Extended L-ensembles: a new representation for determinantal point processes (Q6103978)

Determinantal point processes are by now perhaps the most famous example of repulsive point processes. A key aspect of determinantal point processes is that ``diversity'' is defined relative to a notion of similarity represented by a positive-definite kernel. For instance, if the items are vectors in \(\mathbb{R}^d\), similarity may be defined via the squared-exponential (Gaussian) kernel: \(k(x, y)=\exp\left(-\| x - y\|^2\right)\), where \(x\) and \(y\) are two items, and similarity is a decreasing function of distance. The class of determinantal point processess can be separated into two subclasses: \(L\)-ensembles and the rest. By definition, an \(L\)-ensemble based on the \(n\times n\) kernel matrix \(\mathbf{L} = [k(x_i, x_j)]_{i,j}\) is a distribution over random subsets \(\chi\) such that \(P(\chi = X)\propto \det\mathbf{L}_X\), where \(\mathbf{L}_X\) is the principal submatrix of \(\mathbf{L}\) indexed by \(X\). If two or more points in \(X\) are very similar (in the sense of the kernel function), then the matrix \(\mathbf{L}_X\) has rows that are nearly collinear and the determinant is small. This in turns makes it unlikely that such a set \(\chi\) will be selected by the \(L\)-ensemble. \(L\)-ensembles are thus highly intuitive, and it is easy to design a determinantal point process appropriate for a particular situation, since one only needs to pick an appropriate kernel. Unfortunately, not all determinantal point processes are \(L\)-ensembles. The main contribution of this paper is to provide a unifying framework to write handy, explicit joint distributions for all varying and fixed-size determinantal point processes and thus to fill in the holes of the current theory. To this end, the authors introduce extended \(L\)-ensembles, a novel way of representing the class of determinantal point processes. The authors show that all determinantal point processes are extended \(L\)-ensembles, and vice versa. The paper is organized as follows. Section 1 contains definitions, preliminaries and all of the results are classical [\textit{S. Barthelmé} et al., Bernoulli 25, No. 4B, 3555--3589 (2019; Zbl 1428.62095)]. Section 2 defines extended \(L\)-ensembles while Section 3 gives some of their major properties. As a theoretical application, Section 4 shows how extended \(L\)-ensembles arise in perturbative limits of determinantal point processes. As a practical application, Section 5 shows that some interesting extended \(L\)-ensembles can be constructed via conditionally (semi)-definite positive functions. The authors chose to focus on discrete determinantal point processes, for simplicity. The continuous case is a straightforward generalisation of discrete results, which the authors sketch in the concluding remarks in Section 6.