Energy Generative Modeling: A Lyapunov-based Energy Matching Perspective

Technical Summary: Energy Generative Modeling: A Lyapunov-based Energy Matching Perspective

Problem Statement
Generative modeling typically relies on time-dependent vector fields (e.g., diffusion models, flow matching) to transport a simple noise distribution to a target data distribution. A recent paradigm, Energy Matching, eliminates time conditioning by learning a single static scalar energy function $U_\theta: \mathbb{R}^d \to \mathbb{R}$. Samples are generated by running dynamics on this static landscape, governed by the stochastic differential equation (SDE) $dx_t = -\nabla U(x_t)dt + \sqrt{2}dW_t$. However, the theoretical relationship between the training phase (learning $U_\theta$) and the sampling phase (generating samples) remains fragmented. Furthermore, it is unclear why the stochastic Langevin dynamics is necessary for distributional generation, as deterministic gradient flows on the same energy landscape often lead to "mode collapse" (concentration on local minima) rather than preserving the target distribution.

Methodology: Control-Theoretic Reformulation
The authors unify training and sampling within a single framework: density transport on the Wasserstein space $\mathcal{P}_2(\mathbb{R}^d)$, formulated as a nonlinear control problem.

• State and Control: The state variable is the probability density $\rho_t \in \mathcal{P}_2(\mathbb{R}^d)$. The dynamics are governed by the continuity equation $\partial_t \rho = -\nabla \cdot (\rho u)$, where the velocity field $u$ serves as the control input.
• Lyapunov Design: The Kullback-Leibler (KL) divergence, $V(\rho) = \text{KL}(\rho \| \rho^*)$, is selected as the Lyapunov function candidate, where $\rho^*$ is the target distribution.
• Optimal Control: To ensure convergence ($\dot{V} \leq 0$), the authors derive the optimal control law $u^* = -\nabla U^* - \nabla \log \rho$. Substituting this into the continuity equation yields the Fokker-Planck equation, whose particle-level realization is the Langevin SDE.
• Training vs. Sampling:
  • Training: The neural network learns $U_\theta$ to approximate the target. The target $\rho^*$ is defined as a smoothed version of the empirical data (via kernel density estimation) to ensure the existence of a score and finite KL divergence.
  • Sampling: The trained energy $U^*_\theta$ is used to drive the Langevin dynamics from a noise prior $\rho_0$ to the target $\rho^*_\theta$.

Key Contributions and Analytic Results

1. Unified Framework: The paper establishes that training, sampling, and smoothing are three instances of the same density transport dynamics on $\mathcal{P}_2(\mathbb{R}^d)$, governed by a continuity equation with a specific control law.
2. Impossibility of Deterministic Sampling:
   • The authors prove that for the deterministic gradient flow $\dot{x}_t = -\nabla U^*_\theta(x_t)$, no Lyapunov functional exists that simultaneously satisfies non-negativity, zero at the target, and non-increasing along trajectories.
   • This formally explains the phenomenon of mode collapse: without the entropy gradient term (realized by Brownian noise), the deterministic flow cannot maintain the target Gibbs distribution as a stationary solution; it inevitably concentrates mass on critical points.
3. Finite Step Stopping Criteria:
   • Langevin Sampling: Under the Log-Sobolev inequality, the KL divergence converges exponentially. The paper derives a sufficient sampling time $\tau^*_{\text{Lang}}$ to achieve a target accuracy $\epsilon$. Running the sampler beyond this time does not degrade the bound.
   • Deterministic Sampling: Since no global Lyapunov certificate exists, the authors propose a finite step stopping rule based on the ratio of the Laplacian term to the drift-induced dissipation ($R_t$). The optimal stopping time $\tau^*_{\text{det}}$ is the moment $R_t = 1$, marking the transition from drift-dominated dissipation (convergence) to Laplacian-dominated concentration (mode collapse).
4. Compositional Generation:
   • The framework demonstrates that additive composition of trained scalar energies ($U_{\text{comp}} = \alpha_+ U_+ + \alpha_- U_-$) retains an explicit Gibbs invariant measure.
   • This composition inherits the closed-loop Lyapunov certificate, allowing for the combination of "expert" models (e.g., conjunction, disjunction, negation) without retraining or structural loss of the invariant measure.
5. Diagnostic OOD Scoring:
   • The paper evaluates three scores derived from the closed-loop dynamics for Out-of-Distribution (OOD) detection: the energy value ($U_\theta$), the gradient norm ($\|\nabla U_\theta\|$), and trajectory displacement.
   • Experiments on CIFAR-10 show that the gradient norm ($S_2$) significantly outperforms the energy value and other baselines (like PixelCNN++ and GLOW) on natural OOD datasets (SVHN, DTD), achieving an average AUROC of 0.763.

Results and Validation

• Synthetic Experiments: On an eight-Gaussian target, deterministic flow reaches a near-optimal distribution at step $k=50$ before collapsing into Dirac masses at critical points. The predicted stopping time ($k^*=49$) aligns with the empirical optimum. In contrast, the Langevin sampler converges to the Gibbs distribution by $k \approx 100$ and maintains it, validating the theoretical distinction between the two stopping rules.
• Compositional Experiments: Training two independent energy models on disjoint Gaussian mixtures and combining them via addition, subtraction, and negation (without retraining) successfully produced samples corresponding to the logical operations (e.g., conjunction created a ring of compromise locations; negation successfully repelled samples from one mode set).
• OOD Detection: The gradient norm score demonstrated superior performance in distinguishing in-distribution CIFAR-10 data from SVHN and texture datasets compared to existing generative model baselines.

Significance and Claims
The paper claims to bridge static scalar energy generative modeling with the full toolkit of nonlinear control theory.

• It reframes generative modeling not as a parallel discipline but as a chapter of nonlinear control, where training and sampling are unified under Lyapunov stability principles.
• It provides the first rigorous proof that the stochastic nature of Langevin dynamics is not merely a heuristic but a necessary condition for the existence of a distributional Lyapunov certificate; removing the noise destroys the stability guarantee.
• It opens the door to applying advanced control concepts to generative models, specifically Control Barrier Functions (CBF) for safety-constrained generation (e.g., ensuring generated samples avoid unsafe regions) and contraction metrics for accelerated sampling.
• The authors position this work as a foundational layer that enables future extensions, such as safety guarantees and accelerated sampling, by treating the static energy paradigm as an autonomous control system.