Sunday, November 22, 2026
Science
No Result
View All Result
  • Login
  • HOME
  • SCIENCE NEWS
  • CONTACT US
  • HOME
  • SCIENCE NEWS
  • CONTACT US
No Result
View All Result
Scienmag
No Result
View All Result
Home Science News Mathematics

New Algorithm Computes Matrix Exponential and Gramian Cholesky Factor Simultaneously

November 22, 2026
in Mathematics
Reid Dalton
By Reid Dalton Scienmag Editorial Profile - Applied Mathematics
Reading Time: 6 mins read
0
New Algorithm Computes Matrix Exponential and Gramian Cholesky Factor Simultaneously

New Algorithm Computes Matrix Exponential and Gramian Cholesky Factor Simultaneously

65
SHARES
587
VIEWS
Share on FacebookShare on Twitter
ADVERTISEMENT

Researchers Tony Stillfjord and Filip Tronarp have developed an efficient numerical algorithm that computes both the matrix exponential and the Cholesky factor of a finite horizon controllability Gramian simultaneously without assembling the full Gramian matrix product. Applicable to general dense matrices of moderate size, the newly formulated technique generalizes the traditional scaling-and-squaring framework widely utilized for matrix exponentials by exploiting an analogous doubling relation for finite horizon Gramians. This mathematical formulation directly resolves a fundamental vulnerability in computational control theory, where ill-conditioned or nearly singular Gramians cause conventional Cholesky factorization procedures to fail numerically. By evaluating upper triangular Cholesky factors directly from scaled base approximations, the methodology preserves critical numerical rank while guaranteeing backward error within double-precision round-off limits.

In linear control theory and dynamical systems analysis, continuous linear time-invariant systems defined by matrix triplets governing state evolution and output observations possess two primary integral operators over finite intervals: controllability and observability Gramians. When system inputs are modeled as Gaussian white noise, the underlying state evolves as a Gauss-Markov process whose transition probability density depends directly on the controllability Gramian as its covariance matrix. These integral matrices encode essential geometric properties for state-space estimation, balanced model truncation, and optimal control design. Because finite horizon Gramians are inherently positive semi-definite, each matrix admits an upper triangular Cholesky factorization that remains strictly unique whenever the system is fully controllable. However, obtaining this factored representation reliably in finite-precision arithmetic has historically presented severe numerical obstacles in computational linear algebra.

Standard computational workflows often attempt to evaluate the full un-factored Gramian before applying traditional or pivoted Cholesky factorization algorithms, but this two-stage strategy frequently breaks down. Over short integration intervals, or when dynamical systems possess weakly controllable modes, the Gramian develops extremely small eigenvalues and approaches numerical singularity. Standard Cholesky decomposition fails catastrophically on matrices with eigenvalues near machine precision, while pivoted Cholesky variants cannot guarantee numerical stability when matrix dimension is moderately large and the numerical rank is not small. Alternative formulations relying on algebraic Lyapunov equations require the underlying system matrix to be Hurwitz, meaning all eigenvalues must possess strictly negative real parts. Furthermore, generating Cholesky factors via algebraic equations demands intermediate matrix downdates that are notoriously ill-conditioned and prone to complete arithmetic breakdown.

To circumvent these numerical pitfalls, the authors adapted the scaling-and-squaring framework that forms the cornerstone of modern matrix exponential computations. The classical doubling identity for matrix exponentials allows a scaled matrix exponential to be repeatedly squared across multiple doubling steps. Stillfjord and Tronarp capitalized on an analogous doubling relationship for finite horizon Gramians, splitting the defining definite integral into two equal halves. In their formulation, a scaled Gramian factor is updated across successive squaring steps alongside the matrix exponential, requiring only an auxiliary matrix multiplication and a standard QR decomposition to advance the upper triangular factor. This synchronized doubling recursion ensures that computational overhead remains modest while computing both matrix functions concurrently without assembling the full Gramian. Consequently, the doubling phase for the Gramian achieves computational effort comparable to standard exponential algorithms.

A central technical hurdle in executing the doubling recursion involves constructing suitable base approximations on the initial scaled interval before repeated squarings commence. While previous research applied doubling formulas to evaluate matrix Fréchet derivatives from Padé approximations, analogous initializations for Gramian Cholesky factors were previously unknown. The authors resolved this dilemma by establishing a formal connection between diagonal Padé approximants and a Petrov-Galerkin projection over shifted Legendre polynomial bases on the unit interval. By structuring the expansion coefficients to satisfy a block tridiagonal system of linear equations, the summation of Legendre coefficients exactly matches the classical diagonal Padé approximation of the matrix exponential. Simultaneously, the orthogonal properties of Legendre polynomials enable the initial Gramian to be decomposed directly into a factored upper triangular matrix via QR factorization, establishing the necessary starting values.

A major theoretical milestone of the investigation is a comprehensive backward error analysis establishing that the proposed numerical scheme achieves full accuracy in floating-point arithmetic. By lifting the finite horizon problem into an abstract functional mapping, the authors proved that the computed Gramian corresponds exactly to a slightly perturbed continuous problem. Using the induced supremum norm across time intervals, they demonstrated that relative backward errors in the system dynamic matrix and input matrix are strictly bounded by unit round-off in standard double precision, approximately 1.1 times ten to the negative sixteenth power. Crucially, the mathematical bounds reveal that the backward error is entirely governed by the norm of the system dynamic matrix, leaving the scaling and magnitude of the input matrix completely irrelevant to backward stability. This theoretical guarantee ensures that the numerical approximation behaves like an exact solution to a nearby problem.

Supplementing their backward stability proofs, the researchers formulated explicit forward error bounds by evaluating relative condition numbers for the Gramian functional and its Cholesky factor. Because finite-precision calculations can amplify perturbations in poorly conditioned systems, the authors derived Fréchet derivatives with respect to matrix perturbations in spectral and induced supremum norms. They proved that the absolute condition number of the Gramian functional is bounded by twice the two-norm of the Gramian, yielding an estimated relative forward error proportional to double precision round-off multiplied by one plus the norm of the system matrix. For the Cholesky factor itself, when non-singular, the condition number bound incorporates a mild logarithmic dimension factor that remains strictly bounded even for matrices of astronomical dimensions, confirming robust theoretical forward stability. These forward estimates provide reliable predictive bounds on actual computational errors.

Beyond bounding numerical errors, the authors addressed whether the approximated Gramian mathematically preserves the physical reachability and controllability of the underlying dynamical system. In control theory, complete controllability requires the controllability matrix formed by sequential matrix powers to possess full rank, exactly matching the range space of the exact Gramian. By analyzing the polynomial span generated by the rational Padé basis across multiple doubling iterations, the researchers established that the approximation mathematically reproduces the true controllable subspace whenever the number of squaring steps satisfies a specific logarithmic threshold determined by matrix dimension and polynomial expansion order. This rank-preservation condition acts as an essential safeguard against artificial loss of numerical rank during successive matrix squarings in practical implementations, ensuring that the computed factor correctly captures the controllable dynamics of the target physical system.

The practical algorithm balances backward stability, rank preservation, and computational efficiency by carefully managing the matrix scaling parameter and Padé approximation order. To avoid numerical penalties associated with over-scaling, the procedure restricts expansion orders to odd integers three, five, seven, nine, and thirteen, drawing polynomial coefficients symbolically computed via specialized symbolic mathematics software. For smaller matrices, the algorithm identifies the lowest allowable order that simultaneously bounds matrix norm scaling and satisfies dimensional rank constraints. When high system dimensions or large matrix norms exceed these criteria, the method deploys order thirteen Padé approximations and calculates the minimal integer squarings necessary. Pre-processing steps ensure input matrices with excess control channels are reduced through initial QR decompositions, while default balancing and eigenvalue shifts are explicitly avoided to respect the underlying geometric structure of the Gramian.

The authors validated their methodology across extensive numerical experiments, commencing with a specialized unit test involving nilpotent matrices where exact analytical solutions exist in closed form. In this benchmark, the computed Cholesky factor achieved relative errors nearly half an order of magnitude smaller than the reconstructed Gramian, precisely matching theoretical forward error curves. Subsequent testing on ten-by-ten matrices drawn from the standard MatrixDepot collection demonstrated robust accuracy across fifty randomized input trials per matrix configuration. Across nearly all test cases, the algorithm remained insensitive to randomized input variations and stayed strictly within theoretical error boundaries, encountering mild numerical strain only on an involutory matrix whose unitary positive eigenvalues required an excessive sequence of twenty-five consecutive doublings. Even under such taxing arithmetic demands, the algorithm avoided catastrophic failure and maintained consistent computational stability.

To examine behavior on realistic dynamical systems across increasing dimensions, the researchers benchmarked the algorithm on state-space realizations of Laguerre networks with dimensions ranging from single states up to five thousand. In these Hurwitz systems, reference solutions can be computed independently using specialized Lyapunov algebraic relations. As dimensional size expanded from one thousand to five thousand, observed relative errors grew modestly in close agreement with theoretical estimates, driven solely by increasing matrix norms. Furthermore, numerical rank evaluations revealed that the method maintained full rank in well-conditioned parameter regimes, losing numerical rank only when severe ill-conditioning was intentionally introduced. The scaling-and-squaring Cholesky formulation consistently outperformed alternative eigendecomposition techniques, which exhibited severe relative errors or erratic rank collapse, demonstrating the clear superiority of the direct factorization framework for dense dynamical systems.

The complete numerical method, parameter selection routines, and experimental verification suites have been implemented in the open-source Julia package FiniteHorizonGramians.jl and released under the permissive MIT license. By operating on general dense matrices of moderate size, the package provides a drop-in computational foundation for square-root Kalman filters, state-space smoothers, and robust model order reduction algorithms that previously suffered from numerical instability. The doubling phase requires only one matrix multiplication and one QR decomposition per iteration, matching the computational simplicity and speed of standard matrix exponential solvers while delivering guaranteed backward stability. The authors noted that their backward error analysis and parameter optimization workflows can be directly adapted to other floating-point precisions, establishing a versatile standard for contemporary scientific computing, control system design, and numerical linear algebra.

Subject of Research: Simultaneous numerical computation of matrix exponentials and finite horizon Gramian Cholesky factors

Article Title: Computing the matrix exponential and the Cholesky factor of a related finite horizon Gramian

Article References: Stillfjord, T., & Tronarp, F. (2026). Computing the matrix exponential and the Cholesky factor of a related finite horizon Gramian. BIT Numerical Mathematics, 66(4), Article 65. https://doi.org/10.1007/s10543-026-01160-6

Image Credits: AI Generated

DOI: 10.1007/s10543-026-01160-6

Keywords: matrix exponential, controllability Gramian, Cholesky factor, scaling and squaring, backward error analysis, Petrov-Galerkin method, shifted Legendre polynomials, diagonal Pade approximants, numerical linear algebra, linear time-invariant systems, doubling recursion, FiniteHorizonGramians.jl, Julia programming language, controllable subspace, forward error bounds, Laguerre networks, QR decomposition, condition number estimation

Cite Scienmag News

Reid Dalton. (November 22, 2026). New Algorithm Computes Matrix Exponential and Gramian Cholesky Factor Simultaneously. Scienmag. https://scienmag.com/new-algorithm-computes-matrix-exponential-and-gramian-cholesky-factor-simultaneously/

Reid Dalton. "New Algorithm Computes Matrix Exponential and Gramian Cholesky Factor Simultaneously." Scienmag, 22 November 2026, https://scienmag.com/new-algorithm-computes-matrix-exponential-and-gramian-cholesky-factor-simultaneously/. Accessed 22 November 2026.

Reid Dalton. "New Algorithm Computes Matrix Exponential and Gramian Cholesky Factor Simultaneously." Scienmag. November 22, 2026. https://scienmag.com/new-algorithm-computes-matrix-exponential-and-gramian-cholesky-factor-simultaneously/

Tags: backward error analysisCholesky factorcondition number estimationcontrollability Gramiancontrollable subspacediagonal Pade approximantsdoubling recursionFiniteHorizonGramians.jlforward error boundsJulia programming languageLaguerre networkslinear time-invariant systemsmatrix exponentialnumerical linear algebraPetrov-Galerkin methodQR decompositionscaling and squaringshifted Legendre polynomials
Share26Tweet16
Previous Post

Meta-Predictive Language Model Ensembles Improve Named Entity Recognition in Aerospace Requirements

Next Post

Surjective Isometries on Alexiewicz Normed Spaces Classified via Measure Projections

Related Posts

Distinguished Silver Numbers Generate Aperiodic Half-Line Tilings Through Positional Integers
Mathematics

Distinguished Silver Numbers Generate Aperiodic Half-Line Tilings Through Positional Integers

November 22, 2026
Mathematical Analysis Unifies Stickiness Flavors in Geometric Fréchet Means
Mathematics

Mathematical Analysis Unifies Stickiness Flavors in Geometric Fréchet Means

November 22, 2026
Mathematicians Prove Three Bounded Solutions for Superlinear Double Phase Robin Problems
Mathematics

Mathematicians Prove Three Bounded Solutions for Superlinear Double Phase Robin Problems

November 22, 2026
Surjective Isometries on Alexiewicz Normed Spaces Classified via Measure Projections
Mathematics

Surjective Isometries on Alexiewicz Normed Spaces Classified via Measure Projections

November 22, 2026
Mathematicians Probe Invertibility and Sign Variation in Infinite Totally Positive Matrices
Mathematics

Mathematicians Probe Invertibility and Sign Variation in Infinite Totally Positive Matrices

November 22, 2026
Study Analyzes Milnor Fiber Cohomology for Hypersurfaces with One-Dimensional Singular Sets
Mathematics

Study Analyzes Milnor Fiber Cohomology for Hypersurfaces with One-Dimensional Singular Sets

November 22, 2026
Next Post
Surjective Isometries on Alexiewicz Normed Spaces Classified via Measure Projections

Surjective Isometries on Alexiewicz Normed Spaces Classified via Measure Projections

  • Mothers who receive childcare support from maternal grandparents show more optimized

    Mothers who receive childcare support from maternal grandparents show more parental warmth, finds NTU Singapore study

    27656 shares
    Share 11059 Tweet 6912
  • University of Seville Breaks 120-Year-Old Mystery, Revises a Key Einstein Concept

    1061 shares
    Share 424 Tweet 265
  • Bee body mass, pathogens and local climate influence heat tolerance

    682 shares
    Share 273 Tweet 171
  • Researchers record first-ever images and data of a shark experiencing a boat strike

    546 shares
    Share 218 Tweet 137
  • Groundbreaking Clinical Trial Reveals Lubiprostone Enhances Kidney Function

    531 shares
    Share 212 Tweet 133
Science

Embark on a thrilling journey of discovery with Scienmag.com—your ultimate source for cutting-edge breakthroughs. Immerse yourself in a world where curiosity knows no limits and tomorrow’s possibilities become today’s reality!

RECENT NEWS

  • New co-Editor-in-Chief outlines population health measurement priorities
  • Uncertainty in ATP7B Genetic Sequencing Can Trigger Misdiagnosis of Wilson Disease
  • Traders Abandoning Own Plans for Algorithmic Advice Suffer Reduced Decision Accuracy
  • Surveillance Gaps Mask Thirtyfold Malaria Burden Across Rainy Seasons in Somalia

Categories

  • Agriculture
  • Anthropology
  • Archaeology
  • Athmospheric
  • Biology
  • Biotechnology
  • Blog
  • Bussines
  • Cancer
  • Chemistry
  • Climate
  • Earth Science
  • Editorial Policy
  • Marine
  • Mathematics
  • Medicine
  • Pediatry
  • Policy
  • Psychology & Psychiatry
  • Science Education
  • Science News
  • Social Science
  • Space
  • Technology and Engineering

Subscribe to Blog via Email

Enter your email address to subscribe to this blog and receive notifications of new posts by email.

Join 5,151 other subscribers

© 2026 Scienmag - Science Magazine

Welcome Back!

Login to your account below

Forgotten Password?

Retrieve your password

Please enter your username or email address to reset your password.

Log In
No Result
View All Result
  • HOME
  • SCIENCE NEWS
  • CONTACT US

© 2026 Scienmag - Science Magazine

Discover more from Science

Subscribe now to keep reading and get access to the full archive.

Continue reading