Geometric Methods for Adjoint Systems

Brian K. Tran, Melvin LeokJ. Nonlinear sci 2022

TL;DR

This paper explores the geometric underpinnings of adjoint systems, which are crucial for optimization and sensitivity analysis in systems governed by differential equations. By leveraging symplectic and presymplectic geometry, the authors develop structure-preserving numerical methods that offer a more natural and efficient way to compute sensitivities, particularly for complex differential-algebraic equations.

What Problem Does This Paper Address?

Adjoint systems are indispensable tools for tasks like control, optimization, and design, especially when the parameter space is significantly larger than the system's output dimension. However, understanding and efficiently computing with these adjoint systems, particularly when they arise from differential-algebraic equations (DAEs), can be challenging. The authors aim to provide a unified geometric framework for adjoint systems, revealing their inherent structure and developing numerical methods that respect this structure, leading to more accurate and efficient computations.

Methods

The core of this paper lies in its application of symplectic and Presymplectic Geometry.

  • Symplectic Geometry: Used to analyze adjoint systems derived from ordinary differential equations (ODEs). The authors show how the "quadratic conservation laws," vital for sensitivity analysis, naturally emerge from the symplecticity of these adjoint systems.
  • Presymplectic Geometry: Employed to handle the more complex case of adjoint systems arising from DAEs. This involves using the "presymplectic constraint algorithm" to manage the algebraic constraints inherent in DAEs and relating the index of the DAE to the structure of its adjoint system.
  • Structure-Preserving Numerical Methods: The paper introduces Galerkin Hamiltonian Variational Integrators (both symplectic for ODEs and a presymplectic analogue for DAEs). These methods are designed to discretize adjoint systems while preserving their geometric structure, ensuring that discrete analogues of the quadratic conservation laws are maintained.
  • Naturality: A key concept explored is "naturality," which describes how operations like reduction (simplifying DAEs), forming the adjoint system, and discretization commute. The authors demonstrate this commutativity for specific choices of these operations, leading to a more unified and robust theoretical framework.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Geometric Origin of Conservation Laws: The paper reveals that the crucial "adjoint variational quadratic conservation laws" are not arbitrary but arise directly from the symplectic (for ODEs) or presymplectic (for DAEs) nature of the adjoint system's flow. This provides a deeper understanding of why these laws hold.
  • DAE Index and Adjoint Structure: The index of a DAE is shown to be directly related to the complexity and structure of its adjoint system, particularly through the lens of the presymplectic constraint algorithm. This connection is vital for understanding the solvability of adjoint DAEs.
  • Commuting Operations (Naturality): The demonstration that reduction, adjoint formation, and discretization can commute (under specific conditions) is a significant theoretical result. This "naturality" is crucial for developing reliable numerical methods and error analyses, as it implies that the order in which these operations are performed doesn't affect the final result, simplifying the overall process.
  • Structure-Preserving Integrators for DAEs: The development of presymplectic variational integrators for adjoint DAE systems is a key practical contribution. These integrators are designed to respect the geometric structure, ensuring that discrete conservation laws are preserved, which is essential for accurate sensitivity computations.

Key Results

  • The paper provides a geometric characterization of adjoint systems for both ODEs and DAEs, unifying their treatment under symplectic and presymplectic geometry.
  • It establishes that adjoint variational quadratic conservation laws arise from the (pre)symplecticity of adjoint systems.
  • A framework for constructing structure-preserving numerical integrators (Galerkin Hamiltonian variational integrators) for adjoint ODE and DAE systems is presented. These integrators admit discrete analogues of the quadratic conservation laws.
  • The paper proves a "naturality" result, showing that for index 1 DAEs, the processes of reduction, adjoint formation, and discretization commute under specific choices.
  • A numerical example using a planar pendulum DAE demonstrates that the proposed presymplectic integrator preserves phase space area, unlike a standard backward Euler method, highlighting the benefits of structure preservation.

Limitations and Future Work

  • General Index DAEs: While the paper addresses index 1 DAEs extensively, the treatment of higher index DAEs relies on reduction techniques, which can be numerically challenging. The authors suggest further work on handling general index DAEs more directly.
  • Nonholonomic Constraints: The paper focuses on systems with constraints that are purely algebraic or differential. Future work could extend the framework to nonholonomic constraints, which involve relationships between velocities and configurations, potentially using Dirac structures.
  • Evolutionary PDEs: The authors propose extending the geometric framework to infinite-dimensional systems, specifically evolutionary partial differential equations (PDEs). This would involve exploring the geometry of infinite-dimensional symplectic structures and multisymplectic structures for PDEs.
  • Numerical Stability and Efficiency: While structure-preserving methods are introduced, further investigation into their practical performance, stability, and efficiency for large-scale problems is a natural next step.

Summary updated

Abstract

Adjoint systems are widely used to inform control, optimization, and design in systems described by ordinary differential equations or differential-algebraic equations. In this paper, we explore the geometric properties and develop methods for such adjoint systems. In particular, we utilize symplectic and presymplectic geometry to investigate the properties of adjoint systems associated with ordinary differential equations and differential-algebraic equations, respectively. We show that the adjoint variational quadratic conservation laws, which are key to adjoint sensitivity analysis, arise from (pre)symplecticity of such adjoint systems. We discuss various additional geometric properties of adjoint systems, such as symmetries and variational characterizations. For adjoint systems associated with a differential-algebraic equation, we relate the index of the differential-algebraic equation to the presymplectic constraint algorithm of Gotay et al. [18]. As an application of this geometric framework, we discuss how the adjoint variational quadratic conservation laws can be used to compute sensitivities of terminal or running cost functions. Furthermore, we develop structure-preserving numerical methods for such systems using Galerkin Hamiltonian variational integrators (Leok and Zhang [23]) which admit discrete analogues of these quadratic conservation laws. We additionally show that such methods are natural, in the sense that reduction, forming the adjoint system, and discretization all commute, for suitable choices of these processes. We utilize this naturality to derive a variational error analysis result for the presymplectic variational integrator that we use to discretize the adjoint DAE system. Finally, we discuss the application of adjoint systems in the context of optimal control problems, where we prove a similar naturality result.
BibTeX
@article{tran2022geometric,
  author = {Brian K. Tran and Melvin Leok},
  title = {{Geometric Methods for Adjoint Systems}},
  journal = {J. Nonlinear sci},
  year = {2022},
}
Figure