---
title: Sequential Controlled Langevin Diffusions
url: https://www.emergentmind.com/topics/sequential-controlled-langevin-diffusions-scld
type: topic
---

# Sequential Controlled Langevin Diffusions

Sequential Controlled Langevin Diffusions (SCLD) are a class of sampling and control constructions in which a time-inhomogeneous Langevin diffusion is used to transport particles from an easy prior toward a target distribution, while sequential corrections such as importance weighting, resampling, optional Markov chain Monte Carlo refinement, or state-dependent guidance are applied along the trajectory. In the explicit formulation introduced by "Sequential Controlled Langevin Diffusions" [2412.07081], SCLD combines Sequential Monte Carlo (SMC) and diffusion-based samplers in a continuous-time, path-space framework. Closely related later works interpret diffusion-path samplers and general annealed Langevin schemes through the same lens [2601.21951, 2601.22349], and a related line on controllable generative diffusion recasts classifier-guided reverse diffusion as sequential control of a Langevin-type process under a KL-regularized objective, although it does not use the term SCLD explicitly [2505.21666].

## 1. Definition and conceptual placement

The core sampling problem is to draw from a target distribution
\[
\pi(x)\propto \tilde{\pi}(x),
\]
or, in an equivalent potential form used in related analyses,
\[
\pi(x)\propto e^{-V_\pi(x)},
\]
when the normalizing constant is unknown and samples from the target are unavailable [2412.07081, 2601.21951]. The explicit SCLD formulation positions itself between two established paradigms. Classical SMC or annealed importance sampling transports a population of particles through intermediate tempered distributions using prescribed Markov kernels and resampling; it is asymptotically exact and robust, but its hand-crafted transitions can mix slowly. Diffusion-based samplers instead learn a drift for an SDE that transports an easy prior to the target; they are flexible and target-adaptive, but can be expensive or unstable to train, and can be prone to mode collapse [2412.07081].

SCLD is defined by combining these two viewpoints in continuous time and on path space. The method uses a controlled Langevin diffusion as the forward transport mechanism, splits trajectories into subtrajectories, computes path-space importance weights through a Girsanov-type Radon–Nikodym derivative, and interleaves resampling and optional MCMC steps at chosen times [2412.07081]. In that formulation, SCLD is an SMC algorithm whose Markov kernels are learned controlled Langevin diffusions, trained end-to-end with a path-space variational loss.

A longer-range theoretical precursor appears in the variational characterization of Langevin–Smoluchowski diffusions, where time reversal followed by drift control yields a sequence of stochastic control problems whose values decrease along the relative entropy to the Gibbs measure [2010.04847]. This earlier result does not define SCLD algorithmically, but it places sequential drift control of Langevin dynamics on a path-space, entropy-based foundation.

## 2. Controlled Langevin formulation on path space

The canonical forward SDE in the explicit SCLD framework is
\[
dX_t^u = u(X_t^u,t)\,dt + \sigma(t)\,dW_t,\qquad X_0^u\sim p_{\text{prior}},
\]
with control \(u(x,t)\in\mathbb{R}^d\) and scalar diffusion coefficient \(\sigma(t)\) [2412.07081]. The method uses a Langevin-like parameterization
\[
u(x,t) = \sigma(t)^2\,\widetilde{u}_\theta(x,t) \;+\; \frac{\sigma(t)^2}{2}\,\nabla\log \pi(x,t),
\]
where \(\widetilde{u}_\theta\) is a neural network and \(\pi(\cdot,t)\) is a prescribed path of intermediate distributions,
\[
\pi(x,t) \propto p_{\text{prior}}(x)^{1-\beta(t)} \,\rho_{\text{target}}(x)^{\beta(t)},\quad \beta(0)=0,\ \beta(T)=1.
\]
If \(\widetilde{u}_\theta\equiv 0\), this reduces to annealed Langevin dynamics along the prescribed intermediates [2412.07081].

The corresponding reverse-time construction is tied to the same path \(\pi(\cdot,t)\). With
\[
v(x,t) = u(x,t) - \sigma(t)^2 \nabla\log \pi(x,t),
\]
the reverse SDE is
\[
dY_t^u = (u - \sigma^2 \nabla\log\pi)(Y_t^u,t)\,dt
 + \sigma(t)\,\cev{d}W_t,\qquad Y_T^u \sim p_{\text{target}}.
\]
The forward and backward laws are measures on path space, and the central object is the Radon–Nikodym derivative between them on a subinterval \([s,t]\):
\[
\begin{aligned}
w_{[s,t]}(X) &= \frac{d\cev{P}_{[s,t]}}{d\vec{P}_{[s,t]}}(X) \\
&= \frac{\pi(X_t,t)}{\pi(X_s,s)}\, \exp\Bigg( \int_s^t \frac{\|u\|^2 - \|u - \sigma^2\nabla\log\pi\|^2}{2\sigma^2}(X_\tau,\tau)\,d\tau \\
&\quad + \int_s^t \frac{u - \sigma^2\nabla\log\pi}{\sigma^2}(X_\tau,\tau)\cdot \cev{d}X_\tau
 - \int_s^t \frac{u}{\sigma^2}(X_\tau,\tau)\cdot dX_\tau \Bigg).
\end{aligned}
\]
On a full trajectory,
\[
w(X) = \frac{d\cev{P}}{d\vec{P}}(X) = \prod_{n=1}^N w_{[t_{n-1},t_n]}(X).
\]
This factorization is the continuous-time analogue of incremental importance weights in SMC [2412.07081].

Related work broadens the same controlled-diffusion perspective. One line studies the time-inhomogeneous Langevin diffusion
\[
dX_t = -\nabla U_{\tau(t)}(X_t)\,dt + \sqrt{2}\,dW_t,
\]
where the control is an annealing schedule \(\tau(t)\) selecting an intermediate target \(\pi_{\tau(t)}\) at each time [2601.22349]. Another line defines a diffusion path \(\mu_t\) between a Gaussian base and the target, and runs diffusion-annealed Langevin dynamics
\[
dX_t = \nabla\log\hat\mu_t(X_t)\,dt + \sqrt{2}\,dW_t,\qquad \hat\mu_t=\mu_{t/T},
\]
so that the controlled drift is the time-varying score of the path [2601.21951].

## 3. Sequential structure: subtrajectories, weights, and resampling

The sequential aspect of SCLD is explicit. A time grid
\[
0 = t_0 < t_1 < \dots < t_N = T
\]
partitions the diffusion into subtrajectories. On each interval \([t_{n-1},t_n]\), the sampler simulates the forward SDE as a Markov kernel, computes the path-space subtrajectory weight \(w_{[t_{n-1},t_n]}\), updates cumulative weights via
\[
w_n^{(k)} = w_{n-1}^{(k)}\,w_{[t_{n-1},t_n]}^{(k)},
\]
and optionally resamples particles and applies an MCMC refinement kernel invariant for \(\pi(\cdot,t_n)\) [2412.07081].

The practical implementation uses Euler–Maruyama. If \(L\) discretization steps are used within each subtrajectory and \(h=\tau/L\), then
\[
\widehat{X}^u_{i} = \widehat{X}^u_{i-1} + u(\widehat{X}^u_{i-1}, t_{i-1})\,h + \sigma(t_{i-1})\,\sqrt{h}\,\xi_i,\qquad \xi_i\sim\mathcal{N}(0,I).
\]
The forward and backward one-step transition densities are Gaussian, and the discrete subtrajectory weight is approximated by
\[
w_{[t_{n-1},t_n]}(X) \approx \frac{\pi(X_{t_n},t_n)}{\pi(X_{t_{n-1}},t_{n-1})} \prod_{i=(n-1)L+1}^{nL} \frac{\cev{p}_{t_{i-1}|t_i}(X_{t_{i-1}}\mid X_{t_i})}
     {\vec{p}_{t_i|t_{i-1}}(X_{t_i}\mid X_{t_{i-1}})}.
\]
Effective sample size is monitored through
\[
\mathrm{ESS}_n
= \frac{ \left(\sum_{k=1}^K w_n^{(k)}\right)^2}{\sum_{k=1}^K (w_n^{(k)})^2},
\]
and if \(\mathrm{ESS}_n < \alpha K\) with \(\alpha=0.3\), multinomial resampling is performed and the weights are reset to uniform. After resampling, one Hamiltonian Monte Carlo step with 10 leapfrog updates can be applied as an additional refinement [2412.07081].

This yields a standard SMC structure with unusual forward kernels: the transport between resampling times is no longer a fixed MCMC transition but a learned stochastic flow. The same sequential viewpoint appears in diffusion-path samplers based on auxiliary SMC. There, at each main Langevin step, auxiliary particles approximate a conditional distribution \(\varrho_{t,x}\), and their weighted empirical average yields the score estimate needed for the next controlled Langevin update [2601.21951]. This suggests that “sequential” in SCLD can refer either to particle-level importance-resampling across path segments or to online estimation of the feedback drift along a prescribed path.

## 4. Training objectives, control learning, and theoretical guarantees

In the explicit SCLD sampler, three objects are learned jointly: the control \(u_\theta\) through \(\widetilde{u}_\theta(x,t)\), the annealing schedule \(\beta_\theta(t)\), and the Gaussian prior
\[
p_{\text{prior},\theta} = \mathcal{N}(\mu_\theta,\operatorname{diag}(\exp(2\ell_\theta))).
\]
The annealing schedule is parameterized on the discrete grid by
\[
\beta_\theta(jh) = \frac{\sum_{i=1}^{j} \text{softplus}(\theta_i)}{\sum_{i=1}^{NL} \text{softplus}(\theta_i)}.
\]
Training minimizes a sum of subtrajectory divergences,
\[
\mathcal{L}(u) = \sum_{n=1}^N D\Big( \vec{P}^{u,\pi_{n-1}}_{[t_{n-1},t_n]} \;\Big\|\; \cev{P}^{u,\pi_n}_{[t_{n-1},t_n]} \Big),
\]
with \(\pi_n=\pi(\cdot,t_n)\) [2412.07081].

Two divergence choices are discussed. A KL-based objective,
\[
D_{\text{KL}}(\vec{P}\,\|\,\cev{P}) = -\mathbb{E}_{X\sim \vec{P}}[\log w],
\]
is natural but problematic in high dimension when importance sampling is needed. The default choice is the log-variance divergence,
\[
D_{\text{LV}}^{Q}\Big(\vec{P}^{u,\pi_{n-1}}_{[t_{n-1},t_n]}
  \,\big\|\,\cev{P}^{u,\pi_n}_{[t_{n-1},t_n]}
  \Big)
  = \operatorname{Var}_{X\sim Q}
  \big[ \log w_{[t_{n-1},t_n]}(X) \big],
\]
estimated empirically by
\[
\widehat{D}_{\text{LV},n} = \frac{1}{K}\sum_{k=1}^K \big( \log w_{[t_{n-1},t_n]}^{(k)} - \bar{\ell}_n \big)^2,\quad \bar{\ell}_n = \frac{1}{K}\sum_{k=1}^K \log w_{[t_{n-1},t_n]}^{(k)},
\]
and summed over subtrajectories:
\[
\mathcal{L}(\theta) = \sum_{n=1}^N \widehat{D}_{\text{LV},n}.
\]
Trajectories are detached from the computation graph, and gradients are taken only through \(\log w\), not through the SDE integrator [2412.07081].

The theory attached to this construction has several layers. The path-space identity implies unbiased importance sampling with exact weights, while self-normalized estimators are biased but consistent as the number of particles \(K\to\infty\). If all subtrajectory divergences vanish, then the global divergence between forward and backward path measures is zero. For KL training with importance sampling, the relative error of the Monte Carlo estimator grows exponentially with dimension through a \(\chi^2\)-divergence bound, whereas for the LV divergence one has
\[
D_{\text{LV}}^{Q^{\otimes I}}(\vec{P}^{\otimes I}\,\|\,\cev{P}^{\otimes I}) = I D_{\text{LV}}^{Q}(\vec{P}\,\|\,\cev{P}),
\]
and the variance of its Monte Carlo estimator is \(O(I^2/K)\) [2412.07081].

A complementary convergence theory applies to the broader class of time-inhomogeneous Langevin diffusions
\[
dX_t = -\nabla U_{\tau(t)}(X_t)\,dt + \sqrt{2}\,dW_t.
\]
Under uniform smoothness, dissipativity, and time-regularity assumptions on the intermediate potentials, the forward KL to the final target satisfies
\[
\begin{aligned}
\mathrm{KL}(\mu_t \,\Vert\, \pi) \le\;& c\,\tau(t) \\
& + \exp\!\Bigl(-\int_0^t \frac{4}{LSI(s)}\,ds\Bigr)\mathrm{KL}(\mu_0 \,\Vert\, \pi_{\tau(0)}) \\
& + c\int_0^t\exp\!\Bigl(-\int_s^t\frac{4}{LSI(\sigma)}\,d\sigma\Bigr) |\dot\tau(s)|\,ds,
\end{aligned}
\]
and if \(\tau(t)\to0\) with \(\dot\tau(t)\le0\), then
\[
\lim_{t\to\infty}\mathrm{KL}(\mu_t \,\Vert\, \pi) = 0.
\]
For the Euler–Maruyama discretization,
\[
X_{k+1} = X_k - h_k\nabla U_{\tau(t_k)}(X_k) + \sqrt{2h_k}\,\xi_k,
\]
if \(\sum_k h_k=\infty\) and \(\sum_k h_k^2<\infty\), then
\[
\lim_{k\to\infty} \mathrm{KL}(\hat\mu_k \,\Vert\, \pi) = 0
\]
under the corresponding step-size and regularity conditions [2601.22349]. These results make the control-versus-tracking tradeoff explicit: better intermediate log-Sobolev constants improve mixing, while rapid schedule changes contribute tracking error.

## 5. Related formulations and neighboring uses of the SCLD viewpoint

One major adjacent development is the diffusion-path sampler implemented via auxiliary SMC [2601.21951]. There the path between a Gaussian base \(\nu\) and the target \(\pi\) is defined by
\[
X_t = \sqrt{1-\lambda_t}\,Z + \sqrt{\lambda_t}\,X,\qquad Z\sim\nu,\ X\sim\pi,
\]
with law \(\mu_t\). The controlled dynamics are diffusion-annealed Langevin dynamics,
\[
dX_t = \nabla\log\hat\mu_t(X_t)\,dt + \sqrt{2}\,dW_t,
\]
and the score \(\nabla\log\mu_t(x)\) is estimated from an auxiliary conditional posterior
\[
\varrho_{t,x}(y) \propto \nu\!\bigg(\frac{x-\sqrt{\lambda_t}y}{\sqrt{1-\lambda_t}}\bigg)\pi(y).
\]
The resulting estimator uses SMC particles targeting \(\varrho_{t,x}\) and score identities such as
\[
\nabla\log\mu_t(x)
  = \mathbb{E}_{\varrho_{t,x}}\!\left[
      \frac{1}{\sqrt{1-\lambda_t}}\,
      \nabla\log\nu\!\left(
        \frac{x-\sqrt{\lambda_t}Y}{\sqrt{1-\lambda_t}}
      \right)
  \right]
\]
and
\[
\nabla\log\mu_t(x)
  = \mathbb{E}_{\varrho_{t,x}}\!\left[
      \frac{1}{\sqrt{\lambda_t}}\,
      \nabla\log\pi(Y)
  \right].
\]
Control variates are introduced through matrices \(\mathbf{A}_t\) that combine these identities. The scalar optimum is
\[
\alpha_t^\ast = \frac{\frac{1}{\lambda_t}\,\operatorname{Var}_\pi[\nabla\log\pi(X)]}
 {\frac{1}{1-\lambda_t}\,\operatorname{Var}_\nu[\nabla\log\nu(X)]
  + \frac{1}{\lambda_t}\,\operatorname{Var}_\pi[\nabla\log\pi(X)]},
\]
and the matrix optimum is
\[
\mathbf{A}_t^\ast
= \frac{1}{\lambda_t}\, \mathcal{I}_\pi \left( \frac{1}{1-\lambda_t}\mathcal{I}_\nu + \frac{1}{\lambda_t}\mathcal{I}_\pi \right)^{-1},
\]
where \(\mathcal{I}_p=\operatorname{Cov}_p[\nabla\log p(X)]\). The score estimation error bound
\[
\mathbb{E}\big[\| \nabla\log\hat\mu_{t_k}(X_k)-S_k^N(X_k) \|^2\big] \le 2\,\delta(k,N)^2 \left(
  (\alpha_k^\ast)^2\frac{\lambda_k}{(1-\lambda_k)^2}
 + (1-\alpha_k^\ast)^2\frac{L_\pi^2}{\lambda_k} \right)
\]
feeds into a final KL guarantee for the approximate sampler [2601.21951].

A different but related interpretation appears in controllable generative diffusion. The KL-regularized objective
\[
\max_{p} \; \mathbb{E}_{x \sim p}[r(x)] - \frac{1}{\eta} \mathrm{KL}(p \,\|\, q_0)
\]
has optimal solution
\[
p^{\star}(x)
 = \frac{1}{Z} q_0(x) \exp(\eta r(x)).
\]
The reverse SDE is controlled by a guidance field \(f(x,t)\),
\[
d x_t = \Big[-h(x_t,T-t)
 + g^2(T-t)\big(\nabla \log q_{T-t}(x_t) + f(x_t,t)\big)\Big]dt + g(T-t)\,dw_t,
\]
and the optimal guidance is
\[
f^{\star}(x_t,t)
 := \nabla_{x_t} \log \mathbb{E}_{r\sim R^{\text{prior}}(\cdot\mid x_t,t)}[\exp(\eta r)].
\]
The SLCD procedure learns an approximation to the reward distribution with DAgger-style online data aggregation and obtains a no-regret convergence guarantee in KL to the optimal tilted target. This work does not use the label SCLD explicitly, but it treats the reverse-time diffusion as a sequentially controlled Langevin-type process whose control is applied at every diffusion step [2505.21666].

## 6. Empirical behavior, limitations, and significance

The explicit SCLD sampler was evaluated on 11 targets spanning Bayesian statistics and synthetic multimodal or high-dimensional distributions [2412.07081]. On the ELBO tasks, it achieves the best ELBO on Seeds, Sonar, Credit, and Brownian, and is competitive on LGCP. On the Sinkhorn tasks, the reported distances include GMM40 at approximately \(3.8\text{k}\) versus the best baseline at approximately \(10\text{–}22\text{k}\), MoS at approximately \(656\) versus the best baseline at approximately \(1477\), and Robot1/4 at approximately \(0.3\text{–}0.4\) versus baselines at least \(1.5\text{–}4\). The same study reports that SCLD uses many fewer training iterations than CMCD, with examples of \(3\text{k}\) versus \(40\text{k}\) gradient steps, while often outperforming or matching it, and the abstract states that improved performance is reached on multiple benchmark problems, in many cases using only \(10\%\) of the training budget of previous diffusion-based samplers [2412.07081].

The reported ablations show that increasing the number of SMC subtrajectories is typically beneficial in training and evaluation, that even adding SMC steps only at test time improves diffusion-only methods, that removing MCMC or resampling degrades performance though the method remains competitive, and that replay buffers improve training stability and sample quality, especially in high dimension [2412.07081]. In the diffusion-path formulation, control variate schedules reduce score-estimation mean squared error across the path, improving stability for anisotropic targets, and tempering of auxiliary posteriors helps exploration when initialization is weak [2601.21951]. In the forward-KL analysis of annealed Langevin diffusions, convolutional and Moreau-envelope paths dominate plain ULA, geometric tempering, and dilation on multimodal Gaussian mixtures, with faster KL decay and better mode coverage [2601.22349].

The limitations are correspondingly structural. The explicit SCLD sampler requires gradients of the prior and unnormalized target, simulating many particles with fine time discretization can be computationally heavy, and tuning of noise schedules, step sizes, the number of subtrajectories, and HMC parameters remains necessary [2412.07081]. The diffusion-path formulation requires many auxiliary particles for accurate score estimates, uses a fixed path design, and its non-asymptotic bounds are conservative in practice [2601.21951]. The forward-KL theory is restricted to gradient drifts and assumes intermediate targets satisfy uniform smoothness, dissipativity, and log-Sobolev control [2601.22349].

Taken together, these results place SCLD at the intersection of annealed Monte Carlo, stochastic control, path-space variational inference, and modern diffusion sampling. In its most explicit form, it is an SMC method whose transition kernels are learned controlled Langevin diffusions and whose correction mechanism is path-space importance weighting with resampling. In broader usage, it also names a control perspective in which a time-dependent Langevin drift is designed, learned, or estimated sequentially so that the induced law tracks a path of easier intermediate distributions and converges to the target.

Source: https://www.emergentmind.com/topics/sequential-controlled-langevin-diffusions-scld