Maksim Samoilov

Some note: when linear system has more than one solution, it has a system of basic solutions, and any other solution is a linear combination of elements from that system....

Some other note: this problem of non-linear optimization brings us to (n+m)-dimensional full linear system which has an unique solution. For the theory, see http://www.machinelearning.ru/wiki/images/8/81/MOMO12_ipm.pdf (Russian) and http://cbio.ensmp.fr/~jvert/teaching/2006insead/slides/6_algo2/algo2.pdf (English) So,...

Found a paper on Stanford site: http://see.stanford.edu/materials/lsoeldsee263/08-min-norm.pdf In brief, for full-rank matrices there is a direct expression for minimum norm solution as `a_ln = A*(A*A)^-1 b`. So, we can implement...

Done it.

The problem is to write two methods: one for integer exponents (which could be realized by simple multiplying) and one for floating (which is a huge problem, as we can...

We can change `BigDecimal` to `double` but it may cause loss of precision (if I remember correctly, thats why we used `BigDecimal` class in current implementation).

Correct. This designation means matrix series like standard representation for `e^x`.

The current implementations of some methods, i.e. eachNonZero, need all non-zero values in sparse entry. Are you sure it is worth while?

I have some idea about GC: instead of `gc()` in `SparseVector`/`SparseMatrix` we could implement common `gc()` in `LinearAlgebra`. So, we have to keep links to any `SparseVector` and `SparseMatrix` in...

Can `value != (double)0.0` and `value != 0.0` give different results?