Continuous-Time Evidence Lower Bound

• Continuous-Time ELBO is a variational lower bound for continuous-time SDE models that integrates data likelihoods with stochastic optimal control for irregular time series.
• It employs Doob’s h-transform and neural control parameterizations to derive tractable approximations via a stochastic optimal control framework.
• The approach enables efficient learning and simulation-free inference using piecewise linear drift approximations and modern network architectures.

The continuous-time Evidence Lower Bound (ELBO) is a variational lower bound formulated for probabilistic models that evolve according to continuous-time latent state-space dynamics, notably those driven by stochastic differential equations (SDEs). It provides a foundation for scalable inference and learning in irregularly observed time series, enabling the integration of data likelihoods and pathwise regularization. The continuous-time ELBO considered here arises from a stochastic optimal control (SOC) perspective, establishing a rigorous connection between Doob’s $h$-transform, Feynman–Kac path measures, and amortized variational inference with neural control parameterizations (Park et al., 2024).

1. Feynman–Kac Path Measures and the Posterior in State-space SDEs

Let $(X_t)_{t\in[0,T]}$ denote a $\mathbb{R}^d$-valued diffusion evolving under an SDE

$dX_t = b(t,X_t)\,dt + dW_t,\quad X_0\sim\mu_0,$

where $b$ is the drift and $W_t$ standard Brownian motion. Observations $\{Y_{t_i}\}_{i=1}^k$ are made at irregular time-stamps $0=t_0<t_1<\cdots<t_k=T$, each with likelihood $g_i(y_{t_i}|X_{t_i})$. The joint path-observation posterior, or the Feynman–Kac model, is

$$\mathbb{P}^*\bigl(dX_{0:T}\mid Y_{t_1}=y_{t_1},\dots,Y_{t_k}=y_{t_k}\bigr) = \frac1{Z(y_{1:k})}\,\prod_{i=1}^k g_i\bigl(y_{t_i}\mid X_{t_i}\bigr)\;\mathbb{P}(dX_{0:T}),$$

with $Z=\mathbb{E}_{\mathbb{P}}\bigl[\prod_i g_i\bigr]$ the marginal likelihood. Normalized potentials are defined as $f_i(x) = g_i(y_{t_i}|x) / L_i$, $L_i = \int g_i(y_{t_i}|x)\mathbb{P}(dx)$, with the property $\mathbb{E}_{\mathbb{P}}[\prod_i f_i]=1$.

A multi-marginal Doob's $h$-transform yields the posterior dynamics as

$$dX^*_t = \big[b(t,X^*_t) + \nabla_x \log h(t,X^*_t)\big]\,dt + dW_t,$$

where $h(t,x)$ is a “backward survival” function propagating posterior information. This SDE generates exactly the posterior law $\mathbb{P}^*$ with the correct initial condition $X^*_0 \sim \mu^*_0(dx)\propto h_1(0,x)\,\mu_0(dx)$.

2. Variational Family, Amortization, and Auxiliary Variables

The intractability of $\nabla_x \log h$ motivates a tractable variational family: controlled SDEs parameterized by neural controls,

$$dX^\alpha_t = [b(t,X^\alpha_t) + \alpha(t,X^\alpha_t)]\,dt + dW_t, \quad X^\alpha_0\sim\mu_0,$$

with induced path law $\mathbb{P}^\alpha$. In amortized inference, per-observation latent variables $y_{t_i}$ are encoded via $q_\phi(y_{t_i}|o_{t_i})$ and decoded with $p_\psi(o_{t_i}|y_{t_i})$. The control $\alpha_\theta(t,x)$ is parameterized to depend on latent histories or the full latent collection $\{y_{t_j}\}$, typically via a transformer or RNN.

3. Stochastic Optimal Control Formulation and Dynamic Programming

SOC theory provides a variational foundation for continuous-time inference. Define the cost functional

$$J(t,x;\alpha)= \mathbb{E}_{\mathbb{P}^\alpha|X^\alpha_t=x}\left[ \int_t^T \tfrac12\,\|\alpha(s,X^\alpha_s)\|^2\,ds - \sum_{i:\,t_i\ge t}\log f_i\bigl(X^\alpha_{t_i}\bigr) \right].$$

The value function $V(t,x) = \inf_\alpha J(t,x;\alpha)$ satisfies, on subintervals $[t_{i-1},t_i)$, the Hamilton–Jacobi–Bellman (HJB) PDE

$$\partial_t V + A_t V + \min_{\alpha}\{\nabla_x V^\top \alpha + \tfrac12\|\alpha\|^2\} = 0; \qquad V(t_i^-,x) = -\log f_i(x) + V(t_i^+,x),$$

where $A_t$ is the infinitesimal generator of the prior SDE. The minimizer is $\alpha^*(t,x)=\nabla_x V(t,x)$. The Hopf–Cole transform $V = -\log h$ linearizes the HJB, relating $h$ to the solution of backward Kolmogorov equations and recovering the Doob control.

4. Variational Bound and Continuous-Time ELBO Construction

Using Girsanov’s theorem,

$$\frac{d\mathbb{P}^\alpha}{d\mathbb{P}}(X_{0:T}) = \exp\left\{\int_0^T \alpha^\top\,dW - \tfrac12\int_0^T\|\alpha\|^2\,ds\right\},$$

the KL divergence between the variational path law and the path posterior is

$$D_{KL}(\mathbb{P}^\alpha\|\mathbb{P}^*) = \mathbb{E}_{\mathbb{P}^\alpha}\left[ \int_0^T\tfrac12\|\alpha\|^2\,ds - \sum_i\log f_i(X_{t_i}) \right],$$

excluding a vanishing initial-law term at optimum. Setting

$$\mathcal{L}(\alpha) := \mathbb{E}_{\mathbb{P}^\alpha}\left[ \int_0^T\tfrac12\|\alpha\|^2\,ds - \sum_{i=1}^k\log g_i(Y_{t_i}|X^\alpha_{t_i}) \right],$$

yields a tight variational characterization: $$D_{KL}(\mathbb{P}^\alpha\|\mathbb{P}^*) = \mathcal{L}(\alpha) + \log Z \geq 0$$, so that $-\log Z \leq -\mathcal{L}(\alpha)$. Minimizing $\mathcal{L}(\alpha)$ corresponds to optimal control and tightens the ELBO.

Combining the path-space bound with the standard VAE objective over latent variables, the negative ELBO is given by

$$\text{ELBO}(\psi,\phi,\theta) = \mathbb{E}_{q_\phi}\left[ \sum_i\log p_\psi(o_{t_i}\mid y_{t_i}) - \mathcal{L}(\alpha_\theta) \right],$$

where $\alpha_\theta$ depends on encoded latents $y_{0:k}$. The key object is

$$\mathcal{L}(\alpha_\theta) = \mathbb{E}_{\mathbb{P}^{\alpha_\theta}}\left[ \int_0^T\tfrac12\|\alpha_\theta(t,X_t)\|^2\,dt - \sum_{i=1}^k\log g_i(y_{t_i}|X_{t_i}) \right].$$

This structure enables end-to-end training by maximizing $\text{ELBO}(\psi,\phi,\theta)$ across all variational and generative parameters.

5. Assumptions, Practical Strategies, and Simulation-free ELBO

All drifts $b(t,x)$ and controls $\alpha(t,x)$ are assumed Lipschitz with linear growth to guarantee strong solution existence for SDEs. The Hopf–Cole transform and use of Girsanov require Novikov-type moment conditions for validity. In practical implementation, the optimal drift $\nabla_x\log h$ is replaced by a parameteric neural control $\alpha_\theta$ optimized via the ELBO.

Costly simulation of pathwise SDEs and backpropagation through continuous-time integrators can be circumvented by a piecewise locally linear drift ansatz: $dX_t = (-A_iX_t + \alpha_i)\,dt + dW_t$ for $t\in[t_{i-1},t_i)$, so that state marginals evolve as Gaussian processes with closed-form updates. This approximation enables efficient, simulation-free parallel ELBO computation. Amortized control construction is performed by modern attention-based networks (e.g., transformers) operating over $\{y_{t_i}\}$.

6. Summary and Practical Implementation

The continuous-time ELBO, as realized in the "Amortized Control of Continuous State Space Feynman-Kac Model for Irregular Time Series," synthesizes stochastic optimal control, Feynman–Kac representations, and deep amortized inference, resulting in an end-to-end objective: $$\text{ELBO}(\psi,\phi,\theta) = \mathbb{E}_{q_\phi(y_{1:k}\mid o_{1:k})}\Biggl[ \sum_{i=1}^k\log p_\psi(o_{t_i}\mid y_{t_i}) - \mathbb{E}_{\mathbb{P}^{\alpha_\theta}}\left( \int_0^T\tfrac12\|\alpha_\theta\|^2 - \sum_{i=1}^k\log g_i(y_{t_i}\mid X_{t_i}) \right) \Biggr].$$ All nested expectations are tractable using 1. Sampling of $y_{t_i}$ from the encoder, 2. Neural construction of $\alpha_\theta$ via sequence models, 3. Either numerical SDE simulation or closed-form marginal propagations in the piecewise linear case, 4. Likelihood decoding via $p_\psi$. The formulation provides a theoretically grounded and computationally practical route to sequential data assimilation in continuous time, particularly for irregular time series (Park et al., 2024).