Symplectic model order reduction of port-Hamiltonian systems

Basic Information

• Paper ID: 2203.07751
• Title: Symplectic model order reduction of port-Hamiltonian systems
• Authors: Silke Glas, Mir Mamunuzzaman, Hongliang Mu, Hans Zwart
• Classification: math.OC (Optimization and Control), math.CA (Classical Analysis)
• Publication Date: March 2022 (arXiv preprint)
• Paper Link: https://arxiv.org/abs/2203.07751

Abstract

This paper proposes a novel structure-preserving model order reduction (MOR) method for linear time-invariant port-Hamiltonian (pH) systems. The research objective is to construct a reduced-order pH system that remains interpretable within the physical domain of the original full-order model. Specifically, if the initial high-dimensional pH system is an electrical circuit, the reduced-order model should remain interpretable as an electronic circuit. For mass-spring-damper (MSD) systems, existing MOR methods can guarantee preservation of this specific structure. This paper demonstrates that the new structure-preserving MOR method based on symplectic MOR recovers the known second-order Arnoldi method for MSD systems. However, for circuit pH models and similar models with comparable block structures, the method yields novel model reduction techniques.

Research Background and Motivation

Problem Description

Port-Hamiltonian systems are generalizations of classical Hamiltonian frameworks and are extremely useful in engineering problems because they combine classical Hamiltonian methods with powerful frameworks for modeling open physical systems. However, in control design, controllers typically have the same order as the original system, which motivates the need for reduced-order models that capture the fundamental behavior of the original system.

Limitations of Existing Methods

1. Structure Loss Problem: Direct application of traditional MOR methods loses the pH structure
2. Missing Physical Interpretability: Existing methods cannot guarantee physical domain interpretability of reduced-order models
3. Block Structure Destruction: For pH systems with special block structures, existing methods cannot preserve this structure

Research Motivation

When reformulating physical models into pH form, matrices J, R, and H typically possess special block structures. The goal of this paper is to maintain these block structures during the reduction process, ensuring that the reduced-order model remains interpretable as a physical system of the same type as the full-order model.

Core Contributions

1. Introduction of Symplectic-pH Method: Proposes a novel pH system MOR method based on symplectic MOR applicable to block structure form (2), enabling the reduced-order model to maintain the same form of pH system
2. Theoretical Connection: Proves that in the special case of MSD systems, the proposed method recovers the second-order Arnoldi method (SO-Arnoldi)
3. Numerical Verification: Demonstrates numerical results on electronic circuits (LRCR circuits) and compares with other known methods in the literature
4. Stability Advantages: Discovers that the SO-Arnoldi method produces unstable reduced-order models in certain cases, while the proposed method maintains stability

Methodology Details

Task Definition

Given a linear time-invariant pH system:

ẋ(t) = (J - R)Hx(t) + Bu(t)
y(t) = B*Hx(t)

where x(t) ∈ C²ⁿ is the state, u(t) ∈ Cᵖ is the input, and y(t) ∈ Cᵖ is the output. The objective is to find a system of the same form but with smaller state space dimension such that the two transfer functions are equal at predetermined frequency points.

Core Theoretical Framework

Symplectic Geometry Foundations

Definition 1 (Symplectic Form): Let V be an even-dimensional vector space over C, and symplectic form Ω: V × V → C satisfies:

1. Antisymmetry: Ω(v₁, v₂) = -Ω(v₂, v₁)
2. Non-degeneracy: For all v₁ ∈ V, Ω(v₁, v₂) = 0 if and only if v₂ = 0

Definition 2 (Symplectic Mapping): Matrix Q is called (J₂ₙ, J₂ₖ)-symplectic if:

Q*J₂ₙQ = J₂ₖ

Symplectic MOR for Dissipation-Free Case

Theorem 4: For dissipation-free pH systems (R = 0), if Q ∈ C²ⁿˣ²ᵏ is a (J₂ₙ⁻¹, J₂ₖ⁻¹)-symplectic matrix satisfying:

• ran(B) ⊆ ran(Q)
• (sₘI₂ₙ - J₂ₙH)⁻¹Bμₘ ∈ ran(Q)

then the reduced-order pH system equals the original system at given frequencies.

Symplectic MOR for Dissipative Case

Theorem 5: For dissipative pH systems, construct the extended matrix:

QE := [Q₁  0  ]
      [0   Q₂]

where Q₂ is (J₂ₙ, J₂ₖ)-symplectic and Q₁* is the left inverse of Q₂. Under appropriate conditions, the reduced-order system satisfies interpolation conditions.

Block Structure-Preserving Algorithm

Algorithm 1 (Symplectic-pH MOR):

1. Construct V₁ based on interpolation data
2. Compute V₂ := HV₁
3. Partition V₁ and V₂ according to block structure
4. Construct Ψ₁, Ψ₂, Φ₁, Φ₂ via formula (27)
5. Build Q₁ and Q₂
6. Apply Petrov-Galerkin projection

Construction formulas:

Ψ₂ = orth(V₂₂)
Ψ₁ = V₁₂(Ψ₂*V₁₂)⁻¹
Φ₂ = V₂₁(Ψ₂*Ĵₙ*V₂₁)⁻¹Ĵₖ*
Φ₁ = V₁₁(Φ₂*V₁₁)⁻¹

Technical Innovations

1. Symplectic Structure Preservation: Maintains the fundamental geometric structure of Hamiltonian systems through symplectic mappings
2. Block Diagonal Structure Maintenance: Special construction of Q₁ and Q₂ ensures preservation of physically meaningful block structures
3. Interpolation Conditions: Guarantees exact matching of the reduced-order system to the original system at predetermined frequency points
4. Unified Framework: Incorporates the SO-Arnoldi method for MSD systems as a special case

Experimental Setup

Numerical Example: LRCR Circuit

Consider the LRCR circuit shown in Figure 1 with n=50 units and parameter settings:

• Capacitance: Cᵢ = 1 pF
• Inductance: Lᵢ = 1 nH
• Resistance: Rᵢ,₁ = 1 mΩ, Rᵢ,₂ = 1 kΩ

Evaluation Metrics

• H₂ Error:

  H₂error = √(1/2π ∫ Trace((G(iω) - Gred(iω))*(G(iω) - Gred(iω)))dω)
  

• H∞ Error:

  H∞error = max‖G(iω) - Gred(iω)‖₂
  

Comparison Methods

1. pH-Arnoldi: Structure-preserving method based on Theorem 2
2. SO-Arnoldi: Second-order Arnoldi method
3. Symplectic-pH: The proposed method

Interpolation Point Selection

Uniformly selected interpolation points on logarithmic scale in the interval 10⁻², 10², for example 6 interpolation points: 10⁻²i, -10⁻²i, 10⁰i, -10⁰i, 10²i, -10²i

Experimental Results

Main Results

From Figure 3, the following observations can be made:

1. Symplectic-pH vs pH-Arnoldi: Symplectic-pH provides lower H₂ and H∞ errors in most cases
2. Symplectic-pH vs SO-Arnoldi:
   • Similar performance when reduced dimension < 24
   • SO-Arnoldi performance degrades sharply at reduced dimension = 32
   • SO-Arnoldi produces unstable systems at reduced dimensions 28 and 36

Stability Analysis

Figure 4 shows unstable poles produced by the SO-Arnoldi method at r=28 and r=36, while the symplectic-pH method maintains stability across all tested dimensions.

Theoretical Verification: MSD System Equivalence

Theorem 7 proves that for MSD systems, the symplectic-pH method is equivalent to the SO-Arnoldi method through unitary transformation: There exists a unitary matrix Θ such that zred(t) = Θẑred(t)

Related Work

Development of pH System MOR

1. Early Work: Passive system methods based on positive real transfer functions
2. Moment Matching: Work by Polyuga and van der Schaft
3. Dirac Structure Approximation: Alternative approaches to structure preservation
4. Spectral Decomposition Methods: Methods based on Popov functions
5. Optimization Methods: Methods based on parameter optimization and Riemannian optimization

Development of Symplectic MOR

1. Autonomous Hamiltonian Systems: Pioneering work by Peng and Mohseni
2. Nonlinear Methods: Manifold-based symplectic reduction
3. Data-Driven Methods: Operator inference and system identification

Conclusions and Discussion

Main Conclusions

1. Successfully developed a MOR method that preserves both pH structure and block structure
2. Theoretically unified different MOR methods for MSD systems
3. Numerical experiments verified the effectiveness and stability advantages of the method

Limitations

1. Interpolation Method: Currently only matches transfer function values without considering derivative matching
2. Block Structure Constraints: Method is specifically designed for particular block structures
3. Computational Complexity: Computational cost of constructing symplectic matrices requires further analysis

Future Directions

1. Derivative Matching: Extension to match derivatives of transfer functions for improved fitting quality
2. Adaptive Selection: Development of adaptive selection strategies for interpolation points
3. Nonlinear Extension: Extension to nonlinear pH systems
4. Large-Scale Applications: Algorithm optimization for handling larger-scale systems

In-Depth Evaluation

Strengths

1. Theoretical Rigor: Solid mathematical foundation based on symplectic geometry
2. Structure Preservation: Simultaneously preserves both pH structure and physically meaningful block structures
3. Unification: Unifies existing methods as special cases within a single framework
4. Practicality: Method has clear algorithmic implementation and numerical verification

Weaknesses

1. Limited Applicability: Method is specifically designed for particular block structures with limited generality
2. Computational Efficiency: Computational complexity analysis of symplectic matrix construction is insufficient
3. Large-Scale Verification: Lacks numerical verification on larger-scale systems

Impact

1. Theoretical Contribution: Provides new theoretical perspective for pH system MOR
2. Practical Value: Has direct value in engineering applications such as circuit analysis
3. Methodological Significance: Demonstrates the importance of geometric methods in MOR

Applicable Scenarios

1. Circuit Analysis: Model order reduction of large-scale electronic circuits
2. Mechanical Systems: Control design for mass-spring-damper systems
3. Multi-Physics Systems: Complex physical systems with energy conservation properties

References

The paper contains 31 references covering important works in multiple fields including MOR theory, pH system modeling, symplectic geometry, and control theory, providing a solid theoretical foundation for the research.