[Paper Review] The nonzero gain coefficients of Sobol's sequences are always powers of two

Zexin Pan, Art B. Owen|arXiv (Cornell University)|Jun 19, 2021

Mathematical Approximation and Integration23 references4 citations

TL;DR

This paper proves that all nonzero gain coefficients in digital nets and sequences of Sobol’ in base 2 are powers of two, leading to a tighter bound on the maximal gain coefficient Γ ≤ 2^{t+s−1} for (t,m,s)-nets. The result simplifies computation of gain coefficients and improves error bounds for randomized quasi-Monte Carlo methods, particularly for Sobol’ sequences used in high-dimensional integration.

ABSTRACT

When a plain Monte Carlo estimate on $n$ samples has variance $σ^2/n$, then scrambled digital nets attain a variance that is $o(1/n)$ as $n o\infty$. For finite $n$ and an adversarially selected integrand, the variance of a scrambled $(t,m,s)$-net can be at most $Γσ^2/n$ for a maximal gain coefficient $Γ

Motivation & Objective

• To improve the upper bound on the maximal gain coefficient Γ for scrambled digital nets in base 2.
• To establish that all nonzero gain coefficients in base-2 digital nets are powers of two, a previously unnoticed structural property.
• To develop a simplified and more efficient algorithm for computing gain coefficients using this power-of-two property.
• To refine existing microstructure-based bounds on gain coefficients using the new characterization.

Proposed method

• Leveraging microstructure analysis from Niederreiter and Pirsic (2001), the authors derive a bound Γ ≤ 2^{t+s−1} for base-2 digital nets.
• Proving that the gain coefficient Γ_{u,k} is either zero or a power of two by analyzing the rank of generator matrices modulo 2.
• Using the minimal |k| such that C_{u,k} is rank-deficient to characterize the maximal gain coefficient.
• Applying Corollary 4 to show that Γ_{v,k} = 2^{m - rank(C_{v,k})} when the matrix is rank-deficient, and exploiting the binary structure to show that this value is always a power of two.
• Introducing a refined t*-parameter that allows for a tighter bound than the standard t-parameter in certain cases, such as with shift nets.
• Demonstrating that when C_{u,1_u} has full rank, Γ = 2^{t^*_{1:s} + s - 1}, which is strictly smaller than 2^{t+s−1} in some cases.

Experimental results

Research questions

• RQ1 Are all nonzero gain coefficients in base-2 digital nets necessarily powers of two?
• RQ2 Can the standard bound Γ ≤ 2^t 3^s for Sobol’ sequences be improved for base-2 nets?
• RQ3 Does the microstructure analysis in Niederreiter and Pirsic (2001) imply a tighter bound of Γ ≤ 2^{t+s−1} for base-2 nets?
• RQ4 Can the power-of-two property of gain coefficients be exploited to simplify their computation?
• RQ5 Is there a case where t^*_{1:s} < t, leading to a strictly better bound than 2^{t+s−1}?

Key findings

• All nonzero gain coefficients in base-2 digital nets are powers of two, a structural property previously unobserved.
• The maximal gain coefficient Γ for any base-2 digital net satisfies Γ ≤ 2^{t+s−1}, improving upon the prior bound of 2^t 3^s.
• When the matrix C_{1:s,1_{1:s}} has full rank, the maximal gain coefficient is exactly Γ = 2^{t^*_{1:s} + s - 1}, which can be strictly smaller than 2^{t+s−1}.
• The example of a (1,1,4)-net (shift net) shows that t^*_{1:s} = 0, leading to Γ = 8, which is half of the bound 2^{t+s−1} = 16.
• The power-of-two property enables a simplified and more efficient algorithm for computing gain coefficients, as only powers of two need to be considered.
• The bound Γ ≤ 2^{t+s−1} is tight and can be derived from Niederreiter and Pirsic (2001), though this implication was previously overlooked.