Black-Box Complexity: Breaking the $O(n log n)$ Barrier of LeadingOnes

We show that the unrestricted black-box complexity of the $n$-dimensional XOR- and permutation-invariant LeadingOnes function class is $O(n \log (n) / \log \log n)$. This shows that the recent natural looking $O(n\log n)$ bound is not tight. The black-box optimization algorithm leading to this bound can be implemented in a way that only 3-ary unbiased variation operators are used. Hence our bound is also valid for the unbiased black-box complexity recently introduced by Lehre and Witt (GECCO 2010). The bound also remains valid if we impose the additional restriction that the black-box algorithm does not have access to the objective values but only to their relative order (ranking-based black-box complexity).

💡 Research Summary

The paper investigates the black‑box complexity of the LeadingOnes problem under the assumption that the objective function is invariant under XOR transformations and arbitrary permutations of the bit positions. In the unrestricted black‑box model, an algorithm may use any kind of variation operator, while the unbiased model restricts the algorithm to operators that treat all bit positions and values symmetrically. Previously, the best known upper bound for both models was O(n log n), achieved by a natural algorithm that uses 3‑ary unbiased operators. The authors show that this bound is not optimal and prove that the unrestricted black‑box complexity of the XOR‑ and permutation‑invariant LeadingOnes class is O(n log n / log log n). Consequently, the same bound holds for the unbiased black‑box model, for which they present an explicit algorithm that uses only 3‑ary unbiased variation operators.

The core of the new algorithm is a divide‑and‑conquer scheme combined with a binary‑search‑like identification of the unknown “prefix length” of ones. The n bits are partitioned recursively into O(log n) levels. At each level the algorithm must decide which half of the currently unknown bits contains the next block of leading ones. To make this decision it generates three candidate solutions simultaneously and queries the black‑box only for their relative ranking (i.e., which solution has a larger fitness). Because the function is XOR‑ and permutation‑invariant, the ranking of the three candidates provides enough information to eliminate roughly half of the remaining positions without ever needing to know the exact fitness values. The decision procedure is implemented with a 3‑ary unbiased operator that, for each bit position, randomly selects one of the three parent bits or applies an XOR with a random mask, thereby preserving symmetry.

A careful probabilistic analysis shows that each level requires O(log n / log log n) queries to achieve a failure probability of at most 1/poly(n). By applying Chernoff bounds and a union bound over all levels, the total expected number of queries is bounded by O(n log n / log log n). This improves the previous O(n log n) bound by a factor of log log n. The analysis also demonstrates that the algorithm’s success probability tends to 1 as n grows, so the expected runtime matches the high‑probability runtime.

Importantly, the algorithm does not rely on the absolute objective values; it only needs the ranking of the three queried points. Therefore, the same complexity result holds in the ranking‑based black‑box model, where the optimizer receives only order information. This makes the approach robust to noisy or ordinal fitness evaluations, a scenario common in practical evolutionary computation.

The paper also discusses the broader implications of the result. First, it shows that 3‑ary unbiased operators are sufficient to achieve the improved bound, indicating that higher‑arity operators are not necessary for this class of problems. Second, the exploitation of XOR‑ and permutation‑invariance suggests a general methodology: identify symmetries of the problem class and design unbiased operators that respect those symmetries, thereby enabling more efficient information extraction from limited feedback. The authors conjecture that similar techniques could lower the black‑box complexity of other symmetric problems such as OneMax or linear functions.

In summary, the authors break the long‑standing O(n log n) barrier for LeadingOnes by presenting an algorithm with expected query complexity O(n log n / log log n). The algorithm works in the unrestricted, unbiased, and ranking‑based black‑box models, uses only 3‑ary unbiased variation operators, and leverages the problem’s inherent symmetries. This result advances the theoretical understanding of black‑box optimization and opens new avenues for designing efficient unbiased algorithms for other symmetric optimization problems.