TESTING LINEAR-INVARIANT PROPERTIES

We study the property testing of functions Fnp [R] for fixed prime p and positive integer R. We work in the natural model where we are allowed to query the function on a random subspace of constant dimension. We say that a property is testable if queries of this form can detect the property with one-sided error. Furthermore, a property is proximity oblivious-testable (POtestable) if the test is also independent of the proximity parameter. It is known that a number of natural properties such as linearity and being a low degree polynomial are PO-testable. These properties are examples of linear-invariant properties, meaning that they are preserved under linear automorphisms of the domain. Following work of Kaufman and Sudan, the study of linear-invariant properties has been an important problem in arithmetic property testing. A central conjecture in this field, proposed by Bhattacharyya, Grigorescu, and Shapira, is that a linear-invariant property is testable if and only if it is semi-subspace-hereditary. We prove two results; the first resolves this conjecture and the second classifies PO-testable properties: (1) A linear-invariant property is testable if and only if it is semi-subspace-hereditary. (2) A linear-invariant property is PO-testable if and only if it is locally characterized. Our innovations are twofold. We give a more powerful version of the compactness argument first introduced by Alon and Shapira. This relies on a new strong arithmetic regularity lemma in which one mixes different levels of Gowers uniformity. This allows us to extend the work of Bhattacharyya, Fischer, Hatami, Hatami, and Lovett by removing the bounded complexity restriction in their work. Our second innovation is a novel recoloring technique called patching that builds on earlier work by the authors and Fox. This Ramsey-theoretic technique is critical for working in the linear-invariant setting and allows us to remove the translation-invariant restriction present in previous work.