Stephan Scholz and Lothar Berger have introduced a computational framework combining Faà di Bruno’s formula and Bell polynomials to compute the mathematical structure of high-order derivatives for function compositions, demonstrating its utility in flatness-based open-loop control of diffusion problems. In physical systems governed by partial differential equations, calculating control trajectories requires determining numerous continuous derivatives of composite mathematical expressions. While standard differential calculus employs the elementary chain rule to differentiate composite functions, computing higher-order derivatives manually becomes excessively complex, cumbersome, and susceptible to severe human error. To resolve these challenges, the researchers devised an algorithmic approach that explicitly separates the underlying algebraic structure of composite derivatives from specific function evaluations, enabling automated and reusable derivative construction for large-scale engineering systems.
The investigative context centers on flatness-based control, a mathematical design technique widely utilized to formulate feedforward trajectory planning algorithms for distributed parameter systems and large-scale partial differential equations. Within boundary control formulations for diffusion-convection-reaction phenomena, steering a system between defined states requires generating input signals from smooth reference trajectories. A fundamental mathematical element in these control architectures is the smooth bump function, defined as zero outside a specified time interval and governed by an exponential of an inverted polynomial within that interval. Because the bump function belongs to the Gevrey class of functions, it exhibits exceptional smoothness characteristics that make it ideal for synthesizing open-loop control trajectories. However, generating an exact flatness-based boundary input requires evaluating numerous successive derivatives of this composite function, demanding efficient computational methods.
Prior approaches to evaluating composite derivatives in control engineering have faced notable technical limitations across numerical, symbolic, and algorithmic methodologies. Numerical differentiation schemes, such as finite difference approximations, frequently experience severe stability degradation and numerical inaccuracies when applied to high-order derivatives because their precision depends strictly on the approximation order. Conversely, symbolic differentiation implemented in computer algebra software platforms like Maxima or Mathematica produces exact mathematical formulations, but transferring the resulting unwieldy symbolic expressions into efficient, executable source code for control hardware remains labor-intensive and computationally cumbersome. While algorithmic differentiation has recently advanced machine learning and mathematical optimization by efficiently generating gradients and Hessian matrices, Scholz and Berger sought an alternative direct mathematical formulation that constructs reusable derivative polynomial structures without relying on symbolic computation packages.
The foundational mathematical cornerstone of the new methodology is Faà di Bruno’s formula, which generalizes the traditional elementary chain rule to higher differentiation orders. For a composite function mapping real inputs through an inner function and an outer function, the formula expresses the n-th derivative as a weighted summation across lower derivative orders scaled by partial Bell polynomials. These partial Bell polynomials aggregate products of successive derivative terms of the inner function, reflecting every valid integer partition of the differentiation order. While Bell polynomials can theoretically be expressed through explicit combinatoric formulas involving factorials and products, computing them directly from formal definitions is computationally impractical. Instead, the researchers utilized a classical triangular recursive relation governed by binomial coefficients, which expresses higher-order partial Bell polynomials through sequential linear combinations of preceding lower-order polynomials.
A central conceptual innovation of the presented methodology is the complete decoupling of polynomial structure generation from individual numerical evaluations. In typical dynamic simulations, evaluating composite derivatives at repeated discrete time steps using recursive relations causes massive computational redundancy. To overcome this inefficiency, Scholz and Berger designed an algorithm that executes the recursive polynomial construction exactly once up to a user-defined differentiation order n. The resulting structure stores the exact integer exponents and corresponding rational coefficients for every monomial term within the partial Bell polynomials. Once these static combinatorial templates are established, specific outer and inner function derivatives can be evaluated and substituted into the precomputed structures at any runtime step, ensuring the structural computation remains completely independent of the underlying physical functions.
The algorithmic construction proceeds through three distinct, systematically structured phases to build the polynomial exponent matrices and numerical coefficients. In the initial initialization stage, the algorithm fixes the maximum differentiation order n and initializes base conditions, leveraging the identities that the zero-order Bell polynomial equals one while out-of-bounds polynomial orders evaluate to zero. For each differentiation level, the algorithm automatically populates the first and last monomial entries, which correspond directly to the isolated n-th inner derivative and the n-th power of the first derivative. By establishing these analytical boundaries early, the framework sets up the index arrays and provides exact boundary constraints before initiating the intermediate recursive matrix operations across successive differentiation orders.
During the secondary matrix computation phase, the algorithm executes three nested loops to systematically generate the multidimensional exponent matrices that characterize every monomial. Iterating through the inner differentiation indices, the procedure constructs transformation matrices containing elementary basis unit vectors, interpreting variable multiplication as an additive shift in integer exponent space. The algorithm performs matrix summations combining shifted sub-matrices from earlier recursive iterations, concatenating the intermediate results into temporary storage matrices. Because multiple recursive paths frequently generate identical monomial terms, the algorithm applies an explicit filtering operation at the conclusion of each inner iteration to extract only unique matrix columns. This filtering prevents combinatorial inflation, ensuring that the final exponent matrices retain strictly unique structural representations of the partial Bell polynomials.
In the third procedural step, the algorithm computes the scalar algebraic coefficients that scale each unique monomial term within the Bell polynomials. Drawing directly from the classical combinatorial definition of partial Bell polynomials, the authors derived an exact factorial formulation that evaluates the coefficient using the differentiation order, the monomial exponents, and factorials of the corresponding derivative indices. These scalar coefficients are stored alongside their associated column exponent vectors to complete the analytical representation of the polynomial. Evaluating the full n-th composite derivative then reduces to a straightforward double summation, scaling the outer function derivatives by the evaluated monomials of the inner function derivatives, entirely bypassing the need for runtime recursive tree traversals or symbolic algebraic manipulation.
The authors conducted extensive computational complexity and performance evaluations to quantify the runtime characteristics of their algorithmic implementation. Analyzing the structural loops revealed distinct scaling behaviors between the exponent matrix generation and the scalar coefficient evaluation. The matrix generation process scales quadratically with respect to differentiation order, requiring a predictable number of inner loop iterations governed by a standard second-order polynomial formula. However, the total number of unique monomials within each partial Bell polynomial follows the combinatorial integer sequence A008284. Because this partition sequence grows rapidly as the differentiation order expands, calculating the scalar coefficients demands significantly more processing time than generating the exponent matrices, emerging as the primary computational bottleneck for high-order derivatives exceeding twenty.
To empirically measure execution benchmarks, Scholz and Berger implemented the entire framework in the Julia programming language and conducted systematic trials across differentiation orders up to forty. The experimental performance tests were executed on an Intel Core i5-8350U central processing unit running at 1.70 gigahertz across eight threads on an Ubuntu 64-bit operating system. Each computational trial was repeated across twenty independent cycles to determine mean execution runtimes as well as strict upper and lower variance bounds. The empirical data demonstrated that while initial cycles exhibited brief latency spikes during low differentiation orders, subsequent executions maintained stable timing. Furthermore, because individual monomial coefficients are calculated independently from pre-existing exponent matrices, the authors noted that this stage offers substantial potential for massive parallelization.
To showcase the real-world engineering utility of their algorithm, the authors applied the methodology to a trajectory planning problem for a one-dimensional heat conduction system. The physical setup modeled thermal diffusion across a one-dimensional solid rod of length one meter over a transition interval of ten seconds, with initial temperatures set to zero. The control objective was to steer the measured boundary temperature at the isolated right boundary from zero to one by applying an open-loop heat flux input signal at the left boundary. Governing boundary conditions assumed active thermal actuation at the left edge and complete thermal insulation at the right edge, requiring the synthesis of a smooth flatness-based input trajectory derived from the system’s distributed partial differential equations.
Synthesizing the required boundary control input signal required evaluating a series expansion involving successive even-order derivatives of a Gevrey-class smooth bump function. Although the theoretical series expansion contains an infinite number of derivative terms, the authors truncated the summation at order ten, noting that the series converges rapidly for practical implementation. To calculate the high-order derivatives, the bump function was factored into an inner rational polynomial and an outer exponential function. The Julia implementation executed the structural derivative algorithm once and applied the precomputed structure twice in sequence: first to differentiate the internal rational composition, and subsequently to differentiate the overarching exponential expression. The resulting computed input signal successfully steered the boundary temperature to its target state without experiencing numerical divergence.
Subject of Research: Algorithmic computation of high-order composite function derivatives for flatness-based boundary control of diffusion equations
Article Title: Fast computation of function composition derivatives for flatness-based control of diffusion problems
Article References: Scholz, S., & Berger, L. (2024). Fast computation of function composition derivatives for flatness-based control of diffusion problems. Journal of Mathematics in Industry, 14(1), Article 15. https://doi.org/10.1186/s13362-024-00143-y
Image Credits: AI Generated
DOI: 10.1186/s13362-024-00143-y
Keywords: Faà di Bruno's formula, Bell polynomials, flatness-based control, function compositions, diffusion problems, partial differential equations, bump function, Gevrey-class functions, heat equation, trajectory planning, algorithmic differentiation, symbolic differentiation, Julia programming language, monomial exponents, open-loop control, integer sequence A008284, boundary control, DifferentialEquations.jl
Cite Scienmag News
Reid Dalton. (November 19, 2026). New Algorithm Accelerates High-Order Derivative Calculations for Diffusion Control Systems. Scienmag. https://scienmag.com/new-algorithm-accelerates-high-order-derivative-calculations-for-diffusion-control-systems/
Reid Dalton. "New Algorithm Accelerates High-Order Derivative Calculations for Diffusion Control Systems." Scienmag, 19 November 2026, https://scienmag.com/new-algorithm-accelerates-high-order-derivative-calculations-for-diffusion-control-systems/. Accessed 19 November 2026.
Reid Dalton. "New Algorithm Accelerates High-Order Derivative Calculations for Diffusion Control Systems." Scienmag. November 19, 2026. https://scienmag.com/new-algorithm-accelerates-high-order-derivative-calculations-for-diffusion-control-systems/








