Equilibrium State Estimation

• Equilibrium State Estimation (ESE) is defined as techniques that derive stationary distributions using iterative reweighting and physical or statistical equilibrium constraints.
• ESE methods apply across domains such as molecular simulation, power systems, and forecasting, ensuring estimates align with underlying dynamics and physical laws.
• ESE techniques address challenges like finite sampling and discretization, enabling unbiased state estimation and improved predictive performance in complex systems.

Searching arXiv for the cited ESE-related papers to ground the article in current paper records. {"query":"(Russo et al., 2020) Iterative trajectory reweighting for estimation of equilibrium and non-equilibrium observables", "max_results": 5} {"query":"(Xu et al., 11 Jun 2026) Once-for-All: Scalable Simultaneous Forecasting via Equilibrium State Estimation", "max_results": 5} {"query":"(Russo et al., 2021) Unbiased estimation of equilibrium, rates, and committors from Markov state model analysis", "max_results": 5} Equilibrium State Estimation (ESE) refers, across several research literatures, to estimation procedures in which the target state is defined by stationarity, physical admissibility, or a latent equilibrium relation rather than by a purely local snapshot fit. In molecular simulation, the term is used explicitly for extracting an equilibrium or stationary distribution from short, unbiased trajectory fragments by self-consistent reweighting (Russo et al., 2020). Recent work also uses ESE as a paradigm for simultaneous forecasting of multiple interacting systems by first estimating an equilibrium state across systems and then forecasting from the gap between the current state and that equilibrium (Xu et al., 11 Jun 2026). In power and energy systems, closely related formulations estimate operating points that satisfy circuit equations, power-flow relations, or posterior-distribution constraints, thereby producing physically meaningful states rather than unconstrained observation fits (Li et al., 2019, Mehryar et al., 2019). This suggests that ESE is not a single universally standardized algorithm, but a family of equilibrium-constrained estimation ideas.

1. Terminological scope and domain-specific meanings

In the molecular-dynamics literature, ESE is the first algorithm in "Iterative trajectory reweighting for estimation of equilibrium and non-equilibrium observables" and denotes extraction of a stationary distribution from many short, unbiased trajectories by iterative trajectory reweighting (Russo et al., 2020). In online latent-variable estimation, the same phrase is used for tracking a moving empirical fixed point $\theta^\circ(S)$ induced by a running statistic $S_t$, so that the online problem is decomposed into a frozen batch equilibrium and a tracking lag (Li et al., 9 May 2026). In simultaneous forecasting, ESE is defined as a framework that models multiple interacting systems as a coupled ensemble, estimates an equilibrium state $ES_t$, and then generates joint forecasts in one pass (Xu et al., 11 Jun 2026).

The term is also used near, but not always identical to, equilibrium-oriented state estimation in power systems. "A Circuit-Theoretic Approach to State Estimation" does not name its method ESE, but it explicitly constructs an estimator whose solution is the equilibrium of a circuit model constrained by measurements and network physics (Li et al., 2019). "A Practical Proposal for State Estimation at Balanced, Radial Distribution Systems" similarly estimates a steady-state operating point through angle-free radial power-flow relations (Moutis et al., 2020).

A persistent source of ambiguity is that, in the wireless QoS game literature, ESE denotes Efficient Satisfaction Equilibrium, not an estimation procedure. There, ESE is the refinement of satisfaction equilibrium at which all players adopt the action which requires the lowest effort for satisfaction, or equivalently satisfy their minimum QoS incurring in the lowest effort [(1112.01317); (Perlaza et al., 2010)]. That usage is conceptually distinct from equilibrium state estimation.

2. Stationarity-based trajectory reweighting in molecular simulation

The most explicit ESE algorithm in the supplied literature addresses the following problem: given many short, unbiased trajectory fragments started from arbitrary, possibly far-from-equilibrium initial configurations, estimate the equilibrium or stationary probability distribution over discretized configuration-space bins (Russo et al., 2020). The central idea is a self-consistency condition based on stationarity. If trajectory weights are chosen so that the resulting ensemble is stationary under the dynamics, then the time-averaged occupancy of bins along those weighted trajectories is the equilibrium distribution.

The algorithm initializes trajectory weights with total weight normalized to $1$, gives each bin equal total initial weight, and divides that bin weight equally among all fragments that start in that bin. It then iterates the following map: trajectory weights $\rightarrow$ bin occupancies $\rightarrow$ time-averaged bin weights $\rightarrow$ new trajectory weights. In words, if a bin $i$ has average probability $p_i$ under the current iterate and $K_i$ trajectories start in that bin, then each starting trajectory from bin $S_t$0 receives weight $S_t$1 in the next iterate. Iteration continues until a user-defined convergence threshold is met.

The formal viewpoint is discrete left stationarity. If $S_t$2 is the equilibrium probability vector and $S_t$3 is the transition matrix in a left stochastic convention, then stationarity is

$S_t$4

The ESE iteration exploits this fixed-point condition without estimating $S_t$5 explicitly. Its target is a self-consistent solution in which

$S_t$6

Several properties are emphasized. The method is unbiased, does not rely on computing transition matrices, and makes no Markov assumption about discretized states. It is explicitly presented as not being an MSM. The only systematic error beyond finite sampling is discretization into bins. The assumptions are that trajectories are unbiased, bins are used only for accounting, the fragments provide overlap and communication between bins, unvisited bins receive probability zero, and total probability remains $S_t$7 throughout the iteration. The procedure can also be repeated after progressively trimming the initial time point from trajectories to reduce dependence on the starting distribution.

Once the stationary distribution $S_t$8 is estimated, standard equilibrium observables follow directly. Bin populations are immediate; free-energy profiles are obtained as

$S_t$9

up to a constant; and any bin-averaged observable satisfies

$ES_t$0

The same overall iterative framework can also be extended to non-equilibrium steady-state distributions, committors, and mean first-passage times via the Hill relation.

The paper validates the scheme on a one-dimensional double-well potential and on a $ES_t$1 atomistic Trp-cage folding trajectory from D. E. Shaw Research. In the double-well system, 32 trajectories of length $ES_t$2 steps were discretized into 130 bins, and the converged estimate matched the exact Boltzmann distribution very closely. In the Trp-cage example, the iterative equilibrium estimate agreed reasonably well with simple long-trajectory counts, although a discrepancy in a rightmost bin was flagged as needing further investigation. The conclusion notes that the procedure is formally equivalent to a power method applied to a non-standard transition matrix that accounts for all available trajectory timescales (Russo et al., 2020).

3. MSM, OOM, transfer-operator, and exact-estimation perspectives

A parallel line of work places equilibrium state estimation within coarse-grained Markov modeling. "Unbiased estimation of equilibrium, rates, and committors from Markov state model analysis" shows that the stationary distribution of the coarse MSM transition matrix is an asymptotically unbiased estimator of the correct coarse equilibrium probabilities, provided the underlying trajectories relax under equilibrium boundary conditions (Russo et al., 2021). A central technical contribution is exact treatment of sliding-window averaging over finite trajectories. For one-step lag,

$ES_t$3

and the equilibrium result is

$ES_t$4

The same paper also extends a reweighting idea into an iterative scheme that starts with uniform initial weights, computes sliding-window averages, builds an MSM, solves for its stationary distribution, redistributes the coarse-state probability among constituent microstates, and repeats until convergence.

"Markov State Models from short non-Equilibrium Simulations - Analysis and Correction of Estimation Bias" studies the bias induced when short trajectories are started from a non-equilibrium empirical distribution $ES_t$5 rather than the equilibrium distribution $ES_t$6 (Nüske et al., 2017). The paper derives the error

$ES_t$7

and shows that the bias decays with lag time through factors $ES_t$8, while also depending on discretization quality. Its correction mechanism is based on observable operator models (OOMs), which reconstruct unbiased equilibrium correlations and stationary probabilities from lag-$ES_t$9 and lag-$1$0 statistics even when the data start far from equilibrium. An important diagnostic consequence is that if OOM-based timescales and corrected MSM timescales disagree strongly, the limiting issue is poor discretization rather than initial-condition bias.

At a more abstract level, "Optimal data-driven estimation of generalized Markov state models for non-equilibrium dynamics" formulates equilibrium and non-equilibrium estimation through transfer operators (Koltai et al., 2018). In the equilibrium reversible case, the transfer operator $1$1 satisfies $1$2, dominant eigenfunctions represent slow relaxation processes, and equilibrium state estimation from finite non-equilibrium data is posed as recovery of the invariant density $1$3 from a reference density $1$4. The correction factor is

$1$5

and the leading eigenfunction at eigenvalue $1$6 of the estimated reference operator provides the needed reweighting. Low-rank generalized Markov state models arise as optimal spectral or singular-value approximations of transfer operators.

A related but broader equilibrium-estimation result appears in "Exact Estimation for Markov Chain Equilibrium Expectations" (Glynn et al., 2014). There the target is not a full stationary density but unbiased estimation of equilibrium expectations $1$7 and even the equilibrium distribution function $1$8. The key estimator is the randomized telescoping sum

$1$9

which is unbiased whenever $\rightarrow$0 and $\rightarrow$1. This work does not use ESE as a named framework, but it belongs to the same program of unbiased recovery of stationary quantities without exact stationary sampling.

4. Physically consistent operating-point estimation in power systems

In power systems, equilibrium-oriented state estimation emphasizes network feasibility and physical balance. "A Circuit-Theoretic Approach to State Estimation" recasts power-system state estimation as an equivalent circuit formulation (ECF) in rectangular coordinates, with state vector

$\rightarrow$2

and network constraints expressed through Kirchhoff’s Current Law,

$\rightarrow$3

Measurements from PMUs, RTUs, line-flow meters, and zero-injection buses are embedded as linear circuit elements or controlled sources. The full topology is thereby built directly into the estimator (Li et al., 2019).

The paper contrasts the conventional weighted least squares formulation

$\rightarrow$4

with the ECF-based formulation, whose objective is quadratic and whose constraints are linear. For the illustrative IEEE 5-bus case, the problem is reduced to a convex quadratic program with objective

$\rightarrow$5

and a closed-form global optimum in the sense that it can be solved directly without iterative nonlinear updates. From a probabilistic viewpoint, the method incorporates a topology-based term and is described as applying prior knowledge, thereby producing a more physics-based estimate than a purely measurement-driven maximum likelihood estimator. The paper does not explicitly call this ESE, but it repeatedly frames the estimator as solving for a physically meaningful solution that satisfies circuit equilibrium and measurement constraints.

"A Practical Proposal for State Estimation at Balanced, Radial Distribution Systems" develops an angle-free steady-state estimator tailored to balanced radial feeders (Moutis et al., 2020). The state consists of bus voltage magnitudes and line active and reactive power flows, not voltage angles. Its linearized radial power-flow relations include

$\rightarrow$6

together with nodal power balance and line-flow equalities. The estimator is written abstractly as

$\rightarrow$7

but the defining feature is that angles do not appear anywhere in the state vector or in the measurement equations. In the reported study on a modified IEEE-123 feeder with 61 load/generation buses, 60 lines, 49 DG units, and base power $\rightarrow$8 MVA, voltage magnitude estimation is described as very accurate, while observability depends on a mix of actual measurements and pseudo-measurements. The paper summarizes measurement-coverage tradeoffs by noting that if nodal measurements are preferred, about $\rightarrow$9 of nodes and $\rightarrow$0 of flows are needed, whereas if edge measurements are preferred, about $\rightarrow$1 of nodes and $\rightarrow$2 of flows are needed.

These formulations show a recurring power-systems interpretation of ESE: the estimate is treated as a feasible operating point consistent with network topology, device equations, and available measurements. This differs from the trajectory-stationarity meaning of ESE in molecular simulation, but both are equilibrium-constrained rather than purely local estimators.

5. Moving-equilibrium, Bayesian, and simultaneous-forecasting formulations

A recent extension of equilibrium reasoning treats estimation itself as equilibrium tracking. "Higher-Order Equilibrium Tracking for EM-Compressible Online Estimation" defines the running statistic

$\rightarrow$3

and the frozen empirical equilibrium

$\rightarrow$4

In local coordinates, the equilibrium branch is $\rightarrow$5, the online iterate is $\rightarrow$6, and the tracking error is $\rightarrow$7. The terminal error decomposes as

$\rightarrow$8

The key batch-to-online transfer theorem states that if

$\rightarrow$9

then the online estimator inherits the batch central limit theorem and the sharp first-order risk constant. The paper develops an $\rightarrow$0-th order equilibrium-jet predictor and an order-$\rightarrow$1 frozen corrector, yielding localized tracking rates $\rightarrow$2, and formalizes EM-compressibility and EM-jet$\rightarrow$3-compressibility as the structural conditions under which the equilibrium response can be evaluated from a retained streaming statistic (Li et al., 9 May 2026).

In future-grid state estimation, dynamic equilibrium notions appear as posterior-distribution tracking rather than fixed operating-point recovery. "State Estimation for Future Energy Grids" models the hidden grid state

$\rightarrow$4

as a Hidden Markov Model and argues that state estimation should approximate the true posterior distribution $\rightarrow$5, not a linearized Gaussian surrogate (Mehryar et al., 2019). The paper studies Belief Condensation Filtering, which approximates the posterior by a mixture of exponential-family densities, and Particle Filtering, which approximates it by weighted particles. On the IEEE 14-bus system, average MSE over 500 runs is reported as

$\rightarrow$6

Although the paper does not define ESE as a named framework, it explicitly motivates real-time state estimation under nonlinear physics and non-Gaussian noise.

"State estimation for aoristic models" uses equilibrium in yet another probabilistic sense (Lieshout et al., 2021). The observed data are interval marks generated by an alternating renewal process in equilibrium, and latent event times are assigned a Markov point process prior. Under the equilibrium assumption, the mark law becomes a mixture of an atom at exact observation and a continuous component for interval-censored observations. Posterior inference for latent points is then performed by Metropolis-Hastings, with the prior choice materially affecting where latent events are inferred to lie inside observed intervals.

"Once-for-All: Scalable Simultaneous Forecasting via Equilibrium State Estimation" uses ESE explicitly for multi-system forecasting (Xu et al., 11 Jun 2026). The ensemble of systems is

$\rightarrow$7

with current state $\rightarrow$8 and equilibrium state $\rightarrow$9. The method first estimates $i$0 from attributes and recent history, iteratively updating a correction vector until the estimated equilibrium and the historical state are cointegrated. It then forecasts the ensemble-level total and allocates that forecast across systems using equilibrium proportions. The forecasting model is written as

$i$1

and the separability claim is

$i$2

The paper reports extensive experiments on synthetic datasets, currency exchange, and COVID-19 spread modeling, states that ESE has linear-time complexity, and reports a $i$3–$i$4 speedup when integrated with conventional predictors.

6. Limitations, recurring issues, and common misconceptions

Across the literature, equilibrium-constrained estimators are consistently limited by coverage, model specification, and the quality of the equilibrium surrogate. In the trajectory-reweighting ESE of molecular simulation, the method requires sufficient data overlap across bins, is vulnerable to overlap problems, depends on binning resolution and finite sampling, assigns zero probability to unvisited bins, and can inherit correlation if fragments are extracted as overlapping windows from one long trajectory rather than from independent datasets (Russo et al., 2020). In MSM-based equilibrium estimation, unbiasedness is asymptotic in trajectory length, so finite-$i$5 estimates can retain initial-state bias, and correct boundary conditions are essential when the target observable is not equilibrium itself (Russo et al., 2021). OOM correction removes bias from non-equilibrium starts, but poor discretization can still dominate, in which case corrected MSMs remain inaccurate even though OOM spectral estimates reveal the failure (Nüske et al., 2017).

In power systems, equilibrium-style estimators are only as general as their physical assumptions. The radial distribution estimator is explicitly restricted to balanced, radial networks, relies on observability through a chosen mix of measurements and pseudo-measurements, uses a linearized approximation, and is intended for minute-scale estimation rather than fast transient supervision (Moutis et al., 2020). The circuit-theoretic formulation is designed to avoid the nonconvexity of standard WLS, but its claims of physical plausibility are tied to the equivalent-circuit model and topology constraints it encodes (Li et al., 2019).

Moving-equilibrium formulations introduce different constraints. The online EM framework is local: the predictor-corrector analysis requires entry into a contraction tube and uses a certified restart initializer to guarantee that the iterate begins inside the equilibrium-tracking regime (Li et al., 9 May 2026). The simultaneous-forecasting ESE assumes an interacting ensemble, needs informative attributes, works better with longer inputs, and estimates a statistical equilibrium rather than a true physical or game-theoretic one (Xu et al., 11 Jun 2026). The future-grid Bayesian framework and aoristic interval-censoring framework, although equilibrium-aware, shift the problem from fixed-point computation to posterior approximation, so their performance depends on posterior family choice, particle budget, or prior point-process specification (Mehryar et al., 2019, Lieshout et al., 2021).

A final misconception concerns nomenclature. In the wireless QoS literature, ESE refers to Efficient Satisfaction Equilibrium, a refinement of satisfaction equilibrium in which the satisfactory profile minimizes effort or cost. That concept belongs to game theory, not state estimation [(1112.01317); (Perlaza et al., 2010)]. The coexistence of these meanings underscores that ESE is a domain-dependent acronym whose interpretation must be fixed by context.