Generalized Sample Transition Probability (GSTP)

• GSTP is a method that employs coarse-grained Markov chains to estimate intrinsic transition probabilities from biased MD simulation data without relying on an underlying stochastic process.
• It utilizes flexible kernel functions and reweighting techniques to recover unbiased eigenvalues and eigenfunctions, validating its performance on benchmark systems.
• GSTP bridges enhanced sampling in molecular dynamics with broader generalized transition probability frameworks, highlighting its versatility across different mathematical approaches.

In the current arXiv literature, Generalized Sample Transition Probability (GSTP) primarily denotes a procedure for constructing pairwise transition probabilities between sampled states from biased molecular dynamics (MD) simulation data by means of a coarse-grained Markov chain, with the aim of recovering the kinetic information of the corresponding unbiased system (Wang et al., 6 Aug 2025). In the same data corpus, the term also appears in a broader and less domain-specific sense within the Colombeau-Gsponer framework, where generalized transition probabilities are expressed as mean values of generalized almost periodic functions and are extended to operator-induced overlaps in Hilbert spaces and Fock space (Juriaans et al., 2023). These usages share a common concern with transition structure beyond standard settings, but they arise from distinct mathematical programs and should not be conflated.

1. Domain and problem setting

In the MD setting, GSTP is introduced to address a specific limitation of standard simulation practice: accessing transition probabilities between states is crucial for kinetic information such as reaction paths and rates, yet standard MD simulations are hindered by the capacity to visit the states of interest, which motivates the use of enhanced sampling (Wang et al., 6 Aug 2025). Enhanced sampling accelerates exploration, but the resulting trajectories are biased and therefore do not sample from equilibrium; as a consequence, direct computation of kinetic quantities is invalid without correction.

Within this setting, GSTP is defined as a method that uses a coarse-grained Markov chain to estimate the intrinsic pairwise transition probabilities between states sampled from a biased distribution (Wang et al., 6 Aug 2025). A central claim of the construction is that it can recover transition probabilities without relying on an underlying stochastic process and without specifying the form of the kernel function, in contrast with diffusion map methods that require such structure (Wang et al., 6 Aug 2025).

This suggests that GSTP is positioned as a kinetic reconstruction formalism for ensemble data rather than a direct estimator of dynamical propagators from time series. A plausible implication is that its intended use is strongest in situations where enhanced sampling is indispensable but the induced bias would otherwise obstruct spectral or Markovian kinetic analysis.

2. Mathematical construction in biased molecular dynamics

The construction begins from a feature or configuration space partitioned into small cells centered on sampled states, together with a kernel function $K_s$ that is required only to be non-negative and localized (Wang et al., 6 Aug 2025). In the unbiased case, given samples $\{s_i\}$, the generalized transition matrix is written as

$$P^*_{ij} = \frac{ \dfrac{1}{ \sqrt{\rho(s_j)} } K_s(s_i, s_j) [M(s_j)]^{-1/4} } { \sum_k \dfrac{1}{\sqrt{\rho(s_k)}} K_s(s_i, s_k) [M(s_k)]^{-1/4} } .$$

Here $K_s(s_i,s_j)$ is a generic positive-definite kernel, $\rho(s_j)$ is the equilibrium feature-space density at $s_j$, and $M(s_j)$ is the determinant of the position-dependent metric or diffusion matrix (Wang et al., 6 Aug 2025). The kernel requirements are that, for small kernel width $\sigma$, $k(d^2(s_i,s_j);\sigma)\to \delta(s_i-s_j)$, and that the similarity function $d^2(s_i,s_j)$ be locally quadratic so that the kernel is peaked at $\{s_i\}$0 (Wang et al., 6 Aug 2025).

The main GSTP formula is given for the biased case. If the biased simulation produces samples $\{s_i\}$1 with unbiasing weight $\{s_i\}$2, then the transition matrix elements are approximated by

$\{s_i\}$3

where $\{s_i\}$4 is the reweighting factor, i.e. the ratio of unbiased to biased density (Wang et al., 6 Aug 2025). According to the paper, this recovers the transition probabilities as if the samples were drawn from the unbiased equilibrium distribution, independently of the kernel choice or underlying process (Wang et al., 6 Aug 2025).

For the standard diffusion map or Mahalanobis diffusion map setting with a Gaussian kernel in coordinate space, the formula reduces to

$\{s_i\}$5

This reduction is presented as a special case rather than the defining form of GSTP (Wang et al., 6 Aug 2025).

3. Computational workflow and relation to diffusion maps

The algorithmic workflow stated for GSTP consists of four steps: computing the kernel matrix $\{s_i\}$6, forming the weighted numerators

$\{s_i\}$7

normalizing each row by $\{s_i\}$8, and defining $\{s_i\}$9 so that the rows sum to $$P^*_{ij} = \frac{ \dfrac{1}{ \sqrt{\rho(s_j)} } K_s(s_i, s_j) [M(s_j)]^{-1/4} } { \sum_k \dfrac{1}{\sqrt{\rho(s_k)}} K_s(s_i, s_k) [M(s_k)]^{-1/4} } .$$0 (Wang et al., 6 Aug 2025). The output is a GSTP matrix $$P^*_{ij} = \frac{ \dfrac{1}{ \sqrt{\rho(s_j)} } K_s(s_i, s_j) [M(s_j)]^{-1/4} } { \sum_k \dfrac{1}{\sqrt{\rho(s_k)}} K_s(s_i, s_k) [M(s_k)]^{-1/4} } .$$1 whose eigenvectors and eigenvalues are intended to reflect unbiased kinetics even though the data were biased (Wang et al., 6 Aug 2025).

The comparison with diffusion maps is integral to the method’s framing. Diffusion map (DM) and Mahalanobis diffusion map (MDM) are described as spatial techniques that approximate the generator of a diffusion process by constructing a transition matrix on equilibrium samples using a kernel function (Wang et al., 6 Aug 2025). Their applicability depends on the equilibrium distribution and on the kernel’s connection to an underlying stochastic dynamics. GSTP generalizes this picture by allowing calculation of pairwise transition probabilities from biased simulation data without assuming an underlying diffusion process or a specific kernel form (Wang et al., 6 Aug 2025).

Several distinctions are stated explicitly. GSTP permits any kernel, is not limited to Gaussian choices tied to stochastic processes, and does not require identification of the “correct” underlying stochastic differential equation or generator (Wang et al., 6 Aug 2025). It treats sampled points as centers of Voronoi-like cells and defines transitions as coarse-grained moves between such cells (Wang et al., 6 Aug 2025). It also requires no explicit time information, operating on ensemble data analogously to DM and MDM (Wang et al., 6 Aug 2025).

A plausible implication is that GSTP should be understood as a reweighted geometric Markov construction rather than a direct discretization of a prescribed continuous generator. That interpretation is consistent with the emphasis on coarse-graining, reweighting, and spectral recovery.

4. Validation and empirical scope

The validation reported for GSTP covers three model classes in the MD paper (Wang et al., 6 Aug 2025). The first is a 1D harmonic oscillator, where GSTP is compared under unbiased and temperature-biased simulations with $$P^*_{ij} = \frac{ \dfrac{1}{ \sqrt{\rho(s_j)} } K_s(s_i, s_j) [M(s_j)]^{-1/4} } { \sum_k \dfrac{1}{\sqrt{\rho(s_k)}} K_s(s_i, s_k) [M(s_k)]^{-1/4} } .$$2 versus $$P^*_{ij} = \frac{ \dfrac{1}{ \sqrt{\rho(s_j)} } K_s(s_i, s_j) [M(s_j)]^{-1/4} } { \sum_k \dfrac{1}{\sqrt{\rho(s_k)}} K_s(s_i, s_k) [M(s_k)]^{-1/4} } .$$3, using both Gaussian and student $$P^*_{ij} = \frac{ \dfrac{1}{ \sqrt{\rho(s_j)} } K_s(s_i, s_j) [M(s_j)]^{-1/4} } { \sum_k \dfrac{1}{\sqrt{\rho(s_k)}} K_s(s_i, s_k) [M(s_k)]^{-1/4} } .$$4-kernels (Wang et al., 6 Aug 2025). The reported result is that recovered eigenvalues and eigenfunctions, identified with Hermite polynomials, agree with analytical results, and that GSTP with reweighting accurately matches the unbiased case even with non-Gaussian kernels (Wang et al., 6 Aug 2025).

The second benchmark is alanine dipeptide in vacuum, comparing GSTP from plain MD and from well-tempered metadynamics with bias applied in the backbone torsion variables $$P^*_{ij} = \frac{ \dfrac{1}{ \sqrt{\rho(s_j)} } K_s(s_i, s_j) [M(s_j)]^{-1/4} } { \sum_k \dfrac{1}{\sqrt{\rho(s_k)}} K_s(s_i, s_k) [M(s_k)]^{-1/4} } .$$5 (Wang et al., 6 Aug 2025). Both coordinate-based kernels with atom weights and feature-space kernels using torsion angles and periodic similarities via $$P^*_{ij} = \frac{ \dfrac{1}{ \sqrt{\rho(s_j)} } K_s(s_i, s_j) [M(s_j)]^{-1/4} } { \sum_k \dfrac{1}{\sqrt{\rho(s_k)}} K_s(s_i, s_k) [M(s_k)]^{-1/4} } .$$6 are considered (Wang et al., 6 Aug 2025). The leading GSTP eigenvectors, denoted $$P^*_{ij} = \frac{ \dfrac{1}{ \sqrt{\rho(s_j)} } K_s(s_i, s_j) [M(s_j)]^{-1/4} } { \sum_k \dfrac{1}{\sqrt{\rho(s_k)}} K_s(s_i, s_k) [M(s_k)]^{-1/4} } .$$7 and $$P^*_{ij} = \frac{ \dfrac{1}{ \sqrt{\rho(s_j)} } K_s(s_i, s_j) [M(s_j)]^{-1/4} } { \sum_k \dfrac{1}{\sqrt{\rho(s_k)}} K_s(s_i, s_k) [M(s_k)]^{-1/4} } .$$8, are reported to be consistent between plain MD and unbias-corrected metadynamics data (Wang et al., 6 Aug 2025).

The third benchmark is met-enkephalin in water, where GSTP is applied to two separate enhanced sampling datasets, one from metadynamics and one from TAMD/d-AFED with different collective-variable implementations (Wang et al., 6 Aug 2025). The reported outcome is that the GSTP-generated kinetics, represented by eigenvectors, eigenvalues, and free-energy surfaces in the slowest-variable space, are consistent across both sampling methods (Wang et al., 6 Aug 2025).

These examples are used to support the claim that GSTP effectively recovers the unbiased eigenvalues and eigenstates from biased data and that it is robust with respect to both kernel choice and enhanced-sampling protocol (Wang et al., 6 Aug 2025).

5. Broader meanings of generalized transition probability and the place of GSTP

The acronym GSTP also appears in a broader framework developed for generalized functions. In "Transition Probabilities and Almost Periodic Functions" (Juriaans et al., 2023), generalized transition probability is formulated in the Colombeau-Gsponer setting, where generalized numbers are represented by equivalence classes of moderate nets and transition quantities are defined by averaged overlaps in Hilbert space. In that context, for generalized vectors $$P^*_{ij} = \frac{ \dfrac{1}{ \sqrt{\rho(s_j)} } K_s(s_i, s_j) [M(s_j)]^{-1/4} } { \sum_k \dfrac{1}{\sqrt{\rho(s_k)}} K_s(s_i, s_k) [M(s_k)]^{-1/4} } .$$9 and a net of operators $K_s(s_i,s_j)$0, the generalized transition probability is given by

$K_s(s_i,s_j)$1

The paper states that, in its broader framework, generalized transition probability extends to the GSTP, where a net of almost periodic scalar functions or operator-induced scalar overlaps is assigned a generalized mean value of this form (Juriaans et al., 2023). The same work emphasizes existence results for moderate nets $K_s(s_i,s_j)$2, including selfadjoint Hilbert-Schmidt operators, and discusses possible relevance to Fock space when spectra involve pure infinities or infinitesimals (Juriaans et al., 2023).

This usage differs fundamentally from the MD construction. The MD GSTP is a coarse-grained Markov matrix built from biased samples and reweighting factors (Wang et al., 6 Aug 2025), whereas the Colombeau-Gsponer GSTP is a generalized-number-valued mean of overlap magnitudes for generalized operators and states (Juriaans et al., 2023). The common term “transition probability” therefore masks a substantive difference in both ontology and technical machinery: one is a kinetic estimator on sampled state space, the other a generalized-functional notion of transition amplitude averaging.

For context, the broader literature on generalized transition probability also includes convex-operational and quantum-logical formulations in which transition probability is defined on minimal extreme points, atoms, or generalized state spaces, with symmetry emerging only under additional structural assumptions (Niestegge, 2023, Niestegge, 2022). These works are not about GSTP in the MD sense, but they situate the phrase “generalized transition probability” within a larger mathematical landscape.

6. Interpretation, scope, and terminological cautions

GSTP in the MD literature is presented as a general framework for analyzing kinetic information in complex systems, where biased simulations are necessary to access longer timescales (Wang et al., 6 Aug 2025). Its scope explicitly includes enhanced sampling methods such as metadynamics, umbrella sampling, TAMD, and d-AFED, provided that biasing weights are available (Wang et al., 6 Aug 2025). It is also described as revealing spectral information such as eigenmodes and timescales directly from biased data, with possible use in variational or machine-learning frameworks for collective variables (Wang et al., 6 Aug 2025).

At the same time, the available sources warrant terminological caution. The expression “Generalized Sample Transition Probability” is used explicitly for the biased-simulation method of Wang and collaborators (Wang et al., 6 Aug 2025), but the same acronym is also invoked in the details supplied for the Colombeau-Gsponer paper as a further generalization of mean-value-based transition probabilities (Juriaans et al., 2023). Since these are distinct constructions, the acronym is context dependent.

A common misconception would be to treat GSTP as a single, unified formalism across generalized quantum theory, generalized functions, and molecular simulation. The sources do not support that interpretation. Rather, they support two separate uses of the term: a specific MD methodology for unbiased kinetic recovery from biased ensembles (Wang et al., 6 Aug 2025), and a broader generalized-function extension of transition probability based on Colombeau mean values (Juriaans et al., 2023). Any cross-domain connection beyond that is interpretive rather than explicitly established.

In present usage on arXiv, the most concrete and technically developed meaning of GSTP is therefore the Markov-chain-based reconstruction of unbiased pairwise transition probabilities from biased MD samples (Wang et al., 6 Aug 2025). The broader generalized-function usage remains mathematically distinct and belongs to a different line of research (Juriaans et al., 2023).