A stabilized march approach to adjoint-based sensitivity analysis of chaotic flows
Pranshul Thakur, Siva Nadarajah · 2025
Read A stabilized march approach to adjoint-based sensitivity analysis of chaotic flows by Pranshul Thakur, Siva Nadarajah (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-00838.txt.
Introduction
Pranshul Thakur and Siva Nadarajah compute how sensitive chaotic flows are to design changes, cheaply and correctly, in a 2025 paper. Adjoint methods give sensitivities for many parameters at low cost, which is why engineers love them. Chaos breaks the standard version: tiny changes explode over time.
Existing fixes follow a shadowing trajectory, the nearby path that stays close forever. Finding it means solving a huge space-time least-squares problem. The authors instead march a stabilized adjoint with the right boundary conditions. That reduces the solve to a single backward sweep. Its cost scales with the size of the unstable part of the flow. They prove the computed sensitivities converge to the true ones over long times. The errors stay on the order of the scheme's own accuracy. Tests on the Lorenz system and a chaotic partial differential equation confirmed the claims.
Catalog record
| CREATOR | Pranshul Thakur, Siva Nadarajah |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 59 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2505-00838
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
A stabilized march approach to adjoint-based sensitivity analysis of chaotic flows Thanks: Pranshul Thakur, Siva Nadarajah Abstract Adjoint-based sensitivity analysis is of interest in computational science due to its ability to compute sensitivities at a lower cost with respect to several design parameters. However, conventional sensitivity analysis methods fail in the presence of chaotic flows. Popular approaches to chaotic sensitivity analysis of flows involve the use of the shadowing trajectory. The state-of-the-art approach computes the shadowing trajectory by solving a least squares minimization problem, resulting in a space-time linear system of equations. The current paper computes the adjoint shadowing trajectory using the stabilized march, by specifying the adjoint boundary conditions instead of solving a minimization problem. This approach results in a space-time linear system that can be solved through a single backward substitution of order 𝒪 ( n u 2 ) \mathcal{O}(n_{u}^{2}) with n u n_{u} being the dimension of the unstable subspace. It is proven to compute sensitivities that converge to the true sensitivity for large integration times and that the error in the sensitivity due to the discretization is of the order of the local truncation error of the scheme. The approach is numerically verified on the Lorentz 63 and Kuramoto-Sivasinsky equations. keywords chaos, sensitivity analysis, linear response, adjoint, least squares shadowing, stabilized march Funding. This work was funded by the Fonds de recherche du Québec – Nature et technologies (FRQNT) Doctoral Award, McGill Engineering Doctoral Award (MEDA) and the Natural Sciences and Engineering Research Council of Canada (NSERC) Discovery Grant RGPIN-2019-04791. † † runningheads: Sensitivity analysis of chaotic flows using the adjoint shadowing stabilized march / P. Thakur and S. Nadarajah AMS 34A34, 37A99, 37D20, 37D45, 37N30, 46N40, 65P99, 76F20 1 Introduction Sensitivity analysis of quantities of interest is useful for optimization [ 41 ] , computing improved functional estimates [ 5 , 21 , 44 ] , mesh adaptation [ 16 , 57 , 53 , 20 , 35 , 17 ] etc. The adjoint method of computing sensitivities is efficient when there are a lot fewer functionals than independent parameters, as is common in aerodynamic applications [ 26 ] . However, both the conventional forward and adjoint methods yield unbounded sensitivities when applied to chaotic systems [ 33 ] . The reason is due to the positive Lyapunov exponents in chaotic systems, which cause most nearby trajectories to diverge exponentially. The work of Ruelle [ 48 ] established the differentiability of chaotic hyperbolic systems. The ensemble adjoint approach of Lea et al. [ 19 , 33 ] computes chaotic sensitivities by computing conventional adjoint solutions over short trajectories and taking the ensemble average of the resulting short-window sensitivities. Some approaches replace the initial condition with periodic boundary conditions on the tangent/adjoint equations to yield bounded sensitivities, such as the unstable periodic orbits [ 31 ] and periodic shadowing [ 32 ] . The unstable periodic orbit approach has also been extended to discontinuous periodic orbits to yield regularized linear responses to parameter perturbations [ 25 ] . It is known that hyperbolic systems with a compact attractor have trajectories that shadow the reference trajectory when the function generating the dynamics is perturbed [ 2 , 12 , 24 , 15 ] . Using this idea, Wang et al. [ 55 , 56 ] developed a method of linearizing the governing equation using the shadowing trajectory. In the original least squares shadowing (LSS) approach [ 56 ] , a least squares minimization problem is solved to approximate the shadowing trajectory. To reduce the cost of the LSS approach, there have been approaches that reduce the size of the linear system through multiple shooting [ 10 ] , reduced-order modeling [ 28 , 50 ] or formulating the shadowing algorithm in a Fourier space with truncation [ 4 , 27 ] . The non-intrusive [ 40 ] and adjoint non-intrusive methods [ 8 , 38 ] have been developed which require only the unstable subspace of the dynamical system for forward sensitivity and an additional one-dimensional neutral subspace for adjoint-based sensitivity [ 38 ] methods. Their cost scales with the number of unstable Lyapunov covariant vectors in the system and are efficient for problems with lower dimensional unstable subspaces, but are expensive when the number of positive Lyapunov exponents is large [ 38 ] . We also note that the shadowing-based approaches might compute shadowing trajectories that are non-physical [ 13 ] and approaches based on the linear response [ 48 ] have been investigated [ 14 , 39 , 51 ] to address the issue. Since the shadowing-based approaches are known to yield accurate sensitivities for several test cases involving turbulent fluid flows [ 11 , 38 , 40 ] , the current paper does not aim to address the issue of non-physical shadow trajectories and assumes that the shadow trajectories are representative of the perturbed dynamical system. Moreover, few recent approaches based on the linear response require the computation of the shadow trajectory [ 37 , 39 ] , and the approach developed in this work can be used to that end. The current paper provides a method to compute the adjoint shadowing trajectory through the stabilized march approach [ 3 ] . The method uses QR-decompositions to compute the unstable subspace, similar to the non-intrusive adjoint LSS approach (NILSAS) [ 38 ] . However, instead of solving a least squares minimization problem in the unstable subspace, the current approach chooses the adjoint boundary conditions to make the resulting space-time linear system triangular. We note that an approach similar to the stabilized march has been used for chaotic maps [ 37 ] by modifying adjoint boundary conditions. The current work can be seen as the extension of stabilized march to chaotic continuous-in-time flows. Additionally, while the classical stabilized march requires the exact number of unstable modes for stability [ 3 ] , the method developed in this work is stable even when the number of unstable modes is overestimated at the beginning of the algorithm. The paper is organized as follows: Section 2 discusses the preliminaries and notations used in this work. Section 3 briefly provides the results for the adjoint of LSS, proven in [ 54 ] . Section 4 provides the adjoint boundary conditions used in this work and proves that the sensitivity from the method converges to the true sensitivity at \mathcal{O}\left(\frac{1}{\sqrt{T}}\right) as the integration time T\to\infty . Section 5 shows that the boundary conditions presented in section 4 convert the linear space-time system to a sequence of triangular solves. Section 5.1 provides the stabilized march algorithm when the dimension of the unstable subspace has been overestimated. Section 6 investigates errors due to discretization and proves that the discretization error in the computed sensitivity from stabilized march is of the order of the truncation error of the scheme. Finally, section 7 applies the approach to the test cases involving chaotic flows. 2 Preliminaries Consider the following dynamical system, governing the evolution of a state u with time t : \frac{du}{dt}=f(u(t,s),s),\;\;u(0)=u_{0} (1) where u(t,s)\in\Lambda \forall t , \Lambda\subset X is a compact attractor in a n -dimensional Hilbert space X , s\in\mathbb{R} is the design or control parameter (for example, the shape