Faculty of Natural, Mathematical, and Engineering Sciences

Department of Mathematics

PhD Projects in Mathematics

Applied Mathematics Research: Disordered Systems/Financial

Mathematics/Probability/Theoretical Physics MPhil/PhD Projects
Project: Kasteleyn theory
Supervisor: Sam Johnston
Group: Probability
Summary: Kasteleyn theory is a tool used by probabilists to obtain explicit solutions for the partition
functions and correlation functions for models of random surfaces and random tilings. In this project we
use a new coordinate system to study generalisations of these models in higher dimensions, making
connections with areas of probability, algebra, and statistical physics.

Project: Microstructure of the Kerr black hole

Supervisor: Dionysios Anninos
Group: Theoretical Physics
Summary: We will apply tools from modern theoretical physics to assess mathematical properties of highly
spinning black holes. Such highly spinning black holes have been observed in our own Universe and display
remarkable geometric features.

Project: Cylindridcal Levy processes

Supervisor: Markus Riedle
Group: Probability/ Financial Mathematics
Summary: Cylindridcal Lévy processes are a new approach to model random perturbations of complex
dynamical systems such as partial differential equations. It has already been observed that these cause
completely different behaviour of the dynamical system than if perturbed by a classical stochastic process. In
this project we investigate one of the many aspects which are unknown for cylindrical Levy processes, such
as stability, ergodicity etc.

Project Title: Equilibrium between providers and takers of liquidity in order driven markets
Supervisor: Leonardo Sanchez Betancourt
Group: Financial Mathematics
Summary: We search for optimal trading strategies between takers and makers of liquidity with transaction
costs. We aim to contribute to the literature around stochastic games for algorithmic trading.
Project: Lorentzian and Euclidean Conformal Field Theory
Supervisor: Petr Kravchuk
Group: Theoretical Physics
Summary: Conformal field theories have many facets, and appear in physics in both Lorentzian (or realtime),
and Euclidean (or statistical) versions. From the theory perspective the Lorentzian point of view brings
powerful analytical insights, while the Euclidean one allows for a precise numerical analysis. You will study
the interplay between these two worlds to obtain new results relevant both in condensed-matter and high-
energy physics

Project: Entanglement and universality in inhomogeneous and out-of-equilibrium many body quantum
systems
Supervisor: Paola Ruggiero & Benjamin Doyon
Group: Disordered Systems
Summary: The project lies at the intersection of several different fields: from statistical physics to quantum
information, from high energy to condensed matter physics. The common denominators are the
phenomenon of quantum entanglement and the out-of-equilibrium behaviour of many-body quantum
systems, particularly in low dimensional quantum many body systems. While the emergence of universality
is rather well established for systems which are homogeneous and are at equilibrium, the same is not true for
inhomogeneous systems, both in and out of equilibrium. Those, on the other hand, are the rule rather than
the exception in the context of quantum experiments. Different new theories and tools are currently under
study to investigate what remains of the aforementioned universality in such more complicated situations,
and the projects aims at going deeper in this direction, mainly using the “lens” of quantum entanglement.

Project: Random models for arithmetic correlations

Supervisor: Aled Walker
Group: Number Theory
Summary: The project lies at the intersection of several different fields: number theory, harmonic analysis,
probability, and additive combinatorics, with even some side connections to mathematical physics. The goal
is to investigate the gap distributions of certain types of arithmetic sequences (dilated squares, dilated primes,
dilated integers, etc.), with the aim to understand how closely these distributions agree with a simple random
model, formed by replacing the arithmetic sequence with independent uniform random variables. It has been
conjectured, for example, that the dilated squares can indeed be modelled by a random sequence in this way
(having so-called 'poissonian gap distribution'). This is turn is connected to the phenomenon of quantum
chaos, which concerns the gap distribution of the eigenvalues of certain Hamiltonians. Additive
combinatorics also plays a role: recent work has found new connections between correlation functions
(certain objects associated to gap distributions) and the so-called 'additive energy' of the underlying
sequence. The exact direction of the project would be determined by the student, but a possible first move
concerns establishing a 'density threshold' for the triple correlation function: given a randomly chosen
sequence of integers (with a certain density), can one establish when the triple correlations will be
poissonian?

2
Project: The mathematics of adaptive, stable and robust Artificial Intelligence
Supervisor: Ivan Tyukin
Group: Disordered Systems
Summary: Recent years have seen explosive growth in the applications of Artificial Intelligence (AI).
Notwithstanding significant successes of the new technology, empirical evidence points to numerous
examples of errors in the decisions of state-of-the-art AI systems. Current theoretical works confirm that
modern design practices and classical mathematical frameworks supporting these could suffer from various
fundamental shortcomings including the typicality of adversarial attacks, susceptibility to backdoors and
stealth attacks, instabilities, and hallucinations. At the same time, there is a plethora of under-explored and
seemingly controversial phenomena, such as learning from scarce yet high-dimensional information, which
has been broadly observed experimentally in large-scale AI systems but whose complete mathematical
understanding is yet to be developed. Other factors hindering the performance of current AI systems include
the need to adapt to concept drifts and unforeseen changes in operational conditions. The project aims to
explore and systematically assess major causes of AI errors and instabilities and develop appropriate
mathematical foundations enabling the creation of stable and robust Artificial Intelligence capable of dealing
with errors and concept drifts. The successful candidate will work alongside a team of early career
researchers funded by the UKRI Turing AI Acceleration Fellowship (EP/V025295/2) and will benefit from
the team’s established relationships with industrial partners, academic networks (Turing AI Fellows,
Trustworthy Autonomous Systems Node in Verifiability EP/V026801/2), and other collaborators in
academia, healthcare, defence & security.

Pure Mathematics Research MPhil/PhD Projects

Project: Foliations of 3-dimensional manifoldsSupervisor: Mehdi Yazdi
Group: Geometry
Summary: The student is expected to work in low-dimensional topology; in one of the areas of foliations, 3-
dimensional manifolds, knot theory, mapping class groups of surfaces, or related subjects. Below some of
these terms are briefly explained: A foliation of a 3-dimensional manifold is a decomposition of it into disjoint
surfaces that locally fit together like a stack of papers. Examples of singular foliations in one dimension lower
are marbleised papers and tree rings. The student can for example study an important class of foliations called
taut foliations, which has applications to knot theory. A knot is an embedding of the circle in a 3-dimensional
manifold, considered up to isotopy. In other words, we are allowed to deform the knot without the knot
crossing itself. Knots have been extensively studied both for their own beauty, and as an important class of 3-
dimensional manifolds (the complement of the knot). The mapping class group of a surface is the group of
homeomorphisms of the surface considered up to isotopy. This is an infinite group and has applications in
various parts of geometry and topology including hyperbolic geometry of surfaces and the topology of 3-
dimensional manifolds.

3
Project: Analytic aspects of L-functions and automorphic forms
Supervisor: Stephen Lester
Group: Number Theory
Summary: L-functions and automorphic forms appear prominently in modern number theory and also arise
in other areas of mathematics. The analytic properties of these functions are closely related to arithmetic
problems involving the distribution of the prime numbers, lattice points, and structure of elliptic curves. For
instance, Millennium prize problems such as the Generalised Riemann Hypothesis and Birch and
Swinnerton-Dyer Conjecture are directly related to the theory of these functions. This project aims to
further understand the analytic behaviour of L-functions and automorphic forms, with applications to
problems arising in number theory or mathematical physics. Specific problems will follow the interests of the
student and possible topics include: the distribution of zeros of Hecke cusp forms, moment estimates for
families of L-functions, and the behaviour of Fourier coefficients of automorphic forms.

Project: Birational Geometry and foliations

Supervisor: Calum Spicer
Group: Geometry (pure mathematics)
Summary: The goal of this project is to explore cutting edge applications and interactions between birational
geometry and holomorphic foliation theory. There is a specific focus on the use of the Minimal Model
Program and related techniques as applied to questions in local and global geometry of algebraic varieties and
foliations. The study of the Minimal Model Program is one of the main research directions in modern
algebraic geometry, and in recent years has seen many significant developments. Among these is a new
focus on relations between the Minimal Model Program and foliations, which has resulted in applications of
ideas and techniques of the Minimal Model Program to a wide range of novel situations.

Statistics Research MPhil/PhD Projects

Project: Multi-Objective Optimal Design for Big Data
Supervisor: Kalliopi Mylona and Steven Gilmour
Group: Statistics, Data Analysis, Machine Learning
Big data sets typically arise from observational data and contain many more observations than are needed to
draw inferential conclusions about important research questions. Analysing full datasets can be challenging
due to their size and unrewarding due the vast amount of redundant information available. One approach to
dealing with such datasets, which has received some attention recently, is to use the theory of optimal design
of experiments to select maximally informative subsamples of the data (Deldossi & Tommasi, 2021). Like
most work on optimal design, this emphasises D-optimality, which implies that point estimation of the
parameters of a given model is the main inferential objective. Of course, many studies have more complex
and diverse objective than this. There is currently a large EPSRC-funded grant at King’s on Multi-Objective
Optimal Design of Experiments, on which 3 academic staff and 2 post-doctoral researchers are working. The
aim of this PhD project is to adapt some of the methods of multi-objective optimal design being explored to
the subsampling of big data sets. One of the challenges will be in adapting methods which require estimation
of “pure error”, which can be ensured in experimental studies by replication, to observational data where
there might be no exact replicates.
4
Project: Hierarchical models for function-valued traits in quantitative genetics
Supervisor: Davide Pigoli
Group: Statistics, Mathematical Modelling, Applied Statistics.
Summary: Some biological traits, such as growth trajectories, have a functional nature and are known as
function-valued traits (Kingsolver et al., 2001). These traits can be assessed for continuous genetic variation
using a quantitative genetic approach. In quantitative genetics, one of the goals is to decompose the genetic
(i.e. due to the genes of the individuals) and the environmental (i.e. due to the influence from the external
environment). While some works exist on the analysis of function-valued traits using mixed effects models
(Meyer, 1998; Vazquez et al., 2010), the field could potentially benefit from the use of modern functional
data analysis techniques, such as the ones in Behseta et al. (2005), Scarpa & Dunson (2009) or Scheipl et al.
(2015). The goal of the project is to extend these techniques to the case of quantitative genetics experiments,
where individuals have a known relationship matrix given by their genealogy.

Project: Supersaturated multi-stratum experiments

Supervisor: Kalliopi Mylona
Group: Statistics, Applied Statistics, Engineering Mathematics, Agricultural Sciences
Summary: The objective of this project is to develop a new general methodology for setting up and analysing
informative experiments with both restricted randomisation and a large number of factors. As designs with
restricted randomisation are often much more cost-efficient than completely randomised designs, they are
extremely popular in industry for quality improvement experiments and for experimenting with new
products or processes. Multi-stratum designs with more than two strata are very effective in reducing the
cost of an experiment in the presence of different levels of hard-to-change factors and/or of multi-stage
processes. In addition, supersaturated designs (SSDs) is a large class of factorial designs which can be used
for screening out the important factors from a large set of potentially active variables. The huge advantage of
these designs is that they reduce the experimental cost drastically, but their critical disadvantage is the
confounding involved in the statistical analysis. The goal is to combine the multi-stratum designs (more than
two strata) with the supersaturated designs. This will give the flexibility to the experimenters to use the cost-
efficient SSDs under restricted randomised situations that appear very often in industrial experimentation.
The combination of these two well-known classes of designs is a relatively unexplored research area (see,
Koh et al. [2013], Lee et al. [2009], Chatterjee et al. [2018]). General construction and analysis methods for
supersaturated multi-stratum designs with more than two strata will allow experiments that can extract the
maximum information from the data with the minimum time and cost constraints.

Project: Statistical analysis of spatially and/or temporally dependent manifold-valued data.

Supervisor: Davide Pigoli
Group: Statistics, Stochastic Processes, Mathematical Modelling.
Summary: There is a growing interest in the statistical analysis of data that take value on smooth manifolds
(3D rotations, shapes, directions, positive definite matrices, points on a circle or a sphere, constrained curves
to mention just a few examples). The goal of the project is to develop a framework to analyse manifold-
valued data that are observed over time and/or space, example including molecular dynamics simulations of
protein folding in medicinal chemistry (temporally dependent), permeability tensors in mining engineering
(spatially dependent) or coordinates (latitude and longitude) of earthquakes aftershocks in earth science
5
(spatio-temporally dependent). Possible lines of research include (but are not limited to): - modelling of
stochastic dependence in these non-linear spaces, where usual concepts of co-variability are not valid. -
extension to dependent samples of existing techniques for dimensional reduction (Principal Geodesic
Analysis, Principal Flows) and centerpoint estimation (Fréchet mean) for manifold-valued data. -
multiresolution analysis and prediction for spatio-temporal manifold-valued processes.

Project title : The geometry of probability distributions in computational statistics and machine
learning
Supervisor: Nikolas Nüsken
Group: Statistics
The process of learning can be understood as transforming one probability measure into another, guided by
available data. In order to analyse and construct computational learning routines, three perspectives haven
proved especially powerful: First, such transitions can be described by the dynamics of interacting particle
systems, often using stochastic and/or partial differential equations. Second, one can (heuristically) view the
set of probability distributions as an infinite-dimensional Riemannian manifold, in which the learning
transition evolves as a curve. Algorithmic efficiency can then directly be linked to curvature properties and
their bearings on Riemannian gradient flows and geodesics. Third, desirable transitions can be obtained as
solutions to suitable optimal control or optimal transport problems. The aim of this project is to contribute to
a deeper understanding of the connections between these three approaches and to leverage those to develop
novel and innovative methodology for data science. On the theoretical side, a possible focus could be on the
formulation of evolution equations of second order or with memory (akin to Hamiltonian dynamics). On the
applied side, a major challenge is the seamless combination of large amounts of data with models coming
from science, economics or engineering.

Project: Changepoint detection in Non-Euclidean spaces

Supervisor: Davide Pigoli
Group: Statistics, Stochastic Processes, Applied Statistics.
Summary: This project is concerned with the development and analysis of cumulative sums statistics
(Barnard, 1959) to detect changepoints in sequences where the observations belong to a non-Euclidean
space (Chen, 2019). Examples include time series of directional data (e.g., wind directions), time series of
correlation matrices (e.g., correlation matrices between electroencephalogram signals) or dynamical
networks (e.g., social networks or trade networks). Cumulative sum statistics can be defined based on a
metric in the non-Euclidean space and there are already some results for data on a circle (Lombard et al.,
2020). The purpose of this project is to extend this approach to more general non-Euclidean spaces, and
analyse the performance of changepoint detection procedures based on these statistics both theoretically and
in practice.

Project: Statistical analysis of supersaturated split-plot experiment

6
Supervisor: Kalliopi Mylona
Group: Statistics, Applied Statistics, Data Analysis, Mathematical Modelling, Engineering Mathematics,
Agricultural Sciences
Summary: The objective of this project is to develop a new general methodology for analysing informative
experiments with both restricted randomisation and a large number of factors. Split-plot designs are very
effective in reducing the cost of an experiment in the presence of hard-to-change factors and/or of two-stage
processes. In addition, supersaturated designs (SSDs) is a large class of factorial designs which can be used
for screening out the important factors from a large set of potentially active variables. The huge advantage of
these designs is that they reduce the experimental cost drastically but their critical disadvantage is the
confounding involved in the statistical analysis. The combination of split-plot designs with supersaturated
designs is a relatively unexplored research area, and the statistical analysis of data from this type of
experiments is still not efficient. In this project, the use of Bayesian analysis methods, will be studied.
Gilmour and Goos (2009) recommended a Bayesian analysis, using an informative prior distribution for the
whole plot variance component and implement this by using Markov chain Monte Carlo sampling. The
designs considered in this project have an additional complication compared to the designs used in Gilmour
and Goos (2009) and this is supersaturated property. We are planning to study and adapt several statistical
analysis methods that had been used in the literature of supersaturated designs, such as Dantzig selector
(Candes and Tao, 2007), the LARS (Efron, Hastie, Johnstone and Tibshirany,2004) methods and the
Bayesian methods: SSVS ( Chipman, Hamada and Wu, 1997) and Bayesian LASSO (Park and Casella,
2008).

Project: Stein thinning in practice, with application to cardiac modelling

Supervisor: Marina Riabiz
Group: Statistics
Summary: Markov Chain Monte Carlo (MCMC) methods form the main approach to sample from
probability distributions with intractable normalizing constants, such as the parameter posterior in Bayesian
computation. Whilst MCMC samples converge in the limit to the target distribution, in many realworld
applications it is possible to draw only a finite number of samples, due to restrictions in computing budget. In
this framework, Stein thinning is a recent method developed to optimally select a set of samples from an
MCMC output of fixed length, by minimizing a measure of discrepancy between the empirical
approximation produced and the target. This PhD project aims at exploring different research questions that
are still open, in order to produce a robust version of the algorithm, including the selection of a weighted
kernel function that jointly guarantees weak convergence and convergence of certain moments of the
distribution, based on related work on control variates and optimization of the power of goodness of fit test.
The project will be applied to challenging Bayesian inverse problems arising in cardiac electrophysiology.

Project Title: Bayesian learning for partially specified models

Supervisor: Yu Luo
Group: Statistics
Summary: Bayesian inference shifts from unknown deterministic quantities to studying distributions, which
has been proven increasing powerful in a large spectrum of applications over the decades. The posterior
inference is often thought as a procedure to update prior beliefs in light of the data. However, in the usual
Bayesian setting where a full probabilistic model is required. However, recently some researchers in

7
statistical learning have investigated extensions of the Bayes paradigm to the data-driven procedure, which
one does not make any assumption on the probabilistic model, but rather relies on an agnostic measure
between the model prediction and outcome, resulting a partially specified model. These extensions no longer
abide by the canonical Bayesian rules and may be harder to interpret by practitioners. Therefore, the overall
goal of this project is to help develop a mature general Bayesian decision-making machinery for models
which are partially specified. This methodology advancement has the potential to apply in various fields,
such as machine learning and causal inference.

Project Title: Estimation and testing for extremes of spherical data

Supervisor: Claudia Neves
Group: Statistics
Summary: Extreme value theory (and related statistical methodology) is concerned with the definition,
characterisation and estimation of tale-telling characteristics of extreme and rare events, whilst accounting
for the possible dependence between them. These could be, for example, strong wind and heavy rainfall, or
strong demand for electric energy and little solar radiation. It is the primary aim of this project to develop
statistical inference methodology for assessing and estimating changes in the intensity of extreme events on a
surface area, thus extending current methods considerably, in an actionable way.

Project Title: The geometry of probability distributions in computational statistics and machine
learning
Supervisor: Nikolas Nüsken
Group: Statistics, Disordered Systems, Analysis, Geometry
Summary: The process of learning can be understood as transforming one probability measure into another,
guided by available data. In order to analyse and construct computational learning routines, three
perspectives haven proved especially powerful: First, such transitions can be described by the dynamics of
interacting particle systems, often using stochastic and/or partial differential equations. Second, one can
(heuristically) view the set of probability distributions as an infinite-dimensional Riemannian manifold, in
which the learning transition evolves as a curve. Algorithmic efficiency can then directly be linked to
curvature properties and their bearings on Riemannian gradient flows and geodesics. Third, desirable
transitions can be obtained as solutions to suitable optimal control or optimal transport problems. The aim of
this project is to contribute to a deeper understanding of the connections between these three approaches and
to leverage those to develop novel and innovative methodology for data science. On the theoretical side, a
possible focus could be on the formulation of evolution equations of second order or with memory (akin to
Hamiltonian dynamics). On the applied side, a major challenge is the seamless combination of large amounts
of data with models coming from science, economics or engineering.