Non-Conservative Potential Terms

• Non-conservative potential terms are components in dynamical systems that cannot be derived from a scalar potential, thus violating energy conservation and time-reversal symmetry.
• They enable the modeling of dissipative, frictional, and radiative forces through advanced variational formulations and the decomposition of force fields.
• Applications span nonequilibrium physics, machine learning force fields, and both classical and quantum mechanical systems, providing enhanced simulation capabilities.

A non-conservative potential term is any term contributing to the effective forces or equations of motion of a dynamical system that cannot be derived from a scalar potential function by differentiation, thus explicitly violating energy conservation and, typically, time-reversal symmetry. Such terms manifest in a wide array of physical settings as dissipative, frictional, radiative, gyroscopic, or active forces, as well as in models with curl or non-gradient force fields, and are central to the variational and geometric formulations of contemporary nonequilibrium physics, mechanics, and atomistic simulation.

1. Mathematical Structure and Classification

Non-conservative force fields $F$ on a domain $U \subset \mathbb{R}^n$ are not gradients of any scalar function: $F \neq -\nabla V$ for all $V$. The associated differential $1$-form, known as the work form $\omega = F_i(x)\,dx^i$, admits a unique local decomposition (on star-shaped domains) into an exact (conservative) part and an anti-exact (non-conservative) part via the homotopy operator $H$: $$\omega = d\phi + \psi, \quad \phi=H\omega, \quad \psi=H(d\omega)$$ where $d\phi$ is exact and $\psi$ is anti-exact (i.e., in the kernel of $H$) (Kycia, 13 Jul 2025).

In three or fewer dimensions, classical results such as Darboux’s theorem and the Helmholtz decomposition further organize non-conservative fields:

• In $\mathbb{R}^2$, any non-exact $1$-form (curl force) admits at most two generalized potentials such that $F=-V(x,y)\,\nabla U(x,y)$ (Yavari et al., 22 May 2025).
• In $\mathbb{R}^3$, up to three generalized potentials are required: $F(x) = -V(x)\nabla U(x) - \nabla W(x)$, distinguishing truly three-dimensional curl forces with intrinsic “helicity” (i.e., $\omega \wedge d\omega \neq 0$).

In general, the anti-exact component $\psi$ is hierarchically decomposed via the Frobenius theorem: $\psi = e^{\gamma}(dh + \eta)$ where $h$ and $\gamma$ are “generalized potentials” and $\eta$ is a torsion $1$-form encoding genuine non-integrability of the distribution $\psi=0$ in higher dimensions (Kycia, 13 Jul 2025).

2. Variational Principles and Non-Conservative Potentials

Standard Lagrangian and Hamiltonian formulations, based solely on a conservative potential $V(q)$, are fundamentally inadequate for non-conservative systems. The modern variational approach, introduced and axiomatized by Galley et al., doubles the degrees of freedom and introduces a “nonconservative potential” $K(q_1,q_2, \dot{q}_1, \dot{q}_2, t)$ in the composite action (Galley et al., 2014, Galley, 2012): $$\mathcal{S}[q_1,q_2] = \int_{t_i}^{t_f} dt \left[ L(q_1, \dot{q}_1, t) - L(q_2, \dot{q}_2, t) + K(q_1,q_2,\dot{q}_1,\dot{q}_2,t) \right]$$ All non-conservative effects are generated by $K$, which is antisymmetric under $1 \leftrightarrow 2$ and vanishes for $q_1=q_2$. The resulting equations of motion, after the “physical limit” ($q_1=q_2$), contain a generalized non-conservative force

$$Q(q,\dot q,t) = \left. \frac{\partial K}{\partial q_{-}} - \frac{d}{dt} \frac{\partial K}{\partial \dot q_{-}} \right|_{q_{-}=0}$$

(Galley et al., 2014).

Rayleigh dissipation, memory-friction, the Abraham–Lorentz–Dirac radiation reaction, and novel interaction-induced dissipative binding all fit seamlessly into this structure (Aashish et al., 2016, Lemeshko et al., 2012).

The framework generalizes Noether’s theorem: the presence of $K$ induces a controlled violation of energy and other conservation laws, with explicit nonconservation rates dictated by the structure of $K$ (Galley et al., 2014, Martínez-Pérez et al., 2016).

3. Examples and Applications

Non-conservative potential terms manifest in diverse systems:

• Dissipative Lagrangians: Time-local dissipation is encoded as $K=-q_{-} f(q_{+},\dot q_{+},t)$, reproducing $F = -b\dot q$ for Stokes drag, or $F = -b | \dot q | \dot q$ for nonlinear dissipation (Galley, 2012). For memory-friction, nonlocal $K$ terms arise naturally when integrating out bath degrees of freedom (Galley et al., 2014).
• Electrodynamics: Radiation reaction is generated by a non-conservative potential of the form $K \sim q_{-} \dddot{q}_{+}$, embodying the Abraham–Lorentz force after integrating out the electromagnetic field (Aashish et al., 2016).
• Mechatronic and Control Systems: Fractional calculus-based Lagrangians, incorporating $V_{\rm nc}(q^{(\alpha)})$ (fractional derivative), yield the correct equations for systems with resistors, dampers, or viscoelasticity (Allison et al., 2012).
• Classical and Quantum Stochastic Processes: In nonequilibrium Fokker–Planck dynamics, any drift $b(x)$ is decomposed as $b(x) = -\nabla V(x) + A(x)$; non-conservative $A(x)$ is analogous to an “imaginary vector potential” and enters formal path-integral treatments as a surrogate for magnetism, introducing persistent probability currents even at steady state (Garbaczewski et al., 2023).
• Dissipative Binding and Complex Potentials: Engineered non-conservative potentials, such as those arising from driven open quantum systems, can create effective, generally complex-valued potentials $V(r) - i V_{\rm nc}(r)$ that trap particles at preordained distances, exemplifying dissipative binding via coherent population trapping and spontaneous emission control (Lemeshko et al., 2012).

4. Geometric and Hamiltonian Perspectives

The geometric theory provides a systematic, coordinate-independent view of non-conservative potentials. The work $1$-form and its decomposition into exact and anti-exact parts correspond to the traditional gradient and “curl” (or its generalization in higher dimensions). The anti-exact part is further decomposed via Frobenius’ integrability machinery, delineating the space of generalized potentials necessary to characterize non-conservative content in arbitrary dimensions (Kycia, 13 Jul 2025).

Darboux classification in $\mathbb{R}^2$ and $\mathbb{R}^3$ provides normal forms for curl forces and connects these to auxiliary Hamiltonian systems whose “Hamiltonian” (though not the physical energy) is conserved under the original non-conservative dynamics (Yavari et al., 22 May 2025). Generalizations of the Lagrange and Hamilton formalisms to systems linear in the velocities further allow for the extension of the notions of Hamilton’s equations, Poisson brackets, and Hamilton–Jacobi theory to entirely non-conservative flows, with generalized “energy-like” integrals of motion (Ushveridze, 2022).

5. Diagnostics and Pathologies in Machine Learning Potentials

In atomistic machine learning, non-conservative models (direct-force networks) optionally learn forces $\hat{F}(R)$ directly, eschewing the existence of any scalar potential. The degree of non-conservativity is diagnosed by the asymmetry of the force Jacobian, with the scalar index

$$\lambda(R) = \frac{\| J_{\text{anti}} \|_F}{\| J \|_F}$$

serving as a quantitative indicator ($\lambda=0$ for conservative, $\lambda=1$ for maximally non-conservative). Even tiny $\lambda$ values can produce pathological artifacts in geometry optimization and molecular dynamics: line-search methods may fail, MD shows catastrophic temperature drifts, and no shadow Hamiltonian exists to stabilize finite-step integrators. Hybrid “multi-force” models, combining direct and conservative heads, restore stability and energy conservation while enabling efficient inference via multiple-time stepping (Bigi et al., 2024).

A summary table relating conventional and non-conservative model diagnostics and pathologies:

Model	$\lambda$ (Non-conservativity)	Geometry Optimization	Molecular Dynamics
PET (conservative)	$\sim0$	Robust, convergent	Stable ensembles
PET-NC (direct-force)	$0.004$	Erratic, fluctuating	Rapid heating, unstable
ORB	$0.015$	Erratic, fluctuating	Rapid heating, unstable
SOAP-BPNN-NC	$0.032$	Catastrophic failure	Severe drift
PET-M (hybrid)	$\sim0$	Stable, fast convergence	Stable, fast integration

6. Physical and Theoretical Implications

Non-conservative potential terms generically induce:

• Nontrivial energy flows: No global energy function is conserved; instead, the time-evolution of energy and symmetry currents is governed by the structure of the non-conservative potential (Galley et al., 2014, Martínez-Pérez et al., 2016).
• Breakdown of stationary variational principles: Modified Noether-type theorems quantify the symmetry breaking induced by non-conservative terms.
• Intrinsic irreversibility and instability: In dynamical systems, even infinitesimal non-conservative perturbations can destabilize otherwise convergent flows, as in Nesterov’s ODE, where skew-symmetric (curl) perturbations trigger instability by Floquet theory (Ochoa et al., 22 Jan 2025).
• Extension of classical dynamical structures: Generalized Helmholtz conditions, formulated in terms of semi-basic $1$-forms, provide constructive criteria for the existence and reconstruction of Lagrangians together with non-conservative potentials in second-order systems, including both dissipative and gyroscopic classes (Bucataru et al., 2014).

7. Future Directions and Open Problems

Current research focuses on:

• Full geometric classifications of non-conservative content in high-dimensional force fields, with explicit Frobenius-torsion decompositions (Kycia, 13 Jul 2025).
• Optimization of machine learning potentials that respect, diagnose, or strategically exploit non-conservative contributions while maintaining stable physical predictions (Bigi et al., 2024).
• Extension and quantification of the relationship between non-Hermitian quantum mechanics and stochastic processes with non-conservative drifts, including operator path-integral representations with generalized "magnetic" terms (Garbaczewski et al., 2023).
• Unified variational formulations for complex open systems, nonequilibrium field theories, and quantum-classical correspondences, embedding the non-conservative action functional framework as a generalization of classical physics (Galley et al., 2014, Ushveridze, 2022).

Non-conservative potential terms are now recognized as essential for modeling dissipation, irreversibility, and memory effects in both fundamental and applied settings—bridging gaps between mathematics, classical and quantum physics, materials science, and statistical mechanics.