MathLabs

Sensitivity conjecture

Solved, 2019Combinatorics and discrete mathematicsApplied and computational mathematics
Statement

For every Boolean function f:{0,1}n→{0,1}f : \{0, 1\}^n \to \{0, 1\}, the sensitivity s(f)s(f) and the block sensitivity bs(f)bs(f) are polynomially related; specifically, bs(f)≤C⋅s(f)cbs(f) \le C \cdot s(f)^c for absolute constants C,c>0C, c > 0 (and in fact deg⁡(f)≤s(f)2\deg(f) \le s(f)^2, where deg⁡(f)\deg(f) is the real polynomial degree of ff).

In July 2019, Hao Huang proved that every induced subgraph HH of the nn-dimensional hypercube QnQ_n with 2n−1+12^{n-1} + 1 vertices has maximum degree Δ(H)≥n\Delta(H) \ge \sqrt{n}. Huang inductively constructed a 2n×2n2^n \times 2^n signed adjacency matrix AnA_n of QnQ_n satisfying An2=nI2nA_n^2 = n I_{2^n}, whose eigenvalues are ±n\pm\sqrt{n} each with multiplicity 2n−12^{n-1}. By Cauchy's interlace theorem, any (2n−1+1)×(2n−1+1)(2^{n-1} + 1) \times (2^{n-1} + 1) principal submatrix of AnA_n has largest eigenvalue at least n\sqrt{n}, forcing Δ(H)≥n\Delta(H) \ge \sqrt{n}. Via the Gotsman–Linial equivalence (1992) and Tal's bound (2013), this immediately implies deg⁡(f)≤s(f)2\deg(f) \le s(f)^2 and bs(f)≤s(f)4bs(f) \le s(f)^4.

Prior to Huang's theorem, decision-tree depth D(f)D(f), certificate complexity C(f)C(f), block sensitivity bs(f)bs(f), polynomial degree deg⁡(f)\deg(f), and quantum query complexity Q(f)Q(f) were all known to be polynomially equivalent, with sensitivity s(f)s(f) being the sole holdout. While deg⁡(f)≤s(f)2\deg(f) \le s(f)^2 is tight, the optimal exponent between bs(f)bs(f) and s(f)s(f)—known to lie between 22 (Rubinstein's quadratic separation) and 44 (Huang + Tal)—remains open.

References

  1. Hao Huang (2019). Induced subgraphs of hypercubes and a proof of the Sensitivity Conjecture · DOI:10.4007/annals.2019.190.3.6 · arXiv:1907.00847
  2. Craig Gotsman, Nathan Linial (1992). The equivalence of two problems on the cube · DOI:10.1016/0097-3165(92)90060-8