Series-to-Series Diffusion Bridge Model (S²DBM)

• Series-to-Series Diffusion Bridge Model (S²DBM) is a generative framework that applies closed-form Gaussian bridges to transform paired time series across applications like speech enhancement, forecasting, and recommendations.
• It leverages continuous-time stochastic processes and generalizes score-based diffusion and Schrödinger bridge methods to support both probabilistic and deterministic mappings.
• The model employs diverse deep architectures and advanced training objectives, such as flow-matching and score-matching losses, to achieve state-of-the-art performance with enhanced computational efficiency.

The Series-to-Series Diffusion Bridge Model (S²DBM) is a class of generative models that interpolates between two paired time series via a continuous-time stochastic process parameterized as a diffusion bridge. S²DBM generalizes score-based diffusion, Schrödinger bridge, and flow-matching frameworks to model mappings between source and target time series with closed-form Gaussian bridges, enabling stochastic, deterministic, and conditionally guided transformations across a wide range of domains including speech enhancement, recommendation, and forecasting (Wang et al., 20 Feb 2026, Xie et al., 2024, Yang et al., 2024, Zhou et al., 2023).

1. Mathematical Foundations and Stochastic Bridge Construction

S²DBM posits a probability path $x_t$ indexed by $t \in [0,1]$ or $[0,T]$, constructed such that $x_{t=0}=x_0$ (source series) and $x_{t=1}=x_1$ (target series). The time-marginals of the bridge are parameterized as Gaussian distributions: $$p_t(x_t \mid x_0, x_1) = \mathcal{N}(x_t; \mu_t(x_0, x_1), \sigma_t^2 I)$$ with a linear mean schedule $\mu_t(x_0, x_1) = a_t x_0 + b_t x_1$, and variance $\sigma_t^2 = \sigma^2(t)$, controlled by smooth schedules $a_t, b_t, \sigma^2(t)$ satisfying boundary conditions $a_0=1, b_0=0, \sigma_0=0$ and $a_1=0, b_1=1, \sigma_1=0$ (Wang et al., 20 Feb 2026, Yang et al., 2024, Zhou et al., 2023).

Particular instantiations include Brownian bridges, Ornstein–Uhlenbeck bridges, Schrödinger bridges, and variants corresponding to different physical or optimal transport constraints (Wang et al., 20 Feb 2026, Zhou et al., 2023). For the Brownian bridge, the forward process is: $$x_t = \hat{\alpha}_t x_0 + \hat{\beta}_t \varepsilon + \hat{\gamma}_t, \qquad \varepsilon \sim \mathcal{N}(0, I)$$ where typically $\hat{\alpha}_0=0$, $\hat{\alpha}_T=1$, $\hat{\gamma}_t=1-\hat{\alpha}_t$, $$\hat{\beta}_t = \sqrt{2 \hat{\alpha}_t (1-\hat{\alpha}_t)}$$ (Yang et al., 2024).

The forward diffusion-bridge SDE adopts the Doob h-transform: $$dX_t = f(X_t, t)\,dt + g(t)\,dW_t + g^2(t)\nabla_x \log p(X_T = y | X_t = x)\,dt$$ with the guiding drift ensuring the path reaches $x_1$ at the terminal time (Zhou et al., 2023, Xie et al., 2024).

2. Reverse-Time SDEs, ODEs, and Inference

The generative process in S²DBM is realized by integrating the reverse-time SDE or its ODE equivalent. For a well-trained bridge, the reverse SDE for sampling from $x_1$ back to $x_0$ is: $$dx_t = [\kappa_t^- x_t + (a'_t - a_t \kappa_t^-) x_0 + (b'_t - b_t \kappa_t^-) x_1] dt + g_t d\bar{W}_t$$ where $$\kappa_t^\pm = \sigma'(t)/\sigma(t) \mp g_t^2/(2\sigma(t)^2)$$, and $g_t$ the diffusion coefficient (Wang et al., 20 Feb 2026). The bridge score is $$\nabla_{x_t} \log p_t(x_t|x_0, x_1) = -(x_t - (a_t x_0 + b_t x_1))/\sigma_t^2$$ (Wang et al., 20 Feb 2026, Zhou et al., 2023).

A deterministic sampling algorithm (probability-flow ODE) can be constructed to yield stable point predictions: $dx_t = (\sigma'/\sigma)x_t + m_t x_0 + n_t x_1\,dt$ with $m_t = a'_t - a_t \sigma'/\sigma$, $n_t = b'_t - b_t \sigma'/\sigma$ (Wang et al., 20 Feb 2026).

For time series forecasting, the deterministic update for the linear bridge can be written as: $$x_{t-1} = \kappa_t x_t + \lambda_t \hat{x}_0 + \zeta_t \bar{x}$$ where $\hat{x}_0$ is the network prediction, and $(\kappa_t, \lambda_t, \zeta_t)$ are closed-form bridge coefficients (Yang et al., 2024).

3. Training Objectives and Loss Functions

Training objectives are derived from the marginal statistics of the diffusion bridge. Primary approaches include:

• Flow-matching Loss: For vector field prediction $u_\theta(x_t, x_1, t)$,

$$\mathcal{L}_\mathrm{FM} = \mathbb{E}_{x_0, x_1, t} \|u_\theta(x_t, x_1, t) - [\sigma'/\sigma(x_t - \mu_t) + \mu_t']\|^2$$

(Wang et al., 20 Feb 2026).

• Score-matching / Data-prediction Loss: For data regression or denoising,

$$\mathcal{L}_\mathrm{pred} = \mathbb{E}_{x_0, x_1, t} \|s_\theta(x_t, x_1, t) - x_0\|^2$$

augmented by spectral/time-domain losses in applications such as speech enhancement (Wang et al., 20 Feb 2026).

• Classification or Cross-Entropy Loss: In sequential recommendation, explicitly supervising the predicted embedding to match the target,

$$\mathcal{L}_\mathrm{CE} = -\frac{1}{|\mathcal{U}|} \sum_{u \in \mathcal{U}} \log \frac{\exp(\hat{e}_n^\top e_n)}{\sum_{v \in \mathcal{I}} \exp(\hat{e}_n^\top e_v)}$$

where $\hat{e}_n$ is the reconstructed embedding (Xie et al., 2024).

For forecasting, $\ell_2$ reconstruction losses are used, often formulated over a "label-length" window incorporating history and future (Yang et al., 2024).

4. Model Architectures and Conditioning Strategies

S²DBM accommodates a variety of deep architectures, as long as the function mapping $(x_t, x_1, t)$ or $(x_t, c, t)$ to $x_0$ is expressive. Notable designs include:

• TF-GridNet: Five-block architecture using complex-STFT, time-frequency fusion, sub-band convolutions, LSTMs, and skip connections (Wang et al., 20 Feb 2026).
• Transformer-U-Net Hybrids: Stacking transformer layers with U-Net structures for time series, removing masking when bridging between series (Yang et al., 2024).
• Temporal Convolutional Networks (TCN): Residual 1D convolutions plus time embeddings (Zhou et al., 2023).
• FiLM/Conditioning: Feature-wise affine modulation or concatenation for conditioning on endpoints/history (Wang et al., 20 Feb 2026, Zhou et al., 2023).

In sequential recommendation (e.g., SdifRec), Transformer encoders produce user state vectors, embeddings are clustered for collaborative conditioning, and classifier-free guidance is applied to incorporate user-cluster information in the conditional bridge (Xie et al., 2024).

5. Sampling, Computational Properties, and Efficiency

Sampling is accomplished by discretizing $t$ and iteratively applying the reverse ODE or SDE, with coefficients and network calls computed at each step. Pseudocode for a typical deterministic ODE sampler is:

for n = N to 1: t_n = n/N; t_{n-1} = (n-1)/N x0_hat = s_theta(x, x1, t_n) # Compute alpha, beta, gamma coefficients x = alpha * x + sigma(t_{n-1}) * (beta * x0_hat + gamma * x1)

(Wang et al., 20 Feb 2026). This approach enables fast, low-NFE (number of function evaluation) inference and, in speech enhancement, achieves >10x reduction in both parameter count and compute relative to classic Schrödinger Bridge/score-based models (Wang et al., 20 Feb 2026).

For forecasting, deterministic (DDIM-style) updates correspond to $s=0$ bridge variance, generating smooth, non-oscillatory predictions (Yang et al., 2024). The stochastic case ($s>0$) enables full probabilistic generation and uncertainty quantification.

6. Applications, Generalization, and Empirical Performance

S²DBM has demonstrated efficacy in a range of domains:

• Speech enhancement and denoising: Outperforming flow/diffusion baselines on denoising and dereverberation benchmarks with fewer parameters and reduced computation (Wang et al., 20 Feb 2026).
• Time series forecasting: Superior point forecasting (ranked first/second on the majority of tasks) and competitive probabilistic forecasting (CRPS on par with TMDM and CSDI) (Yang et al., 2024). Deterministic samples are smooth and free from frame-level oscillation observed in other diffusion models.
• Sequential recommendation: SdifRec and con-SdifRec enable conditional user-aware recommendations, unifying user state and collaborative priors via tractable Schrödinger bridges (Xie et al., 2024).
• General series translation: Theoretical and empirical framework extends to finance, medical, and video domains given paired series and appropriate scheduling (Wang et al., 20 Feb 2026, Zhou et al., 2023).

A table of select benchmark results from (Yang et al., 2024):

Application	Metric	S²DBM (Point)	S²DBM (Prob)	Baseline (Best Other)
Weather L=96	MSE	0.397	N/A	0.407 (TimeDiff)
ETTh1 L=96	CRPS_sum	-	on par	on par (TMDM/CSDI)

7. Limitations and Prospects

S²DBM unifies and generalizes bridge-based generative modeling, but faces intrinsic limitations:

• The upper bound of achievable performance is set by the expressiveness of the prediction network; the bridge’s output is a weighted sum of $s_\theta$ predictions across steps, with the final step dominating—this is formalized as "predictive equivalence" (Wang et al., 20 Feb 2026). The framework cannot surpass discriminative models with the same architecture.
• Hand-designed bridge schedules may not optimally exploit data structure; learning schedules via adaptive Schrödinger bridges could offer improvements (Yang et al., 2024).
• Conditioning via simple linear maps may fail for rich exogenous covariates; more sophisticated encoders or sequence state models should be explored (Yang et al., 2024).
• Generalization to irregularly sampled or event-based series is an open problem, as is the integration of nontrivial auxiliary data and hybrid architectures (Yang et al., 2024).

S²DBM establishes a closed-form, flexible, and computationally efficient foundation for series-to-series generative modeling, supporting deterministic, probabilistic, and cross-domain translation tasks under a unified diffusion bridge formalism (Wang et al., 20 Feb 2026, Yang et al., 2024, Zhou et al., 2023, Xie et al., 2024).