What are tasty sushi combinations?

If the authors of a paper do not specify otherwise, when they transfer copyright of a paper to a publisher, that includes its images. That's what most people do.

But, during the copyright transfer process, an author can specify

• if there are original images of independent value that they'd like to retain copyright to, or
• if there are images whose copyright is held by a third party, which they have a license to reuse.

See for example this ACM copyright transfer form.

If you are creating images for papers on which you are not an author, you can retain copyright of your images and just give a license to use in the paper. The authors will have to specify on submission that they are using third party images, and show that they have permission from you to use them.

If you are creating images for your own paper, you may be able to claim them as an "artistic" image and retain the copyright, depending on the publisher.

We can reformulate the problem as follows: given vectors $u_1,\ldots,u_n,v_1,\ldots,v_n \in \mathbb{R}^m$, find $\max | \langle u_i,v_j \rangle |$ over all $i,j \in [n]$.

Now, if an approximate answer is good enough, I think that for small $m$ an approximate answer with multiplicative error of $1+\epsilon$ can be found in time $\approx n m + (1/\epsilon)^m$. Take an $\epsilon$-net of the $m$-dimensional sphere (that is, a set of points on the sphere such that each other point on the sphere is $\epsilon$-close to one of them). Such a net has size $ (1/\epsilon)^{O(m)}$. Now, hash the vectors $u_1,\ldots,u_n$ to the sphere. Let $u'_i=u_i / \|u_i\|$ be the image of $u_i$. Find the point on the net closest to $u'_i$ (trivially this can be done in time linear in the size of the net, but probably there are some geometric algorithms to do this is constant or near constant time). For each point in the net store the vector close to it with maximal norm. Do the same for $v_1,\ldots,v_n$. To finish, go over all pairs of points in the net and compute the inner products of their associated vectors.

It has been a while since I did cross compilation, but I found this description of doing cross compilation quite clear, YMMV.

Depending on your specific target there might be more specific documentation available online.

I would set up two complete new trees (one for 3.4 and one for 3.8) for compiling the kernel and not use the apt-get command as it will not install these headers in the new trees. Any changes you would make in files installed with apt-get would run the risk of breaking some other package relying on them (you obviously don't use any of those now, but you might in the future).