[Paper Review] On the norm of inverses of confluent Vandermonde matrices

Dmitry Batenkov|arXiv (Cornell University)|Dec 2, 2012

Matrix Theory and Algorithms11 references3 citations

TL;DR

This paper presents a new upper bound for the row-wise ℓ₁-norm of the inverse of general confluent Vandermonde matrices, leveraging derivative estimates of rational functions and combinatorial bounds. The key result shows that the norm grows polynomially in the inverse separation distance δ and exponentially in the matrix size N, with explicit dependence on the multiplicity structure and point locations within the unit disk.

ABSTRACT

In this note we present a simple upper bound for the row-wise norm of the inverses of general confluent Vandermonde matrices.

Motivation & Objective

• To derive a general upper bound for the ℓ₁-norm of rows in the inverse of confluent Vandermonde matrices beyond the limited case of multiplicity ≤2.
• To extend Gautschi’s earlier bounds to arbitrary multiplicity configurations ℓ_j ≥ 1.
• To quantify the dependence of inverse matrix row norms on the separation distance δ and the total size N of the matrix.
• To provide a technically self-contained proof using explicit inverse formulas and derivative estimates of rational functions.

Proposed method

• The method uses explicit formulas for entries of the inverse of confluent Vandermonde matrices derived from Hermite interpolation theory.
• It applies a technical lemma estimating the t-th derivative of the rational function h_j(x) = ∏_{i≠j} (x−x_i)^{-ℓ_i} at x_j.
• The bound relies on the Leibniz rule and induction to control the growth of derivatives in terms of δ and N.
• It combines these derivative bounds with coefficient sum estimates for monic polynomials, using (1+|x_j|)^k ≤ 2^k for |x_j|≤1.
• The proof introduces a binomial-type summation over t to bound the total ℓ₁-norm of the row vector u_{j,k}.
• A final simplification uses the inequality (1 + a)^m ≤ (1 + a/m)^m ≤ e^a to bound the sum, yielding a clean exponential form.

Experimental results

Research questions

• RQ1 How does the ℓ₁-norm of rows in the inverse of a confluent Vandermonde matrix scale with the matrix size N and the separation distance δ between nodes?
• RQ2 Can Gautschi’s bounds for low multiplicity (ℓ_j ≤ 2) be generalized to arbitrary multiplicity configurations?
• RQ3 What is the dependence of the inverse matrix row norms on the local multiplicity structure ℓ_j and the position of nodes within the unit disk?
• RQ4 How can the entries of the inverse be bounded using only the separation δ and combinatorial structure of the interpolation points?
• RQ5 Is it possible to derive a uniform bound that depends only on δ and N, independent of the specific configuration beyond separation and boundedness?

Key findings

• The ℓ₁-norm of the row u_{j,k} in V^{-1} is bounded by (2/δ)^N × 2/(k!) × (1/2 + N/δ)^{ℓ_j−1−k}.
• The bound grows exponentially with N, reflecting the ill-conditioning of high-order confluent systems.
• The dependence on δ is polynomial, with (2/δ)^N capturing the sensitivity to node clustering.
• The bound is independent of the specific values of x_j as long as |x_j| ≤ 1 and |x_i − x_j| ≥ δ for i ≠ j.
• The result generalizes Gautschi’s earlier bounds to arbitrary multiplicities ℓ_j, extending their applicability to higher-order numerical methods.
• The bound is sharp in its dependence on δ and N, and explicitly accounts for the local multiplicity structure via the exponent ℓ_j−1−k.