Diffusion bridges have emerged as versatile frameworks for facilitating transformations between arbitrary data distributions. However, these approaches are hindered by an immense optimization space and its underlying high-order structures. This paper explores a variant of diffusion bridges for solving real-world problems in complex vector spaces governed by scaling and rotation laws, such as Hamiltonian structures. To address this, we introduce conformal optimal transport (COT), a complex-valued transportation framework designed for physical systems with geometric and dissipative dynamics under a geometric lifting. Building on this, we propose a new class of dynamical diffusion bridges, called complex Schrödinger bridges (CSBs), whose derived algorithms can handle diffusion-based tasks using a conformal geometric prior. For scalability, we propose image-to-image CSB (I2CSB), a tractable diffusion-bridge technique with an analytic spiral posterior. Finally, we implement a physics-informed CSB algorithm within simulations of quantum electrodynamics, mimicking the logical behavior of quantum logic gates. Our extensive results validate the underlying theory while highlighting advantages of complex-valued diffusion.
Figures & tables
Figure 1: An overview of a complex SB and its associated conformal optimal transport geometry.
Figure 2: Diffusion models approximate a distribution with different values of γ . Their stochastic trajectories exhibit varying dynamics and coverage of the density.
Figure 3: Conformal mappings under hγ(w) for varying coefficients. A Cartesian grid (normalized w -plane) is mapped biholomorphically onto a curvilinear grid, preserving right angles.
Figure 4
Figure 5: Results on two-dimensional control problems. (a) Visualization of the potential function and Langevin dynamics for 3 seconds in the Moons-Circles bridge. (b) Learned CSB dynamics for S-Tunnel, GMM-8, and V-Neck. (c) Statistics on marginal fitting and total costs (5 different seeds).
Figure 6: Learned complex-valued log-potential functions at three distinct time points ( γ=0.5 ).
Figure 9: Complex SB with a synthetic cQED dataset. (a) Physical simulation and prediction for a harmonic oscillator with fluctuation noise. (b) For a long time horizon, our SB representation was fitted to the optimal state. (c) 3D qubit logic gate simulation (averaging over 100 particles).
Table 2: Qubit OT validation at higher, unseen energies for Hadamard and Pauli gates (5 runs).
Appendix figures & tables12 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 10: Conformal mappings in one complex dimension C1 from Moon and Spencer [59] .
Figure 11: Conformal optimal transport geometry for γ={0.1,0.5,1.0,2.5} . Time increases downward. Increasing γ produces greater angular winding through eiγlog(1−t) , while the radial contraction remains 1−t , preserving the self-similarity of the transported geometry.
Figure 12: Moons to circles.
Method
ξ=10−3
ξ=10−2
ξ=10−1
NODE [ 13 ]
4027144
4197223
234321
HNN [ 30 ]
8.37 ±0.55
7.79 ±1.52
9.35 ±0.77
Neural SDE
1.39±2.47
1.54±3.19
2.92±2.63
Ours (potential net)
0.16 ±0.10
0.35 ±0.33
1.61±1.40
Appendix
Table 3: Fitting a harmonic oscillator trajectory from noisy quantum observations (5 seeds).
γ
W2 -rAUC(.94) ↓
MMD2 -rAUC(.94) ↓
W2 -rAUC(.95) ↓
MMD2 -rAUC(.95) ↓
W2 -rAUC(.97) ↓
MMD2 -rAUC(.97) ↓
W2 -rAUC(.99) ↓
MMD2 -rAUC(.99) ↓
0
0.202734
0.051280
0.154899
0.039132
0.063955
0.015614
0.007683
0.001424
1
0.203424
0.051459
0.155543
0.039302
0.064253
0.015706
0.007710
0.001433
3
0.203286
0.051405
0.155257
0.039217
0.064030
0.015637
0.007709
0.001433
5
0.203372
0.051429
0.155256
0.039224
0.064083
0.015658
0.007723
0.001437
10
0.203020
0.051384
0.155370
0.039272
0.064292
0.015720
0.007733
0.001441
30
0.203512
0.051452
0.155503
0.039270
0.064053
0.015648
0.007682
0.001424
Appendix
Table 4: 2-Wasserstein saturation of the analytic posterior for oscillator states of the Hadamard operation. Bold and underlined values denote the best and second-best results.
γ
FID at 0.9
FID at 0.95
FID at 0.99
FID at 0.995
0
31.316002
2.152040
0.081899
0.094650
0.0001
31.315914
2.152060
0.081906
0.094652
0.001
31.315853
2.151930
0.081904
0.094637
0.01
31.321283
2.152645
0.081816
0.094604
0.1
31.520388
2.184362
0.081928
0.094841
1
47.182999
5.063271
0.094534
0.100584
Appendix
Table 5: CIFAR-10 posterior FID scores for different values of γ .
Table 16
CNOT
SWAP
OT-CFM
0.9993
0.9448
SB-CFM
0.9062
0.9991
Ours ( γ=1 )
0.9994
0.9998
Appendix
Table 8: Qubit fidelity of CNOT and SWAP 2-qubit logic gates.
Table 9: MNIST FID.
Method
FID ↓
SSIM ↑
DSBM
23.1730
0.6126
SB Flow
21.4061
0.6312
I 2 SB
15.2626
0.6597
Ours ( γ=0.001 )
12.0784
0.7130
Ours ( γ=0.01 )
12.4734
0.7104
Ours ( γ=0.1 )
12.0135
0.7139
Appendix
Table 10: Magnitude FID and SSIM metrics on the paired 64×64 synthetic brain MRI validation dataset for unseen subjects (source-target slice separation: 24 mm).
Figure 13: MNIST and CIFAR-10 image-to-image translation
Figure 15: Complex MRI slice-transport results for fixed validation examples. Each block shows magnitude, phase, real, and imaginary components with shared display ranges.