| name | conjugacy-based-similarity-analysis |
| description | Conjugacy-based Similarity Analysis (CSA) methodology for comparing dynamical systems in neuroscience and ML. Addresses limitations of Dynamical Similarity Analysis (DSA) by restricting alignments to state-space bijections rather than arbitrary orthogonal matrices. |
| version | 1.0.0 |
| date | 2026-07-08T00:00:00.000Z |
| source | arXiv:2607.04493 |
| authors | Prakhar Godara, Pang Shiang Tay, Marcelo G. Mattar |
| tags | ["dynamical-systems","neuroscience","koopman-theory","conjugacy","similarity-analysis","neural-dynamics"] |
Conjugacy-based Similarity Analysis (CSA)
Overview
CSA is a methodology for comparing whether two dynamical systems implement the same computation despite differences in coordinates or measurements. It addresses fundamental limitations of Dynamical Similarity Analysis (DSA) in neuroscience and machine learning.
Core Problem
Comparing dynamical systems is central to neuroscience and ML:
- Do two neural networks implement the same computation?
- Are two brain regions performing similar operations?
- How do we measure similarity when systems use different coordinates?
DSA Limitations
Dynamical Similarity Analysis (DSA) aligns finite-dimensional Koopman approximations through orthogonal similarity transformations.
Key Insight: Orthogonal alignment is neither necessary nor sufficient for topological conjugacy:
- Conjugate systems may require non-orthogonal basis-transfer matrices that DSA cannot capture
- Non-conjugate systems may have orthogonally equivalent Koopman operators that DSA fails to distinguish
CSA Methodology
Core Principle
CSA restricts alignments to those induced by candidate state-space bijections rather than arbitrary orthogonal matrices.
Mathematical Foundation
- CSA's fitted alignment is the finite-data projection of the composition operator associated with the candidate bijection
- This ensures the alignment respects the actual state-space mapping between systems
Key Theoretical Result
Theorem: CSA's alignment matrix equals the projection of the Koopman composition operator onto the observable dictionary space.
When to Use CSA
Appropriate Scenarios
- Comparing neural recordings from different brain regions or subjects
- Validating computational models against biological data
- Analyzing RNN dynamics to identify functional equivalences
- Studying neural manifolds across experimental conditions
Inappropriate Scenarios
- When systems have fundamentally different state-space dimensions
- When no reasonable bijection hypothesis exists
- For purely statistical similarity without mechanistic interpretation
Implementation Steps
1. Define Candidate Bijection
def candidate_bijection(x, A, b):
"""Affine transformation between state spaces"""
return A @ x + b
2. Compute Koopman Operators
def compute_koopman(data, observables):
"""
data: (T, n_features) time series
observables: function mapping state to observation space
"""
K = least_squares_fit(observables(data[:-1]), observables(data[1:]))
return K
3. Fit CSA Alignment
def csa_alignment(K1, K2, bijection_params):
"""
K1, K2: Koopman operators from two systems
bijection_params: parameters of candidate bijection
"""
C = composition_operator(K1, K2, bijection_params)
alignment = project_to_observables(C)
return alignment
4. Evaluate Conjugacy Quality
def conjugacy_error(K1, K2, alignment):
"""Measure how well alignment satisfies conjugacy relation"""
error = norm(K1 - inv(alignment) @ K2 @ alignment)
return error
Comparison with DSA
| Aspect | DSA | CSA |
|---|
| Alignment constraint | Orthogonal matrices | State-space bijections |
| Captures conjugacy | No (only orthogonal equivalence) | Yes (by construction) |
| False positives | Possible (non-conjugate systems) | Avoided |
| False negatives | Possible (conjugate but non-orthogonal) | Avoided |
| Interpretability | Limited | Direct mechanistic meaning |
Practical Considerations
Observable Dictionary Selection
- Critical: Results depend heavily on observable choice
- Use domain knowledge to select meaningful observables
- Test robustness across different observable sets
Finite Data Effects
- CSA alignment is a finite-data projection
- Convergence guarantees require sufficient data
- Use cross-validation to assess stability
Computational Cost
- More expensive than DSA due to bijection optimization
- Parallelize over candidate bijections
- Use warm starts from DSA solution
Applications in Neuroscience
Neural Population Analysis
Compare population dynamics across:
- Different experimental conditions
- Brain regions performing similar computations
- Species with homologous circuits
Model Validation
Test whether computational models capture:
- Qualitative dynamics (fixed points, limit cycles)
- Quantitative trajectories
- Response to perturbations
Learning and Plasticity
Track how neural representations evolve:
- During learning
- Across development
- After injury/recovery
Pitfalls and Limitations
- Bijection hypothesis: CSA assumes a bijection exists; if wrong, results are meaningless
- Observability: Systems must be sufficiently observed; hidden states break assumptions
- Noise sensitivity: Finite data + noise can obscure true conjugacies
- Computational scaling: High-dimensional systems require careful optimization
Related Methods
- Dynamical Similarity Analysis (DSA): Orthogonal Koopman alignment
- Representational Similarity Analysis (RSA): Compare representational geometries
- Procrustes analysis: Shape comparison under rigid transformations
- Optimal transport: Distribution-level comparisons
References
- Godara, P., Tay, P. S., & Mattar, M. G. (2026). Beyond DSA: Conjugacy-based Comparison of Dynamical Systems. arXiv:2607.04493
- Ostrow, E., et al. (2023). Dynamical Similarity Analysis.
Activation Triggers
Use this skill when:
- Comparing neural dynamics across conditions/subjects
- Validating computational models against neural data
- Analyzing RNN internal representations
- Studying neural manifold structure
- Keywords: dynamical systems, Koopman, conjugacy, similarity analysis, neural comparison