LightSBB-M Algorithm Insights

• LightSBB-M is a dual-method approach that efficiently implements Schrödinger–Bass Bridge transport and sparse Bayesian learning to optimize generative model trajectories.
• It employs a tunable beta parameter to balance deterministic drift and stochastic volatility, achieving lower sample errors and robust performance.
• The algorithm uses dual representations and iterative updates to avoid high-complexity methods like neural ODE rollouts, ensuring efficient computation and improved empirical results.

LightSBB-M refers to two distinct algorithms introduced in separate fields: (1) a simulation-efficient solver for Schrödinger–Bass Bridge (SBB) transport in generative modeling, and (2) a low-complexity BP–MF message-passing approach for sparse Bayesian learning (SBL) on a stretched factor graph. Both algorithms exploit dual representations and efficient iterations to outperform prior baselines in their respective domains. The following account characterizes both definitions in detail, covering their mathematical frameworks, algorithmic workflow, theoretical properties, empirical performance, and technical implementation.

1. Schrödinger–Bass Bridge Formulation: Mathematical Grounding

The SBB problem extends classical Schrödinger Bridge (SB) theory by introducing a tunable parameter $\beta > 0$ that interpolates between deterministic drift (SB, $\beta \rightarrow \infty$) and stochastic volatility (Bass martingale transport, $\beta \rightarrow 0$). The objective is formulated as:

$$J^* = \inf_{\mathbb{P} \in \mathcal{P}(\mu_0,\mu_T)} \mathbb{E}_{\mathbb{P}}\left[ \int_0^T \left( \|\alpha_t\|^2 + \beta \|\sigma_t - \sqrt{\epsilon}I_d\|^2 \right) dt \right]$$

subject to the Itô process: $$dX_t = \alpha_t dt + \sigma_t dW_t,\quad X_0 \sim \mu_0,\ X_T \sim \mu_T$$

A dual variational form leverages Lagrange multipliers $(\psi, v)$ to enforce terminal/final-time constraints and the Hamilton–Jacobi–Bellman PDE: $$J^* = \sup_{\psi,v} \left\{ \int \psi(x)\,\mu_T(dx) - \int v(0,x)\,\mu_0(dx) \right\}$$ with a backward PDE: $$\partial_t v(t,x) + H_\beta^*(\nabla_x v, D^2_x v) = 0,\quad v(T,\cdot) = \psi(\cdot)$$ where

$$H_\beta^*(p, q) = \frac{1}{2}|p|^2 + \frac{\beta \epsilon}{2} I_d : \left( (I_d - q/\beta)^{-1} - I_d \right),\quad q<\beta I_d$$

This duality allows closed-form optimal controls for drift and volatility: $\alpha^*(t,x) = \nabla_x v^*(t,x)$

$\beta \rightarrow \infty$0

These expressions are crucial for highly efficient computation and generative path sampling (Alouadi et al., 27 Jan 2026).

2. Algorithmic Workflow for Generative Diffusion

LightSBB-M operationalizes the SBB framework as follows:

• Initialization: Parameters of a transport map network $\beta \rightarrow \infty$1 and mixture potential $\beta \rightarrow \infty$2 are set. $\beta \rightarrow \infty$3, $\beta \rightarrow \infty$4, and minibatch size $\beta \rightarrow \infty$5 are specified.
• Outer Loop (typically $\beta \rightarrow \infty$6 iterations):

1. Endpoint Push-Forward: Map data samples $\beta \rightarrow \infty$7 to latent codes via $\beta \rightarrow \infty$8.
2. Bridge Sampling: For each $\beta \rightarrow \infty$9, sample intermediate $\beta \rightarrow 0$0 along convex decompositions and apply stochastic perturbation (Brownian noise).
3. Drift Model Update: Minimize the squared error between the analytic drift $\beta \rightarrow 0$1 and the conditional endpoint difference $\beta \rightarrow 0$2 by SGD.
4. Transport Map Update: Minimize the endpoint reconstruction loss for the composed map $\beta \rightarrow 0$3, with $\beta \rightarrow 0$4.

• Inference: For unseen $\beta \rightarrow 0$5, compute $\beta \rightarrow 0$6, sample $\beta \rightarrow 0$7 conditioned on $\beta \rightarrow 0$8, and recover $\beta \rightarrow 0$9.

Convergence occurs within a small number of outer iterations ($$J^* = \inf_{\mathbb{P} \in \mathcal{P}(\mu_0,\mu_T)} \mathbb{E}_{\mathbb{P}}\left[ \int_0^T \left( \|\alpha_t\|^2 + \beta \|\sigma_t - \sqrt{\epsilon}I_d\|^2 \right) dt \right]$$0), and the algorithm avoids high-dimensional convolutions, neural ODE rollouts, or expensive iterative proportional fitting (Alouadi et al., 27 Jan 2026).

3. Role of the Interpolation Parameter $$J^* = \inf_{\mathbb{P} \in \mathcal{P}(\mu_0,\mu_T)} \mathbb{E}_{\mathbb{P}}\left[ \int_0^T \left( \|\alpha_t\|^2 + \beta \|\sigma_t - \sqrt{\epsilon}I_d\|^2 \right) dt \right]$$1

The parameter $$J^* = \inf_{\mathbb{P} \in \mathcal{P}(\mu_0,\mu_T)} \mathbb{E}_{\mathbb{P}}\left[ \int_0^T \left( \|\alpha_t\|^2 + \beta \|\sigma_t - \sqrt{\epsilon}I_d\|^2 \right) dt \right]$$2 controls the drift-versus-volatility tradeoff:

• $$J^* = \inf_{\mathbb{P} \in \mathcal{P}(\mu_0,\mu_T)} \mathbb{E}_{\mathbb{P}}\left[ \int_0^T \left( \|\alpha_t\|^2 + \beta \|\sigma_t - \sqrt{\epsilon}I_d\|^2 \right) dt \right]$$3: Recovers classical Schrödinger Bridge, i.e., optimal coupling is encoded entirely in deterministic drift, volatility remains fixed.
• $$J^* = \inf_{\mathbb{P} \in \mathcal{P}(\mu_0,\mu_T)} \mathbb{E}_{\mathbb{P}}\left[ \int_0^T \left( \|\alpha_t\|^2 + \beta \|\sigma_t - \sqrt{\epsilon}I_d\|^2 \right) dt \right]$$4: Recovers Bass martingale transport, i.e., drift vanishes and adaptive volatility absorbs the distributional constraints.
• Intermediate $$J^* = \inf_{\mathbb{P} \in \mathcal{P}(\mu_0,\mu_T)} \mathbb{E}_{\mathbb{P}}\left[ \int_0^T \left( \|\alpha_t\|^2 + \beta \|\sigma_t - \sqrt{\epsilon}I_d\|^2 \right) dt \right]$$5: Achieves optimality between smooth deterministic flow and stochastic paths, empirically producing lowest sample errors for $$J^* = \inf_{\mathbb{P} \in \mathcal{P}(\mu_0,\mu_T)} \mathbb{E}_{\mathbb{P}}\left[ \int_0^T \left( \|\alpha_t\|^2 + \beta \|\sigma_t - \sqrt{\epsilon}I_d\|^2 \right) dt \right]$$6–$$J^* = \inf_{\mathbb{P} \in \mathcal{P}(\mu_0,\mu_T)} \mathbb{E}_{\mathbb{P}}\left[ \int_0^T \left( \|\alpha_t\|^2 + \beta \|\sigma_t - \sqrt{\epsilon}I_d\|^2 \right) dt \right]$$7 on multimodal transport tasks. Too-small $$J^* = \inf_{\mathbb{P} \in \mathcal{P}(\mu_0,\mu_T)} \mathbb{E}_{\mathbb{P}}\left[ \int_0^T \left( \|\alpha_t\|^2 + \beta \|\sigma_t - \sqrt{\epsilon}I_d\|^2 \right) dt \right]$$8 induces high-variance trajectories; too-large $$J^* = \inf_{\mathbb{P} \in \mathcal{P}(\mu_0,\mu_T)} \mathbb{E}_{\mathbb{P}}\left[ \int_0^T \left( \|\alpha_t\|^2 + \beta \|\sigma_t - \sqrt{\epsilon}I_d\|^2 \right) dt \right]$$9 is suboptimal when $$dX_t = \alpha_t dt + \sigma_t dW_t,\quad X_0 \sim \mu_0,\ X_T \sim \mu_T$$0 is heavy-tailed.

This continuous interpolation is crucial for robust generative modeling and transport accuracy (Alouadi et al., 27 Jan 2026).

4. Theoretical and Empirical Performance

LightSBB-M demonstrates the following performance characteristics:

Task	LightSBB-M 2-Wasserstein	Best SB Baseline	Relative Improvement
N→8-Gaussians	0.241±0.083	0.315	≈23% ↓
Moons→8-Gaussians	0.201±0.034	≥0.295	32% ↓
N→Moons	0.109±0.014	0.144	24% ↓

In high-dimensional image translation (FFHQ adult→child faces, $$dX_t = \alpha_t dt + \sigma_t dW_t,\quad X_0 \sim \mu_0,\ X_T \sim \mu_T$$1 latent),