In the distribution-free property testing model, the distance between functions is measured with respect to an arbitrary and unknown probability distribution D over the input domain. We consider distribution-free testing of several basic Boolean function classes over {0,1\}^n, namely monotone conjunctions, general conjunctions, decision lists, and linear threshold functions. We prove that for each of these function classes, Omega((n/\log n)^{1/5}) oracle calls are required for any distribution-free testing algorithm. Since each of these function classes is known to be distribution-free learnable (and hence testable) using Theta(n) oracle calls, our lower bounds are within a polynomial factor of the best possible.
pdf of full version of
conference paper
Link to open-access
journal version