[Paper Review] Multidimensional Divide-and-Conquer and Weighted Digital Sums

Yun Kuen Cheung, Philippe Flajolet+2 more|arXiv (Cornell University)|Feb 28, 2010

Analytic Number Theory Research19 references3 citations

TL;DR

This paper applies Mellin transform techniques to derive exact asymptotic solutions for three algorithmic problems: multidimensional divide-and-conquer recurrences, weighted digital sums of the first type (TS_M(n)), and weighted digital sums of the second type (TW_M(n)). It shows that all three functions exhibit solutions with periodic fluctuations via Fourier series, extending prior work on divide-and-conquer and digital sum analyses with precise, structured asymptotic forms including polylogarithmic terms and periodic coefficients.

ABSTRACT

This paper studies three types of functions arising separately in the analysis of algorithms that we analyze exactly using similar Mellin transform techniques. The first is the solution to a Multidimensional Divide-and-Conquer (MDC) recurrence that arises when solving problems on points in $d$-dimensional space. The second involves weighted digital sums. Write $n$ in its binary representation $n=(b_i b_{i-1}... b_1 b_0)_2$ and set $S_M(n) = \sum_{t=0}^i t^{\bar{M}} b_t 2^t$. We analyze the average $TS_M(n) = \frac{1}{n}\sum_{j i_2 > ... > i_k\geq 0$ and set $W_M(n) = \sum_{t=1}^k t^M 2^{i_t}$. We analyze the average $TW_M(n) = \frac{1}{n}\sum_{j

Motivation & Objective

• To provide exact asymptotic solutions for multidimensional divide-and-conquer recurrences in d-dimensional space.
• To analyze the average behavior of weighted digital sums S_M(n) and W_M(n) via smoothed averages TS_M(n) and TW_M(n).
• To unify the analysis of these diverse algorithmic functions using Mellin transform techniques.
• To derive closed-form solutions with periodic coefficients expressed as absolutely convergent Fourier series.
• To extend prior work on divide-and-conquer and digital sum analyses to higher-dimensional and more complex weighted settings.

Proposed method

• Use Mellin transforms to analyze the generating functions of the recurrence relations and digital sum functions.
• Apply contour integration and residue calculus to extract asymptotic expansions from complex integrals.
• Express the solutions in terms of periodic functions with period one, derived from Fourier series expansions.
• Handle singularities of Dirichlet generating functions using Laurent series and analytic continuation.
• Leverage known bounds on zeta and polylogarithmic functions to control error terms in the asymptotic expansions.
• Decompose complex generating functions into rational and logarithmic components to enable term-by-term residue evaluation.

Experimental results

Research questions

• RQ1 How can Mellin transform techniques be extended to solve multidimensional divide-and-conquer recurrences in d-dimensional space?
• RQ2 What is the exact asymptotic structure of the average weighted digital sum TS_M(n) for polynomial weights t^M in binary digit positions?
• RQ3 How does the second variant of weighted digital sums W_M(n), based on the exponents in the binary representation, behave on average?
• RQ4 Can the periodic fluctuations in these functions be captured precisely using Fourier series with absolutely convergent coefficients?
• RQ5 What is the relationship between the asymptotic forms of MDC, TS_M(n), and TW_M(n), and why do they all exhibit similar functional structures?

Key findings

• The solution to the multidimensional divide-and-conquer recurrence T_d(n) has the form λ_d n lg^{d-1}n + ∑_{m=0}^{d-2} (n lg^m n) A_{d,m}(lg n) + c_d, where A_{d,m}(u) are periodic functions with period one given by absolutely convergent Fourier series.
• The average weighted digital sum TS_M(n) has the same functional form as T_d(n) with d = M+1, confirming a deep structural similarity between MDC and weighted digital sum problems.
• The average W_M(n) function TW_M(n) admits the solution n G_M(lg n) + d_M lg^M n + ∑_{d=0}^{M-1} (lg^d n) G_{M,d}(lg n), where G_M(u) and G_{M,d}(u) are periodic with period one and defined by absolutely convergent Fourier series.
• The periodic components in all solutions arise from residues of complex functions with regularly spaced poles, and their coefficients are derived from Fourier expansions of the generating functions.
• The analysis reveals that the fluctuations in S_M(n) are too large for direct study, justifying the use of smoothed averages like TS_M(n) and TW_M(n).
• The Mellin transform method successfully handles the Dirichlet generating functions of differences of digital functions, enabling exact asymptotic analysis of complex algorithmic recurrences.