Abstract:For any $\{0,1\}$-valued function $f$, its \emph{$n$-folded XOR} is the function $f^{\oplus n}$ where $f^{\oplus n}(X_1, \ldots, X_n) = f(X_1) \oplus \cdots \oplus f(X_n)$. Given a procedure for computing the function $f$, one can apply a ``naive" approach to compute $f^{\oplus n}$ by computing each $f(X_i)$ independently, followed by XORing the outputs. This approach uses $n$ times the resources required for computing $f$.
In this paper, we prove a strong XOR lemma for \emph{information complexity} in the two-player randomized communication model: if computing $f$ with an error probability of $O(n^{-1})$ requires revealing $I$ bits of information about the players' inputs, then computing $f^{\oplus n}$ with a constant error requires revealing $\Omega(n) \cdot (I - 1 - o_n(1))$ bits of information about the players' inputs. Our result demonstrates that the naive protocol for computing $f^{\oplus n}$ is both information-theoretically optimal and asymptotically tight in error trade-offs.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is, in the two - party random communication model, for any given function \( f \), whether the information complexity of its \( n \)-folded XOR function \( f^{\oplus n} \) can be significantly lower than the information complexity required for directly using \( n \) independent computations of \( f \). Specifically, the paper explores whether, when computing \( f^{\oplus n} \), if a certain error rate is allowed, the amount of information required can be reduced, thereby achieving optimal utilization of resources.
### Background of the Paper and Problem Definition
For a function \( f: X \to \{0, 1\} \) and any natural number \( n \), define \( f^{\oplus n}: X^n \to \{0, 1\} \) as:
\[ f^{\oplus n}(X_1, \ldots, X_n) = f(X_1) \oplus \cdots \oplus f(X_n) \]
In the two - party random communication model, Alice receives the input \( X \in X \), Bob receives the input \( Y \in Y \), and the goal is to compute \( f(X, Y) \) by exchanging messages. To compute \( f^{\oplus n} \), Alice receives \( (X_1, \ldots, X_n) \) and Bob receives \( (Y_1, \ldots, Y_n) \), and their goal is to compute \( f^{\oplus n}(X_1, \ldots, X_n, Y_1, \ldots, Y_n) \) by running a protocol.
### Central Problem
The core problem of the paper is to explore the following problem:
**Problem 1**: For which parameter ranges \((\rho, \rho')\) and resource types, does the resource required to compute \( f^{\oplus n} \) with probability \(\rho'\) need to be at least \(\Omega(n)\) times the resource required to compute \( f \) with probability \(\rho\)?
### Specific Research Objectives
In this paper, the author pays special attention to the case where the resource type is information complexity and the parameter range is \((\rho, \rho')=(1 - \frac{1}{n}, \frac{9}{10})\). Specifically, the paper proves a strong XOR lemma, showing that if \( I \) bits of information need to be revealed when the error rate of computing \( f \) is \( O(n^{-1})\), then \(\Omega(n)\cdot(I - 1 - o_n(1))\) bits of information need to be revealed when computing \( f^{\oplus n}\) with a constant error rate.
### Main Results
**Theorem 2 (Strong XOR Lemma)**: There exists a universal constant \(\lambda\in(0, 1)\) and \( c_1>0 \), for any function \( f: X\times Y \to \{0, 1\} \) and any positive integer \( n \), we have:
\[ I(f^{\oplus n}, \frac{1}{10}) \geq c_1 n\cdot\left( I(f, n^{-1}) - \frac{\log(|X|\cdot|Y|)}{n^\lambda - 1} \right) \]
This result is asymptotically tight (ignoring vanishing additive losses) and is further verified by the following result:
\[ I(f^{\oplus n}, \frac{1}{10}) \leq c_2 n\cdot I(f, n^{-1}) \]
### Technical Overview
To prove the above theorem, the author adopts a "conditional" protocol decomposition method. The specific steps include:
1. **Binary Protocol Decomposition**: Decompose the original protocol \(\pi\) into two...