---
title: UBU Splitting in Kinetic Langevin Sampling
url: https://www.emergentmind.com/topics/ubu-splitting
type: topic
---

# UBU Splitting in Kinetic Langevin Sampling

to=arxiv_search.search  大发彩票官网 无码不卡高清免费 code +="{"query":"UBU splitting kinetic Langevin 2405.09992", "max_results": 10, "sort_by": "relevance"}"
to=arxiv_search.search _人人碰 code +="{"query":"\"UBU\" Langevin splitting sampler", "max_results": 10, "sort_by": "relevance"}"
to=arxiv_search.search  天天中彩票买 code +="{"query":"2405.09992", "max_results": 5, "sort_by": "relevance"}"
to=arxiv_search.search  彩神争霸如何 code +="{"query":"Convergence of kinetic Langevin samplers for non-convex potentials", "max_results": 5, "sort_by": "relevance"}"
to=arxiv_search.search  天天中彩票中了 code +="{"query":"Sampling from Bayesian Neural Network Posteriors with Symmetric Minibatch Splitting Langevin Dynamics UBU", "max_results": 5, "sort_by": "relevance"}"
UBU splitting denotes the palindromic kinetic Langevin composition
\[
\mathcal U_{h/2}\circ \mathcal B_h\circ \mathcal U_{h/2},
\]
where \(\mathcal U\) is the universal Ornstein–Uhlenbeck \(+\) transport subflow and \(\mathcal B\) is the potential or force subflow. In the recent sampling literature, it appears as a symmetric, second-order splitting strategy for underdamped Langevin dynamics, with the most detailed non-convex convergence analysis given in "Convergence of kinetic Langevin samplers for non-convex potentials" [2405.09992]. In that work, UBU is analyzed alongside Euler discretization and the BU scheme, and is identified as the most accurate scheme among the three under stronger regularity assumptions. Related work studies the same splitting in stochastic-gradient and minibatch settings, including SG-UBU, SVRG-UBU, SAGA-UBU, and SMS-UBU [2511.04413; 2410.19780].

## 1. Continuous dynamics and the splitting principle

The underlying continuous-time process is the kinetic Langevin system
\[
\begin{cases}
dX_t = V_t\,dt,\\[2mm]
dV_t = -\nabla U(X_t)\,dt - \gamma V_t\,dt + \sqrt{2\gamma}\,dB_t,
\end{cases}
\]
with friction \(\gamma>0\). Its invariant measure is the Boltzmann-Gibbs law
\[
\mu(dx\,dv)\propto e^{-U(x)-|v|^2/2}\,dx\,dv.
\]
The operator-splitting viewpoint used in the non-convex analysis separates the dynamics into a force step \(\mathcal B\) and a universal Ornstein–Uhlenbeck \(+\) transport step \(\mathcal U\) [2405.09992].

In this formulation, UBU is distinguished by its symmetry. The composition
\[
\mathcal U_{h/2}\circ \mathcal B_h\circ \mathcal U_{h/2}
\]
is a palindromic composition, and the cited analysis attributes to it a higher-quality splitting than Euler. Under additional smoothness assumptions, that structure yields second-order asymptotic bias rather than first-order bias [2405.09992].

The same structural description appears in later stochastic-gradient work, where the notation is written as
\[
\Phi_{h/2}^{\mathfrak U}\circ \Phi_h^{\mathfrak B}\circ \Phi_{h/2}^{\mathfrak U},
\]
with \(\mathfrak U\) the linear Ornstein–Uhlenbeck-like part and \(\mathfrak B\) the force or gradient part. That paper explicitly describes UBU splitting as a Lie–Trotter/Strang-type splitting of the underdamped Langevin SDE into pieces that can be solved exactly or very cheaply [2511.04413].

## 2. Definition of the UBU integrator

For the non-convex kinetic Langevin setting, the force step is
\[
\mathcal B(x,v,h) = (x,\; v-h\nabla U(x)).
\]
The \(\mathcal U\)-step integrates
\[
dx = v\,dt,\qquad dv = -\gamma v\,dt+\sqrt{2\gamma}\,dW_t
\]
in the weak sense. With \(\eta=e^{-\gamma h}\), the map is written using Gaussian random variables \(\xi^{(1)},\xi^{(2)}\sim\mathcal N(0,I_d)\) as
\[
\mathcal U(x,v,h,\xi^{(1)},\xi^{(2)}) = \Big( x+\frac{1-\eta}{\gamma}v+\sqrt{\frac{2}{\gamma}\Big(\mathcal Z^{(1)}-\mathcal Z^{(2)}\Big)}, \; \eta v+\sqrt{2\gamma}\,\mathcal Z^{(2)} \Big),
\]
where \(\mathcal Z^{(1)}\) and \(\mathcal Z^{(2)}\) are auxiliary Gaussian constructions encoding the correlated Brownian increments for the integrated OU dynamics [2405.09992].

The full one-step UBU update is
\[
(x_{k+1},v_{k+1})=\mathcal U\!\left( \mathcal B\!\left(\mathcal U(x_k,v_k,h/2,\xi_{k+1}^{(1)},\xi_{k+1}^{(2)}),h\right), h/2,\xi_{k+1}^{(3)},\xi_{k+1}^{(4)} \right),
\]
that is,
\[
\boxed{\mathcal U_{h/2}\;\to\;\mathcal B_h\;\to\;\mathcal U_{h/2}.}
\]
This exact ordering is the defining feature of the scheme [2405.09992].

A closely related presentation appears in Bayesian posterior sampling, where the same force impulse is written
\[
\mathcal B(x,v,h)=(x,\,v-h\nabla f(x)),
\]
and the \(\mathcal U\)-half-step is expressed through the exact OU semigroup, with damping factor \(\eta=e^{-\gamma h/2}\). That work emphasizes that the exact treatment of the \(\mathcal U\) piece and the overall symmetry give second-order accuracy for the full-gradient scheme without Metropolis correction [2410.19780].

## 3. Assumptions, coordinates, and the custom distance

The non-convex theory assumes that \(U\) is \(L\)-smooth and strongly convex outside a ball. More precisely, there exist \(\kappa>0\) and \(R\ge 0\) such that
\[
(\nabla U(x)-\nabla U(y))\cdot(x-y)\ge \kappa |x-y|^2
\quad \text{whenever } |x-y|>R.
\]
Equivalently,
\[
U(x)=\tfrac12 x^\top Kx + G(x),
\]
where \(K\) is symmetric positive definite with smallest eigenvalue \(\kappa\), and \(G\) is convex outside the ball of radius \(R\) [2405.09992].

For the improved UBU bias and complexity theory, the same paper imposes additional regularity, either a Lipschitz Hessian condition or the stronger strongly Hessian Lipschitz condition. This is the point at which the sharper UBU complexity bounds become available [2405.09992].

A central technical ingredient is a hybrid synchronous/reflection coupling for the difference variables
\[
Z_k=X_k-X_k',\qquad W_k=V_k-V_k',
\]
together with the auxiliary combination
\[
q_k=Z_k+\gamma^{-1}W_k.
\]
The construction uses synchronous coupling when the chains are far apart and a reflection-type coupling in the small-distance regime [2405.09992].

The resulting analysis is organized around two intermediate distances. The large-distance metric is
\[
r_l((x,v),(x',v'))^2 = \gamma^{-2}(x-x')^\top K(x-x') +\frac12\big|(1-2\tau)(x-x')+\gamma^{-1}(v-v')\big|^2 +\frac12\gamma^{-2}|v-v'|^2,
\]
while the small-distance metric is
\[
r_s((x,v),(x',v')) = \alpha |x-x'|+|x-x'+\gamma^{-1}(v-v')|,
\]
with
\[
\tau = \min\Big(\frac18,\ \frac{\kappa}{2\gamma^2}-\frac{L_G^2}{\gamma^4}\Big), \qquad \alpha = 2L\gamma^{-2}.
\]
These are glued into a single metric \(\rho\) through a concave increasing function \(f\),
\[
\rho((x,v),(x',v')) = f\!\left(\Delta((x,v),(x',v'))\wedge D_{\mathcal K+\epsilon r_l}\right),\qquad \Delta=r_s-\epsilon r_l.
\]
The paper proves the equivalence
\[
|(x,v)-(x',v')| \le \mathbf N\,\rho((x,v),(x',v')) \le \mathbf M\, |(x,v)-(x',v')|,
\]
with \(\mathbf M,\mathbf N\) independent of \(d\) and \(h\). This is what permits transfer of contraction estimates in \(\rho\) to standard \(L^1\)-Wasserstein control [2405.09992].

## 4. Contractivity, ergodicity, and invariant measures

The UBU convergence theorem is obtained by combining the BU contraction estimate with control of the two half-\(\mathcal U\) steps under synchronous coupling. The resulting bound is
\[
\mathbb E\big[\rho((X_k,V_k),(X_k',V_k'))\big]
\le \mathbf C\,(1-ch)^k\,\rho((x,v),(x',v')),
\]
where \(\mathbf C>0\) is the prefactor arising from the two half-\(\mathcal U\) corrections [2405.09992].

Via the equivalence of metrics, the same result yields
\[
\mathcal W_1(\mu_k,\nu_k) \le \mathbf C\,\mathbf M\,e^{-chk}\,\mathcal W_1(\mu_0,\nu_0).
\]
The stated consequences are the existence of a unique invariant measure \(\mu_{h,\infty}\) for the UBU chain and exponential convergence to it, with constants independent of dimension and step size under the imposed restrictions [2405.09992].

The paper situates UBU relative to Euler and BU. Euler–Maruyama is described as the simplest scheme to implement, with one gradient evaluation per step but only first-order asymptotic bias. BU uses the splitting \(\mathcal B\) then \(\mathcal U\), and is analyzed with the same custom metric and coupling framework. UBU retains those coupling ideas while adding control of the half-\(\mathcal U\) corrections and achieves the best smooth-regime bias behavior among the three schemes [2405.09992].

A concise comparison is:

| Scheme | Composition | Stated behavior |
|---|---|---|
| Euler | Euler discretization | first-order asymptotic bias |
| BU | \(\mathcal B_h\to \mathcal U_h\) | contraction via the same custom metric/coupling framework |
| UBU | \(\mathcal U_{h/2}\to \mathcal B_h\to \mathcal U_{h/2}\) | higher-order bias; best complexity in the smooth setting |

This suggests that the principal analytical role of UBU is not merely stability, but the conversion of splitting symmetry into stronger invariant-measure control.

## 5. Discretization bias and complexity

The asymptotic bias is measured as
\[
\mathcal W_1(\mu_\infty,\mu_{h,\infty}),
\]
where \(\mu_\infty\) is the true invariant measure of the diffusion and \(\mu_{h,\infty}\) is the invariant law of the UBU discretization. The argument compares the continuous diffusion and UBU over many steps using synchronized Brownian motion and then uses the contraction estimate to convert pathwise discretization discrepancy into a stationary bias bound [2405.09992].

Under only the basic smoothness and non-convex-outside-a-ball assumptions, the paper obtains an \(O(h)\)-type bias:
\[
\mathcal W_\rho(\mu_\infty,\mu_{h,\infty}) \lesssim h\sqrt d.
\]
If \(U\) has Lipschitz Hessian, the bias improves to
\[
\mathcal W_\rho(\mu_\infty,\mu_{h,\infty}) \lesssim h^2\sqrt d,
\]
and under the stronger Hessian-Lipschitz condition the dimension dependence improves further. The paper identifies this as the key point where UBU outperforms Euler: Euler has first-order bias, whereas UBU achieves second-order asymptotic bias with extra smoothness [2405.09992].

The total error is decomposed into discretization bias and mixing error. For \(\varepsilon\)-accuracy in \(L^1\)-Wasserstein distance, the stated complexity bounds are:
\[
\mathcal O\!\left(\frac{\sqrt d}{\varepsilon}\right)
\quad \text{for Euler},
\]
\[
\mathcal O\!\left(\frac{\sqrt d}{\sqrt\varepsilon}\right)
\quad \text{for UBU under Hessian-Lipschitz smoothness},
\]
and, under the stronger Hessian condition,
\[
\boxed{\mathcal O\!\left(\frac{d^{1/4}}{\sqrt\varepsilon}\right).}
\]
The abstract of the same paper presents this \( \mathcal O(d^{1/4}/\sqrt{\varepsilon}) \) rate as the headline UBU complexity guarantee under appropriate regularity assumptions on the target measure [2405.09992].

This improved scaling persists in a related 2026 splitting framework based on an exact harmonic Langevin integrator. That work is not literally classical UBU, but it explicitly states that its second-order scheme has step-size requirements comparable to established splitting schemes such as OBABO or UBU, and it reaches the same \(h^{-1}=O(d^{1/4}\varepsilon^{-1/2})\) scaling for \(\varepsilon\)-accuracy [2605.24070].

## 6. Stochastic-gradient, minibatch, and mean-field extensions

UBU splitting has been extended beyond the full-gradient, single-particle setting. In SG-UBU, the force term \(\nabla U\) is replaced by an unbiased estimator \(b(\bm x,\theta)\), and the update is
\[
(\bm{X}_{k+1},\bm{V}_{k+1})=
\big(\Phi_{h/2}^{\mathfrak U}\circ \Phi_h^{\mathfrak B(\theta_k)}\circ \Phi_{h/2}^{\mathfrak U}\big)(\bm{X}_k,\bm{V}_k).
\]
A notable structural point is that the stochastic gradient is evaluated at the intermediate position \(\bm Y_k\), not at \(\bm X_k\). The mean-square error analysis in "Mean square error analysis of stochastic gradient and variance-reduced sampling algorithms" shows that SG-UBU has first-order numerical bias in \(h\), with leading coefficient proportional to the stochastic-gradient variance, while the pure discretization contribution remains \(O(h^2)\) [2511.04413].

For finite-sum potentials, the same paper studies SVRG-UBU and SAGA-UBU. It reports a phase transition in the numerical bias: for larger step sizes the bias is first order, while below a critical threshold the rate becomes second order. For SVRG-UBU, the threshold is identified as
\[
h \sim \frac{\sqrt{m}\,p}{N}.
\]
A plausible implication is that the second-order structure of UBU remains visible in stochastic-gradient sampling only when variance reduction is strong enough to suppress the dominant first-order gradient-noise term [2511.04413].

A separate minibatch extension, SMS-UBU, combines UBU with a symmetric forward/backward sweep over a random partition of the data. Its main theorem gives
\[
\mathcal W_2(\widetilde\pi_k,\pi) \le \sqrt{2}\,e^{-\frac{mh}{8\gamma}\lfloor k/N_m\rfloor}\mathcal W_{2,a,b}(\pi_0,\bar\pi) + C\,h^2\sqrt d,
\]
and the paper advertises this as bias \(O(h^2 d^{1/2})\) despite using one minibatch per iteration. In Bayesian neural network experiments, SMS-UBU is reported to improve calibration metrics relative to standard training and stochastic weight averaging, while remaining cheaper than metropolized alternatives [2410.19780].

The full-gradient non-convex paper also extends the UBU analysis to interacting particle systems of mean-field type. For
\[
U(\mathbf x)=\sum_{i=1}^N\left(V(x^i)+\frac1N\sum_{j\ne i}W(x^i-x^j)\right),
\]
it introduces the particlewise metric
\[
\rho_N((\mathbf x,\mathbf v),(\mathbf x',\mathbf v')) = \frac1N\sum_{i=1}^N \rho((x^i,v^i),(x^{\prime i},v^{\prime i})),
\]
and states that the contraction and complexity bounds extend with constants independent of \(N\), provided the interaction is sufficiently weak in Lipschitz norm relative to \(\kappa\). Combined with propagation-of-chaos estimates, this yields
\[
\mathcal W_{\ell_1^N}(\mu_k^N,\mu_*^{\otimes N}) \le C\Big( e^{-chk}\mathcal W_1(\mu_0,\mu_*) +\sqrt d\,h^2 +N^{-1/2} \Big)
\]
for the UBU discretization under the stronger smoothness regime [2405.09992].

In current usage, UBU splitting therefore denotes more than a formal symmetric composition. It is a specific second-order kinetic Langevin discretization whose symmetry supports non-convex Wasserstein contraction, second-order invariant-measure bias under higher smoothness, and a family of stochastic-gradient and minibatch variants that attempt to preserve those advantages in scalable regimes [2405.09992; 2511.04413; 2410.19780].

Source: https://www.emergentmind.com/topics/ubu-splitting