The Erdos similarity conjecture, posed decades ago, asks whether for any "small" set A of real numbers, one can find a large set E on the real line that contains no stretched or shifted copy of A. Here "large" means E has nearly full measure locally, that is, in every interval of length 1, the set E occupies at least a fraction (1 minus epsilon) of the length, for any epsilon you choose in advance. The conjecture has resisted proof in general, but this paper establishes it for a broad and natural class of sets A, namely those that can support a Rajchman measure.
A Rajchman measure is a probability measure whose Fourier transform fades away to zero at high frequencies. This includes measures supported on many well-studied fractal and arithmetic sets. The key insight the authors exploit is that when you dilate such a measure by a large scaling factor, the mass becomes equidistributed modulo 1, meaning it spreads out evenly around the circle in the limit. The authors combine this equidistribution property with a clever multiscale construction of "blocker" sets, periodic sparse obstacles built at many scales, to produce a closed set E that is large in the measure sense yet systematically avoids every affine copy (stretch and shift) of A. Crucially, the argument does not require any specific rate at which the Fourier transform decays, only that it eventually goes to zero.
The paper also sharpens a classical result about how small a Rajchman support can be. It shows that for any prescribed gauge function h (a generalized measure of size), one can construct a compact set K that is null with respect to h, yet whose box-counting and packing dimensions, measured with a logarithmic correction, both equal 1, the largest possible value. This means the set is geometrically rich in a dimensional sense while still being tiny in the sense of h-measure, and the value 1 cannot be improved, making the result sharp.