🔢 mathematics

Teleported free energy simulations

This paper introduces "teleported" free energy simulations that overcome the variance floor inherent in continuous-flow escort methods by incorporating teleportation jumps, thereby enabling efficient probability transfer between metastable basins and achieving zero dissipation.

Original author: Davide Carbone

Published 2026-10-30

📖 5 min read🧠 Deep dive

CC BY 4.0

Original author: Davide Carbone

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the author. For technical accuracy, refer to the original paper. Read full disclaimer

In the microscopic world of molecules, scientists often need to know the energy difference between two stable states, such as a protein folding into a specific shape or a drug binding to a target. This energy difference, known as free energy, dictates whether a chemical process will happen spontaneously. To find this value, researchers traditionally run computer simulations that slowly nudge a system from one state to another. However, nature is rarely cooperative; molecules often get stuck in deep energy valleys, separated by high hills that they cannot easily cross. When a simulation tries to force a system over these hills too quickly, it fails to capture the rare but crucial paths the molecules might take, leading to wildly inaccurate results. The challenge has long been to find a way to guide these molecular journeys so that the computer can calculate the energy difference efficiently, even when the system is trapped in separate, disconnected regions.

A recent study by Davide Carbone at the Conservatoire National des Arts et Métiers in Paris addresses this exact bottleneck by introducing a new method called "teleported free energy simulations." The research builds on a technique known as "escorted" simulations, where a computer program adds a gentle, artificial flow to the equations of motion. This flow acts like a current in a river, carrying the molecules along with the changing conditions so they stay close to their natural equilibrium. While this works well when the landscape of the molecule simply shifts or reshapes, it hits a hard wall when the landscape contains two distinct valleys that are separated by a barrier too high to cross. In such cases, the smooth flow cannot move probability from one valley to the other without passing through a region where the molecules almost never exist, making the calculation of energy differences impossible to get right.

Carbone's work demonstrates that this limitation is not just a technical hurdle but a fundamental one: as long as the simulation relies only on smooth movement, the error in the calculation will never drop below a certain floor determined by how much the population of the two valleys changes. To break through this barrier, the author proposes adding "teleportation" to the simulation. Instead of only allowing the system to flow smoothly, the method allows the computer to occasionally jump a molecule from one valley to the other at a specific rate. These jumps are not random; they are timed to match the speed at which the external conditions change, ensuring that the molecules are redistributed exactly as nature would require if the barriers could be crossed.

The paper proves mathematically that this combination of smooth flow and controlled jumps eliminates the lag that causes errors. In a simple model involving two disconnected basins, the author shows that without teleportation, the simulation relies entirely on the rare chance that a molecule starts in the correct valley, leading to massive statistical noise. By adding jumps that move molecules from the empty valley to the populated one, the simulation achieves perfect accuracy with zero variance, meaning every single run of the simulation yields the correct answer. The study further establishes that there is a minimum number of jumps required to achieve this perfection, a number that depends solely on how the populations of the two states change from start to finish.

To test these ideas in a more realistic setting, the author simulated a "tilted double well," a model representing a system with two energy valleys separated by a barrier more than fifty-five times the thermal energy of the system. In this scenario, the barrier is so high that the two valleys are effectively disconnected; a smooth flow cannot move molecules between them. The simulation showed that without teleportation, the error in the energy calculation remained stubbornly high, regardless of how slowly the simulation was run. However, when teleportation was introduced, the error dropped by more than two orders of magnitude. Even when the jump rates were not perfectly tuned, the method still reduced the variance significantly, proving that the approach is robust and does not require perfect knowledge of the system's details.

The findings suggest that this method is particularly useful for systems where the populations of different states change drastically, such as when a ligand binds to a protein or when side chains of a molecule rotate into different positions. The author notes that the "landing spots" for these jumps do not need to be known with extreme precision; even rough approximations of the target state are sufficient to dramatically improve the results. This approach effectively folds the difficult task of estimating how many molecules should be in each state into a single, unified calculation, allowing the simulation to handle both the reshaping of energy valleys and the transfer of molecules between them. By combining the gentle guidance of an escort with the decisive action of a teleport, the method offers a powerful new tool for calculating free energy in complex molecular systems where traditional methods fail.

Technical Summary: Teleported Free Energy Simulations

Problem Statement
Computing free energy differences ( $\Delta F$ ) between equilibrium states is a fundamental task in molecular simulation. While nonequilibrium approaches based on the Jarzynski equality (Eq. 1) are standard, their efficiency is limited by the dissipated work. When a system is driven out of equilibrium, the exponential average is dominated by rare trajectories with atypically low work, causing the number of required trajectories to grow exponentially with dissipation.

Previous work by Vaikuntanathan and Jarzynski introduced "escorted" simulations, where an artificial flow field ( $u$ ) is added to the equations of motion to guide the system along the changing equilibrium distribution, theoretically eliminating dissipation. However, this approach relies on a continuous flow. In systems with metastable basins separated by high energy barriers, a continuous flow cannot transfer probability between basins without traversing regions of vanishingly small density. If the relative weights of these basins change during the switching process (e.g., one basin becomes significantly more populated than the other), a continuous escort cannot eliminate the lag. Consequently, the variance of the free energy estimator is bounded by a "variance floor" determined solely by the change in basin weights, regardless of the sophistication of the flow field design.

Methodology
To overcome the limitations of continuous flows, this paper introduces "teleported" free energy simulations. The methodology supplements the continuous flow field with a jump process (teleportation) that allows the system to instantaneously move between metastable basins.

Modified Dynamics: The system evolution is governed by a generator that includes the physical dynamics ( $L_\lambda$ ), a continuous flow term ( $\dot{\lambda}u$ ), and a jump term ( $\dot{\lambda}J_\lambda$ ). During an increment $d\lambda$ , the system at state $z$ is teleported to a new state $z'$ with probability $a_\lambda(z)d\lambda$ , drawn from a landing density $q_\lambda(z'|z)$ .
Nonequilibrium Identity: The author derives a generalized nonequilibrium identity (Eq. 9) for the free energy difference:
$e^{-\beta \Delta F} = \langle e^{-\beta W} \rangle_{u,k}$
where the work functional $W$ (Eq. 10) includes terms accounting for the jump rates and landing densities. This identity holds for any choice of flow and jump kernels, provided the reverse process is defined consistently.
Zero-Dissipation Condition: The paper derives the condition under which the modified dynamics keeps the system exactly on the equilibrium manifold ( $f = p_{eq}$ ). This requires the flow and jumps to satisfy a specific balance equation (Eq. 11), where the jump term acts as a nonlocal source/sink to redistribute probability between basins, a task impossible for continuity equations alone.
Variance Bounds: Using Jensen's inequality, the author establishes that for any continuous escort without barrier crossings, the relative variance of the estimator is bounded below by the $\chi^2$ divergence between the initial and final basin weight distributions (Eq. 14). To achieve zero variance, the teleportation activity must satisfy a lower bound related to the total variation distance between the initial and final basin weights (Eq. 16).

Key Results
The paper validates these theoretical findings using two models:

Exactly Solvable Two-Basin Model: A system with two disconnected basins where an energy offset is switched off.
- Without teleportation, the variance saturates the theoretical floor, growing exponentially with the energy barrier ( $\sim e^{\beta \Delta}$ ).
- With teleportation tuned to the exact change in basin weights, the variance is reduced to zero, and the minimal number of jumps required matches the theoretical lower bound (Eq. 21).
- The method is robust; even with teleportation rates deviating by a factor of two from the optimal, the variance is reduced by more than three orders of magnitude.
Tilted Double-Well Model: A continuous potential (Sun's model with a linear tilt) where the barrier remains high ( $>55 k_B T$ ) throughout the switching, preventing natural barrier crossing.
- The weight of one well changes from $1.3 \times 10^{-5}$ to $0.5$.
- Standard and escorted simulations fail to converge efficiently, with relative variances hitting the theoretical floor ( $\sim 2 \times 10^4$ ) regardless of switching time $\tau$ .
- Teleported simulations, using crude harmonic approximations for jump rates and landing densities (without prior knowledge of $\Delta F$ ), reduce the variance to $10^{-2}$ to $0.3$.
- In the regime where the rare basin is so empty that no trajectories start there ( $N \pi_R(0) < 1$ ), standard methods yield systematically biased results, whereas teleported simulations recover the exact free energy.

Significance and Claims
The paper claims to lift the fundamental restriction of escorted simulations regarding metastable states. By introducing teleportation, the author demonstrates that:

The variance floor imposed by changing basin weights in continuous flows can be completely removed.
A minimal amount of "teleportation activity" (number of jumps) is sufficient to achieve zero-variance estimation, dictated solely by the change in basin weights.
The approach is robust to imperfect knowledge of the target distributions; even crude approximations of basin weights and shapes yield significant variance reduction.

The author positions this method as a bridge between escorted simulations and jump-based sampling techniques (like annealed importance sampling or sequential Monte Carlo). They suggest the method is particularly valuable for systems with "hidden barriers" (e.g., side-chain rotamers or ligand binding poses) and for training generative AI models where the target distribution shifts rapidly. The approach extends naturally to Crooks' fluctuation theorem and the Hummer–Szabo identity.

More like this