This special issue is devoted to papers from the meeting on Combinatorics, Probability and Computing, held at the Mathematisches Forschungsinstitut in Oberwolfach from the 23$^{\rm rd}$ to the 29$^{\rm th}$ of September 2001. As is typical of meetings at Oberwolfach, this was an exciting and stimulating conference; there was a large number of excellent talks, many of which provoked a great deal of interest and discussion among the participants.

We analyse a randomized pursuit-evasion game played by two players on a graph, a hunter and a rabbit. Let $G$ be any connected, undirected graph with $n$ nodes. The game is played in rounds and in each round both the hunter and the rabbit are located at a node of the graph. Between rounds both the hunter and the rabbit can stay at the current node or move to another node. The hunter is assumed to be restricted to the graph $G$: in every round, the hunter can move using at most one edge. For the rabbit we investigate two models: in one model the rabbit is restricted to the same graph as the hunter, and in the other model the rabbit is unrestricted, i.e., it can jump to an arbitrary node in every round.

We say that the rabbit is caught as soon as hunter and rabbit are located at the same node in a round. The goal of the hunter is to catch the rabbit in as few rounds as possible, whereas the rabbit aims to maximize the number of rounds until it is caught. Given a randomized hunter strategy for $G$, the escape length for that strategy is the worst case expected number of rounds it takes the hunter to catch the rabbit, where the worst case is with regard to all (possibly randomized) rabbit strategies. Our main result is a hunter strategy for general graphs with an escape length of only $\O(n \log (\diam(G)))$ against restricted as well as unrestricted rabbits. This bound is close to optimal since $\Omega(n)$ is a trivial lower bound on the escape length in both models. Furthermore, we prove that our upper bound is optimal up to constant factors against unrestricted rabbits.

It is known that random $k$-Sat instances with at least $(2^k \cdot \ln 2)\cdot n$ random clauses are unsatisfiable with high probability. This result is simply obtained by showing that the expected number of satisfying assignments tends to $0$ when the number of variables $n$ tends to infinity. This proof does not directly provide us with an efficient algorithm certifying the unsatisfiability of a given random formula. Concerning efficient algorithms, it is essentially known that random formulas with $n^\varepsilon \cdot n^{k/2}$ clauses with $k$ literals can be efficiently certified as unsatisfiable. The present paper is the result of trying to lower this bound. We obtain better bounds for some specialized satisfiability problems. These results are based on discrepancy investigations for hypergraphs.

Further, we show that random formulas with a linear number of clauses can be efficiently certified as unsatisfiable in the Not-All-Equal-$3$-Sat sense. A similar result holds for the non-$3$-colourability of random graphs with a linear number of edges. We obtain these results by direct application of approximation algorithms.

In this paper we consider a random star $d$-process which begins with $n$ isolated vertices, and in each step chooses randomly a vertex of current minimum degree $\delta$, and connects it with $d - \delta$ random vertices of degree less than $d$. We show that, for $d \geqslant 3$, the resulting final graph is connected with probability $1 - o(1)$, and moreover that, for suficiently large $d$, it is $d$-connected with probability $1 - o(1)$.

This paper shows that the largest possible contrast $C_{k,n}$ in a $k$-out-of-$n$ secret sharing scheme is approximately $4^{-(k-1)}$. More precisely, we show that $4^{-(k-1)} \leq C_{k,n} \leq 4^{-(k-1)}n^k/(n(n-1)\cdots(n-(k-1)))$. This implies that the largest possible contrast equals $4^{-(k-1)}$ in the limit when $n$ approaches infinity. For large $n$, the above bounds leave almost no gap. For values of $n$ that come close to $k$, we will present alternative bounds (being tight for $n=k$). The proofs of our results proceed by finding a relationship between the largest possible contrast in a secret sharing scheme and the smallest possible approximation error in problems occurring in approximation theory.

Suppose $n$ circular arcs of lengths $\len_i \in [0,1],0\leq i<n$, are placed uniformly at random on a unit length circle. We study the maximum overlap, i.e., the number of arcs that overlap at the same position of the circle. In particular, we give almost exact tail bounds for this random variable. By applying these tail bounds we can characterize the expected maximum overlap exactly up to constant factors in lower order terms. We illustrate the strength of our results by presenting new performance guarantees for three algorithmic applications: minimizing rotational delays for disks, scheduling accesses to parallel disks, and allocating memory blocks to limit cache interference misses.

This is a continuation of our work on quasi-random graph properties. The class of quasi-random graphs is defined by certain equivalent graph properties possessed by random graphs. One of the most important of these properties is that, for fixed $\nu$, every fixed sample graph $L_\nu$ has the same frequency in $G_n$ as in the $p$-random graph. (This holds for both induced and not necessarily induced containment.) In [9] we proved that, if the frequency of just one fixed $L_\nu$ – as a not necessarily induced subgraph – in every ‘large’ induced subgraph $F_h\subseteq G_n$ is the same as for the random graphs, then $(G_n)$ is quasi-random. Here we shall investigate the analogous problem for induced subgraphs $L_\nu$. In such cases $(G_n)$ is not necessarily quasi-random.

We shall prove, among other things, that, for every regular sample graph $L_\nu$, $\nu\geqslant 4$, if the number of induced copies of $L_\nu$ in every induced $F_h\subseteq G_n$ is asymptotically the same as in a $p$-random graph (up to an error term $o(n^\nu)$), then $(G_n)$ is the union of (at most) two quasi-random graph sequences, with possibly distinct attached probabilities (assuming that $p\in (0,1)$, $e(L_\nu)>0$, and $L_\nu\ne K_\nu$).

We conjecture the same conclusion for every $L_\nu$ with $\nu\ge 4$, i.e., even if we drop the assumption of regularity.

We shall reduce the general problem to solving a system of polynomials. This gives a ‘simple’ algorithm to decide the problem for every given $L_\nu$.

The branching random walk on a regular graph turns out to be particularly easy to analyse using results for the corresponding simple random walk. In this way, one can show that there is an intermediate phase of weak survival if and only if the graph is nonamenable. No such simple analysis holds more generally, and it is known that the nonamenability equivalence does not extend to general connected graphs of bounded degree (although we observe that it does hold for such graphs if the branching random walk is modified in a certain natural way). The most important general class of (bounded degree, connected) graphs for which it is thought that the equivalence may hold is that of quasi-transitive graphs: we show that this is indeed the case.