pith. sign in

arxiv: 2606.30451 · v1 · pith:MMSMP444new · submitted 2026-06-29 · ❄️ cond-mat.stat-mech · cond-mat.dis-nn

Estimating Free Energy Differences with Virtually Escorted Trajectories

Sangyun Lee , Christopher Jarzynski This is my paper

Pith reviewed 2026-06-30 03:51 UTC · model grok-4.3

classification ❄️ cond-mat.stat-mech cond-mat.dis-nn
keywords free energy estimationfluctuation theoremvirtual control fieldwork distributionsconvergence optimizationnonequilibrium thermodynamicszero-variance estimator
0
0 comments X

The pith

A virtual control field generates infinitely many work-like quantities from the same trajectories, each satisfying the fluctuation theorem for free energy differences.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

In nonequilibrium processes, free energy differences are estimated from the exponential average of work over trajectories, but this often converges slowly. The paper shows how to introduce a virtual control field that leaves the physical dynamics unchanged while defining a family of different work quantities W_θ. Each of these satisfies the same average equality as the standard work. Parameters in the field can be chosen to reduce the variance of the free energy estimate for a fixed set of trajectories. The method identifies conditions that yield a zero-variance estimator.

Core claim

For a system driven irreversibly from equilibrium state A to B, the free energy difference ΔF can be estimated using the work fluctuation theorem. Adding a virtual control field u_θ that does not modify the underlying dynamics allows construction of work-like quantities W_θ such that the average of exp(-W_θ/T) equals exp(-ΔF/T) for any choice of θ. The parameters θ are selected to optimize convergence of the estimate from a given collection of trajectories, and conditions are identified under which the estimator has zero variance.

What carries the argument

The virtual control field u_θ, which redefines work without changing the physical trajectories, to produce an infinite family of equivalent fluctuation-theorem estimators.

If this is right

  • The same set of trajectories can be reused to obtain improved free energy estimates by selecting an optimal θ after the fact.
  • Variance reduction is achieved without generating additional trajectories or altering the simulation protocol.
  • Under conditions where the virtual field makes W_θ independent of the random work fluctuations, the estimator variance reaches zero.
  • The approach borrows from escorted estimation but applies the escorting only virtually to the work definition.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The method could lower the number of trajectories needed in molecular dynamics calculations of binding affinities.
  • Similar virtual adjustments might be applied to other nonequilibrium theorems such as those for entropy production.
  • The zero-variance condition provides a concrete test case in exactly solvable models like driven harmonic oscillators.
  • Links to diffusion models suggest the parameter optimization step could be automated with machine-learning techniques.

Load-bearing premise

The virtual control field does not modify the underlying dynamics of the system.

What would settle it

A simulation in which the average of exp(-W_θ/T) over trajectories deviates from exp(-ΔF/T) for some choice of θ would falsify the claim that all such W_θ obey the fluctuation theorem.

Figures

Figures reproduced from arXiv: 2606.30451 by Christopher Jarzynski, Sangyun Lee.

Figure 1
Figure 1. Figure 1: FIG. 1. Free energy estimation with ( [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2. Free energy estimation for the quartic potential, [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
read the original abstract

For a process in which a system is driven irreversibly from equilibrium state $A$ toward equilibrium state $B$, the free energy difference $\Delta F = F_B-F_A$ can be estimated using the work fluctuation theorem $\langle e^{-W/T}\rangle = e^{-\Delta F/T}$, where $W$ and $T$ denote work and temperature. The estimate often suffers from poor convergence with the number of trajectories used to calculate the average. Borrowing ideas from escorted free energy estimation, and from diffusion models of machine learning, we show how to construct infinitely many work-like quantities, $W_\theta$, that satisfy $\langle e^{-W_\theta/T}\rangle = e^{-\Delta F/T}$, for the same underlying dynamics. Our method involves a virtual control field ${\boldsymbol u}_\theta$ that does not modify these dynamics. We show how to choose parameter values $\theta$ to optimize convergence of the free energy estimate, for a fixed set of trajectories. We identify conditions under which our method provides a zero-variance estimator of $\Delta F$. We use numerical simulations of model systems to illustrate the gains in convergence that our method can achieve.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper claims to construct infinitely many work-like quantities W_θ using a virtual control field u_θ that does not modify the underlying dynamics, such that ⟨e^{-W_θ/T}⟩ = e^{-ΔF/T} holds exactly on the original trajectory measure. It describes optimization of θ for improved convergence of the free energy estimate from a fixed set of trajectories and identifies conditions for zero-variance estimators, illustrated via numerical simulations on model systems.

Significance. If the construction is valid without hidden dynamical assumptions, the approach would allow post hoc optimization over a family of fluctuation-theorem estimators, potentially reducing the number of trajectories needed for accurate ΔF estimates in nonequilibrium simulations. The zero-variance conditions, if explicitly achievable, represent a notable theoretical strength.

major comments (2)
  1. [Derivation of fluctuation theorem for W_θ (likely §2–3)] The validity of ⟨e^{-W_θ/T}⟩ = e^{-ΔF/T} for virtual W_θ requires an exact path-measure cancellation (via Girsanov or equivalent Radon–Nikodym factor) that is absorbed into the definition of W_θ. The manuscript must provide the explicit derivation, including the assumed SDE form (e.g., overdamped Langevin with additive noise) and any divergence or continuity-equation conditions on u_θ, because this is load-bearing for the central claim that the equality holds for arbitrary u_θ on unmodified trajectories.
  2. [Optimization of θ and zero-variance conditions] The optimization procedure for θ and the explicit functional form of u_θ must be stated with sufficient detail to verify that the resulting estimator remains unbiased and that the zero-variance conditions are attainable without circularity or additional sampling; the abstract provides no equations for these steps.
minor comments (2)
  1. [Introduction] Clarify the precise relationship to prior escorted free-energy methods and diffusion-model literature with one or two additional citations in the introduction.
  2. Ensure all notation for W_θ, u_θ, and the parameter set θ is defined before first use in the main text.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their careful review and constructive suggestions. We address each major comment below and will incorporate clarifications into a revised manuscript.

read point-by-point responses

  1. Referee: [Derivation of fluctuation theorem for W_θ (likely §2–3)] The validity of ⟨e^{-W_θ/T}⟩ = e^{-ΔF/T} for virtual W_θ requires an exact path-measure cancellation (via Girsanov or equivalent Radon–Nikodym factor) that is absorbed into the definition of W_θ. The manuscript must provide the explicit derivation, including the assumed SDE form (e.g., overdamped Langevin with additive noise) and any divergence or continuity-equation conditions on u_θ, because this is load-bearing for the central claim that the equality holds for arbitrary u_θ on unmodified trajectories.

    Authors: We agree that the derivation requires greater explicitness. The manuscript assumes the standard overdamped Langevin dynamics with additive white noise and derives the fluctuation theorem for W_θ via the Girsanov change-of-measure factor absorbed into the definition of the virtual work. In the revision we will add a self-contained derivation in §2 that starts from the SDE, states the precise form of the Radon–Nikodym derivative, and lists the required regularity conditions (divergence of u_θ and continuity-equation compatibility) under which the equality holds exactly for any sufficiently regular u_θ on the original trajectory measure. revision: yes

  2. Referee: [Optimization of θ and zero-variance conditions] The optimization procedure for θ and the explicit functional form of u_θ must be stated with sufficient detail to verify that the resulting estimator remains unbiased and that the zero-variance conditions are attainable without circularity or additional sampling; the abstract provides no equations for these steps.

    Authors: We will expand §3 and the methods section to give the explicit parametric form of u_θ(θ) and the post-hoc optimization criterion (minimization of the sample variance of e^{-W_θ/T} subject to the fixed trajectory ensemble). Because the fluctuation-theorem identity holds for every admissible θ, the estimator remains unbiased for any choice; the optimization merely selects the θ that yields the lowest variance on the given data. The zero-variance conditions are derived analytically as the existence of a u_θ that makes W_θ = ΔF almost surely; they do not require extra sampling and will be stated with the corresponding functional equation in the revision. revision: yes

Circularity Check

0 steps flagged

No circularity: virtual W_θ construction extends fluctuation theorem via independent path-measure adjustment

full rationale

The derivation introduces a virtual control field u_θ that leaves the underlying stochastic trajectories unchanged while redefining the work functional W_θ so that the exponential average remains exactly e^{-ΔF/T}. This adjustment is obtained by explicit construction of the modified work (typically via an additive term whose path integral averages to unity under the original measure), not by fitting parameters to the target ΔF or by re-expressing the same quantity under a new name. The zero-variance conditions are stated as explicit requirements on u_θ that can be verified independently of any particular data set. No load-bearing self-citation, uniqueness theorem imported from the authors' prior work, or ansatz smuggled through citation appears in the chain; the result is therefore self-contained against the standard Jarzynski equality.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 1 invented entities

The central claim rests on the standard work fluctuation theorem plus the ability to define a virtual field that does not affect dynamics. The θ parameters are chosen post hoc to optimize convergence and therefore count as free parameters. The virtual field itself is an invented construct whose independent evidence is not supplied in the abstract.

free parameters (1)
  • θ (control-field parameters)
    Chosen to optimize convergence of the free-energy estimate for a fixed set of trajectories.
axioms (1)
  • domain assumption The work fluctuation theorem ⟨e^{-W/T}⟩ = e^{-ΔF/T} holds for the driven irreversible process from A to B.
    This is the foundational identity that all W_θ must satisfy.
invented entities (1)
  • virtual control field u_θ no independent evidence
    purpose: To define alternative work-like quantities W_θ without changing the actual system dynamics.
    New construct introduced to enable the family of equivalent estimators.

pith-pipeline@v0.9.1-grok · 5733 in / 1394 out tokens · 51533 ms · 2026-06-30T03:51:26.807349+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

52 extracted references · 2 canonical work pages · 2 internal anchors

  1. [1]

    For this process, ∆F= 62.94· · ·atT= 1 [29]

    Asλvaries from 0 to 1, these minima approach one another, coalescing at the origin. For this process, ∆F= 62.94· · ·atT= 1 [29]. This model process illustrates the poor convergence of Eq. 1 [29], and the efficiency gains that can be achieved with escorting [4]. Here, we use it to illustrate virtual escorting, whenρ(x, t) is not known analytically. FIG. 2....

  2. [2]

    Jarzynski, Phys

    C. Jarzynski, Phys. Rev. Lett.78, 2690 (1997)

    1997

  3. [3]

    Jarzynski, Phys

    C. Jarzynski, Phys. Rev. E56, 5018 (1997)

    1997

  4. [4]

    Vaikuntanathan and C

    S. Vaikuntanathan and C. Jarzynski, Europhys. Lett.87, 60005 (2009)

    2009

  5. [5]

    Vaikuntanathan and C

    S. Vaikuntanathan and C. Jarzynski, Phys. Rev. Lett. 100, 190601 (2008)

    2008

  6. [6]

    Wirnsberger, A

    P. Wirnsberger, A. J. Ballard, G. Papamakarios, S. Aber- crombie, S. Racani` ere, A. Pritzel, D. Jimenez Rezende, and C. Blundell, J. Chem. Phys.153, 144112 (2020)

    2020

  7. [7]

    Vargas, S

    F. Vargas, S. Padhy, D. Blessing, and N. N¨ usken, in International Conference on Learning Representations (ICLR)(2024)

    2024

  8. [8]

    Zhong, B

    A. Zhong, B. Kuznets-Speck, and M. R. DeWeese, Phys. Rev. E110, 034121 (2024)

    2024

  9. [9]

    M´ at´ e, F

    B. M´ at´ e, F. Fleuret, and T. Bereau, J. Phys. Chem. Lett.15, 11395 (2024)

    2024

  10. [10]

    M´ at´ e, F

    B. M´ at´ e, F. Fleuret, and T. Bereau, J. Chem. Phys.162, 124107 (2025)

    2025

  11. [11]

    J. L. Rosa-Ra´ ıces and D. T. Limmer, J. Chem. Theory Comput.21, 11455 (2025)

    2025

  12. [12]

    J. He, Y. Du, F. Vargas, Y. Wang, C. P. Gomes, J. M. Hern´ andez-Lobato, and E. Vanden-Eijnden, inAdvances in Neural Information Processing Systems (NeurIPS) (2025)

    2025

  13. [13]

    Holdijk, N

    L. Holdijk, N. M. Anand, M. M. Bronstein, and M. Welling, inInternational Conference on Learning Representations (ICLR)(2026)

    2026

  14. [14]

    Schebek, J

    M. Schebek, J. He, E. Hoffmann, Y. Du, F. No´ e, and J. Rogal, J. Chem. Phys.164, 184111 (2026)

    2026

  15. [15]

    Sohl-Dickstein, E

    J. Sohl-Dickstein, E. A. Weiss, N. Maheswaranathan, and S. Ganguli, inInternational Conference on Machine Learning (ICML)(2015)

    2015

  16. [16]

    Y. Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole, inAdvances in Neural Infor- mation Processing Systems (NeurIPS)(2021)

    2021

  17. [17]

    Albergo, N

    M. Albergo, N. M. Boffi, and E. Vanden-Eijnden, J. Mach. Learn. Technol.26, 1 (2025)

    2025

  18. [18]

    Risken, inThe Fokker-Planck equation: methods of solution and applications(Springer, 1989)

    H. Risken, inThe Fokker-Planck equation: methods of solution and applications(Springer, 1989)

    1989

  19. [19]

    Zwanzig,Nonequilibrium Statistical Mechanics(Ox- 6 ford University Press, 2001)

    R. Zwanzig,Nonequilibrium Statistical Mechanics(Ox- 6 ford University Press, 2001)

    2001

  20. [20]

    C. W. Gardineret al.,Handbook of Stochastic Methods (Springer, 1985)

    1985

  21. [21]

    potential

    In the context of overdamped Langevin dynamics, we refer toU(x, t) as a “potential” rather than using the more general term “energy function”

  22. [22]

    J. L. W. V. Jensen, Acta Mathematica30, 175 (1906)

    1906

  23. [23]

    The Fisher divergence is also known as the relative Fisher information, the score- matching divergence, and the Hyv¨ arinen divergence

    Explicitly,D KL(f||g)≡ R fln(f /g) andD F(f||g)≡ (1/2) R f|∇ln(f /g)| 2. The Fisher divergence is also known as the relative Fisher information, the score- matching divergence, and the Hyv¨ arinen divergence. It differs from (but is related to) the Fisherinformation, which involves derivatives with respect to auxiliary pa- rameters (e.g.θ) rather than sta...

  24. [24]

    X. Ding, H. T. Quan, and Y. Tu, (2026), arXiv:2606.17252

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  25. [25]

    Hyv¨ arinen and P

    A. Hyv¨ arinen and P. Dayan, J. Mach. Learn. Res.6 (2005)

    2005

  26. [26]

    Y. Song, S. Garg, J. Shi, and S. Ermon, inUncertainty in Artificial Intelligence Conference (UAI)(2020)

    2020

  27. [27]

    R. W. Zwanzig, J. Chem. Phys.22, 1420 (1954)

    1954

  28. [28]

    (4) asW[x t] = ∆U(x f , xi)−R τ 0 dt˙xt ◦∂ xU(x t, t), especially whenτ s ∼10 −3

    We found that the numerical evaluation ofWwas im- proved by rewriting Eq. (4) asW[x t] = ∆U(x f , xi)−R τ 0 dt˙xt ◦∂ xU(x t, t), especially whenτ s ∼10 −3

  29. [29]

    Efron,The Jackknife, the Bootstrap, and Other Resampling Plans(Society for Industrial and Applied Mathematics (SIAM), 1982)

    B. Efron,The Jackknife, the Bootstrap, and Other Resampling Plans(Society for Industrial and Applied Mathematics (SIAM), 1982)

    1982

  30. [30]

    Oberhofer, C

    H. Oberhofer, C. Dellago, and P. L. Geissler, J. Phys. Chem. B109, 6902 (2005)

    2005

  31. [31]

    Gaussian Error Linear Units (GELUs)

    D. Hendrycks and K. Gimpel, (2016), arXiv:1606.08415

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  32. [32]

    D. P. Kingma and J. Ba, inInternational Conference on Learning Representations (ICLR)(2015)

    2015

  33. [33]

    Frenkel and A

    D. Frenkel and A. J. Ladd, J. Chem. Phys81, 3188 (1984)

    1984

  34. [34]

    Lipman, R

    Y. Lipman, R. T. Q. Chen, H. Ben-Hamu, M. Nickel, and M. Le, inInternational Conference on Learning Rep- resentations (ICLR)(2023)

    2023

  35. [35]

    Klinger and G

    J. Klinger and G. M. Rotskoff, Phys. Rev. E111, 014114 (2025)

    2025

  36. [36]

    Ikeda, T

    K. Ikeda, T. Uda, D. Okanohara, and S. Ito, Phys. Rev. X15, 031031 (2025)

    2025

  37. [37]

    (13),∇ ·u θ =µT∇ 2Sθ −µ∇ 2U

    In Eq. (13),∇ ·u θ =µT∇ 2Sθ −µ∇ 2U

  38. [38]

    Seifert, Rep

    U. Seifert, Rep. Prog. Phys.75, 126001 (2012)

    2012

  39. [39]

    Peliti and S

    L. Peliti and S. Pigolotti,Stochastic Thermodynamics: An Introduction(Princeton University Press, 2021)

    2021

  40. [40]

    D. T. Limmer,Statistical Mechanics and Stochastic Ther- modynamics(Oxford University Press, 2024)

    2024

  41. [41]

    Seifert,Stochastic Thermodynamics(Cambridge Uni- versity Press, 2025)

    U. Seifert,Stochastic Thermodynamics(Cambridge Uni- versity Press, 2025)

    2025

  42. [42]

    B. D. Anderson, Stoch. Process. Their Appl.12, 313 (1982)

    1982

  43. [43]

    J. Ho, A. Jain, and P. Abbeel, inAdvances in Neural Information Processing Systems (NeurIPS)(2020)

    2020

  44. [44]

    Y. Wang, L. Herron, and P. Tiwary, Proc. Natl. Acad. Sci. U.S.A.119, e2203656119 (2022)

    2022

  45. [45]

    M. Xu, L. Yu, Y. Song, C. Shi, S. Ermon, and J. Tang, inInternational Conference on Learning Representations (ICLR)(2022)

    2022

  46. [46]

    In the DM literature the result traces back to rigorous mathematical analysis [41]

    We have derived this result using standard tools of nonequilibrium statistical physics and stochastic thermo- dynamics [37–40]. In the DM literature the result traces back to rigorous mathematical analysis [41]

  47. [47]

    Hummer and A

    G. Hummer and A. Szabo, Proc. Natl. Acad. Sci. U.S.A. 98, 3658 (2001)

    2001

  48. [48]

    Hummer and A

    G. Hummer and A. Szabo, Acc. Chem. Res.38, 504 (2005)

    2005

  49. [49]

    Ge and D.-Q

    H. Ge and D.-Q. Jiang, J. Stat. Phys.131, 675 (2008)

    2008

  50. [50]

    Note that the sumv θ +u θ describes Hamiltonian flow with friction under the HamiltonianH(x, t)

  51. [51]

    time-reversed

    X. Glorot, A. Bordes, and Y. Bengio, inInternational Conference on Artificial Intelligence and Statistics (AIS- TATS)(2011). 7 Appendix A: Virtual escorted free energy estimation and diffusion models The efficacy of virtual escorted free energy estimation (VE for short) hinges on the ability to infer, up to normal- ization, a time-dependent density,ρ(x, t...

    2011

  52. [52]

    imperfect

    Explicit Construction ofL θ Given a choice ofπ θ(x, t), the validity of Eq. (B11) for Hamiltonian or underdamped Langevin dynamics requires an operatorL θ that satisfies Eqs. (B4) and (B6), for some fielduθ, whereLin Eq. (B6) is the evolution operator governing the unescorted evolution. We now constructL θ explicitly, for the underdamped Langevin case. Th...