Limiting negations in bounded treewidth and upward planar circuits

TLDR: In this paper, it was shown that 8k/2k negation gates are sufficient and sufficient for any Boolean function f to be computed by a monotone one-input-face upward planar circuit.

Abstract: The decrease of a Boolean function f: {0, 1}n → {0, 1}, denoted by d(f) is the maximum number of inverse indices in any increasing chain of inputs x1,..., xl ∈ {0,1}n, where i is an inverse index if f(xi) > f(xi+1). It follows from a theorem of Markov (JACM 1958) that the minimum number of negation gates in a circuit necessary and sufficient to compute any Boolean function f is ⌈log(d(f) + 1)⌉. A recent result due to Morizumi (ICALP 2009) proves that d(f) negations are necessary and sufficient when the computation is done by formulas. We explore the situation in between formulas (directed trees) and general circuits (DAGs), and related models. We obtain the following results: 1. We argue that for any Boolean function f, there is a circuit computing f, that uses ⌈log(d(f) + 1)⌉ negations and has treewidth at most ⌈log(d(f)+1)⌉ + 1. For 1 ≤ k ≤ ⌈log(d(f)+1)⌉, we prove that d(f). 8k/2k negations are sufficient to compute any Boolean function f by circuits of treewidth at most k. Moreover, if there is a circuit family of size s = s(n) and treewidth k = k(n) computing {fn}, then there exists a circuit family of size sċnO(1)ċ2O(min{k, log n}) and treewidth at most 2k which computes {fn} and contains O(max{nk/22k, log n}) negation gates. 2. We obtain tight bounds on the number of negation gates required to compute specific functions such as PARITYn, PARITYn and INVERTERn by one-input-face upward planar circuits. We extend these lower bounds to a larger class of functions (which also includes natural functions like ADD and SUBTRACT) and we show a direct sum theorem for this class with respect to the number of negations. 3. We demonstrate the limitations of the one-input-face constraint in the upward planar circuits by showing the explicit function which can be computed by a monotone upward planar circuit, but cannot be computed by any montone one-input-face upward planar circuit. 4. We prove that for every Boolean function f, there exists a multi-lective upward planar circuit which uses at most ⌈d(f)+1/2⌉ negation gates for computing f.