Past exam of the mathematics course of the University of Cambridge / 2013 / iii / Paper 59 / 2

With a read-only input tape, deterministic space complexity class $\mathrm{SPACE}(s(n))$ consists of languages decidable by deterministic Turing machines using $O(s(n))$ work-tape cells. The nondeterministic space complexity class $\mathrm{NSPACE}(s(n))$ uses Nondeterministic Turing machines, with the bound holding on every computation path and acceptance meaning that some path accepts. Input storage is not charged. We use the standard finite-state, fixed-alphabet model.

The Savitch theorem isFirst suppose a work-space budget $m\ge \lceil {\log }_2(n+2)\rceil$ is available. A configuration graph for computations restricted to $m$ cells has configurations encoded in $O(m)$ bits: work contents, control state and head positions, including $O(\log n)$ bits for the input head. Its size is at most $2^{cm}$ for a machine-dependent constant $c$. Add accept and exit sinks if needed; the size bound merely changes $c$. A reachable configuration has a simple path of length smaller than the number of configurations.

For encoded configurations $u,v$, define $R(u,v,k)$ to ask whether a path of length at most $2^k$ exists. At $k=0$, test equality or one transition. For $k>0$, enumerate every candidate middle configuration $w$ and testAny path of the given length can be split into two halves of length at most $2^{k-1}$; conversely concatenation gives a path of length at most $2^k$. This proves the recursion. Taking $k=\lceil cm\rceil$ suffices. Depth is $O(m)$, and each recursive frame stores only its endpoints, the current middle configuration and counters in $O(m)$ bits. The two recursive calls are made sequentially and reuse their space. The total is $O(m^2)$, even though the time can be very large.

The printed hypothesis does not say that $s$ is computable. The space-bound discovery by exit reachability removes that issue. Begin with $m=\lceil {\log }_2(n+2)\rceil$ and use the preceding recursion to test both whether an accepting state is reachable within the budget and whether a reachable state has a transition leaving the budget. Accept in the first case. If acceptance is absent and an exit is reachable, double $m$ and repeat. If neither is reachable, reject: all computations remain within this finite configuration graph, and none accepts.

Every path of the original machine uses at most $Cs(n)$ cells. Once $m$ reaches that bound there can be no reachable exit, so this procedure terminates. The final budget is $O(s(n))$ by doubling, and all stages reuse the same storage. Hence its deterministic space is $O(s(n{)}^2)$ without requiring a machine that first computes $s(n)$. This proves the stated general form.

Here a directed cycle has at least one edge. If zero-length paths were treated as cycles, an edgeless one-vertex graph would already satisfy the predicate; that convention would give the trivial nonempty-graph predicate instead of the intended directed cycle detection problem.

For NL membership, guess a starting vertex $v$, follow a guessed directed walk for between one and $N=|V|$ edges, and accept if it returns to $v$. Store only the start, current vertex and step counter. A graph with a nonempty closed walk contains a simple directed cycle of at most $N$ edges, including a self-loop if present. Thus the algorithm uses logarithmic space and is complete for the predicate.

For hardness, reduce the directed graph reachability problem to cycle detection. Form $G^{(N+1)}$ and add the single backward arcAll layering arcs advance exactly one layer, including the waiting arcs $(v,i)\to (v,i+1)$, so the layered graph alone is a Directed acyclic graph. Any cycle in the augmented graph must contain the backward arc and therefore contains a path from $(s,1)$ to $(t,N+1)$.

If $t$ is reachable from $s$ in $G$, use a simple path of length at most $N-1$ and pad it with waits to exactly $N$ steps. It becomes the required layered path and closes to a cycle. Conversely, projecting such a layered path and deleting waits gives a walk from $s$ to $t$ in $G$. This includes the case $s=t$, where reachability has a zero-length witness but the augmented graph has a genuinely positive-length cycle.

There are $N(N+1)$ vertices and polynomially many arcs. A transducer enumerates pairs/layers and checks old adjacency by scanning its input, using only $O(\log N)$ bits. Hence this is a logspace many-one reduction, proving