Amortized Circuit Complexity, Formal Complexity Measures, and Catalytic Algorithms

We study the amortized circuit complexity of boolean functions. Given a circuit model $\mathcal{F}$ and a boolean function $f: \{0, 1\}^{n}\rightarrow\{0, 1\}$, the $\mathcal{F}$ -amortized circuit complexity is defined to be the size of the smallest circuit that outputs $m$ copies of $f$ (evaluated on the same input), divided by $m$, as $m\rightarrow\infty$. We prove a general duality theorem that characterizes the amortized circuit complexity in terms of “formal complexity measures”. More precisely, we prove that the amortized circuit complexity in any circuit model composed out of gates from a finite set is equal to the pointwise maximum of the family of “formal complexity measures” associated with $\mathcal{F}$. Our duality theorem captures many of the formal complexity measures that have been previously studied in the literature for proving lower bounds (such as formula complexity measures, submodular complexity measures, and branching program complexity measures), and thus gives a characterization of formal complexity measures in terms of circuit complexity. We also introduce and investigate a related notion of catalytic circuit complexity, which we show is “intermediate” between amortized circuit complexity and standard circuit complexity, and which we also characterize (now, as the best integer solution to a linear program). Finally, using our new duality theorem as a guide, we strengthen the known upper bounds for non-uniform catalytic space, introduced by Buhrman et. al [1] (this is related to, but not the same as, our notion of catalytic circuit size). Potechin [2] proved that for any boolean function $f: \{0, 1\}^{n}\rightarrow\{0, 1\}$, there is a catalytic branching program computing $m=2^{2^{n}-1}$ copies of $f$ with total size $O(mn)$ -that is, linear size per copy — refuting a conjecture of Girard, Koucký and McKenzie [3]. Potechin then asked if the number of copies $m$ can be reduced while retaining the amortized upper bound. We make progress on this question by showing that if $f$ has degree $d$ when represented as polynomial over $\mathbb{F}_{2}$, then there is a catalytic branching program computing $m=2^{\begin{pmatrix}n\\ \leq d\end{pmatrix}}$ copies of $f$ with total size $O(mn)$.

Authors

• Robert Robere
• Jeroen Zuiddam

Keywords

• Computer science
• Boolean functions
• Upper bound
• Logic gates
• Complexity theory
• Integrated circuit modeling
• Standards
• Complexity Measures
• Formal Measures
• Complex Circuits
• Lower Bound
• Total Size
• Linear Programming
• Circuit Model
• Linear Size
• Dual Theory
• Boolean Function
• Circuit Size
• Information Theory
• Matrix Multiplication
• Complex Communication
• Direct Sum
• Semigroup
• Overview Of Techniques
• Notion Of Space
• Multiset
• Turing Machine
• End Of The Computation
• Subadditivity
• Sink Node
• Tensor Rank
• Collection Of Functions
• Boolean Model
• Amortized Complexity
• Branching Programs
• Catalytic Computation
• Comparator Circuits
• Duality
• Formal Complexity Measures
• Submodular