London School of Economics

Hilbert-Schmidt independence criterion (HSIC) is among the most widely-used approaches in machine learning and statistics to measure the independence of random variables. Despite its popularity and success in numerous applications, quite little is known about when HSIC characterizes independence. I am going to provide a complete answer to this question, with conditions which are often easy to verify in practice.

This talk is based on joint work with Bharath Sriperumbudur.

We analyse the consequences of conservative portfolio compression, i.e., netting cycles in financial networks, on systemic risk.  We show that the recovery rate in case of default plays a significant role in determining whether portfolio compression is potentially beneficial.  If recovery rates of defaulting nodes are zero then compression weakly reduces systemic risk. We also provide a necessary condition under which compression strongly reduces systemic risk.  If recovery rates are positive we show that whether compression is potentially beneficial or harmful for individual institutions does not just depend on the network itself but on quantities outside the network as well. In particular we show that  portfolio compression can have negative effects both for institutions that are part of the compression cycle and for those that are not. Furthermore, we show that while a given conservative compression might be beneficial for some shocks it might be detrimental for others. In particular, the distribution of the shock over the network matters and not just its size.  

[[{"fid":"47873","view_mode":"default","fields":{"format":"default"},"type":"media","attributes":{"class":"file media-element file-default"},"link_text":"AD abstract.pdf"}]]

We prove tight upper and lower bounds on an observable of the antiferromagnetic Potts model. From this we deduce the case d=3 of a conjecture of Galvin and Tetali on maximising the number of proper colourings in d-regular graphs.

For a graph $G$, the $k$-colour Ramsey number $R_k(G)$ is the least integer $N$ such that every $k$-colouring of the edges of the complete graph $K_N$ contains a monochromatic copy of $G$. Let $C_n$ denote the cycle on $n$ vertices. We show that for fixed $k\geq2$ and $n$ odd and sufficiently large,
$$
R_k(C_n)=2^{k-1}(n-1)+1.
$$
This resolves a conjecture of Bondy and Erdős for large $n$. The proof is analytic in nature, the first step of which is to use the regularity method to relate this problem in Ramsey theory to one in nonlinear optimisation.  This allows us to prove a stability-type generalisation of the above and establish a correspondence between extremal $k$-colourings for this problem and perfect matchings in the $k$-dimensional hypercube $Q_k$.

In this talk, we will establish existence and uniqueness for a wide class of Markovian systems of backward stochastic differential equations (BSDE) with quadratic nonlinearities. This class is characterized by an abstract structural assumption on the generator, an a-priori local-boundedness property, and a locally-H\"older-continuous terminal condition. We present easily verifiable sufficient conditions for these assumptions and treat several applications, including stochastic equilibria in incomplete financial markets, stochastic differential games, and martingales on Riemannian manifolds. This is a joint work with Gordan Zitkovic.

Dellamonica, Kohayakawa, Rödl and Ruciński showed that for $p=C(\log n/n)^{1/d}$ the random graph $G(n,p)$ contains asymptotically almost surely all spanning graphs $H$ with maximum degree $d$ as subgraphs. In this talk I will discuss a resilience version of this result, which shows that for the same edge density, even if a $(1/k-\epsilon)$-fraction of the edges at every vertex is deleted adversarially from $G(n,p)$, the resulting graph continues to contain asymptotically almost surely all spanning $H$ with maximum degree $d$, with sublinear bandwidth and with at least $C \max\{p^{-2},p^{-1}\log n\}$ vertices not in triangles. Neither the restriction on the bandwidth, nor the condition that not all vertices are allowed to be in triangles can be removed. The proof uses a sparse version of the Blow-Up Lemma. Joint work with Peter Allen, Julia Ehrenmüller, Anusch Taraz.