Principled Coarse-Graining (PCG) Overview

• Principled Coarse-Graining (PCG) is a methodology that defines explicit mappings from fine-scale systems to reduced models using variational or probabilistic targets.
• PCG captures equilibrium and dynamic properties by targeting exact objects like the potential of mean force, rate functionals, and spectral modes through controlled many-body approximations.
• PCG extends across domains by integrating force matching, Bayesian inference, and machine learning to preserve structural fidelity and slow kinetics in coarse representations.

Searching arXiv for recent and foundational papers on principled coarse-graining, force matching, predictive coarse-graining, and spectral/operator-based coarse-graining. Principled Coarse-Graining (PCG) denotes a class of reduction strategies in which the coarse representation is defined by an explicit mapping, projection, variational principle, or probabilistic generative model, and the reduced model is judged against a mathematically specified target induced by the fine-scale system rather than by ad hoc agreement with a single observable. In molecular statistical mechanics, that target is often the marginal equilibrium distribution or its potential of mean force (PMF); in stochastic dynamics it may be a coarse large-deviation functional or a projected generator; in data-driven formulations it may be the predictive distribution of fine-scale observables conditioned on latent coarse variables. Across these settings, PCG treats coarse-graining as a controlled reduction of degrees of freedom, with explicit attention to representability, information loss, and the structure preserved by the map (Scherer et al., 2017, Duong et al., 2014, Schöberl et al., 2016).

1. Definition and conceptual core

A recurring starting point in PCG is that coarse-graining is not merely a dimensionality reduction heuristic but a projection from fine-grained variables onto prescribed coarse-grained degrees of freedom. In the molecular setting, coarse-graining “can be thought of as a projection on coarse-grained DOF,” and for a fixed mapping the exact coarse-grained potential energy surface exists as a many-body PMF consistent with the mapped equilibrium distribution. The exact coarse-grained free-energy surface may be written as

$$U_{\text{eff}}(r_1,\ldots,r_M):=-\frac{1}{\beta}\log\int_{\{T\mathbf{q}=\mathbf{r}\}} dq_1\cdots dq_N\, e^{-\beta U(\mathbf q)}.$$

This formulation makes the mapping primary and the effective interaction secondary: one first specifies the coarse variables, then derives or approximates the induced reduced model (Scherer et al., 2017, Tsourtis et al., 2016).

This viewpoint also appears in broader variational formulations. In the large-deviation framework, one begins from a microscopic rate functional with dual representation

$I(\rho)=\sup_f J(\rho,f),$

introduces a coarse-graining map $\Pi:\mathcal X\to\mathcal Y$, and studies the push-forward $\hat\rho=\Pi_\#\rho$ by restricting test functions to the coarse form $f=g\circ\Pi$. The reduced description is therefore induced by the variational structure itself, not guessed independently. In the generalized mean-field formulation of equilibrium coarse-graining, the effective model is defined by a self-consistency condition that forces selected ensemble averages or correlations of the reduced model to match target averages. These formulations differ technically, but they share the same principle: the reduced model is meaningful only relative to an explicit target determined by the fine-scale theory (Duong et al., 2014, Larini et al., 2010).

A plausible implication is that PCG is best understood not as a single algorithm but as a methodological criterion. What makes a procedure “principled” is not whether it uses force matching, relative entropy, cluster expansion, or a neural network, but whether the map, objective, and retained structure are all explicitly defined.

2. Exact targets: PMFs, rate functionals, and constitutive equations

In equilibrium molecular coarse-graining, the canonical exact target is the PMF or marginal free energy associated with the coarse variables. For pair observables, the pair PMF can be written from the radial force as

$U_{\text{PMF}}(r) = -\int_0^r F_r(r')\,dr',$

with $g(r)\sim \exp[-U_{\text{PMF}}(r)/k_BT]$. In cluster-expansion approaches, the exact coarse-grained Hamiltonian is expanded hierarchically into two-body, three-body, and higher-body terms, with the pair and three-body contributions given by explicit cluster integrals over conditional molecular measures. In this setting, the cluster-expansion pair term is the leading term of the exact pair PMF, so PMF theory and cluster expansion are not competing interpretations but different levels of approximation to the same exact reduced free energy (Scherer et al., 2017, Tsourtis et al., 2016).

The same exact-target logic appears in large-deviation and variational theories. In the large-deviation formulation, the limiting coarse-grained rate functional is defined by

$$\widehat I^0(\hat\rho^0)=\sup_g \widehat J^0(\hat\rho^0,g),$$

so the coarse model is characterized variationally rather than phenomenologically. In the generalized mean-field formulation, one chooses observables or correlation functions and determines an effective Hamiltonian by a stationarity or matching condition. The reduction and inverse problems are thereby unified: one may either infer an effective microscopic model from low-resolution data or derive a reduced model from a known microscopic ensemble (Duong et al., 2014, Larini et al., 2010).

The significance of these exact targets is methodological. PCG does not assume that a reduced pair potential, a low-dimensional Markov model, or a neural surrogate is itself the coarse-grained truth. Rather, the exact target is typically a many-body object, a variational functional, or a marginal distribution, and practical coarse-grained models arise by projecting that target onto a tractable representation.

3. Many-body representability and molecular force fields

The representability problem is central to PCG because the exact coarse-grained free-energy surface is generally many-body, whereas practical force fields are usually truncated. A direct illustration is the analysis of two-body and three-body contributions in liquid water and liquid methanol. Simply adding a short-ranged Stillinger–Weber three-body term to a pair force field and reoptimizing can produce “unphysical parametrizations of the CG PES,” because the three-body term is “not orthogonal” to the pair potential when liquid structure is used as the target observable. In water, the pair PMF obtained in such simultaneous fits can become too attractive in the first coordination shell, with the three-body term supplying a compensating short-range repulsion. The resulting total structure may look satisfactory even though the decomposition of physics between pair and three-body contributions is distorted. A residual-force strategy, described as analogous to Gram–Schmidt orthogonalization, provides a cleaner diagnostic of genuine many-body effects: water requires explicit three-body terms to reproduce structure and thermodynamics satisfactorily, whereas methanol is largely captured by pair terms for the main observables examined (Scherer et al., 2017).

The cluster-expansion construction gives a mathematically explicit hierarchy for this many-body structure. For methane and ethane, the exact coarse-grained Hamiltonian is approximated as

$$U_{\text{eff}} = U^{(2)} + U^{(3)} + O(\epsilon^3),$$

with pair and three-body terms derived from connected clusters. The expansion is controlled in a low-density/high-temperature regime, with convergence for $\rho C(\beta)<c_0$. Empirically, the pair term is already accurate for methane in favorable regimes, three-body terms yield only a small improvement in the liquid regime, and substantially better agreement there requires higher-order interactions. This suggests that low-order truncation is state-point dependent and that even an explicitly derived pair-plus-three-body model may remain inadequate for dense liquids (Tsourtis et al., 2016).

Analytical coarse-graining of polymer melts provides a related but distinct example. Using PRISM and integral-equation theory, the effective mesoscale interaction is derived analytically from liquid-state structure and parameterized by the single quantity $I(\rho)=\sup_f J(\rho,f),$0. The resulting coarse-grained model preserves structural correlations, isothermal compressibility, equation of state, Helmholtz and Gibbs free energies, and potential energy and entropy across multiple resolutions. In that formulation, the soft-particle interaction “becomes a free energy in the coarse-grained coordinates,” preserving the excess free energy from an ideal gas across all levels of description (McCarty et al., 2014).

Taken together, these results show that PCG is inseparable from basis design. Matching one structural observable does not uniquely determine how pair and many-body terms should be partitioned, and low-order coarse-grained models can be accurate, misleading, or insufficient depending on both the mapping and the thermodynamic state.

4. Dynamical fidelity: generators, spectra, and projected Langevin dynamics

A major extension of PCG is the demand that a reduced model preserve slow kinetics, not only equilibrium structure. In spectral matching, one starts from a coarse-graining map $I(\rho)=\sup_f J(\rho,f),$1 and trains a reduced reversible generator by targeting the leading eigenvalue equations of the full generator. The central loss may be written as

$I(\rho)=\sup_f J(\rho,f),$2

Here the objective is not merely to reproduce the stationary measure but to preserve the low-lying spectral content of the generator, and hence the slow timescales $I(\rho)=\sup_f J(\rho,f),$3, metastable sets, and transition pathways. In this formulation, force matching enforces thermodynamic consistency, whereas spectral matching enforces kinetic consistency; the two can be combined by first learning the PMF and then fitting a position-dependent diffusion to restore slow timescales (Nüske et al., 2019).

A projection-based treatment of underdamped Langevin dynamics pushes this logic further. Starting from the full Langevin equation in phase space, a reduced phase-space map

$I(\rho)=\sup_f J(\rho,f),$4

is used together with the Zwanzig conditional-expectation projector to define the coarse generator $I(\rho)=\sup_f J(\rho,f),$5. The resulting effective process remains Langevin-like, with explicit drift and diffusion depending on projected forces and the metric $I(\rho)=\sup_f J(\rho,f),$6. Generator Extended Dynamic Mode Decomposition (gEDMD) is then used to estimate the reduced generator and implied timescales, while thermodynamic interpolation extends the construction across temperatures. On the two-dimensional Lemon Slice potential, the resulting coarse model reproduces the empirical reduced free energy, slow implied timescales, and metastable decomposition across a temperature range (Nateghi et al., 3 Dec 2025).

These developments imply that the notion of “correctness” in PCG is multi-objective. A coarse model may reproduce equilibrium marginals yet fail on kinetics, or preserve slow spectral modes while approximating fast fluctuations poorly. PCG therefore becomes a question of which operator or distribution the reduced model is designed to preserve.

5. Probabilistic, Bayesian, and machine-learned formulations

Probabilistic PCG replaces deterministic restriction by latent-variable modeling. In predictive coarse-graining, the reduced variables are latent generators with prior $I(\rho)=\sup_f J(\rho,f),$7 and probabilistic coarse-to-fine map $I(\rho)=\sup_f J(\rho,f),$8, defining the joint model

$I(\rho)=\sup_f J(\rho,f),$9

The induced fine-scale marginal $\Pi:\mathcal X\to\mathcal Y$0 is learned by maximizing likelihood, equivalently minimizing $\Pi:\mathcal X\to\mathcal Y$1. This construction yields predictive posterior distributions for observables, makes uncertainty due to information loss explicit, and uses a hierarchical ARD prior together with MC-EM for sparse model discovery and parameter inference. The paper presents this as an improvement upon the relative entropy method because the reconstruction map itself is learned probabilistically rather than fixed by a deterministic restriction (Schöberl et al., 2016).

A related state-space formulation for nonequilibrium dynamics introduces latent coarse states with a probabilistic transition law and a probabilistic coarse-to-fine emission law. For random walkers, the coarse variable is a latent discretized density field represented through unconstrained variables passed through a softmax, thereby enforcing nonnegativity and mass conservation. Stochastic Variational Inference and sparse Bayesian learning identify salient features in the coarse evolution law and quantify predictive uncertainty arising from both finite data and the intrinsic information loss of coarse-graining (Felsberger et al., 2018).

Machine learning enters PCG in several distinct ways. In CGSchNet, bottom-up force matching is recast as supervised learning of a scalar coarse-grained energy, with SchNet replacing hand-engineered distances, angles, and torsions by learned graph-based features. The theoretical basis remains the variational force-matching principle, which guarantees thermodynamic consistency in the variational limit (Husic et al., 2020). In Coarse-Grained Boltzmann Generators, a continuous normalizing flow generates samples directly in coarse-grained coordinate space, while a learned PMF $\Pi:\mathcal X\to\mathcal Y$2 provides importance weights

$\Pi:\mathcal X\to\mathcal Y$3

This combines reduced-order modeling with asymptotically exact reweighting, and the PMF is learned by variational force matching, including an enhanced-sampling variant that preserves the same optimum under biased sampling (Chen et al., 11 Feb 2026).

These formulations do not abandon PCG’s classical bottom-up targets. Rather, they embed those targets in latent-variable, Bayesian, or generative architectures that explicitly model uncertainty, reconstruction, and sampling efficiency.

6. Mapping selection, structure preservation, and broader extensions

PCG is not confined to the learning of coarse force fields; it also concerns how the coarse variables themselves are selected and what algebraic structure survives reduction. A graph-theoretic mapping scheme for molecular dynamics represents a molecule as a weighted graph and constructs coarse sites by multilevel edge contractions chosen according to a local variation cost

$\Pi:\mathcal X\to\mathcal Y$4

subject to a per-level threshold $\Pi:\mathcal X\to\mathcal Y$5. The method is explicitly aimed at preserving spectral properties of the molecular graph and exact cut weights, so its contribution is primarily a principled mapping-selection criterion rather than a new statistical-mechanical coarse-grained potential (Mondal et al., 22 Jul 2025).

For finite Markov systems, a clustering-based framework defines a deterministic coarse-graining operator and a reconstruction map

$\Pi:\mathcal X\to\mathcal Y$6

interpreted as a generalized Penrose–Moore inverse adapted to the invariant measure. The resulting reduced Markov matrix preserves positivity, mass, the invariant measure, detailed balance, and compatible tensor-space structures for fluxes and quotient graphs (Stephan, 2021). In partitioned cellular automata, a local coarse-grained dynamics exists only when the effective rule satisfies

$\Pi:\mathcal X\to\mathcal Y$7

with locality requiring $\Pi:\mathcal X\to\mathcal Y$8. The analysis shows that deterministic microscopic rules often induce stochastic effective dynamics after coarse-graining because distinct microscopic histories collapse to the same coarse state but evolve to different coarse futures (Costa et al., 2019).

The idea extends even further. In loop quantum gravity, coarse-graining by gauge-fixing each connected subgraph to a loopy effective vertex preserves spin-network entanglement at the kinematical level and under holonomy-generated evolution, which is presented as a bulk-boundary realization of coarse-graining (Chen, 2022). In supervised learning, “data coarse graining” studies feature elimination by relevance to the target and shows that a high-pass scheme can produce a nonmonotonic risk curve with a minimum at intermediate coarse-graining, whereas low-pass elimination is detrimental (Nguyen et al., 18 Sep 2025). In speech speculative decoding, the term PCG is used for acceptance on Acoustic Similarity Groups, with overlap-aware mass splitting and an exactness guarantee at the group level rather than the token level (Yanuka et al., 5 Nov 2025).

A plausible implication is that the unifying content of PCG across domains is neither a particular equation nor a single physical interpretation. It is the insistence that reduction should come with an explicit map, a defined preserved structure, and a clear account of what exactness means after information has been discarded.