betaFeedback

WorksPhysics and Astronomy

Teleported free energy simulations

Abstract

Escorted free energy simulations reduce dissipation with a flow field that follows the equilibrium distribution. A continuous flow, however, cannot transfer probability between metastable basins. We show that any escort then faces a variance floor set by the change in basin weights, and lift it by adding teleportation: jumps at a rate proportional to the driving rate. We derive the corresponding nonequilibrium identity, the zero-dissipation condition, and the minimal number of jumps it requires, and illustrate the gain on an exactly solvable model and a tilted double well.

Computing the free energy difference ΔF=FB−FA\Delta F = F_B - F_A between two equilibrium states of a system, held at the same temperature β−1\beta^{-1} but at different values λ=A,B\lambda= A, B of an external parameter, is a routine and yet demanding task of molecular simulation. Alongside equilibrium methods such as thermodynamic integration and free energy perturbation, nonequilibrium approaches have become a standard tool: one repeatedly drives λ\lambda from AA to BB in a finite time τ\tau, starting from equilibrium samples, and estimates ΔF\Delta F through the nonequilibrium work relation [1, 2]

e−βΔF=⟨e−βW⟩,(1)e^{-\beta\Delta F} = \langle e^{-\beta W}\rangle, \tag*{(1)}

where WW is the work performed during a single realization and the angular brackets denote an average over realizations.

Equation (1) holds at any switching rate, but its usefulness depends on how fast the exponential average converges. The relevant scale is the dissipated work ⟨W⟩−ΔF≥0\langle W\rangle- \Delta F \ge0. When the system falls out of step with the instantaneous equilibrium distribution, the average becomes dominated by rare realizations with atypically low work, and the number of trajectories needed grows rapidly, typically exponentially, with the dissipated work [3, 4]. Vaikuntanathan and Jarzynski showed that this lag can be attacked at its source [5]. Consider a system with Hamiltonian H(z,λ)H(\boldsymbol{z}, \lambda) and equilibrium distribution peq=e−βH/Zp_{\mathrm{eq}} = e^{-\beta H}/Z, where Z(λ)=∫dz e−βH=e−βFZ(\lambda) = \int d\boldsymbol{z}\,e^{-\beta H} = e^{-\beta F}; we write ZA=Z(A)Z_A = Z(A). One adds to the equations of motion an artificial flow field u\boldsymbol{u} that couples the system coordinates z\boldsymbol{z} directly to changes in λ\lambda,

z˙=v~(z,λ)+λ˙u(z,λ),(2)\dot{\boldsymbol{z}} = \tilde{\boldsymbol{v}}(\boldsymbol{z}, \lambda) + \dot{\lambda}\boldsymbol{u}(\boldsymbol{z}, \lambda), \tag*{(2)}

and corrects the work with the corresponding Jacobian term. The result is an exact identity of the same form as Eq. (1). If u\boldsymbol{u} moves the system exactly as the equilibrium distribution moves, the lag disappears and every trajectory returns W=ΔFW = \Delta F. The idea has since been extended to discrete-time maps [6], to learned flows [7, 8], and to the virtual escorting of unmodified trajectories [9].

An escort, however, is a smooth flow, and a smooth flow can only rearrange probability continuously. This is harmless as long as λ\lambda merely deforms the equilibrium distribution, shifting or reshaping its peaks. It becomes a genuine obstruction when λ\lambda changes the relative population of metastable states separated by high barriers: probability must then be moved from one basin to another. A continuous flow can only do so by sweeping it through a region where the density is essentially zero, at a cost that grows with the barrier, and when the basins are effectively disconnected it cannot do so at all. Tellingly, the double-well benchmark used to illustrate escorted simulations in Refs. [5, 9], Sun’s model [10], is symmetric under q→−qq \to-q. Its two wells always carry the same weight, and the problem never shows up.

In this Letter we lift this restriction by allowing the escort to jump. We supplement the flow with a teleportation kernel, i.e., a jump process whose rate is proportional to λ˙\dot{\lambda}; we derive the corresponding nonequilibrium identity and the condition under which flow and jumps together eliminate the lag. We then show that, whenever the barriers are not crossed during switching, every continuous escort is subject to a variance floor fixed solely by the change in basin weights, and that teleportation removes it, with a minimal number of jumps set by the same weights. An exactly solvable two-basin model and a tilted double well built on Sun’s potential illustrate these results.

We denote by Lλ\mathcal{L}_{\lambda} the generator of the physical dynamics (Hamiltonian, Langevin, or Monte Carlo), which preserves peqp_{\mathrm{eq}} at fixed λ\lambda. As in Ref. [5] we add the flow u\boldsymbol{u}; in addition, we allow the system to jump. During an increment dλd\lambda, a system at z\boldsymbol{z} is teleported with probability aλ(z)dλa_{\lambda}(\boldsymbol{z})d\lambda to a new point z′\boldsymbol{z}' drawn from a landing density qλ(z′∣z)q_{\lambda}(\boldsymbol{z}'|\boldsymbol{z}). The phase-space density f(z,t)f(\boldsymbol{z},t) then evolves according to

∂tf=Lλf−λ˙∇⋅(uf)+λ˙Jλf,(3)\partial_t f = \mathcal{L}_{\lambda}f - \dot{\lambda}\nabla\cdot(\boldsymbol{u}f) + \dot{\lambda}\mathcal{J}_{\lambda}f, \tag*{(3)}
(Jλf)(z)=∫dz′ [kλ(z′→z)f(z′)−kλ(z→z′)f(z)],(4)(\mathcal{J}_{\lambda}f)(\boldsymbol{z}) = \int d\boldsymbol{z}'\,[k_{\lambda}(\boldsymbol{z}'\to\boldsymbol{z})f(\boldsymbol{z}')-k_{\lambda}(\boldsymbol{z}\to\boldsymbol{z}')f(\boldsymbol{z})], \tag*{(4)}
kλ(z→z′)=aλ(z)qλ(z′∣z),Lλe−βHλ=0.(5)k_{\lambda}(\boldsymbol{z}\to\boldsymbol{z}') = a_{\lambda}(\boldsymbol{z})q_{\lambda}(\boldsymbol{z}'|\boldsymbol{z}),\qquad\mathcal{L}_{\lambda}e^{-\beta H_{\lambda}}=0. \tag*{(5)}

Apart from normalization, the kernel is essentially arbitrary: it may be local or global, it need not satisfy detailed balance, and it may depend on λ\lambda in any way. Generators of this type have recently been exploited in generative modeling [11] and in the sampling of discrete distributions [12].

The derivation of an identity for ΔF\Delta F follows Refs. [5, 13] step by step. We define

δHδλ≡∂H∂λ+u⋅∇H−β−1∇⋅u+β−1[∫dz′ kλ(z′→z)eβ[H(z)−H(z′)]−aλ(z)],(6)\frac{\delta H}{\delta\lambda} \equiv\frac{\partial H}{\partial\lambda}+\mathbf{u}\cdot\nabla H-\beta^{-1}\nabla\cdot\mathbf{u} +\beta^{-1}\left[\int\mathrm{d}z'\,k_\lambda(z'\to z)e^{\beta[H(z)-H(z')]}-a_\lambda(z)\right], \tag*{(6)}

and verify by direct substitution that g(z,t)=e−βH(z,λt)/ZAg(z,t)=e^{-\beta H(z,\lambda_t)}/Z_A solves the sink equation ∂tg=Lλg−λ˙∇⋅(ug)+λ˙Jλg−βλ˙(δH/δλ)g\partial_tg=\mathcal{L}_{\lambda}g-\dot{\lambda}\nabla\cdot(\mathbf{u}g)+\dot{\lambda}\mathcal{J}_{\lambda}g-\beta\dot{\lambda}(\delta H/\delta\lambda)g. A second solution follows from the Feynman–Kac theorem, which applies to any Markov process, with or without jumps [13]. Let each trajectory of the modified dynamics, started in equilibrium at λ=A\lambda=A, carry the weight e−βwte^{-\beta w_t}, with

wt=∫0tdt′ λ˙′δHδλ(zt′,λt′).(7)w_t=\int_0^t\mathrm{d}t'\,\dot{\lambda}'\frac{\delta H}{\delta\lambda}(z_{t'},\lambda_{t'}). \tag*{(7)}

The weighted density ⟨δ(z−zt)e−βwt⟩u,k\langle\delta(z-z_t)e^{-\beta w_t}\rangle_{u,k} then obeys the same sink equation: the first three terms move the trajectories, the last one updates their weights. Since the equation is linear and both solutions start from peq(z,A)p_{\mathrm{eq}}(z,A) at t=0t=0, they coincide,

1ZAe−βH(z,λt)=⟨δ(z−zt)e−βwt⟩u,k.(8)\frac{1}{Z_A}e^{-\beta H(z,\lambda_t)}=\langle\delta(z-z_t)e^{-\beta w_t}\rangle_{u,k}. \tag*{(8)}

Setting t=τt=\tau and integrating over zz, we obtain

e−βΔF=⟨e−βW⟩u,k,W≡wτ.(9)e^{-\beta\Delta F}=\langle e^{-\beta W}\rangle_{u,k},\qquad W\equiv w_\tau. \tag*{(9)}

Compared with Ref. [5], the new ingredient is the term in square brackets in Eq. (6), which accounts for the probability removed from zz by jumps (−aλ-a_\lambda) and deposited there from elsewhere (the integral). Since this integral is generally not available in closed form, in practice we use an alternative work functional, obtained by comparing the probability of each trajectory with that of its time reverse, in the spirit of Crooks [2, 14]. Introducing reverse jump rates a~λ\tilde{a}_\lambda and landing densities q~λ\tilde{q}_\lambda, which are again arbitrary up to the usual requirement that every reverse jump can also occur in the forward direction, the work along a trajectory with jumps z−→z+z^-\to z^+ is

βW=β∫dλ[∂λH+u⋅∇H−β−1∇⋅u]−∫dλ[aλ(zλ)−a~λ(zλ)]+∑jumps{β[H(z+)−H(z−)]+ln⁡a(z−) q(z+∣z−)a~(z+) q~(z−∣z+)},(10)\begin{aligned} \beta W={}&\beta\int\mathrm{d}\lambda\left[\partial_\lambda H+\mathbf{u}\cdot\nabla H-\beta^{-1}\nabla\cdot\mathbf{u}\right]\\ &-\int\mathrm{d}\lambda\left[a_\lambda(z_\lambda)-\tilde{a}_\lambda(z_\lambda)\right]\\ &+\sum_{\mathrm{jumps}}\left\{\beta[H(z^+)-H(z^-)]\right.\\ &\left.\qquad+\ln\frac{a(z^-)\,q(z^+|z^-)}{\tilde{a}(z^+)\,\tilde{q}(z^-|z^+)}\right\}, \tag*{(10)} \end{aligned}

and it satisfies the same identity, Eq. (9). Only the Hamiltonian and normalized landing densities are needed; no partition function enters. A derivation, which also yields an identity valid exactly at finite time step, is given in Appendix. Without jumps, Eq. (10) reduces to the escorted work of Ref. [5], and for u=0\mathbf{u}=0 to the ordinary work of Eq. (1); without flow, it is the continuous-time analogue of the incremental weights of sequential Monte Carlo with a backward kernel [15].

As in Ref. [5], the identity is most useful when the modified dynamics keeps the system close to equilibrium. Suppose that flow and jumps together carry peq(⋅,λ)p_{\mathrm{eq}}(\cdot,\lambda) exactly onto peq(⋅,λ+dλ)p_{\mathrm{eq}}(\cdot,\lambda+\mathrm{d}\lambda), so that f=peqf=p_{\mathrm{eq}} solves Eq. (3). Substituting this solution, we find

∂λpeq+∇⋅(upeq)=Jλpeq⟹δHδλ=dFdλ,W=ΔF.(11)\partial_\lambda p_{\mathrm{eq}}+\nabla\cdot(\mathbf{u}p_{\mathrm{eq}})=\mathcal{J}_\lambda p_{\mathrm{eq}} \quad\Longrightarrow\quad \frac{\delta H}{\delta\lambda}=\frac{\mathrm{d}F}{\mathrm{d}\lambda},\qquad W=\Delta F. \tag*{(11)}

In other words, there is no dissipation and a single realization yields the correct free energy difference. The same holds for the trajectory form, Eq. (10), provided the reverse jumps are those of the time-inversed process, a~λ(z′)q~λ(z∣z′)=aλ(z)qλ(z′∣z)peq(z,λ)/peq(z′,λ)\tilde{a}_\lambda(z')\tilde{q}_\lambda(z|z')=a_\lambda(z)q_\lambda(z'|z)p_{\mathrm{eq}}(z,\lambda)/p_{\mathrm{eq}}(z',\lambda): with this choice the jump terms in Eq. (10) vanish identically, a~λ\tilde{a}_\lambda becomes the inflow term of Eq. (6), and the two work functionals coincide trajectory by trajectory. (11) is the analogue, with jumps, of Eq. (15) of Ref. [5]. Its right-hand side acts as a nonlocal source term: it removes probability at one point and creates it at another, which no continuity equation can do.

Why should jumps be needed at all? Consider a system whose configuration space contains metastable basins Ωc\Omega_c, separated by barriers that are not crossed on the time scale τ\tau of the simulation, and let

πc(λ)=∫Ωcdz peq(z,λ).(12)\pi_c(\lambda)=\int_{\Omega_c}\mathrm{d}z\,p_{\mathrm{eq}}(z,\lambda). \tag*{(12)}

be the equilibrium weight of basin cc; we write πA\pi^A and πB\pi^B for the weights at the two end points. A continuous flow field generates a continuous, invertible map of phase space. It could move probability between basins only by dragging it through the barrier region, where the equilibrium density is vanishingly small; excluding this, every escorted trajectory ends in the basin where it started, and the fraction of trajectories in basin cc stays equal to πc(A)\pi_c(A) throughout the process. If the weights change with λ\lambda, the lag cannot be removed, whatever u\mathbf{u} we choose.

This intuition can be made quantitative. Let π^c\hat{\pi}_c be the fraction of trajectories that end in basin cc, for any dynamics, with or without flow and jumps. Integrating Eq. (8) at t=τt=\tau over Ωc\Omega_c gives π^c⟨e−βW∣zτ∈Ωc⟩=e−βΔFπc(B)\hat{\pi}_c\langle e^{-\beta W}\mid z_\tau\in\Omega_c\rangle=e^{-\beta\Delta F}\pi_c(B): the trajectories that end in basin cc must carry, on average, a weight proportional to πc(B)/π^c\pi_c(B)/\hat{\pi}_c. (The same relation holds for the trajectory form of the work, Eq. (10), since the reverse process starts in equilibrium at λ=B\lambda=B; see Appendix.) Applying Jensen’s inequality within each basin, we find

Var⁡[e−βW]⟨e−βW⟩2≥χ2(πB∥π^)=∑cπc(B)2π^c−1.(13)\frac{\operatorname{Var}\left[e^{-\beta W}\right]}{\left\langle e^{-\beta W}\right\rangle^2} \ge\chi^2\left(\pi^B \mathbin\Vert\hat{\pi}\right) = \sum_c \frac{\pi_c(B)^2}{\hat{\pi}_c} - 1. \tag*{(13)}

This is the coarse-grained version of a familiar fact about importance sampling: the relative variance of the weights cannot be smaller than the χ2\chi^2 divergence between the target and the distribution actually sampled at the final time [16]. For a continuous escort without barrier crossings π^=πA\hat{\pi} = \pi^A, and Eq. (13), together with the analogous bound on the mean work obtained by applying Jensen’s inequality to WW itself, yields

Var⁡[e−βW]⟨e−βW⟩2≥χ2(πB∥πA),(14)\frac{\operatorname{Var}\left[e^{-\beta W}\right]}{\left\langle e^{-\beta W}\right\rangle^2} \ge\chi^2\left(\pi^B \mathbin\Vert\pi^A\right), \tag*{(14)}
β⟨W⟩−βΔF≥DKL(πA∥πB).\beta\langle W\rangle- \beta\Delta F \ge D_{\mathrm{KL}}\left(\pi^A \mathbin\Vert\pi^B\right).

These floors, derived in Appendix, depend only on how the basin weights change, not on uu: no amount of ingenuity in the design of the flow can beat them. They are also easy to evaluate. When a basin that is initially almost empty must end up holding a finite fraction of the probability, πc(B)2/πc(A)\pi_c(B)^2/\pi_c(A) is huge, and so is the number of trajectories needed.

Jumps lift the constraint, because a single jump can carry a trajectory from one basin to another. Since each such jump transfers the probability of one trajectory, the mean number ⟨n⟩\langle n\rangle of inter-basin jumps per trajectory obeys

⟨n⟩≥12∑c∣π^c−πc(A)∣,(15)\langle n\rangle\ge\frac{1}{2}\sum_c \left|\hat{\pi}_c-\pi_c(A)\right|, \tag*{(15)}
Var⁡[e−βW]=0 ⇒ ⟨n⟩≥dTV(πA,πB),(16)\operatorname{Var}\left[e^{-\beta W}\right] = 0 \ \Rightarrow\ \langle n\rangle\ge d_{\mathrm{TV}}\left(\pi^A,\pi^B\right), \tag*{(16)}

where the second line follows from Eq. (13), which requires π^=πB\hat{\pi} = \pi^B for zero variance, and dTVd_{\mathrm{TV}} is the total variation distance. A zero-variance estimator thus needs a minimal teleportation activity, fixed once again by the basin weights alone, in the spirit of thermodynamic speed limits involving dynamical activity [17].

These bounds are saturated by a model simple enough to be solved by hand. Take two identical, disconnected basins, LL and RR, and switch off an energy offset on the right one as λ\lambda goes from A=0A = 0 to B=1B = 1:

Hλ(z)=H0(z)+(1−λ)Δ 1R(z),H_\lambda(z) = H_0(z) + (1-\lambda)\Delta\,1_R(z),
πR(λ)=11+eβ(1−λ)Δ.(17)\pi_R(\lambda) = \frac{1}{1+e^{\beta(1-\lambda)\Delta}}. \tag*{(17)}

The right basin starts with weight πR(0)≃e−βΔ\pi_R(0) \simeq e^{-\beta\Delta} and ends with πR(1)=1/2\pi_R(1)=1/2, while the shape of each basin never changes. The best escort is therefore u=0u=0, and the work is W=0W=0 for trajectories in LL and W=−ΔW=-\Delta for those in RR. The resulting variance saturates the floor (14),

Var⁡[e−βW]⟨e−βW⟩2∣escort=sinh⁡2βΔ2,(18)\left.\frac{\operatorname{Var}\left[e^{-\beta W}\right]}{\left\langle e^{-\beta W}\right\rangle^2}\right|_{\mathrm{escort}} = \sinh^2\frac{\beta\Delta}{2}, \tag*{(18)}

which grows as eβΔ/4e^{\beta\Delta}/4: the estimate rests entirely on the rare trajectories that happen to start on the right. Now add teleports from LL to RR, landing on the equilibrium distribution of the right basin,

aλ(z)=βΔ πR(λ) 1L(z),a_\lambda(z)=\beta\Delta\,\pi_R(\lambda)\,1_L(z),
qλ(z′∣z)=peq(z′,λ)πR(λ) 1R(z′).(19)q_\lambda(z'|z)=\frac{p_{\mathrm{eq}}(z',\lambda)}{\pi_R(\lambda)}\,1_R(z'). \tag*{(19)}

The rate is chosen so that the flux of jumps, aλπLa_\lambda\pi_L, equals dπR/dλ=βΔ πLπRd\pi_R/d\lambda=\beta\Delta\,\pi_L\pi_R. One checks directly that Eq. (11) is satisfied, and hence

W=ΔFfor every trajectory,(20)W=\Delta F \quad\text{for every trajectory}, \tag*{(20)}
⟨n⟩=πR(1)−πR(0)=dTV(πA,πB).(21)\langle n\rangle=\pi_R(1)-\pi_R(0)=d_{\mathrm{TV}}\left(\pi^A,\pi^B\right). \tag*{(21)}

The teleport removes the floor completely, and it does so with the smallest number of jumps allowed by Eq. (16). The result is also robust. Rescaling the rate to caλc a_\lambda, so that the kernel is no longer perfect, we computed the variance exactly with the work functional of Eq. (7): for βΔ=10\beta\Delta=10, where the escort gives 5.5×1035.5\times10^3, it is 2.12.1 for c=1/2c=1/2 and 0.550.55 for c=2c=2. Teleports that are wrong by a factor of two still reduce the variance by more than three orders of magnitude.

We now turn to a continuous model in the spirit of Ref. [5]. We freeze Sun’s double well at λ=0\lambda=0 and add a linear tilt that is switched off,

V(q,λ)=q4−16q2+(1−λ)hq,λ:0→1,(22)V(q,\lambda)=q^4-16q^2+(1-\lambda)h q,\qquad\lambda:0\to1, \tag*{(22)}

with β=1\beta=1. The barrier between the wells stays above 55 kBT55\,k_{\mathrm{B}}T throughout, so the wells remain dynamically disconnected. The tilt moves the minima q±(λ)q_\pm(\lambda) only slightly, but it changes the weights of the wells by orders of magnitude: for h=2h=2 the right well holds a fraction πR(0)=1.3×10−5\pi_R(0)=1.3\times10^{-5} of the probability at λ=0\lambda=0 and half of it at λ=1\lambda=1 (Figure 1). Here ΔF=4.978\Delta F=4.978, and Eq. (14) sets a floor χ2=2.0×104\chi^2=2.0\times10^4 for any escort. Unlike Sun’s original model, in which the wells merge at λ=1\lambda=1 and a smooth flow can in principle do the job, here a flow could transfer probability only by dragging it across a barrier of more than 55 kBT55\,k_{\mathrm{B}}T.

Tilted double well and equilibrium distributions

Figure 1. Tilted double well, Eq. (22), with h=2h=2, at λ=0\lambda=0 (solid) and λ=1\lambda=1 (dashed), with the corresponding equilibrium distributions (shaded). The weight of the right well grows from 1.3×10−51.3 \times10^{-5} to 1/21/2; its distribution at λ=0\lambda=0 is shown magnified 10510^5 times. The tilt also shifts the minima slightly, which the escort field compensates. The arrow indicates teleportation.

We simulate overdamped Langevin dynamics with Metropolis-adjusted steps, so that the identity holds exactly at finite time step (Appendix), and vary λ\lambda linearly in a time τ\tau. For the escort we use, in the spirit of Eq. (19) of Ref. [5], a field that moves each well along with its minimum and interpolates smoothly across the barrier over a width w=1/2w=1/2 (dots denote derivatives with respect to λ\lambda),

u(q,λ)=q˙++(q˙+−q˙−)1+tanh⁡(q/w)2,u(q,\lambda)=\dot{q}_+ + (\dot{q}_+ - \dot{q}_-)\frac{1+\tanh(q/w)}{2},
q˙±≡dq±dλ=hV′′(q±).(23)\dot{q}_\pm\equiv\frac{dq_\pm}{d\lambda}=\frac{h}{V''(q_\pm)}. \tag*{(23)}

For the teleport we deliberately use crude ingredients that require no knowledge of ΔF\Delta F. Forward jumps go from the left well to the right one and reverse jumps the other way, with rates set by harmonic estimates of the basin weights, and landing densities given by the harmonic approximation of each well,

aλ=π˙RπLh1q<0,a~λ=π˙RπRh1q>0,(24)a_\lambda=\frac{\dot{\pi}_{R}}{\pi_{L}^{h}}1_{q<0},\qquad\tilde{a}_\lambda=\frac{\dot{\pi}_{R}}{\pi_{R}^{h}}1_{q>0}, \tag*{(24)}
qλ=N(q+,1βV′′(q+)),q~λ=N(q−,1βV′′(q−)),(25)q_\lambda=\mathcal{N}\left(q_{+},\frac{1}{\beta V^{\prime\prime}(q_{+})}\right),\qquad\tilde{q}_\lambda=\mathcal{N}\left(q_{-},\frac{1}{\beta V^{\prime\prime}(q_{-})}\right), \tag*{(25)}
πRh=Z+hZ−h+Z+h,Z±h=e−βV(q±,λ)2πβV′′(q±).(26)\pi_{R}^{h}=\frac{Z_{+}^{h}}{Z_{-}^{h}+Z_{+}^{h}},\qquad Z_{\pm}^{h}=e^{-\beta V(q_{\pm,\lambda})}\sqrt{\frac{2\pi}{\beta V^{\prime\prime}(q_{\pm})}}. \tag*{(26)}

Figure 2(a) shows the estimates for h=2h=2, with N=107N=10^7 trajectories per point for all three methods. About a hundred of them, NπR(0)≃130N\pi_{R}(0)\simeq130, start in the right well, so the rare events that matter are sampled and all estimates are correct on average. The difference lies in the fluctuations. Without teleports the error bars are about 0.040.04 and do not shrink with τ\tau: the measured relative variance of e−βWe^{-\beta W} is 2×1042\times10^4 at every switching time, with and without the escort, right on the floor of Eq. (14). Slower switching does not help, because the barrier is never crossed, and no escort can help either. With teleports the variance drops to between 0.30.3, at the shortest τ\tau, and 8×10−38\times10^{-3}, and the error bars shrink by more than two orders of magnitude; at the longer switching times, an escorted simulation would need more than

Free energy estimates and measured relative variance

Figure 2. (a) Free energy estimates versus switching time τ\tau for the model of Fig. 1 (h=2h=2, N=107N=10^7 trajectories per point): Eq. (1) (open circles), escorted simulations (open squares), and escorted simulations with teleportation (filled circles); dashed line, exact ΔF\Delta F. Error bars are standard errors over 20 batches; for teleported simulations they are smaller than the symbols. (b) Measured relative variance of e−βWe^{-\beta W} versus tilt hh at τ=0.2\tau=0.2, with N=106N=10^6, 10710^7, and 3×1073\times10^7 trajectories for h=1.5h=1.5, 22, and 2.252.25, so that about 10210^2 trajectories start in the right well; dotted line, the floor of Eq. (14), valid for any escort.

10610^6 times as many trajectories to reach the same precision. Figure 2(b) shows that none of this depends on the chosen tilt: as hh grows, the floor rises exponentially and the variance of the plain and escorted estimates follows it, while with teleports it stays around 10−210^{-2}.

When the small well is so rare that no trajectory starts there, the failure becomes systematic. For h=3h=3, πR(0)=4.5×10−8\pi_R(0)=4.5\times10^{-8}, and with N=106N=10^6 trajectories the plain and escorted estimates return, at every τ\tau,

ΔFest⟶ΔF−β−1ln⁡πL(1)=ΔF+β−1ln⁡2,(27)\Delta F_{\mathrm{est}}\longrightarrow\Delta F-\beta^{-1}\ln\pi_L(1)=\Delta F+\beta^{-1}\ln2, \tag*{(27)}

with small but misleading error bars, as follows from the basin relation that leads to Eq. (13) with no trajectory in the right well; teleported simulations give the exact value [18]. The intermediate regime of Fig. 2 shows that this is not a sampling artifact: whenever the rare trajectories are present, the variance of an escorted estimate sits exactly where Eq. (14) puts it.

As in Ref. [5], the key to success is a modified dynamics that mimics the effect of λ\lambda on peqp_{\mathrm{eq}}: escorting when λ\lambda deforms the equilibrium distribution, teleportation when λ\lambda redistributes probability among states that the dynamics cannot connect. The price is that landing densities must be normalized and should resemble the basins they target, much like a proposal in importance sampling. Crude harmonic approximations were enough here; in molecular systems they could come from short equilibrium runs in each basin or from learned generative models [7, 11]. A discontinuous invertible map, admissible in discrete-time escorts [6, 19], can also move probability between basins, but it requires knowing exactly which part of one basin should be mapped onto another.

Jump moves are familiar in sampling, from annealed importance sampling and sequential Monte Carlo [15, 20] to recent schemes that combine transport with jumps or nonlocal proposals [12, 21, 22, 23, 24], and the imbalance of mass between modes is known to control the complexity of annealing [25, 26, 27]. Here we have brought this mechanism into escorted free energy simulations, and identified an escort-independent variance floor together with the minimal teleportation activity needed to remove it. Basin weights could also be estimated separately by reweighting [28]; teleportation folds this step into a single nonequilibrium identity, while the escort handles the deformation within each basin.

Teleportation should be most useful against hidden barriers, for instance when the populations of side-chain rotamers or ligand binding poses differ between the end states of an alchemical transformation, and in discrete systems, where jumps are the only possible escort. It could also help in training generative AI models defined by an energy function: there, each training step requires samples of a model Boltzmann distribution that keeps changing, and Jarzynski weights are used to correct walkers that cannot follow the shifting weights of its modes [29, 30]; teleports would let the walkers follow them. As for the original escort [5], the approach extends directly to Crooks’s fluctuation theorem [2] and, through Eq. (8), to the Hummer–Szabo identity for potentials of mean force [13].

**Data availability—**The code and data that support the findings of this article are openly available in the Supplemental Material.

IA Statement— The author used an AI assistant (Claude, Anthropic) to help polish the text, write and test the simulation code. All content was reviewed and verified by the author, who takes full responsibility for it.

END MATTER

Trajectory form of the work

We derive Eq. (10) as the continuum limit of a discrete-time scheme for which the identity holds exactly at any step size; the same scheme is used in the simulations [18].

Divide the protocol into KK steps λ0=A<⋯<λK=B\lambda_0=A<\cdots<\lambda_K=B and write γk=e−βH(⋅,λk)\gamma_k=e^{-\beta H(\cdot,\lambda_k)}, Zk=∫γkZ_k=\int\gamma_k. Starting from z0∼γ0/Z0\boldsymbol{z}_0\sim\gamma_0/Z_0, step k→k+1k\to k+1 consists of (i) an escort map yk=Mk(zk)\boldsymbol{y}_k=M_k(\boldsymbol{z}_k), a bijection with Jacobian JkJ_k, together with λk→λk+1\lambda_k\to\lambda_{k+1}; (ii) a teleport move xk∼Kk+1(yk→⋅)\boldsymbol{x}_k\sim\mathcal{K}_{k+1}(\boldsymbol{y}_k\to\cdot); (iii) a move zk+1∼Pk+1(xk→⋅)\boldsymbol{z}_{k+1}\sim\mathcal{P}_{k+1}(\boldsymbol{x}_k\to\cdot) of the physical dynamics, which leaves γk+1\gamma_{k+1} invariant. With an arbitrary normalized backward kernel Bk+1\mathcal{B}_{k+1}, such that Kk+1(y→x)>0\mathcal{K}_{k+1}(\boldsymbol{y}\to\boldsymbol{x})>0 whenever γk+1(y)Bk+1(x→y)>0\gamma_{k+1}(\boldsymbol{y})\mathcal{B}_{k+1}(\boldsymbol{x}\to\boldsymbol{y})>0, we define

e−βW=∏k=0K−1γk+1(yk)Jk(zk)γk(zk)γk+1(xk)Bk+1(xk→yk)γk+1(yk)Kk+1(yk→xk).(28)e^{-\beta W}=\prod_{k=0}^{K-1}\frac{\gamma_{k+1}(\boldsymbol{y}_k)J_k(\boldsymbol{z}_k)}{\gamma_k(\boldsymbol{z}_k)}\frac{\gamma_{k+1}(\boldsymbol{x}_k)\mathcal{B}_{k+1}(\boldsymbol{x}_k\to\boldsymbol{y}_k)}{\gamma_{k+1}(\boldsymbol{y}_k)\mathcal{K}_{k+1}(\boldsymbol{y}_k\to\boldsymbol{x}_k)}. \tag*{(28)}

Let wkw_k be the product of the first kk factors. By induction, E[wkf(zk)]=Z0−1∫γkf\mathbb{E}[w_k f(\boldsymbol{z}_k)]=Z_0^{-1}\int\gamma_k f for any test function ff: the first factor turns γk\gamma_k into γk+1\gamma_{k+1} through the change of variables y=Mk(z)\boldsymbol{y}=M_k(\boldsymbol{z}); the second leaves γk+1\gamma_{k+1} unchanged because ∫dy Bk+1(x→y)=1\int\mathrm{d}\boldsymbol{y}\,\mathcal{B}_{k+1}(\boldsymbol{x}\to\boldsymbol{y})=1; and move (iii) preserves γk+1\gamma_{k+1} by assumption. At k=Kk=K,

⟨e−βW⟩=e−βΔF,\left\langle e^{-\beta W}\right\rangle=e^{-\beta\Delta F},
⟨δ(z−zK)e−βW⟩=e−βΔFpeq(z,B).(29)\left\langle\delta(\boldsymbol{z}-\boldsymbol{z}_K)e^{-\beta W}\right\rangle=e^{-\beta\Delta F}p_{\mathrm{eq}}(\boldsymbol{z},B). \tag*{(29)}

exactly, for any KK; only invariance under Pk+1\mathcal{P}_{k+1} is needed, not detailed balance. For δλ=λk+1−λk→0\delta\lambda=\lambda_{k+1}-\lambda_k\to0, take Mk(z)=z+uδλM_k(\boldsymbol{z})=\boldsymbol{z}+\boldsymbol{u}\delta\lambda, so that Jk=∇⋅uδλJ_k=\nabla\cdot\boldsymbol{u}\delta\lambda, and

K(y→x)=(1−a~δλ)δ(x−y)+a~(y)δλ q(x∣y),\mathcal{K}(\boldsymbol{y}\to\boldsymbol{x})=(1-\widetilde{a}\delta\lambda)\delta(\boldsymbol{x}-\boldsymbol{y})+\widetilde{a}(\boldsymbol{y})\delta\lambda\,q(\boldsymbol{x}|\boldsymbol{y}),
B(x→y)=(1−aδλ)δ(y−x)+a~(x)δλ q~(y∣x).(30)\mathcal{B}(\boldsymbol{x}\to\boldsymbol{y})=(1-a\delta\lambda)\delta(\boldsymbol{y}-\boldsymbol{x})+\widetilde{a}(\boldsymbol{x})\delta\lambda\,\widetilde{q}(\boldsymbol{y}|\boldsymbol{x}). \tag*{(30)}

The first factor in Eq. (28) then contributes β[∂λH+u⋅∇H−β−1∇⋅u]δλ\beta[\partial_\lambda H+\boldsymbol{u}\cdot\nabla H-\beta^{-1}\nabla\cdot\boldsymbol{u}]\delta\lambda to βW\beta W. If no jump occurs, only the singular parts of Eq. (30) enter, and the second factor contributes −(a−a~)δλ-(a-\widetilde{a})\delta\lambda; if a jump y→x\boldsymbol{y}\to\boldsymbol{x} occurs, only the regular parts enter, and it contributes β[H(x)−H(y)]+ln⁡[a q(x∣y)/a~ q~(y∣x)]\beta[H(\boldsymbol{x})-H(\boldsymbol{y})]+\ln[a\,q(\boldsymbol{x}|\boldsymbol{y})/\widetilde{a}\,\widetilde{q}(\boldsymbol{y}|\boldsymbol{x})]. Summing over steps gives Eq. (10).

If the backward jumps are the time reversal of the forward ones, a~(x)q~(y∣x)=a(y)q(x∣y)peq(y)/peq(x)\widetilde{a}(\boldsymbol{x})\widetilde{q}(\boldsymbol{y}|\boldsymbol{x})=a(\boldsymbol{y})q(\boldsymbol{x}|\boldsymbol{y})p_{\mathrm{eq}}(\boldsymbol{y})/p_{\mathrm{eq}}(\boldsymbol{x}), every jump term vanishes, and integrating over y\boldsymbol{y} shows that a~(x)\widetilde{a}(\boldsymbol{x}) equals the inflow term of Eq. (6). (10) then reduces to Eq. (7) trajectory by trajectory.

Variance and dissipation floors

Let π^c\widehat{\pi}_c be the probability that a trajectory ends in basin Ωc\Omega_c and ⟨⋅∣c⟩\langle\cdot|c\rangle the corresponding conditional average. Integrating the second relation in Eq. (29), or Eq. (8) at t=τt=\tau, over Ωc\Omega_c gives

π^c⟨e−βW|c⟩=e−βΔFπc(B),(31)\widehat{\pi}_c\left\langle e^{-\beta W}\middle|c\right\rangle=e^{-\beta\Delta F}\pi_c(B), \tag*{(31)}

for any dynamics and either form of the work; in particular π^c>0\widehat{\pi}_c>0 whenever πc(B)>0\pi_c(B)>0. Decomposing the second moment over final basins and using ⟨X2∣c⟩≥⟨X∣c⟩2\langle X^2|c\rangle\ge\langle X|c\rangle^2,

⟨e−2βW⟩≥∑cπ^c⟨e−βW∣c⟩2=e−2βΔF∑cπc(B)2π^c,(32)\left\langle e^{-2\beta W}\right\rangle\ge\sum_c\widehat{\pi}_c\left\langle e^{-\beta W}|c\right\rangle^2=e^{-2\beta\Delta F}\sum_c\frac{\pi_c(B)^2}{\widehat{\pi}_c}, \tag*{(32)}

which, divided by ⟨e−βW⟩2=e−2βΔF\left\langle e^{-\beta W}\right\rangle^2=e^{-2\beta\Delta F}, is Eq. (13). Equality holds if and only if e−βWe^{-\beta W} is constant within each final basin. Similarly, Jensen’s inequality ⟨e−βW∣c⟩≥e−β⟨W∣c⟩\langle e^{-\beta W}|c\rangle\ge e^{-\beta\langle W|c\rangle} and Eq. (31) give $\langle W|c\rangle\ge ΔF−β−1ln⁡[πc(B)/π^c]\Delta F-\beta^{-1}\ln[\pi_c(B)/\hat{\pi}_c], hence

β⟨W⟩−βΔF≥∑cπ^cln⁡π^cπc(B)=DKL(π^∥πB).(33)\beta\langle W\rangle-\beta\Delta F\geq\sum_c\hat{\pi}_c\ln\frac{\hat{\pi}_c}{\pi_c(B)}=D_{\mathrm{KL}}(\hat{\pi}\Vert\pi^B). \tag*{(33)}

A continuous escort generates a homeomorphism of phase space at each step; if neither the flow nor the physical dynamics carries trajectories across the barriers, every trajectory ends in its initial basin, π^=πA\hat{\pi}=\pi^A, and Eq. (14) follows. Without such crossings, basin occupations change only through inter-basin jumps. With ncinn_c^{\mathrm{in}} and ncoutn_c^{\mathrm{out}} the numbers of jumps into and out of basin cc and n=12∑c(ncin+ncout)n=\frac{1}{2}\sum_c(n_c^{\mathrm{in}}+n_c^{\mathrm{out}}), we have π^c−πc(A)=⟨ncin⟩−⟨ncout⟩\hat{\pi}_c-\pi_c(A)=\langle n_c^{\mathrm{in}}\rangle-\langle n_c^{\mathrm{out}}\rangle, so that 12∑c∣π^c−πc(A)∣≤⟨n⟩\frac{1}{2}\sum_c|\hat{\pi}_c-\pi_c(A)|\leq\langle n\rangle, which is Eq. (15). Zero variance requires π^=πB\hat{\pi}=\pi^B, which gives Eq. (16).

References

  1. [1]C. Jarzynski, Nonequilibrium equality for free energy differences, Phys. Rev. Lett. 78, 2690 (1997).DOI
  2. [2]G. E. Crooks, Entropy production fluctuation theorem and the nonequilibrium work relation for free energy differences, Phys. Rev. E 60, 2721 (1999).
  3. [3]C. Jarzynski, Rare events and the convergence of exponentially averaged work values, Phys. Rev. E 73, 046105 (2006).DOI
  4. [4]S. Vaikuntanathan and C. Jarzynski, Dissipation and lag in irreversible processes, Europhys. Lett. 87, 60005 (2009).DOI
  5. [5]S. Vaikuntanathan and C. Jarzynski, Escorted free energy simulations: Improving convergence by reducing dissipation, Phys. Rev. Lett. 100, 190601 (2008).DOI
  6. [6]S. Vaikuntanathan and C. Jarzynski, Escorted free energy simulations, J. Chem. Phys. 134, 054107 (2011).
  7. [7]P. Wirnsberger, A. J. Ballard, G. Papamakarios, S. Abercrombie, S. Racanière, A. Pritzel, D. J. Rezende, and C. Blundell, Targeted free energy estimation via learned mappings, J. Chem. Phys. 153, 144112 (2020).DOI
  8. [8]M. S. Albergo and E. Vanden-Eijnden, NETS: A non-equilibrium transport sampler, arXiv:2410.02711 (2024).arxiv.org/abs/2410.02711
  9. [9]S. Lee and C. Jarzynski, Estimating free energy differences with virtually escorted trajectories, arXiv:2606.30451 (2026).arxiv.org/abs/2606.30451
  10. [10]S. X. Sun, Equilibrium free energies from path sampling of nonequilibrium trajectories, J. Chem. Phys. 118, 5769 (2003).DOI
  11. [11]P. Holderrieth, M. Havasi, J. Yim, N. Shaul, I. Gat, T. Jaakkola, B. Karrer, R. T. Q. Chen, and Y. Lipman, Generator matching: Generative modeling with arbitrary Markov processes, in International Conference on Learning Representations (ICLR) (2025).arxiv.org/abs/2410.20587
  12. [12]P. Holderrieth, M. S. Albergo, and T. Jaakkola, LEAPS: A discrete neural sampler via locally equivariant networks, in Proceedings of the 42nd International Conference on Machine Learning, PMLR, Vol. 267 (2025) p. 23397.arxiv.org/abs/2502.10843
  13. [13]G. Hummer and A. Szabo, Free energy reconstruction from nonequilibrium single-molecule pulling experiments, Proc. Natl. Acad. Sci. U.S.A. 98, 3658 (2001).pubmed.ncbi.nlm.nih.gov/11274384
  14. [14]G. E. Crooks, Nonequilibrium measurements of free energy differences for microscopically reversible Markovian systems, J. Stat. Phys. 90, 1481 (1998).
  15. [15]P. D. Moral, A. Doucet, and A. Jasra, Sequential Monte Carlo samplers, J. R. Stat. Soc. B 68, 411 (2006).
  16. [16]A. Doucet, W. Grathwohl, A. G. D. G. Matthews, and H. Strathmann, Score-based diffusion meets annealed importance sampling, in Advances in Neural Information Processing Systems (NeurIPS) (2022).arxiv.org/abs/2208.07698
  17. [17]N. Shiraishi, K. Funo, and K. Saito, Speed limit for classical stochastic processes, Phys. Rev. Lett. 121, 070601 (2018).DOI
  18. [18]See Supplemental Material, provided as an ancillary file (supplemental.pdf), for technical conditions, details of the two-basin calculation, simulation details, and results for a stronger tilt.
  19. [19]C. Jarzynski, Targeted free energy perturbation, Phys. Rev. E 65, 046122 (2002).pubmed.ncbi.nlm.nih.gov/12005941
  20. [20]R. M. Neal, Annealed importance sampling, Stat. Com- put. 11, 125. (2001).
  21. [21]M. Skreta, T. Akhound-Sadegh, V. Ohanesian, R. Bondesan, A. Aspuru-Guzik, A. Doucet, R. Brekelmans, A. Tong, and K. Neklyudov, Feynman-Kac correctors in diffusion: Annealing, guidance, and product of experts, in Proceedings of the 42nd International Conference on Machine Learning, PMLR, Vol. 267 (2025).arxiv.org/abs/2503.02819
  22. [22]C. Schönle, M. Gabrié, T. Lié​​vre, and G. Stoltz, Sampling metastable systems using collective variables and Jarzynski–Crooks paths, J. Comput. Phys. 527, 113806 (2025).DOI
  23. [23]C. Schönle, D. Carbone, M. Gabrié, T. Lié​​vre, and G. Stoltz, Efficient Monte Carlo sampling of metastable systems using nonlocal collective variable updates, J. Chem. Phys. 164, 154107 (2026).DOI
  24. [24]H. Lee and M. Santana-Gijzen, Sampling from multimodal distributions with warm starts: Non-asymptotic bounds for the reweighted annealed leap-point sampler, arXiv:2512.17977 (2025).arxiv.org/abs/2512.17977
  25. [25]W. Guo, M. Tao, and Y. Chen, Complexity analysis of normalizing constant estimation: from Jarzynski equality to annealed importance sampling and beyond, in International Conference on Learning Representations (ICLR) (2026).arxiv.org/abs/2502.04575
  26. [26]R. Han, G. Iyer, and D. Slepčev, Time-complexity of sampling from a multimodal distribution using sequential Monte Carlo, arXiv:2508.02763 (2025).arxiv.org/abs/2508.02763
  27. [27]A. Agarwal, G. Iyer, A. Jameson, S. Son, and W. Wimmer, Convergence of Langevin AIS for multimodal distributions, arXiv:2604.17526 (2026).arxiv.org/abs/2604.17526
  28. [28]P. Monmarché, General-purpose post-sampling reweighting method for multimodal target measures, arXiv:2602.12027 (2026).arxiv.org/abs/2602.12027
  29. [29]D. Carbone, M. Hua, S. Coste, and E. Vanden-Eijnden, Efficient training of energy-based models using Jarzynski equality, in Advances in Neural Information Processing Systems 36 (NeurIPS 2023) (2023).
  30. [30]D. Carbone, Hitchhiker’s guide on the relation of energy-based models with other generative models, sampling and statistical physics: a comprehensive review, Trans. Mach. Learn. Res. (2025).arxiv.org/abs/2406.13661

Paper details

vlnc:W_r4aka3bb
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 2

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents