How can a Lagrange-Newton method solve optimal control problems on manifolds?

Why linear-space optimal control tools do not transfer

In PDE-constrained optimization, a control is applied inside a partial differential equation, for example as a force field, to drive the PDE solution toward a desired configuration. The objective usually balances control cost against the deviation of the controlled state from that desired configuration. The classical theory for such problems in linear spaces is described as well established: existence of optimal controls, first-order KKT and second-order optimality conditions, and a rich array of numerical methods.

The paper addresses a setting where that theory cannot be applied verbatim. The PDE is a geometric variational equation, meaning its solutions are mappings into a nonlinear manifold, and the state space therefore has no global linear structure. The authors note that such problems appear in continuum mechanics, stationary action problems, elasticity theory and geometry processing, image and shape analysis, liquid crystal modeling, and micromagnetics. A motivating example is actuating an elastic rod or shell into a target shape by an external force field.

The paper also positions itself against prior work. Geometric optimal control has classically treated finite-dimensional systems governed by ordinary differential equations on manifolds, formulating the Pontryagin maximum principle intrinsically via Lie brackets and Hamiltonian flows, and more recently using the Riemannian metric as a control design tool. In infinite dimensions, a Riemannian view has been fruitful in shape optimization, where the shape is treated as a point on a manifold and Lagrange-Newton methods have been developed accordingly. Equality-constrained optimization on manifolds has also been extended to SQP-type methods on Hilbert manifolds and to manifold-valued constraints. The authors state that extending PDE-constrained optimization to the geometric setting is not straightforward, because the state and adjoint state are coupled in a subtle way.

Three coupled equations define the optimality system

The paper formulates a Lagrangian function corresponding to the geometric optimal control problem and derives the resulting first-order optimality conditions. Those conditions comprise three equations: the state equation, an adjoint equation, and a stationarity equation. This optimality system, not the original minimization problem alone, is the object the numerical method is built to solve.

All quantities are derived in weak form, which the authors say allows implementations in terms of manifold-valued finite elements. The state and adjoint state are coupled together in a subtle way, and the paper's stated aim is to preserve, as far as possible, the structural elegance of the classical Lagrangian approach despite the absence of a global linear structure on the state space.

The main difficulty the authors identify is finding a concise and reliable calculus to compute the Newton linear operator, including the quantities related to the curvature of the underlying manifold. That calculus, rather than the derivation of the optimality conditions alone, is where the technical contribution concentrates.

Local splitting makes the Newton step implementable

Building on the optimality system, the authors devise a Lagrange-Newton method for its numerical solution. The implementable algorithm uses a local splitting technique. At each iterate, the method builds a locally splitted Lagrangian by composing the Lagrangian with a vector forward transport, then solves a linear equation for the direction in the state, adjoint, and control components.

The update rule is not a plain addition. The state component is updated through a retraction on the state manifold, the adjoint component is transported by a vector forward transport evaluated at the retracted state before the adjoint increment is added, and the control component is updated additively. The algorithm requires an initial guess in the product space, a vector forward transport on the adjoint bundle, a vector forward transport on the tangent bundle of the state manifold, and a retraction on the state manifold.

The paper is explicit that this construction is not obviously a Newton method. The splitted Lagrangian depends on the current iterate, so the problem being solved appears to change at every step, and the mapping used for the update is stated not to be a retraction in general. The authors therefore set out to prove that the iteration is nevertheless equivalent to a Newton update under suitable choices of connection and retraction.

The splitted iteration is Newton's method in disguise

The central theoretical result is an equivalence. The authors introduce a vector forward transport on the product of the adjoint bundle and the control space, constructed from the local splitting transport and a vector forward transport on the tangent bundle of the state manifold. With this transport, they show that the covariant derivative of the splitted Lagrangian equals the covariant derivative of the original Lagrangian along the corresponding directions.

They then show that a certain mapping defines a retraction on the adjoint bundle, built from the retraction on the state manifold and the vector forward transport. Combining these two results yields the main theorem: the iterates generated by the splitted algorithm coincide with the iterates generated by Newton's method applied to the derivative of the Lagrangian, provided the connection on the adjoint bundle is induced by the constructed vector transport and the retraction on the adjoint bundle is the one defined in the paper.

The convergence statement follows from that equivalence. Under the assumptions that the derivative of the Lagrangian is Newton-differentiable at a zero and that the operator norm of the inverse Newton operator is uniformly bounded by some finite constant, the iterates of the algorithm converge locally superlinearly to that zero. The paper also notes that the second derivative of the Lagrangian is symmetric because the Lagrangian is a twice continuously differentiable function from a manifold into the real numbers, a structural property the authors describe as beneficial when constructing numerical algorithms.

A practical consequence of the equivalence is that neither the connection on the adjoint bundle nor the retraction on the adjoint bundle needs to be computed explicitly for the implementation. The authors state this directly: these quantities are not needed explicitly in the algorithm.

Only the lower triangle of the Newton operator is assembled

To implement the method, a representation of the second covariant derivative of the splitted Lagrangian at the current iterate is required. The paper derives expressions for the partial derivatives of the splitted Lagrangian with respect to the state, adjoint, and control variables and evaluates them at the current iterate.

Because the linear Newton operator is symmetric, it suffices to compute its lower triangle and obtain the remaining entries by taking adjoints, or transposes after discretization. This is a concrete implementation economy that follows from the symmetry property noted above, and it matters because the operator involves curvature-related quantities of the underlying manifold.

The paper states that despite the high complexity of the problem, all quantities can be derived from relatively simple algorithmic ingredients that are available for most standard manifolds and admit an efficient implementation. The authors also state that they concentrated on the detailed description of the overall framework and the derivation of Newton's method, and that aspects concerning discretization with finite elements, including details of the necessary constructions, are planned for a companion paper.

Numerical examples cover geodesics, surfaces, rods, and shells

The theoretical framework is illustrated by numerical examples, including the optimal control of elastic geodesics and surfaces of minimal energy by force fields, and the control of inextensible rods and shells. These examples are demonstrations of the framework rather than a benchmark comparison against alternative solvers.

In the inextensible rod example, the state is a triplet on a product manifold combining a vector field, a field taking values in a Stiefel manifold, and a Lagrange multiplier field, while the control is simply the applied force field. The aim is to minimize the integral over the squared distance between the points of the rod and a fixed desired state, balanced against control cost. The paper reports that the rod winds itself around the desired state when the control cost is small, and shows solutions for three regularization parameters, 10^-5, 10^-7, and 10^-9, with one-sided Dirichlet boundary conditions.

For the rod, the authors use linear Lagrange elements for the state, the Stiefel-valued field, the control, and their adjoint counterparts, and piecewise constant elements for the multiplier and its adjoint on a uniform partition of the unit interval. Because the problem is highly nonlinear, particularly for small regularization parameters, they use a simple homotopy method that decreases the regularization parameter each iteration, starting from 0.1 and reducing it successively by a factor of (0.1)^(1/3) per homotopy step.

For the shell case, the controlled object is a strip of length 5 and width 1, clamped horizontally at one of the short edges, and the aim is to bend the strip upwards so that its integrated square distance to the horizontal plane at height 0.4 is minimized, balanced with control cost. The paper shows the computed result for a regularization parameter of 10^-3, including the bending forces. Discretization uses Crouzeix-Raviart finite elements for the state and its adjoint, linear finite elements augmented by a bubble function for the Stiefel-valued field and its adjoint, and piecewise constant elements for the multiplier and its adjoint. This problem was solved by a damped Newton method with an affine covariant damping strategy for globalization.

Observed convergence shows an active globalization phase first

For the shell example, the paper reports the convergence history of the algorithm and the computed damping factors. The observed pattern is an initial phase in which globalization is active, followed by local superlinear convergence. The convergence plot shows the norm of the combined state, adjoint, and control increment across iterations, and the damping factors are reported as values in the interval from 0 to 1.

This observed behavior is consistent with the theoretical result, which is a local superlinear convergence statement. The theory does not claim global convergence, and the numerical evidence shows that a globalization strategy was needed before the local phase was reached. Readers comparing this method with other approaches should therefore treat the superlinear rate as a local property, conditional on the Newton-differentiability and uniform boundedness assumptions, and as observed in these examples only after damping became inactive.

The paper does not report iteration counts, timings, or accuracy comparisons against alternative discretizations or solvers, so the numerical section should be read as an illustration of the framework rather than as evidence of computational superiority.

Open questions remain before broader use

The authors describe the work as a first step into the algorithmic treatment of problems in geometric PDE-constrained optimization. They state that a number of important theoretical and algorithmic issues remain open and subject to future research. First, a functional analytic treatment of concrete classes of problems needs to be done, and in particular the abstract assumptions, such as the differentiability assumptions, need to be verified for appropriate classes of problems.

A related open question is the integration of inequality constraints into the framework. The authors also state that algorithms more specifically constructed for constrained optimization on manifolds, refining existing SQP-type ideas, need to be developed. Discretization details with finite elements are deferred to a companion paper.

These are the boundaries that matter for a reader deciding whether to build on the method. The equivalence theorem and the local superlinear convergence result are established at the level of the abstract framework, and the numerical examples demonstrate the framework on elastic geodesics, minimal-energy surfaces, inextensible rods, and shells. Applying the method to a new problem class requires verifying the differentiability and boundedness assumptions for that class, and problems with inequality constraints are outside the current scope.

The method's value is a computable Newton step on manifolds

The most decision-relevant contribution of this paper is not a new class of optimal control problems but a way to compute a Newton step for the optimality system of a geometric optimal control problem using only ingredients that are available for most standard manifolds. The local splitting and the moving test function technique reduce the second covariant derivative to partial derivatives of a splitted Lagrangian, and symmetry means only the lower triangle needs to be assembled.

The equivalence theorem is what makes this reduction trustworthy: the splitted iteration is not merely inspired by Newton's method, it coincides with Newton's method applied to the derivative of the Lagrangian under the specified connection and retraction. The practical payoff is that the connection and retraction on the adjoint bundle never have to be formed explicitly.

For a reader weighing whether to adopt this approach, the relevant split is between problems that fit the framework's assumptions and problems that do not. If the state is a mapping into a nonlinear manifold, the constraint is a geometric variational equation, the derivative of the Lagrangian is Newton-differentiable at the solution, and the inverse Newton operator is uniformly bounded, the paper provides both the optimality system and an implementable, locally superlinearly convergent solver. If the problem requires inequality constraints, or if the differentiability and boundedness assumptions have not been verified for that problem class, the paper's own discussion places it in future work rather than in the demonstrated scope.

Sources used in this answer