This is the homepage for the UC Berkeley Student Probability Seminar,
a venue for graduate students
in the departments of mathematics, statistics, and others to
study aspects of modern probability theory.
We meet on Wednesdays in Evans 891 from 2:00 PM - 3:00 PM
Organizers:
Ella Hiesmayr
and
Adam Quinn Jaffe.
The topic for this semester's seminar is the large deviations on random
graphs. Near the beginning of the semester we will primarily follow standard
sources on classical large deviations theory, like
Dembo and Zeitouni [DZ] and
den Hollander [dH],
and in the middle of the semester we will follow some standard texts on
random graphs and graph limits like
Lovasz [L].
Our general structure will follow the textbook by
Chatterjee [C].
The list of talks and abstracts can be found below.
- 24 August,
Initial meeting.
- A casual first meeting to get to know each other and
decide on a topic for the semester.
- 31 August, Ella Hiesmayr,
Introductory Talk.
- We introduce the basic questions motivating the theory
of large deivations for random graphs. Specifically, we describe how
the fundamental ideas of classical large deviations theory do not apply
to the setting of random graphs, because of the lack of independence
and linearity.
- 7 September, Adam Quinn Jaffe,
Crámer's Theorem, DZ Chapter 2.2 and dH Chapter I.3
- In this talk we will prove a simple but important
result of large deviations theory called Crámer's theorem.
Roughly speaking, it states that empirical averages of i.i.d. random
variables concentrate exponentially fast around their common mean,
and that the exponential rate of convergence is related to the
Fenchel-Legendre transform of their common logarithmic moment generating function.
- 14 September, Neo Yin,
Proof of Sanov's Theorem, DZ Chapter 2.1 and dH Chapter I.4
- This week we are going to prove Sanov's Theorem
on a discrete finite alphabet. Sanov’s Theorem establishes the
relative entropy (a.k.a. Kullback-Leibler divergence) as the
large deviation rate function for the random empirical measure of
iid sampling.
- 21 September, Rikhav Shah,
Probabilists love them: Upper bound the rate function using any
one of these three easy tricks, C Chapter 4 and DZ Chapter 4.2
- We'll start by taking a step back to get an intuitive
grasp on what the technical definitions of "large deviation principle"
and "good rate function" really are. This will lead directly to our
first trick: bounding the rate function locally is enough to bound
it globally. Then, we generalize the classic Chernoff inequality
to arbitrary vector spaces using the Fenchel-Legendre transform,
giving a bound on the rate function that is often optimal.
Finally, we show how to transfer LDPs from one space to another;
combined with Sanov's theorem this immediately gives the rate
function for many distributions of interest.
- 28 September, Karissa Huang,
An Introduction to Graphons, C Chapters 3.1, 3.2, and 3.3
- In this talk we summarize some basic results from
graph limit theory. In particular, we will introduce graphons and
homomorphism densities. Then, we will equip the space of graphons
with an appropriate metric (the cut metric), which induces a metric
on the space of equivalence classes of graphons. These results will
allow for the rigorous analysis of limits of finite graph
sequences.
- 05 October, Yang Chu,
Samplings of graphons, L Chapters 10.1, 10.2, 10.3, and 10.4
- We give two schemes of sampling of graphons. One can
either view graphons as weighted graphs with nodes as [0,1], or,
roughly speaking, a collection of clusters of nodes with random
bipartite structure between them. We also investigate finite-sample
concentration of homomorphism densities under sampling visa standard
concentration of measure methods, and give estimates of cut metric
based on these sampling schemes.
- 12 October, Neo Yin,
Compactness of graphon space, L Chapters 9.1, 9.2, and 9.3
- In this talk, we will prove that the (quotient)
space of graphons equipped with the cut metric is compact,
using the regularity theorems.
- 19 October, Gabriel Ramirez Raposo,
Some properties of the rate function, C Chapters 5.1 and 5.2
- We introduce the rate function used in large deviation
bounds for random graphs. We will show a nice representation of the
rate function in terms of a supremum of operators in the set of L2
functions and prove that the rate function is lower semicontinuous
in the space of graphons.
- 26 October, Zack McNulty,
The large deviations upper bound, in the weak topology, C Chapter 5.3
- In this talk we prove a large deviation upper bound
for the probability measure on the space of graphons induced by the
Erdős–Rényi model. We show on weakly closed sets that a natural
extension of the binary relative entropy function acts as the
desired rate function with rate 2/n^2. As the cut topology on W
acts as an extension of the weak topology, this result will be a
stepping stone towards proving a similar result on the quotient
space of graphons.
- 02 November, Mriganka Basu Roy Chowdhury,
The large deviations upper bound, in the cut topology, C Chapter 5.4
- Last week we saw the proof of the upper tail in the
weak topology. This week we extend the proof to the cut topology
on the graphon space, with the same rate function. This is (half of)
the LDP result proved by Chatterjee-Varadhan in their seminal
2011 paper.
- 09 November, Jorge Garza Vargas,
The large deviations lower bound, C Chapter 5.4
- We will prove the lower bound in the LDP for the
Erdős–Rényi model on the graphon space equipped with the cut
metric (the rate function will be the usual one). As in last talk,
the idea is to reduce the problem to proving a local statement for
any graphon h in W. Then, the key idea in showing the local
statement will be to tilt the Erdős–Rényi measure towards h in
some judicious manner, and showing that the relative entropy of
this new measure with respect to the Erdős–Rényi measure, in the
limit, is precisely the rate function at h.
The remainder of the talks were canceled in solidarity with the
academic workers' strike over contract negotiations. Thanks for
the great semester, everyone!