Entropy-Controlled Flow Matching (ECFM)

• Entropy-Controlled Flow Matching (ECFM) is a variational framework that integrates an entropy-rate constraint into flow matching, effectively preventing transient mode collapse and ensuring comprehensive semantic mode coverage.
• It employs convex minimization in Wasserstein space along with KKT conditions and the Pontryagin framework to regulate drift fields and the evolution of differential entropy.
• ECFM bridges optimal transport, Schrödinger bridge problems, and KL-control, offering practical guarantees such as certified lower bounds on mode masses and Lipschitz stability under perturbations.

Entropy-Controlled Flow Matching (ECFM) is a constrained variational framework for transporting probability measures in generative modeling, with a strict control on the entropy evolution along the induced path. ECFM introduces a global entropy-rate constraint to conventional flow matching, directly regulating information geometry during the transformation between distributions. This prevents transient mode collapse and ensures robust coverage over semantic modes, in contrast to standard unconstrained flow matching formulations, which may induce pathological low-entropy bottlenecks. ECFM is formulated as a convex minimization in Wasserstein space, with rigorous links to optimal transport, Schrödinger bridge problems, and stochastic control theory (Maduabuchi, 25 Feb 2026).

1. Variational Principle and Entropy Constraint in Wasserstein Space

Let $\mu_0$, $\mu_T \in \mathcal P_2^{ac}(\mathbb R^d)$ be probability measures corresponding to the initial and terminal distributions. Let $u^\star(x,t)$ denote a reference drift field (teacher). The class of admissible continuity-equation curves is: $$\mathfrak A(\mu_0,\mu_T) = \left\{(\mu_t,v_t):\; \partial_t\mu_t + \nabla\cdot (\mu_t v_t) = 0, \; \mu|_{t=0} = \mu_0, \; \mu|_{t=T} = \mu_T, \; \int_0^T\int\|v_t\|^2\,d\mu_t<\infty\right\}.$$ With differential entropy $\mathcal H(\mu) = \int\rho\log\rho\,dx$ for absolutely continuous $\mu = \rho\,dx$, the entropy rate evolves as $$\dot{\mathcal H}(\mu_t) = \int \nabla\cdot v_t\, d\mu_t$$. ECFM introduces a uniform lower bound: $$\dot{\mathcal H}(\mu_t)\geq -\lambda, \quad\text{almost everywhere on } [0,T], \quad \lambda\geq 0.$$ The primal ECFM problem is: $$\boxed{ (\mathrm{ECFM}_\lambda)\quad \min_{(\mu,v)\in\mathfrak A_\lambda(\mu_0,\mu_T)}\; \frac12\int_0^T\!\!\int\|v_t(x)-u^\star(x,t)\|^2\,d\mu_t(x)\,dt, }$$ where $\mathfrak A_\lambda$ denotes flows satisfying both the continuity equation and the entropy-rate budget.

2. KKT Conditions, Pontryagin System, and Convexity

Enforcement of the entropy-rate constraint leads to additional dual variables. The Lagrange multiplier $\eta(t)\ge0$ handles the entropy constraint, and the adjoint costate potential $\phi(x,t)$ is conjugate to the continuity equation. The resulting KKT (Karush-Kuhn-Tucker) system (Theorem 4.1) comprises:

• Primal feasibility:

$$\partial_t\mu^\lambda + \nabla\cdot(\mu^\lambda v^\lambda)=0,\quad \dot{\mathcal H}(\mu^\lambda_t)\ge -\lambda.$$

• Stationarity in $v$:

$$v^\lambda(x,t) = u^\star(x,t) - \nabla\phi(x,t) - \eta(t)\,\nabla\log\rho_t(x).$$

• Complementarity conditions:

$$\eta(t)\ge0,\quad \dot{\mathcal H}(\mu^\lambda_t)+\lambda\ge0,\quad \eta(t)\big(\dot{\mathcal H}(\mu^\lambda_t)+\lambda\big)=0.$$

• Weak-form adjoint (costate) equation:

$$\int_0^T\int \Bigl[ -\partial_t\phi - \nabla\phi\cdot v^\lambda - \tfrac12\|v^\lambda-u^\star\|^2 - \eta\,\nabla\!\cdot v^\lambda \Bigl]\zeta\,d\mu^\lambda\,dt = 0$$

for all test functions $\zeta$.

In flux variables ($m_t = \rho_t v_t$), this also realizes the Pontryagin Maximum Principle. The Hamiltonian density is

$$\mathcal H(t,\mu,v,\phi,\eta) = -\tfrac12\int|v-u^\star|^2\,d\mu +\int (\nabla\phi-\eta\nabla\log\rho)\cdot v\,d\mu - \eta\,\lambda.$$

The system is convex in $(\mu,v)$ due to the quadratic kinetic energy term.

3. Stochastic Control and Schrödinger Bridge Connections

The stochastic control formulation introduces a reference path law $R$ for the process $dX_t = u^\star(X_t,t) dt + \sqrt{2\varepsilon}dW_t$. The dynamic Schrödinger bridge problem minimizes the relative entropy: $$\min_{P:\,P_0=\mu_0,\,P_T=\mu_T}\; \mathrm{KL}(P\,\|\,R).$$ Using Girsanov's theorem, a controlled law $P^w$ with drift $w_t$ yields

$$\mathrm{KL}(P^w\|R)=\frac{1}{4\varepsilon}\mathbb E_{P^w}\int_0^T\|w_t\|^2dt.$$

Defining current velocity as $v_t = b_t - \varepsilon\nabla\log\rho_t$ for $b_t = u^\star + w_t$, minimization of the ECFM functional reduces, up to an explicit entropy/Fisher correction, to minimizing KL-control in this setup. In the pure transport regime ($\varepsilon\to0$ or $u^\star=0$), this correction vanishes and ECFM exactly recovers the Schrödinger bridge/KL-control problem (Maduabuchi, 25 Feb 2026).

4. Entropic Optimal Transport and $\Gamma$-Convergence

With vanishing reference drift ($u^\star\equiv0$), ECFM specializes to a constrained kinetic action: $$\min_{\dot{\mathcal H}\geq -\lambda}\; \frac12\int\|v\|^2\,d\mu\,dt$$ which, via correspondence, matches the entropic optimal transport (entropic-OT) geodesic at an implicit regularization $\varepsilon(\lambda)$. There is a one-to-one mapping between the entropy budget $\lambda$ and the entropic regularization parameter $\varepsilon$.

As $\lambda\downarrow 0$, the ECFM functional $\mathcal F_\lambda$ $\Gamma$-converges to the classic Benamou–Brenier optimal transport action: $$\Gamma\text{-}\lim_{\lambda\to 0} \mathcal F_\lambda = \mathcal F_0, \qquad \mathcal F_0 = \tfrac12\int\|v\|^2\rho\,dxdt,$$ demonstrating that ECFM interpolations converge to classical OT geodesics as the entropy constraint is relaxed.

5. Mode Coverage, Density Floors, and Lipschitz Stability

ECFM provides explicit, certified lower bounds for mode masses and density floors along the entire transport trajectory. For disjoint sets $A_k$ in $\mathbb{R}^d$ (semantic modes), with mode masses $M_k(t) = \mu_t(A_k)$, ECFM guarantees under the entropy-rate constraint (Theorem 7.1): $$\inf_{t\in[0,T]} M_k(t) \geq \beta_k(\lambda,\mu_0,\mu_T) > 0.$$ If each mode contains a compact core $K_k$ where initial and final densities are positive, then for $x \in K_k$ and $t\in[\delta, T-\delta]$: $\rho_t(x) \geq \underline\rho_k > 0.$ The framework is Lipschitz-stable: under small perturbations of endpoints, reference drift, entropy budget, or additive noise, all lower floors degrade at most linearly in the perturbation size, and the path evolves Lipschitz-continuously in $W_2$.

6. Collapse Counterexamples and Limitations of Unconstrained Flow Matching

Constructed examples in one dimension define flow maps $\Phi_t^{(n)}$ that compress mixture-distributed endpoints into a transient bottleneck, holding the mass near a point before re-expanding it to the target $\mu_T$. These flows fulfill the continuity equation and match endpoints, with near-zero flow-matching loss, yet the path densities $\mu_t$ collapse to a Dirac at intermediate times, causing semantic mode loss. The entropy rate $\dot{\mathcal H} \to -\infty$ during contraction, violating any finite ECFM budget. No uniform lower bounds for mode masses or densities exist in standard, unconstrained flow matching, highlighting the necessity of controlled entropy.

7. Practical Algorithm and Certification Procedure

Algorithm 1 (primal–dual with augmented Lagrangian) operationalizes ECFM for training parameterized flows:

Algorithm: ECFM Primal–Dual Training Input: time grid {t_n}, budgets {λ_n}, penalty ρ>0, step‐sizes α,β Initialize θ⁰, multipliers η⁰_n←0 for k=0,1,2,… do Sample minibatches {x_{i,n}∼μ^{θ^k}_{t_n} Estimate entropy‐rate: Ĥ̇_n = (1/B)∑ₙ div v_{θ^k}(x_{i,n},t_n) (or FP form) Residual g_n = −Ĥ̇_n − λ_n Aug‐Lagrangian loss L_AL(θ,η) = L_FM(θ) + ∑_n[η_n g_n + (ρ/2)(g_n)₊²] θ^{k+1} ← θ^k − α ∇_θ L_AL(θ^k,η^k) η_n^{k+1} ← [η_n^k + β g_n(θ^{k+1})]_+ end for

Here, $$L_{FM}(\theta) = \tfrac12 E_{t,x\sim\mu_t^\theta}\|v_\theta(x,t)-u^\star(x,t)\|^2$$. Divergences $\operatorname{div}v_\theta$ may be estimated using automatic differentiation (JVPs/Hutchinson trace) or via Fokker-Planck score identities. The entropy budget $\lambda_n$ can be uniform or adapted online. At test time, certification involves estimating the empirical entropy rate and its lower confidence bound to verify compliance with the entropy constraint, and computing mode masses with statistical confidence. These combine to provide a verifiable, architecture-agnostic anti-collapse mechanism.

For theorems, technical proofs, and further implementation specifics, see (Maduabuchi, 25 Feb 2026).