A Distance Amplification Lemma for Monotonicity

We show a procedure that, given oracle access to a function $f\colon {0,1}^n\to{0,1}$, produces oracle access to a function $f’\colon {0,1}^{n’}\to{0,1}$ such that if $f$ is monotone, then $f’$ is monotone, and if $f$ is $\varepsilon$-far from monotone, then $f’$ is $Ω(1)$-far from monotone. Moreover, $n’ \leq n 2^{O(1/\varepsilon)}$ and each oracle query to $f’$ can be answered by making $2^{O(1/\varepsilon)}$ oracle queries to $f$. Our lemma is motivated by a recent result of [Chen, Chen, Cui, Pires, Stockwell, arXiv:2511.04558], who showed that for all $c>0$ there exists $\varepsilon_c>0$, such that any (even two-sided, adaptive) algorithm distinguishing between monotone functions and $\varepsilon_c$-far from monotone functions, requires $Ω(n^{1/2-c})$ queries. Combining our lemma with their result implies a similar result, except that the distance from monotonicity is an absolute constant $\varepsilon>0$, and the lower bound is $Ω(n^{1/2-o(1)})$ queries.

💡 Research Summary

This paper introduces a fundamental “Distance Amplification Lemma” for monotonicity testing of Boolean functions, and uses it to strengthen the state-of-the-art lower bounds for adaptive testers.

The core problem is property testing of monotonicity: given oracle access to a function f: {0,1}^n → {0,1}, distinguish with high probability if f is monotone or if it is ε-far from being monotone (requiring at least an ε-fraction of outputs to be changed to become monotone). A recent breakthrough by Chen et al. (Theorem 1.1) established that any two-sided, adaptive tester for this task requires Ω(n^{1/2-c}) queries. However, a key limitation was that the distance parameter ε_c in their lower bound shrinks to zero as c approaches zero (ε_c = 2^{-Θ(1/c)}). This left open the possibility of very efficient testers (e.g., with O(n^{0.49}) queries) for some fixed, small constant ε (like 0.001).

The main contribution of this work is Lemma 1.2, which provides a randomized transformation that amplifies the distance from monotonicity. Given oracle access to an n-variate function f, it produces oracle access to a new function f’ on n’ ≤ n * 2^{O(1/ε)} variables. The transformation has two crucial guarantees: (1) if f is monotone, then f’ is monotone with probability 1; (2) if f is ε-far from monotone, then with high probability f’ is δ-far from monotone for some absolute constant δ > 0 (which can be made arbitrarily close to ln2/2). Furthermore, each query to f’ can be simulated by making 2^{O(1/ε)} queries to f.

The construction uses the well-known “Tribes” function. The transformation works by evaluating the original function f on k = 2^{Θ(1/ε)} independent blocks of input, resulting in a k-bit string. This string is then fed as input to a carefully chosen Tribes function T to determine the output of f’. The Tribes function, a read-once DNF formula with specific parameters, exhibits a valuable combinatorial property: under a suitable biased measure, there is a constant probability (≈ ln2/2) that its output is 1 and this output is “on the vertex-boundary,” meaning it is sensitive to flipping a single specific input bit.

The proof leverages this property. If f is ε-far from monotone, a classical result (Lemma 2.1) implies the existence of a large matching of “violating pairs” in its violation graph. When evaluating f’, the boundary property of the Tribes function creates a scenario where many of the k blocks correspond to endpoints of this matching. By strategically flipping these blocks according to the matching, one can construct a new, large set of violating pairs for f’, thereby proving that f’ is Ω(1)-far from monotone.

The power of this lemma is demonstrated in Theorem 1.3. By applying the amplification lemma to the hard instances from the earlier lower bound (Theorem 1.1), one can “boost” the distance parameter from a vanishing ε_c to a fixed constant δ. Specifically, one sets c to be very small (e.g., O(1/log log n)), so that ε_c is about 1/√(log n). The lemma then transforms an n-variable instance into an n’ ≈ n^{1+o(1)}-variable instance. If there existed an adaptive tester making o(n’^{1/2}) queries for the constant distance δ, one could use it with the transformation to solve the original problem with ε_c distance in o(n^{1/2}) queries, contradicting Theorem 1.1. This yields the final result: any two-sided, adaptive tester distinguishing monotone functions from those δ-far from monotone must make Ω(n^{1/2-o(1)}) queries.

In summary, this paper provides an elegant distance amplification technique using the Tribes function, which effectively eliminates the dependence of the lower bound on a vanishing distance parameter. It solidifies the understanding that testing monotonicity adaptively requires query complexity essentially on the order of √n, even for a constant promise gap.