On the Fourier Tails of Bounded Functions over the Discrete Cube Guy Kindler Microsoft Research Joint work with Irit Dinur, Ehud Friedgut, and Ryan O’Donnell Fourier Analysis Fourier Analysis Fourier representation: can be written as a multilinear polynomial is called the S Fourier coefficient of f. Fourier Analysis Fourier representation: can be written as a multilinear polynomial is called the S Fourier coefficient of f. Many structural properties of f can be inferred from its Fourier representation. Useful in: hardness of approximation, circuit lower bounds, threshold phenomena, metric embeddings, algorithms, learning, communication complexity, complexity,… Boolean vs. Bounded functions Often one needs to study averages of Boolean functions. Question: which properties persist for bounded functions? Our initial motivation: coloring. Ideas used in [KO 05] and [ABHKS 05]. What next: Some technical background Some symmetry breaking phenomena for Boolean functions Main theorem: symmetry breaking for bounded functions Something about the proof. On weights and tails k-tail of f: Low-degree part of f: Weight: k-tail weight: Dinstance: Parseval’s identity. On Juntas and tails A J-junta: a function f that depends on at most J coordinates. Often: having small k-tail weight implies f is junta-ish. f is an (,J)-junta if 9 a J junta g such that breaking. Symmetry [FKN 02] [B 02] ! f is an (O(),1)-junta. ! f is an (0.001,100k)-junta. For majority, the weight of the k-tail is . Tails of bounded functions A J-junta: a function f that depends on at most J coordinates. Often: having small k-tail weight implies f is junta-ish. f is an (,J)-junta if 9 a J junta g such that [FKN 02] [B 02] ! f is an (O(),1)-junta. ! f is an (0.001,100k)-junta. For majority, the weight of the k-tail is . Tails of bounded functions Is a threshold for k-tail bounded function? No: We have symmetric f with Does there really exist a threshold ?? Theorem: If then it is an -junta. what’s next: Some technical background Some symmetry breaking phenomena for Boolean functions Main theorem: symmetry breaking for bounded functions Something about the proof. Proof idea: use large deviations Theorem: If then it is an -junta. Idea: If f<k is smeared over many coordinates then it must obtain large values. So fk must also obtain large values, and therefore have large weight. We need a lower-bound on large deviations for low-degree functions. Large deviation lower bounds , Linear case (folklore): and for all i. Then , Main lemma: and , for all i. Then , Vague idea of the proof f^{=1} f^{=1} f^{=2} f^{=1} f^{=2} f^{=3} f^{=3} f x N_0.1(x) N_0.2(x) N_0.3(x) N_0.4(x) N_0.5(x) Conclusions and questions Bounded functions do show symmetry-breaking phenomena. This happens for different reasons and parameter-range than in the Boolean case. Is there a generalization of Boolean functions where the same symmetry-breaking phenomena hold? Get other bounded-case analogues for Boolean results.