arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2603.15565cs.CC
收藏

更小深度的2线性电路用于不相交矩阵

Improved Depth-2 Linear Circuits for Disjointness via Quenched Lyapunov Exponents

Institute for Interdisciplinary Information Sciences, Tsinghua University(清华大学交叉信息研究院)

Lixi Ye

AI总结:

本文改进了计算第N个不相交矩阵D⊗N的深度2线性电路的上界,通过改进的重平衡框架和凸优化上界,获得了更优的电路规模和度数界。

摘要:

我们证明了两个新的深度2线性电路的上界,用于计算第N个不相交矩阵D⊗N。首先,我们得到一个规模为O(2^{1.24485N})的电路,其次得到一个度数为O(2^{0.3199N})的电路。这些结果改进了Alman和Li之前的界限,即规模O(2^{1.249424N})和度数O(2^{N/3})。我们的起点是Jukna和Sergeev等人的重平衡框架,最终由Alman和Li完成。我们通过两种方式改进了该框架:首先,将之前的“野性”重平衡过程替换为一个驯服、离散化的过程,其几何平均行为由随机矩阵乘积的淬火上Lyapunov指数控制。这使我们能够调用Gharavi和Anantharam的凸优化上界。其次,对于度数界,我们显式地考虑了(p,q)-平面上的成本景观,并证明了不同电路家族在不同区域主导,因此全局最大值仍低于0.3199。

英文摘要:

Let $R_1 := \begin{pmatrix}1&1\\1&0\end{pmatrix}$, and write $R_n := R_1^{\otimes n}$ for the $N \times N$ disjointness matrix, where $N = 2^n$. We construct new depth-2 linear circuits for $R_n$: one of size $O(N^{1.2449})$ and one of maximum input-output degree $O(N^{0.3199})$, improving upon the results of Alman and Li [AL25] (FOCS 2025), who achieved size $O(N^{1.2495})$ and degree $O(N^{1/3})$. In this paper, we develop a number of mechanisms for designing depth-2 linear circuits that compute linear transforms represented by Kronecker powers of a general base matrix. For $R_1$ in particular, we obtain our refined bounds by expressing a certain recursive circuit construction as a random matrix product and analyzing its quenched Lyapunov exponent, a quantity studied in the theory of random dynamical systems. Our new upper bounds imply faster deterministic algorithms for #OV, more efficient online data structures for the Subset Query and Partial Match primitives, and smaller low-depth linear circuits for both disjointness and a class of related non-Kronecker-power matrices.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑