While it seems common knowledge how to formulate necessary optimality conditions backed by the famous Pontryagin’s Maximum Principle, the direct derivation from the very general original formulations is not obvious. Sometimes, authors like me¹ provide some plausibility like obviously this term has to vanish, refer to other sources that miss the derivation too, or attach the complete book by Pontryagin to a specific formula².

In this post, I show

• for the example of a 2D pendulum,
• how a standard dynamic optimization problem is brought into the form for which Pontryagin’s original results apply
• how the optimality conditions are derived directly from Pontryagin’s theorems
• and under which particular choices of bases, they become the set of equations that are typically referred to in the literature and on wikipedia.

The Model

A Unit Pendulum

Consider the pendulum subjected to a potential force that can be adjusted. That is, instead of the gravity

we consider a force field

with scalar functions $g_1$, $g_2$ of time that can act as a control.

With mass $m$, length $L$, and angle $\theta$ (see the picture below), we define the momentum as

This gives the Hamiltonian

or

and the equations of motion as

$\begin{aligned} \quad \quad \quad \dot \theta &= \frac{p}{mL^2}, \\\\\\\\ \dot p &= -mg_1L\cos \theta - mg_2L\sin \theta. \end{aligned}$

Here is a schematic illustration taken from scholarpedia:

Control Setups

With consider two control setups:

Final State Fixed

Fix a time $t_1$, start with
$\quad(\theta(t_0), p(t_0))=(\pi, 0)$,
i.e. pendulum standing up-side-down, and minimize the control effort
$\quad \mathcal J:= \frac{1}{2}\int_{t_0}^{t_1} g_1^2 + g_2^2 ~ dt \to \min$
over all $(g_1, g_2)$ that lead to
$\quad(\theta(t_1), p(t_1))=(0,0)$,
i.e. pendulum hanging down.

Final State as Optimization Target

Fix a time $t_1$, start with
$\quad(\theta(t_0), p(t_0))=(\pi, 0)$,
i.e. pendulum standing up-side-down, and find a control that targets
$\quad (\theta(t_1), p(t_1))=(0,0)$,
i.e. pendulum hanging down while also minimizing the control effort:
$\quad \mathcal J:= \frac{1}{2}(\theta(t\_1)^2 + p(t\_1)^2) + \frac{1}{2}\int_{t_0}^{t_1} g_1^2 + g_2^2 ~ dt \to \min$.

The Maximum Principle

Classically, as treated by Pontryagin itself³ (see the title image for the original statement), one seeks the minimum over all $u$ (here, we have $u=(g_1, g_2)$) that a-priori ensure that the initial value is transferred into the target value (here: ensure that $(\theta(t_1), p(t_1))=(0, 0)$.

In other words, this means that suitable controls are known and that the optimization looks for those with the least magnitude. For people like me, that use optimization because they don’t know a control, this seems not very feasible. But Pontryagin has the answer for that too:

Maximum Principle with Variable Endpoints

One can (partially) abandon the end conditions, see Chapter I.7 in Pontryagin’s book. Similarly, one can abandon the initial conditions which, again, seems odd from an application point of view. Also, simply omitting the end conditions will lead to trivial solutions. Certainly, we need to include them in the optimization. Which leads to the Wikipedia case.

The Wikipedia Case

Commonly used and illustrated on wikipedia:Pontryagin’s maximum principle – the form that leaves the controls free (in the sense that the trajectory does not need to end in the target state) but instead penalizes the endpoint.

Now, let me show how this form derives from Pontryagin’s book.

Connecting Wikipedia and Pontryagin

The derivation of the Wikipedia formulas from the Maximum Principle with Variable Endpoints goes as follows:

1. Transform (see, e.g., Enc. of Math.: Bolza problem) the problem from Mayer form to Lagrange form.

2. Put the resulting system into the form of Pontryagin.

3. Apply the maximum principle with completely free endpoints.

For the pendulum optimal control problem

subject to

with initial conditions

this means:

I. Lagrange form

We get rid of the costs that are put on the terminal values by introducing a variable $\tilde x$. Thus the optimal control problem now reads:

subject to

with one terminal condition

and the old initial conditions

In fact, since $\tilde x$ is constant and fixed by its end condition, one obtains that $\int_{t_0}^{t_1} \tilde x \;dt = (t_1 - t_0)\tilde x(t_1) = \frac{1}{2}(\theta(t_1)^2+p(t_1)^2).$

Once in Lagrange form, the problem can be put into the form used in Pontryagin’s Theorem 1:

… where in particular the bold face $\mathbf x := (x^0, x)$ is the original state $x$ augmented with the integrated running cost $x^0$, so that $\mathbf x$ passes through $\Pi$ at time $t_1$ means feasibility and optimality namely

• that $x$ at $t_1$ assumes the prescribed value $x_1$ and
• that $x^0(t_1)=\int_{t_0}^{t_1}f(x(t), u(t))\;dt$ is a minimum.

Hence, we arrive at the somewhat involved but plausible and in line with the theory formulation in:

II. Pontryagin’s formulation

subject to

with initial conditions

and the end condition

Here, the trick is that the new variable $x^0$ simply integrates the value of the cost functional.

The corresponding adjoint system for the $\psi$ reads⁴

This $\psi$ defines the control Hamiltonian as

III. Pontryagin with variable endpoints

Finally, we can omit parts of the end conditions.

If we let the state $(\tilde x, \theta, p)$ at $t_1$ take on an arbitrary value then it will belong to the hypersurface

Then the transversality condition⁵ that $\psi(t_1)$ has to be orthogonal to the tangent plane to $S_1$ at $(\tilde x(t_1), \theta(t_1), p(t_1))$ gives the following boundary conditions for $\psi$:

These give the Wikipedia terminal condition for the adjoint state with the choice of $\psi_1(t_1) = t_1-t_0$.

Note that these particular end conditions were derived from the particular basis

of the tangent space and note that any other basis can be taken, or equivalently, any linear combination of the boundary conditions are feasible.

References

Theory and the screenshots are from 📙 Pontryagin, Boltyanski, Gamkrelidze, and Mishchenko (1962) The Mathematical Theory of Optimal Processes