---
title: 'MadSBM: Minimal-action Discrete Schrödinger Bridge'
url: https://www.emergentmind.com/topics/minimal-action-discrete-schrodinger-bridge-matching-madsbm
type: topic
---

# MadSBM: Minimal-action Discrete Schrödinger Bridge

Minimal-action discrete Schrödinger Bridge Matching (MadSBM) is a generative modeling framework that addresses the challenge of synthesizing discrete objects—specifically peptide sequences—by formulating generation as a minimal-action transport problem between a simple prior and the data distribution, where all trajectories traverse high-likelihood regions of the discrete state space. The method combines concepts from continuous-time Markov chains, Schrödinger bridge theory, and rate-based optimal control, yielding probabilistic generation processes that circumvent many difficulties of conventional discrete diffusion and flow models [2601.22408].

## 1. Discrete State Space and Markov Process Construction

MadSBM operates on the space of $L$-length peptides, each $x = (x_1, ..., x_L)$ where $x_i \in \mathcal{V}$, the alphabet of $20$ amino acids plus a mask token “$\sqcup$”. The underlying structure is an edit graph: edges $(x \to y)$ exist for all sequences $x, y$ differing at exactly one position. The dynamics over this space are modeled as a time-inhomogeneous continuous-time Markov chain (CTMC) with generator $Q^0(t)$.

The reference transition rate for an edit $(x\to y)$ at time $t$ and position $i$ is given by
$$
Q^0_{x \to y}(t) = (1-t)\cdot \mathrm{softmax}_v[f_\varphi(x)_i(v)]_{v = y_i}
$$
where $f_\varphi(x)$ are masked language modeling logits from a frozen, pre-trained protein language model (ESM-2). This structure encodes domain knowledge: transitions are driven by biologically plausible substitutions reflecting high-likelihood amino acid predictions.

## 2. Discrete Schrödinger Bridge and Minimal-Action Principle

The generative task is formulated as a minimal Kullback-Leibler (KL) divergence transport problem between endpoint marginals: the fully masked prior $\mu_0$ (all positions masked) and the data marginal $\mu_1$ (empirical peptide distribution). Specifically, one seeks a controlled path law $\mathbb{P}_Q$ over trajectories under controlled rate $Q(t)$ that solves
$$
\min_{Q} \ \mathrm{KL}(\mathbb{P}_Q \| \mathbb{P}_0) \quad \text{subject to} \quad (\mathbb{P}_Q)_0 = \mu_0, \ (\mathbb{P}_Q)_1 = \mu_1
$$
where $\mathbb{P}_0$ denotes the reference CTMC path law.

The path-space relative entropy expands (via discrete CTMC Girsanov theory) to a minimal-action functional:
$$
S[Q \| Q^0] = \int_0^1 \sum_{x, y} Q_{x \to y}(t) \log \frac{Q_{x \to y}(t)}{Q^0_{x \to y}(t)} - (Q_{x \to y}(t) - Q^0_{x \to y}(t)) \, dt
$$
This measures the entropic cost of deviating from the reference process, penalizing unnecessary or low-likelihood transitions.

## 3. Optimal Control Field and Schrödinger System

The solution to the minimal-action Schrödinger bridge has a Doob-$h$ transform structure: there exist strictly positive backward potentials $\psi_t(x)$ such that the optimal log-tilt parameter is $u_t^*(x \to y) = \log \psi_t(y) - \log \psi_t(x)$. The optimal control modifies the reference generator multiplicatively:
$$
Q^*_{x \to y}(t) = Q^0_{x \to y}(t) \cdot \exp\left(\log \psi_t(y) - \log \psi_t(x)\right)
$$
The backward potential $\psi_t(x)$ solves the backward Kolmogorov (adjoint) equation:
$$
\partial_t \psi_t(x) + \sum_{y \neq x} Q^0_{x \to y}(t) [\psi_t(y) - \psi_t(x)] = 0, \qquad \psi_1(x) = g(x)
$$
with $g(x)$ chosen so that final-time marginal aligns with the data.

## 4. Simulation-Free Training via Masking Interpolation

MadSBM eschews explicit simulation of CTMCs in training. Instead, it leverages the structure of peptide masking: an observed peptide $x_1$ is partially masked to $x_t$, where each site is masked independently with probability $1-t$. The model parameterizes $u_\theta(x \to y, t)$ and defines time-dependent logits as
$$
\ell_\theta(x_t, v) = u_\theta(x_t \to x_t^{(i) \leftarrow v}, t) + (1-t) f_\varphi(x_t)_i(v)
$$
for each position $i$ and token $v$. The one-step transition is trained to reconstruct the true token $x_1^{(i)}$ using a cross-entropy loss summed across masked sites:
$$
\mathcal{L}(\theta) = - \mathbb{E}_{t,x_t,x_1} \left[\sum_{i: x_t^{(i)} \neq x_1^{(i)}} \log p_\theta(x_1^{(i)} | x_t, t)\right]
$$
Minimization of $\mathcal{L}$ ensures $u_\theta$ converges to the optimal log-tilt up to an additive constant.

## 5. Rate-Based Generation and Jump Sampling Procedure

Generation from the model runs backward in time ($t=1 \to 0$) in $N$ discrete steps. At each step, for each sequence position, exit rates $r_i(v)$ are computed via the learned $u_\theta$ field. A stochastic jump process—parameterized by hyperparameters $(\lambda, \beta, \tau, p)$—decides whether to update each token. If updated, nucleus sampling with temperature is used to draw new tokens. This framework ensures sampled trajectories remain close to high-likelihood or biologically plausible sequence neighborhoods throughout generation [2601.22408].

## 6. Discrete Classifier Guidance

MadSBM introduces discrete classifier guidance for the first time in a Schrödinger bridge context. This mechanism allows generation to be steered toward target properties (e.g., functional affinity) by resampling candidate jumps based on a property classifier $C(x)$. The guided rate is
$$
Q_{\text{guided}}(x \to y, t) \propto Q_\theta(x \to y, t) \exp\bigl[\gamma \Delta \log C \bigr], \quad \Delta \log C \approx \log C(y) - \log C(x)
$$
In practice, M candidate successors are generated per position, scored, and resampled proportionally to the classifier output, implementing guidance through candidate selection rather than full re-tilting.

## 7. Context, Relation to Broader Discrete Schrödinger Bridge Paradigms, and Applications

MadSBM builds conceptually on the discrete Schrödinger bridge approach, which seeks minimal-action or entropy-regularized transport in discrete path space, generalizing frameworks previously developed for molecular graphs [2410.01500], categorical variables [2502.01416], and adversarially-trained generative models [2405.14449]. In contrast to existing discrete diffusion or flow-based methods—which either require simulating long chains through implausible intermediate states or rely on fixed corruption paths—MadSBM constructs transport processes explicitly biased to remain in high-likelihood sequence regions, drawing on pre-trained protein language models for biological prior structure.

MadSBM is explicitly developed for peptide sequence design, synthesizing sequences that interpolate between a fully masked prior and empirical data, and, via classifier guidance, can extend to functional optimization of therapeutic peptides. Empirical results and theoretical justification are provided in [2601.22408].

---

### Table: Key Formal Components of MadSBM

| Component                | Mathematical Description                                                | Biological/Algorithmic Role                 |
|--------------------------|------------------------------------------------------------------------|---------------------------------------------|
| Sequence Space           | $\mathcal{X} = \mathcal{V}^L$ (amino-acid strings + mask)              | Set of all peptides considered              |
| Reference Rates $Q^0$    | From $\mathrm{softmax}$ of ESM-2 logits at masked sites                | Biologically plausible edit dynamics        |
| Control Field $u_\theta$ | Time-dependent parametric tilt on transitions                          | Steers paths toward data distribution       |
| Loss $\mathcal{L}$       | Cross-entropy between predicted/true token at masked sites             | Learns optimal local update rates           |
| Guidance Mechanism       | Resampling using $C(x)$ as property classifier                         | Conditional generation for target function  |

MadSBM represents an integration of minimal-action discrete Schrödinger bridge formulation, domain-informed reference processes, efficient simulation-free training, and function-classifier guidance. This enables efficient, plausible peptide generation while traversing low-entropy regions of the discrete Markov path space [2601.22408].

Source: https://www.emergentmind.com/topics/minimal-action-discrete-schrodinger-bridge-matching-madsbm