

So far, we have seen a bunch of properties and generalizations of exchange-

able random variables (many of them building on the basic theorem of de

Finetti). These highly technical results can be applied to obtain non-trivial

results regarding random edge colorings of suciently regular hypergraphs

on a countably innite vertex set S .

2 Framework

We start by establishing the setting considered during this presentation.

Unless stated otherwise, let S be a countably innite set.

Denition 2.1 (Hypergraphs). A hypergraph on S is a set G P(S). It

is k-uniform if e G |e| = k, and complete k-uniform if |e| = k

e G. By abuse of notation, we denote the latter as S

.

k

Note that in contrast to the usual case k = 2 with nite vertex set,

we describe our hypergraphs implicitly by specifying the edge set (which is

eectively nothing else than an array), dropping S in the denition. Based

on this, colored hypergraphs are now dened in the most intuitive way:

Denition 2.2 (Colored hypergraphs). For a nite set K , a K-colored,

k-uniform hypergraph on S is a map H : Sk K . A K-colored, k-

under coordinate permutations of S k . Denote K as the palette of a coloring

H.

It's not unusual to extend this denition allowing standard Borel spaces

as palettes. However, as the set K corresponds to eligible colors of the

kedges, this nite restriction will do for our needs.

1

3 Structure of random colorings

Building on a denition installed in a previous talk, we can give a brief

description of exchangeability in our new context.

Denition 3.1 (Exchangeable hypergraphs). A random K-colored, k-uniform

hypergraph H (often identied with its law ) is called exchangeable if the

array (He )e(S ) is (jointly) exchangeable (as dened in Raaella's presenta-

k

tion, i.e. if the coloring distribution does not depend on vertex relabeling).

In order to consolidate the numerous denitions to this point, we state

the most basic (non-trivial) examples of exchangeable hypergraphs:

Example 1 (H-sampling random hypergraph). Let H be a xed K-colored,

k-uniform hypergraph with loops on a nite set V . Obtain a random K-

colored, k-uniform hypergraph on S by sampling for each s S a vertex

vs V independently and uniformly at random and setting

S

of the vertices has no eect on a uniform sampling. However, we can not

expect all exchangeable hypergraphs to be obtained by this procedure (as

seen in the following second example).

Example 2 (Intermediate step sampling). Consider . A corresponding

S

3

exchangeable {0, 1} coloring can be constructed in two steps:

For each u S2 , sample independently and uniformly at random a

Identify each e S3 with its three subsets u1 , u2 , u3 S2 of order

H :{0, 1} {0, 1} {0, 1} {0, 1}

(ku1 , ku2 , ku3 ) 7 1P3 kui >0 .

i=1

The general idea behind this extended kind of sampling is to allow for

more sources of randomness by conditioning on the information given by

colorings of lower ranked edges. In this case, the color of e simply is 1 if

at least one of the contained 2-edges is colored 1. For higher k, there is an

obvious extension of this technique by iterating the procedure, resulting in

a sampling construction passing through each rank in between.

Proposition 3.2. There is no H (as described in example 1) such that

= H , where is the intermediate step sampling random hypergraph.

2

Sketch of Proof. Consider a partition of S into two innite subsets (i.e. S =

2 ) and assume there is any H-sampling random hypergraph such that

S1 S

H = . For each s S1 , dene the -algebras T (s) := ((H(e))e({s}S2 ) )

3

and T := sS1 T (s). It can be shown that under the assumed vertex-

W

sampling-only structure of the random variables (H(e))e(S ) are relatively

3

independent over T . On top of that, if |e1 e2 | = 2 for e1 , e2 S3 (i.e. if

they share exactly one 2-edge) then the joint behavior of H(e1 ) and H(e2 )

is already independent from T , this yields a contradiction as in that case

25 7

{H(e1 ) = 1, H(e2 ) = 1} = 6= ( )2 = {H(e1 ) = 1} {H(e2 ) = 1}.

32 8

formulations) to actually fully understand the structure of exchangeable col-

ored hypergraphs, as the quintessence is already captured in the intermediate

step sampling.

Denition 3.3 (Ingredients). A sequence of ingredients is a sequence

(Zi )ki=0 of standard Borel spaces with Zk = K , a probability measure 0

on Z0 and probability kernels

i

(i) (i1)

Pi : Z0 Z1i Z2 2 ... Zi1 Zi

action of Sym(i) on the domain.

Denition 3.4 (Standard recipe). For a sequence of ingredients

(Z0 , 0 ), (Z1 , P1 ), ..., (Zk1 , Pk1 ), (K, Pk )

exchangeable colored hypergraph:

Sample z Z0 according to 0 .

to P2 (z , zs1 , zs2 , ).

..

.

2 k1

3

Moreover, we dene that the ingredients yield upon following the

standard recipe for the induced law on K ( k ) .

S

The central result is that all exchangeable colorings are fully char-

acterized by following the standard recipe for an appropriate sequence of

ingredients:

Theorem 3.5 (Structure of uniform exchangeable hypergraph colorings).

For any k-uniform exchangeable colored random hypergraph there is some

sequence of ingredients

A full proof from scratch of this is assertion can be found in Tim Austin's

"On exchangeable random variables and the statistics of large graphs and

hypergraphs ". Not only is it very similar to the presented bits of the Rep-

resentation theorem for (jointly) exchangeable arrays, both statements are

in fact equivalent. To see the connection between them, we shall give an

explanation on how to translate one representation into another.

Proof. Let be such an exchangeable coloring and f : [0, 1]2 K middle-

k

tion, a collection of ingredients which yields is given by setting:

Zi := [0, 1] for i k 1

, ze = (zb )b2e \e for i k 1

i

Pk (ze , ) := E[1{f (ze ,t)} |z] for all e Sk , ze = (zb )b2e \e where the

k

immediate follows from the corresponding property of f .

A comment on the reverse implication :

As we already know from the representation theorem, every which is in-

duced by following the standard recipe for some sequence of ingredients is

represented by an appropriate function f , simply because it is (jointly) ex-

changeable by construction. The determination of this function from the

ingredients however is not as immediate as the other way round. An insight

is given by the Lemma below.

Lemma 3.6 (Noise outsourcing). Suppose that X and Y are standard Borel

spaces, that is a probability measure on X and that P : X Y is a

probability kernel. Let be the Lebesgue measure on [0, 1]. Then there

1

4

is some Borel measurable map f : X [0, 1] Y such that, endowing

X [0, 1] with the product measure 1 , the kernel P (x, ) is a version of

the conditional distribution of f (x, ) given the rst coordinate x:

Proof. Identify Y with a Borel subset of [0, 1], the function dened by

The Lemma tells us, that the randomness in the choice of y Y according

to P (x, ) for x X can be correctly represented by rst choosing indepen-

dently and uniformly a value for some "noise parameter" t [0, 1], and then

choosing y according to a deterministic Y -valued map on (x, t). By identi-

fying the Zi with intervals [0, 1] inductively and replacing the corresponding

kernels up to Pi with deterministic functions fi (using outsourced noise pa-

rameters) step by step, the representing function f can be constructed from

the sequence of ingredients.

4 Outlook

The notion of hypergraph colorings can be extended and applied in various

ways, a few of them are branched shortly below.

As we are considering complete hypergraphs , some results can be

S

k

carried over to the study of random hypergraphs, by simply setting

K = {0, 1} with the interpretation of an edge being present or absent

depending on its color.

While performing the standard recipe, every edge up to a given size

is assigned to some value in a standard Borel space Zi under some

i , which is not stored in the nal result. With use of a generalized

palette K = (Ki )ki=0 (where Ki nite) and Borel maps i : Zi Ki

one can

obtain similar structural results for an exchangeable coloring

of k

S

:= ki=0 Si .

(consisting of all possible k-edges with endpoints in k distinct partition

classes) one can obtain a similar structural result (using a collection

of ingredients) equivalent to the Representation theorem for separately

exchangeable arrays.

- Borel Sets and Lebesgue MeasureUploaded byread41
- Vlsi Final Notes Unit3Uploaded byRohitParjapat
- lab manual MCSE 101.docxUploaded byJuan Jackson
- Rudin Real Complex SolutionsUploaded byAYAN
- Kramers Kronig Relations Using a Pictorial ProofUploaded byAnand Krish
- Isomorphism of Interval Valued Neutrosophic HypergraphsUploaded byMia Amalia
- Music Through Fourier SpaceUploaded byFederico Larrosa
- Image-Space Modal Bases for Plausible Manipulation of Objects in Video_MITUploaded byAshwin Varma
- series de fourierUploaded byartovolasti
- chap1(2)Uploaded byMichał Gromisz
- Lab-07Uploaded bySobia Shakeel
- Reducing FPGA Algorithm Area by Avoiding Redundant Computation.pdfUploaded byBoppidiSrikanth
- 1st & 2nd Order Edge DetectionUploaded byprakash_oxford

- AnazasisUploaded byFenix862
- Rajib Mall Lecture NotesUploaded byAnuj Nagpal
- Pressure Drop for Flow Through Packed Beds.pdfUploaded byrohl55
- St Dimas Project PresentationUploaded byKarrizzmatic
- Chroma-Tek CTUUploaded byErisie
- Social Research in Developing NationsUploaded byManish Naithani
- Parametric Curve and Surface ModelingUploaded byCandice Canoso
- Tutorial on Factored Language ModelsUploaded byvbsowmya
- Storage and Measurement in crude oil E&PUploaded bySaurabh Sharma
- The Early Chicago Tall Office Building: Artistically and Functionally ConsideredUploaded byDorian Vujnović
- morphological image processingUploaded byappuchoco
- GBPPR 'Zine - Issue #101Uploaded byGBPPR
- Intro to HydrologyUploaded byFred Enea
- cast-256Uploaded byFull Name
- 8_REB500_commisioningUploaded byRetratosDeMiVida
- 2011 OperationUploaded byDaniel Tapia
- 09 September 1991Uploaded byMonitoring Times
- 6.- Tx - Rx FMUploaded byart_ag7
- Solid Rivet InformationUploaded byKukyong Lee
- Solar Water Heating detailed reportUploaded byRahul Rajpal
- Hettich_ProjectSolutions_0412_EN.pdfUploaded byaudithan
- ch14Uploaded bySadaf Khan
- How to Decipher the Most Challenging RCs_ _ GP SPEAKSUploaded bydavidcooper025
- edt 313 - animal track lesson sketch -micks malewitzUploaded byapi-455052049
- maxdbUploaded byPriya Sekar
- Impulsive Hyperbolic Injection from a Circular Park Orbit (Fortran)Uploaded bycdeaglejr
- pp 533-541 HEAT AND MASS TRANSFER IN A SHRINKING CYLINDER DRYINGUploaded byelatani
- Creating Psiphon Cmd Batch FileUploaded bymmakd
- FS-6-episode-4Uploaded byszarielle yumiko
- Saturn automatic transmissionUploaded byDenis