Why has nobody yet suggested spalloc, i.e., a tool designed to allocate memory for a sparse matrix! That is essentially what was requested. spalloc creates an all zero sparse matrix (like sparse can do), but with space allocated for as many elements as you specify.
Regardless, it is a TERRIBLE idea to build a sparse matrix incrementally, even if it is already allocated as such. Stuffing elements into the matrix inside a loop, even if the space is already allocated, will cause MATLAB to re-order the elements. This will STILL take a lot of time to accomplish, unless you are careful enough to add the elements in the proper order. Even so, my guess is there is still a time penalty.
Far better? Create the matrix in ONE step, using sparse. Do so by creating a list of the elements, as well as their row and column indexes in advance. Only when you are all done do you call sparse. If the matrix is representable as a set of diagonals, then learn to use spdiags. Again, only one call.
Be careful of course. Do not build up that list of elements incrementally either, appending them to the end of an array. That also has quadratic time to create that list, since MATLAB must reallocate the array after each append step. If you know how many elements will be created, then preallocate the memory for the list. Then insert your elements one at a time into the list.
If you do not know how many final elements are in the list, then use a tool for incremental array growth like those I have posted here on the file exchange.
In general, build your sparse matrices in one call, or at least as few calls to sparse as possible. There are some nice tricks you can use, some of which you would learn from tools like inpaint_nans , or blktridiag.
