I want to find a curve $\gamma : [0,T] \to \mathbb{R}^n$ that minimizes the following functional

$$J[\gamma]=\int\limits_0^T F(\gamma(t),\gamma'(t))\,dt$$

Suppose I know the endpoints, $a, b \in \mathbb{R}^n$, i.e., $\gamma(0) = a$ and $\gamma(T) = b$. For simplicity, let

$$F(\gamma(t),\gamma'(t)) = ||\gamma'(t)||_2^2 + H(\gamma(t))$$

where $H$ is smooth but highly non-linear.

Computationally, how would I obtain a good $\gamma$?

I am aware of a number of non-linear optimization techniques that can do this (e.g., evolutionary computations), essentially by (smart) brute force. However, this has a nice variational form. Is there some form of variational gradient descent that makes sense here?

The first variation is (or its $i$th component at least):

$$\frac{\delta J}{\delta \gamma_i} = \frac{\partial F}{\partial \gamma_i} - \frac{d}{dt}\frac{\partial F}{\partial \gamma_i'} = \frac{\partial }{\partial \gamma_i}H(\gamma(t)) - \frac{d}{dt}\frac{\partial }{\partial \gamma_i'} \sum_j\gamma_j'(t)^2 = \frac{\partial }{\partial \gamma_i}H(\gamma(t)) - 2\gamma_i''(t)$$

So that

$$\frac{\delta J}{\delta\gamma} = \nabla H(\gamma(t)) - 2\gamma''(t)$$

So now, for instance, suppose I parameterize $\gamma$ by $\theta \in \mathbb{R}^m$, so the curve is $\gamma_\theta(t)$. Not sure what a good parameterization is; maybe the position of some intermediate points we can connect via splines.

On the other hand, I suppose I could ignore the variational stuff and do:

$$\begin{align}\frac{\partial J}{\partial \theta} &= \int\limits_0^T \frac{\partial }{\partial \theta} F(\gamma_\theta(t),\gamma'_\theta(t)) dt \\ &= \int\limits_0^T \frac{\partial }{\partial \theta}||\gamma'(t)||_2^2 + \sum_i \frac{\partial H}{\partial \gamma_i} \frac{\partial\gamma_i}{\partial\theta} dt\\ &= \int\limits_0^T 2\sum_j \gamma_j'(t) \frac{\partial \gamma_j'(t)}{\partial \theta} + \sum_i \frac{\partial H}{\partial \gamma_i} \frac{\partial\gamma_i}{\partial\theta} dt\end{align}$$

with gradient descent. Any suggestions?


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.