Variational principles for conformal geodesics

Conformal geodesics are solutions to a system of third order of equations, which makes a Lagrangian formulation problematic. We show how enlarging the class of allowed variations leads to a variational formulation for this system with a third--order conformally invariant Lagrangian. We also discuss the conformally invariant system of fourth order ODEs arising from this Lagrangian, and show that some of its integral curves are spirals.


Introduction
A geodesic on a (pseudo) Riemannian manifold is uniquely specified by a point and a tangent direction. Moreover, if two points are sufficiently close to each other, then there exists exactly one length minimising (or, in Lorentzian geometry, maximising) curve between these two points, which is geodesic. This formulation relies on the methods of the calculus of variations, and enables a construction of normal neighbourhoods, as well as the analysis of the Jacobi fields determining the existence of conjugate points.
The variational formulation has been missing in conformal geometry, where conformal geodesics arise as solutions to a system of third order ODEs: A conformal geodesic is uniquely specified by a point, a tangent direction, and a perpendicular acceleration [19,3,18]. The odd order of the underlying system of equations is not amenable to the usual methods of calculus of variation, where the resulting Euler-Lagrange equations for nondegenerate Lagrangians are of even order.
One way around this difficulty, which we explore in this paper, is to consider a more general class of variations. As we shall see, this allows to terminate the procedure of integration by parts when the integrand depends on a derivative of a variation. This argument relies on a number of technical steps: carefully controlling boundary terms, respecting conformal invariance, and making sure that the fundamental lemma of the calculus of variations can be applied to the extended class of variations.
In the next section we shall formulate the conformal geodesic equations, and summarise the notation. In §3 we shall propose two ways to deal with third order equations from the variational perspective. In §4 we formulate the main result of our paper (Theorem 4.1) and compute the variation of the conformally invariant functional associated to a third order Lagrangian. While the standard variational procedure leads to a 4th order conformally invariant equation (4.14), looking at the extended class of variations reduces the order of the Euler-Lagrange equations to 3, and gives conformal geodesics as extremal curves. In §5 we focus on the conformally flat case, where the link between the 3rd and 4th order systems is particularly clear, and the Hamiltonian formalism can be constructed. In particular we show that logarithmic spirals arise as solutions to the 4th order system for a particular class of initial conditions. In §6 we show how the Lagrangian of Theorem 4.1 can be interpreted as the 'free particle' Lagrangian on the total space of the tractor bundle. Finally in §7 we construct a degenerate Lagrangian which uses a skew-symmetric tensor. This gives rise, via a Legendre transform, to a Poisson structure on the second-order tangent bundle.
Peter Olver, David Robinson, Paul Townsend and Josef Silhan for correspondence, and the anonymous referee for pointing out a gap in the original proof of Corollary 4.2.

Conformal geodesic equations
A conformal class on an n-dimensional smooth manifold M is an equivalence class of (pseudo) Riemannian metrics, where two metricsĝ and g are equivalent if there exists a nowhere zero function Ω on M such thatĝ = Ω 2 g. (2.1) If a metric g ∈ [g] has been chosen, then X, Y denotes the inner product of two vector fields with respect to this metric. We also set |X| 2 ≡ X, X , and use the notation ψ(X) for the (k − 1)-form arising as a contraction of the k-form ψ with the vector field X. Let γ be a curve of class at least C 3 in M , parametrised by t, and let U be a nowhere vanishing tangent vector to γ such that U (t) = 1. The acceleration vector of γ is A = ∇ U U , where ∇ U ≡ U a ∇ a is the directional derivative along γ, and ∇ is the Levi-Civita connection of g. The conformal geodesic equations in their conformally invariant form given by Bailey and Eastwood [3] are The Schouten tensor P ∈ Γ(T M ⊗ T M ) of g is given by in terms of the Ricci tensor r and the Ricci scalar S of g. The symbol P is an endomorphism of T M defined by P (X), Y = P (X, Y ) for all vector fields X, Y . Changing the metric toĝ as in (2.1) results in changes to the Schouten tensor, the Levi-Civita connection and the acceleration where Υ ≡ Ω −1 dΩ, and Υ is a vector field defined by Υ(X) = Υ , X . It is now a matter of explicit calculation to verify that the conformal geodesic equations (2.2) are conformally invariant.

Lagrangians for third order equations
For a non-degenerate Lagrangian, the order of the resulting Euler-Lagrange equations is equal to twice the order of the highest derivative appearing in the Lagrangian, so that in particular the Euler-Lagrange equations are of even order. Two approaches can be taken to deal with third order systems (while they will also be applicable to systems of other odd orders, for clarity we focus on order 3) (1) To allow Lagrangians which are quadratic in the acceleration, and terminate the procedure of integration by parts at the level of 3rd order derivatives consider variations V which do not keep end points fixed, but only satisfyV (t 0 ) =V (t 1 ) = 0, where t 0 , t 1 are values of the parameter at end points. This enlarges the class of variations of extremal curves, and so reduces the number of these curves as well as the order of the resulting Euler-Lagrange equations from 4 to 3. (2) Consider degenerate Lagrangians which are linear in acceleration, and necessarily involve an anti-symmetric tensor.
To illustrate both approaches with an elementary example, consider a curve γ in R n parametrised by t → X(t), and aim to obtain the third order system ...
from a variational principle. Let Γ : [−1, 1] × [t 0 , t 1 ] → M = R n be a one-parameter family of curves parametrised by s ∈ [−1, 1], such that so that the variation V is also a vector field along γ.
In the first approach we take so that one integration by parts gives If the variationV vanishes at the end points, and is otherwise arbitrary then δI = 0 iff (3.3) holds. There appear to be two immediate obstructions to generalising this approach to the conformal geodesic system (2.2), whereẊ = U,Ẍ = A + . . . , and ...
where . . . denote the lower order terms involving the Christoffel symbols and curvature. Firstly, if there is an explicit X-dependence in the Lagrangian, then the undifferentiated variation V appears in the integrand. SecondlyV = ∇ U V + . . . is not conformally invariant. We shall get around both obstructions by modifying ∇ U V to a first order conformally invariant linear operator along γ This operator differs from the derivative ∇ U along γ by a linear operator which depends on the second jet of γ. It is the unique, up to the reparametrisation of γ, conformally invariant adjustment of ∇ U . In the Proof of Theorem 4.1 we shall see that the linear operator of order zero, D(V ) − ∇ U V , has the effect of canceling some of the V contributions in the variation of the functional, and that all these contributions can be canceled in the conformally flat case.
For the second approach, let Ω ∈ Λ 2 (R n ) be a constant two-form, and set If Γ(s, t) = X(t) + sV (t) + o(s), and the variation V and its derivative vanish at the end-points, then integrating by parts twice give If the dimension n is even, and Ω is non-degenerate (so that Ω is a symplectic form), then this implies (3.3).

Main theorem
Let E be the vector field along a smooth curve γ defined by the equation (2.2). Define a third-order Lagrangian L, and the corresponding functional I[γ] by and The Lagrangian L is conformally invariant, because the expression E is.
Theorem 4.1. The first variation of the functional (4.6) is given by where K is a vector field along γ given, in terms of the Weyl tensor W , by and where D is the operator (3.4).
Proof. The proof is by a cumbersome calculation. We shall list the main steps, and give enough details so that the reader can verify our computations. The third-order term |U | −2 U, ∇ U A in (4.5) can be integrated by parts to give which results in the alternative form The term L 1 coincides, up to a constant multiple, with the Lagrangian considered in [3]. Neither the resulting boundary term, nor L 1 are conformally invariant. We shall therefore focus on L, but use the variation of L 1 as an intermediate step.
First disregard the boundary term in (4.10), and consider variations of the functional The appearance of the Riemann tensor R arises from varrying the metric g in the inner products. We eliminate R in favour of the Weyl tensor W and the Schouten tensor P using the formula We substitute and integrate the following terms, with the given coefficients, by parts: These terms were selected by a systematic, but somewhat tedious procedure, which starts from the highest order term and ensures that, apart from the inner product K, V , only terms involving D(V ) appear in the integrand. The boundary terms arising from these integrations are combined with the boundary term (4.10), which gives an expression for the variation δI. To arrive at the statement (4.7) in the Theorem we use (3.4) to eliminate ∇ U V in favour of D(V ).

4.1.
Conformal geodesic equations. The boundary term B given by (4.9) is conformally invariant, as both D and E are. Furthermore, the vector field K along γ given by (4.8) is conformally invariant, and consequently the integral defines a conformally invariant linear operator acting on variational vector fields along a given curve γ. We shall exploit both B and K to define a class of variations needed in the following corollary if and only if the conformal geodesic equations (2.2) hold.
Proof. The proof relies on a modification of the fundamental lemma of calculus of variations, which we recall following [13] Lemma 4.3 (Fundamental Lemma of Calculus of Variations). If Y : [t 0 , t 1 ] → R n is continuous, and such that In the proof one assumes that W is compactly supported with, and has continuous derivatives up to some specified order. For example all components of W may vanish, except the kth component which is given by a bump function ρ(t), where We now proceed to proving Corollary 4.2. First observe that formula (4.7) immediately proves that I[γ] is stationary under the class of variations such that (4.12) holds, provided that γ is a solution to (2.2). In order to establish the converse we shall proceed by contradiction. For this assume that δI[γ] = 0 but E(t * ) = 0 at some point t * ∈ [t 0 , t 1 ] (by continuity we may assume that t * is in the interior of the interval). Our goal is to construct a variation V satisfying (4.12), and such that which, together with (4.7), and the Fundamental Lemma applied to W = D(V ) will give a contradiction with δI[γ] = 0. In order to complete this task let us notice that there is a vector field V 0 in the kernel of the differential operator D such that V 0 (t * ) = E(t * ) − 2LU (t * ). Indeed, D is a first order differential operator and one can impose arbitrary condition on the value of V 0 at a given point t * when solving the ODE D(V 0 ) = 0. Now, let ρ be a bump function concentrated in a sufficiently small neighbourhood of t * , and letφ = ρ. Set V = φV 0 . Then D( V ) = ρV 0 and we find for some real number ω > 0. If V satisfies (4.12), then the proof is complete. However, this condition is not satisfied in general. Therefore we shall find a correctionV such that V = V +V satisfies (4.12) and (4.13) as well. In order to findV explicitly we exploit the fact that the velocity U satisfies D(U ) = 0 and K, U = 0. Thus, for any function f and can be taken such that it is zero everywhere outside a small neighbourhood of t 0 , the value off (t 0 ) is arbitrary large, and the values of f andḟ are arbitrary small. For instance , c] and f (t) = 0 otherwise satisfies these conditions for appropriately chosen constants κ ∈ R and c > t 0 (function f can be also regularized to be of class C ∞ ). We get thatV can be picked such that The fundamental lemma applied to the arbitrary variation V now gives the fourth order system which we shall call the conformal Mercator equation (the terminology will be justified in the next section) where is the adjoint of D with respect to the L 2 inner product, and K, E, L are given by (4.8), (2.2) and (4.5) respectively. Unlike D, the operator D * is not conformally invariant. The fourth order system (4.14) is nevertheless conformally invariant as the conformal weight −2 of the term |U | −2 balances the contributions from D * if g changes according to (2.1). This equation also arises from the second order Lagrangian L 1 in (4.10) under the assumption that the variation V and its derivative ∇ U V vanish at the end points t 0 and t 1 . Therefore L 1 leads to a boundary value problem for (4.14), where two points on the curve γ and two tangent vectors at these points are specified. This, in n dimensions, gives 4n conditions which is what one would expect for a fourth order system.
If the metric g is conformally flat, so that K = 0, then any solution to the conformal geodesic equation (2.2) is also a solution to (4.14). In the next section we shall explore this conformally flat case in greater detail.

Conformally flat case and spirals
Assume that M = R n , the conformal class is flat, and choose a flat metric representative g with vanishing Christoffel symbols. Then the Schouten tensor P vanishes, and the 4th order system (4.14) arising, for arbitrary variations V such that V and its first derivatives vanish at the end points, is or, eliminating the Ȧ , U term, Picard's existence and uniqueness theorem implies that a solution curve to (5.15) is determined by specifying initial conditions X(0), U (0), A(0),Ȧ(0). Specifying these values also determines C in (5.15), and conversely specifying X(0), U (0), A(0), C determinesȦ(0) by (5.16). There is an advantage in using C instead ofȦ(0) in the initial conditions, as C stays constant along the integral curves of (5.15). This can be used to verify directly that all conformal geodesics are integral curves of (5.15) for special initial conditions. Indeed, setting and substituting this into (5.16) gives the conformal geodesic equations for the flat metriċ We also verify thatĊ = 0, as a consequence of (5.18). Thus (5.17) gives a first integral of the conformal geodesic equations. The general solutions to these equations are projectively parametrised circles where U 0 is a constant unit vector, and U 0 , A 0 = 0. Therefore (5.17) evaluated at t = 0 defines a submanifold in the space of initial conditions which singles out conformal geodesics as integral curves. For generic initial conditions the integral curves of (5.15) are not conformal geodesics. For example, choosing arbitrary values of X(0), U (0), A(0), and setting C = 0 reduces (5.15) toȦ The general solution of this system is 1 t → X(t) = e t cos (ct) P 0 + e t sin (ct) Q 0 + R 0 , (5.22) where P 0 , Q 0 , R 0 are constant vectors such that P 0 , Q 0 = 0 and |P 0 | = |Q 0 |. The curves (5.22) are logarithmic spirals in the plane spanned by (P 0 , Q 0 ) which spiral towards R 0 as t → −∞. If (r, θ) are plane polar coordinates in the plane spanned by P 0 , Q 0 and centered at R 0 , then the unparametrised form of the spirals are r = |P 0 |e θ/c . This contrasts with the behaviour of conformal geodesics, where the spirals are conjectured not to arise even in curved cases [18,5]. The conformal invariance of the fourth order system (4.14) ensures that the inverse images of the logarithmic spirals under the stereographic projection from S n to R n are solutions to (4.14) on the round sphere. These curves are the loxodromes. They cut all meridians at a fixed angle, and correspond to straight lines on the Mercator map -this justifies our terminology. For general initial conditions the integral curves (5.15) are, unlike (5.19) and (5.22), nolonger planar, and their Serret-Frenet torsion is given in terms of different initial jerkȦ(0) (see [9] for other occurrences of equations involving a change of acceleration in physics). ... X: Conformal geodesic (red), logarithmic spiral (blue), and a generic integral curve with non-zero torsion (green). In Figure 1 we plot integral curves of (5.15) in R 3 corresponding to 3 sets of initial conditions. Each set shares the same values of X(0), U (0), A(0) (so the three integral curves are tangent to the second order at X(0)) but has differentȦ(0). IfȦ(0) is determined by (5.18) in terms of the remaining initial conditions, then the (red) integral curves are circles. IfȦ(0) is determined by C = 0, or equivalently by (5.20) then the (blue) curves are spirals. Finally if C = (0, 0, 1) then the numerical solution of (5.15) yields the non planar (green) 1 To find this solution set u = U/|U | so that U =ḣ(t)u, whereḣ(t) = |U |, and s = h(t) is the arc-length.
where X(t) is given by (5.22), and B is a constant vector 2 .
5.1. Hamiltonian formalism. The fourth order system (5.15) arises from a Hamiltonian, and we shall construct the Hamiltonian formulation using the Ostrogradsky approach to higher derivative Lagrangians. Neglecting the boundary term in (4.5) gives the Lagrangian where the curve γ is parametrised by t → X(t) with X ∈ M = R n , and λ ∈ R n is a Lagrange multiplier. Define the momenta P and R conjugate to X and U by so that P − λ = 0 is the set of constraints, and we can eliminateU bẏ In this formula, and below, we abuse notation and use the flat metric to identify the tangent space to R n with its dual, and so regard P and R as both vectors and co-vectors depending on the context. The Hamiltonian is given by the Legendre transform Using the canonical commutation relations R a = {R a , H} = −|R| 2 U a + 2 U, R R a − P a ,Ṗ a = 0 Turning this system of 1st order ODEs to a 4th order system yields (5.16).

Free particle on the tractor bundle
In the tractor approach of [4], the condition for a curve γ to be a conformal geodesic is shown to be equivalent to the condition that the acceleration tractor is constant along this curve, and that its tractor norm is zero. This, at least formally, has a simple mechanical interpretation of a free particle on the total space of the tractor bundle, whose position is given by the velocity tractor U. The natural kinetic Lagrangian given by the squared tractor norm of the acceleration tractor is equal to (4.5).
The details are as follows. The tractor bundle is a rank (n + 2) vector bundle T = denotes a line bundle over M of conformal densities 2 After this paper appeared on the arXiv, Josef Silhan has pointed out to us that the forthcoming work of his and Vojtech Zadnik characterise these curves by their constant conformal curvature, and vanishing conformal torsion. of weight k. Under the conformal rescallings (2.1) a section X = (σ, µ a , ρ) of T transforms according to  The tractor bundle comes equipped with a conformally invariant connection and a metric on the fibres of T defined by the norm X, X T ≡ |µ| 2 + 2σρ, and preserved by the tractor connection.
With the velocity tractor U, and the acceleration tractor A given by

curve is a projectively parametrised conformal geodesic if
where now d/dt is the directional tractor derivative. Given that A =U, the second equation in (6.26) is the Euler-Lagrange equationÜ = 0 for the 'free particle', with position U. This equation arises from a Lagrangian 1/2 U ,U T , and the first equation in (6.26) states that this Lagrangian, or equivalently the kinetic energy taken with respect to the tractor metric, is zero. It can be verified by explicit calculation that this kinetic energy is equal to the 3rd order Lagrangian (4.5), which, at least at this formal level, therefore appears to be natural.

Degenerate Lagrangian
The second approach to the Lagrangian formulation alluded to in §3 is to allow general variations, but consider a degenerate Lagrangian linear in the acceleration. As explained in §3 such Lagrangian must involve a preferred anti-symmetric tensor, which in general is absent in conformal geometry. We shall therefore restrict to structures which admit a Kähler metric in the conformal class. In this case it is advantageous to use a formulation of the conformal geodesic equations which is due to Yano [19], and Tod [18].
This formulation is equivalent to (2.2) after a change of parametrisation. Decomposing the LHS E (2.2) into parts orthogonal and parallel to U gives E ∧ U , and E, U . The second term can be made zero by reparametrizing, and using s = s(t) as a parameter along γ. An explicit calculation then verifies that vanishing of the first term E ∧ U is invariant. If one takes s to be the arc-length so that |dγ/ds| 2 = |U | 2 = 1, and A = ∇ U U is orthogonal to U , then the conformal geodesic equations become ∇ U A = −(|A| 2 + P (U, U ))U + P (U ). (7.27) See [3,18,12] for details.
Proof. Assuming that the metric g is Einstein reduces (7.27) to and the norm |A| 2 of the acceleration is a first integral. If g is in addition Kähler with the complex structure J, then the Kähler magnetic geodesics (i. e. the geodesics of charged particles moving in a magnetic field given by the Kähler form) are also conformal geodesics 3 . Indeed, the equation (7.28) where the constant e is the electric charge, implies where we used ∇J = 0. This is the conformal geodesic equation with |A| = e.
The Kähler magnetic geodesics arise from a Lagrangian where the one-form φ is the magnetic potential, i. e. dφ = Ω, and φ(U ) is the contraction of the vector field U with φ. In the corresponding Hamiltonian formalism the Kähler-magnetic geodesics are integral curves of a Hamiltonian vector field on the phase-space. A choice for the phase space in the conformal geodesic context is the second order tangent bundle T 2 M . It is the union of all second order tangent spaces T 2 x M -the space of equivalence classes (also called 2-jets j 2 (γ)) of curves γ : R → M which agree at x ∈ M up to and including the second derivatives. The space T 2 M = ∪ x∈M T 2 x M is a bundle, but not, in general, a vector bundle over M . In the presence of a linear connection there exists a canonical splitting [20] S : T 2 M → T M ⊕ T M, S(j 2 (γ)) → (γ(0), (∇γ (0)γ )(0)). This splitting equips T 2 M with the structure of rank-2n vector bundle over M , and allows a definition of the acceleration of the curve at γ(0) as A = ∇γ (0)γ (0).
We would like to construct a Lagrangian on T 2 M involving φ, as well as Ω which gives rise to all conformal geodesics on a Kähler-Einstein manifold (M, g, Ω), but even this appears to be out of reach. The following construction works in the flat case with M = R n , where n is even, and Ω is a (chosen) Kähler form.
Consider a second-order Lagrangian The converse statement is not true: there are more unparametrised conformal geodesics (a (3n − 3)dimensional family) than unparametrised Kähler-magnetic geodesics (a (2n − 1)-dimensional family, if one regards the charge as one of the parameters) through each point. For example [1] on CP 2 magnetic-Kähler geodesics lift to circles on S 5 , and all conformal geodesics lift to helices of order 2, 3 and 5. It is however the case that on a hyper-Kähler four-manifold every conformal geodesic is a Kähler-magnetic geodesic for some choice of Kähler structure [10].
where dφ = Ω, or ∂ a φ b − ∂ b φ a = Ω ab , and w is a constant. The Euler-Lagrange equations are which gives, with x a , a = 1, . . . , n denoting the components of X ∈ R n , ...
as Ω ab is invertible.
Proposition 7.2. The Poisson structure on T 2 (R n ) with n even, and coordinates (x a , U a , A a ) induced by the Lagrangian (7.29) is Proof. In order to construct the canonical formalism, rewrite (7.29) as a first order Lagrangian with constraints This gives rise to 3n constraints P a − λ a = 0, ψ a ≡ R a − 1 2 Ω ab U b = 0, S a = 0. (7.33) Impose these constraints in the 6n-dimensional phase-space with coordinates (x, P, U, R, λ, S) with the Poisson structure (and all other brackets vanishing), and compute the Dirac brackets [7] on the 3n-dimensional reduced phase space T M ⊕ T * M with coordinates (x, U, P).
Modifying {U a , U b } = 0 to the Dirac bracket gives where C cd ≡ {ψ c , ψ d } = Ω cd and (C −1 ) dc = Ω cd , where Ω ab Ω ac = δ b c . We also find {U a , ψ c } = δ a c , and finally (dropping * ) {x a , P b } = δ a b , {U a , U b } = Ω ab . The Poisson structure (7.31) does not generalise to curved spaces, as the Jacobi identity is obstructed by the Riemann curvature of g. It is nevertheless possible to make contact with the first integrals of the conformal geodesic equations, and the conformal Killing-Yano tensors under the additional assumption that w 2 = |A| 2 . In this case the Hamiltonian vector field corresponding to (7.32) is given by Therefore Q is is constant along the conformal geodesics, in agreement with the results of [18] and [14].

Conclusions
Conformal geodesics are examples of distinguished curves in parabolic geometries [6,8,12,17]. They also arise naturally in General Relativity as a tool in studying the proprieties of conformal infinity [11,15,2]. Despite their importance, there are few explicit examples known [18,10], and the underlying mathematical theory is not nearly as well developed as that of geodesics. In this paper we have proposed a variational formulation of the conformal geodesic equations. We hope that this will shed light on the integrability properties of these equations [14], as well as the global problems such as trapping and spirals in conformal geometry [5].