TCSLib

23 Boolean Analysis — Circuit Layer Reduction

23.1 Overview

This module proves the core iterative reduction theorem of the LMN argument: under a composed Bernoulli random restriction, a function computed by a bounded-depth, bounded-fan-in circuit has small decision-tree depth with high probability. The proof proceeds by strong induction on the circuit depth \(d\), with a base case (\(d=2\)) handled by the switching lemma and an inductive step (\(d\ge 3\)) that factors the restriction parameter and applies a two-stage argument to the children of the root gate.

23.2 Declarations

Definition 23.1 Composed restriction parameter
#

For a fan-in bound \(w\), a width parameter \(l\), and a depth \(d\), the composed Bernoulli restriction parameter is

\[ \mathrm{composedDelta}(w,l,d) = \frac{1}{40w}\cdot \left(\frac{1}{40l}\right)^{d-2}. \]
Lemma 23.2 Positivity of the composed parameter
#

For \(w\gt 0\) and \(l\gt 0\), the composed restriction parameter \(\mathrm{composedDelta}(w,l,d)\) is positive.

Lemma 23.3 The composed parameter is at most one
#

For \(w\ge 1\), \(l\ge 1\), and \(d\ge 2\), the composed restriction parameter satisfies \(\mathrm{composedDelta}(w,l,d)\le 1\).

Lemma 23.4 Recursive step factorization (left)
#

For natural numbers \(w,l\) and depth \(d\ge 3\),

\[ \mathrm{composedDelta}(w,l,d) = \frac{1}{40w}\cdot \mathrm{composedDelta}(l,l,d-1). \]
Lemma 23.5 Recursive step factorization (right)
#

For \(d\ge 3\) and \(l\gt 0\), the composed parameter factors as the \((d-1)\) case times \(1/(40l)\):

\[ \mathrm{composedDelta}(w,l,d) = \mathrm{composedDelta}(w,l,d-1)\cdot \frac{1}{40l}. \]
Definition 23.6 Layer-2 data
#

A bundle describing the second layer of a circuit over \(n\) variables: a number of gates, a family of DNF formulas indexed by the gates, a common width bound \(\mathrm{width}\gt 0\) such that every gate’s width is at most \(\mathrm{width}\), together with the hypothesis that within each term of each gate distinct literals have distinct variables.

Theorem 23.7 One-step switching for layer-2 data

For layer-2 data and a Bernoulli parameter \(p\) with \(0\lt p\le 1/(40\cdot \mathrm{width})\) and \(p\le 1\), the probability that some gate’s restricted function has decision-tree depth greater than \(l\) is at most

\[ \mathrm{numGates}\cdot \left((1/2)^{l} + \exp (-np/3)\right). \]

For layer-2 data and a Bernoulli parameter \(p\) with \(0\lt p\le 1/(40\cdot \mathrm{width})\) and \(p\le 1\), the probability that it fails that every gate’s restricted function is computed by some width-\(l\) CNF is at most

\[ \mathrm{numGates}\cdot \left((1/2)^{l} + \exp (-np/3)\right). \]
Theorem 23.9 Aggregating per-layer bounds
#

Given layer sizes summing to at most \(s\), if each per-layer quantity is bounded by \(\mathrm{layerSize}(i)\cdot \alpha \) (with \(\alpha \ge 0\)) and a final term is bounded by \(\beta \), then the total \(\sum _i \mathrm{per\_ layer}(i) + \mathrm{final}\) is at most \(s\cdot \alpha +\beta \).

Lemma 23.10 Restriction commutes with restriction composition

For a Boolean function \(f\) and restrictions \(\rho _1,\rho _2\), restricting by the composed restriction equals restricting by \(\rho _1\) and then by \(\rho _2\):

\[ \mathrm{restrictFn}\, f\, (\mathrm{composeRestr}\, \rho _1\, \rho _2) = \mathrm{restrictFn}\, (\mathrm{restrictFn}\, f\, \rho _1)\, \rho _2. \]
Theorem 23.11 Two-stage probability bound

Let \(p_1,p_2\in (0,1]\), let \(E\) and \(A\) be predicates on restrictions, and let \(\beta \ge 0\). If for every \(\rho _1\) with \(\neg A(\rho _1)\) the conditional probability \(\Pr _{p_2}[E(\mathrm{composeRestr}\, \rho _1\, \rho _2)]\le \beta \), then

\[ \Pr _{p_1 p_2}[E] \le \Pr _{p_1}[A] + \beta . \]
Lemma 23.12 Monotonicity in the depth threshold

For \(0\le p\le 1\), a Boolean function \(f\), and thresholds \(l_1\le l_2\), the probability that the restricted function exceeds depth \(l_2\) is at most the probability that it exceeds depth \(l_1\).

Lemma 23.13 Congruence under pointwise-equal functions

If \(f\) and \(g\) agree on every input, then the probability that the restriction of \(f\) exceeds depth \(t\) equals the corresponding probability for \(g\).

If \(f\) is computed by a circuit of depth at most \(2\), size at most \(s\), and fan-in at most \(w\), then for a Bernoulli parameter \(p\) with \(0\lt p\le 1/(40w)\) and \(p\le 1\),

\[ \Pr _{p}\! \left[\mathrm{dtDepth}(\mathrm{restrictFn}\, f\, \rho ) \gt t\right] \le (1/2)^{t} + \exp (-np/3). \]

The base case (\(d=2\)) of the inductive bound: if \(f\) is computed by a depth-\(2\) circuit of size at most \(s\) and fan-in at most \(w\) (with \(s,w,l\gt 0\) and \(n\gt 0\)), then under the composed restriction parameter \(\mathrm{composedDelta}(w,l,2)\),

\[ \Pr \! \left[\mathrm{dtDepth}(\mathrm{restrictFn}\, f\, \rho ) \gt t\right] \le (s-1)(1/2)^{l} + (1/2)^{t} + (s-1)\exp \! \left(-\tfrac {n}{120w}\right) + s\, \exp \! \left(-\tfrac {n}{120l}\right). \]
Lemma 23.16 Union bound over a list of circuits

For \(0\le p\le 1\), a list of circuits \(\mathit{cs}\), and a predicate \(\mathrm{bad}\) on circuits and restrictions, the probability that some circuit in \(\mathit{cs}\) is bad is bounded by the folded sum of the per-circuit probabilities \(\sum _{c\in \mathit{cs}}\Pr _{p}[\mathrm{bad}\, c]\).

The inductive step: for depth \(d\ge 3\), assuming the bound holds for depth \(d-1\) (the hypothesis \(\mathit{ih}\) applied to each child of the root), a circuit of depth at most \(d\), size at most \(s\), and fan-in at most \(w\) satisfies, under \(\mathrm{composedDelta}(w,l,d)\),

\[ \Pr \! \left[\mathrm{dtDepth}(\mathrm{restrictFn}\, f\, \rho ) \gt t\right] \le (s-1)(1/2)^{l} + (1/2)^{t} + (s-1)\exp \! \left(-\tfrac {n}{120w}\right) + s\, \exp \! \left(-\tfrac {n}{120l}\right). \]

By strong induction on the depth \(d\ge 2\): if \(f\) is computed by a circuit of depth at most \(d\), size at most \(s\), and fan-in at most \(w\) (with \(s,w,l,n\gt 0\)), then under \(\mathrm{composedDelta}(w,l,d)\),

\[ \Pr \! \left[\mathrm{dtDepth}(\mathrm{restrictFn}\, f\, \rho ) \gt t\right] \le (s-1)(1/2)^{l} + (1/2)^{t} + (s-1)\exp \! \left(-\tfrac {n}{120w}\right) + s\, \exp \! \left(-\tfrac {n}{120l}\right). \]

Replacing the coefficient \(s-1\) by \(s\) in the previous bound: under the same hypotheses and the parameter \(\mathrm{composedDelta}(w,l,d)\),

\[ \Pr \! \left[\mathrm{dtDepth}(\mathrm{restrictFn}\, f\, \rho ) \gt t\right] \le s(1/2)^{l} + (1/2)^{t} + s\, \exp \! \left(-\tfrac {n}{120w}\right) + s\, \exp \! \left(-\tfrac {n}{120l}\right). \]

The user-facing form of the iterative reduction: for \(f\) computed by a circuit of depth at most \(d\ge 2\), size at most \(s\), and fan-in at most \(w\), under \(\mathrm{composedDelta}(w,l,d)\),

\[ \Pr \! \left[\mathrm{dtDepth}(\mathrm{restrictFn}\, f\, \rho ) \gt t\right] \le s(1/2)^{l} + (1/2)^{t} + s\, \exp \! \left(-\tfrac {n}{120w}\right) + s\, \exp \! \left(-\tfrac {n}{120l}\right). \]