Technical Summary: SplineFlow

Problem Statement

Flow matching has emerged as a scalable generative framework for characterizing continuous normalizing flows, offering simulation-free training and efficient inference compared to score-based diffusion and traditional continuous normalizing flows (CNFs). However, existing state-of-the-art flow matching methods face significant limitations when applied to modeling dynamical systems:

1. Suboptimal Supervision via Linear Interpolants: Current methods construct conditional paths by linearly interpolating between observed data points. While efficient, linear interpolants fail to capture the underlying curvature and complex state evolution of nonlinear, high-curvature, or oscillatory dynamical systems. This results in a biased supervision signal that may not faithfully represent the true dynamics between observation points.
2. Instability in Multi-Marginal Constraints: Constructing unified paths to satisfy multi-marginal constraints (connecting multiple observations across time) using naïve higher-order polynomials often leads to instability and oscillatory behavior, known as Runge's phenomenon.
3. Scalability vs. Fidelity Trade-off: While CNFs can model complex dynamics, they require expensive adjoint simulations at each training iteration, limiting scalability. Existing flow matching approaches avoid this but sacrifice the fidelity of the learned dynamics by relying on restrictive linear paths.

Methodology: SplineFlow

The authors propose SplineFlow, a flow matching algorithm that replaces linear interpolants with B-spline interpolants to construct conditional probability paths. This approach leverages the smoothness, stability, and flexibility of B-splines to model complex dynamics while ensuring multi-marginal constraints are met.

Core Components

1. B-Spline Conditional Paths:

   • Instead of piecewise linear connections, SplineFlow constructs the conditional mean $\mu_t(z)$ of the probability path $p_t(x|z)$ using B-spline functions of degree $m$.
   • The B-spline basis functions are defined recursively via the Cox-de Boor formula. The interpolant $\mu(t)$ is a linear combination of these bases, constrained to pass through the observed trajectory points.
   • Theoretical Advantage: The authors prove (Theorem 3.1) that for a function with bounded $m$-th derivatives, B-spline interpolants achieve an approximation error of $O(n^{-m})$ for the function and $O(n^{-m+1})$ for the first derivative, significantly outperforming linear interpolants ($O(n^{-2})$ and $O(n^{-1})$ respectively). This allows for a more accurate characterization of the underlying velocity field.

2. Deterministic Dynamics (ODEs):

   • The method models the conditional path as a Gaussian distribution $N(x; \mu_t(z), \sigma^2_t I)$ where $\mu_t(z)$ is the B-spline interpolant.
   • The conditional velocity field $u_t(x|z)$ is derived analytically from the B-spline derivative (Theorem 4.3).
   • The neural network $u_\theta$ is trained to regress this conditional velocity field using the standard flow matching regression loss.

3. Stochastic Dynamics (SDEs):

   • SplineFlow is extended to the Score and Flow Matching (SF2M) framework to handle stochastic differential equations with additive noise.
   • It jointly regresses the probability flow velocity field ($u_\theta$) and the score function ($s_\phi$) over B-spline conditional paths (Spline-Bridges).
   • The authors propose a piecewise quadratic variance scheme for the conditional paths in cellular dynamics contexts, which empirically outperforms constant variance schemes in specific settings.

4. Irregular Sampling:

   • The framework explicitly handles irregularly sampled observations (where a fraction $p$ of data points are masked). B-splines are noted for their stability even under irregular sampling regimes, unlike global polynomial interpolants.

Key Contributions

1. Algorithm Development: Introduction of SplineFlow, the first flow matching framework to utilize B-spline interpolants for constructing conditional paths in both deterministic and stochastic dynamical systems.
2. Theoretical Analysis:
   • Formal proof that B-spline interpolants offer superior asymptotic approximation error rates for both function values and derivatives compared to linear interpolants.
   • Derivation of error bounds for the learned velocity field and generated samples, showing that SplineFlow reduces the approximation error component of the total error bound from $O(n^{-1})$ (linear) to $O(n^{-m+1})$ (B-spline).
   • Proof that the B-spline construction satisfies the continuity equation and recovers marginal distributions.
3. Empirical Validation: Comprehensive experiments demonstrating that SplineFlow outperforms baselines (NeuralODE, LatentODE, TFM, MMFM, SF2M) across:
   • Deterministic Systems: Including Exponential Decay, Harmonic/Damped Harmonic Oscillators, Lotka-Volterra, and the chaotic Lorenz system.
   • Stochastic Systems: SDE counterparts of the above systems.
   • Real-World Data: Unpaired cellular trajectory inference tasks (Brain Regeneration and Embryoid Evolution datasets).

Experimental Results

• ODE Dynamics: SplineFlow consistently achieves lower Mean Squared Error (MSE) than baselines, particularly for nonlinear and oscillatory systems (e.g., Lotka-Volterra, Lorenz) and under irregular sampling ($p > 0$). It is significantly faster (100x–600x speedup) than adjoint-based methods like NeuralODE and LatentODE.
• SDE Dynamics: For stochastic systems, SplineFlow (with higher-degree splines) outperforms SF2M, especially in nonlinear (Lotka-Volterra) and oscillatory (Damped Harmonic) regimes.
• Cellular Trajectories: In transcriptomic data tasks, SplineFlow outperforms MOTFM and SF2M in interpolation settings and remains competitive in extrapolation. The method shows particular strength when using PHATE embeddings, which capture nonlinear geometric structures.
• Ablation Studies: The results indicate that higher spline degrees ($m > 1$) are beneficial for nonlinear and irregularly sampled data, while linear degrees ($m=1$) suffice for linear dynamics. A simple cross-validation procedure is shown to effectively select the optimal degree.

Significance and Claims

The paper claims that SplineFlow provides a theoretically grounded and scalable alternative to existing flow matching methods for dynamical systems. Its primary significance lies in:

• Faithful Dynamics Modeling: By moving beyond linear interpolants, SplineFlow captures the true underlying state evolution and curvature of dynamical systems, rather than just generating valid endpoint samples.
• Robustness to Irregularity: The method maintains performance and stability even when observations are sparse or irregularly sampled, a common challenge in real-world biological and physical data.
• Efficiency: It retains the simulation-free training advantage of flow matching while offering higher fidelity than linear-path variants, bridging the gap between the efficiency of flow matching and the modeling capacity required for complex dynamics.

The authors position SplineFlow as a general-purpose framework applicable to diverse domains, from physical simulations (e.g., HopperPhysics) to biological processes (e.g., drug treatment response and cell differentiation), without claiming specific novel applications beyond those tested.