Covering Arrays in Combinatorics

• Covering arrays are combinatorial matrices that ensure every t-wise interaction among parameters appears at least once, providing rigorous test coverage.
• They are constructed using probabilistic methods, group-theoretic techniques, and algorithmic tools like the Lovász Local Lemma to achieve minimal test sizes.
• Applications span software testing, combinatorial design, and extremal set theory, driving research on asymptotic bounds and optimal configurations.

A covering array is a fundamental combinatorial object, central to interaction testing, combinatorial design, and extremal set theory. It is an $N \times k$ array over a $v$-ary alphabet in which every $N \times t$ subarray contains all possible $v^t$ tuples in its rows, guaranteeing exhaustive coverage of all $t$-way interactions among $k$ parameters, each with $v$ levels. The covering array number $\mathrm{CAN}(t, k, v)$ is the minimum $N$ for which such an array exists.

1. Formal Definitions and Basic Properties

A covering array $\mathrm{CA}(N; t, k, v)$ is an $N \times k$ matrix $A$ over $\{0,1,\ldots,v-1\}$ with the property

$$\forall\,T=\{i_1,\ldots,i_t\}\subseteq\{1,\ldots,k\},\quad \text{and}\ x\in\{0,\ldots,v-1\}^t,\ \exists\,r\in\{1,\ldots,N\}\ \text{such that}\ (A_{r,i_1},\ldots,A_{r,i_t}) = x.$$

The covering array number: $$\mathrm{CAN}(t,k,v) = \min\{ N : \mathrm{CA}(N;t,k,v)\ \text{exists} \}.$$ Orthogonal arrays $OA_\lambda(t,k,v)$ require each $t$-tuple to appear exactly $\lambda$ times in every $t$-column subarray; covering arrays require at least one occurrence. Covering arrays generalize orthogonal arrays and provide the minimal test size needed to guarantee $t$-wise coverage (Hiess et al., 20 Oct 2025).

2. Asymptotic Bounds and Constructions

Logarithmic Growth

The classical result (Katona–Kleitman–Godbole) is

$\mathrm{CAN}(t,k,v) = \Theta(\log_2 k)$

for fixed $t$, $v$ and $k \to \infty$ (Francetić et al., 2015).

Probabilistic and Analytical Bounds

The Lovász Local Lemma (LLL) yields: $$\mathrm{CAN}(t,k,v) \leq \frac{(t-1)\,\log_2 k}{\log_2(v^t/(v^t-1))}(1+o(1))$$ with more refined constructions, e.g., fixed-weight columns, improving constants for small $t$ and $v$ (Yuan et al., 2014).

Entropy-compression (algorithmic LLL) further refines constants: $$d(t,v) \leq \frac{v(t-1)}{\log_2(v^{t-1}/(v^{t-1}-1))}$$ and, via multivariable optimization, to (Francetić et al., 2015): $$d(t,v) \leq \frac{v(t-1)}{ (t-1)\log_2( v^v/(v-1)^{v-1} ) - f_0(t,v) }$$ for $$d(t,v) := \limsup_{k \to \infty}\mathrm{CAN}(t,k,v)/\log_2 k$$.

3. Exact Results for Small Parameters and Uniqueness

Binary Arrays and Strength Two

For $q=2$, $t=2$, maximal binary 2-covering arrays are unique up to equivalence, given by the standard maximal array—an $m \times \binom{m-1}{\lfloor m/2 \rfloor-1}$ matrix whose columns are all binary vectors of fixed weight (Choi et al., 2011).

Computational Determination

Modern computational methods (isomorph-free exhaustive search, canonical augmentation) have determined exact optimal covering arrays for 21 strength-two cases with $v>2$, $N>v^2$ (Kokkala et al., 2019). All exhibit uniformity (equal symbol-frequency in each column).

4. Generalizations: Mixed, Hypergraph, and Sequence Covering Arrays

Mixed Covering Arrays

Generalized covering designs handle multi-part alphabets and block sizes, providing lower and upper bounds via Schőnheim's bound, edge-counting, and recursive constructions (Bailey et al., 2010). Partial and mixed-level arrays are realized by partial covering arrays and hypergraph-based models.

Hypergraph Covering Arrays

Covering arrays on $r$-uniform hypergraphs restrict coverage to prescribed interactions, e.g., only certain triples. Inductive construction uses hooking operations (vertex/edge additions), achieving optimally small arrays for $\alpha$-acyclic and conformal hypertrees (Akhtar et al., 2015).

Sequence and Perfect Sequence Covering Arrays

Sequence covering arrays (SCAs) are sets of permutations covering all ordered $k$-subsequences; perfect sequence covering arrays (PSCAs) require exact multiplicity $\lambda$ (Na et al., 2022, Gentle et al., 2022). The minimal such $\lambda$ is denoted $g(n,k)$. For $(n,k) \in \{(5,3),(6,3),(7,3),(7,4)\}$, $g(n,k)=2$; for $(8,3)$, $g(8,3)=3$ (Na et al., 2022). PSCAs are tightly connected to directed designs and deletion-correcting codes.

5. Partial and Relaxed Covering Arrays

Relaxed requirements give rise to partial covering arrays, covering only a fraction of $t$-sets or tuples (Sarkar et al., 2016). Important results:

• Partial covering arrays with fraction $\alpha$:

$$N = O\left( \frac{v^t(t-1)\,\ln k}{v^t - m + 1} \right)$$

• $\epsilon$-almost covering arrays:

$N = O( v^t \ln( v^{t-1}/\epsilon ) )$

Moser–Tardos resampling and Markov-type randomized algorithms ensure efficient generation, matching information-theoretic lower bounds up to constant factors.

6. Arrays with Higher Index (Replication) and Constraints

Covering Arrays of Index $\lambda$

For $\lambda>1$ (every $t$-tuple occurs at least $\lambda$ times), the main asymptotic bound is (Calbert et al., 2022): $$\mathrm{CAN}_\lambda(t,k,v) = \Theta_{v,t}(\log k + \lambda)$$ removing previous $\lambda \log \log k$ terms. Improved leading constants are obtained via the Lovász Local Lemma and two-stage alteration schemes; graph coloring yields further reductions for higher $\lambda$.

SAT/MaxSAT-Based Construction

Satisfiability-based encodings of the covering array problem allow for exact and suboptimal solving, even under additional constraints (forbidden tuples, system-specific restrictions) (Ansótegui et al., 2021). MaxSAT variants minimize test suite size; incomplete MaxSAT is especially effective on large-scale constrained problems.

7. Algebraic and Group-Theoretic Constructions

Finite Field and Group Development Constructions

Maximal sequences (m-sequences), cyclic trace arrays, and group actions (e.g., PGL$(2,q)$) are exploited to yield covering arrays of high strength and efficiency (Maity et al., 2015, Tzanakis, 2017). For strength $t=4$, $g=3$, explicit bounds $4\text{-}CAN(k,3) \le 12k + 3$ are achieved using projective general linear group constructions and starter vectors. For binary arrays, concatenation of cyclic Hamming codes, self-dual sequence families, interleaving, and primitive polynomial periodicity yield near-optimal covering sequences and arrays (Chee et al., 12 Feb 2025, Chee et al., 2024).

8. Open Problems and Current Research Directions

Key open areas include:

• Determining $\mathrm{CAN}(t,k,v)$ for more small ($t,k,v$) and high strength ($t>4$) parameter sets (Hiess et al., 20 Oct 2025).
• Improving leading constants and sharpening lower bounds, especially for partial arrays and arrays of higher index.
• Extending algebraic and group-theoretic constructions to broader parameter ranges, especially using cyclotomy and discrete logarithms (Tzanakis, 2017).
• Classification and existence problems for mixed, hypergraph, and constrained covering arrays.
• Connections to perfect sequence covering arrays, deletion codes, and directed $t$-designs.
• Systematic study of optimal arrays for non-binary alphabets and verification of the uniformity conjecture (Kokkala et al., 2019).
• Efficient SAT/MaxSAT and enumeration algorithms capable of exact or near-optimal constructions in real-world, large-scale systems.

Table: Summary of Covering Array Number Bounds

Type	Primary Bound or Complexity	Reference
Classical ($t$-way, full)	$O((t-1)v^t\log k)$; $\Theta(v^t\log k)$	(Sarkar et al., 2016)
Logarithmic growth ($t$ fixed)	$\Theta(\log_2 k)$	(Francetić et al., 2015)
Probabilistic/LLL bound	$(t-1)\log_2 k/\log_2(v^t/(v^t-1))$	(Yuan et al., 2014)
Entropy compression	$v(t-1)/\log_2[v^{t-1}/(v^{t-1}-1)]$	(Francetić et al., 2015)
Higher index $\lambda$	$\Theta(\log k+\lambda)$	(Calbert et al., 2022)
Small $q=2, t=2$ (exact)	$(m-1\choose \lfloor m/2 \rfloor-1)$ columns	(Choi et al., 2011)
Partial/relaxed coverage	$O(v^{t-1}\log k)$ for small relaxations	(Sarkar et al., 2016)
Algebraic/group development ($t\ge3$)	Polynomial/construction-dependent	(Maity et al., 2015)

Covering arrays remain a rich, continually evolving domain at the intersection of extremal combinatorics, algorithmic theory, algebraic design, and practical test suite optimization.