## Section: Research Program

### Optimal Control and its Geometry

Let us detail our research program concerning optimal control. Relying on Hamiltonian dynamics is now prevalent, instead of the Lagrangian formalism in classical calculus of variations. The two points of view run parallel when computing geodesics and shortest path in Riemannian Geometry for instance, in that there is a clear one-to-one correspondance between the solutions of the geodesic equation in the tangent bundle and the solution of the Pontryagin Maximum Principle in the cotangent bundle. In most optimal control problems, on the contrary, due to the differential constraints (velocities of feasible trajectories do not cover all directions in the state space), the Lagrangian formalism becomes more involved, while the Pontryagin Maximum Principle keeps the same form, its solutions still live in the cotangent bundle, their projections are the extremals, and a minimizing curve must be the projection of such a solution.

**Cut and conjugate loci.**
The cut locus —made of the points where the extremals lose optimality— is obviously crucial in optimal control, but usually out of reach
(even in low dimensions), and anyway does not have an analytic characterization because it is a non-local object. Fortunately, conjugate
points —where the extremals lose *local* optimality— can be effectively computed with high accuracy for many control systems.
Elaborating on the seminal work of the Russian and French schools (see [76], [35], [36] and
[53] among others), efficient algorithms were designed to treat the smooth case.
This was the starting point of a series of papers of members of the team culminating in the outcome of the *cotcot* software
[46], followed by the *Hampath* [55] code.
Over the years, these codes have allowed for the computation of conjugate loci in a wealth of situations including applications to space
mechanics, quantum control, and more recently swimming at low Reynolds number.
With in mind the two-dimensional analytic Riemannian framework, a heuristic approach to the global issue of determining cut points is to
search for singularities of the conjugate loci; this line is however very delicate to follow on problems stemming from applications in three
or more dimensions (see *e.g.* [56] and [43]).
In all these situations, the fundamental object underlying the analysis is the curvature tensor. In Hamiltonian terms, one considers the
dynamics of subspaces (spanned by Jacobi fields) in the Lagrangian Grassmannian [33].
This point of view withstands generalizations far beyond the smooth case: In ${\mathrm{L}}^{1}$-minimization, for instance, discontinuous curves
in the Grassmannian have to be considered (instantaneous rotations of Lagrangian subspaces still obeying symplectic rules
[60]).
The cut locus is a central object in Riemannian geometry, control and optimal transport.
This is the motivation for a series of conferences on “The cut locus: A bridge over differential geometry, optimal control, and
transport”, co-organized by team members and Japanese colleagues, the
next one should take place in Nice in 2020.

**Riemann and Finsler geometry.**
Studying the distance and minimising geodesics in Riemannian Geometry or Finsler Geometry is a particular case of optimal
control, simpler because there are no differential constraints; it is studied in the team for the following two reasons.
On the one hand, after some tranformations, like averaging
(see section 3.2)
or reduction, some more difficult optimal control problems lead to a Riemann or Finsler
geometry problem. On the other hand, optimal control, mostly the Hamiltonian setting, brings a fresh viewpoint on problems in Riemann and Finsler geometry.
On Riemannian ellipsoids of revolution, the optimal control approach
allowed to decide on the convexity of the injectivity domain, which, associated with non-negativity of the
Ma-Trudinger-Wang curvature tensor, ensures continuity of the optimal transport on
the ambient Riemannian manifold [64], [63].
The analysis in the oblate geometry [44] was completed in [59] in the
prolate one,
including a preliminary analysis of non-focal domains associated with conjugate loci.
Averaging in systems coming from space mechanics control (see sections 3.2 and 4.1) with ${\mathrm{L}}^{2}$-minimization
yields a Riemannian metric, thoroughly computed in [42] together with its geodesic flow; in reduced dimension, its conjugate
and cut loci were computed in [45] with Japanese Riemannian geometers.
Averaging the same systems for minimum time yields a Finsler Metric, as noted in [41].
In [51], the geodesic convexity properties of these two types of metrics were compared.
When perturbations (other than the control) are considered, they introduce a “drift”, *i.e.* the Finsler metric is no longer symmetric.

**Sub-Riemannian Geometry.**
Optimal control problems that pertain to sub-Riemannian Geometry bear all the difficulties of optimal control, like the role of singular/abnormal trajectories, while having some useful structure. They lead to many open problems, like smoothness of minimisers, see the recent monograph [69] for an introduction. Let us detail one open question related to these singular trajectories: the Sard conjecture in sub-Riemannian geometry.
Given a totally non-holonomic distribution on a smooth manifold, the Sard Conjecture is concerned with the size of the set of points
that can be reached by singular horizontal paths starting from a given point. In the setting of rank-two distributions in dimension three,
the Sard conjecture is that this set should be a subset of the so-called Martinet surface, indeed small both in measure and in dimension.
In [39], it has been proved that the conjecture holds in the case where the Martinet surface is smooth.
Moreover, the case of singular real-analytic Martinet surfaces was also addressed. In this case, it was shown that the Sard Conjecture holds
true under an assumption of non-transversality of the distribution on the singular set of the Martinet surface.
It is, of course, very intersting to get rid of the remaining technical assumption, or to go to higher dimension.
Note that any Sard-type result has strong consequences on the regularity of sub-Riemannian distance functions and in turn on optimal transport problems in the sub-Riemannian setting.

**Small controls and conservative systems, averaging.**
Using averaging techniques to study small perturbations of integrable Hamiltonian systems is as old an idea as celestial mechanics.
It is very subtle in the case of multiple periods but more elementary in the
single period case, here it boils down to taking the average of the perturbation along each periodic orbit [37], [75].
This line of research stemmed out of applications to space engineering (see section 4.1): the control of the
super-integrable Keplerian motion of a spacecraft orbiting around the Earth is an example of a slow-fast controlled system.
Since weak propulsion is used, the control itself acts as a perturbation, among other perturbations of similar magnitudes: higher order
terms of the Earth potential (including ${J}_{2}$ effect, first), potential of more distant celestial bodies (such as the Sun and the Moon),
atmospheric drag, or even radiation pressure.
Properly qualifying the convergence properties (when the small parameter goes to zero) is important and is made difficult by the presence of control.
In [41], convergence is seen as convergence to a differential inclusion;
this applies to minimum time; a contribution of
this work is to put forward the metric character of the averaged system by yielding a Finsler
metric (see section 3.2).
Proving convergence of the extremals (solutions of the Pontryagin Maximum Principle) is more intricate.
In [58], standard averaging ( [37], [75])
is performed on the minimum time extremal flow after carefully identifying slow
variables of the system thanks to a symplectic reduction. This alternative approach
allows to retrieve the previous metric approximation, and to partly address the
question of convergence. Under suitable assumptions on a given geodesic of the
averaged system (disconjugacy conditions, namely), one proves existence of
a family of quasi-extremals for the original system that converge towards the geodesic
when the small perturbation parameter goes to zero.
This needs to be improved, but convergence of all extremals to extremals of an “averaged Pontryagin Maximum Principle” certainly fails.
In particular, one cannot hope for ${C}^{1}$-regularity on the value function when the small
parameter goes to zero as swallowtail-like singularities due to the
structure of local minima in the problem are expected.
(A preliminary analysis has been made in [57].)

**Optimality of periodic solutions/periodic controls.**
When seeking to minimize a cost with the constraint that the controls and/or part of the
states are periodic (and with other initial and final conditions), the notion of conjugate
points is more difficult than with straightforward fixed initial point.
In [48], for the problem of optimizing the efficiency of the
displacement of some micro-swimmers (see section 4.3) with
periodic deformations, we used the sufficient optimality conditions established by
R. Vinter's group [80], [66] for systems with non unique
minimizers due to the existence of a group of symmetry (always present with a periodic
minimizer-candidate control).
This takes place in a long term collaboration with P. Bettiol (Univ. Bretagne Ouest) on
second order sufficient optimality conditions for periodic solutions, or in the presence
of higher dimensional symmetry groups, following [80], [66].
Another question relevant to locomotion is the following. Observing animals (or humans), or numerically solving the optimal control problem associated with driftless micro-swimmers for various initial and final conditions, we remark that the optimal strategies of deformation seem to be periodic, at least asymptotically for large distances.
This observation is the starting point for characterizing dynamics for which some optimal solutions are periodic, and asymptotically attract other solutions as the final time grows large; this is reminiscent of the “turnpike theorem” (classical, recently applied to nonlinear situations in [79]).

**Software.** These applications (but also the development of theory where numerical experiments can be very enlightening) require many algorithmic and numerical developments that are an important side of the team activity.
The software *Hampath* (see section 6.1) is maintained by former members of the team in close collaboration with McTAO.
We also use direct discretization approaches (such as the Bocop solver developed by COMMANDS) in parallel.
Apart from this, we develop on-demand algorithms and pieces of software, for instance we have to interact with a production software developed by Thales Alenia Space.
A strong asset of the team is the interplay of its expertise in geometric control theory with applications and algorithms (see sections 4.1 to 4.3) on one hand, and with optimal transport, and more recently Hamiltonian
dynamics, on the other.
In 2019, the ADT `ct` (Control Toolbox) has started with a first sprint in
"AMDT mode" with Sophia SED during spring 2019. In addition to McTAO, researchers from
the CAGE team (Inria Paris) and the APO team (CNRS Toulouse) are involved. The idea is
to put together the efforts on BOCOP and HamPath to go towards a reference
toolbox in optimal control. After the first sprint cycle (24 months being planned on
the whole action), some starting points have been addressed including: continuous
integration for BOCOP and HamPath, refresh on collaborative development tools, first
steps of software refactoring, first test of a high-end interface (throught scripting,
notebooks, or an *ad hoc* GUI). The next sprint is planned during spring 2020.