DLP: Gradient-Based Proposals for Discrete Sampling

• Discrete Langevin Proposals are gradient-based kernels defined natively on discrete product spaces, enabling scalable and parallel updates for sampling.
• They derive from a discrete Wasserstein gradient-flow framework and connect to Glauber dynamics, ensuring controlled bias and convergence to the target distribution.
• Advanced variants such as DMALA, DULA, and DREAM incorporate Metropolis correction, replica exchange, and parallel tempering to navigate multimodal landscapes efficiently.

Searching arXiv for recent and foundational papers on discrete Langevin proposals. Discrete Langevin Proposals (DLP) are gradient-based proposal kernels for sampling discrete distributions on finite product spaces such as $\{0,1\}^d$ or categorical grids. In their canonical form, they transplant the algebra of Langevin proposals from continuous spaces to discrete domains by defining a Gaussian-like kernel directly on the discrete state space, with a step-size controlling move magnitude and a coordinate-wise factorization enabling parallel updates. Subsequent work has given DLP several complementary interpretations: as a simple scalable proposal for high-dimensional discrete targets, as a discretization of a discrete Wasserstein gradient flow, and as a first-order approximation to continuous-time Glauber spin-flip dynamics. The resulting family includes unadjusted, Metropolis-adjusted, stochastic-gradient, preconditioned, replica-exchange, and parallel-tempering variants (Zhang et al., 2022, Sun et al., 2022, Gissler et al., 17 Feb 2026).

1. Core construction on discrete product spaces

A standard DLP setup takes a discrete variable $x\in\mathcal X$ with product structure

$\mathcal X=\prod_{i=1}^d \mathcal X_i,$

and target

$\pi(x)=\frac{1}{Z}\exp(U(x)).$

The proposal kernel is defined by

$$q(x' \mid x) = \frac{ \exp\!\bigl(\tfrac12\langle x'-x,\nabla\log\pi(x)\rangle -\tfrac1{2\delta}\|x'-x\|^2\bigr) }{ \sum_{y\in\mathcal X} \exp\!\bigl(\tfrac12\langle y-x,\nabla\log\pi(x)\rangle -\tfrac1{2\delta}\|y-x\|^2\bigr) },$$

where $\delta$ is the step-size. Because $\mathcal X$ factorizes coordinate-wise, the proposal itself factorizes:

$$q(x'\mid x)=\prod_{i=1}^d q_i(x_i'\mid x), \qquad q_i(x_i'\mid x)\propto \exp\!\Bigl( \tfrac12\,\nabla_i\log\pi(x)\,(x_i'-x_i)-\tfrac1{2\delta}(x_i'-x_i)^2 \Bigr).$$

This yields a single-gradient, $O(d)$ update in which all coordinates can be sampled independently and in parallel (Zhang et al., 2022).

A closely related presentation introduces temperature $\tau>0$ and step-size $x\in\mathcal X$0, with

$x\in\mathcal X$1

and proposal

$x\in\mathcal X$2

On factorized domains, the corresponding coordinate-wise log-weight is

$x\in\mathcal X$3

In this parameterization, small $x\in\mathcal X$4 produces local moves, whereas large $x\in\mathcal X$5 permits global jumps at the cost of low Metropolis acceptance; increasing $x\in\mathcal X$6 flattens $x\in\mathcal X$7 and encourages exploration (Zheng et al., 28 Jan 2025).

These constructions are discrete analogues of the continuous Langevin proposal, but they are not obtained by rounding a Gaussian sample after the fact. Instead, the kernel is defined natively on the discrete domain by replacing integration with summation in the normalizing constant. This suggests that DLP is best understood as a discrete-state proposal family rather than as a discretized continuous sampler with an external projection step (Zhang et al., 2022).

2. Wasserstein gradient-flow formulation and discrete Langevin dynamics

A more principled derivation begins from a discrete analogue of the $x\in\mathcal X$8-Wasserstein gradient flow of the Kullback–Leibler divergence. On a finite state space $x\in\mathcal X$9 with symmetric nonnegative weights $\mathcal X=\prod_{i=1}^d \mathcal X_i,$0, one defines

$\mathcal X=\prod_{i=1}^d \mathcal X_i,$1

antisymmetric flows $\mathcal X=\prod_{i=1}^d \mathcal X_i,$2, the discrete continuity equation

$\mathcal X=\prod_{i=1}^d \mathcal X_i,$3

and a Riemannian-type inner product

$\mathcal X=\prod_{i=1}^d \mathcal X_i,$4

where $\mathcal X=\prod_{i=1}^d \mathcal X_i,$5 is a symmetric mobility. The steepest-descent flow of the discrete KL then takes

$\mathcal X=\prod_{i=1}^d \mathcal X_i,$6

so that

$\mathcal X=\prod_{i=1}^d \mathcal X_i,$7

This is the discrete-space counterpart of the Wasserstein gradient-flow picture underlying continuous Langevin Monte Carlo (Sun et al., 2022).

Freezing the mobility prefactor at stationarity and taking $\mathcal X=\prod_{i=1}^d \mathcal X_i,$8 yields a reversible continuous-time Markov chain with off-diagonal generator

$\mathcal X=\prod_{i=1}^d \mathcal X_i,$9

and diagonal

$\pi(x)=\frac{1}{Z}\exp(U(x)).$0

The law evolves as

$\pi(x)=\frac{1}{Z}\exp(U(x)).$1

with transition semigroup

$\pi(x)=\frac{1}{Z}\exp(U(x)).$2

For small $\pi(x)=\frac{1}{Z}\exp(U(x)).$3,

$\pi(x)=\frac{1}{Z}\exp(U(x)).$4

and this first-order kernel is used inside a Metropolis–Hastings wrapper. In product spaces, the global generator decomposes as $\pi(x)=\frac{1}{Z}\exp(U(x)).$5, which yields a factorized proposal

$\pi(x)=\frac{1}{Z}\exp(U(x)).$6

The corresponding acceptance probability is the usual

$\pi(x)=\frac{1}{Z}\exp(U(x)).$7

Within this framework, the paper derives Discrete Langevin Monte Carlo (DLMC), highlights its convenient parallel implementation, and emphasizes its time-uniform sampling and larger jump distances (Sun et al., 2022).

This formulation also clarifies the relationship between proposal geometry and Markov dynamics. Because the underlying object is the exact semigroup $\pi(x)=\frac{1}{Z}\exp(U(x)).$8, the step-size $\pi(x)=\frac{1}{Z}\exp(U(x)).$9 has a direct continuous-time meaning, and the small-$$q(x' \mid x) = \frac{ \exp\!\bigl(\tfrac12\langle x'-x,\nabla\log\pi(x)\rangle -\tfrac1{2\delta}\|x'-x\|^2\bigr) }{ \sum_{y\in\mathcal X} \exp\!\bigl(\tfrac12\langle y-x,\nabla\log\pi(x)\rangle -\tfrac1{2\delta}\|y-x\|^2\bigr) },$$0 approximation is a computational device rather than a redefinition of the time scale. That distinction is central to the claim that DLP-like samplers can be tied to a genuine reversible jump process instead of to an ad hoc discrete update rule (Sun et al., 2022).

3. Score-based formulations and the link to Glauber dynamics

On the binary hypercube $$q(x' \mid x) = \frac{ \exp\!\bigl(\tfrac12\langle x'-x,\nabla\log\pi(x)\rangle -\tfrac1{2\delta}\|x'-x\|^2\bigr) }{ \sum_{y\in\mathcal X} \exp\!\bigl(\tfrac12\langle y-x,\nabla\log\pi(x)\rangle -\tfrac1{2\delta}\|y-x\|^2\bigr) },$$1, DLP can also be formulated as a discretization of classical Glauber dynamics. For a strictly positive target $$q(x' \mid x) = \frac{ \exp\!\bigl(\tfrac12\langle x'-x,\nabla\log\pi(x)\rangle -\tfrac1{2\delta}\|x'-x\|^2\bigr) }{ \sum_{y\in\mathcal X} \exp\!\bigl(\tfrac12\langle y-x,\nabla\log\pi(x)\rangle -\tfrac1{2\delta}\|y-x\|^2\bigr) },$$2, the continuous-time generator is

$$q(x' \mid x) = \frac{ \exp\!\bigl(\tfrac12\langle x'-x,\nabla\log\pi(x)\rangle -\tfrac1{2\delta}\|x'-x\|^2\bigr) }{ \sum_{y\in\mathcal X} \exp\!\bigl(\tfrac12\langle y-x,\nabla\log\pi(x)\rangle -\tfrac1{2\delta}\|y-x\|^2\bigr) },$$3

where

$$q(x' \mid x) = \frac{ \exp\!\bigl(\tfrac12\langle x'-x,\nabla\log\pi(x)\rangle -\tfrac1{2\delta}\|x'-x\|^2\bigr) }{ \sum_{y\in\mathcal X} \exp\!\bigl(\tfrac12\langle y-x,\nabla\log\pi(x)\rangle -\tfrac1{2\delta}\|y-x\|^2\bigr) },$$4

is the Glauber score and

$$q(x' \mid x) = \frac{ \exp\!\bigl(\tfrac12\langle x'-x,\nabla\log\pi(x)\rangle -\tfrac1{2\delta}\|x'-x\|^2\bigr) }{ \sum_{y\in\mathcal X} \exp\!\bigl(\tfrac12\langle y-x,\nabla\log\pi(x)\rangle -\tfrac1{2\delta}\|y-x\|^2\bigr) },$$5

is the logistic sigmoid. This process has invariant law $$q(x' \mid x) = \frac{ \exp\!\bigl(\tfrac12\langle x'-x,\nabla\log\pi(x)\rangle -\tfrac1{2\delta}\|x'-x\|^2\bigr) }{ \sum_{y\in\mathcal X} \exp\!\bigl(\tfrac12\langle y-x,\nabla\log\pi(x)\rangle -\tfrac1{2\delta}\|y-x\|^2\bigr) },$$6 and flips each coordinate independently at rate $$q(x' \mid x) = \frac{ \exp\!\bigl(\tfrac12\langle x'-x,\nabla\log\pi(x)\rangle -\tfrac1{2\delta}\|x'-x\|^2\bigr) }{ \sum_{y\in\mathcal X} \exp\!\bigl(\tfrac12\langle y-x,\nabla\log\pi(x)\rangle -\tfrac1{2\delta}\|y-x\|^2\bigr) },$$7 (Gissler et al., 17 Feb 2026).

Two discrete-time kernels are then obtained by truncating the jump process with $$q(x' \mid x) = \frac{ \exp\!\bigl(\tfrac12\langle x'-x,\nabla\log\pi(x)\rangle -\tfrac1{2\delta}\|x'-x\|^2\bigr) }{ \sum_{y\in\mathcal X} \exp\!\bigl(\tfrac12\langle y-x,\nabla\log\pi(x)\rangle -\tfrac1{2\delta}\|y-x\|^2\bigr) },$$8. The first is DULA:

$$q(x' \mid x) = \frac{ \exp\!\bigl(\tfrac12\langle x'-x,\nabla\log\pi(x)\rangle -\tfrac1{2\delta}\|x'-x\|^2\bigr) }{ \sum_{y\in\mathcal X} \exp\!\bigl(\tfrac12\langle y-x,\nabla\log\pi(x)\rangle -\tfrac1{2\delta}\|y-x\|^2\bigr) },$$9

which is equivalent to flipping bit $\delta$0 independently with probability

$\delta$1

The second is DUPS, a two-stage kernel

$\delta$2

with

$\delta$3

Both depend on a discrete score map $\delta$4 (Gissler et al., 17 Feb 2026).

Three score choices are singled out. The Stein score uses a smooth extension $\delta$5,

$\delta$6

The Glauber score is the finite-difference quantity $\delta$7. The Gibbs-adjusted score is

$\delta$8

When the Gibbs score is used in DULA, the generator coincides exactly with the Glauber $\delta$9. Theoretical analysis gives contraction bounds such as

$\mathcal X$0

and

$\mathcal X$1

As $\mathcal X$2, the invariant measures of unadjusted DULA and DUPS converge to $\mathcal X$3, while Metropolis-adjusted DLP converges exactly to $\mathcal X$4 for any $\mathcal X$5 by detailed balance (Gissler et al., 17 Feb 2026).

This line of work sharpens the continuous-time interpretation of discrete Langevin methods. Rather than treating the gradient merely as a heuristic bias toward promising moves, it identifies finite-difference and reweighted scores under which the discrete proposal is a first-order discretization of a known jump process with the correct target law (Gissler et al., 17 Feb 2026).

4. Bias, Metropolis correction, and invariant laws

In the original product-space formulation, the unadjusted chain that repeatedly samples from $\mathcal X$6 has a stationary distribution $\mathcal X$7. For exact log-quadratic targets

$\mathcal X$8

the chain is reversible with respect to some $\mathcal X$9, and if $$q(x'\mid x)=\prod_{i=1}^d q_i(x_i'\mid x), \qquad q_i(x_i'\mid x)\propto \exp\!\Bigl( \tfrac12\,\nabla_i\log\pi(x)\,(x_i'-x_i)-\tfrac1{2\delta}(x_i'-x_i)^2 \Bigr).$$0 then

$$q(x'\mid x)=\prod_{i=1}^d q_i(x_i'\mid x), \qquad q_i(x_i'\mid x)\propto \exp\!\Bigl( \tfrac12\,\nabla_i\log\pi(x)\,(x_i'-x_i)-\tfrac1{2\delta}(x_i'-x_i)^2 \Bigr).$$1

For targets whose gradient is uniformly close to an affine form $$q(x'\mid x)=\prod_{i=1}^d q_i(x_i'\mid x), \qquad q_i(x_i'\mid x)\propto \exp\!\Bigl( \tfrac12\,\nabla_i\log\pi(x)\,(x_i'-x_i)-\tfrac1{2\delta}(x_i'-x_i)^2 \Bigr).$$2, the bias remains small:

$$q(x'\mid x)=\prod_{i=1}^d q_i(x_i'\mid x), \qquad q_i(x_i'\mid x)\propto \exp\!\Bigl( \tfrac12\,\nabla_i\log\pi(x)\,(x_i'-x_i)-\tfrac1{2\delta}(x_i'-x_i)^2 \Bigr).$$3

To recover exact invariance, one adds a Metropolis–Hastings correction,

$$q(x'\mid x)=\prod_{i=1}^d q_i(x_i'\mid x), \qquad q_i(x_i'\mid x)\propto \exp\!\Bigl( \tfrac12\,\nabla_i\log\pi(x)\,(x_i'-x_i)-\tfrac1{2\delta}(x_i'-x_i)^2 \Bigr).$$4

obtaining the DMALA chain, which is reversible and converges to $$q(x'\mid x)=\prod_{i=1}^d q_i(x_i'\mid x), \qquad q_i(x_i'\mid x)\propto \exp\!\Bigl( \tfrac12\,\nabla_i\log\pi(x)\,(x_i'-x_i)-\tfrac1{2\delta}(x_i'-x_i)^2 \Bigr).$$5 (Zhang et al., 2022).

The same correction appears in temperature-parameterized DLP:

$$q(x'\mid x)=\prod_{i=1}^d q_i(x_i'\mid x), \qquad q_i(x_i'\mid x)\propto \exp\!\Bigl( \tfrac12\,\nabla_i\log\pi(x)\,(x_i'-x_i)-\tfrac1{2\delta}(x_i'-x_i)^2 \Bigr).$$6

Because $$q(x'\mid x)=\prod_{i=1}^d q_i(x_i'\mid x), \qquad q_i(x_i'\mid x)\propto \exp\!\Bigl( \tfrac12\,\nabla_i\log\pi(x)\,(x_i'-x_i)-\tfrac1{2\delta}(x_i'-x_i)^2 \Bigr).$$7 is finite and $$q(x'\mid x)=\prod_{i=1}^d q_i(x_i'\mid x), \qquad q_i(x_i'\mid x)\propto \exp\!\Bigl( \tfrac12\,\nabla_i\log\pi(x)\,(x_i'-x_i)-\tfrac1{2\delta}(x_i'-x_i)^2 \Bigr).$$8, the chain is irreducible and aperiodic, and standard MH theory gives invariance of $$q(x'\mid x)=\prod_{i=1}^d q_i(x_i'\mid x), \qquad q_i(x_i'\mid x)\propto \exp\!\Bigl( \tfrac12\,\nabla_i\log\pi(x)\,(x_i'-x_i)-\tfrac1{2\delta}(x_i'-x_i)^2 \Bigr).$$9. In the two-replica setting, if MH corrections are used, the resulting DREAM chain converges exactly to $O(d)$0 independently of $O(d)$1. Even without MH, for log-quadratic

$O(d)$2

the invariant law $O(d)$3 converges weakly, in total variation, to $O(d)$4 as $O(d)$5 (Zheng et al., 28 Jan 2025).

A persistent theme across these analyses is that unadjusted DLP is not simply either “correct” or “incorrect.” For log-quadratic targets its stationary bias vanishes in the small-step limit, and for nearby targets it is controlled quantitatively; nevertheless, exact stationarity for fixed step-size is obtained only after Metropolis correction (Zhang et al., 2022, Zheng et al., 28 Jan 2025).

5. Replica exchange and parallel tempering for multimodal landscapes

Gradient-based discrete samplers can stagnate in non-convex energy landscapes, and recent work augments DLP with multiple temperatures and replica exchange. DREXEL and DREAM run two DLP chains in parallel: Replica-1 uses small $O(d)$6 and low temperature $O(d)$7 for local sampling, whereas Replica-2 uses larger $O(d)$8 and higher $O(d)$9 for broader exploration. After proposal updates, a history-aware swap is attempted with

$\tau>0$0

designed so that the joint chain on $\tau>0$1 satisfies detailed balance with respect to $\tau>0$2. Under mild conditions, the joint chain is reversible with respect to $\tau>0$3, and DREAM converges exactly to $\tau>0$4 independently of $\tau>0$5 (Zheng et al., 28 Jan 2025).

Empirically, DREAM recovers all modes on 2D synthetic multi-modal energies, while baselines listed as DMALA, AB, and ACS miss many; KL and MMD are reported as down by a factor of $\tau>0$6–$\tau>0$7. On Ising models, DREXEL and DREAM reduce log-RMSE faster than DULA, ACS, and DMALA, and DREAM achieves the best final RMSE. On restricted Boltzmann machines, DREAM is reported to have consistently lowest MMD. On deep energy-based models trained via PCD plus the sampler, AIS test log-likelihood improves by $\tau>0$8–$\tau>0$9 nats over DMALA, with the best test log-likelihood on Static/Dynamic MNIST, Omniglot, and Caltech Silhouettes (Zheng et al., 28 Jan 2025).

Parallel Tempering enhanced DLP generalizes this idea to $x\in\mathcal X$00 replicas with inverse temperatures

$x\in\mathcal X$01

targeting

$x\in\mathcal X$02

Adjacent replicas are swapped using the standard acceptance ratio

$x\in\mathcal X$03

The method also introduces an automatic temperature-schedule and chain-count selection rule based on the round-trip rate

$x\in\mathcal X$04

which is maximized when all neighboring swap probabilities are equal. Theoretical results establish a uniform minorization for each tempered MH-corrected chain, uniform ergodicity of the joint PT-DLP chain,

$x\in\mathcal X$05

and faster mixing than single-chain DLP. Reported experiments show consistently higher Entropic Mode Coverage and lower MMD and forward KL on synthetic multimodal targets, faster convergence on RBMs, and the best or tied best AIS-estimated test log-likelihoods on deep convolutional EBMs (Liang et al., 26 Feb 2025).

6. Relations to other samplers and broader terminology

DLP occupies a crowded methodological neighborhood that includes Gibbs-with-gradients (GWG), locally balanced proposals, discrete MALA variants, and generalized jump-process discretizations. Within the Wasserstein-gradient-flow derivation, GWG and LB–1 arise by taking $x\in\mathcal X$06, restricting moves to nearest-neighbor binary flips, and using a renormalized first-order expansion rather than the exact continuous-time semigroup. Their proposal ratios scale like $x\in\mathcal X$07, matching the off-diagonal rates but ignoring the self-loop correction. DMALA is described there as a Gaussian-like log-linearization that can be viewed as applying Euler’s method to a diffusion rather than to the true jump process, and therefore cannot recover the correct asymptotic ratio $x\in\mathcal X$08 as $x\in\mathcal X$09. By contrast, DLP/DLMC is derived from the exact discrete Wasserstein gradient flow, with a genuine reversible CTMC generator $x\in\mathcal X$10 and a small-time approximation that preserves the self-loop term (Sun et al., 2022).

The term “Discrete Langevin Proposal” is also broader than the discrete-state sampling literature alone. In continuous-state MCMC, it has been used for Metropolized discretizations of overdamped Langevin dynamics. In that setting, modified proposals can improve the strong order from $x\in\mathcal X$11 for standard MALA to $x\in\mathcal X$12, with rejection rate scaling $x\in\mathcal X$13 rather than $x\in\mathcal X$14, while Barker-corrected midpoint or HMC-type proposals reduce transport-coefficient bias from $x\in\mathcal X$15 or $x\in\mathcal X$16 to $x\in\mathcal X$17 (Fathi et al., 2015). High-dimensional AR(1) and $x\in\mathcal X$18-Langevin analyses further show that, under eigenvalue growth assumptions and step-size scaling $x\in\mathcal X$19, the asymptotically optimal acceptance rate is approximately $x\in\mathcal X$20 (Norton et al., 2016).

That terminological overlap is a recurrent source of ambiguity. In contemporary discrete-distribution research, DLP most commonly denotes gradient-based proposals defined directly on discrete state spaces and often factorized across coordinates. In older continuous-state analyses, the same phrase refers to discrete-time Langevin proposals used inside Metropolis–Hastings. The two traditions share Langevin motivation and Metropolization, but they differ in state space, proposal algebra, and the continuous-time limits they approximate (Zhang et al., 2022, Fathi et al., 2015).