This paper introduces a novel framework for diffusion models by moving from real-valued state spaces to the complex plane. By leveraging the geometric properties of complex numbers, the authors propose Complex Schrödinger Bridges (CSBs), a new class of generative models designed to handle systems governed by rotation, oscillation, and phase—dynamics that are notoriously difficult for standard real-valued diffusion models to capture.
TL;DR
The authors extend Schrödinger Bridges into complex-valued vector spaces, introducing Conformal Optimal Transport (COT) to model dynamics as logarithmic spirals rather than straight lines. This approach significantly improves performance in modeling physical systems like quantum circuits and complex-valued data like MRI scans.
What Problem Does This Paper Address?
Diffusion bridges, such as Schrödinger Bridges (SBs), have become foundational in generative modeling for their ability to transform one data distribution into another. However, standard formulations are typically restricted to real-valued state spaces and linear dynamics.
In the real world, many systems—especially in physics and medical imaging—are governed by scaling and rotation laws (Hamiltonian structures). Modeling these using real-valued SDEs requires navigating a massive optimization space because the model lacks an inherent "understanding" of rotational symmetry or phase. The authors aim to embed these physically grounded geometric biases directly into the diffusion framework to make learning more efficient and accurate for complex-valued dynamics.
Methods
The core innovation lies in the "complexification" of the Schrödinger Bridge Problem (SBP):
- Wirtinger Calculus: The authors utilize Wirtinger operators, which provide a succinct way to perform differential calculus on complex variables without requiring the functions to be strictly holomorphic (a condition often too restrictive for neural network training).
- Conformal Optimal Transport (COT): Standard optimal transport seeks the "straightest" path between distributions. COT introduces a geometric lifting where the prior is a logarithmic spiral. This naturally incorporates rotational and dissipative dynamics.
- Complex Schrödinger Bridges (CSBs): Building on COT, the authors derive a new class of dynamical diffusion bridges. They use a complex version of the Hopf-Cole transform to linearize the underlying partial differential equations (PDEs), leading to forward and backward stochastic differential equations (SDEs) that can be solved via neural networks.
- Image-to-Image CSB ($I^2$CSB): To ensure scalability for high-dimensional tasks like image translation, the authors derive a tractable version of the model with an analytic spiral posterior, bypassing the need for costly iterative updates in some settings.
Datasets
The framework was validated across three distinct domains:
- 2D Control Benchmarks: Standard tasks like Moons-to-Circles, S-Tunnel, GMM-8, and V-Neck, modified to include momentum and obstacles.
- Image Translation: Unpaired MNIST generation, and inpainting tasks on CIFAR-10 and ImageNet256.
- Scientific Data:
- Synthetic Brain MRI: A complex-valued dataset generated using the
KomaMRI.jlsimulator, where preserving both magnitude and phase is critical.
- Synthetic Brain MRI: A complex-valued dataset generated using the
- Quantum Electrodynamics (cQED): Simulations of harmonic oscillators and 3D qubit logic gates (Hadamard and Pauli gates) using the
QuantumToolbox.jlpackage.
Key Insights and Ideas
- Phase as a Degree of Freedom: By treating states as complex numbers, the model gains the freedom to explore phase trajectories. This allows it to represent oscillatory motion (like a pendulum or a quantum wave function) much more naturally than a real-valued model could.
- The Spiral Prior: In many physical systems, the "natural" path isn't a straight line but a spiral. By baking this into the optimal transport cost via COT, the model starts with a much better "guess" (prior) for the data's underlying dynamics.
- Conformal Invariance: The use of conformal mappings ensures that angular relationships are preserved during the transformation, which is a powerful inductive bias for systems following Hamiltonian mechanics.
Key Results
- Superior Control: In 2D control tasks, CSB achieved lower mean endpoint 2-Wasserstein ($W_2$) distances and lower total trajectory costs compared to baselines like DeepGSB and GSBM.
- High-Fidelity MRI: On the synthetic brain MRI task, the CSB method significantly outperformed others. It achieved a magnitude FID of 12.01 (compared to 15.26 for $I^2$SB) and a structural similarity (SSIM) of 0.71 (compared to 0.66 for $I^2$SB), showing much better reconstruction of both magnitude and phase.
- Quantum Logic Gates: In cQED simulations, the model successfully mimicked the behavior of quantum logic gates. It outperformed benchmarks in seven out of nine settings when tested at higher, unseen energy levels, demonstrating robust generalization to physical priors.
Limitations and Future Work
- Geometric Rigidity: The authors note that in higher dimensions ($d \ge 2$), conformal mappings become mathematically rigid (limited to Möbius transformations). While they use coordinate-wise operations to maintain flexibility, this remains a theoretical constraint.
- Architecture Exploration: The study primarily used standard neural architectures adapted for complex inputs. The authors suggest that specialized complex-valued architectures (e.g., non-holomorphic networks) are an under-explored area that could further improve results.
- General Manifolds: A promising future direction is extending this complex-valued OT framework to general manifolds and exploring matrix-valued optimal transport problems.
Summary updated