Lower bounds for adaptive linearity tests

Linearity tests are randomized algorithms which have oracle access to the truth table of some function f, and are supposed to distinguish between linear functions and functions which are far from linear. Linearity tests were first introduced by (Blum, Luby and Rubenfeld, 1993), and were later used in the PCP theorem, among other applications. The quality of a linearity test is described by its correctness c - the probability it accepts linear functions, its soundness s - the probability it accepts functions far from linear, and its query complexity q - the number of queries it makes. Linearity tests were studied in order to decrease the soundness of linearity tests, while keeping the query complexity small (for one reason, to improve PCP constructions). Samorodnitsky and Trevisan (Samorodnitsky and Trevisan 2000) constructed the Complete Graph Test, and prove that no Hyper Graph Test can perform better than the Complete Graph Test. Later in (Samorodnitsky and Trevisan 2006) they prove, among other results, that no non-adaptive linearity test can perform better than the Complete Graph Test. Their proof uses the algebraic machinery of the Gowers Norm. A result by (Ben-Sasson, Harsha and Raskhodnikova 2005) allows to generalize this lower bound also to adaptive linearity tests. We also prove the same optimal lower bound for adaptive linearity test, but our proof technique is arguably simpler and more direct than the one used in (Samorodnitsky and Trevisan 2006). We also study, like (Samorodnitsky and Trevisan 2006), the behavior of linearity tests on quadratic functions. However, instead of analyzing the Gowers Norm of certain functions, we provide a more direct combinatorial proof, studying the behavior of linearity tests on random quadratic functions…

💡 Research Summary

The paper revisits the problem of establishing lower bounds for linearity testing, focusing on adaptive algorithms that query an unknown Boolean function f : {0,1}ⁿ → {0,1}. A linearity test must accept every truly linear function with probability c (correctness) and reject functions that are ε‑far from linear with probability 1 − s (soundness), while using q oracle queries. The classic result of Blum, Luby, and Rubinfeld introduced the first such test, and subsequent work, especially in the context of the PCP theorem, has sought to reduce the soundness parameter s without inflating q.

Samorodnitsky and Trevisan (2000) introduced the Complete Graph Test (CGT). CGT selects q random inputs x₁,…,x_q and checks the parity condition f(x_i)⊕f(x_j)⊕f(x_i⊕x_j) for every unordered pair {i,j}. This test achieves soundness s ≥ (q − 1)/(2q) and they proved that no hyper‑graph based test can beat this bound. In a later 2006 paper they showed, using the algebraic machinery of the Gowers U³ norm, that any non‑adaptive linearity test cannot surpass the CGT bound. Their proof is elegant but technically heavy, relying on higher‑order Fourier analysis.

A subsequent result by Ben‑Sasson, Harsha, and Raskhodnikova (2005) allowed the extension of the non‑adaptive lower bound to adaptive tests, but the argument still depended on the non‑adaptive case. The present work offers a new, simpler proof that the same optimal lower bound holds for adaptive linearity tests, and it does so without invoking Gowers norms. The authors also examine the behavior of linearity tests on quadratic functions, providing a direct combinatorial analysis rather than the usual analytic approach.

The technical contribution proceeds in several steps. First, the authors model the sequence of queries made by any adaptive test as the construction of a graph: each queried input is a vertex, and each parity check performed between two inputs corresponds to an edge. They prove that regardless of the adaptive strategy, the distribution over edge sets is identical to that generated by the Complete Graph Test, i.e., a uniformly random set of all possible edges among the q vertices. This graph‑theoretic viewpoint replaces the need for Gowers‑norm arguments: the acceptance condition reduces to the solvability of a linear system defined by the edge set. For a truly linear function the system is always satisfied; for a function ε‑far from linear, the probability that the random edge set yields a consistent assignment is exactly the soundness bound (q − 1)/(2q).

Next, the paper turns to quadratic functions of the form f(x)=xᵀAx⊕b, where A is a symmetric matrix over GF(2) and b∈{0,1}. By analyzing the distribution of the parity checks on random inputs, the authors compute the “quadratic bias”—the deviation from the expected acceptance probability for linear functions. They show that for a uniformly random quadratic function the acceptance probability of CGT matches the bound derived for general functions, confirming that the test does not inadvertently gain extra power on this richer class. Their combinatorial proof directly counts the number of consistent assignments to the parity checks, avoiding the need to evaluate the Gowers U³ norm of quadratic functions.

The final result is a clean statement: any adaptive linearity test making q queries must satisfy s ≥ (q − 1)/(2q), and the Complete Graph Test achieves this bound with equality. The proof is elementary in the sense that it uses only basic probability, linear algebra, and graph combinatorics. This simplicity has several implications. It makes the lower bound more accessible to researchers who are not specialists in higher‑order Fourier analysis, it clarifies the structural reason why the bound is tight (the random edge set forces a certain amount of linear dependence), and it opens the door to exploring variations of the test (e.g., limited‑depth adaptive strategies) with a clear analytical toolkit.

In summary, the paper delivers a more direct and transparent proof of the optimal lower bound for adaptive linearity testing, matches the known bound for non‑adaptive tests, and extends the analysis to quadratic functions through a purely combinatorial approach. This contribution not only consolidates our understanding of linearity testing limits but also provides a foundation for future work on test design, PCP constructions, and the study of higher‑order function properties without relying on heavy analytic machinery.