Symplectic H2 Model Reduction for High-Dimensional Linear Quantum Systems

This paper introduces Quantum IRKA (Q-IRKA), a symplectic Petrov-Galerkin framework that enables scalable $\mathcal{H}_2$ model reduction for high-dimensional linear quantum systems while strictly preserving physical realizability and canonical commutation relations through structure-preserving projection.

The Gist

Imagine you have a massive, incredibly complex orchestra of quantum instruments (like a giant chain of vibrating springs). This orchestra plays a specific song (the system's behavior). While the full orchestra sounds perfect, it's too big to carry around, too expensive to simulate on a computer, and too slow to use for real-time control. You want a small, portable "mini-orchestra" (a reduced model) that plays the same song just as well, but with far fewer instruments.

The problem? In the quantum world, you can't just pick and choose instruments randomly. The laws of physics (specifically, the "rules of the game" known as Physical Realizability) demand that the instruments stay perfectly synchronized in a very specific, rigid way. If you break this synchronization, your mini-orchestra isn't a real quantum system anymore; it's just a mathematical fantasy that violates the laws of nature.

This paper introduces a new method called Q-IRKA (Quantum Iterative Rational Krylov Algorithm) to solve this problem. Here is how it works, using simple analogies:

1. The "Symplectic" Rulebook

In classical engineering, you can shrink a model by simply projecting it onto a smaller space, like taking a photo of a 3D object to get a 2D image. But in quantum systems, the "shape" of the system is defined by a special geometric rule called symplecticity. Think of this like a dance where every partner must hold hands in a specific, mirrored pattern. If you shrink the dance floor but break the hand-holding pattern, the dance falls apart.

The authors' key insight is to build a "dance floor" (a mathematical space) that forces the partners to hold hands correctly by design. They use a special type of projection (a Symplectic Petrov-Galerkin framework) that acts like a mold. No matter how you pour the liquid (the complex system) into this mold, the resulting shape is guaranteed to keep the correct hand-holding pattern. You don't have to check if the rules are followed; the mold ensures they are.

2. The "Smart Search" (Q-IRKA)

How do you find the best small set of instruments to keep?

• The Old Way: You might try to guess the best instruments, check if they work, and then guess again. This is slow and computationally heavy.
• The Q-IRKA Way: The algorithm acts like a smart, iterative tuner.
  1. It starts with a guess for the "notes" (interpolation points) the mini-orchestra should play.
  2. It builds a temporary mini-orchestra using those notes.
  3. It listens to the mini-orchestra, finds the "notes" (poles) it naturally wants to play, and then mirrors them to update its guess.
  4. It repeats this process, refining the mini-orchestra over and over.

Crucially, at every single step of this tuning process, the algorithm uses the "symplectic mold" mentioned above. This means the mini-orchestra never loses its physical validity, even while it is being tweaked. It preserves the "laws of physics" to the limit of the computer's precision (machine precision).

3. The Experiments: Testing the Mini-Orchestra

The authors tested this method on two types of "orchestras":

• The Oscillator Chain: Imagine a row of 100 to 200 pendulums connected together. Some were all identical (homogeneous), while others had different weights and friction (heterogeneous).
• The Kitaev Chain: A more complex setup inspired by real quantum experiments, where energy flows in a specific direction along a chain.

What they found:

• It Works: The method successfully created tiny models (reducing a system of 200 variables down to 20) that played the song almost perfectly.
• Physics is Safe: The "hand-holding" rules (physical realizability) were never broken. The math stayed perfect down to the smallest decimal point.
• Complexity Matters: The quality of the mini-orchestra depended heavily on how the "friction" (dissipation) was arranged. If the system was uniform, the mini-model was very accurate. If the system was messy and uneven (heterogeneous), it was harder to shrink, but the method still worked well.
• Speed: The method was fast enough to handle large systems, making it practical for real-world engineering tasks like designing controllers or filters.

Summary

In short, this paper presents a new "mold" and a "smart tuner" that allow engineers to shrink massive, complex quantum systems into tiny, manageable versions without ever breaking the laws of physics. It ensures that the simplified model is not just a mathematical approximation, but a physically valid quantum system that can be used for design and control.

Technical Summary

Technical Summary: Symplectic H2 Model Reduction for High-Dimensional Linear Quantum Systems

Problem Statement
The paper addresses the $H_2$ model reduction problem for high-dimensional linear quantum systems. A primary challenge in this domain is the constraint of Physical Realizability (PR). Unlike classical systems, linear quantum systems must satisfy specific algebraic identities coupling the system matrices $(A, B, C, D)$ to preserve canonical commutation relations and the quantum input-output structure. Standard projection-based reduction methods often fail to preserve these nonlinear PR constraints, leading to reduced-order models that are physically invalid. Directly enforcing these constraints as optimization conditions results in difficult nonlinear algebraic problems that are not scalable for large systems.

Methodology
The authors propose a Symplectic Petrov-Galerkin framework that ensures PR is satisfied by construction rather than by post-hoc enforcement.

1. Symplectic Projection: The method relies on constructing a reduced-order model via projection onto a symplectic subspace. If the trial basis $V \in \mathbb{R}^{2n \times 2r}$ satisfies the symplectic condition $V^\top J_n V = J_r$ (where $J_k$ is the canonical symplectic matrix), and the test basis is chosen as $U = J_r^{-1} V^\top J_n$, the resulting reduced matrices $(A_r, B_r, C_r, D_r)$ automatically satisfy the PR identities.
2. Quantum IRKA (Q-IRKA): The authors develop a symplectic variant of the Iterative Rational Krylov Algorithm (IRKA). The algorithm operates as a fixed-point iteration:
   • Shift Updates: Interpolation points (shifts) are updated recursively based on the poles of the current reduced-order model. Specifically, shifts are set to the negative of the selected poles (pole-mirroring), respecting the spectral symmetry of the real reduced system.
   • Tangential Interpolation: At each iteration, a candidate pool is generated by solving shifted linear systems $(A - \sigma_i I)^{-1} B t_{i,\ell}$ for prescribed shifts and tangential directions.
   • Symplectic Basis Extraction: From the candidate pool, a symplectic basis is extracted using a Gram-Schmidt-type procedure that pairs vectors with their symplectic conjugates ($J_n^\top v$). A subsequent normalization step ensures the basis strictly satisfies $V^\top J_n V = J_r$.
   • Structure Preservation: The reduced matrices are obtained exclusively through the projection $A_r = UAV$, $B_r = UB$, $C_r = CV$, and $D_r = D$. No independent modification of reduced matrices occurs, guaranteeing PR is preserved to machine precision throughout the iteration.

Key Contributions

• Q-IRKA Algorithm: The introduction of a scalable, structure-preserving recursive algorithm for $H_2$ optimal model reduction of linear quantum systems. It combines the efficiency of IRKA (using shifted linear solves) with strict symplectic constraints.
• Construction-Based PR: The framework eliminates the need for solving constrained nonlinear optimization problems. Physical realizability is an inherent property of the projection geometry, ensuring the reduced model remains a valid quantum system.
• Systematic Numerical Analysis: The paper provides a comprehensive study of how reduction quality depends on physical structural features, specifically:
  • Dissipation Geometry: Homogeneous vs. heterogeneous damping.
  • Channel Placement: Low-channel configurations where multiple oscillators share dissipative pathways.
  • System Heterogeneity: Variations in on-site frequencies and damping strengths.

Results
Numerical experiments were conducted on two benchmark families:

1. Low-Channel Port-Hamiltonian Systems: Oscillator chains with aggregated external channels.
2. Bosonic Kitaev-Chain-Inspired Systems: Structured transport-dominated systems with nearest-neighbor coupling.

Key Findings:

• Accuracy and Stability: Q-IRKA produces reduced-order models with accurate $H_2$ approximation. The symplecticity and PR residuals ($\Delta_{symp}$ and $\Delta_{PR}$) remain at machine precision ($\approx 10^{-14}$ to $10^{-16}$) throughout the iterations.
• Dependence on Heterogeneity: Reduction quality is heavily influenced by the system's dissipation geometry. Homogeneous systems with fast Hankel singular value decay yield low-order approximations with high accuracy ($H_2$ error $\approx 10^{-5}$). Heterogeneous systems exhibit slower spectral decay, resulting in higher effective dynamical ranks and larger errors ($H_2$ error $\approx 10^{-3}$), even with the same reduced order.
• Computational Cost: The computational cost scales primarily with the system dimension $n$ and the reduced order $r$. The dependence on the number of channels $m$ is mild. Heterogeneous cases generally require more iterations for convergence and incur higher computational costs due to slower spectral decay.
• Comparison: In a small-scale comparison, Q-IRKA achieved a lower $H_2$ error (1.2130) compared to quasi-balanced truncation (1.86) and a non-convex optimization method (1.25).

Significance
The paper claims that scalable $H_2$ model reduction for large-scale linear quantum systems can be achieved without sacrificing the underlying physical structure. By embedding the PR constraints directly into the projection geometry, the method avoids the computational intractability of constrained optimization. The results demonstrate that while reduction performance is sensitive to physical parameters like dissipation geometry and heterogeneity, the proposed framework robustly maintains physical realizability and provides accurate reduced-order models suitable for design loops, controller synthesis, and scalability studies in quantum engineering. The framework is noted to extend naturally to non-square linear quantum systems, though this is reserved for future work.