betaFeedback

WorksPhysics and Astronomy

Data-Driven Conservation Laws for Newtonian Systems: Physics-Informed Neural Networks and Eisenhart-Duval Geometry

Abstract

We introduce a Physics-Informed Neural Network approach for the data-driven reconstruction of the symmetries and conservation laws of Newtonian dynamical systems, based on the geometrization of the dynamics through the Eisenhart-Duval lift. Within the Eisenhart-Duval lift, the symmetry conditions reduce to the linear Killing or Conformal Killing equations of the extended metric. We combine a Hamiltonian neural network, which reconstructs the potential function from the data, with a random-feature neural network, which solves the symmetry conditions for the collineations of the extended geometry. From the learned vector fields we construct the Lie algebra and, through Noether's theorem, the corresponding conserved quantities, while the derivation of the curvature invariants for the extended geometry provide allow us to perform an independent consistency test. We apply the method to the one-dimensional Ermakov-Pinney equation and to the one-dimensional harmonic oscillator, with and without a quartic perturbation, by using different level of noise in the phase-space data. For the Ermakov-Pinney equation, the three Killing vectors of the flat extended plane, and the corresponding nonlocal conservation laws, are recovered for, while for the harmonic oscillator, the ten conformal Killing vectors which form the so(3,2)so(3,2) Lie algebra are determined for noise levels up to 10%10\%. When the quartic term is added, the broken symmetries are identified as approximate symmetries.

INTRODUCTION

Newton’s second law expresses the dynamics of physical systems as second-order ordinary differential equations. The solutions of these equations describe the motion in a wide range of physical systems, from celestial mechanics to molecular dynamics, and allow us to make predictions about their evolution. The simplest systems in Newtonian mechanics are the free particle, described by the equation q¨=0\ddot{q}=0, and the harmonic oscillator, whose equation of motion q¨+ω2q=0\ddot{q}+\omega^{2}q=0 describes small oscillations around a stable equilibrium. Here q=q(t)q=q(t) is the dependent variable which describes the position of the particle, tt is the independent variable which determines the time parameter, ω2\omega^{2} is the frequency of the oscillator and a dot represents the derivative with respect to the independent variable.

The solution of the free particle is a linear function of time, while the solution of the oscillator is expressed in terms of the trigonometric functions, which are defined by infinite power series. Despite the different form of their solutions, the two systems are equivalent, that is, they are related by a point transformation, and both admit the maximal number of Lie point symmetries [1–3]. This equivalence is revealed by the symmetries of the equations of motion, and it is a first indication that the symmetry structure of a dynamical system carries information which is not evident from the equations themselves.

A systematic method for the study of differential equations is the symmetry analysis introduced by Sophus Lie [4–6]. A Lie symmetry vector is the generator of a one-parameter group of transformations which leaves the differential equation invariant, that is, it maps solutions into solutions. The main novelty of the Lie symmetry analysis is that the invariants of the symmetry vectors can be used to simplify the differential equation, by reducing the order of an ordinary differential equation, or the number of independent variables for the case of partial differential equations [7–10]. Therefore, the solution of the dynamical system can be expressed in algebraic form or by quadratures, or the system can be transformed into the form of another well-known system, as in the relation between the oscillator and the free particle discussed before. Thus, symmetries have been used to classify dynamical systems according to the admitted group properties [11].

For dynamical systems which follow from a variational principle, the Lie symmetries which also leave the action integral invariant, up to a boundary term, are characterized as Noether symmetries [12] and they are the origin of conserved quantities, the Noether charges. For a recent review on Noether’s theorems we refer the reader to [13]. Conservation laws are fundamental in the study of dynamical systems since they define the surfaces within the phase space on which the motion takes place, they permit the reduction of the dynamical system, and they reveal information about its integrability properties. Since conservation laws follow from symmetries through Noether’s theorem, symmetry analysis is an essential tool for the study of nonlinear differential equations, which is why it has been widely used for the study of nonlinear phenomena, in physical science and beyond, see for instance [14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28] and references therein.

The determination of Lie and Noether symmetries requires the solution of a system of partial differential equations [11], the symmetry conditions, which depend explicitly on the form of the equations of motion. For Newtonian dynamical systems these conditions have a geometric interpretation. Indeed, the symmetry vectors of the equations of motion of a particle moving in a Riemannian space are related to the collineations of the kinetic metric, while the potential function introduces additional constraints [29]. On the other hand, the Noether symmetries for the Newtonian systems with a variational principle are related to the admitted elements of the homothetic algebra of the flat space where the motion occurs [30], or to the elements of the conformal algebra for the background geometry in the case of constrained systems with zero value for the energy [31]. These geometric characteristics become clear in the framework of the Eisenhart-Duval lift [35, 36], where the trajectories of the dynamical system are described by the equivalent set of geodesic equations for an extended manifold, where the potential component has been absorbed in the new extended metric tensor. Therefore, the Lie and Noether symmetries are directly related to the collineations of the extended geometry without any additional constraint. The collineations can be used for the construction of conservation quantities for the geodesic system [32, 33, 34].

In the Eisenhart-Duval approach, the geometrization is achieved by introducing additional degrees of freedom [35], which absorb the potential function of the dynamical system. The resulting equivalent system is described by the geodesic equations of a higher-dimensional Riemannian or Lorentzian manifold. In order to recover the original system, the extended geometry must admit Killing vector fields along the new coordinates. The corresponding conservation laws constrain the evolution of the new degrees of freedom, which is necessary in order to map the geodesics to the solutions of the original system. In [35], Eisenhart introduced the method by adding one new degree of freedom, which leads to the Riemannian lift. Later, in [36], a lift was formulated with two additional degrees of freedom, such that the extended geometry is a Lorentzian spacetime; this is known as the Lorentzian lift. Nevertheless, the lift is not unique, and other formulations can lead to new results [37]. The Eisenhart-Duval lift has been applied to the determination of conservation laws in a wide range of dynamical systems, and it has been shown that it reveals hidden symmetries and conservation laws of the original system [38]. Furthermore, the Eisenhart-Duval lift provides a different direction for the geometric linearization of dynamical systems which fail to satisfy the Lie symmetry conditions for linearization. Indeed, by using the Eisenhart-Duval lift, it was shown in [39] that a large class of gravitational models share a common solution space, that of the free particle in a flat extended geometry, such that each physical system corresponds to a different representation of the common solution. For other applications we refer the reader to [40, 41, 42, 43, 44, 45].

Machine learning techniques have become an important tool for the study of dynamical systems from the data. Physics-Informed Neural Networks [46] are of special interest since they force the network to satisfy the given physical laws of motion [47, 48, 49, 50, 51, 52, 53, 54]. A first application of PINNs is to Hamiltonian mechanics, introduced in [55], where Hamilton’s equations were enforced to reconstruct the Hamiltonian function, and consequently, the conservation law of energy for dynamical systems directly from the data. The Lagrangian formalism within a PINN was introduced later in [56]. Moreover, in [57], conserved quantities have been determined from trajectories without the knowledge of the equations of motion. As far as the admitted symmetries are concerned, there are various approaches for the reconstruction of symmetries directly from the data [58, 59].

The reconstruction of the symmetries and of the conservation laws is performed through the introduction of the residual within the loss function. In this work we introduce a new framework, by employing the Eisenhart-Duval framework for the derivation of symmetries and conservation laws of Newtonian dynamical systems. Within the Eisenhart-Duval lift, the derivation of the symmetries is equivalent to the derivation of the collineations for the extended geometry. This is a linear problem, and can be approached with a linear model such as the random-featured expansion described in [60]. The structure of the paper is as follows.

In Section 2 we present the basic properties and elements of the Eisenhart-Duval lift, for the geometric description of dynamical systems. Moreover, we introduce the symmetry structure of Riemannian and Lorentzian manifolds, and specifically the Killing symmetries, the Homothetic symmetries and the Conformal Killing symmetries, which are related through Noether’s theorem to conservation laws for the massive and the massless (null) geodesic Lagrangians.

We review previous results and we demonstrate the application of the Eisenhart-Duval lift for the derivation of local and hidden conservation laws. Applications to Newtonian Mechanics are presented. In § 4 we introduce the main approach which we employ for the reconstruction of conservation laws for dynamical systems. In particular, we introduce a physics-informed neural network that learns the potential function of a Newtonian Hamiltonian system directly from trajectory data. The learned potential is used to construct the corresponding Eisenhart-Duval metric, in which the dynamical system is described by the geodesic equations. We examine the geometric structure of the extended metric by calculating curvature invariants and tensors, namely the Ricci scalar, the Cotton-York tensor and the Weyl tensor, which reveal important information about the characteristics of the extended geometry. We then introduce a random-feature network that determines the symmetry vector fields from a linear null-space problem, along with a bootstrap criterion that selects the number of symmetries, and we extract the Lie algebra and the corresponding conserved quantities by applying Noether’s theorem for the geodesic Lagrangian. In § 5 we demonstrate the application of the networks for the data-driven determination of conservation laws, using mock data. Specifically, we consider noisy trajectories for the one-dimensional Ermakov-Pinney equation and the one-dimensional harmonic oscillator, two models which lead to maximally symmetric geometries after the application of the Eisenhart-Duval lift. Finally, in § 6 we summarize our results and we draw our conclusions.

Preliminaries

In this section we briefly discuss the main geometric framework used in this study. Specifically we introduce the collineations which provides conservation laws for the geodesic equations, and the Eisenhart-Duval lift for the geometrization of Hamiltonian systems.

Symmetries of the Metric and of the Geodesic Equations

We briefly introduce the symmetries for the metric tensor and how they can be used to construct conservation laws for the geodesic equations.

Symmetries of the Metric

Consider a Riemannian manifold endowed with the metric tensor gμν(xμ)g_{\mu\nu}(x^\mu), then we shall say that the generator ξμ=ξμ(xκ)\xi^\mu=\xi^\mu(x^\kappa) of the infinitesimal one parameter point transformation x′μ=xμ+ϵξμ(xκ)x^{\prime\mu}=x^\mu+\epsilon\xi^\mu(x^\kappa) leaves invariant the length, distances, and angles of geometric objects if and only if [61]

Lξgμν=0.(1)\mathcal{L}_{\xi}g_{\mu\nu}=0. \tag*{(1)}

where Lξ\mathcal{L}_{\xi} denotes the Lie derivative with respect to the vector field ξμ\xi^\mu, which for a second-rank tensor field reads

Lξgμν=gμν;κξκ+gμκξκ,ν+gκνξκ,μ.\mathcal{L}_{\xi}g_{\mu\nu}=g_{\mu\nu;\kappa}\xi^\kappa+g_{\mu\kappa}\xi^\kappa{}_{,\nu}+g_{\kappa\nu}\xi^\kappa{}_{,\mu}.

If condition (1) is true, ξμ\xi^\mu acts as a symmetry of the metric tensor and it is characterized as a Killing vector (KV). For the two dimensional Euclidean space, the Killing symmetries form the e(2)\mathfrak{e}(2) Lie algebra, also known as Bianchi VII0_0 algebra, while for the three-dimensional Euclidian space the Killing vectors form the e(3)\mathfrak{e}(3) Lie algebra. The four-dimensional Minkowski geometry is invariant under the ten elements of the Poincaré group. Killing symmetries characterize the geometry and reveal important features when they exist.

The admitted number of Killing symmetries in Riemannian spaces is not arbitrary; two-dimensional geometries can admit zero, one or three Killing vector fields, with the latter being true for the maximally symmetric spaces. These are the e(2)\mathfrak{e}(2) for the flat space, as we discussed before, and the so(3)\mathfrak{so}(3), so(2,1)\mathfrak{so}(2,1) for the space of constant curvature. In general the nn-dimensional maximally symmetric spaces possess 12n(n+1)\frac{1}{2}n(n+1) Killing symmetries, which form the e(n)\mathfrak{e}(n) or so(n+1)/so(n,1)\mathfrak{so}(n+1)/\mathfrak{so}(n,1) Lie algebras respectively.

Conformal killing vectors (CKVs) are the generators of infinitesimal one parameter point transformations which do not preserve the length and the distances, but they preserve the angles. They are defined as [61]

Lξgμν=2ψ(xκ)gμν,ψ(xκ)=1nξμ;μ.(2)\mathcal{L}_{\xi}g_{\mu\nu}=2\psi\left(x^\kappa\right)g_{\mu\nu},\quad\psi\left(x^\kappa\right)=\frac{1}{n}\xi^\mu{}_{;\mu}. \tag*{(2)}

We remark that when ψ(xκ)=0\psi\left(x^\kappa\right)=0, the Killing equation (1) is recovered, which means that KVs are particular cases of the CKVs. In general if ψ(xκ)≠0\psi\left(x^\kappa\right)\ne0, the field ξμ\xi^\mu is called a proper CKV, while in the case where ψ(xκ)\psi\left(x^\kappa\right) is a constant, the field ξμ\xi^\mu is characterized as a Homothetic killing vector (HKV).

The admitted algebra for the CKVs characterizes the Riemannian manifold. As with the case of KVs, the maximum admitted algebra of CKVs has dimension 12(n+1)(n+2)\frac{1}{2}(n+1)(n+2), for n>2n>2, the so(n+1,1)\mathfrak{so}(n+1,1), admitted by conformally flat spaces. Therefore, the three-dimensional flat space admits ten CKVs, which form the so(4,1)\mathfrak{so}(4,1) Lie algebra. On the other hand, two-dimensional spaces admit an infinite number of CKVs, while the one-dimensional space does not admit any proper CKV.

A Riemannian geometry is conformally flat if and only if, for n>3n>3, it has zero Weyl tensor [62]

Cμνκλ=Rμνκλ−2n−2(gμ[κRλ]ν−gν[κRλ]μ)+2(n−1)(n−2)Rgμ[κgλ]ν,(3)C_{\mu\nu\kappa\lambda}=R_{\mu\nu\kappa\lambda}-\frac{2}{n-2}\left(g_{\mu[\kappa}R_{\lambda]\nu}-g_{\nu[\kappa}R_{\lambda]\mu}\right)+\frac{2}{(n-1)(n-2)}Rg_{\mu[\kappa}g_{\lambda]\nu}, \tag*{(3)}

while for n=3n=3, the components of the Cotton-York tensor [62]

Cμνκ=Rνμ;κ−Rνκ;μ+14(R;μgνκ−R;κgνμ),n=3,(4)C_{\mu\nu\kappa}=R_{\nu\mu;\kappa}-R_{\nu\kappa;\mu}+\frac{1}{4}\left(R_{;\mu}g_{\nu\kappa}-R_{;\kappa}g_{\nu\mu}\right),\quad n=3, \tag*{(4)}

vanish. Finally, all two-dimensional Riemannian manifolds are always conformally flat.

Conservation laws for the Geodesic Equations

It is well known that there exists a unique relation between the symmetries of the metric and the symmetries of the geodesic equations and their corresponding conservation laws [32, 33, 34]. For instance the conservation law of momentum for the free particle follows from the translation symmetry of the Euclidean spacetime. The group of translations and the group of rotations form the group of isometries or Killing vectors of the Euclidean space, while the conservation law for the angular momentum is related to the rotation symmetries of the flat space.

From Noether’s theorem, it follows that the quantity2 I=pμξμI=p_\mu\xi^\mu is a conservation law for the geodesic equations, if ξμ\xi^\mu is a KV, that is, dIdt=I,t+{I,HG}=0\frac{dI}{dt}=I_{,t}+\{I,H_G\}=0, hence {I,HG}=0\{I,H_G\}=0, where HGH_G is the Hamiltonian function for the geodesic equations HG=12gμνpμpνH_G=\frac{1}{2}g^{\mu\nu}p_\mu p_\nu.

Condition dIdt=0\frac{dI}{dt}=0, is equivalent to the Killing equation (1). Nevertheless, the geodesic Hamiltonian is not invariant under conformal transformations except if the “Energy” is zero, that is, when it describes massless (null) geodesics. In this case, under a conformal transformation, the volume continues to be zero under conformal transformations and consequently conservation laws can be constructed. Hence, if ξμ\xi^\mu is a CKV, the symmetry condition is relaxed as dIdt=2ψ(xκ)HG\frac{dI}{dt}=2\psi(x^\kappa)H_G, mod HG=0H_G=0, where ψ(xκ)\psi(x^\kappa) is the conformal factor. Therefore, we end with the condition {I,H}=2ψ(xκ)H\{I,H\}=2\psi(x^\kappa)H, which is equivalent to the conformal condition (2).

Eisenhart-Duval lift

Consider an nn-dimensional dynamical system described by the Hamiltonian function

H=12gμν(qκ)pμpν+V(qκ)=E,(5)\mathcal{H}=\frac{1}{2}g^{\mu\nu}(q^\kappa)p_\mu p_\nu+V(q^\kappa)=E, \tag*{(5)}

where qμq^\mu, pμp_\mu are the position and the momentum terms respectively, V(qμ)V(q^\mu) is the potential function that describes the dynamics and gμν(qκ)g^{\mu\nu}(q^\kappa) is a second-rank tensor that defines the kinetic term. In Newtonian physics, gμν(qκ)g^{\mu\nu}(q^\kappa) represents the nn-dimensional Euclidean flat space, and EE is the value for the energy. The equations of motion read

q˙μ=gμνpν,(6)\dot q^\mu=g^{\mu\nu}p_\nu, \tag*{(6)}
p˙μ=Fμ.(7)\dot p_\mu=F_\mu. \tag*{(7)}

in which Fμ=−V,μF_\mu=-V_{,\mu} is the force term.

We introduce the Eisenhart-Duval lift as a geometric framework for representing dynamical systems, with particular emphasis on Newtonian dynamics. Within this approach, the study of the dynamical system is reformulated in terms of the geometric properties of an extended space. The collineations of the extended metric can be used to construct local and nonlocal conservation laws for the original system [38], while properties such as the linearizability of the dynamical system can be investigated by examining the geometric invariants of the lifted space [39].

Riemannian lift

We proceed with the Riemannian lift [35] and we consider the (n+1)(n+1)-dimensional dynamical system with Hamiltonian function

HR=12gμνpμpν+12V(qκ)pz2=ER.(8)\mathcal{H}_{R}=\frac{1}{2}g^{\mu\nu}p_{\mu}p_{\nu}+\frac{1}{2}V\left(q^{\kappa}\right)p_{z}^{2}=E_{R}. \tag*{(8)}

The latter function describes the geodesic equations for the (n+1)(n+1)-dimensional Riemannian manifold with line element

dsR2=gμν(qκ)dqμdqν+1V(qκ)dz2,(9)ds_{R}^{2}=g_{\mu\nu}\left(q^{\kappa}\right)dq^{\mu}dq^{\nu}+\frac{1}{V\left(q^{\kappa}\right)}dz^{2}, \tag*{(9)}

and geodesic equations

q˙μ=gμνpν,(10)\dot{q}^{\mu}=g^{\mu\nu}p_{\nu}, \tag*{(10)}
z˙=V(qκ)pz,(11)\dot{z}=V\left(q^{\kappa}\right)p_{z}, \tag*{(11)}
p˙μ=12Fμpz2,(12)\dot{p}_{\mu}=\frac{1}{2}F_{\mu}p_{z}^{2}, \tag*{(12)}
p˙z=0.(13)\dot{p}_{z}=0. \tag*{(13)}

The momentum term pzp_{z} is a conservation law for the geodesic equations, related to the Killing symmetry ∂z\partial_{z} for the extended metric (9). Therefore, the solution trajectories of the geodesic equations (10)-(13) are solutions for the original dynamical system (6), (7), when ER=EE_{R}=E and 12pz2=1\frac{1}{2}p_{z}^{2}=1. Due to the Killing symmetry ∂z\partial_{z}, the extended geometry is foliated and the extended degrees of freedom zz, pzp_{z}, play no role in the evolution of the original system.

Lorentzian lift

On the other hand, in the Lorentzian lift approach [36], we consider the (n+2)(n+2)-dimensional dynamical system with Hamiltonian function

HL=12gμνpμpν+V(qκ)pv2+pupv=EL,(14)\mathcal{H}_{L}=\frac{1}{2}g^{\mu\nu}p_{\mu}p_{\nu}+V\left(q^{\kappa}\right)p_{v}^{2}+p_{u}p_{v}=E_{L}, \tag*{(14)}

and equations of motion

q˙μ=gμνpν,(15)\dot{q}^{\mu}=g^{\mu\nu}p_{\nu}, \tag*{(15)}
u˙=2V(qκ)pu+pv,(16)\dot{u}=2V\left(q^{\kappa}\right)p_{u}+p_{v}, \tag*{(16)}
v˙=pu,(17)\dot{v}=p_{u}, \tag*{(17)}
p˙μ=Fμpu2,(18)\dot{p}_{\mu}=F_{\mu}p_{u}^{2}, \tag*{(18)}
p˙u=0,(19)\dot{p}_{u}=0, \tag*{(19)}
p˙v=0,(20)\dot{p}_{v}=0, \tag*{(20)}

where pup_{u} and pvp_{v} are conserved quantities. We remark that the original system (6), (7) is recovered when pu2=1p_{u}^{2}=1 and pupv−EL=Ep_{u}p_{v}-E_{L}=E. Similarly to before, the extended degrees of freedom u,v,puu,v,p_{u} and pvp_{v} play no role in the dynamic evolution for the rest of the variables.

Function (14) is the geodesic Hamiltonian for the (n+2)(n+2)-dimensional Riemannian geometry of pp-wave like form

dsL2=gμν(qκ)dqμdqν+2dudv−2V(qκ)du2.(21)ds_{L}^{2}=g_{\mu\nu}\left(q^{\kappa}\right)dq^{\mu}dq^{\nu}+2dudv-2V\left(q^{\kappa}\right)du^{2}. \tag*{(21)}

Without loss of generality in the Lorentzian lift we can assume EL=0E_{L}=0, such that the Hamiltonian function (14) describes massless (null) geodesic equations. This choice is particularly useful because the null geodesic equations are invariant under conformal transformations, up to a reparametrization of the geodesics. Consequently, conformal symmetries of the lifted geometry can be used to construct conservation laws for the null geodesic dynamics. The non-uniqueness of the lifting approach is discussed in detail in [37], where different families of extended geometries are introduced, leading to new results.

Newtonian systems

We continue with the application of the Eisenhart-lift to two systems of special interest, the Ermakov-Pinney equation, and the harmonic oscillator. These two systems have the property of being linearizable within the Eisenhart-Duval framework.

Ermakov-Pinney Equation

The one-dimensional Ermakov-Pinney equation is described by the Hamiltonian function

HEP=12p2+12μq2=EEP,(22)\mathcal{H}_{EP} = \frac{1}{2}p^2 + \frac{1}{2}\frac{\mu}{q^2} = E_{EP}, \tag*{(22)}

with equations of motion

q˙=p,(23)\dot{q} = p, \tag*{(23)}
p˙=μq3.(24)\dot{p} = \frac{\mu}{q^3}. \tag*{(24)}

with solution q=q12(t)+q22(t)q = \sqrt{q_1^2(t) + q_2^2(t)}, where q1,2(t)q_{1,2}(t) are the two solutions of the free particle q¨1,2=0\ddot{q}_{1,2} = 0.

We employ the Riemannian lift, and we define the equivalent geodesic Hamiltonian

HEP,R=12p2+12(2q2μ)pz2=EEP,R,(25)\mathcal{H}_{EP,R} = \frac{1}{2}p^2 + \frac{1}{2}\left(\frac{2q^2}{\mu}\right)p_z^2 = E_{EP,R}, \tag*{(25)}

which recovers the original system when pz=1p_z = 1 and EEP=EEP,RE_{EP} = E_{EP,R}. The kinetic metric for the geodesic Lagrangian reads

ds2=dq2+(2q2μ)dz2,(26)ds^2 = dq^2 + \left(\frac{2q^2}{\mu}\right) dz^2, \tag*{(26)}

where we calculate the Ricci scalar R=0R = 0. From this we conclude that the space is the flat space and admits three KVs, two translations and one rotation. In Cartesian coordinates X=qcos⁡zX = q \cos z, Y=qsin⁡zY = q \sin z, the vector fields read {∂X,∂Y,Y∂X−X∂Y}\{\partial_X, \partial_Y, Y\partial_X - X\partial_Y\} with conservation laws3

I1=YpX−XpY,(27)I_1 = Yp_X - Xp_Y, \tag*{(27)}
I2=pX,(28)I_2 = p_X, \tag*{(28)}
I3=pY.(29)I_3 = p_Y. \tag*{(29)}

The transformation from the polar to Cartesian coordinates is a nonlocal transformation, which transforms the equations for the free particle X¨=0\ddot{X} = 0, into the Ermakov-Pinney equation. The inverse transformation q=X2+Y2q = \sqrt{X^2 + Y^2} defines the solution.

We show that the Ermakov-Pinney equation can be linearized through the extended geometry.

Harmonic Oscillator

Consider now the harmonic oscillator with Hamiltonian function

HHO=12p2+12ω2q2=EHO(30)\mathcal{H}_{HO} = \frac{1}{2}p^2 + \frac{1}{2}\omega^2q^2 = E_{HO} \tag*{(30)}

and equations of motion

q˙=p,(31)\dot{q} = p, \tag*{(31)}
p˙=−ω2q.(32)\dot{p} = -\omega^2q. \tag*{(32)}

We employ the Lorentzian lift and we introduce the (null) geodesic equivalent Hamiltonian

HHO,L≡12p2+12ω2q2pv2+pupv=0.(33)\mathcal{H}_{\mathrm{HO,L}} \equiv\frac{1}{2}p^{2}+\frac{1}{2}\omega^{2}q^{2}p_{v}^{2}+p_{u}p_{v}=0. \tag*{(33)}

For the three-dimensional extended geometry we calculate the Cotton-York tensor and we find that it has zero components, consequently, the extended space is conformally flat. The extended space possesses a ten dimensional Conformal algebra, the so(3,2)\mathrm{so}(3,2), with elements

K1=cos⁡(ωu) ∂q+qsin⁡(ωu) ∂v(34)K^{1}=\cos(\omega u)\,\partial_{q}+q\sin(\omega u)\,\partial_{v} \tag*{(34)}
K2=sin⁡(ωu) ∂q−qcos⁡(ωu) ∂v(35)K^{2}=\sin(\omega u)\,\partial_{q}-q\cos(\omega u)\,\partial_{v} \tag*{(35)}
K3=∂u(36)K^{3}=\partial_{u} \tag*{(36)}
K4=∂v(37)K^{4}=\partial_{v} \tag*{(37)}

which form the four-dimensional Killing algebra, and

CKV1=q∂q+2v∂u(38)CKV^{1}=q\partial_{q}+2v\partial_{u} \tag*{(38)}
CKV2=qv∂q−q22∂u+(v2−ω2q44)∂v(39)CKV^{2}=qv\partial_{q}-\frac{q^{2}}{2}\partial_{u}+\left(v^{2}-\omega^{2}\frac{q^{4}}{4}\right)\partial_{v} \tag*{(39)}
CKV3=(cos⁡(ωu)+ωq22sin⁡(ωu))∂q−qcos⁡(ωu)∂v+q2(2ωvsin⁡(ωu)−cos⁡(ωu)ω2q2)∂v(40)CKV^{3}=\left(\cos(\omega u)+\frac{\omega q^{2}}{2}\sin(\omega u)\right)\partial_{q}-q\cos(\omega u)\partial_{v}+\frac{q}{2}\left(2\omega v\sin(\omega u)-\cos(\omega u)\omega^{2}q^{2}\right)\partial_{v} \tag*{(40)}
CKV4=(sin⁡(ωu)−ωq22cos⁡(ωu))∂q+qsin⁡(ωu)∂v−q2(2ωvcos⁡(ωu)+sin⁡(ωu)ω2q2)∂v(41)CKV^{4}=\left(\sin(\omega u)-\frac{\omega q^{2}}{2}\cos(\omega u)\right)\partial_{q}+q\sin(\omega u)\partial_{v}-\frac{q}{2}\left(2\omega v\cos(\omega u)+\sin(\omega u)\omega^{2}q^{2}\right)\partial_{v} \tag*{(41)}
CKV5=qωsin⁡(2ωu) ∂q+cos⁡(2ωu) ∂u+ω2q2cos⁡(2ωu) ∂v(42)CKV^{5}=q\omega\sin(2\omega u)\,\partial_{q}+\cos(2\omega u)\,\partial_{u}+\omega^{2}q^{2}\cos(2\omega u)\,\partial_{v} \tag*{(42)}
CKV6=qωcos⁡(2ωu) ∂q+sin⁡(2ωu) ∂u+ω2q2sin⁡(2ωu) ∂v.(43)CKV^{6}=q\omega\cos(2\omega u)\,\partial_{q}+\sin(2\omega u)\,\partial_{u}+\omega^{2}q^{2}\sin(2\omega u)\,\partial_{v}. \tag*{(43)}

At this point it is important to mention that for the nn-dimensional isotropic harmonic oscillator the Lorentzian lift leads always to a (n+2)(n+2)-dimensional geometry which is always conformally flat.

Data-Driven Conservation Laws via Eisenhart–Duval Geometry

In this section we introduce the method that we employ for the determination of the number of symmetries and the reconstruction of conservation laws for dynamical systems directly from data. This framework consists of four steps, (i) a PINN for the reconstruction of the Hamiltonian function; (ii) the application of the Eisenhart-Duval lift and the reconstruction of the extended geometry; (iii) a random-feature neural network that provides the KVs and the CKVs for the extended metric, and as a final step (iv) the construction of the conserved quantities and the determination of the commutators. Figure 1 shows the flow diagram for the framework.

Figure 1. Flow diagram for the data-drive reconstruction of the conservation laws and symmetries for Newtonian systems via the Eisenhart-Duval geometry.

Physics-Informed Neural Network

We make the assumption that the observed data are points on the phase space zk=(q(tk),p(tk))z_k=(q(t_k),p(t_k)) for a Newtonian system with Hamiltonian function (5) and gμνg_{\mu\nu} describes the flat space in Cartesian coordinates4, that is, gμν=δμνg_{\mu\nu}=\delta_{\mu\nu}. The potential function V(q)V(q) is trainable and is learned using a Hamiltonian Neural Network approach. The equations of motion (6), (7) are integrated for a step number KK, by using the Störmer-Verlet approach [63, 64], that is, we integrate as follows

qk+1=qk+pk+12Δt,(44)q_{k+1}=q_k+p_{k+\frac{1}{2}}\Delta t, \tag*{(44)}
pk+12=pk+Δt2F(qk),(45)p_{k+\frac{1}{2}}=p_k+\frac{\Delta t}{2}F(q_k), \tag*{(45)}
pk+1=pk+12+Δt2F(qk+1),(46)p_{k+1}=p_{k+\frac{1}{2}}+\frac{\Delta t}{2}F(q_{k+1}), \tag*{(46)}

where Δt\Delta t is the step of integration.

We consider a dimensionless loss function, defined by the average difference between the trajectories predicted by the learned Hamiltonian and the data, i.e.

L=∑μ⟨(q^μ(tk+K)−qμ(tk+K))2⟩Var⁡(qμ)+∑μ⟨(p^μ(tk+K)−pμ(tk+K))2⟩Var⁡(pμ),(47)\mathcal{L} = \sum_{\mu} \frac{\left\langle\left(\hat{q}^{\mu}(t_{k+K}) - q^{\mu}(t_{k+K})\right)^{2} \right\rangle}{\operatorname{Var}(q^{\mu})} + \sum_{\mu} \frac{\left\langle\left(\hat{p}_{\mu}(t_{k+K}) - p_{\mu}(t_{k+K})\right)^{2} \right\rangle}{\operatorname{Var}(p_{\mu})}, \tag*{(47)}

where {q^μ(tk+K),p^μ(tk+K)}\{\hat{q}^{\mu}(t_{k+K}), \hat{p}_{\mu}(t_{k+K})\} are the predictions of the network starting from the data at tkt_k and Var⁡(qμ)\operatorname{Var}(q^{\mu}), Var⁡(pμ)\operatorname{Var}(p^{\mu}) are the variances of the data points.

Eisenhart-Duval lift

From the reconstructed potential we construct the extended geometries of the Riemannian and of the Lorentzian lift, and we calculate the Ricci scalar, the Cotton-York tensor and the Weyl tensor by automatic differentiation of the metric components. Automatic differentiation is exact up to machine precision. Nevertheless, the curvature tensors depend on the second and third derivatives of the learned potential V(qμ)V(q^{\mu}), which are more sensitive to the errors of the reconstruction than the potential itself.

Random-feature network

For the determination of the KVs and the CKVs of the extended metric we introduce a neural network with random, non-trainable weights. Let ξA\xi^{A} be the generic CKV for the extended geometry expressed in a tanh base [60], such as

ξA(X)=∑j=1NaAj ϕj(X),ϕj(X)=tanh⁡(wj⋅X+bj),(48)\xi^{A}(\mathbf{X}) = \sum_{j=1}^{N} a_{Aj}\,\phi_j(\mathbf{X}), \qquad\phi_j(\mathbf{X}) = \tanh(w_j \cdot\mathbf{X} + b_j), \tag*{(48)}

where wjw_j are the weights and the biases bjb_j are random numbers, and only the coefficients aAja_{Aj} are unknown coefficients. Parameter X\mathbf{X} denotes the normalized coordinates {qμ,z}\{q^\mu,z\} or {qμ,u,v}\{q^\mu,u,v\} within the range X∈[−1,1]\mathbf{X} \in[-1,1].

The symmetry condition S(aAj)S(a_{Aj}) (2) is a linear equation, which means that the derivation of the symmetry vectors is reduced to a problem of optimizing the coefficients aAja_{Aj}. We evaluate the symmetry condition at a large number of collocation points, and we define the least square quantity, for each vector field ξk\xi_k,

σk=S(aAj)2ξk2,ξk2=ξkAξk A,(49)\sigma_k = \sqrt{\frac{S(a_{Aj})^2}{\xi_k^2}}, \qquad\xi_k^2 = \xi_k^{A}\xi_{k\,A}, \tag*{(49)}

and we determine the coefficients aAja_{Aj} which minimize it. We have normalized to ξk2\xi_k^2 in order to avoid the determination of the trivial symmetry ξkA=0\xi_k^{A}=0. A positive definite metric is necessary because for the Lorentzian metric ξk2\xi_k^2 vanishes for the null vector fields, such as the KV ∂v\partial_v.

For an exact symmetry it holds that S(aAj)=0S(a_{Aj})=0, that is, σexact=0\sigma_{\mathrm{exact}}=0. Nevertheless, for a learned metric (the trained) symmetries provide small values for the parameter σ\sigma. For non-symmetry vectors the obtained values σk\sigma_k are far from zero, and when the network finds non-symmetries, the evolution of the σk\sigma_k has a gap. The vector fields with σk\sigma_k below the gap give the number of symmetries and the corresponding KVs/CKVs.

Indeed, the network does not learn exact symmetries, but approximate symmetries where σ<ϵ\sigma< \epsilon and ϵ\epsilon is a small positive number [65].

Lie algebra and conservation laws

From the learned symmetry vectors we can compute the commutators [ξA,ξB][\xi^{A},\xi^{B}] by automatic differentiation. Because the learned vector fields should form a Lie algebra, we calculate the structure constants CBCAC^{A}_{BC} by the least-square solution at the collocation points. This allows us to test if the admitted symmetry vectors form a closed algebra. Finally, we calculate the conservation laws by using Noether’s theorem, that is,

I(ξA)=ξApA.(50)I(\xi^{A}) = \xi^{A}p_A. \tag*{(50)}

Because the network has learned approximate symmetries, the corresponding conserved quantities are approximate conservation laws [65, 66], that is dIdt≃Φ(t)\frac{dI}{dt} \simeq\Phi(t), with Φ2(t)→0\Phi^2(t) \to0, a small varying function.

Experiments

We proceed with the validation of the method described before by using mock data generated from systems where the admitted symmetries and conservation laws are known. Specifically, we consider the one-dimensional Ermakov-Pinney equation described before, and the one-dimensional harmonic oscillator with or without a quartic perturbation term. The last example allows us to explore the application of the method to the construction of approximate conserved quantities. For each of these systems we solve numerically the equations of motion, and we add Gaussian noise on the position-momentum data points, with different noise levels.

For this analysis we consider five independent networks and we report the median obtained from the five different networks. For the Hamiltonian PINN we consider a three layer network of sixty-four (3×643 \times64) neurons each with a tanh function, and the network was trained using the Adam optimizer. For the random-feature network we used a sufficiently large number of collocation points which depends on the dimension of the system, in order to recover the complete number of symmetries.

Ermakov 1D

We consider the Ermakov-Pinney Hamiltonian system (22), with μ=1\mu= 1. We consider twenty trajectories with initial conditions q0∈[0.5,2]q_0 \in[0.5, 2] and p0∈[−0.8,0.2]p_0 \in[-0.8, 0.2] and we integrated so as to create 300 points per trajectory. Then, we introduced Gaussian noise to the data (q(tk),p(tk))(q(t_k), p(t_k)). Specifically, we have added three different levels of noise, of 1%, 5% and 10%, on the same baseline trajectories, as can be seen in Figure 2.

Figure 2. Ermakov 1D: Training data in the phase-space (q,p)(q,p), for the Gaussian noise of 1%, 5% and 10%.

In Figure 3 we present the reconstruction of the potential function, and of the force term, as follows from the training of the Hamiltonian PINN. We observe that over the region covered by the data, the typical error is lower than the noise level. The noise in the force term arises from the automatic differentiation used to compute it. Note that the potential is the function being trained.

Figure 3. Ermakov 1D: Reconstructed potential function and force term from the Hamiltonian PINN for noise levels 1%, 5% and 10%. The lower panels give the relative error.

We employ the Riemannian lift, and for the reconstructed two-dimensional extended metric with line element (8) we calculate the Ricci scalar R(q)R(q) as presented in Figure 4. The Ricci scalar is determined by the second-derivative of the potential V(q)V(q), where the automatic differentiation is employed. For 1% and 5% noise, the Ricci scalar stays close to the zero value with averaged values ∣R∣=0.028|R| = 0.028 and ∣R∣=0.088|R| = 0.088 for the 1% and 5% noise levels respectively. On the other hand, for the data with 10% noise, the reconstructed Ricci scalar had averaged value ∣R∣=0.167|R| = 0.167, which indicates a deviation from the zero value, as expected for the analytic form. The main deviation exists for q<1q < 1, as the trajectories reach the center. We conclude that the reconstructed extended metric is thus flat to within the reconstruction error. This should also be verified by the number of admitted KVs, which should be three and form the e(2)\mathfrak{e}(2) Lie algebra.

Figure 4. Ermakov 1D: Derived Ricci scalar R(q)R(q) for the reconstructed extended line element after the application of the Riemannian lift.

Within the reconstructed extended geometry we apply the random-feature network for the derivation of the KVs. In Figure 5 we present the lowest values of the σ\sigma, for three noise levels. The markers are the median of the twenty independent trajectories. For each noise level we observe a similar behaviour. There are three vector fields with σ<0.1\sigma< 0.1, while for k>3k > 3, the σ≃1\sigma\simeq1, from where we can conclude that the network has reconstructed the three KVs of the two-dimensional flat space. The lower-panel of 5 shows the relative non-conservation of the quantity rk=ddtln⁡(Ik(ξAk))r_k = \frac{d}{dt}\ln\left(I_k\left(\xi_A^k\right)\right), where we see that for k=1,2k = 1, 2 and 3, the values are smaller than 10−110^{-1}, while for k>3k > 3, the non-conservation value is close to 10010^0.

Figure 5. Ermakov 1D: Values σk\sigma_k of the Killing operator of the reconstructed extended metric. In the lower-pane we present the relative non-conservation rkr_k of the quantity Ik=ξkApAI_k = \xi_k^A p_A on different trajectories.

The symmetry which corresponds to k=1k = 1 is the trivial conservation law of the momentum pzp_z, which is the angular momentum and corresponds to the rotation symmetry. The other two vector fields correspond to the two-translations of the flat space. In Figure 6 we plot the interval curves for the reconstructed KVs in Cartesian coordinates, and we present the commutators of the obtained vector fields. Finally, in 7 we present the derivation of the conservation law and its evolution dIdt\frac{dI}{dt} for a specific trajectory, as well as the Casimir invariant of the admitted Lie algebra.

Figure 6. Ermakov 1D: Integral curves of the three learned KVs and commutator relations.

We conclude that the reconstructed space is (approximately) flat.

Harmonic Oscillator 1D

We proceed with the derivation of the conservation laws for the one-dimensional harmonic oscillator. We consider again data from twenty trajectories as presented in the phase-space portraits of Figure 8. The initial conditions of the trajectories are for q0∈[−2,2]q_0 \in[-2, 2], p0∈[−2,2]p_0 \in[-2, 2] and ω2=1\omega^2 = 1. The data are for trajectories of one period.

Figure 8. Harmonic Oscillator 1D: Training data in the phase-space (qq, pp), for the Gaussian noise of 1%, 5% and 10%.

Using the Hamiltonian PINN, we reconstruct the potential function and compute the Cotton–York tensor of the extended geometry (21) obtained from the Lorentzian lift. The latter are shown in Figure 9, from which we see that the network reconstructs the potential function well. Although the Cotton–York tensor is not zero, it has an average value ⟨C⟩\langle C \rangle close to zero with values −0.005-0.005, −0.020-0.020, −0.038-0.038 for the 1%, 5% and 10% noise levels respectively. Hence, the reconstructed extended geometry is asymptotically conformally flat for all the noise levels. The amplitude of the oscillations for the Cotton–York tensor depends on the noise level. This reflects the sensitivity of the Cotton–York tensor to small errors in the learned potential, since the Cotton–York tensor follows from the third derivative of the potential, and repeated differentiation amplifies small deviations of the network from the exact potential.

Figure 9. Harmonic Oscillator 1D: Reconstructed potential function from the Hamiltonian PINN for noise levels 1%, 5% and 10%. The lower panel gives the relative error. Moreover, we present the reconstructed Cotton-York tensor for the Lorentzian extended metric, which is derived with automatic differentiation of the reconstructed potential.

In Figure 10 we present the values for the σ\sigma as derived for the CKV condition, for the reconstructed three-dimensional extended metric. Furthermore, we present the non-conservation rk=(rmst⁡(Ik(ξkA)))−1ddtIk(ξkA)r_k = \left(\operatorname{rmst}\left(I_k\left(\xi_k^A\right)\right)\right)^{-1} \frac{d}{dt} I_k\left(\xi_k^A\right) of the quantities Ik=ξkApAI_k = \xi_k^A p_A. We observe that for k≤10k \le10, the σk≤10−1\sigma_k \le10^{-1}, while for k>10k > 10, σk\sigma_k are close to 10010^0, while for the rkr_k it follows that for k≤10k \le10, rk<10−2r_k < 10^{-2}, while for k>10k > 10 the quantities IkI_k are not conservation laws. The time evolution for the ten admitted conserved quantities is plotted for trained trajectories in Figure 11. Moreover, Figure 12 shows the commutators for the reconstructed CKVs where we see that the commutator relations are independent of the noise level, and they form the so(3,2)\mathfrak{so}(3,2) Lie algebra, as expected for the three-dimensional conformally flat geometry.

Figure 10. Harmonic Oscillator 1D: Values σk\sigma_k of the Conformal Killing operator of the reconstructed extended metric. In the lower-pane we present the relative non-conservation rkr_k of the quantity Ik=ξAkpAI_k = \xi^k_A p_A on different trajectories.

Figure 11. Harmonic Oscillator 1D: Conservation laws for the geodesic equations.

Figure 12. Harmonic Oscillator 1D: Commutators for the reconstructed ten CKVs. The vector fields form the so(3,2)so(3, 2) Lie algebra, independent of the noise level.

Quartic correction

We introduce a quartic term Vcorr(q)=ϵq4V_{\mathrm{corr}}(q)=\epsilon q^4 to the one-dimensional harmonic oscillator and repeat the analysis. We assume two cases, ϵ=10−3\epsilon=10^{-3} and ϵ=10−1\epsilon=10^{-1}. For ϵ≠0\epsilon\ne0, the dynamical system does not admit any nontrivial symmetries, except for the time-translation symmetry with conservation law the Hamiltonian function. Consequently, the extended three-dimensional metric of the Lorentzian lift has only two CKVs, the vector fields ∂u\partial_u and ∂v\partial_v. The other eight symmetries of the oscillator are broken by the correction term and survive only as approximate symmetries, to a degree controlled by ϵ\epsilon.

In Figure 13 we present the values for the σ\sigma as derived for the CKV condition, for the reconstructed three-dimensional extended metric, and the relative non-conservation rkr_k for the corresponding quantities along the (null) geodesic equations. For the small correction term ϵ=10−3\epsilon=10^{-3} and for the time evolution of one period, we see that there are ten vector fields where σ<10−1\sigma<10^{-1}, and the relative non-conservation rkr_k are lower than 10−210^{-2}. On the other hand, for ϵ=10−1\epsilon=10^{-1}, there are only six vector fields with σ<10−1\sigma<10^{-1}, where the corresponding relative non-conservation rkr_k are larger than 10−210^{-2}. For the ϵ=10−3\epsilon= 10^{-3} we observe the gap in the σ\sigma, for k>10k > 10; such a gap does not exist for the ϵ=10−1\epsilon= 10^{-1} case.

Figure 13. Harmonic Oscillator 1D with Quartic Correction: Values σk\sigma_k of the Conformal Killing operator of the reconstructed extended metric. In the lower-pane we present the relative non-conservation rkr_k of the quantity Ik=ξkApAI_k = \xi_k^A p_A on different trajectories.

Figure 7. Ermakov 1D: Conservation laws for the geodesic equations and Casimir invariant.

Conclusions

In this work we introduced a data-driven framework for the derivation of symmetries and of conservation laws that are linear in the momenta, for Newtonian dynamical systems described by Hamiltonian functions with kinetic and potential terms. We employed the Eisenhart-Duval lift to geometrize the dynamical system, which reduces the derivation of the symmetries to the determination of the collineations of an extended geometry, and specifically to the determination of its Killing and conformal Killing algebras.

Our approach consists of four steps. First, a Hamiltonian PINN learns the potential function from phase-space data. Second, the learned potential defines the extended metric of the Riemannian or of the Lorentzian lift, and we calculate the curvature invariants of these geometries, which provide important information about the structure of the geometry and the nature of the geodesic equations. Third, a random-feature network solves the Killing and the conformal Killing equations. Finally, the Lie algebra and the Noetherian conservation laws are determined from the learned vector fields.

The main advantage of the geometric formulation is that the symmetry conditions for the original equations of motion, in which the potential enters through additional constraints, are replaced by the Killing and conformal Killing equations of the extended metric. These equations are linear in the unknown vector fields, so that the symmetry problem becomes a linear algebra problem and the number of symmetries can be read from the spectrum of the operator. Moreover, the conservation laws which follow from the extended geometry are not necessarily local functions of the original variables, which means that the approach also reveals nonlocal conservation laws.

In order to demonstrate the application of the framework, we considered mock data for the one-dimensional Ermakov-Pinney equation and the one-dimensional harmonic oscillator. The two examples were not selected randomly. Through the Eisenhart-Duval lift, both lead to geodesic equations in geometries with the maximal number of symmetries. This allows us to test our framework in the reconstruction of the symmetries of geometries with a maximal symmetry algebra. Specifically, the Ermakov-Pinney equation is equivalent to the geodesic equations of the two-dimensional flat space, which admits the maximal Killing algebra, while the harmonic oscillator is equivalent to the null geodesic equations of a three-dimensional conformally flat Lorentzian geometry, which admits the maximal conformal algebra.

For different levels of Gaussian noise on the solution trajectories of the aforementioned systems, we reconstructed the potential function and the geometric features of the extended geometries up to a certain noise level. Furthermore, we showed that the random-feature neural network reconstructs the full symmetry algebra, of maximal dimension, admitted by the equivalent geodesic equations.

The present analysis has several limitations. First, we assumed a specific form of the Hamiltonian function, in which the kinetic metric is flat and expressed in Cartesian coordinates. Second, we focused only on conservation laws which are linear in the momenta and are constructed from the KVs and the CKVs, and we considered only the Riemannian and the Lorentzian lifts, although other lifts exist which can provide new results [37]. Moreover, the random-feature neural network learns the symmetries locally, so the reconstructed conservation laws are valid only within the region covered by the data.

In order to demonstrate the application of the framework to higher-dimensional systems, we considered the two-dimensional Hénon-Heiles system, as a perturbation of the two-dimensional isotropic harmonic oscillator by the correction term VcorrHH=ϵ(q12q2−q233)V^{HH}_{corr} = \epsilon\left(q_1^2q_2 - \frac{q_2^3}{3}\right). We applied the framework for ϵ=0\epsilon= 0, 10−210^{-2}, 10−110^{-1} and 4×10−14 \times10^{-1}, with 5% noise, for trajectories for half a period of the oscillator. For ϵ=0\epsilon= 0 the extended Lorentzian metric is four-dimensional and conformally flat, and it admits the maximal, fifteen-dimensional, conformal algebra so(4,2)so(4, 2). Figure 14 shows the spectrum σk\sigma_k of the conformal Killing operator and the relative non-conservation rkr_k. For ϵ=0\epsilon= 0 we recover the maximal number of symmetries, fifteen vector fields have σk<10−1\sigma_k < 10^{-1}, and the spectrum has a clear gap at k=16k = 16. For ϵ=10−2\epsilon= 10^{-2} we recover the same number of symmetries, with values of rkr_k comparable to those of the ϵ=0\epsilon= 0 case. For ϵ=10−1\epsilon= 10^{-1} a weaker gap remains at k=16k = 16, while the values of rkr_k increase to approximately 10−210^{-2}. Finally, for ϵ=4×10−1\epsilon= 4 \times10^{-1} the gap disappears, and only the two exact symmetries remain, which correspond to the trivial vector fields ∂u\partial_u and ∂v\partial_v.

Figure 14. Hénon-Heiles: Values σk\sigma_k of the Conformal Killing operator of the reconstructed extended metric. In the lower-pane we present the relative non-conservation rkr_k of the quantity Ik=ξkApAI_k = \xi^A_k p_A on different trajectories.

In a future work we plan to extend this framework to the reconstruction of higher-order conservation laws and to more general forms of the Hamiltonian function. Conservation laws which are quadratic or higher-polynomial functions in the momenta, such as in the the integrable cases of the Hénon-Heiles system, the Lewis-invariant for the higher-dimensional Ermakov-Pinney system, or the Runge-Lenz vector of the Kepler problem, are a result of Killing tensors of the extended metric or conformal Killing tensors for the case of null geodesics.

Finally, the geometric formulation offers a natural route to quantization. The symmetries learned from the classical trajectories can be used for the construction of the symmetry operators and of the conserved quantities of the quantum system, and for the study of the relation between classical and quantum integrability directly from data.

Code Availability Statement: The code is available upon reasonable request.

**Declaration of competing interests:**The author declare no conflict of interests.

Acknowledgments

AP was partially supported by FONDECYT Grant No. 1240514.

References

  1. [14]S.V. Meleshko and V.P. Shapeev, J. Nonl. Math. Phys. 18, 195 (2011)
  2. [15]G.M. Webb and G.P. Zank, J. Math. Phys. A: Math. Theor. 40, 545 (2007)
  3. [16]M.C. Nucci and G. Sanchini, Symmetry 7, 1613 (2015)
  4. [17]A. Paliathanasis, K. Krishnakumar, K.M. Tamizhmani and P.G.L. Leach, Mathematics 4, 28 (2016)
  5. [18]P.G. Estevez, F.J. Herranz, J. de Lucas and C. Sardon, Appl. Math. Comput. 273, 435 (2016)
  6. [19]X. Xin, Acta Phys. Sin. 65, 240202 (2016)
  7. [20]N. Kallinikos and E. Melletidou, J. Phys. A: Math. Theor. 46, 305202 (2013)
  8. [21]S. Jamal and A. Paliathanasis, J. Geom. Phys. 117, 50 (2017)
  9. [22]G.M. Webb, J. Phys A: Math. Gen. 23, 3885 (1990)
  10. [23]P.G.L. Leach, J. Math. Anal. Appl. 348, 487 (2008)
  11. [24]M. Tsamparlis, J. Phys. Conf. 453, 012020 (2013)
  12. [25]S. Moyo and P.G.L. Leach, Proceedings of Institute of NAS of Ukraine 50, 204 (2004)
  13. [26]S. Moyo and P.G.L. Leach, J. Math. Anal. Appl. 306, 35 (2005)
  14. [27]X. Xin, H. Liu, L. Zhang and Z. Wang, Appl. Math. Lett. 88, 132 (2019)
  15. [28]G.E. Prince and C.J. Eliezer, J. Math. Phys.: Math. Gen. 14, 587 (1981)
  16. [29]M. Tsamparlis and A. Paliathanasis, Nonlinear Dynamics 62, 203 (2010)
  17. [30]M. Tsamparlis and A. Paliathanasis, Gen. Rel. Gravit. 43, 1861 (2011)
  18. [31]T. Christodoulakis, N. Dimakis and P.A. Terzis, J. Phys. A. 47, 095202 (2014)
  19. [32]G.H. Katzin, and J. Levine, J. Math. Phys. 15, 1460 (1974)
  20. [33]A.V. Aminova, Sbornik Math. 186, 1711 (1995)
  21. [34]A.V. Aminova and N.A Aminov, Sbornik Math. 201, 631 (2010)
  22. [35]L.P. Eisenhart, Dynamical Trajectories and Geodesics, Annals. Math. 30, 591 (1928)DOI
  23. [36]C. Duval, G. Burdet, H.P. Kunzle and M. Perrin, Bargmann structures and Newton-Cartan theory, Phys. Rev. D 31, 1841 (1985)
  24. [37]A. Paliathanasis, Linearization of Newton’s second law, Int. J. Theor. Phys. 63, 303 (2024)DOI
  25. [38]A. Paliathanasis, Symmetry, 17, 340 (2025)
  26. [39]A. Paliathanasis, J. Geom. Phys. 206, 105338 (2025)
  27. [40]M. Cariglia and F.K. Alves, Eur. Phys. J. Plus 36, 025018 (2015)
  28. [41]B.K. Parida, A. Sen, S. Dhasmana and Z.K. Silagadze, Mod. Phys. Lett. A 38, 2350149 (2023)
  29. [42]N. Kan, T. Aoyama, T. Hasegawa and K. Shiraishi, Phys. Rev. D 104, 086001 (2021)
  30. [43]M. Cariglia, A. Galajinsky, G.W. Gibbons and P.A. Horvathy, Eur. Phys. J. C 78, 314 (2018)
  31. [44]A. Galajinsky and I. Masterov, Phys. Lett. B 765, 86 (2017)
  32. [45]S. Chanda, G.W. Gibbons and P. Guha, J. Math. Phys. 58, 032503 (2017)
  33. [46]M. Raisi, P. Perdikaris and G.E. Karniadakis, J. Compt. Phys. 378, 686 (2019)
  34. [47]S. Brunton, J.L. Proctor and J. Nathan Kutz, PNAS 113, 3932 (2016)
  35. [48]G.E. Karniadakis, I.G. Kevrekidis, L. Lu, P. Perdikaris, S. Wang and L. Yang, Nature Reviews Physics 3, 442 (2021)
  36. [49]J. A. Carretero, J.F. Urbán, F. Abalos, J. Massó, O.A. Reula, J.A. Pons and C. Palenzuela, (2026) [arXiv:2609.38331]arxiv.org/abs/2609.38331
  37. [50]A. Xiang, Z. Shumaylov, P. Zaika, F. Sherry and C.-B. Schonlieb, (2025) [arXiv:2502.00373]
  38. [51]N.S. Akintsov, A.P. Nevecheria, G. Yuan, V.S. Igumnov, S.N. Andreev and Q.-H. Qin, Symmetry 18, 1303 (2026)
  39. [52]A.H. Shah, N.R. Butt, A. Ahmad, M.O. Bin Saeed, (2025) [arXiv:2509.26113]arxiv.org/abs/2509.26113
  40. [53]T. Akhound-Sadegh, L. Perreult-Levasseur, J. Brandsteler, M. Welling and S. Ravanbakhsh, NIPS ’23: Proceedings of the 37th International Conference on Neural Information Processing Systems, 1842, 42468 (2023)
  41. [54]A. Paliathanasis, (2026) [arXiv:2605.30139]arxiv.org/abs/2605.30139
  42. [55]S. Greydanus, M. Dzamba and J. Yosinki, Proceedings of the 33rd International Conference on Neural Information Processing Systems, 1378, 15379 (2019)
  43. [56]M. Crammer, S. Greydanus, S. Hoyer, P. Battaglia, D. Spergel and S. Ho, (2020) [arXiv:2003.04630]
  44. [57]Z. Liu and M. Tegmark, Phys. Rev. Lett. 126, 180604 (2021)
  45. [58]Y.-i. Mototake, Phys. Rev. E 103, 033303 (2021)
  46. [59]K. Desai, B. Nachman and J. Thaler, Phys. Rev. D 105, 096031 (2022)
  47. [60]J. Chen, X. Chi, W. E. and Z. Yang, J. Mach. Learn. 1, 268 (2022)
  48. [61]G.H Katzin, J. Levine and W.R. Davis, J. Math. Phys. 10, 617 (1969)
  49. [62]K. Yano, Lie derivatives and its applications, North-Holland Publishing Co., Amsterdam (1957)
  50. [63]L. Verlet, Phys. Rev. 159, 98 (1967)
  51. [64]E. Hairer, C. Lubich and G. Wanner, Acta Numerica 12, 399 (2003)
  52. [65]A. Paliathanasis and S. Jamal, J. Geom. Phys. 124, 300 (2018)
  53. [66]K.S. Govinder, T.G. Heil and T. Uzer, Phys. Lett. A 240, 127 (1998)

Paper details

vlnc:W_xvmp5cen
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 10

downloads · 0

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