arXiv++ Combinatorics

We consider the upper tail large deviations of subgraph counts for irregular graphs $\mathrm{H}$ in $\mathbb{G}(n,p)$, the sparse Erdős-Rényi graph on $n$ vertices with edge connectivity probability $p \in (0,1)$. For $n^{-1/Δ} \ll p \ll 1$, where $Δ$ is the maximum degree of $\mathrm{H}$, we derive the upper tail large deviations for any irregular graph $\mathrm{H}$. On the other hand, we show that for $p$ such that $1 \ll n^{v_{\mathrm{H}}} p^{e_{\mathrm{H}}} \ll (\log n)^{α^{*}_{\mathrm{H}}/\left(α^{*}_{\mathrm{H}}-1\right)}$, where $v_{\mathrm{H}}$ and $e_{\mathrm{H}}$ denote the number of vertices and edges of $\mathrm{H}$, and $α^*_{\mathrm{H}}$ denotes the fractional independence number, the upper tail large deviations of the number of unlabelled copies of $\mathrm{H}$ in $\mathbb{G}(n,p)$ is given by that of a sequence of Poisson random variables with diverging mean, for any strictly balanced graph $\mathrm{H}$. Restricting to the $r$-armed star graph we further prove a localized behavior in the intermediate range of $p$ (left open by the above two results) and show that the mean-field approximation is asymptotically tight for the logarithm of the upper tail probability. This work further identifies the typical structures of $\mathbb{G}(n,p)$ conditioned on upper tail rare events in the localized regime.

Consider a random symmetric matrix with i.i.d.~entries on and above its diagonal that are products of Bernoulli random variables and random variables with sub-Gaussian tails. Such a matrix will be called a sparse Wigner matrix and can be viewed as the adjacency matrix of a random network with sub-Gaussian weights on its edges. In the regime where the mean degree is at least logarithmic in dimension, the edge eigenvalues of an appropriately scaled sparse Wigner matrix stick to the edges of the support of the semicircle law. We show that in this sparsity regime, the large deviations upper tail event of the largest eigenvalue of a sparse Wigner matrix with sub-Gaussian entries is generated by either the emergence of a high degree vertex with a large vertex weight or that of a clique with large edge weights. Interestingly, the rate function obtained is discontinuous at the typical value of the largest eigenvalue, which accounts for the fact that its large deviation behaviour is generated by finite rank perturbations. This complements the results of Ganguly and Nam, and Ganguly, Hiesmayr, and Nam which considered the case where the mean degree is constant.

We consider an Erdős-Rényi graph $\mathbb{G}(n,p)$ on $n$ vertices with edge probability $p$ such that \[ \sqrt{\frac{\log n}{\log \log n}} \ll np \le n^{1/2-o(1)}, \label{eq:abs} \tag{$\dagger$} \] and derive the upper tail large deviations of $λ(\mathbb{G}(n,p))$, the largest eigenvalue of its adjacency matrix. Within this regime we show that, for $p \gg n^{-2/3}$ the $\log$-probability of the upper tail event of $λ(\mathbb{G}(n,p))$ equals to that of planting a clique of an appropriate size (upon ignoring smaller order terms), while for $p \ll n^{-2/3}$ the same is given by that of the existence of a high degree vertex. This, in particular, shows an emergence of {\em non-planted localized structure} in the latter regime. We also confirm that in the entire regime \eqref{eq:abs} the large deviation probability is asymptotically approximated by the solution of the mean-field variational problem, and further identify the typical structure of $\mathbb{G}(n,p)$ conditioned on the upper tail event of $λ(\mathbb{G}(n,p))$ in a certain sub-regime of $p$. For $p$ such that $\log(np) \gtrsim \log n$ the large deviations of $λ(\mathbb{G}(n,p))$ is deduced from those of ${\rm Hom}(C_{2t}, \mathbb{G}(n,p))$, the homomorphism counts of cycle of length $2t$, for $t \ge 3$ and $p$ such that $n^{1/2-o(1)} \ge np \gg n^{1/t}$. In this latter regime the typical structure of $\mathbb{G}(n,p)$ conditioned on the upper tail of ${\rm Hom}(C_{2t}, \mathbb{G}(n,p))$ is identified and asymptotic tightness of the mean-field approximation is also established.

For a $Δ$-regular connected graph ${\sf H}$ the problem of determining the upper tail large deviation for the number of copies of ${\sf H}$ in $\mathbb{G}(n,p)$, an Erdős-Rényi graph on $n$ vertices with edge probability $p$, has generated significant interests. For $p\ll 1$ and $np^{Δ/2} \gg (\log n)^{1/(v_{\sf H}-2)}$, where $v_{\sf H}$ is the number of vertices in ${\sf H}$, the upper tail large deviation event is believed to occur due to the presence of localized structures. In this regime the large deviation event that the number of copies of ${\sf H}$ in $\mathbb{G}(n,p)$ exceeds its expectation by a constant factor is predicted to hold at a speed $n^2 p^Δ \log (1/p)$ and the rate function is conjectured to be given by the solution of a mean-field variational problem. After a series of developments in recent years, covering progressively broader ranges of $p$, the upper tail large deviations for cliques of fixed size was proved by Harel, Mousset, and Samotij \cite{hms} in the entire localized regime. This paper establishes the conjecture for all connected regular graphs in the whole localized regime.

We consider three different models of sparse random graphs:~undirected and directed Erdős-Rényi graphs, and random bipartite graph with an equal number of left and right vertices. For such graphs we show that if the edge connectivity probability $p \in (0,1)$ satisfies $n p \ge \log n + k(n)$ with $k(n) \to \infty$ as $n \to \infty$, then the adjacency matrix is invertible with probability approaching one (here $n$ is the number of vertices in the two former cases and the number of left and right vertices in the latter case). If $np \le \log n -k(n)$ then these matrices are invertible with probability approaching zero, as $n \to \infty$. In the intermediate region, when $np=\log n + k(n)$, for a bounded sequence $k(n) \in \mathbb{R}$, the event $Ω_0$ that the adjacency matrix has a zero row or a column and its complement both have non-vanishing probability. For such choices of $p$ our results show that conditioned on the event $Ω_0^c$ the matrices are again invertible with probability tending to one. This shows that the primary reason for the non-invertibility of such matrices is the existence of a zero row or a column. The bounds on the probability of the invertibility of these matrices are a consequence of quantitative lower bounds on their smallest singular values. Combining this with an upper bound on the largest singular value of the centered version of these matrices we show that the (modified) condition number is $O(n^{1+o(1)})$ on the event that there is no zero row or column, with large probability. This matches with von Neumann's prediction about the condition number of random matrices up to a factor of $n^{o(1)}$, for the entire range of $p$.

Let $P_n^1,\dots, P_n^d$ be $n\times n$ permutation matrices drawn independently and uniformly at random, and set $S_n^d:=\sum_{\ell=1}^d P_n^\ell$. We show that if $\log^{12}n/(\log \log n)^{4} \le d=O(n)$, then the empirical spectral distribution of $S_n^d/\sqrt{d}$ converges weakly to the circular law in probability as $n \to \infty$.

We consider classes of pseudo-random graphs on $n$ vertices for which the degree of every vertex and the co-degree between every pair of vertices are in the intervals $(np - Cn^δ,np+Cn^δ)$ and $(np^2- C n^δ, np^2 +C n^δ)$ respectively, for some absolute constant $C$, and $p, δ\in (0,1)$. We show that for such pseudo-random graphs the number of induced isomorphic copies of subgraphs of size $s$ are approximately same as that of an Erdős-Réyni random graph with edge connectivity probability $p$ as long as $s \le (((1-δ)\wedge \frac{1}{2})-o(1))\log n/\log (1/p)$, when $p \in (0,1/2]$. When $p \in (1/2,1)$ we obtain a similar result. Our result is applicable for a large class of random and deterministic graphs including exponential random graph models (ERGMs), thresholded graphs from high-dimensional correlation networks, Erdős-Réyni random graphs conditioned on large cliques, random $d$-regular graphs and graphs obtained from vector spaces over binary fields. In the context of the last example, the results obtained are optimal. Straight-forward extensions using the proof techniques in this paper imply strengthening of the above results in the context of larger motifs if a model allows control over higher co-degree type functionals.