TODO

- Sparse matrices
- Blocked or block-diagonal matrices
- Optimizations for the optimizer:
  	- Experiment with stochastic (MCMC-style / tempered) search 
	  algorithms
	- Add clever rewrite rules to reduce required search depth
          (could be done by looking for multistep rewrites that tend
          to occur on the solution path)
	- Optimizer should 'show its work' and return a list of
          rewrite rules applied
        - implement http://en.wikipedia.org/wiki/Matrix_chain_multiplication
- Other matrix operations (with appropriate optimizer rewrite rules):
  	- decompositions: eigenvalue, SVD, QR, LU
	- matrix derivatives
	- Kronecker and Hadamard (pointwise) products
	- "solve" functionality for implicitly defined variables
	  (e.g. solve for X,D in AX=XD, with D diag, should find 
	  eigendecomposition of A)
- Other matrix properties:
  	- unitary
  	- stochastic
	- rotation matrices (always commute with each other)
	- real/complex
	- singular / invertible
- Code generator should be smarter about parenthesis/order of
  operations.
- Explicit consideration of numeric stability (currently we just
  assume that fewer FLOPs is better). Give each matrix a condition
  number, then calculate stability of the resulting expression.
- Web interface for easy public use