关于Pontryagin's Maximum Principle极小化问题必要条件、应用及固定端点自由终止时间场景解释的技术问询

Hey Marco, I totally get where you're coming from—Pontryagin's Maximum Principle (PMP) can feel like a dense wall of math and jargon when you first start digging into minimization problems. Let's break this down into plain, actionable terms, no overly complicated formulations required.

First: What PMP Solves (Quick Context) #

PMP is your go-to tool for optimal control problems: you want to find a control function $u(t)$ that minimizes (or maximizes) a cost function, while respecting the dynamic rules of your system (state equations) and any constraints on controls/states. For minimization problems, we just tweak the core conditions slightly from the maximization version you might see elsewhere.

Necessary Conditions for Minimization Problems (Simplified) #

Think of these as the "rules" any optimal solution must follow. Let's define the basic setup first, then walk through each condition:

• Your system follows state equations: $\dot{x}(t) = f(x(t), u(t), t)$ (how your state $x$ changes over time based on control $u$)
• Your total cost to minimize: $J = \phi(x(T)) + \int_{t_0}^T L(x(t), u(t), t) dt$ (a terminal cost $\phi$ at time $T$, plus an integral of running costs $L$ from start $t_0$ to end $T$)
• $\lambda(t)$ = "costate" vector (think of it as a dynamic Lagrange multiplier that tracks the marginal cost of changing each state variable)

1. Build the Hamiltonian Function #

First, construct the Hamiltonian $H$—this is the core function PMP uses to link costs, states, controls, and costates:

H(x, u, λ, t) = L(x, u, t) + λ(t)^T f(x, u, t)

For minimization, we don't flip the sign here (unlike some maximization formulations)—keep this in mind for the next step.

2. Costate (Adjoint) Equation #

The costate vector follows its own dynamic equation, which tells you how the marginal cost of state variables changes over time:

dot(λ(t)) = -∂H/∂x (x(t), u(t), λ(t), t)

In plain terms: the rate of change of each costate is the negative of how the Hamiltonian changes when you tweak the corresponding state variable.

3. Optimal Control Condition #

At every moment $t$, the optimal control $u^*(t)$ must minimize the Hamiltonian function (this is the key difference from maximization problems where you maximize $H$):

u*(t) = argmin_{u ∈ U} H(x*(t), u, λ*(t), t)

Here, $U$ is the set of all allowed control values (e.g., if your control is a throttle, $U$ might be 0 to 100%). This means for the current state and costate, you pick the control that makes $H$ as small as possible.

4. Terminal (Transversality) Conditions #

These depend on whether your endpoints (start/end states) and end time $T$ are fixed or free. We'll dive into your specific case next.

How to Use These Conditions in an Optimization Problem (Step-by-Step) #

1. Define your problem clearly: Write down your state equations, cost function, control/state constraints, and initial conditions ($x(t_0) = x_0$).
2. Construct the Hamiltonian: Use the formula from condition 1 above.
3. Write the costate equations: Take partial derivatives of $H$ with respect to each state variable, add a negative sign, and set equal to $\dot{\lambda}$.
4. Find the optimal control: Minimize $H$ with respect to $u$—this might involve taking a derivative of $H$ with respect to $u$, setting it to zero, and solving for $u$ in terms of $x$ and $\lambda$. If there are control constraints (e.g., $u ≤ 5$), you might get a bang-bang control (switches between extreme values) instead of a smooth solution.
5. Apply terminal conditions:联立 the state equations, costate equations, and terminal conditions to solve for $x^(t)$, $\lambda^(t)$, $u^(t)$, and (if free) $T^$. For simple problems, you can get analytical solutions; for complex ones, you'll use numerical methods like shooting or collocation.

Fixed Endpoint + Free End Time: What Do the Conditions Say? #

Let's clarify the setup here:

• Fixed endpoint: Initial state $x(t_0) = x_0$ is fixed, and terminal state $x(T) = x_T$ is also fixed (e.g., you need to go from point A to point B exactly).
• Free end time: You can choose the optimal $T^*$ (e.g., you can decide when to arrive at point B to minimize cost).

For this scenario, you keep the first three conditions (Hamiltonian, costate equation, optimal control minimizes $H$), plus two additional terminal conditions:

1. Fixed terminal state constraint: $x(T^*) = x_T$ (you have to end up exactly at the specified state).
2. Hamiltonian equals zero at terminal time:

H(x*(T^*), u*(T^*), λ*(T^*), T^*) = 0

Why this? If $H$ wasn't zero at $T^$, you could adjust the end time slightly to reduce the total cost. For example, if $H(T^) > 0$, ending a little earlier would lower the integral cost; if $H(T^) < 0$, ending a little later would help. The optimal $T^$ is where this tradeoff is balanced, so $H(T^*) = 0$.

Quick Example #

Suppose you're driving from point A (fixed position/speed) to point B (fixed position/speed), and you want to minimize total fuel use with no fixed arrival time. Using PMP:

• States: $x_1$ = position, $x_2$ = speed; $\dot{x}_1 = x_2$, $\dot{x}_2 = a(u)$ (acceleration as a function of throttle $u$)
• Cost: $J = \int_{t_0}^T L(u) dt$ (fuel use rate $L(u)$)
• Hamiltonian: $H = L(u) + \lambda_1 x_2 + \lambda_2 a(u)$
• Costate equations: $\dot{\lambda}_1 = 0$ (so $\lambda_1$ is constant), $\dot{\lambda}_2 = -\lambda_1$ (so $\lambda_2 = -\lambda_1 t + C$)
• Optimal throttle $u^*$ minimizes $H$—if $L(u)$ is quadratic, this gives a smooth control linked to $\lambda_2$.
• Terminal conditions: $x_1(T^) = x_B$, $x_2(T^) = v_B$, and $H(T^*) = 0$. Solving these gives you the optimal speed profile, throttle control, and arrival time.

备注：内容来源于stack exchange，提问作者Marco Di Giacomo