BEGIN:VCALENDAR
VERSION:2.0
PRODID:researchseminars.org
CALSCALE:GREGORIAN
X-WR-CALNAME:researchseminars.org
BEGIN:VEVENT
SUMMARY:Roland Bauerschmidt (University of Cambridge)
DTSTART;VALUE=DATE-TIME:20220921T140000Z
DTEND;VALUE=DATE-TIME:20220921T150000Z
DTSTAMP;VALUE=DATE-TIME:20230205T204654Z
UID:ADPS/1
DESCRIPTION:Title: Log
-Sobolev inequality for near-critical Ising models\nby Roland Bauersch
midt (University of Cambridge) as part of Abu Dhabi Stochastics Seminar\n\
nAbstract: TBA\n
LOCATION:https://researchseminars.org/talk/ADPS/1/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Domenico Marinucci (University of Rome)
DTSTART;VALUE=DATE-TIME:20220928T140000Z
DTEND;VALUE=DATE-TIME:20220928T150000Z
DTSTAMP;VALUE=DATE-TIME:20230205T204654Z
UID:ADPS/2
DESCRIPTION:Title: The
Geometry of Time-Dependent Spherical Random Fields\nby Domenico Marin
ucci (University of Rome) as part of Abu Dhabi Stochastics Seminar\n\nAbst
ract: TBA\n
LOCATION:https://researchseminars.org/talk/ADPS/2/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Reda Chhaibi (Université Paul Sabatier)
DTSTART;VALUE=DATE-TIME:20221026T140000Z
DTEND;VALUE=DATE-TIME:20221026T150000Z
DTSTAMP;VALUE=DATE-TIME:20230205T204654Z
UID:ADPS/3
DESCRIPTION:Title: Fre
e Probability for predicting the performance neural networks\nby Reda
Chhaibi (Université Paul Sabatier) as part of Abu Dhabi Stochastics Semin
ar\n\n\nAbstract\nGradient descent during the learning process of a neural
network can be subject to many instabilities. The spectral density of the
Jacobian is a key component for analyzing stability. Following the works
of Pennington et al.\, such Jacobians are modeled using free multiplicativ
e convolutions from Free Probability Theory (FPT). We make the following
contributions:\n– theoretical: refine the metamodel of Pennington et al.
thanks to the rectangular analogue of free multiplicative convolutions.\n
– numerical: present and benchmark a homotopy method for solving the equ
ations of free probability.\n– empirical: we show that the relevant FPT
metrics computed before training are highly correlated to final test accur
acies – up to 85%.\n
LOCATION:https://researchseminars.org/talk/ADPS/3/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Simona Diaconu (Stanford University)
DTSTART;VALUE=DATE-TIME:20221102T140000Z
DTEND;VALUE=DATE-TIME:20221102T150000Z
DTSTAMP;VALUE=DATE-TIME:20230205T204654Z
UID:ADPS/4
DESCRIPTION:Title: Met
hod of Moments and Edge Eigenvalues\nby Simona Diaconu (Stanford Unive
rsity) as part of Abu Dhabi Stochastics Seminar\n\n\nAbstract\nThe method
of moments is a classical technique for showing weak convergence and follo
ws a simple recipe: for any natural number m\; compute the mth moments of
the random variables of interest\, and prove they tend to the mth moment o
f the claimed limit (this works for some limiting laws\, including Gaussia
n). This approach has been prolific for universality results: for the larg
est eigenvalues of random matrices\, justify their asymptotic behavior dep
ends solely on a random variable\, show the moments of the latter depend (
asymptotically) on few moments of the former\, and use the Gaussian case t
o deduce the limiting behavior. Although Gaussianity can be relaxed consid
erably\, some constraints are indispensable: consider a real-valued Wigner
matrix with i.i.d. entries. When the fourth moment of the entry distribut
ions is infinite (heavy-tailed)\, the largest eigenvalues are known to con
verge to Poisson point processes\, whereas when it is finite (light-tailed
)\, the limits are the same as for Gaussian orthogonal ensembles. This tal
k focuses on a subfamily of edge cases\, distributions at the boundary bet
ween heavy- and lighttailed regimes\, and presents a new application of th
e method of moments\, one that allows to obtain the asymptotics of the lar
gest eigenvalues directly\, without any comparison to the Gaussian case. A
byproduct of this result is a connection between the aforementioned subfa
mily and two other families\, finite-rank perturbations of Wigner matrices
and sparse random matrices.\nThis presentation is based on https://arxiv.
org/pdf/2203.08712.pdf.\n
LOCATION:https://researchseminars.org/talk/ADPS/4/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Barbara Dembin (ETH Zürich)
DTSTART;VALUE=DATE-TIME:20221012T140000Z
DTEND;VALUE=DATE-TIME:20221012T150000Z
DTSTAMP;VALUE=DATE-TIME:20230205T204654Z
UID:ADPS/5
DESCRIPTION:Title: Coa
lescence of geodesics and the BKS midpoint problem in first-passage percol
ation\nby Barbara Dembin (ETH Zürich) as part of Abu Dhabi Stochastic
s Seminar\n\nAbstract: TBA\n
LOCATION:https://researchseminars.org/talk/ADPS/5/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Mauro Mariani (HSE University)
DTSTART;VALUE=DATE-TIME:20221109T140000Z
DTEND;VALUE=DATE-TIME:20221109T150000Z
DTSTAMP;VALUE=DATE-TIME:20230205T204654Z
UID:ADPS/8
DESCRIPTION:Title: Met
astable regimes of diffusion processes\nby Mauro Mariani (HSE Universi
ty) as part of Abu Dhabi Stochastics Seminar\n\n\nAbstract\nI will discuss
a classical example featuring a metastable behavior: finite-dimensional d
iffusion processes in the vanishing noise limit. Exponential estimates wer
e introduced fifty years ago by Freidlin and Wentzell. Recent developments
in potential theory and variational convergence allowed a refinement of t
hose results. I will focus on the non-reversible case\, with motivations c
oming from MonteCarlo methods. In collaboration with C.Landim and I.Seo.\n
LOCATION:https://researchseminars.org/talk/ADPS/8/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Alexander Stauffer (University of Bath)
DTSTART;VALUE=DATE-TIME:20221116T140000Z
DTEND;VALUE=DATE-TIME:20221116T150000Z
DTSTAMP;VALUE=DATE-TIME:20230205T204654Z
UID:ADPS/9
DESCRIPTION:Title: Non
-equilibrium multi-scale analysis and coexistence in competing first-passa
ge percolation\nby Alexander Stauffer (University of Bath) as part of
Abu Dhabi Stochastics Seminar\n\n\nAbstract\nWe consider a natural random
growth process with competition on Z^d called first-passage percolation in
a hostile environment\, that consists of two first-passage percolation pr
ocesses FPP_1 and FPP_\\lambda that compete for the occupancy of sites. In
itially FPP_1 occupies the origin and spreads through the edges of Z^d at
rate 1\, while FPP_\\lambda is initialised at sites called seeds that are
distributed according to a product of Bernoulli measures of parameter p. A
seed remains dormant until FPP_1 or FPP_\\lambda attempts to occupy it\,
after which it spreads through the edges of Z^d at rate \\lambda. We will
discuss the results known for this model and present a recent proof that t
he two types can coexist (concurrently produce an infinite cluster) on Z^d
. We remark that\, though counterintuitive\, the above model is not monoto
ne in the sense that adding a seed of FPP_\\lambda could favor FPP_1. A ce
ntral contribution of our work is the development of a novel multi-scale a
nalysis to analyze this model\, which we call a multi-scale analysis with
non-equilibrium feedback and which we believe could help analyze other mod
els with non-equilibrium dynamics and lack of monotonicity. Based on joint
works with Vladas Sidoravicius and Tom Finn.\n
LOCATION:https://researchseminars.org/talk/ADPS/9/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Mark Rudelson (University of Michigan)
DTSTART;VALUE=DATE-TIME:20230118T140000Z
DTEND;VALUE=DATE-TIME:20230118T150000Z
DTSTAMP;VALUE=DATE-TIME:20230205T204654Z
UID:ADPS/10
DESCRIPTION:Title: Ap
proximately Hadamard matrices and random frames\nby Mark Rudelson (Uni
versity of Michigan) as part of Abu Dhabi Stochastics Seminar\n\n\nAbstrac
t\nAn n by n matrix with plus-minus 1 entries which acts as a scaled isome
try is called Hadamard. Such matrices exist in some\, but not all dimensio
ns. Combining number-theoretic and probabilistic tools we construct matric
es with plus-minus 1 entries which act as approximate scaled isometries fo
r all n. More precisely\, the matrices we construct have condition numbers
bounded by a constant independent of the dimension. We will also discuss
an application in signal processing. A frame is an overcomplete set of vec
tors which allows a robust decomposition of any vector in the space as a l
inear combination of these vectors. Frames are used in signal processing s
ince the loss of a fraction of coordinates does not prevent the recovery o
f the signal. We will discuss a question when a random frame contains a co
py of a nice basis. Joint work with Xiaoyu Dong.\n
LOCATION:https://researchseminars.org/talk/ADPS/10/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Mireille Capitaine (Institut de Mathématiques de Toulouse)
DTSTART;VALUE=DATE-TIME:20230201T140000Z
DTEND;VALUE=DATE-TIME:20230201T150000Z
DTSTAMP;VALUE=DATE-TIME:20230205T204654Z
UID:ADPS/12
DESCRIPTION:Title: St
rong convergence of tensor products of independent GUE matrices\nby Mi
reille Capitaine (Institut de Mathématiques de Toulouse) as part of Abu D
habi Stochastics Seminar\n\n\nAbstract\nGiven tuples of properly normalize
d independent NxN GUE matrices (X_1\,…\,X_r) and (Y_1\,…\,Y_s)\, we pr
oved that the tuple (X_1⊗I\,…\,X_r⊗I\,I⊗Y_1\,…\,I⊗Y_s) of matr
ices converges strongly as N tends to infinity. We will present the key st
eps and ideas of the proof. Note that it was shown by B. Hayes that this r
esult implies that the Peterson-Thom conjecture is true. This is a joint w
ork with Serban Belinschi.\n
LOCATION:https://researchseminars.org/talk/ADPS/12/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Anton Nikitenko (Institute of Science and Technology Austria)
DTSTART;VALUE=DATE-TIME:20230208T140000Z
DTEND;VALUE=DATE-TIME:20230208T150000Z
DTSTAMP;VALUE=DATE-TIME:20230205T204654Z
UID:ADPS/13
DESCRIPTION:Title: Me
asuring shapes with random Delaunay mosaics\nby Anton Nikitenko (Insti
tute of Science and Technology Austria) as part of Abu Dhabi Stochastics S
eminar\n\n\nAbstract\nClassically\, we are used to working with shapes emb
edded into the square lattice. It might be anything\, a pixelated photo of
a maple leaf\, or a 3d-scanned voxelated image of a human organ. And desp
ite being simple and intuitive\, such a way of representation can totally
distort the internal geometry of the object. While it is pretty simple to
estimate the area of the leaf from a pixelated photograph\, there is no st
raightforward way to compute its perimeter: the pixel approximation of the
boundary varies dramatically\, depending on how good the object is “ali
gned” with the lattice. By changing the square lattice for an isotropic
mosaic\, we can expect that the misalignment problem will fade away\, as t
he isotropic background equalizes all directions. In the talk\, we will be
moving objects and mosaics in all possible ways\, and utilizing the archa
ic probability theory approach of counting how many points or lines are th
ere in the space\, to lead us to the precise answer of how an approximatio
n of a p-dimensional object in d-dimensional space distorts its p-dimensio
nal volume.\n
LOCATION:https://researchseminars.org/talk/ADPS/13/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Mohamed Seddik (Technology Innovation Institute\, UAE)
DTSTART;VALUE=DATE-TIME:20230215T140000Z
DTEND;VALUE=DATE-TIME:20230215T150000Z
DTSTAMP;VALUE=DATE-TIME:20230205T204654Z
UID:ADPS/14
DESCRIPTION:by Mohamed Seddik (Technology Innovation Institute\, UAE) as p
art of Abu Dhabi Stochastics Seminar\n\nAbstract: TBA\n
LOCATION:https://researchseminars.org/talk/ADPS/14/
END:VEVENT
END:VCALENDAR