| name | neural-phase-correlation |
| description | Learned generalization of phase correlation that lifts the fixed Fourier basis restriction to discover unknown transformations between observations. Applicable to image registration, non-rigid deformation, and quantum Hamiltonian eigenstate recovery from observation pairs. |
| metadata | {"arxiv_id":"2606.18496","published":"2026-06-16","authors":"Cole Reynolds","tags":["phase-correlation","neural-networks","image-registration","quantum","fourier","correspondence"]} |
Neural Phase Correlation
Description
A learned generalization of phase correlation that lifts the fixed Fourier basis restriction, enabling discovery of unknown transformations between observations. Extends from global translation to dense non-rigid deformations and unitary dynamics. Validated on cardiac MRI, echocardiography, and quantum harmonic oscillator eigenstate recovery.
Activation Keywords
- neural phase correlation
- learned phase correlation
- image registration neural
- fourier correspondence
- transformation discovery
- quantum eigenstate recovery
- 相位相关学习
Core Concepts
Phase Correlation Generalization
- Classical limitation: Standard phase correlation only measures global translation via fixed Fourier basis
- Key insight: Learn the basis on which the transformation decomposes, rather than using fixed Fourier modes
- Algebraic primitive: Same mathematical foundation extends to dense non-rigid deformations and unitary dynamics
- First-class transformation: Architecture represents transformation as explicit object, not implicit through similarity functions
Applications
- Medical imaging registration: ACDC cardiac MRI, CAMUS echocardiography
- Quantum physics: Recovering Hamiltonian eigenstates from observation pairs
- Non-rigid deformation: Dense correspondence beyond global translation
Usage Patterns
Pattern 1: Image Registration
When registering medical images or general observation pairs:
- Encode both observations into shared representation
- Apply learned phase correlation in Fourier-like domain
- Decode transformation from correlation peak
- Achieves state-of-the-art without auxiliary scoring mechanisms
Pattern 2: Quantum State Analysis
When analyzing quantum systems from observation data:
- Apply to pairs of time-evolved wavefunctions
- Framework recovers Hermite-function eigenstates
- Quantized energy levels extracted from observation pairs alone
- No prior knowledge of Hamiltonian required
Pattern 3: Non-rigid Deformation Tracking
When tracking dense non-rigid deformations:
- Same algebraic primitive extends beyond translation
- Learns basis for deformation decomposition
- Matches/exceeds prior baselines on cardiac benchmarks
Methodology
Step 1: Representation Learning
- Learn the decomposition basis from data
- Same architecture handles translation, deformation, and unitary dynamics
Step 2: Correlation Computation
- Apply phase-correlation algebraic primitive on learned basis
- Extract transformation parameters from correlation structure
Step 3: Validation
- Evaluate on paired observations
- For quantum systems: verify eigenstate and energy recovery
Error Handling
Mode Collapse in Learned Basis
- Ensure basis diversity through regularization
- Validate against known ground-truth transformations
High-Frequency Artifacts
- Apply appropriate filtering in learned frequency domain
- Balance resolution with noise robustness
Examples
Example: Quantum Harmonic Oscillator
Applied to time-evolved wavefunction pairs of 1-D quantum harmonic oscillator, the framework recovers Hermite-function eigenstates and quantized energy levels from observation pairs alone — no Hamiltonian specification needed.
Resources
- arXiv: 2606.18496 - "Neural Phase Correlation"
- Related:
quantum-state-engineering, quantum-brain-modeling