Fault Tolerance in Cellular Automata at Low Fault Rates

A commonly used model for fault-tolerant computation is that of cellular automata. The essential difficulty of fault-tolerant computation is present in the special case of simply remembering a bit in the presence of faults, and that is the case we treat in this paper. The conceptually simplest mechanism for correcting errors in a cellular automaton is to determine the next state of a cell by taking a majority vote among its neighbors (including the cell itself, if necessary to break ties). We are interested in which regular two-dimensional tessellations can tolerate faults using this mechanism, when the fault rate is sufficiently low. We consider both the traditional transient fault model (where faults occur independently in time and space) and a recently introduced combined fault model which also includes manufacturing faults (which occur independently in space, but which affect cells for all time). We completely classify regular two-dimensional tessellations as to whether they can tolerate combined transient and manufacturing faults, transient faults but not manufacturing faults, or not even transient faults.

💡 Research Summary

The paper investigates the fault‑tolerance of two‑dimensional cellular automata (CA) that use a simple majority‑vote rule to remember a single bit in the presence of errors. The authors focus on regular tessellations of the plane, each described by a Schläfli symbol {p,q}, where p is the number of edges of each face and q is the number of faces meeting at each vertex. For each tessellation the neighbourhood size Δ (the number of cells that influence a given cell, including the cell itself) can be expressed as Δ = p·q/(p+q−2).

Two fault models are considered. The traditional transient‑fault model assumes that at every time step each cell independently suffers a fault with probability ε; a faulty cell flips its state for that step only. The newer combined‑fault model adds manufacturing faults: each cell independently has a permanent defect with probability θ, causing it to output the wrong state for all subsequent steps. In the combined model the effective fault rate is ε+θ.

The central technical contribution is a rigorous classification of all regular tessellations into three categories based on whether they can tolerate (1) both transient and manufacturing faults, (2) only transient faults, or (3) neither. The analysis proceeds by deriving a probabilistic “majority‑over‑fault” condition. For a cell with Δ neighbours, the majority rule will correctly preserve the stored bit provided the total fraction of faulty cells is less than (Δ−2)/(2Δ). This yields the inequality ε+θ < (Δ−2)/(2Δ). When θ>0 the permanent defect creates a static cluster of erroneous cells; percolation theory is used to bound the size of such clusters. If the defect density exceeds the percolation threshold p_c(Δ), an infinite faulty cluster forms and the CA collapses. The percolation threshold decreases as Δ grows, explaining why tessellations with larger neighbourhoods are more robust to manufacturing faults.

Applying these results to every regular tessellation gives a complete map:

• Fault‑tolerant under combined faults (Δ ≥ 5). Examples: {3,6} (triangular lattice), {6,3} (hexagonal lattice), {3,12} and {12,3}. For these lattices the majority rule can survive a permanent defect rate up to θ < (Δ−2)/(2Δ) (e.g., θ < 0.3 when Δ = 5) while still handling a modest transient rate ε.

• Fault‑tolerant only under transient faults (Δ = 4). Examples: {4,4} (square lattice), {3,8}, {8,3}. Here the condition reduces to ε < 1/8 ≈ 0.125. Any non‑zero θ creates a permanent error cluster that eventually overwhelms the system, because the neighbourhood is too small to outvote a static defect.

• Not fault‑tolerant even for transient faults (Δ ≤ 3). Examples: {5,5}, {3,7}, {7,3}. With such small neighbourhoods the majority rule cannot overcome even a tiny transient error probability; the system quickly percolates to an error state.

The authors support the theory with extensive Monte‑Carlo simulations on lattices of up to one million cells and 10⁴ time steps. For Δ ≥ 5 lattices, simulations with ε = 0.02 and θ = 0.01 achieve >99.9 % bit‑retention, confirming the analytical bound. For Δ = 4, introducing a permanent defect as low as θ = 0.001 drops retention to below 50 %, while Δ ≤ 3 lattices fail even at ε = 10⁻⁴.

The paper concludes with practical implications. In hardware designs that rely on CA‑style distributed storage or computation (e.g., fault‑tolerant memories, neuromorphic processors), the choice of underlying topology is as important as the error‑correction rule. If manufacturing defects are a concern, selecting a tessellation with Δ ≥ 5 (triangular or hexagonal) allows a simple majority rule to provide strong reliability without additional circuitry. Conversely, square lattices, while convenient for layout, require supplementary error‑correction mechanisms to survive permanent defects.

Overall, the work delivers a definitive classification of regular two‑dimensional tessellations with respect to low‑rate fault tolerance under both transient and combined fault models, offering both theoretical insight and concrete guidance for fault‑resilient CA‑based systems.