[Paper Review] Alternative to the Romberg Method of Estimating the Definite Integral

Michael Brandon Youngberg|arXiv (Cornell University)|Jul 25, 2012

Advanced Thermodynamics and Statistical Mechanics1 references3 citations

TL;DR

This paper proposes a novel numerical integration method that uses a weighted average of the trapezoidal and composite trapezoidal rules to achieve faster convergence than Romberg integration. By deriving an extrapolation technique based on predictable error patterns, the method attains high accuracy with fewer subintervals—demonstrated to reach 15-digit precision in 15 approximations versus Romberg’s 21, and effective even with non-power-of-two subintervals.

ABSTRACT

Using elementary methods, we define and derive a particular weighted average of the trapezoidal and composite trapezoidal rules and show that this approximation, as well as its composite, is straightforward in computation. This approximation and its composite, in their general forms, are shown to have predictable error patterns; thus, an extrapolation method can be used to increase the accuracy. We then derive the necessary weights to use an extrapolation method to reduce error and converge more quickly than Romberg integration by allowing for improved accuracy with fewer necessary subintervals. The procedure necessary to implement this alternative method is then carefully described, followed by two examples.

Motivation & Objective

• To develop a numerically stable, computationally efficient alternative to the Romberg method for definite integral estimation.
• To reduce the number of required subintervals while maintaining or improving convergence speed compared to Romberg integration.
• To enable accurate integration using composite approximations even when the number of subintervals is not a power of two, a key limitation of Romberg’s method.
• To derive an extrapolation framework that systematically eliminates leading error terms using weighted combinations of trapezoidal approximations.
• To demonstrate that the method achieves higher accuracy with fewer approximations than Romberg, particularly in non-ideal subinterval counts.

Proposed method

• The method begins with a weighted average of the standard trapezoidal rule and the composite trapezoidal rule, denoted as $ A $, which is derived using elementary calculus and error balancing.
• The initial error $ \alpha $ is computed by weighting the errors of the two component rules using the same coefficients as in the approximation $ A $, enabling systematic error cancellation.
• A composite version $ A_m $ is defined for $ m $ subintervals, and its error structure is analyzed to identify a predictable pattern suitable for extrapolation.
• An extrapolation factor $ \Omega $ is derived to eliminate the leading error term between $ A $ and $ A_m $, and general forms of $ \Omega $ are derived for all feasible combinations.
• Successive extrapolation factors are computed to eliminate higher-order error terms, enabling convergence to $ O(n^2 h^{16}) $ accuracy with appropriate subinterval choices.
• The implementation procedure is formalized, including how to select and combine composites to maximize accuracy with minimal computational cost.

Experimental results

Research questions

• RQ1 Can a weighted average of the trapezoidal and composite trapezoidal rules be constructed such that its error pattern enables effective extrapolation?
• RQ2 How can extrapolation factors be derived to eliminate leading error terms in this new method, and what is their general form?
• RQ3 Does this method achieve faster convergence than Romberg integration, especially when the number of subintervals is not a power of two?
• RQ4 What is the maximum achievable accuracy of the method, and how does it scale with the number of available composites?
• RQ5 Can the method maintain high accuracy with fewer approximations than Romberg while being applicable to arbitrary subinterval counts?

Key findings

• The proposed method achieves $ O(n^2 h^{16}) $ accuracy with 24 or 30 subintervals, a level not attainable by Romberg integration until 64 subintervals.
• The method reaches 15-digit precision in 15 approximations, compared to 21 approximations required by Romberg integration for the same accuracy.
• For the test function $ f(x) = \sin(x) $ on $ [0, \pi] $, the method produced an approximation of $ -2.00000000000133 $ at $ h = \pi/32 $, matching the true value to 15 digits.
• The method remains effective even with odd numbers of subintervals, unlike Romberg integration, which requires doubling subintervals and fails when counts are not powers of two.
• The extrapolation framework successfully eliminates error terms up to $ O(n^2 h^{12}) $, and further improvement is possible with additional unused composites.
• The method’s error structure is predictable and amenable to systematic extrapolation, enabling higher accuracy with fewer required function evaluations than Romberg.